Skip to content

S
swiftshader
Commit c39901ec by Nicolas Capens
Options

Refactor buffer clearing.

Bug 27460431 Change-Id: I90de3285c86c0d3187969270dcbd78a8b02feee9 Reviewed-on: https://swiftshader-review.googlesource.com/4978Tested-by: 's avatarNicolas Capens <capn@google.com> Reviewed-by: 's avatarAlexis Hétu <sugoi@google.com> Reviewed-by: 's avatarNicolas Capens <capn@google.com>
parent dc4ae86d
Showing with 230 additions and 337 deletions
src/D3D9/Direct3DDevice9.cpp
...@@ -383,7 +383,6 @@ namespace D3D9 ...@@ -383,7 +383,6 @@ namespace D3D9
count = 1; count = 1;
D3DRECT rect; D3DRECT rect;
rect.x1 = viewport.X; rect.x1 = viewport.X;
rect.x2 = viewport.X + viewport.Width; rect.x2 = viewport.X + viewport.Width;
rect.y1 = viewport.Y; rect.y1 = viewport.Y;
...@@ -394,36 +393,13 @@ namespace D3D9 ...@@ -394,36 +393,13 @@ namespace D3D9
for(unsigned int i = 0; i < count; i++) for(unsigned int i = 0; i < count; i++)
{ {
D3DRECT rect = rects[i]; sw::SliceRect clearRect(rects[i].x1, rects[i].y1, rects[i].x2, rects[i].y2, 0);
// Clamp against viewport
if(rect.x1 < (int)viewport.X) rect.x1 = viewport.X;
if(rect.x2 < (int)viewport.X) rect.x2 = viewport.X;
if(rect.x1 > (int)viewport.X + (int)viewport.Width) rect.x1 = viewport.X + viewport.Width;
if(rect.x2 > (int)viewport.X + (int)viewport.Width) rect.x2 = viewport.X + viewport.Width;
if(rect.y1 < (int)viewport.Y) rect.y1 = viewport.Y; clearRect.clip(viewport.X, viewport.Y, viewport.X + viewport.Width, viewport.Y + viewport.Height);
if(rect.y2 < (int)viewport.Y) rect.y2 = viewport.Y;
if(rect.y1 > (int)viewport.Y + (int)viewport.Height) rect.y1 = viewport.Y + viewport.Height;
if(rect.y2 > (int)viewport.Y + (int)viewport.Height) rect.y2 = viewport.Y + viewport.Height;
// Clamp against scissor rectangle
if(scissorEnable) if(scissorEnable)
{ {
if(rect.x1 < (int)scissorRect.left) rect.x1 = scissorRect.left; clearRect.clip(scissorRect.left, scissorRect.top, scissorRect.right, scissorRect.bottom);
if(rect.x2 < (int)scissorRect.left) rect.x2 = scissorRect.left;
if(rect.x1 > (int)scissorRect.right) rect.x1 = scissorRect.right;
if(rect.x2 > (int)scissorRect.right) rect.x2 = scissorRect.right;
if(rect.y1 < (int)scissorRect.top) rect.y1 = scissorRect.top;
if(rect.y2 < (int)scissorRect.top) rect.y2 = scissorRect.top;
if(rect.y1 > (int)scissorRect.bottom) rect.y1 = scissorRect.bottom;
if(rect.y2 > (int)scissorRect.bottom) rect.y2 = scissorRect.bottom;
}
if(flags & D3DCLEAR_STENCIL)
{
depthStencil->clearStencilBuffer(stencil, 0xFF, rect.x1, rect.y1, rect.x2 - rect.x1, rect.y2 - rect.y1);
} }
if(flags & D3DCLEAR_TARGET) if(flags & D3DCLEAR_TARGET)
...@@ -448,21 +424,20 @@ namespace D3D9 ...@@ -448,21 +424,20 @@ namespace D3D9
rgba[2] = sw::linearToSRGB(rgba[2]); rgba[2] = sw::linearToSRGB(rgba[2]);
} }
sw::SliceRect sliceRect; renderer->clear(rgba, sw::FORMAT_A32B32G32R32F, renderTarget[index], clearRect, 0xF);
if(renderTarget[index]->getClearRect(rect.x1, rect.y1, rect.x2 - rect.x1, rect.y2 - rect.y1, sliceRect))
{
renderer->clear(rgba, sw::FORMAT_A32B32G32R32F, renderTarget[index], sliceRect, 0xF);
}
} }
} }
} }
if(flags & D3DCLEAR_ZBUFFER) if(flags & D3DCLEAR_ZBUFFER)
{ {
if(z > 1) z = 1; z = sw::clamp01(z);
if(z < 0) z = 0; depthStencil->clearDepth(z, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
}
depthStencil->clearDepthBuffer(z, rect.x1, rect.y1, rect.x2 - rect.x1, rect.y2 - rect.y1); if(flags & D3DCLEAR_STENCIL)
{
depthStencil->clearStencil(stencil, 0xFF, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
} }
} }
......
src/OpenGL/libGL/Device.cpp
...@@ -162,13 +162,13 @@ namespace gl ...@@ -162,13 +162,13 @@ namespace gl
} }
Device::~Device() Device::~Device()
{ {
if(depthStencil) if(depthStencil)
{ {
depthStencil->release(); depthStencil->release();
depthStencil = 0; depthStencil = 0;
} }
if(renderTarget) if(renderTarget)
{ {
renderTarget->release(); renderTarget->release();
...@@ -185,17 +185,11 @@ namespace gl ...@@ -185,17 +185,11 @@ namespace gl
return; return;
} }
int x0 = 0; sw::SliceRect clearRect = renderTarget->getRect();
int y0 = 0;
int width = renderTarget->getWidth();
int height = renderTarget->getHeight();
if(scissorEnable) // Clamp against scissor rectangle if(scissorEnable)
{ {
if(x0 < scissorRect.x0) x0 = scissorRect.x0; clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
if(y0 < scissorRect.y0) y0 = scissorRect.y0;
if(width > scissorRect.x1 - scissorRect.x0) width = scissorRect.x1 - scissorRect.x0;
if(height > scissorRect.y1 - scissorRect.y0) height = scissorRect.y1 - scissorRect.y0;
} }
float rgba[4]; float rgba[4];
...@@ -204,11 +198,7 @@ namespace gl ...@@ -204,11 +198,7 @@ namespace gl
rgba[2] = blue; rgba[2] = blue;
rgba[3] = alpha; rgba[3] = alpha;
sw::SliceRect sliceRect; clear(rgba, FORMAT_A32B32G32R32F, renderTarget, clearRect, rgbaMask);
if(renderTarget->getClearRect(x0, y0, width, height, sliceRect))
{
clear(rgba, FORMAT_A32B32G32R32F, renderTarget, sliceRect, rgbaMask);
}
} }
void Device::clearDepth(float z) void Device::clearDepth(float z)
...@@ -218,23 +208,15 @@ namespace gl ...@@ -218,23 +208,15 @@ namespace gl
return; return;
} }
if(z > 1) z = 1; z = clamp01(z);
if(z < 0) z = 0; sw::SliceRect clearRect = depthStencil->getRect();
int x0 = 0; if(scissorEnable)
int y0 = 0;
int width = depthStencil->getWidth();
int height = depthStencil->getHeight();
if(scissorEnable) // Clamp against scissor rectangle
{ {
if(x0 < scissorRect.x0) x0 = scissorRect.x0; clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
if(y0 < scissorRect.y0) y0 = scissorRect.y0;
if(width > scissorRect.x1 - scissorRect.x0) width = scissorRect.x1 - scissorRect.x0;
if(height > scissorRect.y1 - scissorRect.y0) height = scissorRect.y1 - scissorRect.y0;
} }
depthStencil->clearDepthBuffer(z, x0, y0, width, height); depthStencil->clearDepth(z, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
} }
void Device::clearStencil(unsigned int stencil, unsigned int mask) void Device::clearStencil(unsigned int stencil, unsigned int mask)
...@@ -244,20 +226,14 @@ namespace gl ...@@ -244,20 +226,14 @@ namespace gl
return; return;
} }
int x0 = 0; sw::SliceRect clearRect = depthStencil->getRect();
int y0 = 0;
int width = depthStencil->getWidth();
int height = depthStencil->getHeight();
if(scissorEnable) // Clamp against scissor rectangle if(scissorEnable)
{ {
if(x0 < scissorRect.x0) x0 = scissorRect.x0; clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
if(y0 < scissorRect.y0) y0 = scissorRect.y0;
if(width > scissorRect.x1 - scissorRect.x0) width = scissorRect.x1 - scissorRect.x0;
if(height > scissorRect.y1 - scissorRect.y0) height = scissorRect.y1 - scissorRect.y0;
} }
depthStencil->clearStencilBuffer(stencil, mask, x0, y0, width, height); depthStencil->clearStencil(stencil, mask, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
} }
Image *Device::createDepthStencilSurface(unsigned int width, unsigned int height, sw::Format format, int multiSampleDepth, bool discard) Image *Device::createDepthStencilSurface(unsigned int width, unsigned int height, sw::Format format, int multiSampleDepth, bool discard)
...@@ -267,7 +243,7 @@ namespace gl ...@@ -267,7 +243,7 @@ namespace gl
ERR("Invalid parameters: %dx%d", width, height); ERR("Invalid parameters: %dx%d", width, height);
return 0; return 0;
} }
bool lockable = true; bool lockable = true;
switch(format) switch(format)
...@@ -319,7 +295,7 @@ namespace gl ...@@ -319,7 +295,7 @@ namespace gl
ERR("Out of memory"); ERR("Out of memory");
return 0; return 0;
} }
return surface; return surface;
} }
...@@ -387,7 +363,7 @@ namespace gl ...@@ -387,7 +363,7 @@ namespace gl
} }
setIndexBuffer(0); setIndexBuffer(0);
DrawType drawType; DrawType drawType;
switch(primitiveType) switch(primitiveType)
...@@ -491,7 +467,7 @@ namespace gl ...@@ -491,7 +467,7 @@ namespace gl
vertexShaderConstantF[startRegister + i][2] = constantData[i * 4 + 2]; vertexShaderConstantF[startRegister + i][2] = constantData[i * 4 + 2];
vertexShaderConstantF[startRegister + i][3] = constantData[i * 4 + 3]; vertexShaderConstantF[startRegister + i][3] = constantData[i * 4 + 3];
} }
vertexShaderConstantsFDirty = max(startRegister + count, vertexShaderConstantsFDirty); vertexShaderConstantsFDirty = max(startRegister + count, vertexShaderConstantsFDirty);
vertexShaderDirty = true; // Reload DEF constants vertexShaderDirty = true; // Reload DEF constants
} }
...@@ -508,7 +484,7 @@ namespace gl ...@@ -508,7 +484,7 @@ namespace gl
ERR("Invalid parameters"); ERR("Invalid parameters");
return false; return false;
} }
int sWidth = source->getWidth(); int sWidth = source->getWidth();
int sHeight = source->getHeight(); int sHeight = source->getHeight();
int dWidth = dest->getWidth(); int dWidth = dest->getWidth();
...@@ -619,7 +595,7 @@ namespace gl ...@@ -619,7 +595,7 @@ namespace gl
destBytes[4 * x + 3] = 0xFF; destBytes[4 * x + 3] = 0xFF;
} }
} }
sourceBytes += sourcePitch; sourceBytes += sourcePitch;
destBytes += destPitch; destBytes += destPitch;
} }
...@@ -677,7 +653,7 @@ namespace gl ...@@ -677,7 +653,7 @@ namespace gl
{ {
Renderer::setVertexShaderConstantF(0, vertexShaderConstantF[0], vertexShaderConstantsFDirty); Renderer::setVertexShaderConstantF(0, vertexShaderConstantF[0], vertexShaderConstantsFDirty);
} }
Renderer::setVertexShader(vertexShader); // Loads shader constants set with DEF Renderer::setVertexShader(vertexShader); // Loads shader constants set with DEF
vertexShaderConstantsFDirty = vertexShader->dirtyConstantsF; // Shader DEF'ed constants are dirty vertexShaderConstantsFDirty = vertexShader->dirtyConstantsF; // Shader DEF'ed constants are dirty
} }
...@@ -689,7 +665,7 @@ namespace gl ...@@ -689,7 +665,7 @@ namespace gl
vertexShaderDirty = false; vertexShaderDirty = false;
} }
} }
bool Device::bindViewport() bool Device::bindViewport()
{ {
if(viewport.width <= 0 || viewport.height <= 0) if(viewport.width <= 0 || viewport.height <= 0)
...@@ -709,7 +685,7 @@ namespace gl ...@@ -709,7 +685,7 @@ namespace gl
scissor.x1 = scissorRect.x1; scissor.x1 = scissorRect.x1;
scissor.y0 = scissorRect.y0; scissor.y0 = scissorRect.y0;
scissor.y1 = scissorRect.y1; scissor.y1 = scissorRect.y1;
setScissor(scissor); setScissor(scissor);
} }
else else
...@@ -719,7 +695,7 @@ namespace gl ...@@ -719,7 +695,7 @@ namespace gl
scissor.x1 = viewport.x0 + viewport.width; scissor.x1 = viewport.x0 + viewport.width;
scissor.y0 = viewport.y0; scissor.y0 = viewport.y0;
scissor.y1 = viewport.y0 + viewport.height; scissor.y1 = viewport.y0 + viewport.height;
if(renderTarget) if(renderTarget)
{ {
scissor.x0 = max(scissor.x0, 0); scissor.x0 = max(scissor.x0, 0);
...@@ -746,7 +722,7 @@ namespace gl ...@@ -746,7 +722,7 @@ namespace gl
view.height = (float)viewport.height; view.height = (float)viewport.height;
view.minZ = viewport.minZ; view.minZ = viewport.minZ;
view.maxZ = viewport.maxZ; view.maxZ = viewport.maxZ;
Renderer::setViewport(view); Renderer::setViewport(view);
return true; return true;
......
