Line data Source code
1 : /**
2 : * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3 : * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 : * CANN Open Software License Agreement Version 2.0 (the "License").
5 : * Please refer to the License for details. You may not use this file except in compliance with the License.
6 : * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7 : * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8 : * See LICENSE in the root of the software repository for the full text of the License.
9 : */
10 :
11 : #include "reduce_scatter_mesh_atomic.h"
12 : #include "alg_template_register.h"
13 :
14 : namespace hccl {
15 : using namespace std;
16 :
17 1 : ReduceScatterMeshAtomic::ReduceScatterMeshAtomic(const HcclDispatcher dispatcher)
18 1 : : AlgTemplateBase(dispatcher)
19 1 : {}
20 :
21 2 : ReduceScatterMeshAtomic::~ReduceScatterMeshAtomic() {}
22 :
23 1 : HcclResult ReduceScatterMeshAtomic::Prepare(DeviceMem &inputMem, DeviceMem &outputMem, DeviceMem &scratchMem,
24 : const u64 count, const HcclDataType dataType, const Stream &stream,
25 : const HcclReduceOp reductionOp, const u32 root,
26 : const std::vector<Slice> &slices, const u64 baseOffset,
27 : const u64 reduceAttrBitMap, std::vector<Stream> &meshStreams,
28 : std::vector<std::shared_ptr<LocalNotify>> &meshSignal,
29 : std::vector<std::shared_ptr<LocalNotify>> &meshSignalAux,
30 : u32 userRank, const HcomCollOpInfo *opInfo)
31 : {
32 1 : reduceAttr_ = reduceAttrBitMap;
33 1 : userRank_ = userRank;
34 1 : meshStreams_ = meshStreams;
35 1 : meshSignalPtr_ = &meshSignal;
36 1 : meshSignalAuxPtr_ = &meshSignalAux;
37 : (void)opInfo;
38 4 : return AlgTemplateBase::Prepare(inputMem, outputMem, scratchMem, count, dataType, stream, reductionOp,
39 2 : root, slices, baseOffset);
40 : }
41 :
42 0 : HcclResult ReduceScatterMeshAtomic::RunReduceScatter(const std::vector<LINK> &links)
43 : {
44 : // 拼接所有stream
45 0 : vector<Stream> streamVct;
46 0 : streamVct.reserve(localRankSize_ - 1); // 有ranksize-1个对端，每个对端对应一条stream
47 0 : streamVct.push_back(stream_); // 增加主stream
48 0 : streamVct.insert(streamVct.end(), meshStreams_.begin(), meshStreams_.end()); // 增加从stream
49 :
50 : // 每个stream只负责一个对端的交互
51 0 : for (u32 streamIndex = 0; streamIndex < localRankSize_ - 1; streamIndex++) {
52 0 : u32 remoteRank = (streamIndex + localRank_ + 1) % localRankSize_;
53 0 : const LINK &dstLink = links[remoteRank];
54 0 : Stream &stream = streamVct[streamIndex];
55 :
56 0 : CHK_RET(dstLink->TxAck(stream));
57 0 : CHK_RET(dstLink->RxAck(stream));
58 : }
59 :
60 0 : for (u32 streamIndex = 0; streamIndex < localRankSize_ - 1; streamIndex++) {
61 0 : u32 remoteRank = (streamIndex + localRank_ + 1) % localRankSize_;
62 0 : const LINK &dstLink = links[remoteRank];
63 0 : Stream &stream = streamVct[streamIndex];
64 0 : profilerInput_.streamID = stream.id();
65 0 : profilerInput_.planeID = streamIndex - 1;
66 0 : profilerInput_.step = HCCL_EXEC_STEP_NOT_SET;
67 :
68 0 : Slice &rxSlice = slices_[localRank_];
69 0 : if (streamIndex == 0) {
70 0 : for (u32 signalIndex = 0; signalIndex < localRankSize_ - 2; signalIndex++) { // rankSize-2: stream num
