Skip to content

P
Paddle

夏有凉风，冬有雪 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 297f06c6 编写于 作者: H hedaoyuan 提交者： GitHub
浏览文件
操作

Merge pull request #1132 from hedaoyuan/FunctionTest 

New FunctionTest
上级 010dd20a 1c5a7c43
隐藏空白更改
内联 并排
Showing with 399 addition and 136 deletion
paddle/function/BufferArg.h
浏览文件 @ 297f06c6
......@@ -40,7 +40,6 @@ enum SparseDataFormat { SPARSE_CSR_FORMAT = 0, SPARSE_CSC_FORMAT = 1 };
class BufferArg;
class SequenceArg;
class SparseMatrixArg;
typedef std::shared_ptr<BufferArg> BufferArgPtr;
/**
* \brief BufferArg used as the argument type of Function.
......@@ -51,6 +50,11 @@ typedef std::shared_ptr<BufferArg> BufferArgPtr;
* 3. SequenceArg for a Buffer of sequence data.
* 4. SparseMatrixArg for a Buffer of sparse matrix.
*
* Buffer shape
* For most buffers, the first dimension `shape()[0]` represents
* the size of the mini-batch.
*
* Buffer argType
* There is an ArgType property for the BufferArg used as Function Output.
* Whether the result of the Function calculation is assigned to the
* output Buffer or added to the output Buffer is determined by the
......@@ -72,6 +76,14 @@ public:
ArgType getArgType() const { return argType_; }
public:
BufferArg(ValueType valueType,
const TensorShape& shape,
ArgType argType = UNSPECIFIED)
: buf_(nullptr),
valueType_(valueType),
shape_(shape),
argType_(argType) {}
BufferArg(void* buf,
ValueType valueType,
const TensorShape& shape,
......@@ -177,6 +189,13 @@ protected:
// if a < b then value_.buf_[a] < value_.buf_[b]
class SequenceIdArg : public BufferArg {
public:
SequenceIdArg(const TensorShape& shape, ArgType argType = UNSPECIFIED)
: BufferArg(VALUE_TYPE_INT32, shape, argType) {
CHECK_EQ(shape_.ndims(), (size_t)1);
CHECK_GT(shape_[0], 1);
numSeqs_ = shape_[0] - 1;
}
SequenceIdArg(void* buf,
const TensorShape& shape,
ArgType argType = UNSPECIFIED)
......@@ -199,9 +218,18 @@ private:
size_t numSeqs_;
};
// sequence data {seqId(vec), buf(matrix)}
// sequences data
// For mini-batch calculate,
// one batch can contain more than one sequence of data.
// SequenceArg can be used to represent sequences that contain multiple
// unequal lengths.
class SequenceArg : public BufferArg {
public:
SequenceArg(ValueType valueType,
const TensorShape& shape,
ArgType argType = UNSPECIFIED)
: BufferArg(valueType, shape, argType), startPositions_(TensorShape()) {}
SequenceArg(void* buf,
ValueType valueType,
const TensorShape& shape,
......@@ -223,7 +251,8 @@ public:
void* getIdBuf() const { return startPositions_.data(); }
size_t numSeqs() const { return startPositions_.numSeqs(); }
const SequenceIdArg& getSequenceIds() const { return startPositions_; }
SequenceIdArg& getSequenceId() { return startPositions_; }
const SequenceIdArg& getSequenceId() const { return startPositions_; }
private:
SequenceIdArg startPositions_;
......
