support Baidu Kunlun AI Accelerator (#25959)
* support Baidu AI Accelerator * test=kunlun * minor * test=kunlun * support xpu op in separate file * test=kunlun * update XPU error message and remove duplicated code * test=kunlun * minor * test=kunlun * minor * test=kunluntest_feature_precision_test_c
QingshuChen
6 years ago
committed by
GitHub
parent
17bcaef411
commit
138ecf24aa
@ -0,0 +1,54 @@
|
||||
if (NOT WITH_XPU)
|
||||
return()
|
||||
endif()
|
||||
|
||||
INCLUDE(ExternalProject)
|
||||
SET(XPU_PROJECT "extern_xpu")
|
||||
SET(XPU_URL "https://kunlun1.su.bcebos.com/xpu.tar.gz" CACHE STRING "" FORCE)
|
||||
SET(XPU_SOURCE_DIR "${THIRD_PARTY_PATH}/xpu")
|
||||
SET(XPU_DOWNLOAD_DIR "${XPU_SOURCE_DIR}/src/${XPU_PROJECT}")
|
||||
SET(XPU_INSTALL_DIR "${THIRD_PARTY_PATH}/install/xpu")
|
||||
SET(XPU_API_INC_DIR "${THIRD_PARTY_PATH}/install/xpu/api/include")
|
||||
SET(XPU_RUNTIME_INC_DIR "${THIRD_PARTY_PATH}/install/xpu/runtime/include")
|
||||
SET(XPU_LIB_DIR "${THIRD_PARTY_PATH}/install/xpu/lib")
|
||||
|
||||
SET(XPU_API_LIB_NAME "libxpuapi.so")
|
||||
SET(XPU_RT_LIB_NAME "libxpurt.so")
|
||||
SET(XPU_SIM_LIB_NAME "libxpusim.so")
|
||||
SET(XPU_API_LIB "${XPU_LIB_DIR}/${XPU_API_LIB_NAME}")
|
||||
SET(XPU_RT_LIB "${XPU_LIB_DIR}/${XPU_RT_LIB_NAME}")
|
||||
SET(XPU_SIM_LIB "${XPU_LIB_DIR}/${XPU_SIM_LIB_NAME}")
|
||||
|
||||
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${XPU_INSTALL_DIR}/lib")
|
||||
|
||||
INCLUDE_DIRECTORIES(${XPU_API_INC_DIR})
|
||||
INCLUDE_DIRECTORIES(${XPU_RUNTIME_INC_DIR})
|
||||
|
||||
FILE(WRITE ${XPU_DOWNLOAD_DIR}/CMakeLists.txt
|
||||
"PROJECT(XPU)\n"
|
||||
"cmake_minimum_required(VERSION 3.0)\n"
|
||||
"install(DIRECTORY xpu/api xpu/runtime xpu/lib \n"
|
||||
" DESTINATION ${XPU_INSTALL_DIR})\n")
|
||||
|
||||
ExternalProject_Add(
|
||||
${XPU_PROJECT}
|
||||
${EXTERNAL_PROJECT_LOG_ARGS}
|
||||
PREFIX ${XPU_SOURCE_DIR}
|
||||
DOWNLOAD_DIR ${XPU_DOWNLOAD_DIR}
|
||||
DOWNLOAD_COMMAND wget --no-check-certificate ${XPU_URL} -c -q -O xpu.tar.gz
|
||||
&& tar xvf xpu.tar.gz
|
||||
DOWNLOAD_NO_PROGRESS 1
|
||||
UPDATE_COMMAND ""
|
||||
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XPU_INSTALL_ROOT}
|
||||
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${XPU_INSTALL_ROOT}
|
||||
)
|
||||
|
||||
ADD_LIBRARY(shared_xpuapi SHARED IMPORTED GLOBAL)
|
||||
set_property(TARGET shared_xpuapi PROPERTY IMPORTED_LOCATION "${XPU_API_LIB}")
|
||||
|
||||
# generate a static dummy target to track xpulib dependencies
|
||||
# for cc_library(xxx SRCS xxx.c DEPS xpulib)
|
||||
generate_dummy_static_lib(LIB_NAME "xpulib" GENERATOR "xpu.cmake")
|
||||
|
||||
TARGET_LINK_LIBRARIES(xpulib ${XPU_API_LIB} ${XPU_RT_LIB} ${XPU_SIM_LIB})
|
||||
ADD_DEPENDENCIES(xpulib ${XPU_PROJECT})
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,183 @@
|
||||
/* Copyright (c) 2020 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 <memory>
|
||||
#include <string>
|
||||
#include <unordered_map>
|
||||
#include <vector>
|
||||
