Browse Source

Merge pull request #1410 from brada4/develop

Address warnings #1357
tags/v0.3.0
Martin Kroeker GitHub 8 years ago
parent
commit
42285d8e70
No known key found for this signature in database GPG Key ID: 4AEE18F83AFDEB23
23 changed files with 75 additions and 73 deletions
  1. +3
    -3
      driver/level3/level3.c
  2. +2
    -2
      driver/level3/level3_syr2k.c
  3. +2
    -2
      driver/level3/level3_syrk.c
  4. +2
    -2
      driver/level3/level3_syrk_threaded.c
  5. +2
    -2
      driver/level3/level3_thread.c
  6. +2
    -2
      driver/others/init.c
  7. +1
    -1
      kernel/generic/laswp_ncopy_8.c
  8. +1
    -1
      kernel/generic/neg_tcopy_16.c
  9. +4
    -4
      kernel/generic/neg_tcopy_8.c
  10. +10
    -10
      kernel/generic/trmm_lncopy_8.c
  11. +15
    -15
      kernel/generic/trmm_utcopy_8.c
  12. +4
    -4
      kernel/generic/trsm_uncopy_8.c
  13. +2
    -2
      kernel/generic/trsm_utcopy_8.c
  14. +1
    -1
      kernel/generic/zlaswp_ncopy_2.c
  15. +1
    -1
      kernel/generic/zlaswp_ncopy_4.c
  16. +2
    -2
      kernel/generic/zneg_tcopy_2.c
  17. +7
    -7
      kernel/generic/zneg_tcopy_4.c
  18. +2
    -2
      kernel/generic/zneg_tcopy_8.c
  19. +2
    -2
      kernel/x86_64/dgemv_n_4.c
  20. +2
    -2
      kernel/x86_64/dgemv_t_4.c
  21. +2
    -2
      lapack/getrf/getrf_single.c
  22. +3
    -2
      lapack/trti2/ztrti2_L.c
  23. +3
    -2
      lapack/trti2/ztrti2_U.c

+ 3
- 3
driver/level3/level3.c View File

@@ -251,11 +251,11 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
if ((k == 0) || (alpha == NULL)) return 0; if ((k == 0) || (alpha == NULL)) return 0;


#if !defined(XDOUBLE) || !defined(QUAD_PRECISION) #if !defined(XDOUBLE) || !defined(QUAD_PRECISION)
if ((alpha[0] == ZERO)
if ( alpha[0] == ZERO
#ifdef COMPLEX #ifdef COMPLEX
&& (alpha[1] == ZERO)
&& alpha[1] == ZERO
#endif #endif
) return 0;
) return 0;
#else #else
if (((alpha[0].x[0] | alpha[0].x[1] if (((alpha[0].x[0] | alpha[0].x[1]
#ifdef COMPLEX #ifdef COMPLEX


+ 2
- 2
driver/level3/level3_syr2k.c View File

@@ -154,9 +154,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO


if ((k == 0) || (alpha == NULL)) return 0; if ((k == 0) || (alpha == NULL)) return 0;


if ((alpha[0] == ZERO)
if (alpha[0] == ZERO
#ifdef COMPLEX #ifdef COMPLEX
&& (alpha[1] == ZERO)
&& alpha[1] == ZERO
#endif #endif
) return 0; ) return 0;




+ 2
- 2
driver/level3/level3_syrk.c View File

@@ -158,9 +158,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO


if ((k == 0) || (alpha == NULL)) return 0; if ((k == 0) || (alpha == NULL)) return 0;


if ((alpha[0] == ZERO)
if (alpha[0] == ZERO
#if defined(COMPLEX) && !defined(HERK) #if defined(COMPLEX) && !defined(HERK)
&& (alpha[1] == ZERO)
&& alpha[1] == ZERO
#endif #endif
) return 0; ) return 0;




+ 2
- 2
driver/level3/level3_syrk_threaded.c View File

@@ -200,9 +200,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,


if ((k == 0) || (alpha == NULL)) return 0; if ((k == 0) || (alpha == NULL)) return 0;


if ((alpha[0] == ZERO)
if (alpha[0] == ZERO
#if defined(COMPLEX) && !defined(HERK) #if defined(COMPLEX) && !defined(HERK)
&& (alpha[1] == ZERO)
&& alpha[1] == ZERO
#endif #endif
) return 0; ) return 0;




