fix nbl_glsl_expImaginary, use shift theorem to get rid of coordiante shift when storing to image

Author: devshgraphicsprogramming

examples_tests/49.ComputeFFT/fft_convolve_ifft.comp

@@ -76,9 +76,10 @@ void convolve(in uint item_per_thread_count, in uint ch)
 		uvec3 coords = nbl_glsl_ext_FFT_getCoordinates(tid);
 		//coords &= uvec3(0xffeu);
 		coords = bitfieldReverse(coords)>>uvec3(21u,22u,0u);
-		//nbl_glsl_complex shift = nbl_glsl_expImaginary(-nbl_glsl_PI*float(k+l));
+		const nbl_glsl_complex shift = nbl_glsl_expImaginary(-dot(vec2(coords.xy),vec2(512.f)/vec2(2048.f,1024.f))*nbl_glsl_PI); // TODO: does this shift go away later?
 		//
-		nbl_glsl_ext_FFT_impl_values[t] = nbl_glsl_complex_mul(nbl_glsl_ext_FFT_impl_values[t],texelFetch(NormalizedKernel[ch],ivec2(coords.xy),0).xy);
+		nbl_glsl_complex value = nbl_glsl_complex_mul(nbl_glsl_ext_FFT_impl_values[t],shift);
+		nbl_glsl_ext_FFT_impl_values[t] = nbl_glsl_complex_mul(texelFetch(NormalizedKernel[ch],ivec2(coords.xy),0).xy,value);
 	}
 }
 

examples_tests/49.ComputeFFT/last_fft.comp

@@ -25,7 +25,6 @@ layout(set=0, binding=1, rgba16f) uniform image2D outImage;
 layout(push_constant) uniform PushConstants
 {
 	nbl_glsl_ext_FFT_Parameters_t params;
-    layout (offset = 32) uvec3 kernel_dimension;
 } pc;
 
  nbl_glsl_ext_FFT_Parameters_t nbl_glsl_ext_FFT_getParameters() {
@@ -45,8 +44,8 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getData(in uvec3 coordinate, in uint channel)
 
 void nbl_glsl_ext_FFT_setData(in uvec3 coordinate, in uint channel, in nbl_glsl_complex complex_value)
 {
-    ivec2 coords = ivec2(coordinate.xy) - ivec2(pc.kernel_dimension.xy / 2);
-    //const ivec2 coords = ivec2(coordinate.xy);
+	const ivec2 coords = ivec2(coordinate.xy);
+
     vec4 color_value = imageLoad(outImage, coords);
     color_value[channel] = complex_value.x;
     imageStore(outImage, coords, color_value);

examples_tests/49.ComputeFFT/main.cpp

@@ -160,20 +160,6 @@ static inline core::smart_refctd_ptr<video::IGPUPipelineLayout> getPipelineLayou
 
 	using FFTClass = ext::FFT::FFT;
 
-	static const asset::SPushConstantRange ranges[2] =
-	{
-		{
-			ISpecializedShader::ESS_COMPUTE,
-			0u,
-			sizeof(FFTClass::Parameters_t)
-		},
-		{
-			ISpecializedShader::ESS_COMPUTE,
-			sizeof(FFTClass::Parameters_t),
-			sizeof(uint32_t) * 3
-		},
-	};
-
 	static IGPUDescriptorSetLayout::SBinding bnd[] =
 	{
 		{
@@ -191,8 +177,9 @@ static inline core::smart_refctd_ptr<video::IGPUPipelineLayout> getPipelineLayou
 			nullptr
 		},
 	};
-	
-	core::SRange<const asset::SPushConstantRange> pcRange = {ranges, ranges+2};
+
+	using FFTClass = ext::FFT::FFT;
+	core::SRange<const asset::SPushConstantRange> pcRange = FFTClass::getDefaultPushConstantRanges();
 	core::SRange<const video::IGPUDescriptorSetLayout::SBinding> bindings = {bnd, bnd+sizeof(bnd)/sizeof(IGPUDescriptorSetLayout::SBinding)};;
 
 	return driver->createGPUPipelineLayout(
@@ -362,9 +349,7 @@ int main()
 	}
 	 // TODO: re-examine
 	const VkExtent3D paddedDim = FFTClass::padDimensionToNextPOT(srcDim);
-	uint32_t maxPaddedDimensionSize = core::max(core::max(paddedDim.width, paddedDim.height), paddedDim.depth);
-	auto fftGPUSpecializedShader_SSBOInput = FFTClass::createShader(driver, FFTClass::DataType::SSBO, maxPaddedDimensionSize);
-	auto fftGPUSpecializedShader_ImageInput = FFTClass::createShader(driver, FFTClass::DataType::TEXTURE2D, maxPaddedDimensionSize);
+	auto fftGPUSpecializedShader_ImageInput = FFTClass::createShader(driver, FFTClass::DataType::TEXTURE2D, paddedDim.width);
 	auto fftGPUSpecializedShader_KernelNormalization = [&]() -> auto
 	{
 		IAssetLoader::SAssetLoadParams lp;
@@ -373,7 +358,6 @@ int main()
 		return *stuff->begin();
 	}();
 
