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

Add NaN tests

This commit is contained in:
Martin Kroeker 2023-08-14 17:45:35 +02:00 committed by GitHub
parent 95ce0b0c48
commit 43f5e4251b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 155 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

156
kernel/x86_64/iamax_sse.S
View File

@ -93,6 +93,10 @@
addq INCX, X addq INCX, X
decq M decq M
shufps $0, %xmm0, %xmm0 shufps $0, %xmm0, %xmm0
incq RET
comiss %xmm0, %xmm0
jp .L999
decq RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm0 andps %xmm15, %xmm0
#endif #endif
@ -254,6 +258,10 @@
decq M decq M
addq $SIZE, X addq $SIZE, X
incq RET
comiss %xmm1, %xmm1
jp .L998
decq RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
#endif #endif
@ -268,6 +276,14 @@
movss 0 * SIZE(X), %xmm1 movss 0 * SIZE(X), %xmm1
movss 1 * SIZE(X), %xmm2 movss 1 * SIZE(X), %xmm2
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
decq RET
decq RET
subq $2, M subq $2, M
addq $2 * SIZE, X addq $2 * SIZE, X
@ -332,6 +348,31 @@
movss 5 * SIZE(X), %xmm6 movss 5 * SIZE(X), %xmm6
movss 6 * SIZE(X), %xmm7 movss 6 * SIZE(X), %xmm7
movss 7 * SIZE(X), %xmm8 movss 7 * SIZE(X), %xmm8
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
incq RET
comiss %xmm3, %xmm3
jp .L998
incq RET
comiss %xmm4, %xmm4
jp .L998
incq RET
comiss %xmm5, %xmm5
jp .L998
incq RET
comiss %xmm6, %xmm6
jp .L998
incq RET
comiss %xmm7, %xmm7
jp .L998
incq RET
comiss %xmm8, %xmm8
jp .L998
subq $8, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -378,6 +419,19 @@
movss 1 * SIZE(X), %xmm2 movss 1 * SIZE(X), %xmm2
movss 2 * SIZE(X), %xmm3 movss 2 * SIZE(X), %xmm3
movss 3 * SIZE(X), %xmm4 movss 3 * SIZE(X), %xmm4
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
incq RET
comiss %xmm3, %xmm3
jp .L998
incq RET
comiss %xmm4, %xmm4
jp .L998
subq $4, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -405,6 +459,13 @@
movss 0 * SIZE(X), %xmm1 movss 0 * SIZE(X), %xmm1
movss 1 * SIZE(X), %xmm2 movss 1 * SIZE(X), %xmm2
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
subq $2, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -593,6 +654,31 @@
movss 5 * SIZE(X), %xmm6 movss 5 * SIZE(X), %xmm6
movss 6 * SIZE(X), %xmm7 movss 6 * SIZE(X), %xmm7
movss 7 * SIZE(X), %xmm8 movss 7 * SIZE(X), %xmm8
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
incq RET
comiss %xmm3, %xmm3
jp .L998
incq RET
comiss %xmm4, %xmm4
jp .L998
incq RET
comiss %xmm5, %xmm5
jp .L998
incq RET
comiss %xmm6, %xmm6
jp .L998
incq RET
comiss %xmm7, %xmm7
jp .L998
incq RET
comiss %xmm8, %xmm8
jp .L998
subq $8, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -639,6 +725,19 @@
movss 1 * SIZE(X), %xmm2 movss 1 * SIZE(X), %xmm2
movss 2 * SIZE(X), %xmm3 movss 2 * SIZE(X), %xmm3
movss 3 * SIZE(X), %xmm4 movss 3 * SIZE(X), %xmm4
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
incq RET
comiss %xmm3, %xmm3
jp .L998
incq RET
comiss %xmm4, %xmm4
jp .L998
subq $4, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -666,6 +765,13 @@
movss 0 * SIZE(X), %xmm1 movss 0 * SIZE(X), %xmm1
movss 1 * SIZE(X), %xmm2 movss 1 * SIZE(X), %xmm2
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
subq $2, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -885,6 +991,31 @@
movss 0 * SIZE(X), %xmm2 movss 0 * SIZE(X), %xmm2
subq INCX, X subq INCX, X
movss 0 * SIZE(X), %xmm1 movss 0 * SIZE(X), %xmm1
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
incq RET
comiss %xmm3, %xmm3
jp .L998
incq RET
comiss %xmm4, %xmm4
jp .L998
incq RET
comiss %xmm5, %xmm5
jp .L998
incq RET
comiss %xmm6, %xmm6
jp .L998
incq RET
comiss %xmm7, %xmm7
jp .L998
incq RET
comiss %xmm8, %xmm8
jp .L998
subq $8, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -932,7 +1063,19 @@
addq INCX, X addq INCX, X
movss 0 * SIZE(X), %xmm4 movss 0 * SIZE(X), %xmm4
addq INCX, X addq INCX, X
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
incq RET
comiss %xmm3, %xmm3
jp .L998
incq RET
comiss %xmm4, %xmm4
jp .L998
subq $4, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -961,6 +1104,13 @@
addq INCX, X addq INCX, X
movss 0 * SIZE(X), %xmm2 movss 0 * SIZE(X), %xmm2
addq INCX, X addq INCX, X
incq RET
comiss %xmm1, %xmm1
jp .L998
incq RET
comiss %xmm2, %xmm2
jp .L998
subq $2, RET
#ifdef USE_ABS #ifdef USE_ABS
andps %xmm15, %xmm1 andps %xmm15, %xmm1
andps %xmm15, %xmm2 andps %xmm15, %xmm2
@ -982,5 +1132,9 @@
ret ret
.L998:
// incq RET
jmp .L999
EPILOGUE EPILOGUE