av1_rtcd_defs.pl 48.1 KB
Newer Older
Yaowu Xu's avatar
Yaowu Xu committed
1 2 3 4 5 6 7 8 9 10 11
sub av1_common_forward_decls() {
print <<EOF
/*
 * AV1
 */

#include "aom/aom_integer.h"
#include "av1/common/common.h"
#include "av1/common/enums.h"
#include "av1/common/quant_common.h"
#include "av1/common/filter.h"
12
#include "av1/common/convolve.h"
Yaowu Xu's avatar
Yaowu Xu committed
13
#include "av1/common/av1_txfm.h"
14
#include "av1/common/idct.h"
15
#include "av1/common/odintrin.h"
Yaowu Xu's avatar
Yaowu Xu committed
16 17 18 19 20

struct macroblockd;

/* Encoder forward decls */
struct macroblock;
21 22
struct fwd_txfm_param;
struct inv_txfm_param;
Yaowu Xu's avatar
Yaowu Xu committed
23 24 25 26 27
struct aom_variance_vtable;
struct search_site_config;
struct mv;
union int_mv;
struct yv12_buffer_config;
28
typedef uint16_t od_dering_in;
Yaowu Xu's avatar
Yaowu Xu committed
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
EOF
}
forward_decls qw/av1_common_forward_decls/;

# functions that are 64 bit only.
$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
if ($opts{arch} eq "x86_64") {
  $mmx_x86_64 = 'mmx';
  $sse2_x86_64 = 'sse2';
  $ssse3_x86_64 = 'ssse3';
  $avx_x86_64 = 'avx';
  $avx2_x86_64 = 'avx2';
}

#
# 10/12-tap convolution filters
#
Angie Chiang's avatar
Angie Chiang committed
46 47
add_proto qw/void av1_lowbd_convolve_init/, "void";
specialize qw/av1_lowbd_convolve_init ssse3/;
Angie Chiang's avatar
Angie Chiang committed
48

49
add_proto qw/void av1_convolve_horiz/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params";
Yaowu Xu's avatar
Yaowu Xu committed
50 51
specialize qw/av1_convolve_horiz ssse3/;

52
add_proto qw/void av1_convolve_vert/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params";
Yaowu Xu's avatar
Yaowu Xu committed
53 54
specialize qw/av1_convolve_vert ssse3/;

55
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Angie Chiang's avatar
Angie Chiang committed
56 57
  add_proto qw/void av1_highbd_convolve_init/, "void";
  specialize qw/av1_highbd_convolve_init sse4_1/;
Yaowu Xu's avatar
Yaowu Xu committed
58 59 60 61 62 63 64
  add_proto qw/void av1_highbd_convolve_horiz/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd";
  specialize qw/av1_highbd_convolve_horiz sse4_1/;
  add_proto qw/void av1_highbd_convolve_vert/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd";
  specialize qw/av1_highbd_convolve_vert sse4_1/;
}

#
65
# Inverse dct
Yaowu Xu's avatar
Yaowu Xu committed
66
#
67
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
68
  {
69
    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
70
      specialize qw/av1_iht4x4_16_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
71

72
    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
73
      specialize qw/av1_iht4x8_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
74

75
    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
76
      specialize qw/av1_iht8x4_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
77

78
    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
79
      specialize qw/av1_iht8x16_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
80

81
    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
82
      specialize qw/av1_iht16x8_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
83

84
    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
85
      specialize qw/av1_iht16x32_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
86

87
    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
88
      specialize qw/av1_iht32x16_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
89

90
    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
91

92
    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
93

94
    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
95

96
    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
97

98
    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
99
      specialize qw/av1_iht8x8_64_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
100

101
    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
102
      specialize qw/av1_iht16x16_256_add sse2 avx2/;
103

104
    add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
105 106
  }
} else {
107
  {
108
    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
109 110
    specialize qw/av1_iht4x4_16_add sse2 neon dspr2/;

111
    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
112
      specialize qw/av1_iht4x8_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
113

114
    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
115
      specialize qw/av1_iht8x4_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
116

117
    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
118
      specialize qw/av1_iht8x16_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
119

120
    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
121
      specialize qw/av1_iht16x8_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
122

123
    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
124
      specialize qw/av1_iht16x32_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
125

126
    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
127
      specialize qw/av1_iht32x16_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
128

