vpx_dsp_rtcd_defs.pl revision 7ce0a1d1337c01056ba24006efab21f00e179e04 
1sub vpx_dsp_forward_decls() {
2print <<EOF
3/*
4 * DSP
5 */
6
7#include "vpx/vpx_integer.h"
8#include "vpx_dsp/vpx_dsp_common.h"
9
10EOF
11}
12forward_decls qw/vpx_dsp_forward_decls/;
13
14# x86inc.asm had specific constraints. break it out so it's easy to disable.
15# zero all the variables to avoid tricky else conditions.
16$mmx_x86inc = $sse_x86inc = $sse2_x86inc = $ssse3_x86inc = $avx_x86inc =
17  $avx2_x86inc = '';
18$mmx_x86_64_x86inc = $sse_x86_64_x86inc = $sse2_x86_64_x86inc =
19  $ssse3_x86_64_x86inc = $avx_x86_64_x86inc = $avx2_x86_64_x86inc = '';
20if (vpx_config("CONFIG_USE_X86INC") eq "yes") {
21  $mmx_x86inc = 'mmx';
22  $sse_x86inc = 'sse';
23  $sse2_x86inc = 'sse2';
24  $ssse3_x86inc = 'ssse3';
25  $avx_x86inc = 'avx';
26  $avx2_x86inc = 'avx2';
27  if ($opts{arch} eq "x86_64") {
28    $mmx_x86_64_x86inc = 'mmx';
29    $sse_x86_64_x86inc = 'sse';
30    $sse2_x86_64_x86inc = 'sse2';
31    $ssse3_x86_64_x86inc = 'ssse3';
32    $avx_x86_64_x86inc = 'avx';
33    $avx2_x86_64_x86inc = 'avx2';
34  }
35}
36
37# optimizations which depend on multiple features
38$avx2_ssse3 = '';
39if ((vpx_config("HAVE_AVX2") eq "yes") && (vpx_config("HAVE_SSSE3") eq "yes")) {
40  $avx2_ssse3 = 'avx2';
41}
42
43# functions that are 64 bit only.
44$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
45if ($opts{arch} eq "x86_64") {
46  $mmx_x86_64 = 'mmx';
47  $sse2_x86_64 = 'sse2';
48  $ssse3_x86_64 = 'ssse3';
49  $avx_x86_64 = 'avx';
50  $avx2_x86_64 = 'avx2';
51}
52
53#
54# Intra prediction
55#
56
57if ((vpx_config("CONFIG_VP9") eq "yes") || (vpx_config("CONFIG_VP10") eq "yes")) {
58  add_proto qw/void vpx_d207_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
59  specialize qw/vpx_d207_predictor_4x4/, "$ssse3_x86inc";
60
61  add_proto qw/void vpx_d45_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
62  specialize qw/vpx_d45_predictor_4x4 neon/, "$ssse3_x86inc";
63
64  add_proto qw/void vpx_d63_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
65  specialize qw/vpx_d63_predictor_4x4/, "$ssse3_x86inc";
66
67  add_proto qw/void vpx_h_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
68  specialize qw/vpx_h_predictor_4x4 neon dspr2 msa/, "$ssse3_x86inc";
69
70  add_proto qw/void vpx_d117_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
71  specialize qw/vpx_d117_predictor_4x4/;
72
73  add_proto qw/void vpx_d135_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
74  specialize qw/vpx_d135_predictor_4x4 neon/;
75
76  add_proto qw/void vpx_d153_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
77  specialize qw/vpx_d153_predictor_4x4/, "$ssse3_x86inc";
78
79  add_proto qw/void vpx_v_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
80  specialize qw/vpx_v_predictor_4x4 neon msa/, "$sse_x86inc";
81
82  add_proto qw/void vpx_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
83  specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse_x86inc";
84
85  add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
86  specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon/, "$sse_x86inc";
87
88  add_proto qw/void vpx_dc_top_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
89  specialize qw/vpx_dc_top_predictor_4x4 msa neon/, "$sse_x86inc";
90
91  add_proto qw/void vpx_dc_left_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
92  specialize qw/vpx_dc_left_predictor_4x4 msa neon/, "$sse_x86inc";
93
94  add_proto qw/void vpx_dc_128_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
95  specialize qw/vpx_dc_128_predictor_4x4 msa neon/, "$sse_x86inc";
96
97  add_proto qw/void vpx_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
98  specialize qw/vpx_d207_predictor_8x8/, "$ssse3_x86inc";
99
100  add_proto qw/void vpx_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
101  specialize qw/vpx_d45_predictor_8x8 neon/, "$ssse3_x86inc";
102
103  add_proto qw/void vpx_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
104  specialize qw/vpx_d63_predictor_8x8/, "$ssse3_x86inc";
105
106  add_proto qw/void vpx_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
107  specialize qw/vpx_h_predictor_8x8 neon dspr2 msa/, "$ssse3_x86inc";
108
109  add_proto qw/void vpx_d117_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
110  specialize qw/vpx_d117_predictor_8x8/;
111
112  add_proto qw/void vpx_d135_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
113  specialize qw/vpx_d135_predictor_8x8/;
114
115  add_proto qw/void vpx_d153_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
116  specialize qw/vpx_d153_predictor_8x8/, "$ssse3_x86inc";
117
118  add_proto qw/void vpx_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
119  specialize qw/vpx_v_predictor_8x8 neon msa/, "$sse_x86inc";
120
121  add_proto qw/void vpx_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
122  specialize qw/vpx_tm_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
123
124  add_proto qw/void vpx_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
125  specialize qw/vpx_dc_predictor_8x8 dspr2 neon msa/, "$sse_x86inc";
126
127  add_proto qw/void vpx_dc_top_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
128  specialize qw/vpx_dc_top_predictor_8x8 neon msa/, "$sse_x86inc";
129
130  add_proto qw/void vpx_dc_left_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
131  specialize qw/vpx_dc_left_predictor_8x8 neon msa/, "$sse_x86inc";
132
133  add_proto qw/void vpx_dc_128_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
134  specialize qw/vpx_dc_128_predictor_8x8 neon msa/, "$sse_x86inc";
135
136  add_proto qw/void vpx_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
137  specialize qw/vpx_d207_predictor_16x16/, "$ssse3_x86inc";
138
139  add_proto qw/void vpx_d45_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
140  specialize qw/vpx_d45_predictor_16x16 neon/, "$ssse3_x86inc";
141
142  add_proto qw/void vpx_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
143  specialize qw/vpx_d63_predictor_16x16/, "$ssse3_x86inc";
144
145  add_proto qw/void vpx_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
146  specialize qw/vpx_h_predictor_16x16 neon dspr2 msa/, "$ssse3_x86inc";
147
148  add_proto qw/void vpx_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
149  specialize qw/vpx_d117_predictor_16x16/;
150
151  add_proto qw/void vpx_d135_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
152  specialize qw/vpx_d135_predictor_16x16/;
153
154  add_proto qw/void vpx_d153_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
155  specialize qw/vpx_d153_predictor_16x16/, "$ssse3_x86inc";
156
157  add_proto qw/void vpx_v_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
158  specialize qw/vpx_v_predictor_16x16 neon msa/, "$sse2_x86inc";
159
160  add_proto qw/void vpx_tm_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
161  specialize qw/vpx_tm_predictor_16x16 neon msa/, "$sse2_x86inc";
162
163  add_proto qw/void vpx_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
164  specialize qw/vpx_dc_predictor_16x16 dspr2 neon msa/, "$sse2_x86inc";
165
166  add_proto qw/void vpx_dc_top_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
167  specialize qw/vpx_dc_top_predictor_16x16 neon msa/, "$sse2_x86inc";
168
169  add_proto qw/void vpx_dc_left_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
170  specialize qw/vpx_dc_left_predictor_16x16 neon msa/, "$sse2_x86inc";
171
172  add_proto qw/void vpx_dc_128_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
173  specialize qw/vpx_dc_128_predictor_16x16 neon msa/, "$sse2_x86inc";
174
175  add_proto qw/void vpx_d207_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
176  specialize qw/vpx_d207_predictor_32x32/, "$ssse3_x86inc";
177
178  add_proto qw/void vpx_d45_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
179  specialize qw/vpx_d45_predictor_32x32/, "$ssse3_x86inc";
180
181  add_proto qw/void vpx_d63_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
182  specialize qw/vpx_d63_predictor_32x32/, "$ssse3_x86inc";
183
184  add_proto qw/void vpx_h_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
185  specialize qw/vpx_h_predictor_32x32 neon msa/, "$ssse3_x86inc";
186
187  add_proto qw/void vpx_d117_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
188  specialize qw/vpx_d117_predictor_32x32/;
