1sub vp8_common_forward_decls() {
2print <<EOF
3/*
4 * VP8
5 */
6
7struct blockd;
8struct macroblockd;
9struct loop_filter_info;
10
11/* Encoder forward decls */
12struct block;
13struct macroblock;
14struct variance_vtable;
15union int_mv;
16struct yv12_buffer_config;
17EOF
18}
19forward_decls qw/vp8_common_forward_decls/;
20
21#
22# system state
23#
24add_proto qw/void vp8_clear_system_state/, "";
25specialize qw/vp8_clear_system_state mmx/;
26$vp8_clear_system_state_mmx=vpx_reset_mmx_state;
27
28#
29# Dequant
30#
31add_proto qw/void vp8_dequantize_b/, "struct blockd*, short *dqc";
32specialize qw/vp8_dequantize_b mmx media neon/;
33$vp8_dequantize_b_media=vp8_dequantize_b_v6;
34
35add_proto qw/void vp8_dequant_idct_add/, "short *input, short *dq, unsigned char *output, int stride";
36specialize qw/vp8_dequant_idct_add mmx media neon dspr2/;
37$vp8_dequant_idct_add_media=vp8_dequant_idct_add_v6;
38$vp8_dequant_idct_add_dspr2=vp8_dequant_idct_add_dspr2;
39
40add_proto qw/void vp8_dequant_idct_add_y_block/, "short *q, short *dq, unsigned char *dst, int stride, char *eobs";
41specialize qw/vp8_dequant_idct_add_y_block mmx sse2 media neon dspr2/;
42$vp8_dequant_idct_add_y_block_media=vp8_dequant_idct_add_y_block_v6;
43$vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2;
44
45add_proto qw/void vp8_dequant_idct_add_uv_block/, "short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs";
46specialize qw/vp8_dequant_idct_add_uv_block mmx sse2 media neon dspr2/;
47$vp8_dequant_idct_add_uv_block_media=vp8_dequant_idct_add_uv_block_v6;
48$vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2;
49
50#
51# Loopfilter
52#
53add_proto qw/void vp8_loop_filter_mbv/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
54specialize qw/vp8_loop_filter_mbv mmx sse2 media neon dspr2/;
55$vp8_loop_filter_mbv_media=vp8_loop_filter_mbv_armv6;
56$vp8_loop_filter_mbv_dspr2=vp8_loop_filter_mbv_dspr2;
57
58add_proto qw/void vp8_loop_filter_bv/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
59specialize qw/vp8_loop_filter_bv mmx sse2 media neon dspr2/;
60$vp8_loop_filter_bv_media=vp8_loop_filter_bv_armv6;
61$vp8_loop_filter_bv_dspr2=vp8_loop_filter_bv_dspr2;
62
63add_proto qw/void vp8_loop_filter_mbh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
64specialize qw/vp8_loop_filter_mbh mmx sse2 media neon dspr2/;
65$vp8_loop_filter_mbh_media=vp8_loop_filter_mbh_armv6;
66$vp8_loop_filter_mbh_dspr2=vp8_loop_filter_mbh_dspr2;
67
68add_proto qw/void vp8_loop_filter_bh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
69specialize qw/vp8_loop_filter_bh mmx sse2 media neon dspr2/;
70$vp8_loop_filter_bh_media=vp8_loop_filter_bh_armv6;
71$vp8_loop_filter_bh_dspr2=vp8_loop_filter_bh_dspr2;
72
73
74add_proto qw/void vp8_loop_filter_simple_mbv/, "unsigned char *y, int ystride, const unsigned char *blimit";
75specialize qw/vp8_loop_filter_simple_mbv mmx sse2 media neon/;
76$vp8_loop_filter_simple_mbv_c=vp8_loop_filter_simple_vertical_edge_c;
77$vp8_loop_filter_simple_mbv_mmx=vp8_loop_filter_simple_vertical_edge_mmx;
78$vp8_loop_filter_simple_mbv_sse2=vp8_loop_filter_simple_vertical_edge_sse2;
79$vp8_loop_filter_simple_mbv_media=vp8_loop_filter_simple_vertical_edge_armv6;
80$vp8_loop_filter_simple_mbv_neon=vp8_loop_filter_mbvs_neon;
81
82add_proto qw/void vp8_loop_filter_simple_mbh/, "unsigned char *y, int ystride, const unsigned char *blimit";
83specialize qw/vp8_loop_filter_simple_mbh mmx sse2 media neon/;
84$vp8_loop_filter_simple_mbh_c=vp8_loop_filter_simple_horizontal_edge_c;
85$vp8_loop_filter_simple_mbh_mmx=vp8_loop_filter_simple_horizontal_edge_mmx;
