Skip to content

Matrix::MUL operators using and test Daoyuan's Paddle Function, SparseMatrixArg and Function Test #1147

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tianbingsz merged 13 commits into from
Jan 30, 2017
Merged

Matrix::MUL operators using and test Daoyuan's Paddle Function, SparseMatrixArg and Function Test #1147

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
936301f
Use and test Daoyuan's SparseMatrixArg.
Jan 13, 2017
1f0cbcf
add GpuMatrix::mul, CpuMatrix::mul operators
Jan 17, 2017
2df8eec
Pass Unit test for GpuMatrix::mul(GpuMatrix, GpuMatrix) and CpuMatrix…
Jan 17, 2017
1ca2846
Pass unit test for CpuMatrix::mul(CpuMatrix, CpuSparseMatrix)
Jan 18, 2017
4751cc8
Pass unit test for SparseCpuMatrix::mul(CpuMatrix, CpuMatrix),
Jan 18, 2017
171eaff
clean the code a little bit.
Jan 19, 2017
9ade63e
clean code a little bit.
Jan 19, 2017
316bf75
clean code in function/MulOp.cpp
Jan 19, 2017
077f936
Support SparseMatrixArg unit test using Daoyuan's new Function Test.
Jan 22, 2017
bc5d7bb
Add Sparse = dense * dense unit test with Daoyuan's Function test
Jan 22, 2017
b3be735
Daoyuan's comments.
Jan 23, 2017
999cd14
Further address Daoyuan's comments, clean the code.
Jan 24, 2017
5b1a5c1
Daoyuan's comments.
Jan 25, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 8 additions & 2 deletions paddle/function/BufferArg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,14 +32,20 @@ const SparseMatrixArg& BufferArg::sparse() const {
SparseMatrixArg::SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType)
: BufferArg(sparse, argType),
row_(reinterpret_cast<void*>(sparse.getRows()), VALUE_TYPE_INT32),
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32),
nnz_(sparse.getElementCnt()),
format_(static_cast<SparseDataFormat>(sparse.getFormat())),
type_(static_cast<SparseDataType>(sparse.getValueType())) {
bufferType_ = TENSOR_SPARSE;
}

SparseMatrixArg::SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType)
: BufferArg(sparse, argType),
row_(reinterpret_cast<void*>(sparse.getRows()), VALUE_TYPE_INT32),
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32),
nnz_(sparse.getElementCnt()),
format_(static_cast<SparseDataFormat>(sparse.getFormat())),
type_(static_cast<SparseDataType>(sparse.getValueType())) {
bufferType_ = TENSOR_SPARSE;
}

Expand Down
90 changes: 68 additions & 22 deletions paddle/function/BufferArg.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,13 +30,6 @@ enum BufferType {
TENSOR_SPARSE = 4
};

enum SparseDataType {
SPARSE_NO_VALUE = 0, // do not need value pointer, all values are 1
SPARSE_FLOAT_VALUE = 1
};

enum SparseDataFormat { SPARSE_CSR_FORMAT = 0, SPARSE_CSC_FORMAT = 1 };

class BufferArg;
class SequenceArg;
class SparseMatrixArg;
Expand Down Expand Up @@ -79,19 +72,21 @@ class BufferArg {
BufferArg(ValueType valueType,
const TensorShape& shape,
ArgType argType = UNSPECIFIED)
: buf_(nullptr),
valueType_(valueType),
shape_(shape),
argType_(argType) {}
: buf_(nullptr), valueType_(valueType), shape_(shape), argType_(argType) {
bufferType_ = TENSOR_NORMAL;
}

BufferArg(void* buf,
ValueType valueType,
const TensorShape& shape,
ArgType argType = UNSPECIFIED)
: buf_(buf), valueType_(valueType), shape_(shape), argType_(argType) {}
: buf_(buf), valueType_(valueType), shape_(shape), argType_(argType) {
bufferType_ = TENSOR_NORMAL;
}

BufferArg(void* buf, ValueType valueType)
: buf_(buf), valueType_(valueType) {}
BufferArg(void* buf, ValueType valueType) : buf_(buf), valueType_(valueType) {
bufferType_ = TENSOR_NORMAL;
}

