Star861
1735
代码
代码
Issues166
Pull Requests300
Actions
项目讨论
Wiki
分析
Star861
1735
已合并
Improve Conv3DV2 nullptr check and pointwise validations #4403
LiuZonggu创建于 4月29日
Improve Conv3DV2 nullptr check and pointwise validations #4403
已合并
LiuZonggu创建于 4月29日
从LiuZonggu/ops-nnreview/conv3d_v2-base-tiling合入到cann/ops-nnmaster
共 20 个文件变更+319-225
Mconv/conv3d_v2/op_host/op_tiling/conv3d_api_tiling_algorithm.cpp+12-12
@@ -63,7 +63,7 @@ int64_t Conv3dTilingAlgorithm::GetL1Tiling()
63 BiasL1TilingDecision();63 BiasL1TilingDecision();
64 // get kl0 tiling decision64 // get kl0 tiling decision
65 GetKL0TilingDecision();65 GetKL0TilingDecision();
66- // get if weight can by pass in L166+ // get if weight can bypass in L1
67 WeightBypassDecision();67 WeightBypassDecision();
68 SetL1TilingRes();68 SetL1TilingRes();
69 69 
@@ -393,12 +393,12 @@ int64_t Conv3dTilingAlgorithm::ProcessAllL1FullLoad()
393 393 
394int64_t Conv3dTilingAlgorithm::ProcessFmapL1FullLoad()394int64_t Conv3dTilingAlgorithm::ProcessFmapL1FullLoad()
395{395{
396- // when only fmap full load in L1, nfirset and iter kbl1 then nbl1396+ // when only fmap full load in L1, nfirst and iter kbl1 then nbl1
397 this->l1TilingParams.kAL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *397 this->l1TilingParams.kAL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *
398 this->l1TilingCalc.ci0HkWk;398 this->l1TilingCalc.ci0HkWk;
399 this->l1TilingParams.mAL1Value = tilingIns_->cubeInfo.m0 * tilingIns_->shapeCalc.singleM1;399 this->l1TilingParams.mAL1Value = tilingIns_->cubeInfo.m0 * tilingIns_->shapeCalc.singleM1;
400 this->l1TilingFlag.iterateMNOrder = IterateMNOrder::ITER_N_FST;400 this->l1TilingFlag.iterateMNOrder = IterateMNOrder::ITER_N_FST;
401- // speical case, when min weight can not load in L1, bypass401+ // special case, when min weight can not load in L1, bypass
402 if (CoreL1TilingMinWeightBypass()) {402 if (CoreL1TilingMinWeightBypass()) {
403 this->l1TilingFlag.isWeightBypass = true;403 this->l1TilingFlag.isWeightBypass = true;
404 this->l1TilingParams.kBL1 = DISABLE_DOUBLE_BUFFER;404 this->l1TilingParams.kBL1 = DISABLE_DOUBLE_BUFFER;
@@ -412,7 +412,7 @@ int64_t Conv3dTilingAlgorithm::ProcessFmapL1FullLoad()
412 412 
413int64_t Conv3dTilingAlgorithm::ProcessWeightL1FullLoad()413int64_t Conv3dTilingAlgorithm::ProcessWeightL1FullLoad()
414{414{
415- // when only weight full load in L1, mfirset and iter kal1 then mal1415+ // when only weight full load in L1, mfirst and iter kal1 then mal1
416 this->l1TilingParams.kBL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *416 this->l1TilingParams.kBL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *
417 this->l1TilingCalc.ci0HkWk;417 this->l1TilingCalc.ci0HkWk;
418 this->l1TilingParams.nBL1Value = tilingIns_->cubeInfo.n0 * tilingIns_->shapeCalc.singleCo1;418 this->l1TilingParams.nBL1Value = tilingIns_->cubeInfo.n0 * tilingIns_->shapeCalc.singleCo1;
@@ -761,9 +761,9 @@ void Conv3dTilingAlgorithm::WeightBypassDecision()
761 if (!this->l1TilingFlag.isWeightBypass) {761 if (!this->l1TilingFlag.isWeightBypass) {
762 return;762 return;
763 }763 }
764- // when weight can by pass, BL1 db set to default 1764+ // when weight can bypass, BL1 db set to default 1
765 this->doubleBufferValue.pbBL1 = 1;765 this->doubleBufferValue.pbBL1 = 1;
766- // update L1 Tiling when weight by pass766+ // update L1 Tiling when weight bypass
767 // iter kAL1767 // iter kAL1
768 uint64_t tmpKAL1Idx = this->l1TilingIdx.kAL1Idx;768 uint64_t tmpKAL1Idx = this->l1TilingIdx.kAL1Idx;
769 for (this->l1TilingIdx.kAL1Idx = this->l1TilingRange.kAL1Range.size() - 1; this->l1TilingIdx.kAL1Idx >= tmpKAL1Idx;769 for (this->l1TilingIdx.kAL1Idx = this->l1TilingRange.kAL1Range.size() - 1; this->l1TilingIdx.kAL1Idx >= tmpKAL1Idx;
@@ -1004,13 +1004,13 @@ void Conv3dTilingAlgorithm::GetVecTiling() const
1004 1004 
1005 uint32_t maxLoadSize = QUANT_VEC_MAX_LOAD;1005 uint32_t maxLoadSize = QUANT_VEC_MAX_LOAD;
1006 uint64_t scaleAndBiasDTypeSize = this->scaleDTypeSize + this->biasDTypeSize;1006 uint64_t scaleAndBiasDTypeSize = this->scaleDTypeSize + this->biasDTypeSize;
1007- uint64_t scaleAndbiasLoadLength = INITIAL_SIZE;1007+ uint64_t scaleAndBiasLoadLength = INITIAL_SIZE;
1008 if (scaleAndBiasDTypeSize * tilingIns_->shapeInfo.singleCo <= maxLoadSize) {1008 if (scaleAndBiasDTypeSize * tilingIns_->shapeInfo.singleCo <= maxLoadSize) {
1009 tilingIns_->ubTilingInfo.scaleAndBiasLoadType = SINGLECO_LOAD;1009 tilingIns_->ubTilingInfo.scaleAndBiasLoadType = SINGLECO_LOAD;
1010- scaleAndbiasLoadLength = tilingIns_->shapeInfo.singleCo;1010+ scaleAndBiasLoadLength = tilingIns_->shapeInfo.singleCo;
1011 } else if (scaleAndBiasDTypeSize * tilingIns_->l0TilingInfo.nL0 <= maxLoadSize) {1011 } else if (scaleAndBiasDTypeSize * tilingIns_->l0TilingInfo.nL0 <= maxLoadSize) {
1012 tilingIns_->ubTilingInfo.scaleAndBiasLoadType = NL0_LOAD;1012 tilingIns_->ubTilingInfo.scaleAndBiasLoadType = NL0_LOAD;
1013- scaleAndbiasLoadLength = tilingIns_->l0TilingInfo.nL0;1013+ scaleAndBiasLoadLength = tilingIns_->l0TilingInfo.nL0;
1014 } else {1014 } else {
1015 tilingIns_->ubTilingInfo.scaleAndBiasLoadType = NUB_LOAD;1015 tilingIns_->ubTilingInfo.scaleAndBiasLoadType = NUB_LOAD;
1016 }1016 }
@@ -1023,12 +1023,12 @@ void Conv3dTilingAlgorithm::GetVecTiling() const
1023 uint64_t usedInSize = this->ubInDTypeSize * mUB * nUB;1023 uint64_t usedInSize = this->ubInDTypeSize * mUB * nUB;
1024 uint64_t usedOutSize = this->outputDTypeSize * mUB * nUB;1024 uint64_t usedOutSize = this->outputDTypeSize * mUB * nUB;
1025 if (tilingIns_->ubTilingInfo.scaleAndBiasLoadType == NUB_LOAD) {1025 if (tilingIns_->ubTilingInfo.scaleAndBiasLoadType == NUB_LOAD) {
1026- scaleAndbiasLoadLength = nUB;1026+ scaleAndBiasLoadLength = nUB;
1027 }1027 }
1028- uint64_t usedScaleAndBiasSize = scaleAndBiasDTypeSize * scaleAndbiasLoadLength;1028+ uint64_t usedScaleAndBiasSize = scaleAndBiasDTypeSize * scaleAndBiasLoadLength;
1029 if (tilingIns_->descInfo.biasType.dtype == ConvDtype::BF16 || // bf16 to fp321029 if (tilingIns_->descInfo.biasType.dtype == ConvDtype::BF16 || // bf16 to fp32
1030 tilingIns_->descInfo.biasType.dtype == ConvDtype::FLOAT16) { // fp16 to fp321030 tilingIns_->descInfo.biasType.dtype == ConvDtype::FLOAT16) { // fp16 to fp32
1031- usedScaleAndBiasSize += g_dtypeSizeTab.at(ConvDtype::FLOAT32) * scaleAndbiasLoadLength;1031+ usedScaleAndBiasSize += g_dtypeSizeTab.at(ConvDtype::FLOAT32) * scaleAndBiasLoadLength;
1032 }1032 }
1033 uint64_t usedUBSize = usedInSize + usedOutSize + usedScaleAndBiasSize;1033 uint64_t usedUBSize = usedInSize + usedOutSize + usedScaleAndBiasSize;
1034 1034 
Mconv/conv3d_v2/op_host/op_tiling/conv3d_api_tiling_algorithm_hw_mode.cpp+2-2
@@ -245,14 +245,14 @@ int64_t Conv3dTilingAlgorithmHwMode::ProcessAllL1FullLoad()
245 245 
246int64_t Conv3dTilingAlgorithmHwMode::ProcessFmapL1FullLoad()246int64_t Conv3dTilingAlgorithmHwMode::ProcessFmapL1FullLoad()
247{247{
248- // when only fmap full load in L1, nfirset and iter kbl1 then nbl1248+ // when only fmap full load in L1, nfirst and iter kbl1 then nbl1
249 this->l1TilingParams.kAL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) *249 this->l1TilingParams.kAL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) *
250 tilingIns_->shapeCalc.singleCi1 * this->l1TilingCalc.ci0HkWk;250 tilingIns_->shapeCalc.singleCi1 * this->l1TilingCalc.ci0HkWk;
251 this->l1TilingParams.woAL1Value = static_cast<uint64_t>(tilingIns_->shapeInfo.singleWo) /251 this->l1TilingParams.woAL1Value = static_cast<uint64_t>(tilingIns_->shapeInfo.singleWo) /
252 static_cast<uint64_t>(tilingIns_->cubeInfo.m0) *252 static_cast<uint64_t>(tilingIns_->cubeInfo.m0) *
253 static_cast<uint64_t>(tilingIns_->cubeInfo.m0);253 static_cast<uint64_t>(tilingIns_->cubeInfo.m0);
254 this->l1TilingFlag.iterateMNOrder = IterateMNOrder::ITER_N_FST;254 this->l1TilingFlag.iterateMNOrder = IterateMNOrder::ITER_N_FST;
255- // speical case, when min weight can not load in L1, bypass255+ // special case, when min weight can not load in L1, bypass
256 if (CoreL1TilingMinWeightBypass()) {256 if (CoreL1TilingMinWeightBypass()) {
257 this->l1TilingParams.kBL1 = INITIAL_SIZE;257 this->l1TilingParams.kBL1 = INITIAL_SIZE;
258 this->l1TilingParams.nBL1Value = INITIAL_SIZE;258 this->l1TilingParams.nBL1Value = INITIAL_SIZE;
Mconv/conv3d_v2/op_host/op_tiling/conv3d_api_tiling_algorithm_pointwise.cpp+5-5
@@ -70,10 +70,10 @@ int64_t Conv3dTilingAlgorithmPointWise::InitCalcL1Params()
70 static_cast<uint64_t>(tilingIns_->shapeInfo.orgkH *70 static_cast<uint64_t>(tilingIns_->shapeInfo.orgkH *
71 tilingIns_->shapeInfo.orgkW * tilingIns_->shapeInfo.orgkD);71 tilingIns_->shapeInfo.orgkW * tilingIns_->shapeInfo.orgkD);
72 // cal fmap weight full load in L1 size72 // cal fmap weight full load in L1 size
73- uint64_t fampSizeInL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *73+ uint64_t fmapSizeInL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *
74 tilingIns_->shapeCalc.singleM1 * static_cast<uint64_t>(tilingIns_->cubeInfo.m0 * tilingIns_->cubeInfo.k0);74 tilingIns_->shapeCalc.singleM1 * static_cast<uint64_t>(tilingIns_->cubeInfo.m0 * tilingIns_->cubeInfo.k0);
75- if ((fampSizeInL1 * this->fMapDTypeSize) / this->fMapDTypeSize != fampSizeInL1) {75+ if ((fmapSizeInL1 * this->fMapDTypeSize) / this->fMapDTypeSize != fmapSizeInL1) {
76- TILING_ERROR_LOG("fmap size in l1 is overflow uint64, initcalc l1 params failed!");76+ TILING_ERROR_LOG("fmap size in l1 is overflow uint64, InitCalc l1 params failed!");
77 return INVALID_VALUE;77 return INVALID_VALUE;
78 }78 }
79 this->l1TilingCalc.fmapFullLoadL1Size = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) *79 this->l1TilingCalc.fmapFullLoadL1Size = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) *
@@ -82,7 +82,7 @@ int64_t Conv3dTilingAlgorithmPointWise::InitCalcL1Params()
82 uint64_t weightSizeInL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *82 uint64_t weightSizeInL1 = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *
83 this->l1TilingCalc.ci0HkWk * tilingIns_->shapeCalc.singleCo1 * static_cast<uint64_t>(tilingIns_->cubeInfo.n0);83 this->l1TilingCalc.ci0HkWk * tilingIns_->shapeCalc.singleCo1 * static_cast<uint64_t>(tilingIns_->cubeInfo.n0);
84 if ((weightSizeInL1 * this->weightDTypeSize) / weightDTypeSize != weightSizeInL1) {84 if ((weightSizeInL1 * this->weightDTypeSize) / weightDTypeSize != weightSizeInL1) {
85- TILING_ERROR_LOG("weight size in l1 is overflow uint64, initcalc l1 params failed!");85+ TILING_ERROR_LOG("weight size in l1 is overflow uint64, InitCalc l1 params failed!");
86 return INVALID_VALUE;86 return INVALID_VALUE;
87 }87 }
88 this->l1TilingCalc.weightFullLoadL1Size = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) *88 this->l1TilingCalc.weightFullLoadL1Size = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) *
@@ -98,7 +98,7 @@ int64_t Conv3dTilingAlgorithmPointWise::InitCalcL1Params()
98 this->l1TilingCalc.fmapKL1FullLoadSize = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *98 this->l1TilingCalc.fmapKL1FullLoadSize = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *
99 static_cast<uint64_t>(tilingIns_->cubeInfo.k0) * this->l0TilingParams.mL0 * static_cast<uint64_t>(this->doubleBufferValue.pbAL1) *99 static_cast<uint64_t>(tilingIns_->cubeInfo.k0) * this->l0TilingParams.mL0 * static_cast<uint64_t>(this->doubleBufferValue.pbAL1) *
100 this->fMapDTypeSize;100 this->fMapDTypeSize;
101- // cal min/kfullload weiht size in L1101+ // cal min/kfullload weight size in L1
102 this->l1TilingCalc.weightMinLoadL1Size = this->l1TilingCalc.ci0HkWk * this->l0TilingParams.nL0 *102 this->l1TilingCalc.weightMinLoadL1Size = this->l1TilingCalc.ci0HkWk * this->l0TilingParams.nL0 *
103 this->doubleBufferValue.pbBL1 * this->weightDTypeSize * multiK0;103 this->doubleBufferValue.pbBL1 * this->weightDTypeSize * multiK0;
104 this->l1TilingCalc.weightKL1FullLoadSize = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *104 this->l1TilingCalc.weightKL1FullLoadSize = static_cast<uint64_t>(tilingIns_->shapeInfo.singlekD) * tilingIns_->shapeCalc.singleCi1 *
Mconv/conv3d_v2/op_host/op_tiling/conv3d_api_tiling_base.cpp+12-12
@@ -191,9 +191,9 @@ void Conv3dTilingBase::SetHF32(bool hf32Enable, bool hf32TransMode = false)
191bool Conv3dTilingBase::CalOptGroupParams(const Conv3DOriGroupInfo &oriGroupInfo,191bool Conv3dTilingBase::CalOptGroupParams(const Conv3DOriGroupInfo &oriGroupInfo,
192 Conv3DGroupOptInfo &groupOptInfo) const192 Conv3DGroupOptInfo &groupOptInfo) const
193{193{
194- // user need to pass correct parms.194+ // user need to pass correct params.
