Star181
850
代码
代码
Issues157
Pull Requests329
流水线
项目讨论
Wiki
分析
Star181
850
已合并
foreach_addcdiv/mul.scalarlist, foreach_addcdiv/mul.Tensor代码回退 #1868
AtomGit-Bot创建于 2024年9月24日
foreach_addcdiv/mul.scalarlist, foreach_addcdiv/mul.Tensor代码回退 #1868
已合并
AtomGit-Bot创建于 2024年9月24日
从refs/pull/1868/head合入到master
共 4 个文件变更+36-153
Mop_plugin/ops/opapi/ForeachAddcdivScalarListKernelNpuOpApi.cpp+2-69
@@ -19,68 +19,13 @@
19namespace op_api {19namespace op_api {
20using npu_preparation = at_npu::native::OpPreparation;20using npu_preparation = at_npu::native::OpPreparation;
21 21 
22-void _split_and_exec_npu_cmd_addcdiv(const at::TensorList input,
23- const at::TensorList tensors1,
24- const at::TensorList tensors2,
25- at::ArrayRef<at::Scalar> scalars,
26- at::TensorList result,
27- bool is_inplace)
28-{
29- size_t tensor_count = input.size();
30- size_t max_tensor_count = is_inplace ? 16 : 12;
31- size_t loop_time = tensor_count / max_tensor_count;
32- size_t remaining_count = tensor_count % max_tensor_count;
33- size_t data_count = max_tensor_count;
34- if (remaining_count > 0) {
35- loop_time++;
36- }
37- 
38- if (tensor_count <= max_tensor_count) {
39- EXEC_NPU_CMD(aclnnForeachAddcdivScalarList, input, tensors1, tensors2, scalars, result);
40- return;
41- }
42- for (size_t i = 0; i < loop_time; i++) {
43- if (i == loop_time - 1 && remaining_count > 0)
44- data_count = remaining_count;
45- at::TensorList temp_input(input.data() + i * max_tensor_count, data_count);
46- at::TensorList temp_tensors1(tensors1.data() + i * max_tensor_count, data_count);
47- at::TensorList temp_tensors2(tensors2.data() + i * max_tensor_count, data_count);
48- at::ArrayRef<at::Scalar> temp_scalars(scalars.data() + i * max_tensor_count, data_count);
49- at::TensorList temp_result(result.data() + i * max_tensor_count, data_count);
50- EXEC_NPU_CMD(aclnnForeachAddcdivScalarList, temp_input, temp_tensors1, temp_tensors2, temp_scalars, temp_result);
51- }
52-}
53- 
54std::vector<at::Tensor> _foreach_addcdiv(const at::TensorList input,22std::vector<at::Tensor> _foreach_addcdiv(const at::TensorList input,
55 const at::TensorList tensors1,23 const at::TensorList tensors1,
56 const at::TensorList tensors2,24 const at::TensorList tensors2,
57 at::ArrayRef<at::Scalar> scalars)25 at::ArrayRef<at::Scalar> scalars)
58{26{
59- static const bool is_support_nd_out = (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1 &&
60- c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend310B1) ||
61- (c10_npu::GetSocVersion() > c10_npu::SocVersion::Ascend310B4);
62- if (!is_support_nd_out) {
63- return at::native::foreach_tensor_addcdiv_scalarlist_slow(input, tensors1, tensors2, scalars);
64- }
65- 
66 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);27 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);
67- if (!at::native::can_use_fast_route({input, tensors1, tensors2}) ||28+ return at::native::foreach_tensor_addcdiv_scalarlist_slow(input, tensors1, tensors2, scalars);
68- at::native::has_integral_tensor(input, true)) {
69- return at::native::foreach_tensor_addcdiv_scalarlist_slow(input, tensors1, tensors2, scalars);
70- }
71- 
72- auto scalar_type = input[0].scalar_type();
73- 
74- std::vector<at::Tensor> result;
75- result.reserve(input.size());
76- for (const at::Tensor &tensor : input) {
77- auto output_size = op_infer::input_same_output_size(tensor);
78- result.push_back(npu_preparation::apply_tensor_without_format(output_size, tensor.options().dtype(scalar_type)));
79- }
80- at::TensorList result_ = at::TensorList(result);
81- _split_and_exec_npu_cmd_addcdiv(input, tensors1, tensors2, scalars, result_, false);
82- 
83- return result;
84}29}
85 30 
86void _foreach_addcdiv_(const at::TensorList input,31void _foreach_addcdiv_(const at::TensorList input,
@@ -88,19 +33,7 @@ void _foreach_addcdiv_(const at::TensorList input,
88 const at::TensorList tensors2,33 const at::TensorList tensors2,
