floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

trmm sve copy fucntions for single precision

This commit is contained in:
Bine Brank 2021-11-29 21:25:05 +01:00
parent 86ae89bf33
commit 0de36f7b5c
4 changed files with 66 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
kernel/arm64/trmm_lncopy_sve_v1.c
View File

@ -48,12 +48,17 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
BLASLONG i, js; BLASLONG i, js;
BLASLONG X; BLASLONG X;
svint64_t index = svindex_s64(0LL, lda);
FLOAT *ao;
js = 0; js = 0;
FLOAT *ao;
#ifdef DOUBLE
svint64_t index = svindex_s64(0LL, lda);
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64(js, n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else
svint32_t index = svindex_s32(0, lda);
svbool_t pn = svwhilelt_b32(js, n);
int n_active = svcntp_b32(svptrue_b32(), pn);
#endif
do do
{ {
X = posX; X = posX;
@ -68,7 +73,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
do do
{ {
if (X > posY) { if (X > posY) {
#ifdef DOUBLE
svfloat64_t aj_vec = svld1_gather_index(pn, ao, index); svfloat64_t aj_vec = svld1_gather_index(pn, ao, index);
#else
svfloat32_t aj_vec = svld1_gather_index(pn, ao, index);
#endif
svst1(pn, b, aj_vec); svst1(pn, b, aj_vec);
ao ++; ao ++;
b += n_active; b += n_active;
@ -113,9 +122,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64(js, n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else
pn = svwhilelt_b32(js, n);
n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn));
#endif
return 0; return 0;
} }

15
kernel/arm64/trmm_ltcopy_sve_v1.c
View File

@ -50,8 +50,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
FLOAT *ao; FLOAT *ao;
js = 0; js = 0;
#ifdef DOUBLE
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64(js, n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else
svbool_t pn = svwhilelt_b32(js, n);
int n_active = svcntp_b32(svptrue_b32(), pn);
#endif
do do
{ {
X = posX; X = posX;
@ -72,7 +77,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
i ++; i ++;
} else } else
if (X < posY) { if (X < posY) {
#ifdef DOUBLE
svfloat64_t aj_vec = svld1(pn, ao); svfloat64_t aj_vec = svld1(pn, ao);
#else
svfloat32_t aj_vec = svld1(pn, ao);
#endif
svst1(pn, b, aj_vec); svst1(pn, b, aj_vec);
ao += lda; ao += lda;
b += n_active; b += n_active;
@ -112,9 +121,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64(js, n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else
pn = svwhilelt_b32(js, n);
n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn));
#endif
return 0; return 0;

21
kernel/arm64/trmm_uncopy_sve_v1.c
View File

@ -48,12 +48,17 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
BLASLONG i, js; BLASLONG i, js;
BLASLONG X; BLASLONG X;
svint64_t index = svindex_s64(0LL, lda);
FLOAT *ao;
js = 0; js = 0;
FLOAT *ao;
#ifdef DOUBLE
svint64_t index = svindex_s64(0LL, lda);
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64(js, n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else
svint32_t index = svindex_s32(0, lda);
svbool_t pn = svwhilelt_b32(js, n);
int n_active = svcntp_b32(svptrue_b32(), pn);
#endif
do do
{ {
X = posX; X = posX;
@ -68,7 +73,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
do do
{ {
if (X < posY) { if (X < posY) {
#ifdef DOUBLE
svfloat64_t aj_vec = svld1_gather_index(pn, ao, index); svfloat64_t aj_vec = svld1_gather_index(pn, ao, index);
#else
svfloat32_t aj_vec = svld1_gather_index(pn, ao, index);
#endif
svst1(pn, b, aj_vec); svst1(pn, b, aj_vec);
ao ++; ao ++;
b += n_active; b += n_active;
@ -113,9 +122,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64(js, n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else
pn = svwhilelt_b32(js, n);
n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn));
#endif
return 0; return 0;
} }

15
kernel/arm64/trmm_utcopy_sve_v1.c
View File

@ -50,8 +50,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
FLOAT *ao; FLOAT *ao;
js = 0; js = 0;
#ifdef DOUBLE
svbool_t pn = svwhilelt_b64(js, n); svbool_t pn = svwhilelt_b64(js, n);
int n_active = svcntp_b64(svptrue_b64(), pn); int n_active = svcntp_b64(svptrue_b64(), pn);
#else
svbool_t pn = svwhilelt_b32(js, n);
int n_active = svcntp_b32(svptrue_b32(), pn);
#endif
do do
{ {
X = posX; X = posX;
@ -72,7 +77,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
i ++; i ++;
} else } else
if (X > posY) { if (X > posY) {
#ifdef DOUBLE
svfloat64_t aj_vec = svld1(pn, ao); svfloat64_t aj_vec = svld1(pn, ao);
#else
svfloat32_t aj_vec = svld1(pn, ao);
#endif
svst1(pn, b, aj_vec); svst1(pn, b, aj_vec);
ao += lda; ao += lda;
b += n_active; b += n_active;
@ -111,9 +120,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
posY += n_active; posY += n_active;
js += n_active; js += n_active;
#ifdef DOUBLE
pn = svwhilelt_b64(js, n); pn = svwhilelt_b64(js, n);
n_active = svcntp_b64(svptrue_b64(), pn); n_active = svcntp_b64(svptrue_b64(), pn);
} while (svptest_any(svptrue_b64(), pn)); } while (svptest_any(svptrue_b64(), pn));
#else
pn = svwhilelt_b32(js, n);
n_active = svcntp_b32(svptrue_b32(), pn);
} while (svptest_any(svptrue_b32(), pn));
#endif
return 0; return 0;
} }