merged rwmc, fix conflicts

Author: keptsecret

31_HLSLPathTracer/app_resources/hlsl/pathtracer.hlsl

@@ -5,6 +5,8 @@
 #include <nbl/builtin/hlsl/colorspace/encodeCIEXYZ.hlsl>
 #include <nbl/builtin/hlsl/math/functions.hlsl>
 #include <nbl/builtin/hlsl/bxdf/bxdf_traits.hlsl>
+#include <nbl/builtin/hlsl/vector_utils/vector_traits.hlsl>
+#include <nbl/builtin/hlsl/concepts.hlsl>
 
 #include "rand_gen.hlsl"
 #include "ray_gen.hlsl"
@@ -35,10 +37,33 @@ struct PathTracerCreationParams
     matrix<Scalar, 4, 4> invMVP;
 };
 
-template<class RandGen, class RayGen, class Intersector, class MaterialSystem, /* class PathGuider, */ class NextEventEstimator>
+template<typename OutputTypeVec NBL_PRIMARY_REQUIRES(concepts::FloatingPointVector<OutputTypeVec>)
+struct DefaultAccumulator
+{
+    using output_storage_type = OutputTypeVec;
+    using this_t = DefaultAccumulator<OutputTypeVec>;
+    output_storage_type accumulation;
+
+    static this_t create()
+    {
+        this_t retval;
+        retval.accumulation = promote<OutputTypeVec, float32_t>(0.0f);
+
+        return retval;
+    }
+
+    void addSample(uint32_t sampleCount, float32_t3 sample)
+    {
+        using ScalarType = typename vector_traits<OutputTypeVec>::scalar_type;
+        ScalarType rcpSampleSize = 1.0 / (sampleCount);
+        accumulation += (sample - accumulation) * rcpSampleSize;
+    }
+};
+
+template<class RandGen, class RayGen, class Intersector, class MaterialSystem, /* class PathGuider, */ class NextEventEstimator, class Accumulator>
 struct Unidirectional
 {
-    using this_t = Unidirectional<RandGen, RayGen, Intersector, MaterialSystem, NextEventEstimator>;
+    using this_t = Unidirectional<RandGen, RayGen, Intersector, MaterialSystem, NextEventEstimator, Accumulator>;
     using randgen_type = RandGen;
     using raygen_type = RayGen;
     using intersector_type = Intersector;
@@ -49,6 +74,7 @@ struct Unidirectional
     using vector3_type = vector<scalar_type, 3>;
     using monochrome_type = vector<scalar_type, 1>;
     using measure_type = typename MaterialSystem::measure_type;
+    using output_storage_type = typename Accumulator::output_storage_type; // ?
     using sample_type = typename NextEventEstimator::sample_type;
     using ray_dir_info_type = typename sample_type::ray_dir_info_type;
     using ray_type = typename RayGen::ray_type;
@@ -255,40 +281,33 @@ struct Unidirectional
     }
 
     // Li
-    measure_type getMeasure(uint32_t numSamples, uint32_t depth, NBL_CONST_REF_ARG(scene_type) scene)
+    void sampleMeasure(uint32_t sampleIndex, uint32_t maxDepth, NBL_CONST_REF_ARG(scene_type) scene, NBL_REF_ARG(Accumulator) accumulator)
     {
-        measure_type Li = (measure_type)0.0;
-        scalar_type meanLumaSq = 0.0;
-        for (uint32_t i = 0; i < numSamples; i++)
+        //scalar_type meanLumaSq = 0.0;
+        vector3_type uvw = rand3d(0u, sampleIndex, randGen.rng());    // TODO: take from scramblebuf?
+        ray_type ray = rayGen.generate(uvw);
+
+        // bounces
+        bool hit = true;
+        bool rayAlive = true;
+        for (int d = 1; (d <= maxDepth) && hit && rayAlive; d += 2)
         {
-            vector3_type uvw = rand3d(0u, i, randGen.rng());    // TODO: take from scramblebuf?
-            ray_type ray = rayGen.generate(uvw);
-
-            // bounces
-            bool hit = true;
-            bool rayAlive = true;
-            for (int d = 1; (d <= depth) && hit && rayAlive; d += 2)
-            {
-                ray.intersectionT = numeric_limits<scalar_type>::max;
-                ray.objectID = intersector_type::traceRay(ray, scene);
+            ray.intersectionT = numeric_limits<scalar_type>::max;
+            ray.objectID = intersector_type::traceRay(ray, scene);
 
