Skip to content

N
NiuTrans.Tensor
Commit 137df5d9 by liyinqiao
Options

Merge with HU Chi branch. 

1. Offer global setting for enableGrad.
2. Update XList.
3. Minor error fixed.
parent b3dfbdcd
正在显示 包含 223 行增加 和 243 行删除
source/tensor/XCall.h
......@@ -26,6 +26,9 @@
namespace nts { // namespace nts(NiuTrans.Tensor)
/* default settings */
#define X_ENABLE_GRAD false
/*
* we define the "new and delete" functions below
*/
......@@ -38,13 +41,13 @@ void InitTensorV2(XTensor * tensor,
/* initialize a dense XTensor */
void InitTensor(XTensor * tensor,
const int myOrder, const int * myDimSize, const TENSOR_DATA_TYPE myDataType = X_FLOAT,
const int myDevID = -1, const bool isEnableGrad = true);
const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* initialize a scalar V2 */
void InitTensor0DV2(XTensor * tensor, const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, XMem * myMem = NULL);
/* initialize a scalar */
void InitTensor0D(XTensor * tensor, const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
void InitTensor0D(XTensor * tensor, const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* initialize a dense vector V2 */
void InitTensor1DV2(XTensor * tensor, const int num,
......@@ -52,7 +55,7 @@ void InitTensor1DV2(XTensor * tensor, const int num,
/* initialize a dense vector */
void InitTensor1D(XTensor * tensor, const int num,
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* initialize a dense matrix V2 */
void InitTensor2DV2(XTensor * tensor, const int rowNum, const int colNum,
......@@ -60,7 +63,7 @@ void InitTensor2DV2(XTensor * tensor, const int rowNum, const int colNum,
/* initialize a dense matrix */
void InitTensor2D(XTensor * tensor, const int rowNum, const int colNum,
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* initialize a dense 3d tensor V2 */
void InitTensor3DV2(XTensor * tensor, const int d0, const int d1, const int d2,
......@@ -68,7 +71,7 @@ void InitTensor3DV2(XTensor * tensor, const int d0, const int d1, const int d2,
/* initialize a dense 3d tensor */
void InitTensor3D(XTensor * tensor, const int d0, const int d1, const int d2,
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* initialize a dense 4d tensor V2 */
void InitTensor4DV2(XTensor * tensor, const int d0, const int d1, const int d2, const int d3,
......@@ -76,7 +79,7 @@ void InitTensor4DV2(XTensor * tensor, const int d0, const int d1, const int d2,
/* initialize a dense 4d tensor */
void InitTensor4D(XTensor * tensor, const int d0, const int d1, const int d2, const int d3,
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* initialize a dense 5d tensor V2 */
void InitTensor5DV2(XTensor * tensor, const int d0, const int d1, const int d2, const int d3, const int d4,
......@@ -84,7 +87,7 @@ void InitTensor5DV2(XTensor * tensor, const int d0, const int d1, const int d2,
/* initialize a dense 5d tensor */
void InitTensor5D(XTensor * tensor, const int d0, const int d1, const int d2, const int d3, const int d4,
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* initialize a tensor with a reference tensor V2 */
void InitTensorV2(XTensor * tensor, const XTensor * reference);
......@@ -104,7 +107,7 @@ XTensor * NewTensorV2(const int myOrder, const int * myDimSize, const TENSOR_DAT
/* generate a dense XTensor */
XTensor * NewTensor(const int myOrder, const int * myDimSize, const TENSOR_DATA_TYPE myDataType = X_FLOAT,
const int myDevID = -1, const bool isEnableGrad = true);
const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a XTensor which allocates data on the buffer V2 */
XTensor * NewTensorBufV2(const int myOrder, const int * myDimSize,
......@@ -113,26 +116,26 @@ XTensor * NewTensorBufV2(const int myOrder, const int * myDimSize,
/* generate a dense XTensor which allocates data on the buffer */
XTensor * NewTensorBuf(const int myOrder, const int * myDimSize,
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a XTensor which allocates data on the buffer V2 */
XTensor * NewTensorBufV2(const XTensor * reference, int devID, XMem * myMem);
/* generate a XTensor which allocates data on the buffer */
XTensor * NewTensorBuf(const XTensor * reference, int devID, const bool isEnableGrad = true);
XTensor * NewTensorBuf(const XTensor * reference, int devID, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a scalar V2 */
XTensor * NewTensor0DV2(const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, XMem * myMem = NULL);
/* generate a scalar */
XTensor * NewTensor0D(const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
XTensor * NewTensor0D(const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a dense vector V2 */
XTensor * NewTensor1DV2(const int num, const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1,
XMem * myMem = NULL);
/* generate a dense vector */
XTensor * NewTensor1D(const int num, const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = true);
XTensor * NewTensor1D(const int num, const TENSOR_DATA_TYPE myDataType = X_FLOAT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a dense matrix V2 */
XTensor * NewTensor2DV2(const int rowNum, const int colNum,
......@@ -142,7 +145,7 @@ XTensor * NewTensor2DV2(const int rowNum, const int colNum,
/* generate a dense matrix */
XTensor * NewTensor2D(const int rowNum, const int colNum,
const TENSOR_DATA_TYPE myDataType = X_FLOAT,
const int myDevID = -1, const bool isEnableGrad = true);
const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a dense 3d tensor V2 */
XTensor * NewTensor3DV2(const int d0, const int d1, const int d2,
......@@ -152,7 +155,7 @@ XTensor * NewTensor3DV2(const int d0, const int d1, const int d2,
/* generate a dense 3d tensor */
XTensor * NewTensor3D(const int d0, const int d1, const int d2,
const TENSOR_DATA_TYPE myDataType = X_FLOAT,
const int myDevID = -1, const bool isEnableGrad = true);
const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a dense 4d tensor V2 */
XTensor * NewTensor4DV2(const int d0, const int d1, const int d2, const int d3,
......@@ -162,7 +165,7 @@ XTensor * NewTensor4DV2(const int d0, const int d1, const int d2, const int d3,
/* generate a dense 4d tensor */
XTensor * NewTensor4D(const int d0, const int d1, const int d2, const int d3,
const TENSOR_DATA_TYPE myDataType = X_FLOAT,
const int myDevID = -1, const bool isEnableGrad = true);
const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a dense 5d tensor V2 */
XTensor * NewTensor5DV2(const int d0, const int d1, const int d2, const int d3, const int d4,
......@@ -172,10 +175,10 @@ XTensor * NewTensor5DV2(const int d0, const int d1, const int d2, const int d3,
/* generate a dense 5d tensor */
XTensor * NewTensor5D(const int d0, const int d1, const int d2, const int d3, const int d4,
const TENSOR_DATA_TYPE myDataType = X_FLOAT,
const int myDevID = -1, const bool isEnableGrad = true);
const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a dense vector by range */
XTensor * NewTensorRange(int lower, int upper, int step, const TENSOR_DATA_TYPE myDataType = X_INT, const int myDevID = -1, const bool isEnableGrad = true);
XTensor * NewTensorRange(int lower, int upper, int step, const TENSOR_DATA_TYPE myDataType = X_INT, const int myDevID = -1, const bool isEnableGrad = X_ENABLE_GRAD);
/* generate a copy of XTensor (with a reference to a given tensor) */
XTensor * NewTensor(const XTensor * a, bool isFilledData = true);
......