paddle/function/BufferArgTest.cpp
浏览文件 @ 297f06c6
......@@ -14,9 +14,7 @@ limitations under the License. */
#include "BufferArg.h"
#include <gtest/gtest.h>
#include "Function.h"
#include "paddle/math/MemoryHandle.h"
#include "paddle/math/SparseMatrix.h"
namespace paddle {
......@@ -37,55 +35,4 @@ TEST(BufferTest, SequenceIdArg) {
EXPECT_EQ(buffer.numSeqs(), 9);
}
TEST(BufferTest, asArgument) {
MatrixPtr matrix = Matrix::create(100, 200);
VectorPtr vector = Vector::create(100, false);
CpuSparseMatrix sparse(200, 300, 50);
// prepare arguments
BufferArgs argments;
argments.addArg(*matrix);
argments.addArg(*vector);
argments.addArg(sparse);
// function
auto function = [=](const BufferArgs& inputs) {
EXPECT_EQ(inputs.size(), 3);
// check inputs[0]
EXPECT_EQ(inputs[0].shape().ndims(), 2);
EXPECT_EQ(inputs[0].shape()[0], 100);
EXPECT_EQ(inputs[0].shape()[1], 200);
EXPECT_EQ(inputs[0].data(), matrix->getData());
EXPECT_EQ(inputs[0].matrix<DEVICE_TYPE_CPU>().getHeight(),
matrix->getHeight());
EXPECT_EQ(inputs[0].matrix<DEVICE_TYPE_CPU>().getWidth(),
matrix->getWidth());
EXPECT_EQ(inputs[0].matrix<DEVICE_TYPE_CPU>().getData(), matrix->getData());
// check inputs[1]
EXPECT_EQ(inputs[1].shape().ndims(), 1);
EXPECT_EQ(inputs[1].shape()[0], 100);
EXPECT_EQ(inputs[1].data(), vector->getData());
CpuVector inVector = inputs[1].vector<real, DEVICE_TYPE_CPU>();
EXPECT_EQ(inVector.getSize(), vector->getSize());
EXPECT_EQ(inVector.getData(), vector->getData());
// check inputs[2]
EXPECT_EQ(inputs[2].shape().ndims(), 2);
EXPECT_EQ(inputs[2].shape()[0], 200);
EXPECT_EQ(inputs[2].shape()[1], 300);
EXPECT_EQ(inputs[2].data(), sparse.getData());
// CHECK_EQ(inputs[2].sparse().nnz(), 50);
// CHECK_EQ(inputs[2].sparse().dataFormat(), SPARSE_CSR_FORMAT);
// CHECK_EQ(inputs[2].sparse().dataType(), SPARSE_FLOAT_VALUE);
EXPECT_EQ(inputs[2].sparse().getRowBuf(), sparse.getRows());
EXPECT_EQ(inputs[2].sparse().getColBuf(), sparse.getCols());
};
// call function
function(argments);
}
} // namespace paddle
paddle/function/CMakeLists.txt
浏览文件 @ 297f06c6
......@@ -19,7 +19,7 @@ if(WITH_TESTING)
# TODO:
# file(GLOB test_files . *OpTest.cpp)
# add_executable(${test_bin} EXCLUDE_FROM_ALL ${test_files})
# add_simple_unittest(CrossMapNormalOpTest)
add_simple_unittest(CrossMapNormalOpTest)
add_simple_unittest(TensorShapeTest)
add_simple_unittest(TensorTypeTest)
add_simple_unittest(BufferArgTest)
......
paddle/function/ContextProjectionOp.cpp
浏览文件 @ 297f06c6
......@@ -115,11 +115,10 @@ public:
const auto val_seqs = dynamic_cast<const SequenceArg&>(inputs[0]);
auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
CHECK(out_seq.data() && val_seqs.data() &&
val_seqs.getSequenceIds().data());
CHECK(out_seq.data() && val_seqs.data() && val_seqs.getSequenceId().data());
CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
CHECK_EQ(val_seqs.shape().ndims(), (size_t)2);
CHECK_EQ(val_seqs.getSequenceIds().shape().ndims(), (size_t)1);
CHECK_EQ(val_seqs.getSequenceId().shape().ndims(), (size_t)1);
if (2 == inputs.size()) {
CHECK_EQ(inputs[1].shape().ndims(), (size_t)2);
}
......@@ -139,7 +138,7 @@ public:
(2 == inputs.size())
? inputs[1].matrix<Device>()
: typename Tensor<real, Device>::Matrix(nullptr, 0, 0);
const auto seq_vec = val_seqs.getSequenceIds().vector<int, Device>();
const auto seq_vec = val_seqs.getSequenceId().vector<int, Device>();
ContextProjectionForward<Device>(out_mat,
in_mat,
w_mat,
......@@ -242,11 +241,11 @@ public:
<< "SequenceArg required here";
const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
CHECK(in_seq.data() && in_seq.getSequenceIds().data());
CHECK(in_seq.data() && in_seq.getSequenceId().data());
