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_v_pipeline.h"
12 : #include "alg_template_register.h"
13 :
14 : constexpr u32 STEP_OFFSET_TWO = 2;
15 :
16 : namespace hccl {
17 0 : ReduceScatterVPipeline::ReduceScatterVPipeline(const HcclDispatcher dispatcher)
18 0 : : ReduceScatterPipeline(dispatcher) {}
19 :
20 0 : ReduceScatterVPipeline::~ReduceScatterVPipeline() {}
21 :
22 :
23 0 : HcclResult ReduceScatterVPipeline::RunIntraServer(u32 step, u64 remoteOffset)
24 : {
25 0 : for (u32 i = 1; i < intraRankSize_; i++) {
26 0 : u32 remIntraRankId = (intraRankId_ + i) % intraRankSize_;
27 0 : CHK_RET(intraLinks_[remIntraRankId]->TxAck(subStream_[i]));
28 0 : CHK_RET(intraLinks_[remIntraRankId]->RxAck(subStream_[i]));
29 0 : void* remoteMemPtr = nullptr;
30 0 : CHK_RET(intraLinks_[remIntraRankId]->GetRemoteMem(UserMemType::INPUT_MEM, &remoteMemPtr));
31 :
32 : // 本次机内待发送数据的index坐标
33 0 : u32 index = (((interRankId_ + 1 + step) % interRankSize_) * intraRankSize_ + remIntraRankId);
34 0 : Slice userSlice = slices_[index];
35 0 : u64 srcOffset = userSlice.offset;
36 :
37 0 : u64 offset = srcOffset % HCCL_MIN_SLICE_ALIGN_910B;
38 0 : DeviceMem src = DeviceMem::create(static_cast<u8 *>(usrInMem_) + srcOffset, userSlice.size);
39 0 : DeviceMem dst = DeviceMem::create(static_cast<u8 *>(remoteMemPtr) + remoteOffset + offset, userSlice.size);
40 :
41 0 : CHK_RET(HcclReduceAsync(dispatcher_, src.ptr(), userSlice.size/unitSize_, dataType_, reductionOp_,
42 : subStream_[i], dst.ptr(), intraLinks_[remIntraRankId]->GetRemoteRank(),
43 : intraLinks_[remIntraRankId]->GetLinkType(), INLINE_REDUCE_BIT));
44 :
45 0 : CHK_RET(intraLinks_[remIntraRankId]->TxDataSignal(subStream_[i]));
46 0 : CHK_RET(intraLinks_[remIntraRankId]->RxDataSignal(subStream_[i]));
47 0 : }
48 0 : return HCCL_SUCCESS;
49 : }
50 :
51 0 : HcclResult ReduceScatterVPipeline::RunInterServer(u32 step,
52 : const LINK &prevInterLink,
53 : const LINK &nextInterLink)
54 : {
55 0 : u32 dmaMemSliceNum = dmaMem_.size();
56 0 : u32 rxDMAMemSliceId = (step + 1) % dmaMemSliceNum;
57 0 : u32 txDMAMemSliceId = step % dmaMemSliceNum;
58 :
59 0 : u32 txindex = (((interRankId_ + 1 + step) % interRankSize_) * intraRankSize_ + intraRankId_);
60 0 : Slice txUserSlice = slices_[txindex];
61 0 : u64 txSliceOffset = txUserSlice.offset;
62 0 : u64 offset = txSliceOffset % HCCL_MIN_SLICE_ALIGN_910B;
63 0 : u64 rxInterOffset = rxDMAMemSliceId * blockSize_ + offset;
64 0 : void* txLocalAddr = static_cast<u8 *>(dmaMem_[txDMAMemSliceId].ptr()) + offset;
65 :
66 0 : DeviceMem srcMem = DeviceMem::create(txLocalAddr, txUserSlice.size);
67 0 : CHK_RET(senderInfo_->run(nextInterLink, rxInterOffset, srcMem, subStream_[0])); // 发 srcMem -> rxInterOffset(rem)
68 0 : HCCL_DEBUG("[ReduceScatterVPipeline][RunInterServer] local rank[%u] localOffset[%llu]tx with slice[%llu]",
69 : rankId_, rxInterOffset, txUserSlice.size);
70 :
71 0 : u32 rxindex = (((interRankId_ + 2 + step) % interRankSize_) * intraRankSize_ + intraRankId_);
72 0 : Slice rxUserSlice = slices_[rxindex];
73 0 : u64 rxSliceOffset = rxUserSlice.offset;
74 0 : u64 rxOffset = rxSliceOffset % HCCL_MIN_SLICE_ALIGN_910B;
75 0 : u64 rxMemOffset = txDMAMemSliceId * blockSize_ + rxOffset;
76 0 : void* rxLocalAddr = static_cast<u8 *>(dmaMem_[rxDMAMemSliceId].ptr()) + rxOffset;
