Merge pull request #819 from Devsh-Graphics-Programming/ali_blur3

Prefix sum box blur

Author: devshgraphicsprogramming

include/nbl/asset/ICPUSampler.h

@@ -37,16 +37,16 @@ class ICPUSampler : public ISampler, public IAsset
 				};
 				switch (wrapModes[i])
 				{
-					case ISampler::ETC_REPEAT:
+					case ISampler::E_TEXTURE_CLAMP::ETC_REPEAT:
 						repeat();
 						break;
-					case ISampler::ETC_CLAMP_TO_EDGE:
-						texelCoord[i] = core::clamp<int32_t,int32_t>(texelCoord[i],0,mipLastCoord[i]);
+					case ISampler::E_TEXTURE_CLAMP::ETC_CLAMP_TO_EDGE:
+						texelCoord[i] = core::clamp<int32_t, int32_t>(texelCoord[i], 0, mipLastCoord[i]);
 						break;
-					case ISampler::ETC_MIRROR_CLAMP_TO_EDGE:
-						texelCoord[i] = core::clamp<int32_t,int32_t>(texelCoord[i],-int32_t(mipExtent[i]),mipExtent[i]+mipLastCoord[i]);
+					case ISampler::E_TEXTURE_CLAMP::ETC_MIRROR_CLAMP_TO_EDGE:
+						texelCoord[i] = core::clamp<int32_t, int32_t>(texelCoord[i], -int32_t(mipExtent[i]), mipExtent[i] + mipLastCoord[i]);
 						[[fallthrough]];
-					case ISampler::ETC_MIRROR:
+					case ISampler::E_TEXTURE_CLAMP::ETC_MIRROR:
 						{
 							int32_t repeatID = (originalWasNegative+texelCoord[i])/int32_t(mipExtent[i]);
 							repeat();

include/nbl/asset/ISampler.h

@@ -6,6 +6,7 @@
 #define __NBL_ASSET_I_SAMPLER_H_INCLUDED__
 
 #include "nbl/asset/IDescriptor.h"
+#include "nbl/builtin/hlsl/enums.hlsl"
 
 namespace nbl
 {
@@ -16,23 +17,7 @@ class ISampler : public IDescriptor
 {
 	public:
 		//! Texture coord clamp mode outside [0.0, 1.0]
-		enum E_TEXTURE_CLAMP
-		{
-			//! Texture repeats
-			ETC_REPEAT = 0,
-			//! Texture is clamped to the edge pixel
-			ETC_CLAMP_TO_EDGE,
-			//! Texture is clamped to the border pixel (if exists)
-			ETC_CLAMP_TO_BORDER,
-			//! Texture is alternatingly mirrored (0..1..0..1..0..)
-			ETC_MIRROR,
-			//! Texture is mirrored once and then clamped to edge
-			ETC_MIRROR_CLAMP_TO_EDGE,
-			//! Texture is mirrored once and then clamped to border
-			ETC_MIRROR_CLAMP_TO_BORDER,
-
-			ETC_COUNT
-		};
+		using E_TEXTURE_CLAMP = hlsl::TextureClamp;
 
 		enum E_TEXTURE_BORDER_COLOR
 		{
@@ -75,11 +60,11 @@ class ISampler : public IDescriptor
 		{
 			struct {
 				//! Valeus taken from E_TEXTURE_CLAMP
-				uint32_t TextureWrapU : 3 = ETC_REPEAT;
+				uint32_t TextureWrapU : 3 = E_TEXTURE_CLAMP::ETC_REPEAT;
 				//! Valeus taken from E_TEXTURE_CLAMP
-				uint32_t TextureWrapV : 3 = ETC_REPEAT;
+				uint32_t TextureWrapV : 3 = E_TEXTURE_CLAMP::ETC_REPEAT;
 				//! Valeus taken from E_TEXTURE_CLAMP
-				uint32_t TextureWrapW : 3 = ETC_REPEAT;
+				uint32_t TextureWrapW : 3 = E_TEXTURE_CLAMP::ETC_REPEAT;
 				//! Values taken from E_TEXTURE_BORDER_COLOR
 				uint32_t BorderColor : 3 = ETBC_FLOAT_OPAQUE_BLACK;
 				//! Values taken from E_TEXTURE_FILTER

