Skip to content

N
NiuTrans.Tensor
Commit 77663a3c by Tianzhi
Options

delete amax, asum, etc. delete global variable useBLAS also.

parent e42e0bf7
正在显示 包含 196 行增加 和 43 行删除
3.test 0 → 100644
5932 5359 13569 5 3 11 84 1322 5 42 6918 4 467 208 4 3 28 793 3989 15 4 62 1860 924 126 6 2 ||| 1 2930 10 4 79 17 7439 1268 5 19466 23 2453 25 37 81 5834 130 4 79 17 270 65 9 4 681 73 7 152 129 6 270 65 5 19466 23 343 4 146 27 12 215 3529 6 1186 1112 4 30 17 2270 7 2
63 753 1765 4 2695 44 283 1525 4 454 9 12994 11934 8 407 5 463 7158 7294 1106 4 2376 50 499 60 8 31 5 555 577 4 66 6 2 ||| 1 154 5 3820 27 1443 6 4319 275 5 12 920 264 8 2091 6 6700 733 10 4 3776 8784 222 6 374 222 26 69 1770 9 2738 1421 10 12 2441 2250 4 54 17 6 68 17 649 6 403 7 2
25 1329 6208 4 1452 4 50 61 1719 88 5658 5220 10 4375 4 2973 4 5141 125 1652 2551 2830 5 4312 925 4 12802 4 4690 8 2918 6 2 ||| 1 10 190 9 2689 21 4 1109 1806 5 4 969 3106 218 6 3106 218 545 1318 262 9 9936 6 114 4 13737 5 18993 5 4572 5 6 1424 391 10 4 3 1871 5 4 6359 1871 5 6 4 2054 1871 7 2
Makefile
......@@ -14,7 +14,7 @@ CUDA_LIB_DIR = $(CUDA_ROOT)/lib64
CUDA_INCLUDE = $(CUDA_ROOT)/include
# use MKL
USE_MKL = 0
USE_MKL = 1
INTEL_ROOT = /opt/intel
MKL_ROOT = /opt/intel/mkl
MKL_LIB_DIR = $(MKL_ROOT)/lib/intel64/
......
fbis.train 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
niutrans.tensor.cpu
No preview for this file type
source/sample/fnnlm/FNNLM.cpp
......@@ -121,12 +121,18 @@ int FNNLMMain(int argc, const char ** argv)
/* load arguments */
LoadArgs(argc, argv, model);
printf("After load argu\n");
/* check the setting */
Check(model);
printf("After check setting\n");
/* initialize model parameters */
Init(model);
printf("After init model\n");
/* learn model parameters */
if(strcmp(trainFN, ""))
Train(trainFN, shuffled, model);
......@@ -414,7 +420,9 @@ void Train(const char * train, bool isShuffled, FNNModel &model)
/* make a model to keep gradients */
FNNModel grad;
printf("before copy\n");
Copy(grad, model);
printf("after copy\n");
/* XNet for automatic differentiation */
XNet autoDiffer;
......@@ -455,6 +463,7 @@ void Train(const char * train, bool isShuffled, FNNModel &model)
/* make the gold tensor */
MakeWordBatch(gold, ngrams, ngramNum, model.n - 1, model.vSize, model.devID, model.mem);
printf("after make the gold tensor\n");
if(!autoDiff){
/* prepare an empty network for building the fnn */
......@@ -473,19 +482,23 @@ void Train(const char * train, bool isShuffled, FNNModel &model)
Update(model, grad, learningRate, false);
}
else{
printf("in autodiff\n");
/* gradient = 0 */
Clear(model, true);
printf("after clear\n");
/* forward + backward process */
/* this is implemented by gather function */
ForwardAutoDiff(ngrams, ngramNum, output, model);
printf("after implemented by gather function\n");
/* this is implemented by multiply function */
//ForwardAutoDiff(inputs, output, model);
/* automatic differentiation */
autoDiffer.Backward(output, gold, CROSSENTROPY);
printf("after autodiff\n");
/* update model parameters */
Update(model, grad, learningRate, true);
......@@ -992,6 +1005,7 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model
int size = batch * (n-1);
int * index = new int[size];
printf("in FAutoDiff, before bianli\n");
for(int i = 0; i < batch; i++){
for (int j = 0; j < n-1; j++){
int a = i * (n - 1) + j;
......@@ -999,9 +1013,11 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model
}
}
printf("in FAutoDiff, before init tnesor 1d\n");
InitTensor1D(&words, size, X_INT, model.devID, model.mem);
words.SetData(index, size);
printf("in FAutoDiff, before gather\n");
embeddingBig = Gather(model.embeddingW, words);
delete[] index;
......@@ -1010,12 +1026,15 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model
dimSize[0] = embeddingBig.GetDim(0) / (n - 1);
dimSize[1] = embeddingBig.GetDim(1) * (n - 1);
printf("in FAutoDiff, before reshape\n");
hidden = Reshape(embeddingBig, embeddingBig.order, dimSize);
printf("in FAutoDiff, before hidden layers\n");
/* hidden layers */
for(int i = 0; i < depth; i++)
hidden = HardTanH(MMul(hidden, model.hiddenW[i]) + model.hiddenB[i]);
printf("in FAutoDiff, before output layer\n");
/* output layer */
output = LogSoftmax(MMul(hidden, model.outputW) + model.outputB, 1);
}
......
