| 1 |
;/************************************************************************** |
;/**************************************************************************** |
| 2 |
; * |
; * |
| 3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
| 4 |
; * sse2 sum of absolute difference |
; * - SSE2 optimized SAD operators - |
| 5 |
; * |
; * |
| 6 |
; * This program is free software; you can redistribute it and/or modify |
; * Copyright(C) 2003 Pascal Massimino <skal@planet-d.net> |
| 7 |
; * it under the terms of the GNU General Public License as published by |
; * |
| 8 |
|
; * |
| 9 |
|
; * This program is free software; you can redistribute it and/or modify it |
| 10 |
|
; * under the terms of the GNU General Public License as published by |
| 11 |
; * the Free Software Foundation; either version 2 of the License, or |
; * the Free Software Foundation; either version 2 of the License, or |
| 12 |
; * (at your option) any later version. |
; * (at your option) any later version. |
| 13 |
; * |
; * |
| 18 |
; * |
; * |
| 19 |
; * You should have received a copy of the GNU General Public License |
; * You should have received a copy of the GNU General Public License |
| 20 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
| 21 |
; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
;/************************************************************************** |
|
|
; * |
|
|
; * History: |
|
| 22 |
; * |
; * |
| 23 |
; * 24.05.2002 inital version; (c)2002 Dmitry Rozhdestvensky |
; * $Id: sad_sse2.asm,v 1.17 2008-12-01 14:45:45 Isibaar Exp $ |
| 24 |
; * |
; * |
| 25 |
; *************************************************************************/ |
; ***************************************************************************/ |
| 26 |
|
|
| 27 |
bits 32 |
%include "nasm.inc" |
| 28 |
|
|
| 29 |
%macro cglobal 1 |
;============================================================================= |
| 30 |
%ifdef PREFIX |
; Read only data |
| 31 |
global _%1 |
;============================================================================= |
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
| 32 |
|
|
| 33 |
%define sad_debug 0 ;1=unaligned 2=ref unaligned 3=aligned 0=autodetect |
DATA |
|
%define dev_debug 2 ;1=unaligned 2=aligned 0=autodetect |
|
|
%define test_stride_alignment 0 ;test stride for alignment while autodetect |
|
|
%define early_return 0 ;use early return in sad |
|
| 34 |
|
|
| 35 |
section .data |
ALIGN SECTION_ALIGN |
|
|
|
|
align 64 |
|
|
buffer times 4*8 dd 0 ;8 128-bit words |
|
| 36 |
zero times 4 dd 0 |
zero times 4 dd 0 |
| 37 |
|
|
| 38 |
section .text |