71 0 : CHK_RET(LocalNotify::Wait(stream, dispatcher_, (*meshSignalPtr_)[signalIndex], profilerInput_.stage));
72 : }
73 0 : for (u32 signalIndex = 0; signalIndex < localRankSize_ - 2; signalIndex++) { // rankSize-2: stream num
74 0 : CHK_RET(LocalNotify::Post(stream, dispatcher_, (*meshSignalAuxPtr_)[signalIndex],
75 : profilerInput_.stage));
76 : }
77 : } else {
78 0 : u32 signalIndex = streamIndex - 1;
79 0 : CHK_RET(LocalNotify::Post(stream, dispatcher_, (*meshSignalPtr_)[signalIndex],
80 : profilerInput_.stage));
81 0 : CHK_RET(LocalNotify::Wait(stream, dispatcher_, (*meshSignalAuxPtr_)[signalIndex], profilerInput_.stage));
82 : }
83 :
84 0 : HCCL_DEBUG(
85 : "ReduceScatterMeshAtomic RX rank[%u] inputMemSize[%llu], rxSlice.offset[%llu], rxSlice.size[%llu] "
86 : "baseOffset_[%llu]",
87 : localRank_, inputMem_.size(), rxSlice.offset, rxSlice.size, baseOffset_);
88 0 : DeviceMem dstMem = inputMem_.range(rxSlice.offset, rxSlice.size);
89 0 : void *remoteMem = nullptr;
90 0 : CHK_RET(dstLink->GetRemoteMem(UserMemType::INPUT_MEM, &remoteMem));
91 0 : CHK_RET(HcclReduceAsync(dispatcher_, static_cast<s8 *>(remoteMem) + baseOffset_ + rxSlice.offset,
92 : dstMem.size() / SIZE_TABLE[dataType_], dataType_, reductionOp_, stream,
93 : dstMem.ptr(), dstLink->GetRemoteRank(), dstLink->GetLinkType(), INLINE_REDUCE_BIT));
94 :
95 0 : CHK_RET(dstLink->TxDataSignal(stream));
96 0 : CHK_RET(dstLink->RxDataSignal(stream));
97 0 : }
98 : // 添加空task,保证执行时不乱序
99 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, streamVct[0], dispatcher_));
100 0 : return HCCL_SUCCESS;
101 0 : }
102 :
103 0 : HcclResult ReduceScatterMeshAtomic::RunAsync(const u32 rank, const u32 rankSize, const std::vector<LINK> &links)
104 : {
105 0 : CHK_SMART_PTR_NULL(dispatcher_);
106 0 : CHK_PTR_NULL(stream_.ptr());
107 0 : HCCL_INFO("ReduceScatterMeshAtomic run: rank[%u] totalrank[%u] inputMem[%p] outputMem[%p] count[%llu]", rank,
108 : rankSize, inputMem_.ptr(), outputMem_.ptr(), count_);
109 :
110 0 : localRank_ = rank;
111 0 : localRankSize_ = rankSize;
112 :
113 0 : if (localRankSize_ == 1) {
114 0 : if (inputMem_ != outputMem_) {
115 0 : return HcclD2DMemcpyAsync(dispatcher_, outputMem_, inputMem_, stream_);
116 : }
117 0 : return HCCL_SUCCESS;
118 : }
119 :
120 0 : if (links.size() < rankSize) {
121 0 : HCCL_ERROR("[ReduceScatterMeshAtomic][RunAsync]rank[%u] linksize[%zu] error", rank, links.size());
122 0 : return HCCL_E_INTERNAL;
123 : }
124 0 : CHK_RET(MemSlice());
125 :
126 0 : for (u32 streamIndex = 0; streamIndex < rankSize - 2; streamIndex++) { // rankSize-2: stream num
127 0 : HCCL_DEBUG("rank[%u] streamindex[%u] wait signalaux[%p]",
128 : rank, streamIndex, (*meshSignalAuxPtr_)[streamIndex]->ptr());
129 0 : CHK_RET(LocalNotify::Wait(meshStreams_[streamIndex], dispatcher_, (*meshSignalAuxPtr_)[streamIndex],
130 : profilerInput_.stage));
131 : }
132 0 : for (u32 streamIndex = 0; streamIndex < rankSize - 2; streamIndex++) { // rankSize-2: stream num
133 0 : CHK_RET(LocalNotify::Post(stream_, dispatcher_, (*meshSignalAuxPtr_)[streamIndex],
134 : profilerInput_.stage));
