// This file is generated. Do not edit.

#ifndef AV2_RTCD_H_

#define AV2_RTCD_H_

#ifdef RTCD_C

#define RTCD_EXTERN

#else

#define RTCD_EXTERN extern

#endif

/*

 * AV2

 */

#include "avm/avm_integer.h"

#include "avm_dsp/txfm_common.h"

#include "av2/common/common.h"

#include "av2/common/enums.h"

#include "av2/common/quant_common.h"

#include "av2/common/filter.h"

#include "av2/common/convolve.h"

#include "av2/common/av2_txfm.h"

#include "av2/common/odintrin.h"

#include "av2/common/restoration.h"

struct macroblockd;

/* Encoder forward decls */

struct macroblock;

struct txfm_param;

struct avm_variance_vtable;

struct search_site_config;

struct yv12_buffer_config;

struct NN_CONFIG;

typedef struct NN_CONFIG NN_CONFIG;

struct tcq_node_t;

struct tcq_ctx_t;

struct tcq_lf_ctx_t;

struct prequant_t;

struct tcq_rate_t;

struct tcq_coeff_ctx_t;

struct tcq_param_t;

struct LV_MAP_COEFF_COST;

enum { NONE, RELU, SOFTSIGN, SIGMOID } UENUM1BYTE(ACTIVATION);

#if CONFIG_NN_V2

enum { SOFTMAX_CROSS_ENTROPY } UENUM1BYTE(LOSS);

struct NN_CONFIG_V2;

typedef struct NN_CONFIG_V2 NN_CONFIG_V2;

struct FC_LAYER;

typedef struct FC_LAYER FC_LAYER;

#endif  // CONFIG_NN_V2

struct CNN_CONFIG;

typedef struct CNN_CONFIG CNN_CONFIG;

struct CNN_LAYER_CONFIG;

typedef struct CNN_LAYER_CONFIG CNN_LAYER_CONFIG;

struct CNN_THREAD_DATA;

typedef struct CNN_THREAD_DATA CNN_THREAD_DATA;

struct CNN_BRANCH_CONFIG;

typedef struct CNN_BRANCH_CONFIG CNN_BRANCH_CONFIG;

struct CNN_MULTI_OUT;

typedef struct CNN_MULTI_OUT CNN_MULTI_OUT;

/* Function pointers return by CfL functions */

typedef void (*av2_cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,

                                     uint16_t *output_q3);

typedef void (*av2_cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,

                                   int dst_stride, int alpha_q3, int bd);

typedef void (*av2_cfl_subtract_average_fn)(const uint16_t *src, int16_t *dst);

#ifdef __cplusplus

extern "C" {

#endif

void av2_avg_pooling_pdiff_gradients_c(int16_t *pdiff, const int pstride, int16_t *gx, int16_t *gy, const int gstride, const int bw, const int bh, const int n);

void av2_avg_pooling_pdiff_gradients_avx2(int16_t *pdiff, const int pstride, int16_t *gx, int16_t *gy, const int gstride, const int bw, const int bh, const int n);

RTCD_EXTERN void (*av2_avg_pooling_pdiff_gradients)(int16_t *pdiff, const int pstride, int16_t *gx, int16_t *gy, const int gstride, const int bw, const int bh, const int n);

void av2_bicubic_grad_interpolation_highbd_c(const int16_t *pred_src, int16_t *x_grad, int16_t *y_grad, const int stride, const int blk_width, const int blk_height);

void av2_bicubic_grad_interpolation_highbd_sse4_1(const int16_t *pred_src, int16_t *x_grad, int16_t *y_grad, const int stride, const int blk_width, const int blk_height);

void av2_bicubic_grad_interpolation_highbd_avx2(const int16_t *pred_src, int16_t *x_grad, int16_t *y_grad, const int stride, const int blk_width, const int blk_height);

RTCD_EXTERN void (*av2_bicubic_grad_interpolation_highbd)(const int16_t *pred_src, int16_t *x_grad, int16_t *y_grad, const int stride, const int blk_width, const int blk_height);

void av2_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);

void av2_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);

void av2_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);

RTCD_EXTERN void (*av2_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);

void av2_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint16_t *src0, int src0_stride, const uint16_t *src1, int src1_stride, int h, int w, int bd);

void av2_build_compound_diffwtd_mask_highbd_ssse3(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint16_t *src0, int src0_stride, const uint16_t *src1, int src1_stride, int h, int w, int bd);

void av2_build_compound_diffwtd_mask_highbd_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint16_t *src0, int src0_stride, const uint16_t *src1, int src1_stride, int h, int w, int bd);

RTCD_EXTERN void (*av2_build_compound_diffwtd_mask_highbd)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint16_t *src0, int src0_stride, const uint16_t *src1, int src1_stride, int h, int w, int bd);

void av2_cdef_copy_rect8_16bit_to_16bit_c(uint16_t *const dst, int dstride, const uint16_t *src, int sstride, int v, int h);

void av2_cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *const dst, int dstride, const uint16_t *src, int sstride, int v, int h);

void av2_cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *const dst, int dstride, const uint16_t *src, int sstride, int v, int h);

void av2_cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *const dst, int dstride, const uint16_t *src, int sstride, int v, int h);

void av2_cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *const dst, int dstride, const uint16_t *src, int sstride, int v, int h);

RTCD_EXTERN void (*av2_cdef_copy_rect8_16bit_to_16bit)(uint16_t *const dst, int dstride, const uint16_t *src, int sstride, int v, int h);

void av2_cdef_filter_16_0_c(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_0_sse2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_0_ssse3(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_0_sse4_1(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_0_avx2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

RTCD_EXTERN void (*av2_cdef_filter_16_0)(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_1_c(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_1_sse2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_1_ssse3(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_1_sse4_1(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_1_avx2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

RTCD_EXTERN void (*av2_cdef_filter_16_1)(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_2_c(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_2_sse2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_2_ssse3(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_2_sse4_1(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_2_avx2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

RTCD_EXTERN void (*av2_cdef_filter_16_2)(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_3_c(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_3_sse2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_3_ssse3(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_3_sse4_1(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

void av2_cdef_filter_16_3_avx2(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

RTCD_EXTERN void (*av2_cdef_filter_16_3)(uint16_t *const dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);

int av2_cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);

int av2_cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);

int av2_cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);

int av2_cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);

int av2_cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);

RTCD_EXTERN int (*av2_cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);

void av2_cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);

void av2_cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);

void av2_cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);

void av2_cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);

void av2_cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);

RTCD_EXTERN void (*av2_cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_420_hbd_c(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_420_hbd_ssse3(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_420_hbd_avx2(TX_SIZE tx_size);

RTCD_EXTERN av2_cfl_subsample_hbd_fn (*av2_cfl_get_luma_subsampling_420_hbd)(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_420_hbd_121_c(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_420_hbd_121_avx2(TX_SIZE tx_size);

RTCD_EXTERN av2_cfl_subsample_hbd_fn (*av2_cfl_get_luma_subsampling_420_hbd_121)(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_420_hbd_colocated_c(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_420_hbd_colocated_avx2(TX_SIZE tx_size);

RTCD_EXTERN av2_cfl_subsample_hbd_fn (*av2_cfl_get_luma_subsampling_420_hbd_colocated)(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_422_hbd_c(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_422_hbd_ssse3(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_422_hbd_avx2(TX_SIZE tx_size);

