Skip to content

Commit 7129a64

Browse files
martin-frbgmartin-frbg
authored
Merge pull request #4881 from martin-frbg/issue4805-2
Use fld.d/fst.d in PROLOGUE/EPILOGUE in LOONGSON3R5 GEMM
2 parents 3ee9e9d + 49080b6 commit 7129a64

File tree

4 files changed

+62
-62
lines changed

4 files changed

+62
-62
lines changed

‎kernel/loongarch64/cgemm_kernel_16x4_lasx.S‎

Lines changed: 21 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -196,17 +196,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
196196
SDARG $r25, $sp, 16
197197
SDARG $r26, $sp, 24
198198
SDARG $r27, $sp, 32
199-
ST $f23, $sp, 40
200-
ST $f24, $sp, 48
201-
ST $f25, $sp, 56
202-
ST $f26, $sp, 64
203-
ST $f27, $sp, 72
204-
ST $f28, $sp, 80
205-
ST $f29, $sp, 88
206-
ST $f30, $sp, 96
207-
ST $f31, $sp, 104
208-
ST ALPHA_R,$sp, 112
209-
ST ALPHA_I,$sp, 120
199+
fst.d $f23, $sp, 40
200+
fst.d $f24, $sp, 48
201+
fst.d $f25, $sp, 56
202+
fst.d $f26, $sp, 64
203+
fst.d $f27, $sp, 72
204+
fst.d $f28, $sp, 80
205+
fst.d $f29, $sp, 88
206+
fst.d $f30, $sp, 96
207+
fst.d $f31, $sp, 104
208+
fst.d ALPHA_R,$sp, 112
209+
fst.d ALPHA_I,$sp, 120
210210

211211
xvldrepl.w VALPHAR, $sp, 112
212212
xvldrepl.w VALPHAI, $sp, 120
@@ -3741,17 +3741,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37413741
LDARG $r25, $sp, 16
37423742
LDARG $r26, $sp, 24
37433743
LDARG $r27, $sp, 32
3744-
LD $f23, $sp, 40
3745-
LD $f24, $sp, 48
3746-
LD $f25, $sp, 56
3747-
LD $f26, $sp, 64
3748-
LD $f27, $sp, 72
3749-
LD $f28, $sp, 80
3750-
LD $f29, $sp, 88
3751-
LD $f30, $sp, 96
3752-
LD $f31, $sp, 104
3744+
fld.d $f23, $sp, 40
3745+
fld.d $f24, $sp, 48
3746+
fld.d $f25, $sp, 56
3747+
fld.d $f26, $sp, 64
3748+
fld.d $f27, $sp, 72
3749+
fld.d $f28, $sp, 80
3750+
fld.d $f29, $sp, 88
3751+
fld.d $f30, $sp, 96
3752+
fld.d $f31, $sp, 104
37533753

37543754
addi.d $sp, $sp, 128
37553755
jirl $r0, $r1, 0x0
37563756

3757-
EPILOGUE
3757+
EPILOGUE

‎kernel/loongarch64/dgemm_kernel_16x4.S‎

Lines changed: 19 additions & 19 deletions
Original file line numberDiff line numberDiff line change
@@ -1098,16 +1098,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
10981098
SDARG $r25, $sp, 16
10991099
SDARG $r26, $sp, 24
11001100
SDARG $r27, $sp, 32
1101-
ST $f23, $sp, 40
1102-
ST $f24, $sp, 48
1103-
ST $f25, $sp, 56
1104-
ST $f26, $sp, 64
1105-
ST $f27, $sp, 72
1106-
ST $f28, $sp, 80
1107-
ST $f29, $sp, 88
1108-
ST $f30, $sp, 96
1109-
ST $f31, $sp, 104
1110-
ST ALPHA, $sp, 112
1101+
fst.d $f23, $sp, 40
1102+
fst.d $f24, $sp, 48
1103+
fst.d $f25, $sp, 56
1104+
fst.d $f26, $sp, 64
1105+
fst.d $f27, $sp, 72
1106+
fst.d $f28, $sp, 80
1107+
fst.d $f29, $sp, 88
1108+
fst.d $f30, $sp, 96
1109+
fst.d $f31, $sp, 104
1110+
fst.d ALPHA, $sp, 112
11111111