189
190  add_proto qw/void vpx_d135_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
191  specialize qw/vpx_d135_predictor_32x32/;
192
193  add_proto qw/void vpx_d153_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
194  specialize qw/vpx_d153_predictor_32x32/, "$ssse3_x86inc";
195
196  add_proto qw/void vpx_v_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
197  specialize qw/vpx_v_predictor_32x32 neon msa/, "$sse2_x86inc";
198
199  add_proto qw/void vpx_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
200  specialize qw/vpx_tm_predictor_32x32 neon msa/, "$sse2_x86_64_x86inc";
201
202  add_proto qw/void vpx_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
203  specialize qw/vpx_dc_predictor_32x32 msa neon/, "$sse2_x86inc";
204
205  add_proto qw/void vpx_dc_top_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
206  specialize qw/vpx_dc_top_predictor_32x32 msa neon/, "$sse2_x86inc";
207
208  add_proto qw/void vpx_dc_left_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
209  specialize qw/vpx_dc_left_predictor_32x32 msa neon/, "$sse2_x86inc";
210
211  add_proto qw/void vpx_dc_128_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
212  specialize qw/vpx_dc_128_predictor_32x32 msa neon/, "$sse2_x86inc";
213
214# High bitdepth functions
215  if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
216    add_proto qw/void vpx_highbd_d207_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
217    specialize qw/vpx_highbd_d207_predictor_4x4/;
218
219    add_proto qw/void vpx_highbd_d45_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
220    specialize qw/vpx_highbd_d45_predictor_4x4/;
221
222    add_proto qw/void vpx_highbd_d63_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
223    specialize qw/vpx_highbd_d63_predictor_4x4/;
224
225    add_proto qw/void vpx_highbd_h_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
226    specialize qw/vpx_highbd_h_predictor_4x4/;
227
228    add_proto qw/void vpx_highbd_d117_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
229    specialize qw/vpx_highbd_d117_predictor_4x4/;
230
231    add_proto qw/void vpx_highbd_d135_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
232    specialize qw/vpx_highbd_d135_predictor_4x4/;
233
234    add_proto qw/void vpx_highbd_d153_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
235    specialize qw/vpx_highbd_d153_predictor_4x4/;
236
237    add_proto qw/void vpx_highbd_v_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
238    specialize qw/vpx_highbd_v_predictor_4x4/, "$sse_x86inc";
239
240    add_proto qw/void vpx_highbd_tm_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
241    specialize qw/vpx_highbd_tm_predictor_4x4/, "$sse_x86inc";
242
243    add_proto qw/void vpx_highbd_dc_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
244    specialize qw/vpx_highbd_dc_predictor_4x4/, "$sse_x86inc";
245
246    add_proto qw/void vpx_highbd_dc_top_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
247    specialize qw/vpx_highbd_dc_top_predictor_4x4/;
248
249    add_proto qw/void vpx_highbd_dc_left_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
250    specialize qw/vpx_highbd_dc_left_predictor_4x4/;
251
252    add_proto qw/void vpx_highbd_dc_128_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
253    specialize qw/vpx_highbd_dc_128_predictor_4x4/;
254
255    add_proto qw/void vpx_highbd_d207_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
256    specialize qw/vpx_highbd_d207_predictor_8x8/;
257
258    add_proto qw/void vpx_highbd_d45_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
259    specialize qw/vpx_highbd_d45_predictor_8x8/;
260
261    add_proto qw/void vpx_highbd_d63_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
262    specialize qw/vpx_highbd_d63_predictor_8x8/;
263
264    add_proto qw/void vpx_highbd_h_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
265    specialize qw/vpx_highbd_h_predictor_8x8/;
266
267    add_proto qw/void vpx_highbd_d117_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
268    specialize qw/vpx_highbd_d117_predictor_8x8/;
269
270    add_proto qw/void vpx_highbd_d135_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
271    specialize qw/vpx_highbd_d135_predictor_8x8/;
272
273    add_proto qw/void vpx_highbd_d153_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
274    specialize qw/vpx_highbd_d153_predictor_8x8/;
275
276    add_proto qw/void vpx_highbd_v_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
277    specialize qw/vpx_highbd_v_predictor_8x8/, "$sse2_x86inc";
278
279    add_proto qw/void vpx_highbd_tm_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
280    specialize qw/vpx_highbd_tm_predictor_8x8/, "$sse2_x86inc";
281
282    add_proto qw/void vpx_highbd_dc_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
283    specialize qw/vpx_highbd_dc_predictor_8x8/, "$sse2_x86inc";;
284
285    add_proto qw/void vpx_highbd_dc_top_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
286    specialize qw/vpx_highbd_dc_top_predictor_8x8/;
287
288    add_proto qw/void vpx_highbd_dc_left_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
289    specialize qw/vpx_highbd_dc_left_predictor_8x8/;
290
291    add_proto qw/void vpx_highbd_dc_128_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
292    specialize qw/vpx_highbd_dc_128_predictor_8x8/;
293
294    add_proto qw/void vpx_highbd_d207_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
295    specialize qw/vpx_highbd_d207_predictor_16x16/;
296
297    add_proto qw/void vpx_highbd_d45_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
298    specialize qw/vpx_highbd_d45_predictor_16x16/;
299
300    add_proto qw/void vpx_highbd_d63_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
301    specialize qw/vpx_highbd_d63_predictor_16x16/;
302
303    add_proto qw/void vpx_highbd_h_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
304    specialize qw/vpx_highbd_h_predictor_16x16/;
305
306    add_proto qw/void vpx_highbd_d117_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
307    specialize qw/vpx_highbd_d117_predictor_16x16/;
308
309    add_proto qw/void vpx_highbd_d135_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
310    specialize qw/vpx_highbd_d135_predictor_16x16/;
311
312    add_proto qw/void vpx_highbd_d153_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
313    specialize qw/vpx_highbd_d153_predictor_16x16/;
314
315    add_proto qw/void vpx_highbd_v_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
316    specialize qw/vpx_highbd_v_predictor_16x16/, "$sse2_x86inc";
317
318    add_proto qw/void vpx_highbd_tm_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
319    specialize qw/vpx_highbd_tm_predictor_16x16/, "$sse2_x86_64_x86inc";
320
321    add_proto qw/void vpx_highbd_dc_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
322    specialize qw/vpx_highbd_dc_predictor_16x16/, "$sse2_x86inc";
323
324    add_proto qw/void vpx_highbd_dc_top_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
325    specialize qw/vpx_highbd_dc_top_predictor_16x16/;
326
327    add_proto qw/void vpx_highbd_dc_left_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
328    specialize qw/vpx_highbd_dc_left_predictor_16x16/;
329
330    add_proto qw/void vpx_highbd_dc_128_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
331    specialize qw/vpx_highbd_dc_128_predictor_16x16/;
332
333    add_proto qw/void vpx_highbd_d207_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
334    specialize qw/vpx_highbd_d207_predictor_32x32/;
335
336    add_proto qw/void vpx_highbd_d45_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
337    specialize qw/vpx_highbd_d45_predictor_32x32/;
338
339    add_proto qw/void vpx_highbd_d63_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
340    specialize qw/vpx_highbd_d63_predictor_32x32/;
341
342    add_proto qw/void vpx_highbd_h_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
343    specialize qw/vpx_highbd_h_predictor_32x32/;
344
345    add_proto qw/void vpx_highbd_d117_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
346    specialize qw/vpx_highbd_d117_predictor_32x32/;
347
348    add_proto qw/void vpx_highbd_d135_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
