Merge pull request #717 from Devsh-Graphics-Programming/clustered_rendering_cleanup

Clustered rendering cleanup

Author: devshgraphicsprogramming

examples_tests

@@ -320,7 +320,7 @@ class IRenderpass
             }
             return false;
         }
-        
+
         template<bool InputAttachment>
         static inline bool invalidLayout(const IImage::LAYOUT _layout)
         {

include/nbl/asset/IRenderpass.h

@@ -6,7 +6,7 @@
 
 #define NBL_GLSL_DECLARE_LOWER_BOUND_COMP(ARRAY_NAME,TYPE,COMP) NBL_GLSL_CONCATENATE4(uint lower_bound_,ARRAY_NAME,_,NBL_GLSL_LESS)(uint begin, in uint end, in TYPE value);
 
-#define NBL_GLSL_DECLARE_UPPER_BOUND_COMP(ARRAY_NAME,TYPE,COMP) NBL_GLSL_CONCATENATE4(uint lower_bound_,ARRAY_NAME,_,NBL_GLSL_LESS)(uint begin, in uint end, in TYPE value);
+#define NBL_GLSL_DECLARE_UPPER_BOUND_COMP(ARRAY_NAME,TYPE,COMP) NBL_GLSL_CONCATENATE4(uint upper_bound_,ARRAY_NAME,_,NBL_GLSL_LESS)(uint begin, in uint end, in TYPE value);
 
 
 #define NBL_GLSL_DECLARE_LOWER_BOUND(ARRAY_NAME,TYPE) NBL_GLSL_DECLARE_LOWER_BOUND_COMP(ARRAY_NAME,TYPE,NBL_GLSL_LESS) \

include/nbl/builtin/glsl/algorithm.glsl

@@ -23,6 +23,7 @@ class CLumaMeter : public core::TotalInterface
     public:
 		_NBL_STATIC_INLINE_CONSTEXPR uint32_t DEFAULT_BIN_COUNT = 256u;
 		_NBL_STATIC_INLINE_CONSTEXPR uint32_t DEFAULT_BIN_GLOBAL_REPLICATION = 4u;
+		_NBL_STATIC_INLINE_CONSTEXPR uint32_t DEFAULT_DESCRIPTOR_COUNT = 3u;
 
 		enum E_METERING_MODE
 		{
@@ -90,7 +91,18 @@ class CLumaMeter : public core::TotalInterface
 		static core::SRange<const asset::SPushConstantRange> getDefaultPushConstantRanges();
 
 		//
-		static core::SRange<const video::IGPUDescriptorSetLayout::SBinding> getDefaultBindings(video::IVideoDriver* driver);
+		static core::SRange<const video::IGPUDescriptorSetLayout::SBinding> getDefaultBindings(video::ILogicalDevice* device);
+
+		//
+		static inline core::smart_refctd_ptr<video::IGPUPipelineLayout> getDefaultPipelineLayout(video::ILogicalDevice* device)
+		{
+			auto pcRange = getDefaultPushConstantRanges();
+			auto bindings = getDefaultBindings(device);
+			return device->createGPUPipelineLayout(
+				pcRange.begin(), pcRange.end(),
+				device->createGPUDescriptorSetLayout(bindings.begin(), bindings.end()), nullptr, nullptr, nullptr
+			);
+		}
 
 		//
 		static inline size_t getOutputBufferSize(E_METERING_MODE meterMode, uint32_t arrayLayers=1u)
@@ -118,13 +130,67 @@ class CLumaMeter : public core::TotalInterface
 			E_METERING_MODE meterMode, float minLuma=1.f/2048.f, float maxLuma=65536.f
 		);
 
