已合并
【bugfix】upsample_bicubic2d_backward，upsample_linear1d，sort_without_indices，roll，roll_6d，sin，unique2，multinomial，repeat，ctc_loss #678
AtomGit-Bot创建于 2022年4月14日
【bugfix】upsample_bicubic2d_backward，upsample_linear1d，sort_without_indices，roll，roll_6d，sin，unique2，multinomial，repeat，ctc_loss #678
已合并
从refs/pull/678/head合入到master
共 9 个文件变更+28-19
| @@ -62,13 +62,15 @@ std::tuple<at::Tensor, at::Tensor> NPUNativeFunctions::_ctc_loss( | |||
| 62 | auto outputSizes = ctc_loss_npu_output_size(logProbs, targetsCast, targetLengths, maxLength); | 62 | auto outputSizes = ctc_loss_npu_output_size(logProbs, targetsCast, targetLengths, maxLength); |
| 63 | 63 | ||
| 64 | // construct the output tensor of the NPU | 64 | // construct the output tensor of the NPU |
| 65 | - at::Tensor negLogLikelihood = OpPreparation::ApplyTensorWithSizes( | 65 | + at::Tensor negLogLikelihood = OpPreparation::ApplyTensorWithFormat( |
| 66 | std::get<0>(outputSizes), | 66 | std::get<0>(outputSizes), |
| 67 | - logProbsNeed.options()); | 67 | + logProbsNeed.options(), |
| 68 | + CalcuOpUtil::get_tensor_npu_format(logProbsNeed)); | ||
| 68 | 69 | ||
| 69 | - at::Tensor logAlpha = OpPreparation::ApplyTensorWithSizes( | 70 | + at::Tensor logAlpha = OpPreparation::ApplyTensorWithFormat( |
| 70 | std::get<1>(outputSizes), | 71 | std::get<1>(outputSizes), |
| 71 | - logProbsNeed.options()); | 72 | + logProbsNeed.options(), |
| 73 | + CalcuOpUtil::get_tensor_npu_format(logProbsNeed)); | ||
| 72 | 74 | ||
| 73 | // calculate the output result of the NPU | 75 | // calculate the output result of the NPU |
| 74 | OpCommand cmd; | 76 | OpCommand cmd; |
| @@ -70,8 +70,8 @@ at::Tensor NPUNativeFunctions::multinomial( | |||
| 70 | auto shape = array_to_small_vector(self.sizes()); | 70 | auto shape = array_to_small_vector(self.sizes()); |
| 71 | shape[dim-1] = num_samples; | 71 | shape[dim-1] = num_samples; |
| 72 | 72 | ||
| 73 | - at::Tensor result = OpPreparation::ApplyTensorWithSizes( | 73 | + at::Tensor result = OpPreparation::ApplyTensorWithFormat( |
| 74 | - shape, self.options().dtype(at::kLong)); | 74 | + shape, self.options().dtype(at::kLong), CalcuOpUtil::get_tensor_npu_format(self)); |
| 75 | multinomial_out_npu_nocheck(result, self, num_samples, replacement, gen); | 75 | multinomial_out_npu_nocheck(result, self, num_samples, replacement, gen); |
| 76 | return result; | 76 | return result; |
| 77 | } | 77 | } |
| @@ -50,7 +50,8 @@ at::Tensor NPUNativeFunctions::repeat(const at::Tensor& self, at::IntArrayRef re | |||
| 50 | auto outputSize = repeat_npu_output_size(selfCp, repeats); | 50 | auto outputSize = repeat_npu_output_size(selfCp, repeats); |
| 51 | 51 | ||
| 52 | // construct the output tensor of the NPU | 52 | // construct the output tensor of the NPU |
| 53 | - at::Tensor result = OpPreparation::ApplyTensorWithSizes(outputSize, selfCp.options()); | 53 | + at::Tensor result = OpPreparation::ApplyTensorWithFormat( |
| 54 | + outputSize, selfCp.options(), CalcuOpUtil::get_tensor_npu_format(selfCp)); | ||
| 54 | 55 | ||
| 55 | // calculate the output result of the NPU | 56 | // calculate the output result of the NPU |
