Star1406
1494
代码
代码
Issues148
Pull Requests343
Actions
项目讨论
Wiki
分析
Star1406
1494
已合并
fix cleancode include：assign,mem_set,split_v,tensor_move,adds,power #5264
csz创建于 18 天前
fix cleancode include：assign,mem_set,split_v,tensor_move,adds,power #5264
已合并
csz创建于 18 天前
从chengzhi1120/ops-math0831合入到cann/ops-mathmaster
共 7 个文件变更+114-201
Mconversion/assign/op_host/arch35/assign_tiling_arch35.cpp+1-1
@@ -185,7 +185,7 @@ static ge::graphStatus AssignTilingForAscendC(gert::TilingContext* context)
185 OP_CHECK_IF(CheckShapeForAssign(context, tilingParam) != ge::GRAPH_SUCCESS,185 OP_CHECK_IF(CheckShapeForAssign(context, tilingParam) != ge::GRAPH_SUCCESS,
186 OP_LOGE(context->GetNodeName(), "The shape check failed."), return ge::GRAPH_FAILED);186 OP_LOGE(context->GetNodeName(), "The shape check failed."), return ge::GRAPH_FAILED);
187 187 
188- auto compileInfo = reinterpret_cast<const AssignCompileInfo*>(context->GetCompileInfo());188+ auto compileInfo = context->GetCompileInfo<AssignCompileInfo>();
189 OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo);189 OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo);
190 190 
191 tilingParam.totalCoreNum = compileInfo->coreNum;191 tilingParam.totalCoreNum = compileInfo->coreNum;
Mconversion/mem_set/op_host/arch35/mem_set_tiling_arch35.cpp+19-70
@@ -29,6 +29,9 @@ std::set<ge::DataType> SUPPORT_TYPE_LIST = {ge::DT_INT8, ge::DT_INT32, ge::DT_
29std::set<ge::DataType> SUPPORT_TYPE_INT_LIST = {ge::DT_INT8, ge::DT_INT32, ge::DT_UINT8, ge::DT_INT16,29std::set<ge::DataType> SUPPORT_TYPE_INT_LIST = {ge::DT_INT8, ge::DT_INT32, ge::DT_UINT8, ge::DT_INT16,
30 ge::DT_UINT16, ge::DT_UINT32, ge::DT_INT64, ge::DT_UINT64};30 ge::DT_UINT16, ge::DT_UINT32, ge::DT_INT64, ge::DT_UINT64};
31constexpr uint64_t BLOCK_SIZE = 512;31constexpr uint64_t BLOCK_SIZE = 512;
32+constexpr int TILING_COUNTS[] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, 64, 128, 192, 256};
33+constexpr size_t TILING_COUNTS_NUM = sizeof(TILING_COUNTS) / sizeof(TILING_COUNTS[0]);
34+constexpr int MAX_TILING_COUNT = TILING_COUNTS[TILING_COUNTS_NUM - 1];
32 35 
33template <uint16_t Count>36template <uint16_t Count>
34void MemSetTilingClass::CheckTilingData(MemSetTilingData<Count>* tilingDataPostAtr)37void MemSetTilingClass::CheckTilingData(MemSetTilingData<Count>* tilingDataPostAtr)
@@ -69,82 +72,28 @@ void MemSetTilingClass::PostDo()
69 CheckTilingData<Count>(tilingDataPostAtr);72 CheckTilingData<Count>(tilingDataPostAtr);
70}73}
71 74 
75+template <size_t Idx>
76+void MemSetTilingClass::DispatchPostDo(size_t targetIdx)
77+{
78+ if (targetIdx == Idx) {
79+ PostDo<static_cast<uint16_t>(TILING_COUNTS[Idx])>();
80+ return;
81+ }
82+ if constexpr (Idx + 1 < TILING_COUNTS_NUM) {
83+ DispatchPostDo<Idx + 1>(targetIdx);
84+ }
85+}
86+ 
72ge::graphStatus MemSetTilingClass::PostTiling()87ge::graphStatus MemSetTilingClass::PostTiling()
73{88{
74- // Different sizes of tiling data templates89+ if (inputCount_ > MAX_TILING_COUNT) {
75- const std::vector<int> validNums = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, 64, 128, 192, 256};
76- if (inputCount_ > validNums.back()) {
77 OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(context_->GetNodeName(), "TensorNum", std::to_string(inputCount_).c_str(),90 OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(context_->GetNodeName(), "TensorNum", std::to_string(inputCount_).c_str(),
78 "The value of TensorNum must be within the range [1, 256].");91 "The value of TensorNum must be within the range [1, 256].");
79 return ge::GRAPH_FAILED;92 return ge::GRAPH_FAILED;
80 }93 }
81- auto it = std::lower_bound(validNums.begin(), validNums.end(), inputCount_);94+ auto it = std::lower_bound(TILING_COUNTS, TILING_COUNTS + TILING_COUNTS_NUM, static_cast<int>(inputCount_));
82- int targetNum = *it;95+ size_t targetIdx = static_cast<size_t>(std::distance(TILING_COUNTS, it));
