1 /*
2 * Copyright (c) 2023, Alliance for Open Media. All rights reserved.
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <arm_neon.h>
13 #include <assert.h>
14 #include <math.h>
15 #include <stdbool.h>
16
17 #include "config/aom_config.h"
18 #include "config/av1_rtcd.h"
19
20 #include "aom_dsp/aom_dsp_common.h"
21 #include "aom_dsp/arm/sum_neon.h"
22 #include "av1/common/av1_common_int.h"
23 #include "av1/encoder/cnn.h"
24 #include "av1/encoder/partition_cnn_weights.h"
25
26 // The CNN weights used in av1_cnn_convolve_no_maxpool_padding_valid are
27 // declared (av1_intra_mode_cnn_partition_cnn_layer_[01234]_kernel) in
28 // partition_cnn_weights.h. However, to enable linear memory access, rearrange
29 // the weight tables here.
30 static const float weights_layer_1[] = {
31 0.228403f, 0.031690f, -0.251710f, -0.046230f, 0.413294f, -0.236732f,
32 -0.038291f, 0.210766f, 0.427196f, -0.384319f, -0.439463f, 0.366015f,
33 0.112263f, -0.144168f, -0.075017f, 0.119629f, 0.325200f, -0.678246f,
34 -0.370826f, -0.341362f, -0.503392f, 0.400884f, 0.465214f, -0.360847f,
35 0.187100f, -0.190757f, -0.131906f, 0.121492f, -0.303556f, -0.007658f,
36 0.380077f, -0.066394f, -0.016043f, -1.490730f, -0.120682f, 0.132062f,
37 0.086185f, -0.042766f, -0.087069f, 0.029426f, 0.309583f, -0.029985f,
38 -0.297429f, -0.018139f, -0.688828f, 0.756607f, 0.706410f, -0.696826f,
39 -0.087793f, -0.023304f, -0.012332f, -0.018043f, -0.410268f, 0.352143f,
40 0.391284f, -0.363178f, -0.295034f, 0.160246f, -0.149446f, 0.260145f,
41 -0.252249f, 0.190826f, 0.251206f, -0.270796f, -0.979219f, 0.884880f,
42 0.962057f, -0.847601f, -0.011053f, 0.118765f, -0.028428f, -0.020138f,
43 0.400274f, -0.382845f, -0.462766f, 0.390654f, 0.361223f, -0.320068f,
44 -0.372084f, 0.313196f, 0.241933f, -0.416614f, -0.008722f, -0.255078f,
45 0.078730f, -0.381935f, -0.204577f, 0.159768f, 0.071853f, -0.126294f,
46 -0.036186f, -0.007900f, 0.380071f, -0.298882f, 0.387941f, -0.267350f,
47 -0.586802f, 0.477785f, -0.000013f, 0.197296f, -0.079154f, -0.005811f,
48 -0.044300f, -0.021192f, -0.020879f, -0.005265f, 0.082277f, -0.139132f,
49 -0.239237f, 0.440234f, -0.542342f, 0.378360f, -0.070974f, 0.272702f,
50 -0.278939f, -0.044948f, -0.134197f, -0.007172f, -0.353628f, -0.128091f,
51 0.357458f, -0.037614f, -0.144983f, 0.220623f, -0.003394f, -0.070166f,
52 0.200370f, -0.166037f, 0.224448f, -0.012990f, -0.098853f, 0.008613f,
53 -0.017669f, 0.070641f, 0.174530f, -0.119822f, -0.065096f, 0.118487f,
54 -0.024764f, -0.050466f, 0.066631f, -0.075896f, -0.062363f, 0.212604f,
55 -0.377322f, 0.306306f, -0.399733f, 0.238624f, 0.233571f, -0.344080f,
56 0.462491f, -0.565210f, -0.035074f, -0.010459f, 0.084382f, 0.052294f,
57 0.065714f, 0.013716f, 0.135036f, 0.000588f, 0.181079f, -0.566344f,
58 0.395561f, -0.398509f, 0.450017f, -1.462710f, 1.138280f, -0.447774f,
59 0.247936f, -0.417067f, 0.165997f, -0.458632f, -0.018527f, 0.308461f,
60 0.541266f, 0.162257f, 0.601786f, -1.275840f, -0.373404f, -0.589747f,
61 0.026539f, -0.219327f, 0.142972f, -0.018496f, 0.075204f, -0.775190f,
62 0.237307f, -0.348252f, 0.117792f, -0.094332f, 0.363101f, -0.065025f,
63 0.816662f, 0.590110f, 0.752202f, -0.308599f, 0.258337f, -0.842085f,
64 0.695788f, -0.205615f, 0.093930f, -0.392536f, 0.463093f, -0.432456f,
65 0.041660f, -0.827264f, 0.309128f, -0.354658f, 0.451957f, -1.406640f,
66 0.773192f, -0.892943f, 0.134856f, -0.467808f, 0.306003f, -0.226560f,
67 0.086865f, -0.104102f, 0.148098f, -0.082658f, 0.316655f, -1.028310f,
68 0.741566f, -0.345326f, 0.052379f, -0.275613f, 0.191765f, -0.162391f,
69 0.000976f, 0.093061f, 0.068649f, 0.033582f, 0.239727f, -0.647769f,
70 0.218493f, -0.397120f, 0.268229f, -0.303424f, 0.185393f, -0.314189f,
71 0.101728f, -0.163083f, -0.084989f, 0.136783f, -0.264346f, 0.465914f,
72 0.220395f, -0.252968f, -0.326661f, 0.271483f, 0.374717f, -0.311570f,
73 -0.082119f, 0.020870f, 0.091975f, -0.030582f, -0.487148f, 0.198912f,
74 0.024554f, -0.749363f, -0.102267f, 0.097787f, 0.141459f, -0.110706f,
75 0.079467f, -0.082570f, -0.347567f, 0.341043f, -0.137871f, 0.112319f,
76 0.064733f, -0.082869f, 0.269999f, -0.408184f, -0.183443f, 0.180608f,
77 0.223345f, -0.357376f, -0.244593f, 0.355348f, -0.072701f, -0.034311f,
78 0.096544f, 0.016407f, 0.417550f, -0.367772f, -0.484535f, 0.405977f,
79 0.314243f, -0.099622f, -0.192218f, -0.012780f, 0.434551f, -0.399047f,
80 -0.531499f, 0.484513f, -0.691352f, 0.872823f, 1.207720f, -1.377490f,
81 0.006872f, -0.041453f, 0.007845f, 0.007463f, 0.467299f, -0.476372f,
82 -0.452606f, 0.452357f, 0.447332f, -0.365632f, -0.332435f, 0.300284f,
83 -0.290504f, 0.255410f, 0.310921f, -0.293717f, -0.616299f, 0.594207f,
84 0.461347f, -0.449439f, 0.278455f, 0.285085f, -1.201340f, -0.016463f,
85 0.549095f, 0.610375f, -4.608530f, -1.727390f, 0.150404f, -0.012846f,
86 -0.481148f, -0.182257f, 0.918796f, 0.213872f, 1.050410f, 0.681526f,
87 -0.458777f, -0.710395f, -2.347200f, -0.277197f, 0.213294f, 0.337551f,
88 -0.177710f, -0.152136f, 0.167666f, 0.308403f, -1.248500f, -0.565367f,
89 0.122054f, 0.087874f, -0.476556f, -0.083548f, -0.358734f, -0.073131f,
90 -0.146320f, -2.241960f, 0.697639f, 0.545581f, -1.889700f, -0.267725f,
91 0.433045f, 0.298224f, -0.338508f, 0.250226f, 0.405675f, 0.447201f,
92 -1.184690f, -0.473447f, 0.307403f, 0.711236f, -3.191560f, -1.663980f,
93 0.165201f, 0.101360f, -0.624451f, -0.173269f, 0.089795f, 0.227478f,
94 -0.136664f, 0.007907f, 0.131079f, 0.605374f, -2.991620f, -1.723790f,
95 0.082428f, 0.006781f, -0.348732f, -0.019271f, -0.032040f, -0.067078f,
96 -0.437166f, -0.144472f, 0.069844f, 0.194625f, -0.162284f, -0.374656f,
97 0.056472f, -0.236524f, -0.114241f, -0.029161f, -0.222078f, -0.053435f,
98 -0.313938f, -0.555472f, 1.037550f, 0.689968f, 0.575694f, 0.065826f,
99 -0.659979f, -0.881351f, -0.626417f, -0.953975f, -0.576106f, -0.258708f,
100 0.263004f, -0.229847f, 0.463835f, 1.390960f, -2.614480f, -1.272910f,
101 0.065780f, -0.058603f, 0.015612f, 0.104703f, 0.198028f, 0.262792f,
102 0.253616f, -0.079126f, -0.587381f, -0.739021f, -0.822676f, -0.795512f,
103 0.193644f, 0.234643f, -0.034407f, 0.421478f, -0.572610f, -0.290714f,
104 -0.257803f, -0.644835f, -0.536938f, -0.375899f, -0.651077f, -0.522576f,
105 0.562564f, 0.834616f, 0.513893f, 0.649689f, 0.356530f, 0.400716f,
106 0.300606f, 0.290505f, 0.584608f, 0.671574f, 0.564584f, 0.419870f,
107 0.062061f, 0.018263f, 0.009831f, 0.084103f, -0.128281f, -0.018818f,
108 -0.187244f, 0.067210f, 0.437147f, 0.442029f, 0.444939f, 0.226661f,
109 0.541609f, 0.444280f, 0.302795f, 0.633026f, -0.180374f, 0.265197f,
110 0.210404f, -0.118916f, -0.294013f, -0.692627f, -0.402347f, -0.356287f,
111 0.387578f, 0.385496f, 0.789542f, 0.690396f, -0.203542f, -0.688546f,
112 0.045319f, -0.448747f, -0.157148f, 0.152581f, 0.022360f, 0.058358f,
113 0.593007f, 1.131860f, 0.289006f, 1.015560f, 0.144942f, -0.411577f,
114 0.264794f, -0.085791f, 0.156996f, 0.200340f, 0.169264f, 0.267615f,
115 -0.361015f, -0.601842f, -0.442217f, -0.781086f, 0.112938f, 0.385305f,
116 0.482454f, 0.470268f, 1.193390f, 0.589642f, 0.127638f, -0.640946f,
117 0.540310f, 0.741498f, 0.686937f, 0.435879f, 0.534523f, 0.693119f,
118 0.817577f, 0.783109f, 0.021681f, -0.004973f, 0.201236f, -0.086311f,
119 0.028628f, 0.227871f, 0.462751f, 0.126832f, -0.389997f, -0.553965f,
120 -0.343953f, -0.448517f, 0.053129f, -0.115083f, 0.018138f, -0.067131f,
121 -0.293468f, -0.220700f, 0.074348f, -0.273153f, 0.263637f, 0.122049f,
122 0.153025f, 0.076292f, 0.142320f, 0.286734f, 0.100542f, 0.308660f,
123 -0.759591f, -0.750938f, -0.788799f, -0.853076f, -0.588019f, -0.990063f,
124 -0.692327f, -0.722904f, 0.084736f, 0.151068f, 0.159606f, 0.147715f,
125 1.610180f, 1.950330f, 1.765670f, 2.265110f, 0.008262f, 0.185584f,
126 0.039337f, 0.164721f, 0.479446f, 0.314083f, 0.043969f, 0.291320f,
127 0.003400f, -0.551190f, 0.060158f, -0.147591f, 0.089117f, 0.042994f,
128 0.042802f, 0.127392f, -0.066172f, 0.078370f, 0.051408f, 0.014004f,
129 0.086726f, 0.133334f, -0.046733f, 0.155100f, -0.118223f, -0.100778f,
130 -0.225245f, -0.460397f, 0.892644f, 1.003770f, 0.405155f, 0.517477f,
131 0.184585f, 0.279090f, -0.036477f, 0.198703f, 0.027139f, -0.055728f,
132 -0.022396f, -0.147319f, 2.275540f, 2.014990f, 2.296800f, 2.081730f,
133 -0.088713f, 0.105729f, -0.027871f, -0.095047f, 0.012429f, 0.014244f,
134 -0.014755f, -0.003017f, 1.332700f, 1.300040f, 1.464250f, 1.305030f,
135 0.032568f, 0.118042f, 0.079632f, -0.089405f, 0.163905f, 0.146608f,
136 0.026502f, 0.065307f, -0.056909f, -0.065052f, 0.069851f, -0.082958f,
137 0.023419f, -0.026293f, 0.037616f, -0.048096f, -0.073701f, -0.208295f,
138 -0.782095f, 0.000523f, 0.374131f, 0.420946f, 0.466151f, 0.349651f,
139 -0.679275f, -0.745827f, -0.379918f, -0.900107f, 0.044070f, -0.347536f,
140 -1.224390f, 0.740113f, -0.779966f, 0.510920f, -0.968597f, -0.095630f,
141 0.120805f, 0.676803f, -0.164827f, 0.172996f, -0.106720f, 0.197527f,
142 0.337561f, 0.571094f, -0.279090f, -0.396697f, -0.253083f, -0.690170f,