+ 2
- 2
driver/level3/level3_thread.c View File

@@ -295,9 +295,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,


/* Return early if no more computation is needed */ /* Return early if no more computation is needed */
if ((k == 0) || (alpha == NULL)) return 0; if ((k == 0) || (alpha == NULL)) return 0;
if ((alpha[0] == ZERO)
if (alpha[0] == ZERO
#ifdef COMPLEX #ifdef COMPLEX
&& (alpha[1] == ZERO)
&& alpha[1] == ZERO
#endif #endif
) return 0; ) return 0;




+ 2
- 2
driver/others/init.c View File

@@ -237,7 +237,7 @@ static inline void get_cpumap(int node, unsigned long * node_info) {
if(k!=0){ if(k!=0){
name[k]='\0'; name[k]='\0';
affinity[count++] = strtoul(name, &dummy, 16); affinity[count++] = strtoul(name, &dummy, 16);
k=0;
// k=0;
} }
// 0-63bit -> node_info[0], 64-128bit -> node_info[1] .... // 0-63bit -> node_info[0], 64-128bit -> node_info[1] ....
// revert the sequence // revert the sequence
@@ -293,7 +293,7 @@ static inline void get_share(int cpu, int level, unsigned long * share) {
if(k!=0){ if(k!=0){
name[k]='\0'; name[k]='\0';
affinity[count++] = strtoul(name, &dummy, 16); affinity[count++] = strtoul(name, &dummy, 16);
k=0;
// k=0;
} }
// 0-63bit -> node_info[0], 64-128bit -> node_info[1] .... // 0-63bit -> node_info[0], 64-128bit -> node_info[1] ....
// revert the sequence // revert the sequence


+ 1
- 1
kernel/generic/laswp_ncopy_8.c View File

@@ -288,7 +288,7 @@ int CNAME(BLASLONG n, BLASLONG k1, BLASLONG k2, FLOAT *a, BLASLONG lda, blasint
i++; i++;
} while (i <= k2); } while (i <= k2);


a += lda;
// a += lda;
} }


return 0; return 0;


+ 1
- 1
kernel/generic/neg_tcopy_16.c View File

@@ -379,7 +379,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
if (m & 1){ if (m & 1){
ctemp01 = *(aoffset1 + 0); ctemp01 = *(aoffset1 + 0);
*(boffset + 0) = -ctemp01; *(boffset + 0) = -ctemp01;
boffset += 1;
// boffset += 1;
} }
} }




+ 4
- 4
kernel/generic/neg_tcopy_8.c View File

@@ -719,10 +719,10 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){


if (m & 1){ if (m & 1){
aoffset1 = aoffset; aoffset1 = aoffset;
aoffset += lda;
// aoffset += lda;


boffset1 = boffset; boffset1 = boffset;
boffset += 8;
// boffset += 8;


i = (n >> 3); i = (n >> 3);
if (i > 0){ if (i > 0){
@@ -762,7 +762,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
*(boffset2 + 1) = -ctemp02; *(boffset2 + 1) = -ctemp02;
*(boffset2 + 2) = -ctemp03; *(boffset2 + 2) = -ctemp03;
*(boffset2 + 3) = -ctemp04; *(boffset2 + 3) = -ctemp04;
boffset2 += 4;
// boffset2 += 4;
} }


if (n & 2){ if (n & 2){
@@ -772,7 +772,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){


