Skip to content

GitLab

Commit 2eceefe8 authored by Henrik Gramner's avatar Henrik Gramner
Browse files

x86: AVX-512 deblock_strength 

Also drop the MMX version and make some slight improvements to the SSE2,
SSSE3, AVX, and AVX2 versions.
parent 3081ffa1
Hide whitespace changes
Inline Side-by-side
common/common.h
View file @ 2eceefe8
......@@ -823,10 +823,10 @@ struct x264_t
struct
{
/* real intra4x4_pred_mode if I_4X4 or I_8X8, I_PRED_4x4_DC if mb available, -1 if not */
ALIGNED_8( int8_t intra4x4_pred_mode[X264_SCAN8_LUMA_SIZE] );
ALIGNED_16( int8_t intra4x4_pred_mode[X264_SCAN8_LUMA_SIZE] );
/* i_non_zero_count if available else 0x80 */
ALIGNED_16( uint8_t non_zero_count[X264_SCAN8_SIZE] );
/* i_non_zero_count if available else 0x80. intentionally misaligned by 8 for asm */
ALIGNED_8( uint8_t non_zero_count[X264_SCAN8_SIZE] );
/* -1 if unused, -2 if unavailable */
ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
......
common/deblock.c
View file @ 2eceefe8
......@@ -676,21 +676,21 @@ void x264_deblock_h_chroma_intra_avx ( pixel *pix, intptr_t stride, int alpha, i
void x264_deblock_h_chroma_422_intra_mmx2( pixel *pix, intptr_t stride, int alpha, int beta );
void x264_deblock_h_chroma_422_intra_sse2( pixel *pix, intptr_t stride, int alpha, int beta );
void x264_deblock_h_chroma_422_intra_avx ( pixel *pix, intptr_t stride, int alpha, int beta );
void x264_deblock_strength_mmx2 ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_sse2 ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_ssse3( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_avx ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_avx2 ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_sse2 ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_ssse3 ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_avx ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_avx2 ( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_strength_avx512( uint8_t nnz[X264_SCAN8_SIZE], int8_t ref[2][X264_SCAN8_LUMA_SIZE],
int16_t mv[2][X264_SCAN8_LUMA_SIZE][2], uint8_t bs[2][8][4],
int mvy_limit, int bframe );
void x264_deblock_h_chroma_intra_mbaff_mmx2( pixel *pix, intptr_t stride, int alpha, int beta );
void x264_deblock_h_chroma_intra_mbaff_sse2( pixel *pix, intptr_t stride, int alpha, int beta );
......@@ -803,7 +803,6 @@ void x264_deblock_init( int cpu, x264_deblock_function_t *pf, int b_mbaff )
#if !HIGH_BIT_DEPTH
pf->deblock_chroma_420_intra_mbaff = x264_deblock_h_chroma_intra_mbaff_mmx2;
#endif
pf->deblock_strength = x264_deblock_strength_mmx2;
if( cpu&X264_CPU_SSE2 )
{
pf->deblock_strength = x264_deblock_strength_sse2;
......@@ -852,6 +851,10 @@ void x264_deblock_init( int cpu, x264_deblock_function_t *pf, int b_mbaff )
{
pf->deblock_strength = x264_deblock_strength_avx2;
}
if( cpu&X264_CPU_AVX512 )
{
pf->deblock_strength = x264_deblock_strength_avx512;
}
}
#endif
......