86$vp8_loop_filter_simple_mbh_sse2=vp8_loop_filter_simple_horizontal_edge_sse2;
87$vp8_loop_filter_simple_mbh_media=vp8_loop_filter_simple_horizontal_edge_armv6;
88$vp8_loop_filter_simple_mbh_neon=vp8_loop_filter_mbhs_neon;
89
90add_proto qw/void vp8_loop_filter_simple_bv/, "unsigned char *y, int ystride, const unsigned char *blimit";
91specialize qw/vp8_loop_filter_simple_bv mmx sse2 media neon/;
92$vp8_loop_filter_simple_bv_c=vp8_loop_filter_bvs_c;
93$vp8_loop_filter_simple_bv_mmx=vp8_loop_filter_bvs_mmx;
94$vp8_loop_filter_simple_bv_sse2=vp8_loop_filter_bvs_sse2;
95$vp8_loop_filter_simple_bv_media=vp8_loop_filter_bvs_armv6;
96$vp8_loop_filter_simple_bv_neon=vp8_loop_filter_bvs_neon;
97
98add_proto qw/void vp8_loop_filter_simple_bh/, "unsigned char *y, int ystride, const unsigned char *blimit";
99specialize qw/vp8_loop_filter_simple_bh mmx sse2 media neon/;
100$vp8_loop_filter_simple_bh_c=vp8_loop_filter_bhs_c;
101$vp8_loop_filter_simple_bh_mmx=vp8_loop_filter_bhs_mmx;
102$vp8_loop_filter_simple_bh_sse2=vp8_loop_filter_bhs_sse2;
103$vp8_loop_filter_simple_bh_media=vp8_loop_filter_bhs_armv6;
104$vp8_loop_filter_simple_bh_neon=vp8_loop_filter_bhs_neon;
105
106#
107# IDCT
108#
109#idct16
110add_proto qw/void vp8_short_idct4x4llm/, "short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride";
111specialize qw/vp8_short_idct4x4llm mmx media neon dspr2/;
112$vp8_short_idct4x4llm_media=vp8_short_idct4x4llm_v6_dual;
113$vp8_short_idct4x4llm_dspr2=vp8_short_idct4x4llm_dspr2;
114
115#iwalsh1
116add_proto qw/void vp8_short_inv_walsh4x4_1/, "short *input, short *output";
117specialize qw/vp8_short_inv_walsh4x4_1 dspr2/;
118$vp8_short_inv_walsh4x4_1_dspr2=vp8_short_inv_walsh4x4_1_dspr2;
119# no asm yet
120
121#iwalsh16
122add_proto qw/void vp8_short_inv_walsh4x4/, "short *input, short *output";
123specialize qw/vp8_short_inv_walsh4x4 mmx sse2 media neon dspr2/;
124$vp8_short_inv_walsh4x4_media=vp8_short_inv_walsh4x4_v6;
125$vp8_short_inv_walsh4x4_dspr2=vp8_short_inv_walsh4x4_dspr2;
126
127#idct1_scalar_add
128add_proto qw/void vp8_dc_only_idct_add/, "short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride";
129specialize qw/vp8_dc_only_idct_add	mmx media neon dspr2/;
130$vp8_dc_only_idct_add_media=vp8_dc_only_idct_add_v6;
131$vp8_dc_only_idct_add_dspr2=vp8_dc_only_idct_add_dspr2;
132
133#
134# RECON
135#
136add_proto qw/void vp8_copy_mem16x16/, "unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch";
137specialize qw/vp8_copy_mem16x16 mmx sse2 media neon dspr2/;
138$vp8_copy_mem16x16_media=vp8_copy_mem16x16_v6;
139$vp8_copy_mem16x16_dspr2=vp8_copy_mem16x16_dspr2;
140
141add_proto qw/void vp8_copy_mem8x8/, "unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch";
142specialize qw/vp8_copy_mem8x8 mmx media neon dspr2/;
143$vp8_copy_mem8x8_media=vp8_copy_mem8x8_v6;
144$vp8_copy_mem8x8_dspr2=vp8_copy_mem8x8_dspr2;
145
146add_proto qw/void vp8_copy_mem8x4/, "unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch";
147specialize qw/vp8_copy_mem8x4 mmx media neon dspr2/;
148$vp8_copy_mem8x4_media=vp8_copy_mem8x4_v6;
149$vp8_copy_mem8x4_dspr2=vp8_copy_mem8x4_dspr2;
150
151add_proto qw/void vp8_build_intra_predictors_mby_s/, "struct macroblockd *x, unsigned char * yabove_row, unsigned char * yleft, int left_stride, unsigned char * ypred_ptr, int y_stride";
152specialize qw/vp8_build_intra_predictors_mby_s sse2 ssse3 neon/;
153
154add_proto qw/void vp8_build_intra_predictors_mbuv_s/, "struct macroblockd *x, unsigned char * uabove_row, unsigned char * vabove_row,  unsigned char *uleft, unsigned char *vleft, int left_stride, unsigned char * upred_ptr, unsigned char * vpred_ptr, int pred_stride";