83- switch (targetNum) {96+ DispatchPostDo<0>(targetIdx);
84- case 1:
85- PostDo<1>();
86- break;
87- case 2:
88- PostDo<2>();
89- break;
90- case 3:
91- PostDo<3>();
92- break;
93- case 4:
94- PostDo<4>();
95- break;
96- case 5:
97- PostDo<5>();
98- break;
99- case 6:
100- PostDo<6>();
101- break;
102- case 7:
103- PostDo<7>();
104- break;
105- case 8:
106- PostDo<8>();
107- break;
108- case 9:
109- PostDo<9>();
110- break;
111- case 10:
112- PostDo<10>();
113- break;
114- case 11:
115- PostDo<11>();
116- break;
117- case 12:
118- PostDo<12>();
119- break;
120- case 13:
121- PostDo<13>();
122- break;
123- case 14:
124- PostDo<14>();
125- break;
126- case 15:
127- PostDo<15>();
128- break;
129- case 16:
130- PostDo<16>();
131- break;
132- case 32:
133- PostDo<32>();
134- break;
135- case 64:
136- PostDo<64>();
137- break;
138- case 128:
139- PostDo<128>();
140- break;
141- case 196:
142- PostDo<196>();
143- break;
144- case 256:
145- PostDo<256>();
146- break;
147- }
148 context_->SetBlockDim(aicoreParams_.numBlocks);97 context_->SetBlockDim(aicoreParams_.numBlocks);
149 return ge::GRAPH_SUCCESS;98 return ge::GRAPH_SUCCESS;
150}99}
Mconversion/mem_set/op_host/arch35/mem_set_tiling_arch35.h+26-41
@@ -22,7 +22,6 @@
22#include "platform/platform_ascendc.h"22#include "platform/platform_ascendc.h"
23#include "register/op_impl_registry.h"23#include "register/op_impl_registry.h"
24 24 
25- 
26namespace optiling {25namespace optiling {
27 26 
28struct MemSetCompileInfoArch35 {27struct MemSetCompileInfoArch35 {
@@ -32,26 +31,16 @@ struct MemSetCompileInfoArch35 {
32 31 
33class MemSetTilingClass : public Ops::Base::TilingBaseClass {32class MemSetTilingClass : public Ops::Base::TilingBaseClass {
34public:33public:
35- explicit MemSetTilingClass(gert::TilingContext* context) : Ops::Base::TilingBaseClass(context)34+ explicit MemSetTilingClass(gert::TilingContext* context) : Ops::Base::TilingBaseClass(context) {}
36- {}
37 35 
38- void Reset(gert::TilingContext* context) override36+ void Reset(gert::TilingContext* context) override { Ops::Base::TilingBaseClass::Reset(context); }
39- {
40- Ops::Base::TilingBaseClass::Reset(context);
41- }
42 37 
43protected:38protected:
44 ge::graphStatus GetShapeAttrsInfo() override;39 ge::graphStatus GetShapeAttrsInfo() override;
45 ge::graphStatus GetPlatformInfo() override;40 ge::graphStatus GetPlatformInfo() override;
46- bool IsCapable() override41+ bool IsCapable() override { return true; }
47- {
48- return true;
49- }
50 ge::graphStatus DoOpTiling() override;42 ge::graphStatus DoOpTiling() override;
51- ge::graphStatus DoLibApiTiling() override43+ ge::graphStatus DoLibApiTiling() override { return ge::GRAPH_SUCCESS; }
52- {
53- return ge::GRAPH_SUCCESS;
54- }
55 ge::graphStatus GetWorkspaceSize() override;44 ge::graphStatus GetWorkspaceSize() override;
56 ge::graphStatus PostTiling() override;45 ge::graphStatus PostTiling() override;
57 uint64_t GetTilingKey() const override;46 uint64_t GetTilingKey() const override;
@@ -63,22 +52,24 @@ private:
63 void PostDo();52 void PostDo();
64 template <uint16_t Count>53 template <uint16_t Count>
65 void CheckTilingData(MemSetTilingData<Count>* tilingDataPostAtr);54 void CheckTilingData(MemSetTilingData<Count>* tilingDataPostAtr);
55+ template <size_t Idx>
56+ void DispatchPostDo(size_t targetIdx);
66 57 
67-private:58+private:
68- void* tilingDataPostAtr_ = nullptr;59+ void* tilingDataPostAtr_ = nullptr;
69- std::vector<int64_t> perCoreSizes_;60+ std::vector<int64_t> perCoreSizes_;
70- std::vector<int64_t> lastCoreSizes_;61+ std::vector<int64_t> lastCoreSizes_;
71- std::vector<int64_t> intValue_;62+ std::vector<int64_t> intValue_;
72- std::vector<float> floatValue_;63+ std::vector<float> floatValue_;
73- std::vector<int16_t> listType_;64+ std::vector<int16_t> listType_;
74- std::vector<int16_t> useCore_;65+ std::vector<int16_t> useCore_;