;============================================================================= |
| 39 |
|
; Code |
| 40 |
|
;============================================================================= |
| 41 |
|
|
| 42 |
|
SECTION .rotext align=SECTION_ALIGN |
| 43 |
|
|
| 44 |
cglobal sad16_sse2 |
cglobal sad16_sse2 |
| 45 |
cglobal dev16_sse2 |
cglobal dev16_sse2 |
| 46 |
|
|
| 47 |
;=========================================================================== |
cglobal sad16_sse3 |
| 48 |
; General macros for SSE2 code |
cglobal dev16_sse3 |
|
;=========================================================================== |
|
|
|
|
|
%macro load_stride 1 |
|
|
mov ecx,%1 |
|
|
add ecx,ecx |
|
|
mov edx,ecx |
|
|
add ecx,%1 ;stride*3 |
|
|
add edx,edx ;stride*4 |
|
|
%endmacro |
|
|
|
|
|
%macro sad8lines 1 |
|
|
|
|
|
psadbw xmm0,[%1] |
|
|
psadbw xmm1,[%1+ebx] |
|
|
psadbw xmm2,[%1+ebx*2] |
|
|
psadbw xmm3,[%1+ecx] |
|
|
|
|
|
add %1,edx |
|
|
|
|
|
psadbw xmm4,[%1] |
|
|
psadbw xmm5,[%1+ebx] |
|
|
psadbw xmm6,[%1+ebx*2] |
|
|
psadbw xmm7,[%1+ecx] |
|
|
|
|
|
add %1,edx |
|
|
%endmacro |
|
|
|
|
|
%macro after_sad 1 ; Summarizes 0th and 4th words of all xmm registers |
|
| 49 |
|
|
| 50 |
paddusw xmm0,xmm1 |
;----------------------------------------------------------------------------- |
| 51 |
paddusw xmm2,xmm3 |
; uint32_t sad16_sse2 (const uint8_t * const cur, <- assumed aligned! |
|
paddusw xmm4,xmm5 |
|
|
paddusw xmm6,xmm7 |
|
|
|
|
|
paddusw xmm0,xmm2 |
|
|
paddusw xmm4,xmm6 |
|
|
|
|
|
paddusw xmm4,xmm0 |
|
|
pshufd xmm5,xmm4,11111110b |
|
|
paddusw xmm5,xmm4 |
|
|
|
|
|
pextrw %1,xmm5,0 ;less latency then movd |
|
|
%endmacro |
|
|
|
|
|
%macro restore 1 ;restores used registers |
|
|
|
|
|
%if %1=1 |
|
|
pop ebp |
|
|
%endif |
|
|
pop edi |
|
|
pop esi |
|
|
pop ebx |
|
|
%endmacro |
|
|
|
|
|
;=========================================================================== |
|
|
; |
|
|
; uint32_t sad16_sse2 (const uint8_t * const cur, |
|
| 52 |
; const uint8_t * const ref, |
; const uint8_t * const ref, |
| 53 |
; const uint32_t stride, |
; const uint32_t stride, |
| 54 |
; const uint32_t best_sad); |
; const uint32_t /*ignored*/); |
| 55 |
; |
;----------------------------------------------------------------------------- |
|
; |
|
|
;=========================================================================== |
|
|
|
|
|
align 16 |
|
|
sad16_sse2 |
|
|
push ebx |
|
|
push esi |
|
|
push edi |
|
|
|
|
|
mov ebx,[esp + 3*4 + 12] ;stride |
|
|
|
|
|
%if sad_debug<>0 |
|
|
mov edi,[esp + 3*4 + 4] |
|
|
cglobal sad16_sse2 |
|
|
mov esi,[esp + 3*4 + 8] |
|
|
%endif |
|
|
|
|
|
%if sad_debug=1 |
|
|
