OLD | NEW |
1 sub vp8_common_forward_decls() { | 1 sub vp8_common_forward_decls() { |
2 print <<EOF | 2 print <<EOF |
3 /* | 3 /* |
4 * VP8 | 4 * VP8 |
5 */ | 5 */ |
6 | 6 |
7 struct blockd; | 7 struct blockd; |
8 struct macroblockd; | 8 struct macroblockd; |
9 struct loop_filter_info; | 9 struct loop_filter_info; |
10 | 10 |
(...skipping 20 matching lines...) Expand all Loading... |
31 add_proto qw/void vp8_dequantize_b/, "struct blockd*, short *dqc"; | 31 add_proto qw/void vp8_dequantize_b/, "struct blockd*, short *dqc"; |
32 specialize qw/vp8_dequantize_b mmx media neon/; | 32 specialize qw/vp8_dequantize_b mmx media neon/; |
33 $vp8_dequantize_b_media=vp8_dequantize_b_v6; | 33 $vp8_dequantize_b_media=vp8_dequantize_b_v6; |
34 | 34 |
35 add_proto qw/void vp8_dequant_idct_add/, "short *input, short *dq, unsigned char
*output, int stride"; | 35 add_proto qw/void vp8_dequant_idct_add/, "short *input, short *dq, unsigned char
*output, int stride"; |
36 specialize qw/vp8_dequant_idct_add mmx media neon dspr2/; | 36 specialize qw/vp8_dequant_idct_add mmx media neon dspr2/; |
37 $vp8_dequant_idct_add_media=vp8_dequant_idct_add_v6; | 37 $vp8_dequant_idct_add_media=vp8_dequant_idct_add_v6; |
38 $vp8_dequant_idct_add_dspr2=vp8_dequant_idct_add_dspr2; | 38 $vp8_dequant_idct_add_dspr2=vp8_dequant_idct_add_dspr2; |
39 | 39 |
40 add_proto qw/void vp8_dequant_idct_add_y_block/, "short *q, short *dq, unsigned
char *dst, int stride, char *eobs"; | 40 add_proto qw/void vp8_dequant_idct_add_y_block/, "short *q, short *dq, unsigned
char *dst, int stride, char *eobs"; |
41 specialize qw/vp8_dequant_idct_add_y_block mmx sse2 media neon dspr2/; | 41 specialize qw/vp8_dequant_idct_add_y_block mmx sse2 media neon_asm dspr2/; |
42 $vp8_dequant_idct_add_y_block_media=vp8_dequant_idct_add_y_block_v6; | 42 $vp8_dequant_idct_add_y_block_media=vp8_dequant_idct_add_y_block_v6; |
| 43 $vp8_dequant_idct_add_y_block_neon_asm=vp8_dequant_idct_add_y_block_neon; |
43 $vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2; | 44 $vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2; |
44 | 45 |
45 add_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"; | 46 add_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"; |
46 specialize qw/vp8_dequant_idct_add_uv_block mmx sse2 media neon dspr2/; | 47 specialize qw/vp8_dequant_idct_add_uv_block mmx sse2 media neon_asm dspr2/; |
47 $vp8_dequant_idct_add_uv_block_media=vp8_dequant_idct_add_uv_block_v6; | 48 $vp8_dequant_idct_add_uv_block_media=vp8_dequant_idct_add_uv_block_v6; |
| 49 $vp8_dequant_idct_add_uv_block_neon_asm=vp8_dequant_idct_add_uv_block_neon; |
48 $vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2; | 50 $vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2; |
49 | 51 |
50 # | 52 # |
51 # Loopfilter | 53 # Loopfilter |
52 # | 54 # |
53 add_proto qw/void vp8_loop_filter_mbv/, "unsigned char *y, unsigned char *u, uns
igned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; | 55 add_proto qw/void vp8_loop_filter_mbv/, "unsigned char *y, unsigned char *u, uns
igned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; |
54 specialize qw/vp8_loop_filter_mbv mmx sse2 media neon dspr2/; | 56 specialize qw/vp8_loop_filter_mbv mmx sse2 media neon dspr2/; |
55 $vp8_loop_filter_mbv_media=vp8_loop_filter_mbv_armv6; | 57 $vp8_loop_filter_mbv_media=vp8_loop_filter_mbv_armv6; |
56 $vp8_loop_filter_mbv_dspr2=vp8_loop_filter_mbv_dspr2; | 58 $vp8_loop_filter_mbv_dspr2=vp8_loop_filter_mbv_dspr2; |
57 | 59 |
58 add_proto qw/void vp8_loop_filter_bv/, "unsigned char *y, unsigned char *u, unsi
gned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; | 60 add_proto qw/void vp8_loop_filter_bv/, "unsigned char *y, unsigned char *u, unsi
gned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; |
59 specialize qw/vp8_loop_filter_bv mmx sse2 media neon dspr2/; | 61 specialize qw/vp8_loop_filter_bv mmx sse2 media neon_asm dspr2/; |
