Compositor:

re-optimized the Defocus node.
 * localized MemoryBuffers
 * removed read(x,y) calls
 * shuffled some lines in the execute pixel
 * added a readNoCheck function to the memorybuffer (only use this when
you are certain you are reading a pixel inside the memorybuffer.

5 files changed, 79 insertions, 42 deletions

diff --git a/source/blender/compositor/intern/COM_MemoryBuffer.h b/source/blender/compositor/intern/COM_MemoryBuffer.h
index 51a45efc051..eed0c796cd8 100644
--- a/source/blender/compositor/intern/COM_MemoryBuffer.h
+++ b/source/blender/compositor/intern/COM_MemoryBuffer.h
@@ -140,6 +140,13 @@ public:
 		}
 	}
 
+	inline void readNoCheck(float result[4], int x, int y)  {
+		const int dx = x - this->m_rect.xmin;
+		const int dy = y - this->m_rect.ymin;
+		const int offset = (this->m_chunkWidth * dy + dx) * COM_NUMBER_OF_CHANNELS;
+		copy_v4_v4(result, &this->m_buffer[offset]);
+	}
+	
 	void writePixel(int x, int y, const float color[4]);
 	void addPixel(int x, int y, const float color[4]);
 	inline void readCubic(float result[4], float x, float y)
diff --git a/source/blender/compositor/operations/COM_OpenCLKernels.cl b/source/blender/compositor/operations/COM_OpenCLKernels.cl
index 41838e41fba..cbbb4d0b3f2 100644
--- a/source/blender/compositor/operations/COM_OpenCLKernels.cl
+++ b/source/blender/compositor/operations/COM_OpenCLKernels.cl
@@ -101,16 +101,16 @@ __kernel void defocusKernel(__read_only image2d_t inputImage, __read_only image2
 		float size = read_imagef(inputSize, SAMPLER_NEAREST, inputCoordinate).s0;
 		color_accum = read_imagef(inputImage, SAMPLER_NEAREST, inputCoordinate);
 
-		for (int ny = miny; ny < maxy; ny += step) {
-			for (int nx = minx; nx < maxx; nx += step) {
-				if (nx >= 0 && nx < dimension.s0 && ny >= 0 && ny < dimension.s1) {
-					inputCoordinate.s0 = nx - offsetInput.s0;
-					inputCoordinate.s1 = ny - offsetInput.s1;
-					tempSize = read_imagef(inputSize, SAMPLER_NEAREST, inputCoordinate).s0;
-					if (size > threshold && tempSize > threshold) {
-						float dx = nx - realCoordinate.s0;
-						float dy = ny - realCoordinate.s1;
-						if (dx != 0 || dy != 0) {
+		if (size > threshold) {
+			for (int ny = miny; ny < maxy; ny += step) {
+				inputCoordinate.s1 = ny - offsetInput.s1;
+				float dy = ny - realCoordinate.s1;
+				for (int nx = minx; nx < maxx; nx += step) {
+					float dx = nx - realCoordinate.s0;
+					if (dx != 0 || dy != 0) {
+						inputCoordinate.s0 = nx - offsetInput.s0;
+						tempSize = read_imagef(inputSize, SAMPLER_NEAREST, inputCoordinate).s0;
+						if (tempSize > threshold) {
 							if (tempSize >= fabs(dx) && tempSize >= fabs(dy)) {
 								float2 uv = { 256.0f + dx * 256.0f / tempSize, 256.0f + dy * 256.0f / tempSize};
 								bokeh = read_imagef(bokehImage, SAMPLER_NEAREST, uv);
@@ -121,8 +121,8 @@ __kernel void defocusKernel(__read_only image2d_t inputImage, __read_only image2
 						}
 					}
 				}
-			}
-		} 
+			} 
+		}
 	}
 
