void Processor::Impl::getGpuLut3D(float* lut3d, const GpuShaderDesc & shaderDesc) const
{
if(!lut3d) return;
AutoMutex lock(m_resultsCacheMutex);
if(m_lastShaderDesc != shaderDesc.getCacheID())
{
m_lastShaderDesc = shaderDesc.getCacheID();
m_shader = "";
m_shaderCacheID = "";
m_lut3D.clear();
m_lut3DCacheID = "";
}
int lut3DEdgeLen = shaderDesc.getLut3DEdgeLen();
int lut3DNumPixels = lut3DEdgeLen*lut3DEdgeLen*lut3DEdgeLen;
// Can we write the entire shader using only shader text?
// If so, the lut3D is not needed so clear it.
// This is preferable to identity, as it lets people notice if
// it's accidentally being used.
if(m_gpuOpsCpuLatticeProcess.empty())
{
memset(lut3d, 0, sizeof(float) * 3 * lut3DNumPixels);
return;
}
if(m_lut3D.empty())
{
// Allocate 3dlut image, RGBA
m_lut3D.resize(lut3DNumPixels*4);
GenerateIdentityLut3D(&m_lut3D[0], lut3DEdgeLen, 4);
// Apply the lattice ops to it
for(int i=0; i<(int)m_gpuOpsCpuLatticeProcess.size(); ++i)
{
m_gpuOpsCpuLatticeProcess[i]->apply(&m_lut3D[0], lut3DNumPixels);
}
// Convert the RGBA image to an RGB image, in place.
// Of course, this only works because we're doing it from left to right
// so old pixels are read before they're written over
// TODO: is this bad for memory access patterns?
// see if this is faster with a 2nd temp float array
for(int i=1; i<lut3DNumPixels; ++i) // skip the 1st pixel, it's ok.
{
m_lut3D[3*i+0] = m_lut3D[4*i+0];
m_lut3D[3*i+1] = m_lut3D[4*i+1];
m_lut3D[3*i+2] = m_lut3D[4*i+2];
}
}
// Copy to the destination
memcpy(lut3d, &m_lut3D[0], sizeof(float) * 3 * lut3DNumPixels);
}
void ExponentOp::writeGpuShader(std::ostream & shader,
const std::string & pixelName,
const GpuShaderDesc & shaderDesc) const
{
GpuLanguage lang = shaderDesc.getLanguage();
float zerovec[4] = { 0.0f, 0.0f, 0.0f, 0.0f };
shader << pixelName << " = pow(";
shader << "max(" << pixelName << ", " << GpuTextHalf4(zerovec, lang) << ")";
shader << ", " << GpuTextHalf4(m_exp4, lang) << ");\n";
}
void Processor::Impl::calcGpuShaderText(std::ostream & shader,
const GpuShaderDesc & shaderDesc) const
{
std::string pixelName = "out_pixel";
std::string lut3dName = "lut3d";
WriteShaderHeader(shader, pixelName, shaderDesc);
for(unsigned int i=0; i<m_gpuOpsHwPreProcess.size(); ++i)
{
m_gpuOpsHwPreProcess[i]->writeGpuShader(shader, pixelName, shaderDesc);
}
if(!m_gpuOpsCpuLatticeProcess.empty())
{
// Sample the 3D LUT.
int lut3DEdgeLen = shaderDesc.getLut3DEdgeLen();
shader << pixelName << ".rgb = ";
Write_sampleLut3D_rgb(shader, pixelName,
lut3dName, lut3DEdgeLen,
shaderDesc.getLanguage());
}
#ifdef __APPLE__
else
{
// Force a no-op sampling of the 3d lut on OSX to work around a segfault.
int lut3DEdgeLen = shaderDesc.getLut3DEdgeLen();
shader << "// OSX segfault work-around: Force a no-op sampling of the 3d lut.\n";
Write_sampleLut3D_rgb(shader, pixelName,
lut3dName, lut3DEdgeLen,
shaderDesc.getLanguage());
}
#endif // __APPLE__
for(unsigned int i=0; i<m_gpuOpsHwPostProcess.size(); ++i)
{
m_gpuOpsHwPostProcess[i]->writeGpuShader(shader, pixelName, shaderDesc);
}
WriteShaderFooter(shader, pixelName, shaderDesc);
}
