Skip to content

N
NiuTrans.Tensor
Commit f7c6fb3b by xuchen
Options

remove SumByColumnTV and SumByColumnVT function and add the Rand function for example

parent 695c20a8
正在显示 包含 43 行增加 和 1072 行删除
source/tensor/XTensor.cpp
......@@ -844,6 +844,12 @@ void XTensor::SetData(const void * d, int num, int beg)
XMemCopy((char*)data + beg * unitSize, devID, d, -1, num * unitSize);
}
/* generate data items with a uniform distribution in [0, 1] */
void XTensor::Rand(int rNum, int cNum)
{
_SetDataRand(this, rNum, cNum);
}
/*
set the tensor items by a uniform distribution in range [lower, upper]
>> lower - lower value of the range
......
source/tensor/XTensor.h
......@@ -304,6 +304,9 @@ public:
/* set the tensor with an data array */
void SetData(const void * d, int num, int beg = 0);
/* generate data items with a uniform distribution in [0, 1] */
void Rand(int rNum, int cNum);
/* set tensor items by a uniform distribution */
void SetDataRand(DTYPE lower = 0.0F, DTYPE upper = 1.0F);
......
source/tensor/core/CHeader.h
......@@ -39,8 +39,6 @@
#include "arithmetic/Sub.h"
#include "arithmetic/SubDim.h"
#include "arithmetic/Sum.h"
#include "arithmetic/SumByColumnTV.h"
#include "arithmetic/SumByColumnVT.h"
#include "arithmetic/SumDim.h"
#include "arithmetic/XTensorBLAS.h"
#include "arithmetic/MulAndShift.h"
......
source/tensor/core/arithmetic/Mask.h
......@@ -34,7 +34,7 @@ c(i) = a(i) if mask(i) is non-zero
c(i) = alpha if mask(i) = 0
where i is the index of the element
*/
void _Mask(const XTensor * a, const XTensor * mask, XTensor * c, DTYPE alpha);
void _Mask(const XTensor * a, const XTensor * mask, XTensor * c, DTYPE alpha = 0.0);
/*
mask entries of a given tensor (on site):
......@@ -42,7 +42,7 @@ a(i) = a(i) if mask(i) is non-zero
a(i) = alpha if mask(i) = 0
where i is the index of the element
*/
void _MaskMe(XTensor * a, const XTensor * mask, DTYPE alpha);
void _MaskMe(XTensor * a, const XTensor * mask, DTYPE alpha = 0.0);
/*
mask entries of a given tensor (return an XTensor structure):
......
source/tensor/core/arithmetic/Sign.cu deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: LI Yinqiao (li.yin.qiao.2012@hotmail.com) 2018-7-11
*/
#include "../../XDevice.h"
#include "../../XTensor.h"
#include "Sign.h"
#include "Sign.cuh"
namespace nts { // namespace nts(NiuTrans.Tensor)
#ifdef USE_CUDA
/*
set each entry to its sign value (CUDA Kernel)
>> a - pointer to input data array
>> b - pointer to output data array
>> size - size of the data array
*/
__global__
void KernelSign(DTYPE * a, DTYPE * b, int size)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i < size) {
if (a[i] > 0)
b[i] = 1.0F;
else if (a[i] == 0)
b[i] = 0.0F;
else
b[i] = -1.0F;
}
}
/*
set each entry to its sign value with float16 data type value (CUDA Kernel)
This is for float16 computation
>> a - pointer to input data array
>> b - pointer to output data array
>> size - size of the data array
*/
__global__
void KernelSign(__half * a, __half * b, int size)
{
return;
}
/*
set each entry to its sign value
>> a - input tensor we are processing
>> b - output tensor we are processing
*/
void _CudaSign(const XTensor * a, XTensor * b)
{
CheckNTErrors((XTensor::IsSameShaped(a, b)), "Input tensors should have the same type!");
CheckNTErrors((a->isSparse == false), "TODO!");
int gridSize[3];
int blockSize[3];
GDevs.GetCudaThread(a->devID, a->unitNum, gridSize, blockSize);
dim3 blocks(gridSize[0]);
dim3 threads(blockSize[0]);
int devIDBackup;
ProtectCudaDev(a->devID, devIDBackup);
if (a->dataType == DEFAULT_DTYPE) {
KernelSign << <blocks, threads >> >((DTYPE*)a->data, (DTYPE*)b->data, a->unitNum);
}
else if (a->dataType == X_FLOAT16) {
KernelSign << <blocks, threads >> >((__half*)a->data, (__half*)b->data, a->unitNum);
}
else {
ShowNTErrors("TODO!");
}
BacktoCudaDev(a->devID, devIDBackup);
}
#endif // USE_CUDA
} // namespace nts(NiuTrans.Tensor)
source/tensor/core/arithmetic/SumByColumnTV.cpp deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#include "../../XTensor.h"
#include "SumByColumnTV.h"
#include "SumByColumnTV.cuh"
namespace nts { // namespace nts(NiuTrans.Tensor)
/*
sum of a tensor and a vector (column vector) in a column by column manner
for each column a_col (in a block), we have
c_col = a_col + b * \beta
where b is a vector.