195 if (oriGroupInfo.groups < 1 || oriGroupInfo.cin < 1 || oriGroupInfo.cout < 1) {195 if (oriGroupInfo.groups < 1 || oriGroupInfo.cin < 1 || oriGroupInfo.cout < 1) {
196- TILING_DEBUG_LOG("Conv3D AscendC: unSupported parms in groupOpt:" \196+ TILING_DEBUG_LOG("Conv3D AscendC: Unsupported params in groupOpt:" \
197 " groups: %ld, cin: %ld, cout: %ld, only support greater than zero",197 " groups: %ld, cin: %ld, cout: %ld, only support greater than zero",
198 oriGroupInfo.groups, oriGroupInfo.cin, oriGroupInfo.cout);198 oriGroupInfo.groups, oriGroupInfo.cin, oriGroupInfo.cout);
199 return false;199 return false;
@@ -304,7 +304,7 @@ bool Conv3dTilingBase::CheckInputAttr() const
304 this->attrInfo.padHead < 0 || this->attrInfo.padTail < 0);304 this->attrInfo.padHead < 0 || this->attrInfo.padTail < 0);
305 if (padInvalidFlag) {305 if (padInvalidFlag) {
306 TILING_ERROR_LOG(306 TILING_ERROR_LOG(
307- "Illlegal attrs have set: padTop=%ld, padBottom=%ld, padLeft=%ld, padRight=%ld, padHead=%ld, padTail=%ld,\307+ "Illegal attrs have set: padTop=%ld, padBottom=%ld, padLeft=%ld, padRight=%ld, padHead=%ld, padTail=%ld,\
308 which must >= 0.", this->attrInfo.padTop, this->attrInfo.padBottom, this->attrInfo.padLeft,308 which must >= 0.", this->attrInfo.padTop, this->attrInfo.padBottom, this->attrInfo.padLeft,
309 this->attrInfo.padRight, this->attrInfo.padHead, this->attrInfo.padTail);309 this->attrInfo.padRight, this->attrInfo.padHead, this->attrInfo.padTail);
310 return false;310 return false;
@@ -333,7 +333,7 @@ bool Conv3dTilingBase::CheckInputAttrPointWise() const
333{333{
334 if (this->attrInfo.groups != 1) {334 if (this->attrInfo.groups != 1) {
335 TILING_ERROR_LOG(335 TILING_ERROR_LOG(
336- "[PointWise] Illlegal attrs have set: groups=%ld.",336+ "[PointWise] Illegal attrs have set: groups=%ld.",
337 this->attrInfo.groups);337 this->attrInfo.groups);
338 return false;338 return false;
339 }339 }
@@ -346,7 +346,7 @@ bool Conv3dTilingBase::CheckInputAttrPointWise() const
346 this->attrInfo.padTail != 0);346 this->attrInfo.padTail != 0);
347 if (padInvalidFlag) {347 if (padInvalidFlag) {
348 TILING_ERROR_LOG(348 TILING_ERROR_LOG(
349- "[PointWise] Illlegal attrs have set: padTop=%ld, padBottom=%ld, padLeft=%ld, padRight=%ld, padHead=%ld, padTail=%ld,\349+ "[PointWise] Illegal attrs have set: padTop=%ld, padBottom=%ld, padLeft=%ld, padRight=%ld, padHead=%ld, padTail=%ld,\
350 which must = 0.", this->attrInfo.padTop, this->attrInfo.padBottom, this->attrInfo.padLeft,350 which must = 0.", this->attrInfo.padTop, this->attrInfo.padBottom, this->attrInfo.padLeft,
351 this->attrInfo.padRight, this->attrInfo.padHead, this->attrInfo.padTail);351 this->attrInfo.padRight, this->attrInfo.padHead, this->attrInfo.padTail);
352 return false;352 return false;
@@ -528,13 +528,13 @@ bool Conv3dTilingBase::CheckInputShapePointWise() const
528bool Conv3dTilingBase::CheckInputFormat() const528bool Conv3dTilingBase::CheckInputFormat() const
529{529{
530 if (this->descInfo.weightType.format != ConvFormat::FRACTAL_Z_3D) {530 if (this->descInfo.weightType.format != ConvFormat::FRACTAL_Z_3D) {
531- TILING_ERROR_LOG("unSupported weight format: %s.",531+ TILING_ERROR_LOG("Unsupported weight format: %s.",
532 g_formatToStr.at(this->descInfo.weightType.format).c_str());532 g_formatToStr.at(this->descInfo.weightType.format).c_str());
533 return false;533 return false;
534 }534 }
535 535 
536 if (this->descInfo.fMapType.format != ConvFormat::NDC1HWC0) {536 if (this->descInfo.fMapType.format != ConvFormat::NDC1HWC0) {
537- TILING_ERROR_LOG("unSupported feature map format: %s.",537+ TILING_ERROR_LOG("Unsupported feature map format: %s.",
538 g_formatToStr.at(this->descInfo.fMapType.format).c_str());538 g_formatToStr.at(this->descInfo.fMapType.format).c_str());
539 return false;539 return false;
540 }540 }
@@ -571,7 +571,7 @@ bool Conv3dTilingBase::CheckParamsDtypeHasQuantScale() const
571 return true;571 return true;
572 }572 }
573 }573 }
574- TILING_ERROR_LOG("unSupported params data type [fmap, weight, bias, scale, output]: [%s, %s, %s, %s, %s].",574+ TILING_ERROR_LOG("Unsupported params data type [fmap, weight, bias, scale, output]: [%s, %s, %s, %s, %s].",
575 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),575 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),
576 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),576 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),
577 g_dtypeToStr.at(this->descInfo.biasType.dtype).c_str(),577 g_dtypeToStr.at(this->descInfo.biasType.dtype).c_str(),
@@ -592,7 +592,7 @@ bool Conv3dTilingBase::CheckParamsDtypeHasBias() const
592 return true;592 return true;
593 }593 }
594 }594 }
595- TILING_ERROR_LOG("unSupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",595+ TILING_ERROR_LOG("Unsupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",
596 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),596 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),
597 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),597 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),
598 g_dtypeToStr.at(this->descInfo.biasType.dtype).c_str(),598 g_dtypeToStr.at(this->descInfo.biasType.dtype).c_str(),
@@ -611,7 +611,7 @@ bool Conv3dTilingBase::CheckParamsDtypeEssential() const
611 return true;611 return true;
612 }612 }
613 }613 }
614- TILING_ERROR_LOG("unSupported params data type [fmap, weight, output]: [%s, %s, %s].",614+ TILING_ERROR_LOG("Unsupported params data type [fmap, weight, output]: [%s, %s, %s].",
615 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),615 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),
616 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),616 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),
617 g_dtypeToStr.at(this->descInfo.outputType.dtype).c_str());617 g_dtypeToStr.at(this->descInfo.outputType.dtype).c_str());
@@ -641,7 +641,7 @@ bool Conv3dTilingBase::CheckParamsDtypePointWise() const
641 return true;641 return true;
642 }642 }
643 }643 }
644- TILING_ERROR_LOG("[PointWise] unSupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",644+ TILING_ERROR_LOG("[PointWise] Unsupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",
645 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),645 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),
646 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),646 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),
647 g_dtypeToStr.at(this->descInfo.biasType.dtype).c_str(),647 g_dtypeToStr.at(this->descInfo.biasType.dtype).c_str(),
@@ -657,7 +657,7 @@ bool Conv3dTilingBase::CheckParamsDtypePointWise() const
657 return true;657 return true;
658 }658 }
659 }659 }
660- TILING_ERROR_LOG("[PointWise] unSupported params data type [fmap, weight, output]: [%s, %s, %s].",660+ TILING_ERROR_LOG("[PointWise] Unsupported params data type [fmap, weight, output]: [%s, %s, %s].",
661 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),661 g_dtypeToStr.at(this->descInfo.fMapType.dtype).c_str(),
662 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),662 g_dtypeToStr.at(this->descInfo.weightType.dtype).c_str(),
663 g_dtypeToStr.at(this->descInfo.outputType.dtype).c_str());663 g_dtypeToStr.at(this->descInfo.outputType.dtype).c_str());
Mconv/conv3d_v2/op_host/op_tiling/conv3d_base_tiling.cpp+28-40
@@ -56,31 +56,29 @@ ge::graphStatus Conv3dBaseTiling::GetPlatformInfo()
56 return ge::GRAPH_SUCCESS;56 return ge::GRAPH_SUCCESS;
57}57}
58 58 
59-void Conv3dBaseTiling::GetAttrsInfo()59+bool Conv3dBaseTiling::SyncAttrInfoFromEngine()
60{60{
61- auto stridePtr = context_->GetAttrs()->GetListInt(ATTR_STRIDE_INDEX);61+ if (engine_ == nullptr) {
62- attrInfo_.strideH = static_cast<uint32_t>(stridePtr->GetData()[originalFormat_.FORMAT_DATA_H_INDEX]);62+ OP_LOGE(context_->GetNodeName(), "Conv3D AscendC: engine is null when syncing attr info.");
63- attrInfo_.strideW = static_cast<uint32_t>(stridePtr->GetData()[originalFormat_.FORMAT_DATA_W_INDEX]);63+ return false;
64- attrInfo_.strideD = static_cast<uint32_t>(stridePtr->GetData()[originalFormat_.FORMAT_DATA_D_INDEX]);
65- 
66- auto padPtr = context_->GetAttrs()->GetListInt(ATTR_PAD_INDEX);
67- attrInfo_.padh = static_cast<uint32_t>(padPtr->GetData()[PAD_HEAD_INDEX]);
68- attrInfo_.padt = static_cast<uint32_t>(padPtr->GetData()[PAD_TAIL_INDEX]);
69- attrInfo_.padu = static_cast<uint32_t>(padPtr->GetData()[PAD_UP_INDEX]);
70- attrInfo_.padd = static_cast<uint32_t>(padPtr->GetData()[PAD_DOWN_INDEX]);
71- attrInfo_.padl = static_cast<uint32_t>(padPtr->GetData()[PAD_LEFT_INDEX]);
72- attrInfo_.padr = static_cast<uint32_t>(padPtr->GetData()[PAD_RIGHT_INDEX]);
73- 
74- auto dilationPtr = context_->GetAttrs()->GetListInt(ATTR_DILATION_INDEX);
75- if (dilationPtr != nullptr) {
76- attrInfo_.dilationH = static_cast<uint32_t>(dilationPtr->GetData()[originalFormat_.FORMAT_DATA_H_INDEX]);
77- attrInfo_.dilationW = static_cast<uint32_t>(dilationPtr->GetData()[originalFormat_.FORMAT_DATA_W_INDEX]);
78- attrInfo_.dilationD = static_cast<uint32_t>(dilationPtr->GetData()[originalFormat_.FORMAT_DATA_D_INDEX]);
79- }
80- auto groupPtr = context_->GetAttrs()->GetInt(ATTR_GROUP_INDEX);
81- if (groupPtr != nullptr) {
82- attrInfo_.groups = static_cast<uint32_t>(*groupPtr);
83 }64 }
65+ 
66+ attrInfo_.strideH = static_cast<uint32_t>(engine_->attrInfo_.strideH);
67+ attrInfo_.strideW = static_cast<uint32_t>(engine_->attrInfo_.strideW);
68+ attrInfo_.strideD = static_cast<uint32_t>(engine_->attrInfo_.strideD);
69+ 
70+ attrInfo_.padh = static_cast<uint32_t>(engine_->attrInfo_.padHead);
71+ attrInfo_.padt = static_cast<uint32_t>(engine_->attrInfo_.padTail);
72+ attrInfo_.padu = static_cast<uint32_t>(engine_->attrInfo_.padTop);
