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 215 行增加 和 89 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 ...@@ -14,7 +14,7 @@ CUDA_LIB_DIR = $(CUDA_ROOT)/lib64
CUDA_INCLUDE = $(CUDA_ROOT)/include CUDA_INCLUDE = $(CUDA_ROOT)/include
# use MKL # use MKL
USE_MKL = 0 USE_MKL = 1
INTEL_ROOT = /opt/intel INTEL_ROOT = /opt/intel
MKL_ROOT = /opt/intel/mkl MKL_ROOT = /opt/intel/mkl
MKL_LIB_DIR = $(MKL_ROOT)/lib/intel64/ 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) ...@@ -121,12 +121,18 @@ int FNNLMMain(int argc, const char ** argv)
/* load arguments */ /* load arguments */
LoadArgs(argc, argv, model); LoadArgs(argc, argv, model);
printf("After load argu\n");
/* check the setting */ /* check the setting */
Check(model); Check(model);
printf("After check setting\n");
/* initialize model parameters */ /* initialize model parameters */
Init(model); Init(model);
printf("After init model\n");
/* learn model parameters */ /* learn model parameters */
if(strcmp(trainFN, "")) if(strcmp(trainFN, ""))
Train(trainFN, shuffled, model); Train(trainFN, shuffled, model);
...@@ -414,7 +420,9 @@ void Train(const char * train, bool isShuffled, FNNModel &model) ...@@ -414,7 +420,9 @@ void Train(const char * train, bool isShuffled, FNNModel &model)
/* make a model to keep gradients */ /* make a model to keep gradients */
FNNModel grad; FNNModel grad;
printf("before copy\n");
Copy(grad, model); Copy(grad, model);
printf("after copy\n");
/* XNet for automatic differentiation */ /* XNet for automatic differentiation */
XNet autoDiffer; XNet autoDiffer;
...@@ -455,6 +463,7 @@ void Train(const char * train, bool isShuffled, FNNModel &model) ...@@ -455,6 +463,7 @@ void Train(const char * train, bool isShuffled, FNNModel &model)
/* make the gold tensor */ /* make the gold tensor */
MakeWordBatch(gold, ngrams, ngramNum, model.n - 1, model.vSize, model.devID, model.mem); MakeWordBatch(gold, ngrams, ngramNum, model.n - 1, model.vSize, model.devID, model.mem);
printf("after make the gold tensor\n");
if(!autoDiff){ if(!autoDiff){
/* prepare an empty network for building the fnn */ /* prepare an empty network for building the fnn */
...@@ -473,19 +482,23 @@ void Train(const char * train, bool isShuffled, FNNModel &model) ...@@ -473,19 +482,23 @@ void Train(const char * train, bool isShuffled, FNNModel &model)
Update(model, grad, learningRate, false); Update(model, grad, learningRate, false);
} }
else{ else{
printf("in autodiff\n");
/* gradient = 0 */ /* gradient = 0 */
Clear(model, true); Clear(model, true);
printf("after clear\n");
/* forward + backward process */ /* forward + backward process */
/* this is implemented by gather function */ /* this is implemented by gather function */
ForwardAutoDiff(ngrams, ngramNum, output, model); ForwardAutoDiff(ngrams, ngramNum, output, model);
printf("after implemented by gather function\n");
/* this is implemented by multiply function */ /* this is implemented by multiply function */
//ForwardAutoDiff(inputs, output, model); //ForwardAutoDiff(inputs, output, model);
/* automatic differentiation */ /* automatic differentiation */
autoDiffer.Backward(output, gold, CROSSENTROPY); autoDiffer.Backward(output, gold, CROSSENTROPY);
printf("after autodiff\n");
/* update model parameters */ /* update model parameters */
Update(model, grad, learningRate, true); Update(model, grad, learningRate, true);
...@@ -992,6 +1005,7 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model ...@@ -992,6 +1005,7 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model
int size = batch * (n-1); int size = batch * (n-1);
int * index = new int[size]; int * index = new int[size];
printf("in FAutoDiff, before bianli\n");
for(int i = 0; i < batch; i++){ for(int i = 0; i < batch; i++){
for (int j = 0; j < n-1; j++){ for (int j = 0; j < n-1; j++){
int a = i * (n - 1) + j; int a = i * (n - 1) + j;
...@@ -999,9 +1013,11 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model ...@@ -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); InitTensor1D(&words, size, X_INT, model.devID, model.mem);
words.SetData(index, size); words.SetData(index, size);
printf("in FAutoDiff, before gather\n");
embeddingBig = Gather(model.embeddingW, words); embeddingBig = Gather(model.embeddingW, words);
delete[] index; delete[] index;
...@@ -1010,13 +1026,16 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model ...@@ -1010,13 +1026,16 @@ void ForwardAutoDiff(NGram * ngrams, int batch, XTensor &output, FNNModel &model
dimSize[0] = embeddingBig.GetDim(0) / (n - 1); dimSize[0] = embeddingBig.GetDim(0) / (n - 1);
dimSize[1] = embeddingBig.GetDim(1) * (n - 1); dimSize[1] = embeddingBig.GetDim(1) * (n - 1);
printf("in FAutoDiff, before reshape\n");
hidden = Reshape(embeddingBig, embeddingBig.order, dimSize); hidden = Reshape(embeddingBig, embeddingBig.order, dimSize);
printf("in FAutoDiff, before hidden layers\n");
/* hidden layers */ /* hidden layers */
for(int i = 0; i < depth; i++) for(int i = 0; i < depth; i++)
hidden = HardTanH(MMul(hidden, model.hiddenW[i]) + model.hiddenB[i]); hidden = HardTanH(MMul(hidden, model.hiddenW[i]) + model.hiddenB[i]);
/* output layer */ printf("in FAutoDiff, before output layer\n");
/* output layer */
output = LogSoftmax(MMul(hidden, model.outputW) + model.outputB, 1); 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 ...@@ -186,7 +186,7 @@ void T2TTrainer::Train(const char * fn, const char * validFN, const char * model
if(isShuffled) if(isShuffled)
Shuffle(fn, trainFN); Shuffle(fn, trainFN);
#endif #endif
printf("%s\n",trainFN);
FILE * file = fopen(trainFN, "rb"); FILE * file = fopen(trainFN, "rb");
CheckNTErrors(file, "cannot open training file!"); CheckNTErrors(file, "cannot open training file!");
...@@ -286,7 +286,7 @@ void T2TTrainer::Train(const char * fn, const char * validFN, const char * model ...@@ -286,7 +286,7 @@ void T2TTrainer::Train(const char * fn, const char * validFN, const char * model
break; break;
} }
if (step % 100 == 0) { if (step % 1 == 0) {
double elapsed = GetClockSec() - startT; double elapsed = GetClockSec() - startT;
XPRINT8(0, stderr, "[INFO] elapsed=%.1fs, step=%d, epoch=%d, tword=%d, sword=%d, loss=%.3f, ppl=%.3f, sppl=%.3f", 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)); elapsed, step, epoch, wordCountTotal, wordCountBatch, loss/wordCount, exp(loss/wordCount), exp(-prob/wc));
......