BufferArg(const Matrix& matrix, ArgType argType = UNSPECIFIED)
: buf_(
Expand Down Expand Up @@ -167,8 +162,9 @@ class BufferArg {
ValueType valueType() const { return valueType_; }
BufferType bufferType() const { return bufferType_; }
const TensorShape& shape() const { return shape_; }
bool isSparse() const { return (TENSOR_SPARSE == bufferType_); }
bool isSparseArg() const { return TENSOR_SPARSE == bufferType_; }
bool isSequenceArg() const { return TENSOR_SEQUENCE_DATA == bufferType_; }
virtual size_t numElements() const { return shape_.getElements(); }

const SequenceArg& sequence() const;
const SparseMatrixArg& sparse() const;
Expand All @@ -179,6 +175,7 @@ class BufferArg {
TensorShape shape_;
BufferType bufferType_{TENSOR_UNKNOWN};
ArgType argType_{UNSPECIFIED};
// TODO(tianbing), add deviceType_
// leading dimensions. The size is dims_.size()
// Dims lds_;
};
Expand All @@ -191,6 +188,7 @@ class SequenceIdArg : public BufferArg {
public:
SequenceIdArg(const TensorShape& shape, ArgType argType = UNSPECIFIED)
: BufferArg(VALUE_TYPE_INT32, shape, argType) {
bufferType_ = TENSOR_SEQUENCE_ID;
CHECK_EQ(shape_.ndims(), (size_t)1);
CHECK_GT(shape_[0], 1);
numSeqs_ = shape_[0] - 1;
Expand Down Expand Up @@ -228,7 +226,9 @@ class SequenceArg : public BufferArg {
SequenceArg(ValueType valueType,
const TensorShape& shape,
ArgType argType = UNSPECIFIED)
: BufferArg(valueType, shape, argType), startPositions_(TensorShape()) {}
: BufferArg(valueType, shape, argType), startPositions_(TensorShape()) {
bufferType_ = TENSOR_SEQUENCE_DATA;
}

SequenceArg(void* buf,
ValueType valueType,
Expand Down Expand Up @@ -269,31 +269,75 @@ class SparseMatrixArg : public BufferArg {
const BufferArg& row,
const BufferArg& col,
size_t nnz,
SparseDataFormat format,
SparseDataType type,
SparseFormat format,
Copy link
Contributor

@hedaoyuan hedaoyuan Jan 22, 2017

Choose a reason for hiding this comment

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

SparseFormat和SparseValueType放在Matrix.h里面并不合适；Matrix.h和SparseMatrix.h里面的数据结构最终是不准备用的，所以可以考虑把这两个TensorType.h里面去。

Copy link
Contributor Author

@tianbingsz tianbingsz Jan 24, 2017

Choose a reason for hiding this comment

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

done

SparseValueType type,
ArgType argType = UNSPECIFIED)
: BufferArg(buf, valueType, shape, argType),
row_(row),
col_(col),
nnz_(nnz),
format_(format),
type_(type) {
format_(static_cast<SparseDataFormat>(format)),
type_(static_cast<SparseDataType>(type)) {
bufferType_ = TENSOR_SPARSE;
CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE));
CHECK_EQ(shape_.ndims(), (size_t)2);
CHECK_EQ(row_.shape().ndims(), (size_t)1);
CHECK_EQ(col_.shape().ndims(), (size_t)1);
if (format == SPARSE_CSR_FORMAT) {
if (format_ == T_SPARSE_CSR) {
CHECK_EQ(nnz, col.shape()[0]);
} else if (format == SPARSE_CSC_FORMAT) {
} else if (format_ == T_SPARSE_CSC) {
CHECK_EQ(nnz, row.shape()[0]);
}
}

SparseMatrixArg(ValueType valueType,
const TensorShape& shape,
size_t nnz,
SparseFormat format,
SparseValueType type,
ArgType argType = UNSPECIFIED)
: BufferArg(valueType, shape, argType),
row_(BufferArg(nullptr, VALUE_TYPE_INT32)),
col_(BufferArg(nullptr, VALUE_TYPE_INT32)),
nnz_(nnz),
format_(static_cast<SparseDataFormat>(format)),
type_(static_cast<SparseDataType>(type)) {
bufferType_ = TENSOR_SPARSE;
CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE));
CHECK_EQ(shape_.ndims(), (size_t)2);