source/tensor/XList.cpp
......@@ -249,26 +249,6 @@ inline int TensorListBase<T>::FindFirst(const T& item)
return -1;
}
template <>
inline int TensorListBase<Example>::FindFirst(const Example& item)
{
for (int i = 0; i < count; i++) {
if (item.id == items[i].id)
return i;
}
return -1;
}
template <>
inline int TensorListBase<Result>::FindFirst(const Result& item)
{
for (int i = 0; i < count; i++) {
if (item.id == items[i].id)
return i;
}
return -1;
}
/* clear the data array */
template <typename T>
void TensorListBase<T>::Clear()
......@@ -383,8 +363,7 @@ template struct TensorListBase<long>;
template struct TensorListBase<float>;
template struct TensorListBase<short>;
template struct TensorListBase<XTensor*>;
template struct TensorListBase<Result>;
template struct TensorListBase<Example>;
template struct TensorListBase<uint64_t>;
template struct TensorListBase<void*>;
} /* end of the nts (NiuTrans.Tensor) namespace */
\ No newline at end of file
source/tensor/XList.h
......@@ -26,6 +26,8 @@
#include "XMem.h"
#include "XGlobal.h"
#include <cstdint>
#ifndef __TensorList_H__
#define __TensorList_H__
......@@ -118,7 +120,14 @@ public:
void Shuffle(int nround = 10, int beg = -1, int len = 0);
/* short */
T& operator[] (int i) { return GetItem(i); };
T& operator[] (int i) {
CheckNTErrors(i >= -count && i < count, "Index of a list item is out of scope!");
CheckNTErrors(count > 0, "Cannt index the item in an empty list!");
if (i < 0)
return items[count + i];
else
return items[i];
};
T& Get(int i) { return GetItem(i); };
void Set(int i, T item) { SetItem(i, item); };
};
......@@ -132,19 +141,7 @@ typedef TensorListBase<char*> StrList;
typedef TensorListBase<long> LongList;
typedef TensorListBase<float> FloatList;
typedef TensorListBase<short> ShortList;
struct Example {
int id;
IntList data;
};
struct Result {
int id;
IntList data;
};
typedef TensorListBase<Result> ResultList;
typedef TensorListBase<Example> ExampleList;
typedef TensorListBase<uint64_t> UInt64List;
typedef TensorListBase<XTensor*> TensorList;
} /* end of the nts (NiuTrans.Tensor) namespace */
......
source/tensor/XTensor.cpp
......@@ -64,7 +64,7 @@
#endif
/* the nts (NiuTrans.Tensor) namespace */
namespace nts{
namespace nts {
int tensorIDGlobal = 0;
MUTEX_HANDLE tensorMutex;
......@@ -73,7 +73,7 @@ XTensor NULLTensor;
/* generate a tensor id */
int MakeTensorID()
{
if(tensorIDGlobal == 0)
if (tensorIDGlobal == 0)
MUTEX_INIT(tensorMutex);
MUTEX_LOCK(tensorMutex);
......@@ -97,7 +97,7 @@ XTensor::XTensor()
}
/* constructor */
XTensor::XTensor(const XTensor * reference)
XTensor::XTensor(const XTensor* reference)
{
Init();
SetDataPointer();
......@@ -112,7 +112,7 @@ constructor
>> myDevID - device id
>> myMem - memory pool used to allocating the data array
*/
XTensor::XTensor(const int myOrder, int myDevID, XMem * myMem)
XTensor::XTensor(const int myOrder, int myDevID, XMem* myMem)
{
CheckNTErrors((myOrder >= 0), "Illegal tensor order!");
......@@ -134,8 +134,8 @@ constructor
>> myDevID - device id
>> myMem - memory pool used to allocating the data array
*/
XTensor::XTensor(const int myOrder, const int * myDimSize, const TENSOR_DATA_TYPE myDataType,
const float myDenseRatio, int myDevID, XMem * myMem)
XTensor::XTensor(const int myOrder, const int* myDimSize, const TENSOR_DATA_TYPE myDataType,
const float myDenseRatio, int myDevID, XMem* myMem)
{
Init();
SetDataPointer();
......@@ -145,12 +145,12 @@ XTensor::XTensor(const int myOrder, const int * myDimSize, const TENSOR_DATA_TYP
mem = myMem;
devID = myMem != NULL ? myMem->devID : myDevID;
if(order >= 0)
if (order >= 0)
Resize(myOrder, myDimSize, myDataType, myDenseRatio);
}
/* copy constructor */
XTensor::XTensor(const XTensor &reference)
XTensor::XTensor(const XTensor& reference)
{
Init();
SetDataPointer();
......@@ -159,7 +159,7 @@ XTensor::XTensor(const XTensor &reference)
data = NULL;
dataHost = NULL;
if(reference.isTmp){
if (reference.isTmp) {
devID = reference.devID;
mem = reference.mem;
data = reference.data;
......@@ -172,16 +172,16 @@ XTensor::XTensor(const XTensor &reference)
This is VERY tricky and there might be better solutions :) */
*reference.dataP = NULL;
}
else{
else {
devID = reference.devID;
mem = reference.mem;
InitTensorV2(this, &reference);
_CopyValues(&reference, this);
}
if(reference.isTmp)
if (reference.isTmp)
XLink::Replace(&reference, this);
else{
else {
CheckNTErrors(outgo.tailNum == 0, "The node has outgoing edge to other nodes!");
XLink::CopyIncoming(&reference, this);
}
......@@ -191,7 +191,7 @@ XTensor::XTensor(const XTensor &reference)
}
/* copy constructor (with right value reference) */
XTensor::XTensor(const XTensor &&reference)
XTensor::XTensor(const XTensor&& reference)
{
Init();
SetDataPointer();
......@@ -225,12 +225,12 @@ XTensor::~XTensor()
the connectivity of the graph. To kill memory
leak, we release the data of the new tensor
when its parent is deleted (see ClearIncoming). */
if(outgo.tailNum > 0){
if (outgo.tailNum > 0) {
int dims[MAX_TENSOR_DIM_NUM];
memcpy(dims, dimSize, order * sizeof(int));
dims[0] = -dims[0];
XTensor * newTensor = new XTensor(order, dims, dataType, denseRatio, devID, mem);
XTensor* newTensor = new XTensor(order, dims, dataType, denseRatio, devID, mem);