135 : }
136 :
137 0 : CHK_RET(RunReduceScatter(links));
138 :
139 0 : for (u32 streamIndex = 0; streamIndex < rankSize - 2; streamIndex++) { // rankSize - 2 stream num
140 0 : HCCL_DEBUG("rank[%u] streamindex[%u] wait signal[%p] ",
141 : rank, streamIndex, (*meshSignalPtr_)[streamIndex]->ptr());
142 0 : CHK_RET(LocalNotify::Wait(stream_, dispatcher_, (*meshSignalPtr_)[streamIndex], profilerInput_.stage));
143 0 : CHK_RET(LocalNotify::Post(meshStreams_[streamIndex], dispatcher_, (*meshSignalPtr_)[streamIndex],
144 : profilerInput_.stage));
145 : }
146 :
147 : // 添加空task,保证执行时不乱序
148 0 : CHK_RET(AlgTemplateBase::ExecEmptyTask(inputMem_, outputMem_, stream_, dispatcher_));
149 :
150 0 : if (inputMem_ != outputMem_) {
151 0 : DeviceMem src = inputMem_.range(slices_[localRank_].offset, slices_[localRank_].size);
152 0 : HCCL_DEBUG("rank[%u] copy result from to output[%p] offset[%llu] size[%llu] ", localRank_, outputMem_.ptr(),
153 : slices_[localRank_].offset, slices_[localRank_].size);
154 0 : HcclResult ret = HcclD2DMemcpyAsync(dispatcher_, outputMem_, src, stream_);
155 0 : CHK_PRT_RET(ret != HCCL_SUCCESS,
156 : HCCL_ERROR("[ReduceScatterMeshAtomic][RunAsync]rank[%u] memcpy async from mem[%p] "
157 : "to ouputmem[%p] failed",
158 : rank, src.ptr(), outputMem_.ptr()),
159 : ret);
160 0 : }
161 :
162 0 : HCCL_INFO("ReduceScatterMeshAtomic finished: rank[%u]", rank);
163 0 : return HCCL_SUCCESS;
164 : }
165 :
166 0 : HcclResult ReduceScatterMeshAtomic::MemSlice()
167 : {
168 0 : u32 unitSize = SIZE_TABLE[dataType_];
169 0 : if (unitSize == 0) {
170 0 : HCCL_ERROR("[ReduceScatterMeshAtomic][RunAsync]rank[%u] unit data size is zero", localRank_);
171 0 : return HCCL_E_INTERNAL;
172 : }
173 :
174 0 : if (HcclCheckLogLevel(DLOG_DEBUG)) {
175 0 : for (size_t i = 0; i < slices_.size(); i++) {
176 0 : HCCL_DEBUG("[ReduceScatterMeshAtomic] rank[%u] index[%zu] size[%llu] offset[%llu]", localRank_, i,
177 : slices_[i].size, slices_[i].offset);
178 : }
179 0 : HCCL_DEBUG("[ReduceScatterMeshAtomic] localRankSize[%u]", localRankSize_);
180 : }
181 :
182 0 : if (slices_.size() == 0) {
183 0 : slices_.resize(localRankSize_);
184 0 : u64 sliceSize = count_ * unitSize;
185 0 : for (u32 i = 0; i < localRankSize_; i++) {
186 0 : slices_[i].size = sliceSize;
187 0 : slices_[i].offset = (i * sliceSize);
188 : }
189 : }
190 :
191 0 : scratchSlices_.resize(localRankSize_);
192 0 : for (u32 i = 0; i < localRankSize_; i++) {
193 0 : scratchSlices_[i].size = slices_[i].size;
194 0 : scratchSlices_[i].offset = (scratchMem_.size() < inputMem_.size()) ? 0 : slices_[i].offset;
195 : }
196 :
197 0 : if (HcclCheckLogLevel(DLOG_DEBUG)) {
198 0 : for (size_t i = 0; i < slices_.size(); i++) {
199 0 : HCCL_DEBUG("ReduceScatterMeshAtomic rank[%u] index[%zu] " \
200 : "size[%llu] offset[%llu] scratch_size[%llu] scratch_offset[%llu]",
201 : localRank_, i, slices_[i].size, slices_[i].offset, scratchSlices_[i].size, scratchSlices_[i].offset);
202 : }
203 : }
204 :
205 0 : return HCCL_SUCCESS;
206 : }
207 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_REDUCESCATTER_MESH_ATOMIC, ReduceScatterMeshAtomic);
208 : }
|
