m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/recordio_file_reader

Browse Source
shanyi15-patch-2
Yu Yang 7 years ago
parent db46778bdd 47ca1814f3
commit b536799af0
59 changed files with 1892 additions and 736 deletions
Show Stats Download Patch File Download Diff File

29
benchmark/cluster/vgg16/vgg16_fluid.py
Unescape View File

@ -138,13 +138,14 @@ def main():
avg_cost = fluid.layers.mean(x=cost)
# Evaluator
accuracy = fluid.evaluator.Accuracy(input=predict, label=label)
batch_size = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size)
# inference program
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
test_target = accuracy.metrics + accuracy.states
inference_program = fluid.io.get_inference_program(test_target)
inference_program = fluid.io.get_inference_program(batch_acc)
# Optimization
optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
@ -157,27 +158,30 @@ def main():
# test
def test(exe):
accuracy.reset(exe)
test_pass_acc = fluid.average.WeightedAverage()
for batch_id, data in enumerate(test_reader()):
img_data = np.array(map(lambda x: x[0].reshape(data_shape),
data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
exe.run(inference_program,
feed={"pixel": img_data,
"label": y_data})
outs = exe.run(inference_program,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[batch_acc, batch_size])
test_pass_acc.add(value=np.array(outs[0]), weight=np.array(outs[1]))
return accuracy.eval(exe)
return test_pass_acc.eval()
def train_loop(exe, trainer_prog):
iters = 0
ts = time.time()
train_pass_acc = fluid.average.WeightedAverage()
for pass_id in range(args.num_passes):
# train
start_time = time.time()
num_samples = 0
accuracy.reset(exe)
train_pass_acc.reset()
with profiler.profiler("CPU", 'total') as prof:
for batch_id, data in enumerate(train_reader()):
ts = time.time()
@ -187,13 +191,14 @@ def main():
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([-1, 1])
loss, acc = exe.run(
loss, acc, b_size = exe.run(
trainer_prog,
feed={"pixel": img_data,
"label": y_data},
fetch_list=[avg_cost] + accuracy.metrics)
fetch_list=[avg_cost, batch_acc, batch_size])
iters += 1
num_samples += len(data)
train_pass_acc.add(value=acc, weight=b_size)
print(
"Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed = %.2f img/s"
% (pass_id, iters, loss, acc,
@ -201,7 +206,7 @@ def main():
) # The accuracy is the accumulation of batches, but not the current batch.
pass_elapsed = time.time() - start_time
pass_train_acc = accuracy.eval(exe)
pass_train_acc = train_pass_acc.eval()
pass_test_acc = test(exe)
print(
"Pass = %d, Training performance = %f imgs/s, Train accuracy = %f, Test accuracy = %f\n"

6
cmake/external/openblas.cmake vendored
Unescape View File

@ -78,6 +78,7 @@ IF(NOT ${CBLAS_FOUND})
BUILD_IN_SOURCE 1
BUILD_COMMAND ${CMAKE_MAKE_PROGRAM} ${COMMON_ARGS} ${OPTIONAL_ARGS}
INSTALL_COMMAND ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 NO_LAPACK=1 PREFIX=<INSTALL_DIR>
&& rm -r ${CBLAS_INSTALL_DIR}/lib/cmake ${CBLAS_INSTALL_DIR}/lib/pkgconfig
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
)
@ -100,11 +101,6 @@ IF(NOT ${CBLAS_FOUND})
\"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
)"
)
INSTALL(CODE "execute_process(
COMMAND rm -r ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}/cmake
${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}/pkgconfig
)"
)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})

5
cmake/generic.cmake
Unescape View File

@ -186,7 +186,9 @@ function(cc_library TARGET_NAME)
add_library(${TARGET_NAME} SHARED ${cc_library_SRCS})
else()
add_library(${TARGET_NAME} STATIC ${cc_library_SRCS})
find_fluid_modules(${TARGET_NAME})
endif()
if(cc_library_DEPS)
# Don't need link libwarpctc.so
if("${cc_library_DEPS};" MATCHES "warpctc;")
@ -263,7 +265,8 @@ function(nv_library TARGET_NAME)
if (nv_library_SHARED OR nv_library_shared) # build *.so
cuda_add_library(${TARGET_NAME} SHARED ${nv_library_SRCS})
else()
cuda_add_library(${TARGET_NAME} STATIC ${nv_library_SRCS})
cuda_add_library(${TARGET_NAME} STATIC ${nv_library_SRCS})
find_fluid_modules(${TARGET_NAME})
endif()
if (nv_library_DEPS)
add_dependencies(${TARGET_NAME} ${nv_library_DEPS})