CHECK_EQ(in_seq.shape().ndims(), (size_t)2);
CHECK_EQ(in_seq.getSequenceIds().shape().ndims(), (size_t)1);
CHECK_EQ(in_seq.getSequenceId().shape().ndims(), (size_t)1);
CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
CHECK_EQ(out_seq.getSequenceIds().shape().ndims(), (size_t)1);
CHECK_EQ(out_seq.getSequenceId().shape().ndims(), (size_t)1);
CHECK_EQ(outputs[1].shape().ndims(), (size_t)2);
/// dim of input grad == dim of weight
......@@ -258,7 +257,7 @@ public:
CHECK_EQ(out_seq.getArgType(), ADD_TO);
CHECK_EQ(outputs[1].getArgType(), ADD_TO);
const auto seq_vec = in_seq.getSequenceIds().vector<int, Device>();
const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
const auto out_grad_mat = in_seq.matrix<Device>();
auto in_grad_mat =
!out_seq.data() ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
......@@ -312,10 +311,10 @@ public:
const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
const auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
CHECK(in_seq.data() && out_seq.data() && in_seq.getSequenceIds().data());
CHECK(in_seq.data() && out_seq.data() && in_seq.getSequenceId().data());
CHECK_EQ(static_cast<int>(out_seq.shape().ndims()), 2);
CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
CHECK_EQ(static_cast<int>(in_seq.getSequenceIds().shape().ndims()), 1);
CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
/// output layer grad dim == input layer grad dim * context_length_
CHECK_EQ(in_seq.shape().ndims(), out_seq.shape().ndims() * context_length_);
/// input and output has the same batch_size
......@@ -323,7 +322,7 @@ public:
CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
const auto out_grad_mat = in_seq.matrix<Device>();
const auto seq_vec = in_seq.getSequenceIds().vector<int, Device>();
const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
auto in_grad_mat = out_seq.matrix<Device>();
ContextProjectionBackwardData<Device>(
......@@ -360,16 +359,16 @@ public:
CHECK_EQ(1, static_cast<int>(outputs.size()));
CHECK(inputs[0].isSequenceArg()) << "SequenceArg required here";
const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
CHECK(in_seq.data() && in_seq.getSequenceIds().data() && outputs[0].data());
CHECK(in_seq.data() && in_seq.getSequenceId().data() && outputs[0].data());
CHECK_EQ(static_cast<int>(outputs[0].shape().ndims()), 2);
CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
CHECK_EQ(static_cast<int>(in_seq.getSequenceIds().shape().ndims()), 1);
CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
CHECK_EQ(in_seq.shape()[0], outputs[0].shape()[0]);
/// output layer grad dim == weight dim * context_length_
CHECK_EQ(in_seq.shape()[1], outputs[0].shape()[1] * context_length_);
CHECK_EQ(outputs[0].getArgType(), ADD_TO);
const auto seq_vec = in_seq.getSequenceIds().vector<int, Device>();
const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
const auto out_grad_mat = in_seq.matrix<Device>();
auto w_grad_mat = outputs[0].matrix<Device>();
ContextProjectionBackwardWeight<Device>(out_grad_mat,
......
paddle/function/CrossMapNormalOp.cpp
浏览文件 @ 297f06c6
......@@ -188,8 +188,13 @@ public:
CHECK(inputs[0].shape() == inputs[3].shape());
CHECK(inputs[0].shape() == outputs[0].shape());
// TODO(hedaoyuan): need support ASSIGN_TO mode.
CHECK_EQ(outputs[0].getArgType(), ADD_TO);
if (outputs[0].getArgType() != ADD_TO) {
// Currently, some algorithm implementations are ASSIGN_TO mode,
// if need to support the ADD_TO calculation, need to clear the output.
typename Tensor<real, Device>::Vector tmp(
outputs[0].shape().getElements(), outputs[0].data<real>());
tmp.zero();
}
size_t samples = inputs[0].shape()[0];
size_t channels = inputs[0].shape()[1];
......