349    specialize qw/vpx_highbd_d135_predictor_32x32/;
350
351    add_proto qw/void vpx_highbd_d153_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
352    specialize qw/vpx_highbd_d153_predictor_32x32/;
353
354    add_proto qw/void vpx_highbd_v_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
355    specialize qw/vpx_highbd_v_predictor_32x32/, "$sse2_x86inc";
356
357    add_proto qw/void vpx_highbd_tm_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
358    specialize qw/vpx_highbd_tm_predictor_32x32/, "$sse2_x86_64_x86inc";
359
360    add_proto qw/void vpx_highbd_dc_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
361    specialize qw/vpx_highbd_dc_predictor_32x32/, "$sse2_x86_64_x86inc";
362
363    add_proto qw/void vpx_highbd_dc_top_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
364    specialize qw/vpx_highbd_dc_top_predictor_32x32/;
365
366    add_proto qw/void vpx_highbd_dc_left_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
367    specialize qw/vpx_highbd_dc_left_predictor_32x32/;
368
369    add_proto qw/void vpx_highbd_dc_128_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
370    specialize qw/vpx_highbd_dc_128_predictor_32x32/;
371  }  # CONFIG_VP9_HIGHBITDEPTH
372}  # CONFIG_VP9 || CONFIG_VP10
373
374#
375# Sub Pixel Filters
376#
377add_proto qw/void vpx_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
378specialize qw/vpx_convolve_copy neon dspr2 msa/, "$sse2_x86inc";
379
380add_proto qw/void vpx_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
381specialize qw/vpx_convolve_avg neon dspr2 msa/, "$sse2_x86inc";
382
383add_proto qw/void vpx_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
384specialize qw/vpx_convolve8 sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
385
386add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
387specialize qw/vpx_convolve8_horiz sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
388
389add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
390specialize qw/vpx_convolve8_vert sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
391
392add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
393specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa/;
394
395add_proto qw/void vpx_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
396specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa/;
397
398add_proto qw/void vpx_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
399specialize qw/vpx_convolve8_avg_vert sse2 ssse3 neon dspr2 msa/;
400
401add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
402specialize qw/vpx_scaled_2d ssse3/;
403
404add_proto qw/void vpx_scaled_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
405specialize qw/vpx_scaled_horiz/;
406
407add_proto qw/void vpx_scaled_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
408specialize qw/vpx_scaled_vert/;
409
410add_proto qw/void vpx_scaled_avg_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
411specialize qw/vpx_scaled_avg_2d/;
412
413add_proto qw/void vpx_scaled_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
414specialize qw/vpx_scaled_avg_horiz/;
415
416add_proto qw/void vpx_scaled_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
417specialize qw/vpx_scaled_avg_vert/;
418
419if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
420  #
421  # Sub Pixel Filters
422  #
423  add_proto qw/void vpx_highbd_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
424  specialize qw/vpx_highbd_convolve_copy/;
425
426  add_proto qw/void vpx_highbd_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
427  specialize qw/vpx_highbd_convolve_avg/;
428
429  add_proto qw/void vpx_highbd_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
430  specialize qw/vpx_highbd_convolve8/, "$sse2_x86_64";
431
432  add_proto qw/void vpx_highbd_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
433  specialize qw/vpx_highbd_convolve8_horiz/, "$sse2_x86_64";
434
435  add_proto qw/void vpx_highbd_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
436  specialize qw/vpx_highbd_convolve8_vert/, "$sse2_x86_64";
437
438  add_proto qw/void vpx_highbd_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
439  specialize qw/vpx_highbd_convolve8_avg/, "$sse2_x86_64";
440
441  add_proto qw/void vpx_highbd_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
442  specialize qw/vpx_highbd_convolve8_avg_horiz/, "$sse2_x86_64";
443
444  add_proto qw/void vpx_highbd_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
445  specialize qw/vpx_highbd_convolve8_avg_vert/, "$sse2_x86_64";
446}  # CONFIG_VP9_HIGHBITDEPTH
447
448#
449# Loopfilter
450#
451add_proto qw/void vpx_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
452specialize qw/vpx_lpf_vertical_16 sse2 neon_asm dspr2 msa/;
453$vpx_lpf_vertical_16_neon_asm=vpx_lpf_vertical_16_neon;
454
455add_proto qw/void vpx_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
456specialize qw/vpx_lpf_vertical_16_dual sse2 neon_asm dspr2 msa/;
457$vpx_lpf_vertical_16_dual_neon_asm=vpx_lpf_vertical_16_dual_neon;
458
459add_proto qw/void vpx_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
460specialize qw/vpx_lpf_vertical_8 sse2 neon dspr2 msa/;
461
462add_proto qw/void vpx_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
463specialize qw/vpx_lpf_vertical_8_dual sse2 neon_asm dspr2 msa/;
464$vpx_lpf_vertical_8_dual_neon_asm=vpx_lpf_vertical_8_dual_neon;
465
466add_proto qw/void vpx_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
467specialize qw/vpx_lpf_vertical_4 mmx neon dspr2 msa/;
468
469add_proto qw/void vpx_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
470specialize qw/vpx_lpf_vertical_4_dual sse2 neon dspr2 msa/;
471
472add_proto qw/void vpx_lpf_horizontal_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
473specialize qw/vpx_lpf_horizontal_16 sse2 avx2 neon_asm dspr2 msa/;
474$vpx_lpf_horizontal_16_neon_asm=vpx_lpf_horizontal_16_neon;
475
476add_proto qw/void vpx_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
477specialize qw/vpx_lpf_horizontal_8 sse2 neon dspr2 msa/;
478
479add_proto qw/void vpx_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
480specialize qw/vpx_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
481$vpx_lpf_horizontal_8_dual_neon_asm=vpx_lpf_horizontal_8_dual_neon;
482
483add_proto qw/void vpx_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
484specialize qw/vpx_lpf_horizontal_4 mmx neon dspr2 msa/;
485
486add_proto qw/void vpx_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
487specialize qw/vpx_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
488
489if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
490  add_proto qw/void vpx_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
491  specialize qw/vpx_highbd_lpf_vertical_16 sse2/;
492
493  add_proto qw/void vpx_highbd_lpf_vertical_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
494  specialize qw/vpx_highbd_lpf_vertical_16_dual sse2/;
495
496  add_proto qw/void vpx_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
497  specialize qw/vpx_highbd_lpf_vertical_8 sse2/;
498
499  add_proto qw/void vpx_highbd_lpf_vertical_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
500  specialize qw/vpx_highbd_lpf_vertical_8_dual sse2/;
501
502  add_proto qw/void vpx_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
503  specialize qw/vpx_highbd_lpf_vertical_4 sse2/;
504
505  add_proto qw/void vpx_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
506  specialize qw/vpx_highbd_lpf_vertical_4_dual sse2/;
507
508  add_proto qw/void vpx_highbd_lpf_horizontal_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
509  specialize qw/vpx_highbd_lpf_horizontal_16 sse2/;
510
511  add_proto qw/void vpx_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
512  specialize qw/vpx_highbd_lpf_horizontal_8 sse2/;
513
514  add_proto qw/void vpx_highbd_lpf_horizontal_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
515  specialize qw/vpx_highbd_lpf_horizontal_8_dual sse2/;
516
517  add_proto qw/void vpx_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
518  specialize qw/vpx_highbd_lpf_horizontal_4 sse2/;
519
520  add_proto qw/void vpx_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
521  specialize qw/vpx_highbd_lpf_horizontal_4_dual sse2/;
522}  # CONFIG_VP9_HIGHBITDEPTH
523
524#
525# Encoder functions.