35
cmake/inference_lib.cmake
Unescape View File

@ -1,9 +1,22 @@
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
# make package for paddle fluid shared and static library
function(copy TARGET)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS DSTS DEPS)
cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(inference_lib_dist_dep ${TARGET} ${inference_lib_dist_dep} PARENT_SCOPE)
list(LENGTH copy_lib_SRCS copy_lib_SRCS_len)
list(LENGTH copy_lib_DSTS copy_lib_DSTS_len)
@ -42,13 +55,21 @@ copy(glog_lib
DSTS ${dst_dir} ${dst_dir}/lib
)
IF(NOT PROTOBUF_FOUND)
if(NOT PROTOBUF_FOUND)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf")
copy(protobuf_lib
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LITE_LIBRARY}
SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY}
DSTS ${dst_dir} ${dst_dir}/lib
)
ENDIF(NOT PROTOBUF_FOUND)
endif()
if(NOT CBLAS_FOUND)
set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/openblas")
copy(openblas_lib
SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
DSTS ${dst_dir} ${dst_dir}
)
endif()
# paddle fluid module
set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
@ -66,8 +87,8 @@ copy(memory_lib
)
set(module "inference")
copy(inference_lib DEPENDS paddle_fluid_shared
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.so
copy(inference_lib DEPS paddle_fluid_shared paddle_fluid
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
DSTS ${dst_dir}/${module} ${dst_dir}/${module}
)
@ -83,6 +104,4 @@ copy(string_lib
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
)
add_custom_target(inference_lib_dist DEPENDS
inference_lib framework_lib memory_lib platform_lib string_lib
gflags_lib glog_lib protobuf_lib eigen3_lib)
add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep})

BIN
doc/fluid/howto/optimization/timeline.jpeg
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 69 KiB

27
doc/fluid/howto/optimization/timeline.md
Unescape View File

@ -0,0 +1,27 @@
## how to use timeline tool to do profile
1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number.
```python
with profiler.profiler('All', 'total', '/tmp/profile') as prof:
for pass_id in range(pass_num):
for batch_id, data in enumerate(train_reader()):
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[],
use_program_cache=True)
...
```
1. Run `python paddle/tools/timeline.py` to process `/tmp/profile`, it will generate another
file `/tmp/timeline` by default. You can change the path by cmd parameter, please take a look at
[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py) for details.
1. Open chrome and visit <chrome://tracing/>, use `load` button to load the generated `timeline` file.
![chrome tracing](./tracing.jpeg)
1. The resulting timeline should be like:
![chrome timeline](./timeline.jpeg)

BIN
doc/fluid/howto/optimization/tracing.jpeg
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 30 KiB

2
doc/v2/build_and_install/pip_install_cn.rst
Unescape View File

@ -39,7 +39,7 @@ PaddlePaddle可以使用常用的Python包管理工具
"cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddle.tgz>`_"
"cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"

2
doc/v2/build_and_install/pip_install_en.rst
Unescape View File

@ -42,7 +42,7 @@ If the links below shows up the login form, just click "Log in as guest" to star
"cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
"cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddle.tgz>`_"
"cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
"cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"