source/tensor/Main.cpp
...@@ -30,6 +30,7 @@ ...@@ -30,6 +30,7 @@
#include "XDevice.h" #include "XDevice.h"
#include "./test/Test.h" #include "./test/Test.h"
#include "./core/CHeader.h" #include "./core/CHeader.h"
#include "./XBLAS.h"
//#define CRTDBG_MAP_ALLOC //#define CRTDBG_MAP_ALLOC
//#include <stdlib.h> //#include <stdlib.h>
...@@ -46,6 +47,7 @@ void PowerTest(); ...@@ -46,6 +47,7 @@ void PowerTest();
int main( int argc, const char ** argv ) int main( int argc, const char ** argv )
{ {
LoadBLAS("/opt/Openblas/libopenblas.so");
//PowerTest(); //PowerTest();
//LittleTest(); //LittleTest();
......
source/tensor/XBLAS.cpp
/* NiuTrans.Tensor - an open-source tensor library /* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University. * Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved. * All rights reserved.
* *
* Licensed under the Apache License, Version 2.0 (the "License"); * Licensed under the Apache License, Version 2.0 (the "License");
...@@ -17,10 +17,10 @@ ...@@ -17,10 +17,10 @@
/* /*
* *
* This is a wrapper of the BLAS (Basic Linear Algebra Subprograms http://www.netlib.org/blas/) * This is a wrapper of the BLAS (Basic Linear Algebra Subprograms http://www.netlib.org/blas/)
* libraries. By using BLAS, we can access very fast matrix operations although they * libraries. By using BLAS, we can access very fast matrix operations although they
* are also implemented in NiuTrans in a native manner. To use BLAS, * are also implemented in NiuTrans in a native manner. To use BLAS,
* set USE_BLAS. * set USE_BLAS.
* *
* $Created by: XIAO Tong (xiaotong@mail.neu.edu.cn) 2016-04-08 * $Created by: XIAO Tong (xiaotong@mail.neu.edu.cn) 2016-04-08
* *
...@@ -45,34 +45,48 @@ HINSTANCE hBLASDll; ...@@ -45,34 +45,48 @@ HINSTANCE hBLASDll;
/* single-precision floating matrix-matrix multiplication */ /* single-precision floating matrix-matrix multiplication */
void (*XBLAS_SGEMM)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE, OPENBLAS_CONST enum CBLAS_TRANSPOSE, void (*XBLAS_SGEMM)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE, OPENBLAS_CONST enum CBLAS_TRANSPOSE,
OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float, OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float,
OPENBLAS_CONST float *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float *, OPENBLAS_CONST BLASINT,
OPENBLAS_CONST float *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float, OPENBLAS_CONST float *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float,
float *, OPENBLAS_CONST BLASINT); float *, OPENBLAS_CONST BLASINT);
/* double-precision floating matrix-matrix multiplication */ /* double-precision floating matrix-matrix multiplication */
void (*XBLAS_DGEMM)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE, OPENBLAS_CONST enum CBLAS_TRANSPOSE, void (*XBLAS_DGEMM)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST enum CBLAS_TRANSPOSE, OPENBLAS_CONST enum CBLAS_TRANSPOSE,
OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double, OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double,
OPENBLAS_CONST double *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double *, OPENBLAS_CONST BLASINT,
OPENBLAS_CONST double *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double, OPENBLAS_CONST double *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double,
double *, OPENBLAS_CONST BLASINT); double *, OPENBLAS_CONST BLASINT);
/* single-precision floating vector-vector multiplication (rank-1) */ /* single-precision floating vector-vector multiplication (rank-1) */
void (*XBLAS_SGER)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST float alpha, void (*XBLAS_SGER)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST float alpha,
OPENBLAS_CONST float *Y, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float *Y, OPENBLAS_CONST BLASINT, OPENBLAS_CONST float *, OPENBLAS_CONST BLASINT,
float *, OPENBLAS_CONST BLASINT); float *, OPENBLAS_CONST BLASINT);
/* double-precision floating vector-vector multiplication (rank-1) */ /* double-precision floating vector-vector multiplication (rank-1) */
void (*XBLAS_DGER)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST double alpha, void (*XBLAS_DGER)(OPENBLAS_CONST enum CBLAS_ORDER, OPENBLAS_CONST BLASINT M, OPENBLAS_CONST BLASINT N, OPENBLAS_CONST double alpha,
OPENBLAS_CONST double *Y, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double *, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double *Y, OPENBLAS_CONST BLASINT, OPENBLAS_CONST double *, OPENBLAS_CONST BLASINT,
double *, OPENBLAS_CONST BLASINT); double *, OPENBLAS_CONST BLASINT);
