1 /**************************************************************************
2 *
3 * Copyright 2015 Advanced Micro Devices, Inc.
4 *
5 * SPDX-License-Identifier: MIT
6 *
7 **************************************************************************/
8
9 #include "pipe/p_video_codec.h"
10 #include "radeon_vce.h"
11 #include "radeon_video.h"
12 #include "radeonsi/si_pipe.h"
13 #include "util/u_memory.h"
14 #include "util/u_video.h"
15 #include "vl/vl_video_buffer.h"
16
17 #include <stdio.h>
18
get_rate_control_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)19 static void get_rate_control_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
20 {
21 enc->enc_pic.rc.rc_method = pic->rate_ctrl[0].rate_ctrl_method;
22 enc->enc_pic.rc.target_bitrate = pic->rate_ctrl[0].target_bitrate;
23 enc->enc_pic.rc.peak_bitrate = pic->rate_ctrl[0].peak_bitrate;
24 enc->enc_pic.rc.quant_i_frames = pic->quant_i_frames;
25 enc->enc_pic.rc.quant_p_frames = pic->quant_p_frames;
26 enc->enc_pic.rc.quant_b_frames = pic->quant_b_frames;
27 enc->enc_pic.rc.gop_size = pic->gop_size;
28 enc->enc_pic.rc.frame_rate_num = pic->rate_ctrl[0].frame_rate_num;
29 enc->enc_pic.rc.frame_rate_den = pic->rate_ctrl[0].frame_rate_den;
30 enc->enc_pic.rc.max_qp = 51;
31 enc->enc_pic.rc.vbv_buffer_size = pic->rate_ctrl[0].vbv_buffer_size;
32 enc->enc_pic.rc.vbv_buf_lv = pic->rate_ctrl[0].vbv_buf_lv;
33 enc->enc_pic.rc.fill_data_enable = pic->rate_ctrl[0].fill_data_enable;
34 enc->enc_pic.rc.enforce_hrd = pic->rate_ctrl[0].enforce_hrd;
35 enc->enc_pic.rc.target_bits_picture =
36 enc->pic.rate_ctrl[0].target_bitrate *
37 ((float)enc->pic.rate_ctrl[0].frame_rate_den /
38 enc->pic.rate_ctrl[0].frame_rate_num);
39 enc->enc_pic.rc.peak_bits_picture_integer =
40 enc->pic.rate_ctrl[0].peak_bitrate *
41 ((float)enc->pic.rate_ctrl[0].frame_rate_den /
42 enc->pic.rate_ctrl[0].frame_rate_num);
43 enc->enc_pic.rc.peak_bits_picture_fraction =
44 (((enc->pic.rate_ctrl[0].peak_bitrate *
45 (uint64_t)enc->pic.rate_ctrl[0].frame_rate_den) %
46 enc->pic.rate_ctrl[0].frame_rate_num) << 32) /
47 enc->pic.rate_ctrl[0].frame_rate_num;
48 }
49
get_motion_estimation_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)50 static void get_motion_estimation_param(struct rvce_encoder *enc,
51 struct pipe_h264_enc_picture_desc *pic)
52 {
53 enc->enc_pic.me.motion_est_quarter_pixel = 1;
54 enc->enc_pic.me.enc_disable_sub_mode = 254;
55 enc->enc_pic.me.lsmvert = 2;
56 enc->enc_pic.me.enc_en_ime_overw_dis_subm = 0;
57 enc->enc_pic.me.enc_ime_overw_dis_subm_no = 0;
58 enc->enc_pic.me.enc_ime2_search_range_x = 4;
59 enc->enc_pic.me.enc_ime2_search_range_y = 4;
60 enc->enc_pic.me.enc_ime_decimation_search = 0x00000001;
61 enc->enc_pic.me.motion_est_half_pixel = 0x00000001;
62 enc->enc_pic.me.enc_search_range_x = 0x00000010;
63 enc->enc_pic.me.enc_search_range_y = 0x00000010;
