Compare commits
62 Commits
develop
...
fix_doc_bu
| Author | SHA1 | Date |
|---|---|---|
swtkiwi
|
13e906c162 | 6 years ago |
danleifeng
|
3f8746842b | 6 years ago |
Wilber
|
c008ae696a | 6 years ago |
lidanqing
|
56aba4b747 | 6 years ago |
wangguanzhong
|
86887b9b1f | 6 years ago |
|
wangguanzhong
|
5178d9f832 | 6 years ago |
liu zhengxi
|
5d94a5ca54 | 6 years ago |
李灿
|
4193ae0d02 | 6 years ago |
Zhou Wei
|
dc53ba879f | 6 years ago |
WangXi
|
739043c6a6 | 6 years ago |
Double_V
|
39745d448c | 6 years ago |
LielinJiang
|
172e76a852 | 6 years ago |
tianshuo78520a
|
f63f8d730c | 6 years ago |
 李灿
|
2477ccc378 | 6 years ago |
|
danleifeng
|
61162497b4 | 6 years ago |
|
LielinJiang
|
d4160941f3 | 6 years ago |
yinhaofeng
|
0b294906f9 | 6 years ago |
|
yinhaofeng
|
ea45fb90a4 | 6 years ago |
|
Double_V
|
4f43d51f43 | 6 years ago |
lilong12
|
c4e18dc093 | 6 years ago |
Kaipeng Deng
|
6375ad39cc | 6 years ago |
TeslaZhao
|
11adb0f373 | 6 years ago |
|
lilong12
|
957e6fbe5f | 6 years ago |
Chengmo
|
91727ac899 | 6 years ago |
liuyuhui
|
5c2852a330 | 6 years ago |
huangxu96
|
6bb6cb27b8 | 6 years ago |
xiaoting
|
905b076553 | 6 years ago |
|
Double_V
|
8600f47439 | 6 years ago |
Pei Yang
|
d89deae9e0 | 6 years ago |
|
lilong12
|
316c97c7eb | 6 years ago |
pangyoki
|
46a1f69b3c | 6 years ago |
|
pangyoki
|
b21409e0c0 | 6 years ago |
|
pangyoki
|
69ec13cdf0 | 6 years ago |
|
pangyoki
|
386429beb0 | 6 years ago |
MRXLT
|
e3513a6395 | 6 years ago |
LutaoChu
|
e3a88eb450 | 6 years ago |
littletomatodonkey
|
edbaa027fe | 6 years ago |
|
liuyuhui
|
77eddf9168 | 6 years ago |
|
Kaipeng Deng
|
f05b184f72 | 6 years ago |
|
Zhou Wei
|
e9a1669e92 | 6 years ago |
|
xiaoting
|
5c1babde56 | 6 years ago |
Guo Sheng
|
3f565903e8 | 6 years ago |
Jack Zhou
|
a6c18075d6 | 6 years ago |
chentianyu03
|
2ca53fa65f | 6 years ago |
LiuChiachi
|
4316bd4d5a | 6 years ago |
|
lidanqing
|
ea76fe310c | 6 years ago |
tangwei12
|
c0550b54a5 | 6 years ago |
Aurelius84
|
772c266822 | 6 years ago |
|
chentianyu03
|
77866effeb | 6 years ago |
|
chentianyu03
|
cafa35e1ea | 6 years ago |
|
chentianyu03
|
3251f9c1eb | 6 years ago |
Steffy-zxf
|
164b9aabc1 | 6 years ago |
mapingshuo
|
50d24899cf | 6 years ago |
|
Zhou Wei
|
b57254ed61 | 6 years ago |
123malin
|
c0061ff56f | 6 years ago |
Guanghua Yu
|
51dd268cfe | 6 years ago |
Feiyu Chan
|
429c0b62b1 | 6 years ago |
|
mapingshuo
|
39c31a20e5 | 6 years ago |
Qinghe JING
|
1f45c06e92 | 6 years ago |
Yiqun Liu
|
ef2b12f11c | 6 years ago |
|
tianshuo78520a
|
b7fd4f9224 | 6 years ago |
Thunderbrook
|
ba6a29070a | 6 years ago |
@ -0,0 +1,161 @@
|
|||||||
|
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
|
||||||
|
|
||||||
|
Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
you may not use this file except in compliance with the License.