src/OpenGL/libGLES_CM/Device.cpp
...@@ -150,30 +150,20 @@ namespace es1 ...@@ -150,30 +150,20 @@ namespace es1
return; return;
} }
int x0 = 0;
int y0 = 0;
int width = renderTarget->getWidth();
int height = renderTarget->getHeight();
if(scissorEnable) // Clamp against scissor rectangle
{
if(x0 < scissorRect.x0) x0 = scissorRect.x0;
if(y0 < scissorRect.y0) y0 = scissorRect.y0;
if(width > scissorRect.x1 - scissorRect.x0) width = scissorRect.x1 - scissorRect.x0;
if(height > scissorRect.y1 - scissorRect.y0) height = scissorRect.y1 - scissorRect.y0;
}
float rgba[4]; float rgba[4];
rgba[0] = red; rgba[0] = red;
rgba[1] = green; rgba[1] = green;
rgba[2] = blue; rgba[2] = blue;
rgba[3] = alpha; rgba[3] = alpha;
sw::SliceRect sliceRect; sw::SliceRect clearRect = renderTarget->getRect();
if(renderTarget->getClearRect(x0, y0, width, height, sliceRect))
if(scissorEnable)
{ {
clear(rgba, FORMAT_A32B32G32R32F, renderTarget, sliceRect, rgbaMask); clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
} }
clear(rgba, FORMAT_A32B32G32R32F, renderTarget, clearRect, rgbaMask);
} }
void Device::clearDepth(float z) void Device::clearDepth(float z)
...@@ -183,23 +173,15 @@ namespace es1 ...@@ -183,23 +173,15 @@ namespace es1
return; return;
} }
if(z > 1) z = 1; z = clamp01(z);
if(z < 0) z = 0; sw::SliceRect clearRect = depthBuffer->getRect();
int x0 = 0;
int y0 = 0;
int width = depthBuffer->getWidth();
int height = depthBuffer->getHeight();
if(scissorEnable) // Clamp against scissor rectangle if(scissorEnable)
{ {
if(x0 < scissorRect.x0) x0 = scissorRect.x0; clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
if(y0 < scissorRect.y0) y0 = scissorRect.y0;
if(width > scissorRect.x1 - scissorRect.x0) width = scissorRect.x1 - scissorRect.x0;
if(height > scissorRect.y1 - scissorRect.y0) height = scissorRect.y1 - scissorRect.y0;
} }
depthBuffer->clearDepthBuffer(z, x0, y0, width, height); depthBuffer->clearDepth(z, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
} }
void Device::clearStencil(unsigned int stencil, unsigned int mask) void Device::clearStencil(unsigned int stencil, unsigned int mask)
...@@ -209,20 +191,14 @@ namespace es1 ...@@ -209,20 +191,14 @@ namespace es1
return; return;
} }
int x0 = 0; sw::SliceRect clearRect = stencilBuffer->getRect();
int y0 = 0;
int width = stencilBuffer->getWidth();
int height = stencilBuffer->getHeight();
if(scissorEnable) // Clamp against scissor rectangle if(scissorEnable)
{ {
if(x0 < scissorRect.x0) x0 = scissorRect.x0; clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
if(y0 < scissorRect.y0) y0 = scissorRect.y0;
if(width > scissorRect.x1 - scissorRect.x0) width = scissorRect.x1 - scissorRect.x0;
if(height > scissorRect.y1 - scissorRect.y0) height = scissorRect.y1 - scissorRect.y0;
} }
stencilBuffer->clearStencilBuffer(stencil, mask, x0, y0, width, height); stencilBuffer->clearStencil(stencil, mask, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
} }
egl::Image *Device::createDepthStencilSurface(unsigned int width, unsigned int height, sw::Format format, int multiSampleDepth, bool discard) egl::Image *Device::createDepthStencilSurface(unsigned int width, unsigned int height, sw::Format format, int multiSampleDepth, bool discard)
......
src/OpenGL/libGLESv2/Context.cpp
...@@ -2754,18 +2754,6 @@ void Context::applyScissor(int width, int height) ...@@ -2754,18 +2754,6 @@ void Context::applyScissor(int width, int height)
} }
} }
egl::Image *Context::getScissoredImage(GLint drawbuffer, int &x0, int &y0, int &width, int &height, bool depthStencil)
{
Framebuffer* framebuffer = getDrawFramebuffer();
egl::Image* image = depthStencil ? framebuffer->getDepthStencil() : framebuffer->getRenderTarget(drawbuffer);
applyScissor(image->getWidth(), image->getHeight());
device->getScissoredRegion(image, x0, y0, width, height);
return image;
}
// Applies the render target surface, depth stencil surface, viewport rectangle and scissor rectangle // Applies the render target surface, depth stencil surface, viewport rectangle and scissor rectangle
bool Context::applyRenderTarget() bool Context::applyRenderTarget()
{ {
...@@ -3407,18 +3395,24 @@ void Context::clear(GLbitfield mask) ...@@ -3407,18 +3395,24 @@ void Context::clear(GLbitfield mask)
void Context::clearColorBuffer(GLint drawbuffer, void *value, sw::Format format) void Context::clearColorBuffer(GLint drawbuffer, void *value, sw::Format format)
{ {
unsigned int rgbaMask = getColorMask(); unsigned int rgbaMask = getColorMask();
if(device && rgbaMask && !mState.rasterizerDiscardEnabled) if(rgbaMask && !mState.rasterizerDiscardEnabled)
{ {
int x0(0), y0(0), width(0), height(0); Framebuffer *framebuffer = getDrawFramebuffer();
egl::Image* image = getScissoredImage(drawbuffer, x0, y0, width, height, false); egl::Image *colorbuffer = framebuffer->getRenderTarget(drawbuffer);
sw::SliceRect sliceRect; if(colorbuffer)
if(image->getClearRect(x0, y0, width, height, sliceRect))
{ {
device->clear(value, format, image, sliceRect, rgbaMask); sw::SliceRect clearRect = colorbuffer->getRect();
}
image->release(); if(mState.scissorTestEnabled)
{
clearRect.clip(mState.scissorX, mState.scissorY, mState.scissorX + mState.scissorWidth, mState.scissorY + mState.scissorHeight);
}
device->clear(value, format, colorbuffer, clearRect, rgbaMask);
colorbuffer->release();
}
} }
} }
...@@ -3437,52 +3431,51 @@ void Context::clearColorBuffer(GLint drawbuffer, const GLfloat *value) ...@@ -3437,52 +3431,51 @@ void Context::clearColorBuffer(GLint drawbuffer, const GLfloat *value)
clearColorBuffer(drawbuffer, (void*)value, sw::FORMAT_A32B32G32R32F); clearColorBuffer(drawbuffer, (void*)value, sw::FORMAT_A32B32G32R32F);
} }
void Context::clearDepthBuffer(GLint drawbuffer, const GLfloat *value) void Context::clearDepthBuffer(const GLfloat value)
{ {
if(device && mState.depthMask && !mState.rasterizerDiscardEnabled) if(mState.depthMask && !mState.rasterizerDiscardEnabled)
{ {
int x0(0), y0(0), width(0), height(0); Framebuffer *framebuffer = getDrawFramebuffer();
egl::Image* image = getScissoredImage(drawbuffer, x0, y0, width, height, true); egl::Image *depthbuffer = framebuffer->getDepthStencil();
float depth = clamp01(value[0]);
image->clearDepthBuffer(depth, x0, y0, width, height);
image->release(); if(depthbuffer)
} {
} float depth = clamp01(value);
sw::SliceRect clearRect = depthbuffer->getRect();
void Context::clearStencilBuffer(GLint drawbuffer, const GLint *value) if(mState.scissorTestEnabled)
{ {
if(device && mState.stencilWritemask && !mState.rasterizerDiscardEnabled) clearRect.clip(mState.scissorX, mState.scissorY, mState.scissorX + mState.scissorWidth, mState.scissorY + mState.scissorHeight);
{ }
int x0(0), y0(0), width(0), height(0);
egl::Image* image = getScissoredImage(drawbuffer, x0, y0, width, height, true);
unsigned char stencil = value[0] < 0 ? 0 : static_cast<unsigned char>(value[0] & 0x000000FF); depthbuffer->clearDepth(depth, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
image->clearStencilBuffer(stencil, static_cast<unsigned char>(mState.stencilWritemask), x0, y0, width, height);
image->release(); depthbuffer->release();
}
} }
} }
void Context::clearDepthStencilBuffer(GLint drawbuffer, GLfloat depth, GLint stencil) void Context::clearStencilBuffer(const GLint value)
{ {
if(device && (mState.depthMask || mState.stencilWritemask) && !mState.rasterizerDiscardEnabled) if(mState.stencilWritemask && !mState.rasterizerDiscardEnabled)
{ {
int x0(0), y0(0), width(0), height(0); Framebuffer *framebuffer = getDrawFramebuffer();
egl::Image* image = getScissoredImage(drawbuffer, x0, y0, width, height, true); egl::Image *stencilbuffer = framebuffer->getDepthStencil();
if(mState.stencilWritemask) if(stencilbuffer)
{ {
image->clearStencilBuffer(static_cast<unsigned char>(stencil & 0x000000FF), static_cast<unsigned char>(mState.stencilWritemask), x0, y0, width, height); unsigned char stencil = value < 0 ? 0 : static_cast<unsigned char>(value & 0x000000FF);
} sw::SliceRect clearRect = stencilbuffer->getRect();
if(mState.depthMask) if(mState.scissorTestEnabled)
{ {
image->clearDepthBuffer(clamp01(depth), x0, y0, width, height); clearRect.clip(mState.scissorX, mState.scissorY, mState.scissorX + mState.scissorWidth, mState.scissorY + mState.scissorHeight);
} }
stencilbuffer->clearStencil(stencil, static_cast<unsigned char>(mState.stencilWritemask), clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
image->release(); stencilbuffer->release();
}
} }
} }
......