*(boffset3 + 0) = -ctemp01; *(boffset3 + 0) = -ctemp01;
*(boffset3 + 1) = -ctemp02; *(boffset3 + 1) = -ctemp02;
boffset3 += 2;
// boffset3 += 2;
} }


if (n & 1){ if (n & 1){


+ 10
- 10
kernel/generic/trmm_lncopy_8.c View File

@@ -635,7 +635,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data05 = *(ao1 + 4); data05 = *(ao1 + 4);
data06 = *(ao1 + 5); data06 = *(ao1 + 5);
data07 = *(ao1 + 6); data07 = *(ao1 + 6);
data08 = *(ao1 + 7);
// data08 = *(ao1 + 7);


if (i >= 2) { if (i >= 2) {
#ifndef UNIT #ifndef UNIT
@@ -646,7 +646,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data13 = *(ao2 + 4); data13 = *(ao2 + 4);
data14 = *(ao2 + 5); data14 = *(ao2 + 5);
data15 = *(ao2 + 6); data15 = *(ao2 + 6);
data16 = *(ao2 + 7);
// data16 = *(ao2 + 7);
} }


if (i >= 3) { if (i >= 3) {
@@ -657,7 +657,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data21 = *(ao3 + 4); data21 = *(ao3 + 4);
data22 = *(ao3 + 5); data22 = *(ao3 + 5);
data23 = *(ao3 + 6); data23 = *(ao3 + 6);
data24 = *(ao3 + 7);
// data24 = *(ao3 + 7);
} }


if (i >= 4) { if (i >= 4) {
@@ -667,7 +667,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data29 = *(ao4 + 4); data29 = *(ao4 + 4);
data30 = *(ao4 + 5); data30 = *(ao4 + 5);
data31 = *(ao4 + 6); data31 = *(ao4 + 6);
data32 = *(ao4 + 7);
// data32 = *(ao4 + 7);
} }


if (i >= 5) { if (i >= 5) {
@@ -676,7 +676,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
#endif #endif
data38 = *(ao5 + 5); data38 = *(ao5 + 5);
data39 = *(ao5 + 6); data39 = *(ao5 + 6);
data40 = *(ao5 + 7);
// data40 = *(ao5 + 7);
} }


if (i >= 6) { if (i >= 6) {
@@ -684,14 +684,14 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data46 = *(ao6 + 5); data46 = *(ao6 + 5);
#endif #endif
data47 = *(ao6 + 6); data47 = *(ao6 + 6);
data48 = *(ao6 + 7);
// data48 = *(ao6 + 7);
} }


if (i >= 7) { if (i >= 7) {
#ifndef UNIT #ifndef UNIT
data55 = *(ao7 + 6); data55 = *(ao7 + 6);
#endif #endif
data56 = *(ao7 + 7);
// data56 = *(ao7 + 7);
} }


#ifdef UNIT #ifdef UNIT
@@ -1032,14 +1032,14 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
#endif #endif
data02 = *(ao1 + 1); data02 = *(ao1 + 1);
data03 = *(ao1 + 2); data03 = *(ao1 + 2);
data04 = *(ao1 + 3);
// data04 = *(ao1 + 3);


if (i >= 2) { if (i >= 2) {
#ifndef UNIT #ifndef UNIT
data10 = *(ao2 + 1); data10 = *(ao2 + 1);
#endif #endif
data11 = *(ao2 + 2); data11 = *(ao2 + 2);
data12 = *(ao2 + 3);
// data12 = *(ao2 + 3);


} }


@@ -1047,7 +1047,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
#ifndef UNIT #ifndef UNIT
data19 = *(ao3 + 2); data19 = *(ao3 + 2);
#endif #endif
data20 = *(ao3 + 3);
// data20 = *(ao3 + 3);
} }


#ifdef UNIT #ifdef UNIT


+ 15
- 15
kernel/generic/trmm_utcopy_8.c View File

@@ -685,7 +685,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data33 = *(ao5 + 0); data33 = *(ao5 + 0);
data41 = *(ao6 + 0); data41 = *(ao6 + 0);
data49 = *(ao7 + 0); data49 = *(ao7 + 0);
data57 = *(ao8 + 0);
// data57 = *(ao8 + 0);


if (i >= 2) { if (i >= 2) {
#ifndef UNIT #ifndef UNIT
@@ -696,7 +696,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data34 = *(ao5 + 1); data34 = *(ao5 + 1);
data42 = *(ao6 + 1); data42 = *(ao6 + 1);
data50 = *(ao7 + 1); data50 = *(ao7 + 1);
data58 = *(ao8 + 1);
// data58 = *(ao8 + 1);
} }


