32c613e8ec
* Update Dilithium * Alternative montgomery reduce to avoid i386 functest errors * Explicit casts for msvc * More casts; bump upstream version; fix metadata * another cast
26 lines
615 B
PHP
26 lines
615 B
PHP
.macro shuffle8 r0,r1,r2,r3
|
|
vperm2i128 $0x20,%ymm\r1,%ymm\r0,%ymm\r2
|
|
vperm2i128 $0x31,%ymm\r1,%ymm\r0,%ymm\r3
|
|
.endm
|
|
|
|
.macro shuffle4 r0,r1,r2,r3
|
|
vpunpcklqdq %ymm\r1,%ymm\r0,%ymm\r2
|
|
vpunpckhqdq %ymm\r1,%ymm\r0,%ymm\r3
|
|
.endm
|
|
|
|
.macro shuffle2 r0,r1,r2,r3
|
|
#vpsllq $32,%ymm\r1,%ymm\r2
|
|
vmovsldup %ymm\r1,%ymm\r2
|
|
vpblendd $0xAA,%ymm\r2,%ymm\r0,%ymm\r2
|
|
vpsrlq $32,%ymm\r0,%ymm\r0
|
|
#vmovshdup %ymm\r0,%ymm\r0
|
|
vpblendd $0xAA,%ymm\r1,%ymm\r0,%ymm\r3
|
|
.endm
|
|
|
|
.macro shuffle1 r0,r1,r2,r3
|
|
vpslld $16,%ymm\r1,%ymm\r2
|
|
vpblendw $0xAA,%ymm\r2,%ymm\r0,%ymm\r2
|
|
vpsrld $16,%ymm\r0,%ymm\r0
|
|
vpblendw $0xAA,%ymm\r1,%ymm\r0,%ymm\r3
|
|
.endm
|