75- std::vector<int64_t> sizes_;66+ std::vector<int64_t> sizes_;
76- uint16_t cacheLineSize_ = 0;67+ uint16_t cacheLineSize_ = 0;
77- uint16_t needCore_ = 0;68+ uint16_t needCore_ = 0;
78- int halfUbSize_ = 0;69+ int halfUbSize_ = 0;
79- uint16_t inputCount_ = 0;70+ uint16_t inputCount_ = 0;
80- uint16_t TilingKey_ = 0;71+ uint16_t TilingKey_ = 0;
81- bool isDynamic_ = false;72+ bool isDynamic_ = false;
82};73};
83 74 
84} // namespace optiling75} // namespace optiling
@@ -90,22 +81,16 @@ public:
90 * Get workspaceSize list81 * Get workspaceSize list
91 * @return workspaceSize list82 * @return workspaceSize list
92 */83 */
93- const gert::ContinuousVector* GetCleanWorkspaceSizes() const84+ const gert::ContinuousVector* GetCleanWorkspaceSizes() const { return GetInputPointer<gert::ContinuousVector>(0); }
94- {
95- return GetInputPointer<gert::ContinuousVector>(0);
96- }
97 85 
98 /**86 /**
99 * Get the size of the output memory to be cleared by the output index87 * Get the size of the output memory to be cleared by the output index
100 * @param index output index88 * @param index output index
101 * @return Size of output memory to be cleaned89 * @return Size of output memory to be cleaned
102 */90 */
103- uint64_t GetCleanOutputSize(size_t index) const91+ uint64_t GetCleanOutputSize(size_t index) const { return GetInputValue<uint64_t>(index + 1U); }
104- {
105- return GetInputValue<uint64_t>(index + 1U);
106- }
107};92};
108-static_assert(93+static_assert(std::is_standard_layout<AtomicCleanTilingContext>::value,
109- std::is_standard_layout<AtomicCleanTilingContext>::value, "The class AtomicCleanTilingContext must be a POD");94+ "The class AtomicCleanTilingContext must be a POD");
110} // namespace ops95} // namespace ops
111-#endif96+#endif
Mconversion/split_v/op_host/arch35/split_v_tiling_arch35.cpp+16-18
@@ -74,6 +74,14 @@ constexpr int64_t SIMT_CHUNK_TOTAL_ELEM_THRESHOLD = 131072; // 1024 block * 128
74constexpr int64_t SIMT_REJECT_RELATIVE_SKEW_UPPER = 16; // SIMT拒绝后回退Chunk的相对偏差上界74constexpr int64_t SIMT_REJECT_RELATIVE_SKEW_UPPER = 16; // SIMT拒绝后回退Chunk的相对偏差上界
75constexpr int64_t SIMT_REJECT_ABSOLUTE_SKEW_UPPER = 2048; // SIMT拒绝后回退Chunk的绝对偏差上界75constexpr int64_t SIMT_REJECT_ABSOLUTE_SKEW_UPPER = 2048; // SIMT拒绝后回退Chunk的绝对偏差上界
76 76 
77+// IsDoSplitVSIMT 强制走 SIMT 的特定输入 shape [256, 16] 及 numSplit 上限
78+constexpr size_t SIMT_FORCE_INPUT_DIM_NUM = 2;
79+constexpr int64_t SIMT_FORCE_INPUT_DIM0 = 256;
80+constexpr int64_t SIMT_FORCE_INPUT_DIM1 = 16;
81+constexpr int64_t SIMT_FORCE_MAX_NUM_SPLIT = 16;
82+// host 侧 uint64 fast-div 计算中的 uint64 位宽
83+constexpr int64_t UINT64_BIT_WIDTH = 64;
84+ 
77template <typename T>85template <typename T>
78std::string SplitVTiling::ArrayToString(const T* vec, size_t num) const86std::string SplitVTiling::ArrayToString(const T* vec, size_t num) const
79{87{
@@ -893,8 +901,8 @@ bool SplitVTiling::IsDoSplitVSIMT(int32_t maxCoreNum)
893 if (isSameLenMode_) {901 if (isSameLenMode_) {
894 return false;902 return false;
895 }903 }
896- if (inputShape_.GetDimNum() == 2 && inputShape_.GetDim(0) == 256 && inputShape_.GetDim(1) == 16 && splitDim_ == 1 &&904+ if (inputShape_.GetDimNum() == SIMT_FORCE_INPUT_DIM_NUM && inputShape_.GetDim(0) == SIMT_FORCE_INPUT_DIM0 &&
897- numSplit_ <= 16) {905+ inputShape_.GetDim(1) == SIMT_FORCE_INPUT_DIM1 && splitDim_ == 1 && numSplit_ <= SIMT_FORCE_MAX_NUM_SPLIT) {
898 CalInputDataSize();906 CalInputDataSize();
899 OP_LOGI(context_->GetNodeName(), "Force SIMT for shape [256, 16]");907 OP_LOGI(context_->GetNodeName(), "Force SIMT for shape [256, 16]");
900 return true;908 return true;
@@ -1034,19 +1042,19 @@ static inline void HostGetUintDivMagicAndShift64(uint64_t& magic, uint32_t& shif
1034 shift = 0;1042 shift = 0;
1035 return;1043 return;
1036 }1044 }
1037- // pos = 64 - clz(divisor) = floor(log2(divisor)) + 11045+ // pos = UINT64_BIT_WIDTH - clz(divisor) = floor(log2(divisor)) + 1
