erasure_code: optimize AVX2 GFNI single vector dot product

Signed-off-by: Marcel Cornu <marcel.d.cornu@intel.com>
This commit is contained in:
Marcel Cornu 2023-12-07 22:51:27 +00:00 committed by Tomasz Kantecki
parent 164d9ff1f0
commit 3f87141d03

View File

@ -47,17 +47,17 @@
%define tmp r11
%define tmp2 r10
%define tmp3 r12 ;must be saved and restored
%define tmp4 r13 ;must be saved and restored
%define tmp3 r12 ; must be saved and restored
%define stack_size 1*8
%define func(x) x: endbranch
%macro FUNC_SAVE 0
push r12
push r13
sub rsp, stack_size
mov [rsp + 0*8], r12
%endmacro
%macro FUNC_RESTORE 0
pop r13
pop r12
mov r12, [rsp + 0*8]
add rsp, stack_size
%endmacro
%endif
@ -67,32 +67,37 @@
%define arg2 r8
%define arg3 r9
%define arg4 r12 ; must be saved, loaded and restored
%define arg5 r13 ; must be saved and restored
%define arg4 r12 ; must be saved, loaded and restored
%define arg5 r15 ; must be saved and restored
%define tmp r11
%define tmp2 r10
%define tmp3 r14
%define tmp4 r15
%define stack_size 0*16 + 5*8 ; must be an odd multiple of 8
%define tmp3 r13 ; must be saved and restored
%define stack_size 4*16 + 3*8 ; must be an odd multiple of 8
%define arg(x) [rsp + stack_size + 8 + 8*x]
%define func(x) proc_frame x
%macro FUNC_SAVE 0
alloc_stack stack_size
mov [rsp + 0*8], r12
mov [rsp + 1*8], r13
mov [rsp + 2*8], r14
mov [rsp + 3*8], r15
end_prolog
mov arg4, arg(4)
alloc_stack stack_size
vmovdqa [rsp + 0*16], xmm6
vmovdqa [rsp + 1*16], xmm7
vmovdqa [rsp + 2*16], xmm8
vmovdqa [rsp + 3*16], xmm9
mov [rsp + 4*16 + 0*8], r12
mov [rsp + 4*16 + 1*8], r13
mov [rsp + 4*16 + 2*8], r15
end_prolog
mov arg4, arg(4)
%endmacro
%macro FUNC_RESTORE 0
mov r12, [rsp + 0*8]
mov r13, [rsp + 1*8]
mov r14, [rsp + 2*8]
mov r15, [rsp + 3*8]
add rsp, stack_size
vmovdqa xmm6, [rsp + 0*16]
vmovdqa xmm7, [rsp + 1*16]
vmovdqa xmm8, [rsp + 2*16]
vmovdqa xmm9, [rsp + 3*16]
mov r12, [rsp + 4*16 + 0*8]
mov r13, [rsp + 4*16 + 1*8]
mov r15, [rsp + 4*16 + 2*8]
add rsp, stack_size
%endmacro
%endif
@ -106,7 +111,6 @@
%define vec_i tmp2
%define pos rax
%ifndef EC_ALIGNED_ADDR
;;; Use Un-aligned load/store
%define XLDR vmovdqu
@ -122,95 +126,193 @@
%endif
%endif
%define xgft1 ymm2
%define x0l ymm0
%define x0h ymm1
%define x0x ymm2
%define x0 ymm0
%define xp1 ymm1
%define xp1l ymm3
%define xp1h ymm4
%define xp1x ymm5
%define xgft1 ymm6
%define xgft2 ymm7
%define xgft3 ymm8
%define xtmp1 ymm9
%define x0 x0l
%define xp1 xp1l
%define xp2 xp2l
%define xp3 xp3l
default rel
[bits 64]
section .text
;;
;; Encodes 32 bytes of all "k" sources into 32 bytes (single parity disk)
;; Encodes 96 bytes of all "k" sources into 96 bytes (single parity disk)
;;
%macro ENCODE_32B 0
vpxor xp1, xp1, xp1
mov tmp, mul_array
xor vec_i, vec_i
%macro ENCODE_96B 0
vpxor xp1l, xp1l, xp1l
vpxor xp1h, xp1h, xp1h
vpxor xp1x, xp1x, xp1x
mov tmp, mul_array
xor vec_i, vec_i
%%next_vect:
mov ptr, [src + vec_i]
XLDR x0, [ptr + pos] ;Get next source vector (32 bytes)
add vec_i, 8
;; load next source vector
mov ptr, [src + vec_i]
XLDR x0l, [ptr + pos]
XLDR x0h, [ptr + pos + 32]
XLDR x0x, [ptr + pos + 64]
add vec_i, 8
