Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 0146e8bf 编写于 作者: H hedaoyuan 提交者： GitHub
浏览文件
操作

Merge pull request #3843 from Xreki/build_android_clang 

Enable the use of clang to cross-compile for Android
上级 0140f3f9 3360e9cd
隐藏空白更改
内联 并排
Showing with 351 addition and 229 deletion
Dockerfile.android
浏览文件 @ 0146e8bf
...@@ -4,9 +4,15 @@ MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com> ...@@ -4,9 +4,15 @@ MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
ARG UBUNTU_MIRROR ARG UBUNTU_MIRROR
RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi' RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
# ENV variables
ARG ANDROID_ABI
ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
ENV HOME=/root \ ENV HOME=/root \
ANDROID_NDK_HOME=/opt/android-ndk-linux \ ANDROID_NDK_HOME=/opt/android-ndk-linux \
ANDROID_STANDALONE_TOOLCHAIN=/opt/android-toolchain-gcc ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \
ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain
RUN apt-get update && \ RUN apt-get update && \
apt-get install -y \ apt-get install -y \
...@@ -15,12 +21,11 @@ RUN apt-get update && \ ...@@ -15,12 +21,11 @@ RUN apt-get update && \
apt-get clean -y apt-get clean -y
# Install Go and glide # Install Go and glide
RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \ RUN wget -qO- go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
tar -C /usr/local -xzf go.tgz && \ tar -xz -C /usr/local && \
mkdir /root/gopath && \ mkdir /root/gopath && \
mkdir /root/gopath/bin && \ mkdir /root/gopath/bin && \
mkdir /root/gopath/src && \ mkdir /root/gopath/src
rm go.tgz
ENV GOROOT=/usr/local/go GOPATH=/root/gopath ENV GOROOT=/usr/local/go GOPATH=/root/gopath
# should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT. # should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT.
ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin
...@@ -42,7 +47,8 @@ RUN mkdir /opt/android-ndk-tmp && \ ...@@ -42,7 +47,8 @@ RUN mkdir /opt/android-ndk-tmp && \
wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \ wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
unzip -q android-ndk-r14b-linux-x86_64.zip && \ unzip -q android-ndk-r14b-linux-x86_64.zip && \
mv android-ndk-r14b ${ANDROID_NDK_HOME} && \ mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-21 --install-dir=${ANDROID_STANDALONE_TOOLCHAIN} && \ ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \
${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \
rm -rf /opt/android-ndk-tmp && \ rm -rf /opt/android-ndk-tmp && \
rm -rf ${ANDROID_NDK_HOME} rm -rf ${ANDROID_NDK_HOME}
......
cmake/cross_compiling/android.cmake
浏览文件 @ 0146e8bf
...@@ -20,6 +20,7 @@ ...@@ -20,6 +20,7 @@
# The supported variables are listed belows: # The supported variables are listed belows:
# #
# ANDROID_STANDALONE_TOOLCHAIN # ANDROID_STANDALONE_TOOLCHAIN
# ANDROID_TOOLCHAIN
# ANDROID_ABI # ANDROID_ABI
# ANDROID_NATIVE_API_LEVEL # ANDROID_NATIVE_API_LEVEL
# ANDROID_ARM_MODE # ANDROID_ARM_MODE
...@@ -57,6 +58,10 @@ IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL) ...@@ -57,6 +58,10 @@ IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL)
ENDIF() ENDIF()
ENDIF() ENDIF()
IF(NOT DEFINED ANDROID_TOOLCHAIN)
SET(ANDROID_TOOLCHAIN clang)
ENDIF()
IF(NOT DEFINED ANDROID_ABI) IF(NOT DEFINED ANDROID_ABI)
SET(ANDROID_ABI "armeabi-v7a") SET(ANDROID_ABI "armeabi-v7a")
ENDIF() ENDIF()
...@@ -82,6 +87,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ...@@ -82,6 +87,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
"${CMAKE_VERSION}), when cross-compiling for Android.") "${CMAKE_VERSION}), when cross-compiling for Android.")