/// len of row_ : height + 1 (CSR) or nnz (CSC), buf_ == nullptr
row_ = (format_ == T_SPARSE_CSR
? BufferArg(VALUE_TYPE_INT32, TensorShape{shape_[0] + 1})
: BufferArg(VALUE_TYPE_INT32, TensorShape{nnz}));
/// len of col_ : width + 1 (CSC) or nnz (CSR), buf_ == nullptr
col_ = (format_ == T_SPARSE_CSR
? BufferArg(VALUE_TYPE_INT32, TensorShape{nnz})
: BufferArg(VALUE_TYPE_INT32, TensorShape{shape_[1] + 1}));
}

SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType = UNSPECIFIED);

SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType = UNSPECIFIED);

template <DeviceType DType>
typename Tensor<real, DType>::SparseMatrix SparseMatrix() const {
CHECK(buf_);
CHECK(valueType_ == DataType<real>::value);
// CHECK(deviceType_ == DType);
CHECK_EQ(2, shape_.ndims());
return typename Tensor<real, DType>::SparseMatrix(
reinterpret_cast<real*>(buf_),
reinterpret_cast<int*>(row_.data()),
reinterpret_cast<int*>(col_.data()),
shape_[0],
shape_[1],
nnz_,
static_cast<SparseValueType>(type_),
static_cast<SparseFormat>(format_),
false);
}

~SparseMatrixArg() {}

void* getRowBuf() const { return row_.data(); }
Expand All @@ -302,6 +346,8 @@ class SparseMatrixArg : public BufferArg {

size_t nnz() const { return nnz_; }

size_t numElements() const override { return nnz_; }

SparseDataFormat dataFormat() const { return format_; }

SparseDataType dataType() const { return type_; }
Expand Down
1 change: 1 addition & 0 deletions paddle/function/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@ if(WITH_TESTING)
add_simple_unittest(FunctionTest)
add_simple_unittest(ContextProjectionOpTest)
add_simple_unittest(PadOpTest)
add_simple_unittest(MulOpTest)
endif()
endif()

Expand Down
99 changes: 89 additions & 10 deletions paddle/function/FunctionTest.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */

#include "Function.h"
#include "paddle/math/Vector.h"
#include "paddle/math/Matrix.h"
#include "paddle/math/SparseMatrix.h"
#include "paddle/math/tests/TensorCheck.h"
#include "paddle/testing/TestUtil.h"

Expand Down Expand Up @@ -69,7 +70,7 @@ class FunctionCompare {
}

// output need only contains shape, do not contains data.
void addOutputs(const BufferArg& output) {
void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
size_t size =
output.shape().getElements() * sizeOfValuType(output.valueType());
cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
Expand All @@ -79,12 +80,40 @@ class FunctionCompare {
std::make_shared<BufferArg>(cpuMemory_.back()->getBuf(),
output.valueType(),
output.shape(),
ASSIGN_TO));
argType));
gpuOutputs_.emplace_back(
std::make_shared<BufferArg>(gpuMemory_.back()->getBuf(),
output.valueType(),
output.shape(),
ASSIGN_TO));
argType));
}

/// add and init output sparse matrix
void addOutputs(const SparseMatrixArg& output, ArgType argType = ASSIGN_TO) {
cpuSparse_ = std::make_shared<CpuSparseMatrix>(
output.shape()[0],
output.shape()[1],
output.nnz(),
static_cast<SparseValueType>(output.dataType()),
static_cast<SparseFormat>(output.dataFormat()));

gpuSparse_ = std::make_shared<GpuSparseMatrix>(
output.shape()[0],
output.shape()[1],
output.nnz(),
static_cast<SparseValueType>(output.dataType()),
static_cast<SparseFormat>(output.dataFormat()));