source/sample/transformer/T2TTrainer.cpp
......@@ -186,7 +186,7 @@ void T2TTrainer::Train(const char * fn, const char * validFN, const char * model
if(isShuffled)
Shuffle(fn, trainFN);
#endif
printf("%s\n",trainFN);
FILE * file = fopen(trainFN, "rb");
CheckNTErrors(file, "cannot open training file!");
......@@ -286,7 +286,7 @@ void T2TTrainer::Train(const char * fn, const char * validFN, const char * model
break;
}
if (step % 100 == 0) {
if (step % 1 == 0) {
double elapsed = GetClockSec() - startT;
XPRINT8(0, stderr, "[INFO] elapsed=%.1fs, step=%d, epoch=%d, tword=%d, sword=%d, loss=%.3f, ppl=%.3f, sppl=%.3f",
elapsed, step, epoch, wordCountTotal, wordCountBatch, loss/wordCount, exp(loss/wordCount), exp(-prob/wc));
......
source/tensor/Main.cpp
......@@ -30,6 +30,7 @@
#include "XDevice.h"
#include "./test/Test.h"
#include "./core/CHeader.h"
#include "./XBLAS.h"
//#define CRTDBG_MAP_ALLOC
//#include <stdlib.h>
......@@ -46,6 +47,7 @@ void PowerTest();
int main( int argc, const char ** argv )
{
LoadBLAS("/opt/Openblas/libopenblas.so");
//PowerTest();
//LittleTest();
......
source/tensor/XBLAS.cpp
......@@ -68,11 +68,25 @@ void (*XBLAS_DGER)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST BLASINT M, OP
double *, OPENBLAS_CONST BLASINT);
float (*XBLAS_SASUM)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
float (*XBLAS_ISAMAX)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
double (*XBLAS_DASUM)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
CBLAS_INDEX (*XBLAS_ISAMAX)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
CBLAS_INDEX (*XBLAS_IDAMAX)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
CBLAS_INDEX (*XBLAS_ISAMIN)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
CBLAS_INDEX (*XBLAS_IDAMIN)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
float (*XBLAS_SNRM2)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
void (*XBLAS_SSCAL)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float a,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
double (*XBLAS_DNRM2)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
void (*XBLAS_SSCAL)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float a, float *x,OPENBLAS_CONST BLASINT incx);
void (*XBLAS_DSCAL)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double a, double *x,OPENBLAS_CONST BLASINT incx);
void (*XBLAS_SCOPY)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx,OPENBLAS_CONST float *y,OPENBLAS_CONST BLASINT incy);
void (*XBLAS_DCOPY)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx,OPENBLAS_CONST double *y,OPENBLAS_CONST BLASINT incy);
void (*XBLAS_SAXPY)(OPENBLAS_CONST BLASINT n, OPENBLAS_CONST float a, OPENBLAS_CONST float *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST float *y, OPENBLAS_CONST BLASINT incy);
void (*XBLAS_DAXPY)(OPENBLAS_CONST BLASINT n, OPENBLAS_CONST double a, OPENBLAS_CONST double *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST double *y, OPENBLAS_CONST BLASINT incy);