IF(ANDROID_STANDALONE_TOOLCHAIN) IF(ANDROID_STANDALONE_TOOLCHAIN)
# Use standalone toolchain
SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot") SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot")
IF(NOT CMAKE_SYSTEM_VERSION) IF(NOT CMAKE_SYSTEM_VERSION)
...@@ -96,26 +102,44 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ...@@ -96,26 +102,44 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
ENDIF() ENDIF()
# Toolchain # Toolchain
SET(ANDROID_TOOLCHAIN "gcc")
SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN}) SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN})
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") ELSE(ANDROID_NDK)
SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi) # TODO: use android ndk
IF(ANDROID_ABI STREQUAL "armeabi") ENDIF()
SET(CMAKE_SYSTEM_PROCESSOR armv5te)
ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a") IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
SET(CMAKE_SYSTEM_PROCESSOR armv7-a) SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
ENDIF() IF(ANDROID_ABI STREQUAL "armeabi")
ENDIF() SET(CMAKE_SYSTEM_PROCESSOR armv5te)
IF(ANDROID_ABI STREQUAL "arm64-v8a") SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi)
SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android) ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
SET(CMAKE_SYSTEM_PROCESSOR aarch64) SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi)
ENDIF() ENDIF()
SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-") ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
SET(CMAKE_SYSTEM_PROCESSOR aarch64)
SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android)
ELSE()
MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.")
ENDIF()
SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
IF(ANDROID_TOOLCHAIN STREQUAL clang)
SET(ANDROID_C_COMPILER_NAME clang)
SET(ANDROID_CXX_COMPILER_NAME clang++)
SET(CMAKE_C_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc)
SET(ANDROID_C_COMPILER_NAME gcc)
SET(ANDROID_CXX_COMPILER_NAME g++)
ELSE()
MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}")
ENDIF() ENDIF()
# C compiler # C compiler
IF(NOT CMAKE_C_COMPILER) IF(NOT CMAKE_C_COMPILER)
SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}gcc") SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}")
ELSE() ELSE()
GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM) GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
ENDIF() ENDIF()
...@@ -125,7 +149,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ...@@ -125,7 +149,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
# CXX compiler # CXX compiler
IF(NOT CMAKE_CXX_COMPILER) IF(NOT CMAKE_CXX_COMPILER)
SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}g++") SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}")
ELSE() ELSE()
GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM) GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
ENDIF() ENDIF()
...@@ -137,7 +161,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ...@@ -137,7 +161,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE) SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
# Toolchain and ABI specific flags. # Toolchain and ABI specific flags.
SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections -finline-limit=64") SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections")
SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections") SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections")
IF(ANDROID_ABI STREQUAL "armeabi") IF(ANDROID_ABI STREQUAL "armeabi")
...@@ -145,8 +169,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ...@@ -145,8 +169,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
-march=armv5te -march=armv5te
-mtune=xscale -mtune=xscale
-msoft-float) -msoft-float)
ENDIF() ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
IF(ANDROID_ABI STREQUAL "armeabi-v7a")
LIST(APPEND ANDROID_COMPILER_FLAGS LIST(APPEND ANDROID_COMPILER_FLAGS
-march=armv7-a -march=armv7-a
-mfloat-abi=softfp) -mfloat-abi=softfp)
...@@ -156,6 +179,8 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ...@@ -156,6 +179,8 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16) LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16)
ENDIF() ENDIF()
LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8) LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8)
ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
ENDIF() ENDIF()
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
...@@ -164,10 +189,18 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ...@@ -164,10 +189,18 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
ELSE() ELSE()
LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb) LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb)
ENDIF() ENDIF()
IF(ANDROID_TOOLCHAIN STREQUAL clang)
# Disable integrated-as for better compatibility.
LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as)
ENDIF()
ENDIF() ENDIF()
IF(ANDROID_ABI STREQUAL "arm64-v8a") IF(ANDROID_TOOLCHAIN STREQUAL clang)
LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a) # CMake automatically forwards all compiler flags to the linker,
# and clang doesn't like having -Wa flags being used for linking.
# To prevent CMake from doing this would require meddling with
# the CMAKE_<LANG>_COMPILE_OBJECT rules, which would get quite messy.
LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments)
ENDIF() ENDIF()
STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}") STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")
......
