mirror of
https://github.com/intel/isa-l.git
synced 2024-12-13 09:52:56 +01:00
erasure_code: optimize AVX2 GFNI 2 vector dot product
Signed-off-by: Marcel Cornu <marcel.d.cornu@intel.com>
This commit is contained in:
Marcel Cornu
Tomasz Kantecki
parent
3f87141d03
commit
0052080f53
@ -80,36 +80,42 @@
|
|||||||
%define tmp3 r13 ; must be saved and restored
|
%define tmp3 r13 ; must be saved and restored
|
||||||
%define tmp4 r14 ; must be saved and restored
|
%define tmp4 r14 ; must be saved and restored
|
||||||
%define tmp5 rdi ; must be saved and restored
|
%define tmp5 rdi ; must be saved and restored
|
||||||
%define stack_size 4*16 + 5*8 ; must be an odd multiple of 8
|
%define stack_size 7*16 + 5*8 ; must be an odd multiple of 8
|
||||||
%define arg(x) [rsp + stack_size + 8 + 8*x]
|
%define arg(x) [rsp + stack_size + 8 + 8*x]
|
||||||
|
|
||||||
%define func(x) proc_frame x
|
%define func(x) proc_frame x
|
||||||
%macro FUNC_SAVE 0
|
%macro FUNC_SAVE 0
|
||||||
alloc_stack stack_size
|
alloc_stack stack_size
|
||||||
vmovdqa [rsp + 0*16], xmm6
|
vmovdqa [rsp + 0*16], xmm6
|
||||||
vmovdqa [rsp + 1*16], xmm7
|
vmovdqa [rsp + 1*16], xmm7
|
||||||
vmovdqa [rsp + 2*16], xmm8
|
vmovdqa [rsp + 2*16], xmm8
|
||||||
vmovdqa [rsp + 3*16], xmm9
|
vmovdqa [rsp + 3*16], xmm9
|
||||||
mov [rsp + 4*16 + 0*8], r12
|
vmovdqa [rsp + 4*16], xmm10
|
||||||
mov [rsp + 4*16 + 1*8], r13
|
vmovdqa [rsp + 5*16], xmm11
|
||||||
mov [rsp + 4*16 + 2*8], r14
|
vmovdqa [rsp + 6*16], xmm12
|
||||||
mov [rsp + 4*16 + 3*8], r15
|
mov [rsp + 7*16 + 0*8], r12
|
||||||
mov [rsp + 4*16 + 4*8], rdi
|
mov [rsp + 7*16 + 1*8], r13
|
||||||
end_prolog
|
mov [rsp + 7*16 + 2*8], r14
|
||||||
mov arg4, arg(4)
|
mov [rsp + 7*16 + 3*8], r15
|
||||||
|
mov [rsp + 7*16 + 4*8], rdi
|
||||||
|
end_prolog
|
||||||
|
mov arg4, arg(4)
|
||||||
%endmacro
|
%endmacro
|
||||||
|
|
||||||
%macro FUNC_RESTORE 0
|
%macro FUNC_RESTORE 0
|
||||||
vmovdqa xmm6, [rsp + 0*16]
|
vmovdqa xmm6, [rsp + 0*16]
|
||||||
vmovdqa xmm7, [rsp + 1*16]
|
vmovdqa xmm7, [rsp + 1*16]
|
||||||
vmovdqa xmm8, [rsp + 2*16]
|
vmovdqa xmm8, [rsp + 2*16]
|
||||||
vmovdqa xmm9, [rsp + 3*16]
|
vmovdqa xmm9, [rsp + 3*16]
|
||||||
mov r12, [rsp + 4*16 + 0*8]
|
vmovdqa xmm10, [rsp + 4*16]
|
||||||
mov r13, [rsp + 4*16 + 1*8]
|
vmovdqa xmm11, [rsp + 5*16]
|
||||||
mov r14, [rsp + 4*16 + 2*8]
|
vmovdqa xmm12, [rsp + 6*16]
|
||||||
mov r15, [rsp + 4*16 + 3*8]
|
mov r12, [rsp + 7*16 + 0*8]
|
||||||
mov rdi, [rsp + 4*16 + 4*8]
|
mov r13, [rsp + 7*16 + 1*8]
|
||||||
add rsp, stack_size
|
mov r14, [rsp + 7*16 + 2*8]
|
||||||
|
mov r15, [rsp + 7*16 + 3*8]
|
||||||
|
mov rdi, [rsp + 7*16 + 4*8]
|
||||||
|
add rsp, stack_size
|
||||||
%endmacro
|
%endmacro
|
||||||
%endif
|
%endif
|
||||||
|
|
||||||
@ -142,17 +148,22 @@
|
|||||||
|
|
||||||
%define x0l ymm0
|
%define x0l ymm0
|
||||||
