1 // This file is generated. Do not edit.
2 #ifndef AV1_RTCD_H_
3 #define AV1_RTCD_H_
4
5 #ifdef RTCD_C
6 #define RTCD_EXTERN
7 #else
8 #define RTCD_EXTERN extern
9 #endif
10
11 /*
12 * AV1
13 */
14
15 #include "aom/aom_integer.h"
16 #include "aom_dsp/txfm_common.h"
17 #include "av1/common/common.h"
18 #include "av1/common/enums.h"
19 #include "av1/common/quant_common.h"
20 #include "av1/common/filter.h"
21 #include "av1/common/convolve.h"
22 #include "av1/common/av1_txfm.h"
23 #include "av1/common/odintrin.h"
24 #include "av1/common/restoration.h"
25
26 struct macroblockd;
27
28 /* Encoder forward decls */
29 struct macroblock;
30 struct txfm_param;
31 struct aom_variance_vtable;
32 struct search_site_config;
33 struct yv12_buffer_config;
34
35 /* Function pointers return by CfL functions */
36 typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
37 uint16_t *output_q3);
38
39 typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
40 uint16_t *output_q3);
41
42 typedef void (*cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);
43
44 typedef void (*cfl_predict_lbd_fn)(const int16_t *src, uint8_t *dst,
45 int dst_stride, int alpha_q3);
46
47 typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
48 int dst_stride, int alpha_q3, int bd);
49
50 #ifdef __cplusplus
51 extern "C" {
52 #endif
53
54 void apply_selfguided_restoration_c(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
55 void apply_selfguided_restoration_sse4_1(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
56 void apply_selfguided_restoration_avx2(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
57 RTCD_EXTERN void (*apply_selfguided_restoration)(const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf, int bit_depth, int highbd);
58
59 void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
60 void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
61 void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
62 RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
63
64 void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
65 void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
66 void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
67 RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
68
69 void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
70 void av1_build_compound_diffwtd_mask_highbd_ssse3(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
71 void av1_build_compound_diffwtd_mask_highbd_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
72 RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
73
74 void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
75 void av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
76 void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
77 RTCD_EXTERN void (*av1_convolve_2d_copy_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
78
79 void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params);
80 void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params);
81 RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params);
82
83 void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
84 void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
85 void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
86 RTCD_EXTERN void (*av1_convolve_2d_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
87
88 void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
89 void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
90 RTCD_EXTERN void (*av1_convolve_horiz_rs)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn);
91
92 void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
93 void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
94 void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
95 RTCD_EXTERN void (*av1_convolve_x_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
96
97 void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
98 void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
99 void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
100 RTCD_EXTERN void (*av1_convolve_y_sr)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
101
102 void av1_dr_prediction_z1_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int dx, int dy);
103 #define av1_dr_prediction_z1 av1_dr_prediction_z1_c
104
105 void av1_dr_prediction_z2_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_above, int upsample_left, int dx, int dy);
106 #define av1_dr_prediction_z2 av1_dr_prediction_z2_c
107
108 void av1_dr_prediction_z3_c(uint8_t *dst, ptrdiff_t stride, int bw, int bh, const uint8_t *above, const uint8_t *left, int upsample_left, int dx, int dy);
109 #define av1_dr_prediction_z3 av1_dr_prediction_z3_c
110
111 void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
112 void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength);
113 RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
114
115 void av1_filter_intra_edge_high_c(uint16_t *p, int sz, int strength);
116 void av1_filter_intra_edge_high_sse4_1(uint16_t *p, int sz, int strength);
117 RTCD_EXTERN void (*av1_filter_intra_edge_high)(uint16_t *p, int sz, int strength);
118
119 void av1_filter_intra_predictor_c(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
120 void av1_filter_intra_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
121 RTCD_EXTERN void (*av1_filter_intra_predictor)(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left, int mode);
122
123 void av1_highbd_convolve8_c(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);
124 #define av1_highbd_convolve8 av1_highbd_convolve8_c
125
126 void av1_highbd_convolve8_horiz_c(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);
127 #define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_c
128
129 void av1_highbd_convolve8_vert_c(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);
130 #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_c
131
132 void av1_highbd_convolve_2d_copy_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
133 void av1_highbd_convolve_2d_copy_sr_sse2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
134 void av1_highbd_convolve_2d_copy_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
135 RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
136
137 void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd);
138 void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd);
139 RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd);
140
141 void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
142 void av1_highbd_convolve_2d_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
143 void av1_highbd_convolve_2d_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