newTensor->SetTMPFlag();
newTensor->data = data;
data = NULL;
......@@ -243,12 +243,12 @@ XTensor::~XTensor()
DestroyData();
if(grad != NULL)
if (grad != NULL)
delete grad;
}
/* set the name of the tensor */
void XTensor::SetName(const char * myName)
void XTensor::SetName(const char* myName)
{
strcpy(name, myName);
}
......@@ -280,7 +280,7 @@ void XTensor::Init()
isTmp = false;
isGrad = false;
isVar = false;
enableGrad = true;
enableGrad = X_ENABLE_GRAD;
visitMark = 0;
grad = NULL;
}
......@@ -288,16 +288,16 @@ void XTensor::Init()
/* delete data arrays */
void XTensor::DestroyData()
{
if(data != NULL && mem == NULL && !isShared)
if (data != NULL && mem == NULL && !isShared)
XMemFree(devID, data);
else if(data != NULL && isInGlobalMem)
else if (data != NULL && isInGlobalMem)
FreeData(this, mem);
else if(data != NULL)
else if (data != NULL)
mem->Release(data, GetDataSizeInChar(), signature);
data = NULL;
if(dataHost != NULL)
if (dataHost != NULL)
delete[] (char*)dataHost;
dataHost = NULL;
}
......@@ -307,7 +307,7 @@ shallow copy of the tensor
Note that we do not copy data array here
>> tensor - the source tensor
*/
void XTensor::ShallowCopy(const XTensor &tensor)
void XTensor::ShallowCopy(const XTensor& tensor)
{
strcpy(name, tensor.name);
order = tensor.order;
......@@ -330,12 +330,12 @@ XTensor& XTensor::operator= (const XTensor& tensor)
{
/* we must make a hard copy of the tensor if it is the input
of another node. */
if(outgo.tailNum > 0){
if (outgo.tailNum > 0) {
int dims[MAX_TENSOR_DIM_NUM];
memcpy(dims, dimSize, order * sizeof(int));
dims[0] = -dims[0];
XTensor * newTensor = new XTensor(order, dims, dataType, denseRatio, devID, mem);
XTensor* newTensor = new XTensor(order, dims, dataType, denseRatio, devID, mem);
newTensor->SetTMPFlag();
newTensor->data = data;
newTensor->dataHost = dataHost;
......@@ -350,35 +350,35 @@ XTensor& XTensor::operator= (const XTensor& tensor)
dataHost = NULL;
}
if(false && !tensor.isTmp){
if (false && !tensor.isTmp) {
/* NOTE: this might lead to additional data copy by Mac LLVM compilers */
/* we make an identity transformation here */
if(outgo.tailNum > 0)
if (outgo.tailNum > 0)
XLink::ClearOutgoing(this);
XLink::ClearIncoming(this);
if(!_IsSameShaped(this, &tensor))
if (!_IsSameShaped(this, &tensor))
Resize(tensor.order, tensor.dimSize, tensor.dataType, tensor.denseRatio);
_Identity(&tensor, this);
XLink::MakeLink(&tensor, NULL, this, FUNC_IDENTITY);
}
else{
else {
/* hard copy of the data array */
int size = unitNum * unitSize;
if( isInit && !isSparse && !tensor.isSparse &&
if (isInit && !isSparse && !tensor.isSparse &&
size == tensor.unitNum * tensor.unitSize &&
((devID < 0 && tensor.devID < 0) && devID == tensor.devID) &&
data != NULL)
{
XMemCopy(data, devID, tensor.data, tensor.devID, size);
if(dataHost != NULL && tensor.dataHost != NULL)
if (dataHost != NULL && tensor.dataHost != NULL)
XMemCopy(dataHost, -1, tensor.dataHost, tensor.devID, size);
}
else{
else {
DestroyData();
if(!isInit){
if (!isInit) {
devID = tensor.devID;
mem = tensor.mem;
}
......@@ -407,12 +407,12 @@ XTensor& XTensor::operator= (const XTensor&& tensor)
{
/* we must make a hard copy of the tensor if it is the input
of another node. */
if(outgo.tailNum > 0){
if (outgo.tailNum > 0) {
int dims[MAX_TENSOR_DIM_NUM];
memcpy(dims, dimSize, order * sizeof(int));
dims[0] = -dims[0];
XTensor * newTensor = new XTensor(order, dims, dataType, denseRatio, devID, mem);
XTensor* newTensor = new XTensor(order, dims, dataType, denseRatio, devID, mem);
newTensor->SetTMPFlag();
newTensor->data = data;
newTensor->dataHost = dataHost;
......@@ -500,7 +500,7 @@ XTensor XTensor::operator/ (const XTensor& tensor) const
/* overloading of the division-sign */
XTensor XTensor::operator/ (const DTYPE scale) const
{
return ScaleAndShift(*this, (DTYPE)1/scale, 0);
return ScaleAndShift(*this, (DTYPE)1.0F / scale, 0);
}
/*
......@@ -518,7 +518,7 @@ relocate the data on the target device
>> myDevId - target device id
>> myMem - memory pool on the target device
*/
void XTensor::SetDevice(int myDevId, XMem * myMem)
void XTensor::SetDevice(int myDevId, XMem* myMem)
{
if (myMem == NULL) {
myMem = GMems.GetMem(myDevId);
......@@ -527,9 +527,9 @@ void XTensor::SetDevice(int myDevId, XMem * myMem)
isInGlobalMem = false;
}
bool XTensor::IsReduceShaped(const XTensor * a, const XTensor * b, int dim)
bool XTensor::IsReduceShaped(const XTensor* a, const XTensor* b, int dim)
{
if(a == NULL || b == NULL)
if (a == NULL || b == NULL)
return false;
if ((a->order - 1) != b->order)
......@@ -541,18 +541,18 @@ bool XTensor::IsReduceShaped(const XTensor * a, const XTensor * b, int dim)
return false;
}
else if (i >= dim) {
if (a->dimSize[i+1] != b->dimSize[i])
if (a->dimSize[i + 1] != b->dimSize[i])
return false;
}
}
if(a->dataType != b->dataType)
if (a->dataType != b->dataType)
return false;
if(a->denseRatio != b->denseRatio)
if (a->denseRatio != b->denseRatio)
return false;
if(a->isSparse != b->isSparse)
if (a->isSparse != b->isSparse)
return false;
return true;
......@@ -562,7 +562,7 @@ bool XTensor::IsReduceShaped(const XTensor * a, const XTensor * b, int dim)
set the size of each dimension
>> myDimSize - size of each dimension
*/
void XTensor::SetDim(int * myDimSize)