155specialize qw/vp8_build_intra_predictors_mbuv_s sse2 ssse3 neon/;
156
157add_proto qw/void vp8_intra4x4_predict/, "unsigned char *Above, unsigned char *yleft, int left_stride, int b_mode, unsigned char *dst, int dst_stride, unsigned char top_left";
158specialize qw/vp8_intra4x4_predict media/;
159$vp8_intra4x4_predict_media=vp8_intra4x4_predict_armv6;
160
161#
162# Postproc
163#
164if (vpx_config("CONFIG_POSTPROC") eq "yes") {
165    add_proto qw/void vp8_mbpost_proc_down/, "unsigned char *dst, int pitch, int rows, int cols,int flimit";
166    specialize qw/vp8_mbpost_proc_down mmx sse2/;
167    $vp8_mbpost_proc_down_sse2=vp8_mbpost_proc_down_xmm;
168
169    add_proto qw/void vp8_mbpost_proc_across_ip/, "unsigned char *dst, int pitch, int rows, int cols,int flimit";
170    specialize qw/vp8_mbpost_proc_across_ip sse2/;
171    $vp8_mbpost_proc_across_ip_sse2=vp8_mbpost_proc_across_ip_xmm;
172
173    add_proto qw/void vp8_post_proc_down_and_across_mb_row/, "unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size";
174    specialize qw/vp8_post_proc_down_and_across_mb_row sse2/;
175
176    add_proto qw/void vp8_plane_add_noise/, "unsigned char *s, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int w, unsigned int h, int pitch";
177    specialize qw/vp8_plane_add_noise mmx sse2/;
178    $vp8_plane_add_noise_sse2=vp8_plane_add_noise_wmt;
179
180    add_proto qw/void vp8_blend_mb_inner/, "unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride";
181    # no asm yet
182
183    add_proto qw/void vp8_blend_mb_outer/, "unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride";
184    # no asm yet
185
186    add_proto qw/void vp8_blend_b/, "unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride";
187    # no asm yet
188
189    add_proto qw/void vp8_filter_by_weight16x16/, "unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight";
190    specialize qw/vp8_filter_by_weight16x16 sse2/;
191
192    add_proto qw/void vp8_filter_by_weight8x8/, "unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight";
193    specialize qw/vp8_filter_by_weight8x8 sse2/;
194
195    add_proto qw/void vp8_filter_by_weight4x4/, "unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight";
196    # no asm yet
197}
198
199#
200# Subpixel
201#
202add_proto qw/void vp8_sixtap_predict16x16/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
203specialize qw/vp8_sixtap_predict16x16 mmx sse2 ssse3 media neon dspr2/;
204$vp8_sixtap_predict16x16_media=vp8_sixtap_predict16x16_armv6;
205$vp8_sixtap_predict16x16_dspr2=vp8_sixtap_predict16x16_dspr2;
206
207add_proto qw/void vp8_sixtap_predict8x8/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
208specialize qw/vp8_sixtap_predict8x8 mmx sse2 ssse3 media neon dspr2/;
209$vp8_sixtap_predict8x8_media=vp8_sixtap_predict8x8_armv6;
210$vp8_sixtap_predict8x8_dspr2=vp8_sixtap_predict8x8_dspr2;
211
212add_proto qw/void vp8_sixtap_predict8x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
213specialize qw/vp8_sixtap_predict8x4 mmx sse2 ssse3 media neon dspr2/;
214$vp8_sixtap_predict8x4_media=vp8_sixtap_predict8x4_armv6;
215$vp8_sixtap_predict8x4_dspr2=vp8_sixtap_predict8x4_dspr2;
216
217add_proto qw/void vp8_sixtap_predict4x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
218# Disable neon while investigating https://code.google.com/p/webm/issues/detail?id=817
219specialize qw/vp8_sixtap_predict4x4 mmx ssse3 media dspr2/;
220$vp8_sixtap_predict4x4_media=vp8_sixtap_predict4x4_armv6;
221$vp8_sixtap_predict4x4_dspr2=vp8_sixtap_predict4x4_dspr2;
