Star559
1035
代码
代码
Issues81
Pull Requests386
Actions
项目讨论
Wiki
分析
Star559
1035
已合并
aiv sync perf #1959
chenyia创建于 4月27日
aiv sync perf #1959
已合并
chenyia创建于 4月27日
从chenyia/hcomm430poc合入到cann/hcomm9.1.0-beta.1
共 44 个文件变更+785-40
Mscripts/package/module/ascend/CommLib.xml+1-0
@@ -187,6 +187,7 @@
187 <file value="hccl_aiv_all_to_all_vc_op.o"/>187 <file value="hccl_aiv_all_to_all_vc_op.o"/>
188 <file value="hccl_aiv_reduce_scatter_v_op.o"/>188 <file value="hccl_aiv_reduce_scatter_v_op.o"/>
189 <file value="hccl_aiv_all_gather_v_op.o"/>189 <file value="hccl_aiv_all_gather_v_op.o"/>
190+ <file value="hccl_aiv_sync_op.o"/>
190 <file value="hccl_ag_superkernel_int8.o"/>191 <file value="hccl_ag_superkernel_int8.o"/>
191 <file value="hccl_ag_superkernel_int16.o"/>192 <file value="hccl_ag_superkernel_int16.o"/>
192 <file value="hccl_ag_superkernel_int32.o"/>193 <file value="hccl_ag_superkernel_int32.o"/>
Msrc/algorithm/base/alg_aiv_template/CMakeLists.txt+2-1
@@ -13,4 +13,5 @@ add_subdirectory(all_reduce)
13add_subdirectory(all_to_all)13add_subdirectory(all_to_all)
14add_subdirectory(broadcast)14add_subdirectory(broadcast)
15add_subdirectory(reduce_scatter)15add_subdirectory(reduce_scatter)
16-add_subdirectory(reduce_scatter_v)16+add_subdirectory(reduce_scatter_v)
17+add_subdirectory(aiv_sync)
Msrc/algorithm/base/alg_aiv_template/aiv_communication_base.h+0-22
@@ -165,9 +165,6 @@ hiddenInput, input, output
165hiddenInput165hiddenInput
166 166 
167constexpr uint64_t AIV_FLAG_BUFFER_SIZE = 3 * 1024 * 1024; // aiv算子的flag区域大小167constexpr uint64_t AIV_FLAG_BUFFER_SIZE = 3 * 1024 * 1024; // aiv算子的flag区域大小
168-constexpr uint64_t INFO_EVEN_BUFFER_OFFSET = 3 * 1024 * 1024; // aiv算子偶数tag存放AIV_INFO区域偏移
169-constexpr uint64_t INFO_ODD_BUFFER_OFFSET = 4 * 1024 * 1024; // aiv算子奇数tag存放AIV_INFO区域偏移
170-constexpr uint64_t LOG_LEVEL_OFFSET = 2 * 1024 * 1024; // 存放AIV_INFO日志环境变量的偏移
171constexpr uint64_t CLEAR_BUFFER_OFFSET = 1024 * 1024; // 用于清空的aiv buffer的偏移168constexpr uint64_t CLEAR_BUFFER_OFFSET = 1024 * 1024; // 用于清空的aiv buffer的偏移
172constexpr uint64_t SYNC_BUFFER_OFFSET = 2 * 1024 * 1024; // 用于sync的aiv buffer的偏移169constexpr uint64_t SYNC_BUFFER_OFFSET = 2 * 1024 * 1024; // 用于sync的aiv buffer的偏移
173constexpr uint64_t BUFFER_AREA = 1024 * 1024; // aiv算子的单独功能flag区域大小170constexpr uint64_t BUFFER_AREA = 1024 * 1024; // aiv算子的单独功能flag区域大小
@@ -386,8 +383,6 @@ public:
386 pingpongOffset = multiOffset + DOUBLE * DOUBLE * NUM_BLOCKS_FOUR_PER_RANK_A3 * ATOMIC_FLAG_SIZE * DOUBLE;383 pingpongOffset = multiOffset + DOUBLE * DOUBLE * NUM_BLOCKS_FOUR_PER_RANK_A3 * ATOMIC_FLAG_SIZE * DOUBLE;
387 countOffset = DOUBLE * pingpongOffset;384 countOffset = DOUBLE * pingpongOffset;
388 seperateOffset = countOffset + NUM_BLOCKS_FOUR_PER_RANK_A3 * rankSize_ * FLAG_SIZE;385 seperateOffset = countOffset + NUM_BLOCKS_FOUR_PER_RANK_A3 * rankSize_ * FLAG_SIZE;
389- logLevel_ = GetLogLevel();
390- uint64_t offset = (logLevel_ == 1) ? (tag_ & 1 ? INFO_EVEN_BUFFER_OFFSET : INFO_ODD_BUFFER_OFFSET) : INFO_EVEN_BUFFER_OFFSET;
391 386 
392 pipe.InitBuffer(localFlagBuf, UB_FLAG_SIZE_4);387 pipe.InitBuffer(localFlagBuf, UB_FLAG_SIZE_4);
393 localSetTensor = localFlagBuf.GetWithOffset<int32_t>(UB_FLAG_PAD_COUNT, FLAG_ONE_OFFSET);388 localSetTensor = localFlagBuf.GetWithOffset<int32_t>(UB_FLAG_PAD_COUNT, FLAG_ONE_OFFSET);
@@ -406,10 +401,6 @@ public:
406 401 
407 pipe.InitBuffer(flagInQue, AIV_PING_PONG_FACTOR_TWO, UB_FLAG_SIZE);402 pipe.InitBuffer(flagInQue, AIV_PING_PONG_FACTOR_TWO, UB_FLAG_SIZE);
408 InitOpCounter(headCountMem, tailCountMem, addOneMem, counterMemSize, isEnableCounter);403 InitOpCounter(headCountMem, tailCountMem, addOneMem, counterMemSize, isEnableCounter);
409- if (tag_ == 1) {
410- ClearSyncBuf();
411- pipe_barrier(PIPE_ALL);
412- }
413 }404 }
414 405 
415 __aicore__ inline void Init(GM_ADDR hiddenInput, uint64_t threshold, bool useDoubleBuffer = false)406 __aicore__ inline void Init(GM_ADDR hiddenInput, uint64_t threshold, bool useDoubleBuffer = false)
@@ -627,8 +618,6 @@ public:
627 618
628 __aicore__ inline void ClearSyncBuf();619 __aicore__ inline void ClearSyncBuf();
629 620 
630- __aicore__ inline int32_t GetLogLevel();
631- 
632 __aicore__ inline void InitOpCounter(GM_ADDR headCountMem, GM_ADDR tailCountMem, GM_ADDR addOneMem, uint32_t counterMemSize,621 __aicore__ inline void InitOpCounter(GM_ADDR headCountMem, GM_ADDR tailCountMem, GM_ADDR addOneMem, uint32_t counterMemSize,
633 bool isEnableCounter)622 bool isEnableCounter)
634 {623 {
@@ -669,7 +658,6 @@ public:
669 uint64_t len_;658 uint64_t len_;
670 int32_t tag_;659 int32_t tag_;
671 int32_t numBlocks_;660 int32_t numBlocks_;
672- int32_t logLevel_;
673 uint32_t blockIdx_ = GetBlockIdx(); // 在构造函数中初始化，以免漏初始化661 uint32_t blockIdx_ = GetBlockIdx(); // 在构造函数中初始化，以免漏初始化
674 662 
675 bool useDoubleBuffer_;663 bool useDoubleBuffer_;
@@ -769,16 +757,6 @@ __aicore__ inline void AivCommBase::ClearSyncBuf()
769 BlockSync();757 BlockSync();
770}758}
771 759 
772-__aicore__ inline int32_t AivCommBase::GetLogLevel()
773-{
774- #ifndef OPEN_HCCL_TEST
775- int32_t tmpLogLevel = *((__gm__ int32_t*)(GM_OUT[rank_] + LOG_LEVEL_OFFSET - sizeof(int32_t)));
776- return tmpLogLevel;
777- #else
778- return 0;
779- #endif
780-}
781- 
782__aicore__ inline uint64_t AivCommBase::CeilDiv(uint64_t a, uint64_t b)760__aicore__ inline uint64_t AivCommBase::CeilDiv(uint64_t a, uint64_t b)
783{761{
784 if (b == 0) {762 if (b == 0) {
Msrc/algorithm/base/alg_aiv_template/aiv_crossnode_91093_base.h+0-15
@@ -203,8 +203,6 @@ public:
203 203 
204 __aicore__ inline void IntraSync(int32_t curTag, int32_t offset, int32_t blockIdx, bool ifPingpong = false);204 __aicore__ inline void IntraSync(int32_t curTag, int32_t offset, int32_t blockIdx, bool ifPingpong = false);
205 205 
206- __aicore__ inline int32_t GetLogLevel();
207- 
208 __aicore__ inline void InitOpCounter(GM_ADDR headCountMem, GM_ADDR tailCountMem, GM_ADDR addOneMem, 206 __aicore__ inline void InitOpCounter(GM_ADDR headCountMem, GM_ADDR tailCountMem, GM_ADDR addOneMem,
209 uint32_t counterMemSize, bool isEnableCounter)207 uint32_t counterMemSize, bool isEnableCounter)
210 {208 {
@@ -241,7 +239,6 @@ public:
241 uint32_t usedBlockNum_;239 uint32_t usedBlockNum_;
242 uint32_t blockGroup_;240 uint32_t blockGroup_;
243 bool useDoubleBuffer_;241 bool useDoubleBuffer_;
244- int32_t logLevel_;
245 int32_t tag_;242 int32_t tag_;
246 bool localCopyCores = false;243 bool localCopyCores = false;
247 int32_t clearEnable_ = 0;244 int32_t clearEnable_ = 0;
@@ -392,8 +389,6 @@ __aicore__ inline void AivCrossNode91093Base::CalcNumTargetsAndTargetRanksGroup(
392 389 
393__aicore__ inline void AivCrossNode91093Base::InitSetCheckClearArgsTensor() 390__aicore__ inline void AivCrossNode91093Base::InitSetCheckClearArgsTensor()
394{391{
395- logLevel_ = GetLogLevel();
396- uint64_t offset = (logLevel_ == 1) ? (tag_ & 1 ? INFO_EVEN_BUFFER_OFFSET : INFO_ODD_BUFFER_OFFSET) : INFO_EVEN_BUFFER_OFFSET;
397 pipe.InitBuffer(localFlagBuf, UB_FLAG_SIZE * FLAG_BUF_NUM);392 pipe.InitBuffer(localFlagBuf, UB_FLAG_SIZE * FLAG_BUF_NUM);
398 localSetTensor = localFlagBuf.GetWithOffset<int32_t>(UB_FLAG_PAD_COUNT, 0);393 localSetTensor = localFlagBuf.GetWithOffset<int32_t>(UB_FLAG_PAD_COUNT, 0);
399 localCheckTensor = localFlagBuf.GetWithOffset<int32_t>(UB_FLAG_PAD_COUNT, UB_FLAG_SIZE);394 localCheckTensor = localFlagBuf.GetWithOffset<int32_t>(UB_FLAG_PAD_COUNT, UB_FLAG_SIZE);
@@ -849,16 +844,6 @@ __aicore__ inline void AivCrossNode91093Base::IntraSync(int32_t tag, int32_t off
849 }844 }
850}845}
851 846 
852-__aicore__ inline int32_t AivCrossNode91093Base::GetLogLevel()
853-{
854- #ifndef OPEN_HCCL_TEST
855- int32_t tmpLogLevel = *((__gm__ int32_t*)(flagAddrSelf_ + LOG_LEVEL_OFFSET - sizeof(int32_t)));
856- return tmpLogLevel;
857- #else
858- return 0;
859- #endif
860-}
861- 
862__aicore__ inline void AivCrossNode91093Base::SetSyncRecord(int32_t value, GM_ADDR setAddr,847__aicore__ inline void AivCrossNode91093Base::SetSyncRecord(int32_t value, GM_ADDR setAddr,
863 int32_t highOrderOff, int32_t lowOrderOff, bool ifPingpong)848 int32_t highOrderOff, int32_t lowOrderOff, bool ifPingpong)
864{849{
Asrc/algorithm/base/alg_aiv_template/aiv_sync/CMakeLists.txt+293-0
@@ -0,0 +1,293 @@
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+set_source_files_properties(
11+ aiv_sync_op.h PROPERTIES LANGUAGE CXX
12+)
13+ 
14+if(CCACHE_PROGRAM)
15+ set(BISHENG_CCACHE ${CCACHE_PROGRAM})
16+endif()
17+if(BUILD_OPEN_PROJECT)
18+ set(CCE_DIR ${ASCEND_CANN_PACKAGE_PATH}/tools/ccec_compiler/bin)
19+ add_custom_target(hccl_aiv_sync_op ALL
20+ DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o
21+ COMMAND ${CCE_DIR}/ld.lld -m aicorelinux -Ttext=0 ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o -static -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op.o
22+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
23+ )
24+ 
25+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o
26+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o
27+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 --cce-aicore-arch=dav-c220-vec
28+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
29+ -mllvm -cce-aicore-stack-size=0x8000
30+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
31+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
32+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
33+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
34+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
35+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
36+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
37+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
38+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
39+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
