Skip to content

Commit a53a20e

Browse files
mdcornupablodelara
mdcornu
authored and
pablodelara
committed
erasure_code: add AVX2 5vect mad with GFNI implementation
Signed-off-by: Marcel Cornu <[email protected]>
1 parent 47ed284 commit a53a20e

File tree

4 files changed

+277
-5
lines changed

4 files changed

+277
-5
lines changed

‎Makefile.nmake‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -89,6 +89,7 @@ objs = \
8989
bin\gf_2vect_mad_avx2_gfni.obj \
9090
bin\gf_3vect_mad_avx2_gfni.obj \
9191
bin\gf_4vect_mad_avx2_gfni.obj \
92+
bin\gf_5vect_mad_avx2_gfni.obj \
9293
bin\gf_vect_dot_prod_avx512.obj \
9394
bin\gf_2vect_dot_prod_avx512.obj \
9495
bin\gf_3vect_dot_prod_avx512.obj \

‎erasure_code/Makefile.am‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -82,6 +82,7 @@ lsrc_x86_64 += \
8282
erasure_code/gf_2vect_mad_avx2_gfni.asm \
8383
erasure_code/gf_3vect_mad_avx2_gfni.asm \
8484
erasure_code/gf_4vect_mad_avx2_gfni.asm \
85+
erasure_code/gf_5vect_mad_avx2_gfni.asm \
8586
erasure_code/gf_vect_dot_prod_avx512.asm \
8687
erasure_code/gf_2vect_dot_prod_avx512.asm \
8788
erasure_code/gf_3vect_dot_prod_avx512.asm \

‎erasure_code/ec_highlevel_func.c‎

Lines changed: 10 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -281,6 +281,8 @@ extern void gf_3vect_mad_avx2_gfni(int len, int vec, int vec_i, unsigned char *g
281281
unsigned char *src, unsigned char **dest);
282282
extern void gf_4vect_mad_avx2_gfni(int len, int vec, int vec_i, unsigned char *gftbls,
283283
unsigned char *src, unsigned char **dest);
284+
extern void gf_5vect_mad_avx2_gfni(int len, int vec, int vec_i, unsigned char *gftbls,
285+
unsigned char *src, unsigned char **dest);
284286

285287
void ec_init_tables_gfni(int k, int rows, unsigned char *a, unsigned char *g_tbls)
286288
{
@@ -384,13 +386,16 @@ void ec_encode_data_update_avx2_gfni(int len, int k, int rows, int vec_i,
384386
unsigned char *g_tbls, unsigned char *data,
385387
unsigned char **coding)
386388
{
387-
while (rows >= 4) {
388-
gf_4vect_mad_avx2_gfni(len, k, vec_i, g_tbls, data, coding);
389-
g_tbls += 4 * k * 8;
390-
coding += 4;
391-
rows -= 4;
389+
while (rows >= 5) {
390+
gf_5vect_mad_avx2_gfni(len, k, vec_i, g_tbls, data, coding);
391+
g_tbls += 5 * k * 8;
392+
coding += 5;
393+
rows -= 5;
392394
}
393395
switch (rows) {
396+
case 4:
397+
gf_4vect_mad_avx2_gfni(len, k, vec_i, g_tbls, data, coding);
398+
break;
394399
case 3:
395400
gf_3vect_mad_avx2_gfni(len, k, vec_i, g_tbls, data, coding);
396401
break;

