Skip to content

Commit 5240ac0

Browse files
chenwhqlShixiaowei02
chenwhql
and
Shixiaowei02
authored
Dev/op2func refactor 3 (#30)
* add a candidate dense tensor class, test=develop * remove TensorBase::backend(), test=develop * remove some ops, test=develop * cherry-pick the pr of tensor meta, test=develop * moves the dense tensor and some ops, test=develop * update the linalg operator, test=develop * update other operators, test=develop * fix errors, test=develop * fix bugs, test=develop * try to resolve the problem of windows ci, test=develop * updates codes, test=develop * fix the tensor_utils.cc, test=develop * modify the dense tensor, test=develop * fix the data type, test=develop Co-authored-by: shixiaowei02 <[email protected]>
1 parent e3ed2c6 commit 5240ac0

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

51 files changed

+632
-1106
lines changed

‎paddle/fluid/framework/CMakeLists.txt‎

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -195,10 +195,12 @@ cc_library(unused_var_check SRCS unused_var_check.cc DEPS glog no_need_buffer_va
195195

196196
IF(WITH_XPU)
197197
cc_library(operator SRCS operator.cc DEPS xpu_op_list op_info device_context tensor scope glog trainer_desc_proto data_feed_proto
198-
shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils pten pten_utils)
198+
shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils
199+
pten pten_utils kernel_factory)
199200
ELSE()
200201
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog trainer_desc_proto data_feed_proto
201-
shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils pten pten_utils)
202+
shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type op_call_stack unused_var_check nan_inf_utils
203+
pten pten_utils kernel_factory)
202204
ENDIF()
203205

204206
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
@@ -392,7 +394,7 @@ cc_library(save_load_util SRCS save_load_util.cc DEPS tensor scope layer)
392394
cc_test(save_load_util_test SRCS save_load_util_test.cc DEPS save_load_util tensor scope layer)
393395
cc_library(generator SRCS generator.cc DEPS enforce place)
394396

395-
cc_library(pten_utils SRCS pten_utils.cc DEPS lod_tensor selected_rows place pten var_type_traits)
397+
cc_library(pten_utils SRCS pten_utils.cc DEPS lod_tensor selected_rows place pten var_type_traits pten_hapi_utils)
396398

397399
# Get the current working branch
398400
execute_process(

‎paddle/fluid/framework/operator.cc‎

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -1819,10 +1819,10 @@ pten::KernelContext OperatorWithKernel::BuildPtenKernelContext(
18191819

18201820
paddle::SmallVector<std::shared_ptr<pten::TensorBase>> tmp_inputs;
18211821
for (auto var : ins_vector) {
1822-
auto pt_in = framework::InputVariableToPtenTensor(*var, in_def);
1823-
tmp_inputs.emplace_back(pt_in);
1822+
tmp_inputs.emplace_back(
1823+
experimental::MakePtenTensorBaseFromVar(*var, in_def));
18241824
}
1825-
op_kernel_ctx.EmplaceBackInputs(tmp_inputs);
1825+
op_kernel_ctx.EmplaceBackInputs(std::move(tmp_inputs));
18261826
}
18271827

18281828
for (size_t i = 0; i < output_names.size(); ++i) {
@@ -1831,10 +1831,10 @@ pten::KernelContext OperatorWithKernel::BuildPtenKernelContext(
18311831

18321832
paddle::SmallVector<std::shared_ptr<pten::TensorBase>> tmp_outputs;
18331833
for (auto var : outs_vector) {
1834-
auto pt_out = framework::OutputVariableToPtenTensor(var, out_def);
1835-
tmp_outputs.emplace_back(pt_out);
1834+
tmp_outputs.emplace_back(
1835+
experimental::MakePtenTensorBaseFromVar(var, out_def));
18361836
}
1837-
op_kernel_ctx.EmplaceBackOutputs(tmp_outputs);
1837+
op_kernel_ctx.EmplaceBackOutputs(std::move(tmp_outputs));
18381838
}
18391839