/* set the number of threads */
void (*XBLAS_SET_THREAD_NUM)(int);
......@@ -123,11 +137,25 @@ void LoadBLAS(const char * dllFileName)
(FARPROC&)XBLAS_DGER = GetProcAddress(hBLASDll, "cblas_dger");
(FARPROC&)XBLAS_SASUM = GetProcAddress(hBLASDll, "cblas_sasum");
(FARPROC&)XBLAS_DASUM = GetProcAddress(hBLASDll, "cblas_dasum");
(FARPROC&)XBLAS_ISAMAX = GetProcAddress(hBLASDll, "cblas_isamax");
(FARPROC&)XBLAS_IDAMAX = GetProcAddress(hBLASDll, "cblas_idamax");
(FARPROC&)XBLAS_ISAMIN = GetProcAddress(hBLASDll, "cblas_isamin");
(FARPROC&)XBLAS_IDAMIN = GetProcAddress(hBLASDll, "cblas_idamin");
(FARPROC&)XBLAS_SNRM2 = GetProcAddress(hBLASDll, "cblas_snrm2");
(FARPROC&)XBLAS_DNRM2 = GetProcAddress(hBLASDll, "cblas_dnrm2");
(FARPROC&)XBLAS_SSCAL = GetProcAddress(hBLASDll, "cblas_sscal");
(FARPROC&)XBLAS_DSCAL = GetProcAddress(hBLASDll, "cblas_dscal");
(FARPROC&)XBLAS_SCOPY = GetProcAddress(hBLASDll, "cblas_scopy");
(FARPROC&)XBLAS_DCOPY = GetProcAddress(hBLASDll, "cblas_dcopy");
(FARPROC&)XBLAS_SAXPY = GetProcAddress(hBLASDll, "cblas_saxpy");
(FARPROC&)XBLAS_DAXPY = GetProcAddress(hBLASDll, "cblas_daxpy");
/* multi-threading */
(FARPROC&)XBLAS_SET_THREAD_NUM = GetProcAddress(hBLASDll, "openblas_set_num_threads");
......@@ -163,11 +191,25 @@ void LoadBLAS(const char * dllFileName)
(FARPROC&)XBLAS_DGER = GetProcAddress(hBLASDll, "cblas_dger");
(FARPROC&)XBLAS_SASUM = GetProcAddress(hBLASDll, "cblas_sasum");
(FARPROC&)XBLAS_DASUM = GetProcAddress(hBLASDll, "cblas_dasum");
(FARPROC&)XBLAS_ISAMAX = GetProcAddress(hBLASDll, "cblas_isamax");
(FARPROC&)XBLAS_IDAMAX = GetProcAddress(hBLASDll, "cblas_idamax");
(FARPROC&)XBLAS_ISAMIN = GetProcAddress(hBLASDll, "cblas_isamin");
(FARPROC&)XBLAS_IDAMIN = GetProcAddress(hBLASDll, "cblas_idamin");
(FARPROC&)XBLAS_SNRM2 = GetProcAddress(hBLASDll, "cblas_snrm2");
(FARPROC&)XBLAS_DNRM2 = GetProcAddress(hBLASDll, "cblas_dnrm2");
(FARPROC&)XBLAS_SSCAL = GetProcAddress(hBLASDll, "cblas_sscal");
(FARPROC&)XBLAS_DSCAL = GetProcAddress(hBLASDll, "cblas_dscal");
(FARPROC&)XBLAS_SCOPY = GetProcAddress(hBLASDll, "cblas_scopy");
(FARPROC&)XBLAS_DCOPY = GetProcAddress(hBLASDll, "cblas_dcopy");
(FARPROC&)XBLAS_SAXPY = GetProcAddress(hBLASDll, "cblas_saxpy");
(FARPROC&)XBLAS_DAXPY = GetProcAddress(hBLASDll, "cblas_daxpy");
/* multi-threading */
(FARPROC&)XBLAS_SET_THREAD_NUM = GetProcAddress(hBLASDll, "MKL_Set_Num_Threads");
......@@ -181,11 +223,19 @@ void LoadBLAS(const char * dllFileName)
XBLAS_DGER = &cblas_dger;
XBLAS_SASUM = &cblas_sasum;
XBLAS_DASUM = &cblas_dasum;
XBLAS_ISAMAX = &cblas_isamax;
XBLAS_IDAMAX = &cblas_idamax;
XBLAS_ISAMIN = &cblas_isamin;
XBLAS_IDAMIN = &cblas_idamin;
XBLAS_SNRM2 = &cblas_snrm2;
XBLAS_DNRM2 = &cblas_dnrm2;
XBLAS_SSCAL = &cblas_sscal;
XBLAS_DSCAL = &cblas_dscal;
XBLAS_SCOPY = &cblas_scopy;
XBLAS_DCOPY = &cblas_dcopy;
XBLAS_SAXPY = &cblas_saxpy;
XBLAS_DAXPY = &cblas_daxpy;
#if defined(OPENBLAS)
XBLAS_SET_THREAD_NUM = &openblas_set_num_threads;
......