void XTensor::SetDim(int* myDimSize)
{
for (int i = 0; i < order; i++) {
dimSize[i] = myDimSize[i];
......@@ -579,7 +579,7 @@ int XTensor::GetDim(const int dim) const
CheckNTErrors(dim >= -order, "dimenision is out of range!");
int d = dim;
if(dim < 0)
if (dim < 0)
d = order + dim;
return dimSize[d];
......@@ -590,12 +590,12 @@ reshape the tensor
>> myOrder - order of the tensor
>> myDimSize - size of each dimension
*/
void XTensor::Reshape(const int myOrder, const int * myDimSize)
void XTensor::Reshape(const int myOrder, const int* myDimSize)
{
int dims[MAX_TENSOR_DIM_NUM];
int num = 1;
for(int i = 0; i < myOrder; i++){
for (int i = 0; i < myOrder; i++) {
num *= myDimSize[i];
dims[i] = abs(myDimSize[i]);
}
......@@ -663,7 +663,7 @@ XTensor XTensor::TypeAs(const XTensor input)
/* get the number of items in the data array */
int XTensor::GetSize() const
{
if(isSparse)
if (isSparse)
return unitNumNonZero;
else
return unitNum;
......@@ -672,13 +672,13 @@ int XTensor::GetSize() const
/* get the size of the memory space used */
int XTensor::GetDataSizeInChar() const
{
if(isSparse){
if (isSparse) {
int num = int(unitNum * denseRatio + 1);
int tupleSize = sizeof(int)+sizeof(DTYPE);
int size = sizeof(int) + tupleSize*(num);
int tupleSize = sizeof(int) + sizeof(DTYPE);
int size = sizeof(int) + tupleSize * (num);
return size;
}
else{
else {
return unitNum * unitSize;
}
}
......@@ -690,15 +690,15 @@ get unit size in terms of "dataType"
*/
int XTensor::GetUnitSize(TENSOR_DATA_TYPE myDataType) const
{
if(myDataType == X_INT)
if (myDataType == X_INT)
return sizeof(int);
else if(myDataType == X_FLOAT)
else if (myDataType == X_FLOAT)
return sizeof(float);
else if(myDataType == X_DOUBLE)
else if (myDataType == X_DOUBLE)
return sizeof(double);
else if(myDataType == X_INT8)
else if (myDataType == X_INT8)
return 1;
else if(myDataType == X_FLOAT16)
else if (myDataType == X_FLOAT16)
return 2;
return sizeof(float);
}
......@@ -737,21 +737,21 @@ MTYPE XTensor::GetOffset3D(int d0, int d1, int d2) const
a vector with all entries of 0
>> stream - stream for the job pipeline
*/
void XTensor::SetZeroAll(XStream * stream)
void XTensor::SetZeroAll(XStream* stream)
{
if(data == NULL)
if (data == NULL)
return;
if(isSparse){
if(devID >= 0){
if (isSparse) {
if (devID >= 0) {
#ifdef USE_CUDA
int size = sizeof(int) + (sizeof(int)+sizeof(DTYPE)) * unitNumNonZero;
int size = sizeof(int) + (sizeof(int) + sizeof(DTYPE)) * unitNumNonZero;
int devIDBackup = 0;
cudaGetDevice(&devIDBackup);
cudaSetDevice(devID);
if(stream == NULL)
if (stream == NULL)
cudaMemset(data, 0, size);
else
cudaMemsetAsync(data, 0, size, stream->stream);
......@@ -764,14 +764,14 @@ void XTensor::SetZeroAll(XStream * stream)
unitNumNonZero = 0;
}
else{
if(devID >= 0){
else {
if (devID >= 0) {
#ifdef USE_CUDA
int devIDBackup = 0;
cudaGetDevice(&devIDBackup);
cudaSetDevice(devID);
if(stream == NULL)
if (stream == NULL)
cudaMemset(data, 0, unitNum * unitSize);
else
cudaMemsetAsync(data, 0, unitNum * unitSize, stream->stream);
......@@ -789,9 +789,9 @@ void XTensor::SetZeroAll(XStream * stream)
>> num - number of data items
>> beg - where we start the data copy in the data array of the tensor
*/
void XTensor::SetData(const void * d, int num, int beg)
void XTensor::SetData(const void* d, int num, int beg)
{
if (data == NULL || d ==NULL)
if (data == NULL || d == NULL)
return;
CheckNTErrors(!isSparse, "TODO");
......@@ -830,7 +830,7 @@ void XTensor::SetDataRand(DTYPE lower, DTYPE upper)
// srand((unsigned)time(0));
DTYPE variance = upper - lower;
void * d = NULL;
void* d = NULL;
if (dataType == X_FLOAT) {
d = new float[unitNum];
for (int i = 0; i < unitNum; i++) {
......@@ -868,12 +868,12 @@ double GaussRand(DTYPE mean, DTYPE standardDeviation)
double z;
double pi = 3.141592654;
if (phase == 0){
if (phase == 0) {
u = (rand() + 1.0) / (RAND_MAX + 1.0);
v = (rand() + 1.0) / (RAND_MAX + 1.0);
z = sqrt(-2.0 * log(u))* sin(2.0 * pi * v);
z = sqrt(-2.0 * log(u)) * sin(2.0 * pi * v);
}
else{
else {
z = sqrt(-2.0 * log(u)) * cos(2.0 * pi * v);
}
......@@ -894,7 +894,7 @@ void XTensor::SetDataRandn(DTYPE mean, DTYPE standardDeviation)
return;
// srand((unsigned)time(0));
void * d = NULL;
void* d = NULL;
if (dataType == X_FLOAT) {
d = new float[unitNum];
for (int i = 0; i < unitNum; i++) {
......@@ -927,7 +927,7 @@ set tensor items with an array of offsets
>> value - value for the data items
>> num - number of the data items
*/
void XTensor::SetDataBatched(MTYPE * offsets, DTYPE value, int num)
void XTensor::SetDataBatched(MTYPE* offsets, DTYPE value, int num)
{
_SetDataWithOffset(this, offsets, value, num);
}
......@@ -938,7 +938,7 @@ set tensor items with an array of values
>> values - value for each data item
>> num - number of the data items
*/
void XTensor::SetDataBatchedWithValues(MTYPE * offsets, void * values, int num)
void XTensor::SetDataBatchedWithValues(MTYPE* offsets, void* values, int num)
{
_SetDataWithOffsetAndValue(this, offsets, values, num);
}
......@@ -974,7 +974,7 @@ DTYPE XTensor::Get(int offset) const
CheckNTErrors(data != NULL, "Cannot use an uninitialized tensor!");
CheckNTErrors(denseRatio == 1.0F, "Only dense tensors are supported in Get(offset).");
DTYPE * address = (DTYPE*)data + offset;