| 56 | repeat_out_npu_nocheck(result, selfCp, repeats); | 57 | repeat_out_npu_nocheck(result, selfCp, repeats); |
| @@ -52,9 +52,10 @@ at::Tensor& roll_transpose( | |||
| 52 | std::swap(perm[axis], perm[firstDim]); | 52 | std::swap(perm[axis], perm[firstDim]); |
| 53 | at::Tensor transposeSelf = NPUNativeFunctions::npu_transpose(self, perm); | 53 | at::Tensor transposeSelf = NPUNativeFunctions::npu_transpose(self, perm); |
| 54 | auto outputSize = transpose_npu_output_size(result, perm); | 54 | auto outputSize = transpose_npu_output_size(result, perm); |
| 55 | - at::Tensor transposeResult = OpPreparation::ApplyTensorWithSizes( | 55 | + at::Tensor transposeResult = OpPreparation::ApplyTensorWithFormat( |
| 56 | outputSize, | 56 | outputSize, |
| 57 | - self.options()); | 57 | + self.options(), |
| 58 | + CalcuOpUtil::get_tensor_npu_format(self)); | ||
| 58 | c10::SmallVector<int64_t, SIZE> dim = {firstDim}; | 59 | c10::SmallVector<int64_t, SIZE> dim = {firstDim}; |
| 59 | c10::SmallVector<int64_t, SIZE> shift_bak = {shifts[id]}; | 60 | c10::SmallVector<int64_t, SIZE> shift_bak = {shifts[id]}; |
| 60 | at::IntArrayRef dim_now = at::IntArrayRef(dim); | 61 | at::IntArrayRef dim_now = at::IntArrayRef(dim); |
| @@ -46,7 +46,8 @@ at::Tensor& NPUNativeFunctions::sin_out(const at::Tensor& self, at::Tensor& resu | |||
| 46 | 46 | ||
| 47 | at::Tensor NPUNativeFunctions::sin(const at::Tensor& self) { | 47 | at::Tensor NPUNativeFunctions::sin(const at::Tensor& self) { |
| 48 | // construct the output tensor of the NPU | 48 | // construct the output tensor of the NPU |
| 49 | - at::Tensor result = OpPreparation::ApplyTensorWithSizes(self.sizes(), self.options()); | 49 | + at::Tensor result = OpPreparation::ApplyTensorWithFormat( |
| 50 | + self.sizes(), self.options(), CalcuOpUtil::get_tensor_npu_format(self)); | ||
| 50 | 51 | ||
| 51 | // calculate the output result of the NPU | 52 | // calculate the output result of the NPU |
| 52 | sin_out_npu_nocheck(result, self); | 53 | sin_out_npu_nocheck(result, self); |
| @@ -61,7 +61,7 @@ at::Tensor& NPUNativeFunctions::npu_sort_v2_out( | |||
| 61 | at::Tensor transposeSelf = NPUNativeFunctions::npu_transpose(self, perm); | 61 | at::Tensor transposeSelf = NPUNativeFunctions::npu_transpose(self, perm); |
| 62 | 62 | ||
| 63 | auto outputSize = transpose_npu_output_size(result, perm); | 63 | auto outputSize = transpose_npu_output_size(result, perm); |
| 64 | - at::Tensor transposeResult = OpPreparation::ApplyTensorWithSizes(outputSize, result.options()); | 64 | + at::Tensor transposeResult = OpPreparation::ApplyTensor(result, outputSize); |
| 65 | 65 | ||
| 66 | sort_without_indices_no_transpose(transposeResult, transposeSelf, lastDim, descending); | 66 | sort_without_indices_no_transpose(transposeResult, transposeSelf, lastDim, descending); |
| 67 | NPUNativeFunctions::npu_transpose_out(transposeResult, perm, result); | 67 | NPUNativeFunctions::npu_transpose_out(transposeResult, perm, result); |
| @@ -97,7 +97,7 @@ at::Tensor NPUNativeFunctions::npu_sort_v2( | |||