60 $vp8_loop_filter_bv_media=vp8_loop_filter_bv_armv6; | 62 $vp8_loop_filter_bv_media=vp8_loop_filter_bv_armv6; |
| 63 $vp8_loop_filter_bv_neon_asm=vp8_loop_filter_bv_neon; |
61 $vp8_loop_filter_bv_dspr2=vp8_loop_filter_bv_dspr2; | 64 $vp8_loop_filter_bv_dspr2=vp8_loop_filter_bv_dspr2; |
62 | 65 |
63 add_proto qw/void vp8_loop_filter_mbh/, "unsigned char *y, unsigned char *u, uns
igned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; | 66 add_proto qw/void vp8_loop_filter_mbh/, "unsigned char *y, unsigned char *u, uns
igned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; |
64 specialize qw/vp8_loop_filter_mbh mmx sse2 media neon dspr2/; | 67 specialize qw/vp8_loop_filter_mbh mmx sse2 media neon dspr2/; |
65 $vp8_loop_filter_mbh_media=vp8_loop_filter_mbh_armv6; | 68 $vp8_loop_filter_mbh_media=vp8_loop_filter_mbh_armv6; |
66 $vp8_loop_filter_mbh_dspr2=vp8_loop_filter_mbh_dspr2; | 69 $vp8_loop_filter_mbh_dspr2=vp8_loop_filter_mbh_dspr2; |
67 | 70 |
68 add_proto qw/void vp8_loop_filter_bh/, "unsigned char *y, unsigned char *u, unsi
gned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; | 71 add_proto qw/void vp8_loop_filter_bh/, "unsigned char *y, unsigned char *u, unsi
gned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi"; |
69 specialize qw/vp8_loop_filter_bh mmx sse2 media neon dspr2/; | 72 specialize qw/vp8_loop_filter_bh mmx sse2 media neon_asm dspr2/; |
70 $vp8_loop_filter_bh_media=vp8_loop_filter_bh_armv6; | 73 $vp8_loop_filter_bh_media=vp8_loop_filter_bh_armv6; |
| 74 $vp8_loop_filter_bh_neon_asm=vp8_loop_filter_bh_neon; |
71 $vp8_loop_filter_bh_dspr2=vp8_loop_filter_bh_dspr2; | 75 $vp8_loop_filter_bh_dspr2=vp8_loop_filter_bh_dspr2; |
72 | 76 |
73 | 77 |
74 add_proto qw/void vp8_loop_filter_simple_mbv/, "unsigned char *y, int ystride, c
onst unsigned char *blimit"; | 78 add_proto qw/void vp8_loop_filter_simple_mbv/, "unsigned char *y, int ystride, c
onst unsigned char *blimit"; |
75 specialize qw/vp8_loop_filter_simple_mbv mmx sse2 media neon/; | 79 specialize qw/vp8_loop_filter_simple_mbv mmx sse2 media neon_asm/; |
76 $vp8_loop_filter_simple_mbv_c=vp8_loop_filter_simple_vertical_edge_c; | 80 $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; | 81 $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; | 82 $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; | 83 $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; | 84 $vp8_loop_filter_simple_mbv_neon_asm=vp8_loop_filter_mbvs_neon; |
81 | 85 |
82 add_proto qw/void vp8_loop_filter_simple_mbh/, "unsigned char *y, int ystride, c
onst unsigned char *blimit"; | 86 add_proto qw/void vp8_loop_filter_simple_mbh/, "unsigned char *y, int ystride, c
onst unsigned char *blimit"; |
83 specialize qw/vp8_loop_filter_simple_mbh mmx sse2 media neon/; | 87 specialize qw/vp8_loop_filter_simple_mbh mmx sse2 media neon/; |
84 $vp8_loop_filter_simple_mbh_c=vp8_loop_filter_simple_horizontal_edge_c; | 88 $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; | 89 $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; | 90 $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; | 91 $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; | 92 $vp8_loop_filter_simple_mbh_neon=vp8_loop_filter_mbhs_neon; |
89 | 93 |
90 add_proto qw/void vp8_loop_filter_simple_bv/, "unsigned char *y, int ystride, co
nst unsigned char *blimit"; | 94 add_proto qw/void vp8_loop_filter_simple_bv/, "unsigned char *y, int ystride, co
nst unsigned char *blimit"; |
91 specialize qw/vp8_loop_filter_simple_bv mmx sse2 media neon/; | 95 specialize qw/vp8_loop_filter_simple_bv mmx sse2 media neon_asm/; |
92 $vp8_loop_filter_simple_bv_c=vp8_loop_filter_bvs_c; | 96 $vp8_loop_filter_simple_bv_c=vp8_loop_filter_bvs_c; |
93 $vp8_loop_filter_simple_bv_mmx=vp8_loop_filter_bvs_mmx; | 97 $vp8_loop_filter_simple_bv_mmx=vp8_loop_filter_bvs_mmx; |
94 $vp8_loop_filter_simple_bv_sse2=vp8_loop_filter_bvs_sse2; | 98 $vp8_loop_filter_simple_bv_sse2=vp8_loop_filter_bvs_sse2; |
