Fixed #778. Merge branch 'buffer51-develop' into develop

This commit is contained in:
Zhang Xianyi 2016-02-05 08:39:08 +08:00
commit be4e5fcd20
7 changed files with 36 additions and 34 deletions

View File

@ -11,8 +11,8 @@ endif
ifeq ($(CORE), ARMV7)
ifeq ($(OSNAME), Android)
CCOMMON_OPT += -marm -mfpu=neon -mfloat-abi=hard -march=armv7-a
FCOMMON_OPT += -marm -mfpu=neon -mfloat-abi=hard -march=armv7-a
CCOMMON_OPT += -marm -mfpu=neon -mfloat-abi=hard -march=armv7-a -Wl,--no-warn-mismatch
FCOMMON_OPT += -marm -mfpu=neon -mfloat-abi=hard -march=armv7-a -Wl,--no-warn-mismatch
else
CCOMMON_OPT += -marm -mfpu=vfpv3 -mfloat-abi=hard -march=armv7-a
FCOMMON_OPT += -marm -mfpu=vfpv3 -mfloat-abi=hard -march=armv7-a
@ -29,5 +29,3 @@ ifeq ($(CORE), ARMV5)
CCOMMON_OPT += -marm -march=armv5
FCOMMON_OPT += -marm -march=armv5
endif

View File

@ -296,12 +296,14 @@ endif
ifneq ($(OSNAME), WINNT)
ifneq ($(OSNAME), CYGWIN_NT)
ifneq ($(OSNAME), Interix)
ifneq ($(OSNAME), Android)
ifdef SMP
EXTRALIB += -lpthread
endif
endif
endif
endif
endif
# ifeq logical or
ifeq ($(OSNAME), $(filter $(OSNAME),WINNT CYGWIN_NT Interix))
@ -379,7 +381,7 @@ FCOMMON_OPT += -m128bit-long-double
endif
ifeq ($(C_COMPILER), CLANG)
EXPRECISION = 1
CCOMMON_OPT += -DEXPRECISION
CCOMMON_OPT += -DEXPRECISION
FCOMMON_OPT += -m128bit-long-double
endif
endif
@ -393,7 +395,7 @@ endif
ifeq ($(USE_OPENMP), 1)
#check
#check
ifeq ($(USE_THREAD), 0)
$(error OpenBLAS: Cannot set both USE_OPENMP=1 and USE_THREAD=0. The USE_THREAD=0 is only for building single thread version.)
endif
@ -1185,4 +1187,3 @@ SUNPATH = /opt/sunstudio12.1
else
SUNPATH = /opt/SUNWspro
endif

View File

@ -34,6 +34,13 @@
#ifndef _LAPACKE_CONFIG_H_
#define _LAPACKE_CONFIG_H_
// For Android prior to API 21 (no <complex> include)
#if defined(__ANDROID__)
#if __ANDROID_API__ < 21
#define LAPACK_COMPLEX_STRUCTURE
#endif
#endif
#ifdef __cplusplus
#if defined(LAPACK_COMPLEX_CPP)
#include <complex>

View File

@ -46,7 +46,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
BLASLONG n, lda;
FLOAT *a;
FLOAT temp[2];
FLOAT temp;
BLASLONG i;
n = args -> n;
@ -64,12 +64,11 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
a + i * COMPSIZE, lda, NULL, 0, NULL, 0);
if (i < n - 1) {
temp[0] = DOTC_K(n - i - 1,
temp = CREAL(DOTC_K(n - i - 1,
a + (i + 1 + i * lda) * COMPSIZE, 1,
a + (i + 1 + i * lda) * COMPSIZE, 1);
GET_IMAGE(temp[1]);
a + (i + 1 + i * lda) * COMPSIZE, 1));
*(a + (i + i * lda) * COMPSIZE + 0) += temp[0];
*(a + (i + i * lda) * COMPSIZE + 0) += temp;
*(a + (i + i * lda) * COMPSIZE + 1) = ZERO;
GEMV_U(n - i - 1, i, 0, dp1, ZERO,

View File

@ -46,7 +46,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
BLASLONG n, lda;
FLOAT *a;
FLOAT temp[2];
FLOAT temp;
BLASLONG i;
n = args -> n;
@ -64,10 +64,9 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
a + i * lda * COMPSIZE, 1, NULL, 0, NULL, 0);
if (i < n - 1) {
temp[0] = DOTC_K(n - i - 1, a + (i + (i + 1) * lda) * COMPSIZE, lda, a + (i + (i + 1) * lda) * COMPSIZE, lda);
GET_IMAGE(temp[1]);
temp = CREAL(DOTC_K(n - i - 1, a + (i + (i + 1) * lda) * COMPSIZE, lda, a + (i + (i + 1) * lda) * COMPSIZE, lda));
*(a + (i + i * lda) * COMPSIZE + 0) += temp[0];
*(a + (i + i * lda) * COMPSIZE + 0) += temp;
*(a + (i + i * lda) * COMPSIZE + 1) = ZERO;
GEMV_O(i, n - i - 1, 0, dp1, ZERO,

View File

@ -51,7 +51,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
BLASLONG n, lda;
FLOAT *a;
FLOAT ajj[2];
FLOAT ajj;
FLOAT *aoffset;
BLASLONG i, j;
@ -68,18 +68,17 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
for (j = 0; j < n; j++) {
ajj[0] = DOTC_K(j, a + j * 2, lda, a + j * 2, lda);
GET_IMAGE(ajj[1]);
ajj = CREAL(DOTC_K(j, a + j * 2, lda, a + j * 2, lda));
ajj[0] = *(aoffset + j * 2) - ajj[0];
ajj = *(aoffset + j * 2) - ajj;
if (ajj[0] <= 0){
*(aoffset + j * 2 + 0) = ajj[0];
if (ajj <= 0){
*(aoffset + j * 2 + 0) = ajj;
*(aoffset + j * 2 + 1) = ZERO;
return j + 1;
}
ajj[0] = SQRT(ajj[0]);
*(aoffset + j * 2 + 0) = ajj[0];
ajj = SQRT(ajj);
*(aoffset + j * 2 + 0) = ajj;
*(aoffset + j * 2 + 1) = ZERO;
i = n - j - 1;
@ -90,7 +89,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
a + j * 2, lda,
aoffset + (j + 1) * 2, 1, sb);
SCAL_K(i, 0, 0, ONE / ajj[0], ZERO,
SCAL_K(i, 0, 0, ONE / ajj, ZERO,
aoffset + (j + 1) * 2, 1, NULL, 0, NULL, 0);
}

View File

@ -51,7 +51,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
BLASLONG n, lda;
FLOAT *a;
FLOAT ajj[2];
FLOAT ajj;
BLASLONG i, j;
n = args -> n;
@ -65,19 +65,18 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
for (j = 0; j < n; j++) {
ajj[0] = DOTC_K(j, a, 1, a, 1);
GET_IMAGE(ajj[1]);
ajj = CREAL(DOTC_K(j, a, 1, a, 1));
ajj[0] = *(a + j * 2) - ajj[0];
ajj = *(a + j * 2) - ajj;
if (ajj[0] <= 0){
*(a + j * 2 + 0) = ajj[0];
if (ajj <= 0){
*(a + j * 2 + 0) = ajj;
*(a + j * 2 + 1) = ZERO;
return j + 1;
}
ajj[0] = SQRT(ajj[0]);
*(a + j * 2 + 0) = ajj[0];
ajj = SQRT(ajj);
*(a + j * 2 + 0) = ajj;
*(a + j * 2 + 1) = ZERO;
i = n - j - 1;
@ -88,7 +87,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
a, 1,
a + (j + lda) * 2, lda, sb);
SCAL_K(i, 0, 0, ONE / ajj[0], ZERO,
SCAL_K(i, 0, 0, ONE / ajj, ZERO,
a + (j + lda) * 2, lda, NULL, 0, NULL, 0);
}