paddle/function/CrossMapNormalOpTest.cpp
浏览文件 @ 297f06c6
......@@ -27,15 +27,19 @@ TEST(CrossMapNormal, real) {
<< " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW
<< " size=" << size;
FunctionCompare compare("CrossMapNormal",
FuncConfig()
.set("size", size)
.set("scale", (real)1.5)
.set("pow", (real)0.5));
Dims dims{numSamples, channels, imgSizeH, imgSizeW};
compare.cmpWithArg({Tensor(nullptr, dims)},
{Tensor(nullptr, dims), Tensor(nullptr, dims)},
{});
// init Test object
FunctionCompare test("CrossMapNormal",
FuncConfig()
.set("size", size)
.set("scale", (real)1.5)
.set("pow", (real)0.5));
// prepare input arguments
TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape));
// run Function
test.run();
}
}
}
......@@ -53,18 +57,19 @@ TEST(CrossMapNormalGrad, real) {
<< " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW
<< " size=" << size;
FunctionCompare compare("CrossMapNormalGrad",
FuncConfig()
.set("size", size)
.set("scale", (real)1.5)
.set("pow", (real)0.5));
Dims dims{numSamples, channels, imgSizeH, imgSizeW};
compare.cmpWithArg({Tensor(nullptr, dims),
Tensor(nullptr, dims),
Tensor(nullptr, dims),
Tensor(nullptr, dims)},
{Tensor(nullptr, dims)},
{});
FunctionCompare test("CrossMapNormalGrad",
FuncConfig()
.set("size", size)
.set("scale", (real)1.5)
.set("pow", (real)0.5));
TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape));
// run Function
test.run();
}
}
}
......
paddle/function/Function.cpp
浏览文件 @ 297f06c6
......@@ -79,21 +79,25 @@ FuncConfig& FuncConfig::set<bool>(const std::string& key, bool v) {
void BufferArgs::addArg(const Matrix& arg,
const TensorShape& shape,
ArgType argType) {
args_.push_back(std::make_shared<BufferArg>(arg, shape, argType));
_args_.push_back(new BufferArg(arg, shape, argType));
addArg(*_args_.back());
}
void BufferArgs::addArg(const CpuSparseMatrix& arg, ArgType argType) {
args_.push_back(std::make_shared<SparseMatrixArg>(arg, argType));
_args_.push_back(new SparseMatrixArg(arg, argType));
addArg(*_args_.back());
}
void BufferArgs::addArg(const GpuSparseMatrix& arg, ArgType argType) {
args_.push_back(std::make_shared<SparseMatrixArg>(arg, argType));
_args_.push_back(new SparseMatrixArg(arg, argType));
addArg(*_args_.back());
}
void BufferArgs::addArg(const Matrix& matrix,
const IVector& vector,
ArgType argType) {
args_.push_back(std::make_shared<SequenceArg>(matrix, vector, argType));
_args_.push_back(new SequenceArg(matrix, vector, argType));
addArg(*_args_.back());
}
ClassRegistrar<FunctionBase> FunctionBase::funcRegistrar_;
......
paddle/function/Function.h
浏览文件 @ 297f06c6
......@@ -50,19 +50,44 @@ protected:
* Argument type for Function::calc().
* A BufferArgs contains a set of BufferArg,
* because Function can have multiple inputs and outputs.
*
* addArg() with Matix object used to adapt Layer Argument.
* Will create a BufferArg object in addArg(),
* and free in destructor of BufferArgs.
*
* addArg() with BufferArg object, just save BufferArg object address,
* and the caller needs to guarantee the validity of the BufferArg object
* in the BufferArgs life time.
*/
class BufferArgs {
public:
BufferArgs() {}
~BufferArgs() {
for (auto arg : _args_) {
delete arg;
}
}
size_t size() const { return args_.size(); }
// add argument into BufferArgs
// Tensor can be Matrix, Vector, IVector.
// For inputs, do not need argType.
// For outputs, the argType needs to be specified as ASSIGN_TO or ADD_TO.
template <typename Tensor>
void addArg(const Tensor& arg, ArgType argType = UNSPECIFIED) {
args_.push_back(std::make_shared<BufferArg>(arg, argType));
void addArg(const Matrix& arg, ArgType argType = UNSPECIFIED) {
_args_.push_back(new BufferArg(arg, argType));
addArg(*_args_.back());
}
void addArg(const Vector& arg, ArgType argType = UNSPECIFIED) {
_args_.push_back(new BufferArg(arg, argType));
addArg(*_args_.back());
}
void addArg(const IVector& arg, ArgType argType = UNSPECIFIED) {
_args_.push_back(new BufferArg(arg, argType));
addArg(*_args_.back());
}
// Add arg into BufferArgs and reshape the arg.