95 $vp8_loop_filter_simple_bv_media=vp8_loop_filter_bvs_armv6; | 99 $vp8_loop_filter_simple_bv_media=vp8_loop_filter_bvs_armv6; |
96 $vp8_loop_filter_simple_bv_neon=vp8_loop_filter_bvs_neon; | 100 $vp8_loop_filter_simple_bv_neon_asm=vp8_loop_filter_bvs_neon; |
97 | 101 |
98 add_proto qw/void vp8_loop_filter_simple_bh/, "unsigned char *y, int ystride, co
nst unsigned char *blimit"; | 102 add_proto qw/void vp8_loop_filter_simple_bh/, "unsigned char *y, int ystride, co
nst unsigned char *blimit"; |
99 specialize qw/vp8_loop_filter_simple_bh mmx sse2 media neon/; | 103 specialize qw/vp8_loop_filter_simple_bh mmx sse2 media neon/; |
100 $vp8_loop_filter_simple_bh_c=vp8_loop_filter_bhs_c; | 104 $vp8_loop_filter_simple_bh_c=vp8_loop_filter_bhs_c; |
101 $vp8_loop_filter_simple_bh_mmx=vp8_loop_filter_bhs_mmx; | 105 $vp8_loop_filter_simple_bh_mmx=vp8_loop_filter_bhs_mmx; |
102 $vp8_loop_filter_simple_bh_sse2=vp8_loop_filter_bhs_sse2; | 106 $vp8_loop_filter_simple_bh_sse2=vp8_loop_filter_bhs_sse2; |
103 $vp8_loop_filter_simple_bh_media=vp8_loop_filter_bhs_armv6; | 107 $vp8_loop_filter_simple_bh_media=vp8_loop_filter_bhs_armv6; |
104 $vp8_loop_filter_simple_bh_neon=vp8_loop_filter_bhs_neon; | 108 $vp8_loop_filter_simple_bh_neon=vp8_loop_filter_bhs_neon; |
105 | 109 |
106 # | 110 # |
(...skipping 155 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
262 $vp8_variance16x16_media=vp8_variance16x16_armv6; | 266 $vp8_variance16x16_media=vp8_variance16x16_armv6; |
263 | 267 |
264 # | 268 # |
265 # Sub-pixel Variance | 269 # Sub-pixel Variance |
266 # | 270 # |
267 add_proto qw/unsigned int vp8_sub_pixel_variance4x4/, "const unsigned char *src
_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_p
tr, int Refstride, unsigned int *sse"; | 271 add_proto qw/unsigned int vp8_sub_pixel_variance4x4/, "const unsigned char *src
_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_p
tr, int Refstride, unsigned int *sse"; |
268 specialize qw/vp8_sub_pixel_variance4x4 mmx sse2/; | 272 specialize qw/vp8_sub_pixel_variance4x4 mmx sse2/; |
269 $vp8_sub_pixel_variance4x4_sse2=vp8_sub_pixel_variance4x4_wmt; | 273 $vp8_sub_pixel_variance4x4_sse2=vp8_sub_pixel_variance4x4_wmt; |
270 | 274 |
271 add_proto qw/unsigned int vp8_sub_pixel_variance8x8/, "const unsigned char *src
_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_p
tr, int Refstride, unsigned int *sse"; | 275 add_proto qw/unsigned int vp8_sub_pixel_variance8x8/, "const unsigned char *src
_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_p
tr, int Refstride, unsigned int *sse"; |
272 specialize qw/vp8_sub_pixel_variance8x8 mmx sse2 media neon/; | 276 specialize qw/vp8_sub_pixel_variance8x8 mmx sse2 media neon_asm/; |
273 $vp8_sub_pixel_variance8x8_sse2=vp8_sub_pixel_variance8x8_wmt; | 277 $vp8_sub_pixel_variance8x8_sse2=vp8_sub_pixel_variance8x8_wmt; |
274 $vp8_sub_pixel_variance8x8_media=vp8_sub_pixel_variance8x8_armv6; | 278 $vp8_sub_pixel_variance8x8_media=vp8_sub_pixel_variance8x8_armv6; |
| 279 $vp8_sub_pixel_variance8x8_neon_asm=vp8_sub_pixel_variance8x8_neon; |
275 | 280 |
276 add_proto qw/unsigned int vp8_sub_pixel_variance8x16/, "const unsigned char *sr
c_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_
ptr, int Refstride, unsigned int *sse"; | 281 add_proto qw/unsigned int vp8_sub_pixel_variance8x16/, "const unsigned char *sr
c_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_
ptr, int Refstride, unsigned int *sse"; |
277 specialize qw/vp8_sub_pixel_variance8x16 mmx sse2/; | 282 specialize qw/vp8_sub_pixel_variance8x16 mmx sse2/; |
278 $vp8_sub_pixel_variance8x16_sse2=vp8_sub_pixel_variance8x16_wmt; | 283 $vp8_sub_pixel_variance8x16_sse2=vp8_sub_pixel_variance8x16_wmt; |
279 | 284 |
280 add_proto qw/unsigned int vp8_sub_pixel_variance16x8/, "const unsigned char *sr
c_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_