| 97 | at::Tensor transposeSelf = NPUNativeFunctions::npu_transpose(self, perm); | 97 | at::Tensor transposeSelf = NPUNativeFunctions::npu_transpose(self, perm); |
| 98 | 98 | ||
| 99 | auto outputSize = transpose_npu_output_size(result, perm); | 99 | auto outputSize = transpose_npu_output_size(result, perm); |
| 100 | - at::Tensor transposeResult = OpPreparation::ApplyTensorWithSizes(outputSize, result.options()); | 100 | + at::Tensor transposeResult = OpPreparation::ApplyTensor(result, outputSize); |
| 101 | 101 | ||
| 102 | sort_without_indices_no_transpose(transposeResult, transposeSelf, lastDim, descending); | 102 | sort_without_indices_no_transpose(transposeResult, transposeSelf, lastDim, descending); |
| 103 | NPUNativeFunctions::npu_transpose_out(transposeResult, perm, result); | 103 | NPUNativeFunctions::npu_transpose_out(transposeResult, perm, result); |
| @@ -115,7 +115,8 @@ at::Tensor NPUNativeFunctions::upsample_bicubic2d_backward( | |||
| 115 | c10::optional<double> scales_w) { | 115 | c10::optional<double> scales_w) { |
| 116 | // construct the output tensor of the NPU | 116 | // construct the output tensor of the NPU |
| 117 | auto outputSize = upsample_bicubic2d_backward_npu_output_size(input_size); | 117 | auto outputSize = upsample_bicubic2d_backward_npu_output_size(input_size); |
| 118 | - at::Tensor result = OpPreparation::ApplyTensorWithSizes(outputSize, grad_output.options()); | 118 | + at::Tensor result = OpPreparation::ApplyTensorWithFormat( |
| 119 | + outputSize, grad_output.options(), CalcuOpUtil::get_tensor_npu_format(grad_output)); | ||
| 119 | // calculate the output result of the NPU | 120 | // calculate the output result of the NPU |
| 120 | return upsample_bicubic2d_backward_out_nocheck(grad_output, output_size, input_size, align_corners, scales_h, scales_w, result); | 121 | return upsample_bicubic2d_backward_out_nocheck(grad_output, output_size, input_size, align_corners, scales_h, scales_w, result); |
| 121 | } | 122 | } |
| @@ -131,7 +132,8 @@ at::Tensor NPUNativeFunctions::upsample_bicubic2d_backward( | |||
| 131 | auto scales_w = CalcuOpUtil::get_scale_value(scale_factors, 1); | 132 | auto scales_w = CalcuOpUtil::get_scale_value(scale_factors, 1); |
| 132 | // construct the output tensor of the NPU | 133 | // construct the output tensor of the NPU |
| 133 | auto outputSize = upsample_bicubic2d_backward_npu_output_size(input_size); | 134 | auto outputSize = upsample_bicubic2d_backward_npu_output_size(input_size); |
| 134 | - at::Tensor result = OpPreparation::ApplyTensorWithSizes(outputSize, grad_output.options()); | 135 | + at::Tensor result = OpPreparation::ApplyTensorWithFormat( |
| 136 | + outputSize, grad_output.options(), CalcuOpUtil::get_tensor_npu_format(grad_output)); | ||
| 135 | // calculate the output result of the NPU | 137 | // calculate the output result of the NPU |
| 136 | return upsample_bicubic2d_backward_out_nocheck(grad_output, osize, input_size, align_corners, scales_h, scales_w, result); | 138 | return upsample_bicubic2d_backward_out_nocheck(grad_output, osize, input_size, align_corners, scales_h, scales_w, result); |
