1 #ifndef VP8_RTCD_H_
2 #define VP8_RTCD_H_
3
4 #ifdef RTCD_C
5 #define RTCD_EXTERN
6 #else
7 #define RTCD_EXTERN extern
8 #endif
9
10 /*
11 * VP8
12 */
13
14 struct blockd;
15 struct macroblockd;
16 struct loop_filter_info;
17
18 /* Encoder forward decls */
19 struct block;
20 struct macroblock;
21 struct variance_vtable;
22 union int_mv;
23 struct yv12_buffer_config;
24
25 #ifdef __cplusplus
26 extern "C" {
27 #endif
28
29 void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
30 void vp8_bilinear_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
31 void vp8_bilinear_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
32 RTCD_EXTERN void (*vp8_bilinear_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
33
34 void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
35 void vp8_bilinear_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
36 #define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_mmx
37
38 void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
39 void vp8_bilinear_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
40 #define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_mmx
41
42 void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
43 void vp8_bilinear_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
44 void vp8_bilinear_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
45 RTCD_EXTERN void (*vp8_bilinear_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
46
47 void vp8_blend_b_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
48 #define vp8_blend_b vp8_blend_b_c
49
50 void vp8_blend_mb_inner_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
51 #define vp8_blend_mb_inner vp8_blend_mb_inner_c
52
53 void vp8_blend_mb_outer_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
54 #define vp8_blend_mb_outer vp8_blend_mb_outer_c
55
56 int vp8_block_error_c(short *coeff, short *dqcoeff);
57 int vp8_block_error_sse2(short *coeff, short *dqcoeff);
58 #define vp8_block_error vp8_block_error_sse2
59
60 void vp8_copy32xn_c(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
61 void vp8_copy32xn_sse2(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
62 void vp8_copy32xn_sse3(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
63 RTCD_EXTERN void (*vp8_copy32xn)(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
64
65 void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
66 void vp8_copy_mem16x16_sse2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
67 #define vp8_copy_mem16x16 vp8_copy_mem16x16_sse2
68
69 void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
70 void vp8_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
71 #define vp8_copy_mem8x4 vp8_copy_mem8x4_mmx
72
73 void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
74 void vp8_copy_mem8x8_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
75 #define vp8_copy_mem8x8 vp8_copy_mem8x8_mmx
76
77 void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
78 void vp8_dc_only_idct_add_mmx(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
79 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx
80
81 int vp8_denoiser_filter_c(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);
82 int vp8_denoiser_filter_sse2(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);
83 #define vp8_denoiser_filter vp8_denoiser_filter_sse2
84
85 int vp8_denoiser_filter_uv_c(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);
86 int vp8_denoiser_filter_uv_sse2(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);
87 #define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_sse2
88
89 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
90 void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *output, int stride);
91 #define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
92
93 void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
94 void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
95 #define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_sse2
96
97 void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
98 void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
99 #define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_sse2
100
101 void vp8_dequantize_b_c(struct blockd*, short *dqc);
102 void vp8_dequantize_b_mmx(struct blockd*, short *dqc);
103 #define vp8_dequantize_b vp8_dequantize_b_mmx
104
105 int vp8_diamond_search_sad_c(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);
106 int vp8_diamond_search_sadx4(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);
107 #define vp8_diamond_search_sad vp8_diamond_search_sadx4
108
109 void vp8_fast_quantize_b_c(struct block *, struct blockd *);
110 void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
111 void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *);
112 RTCD_EXTERN void (*vp8_fast_quantize_b)(struct block *, struct blockd *);
113
114 void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
115 void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
116 #define vp8_filter_by_weight16x16 vp8_filter_by_weight16x16_sse2
117
118 void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
119 #define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c
120
121 void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
122 void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
123 #define vp8_filter_by_weight8x8 vp8_filter_by_weight8x8_sse2
124
125 int vp8_full_search_sad_c(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);
126 int vp8_full_search_sadx3(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);
127 int vp8_full_search_sadx8(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);
128 RTCD_EXTERN 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);
129
130 void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
131 void vp8_loop_filter_bh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
132 #define vp8_loop_filter_bh vp8_loop_filter_bh_sse2
133
134 void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
135 void vp8_loop_filter_bv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
136 #define vp8_loop_filter_bv vp8_loop_filter_bv_sse2
137
138 void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
139 void vp8_loop_filter_mbh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
140 #define vp8_loop_filter_mbh vp8_loop_filter_mbh_sse2