ptr, int Refstride, unsigned int *sse"; | 285 add_proto qw/unsigned int vp8_sub_pixel_variance16x8/, "const unsigned char *sr
c_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref_
ptr, int Refstride, unsigned int *sse"; |
281 specialize qw/vp8_sub_pixel_variance16x8 mmx sse2 ssse3/; | 286 specialize qw/vp8_sub_pixel_variance16x8 mmx sse2 ssse3/; |
282 $vp8_sub_pixel_variance16x8_sse2=vp8_sub_pixel_variance16x8_wmt; | 287 $vp8_sub_pixel_variance16x8_sse2=vp8_sub_pixel_variance16x8_wmt; |
283 | 288 |
284 add_proto qw/unsigned int vp8_sub_pixel_variance16x16/, "const unsigned char *s
rc_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref
_ptr, int Refstride, unsigned int *sse"; | 289 add_proto qw/unsigned int vp8_sub_pixel_variance16x16/, "const unsigned char *s
rc_ptr, int source_stride, int xoffset, int yoffset, const unsigned char *ref
_ptr, int Refstride, unsigned int *sse"; |
285 specialize qw/vp8_sub_pixel_variance16x16 mmx sse2 ssse3 media neon/; | 290 specialize qw/vp8_sub_pixel_variance16x16 mmx sse2 ssse3 media neon_asm/; |
286 $vp8_sub_pixel_variance16x16_sse2=vp8_sub_pixel_variance16x16_wmt; | 291 $vp8_sub_pixel_variance16x16_sse2=vp8_sub_pixel_variance16x16_wmt; |
287 $vp8_sub_pixel_variance16x16_media=vp8_sub_pixel_variance16x16_armv6; | 292 $vp8_sub_pixel_variance16x16_media=vp8_sub_pixel_variance16x16_armv6; |
| 293 $vp8_sub_pixel_variance16x16_neon_asm=vp8_sub_pixel_variance16x16_neon; |
288 | 294 |
289 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_h/, "const unsigned char
*src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, unsi
gned int *sse"; | 295 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_h/, "const unsigned char
*src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, unsi
gned int *sse"; |
290 specialize qw/vp8_variance_halfpixvar16x16_h mmx sse2 media neon/; | 296 specialize qw/vp8_variance_halfpixvar16x16_h mmx sse2 media neon_asm/; |
291 $vp8_variance_halfpixvar16x16_h_sse2=vp8_variance_halfpixvar16x16_h_wmt; | 297 $vp8_variance_halfpixvar16x16_h_sse2=vp8_variance_halfpixvar16x16_h_wmt; |
292 $vp8_variance_halfpixvar16x16_h_media=vp8_variance_halfpixvar16x16_h_armv6; | 298 $vp8_variance_halfpixvar16x16_h_media=vp8_variance_halfpixvar16x16_h_armv6; |
| 299 $vp8_variance_halfpixvar16x16_h_neon_asm=vp8_variance_halfpixvar16x16_h_neon; |
293 | 300 |
294 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_v/, "const unsigned char
*src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, unsi
gned int *sse"; | 301 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_v/, "const unsigned char
*src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, unsi
gned int *sse"; |
295 specialize qw/vp8_variance_halfpixvar16x16_v mmx sse2 media neon/; | 302 specialize qw/vp8_variance_halfpixvar16x16_v mmx sse2 media neon_asm/; |
296 $vp8_variance_halfpixvar16x16_v_sse2=vp8_variance_halfpixvar16x16_v_wmt; | 303 $vp8_variance_halfpixvar16x16_v_sse2=vp8_variance_halfpixvar16x16_v_wmt; |
297 $vp8_variance_halfpixvar16x16_v_media=vp8_variance_halfpixvar16x16_v_armv6; | 304 $vp8_variance_halfpixvar16x16_v_media=vp8_variance_halfpixvar16x16_v_armv6; |
| 305 $vp8_variance_halfpixvar16x16_v_neon_asm=vp8_variance_halfpixvar16x16_v_neon; |
298 | 306 |
299 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_hv/, "const unsigned char
*src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, uns
igned int *sse"; | 307 add_proto qw/unsigned int vp8_variance_halfpixvar16x16_hv/, "const unsigned char
*src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, uns
igned int *sse"; |
300 specialize qw/vp8_variance_halfpixvar16x16_hv mmx sse2 media neon/; | 308 specialize qw/vp8_variance_halfpixvar16x16_hv mmx sse2 media neon_asm/; |
301 $vp8_variance_halfpixvar16x16_hv_sse2=vp8_variance_halfpixvar16x16_hv_wmt; | 309 $vp8_variance_halfpixvar16x16_hv_sse2=vp8_variance_halfpixvar16x16_hv_wmt; |
302 $vp8_variance_halfpixvar16x16_hv_media=vp8_variance_halfpixvar16x16_hv_armv6; | 310 $vp8_variance_halfpixvar16x16_hv_media=vp8_variance_halfpixvar16x16_hv_armv6; |