40+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
41+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
42+ -I${ASCEND_CANN_PACKAGE_PATH}/include/hccl/
43+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
44+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
45+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
46+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
47+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
48+ )
49+ 
50+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o
51+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o
52+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=1 --cce-aicore-arch=dav-c220-vec
53+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
54+ -mllvm -cce-aicore-stack-size=0x8000
55+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
56+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
57+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
58+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
59+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
60+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
61+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
62+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
63+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
64+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
65+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
66+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
67+ -I${ASCEND_CANN_PACKAGE_PATH}/include/hccl/
68+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
69+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
70+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
71+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
72+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
73+ )
74+ 
75+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o
76+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o
77+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=2 --cce-aicore-arch=dav-c220-vec
78+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
79+ -mllvm -cce-aicore-stack-size=0x8000
80+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
81+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
82+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
83+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
84+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
85+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
86+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
87+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
88+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
89+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
90+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
91+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
92+ -I${ASCEND_CANN_PACKAGE_PATH}/include/hccl/
93+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
94+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
95+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
96+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
97+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
98+ )
99+ 
100+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o
101+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o
102+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=3 --cce-aicore-arch=dav-c220-vec
103+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
104+ -mllvm -cce-aicore-stack-size=0x8000
105+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
106+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
107+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
108+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
109+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
110+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
111+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
112+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
113+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
114+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
115+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
116+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
117+ -I${ASCEND_CANN_PACKAGE_PATH}/include/hccl/
118+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
119+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
120+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
121+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
122+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
123+ )
124+ 
125+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o
126+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o
127+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=4 --cce-aicore-arch=dav-c220-vec
128+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
129+ -mllvm -cce-aicore-stack-size=0x8000
130+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
131+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
132+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
133+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
134+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
135+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
136+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
137+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
138+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
139+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
140+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
141+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
142+ -I${ASCEND_CANN_PACKAGE_PATH}/include/hccl/
143+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
144+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
145+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
146+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
147+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
148+ )
149+else()
150+ get_filename_component(CCE_DIR "${CMAKE_LINKER}" PATH)
151+ 
152+ set(TILING_GEN_DIR ${CMAKE_BINARY_DIR}/asc/asc-devkit)
153+ 
154+ add_custom_target(hccl_aiv_sync_op
155+ DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o kernel_tiling_headers
156+ COMMAND ${CCE_DIR}/ld.lld -m aicorelinux -Ttext=0 ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o -static -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op.o
157+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
158+ )
159+ 
160+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o
161+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_tmp.o
162+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 --cce-aicore-arch=dav-c220-vec
163+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
164+ -mllvm -cce-aicore-stack-size=0x8000
165+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
166+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
167+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
168+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
169+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
170+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
171+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
172+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
173+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
174+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
175+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
176+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
177+ -I${TOP_DIR}/hcomm/include/hccl/
178+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
179+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
180+ -I${TILING_GEN_DIR}
181+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
182+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
183+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
184+ )
185+ 
186+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o
187+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk1.o
188+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=1 --cce-aicore-arch=dav-c220-vec
189+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
190+ -mllvm -cce-aicore-stack-size=0x8000
191+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
192+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
193+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
194+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
195+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
196+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
197+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
198+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
199+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
200+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
201+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
202+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
203+ -I${TOP_DIR}/hcomm/include/hccl/
204+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
205+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
206+ -I${TILING_GEN_DIR}
207+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
208+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
209+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
210+ )
211+ 
212+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o
213+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk2.o
214+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=2 --cce-aicore-arch=dav-c220-vec
215+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
216+ -mllvm -cce-aicore-stack-size=0x8000
217+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
218+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
219+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
220+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
221+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
222+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
223+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