|
||||||
|
You may obtain a copy of the License at
|
||||||
|
|
||||||
|
http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
|
||||||
|
Unless required by applicable law or agreed to in writing, software
|
||||||
|
distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
See the License for the specific language governing permissions and
|
||||||
|
limitations under the License. */
|
||||||
|
|
||||||
|
#ifdef PADDLE_WITH_XPU
|
||||||
|
#include "paddle/fluid/operators/assign_op.h"
|
||||||
|
|
||||||
|
#include <string>
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace framework {
|
||||||
|
class OpDesc;
|
||||||
|
class Variable;
|
||||||
|
} // namespace framework
|
||||||
|
namespace imperative {
|
||||||
|
class OpBase;
|
||||||
|
} // namespace imperative
|
||||||
|
namespace platform {
|
||||||
|
struct CPUPlace;
|
||||||
|
struct CUDAPlace;
|
||||||
|
struct float16;
|
||||||
|
} // namespace platform
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
class AssignOp : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
AssignOp(const std::string &type, const framework::VariableNameMap &inputs,
|
||||||
|
const framework::VariableNameMap &outputs,
|
||||||
|
const framework::AttributeMap &attrs)
|
||||||
|
: OperatorWithKernel(type, inputs, outputs, attrs) {}
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext *ctx) const override {
|
||||||
|
if (ctx->HasInput("X")) {
|
||||||
|
auto type = ctx->GetInputsVarType("X")[0];
|
||||||
|
if (type == framework::proto::VarType::SELECTED_ROWS ||
|
||||||
|
type == framework::proto::VarType::LOD_TENSOR) {
|
||||||
|
ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
|
||||||
|
if (type == framework::proto::VarType::LOD_TENSOR) {
|
||||||
|
ctx->ShareLoD("X", /*->*/ "Out");
|
||||||
|
}
|
||||||
|
} else if (type == framework::proto::VarType::LOD_TENSOR_ARRAY) {
|
||||||
|
if (ctx->IsRuntime()) {
|
||||||
|
// The runtime output shape is determined in kernel.
|
||||||
|
return;
|
||||||
|
} else {
|
||||||
|
ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
protected:
|
||||||
|
framework::OpKernelType GetKernelTypeForVar(
|
||||||
|
const std::string &var_name, const framework::Tensor &tensor,
|
||||||
|
const framework::OpKernelType &expected_kernel_type) const override {
|
||||||
|
return framework::OpKernelType(expected_kernel_type.data_type_,
|
||||||
|
expected_kernel_type.place_,
|
||||||
|
tensor.layout());
|
||||||
|
}
|
||||||
|
|
||||||
|
framework::OpKernelType GetExpectedKernelType(
|
||||||
|
const framework::ExecutionContext &ctx) const override {
|
||||||
|
const framework::Variable *var = ctx.InputVar("X");
|
||||||
|
if (var->IsType<framework::LoDTensorArray>()) {
|
||||||
|
auto t_arr = var->Get<framework::LoDTensorArray>();
|
||||||
|
// NOTE(liym27): Support an empty tensor array as Input.
|
||||||
|
// And set the kernel type is float.
