Merge pull request #119 from Devsh-Graphics-Programming/failed_pt_refactor2

devshgraphicsprogramming · web-flow · commit 80c3f4c64156 · 2021-06-10T23:07:09.000+02:00
Pack UV and Normal into one attribute
diff --git a/examples_tests/22.RaytracedAO/Renderer.cpp b/examples_tests/22.RaytracedAO/Renderer.cpp
@@ -88,9 +88,9 @@ Renderer::Renderer(IVideoDriver* _driver, IAssetManager* _assetManager, scene::I
 	}
 	{
 #ifndef DISABLE_NEE
-		constexpr auto additionalGlobalDescriptorCount = 6u;
+		constexpr auto additionalGlobalDescriptorCount = 5u;
 #else
-		constexpr auto additionalGlobalDescriptorCount = 4u;
+		constexpr auto additionalGlobalDescriptorCount = 3u;
 #endif
 		IGPUDescriptorSetLayout::SBinding bindings[additionalGlobalDescriptorCount];
 		fillIotaDescriptorBindingDeclarations(bindings,ISpecializedShader::ESS_COMPUTE|ISpecializedShader::ESS_VERTEX|ISpecializedShader::ESS_FRAGMENT,additionalGlobalDescriptorCount,asset::EDT_STORAGE_BUFFER);
@@ -209,7 +209,7 @@ Renderer::InitializationData Renderer::initSceneObjects(const SAssetBundle& mesh
 		info.buffer.offset = 0u;
 		info.desc = std::move(buf);
 	};
-	constexpr uint32_t writeBound = 4u;
+	constexpr uint32_t writeBound = 3u;
 	IGPUDescriptorSet::SWriteDescriptorSet writes[writeBound];
 	auto recordSSBOWrite = [](IGPUDescriptorSet::SWriteDescriptorSet& write, IGPUDescriptorSet::SDescriptorInfo* infos, uint32_t binding, uint32_t count=1u) -> void
 	{
@@ -250,15 +250,18 @@ Renderer::InitializationData Renderer::initSceneObjects(const SAssetBundle& mesh
 
 				constexpr uint8_t kIndicesPerTriangle = 3u;
 				constexpr uint16_t minIndicesBatch = minTrisBatch*kIndicesPerTriangle;
-				
+				 
 				CPUMeshPacker::AllocationParams allocParams;
 				allocParams.vertexBuffSupportedByteSize = 1u<<31u;
 				allocParams.vertexBufferMinAllocByteSize = minTrisBatch*minVertexSize;
 				allocParams.indexBuffSupportedCnt = (allocParams.vertexBuffSupportedByteSize/allocParams.vertexBufferMinAllocByteSize)*minIndicesBatch;
 				allocParams.indexBufferMinAllocCnt = minIndicesBatch;
 				allocParams.MDIDataBuffSupportedCnt = allocParams.indexBuffSupportedCnt/minIndicesBatch;
 				allocParams.MDIDataBuffMinAllocCnt = 1u; //so structs from different meshbuffers are adjacent in memory
-    
+
+				constexpr auto combinedNormalUVAttributeIx = 1;
+				constexpr auto newEnabledAttributeMask = (0x1u<<combinedNormalUVAttributeIx)|0b1;
+
 				auto cpump = core::make_smart_refctd_ptr<CCPUMeshPackerV2<>>(allocParams,minTrisBatch,maxTrisBatch);
 				uint32_t mdiBoundMax=0u,batchInstanceBoundTotal=0u;
 				core::vector<CPUMeshPacker::ReservedAllocationMeshBuffers> allocData;
@@ -274,14 +277,49 @@ Renderer::InitializationData Renderer::initSceneObjects(const SAssetBundle& mesh
 						assert(!meshBuffers.empty());
 						const uint32_t instanceCount = (*meshBuffers.begin())->getInstanceCount();
 						for (auto mbIt=meshBuffers.begin(); mbIt!=meshBuffers.end(); mbIt++)
-							assert((*mbIt)->getInstanceCount()==instanceCount);
+						{
+							auto meshBuffer = *mbIt;
+							assert(meshBuffer->getInstanceCount()==instanceCount);
+							// We'll disable certain attributes to ensure we only copy position, normal and uv attribute
+							SVertexInputParams& vertexInput = meshBuffer->getPipeline()->getVertexInputParams();
+							// but we'll pack normals and UVs together to save one SSBO binding (and quantize UVs to half floats)
+							constexpr auto freeBinding = 15u;
+							vertexInput.attributes[combinedNormalUVAttributeIx].binding = freeBinding;
+							vertexInput.attributes[combinedNormalUVAttributeIx].format = EF_R32G32_UINT;
+							vertexInput.attributes[combinedNormalUVAttributeIx].relativeOffset = 0u;
+							vertexInput.enabledBindingFlags |= 0x1u<<freeBinding;
+							vertexInput.bindings[freeBinding].inputRate = EVIR_PER_VERTEX;
+							vertexInput.bindings[freeBinding].stride = 0u;
+							const auto approxVxCount = IMeshManipulator::upperBoundVertexID(meshBuffer)+meshBuffer->getBaseVertex();
+							struct CombinedNormalUV
+							{
+								uint32_t nml;
+								uint16_t u,v;
+							};
+							auto newBuff = core::make_smart_refctd_ptr<ICPUBuffer>(sizeof(CombinedNormalUV)*approxVxCount);
+							auto* dst = reinterpret_cast<CombinedNormalUV*>(newBuff->getPointer())+meshBuffer->getBaseVertex();
+							meshBuffer->setVertexBufferBinding({0u,newBuff},freeBinding);
+							// copy and pack data
+							const auto normalAttr = meshBuffer->getNormalAttributeIx();
+							vertexInput.attributes[normalAttr].format = EF_R32_UINT;
+							for (auto i=0u; i<approxVxCount; i++)
+							{
+								meshBuffer->getAttribute(&dst[i].nml,normalAttr,i);
+								core::vectorSIMDf uv;
+								meshBuffer->getAttribute(uv,2u,i);
+								dst[i].u = core::Float16Compressor::compress(uv.x);
+								dst[i].v = core::Float16Compressor::compress(uv.y);
+							}
+						}
 
