third_party/libvpx_new/source/config/win/ia32/vp9_rtcd.h

   1 #ifndef VP9_RTCD_H_
   2 #define VP9_RTCD_H_
   3
   4 #ifdef RTCD_C
   5 #define RTCD_EXTERN
   6 #else
   7 #define RTCD_EXTERN extern
   8 #endif
   9
  10 /*
  11  * VP9
  12  */
  13
  14 #include "vpx/vpx_integer.h"
  15 #include "vp9/common/vp9_common.h"
  16 #include "vp9/common/vp9_enums.h"
  17
  18 struct macroblockd;
  19
  20 /* Encoder forward decls */
  21 struct macroblock;
  22 struct vp9_variance_vtable;
  23 struct search_site_config;
  24 struct mv;
  25 union int_mv;
  26 struct yv12_buffer_config;
  27
  28 #ifdef __cplusplus
  29 extern "C" {
  30 #endif
  31
  32 unsigned int vp9_avg_4x4_c(const uint8_t *, int p);
  33 unsigned int vp9_avg_4x4_sse2(const uint8_t *, int p);
  34 RTCD_EXTERN unsigned int (*vp9_avg_4x4)(const uint8_t *, int p);
  35
  36 unsigned int vp9_avg_8x8_c(const uint8_t *, int p);
  37 unsigned int vp9_avg_8x8_sse2(const uint8_t *, int p);
  38 RTCD_EXTERN unsigned int (*vp9_avg_8x8)(const uint8_t *, int p);
  39
  40 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
  41 int64_t vp9_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
  42 int64_t vp9_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
  43 RTCD_EXTERN int64_t (*vp9_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
  44
  45 int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
  46 int64_t vp9_block_error_fp_sse2(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
  47 RTCD_EXTERN int64_t (*vp9_block_error_fp)(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
  48
  49 int vp9_denoiser_filter_c(const uint8_t *sig, int sig_stride, const uint8_t *mc_avg, int mc_avg_stride, uint8_t *avg, int avg_stride, int increase_denoising, BLOCK_SIZE bs, int motion_magnitude);
  50 int vp9_denoiser_filter_sse2(const uint8_t *sig, int sig_stride, const uint8_t *mc_avg, int mc_avg_stride, uint8_t *avg, int avg_stride, int increase_denoising, BLOCK_SIZE bs, int motion_magnitude);
  51 RTCD_EXTERN int (*vp9_denoiser_filter)(const uint8_t *sig, int sig_stride, const uint8_t *mc_avg, int mc_avg_stride, uint8_t *avg, int avg_stride, int increase_denoising, BLOCK_SIZE bs, int motion_magnitude);
  52
  53 int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
  54 #define vp9_diamond_search_sad vp9_diamond_search_sad_c
  55
  56 void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
  57 void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
  58 void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
  59 RTCD_EXTERN void (*vp9_fdct8x8_quant)(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
  60
  61 void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  62 void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  63 RTCD_EXTERN void (*vp9_fht16x16)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  64
  65 void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  66 void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  67 RTCD_EXTERN void (*vp9_fht4x4)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  68
  69 void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  70 void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  71 RTCD_EXTERN void (*vp9_fht8x8)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
  72
  73 void vp9_filter_by_weight16x16_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight);
  74 void vp9_filter_by_weight16x16_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight);
  75 RTCD_EXTERN void (*vp9_filter_by_weight16x16)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight);
  76
  77 void vp9_filter_by_weight8x8_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight);
  78 void vp9_filter_by_weight8x8_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight);
  79 RTCD_EXTERN void (*vp9_filter_by_weight8x8)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight);
  80
  81 int vp9_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
  82 #define vp9_full_range_search vp9_full_range_search_c
  83
  84 int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
  85 int vp9_full_search_sadx3(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
  86 int vp9_full_search_sadx8(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
  87 RTCD_EXTERN int (*vp9_full_search_sad)(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
