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 "scatter_ring_direct.h"
12 : #include "alg_template_register.h"
13 :
14 : namespace hccl {
15 0 : ScatterRingDirect::ScatterRingDirect(const HcclDispatcher dispatcher)
16 0 : : AlgTemplateBase(dispatcher)
17 : {
18 0 : }
19 :
20 0 : ScatterRingDirect::~ScatterRingDirect()
21 : {
22 0 : }
23 :
24 0 : HcclResult ScatterRingDirect::Prepare(HcomCollOpInfo *opInfo, const u32 userRank,
25 : const std::vector<u32> &ringsOrders, const std::vector<Slice> &userMemInputSlices)
26 : {
27 0 : opInfo_ = opInfo;
28 0 : userRank_ = userRank;
29 0 : ringsOrder_ = ringsOrders;
30 0 : userMemInputSlices_ = userMemInputSlices;
31 0 : return HCCL_SUCCESS;
32 : }
33 :
34 : // reduce scatter ring direct算法的函数入口
35 0 : HcclResult ScatterRingDirect::RunAsync(const u32 rank, const u32 rankSize, const std::vector<LINK> &links)
36 : {
37 : // 基本的检查
38 0 : CHK_RET(CheckParameters(rank, rankSize, links));
39 :
40 : // 判断rank_size == 1, 若inputMem_ != outputMem_，才需要搬运
41 0 : if (rankSize == 1) {
42 0 : CHK_RET(OneRankMemcpy());
43 0 : return HCCL_SUCCESS;
44 : }
45 : // 收集邻居信息
46 0 : CHK_RET(GetInitializedNeighborLinks(rank, rankSize, links));
47 : // 填充slice_
48 0 : CHK_RET(SetSlices(rank, rankSize));
49 :
50 : // 运行scatter, ring算法
51 0 : CHK_RET(RunScatter(rank, rankSize));
52 :
53 0 : if (barrierSwitchOn_) {
54 : // 执行barrier，保证数据收发完成
55 0 : CHK_RET(ExecuteBarrier(leftLink_, rightLink_));
56 : }
57 :
58 0 : HCCL_INFO("ScatterRingDirect finished: rank[%u]", rank);
59 0 : return HCCL_SUCCESS;
60 : }
61 :
62 0 : HcclResult ScatterRingDirect::CheckParameters(const u32 rank, const u32 rankSize,
63 : const std::vector<LINK> &links)
64 : {
65 0 : CHK_PTR_NULL(opInfo_);
66 0 : CHK_RET(CheckConcurrentDirectParameters(rank, rankSize, links));
67 : // 判断ringsOrder数量是否正确
68 0 : CHK_PRT_RET(ringsOrder_.size() != rankSize,
69 : HCCL_ERROR("[ScatterRingDirect] ringsOrder size[%u] is not equal to rank size[%u]",
70 : ringsOrder_.size(), rankSize),
71 : HCCL_E_PARA);
72 : // 判断userMemInputSlices数量是否正确
73 0 : CHK_PRT_RET(userMemInputSlices_.size() != rankSize,
74 : HCCL_ERROR("[ScatterRingDirect] userMemInputSlices size[%u] is not equal to rank size[%u]",
75 : userMemInputSlices_.size(), rankSize),
76 : HCCL_E_PARA);
77 0 : HCCL_INFO("ScatterRingDirect CheckParameters success");
78 0 : return HCCL_SUCCESS;
79 : }
80 :
81 0 : HcclResult ScatterRingDirect::OneRankMemcpy()
82 : {
83 0 : const Slice &srcSlice = userMemInputSlices_[0];
84 0 : const Slice &dstSlice = slices_[0];
85 0 : DeviceMem src;
86 0 : DeviceMem dst;
87 0 : if (opInfo_->inputAddr == nullptr) {
88 0 : src = inputMem_.range(srcSlice.offset, srcSlice.size);
89 : }
90 : else {
91 0 : src = DeviceMem::create(static_cast<u8 *>(opInfo_->inputAddr) + srcSlice.offset, srcSlice.size);
92 : }
93 0 : if (opInfo_->outputAddr != nullptr) {
94 : // opInfo_->outputAddr != nullptr指示要将输出发送至user output
95 0 : u64 stepOffset = slices_[ringsOrder_[0]].offset;
96 0 : HCCL_DEBUG("Memcpy operation: stream[main], rank[%u] starts to rcv offset[%llu], size[%llu] at userMemOut_",
97 : userRank_, stepOffset, dstSlice.size);
98 0 : dst = DeviceMem::create(static_cast<u8 *>(opInfo_->outputAddr) + stepOffset, dstSlice.size);