......@@ -87,14 +112,27 @@ public:
return *args_[num];
}
void addArg(BufferArg& arg) { args_.push_back(&arg); }
void addArg(SequenceIdArg& arg) { args_.push_back(&arg); }
void addArg(SequenceArg& arg) { args_.push_back(&arg); }
void addArg(SparseMatrixArg& arg) { args_.push_back(&arg); }
private:
std::vector<BufferArgPtr> args_;
std::vector<BufferArg*> args_;
// The BufferArg object is constructed and freed by BufferArgs.
std::vector<BufferArg*> _args_;
};
/**
* \brief Base class for Function.
* The basic Function implementation requires override init and calc interfaces.
*
* The caller needs to ensure the validity of the arguments
* during Function execution.
*
* Function inputs are readonly, Function outputs have two modes: ASSIGN_TO
* and ADD_TO.
* If output.getArgType() == ASSIGN_TO, this is assign mode, and the calculation
......
paddle/function/FunctionTest.cpp
浏览文件 @ 297f06c6
......@@ -14,6 +14,7 @@ limitations under the License. */
#include "Function.h"
#include <gtest/gtest.h>
#include "paddle/math/SparseMatrix.h"
namespace paddle {
......@@ -56,4 +57,110 @@ TEST(Function, BufferArgs) {
Function<DEVICE_TYPE_GPU>(gpuArgments);
}
/**
* Some tests case are used to check the consistency between the BufferArg type
* argument received by Function and the original type argument.
*
* Use Case:
* TEST() {
* Matrix matrix(...);
* CheckBufferArg lambda = [=](const BufferArg& arg) {
* // check matrix and arg are equivalent
* EXPECT_EQ(matrix, arg);
* }
*
* BufferArgs argments{matrix...};
* std::vector<CheckBufferArg> checkFunc{lambda...};
* testBufferArgs(argments, checkFunc);
* }
*/
typedef std::function<void(const BufferArg&)> CheckBufferArg;
void testBufferArgs(const BufferArgs& inputs,
const std::vector<CheckBufferArg>& check) {
EXPECT_EQ(inputs.size(), check.size());
for (size_t i = 0; i < inputs.size(); i++) {
check[i](inputs[i]);
}
}
void testBufferArgs(const BufferArgs& inputs, const CheckBufferArg& check) {
EXPECT_EQ(inputs.size(), 1);
check(inputs[0]);
}
TEST(Arguments, Matrix) {
MatrixPtr matrix = Matrix::create(100, 200);
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 2);
EXPECT_EQ(arg.shape()[0], 100);
EXPECT_EQ(arg.shape()[1], 200);
EXPECT_EQ(arg.data(), matrix->getData());
EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getHeight(), matrix->getHeight());
EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getWidth(), matrix->getWidth());
EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getData(), matrix->getData());
};
BufferArgs argments;
argments.addArg(*matrix);
std::vector<CheckBufferArg> checkFunc;
checkFunc.push_back(check);
testBufferArgs(argments, checkFunc);
}
TEST(Arguments, Vector) {
VectorPtr vector = Vector::create(100, false);
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 1);
EXPECT_EQ(arg.shape()[0], 100);
EXPECT_EQ(arg.data(), vector->getData());
CpuVector inVector = arg.vector<real, DEVICE_TYPE_CPU>();
EXPECT_EQ(inVector.getSize(), vector->getSize());
EXPECT_EQ(inVector.getData(), vector->getData());
};
BufferArgs argments;
argments.addArg(*vector);
std::vector<CheckBufferArg> checkFunc;
checkFunc.push_back(check);
testBufferArgs(argments, checkFunc);
}
TEST(Arguments, CpuSparseMatrix) {
CpuSparseMatrix sparse(200, 300, 50);
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 2);
EXPECT_EQ(arg.shape()[0], 200);
EXPECT_EQ(arg.shape()[1], 300);
EXPECT_EQ(arg.data(), sparse.getData());
// CHECK_EQ(arg.sparse().nnz(), 50);
// CHECK_EQ(arg.sparse().dataFormat(), SPARSE_CSR_FORMAT);
// CHECK_EQ(arg.sparse().dataType(), SPARSE_FLOAT_VALUE);
EXPECT_EQ(arg.sparse().getRowBuf(), sparse.getRows());
EXPECT_EQ(arg.sparse().getColBuf(), sparse.getCols());
};
BufferArgs argments;
argments.addArg(sparse);
std::vector<CheckBufferArg> checkFunc;
checkFunc.push_back(check);
testBufferArgs(argments, checkFunc);
}
TEST(Arguments, BufferArg) {
BufferArg arg(nullptr, VALUE_TYPE_FLOAT, {1, 2, 3});
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 3);
EXPECT_EQ(arg.shape()[0], 1);
EXPECT_EQ(arg.shape()[1], 2);
EXPECT_EQ(arg.shape()[2], 3);
};
BufferArgs argments;
argments.addArg(arg);
testBufferArgs(argments, check);
}
} // namespace paddle
paddle/function/FunctionTest.h
浏览文件 @ 297f06c6
......@@ -15,57 +15,186 @@ limitations under the License. */
#include "Function.h"
#include "paddle/math/Vector.h"
#include "paddle/math/tests/TensorCheck.h"
#include "paddle/testing/TestUtil.h"
namespace paddle {
typedef std::shared_ptr<BufferArg> BufferArgPtr;
/**
* \brief A class for comparing CPU and GPU implementations of Function.
