Skip to content

Commit 55a42d7

Browse files
pablodelaramdcornu
pablodelara
authored and
mdcornu
committed
raid: add AVX512+GFNI implementation for P+Q gen
Signed-off-by: Pablo de Lara <[email protected]>
1 parent 359e2ac commit 55a42d7

File tree

4 files changed

+245
-1
lines changed

4 files changed

+245
-1
lines changed

‎Makefile.nmake‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -128,6 +128,7 @@ objs = \
128128
bin\pq_gen_avx2.obj \
129129
bin\xor_gen_avx512.obj \
130130
bin\pq_gen_avx512.obj \
131+
bin\pq_gen_avx512_gfni.obj \
131132
bin\raid_multibinary.obj \
132133
bin\crc16_t10dif_01.obj \
133134
bin\crc16_t10dif_by4.obj \

‎raid/Makefile.am‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -46,6 +46,7 @@ lsrc_x86_64 += \
4646
raid/pq_gen_avx2.asm \
4747
raid/xor_gen_avx512.asm \
4848
raid/pq_gen_avx512.asm \
49+
raid/pq_gen_avx512_gfni.asm \
4950
raid/raid_multibinary.asm
5051

5152
lsrc_x86_32 += \

‎raid/pq_gen_avx512_gfni.asm‎

Lines changed: 238 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,238 @@
1+
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
2+
; Copyright(c) 2024 Intel Corporation All rights reserved.
3+
;
4+
; Redistribution and use in source and binary forms, with or without
5+
; modification, are permitted provided that the following conditions
6+
; are met:
7+
; * Redistributions of source code must retain the above copyright
8+
; notice, this list of conditions and the following disclaimer.
9+
; * Redistributions in binary form must reproduce the above copyright
10+
; notice, this list of conditions and the following disclaimer in
11+
; the documentation and/or other materials provided with the
12+
; distribution.
13+
; * Neither the name of Intel Corporation nor the names of its
14+
; contributors may be used to endorse or promote products derived
15+
; from this software without specific prior written permission.
16+
;
17+
; THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
18+
; "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
19+
; LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
20+
; A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
21+
; OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
22+
; SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
23+
; LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24+
; DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25+
; THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26+
; (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27+
; OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28+
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
29+
30+
;;; Optimized pq of N source vectors using AVX512+GFNI
31+
;;; int pq_gen_avx512_gfni(int vects, int len, void **array)
32+
33+
;;; Generates P+Q parity vector from N (vects-2) sources in array of pointers
34+
;;; (**array). Last two pointers are the P and Q destinations respectively.
35+
;;; Vectors must be aligned to 64 bytes if NO_NT_LDST is not defined.
36+
;;; Length must be 32 byte multiple.
37+
38+
%include "reg_sizes.asm"
39+
40+
%ifdef HAVE_AS_KNOWS_AVX512
41+
42+
%ifidn __OUTPUT_FORMAT__, elf64