89 at::ArrayRef<at::Scalar> scalars)34 at::ArrayRef<at::Scalar> scalars)
90{35{
91- static const bool is_support_nd_out = (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1 &&
92- c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend310B1) ||
93- (c10_npu::GetSocVersion() > c10_npu::SocVersion::Ascend310B4);
94- if (!is_support_nd_out) {
95- return at::native::foreach_tensor_addcdiv_scalarlist_slow_(input, tensors1, tensors2, scalars);
96- }
97- 
98 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);36 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);
99- if (!at::native::can_use_fast_route({input, tensors1, tensors2}) ||37+ return at::native::foreach_tensor_addcdiv_scalarlist_slow_(input, tensors1, tensors2, scalars);
100- at::native::has_integral_tensor(input, true)) {
101- return at::native::foreach_tensor_addcdiv_scalarlist_slow_(input, tensors1, tensors2, scalars);
102- }
103- 
104- _split_and_exec_npu_cmd_addcdiv(input, tensors1, tensors2, scalars, input, true);
105}38}
106}39}
Mop_plugin/ops/opapi/ForeachAddcdivTensorKernelNpuOpApi.cpp+10-6
@@ -25,7 +25,7 @@ using npu_preparation = at_npu::native::OpPreparation;
25void _split_and_exec_npu_cmd_addcdiv_tensor(const at::TensorList input,25void _split_and_exec_npu_cmd_addcdiv_tensor(const at::TensorList input,
26 const at::TensorList tensors1,26 const at::TensorList tensors1,
27 const at::TensorList tensors2,27 const at::TensorList tensors2,
28- at::ArrayRef<at::Scalar> scalars,28+ const at::Tensor scalars,
29 at::TensorList result,29 at::TensorList result,
30 bool is_inplace)30 bool is_inplace)
31{31{
@@ -39,7 +39,8 @@ void _split_and_exec_npu_cmd_addcdiv_tensor(const at::TensorList input,
39 }39 }
40 40 
41 if (tensor_count <= max_tensor_count) {41 if (tensor_count <= max_tensor_count) {
42- EXEC_NPU_CMD(aclnnForeachAddcdivScalarList, input, tensors1, tensors2, scalars, result);42+ auto scalar_tensor = npu_preparation::copy_tensor_host_to_device(scalars);
43+ EXEC_NPU_CMD(aclnnForeachAddcdivScalarList, input, tensors1, tensors2, scalar_tensor, result);
43 return;44 return;
44 }45 }
45 for (size_t i = 0; i < loop_time; i++) {46 for (size_t i = 0; i < loop_time; i++) {
@@ -47,9 +48,11 @@ void _split_and_exec_npu_cmd_addcdiv_tensor(const at::TensorList input,
47 at::TensorList temp_input(input.data() + i * max_tensor_count, data_count);48 at::TensorList temp_input(input.data() + i * max_tensor_count, data_count);
48 at::TensorList temp_tensors1(tensors1.data() + i * max_tensor_count, data_count);49 at::TensorList temp_tensors1(tensors1.data() + i * max_tensor_count, data_count);
49 at::TensorList temp_tensors2(tensors2.data() + i * max_tensor_count, data_count);50 at::TensorList temp_tensors2(tensors2.data() + i * max_tensor_count, data_count);
50- at::ArrayRef<at::Scalar> temp_scalars(scalars.data() + i * max_tensor_count, data_count);51+ at::Tensor temp_scalars = scalars.slice(0, i * max_tensor_count, data_count);
51 at::TensorList temp_result(result.data() + i * max_tensor_count, data_count);52 at::TensorList temp_result(result.data() + i * max_tensor_count, data_count);
52- EXEC_NPU_CMD(aclnnForeachAddcdivScalarList, temp_input, temp_tensors1, temp_tensors2, temp_scalars, temp_result);53+ 
54+ auto scalar_tensor = npu_preparation::copy_tensor_host_to_device(temp_scalars);
55+ EXEC_NPU_CMD(aclnnForeachAddcdivScalarList, temp_input, temp_tensors1, temp_tensors2, scalar_tensor, temp_result);
53 }56 }
54}57}
55 58 
@@ -73,6 +76,7 @@ std::vector<at::Tensor> _foreach_addcdiv(const at::TensorList input,
73 at::native::has_integral_tensor(input, true)) {76 at::native::has_integral_tensor(input, true)) {
74 return at::native::foreach_tensor_addcdiv_scalarlist_slow(input, tensors1, tensors2, scalars_);77 return at::native::foreach_tensor_addcdiv_scalarlist_slow(input, tensors1, tensors2, scalars_);
75 }78 }
79+