222
223add_proto qw/void vp8_bilinear_predict16x16/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
224specialize qw/vp8_bilinear_predict16x16 mmx sse2 ssse3 media neon/;
225$vp8_bilinear_predict16x16_media=vp8_bilinear_predict16x16_armv6;
226
227add_proto qw/void vp8_bilinear_predict8x8/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
228specialize qw/vp8_bilinear_predict8x8 mmx sse2 ssse3 media neon/;
229$vp8_bilinear_predict8x8_media=vp8_bilinear_predict8x8_armv6;
230
231add_proto qw/void vp8_bilinear_predict8x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
232specialize qw/vp8_bilinear_predict8x4 mmx media neon/;
233$vp8_bilinear_predict8x4_media=vp8_bilinear_predict8x4_armv6;
234
235add_proto qw/void vp8_bilinear_predict4x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
236specialize qw/vp8_bilinear_predict4x4 mmx media neon/;
237$vp8_bilinear_predict4x4_media=vp8_bilinear_predict4x4_armv6;
238
239#
240# Whole-pixel Variance
241#
242add_proto qw/unsigned int vp8_variance4x4/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
243specialize qw/vp8_variance4x4 mmx sse2/;
244$vp8_variance4x4_sse2=vp8_variance4x4_wmt;
245
246add_proto qw/unsigned int vp8_variance8x8/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
247specialize qw/vp8_variance8x8 mmx sse2 media neon/;
248$vp8_variance8x8_sse2=vp8_variance8x8_wmt;
249$vp8_variance8x8_media=vp8_variance8x8_armv6;
250
251add_proto qw/unsigned int vp8_variance8x16/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
252specialize qw/vp8_variance8x16 mmx sse2 neon/;
253$vp8_variance8x16_sse2=vp8_variance8x16_wmt;
254
255add_proto qw/unsigned int vp8_variance16x8/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
256specialize qw/vp8_variance16x8 mmx sse2 neon/;
257$vp8_variance16x8_sse2=vp8_variance16x8_wmt;
258
259add_proto qw/unsigned int vp8_variance16x16/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
260specialize qw/vp8_variance16x16 mmx sse2 media neon/;
261$vp8_variance16x16_sse2=vp8_variance16x16_wmt;
262$vp8_variance16x16_media=vp8_variance16x16_armv6;
263
264#
265# Sub-pixel Variance
266#
267add_proto qw/unsigned int vp8_sub_pixel_variance4x4/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
268specialize qw/vp8_sub_pixel_variance4x4 mmx sse2/;
269$vp8_sub_pixel_variance4x4_sse2=vp8_sub_pixel_variance4x4_wmt;
270
271add_proto qw/unsigned int vp8_sub_pixel_variance8x8/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
272specialize qw/vp8_sub_pixel_variance8x8 mmx sse2 media neon_asm/;
273$vp8_sub_pixel_variance8x8_sse2=vp8_sub_pixel_variance8x8_wmt;
274$vp8_sub_pixel_variance8x8_media=vp8_sub_pixel_variance8x8_armv6;
275$vp8_sub_pixel_variance8x8_neon_asm=vp8_sub_pixel_variance8x8_neon;
276
277add_proto qw/unsigned int vp8_sub_pixel_variance8x16/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
278specialize qw/vp8_sub_pixel_variance8x16 mmx sse2/;
279$vp8_sub_pixel_variance8x16_sse2=vp8_sub_pixel_variance8x16_wmt;
280
281add_proto qw/unsigned int vp8_sub_pixel_variance16x8/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
282specialize qw/vp8_sub_pixel_variance16x8 mmx sse2 ssse3/;
283$vp8_sub_pixel_variance16x8_sse2=vp8_sub_pixel_variance16x8_wmt;
284
285add_proto qw/unsigned int vp8_sub_pixel_variance16x16/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
286specialize qw/vp8_sub_pixel_variance16x16 mmx sse2 ssse3 media neon_asm/;
287$vp8_sub_pixel_variance16x16_sse2=vp8_sub_pixel_variance16x16_wmt;
288$vp8_sub_pixel_variance16x16_media=vp8_sub_pixel_variance16x16_armv6;
