fft_convolve_ifft shader start

+ cleanups

Author: Erfan-Ahmadi

examples_tests/49.ComputeFFT/fft_convolve_ifft.comp

@@ -0,0 +1,96 @@
+// WorkGroup Size
+
+#define _NBL_GLSL_WORKGROUP_SIZE_ 256
+
+layout(local_size_x=_NBL_GLSL_WORKGROUP_SIZE_, local_size_y=1, local_size_z=1) in;
+
+ 
+#define _NBL_GLSL_EXT_FFT_GET_PARAMETERS_DEFINED_
+#define _NBL_GLSL_EXT_FFT_GET_DATA_DEFINED_
+#define _NBL_GLSL_EXT_FFT_SET_DATA_DEFINED_
+#define _NBL_GLSL_EXT_FFT_GET_PADDED_DATA_DEFINED_
+#include "nbl/builtin/glsl/ext/FFT/fft.glsl"
+
+// Input Descriptor
+
+layout(set=0, binding=0) readonly restrict buffer InputBuffer
+{
+	nbl_glsl_complex inData[];
+};
+
+layout(set=0, binding=1) restrict readonly buffer KernelBuffer
+{
+	nbl_glsl_complex kerData[];
+};
+
+// Output Descriptor
+
+layout(set=0, binding=2) restrict buffer OutputBuffer
+{
+	nbl_glsl_complex outData[];
+};
+
+// Get/Set Data Function
+layout(push_constant) uniform PushConstants
+{
+	layout (offset = 0) 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 ret;
+	 ret = pc.params;
+	 return ret;
+ }
+
+nbl_glsl_complex nbl_glsl_ext_FFT_getData(in uvec3 coordinate, in uint channel)
+{
+	nbl_glsl_complex retValue = nbl_glsl_complex(0, 0);
+	uvec3 dimension = nbl_glsl_ext_FFT_getDimensions();
+	uint index = channel * (dimension.x * dimension.y * dimension.z) + coordinate.z * (dimension.x * dimension.y) + coordinate.y * (dimension.x) + coordinate.x;
+	retValue = inData[index];
+	return retValue;
+}
+
+void nbl_glsl_ext_FFT_setData(in uvec3 coordinate, in uint channel, in nbl_glsl_complex complex_value)
+{
+	uvec3 dimension = nbl_glsl_ext_FFT_getPaddedDimensions();
+	uint index = channel * (dimension.x * dimension.y * dimension.z) + coordinate.z * (dimension.x * dimension.y) + coordinate.y * (dimension.x) + coordinate.x;
+	outData[index] = complex_value;
+}
+
+nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in uvec3 coordinate, in uint channel) {
+
+	uvec3 max_coord = nbl_glsl_ext_FFT_getDimensions() - uvec3(1u);
+	uvec3 clamped_coord = min(coordinate, max_coord);
+	
+	bool is_out_of_range = any(bvec3(coordinate!=clamped_coord));
+
+	uint paddingType = nbl_glsl_ext_FFT_getPaddingType();
+
+	if (_NBL_GLSL_EXT_FFT_FILL_WITH_ZERO_ == paddingType && is_out_of_range) {
+		return nbl_glsl_complex(0, 0);
+	}
+	
+	return nbl_glsl_ext_FFT_getData(clamped_coord, channel);
+}
+
+void convolve() 
+{
+	uint idx = 0;
+    inData[idx] = nbl_glsl_complex_mul(outData[idx], kerData[idx]);
+}
+
+void main()
+{
+	nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_getIsInverse()); // inData->outData
+	
+    barrier();
+    memoryBarrierShared();
+
+	convolve();	// outData+kerData->inData
+	
+    barrier();
+    memoryBarrierShared();
+
+	nbl_glsl_ext_FFT(!nbl_glsl_ext_FFT_getIsInverse()); // inData->outData
+}

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

@@ -62,21 +62,14 @@ layout(set=_NBL_GLSL_EXT_FFT_OUTPUT_SET_DEFINED_, binding=_NBL_GLSL_EXT_FFT_OUTP
 
 // Get/Set Data Function
 
-//TODO: investigate why putting this uint between the 2 other uvec3's don't work
 layout(push_constant) uniform PushConstants
 {
 	layout (offset = 0) nbl_glsl_ext_FFT_Parameters_t params;
-    // layout (offset = 0) uvec3 dimension;
-    // layout (offset = 16) uvec3 padded_dimension;
-	// layout (offset = 32) uint direction_isInverse_paddingType; // packed into a uint
 } pc;
 
  nbl_glsl_ext_FFT_Parameters_t nbl_glsl_ext_FFT_getParameters() {
 	 nbl_glsl_ext_FFT_Parameters_t ret;
 	 ret = pc.params;
-	 // ret.dimension = pc.dimension;
-	 // ret.direction_isInverse_paddingType = pc.direction_isInverse_paddingType;
-	 // ret.padded_dimension = pc.padded_dimension;
 	 return ret;
  }
 
@@ -119,7 +112,7 @@ nbl_glsl_complex nbl_glsl_ext_FFT_getPaddedData(in uvec3 coordinate, in uint cha
 
 void main()
 {
-	nbl_glsl_ext_FFT();
+	nbl_glsl_ext_FFT(nbl_glsl_ext_FFT_getIsInverse());
 }
 
 #endif

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

@@ -147,7 +147,7 @@ uint nbl_glsl_ext_FFT_getDimLength(uvec3 dimension)
     return dimension[direction];
 }
 
-void nbl_glsl_ext_FFT()
+void nbl_glsl_ext_FFT(bool is_inverse)
 {
     nbl_glsl_ext_FFT_Parameters_t params = nbl_glsl_ext_FFT_getParameters();
     // Virtual Threads Calculation
@@ -157,8 +157,6 @@ void nbl_glsl_ext_FFT()
 
 	uint channel = nbl_glsl_ext_FFT_getChannel();
 
-    bool is_inverse = nbl_glsl_ext_FFT_getIsInverse();
-
 	// Pass 0: Bit Reversal
 	uint leadingZeroes = nbl_glsl_clz(dataLength) + 1u;
 	uint logTwo = 32u - leadingZeroes;

include/nbl/ext/FFT/FFT.h

@@ -226,15 +226,11 @@ class FFT : public core::TotalInterface
 			params.dimension.y = inputDimension.height;
 			params.dimension.z = inputDimension.depth;
 			params.dimension.w = packed;
-			// params.direction_isInverse_paddingType = packed;
 			params.padded_dimension.x = paddedInputDimension.width;
 			params.padded_dimension.y = paddedInputDimension.height;
 			params.padded_dimension.z = paddedInputDimension.depth;
 
 			driver->pushConstants(pipelineLayout, nbl::video::IGPUSpecializedShader::ESS_COMPUTE, 0u, sizeof(Parameters_t), &params);
-			// driver->pushConstants(pipelineLayout, nbl::video::IGPUSpecializedShader::ESS_COMPUTE, 0u, sizeof(uint32_t) * 3, &inputDimension);
-			// driver->pushConstants(pipelineLayout, nbl::video::IGPUSpecializedShader::ESS_COMPUTE, sizeof(uint32_t) * 4, sizeof(uint32_t) * 3, &paddedInputDimension);
-			// driver->pushConstants(pipelineLayout, nbl::video::IGPUSpecializedShader::ESS_COMPUTE, sizeof(uint32_t) * 8, sizeof(uint32_t), &packed);
 		}
 
 		// Kernel Normalization