>> a - a tensor
>> b - a vector with the same column size with a
>> c - where we put a+b. we save it in a if c is NULL
>> beta - the scaling factor
*/
void _SumByColumnTV(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
{
CheckNTErrors((a && b && c), "Empty input tensors!");
CheckNTErrors((XTensor::IsSameShaped(a, c)), "Unmatched tensors in addition!");
CheckNTErrors((b->order == 2 && b->dimSizeRDI[0] == 1 && b->dimSizeRDI[1] == a->dimSizeRDI[1]),
"Illegal input vector size!");
int rowNum = a->dimSize[0];
int colNum = a->dimSize[1];
int blockNum = 1;
for (int i = 2; i < a->order; i++)
blockNum *= a->dimSizeRDI[i];
int blockSize = colNum * rowNum;
if (a->devID >= 0 || b->devID >= 0 || c->devID >= 0) {
#ifdef USE_CUDA
_CudaSumByColumnTV(a, b, c, beta);
#endif
}
else {
if (!a->isSparse && !b->isSparse) {
CheckNTErrors(!c->isSparse, "TODO!");
if (a->dataType == DEFAULT_DTYPE &&
b->dataType == DEFAULT_DTYPE &&
c->dataType == DEFAULT_DTYPE)
{
for (int k = 0; k < blockNum; k++) {
for (int i = 0; i < rowNum; i++) {
DTYPE * ap = (DTYPE*)a->data + k * blockSize + i * colNum;
DTYPE * bp = (DTYPE*)b->data;
DTYPE * cp = (DTYPE*)c->data + k * blockSize + i * colNum;
DTYPE v = bp[i];
for (int j = 0; j < colNum; j++)
cp[j] = ap[j] + v * beta;
}
}
}
else {
// TODO!!
ShowNTErrors("TODO!");
}
}
else {
// TODO!!
ShowNTErrors("TODO!");
}
}
}
} // namespace nts(NiuTrans.Tensor)
\ No newline at end of file
source/tensor/core/arithmetic/SumByColumnTV.cu deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#include "../../XDevice.h"
#include "../../XTensor.h"
#include "SumByColumnTV.h"
#include "SumByColumnTV.cuh"
namespace nts { // namespace nts(NiuTrans.Tensor)
#ifdef USE_CUDA
/*
summation of a tensor and a vector (column vector)
c_col = a_col + b * \beta
>> a - a tensor
>> b - a vector with the same column size with a
>> c - where we put a+b. we save it in a
>> colNum - column number (of a block)
>> blockSize - size of a block
>> size - size of the entire data array
>> beta - the scaling factor
*/
__global__
void KernelADDByColumnTV(DTYPE * a, DTYPE * b, DTYPE * c, int colNum, int blockSize, int size, DTYPE beta)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
if (i >= size)
return;
int offset = i % blockSize;
int row = offset / colNum;
c[i] = a[i] + b[row] * beta;
}
/*
summation of a tensor and a vector (column vector)
for each column a_col (in a block), we have
c_col = a_col + b * \beta
where b is a vector.