99 : } else {
100 : // opInfo_->outputAddr == nullptr指示要将输出发送至CCL buffer
101 0 : HCCL_DEBUG("Memcpy operation: stream[main], rank[%u] starts to rcv offset[%llu], size[%llu] at outputMem_",
102 : userRank_, dstSlice.offset, dstSlice.size);
103 0 : dst = outputMem_.range(dstSlice.offset, dstSlice.size);
104 : }
105 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, stream_));
106 0 : return HCCL_SUCCESS;
107 0 : }
108 :
109 0 : HcclResult ScatterRingDirect::GetInitializedNeighborLinks(const u32 rank, const u32 rankSize,
110 : const std::vector<LINK> &links)
111 : {
112 : // 收集左邻居信息
113 0 : leftLink_ = links[(rank + rankSize - 1) % rankSize];
114 0 : CHK_SMART_PTR_NULL(leftLink_);
115 :
116 : // 收集右邻居信息
117 0 : rightLink_ = links[(rank + 1) % rankSize];
118 0 : CHK_SMART_PTR_NULL(rightLink_);
119 0 : HCCL_INFO("ScatterRingDirect finished to GetInitializedNeighborLinks");
120 0 : return HCCL_SUCCESS;
121 : }
122 :
123 0 : HcclResult ScatterRingDirect::SetSlices(const u32 rank, const u32 rankSize)
124 : {
125 0 : if (slices_.size() == 0) {
126 0 : slices_.resize(rankSize);
127 :
128 : // 生成std::vector<Slice> slices_
129 0 : u64 sliceSize = count_ * SIZE_TABLE[dataType_];
130 : ;
131 :
132 0 : for (u32 i = 0; i < rankSize; i++) {
133 0 : slices_[i].size = sliceSize;
134 : // 用于DMA消减过程中，消除src与dst不对位的风险
135 0 : slices_[i].offset = RoundUpWithDivisor(i * sliceSize, HCCL_MIN_SLICE_ALIGN);
136 :
137 0 : HCCL_DEBUG("rank[%u], slices[%u].offset=[%llu], slices[%u].size=[%llu]", rank, i, slices_[i].offset, i,
138 : slices_[i].size);
139 : }
140 : }
141 0 : if (UNLIKELY(HcclCheckLogLevel(HCCL_LOG_DEBUG))) {
142 0 : for (u32 i = 0; i < slices_.size(); i++) {
143 0 : HCCL_DEBUG("[ScatterRingDirect][SetSlices]rank[%u], slices[%u].offset=[%llu], slices[%u].size=[%llu]",
144 : rank, i, slices_[i].offset, i, slices_[i].size);
145 : }
146 : }
147 : // 最后一步搬到userMemOut_的offset, 不同的ring环offset不一样
148 0 : lastStepOffset_ = slices_[ringsOrder_[0]].offset;
149 0 : HCCL_INFO("ScatterRingDirect finished to SetSlices");
150 0 : return HCCL_SUCCESS;
151 : }
152 :
153 0 : HcclResult ScatterRingDirect::RunScatter(const u32 rank, const u32 rankSize)
154 : {
155 0 : HCCL_INFO("ScatterRingDirect starts, the input param rank[%u]", rank);
156 : // 空拷贝用于后续操作附着
157 0 : CHK_RET(ExecEmptyTask(inputMem_, outputMem_, stream_, dispatcher_));
158 : // 例如rank[0,1,2,3]中，rank0的rxSliceIdx = 2，txSliceIdx = 3, subSliceIdx = 1
159 0 : u32 txSliceIdx = (rank + rankSize - 1) % rankSize;
160 0 : u32 rxSliceIdx = (rank + rankSize - DMA_REDUCE_TWO_OFFSET) % rankSize;
161 0 : u32 subSliceIdx = (rank + rankSize - DMA_REDUCE_TWO_OFFSET) % rankSize; // 只存在于根节点
162 0 : u32 stepsFromRank2Root = (rank + rankSize - root_) % rankSize;
163 0 : for (u32 step = 0; step < rankSize - 1; step++) {
164 0 : const Slice &subSlice = userMemInputSlices_[subSliceIdx];
165 0 : const Slice &cclSlice = slices_[subSliceIdx];
166 0 : const Slice &txSlice = slices_[txSliceIdx];
167 0 : const Slice &rxSlice = slices_[rxSliceIdx];
168 :
169 0 : CHK_RET(RunScatterOnRootRank(step, subSlice, cclSlice, rank, rankSize));
170 0 : CHK_RET(RunScatterOnOtherRank(stepsFromRank2Root, step, txSlice, rxSlice, rankSize));
171 :