include/nbl/asset/filters/CBlitImageFilter.h

@@ -34,7 +34,7 @@ class CBlitImageFilterBase : public impl::CSwizzleableAndDitherableFilterBase<Sw
 				uint8_t*							scratchMemory = nullptr;
 				uint32_t							scratchMemoryByteSize = 0u;
 				_NBL_STATIC_INLINE_CONSTEXPR auto	NumWrapAxes = 3;
-				ISampler::E_TEXTURE_CLAMP			axisWraps[NumWrapAxes] = { ISampler::ETC_REPEAT,ISampler::ETC_REPEAT,ISampler::ETC_REPEAT };
+				ISampler::E_TEXTURE_CLAMP			axisWraps[NumWrapAxes] = { ISampler::E_TEXTURE_CLAMP::ETC_REPEAT,ISampler::E_TEXTURE_CLAMP::ETC_REPEAT,ISampler::E_TEXTURE_CLAMP::ETC_REPEAT };
 				ISampler::E_TEXTURE_BORDER_COLOR	borderColor = ISampler::ETBC_FLOAT_TRANSPARENT_BLACK;
 				IBlitUtilities::E_ALPHA_SEMANTIC	alphaSemantic = IBlitUtilities::EAS_NONE_OR_PREMULTIPLIED;
 				double								alphaRefValue = 0.5; // only required to make sense if `alphaSemantic==EAS_REFERENCE_OR_COVERAGE`
@@ -56,7 +56,7 @@ class CBlitImageFilterBase : public impl::CSwizzleableAndDitherableFilterBase<Sw
 				return false;
 
 			for (auto i=0; i<CStateBase::NumWrapAxes; i++)
-			if (state->axisWraps[i]>=ISampler::ETC_COUNT)
+			if (state->axisWraps[i]>=ISampler::E_TEXTURE_CLAMP::ETC_COUNT)
 				return false;
 
 			if (state->borderColor>=ISampler::ETBC_COUNT)

include/nbl/asset/filters/CPaddedCopyImageFilter.h

@@ -30,7 +30,7 @@ class CPaddedCopyImageFilter : public CImageFilter<CPaddedCopyImageFilter>, publ
 				virtual ~CState() {}
 
 				_NBL_STATIC_INLINE_CONSTEXPR auto NumWrapAxes = 3;
-				ISampler::E_TEXTURE_CLAMP axisWraps[NumWrapAxes] = {ISampler::ETC_REPEAT,ISampler::ETC_REPEAT,ISampler::ETC_REPEAT};
+				ISampler::E_TEXTURE_CLAMP axisWraps[NumWrapAxes] = {ISampler::E_TEXTURE_CLAMP::ETC_REPEAT,ISampler::E_TEXTURE_CLAMP::ETC_REPEAT,ISampler::E_TEXTURE_CLAMP::ETC_REPEAT};
 				ISampler::E_TEXTURE_BORDER_COLOR borderColor;
 				VkOffset3D relativeOffset;
 				VkExtent3D paddedExtent;
@@ -57,7 +57,7 @@ class CPaddedCopyImageFilter : public CImageFilter<CPaddedCopyImageFilter>, publ
 			// TODO: eventually remove when we can encode blocks
 			for (auto i=0; i<CState::NumWrapAxes; i++)
 			{
-				if ((isBlockCompressionFormat(inFormat)||isBlockCompressionFormat(outFormat))&&state->axisWraps[i]!=ISampler::ETC_REPEAT)
+				if ((isBlockCompressionFormat(inFormat)||isBlockCompressionFormat(outFormat))&&state->axisWraps[i]!=ISampler::E_TEXTURE_CLAMP::ETC_REPEAT)
 					return false;
 			}
 

include/nbl/builtin/hlsl/enums.hlsl

@@ -32,6 +32,22 @@ enum ShaderStage : uint32_t
     ESS_ALL = 0x7fffffff
 };
 
