1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
|
/*************************************************************************
* Copyright (c) 2015-2020, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#include "devcomm.h"
#include "primitives.h"
#include "collectives.h"
template<class FUNC, typename T, int UNROLL>
class ncclFunction<ncclFuncAllGather, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE, FUNC, T, UNROLL> {
public:
__device__ void run(struct ncclWorkElem* args) {
const int tid = threadIdx.x;
const int nthreads = args->nThreads-WARP_SIZE;
const int bid = args->coll.bid;
const int nChannels = args->coll.nChannels;
struct ncclDevComm* comm = args->comm;
struct ncclChannel* channel = comm->channels+blockIdx.x;
struct ncclRing* ring = &channel->ring;
const int stepSize = comm->buffSizes[NCCL_PROTO_SIMPLE] / (sizeof(T)*NCCL_STEPS);
const int chunkSize = stepSize * ALLGATHER_CHUNKSTEPS;
const int nranks = comm->nRanks;
const ssize_t loopSize = nChannels*(ssize_t)chunkSize;
const ssize_t size = args->coll.count;
// Compute pointers
const T * __restrict__ thisInput = (const T*)args->sendbuff;
T * __restrict__ thisOutput = (T*)args->recvbuff;
ncclPrimitives<UNROLL, ALLGATHER_CHUNKSTEPS/ALLGATHER_SLICESTEPS, ALLGATHER_SLICESTEPS, T, 1, 1, 1, FUNC>
prims(tid, nthreads, &ring->prev, &ring->next, thisOutput, stepSize, channel, comm, ncclShmem->ptrs, 0);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
int realChunkSize = min(chunkSize, DIVUP(size-gridOffset,nChannels));
ALIGN_SIZE(realChunkSize, nthreads*sizeof(uint64_t)/sizeof(T));
ssize_t chunkOffset = gridOffset + bid*realChunkSize;
/////////////// begin AllGather steps ///////////////
ssize_t offset;
int nelem = min(realChunkSize, size-chunkOffset);
int rankDest;
// step 0: push data to next GPU
rankDest = ring->devUserRanks[0];
offset = chunkOffset + rankDest * size;
if (thisInput + chunkOffset == thisOutput + offset) { // In place
prims.directSend(thisInput+chunkOffset, offset, nelem);
} else {
prims.directCopySend(thisInput+chunkOffset, thisOutput+offset, offset, nelem);
}
// k-2 steps: copy to next GPU
for (int j=1; j<nranks-1; ++j) {
rankDest = ring->devUserRanks[nranks-j];
offset = chunkOffset + rankDest * size;
prims.directRecvCopySend(thisOutput+offset, offset, nelem);
}
// Make final copy from buffer to dest.
rankDest = ring->devUserRanks[1];
offset = chunkOffset + rankDest * size;
// Final wait/copy.
prims.directRecv(thisOutput+offset, offset, nelem);
}
}
};
template<class FUNC, typename T, int UNROLL>
class ncclFunction<ncclFuncAllGather, NCCL_ALGO_RING, NCCL_PROTO_LL, FUNC, T, UNROLL> {
public:
__device__ void run(struct ncclWorkElem* args) {
const int tid = threadIdx.x;
const int nthreads = args->nThreads;
const int bid = args->coll.bid;
const int nChannels = args->coll.nChannels;
struct ncclDevComm* comm = args->comm;
struct ncclChannel* channel = comm->channels+blockIdx.x;
struct ncclRing* ring = &channel->ring;
const int stepLines = comm->buffSizes[NCCL_PROTO_LL] / (sizeof(union ncclLLFifoLine)*NCCL_STEPS);
ssize_t chunkSize = stepLines * sizeof(uint64_t) / sizeof(T);
const int nranks = comm->nRanks;
const ssize_t loopSize = nChannels*chunkSize;
const ssize_t size = args->coll.count;
ncclLLPrimitives<T, FUNC, 1, 1> LLprims(tid, nthreads, &ring->prev, &ring->next, stepLines, channel, comm);
// Compute pointers
const T * __restrict__ thisInput = (const T*)args->sendbuff;
T * __restrict__ thisOutput = (T*)args->recvbuff;
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