18401840
for (size_t i = 0; i < attr_names.size(); ++i) {

‎paddle/fluid/framework/pten_utils.cc‎

Lines changed: 0 additions & 142 deletions
Original file line numberDiff line numberDiff line change
@@ -24,148 +24,6 @@ limitations under the License. */
2424
namespace paddle {
2525
namespace framework {
2626

27-
// TODO(chenweihang, shixiaowei): adapt SelectedRows
28-
template <>
29-
std::shared_ptr<pten::DenseTensor> MakeTensorImpl<pten::DenseTensor, LoDTensor>(
30-
const LoDTensor& tensor, pten::Backend backend,
31-
paddle::experimental::DataType dtype,
32-
paddle::experimental::DataLayout layout) {
33-
auto holder = tensor.Holder();
34-
auto tensor_impl = std::make_shared<pten::DenseTensor>(
35-
pten::TensorMeta(tensor.dims(), backend, dtype, layout, tensor.offset()),
36-
pten::TensorStatus());
37-
38-
if (holder != nullptr) {
39-
tensor_impl->ShareAllocation(tensor.Holder());
40-
}
41-
return tensor_impl;
42-
}
43-
44-
template <>
45-
std::shared_ptr<pten::DenseTensor> MakeTensorImpl<pten::DenseTensor, Tensor>(
46-
const Tensor& tensor, pten::Backend backend,
47-
paddle::experimental::DataType dtype,
48-
paddle::experimental::DataLayout layout) {
49-
auto holder = tensor.Holder();
50-
auto tensor_impl = std::make_shared<pten::DenseTensor>(
51-
pten::TensorMeta(tensor.dims(), backend, dtype, layout, tensor.offset()),
52-
pten::TensorStatus());
53-
54-
if (holder != nullptr) {
55-
tensor_impl->ShareAllocation(tensor.Holder());
56-
}
57-
return tensor_impl;
58-
}
59-
60-
template <>
61-
std::shared_ptr<pten::DenseTensor> MakeTensorImpl<pten::DenseTensor>(
62-
const LoDTensor& tensor, const platform::Place& place,
63-
proto::VarType::Type type) {
64-
return MakeTensorImpl<pten::DenseTensor, LoDTensor>(
65-
tensor, pten::TransToPtenBackend(place), pten::TransToPtenDataType(type),
66-
pten::TransToPtenDataLayout(tensor.layout()));
67-
}
68-
69-
template <>
70-
std::shared_ptr<pten::DenseTensor> MakeTensorImpl<pten::DenseTensor>(
71-
const Tensor& tensor, const platform::Place& place,
72-
proto::VarType::Type type) {
73-
return MakeTensorImpl<pten::DenseTensor, Tensor>(
74-
tensor, pten::TransToPtenBackend(place), pten::TransToPtenDataType(type),
75-
pten::TransToPtenDataLayout(tensor.layout()));
76-
}
77-
78-
template <>
79-
void ShareTensorImpl<pten::DenseTensor>(pten::DenseTensor* tensor_impl,
80-
LoDTensor* out) {
81-
out->ResetHolderWithType(tensor_impl->allocation(),
82-
pten::TransToProtoVarType(tensor_impl->data_type()));
83-
}
84-
85-
template <>
86-
void ShareTensorImpl<pten::DenseTensor>(pten::DenseTensor* tensor_impl,
87-
Tensor* out) {
88-
out->ResetHolderWithType(tensor_impl->allocation(),
89-
pten::TransToProtoVarType(tensor_impl->data_type()));
90-
}
91-
92-
std::shared_ptr<pten::TensorBase> InputVariableToPtenTensor(
93-
const framework::Variable& variable, const pten::TensorArgDef& arg_def) {
94-
auto expected_place = pten::TransToFluidPlace(arg_def.backend);
95-
96-
if (variable.template IsType<framework::LoDTensor>()) {
97-
const auto& tensor = variable.template Get<framework::LoDTensor>();
98-
if (!platform::is_same_place(tensor.place(), expected_place)) {
99-
framework::LoDTensor tmp_tensor;
