AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Replace __HIP_PLATFORM_HCC__ with __HIP_PLATFORM_AMD__ (#1154)

Browse Source
This commit is contained in:
Wenkai Du
2024-04-25 07:19:18 -07:00
committed by GitHub
parent dcc75797a1
commit 9e0c9b4ed8
19 changed files with 41 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/device/msccl_kernel_impl.h
+1 -1
View File
@@ -48,7 +48,7 @@ extern __shared__ struct mscclShmemData mscclShmem;
#endif
inline __device__ static void barrier(int nthreads) {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
assert(nthreads == NCCL_MAX_NTHREADS);
__asm__ __volatile__("s_waitcnt vmcnt(0) lgkmcnt(0)\ns_barrier");
#else
src/device/prims_ll.h
+7 -7
View File
@@ -81,7 +81,7 @@ private:
uint64_t* barrier_next;
inline __device__ void barrier() {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
if (nthreads != WARP_SIZE)
barrier_by_group();
#else
@@ -159,7 +159,7 @@ private:
}
#endif
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
union ncclLLFifoLine i4;
do {
i4.v[0] = LL_LOAD(src->v);
@@ -197,7 +197,7 @@ private:
for (int i=BeginIx; i < MaxRecv; i++) {
if (i < fan.nrecv()) {
union ncclLLFifoLine* src = recvPtr(i) + offset;
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
line[i].v[0] = LL_LOAD(src->v);
line[i].v[1] = LL_LOAD(src->v+1);
#else
@@ -219,7 +219,7 @@ private:
#endif
do {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
line[i].v[0] = LL_LOAD(src->v);
line[i].v[1] = LL_LOAD(src->v+1);
#else
@@ -243,7 +243,7 @@ private:
}
__device__ void storeLL(union ncclLLFifoLine* dst, uint64_t val, uint32_t flag) {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
union ncclLLFifoLine i4;
i4.data1 = val & 0xffffffff;
i4.flag1 = flag;
@@ -267,7 +267,7 @@ private:
uint32_t u4;
uint64_t u8;
};
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
if(sizeof(U) == 1)
u1 = LL_LOAD((uint8_t*)src);
else if(sizeof(U) == 2)
@@ -299,7 +299,7 @@ private:
uint64_t u8;
};
elt = val;
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
if(sizeof(U) == 1)
LL_STORE(u1, (uint8_t*)dst);
else if(sizeof(U) == 2)
src/device/prims_ll128.h
+1 -1
View File
@@ -74,7 +74,7 @@ private:
#endif
inline __device__ void barrier() {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
if (nthreads != WARP_SIZE)
barrier_by_group();
#else