cmake/external/openblas.cmake
浏览文件 @ 0146e8bf
...@@ -12,6 +12,10 @@ ...@@ -12,6 +12,10 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
IF(USE_EIGEN_FOR_BLAS)
return()
ENDIF(USE_EIGEN_FOR_BLAS)
INCLUDE(cblas) INCLUDE(cblas)
IF(NOT ${CBLAS_FOUND}) IF(NOT ${CBLAS_FOUND})
......
paddle/cuda/include/hl_cpu_gru.cuh
浏览文件 @ 0146e8bf
...@@ -18,14 +18,6 @@ limitations under the License. */ ...@@ -18,14 +18,6 @@ limitations under the License. */
#ifndef __NVCC__ #ifndef __NVCC__
#include "paddle/math/MathFunctions.h"
#ifndef PADDLE_TYPE_DOUBLE
#define CBLAS_GEMM paddle::gemm<float>
#else
#define CBLAS_GEMM paddle::gemm<double>
#endif
template<class OpResetOutput> template<class OpResetOutput>
void hl_naive_gru_forward_reset_output(OpResetOutput opResetOutput, void hl_naive_gru_forward_reset_output(OpResetOutput opResetOutput,
real *gateValue, real *gateValue,
...@@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput, ...@@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput,
} }
} }
template<class OpResetOutput, class OpFinalOutput>
void hl_cpu_gru_forward(OpResetOutput opResetOutput,
OpFinalOutput opFinalOutput,
hl_gru_value value,
int frameSize,
int batchSize,
hl_activation_mode_t active_node,
hl_activation_mode_t active_gate) {
if (value.prevOutValue) {
CBLAS_GEMM(CblasNoTrans,
CblasNoTrans,
batchSize,
2 * frameSize,
frameSize,
1,
value.prevOutValue,
frameSize,
value.gateWeight,
frameSize * 2,
1,
value.gateValue,
frameSize * 3);
}
forward_reset_output(opResetOutput, value, frameSize, batchSize, active_gate);
if (value.prevOutValue) {
CBLAS_GEMM(CblasNoTrans,
CblasNoTrans,
batchSize,
frameSize,
frameSize,
1,
value.resetOutputValue,
frameSize,
value.stateWeight,
frameSize,
1,
value.gateValue + frameSize * 2,
frameSize * 3);
}
forward_final_output(opFinalOutput, value, frameSize, batchSize, active_node);
}
template<class OpStateGrad> template<class OpStateGrad>
void hl_naive_gru_backward_state_grad(OpStateGrad opStateGrad, void hl_naive_gru_backward_state_grad(OpStateGrad opStateGrad,
real *gateValue, real *gateValue,
...@@ -525,86 +472,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad, ...@@ -525,86 +472,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad,
} }
} }
template<class OpStateGrad, class OpResetGrad>
void hl_cpu_gru_backward(OpStateGrad opStateGrad,
OpResetGrad opResetGrad,
hl_gru_value value,
hl_gru_grad grad,
int frameSize,
int batchSize,
hl_activation_mode_t active_node,
hl_activation_mode_t active_gate) {
backward_state_grad(opStateGrad, value, grad,
frameSize, batchSize, active_node);
if (value.prevOutValue && grad.prevOutGrad) {
CBLAS_GEMM(CblasNoTrans,
CblasTrans,
batchSize,
frameSize,
frameSize,
1,
grad.gateGrad + frameSize * 2,
frameSize * 3,
value.stateWeight,
frameSize,
0,
grad.resetOutputGrad,
frameSize);
if (grad.stateWeightGrad) {
CBLAS_GEMM(CblasTrans,
CblasNoTrans,
frameSize,
frameSize,
batchSize,
1,
value.resetOutputValue,
frameSize,
grad.gateGrad + frameSize * 2,