‎erasure_code/gf_5vect_mad_avx2_gfni.asm‎

Lines changed: 265 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,265 @@
1+
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
2+
; Copyright(c) 2023 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+
;;;
31+
;;; gf_5vect_mad_avx2_gfni(len, vec, vec_i, mul_array, src, dest);
32+
;;;
33+
34+
%include "reg_sizes.asm"
35+
%include "gf_vect_gfni.inc"
36+
%include "memcpy.asm"
37+
38+
%if AS_FEATURE_LEVEL >= 10
39+
40+
%ifidn __OUTPUT_FORMAT__, elf64
41+
%define arg0 rdi
42+
%define arg1 rsi
43+
%define arg2 rdx
44+
%define arg3 rcx
45+
%define arg4 r8
46+
%define arg5 r9
47+
%define tmp r11
48+
%define tmp2 r10
49+
%define tmp3 r12
50+
%define tmp4 r13
51+
%define func(x) x: endbranch
52+
%define stack_size 2*8
53+
%macro FUNC_SAVE 0
54+
sub rsp, stack_size
55+
mov [rsp + 0*8], r12
56+
mov [rsp + 1*8], r13
57+
%endmacro
58+
%macro FUNC_RESTORE 0
59+
mov r12, [rsp + 0*8]
60+
mov r13, [rsp + 1*8]
61+
add rsp, stack_size
62+
%endmacro
63+
%endif
64+
65+
%ifidn __OUTPUT_FORMAT__, win64
66+
%define arg0 rcx
67+
%define arg1 rdx
68+
%define arg2 r8
69+
%define arg3 r9
70+
%define arg4 r12 ; must be saved, loaded and restored
71+
%define arg5 r13 ; must be saved and restored
72+
%define tmp r11
73+
%define tmp2 r10
74+
%define tmp3 r14
75+
%define tmp4 r15
76+
%define stack_size 16*10 + 5*8
77+
%define arg(x) [rsp + stack_size + 8 + 8*x]
78+
%define func(x) proc_frame x
79+
80+
%macro FUNC_SAVE 0
81+
sub rsp, stack_size
82+
vmovdqa [rsp + 0*16], xmm6
83+
vmovdqa [rsp + 1*16], xmm7
84+
vmovdqa [rsp + 2*16], xmm8
85+
vmovdqa [rsp + 3*16], xmm9
86+
vmovdqa [rsp + 4*16], xmm10
87+
vmovdqa [rsp + 5*16], xmm11
88+
vmovdqa [rsp + 6*16], xmm12
89+
vmovdqa [rsp + 7*16], xmm13
90+
vmovdqa [rsp + 8*16], xmm14
91+
vmovdqa [rsp + 9*16], xmm15
92+
mov [rsp + 10*16 + 0*8], r12
93+
mov [rsp + 10*16 + 1*8], r13
94+
mov [rsp + 10*16 + 2*8], r14
95+
mov [rsp + 10*16 + 3*8], r15
96+
end_prolog
97+
mov arg4, arg(4)
98+
mov arg5, arg(5)
99+
%endmacro
100+
101+
%macro FUNC_RESTORE 0
102+
vmovdqa xmm6, [rsp + 0*16]
103+
vmovdqa xmm7, [rsp + 1*16]
104+
vmovdqa xmm8, [rsp + 2*16]
105+
vmovdqa xmm9, [rsp + 3*16]
106+
vmovdqa xmm10, [rsp + 4*16]
107+
vmovdqa xmm11, [rsp + 5*16]
108+
vmovdqa xmm12, [rsp + 6*16]
109+
vmovdqa xmm13, [rsp + 7*16]
110+
vmovdqa xmm14, [rsp + 8*16]
111+
vmovdqa xmm15, [rsp + 9*16]
112+
mov r12, [rsp + 10*16 + 0*8]
113+
mov r13, [rsp + 10*16 + 1*8]
114+
mov r14, [rsp + 10*16 + 2*8]
115+
mov r15, [rsp + 10*16 + 3*8]
116+
add rsp, stack_size
117+
%endmacro
118+
%endif
119+
120+
%define len arg0
121+
%define vec arg1
122+
%define vec_i arg2
123+
%define mul_array arg3
124+
%define src arg4
125+
%define dest1 arg5
126+
%define pos rax
127+
%define dest2 mul_array
128+
%define dest3 vec_i
129+
%define dest4 tmp3
130+
%define dest5 tmp4
131+
132+
%ifndef EC_ALIGNED_ADDR
133+
;;; Use Un-aligned load/store
134+
%define XLDR vmovdqu
135+
%define XSTR vmovdqu
136+
%else
137+
;;; Use Non-temporal load/stor
138+
%ifdef NO_NT_LDST
139+
%define XLDR vmovdqa
140+
%define XSTR vmovdqa
141+
%else
142+
%define XLDR vmovntdqa
143+
%define XSTR vmovntdq
144+
%endif
145+
%endif
146+
147+
default rel
148+
[bits 64]
149+