-                hit = ray.objectID.id != -1;
-                if (hit)
-                    rayAlive = closestHitProgram(1, i, ray, scene);
-            }
-            if (!hit)
-                missProgram(ray);
-
-            measure_type accumulation = ray.payload.accumulation;
-            scalar_type rcpSampleSize = 1.0 / (i + 1);
-            Li += (accumulation - Li) * rcpSampleSize;
+            hit = ray.objectID.id != -1;
+            if (hit)
+                rayAlive = closestHitProgram(1, sampleIndex, ray, scene);
+        }
+        if (!hit)
+            missProgram(ray);
 
-            // TODO: visualize high variance
+        const uint32_t sampleCount = sampleIndex + 1;
+        accumulator.addSample(sampleCount, ray.payload.accumulation);
 
-            // TODO: russian roulette early exit?
-        }
+        // TODO: visualize high variance
 
-        return Li;
+        // TODO: russian roulette early exit?
     }
 
     NBL_CONSTEXPR_STATIC_INLINE uint32_t MAX_DEPTH_LOG2 = 4u;

31_HLSLPathTracer/app_resources/hlsl/present.frag.hlsl

@@ -10,10 +10,10 @@ using namespace nbl::hlsl;
 using namespace ext::FullScreenTriangle;
 
 // binding 0 set 0
-[[vk::combinedImageSampler]] [[vk::binding(0, 0)]] Texture2D texture;
+[[vk::combinedImageSampler]] [[vk::binding(0, 0)]] Texture2DArray texture;
 [[vk::combinedImageSampler]] [[vk::binding(0, 0)]] SamplerState samplerState;
 
 [[vk::location(0)]] float32_t4 main(SVertexAttributes vxAttr) : SV_Target0
 {
-    return float32_t4(texture.Sample(samplerState, vxAttr.uv).rgb, 1.0f);
+    return float32_t4(texture.Sample(samplerState, float3(vxAttr.uv, 0)).rgb, 1.0f);
 }

31_HLSLPathTracer/app_resources/hlsl/render.comp.hlsl

@@ -35,7 +35,31 @@
 #define LIGHT_COUNT 1
 #define BXDF_COUNT 7
 
-#include "render_common.hlsl"
+#include <render_common.hlsl>
+#include <rwmc_global_settings_common.hlsl>
+
+#ifdef RWMC_ENABLED
+#include <nbl/builtin/hlsl/rwmc/CascadeAccumulator.hlsl>
+#include <render_rwmc_common.hlsl>
+#endif
+
+#ifdef RWMC_ENABLED
+[[vk::push_constant]] RenderRWMCPushConstants pc;
+#else
+[[vk::push_constant]] RenderPushConstants pc;
+#endif
+
+[[vk::combinedImageSampler]] [[vk::binding(0, 2)]] Texture2D<float3> envMap;      // unused
+[[vk::combinedImageSampler]] [[vk::binding(0, 2)]] SamplerState envSampler;
+
+[[vk::binding(1, 2)]] Buffer<uint3> sampleSequence;
+
+[[vk::combinedImageSampler]] [[vk::binding(2, 2)]] Texture2D<uint2> scramblebuf; // unused
+[[vk::combinedImageSampler]] [[vk::binding(2, 2)]] SamplerState scrambleSampler;
+
+[[vk::image_format("rgba16f")]] [[vk::binding(0)]] RWTexture2DArray<float32_t4> outImage;
+[[vk::image_format("rgba16f")]] [[vk::binding(1)]] RWTexture2DArray<float32_t4> cascade;
+
 #include "pathtracer.hlsl"
 