 						const uint32_t mdiBound = cpump->calcMDIStructMaxCount(meshBuffers.begin(),meshBuffers.end());
 						mdiBoundMax = core::max(mdiBound,mdiBoundMax);
 						batchInstanceBoundTotal += mdiBound*instanceCount;
 
 						meshBuffersToProcess.insert(meshBuffersToProcess.end(),meshBuffers.begin(),meshBuffers.end());
 					}
+					for (auto meshBuffer : meshBuffersToProcess)
+						const_cast<ICPUMeshBuffer*>(meshBuffer)->getPipeline()->getVertexInputParams().enabledAttribFlags = newEnabledAttributeMask;
 					allocData.resize(meshBuffersToProcess.size());
 
 					cpump->alloc(allocData.data(),meshBuffersToProcess.begin(),meshBuffersToProcess.end());
@@ -356,22 +394,20 @@ Renderer::InitializationData Renderer::initSceneObjects(const SAssetBundle& mesh
 									indexCount/kIndicesPerTriangle
 								);
 
-								const auto normalAttrID = mb->getNormalAttributeIx();
 								const auto thisShapeInstancesBeginIx = rrInstances.size();
 								const auto& batchAABB = mb->getBoundingBox();// TODO: replace with batch AABB
 								for (auto auxIt=instanceAuxData.begin(); auxIt!=instanceAuxData.end(); auxIt++)
 								{
-									constexpr auto UVAttributeIx = 2;
 									const auto batchInstanceGUID = cullData.size();
 
