Star1233
2024
代码
代码
Issues415
Pull Requests514
Actions
项目讨论
Wiki
分析
Star1233
2024
已合并
DynamicQuant 字面常量修改 #8346
dongfei创建于 8月6日
DynamicQuant 字面常量修改 #8346
已合并
dongfei创建于 8月6日
从ASCEND222/ops-nn_qy3master合入到cann/ops-nnmaster
共 1 个文件变更+20-10
Mquant/dynamic_quant/op_host/dynamic_quant_tiling_arch35.cpp+20-10
@@ -75,6 +75,16 @@ constexpr uint32_t PER_CHANNEL_N_BASE_SIZE = 64;
75constexpr uint32_t PER_CHANNEL_RESERVED_FACTOR = 8;75constexpr uint32_t PER_CHANNEL_RESERVED_FACTOR = 8;
76// perchannel float buffer 数量：col_max、col_min、scale、offset76// perchannel float buffer 数量：col_max、col_min、scale、offset
77constexpr uint32_t PER_CHANNEL_FLOAT_BUF_NUM = 4;77constexpr uint32_t PER_CHANNEL_FLOAT_BUF_NUM = 4;
78+// row 对齐到 16 字节
79+constexpr uint32_t ROW_ALIGN_SIZE = 16;
80+// UB 队列平滑 buffer 系数
81+constexpr uint32_t SMOOTH_QUEUE_FACTOR = 2;
82+// UB 队列主 buffer 系数
83+constexpr uint32_t MAIN_QUEUE_FACTOR = 4;
84+// INT4 偶数对齐位掩码
85+constexpr uint32_t EVEN_ALIGN_MASK = 1;
86+// batch mode 调度模式
87+constexpr uint32_t BATCH_SCHEDULE_MODE = 1;
78 88 
79constexpr float HIFLOAT8_MAX_VALUE = 32768.0;89constexpr float HIFLOAT8_MAX_VALUE = 32768.0;
80constexpr float FLT_EPSILON = 1e-6f;90constexpr float FLT_EPSILON = 1e-6f;
@@ -540,7 +550,7 @@ void DynamicQuantRegbaseTiling::CalculateCoreNum(const gert::TilingContext* cont
540void DynamicQuantRegbaseTiling::CalculateTilingData()550void DynamicQuantRegbaseTiling::CalculateTilingData()
541{551{
542 // 每个row的元素数量对齐16552 // 每个row的元素数量对齐16
543- uint32_t alignedRowLen = AlignUp<16>(rowLen);553+ uint32_t alignedRowLen = AlignUp<ROW_ALIGN_SIZE>(rowLen);
544 uint64_t maxUseUbSize = ubSize - RESERVED_LENGTH;554 uint64_t maxUseUbSize = ubSize - RESERVED_LENGTH;
545 uint32_t smoothBuffer = (hasSmooth ? 1UL : 0UL);555 uint32_t smoothBuffer = (hasSmooth ? 1UL : 0UL);
546 uint32_t offsetBuffer = 1UL;556 uint32_t offsetBuffer = 1UL;
@@ -562,9 +572,9 @@ void DynamicQuantRegbaseTiling::CalculateTilingData()
562 CalculateTilingForPertenLargeMulticore(maxUseUbSize, smoothBuffer);572 CalculateTilingForPertenLargeMulticore(maxUseUbSize, smoothBuffer);
563 return;573 return;
564 }574 }
565- uint64_t calcLargeBuf = 2UL * (static_cast<uint64_t>(smoothBuffer) * 2UL +575+ uint64_t calcLargeBuf = DOUBLE * (static_cast<uint64_t>(smoothBuffer) * SMOOTH_QUEUE_FACTOR +
566- 4UL); // for inQueue,outQueue&smoothQueue DB576+ MAIN_QUEUE_FACTOR); // for inQueue,outQueue&smoothQueue DB
567- maxUseUbSize -= RESERVED_LENGTH; // for scaleQueue DB577+ maxUseUbSize -= RESERVED_LENGTH; // for scaleQueue DB
568 useDb = true;578 useDb = true;
569 innerLoopEle = static_cast<uint32_t>(maxUseUbSize) / static_cast<uint32_t>(calcLargeBuf) / FLOAT_NUM_ONE_RPT *579 innerLoopEle = static_cast<uint32_t>(maxUseUbSize) / static_cast<uint32_t>(calcLargeBuf) / FLOAT_NUM_ONE_RPT *
570 FLOAT_NUM_ONE_RPT;580 FLOAT_NUM_ONE_RPT;