26
doc/v2/howto/index_cn.rst
Unescape View File

@ -1,11 +1,37 @@
进阶使用
========
PaddlePaddle支持用户灵活地设置各种命令行参数，以实现对模型训练或预测流程的控制。使用方式请参考：
.. toctree::
:maxdepth: 1
cmd_parameter/index_cn.rst
PaddlePaddle支持在fabric集群、MPI集群、kubernetes集群上分布式训练任务，具体环境配置和使用说明请参考：
.. toctree::
:maxdepth: 1
cluster/index_cn.rst
PaddlePaddle提供了用于预测的C-API，关于C-API的使用，我们提供了如下指南:
.. toctree::
:maxdepth: 1
capi/index_cn.rst
PaddlePaddle支持多种灵活和高效的循环神经网络，具体配置使用方式请参考：
.. toctree::
:maxdepth: 1
rnn/index_cn.rst
关于如何使用内置的定时工具、nvprof 或 nvvp 来运行性能分析和调优，请参考：
.. toctree::
:maxdepth: 1
optimization/gpu_profiling_cn.rst

1
paddle/fluid/framework/data_type_transform.cc
Unescape View File

@ -53,6 +53,7 @@ struct CastDataType {
auto* context = static_cast<const platform::CUDADeviceContext*>(ctx_);
trans(*context, in_begin, in_end, out_begin,
CastDataTypeFunctor<InType, OutType>());
context->Wait();
#endif
} else {
PADDLE_THROW("Unsupported place!");

24
paddle/fluid/framework/data_type_transform_test.cc
Unescape View File

@ -50,13 +50,13 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp32, kernel_fp64, in, &out);
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(i / 3));
EXPECT_EQ(out_data_double[i], static_cast<double>(i / 3));
}
TransDataType(kernel_fp32, kernel_int32, in, &out);
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(i / 3));
EXPECT_EQ(out_data_int[i], static_cast<int>(i / 3));
}
}
@ -76,31 +76,31 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp16, kernel_fp32, in, &out);
float* out_data_float = out.data<float>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
EXPECT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_fp64, in, &out);
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int32, in, &out);
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int64, in, &out);
int64_t* out_data_int64 = out.data<int64_t>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
EXPECT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_bool, in, &out);
bool* out_data_bool = out.data<bool>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
EXPECT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
}
// transform float to float16
@ -112,7 +112,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp32, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
}
// transform double to float16
@ -124,7 +124,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp64, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
}
// transform int to float16
@ -136,7 +136,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_int32, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
}
// transform int64 to float16
@ -148,7 +148,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_int64, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
}
// transform bool to float16
@ -160,7 +160,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_bool, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
}
}
}

33
paddle/fluid/framework/data_type_transform_test.cu
Unescape View File

@ -49,15 +49,16 @@ TEST(DataTypeTransform, GPUTransform) {
float arr[6] = {0, 1, 2, 3, 4, 5};
int data_number = sizeof(arr) / sizeof(arr[0]);
memcpy(in_ptr, arr, sizeof(arr));
TensorCopy(in, gpu_place, context, &in_gpu);
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp32, kernel_fp64, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(arr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(arr[i]));
}
TransDataType(kernel_fp32, kernel_int32, in_gpu, &out_gpu);
@ -66,7 +67,7 @@ TEST(DataTypeTransform, GPUTransform) {
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(arr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(arr[i]));
}
}
@ -83,6 +84,7 @@ TEST(DataTypeTransform, GPUTransform) {
int data_number = sizeof(arr) / sizeof(arr[0]);
memcpy(ptr, arr, sizeof(arr));
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
// transform from float16 to other data types
TransDataType(kernel_fp16, kernel_fp32, in_gpu, &out_gpu);
@ -91,7 +93,7 @@ TEST(DataTypeTransform, GPUTransform) {
float* out_data_float = out.data<float>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
EXPECT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_fp64, in_gpu, &out_gpu);
@ -100,7 +102,7 @@ TEST(DataTypeTransform, GPUTransform) {
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int32, in_gpu, &out_gpu);
@ -109,7 +111,7 @@ TEST(DataTypeTransform, GPUTransform) {
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int64, in_gpu, &out_gpu);
@ -118,7 +120,7 @@ TEST(DataTypeTransform, GPUTransform) {
int64_t* out_data_int64 = out.data<int64_t>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
EXPECT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_bool, in_gpu, &out_gpu);
@ -127,7 +129,7 @@ TEST(DataTypeTransform, GPUTransform) {
bool* out_data_bool = out.data<bool>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
EXPECT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
}
// transform float to float16
@ -137,13 +139,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp32, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
}
// transform double to float16
@ -154,13 +157,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp64, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
}
// transform int to float16
@ -170,13 +174,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_int32, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
}
// transform int64 to float16
@ -187,13 +192,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_int64, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
}
// transform bool to float16
@ -203,13 +209,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_bool, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
}
}
}