%define x0h ymm1
|
%define x0h ymm1
|
||||||
|
%define x0x ymm2
|
||||||
|
|
||||||
%define xgft1 ymm2
|
%define xgft1 ymm3
|
||||||
%define xgft2 ymm3
|
%define xgft2 ymm4
|
||||||
|
|
||||||
%define xtmp1 ymm4
|
%define xtmp1 ymm5
|
||||||
%define xtmp2 ymm5
|
%define xtmp2 ymm6
|
||||||
|
|
||||||
%define xp1l ymm6
|
%define xp1l ymm7
|
||||||
%define xp2l ymm7
|
%define xp2l ymm8
|
||||||
%define xp1h ymm8
|
|
||||||
%define xp2h ymm9
|
%define xp1h ymm9
|
||||||
|
%define xp2h ymm10
|
||||||
|
|
||||||
|
%define xp1x ymm11
|
||||||
|
%define xp2x ymm12
|
||||||
|
|
||||||
%define x0 x0l
|
%define x0 x0l
|
||||||
%define xp1 xp1l
|
%define xp1 xp1l
|
||||||
@ -163,64 +174,105 @@ default rel
|
|||||||
|
|
||||||
section .text
|
section .text
|
||||||
|
|
||||||
|
;;
|
||||||
|
;; Encodes 96 bytes of all "k" sources into 2x 96 bytes (parity disk)
|
||||||
|
;;
|
||||||
|
%macro ENCODE_96B_2 0
|
||||||
|
vpxor xp1l, xp1l, xp1l
|
||||||
|
vpxor xp1h, xp1h, xp1h
|
||||||
|
vpxor xp1x, xp1x, xp1x
|
||||||
|
|
||||||
|
vpxor xp2l, xp2l, xp2l
|
||||||
|
vpxor xp2h, xp2h, xp2h
|
||||||
|
vpxor xp2x, xp2x, xp2x
|
||||||
|
mov tmp, mul_array
|
||||||
|
xor vec_i, vec_i
|
||||||
|
|
||||||
|
%%next_vect:
|
||||||
|
;; 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]
|
||||||
|
vbroadcastsd xgft2, [tmp + vec]
|
||||||
|
|
||||||
|
GF_MUL_XOR VEX, x0l, xgft1, xtmp1, xp1l, xgft2, xtmp2, xp2l
|
||||||
|
GF_MUL_XOR VEX, x0h, xgft1, xtmp1, xp1h, xgft2, xtmp2, xp2h
|
||||||
|
GF_MUL_XOR VEX, x0x, xgft1, xtmp1, xp1x, xgft2, xtmp2, xp2x
|
||||||
|
add tmp, 8
|
||||||
|
|
||||||
|
cmp vec_i, vec
|
||||||
|
jl %%next_vect
|
||||||
|
|
||||||
|
XSTR [dest1 + pos], xp1l
|
||||||
|
XSTR [dest1 + pos + 32], xp1h
|
||||||
|
XSTR [dest1 + pos + 64], xp1x
|
||||||
|
XSTR [dest2 + pos], xp2l
|
||||||
|
XSTR [dest2 + pos + 32], xp2h
|
||||||
|
XSTR [dest2 + pos + 64], xp2x
|
||||||
|
%endmacro
|
||||||
|
|
||||||
;;
|
;;
|
||||||
;; Encodes 64 bytes of all "k" sources into 2x 64 bytes (parity disks)
|
;; Encodes 64 bytes of all "k" sources into 2x 64 bytes (parity disks)
|
||||||
;;
|
;;
|
||||||
%macro ENCODE_64B_2 0
|
%macro ENCODE_64B_2 0
|
||||||
vpxor xp1l, xp1l, xp1l
|
vpxor xp1l, xp1l, xp1l
|
||||||
vpxor xp1h, xp1h, xp1h
|
vpxor xp1h, xp1h, xp1h
|
||||||
vpxor xp2l, xp2l, xp2l
|
vpxor xp2l, xp2l, xp2l
|
||||||
vpxor xp2h, xp2h, xp2h
|
vpxor xp2h, xp2h, xp2h
|
||||||
mov tmp, mul_array
|
mov tmp, mul_array
|
||||||
xor vec_i, vec_i
|
xor vec_i, vec_i
|
||||||
|
|
||||||
%%next_vect:
|
%%next_vect:
|
||||||
mov ptr, [src + vec_i]
|
mov ptr, [src + vec_i]
|
||||||
XLDR x0l, [ptr + pos] ;; Get next source vector low 32 bytes
|
XLDR x0l, [ptr + pos] ;; Get next source vector low 32 bytes
|
||||||
XLDR x0h, [ptr + pos + 32] ;; Get next source vector high 32 bytes
|
XLDR x0h, [ptr + pos + 32] ;; Get next source vector high 32 bytes
