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

Change MSCCL kernel signature to allow kernel arguments be preloaded via SGPR (#911)

Browse Source 
* Adding a script that will download/compile/run TransferBench/RCCL/UCX/RCCL-tests/RCCL-Unittests/hip-mpi-testsuite (#895)

Co-authored-by: Pedram Alizadeh <pmohamma@banff-pla-r27-05.pla.dcgpu>

* Only build gfx941

* demo

* fine tune malloc

* Fix merge errors

* Fix merge errors

* Disable parallel build

* Adopt --amdgpu-kernarg-preload-count

* Revert "Adding a script that will download/compile/run TransferBench/RCCL/UCX/RCCL-tests/RCCL-Unittests/hip-mpi-testsuite (#895)"

This reverts commit f5e252dddf02a41b4d1bc512f306f45f97166304.

* Revert CMake changes.

* NPKIT changes.

* Remove some license declarations.

* Address code review feedbacks on msccl_kernel_impl.h

* Update CMakeLists.txt

* Add CMake logic to check the existence of --amdgpu-kernarg-preload-count

* Fix NPKIT trace logic.

---------

Co-authored-by: Pedram Alizadeh <pmohamma@amd.com>
Co-authored-by: Pedram Alizadeh <pmohamma@banff-pla-r27-05.pla.dcgpu>
Co-authored-by: Ziyue Yang <ziyyang@microsoft.com>

[ROCm/rccl commit: 7ee5c1c28b]
This commit is contained in:
Wen-Heng (Jack) Chung
2023-10-12 20:17:08 -05:00
committed by GitHub
parent 9a0c9ba3e9
commit dfa0d98f9e
7 changed files with 133 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/collectives/device/msccl_kernel_impl.h
+11 -9
View File
@@ -130,7 +130,7 @@ for (int r = 0; r < numloops; r++) { \
template<typename T, typename RedOp, typename Proto>
__device__ __forceinline__ void mscclRunInterpreter(
struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork work) {
struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork* work) {
const int tid = threadIdx.x;
const int bid = blockIdx.x;
const int nthreads = NCCL_MAX_NTHREADS;
@@ -173,7 +173,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
break;
case 2:
dst = &mscclShmem.work;
src = &work;
src = work + blockIdx.x;
bytes = sizeof(mscclWork);
static_assert(sizeof(mscclWork) <= sizeof(uint64_t) * WARP_SIZE, "mscclWork cannot be loaded by a single warp in one insn.");
break;
@@ -192,6 +192,8 @@ __device__ __forceinline__ void mscclRunInterpreter(
if (tid == 0) ncclShmem.event_buffer_head = 0;
#endif
__synclds(); // publish shmem
if (tid == 0)
*mscclShmem.work.workFifoDone = mscclShmem.work.workFifoDoneAck;
#if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_EVENT_TIME_SYNC_CPU)
if (tid == 0) {
@@ -226,7 +228,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
}
}
__synclds(); // publish shmem
// User pointers for primitives
T* thisInput = (T*)mscclShmem.work.sendBuff;
T* thisOutput = (T*)mscclShmem.work.recvBuff;
@@ -253,7 +255,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
}
#endif
const ssize_t sizePerMscclChunk = mscclShmem.work.count / mscclShmem.work.nChunksPerLoop;
const ssize_t sizePerMscclChunk = mscclShmem.work.sizePerMscclChunk;
uint32_t maxAllowedCount = mscclShmem.work.maxAllowedCount;
#if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_EVENT_MSCCL_RUN_ENTRY)
@@ -262,7 +264,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
asm volatile ("s_getreg_b32 %0, hwreg(HW_REG_XCC_ID)" : "=s" (xcc_id));
#endif
NpKit::CollectGpuEventLDS(NPKIT_EVENT_MSCCL_RUN_ENTRY, mscclShmem.work.count*sizeof(T), 0, NPKIT_GET_GPU_TIMESTAMP());
NpKit::CollectGpuEventLDS(NPKIT_EVENT_MSCCL_RUN_ENTRY, mscclShmem.work.sizePerMscclChunk*mscclShmem.work.nChunksPerLoop, 0, NPKIT_GET_GPU_TIMESTAMP());
}
#endif
@@ -435,7 +437,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
}
#if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_EVENT_MSCCL_RUN_EXIT)
if (tid == 0) {
NpKit::CollectGpuEventLDS(NPKIT_EVENT_MSCCL_RUN_EXIT, mscclShmem.work.count*sizeof(T), 0, NPKIT_GET_GPU_TIMESTAMP());
NpKit::CollectGpuEventLDS(NPKIT_EVENT_MSCCL_RUN_EXIT, mscclShmem.work.sizePerMscclChunk*mscclShmem.work.nChunksPerLoop, 0, NPKIT_GET_GPU_TIMESTAMP());
}
#endif
#if defined(ENABLE_NPKIT)
@@ -447,13 +449,13 @@ __device__ __forceinline__ void mscclRunInterpreter(
}
#define MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, type) \
__global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, LL)(struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork work) { \
__global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, LL)(struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork* work) { \
mscclRunInterpreter<type, Func##devredop<type>, ProtoLL>(comm, algo, work); \
} \
__global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, LL128)(struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork work) { \
__global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, LL128)(struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork* work) { \
mscclRunInterpreter<type, Func##devredop<type>, ProtoLL128>(comm, algo, work); \
} \
__global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, Simple)(struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork work) { \
__global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, Simple)(struct ncclDevComm* comm, struct mscclAlgo* algo, struct mscclWork* work) { \
mscclRunInterpreter<type, Func##devredop<type>, ProtoSimple<MSCCL_CHUNKSTEPS/MSCCL_SLICESTEPS, MSCCL_SLICESTEPS>>(comm, algo, work); \
}