@@ -720,8 +730,8 @@ void DynamicQuantRegbaseTiling::CalculateTilingForPertenLargeMulticore(uint64_t
720 730 
721 // 每个 token 分配的核数（至少 2，由前置条件 rowNum <= vectorCoreNum/2 保证）731 // 每个 token 分配的核数（至少 2，由前置条件 rowNum <= vectorCoreNum/2 保证）
722 uint32_t coreNumPerToken = vectorCoreNum / static_cast<uint32_t>(rowNum);732 uint32_t coreNumPerToken = vectorCoreNum / static_cast<uint32_t>(rowNum);
723- if (coreNumPerToken < 2) {733+ if (coreNumPerToken < DOUBLE) {
724- coreNumPerToken = 2;734+ coreNumPerToken = DOUBLE;
725 }735 }
726 736 
727 // 每个核处理的 rowLen 切片大小737 // 每个核处理的 rowLen 切片大小
@@ -730,13 +740,13 @@ void DynamicQuantRegbaseTiling::CalculateTilingForPertenLargeMulticore(uint64_t
730 740 
731 if (yDtype == ge::DT_INT4) {741 if (yDtype == ge::DT_INT4) {
732 if ((elePerTailCore & 1) != 0) {742 if ((elePerTailCore & 1) != 0) {
733- elePerTailCore = elePerTailCore & ~1U;743+ elePerTailCore = elePerTailCore & ~EVEN_ALIGN_MASK;
734 }744 }
735 uint32_t remaining = rowLen - elePerTailCore * coreNumPerToken;745 uint32_t remaining = rowLen - elePerTailCore * coreNumPerToken;
736 if (remaining == 0) {746 if (remaining == 0) {
737 elePerHeadCore = elePerTailCore;747 elePerHeadCore = elePerTailCore;
738 } else {748 } else {
739- elePerHeadCore = elePerTailCore + 2;749+ elePerHeadCore = elePerTailCore + DOUBLE;
740 }750 }
741 }751 }
742 752 
@@ -752,7 +762,7 @@ void DynamicQuantRegbaseTiling::CalculateTilingForPertenLargeMulticore(uint64_t
752 coreNum = coreNumPerToken * static_cast<uint32_t>(rowNum);762 coreNum = coreNumPerToken * static_cast<uint32_t>(rowNum);
753 763 
754 // UB 内循环切分参数764 // UB 内循环切分参数
755- uint64_t calcBuf = 2UL * (static_cast<uint64_t>(smoothBuffer) * 2UL + 4UL);765+ uint64_t calcBuf = DOUBLE * (static_cast<uint64_t>(smoothBuffer) * SMOOTH_QUEUE_FACTOR + MAIN_QUEUE_FACTOR);
756 uint64_t calcMaxUb = maxUseUbSize;766 uint64_t calcMaxUb = maxUseUbSize;
757 calcMaxUb -= RESERVED_LENGTH;767 calcMaxUb -= RESERVED_LENGTH;
758 768 
@@ -908,7 +918,7 @@ ge::graphStatus DynamicQuantRegbaseTiling::RunFusionKernelTiling(gert::TilingCon
908 if (quantMode_ == TPL_PER_TENSOR_FULL_LOAD || quantMode_ == TPL_PER_TENSOR_LARGE_SHAPE ||918 if (quantMode_ == TPL_PER_TENSOR_FULL_LOAD || quantMode_ == TPL_PER_TENSOR_LARGE_SHAPE ||
909 quantMode_ == TPL_MOE_PER_TENSOR_FULL_LOAD || quantMode_ == TPL_MOE_PER_TENSOR_LARGE_SHAPE ||919 quantMode_ == TPL_MOE_PER_TENSOR_FULL_LOAD || quantMode_ == TPL_MOE_PER_TENSOR_LARGE_SHAPE ||
910 quantMode_ == TPL_PERTEN_LARGE_MULTICORE) {920 quantMode_ == TPL_PERTEN_LARGE_MULTICORE) {
911- context->SetScheduleMode(1); // 设置为batch mode模式，所有核同时启动921+ context->SetScheduleMode(BATCH_SCHEDULE_MODE); // 设置为batch mode模式，所有核同时启动
912 }922 }
913 context->SetBlockDim(coreNum);923 context->SetBlockDim(coreNum);
914 return ge::GRAPH_SUCCESS;924 return ge::GRAPH_SUCCESS;