76 auto scalar_type = input[0].scalar_type();80 auto scalar_type = input[0].scalar_type();
77 std::vector<at::Tensor> result(input.size());81 std::vector<at::Tensor> result(input.size());
78 auto iterRes = result.data();82 auto iterRes = result.data();
@@ -82,7 +86,7 @@ std::vector<at::Tensor> _foreach_addcdiv(const at::TensorList input,
82 iterRes[i++] = at_npu::native::OpPreparation::apply_tensor_without_format(output_size, tensor.options().dtype(scalar_type));86 iterRes[i++] = at_npu::native::OpPreparation::apply_tensor_without_format(output_size, tensor.options().dtype(scalar_type));
83 }87 }
84 at::TensorList result_ = at::TensorList(result);88 at::TensorList result_ = at::TensorList(result);
85- _split_and_exec_npu_cmd_addcdiv_tensor(input, tensors1, tensors2, scalars_, result_, false);89+ _split_and_exec_npu_cmd_addcdiv_tensor(input, tensors1, tensors2, scalars, result_, false);
86 return result;90 return result;
87}91}
88 92 
@@ -108,7 +112,7 @@ void _foreach_addcdiv_(const at::TensorList input,
108 }112 }
109 113 
110 at::native::check_foreach_api_restrictions(input, tensors1, tensors2);114 at::native::check_foreach_api_restrictions(input, tensors1, tensors2);
111- _split_and_exec_npu_cmd_addcdiv_tensor(input, tensors1, tensors2, scalars_, input, true);115+ _split_and_exec_npu_cmd_addcdiv_tensor(input, tensors1, tensors2, scalars, input, true);
112}116}
113#endif117#endif
114}118}
Mop_plugin/ops/opapi/ForeachAddcmulScalarListKernelNpuOpApi.cpp+2-67
@@ -19,67 +19,13 @@
19namespace op_api {19namespace op_api {
20using npu_preparation = at_npu::native::OpPreparation;20using npu_preparation = at_npu::native::OpPreparation;
21 21 
22-void _split_and_exec_npu_cmd_addcmul(const at::TensorList input,
23- const at::TensorList tensors1,
24- const at::TensorList tensors2,
25- at::ArrayRef<at::Scalar> scalars,
26- at::TensorList result,
27- bool is_inplace)
28-{
29- size_t tensor_count = input.size();
30- size_t max_tensor_count = is_inplace ? 16 : 12;
31- size_t loop_time = tensor_count / max_tensor_count;
32- size_t remaining_count = tensor_count % max_tensor_count;
33- size_t data_count = max_tensor_count;
34- if (remaining_count > 0)
35- loop_time++;
36- 
37- if (tensor_count <= max_tensor_count) {
38- EXEC_NPU_CMD(aclnnForeachAddcmulScalarList, input, tensors1, tensors2, scalars, result);
39- return;
40- }
41- for (size_t i = 0; i < loop_time; i++) {
42- if (i == loop_time - 1 && remaining_count > 0)
43- data_count = remaining_count;
44- at::TensorList temp_input(input.data() + i * max_tensor_count, data_count);
45- at::TensorList temp_tensors1(tensors1.data() + i * max_tensor_count, data_count);
46- at::TensorList temp_tensors2(tensors2.data() + i * max_tensor_count, data_count);
47- at::ArrayRef<at::Scalar> temp_scalars(scalars.data() + i * max_tensor_count, data_count);
48- at::TensorList temp_result(result.data() + i * max_tensor_count, data_count);
49- EXEC_NPU_CMD(aclnnForeachAddcmulScalarList, temp_input, temp_tensors1, temp_tensors2, temp_scalars, temp_result);
50- }
51-}
52- 
53std::vector<at::Tensor> _foreach_addcmul(const at::TensorList input,22std::vector<at::Tensor> _foreach_addcmul(const at::TensorList input,
54 const at::TensorList tensors1,23 const at::TensorList tensors1,
55 const at::TensorList tensors2,24 const at::TensorList tensors2,
56 at::ArrayRef<at::Scalar> scalars)25 at::ArrayRef<at::Scalar> scalars)
57{26{
58- static const bool is_support_nd_out = (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1 &&
59- c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend310B1) ||
60- (c10_npu::GetSocVersion() > c10_npu::SocVersion::Ascend310B4);
61- if (!is_support_nd_out) {
62- return at::native::foreach_tensor_addcmul_scalarlist_slow(input, tensors1, tensors2, scalars);
63- }
64- 
65 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);27 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);