/// init sparse matrix
hl_stream_t stream(HPPL_STREAM_1);
cpuSparse_->randomizeUniform();
gpuSparse_->copyFrom(*cpuSparse_, stream);
hl_stream_synchronize(stream);

cpuOutputs_.emplace_back(
std::make_shared<SparseMatrixArg>(*cpuSparse_, argType));
gpuOutputs_.emplace_back(
std::make_shared<SparseMatrixArg>(*gpuSparse_, argType));
}

void addInputs(const SequenceArg& input) {
Expand All @@ -107,10 +136,36 @@ class FunctionCompare {
// TODO: need be implemented.
}

void addInputs(const SparseMatrixArg& input) {
cpuSparse_ = std::make_shared<CpuSparseMatrix>(
input.shape()[0],
input.shape()[1],
input.nnz(),
static_cast<SparseValueType>(input.dataType()),
static_cast<SparseFormat>(input.dataFormat()));

gpuSparse_ = std::make_shared<GpuSparseMatrix>(
input.shape()[0],
input.shape()[1],
input.nnz(),
static_cast<SparseValueType>(input.dataType()),
static_cast<SparseFormat>(input.dataFormat()));

/// init sparse matrix
hl_stream_t stream(HPPL_STREAM_1);
cpuSparse_->randomizeUniform();
gpuSparse_->copyFrom(*cpuSparse_, stream);
hl_stream_synchronize(stream);

cpuInputs_.emplace_back(std::make_shared<SparseMatrixArg>(*cpuSparse_));
gpuInputs_.emplace_back(std::make_shared<SparseMatrixArg>(*gpuSparse_));
}

void run() {
// prepare cpu/gpu arguments
initInputs();

initOutputs();
Copy link
Contributor

@hedaoyuan hedaoyuan Jan 23, 2017

Choose a reason for hiding this comment

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

FunctionCompare可能得加个钩子，uniform(0.001, 1)可能支持不了所有Function的参数初始化。

Copy link
Contributor Author

@tianbingsz tianbingsz Jan 24, 2017 •
edited
Loading

Choose a reason for hiding this comment

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

Good idea. Will do it later for the future unit test cases.

// function calculate
auto callFunction = [](FunctionBase* function,
std::vector<BufferArgPtr>& inputs,
Expand All @@ -129,7 +184,7 @@ class FunctionCompare {
callFunction(cpuFunc_.get(), cpuInputs_, cpuOutputs_);
callFunction(gpuFunc_.get(), gpuInputs_, gpuOutputs_);

// check outputs and inouts
// check outputs
compareOutputs();
}

Expand All @@ -140,6 +195,10 @@ class FunctionCompare {
protected:
void initInputs() {
for (size_t i = 0; i < cpuInputs_.size(); i++) {
if (cpuInputs_[i]->isSparseArg()) {
continue; /// sparse matrix already init
}

initArg(*cpuInputs_[i]);

// TODO: Need a BufferCopy used to copy from one BufferArg to another.
Expand All @@ -152,14 +211,32 @@ class FunctionCompare {
}
}

void initOutputs() {
for (size_t i = 0; i < cpuOutputs_.size(); i++) {
if (cpuOutputs_[i]->isSparseArg()) {
continue; /// sparse matrix already init
}

initArg(*cpuOutputs_[i]);

// TODO: Need a BufferCopy used to copy from one BufferArg to another.
CpuVector cpuVector(cpuOutputs_[i]->shape().getElements(),
(real*)cpuOutputs_[i]->data());
GpuVector gpuVector(gpuOutputs_[i]->shape().getElements(),
(real*)gpuOutputs_[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());

const auto cpu = cpuOutputs_[i];
const auto gpu = gpuOutputs_[i];
CHECK_EQ(cpu->numElements(), gpu->numElements());
CpuVector cpuVector(cpu->numElements(), (real*)cpu->data());
GpuVector gpuVector(gpu->numElements(), (real*)gpu->data());
autotest::TensorCheckErr(cpuVector, gpuVector);
}
}
Expand Down Expand Up @@ -195,6 +272,8 @@ class FunctionCompare {
std::vector<BufferArgPtr> cpuOutputs_;
std::vector<BufferArgPtr> gpuInputs_;
std::vector<BufferArgPtr> gpuOutputs_;
std::shared_ptr<CpuSparseMatrix> cpuSparse_;
std::shared_ptr<GpuSparseMatrix> gpuSparse_;
};

} // namespace paddle
Loading