|
||||||
add vec_i, 8
|
add vec_i, 8
|
||||||
|
|
||||||
vbroadcastsd xgft1, [tmp]
|
vbroadcastsd xgft1, [tmp]
|
||||||
vbroadcastsd xgft2, [tmp + vec]
|
vbroadcastsd xgft2, [tmp + vec]
|
||||||
add tmp, 8
|
add tmp, 8
|
||||||
|
|
||||||
GF_MUL_XOR VEX, x0l, xgft1, xtmp1, xp1l, xgft2, xtmp2, xp2l
|
GF_MUL_XOR VEX, x0l, xgft1, xtmp1, xp1l, xgft2, xtmp2, xp2l
|
||||||
GF_MUL_XOR VEX, x0h, xgft1, xgft1, xp1h, xgft2, xgft2, xp2h
|
GF_MUL_XOR VEX, x0h, xgft1, xgft1, xp1h, xgft2, xgft2, xp2h
|
||||||
|
|
||||||
cmp vec_i, vec
|
cmp vec_i, vec
|
||||||
jl %%next_vect
|
jl %%next_vect
|
||||||
|
|
||||||
XSTR [dest1 + pos], xp1l
|
XSTR [dest1 + pos], xp1l
|
||||||
XSTR [dest1 + pos + 32], xp1h
|
XSTR [dest1 + pos + 32], xp1h
|
||||||
XSTR [dest2 + pos], xp2l
|
XSTR [dest2 + pos], xp2l
|
||||||
XSTR [dest2 + pos + 32], xp2h
|
XSTR [dest2 + pos + 32], xp2h
|
||||||
%endmacro
|
%endmacro
|
||||||
|
|
||||||
;;
|
;;
|
||||||
;; Encodes 32 bytes of all "k" sources into 2x 32 bytes (parity disks)
|
;; Encodes 32 bytes of all "k" sources into 2x 32 bytes (parity disks)
|
||||||
;;
|
;;
|
||||||
%macro ENCODE_32B_2 0
|
%macro ENCODE_32B_2 0
|
||||||
vpxor xp1, xp1, xp1
|
vpxor xp1, xp1, xp1
|
||||||
vpxor xp2, xp2, xp2
|
vpxor xp2, xp2, xp2
|
||||||
mov tmp, mul_array
|
mov tmp, mul_array
|
||||||
xor vec_i, vec_i
|
xor vec_i, vec_i
|
||||||
|
|
||||||
%%next_vect:
|
%%next_vect:
|
||||||
mov ptr, [src + vec_i]
|
mov ptr, [src + vec_i]
|
||||||
XLDR x0, [ptr + pos] ;Get next source vector (32 bytes)
|
XLDR x0, [ptr + pos] ;Get next source vector (32 bytes)
|
||||||
add vec_i, 8
|
add vec_i, 8
|
||||||
|
|
||||||
vbroadcastsd xgft1, [tmp]
|
vbroadcastsd xgft1, [tmp]
|
||||||
vbroadcastsd xgft2, [tmp + vec]
|
vbroadcastsd xgft2, [tmp + vec]
|
||||||
add tmp, 8
|
add tmp, 8
|
||||||
|
|
||||||
GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1, xgft2, xgft2, xp2
|
GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1, xgft2, xgft2, xp2
|
||||||
|
|
||||||
cmp vec_i, vec
|
cmp vec_i, vec
|
||||||
jl %%next_vect
|
jl %%next_vect
|
||||||
|
|
||||||
XSTR [dest1 + pos], xp1
|
XSTR [dest1 + pos], xp1
|
||||||
XSTR [dest2 + pos], xp2
|
XSTR [dest2 + pos], xp2
|
||||||
%endmacro
|
%endmacro
|
||||||
|
|
||||||
;;
|
;;
|
||||||
@ -229,73 +281,82 @@ section .text
|
|||||||
%macro ENCODE_LT_32B_2 1
|
%macro ENCODE_LT_32B_2 1
|
||||||
%define %%LEN %1
|
%define %%LEN %1
|
||||||
|
|
||||||
vpxor xp1, xp1, xp1
|
vpxor xp1, xp1, xp1
|
||||||
vpxor xp2, xp2, xp2
|
vpxor xp2, xp2, xp2
|
||||||
xor vec_i, vec_i
|
xor vec_i, vec_i
|
||||||
|
|
||||||
%%next_vect:
|
%%next_vect:
|
||||||
mov ptr, [src + vec_i]
|
mov ptr, [src + vec_i]
|
||||||
simd_load_avx2 x0, ptr + pos, %%LEN, tmp, tmp4 ;Get next source vector
|
simd_load_avx2 x0, ptr + pos, %%LEN, tmp, tmp4 ;Get next source vector
|
||||||
add vec_i, 8
|
