fix coalesced writes on the convolution FFT

prep for hardcoded Parameters_t

Optimization strategies for FFT:
- 2-4-1 optimization for real FFTs
- hardcoded parameters that control loops, branches, coordinate calc and dynamic indexing
- no more dynamic indexing of locals to create axis dependent stuff
- tiled bloom (reduce the log(n) and the register pressure)
- prescale the blur kernel to match the bitreverse permuted spectrum, also use use image or buffer?
- use images as temporary storage (maybe hilbert curve will be nicer for caches)

Author: devshgraphicsprogramming

examples_tests/49.ComputeFFT/fft_convolve_ifft.comp

@@ -1,9 +1,43 @@
 layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) in;
 
-#include <nbl/builtin/glsl/ext/FFT/types.glsl>
-
 layout(set=0, binding=2) uniform sampler2D NormalizedKernel[3];
 
+/* TODO: Hardcode the parameters for the frequent FFTs
+uvec3 nbl_glsl_ext_FFT_Parameters_t_getDimensions()
+{
+	return uvec3(1280u,1024u,1u);
+}
+bool nbl_glsl_ext_FFT_Parameters_t_getIsInverse()
+{
+	return false;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getDirection()
+{
+	return 0u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getMaxChannel()
+{
+    return 2u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getLog2FFTSize()
+{
+    return 11u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getPaddingType()
+{
+    return 3u; // _NBL_GLSL_EXT_FFT_PAD_MIRROR_;
+}
+uvec4 nbl_glsl_ext_FFT_Parameters_t_getInputStrides()
+{
+	return uvec4(1024u,1u,0u,1024u*1280u);
+}
+uvec4 nbl_glsl_ext_FFT_Parameters_t_getOutputStrides()
+{
+	return uvec4(1u,1280u,0u,1280u*1024u);
+}
+#define _NBL_GLSL_EXT_FFT_PARAMETERS_METHODS_DECLARED_
+*/
+
 #include "extra_parameters.glsl"
 layout(push_constant) uniform PushConstants
 {

examples_tests/49.ComputeFFT/image_first_fft.comp

@@ -15,4 +15,42 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in ivec3 coordinate, in uint cha
 #define _NBL_GLSL_EXT_FFT_GET_PADDED_DATA_DEFINED_
 
 
+/* TODO: Hardcode the parameters for the frequent FFTs
+#if _NBL_GLSL_EXT_FFT_MAX_DIM_SIZE_>512
+uvec3 nbl_glsl_ext_FFT_Parameters_t_getDimensions()
+{
+	return uvec3(1280u,720u,1u);
+}
+bool nbl_glsl_ext_FFT_Parameters_t_getIsInverse()
+{
+	return false;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getDirection()
+{
+	return 1u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getMaxChannel()
+{
+    return 2u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getLog2FFTSize()
+{
+    return 10u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getPaddingType()
+{
+    return 3u; // _NBL_GLSL_EXT_FFT_PAD_MIRROR_;
+}
+uvec4 nbl_glsl_ext_FFT_Parameters_t_getInputStrides()
+{
+	return uvec4(0xdeadbeefu);
+}
+uvec4 nbl_glsl_ext_FFT_Parameters_t_getOutputStrides()
+{
+	return uvec4(1024u,1u,0u,1024u*1280u);
+}
+#define _NBL_GLSL_EXT_FFT_PARAMETERS_METHODS_DECLARED_
+#endif
+*/
+
 #include "nbl/builtin/glsl/ext/FFT/default_compute_fft.comp"

examples_tests/49.ComputeFFT/last_fft.comp

@@ -4,6 +4,41 @@ layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) i
 layout(set=0, binding=1, rgba16f) uniform image2D outImage;
 #define _NBL_GLSL_EXT_FFT_OUTPUT_DESCRIPTOR_DEFINED_
 
+/* TODO: Hardcode the parameters for the frequent FFTs
+uvec3 nbl_glsl_ext_FFT_Parameters_t_getDimensions()
+{
+	return uvec3(1280u,1024u,1u);
+}
+bool nbl_glsl_ext_FFT_Parameters_t_getIsInverse()
+{
+	return true;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getDirection()
+{
+	return 1u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getMaxChannel()
+{
+    return 2u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getLog2FFTSize()
+{
+    return 10u;
+}
+uint nbl_glsl_ext_FFT_Parameters_t_getPaddingType()
+{
+    return 3u; // _NBL_GLSL_EXT_FFT_PAD_MIRROR_;
+}
+uvec4 nbl_glsl_ext_FFT_Parameters_t_getInputStrides()
+{
+	return uvec4(1u,1280u,0u,1280u*1024u);
+}
+uvec4 nbl_glsl_ext_FFT_Parameters_t_getOutputStrides()
+{
+	return uvec4(0xdeadbeefu);
+}
+#define _NBL_GLSL_EXT_FFT_PARAMETERS_METHODS_DECLARED_
+*/
 
 #include "extra_parameters.glsl"
 layout(push_constant) uniform PushConstants

examples_tests/49.ComputeFFT/main.cpp

@@ -592,7 +592,12 @@ int main()
 	const ISampler::E_TEXTURE_CLAMP fftPadding[2] = {ISampler::ETC_MIRROR,ISampler::ETC_MIRROR};
 	const auto passes = FFTClass::buildParameters(false,srcNumChannels,srcDim,fftPushConstants,fftDispatchInfo,fftPadding,marginSrcDim);
 	{
-		fftPushConstants[1].output_strides = fftPushConstants[1].input_strides; // override for less work and storage (dont need to store the extra padding of the last axis after iFFT)
+		// override for less work and storage (dont need to store the extra padding of the last axis after iFFT)
+		fftPushConstants[1].output_strides.x = fftPushConstants[0].input_strides.x;
+		fftPushConstants[1].output_strides.y = fftPushConstants[0].input_strides.y;
+		fftPushConstants[1].output_strides.z = fftPushConstants[1].input_strides.z;
+		fftPushConstants[1].output_strides.w = fftPushConstants[1].input_strides.w;
+		// iFFT
 		fftPushConstants[2].input_dimensions = fftPushConstants[1].input_dimensions;
 		{
 			fftPushConstants[2].input_dimensions.w = fftPushConstants[0].input_dimensions.w^0x80000000u;

include/nbl/builtin/glsl/ext/FFT/parameters.glsl

@@ -24,6 +24,8 @@ nbl_glsl_ext_FFT_Parameters_t nbl_glsl_ext_FFT_getParameters();
 #endif
 
 
+#ifndef _NBL_GLSL_EXT_FFT_PARAMETERS_METHODS_DECLARED_
+#define _NBL_GLSL_EXT_FFT_PARAMETERS_METHODS_DECLARED_
 uvec3 nbl_glsl_ext_FFT_Parameters_t_getDimensions()
 {
     nbl_glsl_ext_FFT_Parameters_t params = nbl_glsl_ext_FFT_getParameters();
@@ -67,5 +69,6 @@ uvec4 nbl_glsl_ext_FFT_Parameters_t_getOutputStrides()
     nbl_glsl_ext_FFT_Parameters_t params = nbl_glsl_ext_FFT_getParameters();
     return params.output_strides;
 }
+#endif
 
 #endif