未验证 提交 ccf61b30 编写于 作者: T Tao Luo 提交者: GitHub

Merge pull request #10703 from mozga-intel/mozga-intel/Mul_mkldnn_withdraw

Withdraw MKLDNN Mul operator
/* Copyright (c) 2018 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 "mkldnn.hpp"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/mul_op.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
namespace paddle {
namespace operators {
using paddle::framework::Tensor;
using paddle::platform::MKLDNNDeviceContext;
template <typename Format = mkldnn::memory::format>
mkldnn::memory::desc type(const std::vector<int>& dims, Format&& f) {
return platform::MKLDNNMemDesc(dims, mkldnn::memory::data_type::f32, f);
}
template <typename T>
class MulMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
auto mkldnn_engine = dev_ctx.GetEngine();
auto input = ctx.Input<Tensor>("X");
auto weight = ctx.Input<Tensor>("Y");
PADDLE_ENFORCE(input->dims().size() & (2 | 4),
"Input must be with 2 or 4 dimensions, i.e. NC or NCHW");
PADDLE_ENFORCE(weight->dims().size() & (2 | 4),
"Weights must be with 2 or 4 dimensions, i.e. OI or OIHW");
std::vector<int> w_tz = paddle::framework::vectorize2int(weight->dims());
std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
auto src_md =
src_tz.size() != 2
? type(src_tz, mkldnn::memory::format::nchw)
: type({src_tz[0], src_tz[1]}, mkldnn::memory::format::nc);
auto dst_md = type({src_tz[0], w_tz[1]}, mkldnn::memory::format::nc);
auto weights_md =
src_tz.size() != 2
? type({w_tz[1], src_tz[1], src_tz[2], src_tz[3]},
mkldnn::memory::format::oihw)
: type({w_tz[1], src_tz[1]}, mkldnn::memory::format::oi);
auto output = ctx.Output<Tensor>("Out");
T* output_data = output->mutable_data<T>(ctx.GetPlace());
const std::string key = ctx.op().Output("Out");
const std::string key_fc_pd = key + "@mul_pd";
const T* input_data = input->data<T>();
const T* w_data = weight->data<T>();
auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, output_data);
auto src_memory = mkldnn::memory({src_md, mkldnn_engine},
platform::to_void_cast(input_data));
auto weights_memory = mkldnn::memory({weights_md, mkldnn_engine},
platform::to_void_cast(w_data));
auto pd = platform::MKLDNNFwdPrimitiveDesc<mkldnn::inner_product_forward>(
mkldnn_engine, src_md, weights_md, dst_md);
dev_ctx.SetBlob(key_fc_pd, pd);
auto forward = mkldnn::inner_product_forward(*pd, src_memory,
weights_memory, dst_memory);
std::vector<mkldnn::primitive> pipeline = {forward};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
}
};
template <typename T>
class MulMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
auto mkldnn_engine = dev_ctx.GetEngine();
const Tensor* input = ctx.Input<Tensor>("X");
const Tensor* w = ctx.Input<Tensor>("Y");
const Tensor* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
Tensor* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
Tensor* w_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
const std::string key = ctx.op().Input("Out");
const std::string key_fc_pd = key + "@mul_pd";
const T* input_data = input->data<T>();
const T* w_data = w->data<T>();
const T* out_grad_data = out_grad->data<T>();
T* input_grad_data = nullptr;
T* w_grad_data = nullptr;
if (input_grad) {
input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
}
if (w_grad) {
w_grad_data = w_grad->mutable_data<T>(ctx.GetPlace());
}
std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
std::vector<int> w_tz = paddle::framework::vectorize2int(w->dims());
auto src_md =
src_tz.size() != 2
? type(src_tz, mkldnn::memory::format::nchw)
: type({src_tz[0], src_tz[1]}, mkldnn::memory::format::nc);
auto dst_md = type({src_tz[0], w_tz[1]}, mkldnn::memory::format::nc);
auto weights_md =
src_tz.size() != 2
? type({w_tz[1], src_tz[1], src_tz[2], src_tz[3]},
mkldnn::memory::format::oihw)
: type({w_tz[1], src_tz[1]}, mkldnn::memory::format::oi);
auto src_memory = mkldnn::memory({src_md, mkldnn_engine},
platform::to_void_cast(input_data));
auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine},
platform::to_void_cast(out_grad_data));
auto weight_memory = mkldnn::memory({weights_md, mkldnn_engine},
platform::to_void_cast(w_data));
auto pd =
std::static_pointer_cast<mkldnn::inner_product_forward::primitive_desc>(
dev_ctx.GetBlob(key_fc_pd));