float (*XBLAS_SASUM)(OPENBLAS_CONST BLASINT n,OPENBLAS_CONST float *x,OPENBLAS_CONST BLASINT incx); 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); 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_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_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 */ /* set the number of threads */
void (*XBLAS_SET_THREAD_NUM)(int); void (*XBLAS_SET_THREAD_NUM)(int);
...@@ -123,11 +137,25 @@ void LoadBLAS(const char * dllFileName) ...@@ -123,11 +137,25 @@ void LoadBLAS(const char * dllFileName)
(FARPROC&)XBLAS_DGER = GetProcAddress(hBLASDll, "cblas_dger"); (FARPROC&)XBLAS_DGER = GetProcAddress(hBLASDll, "cblas_dger");
(FARPROC&)XBLAS_SASUM = GetProcAddress(hBLASDll, "cblas_sasum"); (FARPROC&)XBLAS_SASUM = GetProcAddress(hBLASDll, "cblas_sasum");
(FARPROC&)XBLAS_DASUM = GetProcAddress(hBLASDll, "cblas_dasum");
(FARPROC&)XBLAS_ISAMAX = GetProcAddress(hBLASDll, "cblas_isamax"); (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_SNRM2 = GetProcAddress(hBLASDll, "cblas_snrm2");
(FARPROC&)XBLAS_DNRM2 = GetProcAddress(hBLASDll, "cblas_dnrm2");
(FARPROC&)XBLAS_SSCAL = GetProcAddress(hBLASDll, "cblas_sscal"); (FARPROC&)XBLAS_SSCAL = GetProcAddress(hBLASDll, "cblas_sscal");
(FARPROC&)XBLAS_DSCAL = GetProcAddress(hBLASDll, "cblas_dscal");
(FARPROC&)XBLAS_SCOPY = GetProcAddress(hBLASDll, "cblas_scopy"); (FARPROC&)XBLAS_SCOPY = GetProcAddress(hBLASDll, "cblas_scopy");
(FARPROC&)XBLAS_DCOPY = GetProcAddress(hBLASDll, "cblas_dcopy");
(FARPROC&)XBLAS_SAXPY = GetProcAddress(hBLASDll, "cblas_saxpy"); (FARPROC&)XBLAS_SAXPY = GetProcAddress(hBLASDll, "cblas_saxpy");
(FARPROC&)XBLAS_DAXPY = GetProcAddress(hBLASDll, "cblas_daxpy");
/* multi-threading */ /* multi-threading */
(FARPROC&)XBLAS_SET_THREAD_NUM = GetProcAddress(hBLASDll, "openblas_set_num_threads"); (FARPROC&)XBLAS_SET_THREAD_NUM = GetProcAddress(hBLASDll, "openblas_set_num_threads");
...@@ -163,29 +191,51 @@ void LoadBLAS(const char * dllFileName) ...@@ -163,29 +191,51 @@ void LoadBLAS(const char * dllFileName)
(FARPROC&)XBLAS_DGER = GetProcAddress(hBLASDll, "cblas_dger"); (FARPROC&)XBLAS_DGER = GetProcAddress(hBLASDll, "cblas_dger");
(FARPROC&)XBLAS_SASUM = GetProcAddress(hBLASDll, "cblas_sasum"); (FARPROC&)XBLAS_SASUM = GetProcAddress(hBLASDll, "cblas_sasum");
(FARPROC&)XBLAS_DASUM = GetProcAddress(hBLASDll, "cblas_dasum");
(FARPROC&)XBLAS_ISAMAX = GetProcAddress(hBLASDll, "cblas_isamax"); (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_SNRM2 = GetProcAddress(hBLASDll, "cblas_snrm2");
(FARPROC&)XBLAS_DNRM2 = GetProcAddress(hBLASDll, "cblas_dnrm2");
(FARPROC&)XBLAS_SSCAL = GetProcAddress(hBLASDll, "cblas_sscal"); (FARPROC&)XBLAS_SSCAL = GetProcAddress(hBLASDll, "cblas_sscal");
(FARPROC&)XBLAS_DSCAL = GetProcAddress(hBLASDll, "cblas_dscal");
(FARPROC&)XBLAS_SCOPY = GetProcAddress(hBLASDll, "cblas_scopy"); (FARPROC&)XBLAS_SCOPY = GetProcAddress(hBLASDll, "cblas_scopy");
(FARPROC&)XBLAS_DCOPY = GetProcAddress(hBLASDll, "cblas_dcopy");
(FARPROC&)XBLAS_SAXPY = GetProcAddress(hBLASDll, "cblas_saxpy"); (FARPROC&)XBLAS_SAXPY = GetProcAddress(hBLASDll, "cblas_saxpy");
(FARPROC&)XBLAS_DAXPY = GetProcAddress(hBLASDll, "cblas_daxpy");
/* multi-threading */ /* multi-threading */
(FARPROC&)XBLAS_SET_THREAD_NUM = GetProcAddress(hBLASDll, "MKL_Set_Num_Threads"); (FARPROC&)XBLAS_SET_THREAD_NUM = GetProcAddress(hBLASDll, "MKL_Set_Num_Threads");
(FARPROC&)XBLAS_GET_CORE_NUM = GetProcAddress(hBLASDll, "MKL_Get_Max_Threads"); (FARPROC&)XBLAS_GET_CORE_NUM = GetProcAddress(hBLASDll, "MKL_Get_Max_Threads");
#endif // defined(MKL) #endif // defined(MKL)
#else // _WIN32 #else // _WIN32
XBLAS_SGEMM = &cblas_sgemm; XBLAS_SGEMM = &cblas_sgemm;
XBLAS_DGEMM = &cblas_dgemm; XBLAS_DGEMM = &cblas_dgemm;
XBLAS_SGER = &cblas_sger; XBLAS_SGER = &cblas_sger;
XBLAS_DGER = &cblas_dger; XBLAS_DGER = &cblas_dger;
XBLAS_SASUM = &cblas_sasum; XBLAS_SASUM = &cblas_sasum;
XBLAS_DASUM = &cblas_dasum;
XBLAS_ISAMAX = &cblas_isamax; XBLAS_ISAMAX = &cblas_isamax;
XBLAS_IDAMAX = &cblas_idamax;
XBLAS_ISAMIN = &cblas_isamin;
XBLAS_IDAMIN = &cblas_idamin;
XBLAS_SNRM2 = &cblas_snrm2; XBLAS_SNRM2 = &cblas_snrm2;
XBLAS_DNRM2 = &cblas_dnrm2;
XBLAS_SSCAL = &cblas_sscal; XBLAS_SSCAL = &cblas_sscal;
XBLAS_DSCAL = &cblas_dscal;
XBLAS_SCOPY = &cblas_scopy; XBLAS_SCOPY = &cblas_scopy;
XBLAS_DCOPY = &cblas_dcopy;
XBLAS_SAXPY = &cblas_saxpy; XBLAS_SAXPY = &cblas_saxpy;
XBLAS_DAXPY = &cblas_daxpy;
#if defined(OPENBLAS) #if defined(OPENBLAS)
XBLAS_SET_THREAD_NUM = &openblas_set_num_threads; XBLAS_SET_THREAD_NUM = &openblas_set_num_threads;
......
