xref: /linux/drivers/media/platform/rockchip/rkvdec/rkvdec-hevc.c (revision 546b928da0427b0d6c663cbb992bd7bfa9ac7971)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Rockchip Video Decoder HEVC backend
4  *
5  * Copyright (C) 2023 Collabora, Ltd.
6  *      Sebastian Fricke <sebastian.fricke@collabora.com>
7  *
8  * Copyright (C) 2019 Collabora, Ltd.
9  *	Boris Brezillon <boris.brezillon@collabora.com>
10  *
11  * Copyright (C) 2016 Rockchip Electronics Co., Ltd.
12  *	Jeffy Chen <jeffy.chen@rock-chips.com>
13  */
14 
15 #include <media/v4l2-hevc.h>
16 #include <media/v4l2-mem2mem.h>
17 
18 #include "rkvdec.h"
19 #include "rkvdec-regs.h"
20 #include "rkvdec-cabac.h"
21 #include "rkvdec-hevc-common.h"
22 #include "rkvdec-bitwriter.h"
23 
24 /* Size in u8/u32 units. */
25 #define RKV_SCALING_LIST_SIZE		1360
26 #define RKV_PPS_SIZE			(80 / 4)
27 #define RKV_PPS_LEN			64
28 #define RKV_RPS_SIZE			(32 / 4)
29 #define RKV_RPS_LEN			600
30 
31 struct rkvdec_sps_pps_packet {
32 	u32 info[RKV_PPS_SIZE];
33 };
34 
35 struct rkvdec_rps_packet {
36 	u32 info[RKV_RPS_SIZE];
37 };
38 
39 /* SPS */
40 #define VIDEO_PARAMETER_SET_ID				BW_FIELD(0, 4)
41 #define SEQ_PARAMETER_SET_ID				BW_FIELD(4, 4)
42 #define CHROMA_FORMAT_IDC				BW_FIELD(8, 2)
43 #define PIC_WIDTH_IN_LUMA_SAMPLES			BW_FIELD(10, 13)
44 #define PIC_HEIGHT_IN_LUMA_SAMPLES			BW_FIELD(23, 13)
45 #define BIT_DEPTH_LUMA					BW_FIELD(36, 4)
46 #define BIT_DEPTH_CHROMA				BW_FIELD(40, 4)
47 #define LOG2_MAX_PIC_ORDER_CNT_LSB			BW_FIELD(44, 5)
48 #define LOG2_DIFF_MAX_MIN_LUMA_CODING_BLOCK_SIZE	BW_FIELD(49, 2)
49 #define LOG2_MIN_LUMA_CODING_BLOCK_SIZE			BW_FIELD(51, 3)
50 #define LOG2_MIN_TRANSFORM_BLOCK_SIZE			BW_FIELD(54, 3)
51 #define LOG2_DIFF_MAX_MIN_LUMA_TRANSFORM_BLOCK_SIZE	BW_FIELD(57, 2)
52 #define MAX_TRANSFORM_HIERARCHY_DEPTH_INTER		BW_FIELD(59, 3)
53 #define MAX_TRANSFORM_HIERARCHY_DEPTH_INTRA		BW_FIELD(62, 3)
54 #define SCALING_LIST_ENABLED_FLAG			BW_FIELD(65, 1)
55 #define AMP_ENABLED_FLAG				BW_FIELD(66, 1)
56 #define SAMPLE_ADAPTIVE_OFFSET_ENABLED_FLAG		BW_FIELD(67, 1)
57 #define PCM_ENABLED_FLAG				BW_FIELD(68, 1)
58 #define PCM_SAMPLE_BIT_DEPTH_LUMA			BW_FIELD(69, 4)
59 #define PCM_SAMPLE_BIT_DEPTH_CHROMA			BW_FIELD(73, 4)
60 #define PCM_LOOP_FILTER_DISABLED_FLAG			BW_FIELD(77, 1)
61 #define LOG2_DIFF_MAX_MIN_PCM_LUMA_CODING_BLOCK_SIZE	BW_FIELD(78, 3)
62 #define LOG2_MIN_PCM_LUMA_CODING_BLOCK_SIZE		BW_FIELD(81, 3)
63 #define NUM_SHORT_TERM_REF_PIC_SETS			BW_FIELD(84, 7)
64 #define LONG_TERM_REF_PICS_PRESENT_FLAG			BW_FIELD(91, 1)
65 #define NUM_LONG_TERM_REF_PICS_SPS			BW_FIELD(92, 6)
66 #define SPS_TEMPORAL_MVP_ENABLED_FLAG			BW_FIELD(98, 1)
67 #define STRONG_INTRA_SMOOTHING_ENABLED_FLAG		BW_FIELD(99, 1)
68 /* PPS */
69 #define PIC_PARAMETER_SET_ID				BW_FIELD(128, 6)
70 #define PPS_SEQ_PARAMETER_SET_ID			BW_FIELD(134, 4)
71 #define DEPENDENT_SLICE_SEGMENTS_ENABLED_FLAG		BW_FIELD(138, 1)
72 #define OUTPUT_FLAG_PRESENT_FLAG			BW_FIELD(139, 1)
