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 void av1_highbd_convolve8_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, int bps);
125 #define av1_highbd_convolve8 av1_highbd_convolve8_sse2
126 
127 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);
128 void av1_highbd_convolve8_horiz_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, int bps);
129 #define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_sse2
130 
131 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);
132 void av1_highbd_convolve8_vert_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, int bps);
133 #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2
134 
135 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);
136 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);
137 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);
138 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);
139 
140 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);
141 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);
142 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);
143 
144 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);
145 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);
146 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);
147 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);
148 
149 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);
150 #define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
151 
152 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);
153 #define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
154 
155 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);
156 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);
157 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);
158 
159 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);
160 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);
161 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);
162 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);
163 
164 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);
165 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);
166 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);
167 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);
168 
169 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);
170 #define av1_highbd_dr_prediction_z1 av1_highbd_dr_prediction_z1_c
171 
172 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);
173 #define av1_highbd_dr_prediction_z2 av1_highbd_dr_prediction_z2_c
174 
175 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);
176 #define av1_highbd_dr_prediction_z3 av1_highbd_dr_prediction_z3_c
177 
178 void av1_highbd_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
179 void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
180 void av1_highbd_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
181 RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
182 
183 void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
184 void av1_highbd_inv_txfm_add_16x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
185 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
186 
187 void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
188 void av1_highbd_inv_txfm_add_16x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
189 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
190 
191 void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
192 void av1_highbd_inv_txfm_add_32x32_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
193 void av1_highbd_inv_txfm_add_32x32_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
194 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
195 
196 void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
197 void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
198 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
199 
200 void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
201 void av1_highbd_inv_txfm_add_8x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
202 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
203 
204 void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
205 void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
206 RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
207 
208 void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
209 #define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c
210 
211 void av1_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
212 #define av1_highbd_iwht4x4_1_add av1_highbd_iwht4x4_1_add_c
213 
214 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);
215 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);
216 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);
217 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);
218 
219 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);
220 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);
221 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);
222 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);
223 
224 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);
225 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);
226 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);
227 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);
228 
229 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);
230 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);
231 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);
232 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);
233 
234 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);
235 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);
236 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);
237 
238 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);
239 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);
240 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);
241 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);
242 
243 void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
244 #define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
245 
246 void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
247 #define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
248 
249 void av1_inv_txfm2d_add_16x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
250 #define av1_inv_txfm2d_add_16x4 av1_inv_txfm2d_add_16x4_c
251 
252 void av1_inv_txfm2d_add_16x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
253 #define av1_inv_txfm2d_add_16x64 av1_inv_txfm2d_add_16x64_c
254 
255 void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
256 #define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
257 
258 void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
259 #define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
260 
261 void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
262 #define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
263 
264 void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
265 #define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
266 
267 void av1_inv_txfm2d_add_32x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
268 #define av1_inv_txfm2d_add_32x8 av1_inv_txfm2d_add_32x8_c
269 
270 void av1_inv_txfm2d_add_4x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
271 #define av1_inv_txfm2d_add_4x16 av1_inv_txfm2d_add_4x16_c
272 
273 void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
274 void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
275 RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
276 
277 void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
278 #define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
279 
280 void av1_inv_txfm2d_add_64x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
281 #define av1_inv_txfm2d_add_64x16 av1_inv_txfm2d_add_64x16_c
282 
283 void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
284 #define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
285 
286 void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
287 #define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
288 
289 void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
290 #define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
291 
292 void av1_inv_txfm2d_add_8x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
293 #define av1_inv_txfm2d_add_8x32 av1_inv_txfm2d_add_8x32_c
294 
295 void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
296 #define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
297 
298 void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
299 void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
300 RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
301 