RTCD_EXTERN av2_cfl_subsample_hbd_fn (*av2_cfl_get_luma_subsampling_422_hbd)(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_444_hbd_c(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_444_hbd_ssse3(TX_SIZE tx_size);

av2_cfl_subsample_hbd_fn av2_cfl_get_luma_subsampling_444_hbd_avx2(TX_SIZE tx_size);

RTCD_EXTERN av2_cfl_subsample_hbd_fn (*av2_cfl_get_luma_subsampling_444_hbd)(TX_SIZE tx_size);

av2_cfl_predict_hbd_fn av2_cfl_get_predict_hbd_fn_c(TX_SIZE tx_size);

av2_cfl_predict_hbd_fn av2_cfl_get_predict_hbd_fn_avx2(TX_SIZE tx_size);

RTCD_EXTERN av2_cfl_predict_hbd_fn (*av2_cfl_get_predict_hbd_fn)(TX_SIZE tx_size);

av2_cfl_subtract_average_fn av2_cfl_get_subtract_average_fn_c(TX_SIZE tx_size);

av2_cfl_subtract_average_fn av2_cfl_get_subtract_average_fn_sse2(TX_SIZE tx_size);

av2_cfl_subtract_average_fn av2_cfl_get_subtract_average_fn_avx2(TX_SIZE tx_size);

RTCD_EXTERN av2_cfl_subtract_average_fn (*av2_cfl_get_subtract_average_fn)(TX_SIZE tx_size);

void av2_cnn_activate_c( float **input, int channels, int width, int height, int stride, ACTIVATION layer_activation);

#define av2_cnn_activate av2_cnn_activate_c

void av2_cnn_add_c( float **input, int channels, int width, int height, int stride, const float **add);

#define av2_cnn_add av2_cnn_add_c

void av2_cnn_batchnorm_c(float **image, int channels, int width, int height, int stride, const float *gamma, const float *beta, const float *mean, const float *std);

#define av2_cnn_batchnorm av2_cnn_batchnorm_c

void av2_cnn_convolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride, int start_idx, int step);

#define av2_cnn_convolve av2_cnn_convolve_c

void av2_cnn_deconvolve_c( const float **input, int in_width, int in_height, int in_stride, const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride);

#define av2_cnn_deconvolve av2_cnn_deconvolve_c

void av2_cnn_predict_c( const float **input, int in_width, int in_height, int in_stride, const CNN_CONFIG *cnn_config, const CNN_THREAD_DATA *thread_data, CNN_MULTI_OUT *output_struct);

#define av2_cnn_predict av2_cnn_predict_c

void av2_convolve_mixedsymmetric_highbd_c(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_mixedsymmetric_highbd_avx2(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

RTCD_EXTERN void (*av2_convolve_mixedsymmetric_highbd)(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_blk8x8_highbd_c(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_blk8x8_highbd_avx2(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

RTCD_EXTERN void (*av2_convolve_symmetric_blk8x8_highbd)(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_dual_highbd_c(const uint16_t *dgd, int dgd_stride, const uint16_t *dgd_dual, int dgd_dual_stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_dual_highbd_avx2(const uint16_t *dgd, int dgd_stride, const uint16_t *dgd_dual, int dgd_dual_stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

RTCD_EXTERN void (*av2_convolve_symmetric_dual_highbd)(const uint16_t *dgd, int dgd_stride, const uint16_t *dgd_dual, int dgd_dual_stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_dual_subtract_center_highbd_c(const uint16_t *dgd, int dgd_stride, const uint16_t *dgd_dual, int dgd_dual_stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_dual_subtract_center_highbd_avx2(const uint16_t *dgd, int dgd_stride, const uint16_t *dgd_dual, int dgd_dual_stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

RTCD_EXTERN void (*av2_convolve_symmetric_dual_subtract_center_highbd)(const uint16_t *dgd, int dgd_stride, const uint16_t *dgd_dual, int dgd_dual_stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_highbd_c(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_highbd_avx2(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

RTCD_EXTERN void (*av2_convolve_symmetric_highbd)(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_subtract_center_highbd_c(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_convolve_symmetric_subtract_center_highbd_avx2(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

RTCD_EXTERN void (*av2_convolve_symmetric_subtract_center_highbd)(const uint16_t *dgd, int stride, const NonsepFilterConfig *filter_config, const int16_t *filter, uint16_t *dst, int dst_stride, int bit_depth, int block_row_begin, int block_row_end, int block_col_begin, int block_col_end);

void av2_copy_pred_array_highbd_c(const uint16_t *src1, const uint16_t *src2, int src_stride, int16_t *dst1,int16_t *dst2, int bw, int bh, int d0, int d1, int bd, int centered);

void av2_copy_pred_array_highbd_sse4_1(const uint16_t *src1, const uint16_t *src2, int src_stride, int16_t *dst1,int16_t *dst2, int bw, int bh, int d0, int d1, int bd, int centered);

void av2_copy_pred_array_highbd_avx2(const uint16_t *src1, const uint16_t *src2, int src_stride, int16_t *dst1,int16_t *dst2, int bw, int bh, int d0, int d1, int bd, int centered);

RTCD_EXTERN void (*av2_copy_pred_array_highbd)(const uint16_t *src1, const uint16_t *src2, int src_stride, int16_t *dst1,int16_t *dst2, int bw, int bh, int d0, int d1, int bd, int centered);

void av2_dip_matrix_multiplication_c(const uint16_t *A, const uint16_t *B, uint16_t *C, int bd);

void av2_dip_matrix_multiplication_avx2(const uint16_t *A, const uint16_t *B, uint16_t *C, int bd);

RTCD_EXTERN void (*av2_dip_matrix_multiplication)(const uint16_t *A, const uint16_t *B, uint16_t *C, int bd);

void av2_ext_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, int use_warp_bd_box, PadBlock *warp_bd_box);

void av2_ext_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, int use_warp_bd_box, PadBlock *warp_bd_box);

RTCD_EXTERN void (*av2_ext_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, int use_warp_bd_box, PadBlock *warp_bd_box);

void av2_fill_directional_feature_accumulators_c(int dir_feature_accum[NUM_PC_WIENER_FEATURES][PC_WIENER_FEATURE_ACC_SIZE], int *feature_sum_buff[NUM_PC_WIENER_FEATURES], int width, int col_offset, int feature_lead, int feature_lag);

void av2_fill_directional_feature_accumulators_avx2(int dir_feature_accum[NUM_PC_WIENER_FEATURES][PC_WIENER_FEATURE_ACC_SIZE], int *feature_sum_buff[NUM_PC_WIENER_FEATURES], int width, int col_offset, int feature_lead, int feature_lag);

RTCD_EXTERN void (*av2_fill_directional_feature_accumulators)(int dir_feature_accum[NUM_PC_WIENER_FEATURES][PC_WIENER_FEATURE_ACC_SIZE], int *feature_sum_buff[NUM_PC_WIENER_FEATURES], int width, int col_offset, int feature_lead, int feature_lag);

void av2_fill_tskip_feature_accumulator_c(int16_t tskip_feature_accum[PC_WIENER_FEATURE_ACC_SIZE], int8_t* tskip_sum_buff, int width, int col_offset,int tskip_lead, int tskip_lag);

void av2_fill_tskip_feature_accumulator_avx2(int16_t tskip_feature_accum[PC_WIENER_FEATURE_ACC_SIZE], int8_t* tskip_sum_buff, int width, int col_offset,int tskip_lead, int tskip_lag);