224+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
225+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
226+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
227+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
228+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
229+ -I${TOP_DIR}/hcomm/include/hccl/
230+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
231+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
232+ -I${TILING_GEN_DIR}
233+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
234+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
235+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
236+ )
237+ 
238+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o
239+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk3.o
240+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=3 --cce-aicore-arch=dav-c220-vec
241+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
242+ -mllvm -cce-aicore-stack-size=0x8000
243+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
244+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
245+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
246+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
247+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
248+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
249+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
250+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
251+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
252+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
253+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
254+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
255+ -I${TOP_DIR}/hcomm/include/hccl/
256+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
257+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
258+ -I${TILING_GEN_DIR}
259+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
260+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
261+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
262+ )
263+ 
264+ add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o
265+ COMMAND ${BISHENG_CCACHE} ${CCE_DIR}/ccec -c -x cce -O2 ${CMAKE_CURRENT_SOURCE_DIR}/aiv_sync_op.h -o ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op_sk4.o
266+ -DASCENDC_DUMP=${ASCENDC_DUMP} -DTILING_KEY_VAR=0 -D__ASCENDC_ENABLE_SUPER_KERNEL__ -DBUILD_SK_FUNC -DSK_FUNC_ID=4 --cce-aicore-arch=dav-c220-vec
267+ --cce-aicore-input-parameter-size=28000 --cce-aicore-only -mllvm -cce-aicore-function-stack-size=0x8000 -mllvm -cce-aicore-dcci-insert-for-scalar=false
268+ -mllvm -cce-aicore-stack-size=0x8000
269+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/
270+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/basic_api/interface/
271+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/include/adv_api/
272+ -I${ASCEND_CANN_PACKAGE_PATH}/include/ascendc/highlevel_api
273+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/ascendc/include/basic_api/impl/
274+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/
275+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/basic_api/
276+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/simt_api/
277+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/asc/impl/micro_api/
278+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/basic_api/
279+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/simt_api/
280+ -I${ASCEND_CANN_PACKAGE_PATH}/${CMAKE_HOST_SYSTEM_PROCESSOR}-linux/include/micro_api/
281+ -I${TOP_DIR}/hcomm/include/hccl/
282+ -I${CMAKE_CURRENT_SOURCE_DIR}/../
283+ -I${CMAKE_CURRENT_SOURCE_DIR}/../aiv_interface/
284+ -I${TILING_GEN_DIR}
285+ -mllvm -cce-aicore-record-overflow=false -mllvm -cce-aicore-addr-transform -mllvm
286+ --cce-aicore-jump-expand=true -std=c++17 -fstack-protector-all -DONE_CORE_DUMP_SIZE=1024
287+ WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
288+ )
289+endif()
290+ 
291+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/hccl_aiv_sync_op.o OPTIONAL
292+ DESTINATION ${INSTALL_LIBRARY_DIR}
293+)
Asrc/algorithm/base/alg_aiv_template/aiv_sync/aiv_sync_910b.h+61-0
@@ -0,0 +1,61 @@
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 "aiv_communication_base.h"
12+ 
13+using namespace AscendC;
14+ 
15+class AivSync910B : public AivCommBase {
16+public:
17+ __aicore__ inline AivSync910B() {}
18+ __aicore__ inline void SyncBarrier(int32_t tag);
19+ __aicore__ inline void ClearGM();
20+ __aicore__ inline void Process(int32_t tag);
21+};
22+ 
23+__aicore__ inline void AivSync910B::SyncBarrier(int32_t tag)
24+{
25+ // 从0开始，用4个flag
26+ uint32_t flagOffset = SYNC_BUFFER_OFFSET;
27+ flagOffset += ((tag % AIV_PING_PONG_FACTOR_TWO == 0) ? 0 : rankSize_ * FLAG_SIZE);
28+ if (blockIdx_ != rank_) {
29+ // 卡间同步
30+ SetSignalValue((__gm__ int32_t *)(GM_OUT[blockIdx_] + flagOffset + rank_ * FLAG_SIZE), localSetTensor, 1);
31+ WaitSignalValue((__gm__ int32_t *)(GM_OUT[rank_] + flagOffset + blockIdx_ * FLAG_SIZE), localCheckTensor, 1);
32+ PipeBarrier<PIPE_ALL>();
33+ SetSignalValue((__gm__ int32_t *)(GM_OUT[rank_] + flagOffset + blockIdx_ * FLAG_SIZE), localSetTensor, 0);
34+ }
35+}
36+ 
37+__aicore__ inline void AivSync910B::ClearGM()
38+{
39+ uint32_t emptyOffset = CLEAR_BUFFER_OFFSET;
40+ uint32_t blockCount = BUFFER_AREA / rankSize_;
41+ uint32_t blockOffset = blockCount * blockIdx_;
42+ CpGM2GM(GM_OUT[rank_] + blockOffset, GM_OUT[rank_] + blockOffset + emptyOffset, blockCount);
43+}
44+ 
45+__aicore__ inline void AivSync910B::Process(int32_t tag)
46+{
47+ if (tag == 1) {
48+ SyncBarrier(1);
49+ } else if (tag == IDX_2) {
50+ ClearGM();
51+ } else if (tag == IDX_3) {
52+ SyncBarrier(IDX_2);
53+ }
54+}
55+ 
56+__aicore__ inline void aiv_sync_910b_inner(KERNEL_ARGS_DEF)
57+{
58+ AivSync910B op;
59+ op.Init(KERNEL_CLASS_INIT, false);
60+ op.Process(tag);
61+}
Asrc/algorithm/base/alg_aiv_template/aiv_sync/aiv_sync_910b_rdma.h+56-0
@@ -0,0 +1,56 @@
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 "aiv_communication_base.h"
12+
13+using namespace AscendC;
14+
15+class AivSync910BRdma : public AivCommBase {
16+public:
17+ __aicore__ inline AivSync910BRdma() {}
18+ __aicore__ inline void Process(int32_t tag, uint64_t rmaInfo, int32_t serverNum);
19+};
20+
21+__aicore__ inline void AivSync910BRdma::Process(int32_t tag, uint64_t rmaInfo, int32_t serverNum)
22+{
23+ uint32_t rankPerSever = rankSize_ / serverNum;
24+ uint64_t flagOffset = ((tag % AIV_PING_PONG_FACTOR_TWO == 0) ? 0 : 2 * rankSize_ * FLAG_SIZE);
25+ if (blockIdx_ < serverNum) {
26+ for(uint32_t i = 0; i < rankPerSever; i++) {
27+ uint32_t targetRank = blockIdx_ + i * serverNum;
28+ if (targetRank == rank_) {
29+ continue;
30+ }
31+ uint64_t localFlagOffset = 2 * targetRank * FLAG_SIZE;
32+ __gm__ int32_t *ctrlFlagGM = (__gm__ int32_t *)(GM_IN_RDMA[rank_] + flagOffset + localFlagOffset + FLAG_SIZE); // flag标志位
33+ SetSignalValue(ctrlFlagGM, localSetTensor, 1);
34+ PipeBarrier<PIPE_ALL>();
35+ AIVRDMAPostSend((GM_ADDR)((uint64_t)ctrlFlagGM), (GM_ADDR)((uint64_t)(GM_IN_RDMA[targetRank] + flagOffset + 2 * rank_ * FLAG_SIZE)),
36+ targetRank, UB_FLAG_PAD_COUNT, (__gm__ HcclRMAInfo*)rmaInfo, false, false);
37+ PipeBarrier<PIPE_ALL>();
38+ }
39+ } else if (blockIdx_ < (DOUBLE * serverNum)) {
40+ for (uint32_t i = 0; i < rankPerSever; i++) {
41+ uint32_t sourceRank = blockIdx_ % serverNum + i * serverNum;
42+ if (sourceRank == rank_) {
43+ continue;
44+ }
45+ WaitSignalValue((__gm__ int32_t *)(GM_IN_RDMA[rank_] + flagOffset + 2 * sourceRank * FLAG_SIZE), localCheckTensor, 1);
46+ PipeBarrier<PIPE_ALL>();
47+ SetSignalValue((__gm__ int32_t *)(GM_IN_RDMA[rank_] + flagOffset + 2 * sourceRank * FLAG_SIZE), localSetTensor, 0);
48+ }
49+ }
50+}
51+__aicore__ inline void aiv_sync_910b_rdma(KERNEL_ARGS_DEF)
52+{
53+ AivSync910BRdma op;
54+ op.InitForRDMA(KERNEL_CLASS_INIT, false);
55+ op.Process(tag, rmaInfo, serverNum);
56+}
Asrc/algorithm/base/alg_aiv_template/aiv_sync/aiv_sync_op.h+39-0
@@ -0,0 +1,39 @@
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+#ifndef AIV_SYNC_OP_H
12+#define AIV_SYNC_OP_H
13+ 
14+#include "aiv_communication_base.h"
15+ 
16+#include "aiv_sync_910b.h"
17+#include "aiv_sync_910b_rdma.h"
18+ 
19+// aiv sync
20+__aicore__ inline void hccl_aiv_sync_inner(KERNEL_ARGS_DEF) {
21+ return aiv_sync_910b_inner(KERNEL_ARGS_CALL);
22+}
23+ 
24+// aiv sync rdma
25+__aicore__ inline void hccl_aiv_sync_rdma_inner(KERNEL_ARGS_DEF) {
26+ return aiv_sync_910b_rdma(KERNEL_ARGS_CALL);
27+}
28+ 
29+#if defined(BUILD_SK_FUNC) && defined(SK_FUNC_ID)
30+ SK_BIND_FUNC_DEF_A2(hccl_aiv_sync, SK_FUNC_ID)
31+ SK_BIND_FUNC_DEF_A2(hccl_aiv_sync_rdma, SK_FUNC_ID)
32+#else
33+ GLOBAL_FUNC_DEF_A2(hccl_aiv_sync);
34+ GLOBAL_FUNC_DEF_A2(hccl_aiv_sync_rdma);
35+ SuperKernelBindA2(hccl_aiv_sync);
36+ SuperKernelBindA2(hccl_aiv_sync_rdma);
37+#endif
38+ 
39+#endif /* AIV_SYNC_OP_H */
Msrc/algorithm/impl/coll_executor/coll_all_gather/coll_all_gather_aiv_rdma_executor.cc+3-1
@@ -226,7 +226,9 @@ HcclResult CollAllGatherAivRdmaExecutor::KernelRun(const OpParam &param, ExecMem
226 226 
227 struct AivProfilingInfo aivProfilingInfo;227 struct AivProfilingInfo aivProfilingInfo;
228 aivProfilingInfo.counter = opCounter_;228 aivProfilingInfo.counter = opCounter_;
229- 229+ if (aivClearEnable_) {
230+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
231+ }
230 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));232 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
231 HCCL_INFO("[CollAllGatherAivRdmaExecutor][KernelRun]allGather aiv run success.");233 HCCL_INFO("[CollAllGatherAivRdmaExecutor][KernelRun]allGather aiv run success.");
232 return HCCL_SUCCESS;234 return HCCL_SUCCESS;
Msrc/algorithm/impl/coll_executor/coll_all_gather/coll_all_gather_mesh_aiv_executor.cc+3-0