source/tensor/XBLAS.h
......@@ -37,6 +37,7 @@ namespace nts{
//#define OPENBLAS
#define OPENBLAS_CONST const
#define CBLAS_INDEX size_t
typedef int BLASINT;
typedef enum CBLAS_ORDER {CblasRowMajor=101, CblasColMajor=102} CBLAS_ORDER;
typedef enum CBLAS_TRANSPOSE {CblasNoTrans=111, CblasTrans=112, CblasConjTrans=113, CblasConjNoTrans=114} CBLAS_TRANSPOSE;
......@@ -98,12 +99,25 @@ extern "C" void (*XBLAS_DGER)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST BL
double *, OPENBLAS_CONST BLASINT);
extern "C" float (*XBLAS_SASUM)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" float (*XBLAS_ISAMAX)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" float (*XBLAS_ISAMIN)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" double (*XBLAS_DASUM)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX (*XBLAS_ISAMAX)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX (*XBLAS_IDAMAX)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX (*XBLAS_ISAMIN)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX (*XBLAS_IDAMIN)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" float (*XBLAS_SNRM2)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" void (*XBLAS_SSCAL)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float a,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" double (*XBLAS_DNRM2)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" void (*XBLAS_SSCAL)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float a, float *x,OPENBLAS_CONST BLASINT incx);
extern "C" void (*XBLAS_DSCAL)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double a, double *x,OPENBLAS_CONST BLASINT incx);
extern "C" void (*XBLAS_SCOPY)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx,OPENBLAS_CONST float *y,OPENBLAS_CONST BLASINT incy);
extern "C" void (*XBLAS_DCOPY)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx,OPENBLAS_CONST double *y,OPENBLAS_CONST BLASINT incy);
extern "C" void (*XBLAS_SAXPY)(OPENBLAS_CONST BLASINT n, OPENBLAS_CONST float a, OPENBLAS_CONST float *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST float *y, OPENBLAS_CONST BLASINT incy);
extern "C" void (*XBLAS_DAXPY)(OPENBLAS_CONST BLASINT n, OPENBLAS_CONST double a, OPENBLAS_CONST double *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST double *y, OPENBLAS_CONST BLASINT incy);
/* set the number of threads */
......@@ -144,12 +158,25 @@ extern "C" void cblas_dger (OPENBLAS_CONST enum CBLAS_ORDER order, OPENBLAS_CONS
double *A, OPENBLAS_CONST BLASINT lda);
extern "C" float cblas_sasum (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" float cblas_isamax (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" float cblas_isamin (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" double cblas_dasum (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX cblas_isamax (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX cblas_idamax (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX cblas_isamin (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" CBLAS_INDEX cblas_idamin (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" float cblas_snrm2 (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" void cblas_sscal (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float a,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx);
extern "C" double cblas_dnrm2 (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx);
extern "C" void cblas_sscal (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float a, float *x,OPENBLAS_CONST BLASINT incx);
extern "C" void cblas_dscal (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double a, double *x,OPENBLAS_CONST BLASINT incx);
extern "C" void cblas_scopy (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx,OPENBLAS_CONST float *y,OPENBLAS_CONST BLASINT incy);
extern "C" void cblas_dcopy (OPENBLAS_CONST BLASINT n,OPENBLAS_CONST double *x,OPENBLAS_CONST BLASINT incx,OPENBLAS_CONST double *y,OPENBLAS_CONST BLASINT incy);
extern "C" void cblas_saxpy (OPENBLAS_CONST BLASINT n, OPENBLAS_CONST float a, OPENBLAS_CONST float *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST float *y, OPENBLAS_CONST BLASINT incy);
extern "C" void cblas_daxpy (OPENBLAS_CONST BLASINT n, OPENBLAS_CONST double a, OPENBLAS_CONST double *x, OPENBLAS_CONST BLASINT incx, OPENBLAS_CONST double *y, OPENBLAS_CONST BLASINT incy);
#if defined(OPENBLAS)
/* better control of multi-threading */
......