289$vp8_sub_pixel_variance16x16_neon_asm=vp8_sub_pixel_variance16x16_neon;
290
291add_proto qw/unsigned int vp8_variance_halfpixvar16x16_h/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
292specialize qw/vp8_variance_halfpixvar16x16_h mmx sse2 media neon/;
293$vp8_variance_halfpixvar16x16_h_sse2=vp8_variance_halfpixvar16x16_h_wmt;
294$vp8_variance_halfpixvar16x16_h_media=vp8_variance_halfpixvar16x16_h_armv6;
295
296add_proto qw/unsigned int vp8_variance_halfpixvar16x16_v/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
297specialize qw/vp8_variance_halfpixvar16x16_v mmx sse2 media neon/;
298$vp8_variance_halfpixvar16x16_v_sse2=vp8_variance_halfpixvar16x16_v_wmt;
299$vp8_variance_halfpixvar16x16_v_media=vp8_variance_halfpixvar16x16_v_armv6;
300
301add_proto qw/unsigned int vp8_variance_halfpixvar16x16_hv/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
302specialize qw/vp8_variance_halfpixvar16x16_hv mmx sse2 media neon/;
303$vp8_variance_halfpixvar16x16_hv_sse2=vp8_variance_halfpixvar16x16_hv_wmt;
304$vp8_variance_halfpixvar16x16_hv_media=vp8_variance_halfpixvar16x16_hv_armv6;
305
306#
307# Single block SAD
308#
309add_proto qw/unsigned int vp8_sad4x4/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad";
310specialize qw/vp8_sad4x4 mmx sse2 neon/;
311$vp8_sad4x4_sse2=vp8_sad4x4_wmt;
312
313add_proto qw/unsigned int vp8_sad8x8/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad";
314specialize qw/vp8_sad8x8 mmx sse2 neon/;
315$vp8_sad8x8_sse2=vp8_sad8x8_wmt;
316
317add_proto qw/unsigned int vp8_sad8x16/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad";
318specialize qw/vp8_sad8x16 mmx sse2 neon/;
319$vp8_sad8x16_sse2=vp8_sad8x16_wmt;
320
321add_proto qw/unsigned int vp8_sad16x8/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad";
322specialize qw/vp8_sad16x8 mmx sse2 neon/;
323$vp8_sad16x8_sse2=vp8_sad16x8_wmt;
324
325add_proto qw/unsigned int vp8_sad16x16/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad";
326specialize qw/vp8_sad16x16 mmx sse2 sse3 media neon/;
327$vp8_sad16x16_sse2=vp8_sad16x16_wmt;
328$vp8_sad16x16_media=vp8_sad16x16_armv6;
329
330#
331# Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
332#
333add_proto qw/void vp8_sad4x4x3/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array";
334specialize qw/vp8_sad4x4x3 sse3/;
335
336add_proto qw/void vp8_sad8x8x3/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array";
337specialize qw/vp8_sad8x8x3 sse3/;
338
339add_proto qw/void vp8_sad8x16x3/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array";
340specialize qw/vp8_sad8x16x3 sse3/;
341
342add_proto qw/void vp8_sad16x8x3/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array";
343specialize qw/vp8_sad16x8x3 sse3 ssse3/;
344
345add_proto qw/void vp8_sad16x16x3/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array";
346specialize qw/vp8_sad16x16x3 sse3 ssse3/;
347
348# Note the only difference in the following prototypes is that they return into
349# an array of short
350add_proto qw/void vp8_sad4x4x8/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array";
351specialize qw/vp8_sad4x4x8 sse4_1/;
352$vp8_sad4x4x8_sse4_1=vp8_sad4x4x8_sse4;
353
354add_proto qw/void vp8_sad8x8x8/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array";
355specialize qw/vp8_sad8x8x8 sse4_1/;
356$vp8_sad8x8x8_sse4_1=vp8_sad8x8x8_sse4;
357
358add_proto qw/void vp8_sad8x16x8/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array";
359specialize qw/vp8_sad8x16x8 sse4_1/;
360$vp8_sad8x16x8_sse4_1=vp8_sad8x16x8_sse4;
361
362add_proto qw/void vp8_sad16x8x8/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array";