src/OpenGL/libGLESv2/Context.h
...@@ -644,9 +644,8 @@ public: ...@@ -644,9 +644,8 @@ public:
void clearColorBuffer(GLint drawbuffer, const GLint *value); void clearColorBuffer(GLint drawbuffer, const GLint *value);
void clearColorBuffer(GLint drawbuffer, const GLuint *value); void clearColorBuffer(GLint drawbuffer, const GLuint *value);
void clearColorBuffer(GLint drawbuffer, const GLfloat *value); void clearColorBuffer(GLint drawbuffer, const GLfloat *value);
void clearDepthBuffer(GLint drawbuffer, const GLfloat *value); void clearDepthBuffer(const GLfloat value);
void clearStencilBuffer(GLint drawbuffer, const GLint *value); void clearStencilBuffer(const GLint value);
void clearDepthStencilBuffer(GLint drawbuffer, GLfloat depth, GLint stencil);
void drawArrays(GLenum mode, GLint first, GLsizei count, GLsizei instanceCount = 1); void drawArrays(GLenum mode, GLint first, GLsizei count, GLsizei instanceCount = 1);
void drawElements(GLenum mode, GLuint start, GLuint end, GLsizei count, GLenum type, const void *indices, GLsizei instanceCount = 1); void drawElements(GLenum mode, GLuint start, GLuint end, GLsizei count, GLenum type, const void *indices, GLsizei instanceCount = 1);
void finish(); void finish();
...@@ -677,7 +676,6 @@ public: ...@@ -677,7 +676,6 @@ public:
private: private:
virtual ~Context(); virtual ~Context();
egl::Image *getScissoredImage(GLint drawbuffer, int &x0, int &y0, int &width, int &height, bool depthStencil);
void applyScissor(int width, int height); void applyScissor(int width, int height);
bool applyRenderTarget(); bool applyRenderTarget();
void applyState(GLenum drawMode); void applyState(GLenum drawMode);
......
src/OpenGL/libGLESv2/Device.cpp
...@@ -173,22 +173,6 @@ namespace es2 ...@@ -173,22 +173,6 @@ namespace es2
delete context; delete context;
} }
void Device::getScissoredRegion(egl::Image *sourceSurface, int &x0, int &y0, int& width, int& height) const
{
x0 = 0;
y0 = 0;
width = sourceSurface->getWidth();
height = sourceSurface->getHeight();
if(scissorEnable) // Clamp against scissor rectangle
{
if(x0 < scissorRect.x0) x0 = scissorRect.x0;
if(y0 < scissorRect.y0) y0 = scissorRect.y0;
if(width > scissorRect.x1 - scissorRect.x0) width = scissorRect.x1 - scissorRect.x0;
if(height > scissorRect.y1 - scissorRect.y0) height = scissorRect.y1 - scissorRect.y0;
}
}
void Device::clearColor(float red, float green, float blue, float alpha, unsigned int rgbaMask) void Device::clearColor(float red, float green, float blue, float alpha, unsigned int rgbaMask)
{ {
if(!rgbaMask) if(!rgbaMask)
...@@ -206,17 +190,17 @@ namespace es2 ...@@ -206,17 +190,17 @@ namespace es2
{ {
if(renderTarget[i]) if(renderTarget[i])
{ {
int x0(0), y0(0), width(0), height(0); sw::SliceRect clearRect = renderTarget[i]->getRect();
getScissoredRegion(renderTarget[i], x0, y0, width, height);
sw::SliceRect sliceRect; if(scissorEnable)
if(renderTarget[i]->getClearRect(x0, y0, width, height, sliceRect))
{ {
int depth = sw::max(renderTarget[i]->getDepth(), 1); clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
for(sliceRect.slice = 0; sliceRect.slice < depth; ++sliceRect.slice) }
{
clear(rgba, FORMAT_A32B32G32R32F, renderTarget[i], sliceRect, rgbaMask); int depth = sw::max(renderTarget[i]->getDepth(), 1);
} for(clearRect.slice = 0; clearRect.slice < depth; clearRect.slice++)
{
clear(rgba, FORMAT_A32B32G32R32F, renderTarget[i], clearRect, rgbaMask);
} }
} }
} }
...@@ -229,13 +213,15 @@ namespace es2 ...@@ -229,13 +213,15 @@ namespace es2
return; return;
} }
if(z > 1) z = 1; z = clamp01(z);
if(z < 0) z = 0; sw::SliceRect clearRect = depthBuffer->getRect();
int x0(0), y0(0), width(0), height(0); if(scissorEnable)
getScissoredRegion(depthBuffer, x0, y0, width, height); {
clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
}
depthBuffer->clearDepthBuffer(z, x0, y0, width, height); depthBuffer->clearDepth(z, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
} }
void Device::clearStencil(unsigned int stencil, unsigned int mask) void Device::clearStencil(unsigned int stencil, unsigned int mask)
...@@ -245,10 +231,14 @@ namespace es2 ...@@ -245,10 +231,14 @@ namespace es2
return; return;
} }
int x0(0), y0(0), width(0), height(0); sw::SliceRect clearRect = stencilBuffer->getRect();
getScissoredRegion(stencilBuffer, x0, y0, width, height);
if(scissorEnable)
{
clearRect.clip(scissorRect.x0, scissorRect.y0, scissorRect.x1, scissorRect.y1);
}
stencilBuffer->clearStencilBuffer(stencil, mask, x0, y0, width, height); stencilBuffer->clearStencil(stencil, mask, clearRect.x0, clearRect.y0, clearRect.width(), clearRect.height());
} }
egl::Image *Device::createDepthStencilSurface(unsigned int width, unsigned int height, sw::Format format, int multiSampleDepth, bool discard) egl::Image *Device::createDepthStencilSurface(unsigned int width, unsigned int height, sw::Format format, int multiSampleDepth, bool discard)
......
src/OpenGL/libGLESv2/Device.hpp
...@@ -62,8 +62,6 @@ namespace es2 ...@@ -62,8 +62,6 @@ namespace es2
virtual bool stretchCube(sw::Surface *sourceSurface, sw::Surface *destSurface); virtual bool stretchCube(sw::Surface *sourceSurface, sw::Surface *destSurface);
virtual void finish(); virtual void finish();
void getScissoredRegion(egl::Image *sourceSurface, int &x0, int &y0, int& width, int& height) const;
private: private:
sw::Context *const context; sw::Context *const context;
......
src/OpenGL/libGLESv2/libGLESv3.cpp
...@@ -2586,7 +2586,7 @@ GL_APICALL void GL_APIENTRY glClearBufferiv(GLenum buffer, GLint drawbuffer, con ...@@ -2586,7 +2586,7 @@ GL_APICALL void GL_APIENTRY glClearBufferiv(GLenum buffer, GLint drawbuffer, con
} }
else else
{ {
context->clearStencilBuffer(drawbuffer, value); context->clearStencilBuffer(value[0]);
} }
break; break;
default: default:
...@@ -2650,7 +2650,7 @@ GL_APICALL void GL_APIENTRY glClearBufferfv(GLenum buffer, GLint drawbuffer, con ...@@ -2650,7 +2650,7 @@ GL_APICALL void GL_APIENTRY glClearBufferfv(GLenum buffer, GLint drawbuffer, con
} }
else else
{ {
context->clearDepthBuffer(drawbuffer, value); context->clearDepthBuffer(value[0]);
} }
break; break;
default: default:
...@@ -2677,7 +2677,8 @@ GL_APICALL void GL_APIENTRY glClearBufferfi(GLenum buffer, GLint drawbuffer, GLf ...@@ -2677,7 +2677,8 @@ GL_APICALL void GL_APIENTRY glClearBufferfi(GLenum buffer, GLint drawbuffer, GLf
} }
else else
{ {
context->clearDepthStencilBuffer(drawbuffer, depth, stencil); context->clearDepthBuffer(depth);
context->clearStencilBuffer(stencil);
} }
break; break;
default: default:
......
