Perf: Add NO_COMPRESS option to disable slow offload-compress
This commit is contained in:
@@ -1217,12 +1217,12 @@ if (HAVE_PARALLEL_JOBS)
|
|||||||
target_compile_options(rccl PRIVATE -parallel-jobs=12)
|
target_compile_options(rccl PRIVATE -parallel-jobs=12)
|
||||||
endif()
|
endif()
|
||||||
|
|
||||||
if (ROCM_VERSION VERSION_GREATER_EQUAL "60200")
|
if (ROCM_VERSION VERSION_GREATER_EQUAL "60200" AND NOT NO_COMPRESS)
|
||||||
target_compile_options(rccl PRIVATE --offload-compress) # Compress GPU code at compile time.
|
target_compile_options(rccl PRIVATE --offload-compress) # Compress GPU code at compile time.
|
||||||
target_link_libraries(rccl PRIVATE --offload-compress) # Compress GPU code at link time.
|
target_link_libraries(rccl PRIVATE --offload-compress) # Compress GPU code at link time.
|
||||||
message(STATUS "--offload-compress enabled - ROCm version >= 6.2.0")
|
message(STATUS "--offload-compress enabled - ROCm version >= 6.2.0")
|
||||||
else()
|
else()
|
||||||
message(STATUS "--offload-compress disabled - ROCm version < 6.2.0")
|
message(STATUS "--offload-compress disabled (ROCM < 6.2.0 or NO_COMPRESS=ON)")
|
||||||
endif()
|
endif()
|
||||||
|
|
||||||
target_compile_options(rccl PRIVATE -Werror=uninitialized)
|
target_compile_options(rccl PRIVATE -Werror=uninitialized)
|
||||||
|
|||||||
Reference in New Issue
Block a user