77 :
78 0 : DeviceMem rxLocalMem = DeviceMem::create(rxLocalAddr, rxUserSlice.size);
79 0 : CHK_RET(reducerInfo_->run(dispatcher_, prevInterLink, rxMemOffset, rxLocalMem, rxLocalMem, rxLocalMem,// 收 rxMemOffset(rem) -> rxLocalMem
80 : subStream_[0]));
81 0 : return HCCL_SUCCESS;
82 0 : }
83 :
84 0 : HcclResult ReduceScatterVPipeline::CopyToScratchBuffer(u32 step)
85 : {
86 0 : u32 dmaMemSliceNum = dmaMem_.size();
87 0 : u32 dmaMemSliceId = step % dmaMemSliceNum;
88 :
89 0 : u32 index = (((interRankId_ + 1 + step) % interRankSize_) * intraRankSize_ + intraRankId_);
90 0 : Slice userslice = slices_[index];
91 0 : u64 srcOffset = userslice.offset;
92 0 : u64 offset = srcOffset % HCCL_MIN_SLICE_ALIGN_910B;
93 :
94 0 : void* srcAddr = static_cast<u8 *>(usrInMem_) + srcOffset;
95 0 : DeviceMem locSrc = DeviceMem::create(srcAddr, userslice.size);
96 0 : DeviceMem locDst = DeviceMem::create(static_cast<u8 *>(dmaMem_[dmaMemSliceId].ptr()) + offset, userslice.size);
97 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, locDst, locSrc, stream_));
98 0 : return HCCL_SUCCESS;
99 0 : }
100 :
101 0 : HcclResult ReduceScatterVPipeline::RunAsync()
102 : {
103 : // inter ring algo
104 0 : u32 prevInterRankId = (interRankId_ + 1) % interRankSize_;
105 0 : u32 nextInterRankId = (interRankId_ - 1 + interRankSize_) % interRankSize_;
106 0 : LINK prevInterLink = interLinks_[prevInterRankId];
107 0 : LINK nextInterLink = interLinks_[nextInterRankId];
108 : // 当前使用3块DMAMem buffer
109 0 : u32 dmaMemSliceNum = dmaMem_.size();
110 :
111 0 : for (u32 step = 0; step < interRankSize_; step ++) {
112 0 : u32 begin = 0;
113 0 : if (step == 0) {
114 0 : begin = 1;
115 0 : CHK_RET(CopyToScratchBuffer(step));
116 0 : CHK_RET(MainRecordSub(begin));
117 0 : CHK_RET(SubWaitMain(begin));
118 : }
119 : // // server内做SDMA的reduce
120 0 : u64 remoteOffset = (step % dmaMemSliceNum) * blockSize_;
121 0 : CHK_RET(RunIntraServer(step, remoteOffset));
122 0 : CHK_RET(SubRecordMain(begin));
123 0 : CHK_RET(MainWaitSub(begin));
124 0 : if (step < interRankSize_ - 1) {
125 : // 把下一块切片从userIn 做拷贝到CCLBuffer
126 0 : CHK_RET(CopyToScratchBuffer(step + 1));
127 : // // 全部流同步，确保SDMA执行完成
128 0 : CHK_RET(MainRecordSub(0));
129 0 : CHK_RET(SubWaitMain(0));
130 0 : CHK_RET(prevInterLink->TxAck(subStream_[0]));
131 0 : CHK_RET(nextInterLink->RxAck(subStream_[0]));
132 : // server间做RDMA的reduce，可与下一个step的SDMA并发执行
133 0 : CHK_RET(RunInterServer(step, prevInterLink, nextInterLink));
134 0 : CHK_RET(prevInterLink->PostFinAck(subStream_[0]));
135 0 : CHK_RET(nextInterLink->WaitFinAck(subStream_[0]));
136 : // inter的最后一步需要barrier确保数据发完
137 0 : if (step == interRankSize_ - STEP_OFFSET_TWO) {
138 0 : CHK_RET(ExecuteBarrier(prevInterLink, nextInterLink, subStream_[0]));
139 : }
140 : }
141 : }
142 : // 把对应的切片从CCLBuffer拷贝到userOut
143 0 : Slice userSlice = slices_[rankId_];
144 0 : u64 srcOffset = userSlice.offset % HCCL_MIN_SLICE_ALIGN_910B;
145 0 : void* locSrcAddr = static_cast<u8 *>(dmaMem_[(interRankSize_ - 1) % dmaMemSliceNum].ptr()) + srcOffset;
146 :
147 0 : DeviceMem locSrc = DeviceMem::create(locSrcAddr, userSlice.size);
148 0 : DeviceMem locDst = DeviceMem::create(static_cast<u8 *>(usrOutMem_), userSlice.size);
