bugfix: eliminate redundant operations in sort.-op-plugin-AtomGit

已合并

bugfix: eliminate redundant operations in sort. #5130

yvjc创建于 6月8日从已删除 :shangfa 合入到 Ascend/op-plugin： 26.1.0

bugfix: eliminate redundant operations in sort. #5130

已合并

yvjc创建于 6月8日

从已删除 :shangfa合入到Ascend/op-plugin 26.1.0

共 1 个文件变更+0-15

Mop_plugin/ops/opapi/SortKernelNpuOpApi.cpp+0-15

	@@ -27,20 +27,11 @@ std::tuple<at::Tensor &, at::Tensor &> sort_output(const at::Tensor &self, bool
27	return std::tie(values, indices);	27	return std::tie(values, indices);
28	}	28	}
29		29
30	-std::vector<int64_t> compute_sort_strides(const at::Tensor& self) {
31	- return self.is_non_overlapping_and_dense()
32	- ? self.strides().vec()
33	- : at::infer_dense_strides(self.sizes(), self.strides());
34	-}
35	-
36	std::tuple<at::Tensor, at::Tensor> sort(const at::Tensor &self, int64_t dim, bool descending)	30	std::tuple<at::Tensor, at::Tensor> sort(const at::Tensor &self, int64_t dim, bool descending)
37	{	31	{
38	DO_COMPATIBILITY(aclnnSort, acl_op::sort(self, dim, descending));	32	DO_COMPATIBILITY(aclnnSort, acl_op::sort(self, dim, descending));
39	- auto strides = compute_sort_strides(self);
40	at::Tensor values = npu_preparation::apply_tensor_without_format(self);	33	at::Tensor values = npu_preparation::apply_tensor_without_format(self);
41	at::Tensor indices = npu_preparation::apply_tensor_without_format(self.sizes(), self.options().dtype(at::kLong));	34	at::Tensor indices = npu_preparation::apply_tensor_without_format(self.sizes(), self.options().dtype(at::kLong));
42	- values = values.as_strided_(self.sizes(), strides);
43	- indices = indices.as_strided_(self.sizes(), strides);
44	bool stable = false;	35	bool stable = false;
45		36
46	return sort_output(self, stable, dim, descending, values, indices);	37	return sort_output(self, stable, dim, descending, values, indices);
	@@ -49,11 +40,8 @@ std::tuple<at::Tensor, at::Tensor> sort(const at::Tensor &self, int64_t dim, boo
49	std::tuple<at::Tensor, at::Tensor> sort(const at::Tensor &self, at::Dimname dim, bool descending)	40	std::tuple<at::Tensor, at::Tensor> sort(const at::Tensor &self, at::Dimname dim, bool descending)
50	{	41	{
51	DO_COMPATIBILITY(aclnnSort, acl_op::sort(self, dim, descending));	42	DO_COMPATIBILITY(aclnnSort, acl_op::sort(self, dim, descending));
52	- auto strides = compute_sort_strides(self);
53	at::Tensor values = npu_preparation::apply_tensor_without_format(self);	43	at::Tensor values = npu_preparation::apply_tensor_without_format(self);
54	at::Tensor indices = npu_preparation::apply_tensor_without_format(self.sizes(), self.options().dtype(at::kLong));	44	at::Tensor indices = npu_preparation::apply_tensor_without_format(self.sizes(), self.options().dtype(at::kLong));
55	- values = values.as_strided_(self.sizes(), strides);
56	- indices = indices.as_strided_(self.sizes(), strides);
57	bool stable = false;	45	bool stable = false;
58	int64_t argDim = dimname_to_position(self, dim);	46	int64_t argDim = dimname_to_position(self, dim);
59		47
	@@ -90,11 +78,8 @@ std::tuple<at::Tensor, at::Tensor> sort(const at::Tensor &self,
90	auto dtype = self.scalar_type();	78	auto dtype = self.scalar_type();
91	TORCH_CHECK(!(dtype == at::kDouble),	79	TORCH_CHECK(!(dtype == at::kDouble),
92	"Input data type should not be float64 " + OPS_ERROR(ErrCode::TYPE));	80	"Input data type should not be float64 " + OPS_ERROR(ErrCode::TYPE));
93	- auto strides = compute_sort_strides(self);
94	at::Tensor values = npu_preparation::apply_tensor_without_format(self);	81	at::Tensor values = npu_preparation::apply_tensor_without_format(self);
95	at::Tensor indices = npu_preparation::apply_tensor_without_format(self.sizes(), self.options().dtype(at::kLong));	82	at::Tensor indices = npu_preparation::apply_tensor_without_format(self.sizes(), self.options().dtype(at::kLong));
96	- values = values.as_strided_(self.sizes(), strides);
97	- indices = indices.as_strided_(self.sizes(), strides);
98	bool argStable = c10::value_or_else(stable, [] { return false; });	83	bool argStable = c10::value_or_else(stable, [] { return false; });
99	EXEC_NPU_CMD(aclnnSort, self, argStable, dim, descending, values, indices);	84	EXEC_NPU_CMD(aclnnSort, self, argStable, dim, descending, values, indices);
100	return std::tie(values, indices);	85	return std::tie(values, indices);