526#
527
528#
529# Forward transform
530#
531if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
532if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
533  add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
534  specialize qw/vpx_fdct4x4 sse2/;
535
536  add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
537  specialize qw/vpx_fdct4x4_1 sse2/;
538
539  add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
540  specialize qw/vpx_fdct8x8 sse2/;
541
542  add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
543  specialize qw/vpx_fdct8x8_1 sse2/;
544
545  add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
546  specialize qw/vpx_fdct16x16 sse2/;
547
548  add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
549  specialize qw/vpx_fdct16x16_1 sse2/;
550
551  add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
552  specialize qw/vpx_fdct32x32 sse2/;
553
554  add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
555  specialize qw/vpx_fdct32x32_rd sse2/;
556
557  add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
558  specialize qw/vpx_fdct32x32_1 sse2/;
559
560  add_proto qw/void vpx_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
561  specialize qw/vpx_highbd_fdct4x4 sse2/;
562
563  add_proto qw/void vpx_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
564  specialize qw/vpx_highbd_fdct8x8 sse2/;
565
566  add_proto qw/void vpx_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
567  specialize qw/vpx_highbd_fdct8x8_1/;
568
569  add_proto qw/void vpx_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
570  specialize qw/vpx_highbd_fdct16x16 sse2/;
571
572  add_proto qw/void vpx_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
573  specialize qw/vpx_highbd_fdct16x16_1/;
574
575  add_proto qw/void vpx_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
576  specialize qw/vpx_highbd_fdct32x32 sse2/;
577
578  add_proto qw/void vpx_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
579  specialize qw/vpx_highbd_fdct32x32_rd sse2/;
580
581  add_proto qw/void vpx_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
582  specialize qw/vpx_highbd_fdct32x32_1/;
583} else {
584  add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
585  specialize qw/vpx_fdct4x4 sse2 msa/;
586
587  add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
588  specialize qw/vpx_fdct4x4_1 sse2/;
589
590  add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
591  specialize qw/vpx_fdct8x8 sse2 neon msa/, "$ssse3_x86_64_x86inc";
592
593  add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
594  specialize qw/vpx_fdct8x8_1 sse2 neon msa/;
595
596  add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
597  specialize qw/vpx_fdct16x16 sse2 msa/;
598
599  add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
600  specialize qw/vpx_fdct16x16_1 sse2 msa/;
601
602  add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
603  specialize qw/vpx_fdct32x32 sse2 avx2 msa/;
604
605  add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
606  specialize qw/vpx_fdct32x32_rd sse2 avx2 msa/;
607
608  add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
609  specialize qw/vpx_fdct32x32_1 sse2 msa/;
610}  # CONFIG_VP9_HIGHBITDEPTH
611}  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
612
613#
614# Inverse transform
615if ((vpx_config("CONFIG_VP9") eq "yes") || (vpx_config("CONFIG_VP10") eq "yes")) {
616if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
617  # Note as optimized versions of these functions are added we need to add a check to ensure
618  # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
619  add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
620  specialize qw/vpx_idct4x4_1_add/;
621
622  add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
623  specialize qw/vpx_idct4x4_16_add/;
624
625  add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
626  specialize qw/vpx_idct8x8_1_add/;
627
628  add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
629  specialize qw/vpx_idct8x8_64_add/;
630
631  add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
632  specialize qw/vpx_idct8x8_12_add/;
633
634  add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
635  specialize qw/vpx_idct16x16_1_add/;
636
637  add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
638  specialize qw/vpx_idct16x16_256_add/;
639
640  add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
641  specialize qw/vpx_idct16x16_10_add/;
642
643  add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
644  specialize qw/vpx_idct32x32_1024_add/;
645
646  add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
647  specialize qw/vpx_idct32x32_34_add/;
648
649  add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
650  specialize qw/vpx_idct32x32_1_add/;
651
652  add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
653  specialize qw/vpx_iwht4x4_1_add/;
654
655  add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
656  specialize qw/vpx_iwht4x4_16_add/;
657
658  add_proto qw/void vpx_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
659  specialize qw/vpx_highbd_idct4x4_1_add/;
660
661  add_proto qw/void vpx_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
662  specialize qw/vpx_highbd_idct8x8_1_add/;
663
664  add_proto qw/void vpx_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
665  specialize qw/vpx_highbd_idct16x16_1_add/;
666
667  add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
668  specialize qw/vpx_highbd_idct32x32_1024_add/;
669
670  add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
671  specialize qw/vpx_highbd_idct32x32_34_add/;
672
673  add_proto qw/void vpx_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
674  specialize qw/vpx_highbd_idct32x32_1_add/;
675
676  add_proto qw/void vpx_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
677  specialize qw/vpx_highbd_iwht4x4_1_add/;
678
679  add_proto qw/void vpx_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
680  specialize qw/vpx_highbd_iwht4x4_16_add/;
681
682  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
683  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
684    add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
685    specialize qw/vpx_highbd_idct4x4_16_add/;
686
687    add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
688    specialize qw/vpx_highbd_idct8x8_64_add/;
689
690    add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
691    specialize qw/vpx_highbd_idct8x8_10_add/;
692
693    add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
694    specialize qw/vpx_highbd_idct16x16_256_add/;
695
696    add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
697    specialize qw/vpx_highbd_idct16x16_10_add/;
698  } else {
699    add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
700    specialize qw/vpx_highbd_idct4x4_16_add sse2/;
701
702    add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
703    specialize qw/vpx_highbd_idct8x8_64_add sse2/;
704
705    add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
706    specialize qw/vpx_highbd_idct8x8_10_add sse2/;
707
708    add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
709    specialize qw/vpx_highbd_idct16x16_256_add sse2/;
710
711    add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
712    specialize qw/vpx_highbd_idct16x16_10_add sse2/;
713  }  # CONFIG_EMULATE_HARDWARE
714} else {
715  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
716  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
717    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
718    specialize qw/vpx_idct4x4_1_add/;
719
720    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
721    specialize qw/vpx_idct4x4_16_add/;
722
723    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
724    specialize qw/vpx_idct8x8_1_add/;
725
726    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
727    specialize qw/vpx_idct8x8_64_add/;
728
729    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
730    specialize qw/vpx_idct8x8_12_add/;
731
732    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
733    specialize qw/vpx_idct16x16_1_add/;
734
735    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
736    specialize qw/vpx_idct16x16_256_add/;
737
738    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
739    specialize qw/vpx_idct16x16_10_add/;
740
741    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
742    specialize qw/vpx_idct32x32_1024_add/;
743
744    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
745    specialize qw/vpx_idct32x32_34_add/;
746
747    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
748    specialize qw/vpx_idct32x32_1_add/;
749
750    add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
751    specialize qw/vpx_iwht4x4_1_add/;
752
753    add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
754    specialize qw/vpx_iwht4x4_16_add/;
755  } else {
756    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
757    specialize qw/vpx_idct4x4_1_add sse2 neon dspr2 msa/;
758
759    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
760    specialize qw/vpx_idct4x4_16_add sse2 neon dspr2 msa/;
761
762    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
763    specialize qw/vpx_idct8x8_1_add sse2 neon dspr2 msa/;
764
765    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
766    specialize qw/vpx_idct8x8_64_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
767
768    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
769    specialize qw/vpx_idct8x8_12_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
770
771    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
772    specialize qw/vpx_idct16x16_1_add sse2 neon dspr2 msa/;
773
774    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
775    specialize qw/vpx_idct16x16_256_add sse2 neon dspr2 msa/;
776
777    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
778    specialize qw/vpx_idct16x16_10_add sse2 neon dspr2 msa/;
779
780    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
781    specialize qw/vpx_idct32x32_1024_add sse2 neon dspr2 msa/;
782
783    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
784    specialize qw/vpx_idct32x32_34_add sse2 neon_asm dspr2 msa/;
785    # Need to add 34 eob idct32x32 neon implementation.
786    $vpx_idct32x32_34_add_neon_asm=vpx_idct32x32_1024_add_neon;
787
788    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
789    specialize qw/vpx_idct32x32_1_add sse2 neon dspr2 msa/;
790
791    add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
792    specialize qw/vpx_iwht4x4_1_add msa/;
793
794    add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
795    specialize qw/vpx_iwht4x4_16_add msa/, "$sse2_x86inc";
796  }  # CONFIG_EMULATE_HARDWARE
797}  # CONFIG_VP9_HIGHBITDEPTH
798}  # CONFIG_VP9 || CONFIG_VP10
799
800#
801# Quantization
802#
803if ((vpx_config("CONFIG_VP9_ENCODER") eq "yes") || (vpx_config("CONFIG_VP10_ENCODER") eq "yes")) {
804if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