frameSize * 3,
1,
grad.stateWeightGrad,
frameSize);
}
}
backward_reset_grad(opResetGrad, value, grad,
frameSize, batchSize, active_gate);
if (grad.prevOutGrad && value.prevOutValue) {
CBLAS_GEMM(CblasNoTrans,
CblasTrans,
batchSize,
frameSize,
frameSize * 2,
1,
grad.gateGrad,
frameSize * 3,
value.gateWeight,
frameSize * 2,
1,
grad.prevOutGrad,
frameSize);
if (grad.gateWeightGrad) {
CBLAS_GEMM(CblasTrans,
CblasNoTrans,
frameSize,
frameSize * 2,
batchSize,
1,
value.prevOutValue,
frameSize,
grad.gateGrad,
frameSize * 3,
1,
grad.gateWeightGrad,
frameSize * 2);
}
}
}
#endif #endif
#endif // HL_CPU_GRU_CUH_ #endif // HL_CPU_GRU_CUH_
paddle/function/EigenGemm.cpp
浏览文件 @ 0146e8bf
...@@ -83,9 +83,9 @@ struct EigenBlasGemm { ...@@ -83,9 +83,9 @@ struct EigenBlasGemm {
}; };
#ifdef PADDLE_TYPE_DOUBLE #ifdef PADDLE_TYPE_DOUBLE
template class EigenBlasGemm<double>; template struct EigenBlasGemm<double>;
#else #else
template class EigenBlasGemm<float>; template struct EigenBlasGemm<float>;
#endif #endif
} // namespace paddle } // namespace paddle
paddle/function/GruFunctor.h 0 → 100644
浏览文件 @ 0146e8bf
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#pragma once
#include "GemmFunctor.h"
#include "hl_cpu_gru.cuh"
namespace paddle {
template <DeviceType Device, class T>
struct GruFunctor {
template <class OpResetOutput, class OpFinalOutput>
static void compute(OpResetOutput opResetOutput,
OpFinalOutput opFinalOutput,
hl_gru_value value,
int frameSize,
int batchSize,
hl_activation_mode_t active_node,
hl_activation_mode_t active_gate) {
#ifndef __NVCC__
if (value.prevOutValue) {
BlasGemm<Device, T>::compute(false,
false,
batchSize,
2 * frameSize,
frameSize,
1,
value.prevOutValue,
frameSize,
value.gateWeight,
frameSize * 2,
1,
value.gateValue,
frameSize * 3);
}
forward_reset_output(
opResetOutput, value, frameSize, batchSize, active_gate);
if (value.prevOutValue) {
BlasGemm<Device, T>::compute(false,
false,
batchSize,
frameSize,
frameSize,
1,
value.resetOutputValue,
frameSize,
value.stateWeight,
frameSize,
1,
value.gateValue + frameSize * 2,
frameSize * 3);
}
forward_final_output(
opFinalOutput, value, frameSize, batchSize, active_node);
#endif
}
};
template <DeviceType Device, class T>
struct GruGradFunctor {
template <class OpStateGrad, class OpResetGrad>
static void compute(OpStateGrad opStateGrad,
OpResetGrad opResetGrad,
hl_gru_value value,
hl_gru_grad grad,
int frameSize,
int batchSize,
hl_activation_mode_t active_node,
hl_activation_mode_t active_gate) {
#ifndef __NVCC__
backward_state_grad(
opStateGrad, value, grad, frameSize, batchSize, active_node);
if (value.prevOutValue && grad.prevOutGrad) {
BlasGemm<Device, T>::compute(false,
true,
batchSize,
frameSize,
frameSize,
1,
grad.gateGrad + frameSize * 2,
frameSize * 3,
value.stateWeight,
frameSize,
0,
grad.resetOutputGrad,
frameSize);
if (grad.stateWeightGrad) {
BlasGemm<Device, T>::compute(true,
false,
frameSize,
frameSize,
batchSize,
1,