143 -0.363291f, 0.516921f, 0.489391f, -0.920628f, 0.497572f, 0.483864f,
144 -0.125696f, -0.338123f, -0.041517f, -0.534630f, -0.388465f, -0.784554f,
145 0.215227f, 0.055088f, 0.179638f, 0.086997f, 0.569313f, 0.572926f,
146 0.137182f, -0.045485f, 0.118087f, 0.210383f, 0.212664f, 0.482443f,
147 0.151921f, 0.307947f, -0.084656f, -0.386206f, 0.542277f, -0.207005f,
148 0.073792f, -1.013240f, 0.303581f, 0.270527f, 0.265985f, 0.332702f,
149 0.848609f, 0.686757f, 0.767212f, 0.316901f, -0.502460f, -0.567092f,
150 -0.484799f, -0.173350f, -0.426863f, 0.222375f, -0.200267f, -0.523758f,
151 0.265180f, -0.175648f, -0.229754f, 0.148740f, 0.402515f, 0.028243f,
152 -0.366109f, 0.157232f, -0.131564f, 0.055136f, 0.211046f, -0.115542f,
153 0.322379f, -0.137768f, -0.247832f, 0.070394f, 0.058530f, -0.295023f,
154 -0.196022f, -0.109097f, 0.261285f, -0.273585f, -0.240632f, 0.258326f,
155 -0.077364f, 0.071405f, -0.014766f, -0.008751f, -0.203622f, 0.177818f,
156 0.116726f, -0.116735f, -0.723616f, -0.700154f, 0.145082f, -0.184949f,
157 -0.287076f, 0.150405f, 0.258075f, -0.157764f, -0.120909f, 0.105459f,
158 0.113288f, -0.092963f, 0.328183f, -0.300115f, -0.361289f, 0.319792f,
159 -0.048875f, 0.135673f, 0.132539f, -0.162481f, 0.002109f, 0.065048f,
160 -0.135969f, 0.061558f, 1.510670f, -0.884925f, -0.827022f, 0.190311f,
161 -0.060088f, -0.033362f, 0.013354f, 0.002847f, 0.353479f, -0.462538f,
162 -0.319638f, 0.424484f, 0.199540f, -0.073843f, -0.140621f, 0.072133f,
163 -0.098662f, 0.070613f, 0.031150f, -0.021869f, -0.511253f, 0.503412f,
164 0.565963f, -0.576146f, -1.081700f, 0.047670f, 0.266687f, 0.524804f,
165 -2.361150f, 0.147823f, 0.594717f, 0.956842f, -1.048220f, 0.127083f,
166 0.079581f, 0.065419f, 0.176783f, 0.653953f, 0.260967f, 0.537892f,
167 -1.207580f, 0.245983f, -0.727067f, 0.071755f, -0.343025f, -0.173435f,
168 0.215289f, 0.268578f, -1.158560f, 0.039263f, -0.132888f, 0.217132f,
169 -0.622195f, -0.071256f, 0.317333f, 0.157614f, -1.588250f, 0.316432f,
170 -0.736720f, -0.041698f, -1.959280f, 0.083451f, 0.570584f, 0.327620f,
171 -1.262200f, -0.026738f, 0.231198f, 0.326861f, -1.644200f, -0.143833f,
172 -0.079495f, 0.493026f, -2.488090f, -0.034046f, 0.165884f, 1.074260f,
173 -1.076980f, 0.248198f, -0.017987f, 0.421900f, -0.105860f, 0.076710f,
174 0.002072f, 0.070264f, -1.734750f, 0.227145f, 0.209220f, 0.851459f,
175 -0.142369f, 0.066502f, 0.027816f, 0.044321f, -0.186591f, -0.100340f,
176 0.115580f, 0.192252f, -0.892114f, 0.209531f, -0.308243f, 0.367968f,
177 -0.721770f, 0.220224f, -0.062744f, 0.133754f, 0.040416f, 0.190428f,
178 -0.035428f, 0.162974f, 0.116427f, 0.669393f, 0.278891f, 0.856676f,
179 1.060390f, 0.936983f, 0.863355f, 0.990560f, -0.147111f, -0.217883f,
180 0.355794f, -0.186530f, -0.275614f, -0.095719f, 0.167346f, 0.359078f,
181 -0.079223f, -0.581596f, -0.213134f, -0.431123f, -0.516443f, -0.388628f,
182 -0.643821f, -0.202345f, 0.426230f, 0.516923f, 0.548131f, 0.555973f,
183 0.022286f, 0.361170f, 0.980065f, 0.648400f, -0.056813f, -0.100310f,
184 -0.439481f, -0.166454f, 0.412449f, 0.509400f, 0.316208f, 0.470293f,
185 -0.827838f, -1.078380f, -1.047040f, -1.074560f, 0.274555f, -0.316736f,
186 0.128818f, 0.228566f, -0.520967f, -0.731674f, -0.687887f, -0.536388f,
187 -0.031187f, 0.041404f, 0.047821f, 0.064397f, 0.054230f, 0.105059f,
188 -0.178671f, 0.176847f, -0.394797f, -0.260255f, -0.333734f, -0.162345f,
189 -0.444650f, -0.928438f, -0.705840f, -0.833162f, 0.306737f, 0.429699f,
190 0.417298f, 0.478469f, 0.420903f, 0.676871f, 0.429677f, 0.616921f,
191 -0.805199f, -0.643391f, -0.304100f, 0.797599f, -0.172157f, 0.429085f,
192 -0.750676f, 0.149227f, -0.207898f, -0.022534f, -0.341448f, -0.247976f,
193 0.095325f, -0.561120f, 0.599694f, -0.025236f, 0.292346f, -0.312001f,
194 0.517478f, 0.301457f, -0.106415f, 0.226263f, -0.184163f, -0.114419f,
195 -0.322702f, 0.172541f, 0.445573f, 0.157213f, 0.670704f, 0.102174f,
196 -0.234667f, -0.293311f, 0.769852f, 0.038028f, -0.036741f, -0.228060f,
197 -0.253335f, 0.424054f, -0.597980f, 0.221007f, -0.114741f, -0.411557f,
198 -0.592201f, 0.442684f, 0.115491f, -0.106896f, -0.028110f, 0.354751f,
199 -0.248375f, 0.242570f, -0.155856f, 0.280528f, -0.198742f, 0.588725f,
200 0.371065f, 0.078197f, 0.114706f, -0.448021f, 0.065255f, 0.133741f,
201 -0.227522f, -0.047339f, -0.052849f, 0.309480f, 0.597185f, 0.209182f,
202 0.226108f, -0.601036f, -0.431672f, -0.172601f, -0.000174f, 0.194292f,
203 -0.133937f, 0.130676f, 0.059372f, 0.091381f, 0.098751f, -0.150996f,
204 0.170514f, -0.085494f, 0.336576f, 0.484004f, 0.033862f, 0.277473f,
205 -0.231482f, -0.328385f, -0.332739f, -0.626957f, 0.510167f, 0.575861f,
206 0.421494f, 0.482540f, -0.636377f, -0.864661f, -0.694180f, -0.420014f,
207 -0.132781f, 0.017599f, 0.003538f, 0.486934f, 0.133878f, -0.094622f,
208 0.016132f, 0.010117f, 0.156680f, -0.022201f, -0.014621f, 0.228445f,
209 0.190826f, 0.171580f, 0.579923f, 0.245428f, 0.322713f, 0.480101f,
210 0.406320f, 0.412229f, 0.002334f, -0.022349f, 0.074571f, -0.043828f,
211 0.290453f, 0.451749f, 0.530376f, 0.271879f, 0.095144f, 0.169450f,
212 0.049482f, 0.114605f, -0.635634f, -0.700768f, -0.558538f, -0.537625f,
213 0.190255f, -0.308237f, -0.053703f, 0.212489f, 0.056520f, -0.040019f,
214 0.089822f, -0.014155f, -0.376004f, -0.448752f, -0.526717f, -0.571440f,
215 0.116482f, 0.162321f, 0.147895f, 0.280527f, 0.159037f, -0.095958f,
216 0.007931f, -0.086630f, 0.285625f, 0.514914f, 0.208908f, 0.519251f,
217 0.309368f, 0.379777f, 0.350565f, 0.487487f, -0.541494f, -0.421836f,
218 -0.390001f, -0.500696f, -0.905736f, -0.150439f, -0.942304f, -0.566771f,
219 0.484233f, 0.767417f, 0.410477f, 0.670196f, 0.070210f, 0.488836f,
220 0.372805f, 0.197631f, 0.337892f, 0.524423f, 0.777219f, -0.260955f,
221 -0.112981f, -0.060088f, -0.200250f, -0.195671f, 0.007584f, 0.252096f,
222 0.235511f, 0.366612f, -0.304979f, -0.211068f, -0.420683f, -0.085370f,
223 0.085762f, -0.097549f, -0.802509f, -0.468079f, -0.192787f, -0.069670f,
224 -0.235162f, -0.077772f, -0.441671f, -0.348479f, -0.431434f, -0.108256f,
225 -0.133779f, 0.017032f, 0.001964f, -0.120647f, -0.187663f, -0.194985f,
226 -0.231742f, -0.175288f, -0.162639f, 0.245110f, 0.049951f, 0.104229f,
227 -0.159634f, -0.076545f, -0.022496f, -0.036532f, -0.147028f, -0.034215f,
228 0.028213f, -0.059669f, -0.078259f, 0.062993f, -0.124066f, -0.137362f,
229 -0.129977f, -0.010532f, -0.049090f, -0.189401f, 0.495471f, 0.615778f,
230 0.451437f, 0.803526f, 0.523532f, 0.841339f, 0.699528f, 0.745129f,
231 0.246264f, -0.198290f, -0.283620f, 0.189917f, -0.018306f, -0.419097f,
232 0.280363f, -0.098085f, 0.138972f, -0.140867f, -0.117025f, 0.098585f,
233 0.130979f, 0.268133f, -0.161731f, -0.176629f, -0.357677f, -0.126379f,
234 0.553128f, -0.126821f, -0.001511f, -0.010081f, -0.031162f, 0.079203f,
235 -0.157731f, 0.072865f, 0.535830f, -0.529989f, -0.570075f, 0.295795f,
236 0.595613f, -0.449278f, -0.669756f, 0.941452f, 0.356897f, -0.723720f,
237 -0.115203f, -0.134479f, 0.133048f, 0.109860f, -0.024250f, -0.049732f,
238 0.020098f, 0.048356f, -0.048293f, 0.108754f, 0.062548f, -0.238315f,
239 0.182700f, 0.312011f, -0.244377f, -0.118012f, 0.012276f, 0.006089f,
240 0.098068f, -0.079280f, -0.423987f, -0.411931f, -0.027425f, 0.870280f,
241 0.022825f, -0.024481f, -0.036320f, -0.111189f, 0.364539f, -0.244896f,
242 -0.373060f, 0.266345f, -0.141778f, 0.277549f, 0.059834f, -0.178242f,
243 -0.686222f, 0.594535f, 0.354546f, -0.272516f, 1.060730f, -1.059810f,
244 -0.948126f, 0.993267f, 0.116597f, -0.227574f, -0.436144f, -0.333309f,
245 -0.575746f, -0.828102f, 0.284561f, 0.351668f, -0.080164f, -0.762518f,
246 -0.511108f, -0.212855f, 0.293892f, -0.548664f, 0.072057f, 0.006748f,
247 1.485110f, 0.124687f, 0.727211f, 1.557560f, -0.064383f, -0.022242f,
248 0.002921f, -0.151505f, 0.270926f, 0.173632f, -0.640644f, 0.422410f,
249 -0.240699f, -0.361980f, -0.279864f, -0.055165f, -1.084140f, 0.231705f,
250 0.366172f, -0.347698f, -0.097565f, -0.747227f, -0.243033f, 0.941545f,
251 -0.207460f, -0.353913f, 0.104303f, -0.403151f, 0.203177f, 0.335893f,
252 -0.229033f, 0.029096f, -0.409634f, -0.179599f, -0.442397f, 0.649114f,
253 0.460774f, 0.170906f, -0.043857f, 0.402066f, -0.226896f, -0.199624f,
254 0.016650f, 0.207894f, 0.056954f, 0.220329f, 0.374060f, 0.130361f,
255 -0.303960f, -0.078863f, 0.195410f, 0.729438f, 0.246818f, 0.287730f,
256 0.484876f, 0.111488f, -0.168647f, -0.087878f, -0.070089f, -0.341329f,
257 -0.330280f, 0.259943f, -0.364205f, 0.256555f, -0.756804f, -0.086915f,
258 0.777351f, 0.006136f, 0.110348f, 0.248743f, 0.209326f, -0.362741f,
259 -0.184416f, 0.422446f, 0.565193f, 0.310072f, -0.011212f, -0.765226f,
260 0.039466f, 0.301288f, 0.172907f, -1.539450f, 0.606202f, 0.477469f,
261 0.045894f, -0.222180f, -0.013192f, -0.064077f, -0.241551f, 0.192914f,
262 0.028004f, -0.540538f, 0.437440f, 0.179087f, -0.753204f, -0.001374f,
263 1.185930f, -0.151182f, 1.238580f, -1.389900f, 0.277954f, 0.422208f,
