Skip to content

Commit 0de36f7

Browse files
committed
trmm sve copy fucntions for single precision
1 parent 86ae89b commit 0de36f7

4 files changed

Lines changed: 66 additions & 6 deletions

File tree

kernel/arm64/trmm_lncopy_sve_v1.c

Lines changed: 18 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -48,12 +48,17 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
4848
BLASLONG i, js;
4949
BLASLONG X;
5050

51-
svint64_t index = svindex_s64(0LL, lda);
52-
53-
FLOAT *ao;
5451
js = 0;
52+
FLOAT *ao;
53+
#ifdef DOUBLE
54+
svint64_t index = svindex_s64(0LL, lda);
5555
svbool_t pn = svwhilelt_b64(js, n);
5656
int n_active = svcntp_b64(svptrue_b64(), pn);
57+
#else
58+
svint32_t index = svindex_s32(0, lda);
59+
svbool_t pn = svwhilelt_b32(js, n);
60+
int n_active = svcntp_b32(svptrue_b32(), pn);
61+
#endif
5762
do
5863
{
5964
X = posX;
@@ -68,7 +73,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
6873
do
6974
{
7075
if (X > posY) {
76+
#ifdef DOUBLE
7177
svfloat64_t aj_vec = svld1_gather_index(pn, ao, index);
78+
#else
79+
svfloat32_t aj_vec = svld1_gather_index(pn, ao, index);
80+
#endif
7281
svst1(pn, b, aj_vec);
7382
ao ++;
7483
b += n_active;
@@ -113,9 +122,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
113122

114123
posY += n_active;
115124
js += n_active;
125+
#ifdef DOUBLE
116126
pn = svwhilelt_b64(js, n);
117127
n_active = svcntp_b64(svptrue_b64(), pn);
118128
} while (svptest_any(svptrue_b64(), pn));
129+
#else
130+
pn = svwhilelt_b32(js, n);
131+
n_active = svcntp_b32(svptrue_b32(), pn);
132+
} while (svptest_any(svptrue_b32(), pn));
133+
#endif
119134

120135
return 0;
121136
}

kernel/arm64/trmm_ltcopy_sve_v1.c

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -50,8 +50,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
5050

5151
FLOAT *ao;
5252
js = 0;
53+
#ifdef DOUBLE
5354
svbool_t pn = svwhilelt_b64(js, n);
5455
int n_active = svcntp_b64(svptrue_b64(), pn);
56+
#else
57+
svbool_t pn = svwhilelt_b32(js, n);
58+
int n_active = svcntp_b32(svptrue_b32(), pn);
59+
#endif
5560
do
5661
{
5762
X = posX;
@@ -72,7 +77,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
7277
i ++;
7378
} else
7479
if (X < posY) {
80+
#ifdef DOUBLE
7581
svfloat64_t aj_vec = svld1(pn, ao);
82+
#else
83+
svfloat32_t aj_vec = svld1(pn, ao);
84+
#endif
7685
svst1(pn, b, aj_vec);
7786
ao += lda;
7887
b += n_active;
@@ -112,9 +121,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
112121

113122
posY += n_active;
114123
js += n_active;
124+
#ifdef DOUBLE
115125
pn = svwhilelt_b64(js, n);
116126
n_active = svcntp_b64(svptrue_b64(), pn);
117127
} while (svptest_any(svptrue_b64(), pn));
128+
#else
129+
pn = svwhilelt_b32(js, n);
130+
n_active = svcntp_b32(svptrue_b32(), pn);
131+
} while (svptest_any(svptrue_b32(), pn));
132+
#endif
118133

119134

120135
return 0;

kernel/arm64/trmm_uncopy_sve_v1.c

Lines changed: 18 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -48,12 +48,17 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
4848
BLASLONG i, js;
4949
BLASLONG X;
5050

51-
svint64_t index = svindex_s64(0LL, lda);
52-
53-
FLOAT *ao;
5451
js = 0;
52+
FLOAT *ao;
53+
#ifdef DOUBLE
54+
svint64_t index = svindex_s64(0LL, lda);
5555
svbool_t pn = svwhilelt_b64(js, n);
5656
int n_active = svcntp_b64(svptrue_b64(), pn);
57+
#else
58+
svint32_t index = svindex_s32(0, lda);
59+
svbool_t pn = svwhilelt_b32(js, n);
60+
int n_active = svcntp_b32(svptrue_b32(), pn);
61+
#endif
5762
do
5863
{
5964
X = posX;
@@ -68,7 +73,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
6873
do
6974
{
7075
if (X < posY) {
76+
#ifdef DOUBLE
7177
svfloat64_t aj_vec = svld1_gather_index(pn, ao, index);
78+
#else
79+
svfloat32_t aj_vec = svld1_gather_index(pn, ao, index);
80+
#endif
7281
svst1(pn, b, aj_vec);
7382
ao ++;
7483
b += n_active;
@@ -113,9 +122,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
113122

114123
posY += n_active;
115124
js += n_active;
125+
#ifdef DOUBLE
116126
pn = svwhilelt_b64(js, n);
117127
n_active = svcntp_b64(svptrue_b64(), pn);
118128
} while (svptest_any(svptrue_b64(), pn));
129+
#else
130+
pn = svwhilelt_b32(js, n);
131+
n_active = svcntp_b32(svptrue_b32(), pn);
132+
} while (svptest_any(svptrue_b32(), pn));
133+
#endif
119134

120135
return 0;
121136
}

kernel/arm64/trmm_utcopy_sve_v1.c

Lines changed: 15 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -50,8 +50,13 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
5050

5151
FLOAT *ao;
5252
js = 0;
53+
#ifdef DOUBLE
5354
svbool_t pn = svwhilelt_b64(js, n);
5455
int n_active = svcntp_b64(svptrue_b64(), pn);
56+
#else
57+
svbool_t pn = svwhilelt_b32(js, n);
58+
int n_active = svcntp_b32(svptrue_b32(), pn);
59+
#endif
5560
do
5661
{
5762
X = posX;
@@ -72,7 +77,11 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
7277
i ++;
7378
} else
7479
if (X > posY) {
80+
#ifdef DOUBLE
7581
svfloat64_t aj_vec = svld1(pn, ao);
82+
#else
83+
svfloat32_t aj_vec = svld1(pn, ao);
84+
#endif
7685
svst1(pn, b, aj_vec);
7786
ao += lda;
7887
b += n_active;
@@ -111,9 +120,15 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
111120

112121
posY += n_active;
113122
js += n_active;
123+
#ifdef DOUBLE
114124
pn = svwhilelt_b64(js, n);
115125
n_active = svcntp_b64(svptrue_b64(), pn);
116126
} while (svptest_any(svptrue_b64(), pn));
127+
#else
128+
pn = svwhilelt_b32(js, n);
129+
n_active = svcntp_b32(svptrue_b32(), pn);
130+
} while (svptest_any(svptrue_b32(), pn));
131+
#endif
117132

118133
return 0;
119134
}

0 commit comments

Comments
 (0)