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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
22 |
; * |
; * |
23 |
; * $Id: sad_sse2.asm,v 1.9 2004-03-22 22:36:24 edgomez Exp $ |
; * $Id: sad_sse2.asm,v 1.15 2008-11-11 20:46:24 Isibaar Exp $ |
24 |
; * |
; * |
25 |
; ***************************************************************************/ |
; ***************************************************************************/ |
26 |
|
|
28 |
|
|
29 |
%macro cglobal 1 |
%macro cglobal 1 |
30 |
%ifdef PREFIX |
%ifdef PREFIX |
31 |
|
%ifdef MARK_FUNCS |
32 |
|
global _%1:function %1.endfunc-%1 |
33 |
|
%define %1 _%1:function %1.endfunc-%1 |
34 |
|
%define ENDFUNC .endfunc |
35 |
|
%else |
36 |
global _%1 |
global _%1 |
37 |
%define %1 _%1 |
%define %1 _%1 |
38 |
|
%define ENDFUNC |
39 |
|
%endif |
40 |
|
%else |
41 |
|
%ifdef MARK_FUNCS |
42 |
|
global %1:function %1.endfunc-%1 |
43 |
|
%define ENDFUNC .endfunc |
44 |
%else |
%else |
45 |
global %1 |
global %1 |
46 |
|
%define ENDFUNC |
47 |
|
%endif |
48 |
%endif |
%endif |
49 |
%endmacro |
%endmacro |
50 |
|
|
53 |
;============================================================================= |
;============================================================================= |
54 |
|
|
55 |
%ifdef FORMAT_COFF |
%ifdef FORMAT_COFF |
56 |
SECTION .rodata data |
SECTION .rodata |
57 |
%else |
%else |
58 |
SECTION .rodata data align=16 |
SECTION .rodata align=16 |
59 |
%endif |
%endif |
60 |
|
|
61 |
ALIGN 64 |
ALIGN 64 |
70 |
cglobal sad16_sse2 |
cglobal sad16_sse2 |
71 |
cglobal dev16_sse2 |
cglobal dev16_sse2 |
72 |
|
|
73 |
|
cglobal sad16_sse3 |
74 |
|
cglobal dev16_sse3 |
75 |
|
|
76 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
77 |
; uint32_t sad16_sse2 (const uint8_t * const cur, <- assumed aligned! |
; uint32_t sad16_sse2 (const uint8_t * const cur, <- assumed aligned! |
78 |
; const uint8_t * const ref, |
; const uint8_t * const ref, |
81 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
82 |
|
|
83 |
|
|
84 |
%macro SAD_16x16_SSE2 0 |
%macro SAD_16x16_SSE2 1 |
85 |
movdqu xmm0, [edx] |
%1 xmm0, [edx] |
86 |
movdqu xmm1, [edx+ecx] |
%1 xmm1, [edx+ecx] |
87 |
lea edx,[edx+2*ecx] |
lea edx,[edx+2*ecx] |
88 |
movdqa xmm2, [eax] |
movdqa xmm2, [eax] |
89 |
movdqa xmm3, [eax+ecx] |
movdqa xmm3, [eax+ecx] |
94 |
paddusw xmm6,xmm1 |
paddusw xmm6,xmm1 |
95 |
%endmacro |
%endmacro |
96 |
|
|
97 |
ALIGN 16 |
%macro SAD16_SSE2_SSE3 1 |
|
sad16_sse2: |
|
98 |
mov eax, [esp+ 4] ; cur (assumed aligned) |
mov eax, [esp+ 4] ; cur (assumed aligned) |
99 |
mov edx, [esp+ 8] ; ref |
mov edx, [esp+ 8] ; ref |
100 |
mov ecx, [esp+12] ; stride |
mov ecx, [esp+12] ; stride |
101 |
|
|
102 |
pxor xmm6, xmm6 ; accum |
pxor xmm6, xmm6 ; accum |
103 |
|
|
104 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
105 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
106 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
107 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
108 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
109 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
110 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
111 |
SAD_16x16_SSE2 |
SAD_16x16_SSE2 %1 |
112 |
|
|
113 |