RTCD_EXTERN void (*av2_fill_tskip_feature_accumulator)(int16_t tskip_feature_accum[PC_WIENER_FEATURE_ACC_SIZE], int8_t* tskip_sum_buff, int width, int col_offset,int tskip_lead, int tskip_lag);

void av2_fill_tskip_sum_buffer_c(int row, const uint8_t *tskip, int tskip_stride, int8_t *tskip_sum_buffer, int width, int height, int tskip_lead, int tskip_lag, bool use_strict_bounds);

void av2_fill_tskip_sum_buffer_avx2(int row, const uint8_t *tskip, int tskip_stride, int8_t *tskip_sum_buffer, int width, int height, int tskip_lead, int tskip_lag, bool use_strict_bounds);

RTCD_EXTERN void (*av2_fill_tskip_sum_buffer)(int row, const uint8_t *tskip, int tskip_stride, int8_t *tskip_sum_buffer, int width, int height, int tskip_lead, int tskip_lag, bool use_strict_bounds);

void av2_filter_intra_edge_high_c(uint16_t *p, int sz, int strength);

void av2_filter_intra_edge_high_sse4_1(uint16_t *p, int sz, int strength);

RTCD_EXTERN void (*av2_filter_intra_edge_high)(uint16_t *p, int sz, int strength);

void av2_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_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);

void av2_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_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_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_qn, const int x_step_qn, const int subpel_y_qn, const int y_step_qn, ConvolveParams *conv_params, int bd);

void av2_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_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);

void av2_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);

RTCD_EXTERN void (*av2_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);

void av2_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

void av2_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

void av2_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

void av2_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_y, const int subpel_y_qn, int bd);

void av2_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_y, const int subpel_y_qn, int bd);

void av2_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_y, const int subpel_y_qn, int bd);

RTCD_EXTERN void (*av2_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_y, const int subpel_y_qn, int bd);

void av2_highbd_dist_wtd_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_highbd_dist_wtd_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_qn, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_convolve_2d_copy_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_convolve_2d_copy_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_highbd_dist_wtd_convolve_2d_copy)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_highbd_dist_wtd_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 int subpel_x_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_convolve_y_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_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_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dist_wtd_convolve_y_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_highbd_dist_wtd_convolve_y)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, ConvolveParams *conv_params, int bd);

void av2_highbd_dr_prediction_z1_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

RTCD_EXTERN void (*av2_highbd_dr_prediction_z1)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z1_idif_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z1_idif_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

RTCD_EXTERN void (*av2_highbd_dr_prediction_z1_idif)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z2_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

RTCD_EXTERN void (*av2_highbd_dr_prediction_z2)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z2_idif_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z2_idif_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

RTCD_EXTERN void (*av2_highbd_dr_prediction_z2_idif)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z3_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

RTCD_EXTERN void (*av2_highbd_dr_prediction_z3)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z3_idif_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_dr_prediction_z3_idif_avx2(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

RTCD_EXTERN void (*av2_highbd_dr_prediction_z3_idif)(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int dx, int dy, int bd, int mrl_index);

void av2_highbd_ibp_dr_prediction_z1_c(const IbpWeightsType weights[][IBP_WEIGHT_SIZE][DIR_MODES_0_90], int mode_idx, uint16_t *dst, ptrdiff_t stride, uint16_t* second_pred, ptrdiff_t second_stride, int bw, int bh);

#define av2_highbd_ibp_dr_prediction_z1 av2_highbd_ibp_dr_prediction_z1_c

void av2_highbd_ibp_dr_prediction_z3_c(const IbpWeightsType weights[][IBP_WEIGHT_SIZE][DIR_MODES_0_90], int mode_idx, uint16_t *dst, ptrdiff_t stride, uint16_t* second_pred, ptrdiff_t second_stride, int bw, int bh);

#define av2_highbd_ibp_dr_prediction_z3 av2_highbd_ibp_dr_prediction_z3_c

void av2_highbd_inv_txfm_add_c(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

void av2_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

void av2_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

RTCD_EXTERN void (*av2_highbd_inv_txfm_add)(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

void av2_highbd_inv_txfm_add_horz_c(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

#define av2_highbd_inv_txfm_add_horz av2_highbd_inv_txfm_add_horz_c

void av2_highbd_inv_txfm_add_vert_c(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

#define av2_highbd_inv_txfm_add_vert av2_highbd_inv_txfm_add_vert_c

void av2_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint16_t *dest, int dest_stride, int bd);

#define av2_highbd_iwht4x4_16_add av2_highbd_iwht4x4_16_add_c

void av2_highbd_iwht4x4_16_horz_add_c(const tran_low_t *input, uint16_t *dest, int dest_stride, int bd);

#define av2_highbd_iwht4x4_16_horz_add av2_highbd_iwht4x4_16_horz_add_c

void av2_highbd_iwht4x4_16_vert_add_c(const tran_low_t *input, uint16_t *dest, int dest_stride, int bd);

#define av2_highbd_iwht4x4_16_vert_add av2_highbd_iwht4x4_16_vert_add_c

void av2_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint16_t *dest, int dest_stride, int bd);

#define av2_highbd_iwht4x4_1_add av2_highbd_iwht4x4_1_add_c

void av2_highbd_iwht4x4_1_horz_add_c(const tran_low_t *input, uint16_t *dest, int dest_stride, int bd);

#define av2_highbd_iwht4x4_1_horz_add av2_highbd_iwht4x4_1_horz_add_c

void av2_highbd_iwht4x4_1_vert_add_c(const tran_low_t *input, uint16_t *dest, int dest_stride, int bd);

#define av2_highbd_iwht4x4_1_vert_add av2_highbd_iwht4x4_1_vert_add_c

void av2_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);

void av2_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);

void av2_highbd_warp_affine_avx2(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);

RTCD_EXTERN void (*av2_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);

void av2_highbd_wiener_convolve_add_src_c(const uint16_t *src, ptrdiff_t src_stride, uint16_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 WienerConvolveParams *conv_params, int bd);

void av2_highbd_wiener_convolve_add_src_ssse3(const uint16_t *src, ptrdiff_t src_stride, uint16_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 WienerConvolveParams *conv_params, int bd);

void av2_highbd_wiener_convolve_add_src_avx2(const uint16_t *src, ptrdiff_t src_stride, uint16_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 WienerConvolveParams *conv_params, int bd);

RTCD_EXTERN void (*av2_highbd_wiener_convolve_add_src)(const uint16_t *src, ptrdiff_t src_stride, uint16_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 WienerConvolveParams *conv_params, int bd);

void av2_inv_idfm2d_add_4x4_horz_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);

#define av2_inv_idfm2d_add_4x4_horz av2_inv_idfm2d_add_4x4_horz_c

void av2_inv_idfm2d_add_4x4_vert_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);

#define av2_inv_idfm2d_add_4x4_vert av2_inv_idfm2d_add_4x4_vert_c

void av2_lossless_inv_idtx_add_c(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

void av2_lossless_inv_idtx_add_avx2(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

RTCD_EXTERN void (*av2_lossless_inv_idtx_add)(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

void av2_lossless_inv_idtx_add_horz_c(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

#define av2_lossless_inv_idtx_add_horz av2_lossless_inv_idtx_add_horz_c

void av2_lossless_inv_idtx_add_vert_c(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

