1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2022 MediaTek Inc.
4 * Author: Yunfei Dong <yunfei.dong@mediatek.com>
5 */
6
7 #include <linux/module.h>
8 #include <linux/slab.h>
9 #include <media/v4l2-h264.h>
10 #include <media/v4l2-mem2mem.h>
11 #include <media/videobuf2-dma-contig.h>
12
13 #include "../mtk_vcodec_util.h"
14 #include "../mtk_vcodec_dec.h"
15 #include "../mtk_vcodec_intr.h"
16 #include "../vdec_drv_base.h"
17 #include "../vdec_drv_if.h"
18 #include "../vdec_vpu_if.h"
19 #include "vdec_h264_req_common.h"
20
21 /**
22 * enum vdec_h264_core_dec_err_type - core decode error type
23 *
24 * @TRANS_BUFFER_FULL: trans buffer is full
25 * @SLICE_HEADER_FULL: slice header buffer is full
26 */
27 enum vdec_h264_core_dec_err_type {
28 TRANS_BUFFER_FULL = 1,
29 SLICE_HEADER_FULL,
30 };
31
32 /**
33 * struct vdec_h264_slice_lat_dec_param - parameters for decode current frame
34 *
35 * @sps: h264 sps syntax parameters
36 * @pps: h264 pps syntax parameters
37 * @slice_header: h264 slice header syntax parameters
38 * @scaling_matrix: h264 scaling list parameters
39 * @decode_params: decoder parameters of each frame used for hardware decode
40 * @h264_dpb_info: dpb reference list
41 */
42 struct vdec_h264_slice_lat_dec_param {
43 struct mtk_h264_sps_param sps;
44 struct mtk_h264_pps_param pps;
45 struct mtk_h264_slice_hd_param slice_header;
46 struct slice_api_h264_scaling_matrix scaling_matrix;
47 struct slice_api_h264_decode_param decode_params;
48 struct mtk_h264_dpb_info h264_dpb_info[V4L2_H264_NUM_DPB_ENTRIES];
49 };
50
51 /**
52 * struct vdec_h264_slice_info - decode information
53 *
54 * @nal_info: nal info of current picture
55 * @timeout: Decode timeout: 1 timeout, 0 no timeount
56 * @bs_buf_size: bitstream size
57 * @bs_buf_addr: bitstream buffer dma address
58 * @y_fb_dma: Y frame buffer dma address
59 * @c_fb_dma: C frame buffer dma address
60 * @vdec_fb_va: VDEC frame buffer struct virtual address
61 * @crc: Used to check whether hardware's status is right
62 */
63 struct vdec_h264_slice_info {
64 u16 nal_info;
65 u16 timeout;
66 u32 bs_buf_size;
67 u64 bs_buf_addr;
68 u64 y_fb_dma;
69 u64 c_fb_dma;
70 u64 vdec_fb_va;
71 u32 crc[8];
72 };
73
74 /**
75 * struct vdec_h264_slice_vsi - shared memory for decode information exchange
76 * between SCP and Host.
77 *
78 * @wdma_err_addr: wdma error dma address
79 * @wdma_start_addr: wdma start dma address
80 * @wdma_end_addr: wdma end dma address
81 * @slice_bc_start_addr: slice bc start dma address
82 * @slice_bc_end_addr: slice bc end dma address
83 * @row_info_start_addr: row info start dma address
84 * @row_info_end_addr: row info end dma address
85 * @trans_start: trans start dma address
86 * @trans_end: trans end dma address
87 * @wdma_end_addr_offset: wdma end address offset
88 *
89 * @mv_buf_dma: HW working motion vector buffer
90 * dma address (AP-W, VPU-R)
91 * @dec: decode information (AP-R, VPU-W)
92 * @h264_slice_params: decode parameters for hw used
93 */
94 struct vdec_h264_slice_vsi {
95 /* LAT dec addr */
96 u64 wdma_err_addr;
97 u64 wdma_start_addr;
98 u64 wdma_end_addr;
99 u64 slice_bc_start_addr;
100 u64 slice_bc_end_addr;
101 u64 row_info_start_addr;
102 u64 row_info_end_addr;
103 u64 trans_start;
104 u64 trans_end;
105 u64 wdma_end_addr_offset;
106
107 u64 mv_buf_dma[H264_MAX_MV_NUM];
108 struct vdec_h264_slice_info dec;
109 struct vdec_h264_slice_lat_dec_param h264_slice_params;
110 };
111
112 /**
113 * struct vdec_h264_slice_share_info - shared information used to exchange
114 * message between lat and core
115 *
116 * @sps: sequence header information from user space
117 * @dec_params: decoder params from user space
118 * @h264_slice_params: decoder params used for hardware
119 * @trans_start: trans start dma address
120 * @trans_end: trans end dma address
121 * @nal_info: nal info of current picture
122 */
123 struct vdec_h264_slice_share_info {
124 struct v4l2_ctrl_h264_sps sps;
125 struct v4l2_ctrl_h264_decode_params dec_params;
126 struct vdec_h264_slice_lat_dec_param h264_slice_params;
127 u64 trans_start;
128 u64 trans_end;
129 u16 nal_info;
130 };
131
132 /**
133 * struct vdec_h264_slice_inst - h264 decoder instance
134 *
135 * @slice_dec_num: how many picture be decoded
136 * @ctx: point to mtk_vcodec_ctx
137 * @pred_buf: HW working predication buffer
138 * @mv_buf: HW working motion vector buffer
139 * @vpu: VPU instance
140 * @vsi: vsi used for lat
141 * @vsi_core: vsi used for core
142 *
143 * @vsi_ctx: Local VSI data for this decoding context
144 * @h264_slice_param: the parameters that hardware use to decode
145 *
146 * @resolution_changed:resolution changed
147 * @realloc_mv_buf: reallocate mv buffer
148 * @cap_num_planes: number of capture queue plane
149 *
150 * @dpb: decoded picture buffer used to store reference
151 * buffer information
152 *@is_field_bitstream: is field bitstream
153 */
154 struct vdec_h264_slice_inst {
155 unsigned int slice_dec_num;
156 struct mtk_vcodec_ctx *ctx;
157 struct mtk_vcodec_mem pred_buf;
158 struct mtk_vcodec_mem mv_buf[H264_MAX_MV_NUM];
159 struct vdec_vpu_inst vpu;
160 struct vdec_h264_slice_vsi *vsi;
161 struct vdec_h264_slice_vsi *vsi_core;
162
163 struct vdec_h264_slice_vsi vsi_ctx;
164 struct vdec_h264_slice_lat_dec_param h264_slice_param;
165
166 unsigned int resolution_changed;
167 unsigned int realloc_mv_buf;
168 unsigned int cap_num_planes;
169
170 struct v4l2_h264_dpb_entry dpb[16];
171 bool is_field_bitstream;
172 };
173
vdec_h264_slice_fill_decode_parameters(struct vdec_h264_slice_inst * inst,struct vdec_h264_slice_share_info * share_info)174 static int vdec_h264_slice_fill_decode_parameters(struct vdec_h264_slice_inst *inst,
175 struct vdec_h264_slice_share_info *share_info)
176 {
177 struct vdec_h264_slice_lat_dec_param *slice_param = &inst->vsi->h264_slice_params;
178 const struct v4l2_ctrl_h264_decode_params *dec_params;
179 const struct v4l2_ctrl_h264_scaling_matrix *src_matrix;
180 const struct v4l2_ctrl_h264_sps *sps;
181 const struct v4l2_ctrl_h264_pps *pps;
182
183 dec_params =
184 mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_DECODE_PARAMS);
185 if (IS_ERR(dec_params))
186 return PTR_ERR(dec_params);
187
188 src_matrix =
189 mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SCALING_MATRIX);
190 if (IS_ERR(src_matrix))
191 return PTR_ERR(src_matrix);
192
193 sps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SPS);
194 if (IS_ERR(sps))
195 return PTR_ERR(sps);
196
197 pps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_PPS);
198 if (IS_ERR(pps))
199 return PTR_ERR(pps);
200
201 if (dec_params->flags & V4L2_H264_DECODE_PARAM_FLAG_FIELD_PIC) {
202 mtk_vcodec_err(inst, "No support for H.264 field decoding.");
203 inst->is_field_bitstream = true;
204 return -EINVAL;
205 }
206
207 mtk_vdec_h264_copy_sps_params(&slice_param->sps, sps);
208 mtk_vdec_h264_copy_pps_params(&slice_param->pps, pps);
209 mtk_vdec_h264_copy_scaling_matrix(&slice_param->scaling_matrix, src_matrix);
210
211 memcpy(&share_info->sps, sps, sizeof(*sps));
212 memcpy(&share_info->dec_params, dec_params, sizeof(*dec_params));
213
214 return 0;
215 }
216
get_vdec_sig_decode_parameters(struct vdec_h264_slice_inst * inst)217 static int get_vdec_sig_decode_parameters(struct vdec_h264_slice_inst *inst)
218 {
219 const struct v4l2_ctrl_h264_decode_params *dec_params;
220 const struct v4l2_ctrl_h264_sps *sps;
221 const struct v4l2_ctrl_h264_pps *pps;
222 const struct v4l2_ctrl_h264_scaling_matrix *scaling_matrix;
223 struct vdec_h264_slice_lat_dec_param *slice_param = &inst->h264_slice_param;
224 struct v4l2_h264_reflist_builder reflist_builder;
225 struct v4l2_h264_reference v4l2_p0_reflist[V4L2_H264_REF_LIST_LEN];
226 struct v4l2_h264_reference v4l2_b0_reflist[V4L2_H264_REF_LIST_LEN];
227 struct v4l2_h264_reference v4l2_b1_reflist[V4L2_H264_REF_LIST_LEN];
228 u8 *p0_reflist = slice_param->decode_params.ref_pic_list_p0;
229 u8 *b0_reflist = slice_param->decode_params.ref_pic_list_b0;
230 u8 *b1_reflist = slice_param->decode_params.ref_pic_list_b1;
231
232 dec_params =
233 mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_DECODE_PARAMS);
234 if (IS_ERR(dec_params))
235 return PTR_ERR(dec_params);
236
237 sps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SPS);
238 if (IS_ERR(sps))
239 return PTR_ERR(sps);
240
241 pps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_PPS);
242 if (IS_ERR(pps))
243 return PTR_ERR(pps);
244
245 scaling_matrix =
246 mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SCALING_MATRIX);
247 if (IS_ERR(scaling_matrix))
248 return PTR_ERR(scaling_matrix);
249
250 mtk_vdec_h264_update_dpb(dec_params, inst->dpb);
251
252 mtk_vdec_h264_copy_sps_params(&slice_param->sps, sps);
253 mtk_vdec_h264_copy_pps_params(&slice_param->pps, pps);
254 mtk_vdec_h264_copy_scaling_matrix(&slice_param->scaling_matrix, scaling_matrix);
255
256 mtk_vdec_h264_copy_decode_params(&slice_param->decode_params, dec_params, inst->dpb);
257 mtk_vdec_h264_fill_dpb_info(inst->ctx, &slice_param->decode_params,
258 slice_param->h264_dpb_info);
259
260 /* Build the reference lists */
261 v4l2_h264_init_reflist_builder(&reflist_builder, dec_params, sps, inst->dpb);
262 v4l2_h264_build_p_ref_list(&reflist_builder, v4l2_p0_reflist);
263 v4l2_h264_build_b_ref_lists(&reflist_builder, v4l2_b0_reflist, v4l2_b1_reflist);
264
265 /* Adapt the built lists to the firmware's expectations */
266 mtk_vdec_h264_get_ref_list(p0_reflist, v4l2_p0_reflist, reflist_builder.num_valid);
267 mtk_vdec_h264_get_ref_list(b0_reflist, v4l2_b0_reflist, reflist_builder.num_valid);
268 mtk_vdec_h264_get_ref_list(b1_reflist, v4l2_b1_reflist, reflist_builder.num_valid);
269
270 memcpy(&inst->vsi_ctx.h264_slice_params, slice_param,
271 sizeof(inst->vsi_ctx.h264_slice_params));
272
273 return 0;
274 }
275
vdec_h264_slice_fill_decode_reflist(struct vdec_h264_slice_inst * inst,struct vdec_h264_slice_lat_dec_param * slice_param,struct vdec_h264_slice_share_info * share_info)276 static void vdec_h264_slice_fill_decode_reflist(struct vdec_h264_slice_inst *inst,
277 struct vdec_h264_slice_lat_dec_param *slice_param,
278 struct vdec_h264_slice_share_info *share_info)
279 {
280 struct v4l2_ctrl_h264_decode_params *dec_params = &share_info->dec_params;
281 struct v4l2_ctrl_h264_sps *sps = &share_info->sps;
282 struct v4l2_h264_reflist_builder reflist_builder;
283 struct v4l2_h264_reference v4l2_p0_reflist[V4L2_H264_REF_LIST_LEN];
284 struct v4l2_h264_reference v4l2_b0_reflist[V4L2_H264_REF_LIST_LEN];
285 struct v4l2_h264_reference v4l2_b1_reflist[V4L2_H264_REF_LIST_LEN];
286 u8 *p0_reflist = slice_param->decode_params.ref_pic_list_p0;
287 u8 *b0_reflist = slice_param->decode_params.ref_pic_list_b0;
288 u8 *b1_reflist = slice_param->decode_params.ref_pic_list_b1;
289
290 mtk_vdec_h264_update_dpb(dec_params, inst->dpb);
291
292 mtk_vdec_h264_copy_decode_params(&slice_param->decode_params, dec_params,
293 inst->dpb);
294 mtk_vdec_h264_fill_dpb_info(inst->ctx, &slice_param->decode_params,
295 slice_param->h264_dpb_info);
296
297 mtk_v4l2_debug(3, "cur poc = %d\n", dec_params->bottom_field_order_cnt);
298 /* Build the reference lists */
299 v4l2_h264_init_reflist_builder(&reflist_builder, dec_params, sps,
300 inst->dpb);
301 v4l2_h264_build_p_ref_list(&reflist_builder, v4l2_p0_reflist);
302 v4l2_h264_build_b_ref_lists(&reflist_builder, v4l2_b0_reflist, v4l2_b1_reflist);
303
304 /* Adapt the built lists to the firmware's expectations */
305 mtk_vdec_h264_get_ref_list(p0_reflist, v4l2_p0_reflist, reflist_builder.num_valid);
306 mtk_vdec_h264_get_ref_list(b0_reflist, v4l2_b0_reflist, reflist_builder.num_valid);
307 mtk_vdec_h264_get_ref_list(b1_reflist, v4l2_b1_reflist, reflist_builder.num_valid);
308 }
309
vdec_h264_slice_alloc_mv_buf(struct vdec_h264_slice_inst * inst,struct vdec_pic_info * pic)310 static int vdec_h264_slice_alloc_mv_buf(struct vdec_h264_slice_inst *inst,
311 struct vdec_pic_info *pic)
312 {
313 unsigned int buf_sz = mtk_vdec_h264_get_mv_buf_size(pic->buf_w, pic->buf_h);
314 struct mtk_vcodec_mem *mem;
315 int i, err;
316
317 mtk_v4l2_debug(3, "size = 0x%x", buf_sz);
318 for (i = 0; i < H264_MAX_MV_NUM; i++) {
319 mem = &inst->mv_buf[i];
320 if (mem->va)
321 mtk_vcodec_mem_free(inst->ctx, mem);
322 mem->size = buf_sz;
323 err = mtk_vcodec_mem_alloc(inst->ctx, mem);
324 if (err) {
325 mtk_vcodec_err(inst, "failed to allocate mv buf");
326 return err;
327 }
328 }
329
330 return 0;
331 }
332
vdec_h264_slice_free_mv_buf(struct vdec_h264_slice_inst * inst)333 static void vdec_h264_slice_free_mv_buf(struct vdec_h264_slice_inst *inst)
334 {
335 int i;
336 struct mtk_vcodec_mem *mem;
337
338 for (i = 0; i < H264_MAX_MV_NUM; i++) {
339 mem = &inst->mv_buf[i];
340 if (mem->va)
341 mtk_vcodec_mem_free(inst->ctx, mem);
342 }
343 }
344
vdec_h264_slice_get_pic_info(struct vdec_h264_slice_inst * inst)345 static void vdec_h264_slice_get_pic_info(struct vdec_h264_slice_inst *inst)
346 {
347 struct mtk_vcodec_ctx *ctx = inst->ctx;
348 u32 data[3];
349
350 data[0] = ctx->picinfo.pic_w;
351 data[1] = ctx->picinfo.pic_h;
352 data[2] = ctx->capture_fourcc;
353 vpu_dec_get_param(&inst->vpu, data, 3, GET_PARAM_PIC_INFO);
354
355 ctx->picinfo.buf_w = ALIGN(ctx->picinfo.pic_w, VCODEC_DEC_ALIGNED_64);
356 ctx->picinfo.buf_h = ALIGN(ctx->picinfo.pic_h, VCODEC_DEC_ALIGNED_64);
357 ctx->picinfo.fb_sz[0] = inst->vpu.fb_sz[0];
358 ctx->picinfo.fb_sz[1] = inst->vpu.fb_sz[1];
359 inst->cap_num_planes =
360 ctx->q_data[MTK_Q_DATA_DST].fmt->num_planes;
361
362 mtk_vcodec_debug(inst, "pic(%d, %d), buf(%d, %d)",
363 ctx->picinfo.pic_w, ctx->picinfo.pic_h,
364 ctx->picinfo.buf_w, ctx->picinfo.buf_h);
365 mtk_vcodec_debug(inst, "Y/C(%d, %d)", ctx->picinfo.fb_sz[0],
366 ctx->picinfo.fb_sz[1]);
367
368 if (ctx->last_decoded_picinfo.pic_w != ctx->picinfo.pic_w ||
369 ctx->last_decoded_picinfo.pic_h != ctx->picinfo.pic_h) {
370 inst->resolution_changed = true;
371 if (ctx->last_decoded_picinfo.buf_w != ctx->picinfo.buf_w ||
372 ctx->last_decoded_picinfo.buf_h != ctx->picinfo.buf_h)
373 inst->realloc_mv_buf = true;
374
375 mtk_v4l2_debug(1, "resChg: (%d %d) : old(%d, %d) -> new(%d, %d)",
376 inst->resolution_changed,
377 inst->realloc_mv_buf,
378 ctx->last_decoded_picinfo.pic_w,
379 ctx->last_decoded_picinfo.pic_h,
380 ctx->picinfo.pic_w, ctx->picinfo.pic_h);
381 }
382 }
383
vdec_h264_slice_get_crop_info(struct vdec_h264_slice_inst * inst,struct v4l2_rect * cr)384 static void vdec_h264_slice_get_crop_info(struct vdec_h264_slice_inst *inst,
385 struct v4l2_rect *cr)
386 {
387 cr->left = 0;
388 cr->top = 0;
389 cr->width = inst->ctx->picinfo.pic_w;
390 cr->height = inst->ctx->picinfo.pic_h;
391
392 mtk_vcodec_debug(inst, "l=%d, t=%d, w=%d, h=%d",
393 cr->left, cr->top, cr->width, cr->height);
394 }
395
vdec_h264_slice_init(struct mtk_vcodec_ctx * ctx)396 static int vdec_h264_slice_init(struct mtk_vcodec_ctx *ctx)
397 {
398 struct vdec_h264_slice_inst *inst;
399 int err, vsi_size;
400
401 inst = kzalloc(sizeof(*inst), GFP_KERNEL);
402 if (!inst)
403 return -ENOMEM;
404
405 inst->ctx = ctx;
406
407 inst->vpu.id = SCP_IPI_VDEC_LAT;
408 inst->vpu.core_id = SCP_IPI_VDEC_CORE;
409 inst->vpu.ctx = ctx;
410 inst->vpu.codec_type = ctx->current_codec;
411 inst->vpu.capture_type = ctx->capture_fourcc;
412
413 err = vpu_dec_init(&inst->vpu);
414 if (err) {
415 mtk_vcodec_err(inst, "vdec_h264 init err=%d", err);
416 goto error_free_inst;
417 }
418
419 vsi_size = round_up(sizeof(struct vdec_h264_slice_vsi), VCODEC_DEC_ALIGNED_64);
420 inst->vsi = inst->vpu.vsi;
421 inst->vsi_core =
422 (struct vdec_h264_slice_vsi *)(((char *)inst->vpu.vsi) + vsi_size);
423 inst->resolution_changed = true;
424 inst->realloc_mv_buf = true;
425
426 mtk_vcodec_debug(inst, "lat struct size = %d,%d,%d,%d vsi: %d\n",
427 (int)sizeof(struct mtk_h264_sps_param),
428 (int)sizeof(struct mtk_h264_pps_param),
429 (int)sizeof(struct vdec_h264_slice_lat_dec_param),
430 (int)sizeof(struct mtk_h264_dpb_info),
431 vsi_size);
432 mtk_vcodec_debug(inst, "lat H264 instance >> %p, codec_type = 0x%x",
433 inst, inst->vpu.codec_type);
434
435 ctx->drv_handle = inst;
436 return 0;
437
438 error_free_inst:
439 kfree(inst);
440 return err;
441 }
442
vdec_h264_slice_deinit(void * h_vdec)443 static void vdec_h264_slice_deinit(void *h_vdec)
444 {
445 struct vdec_h264_slice_inst *inst = h_vdec;
446
447 mtk_vcodec_debug_enter(inst);
448
449 vpu_dec_deinit(&inst->vpu);
450 vdec_h264_slice_free_mv_buf(inst);
451 vdec_msg_queue_deinit(&inst->ctx->msg_queue, inst->ctx);
452
453 kfree(inst);
454 }
455
vdec_h264_slice_core_decode(struct vdec_lat_buf * lat_buf)456 static int vdec_h264_slice_core_decode(struct vdec_lat_buf *lat_buf)
457 {
458 struct vdec_fb *fb;
459 u64 vdec_fb_va;
460 u64 y_fb_dma, c_fb_dma;
461 int err, timeout, i;
462 struct mtk_vcodec_ctx *ctx = lat_buf->ctx;
463 struct vdec_h264_slice_inst *inst = ctx->drv_handle;
464 struct vb2_v4l2_buffer *vb2_v4l2;
465 struct vdec_h264_slice_share_info *share_info = lat_buf->private_data;
466 struct mtk_vcodec_mem *mem;
467 struct vdec_vpu_inst *vpu = &inst->vpu;
468
469 mtk_vcodec_debug(inst, "[h264-core] vdec_h264 core decode");
470 memcpy(&inst->vsi_core->h264_slice_params, &share_info->h264_slice_params,
471 sizeof(share_info->h264_slice_params));
472
473 fb = ctx->dev->vdec_pdata->get_cap_buffer(ctx);
474 if (!fb) {
475 err = -EBUSY;
476 mtk_vcodec_err(inst, "fb buffer is NULL");
477 goto vdec_dec_end;
478 }
479
480 vdec_fb_va = (unsigned long)fb;
481 y_fb_dma = (u64)fb->base_y.dma_addr;
482 if (ctx->q_data[MTK_Q_DATA_DST].fmt->num_planes == 1)
483 c_fb_dma =
484 y_fb_dma + inst->ctx->picinfo.buf_w * inst->ctx->picinfo.buf_h;
485 else
486 c_fb_dma = (u64)fb->base_c.dma_addr;
487
488 mtk_vcodec_debug(inst, "[h264-core] y/c addr = 0x%llx 0x%llx", y_fb_dma,
489 c_fb_dma);
490
491 inst->vsi_core->dec.y_fb_dma = y_fb_dma;
492 inst->vsi_core->dec.c_fb_dma = c_fb_dma;
493 inst->vsi_core->dec.vdec_fb_va = vdec_fb_va;
494 inst->vsi_core->dec.nal_info = share_info->nal_info;
495 inst->vsi_core->wdma_start_addr =
496 lat_buf->ctx->msg_queue.wdma_addr.dma_addr;
497 inst->vsi_core->wdma_end_addr =
498 lat_buf->ctx->msg_queue.wdma_addr.dma_addr +
499 lat_buf->ctx->msg_queue.wdma_addr.size;
500 inst->vsi_core->wdma_err_addr = lat_buf->wdma_err_addr.dma_addr;
501 inst->vsi_core->slice_bc_start_addr = lat_buf->slice_bc_addr.dma_addr;
502 inst->vsi_core->slice_bc_end_addr = lat_buf->slice_bc_addr.dma_addr +
503 lat_buf->slice_bc_addr.size;
504 inst->vsi_core->trans_start = share_info->trans_start;
505 inst->vsi_core->trans_end = share_info->trans_end;
506 for (i = 0; i < H264_MAX_MV_NUM; i++) {
507 mem = &inst->mv_buf[i];
508 inst->vsi_core->mv_buf_dma[i] = mem->dma_addr;
509 }
510
511 vb2_v4l2 = v4l2_m2m_next_dst_buf(ctx->m2m_ctx);
512 v4l2_m2m_buf_copy_metadata(&lat_buf->ts_info, vb2_v4l2, true);
513
514 vdec_h264_slice_fill_decode_reflist(inst, &inst->vsi_core->h264_slice_params,
515 share_info);
516
517 err = vpu_dec_core(vpu);
518 if (err) {
519 mtk_vcodec_err(inst, "core decode err=%d", err);
520 goto vdec_dec_end;
521 }
522
523 /* wait decoder done interrupt */
524 timeout = mtk_vcodec_wait_for_done_ctx(inst->ctx, MTK_INST_IRQ_RECEIVED,
525 WAIT_INTR_TIMEOUT_MS, MTK_VDEC_CORE);
526 if (timeout)
527 mtk_vcodec_err(inst, "core decode timeout: pic_%d",
528 ctx->decoded_frame_cnt);
529 inst->vsi_core->dec.timeout = !!timeout;
530
531 vpu_dec_core_end(vpu);
532 mtk_vcodec_debug(inst, "pic[%d] crc: 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x",
533 ctx->decoded_frame_cnt,
534 inst->vsi_core->dec.crc[0], inst->vsi_core->dec.crc[1],
535 inst->vsi_core->dec.crc[2], inst->vsi_core->dec.crc[3],
536 inst->vsi_core->dec.crc[4], inst->vsi_core->dec.crc[5],
537 inst->vsi_core->dec.crc[6], inst->vsi_core->dec.crc[7]);
538
539 vdec_dec_end:
540 vdec_msg_queue_update_ube_rptr(&lat_buf->ctx->msg_queue, share_info->trans_end);
541 ctx->dev->vdec_pdata->cap_to_disp(ctx, !!err, lat_buf->src_buf_req);
542 mtk_vcodec_debug(inst, "core decode done err=%d", err);
543 ctx->decoded_frame_cnt++;
544 return 0;
545 }
546
vdec_h264_insert_startcode(struct mtk_vcodec_dev * vcodec_dev,unsigned char * buf,size_t * bs_size,struct mtk_h264_pps_param * pps)547 static void vdec_h264_insert_startcode(struct mtk_vcodec_dev *vcodec_dev, unsigned char *buf,
548 size_t *bs_size, struct mtk_h264_pps_param *pps)
549 {
550 struct device *dev = &vcodec_dev->plat_dev->dev;
551
552 /* Need to add pending data at the end of bitstream when bs_sz is small than
553 * 20 bytes for cavlc bitstream, or lat will decode fail. This pending data is
554 * useful for mt8192 and mt8195 platform.
555 *
556 * cavlc bitstream when entropy_coding_mode_flag is false.
557 */
558 if (pps->entropy_coding_mode_flag || *bs_size > 20 ||
559 !(of_device_is_compatible(dev->of_node, "mediatek,mt8192-vcodec-dec") ||
560 of_device_is_compatible(dev->of_node, "mediatek,mt8195-vcodec-dec")))
561 return;
562
563 buf[*bs_size] = 0;
564 buf[*bs_size + 1] = 0;
565 buf[*bs_size + 2] = 1;
566 buf[*bs_size + 3] = 0xff;
567 (*bs_size) += 4;
568 }
569
vdec_h264_slice_lat_decode(void * h_vdec,struct mtk_vcodec_mem * bs,struct vdec_fb * fb,bool * res_chg)570 static int vdec_h264_slice_lat_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
571 struct vdec_fb *fb, bool *res_chg)
572 {
573 struct vdec_h264_slice_inst *inst = h_vdec;
574 struct vdec_vpu_inst *vpu = &inst->vpu;
575 struct mtk_video_dec_buf *src_buf_info;
576 int nal_start_idx, err, timeout = 0, i;
577 unsigned int data[2];
578 struct vdec_lat_buf *lat_buf;
579 struct vdec_h264_slice_share_info *share_info;
580 unsigned char *buf;
581 struct mtk_vcodec_mem *mem;
582
583 if (vdec_msg_queue_init(&inst->ctx->msg_queue, inst->ctx,
584 vdec_h264_slice_core_decode,
585 sizeof(*share_info)))
586 return -ENOMEM;
587
588 /* bs NULL means flush decoder */
589 if (!bs) {
590 vdec_msg_queue_wait_lat_buf_full(&inst->ctx->msg_queue);
591 return vpu_dec_reset(vpu);
592 }
593
594 if (inst->is_field_bitstream)
595 return -EINVAL;
596
597 lat_buf = vdec_msg_queue_dqbuf(&inst->ctx->msg_queue.lat_ctx);
598 if (!lat_buf) {
599 mtk_vcodec_err(inst, "failed to get lat buffer");
600 return -EINVAL;
601 }
602 share_info = lat_buf->private_data;
603 src_buf_info = container_of(bs, struct mtk_video_dec_buf, bs_buffer);
604
605 buf = (unsigned char *)bs->va;
606 nal_start_idx = mtk_vdec_h264_find_start_code(buf, bs->size);
607 if (nal_start_idx < 0) {
608 err = -EINVAL;
609 goto err_free_fb_out;
610 }
611
612 inst->vsi->dec.nal_info = buf[nal_start_idx];
613 lat_buf->src_buf_req = src_buf_info->m2m_buf.vb.vb2_buf.req_obj.req;
614 v4l2_m2m_buf_copy_metadata(&src_buf_info->m2m_buf.vb, &lat_buf->ts_info, true);
615
616 err = vdec_h264_slice_fill_decode_parameters(inst, share_info);
617 if (err)
618 goto err_free_fb_out;
619
620 vdec_h264_insert_startcode(inst->ctx->dev, buf, &bs->size,
621 &share_info->h264_slice_params.pps);
622
623 inst->vsi->dec.bs_buf_addr = (uint64_t)bs->dma_addr;
624 inst->vsi->dec.bs_buf_size = bs->size;
625
626 *res_chg = inst->resolution_changed;
627 if (inst->resolution_changed) {
628 mtk_vcodec_debug(inst, "- resolution changed -");
629 if (inst->realloc_mv_buf) {
630 err = vdec_h264_slice_alloc_mv_buf(inst, &inst->ctx->picinfo);
631 inst->realloc_mv_buf = false;
632 if (err)
633 goto err_free_fb_out;
634 }
635 inst->resolution_changed = false;
636 }
637 for (i = 0; i < H264_MAX_MV_NUM; i++) {
638 mem = &inst->mv_buf[i];
639 inst->vsi->mv_buf_dma[i] = mem->dma_addr;
640 }
641 inst->vsi->wdma_start_addr = lat_buf->ctx->msg_queue.wdma_addr.dma_addr;
642 inst->vsi->wdma_end_addr = lat_buf->ctx->msg_queue.wdma_addr.dma_addr +
643 lat_buf->ctx->msg_queue.wdma_addr.size;
644 inst->vsi->wdma_err_addr = lat_buf->wdma_err_addr.dma_addr;
645 inst->vsi->slice_bc_start_addr = lat_buf->slice_bc_addr.dma_addr;
646 inst->vsi->slice_bc_end_addr = lat_buf->slice_bc_addr.dma_addr +
647 lat_buf->slice_bc_addr.size;
648
649 inst->vsi->trans_end = inst->ctx->msg_queue.wdma_rptr_addr;
650 inst->vsi->trans_start = inst->ctx->msg_queue.wdma_wptr_addr;
651 mtk_vcodec_debug(inst, "lat:trans(0x%llx 0x%llx) err:0x%llx",
652 inst->vsi->wdma_start_addr,
653 inst->vsi->wdma_end_addr,
654 inst->vsi->wdma_err_addr);
655
656 mtk_vcodec_debug(inst, "slice(0x%llx 0x%llx) rprt((0x%llx 0x%llx))",
657 inst->vsi->slice_bc_start_addr,
658 inst->vsi->slice_bc_end_addr,
659 inst->vsi->trans_start,
660 inst->vsi->trans_end);
661 err = vpu_dec_start(vpu, data, 2);
662 if (err) {
663 mtk_vcodec_debug(inst, "lat decode err: %d", err);
664 goto err_free_fb_out;
665 }
666
667 share_info->trans_end = inst->ctx->msg_queue.wdma_addr.dma_addr +
668 inst->vsi->wdma_end_addr_offset;
669 share_info->trans_start = inst->ctx->msg_queue.wdma_wptr_addr;
670 share_info->nal_info = inst->vsi->dec.nal_info;
671
672 if (IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability)) {
673 memcpy(&share_info->h264_slice_params, &inst->vsi->h264_slice_params,
674 sizeof(share_info->h264_slice_params));
675 vdec_msg_queue_qbuf(&inst->ctx->dev->msg_queue_core_ctx, lat_buf);
676 }
677
678 /* wait decoder done interrupt */
679 timeout = mtk_vcodec_wait_for_done_ctx(inst->ctx, MTK_INST_IRQ_RECEIVED,
680 WAIT_INTR_TIMEOUT_MS, MTK_VDEC_LAT0);
681 if (timeout)
682 mtk_vcodec_err(inst, "lat decode timeout: pic_%d", inst->slice_dec_num);
683 inst->vsi->dec.timeout = !!timeout;
684
685 err = vpu_dec_end(vpu);
686 if (err == SLICE_HEADER_FULL || err == TRANS_BUFFER_FULL) {
687 if (!IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability))
688 vdec_msg_queue_qbuf(&inst->ctx->msg_queue.lat_ctx, lat_buf);
689 inst->slice_dec_num++;
690 mtk_vcodec_err(inst, "lat dec fail: pic_%d err:%d", inst->slice_dec_num, err);
691 return -EINVAL;
692 }
693
694 share_info->trans_end = inst->ctx->msg_queue.wdma_addr.dma_addr +
695 inst->vsi->wdma_end_addr_offset;
696 vdec_msg_queue_update_ube_wptr(&lat_buf->ctx->msg_queue, share_info->trans_end);
697
698 if (!IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability)) {
699 memcpy(&share_info->h264_slice_params, &inst->vsi->h264_slice_params,
700 sizeof(share_info->h264_slice_params));
701 vdec_msg_queue_qbuf(&inst->ctx->dev->msg_queue_core_ctx, lat_buf);
702 }
703 mtk_vcodec_debug(inst, "dec num: %d lat crc: 0x%x 0x%x 0x%x", inst->slice_dec_num,
704 inst->vsi->dec.crc[0], inst->vsi->dec.crc[1], inst->vsi->dec.crc[2]);
705
706 inst->slice_dec_num++;
707 return 0;
708 err_free_fb_out:
709 vdec_msg_queue_qbuf(&inst->ctx->msg_queue.lat_ctx, lat_buf);
710 mtk_vcodec_err(inst, "slice dec number: %d err: %d", inst->slice_dec_num, err);
711 return err;
712 }
713
vdec_h264_slice_single_decode(void * h_vdec,struct mtk_vcodec_mem * bs,struct vdec_fb * unused,bool * res_chg)714 static int vdec_h264_slice_single_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
715 struct vdec_fb *unused, bool *res_chg)
716 {
717 struct vdec_h264_slice_inst *inst = h_vdec;
718 struct vdec_vpu_inst *vpu = &inst->vpu;
719 struct mtk_video_dec_buf *src_buf_info, *dst_buf_info;
720 struct vdec_fb *fb;
721 unsigned char *buf;
722 unsigned int data[2], i;
723 u64 y_fb_dma, c_fb_dma;
724 struct mtk_vcodec_mem *mem;
725 int err, nal_start_idx;
726
727 /* bs NULL means flush decoder */
728 if (!bs)
729 return vpu_dec_reset(vpu);
730
731 fb = inst->ctx->dev->vdec_pdata->get_cap_buffer(inst->ctx);
732 src_buf_info = container_of(bs, struct mtk_video_dec_buf, bs_buffer);
733 dst_buf_info = container_of(fb, struct mtk_video_dec_buf, frame_buffer);
734
735 y_fb_dma = fb ? (u64)fb->base_y.dma_addr : 0;
736 c_fb_dma = fb ? (u64)fb->base_c.dma_addr : 0;
737 mtk_vcodec_debug(inst, "[h264-dec] [%d] y_dma=%llx c_dma=%llx",
738 inst->ctx->decoded_frame_cnt, y_fb_dma, c_fb_dma);
739
740 inst->vsi_ctx.dec.bs_buf_addr = (u64)bs->dma_addr;
741 inst->vsi_ctx.dec.bs_buf_size = bs->size;
742 inst->vsi_ctx.dec.y_fb_dma = y_fb_dma;
743 inst->vsi_ctx.dec.c_fb_dma = c_fb_dma;
744 inst->vsi_ctx.dec.vdec_fb_va = (u64)(uintptr_t)fb;
745
746 v4l2_m2m_buf_copy_metadata(&src_buf_info->m2m_buf.vb,
747 &dst_buf_info->m2m_buf.vb, true);
748 err = get_vdec_sig_decode_parameters(inst);
749 if (err)
750 goto err_free_fb_out;
751
752 buf = (unsigned char *)bs->va;
753 nal_start_idx = mtk_vdec_h264_find_start_code(buf, bs->size);
754 if (nal_start_idx < 0) {
755 err = -EINVAL;
756 goto err_free_fb_out;
757 }
758 inst->vsi_ctx.dec.nal_info = buf[nal_start_idx];
759
760 *res_chg = inst->resolution_changed;
761 if (inst->resolution_changed) {
762 mtk_vcodec_debug(inst, "- resolution changed -");
763 if (inst->realloc_mv_buf) {
764 err = vdec_h264_slice_alloc_mv_buf(inst, &inst->ctx->picinfo);
765 inst->realloc_mv_buf = false;
766 if (err)
767 goto err_free_fb_out;
768 }
769 inst->resolution_changed = false;
770
771 for (i = 0; i < H264_MAX_MV_NUM; i++) {
772 mem = &inst->mv_buf[i];
773 inst->vsi_ctx.mv_buf_dma[i] = mem->dma_addr;
774 }
775 }
776
777 memcpy(inst->vpu.vsi, &inst->vsi_ctx, sizeof(inst->vsi_ctx));
778 err = vpu_dec_start(vpu, data, 2);
779 if (err)
780 goto err_free_fb_out;
781
782 /* wait decoder done interrupt */
783 err = mtk_vcodec_wait_for_done_ctx(inst->ctx, MTK_INST_IRQ_RECEIVED,
784 WAIT_INTR_TIMEOUT_MS, MTK_VDEC_CORE);
785 if (err)
786 mtk_vcodec_err(inst, "decode timeout: pic_%d",
787 inst->ctx->decoded_frame_cnt);
788
789 inst->vsi->dec.timeout = !!err;
790 err = vpu_dec_end(vpu);
791 if (err)
792 goto err_free_fb_out;
793
794 memcpy(&inst->vsi_ctx, inst->vpu.vsi, sizeof(inst->vsi_ctx));
795 mtk_vcodec_debug(inst, "pic[%d] crc: 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x",
796 inst->ctx->decoded_frame_cnt,
797 inst->vsi_ctx.dec.crc[0], inst->vsi_ctx.dec.crc[1],
798 inst->vsi_ctx.dec.crc[2], inst->vsi_ctx.dec.crc[3],
799 inst->vsi_ctx.dec.crc[4], inst->vsi_ctx.dec.crc[5],
800 inst->vsi_ctx.dec.crc[6], inst->vsi_ctx.dec.crc[7]);
801
802 inst->ctx->decoded_frame_cnt++;
803 return 0;
804
805 err_free_fb_out:
806 mtk_vcodec_err(inst, "dec frame number: %d err: %d",
807 inst->ctx->decoded_frame_cnt, err);
808 return err;
809 }
810
vdec_h264_slice_decode(void * h_vdec,struct mtk_vcodec_mem * bs,struct vdec_fb * unused,bool * res_chg)811 static int vdec_h264_slice_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
812 struct vdec_fb *unused, bool *res_chg)
813 {
814 struct vdec_h264_slice_inst *inst = h_vdec;
815 int ret;
816
817 if (!h_vdec)
818 return -EINVAL;
819
820 if (inst->ctx->dev->vdec_pdata->hw_arch == MTK_VDEC_PURE_SINGLE_CORE)
821 ret = vdec_h264_slice_single_decode(h_vdec, bs, unused, res_chg);
822 else
823 ret = vdec_h264_slice_lat_decode(h_vdec, bs, unused, res_chg);
824
825 return ret;
826 }
827
vdec_h264_slice_get_param(void * h_vdec,enum vdec_get_param_type type,void * out)828 static int vdec_h264_slice_get_param(void *h_vdec, enum vdec_get_param_type type,
829 void *out)
830 {
831 struct vdec_h264_slice_inst *inst = h_vdec;
832
833 switch (type) {
834 case GET_PARAM_PIC_INFO:
835 vdec_h264_slice_get_pic_info(inst);
836 break;
837 case GET_PARAM_DPB_SIZE:
838 *(unsigned int *)out = 6;
839 break;
840 case GET_PARAM_CROP_INFO:
841 vdec_h264_slice_get_crop_info(inst, out);
842 break;
843 default:
844 mtk_vcodec_err(inst, "invalid get parameter type=%d", type);
845 return -EINVAL;
846 }
847 return 0;
848 }
849
850 const struct vdec_common_if vdec_h264_slice_multi_if = {
851 .init = vdec_h264_slice_init,
852 .decode = vdec_h264_slice_decode,
853 .get_param = vdec_h264_slice_get_param,
854 .deinit = vdec_h264_slice_deinit,
855 };
856