DTYPE* address = (DTYPE*)data + offset;
return ToCPU(devID, address);
}
......@@ -985,25 +985,25 @@ get the pointer to a cell
>> size - size of index
<< return - pointer to the cell
*/
void * XTensor::GetCell(int index[], int size) const
void* XTensor::GetCell(int index[], int size) const
{
CheckNTErrors((size == order), "Illegal index!");
int offset = index[0];
for(int i = 1; i < size; ++i){
for (int i = 1; i < size; ++i) {
CheckNTErrors((index[i] < dimSize[i]), "Index is out of range!");
offset = offset * dimSize[i] + index[i];
}
if(isSparse){
if (isSparse) {
DTYPE value;
void * p;
if(BinarySearch(offset, value, p))
void* p;
if (BinarySearch(offset, value, p))
return (char*)p + sizeof(int);
else
return NULL;
}
else{
else {
return ((char*)data) + offset * unitSize;
}
}
......@@ -1018,7 +1018,7 @@ DTYPE XTensor::Get0D() const
CheckNTErrors((dataType == DEFAULT_DTYPE), "The tensor is not in default type.");
int dims[1] = {0};
void * value = GetCell(dims, 0);
void* value = GetCell(dims, 0);
return ToCPU(devID, value);
}
......@@ -1035,7 +1035,7 @@ DTYPE XTensor::Get1D(int i) const
CheckNTErrors((dataType == DEFAULT_DTYPE), "The tensor is not in default type.");
int dims[1] = {i};
void * value = GetCell(dims, 1);
void* value = GetCell(dims, 1);
return ToCPU(devID, value);
}
......@@ -1054,7 +1054,7 @@ DTYPE XTensor::Get2D(int ni, int mi) const
CheckNTErrors((dataType == DEFAULT_DTYPE), "The tensor is not in default type.");
int dims[2] = {ni, mi};
void * value = GetCell(dims, 2);
void* value = GetCell(dims, 2);
return ToCPU(devID, value);
}
......@@ -1067,14 +1067,14 @@ get the value of a cell in a 3d tensor
*/
DTYPE XTensor::Get3D(int d0, int d1, int d2) const
{
CheckNTErrors((order == 3), "Cannot get a 2d cell for a tensor whose order is not 3!");
CheckNTErrors((order == 3), "Cannot get a 3d cell for a tensor whose order is not 3!");
CheckNTErrors((d0 >= 0 && d0 < dimSize[0]), "dimension 0 is out of range!");
CheckNTErrors((d1 >= 0 && d1 < dimSize[1]), "dimension 1 is out of range!");
CheckNTErrors((d2 >= 0 && d2 < dimSize[2]), "dimension 2 is out of range!");
CheckNTErrors((dataType == DEFAULT_DTYPE), "The tensor is not in default type.");
int dims[3] = {d0, d1, d2};
void * value = GetCell(dims, 3);
void* value = GetCell(dims, 3);
return ToCPU(devID, value);
}
......@@ -1090,7 +1090,7 @@ int XTensor::GetInt(int offset) const
CheckNTErrors(data != NULL, "Cannot use an uninitialized tensor!");
CheckNTErrors(denseRatio == 1.0F, "Only dense tensors are supported in Get(offset).");
int * address = (int*)data + offset;
int* address = (int*)data + offset;
return ToCPUInt(devID, address);
}
......@@ -1105,7 +1105,7 @@ int XTensor::Get0DInt() const
CheckNTErrors(dataType == X_INT, "The tensor is not in int type.");
int dims[1] = {0};
void * value = GetCell(dims, 0);
void* value = GetCell(dims, 0);
return ToCPUInt(devID, value);
}
......@@ -1122,7 +1122,7 @@ int XTensor::Get1DInt(int i) const
CheckNTErrors(dataType == X_INT, "The tensor is not in int type.");
int dims[1] = {i};
void * value = GetCell(dims, 1);
void* value = GetCell(dims, 1);
return ToCPUInt(devID, value);
}
......@@ -1133,7 +1133,7 @@ get the value of a cell in a 2d tensor in int type
>> mi - column index
<< return - value of cell(ni, mi) in int
*/
int XTensor::Get2DInt(int ni, int mi) const
int XTensor::Get2DInt(int ni, int mi) const
{
CheckNTErrors(order == 2, "Cannot get a 2d cell for a tensor whose order is not 2!");
CheckNTErrors(ni >= 0 && ni < dimSize[0], "dimension 0 is out of range!");
......@@ -1141,7 +1141,7 @@ get the value of a cell in a 2d tensor in int type
CheckNTErrors(dataType == X_INT, "The tensor is not in default type.");
int dims[2] = {ni, mi};
void * value = GetCell(dims, 2);
void* value = GetCell(dims, 2);
return ToCPUInt(devID, value);
}
......@@ -1155,14 +1155,14 @@ get the value of a cell in a 3d tensor in int type
*/
int XTensor::Get3DInt(int d0, int d1, int d2) const
{
CheckNTErrors(order == 3, "Cannot get a 2d cell for a tensor whose order is not 3!");
CheckNTErrors(order == 3, "Cannot get a 3d cell for a tensor whose order is not 3!");
CheckNTErrors(d0 >= 0 && d0 < dimSize[0], "dimension 0 is out of range!");
CheckNTErrors(d1 >= 0 && d1 < dimSize[1], "dimension 1 is out of range!");
CheckNTErrors(d2 >= 0 && d2 < dimSize[2], "dimension 2 is out of range!");
CheckNTErrors(dataType == X_INT, "The tensor is not in default type.");
int dims[3] = {d0, d1, d2};
void * value = GetCell(dims, 3);
void* value = GetCell(dims, 3);
return ToCPUInt(devID, value);
}
......@@ -1177,8 +1177,8 @@ DTYPE XTensor::GetInSparse(int i) const
CheckNTErrors(i >= 0 && i < unitNum, "Index is out of range!");
CheckNTErrors(dataType == DEFAULT_DTYPE, "The tensor is not in default type.");
char * d = (char*)data + sizeof(int);
DTYPE * value = (DTYPE*)(d + (sizeof(int) + sizeof(DTYPE)) * i + sizeof(int));
char* d = (char*)data + sizeof(int);
DTYPE* value = (DTYPE*)(d + (sizeof(int) + sizeof(DTYPE)) * i + sizeof(int));
return ToCPU(devID, value);
}
......@@ -1193,8 +1193,8 @@ int XTensor::GetKeyInSparse(int i) const
CheckNTErrors(i >= 0 && i < unitNum, "Index is out of range!");
CheckNTErrors(dataType == DEFAULT_DTYPE, "The tensor is not in default type.");
char * d = (char*)data + sizeof(int);