>> a - a tensor
>> b - a vector with the same column size with a
>> c - where we put a+b. we save it in a if c is NULL
>> beta - the scaling factor
*/
void _CudaSumByColumnTV(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
{
CheckNTErrors((a && b && c), "Empty input tensors!");
CheckNTErrors((XTensor::IsSameShaped(a, c)), "Unmatched tensors in addition!");
CheckNTErrors((b->order == 2 && b->dimSizeRDI[0] == 1 && b->dimSizeRDI[1] == a->dimSizeRDI[1]),
"Illegal input vector size!");
CheckNTErrors((a->dataType == DEFAULT_DTYPE && b->dataType == DEFAULT_DTYPE &&
c->dataType == DEFAULT_DTYPE), "TODO");
int rowNum = a->dimSize[0];
int colNum = a->dimSize[1];
int blockNum = 1;
for (int i = 2; i < a->order; i++)
blockNum *= a->dimSizeRDI[i];
int cudaGridSize[3];
int cudaBlockSize[3];
GDevs.GetCudaThread(c->devID, a->unitNum, cudaGridSize, cudaBlockSize);
int devIDBackup;
ProtectCudaDev(a->devID, devIDBackup);
KernelADDByColumnTV << <dim3(cudaGridSize[0]), dim3(cudaBlockSize[0]) >> >
((DTYPE*)a->data, (DTYPE*)b->data, (DTYPE*)c->data, colNum, rowNum * colNum, a->unitNum, beta);
BacktoCudaDev(a->devID, devIDBackup);
}
#endif // USE_CUDA
} // namespace nts(NiuTrans.Tensor)
\ No newline at end of file
source/tensor/core/arithmetic/SumByColumnTV.cuh deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#ifndef __REDUCEMAX_CUH__
#define __REDUCEMAX_CUH__
#include "../reduce/ReduceMax.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
#ifdef USE_CUDA
/* summation of a tensor and a vector (column vector) */
void _CudaSumByColumnTV(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta = (DTYPE)1.0);
#endif // USE_CUDA
} // namespace nts(NiuTrans.Tensor)
#endif // __REDUCEMAX_CUH__
source/tensor/core/arithmetic/SumByColumnTV.h deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#ifndef __SUMBYCOLUMNTV_H__
#define __SUMBYCOLUMNTV_H__
#include "../../XTensor.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
/* sum of a tensor and a (column) vector */
void _SumByColumnTV(const XTensor * a, const XTensor * b, XTensor * c = NULL, DTYPE beta = (DTYPE)1.0);
} // namespace nts(NiuTrans.Tensor)
#endif // __SUMBYCOLUMNTV_H__
source/tensor/core/arithmetic/SumByColumnVT.cpp deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#include "../../XTensor.h"
#include "SumByColumnVT.h"
#include "SumByColumnVT.cuh"
namespace nts { // namespace nts(NiuTrans.Tensor)
/*
sum of a vector (column vector) and a tensor in a column by column manner
for each column b_col, we have
c = a + \sum{col} b_col * \beta
where c and a are vectors, and b_col is a column in b.
>> a - a tensor
>> b - a vector with the same column size with a
>> c - where we put a+b. we save it in a if c is NULL
>> beta - the scaling factor
*/
void _SumByColumnVT(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
{
CheckNTErrors((a && b && c), "Empty input tensors!");
CheckNTErrors((XTensor::IsSameShaped(a, c)), "Unmatched tensors in addition!");
CheckNTErrors((a->order == 2 && a->dimSizeRDI[0] == 1 && b->dimSizeRDI[1] == a->dimSizeRDI[1]),
"Illegal input vector size!");
if (a->devID >= 0 || b->devID >= 0 || c->devID >= 0) {
#ifdef USE_CUDA
_CudaSumByColumnVT(a, b, c, beta);
#endif
}
else {
int rowNum = b->dimSize[0];
int colNum = b->dimSize[1];
int blockNum = 1;
for (int i = 2; i < b->order; i++)
blockNum *= b->dimSizeRDI[i];
int blockSize = colNum * rowNum;
if (!a->isSparse && !b->isSparse) {
CheckNTErrors(!c->isSparse, "TODO!");
if (a->dataType == DEFAULT_DTYPE &&
b->dataType == DEFAULT_DTYPE &&
c->dataType == DEFAULT_DTYPE)
{
for (int k = 0; k < blockNum; k++) {
for (int i = 0; i < rowNum; i++) {
DTYPE * ap = (DTYPE*)a->data;
DTYPE * bp = (DTYPE*)b->data + k * blockSize + i * colNum;
DTYPE * cp = (DTYPE*)c->data;
DTYPE sum = 0;
for (int j = 0; j < colNum; j++)
sum += bp[j];
cp[i] = ap[i] + sum * beta;
}
}
}
else {
// TODO!!
ShowNTErrors("TODO!");
}
}
else {
// TODO!!