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. * Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved. * All rights reserved.
* *
...@@ -50,7 +50,7 @@ int CONST_MINUSONE = -1; ...@@ -50,7 +50,7 @@ int CONST_MINUSONE = -1;
bool CONST_TRUE = true; bool CONST_TRUE = true;
int verboseLevel = 0; int verboseLevel = 0;
bool useBLAS = false; bool useBLAS = true;
bool useCUDA = false; bool useCUDA = false;
FILE * tmpLog = NULL; FILE * tmpLog = NULL;
......
source/tensor/XGlobal.h
/* NiuTrans.Tensor - an open-source tensor library /* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University. * Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved. * All rights reserved.
* *
* Licensed under the Apache License, Version 2.0 (the "License"); * Licensed under the Apache License, Version 2.0 (the "License");
...@@ -43,7 +43,7 @@ ...@@ -43,7 +43,7 @@
/* the nts (NiuTrans.Tensor) namespace */ /* the nts (NiuTrans.Tensor) namespace */
namespace nts { namespace nts {
#define _XINLINE_ #define _XINLINE_
//#define DOUBELPRICSION //#define DOUBELPRICSION
...@@ -155,11 +155,28 @@ extern bool useCUDA; ...@@ -155,11 +155,28 @@ extern bool useCUDA;
#define B2I(V) V==0?false:true #define B2I(V) V==0?false:true
#define SCAL XBLAS_SSCAL
/* BLAS interfaces */ /* BLAS interfaces */
#ifdef DOUBELPRICSION #ifdef DOUBELPRICSION
#define GEMM XBLAS_DGEMM #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 #else
#define GEMM XBLAS_SGEMM #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 #endif
extern void InitGlobalAll(); extern void InitGlobalAll();
......
source/tensor/core/arithmetic/MatrixMul.cpp
...@@ -62,6 +62,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA, ...@@ -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 /* we transform a higher order tensor to a matrix to kill the number
of calls of matrix multiplication */ of calls of matrix multiplication */
printf("in MMUL\n");
if(transposedA == X_NOTRANS && a->order > 2 && b->order == 2){ if(transposedA == X_NOTRANS && a->order > 2 && b->order == 2){
int ncolA = a->dimSize[a->order - 1]; int ncolA = a->dimSize[a->order - 1];
int ncolC = c->dimSize[c->order - 1]; int ncolC = c->dimSize[c->order - 1];
...@@ -69,7 +70,9 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA, ...@@ -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); XTensor * c2 = NewTensor2D(c->unitNum/ncolC, -ncolC, c->dataType, c->devID, c->mem);
a2->data = a->data; a2->data = a->data;
c2->data = c->data; c2->data = c->data;
printf("before _MatMul\n");
_MatrixMul2D(a2, transposedA, b, transposedB, c2, alpha, beta, parallelRunner); _MatrixMul2D(a2, transposedA, b, transposedB, c2, alpha, beta, parallelRunner);
printf("after _MatMul\n");
a2->data = NULL; a2->data = NULL;
c2->data = NULL; c2->data = NULL;
delete a2; delete a2;
...@@ -117,6 +120,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA, ...@@ -117,6 +120,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
bool isSparseMul = false; bool isSparseMul = false;
printf("before bianli\n");
for (int p = 0; p < aBlockNum; p++) { for (int p = 0; p < aBlockNum; p++) {
void * ap = (char*)a->data + aRealBlockSize * p; void * ap = (char*)a->data + aRealBlockSize * p;
for (int q = 0; q < bBlockNum; q++) { for (int q = 0; q < bBlockNum; q++) {
...@@ -143,6 +147,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA, ...@@ -143,6 +147,7 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
} }
} }
} }
printf("after bianli\n");
if (isSparseMul) { if (isSparseMul) {
for (int i = 0; i < aList->count; i++) { for (int i = 0; i < aList->count; i++) {
...@@ -174,9 +179,11 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA, ...@@ -174,9 +179,11 @@ void _MatrixMul(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
} }
else { else {
CheckNTErrors((a->dataType == DEFAULT_DTYPE), "TODO!"); CheckNTErrors((a->dataType == DEFAULT_DTYPE), "TODO!");
printf("before _MatMul\n");
_MatrixMulBatchedCPU(aList, transposedA, _MatrixMulBatchedCPU(aList, transposedA,
bList, transposedB, bList, transposedB,
cList, alpha, beta); cList, alpha, beta);
printf("after _MatMul\n");
} }
for (int i = 0; i < aList->count; i++) { 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, ...@@ -82,10 +82,11 @@ void _MatrixMul2D(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
b->dataType == DEFAULT_DTYPE && b->dataType == DEFAULT_DTYPE &&
c->dataType == DEFAULT_DTYPE) c->dataType == DEFAULT_DTYPE)
{ {
if (useBLAS) #if defined(USE_BLAS)
_MatrixMULCPU(a, transposedA, b, transposedB, c, alpha, beta); _MatrixMULCPU(a, transposedA, b, transposedB, c, alpha, beta);
else #else
_MatrixMul2DParallel(a, transposedA, b, transposedB, c, alpha, beta, parallelRunner); _MatrixMul2DParallel(a, transposedA, b, transposedB, c, alpha, beta, parallelRunner);
#endif
} }
else { else {
// TODO!! // TODO!!