int * key = (int*)(d + (sizeof(int) + sizeof(DTYPE)) * i);
char* d = (char*)data + sizeof(int);
int* key = (int*)(d + (sizeof(int) + sizeof(DTYPE)) * i);
return ToCPUInt(devID, key);
}
......@@ -1222,7 +1222,7 @@ bool XTensor::Set(DTYPE value, int offset)
CheckNTErrors(offset >= 0 && offset < unitNum, "Invalid index!");
CheckNTErrors(data != NULL, "Cannot use an uninitialized tensor!");
DTYPE * d = (DTYPE*)data + offset;
DTYPE* d = (DTYPE*)data + offset;
return SetToDevice(devID, d, value);
}
......@@ -1288,7 +1288,7 @@ set the value of a cell in a 3d tensor in default type
*/
bool XTensor::Set3D(DTYPE value, int d0, int d1, int d2)
{
CheckNTErrors(order == 3, "Cannot get a 2d cell for a tensor whose order is not 3!");
CheckNTErrors(order == 3, "Cannot get a 3d cell for a tensor whose order is not 3!");
CheckNTErrors(d0 >= 0 && d0 < dimSize[0], "dimension 0 is out of range!");
CheckNTErrors(d1 >= 0 && d1 < dimSize[1], "dimension 1 is out of range!");
CheckNTErrors(d2 >= 0 && d2 < dimSize[2], "dimension 2 is out of range!");
......@@ -1309,7 +1309,7 @@ bool XTensor::SetInt(int value, int offset)
CheckNTErrors(offset >= 0 && offset < unitNum, "Invalid index!");
CheckNTErrors(data != NULL, "Cannot use an uninitialized tensor!");
int * d = (int*)data + offset;
int* d = (int*)data + offset;
return SetToDeviceInt(devID, d, value);
}
......@@ -1390,7 +1390,7 @@ set the integer value of a cell in a 3d tensor in default type
*/
bool XTensor::Set3DInt(int value, int d0, int d1, int d2)
{
CheckNTErrors(order == 3, "Cannot get a 2d cell for a tensor whose order is not 3!");
CheckNTErrors(order == 3, "Cannot get a 3d cell for a tensor whose order is not 3!");
CheckNTErrors(d0 >= 0 && d0 < dimSize[0], "dimension 0 is out of range!");
CheckNTErrors(d1 >= 0 && d1 < dimSize[1], "dimension 1 is out of range!");
CheckNTErrors(d2 >= 0 && d2 < dimSize[2], "dimension 2 is out of range!");
......@@ -1408,15 +1408,15 @@ increase the value of a cell in a 2d tensor
>> mi - column index
<< return - succeeded or not
*/
bool XTensor::Add2D(DTYPE value, int ni, int mi)
bool XTensor::Add2D(DTYPE value, int ni, int mi)
{
CheckNTErrors(ni >= 0 && ni < dimSize[0], "the row index is out of range!");
CheckNTErrors(mi >= 0 && mi < dimSize[1], "the column index is out of range!");
CheckNTErrors(dataType == DEFAULT_DTYPE, "The tensor is not in default type.");
CheckNTErrors(isSparse == false, "TODO!");
if(devID < 0){
DTYPE * p = (DTYPE*)data + ni * dimSize[1] + mi;
if (devID < 0) {
DTYPE* p = (DTYPE*)data + ni * dimSize[1] + mi;
CheckNTErrors((p != NULL), "No data array is found!");
......@@ -1424,7 +1424,7 @@ increase the value of a cell in a 2d tensor
return true;
}
else{
else {
int dims[2] = {ni, mi};
return SetToDevice(devID, GetCell(dims, 2), Get2D(ni, mi) + value);
}
......@@ -1433,24 +1433,24 @@ increase the value of a cell in a 2d tensor
/* get the number of non-zero elements (in a sparse tensor) */
int XTensor::GetNonzeroSize() const
{
if(!isSparse){
if (!isSparse) {
XPRINT(1, stderr, "WARNING! Counting non-zero elements in a dense tensor might be slow!\n");
CheckNTErrors(devID < 0, "TODO");
if(dataType == DEFAULT_DTYPE){
if (dataType == DEFAULT_DTYPE) {
int count = 0;
for(int i = 0; i < unitNum; i++){
for (int i = 0; i < unitNum; i++) {
DTYPE value = *(DTYPE*)((char*)data + i * sizeof(DTYPE));
if(value == 0)
if (value == 0)
count++;
}
return count;
}
else{
else {
ShowNTErrors("TODO!");
return -1;
}
}
else{
else {
/* return the head of the tuple list */
return unitNumNonZero;
}
......@@ -1481,7 +1481,7 @@ set the tensor as "variable"
void XTensor::SetVarFlag(bool myIsVar)
{
isVar = myIsVar;
if(isVar)
if (isVar)
SetGradFlag(true);
}
......@@ -1493,11 +1493,11 @@ resize a tensor with a specified tensor size
>> myDenseRatio - how often an element has non-zero value
<< return - succeeded or not
*/
bool XTensor::Resize(const int myOrder, const int * myDimSize,
bool XTensor::Resize(const int myOrder, const int* myDimSize,
const TENSOR_DATA_TYPE myDataType, const float myDenseRatio)
{
/* free old mem */
if(data != NULL){
if (data != NULL) {
if (mem == NULL)
XMemFree(devID, data);
else
......@@ -1513,11 +1513,11 @@ bool XTensor::Resize(const int myOrder, const int * myDimSize,
bool filledData = true;
bool zeroData = false;
for(int i = 0; i < order; i++){
for (int i = 0; i < order; i++) {
dimSize[i] = abs(myDimSize[i]);
if(myDimSize[i] < 0)
if (myDimSize[i] < 0)
filledData = false;
if(myDimSize[i] == 0)
if (myDimSize[i] == 0)
zeroData = true;
unitNum *= dimSize[i];
}
......@@ -1528,17 +1528,17 @@ bool XTensor::Resize(const int myOrder, const int * myDimSize,
dataType = myDataType;
unitSize = GetUnitSize(dataType);
if(myDataType != DEFAULT_DTYPE)
if (myDataType != DEFAULT_DTYPE)
isDefaultDType = false;
else
isDefaultDType = true;
if(zeroData){
if (zeroData) {
unitNum = 0;
return false;
}
if(isSparse){
if (isSparse) {
/*
for sparse matrices, we use a list of tuple (key, value),
ordered by key. Take a (2-dimensional) matrix as an example,
......@@ -1557,21 +1557,21 @@ bool XTensor::Resize(const int myOrder, const int * myDimSize,
*/
int num = int(unitNum * denseRatio + 1);
int tupleSize = sizeof(int)+sizeof(DTYPE);
int size = sizeof(int) + tupleSize*(num);