73+ attrInfo_.padd = static_cast<uint32_t>(engine_->attrInfo_.padBottom);
74+ attrInfo_.padl = static_cast<uint32_t>(engine_->attrInfo_.padLeft);
75+ attrInfo_.padr = static_cast<uint32_t>(engine_->attrInfo_.padRight);
76+ 
77+ attrInfo_.dilationH = static_cast<uint32_t>(engine_->attrInfo_.dilationH);
78+ attrInfo_.dilationW = static_cast<uint32_t>(engine_->attrInfo_.dilationW);
79+ attrInfo_.dilationD = static_cast<uint32_t>(engine_->attrInfo_.dilationD);
80+ attrInfo_.groups = static_cast<uint32_t>(engine_->attrInfo_.groups);
81+ return true;
84}82}
85 83 
86void Conv3dBaseTiling::GetConv3DParasHf32Mode(const uint32_t enableHf32Idx, uint32_t& hf32Mode)84void Conv3dBaseTiling::GetConv3DParasHf32Mode(const uint32_t enableHf32Idx, uint32_t& hf32Mode)
@@ -447,14 +445,6 @@ ge::graphStatus Conv3dBaseTiling::DoOpTiling()
447 return ge::GRAPH_SUCCESS;445 return ge::GRAPH_SUCCESS;
448}446}
449 447 
450-void Conv3dBaseTiling::InitPointWiseFlag() {
451- auto fMapDesc = context_->GetInputDesc(INPUT_FMAP_INDEX);
452- ge::Format storageFmapFormat = static_cast<ge::Format>(GetPrimaryFormat(fMapDesc->GetStorageFormat()));
453- if (storageFmapFormat == ge::Format::FORMAT_NCDHW) {
454- isPointWise = true;
455- }
456-}
457- 
458bool Conv3dBaseTiling::ExtractAndPassParamsToEngine()448bool Conv3dBaseTiling::ExtractAndPassParamsToEngine()
459{449{
460 // Set shapes450 // Set shapes
@@ -512,7 +502,6 @@ ge::graphStatus Conv3dBaseTiling::GetShapeAttrsInfo()
512 502 
513 // Step 2: Parse original format (NCDHW vs NDHWC index mapping)503 // Step 2: Parse original format (NCDHW vs NDHWC index mapping)
514 InitConv3dOriginFormat();504 InitConv3dOriginFormat();
515- InitPointWiseFlag();
516 505 
517 // Step 3: Extract parameters from GE context and pass to Engine506 // Step 3: Extract parameters from GE context and pass to Engine
518 if (!ExtractAndPassParamsToEngine()) {507 if (!ExtractAndPassParamsToEngine()) {
@@ -520,14 +509,13 @@ ge::graphStatus Conv3dBaseTiling::GetShapeAttrsInfo()
520 }509 }
521 510 
522 // Step 4: Sync Base copies for legacy paths (AOE repo, PrintTilingInfo, GetTilingInputArgs)511 // Step 4: Sync Base copies for legacy paths (AOE repo, PrintTilingInfo, GetTilingInputArgs)
523- GetAttrsInfo();512+ if (!SyncAttrInfoFromEngine()) {
513+ return ge::GRAPH_FAILED;
514+ }
524 GetShapeInfo();515 GetShapeInfo();
525 GetDescInfo();516 GetDescInfo();
526 517 
527- // Step 5: Sync isPointWise flag to Engine518+ // Step 5: Run Engine parameter validation (includes format, shape, dtype checks via CheckInputFormat())
528- engine_->isPointWise = isPointWise;
529- 
530- // Step 6: Run Engine parameter validation (includes format, shape, dtype checks via CheckInputFormat())
531 if (!engine_->CheckAllParams()) {519 if (!engine_->CheckAllParams()) {
532 OP_LOGE(context_->GetNodeName(), "Conv3D AscendC: Engine parameter check failed in GetShapeAttrsInfo.");520 OP_LOGE(context_->GetNodeName(), "Conv3D AscendC: Engine parameter check failed in GetShapeAttrsInfo.");
533 return ge::GRAPH_FAILED;521 return ge::GRAPH_FAILED;
@@ -607,7 +595,7 @@ bool Conv3dBaseTiling::GetTilingInputArgs(std::shared_ptr<void> &inputArgs, size
607 595 
608bool Conv3dBaseTiling::TranslateAoeTiling(tuningtiling::TuningTilingDefPtr &tuningTiling)596bool Conv3dBaseTiling::TranslateAoeTiling(tuningtiling::TuningTilingDefPtr &tuningTiling)
609{597{
610- auto aoeTiling = std::static_pointer_cast<tuningtiling::Conv3DTunnerTiling>(tuningTiling);598+ auto aoeTiling = std::static_pointer_cast<tuningtiling::Conv3DTunerTiling>(tuningTiling);
611 if (aoeTiling == nullptr) {599 if (aoeTiling == nullptr) {
612 return false;600 return false;
613 }601 }
@@ -618,7 +606,7 @@ bool Conv3dBaseTiling::TranslateAoeTiling(tuningtiling::TuningTilingDefPtr &tuni
618 return true;606 return true;
619}607}
620 608 
621-void Conv3dBaseTiling::TranslateApiTiling(std::shared_ptr<tuningtiling::Conv3DTunnerTiling> aoeTiling)609+void Conv3dBaseTiling::TranslateApiTiling(std::shared_ptr<tuningtiling::Conv3DTunerTiling> aoeTiling)
622{610{
623 tilingData_.convApiTiling.groups = aoeTiling->groups;611 tilingData_.convApiTiling.groups = aoeTiling->groups;
624 tilingData_.convApiTiling.orgCi = aoeTiling->orgCi;612 tilingData_.convApiTiling.orgCi = aoeTiling->orgCi;
@@ -664,7 +652,7 @@ void Conv3dBaseTiling::TranslateApiTiling(std::shared_ptr<tuningtiling::Conv3DTu
664 tilingData_.convApiTiling.offsetx = aoeTiling->offsetx;652 tilingData_.convApiTiling.offsetx = aoeTiling->offsetx;
665}653}
666 654 
667-void Conv3dBaseTiling::TranslateRunInfo(std::shared_ptr<tuningtiling::Conv3DTunnerTiling> aoeTiling)655+void Conv3dBaseTiling::TranslateRunInfo(std::shared_ptr<tuningtiling::Conv3DTunerTiling> aoeTiling)
668{656{
669 tilingData_.convRunInfo.batch = shapeInfo_.batch;657 tilingData_.convRunInfo.batch = shapeInfo_.batch;
670 tilingData_.convRunInfo.cin = aoeTiling->orgCi;658 tilingData_.convRunInfo.cin = aoeTiling->orgCi;
Mconv/conv3d_v2/op_host/op_tiling/conv3d_base_tiling.h+5-7
@@ -61,7 +61,7 @@ struct Conv3DAttrInfo {
61 uint32_t hf32Mode = 0;61 uint32_t hf32Mode = 0;
62};62};
63 63 
64-struct Conv3DOrignalFormat {64+struct Conv3DOriginalFormat {
65 // for fmap65 // for fmap
66 uint32_t FORMAT_FMAP_N_INDEX = Conv3dApiTiling::INITIAL_INDEX;66 uint32_t FORMAT_FMAP_N_INDEX = Conv3dApiTiling::INITIAL_INDEX;
67 uint32_t FORMAT_FMAP_C_INDEX = Conv3dApiTiling::INITIAL_INDEX;67 uint32_t FORMAT_FMAP_C_INDEX = Conv3dApiTiling::INITIAL_INDEX;
@@ -156,8 +156,8 @@ protected:
156 156 
157 bool GetTilingFromRepo();157 bool GetTilingFromRepo();
158 bool TranslateAoeTiling(tuningtiling::TuningTilingDefPtr &tuningTiling);158 bool TranslateAoeTiling(tuningtiling::TuningTilingDefPtr &tuningTiling);
159- void TranslateApiTiling(std::shared_ptr<tuningtiling::Conv3DTunnerTiling> aoeTiling);159+ void TranslateApiTiling(std::shared_ptr<tuningtiling::Conv3DTunerTiling> aoeTiling);
160- void TranslateRunInfo(std::shared_ptr<tuningtiling::Conv3DTunnerTiling> aoeTiling);160+ void TranslateRunInfo(std::shared_ptr<tuningtiling::Conv3DTunerTiling> aoeTiling);
161 bool GetTilingInputArgs(std::shared_ptr<void> &inputArgs, size_t &size);161 bool GetTilingInputArgs(std::shared_ptr<void> &inputArgs, size_t &size);
162 void SetAdditionalTilingInfo();162 void SetAdditionalTilingInfo();
163 163 
@@ -170,21 +170,19 @@ private:
170 Ops::NN::Conv3dV2::Conv3DV2TilingData tilingData_;170 Ops::NN::Conv3dV2::Conv3DV2TilingData tilingData_;
171 Conv3DDescInfo descInfo_;171 Conv3DDescInfo descInfo_;
172 Conv3DTilingFlag flagInfo_;172 Conv3DTilingFlag flagInfo_;
173- Conv3DOrignalFormat originalFormat_;173+ Conv3DOriginalFormat originalFormat_;
174 174 
175 // numblocks decision175 // numblocks decision
176 NumBlocksRes numBlocksRes;176 NumBlocksRes numBlocksRes;
177 177 
178 bool useTilingRepo_ = false;178 bool useTilingRepo_ = false;
179- bool isPointWise = false;
180 int8_t outputOrder_ = 0;179 int8_t outputOrder_ = 0;
181 180 
182 private:181 private:
183 ge::graphStatus SetTilingKey();182 ge::graphStatus SetTilingKey();
184 void InitConv3dOriginFormat();183 void InitConv3dOriginFormat();
185- void InitPointWiseFlag();
186 void GetShapeInfo();184 void GetShapeInfo();
187- void GetAttrsInfo();185+ bool SyncAttrInfoFromEngine();
188 void GetDescInfo();186 void GetDescInfo();
189 void PrintTilingInfo();187 void PrintTilingInfo();
190 void GetConv3DParasHf32Mode(const uint32_t enableHf32Idx, uint32_t& hf32Mode);188 void GetConv3DParasHf32Mode(const uint32_t enableHf32Idx, uint32_t& hf32Mode);
Mconv/conv3d_v2/op_host/op_tiling/conv3d_tiling_engine.cpp+68-12
@@ -39,6 +39,9 @@ Conv3dTilingEngine::Conv3dTilingEngine(const std::string &logTag)
39 numBlocksRes_.minCost = MAX_64_BIT_NUM;39 numBlocksRes_.minCost = MAX_64_BIT_NUM;
40 40 
41 isPointWise = false;41 isPointWise = false;
42+ kernelPointWise_ = false;
43+ outputBatch_ = -1;
44+ outputCOut_ = -1;
42 outputOrder_ = static_cast<uint8_t>(Conv3dApiTiling::M_Mode);45 outputOrder_ = static_cast<uint8_t>(Conv3dApiTiling::M_Mode);
43 46 
44 OP_LOGD(logTag_.c_str(), "Conv3dTilingEngine constructed, call Init() to initialize platform info");47 OP_LOGD(logTag_.c_str(), "Conv3dTilingEngine constructed, call Init() to initialize platform info");
@@ -71,6 +74,19 @@ uint8_t Conv3dTilingEngine::GetOutputOrder() const
71 return outputOrder_;74 return outputOrder_;
72}75}
73 76 
77+bool Conv3dTilingEngine::UsesPointWisePath() const
78+{
79+ return kernelPointWise_ &&
80+ descInfo_.fMapFormat == Conv3dApiTiling::ConvFormat::NCDHW &&
81+ descInfo_.weightFormat == Conv3dApiTiling::ConvFormat::NCDHW &&
82+ descInfo_.outFormat == Conv3dApiTiling::ConvFormat::NCDHW;
83+}
84+ 
85+void Conv3dTilingEngine::UpdatePointWiseMode()
86+{
87+ isPointWise = UsesPointWisePath();
88+}
89+ 
74bool Conv3dTilingEngine::InitPlatformInfoFromAscendC()90bool Conv3dTilingEngine::InitPlatformInfoFromAscendC()
75{91{
76 OP_LOGD(logTag_.c_str(), "Initializing Conv3dTilingEngine with PlatformAscendCManager");92 OP_LOGD(logTag_.c_str(), "Initializing Conv3dTilingEngine with PlatformAscendCManager");
@@ -145,8 +161,9 @@ void Conv3dTilingEngine::SetOrgWeightShape(const std::vector<int64_t> &orgWeight
145 shapeInfo_.kh = static_cast<uint32_t>(orgWeightShapeList[FORMAT_NCDHW_H_INDEX]);161 shapeInfo_.kh = static_cast<uint32_t>(orgWeightShapeList[FORMAT_NCDHW_H_INDEX]);
146 shapeInfo_.kw = static_cast<uint32_t>(orgWeightShapeList[FORMAT_NCDHW_W_INDEX]);162 shapeInfo_.kw = static_cast<uint32_t>(orgWeightShapeList[FORMAT_NCDHW_W_INDEX]);
147 163 
148- // Update pointwise flag - pointwise convolution has 1x1x1 kernel164+ // Pointwise path is only valid when both kernel shape and tensor formats match the pointwise contract.