common/x86/deblock-a.asm
View file @ 2eceefe8
......@@ -28,10 +28,14 @@
%include "x86inc.asm"
%include "x86util.asm"
SECTION_RODATA 32
load_bytes_shuf: times 2 db 3,4,5,6,11,12,13,14,4,5,6,7,12,13,14,15
insert_top_shuf: dd 0,1,4,5,7,2,3,6
SECTION_RODATA 64
load_bytes_zmm_shuf: dd 0x50404032, 0x70606053, 0xd0c0c0b4, 0xf0e0e0d5
dd 0x50404036, 0x70606057, 0xd0c0c0b8, 0xf0e0e0d9
dd 0x50104001, 0x70306023, 0xd090c083, 0xf0b0e0a5
dd 0x50104005, 0x70306027, 0xd090c087, 0xf0b0e0a9
load_bytes_ymm_shuf: dd 0x06050403, 0x0e0d0c1b, 0x07060544, 0x0f0e0d5c
dd 0x06050473, 0x0e0d0c2b, 0x07060534, 0x0f0e0d6c
transpose_shuf: db 0,4,8,12,1,5,9,13,2,6,10,14,3,7,11,15
SECTION .text
......@@ -2276,13 +2280,10 @@ cglobal deblock_h_chroma_intra_mbaff, 4,6,8
RET
%endif ; !HIGH_BIT_DEPTH
;-----------------------------------------------------------------------------
; static void deblock_strength( uint8_t nnz[48], int8_t ref[2][40], int16_t mv[2][40][2],
; uint8_t bs[2][4][4], int mvy_limit, int bframe )
;-----------------------------------------------------------------------------
%define scan8start (4+1*8)
%define nnz r0+scan8start
%define ref r1+scan8start
......@@ -2290,145 +2291,54 @@ cglobal deblock_h_chroma_intra_mbaff, 4,6,8
%define bs0 r3
%define bs1 r3+32
%macro LOAD_BYTES_MMX 1
movd m2, [%1+8*0-1]
movd m0, [%1+8*0]
movd m3, [%1+8*2-1]
movd m1, [%1+8*2]
punpckldq m2, [%1+8*1-1]
punpckldq m0, [%1+8*1]
punpckldq m3, [%1+8*3-1]
punpckldq m1, [%1+8*3]
%endmacro
%macro DEBLOCK_STRENGTH_REFS_MMX 0
LOAD_BYTES_MMX ref
pxor m2, m0
pxor m3, m1
por m2, [bs0+0]
por m3, [bs0+8]
movq [bs0+0], m2
movq [bs0+8], m3
movd m2, [ref-8*1]
movd m3, [ref+8*1]
punpckldq m2, m0 ; row -1, row 0
punpckldq m3, m1 ; row 1, row 2
pxor m0, m2
pxor m1, m3
por m0, [bs1+0]
por m1, [bs1+8]
movq [bs1+0], m0
movq [bs1+8], m1
%endmacro
%macro DEBLOCK_STRENGTH_MVS_MMX 2
mova m0, [mv-%2]
mova m1, [mv-%2+8]
psubw m0, [mv]
psubw m1, [mv+8]
packsswb m0, m1
ABSB m0, m1
psubusb m0, m7
packsswb m0, m0
por m0, [%1]
movd [%1], m0
%endmacro
%macro DEBLOCK_STRENGTH_NNZ_MMX 1
por m2, m0
por m3, m1
mova m4, [%1]
mova m5, [%1+8]
pminub m2, m6
pminub m3, m6
pminub m4, m6 ; mv ? 1 : 0
pminub m5, m6
paddb m2, m2 ; nnz ? 2 : 0
paddb m3, m3
pmaxub m2, m4
pmaxub m3, m5
%endmacro
%macro LOAD_BYTES_XMM 1
movu m2, [%1-4] ; FIXME could be aligned if we changed nnz's allocation
%macro LOAD_BYTES_XMM 2 ; src, aligned
%if %2
mova m2, [%1-4]
mova m1, [%1+12]
%else
movu m2, [%1-4]
movu m1, [%1+12]
pslldq m0, m2, 1
%endif
psllq m0, m2, 8
shufps m2, m1, q3131 ; cur nnz, all rows
pslldq m1, 1
psllq m1, 8
shufps m0, m1, q3131 ; left neighbors
%if cpuflag(avx) || (%2 && cpuflag(ssse3))
palignr m1, m2, [%1-20], 12
%else
pslldq m1, m2, 4
movd m3, [%1-8] ; could be palignr if nnz was aligned
movd m3, [%1-8]
por m1, m3 ; top neighbors
%endif
%endmacro
INIT_MMX mmx2
cglobal deblock_strength, 6,6
; Prepare mv comparison register
shl r4d, 8
add r4d, 3 - (1<<8)
movd m7, r4d
SPLATW m7, m7
mova m6, [pb_1]
pxor m0, m0
mova [bs0+0], m0
mova [bs0+8], m0