jmp sad16_sse2_ul |
|
|
%endif |
|
|
%if sad_debug=2 |
|
|
jmp sad16_sse2_semial |
|
|
%endif |
|
|
%if sad_debug=3 |
|
|
jmp sad16_sse2_al |
|
|
%endif |
|
|
|
|
|
%if test_stride_alignment<>0 |
|
|
test ebx,15 |
|
|
jnz sad16_sse2_ul |
|
|
%endif |
|
|
mov edi,[esp + 3*4 + 4] ;cur (most likely aligned) |
|
|
|
|
|
test edi,15 |
|
|
cmovz esi,[esp + 3*4 + 8] ;load esi if edi is aligned |
|
|
cmovnz esi,edi ;move to esi and load edi |
|
|
cmovnz edi,[esp + 3*4 + 8] ;if not |
|
|
jnz esi_unaligned |
|
|
|
|
|
test esi,15 |
|
|
jnz near sad16_sse2_semial |
|
|
jmp sad16_sse2_al |
|
|
|
|
|
esi_unaligned: test edi,15 |
|
|
jnz near sad16_sse2_ul |
|
|
jmp sad16_sse2_semial |
|
|
|
|
|
;=========================================================================== |
|
|
; Branch requires 16-byte alignment of esi and edi and stride |
|
|
;=========================================================================== |
|
|
|
|
|
%macro sad16x8_al 1 |
|
|
|
|
|
movdqa xmm0,[esi] |
|
|
movdqa xmm1,[esi+ebx] |
|
|
movdqa xmm2,[esi+ebx*2] |
|
|
movdqa xmm3,[esi+ecx] |
|
|
|
|
|
add esi,edx |
|
|
|
|
|
movdqa xmm4,[esi] |
|
|
movdqa xmm5,[esi+ebx] |
|
|
movdqa xmm6,[esi+ebx*2] |
|
|
movdqa xmm7,[esi+ecx] |
|
|
|
|
|
add esi,edx |
|
|
|
|
|
sad8lines edi |
|
|
|
|
|
after_sad %1 |
|
|
|
|
|
%endmacro |
|
|
|
|
|
align 16 |
|
|
sad16_sse2_al |
|
|
|
|
|
load_stride ebx |
|
|
|
|
|
sad16x8_al eax |
|
|
|
|
|
%if early_return=1 |
|
|
cmp eax,[esp + 3*4 + 16] ;best_sad |
|
|
jg continue_al |
|
|
%endif |
|
|
|
|
|
sad16x8_al ebx |
|
|
|
|
|
add eax,ebx |
|
|
|
|
|
continue_al: restore 0 |
|
|
|
|
|
ret |
|
|
|
|
|
;=========================================================================== |
|
|
; Branch requires 16-byte alignment of the edi and stride only |
|
|
;=========================================================================== |
|
|
|
|
|
%macro sad16x8_semial 1 |
|
|
|
|
|
movdqu xmm0,[esi] |
|
|
movdqu xmm1,[esi+ebx] |
|
|
movdqu xmm2,[esi+ebx*2] |
|
|
movdqu xmm3,[esi+ecx] |
|
|
|
|
|
add esi,edx |
|
|
|
|
|
movdqu xmm4,[esi] |
|
|
movdqu xmm5,[esi+ebx] |
|
|
movdqu xmm6,[esi+ebx*2] |
|
|
movdqu xmm7,[esi+ecx] |
|
|
|
|
|
add esi,edx |
|
|
|
|
|
sad8lines edi |
|
| 56 |
|
|
|
after_sad %1 |
|
| 57 |
|
|
| 58 |
%endmacro |
%macro SAD_16x16_SSE2 1 |
| 59 |
|
%1 xmm0, [TMP1] |
| 60 |
align 16 |
%1 xmm1, [TMP1+TMP0] |
| 61 |
sad16_sse2_semial |
lea TMP1,[TMP1+2*TMP0] |
| 62 |
|
movdqa xmm2, [_EAX] |
| 63 |
load_stride ebx |
movdqa xmm3, [_EAX+TMP0] |
| 64 |
|
lea _EAX,[_EAX+2*TMP0] |
| 65 |
sad16x8_semial eax |
psadbw xmm0, xmm2 |
| 66 |
|
paddusw xmm6,xmm0 |
| 67 |
%if early_return=1 |