 using namespace nbl;
@@ -59,15 +83,15 @@ NBL_CONSTEXPR ext::PTPolygonMethod POLYGON_METHOD = ext::PPM_SOLID_ANGLE;
 
 int32_t2 getCoordinates()
 {
-    uint32_t width, height;
-    outImage.GetDimensions(width, height);
+    uint32_t width, height, imageArraySize;
+    outImage.GetDimensions(width, height, imageArraySize);
     return int32_t2(glsl::gl_GlobalInvocationID().x % width, glsl::gl_GlobalInvocationID().x / width);
 }
 
 float32_t2 getTexCoords()
 {
-    uint32_t width, height;
-    outImage.GetDimensions(width, height);
+    uint32_t width, height, imageArraySize;
+    outImage.GetDimensions(width, height, imageArraySize);
     int32_t2 iCoords = getCoordinates();
     return float32_t2(float(iCoords.x) / width, 1.0 - float(iCoords.y) / height);
 }
@@ -97,7 +121,14 @@ using raygen_type = ext::RayGen::Basic<ray_type>;
 using intersector_type = ext::Intersector::Comprehensive<ray_type, light_type, bxdfnode_type>;
 using material_system_type = ext::MaterialSystem::System<diffuse_bxdf_type, conductor_bxdf_type, dielectric_bxdf_type>;
 using nee_type = ext::NextEventEstimator::Estimator<scene_type, ray_type, sample_t, aniso_interaction, ext::IntersectMode::IM_PROCEDURAL, LIGHT_TYPE, POLYGON_METHOD>;
-using pathtracer_type = ext::PathTracer::Unidirectional<randgen_type, raygen_type, intersector_type, material_system_type, nee_type>;
+
+#ifdef RWMC_ENABLED
+using accumulator_type = rwmc::CascadeAccumulator<float32_t3, CascadeCount>;
+#else
+using accumulator_type = ext::PathTracer::DefaultAccumulator<float32_t3>;
+#endif
+
+using pathtracer_type = ext::PathTracer::Unidirectional<randgen_type, raygen_type, intersector_type, material_system_type, nee_type, accumulator_type>;
 
 static const ext::Shape<ext::PST_SPHERE> spheres[SPHERE_COUNT] = {
     ext::Shape<ext::PST_SPHERE>::create(float3(0.0, -100.5, -1.0), 100.0, 0u, light_type::INVALID_ID),
@@ -130,7 +161,7 @@ static const ext::Shape<ext::PST_RECTANGLE> rectangles[1];
 #endif
 
 static const light_type lights[LIGHT_COUNT] = {
-    light_type::create(spectral_t(30.0,25.0,15.0),
+    light_type::create(LightEminence,
 #ifdef SPHERE_LIGHT
         8u,
 #else
@@ -155,11 +186,22 @@ static const ext::Scene<light_type, bxdfnode_type> scene = ext::Scene<light_type
     lights, LIGHT_COUNT, bxdfs, BXDF_COUNT
 );
 
-[numthreads(WorkgroupSize, 1, 1)]
+RenderPushConstants retireveRenderPushConstants()
+{
+#ifdef RWMC_ENABLED
+    return pc.renderPushConstants;
+#else
+    return pc;
+#endif
+}
+
+[numthreads(RenderWorkgroupSize, 1, 1)]
 void main(uint32_t3 threadID : SV_DispatchThreadID)
 {
-    uint32_t width, height;
-    outImage.GetDimensions(width, height);
+    const RenderPushConstants renderPushConstants = retireveRenderPushConstants();
+
+    uint32_t width, height, imageArraySize;
+    outImage.GetDimensions(width, height, imageArraySize);
 #ifdef PERSISTENT_WORKGROUPS
     uint32_t virtualThreadIndex;
     [loop]
@@ -181,10 +223,10 @@ void main(uint32_t3 threadID : SV_DispatchThreadID)
 #endif
     }
 
