Star1033
1935
代码
代码
Issues158
Pull Requests308
Actions
项目讨论
Wiki
分析
Star1033
1935
已合并
refactor: 简化 convolution_backward 空指针防御性校验并清理冗余空行 #9309
Nice try创建于 8月27日
refactor: 简化 convolution_backward 空指针防御性校验并清理冗余空行 #9309
已合并
Nice try创建于 8月27日
从Nice_try/ops-nn-cleancode0827-cc合入到cann/ops-nnmaster
共 2 个文件变更+27-36
Mconv/conv3d_backprop_filter_v2/op_host/op_tiling/arch22/conv3d_backprop_filter_v2_basic_block_tiling.cpp+0-1
@@ -339,7 +339,6 @@ void Conv3DDWV2BasicBlockTiling::AdjustSingleNForStreamK()
339 uint64_t maxStreamKDim = mmInfo_.kValue / blockTiling_.blockBaseK;339 uint64_t maxStreamKDim = mmInfo_.kValue / blockTiling_.blockBaseK;
340 uint64_t batchDoutDim = !context_->GetDeterministic() ? 1 : static_cast<uint64_t>(runInfo_.batch) * runInfo_.dout;340 uint64_t batchDoutDim = !context_->GetDeterministic() ? 1 : static_cast<uint64_t>(runInfo_.batch) * runInfo_.dout;
341 maxStreamKDim = std::max(maxStreamKDim, batchDoutDim);341 maxStreamKDim = std::max(maxStreamKDim, batchDoutDim);
342- 
343 if (maxStreamKDim <= STREAM_K_DIM_MIN || tailCnt + STREAM_K_TAIL_TOLERANCE >= targetCoreNum) {342 if (maxStreamKDim <= STREAM_K_DIM_MIN || tailCnt + STREAM_K_TAIL_TOLERANCE >= targetCoreNum) {
344 return;343 return;
345 }344 }
Mconv/convolution_backward/op_api/aclnn_convolution_backward.cpp+27-35
@@ -3438,23 +3438,19 @@ aclnnStatus aclnnConvolutionBackwardGetWorkspaceSize(
3438 Ops::NN::Conv::ConvolutionBackwardChecker convolutionBackwardChecker(inputTensor, outputTensor, params, npuArch);3438 Ops::NN::Conv::ConvolutionBackwardChecker convolutionBackwardChecker(inputTensor, outputTensor, params, npuArch);
3439 auto ret = convolutionBackwardChecker.CheckParams();3439 auto ret = convolutionBackwardChecker.CheckParams();
3440 CHECK_RET(ret == ACLNN_SUCCESS, ret);3440 CHECK_RET(ret == ACLNN_SUCCESS, ret);
3441+ CHECK_RET(input != nullptr && weight != nullptr && gradOutput != nullptr, ACLNN_ERR_PARAM_NULLPTR);
3441 3442 
3442- if (gradOutput != nullptr) {3443+ auto gradOutputContiguous = l0op::Contiguous(gradOutput, uniqueExecutor.get());
3443- auto gradOutputContiguous = l0op::Contiguous(gradOutput, uniqueExecutor.get());3444+ CHECK_RET(gradOutputContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3444- CHECK_RET(gradOutputContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);3445+ inputTensor.gradOutput = gradOutputContiguous;
3445- inputTensor.gradOutput = gradOutputContiguous;
3446- }
3447 3446 
3448- if (input != nullptr) {3447+ auto inputContiguous = l0op::Contiguous(input, uniqueExecutor.get());
3449- auto inputContiguous = l0op::Contiguous(input, uniqueExecutor.get());3448+ CHECK_RET(inputContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3450- CHECK_RET(inputContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);3449+ inputTensor.input = inputContiguous;
3451- inputTensor.input = inputContiguous;3450+ 
3452- }3451+ auto weightContiguous = l0op::Contiguous(weight, uniqueExecutor.get());
3453- if (weight != nullptr) {3452+ CHECK_RET(weightContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3454- auto weightContiguous = l0op::Contiguous(weight, uniqueExecutor.get());3453+ inputTensor.weight = weightContiguous;
3455- CHECK_RET(weightContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3456- inputTensor.weight = weightContiguous;
3457- }
3458 3454 
3459 // 检查conv3ddw确定性计算3455 // 检查conv3ddw确定性计算
3460 if ((*outputMask)[1] && (input->GetViewShape().GetDimNum() == CONV3DINPUTDIM ||3456 if ((*outputMask)[1] && (input->GetViewShape().GetDimNum() == CONV3DINPUTDIM ||
@@ -3622,27 +3618,23 @@ aclnnStatus aclnnConvTbcBackwardGetWorkspaceSize(const aclTensor* self, const ac
3622 Ops::NN::Conv::ConvTbcBackwardChecker convTbcBackwardChecker(inputTensor, outputTensor, tbcparams, npuArch);3618 Ops::NN::Conv::ConvTbcBackwardChecker convTbcBackwardChecker(inputTensor, outputTensor, tbcparams, npuArch);
3623 auto ret = convTbcBackwardChecker.CheckTbcParams();3619 auto ret = convTbcBackwardChecker.CheckTbcParams();
3624 CHECK_RET(ret == ACLNN_SUCCESS, ret);3620 CHECK_RET(ret == ACLNN_SUCCESS, ret);
3621+ CHECK_RET(self != nullptr && input != nullptr && weight != nullptr && bias != nullptr, ACLNN_ERR_PARAM_NULLPTR);
3625 3622 
3626- if (self != nullptr) {3623+ auto selfContiguous = l0op::Contiguous(self, uniqueExecutor.get());
3627- auto selfContiguous = l0op::Contiguous(self, uniqueExecutor.get());3624+ CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3628- CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);3625+ inputTensor.self = selfContiguous;
3629- inputTensor.self = selfContiguous;3626+ 
3630- }3627+ auto weightContiguous = l0op::Contiguous(weight, uniqueExecutor.get());
3631- if (weight != nullptr) {3628+ CHECK_RET(weightContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3632- auto weightContiguous = l0op::Contiguous(weight, uniqueExecutor.get());3629+ inputTensor.weight = weightContiguous;
3633- CHECK_RET(weightContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);3630+ 
3634- inputTensor.weight = weightContiguous;3631+ auto inputContiguous = l0op::Contiguous(input, uniqueExecutor.get());
3635- }3632+ CHECK_RET(inputContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3636- if (input != nullptr) {3633+ inputTensor.input = inputContiguous;
3637- auto inputContiguous = l0op::Contiguous(input, uniqueExecutor.get());3634+ 
3638- CHECK_RET(inputContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);3635+ auto biasContiguous = l0op::Contiguous(bias, uniqueExecutor.get());
3639- inputTensor.input = inputContiguous;3636+ CHECK_RET(biasContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3640- }3637+ inputTensor.bias = biasContiguous;
3641- if (bias != nullptr) {
3642- auto biasContiguous = l0op::Contiguous(bias, uniqueExecutor.get());
3643- CHECK_RET(biasContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
3644- inputTensor.bias = biasContiguous;
3645- }
3646 3638 
3647 // 设置param3639 // 设置param
3648 FVector<int64_t> newStride = {1};3640 FVector<int64_t> newStride = {1};