value.resetOutputValue,
frameSize,
grad.gateGrad + frameSize * 2,
frameSize * 3,
1,
grad.stateWeightGrad,
frameSize);
}
}
backward_reset_grad(
opResetGrad, value, grad, frameSize, batchSize, active_gate);
if (grad.prevOutGrad && value.prevOutValue) {
BlasGemm<Device, T>::compute(false,
true,
batchSize,
frameSize,
frameSize * 2,
1,
grad.gateGrad,
frameSize * 3,
value.gateWeight,
frameSize * 2,
1,
grad.prevOutGrad,
frameSize);
if (grad.gateWeightGrad) {
BlasGemm<Device, T>::compute(true,
false,
frameSize,
frameSize * 2,
batchSize,
1,
value.prevOutValue,
frameSize,
grad.gateGrad,
frameSize * 3,
1,
grad.gateWeightGrad,
frameSize * 2);
}
}
#endif
}
};
} // namespace paddle
paddle/function/MulOp.cpp
浏览文件 @ 0146e8bf
...@@ -13,18 +13,10 @@ See the License for the specific language governing permissions and ...@@ -13,18 +13,10 @@ See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#include "MulOp.h" #include "MulOp.h"
/// todo(tianbing), delete it #include "GemmFunctor.h"
#include <iostream>
#include "paddle/math/MathFunctions.h"
#include "paddle/math/SIMDFunctions.h" #include "paddle/math/SIMDFunctions.h"
#include "paddle/utils/ThreadLocal.h" #include "paddle/utils/ThreadLocal.h"
#ifndef PADDLE_TYPE_DOUBLE
#define GEMM paddle::gemm<float>
#else
#define GEMM paddle::gemm<double>
#endif
namespace { namespace {
inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) { inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) {
for (unsigned int i = 0; i < len; ++i) { for (unsigned int i = 0; i < len; ++i) {
...@@ -114,19 +106,20 @@ void MulOp<DEVICE_TYPE_CPU>(CpuMatrix& out, ...@@ -114,19 +106,20 @@ void MulOp<DEVICE_TYPE_CPU>(CpuMatrix& out,
real scaleT, real scaleT,
bool aTrans, bool aTrans,
bool bTrans) { bool bTrans) {
GEMM(aTrans ? CblasTrans : CblasNoTrans, BlasGemm<DEVICE_TYPE_CPU, real>::compute(
bTrans ? CblasTrans : CblasNoTrans, aTrans,
out.getHeight(), bTrans,
out.getWidth(), out.getHeight(),
!aTrans ? a.getWidth() : a.getHeight(), out.getWidth(),
scaleAB, !aTrans ? a.getWidth() : a.getHeight(),
a.getData(), scaleAB,
a.getStride(), a.getData(),
b.getData(), a.getStride(),
b.getStride(), b.getData(),
scaleT, b.getStride(),
out.getData(), scaleT,
out.getStride()); out.getData(),
out.getStride());
} }
/// dense matrix (+)= sparse matrix * dense matrix /// dense matrix (+)= sparse matrix * dense matrix
......
paddle/function/neon/neon_util.h
浏览文件 @ 0146e8bf
...@@ -33,12 +33,8 @@ inline float32_t vaddvq_f32(float32x4_t a) { ...@@ -33,12 +33,8 @@ inline float32_t vaddvq_f32(float32x4_t a) {
return vget_lane_f32(vpadd_f32(v, v), 0); return vget_lane_f32(vpadd_f32(v, v), 0);
} }
inline float32x4_t vmlaq_laneq_f32(float32x4_t a, #define vmlaq_laneq_f32(a, b, v, lane) \
float32x4_t b, vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane))
float32x4_t v,
const int lane) {
return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane));
}
#endif #endif
} // namespace neon } // namespace neon
......