if (i >= 3) { if (i >= 3) {
@@ -707,7 +707,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data35 = *(ao5 + 2); data35 = *(ao5 + 2);
data43 = *(ao6 + 2); data43 = *(ao6 + 2);
data51 = *(ao7 + 2); data51 = *(ao7 + 2);
data59 = *(ao8 + 2);
// data59 = *(ao8 + 2);
} }


if (i >= 4) { if (i >= 4) {
@@ -717,7 +717,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data36 = *(ao5 + 3); data36 = *(ao5 + 3);
data44 = *(ao6 + 3); data44 = *(ao6 + 3);
data52 = *(ao7 + 3); data52 = *(ao7 + 3);
data60 = *(ao8 + 3);
// data60 = *(ao8 + 3);
} }


if (i >= 5) { if (i >= 5) {
@@ -726,7 +726,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
#endif #endif
data45 = *(ao6 + 4); data45 = *(ao6 + 4);
data53 = *(ao7 + 4); data53 = *(ao7 + 4);
data61 = *(ao8 + 4);
// data61 = *(ao8 + 4);
} }


if (i >= 6) { if (i >= 6) {
@@ -734,14 +734,14 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
data46 = *(ao6 + 5); data46 = *(ao6 + 5);
#endif #endif
data54 = *(ao7 + 5); data54 = *(ao7 + 5);
data62 = *(ao8 + 5);
// data62 = *(ao8 + 5);
} }


if (i >= 7) { if (i >= 7) {
#ifndef UNIT #ifndef UNIT
data55 = *(ao7 + 6); data55 = *(ao7 + 6);
#endif #endif
data63 = *(ao8 + 6);
// data63 = *(ao8 + 6);
} }


#ifdef UNIT #ifdef UNIT
@@ -1074,27 +1074,27 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
} }
} else { } else {


#ifndef UNIT
/* #ifndef UNIT
data01 = *(ao1 + 0); data01 = *(ao1 + 0);
#endif
#endif */
data09 = *(ao2 + 0); data09 = *(ao2 + 0);
data17 = *(ao3 + 0); data17 = *(ao3 + 0);
data25 = *(ao4 + 0);
// data25 = *(ao4 + 0);


if (i >= 2) { if (i >= 2) {
#ifndef UNIT
/* #ifndef UNIT
data10 = *(ao2 + 1); data10 = *(ao2 + 1);
#endif
#endif */
data18 = *(ao3 + 1); data18 = *(ao3 + 1);
data26 = *(ao4 + 1);
// data26 = *(ao4 + 1);
} }


if (i >= 3) {
/* if (i >= 3) {
#ifndef UNIT #ifndef UNIT
data19 = *(ao3 + 2); data19 = *(ao3 + 2);
#endif #endif
data27 = *(ao4 + 2); data27 = *(ao4 + 2);
}
} */


#ifndef UNIT #ifndef UNIT
b[ 0] = ONE; b[ 0] = ONE;


+ 4
- 4
kernel/generic/trsm_uncopy_8.c View File

@@ -628,13 +628,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT


if (ii < jj) { if (ii < jj) {
data01 = *(a1 + 0); data01 = *(a1 + 0);
data02 = *(a1 + 1);
// data02 = *(a1 + 1);
data09 = *(a2 + 0); data09 = *(a2 + 0);
data10 = *(a2 + 1);
// data10 = *(a2 + 1);
data17 = *(a3 + 0); data17 = *(a3 + 0);
data18 = *(a3 + 1);
// data18 = *(a3 + 1);
data25 = *(a4 + 0); data25 = *(a4 + 0);
data26 = *(a4 + 1);
// data26 = *(a4 + 1);


*(b + 0) = data01; *(b + 0) = data01;
*(b + 1) = data09; *(b + 1) = data09;