144 RTCD_EXTERN void (*av1_highbd_convolve_2d_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
145
146 void av1_highbd_convolve_avg_c(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);
147 #define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
148
149 void av1_highbd_convolve_copy_c(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);
150 #define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
151
152 void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
153 void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
154 RTCD_EXTERN void (*av1_highbd_convolve_horiz_rs)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int x0_qn, int x_step_qn, int bd);
155
156 void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
157 void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
158 void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
159 RTCD_EXTERN void (*av1_highbd_convolve_x_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
160
161 void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
162 void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
163 void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
164 RTCD_EXTERN void (*av1_highbd_convolve_y_sr)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
165
166 void av1_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int dx, int dy, int bd);
167 #define av1_highbd_dr_prediction_z1 av1_highbd_dr_prediction_z1_c
168
169 void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_above, int upsample_left, int dx, int dy, int bd);
170 #define av1_highbd_dr_prediction_z2 av1_highbd_dr_prediction_z2_c
171
172 void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
173 #define av1_highbd_dr_prediction_z3 av1_highbd_dr_prediction_z3_c
174
175 void av1_highbd_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
176 void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
177 void av1_highbd_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
178 RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
179
180 void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
181 void av1_highbd_inv_txfm_add_16x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
182 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
183
184 void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
185 void av1_highbd_inv_txfm_add_16x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
186 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
187
188 void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
189 void av1_highbd_inv_txfm_add_32x32_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
190 void av1_highbd_inv_txfm_add_32x32_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
191 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
192
193 void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
194 void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
195 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
196
197 void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
198 void av1_highbd_inv_txfm_add_8x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
199 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
200
201 void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
202 void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
203 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
204
205 void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
206 #define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c
207
208 void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
209 #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
210
211 void av1_highbd_jnt_convolve_2d_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
212 void av1_highbd_jnt_convolve_2d_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
213 void av1_highbd_jnt_convolve_2d_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
214 RTCD_EXTERN void (*av1_highbd_jnt_convolve_2d)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
215
216 void av1_highbd_jnt_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
217 void av1_highbd_jnt_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
218 void av1_highbd_jnt_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
219 RTCD_EXTERN void (*av1_highbd_jnt_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
220
221 void av1_highbd_jnt_convolve_x_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
222 void av1_highbd_jnt_convolve_x_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
223 void av1_highbd_jnt_convolve_x_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
224 RTCD_EXTERN void (*av1_highbd_jnt_convolve_x)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
225
226 void av1_highbd_jnt_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
227 void av1_highbd_jnt_convolve_y_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
228 void av1_highbd_jnt_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
229 RTCD_EXTERN void (*av1_highbd_jnt_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
230
231 void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
232 void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
233 RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
234
235 void av1_highbd_wiener_convolve_add_src_c(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, const ConvolveParams *conv_params, int bps);
236 void av1_highbd_wiener_convolve_add_src_ssse3(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, const ConvolveParams *conv_params, int bps);
237 void av1_highbd_wiener_convolve_add_src_avx2(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, const ConvolveParams *conv_params, int bps);
238 RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(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, const ConvolveParams *conv_params, int bps);
239
240 void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
241 #define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
242
243 void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
244 #define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
245
246 void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
247 #define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
248
249 void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
250 #define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
251
252 void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
253 #define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
254
255 void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
256 #define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
257
258 void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