src/Renderer/Surface.cpp
...@@ -387,7 +387,7 @@ namespace sw ...@@ -387,7 +387,7 @@ namespace sw
ASSERT(palette); ASSERT(palette);
unsigned int abgr = palette[*(unsigned char*)element]; unsigned int abgr = palette[*(unsigned char*)element];
r = (abgr & 0x000000FF) * (1.0f / 0x000000FF); r = (abgr & 0x000000FF) * (1.0f / 0x000000FF);
g = (abgr & 0x0000FF00) * (1.0f / 0x0000FF00); g = (abgr & 0x0000FF00) * (1.0f / 0x0000FF00);
b = (abgr & 0x00FF0000) * (1.0f / 0x00FF0000); b = (abgr & 0x00FF0000) * (1.0f / 0x00FF0000);
...@@ -399,7 +399,7 @@ namespace sw ...@@ -399,7 +399,7 @@ namespace sw
ASSERT(palette); ASSERT(palette);
unsigned int bgr = palette[((unsigned char*)element)[0]]; unsigned int bgr = palette[((unsigned char*)element)[0]];
r = (bgr & 0x000000FF) * (1.0f / 0x000000FF); r = (bgr & 0x000000FF) * (1.0f / 0x000000FF);
g = (bgr & 0x0000FF00) * (1.0f / 0x0000FF00); g = (bgr & 0x0000FF00) * (1.0f / 0x0000FF00);
b = (bgr & 0x00FF0000) * (1.0f / 0x00FF0000); b = (bgr & 0x00FF0000) * (1.0f / 0x00FF0000);
...@@ -427,7 +427,7 @@ namespace sw ...@@ -427,7 +427,7 @@ namespace sw
case FORMAT_R3G3B2: case FORMAT_R3G3B2:
{ {
unsigned char rgb = *(unsigned char*)element; unsigned char rgb = *(unsigned char*)element;
r = (rgb & 0xE0) * (1.0f / 0xE0); r = (rgb & 0xE0) * (1.0f / 0xE0);
g = (rgb & 0x1C) * (1.0f / 0x1C); g = (rgb & 0x1C) * (1.0f / 0x1C);
b = (rgb & 0x03) * (1.0f / 0x03); b = (rgb & 0x03) * (1.0f / 0x03);
...@@ -436,7 +436,7 @@ namespace sw ...@@ -436,7 +436,7 @@ namespace sw
case FORMAT_A8R3G3B2: case FORMAT_A8R3G3B2:
{ {
unsigned short argb = *(unsigned short*)element; unsigned short argb = *(unsigned short*)element;
a = (argb & 0xFF00) * (1.0f / 0xFF00); a = (argb & 0xFF00) * (1.0f / 0xFF00);
r = (argb & 0x00E0) * (1.0f / 0x00E0); r = (argb & 0x00E0) * (1.0f / 0x00E0);
g = (argb & 0x001C) * (1.0f / 0x001C); g = (argb & 0x001C) * (1.0f / 0x001C);
...@@ -446,7 +446,7 @@ namespace sw ...@@ -446,7 +446,7 @@ namespace sw
case FORMAT_X4R4G4B4: case FORMAT_X4R4G4B4:
{ {
unsigned short rgb = *(unsigned short*)element; unsigned short rgb = *(unsigned short*)element;
r = (rgb & 0x0F00) * (1.0f / 0x0F00); r = (rgb & 0x0F00) * (1.0f / 0x0F00);
g = (rgb & 0x00F0) * (1.0f / 0x00F0); g = (rgb & 0x00F0) * (1.0f / 0x00F0);
b = (rgb & 0x000F) * (1.0f / 0x000F); b = (rgb & 0x000F) * (1.0f / 0x000F);
...@@ -455,7 +455,7 @@ namespace sw ...@@ -455,7 +455,7 @@ namespace sw
case FORMAT_A4R4G4B4: case FORMAT_A4R4G4B4:
{ {
unsigned short argb = *(unsigned short*)element; unsigned short argb = *(unsigned short*)element;
a = (argb & 0xF000) * (1.0f / 0xF000); a = (argb & 0xF000) * (1.0f / 0xF000);
r = (argb & 0x0F00) * (1.0f / 0x0F00); r = (argb & 0x0F00) * (1.0f / 0x0F00);
g = (argb & 0x00F0) * (1.0f / 0x00F0); g = (argb & 0x00F0) * (1.0f / 0x00F0);
...@@ -465,7 +465,7 @@ namespace sw ...@@ -465,7 +465,7 @@ namespace sw
case FORMAT_R4G4B4A4: case FORMAT_R4G4B4A4:
{ {
unsigned short rgba = *(unsigned short*)element; unsigned short rgba = *(unsigned short*)element;
r = (rgba & 0xF000) * (1.0f / 0xF000); r = (rgba & 0xF000) * (1.0f / 0xF000);
g = (rgba & 0x0F00) * (1.0f / 0x0F00); g = (rgba & 0x0F00) * (1.0f / 0x0F00);
b = (rgba & 0x00F0) * (1.0f / 0x00F0); b = (rgba & 0x00F0) * (1.0f / 0x00F0);
...@@ -475,7 +475,7 @@ namespace sw ...@@ -475,7 +475,7 @@ namespace sw
case FORMAT_R5G6B5: case FORMAT_R5G6B5:
{ {
unsigned short rgb = *(unsigned short*)element; unsigned short rgb = *(unsigned short*)element;
r = (rgb & 0xF800) * (1.0f / 0xF800); r = (rgb & 0xF800) * (1.0f / 0xF800);
g = (rgb & 0x07E0) * (1.0f / 0x07E0); g = (rgb & 0x07E0) * (1.0f / 0x07E0);
b = (rgb & 0x001F) * (1.0f / 0x001F); b = (rgb & 0x001F) * (1.0f / 0x001F);
...@@ -484,7 +484,7 @@ namespace sw ...@@ -484,7 +484,7 @@ namespace sw
case FORMAT_A1R5G5B5: case FORMAT_A1R5G5B5:
{ {
unsigned short argb = *(unsigned short*)element; unsigned short argb = *(unsigned short*)element;
a = (argb & 0x8000) * (1.0f / 0x8000); a = (argb & 0x8000) * (1.0f / 0x8000);
r = (argb & 0x7C00) * (1.0f / 0x7C00); r = (argb & 0x7C00) * (1.0f / 0x7C00);
g = (argb & 0x03E0) * (1.0f / 0x03E0); g = (argb & 0x03E0) * (1.0f / 0x03E0);
...@@ -494,7 +494,7 @@ namespace sw ...@@ -494,7 +494,7 @@ namespace sw
case FORMAT_R5G5B5A1: case FORMAT_R5G5B5A1:
{ {
unsigned short rgba = *(unsigned short*)element; unsigned short rgba = *(unsigned short*)element;
r = (rgba & 0xF800) * (1.0f / 0xF800); r = (rgba & 0xF800) * (1.0f / 0xF800);
g = (rgba & 0x07C0) * (1.0f / 0x07C0); g = (rgba & 0x07C0) * (1.0f / 0x07C0);
b = (rgba & 0x003E) * (1.0f / 0x003E); b = (rgba & 0x003E) * (1.0f / 0x003E);
...@@ -504,7 +504,7 @@ namespace sw ...@@ -504,7 +504,7 @@ namespace sw
case FORMAT_X1R5G5B5: case FORMAT_X1R5G5B5:
{ {
unsigned short xrgb = *(unsigned short*)element; unsigned short xrgb = *(unsigned short*)element;
r = (xrgb & 0x7C00) * (1.0f / 0x7C00); r = (xrgb & 0x7C00) * (1.0f / 0x7C00);
g = (xrgb & 0x03E0) * (1.0f / 0x03E0); g = (xrgb & 0x03E0) * (1.0f / 0x03E0);
b = (xrgb & 0x001F) * (1.0f / 0x001F); b = (xrgb & 0x001F) * (1.0f / 0x001F);
...@@ -513,7 +513,7 @@ namespace sw ...@@ -513,7 +513,7 @@ namespace sw
case FORMAT_A8R8G8B8: case FORMAT_A8R8G8B8:
{ {
unsigned int argb = *(unsigned int*)element; unsigned int argb = *(unsigned int*)element;
a = (argb & 0xFF000000) * (1.0f / 0xFF000000); a = (argb & 0xFF000000) * (1.0f / 0xFF000000);
r = (argb & 0x00FF0000) * (1.0f / 0x00FF0000); r = (argb & 0x00FF0000) * (1.0f / 0x00FF0000);
g = (argb & 0x0000FF00) * (1.0f / 0x0000FF00); g = (argb & 0x0000FF00) * (1.0f / 0x0000FF00);
...@@ -523,7 +523,7 @@ namespace sw ...@@ -523,7 +523,7 @@ namespace sw
case FORMAT_X8R8G8B8: case FORMAT_X8R8G8B8:
{ {
unsigned int xrgb = *(unsigned int*)element; unsigned int xrgb = *(unsigned int*)element;
r = (xrgb & 0x00FF0000) * (1.0f / 0x00FF0000); r = (xrgb & 0x00FF0000) * (1.0f / 0x00FF0000);
g = (xrgb & 0x0000FF00) * (1.0f / 0x0000FF00); g = (xrgb & 0x0000FF00) * (1.0f / 0x0000FF00);
b = (xrgb & 0x000000FF) * (1.0f / 0x000000FF); b = (xrgb & 0x000000FF) * (1.0f / 0x000000FF);
...@@ -532,7 +532,7 @@ namespace sw ...@@ -532,7 +532,7 @@ namespace sw
case FORMAT_A8B8G8R8I_SNORM: case FORMAT_A8B8G8R8I_SNORM:
{ {
signed char* abgr = (signed char*)element; signed char* abgr = (signed char*)element;
r = max(abgr[0] * (1.0f / 0x7F), -1.0f); r = max(abgr[0] * (1.0f / 0x7F), -1.0f);
g = max(abgr[1] * (1.0f / 0x7F), -1.0f); g = max(abgr[1] * (1.0f / 0x7F), -1.0f);
b = max(abgr[2] * (1.0f / 0x7F), -1.0f); b = max(abgr[2] * (1.0f / 0x7F), -1.0f);
...@@ -542,7 +542,7 @@ namespace sw ...@@ -542,7 +542,7 @@ namespace sw
case FORMAT_A8B8G8R8: case FORMAT_A8B8G8R8:
{ {
unsigned int abgr = *(unsigned int*)element; unsigned int abgr = *(unsigned int*)element;
a = (abgr & 0xFF000000) * (1.0f / 0xFF000000); a = (abgr & 0xFF000000) * (1.0f / 0xFF000000);
b = (abgr & 0x00FF0000) * (1.0f / 0x00FF0000); b = (abgr & 0x00FF0000) * (1.0f / 0x00FF0000);
g = (abgr & 0x0000FF00) * (1.0f / 0x0000FF00); g = (abgr & 0x0000FF00) * (1.0f / 0x0000FF00);
...@@ -552,7 +552,7 @@ namespace sw ...@@ -552,7 +552,7 @@ namespace sw
case FORMAT_A8B8G8R8I: case FORMAT_A8B8G8R8I:
{ {
signed char* abgr = (signed char*)element; signed char* abgr = (signed char*)element;
r = abgr[0]; r = abgr[0];
g = abgr[1]; g = abgr[1];
b = abgr[2]; b = abgr[2];
...@@ -572,7 +572,7 @@ namespace sw ...@@ -572,7 +572,7 @@ namespace sw
case FORMAT_X8B8G8R8I_SNORM: case FORMAT_X8B8G8R8I_SNORM:
{ {
signed char* bgr = (signed char*)element; signed char* bgr = (signed char*)element;
r = max(bgr[0] * (1.0f / 0x7F), -1.0f); r = max(bgr[0] * (1.0f / 0x7F), -1.0f);
g = max(bgr[1] * (1.0f / 0x7F), -1.0f); g = max(bgr[1] * (1.0f / 0x7F), -1.0f);
b = max(bgr[2] * (1.0f / 0x7F), -1.0f); b = max(bgr[2] * (1.0f / 0x7F), -1.0f);
...@@ -581,7 +581,7 @@ namespace sw ...@@ -581,7 +581,7 @@ namespace sw
case FORMAT_X8B8G8R8: case FORMAT_X8B8G8R8:
{ {
unsigned int xbgr = *(unsigned int*)element; unsigned int xbgr = *(unsigned int*)element;
b = (xbgr & 0x00FF0000) * (1.0f / 0x00FF0000); b = (xbgr & 0x00FF0000) * (1.0f / 0x00FF0000);
g = (xbgr & 0x0000FF00) * (1.0f / 0x0000FF00); g = (xbgr & 0x0000FF00) * (1.0f / 0x0000FF00);
r = (xbgr & 0x000000FF) * (1.0f / 0x000000FF); r = (xbgr & 0x000000FF) * (1.0f / 0x000000FF);
...@@ -590,7 +590,7 @@ namespace sw ...@@ -590,7 +590,7 @@ namespace sw
case FORMAT_X8B8G8R8I: case FORMAT_X8B8G8R8I:
{ {
signed char* bgr = (signed char*)element; signed char* bgr = (signed char*)element;
r = bgr[0]; r = bgr[0];
g = bgr[1]; g = bgr[1];
b = bgr[2]; b = bgr[2];
...@@ -608,7 +608,7 @@ namespace sw ...@@ -608,7 +608,7 @@ namespace sw
case FORMAT_G8R8I_SNORM: case FORMAT_G8R8I_SNORM:
{ {
signed char* gr = (signed char*)element; signed char* gr = (signed char*)element;
