idea... support multi-channel FFT to reuse twiddle factors and reduce data-reads and writes

Author: devshgraphicsprogramming

examples_tests/49.ComputeFFT/main.cpp

@@ -44,7 +44,7 @@ R"===(#version 430 core
 
 	const size_t extraSize = 4u+8u+8u+128u;
 
-	constexpr uint32_t DEFAULT_WORK_GROUP_SIZE = 256u;
+	constexpr uint32_t DEFAULT_WORK_GROUP_SIZE = FFTClass::DEFAULT_WORK_GROUP_SIZE;
 	auto shader = core::make_smart_refctd_ptr<ICPUBuffer>(strlen(sourceFmt)+extraSize+1u);
 	snprintf(
 		reinterpret_cast<char*>(shader->getPointer()),shader->getSize(), sourceFmt,

include/nbl/builtin/glsl/ext/FFT/default_compute_fft.comp

@@ -71,6 +71,7 @@ void nbl_glsl_ext_FFT_setData(in uvec3 coordinate, in uint channel, in nbl_glsl_
 #define _NBL_GLSL_EXT_FFT_SET_DATA_DEFINED_
 #endif
 
+// TODO: pad by using shared memory to broadcast copies of the input around (up to 49% VRAM read gains) ?
 bool nbl_glsl_ext_FFT_wrap_coord(inout ivec3 coordinate)
 {
 	const ivec3 dimensions = ivec3(nbl_glsl_ext_FFT_Parameters_t_getDimensions());

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

@@ -52,7 +52,7 @@ ivec3 nbl_glsl_ext_FFT_getPaddedCoordinates(in uint tidx, in uint log2FFTSize, i
 #include "nbl/builtin/glsl/workgroup/fft.glsl"
 
 
-nbl_glsl_complex nbl_glsl_ext_FFT_impl_values[(_NBL_GLSL_EXT_FFT_MAX_DIM_SIZE_-1u)/_NBL_GLSL_WORKGROUP_SIZE_+1u];
+nbl_glsl_complex nbl_glsl_ext_FFT_impl_values[(_NBL_GLSL_EXT_FFT_MAX_DIM_SIZE_-1u)/_NBL_GLSL_WORKGROUP_SIZE_+1u]; // TODO: hardcoded channel count, multichannel FFT (shared twiddles), maybe even precompute twiddles into a LUT!?
 
 void nbl_glsl_ext_FFT_loop(in bool is_inverse, in uint virtual_thread_count, in uint step)
 {