ShowNTErrors("TODO!");
}
}
}
} // namespace nts(NiuTrans.Tensor)
\ No newline at end of file
source/tensor/core/arithmetic/SumByColumnVT.cu deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#include "../../XDevice.h"
#include "../../XTensor.h"
#include "SumByColumnVT.h"
#include "SumByColumnVT.cuh"
namespace nts { // namespace nts(NiuTrans.Tensor)
#ifdef USE_CUDA
/*
summation of a vector (column vector) and a tensor
c = a + \sum{col} b_col * \beta
>> a - a vector with the same column size with b
>> b - a tensor
>> c - where we put a+b. we save it in a
>> colNum - column number (of a block)
>> blockSize - size of a block
>> size - size of the entire data array
>> beta - the scaling factor
*/
__global__
void KernelADDByColumnVT(DTYPE * a, DTYPE * b, DTYPE * c, int colNum, int rowNum, int blockNum, DTYPE beta)
{
int row = blockDim.x * blockIdx.x + threadIdx.x;
if (row >= rowNum)
return;
DTYPE sum = 0;
for (int k = 0; k < blockNum; k++) {
DTYPE * bp = b + (rowNum * k + row) * colNum;
if (colNum % 4 == 0) {
for (int i = 0; i < colNum; i += 4)
sum += bp[i] + bp[i + 1] + bp[i + 2] + bp[i + 3];
}
else if (colNum % 2 == 0) {
for (int i = 0; i < colNum; i += 2)
sum += bp[i] + bp[i + 1];
}
else {
for (int i = 0; i < colNum; i++)
sum += bp[i];
}
__syncthreads();
}
c[row] = a[row] + beta * sum;
}
/*
summation of a vector (column vector) and a tensor
for each column b_col, we have
c = a + \sum{col} b_col * \beta
where c and a are vectors, and b_col is a column in b.
>> a - a vector with the same column size with b
>> b - a tensor
>> c - where we put a+b. we save it in a if c is NULL
>> beta - the scaling factor
*/
void _CudaSumByColumnVT(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
{
CheckNTErrors((a && b && c), "Empty input tensors!");
CheckNTErrors((XTensor::IsSameShaped(a, c)), "Unmatched tensors in addition!");
CheckNTErrors((a->order == 2 && a->dimSizeRDI[0] == 1 && b->dimSizeRDI[1] == a->dimSizeRDI[1]),
"Illegal input vector size!");
CheckNTErrors((a->dataType == DEFAULT_DTYPE && b->dataType == DEFAULT_DTYPE &&
c->dataType == DEFAULT_DTYPE), "TODO");
int rowNum = b->dimSize[0];
int colNum = b->dimSize[1];
int blockNum = 1;
for (int i = 2; i < b->order; i++)
blockNum *= b->dimSizeRDI[i];
int cudaGridSize[3];
int cudaBlockSize[3];
GDevs.GetCudaThread(c->devID, a->dimSizeRDI[1], cudaGridSize, cudaBlockSize);
int devIDBackup = 0;
ProtectCudaDev(a->devID, devIDBackup);
KernelADDByColumnVT << <dim3(cudaGridSize[0]), dim3(cudaBlockSize[0]) >> >
((DTYPE*)a->data, (DTYPE*)b->data, (DTYPE*)c->data, colNum, rowNum, blockNum, beta);
BacktoCudaDev(a->devID, devIDBackup);
}
#endif // USE_CUDA
} // namespace nts(NiuTrans.Tensor)
\ No newline at end of file
source/tensor/core/arithmetic/SumByColumnVT.cuh deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#ifndef __SUMBYCOLUMNVT_CUH__
#define __SUMBYCOLUMNVT_CUH__
#include "SumByColumnVT.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
#ifdef USE_CUDA
/* summation of a vector (column vector) and a tensor */
void _CudaSumByColumnVT(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta = (DTYPE)1.0);
#endif // USE_CUDA
} // namespace nts(NiuTrans.Tensor)
#endif // __SUMBYCOLUMNVT_CUH__
source/tensor/core/arithmetic/SumByColumnVT.h deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: XIAO Tong (email: xiaotong@mail.neu.edu.cn) 2018-04-24
*/
#ifndef __SUMBYCOLUMNVT_H__
#define __SUMBYCOLUMNVT_H__
#include "../../XTensor.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
/* sum of a (column) vector and a tensor */
void _SumByColumnVT(const XTensor * a, const XTensor * b, XTensor * c = NULL, DTYPE beta = (DTYPE)1.0);
} // namespace nts(NiuTrans.Tensor)
#endif // __SUMBYCOLUMNVT_H__
source/tensor/core/getandset/SetData.cpp
......@@ -466,13 +466,23 @@ void _SetDataLowTri(XTensor * tensor, DTYPE p, int shift)
}
}
/* generate data items with a uniform distribution in [0, 1] */
void _SetDataRand(XTensor * tensor, int rNum, int cNum)
{
if (tensor == NULL || tensor->isInit == false || tensor->order !=2 ) {
InitTensor2D(tensor, rNum, cNum);
}
_SetDataRand(tensor, 0.0F, 1.0F);
}
/*
generate data items with a uniform distribution in [lower, upper]
>> tensor - the tensor whose data array would be initialized
>> lower - lower value of the range
>> upper - upper value of the range
*/
void _SetDataRand(const XTensor * tensor, DTYPE lower, DTYPE upper)
void _SetDataRand(XTensor * tensor, DTYPE lower, DTYPE upper)
{
CheckNTErrors(upper > lower, "the high value must be greater than low value!");
......@@ -525,7 +535,7 @@ the item to a pre-defined value if the item >= p, set the item to 0 otherwise
>> p - the threshold
>> value - the value we intend to assign to the item
*/
void _SetDataRandP(const XTensor * tensor, DTYPE lower, DTYPE upper, DTYPE p, DTYPE value)
void _SetDataRandP(XTensor * tensor, DTYPE lower, DTYPE upper, DTYPE p, DTYPE value)
{
CheckNTErrors(tensor->dataType == DEFAULT_DTYPE, "TODO");
......
