diff --git a/31_HLSLPathTracer/app_resources/hlsl/next_event_estimator.hlsl b/31_HLSLPathTracer/app_resources/hlsl/next_event_estimator.hlsl index 64a06b16f..1597b145e 100644 --- a/31_HLSLPathTracer/app_resources/hlsl/next_event_estimator.hlsl +++ b/31_HLSLPathTracer/app_resources/hlsl/next_event_estimator.hlsl @@ -182,9 +182,7 @@ struct ShapeSampling const vector3_type tri_vertices[3] = {tri.vertex0, tri.vertex1, tri.vertex2}; shapes::SphericalTriangle st = shapes::SphericalTriangle::create(tri_vertices, ray.origin); sampling::ProjectedSphericalTriangle pst = sampling::ProjectedSphericalTriangle::create(st, ray.normalAtOrigin, ray.wasBSDFAtOrigin); - const scalar_type pdf = pst.backwardPdf(L); - // if `pdf` is NAN then the triangle's projected solid angle was close to 0.0, if its close to INF then the triangle was very small - return pdf < numeric_limits::max ? pdf : numeric_limits::max; + return pst.backwardWeight(L); } template @@ -252,6 +250,7 @@ template struct ShapeSampling { using scalar_type = T; + using vector2_type = vector; using vector3_type = vector; static ShapeSampling create(NBL_CONST_REF_ARG(Shape) rect) @@ -268,49 +267,58 @@ struct ShapeSampling matrix rectNormalBasis; vector rectExtents; rect.getNormalBasis(rectNormalBasis, rectExtents); + shapes::SphericalRectangle sphR0; sphR0.origin = rect.offset; sphR0.extents = rectExtents; sphR0.basis = rectNormalBasis; - scalar_type solidAngle = sphR0.solidAngle(ray.origin).value; - if (solidAngle > numeric_limits::min) - pdf = 1.f / solidAngle; - else - pdf = bit_cast(numeric_limits::infinity); - return pdf; + + // 1.f/0.f gives infinity no special checks needed + return 1.f / sphR0.solidAngle(ray.origin).value; } template vector3_type generateAndPdfAndWeight(NBL_REF_ARG(scalar_type) pdf, NBL_REF_ARG(scalar_type) weight, NBL_REF_ARG(scalar_type) newRayMaxT, NBL_CONST_REF_ARG(vector3_type) origin, NBL_CONST_REF_ARG(Aniso) interaction, NBL_CONST_REF_ARG(vector3_type) xi) { - const vector3_type N = rect.getNormalTimesArea(); - const vector3_type origin2origin = rect.offset - origin; - matrix rectNormalBasis; vector rectExtents; rect.getNormalBasis(rectNormalBasis, rectExtents); + shapes::SphericalRectangle sphR0; sphR0.origin = rect.offset; sphR0.extents = rectExtents; sphR0.basis = rectNormalBasis; - vector3_type L = hlsl::promote(0.0); + // sampling::SphericalRectangle ssph = sampling::SphericalRectangle::create(sphR0, origin); - if ( ssph.solidAngle > numeric_limits::min) + typename sampling::SphericalRectangle::cache_type cache; + + vector3_type L = hlsl::promote(0.0); + const bool FastVersion = true; + if (FastVersion) { - typename sampling::SphericalRectangle::cache_type cache; - const vector3_type localDir = ssph.generate(xi.xy, cache); - // not sure if generate() can produce NaN/inf when solidAngle > min - assert(!hlsl::any(hlsl::isinf(localDir) || hlsl::isnan(localDir))); - // transform local direction to world space - L = localDir.x * rectNormalBasis[0] + localDir.y * rectNormalBasis[1] + localDir.z * rectNormalBasis[2]; - pdf = ssph.forwardPdf(xi.xy, cache); - weight = ssph.forwardWeight(xi.xy, cache); + // actually the slowest + //L = ssph.generate(xi.xy, cache); + //newRayMaxT = ssph.computeHitT(L); + + // fastest + const vector3_type localL = ssph.generateNormalizedLocal(xi.xy,cache,newRayMaxT); + assert(!hlsl::any(hlsl::isinf(localL) || hlsl::isnan(localL))); + L = hlsl::mul(hlsl::transpose(ssph.basis),localL); } else - weight = bit_cast(numeric_limits::infinity); + { + L = ssph.generateUnnormalized(xi.xy,cache); + assert(!hlsl::any(hlsl::isinf(L) || hlsl::isnan(L))); + const scalar_type rcpLen = hlsl::rsqrt(hlsl::dot(L,L)); + newRayMaxT = 1.f / rcpLen; + L *= rcpLen; + } + // prevent self intersections against the emitter + newRayMaxT -= 0.0001f; - newRayMaxT = hlsl::dot(N, origin2origin) / hlsl::dot(N, L); + pdf = ssph.forwardPdf(xi.xy,cache); + weight = ssph.forwardWeight(xi.xy,cache); return L; } @@ -329,7 +337,6 @@ struct EffectivePolygonMethod NBL_CONSTEXPR_STATIC_INLINE NEEPolygonMethod value = PPM_SOLID_ANGLE; }; - // Projected solid angle NEE for rectangles using "Practical Warps": // bilinear warp over 4-corner NdotL + spherical rectangle sampling. // Same grazing-angle limitations as the triangle variant -- see comments @@ -359,21 +366,12 @@ struct ShapeSampling sphR0.extents = rectExtents; sphR0.basis = rectNormalBasis; sampling::ProjectedSphericalRectangle psr = sampling::ProjectedSphericalRectangle::create(sphR0, ray.origin, ray.normalAtOrigin, ray.wasBSDFAtOrigin); - // Reconstruct normalized [0,1]^2 position on the rectangle from the ray direction - const vector3_type N = rect.getNormalTimesArea(); - const scalar_type t = hlsl::dot(N, rect.offset - ray.origin) / hlsl::dot(N, ray.direction); - const vector3_type hitPoint = ray.origin + ray.direction * t; - const vector3_type localHit = hitPoint - rect.offset; - const vector p = vector(hlsl::dot(localHit, rectNormalBasis[0]) / rectExtents.x, hlsl::dot(localHit, rectNormalBasis[1]) / rectExtents.y); - const scalar_type pdf = psr.backwardPdf(p); - return pdf < numeric_limits::max ? pdf : numeric_limits::max; + return psr.backwardWeight(ray.direction); } template vector3_type generateAndPdfAndWeight(NBL_REF_ARG(scalar_type) pdf, NBL_REF_ARG(scalar_type) weight, NBL_REF_ARG(scalar_type) newRayMaxT, NBL_CONST_REF_ARG(vector3_type) origin, NBL_CONST_REF_ARG(Aniso) interaction, NBL_CONST_REF_ARG(vector3_type) xi) { - const vector3_type N = rect.getNormalTimesArea(); - const vector3_type origin2origin = rect.offset - origin; matrix rectNormalBasis; vector rectExtents; @@ -382,25 +380,37 @@ struct ShapeSampling sphR0.origin = rect.offset; sphR0.extents = rectExtents; sphR0.basis = rectNormalBasis; - vector3_type L = hlsl::promote(0.0); sampling::ProjectedSphericalRectangle psr = sampling::ProjectedSphericalRectangle::create(sphR0, origin, interaction.getN(), interaction.isMaterialBSDF()); - const scalar_type solidAngle = psr.sphrect.solidAngle; - if (solidAngle > numeric_limits::min) + typename sampling::ProjectedSphericalRectangle::cache_type cache; + + vector3_type L = hlsl::promote(0.0); + const bool FastVersion = true; + if (FastVersion) { - typename sampling::ProjectedSphericalRectangle::cache_type cache; - const vector3_type localDir = psr.generate(xi.xy, cache); - // not sure if generate() can produce NaN/inf when solidAngle > min - assert(!hlsl::any(hlsl::isinf(localDir) || hlsl::isnan(localDir))); - // transform local direction to world space - L = localDir.x * rectNormalBasis[0] + localDir.y * rectNormalBasis[1] + localDir.z * rectNormalBasis[2]; - pdf = psr.forwardPdf(xi.xy, cache); - weight = psr.forwardWeight(xi.xy, cache); + // actually the slowest + //L = psr.generate(xi.xy, cache); + //newRayMaxT = psr.sphrect.computeHitT(L); + + // fastest + const vector3_type localL = psr.generateNormalizedLocal(xi.xy,cache,newRayMaxT); + assert(!hlsl::any(hlsl::isinf(localL) || hlsl::isnan(localL))); + // hopefully CSE kicks in for the `UsePdfAsWeight==true` + L = hlsl::mul(hlsl::transpose(psr.sphrect.basis),localL); } else - weight = bit_cast(numeric_limits::infinity); - // TODO: `improved_spherical_rect` branch merge - newRayMaxT = hlsl::dot(N, origin2origin) / hlsl::dot(N, L); + { + L = psr.generateUnnormalized(xi.xy,cache); + assert(!hlsl::any(hlsl::isinf(L) || hlsl::isnan(L))); + const scalar_type rcpLen = hlsl::rsqrt(hlsl::dot(L,L)); + newRayMaxT = 1.f / rcpLen; + L *= rcpLen; + } + // prevent self intersections against the emitter + newRayMaxT -= 0.0001f; + + pdf = psr.forwardPdf(xi.xy,cache); + weight = psr.forwardWeight(xi.xy,cache); return L; } diff --git a/37_HLSLSamplingTests/CMakeLists.txt b/37_HLSLSamplingTests/CMakeLists.txt index 2ac238c33..78e3ab319 100644 --- a/37_HLSLSamplingTests/CMakeLists.txt +++ b/37_HLSLSamplingTests/CMakeLists.txt @@ -26,7 +26,7 @@ set(DEPENDS app_resources/shaders/projected_spherical_triangle_test.comp.hlsl app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl app_resources/shaders/spherical_rectangle_test.comp.hlsl - app_resources/shaders/alias_table_test.comp.hlsl + app_resources/shaders/packed_alias_test.comp.hlsl app_resources/shaders/cumulative_probability_test.comp.hlsl app_resources/common/linear.hlsl app_resources/common/uniform_hemisphere.hlsl @@ -42,6 +42,7 @@ set(DEPENDS app_resources/common/concentric_mapping.hlsl app_resources/common/polar_mapping.hlsl app_resources/common/discrete_sampler_bench.hlsl + app_resources/common/sampler_bench_pc.hlsl app_resources/common/alias_table.hlsl app_resources/common/cumulative_probability.hlsl ) @@ -91,7 +92,7 @@ endif() set(OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/auto-gen") -set(BENCH_ITERS 2048) +set(BENCH_ITERS 128) set(WORKGROUP_SIZE 64) target_compile_definitions(${EXECUTABLE_NAME} PRIVATE @@ -99,7 +100,7 @@ target_compile_definitions(${EXECUTABLE_NAME} PRIVATE WORKGROUP_SIZE=${WORKGROUP_SIZE} ) -set(BENCH_OPTS "\"-DBENCH_ITERS=${BENCH_ITERS}\", \"-DWORKGROUP_SIZE=${WORKGROUP_SIZE}\"") +set(BENCH_OPTS "\"-DBENCH_ITERS=${BENCH_ITERS}\"") set(JSON " [ @@ -113,8 +114,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/linear_test.comp.hlsl\", - \"KEY\": \"linear_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"linear_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/linear_test.comp.hlsl\", + \"KEY\": \"linear_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/uniform_hemisphere_test.comp.hlsl\", @@ -122,8 +128,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/uniform_hemisphere_test.comp.hlsl\", - \"KEY\": \"uniform_hemisphere_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"uniform_hemisphere_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/uniform_hemisphere_test.comp.hlsl\", + \"KEY\": \"uniform_hemisphere_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/uniform_sphere_test.comp.hlsl\", @@ -131,8 +142,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/uniform_sphere_test.comp.hlsl\", - \"KEY\": \"uniform_sphere_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"uniform_sphere_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/uniform_sphere_test.comp.hlsl\", + \"KEY\": \"uniform_sphere_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/projected_hemisphere_test.comp.hlsl\", @@ -140,8 +156,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/projected_hemisphere_test.comp.hlsl\", - \"KEY\": \"projected_hemisphere_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"projected_hemisphere_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/projected_hemisphere_test.comp.hlsl\", + \"KEY\": \"projected_hemisphere_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/projected_sphere_test.comp.hlsl\", @@ -149,8 +170,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/projected_sphere_test.comp.hlsl\", - \"KEY\": \"projected_sphere_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"projected_sphere_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/projected_sphere_test.comp.hlsl\", + \"KEY\": \"projected_sphere_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/spherical_triangle.comp.hlsl\", @@ -158,8 +184,18 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/spherical_triangle.comp.hlsl\", - \"KEY\": \"spherical_triangle_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"spherical_triangle_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_triangle.comp.hlsl\", + \"KEY\": \"spherical_triangle_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_triangle.comp.hlsl\", + \"KEY\": \"spherical_triangle_bench_create_only\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_CREATE_ONLY\"] }, { \"INPUT\": \"app_resources/shaders/concentric_mapping_test.comp.hlsl\", @@ -167,8 +203,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/concentric_mapping_test.comp.hlsl\", - \"KEY\": \"concentric_mapping_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"concentric_mapping_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/concentric_mapping_test.comp.hlsl\", + \"KEY\": \"concentric_mapping_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/polar_mapping_test.comp.hlsl\", @@ -176,8 +217,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/polar_mapping_test.comp.hlsl\", - \"KEY\": \"polar_mapping_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"polar_mapping_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/polar_mapping_test.comp.hlsl\", + \"KEY\": \"polar_mapping_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/bilinear_test.comp.hlsl\", @@ -185,8 +231,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/bilinear_test.comp.hlsl\", - \"KEY\": \"bilinear_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"bilinear_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/bilinear_test.comp.hlsl\", + \"KEY\": \"bilinear_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/box_muller_transform_test.comp.hlsl\", @@ -194,8 +245,13 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/box_muller_transform_test.comp.hlsl\", - \"KEY\": \"box_muller_transform_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"box_muller_transform_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/box_muller_transform_test.comp.hlsl\", + \"KEY\": \"box_muller_transform_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] }, { \"INPUT\": \"app_resources/shaders/projected_spherical_triangle_test.comp.hlsl\", @@ -203,8 +259,18 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/projected_spherical_triangle_test.comp.hlsl\", - \"KEY\": \"projected_spherical_triangle_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"projected_spherical_triangle_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/projected_spherical_triangle_test.comp.hlsl\", + \"KEY\": \"projected_spherical_triangle_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] + }, + { + \"INPUT\": \"app_resources/shaders/projected_spherical_triangle_test.comp.hlsl\", + \"KEY\": \"projected_spherical_triangle_bench_create_only\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_CREATE_ONLY\"] }, { \"INPUT\": \"app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl\", @@ -212,8 +278,18 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl\", - \"KEY\": \"projected_spherical_rectangle_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"projected_spherical_rectangle_bench_1_1\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"projected_spherical_rectangle_bench_1_16\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] + }, + { + \"INPUT\": \"app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"projected_spherical_rectangle_bench_create_only\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_CREATE_ONLY\"] }, { \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", @@ -221,18 +297,68 @@ set(JSON " }, { \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", - \"KEY\": \"spherical_rectangle_bench\", - \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + \"KEY\": \"spherical_rectangle_bench_1_1_shape_observer\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_1_1_sa_extents\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\", \"-DBENCH_VARIANT_SA_EXTENTS\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_1_1_r0_extents\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=1\", \"-DBENCH_VARIANT_R0_EXTENTS\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_1_16_shape_observer\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_1_16_sa_extents\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\", \"-DBENCH_VARIANT_SA_EXTENTS\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_1_16_r0_extents\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_SAMPLES_PER_CREATE=16\", \"-DBENCH_VARIANT_R0_EXTENTS\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_create_only_shape_observer\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_CREATE_ONLY\"] + }, + { + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_create_only_sa_extents\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_CREATE_ONLY\", \"-DBENCH_VARIANT_SA_EXTENTS\"] }, { - \"INPUT\": \"app_resources/shaders/alias_table_test.comp.hlsl\", - \"KEY\": \"alias_table_test\" + \"INPUT\": \"app_resources/shaders/spherical_rectangle_test.comp.hlsl\", + \"KEY\": \"spherical_rectangle_bench_create_only_r0_extents\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DBENCH_CREATE_ONLY\", \"-DBENCH_VARIANT_R0_EXTENTS\"] }, { - \"INPUT\": \"app_resources/shaders/alias_table_test.comp.hlsl\", - \"KEY\": \"alias_table_bench\", + \"INPUT\": \"app_resources/shaders/packed_alias_test.comp.hlsl\", + \"KEY\": \"packed_alias_a_test\" + }, + { + \"INPUT\": \"app_resources/shaders/packed_alias_test.comp.hlsl\", + \"KEY\": \"packed_alias_b_test\", + \"COMPILE_OPTIONS\": [\"-DNBL_PACKED_ALIAS_B\"] + }, + { + \"INPUT\": \"app_resources/shaders/packed_alias_test.comp.hlsl\", + \"KEY\": \"packed_alias_a_bench\", \"COMPILE_OPTIONS\": [${BENCH_OPTS}] }, + { + \"INPUT\": \"app_resources/shaders/packed_alias_test.comp.hlsl\", + \"KEY\": \"packed_alias_b_bench\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DNBL_PACKED_ALIAS_B\"] + }, { \"INPUT\": \"app_resources/shaders/cumulative_probability_test.comp.hlsl\", \"KEY\": \"cumulative_probability_test\" @@ -241,6 +367,16 @@ set(JSON " \"INPUT\": \"app_resources/shaders/cumulative_probability_test.comp.hlsl\", \"KEY\": \"cumulative_probability_bench\", \"COMPILE_OPTIONS\": [${BENCH_OPTS}] + }, + { + \"INPUT\": \"app_resources/shaders/cumulative_probability_test.comp.hlsl\", + \"KEY\": \"cumulative_probability_yolo_bench\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DNBL_CUMPROB_YOLO_READS\"] + }, + { + \"INPUT\": \"app_resources/shaders/cumulative_probability_test.comp.hlsl\", + \"KEY\": \"cumulative_probability_eytzinger_bench\", + \"COMPILE_OPTIONS\": [${BENCH_OPTS}, \"-DNBL_CUMPROB_EYTZINGER\"] } ] ") @@ -250,7 +386,7 @@ NBL_CREATE_NSC_COMPILE_RULES( LINK_TO ${EXECUTABLE_NAME} BINARY_DIR ${OUTPUT_DIRECTORY} MOUNT_POINT_DEFINE NBL_THIS_EXAMPLE_BUILD_MOUNT_POINT - COMMON_OPTIONS -I ${CMAKE_CURRENT_SOURCE_DIR} -T cs_6_8 + COMMON_OPTIONS -I ${CMAKE_CURRENT_SOURCE_DIR} -T cs_6_8 -DWORKGROUP_SIZE=${WORKGROUP_SIZE} OUTPUT_VAR KEYS INCLUDE nbl/this_example/builtin/build/spirv/keys.hpp NAMESPACE nbl::this_example::builtin::build diff --git a/37_HLSLSamplingTests/app_resources/common/alias_table.hlsl b/37_HLSLSamplingTests/app_resources/common/alias_table.hlsl index da7048a1f..08706408f 100644 --- a/37_HLSLSamplingTests/app_resources/common/alias_table.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/alias_table.hlsl @@ -8,12 +8,28 @@ using namespace nbl::hlsl; NBL_CONSTEXPR uint32_t AliasTestTableSize = 4; +// Log2N = ceil_log2(N) minimises quantisation drift on the stayProb unorm +// (here 30 unorm bits, essentially lossless). +NBL_CONSTEXPR uint32_t AliasTestLog2N = 2; -using AliasTestProbAccessor = ArrayAccessor; -using AliasTestAliasAccessor = ArrayAccessor; -using AliasTestPdfAccessor = ArrayAccessor; +using AliasTestPdfAccessor = ArrayAccessor; +using AliasTestPackedWordAccessor = ArrayAccessor; -using AliasTestSampler = sampling::AliasTable; +// Dedicated struct-valued accessor for PackedAliasEntryB. Field-wise copy +// sidesteps HLSL's struct functional-cast ambiguity. +struct AliasTestEntryBAccessor +{ + using value_type = sampling::PackedAliasEntryB; + + template + void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC + { + val.packedWord = data[i].packedWord; + val.ownPdf = data[i].ownPdf; + } + + value_type data[AliasTestTableSize]; +}; struct AliasTableInputValues { @@ -22,32 +38,64 @@ struct AliasTableInputValues struct AliasTableTestResults { - uint32_t generatedIndex; + uint32_t generatedIndex; float32_t forwardPdf; float32_t backwardPdf; float32_t forwardWeight; float32_t backwardWeight; + float32_t jacobianProduct; }; // Pre-computed alias table for weights {1, 2, 3, 4}: -// pdf = {0.1, 0.2, 0.3, 0.4} -// prob = {0.4, 0.8, 1.0, 0.8} -// alias = {3, 3, 2, 2} -struct AliasTableTestExecutor +// pdf = {0.1, 0.2, 0.3, 0.4} +// stayProb = {0.4, 0.8, 1.0, 0.8} +// alias = {3, 3, 2, 2} +// +// Log2N = 2 unorm encoding (30 bits for stayProb, 2 bits for alias): +// packedWord = (alias & 0x3) | (round(stayProb * ((1u<<30) - 1)) << 2) +// bin 0: (3) | (429496729 << 2) = 0x66666667 +// bin 1: (3) | (858993458 << 2) = 0xCCCCCCCB +// bin 2: (2) | (1073741823 << 2) = 0xFFFFFFFE +// bin 3: (2) | (858993458 << 2) = 0xCCCCCCCA + +struct PackedAliasATestExecutor +{ + void operator()(NBL_CONST_REF_ARG(AliasTableInputValues) input, NBL_REF_ARG(AliasTableTestResults) output) + { + AliasTestPackedWordAccessor wordAcc; + wordAcc.data[0] = 0x66666667u; + wordAcc.data[1] = 0xCCCCCCCBu; + wordAcc.data[2] = 0xFFFFFFFEu; + wordAcc.data[3] = 0xCCCCCCCAu; + + AliasTestPdfAccessor pdfAcc; + pdfAcc.data[0] = 0.1f; + pdfAcc.data[1] = 0.2f; + pdfAcc.data[2] = 0.3f; + pdfAcc.data[3] = 0.4f; + + using Sampler = sampling::PackedAliasTableA; + Sampler sampler = Sampler::create(wordAcc, pdfAcc, AliasTestTableSize); + + Sampler::cache_type cache; + output.generatedIndex = sampler.generate(input.u, cache); + output.forwardPdf = sampler.forwardPdf(input.u, cache); + output.backwardPdf = sampler.backwardPdf(output.generatedIndex); + output.forwardWeight = sampler.forwardWeight(input.u, cache); + output.backwardWeight = sampler.backwardWeight(output.generatedIndex); + output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf; + } +}; + +struct PackedAliasBTestExecutor { void operator()(NBL_CONST_REF_ARG(AliasTableInputValues) input, NBL_REF_ARG(AliasTableTestResults) output) { - AliasTestProbAccessor probAcc; - probAcc.data[0] = 0.4f; - probAcc.data[1] = 0.8f; - probAcc.data[2] = 1.0f; - probAcc.data[3] = 0.8f; - - AliasTestAliasAccessor aliasAcc; - aliasAcc.data[0] = 3u; - aliasAcc.data[1] = 3u; - aliasAcc.data[2] = 2u; - aliasAcc.data[3] = 2u; + AliasTestEntryBAccessor entryAcc; + entryAcc.data[0].packedWord = 0x66666667u; entryAcc.data[0].ownPdf = 0.1f; + entryAcc.data[1].packedWord = 0xCCCCCCCBu; entryAcc.data[1].ownPdf = 0.2f; + entryAcc.data[2].packedWord = 0xFFFFFFFEu; entryAcc.data[2].ownPdf = 0.3f; + entryAcc.data[3].packedWord = 0xCCCCCCCAu; entryAcc.data[3].ownPdf = 0.4f; AliasTestPdfAccessor pdfAcc; pdfAcc.data[0] = 0.1f; @@ -55,14 +103,16 @@ struct AliasTableTestExecutor pdfAcc.data[2] = 0.3f; pdfAcc.data[3] = 0.4f; - AliasTestSampler sampler = AliasTestSampler::create(probAcc, aliasAcc, pdfAcc, AliasTestTableSize); + using Sampler = sampling::PackedAliasTableB; + Sampler sampler = Sampler::create(entryAcc, pdfAcc, AliasTestTableSize); - AliasTestSampler::cache_type cache; - output.generatedIndex = sampler.generate(input.u, cache); - output.forwardPdf = sampler.forwardPdf(input.u, cache); - output.backwardPdf = sampler.backwardPdf(output.generatedIndex); - output.forwardWeight = sampler.forwardWeight(input.u, cache); - output.backwardWeight = sampler.backwardWeight(output.generatedIndex); + Sampler::cache_type cache; + output.generatedIndex = sampler.generate(input.u, cache); + output.forwardPdf = sampler.forwardPdf(input.u, cache); + output.backwardPdf = sampler.backwardPdf(output.generatedIndex); + output.forwardWeight = sampler.forwardWeight(input.u, cache); + output.backwardWeight = sampler.backwardWeight(output.generatedIndex); + output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf; } }; diff --git a/37_HLSLSamplingTests/app_resources/common/array_accessor.hlsl b/37_HLSLSamplingTests/app_resources/common/array_accessor.hlsl index 1f0a68195..5e679c98a 100644 --- a/37_HLSLSamplingTests/app_resources/common/array_accessor.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/array_accessor.hlsl @@ -12,7 +12,6 @@ struct ArrayAccessor using value_type = T; template void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC { val = V(data[i]); } - T operator[](uint32_t i) NBL_CONST_MEMBER_FUNC { return data[i]; } T data[N]; }; diff --git a/37_HLSLSamplingTests/app_resources/common/bilinear.hlsl b/37_HLSLSamplingTests/app_resources/common/bilinear.hlsl index 64a13d3e1..752e547ce 100644 --- a/37_HLSLSamplingTests/app_resources/common/bilinear.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/bilinear.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -19,6 +20,7 @@ struct BilinearTestResults float32_t forwardPdf; float32_t forwardWeight; float32_t backwardWeight; + float32_t jacobianProduct; }; struct BilinearTestExecutor @@ -37,6 +39,10 @@ struct BilinearTestExecutor output.backwardPdf = sampler.backwardPdf(output.generated); output.backwardWeight = sampler.backwardWeight(output.generated); } + // marginFactor = 3: same reasoning as Linear; Bilinear is two Linear stages, so the skewed- + // coefficient inverse-CDF d^2/du^2 divergence near [0,1]^2 boundary applies on both axes. + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 3.0f); + } }; diff --git a/37_HLSLSamplingTests/app_resources/common/box_muller_transform.hlsl b/37_HLSLSamplingTests/app_resources/common/box_muller_transform.hlsl index e8247e259..2b86e8560 100644 --- a/37_HLSLSamplingTests/app_resources/common/box_muller_transform.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/box_muller_transform.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -21,6 +22,7 @@ struct BoxMullerTransformTestResults float32_t forwardWeight; float32_t backwardWeight; float32_t2 separateBackwardPdf; + float32_t jacobianProduct; }; struct BoxMullerTransformTestExecutor @@ -40,6 +42,7 @@ struct BoxMullerTransformTestExecutor output.backwardPdf = sampler.backwardPdf(output.generated); output.backwardWeight = sampler.backwardWeight(output.generated); output.separateBackwardPdf = sampler.separateBackwardPdf(output.generated); + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 10.0f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/concentric_mapping.hlsl b/37_HLSLSamplingTests/app_resources/common/concentric_mapping.hlsl index 67d8e5869..e0c6a570c 100644 --- a/37_HLSLSamplingTests/app_resources/common/concentric_mapping.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/concentric_mapping.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -20,6 +21,7 @@ struct ConcentricMappingTestResults float32_t forwardWeight; float32_t backwardWeight; float32_t jacobianProduct; + float32_t inverseJacobianPdf; float32_t2 roundtripError; }; @@ -39,7 +41,15 @@ struct ConcentricMappingTestExecutor output.backwardWeight = sampling::ConcentricMapping::backwardWeight(input.u); } output.roundtripError = nbl::hlsl::abs(input.u - output.inverted); - output.jacobianProduct = float32_t(1.0 / output.backwardPdf) * output.forwardPdf; + { + sampling::ConcentricMapping sampler; + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 1.0f); + // Disk-center singularity: concentric atan2 blows up as r->0. + const float32_t diskRadius = nbl::hlsl::length(output.mapped); + output.inverseJacobianPdf = diskRadius < 0.1f + ? JACOBIAN_SKIP_CODOMAIN_SINGULARITY + : computeInverseJacobianPdf(sampler, output.mapped, output.backwardPdf, 0.0f, 1e30f); + } } }; diff --git a/37_HLSLSamplingTests/app_resources/common/cumulative_probability.hlsl b/37_HLSLSamplingTests/app_resources/common/cumulative_probability.hlsl index f58a22741..e66cb44fe 100644 --- a/37_HLSLSamplingTests/app_resources/common/cumulative_probability.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/cumulative_probability.hlsl @@ -24,6 +24,7 @@ struct CumProbTestResults float32_t backwardPdf; float32_t forwardWeight; float32_t backwardWeight; + float32_t jacobianProduct; }; // Pre-computed CDF table for weights {1, 2, 3, 4}: @@ -46,6 +47,7 @@ struct CumProbTestExecutor output.backwardPdf = sampler.backwardPdf(output.generatedIndex); output.forwardWeight = sampler.forwardWeight(input.u, cache); output.backwardWeight = sampler.backwardWeight(output.generatedIndex); + output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf; } }; diff --git a/37_HLSLSamplingTests/app_resources/common/discrete_sampler_bench.hlsl b/37_HLSLSamplingTests/app_resources/common/discrete_sampler_bench.hlsl index 9f1fec422..198b72faf 100644 --- a/37_HLSLSamplingTests/app_resources/common/discrete_sampler_bench.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/discrete_sampler_bench.hlsl @@ -5,23 +5,22 @@ using namespace nbl::hlsl; -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 -#endif NBL_CONSTEXPR uint32_t WorkgroupSize = WORKGROUP_SIZE; -struct AliasTablePushConstants +struct CumProbPushConstants { - uint64_t probAddress; // float probability[N] - uint64_t aliasAddress; // uint32_t alias[N] - uint64_t pdfAddress; // float pdf[N] + uint64_t cumProbAddress; // float cumProb[N-1] uint64_t outputAddress; // uint32_t acc[threadCount] uint32_t tableSize; // N }; -struct CumProbPushConstants +// Variants A and B both take the entry array plus a separate pdf[] array +// (A: 4 B words, B: 8 B {packedWord, ownPdf}; pdf[] has the same contents in +// both but is tapped independently by the sampler). +struct PackedAliasABPushConstants { - uint64_t cumProbAddress; // float cumProb[N-1] + uint64_t entriesAddress; // A: uint32_t words[N] (4 B); B: PackedAliasEntryB[N] (8 B) + uint64_t pdfAddress; // float pdf[N] uint64_t outputAddress; // uint32_t acc[threadCount] uint32_t tableSize; // N }; diff --git a/37_HLSLSamplingTests/app_resources/common/jacobian_test.hlsl b/37_HLSLSamplingTests/app_resources/common/jacobian_test.hlsl new file mode 100644 index 000000000..f949f5b86 --- /dev/null +++ b/37_HLSLSamplingTests/app_resources/common/jacobian_test.hlsl @@ -0,0 +1,264 @@ +#ifndef _NBL_EXAMPLES_TESTS_37_SAMPLING_COMMON_JACOBIAN_TEST_INCLUDED_ +#define _NBL_EXAMPLES_TESTS_37_SAMPLING_COMMON_JACOBIAN_TEST_INCLUDED_ + +#include +#include + +using namespace nbl::hlsl; + +// Negative sentinels signal "skipped" to the host verifier; the value encodes the reason. +static const float32_t JACOBIAN_SKIP_U_DOMAIN = -1.0f; +static const float32_t JACOBIAN_SKIP_CREASE = -2.0f; +static const float32_t JACOBIAN_SKIP_HEMI_BOUNDARY = -3.0f; +static const float32_t JACOBIAN_SKIP_BWD_PDF_RANGE = -4.0f; +static const float32_t JACOBIAN_SKIP_CODOMAIN_SINGULARITY = -5.0f; + + +template +struct ForwardJacobianMeasure; + +// Signed step that stays inside [0,1]: flip direction when u is in the upper half so u +/- eps +// never overshoots the domain. Magnitude is what matters (the stencil results take abs/length). +template +T signedEps(T u, T eps) +{ + return u > T(0.5) ? -eps : eps; +} + +template +struct ForwardJacobianMeasure +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + using codomain_type = typename Sampler::codomain_type; + using cache_type = typename Sampler::cache_type; + + static scalar_type compute(Sampler _sampler, domain_type u, scalar_type eps, codomain_type L) + { + cache_type c; + const codomain_type L_x = _sampler.generate(u + signedEps(u, eps), c); + return nbl::hlsl::abs(L_x - L) / eps; + } +}; + +template +struct ForwardJacobianMeasure +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + using codomain_type = typename Sampler::codomain_type; + using cache_type = typename Sampler::cache_type; + + static scalar_type compute(Sampler _sampler, domain_type u, scalar_type eps, codomain_type L) + { + domain_type u_x = u; + u_x[0] += signedEps(u[0], eps); + domain_type u_y = u; + u_y[1] += signedEps(u[1], eps); + cache_type c; + const codomain_type L_x = _sampler.generate(u_x, c); + const codomain_type L_y = _sampler.generate(u_y, c); + using matrix2_type = matrix; + const scalar_type det = nbl::hlsl::determinant(matrix2_type(L_x - L, L_y - L)); + return nbl::hlsl::abs(det) / (eps * eps); + } +}; + +template +struct ForwardJacobianMeasure +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + using codomain_type = typename Sampler::codomain_type; + using cache_type = typename Sampler::cache_type; + + static scalar_type compute(Sampler _sampler, domain_type u, scalar_type eps, codomain_type L) + { + domain_type u_x = u; + u_x[0] += signedEps(u[0], eps); + domain_type u_y = u; + u_y[1] += signedEps(u[1], eps); + cache_type c; + const codomain_type L_x = _sampler.generate(u_x, c); + const codomain_type L_y = _sampler.generate(u_y, c); + return nbl::hlsl::length(nbl::hlsl::cross(L_x - L, L_y - L)) / (eps * eps); + } +}; + +// 3D domain: stencil perturbs u[0] and u[1] only, so the (2,3) body applies unchanged. +template +struct ForwardJacobianMeasure : ForwardJacobianMeasure +{ +}; + + +template +struct DomainMarginCheck; + +template +struct DomainMarginCheck +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + static bool outsideMargin(domain_type u, scalar_type margin) + { + return u < margin || u > scalar_type(1) - margin; + } +}; + +template +struct DomainMarginCheck +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + static bool outsideMargin(domain_type u, scalar_type margin) + { + return u[0] < margin || u[0] > scalar_type(1) - margin || u[1] < margin || u[1] > scalar_type(1) - margin; + } +}; + +// 3D domain: forward stencil only perturbs u[0] and u[1], so u[2] is irrelevant and (2) applies. +template +struct DomainMarginCheck : DomainMarginCheck +{ +}; + +enum JacobianMode : uint32_t +{ + JACOBIAN_PLAIN = 0, + JACOBIAN_CONCENTRIC = 1, // + concentric crease skip + JACOBIAN_CONCENTRIC_UXFOLD = 2 // + crease + u.x=0.5 hemi-boundary skip +}; + +// marginFactor scales the u-domain skip to marginFactor * eps. Use > 1 only for samplers whose +// stencil bias extends past a single eps-step (e.g. Arvo spherical triangle: sinZ ~ sqrt(u.y) +// gives O(h/u.y) forward-diff bias, so u.y in [0, k*eps] must be skipped). +template +float32_t computeJacobianProduct(Sampler _sampler, typename Sampler::domain_type u, float32_t eps, float32_t marginFactor) +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + using codomain_type = typename Sampler::codomain_type; + using cache_type = typename Sampler::cache_type; + + NBL_IF_CONSTEXPR(Mode != JACOBIAN_PLAIN) + { + // Cast via float32_t2 so this block typechecks for scalar / vec2 / vec3 domains alike + // (HLSL splats scalars, identity on vec2, .xy on vec3). 1D samplers never reach here. + const float32_t2 uxy = (float32_t2)u; + const float32_t ux = uxy.x; + const float32_t uy = uxy.y; + + NBL_IF_CONSTEXPR(Mode == JACOBIAN_CONCENTRIC_UXFOLD) + { + if (nbl::hlsl::abs(ux - float32_t(0.5)) <= float32_t(2e-3)) + return JACOBIAN_SKIP_HEMI_BOUNDARY; + } + + const bool uxFold = (Mode == JACOBIAN_CONCENTRIC_UXFOLD); + // Empirical: the concentric C0 crease's stencil bias spreads wider than the 2*eps geometric + // straddle band. Non-uxFold 6e-3 covers the disk-center residual for Projected samplers; + // uxFold 1e-2 accounts for the doubled local_ux rate when u.x is folded. + const float32_t creaseBand = uxFold ? float32_t(1e-2) : float32_t(6e-3); + const float32_t local_ux = uxFold ? nbl::hlsl::abs(float32_t(2) * ux - float32_t(1)) : ux; + const float32_t a = float32_t(2) * local_ux - float32_t(1); + const float32_t b = float32_t(2) * uy - float32_t(1); + if (nbl::hlsl::abs(nbl::hlsl::abs(a) - nbl::hlsl::abs(b)) <= creaseBand) + return JACOBIAN_SKIP_CREASE; + } + + using margin_check_type = DomainMarginCheck::Dimension>; + if (margin_check_type::outsideMargin(u, scalar_type(eps * marginFactor))) + return JACOBIAN_SKIP_U_DOMAIN; + + // Generate on a copy: some samplers mutate u through NBL_REF_ARG (e.g. ProjectedSphere + // consumes u.z for hemisphere selection), and the perturbations below need the original u. + cache_type cache; + domain_type uGen = u; + const codomain_type L = _sampler.generate(uGen, cache); + const scalar_type pdf = _sampler.forwardPdf(uGen, cache); + + using measure_type = ForwardJacobianMeasure::Dimension, vector_traits::Dimension>; + const scalar_type measure = measure_type::compute(_sampler, u, scalar_type(eps), L); + + return pdf * measure; +} + + +template +struct InverseJacobianMeasure; + +template +struct InverseJacobianMeasure +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + using codomain_type = typename Sampler::codomain_type; + + static scalar_type compute(Sampler _sampler, codomain_type x, scalar_type eps) + { + const scalar_type twoEps = scalar_type(2) * eps; + codomain_type x0_lo = x; + x0_lo[0] -= eps; + codomain_type x0_hi = x; + x0_hi[0] += eps; + codomain_type x1_lo = x; + x1_lo[1] -= eps; + codomain_type x1_hi = x; + x1_hi[1] += eps; + domain_type u0_lo = _sampler.generateInverse(x0_lo); + domain_type u0_hi = _sampler.generateInverse(x0_hi); + domain_type u1_lo = _sampler.generateInverse(x1_lo); + domain_type u1_hi = _sampler.generateInverse(x1_hi); + const domain_type dudx0 = (u0_hi - u0_lo) / twoEps; + const domain_type dudx1 = (u1_hi - u1_lo) / twoEps; + using matrix2_type = matrix; + const scalar_type det = nbl::hlsl::determinant(matrix2_type(dudx0, dudx1)); + return nbl::hlsl::abs(det); + } +}; + +template +struct InverseJacobianMeasure +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + using codomain_type = typename Sampler::codomain_type; + + static scalar_type compute(Sampler _sampler, codomain_type x, scalar_type eps) + { + const scalar_type twoEps = scalar_type(2) * eps; + codomain_type t1, t2; + const codomain_type up = nbl::hlsl::abs(x[2]) < scalar_type(0.999) + ? codomain_type(scalar_type(0), scalar_type(0), scalar_type(1)) + : codomain_type(scalar_type(1), scalar_type(0), scalar_type(0)); + t1 = nbl::hlsl::normalize(nbl::hlsl::cross(up, x)); + t2 = nbl::hlsl::cross(x, t1); + domain_type u_t1_lo = _sampler.generateInverse(nbl::hlsl::normalize(x - t1 * eps)); + domain_type u_t1_hi = _sampler.generateInverse(nbl::hlsl::normalize(x + t1 * eps)); + domain_type u_t2_lo = _sampler.generateInverse(nbl::hlsl::normalize(x - t2 * eps)); + domain_type u_t2_hi = _sampler.generateInverse(nbl::hlsl::normalize(x + t2 * eps)); + const domain_type dudt1 = (u_t1_hi - u_t1_lo) / twoEps; + const domain_type dudt2 = (u_t2_hi - u_t2_lo) / twoEps; + using matrix2_type = matrix; + const scalar_type det = nbl::hlsl::determinant(matrix2_type(dudt1, dudt2)); + return nbl::hlsl::abs(det); + } +}; + +template +float32_t computeInverseJacobianPdf(Sampler _sampler, typename Sampler::codomain_type sample, float32_t backwardPdf, float32_t pdfMin, float32_t pdfMax) +{ + using scalar_type = typename Sampler::scalar_type; + using domain_type = typename Sampler::domain_type; + using codomain_type = typename Sampler::codomain_type; + + if (backwardPdf < scalar_type(pdfMin) || backwardPdf > scalar_type(pdfMax)) + return JACOBIAN_SKIP_BWD_PDF_RANGE; + + using measure_type = InverseJacobianMeasure::Dimension, vector_traits::Dimension>; + const scalar_type eps = scalar_type(1e-3); + return measure_type::compute(_sampler, sample, eps); +} + +#endif diff --git a/37_HLSLSamplingTests/app_resources/common/linear.hlsl b/37_HLSLSamplingTests/app_resources/common/linear.hlsl index b27d88e5b..af269ad2f 100644 --- a/37_HLSLSamplingTests/app_resources/common/linear.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/linear.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -19,6 +20,7 @@ struct LinearTestResults float32_t backwardPdf; float32_t forwardWeight; float32_t backwardWeight; + float32_t jacobianProduct; }; struct LinearTestExecutor @@ -37,6 +39,7 @@ struct LinearTestExecutor output.backwardPdf = _sampler.backwardPdf(output.generated); output.backwardWeight = _sampler.backwardWeight(output.generated); } + output.jacobianProduct = computeJacobianProduct(_sampler, input.u, 1e-3f, 3.0f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/polar_mapping.hlsl b/37_HLSLSamplingTests/app_resources/common/polar_mapping.hlsl index 82e020fdc..e4b8ffabb 100644 --- a/37_HLSLSamplingTests/app_resources/common/polar_mapping.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/polar_mapping.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -20,6 +21,7 @@ struct PolarMappingTestResults float32_t forwardWeight; float32_t backwardWeight; float32_t jacobianProduct; + float32_t inverseJacobianPdf; float32_t2 roundtripError; }; @@ -39,7 +41,23 @@ struct PolarMappingTestExecutor output.backwardWeight = sampling::PolarMapping::backwardWeight(input.u); } output.roundtripError = nbl::hlsl::abs(input.u - output.inverted); - output.jacobianProduct = float32_t(1.0 / output.backwardPdf) * output.forwardPdf; + + { + sampling::PolarMapping sampler; + // marginFactor = 3: r = sqrt(u.x) gives O(h/u.x) forward-diff bias near u.x=0, so skip + // u.x within 3*eps of the domain boundary (same reasoning as Linear's skewed-density case). + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 3.0f); + // Two inverse singularities: + // - disk center: atan2 diverges as r -> 0 + // - atan2 branch cut at y=0, x>0: the stencil's +/-eps in y straddles the 2*pi wrap, + // producing du.y/eps ~ 1/eps spikes (seen as test values ~305-862 with eps=1e-3). + const float32_t polarRadius = nbl::hlsl::length(output.mapped); + const bool onCutBand = nbl::hlsl::abs(output.mapped.y) < 5e-3f && output.mapped.x > 0.0f; + output.inverseJacobianPdf = (polarRadius < 0.1f || onCutBand) + ? JACOBIAN_SKIP_CODOMAIN_SINGULARITY + : computeInverseJacobianPdf(sampler, output.mapped, output.backwardPdf, 0.0f, 1e30f); + } + } }; diff --git a/37_HLSLSamplingTests/app_resources/common/projected_hemisphere.hlsl b/37_HLSLSamplingTests/app_resources/common/projected_hemisphere.hlsl index 9697cf0df..c48697b03 100644 --- a/37_HLSLSamplingTests/app_resources/common/projected_hemisphere.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/projected_hemisphere.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -22,6 +23,7 @@ struct ProjectedHemisphereTestResults float32_t backwardWeight; float32_t2 roundtripError; float32_t jacobianProduct; + float32_t inverseJacobianPdf; }; struct ProjectedHemisphereTestExecutor @@ -43,7 +45,11 @@ struct ProjectedHemisphereTestExecutor output.backwardWeight = sampler.backwardWeight(output.generated); } output.roundtripError = nbl::hlsl::abs(input.u - output.inverted); - output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf; + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 5.0f); + const float32_t phDiskR = nbl::hlsl::length((float32_t2)output.generated); + output.inverseJacobianPdf = phDiskR < 0.1f + ? JACOBIAN_SKIP_CODOMAIN_SINGULARITY + : computeInverseJacobianPdf(sampler, output.generated, output.backwardPdf, 1e-3f, 1e30f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/projected_sphere.hlsl b/37_HLSLSamplingTests/app_resources/common/projected_sphere.hlsl index e9886b61d..a78a937f6 100644 --- a/37_HLSLSamplingTests/app_resources/common/projected_sphere.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/projected_sphere.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -20,6 +21,7 @@ struct ProjectedSphereTestResults float32_t backwardPdf; float32_t forwardWeight; float32_t backwardWeight; + float32_t jacobianProduct; }; struct ProjectedSphereTestExecutor @@ -38,6 +40,7 @@ struct ProjectedSphereTestExecutor } output.backwardPdf = sampler.backwardPdf(output.generated); output.backwardWeight = sampler.backwardWeight(output.generated); + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 5.0f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/projected_spherical_rectangle.hlsl b/37_HLSLSamplingTests/app_resources/common/projected_spherical_rectangle.hlsl index 8370952ca..4aed7d9c3 100644 --- a/37_HLSLSamplingTests/app_resources/common/projected_spherical_rectangle.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/projected_spherical_rectangle.hlsl @@ -4,6 +4,7 @@ #include #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -24,12 +25,10 @@ struct ProjectedSphericalRectangleTestResults float32_t2 surfaceOffset; float32_t3 referenceDirection; float32_t forwardPdf; - float32_t backwardPdf; float32_t forwardWeight; float32_t backwardWeight; - float32_t backwardPdfAtGenerated; - float32_t backwardWeightAtGenerated; float32_t2 extents; + float32_t jacobianProduct; }; struct ProjectedSphericalRectangleTestExecutor @@ -46,30 +45,29 @@ struct ProjectedSphericalRectangleTestExecutor output.extents = rect.extents; sampling::ProjectedSphericalRectangle::cache_type cache; + output.generated = sampler.generate(input.u, cache); + output.forwardPdf = sampler.forwardPdf(input.u, cache); + output.forwardWeight = sampler.forwardWeight(input.u, cache); + // backwardWeight now takes a 3D direction; evaluate at generated L. + output.backwardWeight = sampler.backwardWeight(output.generated); + + float32_t2 absXY; { - output.generated = sampler.generate(input.u, cache); - output.forwardPdf = sampler.forwardPdf(input.u, cache); - output.forwardWeight = sampler.forwardWeight(input.u, cache); - } - { - sampling::ProjectedSphericalRectangle::cache_type offsetCache; - output.surfaceOffset = sampler.generateSurfaceOffset(input.u, offsetCache); + typename sampling::Bilinear::cache_type bc; + const float32_t2 warped = sampler.bilinearPatch.generate(input.u, bc); + typename sampling::SphericalRectangle::cache_type sphrectCache; + absXY = sampler.sphrect.generateLocalBasisXY(warped, sphrectCache); + output.surfaceOffset = absXY - float32_t2(sampler.sphrect.r0.x, sampler.sphrect.r0.y); } - // reference direction: reconstruct local 3D point from surfaceOffset and normalize { - const float32_t3 localPoint = sampler.sphrect.r0 + float32_t3(output.surfaceOffset.x, output.surfaceOffset.y, float32_t(0)); - output.referenceDirection = nbl::hlsl::normalize(localPoint); + const float32_t3 localPoint = float32_t3(absXY.x, absXY.y, sampler.sphrect.r0.z); + const float32_t3 localDir = nbl::hlsl::normalize(localPoint); + output.referenceDirection = sampler.sphrect.basis[0] * localDir[0] + + sampler.sphrect.basis[1] * localDir[1] + + sampler.sphrect.basis[2] * localDir[2]; } - // Test backwardPdf/Weight at the rect center: a deterministic interior point - // that avoids amplifying generate's FP errors through backward evaluation. - const float32_t2 center = float32_t2(0.5, 0.5); - output.backwardPdf = sampler.backwardPdf(center); - output.backwardWeight = sampler.backwardWeight(center); - // Use cache.warped (the [0,1]^2 input to the spherical rect warp) for consistency - // checks, NOT generated/extents (the nonlinear warp output). The bilinear in - // forwardPdf evaluates at cache.warped, so backwardPdf must too. - output.backwardPdfAtGenerated = sampler.backwardPdf(cache.warped); - output.backwardWeightAtGenerated = sampler.backwardWeight(cache.warped); + + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 10.0f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/projected_spherical_triangle.hlsl b/37_HLSLSamplingTests/app_resources/common/projected_spherical_triangle.hlsl index 5c81e53e0..0c424590b 100644 --- a/37_HLSLSamplingTests/app_resources/common/projected_spherical_triangle.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/projected_spherical_triangle.hlsl @@ -4,6 +4,7 @@ #include #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -21,11 +22,10 @@ struct ProjectedSphericalTriangleTestResults { float32_t3 generated; float32_t forwardPdf; - float32_t backwardPdf; - float32_t backwardPdfAtGenerated; float32_t forwardWeight; float32_t backwardWeight; float32_t backwardWeightAtGenerated; + float32_t jacobianProduct; }; struct ProjectedSphericalTriangleTestExecutor @@ -43,15 +43,20 @@ struct ProjectedSphericalTriangleTestExecutor output.forwardPdf = sampler.forwardPdf(input.u, cache); output.forwardWeight = sampler.forwardWeight(input.u, cache); } - // Test backwardPdf/Weight at the triangle centroid: a deterministic interior point computed - // from only basic arithmetic + sqrt (IEEE 754 exact), so CPU and GPU agree bit-exactly. - // Using output.generated would amplify generate's transcendental FP errors through - // generateInverse's acos, producing CPU/GPU divergence. const float32_t3 center = nbl::hlsl::normalize(input.vertex0 + input.vertex1 + input.vertex2); - output.backwardPdf = sampler.backwardPdf(center); output.backwardWeight = sampler.backwardWeight(center); - output.backwardPdfAtGenerated = sampler.backwardPdf(output.generated); output.backwardWeightAtGenerated = sampler.backwardWeight(output.generated); + // Check the bilinear-warped (inner) u directly: for skinny triangles with a strongly biased + // receiver normal, outer u well inside [0,1] can still warp to inner u <~ 0.02 where Arvo's + // sqrt(sinZ) noise dominates. Pre-skip on the inner u instead of padding an outer marginFactor. + sampling::Bilinear::cache_type bc; + const float32_t2 innerU = sampler.bilinearPatch.generate(input.u, bc); + const float32_t innerMargin = 0.02f; + const bool innerNearEdge = innerU.x < innerMargin || innerU.x > (1.0f - innerMargin) + || innerU.y < innerMargin || innerU.y > (1.0f - innerMargin); + output.jacobianProduct = innerNearEdge + ? JACOBIAN_SKIP_U_DOMAIN + : computeJacobianProduct(sampler, input.u, 1e-3f, 1.0f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/sampler_bench_pc.hlsl b/37_HLSLSamplingTests/app_resources/common/sampler_bench_pc.hlsl new file mode 100644 index 000000000..ab357e504 --- /dev/null +++ b/37_HLSLSamplingTests/app_resources/common/sampler_bench_pc.hlsl @@ -0,0 +1,15 @@ +#ifndef _NBL_EXAMPLES_TESTS_37_SAMPLING_COMMON_SAMPLER_BENCH_PC_INCLUDED_ +#define _NBL_EXAMPLES_TESTS_37_SAMPLING_COMMON_SAMPLER_BENCH_PC_INCLUDED_ + +#include + +// Implicit-output benchmark push constants. Every sampler bench shader writes +// one uint32_t accumulator per thread to outputAddress[invID]; nothing reads it +// back -- the goal is to keep the optimiser from eliding the sampling work. +// Mirrors the BDA convention from discrete_sampler_bench.hlsl. +struct SamplerBenchPushConstants +{ + uint64_t outputAddress; +}; + +#endif diff --git a/37_HLSLSamplingTests/app_resources/common/spherical_rectangle.hlsl b/37_HLSLSamplingTests/app_resources/common/spherical_rectangle.hlsl index 9ae4df256..68159405a 100644 --- a/37_HLSLSamplingTests/app_resources/common/spherical_rectangle.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/spherical_rectangle.hlsl @@ -4,6 +4,7 @@ #include #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -21,11 +22,17 @@ struct SphericalRectangleTestResults float32_t3 generated; float32_t2 surfaceOffset; float32_t3 referenceDirection; + float32_t3 normalizedLocal; + float32_t hitDist; + float32_t3 unnormalized; + float32_t computedHitT; + float32_t3 normalizedLocalToWorld; float32_t forwardPdf; float32_t backwardPdf; float32_t forwardWeight; float32_t backwardWeight; float32_t2 extents; + float32_t jacobianProduct; }; struct SphericalRectangleTestExecutor @@ -47,17 +54,36 @@ struct SphericalRectangleTestExecutor output.forwardPdf = sampler.forwardPdf(input.u, cache); output.forwardWeight = sampler.forwardWeight(input.u, cache); } + float32_t2 absXY; { sampling::SphericalRectangle::cache_type cache; - output.surfaceOffset = sampler.generateSurfaceOffset(input.u, cache); + absXY = sampler.generateLocalBasisXY(input.u, cache); + output.surfaceOffset = absXY - float32_t2(sampler.r0.x, sampler.r0.y); } - // reference direction: reconstruct local 3D point from surfaceOffset and normalize { - const float32_t3 localPoint = sampler.r0 + float32_t3(output.surfaceOffset.x, output.surfaceOffset.y, float32_t(0)); - output.referenceDirection = nbl::hlsl::normalize(localPoint); + const float32_t3 localDir = nbl::hlsl::normalize(float32_t3(absXY.x, absXY.y, sampler.r0.z)); + output.referenceDirection = sampler.basis[0] * localDir[0] + + sampler.basis[1] * localDir[1] + + sampler.basis[2] * localDir[2]; } + { + sampling::SphericalRectangle::cache_type cache; + output.normalizedLocal = sampler.generateNormalizedLocal(input.u, cache, output.hitDist); + output.normalizedLocalToWorld = sampler.basis[0] * output.normalizedLocal[0] + + sampler.basis[1] * output.normalizedLocal[1] + + sampler.basis[2] * output.normalizedLocal[2]; + } + { + sampling::SphericalRectangle::cache_type cache; + output.unnormalized = sampler.generateUnnormalized(input.u, cache); + } + output.computedHitT = sampler.computeHitT(output.generated); + output.backwardPdf = sampler.backwardPdf(output.generated); output.backwardWeight = sampler.backwardWeight(output.generated); + // marginFactor = 3: __generate's sin_au denominator goes through catastrophic cancellation + // for u.x within ~2*eps of 0 or 1 (au near n*pi), leaving ~0.5% residual at factor 3. + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 3.0f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/spherical_triangle.hlsl b/37_HLSLSamplingTests/app_resources/common/spherical_triangle.hlsl index 291661629..d3cd09326 100644 --- a/37_HLSLSamplingTests/app_resources/common/spherical_triangle.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/spherical_triangle.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -24,6 +25,7 @@ struct SphericalTriangleTestResults float32_t backwardWeight; float32_t2 roundtripError; float32_t jacobianProduct; + float32_t inverseJacobianPdf; // Minimum signed distance to a triangle edge (sin of angular distance to nearest great circle). // Positive = inside, negative = outside. Allows tolerance at boundaries. float32_t generatedInside; @@ -39,7 +41,7 @@ struct SphericalTriangleTestExecutor const float32_t3 verts[3] = { input.vertex0, input.vertex1, input.vertex2 }; shapes::SphericalTriangle shape = shapes::SphericalTriangle::createFromUnitSphereVertices(verts); - sampling::SphericalTriangle sampler = sampling::SphericalTriangle::create(shape); + sampling::SphericalTriangle sampler = sampling::SphericalTriangle::create(shape); // Forward: u -> v { @@ -58,9 +60,7 @@ struct SphericalTriangleTestExecutor } // Roundtrip error: ||u - u'|| output.roundtripError = nbl::hlsl::abs(input.u - output.inverted); - - // Jacobian product: (1/forwardPdf) * backwardPdf should equal 1 for bijective samplers - output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf; + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 20.0f); // Domain preservation: // A point is inside the spherical triangle iff it is on the "inside" half-plane @@ -79,6 +79,13 @@ struct SphericalTriangleTestExecutor float32_t2 u = output.inverted; output.invertedInDomain = nbl::hlsl::min(nbl::hlsl::min(u.x, float32_t(1.0) - u.x), nbl::hlsl::min(u.y, float32_t(1.0) - u.y)); + + const float32_t uMargin = 1e-2f; + const bool nearUBoundary = output.inverted.x < uMargin || output.inverted.x > (1.0f - uMargin) + || output.inverted.y < uMargin || output.inverted.y > (1.0f - uMargin); + output.inverseJacobianPdf = nearUBoundary + ? JACOBIAN_SKIP_CODOMAIN_SINGULARITY + : computeInverseJacobianPdf(sampler, output.generated, output.backwardPdf, 0.1f, 10.0f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/uniform_hemisphere.hlsl b/37_HLSLSamplingTests/app_resources/common/uniform_hemisphere.hlsl index 76a724774..8541bef19 100644 --- a/37_HLSLSamplingTests/app_resources/common/uniform_hemisphere.