1038- int64_t pos = 64 - static_cast<int64_t>(__builtin_clzll(divisor));1046+ int64_t pos = UINT64_BIT_WIDTH - static_cast<int64_t>(__builtin_clzll(divisor));
1039 int64_t cnt1 = static_cast<int64_t>(__builtin_popcountll(divisor));1047 int64_t cnt1 = static_cast<int64_t>(__builtin_popcountll(divisor));
1040 shift = (cnt1 == 1) ? static_cast<uint32_t>(pos - 1) : static_cast<uint32_t>(pos);1048 shift = (cnt1 == 1) ? static_cast<uint32_t>(pos - 1) : static_cast<uint32_t>(pos);
1041 uint64_t dividend = 0;1049 uint64_t dividend = 0;
1042- if (shift < 64) {1050+ if (shift < UINT64_BIT_WIDTH) {
1043 dividend = (1ULL << shift) - divisor;1051 dividend = (1ULL << shift) - divisor;
1044 } else {1052 } else {
1045- // shift == 64, 2^64 - divisor1053+ // shift == UINT64_BIT_WIDTH, 2^64 - divisor
1046 dividend = 0xFFFFFFFFFFFFFFFFULL - divisor + 1;1054 dividend = 0xFFFFFFFFFFFFFFFFULL - divisor + 1;
1047 }1055 }
1048 // magic = floor(2^64 * dividend / divisor) + 11056 // magic = floor(2^64 * dividend / divisor) + 1
1049- __uint128_t num = (static_cast<__uint128_t>(dividend) << 64);1057+ __uint128_t num = (static_cast<__uint128_t>(dividend) << UINT64_BIT_WIDTH);
1050 magic = static_cast<uint64_t>(num / divisor) + 1;1058 magic = static_cast<uint64_t>(num / divisor) + 1;
1051}1059}
1052 1060 
@@ -1103,12 +1111,9 @@ void SplitVTiling::DoSplitVSIMTChunkPreTiling(int32_t maxCoreNum)
1103 // diff >= 0 (欠分配): 差值全部补到拷贝量最大的 output1111 // diff >= 0 (欠分配): 差值全部补到拷贝量最大的 output
1104 // diff < 0 (超分配): 按拷贝量降序逐个扣减, 每个非空 output 保留至少 1 块, 避免下游 q=total/blkCnt 除零1112 // diff < 0 (超分配): 按拷贝量降序逐个扣减, 每个非空 output 保留至少 1 块, 避免下游 q=total/blkCnt 除零
1105 if (maxCopyIdx >= 0) {1113 if (maxCopyIdx >= 0) {
1106- // int32_t diff = logicBlockNum_ - blkSum;
1107 int32_t diff = maxLogicBlockByCore - blkSum;1114 int32_t diff = maxLogicBlockByCore - blkSum;
1108 if (diff >= 0) {1115 if (diff >= 0) {
1109- // blockPrefixBuf_[maxCopyIdx] = static_cast<uint32_t>(static_cast<int32_t>(blockPrefixBuf_[maxCopyIdx]) +1116+ realCoreNum_ = CeilDiv(blkSum, LOGIC_BLOCK_PER_CORE);
1110- // diff);
1111- realCoreNum_ = CeilDiv(blkSum, 16);
1112 } else {1117 } else {
1113 int32_t need = -diff;1118 int32_t need = -diff;
1114 // 收集可扣减的输出(blk > 1), 按拷贝量 M*wi 降序排序后贪心扣减1119 // 收集可扣减的输出(blk > 1), 按拷贝量 M*wi 降序排序后贪心扣减
@@ -1175,7 +1180,6 @@ void SplitVTiling::DoSplitVSIMTChunkPreTiling(int32_t maxCoreNum)
1175 }1180 }
1176 1181 
1177 // 5) 1024 逻辑 block = 64 核 * 161182 // 5) 1024 逻辑 block = 64 核 * 16
1178- // realCoreNum_ = std::min(maxCoreNum, SIMT_VECTOR_CORE_NUM);
1179 blockDim_ = realCoreNum_;1183 blockDim_ = realCoreNum_;
1180 tilingKey_ = SIMT_CHUNK_KEY;1184 tilingKey_ = SIMT_CHUNK_KEY;
1181 1185 
@@ -1226,11 +1230,6 @@ ge::graphStatus SplitVTiling::DoSplitVTiling(int32_t maxCoreNum, uint32_t ubSize
1226 if (IsDoSplitVSIMT(maxCoreNum)) {1230 if (IsDoSplitVSIMT(maxCoreNum)) {
1227 return DoSplitVSIMTTiling(maxCoreNum);1231 return DoSplitVSIMTTiling(maxCoreNum);
1228 }1232 }
1229- // if (IsFallbackToChunk()) {
1230- // DoSplitVSIMTChunkPreTiling(maxCoreNum);
1231- // // DoSplitVSIMTChunkTiling(maxCoreNum);
1232- // return ge::GRAPH_SUCCESS;
1233- // }
1234 if (numSplit_ != 0 && numSplit_ != 1 && isSameLenMode_) {1233 if (numSplit_ != 0 && numSplit_ != 1 && isSameLenMode_) {
1235 // SamLenMode fuse to [M G N]1234 // SamLenMode fuse to [M G N]
1236 FuseInputShapeSameLen();1235 FuseInputShapeSameLen();
@@ -1476,7 +1475,6 @@ void SplitVTiling::CalcSameLenSplitTilingInfo(int64_t halfUbEleNum, int64_t bloc
1476 int64_t newGBF = 0, newGBCount = 0, newGBFT = 0;1475 int64_t newGBF = 0, newGBCount = 0, newGBFT = 0;
1477 int64_t newTiles = 0;1476 int64_t newTiles = 0;
1478 bool smallGAvail = FloorAlign(factorTmp, blockEleNum) >= numSplit_;1477 bool smallGAvail = FloorAlign(factorTmp, blockEleNum) >= numSplit_;
1479- 