363specialize qw/vp8_sad16x8x8 sse4_1/;
364$vp8_sad16x8x8_sse4_1=vp8_sad16x8x8_sse4;
365
366add_proto qw/void vp8_sad16x16x8/, "const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array";
367specialize qw/vp8_sad16x16x8 sse4_1/;
368$vp8_sad16x16x8_sse4_1=vp8_sad16x16x8_sse4;
369
370#
371# Multi-block SAD, comparing a reference to N independent blocks
372#
373add_proto qw/void vp8_sad4x4x4d/, "const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array";
374specialize qw/vp8_sad4x4x4d sse3/;
375
376add_proto qw/void vp8_sad8x8x4d/, "const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array";
377specialize qw/vp8_sad8x8x4d sse3/;
378
379add_proto qw/void vp8_sad8x16x4d/, "const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array";
380specialize qw/vp8_sad8x16x4d sse3/;
381
382add_proto qw/void vp8_sad16x8x4d/, "const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array";
383specialize qw/vp8_sad16x8x4d sse3/;
384
385add_proto qw/void vp8_sad16x16x4d/, "const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array";
386specialize qw/vp8_sad16x16x4d sse3/;
387
388#
389# Encoder functions below this point.
390#
391if (vpx_config("CONFIG_VP8_ENCODER") eq "yes") {
392
393#
394# Sum of squares (vector)
395#
396add_proto qw/unsigned int vp8_get_mb_ss/, "const short *";
397specialize qw/vp8_get_mb_ss mmx sse2/;
398
399#
400# SSE (Sum Squared Error)
401#
402add_proto qw/unsigned int vp8_sub_pixel_mse16x16/, "const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse";
403specialize qw/vp8_sub_pixel_mse16x16 mmx sse2/;
404$vp8_sub_pixel_mse16x16_sse2=vp8_sub_pixel_mse16x16_wmt;
405
406add_proto qw/unsigned int vp8_mse16x16/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse";
407specialize qw/vp8_mse16x16 mmx sse2 media neon/;
408$vp8_mse16x16_sse2=vp8_mse16x16_wmt;
409$vp8_mse16x16_media=vp8_mse16x16_armv6;
410
411add_proto qw/unsigned int vp8_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride";
412specialize qw/vp8_get4x4sse_cs mmx neon/;
413
414#
415# Block copy
416#
417if ($opts{arch} =~ /x86/) {
418    add_proto qw/void vp8_copy32xn/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n";
419    specialize qw/vp8_copy32xn sse2 sse3/;
420}
421
422#
423# Structured Similarity (SSIM)
424#
425if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
426    $opts{arch} eq "x86_64" and $sse2_on_x86_64 = "sse2";
427
428    add_proto qw/void vp8_ssim_parms_8x8/, "unsigned char *s, int sp, unsigned char *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr";
429    specialize qw/vp8_ssim_parms_8x8/, "$sse2_on_x86_64";
430
431    add_proto qw/void vp8_ssim_parms_16x16/, "unsigned char *s, int sp, unsigned char *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr";
432    specialize qw/vp8_ssim_parms_16x16/, "$sse2_on_x86_64";
433}
434
435#
436# Forward DCT
437#
438add_proto qw/void vp8_short_fdct4x4/, "short *input, short *output, int pitch";
439specialize qw/vp8_short_fdct4x4 mmx sse2 media neon/;
440$vp8_short_fdct4x4_media=vp8_short_fdct4x4_armv6;
441
442add_proto qw/void vp8_short_fdct8x4/, "short *input, short *output, int pitch";
443specialize qw/vp8_short_fdct8x4 mmx sse2 media neon/;
444$vp8_short_fdct8x4_media=vp8_short_fdct8x4_armv6;
445
446add_proto qw/void vp8_short_walsh4x4/, "short *input, short *output, int pitch";
447specialize qw/vp8_short_walsh4x4 sse2 media neon/;
448$vp8_short_walsh4x4_media=vp8_short_walsh4x4_armv6;
449
450#
451# Quantizer
452#
453add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *";
454specialize qw/vp8_regular_quantize_b sse2 sse4_1/;
