@@ -39,6 +39,7 @@ CcuContextAllToAllMesh1D2Die::CcuContextAllToAllMesh1D2Die(const CcuCtxArg if (ctxArg->dimSize_.size() > 0) { rankSize_ = ctxArg->dimSize_[0]; } + bitNumPerCKE_ = ctxArg->bitNum_; } void CcuContextAllToAllMesh1D2Die::InitResource() @@ -80,10 +81,8 @@ void CcuContextAllToAllMesh1D2Die::InitResource() } logicRankSize = withMyRank_ ? transports.size() + 1 : transports.size(); - uint16_t logicId = rankId_ % logicRankSize; // topo为 2 * n - selfBit = 1 << logicId; - allBit = withMyRank_ ? ((1 << logicRankSize) - 1) & (~(1 << logicId)) : (1 << logicRankSize) - 1; - + signalNum_ = (rankSize_ + bitNumPerCKE_ - 1) / bitNumPerCKE_; + HCCL_INFO("[CcuContextAlltoAll2Die] CtxArg: rankId_[%u], rankSize_[%u], signalNum_[%u]", rankId_, rankSize_, signalNum_); return; } @@ -104,28 +103,74 @@ void CcuContextAllToAllMesh1D2Die::LoadArgs() void CcuContextAllToAllMesh1D2Die::PreSync() { - for (auto t : transports) { - // （transport, param, paramID, SemID, mask） - WriteVariableWithSignal(*t, output_[virRankSize - 1], OUPUT_XN_ID, CKE_IDX_1, - selfBit); // index = 1，传递output信息 - WriteVariableWithSignal(*t, token_[virRankSize - 1], TOKEN_XN_ID, CKE_IDX_2, - selfBit); // index = 2，传递token信息 + if (withMyRank_) { + uint16_t logicId = rankId_ % logicRankSize; + selfBit = 1 << logicId; + allBit = ((1 << logicRankSize) - 1) & (~(1 << logicId)); + for (auto t : transports) { + // （transport, param, paramID, SemID, mask） + WriteVariableWithSignal(*t, output_[virRankSize - 1], OUPUT_XN_ID, CKE_IDX_1, + selfBit); // index = 1，传递output信息 + WriteVariableWithSignal(*t, token_[virRankSize - 1], TOKEN_XN_ID, CKE_IDX_2, + selfBit); // index = 2，传递token信息 + } + GroupWait(*transportGroup, CKE_IDX_1, allBit); // index = 1，传递output信息 + GroupWait(*transportGroup, CKE_IDX_2, allBit); // index = 2，传递token信息 + } else { + uint16_t selfSignalId = rankId_ / bitNumPerCKE_; + uint16_t selfBit = 1 << (rankId_ % bitNumPerCKE_); + for (auto t : transports) { + // （transport, param, paramID, SemID, mask） + WriteVariableWithSignal(*t, output_[virRankSize - 1], OUPUT_XN_ID, selfSignalId + signalNum_ * CKE_IDX_1, + selfBit); // index = 1，传递output信息 + WriteVariableWithSignal(*t, token_[virRankSize - 1], TOKEN_XN_ID, selfSignalId + signalNum_ * CKE_IDX_2, + selfBit); // index = 2，传递token信息 + } + std::vector waitBitVector(signalNum_, 0); + for (uint16_t sId = 0; sId < waitBitVector.size(); sId++) { + waitBitVector[sId] = (1 << bitNumPerCKE_) - 1; + if (sId == selfSignalId) { + waitBitVector[sId] = 0; + } + GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_1, waitBitVector[sId]); // index = 1，传递output信息 + GroupWait(*transportGroup, sId + signalNum_ * CKE_IDX_2, waitBitVector[sId]); // index = 2，传递token信息 + } } - - GroupWait(*transportGroup, CKE_IDX_1, allBit); // index = 1，传递output信息 - GroupWait(*transportGroup, CKE_IDX_2, allBit); // index = 2，传递token信息 return; } void CcuContextAllToAllMesh1D2Die::PostSync() { - for (auto t : transports) { - if (t == nullptr) { - THROW(StringFormat("CcuContextAllToAllMesh1D2Die::Algorithm transport ptr is null")); + if (withMyRank_) { + uint16_t logicId = rankId_ % logicRankSize; + selfBit = 1 << logicId; + allBit = ((1 << logicRankSize) - 1) & (~(1 << logicId)); + for (auto t : transports) { + if (t == nullptr) { + THROW(StringFormat("CcuContextAllToAllMesh1D2Die::Algorithm transport ptr is null")); + } + RemotePost(*t, CKE_IDX_0, selfBit); + } + GroupWait(*transportGroup, CKE_IDX_0, allBit); + } else { + uint16_t selfSignalId = rankId_ / bitNumPerCKE_; + uint16_t selfBit = 1 << (rankId_ % bitNumPerCKE_); + for (auto t : transports) { + if (t == nullptr) { + THROW(StringFormat("CcuContextAllToAllMesh1D2Die::Algorithm transport ptr is null")); + } + RemotePost(*t, selfSignalId + signalNum_ * CKE_IDX_0, selfBit); + } + std::vector waitBitVector(signalNum_, 0); + for (uint16_t sId = 0; sId < waitBitVector.size(); sId++) { + waitBitVector[sId] = (1 << bitNumPerCKE_) - 1; + if (sId == selfSignalId) { + waitBitVector[sId] = 0; + } + GroupWait(*transportGroup, CKE_IDX_0, allBit); } - RemotePost(*t, CKE_IDX_0, selfBit); } - GroupWait(*transportGroup, CKE_IDX_0, allBit); + return; } @@ -158,7 +203,6 @@ void CcuContextAllToAllMesh1D2Die::DoRepeatAllToAll() dst[r].token = token_[r]; src[r].addr = input_[virRankSize - 1]; - dst[r].addr = output_[r]; dst[r].addr += outputoffset_; for(uint64_t i = 0; i < dstRank; i++){ @@ -166,20 +210,38 @@ void CcuContextAllToAllMesh1D2Die::DoRepeatAllToAll() } } - uint64_t allBit_ = withMyRank_ ? ((1 << logicRankSize) - 1) & (~(1 << transports.size())) : (1 << logicRankSize) - 1; - // 创建CKE，源端保序 - CcuRep::MaskSignal locMask = CreateMaskSignal(); // all2all 数据搬运 u32 transportIdx = 0; - for (uint64_t r = 0; r < logicRankSize; r++) { - if (withMyRank_ && r == logicRankSize - 1) { - LocalCopyByLoopGroup(dst[r], src[r]); - continue; + if (withMyRank_) { + uint64_t allBit_ = withMyRank_ ? ((1 << logicRankSize) - 1) & (~(1 << transports.size())) : (1 << logicRankSize) - 1; + CcuRep::MaskSignal locMask = CreateMaskSignal(); + for (uint64_t r = 0; r < logicRankSize; r++) { + if (withMyRank_ && r == logicRankSize - 1) { + LocalCopyByLoopGroup(dst[r], src[r]); + continue; + } + Write(*transports[transportIdx], dst[r], src[r], sliceSize_, locMask, 1 << r); + transportIdx++; + } + LocalWait(locMask, allBit_); + } else { + vector locMask; + uint16_t signalNum = logicRankSize / bitNumPerCKE_; + std::vector waitBitVector(signalNum, 0); + for (uint16_t sId = 0; sId < signalNum; sId++) { + locMask.push_back(CreateMaskSignal()); + } + for (uint16_t r = 0; r < logicRankSize; r++) { + uint16_t rmtSignalId = r / bitNumPerCKE_; + uint16_t rmtSignalBit = 1 << (r % bitNumPerCKE_); + Write(*transports[transportIdx], dst[r], src[r], sliceSize_, locMask[rmtSignalId], rmtSignalBit); + transportIdx++; + } + for (uint16_t sId = 0; sId < signalNum; sId++) { + waitBitVector[sId] = (1 << bitNumPerCKE_) - 1; + LocalWait(locMask[sId], waitBitVector[sId]); } - Write(*transports[transportIdx], dst[r], src[r], sliceSize_, locMask, 1 << r); - transportIdx++; } - LocalWait(locMask, allBit_); } void CcuContextAllToAllMesh1D2Die::CreateLocalCopyLoop()

hcomm

hcomm