  88
  89 void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
  90 void vp9_fwht4x4_mmx(const int16_t *input, tran_low_t *output, int stride);
  91 RTCD_EXTERN void (*vp9_fwht4x4)(const int16_t *input, tran_low_t *output, int stride);
  92
  93 void vp9_hadamard_16x16_c(int16_t const *src_diff, int src_stride, int16_t *coeff);
  94 void vp9_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride, int16_t *coeff);
  95 RTCD_EXTERN void (*vp9_hadamard_16x16)(int16_t const *src_diff, int src_stride, int16_t *coeff);
  96
  97 void vp9_hadamard_8x8_c(int16_t const *src_diff, int src_stride, int16_t *coeff);
  98 void vp9_hadamard_8x8_sse2(int16_t const *src_diff, int src_stride, int16_t *coeff);
  99 RTCD_EXTERN void (*vp9_hadamard_8x8)(int16_t const *src_diff, int src_stride, int16_t *coeff);
 100
 101 void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
 102 void vp9_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
 103 RTCD_EXTERN void (*vp9_iht16x16_256_add)(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
 104
 105 void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
 106 void vp9_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
 107 RTCD_EXTERN void (*vp9_iht4x4_16_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
 108
 109 void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
 110 void vp9_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
 111 RTCD_EXTERN void (*vp9_iht8x8_64_add)(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
 112
 113 int16_t vp9_int_pro_col_c(uint8_t const *ref, const int width);
 114 int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width);
 115 RTCD_EXTERN int16_t (*vp9_int_pro_col)(uint8_t const *ref, const int width);
 116
 117 void vp9_int_pro_row_c(int16_t *hbuf, uint8_t const *ref, const int ref_stride, const int height);
 118 void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const *ref, const int ref_stride, const int height);
 119 RTCD_EXTERN void (*vp9_int_pro_row)(int16_t *hbuf, uint8_t const *ref, const int ref_stride, const int height);
 120
 121 void vp9_mbpost_proc_across_ip_c(uint8_t *src, int pitch, int rows, int cols, int flimit);
 122 void vp9_mbpost_proc_across_ip_xmm(uint8_t *src, int pitch, int rows, int cols, int flimit);
 123 RTCD_EXTERN void (*vp9_mbpost_proc_across_ip)(uint8_t *src, int pitch, int rows, int cols, int flimit);
 124
 125 void vp9_mbpost_proc_down_c(uint8_t *dst, int pitch, int rows, int cols, int flimit);
 126 void vp9_mbpost_proc_down_xmm(uint8_t *dst, int pitch, int rows, int cols, int flimit);
 127 RTCD_EXTERN void (*vp9_mbpost_proc_down)(uint8_t *dst, int pitch, int rows, int cols, int flimit);
 128
 129 void vp9_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
 130 void vp9_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
 131 RTCD_EXTERN void (*vp9_minmax_8x8)(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
 132
 133 void vp9_plane_add_noise_c(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
 134 void vp9_plane_add_noise_wmt(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
 135 RTCD_EXTERN void (*vp9_plane_add_noise)(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
 136
 137 void vp9_post_proc_down_and_across_c(const uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
 138 void vp9_post_proc_down_and_across_xmm(const uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
 139 RTCD_EXTERN void (*vp9_post_proc_down_and_across)(const uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
 140
 141 void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 142 void vp9_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 143 RTCD_EXTERN void (*vp9_quantize_fp)(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 144
 145 void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 146 #define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
 147
 148 int16_t vp9_satd_c(const int16_t *coeff, int length);
 149 int16_t vp9_satd_sse2(const int16_t *coeff, int length);
 150 RTCD_EXTERN int16_t (*vp9_satd)(const int16_t *coeff, int length);
 151
 152 void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 153 void vp9_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 154 RTCD_EXTERN void (*vp9_temporal_filter_apply)(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 155