int tupleSize = sizeof(int) + sizeof(DTYPE);
int size = sizeof(int) + tupleSize * (num);
if(filledData){
int * d = NULL;
if (filledData) {
int* d = NULL;
if(mem == NULL){
if (mem == NULL) {
d = new int[size];
memset(d, 0, size);
}
else{
else {
d = (int*)mem->Alloc(mem->devID, size);
}
if(d == NULL)
if (d == NULL)
return false;
#if !defined(UNSAFE_BUT_FAST_MEM)
......@@ -1581,10 +1581,10 @@ bool XTensor::Resize(const int myOrder, const int * myDimSize,
}
return true;
}
else{
if(filledData){
else {
if (filledData) {
/* allocate the new one */
if(mem == NULL){
if (mem == NULL) {
data = XMemAlloc(devID, unitNum * unitSize);
#if defined(UNSAFE_BUT_FAST_MEM)
XMemSet(devID, data, 0, unitNum * unitSize);
......@@ -1593,12 +1593,12 @@ bool XTensor::Resize(const int myOrder, const int * myDimSize,
else
data = (void*)mem->Alloc(mem->devID, unitNum * unitSize);
if(data == NULL)
if (data == NULL)
return false;
}
#if !defined(UNSAFE_BUT_FAST_MEM)
if(data != NULL)
if (data != NULL)
XMem::SetZero(data, unitNum * unitSize, mem);
#endif
return true;
......@@ -1609,12 +1609,12 @@ bool XTensor::Resize(const int myOrder, const int * myDimSize,
resize a tensor by another
>> myTensor - tensor for reference
*/
bool XTensor::Resize(const XTensor * myTensor)
bool XTensor::Resize(const XTensor* myTensor)
{
denseRatio = myTensor->denseRatio;
TENSOR_DATA_TYPE myDataType = myTensor->dataType;
if(myDataType != DEFAULT_DTYPE)
if (myDataType != DEFAULT_DTYPE)
isDefaultDType = false;
else
isDefaultDType = true;
......@@ -1630,14 +1630,14 @@ binary search to find an element in a sparse tensor
it is the previous one if there is no hit
<< return - found it or not?
*/
bool XTensor::BinarySearch(int key, DTYPE &value, void * &position) const
bool XTensor::BinarySearch(int key, DTYPE& value, void*& position) const
{
CheckNTErrors((isSparse), "A sparse tensor is required!");
CheckNTErrors((dataType == DEFAULT_DTYPE), "The tensor is not in the default type.");
int * d = (int*)data;
int* d = (int*)data;
if(key < 0 || *d == 0){
if (key < 0 || *d == 0) {
value = 0;
position = NULL;
return false;
......@@ -1647,37 +1647,37 @@ bool XTensor::BinarySearch(int key, DTYPE &value, void * &position) const
int high = *d - 1;
int last = -1;
bool ok = false;
int * k = NULL;
int* k = NULL;
int headSize = sizeof(int);
int tupleSize = sizeof(int)+sizeof(DTYPE);
char * p = (char*)data + headSize;
int tupleSize = sizeof(int) + sizeof(DTYPE);
char* p = (char*)data + headSize;
while (low <= high){
int mid = low + (high-low)/2;
while (low <= high) {
int mid = low + (high - low) / 2;
k = (int*)(p + tupleSize * mid);
if (*k == key){
if (*k == key) {
ok = true;
high = mid -1;
high = mid - 1;
break;
}
else if(*k > key){
high = mid -1;
else if (*k > key) {
high = mid - 1;
}
else{
low = mid +1;
else {
low = mid + 1;
last = mid;
}
}
if(ok){
DTYPE * p = (DTYPE*)((char*)k + sizeof(int));
if (ok) {
DTYPE* p = (DTYPE*)((char*)k + sizeof(int));
value = *p;
position = k;
return true;
}
else{
else {
value = 0;
if(last == -1)
if (last == -1)
position = NULL;
else
position = (char*)data + headSize + tupleSize * last;
......@@ -1693,12 +1693,12 @@ dump data to a file
>> beg - the first item id
>> verbose - verbose level
*/
void XTensor::Dump(FILE * file, const char * label, const int n, const int beg, const int verbose)
void XTensor::Dump(FILE* file, const char* label, const int n, const int beg, const int verbose)
{
if (verbose > verboseLevel)
return;
void * d = data;
void* d = data;
bool isNewData = false;
#ifdef USE_CUDA
......@@ -1716,7 +1716,7 @@ void XTensor::Dump(FILE * file, const char * label, const int n, const int beg,
num *= dimSize[i];
num = int(num * denseRatio + 1);
int tupleSize = sizeof(int) + sizeof(DTYPE);
int size = sizeof(int) + tupleSize*(num);
int size = sizeof(int) + tupleSize * (num);
d = new char[size];
memset(d, 0, size);
......@@ -1731,7 +1731,7 @@ void XTensor::Dump(FILE * file, const char * label, const int n, const int beg,
if (label != NULL)
fprintf(file, "%s ", label);
if(isInit){
if (isInit) {
fprintf(file, "order=%d dimsize=", order);
if (order == 0) {
fprintf(file, "%d,", dimSize[0]);
......@@ -1742,21 +1742,21 @@ void XTensor::Dump(FILE * file, const char * label, const int n, const int beg,
fprintf(file, ",");
}
}
else{
else {
fprintf(file, "order=-1 dimsize=-1");
}
fprintf(file, " dtype=%s dense=%f\n", GetDataTypeName(dataType), denseRatio);
if(!isInit){
if (!isInit) {
fprintf(file, "NULL");
}
if (!isSparse) {
if (dataType == DEFAULT_DTYPE) {
int end = MIN(n > 0 ? beg + n : beg + unitNum, unitNum);
for(int i = beg; i < end; i++){
for (int i = beg; i < end; i++) {
DTYPE f = ((DTYPE*)d)[i];
if(i == beg)
if (i == beg)
fprintf(file, "%e", f);
else
fprintf(file, " %e", f);
......@@ -1765,9 +1765,9 @@ void XTensor::Dump(FILE * file, const char * label, const int n, const int beg,
}
else if (dataType == X_INT) {
int end = MIN(n > 0 ? beg + n : beg + unitNum, unitNum);
for(int i = beg; i < end; i++){
for (int i = beg; i < end; i++) {
int f = ((int*)d)[i];
if(i == beg)
if (i == beg)
fprintf(file, "%d", f);
else
fprintf(file, " %d", f);
......@@ -1807,7 +1807,7 @@ dump data to a file
>> beg - the first item id
>> verbose - verbose level
*/
