Browse Source

LAPACK helpers in C that need care too

tags/v0.3.0
Andrew 8 years ago
parent
commit
d602b99386
11 changed files with 27 additions and 25 deletions
  1. +1
    -1
      kernel/generic/laswp_ncopy_8.c
  2. +1
    -1
      kernel/generic/neg_tcopy_16.c
  3. +4
    -4
      kernel/generic/neg_tcopy_8.c
  4. +1
    -1
      kernel/generic/zlaswp_ncopy_2.c
  5. +1
    -1
      kernel/generic/zlaswp_ncopy_4.c
  6. +2
    -2
      kernel/generic/zneg_tcopy_2.c
  7. +7
    -7
      kernel/generic/zneg_tcopy_4.c
  8. +2
    -2
      kernel/generic/zneg_tcopy_8.c
  9. +2
    -2
      lapack/getrf/getrf_single.c
  10. +3
    -2
      lapack/trti2/ztrti2_L.c
  11. +3
    -2
      lapack/trti2/ztrti2_U.c

+ 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++;
} while (i <= k2);

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

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){
ctemp01 = *(aoffset1 + 0);
*(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){
aoffset1 = aoffset;
aoffset += lda;
// aoffset += lda;

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

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

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

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

if (n & 1){


+ 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 + 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 + 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 + 7) = -ctemp12;

b_offset1 += m * 4;
// b_offset1 += m * 4;
a_offset1 += 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 + 3) = -ctemp4;

b_offset += m * 4;
// b_offset += m * 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 + 7) = -ctemp08;

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

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

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

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

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

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

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){
aoffset1 = aoffset;
aoffset2 = aoffset + lda;
aoffset += 2;
// aoffset += 2;

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

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



+ 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;
blasint *ipiv, iinfo, info;
BLASLONG jb, mn, blocking;
FLOAT *a, *offsetA, *offsetB;
FLOAT *a, *offsetA; //, *offsetB;
BLASLONG range_N[2];

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;

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

range_N[0] = offset + j;
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--) {

ajj_r = ONE;
ajj_i = ZERO;

#ifndef UNIT
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 + 1) = ajj_i;
#else
ajj_r = ONE;
ajj_i = ZERO;
#endif

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++) {

ajj_r = ONE;
ajj_i = ZERO;

#ifndef UNIT
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 + 1) = ajj_i;
#else
ajj_r = ONE;
ajj_i = ZERO;
#endif

ZTRMV (j,


Loading…
Cancel
Save