805  add_proto qw/void vpx_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
806  specialize qw/vpx_quantize_b/;
807
808  add_proto qw/void vpx_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
809  specialize qw/vpx_quantize_b_32x32/;
810
811  add_proto qw/void vpx_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
812  specialize qw/vpx_highbd_quantize_b sse2/;
813
814  add_proto qw/void vpx_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
815  specialize qw/vpx_highbd_quantize_b_32x32 sse2/;
816} else {
817  add_proto qw/void vpx_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
818  specialize qw/vpx_quantize_b sse2/, "$ssse3_x86_64_x86inc";
819
820  add_proto qw/void vpx_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
821  specialize qw/vpx_quantize_b_32x32/, "$ssse3_x86_64_x86inc";
822}  # CONFIG_VP9_HIGHBITDEPTH
823}  # CONFIG_VP9_ENCODER || CONFIG_VP10_ENCODER
824
825if (vpx_config("CONFIG_ENCODERS") eq "yes") {
826#
827# Block subtraction
828#
829add_proto qw/void vpx_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
830specialize qw/vpx_subtract_block neon msa/, "$sse2_x86inc";
831
832#
833# Single block SAD
834#
835add_proto qw/unsigned int vpx_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
836specialize qw/vpx_sad64x64 avx2 neon msa/, "$sse2_x86inc";
837
838add_proto qw/unsigned int vpx_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
839specialize qw/vpx_sad64x32 avx2 msa/, "$sse2_x86inc";
840
841add_proto qw/unsigned int vpx_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
842specialize qw/vpx_sad32x64 avx2 msa/, "$sse2_x86inc";
843
844add_proto qw/unsigned int vpx_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
845specialize qw/vpx_sad32x32 avx2 neon msa/, "$sse2_x86inc";
846
847add_proto qw/unsigned int vpx_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
848specialize qw/vpx_sad32x16 avx2 msa/, "$sse2_x86inc";
849
850add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
851specialize qw/vpx_sad16x32 msa/, "$sse2_x86inc";
852
853add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
854specialize qw/vpx_sad16x16 mmx media neon msa/, "$sse2_x86inc";
855
856add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
857specialize qw/vpx_sad16x8 mmx neon msa/, "$sse2_x86inc";
858
859add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
860specialize qw/vpx_sad8x16 mmx neon msa/, "$sse2_x86inc";
861
862add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
863specialize qw/vpx_sad8x8 mmx neon msa/, "$sse2_x86inc";
864
865add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
866specialize qw/vpx_sad8x4 msa/, "$sse2_x86inc";
867
868add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
869specialize qw/vpx_sad4x8 msa/, "$sse_x86inc";
870
871add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
872specialize qw/vpx_sad4x4 mmx neon msa/, "$sse_x86inc";
873
874#
875# Avg
876#
877add_proto qw/unsigned int vpx_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
878specialize qw/vpx_sad64x64_avg avx2 msa/, "$sse2_x86inc";
879
880add_proto qw/unsigned int vpx_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
881specialize qw/vpx_sad64x32_avg avx2 msa/, "$sse2_x86inc";
882
883add_proto qw/unsigned int vpx_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
884specialize qw/vpx_sad32x64_avg avx2 msa/, "$sse2_x86inc";
885
886add_proto qw/unsigned int vpx_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
887specialize qw/vpx_sad32x32_avg avx2 msa/, "$sse2_x86inc";
888
889add_proto qw/unsigned int vpx_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
890specialize qw/vpx_sad32x16_avg avx2 msa/, "$sse2_x86inc";
891
892add_proto qw/unsigned int vpx_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
893specialize qw/vpx_sad16x32_avg msa/, "$sse2_x86inc";
894
895add_proto qw/unsigned int vpx_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
896specialize qw/vpx_sad16x16_avg msa/, "$sse2_x86inc";
897
898add_proto qw/unsigned int vpx_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
899specialize qw/vpx_sad16x8_avg msa/, "$sse2_x86inc";
900
901add_proto qw/unsigned int vpx_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
902specialize qw/vpx_sad8x16_avg msa/, "$sse2_x86inc";
903
904add_proto qw/unsigned int vpx_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
905specialize qw/vpx_sad8x8_avg msa/, "$sse2_x86inc";
906
907add_proto qw/unsigned int vpx_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
908specialize qw/vpx_sad8x4_avg msa/, "$sse2_x86inc";
909
910add_proto qw/unsigned int vpx_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
911specialize qw/vpx_sad4x8_avg msa/, "$sse_x86inc";
912
913add_proto qw/unsigned int vpx_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
914specialize qw/vpx_sad4x4_avg msa/, "$sse_x86inc";
915
916#
917# Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
918#
919# Blocks of 3
920add_proto qw/void vpx_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
921specialize qw/vpx_sad64x64x3 msa/;
922
923add_proto qw/void vpx_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
924specialize qw/vpx_sad32x32x3 msa/;
925
926add_proto qw/void vpx_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
927specialize qw/vpx_sad16x16x3 sse3 ssse3 msa/;
928
929add_proto qw/void vpx_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
930specialize qw/vpx_sad16x8x3 sse3 ssse3 msa/;
931
932add_proto qw/void vpx_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
933specialize qw/vpx_sad8x16x3 sse3 msa/;
934
935add_proto qw/void vpx_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
936specialize qw/vpx_sad8x8x3 sse3 msa/;
937
938add_proto qw/void vpx_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
939specialize qw/vpx_sad4x4x3 sse3 msa/;
940
941# Blocks of 8
942add_proto qw/void vpx_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
943specialize qw/vpx_sad64x64x8 msa/;
944
945add_proto qw/void vpx_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
946specialize qw/vpx_sad32x32x8 msa/;
947
948add_proto qw/void vpx_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
949specialize qw/vpx_sad16x16x8 sse4_1 msa/;
950
951add_proto qw/void vpx_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
952specialize qw/vpx_sad16x8x8 sse4_1 msa/;
953
954add_proto qw/void vpx_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
955specialize qw/vpx_sad8x16x8 sse4_1 msa/;
956
957add_proto qw/void vpx_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
958specialize qw/vpx_sad8x8x8 sse4_1 msa/;
959
960add_proto qw/void vpx_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
961specialize qw/vpx_sad8x4x8 msa/;
962
963add_proto qw/void vpx_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
964specialize qw/vpx_sad4x8x8 msa/;
965
966add_proto qw/void vpx_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
967specialize qw/vpx_sad4x4x8 sse4_1 msa/;
968
969#
970# Multi-block SAD, comparing a reference to N independent blocks
971#
972add_proto qw/void vpx_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
973specialize qw/vpx_sad64x64x4d avx2 neon msa/, "$sse2_x86inc";
974
975add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
976specialize qw/vpx_sad64x32x4d msa/, "$sse2_x86inc";
977
978add_proto qw/void vpx_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
979specialize qw/vpx_sad32x64x4d msa/, "$sse2_x86inc";
980
981add_proto qw/void vpx_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
982specialize qw/vpx_sad32x32x4d avx2 neon msa/, "$sse2_x86inc";
983
984add_proto qw/void vpx_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
985specialize qw/vpx_sad32x16x4d msa/, "$sse2_x86inc";
986
987add_proto qw/void vpx_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
988specialize qw/vpx_sad16x32x4d msa/, "$sse2_x86inc";
989
990add_proto qw/void vpx_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
991specialize qw/vpx_sad16x16x4d neon msa/, "$sse2_x86inc";
992
993add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
994specialize qw/vpx_sad16x8x4d msa/, "$sse2_x86inc";
995
996add_proto qw/void vpx_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
997specialize qw/vpx_sad8x16x4d msa/, "$sse2_x86inc";
998
999add_proto qw/void vpx_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1000specialize qw/vpx_sad8x8x4d msa/, "$sse2_x86inc";
1001
1002add_proto qw/void vpx_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1003specialize qw/vpx_sad8x4x4d msa/, "$sse2_x86inc";
1004
1005add_proto qw/void vpx_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1006specialize qw/vpx_sad4x8x4d msa/, "$sse_x86inc";
1007
1008add_proto qw/void vpx_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
1009specialize qw/vpx_sad4x4x4d msa/, "$sse_x86inc";
1010
1011#
1012# Structured Similarity (SSIM)
1013#
1014if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
1015    add_proto qw/void vpx_ssim_parms_8x8/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
1016    specialize qw/vpx_ssim_parms_8x8/, "$sse2_x86_64";
1017
1018    add_proto qw/void vpx_ssim_parms_16x16/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
1019    specialize qw/vpx_ssim_parms_16x16/, "$sse2_x86_64";
1020}
1021
1022if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
1023  #
1024  # Block subtraction
1025  #
1026  add_proto qw/void vpx_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd";
1027  specialize qw/vpx_highbd_subtract_block/;
1028
1029  #
1030  # Single block SAD
1031  #
1032  add_proto qw/unsigned int vpx_highbd_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1033  specialize qw/vpx_highbd_sad64x64/, "$sse2_x86inc";
1034
1035  add_proto qw/unsigned int vpx_highbd_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1036  specialize qw/vpx_highbd_sad64x32/, "$sse2_x86inc";
1037
1038  add_proto qw/unsigned int vpx_highbd_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1039  specialize qw/vpx_highbd_sad32x64/, "$sse2_x86inc";
1040
1041  add_proto qw/unsigned int vpx_highbd_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1042  specialize qw/vpx_highbd_sad32x32/, "$sse2_x86inc";
1043
1044  add_proto qw/unsigned int vpx_highbd_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1045  specialize qw/vpx_highbd_sad32x16/, "$sse2_x86inc";
1046
1047  add_proto qw/unsigned int vpx_highbd_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1048  specialize qw/vpx_highbd_sad16x32/, "$sse2_x86inc";
1049
1050  add_proto qw/unsigned int vpx_highbd_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1051  specialize qw/vpx_highbd_sad16x16/, "$sse2_x86inc";
1052
1053  add_proto qw/unsigned int vpx_highbd_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1054  specialize qw/vpx_highbd_sad16x8/, "$sse2_x86inc";
1055
1056  add_proto qw/unsigned int vpx_highbd_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1057  specialize qw/vpx_highbd_sad8x16/, "$sse2_x86inc";
1058
1059  add_proto qw/unsigned int vpx_highbd_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1060  specialize qw/vpx_highbd_sad8x8/, "$sse2_x86inc";