 156 int vp9_vector_var_c(int16_t const *ref, int16_t const *src, const int bwl);
 157 int vp9_vector_var_sse2(int16_t const *ref, int16_t const *src, const int bwl);
 158 RTCD_EXTERN int (*vp9_vector_var)(int16_t const *ref, int16_t const *src, const int bwl);
 159
 160 void vp9_rtcd(void);
 161
 162 #ifdef RTCD_C
 163 #include "vpx_ports/x86.h"
 164 static void setup_rtcd_internal(void)
 165 {
 166     int flags = x86_simd_caps();
 167
 168     (void)flags;
 169
 170     vp9_avg_4x4 = vp9_avg_4x4_c;
 171     if (flags & HAS_SSE2) vp9_avg_4x4 = vp9_avg_4x4_sse2;
 172     vp9_avg_8x8 = vp9_avg_8x8_c;
 173     if (flags & HAS_SSE2) vp9_avg_8x8 = vp9_avg_8x8_sse2;
 174     vp9_block_error = vp9_block_error_c;
 175     if (flags & HAS_SSE2) vp9_block_error = vp9_block_error_sse2;
 176     if (flags & HAS_AVX2) vp9_block_error = vp9_block_error_avx2;
 177     vp9_block_error_fp = vp9_block_error_fp_c;
 178     if (flags & HAS_SSE2) vp9_block_error_fp = vp9_block_error_fp_sse2;
 179     vp9_denoiser_filter = vp9_denoiser_filter_c;
 180     if (flags & HAS_SSE2) vp9_denoiser_filter = vp9_denoiser_filter_sse2;
 181     vp9_fdct8x8_quant = vp9_fdct8x8_quant_c;
 182     if (flags & HAS_SSE2) vp9_fdct8x8_quant = vp9_fdct8x8_quant_sse2;
 183     if (flags & HAS_SSSE3) vp9_fdct8x8_quant = vp9_fdct8x8_quant_ssse3;
 184     vp9_fht16x16 = vp9_fht16x16_c;
 185     if (flags & HAS_SSE2) vp9_fht16x16 = vp9_fht16x16_sse2;
 186     vp9_fht4x4 = vp9_fht4x4_c;
 187     if (flags & HAS_SSE2) vp9_fht4x4 = vp9_fht4x4_sse2;
 188     vp9_fht8x8 = vp9_fht8x8_c;
 189     if (flags & HAS_SSE2) vp9_fht8x8 = vp9_fht8x8_sse2;
 190     vp9_filter_by_weight16x16 = vp9_filter_by_weight16x16_c;
 191     if (flags & HAS_SSE2) vp9_filter_by_weight16x16 = vp9_filter_by_weight16x16_sse2;
 192     vp9_filter_by_weight8x8 = vp9_filter_by_weight8x8_c;
 193     if (flags & HAS_SSE2) vp9_filter_by_weight8x8 = vp9_filter_by_weight8x8_sse2;
 194     vp9_full_search_sad = vp9_full_search_sad_c;
 195     if (flags & HAS_SSE3) vp9_full_search_sad = vp9_full_search_sadx3;
 196     if (flags & HAS_SSE4_1) vp9_full_search_sad = vp9_full_search_sadx8;
 197     vp9_fwht4x4 = vp9_fwht4x4_c;
 198     if (flags & HAS_MMX) vp9_fwht4x4 = vp9_fwht4x4_mmx;
 199     vp9_hadamard_16x16 = vp9_hadamard_16x16_c;
 200     if (flags & HAS_SSE2) vp9_hadamard_16x16 = vp9_hadamard_16x16_sse2;
 201     vp9_hadamard_8x8 = vp9_hadamard_8x8_c;
 202     if (flags & HAS_SSE2) vp9_hadamard_8x8 = vp9_hadamard_8x8_sse2;
 203     vp9_iht16x16_256_add = vp9_iht16x16_256_add_c;
 204     if (flags & HAS_SSE2) vp9_iht16x16_256_add = vp9_iht16x16_256_add_sse2;
 205     vp9_iht4x4_16_add = vp9_iht4x4_16_add_c;
 206     if (flags & HAS_SSE2) vp9_iht4x4_16_add = vp9_iht4x4_16_add_sse2;
 207     vp9_iht8x8_64_add = vp9_iht8x8_64_add_c;
 208     if (flags & HAS_SSE2) vp9_iht8x8_64_add = vp9_iht8x8_64_add_sse2;
 209     vp9_int_pro_col = vp9_int_pro_col_c;
 210     if (flags & HAS_SSE2) vp9_int_pro_col = vp9_int_pro_col_sse2;
 211     vp9_int_pro_row = vp9_int_pro_row_c;
 212     if (flags & HAS_SSE2) vp9_int_pro_row = vp9_int_pro_row_sse2;
 213     vp9_mbpost_proc_across_ip = vp9_mbpost_proc_across_ip_c;
 214     if (flags & HAS_SSE2) vp9_mbpost_proc_across_ip = vp9_mbpost_proc_across_ip_xmm;
 215     vp9_mbpost_proc_down = vp9_mbpost_proc_down_c;
 216     if (flags & HAS_SSE2) vp9_mbpost_proc_down = vp9_mbpost_proc_down_xmm;
 217     vp9_minmax_8x8 = vp9_minmax_8x8_c;
 218     if (flags & HAS_SSE2) vp9_minmax_8x8 = vp9_minmax_8x8_sse2;
 219     vp9_plane_add_noise = vp9_plane_add_noise_c;
 220     if (flags & HAS_SSE2) vp9_plane_add_noise = vp9_plane_add_noise_wmt;
 221     vp9_post_proc_down_and_across = vp9_post_proc_down_and_across_c;
 222     if (flags & HAS_SSE2) vp9_post_proc_down_and_across = vp9_post_proc_down_and_across_xmm;
 223     vp9_quantize_fp = vp9_quantize_fp_c;
 224     if (flags & HAS_SSE2) vp9_quantize_fp = vp9_quantize_fp_sse2;
 225     vp9_satd = vp9_satd_c;
 226     if (flags & HAS_SSE2) vp9_satd = vp9_satd_sse2;
 227     vp9_temporal_filter_apply = vp9_temporal_filter_apply_c;
 228     if (flags & HAS_SSE2) vp9_temporal_filter_apply = vp9_temporal_filter_apply_sse2;
 229     vp9_vector_var = vp9_vector_var_c;
 230     if (flags & HAS_SSE2) vp9_vector_var = vp9_vector_var_sse2;
 231 }
 232 #endif
 233
 234 #ifdef __cplusplus
 235 }  // extern "C"
 236 #endif
 237
 238 #endif