vp9_rtcd_defs.pl 8.57 KB
Newer Older
James Zern's avatar
James Zern committed
1 2 3 4 5 6 7
sub vp9_common_forward_decls() {
print <<EOF
/*
 * VP9
 */

#include "vpx/vpx_integer.h"
8
#include "vp9/common/vp9_common.h"
James Zern's avatar
James Zern committed
9
#include "vp9/common/vp9_enums.h"
10
#include "vp9/common/vp9_filter.h"
James Zern's avatar
James Zern committed
11 12 13 14 15 16

struct macroblockd;

/* Encoder forward decls */
struct macroblock;
struct vp9_variance_vtable;
17
struct search_site_config;
James Zern's avatar
James Zern committed
18 19 20 21 22 23 24
struct mv;
union int_mv;
struct yv12_buffer_config;
EOF
}
forward_decls qw/vp9_common_forward_decls/;

Johann's avatar
Johann committed
25 26
# functions that are 64 bit only.
$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
James Zern's avatar
James Zern committed
27 28 29 30 31 32
if ($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';
33
  $avx512_x86_64 = 'avx512';
James Zern's avatar
James Zern committed
34 35 36 37 38 39
}

#
# post proc
#
if (vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
JackyChen's avatar
JackyChen committed
40
add_proto qw/void vp9_filter_by_weight16x16/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
41
specialize qw/vp9_filter_by_weight16x16 sse2 msa/;
JackyChen's avatar
JackyChen committed
42 43

add_proto qw/void vp9_filter_by_weight8x8/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
44
specialize qw/vp9_filter_by_weight8x8 sse2 msa/;
James Zern's avatar
James Zern committed
45 46 47 48 49
}

#
# dct
#
Linfeng Zhang's avatar
Linfeng Zhang committed
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
# Force C versions if CONFIG_EMULATE_HARDWARE is 1
add_proto qw/void vp9_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride, int tx_type";

add_proto qw/void vp9_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride, int tx_type";

add_proto qw/void vp9_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";

if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
  # Note that there are more specializations appended when
  # CONFIG_VP9_HIGHBITDEPTH is off.
  specialize qw/vp9_iht4x4_16_add sse2/;
  specialize qw/vp9_iht8x8_64_add sse2/;
  specialize qw/vp9_iht16x16_256_add sse2/;
  if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") ne "yes") {
    # Note that these specializations are appended to the above ones.
    specialize qw/vp9_iht4x4_16_add neon dspr2 msa/;
    specialize qw/vp9_iht8x8_64_add neon dspr2 msa/;
    specialize qw/vp9_iht16x16_256_add dspr2 msa/;
68
  }
69
}
James Zern's avatar
James Zern committed
70

71 72
# High bitdepth functions
if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
73 74 75 76 77 78 79 80 81 82 83
  #
  # post proc
  #
  if (vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
    add_proto qw/void vp9_highbd_mbpost_proc_down/, "uint16_t *dst, int pitch, int rows, int cols, int flimit";

    add_proto qw/void vp9_highbd_mbpost_proc_across_ip/, "uint16_t *src, int pitch, int rows, int cols, int flimit";

    add_proto qw/void vp9_highbd_post_proc_down_and_across/, "const uint16_t *src_ptr, uint16_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit";
  }

84 85 86
  #
  # dct
  #
87 88
  # 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.
89
  add_proto qw/void vp9_highbd_iht4x4_16_add/, "const tran_low_t *input, uint16_t *dest, int stride, int tx_type, int bd";
90

91
  add_proto qw/void vp9_highbd_iht8x8_64_add/, "const tran_low_t *input, uint16_t *dest, int stride, int tx_type, int bd";
92

93
  add_proto qw/void vp9_highbd_iht16x16_256_add/, "const tran_low_t *input, uint16_t *output, int pitch, int tx_type, int bd";
94
}
James Zern's avatar
James Zern committed
95 96 97 98 99 100 101 102

#
# Encoder functions below this point.
#
if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {

# ENCODEMB INVOKE

103 104 105 106 107
#
# Denoiser
#
if (vpx_config("CONFIG_VP9_TEMPORAL_DENOISING") eq "yes") {
  add_proto qw/int vp9_denoiser_filter/, "const uint8_t *sig, int sig_stride, const uint8_t *mc_avg, int mc_avg_stride, uint8_t *avg, int avg_stride, int increase_denoising, BLOCK_SIZE bs, int motion_magnitude";
108
  specialize qw/vp9_denoiser_filter neon sse2/;
109 110
}

Linfeng Zhang's avatar
Linfeng Zhang committed
111
add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
Jingning Han's avatar
Jingning Han committed
112

Linfeng Zhang's avatar
Linfeng Zhang committed
113
add_proto qw/int64_t vp9_block_error_fp/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, int block_size";
114

Linfeng Zhang's avatar
Linfeng Zhang committed
115 116 117 118 119
add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_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/vp9_quantize_fp neon sse2/, "$ssse3_x86_64";

add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_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/vp9_quantize_fp_32x32 neon/, "$ssse3_x86_64";
120

Linfeng Zhang's avatar
Linfeng Zhang committed
121 122 123 124
add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_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";

if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
  specialize qw/vp9_block_error avx2 sse2/;
125

Linfeng Zhang's avatar
Linfeng Zhang committed
126
  specialize qw/vp9_block_error_fp sse2/;
James Zern's avatar
James Zern committed
127

128
  specialize qw/vp9_fdct8x8_quant neon ssse3/;
Linfeng Zhang's avatar
Linfeng Zhang committed
129 130 131

  add_proto qw/int64_t vp9_highbd_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd";
  specialize qw/vp9_highbd_block_error sse2/;
132
} else {
Johann's avatar
Johann committed
133
  specialize qw/vp9_block_error avx2 msa sse2/;
134

Johann's avatar
Johann committed
135
  specialize qw/vp9_block_error_fp neon sse2/;
136

137
  specialize qw/vp9_fdct8x8_quant sse2 ssse3 neon/;
138
}
James Zern's avatar
James Zern committed
139 140 141

# fdct functions

Linfeng Zhang's avatar
Linfeng Zhang committed
142
add_proto qw/void vp9_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
143

Linfeng Zhang's avatar
Linfeng Zhang committed
144
add_proto qw/void vp9_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
145

Linfeng Zhang's avatar
Linfeng Zhang committed
146
add_proto qw/void vp9_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
147

Linfeng Zhang's avatar
Linfeng Zhang committed
148
add_proto qw/void vp9_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
149

Linfeng Zhang's avatar
Linfeng Zhang committed
150 151 152 153 154 155 156 157 158 159 160 161
# Note that there are more specializations appended when CONFIG_VP9_HIGHBITDEPTH
# is off.
specialize qw/vp9_fht4x4 sse2/;
specialize qw/vp9_fht8x8 sse2/;
specialize qw/vp9_fht16x16 sse2/;
specialize qw/vp9_fwht4x4 sse2/;
if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") ne "yes") {
  # Note that these specializations are appended to the above ones.
  specialize qw/vp9_fht4x4 msa/;
  specialize qw/vp9_fht8x8 msa/;
  specialize qw/vp9_fht16x16 msa/;
  specialize qw/vp9_fwht4x4 msa/;
162
}
James Zern's avatar
James Zern committed
163 164 165 166

#
# Motion search
#
167
add_proto qw/int vp9_diamond_search_sad/, "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 vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
168
specialize qw/vp9_diamond_search_sad avx/;
James Zern's avatar
James Zern committed
169

170
if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") {
Johann's avatar
Johann committed
171
add_proto qw/void vp9_temporal_filter_apply/, "const uint8_t *frame1, unsigned int stride, const uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, uint32_t *accumulator, uint16_t *count";
172
specialize qw/vp9_temporal_filter_apply sse4_1/;
173
}
James Zern's avatar
James Zern committed
174

175 176
if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {

177 178
  # ENCODEMB INVOKE

Johann's avatar
Johann committed
179
  add_proto qw/void vp9_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_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";
180

Johann's avatar
Johann committed
181
  add_proto qw/void vp9_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *round_ptr, const int16_t *quant_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" ;
182 183

  # fdct functions
184
  add_proto qw/void vp9_highbd_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
185

186
  add_proto qw/void vp9_highbd_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
187

188
  add_proto qw/void vp9_highbd_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
189

190
  add_proto qw/void vp9_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
191

Johann's avatar
Johann committed
192
  add_proto qw/void vp9_highbd_temporal_filter_apply/, "const uint8_t *frame1, unsigned int stride, const uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, uint32_t *accumulator, uint16_t *count";
193

194
}
195
# End vp9_high encoder functions
196

197 198 199
#
# frame based scale
#
200
add_proto qw/void vp9_scale_and_extend_frame/, "const struct yv12_buffer_config *src, struct yv12_buffer_config *dst, INTERP_FILTER filter_type, int phase_scaler";
201
specialize qw/vp9_scale_and_extend_frame neon ssse3/;
202

James Zern's avatar
James Zern committed
203 204 205
}
# end encoder functions
1;