100-
framework::TensorCopySync(tensor, expected_place, &tmp_tensor);
101-
auto pt_in =
102-
framework::MakeTensorImpl<pten::DenseTensor, framework::LoDTensor>(
103-
tmp_tensor, arg_def.backend, arg_def.dtype, arg_def.layout);
104-
return pt_in;
105-
} else {
106-
auto pt_in =
107-
framework::MakeTensorImpl<pten::DenseTensor, framework::LoDTensor>(
108-
tensor, arg_def.backend, arg_def.dtype, arg_def.layout);
109-
return pt_in;
110-
}
111-
} else if (variable.template IsType<framework::SelectedRows>()) {
112-
// TODO(chenweihang): now we don't deal with row and height
113-
// by xiaowei's advice
114-
const auto& tensor = variable.template Get<framework::SelectedRows>();
115-
if (!platform::is_same_place(tensor.value().place(), expected_place)) {
116-
framework::Tensor tmp_tensor;
117-
TensorCopySync(tensor.value(), expected_place, &tmp_tensor);
118-
// TODO(chenweihang): adapt SelectedRows by xiaowei's design
119-
auto pt_in =
120-
framework::MakeTensorImpl<pten::DenseTensor, framework::Tensor>(
121-
tmp_tensor, arg_def.backend, arg_def.dtype, arg_def.layout);
122-
return pt_in;
123-
} else {
124-
auto pt_in =
125-
framework::MakeTensorImpl<pten::DenseTensor, framework::Tensor>(
126-
tensor.value(), arg_def.backend, arg_def.dtype, arg_def.layout);
127-
return pt_in;
128-
}
129-
} else {
130-
PADDLE_THROW(platform::errors::Unimplemented(
131-
"Unsupported shared input `%s` type now when call pt kernel.",
132-
framework::ToTypeName(variable.Type())));
133-
}
134-
return nullptr;
135-
}
136-
137-
std::shared_ptr<pten::TensorBase> OutputVariableToPtenTensor(
138-
framework::Variable* variable, const pten::TensorArgDef& arg_def) {
139-
// mutable_data before run kernel, to avoid share output form
140-
// KernelContext to original tensor
141-
if (variable->template IsType<framework::LoDTensor>()) {
142-
auto* tensor = variable->template GetMutable<framework::LoDTensor>();
143-
tensor->mutable_data(pten::TransToFluidPlace(arg_def.backend),
144-
pten::TransToProtoVarType(arg_def.dtype));
145-
auto pt_out =
146-
framework::MakeTensorImpl<pten::DenseTensor, framework::LoDTensor>(
147-
*tensor, arg_def.backend, arg_def.dtype, arg_def.layout);
148-
return pt_out;
149-
} else if (variable->template IsType<framework::SelectedRows>()) {
150-
auto* tensor = variable->template GetMutable<framework::SelectedRows>();
151-
tensor->mutable_value()->mutable_data(
152-
pten::TransToFluidPlace(arg_def.backend),
153-
pten::TransToProtoVarType(arg_def.dtype));
154-
// TODO(chenweihang): adapt SelectedRows by xiaowei's design,
155-
// here the row and height will lost in output!
156-
auto pt_out =
157-
framework::MakeTensorImpl<pten::DenseTensor, framework::Tensor>(
158-
tensor->value(), arg_def.backend, arg_def.dtype, arg_def.layout);
159-
return pt_out;
160-
} else {
161-
PADDLE_THROW(platform::errors::Unimplemented(
162-
"Unsupported shared output `%s` type now when call pt kernel.",
163-
framework::ToTypeName(variable->Type())));
164-
}
165-
166-
return nullptr;
167-
}
168-
16927
OpKernelType TransPtenKernelKeyToOpKernelType(
17028
const pten::KernelKey& kernel_key) {
17129
proto::VarType::Type data_type =