43+
%define arg0 rdi
44+
%define arg1 rsi
45+
%define arg2 rdx
46+
%define arg3 rcx
47+
%define arg4 r8
48+
%define arg5 r9
49+
%define tmp r11
50+
%define tmp3 arg4
51+
%define return rax
52+
%define func(x) x: endbranch
53+
%define FUNC_SAVE
54+
%define FUNC_RESTORE
55+
%endif
56+
57+
%ifidn __OUTPUT_FORMAT__, win64
58+
%define arg0 rcx
59+
%define arg1 rdx
60+
%define arg2 r8
61+
%define arg3 r9
62+
%define tmp r11
63+
%define tmp3 r10
64+
%define return rax
65+
%define stack_size 1*16 + 8 ; must be an odd multiple of 8
66+
%define func(x) proc_frame x
67+
%macro FUNC_SAVE 0
68+
alloc_stack stack_size
69+
vmovdqa [rsp + 0*16], xmm6
70+
end_prolog
71+
%endmacro
72+
73+
%macro FUNC_RESTORE 0
74+
vmovdqa xmm6, [rsp + 0*16]
75+
add rsp, stack_size
76+
%endmacro
77+
%endif
78+
79+
%define vec arg0
80+
%define len arg1
81+
%define ptr arg3
82+
%define pos rax
83+
84+
%define xp1 zmm0
85+
%define xq1 zmm1
86+
%define xs1 zmm2
87+
88+
%define xp2 zmm3
89+
%define xq2 zmm4
90+
%define xs2 zmm5
91+
92+
%define gfmatrix zmm6
93+
94+
%define xp1y ymm0
95+
%define xq1y ymm1
96+
%define xs1y ymm2
97+
98+
%define gfmatrixy ymm6
99+
100+
%define NO_NT_LDST
101+
;;; Use Non-temporal load/stor
102+
%ifdef NO_NT_LDST
103+
%define XLDR vmovdqu8 ;u8
104+
%define XSTR vmovdqu8
105+
%else
106+
%define XLDR vmovntdqa
107+
%define XSTR vmovntdq
108+
%endif
109+
110+
; Matrix with 0x11d as first column
111+
; and identity matrix shited by 1 (as we are multiplying data by 2, mod 0x11d)
112+
; 0 1 0 0 0 0 0 0
113+
; 0 0 1 0 0 0 0 0
114+
; 0 0 0 1 0 0 0 0
115+
; 0 0 0 0 1 0 0 0
116+
; 1 0 0 0 0 1 0 0
117+
; 1 0 0 0 0 0 1 0
118+
; 0 0 0 0 0 0 0 1
119+
; 1 0 0 0 0 0 0 0
120+
default rel
121+
align 64
122+
gf_matrix:
123+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
124+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
125+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
126+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
127+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
128+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
129+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
130+
db 0x40, 0x20, 0x10, 0x88, 0x84, 0x82, 0x01, 0x80
131+
132+
133+
[bits 64]
134+
section .text
135+
136+
align 16
137+
mk_global pq_gen_avx512_gfni, function
138+
func(pq_gen_avx512_gfni)
139+
FUNC_SAVE
140+
sub vec, 3 ;Keep as offset to last source
141+
jng return_fail ;Must have at least 2 sources
142+
cmp len, 0
143+
je return_pass
144+
test len, (32-1) ;Check alignment of length
145+
jnz return_fail
146+
147+
vmovdqa64 gfmatrix, [rel gf_matrix]
148+
149+
xor pos, pos
150+
cmp len, 128
151+
jl loop32
152+
153+
len_aligned_32bytes:
154+
sub len, 2*64 ;Len points to last block
155+
156+
loop128:
157+
mov ptr, [arg2+vec*8] ;Fetch last source pointer
158+
mov tmp, vec ;Set tmp to point back to last vector
159+
XLDR xs1, [ptr+pos] ;Preload last vector (source)
160+
XLDR xs2, [ptr+pos+64] ;Preload last vector (source)
161+
vpxorq xp1, xp1, xp1 ;p1 = 0
162+
vpxorq xp2, xp2, xp2 ;p2 = 0
163+
vpxorq xq1, xq1, xq1 ;q1 = 0
164+
vpxorq xq2, xq2, xq2 ;q2 = 0
165+
166+
next_vect:
167+