149- isPointWise = (shapeInfo_.kd == 1 && shapeInfo_.kh == 1 && shapeInfo_.kw == 1);165+ kernelPointWise_ = (shapeInfo_.kd == 1 && shapeInfo_.kh == 1 && shapeInfo_.kw == 1);
166+ UpdatePointWiseMode();
150}167}
151 168 
152void Conv3dTilingEngine::SetOrgFmapShape(const std::vector<int64_t> &orgFmapShapeList)169void Conv3dTilingEngine::SetOrgFmapShape(const std::vector<int64_t> &orgFmapShapeList)
@@ -174,6 +191,8 @@ void Conv3dTilingEngine::SetOrgOutputShape(const std::vector<int64_t> &orgOutput
174 return;191 return;
175 }192 }
176 193 
194+ outputBatch_ = orgOutputShapeList[FORMAT_NCDHW_N_INDEX];
195+ outputCOut_ = orgOutputShapeList[FORMAT_NCDHW_C_INDEX];
177 shapeInfo_.dOut = static_cast<uint32_t>(orgOutputShapeList[FORMAT_NCDHW_D_INDEX]);196 shapeInfo_.dOut = static_cast<uint32_t>(orgOutputShapeList[FORMAT_NCDHW_D_INDEX]);
178 shapeInfo_.ho = static_cast<uint64_t>(orgOutputShapeList[FORMAT_NCDHW_H_INDEX]);197 shapeInfo_.ho = static_cast<uint64_t>(orgOutputShapeList[FORMAT_NCDHW_H_INDEX]);
179 shapeInfo_.wo = static_cast<uint64_t>(orgOutputShapeList[FORMAT_NCDHW_W_INDEX]);198 shapeInfo_.wo = static_cast<uint64_t>(orgOutputShapeList[FORMAT_NCDHW_W_INDEX]);
@@ -257,6 +276,8 @@ void Conv3dTilingEngine::SetFormat(Conv3dApiTiling::ConvFormat fmapFormat,
257 Conv3dApiTiling::g_formatToStr.at(fmapFormat).c_str(),276 Conv3dApiTiling::g_formatToStr.at(fmapFormat).c_str(),
258 Conv3dApiTiling::g_formatToStr.at(weightFormat).c_str(),277 Conv3dApiTiling::g_formatToStr.at(weightFormat).c_str(),
259 Conv3dApiTiling::g_formatToStr.at(outFormat).c_str());278 Conv3dApiTiling::g_formatToStr.at(outFormat).c_str());
279+ 
280+ UpdatePointWiseMode();
260}281}
261 282 
262void Conv3dTilingEngine::SetBias(bool hasBias, Conv3dApiTiling::ConvDtype biasDtype)283void Conv3dTilingEngine::SetBias(bool hasBias, Conv3dApiTiling::ConvDtype biasDtype)
@@ -679,7 +700,7 @@ bool CheckPointWiseParamsDtypeWithBias(const char *logTag, const Conv3DEngineDes
679 }700 }
680 701 
681 OP_LOGE(logTag,702 OP_LOGE(logTag,
682- "[PointWise] unSupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",703+ "[PointWise] Unsupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",
683 g_convDtypeToStr[descInfo.fMapDtype].c_str(),704 g_convDtypeToStr[descInfo.fMapDtype].c_str(),
684 g_convDtypeToStr[descInfo.weightDtype].c_str(),705 g_convDtypeToStr[descInfo.weightDtype].c_str(),
685 g_convDtypeToStr[descInfo.biasDtype].c_str(),706 g_convDtypeToStr[descInfo.biasDtype].c_str(),
@@ -701,7 +722,7 @@ bool CheckPointWiseParamsDtypeWithoutBias(const char *logTag, const Conv3DEngine
701 return true;722 return true;
702 }723 }
703 724 
704- OP_LOGE(logTag, "[PointWise] unSupported params data type [fmap, weight, output]: [%s, %s, %s].",725+ OP_LOGE(logTag, "[PointWise] Unsupported params data type [fmap, weight, output]: [%s, %s, %s].",
705 g_convDtypeToStr[descInfo.fMapDtype].c_str(),726 g_convDtypeToStr[descInfo.fMapDtype].c_str(),
706 g_convDtypeToStr[descInfo.weightDtype].c_str(),727 g_convDtypeToStr[descInfo.weightDtype].c_str(),
707 g_convDtypeToStr[descInfo.outDtype].c_str());728 g_convDtypeToStr[descInfo.outDtype].c_str());
@@ -727,7 +748,7 @@ bool CheckParamsDtypeWithScale(const char *logTag, const Conv3DEngineDescInfo &d
727 }748 }
728 749 
729 OP_LOGE(logTag,750 OP_LOGE(logTag,
730- "unSupported params data type [fmap, weight, bias, scale, output]: [%s, %s, %s, %s, %s].",751+ "Unsupported params data type [fmap, weight, bias, scale, output]: [%s, %s, %s, %s, %s].",
731 g_convDtypeToStr[descInfo.fMapDtype].c_str(),752 g_convDtypeToStr[descInfo.fMapDtype].c_str(),
732 g_convDtypeToStr[descInfo.weightDtype].c_str(),753 g_convDtypeToStr[descInfo.weightDtype].c_str(),
733 g_convDtypeToStr[descInfo.biasDtype].c_str(),754 g_convDtypeToStr[descInfo.biasDtype].c_str(),
@@ -752,7 +773,7 @@ bool CheckParamsDtypeWithBias(const char *logTag, const Conv3DEngineDescInfo &de
752 return true;773 return true;
753 }774 }
754 775 
755- OP_LOGE(logTag, "unSupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",776+ OP_LOGE(logTag, "Unsupported params data type [fmap, weight, bias, output]: [%s, %s, %s, %s].",
756 g_convDtypeToStr[descInfo.fMapDtype].c_str(),777 g_convDtypeToStr[descInfo.fMapDtype].c_str(),
757 g_convDtypeToStr[descInfo.weightDtype].c_str(),778 g_convDtypeToStr[descInfo.weightDtype].c_str(),
758 g_convDtypeToStr[descInfo.biasDtype].c_str(),779 g_convDtypeToStr[descInfo.biasDtype].c_str(),
@@ -774,7 +795,7 @@ bool CheckParamsDtypeWithoutBias(const char *logTag, const Conv3DEngineDescInfo
774 return true;795 return true;
775 }796 }
776 797 
777- OP_LOGE(logTag, "unSupported params data type [fmap, weight, output]: [%s, %s, %s].",798+ OP_LOGE(logTag, "Unsupported params data type [fmap, weight, output]: [%s, %s, %s].",
778 g_convDtypeToStr[descInfo.fMapDtype].c_str(),799 g_convDtypeToStr[descInfo.fMapDtype].c_str(),
779 g_convDtypeToStr[descInfo.weightDtype].c_str(),800 g_convDtypeToStr[descInfo.weightDtype].c_str(),
780 g_convDtypeToStr[descInfo.outDtype].c_str());801 g_convDtypeToStr[descInfo.outDtype].c_str());
@@ -831,7 +852,7 @@ bool Conv3dTilingEngine::CheckInputFormat()
831 isPointWise ? "Pointwise" : (flagInfo_.hasScale ? "Quant" : "Regular"));852 isPointWise ? "Pointwise" : (flagInfo_.hasScale ? "Quant" : "Regular"));
832 853 
833 // Validate based on pointwise mode854 // Validate based on pointwise mode
834- if (isPointWise) { 855+ if (isPointWise) {
835 // Pointwise mode: all tensors must be NCDHW856 // Pointwise mode: all tensors must be NCDHW
836 if(!CheckValidFormatCombo(857 if(!CheckValidFormatCombo(
837 Conv3dApiTiling::ConvFormat::NCDHW,858 Conv3dApiTiling::ConvFormat::NCDHW,
@@ -984,6 +1005,39 @@ bool Conv3dTilingEngine::CheckInputShapeWithPad()
984 return true;1005 return true;
985}1006}
986 1007 
1008+bool Conv3dTilingEngine::CheckOutputShapeConsistency()
1009+{
1010+ int64_t expectDo = (static_cast<int64_t>(shapeInfo_.di) +
1011+ attrInfo_.padHead + attrInfo_.padTail -
1012+ attrInfo_.dilationD * (static_cast<int64_t>(shapeInfo_.kd) - 1LL) - 1LL) /
1013+ attrInfo_.strideD + 1LL;
1014+ int64_t expectHo = (static_cast<int64_t>(shapeInfo_.hi) +
1015+ attrInfo_.padTop + attrInfo_.padBottom -
1016+ attrInfo_.dilationH * (static_cast<int64_t>(shapeInfo_.kh) - 1LL) - 1LL) /
1017+ attrInfo_.strideH + 1LL;
1018+ int64_t expectWo = (static_cast<int64_t>(shapeInfo_.wi) +
1019+ attrInfo_.padLeft + attrInfo_.padRight -
1020+ attrInfo_.dilationW * (static_cast<int64_t>(shapeInfo_.kw) - 1LL) - 1LL) /
1021+ attrInfo_.strideW + 1LL;
1022+ 
1023+ if (outputBatch_ != static_cast<int64_t>(shapeInfo_.batch) ||
1024+ outputCOut_ != static_cast<int64_t>(shapeInfo_.cOut) ||
1025+ expectDo != static_cast<int64_t>(shapeInfo_.dOut) ||
1026+ expectHo != static_cast<int64_t>(shapeInfo_.ho) ||
1027+ expectWo != static_cast<int64_t>(shapeInfo_.wo)) {
1028+ OP_LOGE(logTag_.c_str(),
1029+ "Conv3D AscendC: output shape mismatch. expect [N,C,D,H,W]=[%ld,%ld,%ld,%ld,%ld], "
1030+ "actual [%ld,%ld,%ld,%ld,%ld].",
1031+ static_cast<int64_t>(shapeInfo_.batch), static_cast<int64_t>(shapeInfo_.cOut),
1032+ expectDo, expectHo, expectWo,
1033+ outputBatch_, outputCOut_, static_cast<int64_t>(shapeInfo_.dOut),
1034+ static_cast<int64_t>(shapeInfo_.ho), static_cast<int64_t>(shapeInfo_.wo));
1035+ return false;
1036+ }
1037+ 
1038+ return true;
1039+}
1040+ 
987bool Conv3dTilingEngine::CheckBiasShape()1041bool Conv3dTilingEngine::CheckBiasShape()
988{1042{
989 if (!flagInfo_.hasBias) {1043 if (!flagInfo_.hasBias) {
@@ -1197,6 +1251,8 @@ bool Conv3dTilingEngine::CheckAllParams()
1197 // Cross-parameter consistency + hardware limit checks.1251 // Cross-parameter consistency + hardware limit checks.
1198 {&Conv3dTilingEngine::CheckInputShapeWithPad,1252 {&Conv3dTilingEngine::CheckInputShapeWithPad,
1199 "CheckAllParams failed: input shape incompatible with pad/dilation/stride."},1253 "CheckAllParams failed: input shape incompatible with pad/dilation/stride."},
1254+ {&Conv3dTilingEngine::CheckOutputShapeConsistency,
1255+ "CheckAllParams failed: output shape is inconsistent with input/weight/attrs."},
1200 {&Conv3dTilingEngine::CheckLoad3DLimits,1256 {&Conv3dTilingEngine::CheckLoad3DLimits,
1201 "CheckAllParams failed: configuration violates LOAD3D hardware limits."},1257 "CheckAllParams failed: configuration violates LOAD3D hardware limits."},
1202 };1258 };
@@ -1376,11 +1432,11 @@ void Conv3dTilingEngine::CoreNumBlocksDecision()
1376 numBlocksRes_ = numBlocksResTmp;1432 numBlocksRes_ = numBlocksResTmp;
1377 1433 
1378 /**1434 /**
1379- * An optimized core partitioning logic is implemented here to ensure that the dout axis partitioning (doDim) is 1435+ * An optimized core partitioning logic is implemented here to ensure that the dout axis partitioning (doDim) is
1380 * used as fully as possible on the kernel side.1436 * used as fully as possible on the kernel side.