1480 if (smallGAvail) {1478 if (smallGAvail) {
1481 newGBF = numSplit_;1479 newGBF = numSplit_;
1482 int64_t tmp = FloorAlign(halfUbEleNum / fusedShape_[1] / newGBF, blockEleNum);1480 int64_t tmp = FloorAlign(halfUbEleNum / fusedShape_[1] / newGBF, blockEleNum);
Mconversion/tensor_move/op_host/arch35/tensor_move_tiling_arch35.cpp+1-1
@@ -186,7 +186,7 @@ static ge::graphStatus Tiling4TensorMove(gert::TilingContext* context)
186 OP_CHECK_IF(CheckTensorMoveShape(context) != ge::GRAPH_SUCCESS,186 OP_CHECK_IF(CheckTensorMoveShape(context) != ge::GRAPH_SUCCESS,
187 OP_LOGE(context->GetNodeName(), "The shape check failed."), return ge::GRAPH_FAILED);187 OP_LOGE(context->GetNodeName(), "The shape check failed."), return ge::GRAPH_FAILED);
188 188 
189- auto compileInfo = reinterpret_cast<const TensorMoveCompileInfo*>(context->GetCompileInfo());189+ auto compileInfo = context->GetCompileInfo<TensorMoveCompileInfo>();
190 OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo);190 OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo);
191 191 
192 TensorMoveTilingParam tilingParam;192 TensorMoveTilingParam tilingParam;
Mmath/adds/op_kernel/arch35/adds_dag.h+5-12
@@ -16,32 +16,25 @@
16 *16 *
17 * 数据流（bf16/fp16/int16/int32/int64）:17 * 数据流（bf16/fp16/int16/int32/int64）:
18 * input (GM) -> CopyIn<T> -> Cast<T,float> -> Adds<float,scalar> -> Cast<float,T> -> CopyOut<T> -> output (GM)18 * input (GM) -> CopyIn<T> -> Cast<T,float> -> Adds<float,scalar> -> Cast<float,T> -> CopyOut<T> -> output (GM)
19- * 19+ *
20 * 数据流（fp32）:20 * 数据流（fp32）:
21 * input (GM) -> CopyIn<float> -> Adds<float,scalar> -> CopyOut<float> -> output (GM)21 * input (GM) -> CopyIn<float> -> Adds<float,scalar> -> CopyOut<float> -> output (GM)
22 * （Cast fp32→fp32会被模板优化掉）22 * （Cast fp32→fp32会被模板优化掉）
23- * 23+ *
24 * 标量参数通过 Placeholder::Var<float, 0> 占位，Kernel 入口通过 SetVar<float, 0>(value) 注入24 * 标量参数通过 Placeholder::Var<float, 0> 占位，Kernel 入口通过 SetVar<float, 0>(value) 注入
25 */25 */
26 26 
27#ifndef ADDS_DAG_H27#ifndef ADDS_DAG_H
28#define ADDS_DAG_H28#define ADDS_DAG_H
29 29 
30-// Host 编译时 mock __aicore__（Kernel 编译器已内置定义）
31-#ifndef __CCE_AICORE__
32-#ifndef __aicore__
33-#define __aicore__
34-#endif
35-#endif
36- 
37#include "atvoss/util/dag.h"30#include "atvoss/util/dag.h"
38#include "atvoss/util/vec.h"31#include "atvoss/util/vec.h"
39#include "atvoss/util/placeholder.h"32#include "atvoss/util/placeholder.h"
40 33 
41-using namespace Ops::Base;
42- 
43namespace NsAdds {34namespace NsAdds {
44 35 
36+using namespace Ops::Base;
37+ 
45constexpr int CAST_MODE_NONE = 0;38constexpr int CAST_MODE_NONE = 0;
46constexpr int CAST_MODE_RINT = 1;39constexpr int CAST_MODE_RINT = 1;
47constexpr int CAST_MODE_ROUND = 4;40constexpr int CAST_MODE_ROUND = 4;
@@ -88,4 +81,4 @@ struct AddsOp {
88 81 
89} // namespace NsAdds82} // namespace NsAdds
90 83 
91-#endif // ADDS_DAG_H84+#endif // ADDS_DAG_H
Mmath/power/op_host/arch35/power_tiling_arch35.cpp+46-58
@@ -38,6 +38,8 @@ static const size_t POWER_ASCEND_WORKSPACE = 16 * 1024 * 1024;
38// isclose 判等使用的绝对/相对容差，与 math/is_close 默认参数对齐。38// isclose 判等使用的绝对/相对容差，与 math/is_close 默认参数对齐。
39static constexpr float POWER_ISCLOSE_ATOL = 1e-8f;39static constexpr float POWER_ISCLOSE_ATOL = 1e-8f;
40static constexpr float POWER_ISCLOSE_RTOL = 1e-5f;40static constexpr float POWER_ISCLOSE_RTOL = 1e-5f;
41+// 整数幂奇偶性判断的模数: intPow % 2 决定 (-1)^intPow 的符号
42+static constexpr int POWER_PARITY_MODULUS = 2;
41 43 
42// 浮点近似相等：|a - b| <= atol + rtol * |b|。44// 浮点近似相等：|a - b| <= atol + rtol * |b|。
43// 用于在 host 端容忍浮点误差地判断 power/scale/shift 是否落到 0/1/2/3 等关键点。45// 用于在 host 端容忍浮点误差地判断 power/scale/shift 是否落到 0/1/2/3 等关键点。
@@ -65,9 +67,7 @@ ge::graphStatus PowerTiling::CalcInputDtype()
65 OP_CHECK_IF(67 OP_CHECK_IF(
66 this->inputDtype != ge::DT_FLOAT16 && this->inputDtype != ge::DT_BF16 && this->inputDtype != ge::DT_FLOAT,68 this->inputDtype != ge::DT_FLOAT16 && this->inputDtype != ge::DT_BF16 && this->inputDtype != ge::DT_FLOAT,