149 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, locDst, locSrc, stream_));
150 0 : HCCL_INFO("[ReduceScatterVPipeline][RunAsync]ReduceScatterVPipeline finished rankId[%u] ", rankId_);
151 0 : return HCCL_SUCCESS;
152 0 : }
153 :
154 : // 适配新CollExecutor接口
155 0 : HcclResult ReduceScatterVPipeline::Prepare(HcomCollOpInfo *opInfo, DeviceMem &cclBuffer, const u64 bufferSize,
156 : const std::vector<Slice> &slices, const SubCommInfo &level0CommInfo, const SubCommInfo &level1CommInfo,
157 : Stream &mainStream, std::vector<Stream> &subStream, std::vector<std::shared_ptr<LocalNotify>> ¬ifyMain,
158 : std::vector<std::shared_ptr<LocalNotify>> ¬ifySub, u64 reduceAttrBitMap)
159 : {
160 0 : reduceAttr_ = reduceAttrBitMap;
161 0 : opInfo_ = opInfo;
162 :
163 0 : unitSize_ = SIZE_TABLE[opInfo_->dataType];
164 0 : usrInMem_ = opInfo_->inputAddr;
165 0 : usrOutMem_ = opInfo_->outputAddr;
166 0 : reductionOp_ = opInfo_->reduceOp;
167 0 : dataType_ = opInfo_->dataType;
168 0 : slices_ = slices;
169 :
170 0 : stream_ = mainStream;
171 0 : subStream_ = subStream;
172 :
173 0 : intraRankSize_ = level0CommInfo.localRankSize;
174 0 : interRankSize_ = level1CommInfo.localRankSize;
175 0 : intraRankId_ = level0CommInfo.localRank;
176 0 : interRankId_ = level1CommInfo.localRank;
177 0 : rankId_ = intraRankId_ + interRankId_ * intraRankSize_;
178 :
179 0 : streamNotifyMain_ = notifyMain;
180 0 : if (streamNotifyMain_.size() < intraRankSize_) {
181 0 : HCCL_ERROR("[ReduceScatterVPipeline][Prepare]rank[%u] streamNotifyMain_ size [%u] error, is smaller than," \
182 : "intraRankSize_[%u]", rankId_, streamNotifyMain_.size(), intraRankSize_);
183 0 : return HCCL_E_INTERNAL;
184 : }
185 0 : streamNotifySub_ = notifySub;
186 0 : if (streamNotifySub_.size() < intraRankSize_) {
187 0 : HCCL_ERROR("[ReduceScatterVPipeline][Prepare]rank[%u] streamNotifySub_ size [%u] error, is smaller than," \
188 : "intraRankSize_[%u]", rankId_, streamNotifySub_.size(), intraRankSize_);
189 0 : return HCCL_E_INTERNAL;
190 : }
191 :
192 0 : intraLinks_ = level0CommInfo.links;
193 0 : interLinks_ = level1CommInfo.links;
194 :
195 : // 3级流水,使用3块DMAMem
196 0 : cclBuffer_ = cclBuffer;
197 0 : bufferSize_ = bufferSize;
198 :
199 0 : blockSize_ = (bufferSize_ / (HCCL_MIN_SLICE_ALIGN_910B * PIPELINE_DEPTH)) * HCCL_MIN_SLICE_ALIGN_910B;
200 :
201 0 : for (u32 i = 0; i < pipDepth_; i ++) {
202 0 : DeviceMem mem = DeviceMem::create(static_cast<u8 *>(cclBuffer_.ptr()) + blockSize_ * i, blockSize_);
203 0 : dmaMem_.push_back(mem);
204 0 : }
205 :
206 0 : HCCL_INFO("[ReduceScatterVPipeline][Prepare]streamNum[%u], streamNotifyMainNum[%u], streamNotifySubNum[%u]",
207 : subStream_.size(), streamNotifyMain_.size(), streamNotifySub_.size());
208 0 : HCCL_INFO("[ReduceScatterVPipeline][Prepare]interLinksNum[%u], intraLinksNum[%u]",
209 : interLinks_.size(), intraLinks_.size());
210 0 : senderInfo_.reset(new (std::nothrow) Sender(dataType_, reductionOp_, reduceAttr_));
211 0 : CHK_SMART_PTR_NULL(senderInfo_);
212 0 : reducerInfo_.reset(new (std::nothrow) Reducer(dataType_, reductionOp_, reduceAttr_));
213 0 : CHK_SMART_PTR_NULL(reducerInfo_);
214 0 : return HCCL_SUCCESS;
215 : }
216 :
217 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_REDUCESCATTER_V_PIPELINE, ReduceScatterVPipeline);
218 : } // namespace hccl
|