#define av2_lossless_inv_idtx_add_vert av2_lossless_inv_idtx_add_vert_c

void av2_make_bawp_block_c(uint16_t *dst, int dst_stride, int16_t alpha, int32_t beta, int shift, int bw, int bh, int bd);

void av2_make_bawp_block_avx2(uint16_t *dst, int dst_stride, int16_t alpha, int32_t beta, int shift, int bw, int bh, int bd);

RTCD_EXTERN void (*av2_make_bawp_block)(uint16_t *dst, int dst_stride, int16_t alpha, int32_t beta, int shift, int bw, int bh, int bd);

void av2_mhccp_derive_multi_param_hv_c(MACROBLOCKD *const xd, int plane,int above_lines, int left_lines, int ref_width,int ref_height, int dir, int is_top_sb_boundary);

void av2_mhccp_derive_multi_param_hv_avx2(MACROBLOCKD *const xd, int plane,int above_lines, int left_lines, int ref_width,int ref_height, int dir, int is_top_sb_boundary);

RTCD_EXTERN void (*av2_mhccp_derive_multi_param_hv)(MACROBLOCKD *const xd, int plane,int above_lines, int left_lines, int ref_width,int ref_height, int dir, int is_top_sb_boundary);

void av2_mhccp_implicit_fetch_neighbor_chroma_c(const uint16_t *dst, int input_stride, TX_SIZE tx_size, int above_lines, int left_lines, int is_top_sb_boundary, int ref_width, int ref_height, uint16_t *output_q3);

void av2_mhccp_implicit_fetch_neighbor_chroma_avx2(const uint16_t *dst, int input_stride, TX_SIZE tx_size, int above_lines, int left_lines, int is_top_sb_boundary, int ref_width, int ref_height, uint16_t *output_q3);

RTCD_EXTERN void (*av2_mhccp_implicit_fetch_neighbor_chroma)(const uint16_t *dst, int input_stride, TX_SIZE tx_size, int above_lines, int left_lines, int is_top_sb_boundary, int ref_width, int ref_height, uint16_t *output_q3);

void av2_mhccp_implicit_fetch_neighbor_luma_420_c(const uint16_t *input, int input_stride, int above_lines, int left_lines, int is_top_sb_boundary, int ref_width, int ref_height, int sub_y, uint8_t cfl_ds_filter_index, int width, int height, uint16_t *output_q3, int output_stride);

void av2_mhccp_implicit_fetch_neighbor_luma_420_avx2(const uint16_t *input, int input_stride, int above_lines, int left_lines, int is_top_sb_boundary, int ref_width, int ref_height, int sub_y, uint8_t cfl_ds_filter_index, int width, int height, uint16_t *output_q3, int output_stride);

RTCD_EXTERN void (*av2_mhccp_implicit_fetch_neighbor_luma_420)(const uint16_t *input, int input_stride, int above_lines, int left_lines, int is_top_sb_boundary, int ref_width, int ref_height, int sub_y, uint8_t cfl_ds_filter_index, int width, int height, uint16_t *output_q3, int output_stride);

void av2_opfl_mv_refinement_8x8_c(const int16_t *pdiff, int pstride, const int16_t *gx, const int16_t *gy, int gstride, int d0, int d1, int grad_prec_bits, int mv_prec_bits, int *vx0, int *vy0, int *vx1, int *vy1);

void av2_opfl_mv_refinement_8x8_sse4_1(const int16_t *pdiff, int pstride, const int16_t *gx, const int16_t *gy, int gstride, int d0, int d1, int grad_prec_bits, int mv_prec_bits, int *vx0, int *vy0, int *vx1, int *vy1);

RTCD_EXTERN void (*av2_opfl_mv_refinement_8x8)(const int16_t *pdiff, int pstride, const int16_t *gx, const int16_t *gy, int gstride, int d0, int d1, int grad_prec_bits, int mv_prec_bits, int *vx0, int *vy0, int *vx1, int *vy1);

int av2_opfl_mv_refinement_nxn_c( const int16_t *pdiff, int pstride, const int16_t *gx, const int16_t *gy, int gstride, int bw, int bh, int n, int d0, int d1, int grad_prec_bits, int mv_prec_bits, int mi_x, int mi_y, int mi_cols, int mi_rows, int is_decode, int *vx0, int *vy0, int *vx1, int *vy1);

int av2_opfl_mv_refinement_nxn_sse4_1( const int16_t *pdiff, int pstride, const int16_t *gx, const int16_t *gy, int gstride, int bw, int bh, int n, int d0, int d1, int grad_prec_bits, int mv_prec_bits, int mi_x, int mi_y, int mi_cols, int mi_rows, int is_decode, int *vx0, int *vy0, int *vx1, int *vy1);

int av2_opfl_mv_refinement_nxn_avx2( const int16_t *pdiff, int pstride, const int16_t *gx, const int16_t *gy, int gstride, int bw, int bh, int n, int d0, int d1, int grad_prec_bits, int mv_prec_bits, int mi_x, int mi_y, int mi_cols, int mi_rows, int is_decode, int *vx0, int *vy0, int *vx1, int *vy1);

RTCD_EXTERN int (*av2_opfl_mv_refinement_nxn)( const int16_t *pdiff, int pstride, const int16_t *gx, const int16_t *gy, int gstride, int bw, int bh, int n, int d0, int d1, int grad_prec_bits, int mv_prec_bits, int mi_x, int mi_y, int mi_cols, int mi_rows, int is_decode, int *vx0, int *vy0, int *vx1, int *vy1);

void av2_resize_and_extend_frame_c(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, const InterpFilter filter, const int phase, const int num_planes);

#define av2_resize_and_extend_frame av2_resize_and_extend_frame_c

void calc_wienerns_ds_luma_420_c(const uint16_t *src, int src_stride, uint16_t *const dst, int dst_stride, int ds_type, int height_uv, int width_uv, int ss_x, int ss_y, int col_start);

void calc_wienerns_ds_luma_420_avx2(const uint16_t *src, int src_stride, uint16_t *const dst, int dst_stride, int ds_type, int height_uv, int width_uv, int ss_x, int ss_y, int col_start);

RTCD_EXTERN void (*calc_wienerns_ds_luma_420)(const uint16_t *src, int src_stride, uint16_t *const dst, int dst_stride, int ds_type, int height_uv, int width_uv, int ss_x, int ss_y, int col_start);

void ccso_filter_block_hbd_wo_buf_c(const uint16_t *src_y, uint16_t *dst_yuv, const int x, const int y, const int pic_width, const int pic_height, int *src_cls, const int8_t *offset_buf, const int scaled_ext_stride, const int dst_stride, const int y_uv_hscale, const int y_uv_vscale, const int thr, const int neg_thr, const int *src_loc, const int max_val, const int blk_size_x, const int blk_size_y, const bool isSingleBand, const uint8_t shift_bits, const int edge_clf, const uint8_t ccso_bo_only);

void ccso_filter_block_hbd_wo_buf_avx2(const uint16_t *src_y, uint16_t *dst_yuv, const int x, const int y, const int pic_width, const int pic_height, int *src_cls, const int8_t *offset_buf, const int scaled_ext_stride, const int dst_stride, const int y_uv_hscale, const int y_uv_vscale, const int thr, const int neg_thr, const int *src_loc, const int max_val, const int blk_size_x, const int blk_size_y, const bool isSingleBand, const uint8_t shift_bits, const int edge_clf, const uint8_t ccso_bo_only);