1381 * In order to ensure that part of the logic entering this optimized core partitioning does not degrade performance,1437 * In order to ensure that part of the logic entering this optimized core partitioning does not degrade performance,
1382 * the following two constraints are implemented:1438 * the following two constraints are implemented:
1383- * ** 1. This logic only updates the inter-core allocation for different axes (such as doDim) and the amount of 1439+ * ** 1. This logic only updates the inter-core allocation for different axes (such as doDim) and the amount of
1384 * data that each core needs to process (such as singleCoreDo)1440 * data that each core needs to process (such as singleCoreDo)
1385 * ** 2. Optimized core partitioning is only performed for cases that meet DO_DIM_FILTER_THRESHOLD1441 * ** 2. Optimized core partitioning is only performed for cases that meet DO_DIM_FILTER_THRESHOLD
1386 */1442 */
@@ -1390,11 +1446,11 @@ void Conv3dTilingEngine::CoreNumBlocksDecision()
1390 allRanges[NUMBLOCKS_DO_IDX] = numBlocksRanges_.doRange;1446 allRanges[NUMBLOCKS_DO_IDX] = numBlocksRanges_.doRange;
1391 NumBlocksDecisionBackTrack(numBlocksResTmp, allRanges, NUMBLOCKS_BATCH_IDX, dimsRecord);1447 NumBlocksDecisionBackTrack(numBlocksResTmp, allRanges, NUMBLOCKS_BATCH_IDX, dimsRecord);
1392 if (numBlocksRes_.doDim > DO_DIM_FILTER_THRESHOLD) {1448 if (numBlocksRes_.doDim > DO_DIM_FILTER_THRESHOLD) {
1393- OP_LOGD(logTag_.c_str(), "Using original block dimensions: doDim (%u) > threshold (%u), keeping original block dimensions", 1449+ OP_LOGD(logTag_.c_str(), "Using original block dimensions: doDim (%u) > threshold (%u), keeping original block dimensions",
1394 numBlocksRes_.doDim, DO_DIM_FILTER_THRESHOLD);1450 numBlocksRes_.doDim, DO_DIM_FILTER_THRESHOLD);
1395 numBlocksResOpt_ = numBlocksRes_;1451 numBlocksResOpt_ = numBlocksRes_;
1396 } else {1452 } else {
1397- OP_LOGD(logTag_.c_str(), "Entering block dimension optimization logic: original doDim (%u) <= threshold (%u), using filtered optimized doDim (%u)", 1453+ OP_LOGD(logTag_.c_str(), "Entering block dimension optimization logic: original doDim (%u) <= threshold (%u), using filtered optimized doDim (%u)",
1398 numBlocksRes_.doDim, DO_DIM_FILTER_THRESHOLD, numBlocksResTmp.doDim);1454 numBlocksRes_.doDim, DO_DIM_FILTER_THRESHOLD, numBlocksResTmp.doDim);
1399 numBlocksResOpt_ = numBlocksResTmp;1455 numBlocksResOpt_ = numBlocksResTmp;
1400 }1456 }
Mconv/conv3d_v2/op_host/op_tiling/conv3d_tiling_engine.h+6-0
@@ -166,10 +166,15 @@ public:
166private:166private:
167 std::string logTag_ {"Conv3DV2"};167 std::string logTag_ {"Conv3DV2"};
168 bool initOk_ = false;168 bool initOk_ = false;
169+ bool kernelPointWise_ = false;
170+ int64_t outputBatch_ = -1;
171+ int64_t outputCOut_ = -1;
169 std::vector<int64_t> biasShape_;172 std::vector<int64_t> biasShape_;
170 std::vector<int64_t> scaleShape_;173 std::vector<int64_t> scaleShape_;
171 174 
172 bool InitPlatformInfoFromAscendC();175 bool InitPlatformInfoFromAscendC();
176+ void UpdatePointWiseMode();
177+ bool UsesPointWisePath() const;
173 178 
174public:179public:
175 bool CheckStrideLegal();180 bool CheckStrideLegal();
@@ -182,6 +187,7 @@ public:
182 bool CheckPointWiseParams();187 bool CheckPointWiseParams();
183 bool CheckLoad3DLimits();188 bool CheckLoad3DLimits();
184 bool CheckInputShapeWithPad();189 bool CheckInputShapeWithPad();
190+ bool CheckOutputShapeConsistency();
185 bool CheckBiasShape();191 bool CheckBiasShape();
186 bool CheckScaleShape();192 bool CheckScaleShape();
187 bool CheckParamsOverflow();193 bool CheckParamsOverflow();
Mconv/conv3d_v2/op_host/op_tiling/conv3d_tilingdata.h+1-1
@@ -21,7 +21,7 @@
21#include "tiling/tiling_base.h"21#include "tiling/tiling_base.h"
22#include "register/op_impl_registry.h"22#include "register/op_impl_registry.h"
23 23 
24-// sizeof(TConv3DTiling) must algin with 8Byte, type sort24+// sizeof(TConv3DTiling) must align with 8Byte, type sort
25namespace optiling {25namespace optiling {
26} // namespace optiling26} // namespace optiling
27 27 
Mconv/conv3d_v2/op_host/op_tiling/conv3d_tuning_tiling.cpp+1-1
@@ -17,5 +17,5 @@ DECLARE_STRUCT_RELATE_WITH_OP_V2(Conv3D, Conv3DInputArgs, aDtype, bDtype, cDtype
17 strideH, strideW, dilationD, dilationH, dilationW, padHead, padTail, padTop,17 strideH, strideW, dilationD, dilationH, dilationW, padHead, padTail, padTop,
18 padBottom, padLeft, padRight, biasFlag);18 padBottom, padLeft, padRight, biasFlag);
19 19 
20-REGISTER_TUNING_TILING_CLASS(Conv3D, Conv3DTunnerTiling);20+REGISTER_TUNING_TILING_CLASS(Conv3D, Conv3DTunerTiling);
21} // namespace tuningtiling21} // namespace tuningtiling
Mconv/conv3d_v2/op_host/op_tiling/conv3d_tuning_tiling.h+53-53
@@ -52,7 +52,7 @@ struct Conv3DInputArgs {
52 bool biasFlag;52 bool biasFlag;
53};53};
54#pragma pack(pop)54#pragma pack(pop)
55-BEGIN_TUNING_TILING_DEF(Conv3DTunnerTiling)55+BEGIN_TUNING_TILING_DEF(Conv3DTunerTiling)
56TUNING_TILING_DATA_FIELD_DEF(uint64_t, groups);56TUNING_TILING_DATA_FIELD_DEF(uint64_t, groups);
57TUNING_TILING_DATA_FIELD_DEF(uint64_t, singleCoreDo);57TUNING_TILING_DATA_FIELD_DEF(uint64_t, singleCoreDo);
58TUNING_TILING_DATA_FIELD_DEF(uint64_t, singleCoreCo);58TUNING_TILING_DATA_FIELD_DEF(uint64_t, singleCoreCo);
@@ -105,57 +105,57 @@ TUNING_TILING_DATA_FIELD_DEF(uint8_t, groupDim);
105TUNING_TILING_DATA_FIELD_DEF(uint8_t, reserved1);105TUNING_TILING_DATA_FIELD_DEF(uint8_t, reserved1);
106TUNING_TILING_DATA_FIELD_DEF(uint8_t, reserved2);106TUNING_TILING_DATA_FIELD_DEF(uint8_t, reserved2);
107END_TUNING_TILING_DEF107END_TUNING_TILING_DEF
108-DECLARE_SCHEMA(Conv3DTunnerTiling,108+DECLARE_SCHEMA(Conv3DTunerTiling,
109- FIELD(Conv3DTunnerTiling, groups),109+ FIELD(Conv3DTunerTiling, groups),
110- FIELD(Conv3DTunnerTiling, singleCoreDo),110+ FIELD(Conv3DTunerTiling, singleCoreDo),
111- FIELD(Conv3DTunnerTiling, singleCoreCo),111+ FIELD(Conv3DTunerTiling, singleCoreCo),
112- FIELD(Conv3DTunnerTiling, singleCoreM),112+ FIELD(Conv3DTunerTiling, singleCoreM),
113- FIELD(Conv3DTunnerTiling, orgDo),113+ FIELD(Conv3DTunerTiling, orgDo),
114- FIELD(Conv3DTunnerTiling, orgCo),114+ FIELD(Conv3DTunerTiling, orgCo),
115- FIELD(Conv3DTunnerTiling, orgHo),115+ FIELD(Conv3DTunerTiling, orgHo),
116- FIELD(Conv3DTunnerTiling, orgWo),116+ FIELD(Conv3DTunerTiling, orgWo),
117- FIELD(Conv3DTunnerTiling, orgCi),117+ FIELD(Conv3DTunerTiling, orgCi),
118- FIELD(Conv3DTunnerTiling, orgDi),118+ FIELD(Conv3DTunerTiling, orgDi),
119- FIELD(Conv3DTunnerTiling, orgHi),119+ FIELD(Conv3DTunerTiling, orgHi),
120- FIELD(Conv3DTunnerTiling, orgWi),120+ FIELD(Conv3DTunerTiling, orgWi),
121- FIELD(Conv3DTunnerTiling, kernelD),121+ FIELD(Conv3DTunerTiling, kernelD),
122- FIELD(Conv3DTunnerTiling, kernelH),122+ FIELD(Conv3DTunerTiling, kernelH),
123- FIELD(Conv3DTunnerTiling, kernelW),123+ FIELD(Conv3DTunerTiling, kernelW),
124- FIELD(Conv3DTunnerTiling, strideD),124+ FIELD(Conv3DTunerTiling, strideD),
125- FIELD(Conv3DTunnerTiling, strideH),125+ FIELD(Conv3DTunerTiling, strideH),
126- FIELD(Conv3DTunnerTiling, strideW),126+ FIELD(Conv3DTunerTiling, strideW),
127- FIELD(Conv3DTunnerTiling, dilationD),127+ FIELD(Conv3DTunerTiling, dilationD),
128- FIELD(Conv3DTunnerTiling, dilationH),128+ FIELD(Conv3DTunerTiling, dilationH),
129- FIELD(Conv3DTunnerTiling, dilationW),129+ FIELD(Conv3DTunerTiling, dilationW),
130- FIELD(Conv3DTunnerTiling, padHead),130+ FIELD(Conv3DTunerTiling, padHead),
131- FIELD(Conv3DTunnerTiling, padTail),131+ FIELD(Conv3DTunerTiling, padTail),
132- FIELD(Conv3DTunnerTiling, padTop),132+ FIELD(Conv3DTunerTiling, padTop),
133- FIELD(Conv3DTunnerTiling, padBottom),133+ FIELD(Conv3DTunerTiling, padBottom),
134- FIELD(Conv3DTunnerTiling, padLeft),134+ FIELD(Conv3DTunerTiling, padLeft),
135- FIELD(Conv3DTunnerTiling, padRight),135+ FIELD(Conv3DTunerTiling, padRight),
136- FIELD(Conv3DTunnerTiling, mL0),136+ FIELD(Conv3DTunerTiling, mL0),
137- FIELD(Conv3DTunnerTiling, kL0),137+ FIELD(Conv3DTunerTiling, kL0),
138- FIELD(Conv3DTunnerTiling, nL0),138+ FIELD(Conv3DTunerTiling, nL0),
139- FIELD(Conv3DTunnerTiling, kAL1),139+ FIELD(Conv3DTunerTiling, kAL1),
140- FIELD(Conv3DTunnerTiling, kBL1),140+ FIELD(Conv3DTunerTiling, kBL1),
141- FIELD(Conv3DTunnerTiling, nBL1),141+ FIELD(Conv3DTunerTiling, nBL1),
142- FIELD(Conv3DTunnerTiling, mAL1),142+ FIELD(Conv3DTunerTiling, mAL1),
143- FIELD(Conv3DTunnerTiling, pBufferFlag),143+ FIELD(Conv3DTunerTiling, pBufferFlag),
144- FIELD(Conv3DTunnerTiling, offsetx),144+ FIELD(Conv3DTunerTiling, offsetx),
145- FIELD(Conv3DTunnerTiling, bl1FullLoad),145+ FIELD(Conv3DTunerTiling, bl1FullLoad),
146- FIELD(Conv3DTunnerTiling, al1FullLoad),146+ FIELD(Conv3DTunerTiling, al1FullLoad),
147- FIELD(Conv3DTunnerTiling, bl1BypassFlag),147+ FIELD(Conv3DTunerTiling, bl1BypassFlag),
148- FIELD(Conv3DTunnerTiling, iterateMNOrder),148+ FIELD(Conv3DTunerTiling, iterateMNOrder),
149- FIELD(Conv3DTunnerTiling, biasFullLoadFlag),149+ FIELD(Conv3DTunerTiling, biasFullLoadFlag),
150- FIELD(Conv3DTunnerTiling, fixpParamsFullLoadFlag),150+ FIELD(Conv3DTunerTiling, fixpParamsFullLoadFlag),
151- FIELD(Conv3DTunnerTiling, hf32Enable),151+ FIELD(Conv3DTunerTiling, hf32Enable),
152- FIELD(Conv3DTunnerTiling, hf32TransMode),152+ FIELD(Conv3DTunerTiling, hf32TransMode),
153- FIELD(Conv3DTunnerTiling, batchDim),153+ FIELD(Conv3DTunerTiling, batchDim),
154- FIELD(Conv3DTunnerTiling, nDim),154+ FIELD(Conv3DTunerTiling, nDim),
155- FIELD(Conv3DTunnerTiling, mDim),155+ FIELD(Conv3DTunerTiling, mDim),
156- FIELD(Conv3DTunnerTiling, doDim),156+ FIELD(Conv3DTunerTiling, doDim),
157- FIELD(Conv3DTunnerTiling, groupDim),157+ FIELD(Conv3DTunerTiling, groupDim),
158- FIELD(Conv3DTunnerTiling, reserved1),158+ FIELD(Conv3DTunerTiling, reserved1),
159- FIELD(Conv3DTunnerTiling, reserved2));159+ FIELD(Conv3DTunerTiling, reserved2));
160} // namespace tuningtiling160} // namespace tuningtiling
161#endif161#endif
Mconv/conv3d_v2/tests/ut/op_tiling/test_conv3d_engine_tiling.cpp+47-3
@@ -698,9 +698,40 @@ TEST(TestConv3dTilingEngine, CheckAttrLimits_MixedBoundaryValues)
698 EXPECT_TRUE(engine.CheckDilationLegal());698 EXPECT_TRUE(engine.CheckDilationLegal());
699}699}
700 700 
701-// Note about CheckOutputShape:701+TEST(TestConv3dTilingEngine, CheckOutputShapeConsistency_Positive)
702-// CheckOutputShape is declared in the header file but not implemented in the cpp file.702+{
703-// Unit tests for this method are skipped to avoid linking issues or relying on dead declarations.703+ Conv3dTilingEngine engine;
704+ InitSimpleConv3dEngine(engine);
705+ 
706+ EXPECT_TRUE(engine.CheckOutputShapeConsistency());
707+}
708+ 
709+TEST(TestConv3dTilingEngine, CheckOutputShapeConsistency_DimensionMismatch)
710+{
711+ Conv3dTilingEngine engine;
712+ InitSimpleConv3dEngine(engine);
713+ 
714+ engine.SetOrgOutputShape({1, 16, 2, 1, 1});
715+ EXPECT_FALSE(engine.CheckOutputShapeConsistency());
716+}
717+ 
718+TEST(TestConv3dTilingEngine, CheckOutputShapeConsistency_BatchOrChannelMismatch)
719+{
720+ Conv3dTilingEngine engine;
721+ InitSimpleConv3dEngine(engine);
722+ 
723+ engine.SetOrgOutputShape({2, 8, 1, 1, 1});
724+ EXPECT_FALSE(engine.CheckOutputShapeConsistency());
725+}
726+ 
727+TEST(TestConv3dTilingEngine, CheckAllParams_FailsOnOutputShapeMismatch)
728+{
729+ Conv3dTilingEngine engine;
730+ InitSimpleConv3dEngine(engine);
731+ 
732+ engine.SetOrgOutputShape({1, 16, 2, 1, 1});
733+ EXPECT_FALSE(engine.CheckAllParams());
734+}
704 735 
705TEST(TestConv3dTilingEngine, CheckParameterConsistency_GroupChannelMismatch)736TEST(TestConv3dTilingEngine, CheckParameterConsistency_GroupChannelMismatch)
706{737{
@@ -1539,6 +1570,19 @@ TEST(TestConv3dTilingEngine, CheckInputFormat_PointwiseModeDetection)
1539 EXPECT_TRUE(engine.CheckInputFormat());1570 EXPECT_TRUE(engine.CheckInputFormat());
1540}1571}
1541 1572 