if (size-gridOffset < loopSize) {
chunkSize = args->coll.lastChunkSize;
}
ssize_t chunkOffset = gridOffset + bid*chunkSize;
/////////////// begin AllGather steps ///////////////
ssize_t offset;
int nelem = min(chunkSize, size-chunkOffset);
int rankDest;
// step 0: push data to next GPU
rankDest = ring->devUserRanks[0];
offset = chunkOffset + rankDest * size;
if (thisInput + chunkOffset == thisOutput + offset) { // In place
LLprims.send(thisInput+chunkOffset, nelem);
} else {
LLprims.copySend(thisInput+chunkOffset, thisOutput+offset, nelem);
}
// k-2 steps: copy to next GPU
for (int j=1; j<nranks-1; ++j) {
rankDest = ring->devUserRanks[nranks-j];
offset = chunkOffset + rankDest * size;
LLprims.recvCopySend(thisOutput+offset, nelem);
}
// step k-1: final store
rankDest = ring->devUserRanks[1];
offset = chunkOffset + rankDest * size;
LLprims.recv(thisOutput+offset, nelem);
}
}
};
#include "prims_ll128.h"
template<class FUNC, typename T, int UNROLL>
class ncclFunction<ncclFuncAllGather, NCCL_ALGO_RING, NCCL_PROTO_LL128, FUNC, T, UNROLL> {
public:
__device__ void run(struct ncclWorkElem* args) {
const int tid = threadIdx.x;
const int nthreads = args->nThreads;
const int bid = args->coll.bid;
const int nChannels = args->coll.nChannels;
struct ncclDevComm* comm = args->comm;
struct ncclChannel* channel = comm->channels+blockIdx.x;
struct ncclRing* ring = &channel->ring;
const int stepSize = comm->buffSizes[NCCL_PROTO_LL128] / (sizeof(uint64_t)*NCCL_STEPS);
ssize_t chunkSize = stepSize*NCCL_LL128_DATAELEMS*sizeof(uint64_t) / (NCCL_LL128_LINEELEMS*sizeof(T));
// We should not need the final /2 but it makes performance much, much smoother. Might be a bug somewhere.
const ssize_t minChunkSize = (NCCL_LL128_SHMEM_ELEMS_PER_THREAD*nthreads*NCCL_LL128_DATAELEMS*sizeof(uint64_t))/(NCCL_LL128_LINEELEMS*sizeof(T))/2;
const int nranks = comm->nRanks;
const ssize_t loopSize = nChannels*chunkSize;
const ssize_t size = args->coll.count;
ncclLL128Primitives<T, FUNC, 1, 1> LLprims(tid, nthreads, &ring->prev, &ring->next, stepSize, channel, comm);
// Compute pointers
const T * __restrict__ thisInput = (const T*)args->sendbuff;
T * __restrict__ thisOutput = (T*)args->recvbuff;
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
chunkSize = min(DIVUP(size-gridOffset, nChannels*minChunkSize)*minChunkSize, chunkSize);
ssize_t chunkOffset = gridOffset + bid*chunkSize;
/////////////// begin AllGather steps ///////////////
ssize_t offset;
int nelem = min(chunkSize, size-chunkOffset);
int rankDest;
// step 0: push data to next GPU
rankDest = ring->devUserRanks[0];
offset = chunkOffset + rankDest * size;
if (thisInput + chunkOffset == thisOutput + offset) { // In place
LLprims.send(thisInput+chunkOffset, nelem);
} else {
LLprims.copySend(thisInput+chunkOffset, thisOutput+offset, nelem);
}
// k-2 steps: copy to next GPU
for (int j=1; j<nranks-1; ++j) {
rankDest = ring->devUserRanks[nranks-j];
offset = chunkOffset + rankDest * size;
LLprims.recvCopySend(thisOutput+offset, nelem);
}
// step k-1: final store
rankDest = ring->devUserRanks[1];
offset = chunkOffset + rankDest * size;
LLprims.recv(thisOutput+offset, nelem);
}
}
};
template<int PROTO, class FUNC, typename T, int UNROLL>
class ncclFunction<ncclFuncAllGather, NCCL_ALGO_TREE, PROTO, FUNC, T, UNROLL> {
public:
__device__ void run(struct ncclWorkElem* args) {}
};
template<int PROTO, class FUNC, typename T, int UNROLL>
class ncclFunction<ncclFuncAllGather, NCCL_ALGO_COLLNET, PROTO, FUNC, T, UNROLL> {
public:
__device__ void run(struct ncclWorkElem* args) {}
};
|