1061
1062  add_proto qw/unsigned int vpx_highbd_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1063  specialize qw/vpx_highbd_sad8x4/, "$sse2_x86inc";
1064
1065  add_proto qw/unsigned int vpx_highbd_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1066  specialize qw/vpx_highbd_sad4x8/;
1067
1068  add_proto qw/unsigned int vpx_highbd_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
1069  specialize qw/vpx_highbd_sad4x4/;
1070
1071  #
1072  # Avg
1073  #
1074  add_proto qw/unsigned int vpx_highbd_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1075  specialize qw/vpx_highbd_sad64x64_avg/, "$sse2_x86inc";
1076
1077  add_proto qw/unsigned int vpx_highbd_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1078  specialize qw/vpx_highbd_sad64x32_avg/, "$sse2_x86inc";
1079
1080  add_proto qw/unsigned int vpx_highbd_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1081  specialize qw/vpx_highbd_sad32x64_avg/, "$sse2_x86inc";
1082
1083  add_proto qw/unsigned int vpx_highbd_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1084  specialize qw/vpx_highbd_sad32x32_avg/, "$sse2_x86inc";
1085
1086  add_proto qw/unsigned int vpx_highbd_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1087  specialize qw/vpx_highbd_sad32x16_avg/, "$sse2_x86inc";
1088
1089  add_proto qw/unsigned int vpx_highbd_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1090  specialize qw/vpx_highbd_sad16x32_avg/, "$sse2_x86inc";
1091
1092  add_proto qw/unsigned int vpx_highbd_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1093  specialize qw/vpx_highbd_sad16x16_avg/, "$sse2_x86inc";
1094
1095  add_proto qw/unsigned int vpx_highbd_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1096  specialize qw/vpx_highbd_sad16x8_avg/, "$sse2_x86inc";
1097
1098  add_proto qw/unsigned int vpx_highbd_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1099  specialize qw/vpx_highbd_sad8x16_avg/, "$sse2_x86inc";
1100
1101  add_proto qw/unsigned int vpx_highbd_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1102  specialize qw/vpx_highbd_sad8x8_avg/, "$sse2_x86inc";
1103
1104  add_proto qw/unsigned int vpx_highbd_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1105  specialize qw/vpx_highbd_sad8x4_avg/, "$sse2_x86inc";
1106
1107  add_proto qw/unsigned int vpx_highbd_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1108  specialize qw/vpx_highbd_sad4x8_avg/;
1109
1110  add_proto qw/unsigned int vpx_highbd_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
1111  specialize qw/vpx_highbd_sad4x4_avg/;
1112
1113  #
1114  # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
1115  #
1116  # Blocks of 3
1117  add_proto qw/void vpx_highbd_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1118  specialize qw/vpx_highbd_sad64x64x3/;
1119
1120  add_proto qw/void vpx_highbd_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1121  specialize qw/vpx_highbd_sad32x32x3/;
1122
1123  add_proto qw/void vpx_highbd_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1124  specialize qw/vpx_highbd_sad16x16x3/;
1125
1126  add_proto qw/void vpx_highbd_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1127  specialize qw/vpx_highbd_sad16x8x3/;
1128
1129  add_proto qw/void vpx_highbd_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1130  specialize qw/vpx_highbd_sad8x16x3/;
1131
1132  add_proto qw/void vpx_highbd_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1133  specialize qw/vpx_highbd_sad8x8x3/;
1134
1135  add_proto qw/void vpx_highbd_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1136  specialize qw/vpx_highbd_sad4x4x3/;
1137
1138  # Blocks of 8
1139  add_proto qw/void vpx_highbd_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1140  specialize qw/vpx_highbd_sad64x64x8/;
1141
1142  add_proto qw/void vpx_highbd_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1143  specialize qw/vpx_highbd_sad32x32x8/;
1144
1145  add_proto qw/void vpx_highbd_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1146  specialize qw/vpx_highbd_sad16x16x8/;
1147
1148  add_proto qw/void vpx_highbd_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1149  specialize qw/vpx_highbd_sad16x8x8/;
1150
1151  add_proto qw/void vpx_highbd_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1152  specialize qw/vpx_highbd_sad8x16x8/;
1153
1154  add_proto qw/void vpx_highbd_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1155  specialize qw/vpx_highbd_sad8x8x8/;
1156
1157  add_proto qw/void vpx_highbd_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1158  specialize qw/vpx_highbd_sad8x4x8/;
1159
1160  add_proto qw/void vpx_highbd_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1161  specialize qw/vpx_highbd_sad4x8x8/;
1162
1163  add_proto qw/void vpx_highbd_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
1164  specialize qw/vpx_highbd_sad4x4x8/;
1165
1166  #
1167  # Multi-block SAD, comparing a reference to N independent blocks
1168  #
1169  add_proto qw/void vpx_highbd_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1170  specialize qw/vpx_highbd_sad64x64x4d/, "$sse2_x86inc";
1171
1172  add_proto qw/void vpx_highbd_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1173  specialize qw/vpx_highbd_sad64x32x4d/, "$sse2_x86inc";
1174
1175  add_proto qw/void vpx_highbd_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1176  specialize qw/vpx_highbd_sad32x64x4d/, "$sse2_x86inc";
1177
1178  add_proto qw/void vpx_highbd_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1179  specialize qw/vpx_highbd_sad32x32x4d/, "$sse2_x86inc";
1180
1181  add_proto qw/void vpx_highbd_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1182  specialize qw/vpx_highbd_sad32x16x4d/, "$sse2_x86inc";
1183
1184  add_proto qw/void vpx_highbd_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1185  specialize qw/vpx_highbd_sad16x32x4d/, "$sse2_x86inc";
1186
1187  add_proto qw/void vpx_highbd_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1188  specialize qw/vpx_highbd_sad16x16x4d/, "$sse2_x86inc";
1189
1190  add_proto qw/void vpx_highbd_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1191  specialize qw/vpx_highbd_sad16x8x4d/, "$sse2_x86inc";
1192
1193  add_proto qw/void vpx_highbd_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1194  specialize qw/vpx_highbd_sad8x16x4d/, "$sse2_x86inc";
1195
1196  add_proto qw/void vpx_highbd_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1197  specialize qw/vpx_highbd_sad8x8x4d/, "$sse2_x86inc";
1198
1199  add_proto qw/void vpx_highbd_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1200  specialize qw/vpx_highbd_sad8x4x4d/, "$sse2_x86inc";
1201
1202  add_proto qw/void vpx_highbd_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1203  specialize qw/vpx_highbd_sad4x8x4d/, "$sse2_x86inc";
1204
1205  add_proto qw/void vpx_highbd_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
1206  specialize qw/vpx_highbd_sad4x4x4d/, "$sse2_x86inc";
1207
1208  #
1209  # Structured Similarity (SSIM)
1210  #
1211  if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
1212    add_proto qw/void vpx_highbd_ssim_parms_8x8/, "const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
1213    specialize qw/vpx_highbd_ssim_parms_8x8/;
1214  }
1215}  # CONFIG_VP9_HIGHBITDEPTH
1216}  # CONFIG_ENCODERS
1217
1218if (vpx_config("CONFIG_ENCODERS") eq "yes" || vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
1219
1220#
1221# Variance
1222#
1223add_proto qw/unsigned int vpx_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1224  specialize qw/vpx_variance64x64 sse2 avx2 neon msa/;
1225
1226add_proto qw/unsigned int vpx_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1227  specialize qw/vpx_variance64x32 sse2 avx2 neon msa/;
1228
1229add_proto qw/unsigned int vpx_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1230  specialize qw/vpx_variance32x64 sse2 neon msa/;
1231
1232add_proto qw/unsigned int vpx_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1233  specialize qw/vpx_variance32x32 sse2 avx2 neon msa/;
1234
1235add_proto qw/unsigned int vpx_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1236  specialize qw/vpx_variance32x16 sse2 avx2 msa/;
1237
1238add_proto qw/unsigned int vpx_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1239  specialize qw/vpx_variance16x32 sse2 msa/;
1240
1241add_proto qw/unsigned int vpx_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1242  specialize qw/vpx_variance16x16 mmx sse2 avx2 media neon msa/;
1243
1244add_proto qw/unsigned int vpx_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1245  specialize qw/vpx_variance16x8 mmx sse2 neon msa/;
1246
1247add_proto qw/unsigned int vpx_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1248  specialize qw/vpx_variance8x16 mmx sse2 neon msa/;
1249
1250add_proto qw/unsigned int vpx_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1251  specialize qw/vpx_variance8x8 mmx sse2 media neon msa/;
1252
1253add_proto qw/unsigned int vpx_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1254  specialize qw/vpx_variance8x4 sse2 msa/;
1255
1256add_proto qw/unsigned int vpx_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1257  specialize qw/vpx_variance4x8 sse2 msa/;
1258
1259add_proto qw/unsigned int vpx_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1260  specialize qw/vpx_variance4x4 mmx sse2 msa/;
1261
1262#
1263# Specialty Variance
1264#
1265add_proto qw/void vpx_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1266  specialize qw/vpx_get16x16var sse2 avx2 neon msa/;
1267
1268add_proto qw/void vpx_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1269  specialize qw/vpx_get8x8var mmx sse2 neon msa/;
1270
1271add_proto qw/unsigned int vpx_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1272  specialize qw/vpx_mse16x16 mmx sse2 avx2 media neon msa/;
1273
1274add_proto qw/unsigned int vpx_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1275  specialize qw/vpx_mse16x8 sse2 msa/;
1276
1277add_proto qw/unsigned int vpx_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1278  specialize qw/vpx_mse8x16 sse2 msa/;
1279
1280add_proto qw/unsigned int vpx_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1281  specialize qw/vpx_mse8x8 sse2 msa/;
1282
1283add_proto qw/unsigned int vpx_get_mb_ss/, "const int16_t *";
1284  specialize qw/vpx_get_mb_ss mmx sse2 msa/;
1285
1286add_proto qw/unsigned int vpx_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride";
1287  specialize qw/vpx_get4x4sse_cs neon msa/;
1288
1289add_proto qw/void vpx_comp_avg_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
1290
1291#
1292# Subpixel Variance
1293#
1294add_proto qw/uint32_t vpx_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1295  specialize qw/vpx_sub_pixel_variance64x64 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1296
1297add_proto qw/uint32_t vpx_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1298  specialize qw/vpx_sub_pixel_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1299
1300add_proto qw/uint32_t vpx_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1301  specialize qw/vpx_sub_pixel_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1302