sub tmp, 1 ;Inner loop for each source vector
168+
mov ptr, [arg2+tmp*8] ; get pointer to next vect
169+
vpxorq xq1, xq1, xs1 ; q1 ^= s1
170+
vpxorq xq2, xq2, xs2 ; q2 ^= s2
171+
vpxorq xp1, xp1, xs1 ; p1 ^= s1
172+
vpxorq xp2, xp2, xs2 ; p2 ^= s2
173+
XLDR xs1, [ptr+pos] ; Get next vector (source data1)
174+
XLDR xs2, [ptr+pos+64] ; Get next vector (source data2)
175+
vgf2p8affineqb xq1, xq1, gfmatrix, 0x00
176+
vgf2p8affineqb xq2, xq2, gfmatrix, 0x00
177+
jg next_vect ; Loop for each vect except 0
178+
179+
mov ptr, [arg2+8+vec*8] ;Get address of P parity vector
180+
mov tmp, [arg2+(2*8)+vec*8] ;Get address of Q parity vector
181+
vpxorq xp1, xp1, xs1 ;p1 ^= s1[0] - last source is already loaded
182+
vpxorq xq1, xq1, xs1 ;q1 ^= 1 * s1[0]
183+
vpxorq xp2, xp2, xs2 ;p2 ^= s2[0]
184+
vpxorq xq2, xq2, xs2 ;q2 ^= 1 * s2[0]
185+
XSTR [ptr+pos], xp1 ;Write parity P1 vector
186+
XSTR [ptr+pos+64], xp2 ;Write parity P2 vector
187+
XSTR [tmp+pos], xq1 ;Write parity Q1 vector
188+
XSTR [tmp+pos+64], xq2 ;Write parity Q2 vector
189+
add pos, 2*64
190+
cmp pos, len
191+
jle loop128
192+
193+
;; ------------------------------
194+
;; Do last 32 or 64 Bytes remaining
195+
add len, 2*64
196+
cmp pos, len
197+
je return_pass
198+
199+
loop32:
200+
mov ptr, [arg2+vec*8] ;Fetch last source pointer
201+
mov tmp, vec ;Set tmp to point back to last vector
202+
XLDR xs1y, [ptr+pos] ;Preload last vector (source)
203+
vpxorq xp1y, xp1y, xp1y ;p = 0
204+
vpxorq xq1y, xq1y, xq1y ;q = 0
205+
206+
next_vect32:
207+
sub tmp, 1 ;Inner loop for each source vector
208+
mov ptr, [arg2+tmp*8] ; get pointer to next vect
209+
vpxorq xq1y, xq1y, xs1y ; q1 ^= s1
210+
vgf2p8affineqb xq1y, xq1y, gfmatrixy, 0x00
211+
vpxorq xp1y, xp1y, xs1y ; p ^= s
212+
XLDR xs1y, [ptr+pos] ; Get next vector (source data)
213+
jg next_vect32 ; Loop for each vect except 0
214+
215+
mov ptr, [arg2+8+vec*8] ;Get address of P parity vector
216+
mov tmp, [arg2+(2*8)+vec*8] ;Get address of Q parity vector
217+
vpxorq xp1y, xp1y, xs1y ;p ^= s[0] - last source is already loaded
218+
vpxorq xq1y, xq1y, xs1y ;q ^= 1 * s[0]
219+
XSTR [ptr+pos], xp1y ;Write parity P vector
220+
XSTR [tmp+pos], xq1y ;Write parity Q vector
221+
add pos, 32
222+
cmp pos, len
223+
jl loop32
224+
225+
226+
return_pass:
227+
mov return, 0
228+
FUNC_RESTORE
229+
ret
230+
231+
return_fail:
232+
mov return, 1
233+
FUNC_RESTORE
234+
ret
235+
236+
endproc_frame
237+
238+
%endif ; ifdef HAVE_AS_KNOWS_AVX512

‎raid/raid_multibinary.asm‎

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -53,12 +53,16 @@ extern xor_check_sse
5353
extern pq_gen_avx512
5454
%endif
5555

56+
%if (AS_FEATURE_LEVEL) >= 10
57+
extern pq_gen_avx512_gfni
58+
%endif
59+
5660
mbin_interface xor_gen
5761
mbin_interface pq_gen
5862

5963

6064
mbin_dispatch_init6 xor_gen, xor_gen_base, xor_gen_sse, xor_gen_avx, xor_gen_avx, xor_gen_avx512
61-
mbin_dispatch_init6 pq_gen, pq_gen_base, pq_gen_sse, pq_gen_avx, pq_gen_avx2, pq_gen_avx512
65+
mbin_dispatch_init8 pq_gen, pq_gen_base, pq_gen_sse, pq_gen_avx, pq_gen_avx2, pq_gen_avx512, pq_gen_avx2, pq_gen_avx512_gfni
6266

6367
section .data
6468

0 commit comments

Comments
 (0)