455
456add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *";
457specialize qw/vp8_fast_quantize_b sse2 ssse3 media neon_asm/;
458$vp8_fast_quantize_b_media=vp8_fast_quantize_b_armv6;
459$vp8_fast_quantize_b_neon_asm=vp8_fast_quantize_b_neon;
460
461add_proto qw/void vp8_regular_quantize_b_pair/, "struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2";
462# no asm yet
463
464add_proto qw/void vp8_fast_quantize_b_pair/, "struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2";
465specialize qw/vp8_fast_quantize_b_pair neon_asm/;
466$vp8_fast_quantize_b_pair_neon_asm=vp8_fast_quantize_b_pair_neon;
467
468add_proto qw/void vp8_quantize_mb/, "struct macroblock *";
469specialize qw/vp8_quantize_mb neon/;
470
471add_proto qw/void vp8_quantize_mby/, "struct macroblock *";
472specialize qw/vp8_quantize_mby neon/;
473
474add_proto qw/void vp8_quantize_mbuv/, "struct macroblock *";
475specialize qw/vp8_quantize_mbuv neon/;
476
477#
478# Block subtraction
479#
480add_proto qw/int vp8_block_error/, "short *coeff, short *dqcoeff";
481specialize qw/vp8_block_error mmx sse2/;
482$vp8_block_error_sse2=vp8_block_error_xmm;
483
484add_proto qw/int vp8_mbblock_error/, "struct macroblock *mb, int dc";
485specialize qw/vp8_mbblock_error mmx sse2/;
486$vp8_mbblock_error_sse2=vp8_mbblock_error_xmm;
487
488add_proto qw/int vp8_mbuverror/, "struct macroblock *mb";
489specialize qw/vp8_mbuverror mmx sse2/;
490$vp8_mbuverror_sse2=vp8_mbuverror_xmm;
491
492add_proto qw/void vp8_subtract_b/, "struct block *be, struct blockd *bd, int pitch";
493specialize qw/vp8_subtract_b mmx sse2 media neon/;
494$vp8_subtract_b_media=vp8_subtract_b_armv6;
495
496add_proto qw/void vp8_subtract_mby/, "short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride";
497specialize qw/vp8_subtract_mby mmx sse2 media neon/;
498$vp8_subtract_mby_media=vp8_subtract_mby_armv6;
499
500add_proto qw/void vp8_subtract_mbuv/, "short *diff, unsigned char *usrc, unsigned char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pred_stride";
501specialize qw/vp8_subtract_mbuv mmx sse2 media neon/;
502$vp8_subtract_mbuv_media=vp8_subtract_mbuv_armv6;
503
504#
505# Motion search
506#
507add_proto qw/int vp8_full_search_sad/, "struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv";
508specialize qw/vp8_full_search_sad sse3 sse4_1/;
509$vp8_full_search_sad_sse3=vp8_full_search_sadx3;
510$vp8_full_search_sad_sse4_1=vp8_full_search_sadx8;
511
512add_proto qw/int vp8_refining_search_sad/, "struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv";
513specialize qw/vp8_refining_search_sad sse3/;
514$vp8_refining_search_sad_sse3=vp8_refining_search_sadx4;
515
516add_proto qw/int vp8_diamond_search_sad/, "struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv";
517$vp8_diamond_search_sad_sse3=vp8_diamond_search_sadx4;
518
519#
520# Alt-ref Noise Reduction (ARNR)
521#
522if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") {
523    add_proto qw/void vp8_temporal_filter_apply/, "unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count";
524    specialize qw/vp8_temporal_filter_apply sse2/;
525}
526
527#
528# Denoiser filter
529#
530if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") {
531    add_proto qw/int vp8_denoiser_filter/, "unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
532    specialize qw/vp8_denoiser_filter sse2 neon/;
533    add_proto qw/int vp8_denoiser_filter_uv/, "unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
534    specialize qw/vp8_denoiser_filter_uv sse2 neon/;
535}
536
537# End of encoder only functions
538}
5391;
540