av1_rtcd_defs.pl 42.8 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
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
James Zern's avatar
James Zern committed
66 67
  add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
  specialize qw/av1_iht4x4_16_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
68

James Zern's avatar
James Zern committed
69 70
  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
71

James Zern's avatar
James Zern committed
72 73
  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
74

James Zern's avatar
James Zern committed
75 76
  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
77

James Zern's avatar
James Zern committed
78 79
  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
80

James Zern's avatar
James Zern committed
81 82
  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
83

James Zern's avatar
James Zern committed
84 85
  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
86

James Zern's avatar
James Zern committed
87
  add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
88

James Zern's avatar
James Zern committed
89
  add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
90

James Zern's avatar
James Zern committed
91
  add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
92

James Zern's avatar
James Zern committed
93
  add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
94

James Zern's avatar
James Zern committed
95 96
  add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
  specialize qw/av1_iht8x8_64_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
97

James Zern's avatar
James Zern committed
98 99
  add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
  specialize qw/av1_iht16x16_256_add sse2 avx2/;
100

James Zern's avatar
James Zern committed
101
  add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
Yaowu Xu's avatar
Yaowu Xu committed
102
} else {
James Zern's avatar
James Zern committed
103 104 105 106
  add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
  if (aom_config("CONFIG_DAALA_DCT4") ne "yes") {
    specialize qw/av1_iht4x4_16_add sse2 neon/;
  }
Yaowu Xu's avatar
Yaowu Xu committed
107

James Zern's avatar
James Zern committed
108 109
  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
110

James Zern's avatar
James Zern committed
111 112
  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
113

James Zern's avatar
James Zern committed
114 115
  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
116

James Zern's avatar
James Zern committed
117 118
  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
119

James Zern's avatar
James Zern committed
120 121
  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
122

James Zern's avatar
James Zern committed
123 124
  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
125

James Zern's avatar
James Zern committed
126
  add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
127

James Zern's avatar
James Zern committed
128
  add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
129

James Zern's avatar
James Zern committed
130
  add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
131

James Zern's avatar
James Zern committed
132
  add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param";
133

James Zern's avatar
James Zern committed
134 135 136 137
  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_DAALA_DCT8") ne "yes") {
    specialize qw/av1_iht8x8_64_add sse2 neon/;
  }
Yaowu Xu's avatar
Yaowu Xu committed
138

James Zern's avatar
James Zern committed
139 140 141 142
  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_DAALA_DCT16") ne "yes") {
    specialize qw/av1_iht16x16_256_add sse2 avx2/;
  }
Yaowu Xu's avatar
Yaowu Xu committed
143

James Zern's avatar
James Zern committed
144
  add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
145

James Zern's avatar
James Zern committed
146 147 148 149 150 151 152 153 154
  if (aom_config("CONFIG_EXT_TX") ne "yes") {
    if (aom_config("CONFIG_DAALA_DCT4") ne "yes") {
      specialize qw/av1_iht4x4_16_add msa/;
    }
    if (aom_config("CONFIG_DAALA_DCT8") ne "yes") {
      specialize qw/av1_iht8x8_64_add msa/;
    }
    if (aom_config("CONFIG_DAALA_DCT16") ne "yes") {
      specialize qw/av1_iht16x16_256_add msa/;
Yaowu Xu's avatar
Yaowu Xu committed
155 156 157
    }
  }
}
158

159
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
160 161

if (aom_config("CONFIG_TX64X64") eq "yes") {
162
  add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param";
163
}
Yaowu Xu's avatar
Yaowu Xu committed
164 165 166 167 168 169 170 171 172

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";
173 174 175 176 177 178

  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
179 180
}

181 182
# FILTER_INTRA predictor functions
if (aom_config("CONFIG_FILTER_INTRA") eq "yes") {
183 184 185 186 187 188 189 190 191 192
  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_d45_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_d135_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";
  add_proto qw/void av1_d207_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_d63_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_tm_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
193
  # High bitdepth functions
194
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
195 196 197 198 199 200 201 202 203 204
    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_d45_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_d135_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";
    add_proto qw/void av1_highbd_d207_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_d63_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_tm_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
205 206 207 208
  }
}

# High bitdepth functions
209
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
  #
  # 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
  #
238
  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
239