‎paddle/fluid/framework/pten_utils.h‎

Lines changed: 1 addition & 29 deletions
Original file line numberDiff line numberDiff line change
@@ -25,41 +25,13 @@ limitations under the License. */
2525
#include "paddle/fluid/platform/macros.h"
2626
#include "paddle/fluid/platform/place.h"
2727
#include "paddle/pten/api/include/core.h"
28+
#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
2829
#include "paddle/utils/flat_hash_map.h"
2930
#include "paddle/utils/small_vector.h"
3031

3132
namespace paddle {
3233
namespace framework {
3334

34-
/* tensor translate */
35-
36-
template <typename PtenTensorImplT, typename VariableT>
37-
std::shared_ptr<PtenTensorImplT> MakeTensorImpl(
38-
const VariableT& tensor, pten::Backend backend,
39-
paddle::experimental::DataType dtype,
40-
paddle::experimental::DataLayout layout);
41-
42-
template <typename PtenTensorImplT>
43-
std::shared_ptr<PtenTensorImplT> MakeTensorImpl(const LoDTensor& tensor,
44-
const platform::Place& place,
45-
proto::VarType::Type type);
46-
47-
template <typename PtenTensorImplT>
48-
std::shared_ptr<PtenTensorImplT> MakeTensorImpl(const Tensor& tensor,
49-
const platform::Place& place,
50-
proto::VarType::Type type);
51-
52-
template <typename PtenTensorImplT>
53-
void ShareTensorImpl(PtenTensorImplT* tensor_impl, LoDTensor* out);
54-
55-
template <typename PtenTensorImplT>
56-
void ShareTensorImpl(PtenTensorImplT* tensor_impl, Tensor* out);
57-
58-
std::shared_ptr<pten::TensorBase> InputVariableToPtenTensor(
59-
const framework::Variable& variable, const pten::TensorArgDef& arg_def);
60-
std::shared_ptr<pten::TensorBase> OutputVariableToPtenTensor(
61-
framework::Variable* variable, const pten::TensorArgDef& arg_def);
62-
6335
/* Kernel Key translate */
6436

6537
OpKernelType TransPtenKernelKeyToOpKernelType(

‎paddle/fluid/framework/pten_utils_test.cc‎

Lines changed: 0 additions & 60 deletions
Original file line numberDiff line numberDiff line change
@@ -18,66 +18,6 @@ limitations under the License. */
1818
#include "paddle/fluid/framework/selected_rows.h"
1919
#include "paddle/fluid/framework/variable.h"
2020