section .text
150+
151+
%define x0 ymm0
152+
%define xd1 ymm1
153+
%define xd2 ymm2
154+
%define xd3 ymm3
155+
%define xd4 ymm4
156+
%define xd5 ymm5
157+
%define xgft1 ymm6
158+
%define xgft2 ymm7
159+
%define xgft3 ymm8
160+
%define xgft4 ymm9
161+
%define xgft5 ymm10
162+
%define xret1 ymm11
163+
%define xret2 ymm12
164+
%define xret3 ymm13
165+
%define xret4 ymm14
166+
%define xret5 ymm15
167+
168+
;;
169+
;; Encodes 32 bytes of a single source into 5x 32 bytes (parity disks)
170+
;;
171+
%macro ENCODE_32B_5 0
172+
;; get next source vector
173+
XLDR x0, [src + pos]
174+
;; get next dest vectors
175+
XLDR xd1, [dest1 + pos]
176+
XLDR xd2, [dest2 + pos]
177+
XLDR xd3, [dest3 + pos]
178+
XLDR xd4, [dest4 + pos]
179+
XLDR xd5, [dest5 + pos]
180+
181+
GF_MUL_XOR VEX, x0, xgft1, xret1, xd1, xgft2, xret2, xd2, \
182+
xgft3, xret3, xd3, xgft4, xret4, xd4, xgft5, xret5, xd5
183+
184+
XSTR [dest1 + pos], xd1
185+
XSTR [dest2 + pos], xd2
186+
XSTR [dest3 + pos], xd3
187+
XSTR [dest4 + pos], xd4
188+
XSTR [dest5 + pos], xd5
189+
%endmacro
190+
191+
;;
192+
;; Encodes less than 32 bytes of a single source into 5x parity disks
193+
;;
194+
%macro ENCODE_LT_32B_5 1
195+
%define %%LEN %1
196+
;; get next source vector
197+
simd_load_avx2 x0, src + pos, %%LEN, tmp, tmp2
198+
;; get next dest vectors
199+
simd_load_avx2 xd1, dest1 + pos, %%LEN, tmp, tmp2
200+
simd_load_avx2 xd2, dest2 + pos, %%LEN, tmp, tmp2
201+
simd_load_avx2 xd3, dest3 + pos, %%LEN, tmp, tmp2
202+
simd_load_avx2 xd4, dest4 + pos, %%LEN, tmp, tmp2
203+
simd_load_avx2 xd5, dest5 + pos, %%LEN, tmp, tmp2
204+
205+
GF_MUL_XOR VEX, x0, xgft1, xret1, xd1, xgft2, xret2, xd2, \
206+
xgft3, xret3, xd3, xgft4, xret4, xd4, xgft5, xret5, xd5
207+
208+
lea dest1, [dest1 + pos]
209+
simd_store_avx2 dest1, xd1, %%LEN, tmp, tmp2
210+
lea dest2, [dest2 + pos]
211+
simd_store_avx2 dest2, xd2, %%LEN, tmp, tmp2
212+
lea dest3, [dest3 + pos]
213+
simd_store_avx2 dest3, xd3, %%LEN, tmp, tmp2
214+
lea dest4, [dest4 + pos]
215+
simd_store_avx2 dest4, xd4, %%LEN, tmp, tmp2
216+
lea dest5, [dest5 + pos]
217+
simd_store_avx2 dest5, xd5, %%LEN, tmp, tmp2
218+
%endmacro
219+
220+
align 16
221+
mk_global gf_5vect_mad_avx2_gfni, function
222+
func(gf_5vect_mad_avx2_gfni)
223+
FUNC_SAVE
224+
225+
xor pos, pos
226+
shl vec_i, 3 ;Multiply by 8
227+
shl vec, 3 ;Multiply by 8
228+
lea tmp, [mul_array + vec_i]
229+
lea tmp2, [vec*3]
230+
vbroadcastsd xgft1, [tmp]
231+
vbroadcastsd xgft2, [tmp + vec]
232+
vbroadcastsd xgft3, [tmp + vec*2]
233+
vbroadcastsd xgft4, [tmp + tmp2]
234+
vbroadcastsd xgft5, [tmp + vec*4]
235+
mov dest2, [dest1 + 1*8] ; reuse mul_array
236+
mov dest3, [dest1 + 2*8] ; reuse vec_i
237+
mov dest4, [dest1 + 3*8]
238+
mov dest5, [dest1 + 4*8]
239+
mov dest1, [dest1]
240+
241+
cmp len, 32
242+
jb .len_lt_32
243+
244+
.loop32:
245+
ENCODE_32B_5 ;; loop on 32 bytes at a time
246+
247+
add pos, 32
248+
sub len, 32
249+
cmp len, 32
250+
jge .loop32
251+
252+
.len_lt_32:
253+
cmp len, 0
254+
jle .exit
255+
256+
ENCODE_LT_32B_5 len ;; encode final bytes
257+
258+
.exit:
259+
vzeroupper
260+
261+
FUNC_RESTORE
262+
ret
263+
264+
endproc_frame
265+
%endif ; if AS_FEATURE_LEVEL >= 10

0 commit comments

Comments
 (0)