managed to get the work decimation going

Author: devshgraphicsprogramming

examples_tests/49.ComputeFFT/fft_convolve_ifft.comp

@@ -60,10 +60,16 @@ void main()
 		barrier();
 		nbl_glsl_ext_FFT_preloaded(true,log2FFTSize);
 		// write out to main memory
+		// we override the setting to happen with padded coordinates because we dont want the padding to be written at all
 		for(uint t=0u; t<item_per_thread_count; t++)
 		{
 			const uint tid = (t<<_NBL_GLSL_WORKGROUP_SIZE_LOG2_)|gl_LocalInvocationIndex;
-			nbl_glsl_ext_FFT_setData(nbl_glsl_ext_FFT_getCoordinates(tid),ch,nbl_glsl_ext_FFT_impl_values[t]);
+			const uint trueDim = nbl_glsl_ext_FFT_Parameters_t_getDimensions()[nbl_glsl_ext_FFT_Parameters_t_getDirection()];
+			// we also prevent certain threads from writing the memory out
+			const uint padding = ((0x1u<<log2FFTSize)-trueDim)>>1u;
+			const uint shifted = tid-padding;
+			if (tid>=padding && shifted<trueDim)
+				nbl_glsl_ext_FFT_setData(ivec3(nbl_glsl_ext_FFT_getCoordinates(shifted)),ch,nbl_glsl_ext_FFT_impl_values[t]);
 		}
 	}
 }

examples_tests/49.ComputeFFT/last_fft.comp

@@ -8,10 +8,10 @@ layout(set=0, binding=1, rgba16f) uniform image2D outImage;
 void nbl_glsl_ext_FFT_setData(in uvec3 coordinate, in uint channel, in nbl_glsl_complex complex_value)
 {
 // TODO PC
-    const ivec2 padding = imageSize(outImage).x!=512u ? ivec2(384,664):ivec2(0);
+    const ivec2 padding = imageSize(outImage).x!=512u ? ivec2(384,0):ivec2(0);
 	const ivec2 coords = ivec2(coordinate.xy)-padding;
 
-    if (all(lessThanEqual(ivec2(0),coords))&&all(greaterThan(imageSize(outImage),coords)))
+    if (all(lessThanEqual(ivec2(0),coords)) && all(greaterThan(imageSize(outImage),coords)))
     {
         vec4 color_value = imageLoad(outImage, coords);
         color_value[channel] = complex_value.x;

examples_tests/49.ComputeFFT/main.cpp

@@ -614,15 +614,14 @@ 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,paddedSrcDim);
 	{
+		fftPushConstants[1].output_strides = fftPushConstants[1].input_strides; // override for less work and storage (dont need to store the extra Y-slices after iFFT)
 		fftPushConstants[2].input_dimensions = fftPushConstants[1].input_dimensions;
-		fftPushConstants[2].input_dimensions.y = 2048u;
 		{
 			fftPushConstants[2].input_dimensions.w = fftPushConstants[0].input_dimensions.w^0x80000000u;
 			fftPushConstants[2].input_strides = fftPushConstants[1].output_strides;
 			fftPushConstants[2].output_strides = fftPushConstants[0].input_strides;
 		}
 		fftDispatchInfo[2] = fftDispatchInfo[0];
-		fftDispatchInfo[2].workGroupCount[1] = 2048;
 	}
 	assert(passes==2);