| 137 | } | 139 | } |
| @@ -124,7 +124,8 @@ at::Tensor NPUNativeFunctions::upsample_linear1d( | |||
| 124 | self, output_size, align_corners, scales); | 124 | self, output_size, align_corners, scales); |
| 125 | 125 | ||
| 126 | // construct the output tensor of the NPU | 126 | // construct the output tensor of the NPU |
| 127 | - at::Tensor result = OpPreparation::ApplyTensorWithSizes(outputSize, self.options()); | 127 | + at::Tensor result = OpPreparation::ApplyTensorWithFormat( |
| 128 | + outputSize, self.options(), CalcuOpUtil::get_tensor_npu_format(self)); | ||
| 128 | 129 | ||
| 129 | // calculate the output result of the NPU | 130 | // calculate the output result of the NPU |
| 130 | upsample_linear1d_out_nocheck(self, output_size, align_corners, scales, result); | 131 | upsample_linear1d_out_nocheck(self, output_size, align_corners, scales, result); |
| @@ -144,7 +145,8 @@ at::Tensor NPUNativeFunctions::upsample_linear1d( | |||
| 144 | self, osize, align_corners, scales_w); | 145 | self, osize, align_corners, scales_w); |
| 145 | 146 | ||
| 146 | // construct the output tensor of the NPU | 147 | // construct the output tensor of the NPU |
| 147 | - at::Tensor result = OpPreparation::ApplyTensorWithSizes(outputSize, self.options()); | 148 | + at::Tensor result = OpPreparation::ApplyTensorWithFormat( |
| 149 | + outputSize, self.options(), CalcuOpUtil::get_tensor_npu_format(self)); | ||
| 148 | 150 | ||
| 149 | // calculate the output result of the NPU | 151 | // calculate the output result of the NPU |
| 150 | upsample_linear1d_out_nocheck(self, osize, align_corners, scales_w, result); | 152 | upsample_linear1d_out_nocheck(self, osize, align_corners, scales_w, result); |
| @@ -64,9 +64,9 @@ tuple<at::Tensor, at::Tensor, at::Tensor> NPUNativeFunctions::_unique2( | |||
| 64 | } | 64 | } |
| 65 | 65 | ||
| 66 | at::Tensor y = OpPreparation::ApplyTensor(selfCopy, std::get<0>(outputSizes)); | 66 | at::Tensor y = OpPreparation::ApplyTensor(selfCopy, std::get<0>(outputSizes)); |
| 67 | - at::Tensor yOutputSize = OpPreparation::ApplyTensorWithSizes(std::get<1>(outputSizes), self.options().dtype(at::kLong)); | 67 | + at::Tensor yOutputSize = OpPreparation::ApplyTensorWithFormat(std::get<1>(outputSizes), self.options().dtype(at::kLong), ACL_FORMAT_ND); |
| 68 | - at::Tensor yInverse = OpPreparation::ApplyTensorWithSizes(std::get<2>(outputSizes), self.options().dtype(at::kLong)); | 68 | + at::Tensor yInverse = OpPreparation::ApplyTensorWithFormat(std::get<2>(outputSizes), self.options().dtype(at::kLong), ACL_FORMAT_ND); |
| 69 | - at::Tensor yCounts = OpPreparation::ApplyTensorWithSizes(std::get<0>(outputSizes), self.options().dtype(at::kLong)); | 69 | + at::Tensor yCounts = OpPreparation::ApplyTensorWithFormat(std::get<0>(outputSizes), self.options().dtype(at::kLong), ACL_FORMAT_ND); |
| 70 | 70 | ||
| 71 | _unique2_out_npu(y, yOutputSize, yInverse, yCounts, selfCopy, sorted, return_inverse, return_counts); | 71 | _unique2_out_npu(y, yOutputSize, yInverse, yCounts, selfCopy, sorted, return_inverse, return_counts); |
| 72 | 72 | ||