mova [bs1+0], m0
mova [bs1+8], m0
.lists:
DEBLOCK_STRENGTH_REFS_MMX
mov r4d, 4
.mvs:
DEBLOCK_STRENGTH_MVS_MMX bs0, 4
DEBLOCK_STRENGTH_MVS_MMX bs1, 4*8
add r2, 4*8
add r3, 4
dec r4d
jg .mvs
add r1, 40
add r2, 4*8
sub r3, 16
dec r5d
jge .lists
; Check nnz
LOAD_BYTES_MMX nnz
DEBLOCK_STRENGTH_NNZ_MMX bs0
; Transpose column output
SBUTTERFLY bw, 2, 3, 4
SBUTTERFLY bw, 2, 3, 4
mova [bs0+0], m2
mova [bs0+8], m3
movd m2, [nnz-8*1]
movd m3, [nnz+8*1]
punpckldq m2, m0 ; row -1, row 0
punpckldq m3, m1 ; row 1, row 2
DEBLOCK_STRENGTH_NNZ_MMX bs1
mova [bs1+0], m2
mova [bs1+8], m3
RET
%if UNIX64
DECLARE_REG_TMP 5
%else
DECLARE_REG_TMP 4
%endif
%macro DEBLOCK_STRENGTH_XMM 0
cglobal deblock_strength, 6,6,7
cglobal deblock_strength, 5,5,7
; Prepare mv comparison register
shl r4d, 8
add r4d, 3 - (1<<8)
movd m6, r4d
movifnidn t0d, r5m
SPLATW m6, m6
pxor m4, m4 ; bs0
pxor m5, m5 ; bs1
.lists:
; Check refs
LOAD_BYTES_XMM ref
LOAD_BYTES_XMM ref, 0
pxor m0, m2
pxor m1, m2
por m4, m0
por m5, m1
; Check mvs
%if cpuflag(ssse3)
%if cpuflag(ssse3) && notcpuflag(avx)
mova m0, [mv+4*8*0]
mova m1, [mv+4*8*1]
palignr m3, m0, [mv+4*8*0-16], 12
......@@ -2481,11 +2391,11 @@ cglobal deblock_strength, 6,6,7
por m5, m0
add r1, 40
add r2, 4*8*5
dec r5d
dec t0d
jge .lists
; Check nnz
LOAD_BYTES_XMM nnz
LOAD_BYTES_XMM nnz, 1
por m0, m2
por m1, m2
mova m6, [pb_1]
......@@ -2518,68 +2428,121 @@ INIT_XMM avx
DEBLOCK_STRENGTH_XMM
%macro LOAD_BYTES_YMM 1
movu m0, [%1-4] ; ___E FGHI ___J KLMN ___O PQRS ___T UVWX
pshufb m0, [load_bytes_shuf] ; EFGH JKLM FGHI KLMN OPQR TUVW PQRS UVWX
mova m2, [insert_top_shuf]
vpermq m1, m0, q3131 ; FGHI KLMN PQRS UVWX x2
vpermd m0, m2, m0 ; EFGH JKLM OPQR TUVW ____ FGHI KLMN PQRS
vpbroadcastd m2, [%1-8] ; ABCD ....
vpblendd m0, m0, m2, 00010000b ; EFGH JKLM OPQR TUVW ABCD FGHI KLMN PQRS
movu m0, [%1-4] ; ___E FGHI ___J KLMN ___O PQRS ___T UVWX
pshufb m0, m6 ; EFGH JKLM FGHI KLMN OPQR TUVW PQRS UVWX
vpermq m1, m0, q3131 ; FGHI KLMN PQRS UVWX x2
vpbroadcastd m2, [%1-8] ; ABCD ....
vpblendd m0, m0, m2, 0x80
vpermd m0, m7, m0 ; EFGH JKLM OPQR TUVW ABCD FGHI KLMN PQRS
%endmacro
INIT_YMM avx2
cglobal deblock_strength, 6,6,7
cglobal deblock_strength, 5,5,8
mova m6, [load_bytes_ymm_shuf]
; Prepare mv comparison register
shl r4d, 8
add r4d, 3 - (1<<8)
movd xm6, r4d
vpbroadcastw m6, xm6
pxor m5, m5 ; bs0,bs1
shl r4d, 8
add r4d, 3 - (1<<8)
movd xm5, r4d
movifnidn t0d, r5m
vpbroadcastw m5, xm5
psrld m7, m6, 4
pxor m4, m4 ; bs0,bs1
.lists:
; Check refs
LOAD_BYTES_YMM ref
pxor m0, m1
por m5, m0
pxor m0, m1
por m4, m0
; Check mvs
movu xm0, [mv-4+4*8*0]
vinserti128 m0, m0, [mv+4*8*-1], 1
vbroadcasti128 m2, [mv+4*8* 0]
vinserti128 m1, m2, [mv-4+4*8*1], 0
vbroadcasti128 m3, [mv+4*8* 1]
psubw m0, m2
psubw m1, m3
vinserti128 m2, m3, [mv-4+4*8*2], 0
vbroadcasti128 m4, [mv+4*8* 2]
vinserti128 m3, m4, [mv-4+4*8*3], 0
psubw m2, m4
vbroadcasti128 m4, [mv+4*8* 3]
psubw m3, m4
packsswb m0, m1
packsswb m2, m3
pabsb m0, m0
pabsb m2, m2
psubusb m0, m6
psubusb m2, m6
packsswb m0, m2
por m5, m0
add r1, 40