172 : // 更新索引
173 0 : subSliceIdx = (subSliceIdx + rankSize - 1) % rankSize;
174 0 : txSliceIdx = (txSliceIdx + rankSize - 1) % rankSize;
175 0 : rxSliceIdx = (rxSliceIdx + rankSize - 1) % rankSize;
176 : }
177 0 : HCCL_INFO("ScatterRingDirect finished to RunScatter");
178 0 : return HCCL_SUCCESS;
179 : }
180 :
181 0 : HcclResult ScatterRingDirect::RunScatterOnOtherRank(const u32 stepsFromRank2Root, const u32 step,
182 : const Slice &txSlice, const Slice &rxSlice, const u32 rankSize)
183 : {
184 0 : bool needSend = stepsFromRank2Root <= step;
185 0 : bool needReceive = stepsFromRank2Root > 0 && stepsFromRank2Root <= (step + 1);
186 : // Ack
187 0 : if (needReceive) {
188 0 : CHK_RET(leftLink_->TxAck(stream_));
189 : }
190 0 : if (needSend) {
191 0 : CHK_RET(rightLink_->RxAck(stream_));
192 : }
193 :
194 0 : DeviceMem src;
195 : // 不同的rank会在不同的step开始持续发送操作，距离root节点越近，越早step开始发送操作
196 0 : if (needSend) {
197 0 : src = inputMem_.range(txSlice.offset, txSlice.size);
198 0 : CHK_RET(rightLink_->TxAsync(UserMemType::INPUT_MEM, txSlice.offset + baseOffset_, src.ptr(), txSlice.size,
199 : stream_));
200 : }
201 : // 不同的rank会在不同的step开始持续发送操作，距离root节点越近，越早step开始发送操作
202 0 : DeviceMem dst;
203 0 : if (needReceive) {
204 0 : HCCL_DEBUG("MemcpyAsync operation: step[%u] stream[main], src rank[%u] starts to send offset[%llu] size[%llu] "
205 : "from leftMem_",
206 : step, leftLink_->GetRemoteRank(), rxSlice.offset, rxSlice.size);
207 0 : if (step == rankSize - DMA_REDUCE_TWO_OFFSET && opInfo_->outputAddr != nullptr) {
208 0 : HCCL_DEBUG("MemcpyAsync operation: step[%u] stream[main], dst rank[%u] starts to rcv offset[%llu], "
209 : "size[%llu] "
210 : "at userMemOut_",
211 : step, userRank_, lastStepOffset_, rxSlice.size);
212 0 : dst = DeviceMem::create(static_cast<u8 *>(opInfo_->outputAddr) + lastStepOffset_, rxSlice.size);
213 : } else {
214 0 : HCCL_DEBUG("MemcpyAsync operation: step[%u] stream[main], dst rank[%u] starts to rcv offset[%llu], "
215 : "size[%llu] "
216 : "at inputMem_",
217 : step, userRank_, rxSlice.offset, rxSlice.size);
218 0 : dst = inputMem_.range(rxSlice.offset, rxSlice.size);
219 : }
220 0 : CHK_RET(
221 : leftLink_->RxAsync(UserMemType::INPUT_MEM, rxSlice.offset + baseOffset_, dst.ptr(), rxSlice.size, stream_));
222 : }
223 0 : return HCCL_SUCCESS;
224 0 : }
225 :
226 0 : HcclResult ScatterRingDirect::RunScatterOnRootRank(const u32 step, const Slice &subSlice, const Slice &cclSlice,
227 : const u32 rank, const u32 rankSize)
228 : {
229 0 : if (step == rankSize - DMA_REDUCE_TWO_OFFSET && opInfo_->outputAddr != nullptr && rank == root_) {
230 0 : HCCL_DEBUG("MemcpyAsync operation: step[%u] stream[main], dst rank[%u] starts to rcv offset[%llu], "
231 : "size[%llu] at userMemOut_", step, userRank_, lastStepOffset_, subSlice.size);
232 0 : DeviceMem src = inputMem_.range(cclSlice.offset, cclSlice.size);
233 0 : DeviceMem dst = DeviceMem::create(static_cast<u8 *>(opInfo_->outputAddr) + lastStepOffset_, subSlice.size);
234 0 : CHK_RET(HcclD2DMemcpyAsync(dispatcher_, dst, src, stream_));
235 0 : }
236 0 : return HCCL_SUCCESS;
237 : }
238 : REGISTER_TEMPLATE(TemplateType::TEMPLATE_SCATTER_RING_DIRECT, ScatterRingDirect);
239 : } // namespace hccl
|