psadbw xmm1, xmm3 |
| 68 |
cmp eax,[esp + 3*4 + 16] ;best_sad |
paddusw xmm6,xmm1 |
| 69 |
jg cont_semial |
%endmacro |
| 70 |
%endif |
|
| 71 |
|
%macro SAD16_SSE2_SSE3 1 |
| 72 |
sad16x8_semial ebx |
PUSH_XMM6_XMM7 |
| 73 |
|
mov _EAX, prm1 ; cur (assumed aligned) |
| 74 |
add eax,ebx |
mov TMP1, prm2 ; ref |
| 75 |
|
mov TMP0, prm3 ; stride |
| 76 |
cont_semial: restore 0 |
|
| 77 |
|
pxor xmm6, xmm6 ; accum |
| 78 |
|
|
| 79 |
|
SAD_16x16_SSE2 %1 |
| 80 |
|
SAD_16x16_SSE2 %1 |
| 81 |
|
SAD_16x16_SSE2 %1 |
| 82 |
|
SAD_16x16_SSE2 %1 |
| 83 |
|
SAD_16x16_SSE2 %1 |
| 84 |
|
SAD_16x16_SSE2 %1 |
| 85 |
|
SAD_16x16_SSE2 %1 |
| 86 |
|
SAD_16x16_SSE2 %1 |
| 87 |
|
|
| 88 |
|
pshufd xmm5, xmm6, 00000010b |
| 89 |
|
paddusw xmm6, xmm5 |
| 90 |
|
pextrw eax, xmm6, 0 |
| 91 |
|
|
| 92 |
|
POP_XMM6_XMM7 |
| 93 |
ret |
ret |
|
|
|
|
|
|
|
;=========================================================================== |
|
|
; Branch does not require alignment, even stride |
|
|
;=========================================================================== |
|
|
|
|
|
%macro sad16x4_ul 1 |
|
|
|
|
|
movdqu xmm0,[esi] |
|
|
movdqu xmm1,[esi+ebx] |
|
|
movdqu xmm2,[esi+ebx*2] |
|
|
movdqu xmm3,[esi+ecx] |
|
|
|
|
|
add esi,edx |
|
|
|
|
|
movdqu xmm4,[edi] |
|
|
movdqu xmm5,[edi+ebx] |
|
|
movdqu xmm6,[edi+ebx*2] |
|
|
movdqu xmm7,[edi+ecx] |
|
|
|
|
|
add edi,edx |
|
|
|
|
|
psadbw xmm4,xmm0 |
|
|
psadbw xmm5,xmm1 |
|
|
psadbw xmm6,xmm2 |
|
|
psadbw xmm7,xmm3 |
|
|
|
|
|
paddusw xmm4,xmm5 |
|
|
paddusw xmm6,xmm7 |
|
|
|
|
|
paddusw xmm4,xmm6 |
|
|
pshufd xmm7,xmm4,11111110b |
|
|
paddusw xmm7,xmm4 |
|
|
|
|
|
pextrw %1,xmm7,0 |
|
| 94 |
%endmacro |
%endmacro |
| 95 |
|
|
| 96 |
|
ALIGN SECTION_ALIGN |
| 97 |
align 16 |
sad16_sse2: |
| 98 |
sad16_sse2_ul |
SAD16_SSE2_SSE3 movdqu |
| 99 |
|
ENDFUNC |
| 100 |
load_stride ebx |
|
| 101 |
|
|
| 102 |
push ebp |
ALIGN SECTION_ALIGN |
| 103 |
|
sad16_sse3: |
| 104 |
sad16x4_ul eax |
SAD16_SSE2_SSE3 lddqu |
| 105 |
|
ENDFUNC |
| 106 |
%if early_return=1 |
|
| 107 |
cmp eax,[esp + 4*4 + 16] ;best_sad |
|
| 108 |
jg continue_ul |
;----------------------------------------------------------------------------- |
| 109 |
%endif |
; uint32_t dev16_sse2(const uint8_t * const cur, const uint32_t stride); |
| 110 |
|
;----------------------------------------------------------------------------- |
| 111 |
sad16x4_ul ebp |
|
| 112 |
add eax,ebp |
%macro MEAN_16x16_SSE2 1 ; _EAX: src, TMP0:stride, mm7: zero or mean => mm6: result |
| 113 |
|
%1 xmm0, [_EAX] |
| 114 |
%if early_return=1 |
%1 xmm1, [_EAX+TMP0] |
| 115 |
cmp eax,[esp + 4*4 + 16] ;best_sad |