-	auto fftPipelineLayout_SSBOInput = FFTClass::getDefaultPipelineLayout(driver, FFTClass::DataType::SSBO);
 	auto fftPipelineLayout_ImageInput = FFTClass::getDefaultPipelineLayout(driver, FFTClass::DataType::TEXTURE2D);
 	auto fftPipelineLayout_KernelNormalization = [&]() -> auto
 	{
@@ -400,18 +384,17 @@ int main()
 		);
 	}();
 
-	auto fftPipeline_SSBOInput = driver->createGPUComputePipeline(nullptr, core::smart_refctd_ptr(fftPipelineLayout_SSBOInput), std::move(fftGPUSpecializedShader_SSBOInput));
 	auto fftPipeline_ImageInput = driver->createGPUComputePipeline(nullptr, core::smart_refctd_ptr(fftPipelineLayout_ImageInput), std::move(fftGPUSpecializedShader_ImageInput));
 	auto fftPipeline_KernelNormalization = driver->createGPUComputePipeline(nullptr, core::smart_refctd_ptr(fftPipelineLayout_KernelNormalization), std::move(fftGPUSpecializedShader_KernelNormalization));
 
 	auto fftDispatchInfo_Horizontal = FFTClass::buildParameters(paddedDim, FFTClass::Direction::X);
 	auto fftDispatchInfo_Vertical = FFTClass::buildParameters(paddedDim, FFTClass::Direction::Y);
 
-	auto convolveShader = createShader_Convolution(driver, am, maxPaddedDimensionSize);
+	auto convolveShader = createShader_Convolution(driver, am, paddedDim.height);
 	auto convolvePipelineLayout = getPipelineLayout_Convolution(driver);
 	auto convolvePipeline = driver->createGPUComputePipeline(nullptr, core::smart_refctd_ptr(convolvePipelineLayout), std::move(convolveShader));
 
-	auto lastFFTShader = createShader_LastFFT(driver, am, maxPaddedDimensionSize);
+	auto lastFFTShader = createShader_LastFFT(driver, am, paddedDim.width);
 	auto lastFFTPipelineLayout = getPipelineLayout_LastFFT(driver);
 	auto lastFFTPipeline = driver->createGPUComputePipeline(nullptr, core::smart_refctd_ptr(lastFFTPipelineLayout), std::move(lastFFTShader));
 
@@ -452,6 +435,7 @@ int main()
 		FFTClass::updateDescriptorSet(driver, fftDescriptorSet_Ker_FFT_X.get(), kerImageView, fftOutputBuffer_0, ISampler::ETC_CLAMP_TO_BORDER);
 
 		// Ker FFT Y
+		auto fftPipelineLayout_SSBOInput = FFTClass::getDefaultPipelineLayout(driver, FFTClass::DataType::SSBO);
 		auto fftDescriptorSet_Ker_FFT_Y = driver->createGPUDescriptorSet(core::smart_refctd_ptr<const IGPUDescriptorSetLayout>(fftPipelineLayout_SSBOInput->getDescriptorSetLayout(0u)));
 		FFTClass::updateDescriptorSet(driver, fftDescriptorSet_Ker_FFT_Y.get(), fftOutputBuffer_0, fftOutputBuffer_1);
 
@@ -502,6 +486,7 @@ int main()
 		FFTClass::dispatchHelper(driver, fftDispatchInfo_Horizontal);
 
 		// Ker Image FFT Y
+		auto fftPipeline_SSBOInput = driver->createGPUComputePipeline(nullptr, core::smart_refctd_ptr(fftPipelineLayout_SSBOInput), FFTClass::createShader(driver,FFTClass::DataType::SSBO,paddedDim.height));
 		driver->bindComputePipeline(fftPipeline_SSBOInput.get());
 		driver->bindDescriptorSets(EPBP_COMPUTE, fftPipelineLayout_SSBOInput.get(), 0u, 1u, &fftDescriptorSet_Ker_FFT_Y.get(), nullptr);
 		FFTClass::pushConstants(driver, fftPipelineLayout_SSBOInput.get(), paddedDim, paddedDim, FFTClass::Direction::Y, false, srcNumChannels);
@@ -559,7 +544,6 @@ int main()
 		driver->bindComputePipeline(lastFFTPipeline.get());
 		driver->bindDescriptorSets(EPBP_COMPUTE, lastFFTPipelineLayout.get(), 0u, 1u, &lastFFTDescriptorSet.get(), nullptr);
 		FFTClass::pushConstants(driver, lastFFTPipelineLayout.get(), paddedDim, paddedDim, FFTClass::Direction::X, true, srcNumChannels);
-		driver->pushConstants(lastFFTPipelineLayout.get(), nbl::video::IGPUSpecializedShader::ESS_COMPUTE, sizeof(FFTClass::Parameters_t), sizeof(uint32_t) * 3, &kerDim); // numSrcChannels
 		FFTClass::dispatchHelper(driver, fftDispatchInfo_Horizontal);
 
 		if(false == savedToFile) {

include/nbl/builtin/glsl/math/complex.glsl

@@ -17,9 +17,7 @@
 
 nbl_glsl_complex nbl_glsl_expImaginary(in float _theta)
 {
-    nbl_glsl_complex retval;
-    nbl_glsl_sincos(_theta,retval.y,retval.x);
-    return retval;
+    return vec2(cos(_theta),sin(_theta));
 }
 
 nbl_glsl_complex nbl_glsl_complex_mul(in nbl_glsl_complex rhs, in nbl_glsl_complex lhs)