3
paddle/fluid/inference/CMakeLists.txt
Unescape View File

@ -5,7 +5,8 @@ cc_library(paddle_fluid_api
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
# Create static library
cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
cc_library(paddle_fluid DEPS ${fluid_modules})
# Create shared library
cc_library(paddle_fluid_shared SHARED

15
paddle/fluid/inference/io.cc
Unescape View File

@ -22,14 +22,14 @@ namespace paddle {
namespace inference {
void ReadBinaryFile(const std::string& filename, std::string& contents) {
VLOG(3) << "loading model from " << filename;
std::ifstream inputfs(filename, std::ios::in | std::ios::binary);
inputfs.seekg(0, std::ios::end);
std::ifstream fin(filename, std::ios::in | std::ios::binary);
PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", filename);
fin.seekg(0, std::ios::end);
contents.clear();
contents.resize(inputfs.tellg());
inputfs.seekg(0, std::ios::beg);
inputfs.read(&contents[0], contents.size());
inputfs.close();
contents.resize(fin.tellg());
fin.seekg(0, std::ios::beg);
fin.read(&contents[0], contents.size());
fin.close();
}
bool IsPersistable(const framework::VarDesc* var) {
@ -97,6 +97,7 @@ std::unique_ptr<framework::ProgramDesc> Load(framework::Executor& executor,
const std::string& dirname) {
std::string model_filename = dirname + "/__model__";
std::string program_desc_str;
VLOG(3) << "loading model from " << model_filename;
ReadBinaryFile(model_filename, program_desc_str);
std::unique_ptr<framework::ProgramDesc> main_program(

19
paddle/fluid/inference/tests/book/test_inference_image_classification.cc
Unescape View File

@ -17,10 +17,13 @@ limitations under the License. */
#include "paddle/fluid/inference/tests/test_helper.h"
DEFINE_string(dirname, "", "Directory of the inference model.");
DEFINE_int32(batch_size, 1, "Batch size of input data");
DEFINE_int32(repeat, 1, "Running the inference program repeat times");
TEST(inference, image_classification) {
if (FLAGS_dirname.empty()) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model "
"--batch_size=1 --repeat=1";
}
LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
@ -29,13 +32,11 @@ TEST(inference, image_classification) {
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
int64_t batch_size = 1;
paddle::framework::LoDTensor input;
// Use normilized image pixels as input data,
// which should be in the range [0.0, 1.0].
SetupTensor<float>(input,
{batch_size, 3, 32, 32},
{FLAGS_batch_size, 3, 32, 32},
static_cast<float>(0),
static_cast<float>(1));
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
@ -46,7 +47,9 @@ TEST(inference, image_classification) {
cpu_fetchs1.push_back(&output1);
// Run inference on CPU
TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << "--- CPU Runs: ---";
TestInference<paddle::platform::CPUPlace>(
dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
@ -55,7 +58,9 @@ TEST(inference, image_classification) {
cpu_fetchs2.push_back(&output2);
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << "--- GPU Runs: ---";
TestInference<paddle::platform::CUDAPlace>(
dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat);
LOG(INFO) << output2.dims();
CheckError<float>(output1, output2);