1573+TEST(TestConv3dTilingEngine, CheckInputFormat_OneByOneByOneRegularFormatsStaysRegularMode)
1574+{
1575+ using Conv3dApiTiling::ConvFormat;
1576+ 
1577+ Conv3dTilingEngine engine;
1578+ InitSimpleConv3dEngine(engine);
1579+ 
1580+ engine.SetFormat(ConvFormat::NDC1HWC0, ConvFormat::FRACTAL_Z_3D, ConvFormat::NDC1HWC0);
1581+ 
1582+ EXPECT_FALSE(engine.isPointWise);
1583+ EXPECT_TRUE(engine.CheckInputFormat());
1584+}
1585+ 
1542TEST(TestConv3dTilingEngine, CheckInputFormat_AllFormatCombinations)1586TEST(TestConv3dTilingEngine, CheckInputFormat_AllFormatCombinations)
1543{1587{
1544 using Conv3dApiTiling::ConvFormat;1588 using Conv3dApiTiling::ConvFormat;
Mconv/conv3d_v2/tests/ut/op_tiling/test_conv3d_v2_tiling.cpp+4-2
@@ -569,7 +569,6 @@ TEST_F(Conv3DV2TilingRuntime, TestConv3DV2CheckPointWiseSuccess)
569 gert::TilingContext *tilingContext = holder.GetContext<gert::TilingContext>();569 gert::TilingContext *tilingContext = holder.GetContext<gert::TilingContext>();
570 570 
571 optiling::Conv3dOpsTiling::Conv3dBaseTiling conv3dBT(tilingContext);571 optiling::Conv3dOpsTiling::Conv3dBaseTiling conv3dBT(tilingContext);
572- conv3dBT.isPointWise = true;
573 conv3dBT.shapeInfo_.kh = 1;572 conv3dBT.shapeInfo_.kh = 1;
574 conv3dBT.shapeInfo_.kw = 1;573 conv3dBT.shapeInfo_.kw = 1;
575 conv3dBT.shapeInfo_.kd = 1;574 conv3dBT.shapeInfo_.kd = 1;
@@ -1331,7 +1330,8 @@ static Conv3dTilingEngine BuildEngineFromBase(const optiling::Conv3dOpsTiling::C
1331 engine.attrInfo_.groups = static_cast<int64_t>(base.attrInfo_.groups);1330 engine.attrInfo_.groups = static_cast<int64_t>(base.attrInfo_.groups);
1332 engine.attrInfo_.groupOpt = static_cast<int64_t>(base.attrInfo_.groupOpt);1331 engine.attrInfo_.groupOpt = static_cast<int64_t>(base.attrInfo_.groupOpt);
1333 engine.outputOrder_ = base.outputOrder_;1332 engine.outputOrder_ = base.outputOrder_;
1334- engine.isPointWise = base.isPointWise;1333+ engine.outputBatch_ = static_cast<int64_t>(base.shapeInfo_.batch);
1334+ engine.outputCOut_ = static_cast<int64_t>(base.shapeInfo_.cOut);
1335 1335 
1336 // Default dtype/format for tests (align with BF16 NCDHW setup)1336 // Default dtype/format for tests (align with BF16 NCDHW setup)
1337 engine.descInfo_.fMapDtype = Conv3dApiTiling::ConvDtype::BF16;1337 engine.descInfo_.fMapDtype = Conv3dApiTiling::ConvDtype::BF16;
@@ -1340,6 +1340,8 @@ static Conv3dTilingEngine BuildEngineFromBase(const optiling::Conv3dOpsTiling::C
1340 engine.descInfo_.fMapFormat = Conv3dApiTiling::ConvFormat::NCDHW;1340 engine.descInfo_.fMapFormat = Conv3dApiTiling::ConvFormat::NCDHW;
1341 engine.descInfo_.weightFormat = Conv3dApiTiling::ConvFormat::NCDHW;1341 engine.descInfo_.weightFormat = Conv3dApiTiling::ConvFormat::NCDHW;
1342 engine.descInfo_.outFormat = Conv3dApiTiling::ConvFormat::NCDHW;1342 engine.descInfo_.outFormat = Conv3dApiTiling::ConvFormat::NCDHW;
1343+ engine.kernelPointWise_ = engine.shapeInfo_.kd == 1 && engine.shapeInfo_.kh == 1 && engine.shapeInfo_.kw == 1;
1344+ engine.UpdatePointWiseMode();
1343 return engine;1345 return engine;
1344}1346}
1345 1347 
Mconv/convolution_forward/op_host/op_api/aclnn_convolution.cpp+31-31
@@ -1449,12 +1449,12 @@ private:
1449 FVector<int64_t>& padding, bool transposed)1449 FVector<int64_t>& padding, bool transposed)
1450 {1450 {
1451 FVector<int64_t> inputShape, weightShape;1451 FVector<int64_t> inputShape, weightShape;
1452- bool inputChannleLast, weightChannleLast;1452+ bool inputChannelLast, weightChannelLast;
1453 int64_t inputSpaceDimIndex, weightSpaceDimIndex;1453 int64_t inputSpaceDimIndex, weightSpaceDimIndex;
1454 size_t inputSpaceDimNum, weightSpaceDimNum;1454 size_t inputSpaceDimNum, weightSpaceDimNum;
1455 1455 
1456- GetSpatialDimInfo(input, inputChannleLast, inputSpaceDimIndex, inputSpaceDimNum, inputShape);1456+ GetSpatialDimInfo(input, inputChannelLast, inputSpaceDimIndex, inputSpaceDimNum, inputShape);
1457- GetSpatialDimInfo(weight, weightChannleLast, weightSpaceDimIndex, weightSpaceDimNum, weightShape);1457+ GetSpatialDimInfo(weight, weightChannelLast, weightSpaceDimIndex, weightSpaceDimNum, weightShape);
1458 1458 
1459 auto newpad = ConstructPad(padding, inputShape);1459 auto newpad = ConstructPad(padding, inputShape);
1460 for (size_t i = 0; i < inputSpaceDimNum; ++i) {1460 for (size_t i = 0; i < inputSpaceDimNum; ++i) {
@@ -1480,7 +1480,7 @@ private:
1480 if (inputShapeValueAfterPad < 0) {1480 if (inputShapeValueAfterPad < 0) {
1481 OP_LOGE(1481 OP_LOGE(
1482 ACLNN_ERR_PARAM_INVALID,1482 ACLNN_ERR_PARAM_INVALID,
1483- "after pad and dilation, expect input shape[%zu] should be greater than kernerl shape[%zu], "1483+ "after pad and dilation, expect input shape[%zu] should be greater than kernel shape[%zu], "
1484 "(input + pad - dilation * (weight - 1) - 1) should >= 0, actual get: %ld",1484 "(input + pad - dilation * (weight - 1) - 1) should >= 0, actual get: %ld",
1485 i, i, inputShapeValueAfterPad);1485 i, i, inputShapeValueAfterPad);
1486 return ACLNN_ERR_PARAM_INVALID;1486 return ACLNN_ERR_PARAM_INVALID;
@@ -1583,13 +1583,13 @@ private:
1583 1583 
1584 // check space(d h w or l)1584 // check space(d h w or l)
1585 FVector<int64_t> inputShape = input.shape;1585 FVector<int64_t> inputShape = input.shape;
1586- bool inputChannleLast = input.ChannelLast();1586+ bool inputChannelLast = input.ChannelLast();
1587- int64_t inputSpaceDimIndex = inputChannleLast ? 1 : 2; // 空间维度在shape中的起始位置，C维度后置时为1，否则为21587+ int64_t inputSpaceDimIndex = inputChannelLast ? 1 : 2; // 空间维度在shape中的起始位置，C维度后置时为1，否则为2
1588 size_t inputSpaceDimNum = input.shape.size() - 2; // 空间维度大小，1d卷积时为1，2d为2，3d为31588 size_t inputSpaceDimNum = input.shape.size() - 2; // 空间维度大小，1d卷积时为1，2d为2，3d为3
1589 1589 
1590 FVector<int64_t> weightShape = weight.shape;1590 FVector<int64_t> weightShape = weight.shape;
1591- bool weightChannleLast = weight.ChannelLast();1591+ bool weightChannelLast = weight.ChannelLast();
1592- int64_t weightSpaceDimIndex = weightChannleLast ? 1 : 2; // 空间维度在shape中的起始位置，C维度后置时为1，否则为21592+ int64_t weightSpaceDimIndex = weightChannelLast ? 1 : 2; // 空间维度在shape中的起始位置，C维度后置时为1，否则为2
1593 1593 
1594 // 假设是NCL，判断L的值。假设是NCHW，判断HW的值1594 // 假设是NCL，判断L的值。假设是NCHW，判断HW的值
1595 for (size_t i = 0; i < inputSpaceDimNum; ++i) {1595 for (size_t i = 0; i < inputSpaceDimNum; ++i) {
@@ -1649,7 +1649,7 @@ public:
1649 1649 
1650private:1650private:
1651 /*1651 /*
1652- input weight output 的shape均瑶大于等于01652+ input weight output 的shape均需大于等于0
1653 bias（一维）的值要等于channel_out1653 bias（一维）的值要等于channel_out
1654 */1654 */
1655 aclnnStatus CheckShapeTbc(TensorMeta& input, TensorMeta& weight, TensorMeta& output) const1655 aclnnStatus CheckShapeTbc(TensorMeta& input, TensorMeta& weight, TensorMeta& output) const
@@ -1769,12 +1769,12 @@ private:
1769 1769 
1770 // check space(d h w or l)1770 // check space(d h w or l)
1771 FVector<int64_t> inputShape, weightShape;1771 FVector<int64_t> inputShape, weightShape;
1772- bool inputChannleLast, weightChannleLast;1772+ bool inputChannelLast, weightChannelLast;
1773 int64_t inputSpaceDimIndex, weightSpaceDimIndex;1773 int64_t inputSpaceDimIndex, weightSpaceDimIndex;
1774 size_t inputSpaceDimNum, weightSpaceDimNum;1774 size_t inputSpaceDimNum, weightSpaceDimNum;
1775 1775 
1776- GetSpatialDimInfo(input, inputChannleLast, inputSpaceDimIndex, inputSpaceDimNum, inputShape);1776+ GetSpatialDimInfo(input, inputChannelLast, inputSpaceDimIndex, inputSpaceDimNum, inputShape);
1777- GetSpatialDimInfo(weight, weightChannleLast, weightSpaceDimIndex, weightSpaceDimNum, weightShape);1777+ GetSpatialDimInfo(weight, weightChannelLast, weightSpaceDimIndex, weightSpaceDimNum, weightShape);
1778 for (size_t i = 0; i < inputSpaceDimNum; ++i) {1778 for (size_t i = 0; i < inputSpaceDimNum; ++i) {
1779 int64_t inputShapeSpace = inputShape[i + inputSpaceDimIndex]; // 空间维度的值1779 int64_t inputShapeSpace = inputShape[i + inputSpaceDimIndex]; // 空间维度的值
1780 if (inputShapeSpace < 0) {1780 if (inputShapeSpace < 0) {
@@ -1810,12 +1810,12 @@ private:
1810 const FVector<int64_t>& padding, bool transposed)1810 const FVector<int64_t>& padding, bool transposed)
1811 {1811 {
1812 FVector<int64_t> inputShape, weightShape;1812 FVector<int64_t> inputShape, weightShape;
1813- bool inputChannleLast, weightChannleLast;1813+ bool inputChannelLast, weightChannelLast;
1814 int64_t inputSpaceDimIndex, weightSpaceDimIndex;1814 int64_t inputSpaceDimIndex, weightSpaceDimIndex;
1815 size_t inputSpaceDimNum, weightSpaceDimNum;1815 size_t inputSpaceDimNum, weightSpaceDimNum;
1816 1816 
1817- GetSpatialDimInfo(input, inputChannleLast, inputSpaceDimIndex, inputSpaceDimNum, inputShape);1817+ GetSpatialDimInfo(input, inputChannelLast, inputSpaceDimIndex, inputSpaceDimNum, inputShape);
1818- GetSpatialDimInfo(weight, weightChannleLast, weightSpaceDimIndex, weightSpaceDimNum, weightShape);1818+ GetSpatialDimInfo(weight, weightChannelLast, weightSpaceDimIndex, weightSpaceDimNum, weightShape);
1819 1819 
1820 for (size_t i = 0; i < inputSpaceDimNum; ++i) {1820 for (size_t i = 0; i < inputSpaceDimNum; ++i) {
1821 auto inputShapeValue = inputShape[i + inputSpaceDimIndex];1821 auto inputShapeValue = inputShape[i + inputSpaceDimIndex];
@@ -3553,12 +3553,12 @@ aclIntArray* ViewValueAs1d(const int64_t value, aclOpExecutor* executor)
3553 3553 
3554const aclTensor* View1dAs4d(const aclTensor* input, aclOpExecutor* executor)3554const aclTensor* View1dAs4d(const aclTensor* input, aclOpExecutor* executor)
3555{3555{
3556- // input NCL->contigious->unsqueeze(2)->reformat NCHW3556+ // input NCL->contiguous->unsqueeze(2)->reformat NCHW
3557- // 非连续转连续contigious3557+ // 非连续转连续contiguous
3558 auto contiguousInput = l0op::Contiguous(input, executor);3558 auto contiguousInput = l0op::Contiguous(input, executor);
3559 CHECK_RET(contiguousInput != nullptr, nullptr);3559 CHECK_RET(contiguousInput != nullptr, nullptr);
3560 3560 
3561- // unsqeeze(2)3561+ // unsqueeze(2)
3562 constexpr int64_t appendDim[] = {0, 2, 3};3562 constexpr int64_t appendDim[] = {0, 2, 3};
3563 aclIntArray* dim = executor->AllocIntArray(appendDim, 3);3563 aclIntArray* dim = executor->AllocIntArray(appendDim, 3);
3564 auto unsqueezedInput = l0op::UnsqueezeNd(contiguousInput, dim, executor);3564 auto unsqueezedInput = l0op::UnsqueezeNd(contiguousInput, dim, executor);
@@ -3573,12 +3573,12 @@ const aclTensor* View1dAs4d(const aclTensor* input, aclOpExecutor* executor)
3573 3573 
3574static const aclTensor* View3dAs4d(const aclTensor* input, aclOpExecutor* executor)3574static const aclTensor* View3dAs4d(const aclTensor* input, aclOpExecutor* executor)
3575{3575{
3576- // input NCL->contigious->unsqueeze(2)->reformat NCHW3576+ // input NCL->contiguous->unsqueeze(2)->reformat NCHW
3577- // 非连续转连续contigious3577+ // 非连续转连续contiguous
3578 auto contiguousInput = l0op::Contiguous(input, executor);3578 auto contiguousInput = l0op::Contiguous(input, executor);
3579 CHECK_RET(contiguousInput != nullptr, nullptr);3579 CHECK_RET(contiguousInput != nullptr, nullptr);
3580 3580 
3581- // unsqeeze(2)3581+ // unsqueeze(2)
3582 constexpr int64_t appendDim[] = {2};3582 constexpr int64_t appendDim[] = {2};
3583 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);3583 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);
3584 auto unsqueezedInput = l0op::UnsqueezeNd(contiguousInput, dim, executor);3584 auto unsqueezedInput = l0op::UnsqueezeNd(contiguousInput, dim, executor);
@@ -3593,12 +3593,12 @@ static const aclTensor* View3dAs4d(const aclTensor* input, aclOpExecutor* execut
3593 3593 
3594static const aclTensor* View3dAs4dw(const aclTensor* input, aclOpExecutor* executor)3594static const aclTensor* View3dAs4dw(const aclTensor* input, aclOpExecutor* executor)
3595{3595{
3596- // input NCL->contigious->unsqueeze(2)->reshape->reformat NCHW3596+ // input NCL->contiguous->unsqueeze(2)->reshape->reformat NCHW
3597- // 非连续转连续contigious3597+ // 非连续转连续contiguous