259 #define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
260
261 void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
262 #define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
263
264 void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
265 #define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
266
267 void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
268 #define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
269
270 void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
271 void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
272 RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
273
274 void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
275 #define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
276
277 void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
278 #define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
279
280 void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
281 #define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
282
283 void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
284 #define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
285
286 void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
287 #define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
288
289 void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
290 #define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
291
292 void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
293 #define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
294
295 void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
296 void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
297 RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
298
299 void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
300 void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
301 void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
302 RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
303
304 void av1_jnt_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
305 void av1_jnt_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
306 void av1_jnt_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
307 RTCD_EXTERN void (*av1_jnt_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
308
309 void av1_jnt_convolve_2d_copy_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
310 void av1_jnt_convolve_2d_copy_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
311 void av1_jnt_convolve_2d_copy_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
312 RTCD_EXTERN void (*av1_jnt_convolve_2d_copy)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
313
314 void av1_jnt_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
315 void av1_jnt_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
316 void av1_jnt_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
317 RTCD_EXTERN void (*av1_jnt_convolve_x)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
318
319 void av1_jnt_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
320 void av1_jnt_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
321 void av1_jnt_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
322 RTCD_EXTERN void (*av1_jnt_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
323
324 int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
325 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
326 int sgr_params_idx, int bit_depth, int highbd);
327 int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
328 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
329 int sgr_params_idx, int bit_depth, int highbd);
330 int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
331 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
332 int sgr_params_idx, int bit_depth, int highbd);
333 RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
334 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
335 int sgr_params_idx, int bit_depth, int highbd);
336
337 void av1_upsample_intra_edge_c(uint8_t *p, int sz);
338 void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
339 RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
340
341 void av1_upsample_intra_edge_high_c(uint16_t *p, int sz, int bd);
342 void av1_upsample_intra_edge_high_sse4_1(uint16_t *p, int sz, int bd);
343 RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t *p, int sz, int bd);
344
345 void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
346 void av1_warp_affine_sse4_1(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
347 RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
348
349 void av1_wiener_convolve_add_src_c(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, const ConvolveParams *conv_params);
350 void av1_wiener_convolve_add_src_sse2(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, const ConvolveParams *conv_params);
351 void av1_wiener_convolve_add_src_avx2(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, const ConvolveParams *conv_params);
352 RTCD_EXTERN void (*av1_wiener_convolve_add_src)(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, const ConvolveParams *conv_params);
353
354 void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
355 void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
356 void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
357 void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
358 void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
359 RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
360
361 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
362 int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
363 int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
364 int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
365 int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
366 RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
367
368 cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
369 cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);
370 cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);
371 RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
372
373 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
374 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_ssse3(TX_SIZE tx_size);
375 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_avx2(TX_SIZE tx_size);
376 RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
377
378 cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
379 cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);
380 cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);
381 RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
382