RTCD_EXTERN void (*ccso_filter_block_hbd_wo_buf)(const uint16_t *src_y, uint16_t *dst_yuv, const int x, const int y, const int pic_width, const int pic_height, int *src_cls, const int8_t *offset_buf, const int scaled_ext_stride, const int dst_stride, const int y_uv_hscale, const int y_uv_vscale, const int thr, const int neg_thr, const int *src_loc, const int max_val, const int blk_size_x, const int blk_size_y, const bool isSingleBand, const uint8_t shift_bits, const int edge_clf, const uint8_t ccso_bo_only);

void ccso_filter_block_hbd_wo_buf_bo_only_c(const uint16_t *src_y, uint16_t *dts_yuv, const int x, const int y, const int pic_width, const int pic_height, const int8_t *offset_buf, const int src_y_stride, const int dst_stride, const int y_uv_hscale, const int y_uv_vscale, const int max_val, const int blk_size_x, const int blk_size_y, const bool isSingleBand, const uint8_t shift_bits);

void ccso_filter_block_hbd_wo_buf_bo_only_avx2(const uint16_t *src_y, uint16_t *dts_yuv, const int x, const int y, const int pic_width, const int pic_height, const int8_t *offset_buf, const int src_y_stride, const int dst_stride, const int y_uv_hscale, const int y_uv_vscale, const int max_val, const int blk_size_x, const int blk_size_y, const bool isSingleBand, const uint8_t shift_bits);

RTCD_EXTERN void (*ccso_filter_block_hbd_wo_buf_bo_only)(const uint16_t *src_y, uint16_t *dts_yuv, const int x, const int y, const int pic_width, const int pic_height, const int8_t *offset_buf, const int src_y_stride, const int dst_stride, const int y_uv_hscale, const int y_uv_vscale, const int max_val, const int blk_size_x, const int blk_size_y, const bool isSingleBand, const uint8_t shift_bits);

void fill_directional_feature_buffers_highbd_c(int *feature_sum_buffers[], int16_t *feature_line_buffers[], int row, int buffer_row, const uint16_t *dgd, int dgd_stride, int width, int feature_lead, int feature_lag);

void fill_directional_feature_buffers_highbd_avx2(int *feature_sum_buffers[], int16_t *feature_line_buffers[], int row, int buffer_row, const uint16_t *dgd, int dgd_stride, int width, int feature_lead, int feature_lag);

RTCD_EXTERN void (*fill_directional_feature_buffers_highbd)(int *feature_sum_buffers[], int16_t *feature_line_buffers[], int row, int buffer_row, const uint16_t *dgd, int dgd_stride, int width, int feature_lead, int feature_lag);

void gdf_compensation_unit_c(uint16_t* rec_pnt, const int rec_stride, int16_t* err_pnt, const int err_stride, const int err_shift, const int scale, const int pxl_max, const int blk_height, const int blk_width);

void gdf_compensation_unit_avx2(uint16_t* rec_pnt, const int rec_stride, int16_t* err_pnt, const int err_stride, const int err_shift, const int scale, const int pxl_max, const int blk_height, const int blk_width);

RTCD_EXTERN void (*gdf_compensation_unit)(uint16_t* rec_pnt, const int rec_stride, int16_t* err_pnt, const int err_stride, const int err_shift, const int scale, const int pxl_max, const int blk_height, const int blk_width);

void gdf_inference_unit_c(const int i_min, const int i_max, const int j_min, const int j_max, const int qp_idx, const uint16_t* rec_pnt, const int rec_stride, uint16_t *const *gdf_lap_pnt, const int gdf_lap_stride, const uint32_t *gdf_cls_pnt, const int gdf_cls_stride, int16_t* err_pnt, const int err_stride, const int pxl_shift, const int ref_dst_idx);

void gdf_inference_unit_avx2(const int i_min, const int i_max, const int j_min, const int j_max, const int qp_idx, const uint16_t* rec_pnt, const int rec_stride, uint16_t *const *gdf_lap_pnt, const int gdf_lap_stride, const uint32_t *gdf_cls_pnt, const int gdf_cls_stride, int16_t* err_pnt, const int err_stride, const int pxl_shift, const int ref_dst_idx);

RTCD_EXTERN void (*gdf_inference_unit)(const int i_min, const int i_max, const int j_min, const int j_max, const int qp_idx, const uint16_t* rec_pnt, const int rec_stride, uint16_t *const *gdf_lap_pnt, const int gdf_lap_stride, const uint32_t *gdf_cls_pnt, const int gdf_cls_stride, int16_t* err_pnt, const int err_stride, const int pxl_shift, const int ref_dst_idx);

void gdf_set_lap_and_cls_unit_c(const int i_min, const int i_max, const int j_min, const int j_max, const uint16_t *rec_pnt, const int rec_stride, const int bit_depth, uint16_t *const *gdf_lap_y, const int gdf_lap_y_stride, uint32_t *gdf_cls_y, const int gdf_cls_y_stride);

void gdf_set_lap_and_cls_unit_avx2(const int i_min, const int i_max, const int j_min, const int j_max, const uint16_t *rec_pnt, const int rec_stride, const int bit_depth, uint16_t *const *gdf_lap_y, const int gdf_lap_y_stride, uint32_t *gdf_cls_y, const int gdf_cls_y_stride);

RTCD_EXTERN void (*gdf_set_lap_and_cls_unit)(const int i_min, const int i_max, const int j_min, const int j_max, const uint16_t *rec_pnt, const int rec_stride, const int bit_depth, uint16_t *const *gdf_lap_y, const int gdf_lap_y_stride, uint32_t *gdf_cls_y, const int gdf_cls_y_stride);

void inv_stxfm_c(tran_low_t *src, tran_low_t *dst, const PREDICTION_MODE mode, const uint8_t stx_idx, const int size, const int bd);

void inv_stxfm_sse4_1(tran_low_t *src, tran_low_t *dst, const PREDICTION_MODE mode, const uint8_t stx_idx, const int size, const int bd);

void inv_stxfm_avx2(tran_low_t *src, tran_low_t *dst, const PREDICTION_MODE mode, const uint8_t stx_idx, const int size, const int bd);

RTCD_EXTERN void (*inv_stxfm)(tran_low_t *src, tran_low_t *dst, const PREDICTION_MODE mode, const uint8_t stx_idx, const int size, const int bd);

void inv_txfm_c(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

void inv_txfm_avx2(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

RTCD_EXTERN void (*inv_txfm)(const tran_low_t *input, uint16_t *dest, int stride, const TxfmParam *txfm_param);

void mhccp_predict_hv_hbd_c(const uint16_t *input, uint16_t *dst, bool have_top, bool have_left, int dst_stride, int *alpha_q3, int bit_depth, int width, int height, int dir);

void mhccp_predict_hv_hbd_avx2(const uint16_t *input, uint16_t *dst, bool have_top, bool have_left, int dst_stride, int *alpha_q3, int bit_depth, int width, int height, int dir);

RTCD_EXTERN void (*mhccp_predict_hv_hbd)(const uint16_t *input, uint16_t *dst, bool have_top, bool have_left, int dst_stride, int *alpha_q3, int bit_depth, int width, int height, int dir);

void refinemv_highbd_pad_mc_border_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int x0, int y0, int b_w, int b_h, const ReferenceArea *ref_area);

void refinemv_highbd_pad_mc_border_avx2(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int x0, int y0, int b_w, int b_h, const ReferenceArea *ref_area);