+enum TextureClamp : uint16_t
+{
+    //! Texture repeats
+    ETC_REPEAT = 0,
+    //! Texture is clamped to the edge pixel
+    ETC_CLAMP_TO_EDGE,
+    //! Texture is clamped to the border pixel (if exists)
+    ETC_CLAMP_TO_BORDER,
+    //! Texture is alternatingly mirrored (0..1..0..1..0..)
+    ETC_MIRROR,
+    //! Texture is mirrored once and then clamped to edge
+    ETC_MIRROR_CLAMP_TO_EDGE,
+
+    ETC_COUNT
+};
+
 enum SampleCountFlags : uint16_t
 {
     ESCF_1_BIT = 0x01,

include/nbl/builtin/hlsl/prefix_sum_blur/blur.hlsl

@@ -0,0 +1,79 @@
+#include "nbl/builtin/hlsl/cpp_compat.hlsl"
+#include "nbl/builtin/hlsl/glsl_compat/core.hlsl"
+#include "nbl/builtin/hlsl/workgroup/basic.hlsl"
+#include "nbl/builtin/hlsl/workgroup/arithmetic.hlsl"
+#include "nbl/builtin/hlsl/device_capabilities_traits.hlsl"
+#include "nbl/builtin/hlsl/enums.hlsl"
+
+#ifndef _NBL_BUILTIN_PREFIX_SUM_BLUR_INCLUDED_
+#define _NBL_BUILTIN_PREFIX_SUM_BLUR_INCLUDED_
+
+namespace nbl
+{
+namespace hlsl
+{
+namespace prefix_sum_blur
+{
+
+// Prefix-Sum Blur using SAT (Summed Area Table) technique.
+// `scanScract` and `_sampler.prefixSumAccessor` must not to alias.
+template<
+    typename DataAccessor,
+    typename ScanSharedAccessor,
+    typename Sampler,
+    uint16_t WorkgroupSize,
+    class device_capabilities=void> // TODO: define concepts for the Box1D and apply constraints
+struct Blur1D
+{
+    void operator()(
+        NBL_REF_ARG(DataAccessor) data,
+        NBL_REF_ARG(ScanSharedAccessor) scanScratch,
+        NBL_REF_ARG(Sampler) _sampler,
+        const uint16_t channel)
+    {
+        const uint16_t end = data.linearSize();
+        const uint16_t localInvocationIndex = workgroup::SubgroupContiguousIndex();
+
+        // prefix sum
+        // note the dynamically uniform loop condition 
+        for (uint16_t baseIx = 0; baseIx < end;)
+        {
+            const uint16_t ix = localInvocationIndex + baseIx;
+            float32_t input = data.template get<float32_t>(channel, ix);
+            // dynamically uniform condition
+            if (baseIx != 0)
+            {
+                // take result of previous prefix sum and add it to first element here
+                if (localInvocationIndex == 0)
+                    input += _sampler.prefixSumAccessor.template get<float32_t>(baseIx - 1);
+            }
+            const float32_t sum = workgroup::inclusive_scan<plus<float32_t>, WorkgroupSize, device_capabilities>::template __call(input, scanScratch);
+            // loop increment
+            baseIx += WorkgroupSize;
+            // save prefix sum results
+            if (ix < end)
+                _sampler.prefixSumAccessor.template set<float32_t>(ix, sum);
+            // previous prefix sum must have finished before we ask for results
+            _sampler.prefixSumAccessor.workgroupExecutionAndMemoryBarrier();
+        }
+
+        // TODO: split this Blur1D into two separate functors:
+        //     - multi-wg-wide prefix sum
+        //     - the SAT sampling
+        const float32_t last = end - 1;
+        for (float32_t ix = localInvocationIndex; ix < end; ix += WorkgroupSize)
+        {
+            const float32_t result = _sampler(ix, radius, borderColor[channel]);
+            data.template set<float32_t>(channel, uint16_t(ix), result);
+        }
+    }
+
+    vector<float32_t, DataAccessor::Channels> borderColor;
+    float32_t radius;
+};
+
+}
+}
+}
+
+#endif