129
    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
130

131
    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
132

133
    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
134

135
    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
136

137
    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
138
      specialize qw/av1_iht8x8_64_add sse2 neon dspr2/;
Yaowu Xu's avatar
Yaowu Xu committed
139

140
    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
141
      specialize qw/av1_iht16x16_256_add sse2 avx2 dspr2/;
Yaowu Xu's avatar
Yaowu Xu committed
142

143
    add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
144

Yaowu Xu's avatar
Yaowu Xu committed
145 146 147 148 149 150 151
    if (aom_config("CONFIG_EXT_TX") ne "yes") {
      specialize qw/av1_iht4x4_16_add msa/;
      specialize qw/av1_iht8x8_64_add msa/;
      specialize qw/av1_iht16x16_256_add msa/;
    }
  }
}
152

153
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
154 155

if (aom_config("CONFIG_TX64X64") eq "yes") {
156
  add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
157
}
Yaowu Xu's avatar
Yaowu Xu committed
158 159 160 161 162 163 164 165 166

if (aom_config("CONFIG_NEW_QUANT") eq "yes") {
  add_proto qw/void quantize_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

  add_proto qw/void quantize_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

  add_proto qw/void quantize_32x32_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

  add_proto qw/void quantize_32x32_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
167 168 169 170 171 172

  if (aom_config("CONFIG_TX64X64") eq "yes") {
    add_proto qw/void quantize_64x64_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

    add_proto qw/void quantize_64x64_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
  }
Yaowu Xu's avatar
Yaowu Xu committed
173 174
}

175 176
# FILTER_INTRA predictor functions
if (aom_config("CONFIG_FILTER_INTRA") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
177 178 179 180 181 182 183 184 185 186 187
  add_proto qw/void av1_dc_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_v_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_h_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d45_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d135_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d117_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d153_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d207_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d63_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_tm_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  # High bitdepth functions
188
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
189 190 191 192 193 194 195 196 197 198 199 200 201 202
    add_proto qw/void av1_highbd_dc_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_v_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_h_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d45_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d135_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d117_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d153_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d207_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d63_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_tm_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
  }
}

# High bitdepth functions
203
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
  #
  # Sub Pixel Filters
  #
  add_proto qw/void av1_highbd_convolve_copy/, "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";

  add_proto qw/void av1_highbd_convolve_avg/, "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";

  add_proto qw/void av1_highbd_convolve8/, "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";
  specialize qw/av1_highbd_convolve8/, "$sse2_x86_64";

  add_proto qw/void av1_highbd_convolve8_horiz/, "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";
  specialize qw/av1_highbd_convolve8_horiz/, "$sse2_x86_64";

  add_proto qw/void av1_highbd_convolve8_vert/, "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";
  specialize qw/av1_highbd_convolve8_vert/, "$sse2_x86_64";

  add_proto qw/void av1_highbd_convolve8_avg/, "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";
  specialize qw/av1_highbd_convolve8_avg/, "$sse2_x86_64";

  add_proto qw/void av1_highbd_convolve8_avg_horiz/, "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";
  specialize qw/av1_highbd_convolve8_avg_horiz/, "$sse2_x86_64";

  add_proto qw/void av1_highbd_convolve8_avg_vert/, "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";
  specialize qw/av1_highbd_convolve8_avg_vert/, "$sse2_x86_64";

  #
  # dct
  #
232
  add_proto qw/void av1_highbd_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
233

234
  add_proto qw/void av1_highbd_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
235

236
  add_proto qw/void av1_highbd_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
237

238
  add_proto qw/void av1_highbd_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
239

240
  add_proto qw/void av1_highbd_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
241

242
  add_proto qw/void av1_highbd_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
243

244
  add_proto qw/void av1_highbd_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
245

246
  add_proto qw/void av1_highbd_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
247

248
  add_proto qw/void av1_highbd_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
249

250
  add_proto qw/void av1_highbd_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
251

252
  add_proto qw/void av1_highbd_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
253

254
  add_proto qw/void av1_highbd_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
255

256
  add_proto qw/void av1_highbd_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct inv_txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
257 258
}

259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274
#inv txfm
add_proto qw/void av1_inv_txfm2d_add_4x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_8x4/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_8x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_16x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_16x32/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_32x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_4x4/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
specialize qw/av1_inv_txfm2d_add_4x4 sse4_1/;
add_proto qw/void av1_inv_txfm2d_add_8x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
specialize qw/av1_inv_txfm2d_add_8x8 sse4_1/;
add_proto qw/void av1_inv_txfm2d_add_16x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
specialize qw/av1_inv_txfm2d_add_16x16 sse4_1/;
add_proto qw/void av1_inv_txfm2d_add_32x32/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
specialize qw/av1_inv_txfm2d_add_32x32 avx2/;
add_proto qw/void av1_inv_txfm2d_add_64x64/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
275