11121112
#if defined (TRMMKERNEL) && !defined(LEFT)
11131113
sub.d OFF, ZERO, OFFSET
@@ -3504,15 +3504,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35043504
LDARG $r25, $sp, 16
35053505
LDARG $r26, $sp, 24
35063506
LDARG $r27, $sp, 32
3507-
LD $f23, $sp, 40
3508-
LD $f24, $sp, 48
3509-
LD $f25, $sp, 56
3510-
LD $f26, $sp, 64
3511-
LD $f27, $sp, 72
3512-
LD $f28, $sp, 80
3513-
LD $f29, $sp, 88
3514-
LD $f30, $sp, 96
3515-
LD $f31, $sp, 104
3507+
fld.d $f23, $sp, 40
3508+
fld.d $f24, $sp, 48
3509+
fld.d $f25, $sp, 56
3510+
fld.d $f26, $sp, 64
3511+
fld.d $f27, $sp, 72
3512+
fld.d $f28, $sp, 80
3513+
fld.d $f29, $sp, 88
3514+
fld.d $f30, $sp, 96
3515+
fld.d $f31, $sp, 104
35163516
addi.d $sp, $sp, 120
35173517

35183518
jirl $r0, $r1, 0x0

‎kernel/loongarch64/zgemm_kernel_8x4_lasx.S‎

Lines changed: 21 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -196,17 +196,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
196196
SDARG $r25, $sp, 16
197197
SDARG $r26, $sp, 24
198198
SDARG $r27, $sp, 32
199-
ST $f23, $sp, 40
200-
ST $f24, $sp, 48
201-
ST $f25, $sp, 56
202-
ST $f26, $sp, 64
203-
ST $f27, $sp, 72
204-
ST $f28, $sp, 80
205-
ST $f29, $sp, 88
206-
ST $f30, $sp, 96
207-
ST $f31, $sp, 104
208-
ST ALPHA_R,$sp, 112
209-
ST ALPHA_I,$sp, 120
199+
fst.d $f23, $sp, 40
200+
fst.d $f24, $sp, 48
201+
fst.d $f25, $sp, 56
202+
fst.d $f26, $sp, 64
203+
fst.d $f27, $sp, 72
204+
fst.d $f28, $sp, 80
205+
fst.d $f29, $sp, 88
206+
fst.d $f30, $sp, 96
207+
fst.d $f31, $sp, 104
208+
fst.d ALPHA_R,$sp, 112
209+
fst.d ALPHA_I,$sp, 120
210210

211211
xvldrepl.d VALPHAR, $sp, 112
212212
xvldrepl.d VALPHAI, $sp, 120
@@ -3529,17 +3529,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35293529
LDARG $r25, $sp, 16
35303530
LDARG $r26, $sp, 24
35313531
LDARG $r27, $sp, 32
3532-
LD $f23, $sp, 40
3533-
LD $f24, $sp, 48
3534-
LD $f25, $sp, 56
3535-
LD $f26, $sp, 64
3536-
LD $f27, $sp, 72
3537-
LD $f28, $sp, 80
3538-
LD $f29, $sp, 88
3539-
LD $f30, $sp, 96
3540-
LD $f31, $sp, 104
3532+
fld.d $f23, $sp, 40
3533+
fld.d $f24, $sp, 48
3534+
fld.d $f25, $sp, 56
3535+
fld.d $f26, $sp, 64
3536+
fld.d $f27, $sp, 72
3537+
fld.d $f28, $sp, 80
3538+
fld.d $f29, $sp, 88
3539+
fld.d $f30, $sp, 96
3540+
fld.d $f31, $sp, 104
35413541

35423542
addi.d $sp, $sp, 128
35433543
jirl $r0, $r1, 0x0
35443544

3545-
EPILOGUE
3545+
EPILOGUE

‎utest/test_potrs.c‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
3232
**********************************************************************************/
3333

3434
#include "openblas_utest.h"
35-
#pragma GCC optimize("no-gcse")
35+
3636
/*
3737
void BLASFUNC(cpotrf)(char*, BLASINT*, complex float*, BLASINT*, BLASINT*);
3838
void BLASFUNC(zpotrs_(char*, BLASINT*, BLASINT*, complex double*,

0 commit comments

Comments
 (0)