source/tensor/core/getandset/SetData.h
......@@ -63,12 +63,15 @@ void _SetDataIndexed(XTensor * source, XTensor * modify, int dim, int index);
/* generate data as lower triangular matrics for last two dimensions */
void _SetDataLowTri(XTensor * tensor, DTYPE p, int shift);
/* generate data items with a uniform distribution in [0, 1] */
void _SetDataRand(XTensor * tensor, int rNum, int cNum);
/* generate data items with a uniform distribution in [lower, upper] */
void _SetDataRand(const XTensor * tensor, DTYPE lower, DTYPE upper);
void _SetDataRand(XTensor * tensor, DTYPE lower, DTYPE upper);
/* generate data items with a uniform distribution in [lower, upper] and set
the item to a pre-defined value if the item >= p, set the item to 0 otherwise */
void _SetDataRandP(const XTensor * tensor, DTYPE lower, DTYPE upper, DTYPE p, DTYPE value);
void _SetDataRandP(XTensor * tensor, DTYPE lower, DTYPE upper, DTYPE p, DTYPE value);
/* generate data items with a normal distribution with specified mean and standard deviation */
void _SetDataRandN(XTensor * tensor, DTYPE mean = 0.0F, DTYPE standardDeviation = 1.0F);
......
source/tensor/core/math/Compare.cpp
......@@ -69,6 +69,7 @@ XTensor funcName(const XTensor &a, DTYPE number) \
_funcName(&a, &b, number); \
return b; \
}
// I think we needn't to make link.
// XLink::MakeLink(&a, NULL, &b, operationId);
......
source/tensor/core/shape/Merge.cpp
......@@ -273,16 +273,16 @@ void Merge(const XTensor &s, XTensor &t, int whereToMerge, int leadingDim, bool
merge small tensors into a big tensor
>> smalls - the list of the small tensors
>> big - the merged tensor (for return)
>> t - the merged tensor (for return)
>> whereToMerge - the merging operation is along with which dimension
*/
void _Merge(const TensorList * smalls, XTensor * big, int whereToMerge)
void _Merge(const TensorList * smalls, XTensor * t, int whereToMerge)
{
whereToMerge = (whereToMerge < 0 ? big->order - 1 : whereToMerge);
whereToMerge = (whereToMerge < 0 ? t->order - 1 : whereToMerge);
CheckNTErrors((smalls != NULL), "Invalid list!");
CheckNTErrors((smalls->count > 0), "Empty list!");
CheckNTErrors((whereToMerge >= 0 && whereToMerge < big->order), "Wrong range of whereToMerge");
CheckNTErrors((whereToMerge >= 0 && whereToMerge < t->order), "Wrong range of whereToMerge");
bool uniform = true;
......@@ -292,7 +292,7 @@ void _Merge(const TensorList * smalls, XTensor * big, int whereToMerge)
for (int i = 0; i < smalls->count; i++) {
XTensor* smallsItem = smalls->GetItem(i);
CheckNTErrors((big->unitNum == smallsItem->unitNum * mergeNum), "Unmatched tensors!");
CheckNTErrors((t->unitNum == smallsItem->unitNum * mergeNum), "Unmatched tensors!");
if (i > 0) {
XTensor * preItem = smalls->GetItem(i - 1);
......@@ -325,17 +325,17 @@ void _Merge(const TensorList * smalls, XTensor * big, int whereToMerge)
/* merging with fewer data copy operations */
if (mergedNum * gridNum <= MIN_TENSOR_MERGE_LIST_NUM) {
int sPitch = blockSize * s0->unitSize;
int tPtich = blockSize * mergedNum * big->unitSize;
int mSize = blockSize * big->unitSize;
int tPtich = blockSize * mergedNum * t->unitSize;
int mSize = blockSize * t->unitSize;
int n = blockNum;
