Xavier Hallade
9d5164d472
When storing multiple binaries in a fatbinary, ocloc stores Intel Graphics binaries for specific devices and these don't get found when the device doesn't match fully. Blender compiles one binary per generation rather than per device, to gain space and compile time. We introduce a patch at ocloc level to work around this behavior that otherwise leads to binaries not being found and recompilation being attempted. Pull Request: https://projects.blender.org/blender/blender/pulls/115750
15 lines
1.0 KiB
Diff
15 lines
1.0 KiB
Diff
diff --git a/shared/offline_compiler/source/ocloc_fatbinary.cpp b/shared/offline_compiler/source/ocloc_fatbinary.cpp
|
|
index 98a1c0e..4d9b5b0 100644
|
|
--- a/shared/offline_compiler/source/ocloc_fatbinary.cpp
|
|
+++ b/shared/offline_compiler/source/ocloc_fatbinary.cpp
|
|
@@ -286,7 +286,9 @@ int buildFatBinaryForTarget(int retVal, const std::vector<std::string> &argsCopy
|
|
productConfig = ProductConfigHelper::parseMajorMinorRevisionValue(argHelper->productConfigHelper->getProductConfigFromDeviceName(product));
|
|
}
|
|
|
|
- fatbinary.appendFileEntry(pointerSize + "." + productConfig, pCompiler->getPackedDeviceBinaryOutput());
|
|
+ // Storing binaries under the hardware prefix instead of the full architecture version number,
|
|
+ // as they would otherwise be ignored if they do not fully match that of the execution device.
|
|
+ fatbinary.appendFileEntry(pointerSize + "." + NEO::hardwarePrefix[argHelper->productConfigHelper->getProductFamilyFromDeviceName(productConfig)], pCompiler->getPackedDeviceBinaryOutput());
|
|
return retVal;
|
|
}
|