83
paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc
Unescape View File

@ -17,10 +17,13 @@ limitations under the License. */
#include "paddle/fluid/inference/tests/test_helper.h"
DEFINE_string(dirname, "", "Directory of the inference model.");
DEFINE_int32(batch_size, 1, "Batch size of input data");
DEFINE_int32(repeat, 1, "Running the inference program repeat times");
TEST(inference, recognize_digits) {
if (FLAGS_dirname.empty()) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model "
"--batch_size=1 --repeat=1";
}
LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
@ -29,77 +32,39 @@ TEST(inference, recognize_digits) {
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
int64_t batch_size = 1;
paddle::framework::LoDTensor input;
// Use normilized image pixels as input data,
// which should be in the range [-1.0, 1.0].
SetupTensor<float>(input,
{batch_size, 1, 28, 28},
{FLAGS_batch_size, 1, 28, 28},
static_cast<float>(-1),
static_cast<float>(1));
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
cpu_feeds.push_back(&input);
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
for (auto is_combined : {false, true}) {
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
// Run inference on CPU
TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << output1.dims();
// Run inference on CPU
LOG(INFO) << "--- CPU Runs: is_combined=" << is_combined << " ---";
TestInference<paddle::platform::CPUPlace>(
dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << output2.dims();
// Run inference on CUDA GPU
LOG(INFO) << "--- GPU Runs: is_combined=" << is_combined << " ---";
TestInference<paddle::platform::CUDAPlace>(
dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat, is_combined);
LOG(INFO) << output2.dims();
CheckError<float>(output1, output2);
CheckError<float>(output1, output2);
#endif
}
TEST(inference, recognize_digits_combine) {
if (FLAGS_dirname.empty()) {
LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
}
LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
std::string dirname = FLAGS_dirname;
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
paddle::framework::LoDTensor input;
// Use normilized image pixels as input data,
// which should be in the range [-1.0, 1.0].
SetupTensor<float>(
input, {1, 1, 28, 28}, static_cast<float>(-1), static_cast<float>(1));
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
cpu_feeds.push_back(&input);
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
// Run inference on CPU
TestInference<paddle::platform::CPUPlace, true>(
dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace, true>(
dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << output2.dims();
CheckError<float>(output1, output2);
#endif
}

78
paddle/fluid/inference/tests/test_helper.h
Unescape View File

@ -15,6 +15,7 @@ limitations under the License. */
#include <time.h>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/profiler.h"
template <typename T>
void SetupTensor(paddle::framework::LoDTensor& input,
@ -87,31 +88,60 @@ void CheckError(paddle::framework::LoDTensor& output1,
EXPECT_EQ(count, 0U) << "There are " << count << " different elements.";
}
template <typename Place, bool IsCombined = false>
template <typename Place>
void TestInference(const std::string& dirname,
const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
std::vector<paddle::framework::LoDTensor*>& cpu_fetchs,
const int repeat = 1,
const bool is_combined = false) {
// 1. Define place, executor, scope
auto place = Place();
auto executor = paddle::framework::Executor(place);
auto* scope = new paddle::framework::Scope();
// Profile the performance
paddle::platform::ProfilerState state;
if (paddle::platform::is_cpu_place(place)) {
state = paddle::platform::ProfilerState::kCPU;
} else {
#ifdef PADDLE_WITH_CUDA
state = paddle::platform::ProfilerState::kCUDA;
// The default device_id of paddle::platform::CUDAPlace is 0.
// Users can get the device_id using:
// int device_id = place.GetDeviceId();
paddle::platform::SetDeviceId(0);
#else
PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
#endif
}
// Enable the profiler
paddle::platform::EnableProfiler(state);
// 2. Initialize the inference_program and load parameters
std::unique_ptr<paddle::framework::ProgramDesc> inference_program;
if (IsCombined) {
// All parameters are saved in a single file.
// Hard-coding the file names of program and parameters in unittest.
// The file names should be consistent with that used in Python API
// `fluid.io.save_inference_model`.
std::string prog_filename = "__model_combined__";
std::string param_filename = "__params_combined__";
inference_program = paddle::inference::Load(executor,
*scope,
dirname + "/" + prog_filename,
dirname + "/" + param_filename);
} else {
// Parameters are saved in separate files sited in the specified `dirname`.
inference_program = paddle::inference::Load(executor, *scope, dirname);
{
paddle::platform::RecordEvent record_event(
"init_program",
paddle::platform::DeviceContextPool::Instance().Get(place));
if (is_combined) {
// All parameters are saved in a single file.
// Hard-coding the file names of program and parameters in unittest.
// The file names should be consistent with that used in Python API
// `fluid.io.save_inference_model`.
std::string prog_filename = "__model_combined__";
std::string param_filename = "__params_combined__";
inference_program =
paddle::inference::Load(executor,
*scope,
dirname + "/" + prog_filename,
dirname + "/" + param_filename);
} else {
// Parameters are saved in separate files sited in the specified
// `dirname`.
inference_program = paddle::inference::Load(executor, *scope, dirname);
}
}
// 3. Get the feed_target_names and fetch_target_names
@ -134,7 +164,21 @@ void TestInference(const std::string& dirname,
}
// 6. Run the inference program
executor.Run(*inference_program, scope, feed_targets, fetch_targets);
{
// Run repeat times to profile the performance
for (int i = 0; i < repeat; ++i) {
paddle::platform::RecordEvent record_event(
"run_inference",
paddle::platform::DeviceContextPool::Instance().Get(place));
executor.Run(*inference_program, scope, feed_targets, fetch_targets);
}
}
// Disable the profiler and print the timing information
paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
"profiler.txt");
paddle::platform::ResetProfiler();
delete scope;
}

33
paddle/fluid/operators/CMakeLists.txt
Unescape View File

@ -1,5 +1,7 @@
file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
string(REPLACE "_mkldnn" "" GENERAL_OPS "${GENERAL_OPS}")
string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
list(REMOVE_DUPLICATES GENERAL_OPS)
set(DEPS_OPS "")
set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/fluid/pybind/pybind.h)
file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt. DO NOT EDIT!\n\n")
@ -13,6 +15,8 @@ function(op_library TARGET)
set(cu_cc_srcs)
set(cudnn_cu_cc_srcs)
set(CUDNN_FILE)
set(mkldnn_cc_srcs)
set(MKLDNN_FILE)
set(op_common_deps operator op_registry math_function)
set(options "")
set(oneValueArgs "")
@ -36,12 +40,20 @@ function(op_library TARGET)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc)
list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc)
endif()
if(WITH_MKLDNN)
string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc)
endif()
endif()
else()
foreach(src ${op_library_SRCS})
if (${src} MATCHES ".*\\.cu$")
list(APPEND cu_srcs ${src})
elseif(${src} MATCHES ".*_cudnn_op.cu.cc$")
list(APPEND cudnn_cu_cc_srcs ${src})
elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$")
list(APPEND mkldnn_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cu.cc$")
list(APPEND cu_cc_srcs ${src})
elseif(${src} MATCHES ".*\\.cc$")
@ -62,11 +74,11 @@ function(op_library TARGET)
set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
endif()
if (WITH_GPU)
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
else()
cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
endif()
# Define operators that don't need pybind here.
@ -101,7 +113,8 @@ function(op_library TARGET)
# pybind USE_CPU_ONLY_OP
list(LENGTH cu_srcs cu_srcs_len)
list(LENGTH cu_cc_srcs cu_cc_srcs_len)
if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len)
if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
set(pybind_flag 1)
endif()
@ -112,6 +125,11 @@ function(op_library TARGET)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
endif()
# pybind USE_OP_DEVICE_KERNEL for MKLDNN
if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0)
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
endif()
# pybind USE_OP
if (${pybind_flag} EQUAL 0)
file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
@ -172,17 +190,18 @@ op_library(cos_sim_op DEPS cos_sim_functor)
op_library(parallel_do_op DEPS executor)
if (WITH_GPU)
op_library(conv_op DEPS vol2col depthwise_conv)
op_library(conv_op DEPS vol2col depthwise_conv im2col)
else()
op_library(conv_op DEPS vol2col)
op_library(conv_op DEPS vol2col im2col)
endif()
op_library(conv_transpose_op DEPS vol2col)
op_library(conv_transpose_op DEPS vol2col im2col)
# FIXME(typhoonzero): save/load depends lodtensor serialization functions
op_library(save_op DEPS lod_tensor)
op_library(load_op DEPS lod_tensor)
op_library(save_combine_op DEPS lod_tensor)
op_library(load_combine_op DEPS lod_tensor)
op_library(concat_op DEPS concat)
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
foreach(src ${GENERAL_OPS})

9
paddle/fluid/operators/concat_op.cc
Unescape View File

@ -100,7 +100,8 @@ class ConcatOpGrad : public framework::OperatorWithKernel {
namespace ops = paddle::operators;
REGISTER_OP_EX(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad,
ops::ConcatOpGrad, false)
REGISTER_OP_CPU_KERNEL(concat,
ops::ConcatKernel<paddle::platform::CPUPlace, float>)
REGISTER_OP_CPU_KERNEL(concat_grad,
ops::ConcatGradKernel<paddle::platform::CPUPlace, float>)
REGISTER_OP_CPU_KERNEL(
concat, ops::ConcatKernel<paddle::platform::CPUDeviceContext, float>)
REGISTER_OP_CPU_KERNEL(
concat_grad,
ops::ConcatGradKernel<paddle::platform::CPUDeviceContext, float>)

84
paddle/fluid/operators/concat_op.h
Unescape View File

@ -17,6 +17,7 @@ limitations under the License. */
#include <utility>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/concat.h"
#include "paddle/fluid/operators/strided_memcpy.h"
namespace paddle {
@ -27,54 +28,30 @@ class ConcatKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto ins = ctx.MultiInput<framework::Tensor>("X");
auto* out = ctx.Output<framework::Tensor>("Out");
framework::Tensor* out = ctx.Output<framework::Tensor>("Out");
int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
auto place = ctx.GetPlace();
out->mutable_data<T>(place);
auto out_stride = framework::stride_numel(out->dims());
size_t output_offset = 0;
// If axis >=1, copy to out immediately need to call many times
// of cuda memcpy. Copy the input to cpu and do the stride copy,
// then copy to gpu output.
if (platform::is_gpu_place(place) && axis >= 1) {
platform::CPUPlace copy_place;
auto& cpu_ctx = *platform::DeviceContextPool::Instance().Get(copy_place);
framework::Tensor cpu_out;
cpu_out.Resize(out->dims());
cpu_out.mutable_data<T>(copy_place);
auto& dev_ctx = ctx.device_context();
std::vector<std::unique_ptr<framework::Tensor>> cpu_ins;
for (auto* in : ins) {
std::unique_ptr<framework::Tensor> cpu_in(new framework::Tensor);
framework::TensorCopy(*in, copy_place, dev_ctx, cpu_in.get());
cpu_ins.emplace_back(std::move(cpu_in));
}
// TODO(dzhwinter): overlap copy and compute stream
// https://devblogs.nvidia.com/how-overlap-data-transfers-cuda-cc/
dev_ctx.Wait();
for (auto& in : cpu_ins) {
auto& cpu_in = *in.get();
auto in_stride = framework::stride_numel(cpu_in.dims());
StridedNumelCopyWithAxis<T>(
cpu_ctx, axis, cpu_out.data<T>() + output_offset, out_stride,
cpu_in.data<T>(), in_stride, in_stride[axis]);
output_offset += in_stride[axis];
}
framework::TensorCopy(cpu_out, place, dev_ctx, out);
} else {
// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && ins.size() < 10) {
size_t output_offset = 0;
for (auto* in : ins) {
auto in_stride = framework::stride_numel(in->dims());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis,
out->data<T>() + output_offset, out_stride,
in->data<T>(), in_stride, in_stride[axis]);
output_offset += in_stride[axis];
}
} else {
std::vector<framework::Tensor> inputs(ins.size());
for (size_t j = 0; j < ins.size(); ++j) {
inputs[j] = *ins[j];
}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
paddle::operators::math::ConcatFunctor<DeviceContext, T> concat_functor;
concat_functor(dev_ctx, inputs, static_cast<int>(axis), out);
}
}
};
@ -86,16 +63,31 @@ class ConcatGradKernel : public framework::OpKernel<T> {
auto* in = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
size_t input_offset = 0;
auto in_stride = framework::stride_numel(in->dims());
for (auto& out : outs) {
out->mutable_data<T>(ctx.GetPlace());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
out_stride, in->data<T>() + input_offset,
in_stride, out_stride[axis]);
input_offset += out_stride[axis];
// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && outs.size() < 10) {
size_t input_offset = 0;
auto in_stride = framework::stride_numel(in->dims());
for (auto& out : outs) {
out->mutable_data<T>(ctx.GetPlace());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
out_stride, in->data<T>() + input_offset,
in_stride, out_stride[axis]);
input_offset += out_stride[axis];
}
} else {
std::vector<framework::Tensor> outputs(outs.size());
for (size_t j = 0; j < outs.size(); ++j) {
outs[j]->mutable_data<T>(ctx.GetPlace());
outputs[j] = *outs[j];
}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
concat_grad_functor;
concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), outputs);
}
}
};