r = (gr[0] & 0xFF00) * (1.0f / 0xFF00); r = (gr[0] & 0xFF00) * (1.0f / 0xFF00);
g = (gr[1] & 0x00FF) * (1.0f / 0x00FF); g = (gr[1] & 0x00FF) * (1.0f / 0x00FF);
} }
...@@ -616,7 +616,7 @@ namespace sw ...@@ -616,7 +616,7 @@ namespace sw
case FORMAT_G8R8: case FORMAT_G8R8:
{ {
unsigned short gr = *(unsigned short*)element; unsigned short gr = *(unsigned short*)element;
g = (gr & 0xFF00) * (1.0f / 0xFF00); g = (gr & 0xFF00) * (1.0f / 0xFF00);
r = (gr & 0x00FF) * (1.0f / 0x00FF); r = (gr & 0x00FF) * (1.0f / 0x00FF);
} }
...@@ -624,7 +624,7 @@ namespace sw ...@@ -624,7 +624,7 @@ namespace sw
case FORMAT_G8R8I: case FORMAT_G8R8I:
{ {
signed char* gr = (signed char*)element; signed char* gr = (signed char*)element;
r = gr[0]; r = gr[0];
g = gr[1]; g = gr[1];
} }
...@@ -632,7 +632,7 @@ namespace sw ...@@ -632,7 +632,7 @@ namespace sw
case FORMAT_G8R8UI: case FORMAT_G8R8UI:
{ {
unsigned char* gr = (unsigned char*)element; unsigned char* gr = (unsigned char*)element;
r = gr[0]; r = gr[0];
g = gr[1]; g = gr[1];
} }
...@@ -646,7 +646,7 @@ namespace sw ...@@ -646,7 +646,7 @@ namespace sw
case FORMAT_G16R16I: case FORMAT_G16R16I:
{ {
short* gr = (short*)element; short* gr = (short*)element;
r = gr[0]; r = gr[0];
g = gr[1]; g = gr[1];
} }
...@@ -654,7 +654,7 @@ namespace sw ...@@ -654,7 +654,7 @@ namespace sw
case FORMAT_G16R16: case FORMAT_G16R16:
{ {
unsigned int gr = *(unsigned int*)element; unsigned int gr = *(unsigned int*)element;
g = (gr & 0xFFFF0000) * (1.0f / 0xFFFF0000); g = (gr & 0xFFFF0000) * (1.0f / 0xFFFF0000);
r = (gr & 0x0000FFFF) * (1.0f / 0x0000FFFF); r = (gr & 0x0000FFFF) * (1.0f / 0x0000FFFF);
} }
...@@ -662,7 +662,7 @@ namespace sw ...@@ -662,7 +662,7 @@ namespace sw
case FORMAT_G16R16UI: case FORMAT_G16R16UI:
{ {
unsigned short* gr = (unsigned short*)element; unsigned short* gr = (unsigned short*)element;
r = gr[0]; r = gr[0];
g = gr[1]; g = gr[1];
} }
...@@ -670,7 +670,7 @@ namespace sw ...@@ -670,7 +670,7 @@ namespace sw
case FORMAT_A2R10G10B10: case FORMAT_A2R10G10B10:
{ {
unsigned int argb = *(unsigned int*)element; unsigned int argb = *(unsigned int*)element;
a = (argb & 0xC0000000) * (1.0f / 0xC0000000); a = (argb & 0xC0000000) * (1.0f / 0xC0000000);
r = (argb & 0x3FF00000) * (1.0f / 0x3FF00000); r = (argb & 0x3FF00000) * (1.0f / 0x3FF00000);
g = (argb & 0x000FFC00) * (1.0f / 0x000FFC00); g = (argb & 0x000FFC00) * (1.0f / 0x000FFC00);
...@@ -680,7 +680,7 @@ namespace sw ...@@ -680,7 +680,7 @@ namespace sw
case FORMAT_A2B10G10R10: case FORMAT_A2B10G10R10:
{ {
unsigned int abgr = *(unsigned int*)element; unsigned int abgr = *(unsigned int*)element;
a = (abgr & 0xC0000000) * (1.0f / 0xC0000000); a = (abgr & 0xC0000000) * (1.0f / 0xC0000000);
b = (abgr & 0x3FF00000) * (1.0f / 0x3FF00000); b = (abgr & 0x3FF00000) * (1.0f / 0x3FF00000);
g = (abgr & 0x000FFC00) * (1.0f / 0x000FFC00); g = (abgr & 0x000FFC00) * (1.0f / 0x000FFC00);
...@@ -690,7 +690,7 @@ namespace sw ...@@ -690,7 +690,7 @@ namespace sw
case FORMAT_A16B16G16R16I: case FORMAT_A16B16G16R16I:
{ {
short* abgr = (short*)element; short* abgr = (short*)element;
r = abgr[0]; r = abgr[0];
g = abgr[1]; g = abgr[1];
b = abgr[2]; b = abgr[2];
...@@ -706,7 +706,7 @@ namespace sw ...@@ -706,7 +706,7 @@ namespace sw
case FORMAT_A16B16G16R16UI: case FORMAT_A16B16G16R16UI:
{ {
unsigned short* abgr = (unsigned short*)element; unsigned short* abgr = (unsigned short*)element;
r = abgr[0]; r = abgr[0];
g = abgr[1]; g = abgr[1];
b = abgr[2]; b = abgr[2];
...@@ -716,7 +716,7 @@ namespace sw ...@@ -716,7 +716,7 @@ namespace sw
case FORMAT_X16B16G16R16I: case FORMAT_X16B16G16R16I:
{ {
short* bgr = (short*)element; short* bgr = (short*)element;
r = bgr[0]; r = bgr[0];
g = bgr[1]; g = bgr[1];
b = bgr[2]; b = bgr[2];
...@@ -725,7 +725,7 @@ namespace sw ...@@ -725,7 +725,7 @@ namespace sw
case FORMAT_X16B16G16R16UI: case FORMAT_X16B16G16R16UI:
{ {
unsigned short* bgr = (unsigned short*)element; unsigned short* bgr = (unsigned short*)element;
r = bgr[0]; r = bgr[0];
g = bgr[1]; g = bgr[1];
b = bgr[2]; b = bgr[2];
...@@ -734,7 +734,7 @@ namespace sw ...@@ -734,7 +734,7 @@ namespace sw
case FORMAT_A32B32G32R32I: case FORMAT_A32B32G32R32I:
{ {
int* abgr = (int*)element; int* abgr = (int*)element;
r = static_cast<float>(abgr[0]); r = static_cast<float>(abgr[0]);
g = static_cast<float>(abgr[1]); g = static_cast<float>(abgr[1]);
b = static_cast<float>(abgr[2]); b = static_cast<float>(abgr[2]);
...@@ -744,7 +744,7 @@ namespace sw ...@@ -744,7 +744,7 @@ namespace sw
case FORMAT_A32B32G32R32UI: case FORMAT_A32B32G32R32UI:
{ {
unsigned int* abgr = (unsigned int*)element; unsigned int* abgr = (unsigned int*)element;
r = static_cast<float>(abgr[0]); r = static_cast<float>(abgr[0]);
g = static_cast<float>(abgr[1]); g = static_cast<float>(abgr[1]);
b = static_cast<float>(abgr[2]); b = static_cast<float>(abgr[2]);
...@@ -754,7 +754,7 @@ namespace sw ...@@ -754,7 +754,7 @@ namespace sw
case FORMAT_X32B32G32R32I: case FORMAT_X32B32G32R32I:
{ {
int* bgr = (int*)element; int* bgr = (int*)element;
r = static_cast<float>(bgr[0]); r = static_cast<float>(bgr[0]);
g = static_cast<float>(bgr[1]); g = static_cast<float>(bgr[1]);
b = static_cast<float>(bgr[2]); b = static_cast<float>(bgr[2]);
...@@ -763,7 +763,7 @@ namespace sw ...@@ -763,7 +763,7 @@ namespace sw
case FORMAT_X32B32G32R32UI: case FORMAT_X32B32G32R32UI:
{ {
unsigned int* bgr = (unsigned int*)element; unsigned int* bgr = (unsigned int*)element;
r = static_cast<float>(bgr[0]); r = static_cast<float>(bgr[0]);
g = static_cast<float>(bgr[1]); g = static_cast<float>(bgr[1]);
b = static_cast<float>(bgr[2]); b = static_cast<float>(bgr[2]);
...@@ -772,7 +772,7 @@ namespace sw ...@@ -772,7 +772,7 @@ namespace sw
case FORMAT_G32R32I: case FORMAT_G32R32I:
{ {
int* gr = (int*)element; int* gr = (int*)element;
r = static_cast<float>(gr[0]); r = static_cast<float>(gr[0]);
g = static_cast<float>(gr[1]); g = static_cast<float>(gr[1]);
} }
...@@ -780,7 +780,7 @@ namespace sw ...@@ -780,7 +780,7 @@ namespace sw
case FORMAT_G32R32UI: case FORMAT_G32R32UI:
{ {
unsigned int* gr = (unsigned int*)element; unsigned int* gr = (unsigned int*)element;
r = static_cast<float>(gr[0]); r = static_cast<float>(gr[0]);
g = static_cast<float>(gr[1]); g = static_cast<float>(gr[1]);
} }
...@@ -802,7 +802,7 @@ namespace sw ...@@ -802,7 +802,7 @@ namespace sw
case FORMAT_L6V5U5: case FORMAT_L6V5U5:
{ {
unsigned short lvu = *(unsigned short*)element; unsigned short lvu = *(unsigned short*)element;
r = ((int)(lvu & 0x001F) << 27) * (1.0f / 0x78000000); r = ((int)(lvu & 0x001F) << 27) * (1.0f / 0x78000000);
g = ((int)(lvu & 0x03E0) << 22) * (1.0f / 0x78000000); g = ((int)(lvu & 0x03E0) << 22) * (1.0f / 0x78000000);
b = (lvu & 0xFC00) * (1.0f / 0xFC00); b = (lvu & 0xFC00) * (1.0f / 0xFC00);
...@@ -811,7 +811,7 @@ namespace sw ...@@ -811,7 +811,7 @@ namespace sw
case FORMAT_Q8W8V8U8: case FORMAT_Q8W8V8U8:
{ {
unsigned int qwvu = *(unsigned int*)element; unsigned int qwvu = *(unsigned int*)element;
r = ((int)(qwvu & 0x000000FF) << 24) * (1.0f / 0x7F000000); r = ((int)(qwvu & 0x000000FF) << 24) * (1.0f / 0x7F000000);
g = ((int)(qwvu & 0x0000FF00) << 16) * (1.0f / 0x7F000000); g = ((int)(qwvu & 0x0000FF00) << 16) * (1.0f / 0x7F000000);
b = ((int)(qwvu & 0x00FF0000) << 8) * (1.0f / 0x7F000000); b = ((int)(qwvu & 0x00FF0000) << 8) * (1.0f / 0x7F000000);
...@@ -821,7 +821,7 @@ namespace sw ...@@ -821,7 +821,7 @@ namespace sw
case FORMAT_X8L8V8U8: case FORMAT_X8L8V8U8:
{ {
unsigned int xlvu = *(unsigned int*)element; unsigned int xlvu = *(unsigned int*)element;
r = ((int)(xlvu & 0x000000FF) << 24) * (1.0f / 0x7F000000); r = ((int)(xlvu & 0x000000FF) << 24) * (1.0f / 0x7F000000);
g = ((int)(xlvu & 0x0000FF00) << 16) * (1.0f / 0x7F000000); g = ((int)(xlvu & 0x0000FF00) << 16) * (1.0f / 0x7F000000);
b = (xlvu & 0x00FF0000) * (1.0f / 0x00FF0000); b = (xlvu & 0x00FF0000) * (1.0f / 0x00FF0000);
...@@ -840,7 +840,7 @@ namespace sw ...@@ -840,7 +840,7 @@ namespace sw
case FORMAT_V16U16: case FORMAT_V16U16:
{ {
unsigned int vu = *(unsigned int*)element; unsigned int vu = *(unsigned int*)element;
r = ((int)(vu & 0x0000FFFF) << 16) * (1.0f / 0x7FFF0000); r = ((int)(vu & 0x0000FFFF) << 16) * (1.0f / 0x7FFF0000);
g = ((int)(vu & 0xFFFF0000) << 0) * (1.0f / 0x7FFF0000); g = ((int)(vu & 0xFFFF0000) << 0) * (1.0f / 0x7FFF0000);
} }
...@@ -848,7 +848,7 @@ namespace sw ...@@ -848,7 +848,7 @@ namespace sw
case FORMAT_A2W10V10U10: case FORMAT_A2W10V10U10:
{ {
unsigned int awvu = *(unsigned int*)element; unsigned int awvu = *(unsigned int*)element;
r = ((int)(awvu & 0x000003FF) << 22) * (1.0f / 0x7FC00000); r = ((int)(awvu & 0x000003FF) << 22) * (1.0f / 0x7FC00000);