int sStep = 0;
int tStep = blockSize * big->unitSize;
int tStep = blockSize * t->unitSize;
for (int g = 0; g < gridNum; g++) {
char * tData = (char*)big->data + g * blockSize * blockNum * big->unitSize;
char * tData = (char*)t->data + g * blockSize * blockNum * t->unitSize;
for (int k = 0; k < mergedNum; k++) {
XTensor * s = smalls->GetItem(k);
char * sData = (char*)s->data + g * blockSize * blockNum * s->unitSize;
XMemCopy2D(tData + k * tStep, tPtich, big->devID,
XMemCopy2D(tData + k * tStep, tPtich, t->devID,
sData + k * sStep, sPitch, s->devID,
mSize, n);
}
......@@ -358,7 +358,7 @@ void _Merge(const TensorList * smalls, XTensor * big, int whereToMerge)
if (uniform)
dataTMP = smallsItem0->data;
else
dataTMP = mem != NULL ? mem->AllocBuf(mem->devID, size) : XMemAlloc(big->devID, size);
dataTMP = mem != NULL ? mem->AllocBuf(mem->devID, size) : XMemAlloc(t->devID, size);
tensorTMP->data = dataTMP;
......@@ -370,7 +370,7 @@ void _Merge(const TensorList * smalls, XTensor * big, int whereToMerge)
}
}
_Merge(tensorTMP, big, whereToMerge + 1);
_Merge(tensorTMP, t, whereToMerge + 1);
delete[] dimSizeTMP;
......@@ -380,7 +380,7 @@ void _Merge(const TensorList * smalls, XTensor * big, int whereToMerge)
if ((!uniform) && (mem != NULL))
mem->ReleaseBuf(mem->devID, size);
else
XMemFree(big->devID, dataTMP);
XMemFree(t->devID, dataTMP);
}
}
......
source/tensor/core/shape/Merge.h
......@@ -36,7 +36,7 @@ XTensor Merge(const XTensor &s, int whereToMerge, int leadingDim = -1);
void Merge(const XTensor &s, XTensor &t, int whereToMerge, int leadingDim = -1, bool requireLink = false);
/* merge small tensors into a big tensor */
void _Merge(const TensorList * smalls, XTensor * big, int whereToMerge);
void _Merge(const TensorList * smalls, XTensor * t, int whereToMerge);
/* merge small tensors into a big tensor (return an XTensor structure) */
XTensor Merge(const TensorList &smalls, int whereToMerge);
......
source/tensor/test/TSumByColumnTV.cpp deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: Xu Chen (email: hello_master1954@163.com) 2018-07-06
*/
#include "TSumByColumnTV.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
/*
case 1: test SumByColumnTV function
sum of a tensor and a vector (column vector) in a column by column manner
*/
bool TestSumByColumnTV1()
{
/* a tensor of size (2, 4) */
int aOrder = 2;
int * aDimSize = new int[aOrder];
aDimSize[0] = 2;
aDimSize[1] = 4;
int aUnitNum = 1;
for (int i = 0; i < aOrder; i++)
aUnitNum *= aDimSize[i];
/* a tensor of size (2, 1) */
int bOrder = 2;
int * bDimSize = new int[bOrder];
bDimSize[0] = 2;
bDimSize[1] = 1;
int bUnitNum = 1;
for (int i = 0; i < bOrder; i++)
bUnitNum *= bDimSize[i];
/* a tensor of size (2, 4) */
int cOrder = 2;
int * cDimSize = new int[cOrder];
cDimSize[0] = 2;
cDimSize[1] = 4;
int cUnitNum = 1;
for (int i = 0; i < cOrder; i++)
cUnitNum *= cDimSize[i];
DTYPE aData[2][4] = { {0.0F, 1.0F, 2.0F, 3.0F},
{4.0F, 5.0F, 6.0F, 7.0F} };
DTYPE bData[2][1] = { {1.0F},
{0.0F} };
DTYPE answer[2][4] = { {1.0F, 2.0F, 3.0F, 4.0F},
{4.0F, 5.0F, 6.0F, 7.0F} };
/* CPU test */
bool cpuTest = true;
/* create tensors */
XTensor * a = NewTensor(aOrder, aDimSize);
XTensor * b = NewTensor(bOrder, bDimSize);
XTensor * c = NewTensor(cOrder, cDimSize);
/* initialize variables */
a->SetData(aData, aUnitNum);
b->SetData(bData, bUnitNum);
/* call SumByColumnTV function */
_SumByColumnTV(a, b, c);
/* check results */
cpuTest = c->CheckData(answer, cUnitNum);
#ifdef USE_CUDA
/* GPU test */
bool gpuTest = true;
/* create tensor */
XTensor * aGPU = NewTensor(aOrder, aDimSize, X_FLOAT, 1.0F, 0);
XTensor * bGPU = NewTensor(bOrder, bDimSize, X_FLOAT, 1.0F, 0);
XTensor * cGPU = NewTensor(cOrder, cDimSize, X_FLOAT, 1.0F, 0);
/* Initialize variables */
aGPU->SetData(aData, aUnitNum);
bGPU->SetData(bData, bUnitNum);
cGPU->SetZeroAll();
/* call SumByColumnTV function */
_SumByColumnTV(aGPU, bGPU, cGPU);
/* check results */
gpuTest = cGPU->CheckData(answer, cUnitNum);
/* destroy variables */
delete a;
delete b;
delete c;
delete aGPU;
delete bGPU;
delete cGPU;
delete[] aDimSize;
delete[] bDimSize;
delete[] cDimSize;
return cpuTest && gpuTest;
#else
/* destroy variables */
delete a;
delete b;
delete c;
delete[] aDimSize;
delete[] bDimSize;
delete[] cDimSize;
return cpuTest;
#endif // USE_CUDA
}
/* other cases */
/*
TODO!!
*/
/* test for SumByColumnTV Function */
bool TestSumByColumnTV()
{
XPRINT(0, stdout, "[TEST SumByColumnTV] sum of a tensor and a vector (column vector) in a column by column manner \n");
bool returnFlag = true, caseFlag = true;
/* case 1 test */
caseFlag = TestSumByColumnTV1();
if (!caseFlag) {
returnFlag = false;
XPRINT(0, stdout, ">> case 1 failed!\n");
}
else
XPRINT(0, stdout, ">> case 1 passed!\n");
/* other cases test */
/*
TODO!!
*/
if (returnFlag) {
XPRINT(0, stdout, ">> All Passed!\n");
}
else
XPRINT(0, stdout, ">> Failed!\n");
XPRINT(0, stdout, "\n");
return returnFlag;
}
} // namespace nts(NiuTrans.Tensor)
source/tensor/test/TSumByColumnTV.h deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: Xu Chen (email: hello_master1954@163.com) 2018-07-06
*/
#ifndef __TEST_SUMBYCOLUMNTV_H__
#define __TEST_SUMBYCOLUMNTV_H__
#include "../core/arithmetic/SumByColumnTV.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
/* test for SumByColumnTV Function */
extern "C"
bool TestSumByColumnTV();
} // namespace nts(NiuTrans.Tensor)
#endif // __TEST_SUMBYCOLUMNTV_H__
source/tensor/test/TSumByColumnVT.cpp deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: Xu Chen (email: hello_master1954@163.com) 2018-07-06
*/
#include "TSumByColumnVT.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
/*
case 1: test SumByColumnVT function
sum of a vector (column vector) and a tensor in a column by column manner
*/
bool TestSumByColumnVT1()
{
/* a tensor of size (2, 1) */
int aOrder = 2;
int * aDimSize = new int[aOrder];
aDimSize[0] = 2;
aDimSize[1] = 1;
int aUnitNum = 1;
for (int i = 0; i < aOrder; i++)
aUnitNum *= aDimSize[i];
/* a tensor of size (2, 4) */
int bOrder = 2;
int * bDimSize = new int[bOrder];
bDimSize[0] = 2;
bDimSize[1] = 4;
int bUnitNum = 1;
for (int i = 0; i < bOrder; i++)
bUnitNum *= bDimSize[i];
/* a tensor of size (2, 1) */
int cOrder = 2;
int * cDimSize = new int[cOrder];
cDimSize[0] = 2;
cDimSize[1] = 1;
int cUnitNum = 1;
for (int i = 0; i < cOrder; i++)
cUnitNum *= cDimSize[i];
DTYPE aData[2][1] = { {1.0F},
{0.0F} };
DTYPE bData[2][4] = { {0.0F, 1.0F, 2.0F, 3.0F},
{4.0F, 5.0F, 6.0F, 7.0F} };
DTYPE answer[2][1] = { {7.0F},
{22.0F} };
/* CPU test */
bool cpuTest = true;
/* create tensors */
XTensor * a = NewTensor(aOrder, aDimSize);
XTensor * b = NewTensor(bOrder, bDimSize);
XTensor * c = NewTensor(cOrder, cDimSize);
/* initialize variables */
a->SetData(aData, aUnitNum);
b->SetData(bData, bUnitNum);
c->SetZeroAll();
/* call SumByColumnVT function */
_SumByColumnVT(a, b, c);
/* check results */
cpuTest = c->CheckData(answer, cUnitNum);
#ifdef USE_CUDA
/* GPU test */
bool gpuTest = true;
/* create tensor */
XTensor * aGPU = NewTensor(aOrder, aDimSize, X_FLOAT, 1.0F, 0);
XTensor * bGPU = NewTensor(bOrder, bDimSize, X_FLOAT, 1.0F, 0);
XTensor * cGPU = NewTensor(cOrder, cDimSize, X_FLOAT, 1.0F, 0);
/* Initialize variables */
aGPU->SetData(aData, aUnitNum);
bGPU->SetData(bData, bUnitNum);
cGPU->SetZeroAll();
/* call SumByColumnVT function */
_SumByColumnVT(aGPU, bGPU, cGPU);
/* check results */
gpuTest = cGPU->CheckData(answer, cUnitNum);
/* destroy variables */
delete a;
delete b;
delete c;
delete aGPU;
delete bGPU;
delete cGPU;
delete[] aDimSize;
delete[] bDimSize;
delete[] cDimSize;
return cpuTest && gpuTest;
#else
/* destroy variables */
delete a;
delete b;
delete c;
delete[] aDimSize;
delete[] bDimSize;
delete[] cDimSize;
return cpuTest;
#endif // USE_CUDA
}
/* other cases */
/*
TODO!!
*/
/* test for SumByColumnVT Function */
bool TestSumByColumnVT()
{
XPRINT(0, stdout, "[TEST SumByColumnVT] sum of a vector (column vector) and a tensor in a column by column manner \n");
bool returnFlag = true, caseFlag = true;
/* case 1 test */
caseFlag = TestSumByColumnVT1();
if (!caseFlag) {
returnFlag = false;
XPRINT(0, stdout, ">> case 1 failed!\n");
}
else
XPRINT(0, stdout, ">> case 1 passed!\n");
/* other cases test */
/*
TODO!!
*/
if (returnFlag) {
XPRINT(0, stdout, ">> All Passed!\n");
}
else
XPRINT(0, stdout, ">> Failed!\n");
XPRINT(0, stdout, "\n");
return returnFlag;
}
} // namespace nts(NiuTrans.Tensor)
source/tensor/test/TSumByColumnVT.h deleted 100644 → 0
/* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
/*
* $Created by: Xu Chen (email: hello_master1954@163.com) 2018-07-06
*/
#ifndef __TEST_SUMBYCOLUMNVT_H__
#define __TEST_SUMBYCOLUMNVT_H__
#include "../core/arithmetic/SumByColumnVT.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
/* test for SumByColumnVT Function */
extern "C"
bool TestSumByColumnVT();
} // namespace nts(NiuTrans.Tensor)
#endif // __TEST_SUMBYCOLUMNVT_H__
source/tensor/test/Test.cpp
......@@ -71,8 +71,6 @@ bool Test()
wrong = !TestSpread() || wrong;
wrong = !TestSub() || wrong;
wrong = !TestSum() || wrong;
wrong = !TestSumByColumnTV() || wrong;
wrong = !TestSumByColumnVT() || wrong;
wrong = !TestSumDim() || wrong;
wrong = !TestTan() || wrong;
wrong = !TestTranspose() || wrong;
......
source/tensor/test/Test.h
......@@ -64,8 +64,6 @@
#include "TSpread.h"
#include "TSub.h"
#include "TSum.h"
#include "TSumByColumnTV.h"
#include "TSumByColumnVT.h"
#include "TSumDim.h"
#include "TTan.h"
#include "TTranspose.h"
......
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