66- if (!at::native::can_use_fast_route({input, tensors1, tensors2}) ||28+ return at::native::foreach_tensor_addcmul_scalarlist_slow(input, tensors1, tensors2, scalars);
67- at::native::has_integral_tensor(input, true)) {
68- return at::native::foreach_tensor_addcmul_scalarlist_slow(input, tensors1, tensors2, scalars);
69- }
70- 
71- auto scalar_type = input[0].scalar_type();
72- 
73- std::vector<at::Tensor> result;
74- result.reserve(input.size());
75- for (const at::Tensor &tensor : input) {
76- auto output_size = op_infer::input_same_output_size(tensor);
77- result.push_back(npu_preparation::apply_tensor_without_format(output_size, tensor.options().dtype(scalar_type)));
78- }
79- at::TensorList result_ = at::TensorList(result);
80- _split_and_exec_npu_cmd_addcmul(input, tensors1, tensors2, scalars, result_, false);
81- 
82- return result;
83}29}
84 30 
85void _foreach_addcmul_(const at::TensorList input,31void _foreach_addcmul_(const at::TensorList input,
@@ -87,18 +33,7 @@ void _foreach_addcmul_(const at::TensorList input,
87 const at::TensorList tensors2,33 const at::TensorList tensors2,
88 at::ArrayRef<at::Scalar> scalars)34 at::ArrayRef<at::Scalar> scalars)
89{35{
90- static const bool is_support_nd_out = (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1 &&
91- c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend310B1) ||
92- (c10_npu::GetSocVersion() > c10_npu::SocVersion::Ascend310B4);
93- if (!is_support_nd_out) {
94- return at::native::foreach_tensor_addcmul_scalarlist_slow_(input, tensors1, tensors2, scalars);
95- }
96-
97 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);36 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars);
98- if (!at::native::can_use_fast_route({input, tensors1, tensors2}) ||37+ return at::native::foreach_tensor_addcmul_scalarlist_slow_(input, tensors1, tensors2, scalars);
99- at::native::has_integral_tensor(input, true)) {
100- return at::native::foreach_tensor_addcmul_scalarlist_slow_(input, tensors1, tensors2, scalars);
101- }
102- _split_and_exec_npu_cmd_addcmul(input, tensors1, tensors2, scalars, input, true);
103}38}
104}39}
Mop_plugin/ops/opapi/ForeachAddcmulTensorKernelNpuOpApi.cpp+22-11
@@ -25,7 +25,7 @@ using npu_preparation = at_npu::native::OpPreparation;
25void _split_and_exec_npu_cmd_addcmul_tensor(const at::TensorList input,25void _split_and_exec_npu_cmd_addcmul_tensor(const at::TensorList input,
26 const at::TensorList tensors1,26 const at::TensorList tensors1,
27 const at::TensorList tensors2,27 const at::TensorList tensors2,
28- at::ArrayRef<at::Scalar> scalars,28+ at::Tensor scalars,
29 at::TensorList result,29 at::TensorList result,
30 bool is_inplace)30 bool is_inplace)
31{31{
@@ -49,7 +49,7 @@ void _split_and_exec_npu_cmd_addcmul_tensor(const at::TensorList input,
49 at::TensorList temp_input(input.data() + i * max_tensor_count, data_count);49 at::TensorList temp_input(input.data() + i * max_tensor_count, data_count);
50 at::TensorList temp_tensors1(tensors1.data() + i * max_tensor_count, data_count);50 at::TensorList temp_tensors1(tensors1.data() + i * max_tensor_count, data_count);
51 at::TensorList temp_tensors2(tensors2.data() + i * max_tensor_count, data_count);51 at::TensorList temp_tensors2(tensors2.data() + i * max_tensor_count, data_count);
52- at::ArrayRef<at::Scalar> temp_scalars(scalars.data() + i * max_tensor_count, data_count);52+ at::Tensor temp_scalars = scalars.slice(0, i * max_tensor_count, data_count);
53 at::TensorList temp_result(result.data() + i * max_tensor_count, data_count);53 at::TensorList temp_result(result.data() + i * max_tensor_count, data_count);
54 EXEC_NPU_CMD(aclnnForeachAddcmulScalarList, temp_input, temp_tensors1, temp_tensors2, temp_scalars, temp_result);54 EXEC_NPU_CMD(aclnnForeachAddcmulScalarList, temp_input, temp_tensors1, temp_tensors2, temp_scalars, temp_result);
55 }55 }