|
||||||
|
if (t_arr.size() == 0) {
|
||||||
|
return framework::OpKernelType(framework::proto::VarType::FP32,
|
||||||
|
ctx.device_context());
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
return framework::OpKernelType(
|
||||||
|
OperatorWithKernel::IndicateVarDataType(ctx, "X"),
|
||||||
|
ctx.device_context());
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class AssignInferVarType : public framework::VarTypeInference {
|
||||||
|
public:
|
||||||
|
void operator()(framework::InferVarTypeContext *ctx) const override {
|
||||||
|
ctx->SyncTypeAndDataType("X", "Out");
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class AssignKernel {
|
||||||
|
public:
|
||||||
|
void operator()(const framework::ExecutionContext &ctx) const {
|
||||||
|
auto *x = ctx.InputVar("X");
|
||||||
|
if (x == nullptr) {
|
||||||
|
return;
|
||||||
|
}
|
||||||
|
PADDLE_ENFORCE_EQ(
|
||||||
|
ctx.HasOutput("Out"), true,
|
||||||
|
platform::errors::NotFound("Output(Out) of assign_op is not found."));
|
||||||
|
auto *out = ctx.OutputVar("Out");
|
||||||
|
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
|
||||||
|
auto &dev_ctx = *pool.Get(ctx.GetPlace());
|
||||||
|
|
||||||
|
framework::VisitVarType(*x, AssignFunctor(out, dev_ctx));
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
class AssignOpProtoMaker : public framework::OpProtoAndCheckerMaker {
|
||||||
|
public:
|
||||||
|
void Make() override {
|
||||||
|
AddInput("X",
|
||||||
|
"(LoDTensor, SelectedRows or LoDTensorArray) The input variable "
|
||||||
|
"could be LoDTensor, SelectedRows or LoDTensorArray.")
|
||||||
|
.AsDispensable();
|
||||||
|
AddOutput("Out",
|
||||||
|
"(LoDTensor, SelectedRows or LoDTensorArray) The type of output "
|
||||||
|
"is the same as input X.");
|
||||||
|
AddComment(R"DOC(Assign Operator
|
||||||
|
|
||||||
|
Out = X, when type in [LoDTensor/SelectedRows/LoDTensorArray]
|
||||||
|
raise error if the type is not listed above.
|
||||||
|
)DOC");
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class AssignGradMaker : public framework::SingleGradOpMaker<T> {
|
||||||
|
public:
|
||||||
|
using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
|
||||||
|
|
||||||
|
protected:
|
||||||
|
void Apply(GradOpPtr<T> op) const override {
|
||||||
|
op->SetType("assign");
|
||||||
|
op->SetInput("X", this->OutputGrad("Out"));
|
||||||
|
op->SetOutput("Out", this->InputGrad("X"));
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
DECLARE_INPLACE_OP_INFERER(AssignOpInplaceInferer, {"X", "Out"});
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
namespace plat = paddle::platform;
|
||||||
|
|
||||||
|
REGISTER_OP_XPU_KERNEL_FUNCTOR(assign, float, ops::AssignKernel, double,
|
||||||
|
ops::AssignKernel, int, ops::AssignKernel,
|
||||||
|
int64_t, ops::AssignKernel, bool,
|
||||||
|
ops::AssignKernel);
|
||||||
|
#endif
|
||||||
@ -0,0 +1,69 @@
|
|||||||
|
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
|
||||||
|
|
||||||
|
Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
you may not use this file except in compliance with the License.
|
||||||
|
You may obtain a copy of the License at
|
||||||
|
|
||||||
|
http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
|
||||||
|
Unless required by applicable law or agreed to in writing, software
|
||||||
|
distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
See the License for the specific language governing permissions and
|
||||||
|
limitations under the License. */
|
||||||
|
|
||||||
|
#ifdef PADDLE_WITH_XPU
|
||||||
|
#include "paddle/fluid/operators/cast_op.h"
|
||||||
|
#include <memory>
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
#include "paddle/fluid/platform/float16.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
template <typename DeviceContext, typename InT>
|
||||||
|
class CastXPUKernel : public framework::OpKernel<InT> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& context) const override {
|
||||||
|
auto* in = context.Input<framework::Tensor>("X");
|
||||||
|
auto* out = context.Output<framework::Tensor>("Out");
|
||||||
|
auto in_type = static_cast<framework::proto::VarType::Type>(
|
||||||
|
context.Attr<int>("in_dtype"));
|
||||||
|
auto out_type = static_cast<framework::proto::VarType::Type>(
|
||||||
|
context.Attr<int>("out_dtype"));
|
||||||
|
auto* in_data = in->data<InT>();
|
||||||
|
auto numel = in->numel();
|
||||||
|
auto& dev_ctx = context.template device_context<DeviceContext>();
|
||||||
|
int r = -1;
|
||||||
|
if (out_type == framework::proto::VarType::FP32) {
|
||||||
|
auto* out_data = out->mutable_data<float>(context.GetPlace());
|
||||||
|
r = xpu::cast<InT, float>(dev_ctx.x_context(), in_data, out_data, numel);
|
||||||
|
} else if (out_type == framework::proto::VarType::INT32) {
|
||||||
|
auto* out_data = out->mutable_data<int>(context.GetPlace());
|
||||||
|
r = xpu::cast<InT, int>(dev_ctx.x_context(), in_data, out_data, numel);
|
||||||
|
} else if (out_type == framework::proto::VarType::INT64) {
|
||||||
|
auto* out_data = out->mutable_data<int64_t>(context.GetPlace());
|
||||||
|
r = xpu::cast<InT, int64_t>(dev_ctx.x_context(), in_data, out_data,
|
||||||
|
numel);
|
||||||
|
} else {
|
||||||
|
PADDLE_THROW(platform::errors::Unavailable("Not supported cast %d -> %d",
|
||||||
|
in_type, out_type));
|
||||||
|
}
|
||||||
|
PADDLE_ENFORCE_EQ(
|
||||||
|
r, XPU_SUCCESS,
|
||||||
|
platform::errors::External(
|
||||||
|
"XPU API return wrong value[%d], please check whether "
|
||||||
|
"Baidu Kunlun Card is properly installed.",
|
||||||
|
r));
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OP_XPU_KERNEL(
|
||||||
|
cast, ops::CastXPUKernel<paddle::platform::XPUDeviceContext, int>,
|
||||||
|
ops::CastXPUKernel<paddle::platform::XPUDeviceContext, float>,
|
||||||
|
ops::CastXPUKernel<paddle::platform::XPUDeviceContext, int64_t>);
|
||||||
|
#endif
|
||||||
@ -0,0 +1,185 @@
|
|||||||
|
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
|
||||||
|
|
||||||
|
Licensed under the Apache License, Version 2.0 (the "License");
|
||||||
|
you may not use this file except in compliance with the License.
|
||||||
|
You may obtain a copy of the License at
|
||||||
|
|
||||||
|
http://www.apache.org/licenses/LICENSE-2.0
|
||||||
|
|
||||||
|
Unless required by applicable law or agreed to in writing, software
|
||||||
|
distributed under the License is distributed on an "AS IS" BASIS,
|
||||||
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||||
|
See the License for the specific language governing permissions and
|
||||||
|
limitations under the License. */
|
||||||
|
|
||||||
|
#include "paddle/fluid/operators/concat_op.h"
|
||||||
|
|
||||||
|
#include <memory>
|
||||||
|
#include <string>
|
||||||
|
#include <vector>
|
||||||
|
|
||||||
|
#ifdef PADDLE_WITH_MKLDNN
|
||||||
|
#include <paddle/fluid/platform/mkldnn_helper.h>
|