......
source/tensor/core/arithmetic/MatrixMulBatched.cpp
...@@ -201,10 +201,7 @@ CheckNTErrors((a && b && c), "Empty input tensors!"); ...@@ -201,10 +201,7 @@ CheckNTErrors((a && b && c), "Empty input tensors!");
bi->data = (char*)b->data + i * bRealBlockSize; bi->data = (char*)b->data + i * bRealBlockSize;
ci->data = (char*)c->data + i * cRealBlockSize; ci->data = (char*)c->data + i * cRealBlockSize;
#ifdef USE_BLAS #ifdef USE_BLAS
if (useBLAS) _MatrixMULCPU(ai, transposedA, bi, transposedB, ci, alpha, beta);
_MatrixMULCPU(ai, transposedA, bi, transposedB, ci, alpha, beta);
else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#else #else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta); _MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#endif #endif
...@@ -233,6 +230,7 @@ void _MatrixMulBatchedCPU(const XList * a, MATRIX_TRANS_TYPE transposedA, ...@@ -233,6 +230,7 @@ void _MatrixMulBatchedCPU(const XList * a, MATRIX_TRANS_TYPE transposedA,
const XList * b, MATRIX_TRANS_TYPE transposedB, const XList * b, MATRIX_TRANS_TYPE transposedB,
XList * c, DTYPE alpha, DTYPE beta) XList * c, DTYPE alpha, DTYPE beta)
{ {
printf("in _MMULBATCHED\n");
CheckNTErrors(a && b && c, "Empty input lists!"); CheckNTErrors(a && b && c, "Empty input lists!");
CheckNTErrors(a->count == b->count && a->count == c->count, "Input lists must be of the same size!"); 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, ...@@ -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((bi->order == 2), "2d tensor (i.e., matrix) is required!");
CheckNTErrors((ci->order == 2), "2d tensor (i.e., matrix) is required!"); CheckNTErrors((ci->order == 2), "2d tensor (i.e., matrix) is required!");
#ifdef USE_BLAS #ifdef USE_BLAS
if (useBLAS) _MatrixMULCPU(ai, transposedA, bi, transposedB, ci, alpha, beta);
_MatrixMULCPU(ai, transposedA, bi, transposedB, ci, alpha, beta);
else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#else #else
_MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta); _MatrixMul2D(ai, transposedA, bi, transposedB, ci, alpha, beta);
#endif #endif
......
source/tensor/core/arithmetic/Sum.cpp
...@@ -76,7 +76,7 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta) ...@@ -76,7 +76,7 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
else { else {
if (!a->isSparse && !b->isSparse) { if (!a->isSparse && !b->isSparse) {
CheckNTErrors(!c->isSparse, "Illegal use of sparse tensor in addition!"); CheckNTErrors(!c->isSparse, "Illegal use of sparse tensor in addition!");
if (a->dataType == DEFAULT_DTYPE && if (a->dataType == DEFAULT_DTYPE &&
b->dataType == DEFAULT_DTYPE && b->dataType == DEFAULT_DTYPE &&
c->dataType == DEFAULT_DTYPE) c->dataType == DEFAULT_DTYPE)
...@@ -84,12 +84,34 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta) ...@@ -84,12 +84,34 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
DTYPE * ap = (DTYPE*)a->data; DTYPE * ap = (DTYPE*)a->data;
DTYPE * bp = (DTYPE*)b->data; DTYPE * bp = (DTYPE*)b->data;
DTYPE * cp = (DTYPE*)c->data; DTYPE * cp = (DTYPE*)c->data;
// when c != a, OpenBLAS needs to copy a to c first. This operation // when c != a, OpenBLAS needs to copy a to c first. This operation
// slow down the speed, so just use OpenBLAS when c == a // slow down the speed, so just use OpenBLAS when c == a
if(useBLAS && c == a){ #if defined(USE_BLAS)
cblas_saxpy(a->unitNum,1,bp,1,cp,1); 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 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{ #else
/* unrolling */ /* unrolling */
int num = a->unitNum; int num = a->unitNum;
if (num % 4 == 0) { if (num % 4 == 0) {
...@@ -111,8 +133,8 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta) ...@@ -111,8 +133,8 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
cp[i] = ap[i] + bp[i] * beta; cp[i] = ap[i] + bp[i] * beta;
} }
} }
#endif
} }
}
else { else {
// TODO!! // TODO!!
ShowNTErrors("TODO!"); ShowNTErrors("TODO!");
...@@ -124,7 +146,7 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta) ...@@ -124,7 +146,7 @@ void _Sum(const XTensor * a, const XTensor * b, XTensor * c, DTYPE beta)
} }
} }
} }
/* /*
tensor summation a = a + b * \beta (do it on site) tensor summation a = a + b * \beta (do it on site)
keep the result in the tensor a and return nothing keep the result in the tensor a and return nothing
...@@ -138,7 +160,7 @@ void _SumMe(XTensor * a, const XTensor * b, DTYPE beta) ...@@ -138,7 +160,7 @@ void _SumMe(XTensor * a, const XTensor * b, DTYPE beta)
_Sum(a, b, a, beta); _Sum(a, b, a, beta);
} }
/* /*
return a dimension if the sum is performed as SumDim (in more details in SumDim.h) return a dimension if the sum is performed as SumDim (in more details in SumDim.h)
>> a - a tensor >> a - a tensor
>> b - another tensor for sum >> b - another tensor for sum
...@@ -166,7 +188,7 @@ int GetSumDimIndex(const XTensor &a, const XTensor &b) ...@@ -166,7 +188,7 @@ int GetSumDimIndex(const XTensor &a, const XTensor &b)
else else
return -1; return -1;
} }
/* /*
tensor summation c = a + b * \beta (return an XTensor structure) tensor summation c = a + b * \beta (return an XTensor structure)
make a new tensor c to keep the result and return it make a new tensor c to keep the result and return it
...@@ -186,7 +208,7 @@ XTensor Sum(const XTensor &a, const XTensor &b, DTYPE beta) ...@@ -186,7 +208,7 @@ XTensor Sum(const XTensor &a, const XTensor &b, DTYPE beta)
if(n == -1){ if(n == -1){
/* call _Sum function */ /* call _Sum function */
_Sum(&a, &b, &c, beta); _Sum(&a, &b, &c, beta);
/* tensor connections */ /* tensor connections */
XLink::MakeLink(&a, &b, &c, MATH_SUM); XLink::MakeLink(&a, &b, &c, MATH_SUM);
XLink::AddParamToHead(&c, beta); XLink::AddParamToHead(&c, beta);
...@@ -194,7 +216,7 @@ XTensor Sum(const XTensor &a, const XTensor &b, DTYPE beta) ...@@ -194,7 +216,7 @@ XTensor Sum(const XTensor &a, const XTensor &b, DTYPE beta)
else if(n >= 0 && n < a.order){ else if(n >= 0 && n < a.order){
/* call _SumDim function */ /* call _SumDim function */
_SumDim(&a, &b, &c, n, beta); _SumDim(&a, &b, &c, n, beta);
/* tensor connections */ /* tensor connections */
XLink::MakeLink(&a, &b, &c, MATH_SUMDIM); XLink::MakeLink(&a, &b, &c, MATH_SUMDIM);
XLink::AddParamToHeadInt(&c, n); XLink::AddParamToHeadInt(&c, n);
...@@ -203,7 +225,7 @@ XTensor Sum(const XTensor &a, const XTensor &b, DTYPE beta) ...@@ -203,7 +225,7 @@ XTensor Sum(const XTensor &a, const XTensor &b, DTYPE beta)
else{ else{
ShowNTErrors("Something is wrong!"); ShowNTErrors("Something is wrong!");
} }
return c; return c;
} }
......
source/tensor/core/arithmetic/XTensorBLAS.cpp
...@@ -54,7 +54,6 @@ void _MatrixMULCPU(const XTensor * a, MATRIX_TRANS_TYPE transposedA, ...@@ -54,7 +54,6 @@ void _MatrixMULCPU(const XTensor * a, MATRIX_TRANS_TYPE transposedA,
int bm = b->dimSize[1]; int bm = b->dimSize[1];
int cn = c->dimSize[0]; int cn = c->dimSize[0];
int cm = c->dimSize[1]; int cm = c->dimSize[1];
printf("4\n");
if (transposedA == X_NOTRANS && transposedB == X_NOTRANS) 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); 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) else if (transposedA == X_TRANS && transposedB == X_NOTRANS)
......
source/tensor/core/math/ScaleAndShift.cpp
/* NiuTrans.Tensor - an open-source tensor library /* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University. * Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved. * All rights reserved.
* *
* Licensed under the Apache License, Version 2.0 (the "License"); * Licensed under the Apache License, Version 2.0 (the "License");
...@@ -30,7 +30,7 @@ ...@@ -30,7 +30,7 @@
namespace nts{ // namespace nts(NiuTrans.Tensor) namespace nts{ // namespace nts(NiuTrans.Tensor)
/* /*
scale and shift all tensor entires scale and shift all tensor entires
b = a * scale + shift b = a * scale + shift
...@@ -71,8 +71,9 @@ void _ScaleAndShift(const XTensor * a, XTensor * b, DTYPE scale, DTYPE shift) ...@@ -71,8 +71,9 @@ void _ScaleAndShift(const XTensor * a, XTensor * b, DTYPE scale, DTYPE shift)
else{ else{
DTYPE * va = (DTYPE*)a->data; DTYPE * va = (DTYPE*)a->data;
DTYPE * vb = (DTYPE*)b->data; DTYPE * vb = (DTYPE*)b->data;
if(shift == 0 && useBLAS && a==b){ #if defined(USE_BLAS)
cblas_sscal(b->unitNum, scale, vb, 1); if(shift == 0 && a==b){
SCAL(b->unitNum, scale, vb, 1);
} else{ } else{
for(int i = 0; i < b->unitNum; i++){ for(int i = 0; i < b->unitNum; i++){
*vb = *va * scale + shift; *vb = *va * scale + shift;
...@@ -80,10 +81,17 @@ void _ScaleAndShift(const XTensor * a, XTensor * b, DTYPE scale, DTYPE shift) ...@@ -80,10 +81,17 @@ void _ScaleAndShift(const XTensor * a, XTensor * b, DTYPE scale, DTYPE shift)
vb++; vb++;
} }
} }
#else
for(int i = 0; i < b->unitNum; i++){
*vb = *va * scale + shift;
va++;
vb++;
}
#endif
} }
} }
/* /*
scale and shift all tensor entires (do it on site) scale and shift all tensor entires (do it on site)
keep the result in the input tensor a and return nothing keep the result in the input tensor a and return nothing
...@@ -98,7 +106,7 @@ void _ScaleAndShiftMe(XTensor * a, DTYPE scale, DTYPE shift) ...@@ -98,7 +106,7 @@ void _ScaleAndShiftMe(XTensor * a, DTYPE scale, DTYPE shift)
_ScaleAndShift(a, a, scale, shift); _ScaleAndShift(a, a, scale, shift);
} }
/* /*
scale and shift all tensor entires (return an XTensor structure) scale and shift all tensor entires (return an XTensor structure)
make a new tensor to keep the result and return it make a new tensor to keep the result and return it
...@@ -113,15 +121,15 @@ XTensor ScaleAndShift(const XTensor &a, DTYPE scale, DTYPE shift) ...@@ -113,15 +121,15 @@ XTensor ScaleAndShift(const XTensor &a, DTYPE scale, DTYPE shift)
{ {
XTensor b(&a); XTensor b(&a);
b.SetTMPFlag(); b.SetTMPFlag();
/* call _ScaleAndShift function */ /* call _ScaleAndShift function */
_ScaleAndShift(&a, &b, scale, shift); _ScaleAndShift(&a, &b, scale, shift);
/* tensor connections */ /* tensor connections */
XLink::MakeLink(&a, NULL, &b, MATH_SCALEANDSHIFT); XLink::MakeLink(&a, NULL, &b, MATH_SCALEANDSHIFT);
XLink::AddParamToHead(&b, scale); XLink::AddParamToHead(&b, scale);
XLink::AddParamToHead(&b, shift); XLink::AddParamToHead(&b, shift);
return b; return b;
} }
......
source/tensor/core/reduce/ReduceMax.cpp
/* NiuTrans.Tensor - an open-source tensor library /* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University. * Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved. * All rights reserved.