| 311 $vp8_variance_halfpixvar16x16_hv_neon_asm=vp8_variance_halfpixvar16x16_hv_neon; |
303 | 312 |
304 # | 313 # |
305 # Single block SAD | 314 # Single block SAD |
306 # | 315 # |
307 add_proto qw/unsigned int vp8_sad4x4/, "const unsigned char *src_ptr, int src_st
ride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad"; | 316 add_proto qw/unsigned int vp8_sad4x4/, "const unsigned char *src_ptr, int src_st
ride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad"; |
308 specialize qw/vp8_sad4x4 mmx sse2 neon/; | 317 specialize qw/vp8_sad4x4 mmx sse2 neon/; |
309 $vp8_sad4x4_sse2=vp8_sad4x4_wmt; | 318 $vp8_sad4x4_sse2=vp8_sad4x4_wmt; |
310 | 319 |
311 add_proto qw/unsigned int vp8_sad8x8/, "const unsigned char *src_ptr, int src_st
ride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad"; | 320 add_proto qw/unsigned int vp8_sad8x8/, "const unsigned char *src_ptr, int src_st
ride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad"; |
312 specialize qw/vp8_sad8x8 mmx sse2 neon/; | 321 specialize qw/vp8_sad8x8 mmx sse2 neon/; |
(...skipping 82 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
395 specialize qw/vp8_get_mb_ss mmx sse2/; | 404 specialize qw/vp8_get_mb_ss mmx sse2/; |
396 | 405 |
397 # | 406 # |
398 # SSE (Sum Squared Error) | 407 # SSE (Sum Squared Error) |
399 # | 408 # |
400 add_proto qw/unsigned int vp8_sub_pixel_mse16x16/, "const unsigned char *src_pt
r, int source_stride, int xoffset, int yoffset, const unsigned char *ref_ptr,
int Refstride, unsigned int *sse"; | 409 add_proto qw/unsigned int vp8_sub_pixel_mse16x16/, "const unsigned char *src_pt
r, int source_stride, int xoffset, int yoffset, const unsigned char *ref_ptr,
int Refstride, unsigned int *sse"; |
401 specialize qw/vp8_sub_pixel_mse16x16 mmx sse2/; | 410 specialize qw/vp8_sub_pixel_mse16x16 mmx sse2/; |
402 $vp8_sub_pixel_mse16x16_sse2=vp8_sub_pixel_mse16x16_wmt; | 411 $vp8_sub_pixel_mse16x16_sse2=vp8_sub_pixel_mse16x16_wmt; |
403 | 412 |
404 add_proto qw/unsigned int vp8_mse16x16/, "const unsigned char *src_ptr, int sour
ce_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int *sse"; | 413 add_proto qw/unsigned int vp8_mse16x16/, "const unsigned char *src_ptr, int sour
ce_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int *sse"; |
405 specialize qw/vp8_mse16x16 mmx sse2 media neon/; | 414 specialize qw/vp8_mse16x16 mmx sse2 media neon_asm/; |
406 $vp8_mse16x16_sse2=vp8_mse16x16_wmt; | 415 $vp8_mse16x16_sse2=vp8_mse16x16_wmt; |
407 $vp8_mse16x16_media=vp8_mse16x16_armv6; | 416 $vp8_mse16x16_media=vp8_mse16x16_armv6; |
| 417 $vp8_mse16x16_neon_asm=vp8_mse16x16_neon; |
408 | 418 |
409 add_proto qw/unsigned int vp8_get4x4sse_cs/, "const unsigned char *src_ptr, int
source_stride, const unsigned char *ref_ptr, int ref_stride"; | 419 add_proto qw/unsigned int vp8_get4x4sse_cs/, "const unsigned char *src_ptr, int
source_stride, const unsigned char *ref_ptr, int ref_stride"; |
410 specialize qw/vp8_get4x4sse_cs mmx neon/; | 420 specialize qw/vp8_get4x4sse_cs mmx neon_asm/; |
| 421 $vp8_get4x4sse_cs_neon_asm=vp8_get4x4sse_cs_neon; |
411 | 422 |
412 # | 423 # |
413 # Block copy | 424 # Block copy |
414 # | 425 # |
415 if ($opts{arch} =~ /x86/) { | 426 if ($opts{arch} =~ /x86/) { |
416 add_proto qw/void vp8_copy32xn/, "const unsigned char *src_ptr, int source_s
tride, const unsigned char *ref_ptr, int ref_stride, int n"; | 427 add_proto qw/void vp8_copy32xn/, "const unsigned char *src_ptr, int source_s
tride, const unsigned char *ref_ptr, int ref_stride, int n"; |
417 specialize qw/vp8_copy32xn sse2 sse3/; | 428 specialize qw/vp8_copy32xn sse2 sse3/; |
418 } | 429 } |
419 | 430 |
420 # | 431 # |
421 # Structured Similarity (SSIM) | 432 # Structured Similarity (SSIM) |
422 # | 433 # |
423 if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") { | 434 if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") { |