1303add_proto qw/uint32_t vpx_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1304  specialize qw/vpx_sub_pixel_variance32x32 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1305
1306add_proto qw/uint32_t vpx_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1307  specialize qw/vpx_sub_pixel_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1308
1309add_proto qw/uint32_t vpx_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1310  specialize qw/vpx_sub_pixel_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1311
1312add_proto qw/uint32_t vpx_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1313  specialize qw/vpx_sub_pixel_variance16x16 mmx media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1314
1315add_proto qw/uint32_t vpx_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1316  specialize qw/vpx_sub_pixel_variance16x8 mmx msa/, "$sse2_x86inc", "$ssse3_x86inc";
1317
1318add_proto qw/uint32_t vpx_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1319  specialize qw/vpx_sub_pixel_variance8x16 mmx msa/, "$sse2_x86inc", "$ssse3_x86inc";
1320
1321add_proto qw/uint32_t vpx_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1322  specialize qw/vpx_sub_pixel_variance8x8 mmx media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
1323
1324add_proto qw/uint32_t vpx_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1325  specialize qw/vpx_sub_pixel_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1326
1327add_proto qw/uint32_t vpx_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1328  specialize qw/vpx_sub_pixel_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
1329
1330add_proto qw/uint32_t vpx_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1331  specialize qw/vpx_sub_pixel_variance4x4 mmx msa/, "$sse_x86inc", "$ssse3_x86inc";
1332
1333add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1334  specialize qw/vpx_sub_pixel_avg_variance64x64 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1335
1336add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1337  specialize qw/vpx_sub_pixel_avg_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1338
1339add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1340  specialize qw/vpx_sub_pixel_avg_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1341
1342add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1343  specialize qw/vpx_sub_pixel_avg_variance32x32 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1344
1345add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1346  specialize qw/vpx_sub_pixel_avg_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1347
1348add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1349  specialize qw/vpx_sub_pixel_avg_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1350
1351add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1352  specialize qw/vpx_sub_pixel_avg_variance16x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1353
1354add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1355  specialize qw/vpx_sub_pixel_avg_variance16x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1356
1357add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1358  specialize qw/vpx_sub_pixel_avg_variance8x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1359
1360add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1361  specialize qw/vpx_sub_pixel_avg_variance8x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1362
1363add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1364  specialize qw/vpx_sub_pixel_avg_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
1365
1366add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1367  specialize qw/vpx_sub_pixel_avg_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
1368
1369add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1370  specialize qw/vpx_sub_pixel_avg_variance4x4 msa/, "$sse_x86inc", "$ssse3_x86inc";
1371
1372#
1373# Specialty Subpixel
1374#
1375add_proto qw/uint32_t vpx_variance_halfpixvar16x16_h/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
1376  specialize qw/vpx_variance_halfpixvar16x16_h mmx media/;
1377
1378add_proto qw/uint32_t vpx_variance_halfpixvar16x16_v/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
1379  specialize qw/vpx_variance_halfpixvar16x16_v mmx media/;
1380
1381add_proto qw/uint32_t vpx_variance_halfpixvar16x16_hv/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
1382  specialize qw/vpx_variance_halfpixvar16x16_hv mmx media/;
1383
1384if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
1385  add_proto qw/unsigned int vpx_highbd_12_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1386  specialize qw/vpx_highbd_12_variance64x64 sse2/;
1387
1388  add_proto qw/unsigned int vpx_highbd_12_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1389  specialize qw/vpx_highbd_12_variance64x32 sse2/;
1390
1391  add_proto qw/unsigned int vpx_highbd_12_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1392  specialize qw/vpx_highbd_12_variance32x64 sse2/;
1393
1394  add_proto qw/unsigned int vpx_highbd_12_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1395  specialize qw/vpx_highbd_12_variance32x32 sse2/;
1396
1397  add_proto qw/unsigned int vpx_highbd_12_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1398  specialize qw/vpx_highbd_12_variance32x16 sse2/;
1399
1400  add_proto qw/unsigned int vpx_highbd_12_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1401  specialize qw/vpx_highbd_12_variance16x32 sse2/;
1402
1403  add_proto qw/unsigned int vpx_highbd_12_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1404  specialize qw/vpx_highbd_12_variance16x16 sse2/;
1405
1406  add_proto qw/unsigned int vpx_highbd_12_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1407  specialize qw/vpx_highbd_12_variance16x8 sse2/;
1408
1409  add_proto qw/unsigned int vpx_highbd_12_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1410  specialize qw/vpx_highbd_12_variance8x16 sse2/;
1411
1412  add_proto qw/unsigned int vpx_highbd_12_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1413  specialize qw/vpx_highbd_12_variance8x8 sse2/;
1414
1415  add_proto qw/unsigned int vpx_highbd_12_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1416  add_proto qw/unsigned int vpx_highbd_12_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1417  add_proto qw/unsigned int vpx_highbd_12_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1418
1419  add_proto qw/unsigned int vpx_highbd_10_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1420  specialize qw/vpx_highbd_10_variance64x64 sse2/;
1421
1422  add_proto qw/unsigned int vpx_highbd_10_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1423  specialize qw/vpx_highbd_10_variance64x32 sse2/;
1424
1425  add_proto qw/unsigned int vpx_highbd_10_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1426  specialize qw/vpx_highbd_10_variance32x64 sse2/;
1427
1428  add_proto qw/unsigned int vpx_highbd_10_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1429  specialize qw/vpx_highbd_10_variance32x32 sse2/;
1430
1431  add_proto qw/unsigned int vpx_highbd_10_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1432  specialize qw/vpx_highbd_10_variance32x16 sse2/;
1433
1434  add_proto qw/unsigned int vpx_highbd_10_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1435  specialize qw/vpx_highbd_10_variance16x32 sse2/;
1436
1437  add_proto qw/unsigned int vpx_highbd_10_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1438  specialize qw/vpx_highbd_10_variance16x16 sse2/;
1439
1440  add_proto qw/unsigned int vpx_highbd_10_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1441  specialize qw/vpx_highbd_10_variance16x8 sse2/;
1442
1443  add_proto qw/unsigned int vpx_highbd_10_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1444  specialize qw/vpx_highbd_10_variance8x16 sse2/;
1445
1446  add_proto qw/unsigned int vpx_highbd_10_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1447  specialize qw/vpx_highbd_10_variance8x8 sse2/;
1448
1449  add_proto qw/unsigned int vpx_highbd_10_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1450  add_proto qw/unsigned int vpx_highbd_10_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1451  add_proto qw/unsigned int vpx_highbd_10_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1452
1453  add_proto qw/unsigned int vpx_highbd_8_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1454  specialize qw/vpx_highbd_8_variance64x64 sse2/;
1455
1456  add_proto qw/unsigned int vpx_highbd_8_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1457  specialize qw/vpx_highbd_8_variance64x32 sse2/;
1458
1459  add_proto qw/unsigned int vpx_highbd_8_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1460  specialize qw/vpx_highbd_8_variance32x64 sse2/;
1461
1462  add_proto qw/unsigned int vpx_highbd_8_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1463  specialize qw/vpx_highbd_8_variance32x32 sse2/;
1464
1465  add_proto qw/unsigned int vpx_highbd_8_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1466  specialize qw/vpx_highbd_8_variance32x16 sse2/;
1467
1468  add_proto qw/unsigned int vpx_highbd_8_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1469  specialize qw/vpx_highbd_8_variance16x32 sse2/;
1470
1471  add_proto qw/unsigned int vpx_highbd_8_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1472  specialize qw/vpx_highbd_8_variance16x16 sse2/;
1473
1474  add_proto qw/unsigned int vpx_highbd_8_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1475  specialize qw/vpx_highbd_8_variance16x8 sse2/;
1476
1477  add_proto qw/unsigned int vpx_highbd_8_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1478  specialize qw/vpx_highbd_8_variance8x16 sse2/;
1479
1480  add_proto qw/unsigned int vpx_highbd_8_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1481  specialize qw/vpx_highbd_8_variance8x8 sse2/;
1482
1483  add_proto qw/unsigned int vpx_highbd_8_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1484  add_proto qw/unsigned int vpx_highbd_8_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1485  add_proto qw/unsigned int vpx_highbd_8_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
1486
1487  add_proto qw/void vpx_highbd_8_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1488  add_proto qw/void vpx_highbd_8_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1489
1490  add_proto qw/void vpx_highbd_10_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1491  add_proto qw/void vpx_highbd_10_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1492
1493  add_proto qw/void vpx_highbd_12_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1494  add_proto qw/void vpx_highbd_12_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
1495
1496  add_proto qw/unsigned int vpx_highbd_8_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1497  specialize qw/vpx_highbd_8_mse16x16 sse2/;
1498