302 void av1_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
303 void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
304 void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
305 RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
306 
307 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);
308 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);
309 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);
310 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);
311 
312 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);
313 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);
314 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);
315 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);
316 
317 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);
318 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);
319 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);
320 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);
321 
322 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);
323 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);
324 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);
325 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);
326 
327 int av1_selfguided_restoration_c(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_sse4_1(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 int av1_selfguided_restoration_avx2(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 RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
337                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
338                                  int sgr_params_idx, int bit_depth, int highbd);
339 
340 void av1_upsample_intra_edge_c(uint8_t *p, int sz);
341 void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
342 RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
343 
344 void av1_upsample_intra_edge_high_c(uint16_t *p, int sz, int bd);
345 void av1_upsample_intra_edge_high_sse4_1(uint16_t *p, int sz, int bd);
346 RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t *p, int sz, int bd);
347 
348 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);
349 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);
350 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);
351 
352 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);
353 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);
354 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);
355 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);
356 
357 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);
358 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);
359 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);
360 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);
361 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);
362 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);
363 
364 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
365 int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
366 int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
367 int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
368 int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
369 RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
370 
371 cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);
372 cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);
373 cfl_subsample_hbd_fn cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);
374 RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);
375 
376 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_c(TX_SIZE tx_size);
377 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_ssse3(TX_SIZE tx_size);
378 cfl_subsample_lbd_fn cfl_get_luma_subsampling_420_lbd_avx2(TX_SIZE tx_size);
379 RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_420_lbd)(TX_SIZE tx_size);
380 
381 cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);
382 cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);
383 cfl_subsample_hbd_fn cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);
384 RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);
385 
386 cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_c(TX_SIZE tx_size);
387 cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_ssse3(TX_SIZE tx_size);
388 cfl_subsample_lbd_fn cfl_get_luma_subsampling_422_lbd_avx2(TX_SIZE tx_size);
389 RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_422_lbd)(TX_SIZE tx_size);
390 
391 cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);
392 cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);
393 cfl_subsample_hbd_fn cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);
394 RTCD_EXTERN cfl_subsample_hbd_fn (*cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);
395 
396 cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_c(TX_SIZE tx_size);
397 cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_ssse3(TX_SIZE tx_size);
398 cfl_subsample_lbd_fn cfl_get_luma_subsampling_444_lbd_avx2(TX_SIZE tx_size);
399 RTCD_EXTERN cfl_subsample_lbd_fn (*cfl_get_luma_subsampling_444_lbd)(TX_SIZE tx_size);
400 
401 void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
402 void copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
403 void copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
404 void copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
405 void copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
406 RTCD_EXTERN void (*copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
407 
408 void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
409 void copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
410 void copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
411 void copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
412 void copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
413 RTCD_EXTERN void (*copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
414 
415 cfl_predict_hbd_fn get_predict_hbd_fn_c(TX_SIZE tx_size);
416 cfl_predict_hbd_fn get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
417 cfl_predict_hbd_fn get_predict_hbd_fn_avx2(TX_SIZE tx_size);
418 RTCD_EXTERN cfl_predict_hbd_fn (*get_predict_hbd_fn)(TX_SIZE tx_size);
419 
420 cfl_predict_lbd_fn get_predict_lbd_fn_c(TX_SIZE tx_size);
421 cfl_predict_lbd_fn get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
422 cfl_predict_lbd_fn get_predict_lbd_fn_avx2(TX_SIZE tx_size);
423 RTCD_EXTERN cfl_predict_lbd_fn (*get_predict_lbd_fn)(TX_SIZE tx_size);
424 
425 cfl_subtract_average_fn get_subtract_average_fn_c(TX_SIZE tx_size);
426 cfl_subtract_average_fn get_subtract_average_fn_sse2(TX_SIZE tx_size);
427 cfl_subtract_average_fn get_subtract_average_fn_avx2(TX_SIZE tx_size);
428 RTCD_EXTERN cfl_subtract_average_fn (*get_subtract_average_fn)(TX_SIZE tx_size);
429 
430 void av1_rtcd(void);
431 
432 #ifdef RTCD_C
433 #include "aom_ports/x86.h"
setup_rtcd_internal(void)434 static void setup_rtcd_internal(void)
435 {
436     int flags = x86_simd_caps();
437 
438     (void)flags;
439 
440     apply_selfguided_restoration = apply_selfguided_restoration_c;
441     if (flags & HAS_SSE4_1) apply_selfguided_restoration = apply_selfguided_restoration_sse4_1;
442     if (flags & HAS_AVX2) apply_selfguided_restoration = apply_selfguided_restoration_avx2;
443     av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
444     if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
445     if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
446     av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
447     if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
448     if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
449     av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
450     if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
451     if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
452     av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2;
453     if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_avx2;
454     av1_convolve_2d_scale = av1_convolve_2d_scale_c;
455     if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
456     av1_convolve_2d_sr = av1_convolve_2d_sr_sse2;
457     if (flags & HAS_AVX2) av1_convolve_2d_sr = av1_convolve_2d_sr_avx2;
458     av1_convolve_horiz_rs = av1_convolve_horiz_rs_c;
459     if (flags & HAS_SSE4_1) av1_convolve_horiz_rs = av1_convolve_horiz_rs_sse4_1;
460     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_sse2;
463     if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
464     av1_filter_intra_edge = av1_filter_intra_edge_c;
465     if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
466     av1_filter_intra_edge_high = av1_filter_intra_edge_high_c;
