1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Rockchip RK3288 VPU codec driver
4 *
5 * Copyright (c) 2014 Rockchip Electronics Co., Ltd.
6 * Hertz Wong <hertz.wong@rock-chips.com>
7 * Herman Chen <herman.chen@rock-chips.com>
8 *
9 * Copyright (C) 2014 Google, Inc.
10 * Tomasz Figa <tfiga@chromium.org>
11 */
12
13 #include <linux/types.h>
14 #include <media/v4l2-h264.h>
15 #include <media/v4l2-mem2mem.h>
16
17 #include "hantro.h"
18 #include "hantro_hw.h"
19
20 /* Size with u32 units. */
21 #define CABAC_INIT_BUFFER_SIZE (460 * 2)
22 #define POC_BUFFER_SIZE 34
23 #define SCALING_LIST_SIZE (6 * 16 + 2 * 64)
24
25 /* Data structure describing auxiliary buffer format. */
26 struct hantro_h264_dec_priv_tbl {
27 u32 cabac_table[CABAC_INIT_BUFFER_SIZE];
28 u32 poc[POC_BUFFER_SIZE];
29 u8 scaling_list[SCALING_LIST_SIZE];
30 };
31
32 /*
33 * Constant CABAC table.
34 * From drivers/media/platform/rk3288-vpu/rk3288_vpu_hw_h264d.c
35 * in https://chromium.googlesource.com/chromiumos/third_party/kernel,
36 * chromeos-3.14 branch.
37 */
38 static const u32 h264_cabac_table[] = {
39 0x14f10236, 0x034a14f1, 0x0236034a, 0xe47fe968, 0xfa35ff36, 0x07330000,
40 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
41 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
42 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
43 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
44 0x0029003f, 0x003f003f, 0xf7530456, 0x0061f948, 0x0d29033e, 0x000b0137,
45 0x0045ef7f, 0xf3660052, 0xf94aeb6b, 0xe57fe17f, 0xe87fee5f, 0xe57feb72,
46 0xe27fef7b, 0xf473f07a, 0xf573f43f, 0xfe44f154, 0xf368fd46, 0xf85df65a,
47 0xe27fff4a, 0xfa61f95b, 0xec7ffc38, 0xfb52f94c, 0xea7df95d, 0xf557fd4d,
48 0xfb47fc3f, 0xfc44f454, 0xf93ef941, 0x083d0538, 0xfe420140, 0x003dfe4e,
49 0x01320734, 0x0a23002c, 0x0b26012d, 0x002e052c, 0x1f110133, 0x07321c13,
50 0x10210e3e, 0xf36cf164, 0xf365f35b, 0xf45ef658, 0xf054f656, 0xf953f357,
51 0xed5e0146, 0x0048fb4a, 0x123bf866, 0xf164005f, 0xfc4b0248, 0xf54bfd47,
52 0x0f2ef345, 0x003e0041, 0x1525f148, 0x09391036, 0x003e0c48, 0x18000f09,
53 0x08190d12, 0x0f090d13, 0x0a250c12, 0x061d1421, 0x0f1e042d, 0x013a003e,
54 0x073d0c26, 0x0b2d0f27, 0x0b2a0d2c, 0x102d0c29, 0x0a311e22, 0x122a0a37,
55 0x1133112e, 0x00591aed, 0x16ef1aef, 0x1ee71cec, 0x21e925e5, 0x21e928e4,
56 0x26ef21f5, 0x28f129fa, 0x26012911, 0x1efa1b03, 0x1a1625f0, 0x23fc26f8,
57 0x26fd2503, 0x26052a00, 0x23102716, 0x0e301b25, 0x153c0c44, 0x0261fd47,
58 0xfa2afb32, 0xfd36fe3e, 0x003a013f, 0xfe48ff4a, 0xf75bfb43, 0xfb1bfd27,
59 0xfe2c002e, 0xf040f844, 0xf64efa4d, 0xf656f45c, 0xf137f63c, 0xfa3efc41,
60 0xf449f84c, 0xf950f758, 0xef6ef561, 0xec54f54f, 0xfa49fc4a, 0xf356f360,
61 0xf561ed75, 0xf84efb21, 0xfc30fe35, 0xfd3ef347, 0xf64ff456, 0xf35af261,
62 0x0000fa5d, 0xfa54f84f, 0x0042ff47, 0x003efe3c, 0xfe3bfb4b, 0xfd3efc3a,
63 0xf742ff4f, 0x00470344, 0x0a2cf93e, 0x0f240e28, 0x101b0c1d, 0x012c1424,
64 0x1220052a, 0x01300a3e, 0x112e0940, 0xf468f561, 0xf060f958, 0xf855f955,
65 0xf755f358, 0x0442fd4d, 0xfd4cfa4c, 0x0a3aff4c, 0xff53f963, 0xf25f025f,
66 0x004cfb4a, 0x0046f54b, 0x01440041, 0xf249033e, 0x043eff44, 0xf34b0b37,
67 0x05400c46, 0x0f060613, 0x07100c0e, 0x120d0d0b, 0x0d0f0f10, 0x0c170d17,
68 0x0f140e1a, 0x0e2c1128, 0x112f1811, 0x15151916, 0x1f1b161d, 0x13230e32,
69 0x0a39073f, 0xfe4dfc52, 0xfd5e0945, 0xf46d24dd, 0x24de20e6, 0x25e22ce0,
70 0x22ee22f1, 0x28f121f9, 0x23fb2100, 0x2602210d, 0x17230d3a, 0x1dfd1a00,
71 0x161e1ff9, 0x23f122fd, 0x220324ff, 0x2205200b, 0x2305220c, 0x270b1e1d,
72 0x221a1d27, 0x13421f15, 0x1f1f1932, 0xef78ec70, 0xee72f555, 0xf15cf259,
73 0xe647f151, 0xf2500044, 0xf246e838, 0xe944e832, 0xf54a17f3, 0x1af328f1,
74 0x31f22c03, 0x2d062c22, 0x21361352, 0xfd4bff17, 0x0122012b, 0x0036fe37,
75 0x003d0140, 0x0044f75c, 0xf26af361, 0xf15af45a, 0xee58f649, 0xf74ff256,
76 0xf649f646, 0xf645fb42, 0xf740fb3a, 0x023b15f6, 0x18f51cf8, 0x1cff1d03,
77 0x1d092314, 0x1d240e43, 0x14f10236, 0x034a14f1, 0x0236034a, 0xe47fe968,
78 0xfa35ff36, 0x07331721, 0x17021500, 0x01090031, 0xdb760539, 0xf34ef541,
79 0x013e0c31, 0xfc491132, 0x1240092b, 0x1d001a43, 0x105a0968, 0xd27fec68,
80 0x0143f34e, 0xf541013e, 0xfa56ef5f, 0xfa3d092d, 0xfd45fa51, 0xf5600637,
81 0x0743fb56, 0x0258003a, 0xfd4cf65e, 0x05360445, 0xfd510058, 0xf943fb4a,
82 0xfc4afb50, 0xf948013a, 0x0029003f, 0x003f003f, 0xf7530456, 0x0061f948,
83 0x0d29033e, 0x002dfc4e, 0xfd60e57e, 0xe462e765, 0xe943e452, 0xec5ef053,
84 0xea6eeb5b, 0xee66f35d, 0xe37ff95c, 0xfb59f960, 0xf36cfd2e, 0xff41ff39,
85 0xf75dfd4a, 0xf75cf857, 0xe97e0536, 0x063c063b, 0x0645ff30, 0x0044fc45,
86 0xf858fe55, 0xfa4eff4b, 0xf94d0236, 0x0532fd44, 0x0132062a, 0xfc51013f,
87 0xfc460043, 0x0239fe4c, 0x0b230440, 0x013d0b23, 0x12190c18, 0x0d1d0d24,
88 0xf65df949, 0xfe490d2e, 0x0931f964, 0x09350235, 0x0535fe3d, 0x00380038,
89 0xf33ffb3c, 0xff3e0439, 0xfa450439, 0x0e270433, 0x0d440340, 0x013d093f,
90 0x07321027, 0x052c0434, 0x0b30fb3c, 0xff3b003b, 0x1621052c, 0x0e2bff4e,
91 0x003c0945, 0x0b1c0228, 0x032c0031, 0x002e022c, 0x0233002f, 0x0427023e,
92 0x062e0036, 0x0336023a, 0x043f0633, 0x06390735, 0x06340637, 0x0b2d0e24,
93 0x0835ff52, 0x0737fd4e, 0x0f2e161f, 0xff541907, 0x1ef91c03, 0x1c042000,
94 0x22ff1e06, 0x1e062009, 0x1f131a1b, 0x1a1e2514, 0x1c221146, 0x0143053b,
95 0x0943101e, 0x12201223, 0x161d181f, 0x1726122b, 0x14290b3f, 0x093b0940,
96 0xff5efe59, 0xf76cfa4c, 0xfe2c002d, 0x0034fd40, 0xfe3bfc46, 0xfc4bf852,
97 0xef66f74d, 0x0318002a, 0x00300037, 0xfa3bf947, 0xf453f557, 0xe277013a,
98 0xfd1dff24, 0x0126022b, 0xfa37003a, 0x0040fd4a, 0xf65a0046, 0xfc1d051f,
99 0x072a013b, 0xfe3afd48, 0xfd51f561, 0x003a0805, 0x0a0e0e12, 0x0d1b0228,
100 0x003afd46, 0xfa4ff855, 0x0000f36a, 0xf06af657, 0xeb72ee6e, 0xf262ea6e,
101 0xeb6aee67, 0xeb6be96c, 0xe670f660, 0xf45ffb5b, 0xf75dea5e, 0xfb560943,
102 0xfc50f655, 0xff46073c, 0x093a053d, 0x0c320f32, 0x12311136, 0x0a29072e,
103 0xff330731, 0x08340929, 0x062f0237, 0x0d290a2c, 0x06320535, 0x0d31043f,
104 0x0640fe45, 0xfe3b0646, 0x0a2c091f, 0x0c2b0335, 0x0e220a26, 0xfd340d28,
105 0x1120072c, 0x07260d32, 0x0a391a2b, 0x0e0b0b0e, 0x090b120b, 0x150917fe,
106 0x20f120f1, 0x22eb27e9, 0x2adf29e1, 0x2ee426f4, 0x151d2de8, 0x35d330e6,
107 0x41d52bed, 0x27f61e09, 0x121a141b, 0x0039f252, 0xfb4bed61, 0xdd7d1b00,
108 0x1c001ffc, 0x1b062208, 0x1e0a1816, 0x21131620, 0x1a1f1529, 0x1a2c172f,
109 0x10410e47, 0x083c063f, 0x11411518, 0x17141a17, 0x1b201c17, 0x1c181728,
110 0x18201c1d, 0x172a1339, 0x1635163d, 0x0b560c28, 0x0b330e3b, 0xfc4ff947,
111 0xfb45f746, 0xf842f644, 0xed49f445, 0xf046f143, 0xec3eed46, 0xf042ea41,
112 0xec3f09fe, 0x1af721f7, 0x27f929fe, 0x2d033109, 0x2d1b243b, 0xfa42f923,
113 0xf92af82d, 0xfb30f438, 0xfa3cfb3e, 0xf842f84c, 0xfb55fa51, 0xf64df951,
114 0xef50ee49, 0xfc4af653, 0xf747f743, 0xff3df842, 0xf242003b, 0x023b15f3,
115 0x21f227f9, 0x2efe3302, 0x3c063d11, 0x37222a3e, 0x14f10236, 0x034a14f1,
116 0x0236034a, 0xe47fe968, 0xfa35ff36, 0x07331619, 0x22001000, 0xfe090429,
117 0xe3760241, 0xfa47f34f, 0x05340932, 0xfd460a36, 0x1a221316, 0x28003902,
118 0x29241a45, 0xd37ff165, 0xfc4cfa47, 0xf34f0534, 0x0645f35a, 0x0034082b,
119 0xfe45fb52, 0xf660023b, 0x024bfd57, 0xfd640138, 0xfd4afa55, 0x003bfd51,
120 0xf956fb5f, 0xff42ff4d, 0x0146fe56, 0xfb48003d, 0x0029003f, 0x003f003f,
121 0xf7530456, 0x0061f948, 0x0d29033e, 0x0d0f0733, 0x0250d97f, 0xee5bef60,
122 0xe651dd62, 0xe866e961, 0xe577e863, 0xeb6eee66, 0xdc7f0050, 0xfb59f95e,
123 0xfc5c0027, 0x0041f154, 0xdd7ffe49, 0xf468f75b, 0xe17f0337, 0x07380737,
124 0x083dfd35, 0x0044f94a, 0xf758f367, 0xf35bf759, 0xf25cf84c, 0xf457e96e,
125 0xe869f64e, 0xec70ef63, 0xb27fba7f, 0xce7fd27f, 0xfc42fb4e, 0xfc47f848,
126 0x023bff37, 0xf946fa4b, 0xf859de77, 0xfd4b2014, 0x1e16d47f, 0x0036fb3d,
127 0x003aff3c, 0xfd3df843, 0xe754f24a, 0xfb410534, 0x0239003d, 0xf745f546,
128 0x1237fc47, 0x003a073d, 0x09291219, 0x0920052b, 0x092f002c, 0x0033022e,
129 0x1326fc42, 0x0f260c2a, 0x09220059, 0x042d0a1c, 0x0a1f21f5, 0x34d5120f,
130 0x1c0023ea, 0x26e72200, 0x27ee20f4, 0x66a20000, 0x38f121fc, 0x1d0a25fb,
131 0x33e327f7, 0x34de45c6, 0x43c12cfb, 0x200737e3, 0x20010000, 0x1b2421e7,
132 0x22e224e4, 0x26e426e5, 0x22ee23f0, 0x22f220f8, 0x25fa2300, 0x1e0a1c12,
133 0x1a191d29, 0x004b0248, 0x084d0e23, 0x121f1123, 0x151e112d, 0x142a122d,
134 0x1b1a1036, 0x07421038, 0x0b490a43, 0xf674e970, 0xf147f93d, 0x0035fb42,
135 0xf54df750, 0xf754f657, 0xde7feb65, 0xfd27fb35, 0xf93df54b, 0xf14def5b,
136 0xe76be76f, 0xe47af54c, 0xf62cf634, 0xf639f73a, 0xf048f945, 0xfc45fb4a,
137 0xf7560242, 0xf7220120, 0x0b1f0534, 0xfe37fe43, 0x0049f859, 0x03340704,
138 0x0a081108, 0x10130325, 0xff3dfb49, 0xff46fc4e, 0x0000eb7e, 0xe97cec6e,
139 0xe67ee77c, 0xef69e579, 0xe575ef66, 0xe675e574, 0xdf7af65f, 0xf264f85f,
140 0xef6fe472, 0xfa59fe50, 0xfc52f755, 0xf851ff48, 0x05400143, 0x09380045,
141 0x01450745, 0xf945fa43, 0xf04dfe40, 0x023dfa43, 0xfd400239, 0xfd41fd42,
142 0x003e0933, 0xff42fe47, 0xfe4bff46, 0xf7480e3c, 0x1025002f, 0x12230b25,
143 0x0c290a29, 0x02300c29, 0x0d29003b, 0x03321328, 0x03421232, 0x13fa12fa,
144 0x0e001af4, 0x1ff021e7, 0x21ea25e4, 0x27e22ae2, 0x2fd62ddc, 0x31de29ef,
145 0x200945b9, 0x3fc142c0, 0x4db636d9, 0x34dd29f6, 0x240028ff, 0x1e0e1c1a,
146 0x17250c37, 0x0b4125df, 0x27dc28db, 0x26e22edf, 0x2ae228e8, 0x31e326f4,
147 0x28f626fd, 0x2efb1f14, 0x1d1e192c, 0x0c300b31, 0x1a2d1616, 0x17161b15,
148 0x21141a1c, 0x1e181b22, 0x122a1927, 0x12320c46, 0x15360e47, 0x0b531920,
149 0x15311536, 0xfb55fa51, 0xf64df951, 0xef50ee49, 0xfc4af653, 0xf747f743,
150 0xff3df842, 0xf242003b, 0x023b11f6, 0x20f32af7, 0x31fb3500, 0x4003440a,
151 0x421b2f39, 0xfb470018, 0xff24fe2a, 0xfe34f739, 0xfa3ffc41, 0xfc43f952,
152 0xfd51fd4c, 0xf948fa4e, 0xf448f244, 0xfd46fa4c, 0xfb42fb3e, 0x0039fc3d,
153 0xf73c0136, 0x023a11f6, 0x20f32af7, 0x31fb3500, 0x4003440a, 0x421b2f39,
154 0x14f10236, 0x034a14f1, 0x0236034a, 0xe47fe968, 0xfa35ff36, 0x07331d10,
155 0x19000e00, 0xf633fd3e, 0xe5631a10, 0xfc55e866, 0x05390639, 0xef490e39,
156 0x1428140a, 0x1d003600, 0x252a0c61, 0xe07fea75, 0xfe4afc55, 0xe8660539,
157 0xfa5df258, 0xfa2c0437, 0xf559f167, 0xeb741339, 0x143a0454, 0x0660013f,
158 0xfb55f36a, 0x053f064b, 0xfd5aff65, 0x0337fc4f, 0xfe4bf461, 0xf932013c,
159 0x0029003f, 0x003f003f, 0xf7530456, 0x0061f948, 0x0d29033e, 0x0722f758,
160 0xec7fdc7f, 0xef5bf25f, 0xe754e756, 0xf459ef5b, 0xe17ff24c, 0xee67f35a,
161 0xdb7f0b50, 0x054c0254, 0x054efa37, 0x043df253, 0xdb7ffb4f, 0xf568f55b,
162 0xe27f0041, 0xfe4f0048, 0xfc5cfa38, 0x0344f847, 0xf362fc56, 0xf458fb52,
163 0xfd48fc43, 0xf848f059, 0xf745ff3b, 0x05420439, 0xfc47fe47, 0x023aff4a,
164 0xfc2cff45, 0x003ef933, 0xfc2ffa2a, 0xfd29fa35, 0x084cf74e, 0xf5530934,
165 0x0043fb5a, 0x0143f148, 0xfb4bf850, 0xeb53eb40, 0xf31fe740, 0xe35e094b,
166 0x113ff84a, 0xfb23fe1b, 0x0d5b0341, 0xf945084d, 0xf642033e, 0xfd44ec51,
167 0x001e0107, 0xfd17eb4a, 0x1042e97c, 0x11252cee, 0x32deea7f, 0x0427002a,
168 0x07220b1d, 0x081f0625, 0x072a0328, 0x08210d2b, 0x0d24042f, 0x0337023a,
169 0x063c082c, 0x0b2c0e2a, 0x07300438, 0x04340d25, 0x0931133a, 0x0a300c2d,
170 0x00451421, 0x083f23ee, 0x21e71cfd, 0x180a1b00, 0x22f234d4, 0x27e81311,
171 0x1f19241d, 0x1821220f, 0x1e141649, 0x1422131f, 0x1b2c1310, 0x0f240f24,
172 0x151c1915, 0x1e141f0c, 0x1b10182a, 0x005d0e38, 0x0f391a26, 0xe87fe873,
173 0xea52f73e, 0x0035003b, 0xf255f359, 0xf35ef55c, 0xe37feb64, 0xf239f443,
174 0xf547f64d, 0xeb55f058, 0xe968f162, 0xdb7ff652, 0xf830f83d, 0xf842f946,
175 0xf24bf64f, 0xf753f45c, 0xee6cfc4f, 0xea45f04b, 0xfe3a013a, 0xf34ef753,
176 0xfc51f363, 0xf351fa26, 0xf33efa3a, 0xfe3bf049, 0xf64cf356, 0xf753f657,
177 0x0000ea7f, 0xe77fe778, 0xe57fed72, 0xe975e776, 0xe675e871, 0xe476e178,
178 0xdb7cf65e, 0xf166f663, 0xf36ace7f, 0xfb5c1139, 0xfb56f35e, 0xf45bfe4d,
179 0x0047ff49, 0x0440f951, 0x05400f39, 0x01430044, 0xf6430144, 0x004d0240,
180 0x0044fb4e, 0x0737053b, 0x02410e36, 0x0f2c053c, 0x0246fe4c, 0xee560c46,
181 0x0540f446, 0x0b370538, 0x00450241, 0xfa4a0536, 0x0736fa4c, 0xf552fe4d,
182 0xfe4d192a, 0x11f310f7, 0x11f41beb, 0x25e229d8, 0x2ad730d1, 0x27e02ed8,
183 0x34cd2ed7, 0x34d92bed, 0x200b3dc9, 0x38d23ece, 0x51bd2dec, 0x23fe1c0f,
184 0x22012701, 0x1e111426, 0x122d0f36, 0x004f24f0, 0x25f225ef, 0x2001220f,
185 0x1d0f1819, 0x22161f10, 0x23121f1c, 0x2129241c, 0x1b2f153e, 0x121f131a,
186 0x24181817, 0x1b10181e, 0x1f1d1629, 0x162a103c, 0x0f340e3c, 0x034ef07b,
187 0x15351638, 0x193d1521, 0x1332113d, 0xfd4ef84a, 0xf748f648, 0xee4bf447,
188 0xf53ffb46, 0xef4bf248, 0xf043f835, 0xf23bf734, 0xf54409fe, 0x1ef61ffc,
189 0x21ff2107, 0x1f0c2517, 0x1f261440, 0xf747f925, 0xf82cf531, 0xf638f43b,
190 0xf83ff743, 0xfa44f64f, 0xfd4ef84a, 0xf748f648, 0xee4bf447, 0xf53ffb46,
191 0xef4bf248, 0xf043f835, 0xf23bf734, 0xf54409fe, 0x1ef61ffc, 0x21ff2107,
192 0x1f0c2517, 0x1f261440
193 };
194
195 static void
assemble_scaling_list(struct hantro_ctx * ctx)196 assemble_scaling_list(struct hantro_ctx *ctx)
197 {
198 const struct hantro_h264_dec_ctrls *ctrls = &ctx->h264_dec.ctrls;
199 const struct v4l2_ctrl_h264_scaling_matrix *scaling = ctrls->scaling;
200 const struct v4l2_ctrl_h264_pps *pps = ctrls->pps;
201 const size_t num_list_4x4 = ARRAY_SIZE(scaling->scaling_list_4x4);
202 const size_t list_len_4x4 = ARRAY_SIZE(scaling->scaling_list_4x4[0]);
203 const size_t list_len_8x8 = ARRAY_SIZE(scaling->scaling_list_8x8[0]);
204 struct hantro_h264_dec_priv_tbl *tbl = ctx->h264_dec.priv.cpu;
205 u32 *dst = (u32 *)tbl->scaling_list;
206 const u32 *src;
207 int i, j;
208
209 if (!(pps->flags & V4L2_H264_PPS_FLAG_SCALING_MATRIX_PRESENT))
210 return;
211
212 for (i = 0; i < num_list_4x4; i++) {
213 src = (u32 *)&scaling->scaling_list_4x4[i];
214 for (j = 0; j < list_len_4x4 / 4; j++)
215 *dst++ = swab32(src[j]);
216 }
217
218 /* Only Intra/Inter Y lists */
219 for (i = 0; i < 2; i++) {
220 src = (u32 *)&scaling->scaling_list_8x8[i];
221 for (j = 0; j < list_len_8x8 / 4; j++)
222 *dst++ = swab32(src[j]);
223 }
224 }
225
prepare_table(struct hantro_ctx * ctx)226 static void prepare_table(struct hantro_ctx *ctx)
227 {
228 const struct hantro_h264_dec_ctrls *ctrls = &ctx->h264_dec.ctrls;
229 const struct v4l2_ctrl_h264_decode_params *dec_param = ctrls->decode;
230 struct hantro_h264_dec_priv_tbl *tbl = ctx->h264_dec.priv.cpu;
231 const struct v4l2_h264_dpb_entry *dpb = ctx->h264_dec.dpb;
232 int i;
233
234 for (i = 0; i < HANTRO_H264_DPB_SIZE; ++i) {
235 tbl->poc[i * 2] = dpb[i].top_field_order_cnt;
236 tbl->poc[i * 2 + 1] = dpb[i].bottom_field_order_cnt;
237 }
238
239 tbl->poc[32] = dec_param->top_field_order_cnt;
240 tbl->poc[33] = dec_param->bottom_field_order_cnt;
241
242 assemble_scaling_list(ctx);
243 }
244
dpb_entry_match(const struct v4l2_h264_dpb_entry * a,const struct v4l2_h264_dpb_entry * b)245 static bool dpb_entry_match(const struct v4l2_h264_dpb_entry *a,
246 const struct v4l2_h264_dpb_entry *b)
247 {
248 return a->top_field_order_cnt == b->top_field_order_cnt &&
249 a->bottom_field_order_cnt == b->bottom_field_order_cnt;
250 }
251
update_dpb(struct hantro_ctx * ctx)252 static void update_dpb(struct hantro_ctx *ctx)
253 {
254 const struct v4l2_ctrl_h264_decode_params *dec_param;
255 DECLARE_BITMAP(new, ARRAY_SIZE(dec_param->dpb)) = { 0, };
256 DECLARE_BITMAP(used, ARRAY_SIZE(dec_param->dpb)) = { 0, };
257 unsigned int i, j;
258
259 dec_param = ctx->h264_dec.ctrls.decode;
260
261 /* Disable all entries by default. */
262 for (i = 0; i < ARRAY_SIZE(ctx->h264_dec.dpb); i++)
263 ctx->h264_dec.dpb[i].flags &= ~V4L2_H264_DPB_ENTRY_FLAG_ACTIVE;
264
265 /* Try to match new DPB entries with existing ones by their POCs. */
266 for (i = 0; i < ARRAY_SIZE(dec_param->dpb); i++) {
267 const struct v4l2_h264_dpb_entry *ndpb = &dec_param->dpb[i];
268
269 if (!(ndpb->flags & V4L2_H264_DPB_ENTRY_FLAG_ACTIVE))
270 continue;
271
272 /*
273 * To cut off some comparisons, iterate only on target DPB
274 * entries which are not used yet.
275 */
276 for_each_clear_bit(j, used, ARRAY_SIZE(ctx->h264_dec.dpb)) {
277 struct v4l2_h264_dpb_entry *cdpb;
278
279 cdpb = &ctx->h264_dec.dpb[j];
280 if (cdpb->flags & V4L2_H264_DPB_ENTRY_FLAG_ACTIVE ||
281 !dpb_entry_match(cdpb, ndpb))
282 continue;
283
284 *cdpb = *ndpb;
285 set_bit(j, used);
286 break;
287 }
288
289 if (j == ARRAY_SIZE(ctx->h264_dec.dpb))
290 set_bit(i, new);
291 }
292
293 /* For entries that could not be matched, use remaining free slots. */
294 for_each_set_bit(i, new, ARRAY_SIZE(dec_param->dpb)) {
295 const struct v4l2_h264_dpb_entry *ndpb = &dec_param->dpb[i];
296 struct v4l2_h264_dpb_entry *cdpb;
297
298 /*
299 * Both arrays are of the same sizes, so there is no way
300 * we can end up with no space in target array, unless
301 * something is buggy.
302 */
303 j = find_first_zero_bit(used, ARRAY_SIZE(ctx->h264_dec.dpb));
304 if (WARN_ON(j >= ARRAY_SIZE(ctx->h264_dec.dpb)))
305 return;
306
307 cdpb = &ctx->h264_dec.dpb[j];
308 *cdpb = *ndpb;
309 set_bit(j, used);
310 }
311 }
312
hantro_h264_get_ref_buf(struct hantro_ctx * ctx,unsigned int dpb_idx)313 dma_addr_t hantro_h264_get_ref_buf(struct hantro_ctx *ctx,
314 unsigned int dpb_idx)
315 {
316 struct v4l2_h264_dpb_entry *dpb = ctx->h264_dec.dpb;
317 dma_addr_t dma_addr = 0;
318
319 if (dpb[dpb_idx].flags & V4L2_H264_DPB_ENTRY_FLAG_ACTIVE)
320 dma_addr = hantro_get_ref(ctx, dpb[dpb_idx].reference_ts);
321
322 if (!dma_addr) {
323 struct vb2_v4l2_buffer *dst_buf;
324 struct vb2_buffer *buf;
325
326 /*
327 * If a DPB entry is unused or invalid, address of current
328 * destination buffer is returned.
329 */
330 dst_buf = hantro_get_dst_buf(ctx);
331 buf = &dst_buf->vb2_buf;
332 dma_addr = hantro_get_dec_buf_addr(ctx, buf);
333 }
334
335 return dma_addr;
336 }
337
hantro_h264_dec_prepare_run(struct hantro_ctx * ctx)338 int hantro_h264_dec_prepare_run(struct hantro_ctx *ctx)
339 {
340 struct hantro_h264_dec_hw_ctx *h264_ctx = &ctx->h264_dec;
341 struct hantro_h264_dec_ctrls *ctrls = &h264_ctx->ctrls;
342 struct v4l2_h264_reflist_builder reflist_builder;
343
344 hantro_start_prepare_run(ctx);
345
346 ctrls->scaling =
347 hantro_get_ctrl(ctx, V4L2_CID_MPEG_VIDEO_H264_SCALING_MATRIX);
348 if (WARN_ON(!ctrls->scaling))
349 return -EINVAL;
350
351 ctrls->decode =
352 hantro_get_ctrl(ctx, V4L2_CID_MPEG_VIDEO_H264_DECODE_PARAMS);
353 if (WARN_ON(!ctrls->decode))
354 return -EINVAL;
355
356 ctrls->sps =
357 hantro_get_ctrl(ctx, V4L2_CID_MPEG_VIDEO_H264_SPS);
358 if (WARN_ON(!ctrls->sps))
359 return -EINVAL;
360
361 ctrls->pps =
362 hantro_get_ctrl(ctx, V4L2_CID_MPEG_VIDEO_H264_PPS);
363 if (WARN_ON(!ctrls->pps))
364 return -EINVAL;
365
366 /* Update the DPB with new refs. */
367 update_dpb(ctx);
368
369 /* Prepare data in memory. */
370 prepare_table(ctx);
371
372 /* Build the P/B{0,1} ref lists. */
373 v4l2_h264_init_reflist_builder(&reflist_builder, ctrls->decode,
374 ctrls->sps, ctx->h264_dec.dpb);
375 v4l2_h264_build_p_ref_list(&reflist_builder, h264_ctx->reflists.p);
376 v4l2_h264_build_b_ref_lists(&reflist_builder, h264_ctx->reflists.b0,
377 h264_ctx->reflists.b1);
378 return 0;
379 }
380
hantro_h264_dec_exit(struct hantro_ctx * ctx)381 void hantro_h264_dec_exit(struct hantro_ctx *ctx)
382 {
383 struct hantro_dev *vpu = ctx->dev;
384 struct hantro_h264_dec_hw_ctx *h264_dec = &ctx->h264_dec;
385 struct hantro_aux_buf *priv = &h264_dec->priv;
386
387 dma_free_coherent(vpu->dev, priv->size, priv->cpu, priv->dma);
388 }
389
hantro_h264_dec_init(struct hantro_ctx * ctx)390 int hantro_h264_dec_init(struct hantro_ctx *ctx)
391 {
392 struct hantro_dev *vpu = ctx->dev;
393 struct hantro_h264_dec_hw_ctx *h264_dec = &ctx->h264_dec;
394 struct hantro_aux_buf *priv = &h264_dec->priv;
395 struct hantro_h264_dec_priv_tbl *tbl;
396
397 priv->cpu = dma_alloc_coherent(vpu->dev, sizeof(*tbl), &priv->dma,
398 GFP_KERNEL);
399 if (!priv->cpu)
400 return -ENOMEM;
401
402 priv->size = sizeof(*tbl);
403 tbl = priv->cpu;
404 memcpy(tbl->cabac_table, h264_cabac_table, sizeof(tbl->cabac_table));
405
406 return 0;
407 }
408