source/tensor/XGlobal.cpp
/* NiuTrans.Tensor - an open-source tensor library
 /* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved.
*
......@@ -50,7 +50,7 @@ int CONST_MINUSONE = -1;
bool CONST_TRUE = true;
int verboseLevel = 0;
bool useBLAS = false;
bool useBLAS = true;
bool useCUDA = false;
FILE * tmpLog = NULL;
......
source/tensor/XGlobal.h
......@@ -155,11 +155,28 @@ extern bool useCUDA;
#define B2I(V) V==0?false:true
#define SCAL XBLAS_SSCAL
/* BLAS interfaces */
#ifdef DOUBELPRICSION
#define GEMM XBLAS_DGEMM
#define GER XBLAS_DGER
#define ASUM XBLAS_DASUM
#define IAMAX XBLAS_IDAMAX
#define IAMIN XBLAS_IDAMIN
#define NRM2 XBLAS_DNRM2
#define SCAL XBLAS_DSCAL
#define COPY XBLAS_DCOPY
#define AXPY XBLAS_DAXPY
#else
#define GEMM XBLAS_SGEMM
#define GER XBLAS_SGER
#define ASUM XBLAS_SASUM
#define IAMAX XBLAS_ISAMAX
#define IAMIN XBLAS_ISAMIN
#define NRM2 XBLAS_SNRM2
#define SCAL XBLAS_SSCAL
#define COPY XBLAS_SCOPY
#define AXPY XBLAS_SAXPY
#endif
extern void InitGlobalAll();
......
source/tensor/core/arithmetic/MatrixMul.cpp
......@@ -62,6 +62,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
/* we transform a higher order tensor to a matrix to kill the number
of calls of matrix multiplication */
printf("in MMUL\n");
if(transposedA == X_NOTRANS && a->order > 2 && b->order == 2){
int ncolA = a->dimSize[a->order - 1];
int ncolC = c->dimSize[c->order - 1];
......@@ -69,7 +70,9 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
XTensor * c2 = NewTensor2D(c->unitNum/ncolC, -ncolC, c->dataType, c->devID, c->mem);
a2->data = a->data;
c2->data = c->data;
printf("before _MatMul\n");
_MatrixMul2D(a2, transposedA, b, transposedB, c2, alpha, beta, parallelRunner);
printf("after _MatMul\n");
a2->data = NULL;
c2->data = NULL;
delete a2;
......@@ -117,6 +120,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
bool isSparseMul = false;
printf("before bianli\n");
for (int p = 0; p < aBlockNum; p++) {
void * ap = (char*)a->data + aRealBlockSize * p;
for (int q = 0; q < bBlockNum; q++) {
......@@ -143,6 +147,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
}
}
}
printf("after bianli\n");
if (isSparseMul) {
for (int i = 0; i < aList->count; i++) {
......@@ -174,9 +179,11 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
}
else {
CheckNTErrors((a->dataType == DEFAULT_DTYPE), "TODO!");
printf("before _MatMul\n");
_MatrixMulBatchedCPU(aList, transposedA,
bList, transposedB,
cList, alpha, beta);
printf("after _MatMul\n");
}
for (int i = 0; i < aList->count; i++) {
......
source/tensor/core/arithmetic/MatrixMul2D.cpp
......@@ -82,10 +82,11 @@ void _MatrixMul2D(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
b->dataType == DEFAULT_DTYPE &&
c->dataType == DEFAULT_DTYPE)
{
if (useBLAS)
#if defined(USE_BLAS)
_MatrixMULCPU(a, transposedA, b, transposedB, c, alpha, beta);
else
#else
_MatrixMul2DParallel(a, transposedA, b, transposedB, c, alpha, beta, parallelRunner);
#endif
}
else {
// TODO!!