-    if (((pc.depth - 1) >> MAX_DEPTH_LOG2) > 0 || ((pc.sampleCount - 1) >> MAX_SAMPLES_LOG2) > 0)
+    if (((renderPushConstants.depth - 1) >> MAX_DEPTH_LOG2) > 0 || ((renderPushConstants.sampleCount - 1) >> MAX_SAMPLES_LOG2) > 0)
     {
         float32_t4 pixelCol = float32_t4(1.0,0.0,0.0,1.0);
-        outImage[coords] = pixelCol;
+        outImage[uint3(coords.x, coords.y, 0)] = pixelCol;
 #ifdef PERSISTENT_WORKGROUPS
         continue;
 #else
@@ -204,19 +246,33 @@ void main(uint32_t3 threadID : SV_DispatchThreadID)
 
     float4 NDC = float4(texCoord * float2(2.0, -2.0) + float2(-1.0, 1.0), 0.0, 1.0);
     {
-        float4 tmp = mul(pc.invMVP, NDC);
+        float4 tmp = mul(renderPushConstants.invMVP, NDC);
         ptCreateParams.camPos = tmp.xyz / tmp.w;
         NDC.z = 1.0;
     }
 
     ptCreateParams.NDC = NDC;
-    ptCreateParams.invMVP = pc.invMVP;
+    ptCreateParams.invMVP = renderPushConstants.invMVP;
 
     pathtracer_type pathtracer = pathtracer_type::create(ptCreateParams);
 
-    float32_t3 color = pathtracer.getMeasure(pc.sampleCount, pc.depth, scene);
-    float32_t4 pixCol = float32_t4(color, 1.0);
-    outImage[coords] = pixCol;
+#ifdef RWMC_ENABLED
+    accumulator_type accumulator = accumulator_type::create(pc.splattingParameters);
+#else
+    accumulator_type accumulator = accumulator_type::create();
+#endif
+    // path tracing loop
+    for(int i = 0; i < renderPushConstants.sampleCount; ++i)
+        pathtracer.sampleMeasure(i, renderPushConstants.depth, scene, accumulator);
+
+#ifdef RWMC_ENABLED
+    for (uint32_t i = 0; i < CascadeCount; ++i)
+        cascade[uint3(coords.x, coords.y, i)] = float32_t4(accumulator.accumulation.data[i], 1.0f);
+#else
+    outImage[uint3(coords.x, coords.y, 0)] = float32_t4(accumulator.accumulation, 1.0);
+#endif
+
+    
 
 #ifdef PERSISTENT_WORKGROUPS
     }

31_HLSLPathTracer/app_resources/hlsl/render_common.hlsl

@@ -1,23 +1,23 @@
 #ifndef _NBL_HLSL_PATHTRACER_RENDER_COMMON_INCLUDED_
 #define _NBL_HLSL_PATHTRACER_RENDER_COMMON_INCLUDED_
+#include "nbl/builtin/hlsl/cpp_compat.hlsl"
 
-struct SPushConstants
+#ifndef __HLSL_VERSION
+#include "matrix4SIMD.h"
+#endif
+
+struct RenderPushConstants
 {
+#ifdef __HLSL_VERSION
     float32_t4x4 invMVP;
+#else
+    nbl::hlsl::float32_t4x4 invMVP;
+#endif
     int sampleCount;
     int depth;
 };
 
-[[vk::push_constant]] SPushConstants pc;
-
-[[vk::combinedImageSampler]][[vk::binding(0, 2)]] Texture2D<float3> envMap;      // unused
-[[vk::combinedImageSampler]][[vk::binding(0, 2)]] SamplerState envSampler;
-
-[[vk::binding(1, 2)]] Buffer<uint3> sampleSequence;
-
-[[vk::combinedImageSampler]][[vk::binding(2, 2)]] Texture2D<uint2> scramblebuf; // unused
-[[vk::combinedImageSampler]][[vk::binding(2, 2)]] SamplerState scrambleSampler;
-
-[[vk::image_format("rgba16f")]][[vk::binding(0, 0)]] RWTexture2D<float32_t4> outImage;
+NBL_CONSTEXPR nbl::hlsl::float32_t3 LightEminence = nbl::hlsl::float32_t3(30.0f, 25.0f, 15.0f);
+NBL_CONSTEXPR uint32_t RenderWorkgroupSize = 512u;
 