383 cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
384 cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_ssse3(TX_SIZE tx_size);
385 cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_avx2(TX_SIZE tx_size);
386 RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
387
388 cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
389 cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);
390 cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);
391 RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
392
393 cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
394 cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_ssse3(TX_SIZE tx_size);
395 cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_avx2(TX_SIZE tx_size);
396 RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
397
398 void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
399 void copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
400 void copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
401 void copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
402 void copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
403 RTCD_EXTERN void (*copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
404
405 void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
406 void copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
407 void copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
408 void copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
409 void copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
410 RTCD_EXTERN void (*copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
411
412 cfl_predict_hbd_fn get_predict_hbd_fn_c(TX_SIZE tx_size);
413 cfl_predict_hbd_fn get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
414 cfl_predict_hbd_fn get_predict_hbd_fn_avx2(TX_SIZE tx_size);
415 RTCD_EXTERN cfl_predict_hbd_fn (*get_predict_hbd_fn)(TX_SIZE tx_size);
416
417 cfl_predict_lbd_fn get_predict_lbd_fn_c(TX_SIZE tx_size);
418 cfl_predict_lbd_fn get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
419 cfl_predict_lbd_fn get_predict_lbd_fn_avx2(TX_SIZE tx_size);
420 RTCD_EXTERN cfl_predict_lbd_fn (*get_predict_lbd_fn)(TX_SIZE tx_size);
421
422 cfl_subtract_average_fn get_subtract_average_fn_c(TX_SIZE tx_size);
423 cfl_subtract_average_fn get_subtract_average_fn_sse2(TX_SIZE tx_size);
424 cfl_subtract_average_fn get_subtract_average_fn_avx2(TX_SIZE tx_size);
425 RTCD_EXTERN cfl_subtract_average_fn (*get_subtract_average_fn)(TX_SIZE tx_size);
426
427 void av1_rtcd(void);
428
429 #ifdef RTCD_C
430 #include "aom_ports/x86.h"
setup_rtcd_internal(void)431 static void setup_rtcd_internal(void)
432 {
433 int flags = x86_simd_caps();
434
435 (void)flags;
436
437 apply_selfguided_restoration = apply_selfguided_restoration_c;
438 if (flags & HAS_SSE4_1) apply_selfguided_restoration = apply_selfguided_restoration_sse4_1;
439 if (flags & HAS_AVX2) apply_selfguided_restoration = apply_selfguided_restoration_avx2;
440 av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
441 if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
442 if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
443 av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
444 if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
445 if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
446 av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
447 if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
448 if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
449 av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_c;
450 if (flags & HAS_SSE2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2;
451 if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2;
452 av1_convolve_2d_scale = av1_convolve_2d_scale_c;
453 if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
454 av1_convolve_2d_sr = av1_convolve_2d_sr_c;
455 if (flags & HAS_SSE2) av1_convolve_2d_sr = av1_convolve_2d_sr_sse2;
456 if (flags & HAS_AVX2) av1_convolve_2d_sr = av1_convolve_2d_sr_avx2;
457 av1_convolve_horiz_rs = av1_convolve_horiz_rs_c;
458 if (flags & HAS_SSE4_1) av1_convolve_horiz_rs = av1_convolve_horiz_rs_sse4_1;
459 av1_convolve_x_sr = av1_convolve_x_sr_c;
460 if (flags & HAS_SSE2) av1_convolve_x_sr = av1_convolve_x_sr_sse2;
461 if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
462 av1_convolve_y_sr = av1_convolve_y_sr_c;
463 if (flags & HAS_SSE2) av1_convolve_y_sr = av1_convolve_y_sr_sse2;
464 if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
465 av1_filter_intra_edge = av1_filter_intra_edge_c;
466 if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
467 av1_filter_intra_edge_high = av1_filter_intra_edge_high_c;
468 if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1;
469 av1_filter_intra_predictor = av1_filter_intra_predictor_c;
470 if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1;
471 av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_c;
472 if (flags & HAS_SSE2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2;
473 if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2;
474 av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
475 if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
476 av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
477 if (flags & HAS_SSSE3) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_ssse3;
478 if (flags & HAS_AVX2) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_avx2;
479 av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
480 if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_sse4_1;
481 av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
482 if (flags & HAS_SSSE3) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_ssse3;
483 if (flags & HAS_AVX2) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_avx2;
484 av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
485 if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
486 if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
487 av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
488 if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
489 if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
490 av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_c;
491 if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_sse4_1;
492 av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c;
493 if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_sse4_1;
494 av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c;
495 if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_sse4_1;
496 if (flags & HAS_AVX2) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_avx2;
497 av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