*
*
* Use case:
* // Initializes a test object, the corresponding cpu and gpu Function
* // are constructed according to FunctionName and FuncConfig.
* FunctionCompare test(FunctionName, FuncConfig);
* // Prepare inputs and outputs arguments.
* // Here the input and output can not contain real data,
* // only contains the argument type and shape.
* test.addInputs(input1);
* test.addInputs(input2);
* test.addOutputs(output1);
* test.addOutputs(output2);
* // Run.
* // Will according to the type and shape of arguments(inputs_/outputs_),
* // automatic initialization cpu and gpu function required arguments
* // (cpuInputs_/cpuOutputs_/gpuInputs_/gpuOutputs_).
* // Call the CPU and GPU Function calculation results.
* // Compares CPU and GPU calculation results for consistency.
* test.run();
*/
class FunctionCompare {
public:
FunctionCompare(const std::string& name, const FuncConfig& config)
: cpu(FunctionBase::funcRegistrar_.createByType(name + "-CPU")),
gpu(FunctionBase::funcRegistrar_.createByType(name + "-GPU")) {
cpu->init(config);
gpu->init(config);
: cpuFunc_(FunctionBase::funcRegistrar_.createByType(name + "-CPU")),
gpuFunc_(FunctionBase::funcRegistrar_.createByType(name + "-GPU")) {
cpuFunc_->init(config);
gpuFunc_->init(config);
}
void cmpWithArg(const BufferArgs& inputs,
const BufferArgs& outputs,
const BufferArgs& inouts) {
// init cpu and gpu arguments
auto initArgs = [=](
BufferArgs& cpuArgs, BufferArgs& gpuArgs, const BufferArgs& inArgs) {
/// leave it empty to pass the compile of ContextProjectionTest
/// Daoyuan is working on FunctionTest
/// and I will further merge with it
};
initArgs(cpuInputs, gpuInputs, inputs);
initArgs(cpuOutputs, gpuOutputs, outputs);
~FunctionCompare() {}
// input need only contains shape, do not contains data.
void addInputs(const BufferArg& input) {
size_t size =
input.shape().getElements() * sizeOfValuType(input.valueType());
cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
cpuInputs_.emplace_back(std::make_shared<BufferArg>(
cpuMemory_.back()->getBuf(), input.valueType(), input.shape()));
gpuInputs_.emplace_back(std::make_shared<BufferArg>(
gpuMemory_.back()->getBuf(), input.valueType(), input.shape()));
}
// output need only contains shape, do not contains data.
void addOutputs(const BufferArg& output) {
size_t size =
output.shape().getElements() * sizeOfValuType(output.valueType());
cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
cpuOutputs_.emplace_back(
std::make_shared<BufferArg>(cpuMemory_.back()->getBuf(),
output.valueType(),
output.shape(),
ASSIGN_TO));
gpuOutputs_.emplace_back(
std::make_shared<BufferArg>(gpuMemory_.back()->getBuf(),
output.valueType(),
output.shape(),
ASSIGN_TO));
}
void addInputs(const SequenceArg& input) {
size_t batchSize = input.shape()[0];
size_t numSeqs = batchSize / 10 + 1;
size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32);
cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(sizeId));
gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(sizeId));
TensorShape seqsId({numSeqs + 1});
// void* cpuBuffer = cpuMemory_.back()->getBuf();
// void* gpuBuffer = gpuMemory_.back()->getBuf();
size_t size =
input.shape().getElements() * sizeOfValuType(input.valueType());
cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
// TODO: need be implemented.