264 0.041553f, -0.542284f, 0.139019f, -0.148580f, -0.130705f, 0.361830f,
265 0.322953f, -0.092371f, 0.120180f, -0.355299f, -0.028057f, 0.128114f,
266 0.250947f, -0.349926f, -0.684633f, 0.246175f, 0.186731f, -0.676313f,
267 0.060535f, 0.333371f, -0.021172f, -0.421266f, -0.079650f, 0.031359f,
268 -0.303658f, -0.298286f, 0.119016f, 0.655585f, 0.200175f, -0.887182f,
269 -0.197539f, -0.318883f, -0.130250f, 0.522487f, -0.092616f, 0.405930f,
270 -0.281678f, 0.089728f, 0.081814f, -0.781745f, 0.348878f, 0.082274f,
271 -0.914136f, 1.098810f, 0.855321f, -1.078170f, -0.268018f, 0.246440f,
272 0.238347f, -0.027228f, 0.074111f, -0.061197f, -0.063582f, 0.089462f,
273 -0.040347f, 0.117082f, 0.122772f, -0.162816f, -0.148668f, -0.342856f,
274 -0.495604f, -1.453630f, -0.045273f, -0.030463f, 0.043766f, 0.047978f,
275 0.016910f, -0.009700f, 0.006288f, -0.042556f, 0.632896f, -0.845744f,
276 -0.516844f, 0.709439f, 0.486166f, -1.203050f, -0.978381f, 0.631876f,
277 0.000705f, 0.123858f, -0.001187f, -0.172312f, -0.422668f, 0.241838f,
278 0.437400f, -0.268186f, -0.513259f, 0.450209f, 0.542629f, -0.453810f,
279 -0.207119f, 0.072598f, 0.085066f, -0.018986f, -0.149512f, 0.149521f,
280 0.182105f, -0.227200f, -0.363240f, 0.172670f, -0.502932f, 0.689256f,
281 0.093760f, -0.090207f, -0.066803f, 0.056759f, -0.002243f, -0.050662f,
282 -0.059324f, 0.152943f, -0.701150f, 0.712540f, 0.660349f, -0.654970f,
283 0.351772f, -0.303383f, -0.311177f, 0.247653f, 0.013035f, 0.034648f,
284 -0.137832f, 0.041197f, 0.410265f, 0.345129f, 0.653338f, 0.047050f,
285 0.140399f, 0.018613f, -0.012431f, -0.113632f, -0.029928f, 0.051564f,
286 -0.031349f, 0.151944f, -0.160340f, 0.326798f, -0.458067f, 0.636235f,
287 0.243184f, 0.514072f, 2.414450f, 1.421980f, -0.001474f, -0.141389f,
288 -0.104817f, -0.141882f, -0.026395f, 0.053014f, 0.143885f, -0.207774f,
289 -0.563846f, -0.242514f, -0.436574f, -0.456796f, -0.520646f, 0.282550f,
290 -0.684924f, 0.061105f, -0.315884f, -0.392624f, 0.009805f, -0.256597f,
291 -0.146732f, 0.331039f, 0.362342f, 0.270851f, 0.067679f, -0.071331f,
292 -0.222423f, 0.081286f, -0.208192f, -0.193816f, -0.008201f, -0.309340f,
293 0.167556f, 0.106071f, 0.172254f, -0.163790f, -0.142205f, -0.043182f,
294 0.096145f, 0.145037f, -0.066015f, -0.073194f, 0.132237f, -0.088522f,
295 -0.044292f, -0.487128f, 0.033389f, -0.573548f, 0.185449f, 0.273593f,
296 0.147503f, 0.457049f, -0.021539f, 0.090786f, 0.009147f, 0.000899f,
297 0.018088f, 0.115791f, -0.079165f, 0.139388f,
298 };
299
300 static const float weights_layer_2[] = {
301 0.153048f, 0.112901f, 0.136781f, 0.154580f, 0.091610f, 0.045165f,
302 0.088490f, 0.116991f, -0.463766f, -0.596567f, -0.567008f, -0.630565f,
303 0.141874f, 0.095726f, 0.175427f, 0.145027f, -0.969824f, -1.018190f,
304 -1.073300f, -1.041130f, -0.070545f, -0.123600f, -0.114967f, -0.169453f,
305 -0.267458f, -0.147730f, -0.161419f, -0.164894f, -0.117508f, -0.204389f,
306 -0.122695f, -0.163107f, -0.003903f, -0.030470f, -0.037433f, -0.059568f,
307 0.138243f, 0.091019f, 0.160372f, 0.141650f, -0.544565f, -0.620004f,
308 -0.504503f, -0.429979f, -0.099491f, -0.096384f, -0.155265f, -0.188536f,
309 0.084923f, 0.038345f, 0.066706f, 0.122083f, 0.267087f, 0.184419f,
310 0.261478f, 0.255746f, -0.245894f, -0.114980f, -0.193880f, -0.227785f,
311 0.087536f, 0.095712f, 0.106105f, 0.099353f, -0.059473f, -0.173247f,
312 -0.202386f, -0.076010f, 0.125928f, 0.100793f, 0.119638f, 0.129623f,
313 0.136593f, 0.102984f, 0.156550f, 0.140558f, 0.122524f, 0.051596f,
314 0.084164f, 0.123630f, 0.072542f, 0.096063f, 0.083236f, 0.087630f,
315 0.025900f, 0.023738f, 0.036385f, 0.053077f, -0.029501f, 0.010544f,
316 -0.010026f, -0.051268f, 0.086302f, 0.109909f, 0.101385f, 0.127513f,
317 -0.031869f, 0.005340f, -0.056267f, -0.032955f, 0.032748f, 0.023162f,
318 0.092118f, -0.001780f, -0.123612f, -0.183433f, -0.202377f, -0.317516f,
319 0.129052f, 0.208112f, 0.145582f, 0.175502f, 0.018476f, 0.036349f,
320 0.072417f, 0.061194f, 0.086985f, 0.117086f, 0.072465f, 0.129068f,
321 0.020182f, 0.052114f, 0.017878f, 0.010478f, -0.001381f, -0.034644f,
322 0.025135f, -0.037748f, 0.004973f, 0.024778f, 0.041816f, 0.032111f,
323 0.080268f, 0.124998f, 0.105719f, 0.177047f, -0.072114f, -0.011864f,
324 -0.076846f, -0.089840f, 0.069993f, 0.089362f, 0.088035f, 0.120621f,
325 0.065916f, 0.100946f, -0.006784f, -0.007751f, 0.122039f, 0.126482f,
326 0.078629f, 0.140299f, 0.074034f, 0.092464f, 0.089798f, 0.108968f,
327 0.075729f, 0.057128f, 0.013570f, 0.021195f, 0.068901f, 0.054022f,
328 0.029781f, 0.031404f, -0.209998f, -0.208731f, -0.198310f, -0.212454f,
329 -0.579168f, -0.490190f, -0.607567f, -0.520541f, 0.083863f, 0.056612f,
330 0.030366f, 0.061790f, -0.004874f, -0.057203f, -0.060429f, -0.049145f,
331 0.080086f, 0.138602f, 0.223796f, 0.133279f, -0.495954f, -0.612093f,
332 -0.545393f, -0.562310f, 0.070672f, 0.037702f, 0.139013f, 0.080192f,
333 -0.111387f, -0.048165f, 0.074359f, -0.042125f, 0.113633f, 0.106579f,
334 0.042633f, 0.102734f, -0.068220f, 0.128423f, -0.181821f, -0.013260f,
335 -0.108563f, -0.138667f, -0.109304f, -0.131909f, -0.168667f, -0.126870f,
336 -0.132533f, -0.167096f, -0.184741f, -0.140890f, -0.125361f, -0.150632f,
337 0.309013f, 0.364376f, 0.361102f, 0.271566f, 0.116552f, 0.091160f,
338 0.096846f, 0.095954f, 0.046972f, 0.080489f, 0.028766f, -0.012223f,
339 0.071379f, 0.041535f, -0.000668f, 0.033698f, -0.013493f, -0.027535f,
340 -0.025804f, -0.012267f, -0.097465f, -0.099232f, -0.208863f, -0.225201f,
341 -0.475608f, 0.077358f, -0.002872f, 0.163890f, -0.420298f, 0.072114f,
342 0.121601f, -0.016727f, 0.573853f, -0.080196f, 0.193053f, 0.053012f,
343 -0.454179f, 0.058563f, 0.067265f, 0.141154f, 0.412541f, 0.086933f,
344 0.030407f, -0.030413f, 0.478757f, -0.097731f, 0.277072f, -0.086393f,
345 0.552604f, -0.334201f, 0.091765f, -0.270262f, -1.395060f, 0.271837f,
346 -0.005335f, 0.240499f, 0.175442f, -0.326329f, -0.019353f, -0.270338f,
347 -0.459273f, 0.096183f, 0.153046f, 0.135818f, 0.759028f, -0.177673f,
348 -0.099966f, 0.103363f, 0.697289f, -0.234184f, -0.048706f, -0.116099f,
349 -0.282575f, 0.025655f, -0.184759f, 0.040658f, -0.558267f, 0.214087f,
350 -0.095620f, 0.200522f, 0.278996f, 0.031959f, 0.122936f, -0.209196f,
351 -0.308217f, 0.092917f, 0.113269f, 0.136274f, -0.037046f, 0.017263f,
352 -0.194183f, 0.089133f, -0.161244f, 0.042799f, 0.030557f, 0.153545f,
353 -0.355048f, 0.070928f, -0.152852f, 0.102875f, -0.193649f, 0.007916f,
354 -0.062952f, 0.050602f, 0.073671f, 0.143045f, -5.978970f, -7.013850f,
355 0.058713f, 0.076116f, 0.026445f, -0.056599f, -0.005966f, 0.032234f,
356 0.006753f, -0.024528f, 0.120308f, 0.179939f, -6.624630f, -7.638680f,
357 0.026359f, 0.020758f, 0.194274f, 0.051489f, -0.008491f, -0.028248f,
358 -0.061328f, -0.134423f, -0.103951f, -0.110877f, 0.042263f, 0.127016f,
359 0.012473f, -0.008595f, 0.031357f, 0.087476f, -0.084022f, -0.015590f,
360 -0.313546f, 0.120072f, 0.123880f, 0.162148f, -6.596560f, -7.358830f,
361 0.004797f, -0.003415f, 0.048455f, 0.026737f, -0.103702f, 0.034416f,
362 -0.003475f, -0.236827f, 0.005378f, 0.048413f, 0.054612f, -0.079359f,
363 0.043707f, 0.001085f, 0.023380f, 0.007785f, 0.025938f, -0.052856f,
364 -0.033421f, 0.022643f, 0.034161f, 0.127681f, -5.019490f, -5.233580f,
365 -0.128630f, 0.087741f, -0.239834f, -0.377876f, 0.128082f, 0.142730f,
366 -0.086819f, -0.350927f, 0.089849f, 0.155776f, -6.155120f, -5.721720f,
367 0.056110f, 0.008761f, 0.045579f, 0.016762f, -0.134076f, -0.101551f,
368 -0.096058f, -0.117146f, 0.003527f, -0.056942f, -0.005578f, 0.071287f,
369 0.023776f, -0.028003f, -0.075390f, -0.191160f, -0.089672f, -0.104372f,
370 -0.104750f, -0.080813f, -0.249824f, -0.124479f, -0.243593f, -0.244284f,
371 -0.554911f, -0.549095f, -0.564693f, -0.475107f, -0.121771f, -0.143441f,
372 -0.171170f, -0.120920f, 0.109831f, 0.079708f, 0.327295f, 0.308907f,
373 -0.178785f, -0.428316f, -0.418882f, -0.366750f, -0.139296f, -0.129645f,
374 -0.081237f, -0.101533f, -0.006256f, -0.146756f, -0.322110f, -0.338865f,
375 -0.306085f, -0.319592f, -0.454803f, -0.363560f, -0.018557f, 0.006605f,
376 -0.131198f, -0.077708f, 0.138160f, 0.119611f, 0.271098f, 0.232168f,
377 0.027812f, 0.035390f, -0.202503f, -0.091172f, -0.142020f, -0.159929f,
378 -0.106404f, -0.107433f, -0.381743f, -0.353222f, -0.484159f, -0.469926f,
379 -0.234659f, -0.315674f, -0.178327f, -0.213485f, -0.096207f, -0.190944f,
380 -0.118917f, -0.161288f, 0.015996f, 0.060737f, 0.051390f, 0.060876f,
381 0.229289f, 0.282418f, 0.250945f, 0.197273f, 0.045131f, -0.008305f,
382 0.072024f, 0.044547f, -0.050010f, 0.055504f, 0.001343f, -0.014445f,
383 0.254909f, 0.309091f, 0.228249f, 0.274843f, 0.089778f, -0.046581f,
384 0.072714f, 0.126814f, -0.048931f, -0.045743f, -0.151333f, -0.004490f,
385 0.179966f, 0.058150f, -0.178622f, -0.088159f, -0.074416f, -0.005821f,
386 -0.011799f, -0.002225f, -0.069361f, -0.098937f, -0.081575f, -0.034796f,
387 0.253792f, 0.301039f, 0.219163f, 0.256027f, 0.058007f, -0.041431f,