add vec_i, 8
|
||||||
|
|
||||||
vbroadcastsd xgft1, [mul_array]
|
vbroadcastsd xgft1, [mul_array]
|
||||||
vbroadcastsd xgft2, [mul_array + vec]
|
vbroadcastsd xgft2, [mul_array + vec]
|
||||||
add mul_array, 8
|
add mul_array, 8
|
||||||
|
|
||||||
GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1, xgft2, xgft2, xp2
|
GF_MUL_XOR VEX, x0, xgft1, xgft1, xp1, xgft2, xgft2, xp2
|
||||||
|
|
||||||
cmp vec_i, vec
|
cmp vec_i, vec
|
||||||
jl %%next_vect
|
jl %%next_vect
|
||||||
|
|
||||||
;Store updated encoded data
|
;Store updated encoded data
|
||||||
lea ptr, [dest1 + pos]
|
lea ptr, [dest1 + pos]
|
||||||
simd_store_avx2 ptr, xp1, %%LEN, tmp, vec_i
|
simd_store_avx2 ptr, xp1, %%LEN, tmp, tmp4
|
||||||
|
|
||||||
lea ptr, [dest2 + pos]
|
lea ptr, [dest2 + pos]
|
||||||
simd_store_avx2 ptr, xp2, %%LEN, tmp, vec_i
|
simd_store_avx2 ptr, xp2, %%LEN, tmp, tmp4
|
||||||
%endmacro
|
%endmacro
|
||||||
|
|
||||||
align 16
|
align 16
|
||||||
mk_global gf_2vect_dot_prod_avx2_gfni, function
|
mk_global gf_2vect_dot_prod_avx2_gfni, function
|
||||||
func(gf_2vect_dot_prod_avx2_gfni)
|
func(gf_2vect_dot_prod_avx2_gfni)
|
||||||
FUNC_SAVE
|
FUNC_SAVE
|
||||||
|
|
||||||
xor pos, pos
|
xor pos, pos
|
||||||
shl vec, 3 ;; vec *= 8. Make vec_i count by 8
|
shl vec, 3 ;; vec *= 8. Make vec_i count by 8
|
||||||
mov dest1, [dest]
|
mov dest1, [dest]
|
||||||
mov dest2, [dest + 8]
|
mov dest2, [dest + 8]
|
||||||
|
|
||||||
cmp len, 64
|
cmp len, 96
|
||||||
jb .len_lt_64
|
jl .len_lt_96
|
||||||
|
|
||||||
.loop64:
|
.loop96:
|
||||||
ENCODE_64B_2
|
ENCODE_96B_2
|
||||||
|
|
||||||
add pos, 64 ;; Loop on 64 bytes at a time first
|
add pos, 96 ;; Loop on 96 bytes at a time first
|
||||||
sub len, 64
|
sub len, 96
|
||||||
cmp len, 64
|
cmp len, 96
|
||||||
jge .loop64
|
jge .loop96
|
||||||
|
|
||||||
|
.len_lt_96:
|
||||||
|
cmp len, 64
|
||||||
|
jl .len_lt_64
|
||||||
|
|
||||||
|
ENCODE_64B_2
|
||||||
|
|
||||||
|
add pos, 64 ;; encode next 64 bytes
|
||||||
|
sub len, 64
|
||||||
|
|
||||||
.len_lt_64:
|
.len_lt_64:
|
||||||
cmp len, 32
|
cmp len, 32
|
||||||
jb .len_lt_32
|
jl .len_lt_32
|
||||||
|
|
||||||
ENCODE_32B_2
|
ENCODE_32B_2
|
||||||
|
|
||||||
add pos, 32 ;; encode next 32 bytes
|
add pos, 32 ;; encode next 32 bytes
|
||||||
sub len, 32
|
sub len, 32
|
||||||
|
|
||||||
.len_lt_32:
|
.len_lt_32:
|
||||||
cmp len, 0
|
cmp len, 0
|
||||||
jle .exit
|
jle .exit
|
||||||
|
|
||||||
ENCODE_LT_32B_2 len ;; encode remaining bytes
|
ENCODE_LT_32B_2 len ;; encode remaining bytes
|
||||||
|
|
||||||
.exit:
|
.exit:
|
||||||
vzeroupper
|
vzeroupper
|
||||||
|
|
||||||
FUNC_RESTORE
|
FUNC_RESTORE
|
||||||
ret
|
ret
|
||||||
|
|
||||||
endproc_frame
|
endproc_frame
|
||||||
%endif ; if AS_FEATURE_LEVEL >= 10
|
%endif ; if AS_FEATURE_LEVEL >= 10
|
||||||
|
|||||||
Loading…
Reference in New Issue
Block a user