||||||
|
#endif
|
||||||
|
|
||||||
|
#ifdef PADDLE_WITH_XPU
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
using Tensor = framework::Tensor;
|
||||||
|
|
||||||
|
template <typename DeviceContext, typename T>
|
||||||
|
class ConcatXPUKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||||
|
auto ins = ctx.MultiInput<framework::Tensor>("X");
|
||||||
|
framework::Tensor* out = ctx.Output<framework::Tensor>("Out");
|
||||||
|
int axis = ctx.Attr<int>("axis");
|
||||||
|
PADDLE_ENFORCE_NE(ins[0], nullptr, platform::errors::InvalidArgument(
|
||||||
|
"The input should not be null."));
|
||||||
|
PADDLE_ENFORCE_NE(ctx.HasInput("AxisTensor"), true,
|
||||||
|
platform::errors::InvalidArgument(
|
||||||
|
"XPU donot surpport AxisTensor for now"));
|
||||||
|
axis = ComputeAxis(static_cast<int64_t>(axis),
|
||||||
|
static_cast<int64_t>(ins[0]->dims().size()));
|
||||||
|
PADDLE_ENFORCE_GE(
|
||||||
|
axis, 0, platform::errors::InvalidArgument("concat: axis shoud >= 0!"));
|
||||||
|
PADDLE_ENFORCE_LT(axis, ins[0]->dims().size(),
|
||||||
|
platform::errors::InvalidArgument(
|
||||||
|
"concat: axis shoud < ins[0]->dims()!"));
|
||||||
|
auto place = ctx.GetPlace();
|
||||||
|
out->mutable_data<T>(place);
|
||||||
|
std::vector<int> choose_idx;
|
||||||
|
int n = 0;
|
||||||
|
for (unsigned int i = 0; i < ins.size(); ++i) {
|
||||||
|
if (ins[i] && ins[i]->numel() > 0) {
|
||||||
|
choose_idx.push_back(i);
|
||||||
|
n++;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
PADDLE_ENFORCE_LE(n, 8, platform::errors::InvalidArgument(
|
||||||
|
"XPU only surpport at most 8 tensors for now"));
|
||||||
|
PADDLE_ENFORCE_GT(
|
||||||
|
n, 0, platform::errors::InvalidArgument("No tensor need concat?"));
|
||||||
|
int h = 1;
|
||||||
|
int w_except_axis = 1;
|
||||||
|
for (int i = 0; i < axis; ++i) {
|
||||||
|
h *= (ins[choose_idx[0]]->dims())[i];
|
||||||
|
}
|
||||||
|
for (int i = axis + 1; i < ins[0]->dims().size(); ++i) {
|
||||||
|
w_except_axis *= (ins[choose_idx[0]]->dims())[i];
|
||||||
|
}
|
||||||
|
for (int i = 1; i < n; ++i) {
|
||||||
|
int hh = 1;
|
||||||
|
int ww = 1;
|
||||||
|
for (int j = 0; j < axis; ++j) {
|
||||||
|
hh *= (ins[choose_idx[i]]->dims())[j];
|
||||||
|
}
|
||||||
|
for (int j = axis + 1; j < ins[i]->dims().size(); ++j) {
|
||||||
|
ww *= (ins[choose_idx[i]]->dims())[j];
|
||||||
|
}
|
||||||
|
PADDLE_ENFORCE_EQ(hh, h, platform::errors::InvalidArgument(
|
||||||
|
"concat: h should be eual!"));
|
||||||
|
PADDLE_ENFORCE_EQ(ww, w_except_axis,
|
||||||
|
platform::errors::InvalidArgument(
|
||||||
|
"concat: w should be eual except for axis!"));
|
||||||
|
}
|
||||||
|
auto& dev_ctx = ctx.template device_context<DeviceContext>();
|
||||||
|
std::unique_ptr<int[]> in_w_host(new int[n]);
|
||||||
|
std::unique_ptr<const float* []> ptrs(new const float*[n]);
|
||||||
|
for (int i = 0; i < n; ++i) {
|
||||||
|
ptrs[i] = ins[choose_idx[i]]->data<T>();
|
||||||
|
in_w_host[i] = w_except_axis * (ins[choose_idx[i]]->dims())[axis];
|
||||||
|
}
|
||||||
|
int r =
|
||||||
|
xpu::concat<float>(dev_ctx.x_context(), h, (const int*)in_w_host.get(),
|
||||||
|
n, (const float**)ptrs.get(), out->data<T>());
|
||||||
|
PADDLE_ENFORCE_EQ(
|
||||||
|
r, XPU_SUCCESS,
|
||||||
|
platform::errors::External(
|
||||||
|
"XPU API return wrong value[%d], please check whether "
|
||||||
|
"Baidu Kunlun Card is properly installed.",
|
||||||
|
r));
|
||||||
|
}
|
||||||
|
};