......
source/tensor/core/arithmetic/MatrixMulBatched.cpp
......@@ -201,10 +201,7 @@ CheckNTErrors((a && b && c), "Empty input tensors!");
bi->data = (char*)b->data + i * bRealBlockSize;
ci->data = (char*)c->data + i * cRealBlockSize;
#ifdef USE_BLAS
if (useBLAS)
_MatrixMULCPU(ai, transposedA, bi, transposedB, ci, alpha, beta);
else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#endif
......@@ -233,6 +230,7 @@ void _MatrixMulBatchedCPU(const XList * a, MATRIX_TRANS_TYPE transposedA,
const XList * b, MATRIX_TRANS_TYPE transposedB,
XList * c, DTYPE alpha, DTYPE beta)
{
printf("in _MMULBATCHED\n");
CheckNTErrors(a && b && c, "Empty input lists!");
CheckNTErrors(a->count == b->count && a->count == c->count, "Input lists must be of the same size!");
......@@ -264,10 +262,7 @@ void _MatrixMulBatchedCPU(const XList * a, MATRIX_TRANS_TYPE transposedA,
CheckNTErrors((bi->order == 2), "2d tensor (i.e., matrix) is required!");
CheckNTErrors((ci->order == 2), "2d tensor (i.e., matrix) is required!");
#ifdef USE_BLAS
if (useBLAS)
_MatrixMULCPU(ai, transposedA, bi, transposedB, ci, alpha, beta);
else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#endif
......
source/tensor/core/arithmetic/Sum.cpp
......@@ -86,10 +86,32 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
DTYPE * cp = (DTYPE*)c->data;
// when c != a, OpenBLAS needs to copy a to c first. This operation
// slow down the speed, so just use OpenBLAS when c == a
if(useBLAS && c == a){
cblas_saxpy(a->unitNum,1,bp,1,cp,1);
#if defined(USE_BLAS)
if( c == a){
AXPY(a->unitNum,beta,bp,1,cp,1);
} else{
int num = a->unitNum;
if (num % 4 == 0) {
for (int i = 0; i < num; i += 4) {
cp[i] = ap[i] + bp[i] * beta;
cp[i + 1] = ap[i + 1] + bp[i + 1] * beta;
cp[i + 2] = ap[i + 2] + bp[i + 2] * beta;
cp[i + 3] = ap[i + 3] + bp[i + 3] * beta;
}
else{
}
else if (num % 2 == 0) {
for (int i = 0; i < num; i += 2) {
cp[i] = ap[i] + bp[i] * beta;
cp[i + 1] = ap[i + 1] + bp[i + 1] * beta;
}
}
else {
for (int i = 0; i < num; i++) {
cp[i] = ap[i] + bp[i] * beta;
}
}
}
#else
/* unrolling */
int num = a->unitNum;
if (num % 4 == 0) {
......@@ -111,7 +133,7 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
cp[i] = ap[i] + bp[i] * beta;
}
}
}
#endif
}
else {
// TODO!!
......
source/tensor/core/arithmetic/XTensorBLAS.cpp
......@@ -54,7 +54,6 @@ void _MatrixMULCPU(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
int bm = b->dimSize[1];
int cn = c->dimSize[0];
int cm = c->dimSize[1];
printf("4\n");
if (transposedA == X_NOTRANS && transposedB == X_NOTRANS)
GEMM(CblasRowMajor, CblasNoTrans, CblasNoTrans, cn, cm, am, alpha, (DTYPE*)a->data, am, (DTYPE*)b->data, bm, beta, (DTYPE*)c->data, cm);
else if (transposedA == X_TRANS && transposedB == X_NOTRANS)
......