 									const auto instanceID = std::distance(instanceAuxData.begin(),auxIt);
 									*newInstanceData = mbInstanceData[instanceID];
 									assert(instanceData.begin()[instanceID].worldTform==newInstanceData->tform);
-									newInstanceData->padding0 = reinterpret_cast<const uint32_t&>(cdotIt->attribInfo[posAttrID]);
-									newInstanceData->padding1 = reinterpret_cast<const uint32_t&>(cdotIt->attribInfo[normalAttrID]);
+									newInstanceData->padding0 = firstIndex;
+									newInstanceData->padding1 = reinterpret_cast<const uint32_t&>(cdotIt->attribInfo[posAttrID]);
 									newInstanceData->determinantSignBit = core::bitfieldInsert(
 										newInstanceData->determinantSignBit,
-										reinterpret_cast<const uint32_t&>(cdotIt->attribInfo[UVAttributeIx]),
+										reinterpret_cast<const uint32_t&>(cdotIt->attribInfo[combinedNormalUVAttributeIx]),
 										0u,31u
 									);
 									if (frontFaceIsCCW) // compensate for Nabla's default camera being left handed
@@ -446,7 +482,7 @@ Renderer::InitializationData Renderer::initSceneObjects(const SAssetBundle& mesh
 							recordInfoBuffer(infos[i],core::smart_refctd_ptr(dataStore.vertexBuffer));
 							recordSSBOWrite(writes[i],infos+i,i);
 						}
-						recordInfoBuffer(infos[2],core::smart_refctd_ptr(m_indexBuffer));
+						recordInfoBuffer(infos[1],core::smart_refctd_ptr(m_indexBuffer));
 
 						setDstSetOnAllWrites(m_additionalGlobalDS.get());
 						m_driver->updateDescriptorSets(writeBound,writes,0u,nullptr);
@@ -866,7 +902,7 @@ void Renderer::init(const SAssetBundle& meshes,	core::smart_refctd_ptr<ICPUBuffe
 				createFilledBufferAndSetUpInfoFromVector(infos+0,initData.lightCDF);
 				createFilledBufferAndSetUpInfoFromVector(infos+1,initData.lights);
 
-				setDstSetAndDescTypesOnWrites(m_additionalGlobalDS.get(),writes,infos,{EDT_STORAGE_BUFFER,EDT_STORAGE_BUFFER},4u);
+				setDstSetAndDescTypesOnWrites(m_additionalGlobalDS.get(),writes,infos,{EDT_STORAGE_BUFFER,EDT_STORAGE_BUFFER},3u);
 			}
 			m_driver->updateDescriptorSets(descriptorUpdateCounts[0],writes,0u,nullptr);
 #endif
diff --git a/examples_tests/22.RaytracedAO/main.cpp b/examples_tests/22.RaytracedAO/main.cpp
@@ -234,7 +234,7 @@ int main()
 
 		if (generateNewSamples)
 		{
-			/** TODO: move into the renderer and redo the sampling
+			/** TODO: move into the renderer and redo the sampling (compress into R21G21B21_UINT)
 			Locality Level 0: the 3 dimensions consumed for a BxDF or NEE sample
 			Locality Level 1: the k = 3 (1 + NEE) samples which will be consumed in the same invocation
 			Locality Level 2-COMP: the N = k dispatchSPP Resolution samples consumed by a raygen dispatch (another TODO: would be order CS and everything in a morton curve)
diff --git a/examples_tests/22.RaytracedAO/raytraceCommon.glsl b/examples_tests/22.RaytracedAO/raytraceCommon.glsl
@@ -11,11 +11,11 @@ layout(push_constant, row_major) uniform PushConstants
 
 #ifndef DISABLE_NEE
 // lights
-layout(set = 1, binding = 4, std430) restrict readonly buffer CumulativeLightPDF
+layout(set = 1, binding = 3, std430) restrict readonly buffer CumulativeLightPDF
 {
 	uint lightCDF[];
 };
-layout(set = 1, binding = 5, std430, row_major) restrict readonly buffer Lights
+layout(set = 1, binding = 4, std430, row_major) restrict readonly buffer Lights
 {
 	SLight light[];
 };
@@ -262,7 +262,7 @@ for (uint i=1u; i!=vertex_depth; i++)
 	atomicMax(traceIndirect[vertex_depth_mod_2_inv].params.num_groups_x,(baseOutputID+raysToAllocate-1u)/WORKGROUP_SIZE+1u);
 
 	// TODO: improve ray offset (maybe using smooth normal wouldn't be a sin)
-	vec3 geomNormal = cross(dPdBary[0], dPdBary[1]);
+	vec3 geomNormal = cross(dPdBary[0],dPdBary[1]);
 	const vec3 absGeomNormal = abs(geomNormal);
 	geomNormal /= max(max(absGeomNormal.x,absGeomNormal.y),max(absGeomNormal.z,0.001f))*96.f;
 	uint offset = 0u;
diff --git a/examples_tests/22.RaytracedAO/virtualGeometry.glsl b/examples_tests/22.RaytracedAO/virtualGeometry.glsl
@@ -5,15 +5,13 @@
 