@@ -214,6 +214,9 @@ HcclResult CollAllGatherMeshAivExecutor::KernelRun(const OpParam &param, ExecMem
214 algArgs.execTimeOutSet = true;214 algArgs.execTimeOutSet = true;
215 struct AivProfilingInfo aivProfilingInfo;215 struct AivProfilingInfo aivProfilingInfo;
216 aivProfilingInfo.counter = opCounter_;216 aivProfilingInfo.counter = opCounter_;
217+ if (aivClearEnable_) {
218+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
219+ }
217 220 
218 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);221 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
219 CHK_PRT_RET(ret != HCCL_SUCCESS,222 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_all_gather/coll_all_gather_mesh_aiv_for_910_93_executor.cc+3-0
@@ -218,6 +218,9 @@ HcclResult CollAllGatherMeshAivFor91093Executor::KernelRun(const OpParam &param,
218 algArgs.execTimeOutSet = true;218 algArgs.execTimeOutSet = true;
219 struct AivProfilingInfo aivProfilingInfo;219 struct AivProfilingInfo aivProfilingInfo;
220 aivProfilingInfo.counter = opCounter_;220 aivProfilingInfo.counter = opCounter_;
221+ if (aivClearEnable_) {
222+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
223+ }
221 224 
222 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);225 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
223 226 
Msrc/algorithm/impl/coll_executor/coll_all_gather/coll_all_gather_mesh_aiv_smallcount_executor.cc+3-0
@@ -204,6 +204,9 @@ HcclResult CollAllGatherMeshAivSmallCountExecutor::KernelRun(const OpParam &para
204 algArgs.execTimeOutSet = true;204 algArgs.execTimeOutSet = true;
205 struct AivProfilingInfo aivProfilingInfo;205 struct AivProfilingInfo aivProfilingInfo;
206 aivProfilingInfo.counter = opCounter_;206 aivProfilingInfo.counter = opCounter_;
207+ if (aivClearEnable_) {
208+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
209+ }
207 210 
208 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);211 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
209 CHK_PRT_RET(ret != HCCL_SUCCESS,212 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_all_gather_v/coll_all_gatherv_mesh_aiv_executor.cc+3-0
@@ -140,6 +140,9 @@ HcclResult AllGatherVMeshAivExecutor::KernelRun(const OpParam &param, ExecMem &e
140 algArgs.execTimeOutSet = true;140 algArgs.execTimeOutSet = true;
141 struct AivProfilingInfo aivProfilingInfo;141 struct AivProfilingInfo aivProfilingInfo;
142 aivProfilingInfo.counter = opCounter_;142 aivProfilingInfo.counter = opCounter_;
143+ if (aivClearEnable_) {
144+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
145+ }
143 146 
144 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);147 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
145 CHK_PRT_RET(ret != HCCL_SUCCESS,148 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_all_gather_v/coll_all_gatherv_mesh_aiv_smallcount_executor.cc+3-0
@@ -137,6 +137,9 @@ HcclResult CollAllGatherVMeshAivSmallCountExecutor::KernelRun(const OpParam &par
137 algArgs.execTimeOutSet = true;137 algArgs.execTimeOutSet = true;
138 struct AivProfilingInfo aivProfilingInfo;138 struct AivProfilingInfo aivProfilingInfo;
139 aivProfilingInfo.counter = opCounter_;139 aivProfilingInfo.counter = opCounter_;
140+ if (aivClearEnable_) {
141+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
142+ }
140 143 
141 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);144 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
142 CHK_PRT_RET(ret != HCCL_SUCCESS,145 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_aiv_deter_executor.cc+4-0
@@ -148,6 +148,10 @@ HcclResult CollAllReduceAivDeterExecutor::KernelRun(const OpParam &param, ExecMe
148 aivProfilingInfo.counter = opCounter_;148 aivProfilingInfo.counter = opCounter_;
149 HCCL_INFO("[CollAllReduceAivDeterExecutor][KernelRun]AllReduce bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());149 HCCL_INFO("[CollAllReduceAivDeterExecutor][KernelRun]AllReduce bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());
150 150 
151+ if (aivClearEnable_) {
152+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
153+ }
154+ 
151 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);155 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
152 156 
153 ExtraArgs extraArgs;157 ExtraArgs extraArgs;
Msrc/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_aiv_deter_small_executor.cc+4-0
@@ -157,6 +157,10 @@ HcclResult CollAllReduceAivDeterSmallExecutor::KernelRun(const OpParam &param, E
157 157 
158 HCCL_INFO("[CollAllReduceAivDeterSmallExecutor][KernelRun]AllReduce bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());158 HCCL_INFO("[CollAllReduceAivDeterSmallExecutor][KernelRun]AllReduce bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());
159 159 
160+ if (aivClearEnable_) {
161+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
162+ }
163+ 
160 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);164 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
161 165 
162 ExtraArgs extraArgs;166 ExtraArgs extraArgs;
Msrc/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_mesh_aiv_executor.cc+3-0
@@ -222,6 +222,9 @@ HcclResult CollAllReduceMeshAivExecutor::KernelRun(const OpParam &param, ExecMem
222 algArgs.execTimeOutSet = true;222 algArgs.execTimeOutSet = true;
223 struct AivProfilingInfo aivProfilingInfo;223 struct AivProfilingInfo aivProfilingInfo;
224 aivProfilingInfo.counter = opCounter_;224 aivProfilingInfo.counter = opCounter_;
225+ if (aivClearEnable_) {
226+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
227+ }
225 228
226 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);229 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
227 230 
Msrc/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_mesh_aiv_for_910_93_executor.cc+3-0
@@ -265,6 +265,9 @@ HcclResult CollAllReduceMeshAivFor91093Executor::KernelRun(const OpParam &param,
265 algArgs.deterministic = 1;265 algArgs.deterministic = 1;
266 }266 }
267 aivProfilingInfo.counter = opCounter_;267 aivProfilingInfo.counter = opCounter_;
268+ if (aivClearEnable_) {
269+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
270+ }
268 271 
269 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);272 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
270 CHK_PRT_RET(ret != HCCL_SUCCESS,273 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_mesh_aiv_smallcount_executor.cc+3-0
@@ -201,6 +201,9 @@ HcclResult CollAllReduceMeshAivSmallCountExecutor::KernelRun(const OpParam &para
201 struct AivProfilingInfo aivProfilingInfo;201 struct AivProfilingInfo aivProfilingInfo;
202 aivProfilingInfo.counter = opCounter_;202 aivProfilingInfo.counter = opCounter_;
203 HCCL_DEBUG("[CollAllReduceMeshAivSmallCountExecutor][KernelRun]numBlocks is %u", numBlocks_);203 HCCL_DEBUG("[CollAllReduceMeshAivSmallCountExecutor][KernelRun]numBlocks is %u", numBlocks_);
204+ if (aivClearEnable_) {
205+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
206+ }
204 207 
205 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);208 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
206 209 
Msrc/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_mid_count_aiv_rdma_executor.cc+3-0
@@ -167,6 +167,9 @@ HcclResult CollAllReduceMidCountAivRdmaExecutor::KernelRun(const OpParam &param,
167 algArgs.execTimeOutSet = true;167 algArgs.execTimeOutSet = true;
168 struct AivProfilingInfo aivProfilingInfo;168 struct AivProfilingInfo aivProfilingInfo;
169 aivProfilingInfo.counter = opCounter_;169 aivProfilingInfo.counter = opCounter_;
170+ if (aivClearEnable_) {
171+ CHK_RET(ClearAivSyncBuf(flagBuffers, resourceArgs, topoArgs, algArgs));
172+ }
170 173
171 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));174 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
172 175
Msrc/algorithm/impl/coll_executor/coll_all_reduce/coll_all_reduce_small_count_aiv_rdma_executor.cc+3-0
@@ -251,6 +251,9 @@ HcclResult CollAllReduceSmallCountAivRdmaExecutor::KernelRun(const OpParam &para
251 algArgs.execTimeOutSet = true;251 algArgs.execTimeOutSet = true;
252 struct AivProfilingInfo aivProfilingInfo;252 struct AivProfilingInfo aivProfilingInfo;
253 aivProfilingInfo.counter = opCounter_;253 aivProfilingInfo.counter = opCounter_;
254+ if (aivClearEnable_) {
255+ CHK_RET(ClearAivSyncBuf(flagBuffers, resourceArgs, topoArgs, algArgs));
256+ }
254 257
255 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));258 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
256 259
Msrc/algorithm/impl/coll_executor/coll_all_to_all/coll_all_to_all_mesh_aiv_executor.cc+9-0
@@ -242,6 +242,9 @@ HcclResult CollAlltoAllMeshAivExecutor::KernelRun(const OpParam &param, ExecMem
242 HCCL_E_PARA);242 HCCL_E_PARA);
243 numBlocks_ = numBlocks;243 numBlocks_ = numBlocks;
244 resourceArgs.numBlocks = numBlocks_;244 resourceArgs.numBlocks = numBlocks_;
245+ if (aivClearEnable_) {
246+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
247+ }
245 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);248 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
246 } else if (param.opType == HcclCMDType::HCCL_CMD_ALLTOALLVC || param.opType == HcclCMDType::HCCL_CMD_ALLTOALL) {249 } else if (param.opType == HcclCMDType::HCCL_CMD_ALLTOALLVC || param.opType == HcclCMDType::HCCL_CMD_ALLTOALL) {
247 for (u32 i = 0; i < localRankSize; i++) {250 for (u32 i = 0; i < localRankSize; i++) {
@@ -264,6 +267,9 @@ HcclResult CollAlltoAllMeshAivExecutor::KernelRun(const OpParam &param, ExecMem
264 HCCL_E_PARA);267 HCCL_E_PARA);
265 numBlocks_ = numBlocks;268 numBlocks_ = numBlocks;
266 resourceArgs.numBlocks = numBlocks_;269 resourceArgs.numBlocks = numBlocks_;
270+ if (aivClearEnable_) {
271+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
272+ }
267 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);273 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