388 0.040674f, 0.009019f, -0.099670f, -0.099077f, -0.039437f, 0.017946f,
389 0.060717f, 0.045796f, 0.109664f, 0.032138f, -0.071094f, 0.023697f,
390 0.011335f, -0.030465f, 0.068677f, 0.039345f, -0.045078f, 0.084037f,
391 0.135517f, 0.190417f, 0.175578f, 0.155286f, -0.044505f, 0.010826f,
392 0.006717f, -0.134715f, 0.068022f, 0.110095f, 0.079966f, 0.034481f,
393 0.185804f, 0.188273f, 0.227283f, 0.135935f, 0.033447f, 0.031571f,
394 -0.014766f, -0.024565f, 0.021792f, 0.017675f, -0.001333f, -0.040069f,
395 -0.049384f, -0.045256f, -0.014013f, -0.000107f, -0.096928f, -0.111495f,
396 -0.051225f, -0.060449f, 0.071446f, 0.017294f, -0.004822f, 0.006932f,
397 0.020884f, 0.089425f, 0.061097f, -0.038708f, -0.184029f, -0.089541f,
398 -0.158035f, -0.214607f, -0.377947f, -0.318586f, -0.336977f, -0.323908f,
399 0.181612f, 0.140018f, 0.233524f, 0.193366f, -0.254507f, -0.271902f,
400 -0.197144f, -0.119539f, 0.042162f, 0.000320f, 0.014708f, -0.014228f,
401 -0.081119f, -0.089326f, 0.001763f, 0.081009f, -0.142618f, -0.160650f,
402 -0.214597f, -0.202143f, -0.053495f, -0.012819f, -0.071468f, -0.010883f,
403 0.072570f, 0.071507f, 0.091045f, 0.083155f, -0.271237f, -0.289211f,
404 -0.272345f, -0.299411f, 0.031697f, -0.029795f, -0.030045f, -0.013604f,
405 -0.106843f, -0.045212f, -0.122459f, -0.096936f, 0.059793f, 0.006157f,
406 0.028092f, 0.040589f, -0.014560f, -0.008975f, -0.051404f, -0.014309f,
407 -0.016883f, 0.018332f, 0.040114f, 0.050348f, 0.044921f, -0.002445f,
408 -0.112396f, 0.014395f, 0.115160f, 0.145350f, -0.166814f, -0.121449f,
409 0.155573f, -0.099446f, -0.161661f, 0.187251f, 0.004711f, 0.024318f,
410 -0.060871f, -0.028311f, -0.098274f, 0.322030f, -0.069242f, -0.153173f,
411 -0.227428f, -0.293965f, 0.228491f, 0.111413f, -1.354720f, -0.344235f,
412 0.866715f, 0.872344f, 0.078789f, -0.384865f, 0.162388f, 0.109018f,
413 -0.191549f, -0.002638f, 0.305053f, 0.087337f, 0.066506f, -0.055810f,
414 -0.010984f, -0.056160f, -0.114617f, -0.058478f, 0.022059f, -0.124368f,
415 -0.130989f, 0.369432f, -0.248898f, -0.003955f, -0.021578f, 0.115991f,
416 -0.114163f, -0.065232f, 0.339857f, -0.225997f, 0.006282f, -0.125395f,
417 0.235082f, -0.347785f, 0.662321f, -0.529182f, 0.153297f, -0.001326f,
418 -0.026725f, -0.024677f, -0.088065f, -0.116127f, 0.080896f, 0.212542f,
419 0.208421f, 0.032047f, -0.211395f, 0.074997f, 0.096659f, 0.096423f,
420 -0.078643f, 0.106556f, -0.123860f, 0.075609f, 0.066008f, -0.097275f,
421 -1.000020f, -0.780154f, -0.856922f, -0.964007f, 0.083135f, -0.018922f,
422 -0.266214f, -0.151480f, 0.051538f, 0.017802f, 0.066774f, -0.021341f,
423 -0.869494f, -0.935252f, -0.895836f, -0.853871f, -0.160490f, 0.085850f,
424 -0.029670f, -0.056675f, 0.159989f, 0.166872f, 0.129970f, 0.194377f,
425 0.153294f, 0.199593f, 0.037692f, 0.103391f, 0.029335f, -0.085324f,
426 -0.079326f, -0.077216f, 0.501561f, 0.366168f, 0.330196f, 0.296432f,
427 -0.977282f, -0.844295f, -1.014870f, -1.098990f, -0.099858f, -0.129552f,
428 0.090051f, -0.013378f, 0.081330f, 0.194911f, 0.286501f, 0.177363f,
429 -0.148250f, -0.111700f, -0.243081f, -0.102918f, 0.161069f, -0.012655f,
430 -0.071722f, -0.020329f, -0.077828f, -0.041716f, 0.109247f, 0.062229f,
431 -0.759722f, -0.742756f, -0.563713f, -0.631187f, 0.005911f, 0.268154f,
432 -0.263769f, 0.087149f, -0.163623f, -0.359600f, -0.464577f, -0.369352f,
433 -0.515784f, -0.475822f, -0.523485f, -0.649813f, -0.112419f, -0.029285f,
434 0.021061f, -0.041515f, 0.149133f, -0.254428f, 0.115776f, -0.061892f,
435 0.103675f, -0.283363f, 0.005005f, 0.022034f, -0.178454f, 0.035836f,
436 -0.113702f, -0.217823f, 0.209407f, -0.296257f, 0.187976f, -0.157370f,
437 -0.127190f, 0.251780f, 0.055633f, 0.294111f, -0.067773f, 0.467190f,
438 -0.192625f, -0.071084f, -0.445284f, 0.511090f, -0.319728f, 0.267971f,
439 0.494929f, -0.586727f, 0.454543f, -0.520675f, -0.085900f, 0.325989f,
440 -0.131006f, -0.069501f, 0.199927f, -0.218919f, 0.170055f, -0.106538f,
441 0.133312f, 0.127629f, -0.561625f, 0.595666f, -0.090927f, 0.363348f,
442 -0.249246f, 0.063068f, -0.016458f, -0.291045f, -0.040509f, 0.017866f,
443 0.304871f, -0.459214f, 0.214390f, -0.238740f, -0.456541f, 0.545848f,
444 -0.218026f, 0.202475f, 0.128490f, -0.036417f, 0.173885f, -0.049385f,
445 0.235514f, -0.132587f, -0.015066f, 0.164638f, 0.196873f, -0.125330f,
446 0.216912f, -0.109398f, 0.121602f, -0.209374f, 0.164400f, -0.123049f,
447 0.195520f, -0.212932f, -0.015180f, -0.005784f, 0.049726f, -5.822150f,
448 0.124536f, 0.040689f, -0.018560f, -3.155020f, 0.014690f, 0.076202f,
449 -0.154008f, 1.070630f, -0.071606f, 0.051026f, 0.138285f, -5.836340f,
450 0.162173f, 0.085890f, -0.186166f, 0.093221f, 0.019240f, -0.017053f,
451 -0.090144f, 0.236254f, -0.125344f, 0.056235f, -0.089813f, -0.252281f,
452 -0.127406f, -0.155088f, 0.009972f, -0.066449f, 0.044222f, 0.025943f,
453 -0.164921f, 0.165463f, -0.001132f, -0.038386f, 0.115194f, -5.757100f,
454 0.163386f, 0.061226f, 0.024626f, 0.132750f, 0.107279f, -0.001622f,
455 -0.107860f, -0.356009f, -0.138935f, -0.145173f, -0.061198f, -0.646138f,
456 0.034279f, 0.078187f, 0.108138f, -0.490444f, 0.074719f, 0.034984f,
457 -0.109303f, 0.741785f, -0.066939f, 0.015558f, 0.114229f, -4.001080f,
458 0.130772f, 0.044675f, -0.165162f, -0.274810f, -0.042987f, -0.048579f,
459 0.156603f, -1.288370f, 0.076198f, 0.035065f, 0.032043f, -5.002520f,
460 0.086900f, -0.010886f, 0.030850f, -0.782259f, 0.056211f, -0.097759f,
461 0.118988f, 0.106638f, 0.091419f, 0.079920f, 0.062325f, 0.097116f,
462 0.126035f, 0.122530f, -0.278299f, -0.083314f, -0.300563f, -0.197946f,
463 0.081664f, 0.089925f, 0.074754f, 0.074628f, 0.102338f, 0.088845f,
464 0.105841f, 0.102381f, 0.003087f, 0.061599f, 0.098326f, 0.040119f,
465 -0.005298f, -0.028834f, 0.059938f, -0.013668f, -0.585882f, -0.631436f,
466 -0.742673f, -0.736666f, 0.025071f, 0.066851f, 0.075046f, 0.091360f,
467 0.099045f, 0.098261f, 0.106413f, 0.099487f, -0.016742f, -0.097334f,
468 -0.086152f, -0.212444f, -0.028043f, -0.007362f, 0.003914f, -0.055864f,
469 0.034756f, 0.081361f, 0.080183f, 0.061319f, 0.193396f, 0.173716f,
470 0.207765f, 0.231701f, -0.074565f, -0.073257f, -0.086470f, -0.083114f,
471 0.081489f, 0.078477f, 0.033452f, 0.058835f, -0.069665f, -0.031691f,
472 -0.111255f, -0.167754f, 0.184179f, 0.174673f, 0.160288f, 0.190893f,
473 0.110930f, 0.103495f, 0.098408f, 0.102918f, 0.053764f, 0.089994f,
474 0.140308f, 0.124867f, 0.074176f, 0.117460f, -0.160775f, -0.144132f,
475 -0.099373f, -0.035913f, 0.081237f, 0.062247f, -0.166421f, 0.062125f,
476 0.276479f, 0.060955f, 0.066627f, 0.455347f, 0.219953f, 0.109912f,
477 0.273931f, 0.233153f, 0.102236f, 0.447606f, -0.352243f, 0.499236f,
478 -0.931206f, 0.248595f, 0.254047f, 0.061542f, 0.268804f, 0.309517f,
479 -0.084414f, -0.245828f, -0.144882f, -0.296579f, -0.091628f, -0.142202f,
480 -0.541764f, -0.407470f, 0.053481f, 0.238955f, 0.150188f, -0.060598f,
481 0.196118f, -0.215617f, -0.086238f, -0.263420f, 0.206877f, 0.241788f,
482 -0.122544f, -0.448790f, 0.286917f, 0.112063f, -0.268408f, -0.041770f,
483 0.089161f, 0.355811f, -0.078245f, -0.148490f, -0.407301f, -1.296870f,
484 -0.633421f, 0.124253f, 0.275402f, 0.223048f, 0.077016f, 0.160766f,
485 0.115374f, 0.061053f, -0.231872f, -0.515052f, -0.278331f, -0.235912f,
486 -0.416372f, -0.284106f, -0.055942f, 0.110698f, -0.428288f, -0.298137f,
487 -0.018101f, 0.102677f, -0.019639f, 0.013479f, 0.038549f, 0.048682f,
488 0.128684f, 0.116416f, 0.044852f, 0.008133f, 0.061597f, 0.083582f,
489 0.014953f, 0.063716f, -0.155318f, -0.061732f, 0.084855f, 0.129505f,
490 0.068249f, 0.193775f, -0.088631f, -0.446398f, -0.075710f, -0.061327f,
491 0.278715f, 0.540366f, 0.618715f, 0.538374f, -0.037843f, 0.062370f,
492 -0.033184f, 0.119901f, -0.008641f, -0.064789f, 0.087498f, 0.043486f,
493 0.247085f, 0.419992f, 0.299935f, 0.234276f, 0.089283f, 0.070357f,
494 0.068888f, 0.134311f, 0.109823f, 0.072431f, 0.081676f, 0.091366f,
495 -1.707980f, -2.213110f, -2.149930f, -1.556870f, 0.226598f, 0.191675f,
496 0.192207f, 0.159566f, -0.070194f, -0.136070f, -0.015172f, -0.204272f,
497 -0.162191f, -0.043313f, -0.158007f, -0.227210f, 0.040398f, 0.043014f,
498 0.039439f, -0.035439f, 0.245558f, 0.439691f, 0.219659f, 0.138210f,
499 -0.048129f, 0.004954f, -0.102860f, -0.185376f, 0.035548f, 0.006821f,
500 0.079199f, 0.032901f, 0.039218f, 0.068113f, 0.023075f, -0.037582f,
501 0.225181f, 0.164562f, 0.106718f, 0.032684f, 0.013402f, 0.018797f,
502 0.076606f, 0.046512f, -0.070024f, 0.099921f, -0.051231f, 0.074167f,
503 0.173313f, 0.220212f, 0.142665f, 0.069809f, -0.195130f, -0.007912f,
504 -0.006764f, -0.063687f, 0.306374f, 0.402035f, 0.273759f, 0.449469f,
505 0.114597f, 0.210745f, 0.355326f, 0.271307f, -0.109943f, -0.171912f,
506 -0.070726f, -0.128932f, 0.138770f, 0.164971f, 0.308516f, 0.332536f,
507 0.081537f, 0.096939f, 0.054136f, 0.052226f, 0.109489f, 0.010223f,
508 0.168072f, -0.106279f, 0.525568f, 0.704816f, 0.588942f, 0.473398f,
509 0.149497f, 0.120835f, 0.080049f, 0.151340f, -0.182038f, -0.191091f,
