av1_rtcd_defs.pl 39.9 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 54 55 56 57 58 59 60 61
if (aom_config("CONFIG_FRAME_SUPERRES") eq "yes") {
  if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
    add_proto qw/void av1_convolve_horiz_rs/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int interp_taps, const int x0_qn, const int x_step_qn";
    if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
      add_proto qw/void av1_highbd_convolve_horiz_rs/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const int16_t *x_filters, int interp_taps, const int x0_qn, const int x_step_qn, int bd";
    }
  }
}

62
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Angie Chiang's avatar
Angie Chiang committed
63 64
  add_proto qw/void av1_highbd_convolve_init/, "void";
  specialize qw/av1_highbd_convolve_init sse4_1/;
Yaowu Xu's avatar
Yaowu Xu committed
65 66 67 68 69 70 71
  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/;
}

#
72
# Inverse dct
Yaowu Xu's avatar
Yaowu Xu committed
73
#
74
add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
75
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
James Zern's avatar
James Zern committed
76
  specialize qw/av1_iht4x4_16_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
77
} else {
78
  if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
79 80
    specialize qw/av1_iht4x4_16_add sse2 neon/;
  }
81
}
Yaowu Xu's avatar
Yaowu Xu committed
82

83 84
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
85

86 87
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
88

89 90
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
91

92 93
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
94

95 96
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
97

98 99
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
100

101
add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
102

103
add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
104

105
add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
106

107
add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
108

109 110 111 112
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 {
113
  if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
114 115
    specialize qw/av1_iht8x8_64_add sse2 neon/;
  }
116
}
Yaowu Xu's avatar
Yaowu Xu committed
117

118 119 120 121
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 {
122
  if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
123 124
    specialize qw/av1_iht16x16_256_add sse2 avx2/;
  }
125
}
Yaowu Xu's avatar
Yaowu Xu committed
126

127
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
128

129
if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
James Zern's avatar
James Zern committed
130
  if (aom_config("CONFIG_EXT_TX") ne "yes") {
131
    if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
132 133
      specialize qw/av1_iht4x4_16_add msa/;
    }
134
    if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
135 136
      specialize qw/av1_iht8x8_64_add msa/;
    }
137
    if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
138
      specialize qw/av1_iht16x16_256_add msa/;
Yaowu Xu's avatar
Yaowu Xu committed
139 140 141
    }
  }
}
142

143
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
144 145

if (aom_config("CONFIG_TX64X64") eq "yes") {
146
  add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
147 148
  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";
149
}
Yaowu Xu's avatar
Yaowu Xu committed
150 151 152 153 154 155 156 157 158

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";
159 160 161 162 163 164

  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
165 166
}

167 168
# FILTER_INTRA predictor functions
if (aom_config("CONFIG_FILTER_INTRA") eq "yes") {
169 170 171 172 173
  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
174
  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
175
  # High bitdepth functions
176
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
177 178 179 180 181
    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
182
    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
183 184 185 186
  }
}

# High bitdepth functions
187
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215
  #
  # 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
  #
216
  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
217

218
  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
219

220
  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
221

222
  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
223

224
  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
225

226
  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
227

228
  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
229

230
  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";
231

232
  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";
233

234
  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";
235

236
  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";
237

238
  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
239

240
  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
241 242
}

243
#inv txfm
244 245 246 247 248 249 250
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";
251
if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
252 253
  specialize qw/av1_inv_txfm2d_add_4x4 sse4_1/;
}
254
add_proto qw/void av1_inv_txfm2d_add_8x8/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
255
if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
256 257
  specialize qw/av1_inv_txfm2d_add_8x8 sse4_1/;
}
258
add_proto qw/void av1_inv_txfm2d_add_16x16/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
259
if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
260 261
  specialize qw/av1_inv_txfm2d_add_16x16 sse4_1/;
}
262
add_proto qw/void av1_inv_txfm2d_add_32x32/, "const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd";
263
if (aom_config("CONFIG_DAALA_TX32") ne "yes") {
264 265
  specialize qw/av1_inv_txfm2d_add_32x32 avx2/;
}
266
if (aom_config("CONFIG_TX64X64") eq "yes") {
267 268 269
  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";
270
}
271

Yaowu Xu's avatar
Yaowu Xu committed
272 273 274 275 276
#
# Encoder functions below this point.
#
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {

James Zern's avatar
James Zern committed
277
  # ENCODEMB INVOKE
Yaowu Xu's avatar
Yaowu Xu committed
278

James Zern's avatar
James Zern committed
279
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
280 281 282
    # 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";
283
    specialize qw/av1_block_error avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
284 285

    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
286
    specialize qw/av1_quantize_fp sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
287 288

    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
289
    specialize qw/av1_quantize_fp_32x32 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
290

291 292 293
    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
294
  } else {
Yaowu Xu's avatar
Yaowu Xu committed
295 296 297 298 299 300 301
    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
302
    specialize qw/av1_quantize_fp neon sse2 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
303 304

    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
305
    specialize qw/av1_quantize_fp_32x32 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
306

307 308 309 310
    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
311
  }
Yaowu Xu's avatar
Yaowu Xu committed
312

James Zern's avatar
James Zern committed
313
  # fdct functions
Yaowu Xu's avatar
Yaowu Xu committed
314

James Zern's avatar
James Zern committed
315
  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
316
  if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
317 318
    specialize qw/av1_fht4x4 sse2/;
  }
319

James Zern's avatar
James Zern committed
320
  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
321

James Zern's avatar
James Zern committed
322
  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
323
  if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
324 325
    specialize qw/av1_fht8x8 sse2/;
  }
326

James Zern's avatar
James Zern committed
327
  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
328
  if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
329 330
    specialize qw/av1_fht16x16 sse2 avx2/;
  }
331

James Zern's avatar
James Zern committed
332
  add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
333
  if (aom_config("CONFIG_DAALA_TX32") ne "yes") {
James Zern's avatar
James Zern committed
334 335
    specialize qw/av1_fht32x32 sse2 avx2/;
  }
336

James Zern's avatar
James Zern committed
337 338
  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";
339 340
    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
341
  }
342

James Zern's avatar
James Zern committed
343 344
  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
345

James Zern's avatar
James Zern committed
346 347
  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
348

James Zern's avatar
James Zern committed
349 350
  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
351

James Zern's avatar
James Zern committed
352 353
  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
354

James Zern's avatar
James Zern committed
355 356
  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
357

James Zern's avatar
James Zern committed
358 359
  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
360

James Zern's avatar
James Zern committed
361
  add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
362

James Zern's avatar
James Zern committed
363
  add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
364

James Zern's avatar
James Zern committed
365
  add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
366

James Zern's avatar
James Zern committed
367
  add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
368

James Zern's avatar
James Zern committed
369 370
  if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
    if (aom_config("CONFIG_EXT_TX") ne "yes") {
371
      if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
372 373
        specialize qw/av1_fht4x4 msa/;
      }
374
      if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
375 376
        specialize qw/av1_fht8x8 msa/;
      }
377
      if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
378 379
        specialize qw/av1_fht16x16 msa/;
      }
380
    }
381 382
  }

383
  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
384 385

  #fwd txfm
386 387 388 389 390 391 392
  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";
393
  if (aom_config("CONFIG_DAALA_TX4") ne "yes") {
James Zern's avatar
James Zern committed
394 395
    specialize qw/av1_fwd_txfm2d_4x4 sse4_1/;
  }
396
  add_proto qw/void av1_fwd_txfm2d_8x8/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
397
  if (aom_config("CONFIG_DAALA_TX8") ne "yes") {
James Zern's avatar
James Zern committed
398 399
    specialize qw/av1_fwd_txfm2d_8x8 sse4_1/;
  }
400
  add_proto qw/void av1_fwd_txfm2d_16x16/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
401
  if (aom_config("CONFIG_DAALA_TX16") ne "yes") {
James Zern's avatar
James Zern committed
402 403
    specialize qw/av1_fwd_txfm2d_16x16 sse4_1/;
  }
404
  add_proto qw/void av1_fwd_txfm2d_32x32/, "const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd";
405
  if (aom_config("CONFIG_DAALA_TX32") ne "yes") {
James Zern's avatar
James Zern committed
406 407
    specialize qw/av1_fwd_txfm2d_32x32 sse4_1/;
  }
408 409

  if (aom_config("CONFIG_TX64X64") eq "yes") {
410 411 412
    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
413 414 415 416 417 418 419 420
  }
  #
  # 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
421

James Zern's avatar
James Zern committed
422
  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
423

James Zern's avatar
James Zern committed
424
  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
425

James Zern's avatar
James Zern committed
426 427
  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
428

James Zern's avatar
James Zern committed
429 430 431 432 433
  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";
  }
434

435 436 437 438
  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
439
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
440

James Zern's avatar
James Zern committed
441 442 443
    # 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
444

James Zern's avatar
James Zern committed
445
      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
446

James Zern's avatar
James Zern committed
447
      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
448

James Zern's avatar
James Zern committed
449
      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";
450

James Zern's avatar
James Zern committed
451 452
      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";
453

James Zern's avatar
James Zern committed
454 455
        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";
      }
456
    }
Yaowu Xu's avatar
Yaowu Xu committed
457

James Zern's avatar
James Zern committed
458 459
    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
460

James Zern's avatar
James Zern committed
461
    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";
462

James Zern's avatar
James Zern committed
463
  }
464

James Zern's avatar
James Zern committed
465 466
  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/;
467

James Zern's avatar
James Zern committed
468
  add_proto qw/void av1_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
469

James Zern's avatar
James Zern committed
470
  # End av1_high encoder functions
Yaowu Xu's avatar
Yaowu Xu committed
471

472 473 474 475 476 477
  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
478 479 480

}
# end encoder functions
481 482 483

# Deringing Functions

484
if (aom_config("CONFIG_CDEF") eq "yes") {
Steinar Midtskogen's avatar
Steinar Midtskogen committed
485
  add_proto qw/int cdef_find_dir/, "const uint16_t *img, int stride, int32_t *var, int coeff_shift";
486 487 488 489 490 491 492 493 494 495 496 497
  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 {
498
    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, int coeff_shift";
499
  }
500

Michael Bebenita's avatar
Michael Bebenita committed
501 502
  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";
503

James Zern's avatar
James Zern committed
504
  # VS compiling for 32 bit targets does not support vector types in
505 506 507
  # 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.*/) {
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529
    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/;
    }
530
  }
531 532
}

533 534
# WARPED_MOTION / GLOBAL_MOTION functions

535 536
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/;
537

538 539 540
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/;
541 542
}

543
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
544 545 546 547
  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
548 549 550
# LOOP_RESTORATION functions

if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
551
  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
552
  specialize qw/apply_selfguided_restoration sse4_1/;
553

554
  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";
555 556
  specialize qw/av1_selfguided_restoration sse4_1/;

557
  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";
558
  specialize qw/av1_highpass_filter sse4_1/;
559

560
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
561
    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";
562
    specialize qw/apply_selfguided_restoration_highbd sse4_1/;
563

564
    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";
565
    specialize qw/av1_selfguided_restoration_highbd sse4_1/;
566

567
    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";
568
    specialize qw/av1_highpass_filter_highbd sse4_1/;
569
  }
David Barker's avatar
David Barker committed
570 571
}

David Barker's avatar
David Barker committed
572 573 574
# CONVOLVE_ROUND/COMPOUND_ROUND functions

if (aom_config("CONFIG_CONVOLVE_ROUND") eq "yes") {
James Zern's avatar
James Zern committed
575 576 577 578
  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
579

James Zern's avatar
James Zern committed
580
  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";
581 582 583
  if (aom_config("CONFIG_COMPOUND_ROUND") ne "yes") {
    specialize qw/av1_convolve_2d_scale sse4_1/;
  }
584

David Barker's avatar
David Barker committed
585 586 587
  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/;
588 589
    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/;
590

591
    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";
592 593 594
    if (aom_config("CONFIG_COMPOUND_ROUND") ne "yes") {
        specialize qw/av1_highbd_convolve_2d_scale sse4_1/;
    }
David Barker's avatar
David Barker committed
595 596 597
  }
}

598 599 600 601
# 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
602 603
  add_proto qw/void av1_upsample_intra_edge/, "uint8_t *p, int sz";
  specialize qw/av1_upsample_intra_edge sse4_1/;
604 605 606
  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
607 608
    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/;
609 610 611
  }
}

612
1;