paddle/gserver/layers/GruCompute.cpp
浏览文件 @ 0146e8bf
...@@ -14,6 +14,7 @@ limitations under the License. */ ...@@ -14,6 +14,7 @@ limitations under the License. */
#include "GruCompute.h" #include "GruCompute.h"
#include "hl_recurrent_apply.cuh" #include "hl_recurrent_apply.cuh"
#include "paddle/function/GruFunctor.h"
#include "paddle/utils/Util.h" #include "paddle/utils/Util.h"
namespace paddle { namespace paddle {
...@@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) { ...@@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) {
template <> template <>
void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) { void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) {
hl_cpu_gru_forward(hppl::forward::gru_resetOutput(), GruFunctor<DEVICE_TYPE_CPU, real>::compute(hppl::forward::gru_resetOutput(),
hppl::forward::gru_finalOutput(), hppl::forward::gru_finalOutput(),
value, value,
frameSize, frameSize,
batchSize, batchSize,
activeNode_, activeNode_,
activeGate_); activeGate_);
} }
template <> template <>
...@@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value, ...@@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value,
hl_gru_grad grad, hl_gru_grad grad,
int frameSize, int frameSize,
int batchSize) { int batchSize) {
hl_cpu_gru_backward(hppl::backward::gru_stateGrad(), GruGradFunctor<DEVICE_TYPE_CPU, real>::compute(
hppl::backward::gru_resetGrad(), hppl::backward::gru_stateGrad(),
value, hppl::backward::gru_resetGrad(),
grad, value,
frameSize, grad,
batchSize, frameSize,
activeNode_, batchSize,
activeGate_); activeNode_,
activeGate_);
} }
} // namespace paddle } // namespace paddle
paddle/math/MathFunctions.cpp
浏览文件 @ 0146e8bf
...@@ -84,6 +84,7 @@ LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP) ...@@ -84,6 +84,7 @@ LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP)
namespace paddle { namespace paddle {
#ifndef PADDLE_USE_EIGEN_FOR_BLAS
template <> template <>
void gemm<float>(const CBLAS_TRANSPOSE transA, void gemm<float>(const CBLAS_TRANSPOSE transA,
const CBLAS_TRANSPOSE transB, const CBLAS_TRANSPOSE transB,
...@@ -143,6 +144,7 @@ void gemm<double>(const CBLAS_TRANSPOSE transA, ...@@ -143,6 +144,7 @@ void gemm<double>(const CBLAS_TRANSPOSE transA,
C, C,
ldc); ldc);
} }
#endif
template <> template <>
int getrf<float>(const CBLAS_ORDER order, int getrf<float>(const CBLAS_ORDER order,
...@@ -182,6 +184,7 @@ int getri<double>(const CBLAS_ORDER order, ...@@ -182,6 +184,7 @@ int getri<double>(const CBLAS_ORDER order,
return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv); return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv);
} }
#ifndef PADDLE_USE_EIGEN_FOR_BLAS
template <> template <>
void axpy<float>(const int n, const float alpha, const float* x, float* y) { void axpy<float>(const int n, const float alpha, const float* x, float* y) {
cblas_saxpy(n, alpha, x, 1, y, 1); cblas_saxpy(n, alpha, x, 1, y, 1);
...@@ -201,6 +204,7 @@ template <> ...@@ -201,6 +204,7 @@ template <>
double dotProduct<double>(const int n, const double* x, const double* y) { double dotProduct<double>(const int n, const double* x, const double* y) {
return cblas_ddot(n, x, 1, y, 1); return cblas_ddot(n, x, 1, y, 1);
} }
#endif
#if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML) #if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML)
......