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/uniform_hemisphere.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -14,7 +15,6 @@ struct UniformHemisphereInputValues struct UniformHemisphereTestResults { float32_t3 generated; - float32_t pdf; float32_t2 inverted; float32_t forwardPdf; float32_t backwardPdf; @@ -22,6 +22,7 @@ struct UniformHemisphereTestResults float32_t backwardWeight; float32_t2 roundtripError; float32_t jacobianProduct; + float32_t inverseJacobianPdf; }; struct UniformHemisphereTestExecutor @@ -42,7 +43,11 @@ struct UniformHemisphereTestExecutor output.backwardWeight = sampler.backwardWeight(output.generated); } output.roundtripError = nbl::hlsl::abs(input.u - output.inverted); - output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf; + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 1.0f); + const float32_t uhDiskR = nbl::hlsl::length((float32_t2)output.generated); + output.inverseJacobianPdf = uhDiskR < 0.1f + ? JACOBIAN_SKIP_CODOMAIN_SINGULARITY + : computeInverseJacobianPdf(sampler, output.generated, output.backwardPdf, 0.0f, 1e30f); } }; diff --git a/37_HLSLSamplingTests/app_resources/common/uniform_sphere.hlsl b/37_HLSLSamplingTests/app_resources/common/uniform_sphere.hlsl index 3780b82ef..fb4086e44 100644 --- a/37_HLSLSamplingTests/app_resources/common/uniform_sphere.hlsl +++ b/37_HLSLSamplingTests/app_resources/common/uniform_sphere.hlsl @@ -3,6 +3,7 @@ #include #include +#include "jacobian_test.hlsl" using namespace nbl::hlsl; @@ -14,7 +15,6 @@ struct UniformSphereInputValues struct UniformSphereTestResults { float32_t3 generated; - float32_t pdf; float32_t2 inverted; float32_t forwardPdf; float32_t backwardPdf; @@ -22,6 +22,7 @@ struct UniformSphereTestResults float32_t backwardWeight; float32_t2 roundtripError; float32_t jacobianProduct; + float32_t inverseJacobianPdf; }; struct UniformSphereTestExecutor @@ -43,7 +44,12 @@ struct UniformSphereTestExecutor output.backwardWeight = sampler.backwardWeight(output.generated); } output.roundtripError = nbl::hlsl::abs(input.u - output.inverted); - output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf; + output.jacobianProduct = computeJacobianProduct(sampler, input.u, 1e-3f, 1.0f); + const float32_t usDiskR = nbl::hlsl::length((float32_t2)output.generated); + const float32_t absZ = nbl::hlsl::abs(output.generated.z); + output.inverseJacobianPdf = (absZ < 0.1f || usDiskR < 0.1f) + ? JACOBIAN_SKIP_CODOMAIN_SINGULARITY + : computeInverseJacobianPdf(sampler, output.generated, output.backwardPdf, 0.0f, 1e30f); } }; diff --git a/37_HLSLSamplingTests/app_resources/shaders/alias_table_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/alias_table_test.comp.hlsl deleted file mode 100644 index 72c4f1977..000000000 --- a/37_HLSLSamplingTests/app_resources/shaders/alias_table_test.comp.hlsl +++ /dev/null @@ -1,77 +0,0 @@ -#pragma shader_stage(compute) - -#include - -#ifdef BENCH_ITERS -#include "../common/discrete_sampler_bench.hlsl" -#include - -[[vk::push_constant]] AliasTablePushConstants pc; - -struct BdaProbabilityAccessor -{ - template && is_integral_v) - void get(I i, NBL_REF_ARG(V) val) { val = vk::RawBufferLoad(addr + uint64_t(sizeof(V)) * uint64_t(i)); } - uint64_t addr; -}; - -struct BdaAliasIndexAccessor -{ - template && is_integral_v) - void get(I i, NBL_REF_ARG(V) val) { val = vk::RawBufferLoad(addr + uint64_t(sizeof(V)) * uint64_t(i)); } - uint64_t addr; -}; - -struct BdaPdfAccessor -{ - template && is_integral_v) - void get(I i, NBL_REF_ARG(V) val) { val = vk::RawBufferLoad(addr + uint64_t(sizeof(V)) * uint64_t(i)); } - uint64_t addr; -}; - -using BenchAliasTable = sampling::AliasTable; -#else -#include "../common/alias_table.hlsl" - -[[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; -[[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; -#endif - -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 -#endif -[numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] -void main() -{ - const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; - -#ifdef BENCH_ITERS - BdaProbabilityAccessor probAcc; - probAcc.addr = pc.probAddress; - BdaAliasIndexAccessor aliasAcc; - aliasAcc.addr = pc.aliasAddress; - BdaPdfAccessor pdfAcc; - pdfAcc.addr = pc.pdfAddress; - BenchAliasTable sampler = BenchAliasTable::create(probAcc, aliasAcc, pdfAcc, pc.tableSize); - - float32_t xi = float32_t(nbl::hlsl::glsl::bitfieldReverse(invID)) / float32_t(~0u); - NBL_CONSTEXPR float32_t goldenRatio = 0.6180339887498949f; - uint32_t acc = 0u; - uint32_t accPdf = 0u; - - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) - { - float32_t u = frac(xi + float32_t(i) * goldenRatio); - BenchAliasTable::cache_type cache; - uint32_t generated = sampler.generate(u, cache); - acc ^= generated; - accPdf ^= asuint(sampler.forwardPdf(u, cache)); - } - - vk::RawBufferStore(pc.outputAddress + uint64_t(sizeof(uint32_t)) * uint64_t(invID), acc + accPdf); -#else - AliasTableTestExecutor executor; - executor(inputTestValues[invID], outputTestValues[invID]); -#endif -} diff --git a/37_HLSLSamplingTests/app_resources/shaders/bilinear_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/bilinear_test.comp.hlsl index 06aad4fdc..420cbcd0b 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/bilinear_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/bilinear_test.comp.hlsl @@ -5,37 +5,42 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; #ifdef BENCH_ITERS - // Perturb coefficients by invID so the sampler is non-uniform across threads. - const float32_t perturbation = float32_t(invID) * 1.0e-7f; - const float32_t4 coeffs = float32_t4(0.25f, 0.5f, 0.75f, 1.0f) + perturbation; - sampling::Bilinear sampler = sampling::Bilinear::create(coeffs); + const float32_t perturbationBase = float32_t(invID) * 1.0e-7f; nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - sampling::Bilinear::cache_type cache; - float32_t2 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y); - acc ^= asuint(sampler.forwardPdf(u, cache)); + const float32_t perturbation = perturbationBase + float32_t(j) * 1.0e-9f; + const float32_t4 coeffs = float32_t4(0.25f, 0.5f, 0.75f, 1.0f) + perturbation; + sampling::Bilinear sampler = sampling::Bilinear::create(coeffs); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::Bilinear::cache_type cache; + float32_t2 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else BilinearTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/box_muller_transform_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/box_muller_transform_test.comp.hlsl index cf0f4065a..3302db2e9 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/box_muller_transform_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/box_muller_transform_test.comp.hlsl @@ -5,37 +5,42 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; #ifdef BENCH_ITERS - // Perturb stddev by invID so the sampler is non-uniform across threads. - const float32_t perturbation = float32_t(invID) * 1.0e-7f; - sampling::BoxMullerTransform sampler = sampling::BoxMullerTransform::create(1.0f + perturbation); + const float32_t perturbationBase = float32_t(invID) * 1.0e-7f; nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - u.x = max(u.x, 1e-7f); - sampling::BoxMullerTransform::cache_type cache; - float32_t2 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y); - acc ^= asuint(sampler.forwardPdf(u, cache)); + const float32_t perturbation = perturbationBase + float32_t(j) * 1.0e-9f; + sampling::BoxMullerTransform sampler = sampling::BoxMullerTransform::create(1.0f + perturbation); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + u.x = max(u.x, 1e-7f); + sampling::BoxMullerTransform::cache_type cache; + float32_t2 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else BoxMullerTransformTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/concentric_mapping_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/concentric_mapping_test.comp.hlsl index 973aba4fe..058c3ef11 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/concentric_mapping_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/concentric_mapping_test.comp.hlsl @@ -5,17 +5,18 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; @@ -23,15 +24,19 @@ void main() nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - sampling::ConcentricMapping::cache_type cache; - float32_t2 generated = sampling::ConcentricMapping::generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y); - acc ^= asuint(sampling::ConcentricMapping::forwardPdf(generated, cache)); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::ConcentricMapping::cache_type cache; + float32_t2 generated = sampling::ConcentricMapping::generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y); + acc ^= asuint(sampling::ConcentricMapping::forwardPdf(generated, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else ConcentricMappingTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/cumulative_probability_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/cumulative_probability_test.comp.hlsl index 2e48adc4a..f06613b49 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/cumulative_probability_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/cumulative_probability_test.comp.hlsl @@ -12,13 +12,18 @@ struct BdaCumProbAccessor { using value_type = float32_t; template - void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC { val = V(vk::RawBufferLoad(addr + uint64_t(sizeof(value_type)) * uint64_t(i))); } - value_type operator[](uint32_t i) NBL_CONST_MEMBER_FUNC { value_type v; get(i, v); return v; } + void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC { val = V(vk::RawBufferLoad(addr + uint64_t(sizeof(value_type)) * uint64_t(i), sizeof(value_type))); } uint64_t addr; }; -using BenchCumProbSampler = sampling::CumulativeProbabilitySampler; +#if defined(NBL_CUMPROB_EYTZINGER) +using BenchCumProbSampler = sampling::CumulativeProbabilitySampler; +#elif defined(NBL_CUMPROB_YOLO_READS) +using BenchCumProbSampler = sampling::CumulativeProbabilitySampler; +#else +using BenchCumProbSampler = sampling::CumulativeProbabilitySampler; +#endif #else #include "../common/cumulative_probability.hlsl" @@ -26,11 +31,7 @@ using BenchCumProbSampler = sampling::CumulativeProbabilitySampler outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 -#endif [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; @@ -46,10 +47,10 @@ void main() for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) { - float32_t u = frac(xi + float32_t(i) * goldenRatio); + xi = frac(xi + goldenRatio); BenchCumProbSampler::cache_type cache; - uint32_t generated = sampler.generate(u, cache); - acc ^= generated ^ asuint(sampler.forwardPdf(u, cache)); + uint32_t generated = sampler.generate(xi, cache); + acc ^= generated ^ asuint(sampler.forwardPdf(xi, cache)); } vk::RawBufferStore(pc.outputAddress + uint64_t(sizeof(uint32_t)) * uint64_t(invID), acc); diff --git a/37_HLSLSamplingTests/app_resources/shaders/linear_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/linear_test.comp.hlsl index 614f339b4..acf0887e5 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/linear_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/linear_test.comp.hlsl @@ -5,37 +5,42 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; #ifdef BENCH_ITERS - // Perturb coefficients by invID so the sampler is non-uniform across threads. - const float32_t perturbation = float32_t(invID) * 1.0e-7f; - const float32_t2 coeffs = float32_t2(0.2f, 0.8f) + perturbation; - sampling::Linear sampler = sampling::Linear::create(coeffs); + const float32_t perturbationBase = float32_t(invID) * 1.0e-7f; nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t u = float32_t(rng()) * toFloat; - sampling::Linear::cache_type cache; - float32_t generated = sampler.generate(u, cache); - acc ^= asuint(generated); - acc ^= asuint(sampler.forwardPdf(u, cache)); + const float32_t perturbation = perturbationBase + float32_t(j) * 1.0e-9f; + const float32_t2 coeffs = float32_t2(0.2f, 0.8f) + perturbation; + sampling::Linear sampler = sampling::Linear::create(coeffs); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t u = float32_t(rng()) * toFloat; + sampling::Linear::cache_type cache; + float32_t generated = sampler.generate(u, cache); + acc ^= asuint(generated); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else LinearTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/packed_alias_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/packed_alias_test.comp.hlsl new file mode 100644 index 000000000..b0dbeedac --- /dev/null +++ b/37_HLSLSamplingTests/app_resources/shaders/packed_alias_test.comp.hlsl @@ -0,0 +1,114 @@ +#pragma shader_stage(compute) + +#include + +#ifdef BENCH_ITERS +#include "../common/discrete_sampler_bench.hlsl" +#include + +[[vk::push_constant]] PackedAliasABPushConstants pc; + +// Log2N bucket. Covers all sweep sizes up to 2^LOG2N buckets without precision +// loss. The same value must be passed to the host-side packA() / +// packB() call so the bit layouts match. +NBL_CONSTEXPR uint32_t LOG2N_BUCKET = 26; + +// Variant A accessor: 4 B packed words. +struct BdaPackedWordAccessor +{ + using value_type = uint32_t; + + template && is_integral_v) + void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC + { + val = vk::RawBufferLoad(addr + uint64_t(sizeof(V)) * uint64_t(i), sizeof(V)); + } + + uint64_t addr; +}; + +// Variant B accessor: 8 B PackedAliasEntryB. Loads a uint2 and decomposes it +// into the POD entry so DXC never sees a bitfield — avoids the Insert/Extract +// round-trip we observed when the sampler read from a bitfield struct. +struct BdaPackedAliasBAccessor +{ + using value_type = nbl::hlsl::sampling::PackedAliasEntryB; + + template) + void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC + { + const uint64_t loadAddr = addr + uint64_t(8u) * uint64_t(i); + const uint2 raw = vk::RawBufferLoad(loadAddr, 8u); + val.packedWord = raw.x; + val.ownPdf = asfloat(raw.y); + } + + uint64_t addr; +}; + +// Separate 4 B pdf[] accessor. +struct BdaPdfAccessor +{ + using value_type = float32_t; + + template && is_integral_v) + void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC + { + val = vk::RawBufferLoad(addr + uint64_t(sizeof(V)) * uint64_t(i), sizeof(V)); + } + + uint64_t addr; +}; + +#ifdef NBL_PACKED_ALIAS_B +using BenchPackedAlias = nbl::hlsl::sampling::PackedAliasTableB; +#else +using BenchPackedAlias = nbl::hlsl::sampling::PackedAliasTableA; +#endif + +#else +#include "../common/alias_table.hlsl" + +[[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; +[[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; +#endif + +[numthreads(WORKGROUP_SIZE, 1, 1)] +void main() +{ + const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; + +#ifdef BENCH_ITERS +#ifdef NBL_PACKED_ALIAS_B + BdaPackedAliasBAccessor entryAcc; +#else + BdaPackedWordAccessor entryAcc; +#endif + entryAcc.addr = pc.entriesAddress; + BdaPdfAccessor pdfAcc; + pdfAcc.addr = pc.pdfAddress; + BenchPackedAlias sampler = BenchPackedAlias::create(entryAcc, pdfAcc, pc.tableSize); + + float32_t xi = float32_t(nbl::hlsl::glsl::bitfieldReverse(invID)) / float32_t(~0u); + NBL_CONSTEXPR float32_t goldenRatio = 0.6180339887498949f; + uint32_t acc = 0u; + + [loop] + for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + { + xi = frac(xi + goldenRatio); + BenchPackedAlias::cache_type cache; + uint32_t generated = sampler.generate(xi, cache); + acc ^= generated ^ asuint(sampler.forwardPdf(xi, cache)); + } + + vk::RawBufferStore(pc.outputAddress + uint64_t(sizeof(uint32_t)) * uint64_t(invID), acc); +#else +#ifdef NBL_PACKED_ALIAS_B + PackedAliasBTestExecutor executor; +#else + PackedAliasATestExecutor executor; +#endif + executor(inputTestValues[invID], outputTestValues[invID]); +#endif +} diff --git a/37_HLSLSamplingTests/app_resources/shaders/polar_mapping_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/polar_mapping_test.comp.hlsl index db7488acd..b12b276e3 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/polar_mapping_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/polar_mapping_test.comp.hlsl @@ -5,17 +5,18 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; @@ -23,15 +24,19 @@ void main() nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - sampling::PolarMapping::cache_type cache; - float32_t2 generated = sampling::PolarMapping::generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y); - acc ^= asuint(sampling::PolarMapping::forwardPdf(generated, cache)); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::PolarMapping::cache_type cache; + float32_t2 generated = sampling::PolarMapping::generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y); + acc ^= asuint(sampling::PolarMapping::forwardPdf(generated, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else PolarMappingTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/projected_hemisphere_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/projected_hemisphere_test.comp.hlsl index 871444955..9be02b9fd 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/projected_hemisphere_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/projected_hemisphere_test.comp.hlsl @@ -5,17 +5,18 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; @@ -23,16 +24,20 @@ void main() nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; sampling::ProjectedHemisphere sampler; - sampling::ProjectedHemisphere::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::ProjectedHemisphere::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else ProjectedHemisphereTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/projected_sphere_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/projected_sphere_test.comp.hlsl index 67a3fa662..7488dc2d5 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/projected_sphere_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/projected_sphere_test.comp.hlsl @@ -5,17 +5,18 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; @@ -23,16 +24,20 @@ void main() nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t3 u = float32_t3(rng(), rng(), rng()) * toFloat; sampling::ProjectedSphere sampler; - sampling::ProjectedSphere::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t3 u = float32_t3(rng(), rng(), rng()) * toFloat; + sampling::ProjectedSphere::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else ProjectedSphereTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl index 903075804..dd7f62db4 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_rectangle_test.comp.hlsl @@ -5,42 +5,69 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +// Number of generate() calls per create(). Default = BENCH_ITERS (persistent: 1 create total). +// Set to 1 for 1:1, 16 for 1:16 multisampling, etc. Must divide BENCH_ITERS. +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void -main() +void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; #ifdef BENCH_ITERS // Perturb rectangle origin by invID so the sampler is non-uniform across threads. - const float32_t perturbation = float32_t(invID) * 1.0e-7f; - shapes::CompressedSphericalRectangle compressed; - compressed.origin = float32_t3(perturbation, perturbation, -2.0f); - compressed.right = float32_t3(1.0f, 0.0f, 0.0f); - compressed.up = float32_t3(0.0f, 1.0f, 0.0f); - shapes::SphericalRectangle rect = shapes::SphericalRectangle::create(compressed); - sampling::ProjectedSphericalRectangle sampler = sampling::ProjectedSphericalRectangle::create(rect, float32_t3(perturbation, 0.0f, 0.0f), float32_t3(0.0f, 0.0f, perturbation + 0.5), false); + const float32_t perturbationBase = float32_t(invID) * 1.0e-7f; nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; +#ifdef BENCH_CREATE_ONLY for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - sampling::ProjectedSphericalRectangle::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + // Depend on i so the compiler can't hoist create() out of the loop. + const float32_t perturbation = perturbationBase + float32_t(i) * 1.0e-9f; + shapes::CompressedSphericalRectangle compressed; + compressed.origin = float32_t3(perturbation, perturbation, -2.0f); + compressed.right = float32_t3(1.0f, 0.0f, 0.0f); + compressed.up = float32_t3(0.0f, 1.0f, 0.0f); + shapes::SphericalRectangle rect = shapes::SphericalRectangle::create(compressed); + sampling::ProjectedSphericalRectangle sampler = sampling::ProjectedSphericalRectangle::create(rect, float32_t3(0.0f, 0.0f, 0.0f), float32_t3(0.0f, 0.0f, perturbation + 0.5), false); + // Read a cheap function of sampler state so create() can't be elided. + sampling::ProjectedSphericalRectangle::cache_type pdfCache; + sampler.generate(float32_t2(0.5f, 0.5f), pdfCache); + acc ^= asuint(sampler.forwardPdf(float32_t2(0.5f, 0.5f), pdfCache)); } - benchOutput.Store(invID * 4u, acc); +#else + // Unified create:generate loop — one create per BENCH_SAMPLES_PER_CREATE generates. + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) + { + const float32_t perturbation = perturbationBase + float32_t(j) * 1.0e-9f; + shapes::CompressedSphericalRectangle compressed; + compressed.origin = float32_t3(perturbation, perturbation, -2.0f); + compressed.right = float32_t3(1.0f, 0.0f, 0.0f); + compressed.up = float32_t3(0.0f, 1.0f, 0.0f); + shapes::SphericalRectangle rect = shapes::SphericalRectangle::create(compressed); + sampling::ProjectedSphericalRectangle sampler = sampling::ProjectedSphericalRectangle::create(rect, float32_t3(0.0f, 0.0f, 0.0f), float32_t3(0.0f, 0.0f, perturbation + 0.5), false); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::ProjectedSphericalRectangle::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } + } +#endif + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else ProjectedSphericalRectangleTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_triangle_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_triangle_test.comp.hlsl index 83e47b3e1..9ed69291a 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_triangle_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/projected_spherical_triangle_test.comp.hlsl @@ -5,39 +5,57 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; #ifdef BENCH_ITERS - // Perturb vertices and normal by invID so the sampler is non-uniform across threads. - const float32_t perturbation = float32_t(invID) * 1.0e-7f; - const float32_t3 verts[3] = { normalize(float32_t3(1.0f, perturbation, 0.0f)), normalize(float32_t3(0.0f, 1.0f, perturbation)), normalize(float32_t3(perturbation, 0.0f, 1.0f)) }; - shapes::SphericalTriangle shape = shapes::SphericalTriangle::createFromUnitSphereVertices(verts); - sampling::ProjectedSphericalTriangle sampler = sampling::ProjectedSphericalTriangle::create(shape, normalize(float32_t3(perturbation, perturbation, 1.0f)), false); + const float32_t perturbationBase = float32_t(invID) * 1.0e-7f; nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; +#ifdef BENCH_CREATE_ONLY for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - sampling::ProjectedSphericalTriangle::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + const float32_t perturbation = perturbationBase + float32_t(i) * 1.0e-9f; + const float32_t3 verts[3] = { normalize(float32_t3(1.0f, perturbation, 0.0f)), normalize(float32_t3(0.0f, 1.0f, perturbation)), normalize(float32_t3(perturbation, 0.0f, 1.0f)) }; + shapes::SphericalTriangle shape = shapes::SphericalTriangle::createFromUnitSphereVertices(verts); + sampling::ProjectedSphericalTriangle sampler = sampling::ProjectedSphericalTriangle::create(shape, normalize(float32_t3(perturbation, perturbation, 1.0f)), false); + sampling::ProjectedSphericalTriangle::cache_type pdfCache; + sampler.generate(float32_t2(0.5f, 0.5f), pdfCache); + acc ^= asuint(sampler.forwardPdf(float32_t2(0.5f, 0.5f), pdfCache)); } - benchOutput.Store(invID * 4u, acc); +#else + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) + { + const float32_t perturbation = perturbationBase + float32_t(j) * 1.0e-9f; + const float32_t3 verts[3] = { normalize(float32_t3(1.0f, perturbation, 0.0f)), normalize(float32_t3(0.0f, 1.0f, perturbation)), normalize(float32_t3(perturbation, 0.0f, 1.0f)) }; + shapes::SphericalTriangle shape = shapes::SphericalTriangle::createFromUnitSphereVertices(verts); + sampling::ProjectedSphericalTriangle sampler = sampling::ProjectedSphericalTriangle::create(shape, normalize(float32_t3(perturbation, perturbation, 1.0f)), false); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::ProjectedSphericalTriangle::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } + } +#endif + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else ProjectedSphericalTriangleTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/spherical_rectangle_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/spherical_rectangle_test.comp.hlsl index 3e9a6fcae..8cba7fbcb 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/spherical_rectangle_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/spherical_rectangle_test.comp.hlsl @@ -5,42 +5,115 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +// Number of generate() calls per create(). Default = BENCH_ITERS (persistent: 1 create total). +// Set to 1 for 1:1 (create+generate per iter), 16 for 1:16 multisampling, etc. Must divide BENCH_ITERS. +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void -main() +void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; #ifdef BENCH_ITERS - // Perturb rectangle origin by invID so the sampler is non-uniform across threads. - const float32_t perturbation = float32_t(invID) * 1.0e-7f; - shapes::CompressedSphericalRectangle compressed; - compressed.origin = float32_t3(perturbation, perturbation, -2.0f); - compressed.right = float32_t3(1.0f, 0.0f, 0.0f); - compressed.up = float32_t3(0.0f, 1.0f, 0.0f); - shapes::SphericalRectangle rect = shapes::SphericalRectangle::create(compressed); - sampling::SphericalRectangle sampler = sampling::SphericalRectangle::create(rect, float32_t3(perturbation, 0.0f, 0.0f)); + // Observer at origin so origin - observer = (p, p, -2) has no zero components: + // keeps all 4 denorm_n_z components perturbation-dependent (no constant-folding). + const float32_t perturbationBase = float32_t(invID) * 1.0e-7f; + +#if (defined(BENCH_VARIANT_SA_EXTENTS) || defined(BENCH_VARIANT_R0_EXTENTS)) && !defined(BENCH_CREATE_ONLY) + // variants 2/3 pre-build: produce a rect (for its basis, sa, extents) once per thread. + shapes::CompressedSphericalRectangle compressedBase; + compressedBase.origin = float32_t3(perturbationBase, perturbationBase, -2.0f); + compressedBase.right = float32_t3(1.0f, 0.0f, 0.0f); + compressedBase.up = float32_t3(0.0f, 1.0f, 0.0f); + const shapes::SphericalRectangle rectBase = shapes::SphericalRectangle::create(compressedBase); + const typename shapes::SphericalRectangle::solid_angle_type saBase = rectBase.solidAngle(float32_t3(0.0f, 0.0f, 0.0f)); + const float32_t2 extentsBase = rectBase.extents; + const matrix basisBase = rectBase.basis; +#endif nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; +#ifdef BENCH_CREATE_ONLY for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - sampling::SphericalRectangle::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + // Depend on i so the compiler can't hoist create() out of the loop. + const float32_t perturbation = perturbationBase + float32_t(i) * 1.0e-9f; + sampling::SphericalRectangle sampler; + #if defined(BENCH_VARIANT_SA_EXTENTS) + shapes::CompressedSphericalRectangle compressed; + compressed.origin = float32_t3(perturbation, perturbation, -2.0f); + compressed.right = float32_t3(1.0f, 0.0f, 0.0f); + compressed.up = float32_t3(0.0f, 1.0f, 0.0f); + shapes::SphericalRectangle rect = shapes::SphericalRectangle::create(compressed); + typename shapes::SphericalRectangle::solid_angle_type sa = rect.solidAngle(float32_t3(0.0f, 0.0f, 0.0f)); + sampler = sampling::SphericalRectangle::create(rect.basis, sa, rect.extents); + #elif defined(BENCH_VARIANT_R0_EXTENTS) + // Build a basis from the same rect geometry so create(basis, r0, extents) has the right frame. + shapes::CompressedSphericalRectangle compressedR0; + compressedR0.origin = float32_t3(perturbation, perturbation, -2.0f); + compressedR0.right = float32_t3(1.0f, 0.0f, 0.0f); + compressedR0.up = float32_t3(0.0f, 1.0f, 0.0f); + const shapes::SphericalRectangle rectR0 = shapes::SphericalRectangle::create(compressedR0); + const float32_t3 r0 = float32_t3(perturbation, perturbation, -2.0f); + const float32_t2 extents = float32_t2(1.0f, 1.0f); + sampler = sampling::SphericalRectangle::create(rectR0.basis, r0, extents); + #else + shapes::CompressedSphericalRectangle compressed; + compressed.origin = float32_t3(perturbation, perturbation, -2.0f); + compressed.right = float32_t3(1.0f, 0.0f, 0.0f); + compressed.up = float32_t3(0.0f, 1.0f, 0.0f); + shapes::SphericalRectangle rect = shapes::SphericalRectangle::create(compressed); + sampler = sampling::SphericalRectangle::create(rect, float32_t3(0.0f, 0.0f, 0.0f)); + #endif + // Read a cheap function of sampler state so create() can't be elided. + acc ^= asuint(sampler.backwardPdf(float32_t3(0.0f, 0.0f, 1.0f))); } - benchOutput.Store(invID * 4u, acc); +#else + // Unified create:generate loop - one create per BENCH_SAMPLES_PER_CREATE generates. + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) + { + const float32_t perturbation = perturbationBase + float32_t(j) * 1.0e-9f; + sampling::SphericalRectangle sampler; + #if defined(BENCH_VARIANT_SA_EXTENTS) + // variant 2: create(basis, sa, extents). Poison one cosGamma so the sincos_accumulator can't be hoisted. + typename shapes::SphericalRectangle::solid_angle_type sa = saBase; + sa.cosGamma[2] += perturbation; + sampler = sampling::SphericalRectangle::create(basisBase, sa, extentsBase); + #elif defined(BENCH_VARIANT_R0_EXTENTS) + // variant 3: create(basis, r0, extents). r0 matches what variant 1 produces. + const float32_t3 r0 = float32_t3(perturbation, perturbation, -2.0f); + const float32_t2 extents = float32_t2(1.0f, 1.0f); + sampler = sampling::SphericalRectangle::create(basisBase, r0, extents); + #else + // variant 1 (default): create(shape, observer). + shapes::CompressedSphericalRectangle compressed; + compressed.origin = float32_t3(perturbation, perturbation, -2.0f); + compressed.right = float32_t3(1.0f, 0.0f, 0.0f); + compressed.up = float32_t3(0.0f, 1.0f, 0.0f); + shapes::SphericalRectangle rect = shapes::SphericalRectangle::create(compressed); + sampler = sampling::SphericalRectangle::create(rect, float32_t3(0.0f, 0.0f, 0.0f)); + #endif + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::SphericalRectangle::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } + } +#endif + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else SphericalRectangleTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/spherical_triangle.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/spherical_triangle.comp.hlsl index 55991bcb3..14b4843b9 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/spherical_triangle.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/spherical_triangle.comp.hlsl @@ -5,39 +5,56 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; #ifdef BENCH_ITERS - // Perturb vertices by invID so the sampler is non-uniform across threads. - const float32_t perturbation = float32_t(invID) * 1.0e-7f; - const float32_t3 verts[3] = { normalize(float32_t3(1.0f, perturbation, 0.0f)), normalize(float32_t3(0.0f, 1.0f, perturbation)), normalize(float32_t3(perturbation, 0.0f, 1.0f)) }; - shapes::SphericalTriangle shape = shapes::SphericalTriangle::createFromUnitSphereVertices(verts); - sampling::SphericalTriangle sampler = sampling::SphericalTriangle::create(shape); + const float32_t perturbationBase = float32_t(invID) * 1.0e-7f; nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; +#ifdef BENCH_CREATE_ONLY for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; - sampling::SphericalTriangle::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + const float32_t perturbation = perturbationBase + float32_t(i) * 1.0e-9f; + const float32_t3 verts[3] = { normalize(float32_t3(1.0f, perturbation, 0.0f)), normalize(float32_t3(0.0f, 1.0f, perturbation)), normalize(float32_t3(perturbation, 0.0f, 1.0f)) }; + shapes::SphericalTriangle shape = shapes::SphericalTriangle::createFromUnitSphereVertices(verts); + sampling::SphericalTriangle sampler = sampling::SphericalTriangle::create(shape); + acc ^= asuint(sampler.backwardPdf(float32_t3(0.0f, 0.0f, 1.0f))); + } +#else + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) + { + const float32_t perturbation = perturbationBase + float32_t(j) * 1.0e-9f; + const float32_t3 verts[3] = { normalize(float32_t3(1.0f, perturbation, 0.0f)), normalize(float32_t3(0.0f, 1.0f, perturbation)), normalize(float32_t3(perturbation, 0.0f, 1.0f)) }; + shapes::SphericalTriangle shape = shapes::SphericalTriangle::createFromUnitSphereVertices(verts); + sampling::SphericalTriangle sampler = sampling::SphericalTriangle::create(shape); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::SphericalTriangle::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); +#endif + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else SphericalTriangleTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/test_compile.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/test_compile.comp.hlsl index 908520243..3c832e995 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/test_compile.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/test_compile.comp.hlsl @@ -1,4 +1,8 @@ +#pragma shader_stage(compute) + // Compile test: instantiate all sampling types and their concept-required methods to verify DXC compilation +#include +#include #include #include #include @@ -9,12 +13,15 @@ #include #include #include +#include +#include +#include +#include "../common/array_accessor.hlsl" using namespace nbl::hlsl; [[vk::binding(0, 0)]] RWStructuredBuffer output; [numthreads(1, 1, 1)] -[shader("compute")] void main() { float32_t2 u2 = float32_t2(0.5, 0.5); @@ -119,7 +126,7 @@ void main() // Octant triangle: all dot products between vertices are 0, so cos_sides=0, csc_sides=1 const float32_t3 triVerts[3] = {float32_t3(1, 0, 0), float32_t3(0, 1, 0), float32_t3(0, 0, 1)}; shapes::SphericalTriangle shapeTri = shapes::SphericalTriangle::createFromUnitSphereVertices(triVerts); - sampling::SphericalTriangle sphTri = sampling::SphericalTriangle::create(shapeTri); + sampling::SphericalTriangle sphTri = sampling::SphericalTriangle::create(shapeTri); sampling::SphericalTriangle::cache_type sphTriCache; float32_t3 stSample = sphTri.generate(u2, sphTriCache); acc.xyz += stSample; @@ -129,7 +136,7 @@ void main() acc.x += sphTri.backwardPdf(stSample); acc.x += sphTri.backwardWeight(stSample); - // SphericalRectangle — generate, forwardPdf, backwardPdf, forwardWeight, backwardWeight + // SphericalRectangle — generate, generateSurfaceOffset, forwardPdf, backwardPdf, forwardWeight, backwardWeight shapes::CompressedSphericalRectangle csr; csr.origin = float32_t3(0.0, 0.0, -1.0); csr.right = float32_t3(1.0, 0.0, 0.0); @@ -140,20 +147,71 @@ void main() sampling::SphericalRectangle::cache_type sphRectCache; float32_t3 srSample = sphRect.generate(u2, sphRectCache); acc.xyz += srSample; + acc.xy += sphRect.generateLocalBasisXY(u2, sphRectCache); acc.x += sphRect.forwardPdf(u2, sphRectCache); acc.x += sphRect.forwardWeight(u2, sphRectCache); acc.x += sphRect.backwardPdf(srSample); acc.x += sphRect.backwardWeight(srSample); - // ProjectedSphericalTriangle — generate, forwardPdf, backwardPdf, forwardWeight, backwardWeight + // ProjectedSphericalTriangle — generate, forwardPdf, forwardWeight, backwardWeight(L) sampling::ProjectedSphericalTriangle projTri = sampling::ProjectedSphericalTriangle::create(shapeTri, float32_t3(0.0, 0.0, 1.0), false); sampling::ProjectedSphericalTriangle::cache_type projTriCache; float32_t3 ptSample = projTri.generate(u2, projTriCache); acc.xyz += ptSample; acc.x += projTri.forwardPdf(u2, projTriCache); acc.x += projTri.forwardWeight(u2, projTriCache); - acc.x += projTri.backwardPdf(ptSample); acc.x += projTri.backwardWeight(ptSample); + // ProjectedSphericalRectangle (UsePdfAsWeight=true) — generate, forwardPdf, forwardWeight, backwardWeight(L) + const float32_t3 psrNormal = float32_t3(0.0, 0.0, 1.0); + sampling::ProjectedSphericalRectangle projRectPdf = + sampling::ProjectedSphericalRectangle::create(shapeRect, srObserver, psrNormal, false); + sampling::ProjectedSphericalRectangle::cache_type projRectPdfCache; + float32_t3 prPdfSample = projRectPdf.generate(u2, projRectPdfCache); + acc.xyz += prPdfSample; + acc.x += projRectPdf.forwardPdf(u2, projRectPdfCache); + acc.x += projRectPdf.forwardWeight(u2, projRectPdfCache); + acc.x += projRectPdf.backwardWeight(prPdfSample); + + // ProjectedSphericalRectangle (UsePdfAsWeight=false) — exercise the MIS-weight path + sampling::ProjectedSphericalRectangle projRectMis = + sampling::ProjectedSphericalRectangle::create(shapeRect, srObserver, psrNormal, true); + sampling::ProjectedSphericalRectangle::cache_type projRectMisCache; + float32_t3 prMisSample = projRectMis.generate(u2, projRectMisCache); + acc.xyz += prMisSample; + acc.x += projRectMis.forwardPdf(u2, projRectMisCache); + acc.x += projRectMis.forwardWeight(u2, projRectMisCache); + acc.x += projRectMis.backwardWeight(prMisSample); + + // AliasTable — generate (with/without cache), forwardPdf, backwardPdf, forwardWeight, backwardWeight + ArrayAccessor aliasProb; + aliasProb.data[0] = 0.25; aliasProb.data[1] = 0.5; aliasProb.data[2] = 0.75; aliasProb.data[3] = 1.0; + ArrayAccessor aliasIdx; + aliasIdx.data[0] = 1u; aliasIdx.data[1] = 2u; aliasIdx.data[2] = 3u; aliasIdx.data[3] = 0u; + ArrayAccessor aliasPdf; + aliasPdf.data[0] = 0.25; aliasPdf.data[1] = 0.25; aliasPdf.data[2] = 0.25; aliasPdf.data[3] = 0.25; + + // CumulativeProbabilitySampler — generate (with/without cache), forwardPdf, backwardPdf, forwardWeight, backwardWeight + ArrayAccessor cumProb; + cumProb.data[0] = 0.25; cumProb.data[1] = 0.5; cumProb.data[2] = 0.75; + sampling::CumulativeProbabilitySampler > cumSampler = + sampling::CumulativeProbabilitySampler >::create(cumProb, 4u); + sampling::CumulativeProbabilitySampler >::cache_type cumCache; + uint32_t cumBin0 = cumSampler.generate(0.6); + uint32_t cumBin = cumSampler.generate(0.6, cumCache); + acc.x += float32_t(cumBin0 + cumBin); + acc.x += cumSampler.forwardPdf(0.6, cumCache); + acc.x += cumSampler.forwardWeight(0.6, cumCache); + acc.x += cumSampler.backwardPdf(cumBin); + acc.x += cumSampler.backwardWeight(cumBin); + + // PartitionRandVariable — operator() partitions u into a left/right branch + sampling::PartitionRandVariable partition; + partition.leftProb = 0.25; + float32_t partXi = 0.5; + float32_t partRcp; + bool partRight = partition(partXi, partRcp); + acc.x += partXi + partRcp + float32_t(partRight ? 1 : 0); + output[0] = acc; } diff --git a/37_HLSLSamplingTests/app_resources/shaders/uniform_hemisphere_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/uniform_hemisphere_test.comp.hlsl index d0990ef43..50901e481 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/uniform_hemisphere_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/uniform_hemisphere_test.comp.hlsl @@ -5,17 +5,18 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; @@ -23,16 +24,20 @@ void main() nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; sampling::UniformHemisphere sampler; - sampling::UniformHemisphere::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::UniformHemisphere::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else UniformHemisphereTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/app_resources/shaders/uniform_sphere_test.comp.hlsl b/37_HLSLSamplingTests/app_resources/shaders/uniform_sphere_test.comp.hlsl index 0d33f5c11..0351e358f 100644 --- a/37_HLSLSamplingTests/app_resources/shaders/uniform_sphere_test.comp.hlsl +++ b/37_HLSLSamplingTests/app_resources/shaders/uniform_sphere_test.comp.hlsl @@ -5,17 +5,18 @@ #include #ifdef BENCH_ITERS -[[vk::binding(1, 0)]] RWByteAddressBuffer benchOutput; +#include "../common/sampler_bench_pc.hlsl" +[[vk::push_constant]] SamplerBenchPushConstants benchPC; #else [[vk::binding(0, 0)]] RWStructuredBuffer inputTestValues; [[vk::binding(1, 0)]] RWStructuredBuffer outputTestValues; #endif -#ifndef WORKGROUP_SIZE -#define WORKGROUP_SIZE 64 +#if !defined(BENCH_SAMPLES_PER_CREATE) && defined(BENCH_ITERS) +#define BENCH_SAMPLES_PER_CREATE (BENCH_ITERS) #endif + [numthreads(WORKGROUP_SIZE, 1, 1)] -[shader("compute")] void main() { const uint32_t invID = nbl::hlsl::glsl::gl_GlobalInvocationID().x; @@ -23,16 +24,20 @@ void main() nbl::hlsl::Xoroshiro64Star rng = nbl::hlsl::Xoroshiro64Star::construct(uint32_t2(invID, 0u)); const float32_t toFloat = asfloat(0x2f800004u); uint32_t acc = 0u; - for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++) + const uint32_t outerIters = uint32_t(BENCH_ITERS) / uint32_t(BENCH_SAMPLES_PER_CREATE); + for (uint32_t j = 0u; j < outerIters; j++) { - float32_t2 u = float32_t2(rng(), rng()) * toFloat; sampling::UniformSphere sampler; - sampling::UniformSphere::cache_type cache; - float32_t3 generated = sampler.generate(u, cache); - acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); - acc ^= asuint(sampler.forwardPdf(u, cache)); + for (uint32_t k = 0u; k < uint32_t(BENCH_SAMPLES_PER_CREATE); k++) + { + float32_t2 u = float32_t2(rng(), rng()) * toFloat; + sampling::UniformSphere::cache_type cache; + float32_t3 generated = sampler.generate(u, cache); + acc ^= asuint(generated.x) ^ asuint(generated.y) ^ asuint(generated.z); + acc ^= asuint(sampler.forwardPdf(u, cache)); + } } - benchOutput.Store(invID * 4u, acc); + vk::RawBufferStore(benchPC.outputAddress + invID * 4u, acc); #else UniformSphereTestExecutor executor; executor(inputTestValues[invID], outputTestValues[invID]); diff --git a/37_HLSLSamplingTests/benchmarks/CDiscreteSamplerBenchmark.h b/37_HLSLSamplingTests/benchmarks/CDiscreteSamplerBenchmark.h index 8f85545b3..f12ba9421 100644 --- a/37_HLSLSamplingTests/benchmarks/CDiscreteSamplerBenchmark.h +++ b/37_HLSLSamplingTests/benchmarks/CDiscreteSamplerBenchmark.h @@ -6,326 +6,247 @@ #include #include #include "app_resources/common/discrete_sampler_bench.hlsl" +#include "nbl/examples/Benchmark/IBenchmark.h" +#include "nbl/examples/Benchmark/GPUBenchmarkHelper.h" #include using namespace nbl; -// Benchmarks alias table vs cumulative probability sampler on the GPU using BDA. -// Builds both tables from the same weight distribution, uploads via BDA buffers, -// and measures GPU throughput using timestamp queries. -class CDiscreteSamplerBenchmark +class CDiscreteSamplerBenchmark : public GPUBenchmark { public: - struct SetupData + // Declared up-front because it's used as the index domain for m_pipelineIdx[] + // (a member-array bound needs the type complete in declaration order). + enum class SamplerKind : uint32_t { - core::smart_refctd_ptr device; - core::smart_refctd_ptr api; - core::smart_refctd_ptr assetMgr; - core::smart_refctd_ptr logger; - video::IPhysicalDevice* physicalDevice; - std::string aliasShaderKey; - std::string cumProbShaderKey; - uint32_t computeFamilyIndex; - uint32_t dispatchGroupCount; - uint32_t tableSize; + AliasPackedA = 0, + AliasPackedB, + CumProbCompare, + CumProbYolo, + CumProbEytzinger, + Count }; - void setup(const SetupData& data) + struct SetupData { - m_device = data.device; - m_logger = data.logger; - m_dispatchGroupCount = data.dispatchGroupCount; - m_tableSize = data.tableSize; - m_physicalDevice = data.physicalDevice; - - m_queue = m_device->getQueue(data.computeFamilyIndex, 0); - - // Command pool + buffers - m_cmdpool = m_device->createCommandPool(data.computeFamilyIndex, video::IGPUCommandPool::CREATE_FLAGS::RESET_COMMAND_BUFFER_BIT); - m_cmdpool->createCommandBuffers(video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_benchCmdbuf); - m_cmdpool->createCommandBuffers(video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_timestampBeforeCmdbuf); - m_cmdpool->createCommandBuffers(video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_timestampAfterCmdbuf); - - // Timestamp query pool - { - video::IQueryPool::SCreationParams qp = {}; - qp.queryType = video::IQueryPool::TYPE::TIMESTAMP; - qp.queryCount = 2; - qp.pipelineStatisticsFlags = video::IQueryPool::PIPELINE_STATISTICS_FLAGS::NONE; - m_queryPool = m_device->createQueryPool(qp); - } - - // Generate random weights - const uint32_t N = m_tableSize; - std::vector weights(N); - std::mt19937 rng(42); - std::uniform_real_distribution dist(0.001f, 100.0f); - for (uint32_t i = 0; i < N; i++) - weights[i] = dist(rng); - - // Build alias table - std::vector aliasProb(N); - std::vector aliasIdx(N); - std::vector aliasPdf(N); - std::vector workspace(N); - nbl::hlsl::sampling::AliasTableBuilder::build({weights}, aliasProb.data(), aliasIdx.data(), aliasPdf.data(), workspace.data()); - - // Build cumulative probability table - std::vector cumProb(N - 1); - nbl::hlsl::sampling::computeNormalizedCumulativeHistogram({weights}, cumProb.data()); - - // Create BDA buffers and upload data - auto createBdaBuffer = [&](const void* srcData, size_t bytes) -> core::smart_refctd_ptr - { - video::IGPUBuffer::SCreationParams bp = {}; - bp.size = bytes; - bp.usage = core::bitflag(video::IGPUBuffer::EUF_STORAGE_BUFFER_BIT) | - video::IGPUBuffer::EUF_SHADER_DEVICE_ADDRESS_BIT; - auto buf = m_device->createBuffer(std::move(bp)); - - video::IDeviceMemoryBacked::SDeviceMemoryRequirements reqs = buf->getMemoryReqs(); - reqs.memoryTypeBits &= data.physicalDevice->getHostVisibleMemoryTypeBits(); - auto alloc = m_device->allocate(reqs, buf.get(), video::IDeviceMemoryAllocation::EMAF_DEVICE_ADDRESS_BIT); + core::smart_refctd_ptr assetMgr; + // Each pipeline is independent; main.cpp can pick precompiled or runtime per + // pipeline by passing ShaderVariant::Precompiled(get_spirv_key<...