Explain why kernels are compiled with code object v4

Change-Id: Ib0502f005faa513105cf570a0013517d6c6349c3
This commit is contained in:
Konstantin Zhuravlyov
2023-02-03 21:53:59 -05:00
rodzic d9c9d39902
commit e0e5f7336b
+4
Wyświetl plik
@@ -73,6 +73,10 @@ fi
for GFXIP in $TGT_LIST ; do
OBJ_PREF=$GFXIP
OBJ_FILE="${OBJ_PREF}_${OBJ_NAME}.$SO_EXT"
# Kernels are currently compiled with code object version 4. In order to
# switch to code object version 5, kernel argument setup in
# rocprofiler/test/simple_convolution/simple_convolution needs to be updated
# according to https://llvm.org/docs/AMDGPUUsage.html#amdgpu-amdhsa-code-object-kernel-argument-metadata-map-table-v5
$BIN_DIR/clang -cl-std=CL$OCL_VER -include $INC_DIR/opencl-c.h $BITCODE_OPTS -target amdgcn-amd-amdhsa -mcpu=$GFXIP -mcode-object-version=4 $TEST_NAME.cl -o $DST_DIR/$OBJ_FILE
echo "'$OBJ_FILE' generated"
done