3598 auto contiguousInput = l0op::Contiguous(input, executor);3598 auto contiguousInput = l0op::Contiguous(input, executor);
3599 CHECK_RET(contiguousInput != nullptr, nullptr);3599 CHECK_RET(contiguousInput != nullptr, nullptr);
3600 3600 
3601- // unsqeeze(2) 扩w维度3601+ // unsqueeze(2) 扩w维度
3602 constexpr int64_t appendDim[] = {2};3602 constexpr int64_t appendDim[] = {2};
3603 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);3603 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);
3604 auto unsqueezedInput = l0op::UnsqueezeNd(contiguousInput, dim, executor);3604 auto unsqueezedInput = l0op::UnsqueezeNd(contiguousInput, dim, executor);
@@ -3622,11 +3622,11 @@ static const aclTensor* View3dAs4dw(const aclTensor* input, aclOpExecutor* execu
3622 3622 
3623static const aclTensor* View4dAs3d(const aclTensor* input, aclOpExecutor* executor)3623static const aclTensor* View4dAs3d(const aclTensor* input, aclOpExecutor* executor)
3624{3624{
3625- // input NCL->contigious->unsqueeze(2)->reformat NCHW3625+ // input NCL->contiguous->unsqueeze(2)->reformat NCHW
3626- // 非连续转连续contigious3626+ // 非连续转连续contiguous
3627 auto contiguousInput = l0op::Contiguous(input, executor);3627 auto contiguousInput = l0op::Contiguous(input, executor);
3628 CHECK_RET(contiguousInput != nullptr, nullptr);3628 CHECK_RET(contiguousInput != nullptr, nullptr);
3629- // sqeeze(2)3629+ // squeeze(2)
3630 constexpr int64_t appendDim[] = {2};3630 constexpr int64_t appendDim[] = {2};
3631 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);3631 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);
3632 CHECK_RET(dim != nullptr, nullptr);3632 CHECK_RET(dim != nullptr, nullptr);
@@ -3642,8 +3642,8 @@ static const aclTensor* View4dAs3d(const aclTensor* input, aclOpExecutor* execut
3642 3642 
3643static const aclTensor* View4dAs3dw(const aclTensor* input, aclOpExecutor* executor)3643static const aclTensor* View4dAs3dw(const aclTensor* input, aclOpExecutor* executor)
3644{3644{
3645- // input NCL->contigious->Reshape->unsqueeze(2)->reformat NCHW3645+ // input NCL->contiguous->Reshape->unsqueeze(2)->reformat NCHW
3646- // 非连续转连续contigious3646+ // 非连续转连续contiguous
3647 auto contiguousInput = l0op::Contiguous(input, executor);3647 auto contiguousInput = l0op::Contiguous(input, executor);
3648 CHECK_RET(contiguousInput != nullptr, nullptr);3648 CHECK_RET(contiguousInput != nullptr, nullptr);
3649 3649 
@@ -3655,7 +3655,7 @@ static const aclTensor* View4dAs3dw(const aclTensor* input, aclOpExecutor* execu
3655 CHECK_RET(shapeArray != nullptr, nullptr);3655 CHECK_RET(shapeArray != nullptr, nullptr);
3656 contiguousInput = l0op::Reshape(contiguousInput, shapeArray, executor);3656 contiguousInput = l0op::Reshape(contiguousInput, shapeArray, executor);
3657 CHECK_RET(contiguousInput != nullptr, nullptr);3657 CHECK_RET(contiguousInput != nullptr, nullptr);
3658- // sqeeze(3)3658+ // squeeze(3)
3659 constexpr int64_t appendDim[] = {2};3659 constexpr int64_t appendDim[] = {2};
3660 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);3660 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);
3661 CHECK_RET(dim != nullptr, nullptr);3661 CHECK_RET(dim != nullptr, nullptr);
@@ -3671,7 +3671,7 @@ static const aclTensor* View4dAs3dw(const aclTensor* input, aclOpExecutor* execu
3671 3671 
3672static const aclTensor* Permute(const aclTensor* input, FVector<int64_t> dims, aclOpExecutor* executor)3672static const aclTensor* Permute(const aclTensor* input, FVector<int64_t> dims, aclOpExecutor* executor)
3673{3673{
3674- // contigious3674+ // contiguous
3675 auto contiguousInput = l0op::Contiguous(input, executor);3675 auto contiguousInput = l0op::Contiguous(input, executor);
3676 CHECK_RET(contiguousInput != nullptr, nullptr);3676 CHECK_RET(contiguousInput != nullptr, nullptr);
3677 // Transpose3677 // Transpose
Mconv/convolution_forward/op_host/op_api/aclnn_quant_convolution.cpp+15-15
@@ -357,8 +357,8 @@ private:
357 FVector<int64_t> weightShape = meta.weight.shape;357 FVector<int64_t> weightShape = meta.weight.shape;
358 358 
359 auto newPad = ConstructPadding(meta.padding, inputShape);359 auto newPad = ConstructPadding(meta.padding, inputShape);
360- int64_t inferedShapeSize = inputShape.size() - 2;360+ int64_t inferredShapeSize = inputShape.size() - 2;
361- for (int64_t i = 0; i < inferedShapeSize; ++i) {361+ for (int64_t i = 0; i < inferredShapeSize; ++i) {
362 int64_t xOut = (inputShape[i + INPUT_C_INDEX + 1] + newPad[i] - meta.dilation[i] *362 int64_t xOut = (inputShape[i + INPUT_C_INDEX + 1] + newPad[i] - meta.dilation[i] *
363 (weightShape[i + INPUT_C_INDEX + 1] - 1) - 1) / meta.stride[i] + 1;363 (weightShape[i + INPUT_C_INDEX + 1] - 1) - 1) / meta.stride[i] + 1;
364 output.push_back(xOut);364 output.push_back(xOut);
@@ -657,17 +657,17 @@ public:
657 return ACLNN_ERR_PARAM_INVALID;657 return ACLNN_ERR_PARAM_INVALID;
658 }658 }
659private:659private:
660- bool QuantConvDtypesMatch(const std::vector<DataType>& matchedList, const std::vector<DataType>& supporedList,660+ bool QuantConvDtypesMatch(const std::vector<DataType>& matchedList, const std::vector<DataType>& supportedList,
661 size_t checkedLength) const661 size_t checkedLength) const
662 {662 {
663- if (matchedList.size() > supporedList.size()) {663+ if (matchedList.size() > supportedList.size()) {
664 return false;664 return false;
665 }665 }
666- if (checkedLength > supporedList.size()) {666+ if (checkedLength > supportedList.size()) {
667 return false;667 return false;
668 }668 }
669 for (size_t i = 0; i < checkedLength; i++) {669 for (size_t i = 0; i < checkedLength; i++) {
670- if (matchedList[i] != supporedList[i]) {670+ if (matchedList[i] != supportedList[i]) {
671 return false;671 return false;
672 }672 }
673 }673 }
@@ -721,7 +721,7 @@ private:
721 return ACLNN_SUCCESS;721 return ACLNN_SUCCESS;
722 }722 }
723 723 
724- static aclnnStatus CheckVaildString(const string &inputStr)724+ static aclnnStatus CheckValidString(const string &inputStr)
725 {725 {
726 if (inputStr.empty()) {726 if (inputStr.empty()) {
727 return ACLNN_SUCCESS;727 return ACLNN_SUCCESS;
@@ -771,7 +771,7 @@ private:
771 }771 }
772 OP_LOGW("the input roundMode is suggested to be set as a nullptr");772 OP_LOGW("the input roundMode is suggested to be set as a nullptr");
773 roundModeStr = std::string(engine.params.roundMode);773 roundModeStr = std::string(engine.params.roundMode);
774- if (CheckVaildString(roundModeStr) != ACLNN_SUCCESS) {774+ if (CheckValidString(roundModeStr) != ACLNN_SUCCESS) {
775 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "the input roundMode has invalid str");775 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "the input roundMode has invalid str");
776 return ACLNN_ERR_PARAM_INVALID;776 return ACLNN_ERR_PARAM_INVALID;
777 }777 }
@@ -830,7 +830,7 @@ private:
830 static aclnnStatus CheckShapeValue(QuantConvEngine &engine)830 static aclnnStatus CheckShapeValue(QuantConvEngine &engine)
831 {831 {
832 int64_t inputDimN = engine.meta.input.N();832 int64_t inputDimN = engine.meta.input.N();
833- int64_t iuputDimC = engine.meta.input.C();833+ int64_t inputDimC = engine.meta.input.C();
834 int64_t inputDimD = engine.meta.input.D();834 int64_t inputDimD = engine.meta.input.D();
835 int64_t inputDimH = engine.meta.input.H();835 int64_t inputDimH = engine.meta.input.H();
836 int64_t inputDimW = engine.meta.input.W();836 int64_t inputDimW = engine.meta.input.W();
@@ -840,14 +840,14 @@ private:
840 int64_t weightDimH = engine.meta.weight.H();840 int64_t weightDimH = engine.meta.weight.H();
841 int64_t weightDimW = engine.meta.weight.W();841 int64_t weightDimW = engine.meta.weight.W();
842 842 
843- // enbale empty tensor843+ // enable empty tensor
844 CHECK_PARAMS_GT(inputDimN, 0L);844 CHECK_PARAMS_GT(inputDimN, 0L);
845 CHECK_PARAMS_GT(inputDimD, 0L);845 CHECK_PARAMS_GT(inputDimD, 0L);
846 CHECK_PARAMS_GT(inputDimH, 0L);846 CHECK_PARAMS_GT(inputDimH, 0L);
847 CHECK_PARAMS_GT(inputDimW, 0L);847 CHECK_PARAMS_GT(inputDimW, 0L);
848 CHECK_PARAMS_GT(weightDimN, 0L);848 CHECK_PARAMS_GT(weightDimN, 0L);
849 849 
850- CHECK_PARAMS_GT(iuputDimC, 0L);850+ CHECK_PARAMS_GT(inputDimC, 0L);
851 CHECK_PARAMS_GT(weightDimC, 0L);851 CHECK_PARAMS_GT(weightDimC, 0L);
852 CHECK_PARAMS_GT(weightDimD, 0L);852 CHECK_PARAMS_GT(weightDimD, 0L);
853 CHECK_PARAMS_GT(weightDimH, 0L);853 CHECK_PARAMS_GT(weightDimH, 0L);
@@ -932,11 +932,11 @@ private:
932 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "check output value greater than 0 failed");932 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "check output value greater than 0 failed");
933 return ACLNN_ERR_PARAM_INVALID;933 return ACLNN_ERR_PARAM_INVALID;
934 }934 }
935- auto inferedOutputShape = engine.CalcOutputShape();935+ auto inferredOutputShape = engine.CalcOutputShape();
936- for (size_t i = 0; i < inferedOutputShape.size(); i++) {936+ for (size_t i = 0; i < inferredOutputShape.size(); i++) {
937- if (inferedOutputShape[i] != outputShape[i]) {937+ if (inferredOutputShape[i] != outputShape[i]) {
938 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "expected output %zuth dim equal %ld, get %ld", i + 1,938 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "expected output %zuth dim equal %ld, get %ld", i + 1,
939- inferedOutputShape[i], outputShape[i]);939+ inferredOutputShape[i], outputShape[i]);
940 return ACLNN_ERR_PARAM_INVALID;940 return ACLNN_ERR_PARAM_INVALID;
941 }941 }
942 }942 }
Mconv/convolution_forward/op_host/op_api/aclnn_quant_convolution.h+2-2
@@ -74,7 +74,7 @@ ACLNN_API aclnnStatus aclnnQuantConvolutionGetWorkspaceSize(const aclTensor* inp
74 const char* roundMode, aclTensor* output,74 const char* roundMode, aclTensor* output,
75 uint64_t* workspaceSize, aclOpExecutor** executor);75 uint64_t* workspaceSize, aclOpExecutor** executor);
76/**76/**
77- * @brief quant convolution接口，进行kernellaunch77+ * @brief quant convolution接口，进行kernel launch
78 *78 *
79 * @param [in] workspace: 在npu device侧申请的workspace内存起址。79 * @param [in] workspace: 在npu device侧申请的workspace内存起址。
80 * @param [in] workspaceSize: 在npu device侧申请的workspace大小，由aclnnQuantConvolutionGetWorkspaceSize获取。80 * @param [in] workspaceSize: 在npu device侧申请的workspace大小，由aclnnQuantConvolutionGetWorkspaceSize获取。
@@ -101,7 +101,7 @@ ACLNN_API aclnnStatus aclnnQuantConvolutionWeightNzGetWorkspaceSize(const aclTen
101 const char* roundMode, aclTensor* output,101 const char* roundMode, aclTensor* output,
102 uint64_t* workspaceSize, aclOpExecutor** executor);102 uint64_t* workspaceSize, aclOpExecutor** executor);
103/**103/**
104- * @brief quant convolution weightNz接口，进行kernellaunch104+ * @brief quant convolution weightNz接口，进行kernel launch
105 *105 *
106 * @param [in] workspace: 在npu device侧申请的workspace内存起址。106 * @param [in] workspace: 在npu device侧申请的workspace内存起址。
107 * @param [in] workspaceSize: 在npu device侧申请的workspace大小，由aclnnQuantConvolutionWeightNzGetWorkspaceSize获取。107 * @param [in] workspaceSize: 在npu device侧申请的workspace大小，由aclnnQuantConvolutionWeightNzGetWorkspaceSize获取。
Mconv/convolution_forward/op_host/op_api/convolution.cpp+14-14
@@ -102,7 +102,7 @@ const vector<vector<int64_t>> CONV2D_TRANSPOSE_V2_WHITE_LIST =
102 4, 320, 80, 80, // input shape102 4, 320, 80, 80, // input shape
103 320, 320, 3, 3, // filter shape103 320, 320, 3, 3, // filter shape
104 4, 320, 80, 80, // outBackprop shape104 4, 320, 80, 80, // outBackprop shape
105- 1, 1, // stide105+ 1, 1, // stride
106 1, 1, // padding106 1, 1, // padding
107 1, 1, // dilation107 1, 1, // dilation
108 0, 0, // output padding108 0, 0, // output padding
@@ -117,7 +117,7 @@ const vector<vector<int64_t>> CONV3D_TRANSPOSE_V2_WHITE_LIST =
117 1, 256, 62, 66, 66, // input shape117 1, 256, 62, 66, 66, // input shape
118 256, 256, 4, 4, 4, // filter shape118 256, 256, 4, 4, 4, // filter shape
119 1, 256, 120, 128, 128, // outBackprop shape119 1, 256, 120, 128, 128, // outBackprop shape