|
||||||
|
template <typename DeviceContext, typename T>
|
||||||
|
class ConcatGradXPUKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& ctx) const {
|
||||||
|
auto* out_grad =
|
||||||
|
ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
|
||||||
|
auto ins = ctx.MultiInput<framework::LoDTensor>("X");
|
||||||
|
auto out_var_names = ctx.OutputNames(framework::GradVarName("X"));
|
||||||
|
auto outs =
|
||||||
|
ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
|
||||||
|
{
|
||||||
|
auto dx = outs;
|
||||||
|
auto x = ins;
|
||||||
|
for (size_t i = 0; i < dx.size(); ++i) {
|
||||||
|
if (dx[i] != nullptr) {
|
||||||
|
dx[i]->set_lod(x[i]->lod());
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
PADDLE_ENFORCE_NE(ins[0], nullptr, platform::errors::InvalidArgument(
|
||||||
|
"The input should not be null."));
|
||||||
|
auto axis = ctx.Attr<int>("axis");
|
||||||
|
if (ctx.HasInput("AxisTensor")) {
|
||||||
|
auto* axis_tensor = ctx.Input<framework::Tensor>("AxisTensor");
|
||||||
|
axis = GetDataFromTensor<int>(axis_tensor)[0];
|
||||||
|
}
|
||||||
|
axis = ComputeAxis(static_cast<int64_t>(axis),
|
||||||
|
static_cast<int64_t>(ins[0]->dims().size()));
|
||||||
|
// get output tensor that the name is not kEmptyVarName
|
||||||
|
std::vector<framework::Tensor*> outputs;
|
||||||
|
for (size_t j = 0; j < outs.size(); ++j) {
|
||||||
|
if (out_var_names[j] != framework::kEmptyVarName &&
|
||||||
|
outs[j]->numel() != 0UL) {
|
||||||
|
outs[j]->mutable_data<T>(ctx.GetPlace());
|
||||||
|
outputs.push_back(outs[j]);
|
||||||
|
} else {
|
||||||
|
outputs.push_back(nullptr);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
PADDLE_ENFORCE_GE(axis, 0, platform::errors::InvalidArgument(
|
||||||
|
"concat_grad: axis shoud >= 0!"));
|
||||||
|
PADDLE_ENFORCE_LT(axis, out_grad->dims().size(),
|
||||||
|
platform::errors::InvalidArgument(
|
||||||
|
"concat_grad: axis shoud < ins[0]->dims()!"));
|
||||||
|
auto out_grad_stride = framework::stride_numel(out_grad->dims());
|
||||||
|
int n = outputs.size();
|
||||||
|
PADDLE_ENFORCE_LE(n, 16,
|
||||||
|
platform::errors::InvalidArgument(
|
||||||
|
"XPU only surpport at most 16 tensors for now"));
|
||||||
|
int h = out_grad_stride[0] / out_grad_stride[axis];
|
||||||
|
auto& dev_ctx = ctx.template device_context<DeviceContext>();
|
||||||
|
std::unique_ptr<int[]> in_w_host(new int[n]);
|
||||||
|
std::unique_ptr<float* []> ptrs(new float*[n]);
|
||||||
|
for (int i = 0; i < n; ++i) {
|
||||||
|
auto out_stride = framework::stride_numel(outputs[i]->dims());
|
||||||
|
ptrs[i] = outputs[i]->data<T>();
|
||||||
|
in_w_host[i] = out_stride[axis];
|
||||||
|
}
|
||||||
|
int r = xpu::concat_grad(dev_ctx.x_context(), h, in_w_host.get(), n,
|
||||||
|
reinterpret_cast<float**>(ptrs.get()),
|
||||||
|
out_grad->data<T>());
|
||||||
|
PADDLE_ENFORCE_EQ(
|
||||||
|
r, XPU_SUCCESS,
|
||||||
|
platform::errors::External(
|
||||||
|
"XPU API return wrong value[%d], please check whether "
|
||||||
|
"Baidu Kunlun Card is properly installed.",
|
||||||
|
r));
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OP_XPU_KERNEL(
|
||||||
|
concat, ops::ConcatXPUKernel<paddle::platform::XPUDeviceContext, float>);
|
||||||
|
REGISTER_OP_XPU_KERNEL(
|
||||||
|
concat_grad,
|
||||||
|
ops::ConcatGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
|
||||||
|
|
||||||
|
#endif
|
||||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue