TrueCloudLab/tzhash
15
0
Fork 6
Code Issues 3 Pull requests Releases 2 Activity

Optimize AVX implementation

Browse source 
1. Do the same mask trick as with AVX2.
2. Get rid of load, generate constant on the fly.

```
name                    old time/op    new time/op    delta
Sum/AVXInline_digest-8    2.26ms ± 4%    2.17ms ± 5%  -4.05%  (p=0.000 n=19+17)

name                    old speed      new speed      delta
Sum/AVXInline_digest-8  44.3MB/s ± 4%  46.2MB/s ± 5%  +4.25%  (p=0.000 n=19+17)
```

Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
This commit is contained in:
Evgenii Stratonikov 2021-12-29 15:56:02 +03:00 committed by Alex Vanin
parent d4cb61e470
commit 921f8b0579
1 changed files with 22 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

35
tz/avx_amd64.s
View file

@ -6,22 +6,15 @@
VPSLLQ $1, FROM, TO \
VPALIGNR $8, TO, FROM, R2 \
VPSRLQ $63, R2, R2 \
VMOVDQU ·x127x63(SB), R3 \
VANDPD TO, R3, R3 \
VANDPD TO, X14, R3 \
VPUNPCKHQDQ R3, R3, R3 \
VXORPD R2, TO, TO \
VXORPD R3, TO, TO
#define mask(bit, src, tmp, to1, to2) \
MOVQ src, tmp \
SHRQ bit, tmp \
ANDQ $1, tmp \
NEGQ tmp \
MOVQ tmp, to1 \
VSHUFPS $0, to1, to1, to2
// VPBROADCASTB to1, to2
// Can't use VPBROADCASTB because it is AVX2 instruction
//https://software.intel.com/en-us/forums/intel-isa-extensions/topic/301461
#define mask(bit, tmp, to) \
VPSRLW bit, X10, tmp \
VPAND X12, tmp, to \ // to = 0x000<bit>000<bit>...
VPSUBW to, X13, to // to = 0xFFFF.. or 0x0000 depending on bit
#define mulBit(bit) \
VMOVDQU X0, X8 \
@ -30,7 +23,7 @@
VXORPD X1, X5, X0 \
mul2(X2, X5, X6, X7) \
VXORPD X3, X5, X2 \
mask(bit, CX, DX, X6, X5) \
mask(bit, X6, X5) \
VANDPD X0, X5, X1 \
VXORPD X8, X1, X1 \
VANDPD X2, X5, X3 \
@ -49,6 +42,11 @@ TEXT ·mulBitRight(SB),NOSPLIT,$0
MOVQ c11+24(FP), DX
VMOVDQU (DX), X3
VPXOR X13, X13, X13 // Y13 = 0x0000...
VPCMPEQB X14, X14, X14 // Y14 = 0xFFFF...
VPSUBQ X14, X13, X13
VPSLLQ $63, X13, X14
mul2(X0, X5, X6, X7) // c00 *= 2
VXORPD X5, X1, X0 // c00 += c01
mul2(X2, X5, X6, X7) // c10 *= 2
@ -77,8 +75,19 @@ TEXT ·mulByteRight(SB),NOSPLIT,$0
VMOVDQU (CX), X2
MOVQ c11+24(FP), DX
VMOVDQU (DX), X3
MOVQ $0, CX
MOVB b+32(FP), CX
VPXOR X13, X13, X13 // X13 = 0x0000...
VPCMPEQB X14, X14, X14 // X14 = 0xFFFF...
VPSUBQ X14, X13, X10
VPSUBW X14, X13, X12 // X12 = 0x00010001... (packed words of 1)
VPSLLQ $63, X10, X14 // X14 = 0x10000000... (packed quad-words with HSB set)
MOVQ CX, X10
VPSHUFLW $0, X10, X11
VPSHUFD $0, X11, X10
mulBit($7)
mulBit($6)
mulBit($5)