RTCD_EXTERN void (*refinemv_highbd_pad_mc_border)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int x0, int y0, int b_w, int b_h, const ReferenceArea *ref_area);

void resample_output_c(uint16_t *dst, int dst_stride, const uint16_t *above_row, const uint16_t *left_col, uint16_t *ml_output, int bw_log2, int bh_log2, int transpose);

void resample_output_avx2(uint16_t *dst, int dst_stride, const uint16_t *above_row, const uint16_t *left_col, uint16_t *ml_output, int bw_log2, int bh_log2, int transpose);

RTCD_EXTERN void (*resample_output)(uint16_t *dst, int dst_stride, const uint16_t *above_row, const uint16_t *left_col, uint16_t *ml_output, int bw_log2, int bh_log2, int transpose);

void av2_rtcd(void);

#ifdef RTCD_C

#include "avm_ports/x86.h"

static void setup_rtcd_internal(void)

{

    int flags = x86_simd_caps();

    (void)flags;

    av2_avg_pooling_pdiff_gradients = av2_avg_pooling_pdiff_gradients_c;

    if (flags & HAS_AVX2) av2_avg_pooling_pdiff_gradients = av2_avg_pooling_pdiff_gradients_avx2;

    av2_bicubic_grad_interpolation_highbd = av2_bicubic_grad_interpolation_highbd_c;

    if (flags & HAS_SSE4_1) av2_bicubic_grad_interpolation_highbd = av2_bicubic_grad_interpolation_highbd_sse4_1;

    if (flags & HAS_AVX2) av2_bicubic_grad_interpolation_highbd = av2_bicubic_grad_interpolation_highbd_avx2;

    av2_build_compound_diffwtd_mask_d16 = av2_build_compound_diffwtd_mask_d16_c;

    if (flags & HAS_SSE4_1) av2_build_compound_diffwtd_mask_d16 = av2_build_compound_diffwtd_mask_d16_sse4_1;

    if (flags & HAS_AVX2) av2_build_compound_diffwtd_mask_d16 = av2_build_compound_diffwtd_mask_d16_avx2;

    av2_build_compound_diffwtd_mask_highbd = av2_build_compound_diffwtd_mask_highbd_c;

    if (flags & HAS_SSSE3) av2_build_compound_diffwtd_mask_highbd = av2_build_compound_diffwtd_mask_highbd_ssse3;

    if (flags & HAS_AVX2) av2_build_compound_diffwtd_mask_highbd = av2_build_compound_diffwtd_mask_highbd_avx2;

    av2_cdef_copy_rect8_16bit_to_16bit = av2_cdef_copy_rect8_16bit_to_16bit_sse2;

    if (flags & HAS_SSSE3) av2_cdef_copy_rect8_16bit_to_16bit = av2_cdef_copy_rect8_16bit_to_16bit_ssse3;

    if (flags & HAS_SSE4_1) av2_cdef_copy_rect8_16bit_to_16bit = av2_cdef_copy_rect8_16bit_to_16bit_sse4_1;

    if (flags & HAS_AVX2) av2_cdef_copy_rect8_16bit_to_16bit = av2_cdef_copy_rect8_16bit_to_16bit_avx2;

    av2_cdef_filter_16_0 = av2_cdef_filter_16_0_sse2;

    if (flags & HAS_SSSE3) av2_cdef_filter_16_0 = av2_cdef_filter_16_0_ssse3;

    if (flags & HAS_SSE4_1) av2_cdef_filter_16_0 = av2_cdef_filter_16_0_sse4_1;

    if (flags & HAS_AVX2) av2_cdef_filter_16_0 = av2_cdef_filter_16_0_avx2;

    av2_cdef_filter_16_1 = av2_cdef_filter_16_1_sse2;

    if (flags & HAS_SSSE3) av2_cdef_filter_16_1 = av2_cdef_filter_16_1_ssse3;

    if (flags & HAS_SSE4_1) av2_cdef_filter_16_1 = av2_cdef_filter_16_1_sse4_1;

    if (flags & HAS_AVX2) av2_cdef_filter_16_1 = av2_cdef_filter_16_1_avx2;

    av2_cdef_filter_16_2 = av2_cdef_filter_16_2_sse2;

    if (flags & HAS_SSSE3) av2_cdef_filter_16_2 = av2_cdef_filter_16_2_ssse3;

    if (flags & HAS_SSE4_1) av2_cdef_filter_16_2 = av2_cdef_filter_16_2_sse4_1;

    if (flags & HAS_AVX2) av2_cdef_filter_16_2 = av2_cdef_filter_16_2_avx2;

    av2_cdef_filter_16_3 = av2_cdef_filter_16_3_sse2;

    if (flags & HAS_SSSE3) av2_cdef_filter_16_3 = av2_cdef_filter_16_3_ssse3;

    if (flags & HAS_SSE4_1) av2_cdef_filter_16_3 = av2_cdef_filter_16_3_sse4_1;

    if (flags & HAS_AVX2) av2_cdef_filter_16_3 = av2_cdef_filter_16_3_avx2;

    av2_cdef_find_dir = av2_cdef_find_dir_sse2;

    if (flags & HAS_SSSE3) av2_cdef_find_dir = av2_cdef_find_dir_ssse3;

    if (flags & HAS_SSE4_1) av2_cdef_find_dir = av2_cdef_find_dir_sse4_1;

    if (flags & HAS_AVX2) av2_cdef_find_dir = av2_cdef_find_dir_avx2;

    av2_cdef_find_dir_dual = av2_cdef_find_dir_dual_sse2;

    if (flags & HAS_SSSE3) av2_cdef_find_dir_dual = av2_cdef_find_dir_dual_ssse3;

    if (flags & HAS_SSE4_1) av2_cdef_find_dir_dual = av2_cdef_find_dir_dual_sse4_1;

    if (flags & HAS_AVX2) av2_cdef_find_dir_dual = av2_cdef_find_dir_dual_avx2;

    av2_cfl_get_luma_subsampling_420_hbd = av2_cfl_get_luma_subsampling_420_hbd_c;

    if (flags & HAS_SSSE3) av2_cfl_get_luma_subsampling_420_hbd = av2_cfl_get_luma_subsampling_420_hbd_ssse3;

    if (flags & HAS_AVX2) av2_cfl_get_luma_subsampling_420_hbd = av2_cfl_get_luma_subsampling_420_hbd_avx2;

    av2_cfl_get_luma_subsampling_420_hbd_121 = av2_cfl_get_luma_subsampling_420_hbd_121_c;

    if (flags & HAS_AVX2) av2_cfl_get_luma_subsampling_420_hbd_121 = av2_cfl_get_luma_subsampling_420_hbd_121_avx2;

    av2_cfl_get_luma_subsampling_420_hbd_colocated = av2_cfl_get_luma_subsampling_420_hbd_colocated_c;

    if (flags & HAS_AVX2) av2_cfl_get_luma_subsampling_420_hbd_colocated = av2_cfl_get_luma_subsampling_420_hbd_colocated_avx2;

    av2_cfl_get_luma_subsampling_422_hbd = av2_cfl_get_luma_subsampling_422_hbd_c;

    if (flags & HAS_SSSE3) av2_cfl_get_luma_subsampling_422_hbd = av2_cfl_get_luma_subsampling_422_hbd_ssse3;