void XTensor::Dump(const XTensor * tensor, FILE * file, const char * label, const int n, const int beg, const int verbose)
void XTensor::Dump(const XTensor* tensor, FILE* file, const char* label, const int n, const int beg, const int verbose)
{
XTensor a(tensor->order, tensor->dimSize, tensor->dataType, tensor->denseRatio, tensor->devID, tensor->mem);
_CopyValues(tensor, &a);
......@@ -1839,7 +1839,7 @@ read data from a file
>> file - where to load the data
>> label - label of the tensor
*/
void XTensor::Read(FILE * file, const char * label)
void XTensor::Read(FILE* file, const char* label)
{
char typeName[32] = "";
char dimSizeName[128] = "";
......@@ -1872,7 +1872,7 @@ void XTensor::Read(FILE * file, const char * label)
int o = 0;
bool sameSize = true;
char * p = dimSizeName;
char* p = dimSizeName;
while (*p != 0) {
while (*p == ' ' || *p == '\t')
p++;
......@@ -1896,14 +1896,14 @@ void XTensor::Read(FILE * file, const char * label)
if (!sameSize || dRatio > denseRatio || GetDataType(typeName) != dataType)
Resize(dimNum, dims, GetDataType(typeName), dRatio);
void * dataBuf = XMemAlloc(-1, GetDataSizeInChar());
void * dataBackup = data;
void* dataBuf = XMemAlloc(-1, GetDataSizeInChar());
void* dataBackup = data;
data = dataBuf;
if (!isSparse) {
if (dataType == DEFAULT_DTYPE) {
for (int i = 0; i < unitNum; i++) {
DTYPE * f = ((DTYPE*)data) + i;
DTYPE* f = ((DTYPE*)data) + i;
if (fscanf(file, "%e", f) < 1) {
ShowNTErrors("Incorrect tensor format!");
}
......@@ -1956,13 +1956,13 @@ void XTensor::BinaryRead(FILE* file, size_t offset)
fseek(file, offset, 0);
switch (dataType) {
case X_INT: {
int * d = new int[unitNum];
int* d = new int[unitNum];
fread(d, sizeof(int), unitNum, file);
SetData(d, unitNum);
delete[] d;
}
default: {
float * d = new float[unitNum];
float* d = new float[unitNum];
fread(d, sizeof(float), unitNum, file);
SetData(d, unitNum);
delete[] d;
......@@ -1974,7 +1974,7 @@ void XTensor::BinaryRead(FILE* file, size_t offset)
flush the data to the target device
>> targetMem - memory pool on the target device
*/
void XTensor::FlushToMem(XMem * targetMem)
void XTensor::FlushToMem(XMem* targetMem)
{
if (targetMem == NULL)
return;
......@@ -1987,7 +1987,7 @@ void XTensor::FlushToMem(XMem * targetMem)
CudaCPUToGPUFlush(&l, targetMem->devID, targetMem);
}
else if (mem != targetMem) {
void * tmpData = targetMem->Alloc(targetMem->devID, GetDataSizeInChar());
void* tmpData = targetMem->Alloc(targetMem->devID, GetDataSizeInChar());
XMemCopy(tmpData, targetMem->devID, data, devID, GetDataSizeInChar());
data = tmpData;
mem = targetMem;
......@@ -2016,24 +2016,24 @@ allocate the memory space of the tensor (in the global memory)
>> myMem - the memory pool we are using
>> useBuf - indicates whether we use the buffer in the memory pool
*/
void XTensor::AllocateData(XTensor * tensor, XMem * myMem, bool useBuf)
void XTensor::AllocateData(XTensor* tensor, XMem* myMem, bool useBuf)
{
if(tensor == NULL)
if (tensor == NULL)
return;
if(myMem == NULL){
if(tensor->data != NULL)
if (myMem == NULL) {
if (tensor->data != NULL)
FreeData(tensor, NULL, false);
tensor->data = XMemAlloc(tensor->devID, tensor->GetDataSizeInChar());
tensor->isInGlobalMem = true;
}
else{
else {
CheckNTErrors((tensor->data == NULL), "Cannot renew the space for the tensor");
if(useBuf){
if (useBuf) {
tensor->data = myMem->AllocBuf(tensor->devID, tensor->GetDataSizeInChar());
tensor->isInGlobalMem = false;
}
else{
else {
tensor->data = myMem->AllocGlobal(tensor->devID, tensor->GetDataSizeInChar());
tensor->isInGlobalMem = true;
}
......@@ -2048,16 +2048,16 @@ free the memory space of the tensor (in the global memory)
>> myMem - the memory pool we are using
>> useBuf - indicates whether we use the buffer in the memory pool
*/
void XTensor::FreeData(XTensor * tensor, XMem * myMem, bool useBuf)
void XTensor::FreeData(XTensor* tensor, XMem* myMem, bool useBuf)
{
if(tensor == NULL)
if (tensor == NULL)
return;
if(myMem == NULL){
if (myMem == NULL) {
XMemFree(tensor->devID, tensor->data);
}
else{
if(tensor->isInGlobalMem)
else {
if (tensor->isInGlobalMem)
myMem->ReleaseGlobal(tensor->devID, tensor->data);
else
myMem->ReleaseBuf(tensor->devID, tensor->GetDataSizeInChar());
......@@ -2068,27 +2068,27 @@ void XTensor::FreeData(XTensor * tensor, XMem * myMem, bool useBuf)
}
/* overloading of the plus-sign */
XTensor operator+ (const DTYPE shift, const XTensor &tensor)
XTensor operator+ (const DTYPE shift, const XTensor& tensor)
{
return ScaleAndShift(tensor, 1, shift);
}
/* overloading of the minus-sign */
XTensor operator- (const DTYPE shift, const XTensor &tensor)
XTensor operator- (const DTYPE shift, const XTensor& tensor)
{
return ScaleAndShift(tensor, 1, -shift);
}
/* overloading of the multiply-sign */
XTensor operator* (const DTYPE scale, const XTensor &tensor)
XTensor operator* (const DTYPE scale, const XTensor& tensor)
{
return ScaleAndShift(tensor, scale, 0);
}
/* overloading of the division-sign */
XTensor operator/ (const DTYPE scale, const XTensor &tensor)
XTensor operator/ (const DTYPE scale, const XTensor& tensor)
{
return ScaleAndShift(tensor, (DTYPE)1/scale, 0);
return ScaleAndShift(tensor, (DTYPE)1.0F / scale, 0);
}
} /* end of the nts (NiuTrans.Tensor) namespace */
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论