>()) or + // ShaderVariant::FromSource(path, defines) respectively. + GPUBenchmarkHelper::ShaderVariant packedAliasAVariant; + GPUBenchmarkHelper::ShaderVariant packedAliasBVariant; + GPUBenchmarkHelper::ShaderVariant cumProbVariant; + GPUBenchmarkHelper::ShaderVariant cumProbYoloVariant; + GPUBenchmarkHelper::ShaderVariant cumProbEytzingerVariant; + hlsl::uint32_t3 dispatchGroupCount; + uint64_t targetBudgetMs = 400; // wall-clock budget per sweep row + // N values the sweep cycles through. Dispatch count per row is auto-sized + // by runTimedBudgeted to hit the budget. + std::span sweepNs; + }; - const auto allocSize = alloc.memory->getAllocationSize(); - if (alloc.memory->map({0ull, allocSize}, video::IDeviceMemoryAllocation::EMCAF_WRITE)) - { - std::memcpy(alloc.memory->getMappedPointer(), srcData, bytes); - // Flush so GPU can see the written data - video::ILogicalDevice::MappedMemoryRange flushRange(alloc.memory.get(), 0ull, allocSize); - m_device->flushMappedMemoryRanges(1u, &flushRange); - alloc.memory->unmap(); - } - return buf; + // Shape is derivable from SetupData; expose it so the caller can use it + // both to configure the bench and to build the matching RunContext for the + // span that runs this bench + static WorkloadShape shapeFor(const SetupData& data) + { + const uint32_t totalThreads = data.dispatchGroupCount.x * data.dispatchGroupCount.y * data.dispatchGroupCount.z * WORKGROUP_SIZE; + const uint64_t samplesPerDispatch = uint64_t(totalThreads) * uint64_t(BENCH_ITERS); + return { + .workgroupSize = {WORKGROUP_SIZE, 1u, 1u}, + .dispatchGroupCount = data.dispatchGroupCount, + .samplesPerDispatch = samplesPerDispatch, }; + } - const uint32_t totalThreads = m_dispatchGroupCount * WORKGROUP_SIZE; - - // Alias table buffers - m_aliasProbBuf = createBdaBuffer(aliasProb.data(), N * sizeof(float)); - m_aliasIdxBuf = createBdaBuffer(aliasIdx.data(), N * sizeof(uint32_t)); - m_aliasPdfBuf = createBdaBuffer(aliasPdf.data(), N * sizeof(float)); + CDiscreteSamplerBenchmark(Aggregator& aggregator, const SetupData& data) + : GPUBenchmark(aggregator, GPUBenchmark::SetupData{ + .name = {}, // per-row names synthesized at run time + .warmupDispatches = 0, + .shape = shapeFor(data), + .targetBudgetMs = data.targetBudgetMs, + }) + { + const uint32_t totalThreads = data.dispatchGroupCount.x * data.dispatchGroupCount.y * data.dispatchGroupCount.z * WORKGROUP_SIZE; - // CDF buffer - m_cumProbBuf = createBdaBuffer(cumProb.data(), (N - 1) * sizeof(float)); + m_assetMgr = data.assetMgr; + m_sweepNs = data.sweepNs; - // Shared output buffer + for (const uint32_t N : m_sweepNs) { - video::IGPUBuffer::SCreationParams bp = {}; - bp.size = totalThreads * sizeof(uint32_t); - bp.usage = core::bitflag(video::IGPUBuffer::EUF_STORAGE_BUFFER_BIT) | - video::IGPUBuffer::EUF_SHADER_DEVICE_ADDRESS_BIT; - m_outputBuf = m_device->createBuffer(std::move(bp)); - video::IDeviceMemoryBacked::SDeviceMemoryRequirements reqs = m_outputBuf->getMemoryReqs(); - reqs.memoryTypeBits &= data.physicalDevice->getHostVisibleMemoryTypeBits(); - m_device->allocate(reqs, m_outputBuf.get(), video::IDeviceMemoryAllocation::EMAF_DEVICE_ADDRESS_BIT); + const std::string nStr = std::format("N={}", N); + for (const auto& v : kSweepVariants) + registerVariant({nStr, v.family, v.leaf}); } - // Create pipelines (push constants only, no descriptor sets) - auto loadShader = [&](const std::string& key) - { - asset::IAssetLoader::SAssetLoadParams lp = {}; - lp.logger = m_logger.get(); - lp.workingDirectory = "app_resources"; - auto bundle = data.assetMgr->getAsset(key, lp); - auto source = asset::IAsset::castDown(bundle.getContents()[0]); - return m_device->compileShader({.source = source.get()}); - }; - - // Alias table pipeline - { - const asset::SPushConstantRange pcRange = { - .stageFlags = asset::IShader::E_SHADER_STAGE::ESS_COMPUTE, - .offset = 0, - .size = sizeof(AliasTablePushConstants)}; - auto layout = m_device->createPipelineLayout({&pcRange, 1}); - if (!layout) - m_logger->log("CDiscreteSamplerBenchmark: failed to create alias pipeline layout", system::ILogger::ELL_ERROR); - video::IGPUComputePipeline::SCreationParams pp = {}; - pp.layout = layout.get(); - auto shader = loadShader(data.aliasShaderKey); - if (!shader) - m_logger->log("CDiscreteSamplerBenchmark: failed to load alias shader", system::ILogger::ELL_ERROR); - pp.shader.shader = shader.get(); - pp.shader.entryPoint = "main"; - - if (m_device->getEnabledFeatures().pipelineExecutableInfo) - { - pp.flags |= video::IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_STATISTICS | video::IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_INTERNAL_REPRESENTATIONS; - } - - if (!m_device->createComputePipelines(nullptr, {&pp, 1}, &m_aliasPipeline)) - m_logger->log("CDiscreteSamplerBenchmark: failed to create alias compute pipeline", system::ILogger::ELL_ERROR); + // Shared output buffer (size only depends on thread count). GPU writes via BDA and + // nothing reads it on the CPU. + m_outputBuf = createBdaOutputBuffer(totalThreads * sizeof(uint32_t)).buf; + + // Pipelines (N-independent; only push constants change per run). Indices + // into m_pipelines (GPUBenchmarkHelper) are stored in the same order as SamplerKind + // so the sweep's variant table can index by enum directly. + m_pipelineIdx[static_cast(SamplerKind::AliasPackedA)] = createPipeline(data.packedAliasAVariant, m_assetMgr, sizeof(PackedAliasABPushConstants), "alias-packed-A"); + m_pipelineIdx[static_cast(SamplerKind::AliasPackedB)] = createPipeline(data.packedAliasBVariant, m_assetMgr, sizeof(PackedAliasABPushConstants), "alias-packed-B"); + m_pipelineIdx[static_cast(SamplerKind::CumProbCompare)] = createPipeline(data.cumProbVariant, m_assetMgr, sizeof(CumProbPushConstants), "cumprob-comparator"); + m_pipelineIdx[static_cast(SamplerKind::CumProbYolo)] = createPipeline(data.cumProbYoloVariant, m_assetMgr, sizeof(CumProbPushConstants), "cumprob-yolo"); + m_pipelineIdx[static_cast(SamplerKind::CumProbEytzinger)] = createPipeline(data.cumProbEytzingerVariant, m_assetMgr, sizeof(CumProbPushConstants), "cumprob-eytzinger"); + } - if (m_device->getEnabledFeatures().pipelineExecutableInfo) - { - auto report = system::to_string(m_aliasPipeline->getExecutableInfo()); - m_logger->log("Alias Table Sampling Pipeline Executable Report:\n%s", system::ILogger::ELL_PERFORMANCE, report.c_str()); - } - m_aliasPplnLayout = std::move(layout); - } + // Rows are synthesized per (N, variant), not a single named entry, so + // each row checks cli.focusVariants individually. The aggregator's silent + // flag selects which half (focused / unfocused) we contribute to. + void run() override + { + const bool focusedPhase = isFocusPhase(); + // Warmup is small and fixed; budgeted measurement auto-sizes the + // measured-dispatch count to hit getTargetBudgetMs(). + constexpr uint32_t kWarmupDispatches = 64; - // CDF pipeline + for (const uint32_t N : m_sweepNs) { - const asset::SPushConstantRange pcRange = { - .stageFlags = asset::IShader::E_SHADER_STAGE::ESS_COMPUTE, - .offset = 0, - .size = sizeof(CumProbPushConstants)}; - auto layout = m_device->createPipelineLayout({&pcRange, 1}); - if (!layout) - m_logger->log("CDiscreteSamplerBenchmark: failed to create cumprob pipeline layout", system::ILogger::ELL_ERROR); - video::IGPUComputePipeline::SCreationParams pp = {}; - pp.layout = layout.get(); - auto shader = loadShader(data.cumProbShaderKey); - if (!shader) - m_logger->log("CDiscreteSamplerBenchmark: failed to load cumprob shader", system::ILogger::ELL_ERROR); - pp.shader.shader = shader.get(); - pp.shader.entryPoint = "main"; - if (m_device->getEnabledFeatures().pipelineExecutableInfo) + const std::string nStr = std::format("N={}", N); + bool built = false; + for (const auto& [family, leaf, kind] : kSweepVariants) { - pp.flags |= video::IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_STATISTICS | video::IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_INTERNAL_REPRESENTATIONS; + core::vector name = {nStr, family, leaf}; + const bool inFocus = isFocused(name); + const bool shouldRun = focusedPhase ? inFocus : !inFocus; + if (!shouldRun) + continue; + if (!built) + { + buildAndUpload(N); + built = true; + } + runSingle(N, std::move(name), kind, kWarmupDispatches); } - if (!m_device->createComputePipelines(nullptr, {&pp, 1}, &m_cumProbPipeline)) - m_logger->log("CDiscreteSamplerBenchmark: failed to create cumprob compute pipeline", system::ILogger::ELL_ERROR); - if (m_device->getEnabledFeatures().pipelineExecutableInfo) - { - auto report = system::to_string(m_cumProbPipeline->getExecutableInfo()); - m_logger->log("Cumulative Probability Sampling Pipeline Executable Report:\n%s", system::ILogger::ELL_PERFORMANCE, report.c_str()); - } - m_cumProbPplnLayout = std::move(layout); + if (built) + releaseTables(); } } - void run(uint32_t warmupIterations = 500, uint32_t benchmarkIterations = 5000) - { - constexpr uint32_t benchWorkgroupSize = WORKGROUP_SIZE; - const uint32_t totalThreads = m_dispatchGroupCount * benchWorkgroupSize; - m_logger->log("=== GPU Discrete Sampler Benchmark (N=%u, %u dispatches, %u threads/dispatch, %u iters/thread, ps/sample is per all GPU threads) ===", - system::ILogger::ELL_PERFORMANCE, m_tableSize, benchmarkIterations, totalThreads, BENCH_ITERS); - - runSingle("AliasTable", m_aliasPipeline, m_aliasPplnLayout, true, warmupIterations, benchmarkIterations); - runSingle("CumulativeProbability", m_cumProbPipeline, m_cumProbPplnLayout, false, warmupIterations, benchmarkIterations); - } - private: - void runSingle(const char* name, const core::smart_refctd_ptr& pipeline, const core::smart_refctd_ptr& layout, bool isAlias, uint32_t warmupIterations, uint32_t benchmarkIterations) + // (family, leaf, kind) for every variant the sweep runs. + struct SweepVariant { - m_device->waitIdle(); - - // Record benchmark command buffer - m_benchCmdbuf->reset(video::IGPUCommandBuffer::RESET_FLAGS::NONE); - m_benchCmdbuf->begin(video::IGPUCommandBuffer::USAGE::SIMULTANEOUS_USE_BIT); - m_benchCmdbuf->bindComputePipeline(pipeline.get()); - - if (isAlias) - { - AliasTablePushConstants pc = {}; - pc.probAddress = m_aliasProbBuf->getDeviceAddress(); - pc.aliasAddress = m_aliasIdxBuf->getDeviceAddress(); - pc.pdfAddress = m_aliasPdfBuf->getDeviceAddress(); - pc.outputAddress = m_outputBuf->getDeviceAddress(); - pc.tableSize = m_tableSize; - m_benchCmdbuf->pushConstants(layout.get(), asset::IShader::E_SHADER_STAGE::ESS_COMPUTE, 0u, sizeof(pc), &pc); - } - else - { - CumProbPushConstants pc = {}; - pc.cumProbAddress = m_cumProbBuf->getDeviceAddress(); - pc.outputAddress = m_outputBuf->getDeviceAddress(); - pc.tableSize = m_tableSize; - m_benchCmdbuf->pushConstants(layout.get(), asset::IShader::E_SHADER_STAGE::ESS_COMPUTE, 0u, sizeof(pc), &pc); - } - - m_benchCmdbuf->dispatch(m_dispatchGroupCount, 1, 1); - m_benchCmdbuf->end(); - - // Record timestamp command buffers - m_timestampBeforeCmdbuf->reset(video::IGPUCommandBuffer::RESET_FLAGS::NONE); - m_timestampBeforeCmdbuf->begin(video::IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); - m_timestampBeforeCmdbuf->resetQueryPool(m_queryPool.get(), 0, 2); - m_timestampBeforeCmdbuf->writeTimestamp(asset::PIPELINE_STAGE_FLAGS::NONE, m_queryPool.get(), 0); - m_timestampBeforeCmdbuf->end(); - - m_timestampAfterCmdbuf->reset(video::IGPUCommandBuffer::RESET_FLAGS::NONE); - m_timestampAfterCmdbuf->begin(video::IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); - m_timestampAfterCmdbuf->writeTimestamp(asset::PIPELINE_STAGE_FLAGS::NONE, m_queryPool.get(), 1); - m_timestampAfterCmdbuf->end(); - - auto semaphore = m_device->createSemaphore(0u); - uint64_t semCounter = 0u; - - const video::IQueue::SSubmitInfo::SCommandBufferInfo benchCmds[] = {{.cmdbuf = m_benchCmdbuf.get()}}; - const video::IQueue::SSubmitInfo::SCommandBufferInfo beforeCmds[] = {{.cmdbuf = m_timestampBeforeCmdbuf.get()}}; - const video::IQueue::SSubmitInfo::SCommandBufferInfo afterCmds[] = {{.cmdbuf = m_timestampAfterCmdbuf.get()}}; - - auto submitSerial = [&](const video::IQueue::SSubmitInfo::SCommandBufferInfo* cmds, uint32_t count) - { - const video::IQueue::SSubmitInfo::SSemaphoreInfo waitSem[] = { - {.semaphore = semaphore.get(), .value = semCounter, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT}}; - const video::IQueue::SSubmitInfo::SSemaphoreInfo signalSem[] = { - {.semaphore = semaphore.get(), .value = ++semCounter, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT}}; - video::IQueue::SSubmitInfo submit = {}; - submit.commandBuffers = {cmds, count}; - submit.waitSemaphores = waitSem; - submit.signalSemaphores = signalSem; - m_queue->submit({&submit, 1u}); - }; - - for (uint32_t i = 0u; i < warmupIterations; ++i) - submitSerial(benchCmds, 1u); + const char* family; // e.g. "AliasTable" + const char* leaf; // e.g. "packed A, 4 B" + SamplerKind kind; + }; + static constexpr SweepVariant kSweepVariants[] = { + {"AliasTable", "packed A, 4 B", SamplerKind::AliasPackedA}, + {"AliasTable", "packed B, 8 B", SamplerKind::AliasPackedB}, + {"CumulativeProbability", "comparator", SamplerKind::CumProbCompare}, + {"CumulativeProbability", "YOLO", SamplerKind::CumProbYolo}, + {"CumulativeProbability", "Eytzinger", SamplerKind::CumProbEytzinger}, + }; - submitSerial(beforeCmds, 1u); - for (uint32_t i = 0u; i < benchmarkIterations; ++i) - submitSerial(benchCmds, 1u); - submitSerial(afterCmds, 1u); + void buildAndUpload(const uint32_t N) + { + m_currentN = N; - m_device->waitIdle(); + std::vector weights(N); + std::mt19937 rng(42u + N); + std::uniform_real_distribution dist(0.001f, 100.0f); + for (uint32_t i = 0; i < N; i++) + weights[i] = dist(rng); - uint64_t timestamps[2] = {}; - const auto flags = core::bitflag(video::IQueryPool::RESULTS_FLAGS::_64_BIT) | - core::bitflag(video::IQueryPool::RESULTS_FLAGS::WAIT_BIT); - m_device->getQueryPoolResults(m_queryPool.get(), 0, 2, timestamps, sizeof(uint64_t), flags); + // Build the alias table SoA (intermediate form), then pack it for variants A and B. + // Builder may pad PoT N to N+1 for cache-friendly stride; returned size drives + // every downstream buffer / push-constant value. + std::vector aliasProb; + std::vector aliasIdx; + std::vector aliasPdf; + m_aliasTableN = sampling::AliasTableBuilder::build({weights}, aliasProb, aliasIdx, aliasPdf); + + constexpr uint32_t kPackedLog2N = 26u; + std::vector packedA(m_aliasTableN); + std::vector> packedB(m_aliasTableN); + sampling::AliasTableBuilder::packA({aliasProb}, {aliasIdx}, packedA.data()); + sampling::AliasTableBuilder::packB({aliasProb}, {aliasIdx}, {aliasPdf}, packedB.data()); + + // Cumulative probability (N-1 entries, last bucket implicitly 1.0) + std::vector cumProb(N - 1u); + sampling::computeNormalizedCumulativeHistogram({weights}, cumProb.data()); + + // Eytzinger level-order tree: 2*P entries where P = nextPot(N) + const uint32_t eytzingerP = sampling::eytzingerLeafCount(N); + const uint32_t eytzingerTreeSize = 2u * eytzingerP; + std::vector cumProbEytzinger(eytzingerTreeSize); + sampling::buildEytzinger({weights}, cumProbEytzinger.data()); + + m_aliasPdfBuf = createBdaBuffer(aliasPdf.data(), m_aliasTableN * sizeof(float)); + m_packedAliasABuf = createBdaBuffer(packedA.data(), m_aliasTableN * sizeof(uint32_t)); + m_packedAliasBBuf = createBdaBuffer(packedB.data(), m_aliasTableN * sizeof(sampling::PackedAliasEntryB)); + m_cumProbBuf = createBdaBuffer(cumProb.data(), (N - 1u) * sizeof(float)); + m_cumProbEytzingerBuf = createBdaBuffer(cumProbEytzinger.data(), eytzingerTreeSize * sizeof(float)); + } - constexpr uint32_t benchIters = BENCH_ITERS; - constexpr uint32_t benchWorkgroupSize = WORKGROUP_SIZE; - const float64_t timestampPeriod = float64_t(m_physicalDevice->getLimits().timestampPeriodInNanoSeconds); - const float64_t elapsed_ns = float64_t(timestamps[1] - timestamps[0]) * timestampPeriod; - const uint64_t totalThreads = uint64_t(m_dispatchGroupCount) * uint64_t(benchWorkgroupSize); - const uint64_t totalSamples = uint64_t(benchmarkIterations) * totalThreads * uint64_t(benchIters); - const float64_t ps_per_sample = elapsed_ns * 1e3 / float64_t(totalSamples); - const float64_t gsamples_per_s = float64_t(totalSamples) / elapsed_ns; - const float64_t elapsed_ms = elapsed_ns * 1e-6; + void releaseTables() + { + m_aliasPdfBuf = nullptr; + m_packedAliasABuf = nullptr; + m_packedAliasBBuf = nullptr; + m_cumProbBuf = nullptr; + m_cumProbEytzingerBuf = nullptr; + } - m_logger->log("[Benchmark] %-28s: %9.3f ps/sample | %10.3f GSamples/s | %10.3f ms total", system::ILogger::ELL_PERFORMANCE, name, ps_per_sample, gsamples_per_s, elapsed_ms); + void runSingle(uint32_t N, core::vector name, SamplerKind kind, uint32_t warmupIterations) + { + // Pipeline + push constants are bound *once* in bindOnce, the inner loop is just + // dispatch(...). Putting binds inside dispatchOne would inflate ps/sample on the + // tighter samplers. + const PipelineEntry* pe = getPipelineEntry(m_pipelineIdx[size_t(kind)], joinName(name)); + if (!pe) + return; + + const TimingResult timingResult = runTimedBudgeted(warmupIterations, getTargetBudgetMs(), + [&](IGPUCommandBuffer* cb) + { + if (kind == SamplerKind::AliasPackedA || kind == SamplerKind::AliasPackedB) + { + PackedAliasABPushConstants pc = {}; + pc.entriesAddress = (kind == SamplerKind::AliasPackedA ? m_packedAliasABuf : m_packedAliasBBuf)->getDeviceAddress(); + pc.pdfAddress = m_aliasPdfBuf->getDeviceAddress(); + pc.outputAddress = m_outputBuf->getDeviceAddress(); + pc.tableSize = m_aliasTableN; + defaultBindAndPush(cb, *pe, pc); + } + else + { + CumProbPushConstants pc = {}; + const auto& buf = (kind == SamplerKind::CumProbEytzinger) ? m_cumProbEytzingerBuf : m_cumProbBuf; + pc.cumProbAddress = buf->getDeviceAddress(); + pc.outputAddress = m_outputBuf->getDeviceAddress(); + pc.tableSize = N; + defaultBindAndPush(cb, *pe, pc); + } + }, + [this](IGPUCommandBuffer* cb) { defaultDispatch(cb); }, + samplesForCurrentRow()); + + record(std::move(name), timingResult, pe->stats); } - core::smart_refctd_ptr m_device; - core::smart_refctd_ptr m_logger; - core::smart_refctd_ptr m_cmdpool; - core::smart_refctd_ptr m_benchCmdbuf; - core::smart_refctd_ptr m_timestampBeforeCmdbuf; - core::smart_refctd_ptr m_timestampAfterCmdbuf; - core::smart_refctd_ptr m_queryPool; + core::smart_refctd_ptr m_assetMgr; - // Alias table - core::smart_refctd_ptr m_aliasPplnLayout; - core::smart_refctd_ptr m_aliasPipeline; - core::smart_refctd_ptr m_aliasProbBuf; - core::smart_refctd_ptr m_aliasIdxBuf; - core::smart_refctd_ptr m_aliasPdfBuf; + // Indices into m_pipelines (GPUBenchmarkHelper), indexed by SamplerKind. + uint32_t m_pipelineIdx[size_t(SamplerKind::Count)] = {}; - // Cumulative probability - core::smart_refctd_ptr m_cumProbPplnLayout; - core::smart_refctd_ptr m_cumProbPipeline; - core::smart_refctd_ptr m_cumProbBuf; + // Per-N data buffers (rebuilt each sweep step). pdf[] is shared between A and B. + core::smart_refctd_ptr m_aliasPdfBuf; + core::smart_refctd_ptr m_packedAliasABuf; + core::smart_refctd_ptr m_packedAliasBBuf; + core::smart_refctd_ptr m_cumProbBuf; + core::smart_refctd_ptr m_cumProbEytzingerBuf; // Shared - core::smart_refctd_ptr m_outputBuf; - video::IQueue* m_queue = nullptr; - video::IPhysicalDevice* m_physicalDevice = nullptr; - uint32_t m_dispatchGroupCount = 0; - uint32_t m_tableSize = 0; + core::smart_refctd_ptr m_outputBuf; + uint32_t m_currentN = 0; + uint32_t m_aliasTableN = 0; + std::span m_sweepNs; }; #endif diff --git a/37_HLSLSamplingTests/benchmarks/CSamplerBenchmark.h b/37_HLSLSamplingTests/benchmarks/CSamplerBenchmark.h index 3e2092670..7410b7242 100644 --- a/37_HLSLSamplingTests/benchmarks/CSamplerBenchmark.h +++ b/37_HLSLSamplingTests/benchmarks/CSamplerBenchmark.h @@ -1,4 +1,4 @@ -// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O. +// Copyright (C) 2026 - DevSH Graphics Programming Sp. z O.O. // This file is part of the "Nabla Engine". // For conditions of distribution and use, see copyright notice in nabla.h @@ -7,259 +7,56 @@ #include #include "nbl/examples/examples.hpp" +#include "nbl/examples/Benchmark/IBenchmark.h" +#include "nbl/examples/Benchmark/GPUBenchmarkHelper.h" +#include "app_resources/common/sampler_bench_pc.hlsl" using namespace nbl; // Measures GPU execution time of a sampler shader using GPU timestamp queries. -class CSamplerBenchmark +// Output is implicit BDA addressed via SamplerBenchPushConstants. GPU plumbing +// (pipeline / buffer / timestamp queries) comes from GPUBenchmarkHelper; the +// bench-side glue here is PC layout + per-run dispatch + result recording. +class CSamplerBenchmark : public GPUBenchmark { -public: - struct SetupData - { - core::smart_refctd_ptr device; - core::smart_refctd_ptr api; - core::smart_refctd_ptr assetMgr; - core::smart_refctd_ptr logger; - video::IPhysicalDevice* physicalDevice; - uint32_t computeFamilyIndex; - std::string shaderKey; - uint32_t dispatchGroupCount; // workgroup count = testBatchCount - uint32_t samplesPerDispatch; // dispatchGroupCount * WorkgroupSize * benchIters - size_t inputBufferBytes; // sizeof(InputType) * samplesPerDispatch - size_t outputBufferBytes; // sizeof(ResultType) * samplesPerDispatch - }; - - void setup(const SetupData& data) - { - m_device = data.device; - m_logger = data.logger; - m_dispatchGroupCount = data.dispatchGroupCount; - - // Command pool + 3 command buffers: benchmark (multi-submit), before/after timestamp - m_cmdpool = m_device->createCommandPool(data.computeFamilyIndex, video::IGPUCommandPool::CREATE_FLAGS::RESET_COMMAND_BUFFER_BIT); - if (!m_cmdpool->createCommandBuffers(video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_benchmarkCmdbuf)) - m_logger->log("CSamplerBenchmark: failed to create benchmark cmdbuf", system::ILogger::ELL_ERROR); - if (!m_cmdpool->createCommandBuffers(video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_timestampBeforeCmdbuf)) - m_logger->log("CSamplerBenchmark: failed to create timestamp-before cmdbuf", system::ILogger::ELL_ERROR); - if (!m_cmdpool->createCommandBuffers(video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_timestampAfterCmdbuf)) - m_logger->log("CSamplerBenchmark: failed to create timestamp-after cmdbuf", system::ILogger::ELL_ERROR); - - // Timestamp query pool (2 queries: before and after) - { - video::IQueryPool::SCreationParams qparams = {}; - qparams.queryType = video::IQueryPool::TYPE::TIMESTAMP; - qparams.queryCount = 2; - qparams.pipelineStatisticsFlags = video::IQueryPool::PIPELINE_STATISTICS_FLAGS::NONE; - m_queryPool = m_device->createQueryPool(qparams); - if (!m_queryPool) - m_logger->log("CSamplerBenchmark: failed to create query pool", system::ILogger::ELL_ERROR); - } - - // Load and compile shader - core::smart_refctd_ptr shader; - { - asset::IAssetLoader::SAssetLoadParams lp = {}; - lp.logger = m_logger.get(); - lp.workingDirectory = "app_resources"; - auto bundle = data.assetMgr->getAsset(data.shaderKey, lp); - const auto assets = bundle.getContents(); - if (assets.empty()) - { - m_logger->log("CSamplerBenchmark: failed to load shader", system::ILogger::ELL_ERROR); - return; - } - auto source = asset::IAsset::castDown(assets[0]); - shader = m_device->compileShader({ source.get() }); - } - - // Descriptor set layout: binding 0 = input SSBO, binding 1 = output SSBO - video::IGPUDescriptorSetLayout::SBinding bindings[2] = { - { .binding = 0, .type = asset::IDescriptor::E_TYPE::ET_STORAGE_BUFFER, - .createFlags = video::IGPUDescriptorSetLayout::SBinding::E_CREATE_FLAGS::ECF_NONE, - .stageFlags = ShaderStage::ESS_COMPUTE, .count = 1 }, - { .binding = 1, .type = asset::IDescriptor::E_TYPE::ET_STORAGE_BUFFER, - .createFlags = video::IGPUDescriptorSetLayout::SBinding::E_CREATE_FLAGS::ECF_NONE, - .stageFlags = ShaderStage::ESS_COMPUTE, .count = 1 } - }; - auto dsLayout = m_device->createDescriptorSetLayout(bindings); - - m_pplnLayout = m_device->createPipelineLayout({}, core::smart_refctd_ptr(dsLayout)); - - { - video::IGPUComputePipeline::SCreationParams pparams = {}; - pparams.layout = m_pplnLayout.get(); - pparams.shader.entryPoint = "main"; - pparams.shader.shader = shader.get(); - if (m_device->getEnabledFeatures().pipelineExecutableInfo) - { - pparams.flags |= IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_STATISTICS | IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_INTERNAL_REPRESENTATIONS; - } - if (!m_device->createComputePipelines(nullptr, { &pparams, 1 }, &m_pipeline)) - m_logger->log("CSamplerBenchmark: failed to create compute pipeline", system::ILogger::ELL_ERROR); - - if (m_device->getEnabledFeatures().pipelineExecutableInfo) - m_executableReport = system::to_string(m_pipeline->getExecutableInfo()); - } - - // Allocate input buffer (host-visible, zero-filled, correctness irrelevant for benchmarking) - core::smart_refctd_ptr inputBuf; - { - video::IGPUBuffer::SCreationParams bparams = {}; - bparams.size = data.inputBufferBytes; - bparams.usage = video::IGPUBuffer::EUF_STORAGE_BUFFER_BIT; - inputBuf = m_device->createBuffer(std::move(bparams)); - video::IDeviceMemoryBacked::SDeviceMemoryRequirements reqs = inputBuf->getMemoryReqs(); - reqs.memoryTypeBits &= data.physicalDevice->getHostVisibleMemoryTypeBits(); - m_inputAlloc = m_device->allocate(reqs, inputBuf.get(), video::IDeviceMemoryAllocation::EMAF_NONE); - if (!m_inputAlloc.isValid()) - m_logger->log("CSamplerBenchmark: failed to allocate input buffer memory", system::ILogger::ELL_ERROR); - if (m_inputAlloc.memory->map({ 0ull, m_inputAlloc.memory->getAllocationSize() }, video::IDeviceMemoryAllocation::EMCAF_READ)) - { - std::memset(m_inputAlloc.memory->getMappedPointer(), 0, m_inputAlloc.memory->getAllocationSize()); - m_inputAlloc.memory->unmap(); - } - } - - // Allocate output buffer (host-visible, GPU writes garbage, never read back) - core::smart_refctd_ptr outputBuf; - { - video::IGPUBuffer::SCreationParams bparams = {}; - bparams.size = data.outputBufferBytes; - bparams.usage = video::IGPUBuffer::EUF_STORAGE_BUFFER_BIT; - outputBuf = m_device->createBuffer(std::move(bparams)); - video::IDeviceMemoryBacked::SDeviceMemoryRequirements reqs = outputBuf->getMemoryReqs(); - reqs.memoryTypeBits &= data.physicalDevice->getHostVisibleMemoryTypeBits(); - m_outputAlloc = m_device->allocate(reqs, outputBuf.get(), video::IDeviceMemoryAllocation::EMAF_NONE); - if (!m_outputAlloc.isValid()) - m_logger->log("CSamplerBenchmark: failed to allocate output buffer memory", system::ILogger::ELL_ERROR); - } - - // Descriptor set: bind both buffers - auto pool = m_device->createDescriptorPoolForDSLayouts(video::IDescriptorPool::ECF_NONE, { &dsLayout.get(), 1 }); - m_ds = pool->createDescriptorSet(core::smart_refctd_ptr(dsLayout)); - { - video::IGPUDescriptorSet::SDescriptorInfo info[2]; - info[0].desc = core::smart_refctd_ptr(inputBuf); - info[0].info.buffer = { .offset = 0, .size = data.inputBufferBytes }; - info[1].desc = core::smart_refctd_ptr(outputBuf); - info[1].info.buffer = { .offset = 0, .size = data.outputBufferBytes }; - video::IGPUDescriptorSet::SWriteDescriptorSet writes[2] = { - { .dstSet = m_ds.get(), .binding = 0, .arrayElement = 0, .count = 1, .info = &info[0] }, - { .dstSet = m_ds.get(), .binding = 1, .arrayElement = 0, .count = 1, .info = &info[1] } - }; - m_device->updateDescriptorSets(writes, {}); - } - - m_queue = m_device->getQueue(data.computeFamilyIndex, 0); - m_samplesPerDispatch = data.samplesPerDispatch; - m_physicalDevice = data.physicalDevice; - } - - void logPipelineReport(const std::string& name) const + public: + struct SetupData : GPUBenchmark::SetupData { - if (!m_executableReport.empty()) - m_logger->log("%s Sampler Benchmark Pipeline Executable Report:\n%s", ILogger::ELL_PERFORMANCE, name.c_str(), m_executableReport.c_str()); - } + core::smart_refctd_ptr assetMgr; + GPUBenchmarkHelper::ShaderVariant variant; // precompiled key OR source path + defines + size_t outputBufferBytes; // sizeof(uint32_t) * threadsPerDispatch + }; - // Runs warmupIterations submits (unclocked), then benchmarkIterations submits under GPU timestamps. - void run(const std::string& samplerName, uint32_t warmupIterations = 500, uint32_t benchmarkIterations = 5000) - { - m_device->waitIdle(); - recordBenchmarkCmdBuf(); - recordTimestampCmdBufs(); - - auto semaphore = m_device->createSemaphore(0u); - uint64_t semCounter = 0u; - - const video::IQueue::SSubmitInfo::SCommandBufferInfo benchCmds[] = { {.cmdbuf = m_benchmarkCmdbuf.get()} }; - const video::IQueue::SSubmitInfo::SCommandBufferInfo beforeCmds[] = { {.cmdbuf = m_timestampBeforeCmdbuf.get()} }; - const video::IQueue::SSubmitInfo::SCommandBufferInfo afterCmds[] = { {.cmdbuf = m_timestampAfterCmdbuf.get()} }; - - // Chains submissions via a timeline semaphore so they execute strictly in order - auto submitSerial = [&](const video::IQueue::SSubmitInfo::SCommandBufferInfo* cmds, uint32_t count) - { - const video::IQueue::SSubmitInfo::SSemaphoreInfo waitSem[] = { - {.semaphore = semaphore.get(), .value = semCounter, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT} - }; - const video::IQueue::SSubmitInfo::SSemaphoreInfo signalSem[] = { - {.semaphore = semaphore.get(), .value = ++semCounter, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT} - }; - video::IQueue::SSubmitInfo submit = {}; - submit.commandBuffers = {cmds, count}; - submit.waitSemaphores = waitSem; - submit.signalSemaphores = signalSem; - m_queue->submit({&submit, 1u}); - }; - - for (uint32_t i = 0u; i < warmupIterations; ++i) - submitSerial(benchCmds, 1u); - - submitSerial(beforeCmds, 1u); - for (uint32_t i = 0u; i < benchmarkIterations; ++i) - submitSerial(benchCmds, 1u); - submitSerial(afterCmds, 1u); - - m_device->waitIdle(); - - uint64_t timestamps[2] = {}; - const auto flags = core::bitflag(video::IQueryPool::RESULTS_FLAGS::_64_BIT) | - core::bitflag(video::IQueryPool::RESULTS_FLAGS::WAIT_BIT); - m_device->getQueryPoolResults(m_queryPool.get(), 0, 2, timestamps, sizeof(uint64_t), flags); - - const float64_t timestampPeriod = float64_t(m_physicalDevice->getLimits().timestampPeriodInNanoSeconds); - const float64_t elapsed_ns = float64_t(timestamps[1] - timestamps[0]) * timestampPeriod; - const uint64_t total_samples = uint64_t(benchmarkIterations) * uint64_t(m_samplesPerDispatch); - const float64_t ps_per_sample = elapsed_ns * 1e3 / float64_t(total_samples); - const float64_t gsamples_per_s = float64_t(total_samples) / elapsed_ns; - const float64_t elapsed_ms = elapsed_ns * 1e-6; - - m_logger->log("[Benchmark] %-28s: %9.3f ps/sample | %10.3f GSamples/s | %10.3f ms total", - system::ILogger::ELL_PERFORMANCE, - samplerName.c_str(), ps_per_sample, gsamples_per_s, elapsed_ms); - } - -private: - void recordBenchmarkCmdBuf() - { - m_benchmarkCmdbuf->reset(video::IGPUCommandBuffer::RESET_FLAGS::NONE); - m_benchmarkCmdbuf->begin(video::IGPUCommandBuffer::USAGE::SIMULTANEOUS_USE_BIT); - m_benchmarkCmdbuf->bindComputePipeline(m_pipeline.get()); - m_benchmarkCmdbuf->bindDescriptorSets(asset::EPBP_COMPUTE, m_pplnLayout.get(), 0, 1, &m_ds.get()); - m_benchmarkCmdbuf->dispatch(m_dispatchGroupCount, 1, 1); - m_benchmarkCmdbuf->end(); - } - - void recordTimestampCmdBufs() - { - m_timestampBeforeCmdbuf->reset(video::IGPUCommandBuffer::RESET_FLAGS::NONE); - m_timestampBeforeCmdbuf->begin(video::IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); - m_timestampBeforeCmdbuf->resetQueryPool(m_queryPool.get(), 0, 2); - m_timestampBeforeCmdbuf->writeTimestamp(asset::PIPELINE_STAGE_FLAGS::NONE, m_queryPool.get(), 0); - m_timestampBeforeCmdbuf->end(); + CSamplerBenchmark(Aggregator& aggregator, const SetupData& data) + : GPUBenchmark(aggregator, data) // slicing-copy of the GPUBenchmark::SetupData base + { + auto bda = createBdaOutputBuffer(data.outputBufferBytes); + m_outputBuf = std::move(bda.buf); + m_outputAddress = bda.address; - m_timestampAfterCmdbuf->reset(video::IGPUCommandBuffer::RESET_FLAGS::NONE); - m_timestampAfterCmdbuf->begin(video::IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); - m_timestampAfterCmdbuf->writeTimestamp(asset::PIPELINE_STAGE_FLAGS::NONE, m_queryPool.get(), 1); - m_timestampAfterCmdbuf->end(); - } + m_pipelineIdx = createPipeline(data.variant, data.assetMgr, sizeof(SamplerBenchPushConstants), joinName(data.name)); + } - core::smart_refctd_ptr m_device; - core::smart_refctd_ptr m_logger; - core::smart_refctd_ptr m_cmdpool; - core::smart_refctd_ptr m_benchmarkCmdbuf; - core::smart_refctd_ptr m_timestampBeforeCmdbuf; - core::smart_refctd_ptr m_timestampAfterCmdbuf; - core::smart_refctd_ptr m_queryPool; - core::smart_refctd_ptr m_pplnLayout; - core::smart_refctd_ptr m_pipeline; - core::smart_refctd_ptr m_ds; - video::IDeviceMemoryAllocator::SAllocation m_inputAlloc = {}; - video::IDeviceMemoryAllocator::SAllocation m_outputAlloc = {}; - video::IQueue* m_queue = nullptr; - video::IPhysicalDevice* m_physicalDevice = nullptr; - uint32_t m_dispatchGroupCount = 0; - uint32_t m_samplesPerDispatch = 0; - std::string m_executableReport; + void doRun() override + { + const PipelineEntry* pe = getPipelineEntry(m_pipelineIdx, joinName(m_name)); + if (!pe) + return; + SamplerBenchPushConstants pc = {}; + pc.outputAddress = m_outputAddress; + + const TimingResult t = runTimedBudgeted(getWarmupDispatches(), getTargetBudgetMs(), + [&](video::IGPUCommandBuffer* cb) { defaultBindAndPush(cb, *pe, pc); }, + [this](video::IGPUCommandBuffer* cb) { defaultDispatch(cb); }, + samplesForCurrentRow()); + + record(m_name, t, pe->stats); + } + + private: + core::smart_refctd_ptr m_outputBuf; + uint64_t m_outputAddress = 0; + uint32_t m_pipelineIdx = 0; }; #endif diff --git a/37_HLSLSamplingTests/main.cpp b/37_HLSLSamplingTests/main.cpp index 98ea127cc..9c66ce2e9 100644 --- a/37_HLSLSamplingTests/main.cpp +++ b/37_HLSLSamplingTests/main.cpp @@ -1,5 +1,8 @@ #include +#include +#include + #include "nbl/examples/examples.hpp" #include "nbl/this_example/builtin/build/spirv/keys.hpp" @@ -49,14 +52,14 @@ using namespace nbl::examples; #include "benchmarks/CSamplerBenchmark.h" #include "benchmarks/CDiscreteSamplerBenchmark.h" +#include "nbl/examples/Tester/FailureManifest.h" #include "tests/property/CSamplerPropertyTester.h" -constexpr bool DoBenchmark = true; class HLSLSamplingTests final : public application_templates::MonoDeviceApplication, public BuiltinResourcesApplication { using device_base_t = application_templates::MonoDeviceApplication; - using asset_base_t = BuiltinResourcesApplication; + using asset_base_t = BuiltinResourcesApplication; public: HLSLSamplingTests(const path& _localInputCWD, const path& _localOutputCWD, const path& _sharedInputCWD, const path& _sharedOutputCWD) @@ -64,7 +67,7 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat virtual SPhysicalDeviceFeatures getPreferredDeviceFeatures() const override { - auto retval = device_base_t::getPreferredDeviceFeatures(); + auto retval = device_base_t::getPreferredDeviceFeatures(); retval.pipelineExecutableInfo = true; return retval; } @@ -80,10 +83,10 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat // test compile with dxc { IAssetLoader::SAssetLoadParams lp = {}; - lp.logger = m_logger.get(); - lp.workingDirectory = "app_resources"; - auto key = nbl::this_example::builtin::build::get_spirv_key<"shader">(m_device.get()); - auto bundle = m_assetMgr->getAsset(key.c_str(), lp); + lp.logger = m_logger.get(); + lp.workingDirectory = "app_resources"; + auto key = nbl::this_example::builtin::build::get_spirv_key<"shader">(m_device.get()); + auto bundle = m_assetMgr->getAsset(key.c_str(), lp); const auto assets = bundle.getContents(); if (assets.empty()) @@ -110,12 +113,19 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat // Note: all samplers almost satisfy BasicSampler, but they have cache parameters in generate(). static_assert(sampling::concepts::BasicSampler>); static_assert(sampling::concepts::BasicSampler>); - static_assert(sampling::concepts::BasicSampler); - static_assert(sampling::concepts::BasicSampler); + static_assert(sampling::concepts::BasicSampler, sampling::TRACKING>>); + static_assert(sampling::concepts::BasicSampler, sampling::YOLO>>); + static_assert(sampling::concepts::BasicSampler, sampling::EYTZINGER>>); + static_assert(sampling::concepts::BasicSampler, ReadOnlyAccessor, 26>>); + static_assert(sampling::concepts::BasicSampler, 4>, ReadOnlyAccessor, 26>>); // --- TractableSampler (level 2) --- generate(domain_type, out cache_type) -> codomain_type, forwardPdf(domain_type, cache_type) -> density_type - static_assert(sampling::concepts::TractableSampler); - static_assert(sampling::concepts::TractableSampler); + ; + static_assert(sampling::concepts::TractableSampler, sampling::TRACKING>>); + static_assert(sampling::concepts::TractableSampler, sampling::YOLO>>); + static_assert(sampling::concepts::TractableSampler, sampling::EYTZINGER>>); + static_assert(sampling::concepts::TractableSampler, ReadOnlyAccessor, 26>>); + static_assert(sampling::concepts::TractableSampler, 4>, ReadOnlyAccessor, 26>>); static_assert(sampling::concepts::TractableSampler>); static_assert(sampling::concepts::TractableSampler>); static_assert(sampling::concepts::TractableSampler>); @@ -131,8 +141,11 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat static_assert(sampling::concepts::TractableSampler>); // --- ResamplableSampler (level 3, parallel) --- generate(domain_type, out cache_type) -> codomain_type, forwardWeight(domain_type, cache_type), backwardWeight(codomain_type) - static_assert(sampling::concepts::ResamplableSampler); - static_assert(sampling::concepts::ResamplableSampler); + static_assert(sampling::concepts::ResamplableSampler, sampling::TRACKING>>); + static_assert(sampling::concepts::ResamplableSampler, sampling::YOLO>>); + static_assert(sampling::concepts::ResamplableSampler, sampling::EYTZINGER>>); + static_assert(sampling::concepts::ResamplableSampler, ReadOnlyAccessor, 26>>); + static_assert(sampling::concepts::ResamplableSampler, 4>, ReadOnlyAccessor, 26>>); static_assert(sampling::concepts::ResamplableSampler>); static_assert(sampling::concepts::ResamplableSampler>); static_assert(sampling::concepts::ResamplableSampler>); @@ -155,8 +168,8 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat static_assert(sampling::concepts::BackwardTractableSampler>); static_assert(sampling::concepts::BackwardTractableSampler>); static_assert(sampling::concepts::BackwardTractableSampler>); - static_assert(sampling::concepts::BackwardTractableSampler>); - static_assert(sampling::concepts::BackwardTractableSampler>); + //static_assert(sampling::concepts::BackwardTractableSampler>); // no backwardPdf + //static_assert(sampling::concepts::BackwardTractableSampler>); // no backwardPdf static_assert(sampling::concepts::BackwardTractableSampler>); static_assert(sampling::concepts::BackwardTractableSampler>); static_assert(sampling::concepts::BackwardTractableSampler>); @@ -166,7 +179,7 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat static_assert(sampling::concepts::BijectiveSampler>); static_assert(sampling::concepts::BijectiveSampler>); static_assert(sampling::concepts::BijectiveSampler>); - static_assert(sampling::concepts::BijectiveSampler>); + static_assert(sampling::concepts::BijectiveSampler>); static_assert(sampling::concepts::BijectiveSampler>); static_assert(sampling::concepts::BijectiveSampler>); @@ -177,92 +190,175 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat m_logger->log("All sampling concept tests passed.", ILogger::ELL_INFO); + const auto runControl = nbl::examples::testing::parseRunControl(this->argv, m_logger.get()); + if (!runControl.valid) + return false; + + nbl::examples::testing::FailureManifest failureManifest("37_HLSLSamplingTests"); + // ====================================================================== // GPU throughput benchmarks // ====================================================================== - const uint32_t testBatchCount = 1024; + constexpr uint32_t benchWorkgroupsCount = 4096; + constexpr bool DoBenchmark = true; if constexpr (DoBenchmark) { - constexpr uint32_t benchWorkgroupSize = WORKGROUP_SIZE; - constexpr uint32_t totalThreadsPerDispatch = testBatchCount * benchWorkgroupSize; - constexpr uint32_t iterationsPerThread = BENCH_ITERS; - constexpr uint32_t benchSamplesPerDispatch = totalThreadsPerDispatch * iterationsPerThread; - - struct BenchEntry + if (runControl.skipBenchmarks) { - CSamplerBenchmark bench; - std::string name; + m_logger->log("Skipping benchmark phase due to CLI.", ILogger::ELL_INFO); + } + else + { + constexpr uint32_t benchWorkgroupSize = WORKGROUP_SIZE; + constexpr uint32_t totalThreadsPerDispatch = benchWorkgroupsCount * benchWorkgroupSize; + constexpr uint32_t iterationsPerThread = BENCH_ITERS; + constexpr uint32_t benchSamplesPerDispatch = totalThreadsPerDispatch * iterationsPerThread; + constexpr uint32_t warmupDispatches = 300; // unmeasured warmup + cooldown around the timing window + constexpr uint64_t targetBudgetMs = 400; // wall-clock per row; runTimedBudgeted sizes dispatches + + std::vector benchmarks; + + // Single Aggregator owns results, baselines, formatting, and reporting + // for both bench classes. Passed by reference into each bench's ctor. + Aggregator agg(m_logger, m_device, m_physicalDevice, getComputeQueue()->getFamilyIndex()); + const auto cli = agg.applyCli({ + .argv = this->argv, + .defaultOutputPath = "SamplerBench.json", + .appName = "37_HLSLSamplingTests", + }); + + // One context for the whole sampler-bench span: drives both the per-bench + // shape/budget and the banner that runSessionAndReport prints. + const RunContext samplerCtx = { + .shape = { + .workgroupSize = {benchWorkgroupSize, 1u, 1u}, + .dispatchGroupCount = {benchWorkgroupsCount, 1u, 1u}, + .samplesPerDispatch = benchSamplesPerDispatch, + }, + .targetBudgetMs = targetBudgetMs, + .sectionLabel = "GPU Sampler Benchmarks", }; - std::vector benchmarks; - auto addBench = [&](const char* name, const std::string& shaderKey, size_t inputSize, size_t outputSize) + auto addBench = [&](const std::initializer_list name, GPUBenchmarkHelper::ShaderVariant variant, size_t outputSize) { - auto& entry = benchmarks.emplace_back(); - entry.name = name; - CSamplerBenchmark::SetupData data; - data.device = m_device; - data.api = m_api; - data.assetMgr = m_assetMgr; - data.logger = m_logger; - data.physicalDevice = m_physicalDevice; - data.computeFamilyIndex = getComputeQueue()->getFamilyIndex(); - data.shaderKey = shaderKey; - data.dispatchGroupCount = testBatchCount; - data.samplesPerDispatch = benchSamplesPerDispatch; - data.inputBufferBytes = inputSize; + data.assetMgr = m_assetMgr; + data.name = name; + data.variant = std::move(variant); data.outputBufferBytes = outputSize; - entry.bench.setup(data); + data.warmupDispatches = warmupDispatches; + data.shape = samplerCtx.shape; + data.targetBudgetMs = samplerCtx.targetBudgetMs; + + benchmarks.emplace_back(agg, data); }; - // Bench shaders don't read input (hardcoded values) and write a single uint32_t per thread via RWByteAddressBuffer - constexpr size_t benchInputBytes = sizeof(uint32_t); // unused but binding must exist, didn't bother removing because some samplers need more complex inputs and it's easier to have a consistent buffer setup for all benchmarks - constexpr size_t benchOutputBytes = sizeof(uint32_t) * totalThreadsPerDispatch; - addBench("Linear", nbl::this_example::builtin::build::get_spirv_key<"linear_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("Bilinear", nbl::this_example::builtin::build::get_spirv_key<"bilinear_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("BoxMullerTransform", nbl::this_example::builtin::build::get_spirv_key<"box_muller_transform_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("UniformHemisphere", nbl::this_example::builtin::build::get_spirv_key<"uniform_hemisphere_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("UniformSphere", nbl::this_example::builtin::build::get_spirv_key<"uniform_sphere_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("ConcentricMapping", nbl::this_example::builtin::build::get_spirv_key<"concentric_mapping_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("PolarMapping", nbl::this_example::builtin::build::get_spirv_key<"polar_mapping_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("ProjectedHemisphere", nbl::this_example::builtin::build::get_spirv_key<"projected_hemisphere_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("ProjectedSphere", nbl::this_example::builtin::build::get_spirv_key<"projected_sphere_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("SphericalRectangle", nbl::this_example::builtin::build::get_spirv_key<"spherical_rectangle_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("ProjectedSphericalRectangle", nbl::this_example::builtin::build::get_spirv_key<"projected_spherical_rectangle_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("SphericalTriangle", nbl::this_example::builtin::build::get_spirv_key<"spherical_triangle_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - addBench("ProjectedSphericalTriangle", nbl::this_example::builtin::build::get_spirv_key<"projected_spherical_triangle_bench">(m_device.get()), benchInputBytes, benchOutputBytes); - - // Print all pipeline reports first - for (auto& entry : benchmarks) - entry.bench.logPipelineReport(entry.name); + // Convenience wrappers so the 35+ existing precompiled-key calls below stay + // one line each, and adding a new runtime variant is also a one-liner without + // CMake JSON edits. Both go through the same addBench, just construct the + // ShaderVariant differently. + auto addPrecompiled = [&](std::initializer_list name, size_t outputSize) + { + auto shader = nbl::this_example::builtin::build::get_spirv_key(m_device.get()); + addBench(name, GPUBenchmarkHelper::ShaderVariant::Precompiled(std::move(shader)), outputSize); + }; + auto addRuntime = [&](std::initializer_list name, const char* sourcePath, std::vector defines, size_t outputSize) + { + // Mirror CMake's COMMON_OPTIONS so runtime variants see the same baseline + // as precompiled ones. + std::vector all = { + {"WORKGROUP_SIZE", std::to_string(WORKGROUP_SIZE)}, + {"BENCH_ITERS", std::to_string(BENCH_ITERS)}, + }; + all.insert(all.end(), std::make_move_iterator(defines.begin()), std::make_move_iterator(defines.end())); + addBench(name, GPUBenchmarkHelper::ShaderVariant::FromSource(sourcePath, std::move(all)), outputSize); + }; + + // Bench shaders don't read input -- output is BDA via push constants. + if constexpr (true) + { + constexpr size_t benchOutputBytes = sizeof(uint32_t) * totalThreadsPerDispatch; + addPrecompiled.operator()<"linear_bench_1_1">({"Linear", "Linear", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"linear_bench_1_16">({"Linear", "Linear", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"bilinear_bench_1_1">({"Linear", "Bilinear", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"bilinear_bench_1_16">({"Linear", "Bilinear", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"box_muller_transform_bench_1_1">({"Gaussian", "BoxMullerTransform", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"box_muller_transform_bench_1_16">({"Gaussian", "BoxMullerTransform", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"uniform_hemisphere_bench_1_1">({"SphereSampling", "UniformHemisphere", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"uniform_hemisphere_bench_1_16">({"SphereSampling", "UniformHemisphere", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"uniform_sphere_bench_1_1">({"SphereSampling", "UniformSphere", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"uniform_sphere_bench_1_16">({"SphereSampling", "UniformSphere", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"projected_hemisphere_bench_1_1">({"SphereSampling", "ProjectedHemisphere", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"projected_hemisphere_bench_1_16">({"SphereSampling", "ProjectedHemisphere", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"projected_sphere_bench_1_1">({"SphereSampling", "ProjectedSphere", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"projected_sphere_bench_1_16">({"SphereSampling", "ProjectedSphere", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"concentric_mapping_bench_1_1">({"DiskMappers", "ConcentricMapping", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"concentric_mapping_bench_1_16">({"DiskMappers", "ConcentricMapping", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"polar_mapping_bench_1_1">({"DiskMappers", "PolarMapping", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"polar_mapping_bench_1_16">({"DiskMappers", "PolarMapping", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_1_1_shape_observer">({"SphShapes", "SphRect", "1:1", "shape,observer"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_1_1_sa_extents">({"SphShapes", "SphRect", "1:1", "sa,extents"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_1_1_r0_extents">({"SphShapes", "SphRect", "1:1", "r0,extents"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_1_16_shape_observer">({"SphShapes", "SphRect", "1:16", "shape,observer"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_1_16_sa_extents">({"SphShapes", "SphRect", "1:16", "sa,extents"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_1_16_r0_extents">({"SphShapes", "SphRect", "1:16", "r0,extents"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_create_only_shape_observer">({"SphShapes", "SphRect", "create-only", "shape,observer"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_create_only_sa_extents">({"SphShapes", "SphRect", "create-only", "sa,extents"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_rectangle_bench_create_only_r0_extents">({"SphShapes", "SphRect", "create-only", "r0,extents"}, benchOutputBytes); + addPrecompiled.operator()<"projected_spherical_rectangle_bench_1_1">({"SphShapes", "ProjSphRect", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"projected_spherical_rectangle_bench_1_16">({"SphShapes", "ProjSphRect", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"projected_spherical_rectangle_bench_create_only">({"SphShapes", "ProjSphRect", "create-only"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_triangle_bench_1_1">({"SphShapes", "SphTri", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_triangle_bench_1_16">({"SphShapes", "SphTri", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"spherical_triangle_bench_create_only">({"SphShapes", "SphTri", "create-only"}, benchOutputBytes); + addPrecompiled.operator()<"projected_spherical_triangle_bench_1_1">({"SphShapes", "ProjSphTri", "1:1"}, benchOutputBytes); + addPrecompiled.operator()<"projected_spherical_triangle_bench_1_16">({"SphShapes", "ProjSphTri", "1:16"}, benchOutputBytes); + addPrecompiled.operator()<"projected_spherical_triangle_bench_create_only">({"SphShapes", "ProjSphTri", "create-only"}, benchOutputBytes); + // ---- Runtime-compiled demo variants (no CMake JSON edit needed) ---- + // Same .hlsl source as the precompiled "linear_bench_1_*" entries, but with + // a `BENCH_SAMPLES_PER_CREATE` value that has no JSON entry. Add as many + // here as you want -- each is a one-liner, no reconfigure required. + //addRuntime({"Linear", "Linear", "1:4 (rt)"}, "shaders/linear_test.comp.hlsl", {{"BENCH_SAMPLES_PER_CREATE", "4"}}, benchOutputBytes); + //addRuntime({"Linear", "Linear", "1:8 (rt)"}, "shaders/linear_test.comp.hlsl", {{"BENCH_SAMPLES_PER_CREATE", "8"}}, benchOutputBytes); + } // Discrete sampler benchmark: alias table vs cumulative probability (BDA) { CDiscreteSamplerBenchmark::SetupData dsData; - dsData.device = m_device; - dsData.api = m_api; - dsData.assetMgr = m_assetMgr; - dsData.logger = m_logger; - dsData.physicalDevice = m_physicalDevice; - dsData.computeFamilyIndex = getComputeQueue()->getFamilyIndex(); - dsData.aliasShaderKey = nbl::this_example::builtin::build::get_spirv_key<"alias_table_bench">(m_device.get()); - dsData.cumProbShaderKey = nbl::this_example::builtin::build::get_spirv_key<"cumulative_probability_bench">(m_device.get()); - dsData.dispatchGroupCount = testBatchCount; - dsData.tableSize = 1024; - - CDiscreteSamplerBenchmark discreteBench; - discreteBench.setup(dsData); - - // Then run all benchmarks here so the reports are at the top of the log, followed by timings - constexpr uint32_t warmupDispatches = 500; - constexpr uint32_t benchDispatches = 5000; - m_logger->log("=== GPU Sampler Benchmarks (%u dispatches, %u threads/dispatch, %u iters/thread, ps/sample is per all GPU threads) ===", - ILogger::ELL_PERFORMANCE, benchDispatches, totalThreadsPerDispatch, iterationsPerThread); - for (auto& entry : benchmarks) - entry.bench.run(entry.name, warmupDispatches, benchDispatches); - - discreteBench.run(warmupDispatches, benchDispatches); + dsData.assetMgr = m_assetMgr; + dsData.packedAliasAVariant = GPUBenchmarkHelper::ShaderVariant::Precompiled(nbl::this_example::builtin::build::get_spirv_key<"packed_alias_a_bench">(m_device.get())); + dsData.packedAliasBVariant = GPUBenchmarkHelper::ShaderVariant::Precompiled(nbl::this_example::builtin::build::get_spirv_key<"packed_alias_b_bench">(m_device.get())); + dsData.cumProbVariant = GPUBenchmarkHelper::ShaderVariant::Precompiled(nbl::this_example::builtin::build::get_spirv_key<"cumulative_probability_bench">(m_device.get())); + dsData.cumProbYoloVariant = GPUBenchmarkHelper::ShaderVariant::Precompiled(nbl::this_example::builtin::build::get_spirv_key<"cumulative_probability_yolo_bench">(m_device.get())); + dsData.cumProbEytzingerVariant = GPUBenchmarkHelper::ShaderVariant::Precompiled(nbl::this_example::builtin::build::get_spirv_key<"cumulative_probability_eytzinger_bench">(m_device.get())); + dsData.dispatchGroupCount = {benchWorkgroupsCount, 1u, 1u}; + dsData.targetBudgetMs = targetBudgetMs; + + // Just the N values now -- runTimedBudgeted sizes dispatches per + // row to hit the budget. The old per-N tuning table is gone. + static constexpr uint32_t kSweepNs[] = { + 2u, 4u, 8u, 16u, 32u, 64u, 100u, 128u, 256u, 400u, + 512u, 1024u, 2048u, 2049u, 3000u, 4096u, 7000u, 8192u, 10'000u, 16'384u, 32'768u, + 65'536u, 131'072u, 262'144u, 524'288u, 1'000'000u, 1'048'576u, 2'097'152u, 16'777'216u, 20'971'520u, 25'165'824u, 33'554'432u}; + dsData.sweepNs = kSweepNs; + + CDiscreteSamplerBenchmark discreteBench(agg, dsData); + + const RunContext discreteCtx = { + .shape = CDiscreteSamplerBenchmark::shapeFor(dsData), + .targetBudgetMs = targetBudgetMs, + .sectionLabel = "Discrete Sampler Sweep", + }; + + // Single call. Each span contributes its own focus rows first, then + // every span's unfocused rows -- the aggregator iterates both packs + // in each phase. CDiscrete's overridden run() does per-row filtering + // against cli.focusVariants since its rows aren't a flat list. + agg.runSessionAndReport( + Aggregator::makeSpan(benchmarks, samplerCtx), + Aggregator::makeSpan(discreteBench, discreteCtx)); + } } } @@ -270,57 +366,81 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat // Runtime CPU/GPU comparison tests using ITester harness // ================================================================ bool pass = true; - const uint32_t workgroupSize = WORKGROUP_SIZE; - + constexpr uint32_t testWorkgroupsCount = 4096; + bool samplerPass = true; // generic lambda to run a GPU sampler test - auto runSamplerTest = [&](const char* testName, auto spirvKey, const char* logFile) + auto runSamplerTest = [&](const char* id, const char* testName, const char* logFile) { + if (!runControl.filter.shouldRun(id)) + { + m_logger->log("Skipping %s tests due to filter.", ILogger::ELL_INFO, testName); + return; + } + m_logger->log("Running %s tests...", ILogger::ELL_INFO, testName); typename Tester::PipelineSetupData data; - data.device = m_device; - data.api = m_api; - data.assetMgr = m_assetMgr; - data.logger = m_logger; - data.physicalDevice = m_physicalDevice; + data.device = m_device; + data.api = m_api; + data.assetMgr = m_assetMgr; + data.logger = m_logger; + data.physicalDevice = m_physicalDevice; data.computeFamilyIndex = getComputeQueue()->getFamilyIndex(); - data.shaderKey = spirvKey; - Tester tester(testBatchCount, workgroupSize); + data.shaderKey = std::move(nbl::this_example::builtin::build::get_spirv_key(m_device.get())); + Tester tester(testWorkgroupsCount); tester.setupPipeline(data); - pass &= tester.performTestsAndVerifyResults(logFile); + tester.setFailureRecordContext(&failureManifest, "sampler", id, testName); + if (const auto seed = runControl.filter.seedFor(id); seed.has_value()) + samplerPass &= tester.performTestsAndVerifyResults(logFile, *seed); + else + samplerPass &= tester.performTestsAndVerifyResults(logFile); }; // --- Sampler tests --- if constexpr (true) { - runSamplerTest.operator()("Linear sampler", nbl::this_example::builtin::build::get_spirv_key<"linear_test">(m_device.get()), "LinearTestLog.txt"); - runSamplerTest.operator()("Bilinear sampler", nbl::this_example::builtin::build::get_spirv_key<"bilinear_test">(m_device.get()), "BilinearTestLog.txt"); - runSamplerTest.operator()("UniformHemisphere sampler", nbl::this_example::builtin::build::get_spirv_key<"uniform_hemisphere_test">(m_device.get()), "UniformHemisphereTestLog.txt"); - runSamplerTest.operator()("UniformSphere sampler", nbl::this_example::builtin::build::get_spirv_key<"uniform_sphere_test">(m_device.get()), "UniformSphereTestLog.txt"); - runSamplerTest.operator()("ProjectedHemisphere sampler", nbl::this_example::builtin::build::get_spirv_key<"projected_hemisphere_test">(m_device.get()), "ProjectedHemisphereTestLog.txt"); - runSamplerTest.operator()("ProjectedSphere sampler", nbl::this_example::builtin::build::get_spirv_key<"projected_sphere_test">(m_device.get()), "ProjectedSphereTestLog.txt"); - runSamplerTest.operator()("ConcentricMapping sampler", nbl::this_example::builtin::build::get_spirv_key<"concentric_mapping_test">(m_device.get()), "ConcentricMappingTestLog.txt"); - runSamplerTest.operator()("PolarMapping sampler", nbl::this_example::builtin::build::get_spirv_key<"polar_mapping_test">(m_device.get()), "PolarMappingTestLog.txt"); - runSamplerTest.operator()("BoxMullerTransform sampler", nbl::this_example::builtin::build::get_spirv_key<"box_muller_transform_test">(m_device.get()), "BoxMullerTransformTestLog.txt"); - runSamplerTest.operator()("SphericalTriangle", nbl::this_example::builtin::build::get_spirv_key<"spherical_triangle">(m_device.get()), "SphericalTriangleTestLog.txt"); - runSamplerTest.operator()("ProjectedSphericalTriangle sampler", nbl::this_example::builtin::build::get_spirv_key<"projected_spherical_triangle_test">(m_device.get()), "ProjectedSphericalTriangleTestLog.txt"); - runSamplerTest.operator()("SphericalRectangle sampler", nbl::this_example::builtin::build::get_spirv_key<"spherical_rectangle_test">(m_device.get()), "SphericalRectangleTestLog.txt"); - runSamplerTest.operator()("ProjectedSphericalRectangle sampler", nbl::this_example::builtin::build::get_spirv_key<"projected_spherical_rectangle_test">(m_device.get()), "ProjectedSphericalRectangleTestLog.txt"); + runSamplerTest.operator()("sampler/Linear", "Linear sampler", "LinearTestLog.txt"); + runSamplerTest.operator()("sampler/Bilinear", "Bilinear sampler", "BilinearTestLog.txt"); + runSamplerTest.operator()("sampler/UniformHemisphere", "UniformHemisphere sampler", "UniformHemisphereTestLog.txt"); + runSamplerTest.operator()("sampler/UniformSphere", "UniformSphere sampler", "UniformSphereTestLog.txt"); + runSamplerTest.operator()("sampler/ProjectedHemisphere", "ProjectedHemisphere sampler", "ProjectedHemisphereTestLog.txt"); + runSamplerTest.operator()("sampler/ProjectedSphere", "ProjectedSphere sampler", "ProjectedSphereTestLog.txt"); + runSamplerTest.operator()("sampler/ConcentricMapping", "ConcentricMapping sampler", "ConcentricMappingTestLog.txt"); + runSamplerTest.operator()("sampler/PolarMapping", "PolarMapping sampler", "PolarMappingTestLog.txt"); + runSamplerTest.operator()("sampler/BoxMullerTransform", "BoxMullerTransform sampler", "BoxMullerTransformTestLog.txt"); + runSamplerTest.operator()("sampler/SphericalTriangle", "SphericalTriangle", "SphericalTriangleTestLog.txt"); + runSamplerTest.operator()("sampler/ProjectedSphericalTriangle", "ProjectedSphericalTriangle sampler", "ProjectedSphericalTriangleTestLog.txt"); + runSamplerTest.operator()("sampler/SphericalRectangle", "SphericalRectangle sampler", "SphericalRectangleTestLog.txt"); + runSamplerTest.operator()("sampler/ProjectedSphericalRectangle", "ProjectedSphericalRectangle sampler", "ProjectedSphericalRectangleTestLog.txt"); } if constexpr (true) { // --- Discrete table construction (CPU) --- { - m_logger->log("Running discrete table builder tests (CPU)...", ILogger::ELL_INFO); - CDiscreteTableTester tableTester(m_logger.get()); - pass &= tableTester.run(); + constexpr const char* id = "sampler/DiscreteTableBuilder"; + if (!runControl.filter.shouldRun(id)) + { + m_logger->log("Skipping discrete table builder tests due to filter.", ILogger::ELL_INFO); + } + else + { + m_logger->log("Running discrete table builder tests (CPU)...", ILogger::ELL_INFO); + CDiscreteTableTester tableTester(m_logger.get()); + const bool ok = tableTester.run(); + samplerPass &= ok; + if (!ok) + failureManifest.addGroupFailure("sampler", id, "Discrete table builder"); + } } // --- GPU table sampler tests --- - runSamplerTest.operator()("AliasTable GPU sampler", nbl::this_example::builtin::build::get_spirv_key<"alias_table_test">(m_device.get()), "AliasTableTestLog.txt"); - runSamplerTest.operator()("CumulativeProbability GPU sampler", nbl::this_example::builtin::build::get_spirv_key<"cumulative_probability_test">(m_device.get()), "CumulativeProbabilityTestLog.txt"); + runSamplerTest.operator()("sampler/PackedAliasA", "PackedAliasA GPU sampler", "PackedAliasATestLog.txt"); + runSamplerTest.operator()("sampler/PackedAliasB", "PackedAliasB GPU sampler", "PackedAliasBTestLog.txt"); + runSamplerTest.operator()("sampler/CumulativeProbability", "CumulativeProbability GPU sampler", "CumulativeProbabilityTestLog.txt"); } - if (pass) + logJacobianSkipCounts(m_logger.get()); + pass &= samplerPass; + if (samplerPass) m_logger->log("All sampling tests PASSED.", ILogger::ELL_INFO); else m_logger->log("Some sampling tests FAILED. Check log files for details.", ILogger::ELL_ERROR); @@ -330,66 +450,55 