240
  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
241

242
  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
243

244
  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
245

246
  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
247

248
  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
249

250
  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
251

252
  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";
253

254
  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";
255

256
  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";
257

258
  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";
259

260
  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
261

262
  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
263 264
}

265 266 267 268 269 270 271 272
#inv txfm
add_proto qw/void av1_inv_txfm2d_add_4x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_8x4/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_8x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_16x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_16x32/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_32x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
add_proto qw/void av1_inv_txfm2d_add_4x4/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
273 274 275
if (aom_config("CONFIG_DAALA_DCT4") ne "yes") {
  specialize qw/av1_inv_txfm2d_add_4x4 sse4_1/;
}
276
add_proto qw/void av1_inv_txfm2d_add_8x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
277 278 279
if (aom_config("CONFIG_DAALA_DCT8") ne "yes") {
  specialize qw/av1_inv_txfm2d_add_8x8 sse4_1/;
}
280
add_proto qw/void av1_inv_txfm2d_add_16x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
281 282 283
if (aom_config("CONFIG_DAALA_DCT16") ne "yes") {
  specialize qw/av1_inv_txfm2d_add_16x16 sse4_1/;
}
284
add_proto qw/void av1_inv_txfm2d_add_32x32/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
285 286 287
if (aom_config("CONFIG_DAALA_DCT32") ne "yes") {
  specialize qw/av1_inv_txfm2d_add_32x32 avx2/;
}
288
add_proto qw/void av1_inv_txfm2d_add_64x64/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
289

Yaowu Xu's avatar
Yaowu Xu committed
290 291 292 293 294
#
# Encoder functions below this point.
#
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {

James Zern's avatar
James Zern committed
295
  # ENCODEMB INVOKE
Yaowu Xu's avatar
Yaowu Xu committed
296

James Zern's avatar
James Zern committed
297
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
298 299 300
    # 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";
301
    specialize qw/av1_block_error avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
302 303

    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
304
    specialize qw/av1_quantize_fp sse2 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
305 306

    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
307
    specialize qw/av1_quantize_fp_32x32 avx2/;
Yaowu Xu's avatar
Yaowu Xu committed
308

309 310 311
    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
312
  } else {
Yaowu Xu's avatar
Yaowu Xu committed
313 314 315 316 317 318 319
    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
320
    specialize qw/av1_quantize_fp neon sse2 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
321 322

    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
323
    specialize qw/av1_quantize_fp_32x32 avx2/, "$ssse3_x86_64";
Yaowu Xu's avatar
Yaowu Xu committed
324

325 326 327 328
    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
329
  }
Yaowu Xu's avatar
Yaowu Xu committed
330

James Zern's avatar
James Zern committed
331
  # fdct functions
Yaowu Xu's avatar
Yaowu Xu committed
332

James Zern's avatar
James Zern committed
333 334 335 336
  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  if (aom_config("CONFIG_DAALA_DCT4") ne "yes") {
    specialize qw/av1_fht4x4 sse2/;
  }
337

James Zern's avatar
James Zern committed
338
  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
339

James Zern's avatar
James Zern committed
340 341 342 343
  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  if (aom_config("CONFIG_DAALA_DCT8") ne "yes") {
    specialize qw/av1_fht8x8 sse2/;
  }
344

James Zern's avatar
James Zern committed
345 346 347 348
  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  if (aom_config("CONFIG_DAALA_DCT16") ne "yes") {
    specialize qw/av1_fht16x16 sse2 avx2/;
  }
349

James Zern's avatar
James Zern committed
350 351 352 353
  add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
  if (aom_config("CONFIG_DAALA_DCT32") ne "yes") {
    specialize qw/av1_fht32x32 sse2 avx2/;
  }
354

James Zern's avatar
James Zern committed
355 356 357
  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";
  }
358

James Zern's avatar
James Zern committed
359 360
  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
361

James Zern's avatar
James Zern committed
362 363
  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
364

James Zern's avatar
James Zern committed
365 366
  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
367

James Zern's avatar
James Zern committed
368 369
  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
370

James Zern's avatar
James Zern committed
371 372
  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
373

James Zern's avatar
James Zern committed
374 375
  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
376

James Zern's avatar
James Zern committed
377
  add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
378

James Zern's avatar
James Zern committed
379
  add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
380

James Zern's avatar
James Zern committed
381
  add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
382

James Zern's avatar
James Zern committed
383
  add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
384

James Zern's avatar
James Zern committed
385 386 387 388 389 390 391 392 393 394 395
  if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
    if (aom_config("CONFIG_EXT_TX") ne "yes") {
      if (aom_config("CONFIG_DAALA_DCT4") ne "yes") {
        specialize qw/av1_fht4x4 msa/;
      }
      if (aom_config("CONFIG_DAALA_DCT8") ne "yes") {
        specialize qw/av1_fht8x8 msa/;
      }
      if (aom_config("CONFIG_DAALA_DCT16") ne "yes") {
        specialize qw/av1_fht16x16 msa/;
      }
396
    }
397 398
  }

James Zern's avatar
James Zern committed
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434
  add_proto qw/void av1_fwd_idtx/, "const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type";

  #fwd txfm
  add_proto qw/void av1_fwd_txfm2d_4x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_8x4/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_8x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_16x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_16x32/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_32x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  add_proto qw/void av1_fwd_txfm2d_4x4/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  if (aom_config("CONFIG_DAALA_DCT4") ne "yes") {
    specialize qw/av1_fwd_txfm2d_4x4 sse4_1/;
  }
  add_proto qw/void av1_fwd_txfm2d_8x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  if (aom_config("CONFIG_DAALA_DCT8") ne "yes") {
    specialize qw/av1_fwd_txfm2d_8x8 sse4_1/;
  }
  add_proto qw/void av1_fwd_txfm2d_16x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  if (aom_config("CONFIG_DAALA_DCT16") ne "yes") {
    specialize qw/av1_fwd_txfm2d_16x16 sse4_1/;
  }
  add_proto qw/void av1_fwd_txfm2d_32x32/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  if (aom_config("CONFIG_DAALA_DCT32") ne "yes") {
    specialize qw/av1_fwd_txfm2d_32x32 sse4_1/;
  }
  add_proto qw/void av1_fwd_txfm2d_64x64/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  if (aom_config("CONFIG_DAALA_DCT64") ne "yes") {
    specialize qw/av1_fwd_txfm2d_64x64 sse4_1/;
  }
  #
  # 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
435

James Zern's avatar
James Zern committed
436
  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
437

James Zern's avatar
James Zern committed
438
  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
439

James Zern's avatar
James Zern committed
440 441
  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
442

James Zern's avatar
James Zern committed
443 444 445 446 447
  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";
  }
448

James Zern's avatar
James Zern committed
449
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
450

James Zern's avatar
James Zern committed
451 452 453
    # 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
454

James Zern's avatar
James Zern committed
455
      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
456

James Zern's avatar
James Zern committed
457
      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
458

James Zern's avatar
James Zern committed
459
      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";
460

James Zern's avatar
James Zern committed
461 462
      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";
463

James Zern's avatar
James Zern committed
464 465
        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";
      }
466
    }
Yaowu Xu's avatar
Yaowu Xu committed
467

James Zern's avatar
James Zern committed
468 469
    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
470

James Zern's avatar
James Zern committed
471
    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";
472

James Zern's avatar
James Zern committed
473
  }
474

James Zern's avatar
James Zern committed
475 476
  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/;
477

James Zern's avatar
James Zern committed
478
  add_proto qw/void av1_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
479

James Zern's avatar
James Zern committed
480
  # End av1_high encoder functions
Yaowu Xu's avatar
Yaowu Xu committed
481

James Zern's avatar
James Zern committed
482 483 484 485 486 487 488 489
  if (aom_config("CONFIG_EXT_INTER") eq "yes") {
    add_proto qw/uint64_t av1_wedge_sse_from_residuals/, "const int16_t *r1, const int16_t *d, const uint8_t *m, int N";
    specialize qw/av1_wedge_sse_from_residuals sse2/;
    add_proto qw/int av1_wedge_sign_from_residuals/, "const int16_t *ds, const uint8_t *m, int N, int64_t limit";
    specialize qw/av1_wedge_sign_from_residuals sse2/;
    add_proto qw/void av1_wedge_compute_delta_squares/, "int16_t *d, const int16_t *a, const int16_t *b, int N";
    specialize qw/av1_wedge_compute_delta_squares sse2/;
  }
Yaowu Xu's avatar
Yaowu Xu committed
490 491 492

}
# end encoder functions
493

494 495
# If PVQ is enabled, fwd transforms are required by decoder
if (aom_config("CONFIG_PVQ") eq "yes") {
James Zern's avatar
James Zern committed
496
  # fdct functions
497

James Zern's avatar
James Zern committed
498 499 500
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
    add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
    specialize qw/av1_fht4x4 sse2/;
501

James Zern's avatar
James Zern committed
502 503
    add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
    specialize qw/av1_fht8x8 sse2/;
504

James Zern's avatar
James Zern committed
505 506
    add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
    specialize qw/av1_fht16x16 sse2/;
507

James Zern's avatar
James Zern committed
508 509 510 511 512
    add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/av1_fwht4x4 sse2/;
  } else {
    add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
    specialize qw/av1_fht4x4 sse2 msa/;
513

James Zern's avatar
James Zern committed
514 515
    add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
    specialize qw/av1_fht8x8 sse2 msa/;
516

James Zern's avatar
James Zern committed
517 518
    add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param";
    specialize qw/av1_fht16x16 sse2 msa/;
519

James Zern's avatar
James Zern committed
520 521 522
    add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
    specialize qw/av1_fwht4x4 msa sse2/;
  }
523 524 525

}

526 527
# Deringing Functions

528
if (aom_config("CONFIG_CDEF") eq "yes") {
Steinar Midtskogen's avatar
Steinar Midtskogen committed
529
  add_proto qw/int cdef_find_dir/, "const uint16_t *img, int stride, int32_t *var, int coeff_shift";
530 531 532 533 534 535 536 537 538 539 540 541 542 543
  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";
  }
544

Michael Bebenita's avatar
Michael Bebenita committed
545 546
  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";
547

James Zern's avatar
James Zern committed
548
  # VS compiling for 32 bit targets does not support vector types in
549 550 551
  # 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.*/) {
552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573
    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/;
    }
574
  }
575 576
}

577 578 579
# PVQ Functions

if (aom_config("CONFIG_PVQ") eq "yes") {
580
  add_proto qw/double pvq_search_rdo_double/, "const od_val16 *xcoeff, int n, int k, int *ypulse, double g2, double pvq_norm_lambda, int prev_k";
581 582 583
  specialize qw/pvq_search_rdo_double sse4_1/;
}

584 585 586 587
# WARPED_MOTION / GLOBAL_MOTION functions

if ((aom_config("CONFIG_WARPED_MOTION") eq "yes") ||
    (aom_config("CONFIG_GLOBAL_MOTION") eq "yes")) {
588
  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";
589
  specialize qw/av1_warp_affine sse2 ssse3/;
590

591
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
592
    add_proto qw/void av1_highbd_warp_affine/, "const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta";
David Barker's avatar
David Barker committed
593
    specialize qw/av1_highbd_warp_affine ssse3/;
594
  }
595 596
}

597
if (aom_config("CONFIG_GLOBAL_MOTION") eq "yes" &&
598 599 600 601 602
    aom_config("CONFIG_AV1_ENCODER") eq "yes") {
  add_proto qw/double compute_cross_correlation/, "unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2";
  specialize qw/compute_cross_correlation sse4_1/;
}

David Barker's avatar
David Barker committed
603 604 605
# LOOP_RESTORATION functions

if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
606
  add_proto qw/void apply_selfguided_restoration/, "uint8_t *dat, int width, int height, int stride, int eps, int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf";
David Barker's avatar
David Barker committed
607
  specialize qw/apply_selfguided_restoration sse4_1/;
608

609
  add_proto qw/void av1_selfguided_restoration/, "uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
610 611
  specialize qw/av1_selfguided_restoration sse4_1/;

612 613
  add_proto qw/void av1_highpass_filter/, "uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
  specialize qw/av1_highpass_filter sse4_1/;
614

615
  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
616
    add_proto qw/void apply_selfguided_restoration_highbd/, "uint16_t *dat, int width, int height, int stride, int bit_depth, int eps, int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf";
617
    specialize qw/apply_selfguided_restoration_highbd sse4_1/;
618

619
    add_proto qw/void av1_selfguided_restoration_highbd/, "uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps";
620
    specialize qw/av1_selfguided_restoration_highbd sse4_1/;
621

622 623
    add_proto qw/void av1_highpass_filter_highbd/, "uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
    specialize qw/av1_highpass_filter_highbd sse4_1/;
624
  }
David Barker's avatar
David Barker committed
625 626
}

David Barker's avatar
David Barker committed
627 628 629
# CONVOLVE_ROUND/COMPOUND_ROUND functions

if (aom_config("CONFIG_CONVOLVE_ROUND") eq "yes") {
James Zern's avatar
James Zern committed
630 631 632 633
  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
634

James Zern's avatar
James Zern committed
635
  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";
636

David Barker's avatar
David Barker committed
637 638 639
  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/;
640 641
    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/;
642
    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";
David Barker's avatar
David Barker committed
643 644 645
  }
}

646 647 648 649
# 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
650 651
  add_proto qw/void av1_upsample_intra_edge/, "uint8_t *p, int sz";
  specialize qw/av1_upsample_intra_edge sse4_1/;
652 653 654
  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
655 656
    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/;
657 658 659
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
660
1;