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

Merge remote-tracking branch 'nccl/master' into no-target-id

Browse Source
This commit is contained in:
Wenkai Du
2020-12-01 11:33:47 -05:00
parent 2e8b3a0857 920dbe5b35
commit d469947641
106 changed files with 11943 additions and 4104 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/collectives.h
+44 -56
View File
@@ -8,63 +8,60 @@
#ifndef NCCL_COLLECTIVES_H_
#define NCCL_COLLECTIVES_H_
#define FUNC_INDEX_P2P (4+NCCL_NUM_FUNCTIONS*NCCL_NUM_ALGORITHMS*NCCL_NUM_PROTOCOLS*ncclNumTypes*ncclNumOps)
#define FUNC_INDEX(coll, redop, dtype, al, pr) ((coll >= NCCL_NUM_FUNCTIONS) \
? (coll-NCCL_NUM_FUNCTIONS+NCCL_NUM_FUNCTIONS*NCCL_NUM_ALGORITHMS*NCCL_NUM_PROTOCOLS*ncclNumTypes*ncclNumOps) \
: ((((((coll)*ncclNumOps + (redop))*ncclNumTypes) + (dtype))*NCCL_NUM_ALGORITHMS+(al))*NCCL_NUM_PROTOCOLS+(pr)))
#define FUNC_INDEX_P2P 1800
#define FUNC_INDEX(func, redop, ncclType, al, pr) ((((((func)*ncclNumOps + (redop))*ncclNumTypes) + (ncclType))*NCCL_NUM_ALGORITHMS+(al))*NCCL_NUM_PROTOCOLS+(pr))
#define NCCL_COLL_NAME(coll, op, dtype) \
coll##_##op##_##dtype
#define NCCL_FUNC_NAME(func, algo, proto, redop, type) \
ncclFunction_##func##_##algo##_##proto##_##redop##_##type
#define NCCL_KERN_NAME(coll, op, dtype) \
coll##Kernel_##op##_##dtype
#define NCCL_KERN_NAME(func, algo, proto, redop, type) \
ncclKernel_##func##_##algo##_##proto##_##redop##_##type
#define NCCL_IMPL_NAME(func, algo, proto) \
nccl##func##algo##proto
/* Declare all collective operations */
#define DECL_COLL5(coll, op, dtype) \
extern __device__ __attribute__((noinline)) void NCCL_COLL_NAME(coll, op, dtype)(struct CollectiveArgs* args); \
extern __global__ void NCCL_KERN_NAME(coll, op, dtype)(struct ncclDevComm* comm); \
#define DECL5(func, algo, proto, redop, type) \
extern __device__ __attribute__((noinline)) void NCCL_FUNC_NAME(func, algo, proto, redop, type)(struct ncclWorkElem* args); \
extern __global__ void NCCL_KERN_NAME(func, algo, proto, redop, type)(struct ncclWorkElem first); \
#define DECL_COLL4(coll, op, dtype) \
DECL_COLL5(coll, op, dtype) \
DECL_COLL5(coll##LL, op, dtype) \
DECL_COLL5(coll##LL128, op, dtype)
#define DECL4(func, algo, redop, type) \
DECL5(func, algo, SIMPLE, redop, type) \
DECL5(func, algo, LL, redop, type) \
DECL5(func, algo, LL128, redop, type)
#define DECL_COLL3(coll, op, dtype) \
DECL_COLL4(coll##Ring, op, dtype) \
DECL_COLL4(coll##Tree, op, dtype) \
DECL_COLL4(coll##CollNet, op, dtype)
#define DECL3(func, redop, type) \
DECL4(func, RING, redop, type) \
DECL4(func, TREE, redop, type) \
DECL4(func, COLLNET, redop, type)
#define DECL_COLL2(coll, op) \
DECL_COLL3(coll, op, i8) \
DECL_COLL3(coll, op, u8) \
DECL_COLL3(coll, op, i32) \
DECL_COLL3(coll, op, u32) \
DECL_COLL3(coll, op, i64) \
DECL_COLL3(coll, op, u64) \
DECL_COLL3(coll, op, f16) \
DECL_COLL3(coll, op, f32) \
DECL_COLL3(coll, op, f64) \
DECL_COLL3(coll, op, b16)
#define DECL2(func, redop) \
DECL3(func, redop, int8_t) \
DECL3(func, redop, uint8_t) \
DECL3(func, redop, int32_t) \
DECL3(func, redop, uint32_t) \
DECL3(func, redop, int64_t) \
DECL3(func, redop, uint64_t) \
DECL3(func, redop, half) \
DECL3(func, redop, float) \
DECL3(func, redop, double) \
DECL3(func, redop, rccl_bfloat16)
#define DECL_COLL(coll) \
DECL_COLL2(coll, sum) \
DECL_COLL2(coll, prod) \
DECL_COLL2(coll, min) \
DECL_COLL2(coll, max)
#define DECL(func) \
DECL2(func, Sum) \
DECL2(func, Prod) \
DECL2(func, Min) \
DECL2(func, Max)
#define DECL_ALL_COLLS \
DECL_COLL2(ncclBroadcast, copy) \
DECL_COLL(ncclReduce) \
DECL_COLL2(ncclAllGather, copy) \
DECL_COLL(ncclReduceScatter) \
DECL_COLL(ncclAllReduce) \
DECL_COLL5(ncclGather, copy, i8) \
DECL_COLL5(ncclScatter, copy, i8) \
DECL_COLL5(ncclAllToAll, copy, i8) \
DECL_COLL5(ncclAllToAllv, copy, i8) \
DECL_COLL5(ncclSendRecv, copy, i8) \
#define DECL_ALL \
DECL2(Broadcast, Sum) \
DECL(Reduce) \
DECL2(AllGather, Sum) \
DECL(ReduceScatter) \
DECL(AllReduce) \
DECL5(SendRecv, RING, SIMPLE, Sum, int8_t) \
DECL_ALL_COLLS
DECL_ALL
// CHUNKSIZE must be a multiple of SLICESIZE
//#define ALLREDUCE_SLICESTEPS (NCCL_STEPS/4)
@@ -84,13 +81,4 @@ DECL_ALL_COLLS
#define REDUCE_SLICESTEPS 1
#define REDUCE_CHUNKSTEPS 1
#define SENDRECV_SLICEFACTOR 1
#define GATHER_SLICESTEPS 4
#define GATHER_CHUNKSTEPS 4
#define SCATTER_SLICESTEPS 4
#define SCATTER_CHUNKSTEPS 4
#define ALLTOALL_SLICESTEPS 4
#define ALLTOALL_CHUNKSTEPS 4
#define ALLTOALLV_SLICESTEPS 4
#define ALLTOALLV_CHUNKSTEPS 4
#endif