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 1.17 2008-12-01 14:45:45 Isibaar Exp $ |
; * $Id: sad_sse2.asm,v 1.18 2008-12-02 13:44:55 Isibaar Exp $ |
24 |
; * |
; * |
25 |
; ***************************************************************************/ |
; ***************************************************************************/ |
26 |
|
|
63 |
movdqa xmm3, [_EAX+TMP0] |
movdqa xmm3, [_EAX+TMP0] |
64 |
lea _EAX,[_EAX+2*TMP0] |
lea _EAX,[_EAX+2*TMP0] |
65 |
psadbw xmm0, xmm2 |
psadbw xmm0, xmm2 |
66 |
paddusw xmm6,xmm0 |
paddusw xmm4,xmm0 |
67 |
psadbw xmm1, xmm3 |
psadbw xmm1, xmm3 |
68 |
paddusw xmm6,xmm1 |
paddusw xmm4,xmm1 |
69 |
%endmacro |
%endmacro |
70 |
|
|
71 |
%macro SAD16_SSE2_SSE3 1 |
%macro SAD16_SSE2_SSE3 1 |
|
PUSH_XMM6_XMM7 |
|
72 |
mov _EAX, prm1 ; cur (assumed aligned) |
mov _EAX, prm1 ; cur (assumed aligned) |
73 |
mov TMP1, prm2 ; ref |
mov TMP1, prm2 ; ref |
74 |
mov TMP0, prm3 ; stride |
mov TMP0, prm3 ; stride |
75 |
|
|
76 |
pxor xmm6, xmm6 ; accum |
pxor xmm4, xmm4 ; accum |
77 |
|
|
78 |
SAD_16x16_SSE2 %1 |
SAD_16x16_SSE2 %1 |
79 |
SAD_16x16_SSE2 %1 |
SAD_16x16_SSE2 %1 |
84 |
SAD_16x16_SSE2 %1 |
SAD_16x16_SSE2 %1 |
85 |
SAD_16x16_SSE2 %1 |
SAD_16x16_SSE2 %1 |
86 |
|
|
87 |
pshufd xmm5, xmm6, 00000010b |
pshufd xmm5, xmm4, 00000010b |
88 |
paddusw xmm6, xmm5 |
paddusw xmm4, xmm5 |
89 |
pextrw eax, xmm6, 0 |
pextrw eax, xmm4, 0 |
90 |
|
|
|
POP_XMM6_XMM7 |
|
91 |
ret |
ret |
92 |
%endmacro |
%endmacro |
93 |
|
|
111 |
%1 xmm0, [_EAX] |
%1 xmm0, [_EAX] |
112 |
%1 xmm1, [_EAX+TMP0] |
%1 xmm1, [_EAX+TMP0] |
113 |
lea _EAX, [_EAX+2*TMP0] ; + 2*stride |
lea _EAX, [_EAX+2*TMP0] ; + 2*stride |
114 |
psadbw xmm0, xmm7 |
psadbw xmm0, xmm5 |
115 |
paddusw xmm6, xmm0 |
paddusw xmm4, xmm0 |
116 |
psadbw xmm1, xmm7 |
psadbw xmm1, xmm5 |
117 |
paddusw xmm6, xmm1 |
paddusw xmm4, xmm1 |
118 |
%endmacro |
%endmacro |
119 |
|
|
120 |
|
|
121 |
%macro MEAN16_SSE2_SSE3 1 |
%macro MEAN16_SSE2_SSE3 1 |
|
PUSH_XMM6_XMM7 |
|
122 |
mov _EAX, prm1 ; src |
mov _EAX, prm1 ; src |
123 |
mov TMP0, prm2 ; stride |
mov TMP0, prm2 ; stride |
124 |
|
|
125 |
pxor xmm6, xmm6 ; accum |
pxor xmm4, xmm4 ; accum |
126 |
pxor xmm7, xmm7 ; zero |
pxor xmm5, xmm5 ; zero |
127 |
|
|
128 |
MEAN_16x16_SSE2 %1 |
MEAN_16x16_SSE2 %1 |
129 |
MEAN_16x16_SSE2 %1 |
MEAN_16x16_SSE2 %1 |
137 |
|
|
138 |
mov _EAX, prm1 ; src again |
mov _EAX, prm1 ; src again |
139 |
|
|
140 |
pshufd xmm7, xmm6, 10b |
pshufd xmm5, xmm4, 10b |
141 |
paddusw xmm7, xmm6 |
paddusw xmm5, xmm4 |
142 |
pxor xmm6, xmm6 ; zero accum |
pxor xmm4, xmm4 ; zero accum |
143 |
psrlw xmm7, 8 ; => Mean |
psrlw xmm5, 8 ; => Mean |
144 |
pshuflw xmm7, xmm7, 0 ; replicate Mean |
pshuflw xmm5, xmm5, 0 ; replicate Mean |
145 |
packuswb xmm7, xmm7 |
packuswb xmm5, xmm5 |
146 |
pshufd xmm7, xmm7, 00000000b |
pshufd xmm5, xmm5, 00000000b |
147 |
|
|
148 |
MEAN_16x16_SSE2 %1 |
MEAN_16x16_SSE2 %1 |
149 |
MEAN_16x16_SSE2 %1 |
MEAN_16x16_SSE2 %1 |
155 |
MEAN_16x16_SSE2 %1 |
MEAN_16x16_SSE2 %1 |
156 |
MEAN_16x16_SSE2 %1 |
MEAN_16x16_SSE2 %1 |
157 |
|
|
158 |
pshufd xmm7, xmm6, 10b |
pshufd xmm5, xmm4, 10b |
159 |
paddusw xmm7, xmm6 |
paddusw xmm5, xmm4 |
160 |
pextrw eax, xmm7, 0 |
pextrw eax, xmm5, 0 |
161 |
|
|
|
POP_XMM6_XMM7 |
|
162 |
ret |
ret |
163 |
%endmacro |
%endmacro |
164 |
|
|