av1_rtcd_defs.pl 47.5 KB
Newer Older
Yaowu Xu's avatar
Yaowu Xu committed
1
2
3
4
5
6
7
8
9
10
11
sub av1_common_forward_decls() {
print <<EOF
/*
 * AV1
 */

#include "aom/aom_integer.h"
#include "av1/common/common.h"
#include "av1/common/enums.h"
#include "av1/common/quant_common.h"
#include "av1/common/filter.h"
12
#include "av1/common/convolve.h"
Yaowu Xu's avatar
Yaowu Xu committed
13
#include "av1/common/av1_txfm.h"
14
#include "av1/common/odintrin.h"
Yaowu Xu's avatar
Yaowu Xu committed
15
16
17
18
19
20
21
22
23
24

struct macroblockd;

/* Encoder forward decls */
struct macroblock;
struct aom_variance_vtable;
struct search_site_config;
struct mv;
union int_mv;
struct yv12_buffer_config;
25
typedef uint16_t od_dering_in;
Yaowu Xu's avatar
Yaowu Xu committed
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
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
43
44
add_proto qw/void av1_lowbd_convolve_init/, "void";
specialize qw/av1_lowbd_convolve_init ssse3/;
Angie Chiang's avatar
Angie Chiang committed
45

46
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
47
48
specialize qw/av1_convolve_horiz ssse3/;

49
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
50
51
52
specialize qw/av1_convolve_vert ssse3/;

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

#
62
# Inverse dct
Yaowu Xu's avatar
Yaowu Xu committed
63
64
65
66
67
68
69
#
if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
  # Note as optimized versions of these functions are added we need to add a check to ensure
  # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
  if (aom_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

70
    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
71

72
    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
73

74
    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
75

76
    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
77

78
    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
79

80
    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
81

82
83
84
85
86
87
88
89
    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

Yaowu Xu's avatar
Yaowu Xu committed
90
91
92
93
94
    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
  } else {
    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
95
      specialize qw/av1_iht4x4_16_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
96

97
    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
98
      specialize qw/av1_iht4x8_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
99

100
    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
101
      specialize qw/av1_iht8x4_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
102

103
    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
104
      specialize qw/av1_iht8x16_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
105

106
    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
107
      specialize qw/av1_iht16x8_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
108

109
    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
110
      specialize qw/av1_iht16x32_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
111

112
    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
113
      specialize qw/av1_iht32x16_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
114

115
116
117
118
119
120
121
122
    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

Yaowu Xu's avatar
Yaowu Xu committed
123
    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
124
      specialize qw/av1_iht8x8_64_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
125
126

    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
127
      specialize qw/av1_iht16x16_256_add sse2 avx2/;
128
129

    add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
130
131
132
133
134
135
  }
} else {
  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
  if (aom_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

136
    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
137

138
    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
139

140
    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
141

142
    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
143

144
    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
145

146
    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
Yaowu Xu's avatar
Yaowu Xu committed
147

148
149
150
151
152
153
154
155
    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

Yaowu Xu's avatar
Yaowu Xu committed
156
157
158
    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
159
160
161

    add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";

Yaowu Xu's avatar
Yaowu Xu committed
162
163
164
165
  } else {
    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
    specialize qw/av1_iht4x4_16_add sse2 neon dspr2/;

166
    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
167
      specialize qw/av1_iht4x8_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
168

169
    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
170
      specialize qw/av1_iht8x4_32_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
171

172
    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
173
      specialize qw/av1_iht8x16_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
174

175
    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
176
      specialize qw/av1_iht16x8_128_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
177

178
    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
179
      specialize qw/av1_iht16x32_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
180

181
    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
182
      specialize qw/av1_iht32x16_512_add sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
183

184
185
186
187
188
189
190
191
    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";

Yaowu Xu's avatar
Yaowu Xu committed
192
    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
193
      specialize qw/av1_iht8x8_64_add sse2 neon dspr2/;
Yaowu Xu's avatar
Yaowu Xu committed
194
195

    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
196
      specialize qw/av1_iht16x16_256_add sse2 avx2 dspr2/;
Yaowu Xu's avatar
Yaowu Xu committed
197

198
199
    add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";

Yaowu Xu's avatar
Yaowu Xu committed
200
201
202
203
204
205
206
    if (aom_config("CONFIG_EXT_TX") ne "yes") {
      specialize qw/av1_iht4x4_16_add msa/;
      specialize qw/av1_iht8x8_64_add msa/;
      specialize qw/av1_iht16x16_256_add msa/;
    }
  }
}
207