Yaowu Xu's avatar
Yaowu Xu committed
276 277 278 279 280 281 282 283
#
# Encoder functions below this point.
#
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {

# ENCODEMB INVOKE

if (aom_config("CONFIG_AOM_QM") eq "yes") {
284
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
285 286 287
    # the transform coefficients are held in 32-bit
    # values, so the assembler code for  av1_block_error can no longer be used.
    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
288
    specialize qw/av1_block_error avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
289 290 291 292

    add_proto qw/void av1_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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";

    add_proto qw/void av1_quantize_fp_32x32/, "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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
293 294 295 296

    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
    }
Yaowu Xu's avatar
Yaowu Xu committed
297 298 299 300 301 302 303 304 305 306 307 308

    add_proto qw/void av1_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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
  } else {
    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
    specialize qw/av1_block_error avx2 msa/, "$sse2_x86inc";

    add_proto qw/int64_t av1_block_error_fp/, "const int16_t *coeff, const int16_t *dqcoeff, int block_size";
    specialize qw/av1_block_error_fp neon/, "$sse2_x86inc";

    add_proto qw/void av1_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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";

    add_proto qw/void av1_quantize_fp_32x32/, "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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
309 310 311 312

    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
    }
Yaowu Xu's avatar
Yaowu Xu committed
313 314 315 316

    add_proto qw/void av1_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, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
  }
} else {
317
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
318 319 320
    # the transform coefficients are held in 32-bit
    # values, so the assembler code for  av1_block_error can no longer be used.
    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
321
    specialize qw/av1_block_error avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
322 323

    add_proto qw/void av1_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";
Yi Luo's avatar
Yi Luo committed
324
    specialize qw/av1_quantize_fp sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
325 326

    add_proto qw/void av1_quantize_fp_32x32/, "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";
Yi Luo's avatar
Yi Luo committed
327
    specialize qw/av1_quantize_fp_32x32 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
328

329 330 331 332
    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "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";
    }

Yaowu Xu's avatar
Yaowu Xu committed
333 334 335 336 337 338 339 340 341
    add_proto qw/void av1_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";
  } else {
    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
    specialize qw/av1_block_error sse2 avx2 msa/;

    add_proto qw/int64_t av1_block_error_fp/, "const int16_t *coeff, const int16_t *dqcoeff, int block_size";
    specialize qw/av1_block_error_fp neon sse2/;

    add_proto qw/void av1_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";
Yi Luo's avatar
Yi Luo committed
342
    specialize qw/av1_quantize_fp neon sse2 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
343 344

    add_proto qw/void av1_quantize_fp_32x32/, "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";
Yi Luo's avatar
Yi Luo committed
345
    specialize qw/av1_quantize_fp_32x32 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
346

347 348 349 350
    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "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";
    }

Yaowu Xu's avatar
Yaowu Xu committed
351 352 353 354 355 356 357 358
    add_proto qw/void av1_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";
    specialize qw/av1_fdct8x8_quant sse2 ssse3 neon/;
  }

}

# fdct functions

359
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
360 361 362 363
specialize qw/av1_fht4x4 sse2/;

add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";

364
add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
365 366
specialize qw/av1_fht8x8 sse2/;

367
add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
368
specialize qw/av1_fht16x16 sse2 avx2/;
369

370
add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
Yi Luo's avatar
Yi Luo committed
371
specialize qw/av1_fht32x32 sse2 avx2/;
372

373
if (aom_config("CONFIG_TX64X64") eq "yes") {
374
  add_proto qw/void av1_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
375 376
}

377
add_proto qw/void av1_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
378
specialize qw/av1_fht4x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
379

380
add_proto qw/void av1_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
381
specialize qw/av1_fht8x4 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
382

383
add_proto qw/void av1_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
384
specialize qw/av1_fht8x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
385

386
add_proto qw/void av1_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
387
specialize qw/av1_fht16x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
388

389
add_proto qw/void av1_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
390
specialize qw/av1_fht16x32 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
391

392
add_proto qw/void av1_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
393
specialize qw/av1_fht32x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
394

