lay the groundwork for half-float storage choice

devshgraphicsprogramming · devshgraphicsprogramming · commit c3326d20e668 · 2021-03-23T22:25:45.000+01:00
diff --git a/examples_tests/49.ComputeFFT/fft_convolve_ifft.comp b/examples_tests/49.ComputeFFT/fft_convolve_ifft.comp
@@ -1,11 +1,11 @@
 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>
 
-#include <nbl/builtin/glsl/math/complex.glsl>
 // Input and Output Descriptor
 layout(set=0, binding=0) buffer restrict InputOutputBuffer
 {
-	nbl_glsl_complex inoutData[];
+	nbl_glsl_ext_FFT_storage_t inoutData[];
 };
 #define _NBL_GLSL_EXT_FFT_INPUT_DESCRIPTOR_DEFINED_
 #define _NBL_GLSL_EXT_FFT_OUTPUT_DESCRIPTOR_DEFINED_
@@ -25,6 +25,7 @@ nbl_glsl_ext_FFT_Parameters_t nbl_glsl_ext_FFT_getParameters()
 }
 #define _NBL_GLSL_EXT_FFT_GET_PARAMETERS_DEFINED_
 
+// do I need to unrestrict these buffers?
 #define inData inoutData
 #define outData inoutData
 #define _NBL_GLSL_EXT_FFT_MAIN_DEFINED_
diff --git a/examples_tests/49.ComputeFFT/main.cpp b/examples_tests/49.ComputeFFT/main.cpp
@@ -35,6 +35,7 @@ R"===(#version 430 core
 
 #define _NBL_GLSL_WORKGROUP_SIZE_ %u
 #define _NBL_GLSL_EXT_FFT_MAX_DIM_SIZE_ %u
+#define _NBL_GLSL_EXT_FFT_HALF_STORAGE_ %u
  
 #include "%s"
 
@@ -48,6 +49,7 @@ R"===(#version 430 core
 		reinterpret_cast<char*>(shader->getPointer()),shader->getSize(), sourceFmt,
 		DEFAULT_WORK_GROUP_SIZE,
 		maxPaddedDimensionSize,
+		0u,
 		includeMainName
 	);
 
@@ -596,10 +598,12 @@ int main()
 	const FFTClass::PaddingType fftPadding[2] = {FFTClass::PaddingType::CLAMP_TO_EDGE,FFTClass::PaddingType::CLAMP_TO_EDGE}; // TODO
 	const auto passes = FFTClass::buildParameters(false,srcNumChannels,srcDim,fftPushConstants,fftDispatchInfo,fftPadding);
 	{
+		fftPushConstants[1].output_strides = fftPushConstants[1].input_strides;
 		fftPushConstants[2] = fftPushConstants[0];
 		{
 			fftPushConstants[2].input_dimensions.w ^= 0x80000000u;
 			fftPushConstants[2].input_dimensions.w &= 0xfffffffdu;
+			fftPushConstants[2].input_strides = fftPushConstants[1].output_strides;
 		}
 		fftDispatchInfo[2] = fftDispatchInfo[0];
 	}
diff --git a/examples_tests/49.ComputeFFT/normalization.comp b/examples_tests/49.ComputeFFT/normalization.comp
@@ -1,11 +1,11 @@
 #version 430 core
 layout(local_size_x=16, local_size_y=16, local_size_z=1) in;
  
-#include "nbl/builtin/glsl/math/complex.glsl"
+#include <nbl/builtin/glsl/ext/FFT/types.glsl>
 
 layout(set=0, binding=0) restrict readonly buffer InBuffer
 {
-	nbl_glsl_complex in_data[];
+	nbl_glsl_ext_FFT_storage_t in_data[];
 };
 
 layout(set=0, binding=1, rg16f) uniform image2D NormalizedKernel[3];
@@ -18,9 +18,9 @@ layout(push_constant) uniform PushConstants
 
 void main()
 {
-	nbl_glsl_complex value = in_data[nbl_glsl_dot(gl_GlobalInvocationID,pc.strides.xyz)];
+	nbl_glsl_complex value = nbl_glsl_ext_FFT_storage_t_get(in_data[nbl_glsl_dot(gl_GlobalInvocationID,pc.strides.xyz)]);
 	
-	const float power = length(in_data[0]);
+	const float power = length(nbl_glsl_ext_FFT_storage_t_get(in_data[0]));
 
 	const uvec2 coord = bitfieldReverse(gl_GlobalInvocationID.xy)>>pc.bitreverse_shift.xy;
 	const nbl_glsl_complex shift = nbl_glsl_expImaginary(-nbl_glsl_PI*float(coord.x+coord.y));
diff --git a/include/nbl/builtin/glsl/ext/FFT/default_compute_fft.comp b/include/nbl/builtin/glsl/ext/FFT/default_compute_fft.comp
@@ -4,7 +4,7 @@ layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) i
 #endif
 
 
-#include <nbl/builtin/glsl/math/complex.glsl>
+#include <nbl/builtin/glsl/ext/FFT/types.glsl>
 
 
 // Input Descriptor
@@ -22,7 +22,7 @@ layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) i
 
 layout(set=_NBL_GLSL_EXT_FFT_INPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_INPUT_BINDING_DEFINED_) readonly restrict buffer InputBuffer
 {
-	nbl_glsl_complex inData[];
+	nbl_glsl_ext_FFT_storage_t inData[];
 };
 #endif
 
@@ -37,16 +37,13 @@ layout(set=_NBL_GLSL_EXT_FFT_INPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_INPUT
 
 #ifndef _NBL_GLSL_EXT_FFT_OUTPUT_DESCRIPTOR_DEFINED_
 #define _NBL_GLSL_EXT_FFT_OUTPUT_DESCRIPTOR_DEFINED_
-layout(set=_NBL_GLSL_EXT_FFT_OUTPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_OUTPUT_BINDING_DEFINED_) restrict buffer OutputBuffer
+layout(set=_NBL_GLSL_EXT_FFT_OUTPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_OUTPUT_BINDING_DEFINED_) writeonly restrict buffer OutputBuffer
 {
-	nbl_glsl_complex outData[];
+	nbl_glsl_ext_FFT_storage_t outData[];
 };
 #endif
 
 
-#include <nbl/builtin/glsl/ext/FFT/parameters.glsl>
-
-
 #ifndef _NBL_GLSL_EXT_FFT_PUSH_CONSTANTS_DEFINED_
 layout(push_constant) uniform PushConstants
 {
@@ -67,7 +64,7 @@ layout(push_constant) uniform PushConstants
 void nbl_glsl_ext_FFT_setData(in uvec3 coordinate, in uint channel, in nbl_glsl_complex complex_value)
 {
 	const uint index = nbl_glsl_dot(uvec4(coordinate,channel),nbl_glsl_ext_FFT_Parameters_t_getOutputStrides());
-	outData[index] = complex_value;
+	nbl_glsl_ext_FFT_storage_t_set(outData[index],complex_value);
 }
 #define _NBL_GLSL_EXT_FFT_SET_DATA_DEFINED_
 #endif
@@ -87,7 +84,7 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in uvec3 coordinate, in uint cha
 	}
 	
 	const uint index = nbl_glsl_dot(uvec4(coordinate,channel),nbl_glsl_ext_FFT_Parameters_t_getInputStrides());
-	return inData[index];
+	return nbl_glsl_ext_FFT_storage_t_get(inData[index]);
 }
 #define _NBL_GLSL_EXT_FFT_GET_PADDED_DATA_DEFINED_
 #endif
@@ -101,6 +98,6 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in uvec3 coordinate, in uint cha
 void main()
 {
 	for(uint ch=0u; ch<=nbl_glsl_ext_FFT_Parameters_t_getMaxChannel(); ++ch)
-		nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_Parameters_t_getIsInverse(), ch);
+		nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_Parameters_t_getIsInverse(),ch);
 }
 #endif
diff --git a/include/nbl/builtin/glsl/ext/FFT/types.glsl b/include/nbl/builtin/glsl/ext/FFT/types.glsl
@@ -0,0 +1,35 @@
+// Copyright (C) 2018-2020 - DevSH Graphics Programming Sp. z O.O.
+// This file is part of the "Nabla Engine".
+// For conditions of distribution and use, see copyright notice in nabla.h
+#ifndef _NBL_GLSL_EXT_FFT_TYPES_INCLUDED_
+#define _NBL_GLSL_EXT_FFT_TYPES_INCLUDED_
+
+#include <nbl/builtin/glsl/math/complex.glsl>
+
+#if _NBL_GLSL_EXT_FFT_HALF_STORAGE_!=0
+#define nbl_glsl_ext_FFT_storage_t uint
+
+nbl_glsl_complex nbl_glsl_ext_FFT_storage_t_get(in nbl_glsl_ext_FFT_storage_t _in)
+{
+	return unpackHalf2x16(_in);
+}
+void nbl_glsl_ext_FFT_storage_t_set(out nbl_glsl_ext_FFT_storage_t _out, in nbl_glsl_complex _in)
+{
+	_out = packHalf2x16(_in);
+}
+#else
+#define nbl_glsl_ext_FFT_storage_t nbl_glsl_complex
+
+nbl_glsl_complex nbl_glsl_ext_FFT_storage_t_get(in nbl_glsl_ext_FFT_storage_t _in)
+{
+	return _in;
+}
+void nbl_glsl_ext_FFT_storage_t_set(out nbl_glsl_ext_FFT_storage_t _out, in nbl_glsl_complex _in)
+{
+	_out = _in;
+}
+#endif
+
+#include <nbl/builtin/glsl/ext/FFT/parameters.glsl>
+
+#endif
diff --git a/src/nbl/builtin/CMakeLists.txt b/src/nbl/builtin/CMakeLists.txt
@@ -139,6 +139,7 @@ set(nbl_resources_to_embed
   "nbl/builtin/glsl/ext/FFT/fft.glsl"
   "nbl/builtin/glsl/ext/FFT/parameters_struct.glsl"
   "nbl/builtin/glsl/ext/FFT/parameters.glsl"
+  "nbl/builtin/glsl/ext/FFT/types.glsl"
   "nbl/builtin/glsl/ext/LumaMeter/common.glsl"
   "nbl/builtin/glsl/ext/LumaMeter/impl.glsl"
   "nbl/builtin/glsl/ext/ToneMapper/operators.glsl"
diff --git a/src/nbl/ext/FFT/FFT.cpp b/src/nbl/ext/FFT/FFT.cpp
@@ -70,19 +70,21 @@ R"===(#version 430 core
 
 #define _NBL_GLSL_WORKGROUP_SIZE_ %u
 #define _NBL_GLSL_EXT_FFT_MAX_DIM_SIZE_ %u
+#define _NBL_GLSL_EXT_FFT_HALF_STORAGE_ %u
  
 layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) in;
 #include "nbl/builtin/glsl/ext/FFT/default_compute_fft.comp"
 
 )===";
 
-	constexpr size_t extraSize = 8u*2u;
+	constexpr size_t extraSize = 8u*2u+1u;
 
 	auto source = core::make_smart_refctd_ptr<ICPUBuffer>(strlen(sourceFmt)+extraSize+1u);
 	snprintf(
 		reinterpret_cast<char*>(source->getPointer()),source->getSize(), sourceFmt,
 		DEFAULT_WORK_GROUP_SIZE,
-		maxPaddedDimensionSize
+		maxPaddedDimensionSize,
+		0u
 	);
 
 	auto shader = driver->createGPUShader(core::make_smart_refctd_ptr<ICPUShader>(std::move(source),asset::ICPUShader::buffer_contains_glsl));

Original file line number	Diff line number	Diff line change
`@@ -35,6 +35,7 @@ R"===(#version 430 core`
`35`	`35`
`36`	`36`	`#define _NBL_GLSL_WORKGROUP_SIZE_ %u`
`37`	`37`	`#define _NBL_GLSL_EXT_FFT_MAX_DIM_SIZE_ %u`
	`38`	`+#define _NBL_GLSL_EXT_FFT_HALF_STORAGE_ %u`
`38`	`39`
`39`	`40`	`#include "%s"`
`40`	`41`
`@@ -48,6 +49,7 @@ R"===(#version 430 core`
`48`	`49`	`reinterpret_cast<char*>(shader->getPointer()),shader->getSize(), sourceFmt,`
`49`	`50`	`DEFAULT_WORK_GROUP_SIZE,`
`50`	`51`	`maxPaddedDimensionSize,`
	`52`	`+ 0u,`
`51`	`53`	`includeMainName`
`52`	`54`	`);`
`53`	`55`
`@@ -596,10 +598,12 @@ int main()`
`596`	`598`	`const FFTClass::PaddingType fftPadding[2] = {FFTClass::PaddingType::CLAMP_TO_EDGE,FFTClass::PaddingType::CLAMP_TO_EDGE}; // TODO`
`597`	`599`	`const auto passes = FFTClass::buildParameters(false,srcNumChannels,srcDim,fftPushConstants,fftDispatchInfo,fftPadding);`
`598`	`600`	`{`
	`601`	`+ fftPushConstants[1].output_strides = fftPushConstants[1].input_strides;`
`599`	`602`	`fftPushConstants[2] = fftPushConstants[0];`
`600`	`603`	`{`
`601`	`604`	`fftPushConstants[2].input_dimensions.w ^= 0x80000000u;`
`602`	`605`	`fftPushConstants[2].input_dimensions.w &= 0xfffffffdu;`
	`606`	`+ fftPushConstants[2].input_strides = fftPushConstants[1].output_strides;`
`603`	`607`	`}`
`604`	`608`	`fftDispatchInfo[2] = fftDispatchInfo[0];`
`605`	`609`	`}`
Original file line number	Diff line number	Diff line change
`@@ -4,7 +4,7 @@ layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) i`
`4`	`4`	`#endif`
`5`	`5`
`6`	`6`
`7`		`-#include <nbl/builtin/glsl/math/complex.glsl>`
	`7`	`+#include <nbl/builtin/glsl/ext/FFT/types.glsl>`
`8`	`8`
`9`	`9`
`10`	`10`	`// Input Descriptor`
`@@ -22,7 +22,7 @@ layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) i`
`22`	`22`
`23`	`23`	`layout(set=_NBL_GLSL_EXT_FFT_INPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_INPUT_BINDING_DEFINED_) readonly restrict buffer InputBuffer`
`24`	`24`	`{`
`25`		`- nbl_glsl_complex inData[];`
	`25`	`+ nbl_glsl_ext_FFT_storage_t inData[];`
`26`	`26`	`};`
`27`	`27`	`#endif`
`28`	`28`
`@@ -37,16 +37,13 @@ layout(set=_NBL_GLSL_EXT_FFT_INPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_INPUT`
`37`	`37`
`38`	`38`	`#ifndef _NBL_GLSL_EXT_FFT_OUTPUT_DESCRIPTOR_DEFINED_`
`39`	`39`	`#define _NBL_GLSL_EXT_FFT_OUTPUT_DESCRIPTOR_DEFINED_`
`40`		`-layout(set=_NBL_GLSL_EXT_FFT_OUTPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_OUTPUT_BINDING_DEFINED_) restrict buffer OutputBuffer`
	`40`	`+layout(set=_NBL_GLSL_EXT_FFT_OUTPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_OUTPUT_BINDING_DEFINED_) writeonly restrict buffer OutputBuffer`
`41`	`41`	`{`
`42`		`- nbl_glsl_complex outData[];`
	`42`	`+ nbl_glsl_ext_FFT_storage_t outData[];`
`43`	`43`	`};`
`44`	`44`	`#endif`
`45`	`45`
`46`	`46`
`47`		`-#include <nbl/builtin/glsl/ext/FFT/parameters.glsl>`
`48`		`-`
`49`		`-`
`50`	`47`	`#ifndef _NBL_GLSL_EXT_FFT_PUSH_CONSTANTS_DEFINED_`
`51`	`48`	`layout(push_constant) uniform PushConstants`
`52`	`49`	`{`
`@@ -67,7 +64,7 @@ layout(push_constant) uniform PushConstants`
`67`	`64`	`void nbl_glsl_ext_FFT_setData(in uvec3 coordinate, in uint channel, in nbl_glsl_complex complex_value)`
`68`	`65`	`{`
`69`	`66`	`const uint index = nbl_glsl_dot(uvec4(coordinate,channel),nbl_glsl_ext_FFT_Parameters_t_getOutputStrides());`
`70`		`- outData[index] = complex_value;`
	`67`	`+ nbl_glsl_ext_FFT_storage_t_set(outData[index],complex_value);`
`71`	`68`	`}`
`72`	`69`	`#define _NBL_GLSL_EXT_FFT_SET_DATA_DEFINED_`
`73`	`70`	`#endif`
`@@ -87,7 +84,7 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in uvec3 coordinate, in uint cha`
`87`	`84`	`}`
`88`	`85`
`89`	`86`	`const uint index = nbl_glsl_dot(uvec4(coordinate,channel),nbl_glsl_ext_FFT_Parameters_t_getInputStrides());`
`90`		`- return inData[index];`
	`87`	`+ return nbl_glsl_ext_FFT_storage_t_get(inData[index]);`
`91`	`88`	`}`
`92`	`89`	`#define _NBL_GLSL_EXT_FFT_GET_PADDED_DATA_DEFINED_`
`93`	`90`	`#endif`
`@@ -101,6 +98,6 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in uvec3 coordinate, in uint cha`
`101`	`98`	`void main()`
`102`	`99`	`{`
`103`	`100`	`for(uint ch=0u; ch<=nbl_glsl_ext_FFT_Parameters_t_getMaxChannel(); ++ch)`
`104`		`- nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_Parameters_t_getIsInverse(), ch);`
	`101`	`+ nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_Parameters_t_getIsInverse(),ch);`
`105`	`102`	`}`
`106`	`103`	`#endif`