vbroadcastsd xgft1, [tmp]
add tmp, 8
add tmp, 8
GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1
GF_MUL_XOR VEX, x0l, xgft1, xtmp1, xp1l
GF_MUL_XOR VEX, x0h, xgft1, xtmp1, xp1h
GF_MUL_XOR VEX, x0x, xgft1, xtmp1, xp1x
cmp vec_i, vec
jl %%next_vect
cmp vec_i, vec
jl %%next_vect
XSTR [dest1 + pos], xp1
XSTR [dest1 + pos], xp1l
XSTR [dest1 + pos + 32], xp1h
XSTR [dest1 + pos + 64], xp1x
%endmacro
;;
;; Encodes less than 32 bytes of all "k" sources and updates single parity disk
;; Encodes 64 bytes of all "k" sources into 64 bytes (single parity disk)
;;
%macro ENCODE_LT_32B 1
%define %%LEN %1
vpxor xp1, xp1, xp1
mov tmp, mul_array
xor vec_i, vec_i
%macro ENCODE_64B 0
vpxor xp1l, xp1l, xp1l
vpxor xp1h, xp1h, xp1h
mov tmp, mul_array
xor vec_i, vec_i
%%next_vect:
mov ptr, [src + vec_i]
simd_load_avx2 x0, ptr + pos, %%LEN, tmp3, tmp4 ;Get next source vector
add vec_i, 8
;; load next source vector
mov ptr, [src + vec_i]
XLDR x0l, [ptr + pos]
XLDR x0h, [ptr + pos + 32]
add vec_i, 8
vbroadcastsd xgft1, [tmp]
add tmp, 8
add tmp, 8
GF_MUL_XOR VEX, x0l, xgft1, xtmp1, xp1l
GF_MUL_XOR VEX, x0h, xgft1, xgft1, xp1h
cmp vec_i, vec
jl %%next_vect
XSTR [dest1 + pos], xp1l
XSTR [dest1 + pos + 32], xp1h
%endmacro
;;
;; Encodes 32 bytes of all "k" sources into 32 bytes (single parity disks)
;;
%macro ENCODE_32B 0
vpxor xp1, xp1, xp1
mov tmp, mul_array
xor vec_i, vec_i
%%next_vect:
;; load next source vector
mov ptr, [src + vec_i]
XLDR x0, [ptr + pos]
add vec_i, 8
vbroadcastsd xgft1, [tmp]
add tmp, 8
GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1
cmp vec_i, vec
jl %%next_vect
cmp vec_i, vec
jl %%next_vect
lea tmp, [dest1 + pos]
simd_store_avx2 tmp, xp1, %%LEN, tmp3, tmp4 ;Store updated encoded data
XSTR [dest1 + pos], xp1
%endmacro
;;
;; Encodes less than 32 bytes of all "k" sources into single parity disks
;;
%macro ENCODE_LT_32B 1
%define %%LEN %1
vpxor xp1, xp1, xp1
xor vec_i, vec_i
%%next_vect:
; get next source vector
mov ptr, [src + vec_i]
simd_load_avx2 x0, ptr + pos, %%LEN, tmp, tmp3
add vec_i, 8
vbroadcastsd xgft1, [mul_array]
add mul_array, 8
GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1
cmp vec_i, vec
jl %%next_vect
;; Store updated encoded data
lea ptr, [dest1 + pos]
simd_store_avx2 ptr, xp1, %%LEN, tmp, vec_i
%endmacro
align 16
mk_global gf_vect_dot_prod_avx2_gfni, function
func(gf_vect_dot_prod_avx2_gfni)
FUNC_SAVE
xor pos, pos
shl vec, 3 ;vec *= 8. Make vec_i count by 8
FUNC_SAVE
cmp len, 32
jb .len_lt_32
xor pos, pos
shl vec, 3 ;; vec *= 8. Make vec_i count by 8
.loop32:
cmp len, 96
jl .len_lt_96
.loop96:
ENCODE_96B
add pos, 96 ;; Loop on 96 bytes at a time first
sub len, 96
cmp len, 96
jge .loop96
.len_lt_96:
cmp len, 64
jl .len_lt_64
ENCODE_64B
add pos, 64 ;; encode next 64 bytes
sub len, 64
.len_lt_64:
cmp len, 32
jl .len_lt_32
ENCODE_32B
add pos, 32 ;Loop on 32 bytes at a time
add pos, 32 ;; encode next 32 bytes
sub len, 32
cmp len, 32
jge .loop32
.len_lt_32:
cmp len, 0
jle .exit
ENCODE_LT_32B len
ENCODE_LT_32B len ;; encode final bytes
.exit:
vzeroupper
FUNC_RESTORE
ret
FUNC_RESTORE
ret
endproc_frame
%endif ; if AS_FEATURE_LEVEL >= 10