Enable Ll128 on gfx90a (#823)

This commit is contained in:
Wenkai Du
2023-07-26 11:44:15 -07:00
committed by GitHub
parent 1c1ec096e2
commit 420f8af6a0
+3 -3
View File
@@ -28,7 +28,7 @@
{ __atomic_store_n((DST), (SRC), __ATOMIC_SEQ_CST); }
#endif
#ifdef ENABLE_LL128
#ifdef __gfx90a__
#define NCCL_FUNC5(func, algo, devredop, type, nullify) \
MACRO_IF(nullify, nullptr, NCCL_FUNC_NAME(func, algo, LL, devredop, type)), \
MACRO_IF(nullify, nullptr, NCCL_FUNC_NAME(func, algo, LL128, devredop, type)), \
@@ -549,7 +549,7 @@ __forceinline__ __device__ void ncclKernel(
#ifdef USE_INDIRECT_FUNCTION_CALL
ncclFuncs[ncclShmem.work.header.funcIndex]();
#else
#ifdef ENABLE_LL128
#ifdef __gfx90a__
NCCL_CALL_FUNCTIONS<1>(ncclShmem.work.header.funcIndex);
#else
NCCL_CALL_FUNCTIONS<0>(ncclShmem.work.header.funcIndex);
@@ -618,7 +618,7 @@ __device__ __attribute__((noinline)) void NCCL_FUNC_NAME(func, algo, proto, dev
#endif
// Only generate inline kernels for LL
#ifdef ENABLE_LL128
#ifdef __gfx90a__
#define IMPL_COLL4(func, algo, devredop, type) \
IMPL_COLL_FUNC(func, algo, LL, devredop, type) \
IMPL_COLL_FUNC(func, algo, LL128, devredop, type) \