 	color = color_accum * (1.0f / multiplier_accum);
diff --git a/source/blender/compositor/operations/COM_OpenCLKernels.cl.h b/source/blender/compositor/operations/COM_OpenCLKernels.cl.h
index d57aa1366de..cc18039c5b1 100644
--- a/source/blender/compositor/operations/COM_OpenCLKernels.cl.h
+++ b/source/blender/compositor/operations/COM_OpenCLKernels.cl.h
@@ -103,16 +103,16 @@ const char * clkernelstoh_COM_OpenCLKernels_cl = "/*\n" \
 "		float size = read_imagef(inputSize, SAMPLER_NEAREST, inputCoordinate).s0;\n" \
 "		color_accum = read_imagef(inputImage, SAMPLER_NEAREST, inputCoordinate);\n" \
 "\n" \
-"		for (int ny = miny; ny < maxy; ny += step) {\n" \
-"			for (int nx = minx; nx < maxx; nx += step) {\n" \
-"				if (nx >= 0 && nx < dimension.s0 && ny >= 0 && ny < dimension.s1) {\n" \
-"					inputCoordinate.s0 = nx - offsetInput.s0;\n" \
-"					inputCoordinate.s1 = ny - offsetInput.s1;\n" \
-"					tempSize = read_imagef(inputSize, SAMPLER_NEAREST, inputCoordinate).s0;\n" \
-"					if (size > threshold && tempSize > threshold) {\n" \
-"						float dx = nx - realCoordinate.s0;\n" \
-"						float dy = ny - realCoordinate.s1;\n" \
-"						if (dx != 0 || dy != 0) {\n" \
+"		if (size > threshold) {\n" \
+"			for (int ny = miny; ny < maxy; ny += step) {\n" \
+"				inputCoordinate.s1 = ny - offsetInput.s1;\n" \
+"				float dy = ny - realCoordinate.s1;\n" \
+"				for (int nx = minx; nx < maxx; nx += step) {\n" \
+"					float dx = nx - realCoordinate.s0;\n" \
+"					if (dx != 0 || dy != 0) {\n" \
+"						inputCoordinate.s0 = nx - offsetInput.s0;\n" \
+"						tempSize = read_imagef(inputSize, SAMPLER_NEAREST, inputCoordinate).s0;\n" \
+"						if (tempSize > threshold) {\n" \
 "							if (tempSize >= fabs(dx) && tempSize >= fabs(dy)) {\n" \
 "								float2 uv = { 256.0f + dx * 256.0f / tempSize, 256.0f + dy * 256.0f / tempSize};\n" \
 "								bokeh = read_imagef(bokehImage, SAMPLER_NEAREST, uv);\n" \
diff --git a/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.cpp b/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.cpp
index 61538fde258..5d17526185b 100644
--- a/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.cpp
+++ b/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.cpp
@@ -62,8 +62,29 @@ void VariableSizeBokehBlurOperation::initExecution()
 	QualityStepHelper::initExecution(COM_QH_INCREASE);
 }
 
+void *VariableSizeBokehBlurOperation::initializeTileData(rcti *rect, MemoryBuffer **memoryBuffers)
+{
+	MemoryBuffer** result = new MemoryBuffer*[3];
+	result[0] = (MemoryBuffer*)this->m_inputProgram->initializeTileData(rect, memoryBuffers);
+	result[1] = (MemoryBuffer*)this->m_inputBokehProgram->initializeTileData(rect, memoryBuffers);
+	result[2] = (MemoryBuffer*)this->m_inputSizeProgram->initializeTileData(rect, memoryBuffers);
+	return result;
+}
+
+void VariableSizeBokehBlurOperation::deinitializeTileData(rcti *rect, MemoryBuffer **memoryBuffers, void *data)
+{
+	MemoryBuffer** result = (MemoryBuffer**)data;
+	delete[] result;
+}
+
 void VariableSizeBokehBlurOperation::executePixel(float *color, int x, int y, MemoryBuffer *inputBuffers[], void *data)
 {
+	MemoryBuffer** buffers = (MemoryBuffer**)data;
+	MemoryBuffer* inputProgramBuffer = buffers[0];
+	MemoryBuffer* inputBokehBuffer = buffers[1];
+	MemoryBuffer* inputSizeBuffer = buffers[2];
+	float* inputSizeFloatBuffer = inputSizeBuffer->getBuffer();
+	float* inputProgramFloatBuffer = inputProgramBuffer->getBuffer();
 	float readColor[4];
 	float bokeh[4];
 	float tempSize[4];
@@ -84,32 +105,37 @@ void VariableSizeBokehBlurOperation::executePixel(float *color, int x, int y, Me
 	int maxy = MIN2(y + this->m_maxBlur, m_height);
 #endif
 	{
-		this->m_inputSizeProgram->read(tempSize, x, y, COM_PS_NEAREST, inputBuffers);
-		this->m_inputProgram->read(readColor, x, y, COM_PS_NEAREST, inputBuffers);
+		inputSizeBuffer->readNoCheck(tempSize, x, y);
+		inputProgramBuffer->readNoCheck(readColor, x, y);
+
 		add_v4_v4(color_accum, readColor);
 		add_v4_fl(multiplier_accum, 1.0f);
 		float sizeCenter = tempSize[0];
 		