424 $opts{arch} eq "x86_64" and $sse2_on_x86_64 = "sse2"; | 435 $opts{arch} eq "x86_64" and $sse2_on_x86_64 = "sse2"; |
425 | 436 |
426 add_proto qw/void vp8_ssim_parms_8x8/, "unsigned char *s, int sp, unsigned c
har *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_s
q_s, unsigned long *sum_sq_r, unsigned long *sum_sxr"; | 437 add_proto qw/void vp8_ssim_parms_8x8/, "unsigned char *s, int sp, unsigned c
har *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_s
q_s, unsigned long *sum_sq_r, unsigned long *sum_sxr"; |
427 specialize qw/vp8_ssim_parms_8x8/, "$sse2_on_x86_64"; | 438 specialize qw/vp8_ssim_parms_8x8/, "$sse2_on_x86_64"; |
428 | 439 |
429 add_proto qw/void vp8_ssim_parms_16x16/, "unsigned char *s, int sp, unsigned
char *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum
_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr"; | 440 add_proto qw/void vp8_ssim_parms_16x16/, "unsigned char *s, int sp, unsigned
char *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum
_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr"; |
430 specialize qw/vp8_ssim_parms_16x16/, "$sse2_on_x86_64"; | 441 specialize qw/vp8_ssim_parms_16x16/, "$sse2_on_x86_64"; |
431 } | 442 } |
432 | 443 |
433 # | 444 # |
434 # Forward DCT | 445 # Forward DCT |
435 # | 446 # |
436 add_proto qw/void vp8_short_fdct4x4/, "short *input, short *output, int pitch"; | 447 add_proto qw/void vp8_short_fdct4x4/, "short *input, short *output, int pitch"; |
437 specialize qw/vp8_short_fdct4x4 mmx sse2 media neon/; | 448 specialize qw/vp8_short_fdct4x4 mmx sse2 media neon_asm/; |
438 $vp8_short_fdct4x4_media=vp8_short_fdct4x4_armv6; | 449 $vp8_short_fdct4x4_media=vp8_short_fdct4x4_armv6; |
| 450 $vp8_short_fdct4x4_neon_asm=vp8_short_fdct4x4_neon; |
439 | 451 |
440 add_proto qw/void vp8_short_fdct8x4/, "short *input, short *output, int pitch"; | 452 add_proto qw/void vp8_short_fdct8x4/, "short *input, short *output, int pitch"; |
441 specialize qw/vp8_short_fdct8x4 mmx sse2 media neon/; | 453 specialize qw/vp8_short_fdct8x4 mmx sse2 media neon_asm/; |
442 $vp8_short_fdct8x4_media=vp8_short_fdct8x4_armv6; | 454 $vp8_short_fdct8x4_media=vp8_short_fdct8x4_armv6; |
| 455 $vp8_short_fdct8x4_neon_asm=vp8_short_fdct8x4_neon; |
443 | 456 |
444 add_proto qw/void vp8_short_walsh4x4/, "short *input, short *output, int pitch"; | 457 add_proto qw/void vp8_short_walsh4x4/, "short *input, short *output, int pitch"; |
445 specialize qw/vp8_short_walsh4x4 sse2 media neon/; | 458 specialize qw/vp8_short_walsh4x4 sse2 media neon_asm/; |
446 $vp8_short_walsh4x4_media=vp8_short_walsh4x4_armv6; | 459 $vp8_short_walsh4x4_media=vp8_short_walsh4x4_armv6; |
| 460 $vp8_short_walsh4x4_neon_asm=vp8_short_walsh4x4_neon; |
447 | 461 |
448 # | 462 # |
449 # Quantizer | 463 # Quantizer |
450 # | 464 # |
451 add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *"; | 465 add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *"; |
452 specialize qw/vp8_regular_quantize_b sse2/; | 466 specialize qw/vp8_regular_quantize_b sse2/; |
453 # TODO(johann) Update sse4 implementation and re-enable | 467 # TODO(johann) Update sse4 implementation and re-enable |
454 #$vp8_regular_quantize_b_sse4_1=vp8_regular_quantize_b_sse4; | 468 #$vp8_regular_quantize_b_sse4_1=vp8_regular_quantize_b_sse4; |
455 | 469 |
456 add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *"; | 470 add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *"; |
457 specialize qw/vp8_fast_quantize_b sse2 ssse3 media neon/; | 471 specialize qw/vp8_fast_quantize_b sse2 ssse3 media neon_asm/; |
458 $vp8_fast_quantize_b_media=vp8_fast_quantize_b_armv6; | 472 $vp8_fast_quantize_b_media=vp8_fast_quantize_b_armv6; |
| 473 $vp8_fast_quantize_b_neon_asm=vp8_fast_quantize_b_neon; |
459 | 474 |
460 add_proto qw/void vp8_regular_quantize_b_pair/, "struct block *b1, struct block