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat // ================================================================ if constexpr (true) { + bool propertyPass = true; m_logger->log("Running sampler property tests (CPU)...", ILogger::ELL_INFO); m_logger->log("WARNING: CPU math may use higher intermediate precision than GPU shaders. Tolerances that pass here may be too tight for GPU.", ILogger::ELL_WARNING); - CSamplerPropertyTester linearProps(m_logger.get()); - pass &= linearProps.run(); - - CSamplerPropertyTester bilinearProps(m_logger.get()); - pass &= bilinearProps.run(); - - CSamplerPropertyTester uniformHemiProps(m_logger.get()); - pass &= uniformHemiProps.run(); - - CSamplerPropertyTester uniformSphereProps(m_logger.get()); - pass &= uniformSphereProps.run(); - - CSamplerPropertyTester projHemiProps(m_logger.get()); - pass &= projHemiProps.run(); - - CSamplerPropertyTester projSphereProps(m_logger.get()); - pass &= projSphereProps.run(); - - CSamplerPropertyTester concentricProps(m_logger.get()); - pass &= concentricProps.run(); - - CSamplerPropertyTester polarProps(m_logger.get()); - pass &= polarProps.run(); - - CSamplerPropertyTester boxMullerProps(m_logger.get()); - pass &= boxMullerProps.run(); - - CSamplerPropertyTester sphTriProps(m_logger.get()); - pass &= sphTriProps.run(); - - CSamplerPropertyTester projSphTriProps(m_logger.get()); - pass &= projSphTriProps.run(); - - CSamplerPropertyTester sphRectProps(m_logger.get()); - pass &= sphRectProps.run(); + auto check = [&]() + { + const std::string id = std::string("property/") + Config::name(); + if (!runControl.filter.shouldRun(id)) + { + m_logger->log("Skipping %s property tests due to filter.", ILogger::ELL_INFO, Config::name()); + return; + } + + CSamplerPropertyTester tester(m_logger.get(), runControl.filter.seedFor(id)); + const bool ok = tester.run(); + propertyPass &= ok; + if (!ok) + { + failureManifest.addGroupFailure("property", id, Config::name()); + if (const auto seed = tester.failureSeed(); seed.has_value()) + failureManifest.addCase("property", id, Config::name(), "property", "CPU", 0, *seed, 0.0, 0.0); + } + }; - CSamplerPropertyTester projSphRectProps(m_logger.get()); - pass &= projSphRectProps.run(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); // Stress tests: extreme coefficient ratios - CSamplerPropertyTester linearStress(m_logger.get()); - pass &= linearStress.run(); - - CSamplerPropertyTester bilinearStress(m_logger.get()); - pass &= bilinearStress.run(); - - CSamplerPropertyTester bilinearPST(m_logger.get()); - pass &= bilinearPST.run(); - - CSamplerPropertyTester sphTriStress(m_logger.get()); - pass &= sphTriStress.run(); + check.operator()(); + check.operator()(); + check.operator()(); + check.operator()(); // Grazing angle tests - CSamplerPropertyTester grazingProps(m_logger.get()); - pass &= grazingProps.run(); + check.operator()(); - if (pass) + pass &= propertyPass; + if (propertyPass) m_logger->log("All sampler property tests PASSED.", ILogger::ELL_INFO); else m_logger->log("Some sampler property tests FAILED.", ILogger::ELL_ERROR); @@ -398,34 +507,43 @@ class HLSLSamplingTests final : public application_templates::MonoDeviceApplicat // ================================================================ // Solid angle accuracy and small triangle convergence tests (CPU-only) // ================================================================ + if constexpr (true) { + bool geometryPass = true; m_logger->log("Running geometry tests (CPU)...", ILogger::ELL_INFO); m_logger->log("WARNING: CPU math may use higher intermediate precision than GPU shaders. Tolerances that pass here may be too tight for GPU.", ILogger::ELL_WARNING); - CSolidAngleAccuracyTester solidAngleTester(m_logger.get()); - pass &= solidAngleTester.run(); - - CSphericalTriangleGenerateTester sphTriGenTester(m_logger.get()); - pass &= sphTriGenTester.run(); - - CSphericalRectangleGenerateTester sphRectGenTester(m_logger.get()); - pass &= sphRectGenTester.run(); - - CProjectedSphericalRectangleGenerateTester projRectGenTester(m_logger.get()); - pass &= projRectGenTester.run(); - - CProjectedSphericalRectangleGeometricTester projRectGeoTester(m_logger.get()); - pass &= projRectGeoTester.run(); + auto check = [&](const char* id, const char* name) + { + if (!runControl.filter.shouldRun(id)) + { + m_logger->log("Skipping %s geometry tests due to filter.", ILogger::ELL_INFO, name); + return; + } + + const bool ok = Tester(m_logger.get()).run(); + geometryPass &= ok; + if (!ok) + failureManifest.addGroupFailure("geometry", id, name); + }; - CProjectedSphericalTriangleGeometricTester pstTester(m_logger.get()); - pass &= pstTester.run(); + check.template operator()("geometry/SolidAngleAccuracy", "SolidAngleAccuracy"); + check.template operator()("geometry/SphericalTriangleGenerate", "SphericalTriangleGenerate"); + check.template operator()("geometry/SphericalRectangleGenerate", "SphericalRectangleGenerate"); + check.template operator()("geometry/ProjectedSphericalRectangleGenerate", "ProjectedSphericalRectangleGenerate"); + check.template operator()("geometry/ProjectedSphericalRectangle", "ProjectedSphericalRectangle"); + check.template operator()("geometry/ProjectedSphericalTriangle", "ProjectedSphericalTriangle"); - if (pass) + pass &= geometryPass; + if (geometryPass) m_logger->log("All geometry tests PASSED.", ILogger::ELL_INFO); else m_logger->log("Some geometry tests FAILED.", ILogger::ELL_ERROR); } + if (!runControl.failedOutPath.empty()) + pass &= nbl::examples::testing::writeFailureManifestFile(failureManifest, runControl.failedOutPath, m_logger.get()); + return pass; } diff --git a/37_HLSLSamplingTests/tests/CAliasTableGPUTester.h b/37_HLSLSamplingTests/tests/CAliasTableGPUTester.h index 87aac65ba..7665ebbb7 100644 --- a/37_HLSLSamplingTests/tests/CAliasTableGPUTester.h +++ b/37_HLSLSamplingTests/tests/CAliasTableGPUTester.h @@ -6,13 +6,31 @@ #include "nbl/examples/Tester/ITester.h" #include "SamplerTestHelpers.h" -class CAliasTableGPUTester final : public ITester +// Shared GPU correctness harness for the packed alias variants. Labels for +// failed-field messages are selected from the Executor type at compile time. +template +class CPackedAliasTableGPUTester final : public ITester { - using base_t = ITester; - using R = AliasTableTestResults; + using base_t = ITester; + using R = AliasTableTestResults; + + using typename base_t::TestType; + using base_t::getRandomEngine; + using base_t::verifyTestValue; + using base_t::printTestFail; + + static constexpr bool kIsA = std::is_same_v; + static constexpr const char* kGeneratedIdxName = kIsA ? "PackedAliasA::generatedIndex" : "PackedAliasB::generatedIndex"; + static constexpr const char* kForwardPdfName = kIsA ? "PackedAliasA::forwardPdf" : "PackedAliasB::forwardPdf"; + static constexpr const char* kBackwardPdfName = kIsA ? "PackedAliasA::backwardPdf" : "PackedAliasB::backwardPdf"; + static constexpr const char* kForwardWeightName = kIsA ? "PackedAliasA::forwardWeight" : "PackedAliasB::forwardWeight"; + static constexpr const char* kBackwardWeightName = kIsA ? "PackedAliasA::backwardWeight" : "PackedAliasB::backwardWeight"; + static constexpr const char* kJacobianName = kIsA ? "PackedAliasA::jacobianProduct" : "PackedAliasB::jacobianProduct"; + static constexpr const char* kPdfConsistencyName = kIsA ? "PackedAliasA::pdf consistency" : "PackedAliasB::pdf consistency"; + static constexpr const char* kWeightConsistencyName = kIsA ? "PackedAliasA::weight consistency" : "PackedAliasB::weight consistency"; public: - CAliasTableGPUTester(const uint32_t testBatchCount, const uint32_t workgroupSize) : base_t(testBatchCount, workgroupSize) {} + CPackedAliasTableGPUTester(const uint32_t testBatchCount) : base_t(testBatchCount, WORKGROUP_SIZE) {} private: AliasTableInputValues generateInputTestValues() override @@ -27,7 +45,7 @@ class CAliasTableGPUTester final : public ITester; +using CPackedAliasBGPUTester = CPackedAliasTableGPUTester; + #endif diff --git a/37_HLSLSamplingTests/tests/CBilinearTester.h b/37_HLSLSamplingTests/tests/CBilinearTester.h index 68605e90a..f5bea6896 100644 --- a/37_HLSLSamplingTests/tests/CBilinearTester.h +++ b/37_HLSLSamplingTests/tests/CBilinearTester.h @@ -14,7 +14,7 @@ class CBilinearTester final : public ITester #include #include +#include // Generic ReadOnly accessor wrapping a raw pointer template + requires std::is_arithmetic_v struct ReadOnlyAccessor { - using value_type = T; - template requires std::is_arithmetic_v - void get(I i, V& val) const { val = V(data[i]); } - T operator[](uint32_t i) const { return data[i]; } + using value_type = T; + template + requires std::is_arithmetic_v + void get(I i, V& val) const { val = V(data[i]); } - const T* data; + const T* data; }; -using ProbabilityAccessor = ReadOnlyAccessor; -using AliasIndexAccessor = ReadOnlyAccessor; -using PdfAccessor = ReadOnlyAccessor; - -using TestAliasTable = nbl::hlsl::sampling::AliasTable; -using TestCumulativeProbabilitySampler = nbl::hlsl::sampling::CumulativeProbabilitySampler>; - // Tests table construction for both alias method and cumulative probability. // Sampler generate/pdf correctness is verified by GPU testers (CAliasTableGPUTester, CCumulativeProbabilityGPUTester). class CDiscreteTableTester { -public: - CDiscreteTableTester(system::ILogger* logger) : m_logger(logger) {} - - bool run() - { - bool pass = true; - auto cases = createTestCases(); - - m_logger->log("AliasTableBuilder tests:", system::ILogger::ELL_INFO); - for (const auto& tc : cases) - pass &= testAliasTable(tc.name, tc.weights); - - m_logger->log("CumulativeProbability tests:", system::ILogger::ELL_INFO); - for (const auto& tc : cases) - pass &= testCumulativeProbability(tc.name, tc.weights); - - return pass; - } - -private: - struct TestCase - { - const char* name; - std::vector weights; - }; - - static std::vector createTestCases() - { - std::vector cases; - cases.push_back({"Uniform(4)", {1.0f, 1.0f, 1.0f, 1.0f}}); - cases.push_back({"NonUniform(1,2,3,4)", {1.0f, 2.0f, 3.0f, 4.0f}}); - - { - std::vector w(32, 1.0f); - w[31] = 97.0f; - cases.push_back({"SingleDominant(32)", std::move(w)}); - } - { - std::vector w(64); - for (uint32_t i = 0; i < 64; i++) - w[i] = 1.0f / float(i + 1); - cases.push_back({"PowerLaw(64)", std::move(w)}); - } - - cases.push_back({"SingleNonZero(4)", {0.0f, 0.0f, 5.0f, 0.0f}}); - - { - std::vector w(1024); - std::mt19937 rng(42); - std::uniform_real_distribution dist(0.001f, 100.0f); - for (uint32_t i = 0; i < 1024; i++) - w[i] = dist(rng); - cases.push_back({"Random(1024)", std::move(w)}); - } - - return cases; - } - - // Verify all values in array are in [0, 1] - bool verifyRange01(const char* prefix, const char* name, const char* arrayName, const float* data, uint32_t count) const - { - bool pass = true; - for (uint32_t i = 0; i < count; i++) - { - if (data[i] < 0.0f || data[i] > 1.0f + 1e-6f) - { - m_logger->log("%s[%s] %s[%u] = %f out of range [0, 1]", - system::ILogger::ELL_ERROR, prefix, name, arrayName, i, data[i]); - pass = false; - } - } - return pass; - } - - // Shared: verify PDFs sum to 1 and each matches weight/totalWeight - bool verifyPdf(const char* prefix, const char* name, const float* pdf, const std::vector& weights) const - { - const uint32_t N = static_cast(weights.size()); - float totalWeight = 0.0f; - for (uint32_t i = 0; i < N; i++) - totalWeight += weights[i]; - - bool pass = true; - - float pdfSum = 0.0f; - for (uint32_t i = 0; i < N; i++) - pdfSum += pdf[i]; - - if (std::abs(pdfSum - 1.0f) > 1e-5f) - { - m_logger->log("%s[%s] PDF sum: expected 1.0, got %f", system::ILogger::ELL_ERROR, prefix, name, pdfSum); - pass = false; - } - - for (uint32_t i = 0; i < N; i++) - { - const float expected = weights[i] / totalWeight; - const float err = std::abs(expected - pdf[i]); - if (err > 1e-6f) - { - m_logger->log("%s[%s] pdf[%u]: expected %f, got %f (err=%e)", system::ILogger::ELL_ERROR, prefix, name, i, expected, pdf[i], err); - pass = false; - } - } - - return pass; - } - - // Verify alias table builder output: - // - bucket contributions reconstruct correct probabilities - // - PDFs sum to 1 and match weight/totalWeight - // - alias indices in range, probabilities in [0, 1] - bool testAliasTable(const char* name, const std::vector& weights) const - { - const uint32_t N = static_cast(weights.size()); - - std::vector outProbability(N); - std::vector outAlias(N); - std::vector outPdf(N); - std::vector workspace(N); - - nbl::hlsl::sampling::AliasTableBuilder::build({ weights },outProbability.data(), outAlias.data(), outPdf.data(), workspace.data()); - - // Accumulate bucket contributions - std::vector dest(N, 0.0f); - for (uint32_t i = 0; i < N; i++) - { - dest[i] += outProbability[i]; - dest[outAlias[i]] += (1.0f - outProbability[i]); - } - - bool pass = true; - - float totalWeight = 0.0f; - for (uint32_t i = 0; i < N; i++) - totalWeight += weights[i]; - - for (uint32_t i = 0; i < N; i++) - { - const float expected = weights[i] / totalWeight * float(N); - const float err = std::abs(expected - dest[i]); - const float tolerance = std::max(1e-5f * float(N), 1e-4f); - - if (err > tolerance) - { - m_logger->log("AliasTable[%s] bucket %u: expected %f, got %f (err=%e)", - system::ILogger::ELL_ERROR, name, i, expected, dest[i], err); - pass = false; - } - } - - // Alias indices in range - for (uint32_t i = 0; i < N; i++) - { - if (outAlias[i] >= N) - { - m_logger->log("AliasTable[%s] alias[%u] = %u out of range [0, %u)", - system::ILogger::ELL_ERROR, name, i, outAlias[i], N); - pass = false; - } - } - - pass &= verifyPdf("AliasTable", name, outPdf.data(), weights); - pass &= verifyRange01("AliasTable", name, "probability", outProbability.data(), N); - - if (pass) - m_logger->log(" [%s] PASSED", system::ILogger::ELL_PERFORMANCE, name); - - return pass; - } - - // Verify CDF table construction: - // - cumulative probabilities are monotonically non-decreasing - // - PDFs match weight/totalWeight - // - PDFs sum to 1 - bool testCumulativeProbability(const char* name, const std::vector& weights) const - { - const uint32_t N = static_cast(weights.size()); - - std::vector cumProb(N - 1); - - nbl::hlsl::sampling::computeNormalizedCumulativeHistogram( - std::span(weights), - cumProb.data()); - - bool pass = true; - - // Monotonically non-decreasing - for (uint32_t i = 1; i < N - 1; i++) - { - if (cumProb[i] < cumProb[i - 1] - 1e-7f) - { - m_logger->log("CumProb[%s] non-monotonic at %u: cumProb[%u]=%f < cumProb[%u]=%f", - system::ILogger::ELL_ERROR, name, i, i, cumProb[i], i - 1, cumProb[i - 1]); - pass = false; - } - } - - // Last stored entry should be < 1.0 (the Nth bucket is implicitly 1.0) - if (N > 1 && cumProb[N - 2] >= 1.0f + 1e-6f) - { - m_logger->log("CumProb[%s] last stored entry %f >= 1.0", - system::ILogger::ELL_ERROR, name, cumProb[N - 2]); - pass = false; - } - - // Derive PDF from CDF for verification - std::vector pdf(N); - for (uint32_t i = 0; i < N; i++) - { - const float cur = (i < N - 1) ? cumProb[i] : 1.0f; - const float prev = (i > 0) ? cumProb[i - 1] : 0.0f; - pdf[i] = cur - prev; - } - - pass &= verifyPdf("CumProb", name, pdf.data(), weights); - pass &= verifyRange01("CumProb", name, "cumProb", cumProb.data(), N - 1); - - if (pass) - m_logger->log(" [%s] PASSED", system::ILogger::ELL_PERFORMANCE, name); - - return pass; - } - - system::ILogger* m_logger; + public: + CDiscreteTableTester(system::ILogger* logger) : m_logger(logger) {} + + bool run() + { + bool pass = true; + auto cases = createTestCases(); + + m_logger->log("AliasTableBuilder tests:", system::ILogger::ELL_INFO); + for (const auto& tc : cases) + pass &= testAliasTable(tc.name, tc.weights); + + m_logger->log("CumulativeProbability tests:", system::ILogger::ELL_INFO); + for (const auto& tc : cases) + pass &= testCumulativeProbability(tc.name, tc.weights); + + m_logger->log("CumulativeProbabilitySampler tests (TRACKING / YOLO / EYTZINGER):", system::ILogger::ELL_INFO); + for (const auto& tc : cases) + pass &= testSamplers(tc.name, tc.weights); + + return pass; + } + + private: + struct TestCase + { + const char* name; + std::vector weights; + }; + + static std::vector createTestCases() + { + std::vector cases; + cases.push_back({"Uniform(4)", {1.0f, 1.0f, 1.0f, 1.0f}}); + cases.push_back({"NonUniform(1,2,3,4)", {1.0f, 2.0f, 3.0f, 4.0f}}); + + { + std::vector w(32, 1.0f); + w[31] = 97.0f; + cases.push_back({"SingleDominant(32)", std::move(w)}); + } + { + std::vector w(64); + for (uint32_t i = 0; i < 64; i++) + w[i] = 1.0f / float(i + 1); + cases.push_back({"PowerLaw(64)", std::move(w)}); + } + + cases.push_back({"SingleNonZero(4)", {0.0f, 0.0f, 5.0f, 0.0f}}); + + { + std::vector w(1024); + std::mt19937 rng(42); + std::uniform_real_distribution dist(0.001f, 100.0f); + for (uint32_t i = 0; i < 1024; i++) + w[i] = dist(rng); + cases.push_back({"Random(1024)", std::move(w)}); + } + + // NPoT cases exercise EYTZINGER padded-leaf territory (P > N). + cases.push_back({"NonPot(7)", {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f}}); + { + std::vector w(1000); + std::mt19937 rng(4242); + std::uniform_real_distribution dist(0.001f, 100.0f); + for (uint32_t i = 0; i < 1000; i++) + w[i] = dist(rng); + cases.push_back({"Random(1000)", std::move(w)}); + } + + return cases; + } + + // Verify all values in array are in [0, 1] + bool verifyRange01(const char* prefix, const char* name, const char* arrayName, const float* data, uint32_t count) const + { + bool pass = true; + for (uint32_t i = 0; i < count; i++) + { + if (data[i] < 0.0f || data[i] > 1.0f + 1e-6f) + { + m_logger->log("%s[%s] %s[%u] = %f out of range [0, 1]", + system::ILogger::ELL_ERROR, prefix, name, arrayName, i, data[i]); + pass = false; + } + } + return pass; + } + + // Shared: verify PDFs sum to 1 and each matches weight/totalWeight + bool verifyPdf(const char* prefix, const char* name, const float* pdf, const std::vector& weights) const + { + const uint32_t N = static_cast(weights.size()); + float totalWeight = 0.0f; + for (uint32_t i = 0; i < N; i++) + totalWeight += weights[i]; + + bool pass = true; + + float pdfSum = 0.0f; + for (uint32_t i = 0; i < N; i++) + pdfSum += pdf[i]; + + if (std::abs(pdfSum - 1.0f) > 1e-5f) + { + m_logger->log("%s[%s] PDF sum: expected 1.0, got %f", system::ILogger::ELL_ERROR, prefix, name, pdfSum); + pass = false; + } + + for (uint32_t i = 0; i < N; i++) + { + const float expected = weights[i] / totalWeight; + const float err = std::abs(expected - pdf[i]); + if (err > 1e-6f) + { + m_logger->log("%s[%s] pdf[%u]: expected %f, got %f (err=%e)", system::ILogger::ELL_ERROR, prefix, name, i, expected, pdf[i], err); + pass = false; + } + } + + return pass; + } + + // Verify alias table builder output: + // - bucket contributions reconstruct correct scaled probabilities + // - PDFs sum to 1 and match weight/totalWeight + // - alias indices in range, probabilities in [0, 1] + // Builder transparently pads PoT N to N+1; actual table size comes back + // as `tableN` and is what gets compared against. + bool testAliasTable(const char* name, const std::vector& weights) const + { + const uint32_t userN = static_cast(weights.size()); + + std::vector outProbability; + std::vector outAlias; + std::vector outPdf; + const uint32_t tableN = nbl::hlsl::sampling::AliasTableBuilder::build({weights}, outProbability, outAlias, outPdf); + + // Accumulate bucket contributions over the full (possibly padded) table + std::vector dest(tableN, 0.0f); + for (uint32_t i = 0; i < tableN; i++) + { + dest[i] += outProbability[i]; + dest[outAlias[i]] += (1.0f - outProbability[i]); + } + + bool pass = true; + + float totalWeight = 0.0f; + for (uint32_t i = 0; i < userN; i++) + totalWeight += weights[i]; + + // Real buckets: expected scaled prob = weight/total * tableN + for (uint32_t i = 0; i < userN; i++) + { + const float expected = weights[i] / totalWeight * float(tableN); + const float err = std::abs(expected - dest[i]); + const float tolerance = std::max(1e-5f * float(tableN), 1e-4f); + + if (err > tolerance) + { + m_logger->log("AliasTable[%s] bucket %u: expected %f, got %f (err=%e)", + system::ILogger::ELL_ERROR, name, i, expected, dest[i], err); + pass = false; + } + } + + // Dummy bucket (only when padded): no real bucket aliases to it -> dest[userN] should be 0. + if (tableN != userN && std::abs(dest[userN]) > 1e-4f) + { + m_logger->log("AliasTable[%s] dummy bucket %u has non-zero reconstructed probability %f", + system::ILogger::ELL_ERROR, name, userN, dest[userN]); + pass = false; + } + + // Alias indices in range [0, tableN) + for (uint32_t i = 0; i < tableN; i++) + { + if (outAlias[i] >= tableN) + { + m_logger->log("AliasTable[%s] alias[%u] = %u out of range [0, %u)", + system::ILogger::ELL_ERROR, name, i, outAlias[i], tableN); + pass = false; + } + } + + pass &= verifyPdf("AliasTable", name, outPdf.data(), weights); + pass &= verifyRange01("AliasTable", name, "probability", outProbability.data(), tableN); + + if (pass) + m_logger->log(" [%s] PASSED", system::ILogger::ELL_PERFORMANCE, name); + + return pass; + } + + // Verify CDF table construction: monotonicity, implicit-1.0 invariant, and + // stored entries in [0, 1]. PDF-from-CDF correctness is covered by the + // TRACKING sampler test below (same cdf[i] - cdf[i-1] derivation via + // sampler.backwardPdf), so it's not repeated here. + bool testCumulativeProbability(const char* name, const std::vector& weights) const + { + const uint32_t N = static_cast(weights.size()); + + std::vector cumProb(N - 1); + + nbl::hlsl::sampling::computeNormalizedCumulativeHistogram(std::span(weights), cumProb.data()); + + bool pass = true; + + // Monotonically non-decreasing + for (uint32_t i = 1; i < N - 1; i++) + { + if (cumProb[i] < cumProb[i - 1] - 1e-7f) + { + m_logger->log("CumProb[%s] non-monotonic at %u: cumProb[%u]=%f < cumProb[%u]=%f", + system::ILogger::ELL_ERROR, name, i, i, cumProb[i], i - 1, cumProb[i - 1]); + pass = false; + } + } + + // Last stored entry should be < 1.0 (the Nth bucket is implicitly 1.0) + if (N > 1 && cumProb[N - 2] >= 1.0f + 1e-6f) + { + m_logger->log("CumProb[%s] last stored entry %f >= 1.0", system::ILogger::ELL_ERROR, name, cumProb[N - 2]); + pass = false; + } + + pass &= verifyRange01("CumProb", name, "cumProb", cumProb.data(), N - 1); + + if (pass) + m_logger->log(" [%s] PASSED", system::ILogger::ELL_PERFORMANCE, name); + + return pass; + } + + // Reference binary search over the full N-entry CDF (last entry == 1.0). + static uint32_t referenceUpperBound(const std::vector& fullCdf, float u) + { + auto it = std::upper_bound(fullCdf.begin(), fullCdf.end(), u); + return static_cast(std::distance(fullCdf.begin(), it)); + } + + // Run TRACKING, YOLO, and EYTZINGER samplers against the same reference + // distribution. Each mode is instantiated via the dual-compile sampler and + // exercised entirely on the CPU. + bool testSamplers(const char* name, const std::vector& weights) const + { + const uint32_t N = static_cast(weights.size()); + if (N < 2) + return true; + + float totalWeight = 0.0f; + for (uint32_t i = 0; i < N; i++) + totalWeight += weights[i]; + const float rcpTotal = 1.0f / totalWeight; + + std::vector pdfRef(N); + std::vector fullCdf(N); + float acc = 0.0f; + for (uint32_t i = 0; i < N; i++) + { + pdfRef[i] = weights[i] * rcpTotal; + acc += pdfRef[i]; + fullCdf[i] = acc; + } + fullCdf[N - 1] = 1.0f; // pin the last entry; reference must treat it as exact + + // Storage for TRACKING / YOLO (N-1 entries, last bucket implicit at 1.0). + std::vector cdfStorage(N - 1); + nbl::hlsl::sampling::computeNormalizedCumulativeHistogram({weights}, cdfStorage.data()); + + // Storage for EYTZINGER (2*P entries, level-order implicit binary tree). + const uint32_t P = nbl::hlsl::sampling::eytzingerLeafCount(N); + std::vector treeStorage(2u * P, 0.0f); + nbl::hlsl::sampling::buildEytzinger({weights}, treeStorage.data()); + + bool pass = true; + pass &= testSamplerMode("TRACKING", name, N, pdfRef, fullCdf, cdfStorage.data()); + pass &= testSamplerMode("YOLO", name, N, pdfRef, fullCdf, cdfStorage.data()); + pass &= testSamplerMode("EYTZINGER", name, N, pdfRef, fullCdf, treeStorage.data()); + return pass; + } + + template + bool testSamplerMode(const char* modeName, const char* caseName, uint32_t N, + const std::vector& pdfRef, const std::vector& fullCdf, const float* accessorData) const + { + using Sampler = nbl::hlsl::sampling::CumulativeProbabilitySampler< + float, float, uint32_t, ReadOnlyAccessor, Mode>; + + ReadOnlyAccessor accessor {accessorData}; + Sampler sampler = Sampler::create(accessor, N); + + bool pass = true; + + // backwardPdf(v) == pdfRef[v], and the implied PDF sums to 1. + float backwardSum = 0.0f; + for (uint32_t v = 0; v < N; v++) + { + const float got = sampler.backwardPdf(v); + const float expected = pdfRef[v]; + const float err = std::abs(got - expected); + const float tol = 1e-5f; + if (err > tol) + { + m_logger->log("Sampler[%s][%s] backwardPdf[%u]: expected %e, got %e (err=%e)", + system::ILogger::ELL_ERROR, modeName, caseName, v, expected, got, err); + pass = false; + } + backwardSum += got; + } + if (std::abs(backwardSum - 1.0f) > 1e-5f) + { + m_logger->log("Sampler[%s][%s] backwardPdf sum: expected 1.0, got %f", + system::ILogger::ELL_ERROR, modeName, caseName, backwardSum); + pass = false; + } + + // generate(u) lands in the correct bucket for a grid of u values, and + // generate(u, cache) produces forwardPdf matching backwardPdf(result). + std::mt19937 rng(1234u + N); + std::uniform_real_distribution udist(0.0f, std::nextafter(1.0f, 0.0f)); + constexpr uint32_t kTrials = 2048; + + for (uint32_t k = 0; k < kTrials; k++) + { + const float u = udist(rng); + const uint32_t ref = referenceUpperBound(fullCdf, u); + + const uint32_t idx = sampler.generate(u); + if (idx != ref) + { + m_logger->log("Sampler[%s][%s] generate(%.7f): expected bucket %u, got %u", + system::ILogger::ELL_ERROR, modeName, caseName, u, ref, idx); + pass = false; + continue; + } + + typename Sampler::cache_type cache; + const uint32_t idxCache = sampler.generate(u, cache); + if (idxCache != ref) + { + m_logger->log("Sampler[%s][%s] generate(u,cache)(%.7f): expected %u, got %u", + system::ILogger::ELL_ERROR, modeName, caseName, u, ref, idxCache); + pass = false; + continue; + } + + const float forwardP = sampler.forwardPdf(u, cache); + const float backwardP = sampler.backwardPdf(idxCache); + if (std::abs(forwardP - backwardP) > 1e-6f) + { + m_logger->log("Sampler[%s][%s] fwd/bwd pdf mismatch at u=%.7f bucket=%u: fwd=%e bwd=%e", + system::ILogger::ELL_ERROR, modeName, caseName, u, idxCache, forwardP, backwardP); + pass = false; + } + } + + if (pass) + m_logger->log(" [%-9s %s] PASSED", system::ILogger::ELL_PERFORMANCE, modeName, caseName); + return pass; + } + + system::ILogger* m_logger; }; #endif diff --git a/37_HLSLSamplingTests/tests/CLinearTester.h b/37_HLSLSamplingTests/tests/CLinearTester.h index 631151f00..394b68721 100644 --- a/37_HLSLSamplingTests/tests/CLinearTester.h +++ b/37_HLSLSamplingTests/tests/CLinearTester.h @@ -14,7 +14,7 @@ class CLinearTester final : public ITesterlog(" coeffStart=%s coeffEnd=%s", nbl::system::ILogger::ELL_ERROR, - to_string(s.linearCoeffStart).c_str(), to_string(s.linearCoeffEnd).c_str()); + to_string(s.normalizedCoeffStart).c_str(), to_string(s.normalizedCoeffEnd).c_str()); } }; @@ -140,7 +141,7 @@ struct LinearStressConfig { using nbl::system::to_string; logger->log(" coeffStart=%s coeffEnd=%s", nbl::system::ILogger::ELL_ERROR, - to_string(s.linearCoeffStart).c_str(), to_string(s.linearCoeffEnd).c_str()); + to_string(s.normalizedCoeffStart).c_str(), to_string(s.normalizedCoeffEnd).c_str()); } }; diff --git a/37_HLSLSamplingTests/tests/CPolarMappingTester.h b/37_HLSLSamplingTests/tests/CPolarMappingTester.h index f7009176b..13971e186 100644 --- a/37_HLSLSamplingTests/tests/CPolarMappingTester.h +++ b/37_HLSLSamplingTests/tests/CPolarMappingTester.h @@ -14,7 +14,7 @@ class CPolarMappingTester final : public ITester sizeDist(0.5f, 3.0f); std::uniform_real_distribution uDist(0.0f, 1.0f); - ProjectedSphericalRectangleInputValues input; - // Observer at origin, rect placed in front (negative Z) so the solid angle is valid. - input.observer = nbl::hlsl::float32_t3(0.0f, 0.0f, 0.0f); - const float width = sizeDist(getRandomEngine()); - const float height = sizeDist(getRandomEngine()); - input.rectOrigin = nbl::hlsl::float32_t3(0.0f, 0.0f, -2.0f); - input.right = nbl::hlsl::float32_t3(width, 0.0f, 0.0f); - input.up = nbl::hlsl::float32_t3(0.0f, height, 0.0f); - - // Build shape to use centralized corner check nbl::hlsl::shapes::CompressedSphericalRectangle compressed; - compressed.origin = input.rectOrigin; - compressed.right = input.right; - compressed.up = input.up; + nbl::hlsl::float32_t3 observer; + generateRandomRectangle(getRandomEngine(), compressed, observer); + + ProjectedSphericalRectangleInputValues input; + input.observer = observer; + input.rectOrigin = compressed.origin; + input.right = compressed.right; + input.up = compressed.up; + auto shape = nbl::hlsl::shapes::SphericalRectangle::create(compressed); // Ensure the receiver normal has positive projection onto at least one vertex, @@ -63,25 +58,25 @@ class CProjectedSphericalRectangleTester final : public ITester actual.extents.x || - actual.surfaceOffset.y < 0.0f || actual.surfaceOffset.y > actual.extents.y) + PdfCheck {"ProjectedSphericalRectangle::forwardPdf", &R::forwardPdf}); + VERIFY_JACOBIAN_OR_SKIP(pass, "ProjectedSphericalRectangle::jacobianProduct", 1.0f, actual.jacobianProduct, iteration, seed, testType, 5e-2, 5e-2); + + constexpr float boundsEps = 1e-5f; + if (actual.surfaceOffset.x < -boundsEps || actual.surfaceOffset.x > actual.extents.x + boundsEps || + actual.surfaceOffset.y < -boundsEps || actual.surfaceOffset.y > actual.extents.y + boundsEps) { pass = false; - printTestFail("ProjectedSphericalRectangle::generateSurfaceOffset (inside rect bounds)", actual.extents, actual.surfaceOffset, iteration, seed, testType, 0.0, 0.0); + printTestFail("ProjectedSphericalRectangle::generateSurfaceOffset (inside rect bounds)", actual.extents, actual.surfaceOffset, iteration, seed, testType, 0.0, boundsEps); } // generate must be unit length @@ -90,7 +85,7 @@ class CProjectedSphericalRectangleTester final : public ITester createProjectedRectSampler( +inline nbl::hlsl::sampling::ProjectedSphericalRectangle createProjectedRectSampler( std::mt19937& rng, nbl::hlsl::shapes::CompressedSphericalRectangle& compressed, nbl::hlsl::float32_t3& observer, @@ -121,15 +116,16 @@ inline nbl::hlsl::sampling::ProjectedSphericalRectangle cr outNormal = generateRandomUnitVector(rng); } while (!anyRectCornerAboveHorizon(shape, observer, outNormal)); - return sampling::ProjectedSphericalRectangle::create(shape, observer, outNormal, false); + return sampling::ProjectedSphericalRectangle::create(shape, observer, outNormal, false); } struct ProjectedSphericalRectanglePropertyConfig { - using sampler_type = nbl::hlsl::sampling::ProjectedSphericalRectangle; + // UsePdfAsWeight=false so receiverNormal and projSolidAngle are populated for logSamplerInfo. + using sampler_type = nbl::hlsl::sampling::ProjectedSphericalRectangle; static constexpr uint32_t numConfigurations = 200; - static constexpr uint32_t samplesPerConfig = 20000; + static constexpr uint32_t samplesPerConfig = 50000; static constexpr bool hasMCNormalization = true; static constexpr bool hasGridIntegration = false; static constexpr float64_t mcNormalizationRelTol = 0.08; @@ -155,23 +151,20 @@ struct ProjectedSphericalRectanglePropertyConfig static void logSamplerInfo(nbl::system::ILogger* logger, const sampler_type& s) { using nbl::system::to_string; - logger->log(" r0=%s extents=%s solidAngle=%s rcpSolidAngle=%s rcpProjSolidAngle=%s", + logger->log(" r0=%s extents=%s solidAngle=%s projSolidAngle=%s receiverNormal=%s", nbl::system::ILogger::ELL_ERROR, to_string(s.sphrect.r0).c_str(), to_string(s.sphrect.extents).c_str(), to_string(s.sphrect.solidAngle).c_str(), - to_string(s.rcpSolidAngle).c_str(), - to_string(s.rcpProjSolidAngle).c_str()); - logger->log(" localReceiverNormal=%s receiverWasBSDF=%u", - nbl::system::ILogger::ELL_ERROR, - to_string(s.localReceiverNormal).c_str(), - static_cast(s.receiverWasBSDF)); + to_string(s.projSolidAngle).c_str(), + to_string(s.receiverNormal).c_str()); } }; struct ProjectedSphericalRectangleGrazingConfig { - using sampler_type = nbl::hlsl::sampling::ProjectedSphericalRectangle; + // UsePdfAsWeight=false so receiverNormal and projSolidAngle are populated for logSamplerInfo. + using sampler_type = nbl::hlsl::sampling::ProjectedSphericalRectangle; static constexpr uint32_t numConfigurations = 200; static constexpr uint32_t samplesPerConfig = 20000; @@ -202,17 +195,13 @@ struct ProjectedSphericalRectangleGrazingConfig static void logSamplerInfo(nbl::system::ILogger* logger, const sampler_type& s) { using nbl::system::to_string; - logger->log(" r0=%s extents=%s solidAngle=%s rcpSolidAngle=%s rcpProjSolidAngle=%s", + logger->log(" r0=%s extents=%s solidAngle=%s projSolidAngle=%s receiverNormal=%s", nbl::system::ILogger::ELL_ERROR, to_string(s.sphrect.r0).c_str(), to_string(s.sphrect.extents).c_str(), to_string(s.sphrect.solidAngle).c_str(), - to_string(s.rcpSolidAngle).c_str(), - to_string(s.rcpProjSolidAngle).c_str()); - logger->log(" localReceiverNormal=%s receiverWasBSDF=%u", - nbl::system::ILogger::ELL_ERROR, - to_string(s.localReceiverNormal).c_str(), - static_cast(s.receiverWasBSDF)); + to_string(s.projSolidAngle).c_str(), + to_string(s.receiverNormal).c_str()); } }; diff --git a/37_HLSLSamplingTests/tests/CProjectedSphericalTriangleTester.h b/37_HLSLSamplingTests/tests/CProjectedSphericalTriangleTester.h index 31f85ba02..611fa1f3c 100644 --- a/37_HLSLSamplingTests/tests/CProjectedSphericalTriangleTester.h +++ b/37_HLSLSamplingTests/tests/CProjectedSphericalTriangleTester.h @@ -14,7 +14,7 @@ class CProjectedSphericalTriangleTester final : public ITester; + // UsePdfAsWeight=false so receiverNormal is populated for logSamplerInfo. + using sampler_type = nbl::hlsl::sampling::ProjectedSphericalTriangle; static constexpr uint32_t numConfigurations = 200; static constexpr uint32_t samplesPerConfig = 20000; @@ -117,18 +120,19 @@ struct ProjectedSphericalTrianglePropertyConfig // E[1/pdf] = solidAngle * E[1/bilinearPdf] = solidAngle * 1.0 = solidAngle static float64_t expectedCodomainMeasure(const sampler_type& s) { - return 1.0 / static_cast(s.sphtri.base.rcpSolidAngle); + return 1.0 / static_cast(s.sphtri.rcpSolidAngle); } static void logSamplerInfo(nbl::system::ILogger* logger, const sampler_type& s) { - logTriangleInfo(logger, s.sphtri.base.tri_vertices[0], s.sphtri.base.tri_vertices[1], s.sphtri.vertexC, s.receiverNormal); + logTriangleInfo(logger, s.sphtri.tri_vertices[0], s.sphtri.tri_vertices[1], s.sphtri.APlusC - s.sphtri.tri_vertices[0], s.receiverNormal); } }; struct ProjectedSphericalTriangleGrazingConfig { - using sampler_type = nbl::hlsl::sampling::ProjectedSphericalTriangle; + // UsePdfAsWeight=false so receiverNormal is populated for logSamplerInfo. + using sampler_type = nbl::hlsl::sampling::ProjectedSphericalTriangle; static constexpr uint32_t numConfigurations = 200; static constexpr uint32_t samplesPerConfig = 20000; @@ -169,12 +173,12 @@ struct ProjectedSphericalTriangleGrazingConfig static float64_t expectedCodomainMeasure(const sampler_type& s) { - return 1.0 / static_cast(s.sphtri.base.rcpSolidAngle); + return 1.0 / static_cast(s.sphtri.rcpSolidAngle); } static void logSamplerInfo(nbl::system::ILogger* logger, const sampler_type& s) { - logTriangleInfo(logger, s.sphtri.base.tri_vertices[0], s.sphtri.base.tri_vertices[1], s.sphtri.vertexC, s.receiverNormal); + logTriangleInfo(logger, s.sphtri.tri_vertices[0], s.sphtri.tri_vertices[1], s.sphtri.APlusC - s.sphtri.tri_vertices[0], s.receiverNormal); } }; diff --git a/37_HLSLSamplingTests/tests/CSphericalRectangleTester.h b/37_HLSLSamplingTests/tests/CSphericalRectangleTester.h index 2a6030b78..7aabc48ec 100644 --- a/37_HLSLSamplingTests/tests/CSphericalRectangleTester.h +++ b/37_HLSLSamplingTests/tests/CSphericalRectangleTester.h @@ -15,22 +15,22 @@ class CSphericalRectangleTester final : public ITester sizeDist(0.5f, 3.0f); std::uniform_real_distribution uDist(0.0f, 1.0f); + nbl::hlsl::shapes::CompressedSphericalRectangle compressed; + nbl::hlsl::float32_t3 observer; + generateRandomRectangle(getRandomEngine(), compressed, observer); + SphericalRectangleInputValues input; - // Observer at origin, rect placed in front (negative Z) so the solid angle is valid. - input.observer = nbl::hlsl::float32_t3(0.0f, 0.0f, 0.0f); - const float width = sizeDist(getRandomEngine()); - const float height = sizeDist(getRandomEngine()); - input.rectOrigin = nbl::hlsl::float32_t3(0.0f, 0.0f, -2.0f); - input.right = nbl::hlsl::float32_t3(width, 0.0f, 0.0f); - input.up = nbl::hlsl::float32_t3(0.0f, height, 0.0f); + input.observer = observer; + input.rectOrigin = compressed.origin; + input.right = compressed.right; + input.up = compressed.up; input.u = nbl::hlsl::float32_t2(uDist(getRandomEngine()), uDist(getRandomEngine())); m_inputs.push_back(input); return input; @@ -48,16 +48,25 @@ class CSphericalRectangleTester final : public ITester world == generate", actual.generated, actual.normalizedLocalToWorld, iteration, seed, testType, 5e-5, 5e-3); + // computeHitT(generated) must equal hitDist returned by generateNormalizedLocal + pass &= verifyTestValue("SphericalRectangle::computeHitT == hitDist", actual.computedHitT, actual.hitDist, iteration, seed, testType, 5e-4, 2e-2); + // generateUnnormalized direction must be parallel to generate() (cross product near zero) + { + const nbl::hlsl::float32_t3 c = nbl::hlsl::cross(actual.unnormalized, actual.generated); + pass &= verifyTestValue("SphericalRectangle::generateUnnormalized parallel to generate", c, nbl::hlsl::float32_t3(0.0f, 0.0f, 0.0f), iteration, seed, testType, 1e-3, 5e-2); + } + // |generateUnnormalized| must equal hitDist (distance to hitpoint along the unit ray) + { + const float ulen = nbl::hlsl::length(actual.unnormalized); + pass &= verifyTestValue("SphericalRectangle::|generateUnnormalized| == hitDist", ulen, actual.hitDist, iteration, seed, testType, 5e-4, 2e-2); + } + if (!pass && iteration < m_inputs.size()) logFailedInput(m_logger.get(), m_inputs[iteration]); diff --git a/37_HLSLSamplingTests/tests/CSphericalTriangleTester.h b/37_HLSLSamplingTests/tests/CSphericalTriangleTester.h index fd8a0f63e..68dd2310b 100644 --- a/37_HLSLSamplingTests/tests/CSphericalTriangleTester.h +++ b/37_HLSLSamplingTests/tests/CSphericalTriangleTester.h @@ -14,7 +14,7 @@ class CSphericalTriangleTester final : public ITester; + using sampler_type = nbl::hlsl::sampling::SphericalTriangle; static constexpr uint32_t numConfigurations = 500; static constexpr uint32_t samplesPerConfig = 20000; @@ -121,7 +124,7 @@ struct SphericalTrianglePropertyConfig static void logSamplerInfo(nbl::system::ILogger* logger, const sampler_type& s) { - logTriangleInfo(logger, s.base.tri_vertices[0], s.base.tri_vertices[1], s.vertexC); + logTriangleInfo(logger, s.tri_vertices[0], s.tri_vertices[1], s.APlusC - s.tri_vertices[0]); } }; @@ -130,7 +133,7 @@ struct SphericalTrianglePropertyConfig // These stress the C_s great-circle intersection and v-recovery in generateInverse. struct SphericalTriangleStressConfig { - using sampler_type = nbl::hlsl::sampling::SphericalTriangle; + using sampler_type = nbl::hlsl::sampling::SphericalTriangle; static constexpr uint32_t numConfigurations = 500; static constexpr uint32_t samplesPerConfig = 20000; @@ -218,7 +221,7 @@ struct SphericalTriangleStressConfig static void logSamplerInfo(nbl::system::ILogger* logger, const sampler_type& s) { - logTriangleInfo(logger, s.base.tri_vertices[0], s.base.tri_vertices[1], s.vertexC); + logTriangleInfo(logger, s.tri_vertices[0], s.tri_vertices[1], s.APlusC - s.tri_vertices[0]); } }; diff --git a/37_HLSLSamplingTests/tests/CUniformHemisphereTester.h b/37_HLSLSamplingTests/tests/CUniformHemisphereTester.h index 29994511f..4f80ecbaf 100644 --- a/37_HLSLSamplingTests/tests/CUniformHemisphereTester.h +++ b/37_HLSLSamplingTests/tests/CUniformHemisphereTester.h @@ -12,7 +12,7 @@ class CUniformHemisphereTester final : public ITester #include +#include + // ============================================================================ // Declarative field verification helpers // @@ -34,30 +36,126 @@ struct PdfCheck // Verify expected.*field vs actual.*field for each FieldCheck. // Must be called from within a method that has access to verifyTestValue. -#define VERIFY_FIELDS(pass, expected, actual, iteration, seed, testType, ...) \ - do \ - { \ - auto _checks = std::make_tuple(__VA_ARGS__); \ - std::apply([&](const auto&... c) { ((pass &= verifyTestValue(c.name, (expected).*c.field, (actual).*c.field, \ - iteration, seed, testType, c.relTol, c.absTol)), \ +#define VERIFY_FIELDS(pass, expected, actual, iteration, seed, testType, ...) \ + do \ + { \ + auto _checks = std::make_tuple(__VA_ARGS__); \ + std::apply([&](const auto&... c) { ((pass &= verifyTestValue(c.name, (expected).*c.field, (actual).*c.field, \ + iteration, seed, testType, c.relTol, c.absTol)), \ ...); }, _checks); \ } while (0) +// ============================================================================ +// Jacobian skip tracking +// +// The device-side sampler writes a reason-encoded skip sentinel (see +// jacobian_test.hlsl) instead of a jacobianProduct value when it cannot test +// a sample honestly. The host recognizes the sentinel, bins it by reason, +// and NEVER counts it as a pass. After all tests run, logJacobianSkipCounts() +// reports per-reason counts so nothing silently inflates pass rates. +// ============================================================================ + +namespace detail +{ +struct JacobianStats +{ + uint64_t total = 0; // total VERIFY_JACOBIAN_OR_SKIP invocations (= samples evaluated) + uint64_t skipUDomain = 0; // JACOBIAN_SKIP_U_DOMAIN = -1.0f + uint64_t skipCrease = 0; // JACOBIAN_SKIP_CREASE = -2.0f + uint64_t skipHemiBoundary = 0; // JACOBIAN_SKIP_HEMI_BOUNDARY = -3.0f + uint64_t skipBwdPdfRange = 0; // JACOBIAN_SKIP_BWD_PDF_RANGE = -4.0f + uint64_t skipCodomainSingularity = 0; // JACOBIAN_SKIP_CODOMAIN_SINGULARITY = -5.0f +}; + +inline nbl::core::map& jacobianStats() +{ + static nbl::core::map s; + return s; +} +} // namespace detail + +inline void logJacobianSkipCounts(nbl::system::ILogger* logger) +{ + auto& stats = detail::jacobianStats(); + if (stats.empty()) + return; + logger->log("Jacobian skip summary (skipped samples are NOT counted as passes):", nbl::system::ILogger::ELL_INFO); + for (const auto& [name, s] : stats) + { + const uint64_t skipped = s.skipUDomain + s.skipCrease + s.skipHemiBoundary + s.skipBwdPdfRange + s.skipCodomainSingularity; + if (skipped == 0) + continue; + const double percentage = s.total ? (100.0 * double(skipped) / double(s.total)) : 0.0; + logger->log(" [JacobianSkip] %s: %llu / %llu skipped (%.2f%%) -- u-domain=%llu, crease=%llu, hemi-boundary=%llu, bwd-pdf-range=%llu, codomain-singularity=%llu", + nbl::system::ILogger::ELL_WARNING, + name.c_str(), + skipped, + s.total, + percentage, + s.skipUDomain, + s.skipCrease, + s.skipHemiBoundary, + s.skipBwdPdfRange, + s.skipCodomainSingularity); + } +} + +// Verify a jacobianProduct value OR bin it by reason if it is a skip sentinel (< 0). +// Skipped samples are counted by reason and NEVER counted as a pass. +// Must be called from a method that has access to verifyTestValue. +#define VERIFY_JACOBIAN_OR_SKIP(pass, name, expected, actual, iteration, seed, testType, relTol, absTol) \ + do \ + { \ + auto& _jstats = detail::jacobianStats()[(name)]; \ + ++_jstats.total; \ + const float _jval = (actual); \ + if (_jval < 0.0f) \ + { \ + /* Sentinel values are integers at -1..-5, so round-to-nearest on _jval picks the bin. */ \ + const int _bin = static_cast(-_jval + 0.5f); \ + switch (_bin) \ + { \ + case 1: \ + ++_jstats.skipUDomain; \ + break; \ + case 2: \ + ++_jstats.skipCrease; \ + break; \ + case 3: \ + ++_jstats.skipHemiBoundary; \ + break; \ + case 4: \ + ++_jstats.skipBwdPdfRange; \ + break; \ + case 5: \ + ++_jstats.skipCodomainSingularity; \ + break; \ + default: \ + ++_jstats.skipUDomain; \ + break; /* fall-through bucket */ \ + } \ + } \ + else \ + { \ + pass &= verifyTestValue((name), (expected), _jval, (iteration), (seed), (testType), (relTol), (absTol)); \ + } \ + } while (0) + // Check that each PDF field is positive and finite. // Must be called from within a method that has access to printTestFail. -#define VERIFY_PDFS_POSITIVE(pass, actual, iteration, seed, testType, ...) \ - do \ - { \ - auto _pdfChecks = std::make_tuple(__VA_ARGS__); \ - std::apply([&](const auto&... c) { (([&] { \ +#define VERIFY_PDFS_POSITIVE(pass, actual, iteration, seed, testType, ...) \ + do \ + { \ + auto _pdfChecks = std::make_tuple(__VA_ARGS__); \ + std::apply([&](const auto&... c) { (([&] { \ if (!((actual).*c.field > 0.0f) || !std::isfinite((actual).*c.field)) \ - { \ - pass = false; \ - printTestFail(std::string(c.name) + " (positive & finite)", \ - 1.0f, (actual).*c.field, iteration, seed, testType, 0.0, 0.0); \ - } \ - }()), \ - ...); }, _pdfChecks); \ + { \ + pass = false; \ + printTestFail(std::string(c.name) + " (positive & finite)", \ + 1.0f, (actual).*c.field, iteration, seed, testType, 0.0, 0.0); \ + } \ + }()), \ + ...); }, _pdfChecks); \ } while (0) // ============================================================================ @@ -139,7 +237,7 @@ inline float64_t gridIntegratePdf1D(const auto& sampler, uint32_t N = 100000) // 2D grid integration of backwardPdf over [0,1]^2 inline float64_t gridIntegratePdf2D(const auto& sampler, uint32_t N = 1000) { - float64_t sum = 0.0; + float64_t sum = 0.0; const float64_t cellArea = 1.0 / static_cast(N * N); for (uint32_t iy = 0; iy < N; iy++) { @@ -190,17 +288,15 @@ inline void buildTangentFrame(nbl::hlsl::float32_t3 dir, nbl::hlsl::float32_t3& // Generate a small equilateral triangle on the unit sphere around baseDir with given half-angle. // Also generates a random normal with decent projection onto the triangle. -inline void generateSmallTriangle(std::mt19937& rng, float halfAngle, - nbl::hlsl::float32_t3& v0, nbl::hlsl::float32_t3& v1, nbl::hlsl::float32_t3& v2, - nbl::hlsl::float32_t3& baseDir, nbl::hlsl::float32_t3& normal) +inline void generateSmallTriangle(std::mt19937& rng, float halfAngle, nbl::hlsl::float32_t3& v0, nbl::hlsl::float32_t3& v1, nbl::hlsl::float32_t3& v2, nbl::hlsl::float32_t3& baseDir, nbl::hlsl::float32_t3& normal) { using namespace nbl::hlsl; baseDir = generateRandomUnitVector(rng); float32_t3 t1, t2; buildTangentFrame(baseDir, t1, t2); - v0 = normalize(baseDir + t1 * halfAngle); - v1 = normalize(baseDir - t1 * (halfAngle * 0.5f) + t2 * (halfAngle * 0.866f)); - v2 = normalize(baseDir - t1 * (halfAngle * 0.5f) - t2 * (halfAngle * 0.866f)); + v0 = normalize(baseDir + t1 * halfAngle); + v1 = normalize(baseDir - t1 * (halfAngle * 0.5f) + t2 * (halfAngle * 0.866f)); + v2 = normalize(baseDir - t1 * (halfAngle * 0.5f) - t2 * (halfAngle * 0.866f)); normal = generateRandomUnitVector(rng); if (dot(normal, baseDir) < 0.1f) normal = normalize(normal + baseDir * 2.0f); @@ -221,10 +317,10 @@ inline void generateStressTriangleVertices(std::mt19937& rng, nbl::hlsl::float32 float32_t3 t1, t2; buildTangentFrame(base, t1, t2); float spread = 0.15f + angleDist(rng) * 0.2f; - v0 = normalize(base + t1 * spread); - v1 = normalize(base - t1 * spread); - float far_ = 0.8f + angleDist(rng) * 0.8f; - v2 = normalize(base * std::cos(far_) + t2 * std::sin(far_)); + v0 = normalize(base + t1 * spread); + v1 = normalize(base - t1 * spread); + float far_ = 0.8f + angleDist(rng) * 0.8f; + v2 = normalize(base * std::cos(far_) + t2 * std::sin(far_)); break; } case 1: // Nearly coplanar @@ -233,12 +329,12 @@ inline void generateStressTriangleVertices(std::mt19937& rng, nbl::hlsl::float32 float32_t3 t1, t2; buildTangentFrame(pole, t1, t2); float offset = 0.05f + angleDist(rng) * 0.1f; - float a1 = angleDist(rng) * 6.2832f; - float a2 = a1 + 0.8f + angleDist(rng); - float a3 = a2 + 0.8f + angleDist(rng); - v0 = normalize(t1 * std::cos(a1) + t2 * std::sin(a1) + pole * offset); - v1 = normalize(t1 * std::cos(a2) + t2 * std::sin(a2) - pole * offset * 0.5f); - v2 = normalize(t1 * std::cos(a3) + t2 * std::sin(a3) + pole * offset * 0.3f); + float a1 = angleDist(rng) * 6.2832f; + float a2 = a1 + 0.8f + angleDist(rng); + float a3 = a2 + 0.8f + angleDist(rng); + v0 = normalize(t1 * std::cos(a1) + t2 * std::sin(a1) + pole * offset); + v1 = normalize(t1 * std::cos(a2) + t2 * std::sin(a2) - pole * offset * 0.5f); + v2 = normalize(t1 * std::cos(a3) + t2 * std::sin(a3) + pole * offset * 0.3f); break; } default: // One short edge @@ -247,9 +343,9 @@ inline void generateStressTriangleVertices(std::mt19937& rng, nbl::hlsl::float32 float32_t3 t1, t2; buildTangentFrame(base, t1, t2); float shortAngle = 0.32f + angleDist(rng) * 0.1f; - v0 = normalize(base + t1 * shortAngle * 0.5f); - v1 = normalize(base - t1 * shortAngle * 0.5f); - v2 = normalize(t2 + base * (0.3f + angleDist(rng) * 0.5f)); + v0 = normalize(base + t1 * shortAngle * 0.5f); + v1 = normalize(base - t1 * shortAngle * 0.5f); + v2 = normalize(t2 + base * (0.3f + angleDist(rng) * 0.5f)); break; } } @@ -262,65 +358,114 @@ inline void generateStressTriangleVertices(std::mt19937& rng, nbl::hlsl::float32 inline void makeEquilateralTriangle(float64_t theta, nbl::hlsl::float32_t3 verts[3]) { using namespace nbl::hlsl; - const float32_t st = static_cast(std::sin(theta)); - const float32_t ct = static_cast(std::cos(theta)); + const float32_t st = static_cast(std::sin(theta)); + const float32_t ct = static_cast(std::cos(theta)); constexpr float64_t twoPiOver3 = 2.0 * numbers::pi / 3.0; - verts[0] = float32_t3(st, 0.0f, ct); - verts[1] = float32_t3(static_cast(st * std::cos(twoPiOver3)), + verts[0] = float32_t3(st, 0.0f, ct); + verts[1] = float32_t3(static_cast(st * std::cos(twoPiOver3)), static_cast(st * std::sin(twoPiOver3)), ct); - verts[2] = float32_t3(static_cast(st * std::cos(2.0 * twoPiOver3)), + verts[2] = float32_t3(static_cast(st * std::cos(2.0 * twoPiOver3)), static_cast(st * std::sin(2.0 * twoPiOver3)), ct); } -// Monte Carlo estimate of projected solid angle: E[abs(dot(L, normal))] * solidAngle. -// Uses abs() to match the BSDF projected solid angle formula (which uses abs so that -// triangles straddling the horizon contribute positively from both hemispheres). -// Samples L uniformly from the spherical triangle. -inline float64_t mcEstimatePSA(const nbl::hlsl::shapes::SphericalTriangle& shape, nbl::hlsl::float32_t3 normal, uint32_t N, std::mt19937& rng) +// Grid estimate of projected solid angle: mean of abs(dot(L, normal)) over a regular +// [0,1]^2 grid, times solidAngle. Uses abs() to match the BSDF projected solid angle +// formula (triangles/rects straddling the horizon contribute from both hemispheres). +// `N` is the total number of samples; the grid side is ceil(sqrt(N)). Grid integration +// is deterministic and has much lower variance than MC at the same sample count, +// so it's a tighter ground truth for PSA-vs-formula comparisons. +inline float64_t gridEstimatePSA(const nbl::hlsl::shapes::SphericalTriangle& shape, nbl::hlsl::float32_t3 normal, uint32_t N) { using namespace nbl::hlsl; - auto sampler = sampling::SphericalTriangle::create(shape); - std::uniform_real_distribution uDist(0.0f, 1.0f); - float64_t sum = 0.0; - for (uint32_t i = 0; i < N; i++) + auto sampler = sampling::SphericalTriangle::create(shape); + const uint32_t gridSide = static_cast(std::ceil(std::sqrt(static_cast(N)))); + const float invSide = 1.0f / static_cast(gridSide); + float64_t sum = 0.0; + for (uint32_t iy = 0; iy < gridSide; iy++) + { + const float uy = (static_cast(iy) + 0.5f) * invSide; + for (uint32_t ix = 0; ix < gridSide; ix++) + { + const float ux = (static_cast(ix) + 0.5f) * invSide; + typename sampling::SphericalTriangle::cache_type cache; + const float32_t3 L = sampler.generate(float32_t2(ux, uy), cache); + sum += static_cast(hlsl::abs(dot(normal, L))); + } + } + return sum / static_cast(gridSide * gridSide) * static_cast(shape.solid_angle); +} + +// Sampler-independent PSA reference for rectangles. Integrates the projected-solid-angle integral +// PSA = integral over rect surface of |cos(theta_receiver)| * |cos(theta_rect)| / d^2 dA +// on a uniform surface grid in (s, t) in [0, extents.x] x [0, extents.y]. No sampler involved, +// so disagreement with a sampler-derived PSA isolates the sampler / formula. +inline float64_t surfaceGridEstimatePSA( + const nbl::hlsl::shapes::SphericalRectangle& shape, + const nbl::hlsl::float32_t3& observer, + const nbl::hlsl::float32_t3& normal, + uint32_t N) +{ + using namespace nbl::hlsl; + const float32_t3 rdir = shape.basis[0]; + const float32_t3 udir = shape.basis[1]; + const float32_t3 rectNormal = shape.basis[2]; + const float32_t width = shape.extents.x; + const float32_t height = shape.extents.y; + const uint32_t gridSide = static_cast(std::ceil(std::sqrt(static_cast(N)))); + const float64_t cellArea = static_cast(width) * static_cast(height) / static_cast(gridSide * gridSide); + float64_t sum = 0.0; + for (uint32_t iy = 0; iy < gridSide; iy++) { - float32_t2 u(uDist(rng), uDist(rng)); - typename sampling::SphericalTriangle::cache_type cache; - float32_t3 L = sampler.generate(u, cache); - sum += static_cast(hlsl::abs(dot(normal, L))); + const float32_t t = (static_cast(iy) + 0.5f) * height / static_cast(gridSide); + for (uint32_t ix = 0; ix < gridSide; ix++) + { + const float32_t s = (static_cast(ix) + 0.5f) * width / static_cast(gridSide); + const float32_t3 worldPt = shape.origin + rdir * s + udir * t; + const float32_t3 toSurf = worldPt - observer; + const float64_t d2 = static_cast(dot(toSurf, toSurf)); + const float64_t d = std::sqrt(d2); + const float32_t3 L = toSurf * static_cast(1.0 / d); + const float64_t cosRx = static_cast(hlsl::abs(dot(normal, L))); + const float64_t cosRt = static_cast(hlsl::abs(dot(rectNormal, L))); + sum += cosRx * cosRt / d2; + } } - return sum / static_cast(N) * static_cast(shape.solid_angle); + return sum * cellArea; } -// Monte Carlo estimate of projected solid angle for a