19 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
20 |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
21 |
; * |
; * |
22 |
; * $Id: sad_3dn.asm,v 1.8 2004-08-22 11:46:10 edgomez Exp $ |
; * $Id: sad_3dn.asm,v 1.12 2008-11-26 01:04:34 Isibaar Exp $ |
23 |
; * |
; * |
24 |
; ***************************************************************************/ |
; ***************************************************************************/ |
25 |
|
|
26 |
BITS 32 |
%include "nasm.inc" |
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
%ifdef MARK_FUNCS |
|
|
global _%1:function |
|
|
%define %1 _%1:function |
|
|
%else |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%endif |
|
|
%else |
|
|
%ifdef MARK_FUNCS |
|
|
global %1:function |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endif |
|
|
%endmacro |
|
27 |
|
|
28 |
;============================================================================= |
;============================================================================= |
29 |
; Read only data |
; Read only data |
30 |
;============================================================================= |
;============================================================================= |
31 |
|
|
32 |
%ifdef FORMAT_COFF |
DATA |
|
SECTION .rodata |
|
|
%else |
|
|
SECTION .rodata align=16 |
|
|
%endif |
|
33 |
|
|
34 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
35 |
mmx_one: |
mmx_one: |
36 |
times 4 dw 1 |
times 4 dw 1 |
37 |
|
|
39 |
; Helper macros |
; Helper macros |
40 |
;============================================================================= |
;============================================================================= |
41 |
%macro SADBI_16x16_3DN 0 |
%macro SADBI_16x16_3DN 0 |
42 |
movq mm0, [eax] ; src |
movq mm0, [_EAX] ; src |
43 |
movq mm2, [eax+8] |
movq mm2, [_EAX+8] |
44 |
|
|
45 |
movq mm1, [edx] ; ref1 |
movq mm1, [TMP1] ; ref1 |
46 |
movq mm3, [edx+8] |
movq mm3, [TMP1+8] |
47 |
pavgusb mm1, [ebx] ; ref2 |
pavgusb mm1, [_EBX] ; ref2 |
48 |
lea edx, [edx+ecx] |
lea TMP1, [TMP1+TMP0] |
49 |
pavgusb mm3, [ebx+8] |
pavgusb mm3, [_EBX+8] |
50 |
lea ebx, [ebx+ecx] |
lea _EBX, [_EBX+TMP0] |
51 |
|
|
52 |
movq mm4, mm0 |
movq mm4, mm0 |
53 |
lea eax, [eax+ecx] |
lea _EAX, [_EAX+TMP0] |
54 |
psubusb mm0, mm1 |
psubusb mm0, mm1 |
55 |
movq mm5, mm2 |
movq mm5, mm2 |
56 |
psubusb mm2, mm3 |
psubusb mm2, mm3 |
75 |
%endmacro |
%endmacro |
76 |
|
|
77 |
%macro SADBI_8x8_3DN 0 |
%macro SADBI_8x8_3DN 0 |
78 |
movq mm0, [eax] ; src |
movq mm0, [_EAX] ; src |
79 |
movq mm2, [eax+ecx] |
movq mm2, [_EAX+TMP0] |
80 |
|
|
81 |
movq mm1, [edx] ; ref1 |
movq mm1, [TMP1] ; ref1 |
82 |
movq mm3, [edx+ecx] |
movq mm3, [TMP1+TMP0] |
83 |
pavgusb mm1, [ebx] ; ref2 |
pavgusb mm1, [_EBX] ; ref2 |
84 |
lea edx, [edx+2*ecx] |