268 } else {274 } else {
269 for (u32 i = 0; i < localRankSize; i++) {275 for (u32 i = 0; i < localRankSize; i++) {
@@ -280,6 +286,9 @@ HcclResult CollAlltoAllMeshAivExecutor::KernelRun(const OpParam &param, ExecMem
280 HCCL_E_PARA);286 HCCL_E_PARA);
281 numBlocks_ = numBlocks;287 numBlocks_ = numBlocks;
282 resourceArgs.numBlocks = numBlocks_;288 resourceArgs.numBlocks = numBlocks_;
289+ if (aivClearEnable_) {
290+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
291+ }
283 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);292 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
284 }293 }
285 CHK_PRT_RET(ret != HCCL_SUCCESS,294 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_all_to_all/coll_all_to_all_mesh_aiv_for_910_93_executor.cc+6-0
@@ -208,9 +208,15 @@ HcclResult CollAlltoAllMeshAivFor91093Executor::KernelRun(const OpParam &param,
208 if (topoArgs.serverNum == 1) {208 if (topoArgs.serverNum == 1) {
209 topoArgs.serverNum = TWO_SERVER_NUM;209 topoArgs.serverNum = TWO_SERVER_NUM;
210 }210 }
211+ if (aivClearEnable_) {
212+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
213+ }
211 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);214 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
212 } else {215 } else {
213 algArgs.argsType = KernelArgsType::ARGS_TYPE_SUPERPOD;216 algArgs.argsType = KernelArgsType::ARGS_TYPE_SUPERPOD;
217+ if (aivClearEnable_) {
218+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
219+ }
214 ExtraArgsV2 extraArgs;220 ExtraArgsV2 extraArgs;
215 if (param.opType == HcclCMDType::HCCL_CMD_ALLTOALLVC) {221 if (param.opType == HcclCMDType::HCCL_CMD_ALLTOALLVC) {
216 for (u32 i = 0; i < localRankSize; i++) {222 for (u32 i = 0; i < localRankSize; i++) {
Msrc/algorithm/impl/coll_executor/coll_all_to_all/coll_all_to_all_mesh_aiv_smallcount_executor.cc+3-0
@@ -203,6 +203,9 @@ HcclResult CollAlltoAllMeshAivSmallCountExecutor::KernelRun(const OpParam &param
203 algArgs.execTimeOutSet = true;203 algArgs.execTimeOutSet = true;
204 struct AivProfilingInfo aivProfilingInfo;204 struct AivProfilingInfo aivProfilingInfo;
205 aivProfilingInfo.counter = opCounter_;205 aivProfilingInfo.counter = opCounter_;
206+ if (aivClearEnable_) {
207+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
208+ }
206 209 
207 // AllToAll pingpong 图模式走单算子归一流程 或者 单算子模式210 // AllToAll pingpong 图模式走单算子归一流程 或者 单算子模式
208 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);211 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
Msrc/algorithm/impl/coll_executor/coll_all_to_all/coll_all_to_all_staged_aiv_rdma_executor.cc+3-0
@@ -136,6 +136,9 @@ HcclResult CollRunAlltoAllStagedAivRdmaExecutor::RunAlltoAllStaged1InAIV(const O
136 struct AivProfilingInfo aivProfilingInfo;136 struct AivProfilingInfo aivProfilingInfo;
137 aivProfilingInfo.counter = opCounter_;137 aivProfilingInfo.counter = opCounter_;
138 HCCL_DEBUG("[CollRunAlltoAllStagedAivRdmaExecutor]RunAlltoAllStaged1InAIV for numBlocks is %u", numBlocks_);138 HCCL_DEBUG("[CollRunAlltoAllStagedAivRdmaExecutor]RunAlltoAllStaged1InAIV for numBlocks is %u", numBlocks_);
139+ if (aivClearEnable_) {
140+ CHK_RET(ClearAivSyncBuf(flagBuffers, resourceArgs, topoArgs, algArgs));
141+ }
139 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));142 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
140 return HCCL_SUCCESS;143 return HCCL_SUCCESS;
141}144}
Msrc/algorithm/impl/coll_executor/coll_broadcast/coll_broadcast_mesh_aiv_executor.cc+3-0
@@ -147,6 +147,9 @@ HcclResult CollBroadcastMeshAivExecutor::KernelRun(const OpParam &param, ExecMem
147 algArgs.execTimeOutSet = true;147 algArgs.execTimeOutSet = true;
148 struct AivProfilingInfo aivProfilingInfo;148 struct AivProfilingInfo aivProfilingInfo;
149 aivProfilingInfo.counter = opCounter_;149 aivProfilingInfo.counter = opCounter_;
150+ if (aivClearEnable_) {
151+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
152+ }
150 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo); // 执行kernelLaunch153 ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo); // 执行kernelLaunch
151 CHK_PRT_RET(ret != HCCL_SUCCESS,154 CHK_PRT_RET(ret != HCCL_SUCCESS,
152 HCCL_ERROR("[CollBroadcastMeshAivExecutor][KernelRun]broadcast aiv failed, return[%d]", ret),155 HCCL_ERROR("[CollBroadcastMeshAivExecutor][KernelRun]broadcast aiv failed, return[%d]", ret),
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter/coll_reduce_scatter_aiv_deter_executor.cc+4-0
@@ -191,6 +191,10 @@ HcclResult CollReduceScatterAivDeterExecutor::KernelRun(const OpParam &param, Ex
191 aivProfilingInfo.counter = opCounter_;191 aivProfilingInfo.counter = opCounter_;
192 HCCL_INFO("[CollReduceScatterAivDeterExecutor][KernelRun]ReduceScatter bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());192 HCCL_INFO("[CollReduceScatterAivDeterExecutor][KernelRun]ReduceScatter bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());
193 193 
194+ if (aivClearEnable_) {
195+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
196+ }
197+ 
194 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);198 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
195 199
196 CHK_PRT_RET(ret != HCCL_SUCCESS,200 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter/coll_reduce_scatter_aiv_deter_small_executor.cc+4-0
@@ -198,6 +198,10 @@ HcclResult CollReduceScatterAivDeterSmallExecutor::KernelRun(const OpParam &para
198 aivProfilingInfo.counter = opCounter_;198 aivProfilingInfo.counter = opCounter_;
199 HCCL_INFO("[CollReduceScatterAivDeterSmallExecutor][KernelRun]ReduceScatter bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());199 HCCL_INFO("[CollReduceScatterAivDeterSmallExecutor][KernelRun]ReduceScatter bufferin[%d] bufferout[%d]",execMem.inputMem.size(), execMem.outputMem.size());
200 200 
201+ if (aivClearEnable_) {
202+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
203+ }
204+ 
201 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);205 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
202 206
203 CHK_PRT_RET(ret != HCCL_SUCCESS,207 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter/coll_reduce_scatter_aiv_rdma_executor.cc+3-0
@@ -177,6 +177,9 @@ HcclResult CollReduceScatterAivRdmaExecutor::KernelRun(const OpParam &param, Exe
177 algArgs.execTimeOutSet = true;177 algArgs.execTimeOutSet = true;
178 struct AivProfilingInfo aivProfilingInfo;178 struct AivProfilingInfo aivProfilingInfo;
179 aivProfilingInfo.counter = opCounter_;179 aivProfilingInfo.counter = opCounter_;
180+ if (aivClearEnable_) {
181+ CHK_RET(ClearAivSyncBuf(flagBuffers, resourceArgs, topoArgs, algArgs));
182+ }
180 183 
181 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));184 CHK_RET(ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo));
182 /* 第二步 节点间RS */185 /* 第二步 节点间RS */
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter/coll_reduce_scatter_mesh_aiv_executor.cc+3-0
@@ -221,6 +221,9 @@ HcclResult CollReduceScatterMeshAivExecutor::KernelRun(const OpParam &param, Exe
221 algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();221 algArgs.execTimeOut = topoMatcher_->GetExecTimeOutConfig();
222 algArgs.execTimeOutSet = true;222 algArgs.execTimeOutSet = true;
223 aivProfilingInfo.counter = opCounter_;223 aivProfilingInfo.counter = opCounter_;
224+ if (aivClearEnable_) {
225+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
226+ }
224 227 
225 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);228 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
226 CHK_PRT_RET(ret != HCCL_SUCCESS,229 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter/coll_reduce_scatter_mesh_aiv_for_910_93_executor.cc+3-0
@@ -334,6 +334,9 @@ HcclResult CollReduceScatterMeshAivFor91093Executor::KernelRun(const OpParam &pa
334 algArgs.execTimeOutSet = true;334 algArgs.execTimeOutSet = true;
335 struct AivProfilingInfo aivProfilingInfo;335 struct AivProfilingInfo aivProfilingInfo;
336 aivProfilingInfo.counter = opCounter_;336 aivProfilingInfo.counter = opCounter_;
337+ if (aivClearEnable_) {
338+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
339+ }
337 340 
338 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);341 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
339 CHK_PRT_RET(ret != HCCL_SUCCESS,342 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter/coll_reduce_scatter_mesh_aiv_smallcount_executor.cc+3-0
@@ -213,6 +213,9 @@ HcclResult CollReduceScatterMeshAivSmallCountExecutor::KernelRun(const OpParam &
213 algArgs.execTimeOutSet = true;213 algArgs.execTimeOutSet = true;
214 struct AivProfilingInfo aivProfilingInfo;214 struct AivProfilingInfo aivProfilingInfo;
215 aivProfilingInfo.counter = opCounter_;215 aivProfilingInfo.counter = opCounter_;
216+ if (aivClearEnable_) {
217+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
218+ }
216 219 
217 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);220 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo);
218 CHK_PRT_RET(ret != HCCL_SUCCESS,221 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter_v/coll_reduce_scatter_v_aiv_big_count_executor.cc+3-0
@@ -153,6 +153,9 @@ HcclResult CollReduceScatterVAIVBigCountExecutor::KernelRun(const OpParam &param
153 algArgs.execTimeOutSet = true;153 algArgs.execTimeOutSet = true;
154 struct AivProfilingInfo aivProfilingInfo;154 struct AivProfilingInfo aivProfilingInfo;
155 aivProfilingInfo.counter = opCounter_;155 aivProfilingInfo.counter = opCounter_;
156+ if (aivClearEnable_) {
157+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
158+ }
156 159 
157 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);160 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
158 CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[CollReduceScatterVAIVBigCountExecutor][KernelRun]"161 CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[CollReduceScatterVAIVBigCountExecutor][KernelRun]"