395
add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
396

397
add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
398

399
add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
400

401
add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
402

403
if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
404 405 406 407 408 409 410
  if (aom_config("CONFIG_EXT_TX") ne "yes") {
    specialize qw/av1_fht4x4 msa/;
    specialize qw/av1_fht8x8 msa/;
    specialize qw/av1_fht16x16 msa/;
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
411 412
add_proto qw/void av1_fwd_idtx/, "const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type";

hui su's avatar
hui su committed
413 414 415
if (aom_config("CONFIG_DPCM_INTRA") eq "yes") {
  @sizes = (4, 8, 16, 32);
  foreach $size (@sizes) {
416 417 418
    if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
      add_proto "void", "av1_hbd_dpcm_ft$size", "const int16_t *input, int stride, TX_TYPE_1D tx_type, tran_low_t *output, int dir";
    }
hui su's avatar
hui su committed
419 420 421 422
    add_proto "void", "av1_dpcm_ft$size", "const int16_t *input, int stride, TX_TYPE_1D tx_type, tran_low_t *output";
  }
}

423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
#fwd txfm
add_proto qw/void av1_fwd_txfm2d_4x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_fwd_txfm2d_8x4/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_fwd_txfm2d_8x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_fwd_txfm2d_16x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_fwd_txfm2d_16x32/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_fwd_txfm2d_32x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_fwd_txfm2d_4x4/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
specialize qw/av1_fwd_txfm2d_4x4 sse4_1/;
add_proto qw/void av1_fwd_txfm2d_8x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
specialize qw/av1_fwd_txfm2d_8x8 sse4_1/;
add_proto qw/void av1_fwd_txfm2d_16x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
specialize qw/av1_fwd_txfm2d_16x16 sse4_1/;
add_proto qw/void av1_fwd_txfm2d_32x32/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
specialize qw/av1_fwd_txfm2d_32x32 sse4_1/;
add_proto qw/void av1_fwd_txfm2d_64x64/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
specialize qw/av1_fwd_txfm2d_64x64 sse4_1/;
Yaowu Xu's avatar
Yaowu Xu committed
440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455

#
# Motion search
#
add_proto qw/int av1_full_search_sad/, "const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv";
specialize qw/av1_full_search_sad sse3 sse4_1/;
$av1_full_search_sad_sse3=av1_full_search_sadx3;
$av1_full_search_sad_sse4_1=av1_full_search_sadx8;

add_proto qw/int av1_diamond_search_sad/, "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 aom_variance_vtable *fn_ptr, const struct mv *center_mv";

add_proto qw/int av1_full_range_search/, "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 aom_variance_vtable *fn_ptr, const struct mv *center_mv";

add_proto qw/void av1_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";
specialize qw/av1_temporal_filter_apply sse2 msa/;

456 457 458 459 460 461
if (aom_config("CONFIG_AOM_QM") eq "yes") {
  add_proto qw/void av1_quantize_b/, "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, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
} else {
  add_proto qw/void av1_quantize_b/, "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, int log_scale";
}

462
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
463 464 465 466 467 468 469 470 471 472

  # ENCODEMB INVOKE
  if (aom_config("CONFIG_NEW_QUANT") eq "yes") {
    add_proto qw/void highbd_quantize_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

    add_proto qw/void highbd_quantize_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

    add_proto qw/void highbd_quantize_32x32_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

    add_proto qw/void highbd_quantize_32x32_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
473 474 475 476 477 478

    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void highbd_quantize_64x64_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";

      add_proto qw/void highbd_quantize_64x64_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
    }
Yaowu Xu's avatar
Yaowu Xu committed
479 480 481 482 483
  }

  add_proto qw/int64_t av1_highbd_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd";
  specialize qw/av1_highbd_block_error sse2/;

484 485 486 487
  # fdct functions
  if (aom_config("CONFIG_TX64X64") eq "yes") {
    add_proto qw/void av1_highbd_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  }
Yaowu Xu's avatar
Yaowu Xu committed
488

489
  add_proto qw/void av1_highbd_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";
490

491
}
492

493 494
if (aom_config("CONFIG_AOM_QM") eq "yes") {
  add_proto qw/void av1_highbd_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, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
Yaowu Xu's avatar
Yaowu Xu committed
495

496
  add_proto qw/void av1_highbd_quantize_fp_32x32/, "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, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
Yaowu Xu's avatar
Yaowu Xu committed
497

498
  if (aom_config("CONFIG_TX64X64") eq "yes") {
499
    add_proto qw/void av1_highbd_quantize_fp_64x64/, "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, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
500 501
  }

502 503 504 505
  add_proto qw/void av1_highbd_quantize_b/, "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, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
} else {
  add_proto qw/void av1_highbd_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, int log_scale";
  specialize qw/av1_highbd_quantize_fp sse4_1 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
506

507
  add_proto qw/void av1_highbd_quantize_b/, "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, int log_scale";
Yaowu Xu's avatar
Yaowu Xu committed
508
}
509 510 511

add_proto qw/void av1_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";

Yaowu Xu's avatar
Yaowu Xu committed
512 513 514 515 516 517 518 519 520 521 522 523 524
# End av1_high encoder functions

if (aom_config("CONFIG_EXT_INTER") eq "yes") {
  add_proto qw/uint64_t av1_wedge_sse_from_residuals/, "const int16_t *r1, const int16_t *d, const uint8_t *m, int N";
  specialize qw/av1_wedge_sse_from_residuals sse2/;
  add_proto qw/int av1_wedge_sign_from_residuals/, "const int16_t *ds, const uint8_t *m, int N, int64_t limit";
  specialize qw/av1_wedge_sign_from_residuals sse2/;
  add_proto qw/void av1_wedge_compute_delta_squares/, "int16_t *d, const int16_t *a, const int16_t *b, int N";
  specialize qw/av1_wedge_compute_delta_squares sse2/;
}

}
# end encoder functions
525

526 527 528 529
# If PVQ is enabled, fwd transforms are required by decoder
if (aom_config("CONFIG_PVQ") eq "yes") {
# fdct functions

530
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
531
  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
532 533
  specialize qw/av1_fht4x4 sse2/;

534
  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
535 536
  specialize qw/av1_fht8x8 sse2/;

537
  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
538 539 540 541 542
  specialize qw/av1_fht16x16 sse2/;

  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
  specialize qw/av1_fwht4x4 sse2/;
} else {
543
  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
544 545
  specialize qw/av1_fht4x4 sse2 msa/;

546
  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
547 548
  specialize qw/av1_fht8x8 sse2 msa/;

549
  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct fwd_txfm_param *param";
550 551 552 553 554 555 556 557
  specialize qw/av1_fht16x16 sse2 msa/;

  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
  specialize qw/av1_fwht4x4 msa sse2/;
}

}

558 559
# Deringing Functions

560
if (aom_config("CONFIG_CDEF") eq "yes") {
561 562 563 564
  add_proto qw/void aom_clpf_block_hbd/, "uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
  add_proto qw/void aom_clpf_hblock_hbd/, "uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
  add_proto qw/void aom_clpf_block/, "uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
  add_proto qw/void aom_clpf_hblock/, "uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
565
  add_proto qw/int od_dir_find8/, "const od_dering_in *img, int stride, int32_t *var, int coeff_shift";
566 567
  add_proto qw/void od_filter_dering_direction_4x4/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping";
  add_proto qw/void od_filter_dering_direction_8x8/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping";
568 569 570 571 572

  add_proto qw/void copy_8x8_16bit_to_8bit/, "uint8_t *dst, int dstride, const uint16_t *src, int sstride";
  add_proto qw/void copy_4x4_16bit_to_8bit/, "uint8_t *dst, int dstride, const uint16_t *src, int sstride";
  add_proto qw/void copy_8x8_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride";
  add_proto qw/void copy_4x4_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride";
Michael Bebenita's avatar
Michael Bebenita committed
573 574
  add_proto qw/void copy_rect8_8bit_to_16bit/, "uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h";
  add_proto qw/void copy_rect8_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h";
575 576

# VS compiling for 32 bit targets does not support vector types in
577 578 579 580 581 582 583 584 585 586
  # structs as arguments, which makes the v256 type of the intrinsics
  # hard to support, so optimizations for this target are disabled.
  if ($opts{config} !~ /libs-x86-win32-vs.*/) {
    specialize qw/aom_clpf_block_hbd sse2 ssse3 sse4_1 neon/;
    specialize qw/aom_clpf_hblock_hbd sse2 ssse3 sse4_1 neon/;
    specialize qw/aom_clpf_block sse2 ssse3 sse4_1 neon/;
    specialize qw/aom_clpf_hblock sse2 ssse3 sse4_1 neon/;
    specialize qw/od_dir_find8 sse2 ssse3 sse4_1 neon/;
    specialize qw/od_filter_dering_direction_4x4 sse2 ssse3 sse4_1 neon/;
    specialize qw/od_filter_dering_direction_8x8 sse2 ssse3 sse4_1 neon/;
587 588 589 590 591

    specialize qw/copy_8x8_16bit_to_8bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_4x4_16bit_to_8bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_8x8_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_4x4_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
Michael Bebenita's avatar
Michael Bebenita committed
592 593
    specialize qw/copy_rect8_8bit_to_16bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_rect8_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
594
  }
595 596
}

597 598 599
# PVQ Functions

if (aom_config("CONFIG_PVQ") eq "yes") {
600
  add_proto qw/double pvq_search_rdo_double/, "const od_val16 *xcoeff, int n, int k, int *ypulse, double g2, double pvq_norm_lambda, int prev_k";
601 602 603
  specialize qw/pvq_search_rdo_double sse4_1/;
}

604 605 606 607
# WARPED_MOTION / GLOBAL_MOTION functions

if ((aom_config("CONFIG_WARPED_MOTION") eq "yes") ||
    (aom_config("CONFIG_GLOBAL_MOTION") eq "yes")) {
608
  add_proto qw/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";
609
  specialize qw/av1_warp_affine sse2 ssse3/;
610

611 612 613 614
  if (aom_config("CONFIG_CONVOLVE_ROUND") eq "yes") {
    add_proto qw/void av1_warp_affine_post_round/, "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";
  }

615
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
616
    add_proto qw/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";
David Barker's avatar
David Barker committed
617
    specialize qw/av1_highbd_warp_affine ssse3/;
618 619 620
    if (aom_config("CONFIG_CONVOLVE_ROUND") eq "yes") {
      add_proto qw/void av1_highbd_warp_affine_post_round/, "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";
    }
621
  }
622 623
}

624
if (aom_config("CONFIG_GLOBAL_MOTION") eq "yes" &&
625 626 627 628 629
    aom_config("CONFIG_AV1_ENCODER") eq "yes") {
  add_proto qw/double compute_cross_correlation/, "unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2";
  specialize qw/compute_cross_correlation sse4_1/;
}

David Barker's avatar
David Barker committed
630 631 632
# LOOP_RESTORATION functions

if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
633
  add_proto qw/void apply_selfguided_restoration/, "uint8_t *dat, int width, int height, int stride, int eps, int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf";
David Barker's avatar
David Barker committed
634
  specialize qw/apply_selfguided_restoration sse4_1/;
635

636
  add_proto qw/void av1_selfguided_restoration/, "uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps, int32_t *tmpbuf";
637 638
  specialize qw/av1_selfguided_restoration sse4_1/;

639 640
  add_proto qw/void av1_highpass_filter/, "uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
  specialize qw/av1_highpass_filter sse4_1/;
641

642
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
643
    add_proto qw/void apply_selfguided_restoration_highbd/, "uint16_t *dat, int width, int height, int stride, int bit_depth, int eps, int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf";
644
    specialize qw/apply_selfguided_restoration_highbd sse4_1/;
645 646

    add_proto qw/void av1_selfguided_restoration_highbd/, "uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps, int32_t *tmpbuf";
647
    specialize qw/av1_selfguided_restoration_highbd sse4_1/;
648

649 650
    add_proto qw/void av1_highpass_filter_highbd/, "uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
    specialize qw/av1_highpass_filter_highbd sse4_1/;
651
  }
David Barker's avatar
David Barker committed
652 653
}

David Barker's avatar
David Barker committed
654 655 656 657 658
# CONVOLVE_ROUND/COMPOUND_ROUND functions

if (aom_config("CONFIG_CONVOLVE_ROUND") eq "yes") {
    add_proto qw/void av1_convolve_2d/, "const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params";
    specialize qw/av1_convolve_2d sse2/;
659 660
    add_proto qw/void av1_convolve_rounding/, "const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits";
    specialize qw/av1_convolve_rounding avx2/;
David Barker's avatar
David Barker committed
661 662 663 664

  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
    add_proto qw/void av1_highbd_convolve_2d/, "const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
    specialize qw/av1_highbd_convolve_2d ssse3/;
665 666
    add_proto qw/void av1_highbd_convolve_rounding/, "const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits, int bd";
    specialize qw/av1_highbd_convolve_rounding avx2/;
David Barker's avatar
David Barker committed
667 668 669
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
670
1;