*b2, struct blockd *d1, struct blockd *d2"; | 475 add_proto qw/void vp8_regular_quantize_b_pair/, "struct block *b1, struct block
*b2, struct blockd *d1, struct blockd *d2"; |
461 # no asm yet | 476 # no asm yet |
462 | 477 |
463 add_proto qw/void vp8_fast_quantize_b_pair/, "struct block *b1, struct block *b2
, struct blockd *d1, struct blockd *d2"; | 478 add_proto qw/void vp8_fast_quantize_b_pair/, "struct block *b1, struct block *b2
, struct blockd *d1, struct blockd *d2"; |
464 specialize qw/vp8_fast_quantize_b_pair neon/; | 479 specialize qw/vp8_fast_quantize_b_pair neon_asm/; |
| 480 $vp8_fast_quantize_b_pair_neon_asm=vp8_fast_quantize_b_pair_neon; |
465 | 481 |
466 add_proto qw/void vp8_quantize_mb/, "struct macroblock *"; | 482 add_proto qw/void vp8_quantize_mb/, "struct macroblock *"; |
467 specialize qw/vp8_quantize_mb neon/; | 483 specialize qw/vp8_quantize_mb neon/; |
468 | 484 |
469 add_proto qw/void vp8_quantize_mby/, "struct macroblock *"; | 485 add_proto qw/void vp8_quantize_mby/, "struct macroblock *"; |
470 specialize qw/vp8_quantize_mby neon/; | 486 specialize qw/vp8_quantize_mby neon/; |
471 | 487 |
472 add_proto qw/void vp8_quantize_mbuv/, "struct macroblock *"; | 488 add_proto qw/void vp8_quantize_mbuv/, "struct macroblock *"; |
473 specialize qw/vp8_quantize_mbuv neon/; | 489 specialize qw/vp8_quantize_mbuv neon/; |
474 | 490 |
475 # | 491 # |
476 # Block subtraction | 492 # Block subtraction |
477 # | 493 # |
478 add_proto qw/int vp8_block_error/, "short *coeff, short *dqcoeff"; | 494 add_proto qw/int vp8_block_error/, "short *coeff, short *dqcoeff"; |
479 specialize qw/vp8_block_error mmx sse2/; | 495 specialize qw/vp8_block_error mmx sse2/; |
480 $vp8_block_error_sse2=vp8_block_error_xmm; | 496 $vp8_block_error_sse2=vp8_block_error_xmm; |
481 | 497 |
482 add_proto qw/int vp8_mbblock_error/, "struct macroblock *mb, int dc"; | 498 add_proto qw/int vp8_mbblock_error/, "struct macroblock *mb, int dc"; |
483 specialize qw/vp8_mbblock_error mmx sse2/; | 499 specialize qw/vp8_mbblock_error mmx sse2/; |
484 $vp8_mbblock_error_sse2=vp8_mbblock_error_xmm; | 500 $vp8_mbblock_error_sse2=vp8_mbblock_error_xmm; |
485 | 501 |
486 add_proto qw/int vp8_mbuverror/, "struct macroblock *mb"; | 502 add_proto qw/int vp8_mbuverror/, "struct macroblock *mb"; |
487 specialize qw/vp8_mbuverror mmx sse2/; | 503 specialize qw/vp8_mbuverror mmx sse2/; |
488 $vp8_mbuverror_sse2=vp8_mbuverror_xmm; | 504 $vp8_mbuverror_sse2=vp8_mbuverror_xmm; |
489 | 505 |
490 add_proto qw/void vp8_subtract_b/, "struct block *be, struct blockd *bd, int pit
ch"; | 506 add_proto qw/void vp8_subtract_b/, "struct block *be, struct blockd *bd, int pit
ch"; |
491 specialize qw/vp8_subtract_b mmx sse2 media neon/; | 507 specialize qw/vp8_subtract_b mmx sse2 media neon_asm/; |
492 $vp8_subtract_b_media=vp8_subtract_b_armv6; | 508 $vp8_subtract_b_media=vp8_subtract_b_armv6; |
| 509 $vp8_subtract_b_neon_asm=vp8_subtract_b_neon; |
493 | 510 |
494 add_proto qw/void vp8_subtract_mby/, "short *diff, unsigned char *src, int src_s
tride, unsigned char *pred, int pred_stride"; | 511 add_proto qw/void vp8_subtract_mby/, "short *diff, unsigned char *src, int src_s
tride, unsigned char *pred, int pred_stride"; |
495 specialize qw/vp8_subtract_mby mmx sse2 media neon/; | 512 specialize qw/vp8_subtract_mby mmx sse2 media neon_asm/; |
496 $vp8_subtract_mby_media=vp8_subtract_mby_armv6; | 513 $vp8_subtract_mby_media=vp8_subtract_mby_armv6; |
| 514 $vp8_subtract_mby_neon_asm=vp8_subtract_mby_neon; |
497 | 515 |
498 add_proto qw/void vp8_subtract_mbuv/, "short *diff, unsigned char *usrc, unsigne
d char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pr
ed_stride"; | 516 add_proto qw/void vp8_subtract_mbuv/, "short *diff, unsigned char *usrc, unsigne
d char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pr
ed_stride"; |
499 specialize qw/vp8_subtract_mbuv mmx sse2 media neon/; | 517 specialize qw/vp8_subtract_mbuv mmx sse2 media neon_asm/; |