PADDLE_ENFORCE(pd != nullptr, "Fail to find pd in device context");
if (w_grad) {
auto weights_grad_memory = mkldnn::memory(
{weights_md, mkldnn_engine}, platform::to_void_cast(w_grad_data));
auto bwd_weight_pd = platform::MKLDNNBwdPrimitiveDesc<
mkldnn::inner_product_backward_weights>(mkldnn_engine, *pd, src_md,
weights_md, dst_md);
auto bwd_weights_prim = mkldnn::inner_product_backward_weights(
bwd_weight_pd, src_memory, dst_memory, weights_grad_memory);
std::vector<mkldnn::primitive> pipeline{bwd_weights_prim};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
}
if (input_grad) {
auto src_grad_memory = mkldnn::memory(
{src_md, mkldnn_engine}, platform::to_void_cast(input_grad_data));
auto bwd_data_pd =
platform::MKLDNNBwdPrimitiveDesc<mkldnn::inner_product_backward_data>(
mkldnn_engine, *pd, src_md, weights_md, dst_md);
auto bwd_data_prim = mkldnn::inner_product_backward_data(
bwd_data_pd, dst_memory, weight_memory, src_grad_memory);
std::vector<mkldnn::primitive> pipeline{bwd_data_prim};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
}
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_KERNEL(mul, MKLDNN, ::paddle::platform::CPUPlace,
paddle::operators::MulMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(mul_grad, MKLDNN, ::paddle::platform::CPUPlace,
paddle::operators::MulMKLDNNGradOpKernel<float>);
...@@ -16,10 +16,6 @@ limitations under the License. */ ...@@ -16,10 +16,6 @@ limitations under the License. */
#include <string> #include <string>
#include <vector> #include <vector>
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle { namespace paddle {
namespace operators { namespace operators {
...@@ -76,22 +72,6 @@ class MulOp : public framework::OperatorWithKernel { ...@@ -76,22 +72,6 @@ class MulOp : public framework::OperatorWithKernel {
ctx->SetOutputDim("Out", framework::make_ddim(output_dims)); ctx->SetOutputDim("Out", framework::make_ddim(output_dims));
ctx->ShareLoD("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out");
} }
private:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
framework::LibraryType library{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_MKLDNN
if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library = framework::LibraryType::kMKLDNN;
}
#endif
framework::DataLayout layout{framework::DataLayout::kAnyLayout};
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout, library);
}
}; };
class MulOpMaker : public framework::OpProtoAndCheckerMaker { class MulOpMaker : public framework::OpProtoAndCheckerMaker {
...@@ -120,9 +100,6 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -120,9 +100,6 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker {
)DOC") )DOC")
.SetDefault(1) .SetDefault(1)
.EqualGreaterThan(1); .EqualGreaterThan(1);
AddAttr<bool>("use_mkldnn",
"(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddAttr<int>( AddAttr<int>(
"y_num_col_dims", "y_num_col_dims",
R"DOC((int, default 1), The mul_op can take tensors with more than two, R"DOC((int, default 1), The mul_op can take tensors with more than two,
...@@ -177,22 +154,6 @@ class MulGradOp : public framework::OperatorWithKernel { ...@@ -177,22 +154,6 @@ class MulGradOp : public framework::OperatorWithKernel {
ctx->SetOutputDim(y_grad_name, y_dims); ctx->SetOutputDim(y_grad_name, y_dims);
} }
} }
private:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
framework::LibraryType library{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_MKLDNN
if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library = framework::LibraryType::kMKLDNN;
}
#endif
framework::DataLayout layout{framework::DataLayout::kAnyLayout};
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace(),
layout, library);
}
}; };
} // namespace operators } // namespace operators
......
...@@ -179,11 +179,8 @@ def fc(input, ...@@ -179,11 +179,8 @@ def fc(input,
inputs={"X": input_var, inputs={"X": input_var,
"Y": w}, "Y": w},
outputs={"Out": tmp}, outputs={"Out": tmp},
attrs={ attrs={"x_num_col_dims": num_flatten_dims,
"x_num_col_dims": num_flatten_dims, "y_num_col_dims": 1})
"y_num_col_dims": 1,
"use_mkldnn": use_mkldnn
})
mul_results.append(tmp) mul_results.append(tmp)
if len(mul_results) == 1: if len(mul_results) == 1:
......