+ 2
- 2
kernel/generic/trsm_utcopy_8.c View File

@@ -649,14 +649,14 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG offset, FLOAT


if (m & 2) { if (m & 2) {
if (ii == jj) { if (ii == jj) {
#ifndef UNIT
/* #ifndef UNIT
data01 = *(a1 + 0); data01 = *(a1 + 0);
#endif #endif


data09 = *(a2 + 0); data09 = *(a2 + 0);
#ifndef UNIT #ifndef UNIT
data10 = *(a2 + 1); data10 = *(a2 + 1);
#endif
#endif */
} }


if (ii > jj) { if (ii > jj) {


+ 1
- 1
kernel/generic/zlaswp_ncopy_2.c View File

@@ -372,7 +372,7 @@ int CNAME(BLASLONG n, BLASLONG k1, BLASLONG k2, FLOAT *a, BLASLONG lda, blasint
*(b1 + 0) = A1; *(b1 + 0) = A1;
*(b1 + 1) = A2; *(b1 + 1) = A2;
} }
buffer += 2;
// buffer += 2;
} }
} }




+ 1
- 1
kernel/generic/zlaswp_ncopy_4.c View File

@@ -702,7 +702,7 @@ int CNAME(BLASLONG n, BLASLONG k1, BLASLONG k2, FLOAT *a, BLASLONG lda, blasint
*(b1 + 0) = A1; *(b1 + 0) = A1;
*(b1 + 1) = A2; *(b1 + 1) = A2;
} }
buffer += 2;
// buffer += 2;
} }
} }




+ 2
- 2
kernel/generic/zneg_tcopy_2.c View File

@@ -140,7 +140,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
*(b_offset1 + 6) = -ctemp11; *(b_offset1 + 6) = -ctemp11;
*(b_offset1 + 7) = -ctemp12; *(b_offset1 + 7) = -ctemp12;


b_offset1 += m * 4;
// b_offset1 += m * 4;
a_offset1 += 4; a_offset1 += 4;
a_offset2 += 4; a_offset2 += 4;
} }
@@ -204,7 +204,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
*(b_offset + 2) = -ctemp3; *(b_offset + 2) = -ctemp3;
*(b_offset + 3) = -ctemp4; *(b_offset + 3) = -ctemp4;


b_offset += m * 4;
// b_offset += m * 4;
a_offset += 4; a_offset += 4;
} }




+ 7
- 7
kernel/generic/zneg_tcopy_4.c View File

@@ -233,10 +233,10 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
*(boffset3 + 6) = -ctemp07; *(boffset3 + 6) = -ctemp07;
*(boffset3 + 7) = -ctemp08; *(boffset3 + 7) = -ctemp08;


aoffset1 += 2;
/* aoffset1 += 2;
aoffset2 += 2; aoffset2 += 2;
aoffset3 += 2; aoffset3 += 2;
aoffset4 += 2;
aoffset4 += 2; */


boffset3 += 8; boffset3 += 8;
} }
@@ -293,8 +293,8 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){


aoffset1 += 8; aoffset1 += 8;
aoffset2 += 8; aoffset2 += 8;
aoffset3 += 8;
aoffset4 += 8;
/* aoffset3 += 8;
aoffset4 += 8; */


boffset1 += m * 8; boffset1 += m * 8;
i --; i --;
@@ -338,8 +338,8 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
*(boffset3 + 2) = -ctemp03; *(boffset3 + 2) = -ctemp03;
*(boffset3 + 3) = -ctemp04; *(boffset3 + 3) = -ctemp04;


aoffset1 += 2;
aoffset2 += 2;
/* aoffset1 += 2;
aoffset2 += 2; */
boffset3 += 4; boffset3 += 4;
} }
} }
@@ -387,7 +387,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
*(boffset2 + 3) = -ctemp04; *(boffset2 + 3) = -ctemp04;


aoffset1 += 4; aoffset1 += 4;
boffset2 += 4;
// boffset2 += 4;
} }


if (n & 1){ if (n & 1){


+ 2
- 2
kernel/generic/zneg_tcopy_8.c View File

@@ -324,7 +324,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
if (n & 1){ if (n & 1){
aoffset1 = aoffset; aoffset1 = aoffset;
aoffset2 = aoffset + lda; aoffset2 = aoffset + lda;
aoffset += 2;
// aoffset += 2;


i = (m >> 1); i = (m >> 1);
if (i > 0){ if (i > 0){
@@ -353,7 +353,7 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){