* *
* Licensed under the Apache License, Version 2.0 (the "License"); * Licensed under the Apache License, Version 2.0 (the "License");
...@@ -28,7 +28,7 @@ ...@@ -28,7 +28,7 @@
namespace nts{ // namespace nts(NiuTrans.Tensor) namespace nts{ // namespace nts(NiuTrans.Tensor)
/* /*
get the max value of the items along a dimension of the tensor get the max value of the items along a dimension of the tensor
>> input - the input tensor >> input - the input tensor
...@@ -37,23 +37,23 @@ get the max value of the items along a dimension of the tensor ...@@ -37,23 +37,23 @@ get the max value of the items along a dimension of the tensor
*/ */
void _ReduceMax(const XTensor * input, XTensor * output, int dim) void _ReduceMax(const XTensor * input, XTensor * output, int dim)
{ {
CheckNTErrors((input->devID == output->devID || (input->devID < 0 && output->devID < 0)), CheckNTErrors((input->devID == output->devID || (input->devID < 0 && output->devID < 0)),
"This code must be run on the same device!"); "This code must be run on the same device!");
CheckNTErrors((input && output), "Empty input or output tensors!"); CheckNTErrors((input && output), "Empty input or output tensors!");
CheckNTErrors((input->order == output->order + 1), "Incorrect tensor sizes!"); CheckNTErrors((input->order == output->order + 1), "Incorrect tensor sizes!");
CheckNTErrors((input->order > dim && dim >=0), "Illegal dimension to reduce!"); CheckNTErrors((input->order > dim && dim >=0), "Illegal dimension to reduce!");
CheckNTErrors((input->dataType == output->dataType), "Unmatched data types!"); CheckNTErrors((input->dataType == output->dataType), "Unmatched data types!");
int dimRDI = input->order - dim - 1; int dimRDI = input->order - dim - 1;
CheckNTErrors(dimRDI >= 0, "Wrong dimension!"); CheckNTErrors(dimRDI >= 0, "Wrong dimension!");
for(int i = 0; i < input->order; i++){ for(int i = 0; i < input->order; i++){
if(i < dimRDI){ if(i < dimRDI){
CheckNTErrors((input->dimSizeRDI[i] == output->dimSizeRDI[i]), CheckNTErrors((input->dimSizeRDI[i] == output->dimSizeRDI[i]),
"Unmatched tensors!"); "Unmatched tensors!");
} }
else if(i > dimRDI){ else if(i > dimRDI){
CheckNTErrors((input->dimSizeRDI[i] == output->dimSizeRDI[i - 1]), CheckNTErrors((input->dimSizeRDI[i] == output->dimSizeRDI[i - 1]),
"Unmatched tensors!"); "Unmatched tensors!");
} }
} }
...@@ -82,9 +82,9 @@ void _ReduceMax(const XTensor * input, XTensor * output, int dim) ...@@ -82,9 +82,9 @@ void _ReduceMax(const XTensor * input, XTensor * output, int dim)
DTYPE * ip = (DTYPE*)input->data + blockSize * k; DTYPE * ip = (DTYPE*)input->data + blockSize * k;
DTYPE * op = (DTYPE*)output->data + stride * k; DTYPE * op = (DTYPE*)output->data + stride * k;
for(int i = 0; i < stride; i++){ for(int i = 0; i < stride; i++){
if(useBLAS){ //#if defined(USE_BLAS)
*(op + i) = cblas_isamax(strideNum, ip + i, stride); // *(op + i) = *(ip + i + (int)(stride * IAMAX(strideNum, ip + i, stride)));
} else{ //#else
DTYPE max = FLOAT_MIN; DTYPE max = FLOAT_MIN;
DTYPE * ipe = ip + blockSize; DTYPE * ipe = ip + blockSize;
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){ for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){
...@@ -93,13 +93,13 @@ void _ReduceMax(const XTensor * input, XTensor * output, int dim) ...@@ -93,13 +93,13 @@ void _ReduceMax(const XTensor * input, XTensor * output, int dim)
max = v; max = v;
} }
*(op + i) = max; *(op + i) = max;
} //#endif
} }
} }
} }
} }
/* /*
get the max value of the items along a dimension of the tensor (return an XTensor structure). get the max value of the items along a dimension of the tensor (return an XTensor structure).
make a new tensor to keep the result and return it make a new tensor to keep the result and return it
...@@ -110,7 +110,7 @@ make a new tensor to keep the result and return it ...@@ -110,7 +110,7 @@ make a new tensor to keep the result and return it
XTensor ReduceMax(const XTensor &input, int dim) XTensor ReduceMax(const XTensor &input, int dim)
{ {
CheckNTErrors(dim >= 0 && dim < input.order, "Illegal dimension to reduce!"); CheckNTErrors(dim >= 0 && dim < input.order, "Illegal dimension to reduce!");
int order = input.order - 1; int order = input.order - 1;
int * dimSize = new int[order]; int * dimSize = new int[order];
for(int i = 0; i < order; i++){ for(int i = 0; i < order; i++){
...@@ -126,7 +126,7 @@ XTensor ReduceMax(const XTensor &input, int dim) ...@@ -126,7 +126,7 @@ XTensor ReduceMax(const XTensor &input, int dim)
/* call _ReduceMax function */ /* call _ReduceMax function */
_ReduceMax(&input, &output, dim); _ReduceMax(&input, &output, dim);
/* tensor connection */ /* tensor connection */
XLink::MakeLink(&input, NULL, &output, REDUCE_REDUCEMAX); XLink::MakeLink(&input, NULL, &output, REDUCE_REDUCEMAX);
XLink::AddParamToHeadInt(&output, dim); XLink::AddParamToHeadInt(&output, dim);
......
source/tensor/core/reduce/ReduceSum.cpp
/* NiuTrans.Tensor - an open-source tensor library /* NiuTrans.Tensor - an open-source tensor library
* Copyright (C) 2017, Natural Language Processing Lab, Northestern University. * Copyright (C) 2017, Natural Language Processing Lab, Northestern University.
* All rights reserved. * All rights reserved.