-		// we expect user binds correct pipeline, descriptor sets and pushes the push constants by themselves
-		static inline void dispatchHelper(video::IVideoDriver* driver, const DispatchInfo_t& dispatchInfo, bool issueDefaultBarrier=true)
+		template <E_METERING_MODE mode>
+		static inline void updateDescriptorSet(
+			video::IGPUDescriptorSet* ds,
+			const core::smart_refctd_ptr<video::IGPUBuffer> paramUbo,
+			const asset::SBufferRange<video::IGPUBuffer>& outputSSBORange,
+			const core::smart_refctd_ptr<video::IGPUImageView> inputImageView,
+			video::ILogicalDevice* logicalDevice)
 		{
-			driver->dispatch(dispatchInfo.workGroupCount[0], dispatchInfo.workGroupCount[1], dispatchInfo.workGroupCount[2]);
+			static_assert(mode < EMM_COUNT, "Invalide meter mode!");
+
+			video::IGPUDescriptorSet::SWriteDescriptorSet writes[DEFAULT_DESCRIPTOR_COUNT] = {};
+			video::IGPUDescriptorSet::SDescriptorInfo infos[DEFAULT_DESCRIPTOR_COUNT] = {};
+
+			const asset::E_DESCRIPTOR_TYPE descriptorTypes[DEFAULT_DESCRIPTOR_COUNT] =
+			{
+				asset::EDT_UNIFORM_BUFFER, // luma input params
+				asset::EDT_STORAGE_BUFFER, // luma output buffer
+				asset::EDT_COMBINED_IMAGE_SAMPLER // input image
+			};
+
+			infos[0].desc = paramUbo;
+			infos[0].buffer.offset = 0ull;
+			infos[0].buffer.size = sizeof(Uniforms_t<mode>);
+
+			infos[1].desc = outputSSBORange.buffer;
+			infos[1].buffer.offset = outputSSBORange.offset;
+			infos[1].buffer.size = outputSSBORange.size - infos[1].buffer.offset;
 
-			if (issueDefaultBarrier)
-				defaultBarrier();
+			infos[2].desc = inputImageView;
+			infos[2].image.sampler = nullptr;
+			infos[2].image.imageLayout = asset::EIL_SHADER_READ_ONLY_OPTIMAL;
+
+			for (uint32_t binding = 0u; binding < DEFAULT_DESCRIPTOR_COUNT; ++binding)
+			{
+				writes[binding].dstSet = ds;
+				writes[binding].binding = binding;
+				writes[binding].arrayElement = 0u;
+				writes[binding].count = 1u;
+				writes[binding].descriptorType = descriptorTypes[binding];
+				writes[binding].info = infos + binding;
+			}
+
+			logicalDevice->updateDescriptorSets(DEFAULT_DESCRIPTOR_COUNT, writes, 0u, nullptr);
+		}
+
+		// we expect user binds correct pipeline, descriptor sets and pushes the push constants by themselves
+		static inline void dispatchHelper(
+			video::IGPUCommandBuffer* cmdbuf, 
+			const DispatchInfo_t& dispatchInfo,
+			const asset::E_PIPELINE_STAGE_FLAGS srcStageMask,
+			const uint32_t srcBarrierCount,
+			const video::IGPUCommandBuffer::SBufferMemoryBarrier* srcBarriers,
+			const asset::E_PIPELINE_STAGE_FLAGS dstStageMask,
+			const uint32_t dstBarrierCount,
+			const video::IGPUCommandBuffer::SBufferMemoryBarrier* dstBarriers)
+		{
+			if (srcStageMask != asset::E_PIPELINE_STAGE_FLAGS::EPSF_TOP_OF_PIPE_BIT && srcBarrierCount)
+				cmdbuf->pipelineBarrier(srcStageMask, asset::EPSF_COMPUTE_SHADER_BIT, asset::EDF_NONE, 0u, nullptr, srcBarrierCount, srcBarriers, 0u, nullptr);
+			cmdbuf->dispatch(dispatchInfo.workGroupCount[0], dispatchInfo.workGroupCount[1], dispatchInfo.workGroupCount[2]);
+			if (dstStageMask != asset::E_PIPELINE_STAGE_FLAGS::EPSF_BOTTOM_OF_PIPE_BIT && dstBarrierCount)
+				cmdbuf->pipelineBarrier(asset::EPSF_COMPUTE_SHADER_BIT, dstStageMask, asset::EDF_NONE, 0u, nullptr, dstBarrierCount, dstBarriers, 0u, nullptr);
 		}
 
     private:
@@ -155,8 +221,6 @@ class CLumaMeter : public core::TotalInterface
 			}
 			return retval;
 		}
-
-		static void defaultBarrier();
 };