}
void run() {
// prepare cpu/gpu arguments
initInputs();
// function calculate
cpu->calc(cpuInputs, cpuOutputs);
gpu->calc(gpuInputs, gpuOutputs);
auto callFunction = [](FunctionBase* function,
std::vector<BufferArgPtr>& inputs,
std::vector<BufferArgPtr>& outputs) {
BufferArgs inArgs;
BufferArgs outArgs;
for (auto arg : inputs) {
inArgs.addArg(*arg);
}
for (auto arg : outputs) {
outArgs.addArg(*arg);
}
function->calc(inArgs, outArgs);
};
callFunction(cpuFunc_.get(), cpuInputs_, cpuOutputs_);
callFunction(gpuFunc_.get(), gpuInputs_, gpuOutputs_);
// check outputs and inouts
auto checkArgs = [=](const BufferArgs& cpuArgs, const BufferArgs& gpuArgs) {
/// leave it open
};
checkArgs(cpuOutputs, gpuOutputs);
compareOutputs();
}
std::shared_ptr<FunctionBase> getCpuFunction() const { return cpu; }
std::shared_ptr<FunctionBase> getCpuFunction() const { return cpuFunc_; }
std::shared_ptr<FunctionBase> getGpuFunction() const { return gpuFunc_; }
std::shared_ptr<FunctionBase> getGpuFunction() const { return gpu; }
protected:
void initInputs() {
for (size_t i = 0; i < cpuInputs_.size(); i++) {
initArg(*cpuInputs_[i]);
// TODO: Need a BufferCopy used to copy from one BufferArg to another.
CpuVector cpuVector(cpuInputs_[i]->shape().getElements(),
(real*)cpuInputs_[i]->data());
GpuVector gpuVector(gpuInputs_[i]->shape().getElements(),
(real*)gpuInputs_[i]->data());
gpuVector.copyFrom(cpuVector);
}
}
void compareOutputs() {
for (size_t i = 0; i < cpuOutputs_.size(); i++) {
// TODO, Need a BufferCheck used to compare the two buffers.
auto cpu = cpuOutputs_[i];
auto gpu = gpuOutputs_[i];
CpuVector cpuVector(cpu->shape().getElements(), (real*)cpu->data());
GpuVector gpuVector(cpu->shape().getElements(), (real*)gpu->data());
autotest::TensorCheckErr(cpuVector, gpuVector);
}
}
// only init cpu argument, gpu argument copy from cpu argument.
void initArg(BufferArg& arg) {
CpuVector vector(arg.shape().getElements(), (real*)arg.data());
vector.uniform(0.001, 1);
}
void initArg(SequenceIdArg& arg, size_t batchSize) {
size_t numSeqs = arg.numSeqs();
int* buf = reinterpret_cast<int*>(arg.data());
int pos = 0;
size_t maxLen = 2 * batchSize / numSeqs;
for (int i = 0; i < (int)numSeqs; ++i) {
int len = uniformRandom(
std::min<int64_t>(maxLen, batchSize - pos - numSeqs + i)) +
1;
buf[i] = pos;
pos += len;
VLOG(1) << " len=" << len;
}
buf[numSeqs] = batchSize;
}
protected:
std::shared_ptr<FunctionBase> cpu;
std::shared_ptr<FunctionBase> gpu;
std::vector<CpuMemHandlePtr> cpuMemory;
std::vector<GpuMemHandlePtr> gpuMemory;
BufferArgs cpuInputs;
BufferArgs cpuOutputs;
BufferArgs cpuInouts;
BufferArgs gpuInputs;
BufferArgs gpuOutputs;
BufferArgs gpuInouts;
std::shared_ptr<FunctionBase> cpuFunc_;
std::shared_ptr<FunctionBase> gpuFunc_;
std::vector<CpuMemHandlePtr> cpuMemory_;
std::vector<GpuMemHandlePtr> gpuMemory_;
std::vector<BufferArgPtr> cpuInputs_;
std::vector<BufferArgPtr> cpuOutputs_;
std::vector<BufferArgPtr> gpuInputs_;
std::vector<BufferArgPtr> gpuOutputs_;
};
} // namespace paddle
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