141
142 void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
143 void vp8_loop_filter_mbv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
144 #define vp8_loop_filter_mbv vp8_loop_filter_mbv_sse2
145
146 void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
147 void vp8_loop_filter_bhs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
148 #define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_sse2
149
150 void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
151 void vp8_loop_filter_bvs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
152 #define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_sse2
153
154 void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
155 void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
156 #define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_sse2
157
158 void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
159 void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
160 #define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_sse2
161
162 int vp8_mbblock_error_c(struct macroblock *mb, int dc);
163 int vp8_mbblock_error_sse2(struct macroblock *mb, int dc);
164 #define vp8_mbblock_error vp8_mbblock_error_sse2
165
166 int vp8_mbuverror_c(struct macroblock *mb);
167 int vp8_mbuverror_sse2(struct macroblock *mb);
168 #define vp8_mbuverror vp8_mbuverror_sse2
169
170 int vp8_refining_search_sad_c(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);
171 int vp8_refining_search_sadx4(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);
172 #define vp8_refining_search_sad vp8_refining_search_sadx4
173
174 void vp8_regular_quantize_b_c(struct block *, struct blockd *);
175 void vp8_regular_quantize_b_sse2(struct block *, struct blockd *);
176 void vp8_regular_quantize_b_sse4_1(struct block *, struct blockd *);
177 RTCD_EXTERN void (*vp8_regular_quantize_b)(struct block *, struct blockd *);
178
179 void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
180 void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch);
181 #define vp8_short_fdct4x4 vp8_short_fdct4x4_sse2
182
183 void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
184 void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch);
185 #define vp8_short_fdct8x4 vp8_short_fdct8x4_sse2
186
187 void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
188 void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
189 #define vp8_short_idct4x4llm vp8_short_idct4x4llm_mmx
190
191 void vp8_short_inv_walsh4x4_c(short *input, short *output);
192 void vp8_short_inv_walsh4x4_sse2(short *input, short *output);
193 #define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_sse2
194
195 void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
196 #define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
197
198 void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
199 void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch);
200 #define vp8_short_walsh4x4 vp8_short_walsh4x4_sse2
201
202 void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
203 void vp8_sixtap_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
204 void vp8_sixtap_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
205 RTCD_EXTERN void (*vp8_sixtap_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
206
207 void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
208 void vp8_sixtap_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
209 void vp8_sixtap_predict4x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
210 RTCD_EXTERN void (*vp8_sixtap_predict4x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
211
212 void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
213 void vp8_sixtap_predict8x4_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
214 void vp8_sixtap_predict8x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
215 RTCD_EXTERN void (*vp8_sixtap_predict8x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
216
217 void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
218 void vp8_sixtap_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
219 void vp8_sixtap_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
220 RTCD_EXTERN void (*vp8_sixtap_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
221
222 void vp8_temporal_filter_apply_c(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count);
223 void vp8_temporal_filter_apply_sse2(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count);
224 #define vp8_temporal_filter_apply vp8_temporal_filter_apply_sse2
225
226 void vp8_rtcd(void);
227
228 #ifdef RTCD_C
229 #include "vpx_ports/x86.h"
setup_rtcd_internal(void)230 static void setup_rtcd_internal(void)
231 {
232 int flags = x86_simd_caps();
233
234 (void)flags;
235
236 vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_sse2;
237 if (flags & HAS_SSSE3) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_ssse3;
238 vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_sse2;
239 if (flags & HAS_SSSE3) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_ssse3;
240 vp8_copy32xn = vp8_copy32xn_sse2;
241 if (flags & HAS_SSE3) vp8_copy32xn = vp8_copy32xn_sse3;
242 vp8_fast_quantize_b = vp8_fast_quantize_b_sse2;
243 if (flags & HAS_SSSE3) vp8_fast_quantize_b = vp8_fast_quantize_b_ssse3;
244 vp8_full_search_sad = vp8_full_search_sad_c;
245 if (flags & HAS_SSE3) vp8_full_search_sad = vp8_full_search_sadx3;
246 if (flags & HAS_SSE4_1) vp8_full_search_sad = vp8_full_search_sadx8;
247 vp8_regular_quantize_b = vp8_regular_quantize_b_sse2;
248 if (flags & HAS_SSE4_1) vp8_regular_quantize_b = vp8_regular_quantize_b_sse4_1;
249 vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_sse2;
250 if (flags & HAS_SSSE3) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_ssse3;
251 vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_mmx;
252 if (flags & HAS_SSSE3) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_ssse3;
253 vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_sse2;
254 if (flags & HAS_SSSE3) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_ssse3;
255 vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_sse2;
256 if (flags & HAS_SSSE3) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_ssse3;
257 }
258 #endif
259
260 #ifdef __cplusplus
261 } // extern "C"
262 #endif
263
264 #endif
265