67 OP_LOGE_FOR_INVALID_DTYPE_WITH_REASON(69 OP_LOGE_FOR_INVALID_DTYPE_WITH_REASON(
68- tilingContext->GetNodeName(),70+ tilingContext->GetNodeName(), "x", Ops::Base::ToString(this->inputDtype).c_str(),
69- "x",
70- Ops::Base::ToString(this->inputDtype).c_str(),
71 "The dtype of x must be within the range [DT_FLOAT16, DT_BF16, DT_FLOAT]."),71 "The dtype of x must be within the range [DT_FLOAT16, DT_BF16, DT_FLOAT]."),
72 return ge::GRAPH_FAILED);72 return ge::GRAPH_FAILED);
73 return ge::GRAPH_SUCCESS;73 return ge::GRAPH_SUCCESS;
@@ -79,14 +79,12 @@ ge::graphStatus PowerTiling::CalcOutputDtype()
79 auto outputDesc = tilingContext->GetOutputDesc(0);79 auto outputDesc = tilingContext->GetOutputDesc(0);
80 OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputDesc);80 OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputDesc);
81 this->outputDtype = outputDesc->GetDataType();81 this->outputDtype = outputDesc->GetDataType();
82- OP_CHECK_IF(82+ OP_CHECK_IF(this->outputDtype != this->inputDtype,
83- this->outputDtype != this->inputDtype,83+ OP_LOGE_FOR_INVALID_DTYPES_WITH_REASON(
84- OP_LOGE_FOR_INVALID_DTYPES_WITH_REASON(84+ tilingContext->GetNodeName(), "x, y",
85- tilingContext->GetNodeName(),85+ (Ops::Base::ToString(this->inputDtype) + ", " + Ops::Base::ToString(this->outputDtype)).c_str(),
86- "x, y",86+ "The dtypes of x and y must be the same."),
87- (Ops::Base::ToString(this->inputDtype) + ", " + Ops::Base::ToString(this->outputDtype)).c_str(),87+ return ge::GRAPH_FAILED);
88- "The dtypes of x and y must be the same."),
89- return ge::GRAPH_FAILED);
90 return ge::GRAPH_SUCCESS;88 return ge::GRAPH_SUCCESS;
91}89}
92 90 
@@ -102,14 +100,12 @@ ge::graphStatus PowerTiling::CheckShape()
102 OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputStorageShape);100 OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputStorageShape);
103 const gert::Shape& yShape = Ops::Base::EnsureNotScalar(outputStorageShape->GetStorageShape());101 const gert::Shape& yShape = Ops::Base::EnsureNotScalar(outputStorageShape->GetStorageShape());
104 102 
105- OP_CHECK_IF(103+ OP_CHECK_IF(xShape != yShape,
106- xShape != yShape,104+ OP_LOGE_FOR_INVALID_SHAPES_WITH_REASON(
107- OP_LOGE_FOR_INVALID_SHAPES_WITH_REASON(105+ tilingContext->GetNodeName(), "x, y",
108- tilingContext->GetNodeName(),106+ (Ops::Base::ToString(xShape) + ", " + Ops::Base::ToString(yShape)).c_str(),
109- "x, y",107+ "The shapes of x and y must be the same."),
110- (Ops::Base::ToString(xShape) + ", " + Ops::Base::ToString(yShape)).c_str(),108+ return ge::GRAPH_FAILED);
111- "The shapes of x and y must be the same."),
112- return ge::GRAPH_FAILED);
113 return ge::GRAPH_SUCCESS;109 return ge::GRAPH_SUCCESS;
114}110}
115 111 
@@ -129,18 +125,17 @@ ge::graphStatus PowerTiling::SetAttr()
129 const float* shiftPtr = attrs->GetAttrPointer<float>(2);125 const float* shiftPtr = attrs->GetAttrPointer<float>(2);
130 this->attrShift = shiftPtr != nullptr ? *shiftPtr : 0.0f;126 this->attrShift = shiftPtr != nullptr ? *shiftPtr : 0.0f;
131 127 
132- OP_LOGD(128+ OP_LOGD(tilingContext->GetNodeName(), "Power attrs: power=%f, scale=%f, shift=%f", this->attrPower, this->attrScale,
133- tilingContext->GetNodeName(), "Power attrs: power=%f, scale=%f, shift=%f",129+ this->attrShift);
134- this->attrPower, this->attrScale, this->attrShift);
135 return ge::GRAPH_SUCCESS;130 return ge::GRAPH_SUCCESS;
136}131}
137 132 
138// -----------------------------------------------------------------------------133// -----------------------------------------------------------------------------
139// 模板辅助函数：根据 dtype 分发 DoTiling 调用（单模板参数 DAG）。134// 模板辅助函数：根据 dtype 分发 DoTiling 调用（单模板参数 DAG）。
140// -----------------------------------------------------------------------------135// -----------------------------------------------------------------------------