lea _EAX, [_EAX+2*TMP0] ; + 2*stride |
| 116 |
jg continue_ul |
psadbw xmm0, xmm7 |
| 117 |
%endif |
paddusw xmm6, xmm0 |
| 118 |
|
psadbw xmm1, xmm7 |
| 119 |
sad16x4_ul ebp |
paddusw xmm6, xmm1 |
|
add eax,ebp |
|
|
|
|
|
%if early_return=1 |
|
|
cmp eax,[esp + 4*4 + 16] ;best_sad |
|
|
jg continue_ul |
|
|
%endif |
|
|
|
|
|
sad16x4_ul ebp |
|
|
add eax,ebp |
|
|
|
|
|
continue_ul: restore 1 |
|
|
|
|
|
ret |
|
|
|
|
|
;=========================================================================== |
|
|
; |
|
|
; uint32_t dev16_sse2(const uint8_t * const cur, |
|
|
; const uint32_t stride); |
|
|
; |
|
|
; experimental! |
|
|
; |
|
|
;=========================================================================== |
|
|
|
|
|
align 16 |
|
|
dev16_sse2 |
|
|
|
|
|
push ebx |
|
|
push esi |
|
|
push edi |
|
|
push ebp |
|
|
|
|
|
mov esi, [esp + 4*4 + 4] ; cur |
|
|
mov ebx, [esp + 4*4 + 8] ; stride |
|
|
cglobal dev16_sse2 |
|
|
mov edi, buffer |
|
|
|
|
|
%if dev_debug=1 |
|
|
jmp dev16_sse2_ul |
|
|
%endif |
|
|
|
|
|
%if dev_debug=2 |
|
|
jmp dev16_sse2_al |
|
|
%endif |
|
|
|
|
|
test esi,15 |
|
|
jnz near dev16_sse2_ul |
|
|
|
|
|
%if test_stride_alignment=1 |
|
|
test ebx,15 |
|
|
jnz dev16_sse2_ul |
|
|
%endif |
|
|
|
|
|
mov edi,esi |
|
|
jmp dev16_sse2_al |
|
|
|
|
|
;=========================================================================== |
|
|
; Branch requires alignment of both the cur and stride |
|
|
;=========================================================================== |
|
|
|
|
|
%macro make_mean 0 |
|
|
add eax,ebp ;mean 16-bit |
|
|
mov al,ah ;eax= {0 0 mean/256 mean/256} |
|
|
mov ebp,eax |
|
|
shl ebp,16 |
|
|
or eax,ebp |
|
| 120 |
%endmacro |
%endmacro |
| 121 |
|
|
|
%macro sad_mean16x8_al 3 ;destination,0=zero,1=mean from eax,source |
|
| 122 |
|
|
| 123 |
%if %2=0 |
%macro MEAN16_SSE2_SSE3 1 |
| 124 |
pxor xmm0,xmm0 |
PUSH_XMM6_XMM7 |
| 125 |
%else |
mov _EAX, prm1 ; src |
| 126 |
movd xmm0,eax |
mov TMP0, prm2 ; stride |
|
pshufd xmm0,xmm0,0 |
|
|
%endif |
|
|
movdqa xmm1,xmm0 |
|
|
movdqa xmm2,xmm0 |
|
|
movdqa xmm3,xmm0 |
|
|
movdqa xmm4,xmm0 |
|
|
movdqa xmm5,xmm0 |
|
|
movdqa xmm6,xmm0 |
|
|
movdqa xmm7,xmm0 |
|
|
|
|
|
sad8lines %3 |
|
| 127 |
|
|
| 128 |
after_sad %1 |
pxor xmm6, xmm6 ; accum |
| 129 |
|
pxor xmm7, xmm7 ; zero |
|
%endmacro |
|
| 130 |
|
|
| 131 |
align 16 |
MEAN_16x16_SSE2 %1 |
| 132 |
dev16_sse2_al |
MEAN_16x16_SSE2 %1 |
| 133 |
|
MEAN_16x16_SSE2 %1 |
| 134 |
|
MEAN_16x16_SSE2 %1 |
| 135 |
|
|
| 136 |
load_stride ebx |
MEAN_16x16_SSE2 %1 |
| 137 |
|
MEAN_16x16_SSE2 %1 |
| 138 |
|
MEAN_16x16_SSE2 %1 |
| 139 |
|
MEAN_16x16_SSE2 %1 |
| 140 |
|