g = ((int)(awvu & 0x000FFC00) << 12) * (1.0f / 0x7FC00000); g = ((int)(awvu & 0x000FFC00) << 12) * (1.0f / 0x7FC00000);
b = ((int)(awvu & 0x3FF00000) << 2) * (1.0f / 0x7FC00000); b = ((int)(awvu & 0x3FF00000) << 2) * (1.0f / 0x7FC00000);
...@@ -875,7 +875,7 @@ namespace sw ...@@ -875,7 +875,7 @@ namespace sw
case FORMAT_A4L4: case FORMAT_A4L4:
{ {
unsigned char al = *(unsigned char*)element; unsigned char al = *(unsigned char*)element;
r = r =
g = g =
b = (al & 0x0F) * (1.0f / 0x0F); b = (al & 0x0F) * (1.0f / 0x0F);
...@@ -1755,7 +1755,7 @@ namespace sw ...@@ -1755,7 +1755,7 @@ namespace sw
{ {
// ASSERT(source.lock != LOCK_UNLOCKED); // ASSERT(source.lock != LOCK_UNLOCKED);
// ASSERT(destination.lock != LOCK_UNLOCKED); // ASSERT(destination.lock != LOCK_UNLOCKED);
if(destination.buffer != source.buffer) if(destination.buffer != source.buffer)
{ {
ASSERT(source.dirty && !destination.dirty); ASSERT(source.dirty && !destination.dirty);
...@@ -1918,7 +1918,7 @@ namespace sw ...@@ -1918,7 +1918,7 @@ namespace sw
for(int x = 0; x < destination.width && x < source.width; x++) for(int x = 0; x < destination.width && x < source.width; x++)
{ {
unsigned int xrgb = *(unsigned short*)sourceElement; unsigned int xrgb = *(unsigned short*)sourceElement;
unsigned int r = (((xrgb & 0x7C00) * 134771 + 0x800000) >> 8) & 0x00FF0000; unsigned int r = (((xrgb & 0x7C00) * 134771 + 0x800000) >> 8) & 0x00FF0000;
unsigned int g = (((xrgb & 0x03E0) * 16846 + 0x8000) >> 8) & 0x0000FF00; unsigned int g = (((xrgb & 0x03E0) * 16846 + 0x8000) >> 8) & 0x0000FF00;
unsigned int b = (((xrgb & 0x001F) * 2106 + 0x80) >> 8); unsigned int b = (((xrgb & 0x001F) * 2106 + 0x80) >> 8);
...@@ -1956,7 +1956,7 @@ namespace sw ...@@ -1956,7 +1956,7 @@ namespace sw
for(int x = 0; x < destination.width && x < source.width; x++) for(int x = 0; x < destination.width && x < source.width; x++)
{ {
unsigned int argb = *(unsigned short*)sourceElement; unsigned int argb = *(unsigned short*)sourceElement;
unsigned int a = (argb & 0x8000) * 130560; unsigned int a = (argb & 0x8000) * 130560;
unsigned int r = (((argb & 0x7C00) * 134771 + 0x800000) >> 8) & 0x00FF0000; unsigned int r = (((argb & 0x7C00) * 134771 + 0x800000) >> 8) & 0x00FF0000;
unsigned int g = (((argb & 0x03E0) * 16846 + 0x8000) >> 8) & 0x0000FF00; unsigned int g = (((argb & 0x03E0) * 16846 + 0x8000) >> 8) & 0x0000FF00;
...@@ -1995,7 +1995,7 @@ namespace sw ...@@ -1995,7 +1995,7 @@ namespace sw
for(int x = 0; x < destination.width && x < source.width; x++) for(int x = 0; x < destination.width && x < source.width; x++)
{ {
unsigned int xrgb = *(unsigned short*)sourceElement; unsigned int xrgb = *(unsigned short*)sourceElement;
unsigned int r = ((xrgb & 0x0F00) * 0x00001100) & 0x00FF0000; unsigned int r = ((xrgb & 0x0F00) * 0x00001100) & 0x00FF0000;
unsigned int g = ((xrgb & 0x00F0) * 0x00000110) & 0x0000FF00; unsigned int g = ((xrgb & 0x00F0) * 0x00000110) & 0x0000FF00;
unsigned int b = (xrgb & 0x000F) * 0x00000011; unsigned int b = (xrgb & 0x000F) * 0x00000011;
...@@ -2033,7 +2033,7 @@ namespace sw ...@@ -2033,7 +2033,7 @@ namespace sw
for(int x = 0; x < destination.width && x < source.width; x++) for(int x = 0; x < destination.width && x < source.width; x++)
{ {
unsigned int argb = *(unsigned short*)sourceElement; unsigned int argb = *(unsigned short*)sourceElement;
unsigned int a = ((argb & 0xF000) * 0x00011000) & 0xFF000000; unsigned int a = ((argb & 0xF000) * 0x00011000) & 0xFF000000;
unsigned int r = ((argb & 0x0F00) * 0x00001100) & 0x00FF0000; unsigned int r = ((argb & 0x0F00) * 0x00001100) & 0x00FF0000;
unsigned int g = ((argb & 0x00F0) * 0x00000110) & 0x0000FF00; unsigned int g = ((argb & 0x00F0) * 0x00000110) & 0x0000FF00;
...@@ -2235,7 +2235,7 @@ namespace sw ...@@ -2235,7 +2235,7 @@ namespace sw
a[0] = source->a0; a[0] = source->a0;
a[1] = source->a1; a[1] = source->a1;
if(a[0] > a[1]) if(a[0] > a[1])
{ {
a[2] = (byte)((6 * (word)a[0] + 1 * (word)a[1] + 3) / 7); a[2] = (byte)((6 * (word)a[0] + 1 * (word)a[1] + 3) / 7);
...@@ -2261,7 +2261,7 @@ namespace sw ...@@ -2261,7 +2261,7 @@ namespace sw
{ {
unsigned int alpha = (unsigned int)a[(unsigned int)(source->alut >> (16 + 3 * (i + j * 4))) % 8] << 24; unsigned int alpha = (unsigned int)a[(unsigned int)(source->alut >> (16 + 3 * (i + j * 4))) % 8] << 24;
unsigned int color = (c[(source->clut >> 2 * (i + j * 4)) % 4] & 0x00FFFFFF) | alpha; unsigned int color = (c[(source->clut >> 2 * (i + j * 4)) % 4] & 0x00FFFFFF) | alpha;
dest[(x + i) + (y + j) * internal.width] = color; dest[(x + i) + (y + j) * internal.width] = color;
} }
} }
...@@ -2292,7 +2292,7 @@ namespace sw ...@@ -2292,7 +2292,7 @@ namespace sw
r[0] = source->r0; r[0] = source->r0;
r[1] = source->r1; r[1] = source->r1;
if(r[0] > r[1]) if(r[0] > r[1])
{ {
r[2] = (byte)((6 * (word)r[0] + 1 * (word)r[1] + 3) / 7); r[2] = (byte)((6 * (word)r[0] + 1 * (word)r[1] + 3) / 7);
...@@ -2345,7 +2345,7 @@ namespace sw ...@@ -2345,7 +2345,7 @@ namespace sw
X[0] = source->x0; X[0] = source->x0;
X[1] = source->x1; X[1] = source->x1;
if(X[0] > X[1]) if(X[0] > X[1])
{ {
X[2] = (byte)((6 * (word)X[0] + 1 * (word)X[1] + 3) / 7); X[2] = (byte)((6 * (word)X[0] + 1 * (word)X[1] + 3) / 7);
...@@ -2369,7 +2369,7 @@ namespace sw ...@@ -2369,7 +2369,7 @@ namespace sw
Y[0] = source->y0; Y[0] = source->y0;
Y[1] = source->y1; Y[1] = source->y1;
if(Y[0] > Y[1]) if(Y[0] > Y[1])
{ {
Y[2] = (byte)((6 * (word)Y[0] + 1 * (word)Y[1] + 3) / 7); Y[2] = (byte)((6 * (word)Y[0] + 1 * (word)Y[1] + 3) / 7);
...@@ -2704,7 +2704,7 @@ namespace sw ...@@ -2704,7 +2704,7 @@ namespace sw
default: default:
ASSERT(false); ASSERT(false);
} }
return false; return false;
} }
...@@ -2782,7 +2782,7 @@ namespace sw ...@@ -2782,7 +2782,7 @@ namespace sw
default: default:
ASSERT(false); ASSERT(false);
} }
return false; return false;
} }
...@@ -3030,7 +3030,7 @@ namespace sw ...@@ -3030,7 +3030,7 @@ namespace sw
} }
__m128 quad = _mm_set_ps1((float&)pattern); __m128 quad = _mm_set_ps1((float&)pattern);
float *pointer = (float*)buffer; float *pointer = (float*)buffer;
int qxwords = bytes / 64; int qxwords = bytes / 64;
bytes -= qxwords * 64; bytes -= qxwords * 64;
...@@ -3075,29 +3075,12 @@ namespace sw ...@@ -3075,29 +3075,12 @@ namespace sw
return (rect.x0 == 0 && rect.y0 == 0 && rect.x1 == internal.width && rect.y1 == internal.height && internal.depth == 1); return (rect.x0 == 0 && rect.y0 == 0 && rect.x1 == internal.width && rect.y1 == internal.height && internal.depth == 1);
} }
bool Surface::getClearRect(int x0, int y0, int width, int height, SliceRect& rect) const SliceRect Surface::getRect() const
{ {
// Not overlapping return SliceRect(0, 0, internal.width, internal.height, 0);
if(x0 > internal.width) return false;
if(y0 > internal.height) return false;
if(x0 + width < 0) return false;
if(y0 + height < 0) return false;
// Clip against dimensions
if(x0 < 0) { width += x0; x0 = 0; }
if(x0 + width > internal.width) width = internal.width - x0;
if(y0 < 0) { height += y0; y0 = 0; }
if(y0 + height > internal.height) height = internal.height - y0;
rect.x0 = x0;
rect.x1 = x0 + width;
rect.y0 = y0;
rect.y1 = y0 + height;
return true;
} }
void Surface::clearDepthBuffer(float depth, int x0, int y0, int width, int height) void Surface::clearDepth(float depth, int x0, int y0, int width, int height)
{ {
if(width == 0 || height == 0) return; if(width == 0 || height == 0) return;
...@@ -3157,7 +3140,7 @@ namespace sw ...@@ -3157,7 +3140,7 @@ namespace sw
for(int y = y0; y < y1; y++) for(int y = y0; y < y1; y++)
{ {
float *target = buffer + (y & ~1) * width2 + (y & 1) * 2; float *target = buffer + (y & ~1) * width2 + (y & 1) * 2;
if((y & 1) == 0 && y + 1 < y1) // Fill quad line at once if((y & 1) == 0 && y + 1 < y1) // Fill quad line at once
{ {
if((x0 & 1) != 0) if((x0 & 1) != 0)
...@@ -3222,7 +3205,7 @@ namespace sw ...@@ -3222,7 +3205,7 @@ namespace sw
} }
} }
void Surface::clearStencilBuffer(unsigned char s, unsigned char mask, int x0, int y0, int width, int height) void Surface::clearStencil(unsigned char s, unsigned char mask, int x0, int y0, int width, int height)
{ {
if(mask == 0 || width == 0 || height == 0) return; if(mask == 0 || width == 0 || height == 0) return;
...@@ -3299,7 +3282,7 @@ namespace sw ...@@ -3299,7 +3282,7 @@ namespace sw
{ {
unsigned char *row; unsigned char *row;
Buffer *buffer; Buffer *buffer;
if(internal.dirty) if(internal.dirty)
{ {
row = (unsigned char*)lockInternal(x0, y0, 0, LOCK_WRITEONLY, PUBLIC); row = (unsigned char*)lockInternal(x0, y0, 0, LOCK_WRITEONLY, PUBLIC);
...@@ -3393,7 +3376,7 @@ namespace sw ...@@ -3393,7 +3376,7 @@ namespace sw
{ {
return isStencil(external.format); return isStencil(external.format);
} }
bool Surface::hasDepth() const bool Surface::hasDepth() const
{ {
return isDepth(external.format); return isDepth(external.format);
...@@ -3708,7 +3691,7 @@ namespace sw ...@@ -3708,7 +3691,7 @@ namespace sw
{ {
__m128i c0 = _mm_load_si128((__m128i*)(source0 + 4 * x)); __m128i c0 = _mm_load_si128((__m128i*)(source0 + 4 * x));
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x));
c0 = _mm_avg_epu8(c0, c1); c0 = _mm_avg_epu8(c0, c1);
_mm_store_si128((__m128i*)(source0 + 4 * x), c0); _mm_store_si128((__m128i*)(source0 + 4 * x), c0);
...@@ -3728,7 +3711,7 @@ namespace sw ...@@ -3728,7 +3711,7 @@ namespace sw
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x));
__m128i c2 = _mm_load_si128((__m128i*)(source2 + 4 * x)); __m128i c2 = _mm_load_si128((__m128i*)(source2 + 4 * x));
__m128i c3 = _mm_load_si128((__m128i*)(source3 + 4 * x)); __m128i c3 = _mm_load_si128((__m128i*)(source3 + 4 * x));
c0 = _mm_avg_epu8(c0, c1); c0 = _mm_avg_epu8(c0, c1);
c2 = _mm_avg_epu8(c2, c3); c2 = _mm_avg_epu8(c2, c3);
c0 = _mm_avg_epu8(c0, c2); c0 = _mm_avg_epu8(c0, c2);
...@@ -3756,7 +3739,7 @@ namespace sw ...@@ -3756,7 +3739,7 @@ namespace sw
__m128i c5 = _mm_load_si128((__m128i*)(source5 + 4 * x)); __m128i c5 = _mm_load_si128((__m128i*)(source5 + 4 * x));
__m128i c6 = _mm_load_si128((__m128i*)(source6 + 4 * x)); __m128i c6 = _mm_load_si128((__m128i*)(source6 + 4 * x));
__m128i c7 = _mm_load_si128((__m128i*)(source7 + 4 * x)); __m128i c7 = _mm_load_si128((__m128i*)(source7 + 4 * x));
c0 = _mm_avg_epu8(c0, c1); c0 = _mm_avg_epu8(c0, c1);
c2 = _mm_avg_epu8(c2, c3); c2 = _mm_avg_epu8(c2, c3);
c4 = _mm_avg_epu8(c4, c5); c4 = _mm_avg_epu8(c4, c5);
...@@ -3999,7 +3982,7 @@ namespace sw ...@@ -3999,7 +3982,7 @@ namespace sw
{ {
__m128i c0 = _mm_load_si128((__m128i*)(source0 + 4 * x)); __m128i c0 = _mm_load_si128((__m128i*)(source0 + 4 * x));
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x));
c0 = _mm_avg_epu16(c0, c1); c0 = _mm_avg_epu16(c0, c1);
_mm_store_si128((__m128i*)(source0 + 4 * x), c0); _mm_store_si128((__m128i*)(source0 + 4 * x), c0);
...@@ -4019,7 +4002,7 @@ namespace sw ...@@ -4019,7 +4002,7 @@ namespace sw
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 4 * x));
__m128i c2 = _mm_load_si128((__m128i*)(source2 + 4 * x)); __m128i c2 = _mm_load_si128((__m128i*)(source2 + 4 * x));
__m128i c3 = _mm_load_si128((__m128i*)(source3 + 4 * x)); __m128i c3 = _mm_load_si128((__m128i*)(source3 + 4 * x));
c0 = _mm_avg_epu16(c0, c1); c0 = _mm_avg_epu16(c0, c1);
c2 = _mm_avg_epu16(c2, c3); c2 = _mm_avg_epu16(c2, c3);
c0 = _mm_avg_epu16(c0, c2); c0 = _mm_avg_epu16(c0, c2);
...@@ -4047,7 +4030,7 @@ namespace sw ...@@ -4047,7 +4030,7 @@ namespace sw
__m128i c5 = _mm_load_si128((__m128i*)(source5 + 4 * x)); __m128i c5 = _mm_load_si128((__m128i*)(source5 + 4 * x));
__m128i c6 = _mm_load_si128((__m128i*)(source6 + 4 * x)); __m128i c6 = _mm_load_si128((__m128i*)(source6 + 4 * x));
__m128i c7 = _mm_load_si128((__m128i*)(source7 + 4 * x)); __m128i c7 = _mm_load_si128((__m128i*)(source7 + 4 * x));
c0 = _mm_avg_epu16(c0, c1); c0 = _mm_avg_epu16(c0, c1);
c2 = _mm_avg_epu16(c2, c3); c2 = _mm_avg_epu16(c2, c3);
c4 = _mm_avg_epu16(c4, c5); c4 = _mm_avg_epu16(c4, c5);
...@@ -4290,7 +4273,7 @@ namespace sw ...@@ -4290,7 +4273,7 @@ namespace sw
{ {
__m128i c0 = _mm_load_si128((__m128i*)(source0 + 8 * x)); __m128i c0 = _mm_load_si128((__m128i*)(source0 + 8 * x));
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 8 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 8 * x));
c0 = _mm_avg_epu16(c0, c1); c0 = _mm_avg_epu16(c0, c1);
_mm_store_si128((__m128i*)(source0 + 8 * x), c0); _mm_store_si128((__m128i*)(source0 + 8 * x), c0);
...@@ -4310,7 +4293,7 @@ namespace sw ...@@ -4310,7 +4293,7 @@ namespace sw
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 8 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 8 * x));
__m128i c2 = _mm_load_si128((__m128i*)(source2 + 8 * x)); __m128i c2 = _mm_load_si128((__m128i*)(source2 + 8 * x));
__m128i c3 = _mm_load_si128((__m128i*)(source3 + 8 * x)); __m128i c3 = _mm_load_si128((__m128i*)(source3 + 8 * x));
c0 = _mm_avg_epu16(c0, c1); c0 = _mm_avg_epu16(c0, c1);
c2 = _mm_avg_epu16(c2, c3); c2 = _mm_avg_epu16(c2, c3);
c0 = _mm_avg_epu16(c0, c2); c0 = _mm_avg_epu16(c0, c2);
...@@ -4338,7 +4321,7 @@ namespace sw ...@@ -4338,7 +4321,7 @@ namespace sw
__m128i c5 = _mm_load_si128((__m128i*)(source5 + 8 * x)); __m128i c5 = _mm_load_si128((__m128i*)(source5 + 8 * x));
__m128i c6 = _mm_load_si128((__m128i*)(source6 + 8 * x)); __m128i c6 = _mm_load_si128((__m128i*)(source6 + 8 * x));
__m128i c7 = _mm_load_si128((__m128i*)(source7 + 8 * x)); __m128i c7 = _mm_load_si128((__m128i*)(source7 + 8 * x));
c0 = _mm_avg_epu16(c0, c1); c0 = _mm_avg_epu16(c0, c1);
c2 = _mm_avg_epu16(c2, c3); c2 = _mm_avg_epu16(c2, c3);
c4 = _mm_avg_epu16(c4, c5); c4 = _mm_avg_epu16(c4, c5);
...@@ -4581,9 +4564,9 @@ namespace sw ...@@ -4581,9 +4564,9 @@ namespace sw
{ {
__m128 c0 = _mm_load_ps((float*)(source0 + 4 * x)); __m128 c0 = _mm_load_ps((float*)(source0 + 4 * x));
__m128 c1 = _mm_load_ps((float*)(source1 + 4 * x)); __m128 c1 = _mm_load_ps((float*)(source1 + 4 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f));
_mm_store_ps((float*)(source0 + 4 * x), c0); _mm_store_ps((float*)(source0 + 4 * x), c0);
} }
...@@ -4602,11 +4585,11 @@ namespace sw ...@@ -4602,11 +4585,11 @@ namespace sw
__m128 c1 = _mm_load_ps((float*)(source1 + 4 * x)); __m128 c1 = _mm_load_ps((float*)(source1 + 4 * x));
__m128 c2 = _mm_load_ps((float*)(source2 + 4 * x)); __m128 c2 = _mm_load_ps((float*)(source2 + 4 * x));
__m128 c3 = _mm_load_ps((float*)(source3 + 4 * x)); __m128 c3 = _mm_load_ps((float*)(source3 + 4 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c2 = _mm_add_ps(c2, c3); c2 = _mm_add_ps(c2, c3);
c0 = _mm_add_ps(c0, c2); c0 = _mm_add_ps(c0, c2);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f));
_mm_store_ps((float*)(source0 + 4 * x), c0); _mm_store_ps((float*)(source0 + 4 * x), c0);
} }
...@@ -4631,7 +4614,7 @@ namespace sw ...@@ -4631,7 +4614,7 @@ namespace sw
__m128 c5 = _mm_load_ps((float*)(source5 + 4 * x)); __m128 c5 = _mm_load_ps((float*)(source5 + 4 * x));
__m128 c6 = _mm_load_ps((float*)(source6 + 4 * x)); __m128 c6 = _mm_load_ps((float*)(source6 + 4 * x));
__m128 c7 = _mm_load_ps((float*)(source7 + 4 * x)); __m128 c7 = _mm_load_ps((float*)(source7 + 4 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c2 = _mm_add_ps(c2, c3); c2 = _mm_add_ps(c2, c3);
c4 = _mm_add_ps(c4, c5); c4 = _mm_add_ps(c4, c5);
...@@ -4639,7 +4622,7 @@ namespace sw ...@@ -4639,7 +4622,7 @@ namespace sw
c0 = _mm_add_ps(c0, c2); c0 = _mm_add_ps(c0, c2);
c4 = _mm_add_ps(c4, c6); c4 = _mm_add_ps(c4, c6);
c0 = _mm_add_ps(c0, c4); c0 = _mm_add_ps(c0, c4);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f));
_mm_store_ps((float*)(source0 + 4 * x), c0); _mm_store_ps((float*)(source0 + 4 * x), c0);
} }
...@@ -4692,7 +4675,7 @@ namespace sw ...@@ -4692,7 +4675,7 @@ namespace sw
c0 = _mm_add_ps(c0, c4); c0 = _mm_add_ps(c0, c4);
c8 = _mm_add_ps(c8, cC); c8 = _mm_add_ps(c8, cC);
c0 = _mm_add_ps(c0, c8); c0 = _mm_add_ps(c0, c8);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f));
_mm_store_ps((float*)(source0 + 4 * x), c0); _mm_store_ps((float*)(source0 + 4 * x), c0);
} }
...@@ -4876,9 +4859,9 @@ namespace sw ...@@ -4876,9 +4859,9 @@ namespace sw
{ {
__m128 c0 = _mm_load_ps((float*)(source0 + 8 * x)); __m128 c0 = _mm_load_ps((float*)(source0 + 8 * x));
__m128 c1 = _mm_load_ps((float*)(source1 + 8 * x)); __m128 c1 = _mm_load_ps((float*)(source1 + 8 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f));
_mm_store_ps((float*)(source0 + 8 * x), c0); _mm_store_ps((float*)(source0 + 8 * x), c0);
} }
...@@ -4897,11 +4880,11 @@ namespace sw ...@@ -4897,11 +4880,11 @@ namespace sw
__m128 c1 = _mm_load_ps((float*)(source1 + 8 * x)); __m128 c1 = _mm_load_ps((float*)(source1 + 8 * x));
__m128 c2 = _mm_load_ps((float*)(source2 + 8 * x)); __m128 c2 = _mm_load_ps((float*)(source2 + 8 * x));
__m128 c3 = _mm_load_ps((float*)(source3 + 8 * x)); __m128 c3 = _mm_load_ps((float*)(source3 + 8 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c2 = _mm_add_ps(c2, c3); c2 = _mm_add_ps(c2, c3);
c0 = _mm_add_ps(c0, c2); c0 = _mm_add_ps(c0, c2);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f));
_mm_store_ps((float*)(source0 + 8 * x), c0); _mm_store_ps((float*)(source0 + 8 * x), c0);
} }
...@@ -4926,7 +4909,7 @@ namespace sw ...@@ -4926,7 +4909,7 @@ namespace sw
__m128 c5 = _mm_load_ps((float*)(source5 + 8 * x)); __m128 c5 = _mm_load_ps((float*)(source5 + 8 * x));
__m128 c6 = _mm_load_ps((float*)(source6 + 8 * x)); __m128 c6 = _mm_load_ps((float*)(source6 + 8 * x));
__m128 c7 = _mm_load_ps((float*)(source7 + 8 * x)); __m128 c7 = _mm_load_ps((float*)(source7 + 8 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c2 = _mm_add_ps(c2, c3); c2 = _mm_add_ps(c2, c3);
c4 = _mm_add_ps(c4, c5); c4 = _mm_add_ps(c4, c5);
