Matt Arsenault
96d4140609
There are a couple problems with this. First, llvm-dis is an unstable llvm development tool and 3rd party users should generally not rely on it. The text format is unstable, and the regex here isn't even explicitly looking for the target triple field, so it could accidentally find something else. Second, picking the target to compile based on the library you are linking is a fundamentally backwards decision. The target you're compiling for changes the library you would want to link. The device libraries are only ever compiled with amdgcn-amd-amdhsa. If we had a second triple, this should be explicitly building for any it cares about. Change-Id: I3bae8398f60f78df61ab2177aa9e83f47ec6dea4
OVERVIEW
This directory contains the CMakeLists.txt for automatically generating the ASCII code object file, "opencl_blit_objects.cpp", which contains the blobs of the code object of the Image BLIT kernels for the devices supported on ROCm. The blobs are loaded by the image library and required to update whenever a new device is introduced.
ADD NEW DEVICE
To add a new supported device, the following steps are required:
- Declare an extern variable of the device XXX, by adding the line of "extern uint32_t ocl_blit_object_gfxNNN[];" in "blit_kernel.cpp"
- Update the BlitKernel::GetPatchedBlitObject() function to support the device by assigning "blit_code_object" to "ocl_blit_object_gfxNNN[]"
- Add the gfxNNN to the TARGET_DEVICES list in CMakeLists.txt
- If the new device requires XNACK, add it to the XNACK_DEVS list in CMakeLists.txt
- Rebuild the image library
REQUIREMENT
In order to create the code object file, the bitcodes of the kernels are generated by the compiler and the following bitcode libraries are required,
opencl.amdgcn.bc ocml.amdgcn.bc irif.amdgcn.bc oclc_correctly_rounded_sqrt_off.amdgcn.bc oclc_daz_opt_on.amdgcn.bc oclc_finite_only_off.amdgcn.bc oclc_isa_version_.amdgcn.bc oclc_unsafe_math_off.amdgcn.bc
where is the gfxip number of the GPU. The directory contains the bitcode libraries is specified in a CMake varaible.
There are several variables are required for CMake to build the code object file. All of them have default values, and defined as following:
OPENCL_DIR - the location of installed OpenCL
(Default: /opt/rocm/opencl)
BITCODE_DIR - the directory contains the bitcode library
(Default: ${OPENCL_DIR}/lib/x86_64/bitcode)
LLVM_DIR - the directory contains the clang, llvm-link and llvm-dis
executables
(Default: ${PROJECT_BUILD_DIR}/../lightning/bin)
TARGET_DEVICES - list of gpu types for kernel builds (eg. "gfx900;gfx902") (Default: "gfx900;gfx902;gfx904")
STEPS TO BUILD
$ make build $ cd build $ cmake -D${OPENCL_DIR} -D${BITCODE_DIR} -D${LLVM_DIR} -D${TARGET_DEVICES} .. $ make opencl_blit_objects.cpp