208
209
210
211
212
add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";

if (aom_config("CONFIG_TX64X64") eq "yes") {
  add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
}
Yaowu Xu's avatar
Yaowu Xu committed
213
214
215
216
217
218
219
220
221

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";
222
223
224
225
226
227

  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
228
229
}

230
231
# FILTER_INTRA predictor functions
if (aom_config("CONFIG_FILTER_INTRA") eq "yes") {
Yaowu Xu's avatar
Yaowu Xu committed
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
  add_proto qw/void av1_dc_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_v_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_h_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d45_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d135_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d117_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d153_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d207_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_d63_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  add_proto qw/void av1_tm_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
  # High bitdepth functions
  if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
    add_proto qw/void av1_highbd_dc_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_v_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_h_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d45_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d135_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d117_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d153_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d207_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_d63_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
    add_proto qw/void av1_highbd_tm_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
  }
}

# High bitdepth functions
if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
  #
  # 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
  #
  # Note as optimized versions of these functions are added we need to add a check to ensure
  # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
  add_proto qw/void av1_highbd_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";

291
  add_proto qw/void av1_highbd_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
Yaowu Xu's avatar
Yaowu Xu committed
292

293
  add_proto qw/void av1_highbd_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
Yaowu Xu's avatar
Yaowu Xu committed
294

295
  add_proto qw/void av1_highbd_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
Yaowu Xu's avatar
Yaowu Xu committed
296

297
  add_proto qw/void av1_highbd_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
Yaowu Xu's avatar
Yaowu Xu committed
298

299
  add_proto qw/void av1_highbd_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
Yaowu Xu's avatar
Yaowu Xu committed
300

301
  add_proto qw/void av1_highbd_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
Yaowu Xu's avatar
Yaowu Xu committed
302

303
304
305
306
307
308
309
310
  add_proto qw/void av1_highbd_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";

  add_proto qw/void av1_highbd_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";

  add_proto qw/void av1_highbd_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";

  add_proto qw/void av1_highbd_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";

Yaowu Xu's avatar
Yaowu Xu committed
311
312
313
314
315
  add_proto qw/void av1_highbd_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";

  add_proto qw/void av1_highbd_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type, int bd";
}

316
317
318
319
320
321
322
323
324
325
326
327
328
if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
  #inv txfm
  add_proto qw/void av1_inv_txfm2d_add_4x4/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_inv_txfm2d_add_4x4 sse4_1/;
  add_proto qw/void av1_inv_txfm2d_add_8x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_inv_txfm2d_add_8x8 sse4_1/;
  add_proto qw/void av1_inv_txfm2d_add_16x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_inv_txfm2d_add_16x16 sse4_1/;
  add_proto qw/void av1_inv_txfm2d_add_32x32/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_inv_txfm2d_add_32x32 avx2/;
  add_proto qw/void av1_inv_txfm2d_add_64x64/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
}

Yaowu Xu's avatar
Yaowu Xu committed
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
#
# Encoder functions below this point.
#
if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {

# ENCODEMB INVOKE

if (aom_config("CONFIG_AOM_QM") eq "yes") {
  if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
    # 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";

    add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";

    add_proto qw/void av1_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
345
346
347
348

    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
    }
Yaowu Xu's avatar
Yaowu Xu committed
349
350
351
352
353
354
355
356
357
358
359
360

    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
  } else {
    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
    specialize qw/av1_block_error avx2 msa/, "$sse2_x86inc";

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

    add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";

    add_proto qw/void av1_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
361
362
363
364

    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
    }
Yaowu Xu's avatar
Yaowu Xu committed
365
366
367
368
369
370
371
372
373
374
375
376
377

    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
  }
} else {
  if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
    # 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";

    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";

    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";

378
379
380
381
    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
    }

Yaowu Xu's avatar
Yaowu Xu committed
382
383
384
385
386
387
388
389
390
391
392
393
394
395
    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
  } else {
    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
    specialize qw/av1_block_error sse2 avx2 msa/;

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

    add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
    specialize qw/av1_quantize_fp neon sse2/, "$ssse3_x86_64";

    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";
    specialize qw/av1_quantize_fp_32x32/, "$ssse3_x86_64";

396
397
398
399
    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
    }

Yaowu Xu's avatar
Yaowu Xu committed
400
401
402
403
404
405
406
407
    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
    specialize qw/av1_fdct8x8_quant sse2 ssse3 neon/;
  }

}

# fdct functions

408
409
410
411
412
413
414
415
416
add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht4x4 sse2/;

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

add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht8x8 sse2/;

add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
417
specialize qw/av1_fht16x16 sse2 avx2/;
418

419
add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
Yi Luo's avatar
Yi Luo committed
420
specialize qw/av1_fht32x32 sse2 avx2/;
421

422
423
424
425
if (aom_config("CONFIG_TX64X64") eq "yes") {
  add_proto qw/void av1_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
}

426
427
add_proto qw/void av1_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht4x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
428

429
430
add_proto qw/void av1_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht8x4 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
431

432
433
add_proto qw/void av1_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht8x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
434

435
436
add_proto qw/void av1_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht16x8 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
437

438
439
add_proto qw/void av1_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht16x32 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
440

441
442
add_proto qw/void av1_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
specialize qw/av1_fht32x16 sse2/;
Yaowu Xu's avatar
Yaowu Xu committed
443

444
445
446
447
448
449
450
451
add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

452
453
454
455
456
457
458
459
if (aom_config("CONFIG_AOM_HIGHBITDEPTH") ne "yes") {
  if (aom_config("CONFIG_EXT_TX") ne "yes") {
    specialize qw/av1_fht4x4 msa/;
    specialize qw/av1_fht8x8 msa/;
    specialize qw/av1_fht16x16 msa/;
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
add_proto qw/void av1_fwd_idtx/, "const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type";

if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
  #fwd txfm
  add_proto qw/void av1_fwd_txfm2d_4x4/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_fwd_txfm2d_4x4 sse4_1/;
  add_proto qw/void av1_fwd_txfm2d_8x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_fwd_txfm2d_8x8 sse4_1/;
  add_proto qw/void av1_fwd_txfm2d_16x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_fwd_txfm2d_16x16 sse4_1/;
  add_proto qw/void av1_fwd_txfm2d_32x32/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_fwd_txfm2d_32x32 sse4_1/;
  add_proto qw/void av1_fwd_txfm2d_64x64/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
  specialize qw/av1_fwd_txfm2d_64x64 sse4_1/;
}

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

add_proto qw/int av1_diamond_search_sad/, "struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv";

add_proto qw/int av1_full_range_search/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv";

add_proto qw/void av1_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
specialize qw/av1_temporal_filter_apply sse2 msa/;

491
492
493
494
495
496
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";
}

Yaowu Xu's avatar
Yaowu Xu committed
497
498
499
500
501
502
503
504
505
506
507
if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {

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

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

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

    add_proto qw/void highbd_quantize_32x32_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
508
509
510
511
512
513

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

      add_proto qw/void highbd_quantize_64x64_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
    }
Yaowu Xu's avatar
Yaowu Xu committed
514
515
516
517
518
519
  }

  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/;

  if (aom_config("CONFIG_AOM_QM") eq "yes") {
520
    add_proto qw/void av1_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
Yaowu Xu's avatar
Yaowu Xu committed
521

522
523
    add_proto qw/void av1_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";

524
525
526
527
    if (aom_config("CONFIG_TX64X64") eq "yes") {
      add_proto qw/void av1_highbd_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
    }

528
    add_proto qw/void av1_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
Yaowu Xu's avatar
Yaowu Xu committed
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
  } else {
    add_proto qw/void av1_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale";
    specialize qw/av1_highbd_quantize_fp sse4_1/;

    add_proto qw/void av1_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale";
  }

  # fdct functions
  add_proto qw/void av1_highbd_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  specialize qw/av1_highbd_fht4x4 sse4_1/;

  add_proto qw/void av1_highbd_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

552
553
554
555
556
557
558
559
  add_proto qw/void av1_highbd_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

Yaowu Xu's avatar
Yaowu Xu committed
560
561
562
563
564
565
  add_proto qw/void av1_highbd_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

  add_proto qw/void av1_highbd_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";

566
567
568
569
  if (aom_config("CONFIG_TX64X64") eq "yes") {
    add_proto qw/void av1_highbd_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  }

Yaowu Xu's avatar
Yaowu Xu committed
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
  add_proto qw/void av1_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";

  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";

}
# End av1_high encoder functions

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

}
# end encoder functions
588

589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
# If PVQ is enabled, fwd transforms are required by decoder
if (aom_config("CONFIG_PVQ") eq "yes") {
# fdct functions

if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  specialize qw/av1_fht4x4 sse2/;

  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  specialize qw/av1_fht8x8 sse2/;

  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  specialize qw/av1_fht16x16 sse2/;

  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
  specialize qw/av1_fwht4x4 sse2/;
} else {
  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  specialize qw/av1_fht4x4 sse2 msa/;

  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  specialize qw/av1_fht8x8 sse2 msa/;

  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
  specialize qw/av1_fht16x16 sse2 msa/;

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

}

621
622
# Deringing Functions

623
if (aom_config("CONFIG_CDEF") eq "yes") {
624
625
626
627
  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";
628
  add_proto qw/int od_dir_find8/, "const od_dering_in *img, int stride, int32_t *var, int coeff_shift";
629
630
  add_proto qw/int od_filter_dering_direction_4x4/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir";
  add_proto qw/int od_filter_dering_direction_8x8/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir";
631
632
633
634
635

  add_proto qw/void copy_8x8_16bit_to_8bit/, "uint8_t *dst, int dstride, const uint16_t *src, int sstride";
  add_proto qw/void copy_4x4_16bit_to_8bit/, "uint8_t *dst, int dstride, const uint16_t *src, int sstride";
  add_proto qw/void copy_8x8_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride";
  add_proto qw/void copy_4x4_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride";
Michael Bebenita's avatar
Michael Bebenita committed
636
637
  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";
638
639

# VS compiling for 32 bit targets does not support vector types in
640
641
642
643
644
645
646
647
648
649
  # structs as arguments, which makes the v256 type of the intrinsics
  # hard to support, so optimizations for this target are disabled.
  if ($opts{config} !~ /libs-x86-win32-vs.*/) {
    specialize qw/aom_clpf_block_hbd sse2 ssse3 sse4_1 neon/;
    specialize qw/aom_clpf_hblock_hbd sse2 ssse3 sse4_1 neon/;
    specialize qw/aom_clpf_block sse2 ssse3 sse4_1 neon/;
    specialize qw/aom_clpf_hblock sse2 ssse3 sse4_1 neon/;
    specialize qw/od_dir_find8 sse2 ssse3 sse4_1 neon/;
    specialize qw/od_filter_dering_direction_4x4 sse2 ssse3 sse4_1 neon/;
    specialize qw/od_filter_dering_direction_8x8 sse2 ssse3 sse4_1 neon/;
650
651
652
653
654

    specialize qw/copy_8x8_16bit_to_8bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_4x4_16bit_to_8bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_8x8_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_4x4_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
Michael Bebenita's avatar
Michael Bebenita committed
655
656
    specialize qw/copy_rect8_8bit_to_16bit sse2 ssse3 sse4_1 neon/;
    specialize qw/copy_rect8_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
657
  }
658
659
}

660
661
662
# PVQ Functions

if (aom_config("CONFIG_PVQ") eq "yes") {
663
  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";
664
665
666
  specialize qw/pvq_search_rdo_double sse4_1/;
}

667
668
669
670
# WARPED_MOTION / GLOBAL_MOTION functions

if ((aom_config("CONFIG_WARPED_MOTION") eq "yes") ||
    (aom_config("CONFIG_GLOBAL_MOTION") eq "yes")) {
671
672
  add_proto qw/void av1_warp_affine/, "int32_t *mat, 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, int ref_frm, int32_t alpha, int32_t beta, int32_t gamma, int32_t delta";
  specialize qw/av1_warp_affine sse2/;
673
674
}

David Barker's avatar
David Barker committed
675
676
677
# LOOP_RESTORATION functions

if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
678
  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
679
  specialize qw/apply_selfguided_restoration sse4_1/;
680

681
  add_proto qw/void av1_selfguided_restoration/, "uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps, int32_t *tmpbuf";
682
683
  specialize qw/av1_selfguided_restoration sse4_1/;

684
685
  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/;
686

687
688
  if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
    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";
689
    specialize qw/apply_selfguided_restoration_highbd sse4_1/;
690
691

    add_proto qw/void av1_selfguided_restoration_highbd/, "uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps, int32_t *tmpbuf";
692
    specialize qw/av1_selfguided_restoration_highbd sse4_1/;
693

694
695
    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/;
696
  }
David Barker's avatar
David Barker committed
697
698
}

Yaowu Xu's avatar
Yaowu Xu committed
699
1;