 #define _NBL_VG_USE_SSBO
 #define _NBL_VG_SSBO_DESCRIPTOR_SET 1
-#define _NBL_VG_USE_SSBO_UINT
-#define _NBL_VG_SSBO_UINT_BINDING 0
-#define _NBL_VG_USE_SSBO_UVEC3
-#define _NBL_VG_SSBO_UVEC3_BINDING 1
-#define _NBL_VG_USE_SSBO_INDEX
-#define _NBL_VG_SSBO_INDEX_BINDING 2
-// TODO: remove after all quantization optimizations in CSerializedLoader and the like
 #define _NBL_VG_USE_SSBO_UVEC2
-#define _NBL_VG_SSBO_UVEC2_BINDING 3
+#define _NBL_VG_SSBO_UVEC2_BINDING 0
+#define _NBL_VG_USE_SSBO_INDEX
+#define _NBL_VG_SSBO_INDEX_BINDING 1
+// TODO: remove after Doom Eternal position quantization trick
+#define _NBL_VG_USE_SSBO_UVEC3
+#define _NBL_VG_SSBO_UVEC3_BINDING 2
 #include <nbl/builtin/glsl/virtual_geometry/virtual_attribute_fetch.glsl>
 
 
@@ -22,20 +20,22 @@
 
 vec3 nbl_glsl_fetchVtxPos(in uint vtxID, in nbl_glsl_ext_Mitsuba_Loader_instance_data_t batchInstanceData)
 {
-    nbl_glsl_VG_VirtualAttributePacked_t va = batchInstanceData.padding0;
+    nbl_glsl_VG_VirtualAttributePacked_t va = batchInstanceData.padding1;
     return nbl_glsl_VG_attribFetch_RGB32_SFLOAT(va,vtxID);
 }
 
 vec3 nbl_glsl_fetchVtxNormal(in uint vtxID, in nbl_glsl_ext_Mitsuba_Loader_instance_data_t batchInstanceData)
 {
-    nbl_glsl_VG_VirtualAttributePacked_t va = batchInstanceData.padding1;
-    return normalize(nbl_glsl_VG_attribFetch_RGB10A2_SNORM(va,vtxID).xyz);
+    nbl_glsl_VG_VirtualAttributePacked_t va = batchInstanceData.determinantSignBit;
+    const uint codedNormal = nbl_glsl_VG_attribFetch2u(va,vtxID)[0];
+    return normalize(nbl_glsl_decodeRGB10A2_SNORM(codedNormal).xyz);
 }
 
 vec2 nbl_glsl_fetchVtxUV(in uint vtxID, in nbl_glsl_ext_Mitsuba_Loader_instance_data_t batchInstanceData)
 {
     nbl_glsl_VG_VirtualAttributePacked_t va = batchInstanceData.determinantSignBit;
-    return nbl_glsl_VG_attribFetch_RG32_SFLOAT(va,vtxID);
+    const uint codedUV = nbl_glsl_VG_attribFetch2u(va,vtxID)[1];
+    return unpackHalf2x16(codedUV).xy;
 }
 
 

Original file line number	Diff line number	Diff line change
`@@ -234,7 +234,7 @@ int main()`
`234`	`234`
`235`	`235`	`if (generateNewSamples)`
`236`	`236`	`{`
`237`		`- /** TODO: move into the renderer and redo the sampling`
	`237`	`+ /** TODO: move into the renderer and redo the sampling (compress into R21G21B21_UINT)`
`238`	`238`	`Locality Level 0: the 3 dimensions consumed for a BxDF or NEE sample`
`239`	`239`	`Locality Level 1: the k = 3 (1 + NEE) samples which will be consumed in the same invocation`
`240`	`240`	`Locality Level 2-COMP: the N = k dispatchSPP Resolution samples consumed by a raygen dispatch (another TODO: would be order CS and everything in a morton curve)`