#include "paddle/fluid/operators/mul_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using framework::OpKernelType;
|
||||
using framework::Tensor;
|
||||
|
||||
template <typename DeviceContext, typename T>
|
||||
class MulXPUKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& context) const override {
|
||||
const Tensor* x = context.Input<Tensor>("X");
|
||||
const Tensor* y = context.Input<Tensor>("Y");
|
||||
Tensor* z = context.Output<Tensor>("Out");
|
||||
const Tensor x_matrix =
|
||||
x->dims().size() > 2
|
||||
? framework::ReshapeToMatrix(
|
||||
*x, context.template Attr<int>("x_num_col_dims"))
|
||||
: *x;
|
||||
const Tensor y_matrix =
|
||||
y->dims().size() > 2
|
||||
? framework::ReshapeToMatrix(
|
||||
*y, context.template Attr<int>("y_num_col_dims"))
|
||||
: *y;
|
||||
z->mutable_data<T>(context.GetPlace());
|
||||
auto z_dim = z->dims();
|
||||
if (z_dim.size() != 2) {
|
||||
z->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
|
||||
}
|
||||
bool trans_a = false;
|
||||
bool trans_b = false;
|
||||
int m = x_matrix.dims()[0];
|
||||
int k = x_matrix.dims()[1];
|
||||
int k1 = y_matrix.dims()[0];
|
||||
int n = y_matrix.dims()[1];
|
||||
PADDLE_ENFORCE_EQ(
|
||||
k, k1, platform::errors::InvalidArgument("Shape mistake in mul_op"));
|
||||
T alpha = static_cast<T>(1.0);
|
||||
T beta = static_cast<T>(0.0);
|
||||
const T* data_a = x_matrix.data<T>();
|
||||
const T* data_b = y_matrix.data<T>();
|
||||
T* data_c = z->data<T>();
|
||||
auto& dev_ctx = context.template device_context<DeviceContext>();
|
||||
int ret = xpu::fc_int16(dev_ctx.x_context(), trans_a, trans_b, m, n, k,
|
||||
alpha, data_a, data_b, beta, data_c);
|
||||
PADDLE_ENFORCE_EQ(
|
||||
ret, XPU_SUCCESS,
|
||||
platform::errors::External(
|
||||
"XPU API return wrong value[%d], please check whether "
|
||||
"Baidu Kunlun Card is properly installed.",
|
||||
ret));
|
||||
if (z_dim.size() != 2) {
|
||||
z->Resize(z_dim);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template <typename DeviceContext, typename T>
|
||||
class MulGradXPUKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
int x_num_col_dims = ctx.template Attr<int>("x_num_col_dims");
|
||||
int y_num_col_dims = ctx.template Attr<int>("y_num_col_dims");
|
||||
auto* x = ctx.Input<framework::LoDTensor>("X");
|
||||
auto* y = ctx.Input<framework::LoDTensor>("Y");
|
||||
auto x_matrix = x->dims().size() > 2
|
||||
? framework::ReshapeToMatrix(*x, x_num_col_dims)
|
||||
: static_cast<const Tensor&>(*x);
|
||||
auto y_matrix = y->dims().size() > 2
|
||||
? framework::ReshapeToMatrix(*y, y_num_col_dims)
|
||||
: static_cast<const Tensor&>(*y);
|
||||
auto* dout = ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
|
||||
Tensor dout_mat;
|
||||
dout_mat.Resize({framework::flatten_to_2d(x->dims(), x_num_col_dims)[0],
|
||||
framework::flatten_to_2d(y->dims(), y_num_col_dims)[1]});
|
||||
auto* dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
|
||||
auto* dy = ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
|
