Skip to content

T
Tensor.LowPrecision
Commit b5c4aa4e by linye
Options

update divdim

parent 2ffea05e
正在显示 包含 253 行增加 和 2 行删除
source/tensor/core/arithmetic/DivDim.cu
......@@ -17,6 +17,7 @@
/*
* $Created by: Xu Chen (email: hello_master1954@163.com) 2018-08-15
* $Update by: Lin Ye (email: linye2015@outlook.com) 2019-07-15 float16 added
*/
#include "DivDim.cuh"
......@@ -168,6 +169,35 @@ void _CudaDivDim(const XTensor * a, const XTensor * b, XTensor * c, int n, DTYPE
ShowNTErrors("Something is wrong!");
}
}
else if (a->dataType == X_FLOAT16) {
unsigned short temp = FloatToFloat16(alpha);
half alpha1 = *((half *)&temp);
if (stride > 1){
GDevs.GetCudaThread2D(a->devID, stride * blockNum, blockSize, MAX_INT, cudaGrids, cudaBlocks);
if (alpha == (DTYPE)0.0F)
KernelDivWithCol<__half, false> <<<dim3(cudaGrids[0], cudaGrids[1]), dim3(cudaBlocks[0], cudaBlocks[1])>>>
((__half*)a->data, (__half*)b->data, (__half*)c->data,
blockSize, stride, blockSize * stride, blockNum, alpha1);
else
KernelDivWithCol<__half, true> <<<dim3(cudaGrids[0], cudaGrids[1]), dim3(cudaBlocks[0], cudaBlocks[1])>>>
((__half*)a->data, (__half*)b->data, (__half*)c->data,
blockSize, stride, blockSize * stride, blockNum, alpha1);
}
else if (stride == 1){
GDevs.GetCudaThread2D(a->devID, blockSize, blockNum, MAX_INT, cudaGrids, cudaBlocks);
if (alpha == (DTYPE)0.0F)
KernelDivWithRow<__half, false> <<<dim3(cudaGrids[0], cudaGrids[1]), dim3(cudaBlocks[0], cudaBlocks[1])>>>
((__half*)a->data, (__half*)b->data, (__half*)c->data,
blockNum, blockSize, alpha1);
else
KernelDivWithRow<__half, true> <<<dim3(cudaGrids[0], cudaGrids[1]), dim3(cudaBlocks[0], cudaBlocks[1])>>>
((__half*)a->data, (__half*)b->data, (__half*)c->data,
blockNum, blockSize, alpha1);
}
else {
ShowNTErrors("Something is wrong!");
}
}
else {
ShowNTErrors("TODO!");
}
......
source/tensor/test/TDivDim.cpp
......@@ -17,11 +17,13 @@
/*
* $Created by: Xu Chen (email: hello_master1954@163.com) 2018-08-14
* $Update by: Lin Ye (email: linye2015@outlook.com) 2019-07-15 float16 added
*/
#include "TDivDim.h"
#include "../core/arithmetic/DivDim.h"
#include "../XTensor.h"
#include "../core/getandset/ConvertDataType.h"
namespace nts { // namespace nts(NiuTrans.Tensor)
......@@ -251,6 +253,207 @@ bool TestDivDim2()
#endif // USE_CUDA
}
/*
case 3: float16 tensor division c = a/b + \alpha * c
where the size of b is equal to the n-th dimension of a,
i.e., a is divided with b by broadcasting.
In this case, (2, 4) / (2) = (2, 4), n = 0, alpha = 0.0.
*/
bool TestDivDim3()
{
/* 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) */
int bOrder = 1;
int * bDimSize = new int[bOrder];
bDimSize[0] = 2;
int bUnitNum = 1;
for (int i = 0; i < bOrder; i++)
bUnitNum *= bDimSize[i];
DTYPE aData[2][4] = { {0.0F, 1.0F, 2.0F, 3.0F},
{4.0F, 5.0F, 6.0F, 7.0F} };
DTYPE bData[2] = {1.0F, -1.0F};
DTYPE answer[2][4] = { {0.0F, 1.0F, 2.0F, 3.0F},
{-4.0F, -5.0F, -6.0F, -7.0F} };
/* CPU test */
bool cpuTest = true;
#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(aOrder, aDimSize, X_FLOAT, 1.0F, 0);
XTensor * cMeGPU = NewTensor(aOrder, aDimSize, X_FLOAT, 1.0F, 0);
XTensor cUserGPU;
/* create float16 tensor */
XTensor aHalfGPU;
XTensor bHalfGPU;
XTensor cHalfGPU;
XTensor cMeHalfGPU;
XTensor cUserHalfGPU;
/* Initialize variables */
aGPU->SetData(aData, aUnitNum);
cMeGPU->SetData(aData, aUnitNum);
bGPU->SetData(bData, bUnitNum);
cGPU->SetZeroAll();
/* convert data type from float to float16 */
aHalfGPU = ConvertDataType(*aGPU, X_FLOAT16);
bHalfGPU = ConvertDataType(*bGPU, X_FLOAT16);
cHalfGPU = ConvertDataType(*cGPU, X_FLOAT16);
cMeHalfGPU = ConvertDataType(*cMeGPU, X_FLOAT16);
/* call sum function */
_DivDim(&aHalfGPU, &bHalfGPU, &cHalfGPU, 0);
_DivDim(&cMeHalfGPU, &bHalfGPU, 0);
cUserHalfGPU = DivDim(aHalfGPU, bHalfGPU, 0);
/* convert data type from float16 to float */