*(boffset + 0) = -ctemp01; *(boffset + 0) = -ctemp01;
*(boffset + 1) = -ctemp02; *(boffset + 1) = -ctemp02;
boffset += 2;
// boffset += 2;
} }
} }




+ 2
- 2
kernel/x86_64/dgemv_n_4.c View File

@@ -281,8 +281,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
if ( n2 & 1 ) if ( n2 & 1 )
{ {
dgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer,&alpha); dgemv_kernel_4x1(NB,a_ptr,x_ptr,ybuffer,&alpha);
a_ptr += lda;
x_ptr += 1;
/* a_ptr += lda;
x_ptr += 1; */


} }




+ 2
- 2
kernel/x86_64/dgemv_t_4.c View File

@@ -393,9 +393,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
{ {


dgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer); dgemv_kernel_4x1(NB,a_ptr,xbuffer,ybuffer);
a_ptr += lda;
// a_ptr += lda;
*y_ptr += ybuffer[0] * alpha; *y_ptr += ybuffer[0] * alpha;
y_ptr += inc_y;
// y_ptr += inc_y;


} }
a += NB; a += NB;


+ 2
- 2
lapack/getrf/getrf_single.c View File

@@ -51,7 +51,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
BLASLONG jjs, min_jj; BLASLONG jjs, min_jj;
blasint *ipiv, iinfo, info; blasint *ipiv, iinfo, info;
BLASLONG jb, mn, blocking; BLASLONG jb, mn, blocking;
FLOAT *a, *offsetA, *offsetB;
FLOAT *a, *offsetA; //, *offsetB;
BLASLONG range_N[2]; BLASLONG range_N[2];


FLOAT *sbb; FLOAT *sbb;
@@ -99,7 +99,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
if (jb > blocking) jb = blocking; if (jb > blocking) jb = blocking;


offsetA = a + j * lda * COMPSIZE; offsetA = a + j * lda * COMPSIZE;
offsetB = a + (j + jb) * lda * COMPSIZE;
// offsetB = a + (j + jb) * lda * COMPSIZE;


range_N[0] = offset + j; range_N[0] = offset + j;
range_N[1] = offset + j + jb; range_N[1] = offset + j + jb;


+ 3
- 2
lapack/trti2/ztrti2_L.c View File

@@ -67,8 +67,6 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,


for (j = n - 1; j >= 0; j--) { for (j = n - 1; j >= 0; j--) {


ajj_r = ONE;
ajj_i = ZERO;


#ifndef UNIT #ifndef UNIT
ajj_r = *(a + (j + j * lda) * COMPSIZE + 0); ajj_r = *(a + (j + j * lda) * COMPSIZE + 0);
@@ -88,6 +86,9 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,


*(a + (j + j * lda) * COMPSIZE + 0) = ajj_r; *(a + (j + j * lda) * COMPSIZE + 0) = ajj_r;
*(a + (j + j * lda) * COMPSIZE + 1) = ajj_i; *(a + (j + j * lda) * COMPSIZE + 1) = ajj_i;
#else
ajj_r = ONE;
ajj_i = ZERO;
#endif #endif


ZTRMV (n - j - 1, ZTRMV (n - j - 1,


+ 3
- 2
lapack/trti2/ztrti2_U.c View File

@@ -67,8 +67,6 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,


for (j = 0; j < n; j++) { for (j = 0; j < n; j++) {


ajj_r = ONE;
ajj_i = ZERO;


#ifndef UNIT #ifndef UNIT
ajj_r = *(a + (j + j * lda) * COMPSIZE + 0); ajj_r = *(a + (j + j * lda) * COMPSIZE + 0);
@@ -89,6 +87,9 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,


*(a + (j + j * lda) * COMPSIZE + 0) = ajj_r; *(a + (j + j * lda) * COMPSIZE + 0) = ajj_r;
*(a + (j + j * lda) * COMPSIZE + 1) = ajj_i; *(a + (j + j * lda) * COMPSIZE + 1) = ajj_i;
#else
ajj_r = ONE;
ajj_i = ZERO;
#endif #endif


ZTRMV (j, ZTRMV (j,


Loading…
Cancel
Save