498 if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
499 av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_c;
500 if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_sse4_1;
501 av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
502 if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
503 av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_c;
504 if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_sse4_1;
505 if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_avx2;
506 av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_c;
507 if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_sse4_1;
508 if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_avx2;
509 av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_c;
510 if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_sse4_1;
511 if (flags & HAS_AVX2) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_avx2;
512 av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_c;
513 if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_sse4_1;
514 if (flags & HAS_AVX2) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_avx2;
515 av1_highbd_warp_affine = av1_highbd_warp_affine_c;
516 if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
517 av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
518 if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
519 if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
520 av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
521 if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
522 av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
523 if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
524 av1_inv_txfm_add = av1_inv_txfm_add_c;
525 if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
526 if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
527 av1_jnt_convolve_2d = av1_jnt_convolve_2d_c;
528 if (flags & HAS_SSSE3) av1_jnt_convolve_2d = av1_jnt_convolve_2d_ssse3;
529 if (flags & HAS_AVX2) av1_jnt_convolve_2d = av1_jnt_convolve_2d_avx2;
530 av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_c;
531 if (flags & HAS_SSE2) av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_sse2;
532 if (flags & HAS_AVX2) av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_avx2;
533 av1_jnt_convolve_x = av1_jnt_convolve_x_c;
534 if (flags & HAS_SSE2) av1_jnt_convolve_x = av1_jnt_convolve_x_sse2;
535 if (flags & HAS_AVX2) av1_jnt_convolve_x = av1_jnt_convolve_x_avx2;
536 av1_jnt_convolve_y = av1_jnt_convolve_y_c;
537 if (flags & HAS_SSE2) av1_jnt_convolve_y = av1_jnt_convolve_y_sse2;
538 if (flags & HAS_AVX2) av1_jnt_convolve_y = av1_jnt_convolve_y_avx2;
539 av1_selfguided_restoration = av1_selfguided_restoration_c;
540 if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
541 if (flags & HAS_AVX2) av1_selfguided_restoration = av1_selfguided_restoration_avx2;
542 av1_upsample_intra_edge = av1_upsample_intra_edge_c;
543 if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
544 av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
545 if (flags & HAS_SSE4_1) av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
546 av1_warp_affine = av1_warp_affine_c;
547 if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
548 av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_c;
549 if (flags & HAS_SSE2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
550 if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
551 cdef_filter_block = cdef_filter_block_c;
552 if (flags & HAS_SSE2) cdef_filter_block = cdef_filter_block_sse2;
553 if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3;
554 if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1;
555 if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2;
556 cdef_find_dir = cdef_find_dir_c;
557 if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2;
558 if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
559 if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
560 if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
561 cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
562 if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3;
563 if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2;
564 cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
565 if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_ssse3;
566 if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_avx2;
567 cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
568 if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_ssse3;
569 if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_avx2;
570 cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
571 if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_ssse3;
572 if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_avx2;
573 cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
574 if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_ssse3;
575 if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_avx2;
576 cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
577 if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_ssse3;
578 if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_avx2;
579 copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_c;
580 if (flags & HAS_SSE2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse2;
581 if (flags & HAS_SSSE3) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_ssse3;
582 if (flags & HAS_SSE4_1) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse4_1;
583 if (flags & HAS_AVX2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_avx2;
584 copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_c;
585 if (flags & HAS_SSE2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse2;
586 if (flags & HAS_SSSE3) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_ssse3;
587 if (flags & HAS_SSE4_1) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse4_1;
588 if (flags & HAS_AVX2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_avx2;
589 get_predict_hbd_fn = get_predict_hbd_fn_c;
590 if (flags & HAS_SSSE3) get_predict_hbd_fn = get_predict_hbd_fn_ssse3;
591 if (flags & HAS_AVX2) get_predict_hbd_fn = get_predict_hbd_fn_avx2;
592 get_predict_lbd_fn = get_predict_lbd_fn_c;
593 if (flags & HAS_SSSE3) get_predict_lbd_fn = get_predict_lbd_fn_ssse3;
594 if (flags & HAS_AVX2) get_predict_lbd_fn = get_predict_lbd_fn_avx2;
595 get_subtract_average_fn = get_subtract_average_fn_c;
596 if (flags & HAS_SSE2) get_subtract_average_fn = get_subtract_average_fn_sse2;
597 if (flags & HAS_AVX2) get_subtract_average_fn = get_subtract_average_fn_avx2;
598 }
599 #endif
600
601 #ifdef __cplusplus
602 } // extern "C"
603 #endif
604
605 #endif
606