-		for (int ny = miny; ny < maxy; ny += QualityStepHelper::getStep()) {
-			for (int nx = minx; nx < maxx; nx += QualityStepHelper::getStep()) {
-				if (nx >= 0 && nx < this->getWidth() && ny >= 0 && ny < getHeight()) {
-					this->m_inputSizeProgram->read(tempSize, nx, ny, COM_PS_NEAREST, inputBuffers);
-					float size = tempSize[0];
-					float fsize = fabsf(size);
-					if (sizeCenter > this->m_threshold && size > this->m_threshold) {
-						float dx = nx - x;
-						float dy = ny - y;
-						if (nx == x && ny == y) {
-						}
-						else if (fsize > fabsf(dx) && fsize > fabsf(dy)) {
-							float u = (256 + (dx/size) * 256);
-							float v = (256 + (dy/size) * 256);
-							this->m_inputBokehProgram->read(bokeh, u, v, COM_PS_NEAREST, inputBuffers);
-							this->m_inputProgram->read(readColor, nx, ny, COM_PS_NEAREST, inputBuffers);
-							madd_v4_v4v4(color_accum, bokeh, readColor);
-							add_v4_v4(multiplier_accum, bokeh);
+		const int addXStep = QualityStepHelper::getStep()*COM_NUMBER_OF_CHANNELS;
+		
+		if (sizeCenter > this->m_threshold) {
+			for (int ny = miny; ny < maxy; ny += QualityStepHelper::getStep()) {
+				float dy = ny - y;
+				int offsetNy = ny * inputSizeBuffer->getWidth() * COM_NUMBER_OF_CHANNELS;
+				int offsetNxNy = offsetNy + (minx*COM_NUMBER_OF_CHANNELS);
+				for (int nx = minx; nx < maxx; nx += QualityStepHelper::getStep()) {
+					if (nx != x || ny != y) 
+					{
+						float size = inputSizeFloatBuffer[offsetNxNy];
+						if (size > this->m_threshold) {
+							float fsize = fabsf(size);
+							float dx = nx - x;
+							if (fsize > fabsf(dx) && fsize > fabsf(dy)) {
+								float u = (256.0f + (dx/size) * 256.0f);
+								float v = (256.0f + (dy/size) * 256.0f);
+								inputBokehBuffer->readNoCheck(bokeh, u, v);
+								madd_v4_v4v4(color_accum, bokeh, &inputProgramFloatBuffer[offsetNxNy]);
+								add_v4_v4(multiplier_accum, bokeh);
+							}
 						}
 					}
+					offsetNxNy += addXStep;
 				}
 			}
 		}
diff --git a/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.h b/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.h
index 6c9196c3eab..0ecfb5a542c 100644
--- a/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.h
+++ b/source/blender/compositor/operations/COM_VariableSizeBokehBlurOperation.h
@@ -50,6 +50,10 @@ public:
 	 */
 	void initExecution();
 	
+	void *initializeTileData(rcti *rect, MemoryBuffer **memoryBuffers);
+	
+	void deinitializeTileData(rcti *rect, MemoryBuffer **memoryBuffers, void *data);
+	
 	/**
 	 * Deinitialize the execution
 	 */