|
| 141 |
sad_mean16x8_al eax,0,esi |
mov _EAX, prm1 ; src again |
|
sad_mean16x8_al ebp,0,esi |
|
| 142 |
|
|
| 143 |
make_mean |
pshufd xmm7, xmm6, 10b |
| 144 |
|
paddusw xmm7, xmm6 |
| 145 |
|
pxor xmm6, xmm6 ; zero accum |
| 146 |
|
psrlw xmm7, 8 ; => Mean |
| 147 |
|
pshuflw xmm7, xmm7, 0 ; replicate Mean |
| 148 |
|
packuswb xmm7, xmm7 |
| 149 |
|
pshufd xmm7, xmm7, 00000000b |
| 150 |
|
|
| 151 |
sad_mean16x8_al ebp,1,edi |
MEAN_16x16_SSE2 %1 |
| 152 |
sad_mean16x8_al eax,1,edi |
MEAN_16x16_SSE2 %1 |
| 153 |
|
MEAN_16x16_SSE2 %1 |
| 154 |
|
MEAN_16x16_SSE2 %1 |
| 155 |
|
|
| 156 |
add eax,ebp |
MEAN_16x16_SSE2 %1 |
| 157 |
|
MEAN_16x16_SSE2 %1 |
| 158 |
|
MEAN_16x16_SSE2 %1 |
| 159 |
|
MEAN_16x16_SSE2 %1 |
| 160 |
|
|
| 161 |
restore 1 |
pshufd xmm7, xmm6, 10b |
| 162 |
|
paddusw xmm7, xmm6 |
| 163 |
|
pextrw eax, xmm7, 0 |
| 164 |
|
|
| 165 |
|
POP_XMM6_XMM7 |
| 166 |
ret |
ret |
|
|
|
|
;=========================================================================== |
|
|
; Branch does not require alignment |
|
|
;=========================================================================== |
|
|
|
|
|
%macro sad_mean16x8_ul 2 |
|
|
|
|
|
pxor xmm7,xmm7 |
|
|
|
|
|
movdqu xmm0,[%1] |
|
|
movdqu xmm1,[%1+ebx] |
|
|
movdqu xmm2,[%1+ebx*2] |
|
|
movdqu xmm3,[%1+ecx] |
|
|
|
|
|
add %1,edx |
|
|
|
|
|
movdqa [buffer+16*0],xmm0 |
|
|
movdqa [buffer+16*1],xmm1 |
|
|
movdqa [buffer+16*2],xmm2 |
|
|
movdqa [buffer+16*3],xmm3 |
|
|
|
|
|
movdqu xmm4,[%1] |
|
|
movdqu xmm5,[%1+ebx] |
|
|
movdqu xmm6,[%1+ebx*2] |
|
|
movdqa [buffer+16*4],xmm4 |
|
|
movdqa [buffer+16*5],xmm5 |
|
|
movdqa [buffer+16*6],xmm6 |
|
|
|
|
|
psadbw xmm0,xmm7 |
|
|
psadbw xmm1,xmm7 |
|
|
psadbw xmm2,xmm7 |
|
|
psadbw xmm3,xmm7 |
|
|
psadbw xmm4,xmm7 |
|
|
psadbw xmm5,xmm7 |
|
|
psadbw xmm6,xmm7 |
|
|
|
|
|
movdqu xmm7,[%1+ecx] |
|
|
movdqa [buffer+16*7],xmm7 |
|
|
psadbw xmm7,[zero] |
|
|
|
|
|
add %1,edx |
|
|
|
|
|
after_sad %2 |
|
| 167 |
%endmacro |
%endmacro |
| 168 |
|
|
| 169 |
align 16 |
ALIGN SECTION_ALIGN |
| 170 |
dev16_sse2_ul |
dev16_sse2: |
| 171 |
|
MEAN16_SSE2_SSE3 movdqu |
| 172 |
load_stride ebx |
ENDFUNC |
| 173 |
|
|
| 174 |
sad_mean16x8_ul esi,eax |
ALIGN SECTION_ALIGN |
| 175 |
sad_mean16x8_ul esi,ebp |
dev16_sse3: |
| 176 |
|
MEAN16_SSE2_SSE3 lddqu |
| 177 |
|
ENDFUNC |
| 178 |
|
|
|
make_mean |
|
| 179 |
|
|
| 180 |
sad_mean16x8_al ebp,1,edi |
%ifidn __OUTPUT_FORMAT__,elf |
| 181 |
sad_mean16x8_al eax,1,edi |
section ".note.GNU-stack" noalloc noexec nowrite progbits |
| 182 |
|
%endif |
|
add eax,ebp |
|
|
|
|
|
restore 1 |
|
| 183 |
|
|
|
ret |
|
Parent Directory
| 
Revision Log
| 
Patch