Msrc/algorithm/impl/coll_executor/coll_reduce_scatter_v/coll_reduce_scatter_v_mesh_aiv_smallcount_executor.cc+3-0
@@ -137,6 +137,9 @@ HcclResult CollReduceScatterVMeshAivSmallCountExecutor::KernelRun(const OpParam
137 algArgs.execTimeOutSet = true;137 algArgs.execTimeOutSet = true;
138 struct AivProfilingInfo aivProfilingInfo;138 struct AivProfilingInfo aivProfilingInfo;
139 aivProfilingInfo.counter = opCounter_;139 aivProfilingInfo.counter = opCounter_;
140+ if (aivClearEnable_) {
141+ CHK_RET(ClearAivSyncBuf(buffersOut, resourceArgs, topoArgs, algArgs));
142+ }
140 143 
141 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);144 HcclResult ret = ExecuteKernelLaunch(opArgs, topoArgs, resourceArgs, algArgs, extraArgs, aivProfilingInfo);
142 CHK_PRT_RET(ret != HCCL_SUCCESS,145 CHK_PRT_RET(ret != HCCL_SUCCESS,
Msrc/algorithm/impl/hccl_aiv.cc+165-0
@@ -36,6 +36,9 @@ constexpr u32 RANK_FIVE = 5;
36constexpr u32 RANK_SIX = 6;36constexpr u32 RANK_SIX = 6;
37constexpr u32 RANK_SEVEN = 7;37constexpr u32 RANK_SEVEN = 7;
38constexpr u32 MAX_ARGS_SIZE_A3_STRUCT = 9;38constexpr u32 MAX_ARGS_SIZE_A3_STRUCT = 9;
39+constexpr u32 STEP_ONE = 1;
40+constexpr u32 STEP_TWO = 2;
41+constexpr u32 STEP_THREE = 3;
39 42 
40constexpr u32 AIV_BUFFER_PING_PONG_FACTOR = 2;43constexpr u32 AIV_BUFFER_PING_PONG_FACTOR = 2;
41 44 
@@ -202,6 +205,13 @@ static std::vector<AivKernelInfo> g_allgathervAivKernelInfoList = {
202};205};
203static std::string g_allgathervAivBinaryName = "hccl_aiv_all_gather_v_op.o";206static std::string g_allgathervAivBinaryName = "hccl_aiv_all_gather_v_op.o";
204 207 
208+// 同步
209+static std::vector<AivKernelInfo> g_syncAivKernelInfoList = {
210+ {"hccl_aiv_sync", HcclDataType::HCCL_DATA_TYPE_RESERVED},
211+ {"hccl_aiv_sync_rdma", HcclDataType::HCCL_DATA_TYPE_RESERVED, KernelArgsType::ARGS_TYPE_SUPERPOD},
212+};
213+static std::string g_syncAivBinaryName = "hccl_aiv_sync_op.o";
214+ 
205static std::map<HcclCMDType, std::pair<std::string, std::vector<AivKernelInfo>>> g_aivKernelInfoMap = {215static std::map<HcclCMDType, std::pair<std::string, std::vector<AivKernelInfo>>> g_aivKernelInfoMap = {
206 {HcclCMDType::HCCL_CMD_ALLREDUCE, {g_allreduceAivBinaryName, g_allreduceAivKernelInfoList}},216 {HcclCMDType::HCCL_CMD_ALLREDUCE, {g_allreduceAivBinaryName, g_allreduceAivKernelInfoList}},
207 {HcclCMDType::HCCL_CMD_REDUCE_SCATTER, {g_reducescatterAivBinaryName, g_reducescatterAivKernelInfoList}},217 {HcclCMDType::HCCL_CMD_REDUCE_SCATTER, {g_reducescatterAivBinaryName, g_reducescatterAivKernelInfoList}},
@@ -212,6 +222,7 @@ static std::map<HcclCMDType, std::pair<std::string, std::vector<AivKernelInfo>>>
212 {HcclCMDType::HCCL_CMD_ALLTOALLVC, {g_alltoallvcAivBinaryName, g_alltoallvcAivKernelInfoList}},222 {HcclCMDType::HCCL_CMD_ALLTOALLVC, {g_alltoallvcAivBinaryName, g_alltoallvcAivKernelInfoList}},
213 {HcclCMDType::HCCL_CMD_REDUCE_SCATTER_V, {g_reducescattervAivBinaryName, g_reducescattervAivKernelInfoList}},223 {HcclCMDType::HCCL_CMD_REDUCE_SCATTER_V, {g_reducescattervAivBinaryName, g_reducescattervAivKernelInfoList}},
214 {HcclCMDType::HCCL_CMD_ALLGATHER_V, {g_allgathervAivBinaryName, g_allgathervAivKernelInfoList}},224 {HcclCMDType::HCCL_CMD_ALLGATHER_V, {g_allgathervAivBinaryName, g_allgathervAivKernelInfoList}},
225+ {HcclCMDType::HCCL_CMD_INVALID, {g_syncAivBinaryName, g_syncAivKernelInfoList}},
215};226};
216 227 
217using AivKernelArgs = struct AivKernelArgsDef {228using AivKernelArgs = struct AivKernelArgsDef {
@@ -572,6 +583,153 @@ u32 GetAivTimeout(s32 execTimeOut, bool isSetByConfig) {
572 return timeout < AIV_TIMEOUT_MAX_US ? timeout : AIV_TIMEOUT_MAX_US;583 return timeout < AIV_TIMEOUT_MAX_US ? timeout : AIV_TIMEOUT_MAX_US;
573}584}
574 585 
586+HcclResult BarrierForMulServer(const AivResourceArgs &resourceArgs, s32 step, const AivTopoArgs &topoArgs, void* args,
587+ u32 argsSize, s32 execTimeOut, bool execTimeOutSet)
588+{
589+ uint64_t beginTime = 0;
590+ SetAivProfilingInfoBeginTime(beginTime);
591+ HcclResult ret = HcclResult::HCCL_E_PARA;
592+ aclrtLaunchKernelCfg cfg;
593+ aclrtLaunchKernelAttr attr[AIV_ATTRNUM_THREE];
594+ attr[0].id = ACL_RT_LAUNCH_KERNEL_ATTR_SCHEM_MODE;
595+ attr[0].value.schemMode = 1;
596+ attr[1].id = ACL_RT_LAUNCH_KERNEL_ATTR_TIMEOUT_US;
597+ attr[1].value.timeoutUs.timeoutLow = GetAivTimeout(execTimeOut, execTimeOutSet);
598+ attr[1].value.timeoutUs.timeoutHigh = 0;
599+ attr[2].id = ACL_RT_LAUNCH_KERNEL_ATTR_ENGINE_TYPE;
600+ attr[2].value.engineType = ACL_RT_ENGINE_TYPE_AIV;
601+ cfg.numAttrs = AIV_ATTRNUM_THREE;
602+ cfg.attrs = attr;
603+
604+ u32 rank = topoArgs.rank;
605+ u8* flagAddr = static_cast<u8 *>(resourceArgs.buffersIn[0]);
606+
607+ aclrtFuncHandle funcHandle;
608+ s8* stubFunc = GetStubFunc(HcclCMDType::HCCL_CMD_INVALID, HcclDataType::HCCL_DATA_TYPE_RESERVED,
609+ KernelArgsType::ARGS_TYPE_SUPERPOD);
610+ ret = GetKernelFunc(funcHandle, stubFunc);
611+ CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[AIV][BarrierForMulServer] errNo[0x%016llx] GetKernelFunc failed, "
612+ "return[%d]", HCCL_ERROR_CODE(HCCL_E_RUNTIME), ret), HCCL_E_RUNTIME);
613+ aclError aclRet = aclrtLaunchKernelWithHostArgs(funcHandle, resourceArgs.numBlocks, resourceArgs.stream,
614+ &cfg, args, argsSize, nullptr, 0);
615+ if (aclRet == ACL_ERROR_RT_INVALID_HANDLE) {
616+ aclError aclGetRet = aclrtBinaryGetFunction(g_binHandleMap[HcclCMDType::HCCL_CMD_INVALID], g_aivNameMap[stubFunc].c_str(), &funcHandle);
617+ CHK_PRT_RET(aclGetRet != ACL_SUCCESS, HCCL_ERROR("[RegisterBinaryKernel]errNo[0x%016llx] get function from binary error.", aclRet),
618+ HCCL_E_NOT_FOUND);
619+ aclRet = aclrtLaunchKernelWithHostArgs(funcHandle, resourceArgs.numBlocks, resourceArgs.stream,
620+ &cfg, args, argsSize, nullptr, 0);
621+ }
622+ CHK_PRT_RET(aclRet != ACL_SUCCESS, HCCL_ERROR("[ExecuteKernelLaunchInner]errNo[0x%016llx] aclrtLaunchKernelWithHostArgs error[%d].",
623+ HCCL_ERROR_CODE(HCCL_E_RUNTIME), aclRet), HCCL_E_RUNTIME);
624+
625+ struct TaskParaGeneral taskParaGeneral;
626+ TaskParaAiv taskParaAiv(HcclCMDType::HCCL_CMD_INVALID, step, 0, resourceArgs.numBlocks, topoArgs.rankSize, -1,
627+ flagAddr, rank);
628+ taskParaGeneral.isMainStream = true;
629+ taskParaGeneral.stream = resourceArgs.stream;
630+ taskParaGeneral.beginTime = beginTime;
631+ taskParaGeneral.aiv = taskParaAiv;
632+
633+ AlgWrap::GetInstance().TaskAivProfiler(topoArgs.identify, taskParaGeneral);
634+
635+ CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[AIV][BarrierForMulServer] errNo[0x%016llx] rtKernelLaunch aiv fail, "
636+ "return[%d]", HCCL_ERROR_CODE(HCCL_E_RUNTIME), ret), HCCL_E_RUNTIME);
637+ return HCCL_SUCCESS;
638+}
639+ 
640+HcclResult LaunchSyncKernel(aclrtLaunchKernelCfg &cfg, AivKernelArgs &aivKernelArgs, rtStream_t stream, s32 step, const std::string& comm)
641+{
642+ uint64_t beginTime = 0;
643+ SetAivProfilingInfoBeginTime(beginTime);
644+ 
645+ aclrtFuncHandle funcHandle;
646+ s8* stubFunc = GetStubFunc(HcclCMDType::HCCL_CMD_INVALID, HcclDataType::HCCL_DATA_TYPE_RESERVED);
647+ HcclResult ret = GetKernelFunc(funcHandle, stubFunc);
648+ CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[AIV][Barrier] errNo[0x%016llx] GetKernelFunc failed, "
649+ "return[%d]", HCCL_ERROR_CODE(HCCL_E_RUNTIME), ret), HCCL_E_RUNTIME);
650+
651+ aivKernelArgs.tag = step;
652+ aclError aclRet = aclrtLaunchKernelWithHostArgs(funcHandle, aivKernelArgs.rankSize, stream,
653+ &cfg, &aivKernelArgs, sizeof(aivKernelArgs), nullptr, 0);
654+ if (aclRet == ACL_ERROR_RT_INVALID_HANDLE) {
655+ aclError aclGetRet = aclrtBinaryGetFunction(g_binHandleMap[HcclCMDType::HCCL_CMD_INVALID], g_aivNameMap[stubFunc].c_str(), &funcHandle);
656+ CHK_PRT_RET(aclGetRet != ACL_SUCCESS, HCCL_ERROR("[RegisterBinaryKernel]errNo[0x%016llx] get function from binary error.", aclRet),
657+ HCCL_E_NOT_FOUND);
658+ aclRet = aclrtLaunchKernelWithHostArgs(funcHandle, aivKernelArgs.rankSize, stream,
659+ &cfg, &aivKernelArgs, sizeof(aivKernelArgs), nullptr, 0);
660+ }
661+ CHK_PRT_RET(aclRet != ACL_SUCCESS, HCCL_ERROR("[RegisterBinaryKernel]errNo[0x%016llx] aclrtLaunchKernelWithHostArgs error[%d].",
662+ HCCL_ERROR_CODE(HCCL_E_RUNTIME), aclRet), HCCL_E_RUNTIME);
663+
664+ struct TaskParaGeneral taskParaGeneral;
665+ u8* flagAddr = static_cast<u8 *>(const_cast<void*>(aivKernelArgs.buffersOut[aivKernelArgs.rank]));
666+ TaskParaAiv taskParaAiv(HcclCMDType::HCCL_CMD_INVALID, step, 0, aivKernelArgs.rankSize, aivKernelArgs.rankSize, -1, flagAddr, aivKernelArgs.rank);
667+ taskParaGeneral.isMainStream = true;
668+ taskParaGeneral.stream = stream;
669+ taskParaGeneral.beginTime = beginTime;
670+ taskParaGeneral.aiv = taskParaAiv;
671+ 
672+ AlgWrap::GetInstance().TaskAivProfiler(comm, taskParaGeneral);
673+ return HCCL_SUCCESS;
674+}
675+ 
676+HcclResult ClearAivSyncBuf(void** cclBuffersOut, const AivResourceArgs &resourceArgs, const AivTopoArgs &topoArgs, AivAlgArgs algArgs)
677+{
678+ if (algArgs.argsType == KernelArgsType::ARGS_TYPE_SERVER && (topoArgs.devType != DevType::DEV_TYPE_910_93 || topoArgs.serverNum == 1)) {
679+ u32 rankSize = topoArgs.rankSize;
680+ rtStream_t stream = resourceArgs.stream;
681+ const std::string& comm = topoArgs.identify;
682+ u32 numBlocks = resourceArgs.numBlocks;
683+ s32 execTimeOut = algArgs.execTimeOut;
684+ bool execTimeOutSet = algArgs.execTimeOutSet;
685+ 
686+ CHK_PRT_RET(numBlocks < rankSize,
687+ HCCL_ERROR("[ClearAivSyncBuf]aivCore[%u] is invalid, at least need [%u].", numBlocks, rankSize), HCCL_E_PARA);
688+ 
689+ aclrtLaunchKernelCfg cfg;
690+ aclrtLaunchKernelAttr attr[AIV_ATTRNUM_THREE];
691+ attr[0].id = ACL_RT_LAUNCH_KERNEL_ATTR_SCHEM_MODE;
692+ attr[0].value.schemMode = 1;
693+ attr[1].id = ACL_RT_LAUNCH_KERNEL_ATTR_TIMEOUT_US;