64 enc->enc_pic.me.enc_search1_range_x = 0x00000010;
65 enc->enc_pic.me.enc_search1_range_y = 0x00000010;
66 }
67
get_pic_control_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)68 static void get_pic_control_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
69 {
70 unsigned encNumMBsPerSlice;
71 encNumMBsPerSlice = align(enc->base.width, 16) / 16;
72 encNumMBsPerSlice *= align(enc->base.height, 16) / 16;
73 if (pic->seq.enc_frame_cropping_flag) {
74 enc->enc_pic.pc.enc_crop_left_offset = pic->seq.enc_frame_crop_left_offset;
75 enc->enc_pic.pc.enc_crop_right_offset = pic->seq.enc_frame_crop_right_offset;
76 enc->enc_pic.pc.enc_crop_top_offset = pic->seq.enc_frame_crop_top_offset;
77 enc->enc_pic.pc.enc_crop_bottom_offset = pic->seq.enc_frame_crop_bottom_offset;
78 } else {
79 enc->enc_pic.pc.enc_crop_right_offset = (align(enc->base.width, 16) - enc->base.width) >> 1;
80 enc->enc_pic.pc.enc_crop_bottom_offset =
81 (align(enc->base.height, 16) - enc->base.height) >> 1;
82 }
83 enc->enc_pic.pc.enc_num_mbs_per_slice = encNumMBsPerSlice;
84 enc->enc_pic.pc.enc_b_pic_pattern = MAX2(enc->base.max_references, 1) - 1;
85 enc->enc_pic.pc.enc_number_of_reference_frames = MIN2(enc->base.max_references, 1);
86 enc->enc_pic.pc.enc_max_num_ref_frames = enc->base.max_references + 1;
87 enc->enc_pic.pc.enc_num_default_active_ref_l0 = 0x00000001;
88 enc->enc_pic.pc.enc_num_default_active_ref_l1 = 0x00000001;
89 enc->enc_pic.pc.enc_cabac_enable = pic->pic_ctrl.enc_cabac_enable;
90 enc->enc_pic.pc.enc_constraint_set_flags = 0x00000040;
91 }
92
get_task_info_param(struct rvce_encoder * enc)93 static void get_task_info_param(struct rvce_encoder *enc)
94 {
95 enc->enc_pic.ti.offset_of_next_task_info = 0xffffffff;
96 }
97
get_feedback_buffer_param(struct rvce_encoder * enc,struct pipe_enc_feedback_metadata * metadata)98 static void get_feedback_buffer_param(struct rvce_encoder *enc, struct pipe_enc_feedback_metadata* metadata)
99 {
100 enc->enc_pic.fb.feedback_ring_size = 0x00000001;
101 }
102
get_config_ext_param(struct rvce_encoder * enc)103 static void get_config_ext_param(struct rvce_encoder *enc)
104 {
105 enc->enc_pic.ce.enc_enable_perf_logging = 0x00000003;
106 }
107
get_vui_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)108 static void get_vui_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
109 {
110 enc->enc_pic.enable_vui = pic->seq.vui_parameters_present_flag;
111 enc->enc_pic.vui.aspect_ratio_info_present_flag =
112 pic->seq.vui_flags.aspect_ratio_info_present_flag;
113 enc->enc_pic.vui.aspect_ratio_idc = pic->seq.aspect_ratio_idc;
114 enc->enc_pic.vui.sar_width = pic->seq.sar_width;
115 enc->enc_pic.vui.sar_height = pic->seq.sar_height;
116 enc->enc_pic.vui.video_signal_type_present_flag =
117 pic->seq.vui_flags.video_signal_type_present_flag;
118 enc->enc_pic.vui.video_format = pic->seq.video_format;
119 enc->enc_pic.vui.video_full_range_flag = pic->seq.video_full_range_flag;
120 enc->enc_pic.vui.color_description_present_flag =
121 pic->seq.vui_flags.colour_description_present_flag;
122 enc->enc_pic.vui.color_prim = pic->seq.colour_primaries;
123 enc->enc_pic.vui.transfer_char = pic->seq.transfer_characteristics;
124 enc->enc_pic.vui.matrix_coef = pic->seq.matrix_coefficients;
125 enc->enc_pic.vui.chroma_loc_info_present_flag =
126 pic->seq.vui_flags.chroma_loc_info_present_flag;
127 enc->enc_pic.vui.chroma_loc_top = pic->seq.chroma_sample_loc_type_top_field;
128 enc->enc_pic.vui.chroma_loc_bottom = pic->seq.chroma_sample_loc_type_bottom_field;
129 enc->enc_pic.vui.timing_info_present_flag = pic->seq.vui_flags.timing_info_present_flag;
130 enc->enc_pic.vui.num_units_in_tick = pic->rate_ctrl[0].frame_rate_den;
131 enc->enc_pic.vui.time_scale = pic->rate_ctrl[0].frame_rate_num * 2;
132 enc->enc_pic.vui.fixed_frame_rate_flag = 0x00000001;
133 enc->enc_pic.vui.bit_rate_scale = 0x00000004;
134 enc->enc_pic.vui.cpb_size_scale = 0x00000006;
135 enc->enc_pic.vui.initial_cpb_removal_delay_length_minus1 = 0x00000017;
136 enc->enc_pic.vui.cpb_removal_delay_length_minus1 = 0x00000017;
137 enc->enc_pic.vui.dpb_output_delay_length_minus1 = 0x00000017;
138 enc->enc_pic.vui.time_offset_length = 0x00000018;
139 enc->enc_pic.vui.motion_vectors_over_pic_boundaries_flag = 0x00000001;
140 enc->enc_pic.vui.max_bytes_per_pic_denom = 0x00000002;
141 enc->enc_pic.vui.max_bits_per_mb_denom = 0x00000001;
142 enc->enc_pic.vui.log2_max_mv_length_hori = 0x00000010;
143 enc->enc_pic.vui.log2_max_mv_length_vert = 0x00000010;
144 enc->enc_pic.vui.num_reorder_frames = 0x00000003;
145 enc->enc_pic.vui.max_dec_frame_buffering = 0x00000003;
146 }
147
si_vce_52_get_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)148 void si_vce_52_get_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
149 {
150 get_rate_control_param(enc, pic);
151 get_motion_estimation_param(enc, pic);
152 get_pic_control_param(enc, pic);
153 get_task_info_param(enc);
154 get_feedback_buffer_param(enc, NULL);
155 get_vui_param(enc, pic);
156 get_config_ext_param(enc);
157
158 enc->enc_pic.picture_type = pic->picture_type;
159 enc->enc_pic.frame_num = pic->frame_num;
160 enc->enc_pic.frame_num_cnt = pic->frame_num_cnt;
161 enc->enc_pic.p_remain = pic->p_remain;
162 enc->enc_pic.i_remain = pic->i_remain;
163 enc->enc_pic.gop_cnt = pic->gop_cnt;
164 enc->enc_pic.pic_order_cnt = pic->pic_order_cnt;
165 enc->enc_pic.ref_idx_l0 = pic->ref_idx_l0_list[0];
166 enc->enc_pic.ref_idx_l1 = pic->ref_idx_l1_list[0];
167 enc->enc_pic.not_referenced = pic->not_referenced;
168 if (enc->dual_inst)
169 enc->enc_pic.addrmode_arraymode_disrdo_distwoinstants = 0x00000201;
170 else
171 enc->enc_pic.addrmode_arraymode_disrdo_distwoinstants = 0x01000201;
172 enc->enc_pic.is_idr = (pic->picture_type == PIPE_H2645_ENC_PICTURE_TYPE_IDR);
173 }
174
create(struct rvce_encoder * enc)175 static void create(struct rvce_encoder *enc)
176 {
177 struct si_screen *sscreen = (struct si_screen *)enc->screen;
178 enc->task_info(enc, 0x00000000, 0, 0, 0);
179
180 RVCE_BEGIN(0x01000001); // create cmd
181 RVCE_CS(enc->enc_pic.ec.enc_use_circular_buffer);
182 RVCE_CS(u_get_h264_profile_idc(enc->base.profile)); // encProfile
183 RVCE_CS(enc->pic.seq.level_idc); // encLevel
184 RVCE_CS(enc->enc_pic.ec.enc_pic_struct_restriction);
185 RVCE_CS(enc->base.width); // encImageWidth
186 RVCE_CS(enc->base.height); // encImageHeight
187
188 if (sscreen->info.gfx_level < GFX9) {
189 RVCE_CS(enc->luma->u.legacy.level[0].nblk_x * enc->luma->bpe); // encRefPicLumaPitch
190 RVCE_CS(enc->chroma->u.legacy.level[0].nblk_x * enc->chroma->bpe); // encRefPicChromaPitch
191 RVCE_CS(align(enc->luma->u.legacy.level[0].nblk_y, 16) / 8); // encRefYHeightInQw
192 } else {
193 RVCE_CS(enc->luma->u.gfx9.surf_pitch * enc->luma->bpe); // encRefPicLumaPitch
194 RVCE_CS(enc->chroma->u.gfx9.surf_pitch * enc->chroma->bpe); // encRefPicChromaPitch
195 RVCE_CS(align(enc->luma->u.gfx9.surf_height, 16) / 8); // encRefYHeightInQw
196 }
197
198 RVCE_CS(enc->enc_pic.addrmode_arraymode_disrdo_distwoinstants);
199
200 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_context_buffer_offset);
201 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_input_luma_buffer_offset);
202 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_input_chroma_buffer_offset);
203 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_mode_chromaflag_vbaqmode_scenechangesensitivity);
204 RVCE_END();
205 }
206
encode(struct rvce_encoder * enc)207 static void encode(struct rvce_encoder *enc)
208 {
209 struct si_screen *sscreen = (struct si_screen *)enc->screen;
210 signed luma_offset, chroma_offset, bs_offset;
211 unsigned dep, bs_idx = enc->bs_idx++;
212 int i;
213
214 if (enc->dual_inst) {
215 if (bs_idx == 0)
216 dep = 1;
217 else if (enc->enc_pic.picture_type == PIPE_H2645_ENC_PICTURE_TYPE_IDR)
218 dep = 0;
219 else
220 dep = 2;
221 } else
222 dep = 0;
223
224 enc->task_info(enc, 0x00000003, dep, 0, bs_idx);
225
226 RVCE_BEGIN(0x05000001); // context buffer
227 RVCE_READWRITE(enc->cpb.res->buf, enc->cpb.res->domains, 0); // encodeContextAddressHi/Lo
228 RVCE_END();
229
230 bs_offset = -(signed)(bs_idx * enc->bs_size);
231
232 RVCE_BEGIN(0x05000004); // video bitstream buffer
233 RVCE_WRITE(enc->bs_handle, RADEON_DOMAIN_GTT, bs_offset); // videoBitstreamRingAddressHi/Lo
234 RVCE_CS(enc->bs_size); // videoBitstreamRingSize
235 RVCE_END();
236
237 if (enc->dual_pipe) {
238 unsigned aux_offset =
239 enc->cpb.res->buf->size - RVCE_MAX_AUX_BUFFER_NUM * RVCE_MAX_BITSTREAM_OUTPUT_ROW_SIZE * 2;
240 RVCE_BEGIN(0x05000002); // auxiliary buffer
241 for (i = 0; i < 8; ++i) {
242 RVCE_CS(aux_offset);
243 aux_offset += RVCE_MAX_BITSTREAM_OUTPUT_ROW_SIZE;
244 }
245 for (i = 0; i < 8; ++i)
246 RVCE_CS(RVCE_MAX_BITSTREAM_OUTPUT_ROW_SIZE);
247 RVCE_END();
248 }
249
250 RVCE_BEGIN(0x03000001); // encode
251 RVCE_CS(enc->enc_pic.frame_num ? 0x0 : 0x11); // insertHeaders
252 RVCE_CS(enc->enc_pic.eo.picture_structure);
253 RVCE_CS(enc->bs_size); // allowedMaxBitstreamSize
254 RVCE_CS(enc->enc_pic.eo.force_refresh_map);
255 RVCE_CS(enc->enc_pic.eo.insert_aud);
256 RVCE_CS(enc->enc_pic.eo.end_of_sequence);
257 RVCE_CS(enc->enc_pic.eo.end_of_stream);
258
259 if (sscreen->info.gfx_level < GFX9) {
260 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
261 (uint64_t)enc->luma->u.legacy.level[0].offset_256B * 256); // inputPictureLumaAddressHi/Lo
262 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
263 (uint64_t)enc->chroma->u.legacy.level[0].offset_256B * 256); // inputPictureChromaAddressHi/Lo
264 RVCE_CS(align(enc->luma->u.legacy.level[0].nblk_y, 16)); // encInputFrameYPitch
265 RVCE_CS(enc->luma->u.legacy.level[0].nblk_x * enc->luma->bpe); // encInputPicLumaPitch
266 RVCE_CS(enc->chroma->u.legacy.level[0].nblk_x * enc->chroma->bpe); // encInputPicChromaPitch
267 } else {
268 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
269 enc->luma->u.gfx9.surf_offset); // inputPictureLumaAddressHi/Lo
270 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
271 enc->chroma->u.gfx9.surf_offset); // inputPictureChromaAddressHi/Lo
272 RVCE_CS(align(enc->luma->u.gfx9.surf_height, 16)); // encInputFrameYPitch
273 RVCE_CS(enc->luma->u.gfx9.surf_pitch * enc->luma->bpe); // encInputPicLumaPitch
274 RVCE_CS(enc->chroma->u.gfx9.surf_pitch * enc->chroma->bpe); // encInputPicChromaPitch
275 }
276
277 if (enc->dual_pipe)
278 enc->enc_pic.eo.enc_input_pic_addr_array_disable2pipe_disablemboffload = 0x00000000;
279 else
280 enc->enc_pic.eo.enc_input_pic_addr_array_disable2pipe_disablemboffload = 0x00010000;
281 RVCE_CS(enc->enc_pic.eo.enc_input_pic_addr_array_disable2pipe_disablemboffload);
282 RVCE_CS(enc->enc_pic.eo.enc_input_pic_tile_config);
283 RVCE_CS(enc->enc_pic.picture_type); // encPicType
284 RVCE_CS(enc->enc_pic.picture_type == PIPE_H2645_ENC_PICTURE_TYPE_IDR); // encIdrFlag
285 if ((enc->enc_pic.picture_type == PIPE_H2645_ENC_PICTURE_TYPE_IDR) &&
286 (enc->enc_pic.eo.enc_idr_pic_id != 0))
287 enc->enc_pic.eo.enc_idr_pic_id = enc->enc_pic.idr_pic_id - 1;
288 else
289 enc->enc_pic.eo.enc_idr_pic_id = 0x00000000;
290 RVCE_CS(enc->enc_pic.eo.enc_idr_pic_id);
291 RVCE_CS(enc->enc_pic.eo.enc_mgs_key_pic);
292 RVCE_CS(!enc->enc_pic.not_referenced);
293 RVCE_CS(enc->enc_pic.eo.enc_temporal_layer_index);
294 RVCE_CS(enc->enc_pic.eo.num_ref_idx_active_override_flag);
295 RVCE_CS(enc->enc_pic.eo.num_ref_idx_l0_active_minus1);