21-
TEST(PtenUtils, FluidTensorToPtenTensor) {
22-
// 1. create tensor
23-
paddle::framework::LoDTensor x;
24-
paddle::framework::Tensor x2;
25-
x.Resize({2});
26-
x.mutable_data<float>(paddle::platform::CPUPlace());
27-
x.data<float>()[0] = 0.2;
28-
x.data<float>()[1] = 0.5;
29-
30-
// 2. test API
31-
auto dense_x = paddle::framework::MakeTensorImpl<pten::DenseTensor>(
32-
x, x.place(), x.type());
33-
34-
// 3. check result
35-
std::vector<float> expect_value = {0.2, 0.5};
36-
ASSERT_EQ(dense_x->data<float>()[0], expect_value[0]);
37-
ASSERT_EQ(dense_x->data<float>()[1], expect_value[1]);
38-
ASSERT_EQ(dense_x->backend(), pten::Backend::CPU);
39-
ASSERT_EQ(dense_x->data_type(), pten::DataType::FLOAT32);
40-
}
41-
42-
TEST(PtenUtils, VarToPtenTensor) {
43-
// 1. create Variable
44-
paddle::framework::Variable v;
45-
auto selected_rows = v.GetMutable<paddle::framework::SelectedRows>();
46-
paddle::framework::Tensor* value = selected_rows->mutable_value();
47-
auto* data = value->mutable_data<int>(paddle::framework::make_ddim({1, 1}),
48-
paddle::platform::CPUPlace());
49-
data[0] = 123;
50-
pten::Backend expect_backend = pten::Backend::CPU;
51-
52-
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
53-
expect_backend = pten::Backend::CUDA;
54-
#endif
55-
auto tensor_def = pten::TensorArgDef(expect_backend, pten::DataLayout::NCHW,
56-
pten::DataType::INT32);
57-
// 2. test API
58-
auto tensor_x = paddle::framework::InputVariableToPtenTensor(v, tensor_def);
59-
// 3. check result
60-
ASSERT_EQ(tensor_x->backend(), expect_backend);
61-
ASSERT_EQ(tensor_x->data_type(), pten::DataType::INT32);
62-
}
63-
64-
TEST(PtenUtils, PtenTensorToFluidTensor) {
65-
pten::DenseTensor dense_tensor(
66-
pten::TensorMeta(paddle::framework::make_ddim({1, 1}), pten::Backend::CPU,
67-
pten::DataType::FLOAT32, pten::DataLayout::ANY),
68-
pten::TensorStatus());
69-
auto* data_ptr = dense_tensor.mutable_data<float>();
70-
data_ptr[0] = 0.5;
71-
// share allocation into fluid Tensor
72-
paddle::framework::Tensor tensor;
73-
paddle::framework::LoDTensor lod_tensor;
74-
paddle::framework::ShareTensorImpl(&dense_tensor, &tensor);
75-
paddle::framework::ShareTensorImpl(&dense_tensor, &lod_tensor);
76-
// compare
77-
ASSERT_EQ(tensor.data<float>()[0], 0.5);
78-
ASSERT_EQ(lod_tensor.data<float>()[0], 0.5);
79-
}
80-
8121
TEST(PtenUtils, TransPtenKernelKeyToOpKernelType) {
8222
pten::KernelKey kernel_key(pten::Backend::CPU, pten::DataLayout::NCHW,
8323
pten::DataType::FLOAT32);

‎paddle/fluid/imperative/prepared_operator.cc‎

Lines changed: 6 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -293,11 +293,10 @@ static pten::KernelContext BuildDygraphPtenKernelContext(
293293
paddle::SmallVector<std::shared_ptr<pten::TensorBase>> tmp_inputs;
294294
for (auto var : ins_vector) {
295295
const auto& variable = var->Var();
296-
297-
auto pt_in = framework::InputVariableToPtenTensor(variable, in_def);
298-
tmp_inputs.emplace_back(pt_in);
296+
tmp_inputs.emplace_back(
297+
experimental::MakePtenTensorBaseFromVar(variable, in_def));
299298
}
300-
op_kernel_ctx.EmplaceBackInputs(tmp_inputs);
299+
op_kernel_ctx.EmplaceBackInputs(std::move(tmp_inputs));
301300
}
302301

303302
for (size_t i = 0; i < output_names.size(); ++i) {
@@ -307,11 +306,10 @@ static pten::KernelContext BuildDygraphPtenKernelContext(
307306
paddle::SmallVector<std::shared_ptr<pten::TensorBase>> tmp_outputs;
308307
for (auto var : outs_vector) {
309308
auto* variable = var->MutableVar();
310-
311-
auto pt_out = framework::OutputVariableToPtenTensor(variable, out_def);
312-
tmp_outputs.emplace_back(pt_out);
309+
tmp_outputs.emplace_back(
310+
experimental::MakePtenTensorBaseFromVar(variable, out_def));
313311
}
314-
op_kernel_ctx.EmplaceBackOutputs(tmp_outputs);
312+
op_kernel_ctx.EmplaceBackOutputs(std::move(tmp_outputs));
315313
}
316314