add r2, 4*8*5
dec r5d
movu xm0, [mv+0*4*8-4]
vinserti128 m0, m0, [mv-1*4*8 ], 1
vbroadcasti128 m2, [mv+0*4*8 ]
vinserti128 m1, m2, [mv+1*4*8-4], 0
psubw m0, m2
vbroadcasti128 m2, [mv+1*4*8 ]
psubw m1, m2
packsswb m0, m1
vinserti128 m1, m2, [mv+2*4*8-4], 0
vbroadcasti128 m3, [mv+2*4*8 ]
vinserti128 m2, m3, [mv+3*4*8-4], 0
psubw m1, m3
vbroadcasti128 m3, [mv+3*4*8 ]
psubw m2, m3
packsswb m1, m2
pabsb m0, m0
pabsb m1, m1
psubusb m0, m5
psubusb m1, m5
packsswb m0, m1
por m4, m0
add r1, 40
add r2, 4*8*5
dec t0d
jge .lists
; Check nnz
LOAD_BYTES_YMM nnz
por m0, m1
mova m6, [pb_1]
pminub m0, m6
pminub m5, m6 ; mv ? 1 : 0
paddb m0, m0 ; nnz ? 2 : 0
pmaxub m5, m0
vextracti128 [bs1], m5, 1
pshufb xm5, [transpose_shuf]
mova [bs0], xm5
mova m2, [pb_1]
por m0, m1
pminub m0, m2
pminub m4, m2 ; mv ? 1 : 0
paddb m0, m0 ; nnz ? 2 : 0
pmaxub m0, m4
vextracti128 [bs1], m0, 1
pshufb xm0, [transpose_shuf]
mova [bs0], xm0
RET
%macro LOAD_BYTES_ZMM 1
vpermd m1, m6, [%1-12]
pshufb m1, m7 ; EF FG GH HI JK KL LM MN OP PQ QR RS TU UV VW WX
%endmacro ; AF BG CH DI FK GL HM IN KP LQ MR NS PU QV RW SX
INIT_ZMM avx512
cglobal deblock_strength, 5,5
mova m6, [load_bytes_zmm_shuf]
shl r4d, 8
add r4d, 3 - (1<<8)
vpbroadcastw m5, r4d
mov r4d, 0x34cc34cc ; {1,-1} * 11001100b
kmovb k1, r4d
vpbroadcastd m4, r4d
movifnidn t0d, r5m
psrld m7, m6, 4
pxor xm3, xm3
.lists:
vbroadcasti64x2 m2, [mv+32]
vinserti64x2 m0, m2, [mv-32], 2
vbroadcasti64x2 m1, [mv+ 0]
vinserti64x2 m0, m0, [mv- 4], 0
vbroadcasti64x2 m1 {k1}, [mv+64]
vinserti64x2 m0, m0, [mv+60], 1
psubw m0, m1
vinserti64x2 m1, m1, [mv+28], 0
vbroadcasti64x2 m2 {k1}, [mv+96]
vinserti64x2 m1, m1, [mv+92], 1
psubw m1, m2
packsswb m0, m1
pabsb m0, m0
psubusb m0, m5
LOAD_BYTES_ZMM ref
pmaddubsw m1, m4 ; E-F F-G G-H H-I ...
vpternlogd m3, m0, m1, 0xfe ; m3 | m0 | m1
add r1, 40
add r2, 4*8*5
dec t0d
jge .lists
LOAD_BYTES_ZMM nnz
mova ym2, [pb_1]
vptestmw k1, m1, m1
vptestmw k2, m3, m3
vpaddb ym0 {k1}{z}, ym2, ym2 ; nnz ? 2 : 0
vpmaxub ym0 {k2}, ym2 ; mv ? 1 : 0
vextracti128 [bs1], ym0, 1
pshufb xm0, [transpose_shuf]
mova [bs0], xm0
RET
tools/checkasm.c
View file @ 2eceefe8
......@@ -1961,9 +1961,12 @@ static int check_deblock( int cpu_ref, int cpu_new )
if( db_a.deblock_strength != db_ref.deblock_strength )
{
set_func_name( "deblock_strength" );
used_asm = 1;
for( int i = 0; i < 100; i++ )
{
ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
ALIGNED_ARRAY_16( uint8_t, nnz_buf, [X264_SCAN8_SIZE+8] );
uint8_t *nnz = &nnz_buf[8];
ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
ALIGNED_ARRAY_32( uint8_t, bs, [2],[2][8][4] );
......@@ -1977,7 +1980,6 @@ static int check_deblock( int cpu_ref, int cpu_new )
for( int l = 0; l < 2; l++ )
mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
}
set_func_name( "deblock_strength" );
call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1) );
call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1) );
if( memcmp( bs[0], bs[1], sizeof(uint8_t)*2*4*8 ) )
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment