Skip to content

Commit

Permalink
Merge branch 'develop' into opt_softmax_bwd
Browse files Browse the repository at this point in the history
  • Loading branch information
@ZzSean
ZzSean committed Feb 10, 2022
2 parents 874560e + 238f3c8 commit 24e82bb
Show file tree
Hide file tree
Showing 73 changed files with 4,173 additions and 1,139 deletions.
114 changes: 76 additions & 38 deletions paddle/fluid/framework/custom_kernel_test.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -35,13 +35,12 @@ limitations under the License. */
// user kernel function
namespace custom_kernel {

// Here we use dot <CPU, ANY, UINT8> for test
// This test will fail when these two kernels are aupported in framework
// Here we use fake_dot for test
// input 3: two Tensors and one std::vector<Tensor>
// attribute 11: fake_attributes
// output 2: one Tensor* and one std::vector<Tensor*>
template <typename T>
void FakeDot(const paddle::CPUContext& dev_ctx, const paddle::Tensor& x,
template <typename T, typename Context>
void FakeDot(const Context& dev_ctx, const paddle::Tensor& x,
const paddle::Tensor& y,
const std::vector<paddle::Tensor>& fake_input_vec,
bool fake_attr_bool, int fake_attr_int, float fake_attr_float,
Expand Down Expand Up @@ -93,53 +92,91 @@ void FakeDot(const paddle::CPUContext& dev_ctx, const paddle::Tensor& x,
}
} // namespace custom_kernel

PD_REGISTER_KERNEL(dot, CPU, ALL_LAYOUT, UINT8,
custom_kernel::FakeDot<uint8_t>) {
/* do some args define here
* the only param can be used is OpKernelInfo* kernel */
kernel->OutputAt(0).SetDataType(paddle::experimental::DataType::UINT8);
}
PD_REGISTER_KERNEL(fake_dot, CPU, ALL_LAYOUT, custom_kernel::FakeDot, float,
double, int, int64_t, int8_t, uint8_t) {}

// Upper code will store dot kernels info into OpKernelInfoMap
TEST(CustomKernel, custom_kernel_dot) {
std::string op_name = "dot";
std::string op_name = "fake_dot";
pten::Backend backend = pten::Backend::CPU;
pten::DataLayout layout = pten::DataLayout::ANY;
pten::DataType dtype = pten::DataType::UINT8;
pten::DataLayout layout = pten::DataLayout::ALL_LAYOUT;

// 1.custom kernel info parsed and store
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance().GetMap().find("dot") !=
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance().GetMap().find(op_name) !=
paddle::OpKernelInfoMap::Instance().GetMap().end());

// 2.info check
EXPECT_EQ(
1, static_cast<int>(paddle::OpKernelInfoMap::Instance()["dot"].size()));
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()["dot"][0].GetBackend() ==
6, static_cast<int>(paddle::OpKernelInfoMap::Instance()[op_name].size()));
// index 0
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()[op_name][0].GetBackend() ==
backend);
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()["dot"][0].GetDataLayout() ==
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()[op_name][0].GetDataLayout() ==
layout);
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()["dot"][0].GetDataType() ==
dtype);

// 3.register
EXPECT_TRUE(pten::KernelFactory::Instance().kernels().end() !=
pten::KernelFactory::Instance().kernels().find("dot"));

pten::KernelKey kernel_key(backend, layout, dtype);
EXPECT_TRUE(
pten::KernelFactory::Instance().kernels()["dot"].find(kernel_key) ==
pten::KernelFactory::Instance().kernels()["dot"].end());

EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()[op_name][0].GetDataType() ==
pten::DataType::FLOAT32);
// index 5
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()[op_name][5].GetBackend() ==
backend);
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()[op_name][5].GetDataLayout() ==
layout);
EXPECT_TRUE(paddle::OpKernelInfoMap::Instance()[op_name][5].GetDataType() ==
pten::DataType::UINT8);

// 3.before register
auto& kernel_factory_instance = pten::KernelFactory::Instance();
auto& kernels = pten::KernelFactory::Instance().kernels();
EXPECT_TRUE(!kernel_factory_instance.HasCompatiblePtenKernel(op_name));

// mock fake_dot is supported by pten for HasCompatiblePtenKernel check while
// registering
auto& fake_dot_kernels = kernels[op_name];

EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::FLOAT32)) ==
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::FLOAT64)) ==
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::INT32)) ==
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::INT64)) ==
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::INT8)) ==
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::UINT8)) ==
fake_dot_kernels.end());

// register
paddle::framework::RegisterKernelWithMetaInfoMap(
paddle::OpKernelInfoMap::Instance());

EXPECT_TRUE(
pten::KernelFactory::Instance().kernels()["dot"].find(kernel_key) !=
pten::KernelFactory::Instance().kernels()["dot"].end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::FLOAT32)) !=
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::FLOAT64)) !=
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::INT32)) !=
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::INT64)) !=
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::INT8)) !=
fake_dot_kernels.end());
EXPECT_TRUE(fake_dot_kernels.find(
pten::KernelKey(backend, layout, pten::DataType::UINT8)) !=
fake_dot_kernels.end());

// 4.kernel select
auto kernel = pten::KernelFactory::Instance().SelectKernelOrThrowError(
op_name, kernel_key);
auto kernel = kernel_factory_instance.SelectKernelOrThrowError(
op_name, pten::KernelKey(backend, layout, pten::DataType::UINT8));

