OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Add NaN tests

pull/4191/head
Martin Kroeker GitHub 2 years ago
parent
562ef5fdca
commit
95ce0b0c48
No known key found for this signature in database GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 162 additions and 5 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +162
    -5
      kernel/x86_64/iamax_sse2.S

+ 162
- 5
kernel/x86_64/iamax_sse2.S View File

@@ -79,6 +79,8 @@
movsd (X), %xmm0 movsd (X), %xmm0
addq INCX, X addq INCX, X
decq M decq M
comisd %xmm0, %xmm0
jp .L987
#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm0 andpd %xmm15, %xmm0
#endif #endif
@@ -269,6 +271,11 @@
je .L21 je .L21


movsd 0 * SIZE(X), %xmm1 movsd 0 * SIZE(X), %xmm1
incq RET
comisd %xmm1, %xmm1
jp .L987
decq RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
#endif #endif
@@ -342,6 +349,32 @@
movsd 5 * SIZE(X), %xmm6 movsd 5 * SIZE(X), %xmm6
movsd 6 * SIZE(X), %xmm7 movsd 6 * SIZE(X), %xmm7
movsd 7 * SIZE(X), %xmm8 movsd 7 * SIZE(X), %xmm8
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
incq RET
comisd %xmm3, %xmm3
jp .L987
incq RET
comisd %xmm4, %xmm4
jp .L987
incq RET
comisd %xmm5, %xmm5
jp .L987
incq RET
comisd %xmm6, %xmm6
jp .L987
incq RET
comisd %xmm7, %xmm7
jp .L987
incq RET
comisd %xmm8, %xmm8
jp .L987
subq $8, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -374,9 +407,9 @@
comisd %xmm0, %xmm6 comisd %xmm0, %xmm6
je .L999 je .L999
incq RET incq RET
comisd %xmm0, %xmm7
je .L999
incq RET
comisd %xmm0, %xmm7
je .L999
incq RET
jmp .L999 jmp .L999
ALIGN_3 ALIGN_3


@@ -388,6 +421,21 @@
movsd 1 * SIZE(X), %xmm2 movsd 1 * SIZE(X), %xmm2
movsd 2 * SIZE(X), %xmm3 movsd 2 * SIZE(X), %xmm3
movsd 3 * SIZE(X), %xmm4 movsd 3 * SIZE(X), %xmm4

incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
incq RET
comisd %xmm3, %xmm3
jp .L987
incq RET
comisd %xmm4, %xmm4
jp .L987
subq $4, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -415,6 +463,14 @@


movsd 0 * SIZE(X), %xmm1 movsd 0 * SIZE(X), %xmm1
movsd 1 * SIZE(X), %xmm2 movsd 1 * SIZE(X), %xmm2
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
subq $2, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -670,6 +726,32 @@
movsd 5 * SIZE(X), %xmm6 movsd 5 * SIZE(X), %xmm6
movsd 6 * SIZE(X), %xmm7 movsd 6 * SIZE(X), %xmm7
movsd 7 * SIZE(X), %xmm8 movsd 7 * SIZE(X), %xmm8
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
incq RET
comisd %xmm3, %xmm3
jp .L987
incq RET
comisd %xmm4, %xmm4
jp .L987
incq RET
comisd %xmm5, %xmm5
jp .L987
incq RET
comisd %xmm6, %xmm6
jp .L987
incq RET
comisd %xmm7, %xmm7
jp .L987
incq RET
comisd %xmm8, %xmm8
jp .L987
subq $8, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -716,6 +798,20 @@
movsd 1 * SIZE(X), %xmm2 movsd 1 * SIZE(X), %xmm2
movsd 2 * SIZE(X), %xmm3 movsd 2 * SIZE(X), %xmm3
movsd 3 * SIZE(X), %xmm4 movsd 3 * SIZE(X), %xmm4
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
incq RET
comisd %xmm3, %xmm3
jp .L987
incq RET
comisd %xmm4, %xmm4
jp .L987
subq $4, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -743,11 +839,21 @@


movsd 0 * SIZE(X), %xmm1 movsd 0 * SIZE(X), %xmm1
movsd 1 * SIZE(X), %xmm2 movsd 1 * SIZE(X), %xmm2
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
subq $2, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
#endif #endif
addq $2 * SIZE, X addq $2 * SIZE, X
comisd %xmm0, %xmm0
jp .L987
incq RET incq RET
comisd %xmm0, %xmm1 comisd %xmm0, %xmm1
je .L999 je .L999
@@ -962,6 +1068,7 @@
ALIGN_4 ALIGN_4


.L92: .L92:

movsd 0 * SIZE(X), %xmm1 movsd 0 * SIZE(X), %xmm1
addq INCX, X addq INCX, X
movhpd 0 * SIZE(X), %xmm1 movhpd 0 * SIZE(X), %xmm1
@@ -1033,6 +1140,32 @@
movsd 0 * SIZE(X), %xmm2 movsd 0 * SIZE(X), %xmm2
subq INCX, X subq INCX, X
movsd 0 * SIZE(X), %xmm1 movsd 0 * SIZE(X), %xmm1
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
incq RET
comisd %xmm3, %xmm3
jp .L987
incq RET
comisd %xmm4, %xmm4
jp .L987
incq RET
comisd %xmm5, %xmm5
jp .L987
incq RET
comisd %xmm6, %xmm6
jp .L987
incq RET
comisd %xmm7, %xmm7
jp .L987
incq RET
comisd %xmm8, %xmm8
jp .L987
subq $8, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -1083,6 +1216,20 @@
addq INCX, X addq INCX, X
movsd 0 * SIZE(X), %xmm4 movsd 0 * SIZE(X), %xmm4
addq INCX, X addq INCX, X
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
incq RET
comisd %xmm3, %xmm3
jp .L987
incq RET
comisd %xmm4, %xmm4
jp .L987
subq $4, RET

#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -1111,6 +1258,14 @@
addq INCX, X addq INCX, X
movsd 0 * SIZE(X), %xmm2 movsd 0 * SIZE(X), %xmm2
addq INCX, X addq INCX, X
incq RET
comisd %xmm1, %xmm1
jp .L987
incq RET
comisd %xmm2, %xmm2
jp .L987
decq RET
decq RET
#ifdef USE_ABS #ifdef USE_ABS
andpd %xmm15, %xmm1 andpd %xmm15, %xmm1
andpd %xmm15, %xmm2 andpd %xmm15, %xmm2
@@ -1122,7 +1277,6 @@
comisd %xmm0, %xmm2 comisd %xmm0, %xmm2
je .L999 je .L999
ALIGN_3 ALIGN_3

.L98: .L98:
incq RET incq RET
ALIGN_3 ALIGN_3
@@ -1132,5 +1286,8 @@


ret ret


EPILOGUE
.L987:
incq RET //count from xmm0
jmp .L999


EPILOGUE

Write Preview
Loading…
Cancel
Save