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

removed gfx940 and gfx941 (#1606)

Browse Source 
* removed gfx940 and gfx941

* removed gfx940 and gfx941

* Update "gfx94" to "gfx942" in init.cc

* Updated remaining "gfx94" updates to "gfx942"

* Update filenames and variables from gfx940 to gfx942

---------

Co-authored-by: akolliasAMD <akollias@amd.com>

[ROCm/rccl commit: 6505639cf4]
This commit is contained in:
corey-derochie-amd
2025-03-20 09:34:53 -06:00
committed by GitHub
parent a95b2c9fb7
commit e95578ef4c
30 changed files with 74 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/device/all_gather.h
+1 -1
View File
@@ -11,7 +11,7 @@
namespace {
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
projects/rccl/src/device/all_reduce.h
+3 -3
View File
@@ -15,7 +15,7 @@
namespace {
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
@@ -209,7 +209,7 @@ namespace {
}
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runTreeUpDown(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runTreeUpDown(int tid, int nthreads, struct ncclDevWorkColl* work) {
@@ -357,7 +357,7 @@ namespace {
}
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runTreeSplit(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runTreeSplit(int tid, int nthreads, struct ncclDevWorkColl* work) {
projects/rccl/src/device/alltoall_pivot.h
+1 -1
View File
@@ -10,7 +10,7 @@
namespace {
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
projects/rccl/src/device/broadcast.h
+1 -1
View File
@@ -10,7 +10,7 @@
namespace {
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
projects/rccl/src/device/msccl_kernel_impl.h
+1 -1
View File
@@ -152,7 +152,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
int npKitCtxIdx = bid;
int xcc_id = 0;
if (tid == 0) {
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile ("s_getreg_b32 %0, hwreg(HW_REG_XCC_ID)" : "=s" (xcc_id));
#endif
}
projects/rccl/src/device/onerank.cu
+1 -1
View File
@@ -11,7 +11,7 @@
#include "common.h"
#include <cuda_runtime.h>
#if defined(__gfx908__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#if defined(__gfx908__) || defined(__gfx942__) || defined(__gfx950__)
#define COLL_UNROLL 2
#else
#define COLL_UNROLL 4
projects/rccl/src/device/op128.h
+1 -1
View File
@@ -124,7 +124,7 @@ union alignas(16) BytePack<16> {
uint32_t u32[4];
uint64_t u64[2];
ulong2 ul2, native;
#if !defined(USE_INDIRECT_FUNCTION_CALL) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#if !defined(USE_INDIRECT_FUNCTION_CALL) || defined(__gfx942__) || defined(__gfx950__)
inline __device__ BytePack<16>& operator=(BytePack<16> other) {
u64[0] = other.u64[0];
u64[1] = other.u64[1];
projects/rccl/src/device/primitives.h
+1 -1
View File
@@ -15,7 +15,7 @@
#define NCCL_SPINS_BEFORE_CHECK_ABORT 1000000
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#if defined(__gfx942__) || defined(__gfx950__)
#define __THREAD_FENCE __threadfence_block()
#else
#define __THREAD_FENCE __threadfence()
projects/rccl/src/device/reduce.h
+1 -1
View File
@@ -11,7 +11,7 @@
namespace {
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
projects/rccl/src/device/reduce_scatter.h
+2 -2
View File
@@ -11,7 +11,7 @@
namespace {
template<typename T, typename RedOp, typename Proto>
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
#else
__device__ __attribute__((noinline)) void runRing(int tid, int nthreads, struct ncclDevWorkColl* work) {
@@ -165,7 +165,7 @@ struct RunWorkColl<ncclFuncReduceScatter, T, RedOp, NCCL_ALGO_NVLS, NCCL_PROTO_S
size_t offset;
int nelem;
/* if we are direct NVLS, we only need to allocate 1 warp to scatter for sync;
/* if we are direct NVLS, we only need to allocate 1 warp to scatter for sync;
* if not, based on #ranks, we allocate 7 or 5 warps to reduce to saturate bandwidth
* and the rest are allocated to scatter. */
const int nThreadsReduce = work->regUsed ? (NCCL_MAX_NTHREADS - WARP_SIZE) : (nranks <= 6 ? 7 * WARP_SIZE : 5 * WARP_SIZE);
projects/rccl/src/device/sendrecv.h
+5 -5
View File
@@ -20,7 +20,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
__device__ void runSend(int tid, int tn, int group, struct ncclDevWorkP2p* work) {
size_t bytes = work->sendBytes;
int chunkSize = u32fp8Decode(work->sendChunkSize_u32fp8);
#if defined(ENABLE_NPKIT)
bool isNpKitThread = (tid == 0);
int npKitCtxIdx = blockIdx.x + group;
@@ -132,7 +132,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
#endif
}
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__) && !defined(__gfx950__)
#if defined(USE_INDIRECT_FUNCTION_CALL) && !defined(__gfx942__) && !defined(__gfx950__)
__device__ void run() {
#else
__device__ __attribute__((noinline)) void run() {
@@ -241,7 +241,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
}
if (isCopy) {
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#if defined(__gfx942__) || defined(__gfx950__)
reduceCopy<COLL_UNROLL*2, RedOp, T, 0,1,1, 0,1,1, /*PreOpSrcs=*/0>
(subtid, subtn, 0, nullptr, false, 1, &work->sendAddr, 1, &work->recvAddr, (ssize_t)work->sendBytes);
#else
@@ -254,7 +254,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
} else {
#if defined(__gfx90a__)
runSend<ProtoSimple<1,1,8>>(subtid, subtn, group, work);
#elif defined(__gfx908__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#elif defined(__gfx908__) || defined(__gfx942__) || defined(__gfx950__)
runSend<ProtoSimple<1,1,4>>(subtid, subtn, group, work);
#else
runSend<ProtoSimple<1,1>>(subtid, subtn, group, work);
@@ -266,7 +266,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
} else {
#if defined(__gfx90a__)
runRecv<ProtoSimple<1,1,8>>(subtid, subtn, group, work);
#elif defined(__gfx908__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
#elif defined(__gfx908__) || defined(__gfx942__) || defined(__gfx950__)
runRecv<ProtoSimple<1,1,4>>(subtid, subtn, group, work);
#else
runRecv<ProtoSimple<1,1>>(subtid, subtn, group, work);