@@ -70,20 +70,25 @@ std::vector<at::Tensor> _foreach_addcmul(const at::TensorList input,
70 }70 }
71 71 
72 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars_);72 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars_);
73- if (!at::native::can_use_fast_route({input, tensors1, tensors2}) ||73+ if (!at_npu::native::env::CheckJitDisable() ||
74+ !at::native::can_use_fast_route({input, tensors1, tensors2}) ||
74 at::native::has_integral_tensor(input, true)) {75 at::native::has_integral_tensor(input, true)) {
75 return at::native::foreach_tensor_addcmul_scalarlist_slow(input, tensors1, tensors2, scalars_);76 return at::native::foreach_tensor_addcmul_scalarlist_slow(input, tensors1, tensors2, scalars_);
76 }77 }
78+ 
77 auto scalar_type = input[0].scalar_type();79 auto scalar_type = input[0].scalar_type();
78- std::vector<at::Tensor> result(input.size());80+ if (scalar_type != at::ScalarType::Half && scalar_type != at::ScalarType::Float && scalar_type != at::ScalarType::Int) {
79- auto iterRes = result.data();81+ TORCH_CHECK(false, "input must be half, float or int32" + OPS_ERROR(ErrCode::TYPE));
80- int i = 0;82+ }
83+ std::vector<at::Tensor> result;
84+ result.reserve(input.size());
81 for (const at::Tensor &tensor : input) {85 for (const at::Tensor &tensor : input) {
82 auto output_size = op_infer::input_same_output_size(tensor);86 auto output_size = op_infer::input_same_output_size(tensor);
83- iterRes[i++] = at_npu::native::OpPreparation::apply_tensor_without_format(output_size, tensor.options().dtype(scalar_type));87+ result.push_back(npu_preparation::apply_tensor_without_format(output_size, tensor.options().dtype(scalar_type)));
84 }88 }
85 at::TensorList result_ = at::TensorList(result);89 at::TensorList result_ = at::TensorList(result);
86- _split_and_exec_npu_cmd_addcmul_tensor(input, tensors1, tensors2, scalars_, result_, false);90+ auto scalar_tensor = npu_preparation::copy_tensor_host_to_device(scalars);
91+ _split_and_exec_npu_cmd_addcmul_tensor(input, tensors1, tensors2, scalar_tensor, result_, false);
87 return result;92 return result;
88}93}
89 94 
@@ -100,16 +105,22 @@ void _foreach_addcmul_(const at::TensorList input,
100 if (!is_support_nd_out) {105 if (!is_support_nd_out) {
101 return at::native::foreach_tensor_addcmul_scalarlist_slow_(input, tensors1, tensors2, scalars_);106 return at::native::foreach_tensor_addcmul_scalarlist_slow_(input, tensors1, tensors2, scalars_);
102 }107 }
103- 108+ 
104 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars_);109 at::native::check_foreach_api_restrictions(input, tensors1, tensors2, scalars_);
105- if (!at::native::can_use_fast_route({input, tensors1, tensors2}) ||110+ if (!at_npu::native::env::CheckJitDisable() ||
111+ !at::native::can_use_fast_route({input, tensors1, tensors2}) ||
106 at::native::has_integral_tensor(input, true)) {112 at::native::has_integral_tensor(input, true)) {
107 return at::native::foreach_tensor_addcmul_scalarlist_slow_(input, tensors1, tensors2, scalars_);113 return at::native::foreach_tensor_addcmul_scalarlist_slow_(input, tensors1, tensors2, scalars_);
108 }114 }
109 115 
110 at::native::check_foreach_api_restrictions(input, tensors1, tensors2);116 at::native::check_foreach_api_restrictions(input, tensors1, tensors2);
117+ auto scalar_type = input[0].scalar_type();
118+ if (scalar_type != at::ScalarType::Half && scalar_type != at::ScalarType::Float && scalar_type != at::ScalarType::Int) {
119+ TORCH_CHECK(false, "input must be half, float or int32" + OPS_ERROR(ErrCode::TYPE));
120+ }
121+ auto scalar_tensor = npu_preparation::copy_tensor_host_to_device(scalars);
111 122 
112- _split_and_exec_npu_cmd_addcmul_tensor(input, tensors1, tensors2, scalars_, input, true);123+ _split_and_exec_npu_cmd_addcmul_tensor(input, tensors1, tensors2, scalar_tensor, input, true);
113}124}
114#endif125#endif
115} // namespace op_api126} // namespace op_api