// 5.prepare parameters for kernel
const auto alloc = std::make_unique<paddle::experimental::DefaultAllocator>(
Expand Down Expand Up @@ -252,10 +289,10 @@ TEST(CustomKernel, custom_kernel_dot) {
// test OpKernelInfoHelper
TEST(OpKernelInfoHelper, op_kernel_info_help_getters) {
using OpKernelInfoHelper = paddle::framework::OpKernelInfoHelper;
std::string op_name = "dot";
std::string op_name = "fake_dot";
pten::Backend backend = pten::Backend::CPU;
pten::DataLayout layout = pten::DataLayout::ANY;
pten::DataType dtype = pten::DataType::UINT8;
pten::DataType dtype = pten::DataType::FLOAT32;

auto op_kernel_info = paddle::OpKernelInfoMap::Instance()[op_name][0];

Expand All @@ -268,10 +305,11 @@ TEST(OpKernelInfoHelper, op_kernel_info_help_getters) {
OpKernelInfoHelper::GetKernelKey(op_kernel_info));

paddle::CustomKernelFunc kernel_fn =
PD_PT_KERNEL(custom_kernel::FakeDot<uint8_t>);
PD_PT_KERNEL(custom_kernel::FakeDot<float, paddle::CPUContext>);
EXPECT_EQ(kernel_fn, OpKernelInfoHelper::GetKernelFn(op_kernel_info));

void* variadic_func = PD_PT_VARIADIC_KERNEL(custom_kernel::FakeDot<uint8_t>);
void* variadic_func =
PD_PT_VARIADIC_KERNEL(custom_kernel::FakeDot<float, paddle::CPUContext>);
EXPECT_EQ(variadic_func,
OpKernelInfoHelper::GetVariadicKernelFn(op_kernel_info));

Expand Down
27 changes: 27 additions & 0 deletions paddle/fluid/inference/api/details/zero_copy_tensor.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -429,6 +429,33 @@ std::vector<int> Tensor::shape() const {
PADDLE_ENFORCE_NOT_NULL(
tensor_, paddle::platform::errors::PreconditionNotMet(
"Not found tensor called %s in the scope", name_));
// mkldnn may does layout transform internally, so need to reorder before
// return
#ifdef PADDLE_WITH_MKLDNN
if (tensor->layout() == paddle::framework::DataLayout::kMKLDNN) {
paddle::framework::DataLayout out_layout =
paddle::platform::MKLDNNDeviceContext::tls()
.get_cur_paddle_data_layout();
// Set default as NCHW in case not specified
out_layout = out_layout == paddle::framework::DataLayout::kAnyLayout
? paddle::framework::DataLayout::kNCHW
: out_layout;
// In these data layouts, channel dimension is either on 2nd position: nChw
// or
// at last nhwC, so for dim==2 these layouts are the same and nothing should
// be done. Similarly for dim==1 when you have just one possible
// combination.
if (tensor->dims().size() < 3)
return paddle::framework::vectorize<int>(tensor->dims());
if (out_layout == paddle::framework::DataLayout::kNHWC) {
auto dims = paddle::framework::vectorize<int>(tensor->dims());
std::rotate(dims.begin() + 1, dims.begin() + 2, dims.end());
return dims;
} else {
return paddle::framework::vectorize<int>(tensor->dims());
}
}
#endif
return paddle::framework::vectorize<int>(tensor->dims());
}

Expand Down
5 changes: 5 additions & 0 deletions paddle/fluid/inference/experimental/javaapi/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
include_directories($ENV{jni_path} $ENV{jni_sub_path} $ENV{paddle_path})
find_library(PADDLE_INFERENCE_C libpaddle_inference_c.so HINTS $ENV{paddle_inference_lib})
aux_source_directory(native JNI_SRCS)
add_library(paddle_inference SHARED ${JNI_SRCS})
target_link_libraries(paddle_inference ${PADDLE_INFERENCE_C})
23 changes: 23 additions & 0 deletions paddle/fluid/inference/experimental/javaapi/build.sh
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,23 @@
#!/bin/bash
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#

mkdir build && cd build
export library_path=$1
export jni_path=$2
export jni_sub_path=$3
mkldnn_lib=$library_path"/third_party/install/mkldnn/lib"
mklml_lib=$library_path"/third_party/install/mklml/lib"
export paddle_inference_lib=$library_path"/paddle/lib"
export paddle_path=$library_path"/paddle/include"
export LD_LIBRARY_PATH=mkldnn_lib:mklml_lib:paddle_inference_lib
cmake .. && make
#g++ -fPIC -D_REENTRANT -I $jni_path -I $jni_sub_path -I $paddle_path -L $paddle_inference_lib -c com_baidu_paddle_inference_Predictor.cpp com_baidu_paddle_inference_Config.cpp com_baidu_paddle_inference_Tensor.cpp
#g++ -shared -I $paddle_path -L $paddle_inference_lib com_baidu_paddle_inference_Config.o com_baidu_paddle_inference_Predictor.o com_baidu_paddle_inference_Tensor.o -o libpaddle_inference.so -lpaddle_inference_c

cd ../src/main/java/com/baidu/paddle/inference
javac Config.java Predictor.java Tensor.java
cd ../../../../../../../
cp ./native/libpaddle_inference.so libpaddle_inference.so
pwd
jar cvf JavaInference.jar -C src/main/java/ .
Loading

1 comment on commit 24e82bb

@paddle-bot-old
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅Congratulation! Your pull request passed all required CI. You could ask reviewer(s) to approve and merge. 🎉

Please sign in to comment.