pshufd xmm5, xmm6, 00000010b |
pshufd xmm5, xmm6, 00000010b |
114 |
paddusw xmm6, xmm5 |
paddusw xmm6, xmm5 |
115 |
pextrw eax, xmm6, 0 |
pextrw eax, xmm6, 0 |
116 |
ret |
ret |
117 |
|
%endmacro |
118 |
|
|
119 |
|
ALIGN 16 |
120 |
|
sad16_sse2: |
121 |
|
SAD16_SSE2_SSE3 movdqu |
122 |
|
ENDFUNC |
123 |
|
|
124 |
|
|
125 |
|
ALIGN 16 |
126 |
|
sad16_sse3: |
127 |
|
SAD16_SSE2_SSE3 lddqu |
128 |
|
ENDFUNC |
129 |
|
|
130 |
|
|
131 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
132 |
; uint32_t dev16_sse2(const uint8_t * const cur, const uint32_t stride); |
; uint32_t dev16_sse2(const uint8_t * const cur, const uint32_t stride); |
133 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
134 |
|
|
135 |
%macro MEAN_16x16_SSE2 0 ; eax: src, ecx:stride, mm7: zero or mean => mm6: result |
%macro MEAN_16x16_SSE2 1 ; eax: src, ecx:stride, mm7: zero or mean => mm6: result |
136 |
movdqu xmm0, [eax] |
%1 xmm0, [eax] |
137 |
movdqu xmm1, [eax+ecx] |
%1 xmm1, [eax+ecx] |
138 |
lea eax, [eax+2*ecx] ; + 2*stride |
lea eax, [eax+2*ecx] ; + 2*stride |
139 |
psadbw xmm0, xmm7 |
psadbw xmm0, xmm7 |
140 |
paddusw xmm6, xmm0 |
paddusw xmm6, xmm0 |
143 |
%endmacro |
%endmacro |
144 |
|
|
145 |
|
|
146 |
ALIGN 16 |
%macro MEAN16_SSE2_SSE3 1 |
|
dev16_sse2: |
|
147 |
mov eax, [esp+ 4] ; src |
mov eax, [esp+ 4] ; src |
148 |
mov ecx, [esp+ 8] ; stride |
mov ecx, [esp+ 8] ; stride |
149 |
|
|
150 |
pxor xmm6, xmm6 ; accum |
pxor xmm6, xmm6 ; accum |
151 |
pxor xmm7, xmm7 ; zero |
pxor xmm7, xmm7 ; zero |
152 |
|
|
153 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
154 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
155 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
156 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
157 |
|
|
158 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
159 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
160 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
161 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
162 |
|
|
163 |
mov eax, [esp+ 4] ; src again |
mov eax, [esp+ 4] ; src again |
164 |
|
|
170 |
packuswb xmm7, xmm7 |
packuswb xmm7, xmm7 |
171 |
pshufd xmm7, xmm7, 00000000b |
pshufd xmm7, xmm7, 00000000b |
172 |
|
|
173 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
174 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
175 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
176 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
177 |
|
|
178 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
179 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
180 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
181 |
MEAN_16x16_SSE2 |
MEAN_16x16_SSE2 %1 |
182 |
|
|
183 |
pshufd xmm7, xmm6, 10b |
pshufd xmm7, xmm6, 10b |
184 |
paddusw xmm7, xmm6 |
paddusw xmm7, xmm6 |
185 |
pextrw eax, xmm7, 0 |
pextrw eax, xmm7, 0 |
186 |
ret |
ret |
187 |
|
%endmacro |
188 |
|
|
189 |
|
ALIGN 16 |
190 |
|
dev16_sse2: |
191 |
|
MEAN16_SSE2_SSE3 movdqu |
192 |
|
ENDFUNC |
193 |
|
|
194 |
|
ALIGN 16 |
195 |
|
dev16_sse3: |
196 |
|
MEAN16_SSE2_SSE3 lddqu |
197 |
|
ENDFUNC |
198 |
|
|
199 |
|
|
200 |
|
%ifidn __OUTPUT_FORMAT__,elf |
201 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
202 |
|
%endif |
203 |
|
|