av1_rtcd_defs.pl 39.3 KB
Newer Older
Yaowu Xu's avatar
Yaowu Xu committed
1 2 3 4 5 6 7
sub av1_common_forward_decls() {
print <<EOF
/*
 * AV1
 */

#include "aom/aom_integer.h"
8
#include "aom_dsp/txfm_common.h"
Yaowu Xu's avatar
Yaowu Xu committed
9 10 11 12
#include "av1/common/common.h"
#include "av1/common/enums.h"
#include "av1/common/quant_common.h"
#include "av1/common/filter.h"
13
#include "av1/common/convolve.h"
Yaowu Xu's avatar
Yaowu Xu committed
14
#include "av1/common/av1_txfm.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
struct txfm_param;
Yaowu Xu's avatar
Yaowu Xu committed
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
struct aom_variance_vtable;
struct search_site_config;
struct mv;
union int_mv;
struct yv12_buffer_config;
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
44 45
add_proto qw/void av1_lowbd_convolve_init/, "void";
specialize qw/av1_lowbd_convolve_init ssse3/;
Angie Chiang's avatar
Angie Chiang committed
46

47
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
48 49
specialize qw/av1_convolve_horiz ssse3/;

50
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
51 52
specialize qw/av1_convolve_vert ssse3/;

53
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Angie Chiang's avatar
Angie Chiang committed
54 55
  add_proto qw/void av1_highbd_convolve_init/, "void";
  specialize qw/av1_highbd_convolve_init sse4_1/;
Yaowu Xu's avatar
Yaowu Xu committed
56 57 58 59 60 61 62
  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/;
}

#
63
# Inverse dct
Yaowu Xu's avatar
Yaowu Xu committed
64
#
65
add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
66
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
James Zern's avatar
James Zern committed
67
  specialize qw/av1_iht4x4_16_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
68
} else {
69
  if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
70 71
    specialize qw/av1_iht4x4_16_add sse2 neon/;
  }
72
}
Yaowu Xu's avatar
Yaowu Xu committed
73

74 75
add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
specialize qw/av1_iht4x8_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
76

77 78
add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
specialize qw/av1_iht8x4_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
79

80 81
add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
specialize qw/av1_iht8x16_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
82

83 84
add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
specialize qw/av1_iht16x8_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
85

86 87
add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
specialize qw/av1_iht16x32_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
88

89 90
add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
specialize qw/av1_iht32x16_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
91

92
add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
93

94
add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
95

96
add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
97

98
add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
99

100 101 102 103
add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
  specialize qw/av1_iht8x8_64_add sse2/;
} else {
104
  if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
105 106
    specialize qw/av1_iht8x8_64_add sse2 neon/;
  }
107
}
Yaowu Xu's avatar
Yaowu Xu committed
108

109 110 111 112
add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
  specialize qw/av1_iht16x16_256_add sse2 avx2/;
} else {
113
  if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
114 115
    specialize qw/av1_iht16x16_256_add sse2 avx2/;
  }
116
}
Yaowu Xu's avatar
Yaowu Xu committed
117

118
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
119

120
if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
James Zern's avatar
James Zern committed
121
  if (aom_config("CONFIG_EXT_TX") ne "yes") {
122
    if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
123 124
      specialize qw/av1_iht4x4_16_add msa/;
    }
125
    if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
126 127
      specialize qw/av1_iht8x8_64_add msa/;
    }
128
    if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
129
      specialize qw/av1_iht16x16_256_add msa/;
Yaowu Xu's avatar
Yaowu Xu committed
130 131 132
    }
  }
}
133

134
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
135 136

if (aom_config("CONFIG_TX64X64") eq "yes") {
137
  add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
138 139
  add_proto qw/void av1_iht32x64_2048_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
  add_proto qw/void av1_iht64x32_2048_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
140
}
Yaowu Xu's avatar
Yaowu Xu committed
141 142 143 144 145 146 147 148 149

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";
150 151 152 153 154 155

  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
156 157
}

158 159
# FILTER_INTRA predictor functions
if (aom_config("CONFIG_FILTER_INTRA") eq "yes") {
160 161 162 163 164
  add_proto qw/void av1_dc_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_v_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_h_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d117_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d153_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left";
Urvang Joshi's avatar
Urvang Joshi committed
165
  add_proto qw/void av1_paeth_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint8_t *above, const uint8_t *left";
Yaowu Xu's avatar
Yaowu Xu committed
166
  # High bitdepth functions
167
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
168 169 170 171 172
    add_proto qw/void av1_highbd_dc_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, TX_SIZE tx_size, 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, TX_SIZE tx_size, 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, TX_SIZE tx_size, 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, TX_SIZE tx_size, 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, TX_SIZE tx_size, const uint16_t *above, const uint16_t *left, int bd";
Urvang Joshi's avatar
Urvang Joshi committed
173
    add_proto qw/void av1_highbd_paeth_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, TX_SIZE tx_size, const uint16_t *above, const uint16_t *left, int bd";
Yaowu Xu's avatar
Yaowu Xu committed
174 175 176 177
  }
}

# High bitdepth functions
178
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
  #
  # 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
  #
207
  add_proto qw/void av1_highbd_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
208

209
  add_proto qw/void av1_highbd_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
210

211
  add_proto qw/void av1_highbd_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
212

213
  add_proto qw/void av1_highbd_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
214

215
  add_proto qw/void av1_highbd_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
216

217
  add_proto qw/void av1_highbd_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
218

219
  add_proto qw/void av1_highbd_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
220

221
  add_proto qw/void av1_highbd_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
222

223
  add_proto qw/void av1_highbd_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
224

225
  add_proto qw/void av1_highbd_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
226

227
  add_proto qw/void av1_highbd_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
228

229
  add_proto qw/void av1_highbd_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
230

231
  add_proto qw/void av1_highbd_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
232 233
}

234
#inv txfm
235 236 237 238 239 240 241
add_proto qw/void av1_inv_txfm2d_add_4x8/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_8x4/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_8x16/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_16x8/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_16x32/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_32x16/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_4x4/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
242
if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
243 244
  specialize qw/av1_inv_txfm2d_add_4x4 sse4_1/;
}
245
add_proto qw/void av1_inv_txfm2d_add_8x8/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
246
if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
247 248
  specialize qw/av1_inv_txfm2d_add_8x8 sse4_1/;
}
249
add_proto qw/void av1_inv_txfm2d_add_16x16/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
250
if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
251 252
  specialize qw/av1_inv_txfm2d_add_16x16 sse4_1/;
}
253
add_proto qw/void av1_inv_txfm2d_add_32x32/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
254
if (aom_config("CONFIG_DAALA_TX32") ne "yes") {
255 256
  specialize qw/av1_inv_txfm2d_add_32x32 avx2/;
}
257
if (aom_config("CONFIG_TX64X64") eq "yes") {
258 259 260
  add_proto qw/void av1_inv_txfm2d_add_64x64/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_inv_txfm2d_add_64x32/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_inv_txfm2d_add_32x64/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
261
}
262

Yaowu Xu's avatar
Yaowu Xu committed
263 264 265 266 267
#
# Encoder functions below this point.
#
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {

James Zern's avatar
James Zern committed
268
  # ENCODEMB INVOKE
Yaowu Xu's avatar
Yaowu Xu committed
269

James Zern's avatar
James Zern committed
270
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
271 272 273
    # 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";
274
    specialize qw/av1_block_error avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
275 276

    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
277
    specialize qw/av1_quantize_fp sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
278 279

    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
280
    specialize qw/av1_quantize_fp_32x32 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
281

282 283 284
    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";
    }
James Zern's avatar
James Zern committed
285
  } else {
Yaowu Xu's avatar
Yaowu Xu committed
286 287 288 289 290 291 292
    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
293
    specialize qw/av1_quantize_fp neon sse2 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
294 295

    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
296
    specialize qw/av1_quantize_fp_32x32 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
297

298 299 300 301
    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";
    }

James Zern's avatar
James Zern committed
302
  }
Yaowu Xu's avatar
Yaowu Xu committed
303

James Zern's avatar
James Zern committed
304
  # fdct functions
Yaowu Xu's avatar
Yaowu Xu committed
305

James Zern's avatar
James Zern committed
306
  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
307
  if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
308 309
    specialize qw/av1_fht4x4 sse2/;
  }
310

James Zern's avatar
James Zern committed
311
  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
312

James Zern's avatar
James Zern committed
313
  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
314
  if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
315 316
    specialize qw/av1_fht8x8 sse2/;
  }
317

James Zern's avatar
James Zern committed
318
  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
319
  if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
320 321
    specialize qw/av1_fht16x16 sse2 avx2/;
  }
322

James Zern's avatar
James Zern committed
323
  add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
324
  if (aom_config("CONFIG_DAALA_TX32") ne "yes") {
James Zern's avatar
James Zern committed
325 326
    specialize qw/av1_fht32x32 sse2 avx2/;
  }
327

James Zern's avatar
James Zern committed
328 329
  if (aom_config("CONFIG_TX64X64") eq "yes") {
    add_proto qw/void av1_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
330 331
    add_proto qw/void av1_fht64x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
    add_proto qw/void av1_fht32x64/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
James Zern's avatar
James Zern committed
332
  }
333

James Zern's avatar
James Zern committed
334 335
  add_proto qw/void av1_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  specialize qw/av1_fht4x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
336

James Zern's avatar
James Zern committed
337 338
  add_proto qw/void av1_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  specialize qw/av1_fht8x4 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
339

James Zern's avatar
James Zern committed
340 341
  add_proto qw/void av1_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  specialize qw/av1_fht8x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
342

James Zern's avatar
James Zern committed
343 344
  add_proto qw/void av1_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  specialize qw/av1_fht16x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
345

James Zern's avatar
James Zern committed
346 347
  add_proto qw/void av1_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  specialize qw/av1_fht16x32 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
348

James Zern's avatar
James Zern committed
349 350
  add_proto qw/void av1_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  specialize qw/av1_fht32x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
351

James Zern's avatar
James Zern committed
352
  add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
353

James Zern's avatar
James Zern committed
354
  add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
355

James Zern's avatar
James Zern committed
356
  add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
357

James Zern's avatar
James Zern committed
358
  add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
359

James Zern's avatar
James Zern committed
360 361
  if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
    if (aom_config("CONFIG_EXT_TX") ne "yes") {
362
      if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
363 364
        specialize qw/av1_fht4x4 msa/;
      }
365
      if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
366 367
        specialize qw/av1_fht8x8 msa/;
      }
368
      if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
369 370
        specialize qw/av1_fht16x16 msa/;
      }
371
    }
372 373
  }

374
  add_proto qw/void av1_fwd_idtx/, "const int16_t *src_diff, tran_low_t *coeff, int stride, int bsx, int bsy, TX_TYPE tx_type";
James Zern's avatar
James Zern committed
375 376

  #fwd txfm
377 378 379 380 381 382 383
  add_proto qw/void av1_fwd_txfm2d_4x8/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_8x4/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_8x16/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_16x8/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_16x32/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_32x16/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_4x4/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
384
  if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
385 386
    specialize qw/av1_fwd_txfm2d_4x4 sse4_1/;
  }
387
  add_proto qw/void av1_fwd_txfm2d_8x8/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
388
  if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
389 390
    specialize qw/av1_fwd_txfm2d_8x8 sse4_1/;
  }
391
  add_proto qw/void av1_fwd_txfm2d_16x16/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
392
  if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
393 394
    specialize qw/av1_fwd_txfm2d_16x16 sse4_1/;
  }
395
  add_proto qw/void av1_fwd_txfm2d_32x32/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
396
  if (aom_config("CONFIG_DAALA_TX32") ne "yes") {
James Zern's avatar
James Zern committed
397 398
    specialize qw/av1_fwd_txfm2d_32x32 sse4_1/;
  }
399 400

  if (aom_config("CONFIG_TX64X64") eq "yes") {
401 402 403
    add_proto qw/void av1_fwd_txfm2d_32x64/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
    add_proto qw/void av1_fwd_txfm2d_64x32/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
    add_proto qw/void av1_fwd_txfm2d_64x64/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
James Zern's avatar
James Zern committed
404 405 406 407 408 409 410 411
  }
  #
  # 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;
Yaowu Xu's avatar
Yaowu Xu committed
412

James Zern's avatar
James Zern committed
413
  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";
Yaowu Xu's avatar
Yaowu Xu committed
414

James Zern's avatar
James Zern committed
415
  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";
Yaowu Xu's avatar
Yaowu Xu committed
416

James Zern's avatar
James Zern committed
417 418
  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/;
Yaowu Xu's avatar
Yaowu Xu committed
419

James Zern's avatar
James Zern committed
420 421 422 423 424
  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";
  }
425

426 427 428 429
  if (aom_config("CONFIG_LGT_FROM_PRED") eq "yes") {
    add_proto qw/void flgt2d_from_pred/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  }

James Zern's avatar
James Zern committed
430
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
431

James Zern's avatar
James Zern committed
432 433 434
    # 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";
Yaowu Xu's avatar
Yaowu Xu committed
435

James Zern's avatar
James Zern committed
436
      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";
Yaowu Xu's avatar
Yaowu Xu committed
437

James Zern's avatar
James Zern committed
438
      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";
Yaowu Xu's avatar
Yaowu Xu committed
439

James Zern's avatar
James Zern committed
440
      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";
441

James Zern's avatar
James Zern committed
442 443
      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";
444

James Zern's avatar
James Zern committed
445 446
        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";
      }
447
    }
Yaowu Xu's avatar
Yaowu Xu committed
448

James Zern's avatar
James Zern committed
449 450
    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/;
Yaowu Xu's avatar
Yaowu Xu committed
451

James Zern's avatar
James Zern committed
452
    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";
453

James Zern's avatar
James Zern committed
454
  }
455

James Zern's avatar
James Zern committed
456 457
  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/;
458

James Zern's avatar
James Zern committed
459
  add_proto qw/void av1_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
460

James Zern's avatar
James Zern committed
461
  # End av1_high encoder functions
Yaowu Xu's avatar
Yaowu Xu committed
462

463 464 465 466 467 468
  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/;
Yaowu Xu's avatar
Yaowu Xu committed
469 470 471

}
# end encoder functions
472 473 474

# Deringing Functions

475
if (aom_config("CONFIG_CDEF") eq "yes") {
Steinar Midtskogen's avatar
Steinar Midtskogen committed
476
  add_proto qw/int cdef_find_dir/, "const uint16_t *img, int stride, int32_t *var, int coeff_shift";
477 478 479 480 481 482 483 484 485 486 487 488 489 490
  if (aom_config("CONFIG_CDEF_SINGLEPASS") ne "yes") {
    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";
    add_proto qw/void cdef_direction_4x4/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping";
    add_proto qw/void cdef_direction_8x8/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping";
    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";
  } else {
    add_proto qw/void cdef_filter_block/, "uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max";
  }
491

Michael Bebenita's avatar
Michael Bebenita committed
492 493
  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";
494

James Zern's avatar
James Zern committed
495
  # VS compiling for 32 bit targets does not support vector types in
496 497 498
  # 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.*/) {
499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520
    if (aom_config("CONFIG_CDEF_SINGLEPASS") eq "yes") {
      specialize qw/cdef_find_dir sse2 ssse3 sse4_1 avx2 neon/;
      specialize qw/cdef_filter_block sse2 ssse3 sse4_1 avx2 neon/;
      specialize qw/copy_rect8_8bit_to_16bit sse2 ssse3 sse4_1 avx2 neon/;
      specialize qw/copy_rect8_16bit_to_16bit sse2 ssse3 sse4_1 avx2 neon/;
    } else {
      specialize qw/cdef_find_dir sse2 ssse3 sse4_1 neon/;
      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/cdef_find_dir sse2 ssse3 sse4_1 neon/;
      specialize qw/cdef_direction_4x4 sse2 ssse3 sse4_1 neon/;
      specialize qw/cdef_direction_8x8 sse2 ssse3 sse4_1 neon/;

      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/;
      specialize qw/copy_rect8_8bit_to_16bit sse2 ssse3 sse4_1 neon/;
      specialize qw/copy_rect8_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
    }
521
  }
522 523
}

524 525
# WARPED_MOTION / GLOBAL_MOTION functions

526 527
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";
specialize qw/av1_warp_affine sse2 ssse3/;
528

529 530 531
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
  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";
  specialize qw/av1_highbd_warp_affine ssse3/;
532 533
}

534
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
535 536 537 538
  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
539 540 541
# LOOP_RESTORATION functions

if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
542
  add_proto qw/void apply_selfguided_restoration/, "const uint8_t *dat, int width, int height, int stride, int eps, const int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf";
David Barker's avatar
David Barker committed
543
  specialize qw/apply_selfguided_restoration sse4_1/;
544

545
  add_proto qw/void av1_selfguided_restoration/, "const uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
546 547
  specialize qw/av1_selfguided_restoration sse4_1/;

548
  add_proto qw/void av1_highpass_filter/, "const uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
549
  specialize qw/av1_highpass_filter sse4_1/;
550

551
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
552
    add_proto qw/void apply_selfguided_restoration_highbd/, "const uint16_t *dat, int width, int height, int stride, int bit_depth, int eps, const int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf";
553
    specialize qw/apply_selfguided_restoration_highbd sse4_1/;
554

555
    add_proto qw/void av1_selfguided_restoration_highbd/, "const uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps";
556
    specialize qw/av1_selfguided_restoration_highbd sse4_1/;
557

558
    add_proto qw/void av1_highpass_filter_highbd/, "const uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
559
    specialize qw/av1_highpass_filter_highbd sse4_1/;
560
  }
David Barker's avatar
David Barker committed
561 562
}

David Barker's avatar
David Barker committed
563 564 565
# CONVOLVE_ROUND/COMPOUND_ROUND functions

if (aom_config("CONFIG_CONVOLVE_ROUND") eq "yes") {
James Zern's avatar
James Zern committed
566 567 568 569
  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/;
  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
570

James Zern's avatar
James Zern committed
571
  add_proto qw/void av1_convolve_2d_scale/, "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_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params";
572 573 574
  if (aom_config("CONFIG_COMPOUND_ROUND") ne "yes") {
    specialize qw/av1_convolve_2d_scale sse4_1/;
  }
575

David Barker's avatar
David Barker committed
576 577 578
  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/;
579 580
    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/;
581

582
    add_proto qw/void av1_highbd_convolve_2d_scale/, "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 x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd";
583 584 585
    if (aom_config("CONFIG_COMPOUND_ROUND") ne "yes") {
        specialize qw/av1_highbd_convolve_2d_scale sse4_1/;
    }
David Barker's avatar
David Barker committed
586 587 588
  }
}

589 590 591 592
# INTRA_EDGE functions
if (aom_config("CONFIG_INTRA_EDGE") eq "yes") {
  add_proto qw/void av1_filter_intra_edge/, "uint8_t *p, int sz, int strength";
  specialize qw/av1_filter_intra_edge sse4_1/;
Joe Young's avatar
Joe Young committed
593 594
  add_proto qw/void av1_upsample_intra_edge/, "uint8_t *p, int sz";
  specialize qw/av1_upsample_intra_edge sse4_1/;
595 596 597
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
    add_proto qw/void av1_filter_intra_edge_high/, "uint16_t *p, int sz, int strength";
    specialize qw/av1_filter_intra_edge_high sse4_1/;
Joe Young's avatar
Joe Young committed
598 599
    add_proto qw/void av1_upsample_intra_edge_high/, "uint16_t *p, int sz, int bd";
    specialize qw/av1_upsample_intra_edge_high sse4_1/;
600 601 602
  }
}

603
1;