paddle/math/MathFunctions.h
浏览文件 @ 0146e8bf
...@@ -40,7 +40,14 @@ extern "C" { ...@@ -40,7 +40,14 @@ extern "C" {
#ifndef LAPACK_FOUND #ifndef LAPACK_FOUND
extern "C" { extern "C" {
#ifndef PADDLE_USE_EIGEN_FOR_BLAS
#include <cblas.h> #include <cblas.h>
#else
typedef enum CBLAS_ORDER {
CblasRowMajor = 101,
CblasColMajor = 102
} CBLAS_ORDER;
#endif
int LAPACKE_sgetrf( int LAPACKE_sgetrf(
int matrix_layout, int m, int n, float* a, int lda, int* ipiv); int matrix_layout, int m, int n, float* a, int lda, int* ipiv);
int LAPACKE_dgetrf( int LAPACKE_dgetrf(
...@@ -56,6 +63,7 @@ int LAPACKE_dgetri( ...@@ -56,6 +63,7 @@ int LAPACKE_dgetri(
namespace paddle { namespace paddle {
#ifndef PADDLE_USE_EIGEN_FOR_BLAS
template <class T> template <class T>
void gemm(const CBLAS_TRANSPOSE transA, void gemm(const CBLAS_TRANSPOSE transA,
const CBLAS_TRANSPOSE transB, const CBLAS_TRANSPOSE transB,
...@@ -70,6 +78,7 @@ void gemm(const CBLAS_TRANSPOSE transA, ...@@ -70,6 +78,7 @@ void gemm(const CBLAS_TRANSPOSE transA,
const T beta, const T beta,
T* C, T* C,
const int ldc); const int ldc);
#endif
template <class T> template <class T>
int getrf(const CBLAS_ORDER Order, int getrf(const CBLAS_ORDER Order,
...@@ -84,10 +93,21 @@ int getri( ...@@ -84,10 +93,21 @@ int getri(
const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv); const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv);
template <class T> template <class T>
void axpy(const int n, const T alpha, const T* x, T* y); void axpy(const int n, const T alpha, const T* x, T* y) {
/// y = y + alpha * x
for (int i = 0; i < n; i++) {
y[i] = y[i] + alpha * x[i];
}
}
template <class T> template <class T>
T dotProduct(const int n, const T* x, const T* y); T dotProduct(const int n, const T* x, const T* y) {
T result = static_cast<T>(0);
for (int i = 0; i < n; i++) {
result += x[i] * y[i];
}
return result;
}
template <class T> template <class T>
void vExp(const int n, const T* a, T* r); void vExp(const int n, const T* a, T* r);
......
paddle/math/Matrix.cpp
浏览文件 @ 0146e8bf
...@@ -28,6 +28,7 @@ limitations under the License. */ ...@@ -28,6 +28,7 @@ limitations under the License. */
#include "hl_top_k.h" #include "hl_top_k.h"
#include "paddle/utils/Logging.h" #include "paddle/utils/Logging.h"
#include "paddle/function/GemmFunctor.h"
#include "paddle/utils/ThreadLocal.h" #include "paddle/utils/ThreadLocal.h"
#include "SIMDFunctions.h" #include "SIMDFunctions.h"
...@@ -2773,24 +2774,24 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { ...@@ -2773,24 +2774,24 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
CHECK(!isTransposed()) << "Not supported"; CHECK(!isTransposed()) << "Not supported";
size_t a_col, b_col, a_row, b_row; size_t a_col, b_col, a_row, b_row;
CBLAS_TRANSPOSE a_trans, b_trans; bool a_trans, b_trans;
if (!a->isTransposed()) { if (!a->isTransposed()) {
a_col = a->getWidth(); a_col = a->getWidth();
a_row = a->getHeight(); a_row = a->getHeight();
a_trans = CblasNoTrans; a_trans = false;
} else { } else {
a_col = a->getHeight(); a_col = a->getHeight();
a_row = a->getWidth(); a_row = a->getWidth();
a_trans = CblasTrans; a_trans = true;
} }
if (!b->isTransposed()) { if (!b->isTransposed()) {
b_col = b->getWidth(); b_col = b->getWidth();
b_row = b->getHeight(); b_row = b->getHeight();
b_trans = CblasNoTrans; b_trans = false;
} else { } else {
b_col = b->getHeight(); b_col = b->getHeight();
b_row = b->getWidth(); b_row = b->getWidth();
b_trans = CblasTrans; b_trans = true;
} }
CHECK_EQ(a_col, b_row); CHECK_EQ(a_col, b_row);
...@@ -2807,7 +2808,7 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { ...@@ -2807,7 +2808,7 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
int lda = a->getStride(); int lda = a->getStride();
int ldb = b->getStride(); int ldb = b->getStride();
int ldc = getStride(); int ldc = getStride();
gemm<real>( BlasGemm<DEVICE_TYPE_CPU, real>::compute(
a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc); a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc);
} }
......