1499  add_proto qw/unsigned int vpx_highbd_8_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1500  add_proto qw/unsigned int vpx_highbd_8_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1501  add_proto qw/unsigned int vpx_highbd_8_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1502  specialize qw/vpx_highbd_8_mse8x8 sse2/;
1503
1504  add_proto qw/unsigned int vpx_highbd_10_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1505  specialize qw/vpx_highbd_10_mse16x16 sse2/;
1506
1507  add_proto qw/unsigned int vpx_highbd_10_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1508  add_proto qw/unsigned int vpx_highbd_10_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1509  add_proto qw/unsigned int vpx_highbd_10_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1510  specialize qw/vpx_highbd_10_mse8x8 sse2/;
1511
1512  add_proto qw/unsigned int vpx_highbd_12_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1513  specialize qw/vpx_highbd_12_mse16x16 sse2/;
1514
1515  add_proto qw/unsigned int vpx_highbd_12_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1516  add_proto qw/unsigned int vpx_highbd_12_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1517  add_proto qw/unsigned int vpx_highbd_12_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
1518  specialize qw/vpx_highbd_12_mse8x8 sse2/;
1519
1520  add_proto qw/void vpx_highbd_comp_avg_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride";
1521
1522  #
1523  # Subpixel Variance
1524  #
1525  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1526  specialize qw/vpx_highbd_12_sub_pixel_variance64x64/, "$sse2_x86inc";
1527
1528  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1529  specialize qw/vpx_highbd_12_sub_pixel_variance64x32/, "$sse2_x86inc";
1530
1531  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1532  specialize qw/vpx_highbd_12_sub_pixel_variance32x64/, "$sse2_x86inc";
1533
1534  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1535  specialize qw/vpx_highbd_12_sub_pixel_variance32x32/, "$sse2_x86inc";
1536
1537  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1538  specialize qw/vpx_highbd_12_sub_pixel_variance32x16/, "$sse2_x86inc";
1539
1540  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1541  specialize qw/vpx_highbd_12_sub_pixel_variance16x32/, "$sse2_x86inc";
1542
1543  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1544  specialize qw/vpx_highbd_12_sub_pixel_variance16x16/, "$sse2_x86inc";
1545
1546  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1547  specialize qw/vpx_highbd_12_sub_pixel_variance16x8/, "$sse2_x86inc";
1548
1549  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1550  specialize qw/vpx_highbd_12_sub_pixel_variance8x16/, "$sse2_x86inc";
1551
1552  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1553  specialize qw/vpx_highbd_12_sub_pixel_variance8x8/, "$sse2_x86inc";
1554
1555  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1556  specialize qw/vpx_highbd_12_sub_pixel_variance8x4/, "$sse2_x86inc";
1557
1558  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1559  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1560
1561  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1562  specialize qw/vpx_highbd_10_sub_pixel_variance64x64/, "$sse2_x86inc";
1563
1564  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1565  specialize qw/vpx_highbd_10_sub_pixel_variance64x32/, "$sse2_x86inc";
1566
1567  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1568  specialize qw/vpx_highbd_10_sub_pixel_variance32x64/, "$sse2_x86inc";
1569
1570  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1571  specialize qw/vpx_highbd_10_sub_pixel_variance32x32/, "$sse2_x86inc";
1572
1573  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1574  specialize qw/vpx_highbd_10_sub_pixel_variance32x16/, "$sse2_x86inc";
1575
1576  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1577  specialize qw/vpx_highbd_10_sub_pixel_variance16x32/, "$sse2_x86inc";
1578
1579  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1580  specialize qw/vpx_highbd_10_sub_pixel_variance16x16/, "$sse2_x86inc";
1581
1582  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1583  specialize qw/vpx_highbd_10_sub_pixel_variance16x8/, "$sse2_x86inc";
1584
1585  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1586  specialize qw/vpx_highbd_10_sub_pixel_variance8x16/, "$sse2_x86inc";
1587
1588  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1589  specialize qw/vpx_highbd_10_sub_pixel_variance8x8/, "$sse2_x86inc";
1590
1591  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1592  specialize qw/vpx_highbd_10_sub_pixel_variance8x4/, "$sse2_x86inc";
1593
1594  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1595  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1596
1597  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1598  specialize qw/vpx_highbd_8_sub_pixel_variance64x64/, "$sse2_x86inc";
1599
1600  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1601  specialize qw/vpx_highbd_8_sub_pixel_variance64x32/, "$sse2_x86inc";
1602
1603  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1604  specialize qw/vpx_highbd_8_sub_pixel_variance32x64/, "$sse2_x86inc";
1605
1606  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1607  specialize qw/vpx_highbd_8_sub_pixel_variance32x32/, "$sse2_x86inc";
1608
1609  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1610  specialize qw/vpx_highbd_8_sub_pixel_variance32x16/, "$sse2_x86inc";
1611
1612  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1613  specialize qw/vpx_highbd_8_sub_pixel_variance16x32/, "$sse2_x86inc";
1614
1615  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1616  specialize qw/vpx_highbd_8_sub_pixel_variance16x16/, "$sse2_x86inc";
1617
1618  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1619  specialize qw/vpx_highbd_8_sub_pixel_variance16x8/, "$sse2_x86inc";
1620
1621  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1622  specialize qw/vpx_highbd_8_sub_pixel_variance8x16/, "$sse2_x86inc";
1623
1624  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1625  specialize qw/vpx_highbd_8_sub_pixel_variance8x8/, "$sse2_x86inc";
1626
1627  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1628  specialize qw/vpx_highbd_8_sub_pixel_variance8x4/, "$sse2_x86inc";
1629
1630  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1631  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
1632
1633  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1634  specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1635
1636  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1637  specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1638
1639  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1640  specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1641
1642  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1643  specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1644
1645  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1646  specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1647
1648  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1649  specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1650
1651  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1652  specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1653
1654  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1655  specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1656
1657  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1658  specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1659
1660  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1661  specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1662
1663  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1664  specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1665
1666  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1667  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1668
1669  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1670  specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1671
1672  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1673  specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1674
1675  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1676  specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1677
1678  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1679  specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1680
1681  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1682  specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1683
1684  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1685  specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1686
1687  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1688  specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1689
1690  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1691  specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1692
1693  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1694  specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1695
1696  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1697  specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1698
1699  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1700  specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1701
1702  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1703  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1704
1705  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1706  specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
1707
1708  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1709  specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
1710
1711  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1712  specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
1713
1714  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1715  specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
1716
1717  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1718  specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
1719
1720  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1721  specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
1722
1723  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1724  specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
1725
1726  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1727  specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
1728
1729  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1730  specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
1731
1732  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1733  specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
1734
1735  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1736  specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
1737
1738  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1739  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
1740
1741}  # CONFIG_VP9_HIGHBITDEPTH
1742}  # CONFIG_ENCODERS || CONFIG_POSTPROC || CONFIG_VP9_POSTPROC
1743
17441;
1745