510 -0.196505f, -0.198309f, -0.801819f, -1.441620f, -1.107780f, -1.025650f,
511 0.035750f, 0.018049f, -0.029033f, -0.067255f, 0.192049f, 0.009664f,
512 -0.043741f, 0.051557f, 0.082815f, 0.069547f, -0.073379f, 0.010584f,
513 0.192128f, 0.208586f, 0.141904f, 0.100763f, 0.046183f, 0.044776f,
514 -0.033611f, -0.005812f, 0.012966f, 0.030301f, 0.100665f, 0.103641f,
515 -0.294776f, -0.361573f, -0.420156f, -0.388743f, 0.239287f, 0.191975f,
516 0.089644f, 0.117591f, 0.069563f, 0.021480f, 0.100287f, 0.174159f,
517 -0.013571f, 0.090960f, 0.010232f, -0.034760f, -0.077205f, 0.060632f,
518 -0.145527f, -0.391110f, -0.143052f, -0.236448f, -0.103902f, -0.188463f,
519 0.071311f, -0.080171f, 0.021987f, 0.041767f, -0.419487f, -0.515479f,
520 -0.205470f, -0.732132f, 0.150901f, 0.107202f, 0.156307f, 0.143672f,
521 0.474682f, 0.178137f, 0.150063f, 0.414515f, 0.559891f, 0.697019f,
522 0.541231f, 0.505310f, -0.478101f, -0.444267f, -0.586539f, -0.445996f,
523 -0.451873f, -0.530085f, -0.447980f, -0.364955f, 0.372435f, 0.318894f,
524 0.351211f, 0.193961f, 0.212295f, 0.212842f, 0.220003f, 0.243743f,
525 -0.388628f, -0.789620f, -0.536618f, -0.430691f, 0.247004f, 0.266489f,
526 0.261033f, 0.263692f, 0.050089f, 0.048958f, 0.065207f, 0.120180f,
527 -0.526230f, -0.481969f, -0.422411f, -0.272292f, 0.155593f, 0.229614f,
528 0.139579f, 0.171805f, -0.251924f, -0.302067f, -0.126157f, -0.346650f,
529 -1.195450f, -1.281100f, -0.987911f, -1.478440f, 0.285667f, 0.284802f,
530 0.301887f, 0.259556f, -0.194127f, -0.090440f, -0.257959f, -0.259572f,
531 -0.012273f, -0.049993f, -0.099431f, 0.012506f, 0.081526f, 0.166279f,
532 0.042594f, 0.185121f, 0.148830f, 0.073161f, 0.201728f, 0.125747f,
533 -0.295065f, -0.187585f, -0.333066f, -0.312291f, 0.253458f, 0.321585f,
534 0.178844f, 0.219944f, -0.763475f, -0.943374f, -0.816825f, -0.709901f,
535 -0.166132f, 0.129186f, 0.015405f, -0.065623f, -0.246006f, -0.340385f,
536 -0.118155f, -0.384905f, -0.233883f, -0.400666f, -0.228597f, -0.228428f,
537 -0.559083f, -0.377784f, -0.541458f, -0.542870f, 0.067400f, 0.122987f,
538 0.180901f, 0.186004f, -0.482910f, -0.424823f, -0.477831f, -0.394719f,
539 0.091558f, 0.049248f, 0.049370f, 0.160429f, 0.133641f, 0.096625f,
540 0.104429f, 0.100782f, -0.238252f, -0.221459f, -0.196974f, -0.250393f,
541 -3.071750f, -2.418450f, -0.861410f, -1.051580f, 0.071263f, 0.118014f,
542 -0.028430f, -0.072073f, -0.074463f, 0.034168f, 0.044089f, -0.091109f,
543 -3.153840f, -2.945850f, -1.977360f, -1.498850f, -0.083429f, 0.131835f,
544 -0.063865f, -0.065785f, -0.069346f, -0.015520f, -0.119551f, 0.044881f,
545 -0.105280f, 0.127516f, 0.005255f, -0.142777f, 0.061055f, -0.117250f,
546 0.020454f, 0.157879f, -0.213812f, -0.151783f, 0.028583f, 0.137759f,
547 -3.248250f, -3.005940f, -1.510540f, -1.475390f, 0.081874f, -0.171465f,
548 -0.135690f, -0.001989f, -0.227574f, -0.132799f, -0.359742f, -0.137197f,
549 0.066324f, 0.039194f, -0.050857f, 0.095166f, 0.044475f, 0.011221f,
550 0.054904f, 0.061414f, -0.039189f, 0.123751f, -0.017171f, -0.008494f,
551 -2.598220f, -2.832670f, -1.622030f, -1.201990f, 0.154313f, -0.021436f,
552 0.042190f, 0.143947f, -0.090623f, 0.086853f, 0.143137f, 0.099821f,
553 -1.732820f, -1.429730f, -0.775125f, -0.648036f, 0.082176f, 0.079448f,
554 -0.040575f, 0.024511f, -0.064105f, -0.117122f, -0.190323f, -0.182589f,
555 -0.076430f, -0.095615f, -0.112513f, -0.101581f, 0.143037f, 0.148180f,
556 0.430958f, 0.359225f, 0.001403f, -0.080541f, -0.295001f, -0.156706f,
557 0.426623f, 0.475597f, 0.455210f, 0.454352f, 0.074365f, 0.099440f,
558 0.066348f, -0.007078f, 0.008335f, -0.097116f, -0.133687f, -0.110535f,
559 0.204145f, 0.281478f, 0.078886f, 0.112857f, -0.103620f, -0.068247f,
560 0.191147f, 0.227593f, -0.011816f, -0.058755f, -0.149477f, -0.101828f,
561 0.079878f, 0.304949f, 0.557555f, 0.305288f, -0.150955f, -0.118610f,
562 0.052073f, 0.064707f, -0.121728f, -0.151132f, -0.193987f, -0.175046f,
563 0.043655f, 0.105270f, -0.120715f, -0.040976f, 0.047776f, -0.004443f,
564 0.149606f, 0.111240f, -0.047502f, -0.064146f, -0.151858f, -0.151872f,
565 -0.160207f, -0.113846f, -0.081585f, -0.006708f, -0.203760f, -0.068597f,
566 -0.179979f, -0.127779f, -0.062460f, -0.064513f, -0.121479f, -0.111122f,
567 -0.212384f, -0.229157f, -0.283428f, -0.184891f,
568 };
569
570 static const float weights_layer_3[] = {
571 -0.039388f, 0.033048f, -0.113003f, -0.011642f, 0.170478f, 0.145713f,
572 0.040189f, -0.280129f, -0.049050f, -0.043788f, -0.157425f, 0.323829f,
573 -0.250725f, -0.166349f, 0.101650f, -0.049690f, 0.205606f, 0.281131f,
574 0.623204f, 0.993452f, -0.015115f, -0.138995f, 0.009473f, 0.157673f,
575 -0.024687f, -0.067214f, 0.125566f, -0.317619f, 0.057002f, 0.031202f,
576 -0.018167f, 0.068542f, 0.011609f, -0.020233f, -0.000428f, -0.035956f,
577 -0.843274f, -0.800587f, -0.214917f, -0.221250f, 0.031255f, -0.077330f,
578 -0.074902f, -0.063979f, -0.055562f, 0.679495f, 0.146609f, 1.315330f,
579 -0.118399f, -0.034539f, -0.050377f, 0.172867f, -0.204607f, -0.034930f,
580 0.176014f, 0.089747f, -0.003889f, 0.044980f, 0.002386f, -0.141723f,
581 -0.035828f, -0.204701f, 0.099813f, 0.123580f, 0.209851f, -0.110989f,
582 -0.043655f, -0.461118f, -0.139664f, 0.026855f, -0.081714f, 0.207623f,
583 0.089942f, 0.253082f, 0.680568f, 0.811360f, -0.090528f, -0.116818f,
584 -0.432361f, -0.075588f, -0.269924f, -0.276810f, -0.289192f, -0.282570f,
585 0.245566f, 0.267216f, 0.238622f, 0.286528f, -0.157605f, -0.200401f,
586 -0.138924f, -0.185006f, 0.215203f, 0.203316f, 0.209532f, 0.293135f,
587 0.928046f, 0.733323f, -0.094120f, 0.036918f, -0.126643f, -0.083371f,
588 -0.147530f, -0.153195f, 0.097097f, 0.101852f, 0.109160f, 0.105129f,
589 -0.051869f, -0.064359f, -0.073469f, -0.059591f, 0.102431f, 0.109444f,
590 0.113614f, 0.105617f, 0.383311f, 0.325783f, 0.393234f, 0.382508f,
591 0.194720f, 0.189672f, 0.217477f, 0.177786f, 0.326461f, 0.114789f,
592 0.317061f, 0.048291f, -0.061143f, -0.134641f, -0.067895f, -0.108446f,
593 0.082592f, 0.029918f, -0.006580f, 0.015533f, -0.053583f, -0.055540f,
594 -0.063395f, -0.023157f, -0.064955f, -0.073981f, -0.115452f, -0.086626f,
595 -0.036616f, 0.008454f, 0.012029f, -0.008039f, -0.207395f, -0.216419f,
596 -0.205363f, -0.249099f, 0.343308f, 0.413215f, -0.009918f, -0.109978f,
597 -0.059711f, -0.045089f, -0.029130f, -0.038483f, -0.070323f, -0.099409f,
598 -0.008849f, -0.063527f, 0.175963f, 0.185335f, 0.149151f, 0.199997f,
599 -0.027516f, -0.039812f, -0.027760f, -0.047910f, -0.007337f, 0.071065f,
600 0.086225f, 0.125539f, 0.151390f, 0.215488f, 0.203450f, 0.045380f,
601 0.095761f, 0.107809f, 0.103918f, 0.122383f, 0.116287f, 0.135455f,
602 0.115446f, 0.155673f, -0.044648f, -0.027455f, -0.015473f, -0.026657f,
603 0.089852f, 0.077459f, 0.077631f, 0.082507f, -0.102761f, -0.054669f,
604 -0.132223f, -0.024768f, 0.111573f, 0.060467f, 0.107883f, 0.056621f,
605 0.219357f, -0.161153f, 0.074379f, -0.118743f, -0.169931f, -0.153995f,
606 -0.220003f, -0.200186f, 0.032318f, -0.060687f, -0.087550f, -0.038022f,
607 0.026633f, -0.005534f, 0.029532f, 0.027081f, 0.011926f, 0.058412f,
608 0.010631f, 0.003068f, -0.014911f, 0.063070f, 0.065271f, 0.089550f,
609 0.012885f, 0.005320f, -0.037494f, -0.019849f, -0.009624f, -0.059090f,
610 -0.021222f, -0.088033f, -0.055261f, -0.055113f, -0.047598f, -0.055478f,
611 -0.023648f, -0.046827f, -0.036572f, -0.057655f, 0.104194f, 0.179800f,
612 0.175751f, 0.192851f, -0.016950f, -0.073650f, -0.028592f, -0.088219f,
613 0.011130f, 0.061825f, 0.025643f, 0.034183f, 0.095548f, 0.001457f,
614 -0.132869f, 0.032981f, -0.140178f, -0.105343f, -0.161799f, -0.161983f,
615 0.177746f, 0.132903f, 0.135627f, 0.152489f, -0.012532f, -0.068747f,
616 -0.085849f, -0.095434f, 0.087037f, 0.139497f, 0.111899f, 0.100189f,
617 -0.024649f, -0.092003f, 0.020783f, -0.115807f, 0.092039f, 0.093943f,
618 0.109466f, 0.049639f, -0.133727f, 0.128430f, -0.050546f, 0.190632f,
619 0.123733f, 0.082305f, 0.114878f, 0.122572f, 0.201618f, 0.137588f,
620 0.065582f, 0.125161f, -0.095179f, -0.120719f, -0.127126f, -0.101961f,
621 -0.118120f, -0.104833f, -0.179632f, -0.131764f, -0.138096f, -0.147861f,
622 -0.131512f, -0.153905f, -0.201816f, -0.206641f, -0.196707f, -0.160013f,
623 -0.212605f, -0.093998f, -0.186258f, -0.076137f, -0.065340f, -0.006969f,
624 -0.071383f, -0.075005f,
625 };
626
627 static const float weights_layer_4[] = {
628 -0.016102f, -0.022836f, 0.624049f, 0.273485f, 0.222800f, -0.290175f,
629 -0.518415f, 0.413484f, -0.264495f, 0.498083f, -0.450145f, -0.106419f,
630 0.095103f, -0.187451f, 0.145933f, -0.371542f, -0.088871f, 0.184017f,
631 -0.429625f, -0.110882f, 0.292781f, 0.289588f, 0.185127f, 0.326017f,
632 -0.432009f, -0.342663f, -0.312206f, 0.004004f, -1.114290f, 0.028497f,
633 -0.264944f, -0.419611f, 0.046336f, 0.138232f, -0.869528f, 0.425557f,
634 -0.954838f, -0.186830f, -0.464622f, -0.757107f, -0.432686f, -0.125978f,
635 -0.402633f, -0.172266f, -0.041749f, -0.822238f, -0.118486f, 0.238617f,
636 -0.198037f, 0.146347f, 0.405257f, 0.513303f, -0.078876f, -0.300385f,
637 -0.010293f, -0.183962f, 0.155738f, 0.186797f, -0.086814f, 0.000179f,
638 0.123467f, 0.362523f, 0.068805f, 0.371834f, 0.038122f, -0.117867f,
639 -0.120445f, -0.422322f, -0.131402f, 0.285449f, 0.038957f, 0.008844f,
640 -0.020197f, 0.187723f, 0.190433f, 0.146532f, -0.091068f, -0.270865f,
641 -0.194231f, -0.226777f, 0.013548f, 0.248351f, 0.537685f, 0.056316f,
642 -0.171540f, -0.003865f, 0.406439f, 0.126507f, 0.192780f, 0.149335f,
643 -0.149602f, 0.255202f, -0.015426f, 0.032335f, -1.791330f, -0.894602f,
644 -0.196641f, -0.282846f, -0.391100f, -0.040969f, 0.049934f, 0.056348f,
645 -0.041426f, -0.075159f, -0.658335f, -0.827270f, -0.175029f, -0.427235f,
646 0.311201f, 0.560413f, 0.363408f, 0.374580f, -0.433531f, -0.180580f,
647 0.142142f, 0.194768f, -0.054118f, -0.376541f, -0.366185f, -0.308782f,
648 -0.273143f, -0.074097f, 0.009000f, -0.182198f, -0.015616f, -0.003882f,
649 -0.174340f, -0.354866f, 0.527972f, 0.348355f, 0.091381f, -0.419828f,
650 -0.530529f, 0.159899f, -0.511867f, -0.104237f, -0.286079f, -0.659039f,
651 -0.266596f, -0.256557f, -0.600437f, -0.446333f, -0.229629f, 0.024931f,
652 -0.143716f, -0.415754f, -0.003760f, -0.107195f, -0.666165f, -0.697312f,
653 -0.650255f, -0.703877f, 0.243402f, 0.426710f, 0.217210f, 0.260255f,
654 0.027416f, 0.163147f, 0.132188f, 0.142374f, 0.558627f, 0.065717f,
655 0.382781f, -1.192240f, 0.195492f, 0.028439f, 0.278252f, -0.491806f,
656 0.497701f, -0.448835f, -0.245079f, -0.014336f, -0.174907f, -0.409633f,
657 0.207548f, 0.433813f, 0.459889f, 0.431728f, 0.605050f, 0.485520f,
658 0.218548f, 0.437307f, 0.027023f, -0.204251f, 0.012100f, 0.150677f,
659 -1.097980f, 0.086866f, -1.293130f, -0.372575f, -0.876264f, -0.021818f,
660 0.322864f, -0.231043f, -0.271608f, 0.132782f, -0.314895f, 0.396800f,
661 0.262788f, -0.317212f, -0.666308f, 0.830742f, 0.319409f, -0.564373f,
662 -0.178656f, 0.306993f, 0.265634f, -0.332480f, -0.491514f, -0.186745f,
663 -0.063044f, -0.009321f, 0.074944f, -0.372082f, -0.029479f, 0.081548f,
664 0.028172f, -0.233148f, -0.337938f, -0.087695f, 0.596556f, 0.559530f,
665 0.139332f, 0.107223f, -0.190915f, 0.137401f, -0.150625f, -0.225484f,
666 -0.191344f, -0.232535f, 0.126510f, 0.296323f, -0.547901f, -0.653080f,
667 0.358514f, 0.726289f, -0.421725f, -0.243620f, 0.236206f, 0.390823f,
668 -0.076560f, -0.282329f, -0.012460f, -0.428484f, 0.349469f, 0.394629f,
669 0.421537f, 0.219632f, -0.117550f, -0.087894f, 0.077155f, 0.016000f,
670 -0.289137f, -0.092937f, -0.014518f, -0.027111f, 0.210329f, -0.159678f,
671 0.013288f, -0.039268f, 0.008112f, 0.003152f, 0.030084f, -0.039859f,
672 0.322028f, -0.407797f, 0.447087f, -0.381562f, 0.529297f, -0.520298f,
673 0.562865f, -0.616878f, 0.689389f, 0.754262f, 0.138475f, 0.750697f,
674 -0.760157f, -0.383740f, 0.074219f, 0.556257f, 0.087827f, -0.511826f,
675 -0.305507f, -0.638214f, 0.114833f, -0.444022f, 0.526612f, -0.604984f,
676 -0.100415f, 0.037824f, -0.106264f, 0.337615f, 0.070743f, 0.031129f,
677 0.281954f, 0.176144f, -0.032833f, -0.073902f, -0.285492f, -0.803803f,
678 -0.015589f, 0.186077f, -0.033351f, 0.517269f, -1.878800f, -1.685210f,
679 -0.416581f, 0.158476f, -0.071929f, -0.624353f, -0.122069f, -0.075065f,
680 0.311816f, 0.506305f, 0.383896f, 0.259450f, -0.308232f, -0.094221f,
681 -0.421885f, -0.293573f,
682 };
683
684 static const float weights_layer_5[] = {
685 0.131894f, 0.078431f, 0.323121f, -0.230680f, -0.684740f, 0.020895f,
686 0.364983f, 0.121656f, 0.132448f, -0.731198f, 0.071148f, 0.739642f,
687 0.318437f, -0.033021f, -1.037080f, 0.135335f, 0.383582f, 0.287332f,
688 0.054042f, -0.825482f, 0.418533f, 0.305606f, 0.041549f, 0.432422f,
689 -0.826878f, -0.593536f, 0.105657f, 0.125357f, 0.408567f, -0.293338f,
690 0.233905f, -0.039609f, 0.547727f, -0.435806f, 0.036160f, 0.220275f,
691 -0.020337f, -0.619403f, -0.455858f, 0.681455f, 0.543846f, -0.495084f,
692 0.251496f, -0.085686f, 0.091395f, -0.476696f, 0.453628f, -0.109663f,
693 0.383493f, -0.456563f, -0.212935f, 0.020567f, -0.719564f, -0.377813f,
694 -0.737511f, 0.765965f, 0.624309f, -0.063679f, -0.055681f, -0.475969f,
695 -0.069902f, 0.725690f, 0.641094f, 0.439922f, -0.111544f, -0.309061f,
696 0.280091f, 0.381416f, 0.481168f, 0.483543f, -0.901267f, -0.499230f,
697 0.043449f, -0.372395f, 0.021216f, -0.002200f, -0.524089f, -0.071485f,
698 -0.273974f, -0.462654f, 0.042369f, -0.138679f, -0.330060f, 0.021886f,
699 -0.306075f, -0.011130f, -0.260224f, -0.288435f, -0.104039f, -0.183563f,
700 0.118990f, -0.531160f, 0.339632f, -0.028374f, 0.159084f, -0.008824f,
701 -0.791388f, 0.245242f, 0.356510f, 0.469867f, -0.396949f, -0.476146f,
702 -0.168472f, 1.068400f, 0.474629f, -0.117554f, -0.142453f, -0.306604f,
703 0.348525f, -0.111929f, -0.435384f, 0.019952f, -0.260185f, 0.373376f,
704 0.109729f, -0.639168f, 0.033392f, -0.082573f, -0.196018f, 0.301637f,
705 -0.124210f, -0.202515f, -1.221920f, -0.253690f, -0.144864f, 0.287753f,
706 -0.161206f, -0.213246f, 0.373968f, 0.141397f, -0.248237f, 0.283090f,
707 -0.008977f, -0.172960f, -0.234146f, -0.720014f, -0.322451f, 0.181083f,
708 0.310659f, -0.422646f, -0.719994f, -0.354339f, 0.352739f, 0.230923f,
709 0.427013f, -0.660316f, 0.232140f, 0.685896f, 0.660208f, 0.225748f,
710 -0.918750f, -0.650790f, -0.674525f, -0.450305f, -0.152529f, 0.498480f,
711 0.895092f, 0.688242f, 0.669057f, 0.612669f, 0.593484f, 0.318204f,
712 -0.169294f, 0.388789f, -0.529777f, -0.219706f, -0.044916f, 0.161697f,
713 -0.145288f, 0.196153f, -0.022212f, -0.434209f, -0.208115f, -0.117745f,
714 -0.279029f, -0.009506f, 0.137474f, 0.330148f, 0.439258f, 0.345879f,
715 -0.845131f, -0.215713f, 0.094463f, 0.638604f, 0.882254f, -0.964082f,
716 -0.383920f, 0.292645f, 0.266341f, 0.747473f, -0.645631f, -0.538896f,
717 -0.319764f, 0.521880f, 0.460091f, -0.470898f, -0.778283f, -0.061622f,
718 -0.142433f, 0.210520f, 0.804197f, 0.285840f, -0.138414f, -0.381846f,
719 -0.499991f, 0.223648f, 0.439025f, 0.321508f, -0.099560f, -0.622893f,
720 0.750925f, 0.740994f, 0.140405f, 0.074631f, -0.270223f, -0.829049f,
721 -0.753355f, -0.258015f, 0.006285f, -0.730573f, -1.107390f, -0.538015f,
722 -1.005520f, -0.724115f, -0.440183f, -0.395239f, 0.508768f, 0.204620f,
723 -0.267331f, 0.001740f, -0.838709f, 0.659333f, 0.043739f, -0.024099f,
724 0.262431f, 0.252433f, -0.265215f, 0.057289f, -0.428192f, -0.114350f,
725 -0.011475f, 0.463995f, 0.668833f, -0.604556f, -0.122780f, -0.441645f,
726 0.145769f, 0.310450f, -1.003500f, 0.936069f, 0.516604f, -0.643386f,
727 -0.518571f, 0.306130f, 0.337387f, 0.583400f, -0.366025f, -0.560035f,
728 -0.262332f, 0.465242f, 0.964332f, -0.545410f, -0.637428f, -0.202695f,
729 0.378931f, 0.834604f, 0.000970f, -0.553303f, -0.562879f, 0.221665f,
730 0.395160f, 0.446281f, -0.184394f, -0.591780f, 0.170595f, 1.164390f,
731 0.227068f, -0.150910f, -0.393690f, -0.131151f, 0.309956f, -0.413518f,
732 -0.768334f, -0.548975f, 0.245384f, -0.256904f, -0.514790f, -0.102616f,
733 -0.347625f, 0.420456f, 0.037804f, -0.283200f, -0.578815f, 0.319282f,
734 0.674622f, -0.011791f, -0.339329f, 0.466705f, 0.563444f, 0.409660f,
735 0.445784f, -0.899507f, -0.605116f, 0.622438f, 0.427385f, -0.062509f,
736 0.666570f, 0.057105f, 0.357894f, -0.811016f, -0.421715f, -0.458397f,
737 0.288955f, 0.005857f, 0.236331f, 0.107957f, 0.587276f, -0.375800f,
738 0.323799f, -0.623363f, 0.254122f, -0.198478f, -0.098436f, -0.282531f,
739 0.452453f, -0.163349f, -0.413382f, -0.448732f, -0.528770f, -0.457449f,
740 -0.619619f, -0.265919f, -0.042760f, 0.438730f, 0.501798f, -0.403851f,
741 0.519564f, 0.817314f, 0.366203f, 0.492610f, 0.546929f, 0.853094f,
742 0.289000f, 0.453941f, -0.076152f, 0.007226f, -0.183717f, -0.506252f,
743 -0.599989f, -0.576006f, 0.746488f, 0.631466f, -0.475599f, -0.334991f,
744 -0.879614f, 0.918957f, 0.473471f, -0.043781f, -0.688234f, -0.925875f,
745 -0.188081f, 0.050918f, 0.116855f, 0.221413f, -0.066680f, -0.674395f,
746 -0.481985f, 0.247368f, 0.271129f, 0.637979f, -1.006970f, -0.855441f,
747 0.144874f, 0.507424f, 1.506960f, -0.338910f, 0.398203f, 0.738000f,
748 0.263193f, -0.425908f, 0.358271f, -1.072900f, -0.816209f, -0.425519f,
749 0.264373f, 0.694014f, 0.036333f, 0.635532f, 0.518856f, 0.047585f,
750 -0.854817f, -0.138202f, 0.006811f, -0.052020f, -0.468498f, 0.489080f,
751 -0.105778f, 0.357038f, -0.782875f, 0.649049f, -0.562652f, -0.544392f,
752 -0.328526f, -0.402121f, -0.263172f, -0.668459f, -0.526702f, -0.395829f,
753 0.190986f, 0.307766f, -1.001830f, -0.293051f, 0.283334f, 0.572450f,
754 0.906095f, -1.144300f, 0.180989f, 0.421092f, 0.684571f, 0.527276f,
755 -0.122287f, 0.575067f, 0.675221f, 0.755029f, 0.094957f, 0.481403f,
756 0.825155f, 0.755035f, 0.641420f, 0.034497f, 0.518783f, 0.283800f,
757 0.293733f, -0.074778f, -0.268720f, 0.798921f, 0.317714f, -0.236391f,
758 -0.375071f, -0.414600f, 0.223413f, -0.349044f, -0.191033f, -0.391779f,
759 -0.596894f, -0.378608f, -0.185920f, -0.822171f, -0.754962f, -0.167706f,
760 0.755378f, 0.671847f, 0.969414f, 0.793048f, 1.078610f, -0.418963f,
761 0.367648f, 0.217645f, 0.294232f, 0.113027f, 0.060312f, -0.327488f,
762 -0.305035f, -0.243600f, -0.020588f, -0.326324f, -0.417534f, -0.425868f,
763 -0.404614f, -0.346750f, -0.339145f, -0.348094f, -0.527290f, -0.617825f,
764 -0.258342f, -0.200753f, -0.249779f, -0.321039f, -0.023117f, -0.004167f,
765 -0.206788f, -0.612420f, -0.646428f, -0.548969f, -0.158875f, 0.213814f,
766 -0.084040f, -0.217365f, -0.511895f, -0.653285f, 0.440971f, 0.455591f,
767 -0.123900f, 0.134097f, -0.251241f, 0.682463f, 0.740614f, 0.991212f,
768 0.565984f, 0.592690f,
769 };
770
add_f32x4_x4(const float32x4_t a[4])771 static inline float32x4_t add_f32x4_x4(const float32x4_t a[4]) {
772 float32x4_t sum01 = vaddq_f32(a[0], a[1]);
773 float32x4_t sum23 = vaddq_f32(a[2], a[3]);
774 return vaddq_f32(sum01, sum23);
775 }
776
av1_cnn_convolve_no_maxpool_padding_valid_2x2_large_neon(const float ** input,int in_width,int in_height,int in_stride,const float * bias,const int skip_width,const int skip_height,const int filter_width,const int filter_height,const int in_channels,const int out_channels,float ** output,int out_stride,int start_idx,const float * weights)777 static inline void av1_cnn_convolve_no_maxpool_padding_valid_2x2_large_neon(
778 const float **input, int in_width, int in_height, int in_stride,
779 const float *bias, const int skip_width, const int skip_height,
780 const int filter_width, const int filter_height, const int in_channels,
781 const int out_channels, float **output, int out_stride, int start_idx,
782 const float *weights) {
783 assert(filter_height == 2 && filter_width == 2);
784 assert(skip_width == 2 && skip_height == 2);
785 assert(in_width >= 16);
786 const int in_size = in_height * in_width;
787
788 do {
789 const float32x4_t bias_v = vdupq_n_f32(bias[0]);
790 const float *weight_ptr0 = weights;
791 const float *in_ptr0 = *input;
792 float *out_ptr0 = *output;
793 int h = 0;
794
795 do {
796 const float *in_ptr1 = in_ptr0;
797 float *out_ptr1 = out_ptr0;
798 int w = 0;
799
800 do {
801 const float *weight_ptr1 = weight_ptr0;
802 const float *in_ptr2 = in_ptr1;
803 int k = 0;
804 float32x4_t sum0[4] = { bias_v, vdupq_n_f32(0), vdupq_n_f32(0),
805 vdupq_n_f32(0) };
806 float32x4_t sum1[4] = { bias_v, vdupq_n_f32(0), vdupq_n_f32(0),
807 vdupq_n_f32(0) };
808
809 do {
810 const float32x4_t weights0 = vld1q_f32(weight_ptr1);
811 const float32x4_t weights1 = vld1q_f32(weight_ptr1 + 4);
812 const float32x2_t weights0_lo = vget_low_f32(weights0);
813 const float32x2_t weights0_hi = vget_high_f32(weights0);
814 const float32x2_t weights1_lo = vget_low_f32(weights1);
815 const float32x2_t weights1_hi = vget_high_f32(weights1);
816
817 const float32x4x2_t in0_lo_0 = vld2q_f32(in_ptr2);
818 const float32x4x2_t in0_hi_0 = vld2q_f32(in_ptr2 + in_stride);
819 const float32x4x2_t in1_lo_0 = vld2q_f32(in_ptr2 + in_size);
820 const float32x4x2_t in1_hi_0 =
821 vld2q_f32(in_ptr2 + in_size + in_stride);
822
823 sum0[0] = vmlaq_lane_f32(sum0[0], in0_lo_0.val[0], weights0_lo, 0);
824 sum0[0] = vmlaq_lane_f32(sum0[0], in0_lo_0.val[1], weights0_lo, 1);
825
826 sum0[1] = vmlaq_lane_f32(sum0[1], in0_hi_0.val[0], weights0_hi, 0);
827 sum0[1] = vmlaq_lane_f32(sum0[1], in0_hi_0.val[1], weights0_hi, 1);
828
829 sum0[2] = vmlaq_lane_f32(sum0[2], in1_lo_0.val[0], weights1_lo, 0);
830 sum0[2] = vmlaq_lane_f32(sum0[2], in1_lo_0.val[1], weights1_lo, 1);
831
832 sum0[3] = vmlaq_lane_f32(sum0[3], in1_hi_0.val[0], weights1_hi, 0);
833 sum0[3] = vmlaq_lane_f32(sum0[3], in1_hi_0.val[1], weights1_hi, 1);
834
835 const float32x4x2_t in0_lo_1 = vld2q_f32(in_ptr2 + 8);
836 const float32x4x2_t in0_hi_1 = vld2q_f32(in_ptr2 + in_stride + 8);
837 const float32x4x2_t in1_lo_1 = vld2q_f32(in_ptr2 + in_size + 8);
838 const float32x4x2_t in1_hi_1 =
839 vld2q_f32(in_ptr2 + in_size + in_stride + 8);
840
841 sum1[0] = vmlaq_lane_f32(sum1[0], in0_lo_1.val[0], weights0_lo, 0);
842 sum1[0] = vmlaq_lane_f32(sum1[0], in0_lo_1.val[1], weights0_lo, 1);
843
844 sum1[1] = vmlaq_lane_f32(sum1[1], in0_hi_1.val[0], weights0_hi, 0);
845 sum1[1] = vmlaq_lane_f32(sum1[1], in0_hi_1.val[1], weights0_hi, 1);
846
847 sum1[2] = vmlaq_lane_f32(sum1[2], in1_lo_1.val[0], weights1_lo, 0);
848 sum1[2] = vmlaq_lane_f32(sum1[2], in1_lo_1.val[1], weights1_lo, 1);
849
850 sum1[3] = vmlaq_lane_f32(sum1[3], in1_hi_1.val[0], weights1_hi, 0);
851 sum1[3] = vmlaq_lane_f32(sum1[3], in1_hi_1.val[1], weights1_hi, 1);
852
853 weight_ptr1 += 8;
854 in_ptr2 += 2 * in_size;
855 k += 2;
856 } while (k < in_channels);
857
858 vst1q_f32(out_ptr1, add_f32x4_x4(sum0));
859 vst1q_f32(out_ptr1 + 4, add_f32x4_x4(sum1));
860
861 out_ptr1 += 8;
862 in_ptr1 += 8 * skip_width;
863 w += 8 * skip_width;
864 } while (w < in_width - filter_width + 1);
865
866 out_ptr0 += out_stride;
867 in_ptr0 += skip_height * in_stride;
868 h += skip_height;
869 } while (h < in_height - filter_height + 1);
870
871 ++bias;
872 ++output;
873 weights += in_channels * filter_height * filter_width;
874 } while (++start_idx < out_channels);
875 }
876
av1_cnn_convolve_no_maxpool_padding_valid_2x2_neon(const float ** input,int in_width,int in_height,int in_stride,const float * bias,const int skip_width,const int skip_height,const int filter_width,const int filter_height,const int in_channels,const int out_channels,float ** output,int out_stride,int start_idx,const float * weights)877 static inline void av1_cnn_convolve_no_maxpool_padding_valid_2x2_neon(
878 const float **input, int in_width, int in_height, int in_stride,
879 const float *bias, const int skip_width, const int skip_height,
880 const int filter_width, const int filter_height, const int in_channels,
881 const int out_channels, float **output, int out_stride, int start_idx,
882 const float *weights) {
883 assert(filter_height == 2 && filter_width == 2);
884 assert(skip_width == 2 && skip_height == 2);
885 assert(in_width == 8);
886 const int in_size = in_height * in_width;
887 do {
888 const float32x4_t bias_v = vdupq_n_f32(*bias);
889 const float *weight_ptr0 = weights;
890 const float *in_ptr0 = *input;
891 float *out_ptr0 = *output;
892 int h = 0;
893
894 do {
895 const float *in_ptr1 = in_ptr0;
896 float *out_ptr1 = out_ptr0;
897 int w = 0;
898
899 do {
900 const float *weight_ptr1 = weight_ptr0;
901 const float *in_ptr2 = in_ptr1;
902 int k = 0;
903 float32x4_t sum[4] = { bias_v, vdupq_n_f32(0), vdupq_n_f32(0),
904 vdupq_n_f32(0) };
905
906 do {
907 const float32x4_t weights0 = vld1q_f32(weight_ptr1);
908 const float32x4_t weights1 = vld1q_f32(weight_ptr1 + 4);
909 const float32x2_t weights0_lo = vget_low_f32(weights0);
910 const float32x2_t weights0_hi = vget_high_f32(weights0);
911 const float32x2_t weights1_lo = vget_low_f32(weights1);
912 const float32x2_t weights1_hi = vget_high_f32(weights1);
913
914 const float32x4x2_t in0_lo = vld2q_f32(in_ptr2);
915 const float32x4x2_t in0_hi = vld2q_f32(in_ptr2 + in_stride);
916 const float32x4x2_t in1_lo = vld2q_f32(in_ptr2 + in_size);
917 const float32x4x2_t in1_hi = vld2q_f32(in_ptr2 + in_size + in_stride);
918
919 sum[0] = vmlaq_lane_f32(sum[0], in0_lo.val[0], weights0_lo, 0);
920 sum[0] = vmlaq_lane_f32(sum[0], in0_lo.val[1], weights0_lo, 1);
921
922 sum[1] = vmlaq_lane_f32(sum[1], in0_hi.val[0], weights0_hi, 0);
923 sum[1] = vmlaq_lane_f32(sum[1], in0_hi.val[1], weights0_hi, 1);
924
925 sum[2] = vmlaq_lane_f32(sum[2], in1_lo.val[0], weights1_lo, 0);
926 sum[2] = vmlaq_lane_f32(sum[2], in1_lo.val[1], weights1_lo, 1);
927
928 sum[3] = vmlaq_lane_f32(sum[3], in1_hi.val[0], weights1_hi, 0);
929 sum[3] = vmlaq_lane_f32(sum[3], in1_hi.val[1], weights1_hi, 1);
930
931 weight_ptr1 += 8;
932 in_ptr2 += 2 * in_size;
933 k += 2;
934 } while (k < in_channels);
935
936 vst1q_f32(out_ptr1, add_f32x4_x4(sum));
937
938 out_ptr1 += 4;
939 in_ptr1 += 4 * skip_width;
940 w += 4 * skip_width;
941 } while (w < in_width - filter_width + 1);
942
943 out_ptr0 += out_stride;
944 in_ptr0 += skip_height * in_stride;
945 h += skip_height;
946 } while (h < in_height - filter_height + 1);
947
948 ++bias;
949 ++output;
950 weights += in_channels * filter_height * filter_width;
951 } while (++start_idx < out_channels);
952 }
953
av1_cnn_convolve_no_maxpool_padding_valid_5x5_neon(const float ** input,int in_width,int in_height,int in_stride,const float * bias,const int skip_width,const int skip_height,const int filter_width,const int filter_height,const int in_channels,const int out_channels,float ** output,int out_stride,int start_idx,const float * weights)954 static inline void av1_cnn_convolve_no_maxpool_padding_valid_5x5_neon(
955 const float **input, int in_width, int in_height, int in_stride,
956 const float *bias, const int skip_width, const int skip_height,
957 const int filter_width, const int filter_height, const int in_channels,
958 const int out_channels, float **output, int out_stride, int start_idx,
959 const float *weights) {
960 assert(filter_height == 5 && filter_width == 5);
961 assert(skip_width == 4 && skip_height == 4);
962 assert(in_width >= 16);
963 assert(in_channels == 1);
964 (void)in_channels;
965
966 do {
967 const float32x4_t bias_v = vdupq_n_f32(*bias);
968 const float *in_ptr0 = *input;
969 const float *weights_ptr0 = weights;
970 float *out_ptr0 = *output;
971 int h = 0;
972
973 do {
974 const float *in_ptr1 = in_ptr0;
975 float *out_ptr1 = out_ptr0;
976 int w = 0;
977
978 do {
979 float32x4_t sum[2] = { bias_v, vdupq_n_f32(0) };
980
981 const float32x4_t weight_0_3 = vld1q_f32(weights_ptr0);
982 const float32x4_t weight_4_7 = vld1q_f32(weights_ptr0 + 4);
983 const float32x4_t weight_8_11 = vld1q_f32(weights_ptr0 + 8);
984 const float32x4_t weight_12_15 = vld1q_f32(weights_ptr0 + 12);
985 const float32x4_t weight_16_19 = vld1q_f32(weights_ptr0 + 16);
986 const float32x4_t weight_20_23 = vld1q_f32(weights_ptr0 + 20);
987
988 const float32x2_t weight_0_3_lo = vget_low_f32(weight_0_3);
989 const float32x2_t weight_0_3_hi = vget_high_f32(weight_0_3);
990 const float32x2_t weight_4_7_lo = vget_low_f32(weight_4_7);
991 const float32x2_t weight_4_7_hi = vget_high_f32(weight_4_7);
992 const float32x2_t weight_8_11_lo = vget_low_f32(weight_8_11);
993 const float32x2_t weight_8_11_hi = vget_high_f32(weight_8_11);
994 const float32x2_t weight_12_15_lo = vget_low_f32(weight_12_15);
995 const float32x2_t weight_12_15_hi = vget_high_f32(weight_12_15);
996 const float32x2_t weight_16_19_lo = vget_low_f32(weight_16_19);
997 const float32x2_t weight_16_19_hi = vget_high_f32(weight_16_19);
998 const float32x2_t weight_20_23_lo = vget_low_f32(weight_20_23);
999 const float32x2_t weight_20_23_hi = vget_high_f32(weight_20_23);
1000
1001 const float32x4x4_t in0 = vld4q_f32(in_ptr1 + 0 * in_stride);
1002 const float32x4x4_t in1 = vld4q_f32(in_ptr1 + 1 * in_stride);
1003 const float32x4x4_t in2 = vld4q_f32(in_ptr1 + 2 * in_stride);
1004 const float32x4x4_t in3 = vld4q_f32(in_ptr1 + 3 * in_stride);
1005 const float32x4x4_t in4 = vld4q_f32(in_ptr1 + 4 * in_stride);
1006
1007 const float32x4_t in0_4 = vextq_f32(
1008 in0.val[0], vdupq_n_f32(*(in_ptr1 + 16 + 0 * in_stride)), 1);
1009 const float32x4_t in1_4 = vextq_f32(
1010 in1.val[0], vdupq_n_f32(*(in_ptr1 + 16 + 1 * in_stride)), 1);
1011 const float32x4_t in2_4 = vextq_f32(
1012 in2.val[0], vdupq_n_f32(*(in_ptr1 + 16 + 2 * in_stride)), 1);
1013 const float32x4_t in3_4 = vextq_f32(
1014 in3.val[0], vdupq_n_f32(*(in_ptr1 + 16 + 3 * in_stride)), 1);
1015 const float32x4_t in4_4 = vextq_f32(
1016 in4.val[0], vdupq_n_f32(*(in_ptr1 + 16 + 4 * in_stride)), 1);
1017
1018 // Kernel row 0.
1019 sum[0] = vmlaq_lane_f32(sum[0], in0.val[0], weight_0_3_lo, 0);
1020 sum[1] = vmlaq_lane_f32(sum[1], in0.val[1], weight_0_3_lo, 1);
1021 sum[0] = vmlaq_lane_f32(sum[0], in0.val[2], weight_0_3_hi, 0);
1022 sum[1] = vmlaq_lane_f32(sum[1], in0.val[3], weight_0_3_hi, 1);
1023 sum[0] = vmlaq_lane_f32(sum[0], in0_4, weight_4_7_lo, 0);
1024
1025 // Kernel row 1.
1026 sum[1] = vmlaq_lane_f32(sum[1], in1.val[0], weight_4_7_lo, 1);
1027 sum[0] = vmlaq_lane_f32(sum[0], in1.val[1], weight_4_7_hi, 0);
1028 sum[1] = vmlaq_lane_f32(sum[1], in1.val[2], weight_4_7_hi, 1);
1029 sum[0] = vmlaq_lane_f32(sum[0], in1.val[3], weight_8_11_lo, 0);
1030 sum[1] = vmlaq_lane_f32(sum[1], in1_4, weight_8_11_lo, 1);
1031
1032 // Kernel row 2.
1033 sum[0] = vmlaq_lane_f32(sum[0], in2.val[0], weight_8_11_hi, 0);
1034 sum[1] = vmlaq_lane_f32(sum[1], in2.val[1], weight_8_11_hi, 1);
1035 sum[0] = vmlaq_lane_f32(sum[0], in2.val[2], weight_12_15_lo, 0);
1036 sum[1] = vmlaq_lane_f32(sum[1], in2.val[3], weight_12_15_lo, 1);
1037 sum[0] = vmlaq_lane_f32(sum[0], in2_4, weight_12_15_hi, 0);
1038
1039 // Kernel row 3.
1040 sum[1] = vmlaq_lane_f32(sum[1], in3.val[0], weight_12_15_hi, 1);
1041 sum[0] = vmlaq_lane_f32(sum[0], in3.val[1], weight_16_19_lo, 0);
1042 sum[1] = vmlaq_lane_f32(sum[1], in3.val[2], weight_16_19_lo, 1);
1043 sum[0] = vmlaq_lane_f32(sum[0], in3.val[3], weight_16_19_hi, 0);
1044 sum[1] = vmlaq_lane_f32(sum[1], in3_4, weight_16_19_hi, 1);
1045
1046 // Kernel row 4.
1047 sum[0] = vmlaq_lane_f32(sum[0], in4.val[0], weight_20_23_lo, 0);
1048 sum[1] = vmlaq_lane_f32(sum[1], in4.val[1], weight_20_23_lo, 1);
1049 sum[0] = vmlaq_lane_f32(sum[0], in4.val[2], weight_20_23_hi, 0);
1050 sum[1] = vmlaq_lane_f32(sum[1], in4.val[3], weight_20_23_hi, 1);
1051 sum[0] = vmlaq_f32(sum[0], vdupq_n_f32(*(weights_ptr0 + 24)), in4_4);
1052
1053 vst1q_f32(out_ptr1, vaddq_f32(sum[0], sum[1]));
1054
1055 out_ptr1 += 4;
1056 in_ptr1 += 4 * skip_width;
1057 w += 4 * skip_width;
1058 } while (w < in_width - filter_width + 1);
1059
1060 out_ptr0 += out_stride;
1061 in_ptr0 += skip_height * in_stride;
1062 h += skip_height;
1063 } while (h < in_height - filter_height + 1);
1064
1065 ++output;
1066 ++bias;
1067 weights += 25;
1068 } while (++start_idx < out_channels);
1069 }
1070
1071 // Neon variant of av1_cnn_convolve_no_maxpool_padding_valid_c().
1072 // As per the current encoder, av1_cnn_convolve function gets called for
1073 // block size equal to 64x64. av1_cnn_convolve() uses layer config values
1074 // set by av1_intra_mode_cnn_partition_cnn_config. The following are a few
1075 // details related to each layer's config parameters.
1076 // Layer_Number in_size out_size filter_wd filter_ht skip_wd skip_ht
1077 // 0 64x64 16x16 5 5 4 4
1078 // 1 16x16 8x8 2 2 2 2
1079 // 2 8x8 4x4 2 2 2 2
1080 // 3 4x4 2x2 2 2 2 2
1081 // 4 2x2 1x1 2 2 2 2
1082 // Here,
1083 // filter_wd = filter_width and filter_ht = filter_height,
1084 // skip_wd = skip_width and skip_ht = skip_height.
av1_cnn_convolve_no_maxpool_padding_valid_neon(const float ** input,int in_width,int in_height,int in_stride,const CNN_LAYER_CONFIG * layer_config,float ** output,int out_stride,int start_idx,int cstep,int channel_step)1085 void av1_cnn_convolve_no_maxpool_padding_valid_neon(
1086 const float **input, int in_width, int in_height, int in_stride,
1087 const CNN_LAYER_CONFIG *layer_config, float **output, int out_stride,
1088 int start_idx, int cstep, int channel_step) {
1089 assert((layer_config->skip_height == 1 && layer_config->skip_width == 1) ||
1090 !layer_config->maxpool);
1091 assert(layer_config->filter_height > 1 || layer_config->filter_width > 1);
1092 assert(layer_config->pad == PADDING_VALID);
1093 assert(channel_step == 1);
1094 assert(cstep == layer_config->in_channels * layer_config->out_channels);
1095
1096 if (layer_config->filter_width == 5 && layer_config->filter_height == 5 &&
1097 layer_config->skip_width == 4 && layer_config->skip_height == 4) {
1098 av1_cnn_convolve_no_maxpool_padding_valid_5x5_neon(
1099 input, in_width, in_height, in_stride, layer_config->bias,
1100 layer_config->skip_width, layer_config->skip_height,
1101 layer_config->filter_width, layer_config->filter_height,
1102 layer_config->in_channels, layer_config->out_channels, output,
1103 out_stride, start_idx, weights_layer_5);
1104 } else if (layer_config->filter_width == 2 &&
1105 layer_config->filter_height == 2 &&
1106 layer_config->skip_width == 2 && layer_config->skip_height == 2) {
1107 const float *weights = weights_layer_1;
1108 if (layer_config->output_num ==
1109 av1_intra_mode_cnn_partition_cnn_config.layer_config[2].output_num) {
1110 weights = weights_layer_2;
1111 } else if ((layer_config->output_num ==
1112 av1_intra_mode_cnn_partition_cnn_config.layer_config[3]
1113 .output_num)) {
1114 weights = weights_layer_3;
1115 } else if ((layer_config->output_num ==
1116 av1_intra_mode_cnn_partition_cnn_config.layer_config[4]
1117 .output_num)) {
1118 weights = weights_layer_4;
1119 }
1120 if (in_width >= 16) {
1121 av1_cnn_convolve_no_maxpool_padding_valid_2x2_large_neon(
1122 input, in_width, in_height, in_stride, layer_config->bias,
1123 layer_config->skip_width, layer_config->skip_height,
1124 layer_config->filter_width, layer_config->filter_height,
1125 layer_config->in_channels, layer_config->out_channels, output,
1126 out_stride, start_idx, weights);
1127 } else if (in_width == 8) {
1128 av1_cnn_convolve_no_maxpool_padding_valid_2x2_neon(
1129 input, in_width, in_height, in_stride, layer_config->bias,
1130 layer_config->skip_width, layer_config->skip_height,
1131 layer_config->filter_width, layer_config->filter_height,
1132 layer_config->in_channels, layer_config->out_channels, output,
1133 out_stride, start_idx, weights);
1134 } else {
1135 av1_cnn_convolve_no_maxpool_padding_valid_c(
1136 input, in_width, in_height, in_stride, layer_config, output,
1137 out_stride, start_idx, cstep, channel_step);
1138 }
1139 } else {
1140 av1_cnn_convolve_no_maxpool_padding_valid_c(
1141 input, in_width, in_height, in_stride, layer_config, output, out_stride,
1142 start_idx, cstep, channel_step);
1143 }
1144 }
1145