141-template<template<typename> class DagT>136+template <template <typename> class DagT>
142-ge::graphStatus DispatchTilingByDtype(137+ge::graphStatus DispatchTilingByDtype(ElewiseBaseTiling& tiling, ge::DataType dtype,
143- ElewiseBaseTiling& tiling, ge::DataType dtype, Ops::Base::EleBaseTilingData& baseTiling)138+ Ops::Base::EleBaseTilingData& baseTiling)
144{139{
145 if (dtype == ge::DT_FLOAT16) {140 if (dtype == ge::DT_FLOAT16) {
146 return tiling.DoTiling<typename DagT<half>::OpDag>(baseTiling);141 return tiling.DoTiling<typename DagT<half>::OpDag>(baseTiling);
@@ -154,9 +149,9 @@ ge::graphStatus DispatchTilingByDtype(
154// -----------------------------------------------------------------------------149// -----------------------------------------------------------------------------
155// 模板辅助函数：根据 dtype 分发 DoTiling 调用（双模板参数 DAG，用于 GENERIC_POW）。150// 模板辅助函数：根据 dtype 分发 DoTiling 调用（双模板参数 DAG，用于 GENERIC_POW）。
156// -----------------------------------------------------------------------------151// -----------------------------------------------------------------------------
157-template<template<typename, int> class DagT, int PowSign>152+template <template <typename, int> class DagT, int PowSign>
158-ge::graphStatus DispatchTilingByDtypeGeneric(153+ge::graphStatus DispatchTilingByDtypeGeneric(ElewiseBaseTiling& tiling, ge::DataType dtype,
159- ElewiseBaseTiling& tiling, ge::DataType dtype, Ops::Base::EleBaseTilingData& baseTiling)154+ Ops::Base::EleBaseTilingData& baseTiling)
160{155{
161 if (dtype == ge::DT_FLOAT16) {156 if (dtype == ge::DT_FLOAT16) {
162 return tiling.DoTiling<typename DagT<half, PowSign>::OpDag>(baseTiling);157 return tiling.DoTiling<typename DagT<half, PowSign>::OpDag>(baseTiling);
@@ -225,9 +220,11 @@ bool PowerTiling::DecideCulType()
225 }220 }
226 221 
227 // 分支2：优化整数幂 1/2/3，避免 exp/log 开销222 // 分支2：优化整数幂 1/2/3，避免 exp/log 开销
228- if (IsCloseScalar(this->attrPower, 1.0f) || IsCloseScalar(this->attrPower, 2.0f) || IsCloseScalar(this->attrPower, 3.0f)) {223+ if (IsCloseScalar(this->attrPower, 1.0f) || IsCloseScalar(this->attrPower, 2.0f) ||
224+ IsCloseScalar(this->attrPower, 3.0f)) {
229 culType = IsCloseScalar(this->attrPower, 1.0f) ? CulTypeEnum::LINEAR :225 culType = IsCloseScalar(this->attrPower, 1.0f) ? CulTypeEnum::LINEAR :
230- IsCloseScalar(this->attrPower, 2.0f) ? CulTypeEnum::SQUARE : CulTypeEnum::CUBE;226+ IsCloseScalar(this->attrPower, 2.0f) ? CulTypeEnum::SQUARE :
227+ CulTypeEnum::CUBE;
231 scalar0 = this->attrScale;228 scalar0 = this->attrScale;
232 scalar1 = this->attrShift;229 scalar1 = this->attrShift;
233 return true;230 return true;
@@ -239,7 +236,7 @@ bool PowerTiling::DecideCulType()
239 scalar1 = this->attrShift;236 scalar1 = this->attrShift;
240 scalar2 = this->attrPower;237 scalar2 = this->attrPower;
241 long long intPow = static_cast<long long>(std::llround(this->attrPower));238 long long intPow = static_cast<long long>(std::llround(this->attrPower));
242- scalar3 = powerIsInt ? ((std::llabs(intPow) % 2 == 0) ? 1.0f : -1.0f) : kNaN;239+ scalar3 = powerIsInt ? ((std::llabs(intPow) % POWER_PARITY_MODULUS == 0) ? 1.0f : -1.0f) : kNaN;
243 return true;240 return true;
244}241}
245 242 
@@ -251,20 +248,16 @@ bool PowerTiling::DecideCulType()
251// -----------------------------------------------------------------------------248// -----------------------------------------------------------------------------
252ge::graphStatus PowerTiling::PerformValidationChecks()249ge::graphStatus PowerTiling::PerformValidationChecks()
253{250{
254- OP_CHECK_IF(251+ OP_CHECK_IF(CalcInputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "get input dtype failed"),
255- CalcInputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "get input dtype failed"),252+ return ge::GRAPH_FAILED);