313
paddle/fluid/operators/conv_mkldnn_op.cc
View File

File diff suppressed because it is too large Load Diff

51
paddle/fluid/operators/conv_op.cc
Unescape View File

@ -13,6 +13,12 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/conv_op.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cudnn_helper.h"
#endif
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle {
namespace operators {
@ -64,22 +70,21 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOp::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
framework::LibraryType library_{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
framework::LibraryType library_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
} else {
library_ = framework::LibraryType::kPlain;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
}
#endif
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
@ -131,6 +136,9 @@ Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
"use_cudnn",
"(bool, default false) Only used in cudnn kernel, need install cudnn")
.SetDefault(false);
AddAttr<bool>("use_mkldnn",
"(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddAttr<std::string>(
"data_format",
"(string, default NCHW) Only used in "
@ -224,6 +232,9 @@ Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
"use_cudnn",
"(bool, default false) Only used in cudnn kernel, need install cudnn")
.SetDefault(false);
AddAttr<bool>("use_mkldnn",
"(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddAttr<std::string>(
"data_format",
"(string, default NCHW) Only used in "
@ -284,23 +295,21 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
const framework::ExecutionContext& ctx) const {
bool use_cudnn = ctx.Attr<bool>("use_cudnn");
use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
framework::LibraryType library_{framework::LibraryType::kPlain};
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(ctx.GetPlace())) {
auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
if (platform::CanCUDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kCUDNN;
}
#endif
framework::LibraryType library_;
if (use_cudnn) {
library_ = framework::LibraryType::kCUDNN;
} else {
library_ = framework::LibraryType::kPlain;
#ifdef PADDLE_WITH_MKLDNN
if (library_ == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library_ = framework::LibraryType::kMKLDNN;
}
#endif
std::string data_format = ctx.Attr<std::string>("data_format");
// TODO(pzelazko-intel): enable MKLDNN layout when it's ready
framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),

2
paddle/fluid/operators/detection_map_op.cc
Unescape View File

@ -71,7 +71,7 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
return framework::OpKernelType(
framework::ToDataType(
ctx.Input<framework::Tensor>("DetectRes")->type()),
ctx.device_context());
platform::CPUPlace());
}
};

79
paddle/fluid/operators/elementwise_div_op.h
Unescape View File

@ -41,77 +41,14 @@ class ElementwiseDivKernel : public framework::OpKernel<T> {
};
template <typename T>
struct ElementwiseDivGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto z_e = framework::EigenVector<T>::Flatten(*z);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e / y_e;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = -1.0 * dz_e * z_e / y_e;
}
}
};
template <typename T>
struct ElementwiseDivBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e / y_e_bcast;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
struct DivGradDX {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout / y; }
};
template <typename T>
struct ElementwiseDivBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e / y_e_bcast;
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
struct DivGradDY {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
return -dout * x / (y * y);
}
};
@ -128,10 +65,8 @@ class ElementwiseDivGradKernel : public framework::OpKernel<T> {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
int axis = ctx.Attr<int>("axis");
ElementwiseGradCompute<DeviceContext, T, ElementwiseDivGradFunctor<T>,
ElementwiseDivBroadCastGradFunctor<T>,
ElementwiseDivBroadCast2GradFunctor<T>>(
ctx, x, y, out, dout, axis, dx, dy);
ElemwiseGradCompute<DeviceContext, T, DivGradDX<T>, DivGradDY<T>>(
ctx, *x, *y, *out, *dout, axis, dx, dy, DivGradDX<T>(), DivGradDY<T>());
}
};

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save