73 #define NUM_EXTRA_SLICE_HEADER_BITS			BW_FIELD(140, 13)
74 #define SIGN_DATA_HIDING_ENABLED_FLAG			BW_FIELD(153, 1)
75 #define CABAC_INIT_PRESENT_FLAG				BW_FIELD(154, 1)
76 #define NUM_REF_IDX_L0_DEFAULT_ACTIVE			BW_FIELD(155, 4)
77 #define NUM_REF_IDX_L1_DEFAULT_ACTIVE			BW_FIELD(159, 4)
78 #define INIT_QP_MINUS26					BW_FIELD(163, 7)
79 #define CONSTRAINED_INTRA_PRED_FLAG			BW_FIELD(170, 1)
80 #define TRANSFORM_SKIP_ENABLED_FLAG			BW_FIELD(171, 1)
81 #define CU_QP_DELTA_ENABLED_FLAG			BW_FIELD(172, 1)
82 #define LOG2_MIN_CU_QP_DELTA_SIZE			BW_FIELD(173, 3)
83 #define PPS_CB_QP_OFFSET				BW_FIELD(176, 5)
84 #define PPS_CR_QP_OFFSET				BW_FIELD(181, 5)
85 #define PPS_SLICE_CHROMA_QP_OFFSETS_PRESENT_FLAG	BW_FIELD(186, 1)
86 #define WEIGHTED_PRED_FLAG				BW_FIELD(187, 1)
87 #define WEIGHTED_BIPRED_FLAG				BW_FIELD(188, 1)
88 #define TRANSQUANT_BYPASS_ENABLED_FLAG			BW_FIELD(189, 1)
89 #define TILES_ENABLED_FLAG				BW_FIELD(190, 1)
90 #define ENTROPY_CODING_SYNC_ENABLED_FLAG		BW_FIELD(191, 1)
91 #define PPS_LOOP_FILTER_ACROSS_SLICES_ENABLED_FLAG	BW_FIELD(192, 1)
92 #define LOOP_FILTER_ACROSS_TILES_ENABLED_FLAG		BW_FIELD(193, 1)
93 #define DEBLOCKING_FILTER_OVERRIDE_ENABLED_FLAG		BW_FIELD(194, 1)
94 #define PPS_DEBLOCKING_FILTER_DISABLED_FLAG		BW_FIELD(195, 1)
95 #define PPS_BETA_OFFSET_DIV2				BW_FIELD(196, 4)
96 #define PPS_TC_OFFSET_DIV2				BW_FIELD(200, 4)
97 #define LISTS_MODIFICATION_PRESENT_FLAG			BW_FIELD(204, 1)
98 #define LOG2_PARALLEL_MERGE_LEVEL			BW_FIELD(205, 3)
99 #define SLICE_SEGMENT_HEADER_EXTENSION_PRESENT_FLAG	BW_FIELD(208, 1)
100 #define NUM_TILE_COLUMNS				BW_FIELD(212, 5)
101 #define NUM_TILE_ROWS					BW_FIELD(217, 5)
102 #define COLUMN_WIDTH(i)					BW_FIELD(256 + ((i) * 8), 8)
103 #define ROW_HEIGHT(i)					BW_FIELD(416 + ((i) * 8), 8)
104 #define SCALING_LIST_ADDRESS				BW_FIELD(592, 32)
105 
106 /* Data structure describing auxiliary buffer format. */
107 struct rkvdec_hevc_priv_tbl {
108 	u8 cabac_table[RKV_HEVC_CABAC_TABLE_SIZE];
109 	struct scaling_factor scaling_list;
110 	struct rkvdec_sps_pps_packet param_set[RKV_PPS_LEN];
111 	struct rkvdec_rps_packet rps[RKV_RPS_LEN];
112 };
113 
114 struct rkvdec_hevc_ctx {
115 	struct rkvdec_aux_buf priv_tbl;
116 	struct v4l2_ctrl_hevc_scaling_matrix scaling_matrix_cache;
117 	struct rkvdec_regs regs;
118 };
119 
assemble_hw_pps(struct rkvdec_ctx * ctx,struct rkvdec_hevc_run * run)120 static void assemble_hw_pps(struct rkvdec_ctx *ctx,
121 			    struct rkvdec_hevc_run *run)
122 {
123 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
124 	const struct v4l2_ctrl_hevc_sps *sps = run->sps;
125 	const struct v4l2_ctrl_hevc_pps *pps = run->pps;
126 	struct rkvdec_hevc_priv_tbl *priv_tbl = hevc_ctx->priv_tbl.cpu;
127 	struct rkvdec_sps_pps_packet *hw_ps;
128 	u32 min_cb_log2_size_y, ctb_log2_size_y, ctb_size_y;
129 	u32 log2_min_cu_qp_delta_size, scaling_distance;
130 	dma_addr_t scaling_list_address;
131 	int i;
132 
133 	/*
134 	 * HW read the SPS/PPS information from PPS packet index by PPS id.
135 	 * offset from the base can be calculated by PPS_id * 80 (size per PPS
136 	 * packet unit). so the driver copy SPS/PPS information to the exact PPS
137 	 * packet unit for HW accessing.
138 	 */
139 	if (pps->pic_parameter_set_id >= ARRAY_SIZE(priv_tbl->param_set))
140 		return;
141 	hw_ps = &priv_tbl->param_set[pps->pic_parameter_set_id];
142 	memset(hw_ps, 0, sizeof(*hw_ps));
143 
144 #define WRITE_PPS(value, field) rkvdec_set_bw_field(hw_ps->info, field, value)
145 	/* write sps */
146 	WRITE_PPS(sps->video_parameter_set_id, VIDEO_PARAMETER_SET_ID);
147 	WRITE_PPS(sps->seq_parameter_set_id, SEQ_PARAMETER_SET_ID);
148 	WRITE_PPS(sps->chroma_format_idc, CHROMA_FORMAT_IDC);
149 	WRITE_PPS(sps->pic_width_in_luma_samples, PIC_WIDTH_IN_LUMA_SAMPLES);
150 	WRITE_PPS(sps->pic_height_in_luma_samples, PIC_HEIGHT_IN_LUMA_SAMPLES);
151 	WRITE_PPS(sps->bit_depth_luma_minus8 + 8, BIT_DEPTH_LUMA);
152 	WRITE_PPS(sps->bit_depth_chroma_minus8 + 8, BIT_DEPTH_CHROMA);
153 	WRITE_PPS(sps->log2_max_pic_order_cnt_lsb_minus4 + 4,
154 		  LOG2_MAX_PIC_ORDER_CNT_LSB);
155 	WRITE_PPS(sps->log2_diff_max_min_luma_coding_block_size,
156 		  LOG2_DIFF_MAX_MIN_LUMA_CODING_BLOCK_SIZE);
157 	WRITE_PPS(sps->log2_min_luma_coding_block_size_minus3 + 3,
158 		  LOG2_MIN_LUMA_CODING_BLOCK_SIZE);
159 	WRITE_PPS(sps->log2_min_luma_transform_block_size_minus2 + 2,
160 		  LOG2_MIN_TRANSFORM_BLOCK_SIZE);
161 	WRITE_PPS(sps->log2_diff_max_min_luma_transform_block_size,
162 		  LOG2_DIFF_MAX_MIN_LUMA_TRANSFORM_BLOCK_SIZE);
163 	WRITE_PPS(sps->max_transform_hierarchy_depth_inter,
164 		  MAX_TRANSFORM_HIERARCHY_DEPTH_INTER);
165 	WRITE_PPS(sps->max_transform_hierarchy_depth_intra,
166 		  MAX_TRANSFORM_HIERARCHY_DEPTH_INTRA);
167 	WRITE_PPS(!!(sps->flags & V4L2_HEVC_SPS_FLAG_SCALING_LIST_ENABLED),
168 		  SCALING_LIST_ENABLED_FLAG);
169 	WRITE_PPS(!!(sps->flags & V4L2_HEVC_SPS_FLAG_AMP_ENABLED),
170 		  AMP_ENABLED_FLAG);
171 	WRITE_PPS(!!(sps->flags & V4L2_HEVC_SPS_FLAG_SAMPLE_ADAPTIVE_OFFSET),
172 		  SAMPLE_ADAPTIVE_OFFSET_ENABLED_FLAG);
173 	if (sps->flags & V4L2_HEVC_SPS_FLAG_PCM_ENABLED) {
174 		WRITE_PPS(1, PCM_ENABLED_FLAG);
175 		WRITE_PPS(sps->pcm_sample_bit_depth_luma_minus1 + 1,
176 			  PCM_SAMPLE_BIT_DEPTH_LUMA);
177 		WRITE_PPS(sps->pcm_sample_bit_depth_chroma_minus1 + 1,
178 			  PCM_SAMPLE_BIT_DEPTH_CHROMA);
179 		WRITE_PPS(!!(sps->flags & V4L2_HEVC_SPS_FLAG_PCM_LOOP_FILTER_DISABLED),
180 			  PCM_LOOP_FILTER_DISABLED_FLAG);
181 		WRITE_PPS(sps->log2_diff_max_min_pcm_luma_coding_block_size,
182 			  LOG2_DIFF_MAX_MIN_PCM_LUMA_CODING_BLOCK_SIZE);
183 		WRITE_PPS(sps->log2_min_pcm_luma_coding_block_size_minus3 + 3,
184 			  LOG2_MIN_PCM_LUMA_CODING_BLOCK_SIZE);
185 	}
186 	WRITE_PPS(sps->num_short_term_ref_pic_sets, NUM_SHORT_TERM_REF_PIC_SETS);
187 	WRITE_PPS(!!(sps->flags & V4L2_HEVC_SPS_FLAG_LONG_TERM_REF_PICS_PRESENT),
188 		  LONG_TERM_REF_PICS_PRESENT_FLAG);
189 	WRITE_PPS(sps->num_long_term_ref_pics_sps, NUM_LONG_TERM_REF_PICS_SPS);
190 	WRITE_PPS(!!(sps->flags & V4L2_HEVC_SPS_FLAG_SPS_TEMPORAL_MVP_ENABLED),
191 		  SPS_TEMPORAL_MVP_ENABLED_FLAG);
192 	WRITE_PPS(!!(sps->flags & V4L2_HEVC_SPS_FLAG_STRONG_INTRA_SMOOTHING_ENABLED),
193 		  STRONG_INTRA_SMOOTHING_ENABLED_FLAG);
194 
195 	/* write pps */
196 	WRITE_PPS(pps->pic_parameter_set_id, PIC_PARAMETER_SET_ID);
197 	WRITE_PPS(sps->seq_parameter_set_id, PPS_SEQ_PARAMETER_SET_ID);
198 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_DEPENDENT_SLICE_SEGMENT_ENABLED),
199 		  DEPENDENT_SLICE_SEGMENTS_ENABLED_FLAG);
200 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_OUTPUT_FLAG_PRESENT),
201 		  OUTPUT_FLAG_PRESENT_FLAG);
202 	WRITE_PPS(pps->num_extra_slice_header_bits, NUM_EXTRA_SLICE_HEADER_BITS);
203 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_SIGN_DATA_HIDING_ENABLED),
204 		  SIGN_DATA_HIDING_ENABLED_FLAG);
205 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_CABAC_INIT_PRESENT),
206 		  CABAC_INIT_PRESENT_FLAG);
207 	WRITE_PPS(pps->num_ref_idx_l0_default_active_minus1 + 1,
208 		  NUM_REF_IDX_L0_DEFAULT_ACTIVE);
209 	WRITE_PPS(pps->num_ref_idx_l1_default_active_minus1 + 1,
210 		  NUM_REF_IDX_L1_DEFAULT_ACTIVE);
211 	WRITE_PPS(pps->init_qp_minus26, INIT_QP_MINUS26);
212 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_CONSTRAINED_INTRA_PRED),
213 		  CONSTRAINED_INTRA_PRED_FLAG);
214 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_TRANSFORM_SKIP_ENABLED),
215 		  TRANSFORM_SKIP_ENABLED_FLAG);
216 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_CU_QP_DELTA_ENABLED),
217 		  CU_QP_DELTA_ENABLED_FLAG);
218 
219 	min_cb_log2_size_y = sps->log2_min_luma_coding_block_size_minus3 + 3;
220 	ctb_log2_size_y = min_cb_log2_size_y +
221 		sps->log2_diff_max_min_luma_coding_block_size;
222 	ctb_size_y = 1 << ctb_log2_size_y;
223 	log2_min_cu_qp_delta_size = ctb_log2_size_y - pps->diff_cu_qp_delta_depth;
224 	WRITE_PPS(log2_min_cu_qp_delta_size, LOG2_MIN_CU_QP_DELTA_SIZE);
225 	WRITE_PPS(pps->pps_cb_qp_offset, PPS_CB_QP_OFFSET);
226 	WRITE_PPS(pps->pps_cr_qp_offset, PPS_CR_QP_OFFSET);
227 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_PPS_SLICE_CHROMA_QP_OFFSETS_PRESENT),
228 		  PPS_SLICE_CHROMA_QP_OFFSETS_PRESENT_FLAG);
229 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_WEIGHTED_PRED),
230 		  WEIGHTED_PRED_FLAG);
231 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_WEIGHTED_BIPRED),
232 		  WEIGHTED_BIPRED_FLAG);
233 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_TRANSQUANT_BYPASS_ENABLED),
234 		  TRANSQUANT_BYPASS_ENABLED_FLAG);
235 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_TILES_ENABLED),
236 		  TILES_ENABLED_FLAG);
237 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_ENTROPY_CODING_SYNC_ENABLED),
238 		  ENTROPY_CODING_SYNC_ENABLED_FLAG);
239 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_PPS_LOOP_FILTER_ACROSS_SLICES_ENABLED),
240 		  PPS_LOOP_FILTER_ACROSS_SLICES_ENABLED_FLAG);
241 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_LOOP_FILTER_ACROSS_TILES_ENABLED),
242 		  LOOP_FILTER_ACROSS_TILES_ENABLED_FLAG);
243 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_DEBLOCKING_FILTER_OVERRIDE_ENABLED),
244 		  DEBLOCKING_FILTER_OVERRIDE_ENABLED_FLAG);
245 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_PPS_DISABLE_DEBLOCKING_FILTER),
246 		  PPS_DEBLOCKING_FILTER_DISABLED_FLAG);
247 	WRITE_PPS(pps->pps_beta_offset_div2, PPS_BETA_OFFSET_DIV2);
248 	WRITE_PPS(pps->pps_tc_offset_div2, PPS_TC_OFFSET_DIV2);
249 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_LISTS_MODIFICATION_PRESENT),
250 		  LISTS_MODIFICATION_PRESENT_FLAG);
251 	WRITE_PPS(pps->log2_parallel_merge_level_minus2 + 2, LOG2_PARALLEL_MERGE_LEVEL);
252 	WRITE_PPS(!!(pps->flags & V4L2_HEVC_PPS_FLAG_SLICE_SEGMENT_HEADER_EXTENSION_PRESENT),
253 		  SLICE_SEGMENT_HEADER_EXTENSION_PRESENT_FLAG);
254 	WRITE_PPS(pps->num_tile_columns_minus1 + 1, NUM_TILE_COLUMNS);
255 	WRITE_PPS(pps->num_tile_rows_minus1 + 1, NUM_TILE_ROWS);
256 
257 	if (pps->flags & V4L2_HEVC_PPS_FLAG_TILES_ENABLED) {
258 		/* Userspace also provide column width and row height for uniform spacing */
259 		for (i = 0; i < v4l2_hevc_pps_num_tile_columns(pps); i++)
260 			WRITE_PPS(pps->column_width_minus1[i], COLUMN_WIDTH(i));
261 		for (i = 0; i < v4l2_hevc_pps_num_tile_rows(pps); i++)
262 			WRITE_PPS(pps->row_height_minus1[i], ROW_HEIGHT(i));
263 	} else {
264 		WRITE_PPS(DIV_ROUND_UP(sps->pic_width_in_luma_samples, ctb_size_y) - 1,
265 			  COLUMN_WIDTH(0));
266 		WRITE_PPS(DIV_ROUND_UP(sps->pic_height_in_luma_samples, ctb_size_y) - 1,
267 			  ROW_HEIGHT(0));
268 	}
269 
270 	scaling_distance = offsetof(struct rkvdec_hevc_priv_tbl, scaling_list);
271 	scaling_list_address = hevc_ctx->priv_tbl.dma + scaling_distance;
272 	WRITE_PPS(scaling_list_address, SCALING_LIST_ADDRESS);
273 }
274 
275 /*
276  * Creation of the Reference Picture Set memory blob for the hardware.
277  * The layout looks like this:
278  * [0] 32 bits for L0 (6 references + 2 bits of the 7th reference)
279  * [1] 32 bits for L0 (remaining 3 bits of the 7th reference + 5 references
280  *     + 4 bits of the 13th reference)
281  * [2] 11 bits for L0 (remaining bit for 13 and 2 references) and
282  *     21 bits for L1 (4 references + first bit of 5)
283  * [3] 32 bits of padding with 0s
284  * [4] 32 bits for L1 (remaining 4 bits for 5 + 5 references + 3 bits of 11)
285  * [5] 22 bits for L1 (remaining 2 bits of 11 and 4 references)
286  *     lowdelay flag (bit 23), rps bit offset long term (bit 24 - 32)
287  * [6] rps bit offset long term (bit 1 - 3),  rps bit offset short term (bit 4 - 12)
288  *     number of references (bit 13 - 16), remaining 16 bits of padding with 0s
289  * [7] 32 bits of padding with 0s
290  *
291  * Thus we have to set up padding in between reference 5 of the L1 list.
292  */
assemble_sw_rps(struct rkvdec_ctx * ctx,struct rkvdec_hevc_run * run)293 static void assemble_sw_rps(struct rkvdec_ctx *ctx,
294 			    struct rkvdec_hevc_run *run)
295 {
296 	const struct v4l2_ctrl_hevc_decode_params *decode_params = run->decode_params;
297 	const struct v4l2_ctrl_hevc_sps *sps = run->sps;
298 	const struct v4l2_ctrl_hevc_slice_params *sl_params;
299 	const struct v4l2_hevc_dpb_entry *dpb;
300 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
301 	struct rkvdec_hevc_priv_tbl *priv_tbl = hevc_ctx->priv_tbl.cpu;
302 	struct rkvdec_rps_packet *hw_ps;
303 	int i, j;
304 	unsigned int lowdelay;
305 
306 #define WRITE_RPS(value, field) rkvdec_set_bw_field(hw_ps->info, field, value)
307 
308 #define REF_PIC_LONG_TERM_L0(n)			BW_FIELD((n) * 5, 1)
309 #define REF_PIC_IDX_L0(n)			BW_FIELD(1 + ((n) * 5), 4)
310 #define REF_PIC_LONG_TERM_L1(n)			BW_FIELD(((n) < 5 ? 75 : 132) + ((n) * 5), 1)
311 #define REF_PIC_IDX_L1(n)			BW_FIELD(((n) < 4 ? 76 : 128) + ((n) * 5), 4)
312 
313 #define LOWDELAY				BW_FIELD(182, 1)
314 #define LONG_TERM_RPS_BIT_OFFSET		BW_FIELD(183, 10)
315 #define SHORT_TERM_RPS_BIT_OFFSET		BW_FIELD(193, 9)
316 #define NUM_RPS_POC				BW_FIELD(202, 4)
317 
318 	for (j = 0; j < run->num_slices; j++) {
319 		uint st_bit_offset = 0;
320 		uint num_l0_refs = 0;
321 		uint num_l1_refs = 0;
322 
323 		sl_params = &run->slices_params[j];
324 		dpb = decode_params->dpb;
325 
326 		if (sl_params->slice_type != V4L2_HEVC_SLICE_TYPE_I) {
327 			num_l0_refs = sl_params->num_ref_idx_l0_active_minus1 + 1;
328 
329 			if (sl_params->slice_type == V4L2_HEVC_SLICE_TYPE_B)
330 				num_l1_refs = sl_params->num_ref_idx_l1_active_minus1 + 1;
331 
332 			lowdelay = 1;
333 		} else {
334 			lowdelay = 0;
335 		}
336 
337 		hw_ps = &priv_tbl->rps[j];
338 		memset(hw_ps, 0, sizeof(*hw_ps));
339 
340 		for (i = 0; i < num_l0_refs; i++) {
341 			const struct v4l2_hevc_dpb_entry dpb_l0 = dpb[sl_params->ref_idx_l0[i]];
342 
343 			WRITE_RPS(!!(dpb_l0.flags & V4L2_HEVC_DPB_ENTRY_LONG_TERM_REFERENCE),
344 				  REF_PIC_LONG_TERM_L0(i));
345 			WRITE_RPS(sl_params->ref_idx_l0[i], REF_PIC_IDX_L0(i));
346 
347 			if (dpb_l0.pic_order_cnt_val > sl_params->slice_pic_order_cnt)
348 				lowdelay = 0;
349 		}
350 
351 		for (i = 0; i < num_l1_refs; i++) {
352 			const struct v4l2_hevc_dpb_entry dpb_l1 = dpb[sl_params->ref_idx_l1[i]];
353 			int is_long_term =
354 				!!(dpb_l1.flags & V4L2_HEVC_DPB_ENTRY_LONG_TERM_REFERENCE);
355 
356 			WRITE_RPS(is_long_term, REF_PIC_LONG_TERM_L1(i));
357 			WRITE_RPS(sl_params->ref_idx_l1[i], REF_PIC_IDX_L1(i));
358 
359 			if (dpb_l1.pic_order_cnt_val > sl_params->slice_pic_order_cnt)
360 				lowdelay = 0;
361 		}
362 
363 		WRITE_RPS(lowdelay, LOWDELAY);
364 
365 		if (!(decode_params->flags & V4L2_HEVC_DECODE_PARAM_FLAG_IDR_PIC)) {
366 			if (sl_params->short_term_ref_pic_set_size)
367 				st_bit_offset = sl_params->short_term_ref_pic_set_size;
368 			else if (sps->num_short_term_ref_pic_sets > 1)
369 				st_bit_offset = fls(sps->num_short_term_ref_pic_sets - 1);
370 		}
371 
372 		WRITE_RPS(st_bit_offset + sl_params->long_term_ref_pic_set_size,
373 			  LONG_TERM_RPS_BIT_OFFSET);
374 		WRITE_RPS(sl_params->short_term_ref_pic_set_size,
375 			  SHORT_TERM_RPS_BIT_OFFSET);
376 
377 		WRITE_RPS(decode_params->num_poc_st_curr_before +
378 			  decode_params->num_poc_st_curr_after +
379 			  decode_params->num_poc_lt_curr,
380 			  NUM_RPS_POC);
381 	}
382 }
383 
config_registers(struct rkvdec_ctx * ctx,struct rkvdec_hevc_run * run)384 static void config_registers(struct rkvdec_ctx *ctx,
385 			     struct rkvdec_hevc_run *run)
386 {
387 	struct rkvdec_dev *rkvdec = ctx->dev;
388 	const struct v4l2_ctrl_hevc_decode_params *decode_params = run->decode_params;
389 	const struct v4l2_ctrl_hevc_sps *sps = run->sps;
390 	const struct v4l2_ctrl_hevc_slice_params *sl_params = &run->slices_params[0];
391 	const struct v4l2_hevc_dpb_entry *dpb = decode_params->dpb;
392 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
393 	struct rkvdec_regs *regs = &hevc_ctx->regs;
394 	dma_addr_t priv_start_addr = hevc_ctx->priv_tbl.dma;
395 	const struct v4l2_pix_format_mplane *dst_fmt;
396 	struct vb2_v4l2_buffer *src_buf = run->base.bufs.src;
397 	struct vb2_v4l2_buffer *dst_buf = run->base.bufs.dst;
398 	const struct v4l2_format *f;
399 	dma_addr_t rlc_addr;
400 	dma_addr_t refer_addr;
401 	u32 rlc_len;
402 	u32 hor_virstride;
403 	u32 ver_virstride;
404 	u32 y_virstride;
405 	u32 yuv_virstride = 0;
406 	u32 offset;
407 	dma_addr_t dst_addr;
408 	u32 reg, i;
409 
410 	memset(regs, 0, sizeof(*regs));
411 
412 	regs->common.reg02.dec_mode = RKVDEC_MODE_HEVC;
413 
414 	f = &ctx->decoded_fmt;
415 	dst_fmt = &f->fmt.pix_mp;
416 	hor_virstride = dst_fmt->plane_fmt[0].bytesperline;
417 	ver_virstride = dst_fmt->height;
418 	y_virstride = hor_virstride * ver_virstride;
419 
420 	if (sps->chroma_format_idc == 0)
421 		yuv_virstride = y_virstride;
422 	else if (sps->chroma_format_idc == 1)
423 		yuv_virstride = y_virstride + y_virstride / 2;
424 	else if (sps->chroma_format_idc == 2)
425 		yuv_virstride = 2 * y_virstride;
426 
427 	regs->common.reg03.slice_num_lowbits = run->num_slices;
428 	regs->common.reg03.uv_hor_virstride = hor_virstride / 16;
429 	regs->common.reg03.y_hor_virstride = hor_virstride / 16;
430 
431 	/* config rlc base address */
432 	rlc_addr = vb2_dma_contig_plane_dma_addr(&src_buf->vb2_buf, 0);
433 	regs->common.strm_rlc_base = rlc_addr;
434 
435 	rlc_len = vb2_get_plane_payload(&src_buf->vb2_buf, 0);
436 	regs->common.stream_len = round_up(rlc_len, 16) + 64;
437 
438 	/* config cabac table */
439 	offset = offsetof(struct rkvdec_hevc_priv_tbl, cabac_table);
440 	regs->common.cabactbl_base = priv_start_addr + offset;
441 
442 	/* config output base address */
443 	dst_addr = vb2_dma_contig_plane_dma_addr(&dst_buf->vb2_buf, 0);
444 	regs->common.decout_base = dst_addr;
445 
446 	regs->common.reg08.y_virstride = y_virstride / 16;
447 	regs->common.reg09.yuv_virstride = yuv_virstride / 16;
448 
449 	/* config ref pic address */
450 	for (i = 0; i < 15; i++) {
451 		struct vb2_buffer *vb_buf = get_ref_buf(ctx, run, i);
452 
453 		if (i < 4 && decode_params->num_active_dpb_entries) {
454 			reg = GENMASK(decode_params->num_active_dpb_entries - 1, 0);
455 			reg = (reg >> (i * 4)) & 0xf;
456 		} else {
457 			reg = 0;
458 		}
459 
460 		refer_addr = vb2_dma_contig_plane_dma_addr(vb_buf, 0);
461 
462 		regs->h26x.ref0_14_base[i].base_addr = refer_addr >> 4;
463 		regs->h26x.ref0_14_base[i].field_ref = !!(reg & 1);
464 		regs->h26x.ref0_14_base[i].topfield_used_ref = !!(reg & 2);
465 		regs->h26x.ref0_14_base[i].botfield_used_ref = !!(reg & 4);
466 		regs->h26x.ref0_14_base[i].colmv_use_flag_ref = !!(reg & 8);
467 
468 		regs->h26x.ref0_14_poc[i] = i < decode_params->num_active_dpb_entries
469 					    ? dpb[i].pic_order_cnt_val
470 					    : 0;
471 	}
472 
473 	regs->h26x.cur_poc = sl_params->slice_pic_order_cnt;
474 
475 	/* config hw pps address */
476 	offset = offsetof(struct rkvdec_hevc_priv_tbl, param_set);
477 	regs->h26x.pps_base = priv_start_addr + offset;
478 
479 	/* config hw rps address */
480 	offset = offsetof(struct rkvdec_hevc_priv_tbl, rps);
481 	regs->h26x.rps_base = priv_start_addr + offset;
482 
483 	rkvdec_memcpy_toio(rkvdec->regs, regs,
484 			   MIN(sizeof(*regs), sizeof(u32) * rkvdec->variant->num_regs));
485 }
486 
rkvdec_hevc_validate_sps(struct rkvdec_ctx * ctx,const struct v4l2_ctrl_hevc_sps * sps)487 static int rkvdec_hevc_validate_sps(struct rkvdec_ctx *ctx,
488 				    const struct v4l2_ctrl_hevc_sps *sps)
489 {
490 	if (sps->chroma_format_idc > 1)
491 		/* Only 4:0:0 and 4:2:0 are supported */
492 		return -EINVAL;
493 	if (sps->bit_depth_luma_minus8 != sps->bit_depth_chroma_minus8)
494 		/* Luma and chroma bit depth mismatch */
495 		return -EINVAL;
496 	if (sps->bit_depth_luma_minus8 != 0 && sps->bit_depth_luma_minus8 != 2)
497 		/* Only 8-bit and 10-bit are supported */
498 		return -EINVAL;
499 
500 	if (sps->pic_width_in_luma_samples > ctx->coded_fmt.fmt.pix_mp.width ||
501 	    sps->pic_height_in_luma_samples > ctx->coded_fmt.fmt.pix_mp.height)
502 		return -EINVAL;
503 
504 	return 0;
505 }
506 
rkvdec_hevc_start(struct rkvdec_ctx * ctx)507 static int rkvdec_hevc_start(struct rkvdec_ctx *ctx)
508 {
509 	struct rkvdec_dev *rkvdec = ctx->dev;
510 	struct rkvdec_hevc_priv_tbl *priv_tbl;
511 	struct rkvdec_hevc_ctx *hevc_ctx;
512 
513 	hevc_ctx = kzalloc_obj(*hevc_ctx);
514 	if (!hevc_ctx)
515 		return -ENOMEM;
516 
517 	priv_tbl = dma_alloc_coherent(rkvdec->dev, sizeof(*priv_tbl),
518 				      &hevc_ctx->priv_tbl.dma, GFP_KERNEL);
519 	if (!priv_tbl) {
520 		kfree(hevc_ctx);
521 		return -ENOMEM;
522 	}
523 
524 	hevc_ctx->priv_tbl.size = sizeof(*priv_tbl);
525 	hevc_ctx->priv_tbl.cpu = priv_tbl;
526 	memcpy(priv_tbl->cabac_table, rkvdec_hevc_cabac_table,
527 	       sizeof(rkvdec_hevc_cabac_table));
528 
529 	ctx->priv = hevc_ctx;
530 	return 0;
531 }
532 
rkvdec_hevc_stop(struct rkvdec_ctx * ctx)533 static void rkvdec_hevc_stop(struct rkvdec_ctx *ctx)
534 {
535 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
536 	struct rkvdec_dev *rkvdec = ctx->dev;
537 
538 	dma_free_coherent(rkvdec->dev, hevc_ctx->priv_tbl.size,
539 			  hevc_ctx->priv_tbl.cpu, hevc_ctx->priv_tbl.dma);
540 	kfree(hevc_ctx);
541 }
542 
rkvdec_hevc_run(struct rkvdec_ctx * ctx)543 static int rkvdec_hevc_run(struct rkvdec_ctx *ctx)
544 {
545 	struct rkvdec_dev *rkvdec = ctx->dev;
546 	struct rkvdec_hevc_run run;
547 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
548 	struct rkvdec_hevc_priv_tbl *tbl = hevc_ctx->priv_tbl.cpu;
549 	u32 reg;
550 
551 	rkvdec_hevc_run_preamble(ctx, &run);
552 
553 	rkvdec_hevc_assemble_hw_scaling_list(ctx, &run, &tbl->scaling_list,
554 					     &hevc_ctx->scaling_matrix_cache);
555 	assemble_hw_pps(ctx, &run);
556 	assemble_sw_rps(ctx, &run);
557 	config_registers(ctx, &run);
558 
559 	rkvdec_run_postamble(ctx, &run.base);
560 
561 	schedule_delayed_work(&rkvdec->watchdog_work, msecs_to_jiffies(2000));
562 
563 	writel(1, rkvdec->regs + RKVDEC_REG_PREF_LUMA_CACHE_COMMAND);
564 	writel(1, rkvdec->regs + RKVDEC_REG_PREF_CHR_CACHE_COMMAND);
565 
566 	if (rkvdec->variant->quirks & RKVDEC_QUIRK_DISABLE_QOS)
567 		rkvdec_quirks_disable_qos(ctx);
568 
569 	/* Start decoding! */
570 	reg = (run.pps->flags & V4L2_HEVC_PPS_FLAG_TILES_ENABLED) ?
571 		0 : RKVDEC_WR_DDR_ALIGN_EN;
572 	writel(RKVDEC_INTERRUPT_DEC_E | RKVDEC_CONFIG_DEC_CLK_GATE_E |
573 	       RKVDEC_TIMEOUT_E | RKVDEC_BUF_EMPTY_E | reg,
574 	       rkvdec->regs + RKVDEC_REG_INTERRUPT);
575 
576 	return 0;
577 }
578 
rkvdec_hevc_try_ctrl(struct rkvdec_ctx * ctx,struct v4l2_ctrl * ctrl)579 static int rkvdec_hevc_try_ctrl(struct rkvdec_ctx *ctx, struct v4l2_ctrl *ctrl)
580 {
581 	if (ctrl->id == V4L2_CID_STATELESS_HEVC_SPS)
582 		return rkvdec_hevc_validate_sps(ctx, ctrl->p_new.p_hevc_sps);
583 
584 	return 0;
585 }
586 
587 const struct rkvdec_coded_fmt_ops rkvdec_hevc_fmt_ops = {
588 	.adjust_fmt = rkvdec_hevc_adjust_fmt,
589 	.start = rkvdec_hevc_start,
590 	.stop = rkvdec_hevc_stop,
591 	.run = rkvdec_hevc_run,
592 	.try_ctrl = rkvdec_hevc_try_ctrl,
593 	.get_image_fmt = rkvdec_hevc_get_image_fmt,
594 };
595