rectangle: E[abs(dot(L, normal))] * solidAngle. -// Uses abs() to match the BSDF projected solid angle formula. -// Samples uniformly from the spherical rectangle, reconstructs world-space direction. -inline float64_t mcEstimatePSA( +// Grid estimate of projected solid angle for a rectangle: mean of abs(dot(L, normal)) +// over a regular [0,1]^2 grid, times solidAngle. See the triangle overload above. +inline float64_t gridEstimatePSA( const nbl::hlsl::shapes::SphericalRectangle& shape, const nbl::hlsl::float32_t3& observer, const nbl::hlsl::float32_t3& normal, - uint32_t N, std::mt19937& rng) + uint32_t N) { using namespace nbl::hlsl; auto sampler = sampling::SphericalRectangle::create(shape, observer); if (sampler.solidAngle <= 0.0f || !std::isfinite(sampler.solidAngle)) return 0.0; - std::uniform_real_distribution uDist(0.0f, 1.0f); - float64_t sum = 0.0; - for (uint32_t i = 0; i < N; i++) + const uint32_t gridSide = static_cast(std::ceil(std::sqrt(static_cast(N)))); + const float invSide = 1.0f / static_cast(gridSide); + float64_t sum = 0.0; + for (uint32_t iy = 0; iy < gridSide; iy++) { - float32_t2 u(uDist(rng), uDist(rng)); - typename sampling::SphericalRectangle::cache_type cache; - float32_t2 gen = sampler.generateSurfaceOffset(u, cache); - // Reconstruct world-space direction from rectangle offset - float32_t3 worldPt = shape.origin - + shape.basis[0] * gen.x - + shape.basis[1] * gen.y; - float32_t3 L = normalize(worldPt - observer); - sum += static_cast(hlsl::abs(dot(normal, L))); + const float uy = (static_cast(iy) + 0.5f) * invSide; + for (uint32_t ix = 0; ix < gridSide; ix++) + { + const float ux = (static_cast(ix) + 0.5f) * invSide; + typename sampling::SphericalRectangle::cache_type cache; + // `generateLocalBasisXY` returns absolute (xu, yv) on the rectangle surface; subtract r0.xy + // to get the offset-from-r0 that the world-space reconstruction below expects. + const float32_t2 absXY = sampler.generateLocalBasisXY(float32_t2(ux, uy), cache); + const float32_t2 gen = absXY - float32_t2(sampler.r0.x, sampler.r0.y); + const float32_t3 worldPt = shape.origin + shape.basis[0] * gen.x + shape.basis[1] * gen.y; + const float32_t3 L = normalize(worldPt - observer); + sum += static_cast(hlsl::abs(dot(normal, L))); + } } - return sum / static_cast(N) * static_cast(sampler.solidAngle); + return sum / static_cast(gridSide * gridSide) * static_cast(sampler.solidAngle); } // Bundles seed + rng + failCount for randomized property tests. @@ -332,7 +477,7 @@ struct SeededTestContext std::mt19937 rng; uint32_t failCount = 0; - SeededTestContext() : seed(std::random_device {}()), rng(seed) {} + SeededTestContext(std::optional seedOverride = {}) : seed(seedOverride.value_or(std::random_device {}())), rng(seed) {} // Log "reproduce with seed" if failCount > 0, return failCount == 0 bool finalize(nbl::system::ILogger* logger, const char* tag) const @@ -357,14 +502,18 @@ struct SeededTestContext } }; -// Generic PSA vs MC comparison. -// ConfigGen: void(std::mt19937& rng, uint32_t index, float64_t& formulaPSA, float64_t& mcPSA, InfoLogger& info) -// Must set formulaPSA and mcPSA for config `index`, or set both to 0 to skip. +// Generic PSA vs grid-integration comparison. +// ConfigGen: void(std::mt19937& rng, uint32_t index, float64_t& formulaPSA, float64_t& gridPSA, InfoLogger& info) +// Must set formulaPSA and gridPSA for config `index`, or set both to 0 to skip. // `info` is a callable: void(nbl::system::ILogger*, nbl::system::ILogger::E_LOG_LEVEL) that logs // sampler/shape details for the current config. Called on mismatch. -// When diagnostic=true, failures log at ELL_WARNING instead of ELL_ERROR (non-hard-fail). +// Two-tier tolerance: +// - (relTol, absTol): soft threshold. Exceedance counts as a mismatch. With diagnostic=true +// the run still returns true (known-limitation noise); with diagnostic=false it hard-fails. +// - (hardRelTol, hardAbsTol): egregious threshold. Always hard-fails regardless of diagnostic, +// so a catastrophic regression can't hide inside the warning stream. template -inline bool testPSAVersusMonteCarlo( +inline bool testPSAVersusGrid( nbl::system::ILogger* logger, const char* tag, const char* label, @@ -372,49 +521,78 @@ inline bool testPSAVersusMonteCarlo( uint32_t numConfigs, float64_t relTol, float64_t absTol, + float64_t hardRelTol, + float64_t hardAbsTol, bool diagnostic = false) { - const auto failLevel = diagnostic ? nbl::system::ILogger::ELL_WARNING : nbl::system::ILogger::ELL_ERROR; + const auto softFailLevel = diagnostic ? nbl::system::ILogger::ELL_WARNING : nbl::system::ILogger::ELL_ERROR; SeededTestContext ctx; + uint32_t hardFailCount = 0; + uint32_t testedCount = 0; for (uint32_t c = 0; c < numConfigs; c++) { - float64_t formulaPSA = 0.0, mcPSA = 0.0; + float64_t formulaPSA = 0.0, gridPSA = 0.0; std::function logInfo = - [](nbl::system::ILogger*, nbl::system::ILogger::E_LOG_LEVEL) {}; - configGenerator(ctx.rng, c, formulaPSA, mcPSA, logInfo); + [](nbl::system::ILogger*, nbl::system::ILogger::E_LOG_LEVEL) { + }; + configGenerator(ctx.rng, c, formulaPSA, gridPSA, logInfo); - if (mcPSA == 0.0 && formulaPSA == 0.0) + if (gridPSA == 0.0 && formulaPSA == 0.0) continue; + testedCount++; - const float64_t absErr = std::abs(formulaPSA - mcPSA); - const float64_t relErr = (std::abs(mcPSA) > 1e-10) ? absErr / std::abs(mcPSA) : 0.0; + const float64_t absErr = std::abs(formulaPSA - gridPSA); + const float64_t relErr = (std::abs(gridPSA) > 1e-10) ? absErr / std::abs(gridPSA) : 0.0; - if (relErr > relTol && absErr > absTol) + const bool softFail = relErr > relTol && absErr > absTol; + const bool hardFail = relErr > hardRelTol && absErr > hardAbsTol; + + if (softFail) { ctx.failCount++; + if (hardFail) + hardFailCount++; if (ctx.failCount <= 5) { - logger->log(" [%s] %s mismatch: formula=%f expected(MC)=%f relErr=%e absErr=%e config %u", - failLevel, tag, label, formulaPSA, mcPSA, relErr, absErr, c); - logInfo(logger, failLevel); + const auto level = hardFail ? nbl::system::ILogger::ELL_ERROR : softFailLevel; + logger->log(" [%s] %s %s: formula=%f expected(grid)=%f relErr=%e absErr=%e config %u", + level, tag, label, hardFail ? "HARD mismatch" : "mismatch", + formulaPSA, gridPSA, relErr, absErr, c); + logInfo(logger, level); } } } + const uint32_t skippedCount = numConfigs - testedCount; + if (ctx.failCount == 0) - logger->log(" [%s] %s PASSED (%u configs, relTol=%e absTol=%e)", - nbl::system::ILogger::ELL_PERFORMANCE, tag, label, numConfigs, relTol, absTol); - else { - logger->log(" [%s] %s FAILED (%u/%u configs exceeded tolerance, relTol=%e absTol=%e)", - failLevel, tag, label, ctx.failCount, numConfigs, relTol, absTol); - if (diagnostic) - logger->log(" [%s] reproduce with seed=%u (diagnostic only, not a hard failure)", - nbl::system::ILogger::ELL_WARNING, tag, ctx.seed); + logger->log(" [%s] %s PASSED (%u tested, %u skipped of %u requested, relTol=%e absTol=%e)", + nbl::system::ILogger::ELL_PERFORMANCE, tag, label, + testedCount, skippedCount, numConfigs, relTol, absTol); + return true; } - return diagnostic ? true : ctx.finalize(logger, tag); + const bool hardFailed = hardFailCount > 0; + const auto summaryLevel = hardFailed ? nbl::system::ILogger::ELL_ERROR : softFailLevel; + if (hardFailed) + logger->log(" [%s] %s FAILED (%u/%u exceeded soft tol, %u/%u exceeded HARD tol, %u skipped of %u, hardRelTol=%e hardAbsTol=%e)", + summaryLevel, tag, label, ctx.failCount, testedCount, hardFailCount, testedCount, + skippedCount, numConfigs, hardRelTol, hardAbsTol); + else + logger->log(" [%s] %s FAILED (%u/%u configs exceeded tolerance, %u skipped of %u, relTol=%e absTol=%e)", + summaryLevel, tag, label, ctx.failCount, testedCount, skippedCount, numConfigs, relTol, absTol); + + const bool shouldHardFail = hardFailed || !diagnostic; + if (shouldHardFail) + logger->log(" [%s] reproduce with seed=%u", + nbl::system::ILogger::ELL_ERROR, tag, ctx.seed); + else + logger->log(" [%s] reproduce with seed=%u (diagnostic only, not a hard failure)", + nbl::system::ILogger::ELL_WARNING, tag, ctx.seed); + + return !shouldHardFail; } // ============================================================================ @@ -435,23 +613,21 @@ inline void generateRandomRectangle(std::mt19937& rng, float32_t3 t1, t2; buildTangentFrame(normal, t1, t2); - const float width = sizeDist(rng); + const float width = sizeDist(rng); const float height = sizeDist(rng); - const float dist = distDist(rng); + const float dist = distDist(rng); - observer = float32_t3(offsetDist(rng), offsetDist(rng), offsetDist(rng)); + observer = float32_t3(offsetDist(rng), offsetDist(rng), offsetDist(rng)); compressed.origin = observer - normal * dist + t1 * offsetDist(rng) + t2 * offsetDist(rng); - compressed.right = t1 * width; - compressed.up = t2 * height; + compressed.right = t1 * width; + compressed.up = t2 * height; } // Stress rectangles: ill-conditioned geometries that exercise edge cases. // - Extreme aspect ratio (10:1 to 20:1) // - Grazing angle (observer nearly in the rectangle plane) // - Observer near corner (most of the rectangle off to one side) -inline void generateStressRectangle(std::mt19937& rng, - nbl::hlsl::shapes::CompressedSphericalRectangle& compressed, - nbl::hlsl::float32_t3& observer) +inline void generateStressRectangle(std::mt19937& rng, nbl::hlsl::shapes::CompressedSphericalRectangle& compressed, nbl::hlsl::float32_t3& observer) { using namespace nbl::hlsl; std::uniform_real_distribution uDist(0.0f, 1.0f); @@ -464,39 +640,39 @@ inline void generateStressRectangle(std::mt19937& rng, switch (caseDist(rng)) { case 0: // Extreme aspect ratio - { - const float longSide = 3.0f + uDist(rng) * 5.0f; - const float shortSide = 0.1f + uDist(rng) * 0.2f; - const float dist = 1.5f + uDist(rng) * 2.0f; - observer = float32_t3(0.0f, 0.0f, 0.0f); - compressed.origin = -normal * dist - t1 * (longSide * 0.5f) - t2 * (shortSide * 0.5f); - compressed.right = t1 * longSide; - compressed.up = t2 * shortSide; - break; - } + { + const float longSide = 3.0f + uDist(rng) * 5.0f; + const float shortSide = 0.1f + uDist(rng) * 0.2f; + const float dist = 1.5f + uDist(rng) * 2.0f; + observer = float32_t3(0.0f, 0.0f, 0.0f); + compressed.origin = -normal * dist - t1 * (longSide * 0.5f) - t2 * (shortSide * 0.5f); + compressed.right = t1 * longSide; + compressed.up = t2 * shortSide; + break; + } case 1: // Grazing angle (observer nearly in the rectangle plane) - { - const float width = 1.0f + uDist(rng) * 2.0f; - const float height = 1.0f + uDist(rng) * 2.0f; - const float normalDist = 0.05f + uDist(rng) * 0.15f; - const float tangentOffset = 0.5f + uDist(rng) * 1.0f; - observer = float32_t3(0.0f, 0.0f, 0.0f); - compressed.origin = -normal * normalDist + t1 * tangentOffset - t2 * (height * 0.5f); - compressed.right = t1 * width; - compressed.up = t2 * height; - break; - } + { + const float width = 1.0f + uDist(rng) * 2.0f; + const float height = 1.0f + uDist(rng) * 2.0f; + const float normalDist = 0.05f + uDist(rng) * 0.15f; + const float tangentOffset = 0.5f + uDist(rng) * 1.0f; + observer = float32_t3(0.0f, 0.0f, 0.0f); + compressed.origin = -normal * normalDist + t1 * tangentOffset - t2 * (height * 0.5f); + compressed.right = t1 * width; + compressed.up = t2 * height; + break; + } default: // Observer near corner - { - const float width = 2.0f + uDist(rng) * 3.0f; - const float height = 2.0f + uDist(rng) * 3.0f; - const float dist = 0.5f + uDist(rng) * 1.0f; - observer = float32_t3(0.0f, 0.0f, 0.0f); - compressed.origin = -normal * dist - t1 * (0.05f + uDist(rng) * 0.1f) - t2 * (0.05f + uDist(rng) * 0.1f); - compressed.right = t1 * width; - compressed.up = t2 * height; - break; - } + { + const float width = 2.0f + uDist(rng) * 3.0f; + const float height = 2.0f + uDist(rng) * 3.0f; + const float dist = 0.5f + uDist(rng) * 1.0f; + observer = float32_t3(0.0f, 0.0f, 0.0f); + compressed.origin = -normal * dist - t1 * (0.05f + uDist(rng) * 0.1f) - t2 * (0.05f + uDist(rng) * 0.1f); + compressed.right = t1 * width; + compressed.up = t2 * height; + break; + } } } @@ -590,10 +766,10 @@ inline void logRectInfo( { using namespace nbl::system; using namespace nbl::hlsl; - const float width = length(compressed.right); - const float height = length(compressed.up); + const float width = length(compressed.right); + const float height = length(compressed.up); const float32_t3 normal = normalize(cross(compressed.right, compressed.up)); - const float dist = length(compressed.origin - observer); + const float dist = length(compressed.origin - observer); logger->log(" origin=%s right=%s up=%s observer=%s", ILogger::ELL_ERROR, to_string(compressed.origin).c_str(), @@ -617,14 +793,14 @@ inline bool anyRectCornerAboveHorizon( const nbl::hlsl::float32_t3& normal) { using namespace nbl::hlsl; - const float32_t3 r0 = mul(shape.basis, shape.origin - observer); + const float32_t3 r0 = mul(shape.basis, shape.origin - observer); const float32_t3 localN = mul(shape.basis, normal); - const float32_t3 v0 = normalize(r0); - const float32_t3 v1 = normalize(r0 + float32_t3(shape.extents.x, 0.0f, 0.0f)); - const float32_t3 v2 = normalize(r0 + float32_t3(shape.extents.x, shape.extents.y, 0.0f)); - const float32_t3 v3 = normalize(r0 + float32_t3(0.0f, shape.extents.y, 0.0f)); + const float32_t3 v0 = normalize(r0); + const float32_t3 v1 = normalize(r0 + float32_t3(shape.extents.x, 0.0f, 0.0f)); + const float32_t3 v2 = normalize(r0 + float32_t3(shape.extents.x, shape.extents.y, 0.0f)); + const float32_t3 v3 = normalize(r0 + float32_t3(0.0f, shape.extents.y, 0.0f)); return dot(localN, v0) > 0.0f || dot(localN, v1) > 0.0f || - dot(localN, v2) > 0.0f || dot(localN, v3) > 0.0f; + dot(localN, v2) > 0.0f || dot(localN, v3) > 0.0f; } // True if all rectangle corners have positive NdotL with the given normal. @@ -635,14 +811,14 @@ inline bool allRectCornersAboveHorizon( const nbl::hlsl::float32_t3& normal) { using namespace nbl::hlsl; - const float32_t3 r0 = mul(shape.basis, shape.origin - observer); + const float32_t3 r0 = mul(shape.basis, shape.origin - observer); const float32_t3 localN = mul(shape.basis, normal); - const float32_t3 v0 = normalize(r0); - const float32_t3 v1 = normalize(r0 + float32_t3(shape.extents.x, 0.0f, 0.0f)); - const float32_t3 v2 = normalize(r0 + float32_t3(shape.extents.x, shape.extents.y, 0.0f)); - const float32_t3 v3 = normalize(r0 + float32_t3(0.0f, shape.extents.y, 0.0f)); + const float32_t3 v0 = normalize(r0); + const float32_t3 v1 = normalize(r0 + float32_t3(shape.extents.x, 0.0f, 0.0f)); + const float32_t3 v2 = normalize(r0 + float32_t3(shape.extents.x, shape.extents.y, 0.0f)); + const float32_t3 v3 = normalize(r0 + float32_t3(0.0f, shape.extents.y, 0.0f)); return dot(localN, v0) > 0.0f && dot(localN, v1) > 0.0f && - dot(localN, v2) > 0.0f && dot(localN, v3) > 0.0f; + dot(localN, v2) > 0.0f && dot(localN, v3) > 0.0f; } #endif diff --git a/37_HLSLSamplingTests/tests/property/CSamplerPropertyTester.h b/37_HLSLSamplingTests/tests/property/CSamplerPropertyTester.h index cb28b63fc..b20ba88f9 100644 --- a/37_HLSLSamplingTests/tests/property/CSamplerPropertyTester.h +++ b/37_HLSLSamplingTests/tests/property/CSamplerPropertyTester.h @@ -78,7 +78,9 @@ class CSamplerPropertyTester } public: - CSamplerPropertyTester(system::ILogger* logger) : m_logger(logger) {} + CSamplerPropertyTester(system::ILogger* logger, std::optional seedOverride = {}) : m_logger(logger), m_seedOverride(seedOverride) {} + + std::optional failureSeed() const { return m_failureSeed; } bool run() { @@ -96,7 +98,7 @@ class CSamplerPropertyTester // If the PDF normalization is wrong by factor k, this will be off by 1/k. bool testMonteCarloPdfNormalization() { - SeededTestContext ctx; + SeededTestContext ctx(m_seedOverride); uint32_t evaluatedConfigs = 0; for (uint32_t c = 0; c < Config::numConfigurations; c++) @@ -159,7 +161,10 @@ class CSamplerPropertyTester m_logger->log(" [%s] MC normalization FAILED (%u/%u evaluated configs failed, %u/%u configs evaluated, %u samples/config, relTol=%e)", system::ILogger::ELL_ERROR, Config::name(), ctx.failCount, evaluatedConfigs, evaluatedConfigs, Config::numConfigurations, Config::samplesPerConfig, Config::mcNormalizationRelTol); - return ctx.finalize(m_logger, Config::name()); + const bool passed = ctx.finalize(m_logger, Config::name()); + if (!passed) + m_failureSeed = ctx.seed; + return passed; } // Test 4: Grid integration of backwardPdf over [0,1]^d codomain @@ -167,7 +172,7 @@ class CSamplerPropertyTester // integral of backwardPdf over codomain should equal 1.0. bool testGridPdfNormalization() { - SeededTestContext ctx; + SeededTestContext ctx(m_seedOverride); for (uint32_t c = 0; c < Config::numConfigurations; c++) { @@ -191,10 +196,15 @@ class CSamplerPropertyTester m_logger->log(" [%s] grid PDF normalization FAILED (%u/%u configs exceeded absTol=%e)", system::ILogger::ELL_ERROR, Config::name(), ctx.failCount, Config::numConfigurations, Config::gridNormalizationAbsTol); - return ctx.finalize(m_logger, Config::name()); + const bool passed = ctx.finalize(m_logger, Config::name()); + if (!passed) + m_failureSeed = ctx.seed; + return passed; } system::ILogger* m_logger; + std::optional m_seedOverride; + std::optional m_failureSeed; }; @@ -414,6 +424,12 @@ class CSphericalTriangleGenerateTester auto sampler = sampling::SphericalTriangle::create(shape); const float64_t SA = static_cast(shape.solid_angle); + // Float32 solid angle (acos sum - pi) loses precision for small + // triangles due to catastrophic cancellation, making the expected + // sub-solid-angle ratio unreliable as a reference value. + // At SA ~ 0.003, the relative error in float32 solid angles reaches + // ~1-3%, comparable to the half-space counting tolerance. + const bool tinyTriangle = SA < 4e-3; // For each cut: pick a vertex and a point on the opposite edge, // forming a great circle that splits the triangle in two. @@ -482,12 +498,20 @@ class CSphericalTriangleGenerateTester testedCuts++; if (absErr > relTol) { - ctx.failCount++; - if (ctx.failCount <= 5) + if (tinyTriangle) { - m_logger->log("[SphericalTriangle::generate] %s half-space: observed=%f expected=%f absErr=%e (tol=%e) tri %u cut %u", - system::ILogger::ELL_ERROR, label, observedFraction, expectedFraction, absErr, relTol, t, c); - logTriangleInfo(m_logger, v0, v1, v2); + m_logger->log("[SphericalTriangle::generate] %s half-space: observed=%f expected=%f absErr=%e (tol=%e) tri %u cut %u -- solid angle %e too small for float32, especially on GPU", + system::ILogger::ELL_WARNING, label, observedFraction, expectedFraction, absErr, relTol, t, c, SA); + } + else + { + ctx.failCount++; + if (ctx.failCount <= 5) + { + m_logger->log("[SphericalTriangle::generate] %s half-space: observed=%f expected=%f absErr=%e (tol=%e) tri %u cut %u", + system::ILogger::ELL_ERROR, label, observedFraction, expectedFraction, absErr, relTol, t, c); + logTriangleInfo(m_logger, v0, v1, v2); + } } } } @@ -504,12 +528,20 @@ class CSphericalTriangleGenerateTester } // ------------------------------------------------------------------------- - // Moment matching: E[dot(generate(u), N)] should equal PSA(N) / SA. + // Moment matching: E[dot(generate(u), N)] should equal signedPSA(N) / SA. // // For a uniform distribution over a spherical triangle: // E[f(L)] = (1/SA) * integral_triangle f(L) dw // - // Choosing f(L) = dot(L, N) gives E[dot(L, N)] = PSA(N) / SA. + // Choosing f(L) = dot(L, N) gives E[dot(L, N)] = signedPSA(N) / SA, + // where signedPSA is the exact signed projected solid angle computed + // via the Kelvin-Stokes theorem: + // signedPSA(N) = 0.5 * sum_edges dot(edgeNormal_i, N) * edgeArcLength_i + // + // Note: shapes::SphericalTriangle::projectedSolidAngle() returns a signed result + // (Kelvin-Stokes signed sum); tests abs() the return to compare against the + // |cos(theta)| (BSDF) PSA integral reference. + // // If generate() has a systematic bias (e.g., concentrating samples // near one vertex), this moment will be wrong for most directions N. // Testing multiple random N per triangle makes it very unlikely that @@ -533,11 +565,34 @@ class CSphericalTriangleGenerateTester auto sampler = sampling::SphericalTriangle::create(shape); const float64_t SA = static_cast(shape.solid_angle); + // Precompute edge normals and arc lengths for the signed PSA formula. + // cross(v_j, v_k) * csc_sides[i] gives outward-pointing edge normals + // only when the vertices are CCW as seen from outside the sphere. + // The sign of the triple product dot(v0, cross(v1, v2)) tells us the + // winding: positive = CCW (outward normals), negative = CW (inward). + const float32_t3 crossBC = hlsl::cross(shape.vertices[1], shape.vertices[2]); + const float64_t windingSign = (hlsl::dot(shape.vertices[0], crossBC) >= 0.0f) ? 1.0 : -1.0; + const float32_t3 edgeNormals[3] = { + crossBC * shape.csc_sides[0], + hlsl::cross(shape.vertices[2], shape.vertices[0]) * shape.csc_sides[1], + hlsl::cross(shape.vertices[0], shape.vertices[1]) * shape.csc_sides[2] + }; + const float64_t edgeAngles[3] = { + std::acos(static_cast(hlsl::clamp(shape.cos_sides[0], -1.0f, 1.0f))), + std::acos(static_cast(hlsl::clamp(shape.cos_sides[1], -1.0f, 1.0f))), + std::acos(static_cast(hlsl::clamp(shape.cos_sides[2], -1.0f, 1.0f))) + }; + for (uint32_t n = 0; n < numNormals; n++) { float32_t3 N = generateRandomUnitVector(ctx.rng); - const float64_t psa = static_cast(shape.projectedSolidAngle(N)); - const float64_t expected = psa / SA; + + // Signed PSA via Kelvin-Stokes: exact for integral dot(L,N) dOmega + float64_t signedPSA = 0.0; + for (uint32_t e = 0; e < 3; e++) + signedPSA += static_cast(hlsl::dot(edgeNormals[e], N)) * edgeAngles[e]; + signedPSA *= 0.5 * windingSign; + const float64_t expected = signedPSA / SA; float64_t sum = 0.0; std::uniform_real_distribution uDist(0.0f, 1.0f); @@ -546,7 +601,7 @@ class CSphericalTriangleGenerateTester float32_t2 u(uDist(ctx.rng), uDist(ctx.rng)); typename sampling::SphericalTriangle::cache_type cache; float32_t3 L = sampler.generate(u, cache); - sum += static_cast(hlsl::abs(dot(L, N))); + sum += static_cast(dot(L, N)); } const float64_t mcEstimate = sum / static_cast(numSamples); @@ -601,7 +656,7 @@ class CSphericalTriangleGenerateTester if (shape.solid_angle <= 0.0f || !std::isfinite(shape.solid_angle)) continue; - auto sampler = sampling::SphericalTriangle::create(shape); + auto sampler = sampling::SphericalTriangle::create(shape); std::uniform_real_distribution uDist(0.0f, 1.0f); for (uint32_t i = 0; i < samplesPerTriangle; i++) @@ -742,7 +797,7 @@ class CSphericalTriangleGenerateTester // Tests two aspects of projected spherical triangles: // // 1. PSA formula accuracy: shapes::SphericalTriangle::projectedSolidAngle -// against Monte Carlo ground truth (PSA = integral_{tri} abs(dot(L,N)) dOmega). +// against grid-integration ground truth (PSA = integral_{tri} abs(dot(L,N)) dOmega). // // 2. PST sampler accuracy: how well ProjectedSphericalTriangle's bilinear // importance sampling approximates the true NdotL distribution, and @@ -767,18 +822,21 @@ class CProjectedSphericalTriangleGeometricTester // when edge normals have mixed signs, even when all vertices are above the horizon. // These tests are diagnostic-only until proper hemisphere clipping is implemented. // TODO: make these hard failures once projectedSolidAngle clips to the hemisphere. - testPSAVersusMonteCarlo("random MC", [](std::mt19937& rng, uint32_t, float32_t3& v0, float32_t3& v1, float32_t3& v2, float32_t3& normal) + // Hard-fail thresholds: relErr > 3.0 AND absErr > 0.3 means the formula is catastrophically + // wrong, not just affected by the known abs()-overcount limitation. Catches regressions that + // would otherwise hide in the warning stream. + pass &= testPSAVersusGrid("random", [](std::mt19937& rng, uint32_t, float32_t3& v0, float32_t3& v1, float32_t3& v2, float32_t3& normal) { generateRandomTriangleVertices(rng, v0, v1, v2); - normal = generateRandomUnitVector(rng); }, 200, 500000, 0.05, 0.01, true); - testPSAVersusMonteCarlo("grazing MC", [](std::mt19937& rng, uint32_t, float32_t3& v0, float32_t3& v1, float32_t3& v2, float32_t3& normal) + normal = generateRandomUnitVector(rng); }, 200, 500000, 0.05, 0.01, 3.0, 0.3, true); + pass &= testPSAVersusGrid("grazing", [](std::mt19937& rng, uint32_t, float32_t3& v0, float32_t3& v1, float32_t3& v2, float32_t3& normal) { generateRandomTriangleVertices(rng, v0, v1, v2); float32_t3 triCenter = normalize(v0 + v1 + v2); float32_t3 tangent, unused; buildTangentFrame(triCenter, tangent, unused); std::uniform_real_distribution grazeDist(0.02f, 0.15f); - normal = normalize(tangent + triCenter * grazeDist(rng)); }, 200, 500000, 0.1, 0.01, true); + normal = normalize(tangent + triCenter * grazeDist(rng)); }, 200, 500000, 0.1, 0.01, 3.0, 0.3, true); // Also diagnostic -- same abs() issue affects small triangles testPSASmallTriangle(); @@ -860,7 +918,7 @@ class CProjectedSphericalTriangleGeometricTester // Known analytic cases bool testPSAKnownCases() { - constexpr float64_t psaOctantMCRelTol = 0.05; + constexpr float64_t psaOctantGridRelTol = 0.05; constexpr float64_t psaSymmetryRelTol = 1e-4; SeededTestContext ctx; @@ -872,51 +930,52 @@ class CProjectedSphericalTriangleGeometricTester // By Kelvin-Stokes / direct integration, PSA = pi/4 for any axis-aligned normal. { auto shape = createSphericalTriangleShape(float32_t3(1, 0, 0), float32_t3(0, 1, 0), float32_t3(0, 0, 1)); - const float64_t psaZ = static_cast(shape.projectedSolidAngle(float32_t3(0, 0, 1))); + const float64_t psaZ = std::abs(static_cast(shape.projectedSolidAngle(float32_t3(0, 0, 1)))); - // MC verification: sample many points uniformly from the octant triangle - const float64_t mcPSA = mcEstimatePSA(shape, float32_t3(0, 0, 1), 1000000, ctx.rng); + // Grid verification: evaluate abs(N.L) over a dense grid on the octant triangle + const float64_t gridPSA = gridEstimatePSA(shape, float32_t3(0, 0, 1), 1000000); - const float64_t formulaVsMC = std::abs(psaZ - mcPSA) / std::abs(mcPSA); - m_logger->log(" [PSA] octant z-normal: formula=%f expected(pi/4)=%f reference=%f relErr=%e", - system::ILogger::ELL_PERFORMANCE, psaZ, nbl::hlsl::numbers::pi / 4.0, mcPSA, formulaVsMC); + const float64_t formulaVsGrid = std::abs(psaZ - gridPSA) / std::abs(gridPSA); + m_logger->log(" [TriPSA] octant z-normal: formula=%f expected(pi/4)=%f reference=%f relErr=%e", + system::ILogger::ELL_PERFORMANCE, psaZ, nbl::hlsl::numbers::pi / 4.0, gridPSA, formulaVsGrid); - if (formulaVsMC > psaOctantMCRelTol) + if (formulaVsGrid > psaOctantGridRelTol) { - m_logger->log(" [PSA] octant z-normal FAILED: formula=%f expected(reference)=%f relErr=%e relTol=%e", - system::ILogger::ELL_ERROR, psaZ, mcPSA, formulaVsMC, psaOctantMCRelTol); + m_logger->log(" [TriPSA] octant z-normal FAILED: formula=%f expected(reference)=%f relErr=%e relTol=%e", + system::ILogger::ELL_ERROR, psaZ, gridPSA, formulaVsGrid, psaOctantGridRelTol); pass = false; } // Same octant, normal = (1,0,0): by symmetry same result as z-normal - const float64_t psaX = static_cast(shape.projectedSolidAngle(float32_t3(1, 0, 0))); + const float64_t psaX = std::abs(static_cast(shape.projectedSolidAngle(float32_t3(1, 0, 0)))); const float64_t relDiff = std::abs(psaZ - psaX) / std::max(psaZ, psaX); - m_logger->log(" [PSA] octant symmetry: psaZ=%f psaX=%f relDiff=%e", + m_logger->log(" [TriPSA] octant symmetry: psaZ=%f psaX=%f relDiff=%e", system::ILogger::ELL_PERFORMANCE, psaZ, psaX, relDiff); if (relDiff > psaSymmetryRelTol) { - m_logger->log(" [PSA] octant symmetry FAILED: psaZ=%f psaX=%f relDiff=%e relTol=%e", + m_logger->log(" [TriPSA] octant symmetry FAILED: psaZ=%f psaX=%f relDiff=%e relTol=%e", system::ILogger::ELL_ERROR, psaZ, psaX, relDiff, psaSymmetryRelTol); pass = false; } } if (pass) - m_logger->log(" [PSA] known cases PASSED (octant z-normal vs MC relTol=%e, octant symmetry z vs x relTol=%e)", - system::ILogger::ELL_PERFORMANCE, psaOctantMCRelTol, psaSymmetryRelTol); + m_logger->log(" [TriPSA] known cases PASSED (octant z-normal vs grid relTol=%e, octant symmetry z vs x relTol=%e)", + system::ILogger::ELL_PERFORMANCE, psaOctantGridRelTol, psaSymmetryRelTol); - return ctx.finalize(pass, m_logger, "PSA"); + return ctx.finalize(pass, m_logger, "TriPSA"); } - // Helper: run MC comparison of formulaPSA vs E[dot(L,N)]*SA for a set of triangle configs. + // Helper: run grid-integration comparison of formulaPSA vs PSA reference for a set of triangle configs. // TriConfigGen: void(rng, index, v0, v1, v2, normal) — generates triangle vertices + normal. template - bool testPSAVersusMonteCarlo(const char* label, TriConfigGen triConfigGenerator, uint32_t numConfigs, uint32_t mcSamples, float64_t relTol, float64_t absTol, bool diagnostic = false) + bool testPSAVersusGrid(const char* label, TriConfigGen triConfigGenerator, uint32_t numConfigs, uint32_t gridSamples, + float64_t relTol, float64_t absTol, float64_t hardRelTol, float64_t hardAbsTol, bool diagnostic = false) { - return ::testPSAVersusMonteCarlo(m_logger, "PSA", label, - [&](std::mt19937& rng, uint32_t c, float64_t& formulaPSA, float64_t& mcPSA, auto& logInfo) + return ::testPSAVersusGrid(m_logger, "TriPSA", label, + [&](std::mt19937& rng, uint32_t c, float64_t& formulaPSA, float64_t& gridPSA, auto& logInfo) { float32_t3 v0, v1, v2, normal; triConfigGenerator(rng, c, v0, v1, v2, normal); @@ -925,8 +984,8 @@ class CProjectedSphericalTriangleGeometricTester if (shape.solid_angle <= 0.0f || !std::isfinite(shape.solid_angle)) return; - formulaPSA = static_cast(shape.projectedSolidAngle(normal)); - mcPSA = mcEstimatePSA(shape, normal, mcSamples, rng); + formulaPSA = std::abs(static_cast(shape.projectedSolidAngle(normal))); + gridPSA = gridEstimatePSA(shape, normal, gridSamples); logInfo = [=](system::ILogger* logger, system::ILogger::E_LOG_LEVEL level) { using nbl::system::to_string; @@ -935,14 +994,14 @@ class CProjectedSphericalTriangleGeometricTester to_string(normal).c_str(), to_string(shape.solid_angle).c_str()); }; }, - numConfigs, relTol, absTol, diagnostic); + numConfigs, relTol, absTol, hardRelTol, hardAbsTol, diagnostic); } - // Small triangles -- PSA should approach MC ground truth + // Small triangles -- PSA should approach grid ground truth bool testPSASmallTriangle() { constexpr float64_t smallTriMeanRelErrTol = 0.1; - constexpr uint32_t smallTriMCSamples = 100000; + constexpr uint32_t smallTriGridSamples = 100000; SeededTestContext ctx; bool pass = true; @@ -973,27 +1032,27 @@ class CProjectedSphericalTriangleGeometricTester if (shape.solid_angle <= 0.0f || !std::isfinite(shape.solid_angle)) continue; - const float64_t formulaPSA = static_cast(shape.projectedSolidAngle(normal)); + const float64_t formulaPSA = std::abs(static_cast(shape.projectedSolidAngle(normal))); const float64_t sa = static_cast(shape.solid_angle); const float64_t centerNdotL = static_cast(dot(normal, baseDir)); if (std::abs(centerNdotL) < 0.1 || sa < 1e-10) continue; - // MC ground truth: E[abs(dot(L, N))] * solidAngle - const float64_t mcPSA = mcEstimatePSA(shape, normal, smallTriMCSamples, ctx.rng); + // Grid ground truth: mean over regular [0,1]^2 grid of abs(dot(L, N)) * solidAngle + const float64_t gridPSA = gridEstimatePSA(shape, normal, smallTriGridSamples); - if (std::abs(mcPSA) < 1e-10) + if (std::abs(gridPSA) < 1e-10) continue; - const float64_t relErr = (formulaPSA - mcPSA) / mcPSA; + const float64_t relErr = (formulaPSA - gridPSA) / gridPSA; sumRelErrPerSize[s] += relErr; validTrials[s]++; } } - m_logger->log(" [PSA] small triangle PSA vs MC (signed relErr, positive=overestimate):", system::ILogger::ELL_PERFORMANCE); + m_logger->log(" [TriPSA] small triangle PSA vs grid (signed relErr, positive=overestimate):", system::ILogger::ELL_PERFORMANCE); for (uint32_t s = 0; s < numSizes; s++) { if (validTrials[s] > 0) @@ -1005,14 +1064,14 @@ class CProjectedSphericalTriangleGeometricTester // Skip halfAngle=0.01 (s==5): float32 solid angle precision collapses if (s == 4 && std::abs(meanRelErr) > smallTriMeanRelErrTol) { - m_logger->log(" [PSA] small triangle exceeded tolerance at halfAngle=%.3f meanRelErr=%+e meanRelErrTol=%e (%u trials)", + m_logger->log(" [TriPSA] small triangle exceeded tolerance at halfAngle=%.3f meanRelErr=%+e meanRelErrTol=%e (%u trials)", system::ILogger::ELL_WARNING, halfAngles[s], meanRelErr, smallTriMeanRelErrTol, validTrials[s]); } } } - m_logger->log(" [PSA] small triangle test complete (%u trials across %u sizes, %u MC samples each, meanRelErrTol=%e) -- diagnostic only", - system::ILogger::ELL_PERFORMANCE, numTrials, numSizes, smallTriMCSamples, smallTriMeanRelErrTol); + m_logger->log(" [TriPSA] small triangle test complete (%u trials across %u sizes, %u grid samples each, meanRelErrTol=%e) -- diagnostic only", + system::ILogger::ELL_PERFORMANCE, numTrials, numSizes, smallTriGridSamples, smallTriMeanRelErrTol); return true; // diagnostic only -- abs()-based PSA overestimates, not a hard failure } @@ -1076,7 +1135,7 @@ class CProjectedSphericalTriangleGeometricTester if (!std::isfinite(sampler.sphtri.rcpSolidAngle) || sampler.sphtri.rcpSolidAngle <= 0.0f) continue; - const float64_t projSA = static_cast(shape.projectedSolidAngle(cfg.normal)); + const float64_t projSA = std::abs(static_cast(shape.projectedSolidAngle(cfg.normal))); const bool hasPSA = projSA > 0.0 && std::isfinite(projSA); const float64_t rcpPSA = hasPSA ? 1.0 / projSA : 0.0; MISStats& mis = isGrazing ? grazingMIS : normalMIS; @@ -1090,7 +1149,7 @@ class CProjectedSphericalTriangleGeometricTester float32_t3 L = sampler.generate(u, cache); const float64_t trueNdotL = std::max(0.0, static_cast(dot(cfg.normal, L))); - const float64_t bilinearNdotL = static_cast(cache.abs_cos_theta); + const float64_t bilinearNdotL = std::numeric_limits::quiet_NaN(); const float64_t pstPdf = static_cast(sampler.forwardPdf(u, cache)); // Bilinear vs true NdotL @@ -1323,7 +1382,7 @@ class CProjectedSphericalTriangleGeometricTester continue; auto sampler = createSampler(cfg); - const float64_t projSA = static_cast(shape.projectedSolidAngle(cfg.normal)); + const float64_t projSA = std::abs(static_cast(shape.projectedSolidAngle(cfg.normal))); if (projSA <= 0.0 || !std::isfinite(projSA) || !std::isfinite(sampler.sphtri.rcpSolidAngle) || sampler.sphtri.rcpSolidAngle <= 0.0f) @@ -1344,7 +1403,11 @@ class CProjectedSphericalTriangleGeometricTester if (trueNdotL < 1e-6) continue; - const float64_t pstPdf = static_cast(sampler.backwardPdf(L)); + // No direct backwardPdf; evaluate forwardPdf at the inverted u to recover pdf(L). + const float32_t2 uInv = sampler.sphtri.generateInverse(L); + typename sampling::ProjectedSphericalTriangle::cache_type pdfCache; + sampler.generate(uInv, pdfCache); + const float64_t pstPdf = static_cast(sampler.forwardPdf(uInv, pdfCache)); const float64_t idealPdf = trueNdotL * rcpPSA; if (!std::isfinite(pstPdf) || pstPdf <= 0.0 || idealPdf <= 0.0) @@ -1416,6 +1479,15 @@ struct UniformRectSamplerPolicy return sampler_type::create(shape, observer); } + // Returns offset-from-r0 on the rectangle surface. Goes through generateLocalBasisXY + // (absolute xy) and subtracts r0.xy so the [0, extents] bounds check still applies. + static float32_t2 generateOffset(sampler_type& s, const float32_t2& u) + { + typename sampler_type::cache_type cache; + const float32_t2 absXY = s.generateLocalBasisXY(u, cache); + return absXY - float32_t2(s.r0.x, s.r0.y); + } + static float getSolidAngle(const sampler_type& s) { return s.solidAngle; } static const char* name() { return "SphericalRectangle"; } @@ -1425,7 +1497,8 @@ struct UniformRectSamplerPolicy struct ProjectedRectSamplerPolicy { - using sampler_type = sampling::ProjectedSphericalRectangle; + // UsePdfAsWeight=false so receiverNormal and projSolidAngle are populated for diagnostic logs. + using sampler_type = sampling::ProjectedSphericalRectangle; static sampler_type createSampler(shapes::SphericalRectangle& shape, const float32_t3& observer, std::mt19937& rng) @@ -1439,6 +1512,17 @@ struct ProjectedRectSamplerPolicy return sampler_type::create(shape, observer, receiverNormal, false); } + // Run u through the bilinear warp then the inner sphrect's generateLocalBasisXY, and subtract + // r0.xy to get offset-from-r0 on the rectangle surface. + static float32_t2 generateOffset(sampler_type& s, const float32_t2& u) + { + typename sampling::Bilinear::cache_type bc; + const float32_t2 warped = s.bilinearPatch.generate(u, bc); + typename sampling::SphericalRectangle::cache_type sphrectCache; + const float32_t2 absXY = s.sphrect.generateLocalBasisXY(warped, sphrectCache); + return absXY - float32_t2(s.sphrect.r0.x, s.sphrect.r0.y); + } + static float getSolidAngle(const sampler_type& s) { return s.sphrect.solidAngle; } static const char* name() { return "ProjectedSphericalRectangle"; } @@ -1635,8 +1719,7 @@ class CRectangleGenerateTester for (uint32_t i = 0; i < numSamples; i++) { float32_t2 u(uDist(ctx.rng), uDist(ctx.rng)); - typename sampler_type::cache_type cache; - float32_t2 gen = sampler.generateSurfaceOffset(u, cache); + float32_t2 gen = Policy::generateOffset(sampler, u); const float coord = cutAlongX ? gen.x : gen.y; if (coord < cutThreshold) countInSub++; @@ -1714,8 +1797,7 @@ class CRectangleGenerateTester for (uint32_t i = 0; i < numSamples; i++) { float32_t2 u(uDist(ctx.rng), uDist(ctx.rng)); - typename sampler_type::cache_type cache; - float32_t2 gen = sampler.generateSurfaceOffset(u, cache); + float32_t2 gen = Policy::generateOffset(sampler, u); float32_t3 dir = reconstructDirection(compressed, shape.extents, observer, gen); sum += static_cast(dot(dir, N)); } @@ -1778,8 +1860,7 @@ class CRectangleGenerateTester for (uint32_t i = 0; i < numSamples; i++) { float32_t2 u(uDist(ctx.rng), uDist(ctx.rng)); - typename sampler_type::cache_type cache; - float32_t2 gen = sampler.generateSurfaceOffset(u, cache); + float32_t2 gen = Policy::generateOffset(sampler, u); if (gen.x < -1e-5f || gen.x > extX + 1e-5f || gen.y < -1e-5f || gen.y > extY + 1e-5f) { @@ -1891,9 +1972,9 @@ using CProjectedSphericalRectangleGenerateTester = CRectangleGenerateTester 3.0 AND absErr > 0.3) still catch catastrophic regressions. + bool pass = true; + pass &= testPSAVersusGrid("random", generateRandomRectangle, 200, 500000, 0.05, 0.01, 3.0, 0.3); + pass &= testPSAVersusGrid("grazing", generateStressRectangle, 200, 500000, 0.1, 0.01, 3.0, 0.3); + return pass; } private: // Reuse rectangle generators from CRectangleGenerateTester using RectGen = void(*)(std::mt19937&, shapes::CompressedSphericalRectangle&, float32_t3&); - bool testPSAVersusMonteCarlo(const char* label, RectGen rectGen, uint32_t numConfigs, uint32_t mcSamples, float64_t relTol, float64_t absTol) + bool testPSAVersusGrid(const char* label, RectGen rectGen, uint32_t numConfigs, uint32_t gridSamples, + float64_t relTol, float64_t absTol, float64_t hardRelTol, float64_t hardAbsTol) { - return ::testPSAVersusMonteCarlo(m_logger, "RectPSA", label, - [&](std::mt19937& rng, uint32_t, float64_t& formulaPSA, float64_t& mcPSA, auto& logInfo) + return ::testPSAVersusGrid(m_logger, "RectPSA", label, + [&](std::mt19937& rng, uint32_t, float64_t& formulaPSA, float64_t& gridPSA, auto& logInfo) { shapes::CompressedSphericalRectangle compressed; float32_t3 observer; @@ -1932,7 +2016,9 @@ class CProjectedSphericalRectangleGeometricTester float32_t3 normal = generateRandomUnitVector(rng); formulaPSA = static_cast(shape.projectedSolidAngle(observer, normal)); - mcPSA = mcEstimatePSA(shape, observer, normal, mcSamples, rng); + // surfaceGridEstimatePSA integrates over the rectangle surface directly (no sampler in + // the loop), so a formula-vs-reference mismatch here isolates the PSA formula. + gridPSA = surfaceGridEstimatePSA(shape, observer, normal, gridSamples); logInfo = [compressed, observer, normal, saValue = sa.value](system::ILogger* logger, system::ILogger::E_LOG_LEVEL level) { using nbl::system::to_string; @@ -1945,7 +2031,7 @@ class CProjectedSphericalRectangleGeometricTester to_string(saValue).c_str()); }; }, - numConfigs, relTol, absTol, true); + numConfigs, relTol, absTol, hardRelTol, hardAbsTol, true); } system::ILogger* m_logger; diff --git a/64_EmulatedFloatTest/main.cpp b/64_EmulatedFloatTest/main.cpp index 7919f68c5..549596bac 100644 --- a/64_EmulatedFloatTest/main.cpp +++ b/64_EmulatedFloatTest/main.cpp @@ -6,6 +6,8 @@ #include "nbl/examples/examples.hpp" #include +#include +#include #include #include #include @@ -17,6 +19,8 @@ #include +#include "nbl/examples/Benchmark/IBenchmark.h" +#include "nbl/examples/Benchmark/GPUBenchmarkHelper.h" using namespace nbl::core; using namespace nbl::hlsl; @@ -26,1195 +30,1031 @@ using namespace nbl::video; using namespace nbl::application_templates; using namespace nbl::examples; -constexpr bool DoTests = true; +constexpr bool DoTests = true; constexpr bool DoBenchmark = true; +// One row per EF64_BENCHMARK_MODE. Each instance owns its own write-sink +// buffer + descriptor set; the framework's GPUBenchmarkHelper handles +// cmdbuf / queryPool / pipeline-stats capture / runTimed timing, IBenchmark +// routes the result through the Aggregator. The shader binds an SSBO at +// set 0 / binding 0, so we pass an explicit dsLayout to createPipeline. +class CEF64Benchmark : public GPUBenchmark +{ + public: + static constexpr const char* kSectionLabel = "EF64 Benchmarks"; + + struct SetupData + { + smart_refctd_ptr assetMgr; + core::vector name; // hierarchical row name + EF64_BENCHMARK_MODE mode; // pushed each run() via PC + GPUBenchmarkHelper::ShaderVariant variant; // precompiled "benchmark" SPIRV + uint32_t warmupDispatches; + uint64_t targetBudgetMs; + }; + + // Shape is fixed by the BENCHMARK_WORKGROUP_* macros; expose it so the + // caller uses the same shape both to construct the bench and to build the + // RunContext for its span. + static WorkloadShape shape() + { + const hlsl::uint32_t3 wg = { + BENCHMARK_WORKGROUP_DIMENSION_SIZE_X, + BENCHMARK_WORKGROUP_DIMENSION_SIZE_Y, + BENCHMARK_WORKGROUP_DIMENSION_SIZE_Z}; + const hlsl::uint32_t3 dgc = {BENCHMARK_WORKGROUP_COUNT, 1u, 1u}; + // Shader writes one float64 per thread per dispatch; "sample" == "thread output". + const uint64_t samplesPerDispatch = uint64_t(dgc.x) * dgc.y * dgc.z * wg.x * wg.y * wg.z; + return {.workgroupSize = wg, .dispatchGroupCount = dgc, .samplesPerDispatch = samplesPerDispatch}; + } + + CEF64Benchmark(Aggregator& aggregator, const SetupData& data) + : GPUBenchmark(aggregator, GPUBenchmark::SetupData{ + .name = data.name, + .warmupDispatches = data.warmupDispatches, + .shape = shape(), + .targetBudgetMs = data.targetBudgetMs, + }) + , m_mode(data.mode) + { + // Buffer the shader writes to (descriptor-bound; not BDA). Sized for one + // float64 per thread; the GPU never reads it back to host. + m_buffer = createOutputBuffer(getShape().samplesPerDispatch * sizeof(float64_t)); + + // One SSBO at set 0 / binding 0. createSingleBindingDS wires the + // layout + pool + DS + write descriptor in one call. + auto ds = createSingleBindingDS(m_buffer); + m_dsLayout = std::move(ds.layout); + m_ds = std::move(ds.set); + m_pipelineIdx = createPipeline(data.variant, data.assetMgr, sizeof(BenchmarkPushConstants), joinName(data.name), m_dsLayout); + } + + void doRun() override + { + const PipelineEntry* pe = getPipelineEntry(m_pipelineIdx, joinName(m_name)); + if (!pe) + return; + BenchmarkPushConstants pc = {}; + pc.benchmarkMode = m_mode; + + const TimingResult t = runTimedBudgeted(getWarmupDispatches(), getTargetBudgetMs(), + [&](IGPUCommandBuffer* cb) + { + cb->bindDescriptorSets(EPBP_COMPUTE, pe->layout.get(), 0, 1, &m_ds.get()); + defaultBindAndPush(cb, *pe, pc); + }, + [this](IGPUCommandBuffer* cb) { defaultDispatch(cb); }, + samplesForCurrentRow()); + + record(m_name, t, pe->stats); + } + + private: + EF64_BENCHMARK_MODE m_mode = EF64_BENCHMARK_MODE::NATIVE; + smart_refctd_ptr m_buffer; + smart_refctd_ptr m_dsLayout; + smart_refctd_ptr m_ds; + uint32_t m_pipelineIdx = 0; +}; + class CompatibilityTest final : public MonoDeviceApplication, public BuiltinResourcesApplication { - using device_base_t = MonoDeviceApplication; - using asset_base_t = BuiltinResourcesApplication; -public: - CompatibilityTest(const path& _localInputCWD, const path& _localOutputCWD, const path& _sharedInputCWD, const path& _sharedOutputCWD) : - IApplicationFramework(_localInputCWD, _localOutputCWD, _sharedInputCWD, _sharedOutputCWD) {} - - virtual SPhysicalDeviceFeatures getPreferredDeviceFeatures() const override - { - auto retval = device_base_t::getPreferredDeviceFeatures(); - retval.pipelineExecutableInfo = true; - return retval; - } - - bool onAppInitialized(smart_refctd_ptr&& system) override - { - // since emulated_float64_t rounds to zero - std::fesetround(FE_TOWARDZERO); - - if (!device_base_t::onAppInitialized(smart_refctd_ptr(system))) - return false; - if (!asset_base_t::onAppInitialized(std::move(system))) - return false; - - return true; - } - - void onAppTerminated_impl() override - { - m_device->waitIdle(); - } - - void workLoopBody() override - { - if constexpr (DoTests) - { - emulated_float64_tests(); - } - if constexpr (DoBenchmark) - { - EF64Benchmark benchmark(*this); - benchmark.run(); - } - - m_keepRunning = false; - } - - bool keepRunning() override - { - return m_keepRunning; - } - - -private: - - bool m_keepRunning = true; - - constexpr static inline uint32_t EmulatedFloat64TestIterations = 1000u; - - enum class EmulatedFloatTestDevice - { - CPU, - GPU - }; - - template - bool compareEmulatedFloat64TestValues(const TestValues& expectedValues, const TestValues& testValues) - { - bool success = true; - - auto printOnFailure = [this](EmulatedFloatTestDevice device) - { - std::string errorMsgPrefix = ""; - if (device == EmulatedFloatTestDevice::CPU) - errorMsgPrefix = "CPU test fail:"; - else - errorMsgPrefix = "GPU test fail:"; - - m_logger->log("%s", ILogger::ELL_ERROR, errorMsgPrefix.c_str()); - m_logFile << errorMsgPrefix << '\n'; - }; - - auto printOnArithmeticFailure = [this](const char* valName, uint64_t expectedValue, uint64_t testValue, uint64_t a, uint64_t b) - { - double expectedAsDouble = reinterpret_cast(expectedValue); - double testAsDouble = reinterpret_cast(testValue); - double error = std::abs(expectedAsDouble - testAsDouble); - - std::stringstream ss; - ss << "for input values: A = " << reinterpret_cast(a) << " B = " << reinterpret_cast(b) << '\n'; - ss << valName << " not equal!"; - ss << "\nexpected value: " << std::fixed << std::setprecision(20) << expectedAsDouble; - ss << "\ntest value: " << std::fixed << std::setprecision(20) << testAsDouble; - ss << "\nerror = " << error << '\n'; - ss << "bit representations: \n"; - ss << "seeeeeeeeeeemmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm\n"; - ss << std::bitset<64>(expectedValue) << " - expectedValue bit pattern\n"; - ss << std::bitset<64>(testValue) << " - testValue bit pattern \n"; - - m_logger->log("%s", ILogger::ELL_ERROR, ss.str().c_str()); - m_logFile << ss.str() << '\n'; - - //std::cout << "ULP error: " << std::max(expectedValue, testValue) - std::min(expectedValue, testValue) << "\n\n"; - - }; - - auto calcULPError = [](emulated_float64_t::storage_t expectedValue, emulated_float64_t::storage_t testValue) - { - return std::max(expectedValue, testValue) - std::min(expectedValue, testValue); - }; - - auto printOnComparisonFailure = [this](const char* valName, int expectedValue, int testValue, double a, double b) - { - std::string inputValuesStr = std::string("for input values: A = ") + std::to_string(a) + std::string(" B = ") + std::to_string(b); - - m_logger->log("%s", ILogger::ELL_ERROR, inputValuesStr.c_str()); - m_logFile << inputValuesStr << '\n'; - - std::stringstream ss; - ss << valName << " not equal!"; - ss << "\nexpected value: " << std::boolalpha << bool(expectedValue); - ss << "\ntest value: " << std::boolalpha << bool(testValue); - - m_logger->log("%s", ILogger::ELL_ERROR, ss.str().c_str()); - m_logFile << ss.str() << '\n'; - }; - - if (calcULPError(expectedValues.int32CreateVal, testValues.int32CreateVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("int32CreateVal", expectedValues.int32CreateVal, testValues.int32CreateVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.int64CreateVal, testValues.int64CreateVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("int64CreateVal", expectedValues.int64CreateVal, testValues.int64CreateVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.uint32CreateVal, testValues.uint32CreateVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("uint32CreateVal", expectedValues.uint32CreateVal, testValues.uint32CreateVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.uint64CreateVal, testValues.uint64CreateVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("uint64CreateVal", expectedValues.uint64CreateVal, testValues.uint64CreateVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.float32CreateVal, testValues.float32CreateVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("float32CreateVal", expectedValues.float32CreateVal, testValues.float32CreateVal, expectedValues.a, expectedValues.b); - success = false; - } - if (expectedValues.float64CreateVal != testValues.float64CreateVal) - { - printOnFailure(Device); - printOnArithmeticFailure("float64CreateVal", expectedValues.float64CreateVal, testValues.float64CreateVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.additionVal, testValues.additionVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("additionVal", expectedValues.additionVal, testValues.additionVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.substractionVal, testValues.substractionVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("substractionVal", expectedValues.substractionVal, testValues.substractionVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.multiplicationVal, testValues.multiplicationVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("multiplicationVal", expectedValues.multiplicationVal, testValues.multiplicationVal, expectedValues.a, expectedValues.b); - success = false; - } - if (calcULPError(expectedValues.divisionVal, testValues.divisionVal) > 1u) - { - printOnFailure(Device); - printOnArithmeticFailure("divisionVal", expectedValues.divisionVal, testValues.divisionVal, expectedValues.a, expectedValues.b); - success = false; - } - if (expectedValues.lessOrEqualVal != testValues.lessOrEqualVal) - { - printOnFailure(Device); - printOnComparisonFailure("lessOrEqualVal", expectedValues.lessOrEqualVal, testValues.lessOrEqualVal, expectedValues.a, expectedValues.b); - success = false; - } - if (expectedValues.greaterOrEqualVal != testValues.greaterOrEqualVal) - { - printOnFailure(Device); - printOnComparisonFailure("greaterOrEqualVal", expectedValues.greaterOrEqualVal, testValues.greaterOrEqualVal, expectedValues.a, expectedValues.b); - success = false; - } - if (expectedValues.equalVal != testValues.equalVal) - { - printOnFailure(Device); - printOnComparisonFailure("equalVal", expectedValues.equalVal, testValues.equalVal, expectedValues.a, expectedValues.b); - success = false; - } - if (expectedValues.notEqualVal != testValues.notEqualVal) - { - printOnFailure(Device); - printOnComparisonFailure("notEqualVal", expectedValues.notEqualVal, testValues.notEqualVal, expectedValues.a, expectedValues.b); - success = false; - } - if (expectedValues.lessVal != testValues.lessVal) - { - printOnFailure(Device); - printOnComparisonFailure("lessVal", expectedValues.lessVal, testValues.lessVal, expectedValues.a, expectedValues.b); - success = false; - } - if (expectedValues.greaterVal != testValues.greaterVal) - { - printOnFailure(Device); - printOnComparisonFailure("greaterVal", expectedValues.greaterVal, testValues.greaterVal, expectedValues.a, expectedValues.b); - success = false; - } - - return success; - }; - - class EF64Submitter - { - public: - EF64Submitter(CompatibilityTest& base) - :m_base(base), m_pushConstants({}), m_semaphoreCounter(0) - { - // setting up pipeline in the constructor - m_queueFamily = base.getComputeQueue()->getFamilyIndex(); - m_semaphore = base.m_device->createSemaphore(0); - m_cmdpool = base.m_device->createCommandPool(m_queueFamily, IGPUCommandPool::CREATE_FLAGS::RESET_COMMAND_BUFFER_BIT); - if (!m_cmdpool->createCommandBuffers(IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_cmdbuf)) - base.logFail("Failed to create Command Buffers!\n"); - - // Load shaders, set up pipeline + using device_base_t = MonoDeviceApplication; + using asset_base_t = BuiltinResourcesApplication; + + public: + CompatibilityTest(const path& _localInputCWD, const path& _localOutputCWD, const path& _sharedInputCWD, const path& _sharedOutputCWD) : IApplicationFramework(_localInputCWD, _localOutputCWD, _sharedInputCWD, _sharedOutputCWD) {} + + virtual SPhysicalDeviceFeatures getPreferredDeviceFeatures() const override + { + auto retval = device_base_t::getPreferredDeviceFeatures(); + retval.pipelineExecutableInfo = true; + return retval; + } + + bool onAppInitialized(smart_refctd_ptr&& system) override + { + // since emulated_float64_t rounds to zero + std::fesetround(FE_TOWARDZERO); + + if (!device_base_t::onAppInitialized(smart_refctd_ptr(system))) + return false; + if (!asset_base_t::onAppInitialized(std::move(system))) + return false; + + return true; + } + + void onAppTerminated_impl() override + { + m_device->waitIdle(); + } + + void workLoopBody() override + { + if constexpr (DoTests) + { + emulated_float64_tests(); + } + if constexpr (DoBenchmark) + { + runEF64Benchmarks(); + } + + m_keepRunning = false; + } + + bool keepRunning() override + { + return m_keepRunning; + } + + + private: + bool m_keepRunning = true; + + constexpr static inline uint32_t EmulatedFloat64TestIterations = 1000u; + + enum class EmulatedFloatTestDevice + { + CPU, + GPU + }; + + template + bool compareEmulatedFloat64TestValues(const TestValues& expectedValues, const TestValues& testValues) + { + bool success = true; + + auto printOnFailure = [this](EmulatedFloatTestDevice device) + { + std::string errorMsgPrefix = ""; + if (device == EmulatedFloatTestDevice::CPU) + errorMsgPrefix = "CPU test fail:"; + else + errorMsgPrefix = "GPU test fail:"; + + m_logger->log("%s", ILogger::ELL_ERROR, errorMsgPrefix.c_str()); + m_logFile << errorMsgPrefix << '\n'; + }; + + auto printOnArithmeticFailure = [this](const char* valName, uint64_t expectedValue, uint64_t testValue, uint64_t a, uint64_t b) + { + double expectedAsDouble = reinterpret_cast(expectedValue); + double testAsDouble = reinterpret_cast(testValue); + double error = std::abs(expectedAsDouble - testAsDouble); + + std::stringstream ss; + ss << "for input values: A = " << reinterpret_cast(a) << " B = " << reinterpret_cast(b) << '\n'; + ss << valName << " not equal!"; + ss << "\nexpected value: " << std::fixed << std::setprecision(20) << expectedAsDouble; + ss << "\ntest value: " << std::fixed << std::setprecision(20) << testAsDouble; + ss << "\nerror = " << error << '\n'; + ss << "bit representations: \n"; + ss << "seeeeeeeeeeemmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm\n"; + ss << std::bitset<64>(expectedValue) << " - expectedValue bit pattern\n"; + ss << std::bitset<64>(testValue) << " - testValue bit pattern \n"; + + m_logger->log("%s", ILogger::ELL_ERROR, ss.str().c_str()); + m_logFile << ss.str() << '\n'; + + //std::cout << "ULP error: " << std::max(expectedValue, testValue) - std::min(expectedValue, testValue) << "\n\n"; + }; + + auto calcULPError = [](emulated_float64_t::storage_t expectedValue, emulated_float64_t::storage_t testValue) + { + return std::max(expectedValue, testValue) - std::min(expectedValue, testValue); + }; + + auto printOnComparisonFailure = [this](const char* valName, int expectedValue, int testValue, double a, double b) + { + std::string inputValuesStr = std::string("for input values: A = ") + std::to_string(a) + std::string(" B = ") + std::to_string(b); + + m_logger->log("%s", ILogger::ELL_ERROR, inputValuesStr.c_str()); + m_logFile << inputValuesStr << '\n'; + + std::stringstream ss; + ss << valName << " not equal!"; + ss << "\nexpected value: " << std::boolalpha << bool(expectedValue); + ss << "\ntest value: " << std::boolalpha << bool(testValue); + + m_logger->log("%s", ILogger::ELL_ERROR, ss.str().c_str()); + m_logFile << ss.str() << '\n'; + }; + + if (calcULPError(expectedValues.int32CreateVal, testValues.int32CreateVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("int32CreateVal", expectedValues.int32CreateVal, testValues.int32CreateVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.int64CreateVal, testValues.int64CreateVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("int64CreateVal", expectedValues.int64CreateVal, testValues.int64CreateVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.uint32CreateVal, testValues.uint32CreateVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("uint32CreateVal", expectedValues.uint32CreateVal, testValues.uint32CreateVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.uint64CreateVal, testValues.uint64CreateVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("uint64CreateVal", expectedValues.uint64CreateVal, testValues.uint64CreateVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.float32CreateVal, testValues.float32CreateVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("float32CreateVal", expectedValues.float32CreateVal, testValues.float32CreateVal, expectedValues.a, expectedValues.b); + success = false; + } + if (expectedValues.float64CreateVal != testValues.float64CreateVal) + { + printOnFailure(Device); + printOnArithmeticFailure("float64CreateVal", expectedValues.float64CreateVal, testValues.float64CreateVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.additionVal, testValues.additionVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("additionVal", expectedValues.additionVal, testValues.additionVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.substractionVal, testValues.substractionVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("substractionVal", expectedValues.substractionVal, testValues.substractionVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.multiplicationVal, testValues.multiplicationVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("multiplicationVal", expectedValues.multiplicationVal, testValues.multiplicationVal, expectedValues.a, expectedValues.b); + success = false; + } + if (calcULPError(expectedValues.divisionVal, testValues.divisionVal) > 1u) + { + printOnFailure(Device); + printOnArithmeticFailure("divisionVal", expectedValues.divisionVal, testValues.divisionVal, expectedValues.a, expectedValues.b); + success = false; + } + if (expectedValues.lessOrEqualVal != testValues.lessOrEqualVal) + { + printOnFailure(Device); + printOnComparisonFailure("lessOrEqualVal", expectedValues.lessOrEqualVal, testValues.lessOrEqualVal, expectedValues.a, expectedValues.b); + success = false; + } + if (expectedValues.greaterOrEqualVal != testValues.greaterOrEqualVal) + { + printOnFailure(Device); + printOnComparisonFailure("greaterOrEqualVal", expectedValues.greaterOrEqualVal, testValues.greaterOrEqualVal, expectedValues.a, expectedValues.b); + success = false; + } + if (expectedValues.equalVal != testValues.equalVal) + { + printOnFailure(Device); + printOnComparisonFailure("equalVal", expectedValues.equalVal, testValues.equalVal, expectedValues.a, expectedValues.b); + success = false; + } + if (expectedValues.notEqualVal != testValues.notEqualVal) + { + printOnFailure(Device); + printOnComparisonFailure("notEqualVal", expectedValues.notEqualVal, testValues.notEqualVal, expectedValues.a, expectedValues.b); + success = false; + } + if (expectedValues.lessVal != testValues.lessVal) + { + printOnFailure(Device); + printOnComparisonFailure("lessVal", expectedValues.lessVal, testValues.lessVal, expectedValues.a, expectedValues.b); + success = false; + } + if (expectedValues.greaterVal != testValues.greaterVal) + { + printOnFailure(Device); + printOnComparisonFailure("greaterVal", expectedValues.greaterVal, testValues.greaterVal, expectedValues.a, expectedValues.b); + success = false; + } + + return success; + }; + + class EF64Submitter + { + public: + EF64Submitter(CompatibilityTest& base) + : m_base(base), m_pushConstants({}), m_semaphoreCounter(0) + { + // setting up pipeline in the constructor + m_queueFamily = base.getComputeQueue()->getFamilyIndex(); + m_semaphore = base.m_device->createSemaphore(0); + m_cmdpool = base.m_device->createCommandPool(m_queueFamily, IGPUCommandPool::CREATE_FLAGS::RESET_COMMAND_BUFFER_BIT); + if (!m_cmdpool->createCommandBuffers(IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_cmdbuf)) + base.logFail("Failed to create Command Buffers!\n"); + + // Load shaders, set up pipeline + { + smart_refctd_ptr shader; { - smart_refctd_ptr shader; - { - IAssetLoader::SAssetLoadParams lp = {}; - lp.logger = base.m_logger.get(); - lp.workingDirectory = "app_resources"; // virtual root - - auto key = nbl::this_example::builtin::build::get_spirv_key<"test">(base.m_device.get()); - auto assetBundle = base.m_assetMgr->getAsset(key.data(), lp); - const auto assets = assetBundle.getContents(); - if (assets.empty()) - { - base.logFail("Could not load shader!"); - assert(0); - } - - // It would be super weird if loading a shader from a file produced more than 1 asset - assert(assets.size() == 1); - shader = IAsset::castDown(assets[0]); - } - - if (!shader) - base.logFail("Failed to load precompiled \"test\" shader!\n"); - - nbl::video::IGPUDescriptorSetLayout::SBinding bindings[1] = { - { - .binding = 0, - .type = nbl::asset::IDescriptor::E_TYPE::ET_STORAGE_BUFFER, - .createFlags = IGPUDescriptorSetLayout::SBinding::E_CREATE_FLAGS::ECF_NONE, - .stageFlags = ShaderStage::ESS_COMPUTE, - .count = 1 - } - }; - smart_refctd_ptr dsLayout = base.m_device->createDescriptorSetLayout(bindings); - if (!dsLayout) - base.logFail("Failed to create a Descriptor Layout!\n"); - - SPushConstantRange pushConstantRanges[] = { - { - .stageFlags = ShaderStage::ESS_COMPUTE, - .offset = 0, - .size = sizeof(PushConstants) - } - }; - m_pplnLayout = base.m_device->createPipelineLayout(pushConstantRanges, smart_refctd_ptr(dsLayout)); - if (!m_pplnLayout) - base.logFail("Failed to create a Pipeline Layout!\n"); - - { - IGPUComputePipeline::SCreationParams params = {}; - params.layout = m_pplnLayout.get(); - params.shader.entryPoint = "main"; - params.shader.shader = shader.get(); - if (base.m_device->getEnabledFeatures().pipelineExecutableInfo) - { - params.flags |= IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_STATISTICS; - params.flags |= IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_INTERNAL_REPRESENTATIONS; - } - if (!base.m_device->createComputePipelines(nullptr, { ¶ms,1 }, &m_pipeline)) - base.logFail("Failed to create pipelines (compile & link shaders)!\n"); - - if (base.m_device->getEnabledFeatures().pipelineExecutableInfo) - { - auto report = system::to_string(m_pipeline->getExecutableInfo()); - base.m_logger->log("EF64Submitter Pipeline Executable Report:\n%s", ILogger::ELL_PERFORMANCE, report.c_str()); - } - } - - // Allocate the memory - { - constexpr size_t BufferSize = sizeof(TestValues); - - nbl::video::IGPUBuffer::SCreationParams params = {}; - params.size = BufferSize; - params.usage = IGPUBuffer::EUF_STORAGE_BUFFER_BIT; - smart_refctd_ptr outputBuff = base.m_device->createBuffer(std::move(params)); - if (!outputBuff) - base.logFail("Failed to create a GPU Buffer of size %d!\n", params.size); - - outputBuff->setObjectDebugName("emulated_float64_t output buffer"); - - nbl::video::IDeviceMemoryBacked::SDeviceMemoryRequirements reqs = outputBuff->getMemoryReqs(); - reqs.memoryTypeBits &= base.m_physicalDevice->getHostVisibleMemoryTypeBits(); - - m_allocation = base.m_device->allocate(reqs, outputBuff.get(), nbl::video::IDeviceMemoryAllocation::EMAF_NONE); - if (!m_allocation.isValid()) - base.logFail("Failed to allocate Device Memory compatible with our GPU Buffer!\n"); - - assert(outputBuff->getBoundMemory().memory == m_allocation.memory.get()); - smart_refctd_ptr pool = base.m_device->createDescriptorPoolForDSLayouts(IDescriptorPool::ECF_NONE, { &dsLayout.get(),1 }); - - m_ds = pool->createDescriptorSet(std::move(dsLayout)); - { - IGPUDescriptorSet::SDescriptorInfo info[1]; - info[0].desc = smart_refctd_ptr(outputBuff); - info[0].info.buffer = { .offset = 0,.size = BufferSize }; - IGPUDescriptorSet::SWriteDescriptorSet writes[1] = { - {.dstSet = m_ds.get(),.binding = 0,.arrayElement = 0,.count = 1,.info = info} - }; - base.m_device->updateDescriptorSets(writes, {}); - } - } - - if (!m_allocation.memory->map({ 0ull,m_allocation.memory->getAllocationSize() }, IDeviceMemoryAllocation::EMCAF_READ)) - base.logFail("Failed to map the Device Memory!\n"); + IAssetLoader::SAssetLoadParams lp = {}; + lp.logger = base.m_logger.get(); + lp.workingDirectory = "app_resources"; // virtual root + + auto key = nbl::this_example::builtin::build::get_spirv_key<"test">(base.m_device.get()); + auto assetBundle = base.m_assetMgr->getAsset(key.data(), lp); + const auto assets = assetBundle.getContents(); + if (assets.empty()) + { + base.logFail("Could not load shader!"); + assert(0); + } + + // It would be super weird if loading a shader from a file produced more than 1 asset + assert(assets.size() == 1); + shader = IAsset::castDown(assets[0]); } - // if the mapping is not coherent the range needs to be invalidated to pull in new data for the CPU's caches - const ILogicalDevice::MappedMemoryRange memoryRange(m_allocation.memory.get(), 0ull, m_allocation.memory->getAllocationSize()); - if (!m_allocation.memory->getMemoryPropertyFlags().hasFlags(IDeviceMemoryAllocation::EMPF_HOST_COHERENT_BIT)) - base.m_device->invalidateMappedMemoryRanges(1, &memoryRange); - - assert(memoryRange.valid() && memoryRange.length >= sizeof(TestValues)); - - m_queue = m_base.m_device->getQueue(m_queueFamily, 0); - } - - ~EF64Submitter() - { - m_allocation.memory->unmap(); - } - - void setPushConstants(PushConstants& pc) - { - m_pushConstants = pc; - } - - TestValues submitGetGPUTestValues() - { - // record command buffer - m_cmdbuf->reset(IGPUCommandBuffer::RESET_FLAGS::NONE); - m_cmdbuf->begin(IGPUCommandBuffer::USAGE::NONE); - m_cmdbuf->beginDebugMarker("emulated_float64_t compute dispatch", vectorSIMDf(0, 1, 0, 1)); - m_cmdbuf->bindComputePipeline(m_pipeline.get()); - m_cmdbuf->bindDescriptorSets(nbl::asset::EPBP_COMPUTE, m_pplnLayout.get(), 0, 1, &m_ds.get()); - m_cmdbuf->pushConstants(m_pplnLayout.get(), IShader::E_SHADER_STAGE::ESS_COMPUTE, 0, sizeof(PushConstants), &m_pushConstants); - m_cmdbuf->dispatch(WORKGROUP_SIZE, 1, 1); - m_cmdbuf->endDebugMarker(); - m_cmdbuf->end(); - - IQueue::SSubmitInfo submitInfos[1] = {}; - const IQueue::SSubmitInfo::SCommandBufferInfo cmdbufs[] = { {.cmdbuf = m_cmdbuf.get()}}; - submitInfos[0].commandBuffers = cmdbufs; - const IQueue::SSubmitInfo::SSemaphoreInfo signals[] = { {.semaphore = m_semaphore.get(), .value = ++m_semaphoreCounter, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT}}; - submitInfos[0].signalSemaphores = signals; - - m_base.m_api->startCapture(); - m_queue->submit(submitInfos); - m_base.m_api->endCapture(); - - m_base.m_device->waitIdle(); - TestValues output; - std::memcpy(&output, static_cast*>(m_allocation.memory->getMappedPointer()), sizeof(TestValues)); - m_base.m_device->waitIdle(); - - return output; - } - - private: - uint32_t m_queueFamily; - nbl::video::IDeviceMemoryAllocator::SAllocation m_allocation = {}; - smart_refctd_ptr m_cmdbuf = nullptr; - smart_refctd_ptr m_cmdpool = nullptr; - smart_refctd_ptr m_ds = nullptr; - smart_refctd_ptr m_pplnLayout = nullptr; - PushConstants m_pushConstants; - CompatibilityTest& m_base; - smart_refctd_ptr m_pipeline; - smart_refctd_ptr m_semaphore; - IQueue* m_queue; - uint64_t m_semaphoreCounter; - }; - - void emulated_float64_tests() - { - EF64Submitter submitter(*this); - - auto printTestOutput = [this](const std::string& functionName, const EmulatedFloat64TestOutput& testResult) - { - std::cout << functionName << ": " << std::endl; - - if (!testResult.cpuTestsSucceed) - logFail("Incorrect CPU determinated values!"); - else - m_logger->log("Correct CPU determinated values!", ILogger::ELL_PERFORMANCE); - - if (!testResult.gpuTestsSucceed) - logFail("Incorrect GPU determinated values!"); - else - m_logger->log("Correct GPU determinated values!", ILogger::ELL_PERFORMANCE); - }; - - m_logFile.open("EmulatedFloatTestLog.txt", std::ios::out | std::ios::trunc); - if (!m_logFile.is_open()) - m_logger->log("Failed to open log file!", system::ILogger::ELL_ERROR); - - printTestOutput("emulatedFloat64RandomValuesTest", emulatedFloat64RandomValuesTest(submitter)); - printTestOutput("emulatedFloat64RandomValuesTestContrastingExponents", emulatedFloat64RandomValuesTestContrastingExponents(submitter)); - printTestOutput("emulatedFloat64NegAndPosZeroTest", emulatedFloat64NegAndPosZeroTest(submitter)); - printTestOutput("emulatedFloat64BothValuesInfTest", emulatedFloat64BothValuesInfTest(submitter)); - printTestOutput("emulatedFloat64BothValuesNegInfTest", emulatedFloat64BothValuesNegInfTest(submitter)); - printTestOutput("emulatedFloat64OneValIsInfOtherIsNegInfTest", emulatedFloat64OneValIsInfOtherIsNegInfTest(submitter)); - printTestOutput("emulatedFloat64OneValIsInfTest", emulatedFloat64OneValIsInfTest(submitter)); - printTestOutput("emulatedFloat64OneValIsNegInfTest", emulatedFloat64OneValIsNegInfTest(submitter)); - if(false) // doesn't work for some reason + fast math is enabled by default - printTestOutput("emulatedFloat64BNaNTest", emulatedFloat64BNaNTest(submitter)); - printTestOutput("emulatedFloat64BInfTest", emulatedFloat64OneValIsZeroTest(submitter)); - printTestOutput("emulatedFloat64BNegInfTest", emulatedFloat64OneValIsNegZeroTest(submitter)); - - m_logFile.close(); - } - - template - struct EmulatedFloat64TestValuesInfo - { - emulated_float64_t a; - emulated_float64_t b; - ConstructorTestValues constrTestValues; - TestValues expectedTestValues; - - void fillExpectedTestValues() - { - double aAsDouble = reinterpret_cast(a); - double bAsDouble = reinterpret_cast(b); - - expectedTestValues.a = a.data; - expectedTestValues.b = b.data; - - expectedTestValues.int32CreateVal = bit_cast(double(constrTestValues.int32)); - expectedTestValues.int64CreateVal = bit_cast(double(constrTestValues.int64)); - expectedTestValues.uint32CreateVal = bit_cast(double(constrTestValues.uint32)); - expectedTestValues.uint64CreateVal = bit_cast(double(constrTestValues.uint64)); - expectedTestValues.float32CreateVal = bit_cast(double(constrTestValues.float32)); - expectedTestValues.float64CreateVal = bit_cast(constrTestValues.float64); - expectedTestValues.additionVal = emulated_float64_t::create(aAsDouble + bAsDouble).data; - expectedTestValues.substractionVal = emulated_float64_t::create(aAsDouble - bAsDouble).data; - expectedTestValues.multiplicationVal = emulated_float64_t::create(aAsDouble * bAsDouble).data; - expectedTestValues.divisionVal = emulated_float64_t::create(aAsDouble / bAsDouble).data; - expectedTestValues.lessOrEqualVal = aAsDouble <= bAsDouble; - expectedTestValues.greaterOrEqualVal = aAsDouble >= bAsDouble; - expectedTestValues.equalVal = aAsDouble == bAsDouble; - expectedTestValues.notEqualVal = aAsDouble != bAsDouble; - expectedTestValues.lessVal = aAsDouble < bAsDouble; - expectedTestValues.greaterVal = aAsDouble > bAsDouble; - } - }; - - struct EmulatedFloat64TestOutput - { - bool cpuTestsSucceed; - bool gpuTestsSucceed; - }; - - EmulatedFloat64TestOutput emulatedFloat64LoopedTests_impl(EF64Submitter& submitter, - const uint32_t iterations, - const std::function& determineValueA, - const std::function& determineValueB) - { - EmulatedFloat64TestOutput output = { true, true }; - - std::uniform_int_distribution i32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_int_distribution i64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_int_distribution u32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_int_distribution u64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_real_distribution fDistribution(-100000.0, 100000.0); - - std::random_device rd; - std::mt19937 mt(rd()); - - for (uint32_t i = 0u; i < iterations; ++i) - { - // generate random test values - EmulatedFloat64TestValuesInfo testValInfo; - double aTmp = determineValueA(); - double bTmp = determineValueB(); - testValInfo.a.data = reinterpret_cast::storage_t&>(aTmp); - testValInfo.b.data = reinterpret_cast::storage_t&>(bTmp); - testValInfo.constrTestValues.int32 = i32Distribution(mt); - testValInfo.constrTestValues.int64 = i64Distribution(mt); - testValInfo.constrTestValues.uint32 = u32Distribution(mt); - testValInfo.constrTestValues.uint64 = u64Distribution(mt); - testValInfo.constrTestValues.float32 = fDistribution(mt); - testValInfo.constrTestValues.float64 = fDistribution(mt); - - testValInfo.fillExpectedTestValues(); - auto singleTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); - - if (!singleTestOutput.cpuTestsSucceed) - output.cpuTestsSucceed = false; - if (!singleTestOutput.gpuTestsSucceed) - output.gpuTestsSucceed = false; - } - - return output; - } - - EmulatedFloat64TestOutput emulatedFloat64RandomValuesTest(EF64Submitter& submitter) - { - auto getRandomFloat64 = []() - { - static std::random_device rd; - static std::mt19937 mt(rd()); - static std::uniform_real_distribution distribution(-100000.0, 100000.0); - - - return distribution(mt); - }; - - return emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations, getRandomFloat64, getRandomFloat64); - } - - EmulatedFloat64TestOutput emulatedFloat64RandomValuesTestContrastingExponents(EF64Submitter& submitter) - { - auto getRandomSmallFloat64 = []() - { - static std::random_device rd; - static std::mt19937 mt(rd()); - static std::uniform_real_distribution distribution(-0.01, 0.01); - - return distribution(mt); - }; - - auto getRandomLargeFloat64 = []() - { - static std::random_device rd; - static std::mt19937 mt(rd()); - static std::uniform_real_distribution distribution(1000000000.0, 2000000000.0); - static std::uniform_int_distribution coinFlipDistribution(0, 1); - - double output = distribution(mt); - if (coinFlipDistribution(mt)) - output = -output; - - return output; - }; - - EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomSmallFloat64, getRandomLargeFloat64); - EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomLargeFloat64, getRandomSmallFloat64); - - EmulatedFloat64TestOutput output; - output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; - output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; - return output; - } - - EmulatedFloat64TestOutput emulatedFloat64BothValuesNaNTest(EF64Submitter& submitter) - { - smart_refctd_ptr semaphore = m_device->createSemaphore(0); - - EmulatedFloat64TestValuesInfo testValInfo; - const float32_t nan32 = std::numeric_limits::quiet_NaN(); - const float64_t nan64 = std::numeric_limits::quiet_NaN(); - testValInfo.a = emulated_float64_t::create(nan64); - testValInfo.b = emulated_float64_t::create(nan64); - testValInfo.constrTestValues = { - .int32 = std::bit_cast(nan32), - .int64 = std::bit_cast(nan64), - .uint32 = std::bit_cast(nan32), - .uint64 = std::bit_cast(nan64), - .float32 = nan32 - //.float64 = nan64 - }; - - testValInfo.fillExpectedTestValues(); - return performEmulatedFloat64Tests(testValInfo, submitter); - } - - EmulatedFloat64TestOutput emulatedFloat64NegAndPosZeroTest(EF64Submitter& submitter) - { - smart_refctd_ptr semaphore = m_device->createSemaphore(0); - - EmulatedFloat64TestValuesInfo testValInfo; - testValInfo.a = emulated_float64_t::create(ieee754::traits::signMask); - testValInfo.b = emulated_float64_t::create(std::bit_cast(0.0)); - testValInfo.constrTestValues = { - .int32 = 0, - .int64 = 0, - .uint32 = 0, - .uint64 = 0, - .float32 = 0 - }; - - testValInfo.fillExpectedTestValues(); - auto firstTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); - std::swap(testValInfo.a, testValInfo.b); - testValInfo.fillExpectedTestValues(); - auto secondTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); - - return { firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed, firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed }; - } - - EmulatedFloat64TestOutput emulatedFloat64BothValuesInfTest(EF64Submitter& submitter) - { - smart_refctd_ptr semaphore = m_device->createSemaphore(0); - - EmulatedFloat64TestValuesInfo testValInfo; - const float32_t inf32 = std::numeric_limits::infinity(); - const float64_t inf64 = std::numeric_limits::infinity(); - testValInfo.a = emulated_float64_t::create(inf64); - testValInfo.b = emulated_float64_t::create(inf64); - testValInfo.constrTestValues = { - .int32 = 0, - .int64 = 0, - .uint32 = 0, - .uint64 = 0, - .float32 = inf32 - //.float64 = inf64 - }; - - testValInfo.fillExpectedTestValues(); - return performEmulatedFloat64Tests(testValInfo, submitter); - } - - EmulatedFloat64TestOutput emulatedFloat64BothValuesNegInfTest(EF64Submitter& submitter) - { - smart_refctd_ptr semaphore = m_device->createSemaphore(0); - - EmulatedFloat64TestValuesInfo testValInfo; - const float32_t inf32 = -std::numeric_limits::infinity(); - const float64_t inf64 = -std::numeric_limits::infinity(); - testValInfo.a = emulated_float64_t::create(inf64); - testValInfo.b = emulated_float64_t::create(inf64); - testValInfo.constrTestValues = { - .int32 = 0, - .int64 = 0, - .uint32 = 0, - .uint64 = 0, - .float32 = inf32 - //.float64 = inf64 - }; - - testValInfo.fillExpectedTestValues(); - return performEmulatedFloat64Tests(testValInfo, submitter); - } - - EmulatedFloat64TestOutput emulatedFloat64OneValIsInfOtherIsNegInfTest(EF64Submitter& submitter) - { - smart_refctd_ptr semaphore = m_device->createSemaphore(0); - - EmulatedFloat64TestValuesInfo testValInfo; - const float64_t inf64 = -std::numeric_limits::infinity(); - testValInfo.a = emulated_float64_t::create(inf64); - testValInfo.b = emulated_float64_t::create(inf64); - testValInfo.constrTestValues = { - .int32 = 0, - .int64 = 0, - .uint32 = 0, - .uint64 = 0, - .float32 = 0 - //.float64 = inf64 - }; - - testValInfo.fillExpectedTestValues(); - auto firstTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); - std::swap(testValInfo.a, testValInfo.b); - testValInfo.fillExpectedTestValues(); - auto secondTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); - - return { firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed, firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed }; - } - - // TODO: fix - EmulatedFloat64TestOutput emulatedFloat64BNaNTest(EF64Submitter& submitter) - { - EmulatedFloat64TestOutput output = { true, true }; - smart_refctd_ptr semaphore = m_device->createSemaphore(0); - - for (uint32_t i = 0u; i < EmulatedFloat64TestIterations; ++i) - { - std::random_device rd; - std::mt19937 mt(rd()); - - std::uniform_int_distribution i32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_int_distribution i64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_int_distribution u32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_int_distribution u64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); - std::uniform_real_distribution f32Distribution(-100000.0f, 100000.0f); - std::uniform_real_distribution f64Distribution(-100000.0, 100000.0); - - EmulatedFloat64TestValuesInfo testValInfo; - double aTmp = f64Distribution(mt); - double bTmp = std::numeric_limits::quiet_NaN(); - testValInfo.a.data = reinterpret_cast::storage_t&>(aTmp); - testValInfo.b.data = reinterpret_cast::storage_t&>(bTmp); - testValInfo.constrTestValues.int32 = i32Distribution(mt); - testValInfo.constrTestValues.int64 = i64Distribution(mt); - testValInfo.constrTestValues.uint32 = u32Distribution(mt); - testValInfo.constrTestValues.uint64 = u64Distribution(mt); - testValInfo.constrTestValues.float32 = f32Distribution(mt); - //testValInfo.constrTestValues.float64 = f64Distribution(mt); - - testValInfo.fillExpectedTestValues(); - auto singleTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); - - if (!singleTestOutput.cpuTestsSucceed) - output.cpuTestsSucceed = false; - if (!singleTestOutput.gpuTestsSucceed) - output.gpuTestsSucceed = false; - } - - return output; - } - - EmulatedFloat64TestOutput emulatedFloat64OneValIsInfTest(EF64Submitter& submitter) - { - auto getRandomFloat64 = []() - { - static std::random_device rd; - static std::mt19937 mt(rd()); - static std::uniform_real_distribution distribution(-100000.0, 100000.0); - - return distribution(mt); - }; - - auto getInfinity = []() - { - return std::numeric_limits::infinity(); - }; - - EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getInfinity); - EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getInfinity, getRandomFloat64); - - EmulatedFloat64TestOutput output; - output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; - output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; - return output; - } - - EmulatedFloat64TestOutput emulatedFloat64OneValIsNegInfTest(EF64Submitter& submitter) - { - auto getRandomFloat64 = []() - { - static std::random_device rd; - static std::mt19937 mt(rd()); - static std::uniform_real_distribution distribution(-100000.0, 100000.0); - + if (!shader) + base.logFail("Failed to load precompiled \"test\" shader!\n"); + + nbl::video::IGPUDescriptorSetLayout::SBinding bindings[1] = { + {.binding = 0, + .type = nbl::asset::IDescriptor::E_TYPE::ET_STORAGE_BUFFER, + .createFlags = IGPUDescriptorSetLayout::SBinding::E_CREATE_FLAGS::ECF_NONE, + .stageFlags = ShaderStage::ESS_COMPUTE, + .count = 1}}; + smart_refctd_ptr dsLayout = base.m_device->createDescriptorSetLayout(bindings); + if (!dsLayout) + base.logFail("Failed to create a Descriptor Layout!\n"); + + SPushConstantRange pushConstantRanges[] = { + {.stageFlags = ShaderStage::ESS_COMPUTE, + .offset = 0, + .size = sizeof(PushConstants)}}; + m_pplnLayout = base.m_device->createPipelineLayout(pushConstantRanges, smart_refctd_ptr(dsLayout)); + if (!m_pplnLayout) + base.logFail("Failed to create a Pipeline Layout!\n"); - return distribution(mt); - }; - - auto getNegInfinity = []() - { - return -std::numeric_limits::infinity(); - }; - - EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getNegInfinity); - EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getNegInfinity, getRandomFloat64); - - EmulatedFloat64TestOutput output; - output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; - output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; - return output; - } - - EmulatedFloat64TestOutput emulatedFloat64OneValIsZeroTest(EF64Submitter& submitter) - { - auto getRandomFloat64 = []() - { - static std::random_device rd; - static std::mt19937 mt(rd()); - static std::uniform_real_distribution distribution(-100000.0, 100000.0); - - return distribution(mt); - }; - - auto getZero = []() - { - return 0.0; - }; - - EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getZero); - EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getZero, getRandomFloat64); - - EmulatedFloat64TestOutput output; - output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; - output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; - return output; - } - - EmulatedFloat64TestOutput emulatedFloat64OneValIsNegZeroTest(EF64Submitter& submitter) - { - auto getRandomFloat64 = []() - { - static std::random_device rd; - static std::mt19937 mt(rd()); - static std::uniform_real_distribution distribution(-100000.0, 100000.0); - - return distribution(mt); - }; - - auto getNegZero = []() - { - return -0.0; - }; - - EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getNegZero); - EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getNegZero, getRandomFloat64); - - EmulatedFloat64TestOutput output; - output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; - output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; - return output; - } - - template - EmulatedFloat64TestOutput performEmulatedFloat64Tests(EmulatedFloat64TestValuesInfo& testValInfo, EF64Submitter& submitter) - { - emulated_float64_t a = testValInfo.a; - emulated_float64_t b = testValInfo.b; - - const TestValues cpuTestValues = { - .int32CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.int32).data, - .int64CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.int64).data, - .uint32CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.uint32).data, - .uint64CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.uint64).data, - .float32CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.float32).data, - .float64CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.float64).data, - .additionVal = (a + b).data, - .substractionVal = (a - b).data, - .multiplicationVal = (a * b).data, - .divisionVal = (a / b).data, - .lessOrEqualVal = a <= b, - .greaterOrEqualVal = a >= b, - .equalVal = a == b, - .notEqualVal = a != b, - .lessVal = a < b, - .greaterVal = a > b - }; - - EmulatedFloat64TestOutput output; - - // cpu validation - output.cpuTestsSucceed = compareEmulatedFloat64TestValues(testValInfo.expectedTestValues, cpuTestValues); - - // gpu validation - PushConstants pc; - pc.a = reinterpret_cast(a); - pc.b = reinterpret_cast(b); - pc.constrTestVals = testValInfo.constrTestValues; - - submitter.setPushConstants(pc); - auto gpuTestValues = submitter.submitGetGPUTestValues(); - - output.gpuTestsSucceed = compareEmulatedFloat64TestValues(testValInfo.expectedTestValues, gpuTestValues); - - return output; - } - - class EF64Benchmark final - { - public: - EF64Benchmark(CompatibilityTest& base) - { - m_device = base.m_device; - m_logger = base.m_logger; - m_api = base.m_api; - - // setting up pipeline in the constructor - m_queueFamily = base.getComputeQueue()->getFamilyIndex(); - m_cmdpool = base.m_device->createCommandPool(m_queueFamily, IGPUCommandPool::CREATE_FLAGS::RESET_COMMAND_BUFFER_BIT); - //core::smart_refctd_ptr* cmdBuffs[] = { &m_cmdbuf, &m_timestampBeforeCmdBuff, &m_timestampAfterCmdBuff }; - if (!m_cmdpool->createCommandBuffers(IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_cmdbuf)) - base.logFail("Failed to create Command Buffers!\n"); - if (!m_cmdpool->createCommandBuffers(IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_timestampBeforeCmdBuff)) - base.logFail("Failed to create Command Buffers!\n"); - if (!m_cmdpool->createCommandBuffers(IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_timestampAfterCmdBuff)) - base.logFail("Failed to create Command Buffers!\n"); - - // Load shaders, set up pipeline { - smart_refctd_ptr shader; - { - IAssetLoader::SAssetLoadParams lp = {}; - lp.logger = base.m_logger.get(); - lp.workingDirectory = "app_resources"; // virtual root - // this time we load a shader directly from a file - auto key = nbl::this_example::builtin::build::get_spirv_key<"benchmark">(m_device.get()); - auto assetBundle = base.m_assetMgr->getAsset(key.data(), lp); - const auto assets = assetBundle.getContents(); - if (assets.empty()) - { - base.logFail("Could not load shader!"); - assert(0); - } - - // It would be super weird if loading a shader from a file produced more than 1 asset - assert(assets.size() == 1); - shader = IAsset::castDown(assets[0]); - } - - if (!shader) - base.logFail("Failed to load precompiled \"benchmark\" shader!\n"); - - nbl::video::IGPUDescriptorSetLayout::SBinding bindings[1] = { - { - .binding = 0, - .type = nbl::asset::IDescriptor::E_TYPE::ET_STORAGE_BUFFER, - .createFlags = IGPUDescriptorSetLayout::SBinding::E_CREATE_FLAGS::ECF_NONE, - .stageFlags = ShaderStage::ESS_COMPUTE, - .count = 1 - } - }; - smart_refctd_ptr dsLayout = base.m_device->createDescriptorSetLayout(bindings); - if (!dsLayout) - base.logFail("Failed to create a Descriptor Layout!\n"); - - SPushConstantRange pushConstantRanges[] = { - { - .stageFlags = ShaderStage::ESS_COMPUTE, - .offset = 0, - .size = sizeof(BenchmarkPushConstants) - } - }; - m_pplnLayout = base.m_device->createPipelineLayout(pushConstantRanges, smart_refctd_ptr(dsLayout)); - if (!m_pplnLayout) - base.logFail("Failed to create a Pipeline Layout!\n"); - - { - IGPUComputePipeline::SCreationParams params = {}; - params.layout = m_pplnLayout.get(); - params.shader.entryPoint = "main"; - params.shader.shader = shader.get(); - if (base.m_device->getEnabledFeatures().pipelineExecutableInfo) - { - params.flags |= IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_STATISTICS; - params.flags |= IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_INTERNAL_REPRESENTATIONS; - } - if (!base.m_device->createComputePipelines(nullptr, { ¶ms,1 }, &m_pipeline)) - base.logFail("Failed to create pipelines (compile & link shaders)!\n"); - - if (base.m_device->getEnabledFeatures().pipelineExecutableInfo) - { - auto report = system::to_string(m_pipeline->getExecutableInfo()); - base.m_logger->log("EF64Benchmark Pipeline Executable Report:\n%s", ILogger::ELL_PERFORMANCE, report.c_str()); - } - } - - // Allocate the memory - { - static_assert(sizeof(float64_t) == sizeof(benchmark_emulated_float64_t)); - constexpr size_t BufferSize = BENCHMARK_WORKGROUP_COUNT * BENCHMARK_WORKGROUP_DIMENSION_SIZE_X * - BENCHMARK_WORKGROUP_DIMENSION_SIZE_Y * BENCHMARK_WORKGROUP_DIMENSION_SIZE_Z * sizeof(float64_t); - - nbl::video::IGPUBuffer::SCreationParams params = {}; - params.size = BufferSize; - params.usage = IGPUBuffer::EUF_STORAGE_BUFFER_BIT; - smart_refctd_ptr dummyBuff = base.m_device->createBuffer(std::move(params)); - if (!dummyBuff) - base.logFail("Failed to create a GPU Buffer of size %d!\n", params.size); - - dummyBuff->setObjectDebugName("benchmark buffer"); - - nbl::video::IDeviceMemoryBacked::SDeviceMemoryRequirements reqs = dummyBuff->getMemoryReqs(); - - m_allocation = base.m_device->allocate(reqs, dummyBuff.get(), nbl::video::IDeviceMemoryAllocation::EMAF_NONE); - if (!m_allocation.isValid()) - base.logFail("Failed to allocate Device Memory compatible with our GPU Buffer!\n"); - - assert(dummyBuff->getBoundMemory().memory == m_allocation.memory.get()); - smart_refctd_ptr pool = base.m_device->createDescriptorPoolForDSLayouts(IDescriptorPool::ECF_NONE, { &dsLayout.get(),1 }); - - m_ds = pool->createDescriptorSet(std::move(dsLayout)); - { - IGPUDescriptorSet::SDescriptorInfo info[1]; - info[0].desc = smart_refctd_ptr(dummyBuff); - info[0].info.buffer = { .offset = 0,.size = BufferSize }; - IGPUDescriptorSet::SWriteDescriptorSet writes[1] = { - {.dstSet = m_ds.get(),.binding = 0,.arrayElement = 0,.count = 1,.info = info} - }; - base.m_device->updateDescriptorSets(writes, {}); - } - } + IGPUComputePipeline::SCreationParams params = {}; + params.layout = m_pplnLayout.get(); + params.shader.entryPoint = "main"; + params.shader.shader = shader.get(); + if (base.m_device->getEnabledFeatures().pipelineExecutableInfo) + { + params.flags |= IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_STATISTICS; + params.flags |= IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_INTERNAL_REPRESENTATIONS; + } + if (!base.m_device->createComputePipelines(nullptr, {¶ms, 1}, &m_pipeline)) + base.logFail("Failed to create pipelines (compile & link shaders)!\n"); + + if (base.m_device->getEnabledFeatures().pipelineExecutableInfo) + { + auto report = system::to_string(m_pipeline->getExecutableInfo()); + base.m_logger->log("EF64Submitter Pipeline Executable Report:\n%s", ILogger::ELL_PERFORMANCE, report.c_str()); + } } - IQueryPool::SCreationParams queryPoolCreationParams{}; - queryPoolCreationParams.queryType = IQueryPool::TYPE::TIMESTAMP; - queryPoolCreationParams.queryCount = 2; - queryPoolCreationParams.pipelineStatisticsFlags = IQueryPool::PIPELINE_STATISTICS_FLAGS::NONE; - m_queryPool = m_device->createQueryPool(queryPoolCreationParams); - - m_computeQueue = m_device->getQueue(m_queueFamily, 0); - } - - void run() - { - m_logger->log("\n\nfloat64_t benchmark result:", ILogger::ELL_PERFORMANCE); - performBenchmark(EF64_BENCHMARK_MODE::NATIVE); - m_logger->log("emulated_float64_t benchmark, fast math enabled result:", ILogger::ELL_PERFORMANCE); - performBenchmark(EF64_BENCHMARK_MODE::EF64_FAST_MATH_ENABLED); - m_logger->log("emulated_float64_t benchmark, fast math disabled result:", ILogger::ELL_PERFORMANCE); - performBenchmark(EF64_BENCHMARK_MODE::EF64_FAST_MATH_DISABLED); - // every subgroup with even ID do calculations with the `emulated_float64_t` type, other subgroups do calculations with float64_t - m_logger->log("emulated_float64_t benchmark, subgroup divided work result:", ILogger::ELL_PERFORMANCE); - performBenchmark(EF64_BENCHMARK_MODE::SUBGROUP_DIVIDED_WORK); - // every item does calculations with both emulated and native types - m_logger->log("emulated_float64_t benchmark, interleaved result:", ILogger::ELL_PERFORMANCE); - performBenchmark(EF64_BENCHMARK_MODE::INTERLEAVED); - } - - private: - void performBenchmark(EF64_BENCHMARK_MODE mode) - { - m_device->waitIdle(); - - recordTimestampQueryCmdBuffers(); - - uint64_t semaphoreCounter = 0; - smart_refctd_ptr semaphore = m_device->createSemaphore(semaphoreCounter); - - IQueue::SSubmitInfo::SSemaphoreInfo signals[] = { {.semaphore = semaphore.get(), .value = 0u, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT} }; - IQueue::SSubmitInfo::SSemaphoreInfo waits[] = { {.semaphore = semaphore.get(), .value = 0u, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT } }; - - IQueue::SSubmitInfo beforeTimestapSubmitInfo[1] = {}; - const IQueue::SSubmitInfo::SCommandBufferInfo cmdbufsBegin[] = { {.cmdbuf = m_timestampBeforeCmdBuff.get()} }; - beforeTimestapSubmitInfo[0].commandBuffers = cmdbufsBegin; - beforeTimestapSubmitInfo[0].signalSemaphores = signals; - beforeTimestapSubmitInfo[0].waitSemaphores = waits; - - IQueue::SSubmitInfo afterTimestapSubmitInfo[1] = {}; - const IQueue::SSubmitInfo::SCommandBufferInfo cmdbufsEnd[] = { {.cmdbuf = m_timestampAfterCmdBuff.get()} }; - afterTimestapSubmitInfo[0].commandBuffers = cmdbufsEnd; - afterTimestapSubmitInfo[0].signalSemaphores = signals; - afterTimestapSubmitInfo[0].waitSemaphores = waits; - - IQueue::SSubmitInfo benchmarkSubmitInfos[1] = {}; - const IQueue::SSubmitInfo::SCommandBufferInfo cmdbufs[] = { {.cmdbuf = m_cmdbuf.get()} }; - benchmarkSubmitInfos[0].commandBuffers = cmdbufs; - benchmarkSubmitInfos[0].signalSemaphores = signals; - benchmarkSubmitInfos[0].waitSemaphores = waits; - - - m_pushConstants.benchmarkMode = mode; - recordCmdBuff(); - - // warmup runs - for (int i = 0; i < WarmupIterations; ++i) - { - if(i == 0) - m_api->startCapture(); - waits[0].value = semaphoreCounter; - signals[0].value = ++semaphoreCounter; - m_computeQueue->submit(benchmarkSubmitInfos); - if (i == 0) - m_api->endCapture(); - } - - waits[0].value = semaphoreCounter; - signals[0].value = ++semaphoreCounter; - m_computeQueue->submit(beforeTimestapSubmitInfo); - - // actual benchmark runs - for (int i = 0; i < Iterations; ++i) - { - waits[0].value = semaphoreCounter; - signals[0].value = ++semaphoreCounter; - m_computeQueue->submit(benchmarkSubmitInfos); - } - - waits[0].value = semaphoreCounter; - signals[0].value = ++semaphoreCounter; - m_computeQueue->submit(afterTimestapSubmitInfo); - - m_device->waitIdle(); - - const uint64_t nativeBenchmarkTimeElapsedNanoseconds = calcTimeElapsed(); - const float nativeBenchmarkTimeElapsedSeconds = double(nativeBenchmarkTimeElapsedNanoseconds) / 1000000000.0; - - m_logger->log("%llu ns, %f s", ILogger::ELL_PERFORMANCE, nativeBenchmarkTimeElapsedNanoseconds, nativeBenchmarkTimeElapsedSeconds); - } - - void recordCmdBuff() - { - m_cmdbuf->begin(IGPUCommandBuffer::USAGE::SIMULTANEOUS_USE_BIT); - m_cmdbuf->beginDebugMarker("emulated_float64_t compute dispatch", vectorSIMDf(0, 1, 0, 1)); - m_cmdbuf->bindComputePipeline(m_pipeline.get()); - m_cmdbuf->bindDescriptorSets(nbl::asset::EPBP_COMPUTE, m_pplnLayout.get(), 0, 1, &m_ds.get()); - m_cmdbuf->pushConstants(m_pplnLayout.get(), IShader::E_SHADER_STAGE::ESS_COMPUTE, 0, sizeof(BenchmarkPushConstants), &m_pushConstants); - m_cmdbuf->dispatch(BENCHMARK_WORKGROUP_COUNT, 1, 1); - m_cmdbuf->endDebugMarker(); - m_cmdbuf->end(); - } - - void recordTimestampQueryCmdBuffers() - { - static bool firstInvocation = true; - - if (!firstInvocation) + // Allocate the memory { - m_timestampBeforeCmdBuff->reset(IGPUCommandBuffer::RESET_FLAGS::NONE); - m_timestampBeforeCmdBuff->reset(IGPUCommandBuffer::RESET_FLAGS::NONE); + constexpr size_t BufferSize = sizeof(TestValues); + + nbl::video::IGPUBuffer::SCreationParams params = {}; + params.size = BufferSize; + params.usage = IGPUBuffer::EUF_STORAGE_BUFFER_BIT; + smart_refctd_ptr outputBuff = base.m_device->createBuffer(std::move(params)); + if (!outputBuff) + base.logFail("Failed to create a GPU Buffer of size %d!\n", params.size); + + outputBuff->setObjectDebugName("emulated_float64_t output buffer"); + + nbl::video::IDeviceMemoryBacked::SDeviceMemoryRequirements reqs = outputBuff->getMemoryReqs(); + reqs.memoryTypeBits &= base.m_physicalDevice->getHostVisibleMemoryTypeBits(); + + m_allocation = base.m_device->allocate(reqs, outputBuff.get(), nbl::video::IDeviceMemoryAllocation::EMAF_NONE); + if (!m_allocation.isValid()) + base.logFail("Failed to allocate Device Memory compatible with our GPU Buffer!\n"); + + assert(outputBuff->getBoundMemory().memory == m_allocation.memory.get()); + smart_refctd_ptr pool = base.m_device->createDescriptorPoolForDSLayouts(IDescriptorPool::ECF_NONE, {&dsLayout.get(), 1}); + + m_ds = pool->createDescriptorSet(std::move(dsLayout)); + { + IGPUDescriptorSet::SDescriptorInfo info[1]; + info[0].desc = smart_refctd_ptr(outputBuff); + info[0].info.buffer = {.offset = 0, .size = BufferSize}; + IGPUDescriptorSet::SWriteDescriptorSet writes[1] = { + {.dstSet = m_ds.get(), .binding = 0, .arrayElement = 0, .count = 1, .info = info}}; + base.m_device->updateDescriptorSets(writes, {}); + } } - m_timestampBeforeCmdBuff->begin(IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); - m_timestampBeforeCmdBuff->resetQueryPool(m_queryPool.get(), 0, 2); - m_timestampBeforeCmdBuff->writeTimestamp(PIPELINE_STAGE_FLAGS::NONE, m_queryPool.get(), 0); - m_timestampBeforeCmdBuff->end(); - - m_timestampAfterCmdBuff->begin(IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); - m_timestampAfterCmdBuff->writeTimestamp(PIPELINE_STAGE_FLAGS::NONE, m_queryPool.get(), 1); - m_timestampAfterCmdBuff->end(); - - firstInvocation = false; - } - - uint64_t calcTimeElapsed() - { - uint64_t timestamps[2]; - const core::bitflag flags = core::bitflag(IQueryPool::RESULTS_FLAGS::_64_BIT) | core::bitflag(IQueryPool::RESULTS_FLAGS::WAIT_BIT); - m_device->getQueryPoolResults(m_queryPool.get(), 0, 2, ×tamps, sizeof(uint64_t), flags); - return timestamps[1] - timestamps[0]; - } - - private: - core::smart_refctd_ptr m_api; - smart_refctd_ptr m_device; - smart_refctd_ptr m_logger; - - nbl::video::IDeviceMemoryAllocator::SAllocation m_allocation = {}; - smart_refctd_ptr m_cmdpool = nullptr; - smart_refctd_ptr m_cmdbuf = nullptr; - smart_refctd_ptr m_ds = nullptr; - smart_refctd_ptr m_pplnLayout = nullptr; - BenchmarkPushConstants m_pushConstants; - smart_refctd_ptr m_pipeline; - - smart_refctd_ptr m_timestampBeforeCmdBuff = nullptr; - smart_refctd_ptr m_timestampAfterCmdBuff = nullptr; - smart_refctd_ptr m_queryPool = nullptr; - - uint32_t m_queueFamily; - IQueue* m_computeQueue; - static constexpr int WarmupIterations = 1000; - static constexpr int Iterations = 1000; - using benchmark_emulated_float64_t = emulated_float64_t; - }; - - template - inline bool logFail(const char* msg, Args&&... args) - { - m_logger->log(msg, ILogger::ELL_ERROR, std::forward(args)...); - return false; - } - - std::ofstream m_logFile; + if (!m_allocation.memory->map({0ull, m_allocation.memory->getAllocationSize()}, IDeviceMemoryAllocation::EMCAF_READ)) + base.logFail("Failed to map the Device Memory!\n"); + } + + // if the mapping is not coherent the range needs to be invalidated to pull in new data for the CPU's caches + const ILogicalDevice::MappedMemoryRange memoryRange(m_allocation.memory.get(), 0ull, m_allocation.memory->getAllocationSize()); + if (!m_allocation.memory->getMemoryPropertyFlags().hasFlags(IDeviceMemoryAllocation::EMPF_HOST_COHERENT_BIT)) + base.m_device->invalidateMappedMemoryRanges(1, &memoryRange); + + assert(memoryRange.valid() && memoryRange.length >= sizeof(TestValues)); + + m_queue = m_base.m_device->getQueue(m_queueFamily, 0); + } + + ~EF64Submitter() + { + m_allocation.memory->unmap(); + } + + void setPushConstants(PushConstants& pc) + { + m_pushConstants = pc; + } + + TestValues submitGetGPUTestValues() + { + // record command buffer + m_cmdbuf->reset(IGPUCommandBuffer::RESET_FLAGS::NONE); + m_cmdbuf->begin(IGPUCommandBuffer::USAGE::NONE); + m_cmdbuf->beginDebugMarker("emulated_float64_t compute dispatch", vectorSIMDf(0, 1, 0, 1)); + m_cmdbuf->bindComputePipeline(m_pipeline.get()); + m_cmdbuf->bindDescriptorSets(nbl::asset::EPBP_COMPUTE, m_pplnLayout.get(), 0, 1, &m_ds.get()); + m_cmdbuf->pushConstants(m_pplnLayout.get(), IShader::E_SHADER_STAGE::ESS_COMPUTE, 0, sizeof(PushConstants), &m_pushConstants); + m_cmdbuf->dispatch(WORKGROUP_SIZE, 1, 1); + m_cmdbuf->endDebugMarker(); + m_cmdbuf->end(); + + IQueue::SSubmitInfo submitInfos[1] = {}; + const IQueue::SSubmitInfo::SCommandBufferInfo cmdbufs[] = {{.cmdbuf = m_cmdbuf.get()}}; + submitInfos[0].commandBuffers = cmdbufs; + const IQueue::SSubmitInfo::SSemaphoreInfo signals[] = {{.semaphore = m_semaphore.get(), .value = ++m_semaphoreCounter, .stageMask = asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT}}; + submitInfos[0].signalSemaphores = signals; + + m_base.m_api->startCapture(); + m_queue->submit(submitInfos); + m_base.m_api->endCapture(); + + m_base.m_device->waitIdle(); + TestValues output; + std::memcpy(&output, static_cast*>(m_allocation.memory->getMappedPointer()), sizeof(TestValues)); + m_base.m_device->waitIdle(); + + return output; + } + + private: + uint32_t m_queueFamily; + nbl::video::IDeviceMemoryAllocator::SAllocation m_allocation = {}; + smart_refctd_ptr m_cmdbuf = nullptr; + smart_refctd_ptr m_cmdpool = nullptr; + smart_refctd_ptr m_ds = nullptr; + smart_refctd_ptr m_pplnLayout = nullptr; + PushConstants m_pushConstants; + CompatibilityTest& m_base; + smart_refctd_ptr m_pipeline; + smart_refctd_ptr m_semaphore; + IQueue* m_queue; + uint64_t m_semaphoreCounter; + }; + + void emulated_float64_tests() + { + EF64Submitter submitter(*this); + + auto printTestOutput = [this](const std::string& functionName, const EmulatedFloat64TestOutput& testResult) + { + std::cout << functionName << ": " << std::endl; + + if (!testResult.cpuTestsSucceed) + logFail("Incorrect CPU determinated values!"); + else + m_logger->log("Correct CPU determinated values!", ILogger::ELL_PERFORMANCE); + + if (!testResult.gpuTestsSucceed) + logFail("Incorrect GPU determinated values!"); + else + m_logger->log("Correct GPU determinated values!", ILogger::ELL_PERFORMANCE); + }; + + m_logFile.open("EmulatedFloatTestLog.txt", std::ios::out | std::ios::trunc); + if (!m_logFile.is_open()) + m_logger->log("Failed to open log file!", system::ILogger::ELL_ERROR); + + printTestOutput("emulatedFloat64RandomValuesTest", emulatedFloat64RandomValuesTest(submitter)); + printTestOutput("emulatedFloat64RandomValuesTestContrastingExponents", emulatedFloat64RandomValuesTestContrastingExponents(submitter)); + printTestOutput("emulatedFloat64NegAndPosZeroTest", emulatedFloat64NegAndPosZeroTest(submitter)); + printTestOutput("emulatedFloat64BothValuesInfTest", emulatedFloat64BothValuesInfTest(submitter)); + printTestOutput("emulatedFloat64BothValuesNegInfTest", emulatedFloat64BothValuesNegInfTest(submitter)); + printTestOutput("emulatedFloat64OneValIsInfOtherIsNegInfTest", emulatedFloat64OneValIsInfOtherIsNegInfTest(submitter)); + printTestOutput("emulatedFloat64OneValIsInfTest", emulatedFloat64OneValIsInfTest(submitter)); + printTestOutput("emulatedFloat64OneValIsNegInfTest", emulatedFloat64OneValIsNegInfTest(submitter)); + if (false) // doesn't work for some reason + fast math is enabled by default + printTestOutput("emulatedFloat64BNaNTest", emulatedFloat64BNaNTest(submitter)); + printTestOutput("emulatedFloat64BInfTest", emulatedFloat64OneValIsZeroTest(submitter)); + printTestOutput("emulatedFloat64BNegInfTest", emulatedFloat64OneValIsNegZeroTest(submitter)); + + m_logFile.close(); + } + + template + struct EmulatedFloat64TestValuesInfo + { + emulated_float64_t a; + emulated_float64_t b; + ConstructorTestValues constrTestValues; + TestValues expectedTestValues; + + void fillExpectedTestValues() + { + double aAsDouble = reinterpret_cast(a); + double bAsDouble = reinterpret_cast(b); + + expectedTestValues.a = a.data; + expectedTestValues.b = b.data; + + expectedTestValues.int32CreateVal = bit_cast(double(constrTestValues.int32)); + expectedTestValues.int64CreateVal = bit_cast(double(constrTestValues.int64)); + expectedTestValues.uint32CreateVal = bit_cast(double(constrTestValues.uint32)); + expectedTestValues.uint64CreateVal = bit_cast(double(constrTestValues.uint64)); + expectedTestValues.float32CreateVal = bit_cast(double(constrTestValues.float32)); + expectedTestValues.float64CreateVal = bit_cast(constrTestValues.float64); + expectedTestValues.additionVal = emulated_float64_t::create(aAsDouble + bAsDouble).data; + expectedTestValues.substractionVal = emulated_float64_t::create(aAsDouble - bAsDouble).data; + expectedTestValues.multiplicationVal = emulated_float64_t::create(aAsDouble * bAsDouble).data; + expectedTestValues.divisionVal = emulated_float64_t::create(aAsDouble / bAsDouble).data; + expectedTestValues.lessOrEqualVal = aAsDouble <= bAsDouble; + expectedTestValues.greaterOrEqualVal = aAsDouble >= bAsDouble; + expectedTestValues.equalVal = aAsDouble == bAsDouble; + expectedTestValues.notEqualVal = aAsDouble != bAsDouble; + expectedTestValues.lessVal = aAsDouble < bAsDouble; + expectedTestValues.greaterVal = aAsDouble > bAsDouble; + } + }; + + struct EmulatedFloat64TestOutput + { + bool cpuTestsSucceed; + bool gpuTestsSucceed; + }; + + EmulatedFloat64TestOutput emulatedFloat64LoopedTests_impl(EF64Submitter& submitter, + const uint32_t iterations, + const std::function& determineValueA, + const std::function& determineValueB) + { + EmulatedFloat64TestOutput output = {true, true}; + + std::uniform_int_distribution i32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_int_distribution i64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_int_distribution u32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_int_distribution u64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_real_distribution fDistribution(-100000.0, 100000.0); + + std::random_device rd; + std::mt19937 mt(rd()); + + for (uint32_t i = 0u; i < iterations; ++i) + { + // generate random test values + EmulatedFloat64TestValuesInfo testValInfo; + double aTmp = determineValueA(); + double bTmp = determineValueB(); + testValInfo.a.data = reinterpret_cast::storage_t&>(aTmp); + testValInfo.b.data = reinterpret_cast::storage_t&>(bTmp); + testValInfo.constrTestValues.int32 = i32Distribution(mt); + testValInfo.constrTestValues.int64 = i64Distribution(mt); + testValInfo.constrTestValues.uint32 = u32Distribution(mt); + testValInfo.constrTestValues.uint64 = u64Distribution(mt); + testValInfo.constrTestValues.float32 = fDistribution(mt); + testValInfo.constrTestValues.float64 = fDistribution(mt); + + testValInfo.fillExpectedTestValues(); + auto singleTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); + + if (!singleTestOutput.cpuTestsSucceed) + output.cpuTestsSucceed = false; + if (!singleTestOutput.gpuTestsSucceed) + output.gpuTestsSucceed = false; + } + + return output; + } + + EmulatedFloat64TestOutput emulatedFloat64RandomValuesTest(EF64Submitter& submitter) + { + auto getRandomFloat64 = []() + { + static std::random_device rd; + static std::mt19937 mt(rd()); + static std::uniform_real_distribution distribution(-100000.0, 100000.0); + + + return distribution(mt); + }; + + return emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations, getRandomFloat64, getRandomFloat64); + } + + EmulatedFloat64TestOutput emulatedFloat64RandomValuesTestContrastingExponents(EF64Submitter& submitter) + { + auto getRandomSmallFloat64 = []() + { + static std::random_device rd; + static std::mt19937 mt(rd()); + static std::uniform_real_distribution distribution(-0.01, 0.01); + + return distribution(mt); + }; + + auto getRandomLargeFloat64 = []() + { + static std::random_device rd; + static std::mt19937 mt(rd()); + static std::uniform_real_distribution distribution(1000000000.0, 2000000000.0); + static std::uniform_int_distribution coinFlipDistribution(0, 1); + + double output = distribution(mt); + if (coinFlipDistribution(mt)) + output = -output; + + return output; + }; + + EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomSmallFloat64, getRandomLargeFloat64); + EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomLargeFloat64, getRandomSmallFloat64); + + EmulatedFloat64TestOutput output; + output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; + output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; + return output; + } + + EmulatedFloat64TestOutput emulatedFloat64BothValuesNaNTest(EF64Submitter& submitter) + { + smart_refctd_ptr semaphore = m_device->createSemaphore(0); + + EmulatedFloat64TestValuesInfo testValInfo; + const float32_t nan32 = std::numeric_limits::quiet_NaN(); + const float64_t nan64 = std::numeric_limits::quiet_NaN(); + testValInfo.a = emulated_float64_t::create(nan64); + testValInfo.b = emulated_float64_t::create(nan64); + testValInfo.constrTestValues = { + .int32 = std::bit_cast(nan32), + .int64 = std::bit_cast(nan64), + .uint32 = std::bit_cast(nan32), + .uint64 = std::bit_cast(nan64), + .float32 = nan32 + //.float64 = nan64 + }; + + testValInfo.fillExpectedTestValues(); + return performEmulatedFloat64Tests(testValInfo, submitter); + } + + EmulatedFloat64TestOutput emulatedFloat64NegAndPosZeroTest(EF64Submitter& submitter) + { + smart_refctd_ptr semaphore = m_device->createSemaphore(0); + + EmulatedFloat64TestValuesInfo testValInfo; + testValInfo.a = emulated_float64_t::create(ieee754::traits::signMask); + testValInfo.b = emulated_float64_t::create(std::bit_cast(0.0)); + testValInfo.constrTestValues = { + .int32 = 0, + .int64 = 0, + .uint32 = 0, + .uint64 = 0, + .float32 = 0}; + + testValInfo.fillExpectedTestValues(); + auto firstTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); + std::swap(testValInfo.a, testValInfo.b); + testValInfo.fillExpectedTestValues(); + auto secondTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); + + return {firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed, firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed}; + } + + EmulatedFloat64TestOutput emulatedFloat64BothValuesInfTest(EF64Submitter& submitter) + { + smart_refctd_ptr semaphore = m_device->createSemaphore(0); + + EmulatedFloat64TestValuesInfo testValInfo; + const float32_t inf32 = std::numeric_limits::infinity(); + const float64_t inf64 = std::numeric_limits::infinity(); + testValInfo.a = emulated_float64_t::create(inf64); + testValInfo.b = emulated_float64_t::create(inf64); + testValInfo.constrTestValues = { + .int32 = 0, + .int64 = 0, + .uint32 = 0, + .uint64 = 0, + .float32 = inf32 + //.float64 = inf64 + }; + + testValInfo.fillExpectedTestValues(); + return performEmulatedFloat64Tests(testValInfo, submitter); + } + + EmulatedFloat64TestOutput emulatedFloat64BothValuesNegInfTest(EF64Submitter& submitter) + { + smart_refctd_ptr semaphore = m_device->createSemaphore(0); + + EmulatedFloat64TestValuesInfo testValInfo; + const float32_t inf32 = -std::numeric_limits::infinity(); + const float64_t inf64 = -std::numeric_limits::infinity(); + testValInfo.a = emulated_float64_t::create(inf64); + testValInfo.b = emulated_float64_t::create(inf64); + testValInfo.constrTestValues = { + .int32 = 0, + .int64 = 0, + .uint32 = 0, + .uint64 = 0, + .float32 = inf32 + //.float64 = inf64 + }; + + testValInfo.fillExpectedTestValues(); + return performEmulatedFloat64Tests(testValInfo, submitter); + } + + EmulatedFloat64TestOutput emulatedFloat64OneValIsInfOtherIsNegInfTest(EF64Submitter& submitter) + { + smart_refctd_ptr semaphore = m_device->createSemaphore(0); + + EmulatedFloat64TestValuesInfo testValInfo; + const float64_t inf64 = -std::numeric_limits::infinity(); + testValInfo.a = emulated_float64_t::create(inf64); + testValInfo.b = emulated_float64_t::create(inf64); + testValInfo.constrTestValues = { + .int32 = 0, + .int64 = 0, + .uint32 = 0, + .uint64 = 0, + .float32 = 0 + //.float64 = inf64 + }; + + testValInfo.fillExpectedTestValues(); + auto firstTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); + std::swap(testValInfo.a, testValInfo.b); + testValInfo.fillExpectedTestValues(); + auto secondTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); + + return {firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed, firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed}; + } + + // TODO: fix + EmulatedFloat64TestOutput emulatedFloat64BNaNTest(EF64Submitter& submitter) + { + EmulatedFloat64TestOutput output = {true, true}; + smart_refctd_ptr semaphore = m_device->createSemaphore(0); + + for (uint32_t i = 0u; i < EmulatedFloat64TestIterations; ++i) + { + std::random_device rd; + std::mt19937 mt(rd()); + + std::uniform_int_distribution i32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_int_distribution i64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_int_distribution u32Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_int_distribution u64Distribution(-std::numeric_limits::max(), std::numeric_limits::max()); + std::uniform_real_distribution f32Distribution(-100000.0f, 100000.0f); + std::uniform_real_distribution f64Distribution(-100000.0, 100000.0); + + EmulatedFloat64TestValuesInfo testValInfo; + double aTmp = f64Distribution(mt); + double bTmp = std::numeric_limits::quiet_NaN(); + testValInfo.a.data = reinterpret_cast::storage_t&>(aTmp); + testValInfo.b.data = reinterpret_cast::storage_t&>(bTmp); + testValInfo.constrTestValues.int32 = i32Distribution(mt); + testValInfo.constrTestValues.int64 = i64Distribution(mt); + testValInfo.constrTestValues.uint32 = u32Distribution(mt); + testValInfo.constrTestValues.uint64 = u64Distribution(mt); + testValInfo.constrTestValues.float32 = f32Distribution(mt); + //testValInfo.constrTestValues.float64 = f64Distribution(mt); + + testValInfo.fillExpectedTestValues(); + auto singleTestOutput = performEmulatedFloat64Tests(testValInfo, submitter); + + if (!singleTestOutput.cpuTestsSucceed) + output.cpuTestsSucceed = false; + if (!singleTestOutput.gpuTestsSucceed) + output.gpuTestsSucceed = false; + } + + return output; + } + + EmulatedFloat64TestOutput emulatedFloat64OneValIsInfTest(EF64Submitter& submitter) + { + auto getRandomFloat64 = []() + { + static std::random_device rd; + static std::mt19937 mt(rd()); + static std::uniform_real_distribution distribution(-100000.0, 100000.0); + + return distribution(mt); + }; + + auto getInfinity = []() + { + return std::numeric_limits::infinity(); + }; + + EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getInfinity); + EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getInfinity, getRandomFloat64); + + EmulatedFloat64TestOutput output; + output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; + output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; + return output; + } + + EmulatedFloat64TestOutput emulatedFloat64OneValIsNegInfTest(EF64Submitter& submitter) + { + auto getRandomFloat64 = []() + { + static std::random_device rd; + static std::mt19937 mt(rd()); + static std::uniform_real_distribution distribution(-100000.0, 100000.0); + + + return distribution(mt); + }; + + auto getNegInfinity = []() + { + return -std::numeric_limits::infinity(); + }; + + EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getNegInfinity); + EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getNegInfinity, getRandomFloat64); + + EmulatedFloat64TestOutput output; + output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; + output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; + return output; + } + + EmulatedFloat64TestOutput emulatedFloat64OneValIsZeroTest(EF64Submitter& submitter) + { + auto getRandomFloat64 = []() + { + static std::random_device rd; + static std::mt19937 mt(rd()); + static std::uniform_real_distribution distribution(-100000.0, 100000.0); + + return distribution(mt); + }; + + auto getZero = []() + { + return 0.0; + }; + + EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getZero); + EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getZero, getRandomFloat64); + + EmulatedFloat64TestOutput output; + output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; + output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; + return output; + } + + EmulatedFloat64TestOutput emulatedFloat64OneValIsNegZeroTest(EF64Submitter& submitter) + { + auto getRandomFloat64 = []() + { + static std::random_device rd; + static std::mt19937 mt(rd()); + static std::uniform_real_distribution distribution(-100000.0, 100000.0); + + return distribution(mt); + }; + + auto getNegZero = []() + { + return -0.0; + }; + + EmulatedFloat64TestOutput firstTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getRandomFloat64, getNegZero); + EmulatedFloat64TestOutput secondTestOutput = emulatedFloat64LoopedTests_impl(submitter, EmulatedFloat64TestIterations / 2, getNegZero, getRandomFloat64); + + EmulatedFloat64TestOutput output; + output.cpuTestsSucceed = firstTestOutput.cpuTestsSucceed && secondTestOutput.cpuTestsSucceed; + output.gpuTestsSucceed = firstTestOutput.gpuTestsSucceed && secondTestOutput.gpuTestsSucceed; + return output; + } + + template + EmulatedFloat64TestOutput performEmulatedFloat64Tests(EmulatedFloat64TestValuesInfo& testValInfo, EF64Submitter& submitter) + { + emulated_float64_t a = testValInfo.a; + emulated_float64_t b = testValInfo.b; + + const TestValues cpuTestValues = { + .int32CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.int32).data, + .int64CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.int64).data, + .uint32CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.uint32).data, + .uint64CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.uint64).data, + .float32CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.float32).data, + .float64CreateVal = emulated_float64_t::create(testValInfo.constrTestValues.float64).data, + .additionVal = (a + b).data, + .substractionVal = (a - b).data, + .multiplicationVal = (a * b).data, + .divisionVal = (a / b).data, + .lessOrEqualVal = a <= b, + .greaterOrEqualVal = a >= b, + .equalVal = a == b, + .notEqualVal = a != b, + .lessVal = a + b}; + + EmulatedFloat64TestOutput output; + + // cpu validation + output.cpuTestsSucceed = compareEmulatedFloat64TestValues(testValInfo.expectedTestValues, cpuTestValues); + + // gpu validation + PushConstants pc; + pc.a = reinterpret_cast(a); + pc.b = reinterpret_cast(b); + pc.constrTestVals = testValInfo.constrTestValues; + + submitter.setPushConstants(pc); + auto gpuTestValues = submitter.submitGetGPUTestValues(); + + output.gpuTestsSucceed = compareEmulatedFloat64TestValues(testValInfo.expectedTestValues, gpuTestValues); + + return output; + } + + void runEF64Benchmarks() + { + constexpr uint32_t WarmupDispatches = 1000; + constexpr uint64_t TargetBudgetMs = 400; // ~400ms per row + + Aggregator agg(m_logger, m_device, m_physicalDevice, getComputeQueue()->getFamilyIndex()); + agg.applyCli({ + .argv = this->argv, + .defaultOutputPath = "EF64Bench.json", + .appName = "64_EmulatedFloatTest", + }); + + const auto shaderKey = nbl::this_example::builtin::build::get_spirv_key<"benchmark">(m_device.get()); + auto shaderVariant = GPUBenchmarkHelper::ShaderVariant::Precompiled(shaderKey); + + // One bench instance per mode -> one report row per mode. std::array + // gives stack-allocated, pointer-stable storage; no parallel + // benchPtrs vector needed since the aggregator iterates the span + // directly. + constexpr std::pair kModes[] = { + {EF64_BENCHMARK_MODE::NATIVE, "native"}, + {EF64_BENCHMARK_MODE::EF64_FAST_MATH_ENABLED, "emulated, fast-math"}, + {EF64_BENCHMARK_MODE::EF64_FAST_MATH_DISABLED, "emulated, strict"}, + {EF64_BENCHMARK_MODE::SUBGROUP_DIVIDED_WORK, "subgroup-divided"}, + {EF64_BENCHMARK_MODE::INTERLEAVED, "interleaved"}, + }; + constexpr size_t N = std::size(kModes); + std::vector benches; + benches.reserve(N); + for (size_t i = 0; i < N; ++i) + { + const auto& [mode, leaf] = kModes[i]; + benches.emplace_back(agg, CEF64Benchmark::SetupData{ + .assetMgr = m_assetMgr, + .name = {"EF64", leaf}, + .mode = mode, + .variant = shaderVariant, + .warmupDispatches = WarmupDispatches, + .targetBudgetMs = TargetBudgetMs, + }); + } + + const RunContext ctx = { + .shape = CEF64Benchmark::shape(), + .targetBudgetMs = TargetBudgetMs, + .sectionLabel = CEF64Benchmark::kSectionLabel, + }; + agg.runSessionAndReport(Aggregator::makeSpan(benches, ctx)); + } + + + template + inline bool logFail(const char* msg, Args&&... args) + { + m_logger->log(msg, ILogger::ELL_ERROR, std::forward(args)...); + return false; + } + + std::ofstream m_logFile; }; -NBL_MAIN_FUNC(CompatibilityTest) \ No newline at end of file +NBL_MAIN_FUNC(CompatibilityTest) diff --git a/common/include/nbl/examples/Benchmark/BenchmarkCli.h b/common/include/nbl/examples/Benchmark/BenchmarkCli.h new file mode 100644 index 000000000..abb0912da --- /dev/null +++ b/common/include/nbl/examples/Benchmark/BenchmarkCli.h @@ -0,0 +1,125 @@ +// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O. +// This file is part of the "Nabla Engine". +// For conditions of distribution and use, see copyright notice in nabla.h + +#ifndef _NBL_COMMON_BENCHMARK_CLI_INCLUDED_ +#define _NBL_COMMON_BENCHMARK_CLI_INCLUDED_ + +#include +#include "nbl/examples/Benchmark/BenchmarkTypes.h" + +#include +#include +#include +#include +#include +#include +#include +#include + +namespace benchmark_cli +{ + +struct ParsedArgs +{ + std::string outputPath; + bool noBaseline = false; + bool noColor = false; + bool helpRequested = false; + std::vector> baselines; // (label, path) + nbl::core::vector> focus; + // Median-of-K window count used for focused rows (see + // IBenchmark::samplesForCurrentRow). Default 3 trades 3 * targetBudgetMs + // wall time for jitter-robust comparisons. + uint32_t focusSamples = 3; +}; + +// Pure: parse argv into a ParsedArgs. Unknown flags are silently ignored; +// the caller decides what to do on help / no-baseline / per-load failure. +inline ParsedArgs parseArgs(std::span argv, std::string defaultOutputPath) +{ + ParsedArgs out; + out.outputPath = std::move(defaultOutputPath); + + for (size_t i = 1; i < argv.size(); ++i) + { + if (argv[i] == "--output" && i + 1 < argv.size()) + out.outputPath = argv[++i]; + else if (argv[i] == "--no-baseline") + out.noBaseline = true; + else if (argv[i] == "--no-color") + out.noColor = true; + else if (argv[i] == "--baseline" && i + 1 < argv.size()) + { + const std::string& spec = argv[++i]; + const auto eq = spec.find('='); + std::string label, path; + if (eq == std::string::npos) + { + path = spec; + const auto stem = std::filesystem::path(path).stem().string(); + label = stem.empty() ? std::string("baseline") : stem; + } + else + { + label = spec.substr(0, eq); + path = spec.substr(eq + 1); + } + out.baselines.emplace_back(std::move(label), std::move(path)); + } + else if (argv[i] == "--focus" && i + 1 < argv.size()) + { + out.focus.push_back(splitFocusSpec(argv[++i])); + } + else if (argv[i] == "--focus-samples" && i + 1 < argv.size()) + { + // Clamp to [1, 32]: 1 disables the median+outlier path, 32 is well past + // the point of diminishing returns (variance of the trimmed mean drops + // ~1/sqrt(K)). from_chars instead of stol to stay no-exceptions per + // Nabla style; malformed input leaves the default in place. + const std::string& s = argv[++i]; + long v = 0; + const auto [_, ec] = std::from_chars(s.data(), s.data() + s.size(), v); + if (ec == std::errc{}) + out.focusSamples = uint32_t(std::clamp(v, 1, 32)); + } + else if (argv[i] == "--help" || argv[i] == "-h") + { + out.helpRequested = true; + } + } + return out; +} + +inline void printHelp(nbl::system::ILogger* logger, std::string_view appName, std::string_view defaultOutputPath) +{ + benchLogFmt(logger, nbl::system::ILogger::ELL_INFO, + "{} CLI:\n" + " --output PATH write this run's report to PATH (default: {})\n" + " --baseline [LABEL=]PATH load PATH as a baseline; LABEL becomes the column header ('vs LABEL').\n" + " repeatable. If LABEL= is omitted, the file's stem is used\n" + " (e.g. main.json -> 'main'). '=' is used instead of ':' so Windows\n" + " drive letters in paths don't collide with the separator.\n" + " --no-baseline skip the default auto-load of the output path\n" + " --no-color disable ANSI color in the live table (also honored: NO_COLOR=1 env var)\n" + " --focus NAME print a focused baseline-comparison table for NAME before the run.\n" + " NAME is the hierarchical name with '>' between segments (whitespace\n" + " around '>' is optional). Repeatable; one row per --focus. The first\n" + " loaded baseline is the reference for inline deltas in this table.\n" + " Example: --focus \"Linear > Linear > 1:1\"\n" + " --focus-samples N run each focused row N times (median + outlier rejection) for\n" + " jitter-robust comparisons. Default 3; clamped to [1, 32]. N=1\n" + " matches the rest-phase single-shot path. Wall time per focused\n" + " row scales linearly with N.\n" + " --help, -h print this help\n" + "\n" + "Default behaviour: with no flags, the prior run's output (if present) is loaded as the single\n" + " 'baseline', and a fresh one is written at the end; iterate-and-compare with no flags needed.\n" + "\n" + "Failed loads (missing/corrupt file) log a warning and continue; the corresponding column reads 'n/a'.", + appName, defaultOutputPath); +} + +} + +#endif diff --git a/common/include/nbl/examples/Benchmark/BenchmarkConsole.h b/common/include/nbl/examples/Benchmark/BenchmarkConsole.h new file mode 100644 index 000000000..e857c36d4 --- /dev/null +++ b/common/include/nbl/examples/Benchmark/BenchmarkConsole.h @@ -0,0 +1,526 @@ +// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O. +// This file is part of the "Nabla Engine". +// For conditions of distribution and use, see copyright notice in nabla.h + +#ifndef _NBL_COMMON_BENCHMARK_CONSOLE_INCLUDED_ +#define _NBL_COMMON_BENCHMARK_CONSOLE_INCLUDED_ + +#include +#include "nbl/examples/Benchmark/BenchmarkTypes.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +// Methods templated on the baselines range must expose `.label` and `.rowsByName`. +class BenchmarkConsole +{ + public: + BenchmarkConsole() + { + // https://no-color.org + if (const char* nc = std::getenv("NO_COLOR"); nc && nc[0] != '\0') + m_useAnsi = false; + } + explicit BenchmarkConsole(nbl::core::smart_refctd_ptr logger) + : BenchmarkConsole() + { + m_logger = std::move(logger); + } + + void setLogger(nbl::core::smart_refctd_ptr logger) { m_logger = std::move(logger); } + nbl::system::ILogger* getLogger() const { return m_logger.get(); } + + void setSilent(bool s) { m_silent = s; } + bool silent() const { return m_silent; } + + void setColorEnabled(bool e) { m_useAnsi = e; } + bool colorEnabled() const { return m_useAnsi; } + + // `neutral` is ELL_PERFORMANCE blue (not a full reset) so uncolored cell + // parts inherit the logger's line-wrap color. Only correct because rows / + // banners are all logged at ELL_PERFORMANCE. + struct Ansi + { + static constexpr std::string_view neutral = "\033[34m"; + static constexpr std::string_view reset = "\033[0m"; + static constexpr std::string_view red = "\033[31m"; + static constexpr std::string_view green = "\033[32m"; + static constexpr std::string_view yellow = "\033[33m"; + static constexpr std::string_view cyan = "\033[36m"; + static constexpr std::string_view bold = "\033[1m"; + }; + + // visualWidth excludes ANSI escape bytes (std::format's `{:>{}}` counts + // bytes), so colored cells must be padded manually via padCell. + struct CellOut + { + std::string text; + size_t visualWidth = 0; + }; + + const Format::Widths& widths() const { return m_widths; } + void growWidthFor(std::string_view joined) { m_widths.grow(joined); } + + // Sizes int columns to unchanged-value width, float columns to "value + // (+/-delta)" with delta=0. Changed-int rows overflow; padding every row + // for worst-case wastes ~40% horizontal space on stable runs. + void growForBaseline(const BaselineRow& b) + { + const auto growInt = [&](size_t& w, uint64_t v) + { + if (v == BaselineRow::kAbsent) + return; + w = std::max(w, std::format("{}", v).size()); + }; + growInt(m_widths.regs, b.registerCount); + growInt(m_widths.code, b.codeSizeBytes); + growInt(m_widths.shared, b.sharedMemBytes); + growInt(m_widths.local, b.privateMemBytes); + + if (b.psPerSample > 0.0) + { + m_widths.psSample = std::max(m_widths.psSample, floatCellPlainText(b.psPerSample, 0.0).size()); + const double gsBase = 1000.0 / b.psPerSample; + m_widths.gsamples = std::max(m_widths.gsamples, floatCellPlainText(gsBase, 0.0).size()); + } + } + + // Pre-register so the header (logged once up front) doesn't stay narrower than later rows. + void registerVariant(std::span name) { m_widths.grow(joinName(name)); } + void registerVariant(std::initializer_list name) + { + std::vector tmp; + tmp.reserve(name.size()); + for (auto s : name) + tmp.emplace_back(s); + m_widths.grow(joinName(tmp)); + } + + void logSectionBanner(std::string_view banner) const + { + if (banner.empty()) + return; + if (m_useAnsi) + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, "{}{}{}{}", Ansi::bold, Ansi::cyan, banner, Ansi::reset); + else + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, "{}", banner); + } + + // Once per session, not per span, otherwise readers see the same text N times. + template + void logBannerNotes(const Baselines& baselines) const + { + if (std::empty(baselines)) + return; + const auto& primary = *std::begin(baselines); + const bool multi = std::distance(std::begin(baselines), std::end(baselines)) > 1; + const std::string primaryLabel = primary.label; + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, + "Note: ps/sample lower = faster; GSamples/s higher = faster. Inline annotations compare to primary baseline '{}': " + "floats show 'value (+/-delta)' always; ints show 'old -> new' only when changed.", + primaryLabel); + if (multi) + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, + "Note: trailing 'vs LABEL' columns carry raw ps/sample deltas against secondary baselines (primary skipped, shown inline)."); + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, + "Note: '[WG!]' on a delta = baseline's workload shape (workgroup / dispatch / samplesPerDispatch) differs from this run, comparison is apples-to-oranges."); + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, + "Note: float deltas only get green/red coloring when the relative change is >= {:.0f}% (typical GPU jitter is 1-2%); smaller deltas stay neutral.", + kFloatColorThreshold * 100.0); + } + + template + void logHeader(const Baselines& baselines) const + { + std::string line = std::format("{:<{}} | {:>{}} | {:>{}} | {:>{}} | {:>{}} | {:>{}} | {:>{}}", + "Name", m_widths.name, + "ps/sample", m_widths.psSample, + "GSamples/s", m_widths.gsamples, + "regs", m_widths.regs, + "code(B)", m_widths.code, + "shared(B)", m_widths.shared, + "local(B)", m_widths.local); + // Primary is shown inline on every value column; only secondaries get trailing columns. + size_t idx = 0; + for (const auto& b : baselines) + { + if (idx++ == 0) + continue; + const std::string col = std::format("vs {}", b.label); + line += std::format(" | {:>{}}", col, baselineColWidth(b.label)); + } + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, "{}", line); + } + + template + void logRow(std::span name, std::string_view joinedName, + const TimingResult& t, const PipelineStats& s, + const std::unordered_map& rowBaselines, + const Baselines& baselines) const + { + if (!m_logger || m_silent) + return; + + const BaselineRow* primary = nullptr; + if (!std::empty(baselines)) + { + const std::string key = makeKey(name); + const auto& b0 = *std::begin(baselines); + if (auto it = b0.rowsByName.find(key); it != b0.rowsByName.end()) + primary = &it->second; + } + + // ps_per_sample * GSamples/s == 1000 (see runTimed), so GSamples is derived not stored. + const auto baselineGSamples = primary ? std::optional{primary->psPerSample > 0.0 ? 1000.0 / primary->psPerSample : 0.0} : std::nullopt; + + std::string line = std::format("{:<{}}", joinedName, m_widths.name); + line += " | " + padCell(formatFloatCell(t.ps_per_sample, primary ? std::optional{primary->psPerSample} : std::nullopt, true), m_widths.psSample); + line += " | " + padCell(formatFloatCell(t.gsamples_per_s, baselineGSamples, false), m_widths.gsamples); + line += " | " + padCell(formatIntCell(s.registerCount, primary ? primary->registerCount : BaselineRow::kAbsent), m_widths.regs); + line += " | " + padCell(formatIntCell(s.codeSizeBytes, primary ? primary->codeSizeBytes : BaselineRow::kAbsent), m_widths.code); + line += " | " + padCell(formatIntCell(s.sharedMemBytes, primary ? primary->sharedMemBytes : BaselineRow::kAbsent), m_widths.shared); + line += " | " + padCell(formatIntCell(s.privateMemBytes, primary ? primary->privateMemBytes : BaselineRow::kAbsent), m_widths.local); + + size_t idx = 0; + for (const auto& b : baselines) + { + if (idx++ == 0) + continue; + std::string plain; + bool better = false; + bool significant = false; + bool haveValue = false; + bool flagShape = false; + if (auto it = rowBaselines.find(b.label); it != rowBaselines.end() && it->second.psPerSample > 0.0) + { + const double delta = t.ps_per_sample - it->second.psPerSample; + plain = std::format("{:+.3f}", delta); + better = delta < 0.0; + significant = std::abs(delta) / it->second.psPerSample >= kFloatColorThreshold; + haveValue = true; + flagShape = it->second.shapeMismatch; + } + else + { + plain = "n/a"; + } + std::string suffix = flagShape ? std::string(" [WG!]") : std::string(); + CellOut cell; + cell.visualWidth = plain.size() + suffix.size(); + if (!m_useAnsi) + { + cell.text = plain + suffix; + } + else + { + const bool paint = haveValue && significant; + const std::string_view col = paint ? (better ? Ansi::green : Ansi::red) : std::string_view{}; + std::string coloredPlain = paint + ? std::format("{}{}{}", col, plain, Ansi::neutral) + : plain; + std::string coloredSuffix = flagShape + ? std::format("{}{}{}{}", Ansi::bold, Ansi::red, suffix, Ansi::neutral) + : std::string(); + cell.text = coloredPlain + coloredSuffix; + } + line += " | " + padCell(cell, baselineColWidth(b.label)); + } + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, "{}", line); + } + + // Flat table, one row per (variant, stat); each baseline gets one delta column: + // + // Name | stat | current | vs iter47 | vs iter48 + // X | ps/sample | 2.151 | -0.044 | +0.123 + // X | GSamples/s | 464.9 | +9.456 | -7.234 + // X | regs | 40 | +0 | +0 + // X | code(B) | 4992 | +128 | 0 + template + void printBaselineComparison(std::span> names, + const Baselines& baselines, const Results& results) const + { + if (!m_logger || names.empty()) + return; + if (std::empty(baselines)) + { + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_WARNING, + "--focus requested {} variant(s) but no baselines are loaded, nothing to compare against. " + "Did your --baseline paths fail to load?", + names.size()); + return; + } + + struct Current + { + TimingResult t; + PipelineStats s; + Workload w; + bool present = false; + }; + std::unordered_map currentByKey; + currentByKey.reserve(std::size(results)); + for (const auto& r : results) + currentByKey[makeKey(r.name)] = {r.timing, r.stats, r.workload, true}; + + const size_t baselineCount = static_cast(std::distance(std::begin(baselines), std::end(baselines))); + + std::vector> rows; + rows.reserve(1 + names.size() * 6); + + { + auto plainCell = [](std::string s) -> CellOut { const size_t w = s.size(); return {std::move(s), w}; }; + std::vector header; + header.reserve(3 + baselineCount); + header.push_back(plainCell("Name")); + header.push_back(plainCell("stat")); + header.push_back(plainCell("current")); + for (const auto& b : baselines) + header.push_back(plainCell(std::format("vs {}", b.label))); + rows.push_back(std::move(header)); + } + + auto floatStatRow = [&](const char* label, std::string_view joined, bool have, double curV, + const Workload& curW, const std::string& key, + auto baselineLookup /*BaselineRow -> double*/, bool lowerIsBetter) + { + auto plainCell = [](std::string s) -> CellOut { const size_t w = s.size(); return {std::move(s), w}; }; + std::vector row; + row.reserve(3 + baselineCount); + row.push_back(plainCell(std::string(joined))); + row.push_back(plainCell(label)); + row.push_back(have ? plainCell(formatFloat5(curV)) : plainCell("n/a")); + + for (const auto& b : baselines) + { + auto bit = b.rowsByName.find(key); + if (!have || bit == b.rowsByName.end()) + { + row.push_back(plainCell("n/a")); + continue; + } + const double baseV = baselineLookup(bit->second); + if (baseV <= 0.0) + { + row.push_back(plainCell("n/a")); + continue; + } + const bool shapeMismatch = curW.present() && bit->second.workload.present() && (curW.shape != bit->second.workload.shape); + const double delta = curV - baseV; + const std::string deltaStr = std::format("{}{}", delta >= 0 ? "+" : "-", formatFloat5(std::abs(delta))); + const bool significant = std::abs(delta) / baseV >= kFloatColorThreshold; + const std::string suffix = shapeMismatch ? std::string(" [WG!]") : std::string(); + CellOut cell; + cell.visualWidth = deltaStr.size() + suffix.size(); + if (!m_useAnsi || !significant) + { + cell.text = m_useAnsi && shapeMismatch + ? std::format("{}{}{}{}{}", deltaStr, Ansi::bold, Ansi::red, suffix, Ansi::neutral) + : deltaStr + suffix; + } + else + { + const bool better = (lowerIsBetter && delta < 0.0) || (!lowerIsBetter && delta > 0.0); + const std::string_view col = better ? Ansi::green : Ansi::red; + std::string coloredDelta = std::format("{}{}{}", col, deltaStr, Ansi::neutral); + std::string coloredSuffix = shapeMismatch + ? std::format("{}{}{}{}", Ansi::bold, Ansi::red, suffix, Ansi::neutral) + : std::string(); + cell.text = coloredDelta + coloredSuffix; + } + row.push_back(std::move(cell)); + } + rows.push_back(std::move(row)); + }; + + auto intStatRow = [&](const char* label, std::string_view joined, bool have, uint64_t curV, + const Workload& curW, const std::string& key, uint64_t BaselineRow::* baseField) + { + auto plainCell = [](std::string s) -> CellOut { const size_t w = s.size(); return {std::move(s), w}; }; + std::vector row; + row.reserve(3 + baselineCount); + row.push_back(plainCell(std::string(joined))); + row.push_back(plainCell(label)); + row.push_back(have ? plainCell(std::format("{}", curV)) : plainCell("n/a")); + + for (const auto& b : baselines) + { + auto bit = b.rowsByName.find(key); + if (!have || bit == b.rowsByName.end()) + { + row.push_back(plainCell("n/a")); + continue; + } + const uint64_t baseV = bit->second.*baseField; + if (baseV == BaselineRow::kAbsent) + { + row.push_back(plainCell("n/a")); + continue; + } + const bool shapeMismatch = curW.present() && bit->second.workload.present() && (curW.shape != bit->second.workload.shape); + const int64_t delta = int64_t(curV) - int64_t(baseV); + const std::string deltaStr = std::format("{:+d}", delta); + const std::string suffix = shapeMismatch ? std::string(" [WG!]") : std::string(); + CellOut cell; + cell.visualWidth = deltaStr.size() + suffix.size(); + if (!m_useAnsi) + { + cell.text = deltaStr + suffix; + } + else + { + std::string coloredDelta = delta != 0 + ? std::format("{}{}{}", Ansi::yellow, deltaStr, Ansi::neutral) + : deltaStr; + std::string coloredSuffix = shapeMismatch + ? std::format("{}{}{}{}", Ansi::bold, Ansi::red, suffix, Ansi::neutral) + : std::string(); + cell.text = coloredDelta + coloredSuffix; + } + row.push_back(std::move(cell)); + } + rows.push_back(std::move(row)); + }; + + for (const auto& nameVec : names) + { + const std::string joined = joinName(nameVec); + const std::string key = makeKey(nameVec); + const auto cit = currentByKey.find(key); + const bool have = (cit != currentByKey.end()) && cit->second.present; + const auto& t = have ? cit->second.t : TimingResult {}; + const auto& s = have ? cit->second.s : PipelineStats {}; + const auto& w = have ? cit->second.w : Workload {}; + + floatStatRow("ps/sample", joined, have, t.ps_per_sample, w, key, + [](const BaselineRow& b) { return b.psPerSample; }, true); + floatStatRow("GSamples/s", joined, have, t.gsamples_per_s, w, key, + [](const BaselineRow& b) { return b.psPerSample > 0.0 ? 1000.0 / b.psPerSample : 0.0; }, false); + intStatRow("regs", joined, have, s.registerCount, w, key, &BaselineRow::registerCount); + intStatRow("code(B)", joined, have, s.codeSizeBytes, w, key, &BaselineRow::codeSizeBytes); + intStatRow("shared(B)", joined, have, s.sharedMemBytes, w, key, &BaselineRow::sharedMemBytes); + intStatRow("local(B)", joined, have, s.privateMemBytes, w, key, &BaselineRow::privateMemBytes); + } + + const size_t nCols = 3 + baselineCount; + std::vector colWidths(nCols, 0); + for (const auto& r : rows) + for (size_t i = 0; i < r.size() && i < nCols; ++i) + colWidths[i] = std::max(colWidths[i], r[i].visualWidth); + + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, + "=== Focus comparison ({} variant(s) vs {} baseline(s); ps/sample lower is better, integer deltas are absolute) ===", + names.size(), baselineCount); + auto leftPad = [](const CellOut& c, size_t targetWidth) -> std::string + { + if (c.visualWidth >= targetWidth) + return c.text; + return c.text + std::string(targetWidth - c.visualWidth, ' '); + }; + for (size_t ri = 0; ri < rows.size(); ++ri) + { + std::string line; + for (size_t ci = 0; ci < rows[ri].size(); ++ci) + { + if (ci) + line.append(" | "); + if (ci <= 1) + line += leftPad(rows[ri][ci], colWidths[ci]); + else + line += padCell(rows[ri][ci], colWidths[ci]); + } + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_PERFORMANCE, "{}", line); + } + } + + private: + static constexpr size_t kBaselineMinColWidth = 10; + size_t baselineColWidth(std::string_view label) const + { + return std::max(kBaselineMinColWidth, std::string_view("vs ").size() + label.size()); + } + + // Typical GPU jitter is 1-2%; coloring below 5% would mostly highlight noise. + static constexpr double kFloatColorThreshold = 0.05; + + // std::format counts ANSI escape bytes, so `{:>N}` can't pad colored cells. + std::string padCell(const CellOut& c, size_t targetWidth) const + { + if (c.visualWidth >= targetWidth) + return c.text; + return std::string(targetWidth - c.visualWidth, ' ') + c.text; + } + + // "regs 40 -> 54" is more useful than "+14 from somewhere", show both endpoints. + CellOut formatIntCell(uint64_t current, uint64_t baseline) const + { + if (baseline == BaselineRow::kAbsent || baseline == current) + { + auto s = std::format("{}", current); + const size_t w = s.size(); + return {std::move(s), w}; + } + const std::string baseStr = std::format("{}", baseline); + const std::string curStr = std::format("{}", current); + const std::string plain = std::format("{} -> {}", baseStr, curStr); + const size_t visW = plain.size(); + if (!m_useAnsi) + return {plain, visW}; + auto colored = std::format("{}{} -> {}{}", Ansi::yellow, baseStr, curStr, Ansi::neutral); + return {std::move(colored), visW}; + } + + // ~5 chars including the decimal point, so column widths stay predictable + // across ps/sample (0.5..100) and GSamples/s (0.03..1000+). + static std::string formatFloat5(double v) + { + const double mag = std::abs(v); + if (mag >= 10000.0) return std::format("{:.0f}", v); + if (mag >= 1000.0) return std::format("{:.1f}", v); + if (mag >= 100.0) return std::format("{:.1f}", v); + if (mag >= 10.0) return std::format("{:.2f}", v); + return std::format("{:.3f}", v); + } + + static std::string floatCellPlainText(double value, double delta) + { + const std::string deltaStr = std::format("{}{}", delta >= 0 ? "+" : "-", formatFloat5(std::abs(delta))); + return std::format("{} ({})", formatFloat5(value), deltaStr); + } + + CellOut formatFloatCell(double current, std::optional baseline, bool lowerIsBetter) const + { + if (!baseline.has_value() || *baseline <= 0.0) + { + auto s = formatFloat5(current); + const size_t w = s.size(); + return {std::move(s), w}; + } + const double delta = current - *baseline; + const std::string plain = floatCellPlainText(current, delta); + const size_t visW = plain.size(); + const bool significant = std::abs(delta) / *baseline >= kFloatColorThreshold; + if (!m_useAnsi || !significant) + return {plain, visW}; + const std::string valStr = formatFloat5(current); + const std::string deltaStr = std::format("{}{}", delta >= 0 ? "+" : "-", formatFloat5(std::abs(delta))); + const bool better = (lowerIsBetter && delta < 0.0) || (!lowerIsBetter && delta > 0.0); + const std::string_view color = better ? Ansi::green : Ansi::red; + auto colored = std::format("{} ({}{}{})", valStr, color, deltaStr, Ansi::neutral); + return {std::move(colored), visW}; + } + + nbl::core::smart_refctd_ptr m_logger; + Format::Widths m_widths; + bool m_silent = false; + bool m_useAnsi = true; +}; + +#endif diff --git a/common/include/nbl/examples/Benchmark/BenchmarkJson.h b/common/include/nbl/examples/Benchmark/BenchmarkJson.h new file mode 100644 index 000000000..e6d3fff24 --- /dev/null +++ b/common/include/nbl/examples/Benchmark/BenchmarkJson.h @@ -0,0 +1,306 @@ +// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O. +// This file is part of the "Nabla Engine". +// For conditions of distribution and use, see copyright notice in nabla.h + +#ifndef _NBL_COMMON_BENCHMARK_JSON_INCLUDED_ +#define _NBL_COMMON_BENCHMARK_JSON_INCLUDED_ + +#include +#include "nbl/examples/Benchmark/BenchmarkTypes.h" +#include "nlohmann/json.hpp" + +#include +#include +#include +#include +#include +#include +#include + +namespace benchmark_json +{ + +// Builds the "device" JSON object from a physical device, or null if dev is null. +inline nlohmann::json buildDeviceMetadata(const nbl::video::IPhysicalDevice* dev) +{ + if (!dev) + return nullptr; + const auto& p = dev->getProperties(); + nlohmann::json out = nlohmann::json::object(); + out["name"] = std::string(p.deviceName); + out["vendorID"] = p.vendorID; + out["deviceID"] = p.deviceID; + out["driverID"] = static_cast(p.driverID); + out["driverName"] = std::string(p.driverName); + out["driverInfo"] = std::string(p.driverInfo); + out["driverVersion"] = p.driverVersion; + out["deviceUUID"] = std::vector(p.deviceUUID, p.deviceUUID + 16); + out["driverUUID"] = std::vector(p.driverUUID, p.driverUUID + 16); + return out; +} + +// Parses a JSON report file into a Baseline. Returns nullopt on missing / +// unparseable / empty file. Caller is responsible for appending / replacing +// in their baseline store and for feeding rows into BenchmarkConsole widths. +inline std::optional loadBaselineFile(std::string label, const std::string& path) +{ + std::ifstream f(path); + if (!f.is_open()) + return std::nullopt; + + nlohmann::json j; + try + { + f >> j; + } + catch (const std::exception&) + { + return std::nullopt; + } + + const auto resultsIt = j.find("results"); + if (resultsIt == j.end() || !resultsIt->is_array()) + return std::nullopt; + + std::unordered_map rowsByName; + for (const auto& r : *resultsIt) + { + const auto n = r.find("name"); + const auto ps = r.find("ps_per_sample"); + if (n == r.end() || ps == r.end()) + continue; + if (!n->is_array() || !ps->is_number()) + continue; + std::vector nameVec; + nameVec.reserve(n->size()); + for (const auto& seg : *n) + { + if (!seg.is_string()) + { + nameVec.clear(); + break; + } + nameVec.emplace_back(seg.get()); + } + if (nameVec.empty()) + continue; + + BaselineRow row; + try + { + row.psPerSample = ps->get(); + } + catch (const std::exception&) + { + continue; + } + + auto readU64 = [&](const char* key, uint64_t& out) + { + const auto it = r.find(key); + if (it != r.end() && it->is_number_unsigned()) + out = it->get(); + }; + readU64("regs", row.registerCount); + readU64("code_bytes", row.codeSizeBytes); + readU64("shared_mem_bytes", row.sharedMemBytes); + readU64("local_mem_bytes", row.privateMemBytes); + readU64("stack_bytes", row.stackBytes); + readU64("subgroup_size", row.subgroupSize); + + auto readUvec3 = [&](const char* key, nbl::hlsl::uint32_t3& out) + { + const auto it = r.find(key); + if (it == r.end() || !it->is_array() || it->size() != 3) + return; + const auto& a = *it; + if (!a[0].is_number_unsigned() || !a[1].is_number_unsigned() || !a[2].is_number_unsigned()) + return; + out.x = a[0].get(); + out.y = a[1].get(); + out.z = a[2].get(); + }; + readUvec3("workgroup_size", row.workload.shape.workgroupSize); + readUvec3("dispatch_groups", row.workload.shape.dispatchGroupCount); + readU64("samples_per_dispatch", row.workload.shape.samplesPerDispatch); + if (const auto it = r.find("bench_dispatches"); it != r.end() && it->is_number_unsigned()) + row.workload.benchDispatches = it->get(); + + rowsByName[makeKey(nameVec)] = row; + } + if (rowsByName.empty()) + return std::nullopt; + + return Baseline {std::move(label), path, j.contains("device") ? j["device"] : nullptr, std::move(rowsByName)}; +} + +// Writes a JSON report. Preserves rows in the prior file whose names weren't +// re-measured this run, so writeReportFile can be an intermediate checkpoint +// during a multi-bench-class session. Returns preservedCount via out-param. +inline bool writeReportFile(const std::string& path, const nlohmann::json& deviceMetadata, const std::vector& baselines, const std::vector& results, nbl::system::ILogger* logger, size_t* outPreservedCount = nullptr) +{ + nlohmann::json doc; + doc["version"] = 1; + + if (!deviceMetadata.is_null()) + doc["device"] = deviceMetadata; + + if (!baselines.empty()) + { + auto& baselinesNode = doc["baselines"] = nlohmann::json::object(); + for (const auto& b : baselines) + baselinesNode[b.label] = b.path; + } + auto& resultsNode = doc["results"] = nlohmann::json::array(); + + std::unordered_set currentKeys; + currentKeys.reserve(results.size()); + for (const auto& r : results) + currentKeys.insert(makeKey(r.name)); + + for (const auto& r : results) + { + nlohmann::json row; + row["name"] = r.name; + row["ps_per_sample"] = r.timing.ps_per_sample; + row["gsamples_per_s"] = r.timing.gsamples_per_s; + row["ms_total"] = r.timing.ms_total; + row["regs"] = r.stats.registerCount; + row["code_bytes"] = r.stats.codeSizeBytes; + row["shared_mem_bytes"] = r.stats.sharedMemBytes; + row["local_mem_bytes"] = r.stats.privateMemBytes; + row["stack_bytes"] = r.stats.stackBytes; + row["subgroup_size"] = r.stats.subgroupSize; + + // Structured so JSON preserves the exact numeric type. + if (!r.stats.unknowns.empty()) + { + using F = nbl::video::IGPUPipelineBase::SExecutableStatistic::FORMAT; + auto& arr = row["unknown_stats"] = nlohmann::json::array(); + for (const auto& s : r.stats.unknowns) + { + nlohmann::json entry; + entry["name"] = s.name; + switch (s.format) + { + case F::BOOL32: + entry["type"] = "bool"; + entry["value"] = s.value.b32; + break; + case F::INT64: + entry["type"] = "int"; + entry["value"] = s.value.i64; + break; + case F::UINT64: + entry["type"] = "uint"; + entry["value"] = s.value.u64; + break; + case F::FLOAT64: + entry["type"] = "float"; + entry["value"] = s.value.f64; + break; + } + arr.push_back(std::move(entry)); + } + } + + row["workgroup_size"] = {r.workload.shape.workgroupSize.x, r.workload.shape.workgroupSize.y, r.workload.shape.workgroupSize.z}; + row["dispatch_groups"] = {r.workload.shape.dispatchGroupCount.x, r.workload.shape.dispatchGroupCount.y, r.workload.shape.dispatchGroupCount.z}; + row["samples_per_dispatch"] = r.workload.shape.samplesPerDispatch; + row["bench_dispatches"] = r.workload.benchDispatches; + + resultsNode.push_back(std::move(row)); + } + + // Caveat: renamed/removed variants linger forever. Delete the output JSON + // to get a clean slate. + size_t preservedCount = 0; + { + std::ifstream in(path); + if (in.is_open()) + { + nlohmann::json existing; + try + { + in >> existing; + } + catch (const std::exception&) + { + existing = nullptr; + } + const auto rIt = existing.find("results"); + if (rIt != existing.end() && rIt->is_array()) + { + for (const auto& priorRow : *rIt) + { + const auto n = priorRow.find("name"); + if (n == priorRow.end() || !n->is_array()) + continue; + std::vector nameVec; + bool ok = true; + for (const auto& seg : *n) + { + if (!seg.is_string()) + { + ok = false; + break; + } + nameVec.emplace_back(seg.get()); + } + if (!ok || nameVec.empty()) + continue; + if (currentKeys.find(makeKey(nameVec)) != currentKeys.end()) + continue; // re-measured this run + + resultsNode.push_back(priorRow); + ++preservedCount; + } + } + } + } + + std::ofstream f(path, std::ios::out | std::ios::trunc); + if (!f.is_open()) + { + benchLogFmt(logger, nbl::system::ILogger::ELL_ERROR, "benchmark_json::writeReportFile: failed to open '{}'", path); + return false; + } + + // One result per line keeps `git diff` showing one row per change instead + // of N lines per row. + f << "{\n"; + f << " \"version\": " << doc["version"].dump() << ",\n"; + if (doc.contains("device")) + { + // Compact value render so byte arrays (deviceUUID etc.) stay inline. + const auto& dev = doc["device"]; + f << " \"device\": {\n"; + bool first = true; + for (auto it = dev.begin(); it != dev.end(); ++it) + { + if (!first) + f << ",\n"; + first = false; + f << " \"" << it.key() << "\": " << it.value().dump(); + } + f << "\n },\n"; + } + if (doc.contains("baselines")) + f << " \"baselines\": " << doc["baselines"].dump() << ",\n"; + f << " \"results\": ["; + for (size_t i = 0; i < resultsNode.size(); ++i) + { + f << (i ? ",\n " : "\n "); + f << resultsNode[i].dump(); + } + f << (resultsNode.empty() ? "]\n" : "\n ]\n"); + f << "}\n"; + + if (outPreservedCount) + *outPreservedCount = preservedCount; + return true; +} + +} // namespace benchmark_json + +#endif diff --git a/common/include/nbl/examples/Benchmark/BenchmarkTypes.h b/common/include/nbl/examples/Benchmark/BenchmarkTypes.h new file mode 100644 index 000000000..274c19514 --- /dev/null +++ b/common/include/nbl/examples/Benchmark/BenchmarkTypes.h @@ -0,0 +1,211 @@ +// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O. +// This file is part of the "Nabla Engine". +// For conditions of distribution and use, see copyright notice in nabla.h + +#ifndef _NBL_COMMON_BENCHMARK_TYPES_INCLUDED_ +#define _NBL_COMMON_BENCHMARK_TYPES_INCLUDED_ + +#include +#include "nlohmann/json.hpp" + +#include +#include +#include +#include +#include +#include +#include +#include + +struct PipelineStats +{ + uint64_t registerCount = 0; + uint64_t codeSizeBytes = 0; + uint64_t sharedMemBytes = 0; + uint64_t privateMemBytes = 0; + uint64_t stackBytes = 0; + uint32_t subgroupSize = 0; + std::string raw; + + // Driver stats matchStat didn't recognise. Structured (not lossy-stringified + // into `raw`) so JSON round-trips the correct numeric type. + std::vector unknowns; +}; + +struct TimingResult +{ + float64_t elapsed_ns = 0.0; + uint64_t totalSamples = 0; + float64_t ps_per_sample = 0.0; + float64_t gsamples_per_s = 0.0; + float64_t ms_total = 0.0; +}; + +struct Format +{ + struct Widths + { + size_t name = std::string_view("Name").size(); + size_t psSample = std::string_view("ps/sample").size(); + size_t gsamples = std::string_view("GSamples/s").size(); + size_t regs = std::string_view("regs").size(); + size_t code = std::string_view("code(B)").size(); + size_t shared = std::string_view("shared(B)").size(); + size_t local = std::string_view("local(B)").size(); + + void grow(std::string_view joinedName) { name = std::max(name, joinedName.size()); } + }; + + static std::string headerBase(const Widths& w = {}) + { + return std::format("{:<{}} | {:>12} | {:>12} | {:>6} | {:>8} | {:>12} | {:>12}", + "Name", w.name, "ps/sample", "GSamples/s", "regs", "code(B)", "shared(B)", "local(B)"); + } + + static std::string dataBase(const Widths& w, std::string_view joinedName, const TimingResult& t, const PipelineStats& s) + { + return std::format("{:<{}} | {:>12.3f} | {:>12.3f} | {:>6} | {:>8} | {:>12} | {:>12}", + joinedName, w.name, t.ps_per_sample, t.gsamples_per_s, s.registerCount, s.codeSizeBytes, s.sharedMemBytes, s.privateMemBytes); + } +}; + +// The "what was measured" part of a workload. Workload (adds benchDispatches) +// and RunContext (adds banner label + budget) both embed a WorkloadShape, so +// the shape can be sliced into either from the other. +struct WorkloadShape +{ + nbl::hlsl::uint32_t3 workgroupSize = {0, 0, 0}; + nbl::hlsl::uint32_t3 dispatchGroupCount = {0, 0, 0}; + uint64_t samplesPerDispatch = 0; + + inline bool operator==(const WorkloadShape& other) const + { + return workgroupSize == other.workgroupSize && dispatchGroupCount == other.dispatchGroupCount && samplesPerDispatch == other.samplesPerDispatch; + } + + inline bool operator!=(const WorkloadShape& other) const + { + return !(*this == other); + } +}; + +struct Workload +{ + WorkloadShape shape; + uint32_t benchDispatches = 0; + + // Default-constructed (all zeros) signals "not recorded". + bool present() const { return shape.samplesPerDispatch != 0; } +}; + +struct BaselineRow +{ + // UINT64_MAX sentinel: no real pipeline stat reaches that magnitude, so an + // "absent" field can't collide with a real value. The current run can also + // produce kAbsent when a driver doesn't expose a given stat. + static constexpr uint64_t kAbsent = std::numeric_limits::max(); + + float64_t psPerSample = 0.0; + uint64_t registerCount = kAbsent; + uint64_t codeSizeBytes = kAbsent; + uint64_t sharedMemBytes = kAbsent; + uint64_t privateMemBytes = kAbsent; + uint64_t stackBytes = kAbsent; + uint64_t subgroupSize = kAbsent; // uint64_t (not 32) to share kAbsent semantics + Workload workload {}; +}; + +// Per-baseline reference for a single row: the baseline's ps/sample plus +// whether its recorded workload shape differs from this run (renders the +// "[WG!]" marker so the reader knows the comparison is questionable). +struct BaselineRef +{ + float64_t psPerSample = 0.0; + bool shapeMismatch = false; +}; + +struct Result +{ + // Hierarchical name, outermost first. Tooling can group by any prefix; the + // console joins with " > ". + nbl::core::vector name; + TimingResult timing {}; + PipelineStats stats {}; + Workload workload {}; + std::unordered_map baselines; +}; + +inline std::string joinName(std::span name, std::string_view sep = " > ") +{ + std::string out; + for (size_t i = 0; i < name.size(); ++i) + { + if (i) + out.append(sep); + out.append(name[i]); + } + return out; +} + +// Unit-separator (\x1f) between segments so makeKey can't collide with any +// user-supplied content. +inline std::string makeKey(std::span name) +{ + std::string k; + size_t total = 0; + for (const auto& s : name) + total += s.size() + 1; + k.reserve(total); + for (size_t i = 0; i < name.size(); ++i) + { + if (i) + k.push_back('\x1f'); + k.append(name[i]); + } + return k; +} + +inline nbl::core::vector splitFocusSpec(std::string_view spec) +{ + auto trim = [](std::string_view s) + { + while (!s.empty() && (s.front() == ' ' || s.front() == '\t')) + s.remove_prefix(1); + while (!s.empty() && (s.back() == ' ' || s.back() == '\t')) + s.remove_suffix(1); + return s; + }; + nbl::core::vector out; + size_t start = 0; + while (start <= spec.size()) + { + size_t end = spec.find('>', start); + if (end == std::string_view::npos) + end = spec.size(); + const auto seg = trim(spec.substr(start, end - start)); + if (!seg.empty()) + out.emplace_back(seg); + if (end == spec.size()) + break; + start = end + 1; + } + return out; +} + +struct Baseline +{ + std::string label; + std::string path; + nlohmann::json device; // top-level "device" field from the file, or null if absent + std::unordered_map rowsByName; // makeKey(name) -> stats +}; + +template +inline void benchLogFmt(nbl::system::ILogger* logger, nbl::system::ILogger::E_LOG_LEVEL level, std::string_view fmt, const Args&... args) +{ + if (!logger) + return; + logger->log("%s", level, std::vformat(fmt, std::make_format_args(args...)).c_str()); +} + +#endif diff --git a/common/include/nbl/examples/Benchmark/GPUBenchmarkHelper.h b/common/include/nbl/examples/Benchmark/GPUBenchmarkHelper.h new file mode 100644 index 000000000..553e5a21b --- /dev/null +++ b/common/include/nbl/examples/Benchmark/GPUBenchmarkHelper.h @@ -0,0 +1,784 @@ +// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O. +// This file is part of the "Nabla Engine". +// For conditions of distribution and use, see copyright notice in nabla.h + +#ifndef _NBL_COMMON_GPU_BENCHMARK_HELPER_INCLUDED_ +#define _NBL_COMMON_GPU_BENCHMARK_HELPER_INCLUDED_ + +#include +#include "nbl/examples/examples.hpp" +#include "nbl/examples/Benchmark/BenchmarkTypes.h" +#include "nbl/asset/utils/CCompilerSet.h" +#include "nbl/asset/utils/IShaderCompiler.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +class GPUBenchmarkHelper +{ +public: + struct InitData + { + nbl::core::smart_refctd_ptr device; + nbl::core::smart_refctd_ptr logger; + nbl::video::IPhysicalDevice* physicalDevice = nullptr; + uint32_t computeFamilyIndex = 0; + nbl::hlsl::uint32_t3 dispatchGroupCount = {0, 0, 0}; + uint64_t samplesPerDispatch = 0; + }; + + // One shader source for a benchmark variant. Picks ONE of two paths: + // * Precompiled: `precompiledKey` is a SPIRV asset key from CMake-time + // NBL_CREATE_NSC_COMPILE_RULES. `defines` is ignored. + // * Runtime: `sourcePath` is an .hlsl file resolved against "app_resources", + // compiled at load time with `defines` as -D macros. Use this for fast + // variant iteration without reconfiguring CMake. + struct ShaderVariant + { + // SMacroDefinition uses string_view; this struct owns the backing strings. + struct Define + { + std::string identifier; + std::string definition; + }; + + std::string sourcePath; + std::string precompiledKey; + std::vector defines; + nbl::asset::IShader::E_SHADER_STAGE stage = nbl::asset::IShader::E_SHADER_STAGE::ESS_COMPUTE; + + static ShaderVariant Precompiled(std::string key) + { + ShaderVariant v; + v.precompiledKey = std::move(key); + return v; + } + static ShaderVariant FromSource(std::string path, std::vector defs = {}, nbl::asset::IShader::E_SHADER_STAGE stage = nbl::asset::IShader::E_SHADER_STAGE::ESS_COMPUTE) + { + ShaderVariant v; + v.sourcePath = std::move(path); + v.defines = std::move(defs); + v.stage = stage; + return v; + } + + bool isRuntime() const { return !sourcePath.empty() && precompiledKey.empty(); } + bool isPrecompiled() const { return !precompiledKey.empty(); } + }; + + // Logical layout: [warmup x dispatchOne][ts0][bench x dispatchOne][ts1][cooldown x dispatchOne] + // Warmup/cooldown can be split into shorter submissions and the measured window stays intact. + // Putting binds inside dispatchOne adds per-iteration cmdbuf overhead that + // shows up in ps/sample on tight shaders. + using DispatchFn = std::function; + + // Input choice for createBindings(). Output is always implicit BDA. + enum class InputBuffer : uint8_t + { + None, + BDA, + SSBO, + UBO, + }; + + struct BindingsConfig + { + size_t outputBytes = 0; + size_t pushConstantBytes = 0; + size_t inputBytes = 0; + InputBuffer inputMode = InputBuffer::None; + }; + + struct Bindings + { + nbl::core::smart_refctd_ptr outputBuf; + uint64_t outputAddress = 0; + nbl::core::smart_refctd_ptr pipelineLayout; + + nbl::core::smart_refctd_ptr inputBuf; + uint64_t inputAddress = 0; // BDA mode only + + nbl::core::smart_refctd_ptr dsLayout; + nbl::core::smart_refctd_ptr ds; + }; + + struct PipelineEntry + { + nbl::core::smart_refctd_ptr pipeline; + nbl::core::smart_refctd_ptr layout; + PipelineStats stats; + std::string tag; + }; + + // Common bindOnce body: bind pipeline + upload push constants. Most benches + // have nothing else in bindOnce; the few that bind descriptor sets too call + // cb->bindDescriptorSets() before/after this. + template + static void defaultBindAndPush(nbl::video::IGPUCommandBuffer* cb, const PipelineEntry& pe, const PC& pc) + { + cb->bindComputePipeline(pe.pipeline.get()); + cb->pushConstants(pe.layout.get(), nbl::asset::IShader::E_SHADER_STAGE::ESS_COMPUTE, 0, sizeof(PC), &pc); + } + + // Dispatch using m_dispatchGroupCount (the setup-time shape). + void defaultDispatch(nbl::video::IGPUCommandBuffer* cb) const + { + cb->dispatch(m_dispatchGroupCount.x, m_dispatchGroupCount.y, m_dispatchGroupCount.z); + } + + bool init(const InitData& data) + { + m_device = data.device; + m_logger = data.logger; + m_physicalDevice = data.physicalDevice; + m_queue = m_device->getQueue(data.computeFamilyIndex, 0); + m_dispatchGroupCount = data.dispatchGroupCount; + m_samplesPerDispatch = data.samplesPerDispatch; + + m_cmdpool = m_device->createCommandPool(data.computeFamilyIndex, + nbl::video::IGPUCommandPool::CREATE_FLAGS::RESET_COMMAND_BUFFER_BIT); + if (!m_cmdpool->createCommandBuffers(nbl::video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &m_cmdbuf)) + { + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_ERROR, "GPUBenchmarkHelper: failed to create cmdbuf"); + return false; + } + + nbl::video::IQueryPool::SCreationParams qparams = {}; + qparams.queryType = nbl::video::IQueryPool::TYPE::TIMESTAMP; + qparams.queryCount = 2; + qparams.pipelineStatisticsFlags = nbl::video::IQueryPool::PIPELINE_STATISTICS_FLAGS::NONE; + m_queryPool = m_device->createQueryPool(qparams); + if (!m_queryPool) + { + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_ERROR, "GPUBenchmarkHelper: failed to create timestamp query pool"); + return false; + } + return true; + } + + // Load (precompiled path) or load+compile (runtime path) a variant's SPIRV. + nbl::core::smart_refctd_ptr loadShader(const ShaderVariant& variant, nbl::core::smart_refctd_ptr assetMgr) const + { + using namespace nbl; + if (!variant.isRuntime() && !variant.isPrecompiled()) + { + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "GPUBenchmarkHelper::loadShader: variant has neither sourcePath nor precompiledKey"); + return nullptr; + } + + asset::IAssetLoader::SAssetLoadParams lp = {}; + lp.logger = m_logger.get(); + + std::string key; + if (variant.isPrecompiled()) + { + lp.workingDirectory = "app_resources"; + key = variant.precompiledKey; + } + else + { + lp.workingDirectory = ""; + key = "app_resources/" + variant.sourcePath; + } + auto bundle = assetMgr->getAsset(key, lp); + const auto assets = bundle.getContents(); + if (assets.empty()) + { + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "GPUBenchmarkHelper::loadShader: failed to load '{}'", key); + return nullptr; + } + auto source = asset::IAsset::castDown(assets[0]); + if (!source) + { + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "GPUBenchmarkHelper::loadShader: '{}' is not an IShader asset", key); + return nullptr; + } + + if (variant.isPrecompiled()) + return source; + + auto* compilerSet = assetMgr->getCompilerSet(); + auto compiler = compilerSet->getShaderCompiler(source->getContentType()); + if (!compiler) + { + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "GPUBenchmarkHelper::loadShader: no compiler for content type of '{}'", variant.sourcePath); + return nullptr; + } + + std::vector wireDefines; + wireDefines.reserve(variant.defines.size()); + for (const auto& d : variant.defines) + wireDefines.push_back({d.identifier, d.definition}); + + asset::IShaderCompiler::SCompilerOptions options = {}; + options.stage = variant.stage; + options.preprocessorOptions.targetSpirvVersion = m_device->getPhysicalDevice()->getLimits().spirvVersion; + options.preprocessorOptions.sourceIdentifier = source->getFilepathHint(); + options.preprocessorOptions.logger = m_logger.get(); + options.preprocessorOptions.includeFinder = compiler->getDefaultIncludeFinder(); + options.preprocessorOptions.extraDefines = {wireDefines.data(), wireDefines.size()}; + + auto spirv = compilerSet->compileToSPIRV(source.get(), options); + if (!spirv) + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "GPUBenchmarkHelper::loadShader: runtime compile failed for '{}'", variant.sourcePath); + return spirv; + } + + nbl::core::smart_refctd_ptr allocateDeviceLocalBuffer(nbl::video::IGPUBuffer::SCreationParams bp, const char* label, + nbl::video::IDeviceMemoryAllocation::E_MEMORY_ALLOCATE_FLAGS allocFlags = nbl::video::IDeviceMemoryAllocation::EMAF_NONE) + { + auto buf = m_device->createBuffer(std::move(bp)); + auto reqs = buf->getMemoryReqs(); + reqs.memoryTypeBits &= m_physicalDevice->getDeviceLocalMemoryTypeBits(); + auto alloc = m_device->allocate(reqs, buf.get(), allocFlags); + if (!alloc.isValid()) + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_ERROR, "GPUBenchmarkHelper: failed to allocate {}", label); + return buf; + } + + struct SingleBindingDS + { + nbl::core::smart_refctd_ptr layout; + nbl::core::smart_refctd_ptr set; + }; + + SingleBindingDS createSingleBindingDS( + nbl::core::smart_refctd_ptr buffer, + nbl::asset::IDescriptor::E_TYPE type = nbl::asset::IDescriptor::E_TYPE::ET_STORAGE_BUFFER, + uint32_t binding = 0, + nbl::hlsl::ShaderStage stages = nbl::hlsl::ShaderStage::ESS_COMPUTE) + { + using namespace nbl; + const size_t bufferBytes = buffer->getSize(); + + video::IGPUDescriptorSetLayout::SBinding b = { + .binding = binding, + .type = type, + .createFlags = video::IGPUDescriptorSetLayout::SBinding::E_CREATE_FLAGS::ECF_NONE, + .stageFlags = stages, + .count = 1, + }; + SingleBindingDS out; + out.layout = m_device->createDescriptorSetLayout({&b, 1}); + auto pool = m_device->createDescriptorPoolForDSLayouts(video::IDescriptorPool::ECF_NONE, {&out.layout.get(), 1}); + out.set = pool->createDescriptorSet(core::smart_refctd_ptr(out.layout)); + + video::IGPUDescriptorSet::SDescriptorInfo info = {}; + info.desc = std::move(buffer); + info.info.buffer = {.offset = 0, .size = bufferBytes}; + video::IGPUDescriptorSet::SWriteDescriptorSet w = { + .dstSet = out.set.get(), + .binding = binding, + .arrayElement = 0, + .count = 1, + .info = &info, + }; + m_device->updateDescriptorSets({&w, 1}, {}); + return out; + } + + nbl::core::smart_refctd_ptr createOutputBuffer( + size_t bytes, + nbl::core::bitflag extraUsage = nbl::video::IGPUBuffer::E_USAGE_FLAGS::EUF_NONE, + nbl::video::IDeviceMemoryAllocation::E_MEMORY_ALLOCATE_FLAGS allocFlags = nbl::video::IDeviceMemoryAllocation::EMAF_NONE) + { + nbl::video::IGPUBuffer::SCreationParams bp = {}; + bp.size = bytes; + bp.usage = nbl::core::bitflag(nbl::video::IGPUBuffer::EUF_STORAGE_BUFFER_BIT) | extraUsage; + return allocateDeviceLocalBuffer(std::move(bp), "output buffer", allocFlags); + } + + // Buffer must have been created with EUF_TRANSFER_DST_BIT. + void submitFillZero(nbl::core::smart_refctd_ptr buf, size_t bytes) const + { + nbl::core::smart_refctd_ptr initCmdbuf; + m_cmdpool->createCommandBuffers(nbl::video::IGPUCommandPool::BUFFER_LEVEL::PRIMARY, 1u, &initCmdbuf); + initCmdbuf->begin(nbl::video::IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); + const nbl::asset::SBufferRange range = {.offset = 0, .size = bytes, .buffer = std::move(buf)}; + initCmdbuf->fillBuffer(range, 0u); + initCmdbuf->end(); + + const nbl::video::IQueue::SSubmitInfo::SCommandBufferInfo cmds[] = {{.cmdbuf = initCmdbuf.get()}}; + nbl::video::IQueue::SSubmitInfo submit = {}; + submit.commandBuffers = cmds; + m_queue->submit({&submit, 1u}); + m_device->waitIdle(); + } + + nbl::core::smart_refctd_ptr createInputBufferZeroFilled(size_t bytes) + { + auto buf = createOutputBuffer(bytes, nbl::video::IGPUBuffer::EUF_TRANSFER_DST_BIT); + if (buf) + submitFillZero(buf, bytes); + return buf; + } + + // BDA buffer staged into device-local VRAM via IUtilities. + nbl::core::smart_refctd_ptr createBdaBuffer(const void* srcData, size_t bytes) + { + using namespace nbl; + if (!m_utils) + m_utils = video::IUtilities::create(core::smart_refctd_ptr(m_device), core::smart_refctd_ptr(m_logger)); + + video::IGPUBuffer::SCreationParams bp = {}; + bp.size = bytes; + bp.usage = core::bitflag(video::IGPUBuffer::EUF_STORAGE_BUFFER_BIT) | video::IGPUBuffer::EUF_SHADER_DEVICE_ADDRESS_BIT | video::IGPUBuffer::EUF_TRANSFER_DST_BIT; + core::smart_refctd_ptr buf; + auto future = m_utils->createFilledDeviceLocalBufferOnDedMem( + video::SIntendedSubmitInfo {.queue = m_queue}, std::move(bp), srcData); + future.move_into(buf); + return buf; + } + + uint32_t createPipeline(const ShaderVariant& variant, + nbl::core::smart_refctd_ptr assetMgr, + size_t pushConstantSize, + std::string tag = "", + nbl::core::smart_refctd_ptr dsLayout = nullptr) + { + using namespace nbl; + PipelineEntry slot = {.tag = tag}; + + const asset::SPushConstantRange pcRange = { + .stageFlags = asset::IShader::E_SHADER_STAGE::ESS_COMPUTE, + .offset = 0, + .size = uint32_t(pushConstantSize), + }; + auto layout = dsLayout + ? m_device->createPipelineLayout({&pcRange, 1}, core::smart_refctd_ptr(dsLayout)) + : m_device->createPipelineLayout({&pcRange, 1}); + if (!layout) + { + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "createPipeline({}): pipeline layout creation failed", tag); + return InvalidPipelineIndex; + } + + auto source = loadShader(variant, std::move(assetMgr)); + auto shader = source ? m_device->compileShader({.source = source.get()}) : nullptr; + if (!shader) + { + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "createPipeline({}): shader load/compile failed", tag); + return InvalidPipelineIndex; + } + + video::IGPUComputePipeline::SCreationParams pp = {}; + pp.layout = layout.get(); + pp.shader.shader = shader.get(); + pp.shader.entryPoint = "main"; + if (m_device->getEnabledFeatures().pipelineExecutableInfo) + pp.flags |= video::IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_STATISTICS | video::IGPUComputePipeline::SCreationParams::FLAGS::CAPTURE_INTERNAL_REPRESENTATIONS; + + core::smart_refctd_ptr pipeline; + if (!m_device->createComputePipelines(nullptr, {&pp, 1}, &pipeline) || !pipeline) + { + benchLogFmt(m_logger.get(), system::ILogger::ELL_ERROR, "createPipeline({}): createComputePipelines failed", tag); + return InvalidPipelineIndex; + } + + if (m_device->getEnabledFeatures().pipelineExecutableInfo) + { + auto infos = pipeline->getExecutableInfo(); + slot.stats.raw = nbl::system::to_string(infos); + + uint64_t vgpr = 0, sgpr = 0; + for (const auto& info : infos) + { + if (info.subgroupSize) + slot.stats.subgroupSize = std::max(slot.stats.subgroupSize, info.subgroupSize); + for (const auto& stat : info.structuredStatistics) + matchStat(stat, slot.stats, vgpr, sgpr); + } + // AMD-style drivers expose VGPR/SGPR separately without a combined + // register count, so fall back to the sum. + if (slot.stats.registerCount == 0 && (vgpr || sgpr)) + slot.stats.registerCount = vgpr + sgpr; + + if (!slot.stats.raw.empty()) + benchLogFmt(m_logger.get(), system::ILogger::ELL_PERFORMANCE, "{} pipeline executable report:\n{}", tag, slot.stats.raw); + } + + slot.layout = std::move(layout); + slot.pipeline = std::move(pipeline); + const uint32_t idx = uint32_t(m_pipelines.size()); + m_pipelines.push_back(std::move(slot)); + return idx; + } + + Bindings createBindings(const BindingsConfig& cfg) + { + using namespace nbl; + Bindings out; + + out.outputBuf = createOutputBuffer(cfg.outputBytes, video::IGPUBuffer::EUF_SHADER_DEVICE_ADDRESS_BIT, video::IDeviceMemoryAllocation::EMAF_DEVICE_ADDRESS_BIT); + out.outputAddress = out.outputBuf->getDeviceAddress(); + + if (cfg.inputMode != InputBuffer::None && cfg.inputBytes > 0) + { + const bool useBDA = cfg.inputMode == InputBuffer::BDA; + const bool useUBO = cfg.inputMode == InputBuffer::UBO; + const bool useSSBO = cfg.inputMode == InputBuffer::SSBO; + + video::IGPUBuffer::SCreationParams bp = {}; + bp.size = cfg.inputBytes; + bp.usage = core::bitflag(video::IGPUBuffer::EUF_TRANSFER_DST_BIT); + if (useBDA || useSSBO) + bp.usage |= video::IGPUBuffer::EUF_STORAGE_BUFFER_BIT; + if (useBDA) + bp.usage |= video::IGPUBuffer::EUF_SHADER_DEVICE_ADDRESS_BIT; + if (useUBO) + bp.usage |= video::IGPUBuffer::EUF_UNIFORM_BUFFER_BIT; + + out.inputBuf = allocateDeviceLocalBuffer(std::move(bp), "input buffer", + useBDA ? video::IDeviceMemoryAllocation::EMAF_DEVICE_ADDRESS_BIT : video::IDeviceMemoryAllocation::EMAF_NONE); + + if (useBDA) + out.inputAddress = out.inputBuf->getDeviceAddress(); + + submitFillZero(out.inputBuf, cfg.inputBytes); + + if (useSSBO || useUBO) + { + video::IGPUDescriptorSetLayout::SBinding b = { + .binding = 0, + .type = useSSBO ? asset::IDescriptor::E_TYPE::ET_STORAGE_BUFFER : asset::IDescriptor::E_TYPE::ET_UNIFORM_BUFFER, + .createFlags = video::IGPUDescriptorSetLayout::SBinding::E_CREATE_FLAGS::ECF_NONE, + .stageFlags = nbl::hlsl::ShaderStage::ESS_COMPUTE, + .count = 1, + }; + out.dsLayout = m_device->createDescriptorSetLayout({&b, 1}); + + auto pool = m_device->createDescriptorPoolForDSLayouts(video::IDescriptorPool::ECF_NONE, {&out.dsLayout.get(), 1}); + out.ds = pool->createDescriptorSet(core::smart_refctd_ptr(out.dsLayout)); + + video::IGPUDescriptorSet::SDescriptorInfo info = {}; + info.desc = core::smart_refctd_ptr(out.inputBuf); + info.info.buffer = {.offset = 0, .size = cfg.inputBytes}; + video::IGPUDescriptorSet::SWriteDescriptorSet w = { + .dstSet = out.ds.get(), + .binding = 0, + .arrayElement = 0, + .count = 1, + .info = &info, + }; + m_device->updateDescriptorSets({&w, 1}, {}); + } + } + + { + const asset::SPushConstantRange pc = { + .stageFlags = nbl::hlsl::ShaderStage::ESS_COMPUTE, + .offset = 0, + .size = uint32_t(cfg.pushConstantBytes), + }; + std::span pcRange = cfg.pushConstantBytes > 0 ? std::span(&pc, 1) : std::span {}; + + if (out.dsLayout) + out.pipelineLayout = m_device->createPipelineLayout(pcRange, core::smart_refctd_ptr(out.dsLayout)); + else + out.pipelineLayout = m_device->createPipelineLayout(pcRange); + } + + return out; + } + + struct BdaBuffer + { + nbl::core::smart_refctd_ptr buf; + uint64_t address = 0; + }; + + BdaBuffer createBdaOutputBuffer(size_t bytes) + { + BdaBuffer out; + out.buf = createOutputBuffer(bytes, nbl::video::IGPUBuffer::EUF_SHADER_DEVICE_ADDRESS_BIT, nbl::video::IDeviceMemoryAllocation::EMAF_DEVICE_ADDRESS_BIT); + out.address = out.buf ? out.buf->getDeviceAddress() : 0; + return out; + } + + // Auto-sizes the dispatch count so the measured window covers ~targetBudgetMs + // of GPU work. Pilots with a small N, then either scales to the budget or + // doubles when the pilot is too noisy (sub-millisecond) to extrapolate. + // + // `samples` controls jitter robustness: values >1 take K independent + // budget-sized timing windows and return the MEDIAN window, costing ~K * + // targetBudgetMs of wall time. Median (not min) is used because GPU + // measurement noise can be two-sided in practice. + TimingResult runTimedBudgeted(uint32_t warmupDispatches, uint64_t targetBudgetMs, const DispatchFn& bindOnce, const DispatchFn& dispatchOne, uint32_t samples) + { + const uint64_t targetBudgetNs = targetBudgetMs * 1'000'000ull; + constexpr uint32_t kPilotN = 64; + constexpr uint32_t kMaxN = 1u << 24; // safety cap for ultra-fast shaders + uint32_t dispatchesPerSubmit = 1u; + TimingResult r = runTimed(warmupDispatches, kPilotN, bindOnce, dispatchOne, dispatchesPerSubmit); + dispatchesPerSubmit = estimateDispatchesPerSubmit(r, kPilotN); + uint32_t lastN = kPilotN; + while (r.elapsed_ns > targetBudgetNs && lastN > 1u) + { + const double scale = double(targetBudgetNs) / r.elapsed_ns; + uint32_t nextN = uint32_t(std::max(1.0, std::floor(double(lastN) * scale))); + if (nextN >= lastN) + nextN = lastN - 1u; + + r = runTimed(warmupDispatches, nextN, bindOnce, dispatchOne, dispatchesPerSubmit); + dispatchesPerSubmit = estimateDispatchesPerSubmit(r, nextN); + lastN = nextN; + } + + while (r.elapsed_ns < targetBudgetNs && lastN < kMaxN) + { + uint32_t nextN; + if (r.elapsed_ns > 1'000'000ull) // > 1 ms, stable enough to scale + { + const double scale = double(targetBudgetNs) / double(r.elapsed_ns); + nextN = uint32_t(std::min(double(kMaxN), std::ceil(double(lastN) * scale))); + } + else + { + nextN = std::min(kMaxN, lastN * 2); + } + if (nextN <= lastN) + break; // converged + r = runTimed(warmupDispatches, nextN, bindOnce, dispatchOne, dispatchesPerSubmit); + dispatchesPerSubmit = estimateDispatchesPerSubmit(r, nextN); + lastN = nextN; + } + + if (samples <= 1) + return r; + + // Reuse the convergence's final measurement as one of the K samples + // (it's already a budget-sized window at lastN). Run K-1 more at the + // same N. All windows measure the same dispatch count, so the per-window + // elapsed_ns values are directly comparable. + std::vector ns; + ns.reserve(samples); + ns.push_back(r.elapsed_ns); + for (uint32_t i = 1; i < samples; ++i) + { + const TimingResult ri = runTimed(warmupDispatches, lastN, bindOnce, dispatchOne, dispatchesPerSubmit); + ns.push_back(ri.elapsed_ns); + } + std::sort(ns.begin(), ns.end()); + + // Outlier rejection: GPU jitter is usually a one-sided spike + const double median = ns[ns.size() / 2]; + const double dLow = median - ns.front(); + const double dHigh = ns.back() - median; + const double dCloser = std::min(dLow, dHigh); + const double dFar = std::max(dLow, dHigh); + size_t lo = 0; + size_t hi = ns.size(); + if (dCloser > 0.0 && dFar > 2.0 * dCloser) + { + if (dHigh > dLow) + --hi; // top sample is the spike + else + ++lo; // bottom sample is the spike (rare on GPU but cheap to handle) + } + + double sum = 0.0; + for (size_t i = lo; i < hi; ++i) + sum += ns[i]; + const double resultNs = sum / double(hi - lo); + + TimingResult m {}; + m.elapsed_ns = resultNs; + m.totalSamples = uint64_t(lastN) * m_samplesPerDispatch; + m.ps_per_sample = m.totalSamples ? resultNs * 1e3 / double(m.totalSamples) : 0.0; + m.gsamples_per_s = resultNs > 0.0 ? double(m.totalSamples) / resultNs : 0.0; + m.ms_total = resultNs * 1e-6; + return m; + } + + TimingResult runTimed(uint32_t warmupDispatches, uint32_t benchDispatches, const DispatchFn& bindOnce, const DispatchFn& dispatchOne, uint32_t maxDispatchesPerSubmit) + { + if (m_device->waitIdle() != nbl::video::IQueue::RESULT::SUCCESS) + return {}; + + const uint32_t cooldownDispatches = warmupDispatches; + + if (!runUntimedDispatches(warmupDispatches, bindOnce, dispatchOne, maxDispatchesPerSubmit)) + return {}; + + double elapsedNs = 0.0; + uint32_t remaining = benchDispatches; + while (remaining > 0u) + { + const uint32_t batch = std::min(remaining, std::max(1u, maxDispatchesPerSubmit)); + + m_cmdbuf->reset(nbl::video::IGPUCommandBuffer::RESET_FLAGS::NONE); + m_cmdbuf->begin(nbl::video::IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); + m_cmdbuf->resetQueryPool(m_queryPool.get(), 0, 2); + + if (bindOnce) + bindOnce(m_cmdbuf.get()); + + m_cmdbuf->writeTimestamp(nbl::asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT, m_queryPool.get(), 0); + for (uint32_t i = 0u; i < batch; ++i) + dispatchOne(m_cmdbuf.get()); + m_cmdbuf->writeTimestamp(nbl::asset::PIPELINE_STAGE_FLAGS::COMPUTE_SHADER_BIT, m_queryPool.get(), 1); + m_cmdbuf->end(); + + if (!submitAndWait()) + return {}; + + uint64_t timestamps[2] = {}; + const auto flags = nbl::core::bitflag(nbl::video::IQueryPool::RESULTS_FLAGS::_64_BIT) | nbl::core::bitflag(nbl::video::IQueryPool::RESULTS_FLAGS::WAIT_BIT); + if (!m_device->getQueryPoolResults(m_queryPool.get(), 0, 2, timestamps, sizeof(uint64_t), flags)) + return {}; + + const double timestampPeriod = double(m_physicalDevice->getLimits().timestampPeriodInNanoSeconds); + elapsedNs += double(timestamps[1] - timestamps[0]) * timestampPeriod; + remaining -= batch; + } + + if (!runUntimedDispatches(cooldownDispatches, bindOnce, dispatchOne, maxDispatchesPerSubmit)) + return {}; + + TimingResult r {}; + r.elapsed_ns = elapsedNs; + r.totalSamples = uint64_t(benchDispatches) * m_samplesPerDispatch; + r.ps_per_sample = r.totalSamples ? r.elapsed_ns * 1e3 / double(r.totalSamples) : 0.0; + r.gsamples_per_s = r.elapsed_ns > 0.0 ? double(r.totalSamples) / r.elapsed_ns : 0.0; + r.ms_total = r.elapsed_ns * 1e-6; + return r; + } + +protected: + static constexpr uint32_t InvalidPipelineIndex = std::numeric_limits::max(); + + const PipelineEntry* getPipelineEntry(uint32_t idx, std::string_view context) const + { + if (idx == InvalidPipelineIndex || idx >= m_pipelines.size() || !m_pipelines[idx].pipeline) + { + benchLogFmt(m_logger.get(), nbl::system::ILogger::ELL_ERROR, "{}: pipeline is not available", context); + return nullptr; + } + return &m_pipelines[idx]; + } + + std::vector m_pipelines; + +private: + // Soft target for one queue submit, estimated from timings on the current GPU. + // Benchmark budgets still control measured work. This only chunks submits. + static constexpr double SubmitChunkTargetNs = 250'000'000.0; + + static uint32_t estimateDispatchesPerSubmit(const TimingResult& r, uint32_t dispatches) + { + if (dispatches == 0u || r.elapsed_ns <= 0.0) + return 1u; + + const double nsPerDispatch = r.elapsed_ns / double(dispatches); + if (nsPerDispatch <= 0.0) + return 1u; + + const double maxDispatches = std::floor(SubmitChunkTargetNs / nsPerDispatch); + return uint32_t(std::clamp(maxDispatches, 1.0, double(std::numeric_limits::max()))); + } + + bool submitAndWait() + { + auto semaphore = m_device->createSemaphore(0u); + if (!semaphore) + return false; + + const nbl::video::IQueue::SSubmitInfo::SCommandBufferInfo cmds[] = {{.cmdbuf = m_cmdbuf.get()}}; + const nbl::video::IQueue::SSubmitInfo::SSemaphoreInfo done[] = { + {.semaphore = semaphore.get(), .value = 1u, .stageMask = nbl::asset::PIPELINE_STAGE_FLAGS::ALL_COMMANDS_BITS}}; + nbl::video::IQueue::SSubmitInfo submit = {}; + submit.commandBuffers = cmds; + submit.signalSemaphores = done; + if (m_queue->submit({&submit, 1u}) != nbl::video::IQueue::RESULT::SUCCESS) + return false; + + const nbl::video::ISemaphore::SWaitInfo wait[] = {{.semaphore = semaphore.get(), .value = 1u}}; + return m_device->blockForSemaphores(wait) == nbl::video::ISemaphore::WAIT_RESULT::SUCCESS; + } + + bool runUntimedDispatches(uint32_t dispatches, const DispatchFn& bindOnce, const DispatchFn& dispatchOne, uint32_t maxDispatchesPerSubmit) + { + while (dispatches > 0u) + { + const uint32_t batch = std::min(dispatches, std::max(1u, maxDispatchesPerSubmit)); + + m_cmdbuf->reset(nbl::video::IGPUCommandBuffer::RESET_FLAGS::NONE); + m_cmdbuf->begin(nbl::video::IGPUCommandBuffer::USAGE::ONE_TIME_SUBMIT_BIT); + if (bindOnce) + bindOnce(m_cmdbuf.get()); + for (uint32_t i = 0u; i < batch; ++i) + dispatchOne(m_cmdbuf.get()); + m_cmdbuf->end(); + + if (!submitAndWait()) + return false; + dispatches -= batch; + } + return true; + } + + static void matchStat(const nbl::video::IGPUPipelineBase::SExecutableStatistic& stat, PipelineStats& out, uint64_t& vgpr, uint64_t& sgpr) + { + const uint64_t v = stat.asUint(); + + auto contains = [&](std::string_view kw) + { + const auto it = std::ranges::search(stat.name, kw, + [&](char a, char b) + { return std::tolower(a) == std::tolower(b); }) + .begin(); + return it != stat.name.end(); + }; + + // Order matters: more specific keys first. + + if (contains("subgroup size") || contains("subgroupsize") || contains("warp size") || contains("wave size")) + out.subgroupSize = std::max(out.subgroupSize, uint32_t(v)); + + else if (contains("vgpr")) + vgpr = std::max(vgpr, v); + else if (contains("sgpr")) + sgpr = std::max(sgpr, v); + else if (contains("register")) + out.registerCount = std::max(out.registerCount, v); + + else if (contains("binary size") || contains("binarysize") || contains("codesize") || contains("code size") || contains("isa size")) + out.codeSizeBytes = std::max(out.codeSizeBytes, v); + else if (contains("instructioncount") || contains("instruction count") || contains("numinstructions")) + out.codeSizeBytes = std::max(out.codeSizeBytes, v); // proxy when no byte size + + else if (contains("shared memory") || contains("sharedmemory") || contains("groupshared") || contains("lds")) + out.sharedMemBytes = std::max(out.sharedMemBytes, v); + + else if (contains("stack size") || contains("stacksize")) + out.stackBytes = std::max(out.stackBytes, v); + + else if (contains("local memory") || contains("localmemory") || contains("scratch") || contains("private memory") || contains("privatememory") || contains("stack")) + out.privateMemBytes = std::max(out.privateMemBytes, v); + + // Vendor-specific stats + // get a structured copy so JSON round-trips the right numeric type. + else + out.unknowns.push_back(stat); + } + + nbl::core::smart_refctd_ptr m_device; + nbl::core::smart_refctd_ptr m_logger; + nbl::video::IPhysicalDevice* m_physicalDevice = nullptr; + nbl::video::IQueue* m_queue = nullptr; + nbl::hlsl::uint32_t3 m_dispatchGroupCount {}; + uint64_t m_samplesPerDispatch = 0; + nbl::core::smart_refctd_ptr m_cmdpool; + nbl::core::smart_refctd_ptr m_cmdbuf; + nbl::core::smart_refctd_ptr m_queryPool; + nbl::core::smart_refctd_ptr m_utils; // lazy, only built on first createBdaBuffer call +}; + +#endif diff --git a/common/include/nbl/examples/Benchmark/IBenchmark.h b/common/include/nbl/examples/Benchmark/IBenchmark.h new file mode 100644 index 000000000..93493c2c6 --- /dev/null +++ b/common/include/nbl/examples/Benchmark/IBenchmark.h @@ -0,0 +1,409 @@ +// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O. +// This file is part of the "Nabla Engine". +// For conditions of distribution and use, see copyright notice in nabla.h + +#ifndef _NBL_COMMON_I_BENCHMARK_INCLUDED_ +#define _NBL_COMMON_I_BENCHMARK_INCLUDED_ + +#include +#include "nbl/examples/Benchmark/BenchmarkTypes.h" +#include "nbl/examples/Benchmark/BenchmarkConsole.h" +#include "nbl/examples/Benchmark/GPUBenchmarkHelper.h" +#include "nbl/examples/Benchmark/BenchmarkJson.h" +#include "nbl/examples/Benchmark/BenchmarkCli.h" +#include "nlohmann/json.hpp" + +#include +#include +#include +#include +#include +#include +#include +#include + + +struct RunContext +{ + WorkloadShape shape; + uint64_t targetBudgetMs = 400; // wall-clock budget per row + std::string sectionLabel = "Benchmarks"; +}; + +// Typical use: +// +// Aggregator agg(logger, logicalDevice, physicalDevice, computeFamilyIndex); +// agg.applyCli({.argv = argv, .defaultOutputPath = "Bench.json"}); +// const RunContext myCtx{.shape = ..., .targetBudgetMs = 400, .sectionLabel = "..."}; +// std::vector benches; +// for (...) benches.emplace_back(agg, MyBench::SetupData{...}); +// MyOtherBench other(agg, MyOtherBench::SetupData{...}); +// agg.runSessionAndReport( +// Aggregator::Span{std::span(benches), myCtx}, +// Aggregator::Span{std::span(&other, 1), otherCtx}); +class Aggregator +{ + friend class IBenchmark; + +public: + Aggregator() = default; + + Aggregator(nbl::core::smart_refctd_ptr logger, + nbl::core::smart_refctd_ptr logicalDevice, + nbl::video::IPhysicalDevice* physicalDevice, + uint32_t computeFamilyIndex) + { + m_console.setLogger(std::move(logger)); + m_logicalDevice = std::move(logicalDevice); + m_physicalDevicePtr = physicalDevice; + m_computeFamilyIndex = computeFamilyIndex; + setDevice(physicalDevice); + } + + void setSilent(bool silent) { m_console.setSilent(silent); } + + const nbl::core::smart_refctd_ptr& getLogicalDevice() const { return m_logicalDevice; } + nbl::video::IPhysicalDevice* getPhysicalDevice() const { return m_physicalDevicePtr; } + uint32_t getComputeFamilyIndex() const { return m_computeFamilyIndex; } + nbl::core::smart_refctd_ptr getLogger() const + { + return nbl::core::smart_refctd_ptr(m_console.getLogger()); + } + + bool loadBaseline(std::string label, const std::string& path) + { + auto b = benchmark_json::loadBaselineFile(label, path); + if (!b) + return false; + + for (const auto& [_, row] : b->rowsByName) + m_console.growForBaseline(row); + + // Vector (not map) so delta columns print in load order. + auto it = std::find_if(m_baselines.begin(), m_baselines.end(), + [&](const Baseline& existing) { return existing.label == label; }); + if (it != m_baselines.end()) + *it = std::move(*b); + else + m_baselines.push_back(std::move(*b)); + return true; + } + + bool loadBaseline(const std::string& path) { return loadBaseline("baseline", path); } + + bool writeReport(const std::string& path) + { + size_t preservedCount = 0; + if (!benchmark_json::writeReportFile(path, m_device, m_baselines, m_results, m_console.getLogger(), &preservedCount)) + return false; + + if (preservedCount > 0) + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_INFO, + "Wrote benchmark report to {} ({} new + {} preserved from prior file)", + path, m_results.size(), preservedCount); + else + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_INFO, + "Wrote benchmark report to {} ({} rows)", path, m_results.size()); + return true; + } + + // Captured for the UUID-mismatch warning in applyCli. + void setDevice(const nbl::video::IPhysicalDevice* dev) { m_device = benchmark_json::buildDeviceMetadata(dev); } + + struct CliResult + { + std::string outputPath; + nbl::core::vector> focusVariants; + uint32_t focusSamples = 3; // --focus-samples, see samplesForCurrentRow + + bool isFocused(const nbl::core::vector& name) const + { + return std::ranges::find(focusVariants, name) != focusVariants.end(); + } + }; + + template + struct Span + { + std::span benches; + RunContext context; + }; + + // Two overloads so a single bench doesn't need `std::span(&bench, 1)`. + template + requires requires (Range& r) { std::data(r); std::size(r); } + static auto makeSpan(Range& benches, RunContext context) + { + using T = std::remove_reference_t; + return Span{std::span(std::data(benches), std::size(benches)), std::move(context)}; + } + + template + requires std::derived_from + static Span makeSpan(T& bench, RunContext context) + { + return Span{std::span(&bench, 1), std::move(context)}; + } + + static std::string describe(const RunContext& ctx) + { + const auto& sh = ctx.shape; + const uint32_t wgThreads = sh.workgroupSize.x * sh.workgroupSize.y * sh.workgroupSize.z; + const uint32_t threadsPerDisp = sh.dispatchGroupCount.x * sh.dispatchGroupCount.y * sh.dispatchGroupCount.z * wgThreads; + const uint64_t itersPerThread = threadsPerDisp ? sh.samplesPerDispatch / threadsPerDisp : 0; + const double budgetMs = double(ctx.targetBudgetMs); + return std::format("=== {} (~{:.0f}ms/row, {} threads/dispatch, {} iters/thread; wg={}x{}x{}; ps/sample is per all GPU threads) ===", + ctx.sectionLabel, budgetMs, threadsPerDisp, itersPerThread, sh.workgroupSize.x, sh.workgroupSize.y, sh.workgroupSize.z); + } + + // Order: banner -> focus(spans...) -> comparison table -> banner -> + // column header -> rest(spans...) -> writeReport. + // All focus rows print globally first, then all rest rows; banner printed + // twice so each chunk reads in isolation when scrolling back. + template + requires(std::derived_from && ...) + void runSessionAndReport(Span... spans) + { + // Templated lambda (not `auto& s`) so only Span deduces -- a future + // signature change can't silently start passing arbitrary types through. + auto runSpan = [this](Span& s, bool silent) + { + if (s.benches.empty()) + return; + if (!silent) + { + m_console.logSectionBanner(describe(s.context)); + m_console.logHeader(m_baselines); + } + for (auto& e : s.benches) + e.run(); + // Flush after each rest span: if span N+1 dies mid-way, span N's + // rows are already on disk. Trailing flush is also the final write. + if (!silent) + writeReport(m_cli.outputPath); + }; + + m_console.logBannerNotes(m_baselines); + if (!m_cli.focusVariants.empty()) + { + m_console.setSilent(true); // benches read this to know they're in the focused-rows half + (runSpan(spans, true), ...); + m_console.setSilent(false); + m_console.printBaselineComparison(std::span>(m_focusNames), m_baselines, m_results); + } + (runSpan(spans, false), ...); + } + + struct CliConfig + { + std::span argv; // feed from IApplicationFramework::argv + std::string defaultOutputPath = "Bench.json"; + std::string appName = "benchmark"; + }; + + CliResult applyCli(const CliConfig& cfg) + { + auto parsed = benchmark_cli::parseArgs(cfg.argv, cfg.defaultOutputPath); + if (parsed.helpRequested) + { + benchmark_cli::printHelp(m_console.getLogger(), cfg.appName, cfg.defaultOutputPath); + exit(0); + } + if (parsed.noColor) + m_console.setColorEnabled(false); + + CliResult res; + res.outputPath = parsed.outputPath; + + if (!parsed.baselines.empty()) + { + size_t succeeded = 0; + for (const auto& [label, path] : parsed.baselines) + { + if (loadBaseline(label, path)) + { + ++succeeded; + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_INFO, + "Loaded baseline '{}' from {} ({} rows)", label, path, m_baselines.back().rowsByName.size()); + } + else + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_WARNING, + "Failed to load baseline '{}' from {}, skipped", label, path); + } + if (succeeded == 0) + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_WARNING, + "All {} --baseline load(s) failed. delta columns and --focus will be empty. " + "Check the paths above; default auto-load of '{}' is suppressed once any --baseline is specified, " + "drop the --baseline flag(s) or use --no-baseline to silence this warning.", + parsed.baselines.size(), res.outputPath); + else if (succeeded < parsed.baselines.size()) + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_WARNING, + "{} of {} --baseline load(s) failed; continuing with {} loaded.", + parsed.baselines.size() - succeeded, parsed.baselines.size(), succeeded); + } + else if (!parsed.noBaseline) + { + if (loadBaseline(res.outputPath)) + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_INFO, + "Loaded baseline from {} ({} rows)", res.outputPath, + m_baselines.empty() ? size_t {0} : m_baselines.back().rowsByName.size()); + else + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_INFO, + "No baseline at {}, delta column will read 'n/a'", res.outputPath); + } + + warnDeviceMismatch(); + + res.focusVariants = std::move(parsed.focus); + res.focusSamples = parsed.focusSamples; + m_cli = res; + return res; + } + +private: + void warnDeviceMismatch() const + { + if (!m_device.is_object() || !m_device.contains("deviceUUID")) + return; + const auto& currentUUID = m_device["deviceUUID"]; + for (const auto& b : m_baselines) + { + if (!b.device.is_object() || !b.device.contains("deviceUUID")) + continue; + if (b.device["deviceUUID"] == currentUUID) + continue; + const std::string baselineDevName = b.device.value("name", std::string {""}); + const std::string currentDevName = m_device.value("name", std::string {""}); + benchLogFmt(m_console.getLogger(), nbl::system::ILogger::ELL_WARNING, + "Baseline '{}' (from {}) was measured on a different GPU ('{}' vs current '{}'). " + "Delta values will be apples-to-oranges.", + b.label, b.path, baselineDevName, currentDevName); + } + } + + // In focus phase (silent), captures the row's name into m_focusNames so + // runSessionAndReport can build the comparison table without main.cpp + // threading names back through each bench class. + void appendAndLog(Result&& r) + { + const std::string joined = joinName(r.name); + if (!m_baselines.empty()) + { + const std::string key = makeKey(r.name); + for (const auto& b : m_baselines) + { + auto it = b.rowsByName.find(key); + if (it == b.rowsByName.end()) + continue; + const bool shapeMismatch = r.workload.present() && it->second.workload.present() && (r.workload.shape != it->second.workload.shape); + r.baselines[b.label] = {it->second.psPerSample, shapeMismatch}; + } + } + m_console.growWidthFor(joined); + if (m_console.silent()) + m_focusNames.push_back(r.name); + m_results.push_back(std::move(r)); + m_console.logRow(std::span(m_results.back().name), joined, m_results.back().timing, m_results.back().stats, m_results.back().baselines, m_baselines); + } + + std::vector m_results; + std::vector m_baselines; + nbl::core::vector> m_focusNames; + nlohmann::json m_device; + CliResult m_cli; + BenchmarkConsole m_console; + nbl::core::smart_refctd_ptr m_logicalDevice; + nbl::video::IPhysicalDevice* m_physicalDevicePtr = nullptr; + uint32_t m_computeFamilyIndex = 0; +}; + +class IBenchmark +{ +public: + virtual ~IBenchmark() = default; + + // Single-named benches override doRun() and inherit this default filter. + // Sweep-style benches synthesize per-row names; they override run() and + // do per-row filtering themselves. + virtual void run() + { + const bool silent = isFocusPhase(); + const bool inFocus = isFocused(m_name); + const bool shouldRun = silent ? inFocus : !inFocus; + if (shouldRun) + doRun(); + } + + uint32_t getWarmupDispatches() const { return m_warmupDispatches; } + uint64_t getTargetBudgetMs() const { return m_targetBudgetMs; } + const WorkloadShape& getShape() const { return m_workloadShape; } + + // Pass this to runTimedBudgeted so only --focus rows pay the K * budget cost. + uint32_t samplesForCurrentRow() const { return isFocusPhase() ? m_aggregator.m_cli.focusSamples : 1u; } + +protected: + // Banner label is NOT taken here; it belongs to the span (see Aggregator::Span). + IBenchmark(Aggregator& aggregator, core::vector name, uint32_t warmupDispatches, const WorkloadShape& shape, uint64_t targetBudgetMs) + : m_name(std::move(name)) + , m_aggregator(aggregator) + , m_warmupDispatches(warmupDispatches) + , m_targetBudgetMs(targetBudgetMs) + , m_workloadShape(shape) + { + registerVariant(m_name); + } + + virtual void doRun() {} + + bool isFocusPhase() const { return m_aggregator.m_console.silent(); } + bool isFocused(const core::vector& name) const { return m_aggregator.m_cli.isFocused(name); } + void registerVariant(std::span name) { m_aggregator.m_console.registerVariant(name); } + void registerVariant(std::initializer_list name) { m_aggregator.m_console.registerVariant(name); } + + void record(core::vector name, const TimingResult& t, const PipelineStats& s) + { + Workload w{.shape = m_workloadShape}; + w.benchDispatches = w.shape.samplesPerDispatch ? uint32_t(t.totalSamples / w.shape.samplesPerDispatch) : 0; + + Result r; + r.name = std::move(name); + r.timing = t; + r.stats = s; + r.workload = w; + m_aggregator.appendAndLog(std::move(r)); + } + + core::vector m_name; + Aggregator& m_aggregator; // non-owning, outlives this bench + uint32_t m_warmupDispatches; + uint64_t m_targetBudgetMs; + WorkloadShape m_workloadShape; +}; + +class GPUBenchmark : public IBenchmark, public GPUBenchmarkHelper +{ +public: + struct SetupData + { + core::vector name; + uint32_t warmupDispatches = 0; + WorkloadShape shape = {}; + uint64_t targetBudgetMs = 400; + }; + +protected: + GPUBenchmark(Aggregator& aggregator, const SetupData& data) + : IBenchmark(aggregator, data.name, data.warmupDispatches, data.shape, data.targetBudgetMs) + { + GPUBenchmarkHelper::init({ + .device = aggregator.getLogicalDevice(), + .logger = aggregator.getLogger(), + .physicalDevice = aggregator.getPhysicalDevice(), + .computeFamilyIndex = aggregator.getComputeFamilyIndex(), + .dispatchGroupCount = data.shape.dispatchGroupCount, + .samplesPerDispatch = data.shape.samplesPerDispatch, + }); + } +}; + +#endif diff --git a/common/include/nbl/examples/Tester/FailureManifest.h b/common/include/nbl/examples/Tester/FailureManifest.h new file mode 100644 index 000000000..a703e933e --- /dev/null +++ b/common/include/nbl/examples/Tester/FailureManifest.h @@ -0,0 +1,331 @@ +#ifndef _NBL_COMMON_TESTER_FAILURE_MANIFEST_INCLUDED_ +#define _NBL_COMMON_TESTER_FAILURE_MANIFEST_INCLUDED_ + +#include + +#include "nlohmann/json.hpp" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace nbl::examples::testing +{ + +struct FailureCase +{ + std::string check; + std::string side; + uint64_t iteration = 0; + uint32_t seed = 0; + double maxRelative = 0.0; + double maxAbsolute = 0.0; +}; + +struct FailureGroup +{ + std::string phase; + std::string id; + std::string name; + std::string logFile; + std::vector cases; + uint32_t omittedCases = 0; +}; + +class FailureManifest +{ + public: + explicit FailureManifest(std::string suite = {}) : m_suite(std::move(suite)) {} + + void setSuite(std::string suite) { m_suite = std::move(suite); } + + void addGroupFailure(std::string_view phase, std::string_view id, std::string_view name, std::string_view logFile = {}) + { + auto& group = groupFor(phase, id, name); + if (!logFile.empty()) + group.logFile = std::string(logFile); + } + + void addCase(std::string_view phase, std::string_view id, std::string_view name, std::string_view check, std::string_view side, + uint64_t iteration, uint32_t seed, double maxRelative, double maxAbsolute) + { + auto& group = groupFor(phase, id, name); + if (group.cases.size() >= MaxCasesPerGroup) + { + ++group.omittedCases; + return; + } + + group.cases.push_back(FailureCase{ + .check = std::string(check), + .side = std::string(side), + .iteration = iteration, + .seed = seed, + .maxRelative = maxRelative, + .maxAbsolute = maxAbsolute, + }); + } + + const std::vector& failures() const { return m_failures; } + + nlohmann::json toJson() const + { + nlohmann::json doc; + doc["version"] = 1; + doc["suite"] = m_suite; + auto& failures = doc["failures"] = nlohmann::json::array(); + + for (const auto& group : m_failures) + { + nlohmann::json g; + g["phase"] = group.phase; + g["id"] = group.id; + g["name"] = group.name; + if (!group.logFile.empty()) + g["log_file"] = group.logFile; + + auto& cases = g["cases"] = nlohmann::json::array(); + for (const auto& c : group.cases) + { + nlohmann::json row; + row["check"] = c.check; + row["side"] = c.side; + row["iteration"] = c.iteration; + row["seed"] = c.seed; + row["max_relative"] = c.maxRelative; + row["max_absolute"] = c.maxAbsolute; + cases.push_back(std::move(row)); + } + + if (group.omittedCases > 0) + g["omitted_cases"] = group.omittedCases; + + failures.push_back(std::move(g)); + } + + return doc; + } + + private: + static constexpr size_t MaxCasesPerGroup = 64; + + FailureGroup& groupFor(std::string_view phase, std::string_view id, std::string_view name) + { + const std::string idString(id); + auto it = std::find_if(m_failures.begin(), m_failures.end(), [&](const FailureGroup& g) { return g.id == idString; }); + if (it != m_failures.end()) + { + if (it->name.empty()) + it->name = std::string(name); + if (it->phase.empty()) + it->phase = std::string(phase); + return *it; + } + + m_failures.push_back(FailureGroup{ + .phase = std::string(phase), + .id = idString, + .name = std::string(name), + }); + return m_failures.back(); + } + + std::string m_suite; + std::vector m_failures; +}; + +class TestFilter +{ + public: + bool enabled() const { return m_enabled; } + + void enable() { m_enabled = true; } + + bool shouldRun(std::string_view id) const + { + return !m_enabled || m_ids.contains(std::string(id)); + } + + void add(std::string_view id) + { + m_enabled = true; + const auto first = id.find_first_not_of(" \t\r\n"); + if (first == std::string_view::npos) + return; + const auto last = id.find_last_not_of(" \t\r\n"); + m_ids.insert(std::string(id.substr(first, last - first + 1))); + } + + void addSeed(std::string_view id, uint32_t seed) + { + add(id); + m_seeds[std::string(id)] = seed; + } + + void addList(std::string_view ids) + { + m_enabled = true; + while (!ids.empty()) + { + const auto comma = ids.find(','); + add(ids.substr(0, comma)); + if (comma == std::string_view::npos) + return; + ids.remove_prefix(comma + 1); + } + } + + std::optional seedFor(std::string_view id) const + { + auto it = m_seeds.find(std::string(id)); + if (it == m_seeds.end()) + return {}; + return it->second; + } + + private: + bool m_enabled = false; + std::set m_ids; + std::map m_seeds; +}; + +struct RunControl +{ + bool valid = true; + bool skipBenchmarks = false; + std::string failedOutPath; + TestFilter filter; +}; + +inline bool addFailedIdsFromFile(TestFilter& filter, const std::string& path, nbl::system::ILogger* logger) +{ + filter.enable(); + std::ifstream in(path); + if (!in.is_open()) + { + if (logger) + logger->log("Failed to open failed-test manifest '%s'", nbl::system::ILogger::ELL_ERROR, path.c_str()); + return false; + } + + nlohmann::json doc; + try + { + in >> doc; + } + catch (const std::exception& e) + { + if (logger) + logger->log("Failed to parse failed-test manifest '%s': %s", nbl::system::ILogger::ELL_ERROR, path.c_str(), e.what()); + return false; + } + + const auto failuresIt = doc.find("failures"); + if (failuresIt == doc.end() || !failuresIt->is_array()) + { + if (logger) + logger->log("Failed-test manifest '%s' does not contain a failures array", nbl::system::ILogger::ELL_ERROR, path.c_str()); + return false; + } + + for (const auto& failure : *failuresIt) + { + if (!failure.is_object()) + continue; + const auto idIt = failure.find("id"); + if (idIt != failure.end() && idIt->is_string()) + { + const std::string id = idIt->get(); + const auto casesIt = failure.find("cases"); + if (casesIt != failure.end() && casesIt->is_array()) + { + const auto seedIt = std::find_if(casesIt->begin(), casesIt->end(), [](const nlohmann::json& row) { + if (!row.is_object()) + return false; + const auto it = row.find("seed"); + return it != row.end() && it->is_number_integer(); + }); + if (seedIt != casesIt->end()) + { + filter.addSeed(id, (*seedIt)["seed"].get()); + continue; + } + } + filter.add(id); + } + } + + return true; +} + +inline RunControl parseRunControl(std::span argv, nbl::system::ILogger* logger) +{ + RunControl out; + + for (size_t i = 1; i < argv.size(); ++i) + { + const std::string& arg = argv[i]; + if (arg == "--skip-benchmarks") + out.skipBenchmarks = true; + else if (arg == "--failed-out" && i + 1 < argv.size()) + out.failedOutPath = argv[++i]; + else if (arg.starts_with("--failed-out=")) + out.failedOutPath = arg.substr(std::string("--failed-out=").size()); + else if (arg == "--test" && i + 1 < argv.size()) + out.filter.addList(argv[++i]); + else if (arg.starts_with("--test=")) + out.filter.addList(std::string_view(arg).substr(std::string_view("--test=").size())); + else if (arg == "--rerun-failed" && i + 1 < argv.size()) + { + if (!addFailedIdsFromFile(out.filter, argv[++i], logger)) + out.valid = false; + } + else if (arg.starts_with("--rerun-failed=")) + { + if (!addFailedIdsFromFile(out.filter, arg.substr(std::string("--rerun-failed=").size()), logger)) + out.valid = false; + } + } + + if (out.filter.enabled()) + out.skipBenchmarks = true; + + return out; +} + +inline bool writeFailureManifestFile(const FailureManifest& manifest, const std::string& path, nbl::system::ILogger* logger) +{ + std::ofstream out(path, std::ios::out | std::ios::trunc); + if (!out.is_open()) + { + if (logger) + logger->log("Failed to open failed-test manifest '%s' for writing", nbl::system::ILogger::ELL_ERROR, path.c_str()); + return false; + } + + out << manifest.toJson().dump(3) << '\n'; + if (!out.good()) + { + if (logger) + logger->log("Failed to write failed-test manifest '%s'", nbl::system::ILogger::ELL_ERROR, path.c_str()); + return false; + } + + if (logger) + logger->log("Wrote failed-test manifest '%s' with %llu failed groups", nbl::system::ILogger::ELL_INFO, + path.c_str(), static_cast(manifest.failures().size())); + return true; +} + +} // namespace nbl::examples::testing + +#endif diff --git a/common/include/nbl/examples/Tester/ITester.h b/common/include/nbl/examples/Tester/ITester.h index 8fd4c6639..bdb85fa82 100644 --- a/common/include/nbl/examples/Tester/ITester.h +++ b/common/include/nbl/examples/Tester/ITester.h @@ -3,6 +3,7 @@ #include #include +#include #include #include #include @@ -169,40 +170,43 @@ class ITester m_queue = m_device->getQueue(m_queueFamily, 0); } + /** + * @brief Runs tests and verifies their results using the current seed (a fresh random seed is generated for the next call). + * + * @param logFileName Name of the file where test logs will be saved. + * @return true if all tests pass and results are valid, false otherwise. + */ bool performTestsAndVerifyResults(const std::string& logFileName) { - m_logFile.open(logFileName, std::ios::out | std::ios::trunc); - if (!m_logFile.is_open()) - m_logger->log("Failed to open log file!", system::ILogger::ELL_ERROR); - - core::vector inputTestValues; - core::vector exceptedTestResults; - - inputTestValues.reserve(m_testIterationCount); - exceptedTestResults.reserve(m_testIterationCount); - - m_logger->log("TESTS:", system::ILogger::ELL_PERFORMANCE); - for (int i = 0; i < m_testIterationCount; ++i) - { - // Set input thest values that will be used in both CPU and GPU tests - InputTestValues testInput = generateInputTestValues(); - // use std library or glm functions to determine expected test values, the output of functions from intrinsics.hlsl will be verified against these values - TestResults expected = determineExpectedResults(testInput); - - inputTestValues.push_back(testInput); - exceptedTestResults.push_back(expected); - } - - core::vector cpuTestResults = performCpuTests(inputTestValues); - core::vector gpuTestResults = performGpuTests(inputTestValues); + reloadSeed(); + return performTestsAndVerifyResults_impl(logFileName); + } - bool pass = verifyAllTestResults(cpuTestResults, gpuTestResults, exceptedTestResults); + /** + * @brief Runs tests and verifies their results using a user-provided seed for test value generation. + * + * @param logFileName Name of the file where test logs will be saved. + * @param seed Custom seed used for generating test values, ensures deterministic and reproducible results. + * @return true if all tests pass and results are valid, false otherwise. + */ + bool performTestsAndVerifyResults(const std::string& logFileName, const uint32_t seed) + { + setSeed(seed); + return performTestsAndVerifyResults_impl(logFileName); + } - m_logger->log("TESTS DONE.", system::ILogger::ELL_PERFORMANCE); - reloadSeed(); + void setFailureRecordContext(nbl::examples::testing::FailureManifest* manifest, std::string phase, std::string id, std::string name) + { + m_failureManifest = manifest; + m_failurePhase = std::move(phase); + m_failureId = std::move(id); + m_failureName = std::move(name); + } - m_logFile.close(); - return pass; + void setSeed(uint32_t seed) + { + m_seed = seed; + m_mersenneTwister = std::mt19937(m_seed); } virtual ~ITester() @@ -223,7 +227,6 @@ class ITester ITester(const uint32_t testBatchCount, const uint32_t workgroupSize = 256) : m_WorkgroupSize(workgroupSize), m_testBatchCount(testBatchCount), m_testIterationCount(testBatchCount * m_WorkgroupSize) { - reloadSeed(); }; virtual bool verifyTestResults(const TestResults& expectedTestValues, const TestResults& testValues, const size_t testIteration, const uint32_t seed, TestType testType) = 0; @@ -339,10 +342,55 @@ class ITester ss << " DIFFERENCE: " << system::to_string(hlsl::abs(expectedVal - testVal)); ss << " MAX RELATIVE: " << system::to_string(maxRelativeDifference) << " MAX ABSOLUTE " << system::to_string(maxAbsoluteDifference) << '\n'; + if (m_failureManifest) + { + const char* side = testType == TestType::CPU ? "CPU" : "GPU"; + m_failureManifest->addCase(m_failurePhase, m_failureId, m_failureName, memberName, side, + testIteration, seed, maxRelativeDifference, maxAbsoluteDifference); + } + m_logger->log("%s", system::ILogger::ELL_ERROR, ss.str().c_str()); m_logFile << ss.str() << '\n'; } + bool performTestsAndVerifyResults_impl(const std::string& logFileName) + { + m_failureLogFile = logFileName; + m_logFile.open(logFileName, std::ios::out | std::ios::trunc); + if (!m_logFile.is_open()) + m_logger->log("Failed to open log file!", system::ILogger::ELL_ERROR); + + core::vector inputTestValues; + core::vector exceptedTestResults; + + inputTestValues.reserve(m_testIterationCount); + exceptedTestResults.reserve(m_testIterationCount); + + m_logger->log("TESTS:", system::ILogger::ELL_PERFORMANCE); + for (int i = 0; i < m_testIterationCount; ++i) + { + // Set input thest values that will be used in both CPU and GPU tests + InputTestValues testInput = generateInputTestValues(); + // use std library or glm functions to determine expected test values, the output of functions from intrinsics.hlsl will be verified against these values + TestResults expected = determineExpectedResults(testInput); + + inputTestValues.push_back(testInput); + exceptedTestResults.push_back(expected); + } + + core::vector cpuTestResults = performCpuTests(inputTestValues); + core::vector gpuTestResults = performGpuTests(inputTestValues); + + bool pass = verifyAllTestResults(cpuTestResults, gpuTestResults, exceptedTestResults); + if (!pass && m_failureManifest) + m_failureManifest->addGroupFailure(m_failurePhase, m_failureId, m_failureName, m_failureLogFile); + + m_logger->log("TESTS DONE.", system::ILogger::ELL_PERFORMANCE); + + m_logFile.close(); + return pass; + } + private: template inline void logFail(const char* msg, Args&&... args) @@ -439,6 +487,11 @@ class ITester uint32_t m_seed; std::ofstream m_logFile; core::unordered_map m_maxErrors; + nbl::examples::testing::FailureManifest* m_failureManifest = nullptr; + std::string m_failurePhase; + std::string m_failureId; + std::string m_failureName; + std::string m_failureLogFile; }; -#endif \ No newline at end of file +#endif