256- return ge::GRAPH_FAILED);253+ OP_CHECK_IF(CalcOutputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "get output dtype failed"),
257- OP_CHECK_IF(254+ return ge::GRAPH_FAILED);
258- CalcOutputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "get output dtype failed"),255+ OP_CHECK_IF(CheckShape() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "check shape failed"),
259- return ge::GRAPH_FAILED);256+ return ge::GRAPH_FAILED);
260- OP_CHECK_IF(257+ OP_CHECK_IF(SetAttr() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "set attrs failed"),
261- CheckShape() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "check shape failed"),258+ return ge::GRAPH_FAILED);
262- return ge::GRAPH_FAILED);259+ OP_CHECK_IF(!DecideCulType(), OP_LOGE(tilingContext->GetNodeName(), "DecideCulType failed"),
263- OP_CHECK_IF(260+ return ge::GRAPH_FAILED);
264- SetAttr() == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "set attrs failed"),
265- return ge::GRAPH_FAILED);
266- OP_CHECK_IF(
267- !DecideCulType(), OP_LOGE(tilingContext->GetNodeName(), "DecideCulType failed"), return ge::GRAPH_FAILED);
268 return ge::GRAPH_SUCCESS;261 return ge::GRAPH_SUCCESS;
269}262}
270 263 
@@ -278,9 +271,7 @@ ge::graphStatus PowerTiling::MapOutputDtypeToTplKey()
278 dType = POWER_TPL_DTYPE_FP32;271 dType = POWER_TPL_DTYPE_FP32;
279 } else {272 } else {
280 OP_LOGE_FOR_INVALID_DTYPE_WITH_REASON(273 OP_LOGE_FOR_INVALID_DTYPE_WITH_REASON(
281- tilingContext->GetNodeName(),274+ tilingContext->GetNodeName(), "y", Ops::Base::ToString(this->outputDtype).c_str(),
282- "y",
283- Ops::Base::ToString(this->outputDtype).c_str(),
284 "The dtype of y must be within the range [DT_FLOAT16, DT_BF16, DT_FLOAT].");275 "The dtype of y must be within the range [DT_FLOAT16, DT_BF16, DT_FLOAT].");
285 return ge::GRAPH_FAILED;276 return ge::GRAPH_FAILED;
286 }277 }
@@ -289,9 +280,9 @@ ge::graphStatus PowerTiling::MapOutputDtypeToTplKey()
289 280 
290ge::graphStatus PowerTiling::SetTilingResults(PowerOp::PowerTilingData* powerTilingData)281ge::graphStatus PowerTiling::SetTilingResults(PowerOp::PowerTilingData* powerTilingData)
291{282{
292- powerTilingData->scale = scalar0;283+ powerTilingData->scale = scalar0;
293- powerTilingData->shift = scalar1;284+ powerTilingData->shift = scalar1;
294- powerTilingData->power = scalar2;285+ powerTilingData->power = scalar2;
295 powerTilingData->negScalar = scalar3;286 powerTilingData->negScalar = scalar3;
296 287 
297 size_t* currentWorkspace = tilingContext->GetWorkspaceSizes(1);288 size_t* currentWorkspace = tilingContext->GetWorkspaceSizes(1);
@@ -318,19 +309,16 @@ ge::graphStatus PowerTiling::RunTiling()
318 }309 }
319 310 
320 culTypeKey = static_cast<uint64_t>(culType);311 culTypeKey = static_cast<uint64_t>(culType);
321- OP_LOGD(312+ OP_LOGD(tilingContext->GetNodeName(), "Power culType=%lu (scalar0=%f, scalar1=%f, scalar2=%f, scalar3=%f)",
322- tilingContext->GetNodeName(),313+ culTypeKey, scalar0, scalar1, scalar2, scalar3);
323- "Power culType=%lu (scalar0=%f, scalar1=%f, scalar2=%f, scalar3=%f)",
324- culTypeKey, scalar0, scalar1, scalar2, scalar3);
325 314 
326 if (MapOutputDtypeToTplKey() != ge::GRAPH_SUCCESS) {315 if (MapOutputDtypeToTplKey() != ge::GRAPH_SUCCESS) {
327 return ge::GRAPH_FAILED;316 return ge::GRAPH_FAILED;
328 }317 }
329 318 
330 ge::graphStatus baseTilingResult = DispatchTilingByCulType(elewiseBaseTiling, powerTilingData);319 ge::graphStatus baseTilingResult = DispatchTilingByCulType(elewiseBaseTiling, powerTilingData);
331- OP_CHECK_IF(320+ OP_CHECK_IF(baseTilingResult == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "elewiseBaseTiling failed"),
332- baseTilingResult == ge::GRAPH_FAILED, OP_LOGE(tilingContext->GetNodeName(), "elewiseBaseTiling failed"),321+ return ge::GRAPH_FAILED);
333- return ge::GRAPH_FAILED);
334 322 
335 return SetTilingResults(powerTilingData);323 return SetTilingResults(powerTilingData);
336}324}