* *
* Licensed under the Apache License, Version 2.0 (the "License"); * Licensed under the Apache License, Version 2.0 (the "License");
...@@ -25,10 +25,11 @@ ...@@ -25,10 +25,11 @@
#include "../../XName.h" #include "../../XName.h"
#include "../../XBLAS.h" #include "../../XBLAS.h"
#include "../arithmetic/XTensorBLAS.h" #include "../arithmetic/XTensorBLAS.h"
#include <iostream>
namespace nts{ // namespace nts(NiuTrans.Tensor) namespace nts{ // namespace nts(NiuTrans.Tensor)
/* /*
sum the items along a dimension of the tensor sum the items along a dimension of the tensor
For a 1-dimensional data array a, For a 1-dimensional data array a,
...@@ -44,7 +45,7 @@ sum = \sum_i exp((a_i - shift)^power) if isExp == true ...@@ -44,7 +45,7 @@ sum = \sum_i exp((a_i - shift)^power) if isExp == true
*/ */
void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor * shift, DTYPE power, bool isExp) void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor * shift, DTYPE power, bool isExp)
{ {
CheckNTErrors((input->devID == output->devID || (input->devID < 0 && output->devID < 0)), CheckNTErrors((input->devID == output->devID || (input->devID < 0 && output->devID < 0)),
"This code must be run on the same device!"); "This code must be run on the same device!");
CheckNTErrors((input && output), "Empty input or output tensors!"); CheckNTErrors((input && output), "Empty input or output tensors!");
CheckNTErrors((input->order == output->order + 1), "Incorrect tensor sizes!"); CheckNTErrors((input->order == output->order + 1), "Incorrect tensor sizes!");
...@@ -145,22 +146,23 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor ...@@ -145,22 +146,23 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor
else{ else{
if(bias == 0){ if(bias == 0){
if(power == (DTYPE)1.0){ if(power == (DTYPE)1.0){
if(useBLAS) //#if defined(USE_BLAS)
sum = cblas_sasum(strideNum, ip + i, stride); // sum = ASUM(strideNum, ip + i, stride);
else //#else
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride) for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride)
sum += *ipb; sum += *ipb;
//#endif
} }
else if(power == (DTYPE)2.0){ else if(power == (DTYPE)2.0){
if(useBLAS){ //#if defined(USE_BLAS)
sum = cblas_snrm2(strideNum, ip + i, stride); // sum = NRM2(strideNum, ip + i, stride);
sum = sum * sum; // sum = sum * sum;
} else{ //#else
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){ for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){
DTYPE value = (*ipb); DTYPE value = (*ipb);
sum += value * value; sum += value * value;
} }
} //#endif
} }
else if(power == (DTYPE)0.5){ else if(power == (DTYPE)0.5){
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){ for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride){
...@@ -177,11 +179,12 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor ...@@ -177,11 +179,12 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor
} }
else{ else{
if(power == (DTYPE)1.0){ if(power == (DTYPE)1.0){
if(useBLAS) //#if defined(USE_BLAS)
sum = cblas_sasum(strideNum, ip + i, stride); // sum = ASUM(strideNum, ip + i, stride);
else //#else
for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride) for(DTYPE * ipb = ip + i; ipb < ipe; ipb += stride)
sum += *ipb; sum += *ipb;
//#endif
sum -= strideNum * bias; sum -= strideNum * bias;
} }
else if(power == (DTYPE)2.0){ else if(power == (DTYPE)2.0){
...@@ -210,7 +213,7 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor ...@@ -210,7 +213,7 @@ void _ReduceSum(const XTensor * input, XTensor * output, int dim, const XTensor
} }
} }
/* /*
sum the items along a dimension of the tensor (return an XTensor structure) sum the items along a dimension of the tensor (return an XTensor structure)
make a new tensor to keep the result and return it make a new tensor to keep the result and return it
...@@ -228,7 +231,7 @@ sum = \sum_i exp((a_i - shift)^power) if isExp == true ...@@ -228,7 +231,7 @@ sum = \sum_i exp((a_i - shift)^power) if isExp == true
XTensor ReduceSum(const XTensor &input, int dim, const XTensor &shift, DTYPE power, bool isExp) XTensor ReduceSum(const XTensor &input, int dim, const XTensor &shift, DTYPE power, bool isExp)
{ {
CheckNTErrors(dim >= 0 && dim < input.order, "Illegal dimension to reduce!"); CheckNTErrors(dim >= 0 && dim < input.order, "Illegal dimension to reduce!");
int order = input.order - 1; int order = input.order - 1;
int * dimSize = new int[order]; int * dimSize = new int[order];
for(int i = 0; i < order; i++){ for(int i = 0; i < order; i++){
...@@ -244,7 +247,7 @@ XTensor ReduceSum(const XTensor &input, int dim, const XTensor &shift, DTYPE pow ...@@ -244,7 +247,7 @@ XTensor ReduceSum(const XTensor &input, int dim, const XTensor &shift, DTYPE pow
/* call _ReduceSum function */ /* call _ReduceSum function */
_ReduceSum(&input, &output, dim, &shift, power, isExp); _ReduceSum(&input, &output, dim, &shift, power, isExp);
/* tensor connection */ /* tensor connection */
XLink::MakeLink(&input, &shift, &output, REDUCE_REDUCESUM); XLink::MakeLink(&input, &shift, &output, REDUCE_REDUCESUM);
XLink::AddParamToHeadInt(&output, dim); XLink::AddParamToHeadInt(&output, dim);
...@@ -257,7 +260,7 @@ XTensor ReduceSum(const XTensor &input, int dim, const XTensor &shift, DTYPE pow ...@@ -257,7 +260,7 @@ XTensor ReduceSum(const XTensor &input, int dim, const XTensor &shift, DTYPE pow
return output; return output;
} }
/* /*
sum the items along a dimension of the tensor (return an XTensor structure) sum the items along a dimension of the tensor (return an XTensor structure)
make a new tensor to keep the result and return it make a new tensor to keep the result and return it
...@@ -274,7 +277,7 @@ sum = \sum_i exp((a_i)^power) if isExp == true ...@@ -274,7 +277,7 @@ sum = \sum_i exp((a_i)^power) if isExp == true
XTensor ReduceSum(const XTensor &input, int dim, DTYPE power, bool isExp) XTensor ReduceSum(const XTensor &input, int dim, DTYPE power, bool isExp)
{ {
CheckNTErrors(dim >= 0 && dim < input.order, "Illegal dimension to reduce!"); CheckNTErrors(dim >= 0 && dim < input.order, "Illegal dimension to reduce!");
int order = input.order - 1; int order = input.order - 1;
int * dimSize = new int[order]; int * dimSize = new int[order];
for(int i = 0; i < order; i++){ for(int i = 0; i < order; i++){
...@@ -290,7 +293,7 @@ XTensor ReduceSum(const XTensor &input, int dim, DTYPE power, bool isExp) ...@@ -290,7 +293,7 @@ XTensor ReduceSum(const XTensor &input, int dim, DTYPE power, bool isExp)
/* call _ReduceSum function */ /* call _ReduceSum function */
_ReduceSum(&input, &output, dim, NULL, power, isExp); _ReduceSum(&input, &output, dim, NULL, power, isExp);
/* tensor connection */ /* tensor connection */
XLink::MakeLink(&input, NULL, &output, REDUCE_REDUCESUM); XLink::MakeLink(&input, NULL, &output, REDUCE_REDUCESUM);
XLink::AddParamToHeadInt(&output, dim); XLink::AddParamToHeadInt(&output, dim);
......
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