# Copyright (c) 2018 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.
import unittest
from test_mul_op import TestMulOp, TestMulOp2, TestFP16MulOp1, TestFP16MulOp2
class TestMKLDNNMulOp(TestMulOp):
def init_op_test(self):
super(TestMKLDNNMulOp, self).setUp()
self.attrs = {"use_mkldnn": True}
class TestMKLDNNMulOp2(TestMulOp2):
def init_op_test(self):
super(TestMKLDNNMulOp2, self).setUp()
self.attrs = {"use_mkldnn": True}
class TestMKLDNNFP16MulOp1(TestFP16MulOp1):
def init_op_test(self):
super(TestMKLDNNFP16MulOp1, self).setUp()
self.attrs = {"use_mkldnn": True}
class TestMKLDNNFP16MulOp2(TestFP16MulOp2):
def init_op_test(self):
super(TestMKLDNNFP16MulOp2, self).setUp()
self.attrs = {"use_mkldnn": True}
if __name__ == "__main__":
unittest.main()
...@@ -21,12 +21,10 @@ from op_test import OpTest ...@@ -21,12 +21,10 @@ from op_test import OpTest
class TestMulOp(OpTest): class TestMulOp(OpTest):
def setUp(self): def setUp(self):
self.op_type = "mul" self.op_type = "mul"
self.use_mkldnn = False
self.inputs = { self.inputs = {
'X': np.random.random((32, 84)).astype("float32"), 'X': np.random.random((32, 84)).astype("float32"),
'Y': np.random.random((84, 100)).astype("float32") 'Y': np.random.random((84, 100)).astype("float32")
} }
self.attrs = {'use_mkldnn': self.use_mkldnn}
self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
def test_check_output(self): def test_check_output(self):
...@@ -47,16 +45,11 @@ class TestMulOp(OpTest): ...@@ -47,16 +45,11 @@ class TestMulOp(OpTest):
class TestMulOp2(OpTest): class TestMulOp2(OpTest):
def setUp(self): def setUp(self):
self.op_type = "mul" self.op_type = "mul"
self.use_mkldnn = False
self.inputs = { self.inputs = {
'X': np.random.random((15, 4, 12, 10)).astype("float32"), 'X': np.random.random((15, 4, 12, 10)).astype("float32"),
'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32") 'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32")
} }
self.attrs = { self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2}
'x_num_col_dims': 2,
'y_num_col_dims': 2,
'use_mkldnn': self.use_mkldnn
}
result = np.dot(self.inputs['X'].reshape(15 * 4, 12 * 10), result = np.dot(self.inputs['X'].reshape(15 * 4, 12 * 10),
self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9)) self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9))
result = result.reshape(15, 4, 8, 2, 9) result = result.reshape(15, 4, 8, 2, 9)
...@@ -80,11 +73,9 @@ class TestMulOp2(OpTest): ...@@ -80,11 +73,9 @@ class TestMulOp2(OpTest):
class TestFP16MulOp1(OpTest): class TestFP16MulOp1(OpTest):
def setUp(self): def setUp(self):
self.op_type = "mul" self.op_type = "mul"
self.use_mkldnn = False
x = np.random.random((32, 84)).astype("float16") x = np.random.random((32, 84)).astype("float16")
y = np.random.random((84, 100)).astype("float16") y = np.random.random((84, 100)).astype("float16")
self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)}
self.attrs = {'use_mkldnn': self.use_mkldnn}
self.outputs = {'Out': np.dot(x, y)} self.outputs = {'Out': np.dot(x, y)}
def test_check_output(self): def test_check_output(self):
...@@ -97,15 +88,10 @@ class TestFP16MulOp1(OpTest): ...@@ -97,15 +88,10 @@ class TestFP16MulOp1(OpTest):
class TestFP16MulOp2(OpTest): class TestFP16MulOp2(OpTest):
def setUp(self): def setUp(self):
self.op_type = "mul" self.op_type = "mul"
self.use_mkldnn = False
x = np.random.random((15, 4, 12, 10)).astype("float16") x = np.random.random((15, 4, 12, 10)).astype("float16")
y = np.random.random((4, 30, 8, 2, 9)).astype("float16") y = np.random.random((4, 30, 8, 2, 9)).astype("float16")
self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)}
self.attrs = { self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2}
'x_num_col_dims': 2,
'y_num_col_dims': 2,
'use_mkldnn': self.use_mkldnn
}
result = np.dot( result = np.dot(
x.reshape(15 * 4, 12 * 10), y.reshape(4 * 30, 8 * 2 * 9)) x.reshape(15 * 4, 12 * 10), y.reshape(4 * 30, 8 * 2 * 9))
result = result.reshape(15, 4, 8, 2, 9) result = result.reshape(15, 4, 8, 2, 9)
......
...@@ -63,10 +63,7 @@ class TestOperator(unittest.TestCase): ...@@ -63,10 +63,7 @@ class TestOperator(unittest.TestCase):
self.assertEqual(mul_op.output("Out"), ["mul.out"]) self.assertEqual(mul_op.output("Out"), ["mul.out"])
self.assertEqual( self.assertEqual(
set(mul_op.attr_names), set(mul_op.attr_names),
set([ set(["x_num_col_dims", "y_num_col_dims", "op_role", "op_role_var"]))
"x_num_col_dims", "y_num_col_dims", "use_mkldnn", "op_role",
"op_role_var"
]))
self.assertEqual(mul_op.has_attr("x_num_col_dims"), True) self.assertEqual(mul_op.has_attr("x_num_col_dims"), True)
self.assertEqual(mul_op.attr_type("x_num_col_dims"), core.AttrType.INT) self.assertEqual(mul_op.attr_type("x_num_col_dims"), core.AttrType.INT)
self.assertEqual(mul_op.attr("x_num_col_dims"), 1) self.assertEqual(mul_op.attr("x_num_col_dims"), 1)
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册