 #endif

31_HLSLPathTracer/app_resources/hlsl/render_rwmc_common.hlsl

@@ -0,0 +1,17 @@
+#ifndef _NBL_HLSL_PATHTRACER_RENDER_RWMC_COMMON_INCLUDED_
+#define _NBL_HLSL_PATHTRACER_RENDER_RWMC_COMMON_INCLUDED_
+#include "nbl/builtin/hlsl/cpp_compat.hlsl"
+#include "nbl/builtin/hlsl/rwmc/SplattingParameters.hlsl"
+#include "render_common.hlsl"
+
+#ifndef __HLSL_VERSION
+#include "matrix4SIMD.h"
+#endif
+
+struct RenderRWMCPushConstants
+{
+	RenderPushConstants renderPushConstants;
+	nbl::hlsl::rwmc::SplattingParameters splattingParameters;
+};
+
+#endif

31_HLSLPathTracer/app_resources/hlsl/resolve.comp.hlsl

@@ -0,0 +1,41 @@
+#include <nbl/builtin/hlsl/rwmc/resolve.hlsl>
+#include "resolve_common.hlsl"
+#include "rwmc_global_settings_common.hlsl"
+#ifdef PERSISTENT_WORKGROUPS
+#include "nbl/builtin/hlsl/math/morton.hlsl"
+#endif
+
+[[vk::push_constant]] ResolvePushConstants pc;
+[[vk::image_format("rgba16f")]] [[vk::binding(0)]] RWTexture2DArray<float32_t4> outImage;
+[[vk::image_format("rgba16f")]] [[vk::binding(1)]] RWTexture2DArray<float32_t4> cascade;
+
+using namespace nbl;
+using namespace hlsl;
+
+int32_t2 getImageExtents()
+{
+    uint32_t width, height, imageArraySize;
+    outImage.GetDimensions(width, height, imageArraySize);
+    return int32_t2(width, height);
+}
+
+[numthreads(ResolveWorkgroupSizeX, ResolveWorkgroupSizeY, 1)]
+void main(uint32_t3 threadID : SV_DispatchThreadID)
+{
+    const int32_t2 coords = int32_t2(threadID.x, threadID.y);
+    const int32_t2 imageExtents = getImageExtents();
+    if (coords.x >= imageExtents.x || coords.y >= imageExtents.y)
+        return;
+
+    using ResolveAccessorAdaptorType = rwmc::ResolveAccessorAdaptor<float>;
+    using ResolverType = rwmc::Resolver<ResolveAccessorAdaptorType, float32_t3>;
+    ResolveAccessorAdaptorType accessor;
+    accessor.cascade = cascade;
+    ResolverType resolve = ResolverType::create(pc.resolveParameters);
+
+    float32_t3 color = resolve(accessor, int16_t2(coords.x, coords.y));
+
+    //float32_t3 color = rwmc::reweight<rwmc::ResolveAccessorAdaptor<float> >(pc.resolveParameters, cascade, coords);
+
+    outImage[uint3(coords.x, coords.y, 0)] = float32_t4(color, 1.0f);
+}

31_HLSLPathTracer/app_resources/hlsl/resolve_common.hlsl

@@ -0,0 +1,15 @@
+#ifndef _NBL_HLSL_PATHTRACER_RESOLVE_COMMON_INCLUDED_
+#define _NBL_HLSL_PATHTRACER_RESOLVE_COMMON_INCLUDED_
+#include "nbl/builtin/hlsl/cpp_compat.hlsl"
+#include "nbl/builtin/hlsl/rwmc/ResolveParameters.hlsl"
+
+struct ResolvePushConstants
+{
+	uint32_t sampleCount;
+	nbl::hlsl::rwmc::ResolveParameters resolveParameters;
+};
+
+NBL_CONSTEXPR uint32_t ResolveWorkgroupSizeX = 32u;
+NBL_CONSTEXPR uint32_t ResolveWorkgroupSizeY = 16u;
+
+#endif