    if (flags & HAS_AVX2) av2_cfl_get_luma_subsampling_422_hbd = av2_cfl_get_luma_subsampling_422_hbd_avx2;

    av2_cfl_get_luma_subsampling_444_hbd = av2_cfl_get_luma_subsampling_444_hbd_c;

    if (flags & HAS_SSSE3) av2_cfl_get_luma_subsampling_444_hbd = av2_cfl_get_luma_subsampling_444_hbd_ssse3;

    if (flags & HAS_AVX2) av2_cfl_get_luma_subsampling_444_hbd = av2_cfl_get_luma_subsampling_444_hbd_avx2;

    av2_cfl_get_predict_hbd_fn = av2_cfl_get_predict_hbd_fn_c;

    if (flags & HAS_AVX2) av2_cfl_get_predict_hbd_fn = av2_cfl_get_predict_hbd_fn_avx2;

    av2_cfl_get_subtract_average_fn = av2_cfl_get_subtract_average_fn_sse2;

    if (flags & HAS_AVX2) av2_cfl_get_subtract_average_fn = av2_cfl_get_subtract_average_fn_avx2;

    av2_convolve_mixedsymmetric_highbd = av2_convolve_mixedsymmetric_highbd_c;

    if (flags & HAS_AVX2) av2_convolve_mixedsymmetric_highbd = av2_convolve_mixedsymmetric_highbd_avx2;

    av2_convolve_symmetric_blk8x8_highbd = av2_convolve_symmetric_blk8x8_highbd_c;

    if (flags & HAS_AVX2) av2_convolve_symmetric_blk8x8_highbd = av2_convolve_symmetric_blk8x8_highbd_avx2;

    av2_convolve_symmetric_dual_highbd = av2_convolve_symmetric_dual_highbd_c;

    if (flags & HAS_AVX2) av2_convolve_symmetric_dual_highbd = av2_convolve_symmetric_dual_highbd_avx2;

    av2_convolve_symmetric_dual_subtract_center_highbd = av2_convolve_symmetric_dual_subtract_center_highbd_c;

    if (flags & HAS_AVX2) av2_convolve_symmetric_dual_subtract_center_highbd = av2_convolve_symmetric_dual_subtract_center_highbd_avx2;

    av2_convolve_symmetric_highbd = av2_convolve_symmetric_highbd_c;

    if (flags & HAS_AVX2) av2_convolve_symmetric_highbd = av2_convolve_symmetric_highbd_avx2;

    av2_convolve_symmetric_subtract_center_highbd = av2_convolve_symmetric_subtract_center_highbd_c;

    if (flags & HAS_AVX2) av2_convolve_symmetric_subtract_center_highbd = av2_convolve_symmetric_subtract_center_highbd_avx2;

    av2_copy_pred_array_highbd = av2_copy_pred_array_highbd_c;

    if (flags & HAS_SSE4_1) av2_copy_pred_array_highbd = av2_copy_pred_array_highbd_sse4_1;

    if (flags & HAS_AVX2) av2_copy_pred_array_highbd = av2_copy_pred_array_highbd_avx2;

    av2_dip_matrix_multiplication = av2_dip_matrix_multiplication_c;

    if (flags & HAS_AVX2) av2_dip_matrix_multiplication = av2_dip_matrix_multiplication_avx2;

    av2_ext_highbd_warp_affine = av2_ext_highbd_warp_affine_c;

    if (flags & HAS_SSE4_1) av2_ext_highbd_warp_affine = av2_ext_highbd_warp_affine_sse4_1;

    av2_fill_directional_feature_accumulators = av2_fill_directional_feature_accumulators_c;

    if (flags & HAS_AVX2) av2_fill_directional_feature_accumulators = av2_fill_directional_feature_accumulators_avx2;

    av2_fill_tskip_feature_accumulator = av2_fill_tskip_feature_accumulator_c;

    if (flags & HAS_AVX2) av2_fill_tskip_feature_accumulator = av2_fill_tskip_feature_accumulator_avx2;

    av2_fill_tskip_sum_buffer = av2_fill_tskip_sum_buffer_c;

    if (flags & HAS_AVX2) av2_fill_tskip_sum_buffer = av2_fill_tskip_sum_buffer_avx2;

    av2_filter_intra_edge_high = av2_filter_intra_edge_high_c;

    if (flags & HAS_SSE4_1) av2_filter_intra_edge_high = av2_filter_intra_edge_high_sse4_1;

    av2_highbd_convolve_2d_scale = av2_highbd_convolve_2d_scale_c;

    if (flags & HAS_SSE4_1) av2_highbd_convolve_2d_scale = av2_highbd_convolve_2d_scale_sse4_1;

    av2_highbd_convolve_2d_sr = av2_highbd_convolve_2d_sr_c;

    if (flags & HAS_SSSE3) av2_highbd_convolve_2d_sr = av2_highbd_convolve_2d_sr_ssse3;

    if (flags & HAS_AVX2) av2_highbd_convolve_2d_sr = av2_highbd_convolve_2d_sr_avx2;

    av2_highbd_convolve_horiz_rs = av2_highbd_convolve_horiz_rs_c;

    if (flags & HAS_SSE4_1) av2_highbd_convolve_horiz_rs = av2_highbd_convolve_horiz_rs_sse4_1;

    av2_highbd_convolve_x_sr = av2_highbd_convolve_x_sr_c;

    if (flags & HAS_SSSE3) av2_highbd_convolve_x_sr = av2_highbd_convolve_x_sr_ssse3;

    if (flags & HAS_AVX2) av2_highbd_convolve_x_sr = av2_highbd_convolve_x_sr_avx2;

    av2_highbd_convolve_y_sr = av2_highbd_convolve_y_sr_c;

    if (flags & HAS_SSSE3) av2_highbd_convolve_y_sr = av2_highbd_convolve_y_sr_ssse3;

    if (flags & HAS_AVX2) av2_highbd_convolve_y_sr = av2_highbd_convolve_y_sr_avx2;

    av2_highbd_dist_wtd_convolve_2d = av2_highbd_dist_wtd_convolve_2d_c;

    if (flags & HAS_SSE4_1) av2_highbd_dist_wtd_convolve_2d = av2_highbd_dist_wtd_convolve_2d_sse4_1;

    if (flags & HAS_AVX2) av2_highbd_dist_wtd_convolve_2d = av2_highbd_dist_wtd_convolve_2d_avx2;

    av2_highbd_dist_wtd_convolve_2d_copy = av2_highbd_dist_wtd_convolve_2d_copy_c;

    if (flags & HAS_SSE4_1) av2_highbd_dist_wtd_convolve_2d_copy = av2_highbd_dist_wtd_convolve_2d_copy_sse4_1;

    if (flags & HAS_AVX2) av2_highbd_dist_wtd_convolve_2d_copy = av2_highbd_dist_wtd_convolve_2d_copy_avx2;

    av2_highbd_dist_wtd_convolve_x = av2_highbd_dist_wtd_convolve_x_c;

    if (flags & HAS_SSE4_1) av2_highbd_dist_wtd_convolve_x = av2_highbd_dist_wtd_convolve_x_sse4_1;

    if (flags & HAS_AVX2) av2_highbd_dist_wtd_convolve_x = av2_highbd_dist_wtd_convolve_x_avx2;