lea TMP1, [TMP1+2*TMP0] |
85 |
pavgusb mm3, [ebx+ecx] |
pavgusb mm3, [_EBX+TMP0] |
86 |
lea ebx, [ebx+2*ecx] |
lea _EBX, [_EBX+2*TMP0] |
87 |
|
|
88 |
movq mm4, mm0 |
movq mm4, mm0 |
89 |
lea eax, [eax+2*ecx] |
lea _EAX, [_EAX+2*TMP0] |
90 |
psubusb mm0, mm1 |
psubusb mm0, mm1 |
91 |
movq mm5, mm2 |
movq mm5, mm2 |
92 |
psubusb mm2, mm3 |
psubusb mm2, mm3 |
114 |
; Code |
; Code |
115 |
;============================================================================= |
;============================================================================= |
116 |
|
|
117 |
SECTION .text |
SECTION .rotext align=SECTION_ALIGN |
118 |
|
|
119 |
cglobal sad16bi_3dn |
cglobal sad16bi_3dn |
120 |
cglobal sad8bi_3dn |
cglobal sad8bi_3dn |
128 |
; |
; |
129 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
130 |
|
|
131 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
132 |
sad16bi_3dn: |
sad16bi_3dn: |
133 |
push ebx |
mov _EAX, prm1 ; Src |
134 |
mov eax, [esp+4+ 4] ; Src |
mov TMP1, prm2 ; Ref1 |
135 |
mov edx, [esp+4+ 8] ; Ref1 |
mov TMP0, prm4 ; Stride |
136 |
mov ebx, [esp+4+12] ; Ref2 |
|
137 |
mov ecx, [esp+4+16] ; Stride |
push _EBX |
138 |
|
%ifdef ARCH_IS_X86_64 |
139 |
|
mov _EBX, prm3 |
140 |
|
%else |
141 |
|
mov _EBX, [_ESP+4+12] ; Ref2 |
142 |
|
%endif |
143 |
|
|
144 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
145 |
pxor mm7, mm7 |
pxor mm7, mm7 |
146 |
.Loop |
.Loop: |
147 |
SADBI_16x16_3DN |
SADBI_16x16_3DN |
148 |
SADBI_16x16_3DN |
SADBI_16x16_3DN |
149 |
SADBI_16x16_3DN |
SADBI_16x16_3DN |
169 |
|
|
170 |
movd eax, mm6 |
movd eax, mm6 |
171 |
|
|
172 |
pop ebx |
pop _EBX |
173 |
|
|
174 |
ret |
ret |
175 |
|
ENDFUNC |
176 |
|
|
177 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
178 |
; |
; |
183 |
; |
; |
184 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
185 |
|
|
186 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
187 |
sad8bi_3dn: |
sad8bi_3dn: |
188 |
push ebx |
mov _EAX, prm1 ; Src |
189 |
mov eax, [esp+4+ 4] ; Src |
mov TMP1, prm2 ; Ref1 |
190 |
mov edx, [esp+4+ 8] ; Ref1 |
mov TMP0, prm4 ; Stride |
191 |
mov ebx, [esp+4+12] ; Ref2 |
|
192 |
mov ecx, [esp+4+16] ; Stride |
push _EBX |
193 |
|
%ifdef ARCH_IS_X86_64 |
194 |
|
mov _EBX, prm3 |
195 |
|
%else |
196 |
|
mov _EBX, [_ESP+4+12] ; Ref2 |
197 |
|
%endif |
198 |
|
|
199 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
200 |
pxor mm7, mm7 |
pxor mm7, mm7 |
201 |
.Loop |
.Loop: |
202 |
SADBI_8x8_3DN |
SADBI_8x8_3DN |
203 |
SADBI_8x8_3DN |
SADBI_8x8_3DN |
204 |
SADBI_8x8_3DN |
SADBI_8x8_3DN |
211 |
|
|
212 |
movd eax, mm6 |
movd eax, mm6 |
213 |
|
|
214 |
pop ebx |
pop _EBX |
215 |
|
|
216 |
ret |
ret |
217 |
|
ENDFUNC |
218 |
|
|
219 |
|
|
220 |
|
%ifidn __OUTPUT_FORMAT__,elf |
221 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
222 |
|
%endif |
223 |
|
|