take a step toward not doing stupid reads in the convolution

Author: devshgraphicsprogramming

examples_tests/49.ComputeFFT/fft_convolve_ifft.comp

@@ -26,10 +26,11 @@ layout(set=0, binding=1) restrict readonly buffer KernelBuffer
 // Get/Set Data Function
 layout(push_constant) uniform PushConstants
 {
-	layout (offset = 0) nbl_glsl_ext_FFT_Parameters_t params;
+	nbl_glsl_ext_FFT_Parameters_t params;
 } pc;
 
- nbl_glsl_ext_FFT_Parameters_t nbl_glsl_ext_FFT_getParameters() {
+ nbl_glsl_ext_FFT_Parameters_t nbl_glsl_ext_FFT_getParameters()
+ {
 	 nbl_glsl_ext_FFT_Parameters_t ret;
 	 ret = pc.params;
 	 return ret;
@@ -67,24 +68,18 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in uvec3 coordinate, in uint cha
 	return nbl_glsl_ext_FFT_getData(clamped_coord, channel);
 }
 
-void convolve() 
+void convolve(in uint ch) 
 {
-	uint numChannels = nbl_glsl_ext_FFT_Parameters_t_getNumChannels();
 	uvec3 dimension = nbl_glsl_ext_FFT_Parameters_t_getDimensions();
-    uint dataLength = nbl_glsl_ext_FFT_Parameters_t_getFFTLength();
-
-    uint thread_offset = gl_LocalInvocationIndex;
-    uint num_virtual_threads = (dataLength-1u)/(_NBL_GLSL_WORKGROUP_SIZE_)+1u;
-
-	for(uint ch = 0u; ch < numChannels; ++ch) {
-		for(uint t = 0u; t < num_virtual_threads; t++)
-		{
-			uint tid = thread_offset + t * _NBL_GLSL_WORKGROUP_SIZE_;
-			uvec3 coords = nbl_glsl_ext_FFT_getCoordinates(tid);
-			uint idx = ch * (dimension.x * dimension.y * dimension.z) + coords.z * (dimension.x * dimension.y) + coords.y * (dimension.x) + coords.x;
-			nbl_glsl_complex temp = inoutData[idx];
-			inoutData[idx] = nbl_glsl_complex_mul(temp, kerData[idx]);
-		}
+
+    const uint item_per_thread_count = nbl_glsl_ext_FFT_Parameters_t_getFFTLength()>>_NBL_GLSL_WORKGROUP_SIZE_LOG2_;
+	for(uint t = 0u; t < item_per_thread_count; t++)
+	{
+		uint tid = gl_LocalInvocationIndex + t * _NBL_GLSL_WORKGROUP_SIZE_;
+		uvec3 coords = nbl_glsl_ext_FFT_getCoordinates(tid);
+		uint idx = ch * (dimension.x * dimension.y * dimension.z) + coords.z * (dimension.x * dimension.y) + coords.y * (dimension.x) + coords.x;
+		nbl_glsl_complex temp = inoutData[idx];
+		inoutData[idx] = nbl_glsl_complex_mul(temp, kerData[idx]);
 	}
 }
 
@@ -93,17 +88,12 @@ void main()
     const uint numChannels = nbl_glsl_ext_FFT_Parameters_t_getNumChannels();
 	for(uint ch = 0u; ch < numChannels; ++ch)
 	{
-		nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_Parameters_t_getIsInverse(), ch);
-	}
-
-    barrier();
+		nbl_glsl_ext_FFT(false, ch);
+		barrier();
 
-	convolve();	// inoutData+kerData->inoutData
+		convolve(ch);	// inoutData+kerData->inoutData
 
-    barrier();
-
-	for(uint ch = 0u; ch < numChannels; ++ch)
-	{
-		nbl_glsl_ext_FFT(!nbl_glsl_ext_FFT_Parameters_t_getIsInverse(), ch); // inoutData->inoutData
+		barrier();
+		nbl_glsl_ext_FFT(true, ch); // inoutData->inoutData
 	}
 }

examples_tests/49.ComputeFFT/last_fft.comp

@@ -24,7 +24,7 @@ layout(set=0, binding=1, rgba16f) uniform image2D outImage;
 
 layout(push_constant) uniform PushConstants
 {
-	layout (offset = 0) nbl_glsl_ext_FFT_Parameters_t params;
+	nbl_glsl_ext_FFT_Parameters_t params;
     layout (offset = 32) uvec3 kernel_dimension;
 } pc;
 
@@ -72,6 +72,6 @@ void main()
     const uint numChannels = nbl_glsl_ext_FFT_Parameters_t_getNumChannels();
 	for(uint ch = 0u; ch < numChannels; ++ch)
 	{
-		nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_Parameters_t_getIsInverse(), ch);
+		nbl_glsl_ext_FFT(true, ch);
 	}
 }

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

@@ -5,12 +5,12 @@ layout(local_size_x=256, local_size_y=1, local_size_z=1) in;
 
 layout(set=0, binding=0) restrict readonly buffer InBuffer
 {
-	complex_value in_data[];
+	nbl_glsl_complex in_data[];
 };
 
 layout(set=0, binding=1) restrict buffer OutBuffer
 {
-	complex_value out_data[];
+	nbl_glsl_complex out_data[];
 };
 
 void main()