||||
if (dx != nullptr) {
|
||||
dx->set_lod(x->lod());
|
||||
}
|
||||
if (dy != nullptr) {
|
||||
dy->set_lod(y->lod());
|
||||
}
|
||||
auto& dev_ctx = ctx.template device_context<DeviceContext>();
|
||||
if (dx) {
|
||||
dx->mutable_data<T>(ctx.GetPlace());
|
||||
Tensor dx_matrix = dx->dims().size() > 2
|
||||
? framework::ReshapeToMatrix(*dx, x_num_col_dims)
|
||||
: *dx;
|
||||
// dx = dout * y'. dx: M x K, dout : M x N, y : K x N
|
||||
// blas.MatMul(dout_mat, false, y_matrix, true, &dx_matrix);
|
||||
bool trans_a = false;
|
||||
bool trans_b = true;
|
||||
int m = dout_mat.dims()[0];
|
||||
int k = dout_mat.dims()[1];
|
||||
int n = y_matrix.dims()[0];
|
||||
int k1 = y_matrix.dims()[1];
|
||||
PADDLE_ENFORCE_EQ(
|
||||
k, k1, platform::errors::InvalidArgument("Shape mistake in mul_op"));
|
||||
int lda = (!trans_a) ? k : m;
|
||||
int ldb = (!trans_b) ? n : k;
|
||||
int ldc = n;
|
||||
T alpha = static_cast<T>(1.0);
|
||||
T beta = static_cast<T>(0.0);
|
||||
const T* data_a = dout->data<T>();
|
||||
const T* data_b = y_matrix.data<T>();
|
||||
T* data_c = dx_matrix.data<T>();
|
||||
int ret =
|
||||
xpu::gemm_int16(dev_ctx.x_context(), trans_a, trans_b, m, n, k, alpha,
|
||||
data_a, lda, data_b, ldb, beta, data_c, ldc);
|
||||
PADDLE_ENFORCE_EQ(ret, XPU_SUCCESS,
|
||||
platform::errors::External(
|
||||
"XPU API return wrong value[%d], please check "
|
||||
"where Baidu Kunlun Card is properly installed.",
|
||||
ret));
|
||||
}
|
||||
|
||||
if (dy) {
|
||||
dy->mutable_data<T>(ctx.GetPlace());
|
||||
Tensor dy_matrix = dy->dims().size() > 2
|
||||
? framework::ReshapeToMatrix(*dy, y_num_col_dims)
|
||||
: *dy;
|
||||
// dy = x' * dout. dy K x N, dout : M x N, x : M x K
|
||||
// blas.MatMul(x_matrix, true, dout_mat, false, &dy_matrix);
|
||||
bool trans_a = true;
|
||||
bool trans_b = false;
|
||||
int k = x_matrix.dims()[0];
|
||||
int m = x_matrix.dims()[1];
|
||||
int k1 = dout_mat.dims()[0];
|
||||
int n = dout_mat.dims()[1];
|
||||
PADDLE_ENFORCE_EQ(
|
||||
k, k1, platform::errors::InvalidArgument("Shape mistake in mul_op"));
|
||||
int lda = (!trans_a) ? k : m;
|
||||
int ldb = (!trans_b) ? n : k;
|
||||
int ldc = n;
|
||||
T alpha = static_cast<T>(1.0);
|
||||
T beta = static_cast<T>(0.0);
|
||||
const T* data_a = x_matrix.data<T>();
|
||||
const T* data_b = dout->data<T>();
|
||||
T* data_c = dy_matrix.data<T>();
|
||||
int ret =
|
||||
xpu::gemm_int16(dev_ctx.x_context(), trans_a, trans_b, m, n, k, alpha,
|
||||
data_a, lda, data_b, ldb, beta, data_c, ldc);
|
||||
PADDLE_ENFORCE_EQ(ret, XPU_SUCCESS,
|
||||
platform::errors::External(
|
||||
"XPU API return wrong value[%d], please check "
|
||||
"where Baidu Kunlun Card is properly installed.",
|
||||
ret));
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
|
||||
REGISTER_OP_XPU_KERNEL(
|
||||
mul, ops::MulXPUKernel<paddle::platform::XPUDeviceContext, float>);
|
||||
REGISTER_OP_XPU_KERNEL(
|
||||
mul_grad, ops::MulGradXPUKernel<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