xref: /linux/drivers/media/platform/rockchip/rkvdec/rkvdec-vdpu381-hevc.c (revision 546b928da0427b0d6c663cbb992bd7bfa9ac7971)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Rockchip VDPU381 HEVC backend
4  *
5  * Copyright (C) 2025 Collabora, Ltd.
6  *  Detlev Casanova <detlev.casanova@collabora.com>
7  */
8 
9 #include <media/v4l2-mem2mem.h>
10 
11 #include "rkvdec.h"
12 #include "rkvdec-cabac.h"
13 #include "rkvdec-rcb.h"
14 #include "rkvdec-hevc-common.h"
15 #include "rkvdec-vdpu381-regs.h"
16 
17 // SPS
18 struct rkvdec_hevc_sps {
19 	u16 video_parameters_set_id			: 4;
20 	u16 seq_parameters_set_id_sps			: 4;
21 	u16 chroma_format_idc				: 2;
22 	u16 width					: 16;
23 	u16 height					: 16;
24 	u16 bit_depth_luma				: 4;
25 	u16 bit_depth_chroma				: 4;
26 	u16 max_pic_order_count_lsb			: 5;
27 	u16 diff_max_min_luma_coding_block_size		: 2;
28 	u16 min_luma_coding_block_size			: 3;
29 	u16 min_transform_block_size			: 3;
30 	u16 diff_max_min_transform_block_size		: 2;
31 	u16 max_transform_hierarchy_depth_inter		: 3;
32 	u16 max_transform_hierarchy_depth_intra		: 3;
33 	u16 scaling_list_enabled_flag			: 1;
34 	u16 amp_enabled_flag				: 1;
35 	u16 sample_adaptive_offset_enabled_flag		: 1;
36 	u16 pcm_enabled_flag				: 1;
37 	u16 pcm_sample_bit_depth_luma			: 4;
38 	u16 pcm_sample_bit_depth_chroma			: 4;
39 	u16 pcm_loop_filter_disabled_flag		: 1;
40 	u16 diff_max_min_pcm_luma_coding_block_size	: 3;
41 	u16 min_pcm_luma_coding_block_size		: 3;
42 	u16 num_short_term_ref_pic_sets			: 7;
43 	u16 long_term_ref_pics_present_flag		: 1;
44 	u16 num_long_term_ref_pics_sps			: 6;
45 	u16 sps_temporal_mvp_enabled_flag		: 1;
46 	u16 strong_intra_smoothing_enabled_flag		: 1;
47 	u16 reserved_0					: 7;
48 	u16 sps_max_dec_pic_buffering_minus1		: 4;
49 	u16 reserved_0_2				: 3;
50 	u16 reserved_f					: 8;
51 } __packed;
52 
53 //PPS
54 struct rkvdec_hevc_pps {
55 	u16 picture_parameters_set_id			: 6;
56 	u16 seq_parameters_set_id_pps			: 4;
57 	u16 dependent_slice_segments_enabled_flag	: 1;
58 	u16 output_flag_present_flag			: 1;
59 	u16 num_extra_slice_header_bits			: 13;
60 	u16 sign_data_hiding_enabled_flag		: 1;
61 	u16 cabac_init_present_flag			: 1;
62 	u16 num_ref_idx_l0_default_active		: 4;
63 	u16 num_ref_idx_l1_default_active		: 4;
64 	u16 init_qp_minus26				: 7;
65 	u16 constrained_intra_pred_flag			: 1;
66 	u16 transform_skip_enabled_flag			: 1;
67 	u16 cu_qp_delta_enabled_flag			: 1;
68 	u16 log2_min_cb_size				: 3;
69 	u16 pps_cb_qp_offset				: 5;
70 	u16 pps_cr_qp_offset				: 5;
71 	u16 pps_slice_chroma_qp_offsets_present_flag	: 1;
72 	u16 weighted_pred_flag				: 1;
73 	u16 weighted_bipred_flag			: 1;
74 	u16 transquant_bypass_enabled_flag		: 1;
75 	u16 tiles_enabled_flag				: 1;
76 	u16 entropy_coding_sync_enabled_flag		: 1;
77 	u16 pps_loop_filter_across_slices_enabled_flag	: 1;
78 	u16 loop_filter_across_tiles_enabled_flag	: 1;
79 	u16 deblocking_filter_override_enabled_flag	: 1;
80 	u16 pps_deblocking_filter_disabled_flag		: 1;
81 	u16 pps_beta_offset_div2			: 4;
82 	u16 pps_tc_offset_div2				: 4;
83 	u16 lists_modification_present_flag		: 1;
84 	u16 log2_parallel_merge_level			: 3;
85 	u16 slice_segment_header_extension_present_flag	: 1;
86 	u16 zeroes					: 3;
87 	u16 num_tile_columns				: 5;
88 	u16 num_tile_rows				: 5;
89 	u16 sps_pps_mode				: 4;
90 	u16 reserved_bits				: 14;
91 	u16 reserved;
92 } __packed;
93 
94 struct rkvdec_hevc_tile {
95 	u16 value0	: 12;
96 	u16 value1	: 12;
97 } __packed;
98 
99 struct rkvdec_sps_pps_packet {
100 	struct rkvdec_hevc_sps sps;
101 	struct rkvdec_hevc_pps pps;
102 	struct rkvdec_hevc_tile column_width[10];
103 	struct rkvdec_hevc_tile row_height[11];
104 	u32 zeroes[3];
105 	u32 zeroes_bits		: 6;
106 	u32 padding_bits	: 2;
107 	u32 padding;
108 } __packed;
109 
110 struct rkvdec_hevc_priv_tbl {
111 	struct rkvdec_sps_pps_packet param_set[64];
112 	struct rkvdec_rps rps;
113 	struct scaling_factor scaling_list;
114 	u8 cabac_table[27456];
115 };
116 
117 struct rkvdec_hevc_ctx {
118 	struct rkvdec_aux_buf			priv_tbl;
119 	struct v4l2_ctrl_hevc_scaling_matrix	scaling_matrix_cache;
120 	struct v4l2_ctrl_hevc_ext_sps_st_rps	st_cache;
121 	struct rkvdec_vdpu381_regs_hevc		regs;
122 };
123 
assemble_hw_pps(struct rkvdec_ctx * ctx,struct rkvdec_hevc_run * run)124 static void assemble_hw_pps(struct rkvdec_ctx *ctx,
125 			    struct rkvdec_hevc_run *run)
126 {
127 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
128 	const struct v4l2_ctrl_hevc_sps *sps = run->sps;
129 	const struct v4l2_ctrl_hevc_pps *pps = run->pps;
130 	struct rkvdec_hevc_priv_tbl *priv_tbl = hevc_ctx->priv_tbl.cpu;
131 	struct rkvdec_sps_pps_packet *hw_ps;
132 	bool tiles_enabled;
133 	s32 max_cu_width;
134 	s32 pic_in_cts_width;
135 	s32 pic_in_cts_height;
136 	u16 log2_min_cb_size, width, height;
137 	u16 column_width[20];
138 	u16 row_height[22];
139 	u8 pcm_enabled;
140 	u32 i;
141 
142 	/*
143 	 * HW read the SPS/PPS information from PPS packet index by PPS id.
144 	 * offset from the base can be calculated by PPS_id * 32 (size per PPS
145 	 * packet unit). so the driver copy SPS/PPS information to the exact PPS
146 	 * packet unit for HW accessing.
147 	 */
148 	if (pps->pic_parameter_set_id >= ARRAY_SIZE(priv_tbl->param_set))
149 		return;
150 	hw_ps = &priv_tbl->param_set[pps->pic_parameter_set_id];
151 	memset(hw_ps, 0, sizeof(*hw_ps));
152 
153 	/* write sps */
154 	hw_ps->sps.video_parameters_set_id = sps->video_parameter_set_id;
155 	hw_ps->sps.seq_parameters_set_id_sps = sps->seq_parameter_set_id;
156 	hw_ps->sps.chroma_format_idc = sps->chroma_format_idc;
157 
158 	log2_min_cb_size = sps->log2_min_luma_coding_block_size_minus3 + 3;
159 	width = sps->pic_width_in_luma_samples;
160 	height = sps->pic_height_in_luma_samples;
161 	hw_ps->sps.width = width;
162 	hw_ps->sps.height = height;
163 	hw_ps->sps.bit_depth_luma = sps->bit_depth_luma_minus8 + 8;
164 	hw_ps->sps.bit_depth_chroma = sps->bit_depth_chroma_minus8 + 8;
165 	hw_ps->sps.max_pic_order_count_lsb = sps->log2_max_pic_order_cnt_lsb_minus4 + 4;
166 	hw_ps->sps.diff_max_min_luma_coding_block_size =
167 		sps->log2_diff_max_min_luma_coding_block_size;
168 	hw_ps->sps.min_luma_coding_block_size = sps->log2_min_luma_coding_block_size_minus3 + 3;
169 	hw_ps->sps.min_transform_block_size = sps->log2_min_luma_transform_block_size_minus2 + 2;
170 	hw_ps->sps.diff_max_min_transform_block_size =
171 		sps->log2_diff_max_min_luma_transform_block_size;
172 	hw_ps->sps.max_transform_hierarchy_depth_inter = sps->max_transform_hierarchy_depth_inter;
173 	hw_ps->sps.max_transform_hierarchy_depth_intra = sps->max_transform_hierarchy_depth_intra;
174 	hw_ps->sps.scaling_list_enabled_flag =
175 		!!(sps->flags & V4L2_HEVC_SPS_FLAG_SCALING_LIST_ENABLED);
176 	hw_ps->sps.amp_enabled_flag = !!(sps->flags & V4L2_HEVC_SPS_FLAG_AMP_ENABLED);
177 	hw_ps->sps.sample_adaptive_offset_enabled_flag =
178 		!!(sps->flags & V4L2_HEVC_SPS_FLAG_SAMPLE_ADAPTIVE_OFFSET);
179 
180 	pcm_enabled = !!(sps->flags & V4L2_HEVC_SPS_FLAG_PCM_ENABLED);
181 	hw_ps->sps.pcm_enabled_flag = pcm_enabled;
182 	hw_ps->sps.pcm_sample_bit_depth_luma =
183 		pcm_enabled ? sps->pcm_sample_bit_depth_luma_minus1 + 1 : 0;
184 	hw_ps->sps.pcm_sample_bit_depth_chroma =
185 		pcm_enabled ? sps->pcm_sample_bit_depth_chroma_minus1 + 1 : 0;
186 	hw_ps->sps.pcm_loop_filter_disabled_flag =
187 		!!(sps->flags & V4L2_HEVC_SPS_FLAG_PCM_LOOP_FILTER_DISABLED);
188 	hw_ps->sps.diff_max_min_pcm_luma_coding_block_size =
189 		sps->log2_diff_max_min_pcm_luma_coding_block_size;
190 	hw_ps->sps.min_pcm_luma_coding_block_size =
191 		pcm_enabled ? sps->log2_min_pcm_luma_coding_block_size_minus3 + 3 : 0;
192 	hw_ps->sps.num_short_term_ref_pic_sets = sps->num_short_term_ref_pic_sets;
193 	hw_ps->sps.long_term_ref_pics_present_flag =
194 		!!(sps->flags & V4L2_HEVC_SPS_FLAG_LONG_TERM_REF_PICS_PRESENT);
195 	hw_ps->sps.num_long_term_ref_pics_sps = sps->num_long_term_ref_pics_sps;
196 	hw_ps->sps.sps_temporal_mvp_enabled_flag =
197 		!!(sps->flags & V4L2_HEVC_SPS_FLAG_SPS_TEMPORAL_MVP_ENABLED);
198 	hw_ps->sps.strong_intra_smoothing_enabled_flag =
199 		!!(sps->flags & V4L2_HEVC_SPS_FLAG_STRONG_INTRA_SMOOTHING_ENABLED);
200 	hw_ps->sps.sps_max_dec_pic_buffering_minus1 = sps->sps_max_dec_pic_buffering_minus1;
201 	hw_ps->sps.reserved_f = 0xff;
202 
203 	/* write pps */
204 	hw_ps->pps.picture_parameters_set_id = pps->pic_parameter_set_id;
205 	hw_ps->pps.seq_parameters_set_id_pps = sps->seq_parameter_set_id;
206 	hw_ps->pps.dependent_slice_segments_enabled_flag =
207 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_DEPENDENT_SLICE_SEGMENT_ENABLED);
208 	hw_ps->pps.output_flag_present_flag =
209 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_OUTPUT_FLAG_PRESENT);
210 	hw_ps->pps.num_extra_slice_header_bits = pps->num_extra_slice_header_bits;
211 	hw_ps->pps.sign_data_hiding_enabled_flag =
212 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_SIGN_DATA_HIDING_ENABLED);
213 	hw_ps->pps.cabac_init_present_flag =
214 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_CABAC_INIT_PRESENT);
215 	hw_ps->pps.num_ref_idx_l0_default_active = pps->num_ref_idx_l0_default_active_minus1 + 1;
216 	hw_ps->pps.num_ref_idx_l1_default_active = pps->num_ref_idx_l1_default_active_minus1 + 1;
217 	hw_ps->pps.init_qp_minus26 = pps->init_qp_minus26;
218 	hw_ps->pps.constrained_intra_pred_flag =
219 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_CONSTRAINED_INTRA_PRED);
220 	hw_ps->pps.transform_skip_enabled_flag =
221 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_TRANSFORM_SKIP_ENABLED);
222 	hw_ps->pps.cu_qp_delta_enabled_flag =
223 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_CU_QP_DELTA_ENABLED);
224 	hw_ps->pps.log2_min_cb_size = log2_min_cb_size +
225 				      sps->log2_diff_max_min_luma_coding_block_size -
226 				      pps->diff_cu_qp_delta_depth;
227 	hw_ps->pps.pps_cb_qp_offset = pps->pps_cb_qp_offset;
228 	hw_ps->pps.pps_cr_qp_offset = pps->pps_cr_qp_offset;
229 	hw_ps->pps.pps_slice_chroma_qp_offsets_present_flag =
230 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_PPS_SLICE_CHROMA_QP_OFFSETS_PRESENT);
231 	hw_ps->pps.weighted_pred_flag = !!(pps->flags & V4L2_HEVC_PPS_FLAG_WEIGHTED_PRED);
232 	hw_ps->pps.weighted_bipred_flag = !!(pps->flags & V4L2_HEVC_PPS_FLAG_WEIGHTED_BIPRED);
233 	hw_ps->pps.transquant_bypass_enabled_flag =
234 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_TRANSQUANT_BYPASS_ENABLED);
235 
236 	tiles_enabled = !!(pps->flags & V4L2_HEVC_PPS_FLAG_TILES_ENABLED);
237 	hw_ps->pps.tiles_enabled_flag = tiles_enabled;
238 	hw_ps->pps.entropy_coding_sync_enabled_flag =
239 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_ENTROPY_CODING_SYNC_ENABLED);
240 	hw_ps->pps.pps_loop_filter_across_slices_enabled_flag =
241 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_PPS_LOOP_FILTER_ACROSS_SLICES_ENABLED);
242 	hw_ps->pps.loop_filter_across_tiles_enabled_flag =
243 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_LOOP_FILTER_ACROSS_TILES_ENABLED);
244 	hw_ps->pps.deblocking_filter_override_enabled_flag =
245 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_DEBLOCKING_FILTER_OVERRIDE_ENABLED);
246 	hw_ps->pps.pps_deblocking_filter_disabled_flag =
247 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_PPS_DISABLE_DEBLOCKING_FILTER);
248 	hw_ps->pps.pps_beta_offset_div2 = pps->pps_beta_offset_div2;
249 	hw_ps->pps.pps_tc_offset_div2 = pps->pps_tc_offset_div2;
250 	hw_ps->pps.lists_modification_present_flag =
251 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_LISTS_MODIFICATION_PRESENT);
252 	hw_ps->pps.log2_parallel_merge_level = pps->log2_parallel_merge_level_minus2 + 2;
253 	hw_ps->pps.slice_segment_header_extension_present_flag =
254 		!!(pps->flags & V4L2_HEVC_PPS_FLAG_SLICE_SEGMENT_HEADER_EXTENSION_PRESENT);
255 	hw_ps->pps.num_tile_columns = tiles_enabled ? pps->num_tile_columns_minus1 + 1 : 0;
256 	hw_ps->pps.num_tile_rows = tiles_enabled ? pps->num_tile_rows_minus1 + 1 : 0;
257 	hw_ps->pps.sps_pps_mode = 0;
258 	hw_ps->pps.reserved_bits = 0x3fff;
259 	hw_ps->pps.reserved = 0xffff;
260 
261 	// Setup tiles information
262 	memset(column_width, 0, sizeof(column_width));
263 	memset(row_height, 0, sizeof(row_height));
264 
265 	max_cu_width = 1 << (sps->log2_diff_max_min_luma_coding_block_size + log2_min_cb_size);
266 	pic_in_cts_width = DIV_ROUND_UP(width, max_cu_width);
267 	pic_in_cts_height = DIV_ROUND_UP(height, max_cu_width);
268 
269 	if (pps->flags & V4L2_HEVC_PPS_FLAG_TILES_ENABLED) {
270 		if (pps->flags & V4L2_HEVC_PPS_FLAG_UNIFORM_SPACING) {
271 			compute_tiles_uniform(run, log2_min_cb_size, width, height,
272 					      pic_in_cts_width, pic_in_cts_height,
273 					      column_width, row_height);
274 		} else {
275 			compute_tiles_non_uniform(run, log2_min_cb_size, width, height,
276 						  pic_in_cts_width, pic_in_cts_height,
277 						  column_width, row_height);
278 		}
279 	} else {
280 		column_width[0] = DIV_ROUND_UP(width, max_cu_width);
281 		row_height[0] = DIV_ROUND_UP(height, max_cu_width);
282 	}
283 
284 	for (i = 0; i < 20; i++) {
285 		if (column_width[i] > 0)
286 			column_width[i]--;
287 
288 		if (i & 1)
289 			hw_ps->column_width[i / 2].value1 = column_width[i];
290 		else
291 			hw_ps->column_width[i / 2].value0 = column_width[i];
292 	}
293 
294 	for (i = 0; i < 22; i++) {
295 		if (row_height[i] > 0)
296 			row_height[i]--;
297 
298 		if (i & 1)
299 			hw_ps->row_height[i / 2].value1 = row_height[i];
300 		else
301 			hw_ps->row_height[i / 2].value0 = row_height[i];
302 	}
303 
304 	hw_ps->padding = 0xffffffff;
305 	hw_ps->padding_bits = 0x3;
306 }
307 
set_ref_valid(struct rkvdec_vdpu381_regs_hevc * regs,int id,u32 valid)308 static void set_ref_valid(struct rkvdec_vdpu381_regs_hevc *regs, int id, u32 valid)
309 {
310 	switch (id) {
311 	case 0:
312 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_0 = valid;
313 		break;
314 	case 1:
315 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_1 = valid;
316 		break;
317 	case 2:
318 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_2 = valid;
319 		break;
320 	case 3:
321 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_3 = valid;
322 		break;
323 	case 4:
324 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_4 = valid;
325 		break;
326 	case 5:
327 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_5 = valid;
328 		break;
329 	case 6:
330 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_6 = valid;
331 		break;
332 	case 7:
333 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_7 = valid;
334 		break;
335 	case 8:
336 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_8 = valid;
337 		break;
338 	case 9:
339 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_9 = valid;
340 		break;
341 	case 10:
342 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_10 = valid;
343 		break;
344 	case 11:
345 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_11 = valid;
346 		break;
347 	case 12:
348 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_12 = valid;
349 		break;
350 	case 13:
351 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_13 = valid;
352 		break;
353 	case 14:
354 		regs->hevc_param.reg099_hevc_ref_valid.hevc_ref_valid_14 = valid;
355 		break;
356 	}
357 }
358 
rkvdec_write_regs(struct rkvdec_ctx * ctx)359 static void rkvdec_write_regs(struct rkvdec_ctx *ctx)
360 {
361 	struct rkvdec_dev *rkvdec = ctx->dev;
362 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
363 
364 	rkvdec_memcpy_toio(rkvdec->regs + OFFSET_COMMON_REGS,
365 			   &hevc_ctx->regs.common,
366 			   sizeof(hevc_ctx->regs.common));
367 	rkvdec_memcpy_toio(rkvdec->regs + OFFSET_CODEC_PARAMS_REGS,
368 			   &hevc_ctx->regs.hevc_param,
369 			   sizeof(hevc_ctx->regs.hevc_param));
370 	rkvdec_memcpy_toio(rkvdec->regs + OFFSET_COMMON_ADDR_REGS,
371 			   &hevc_ctx->regs.common_addr,
372 			   sizeof(hevc_ctx->regs.common_addr));
373 	rkvdec_memcpy_toio(rkvdec->regs + OFFSET_CODEC_ADDR_REGS,
374 			   &hevc_ctx->regs.hevc_addr,
375 			   sizeof(hevc_ctx->regs.hevc_addr));
376 	rkvdec_memcpy_toio(rkvdec->regs + OFFSET_POC_HIGHBIT_REGS,
377 			   &hevc_ctx->regs.hevc_highpoc,
378 			   sizeof(hevc_ctx->regs.hevc_highpoc));
379 }
380 
config_registers(struct rkvdec_ctx * ctx,struct rkvdec_hevc_run * run)381 static void config_registers(struct rkvdec_ctx *ctx,
382 			     struct rkvdec_hevc_run *run)
383 {
384 	const struct v4l2_ctrl_hevc_decode_params *dec_params = run->decode_params;
385 	const struct v4l2_hevc_dpb_entry *dpb = dec_params->dpb;
386 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
387 	struct rkvdec_vdpu381_regs_hevc *regs = &hevc_ctx->regs;
388 	dma_addr_t priv_start_addr = hevc_ctx->priv_tbl.dma;
389 	const struct v4l2_pix_format_mplane *dst_fmt;
390 	struct vb2_v4l2_buffer *src_buf = run->base.bufs.src;
391 	struct vb2_v4l2_buffer *dst_buf = run->base.bufs.dst;
392 	const struct v4l2_format *f;
393 	dma_addr_t rlc_addr;
394 	u32 hor_virstride = 0;
395 	u32 ver_virstride = 0;
396 	u32 y_virstride = 0;
397 	u32 offset;
398 	u32 pixels;
399 	dma_addr_t dst_addr;
400 	u32 i;
401 
402 	memset(regs, 0, sizeof(*regs));
403 
404 	/* Set HEVC mode */
405 	regs->common.reg009_dec_mode.dec_mode = VDPU381_MODE_HEVC;
406 
407 	/* Set config */
408 	regs->common.reg011_important_en.buf_empty_en = 1;
409 	regs->common.reg011_important_en.dec_clkgate_e = 1;
410 	regs->common.reg011_important_en.dec_timeout_e = 1;
411 	regs->common.reg011_important_en.pix_range_det_e = 1;
412 
413 	/* Set IDR flag */
414 	regs->common.reg013_en_mode_set.cur_pic_is_idr =
415 		!!(dec_params->flags & V4L2_HEVC_DECODE_PARAM_FLAG_IDR_PIC);
416 
417 	/* Set input stream length */
418 	regs->common.reg016_stream_len = vb2_get_plane_payload(&src_buf->vb2_buf, 0);
419 
420 	/* Set max slice number */
421 	regs->common.reg017_slice_number.slice_num = 1;
422 
423 	/* Set strides */
424 	f = &ctx->decoded_fmt;
425 	dst_fmt = &f->fmt.pix_mp;
426 	hor_virstride = dst_fmt->plane_fmt[0].bytesperline;
427 	ver_virstride = dst_fmt->height;
428 	y_virstride = hor_virstride * ver_virstride;
429 
430 	regs->common.reg018_y_hor_stride.y_hor_virstride = hor_virstride / 16;
431 	regs->common.reg019_uv_hor_stride.uv_hor_virstride = hor_virstride / 16;
432 	regs->common.reg020_y_stride.y_virstride = y_virstride / 16;
433 
434 	/* Activate block gating */
435 	regs->common.reg026_block_gating_en.inter_auto_gating_e = 1;
436 	regs->common.reg026_block_gating_en.filterd_auto_gating_e = 1;
437 	regs->common.reg026_block_gating_en.strmd_auto_gating_e = 1;
438 	regs->common.reg026_block_gating_en.mcp_auto_gating_e = 1;
439 	regs->common.reg026_block_gating_en.busifd_auto_gating_e = 0;
440 	regs->common.reg026_block_gating_en.dec_ctrl_auto_gating_e = 1;
441 	regs->common.reg026_block_gating_en.intra_auto_gating_e = 1;
442 	regs->common.reg026_block_gating_en.mc_auto_gating_e = 1;
443 	regs->common.reg026_block_gating_en.transd_auto_gating_e = 1;
444 	regs->common.reg026_block_gating_en.sram_auto_gating_e = 1;
445 	regs->common.reg026_block_gating_en.cru_auto_gating_e = 1;
446 	regs->common.reg026_block_gating_en.reg_cfg_gating_en = 1;
447 
448 	/* Set timeout threshold */
449 	pixels = dst_fmt->height * dst_fmt->width;
450 	if (pixels < RKVDEC_1080P_PIXELS)
451 		regs->common.reg032_timeout_threshold = RKVDEC_TIMEOUT_1080p;
452 	else if (pixels < RKVDEC_4K_PIXELS)
453 		regs->common.reg032_timeout_threshold = RKVDEC_TIMEOUT_4K;
454 	else if (pixels < RKVDEC_8K_PIXELS)
455 		regs->common.reg032_timeout_threshold = RKVDEC_TIMEOUT_8K;
456 	else
457 		regs->common.reg032_timeout_threshold = RKVDEC_TIMEOUT_MAX;
458 
459 	/* Set POC val */
460 	regs->hevc_param.reg065_cur_top_poc = dec_params->pic_order_cnt_val;
461 
462 	/* Set ref pic address & poc */
463 	for (i = 0; i < ARRAY_SIZE(dec_params->dpb); i++) {
464 		struct vb2_buffer *vb_buf = get_ref_buf(ctx, run, i);
465 		dma_addr_t buf_dma = vb2_dma_contig_plane_dma_addr(vb_buf, 0);
466 		u32 valid = !!(dec_params->num_active_dpb_entries > i);
467 
468 		/* Set reference addresses */
469 		regs->hevc_addr.reg164_180_ref_base[i] = buf_dma;
470 
471 		/* Set COLMV addresses */
472 		regs->hevc_addr.reg182_198_colmv_base[i] = buf_dma + ctx->colmv_offset;
473 
474 		regs->hevc_param.reg067_082_ref_poc[i] =
475 			dpb[i].pic_order_cnt_val;
476 
477 		set_ref_valid(regs, i, valid);
478 		regs->hevc_param.reg103_hevc_mvc0.ref_pic_layer_same_with_cur |= 1 << i;
479 	}
480 
481 	/* Set rlc base address (input stream) */
482 	rlc_addr = vb2_dma_contig_plane_dma_addr(&src_buf->vb2_buf, 0);
483 	regs->common_addr.rlc_base = rlc_addr;
484 	regs->common_addr.rlcwrite_base = rlc_addr;
485 
486 	/* Set output base address */
487 	dst_addr = vb2_dma_contig_plane_dma_addr(&dst_buf->vb2_buf, 0);
488 	regs->common_addr.decout_base = dst_addr;
489 	regs->common_addr.error_ref_base = dst_addr;
490 
491 	/* Set colmv address */
492 	regs->common_addr.colmv_cur_base = dst_addr + ctx->colmv_offset;
493 
494 	/* Set RCB addresses */
495 	for (i = 0; i < rkvdec_rcb_buf_count(ctx); i++)
496 		regs->common_addr.rcb_base[i] = rkvdec_rcb_buf_dma_addr(ctx, i);
497 
498 	/* Set hw pps address */
499 	offset = offsetof(struct rkvdec_hevc_priv_tbl, param_set);
500 	regs->hevc_addr.reg161_pps_base = priv_start_addr + offset;
501 
502 	/* Set hw rps address */
503 	offset = offsetof(struct rkvdec_hevc_priv_tbl, rps);
504 	regs->hevc_addr.reg163_rps_base = priv_start_addr + offset;
505 
506 	/* Set cabac table */
507 	offset = offsetof(struct rkvdec_hevc_priv_tbl, cabac_table);
508 	regs->hevc_addr.reg199_cabactbl_base = priv_start_addr + offset;
509 
510 	/* Set scaling matrix */
511 	offset = offsetof(struct rkvdec_hevc_priv_tbl, scaling_list);
512 	regs->hevc_addr.reg181_scanlist_addr = priv_start_addr + offset;
513 
514 	rkvdec_write_regs(ctx);
515 }
516 
rkvdec_hevc_validate_sps(struct rkvdec_ctx * ctx,const struct v4l2_ctrl_hevc_sps * sps)517 static int rkvdec_hevc_validate_sps(struct rkvdec_ctx *ctx,
518 				    const struct v4l2_ctrl_hevc_sps *sps)
519 {
520 	if (sps->chroma_format_idc != 1)
521 		/* Only 4:2:0 is supported */
522 		return -EINVAL;
523 
524 	if (sps->bit_depth_luma_minus8 != sps->bit_depth_chroma_minus8)
525 		/* Luma and chroma bit depth mismatch */
526 		return -EINVAL;
527 
528 	if (sps->bit_depth_luma_minus8 != 0 && sps->bit_depth_luma_minus8 != 2)
529 		/* Only 8-bit and 10-bit are supported */
530 		return -EINVAL;
531 
532 	if (sps->pic_width_in_luma_samples > ctx->coded_fmt.fmt.pix_mp.width ||
533 	    sps->pic_height_in_luma_samples > ctx->coded_fmt.fmt.pix_mp.height)
534 		return -EINVAL;
535 
536 	return 0;
537 }
538 
rkvdec_hevc_start(struct rkvdec_ctx * ctx)539 static int rkvdec_hevc_start(struct rkvdec_ctx *ctx)
540 {
541 	struct rkvdec_dev *rkvdec = ctx->dev;
542 	struct rkvdec_hevc_priv_tbl *priv_tbl;
543 	struct rkvdec_hevc_ctx *hevc_ctx;
544 	struct v4l2_ctrl *ctrl;
545 	int ret;
546 
547 	ctrl = v4l2_ctrl_find(&ctx->ctrl_hdl,
548 			      V4L2_CID_STATELESS_HEVC_SPS);
549 	if (!ctrl)
550 		return -EINVAL;
551 
552 	ret = rkvdec_hevc_validate_sps(ctx, ctrl->p_new.p_hevc_sps);
553 	if (ret)
554 		return ret;
555 
556 	hevc_ctx = kzalloc_obj(*hevc_ctx);
557 	if (!hevc_ctx)
558 		return -ENOMEM;
559 
560 	priv_tbl = dma_alloc_coherent(rkvdec->dev, sizeof(*priv_tbl),
561 				      &hevc_ctx->priv_tbl.dma, GFP_KERNEL);
562 	if (!priv_tbl) {
563 		ret = -ENOMEM;
564 		goto err_free_ctx;
565 	}
566 
567 	hevc_ctx->priv_tbl.size = sizeof(*priv_tbl);
568 	hevc_ctx->priv_tbl.cpu = priv_tbl;
569 	memcpy(priv_tbl->cabac_table, rkvdec_hevc_cabac_table,
570 	       sizeof(rkvdec_hevc_cabac_table));
571 
572 	ctx->priv = hevc_ctx;
573 	return 0;
574 
575 err_free_ctx:
576 	kfree(hevc_ctx);
577 	return ret;
578 }
579 
rkvdec_hevc_stop(struct rkvdec_ctx * ctx)580 static void rkvdec_hevc_stop(struct rkvdec_ctx *ctx)
581 {
582 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
583 	struct rkvdec_dev *rkvdec = ctx->dev;
584 
585 	dma_free_coherent(rkvdec->dev, hevc_ctx->priv_tbl.size,
586 			  hevc_ctx->priv_tbl.cpu, hevc_ctx->priv_tbl.dma);
587 	kfree(hevc_ctx);
588 }
589 
rkvdec_hevc_run(struct rkvdec_ctx * ctx)590 static int rkvdec_hevc_run(struct rkvdec_ctx *ctx)
591 {
592 	struct rkvdec_dev *rkvdec = ctx->dev;
593 	struct rkvdec_hevc_run run;
594 	struct rkvdec_hevc_ctx *hevc_ctx = ctx->priv;
595 	struct rkvdec_hevc_priv_tbl *tbl = hevc_ctx->priv_tbl.cpu;
596 
597 	rkvdec_hevc_run_preamble(ctx, &run);
598 
599 	rkvdec_hevc_assemble_hw_scaling_list(ctx, &run, &tbl->scaling_list,
600 					     &hevc_ctx->scaling_matrix_cache);
601 	assemble_hw_pps(ctx, &run);
602 
603 	/*
604 	 * On vdpu381, not setting the long and short term ref sets will just output wrong frames.
605 	 * Let's just warn about it and let the decoder run anyway.
606 	 */
607 	if ((!ctx->has_sps_lt_rps && run.sps->num_long_term_ref_pics_sps) ||
608 	    (!ctx->has_sps_st_rps && run.sps->num_short_term_ref_pic_sets)) {
609 		dev_warn_ratelimited(rkvdec->dev, "Long and short term RPS not set\n");
610 	} else {
611 		rkvdec_hevc_assemble_hw_rps(&run, &tbl->rps, &hevc_ctx->st_cache);
612 	}
613 
614 	config_registers(ctx, &run);
615 
616 	rkvdec_run_postamble(ctx, &run.base);
617 
618 	rkvdec_schedule_watchdog(rkvdec, hevc_ctx->regs.common.reg032_timeout_threshold);
619 
620 	/* Start decoding! */
621 	writel(VDPU381_DEC_E_BIT, rkvdec->regs + VDPU381_REG_DEC_E);
622 
623 	return 0;
624 }
625 
rkvdec_hevc_try_ctrl(struct rkvdec_ctx * ctx,struct v4l2_ctrl * ctrl)626 static int rkvdec_hevc_try_ctrl(struct rkvdec_ctx *ctx, struct v4l2_ctrl *ctrl)
627 {
628 	if (ctrl->id == V4L2_CID_STATELESS_HEVC_SPS)
629 		return rkvdec_hevc_validate_sps(ctx, ctrl->p_new.p_hevc_sps);
630 
631 	return 0;
632 }
633 
634 const struct rkvdec_coded_fmt_ops rkvdec_vdpu381_hevc_fmt_ops = {
635 	.adjust_fmt = rkvdec_hevc_adjust_fmt,
636 	.start = rkvdec_hevc_start,
637 	.stop = rkvdec_hevc_stop,
638 	.run = rkvdec_hevc_run,
639 	.try_ctrl = rkvdec_hevc_try_ctrl,
640 	.get_image_fmt = rkvdec_hevc_get_image_fmt,
641 };
642