source/tensor/core/math/ScaleAndShift.cpp
......@@ -71,8 +71,9 @@ void _ScaleAndShift(const XTensor * a, XTensor * b, DTYPE scale, DTYPE shift)
else{
DTYPE * va = (DTYPE*)a->data;
DTYPE * vb = (DTYPE*)b->data;
if(shift == 0 && useBLAS && a==b){
cblas_sscal(b->unitNum, scale, vb, 1);
#if defined(USE_BLAS)
if(shift == 0 && a==b){
SCAL(b->unitNum, scale, vb, 1);
} else{
for(int i = 0; i < b->unitNum; i++){
*vb = *va * scale + shift;
......@@ -80,6 +81,13 @@ void _ScaleAndShift(const XTensor * a, XTensor * b, DTYPE scale, DTYPE shift)
vb++;
}
}
#else
for(int i = 0; i < b->unitNum; i++){
*vb = *va * scale + shift;
va++;
vb++;
}
#endif
}
}
......
source/tensor/core/reduce/ReduceMax.cpp
......@@ -82,9 +82,9 @@ void _ReduceMax(const XTensor * input, XTensor * output, int dim)
DTYPE * ip = (DTYPE*)input->data + blockSize * k;
DTYPE * op = (DTYPE*)output->data + stride * k;
for(int i = 0; i < stride; i++){
if(useBLAS){
*(op + i) = cblas_isamax(strideNum, ip + i, stride);
} else{
//#if defined(USE_BLAS)
// *(op + i) = *(ip + i + (int)(stride * IAMAX(strideNum, ip + i, stride)));
//#else
DTYPE max = FLOAT_MIN;
DTYPE * ipe = ip + blockSize;
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){
......@@ -93,7 +93,7 @@ void _ReduceMax(const XTensor * input, XTensor * output, int dim)
max = v;
}
*(op + i) = max;
}
//#endif
}
}
}
......
source/tensor/core/reduce/ReduceSum.cpp
......@@ -25,6 +25,7 @@
#include "../../XName.h"
#include "../../XBLAS.h"
#include "../arithmetic/XTensorBLAS.h"
#include <iostream>
namespace nts{ // namespace nts(NiuTrans.Tensor)
......@@ -145,22 +146,23 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor
else{
if(bias == 0){
if(power == (DTYPE)1.0){
if(useBLAS)
sum = cblas_sasum(strideNum, ip + i, stride);
else
//#if defined(USE_BLAS)
// sum = ASUM(strideNum, ip + i, stride);
//#else
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride)
sum += *ipb;
//#endif
}
else if(power == (DTYPE)2.0){
if(useBLAS){
sum = cblas_snrm2(strideNum, ip + i, stride);
sum = sum * sum;
} else{
//#if defined(USE_BLAS)
// sum = NRM2(strideNum, ip + i, stride);
// sum = sum * sum;
//#else
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){
DTYPE value = (*ipb);
sum += value * value;
}
}
//#endif
}
else if(power == (DTYPE)0.5){
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){
......@@ -177,11 +179,12 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor
}
else{
if(power == (DTYPE)1.0){
if(useBLAS)
sum = cblas_sasum(strideNum, ip + i, stride);
else
//#if defined(USE_BLAS)
// sum = ASUM(strideNum, ip + i, stride);
//#else
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride)
sum += *ipb;
//#endif
sum -= strideNum * bias;
}
else if(power == (DTYPE)2.0){
......
tmp.txt 0 → 100644
差异被折叠。 点击展开。
venv/lib/python3.6/site-packages/pip-10.0.1-py3.6.egg/pip/_vendor/certifi/cacert.pem 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
venv/lib/python3.6/site-packages/pip-10.0.1-py3.6.egg/pip/_vendor/html5lib/html5parser.py 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
venv/lib/python3.6/site-packages/pip-10.0.1-py3.6.egg/pip/_vendor/idna/uts46data.py 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
venv/lib/python3.6/site-packages/pip-10.0.1-py3.6.egg/pip/_vendor/pkg_resources/__init__.py 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
venv/lib/python3.6/site-packages/pip-10.0.1-py3.6.egg/pip/_vendor/pyparsing.py 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
wsj.test 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
wsj.train 0 → 100644
This source diff could not be displayed because it is too large. You can view the blob instead.
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