467     if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1;
468     av1_filter_intra_predictor = av1_filter_intra_predictor_c;
469     if (flags & HAS_SSE4_1) av1_filter_intra_predictor = av1_filter_intra_predictor_sse4_1;
470     av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2;
471     if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_avx2;
472     av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
473     if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
474     av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_c;
475     if (flags & HAS_SSSE3) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_ssse3;
476     if (flags & HAS_AVX2) av1_highbd_convolve_2d_sr = av1_highbd_convolve_2d_sr_avx2;
477     av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_c;
478     if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz_rs = av1_highbd_convolve_horiz_rs_sse4_1;
479     av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_c;
480     if (flags & HAS_SSSE3) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_ssse3;
481     if (flags & HAS_AVX2) av1_highbd_convolve_x_sr = av1_highbd_convolve_x_sr_avx2;
482     av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
483     if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
484     if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
485     av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
486     if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
487     if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
488     av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_c;
489     if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_sse4_1;
490     av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c;
491     if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_sse4_1;
492     av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c;
493     if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_sse4_1;
494     if (flags & HAS_AVX2) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_avx2;
495     av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
496     if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
497     av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_c;
498     if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_sse4_1;
499     av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
500     if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
501     av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_c;
502     if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_sse4_1;
503     if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_avx2;
504     av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_c;
505     if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_sse4_1;
506     if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_avx2;
507     av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_c;
508     if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_sse4_1;
509     if (flags & HAS_AVX2) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_avx2;
510     av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_c;
511     if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_sse4_1;
512     if (flags & HAS_AVX2) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_avx2;
513     av1_highbd_warp_affine = av1_highbd_warp_affine_c;
514     if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
515     av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
516     if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
517     if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
518     av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
519     if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
520     av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
521     if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
522     av1_inv_txfm_add = av1_inv_txfm_add_c;
523     if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
524     if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
525     av1_jnt_convolve_2d = av1_jnt_convolve_2d_c;
526     if (flags & HAS_SSSE3) av1_jnt_convolve_2d = av1_jnt_convolve_2d_ssse3;
527     if (flags & HAS_AVX2) av1_jnt_convolve_2d = av1_jnt_convolve_2d_avx2;
528     av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_sse2;
529     if (flags & HAS_AVX2) av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_avx2;
530     av1_jnt_convolve_x = av1_jnt_convolve_x_sse2;
531     if (flags & HAS_AVX2) av1_jnt_convolve_x = av1_jnt_convolve_x_avx2;
532     av1_jnt_convolve_y = av1_jnt_convolve_y_sse2;
533     if (flags & HAS_AVX2) av1_jnt_convolve_y = av1_jnt_convolve_y_avx2;
534     av1_selfguided_restoration = av1_selfguided_restoration_c;
535     if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
536     if (flags & HAS_AVX2) av1_selfguided_restoration = av1_selfguided_restoration_avx2;
537     av1_upsample_intra_edge = av1_upsample_intra_edge_c;
538     if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
539     av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
540     if (flags & HAS_SSE4_1) av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
541     av1_warp_affine = av1_warp_affine_c;
542     if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
543     av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
544     if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
545     cdef_filter_block = cdef_filter_block_sse2;
546     if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3;
547     if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1;
548     if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2;
549     cdef_find_dir = cdef_find_dir_sse2;
550     if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
551     if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
552     if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
553     cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
554     if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_ssse3;
555     if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_avx2;
556     cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_c;
557     if (flags & HAS_SSSE3) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_ssse3;
558     if (flags & HAS_AVX2) cfl_get_luma_subsampling_420_lbd = cfl_get_luma_subsampling_420_lbd_avx2;
559     cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_c;
560     if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_ssse3;
561     if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_hbd = cfl_get_luma_subsampling_422_hbd_avx2;
562     cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_c;
563     if (flags & HAS_SSSE3) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_ssse3;
564     if (flags & HAS_AVX2) cfl_get_luma_subsampling_422_lbd = cfl_get_luma_subsampling_422_lbd_avx2;
565     cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_c;
566     if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_ssse3;
567     if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_hbd = cfl_get_luma_subsampling_444_hbd_avx2;
568     cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_c;
569     if (flags & HAS_SSSE3) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_ssse3;
570     if (flags & HAS_AVX2) cfl_get_luma_subsampling_444_lbd = cfl_get_luma_subsampling_444_lbd_avx2;
571     copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse2;
572     if (flags & HAS_SSSE3) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_ssse3;
573     if (flags & HAS_SSE4_1) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse4_1;
574     if (flags & HAS_AVX2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_avx2;
575     copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse2;
576     if (flags & HAS_SSSE3) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_ssse3;
577     if (flags & HAS_SSE4_1) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse4_1;
578     if (flags & HAS_AVX2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_avx2;
579     get_predict_hbd_fn = get_predict_hbd_fn_c;
580     if (flags & HAS_SSSE3) get_predict_hbd_fn = get_predict_hbd_fn_ssse3;
581     if (flags & HAS_AVX2) get_predict_hbd_fn = get_predict_hbd_fn_avx2;
582     get_predict_lbd_fn = get_predict_lbd_fn_c;
583     if (flags & HAS_SSSE3) get_predict_lbd_fn = get_predict_lbd_fn_ssse3;
584     if (flags & HAS_AVX2) get_predict_lbd_fn = get_predict_lbd_fn_avx2;
585     get_subtract_average_fn = get_subtract_average_fn_sse2;
586     if (flags & HAS_AVX2) get_subtract_average_fn = get_subtract_average_fn_avx2;
587 }
588 #endif
589 
590 #ifdef __cplusplus
591 }  // extern "C"
592 #endif
593 
594 #endif
595