    av2_highbd_dist_wtd_convolve_y = av2_highbd_dist_wtd_convolve_y_c;

    if (flags & HAS_SSE4_1) av2_highbd_dist_wtd_convolve_y = av2_highbd_dist_wtd_convolve_y_sse4_1;

    if (flags & HAS_AVX2) av2_highbd_dist_wtd_convolve_y = av2_highbd_dist_wtd_convolve_y_avx2;

    av2_highbd_dr_prediction_z1 = av2_highbd_dr_prediction_z1_c;

    if (flags & HAS_AVX2) av2_highbd_dr_prediction_z1 = av2_highbd_dr_prediction_z1_avx2;

    av2_highbd_dr_prediction_z1_idif = av2_highbd_dr_prediction_z1_idif_c;

    if (flags & HAS_AVX2) av2_highbd_dr_prediction_z1_idif = av2_highbd_dr_prediction_z1_idif_avx2;

    av2_highbd_dr_prediction_z2 = av2_highbd_dr_prediction_z2_c;

    if (flags & HAS_AVX2) av2_highbd_dr_prediction_z2 = av2_highbd_dr_prediction_z2_avx2;

    av2_highbd_dr_prediction_z2_idif = av2_highbd_dr_prediction_z2_idif_c;

    if (flags & HAS_AVX2) av2_highbd_dr_prediction_z2_idif = av2_highbd_dr_prediction_z2_idif_avx2;

    av2_highbd_dr_prediction_z3 = av2_highbd_dr_prediction_z3_c;

    if (flags & HAS_AVX2) av2_highbd_dr_prediction_z3 = av2_highbd_dr_prediction_z3_avx2;

    av2_highbd_dr_prediction_z3_idif = av2_highbd_dr_prediction_z3_idif_c;

    if (flags & HAS_AVX2) av2_highbd_dr_prediction_z3_idif = av2_highbd_dr_prediction_z3_idif_avx2;

    av2_highbd_inv_txfm_add = av2_highbd_inv_txfm_add_c;

    if (flags & HAS_SSE4_1) av2_highbd_inv_txfm_add = av2_highbd_inv_txfm_add_sse4_1;

    if (flags & HAS_AVX2) av2_highbd_inv_txfm_add = av2_highbd_inv_txfm_add_avx2;

    av2_highbd_warp_affine = av2_highbd_warp_affine_c;

    if (flags & HAS_SSE4_1) av2_highbd_warp_affine = av2_highbd_warp_affine_sse4_1;

    if (flags & HAS_AVX2) av2_highbd_warp_affine = av2_highbd_warp_affine_avx2;

    av2_highbd_wiener_convolve_add_src = av2_highbd_wiener_convolve_add_src_c;

    if (flags & HAS_SSSE3) av2_highbd_wiener_convolve_add_src = av2_highbd_wiener_convolve_add_src_ssse3;

    if (flags & HAS_AVX2) av2_highbd_wiener_convolve_add_src = av2_highbd_wiener_convolve_add_src_avx2;

    av2_lossless_inv_idtx_add = av2_lossless_inv_idtx_add_c;

    if (flags & HAS_AVX2) av2_lossless_inv_idtx_add = av2_lossless_inv_idtx_add_avx2;

    av2_make_bawp_block = av2_make_bawp_block_c;

    if (flags & HAS_AVX2) av2_make_bawp_block = av2_make_bawp_block_avx2;

    av2_mhccp_derive_multi_param_hv = av2_mhccp_derive_multi_param_hv_c;

    if (flags & HAS_AVX2) av2_mhccp_derive_multi_param_hv = av2_mhccp_derive_multi_param_hv_avx2;

    av2_mhccp_implicit_fetch_neighbor_chroma = av2_mhccp_implicit_fetch_neighbor_chroma_c;

    if (flags & HAS_AVX2) av2_mhccp_implicit_fetch_neighbor_chroma = av2_mhccp_implicit_fetch_neighbor_chroma_avx2;

    av2_mhccp_implicit_fetch_neighbor_luma_420 = av2_mhccp_implicit_fetch_neighbor_luma_420_c;

    if (flags & HAS_AVX2) av2_mhccp_implicit_fetch_neighbor_luma_420 = av2_mhccp_implicit_fetch_neighbor_luma_420_avx2;

    av2_opfl_mv_refinement_8x8 = av2_opfl_mv_refinement_8x8_c;

    if (flags & HAS_SSE4_1) av2_opfl_mv_refinement_8x8 = av2_opfl_mv_refinement_8x8_sse4_1;

    av2_opfl_mv_refinement_nxn = av2_opfl_mv_refinement_nxn_c;

    if (flags & HAS_SSE4_1) av2_opfl_mv_refinement_nxn = av2_opfl_mv_refinement_nxn_sse4_1;

    if (flags & HAS_AVX2) av2_opfl_mv_refinement_nxn = av2_opfl_mv_refinement_nxn_avx2;

    calc_wienerns_ds_luma_420 = calc_wienerns_ds_luma_420_c;

    if (flags & HAS_AVX2) calc_wienerns_ds_luma_420 = calc_wienerns_ds_luma_420_avx2;

    ccso_filter_block_hbd_wo_buf = ccso_filter_block_hbd_wo_buf_c;

    if (flags & HAS_AVX2) ccso_filter_block_hbd_wo_buf = ccso_filter_block_hbd_wo_buf_avx2;

    ccso_filter_block_hbd_wo_buf_bo_only = ccso_filter_block_hbd_wo_buf_bo_only_c;

    if (flags & HAS_AVX2) ccso_filter_block_hbd_wo_buf_bo_only = ccso_filter_block_hbd_wo_buf_bo_only_avx2;

    fill_directional_feature_buffers_highbd = fill_directional_feature_buffers_highbd_c;

    if (flags & HAS_AVX2) fill_directional_feature_buffers_highbd = fill_directional_feature_buffers_highbd_avx2;

    gdf_compensation_unit = gdf_compensation_unit_c;

    if (flags & HAS_AVX2) gdf_compensation_unit = gdf_compensation_unit_avx2;

    gdf_inference_unit = gdf_inference_unit_c;

    if (flags & HAS_AVX2) gdf_inference_unit = gdf_inference_unit_avx2;

    gdf_set_lap_and_cls_unit = gdf_set_lap_and_cls_unit_c;

    if (flags & HAS_AVX2) gdf_set_lap_and_cls_unit = gdf_set_lap_and_cls_unit_avx2;

    inv_stxfm = inv_stxfm_c;

    if (flags & HAS_SSE4_1) inv_stxfm = inv_stxfm_sse4_1;

    if (flags & HAS_AVX2) inv_stxfm = inv_stxfm_avx2;

    inv_txfm = inv_txfm_c;

    if (flags & HAS_AVX2) inv_txfm = inv_txfm_avx2;

    mhccp_predict_hv_hbd = mhccp_predict_hv_hbd_c;

    if (flags & HAS_AVX2) mhccp_predict_hv_hbd = mhccp_predict_hv_hbd_avx2;

    refinemv_highbd_pad_mc_border = refinemv_highbd_pad_mc_border_c;

    if (flags & HAS_AVX2) refinemv_highbd_pad_mc_border = refinemv_highbd_pad_mc_border_avx2;

    resample_output = resample_output_c;

    if (flags & HAS_AVX2) resample_output = resample_output_avx2;

}

#endif

#ifdef __cplusplus

}  // extern "C"

#endif

#endif