694+ attr[1].value.timeoutUs.timeoutLow = GetAivTimeout(execTimeOut, execTimeOutSet);
695+ attr[1].value.timeoutUs.timeoutHigh = 0;
696+ attr[2].id = ACL_RT_LAUNCH_KERNEL_ATTR_ENGINE_TYPE;
697+ attr[2].value.engineType = ACL_RT_ENGINE_TYPE_AIV;
698+ cfg.numAttrs = AIV_ATTRNUM_THREE;
699+ cfg.attrs = attr;
700+ 
701+ AivKernelArgs aivKernelArgs {
702+ cclBuffersOut, cclBuffersOut, nullptr, nullptr, topoArgs.rank, rankSize, 0,
703+ HcclDataType::HCCL_DATA_TYPE_RESERVED, HcclReduceOp::HCCL_REDUCE_RESERVED, 0, 0, rankSize
704+ };
705+ 
706+ CHK_RET(LaunchSyncKernel(cfg, aivKernelArgs, stream, STEP_ONE, comm));
707+ CHK_RET(LaunchSyncKernel(cfg, aivKernelArgs, stream, STEP_TWO, comm));
708+ CHK_RET(LaunchSyncKernel(cfg, aivKernelArgs, stream, STEP_THREE, comm));
709+ }
710+ HCCL_INFO("[AIV][ClearAivSyncBuf] clearaiv done.");
711+ return HCCL_SUCCESS;
712+}
713+ 
714+HcclResult ClearAivSyncBufForMulServer(const AivResourceArgs &resourceArgs, const AivTopoArgs &topoArgs, void* args,
715+ u32 argsSize, AivAlgArgs algArgs)
716+{
717+ CHK_RET(BarrierForMulServer(resourceArgs, 1, topoArgs, args, argsSize, algArgs.execTimeOut, algArgs.execTimeOutSet));
718+
719+ u8* flagAddr = static_cast<u8 *>(resourceArgs.buffersOut[0]);
720+ DeviceMem zeroMem = DeviceMem::create(flagAddr + AIV_FLAG_AREA_SIZE, AIV_FLAG_AREA_SIZE);
721+ DeviceMem flagMem = DeviceMem::create(flagAddr, AIV_FLAG_AREA_SIZE);
722+
723+ CHK_RET(hrtMemAsyncCopy(flagMem.ptr(), AIV_FLAG_AREA_SIZE, zeroMem.ptr(), AIV_FLAG_AREA_SIZE,
724+ HcclRtMemcpyKind::HCCL_RT_MEMCPY_KIND_DEVICE_TO_DEVICE, resourceArgs.stream));
725+
726+ CHK_RET(BarrierForMulServer(resourceArgs, RESET_TAIL_SYNC_TAG, topoArgs, args, argsSize, algArgs.execTimeOut,
727+ algArgs.execTimeOutSet));
728+
729+ HCCL_INFO("[AIV][ClearAivSyncBufForMulServer] clearaiv done");
730+ return HCCL_SUCCESS;
731+}
732+ 
575void TaskAivProfilerWrap(const AivOpArgs& opArgs, const AivTopoArgs& topoArgs,733void TaskAivProfilerWrap(const AivOpArgs& opArgs, const AivTopoArgs& topoArgs,
576 const AivResourceArgs& resourceArgs, const AivAlgArgs& algArgs, const AivProfilingInfo& aivProfilingInfo,734 const AivResourceArgs& resourceArgs, const AivAlgArgs& algArgs, const AivProfilingInfo& aivProfilingInfo,
577 void* flagMem)735 void* flagMem)
@@ -649,6 +807,13 @@ HcclResult ExecuteKernelLaunchInner(const AivOpArgs &opArgs, const AivTopoArgs &
649 807 
650 TaskAivProfilerWrap(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo,808 TaskAivProfilerWrap(opArgs, topoArgs, resourceArgs, algArgs, aivProfilingInfo,
651 (algArgs.argsType != KernelArgsType::ARGS_TYPE_SERVER) ? resourceArgs.buffersOut[0]: resourceArgs.buffersOut[topoArgs.rank]);809 (algArgs.argsType != KernelArgsType::ARGS_TYPE_SERVER) ? resourceArgs.buffersOut[0]: resourceArgs.buffersOut[topoArgs.rank]);
810+ if (resourceArgs.aivTag == TAG_RESET_COUNT) {
811+ if (algArgs.isNpuDirectRoce) {
812+ CHK_RET(ClearAivSyncBufForMulServer(resourceArgs, topoArgs, args, argsSize, algArgs));
813+ } else {
814+ CHK_RET(ClearAivSyncBuf(resourceArgs.buffersOut, resourceArgs, topoArgs, algArgs));
815+ }
816+ }
652 817 
653 CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[AIV][ExecuteKernelLaunchInner] errNo[0x%016llx] rtKernelLaunch aiv fail, "818 CHK_PRT_RET(ret != HCCL_SUCCESS, HCCL_ERROR("[AIV][ExecuteKernelLaunchInner] errNo[0x%016llx] rtKernelLaunch aiv fail, "
654 "return[%d]", HCCL_ERROR_CODE(HCCL_E_RUNTIME), ret), HCCL_E_RUNTIME);819 "return[%d]", HCCL_ERROR_CODE(HCCL_E_RUNTIME), ret), HCCL_E_RUNTIME);
Msrc/algorithm/impl/hccl_aiv_device.cc+9-0
@@ -33,6 +33,15 @@ HcclResult UnRegisterAivKernel()
33 return HCCL_SUCCESS;33 return HCCL_SUCCESS;
34}34}
35 35 
36+HcclResult ClearAivSyncBuf(void** cclBuffersOut, const AivResourceArgs &resourceArgs, const AivTopoArgs &topoArgs, AivAlgArgs algArgs)
37+{
38+ (void) cclBuffersOut;
39+ (void) resourceArgs;
40+ (void) topoArgs;
41+ (void) algArgs;
42+ return HCCL_SUCCESS;
43+}
44+ 
36// KernelLaunch内部接口45// KernelLaunch内部接口
37HcclResult ExecuteKernelLaunchInner(const AivOpArgs &opArgs, const AivTopoArgs &topoArgs,46HcclResult ExecuteKernelLaunchInner(const AivOpArgs &opArgs, const AivTopoArgs &topoArgs,
38 const AivResourceArgs &resourceArgs, const AivAlgArgs &algArgs, void* args, u32 argsSize, 47 const AivResourceArgs &resourceArgs, const AivAlgArgs &algArgs, void* args, u32 argsSize,
Msrc/algorithm/pub_inc/hccl_aiv.h+4-1
@@ -317,7 +317,10 @@ HcclResult ExecuteKernelLaunchImpl(const AivOpArgs &opArgs, const AivTopoArgs &t
317 317 
318HcclResult RegisterKernel(DevType deviceType);318HcclResult RegisterKernel(DevType deviceType);
319HcclResult UnRegisterAivKernel();319HcclResult UnRegisterAivKernel();
320- 320+ 
321+HcclResult ClearAivSyncBuf(void** cclBuffersOut, const AivResourceArgs &resourceArgs,
322+ const AivTopoArgs &topoArgs, AivAlgArgs algArgs = AivAlgArgs{});
323+ 
321HcclResult ClearAivSyncBufForMulServer(const AivResourceArgs &resourceArgs, const AivTopoArgs &topoArgs, void* args,324HcclResult ClearAivSyncBufForMulServer(const AivResourceArgs &resourceArgs, const AivTopoArgs &topoArgs, void* args,
322 u32 argsSize);325 u32 argsSize);
323 326 
Msrc/framework/communicator/impl/hccl_communicator_host.cc+3-0
@@ -4155,6 +4155,9 @@ namespace hccl
4155 CHK_RET(RegisterDfxInfo(opParam, algType, resMap_[newTag].slaveStreams, selectAivAlg));4155 CHK_RET(RegisterDfxInfo(opParam, algType, resMap_[newTag].slaveStreams, selectAivAlg));
4156 // 头计数4156 // 头计数
4157 CHK_RET(StarsCounter(dispatcher_, opParam.stream, HEAD, opParam.aicpuUnfoldMode, retryEnable_, selectAivAlg));4157 CHK_RET(StarsCounter(dispatcher_, opParam.stream, HEAD, opParam.aicpuUnfoldMode, retryEnable_, selectAivAlg));
4158+ if (aivClearEnable_) {
4159+ CHK_RET(ClearAivSyncBuf(cacheInfo.buffersOut, cacheInfo.resourceArgs, cacheInfo.topoArgs, cacheInfo.algArgs));
4160+ }
4158 u64 dataSize = (opParam.opType == HcclCMDType::HCCL_CMD_ALLTOALL ?4161 u64 dataSize = (opParam.opType == HcclCMDType::HCCL_CMD_ALLTOALL ?
4159 opParam.All2AllDataDes.sendCount * SIZE_TABLE[opParam.All2AllDataDes.sendType] : 0);4162 opParam.All2AllDataDes.sendCount * SIZE_TABLE[opParam.All2AllDataDes.sendType] : 0);
4160 if (opType == HcclCMDType::HCCL_CMD_REDUCE_SCATTER_V || opType == HcclCMDType::HCCL_CMD_ALLGATHER_V ||4163 if (opType == HcclCMDType::HCCL_CMD_REDUCE_SCATTER_V || opType == HcclCMDType::HCCL_CMD_ALLGATHER_V ||
Mtest/st/algorithm/testcase/testcase/testcase_all_gather.cc+1-0
@@ -49,6 +49,7 @@ protected:
49 std::string caseName = "analysis_result_" + std::string(test_info->test_case_name()) + "_" + std::string(test_info->name());49 std::string caseName = "analysis_result_" + std::string(test_info->test_case_name()) + "_" + std::string(test_info->name());
50 Checker::SetDumpFileName(caseName);50 Checker::SetDumpFileName(caseName);
51 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));51 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
52+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
52 }53 }
53 54 
54 virtual void TearDown()55 virtual void TearDown()
Mtest/st/algorithm/testcase/testcase/testcase_all_reduce.cc+7-0
@@ -53,6 +53,7 @@ protected:
53 std::string caseName = "analysis_result_" + std::string(test_info->test_case_name()) + "_" + std::string(test_info->name());53 std::string caseName = "analysis_result_" + std::string(test_info->test_case_name()) + "_" + std::string(test_info->name());
54 Checker::SetDumpFileName(caseName);54 Checker::SetDumpFileName(caseName);
55 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));55 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
56+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
56 }57 }
57 58 
58 virtual void TearDown()59 virtual void TearDown()
@@ -253,6 +254,7 @@ TEST_F(AllReduceTest, allreduce_allReduceComm_executor_test)
253TEST_F(AllReduceTest, allreduce_executor_test_loop)254TEST_F(AllReduceTest, allreduce_executor_test_loop)
254{255{
255 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));256 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
257+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
256 RankTable_For_LLT gen;258 RankTable_For_LLT gen;
257 TopoMeta topoMeta;259 TopoMeta topoMeta;
258 gen.GenTopoMeta(topoMeta, 1, 1, 8);260 gen.GenTopoMeta(topoMeta, 1, 1, 8);
@@ -1303,6 +1305,7 @@ TEST_F(AllReduceTest, allreduce_mix_AllReduceMixExecutor_deter_comm)
1303TEST_F(AllReduceTest, allreduce_aiv_a2_AllReduceMeshAivExecutor)1305TEST_F(AllReduceTest, allreduce_aiv_a2_AllReduceMeshAivExecutor)
1304{1306{
1305 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));1307 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
1308+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
1306 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));1309 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
1307 RankTable_For_LLT gen;1310 RankTable_For_LLT gen;
1308 TopoMeta topoMeta;1311 TopoMeta topoMeta;
@@ -1328,6 +1331,7 @@ TEST_F(AllReduceTest, allreduce_aiv_a2_AllReduceMeshAivExecutor)
1328TEST_F(AllReduceTest, allreduce_aiv_a3_graph_AllReduceMeshAivSmallCountExecutor)1331TEST_F(AllReduceTest, allreduce_aiv_a3_graph_AllReduceMeshAivSmallCountExecutor)
1329{1332{
1330 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));1333 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
1334+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
1331 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));1335 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
1332 RankTable_For_LLT gen;1336 RankTable_For_LLT gen;
1333 TopoMeta topoMeta;1337 TopoMeta topoMeta;
@@ -1721,6 +1725,7 @@ TEST_F(AllReduceTest, allreduce_aiv_determinstic_small_test)
1721 setenv("HCCL_DETERMINISTIC", "true", 1);1725 setenv("HCCL_DETERMINISTIC", "true", 1);