...@@ -4934,7 +4917,7 @@ namespace sw ...@@ -4934,7 +4917,7 @@ namespace sw
c0 = _mm_add_ps(c0, c2); c0 = _mm_add_ps(c0, c2);
c4 = _mm_add_ps(c4, c6); c4 = _mm_add_ps(c4, c6);
c0 = _mm_add_ps(c0, c4); c0 = _mm_add_ps(c0, c4);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f));
_mm_store_ps((float*)(source0 + 8 * x), c0); _mm_store_ps((float*)(source0 + 8 * x), c0);
} }
...@@ -4987,7 +4970,7 @@ namespace sw ...@@ -4987,7 +4970,7 @@ namespace sw
c0 = _mm_add_ps(c0, c4); c0 = _mm_add_ps(c0, c4);
c8 = _mm_add_ps(c8, cC); c8 = _mm_add_ps(c8, cC);
c0 = _mm_add_ps(c0, c8); c0 = _mm_add_ps(c0, c8);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f));
_mm_store_ps((float*)(source0 + 8 * x), c0); _mm_store_ps((float*)(source0 + 8 * x), c0);
} }
...@@ -5171,9 +5154,9 @@ namespace sw ...@@ -5171,9 +5154,9 @@ namespace sw
{ {
__m128 c0 = _mm_load_ps((float*)(source0 + 16 * x)); __m128 c0 = _mm_load_ps((float*)(source0 + 16 * x));
__m128 c1 = _mm_load_ps((float*)(source1 + 16 * x)); __m128 c1 = _mm_load_ps((float*)(source1 + 16 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 2.0f));
_mm_store_ps((float*)(source0 + 16 * x), c0); _mm_store_ps((float*)(source0 + 16 * x), c0);
} }
...@@ -5192,11 +5175,11 @@ namespace sw ...@@ -5192,11 +5175,11 @@ namespace sw
__m128 c1 = _mm_load_ps((float*)(source1 + 16 * x)); __m128 c1 = _mm_load_ps((float*)(source1 + 16 * x));
__m128 c2 = _mm_load_ps((float*)(source2 + 16 * x)); __m128 c2 = _mm_load_ps((float*)(source2 + 16 * x));
__m128 c3 = _mm_load_ps((float*)(source3 + 16 * x)); __m128 c3 = _mm_load_ps((float*)(source3 + 16 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c2 = _mm_add_ps(c2, c3); c2 = _mm_add_ps(c2, c3);
c0 = _mm_add_ps(c0, c2); c0 = _mm_add_ps(c0, c2);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 4.0f));
_mm_store_ps((float*)(source0 + 16 * x), c0); _mm_store_ps((float*)(source0 + 16 * x), c0);
} }
...@@ -5221,7 +5204,7 @@ namespace sw ...@@ -5221,7 +5204,7 @@ namespace sw
__m128 c5 = _mm_load_ps((float*)(source5 + 16 * x)); __m128 c5 = _mm_load_ps((float*)(source5 + 16 * x));
__m128 c6 = _mm_load_ps((float*)(source6 + 16 * x)); __m128 c6 = _mm_load_ps((float*)(source6 + 16 * x));
__m128 c7 = _mm_load_ps((float*)(source7 + 16 * x)); __m128 c7 = _mm_load_ps((float*)(source7 + 16 * x));
c0 = _mm_add_ps(c0, c1); c0 = _mm_add_ps(c0, c1);
c2 = _mm_add_ps(c2, c3); c2 = _mm_add_ps(c2, c3);
c4 = _mm_add_ps(c4, c5); c4 = _mm_add_ps(c4, c5);
...@@ -5229,7 +5212,7 @@ namespace sw ...@@ -5229,7 +5212,7 @@ namespace sw
c0 = _mm_add_ps(c0, c2); c0 = _mm_add_ps(c0, c2);
c4 = _mm_add_ps(c4, c6); c4 = _mm_add_ps(c4, c6);
c0 = _mm_add_ps(c0, c4); c0 = _mm_add_ps(c0, c4);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 8.0f));
_mm_store_ps((float*)(source0 + 16 * x), c0); _mm_store_ps((float*)(source0 + 16 * x), c0);
} }
...@@ -5282,7 +5265,7 @@ namespace sw ...@@ -5282,7 +5265,7 @@ namespace sw
c0 = _mm_add_ps(c0, c4); c0 = _mm_add_ps(c0, c4);
c8 = _mm_add_ps(c8, cC); c8 = _mm_add_ps(c8, cC);
c0 = _mm_add_ps(c0, c8); c0 = _mm_add_ps(c0, c8);
c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f)); c0 = _mm_mul_ps(c0, _mm_set1_ps(1.0f / 16.0f));
_mm_store_ps((float*)(source0 + 16 * x), c0); _mm_store_ps((float*)(source0 + 16 * x), c0);
} }
...@@ -5466,7 +5449,7 @@ namespace sw ...@@ -5466,7 +5449,7 @@ namespace sw
{ {
__m128i c0 = _mm_load_si128((__m128i*)(source0 + 2 * x)); __m128i c0 = _mm_load_si128((__m128i*)(source0 + 2 * x));
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 2 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 2 * x));
static const ushort8 r_b = {0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F}; static const ushort8 r_b = {0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F};
static const ushort8 _g_ = {0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0}; static const ushort8 _g_ = {0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0};
__m128i c0_r_b = _mm_and_si128(c0, reinterpret_cast<const __m128i&>(r_b)); __m128i c0_r_b = _mm_and_si128(c0, reinterpret_cast<const __m128i&>(r_b));
...@@ -5497,7 +5480,7 @@ namespace sw ...@@ -5497,7 +5480,7 @@ namespace sw
__m128i c1 = _mm_load_si128((__m128i*)(source1 + 2 * x)); __m128i c1 = _mm_load_si128((__m128i*)(source1 + 2 * x));
__m128i c2 = _mm_load_si128((__m128i*)(source2 + 2 * x)); __m128i c2 = _mm_load_si128((__m128i*)(source2 + 2 * x));
__m128i c3 = _mm_load_si128((__m128i*)(source3 + 2 * x)); __m128i c3 = _mm_load_si128((__m128i*)(source3 + 2 * x));
static const ushort8 r_b = {0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F}; static const ushort8 r_b = {0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F};
static const ushort8 _g_ = {0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0}; static const ushort8 _g_ = {0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0};
__m128i c0_r_b = _mm_and_si128(c0, reinterpret_cast<const __m128i&>(r_b)); __m128i c0_r_b = _mm_and_si128(c0, reinterpret_cast<const __m128i&>(r_b));
...@@ -5542,7 +5525,7 @@ namespace sw ...@@ -5542,7 +5525,7 @@ namespace sw
__m128i c5 = _mm_load_si128((__m128i*)(source5 + 2 * x)); __m128i c5 = _mm_load_si128((__m128i*)(source5 + 2 * x));
__m128i c6 = _mm_load_si128((__m128i*)(source6 + 2 * x)); __m128i c6 = _mm_load_si128((__m128i*)(source6 + 2 * x));
__m128i c7 = _mm_load_si128((__m128i*)(source7 + 2 * x)); __m128i c7 = _mm_load_si128((__m128i*)(source7 + 2 * x));
static const ushort8 r_b = {0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F}; static const ushort8 r_b = {0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F, 0xF81F};
static const ushort8 _g_ = {0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0}; static const ushort8 _g_ = {0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0, 0x07E0};
__m128i c0_r_b = _mm_and_si128(c0, reinterpret_cast<const __m128i&>(r_b)); __m128i c0_r_b = _mm_and_si128(c0, reinterpret_cast<const __m128i&>(r_b));
......
src/Renderer/Surface.hpp
...@@ -27,6 +27,9 @@ namespace sw ...@@ -27,6 +27,9 @@ namespace sw
void clip(int minX, int minY, int maxX, int maxY); void clip(int minX, int minY, int maxX, int maxY);
int width() const { return x1 - x0; }
int height() const { return y1 - y0; }
int x0; // Inclusive int x0; // Inclusive
int y0; // Inclusive int y0; // Inclusive
int x1; // Exclusive int x1; // Exclusive
...@@ -154,7 +157,7 @@ namespace sw ...@@ -154,7 +157,7 @@ namespace sw
FORMAT_R32F, FORMAT_R32F,
FORMAT_G32R32F, FORMAT_G32R32F,
FORMAT_B32G32R32F, FORMAT_B32G32R32F,
FORMAT_A32B32G32R32F, FORMAT_A32B32G32R32F,
// Bump map formats // Bump map formats
FORMAT_V8U8, FORMAT_V8U8,
FORMAT_L6V5U5, FORMAT_L6V5U5,
...@@ -283,9 +286,9 @@ namespace sw ...@@ -283,9 +286,9 @@ namespace sw
inline int getSuperSampleCount() const; inline int getSuperSampleCount() const;
bool isEntire(const SliceRect& rect) const; bool isEntire(const SliceRect& rect) const;
bool getClearRect(int x0, int y0, int width, int height, SliceRect& rect) const; SliceRect getRect() const;
void clearDepthBuffer(float depth, int x0, int y0, int width, int height); void clearDepth(float depth, int x0, int y0, int width, int height);
void clearStencilBuffer(unsigned char stencil, unsigned char mask, int x0, int y0, int width, int height); void clearStencil(unsigned char stencil, unsigned char mask, int x0, int y0, int width, int height);
void fill(const Color<float> &color, int x0, int y0, int width, int height); void fill(const Color<float> &color, int x0, int y0, int width, int height);
Color<float> readExternal(int x, int y, int z) const; Color<float> readExternal(int x, int y, int z) const;
...@@ -294,7 +297,7 @@ namespace sw ...@@ -294,7 +297,7 @@ namespace sw
Color<float> sampleExternal(float x, float y) const; Color<float> sampleExternal(float x, float y) const;
void writeExternal(int x, int y, int z, const Color<float> &color); void writeExternal(int x, int y, int z, const Color<float> &color);
void writeExternal(int x, int y, const Color<float> &color); void writeExternal(int x, int y, const Color<float> &color);
void copyInternal(const Surface* src, int x, int y, float srcX, float srcY, bool filter); void copyInternal(const Surface* src, int x, int y, float srcX, float srcY, bool filter);
void copyInternal(const Surface* src, int x, int y, int z, float srcX, float srcY, float srcZ, bool filter); void copyInternal(const Surface* src, int x, int y, int z, float srcX, float srcY, float srcZ, bool filter);
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment