get rid of stupid triple roundtrip in convolutiion

Author: devshgraphicsprogramming

examples_tests/49.ComputeFFT/fft_convolve_ifft.comp

@@ -68,32 +68,46 @@ 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(in uint ch) 
+void convolve(in uint item_per_thread_count, in uint ch) 
 {
+// TODO: decouple kernel size from image size
 	uvec3 dimension = nbl_glsl_ext_FFT_Parameters_t_getDimensions();
-
-    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++)
+	
+	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);
+		//coords &= uvec3(0xffeu);
 		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]);
+		nbl_glsl_ext_FFT_impl_values[t] = nbl_glsl_complex_mul(nbl_glsl_ext_FFT_impl_values[t],kerData[idx]);
 	}
 }
 
 void main()
 {
+    const uint dataLength = nbl_glsl_ext_FFT_Parameters_t_getFFTLength();
+    const uint item_per_thread_count = dataLength>>_NBL_GLSL_WORKGROUP_SIZE_LOG2_;
     const uint numChannels = nbl_glsl_ext_FFT_Parameters_t_getNumChannels();
 	for(uint ch = 0u; ch < numChannels; ++ch)
 	{
-		nbl_glsl_ext_FFT(false, ch);
+		// Load Values into local memory
+		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_impl_values[t] = nbl_glsl_ext_FFT_getPaddedData(nbl_glsl_ext_FFT_getCoordinates(tid),ch);
+		}
+		nbl_glsl_ext_FFT_preloaded(false,dataLength);
 		barrier();
 
-		convolve(ch);	// inoutData+kerData->inoutData
+		convolve(item_per_thread_count,ch);
 
 		barrier();
-		nbl_glsl_ext_FFT(true, ch); // inoutData->inoutData
+		nbl_glsl_ext_FFT_preloaded(true,dataLength);
+		// write out to main memory
+		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]);
+		}
 	}
 }

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

@@ -84,23 +84,12 @@ void nbl_glsl_ext_FFT_loop(in bool is_inverse, in uint virtual_thread_count, in
     }
 }
 
-void nbl_glsl_ext_FFT(bool is_inverse, uint channel)
+void nbl_glsl_ext_FFT_preloaded(bool is_inverse, in uint dataLength)
 {
     // Virtual Threads Calculation
-    const uint dataLength = nbl_glsl_ext_FFT_Parameters_t_getFFTLength();
-    const uint item_per_thread_count = dataLength>>_NBL_GLSL_WORKGROUP_SIZE_LOG2_;
-
     const uint halfDataLength = dataLength>>1u;
-    const uint virtual_thread_count = item_per_thread_count>>1u;
+    const uint virtual_thread_count = halfDataLength>>_NBL_GLSL_WORKGROUP_SIZE_LOG2_;
 
-    // Load Values into local memory
-    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_impl_values[t] = nbl_glsl_ext_FFT_getPaddedData(nbl_glsl_ext_FFT_getCoordinates(tid),channel);
-        if (is_inverse)
-            nbl_glsl_ext_FFT_impl_values[t] /= float(virtual_thread_count);
-    }
     // special forward steps
     if (!is_inverse)
     for (uint step=halfDataLength; step>_NBL_GLSL_WORKGROUP_SIZE_; step>>=1u)
@@ -111,11 +100,32 @@ void nbl_glsl_ext_FFT(bool is_inverse, uint channel)
         const uint lo_ix = t<<1u;
         const uint hi_ix = lo_ix|1u;
         nbl_glsl_workgroupFFT(is_inverse,nbl_glsl_ext_FFT_impl_values[lo_ix],nbl_glsl_ext_FFT_impl_values[hi_ix]);
+        if (is_inverse)
+        {
+            nbl_glsl_ext_FFT_impl_values[lo_ix] /= float(virtual_thread_count);
+            nbl_glsl_ext_FFT_impl_values[hi_ix] /= float(virtual_thread_count);
+        }
     }
     // special inverse steps
     if (is_inverse)
     for (uint step=_NBL_GLSL_WORKGROUP_SIZE_<<1u; step<dataLength; step<<=1u)
         nbl_glsl_ext_FFT_loop(true,virtual_thread_count,step);
+}
+
+void nbl_glsl_ext_FFT(bool is_inverse, uint channel)
+{
+    // Virtual Threads Calculation
+    const uint dataLength = nbl_glsl_ext_FFT_Parameters_t_getFFTLength();
+    const uint item_per_thread_count = dataLength>>_NBL_GLSL_WORKGROUP_SIZE_LOG2_;
+
+    // Load Values into local memory
+    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_impl_values[t] = nbl_glsl_ext_FFT_getPaddedData(nbl_glsl_ext_FFT_getCoordinates(tid),channel);
+    }
+    // do FFT
+    nbl_glsl_ext_FFT_preloaded(is_inverse,dataLength);
     // write out to main memory
     for(uint t=0u; t<item_per_thread_count; t++)
     {

include/nbl/builtin/glsl/math/complex.glsl

@@ -43,32 +43,6 @@ nbl_glsl_complex nbl_glsl_complex_conjugate(in nbl_glsl_complex complex) {
 
 
 // FFT
-nbl_glsl_complex nbl_glsl_FFT_half_twiddle(in uint k, in float N)
-{
-    const float arg = -2.f*nbl_glsl_PI*float(k)/N;
-    nbl_glsl_complex retval;
-    retval.x = cos(arg);
-    retval.y = sqrt(1.f-retval.x*retval.x); // twiddle is always half the range, so no conditional -1.f needed
-    return retval;
-}
-nbl_glsl_complex nbl_glsl_FFT_half_twiddle(in uint k, in uint logTwoN)
-{
-    return nbl_glsl_FFT_half_twiddle(k,float(1<<logTwoN));
-}
-
-nbl_glsl_complex nbl_glsl_FFT_half_twiddle(in bool is_inverse, in uint k, in float N)
-{
-    nbl_glsl_complex twiddle = nbl_glsl_FFT_half_twiddle(k,N);
-    if (is_inverse)
-        return nbl_glsl_complex_conjugate(twiddle);
-    return twiddle;
-}
-nbl_glsl_complex nbl_glsl_FFT_half_twiddle(in bool is_inverse, in uint k, in uint logTwoN)
-{
-    return nbl_glsl_FFT_half_twiddle(is_inverse,k,float(1<<logTwoN));
-}
-
-
 nbl_glsl_complex nbl_glsl_FFT_twiddle(in uint k, in float N)
 {
     nbl_glsl_complex retval;