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_dec.h"
14 #include "../../common/mtk_vcodec_intr.h"
15 #include "../vdec_drv_base.h"
16 #include "../vdec_drv_if.h"
17 #include "../vdec_vpu_if.h"
18 #include "vdec_h264_req_common.h"
19 
20 /**
21  * enum vdec_h264_core_dec_err_type  - core decode error type
22  *
23  * @TRANS_BUFFER_FULL: trans buffer is full
24  * @SLICE_HEADER_FULL: slice header buffer is full
25  */
26 enum vdec_h264_core_dec_err_type {
27 	TRANS_BUFFER_FULL = 1,
28 	SLICE_HEADER_FULL,
29 };
30 
31 /**
32  * struct vdec_h264_slice_lat_dec_param  - parameters for decode current frame
33  *
34  * @sps:		h264 sps syntax parameters
35  * @pps:		h264 pps syntax parameters
36  * @slice_header:	h264 slice header syntax parameters
37  * @scaling_matrix:	h264 scaling list parameters
38  * @decode_params:	decoder parameters of each frame used for hardware decode
39  * @h264_dpb_info:	dpb reference list
40  */
41 struct vdec_h264_slice_lat_dec_param {
42 	struct mtk_h264_sps_param sps;
43 	struct mtk_h264_pps_param pps;
44 	struct mtk_h264_slice_hd_param slice_header;
45 	struct slice_api_h264_scaling_matrix scaling_matrix;
46 	struct slice_api_h264_decode_param decode_params;
47 	struct mtk_h264_dpb_info h264_dpb_info[V4L2_H264_NUM_DPB_ENTRIES];
48 };
49 
50 /**
51  * struct vdec_h264_slice_info - decode information
52  *
53  * @nal_info:		nal info of current picture
54  * @timeout:		Decode timeout: 1 timeout, 0 no timeount
55  * @bs_buf_size:	bitstream size
56  * @bs_buf_addr:	bitstream buffer dma address
57  * @y_fb_dma:		Y frame buffer dma address
58  * @c_fb_dma:		C frame buffer dma address
59  * @vdec_fb_va:	VDEC frame buffer struct virtual address
60  * @crc:		Used to check whether hardware's status is right
61  */
62 struct vdec_h264_slice_info {
63 	u16 nal_info;
64 	u16 timeout;
65 	u32 bs_buf_size;
66 	u64 bs_buf_addr;
67 	u64 y_fb_dma;
68 	u64 c_fb_dma;
69 	u64 vdec_fb_va;
70 	u32 crc[8];
71 };
72 
73 /**
74  * struct vdec_h264_slice_vsi - shared memory for decode information exchange
75  *        between SCP and Host.
76  *
77  * @wdma_err_addr:		wdma error dma address
78  * @wdma_start_addr:		wdma start dma address
79  * @wdma_end_addr:		wdma end dma address
80  * @slice_bc_start_addr:	slice bc start dma address
81  * @slice_bc_end_addr:		slice bc end dma address
82  * @row_info_start_addr:	row info start dma address
83  * @row_info_end_addr:		row info end dma address
84  * @trans_start:		trans start dma address
85  * @trans_end:			trans end dma address
86  * @wdma_end_addr_offset:	wdma end address offset
87  *
88  * @mv_buf_dma:		HW working motion vector buffer
89  *				dma address (AP-W, VPU-R)
90  * @dec:			decode information (AP-R, VPU-W)
91  * @h264_slice_params:		decode parameters for hw used
92  */
93 struct vdec_h264_slice_vsi {
94 	/* LAT dec addr */
95 	u64 wdma_err_addr;
96 	u64 wdma_start_addr;
97 	u64 wdma_end_addr;
98 	u64 slice_bc_start_addr;
99 	u64 slice_bc_end_addr;
100 	u64 row_info_start_addr;
101 	u64 row_info_end_addr;
102 	u64 trans_start;
103 	u64 trans_end;
104 	u64 wdma_end_addr_offset;
105 
106 	u64 mv_buf_dma[H264_MAX_MV_NUM];
107 	struct vdec_h264_slice_info dec;
108 	struct vdec_h264_slice_lat_dec_param h264_slice_params;
109 };
110 
111 /**
112  * struct vdec_h264_slice_share_info - shared information used to exchange
113  *                                     message between lat and core
114  *
115  * @sps:		sequence header information from user space
116  * @dec_params:	decoder params from user space
117  * @h264_slice_params:	decoder params used for hardware
118  * @trans_start:	trans start dma address
119  * @trans_end:		trans end dma address
120  * @nal_info:		nal info of current picture
121  */
122 struct vdec_h264_slice_share_info {
123 	struct v4l2_ctrl_h264_sps sps;
124 	struct v4l2_ctrl_h264_decode_params dec_params;
125 	struct vdec_h264_slice_lat_dec_param h264_slice_params;
126 	u64 trans_start;
127 	u64 trans_end;
128 	u16 nal_info;
129 };
130 
131 /**
132  * struct vdec_h264_slice_inst - h264 decoder instance
133  *
134  * @slice_dec_num:	how many picture be decoded
135  * @ctx:		point to mtk_vcodec_dec_ctx
136  * @pred_buf:		HW working predication buffer
137  * @mv_buf:		HW working motion vector buffer
138  * @vpu:		VPU instance
139  * @vsi:		vsi used for lat
140  * @vsi_core:		vsi used for core
141  *
142  * @vsi_ctx:		Local VSI data for this decoding context
143  * @h264_slice_param:	the parameters that hardware use to decode
144  *
145  * @resolution_changed:resolution changed
146  * @realloc_mv_buf:	reallocate mv buffer
147  * @cap_num_planes:	number of capture queue plane
148  *
149  * @dpb:		decoded picture buffer used to store reference
150  *			buffer information
151  *@is_field_bitstream:	is field bitstream
152  */
153 struct vdec_h264_slice_inst {
154 	unsigned int slice_dec_num;
155 	struct mtk_vcodec_dec_ctx *ctx;
156 	struct mtk_vcodec_mem pred_buf;
157 	struct mtk_vcodec_mem mv_buf[H264_MAX_MV_NUM];
158 	struct vdec_vpu_inst vpu;
159 	struct vdec_h264_slice_vsi *vsi;
160 	struct vdec_h264_slice_vsi *vsi_core;
161 
162 	struct vdec_h264_slice_vsi vsi_ctx;
163 	struct vdec_h264_slice_lat_dec_param h264_slice_param;
164 
165 	unsigned int resolution_changed;
166 	unsigned int realloc_mv_buf;
167 	unsigned int cap_num_planes;
168 
169 	struct v4l2_h264_dpb_entry dpb[16];
170 	bool is_field_bitstream;
171 };
172 
vdec_h264_slice_fill_decode_parameters(struct vdec_h264_slice_inst * inst,struct vdec_h264_slice_share_info * share_info)173 static int vdec_h264_slice_fill_decode_parameters(struct vdec_h264_slice_inst *inst,
174 						  struct vdec_h264_slice_share_info *share_info)
175 {
176 	struct vdec_h264_slice_lat_dec_param *slice_param = &inst->vsi->h264_slice_params;
177 	const struct v4l2_ctrl_h264_decode_params *dec_params;
178 	const struct v4l2_ctrl_h264_scaling_matrix *src_matrix;
179 	const struct v4l2_ctrl_h264_sps *sps;
180 	const struct v4l2_ctrl_h264_pps *pps;
181 
182 	dec_params =
183 		mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_DECODE_PARAMS);
184 	if (IS_ERR(dec_params))
185 		return PTR_ERR(dec_params);
186 
187 	src_matrix =
188 		mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SCALING_MATRIX);
189 	if (IS_ERR(src_matrix))
190 		return PTR_ERR(src_matrix);
191 
192 	sps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SPS);
193 	if (IS_ERR(sps))
194 		return PTR_ERR(sps);
195 
196 	pps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_PPS);
197 	if (IS_ERR(pps))
198 		return PTR_ERR(pps);
199 
200 	if (dec_params->flags & V4L2_H264_DECODE_PARAM_FLAG_FIELD_PIC) {
201 		mtk_vdec_err(inst->ctx, "No support for H.264 field decoding.");
202 		inst->is_field_bitstream = true;
203 		return -EINVAL;
204 	}
205 
206 	mtk_vdec_h264_copy_sps_params(&slice_param->sps, sps);
207 	mtk_vdec_h264_copy_pps_params(&slice_param->pps, pps);
208 	mtk_vdec_h264_copy_scaling_matrix(&slice_param->scaling_matrix, src_matrix);
209 
210 	memcpy(&share_info->sps, sps, sizeof(*sps));
211 	memcpy(&share_info->dec_params, dec_params, sizeof(*dec_params));
212 
213 	return 0;
214 }
215 
get_vdec_sig_decode_parameters(struct vdec_h264_slice_inst * inst)216 static int get_vdec_sig_decode_parameters(struct vdec_h264_slice_inst *inst)
217 {
218 	const struct v4l2_ctrl_h264_decode_params *dec_params;
219 	const struct v4l2_ctrl_h264_sps *sps;
220 	const struct v4l2_ctrl_h264_pps *pps;
221 	const struct v4l2_ctrl_h264_scaling_matrix *scaling_matrix;
222 	struct vdec_h264_slice_lat_dec_param *slice_param = &inst->h264_slice_param;
223 	struct v4l2_h264_reflist_builder reflist_builder;
224 	struct v4l2_h264_reference v4l2_p0_reflist[V4L2_H264_REF_LIST_LEN];
225 	struct v4l2_h264_reference v4l2_b0_reflist[V4L2_H264_REF_LIST_LEN];
226 	struct v4l2_h264_reference v4l2_b1_reflist[V4L2_H264_REF_LIST_LEN];
227 	u8 *p0_reflist = slice_param->decode_params.ref_pic_list_p0;
228 	u8 *b0_reflist = slice_param->decode_params.ref_pic_list_b0;
229 	u8 *b1_reflist = slice_param->decode_params.ref_pic_list_b1;
230 
231 	dec_params =
232 		mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_DECODE_PARAMS);
233 	if (IS_ERR(dec_params))
234 		return PTR_ERR(dec_params);
235 
236 	sps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SPS);
237 	if (IS_ERR(sps))
238 		return PTR_ERR(sps);
239 
240 	pps = mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_PPS);
241 	if (IS_ERR(pps))
242 		return PTR_ERR(pps);
243 
244 	scaling_matrix =
245 		mtk_vdec_h264_get_ctrl_ptr(inst->ctx, V4L2_CID_STATELESS_H264_SCALING_MATRIX);
246 	if (IS_ERR(scaling_matrix))
247 		return PTR_ERR(scaling_matrix);
248 
249 	mtk_vdec_h264_update_dpb(dec_params, inst->dpb);
250 
251 	mtk_vdec_h264_copy_sps_params(&slice_param->sps, sps);
252 	mtk_vdec_h264_copy_pps_params(&slice_param->pps, pps);
253 	mtk_vdec_h264_copy_scaling_matrix(&slice_param->scaling_matrix, scaling_matrix);
254 
255 	mtk_vdec_h264_copy_decode_params(&slice_param->decode_params, dec_params, inst->dpb);
256 	mtk_vdec_h264_fill_dpb_info(inst->ctx, &slice_param->decode_params,
257 				    slice_param->h264_dpb_info);
258 
259 	/* Build the reference lists */
260 	v4l2_h264_init_reflist_builder(&reflist_builder, dec_params, sps, inst->dpb);
261 	v4l2_h264_build_p_ref_list(&reflist_builder, v4l2_p0_reflist);
262 	v4l2_h264_build_b_ref_lists(&reflist_builder, v4l2_b0_reflist, v4l2_b1_reflist);
263 
264 	/* Adapt the built lists to the firmware's expectations */
265 	mtk_vdec_h264_get_ref_list(p0_reflist, v4l2_p0_reflist, reflist_builder.num_valid);
266 	mtk_vdec_h264_get_ref_list(b0_reflist, v4l2_b0_reflist, reflist_builder.num_valid);
267 	mtk_vdec_h264_get_ref_list(b1_reflist, v4l2_b1_reflist, reflist_builder.num_valid);
268 
269 	memcpy(&inst->vsi_ctx.h264_slice_params, slice_param,
270 	       sizeof(inst->vsi_ctx.h264_slice_params));
271 
272 	return 0;
273 }
274 
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)275 static void vdec_h264_slice_fill_decode_reflist(struct vdec_h264_slice_inst *inst,
276 						struct vdec_h264_slice_lat_dec_param *slice_param,
277 						struct vdec_h264_slice_share_info *share_info)
278 {
279 	struct v4l2_ctrl_h264_decode_params *dec_params = &share_info->dec_params;
280 	struct v4l2_ctrl_h264_sps *sps = &share_info->sps;
281 	struct v4l2_h264_reflist_builder reflist_builder;
282 	struct v4l2_h264_reference v4l2_p0_reflist[V4L2_H264_REF_LIST_LEN];
283 	struct v4l2_h264_reference v4l2_b0_reflist[V4L2_H264_REF_LIST_LEN];
284 	struct v4l2_h264_reference v4l2_b1_reflist[V4L2_H264_REF_LIST_LEN];
285 	u8 *p0_reflist = slice_param->decode_params.ref_pic_list_p0;
286 	u8 *b0_reflist = slice_param->decode_params.ref_pic_list_b0;
287 	u8 *b1_reflist = slice_param->decode_params.ref_pic_list_b1;
288 
289 	mtk_vdec_h264_update_dpb(dec_params, inst->dpb);
290 
291 	mtk_vdec_h264_copy_decode_params(&slice_param->decode_params, dec_params,
292 					 inst->dpb);
293 	mtk_vdec_h264_fill_dpb_info(inst->ctx, &slice_param->decode_params,
294 				    slice_param->h264_dpb_info);
295 
296 	mtk_v4l2_vdec_dbg(3, inst->ctx, "cur poc = %d\n", dec_params->bottom_field_order_cnt);
297 	/* Build the reference lists */
298 	v4l2_h264_init_reflist_builder(&reflist_builder, dec_params, sps,
299 				       inst->dpb);
300 	v4l2_h264_build_p_ref_list(&reflist_builder, v4l2_p0_reflist);
301 	v4l2_h264_build_b_ref_lists(&reflist_builder, v4l2_b0_reflist, v4l2_b1_reflist);
302 
303 	/* Adapt the built lists to the firmware's expectations */
304 	mtk_vdec_h264_get_ref_list(p0_reflist, v4l2_p0_reflist, reflist_builder.num_valid);
305 	mtk_vdec_h264_get_ref_list(b0_reflist, v4l2_b0_reflist, reflist_builder.num_valid);
306 	mtk_vdec_h264_get_ref_list(b1_reflist, v4l2_b1_reflist, reflist_builder.num_valid);
307 }
308 
vdec_h264_slice_alloc_mv_buf(struct vdec_h264_slice_inst * inst,struct vdec_pic_info * pic)309 static int vdec_h264_slice_alloc_mv_buf(struct vdec_h264_slice_inst *inst,
310 					struct vdec_pic_info *pic)
311 {
312 	unsigned int buf_sz = mtk_vdec_h264_get_mv_buf_size(pic->buf_w, pic->buf_h);
313 	struct mtk_vcodec_mem *mem;
314 	int i, err;
315 
316 	mtk_v4l2_vdec_dbg(3, inst->ctx, "size = 0x%x", buf_sz);
317 	for (i = 0; i < H264_MAX_MV_NUM; i++) {
318 		mem = &inst->mv_buf[i];
319 		if (mem->va)
320 			mtk_vcodec_mem_free(inst->ctx, mem);
321 		mem->size = buf_sz;
322 		err = mtk_vcodec_mem_alloc(inst->ctx, mem);
323 		if (err) {
324 			mtk_vdec_err(inst->ctx, "failed to allocate mv buf");
325 			return err;
326 		}
327 	}
328 
329 	return 0;
330 }
331 
vdec_h264_slice_free_mv_buf(struct vdec_h264_slice_inst * inst)332 static void vdec_h264_slice_free_mv_buf(struct vdec_h264_slice_inst *inst)
333 {
334 	int i;
335 	struct mtk_vcodec_mem *mem;
336 
337 	for (i = 0; i < H264_MAX_MV_NUM; i++) {
338 		mem = &inst->mv_buf[i];
339 		if (mem->va)
340 			mtk_vcodec_mem_free(inst->ctx, mem);
341 	}
342 }
343 
vdec_h264_slice_get_pic_info(struct vdec_h264_slice_inst * inst)344 static void vdec_h264_slice_get_pic_info(struct vdec_h264_slice_inst *inst)
345 {
346 	struct mtk_vcodec_dec_ctx *ctx = inst->ctx;
347 	u32 data[3];
348 
349 	data[0] = ctx->picinfo.pic_w;
350 	data[1] = ctx->picinfo.pic_h;
351 	data[2] = ctx->capture_fourcc;
352 	vpu_dec_get_param(&inst->vpu, data, 3, GET_PARAM_PIC_INFO);
353 
354 	ctx->picinfo.buf_w = ALIGN(ctx->picinfo.pic_w, VCODEC_DEC_ALIGNED_64);
355 	ctx->picinfo.buf_h = ALIGN(ctx->picinfo.pic_h, VCODEC_DEC_ALIGNED_64);
356 	ctx->picinfo.fb_sz[0] = inst->vpu.fb_sz[0];
357 	ctx->picinfo.fb_sz[1] = inst->vpu.fb_sz[1];
358 	inst->cap_num_planes =
359 		ctx->q_data[MTK_Q_DATA_DST].fmt->num_planes;
360 
361 	mtk_vdec_debug(ctx, "pic(%d, %d), buf(%d, %d)",
362 		       ctx->picinfo.pic_w, ctx->picinfo.pic_h,
363 		       ctx->picinfo.buf_w, ctx->picinfo.buf_h);
364 	mtk_vdec_debug(ctx, "Y/C(%d, %d)", ctx->picinfo.fb_sz[0],
365 		       ctx->picinfo.fb_sz[1]);
366 
367 	if (ctx->last_decoded_picinfo.pic_w != ctx->picinfo.pic_w ||
368 	    ctx->last_decoded_picinfo.pic_h != ctx->picinfo.pic_h) {
369 		inst->resolution_changed = true;
370 		if (ctx->last_decoded_picinfo.buf_w != ctx->picinfo.buf_w ||
371 		    ctx->last_decoded_picinfo.buf_h != ctx->picinfo.buf_h)
372 			inst->realloc_mv_buf = true;
373 
374 		mtk_v4l2_vdec_dbg(1, inst->ctx, "resChg: (%d %d) : old(%d, %d) -> new(%d, %d)",
375 				  inst->resolution_changed,
376 				  inst->realloc_mv_buf,
377 				  ctx->last_decoded_picinfo.pic_w,
378 				  ctx->last_decoded_picinfo.pic_h,
379 				  ctx->picinfo.pic_w, ctx->picinfo.pic_h);
380 	}
381 }
382 
vdec_h264_slice_get_crop_info(struct vdec_h264_slice_inst * inst,struct v4l2_rect * cr)383 static void vdec_h264_slice_get_crop_info(struct vdec_h264_slice_inst *inst,
384 					  struct v4l2_rect *cr)
385 {
386 	cr->left = 0;
387 	cr->top = 0;
388 	cr->width = inst->ctx->picinfo.pic_w;
389 	cr->height = inst->ctx->picinfo.pic_h;
390 
391 	mtk_vdec_debug(inst->ctx, "l=%d, t=%d, w=%d, h=%d",
392 		       cr->left, cr->top, cr->width, cr->height);
393 }
394 
vdec_h264_slice_init(struct mtk_vcodec_dec_ctx * ctx)395 static int vdec_h264_slice_init(struct mtk_vcodec_dec_ctx *ctx)
396 {
397 	struct vdec_h264_slice_inst *inst;
398 	int err, vsi_size;
399 
400 	inst = kzalloc(sizeof(*inst), GFP_KERNEL);
401 	if (!inst)
402 		return -ENOMEM;
403 
404 	inst->ctx = ctx;
405 
406 	inst->vpu.id = SCP_IPI_VDEC_LAT;
407 	inst->vpu.core_id = SCP_IPI_VDEC_CORE;
408 	inst->vpu.ctx = ctx;
409 	inst->vpu.codec_type = ctx->current_codec;
410 	inst->vpu.capture_type = ctx->capture_fourcc;
411 
412 	err = vpu_dec_init(&inst->vpu);
413 	if (err) {
414 		mtk_vdec_err(ctx, "vdec_h264 init err=%d", err);
415 		goto error_free_inst;
416 	}
417 
418 	vsi_size = round_up(sizeof(struct vdec_h264_slice_vsi), VCODEC_DEC_ALIGNED_64);
419 	inst->vsi = inst->vpu.vsi;
420 	inst->vsi_core =
421 		(struct vdec_h264_slice_vsi *)(((char *)inst->vpu.vsi) + vsi_size);
422 	inst->resolution_changed = true;
423 	inst->realloc_mv_buf = true;
424 
425 	mtk_vdec_debug(ctx, "lat struct size = %d,%d,%d,%d vsi: %d\n",
426 		       (int)sizeof(struct mtk_h264_sps_param),
427 		       (int)sizeof(struct mtk_h264_pps_param),
428 		       (int)sizeof(struct vdec_h264_slice_lat_dec_param),
429 		       (int)sizeof(struct mtk_h264_dpb_info),
430 		       vsi_size);
431 	mtk_vdec_debug(ctx, "lat H264 instance >> %p, codec_type = 0x%x",
432 		       inst, inst->vpu.codec_type);
433 
434 	ctx->drv_handle = inst;
435 	return 0;
436 
437 error_free_inst:
438 	kfree(inst);
439 	return err;
440 }
441 
vdec_h264_slice_deinit(void * h_vdec)442 static void vdec_h264_slice_deinit(void *h_vdec)
443 {
444 	struct vdec_h264_slice_inst *inst = h_vdec;
445 
446 	vpu_dec_deinit(&inst->vpu);
447 	vdec_h264_slice_free_mv_buf(inst);
448 	vdec_msg_queue_deinit(&inst->ctx->msg_queue, inst->ctx);
449 
450 	kfree(inst);
451 }
452 
vdec_h264_slice_core_decode(struct vdec_lat_buf * lat_buf)453 static int vdec_h264_slice_core_decode(struct vdec_lat_buf *lat_buf)
454 {
455 	struct vdec_fb *fb;
456 	u64 vdec_fb_va;
457 	u64 y_fb_dma, c_fb_dma;
458 	int err, timeout, i;
459 	struct mtk_vcodec_dec_ctx *ctx = lat_buf->ctx;
460 	struct vdec_h264_slice_inst *inst = ctx->drv_handle;
461 	struct vb2_v4l2_buffer *vb2_v4l2;
462 	struct vdec_h264_slice_share_info *share_info = lat_buf->private_data;
463 	struct mtk_vcodec_mem *mem;
464 	struct vdec_vpu_inst *vpu = &inst->vpu;
465 
466 	mtk_vdec_debug(ctx, "[h264-core] vdec_h264 core decode");
467 	memcpy(&inst->vsi_core->h264_slice_params, &share_info->h264_slice_params,
468 	       sizeof(share_info->h264_slice_params));
469 
470 	fb = ctx->dev->vdec_pdata->get_cap_buffer(ctx);
471 	if (!fb) {
472 		err = -EBUSY;
473 		mtk_vdec_err(ctx, "fb buffer is NULL");
474 		goto vdec_dec_end;
475 	}
476 
477 	vdec_fb_va = (unsigned long)fb;
478 	y_fb_dma = (u64)fb->base_y.dma_addr;
479 	if (ctx->q_data[MTK_Q_DATA_DST].fmt->num_planes == 1)
480 		c_fb_dma =
481 			y_fb_dma + inst->ctx->picinfo.buf_w * inst->ctx->picinfo.buf_h;
482 	else
483 		c_fb_dma = (u64)fb->base_c.dma_addr;
484 
485 	mtk_vdec_debug(ctx, "[h264-core] y/c addr = 0x%llx 0x%llx", y_fb_dma, c_fb_dma);
486 
487 	inst->vsi_core->dec.y_fb_dma = y_fb_dma;
488 	inst->vsi_core->dec.c_fb_dma = c_fb_dma;
489 	inst->vsi_core->dec.vdec_fb_va = vdec_fb_va;
490 	inst->vsi_core->dec.nal_info = share_info->nal_info;
491 	inst->vsi_core->wdma_start_addr =
492 		lat_buf->ctx->msg_queue.wdma_addr.dma_addr;
493 	inst->vsi_core->wdma_end_addr =
494 		lat_buf->ctx->msg_queue.wdma_addr.dma_addr +
495 		lat_buf->ctx->msg_queue.wdma_addr.size;
496 	inst->vsi_core->wdma_err_addr = lat_buf->wdma_err_addr.dma_addr;
497 	inst->vsi_core->slice_bc_start_addr = lat_buf->slice_bc_addr.dma_addr;
498 	inst->vsi_core->slice_bc_end_addr = lat_buf->slice_bc_addr.dma_addr +
499 		lat_buf->slice_bc_addr.size;
500 	inst->vsi_core->trans_start = share_info->trans_start;
501 	inst->vsi_core->trans_end = share_info->trans_end;
502 	for (i = 0; i < H264_MAX_MV_NUM; i++) {
503 		mem = &inst->mv_buf[i];
504 		inst->vsi_core->mv_buf_dma[i] = mem->dma_addr;
505 	}
506 
507 	vb2_v4l2 = v4l2_m2m_next_dst_buf(ctx->m2m_ctx);
508 	v4l2_m2m_buf_copy_metadata(&lat_buf->ts_info, vb2_v4l2, true);
509 
510 	vdec_h264_slice_fill_decode_reflist(inst, &inst->vsi_core->h264_slice_params,
511 					    share_info);
512 
513 	err = vpu_dec_core(vpu);
514 	if (err) {
515 		mtk_vdec_err(ctx, "core decode err=%d", err);
516 		goto vdec_dec_end;
517 	}
518 
519 	/* wait decoder done interrupt */
520 	timeout = mtk_vcodec_wait_for_done_ctx(inst->ctx, MTK_INST_IRQ_RECEIVED,
521 					       WAIT_INTR_TIMEOUT_MS, MTK_VDEC_CORE);
522 	if (timeout)
523 		mtk_vdec_err(ctx, "core decode timeout: pic_%d", ctx->decoded_frame_cnt);
524 	inst->vsi_core->dec.timeout = !!timeout;
525 
526 	vpu_dec_core_end(vpu);
527 	mtk_vdec_debug(ctx, "pic[%d] crc: 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x",
528 		       ctx->decoded_frame_cnt,
529 		       inst->vsi_core->dec.crc[0], inst->vsi_core->dec.crc[1],
530 		       inst->vsi_core->dec.crc[2], inst->vsi_core->dec.crc[3],
531 		       inst->vsi_core->dec.crc[4], inst->vsi_core->dec.crc[5],
532 		       inst->vsi_core->dec.crc[6], inst->vsi_core->dec.crc[7]);
533 
534 vdec_dec_end:
535 	vdec_msg_queue_update_ube_rptr(&lat_buf->ctx->msg_queue, share_info->trans_end);
536 	ctx->dev->vdec_pdata->cap_to_disp(ctx, !!err, lat_buf->src_buf_req);
537 	mtk_vdec_debug(ctx, "core decode done err=%d", err);
538 	ctx->decoded_frame_cnt++;
539 	return 0;
540 }
541 
vdec_h264_insert_startcode(struct mtk_vcodec_dec_dev * vcodec_dev,unsigned char * buf,size_t * bs_size,struct mtk_h264_pps_param * pps)542 static void vdec_h264_insert_startcode(struct mtk_vcodec_dec_dev *vcodec_dev, unsigned char *buf,
543 				       size_t *bs_size, struct mtk_h264_pps_param *pps)
544 {
545 	struct device *dev = &vcodec_dev->plat_dev->dev;
546 
547 	/* Need to add pending data at the end of bitstream when bs_sz is small than
548 	 * 20 bytes for cavlc bitstream, or lat will decode fail. This pending data is
549 	 * useful for mt8192 and mt8195 platform.
550 	 *
551 	 * cavlc bitstream when entropy_coding_mode_flag is false.
552 	 */
553 	if (pps->entropy_coding_mode_flag || *bs_size > 20 ||
554 	    !(of_device_is_compatible(dev->of_node, "mediatek,mt8192-vcodec-dec") ||
555 	    of_device_is_compatible(dev->of_node, "mediatek,mt8195-vcodec-dec")))
556 		return;
557 
558 	buf[*bs_size] = 0;
559 	buf[*bs_size + 1] = 0;
560 	buf[*bs_size + 2] = 1;
561 	buf[*bs_size + 3] = 0xff;
562 	(*bs_size) += 4;
563 }
564 
vdec_h264_slice_lat_decode(void * h_vdec,struct mtk_vcodec_mem * bs,struct vdec_fb * fb,bool * res_chg)565 static int vdec_h264_slice_lat_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
566 				      struct vdec_fb *fb, bool *res_chg)
567 {
568 	struct vdec_h264_slice_inst *inst = h_vdec;
569 	struct vdec_vpu_inst *vpu = &inst->vpu;
570 	struct mtk_video_dec_buf *src_buf_info;
571 	int nal_start_idx, err, timeout = 0, i;
572 	unsigned int data[2];
573 	struct vdec_lat_buf *lat_buf;
574 	struct vdec_h264_slice_share_info *share_info;
575 	unsigned char *buf;
576 	struct mtk_vcodec_mem *mem;
577 
578 	if (vdec_msg_queue_init(&inst->ctx->msg_queue, inst->ctx,
579 				vdec_h264_slice_core_decode,
580 				sizeof(*share_info)))
581 		return -ENOMEM;
582 
583 	/* bs NULL means flush decoder */
584 	if (!bs) {
585 		vdec_msg_queue_wait_lat_buf_full(&inst->ctx->msg_queue);
586 		return vpu_dec_reset(vpu);
587 	}
588 
589 	if (inst->is_field_bitstream)
590 		return -EINVAL;
591 
592 	lat_buf = vdec_msg_queue_dqbuf(&inst->ctx->msg_queue.lat_ctx);
593 	if (!lat_buf) {
594 		mtk_vdec_debug(inst->ctx, "failed to get lat buffer");
595 		return -EAGAIN;
596 	}
597 	share_info = lat_buf->private_data;
598 	src_buf_info = container_of(bs, struct mtk_video_dec_buf, bs_buffer);
599 
600 	buf = (unsigned char *)bs->va;
601 	nal_start_idx = mtk_vdec_h264_find_start_code(buf, bs->size);
602 	if (nal_start_idx < 0) {
603 		err = -EINVAL;
604 		goto err_free_fb_out;
605 	}
606 
607 	inst->vsi->dec.nal_info = buf[nal_start_idx];
608 	lat_buf->src_buf_req = src_buf_info->m2m_buf.vb.vb2_buf.req_obj.req;
609 	v4l2_m2m_buf_copy_metadata(&src_buf_info->m2m_buf.vb, &lat_buf->ts_info, true);
610 
611 	err = vdec_h264_slice_fill_decode_parameters(inst, share_info);
612 	if (err)
613 		goto err_free_fb_out;
614 
615 	vdec_h264_insert_startcode(inst->ctx->dev, buf, &bs->size,
616 				   &share_info->h264_slice_params.pps);
617 
618 	inst->vsi->dec.bs_buf_addr = (uint64_t)bs->dma_addr;
619 	inst->vsi->dec.bs_buf_size = bs->size;
620 
621 	*res_chg = inst->resolution_changed;
622 	if (inst->resolution_changed) {
623 		mtk_vdec_debug(inst->ctx, "- resolution changed -");
624 		if (inst->realloc_mv_buf) {
625 			err = vdec_h264_slice_alloc_mv_buf(inst, &inst->ctx->picinfo);
626 			inst->realloc_mv_buf = false;
627 			if (err)
628 				goto err_free_fb_out;
629 		}
630 		inst->resolution_changed = false;
631 	}
632 	for (i = 0; i < H264_MAX_MV_NUM; i++) {
633 		mem = &inst->mv_buf[i];
634 		inst->vsi->mv_buf_dma[i] = mem->dma_addr;
635 	}
636 	inst->vsi->wdma_start_addr = lat_buf->ctx->msg_queue.wdma_addr.dma_addr;
637 	inst->vsi->wdma_end_addr = lat_buf->ctx->msg_queue.wdma_addr.dma_addr +
638 		lat_buf->ctx->msg_queue.wdma_addr.size;
639 	inst->vsi->wdma_err_addr = lat_buf->wdma_err_addr.dma_addr;
640 	inst->vsi->slice_bc_start_addr = lat_buf->slice_bc_addr.dma_addr;
641 	inst->vsi->slice_bc_end_addr = lat_buf->slice_bc_addr.dma_addr +
642 		lat_buf->slice_bc_addr.size;
643 
644 	inst->vsi->trans_end = inst->ctx->msg_queue.wdma_rptr_addr;
645 	inst->vsi->trans_start = inst->ctx->msg_queue.wdma_wptr_addr;
646 	mtk_vdec_debug(inst->ctx, "lat:trans(0x%llx 0x%llx) err:0x%llx",
647 		       inst->vsi->wdma_start_addr,
648 		       inst->vsi->wdma_end_addr,
649 		       inst->vsi->wdma_err_addr);
650 
651 	mtk_vdec_debug(inst->ctx, "slice(0x%llx 0x%llx) rprt((0x%llx 0x%llx))",
652 		       inst->vsi->slice_bc_start_addr,
653 		       inst->vsi->slice_bc_end_addr,
654 		       inst->vsi->trans_start,
655 		       inst->vsi->trans_end);
656 	err = vpu_dec_start(vpu, data, 2);
657 	if (err) {
658 		mtk_vdec_debug(inst->ctx, "lat decode err: %d", err);
659 		goto err_free_fb_out;
660 	}
661 
662 	share_info->trans_end = inst->ctx->msg_queue.wdma_addr.dma_addr +
663 		inst->vsi->wdma_end_addr_offset;
664 	share_info->trans_start = inst->ctx->msg_queue.wdma_wptr_addr;
665 	share_info->nal_info = inst->vsi->dec.nal_info;
666 
667 	if (IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability)) {
668 		memcpy(&share_info->h264_slice_params, &inst->vsi->h264_slice_params,
669 		       sizeof(share_info->h264_slice_params));
670 		vdec_msg_queue_qbuf(&inst->ctx->msg_queue.core_ctx, lat_buf);
671 	}
672 
673 	/* wait decoder done interrupt */
674 	timeout = mtk_vcodec_wait_for_done_ctx(inst->ctx, MTK_INST_IRQ_RECEIVED,
675 					       WAIT_INTR_TIMEOUT_MS, MTK_VDEC_LAT0);
676 	if (timeout)
677 		mtk_vdec_err(inst->ctx, "lat decode timeout: pic_%d", inst->slice_dec_num);
678 	inst->vsi->dec.timeout = !!timeout;
679 
680 	err = vpu_dec_end(vpu);
681 	if (err == SLICE_HEADER_FULL || err == TRANS_BUFFER_FULL) {
682 		if (!IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability))
683 			vdec_msg_queue_qbuf(&inst->ctx->msg_queue.lat_ctx, lat_buf);
684 		inst->slice_dec_num++;
685 		mtk_vdec_err(inst->ctx, "lat dec fail: pic_%d err:%d", inst->slice_dec_num, err);
686 		return -EINVAL;
687 	}
688 
689 	share_info->trans_end = inst->ctx->msg_queue.wdma_addr.dma_addr +
690 		inst->vsi->wdma_end_addr_offset;
691 	vdec_msg_queue_update_ube_wptr(&lat_buf->ctx->msg_queue, share_info->trans_end);
692 
693 	if (!IS_VDEC_INNER_RACING(inst->ctx->dev->dec_capability)) {
694 		memcpy(&share_info->h264_slice_params, &inst->vsi->h264_slice_params,
695 		       sizeof(share_info->h264_slice_params));
696 		vdec_msg_queue_qbuf(&inst->ctx->msg_queue.core_ctx, lat_buf);
697 	}
698 	mtk_vdec_debug(inst->ctx, "dec num: %d lat crc: 0x%x 0x%x 0x%x", inst->slice_dec_num,
699 		       inst->vsi->dec.crc[0], inst->vsi->dec.crc[1], inst->vsi->dec.crc[2]);
700 
701 	inst->slice_dec_num++;
702 	return 0;
703 err_free_fb_out:
704 	vdec_msg_queue_qbuf(&inst->ctx->msg_queue.lat_ctx, lat_buf);
705 	mtk_vdec_err(inst->ctx, "slice dec number: %d err: %d", inst->slice_dec_num, err);
706 	return err;
707 }
708 
vdec_h264_slice_single_decode(void * h_vdec,struct mtk_vcodec_mem * bs,struct vdec_fb * unused,bool * res_chg)709 static int vdec_h264_slice_single_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
710 					 struct vdec_fb *unused, bool *res_chg)
711 {
712 	struct vdec_h264_slice_inst *inst = h_vdec;
713 	struct vdec_vpu_inst *vpu = &inst->vpu;
714 	struct mtk_video_dec_buf *src_buf_info, *dst_buf_info;
715 	struct vdec_fb *fb;
716 	unsigned char *buf;
717 	unsigned int data[2], i;
718 	u64 y_fb_dma, c_fb_dma;
719 	struct mtk_vcodec_mem *mem;
720 	int err, nal_start_idx;
721 
722 	/* bs NULL means flush decoder */
723 	if (!bs)
724 		return vpu_dec_reset(vpu);
725 
726 	fb = inst->ctx->dev->vdec_pdata->get_cap_buffer(inst->ctx);
727 	if (!fb) {
728 		mtk_vdec_err(inst->ctx, "fb buffer is NULL");
729 		return -ENOMEM;
730 	}
731 
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->base_y.dma_addr;
736 	c_fb_dma = fb->base_c.dma_addr;
737 	mtk_vdec_debug(inst->ctx, "[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_vdec_debug(inst->ctx, "- 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_vdec_err(inst->ctx, "decode timeout: pic_%d", inst->ctx->decoded_frame_cnt);
787 
788 	inst->vsi->dec.timeout = !!err;
789 	err = vpu_dec_end(vpu);
790 	if (err)
791 		goto err_free_fb_out;
792 
793 	memcpy(&inst->vsi_ctx, inst->vpu.vsi, sizeof(inst->vsi_ctx));
794 	mtk_vdec_debug(inst->ctx, "pic[%d] crc: 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x 0x%x",
795 		       inst->ctx->decoded_frame_cnt,
796 		       inst->vsi_ctx.dec.crc[0], inst->vsi_ctx.dec.crc[1],
797 		       inst->vsi_ctx.dec.crc[2], inst->vsi_ctx.dec.crc[3],
798 		       inst->vsi_ctx.dec.crc[4], inst->vsi_ctx.dec.crc[5],
799 		       inst->vsi_ctx.dec.crc[6], inst->vsi_ctx.dec.crc[7]);
800 
801 	inst->ctx->decoded_frame_cnt++;
802 	return 0;
803 
804 err_free_fb_out:
805 	mtk_vdec_err(inst->ctx, "dec frame number: %d err: %d", inst->ctx->decoded_frame_cnt, err);
806 	return err;
807 }
808 
vdec_h264_slice_decode(void * h_vdec,struct mtk_vcodec_mem * bs,struct vdec_fb * unused,bool * res_chg)809 static int vdec_h264_slice_decode(void *h_vdec, struct mtk_vcodec_mem *bs,
810 				  struct vdec_fb *unused, bool *res_chg)
811 {
812 	struct vdec_h264_slice_inst *inst = h_vdec;
813 	int ret;
814 
815 	if (!h_vdec)
816 		return -EINVAL;
817 
818 	if (inst->ctx->dev->vdec_pdata->hw_arch == MTK_VDEC_PURE_SINGLE_CORE)
819 		ret = vdec_h264_slice_single_decode(h_vdec, bs, unused, res_chg);
820 	else
821 		ret = vdec_h264_slice_lat_decode(h_vdec, bs, unused, res_chg);
822 
823 	return ret;
824 }
825 
vdec_h264_slice_get_param(void * h_vdec,enum vdec_get_param_type type,void * out)826 static int vdec_h264_slice_get_param(void *h_vdec, enum vdec_get_param_type type,
827 				     void *out)
828 {
829 	struct vdec_h264_slice_inst *inst = h_vdec;
830 
831 	switch (type) {
832 	case GET_PARAM_PIC_INFO:
833 		vdec_h264_slice_get_pic_info(inst);
834 		break;
835 	case GET_PARAM_DPB_SIZE:
836 		*(unsigned int *)out = 6;
837 		break;
838 	case GET_PARAM_CROP_INFO:
839 		vdec_h264_slice_get_crop_info(inst, out);
840 		break;
841 	default:
842 		mtk_vdec_err(inst->ctx, "invalid get parameter type=%d", type);
843 		return -EINVAL;
844 	}
845 	return 0;
846 }
847 
848 const struct vdec_common_if vdec_h264_slice_multi_if = {
849 	.init		= vdec_h264_slice_init,
850 	.decode		= vdec_h264_slice_decode,
851 	.get_param	= vdec_h264_slice_get_param,
852 	.deinit		= vdec_h264_slice_deinit,
853 };
854