120- 2, 2, 2, // stide120+ 2, 2, 2, // stride
121 3, 3, 3, // padding121 3, 3, 3, // padding
122 1, 1, 1, // dilation122 1, 1, 1, // dilation
123 0, 0, 0, // output padding123 0, 0, 0, // output padding
@@ -473,17 +473,17 @@ static aclIntArray* ConstructConv2DNewStride(const aclTensor *input, const aclIn
473 473 
474static aclIntArray* ConstructConv2DNewDilation(const aclTensor *input, const aclIntArray *dilation, aclOpExecutor *executor)474static aclIntArray* ConstructConv2DNewDilation(const aclTensor *input, const aclIntArray *dilation, aclOpExecutor *executor)
475{475{
476- FVector<int64_t> newDalition;476+ FVector<int64_t> newDilation;
477 if (dilation->Size() < DIM_2) {477 if (dilation->Size() < DIM_2) {
478- newDalition = {0};478+ newDilation = {0};
479- return executor->AllocIntArray(newDalition.data(), newDalition.size());479+ return executor->AllocIntArray(newDilation.data(), newDilation.size());
480 }480 }
481 if (input->GetOriginalFormat() == op::Format::FORMAT_NCHW) {481 if (input->GetOriginalFormat() == op::Format::FORMAT_NCHW) {
482- newDalition = {1, 1, (*dilation)[0], (*dilation)[1]};482+ newDilation = {1, 1, (*dilation)[0], (*dilation)[1]};
483 } else {483 } else {
484- newDalition = {1, (*dilation)[0], (*dilation)[1], 1};484+ newDilation = {1, (*dilation)[0], (*dilation)[1], 1};
485 }485 }
486- return executor->AllocIntArray(newDalition.data(), conv2dDimNum);486+ return executor->AllocIntArray(newDilation.data(), conv2dDimNum);
487}487}
488 488 
489static aclnnStatus Conv2dV2InferShapeAndAddLauncher(const aclTensor *input, const aclTensor *weight,489static aclnnStatus Conv2dV2InferShapeAndAddLauncher(const aclTensor *input, const aclTensor *weight,
@@ -632,14 +632,14 @@ static aclnnStatus Conv3dWithFlag(const aclTensor *input, const aclTensor *weigh
632 aclIntArray *dilation5;632 aclIntArray *dilation5;
633 if (input->GetOriginalFormat() == op::Format::FORMAT_NCDHW) {633 if (input->GetOriginalFormat() == op::Format::FORMAT_NCDHW) {
634 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1], (*stride)[2]};634 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1], (*stride)[2]};
635- FVector<int64_t> newDalition{1, 1, (*dilation)[0], (*dilation)[1], (*dilation)[2]};635+ FVector<int64_t> newDilation{1, 1, (*dilation)[0], (*dilation)[1], (*dilation)[2]};
636 stride5 = executor->AllocIntArray(newStrides.data(), conv3dDimNum);636 stride5 = executor->AllocIntArray(newStrides.data(), conv3dDimNum);
637- dilation5 = executor->AllocIntArray(newDalition.data(), conv3dDimNum);637+ dilation5 = executor->AllocIntArray(newDilation.data(), conv3dDimNum);
638 } else {638 } else {
639 FVector<int64_t> newStrides{1, (*stride)[0], (*stride)[1], (*stride)[2], 1};639 FVector<int64_t> newStrides{1, (*stride)[0], (*stride)[1], (*stride)[2], 1};
640- FVector<int64_t> newDalition{1, (*dilation)[0], (*dilation)[1], (*dilation)[2], 1};640+ FVector<int64_t> newDilation{1, (*dilation)[0], (*dilation)[1], (*dilation)[2], 1};
641 stride5 = executor->AllocIntArray(newStrides.data(), conv3dDimNum);641 stride5 = executor->AllocIntArray(newStrides.data(), conv3dDimNum);
642- dilation5 = executor->AllocIntArray(newDalition.data(), conv3dDimNum);642+ dilation5 = executor->AllocIntArray(newDilation.data(), conv3dDimNum);
643 }643 }
644 644 
645 FVector<int64_t> newPad{(*padding)[0], (*padding)[0], (*padding)[1], (*padding)[1], (*padding)[2], (*padding)[2]};645 FVector<int64_t> newPad{(*padding)[0], (*padding)[0], (*padding)[1], (*padding)[1], (*padding)[2], (*padding)[2]};
@@ -731,9 +731,9 @@ static aclnnStatus Conv3dv2WithFlag(const aclTensor *input, const aclTensor *wei
731 return ACLNN_ERR_INNER;731 return ACLNN_ERR_INNER;
732 }732 }
733 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1], (*stride)[2]};733 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1], (*stride)[2]};
734- FVector<int64_t> newDalition{1, 1, (*dilation)[0], (*dilation)[1], (*dilation)[2]};734+ FVector<int64_t> newDilation{1, 1, (*dilation)[0], (*dilation)[1], (*dilation)[2]};
735 stride5 = executor->AllocIntArray(newStrides.data(), conv3dDimNum);735 stride5 = executor->AllocIntArray(newStrides.data(), conv3dDimNum);
736- dilation5 = executor->AllocIntArray(newDalition.data(), conv3dDimNum);736+ dilation5 = executor->AllocIntArray(newDilation.data(), conv3dDimNum);
737 737 
738 aclIntArray *pad6 = ConstructConv3DNewPad(padding, executor);738 aclIntArray *pad6 = ConstructConv3DNewPad(padding, executor);
739 if (pad6->Size() != PAD_DIM_6) {739 if (pad6->Size() != PAD_DIM_6) {
Mconv/convolution_forward/op_host/op_api/convolution_util.cpp+8-8
@@ -224,7 +224,7 @@ bool Conv2DSplitWInfo::CheckLoadL1InSplitW(const aclTensor* bias, aclTensor* out
224 return true;224 return true;
225}225}
226 226 
227-aclIntArray* View2dAs3dForAttr(const aclIntArray* intArray, int64_t expendValue, aclOpExecutor* executor, bool isPad)227+aclIntArray* View2dAs3dForAttr(const aclIntArray* intArray, int64_t expandValue, aclOpExecutor* executor, bool isPad)
228{228{
229 int64_t data[SplitWInfo::CONV3D_ATTR_NUM];229 int64_t data[SplitWInfo::CONV3D_ATTR_NUM];
230 uint64_t size = intArray->Size();230 uint64_t size = intArray->Size();
@@ -232,7 +232,7 @@ aclIntArray* View2dAs3dForAttr(const aclIntArray* intArray, int64_t expendValue,
232 if (!isPad && (size != static_cast<uint64_t>(SplitWInfo::CONV3D_ATTR_NUM - 1))) {232 if (!isPad && (size != static_cast<uint64_t>(SplitWInfo::CONV3D_ATTR_NUM - 1))) {
233 return nullptr;233 return nullptr;
234 }234 }
235- data[0] = expendValue;235+ data[0] = expandValue;
236 data[1] = (*intArray)[0];236 data[1] = (*intArray)[0];
237 if (isPad) {237 if (isPad) {
238 data[SplitWInfo::W_INDEX_ATTR_CONV3D] = (*intArray)[SplitWInfo::LEFT_INDEX_ATTR];238 data[SplitWInfo::W_INDEX_ATTR_CONV3D] = (*intArray)[SplitWInfo::LEFT_INDEX_ATTR];
@@ -272,12 +272,12 @@ const aclTensor* View4DSwapHWForTensor(const aclTensor* input, aclOpExecutor* ex
272 272 
273const aclTensor* View4dAs5dForInput(const aclTensor* input, aclOpExecutor* executor)273const aclTensor* View4dAs5dForInput(const aclTensor* input, aclOpExecutor* executor)
274{274{
275- // input NCHW->contigious->unsqueeze(2)->reformat NCDHW275+ // input NCHW->contiguous->unsqueeze(2)->reformat NCDHW
276- // 非连续转连续contigious276+ // 非连续转连续contiguous
277 auto contiguousInput = l0op::Contiguous(input, executor);277 auto contiguousInput = l0op::Contiguous(input, executor);
278 CHECK_RET(contiguousInput != nullptr, nullptr);278 CHECK_RET(contiguousInput != nullptr, nullptr);
279 279 
280- // unsqeeze(2)280+ // unsqueeze(2)
281 const int64_t appendDim[] = {SplitWInfo::HI_INDEX};281 const int64_t appendDim[] = {SplitWInfo::HI_INDEX};
282 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);282 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);
283 CHECK_RET(dim != nullptr, nullptr);283 CHECK_RET(dim != nullptr, nullptr);
@@ -314,11 +314,11 @@ aclnnStatus ChangeConv2dInputToConv3d(const aclTensor* &input, const aclTensor*
314 314 
315const aclTensor* View5dAs4dForOutput(const aclTensor* input, aclOpExecutor* executor)315const aclTensor* View5dAs4dForOutput(const aclTensor* input, aclOpExecutor* executor)
316{316{
317- // input NCDHW->contigious->squeeze(2)->reformat NCHW317+ // input NCDHW->contiguous->squeeze(2)->reformat NCHW
318- // 非连续转连续contigious318+ // 非连续转连续contiguous
319 auto contiguousInput = l0op::Contiguous(input, executor);319 auto contiguousInput = l0op::Contiguous(input, executor);
320 CHECK_RET(contiguousInput != nullptr, nullptr);320 CHECK_RET(contiguousInput != nullptr, nullptr);
321- // sqeeze(2)321+ // squeeze(2)
322 const int64_t appendDim[] = {SplitWInfo::HI_INDEX};322 const int64_t appendDim[] = {SplitWInfo::HI_INDEX};
323 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);323 aclIntArray* dim = executor->AllocIntArray(appendDim, 1);
324 CHECK_RET(dim != nullptr, nullptr);324 CHECK_RET(dim != nullptr, nullptr);
Mconv/convolution_forward/op_host/op_api/convolution_util.h+1-1
@@ -78,7 +78,7 @@ aclnnStatus ChangeConv2dAttrToConv3d(const aclIntArray* &stride, const aclIntArr
78aclnnStatus ChangeConv2dInputToConv3d(const aclTensor* &input, const aclTensor* &weight, aclOpExecutor* executor);78aclnnStatus ChangeConv2dInputToConv3d(const aclTensor* &input, const aclTensor* &weight, aclOpExecutor* executor);
79const aclTensor* View4dAs5dForInput(const aclTensor* input, aclOpExecutor* executor);79const aclTensor* View4dAs5dForInput(const aclTensor* input, aclOpExecutor* executor);
80const aclTensor* View5dAs4dForOutput(const aclTensor* input, aclOpExecutor* executor);80const aclTensor* View5dAs4dForOutput(const aclTensor* input, aclOpExecutor* executor);
81-aclIntArray* View2dAs3dForAttr(const aclIntArray* intArray, int64_t expendValue, aclOpExecutor* executor, bool isPad);81+aclIntArray* View2dAs3dForAttr(const aclIntArray* intArray, int64_t expandValue, aclOpExecutor* executor, bool isPad);
82aclIntArray* View2DSwapHWForAttr(const aclIntArray* intArray, aclOpExecutor* executor);82aclIntArray* View2DSwapHWForAttr(const aclIntArray* intArray, aclOpExecutor* executor);
83const aclTensor* View4DSwapHWForTensor(const aclTensor* input, aclOpExecutor* executor);83const aclTensor* View4DSwapHWForTensor(const aclTensor* input, aclOpExecutor* executor);
84bool CheckDisContinuousStride(const aclTensor* input, const std::vector<int64_t>& newStrides, uint32_t dims);84bool CheckDisContinuousStride(const aclTensor* input, const std::vector<int64_t>& newStrides, uint32_t dims);
Mconv/convolution_forward/op_host/op_api/quant_convolution.cpp+4-4
@@ -92,10 +92,10 @@ static aclnnStatus ExtendConv2dL0Inner(const aclTensor *input, const aclTensor *
92 aclIntArray *dilation4;92 aclIntArray *dilation4;
93 aclIntArray *pad4;93 aclIntArray *pad4;
94 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1]};94 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1]};
95- FVector<int64_t> newDilaition{1, 1, (*dilation)[0], (*dilation)[1]};95+ FVector<int64_t> newDilation{1, 1, (*dilation)[0], (*dilation)[1]};
96 96 
97 stride4 = executor->AllocIntArray(newStrides.data(), QUANT_CONV2D_DIM);97 stride4 = executor->AllocIntArray(newStrides.data(), QUANT_CONV2D_DIM);
98- dilation4 = executor->AllocIntArray(newDilaition.data(), QUANT_CONV2D_DIM);98+ dilation4 = executor->AllocIntArray(newDilation.data(), QUANT_CONV2D_DIM);
99 pad4 = ConstructQuantConvNewPad(padding, executor);99 pad4 = ConstructQuantConvNewPad(padding, executor);
100 if (pad4->Size() != QUANT_CONV2D_PAD_DIM_4) {100 if (pad4->Size() != QUANT_CONV2D_PAD_DIM_4) {
101 OP_LOGE(ACLNN_ERR_INNER, "L0 func construct quant conv2d new pad failed.");101 OP_LOGE(ACLNN_ERR_INNER, "L0 func construct quant conv2d new pad failed.");
@@ -147,9 +147,9 @@ static aclnnStatus QuantConv3dL0Inner(const aclTensor *input, const aclTensor *w
147 aclIntArray *dilation5;147 aclIntArray *dilation5;
148 aclIntArray *pad6;148 aclIntArray *pad6;
149 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1], (*stride)[2]};149 FVector<int64_t> newStrides{1, 1, (*stride)[0], (*stride)[1], (*stride)[2]};
150- FVector<int64_t> newDilaition{1, 1, (*dilation)[0], (*dilation)[1], (*dilation)[2]};150+ FVector<int64_t> newDilation{1, 1, (*dilation)[0], (*dilation)[1], (*dilation)[2]};
151 stride5 = executor->AllocIntArray(newStrides.data(), QUANT_CONV3D_DIM);151 stride5 = executor->AllocIntArray(newStrides.data(), QUANT_CONV3D_DIM);
152- dilation5 = executor->AllocIntArray(newDilaition.data(), QUANT_CONV3D_DIM);152+ dilation5 = executor->AllocIntArray(newDilation.data(), QUANT_CONV3D_DIM);
153 pad6 = ConstructQuantConvNewPad(padding, executor);153 pad6 = ConstructQuantConvNewPad(padding, executor);
154 if (pad6->Size() != QUANT_CONV3D_PAD_DIM_6) {154 if (pad6->Size() != QUANT_CONV3D_PAD_DIM_6) {
155 OP_LOGE(ACLNN_ERR_INNER, "L0 func construct quant conv3d new pad failed.");155 OP_LOGE(ACLNN_ERR_INNER, "L0 func construct quant conv3d new pad failed.");