500 $vp8_subtract_mbuv_media=vp8_subtract_mbuv_armv6; | 518 $vp8_subtract_mbuv_media=vp8_subtract_mbuv_armv6; |
| 519 $vp8_subtract_mbuv_neon_asm=vp8_subtract_mbuv_neon; |
501 | 520 |
502 # | 521 # |
503 # Motion search | 522 # Motion search |
504 # | 523 # |
505 add_proto qw/int vp8_full_search_sad/, "struct macroblock *x, struct block *b, s
truct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct var
iance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv"; | 524 add_proto qw/int vp8_full_search_sad/, "struct macroblock *x, struct block *b, s
truct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct var
iance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv"; |
506 specialize qw/vp8_full_search_sad sse3 sse4_1/; | 525 specialize qw/vp8_full_search_sad sse3 sse4_1/; |
507 $vp8_full_search_sad_sse3=vp8_full_search_sadx3; | 526 $vp8_full_search_sad_sse3=vp8_full_search_sadx3; |
508 $vp8_full_search_sad_sse4_1=vp8_full_search_sadx8; | 527 $vp8_full_search_sad_sse4_1=vp8_full_search_sadx8; |
509 | 528 |
510 add_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"; | 529 add_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"; |
511 specialize qw/vp8_refining_search_sad sse3/; | 530 specialize qw/vp8_refining_search_sad sse3/; |
512 $vp8_refining_search_sad_sse3=vp8_refining_search_sadx4; | 531 $vp8_refining_search_sad_sse3=vp8_refining_search_sadx4; |
513 | 532 |
514 add_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_para
m, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2],
union int_mv *center_mv"; | 533 add_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_para
m, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2],
union int_mv *center_mv"; |
515 $vp8_diamond_search_sad_sse3=vp8_diamond_search_sadx4; | 534 $vp8_diamond_search_sad_sse3=vp8_diamond_search_sadx4; |
516 | 535 |
517 # | 536 # |
518 # Alt-ref Noise Reduction (ARNR) | 537 # Alt-ref Noise Reduction (ARNR) |
519 # | 538 # |
520 if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") { | 539 if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") { |
521 add_proto qw/void vp8_temporal_filter_apply/, "unsigned char *frame1, unsign
ed int stride, unsigned char *frame2, unsigned int block_size, int strength, int
filter_weight, unsigned int *accumulator, unsigned short *count"; | 540 add_proto qw/void vp8_temporal_filter_apply/, "unsigned char *frame1, unsign
ed int stride, unsigned char *frame2, unsigned int block_size, int strength, int
filter_weight, unsigned int *accumulator, unsigned short *count"; |
522 specialize qw/vp8_temporal_filter_apply sse2/; | 541 specialize qw/vp8_temporal_filter_apply sse2/; |
523 } | 542 } |
524 | 543 |
525 # | 544 # |
526 # Pick Loopfilter | 545 # Pick Loopfilter |
527 # | 546 # |
528 add_proto qw/void vp8_yv12_copy_partial_frame/, "struct yv12_buffer_config *src_
ybc, struct yv12_buffer_config *dst_ybc"; | 547 add_proto qw/void vp8_yv12_copy_partial_frame/, "struct yv12_buffer_config *src_
ybc, struct yv12_buffer_config *dst_ybc"; |
529 specialize qw/vp8_yv12_copy_partial_frame neon/; | 548 specialize qw/vp8_yv12_copy_partial_frame neon_asm/; |
| 549 $vp8_yv12_copy_partial_frame_neon_asm=vp8_yv12_copy_partial_frame_neon; |
530 | 550 |
531 # | 551 # |
532 # Denoiser filter | 552 # Denoiser filter |
533 # | 553 # |
534 if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") { | 554 if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") { |
535 add_proto qw/int vp8_denoiser_filter/, "struct yv12_buffer_config* mc_runnin
g_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsign
ed int motion_magnitude2, int y_offset, int uv_offset"; | 555 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"; |
536 specialize qw/vp8_denoiser_filter sse2 neon/; | 556 specialize qw/vp8_denoiser_filter sse2 neon/; |
537 } | 557 } |
538 | 558 |
539 # End of encoder only functions | 559 # End of encoder only functions |
540 } | 560 } |
541 1; | 561 1; |
OLD | NEW |