317315
for (size_t i = 0; i < attr_names.size(); ++i) {

‎paddle/fluid/operators/CMakeLists.txt‎

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -80,8 +80,9 @@ if(WITH_UNITY_BUILD)
8080
endif()
8181

8282
set(OP_HEADER_DEPS ${OP_HEADER_DEPS} pten)
83-
set(OP_HEADER_DEPS ${OP_HEADER_DEPS} pten_utils)
84-
register_operators(EXCLUDES py_layer_op py_func_op warpctc_op dgc_op load_combine_op lstm_op run_program_op eye_op
83+
#set(OP_HEADER_DEPS ${OP_HEADER_DEPS} pten_utils)
84+
register_operators(EXCLUDES
85+
py_layer_op py_func_op warpctc_op dgc_op load_combine_op lstm_op run_program_op eye_op
8586
recurrent_op save_combine_op sparse_attention_op sync_batch_norm_op spectral_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS})
8687

8788
op_library(run_program_op SRCS run_program_op.cc run_program_op.cu.cc DEPS executor_cache ${OP_HEADER_DEPS})

‎paddle/fluid/operators/dot_op.h‎

Lines changed: 4 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -16,13 +16,13 @@
1616

1717
#include "paddle/fluid/framework/op_registry.h"
1818
#include "paddle/fluid/framework/operator.h"
19-
#include "paddle/fluid/framework/pten_utils.h"
2019
#include "paddle/fluid/operators/math/complex_functors.h"
2120
#include "paddle/fluid/platform/for_range.h"
2221

2322
// only can include the headers in paddle/pten/api dirs
2423
#include "paddle/pten/api/include/core.h"
2524
#include "paddle/pten/api/include/linalg.h"
25+
#include "paddle/pten/hapi/lib/utils/tensor_utils.h"
2626

2727
namespace paddle {
2828
namespace operators {
@@ -244,12 +244,9 @@ class DotKernel : public framework::OpKernel<T> {
244244
auto& dev_ctx = ctx.device_context<DeviceContext>();
245245
out->mutable_data<T>(x->place());
246246

247-
auto pt_x =
248-
framework::MakeTensorImpl<pten::DenseTensor>(*x, x->place(), x->type());
249-
auto pt_y =
250-
framework::MakeTensorImpl<pten::DenseTensor>(*y, y->place(), y->type());
251-
auto pt_out = framework::MakeTensorImpl<pten::DenseTensor>(*out, x->place(),
252-
x->type());
247+
auto pt_x = paddle::experimental::MakePtenDenseTensor(*x);
248+
auto pt_y = paddle::experimental::MakePtenDenseTensor(*y);
249+
auto pt_out = paddle::experimental::MakePtenDenseTensor(*out);
253250

254251
// call new kernel
255252
pten::Dot<T>(dev_ctx, *pt_x.get(), *pt_y.get(), pt_out.get());

‎paddle/fluid/operators/fill_any_like_op.h‎

Lines changed: 2 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -62,10 +62,8 @@ class FillAnyLikeKernel : public framework::OpKernel<T> {
6262
std::isnan(value), false,
6363
platform::errors::InvalidArgument("The filled value is NaN."));
6464

65-
auto pt_x = framework::MakeTensorImpl<pten::DenseTensor>(*in, in->place(),
66-
in->type());
67-
auto pt_out = framework::MakeTensorImpl<pten::DenseTensor>(
68-
*out, out->place(), out->type());
65+
auto pt_x = paddle::experimental::MakePtenDenseTensor(*in);
66+
auto pt_out = paddle::experimental::MakePtenDenseTensor(*out);
6967

7068
const auto& dev_ctx = context.template device_context<DeviceContext>();
7169
// call new kernel

0 commit comments

Comments
 (0)