paddle/scripts/docker/build_android.sh
浏览文件 @ 0146e8bf
...@@ -2,22 +2,58 @@ ...@@ -2,22 +2,58 @@
set -xe set -xe
mkdir -p /paddle/build_android BUILD_ROOT=/paddle/build_android
cd /paddle/build_android DEST_ROOT=/paddle/install
rm -rf /paddle/install 2>/dev/null || true
cmake -DCMAKE_SYSTEM_NAME=Android \ rm -rf $BUILD_ROOT 2>/dev/null || true
-DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \ mkdir -p $BUILD_ROOT
-DANDROID_ABI=armeabi-v7a \ cd $BUILD_ROOT
-DANDROID_ARM_NEON=ON \
-DANDROID_ARM_MODE=ON \ if [ $ANDROID_ABI == "armeabi-v7a" ]; then
-DHOST_C_COMPILER=/usr/bin/gcc \ cmake -DCMAKE_SYSTEM_NAME=Android \
-DHOST_CXX_COMPILER=/usr/bin/g++ \ -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
-DCMAKE_INSTALL_PREFIX=/paddle/install \ -DANDROID_ABI=$ANDROID_ABI \
-DCMAKE_BUILD_TYPE=RelWithDebInfo \ -DANDROID_ARM_NEON=ON \
-DCMAKE_C_FLAGS_RELWITHDEBINFO="-O3" \ -DANDROID_ARM_MODE=ON \
-DCMAKE_CXX_FLAGS_RELWITHDEBINFO="-O3" \ -DHOST_C_COMPILER=/usr/bin/gcc \
-DWITH_C_API=ON \ -DHOST_CXX_COMPILER=/usr/bin/g++ \
-DWITH_SWIG_PY=OFF \ -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
.. -DCMAKE_BUILD_TYPE=Release \
-DUSE_EIGEN_FOR_BLAS=ON \
-DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \
-DWITH_STYLE_CHECK=OFF \
..
elif [ $ANDROID_ABI == "arm64-v8a" ]; then
cmake -DCMAKE_SYSTEM_NAME=Android \
-DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM64_STANDALONE_TOOLCHAIN \
-DANDROID_ABI=$ANDROID_ABI \
-DANDROID_ARM_MODE=ON \
-DHOST_C_COMPILER=/usr/bin/gcc \
-DHOST_CXX_COMPILER=/usr/bin/g++ \
-DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
-DCMAKE_BUILD_TYPE=Release \
-DUSE_EIGEN_FOR_BLAS=OFF \
-DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \
-DWITH_STYLE_CHECK=OFF \
..
elif [ $ANDROID_ABI == "armeabi" ]; then
cmake -DCMAKE_SYSTEM_NAME=Android \
-DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
-DANDROID_ABI=$ANDROID_ABI \
-DANDROID_ARM_MODE=ON \
-DHOST_C_COMPILER=/usr/bin/gcc \
-DHOST_CXX_COMPILER=/usr/bin/g++ \
-DCMAKE_INSTALL_PREFIX=/paddle/install \
-DCMAKE_BUILD_TYPE=Release \
-DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \
-DWITH_STYLE_CHECK=OFF \
..
else
echo "Invalid ANDROID_ABI: $ANDROID_ABI"
fi
make -j `nproc` make -j `nproc`
make install -j `nproc` make install -j `nproc`
paddle/scripts/travis/build_android.sh
浏览文件 @ 0146e8bf
...@@ -22,6 +22,7 @@ cmake -DCMAKE_SYSTEM_NAME=Android \ ...@@ -22,6 +22,7 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
-DANDROID_ABI=armeabi-v7a \ -DANDROID_ABI=armeabi-v7a \
-DANDROID_ARM_NEON=ON \ -DANDROID_ARM_NEON=ON \
-DANDROID_ARM_MODE=ON \ -DANDROID_ARM_MODE=ON \
-DUSE_EIGEN_FOR_BLAS=ON \
-DWITH_C_API=ON \ -DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \ -DWITH_SWIG_PY=OFF \
-DWITH_STYLE_CHECK=OFF \ -DWITH_STYLE_CHECK=OFF \
......
Markdown is supported
0% 或 .
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 或