1722 setenv("HCCL_OP_EXPANSION_MODE", "AIV", 1);1726 setenv("HCCL_OP_EXPANSION_MODE", "AIV", 1);
1723 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));1727 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
1728+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
1724 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));1729 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
1725 1730 
1726 CheckerOpParam checkerOpParam;1731 CheckerOpParam checkerOpParam;
@@ -1749,6 +1754,7 @@ TEST_F(AllReduceTest, allreduce_aiv_determinstic_test)
1749 setenv("HCCL_DETERMINISTIC", "true", 1);1754 setenv("HCCL_DETERMINISTIC", "true", 1);
1750 setenv("HCCL_OP_EXPANSION_MODE", "AIV", 1);1755 setenv("HCCL_OP_EXPANSION_MODE", "AIV", 1);
1751 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));1756 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
1757+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
1752 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));1758 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
1753 1759 
1754 CheckerOpParam checkerOpParam;1760 CheckerOpParam checkerOpParam;
@@ -1777,6 +1783,7 @@ TEST_F(AllReduceTest, allreduce_aiv_determinstic_mid_test)
1777 setenv("HCCL_DETERMINISTIC", "true", 1);1783 setenv("HCCL_DETERMINISTIC", "true", 1);
1778 setenv("HCCL_OP_EXPANSION_MODE", "AIV", 1);1784 setenv("HCCL_OP_EXPANSION_MODE", "AIV", 1);
1779 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));1785 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
1786+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
1780 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));1787 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
1781 1788 
1782 CheckerOpParam checkerOpParam;1789 CheckerOpParam checkerOpParam;
Mtest/st/algorithm/testcase/testcase/testcase_all_to_all.cc+4-0
@@ -574,6 +574,7 @@ TEST_F(AllToAllTest, alltoall_test_91093_opbase_AlltoAllMeshAivExecutor)
574{574{
575 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));575 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
576 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));576 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
577+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
577 RankTable_For_LLT gen;578 RankTable_For_LLT gen;
578 TopoMeta topoMeta;579 TopoMeta topoMeta;
579 gen.GenTopoMeta(topoMeta, 1, 1, 16);580 gen.GenTopoMeta(topoMeta, 1, 1, 16);
@@ -603,6 +604,7 @@ TEST_F(AllToAllTest, alltoall_test_a2_opbase_AlltoAllStagedAIVRdmaExecutor)
603{604{
604 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));605 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
605 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));606 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
607+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
606 RankTable_For_LLT gen;608 RankTable_For_LLT gen;
607 TopoMeta topoMeta;609 TopoMeta topoMeta;
608 gen.GenTopoMeta(topoMeta, 1, 2, 8);610 gen.GenTopoMeta(topoMeta, 1, 2, 8);
@@ -631,6 +633,7 @@ TEST_F(AllToAllTest, alltoall_test_91093_graph_AlltoAllMeshAivExecutor)
631{633{
632 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));634 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
633 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));635 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
636+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
634 RankTable_For_LLT gen;637 RankTable_For_LLT gen;
635 TopoMeta topoMeta;638 TopoMeta topoMeta;
636 gen.GenTopoMeta(topoMeta, 1, 1, 16);639 gen.GenTopoMeta(topoMeta, 1, 1, 16);
@@ -662,6 +665,7 @@ TEST_F(AllToAllTest, alltoall_test_91093_opbase_AlltoAllMeshAivFor91093Executor)
662 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));665 MOCKER(GetExternalInputHcclAivMode).stubs().will(returnValue(true));
663 MOCKER_CPP(&AlltoAllOperator::IsSatisfyAlltoAllAivCondition).stubs().will(returnValue(true));666 MOCKER_CPP(&AlltoAllOperator::IsSatisfyAlltoAllAivCondition).stubs().will(returnValue(true));
664 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));667 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
668+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
665 RankTable_For_LLT gen;669 RankTable_For_LLT gen;
666 TopoMeta topoMeta;670 TopoMeta topoMeta;
667 gen.GenTopoMeta(topoMeta, 1, 2, 16);671 gen.GenTopoMeta(topoMeta, 1, 2, 16);
Mtest/st/algorithm/testcase/testcase/testcase_reduce_scatter.cc+1-0
@@ -49,6 +49,7 @@ protected:
49 std::string caseName = "analysis_result_" + std::string(test_info->test_case_name()) + "_" + std::string(test_info->name());49 std::string caseName = "analysis_result_" + std::string(test_info->test_case_name()) + "_" + std::string(test_info->name());
50 Checker::SetDumpFileName(caseName);50 Checker::SetDumpFileName(caseName);
51 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));51 MOCKER(ExecuteKernelLaunch).stubs().will(returnValue(HCCL_SUCCESS));
52+ MOCKER(ClearAivSyncBuf).stubs().will(returnValue(HCCL_SUCCESS));
52 }53 }
53 54 
54 virtual void TearDown()55 virtual void TearDown()
Mtest/st/algorithm/utils/adapter_v1/hccl_aiv_stub.cc+9-0
@@ -210,6 +210,8 @@ static std::vector<AivKernelInfo> g_aivKernelInfoList = {
210 {"aiv_broadcast_int8_t", HcclCMDType::HCCL_CMD_BROADCAST, HcclDataType::HCCL_DATA_TYPE_INT8},210 {"aiv_broadcast_int8_t", HcclCMDType::HCCL_CMD_BROADCAST, HcclDataType::HCCL_DATA_TYPE_INT8},
211 {"aiv_broadcast_uint8_t", HcclCMDType::HCCL_CMD_BROADCAST, HcclDataType::HCCL_DATA_TYPE_UINT8},211 {"aiv_broadcast_uint8_t", HcclCMDType::HCCL_CMD_BROADCAST, HcclDataType::HCCL_DATA_TYPE_UINT8},
212 {"aiv_broadcast_bfloat16_t", HcclCMDType::HCCL_CMD_BROADCAST, HcclDataType::HCCL_DATA_TYPE_BFP16},212 {"aiv_broadcast_bfloat16_t", HcclCMDType::HCCL_CMD_BROADCAST, HcclDataType::HCCL_DATA_TYPE_BFP16},
213+ // 同步
214+ {"hccl_aiv_sync", HcclCMDType::HCCL_CMD_INVALID, HcclDataType::HCCL_DATA_TYPE_RESERVED},
213};215};
214 216 
215extern "C" {217extern "C" {
@@ -252,6 +254,7 @@ extern "C" {
252 extern void aiv_broadcast_int8_t(KERNEL_ARGS_DEF);254 extern void aiv_broadcast_int8_t(KERNEL_ARGS_DEF);
253 extern void aiv_broadcast_uint8_t(KERNEL_ARGS_DEF);255 extern void aiv_broadcast_uint8_t(KERNEL_ARGS_DEF);
254 extern void aiv_broadcast_bfloat16_t(KERNEL_ARGS_DEF);256 extern void aiv_broadcast_bfloat16_t(KERNEL_ARGS_DEF);
257+ extern void hccl_aiv_sync(KERNEL_ARGS_DEF);
255 extern void aiv_all_to_all_half(EXTERN_KERNEL_ARGS_DEF);258 extern void aiv_all_to_all_half(EXTERN_KERNEL_ARGS_DEF);
256 extern void aiv_all_to_all_int16_t(EXTERN_KERNEL_ARGS_DEF);259 extern void aiv_all_to_all_int16_t(EXTERN_KERNEL_ARGS_DEF);
257 extern void aiv_all_to_all_uint16_t(EXTERN_KERNEL_ARGS_DEF);260 extern void aiv_all_to_all_uint16_t(EXTERN_KERNEL_ARGS_DEF);
@@ -369,6 +372,7 @@ std::unordered_map<const char*, aivFunc> aivFuncMap = {
369 {"aiv_broadcast_int8_t", aiv_broadcast_int8_t},372 {"aiv_broadcast_int8_t", aiv_broadcast_int8_t},
370 {"aiv_broadcast_uint8_t", aiv_broadcast_uint8_t},373 {"aiv_broadcast_uint8_t", aiv_broadcast_uint8_t},
371 {"aiv_broadcast_bfloat16_t", aiv_broadcast_bfloat16_t},374 {"aiv_broadcast_bfloat16_t", aiv_broadcast_bfloat16_t},
375+ {"hccl_aiv_sync", hccl_aiv_sync},
372};376};
373 377 
374std::unordered_map<const char*, aivFuncExtra> aivFuncExtraMap = {378std::unordered_map<const char*, aivFuncExtra> aivFuncExtraMap = {
@@ -695,4 +699,9 @@ void TaskAivProfilerWrap(const AivOpArgs& opArgs, const AivTopoArgs& topoArgs,
695{699{
696}700}
697 701
702+HcclResult ClearAivSyncBuf(void** cclBuffersOut, const AivResourceArgs &resourceArgs, const AivTopoArgs &topoArgs, AivAlgArgs algArgs)
703+{
704+ return HCCL_SUCCESS;
705+}
706+ 
698}707}
Mtest/ut/impl/ut_hccl_impl_new.cc+39-0
@@ -385,6 +385,45 @@ TEST_F(HcclImplTest, ut_SelectAlg_when_broadcast_910C_Expect_ReturnIs_BroadcastM
385 std::unique_ptr<BroadCastOperator> operation(new (std::nothrow) BroadCastOperator(algConfigurator.get(), cclBufferManager, dispatcher, topoMatcher));385 std::unique_ptr<BroadCastOperator> operation(new (std::nothrow) BroadCastOperator(algConfigurator.get(), cclBufferManager, dispatcher, topoMatcher));
386 ret = operation->SelectAlg("", opParam, algName, newTag);386 ret = operation->SelectAlg("", opParam, algName, newTag);
387 EXPECT_TRUE(algName == "BroadcastMeshAivExecutor");387 EXPECT_TRUE(algName == "BroadcastMeshAivExecutor");
388+ 
389+ MOCKER_CPP(&HcclCommunicator::HandleAclGraphFirstOpAivBuff)
390+ .stubs()
391+ .with(any())
392+ .will(returnValue(HCCL_SUCCESS));
393+ MOCKER_CPP(&HcclCommunicator::RegisterDfxInfo)
394+ .stubs()
395+ .with(any())
396+ .will(returnValue(HCCL_SUCCESS));
397+ MOCKER(StarsCounter).stubs().will(returnValue(HCCL_SUCCESS));
398+ AivOpArgs opArgs{HcclCMDType::HCCL_CMD_BROADCAST,
399+ opParam.inputPtr,
400+ opParam.outputPtr,
401+ opParam.DataDes.count,
402+ opParam.DataDes.dataType,
403+ opParam.reduceType,
404+ opParam.root,
405+ 0};
406+ AivTopoArgs topoArgs{0, 2};
407+ topoArgs.identify = "test";
408+ opParam.aivTag = 1;
409+ AivResourceArgs resourceArgs {
410+ opParam.tag, opParam.stream.ptr(), nullptr, nullptr, 4096, 1, opParam.aivTag
411+ };
412+ AivAlgArgs algArgs{};
413+ algArgs.execTimeOut = 1;
414+ algArgs.execTimeOutSet = true;
415+ struct AivProfilingInfo aivProfilingInfo;
416+ ExtraArgs extraArgs;
417+ HcclCacheInfo cacheInfo;
418+ cacheInfo.opArgs = opArgs;
419+ cacheInfo.topoArgs = topoArgs;
420+ cacheInfo.resourceArgs = resourceArgs;
421+ cacheInfo.algArgs = algArgs;
422+ cacheInfo.profilingInfo = aivProfilingInfo;
423+ cacheInfo.extraArgs = extraArgs;
424+ cacheInfo.isUseCache = true;
425+ implBase->SetClearAivSyncBuf(true);
426+ implBase->ExecOpCache(HcclCMDType::HCCL_CMD_BROADCAST, opParam, cacheInfo);
388 operation = nullptr;427 operation = nullptr;
389 GlobalMockObject::verify();428 GlobalMockObject::verify();
390}429}