_ConvertDataType(&cHalfGPU, cGPU);
_ConvertDataType(&cMeHalfGPU, cMeGPU);
cUserGPU = ConvertDataType(cUserHalfGPU, X_FLOAT);
/* check results */
gpuTest = cGPU->CheckData(answer, aUnitNum) &&
cMeGPU->CheckData(answer, aUnitNum) &&
cUserGPU.CheckData(answer, aUnitNum);
/* destroy variables */
delete aGPU;
delete bGPU;
delete cGPU;
delete cMeGPU;
delete[] aDimSize;
delete[] bDimSize;
return cpuTest && gpuTest;
#else
/* destroy variables */
delete[] aDimSize;
delete[] bDimSize;
return cpuTest;
#endif // USE_CUDA
}
/*
case 4: float16 tensor division c = a/b + \alpha * c
where the size of b is equal to the n-th dimension of a,
i.e., a is divided with b by broadcasting.
In this case, (2, 4) / (2, 2) = (2, 4), n = 1.
*/
bool TestDivDim4()
{
/* 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, 2) */
int bOrder = 2;
int * bDimSize = new int[bOrder];
bDimSize[0] = 2;
bDimSize[1] = 2;
int bUnitNum = 1;
for (int i = 0; i < bOrder; i++)
bUnitNum *= bDimSize[i];
DTYPE aData[2][4] = { {0.0F, 1.0F, 2.0F, 3.0F},
{4.0F, 5.0F, 6.0F, 7.0F} };
DTYPE bData[2][2] = { {1.0F, -1.0F},
{-1.0F, 1.0F} };
DTYPE answer[2][4] = { {0.0F, -1.0F, -2.0F, 3.0F},
{4.0F, -5.0F, -6.0F, 7.0F} };
/* CPU test */
bool cpuTest = true;
#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(aOrder, aDimSize, X_FLOAT, 1.0F, 0);
XTensor * cMeGPU = NewTensor(aOrder, aDimSize, X_FLOAT, 1.0F, 0);
XTensor cUserGPU;
/* create float16 tensor */
XTensor aHalfGPU;
XTensor bHalfGPU;
XTensor cHalfGPU;
XTensor cMeHalfGPU;
XTensor cUserHalfGPU;
/* Initialize variables */
aGPU->SetData(aData, aUnitNum);
cMeGPU->SetData(aData, aUnitNum);
bGPU->SetData(bData, bUnitNum);
cGPU->SetZeroAll();
/* convert data type from float to float16 */
aHalfGPU = ConvertDataType(*aGPU, X_FLOAT16);
bHalfGPU = ConvertDataType(*bGPU, X_FLOAT16);
cHalfGPU = ConvertDataType(*cGPU, X_FLOAT16);
cMeHalfGPU = ConvertDataType(*cMeGPU, X_FLOAT16);
/* call sum function */
_DivDim(&aHalfGPU, &bHalfGPU, &cHalfGPU, 1);
_DivDim(&cMeHalfGPU, &bHalfGPU, 1);
cUserHalfGPU = DivDim(aHalfGPU, bHalfGPU, 1);
/* convert data type from float16 to float */
_ConvertDataType(&cHalfGPU, cGPU);
_ConvertDataType(&cMeHalfGPU, cMeGPU);
cUserGPU = ConvertDataType(cUserHalfGPU, X_FLOAT);
/* check results */
gpuTest = cGPU->CheckData(answer, aUnitNum) &&
cMeGPU->CheckData(answer, aUnitNum) &&
cUserGPU.CheckData(answer, aUnitNum);
/* destroy variables */
delete aGPU;
delete bGPU;
delete cGPU;
delete cMeGPU;
delete[] aDimSize;
delete[] bDimSize;
return cpuTest && gpuTest;
#else
/* destroy variables */
delete[] aDimSize;
delete[] bDimSize;
return cpuTest;
#endif // USE_CUDA
}
/* other cases */
/*
TODO!!
......@@ -280,6 +483,24 @@ bool TestDivDim()
else
XPRINT(0, stdout, ">> case 2 passed!\n");
/* case 3 test */
caseFlag = TestDivDim3();
if (!caseFlag) {
returnFlag = false;
XPRINT(0, stdout, ">> case 3 failed!\n");
}
else
XPRINT(0, stdout, ">> case 3 passed!\n");
/* case 4 test */
caseFlag = TestDivDim4();
if (!caseFlag) {
returnFlag = false;
XPRINT(0, stdout, ">> case 4 failed!\n");
}
else
XPRINT(0, stdout, ">> case 4 passed!\n");
/* other cases test */
/*
TODO!!
......
source/tensor/test/Test.cpp
......@@ -39,7 +39,7 @@ bool Test()
//wrong = !TestCopyIndexed() || wrong;
//wrong = !TestCopyValues() || wrong;
//wrong = !TestDiv() || wrong;
//wrong = !TestDivDim() || wrong;
wrong = !TestDivDim() || wrong;
//wrong = !TestExp() || wrong;
//wrong = !TestGather() || wrong;
//wrong = !TestLog() || wrong;
......@@ -82,7 +82,7 @@ bool Test()
//wrong = !TestCrossEntropy() || wrong;
//wrong = !TestDropout() || wrong;
wrong = !TestHardTanH() || wrong;
//wrong = !TestHardTanH() || wrong;
//wrong = !TestIdentity() || wrong;
//wrong = !TestLogSoftmax() || wrong;
//wrong = !TestLoss() || wrong;
......
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