296 RVCE_CS(enc->enc_pic.eo.num_ref_idx_l1_active_minus1);
297
298 i = enc->enc_pic.frame_num - enc->enc_pic.ref_idx_l0;
299 if (i > 1 && enc->enc_pic.picture_type == PIPE_H2645_ENC_PICTURE_TYPE_P) {
300 enc->enc_pic.eo.enc_ref_list_modification_op = 0x00000001;
301 enc->enc_pic.eo.enc_ref_list_modification_num = i - 1;
302 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_op);
303 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_num);
304 } else {
305 enc->enc_pic.eo.enc_ref_list_modification_op = 0x00000000;
306 enc->enc_pic.eo.enc_ref_list_modification_num = 0x00000000;
307 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_op);
308 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_num);
309 }
310
311 for (i = 0; i < 3; ++i) {
312 enc->enc_pic.eo.enc_ref_list_modification_op = 0x00000000;
313 enc->enc_pic.eo.enc_ref_list_modification_num = 0x00000000;
314 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_op);
315 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_num);
316 }
317 for (i = 0; i < 4; ++i) {
318 RVCE_CS(enc->enc_pic.eo.enc_decoded_picture_marking_op);
319 RVCE_CS(enc->enc_pic.eo.enc_decoded_picture_marking_num);
320 RVCE_CS(enc->enc_pic.eo.enc_decoded_picture_marking_idx);
321 RVCE_CS(enc->enc_pic.eo.enc_decoded_ref_base_picture_marking_op);
322 RVCE_CS(enc->enc_pic.eo.enc_decoded_ref_base_picture_marking_num);
323 }
324
325 // encReferencePictureL0[0]
326 RVCE_CS(0x00000000); // pictureStructure
327 if (enc->enc_pic.picture_type == PIPE_H2645_ENC_PICTURE_TYPE_P ||
328 enc->enc_pic.picture_type == PIPE_H2645_ENC_PICTURE_TYPE_B) {
329 struct rvce_cpb_slot *l0 = si_l0_slot(enc);
330 si_vce_frame_offset(enc, l0, &luma_offset, &chroma_offset);
331 RVCE_CS(l0->picture_type);
332 RVCE_CS(l0->frame_num);
333 RVCE_CS(l0->pic_order_cnt);
334 RVCE_CS(luma_offset);
335 RVCE_CS(chroma_offset);
336 } else {
337 enc->enc_pic.eo.l0_enc_pic_type = 0x00000000;
338 enc->enc_pic.eo.l0_frame_number = 0x00000000;
339 enc->enc_pic.eo.l0_picture_order_count = 0x00000000;
340 enc->enc_pic.eo.l0_luma_offset = 0xffffffff;
341 enc->enc_pic.eo.l0_chroma_offset = 0xffffffff;
342 RVCE_CS(enc->enc_pic.eo.l0_enc_pic_type);
343 RVCE_CS(enc->enc_pic.eo.l0_frame_number);
344 RVCE_CS(enc->enc_pic.eo.l0_picture_order_count);
345 RVCE_CS(enc->enc_pic.eo.l0_luma_offset);
346 RVCE_CS(enc->enc_pic.eo.l0_chroma_offset);
347 }
348
349 // encReferencePictureL0[1]
350 enc->enc_pic.eo.l0_picture_structure = 0x00000000;
351 enc->enc_pic.eo.l0_enc_pic_type = 0x00000000;
352 enc->enc_pic.eo.l0_frame_number = 0x00000000;
353 enc->enc_pic.eo.l0_picture_order_count = 0x00000000;
354 enc->enc_pic.eo.l0_luma_offset = 0xffffffff;
355 enc->enc_pic.eo.l0_chroma_offset = 0xffffffff;
356 RVCE_CS(enc->enc_pic.eo.l0_picture_structure);
357 RVCE_CS(enc->enc_pic.eo.l0_enc_pic_type);
358 RVCE_CS(enc->enc_pic.eo.l0_frame_number);
359 RVCE_CS(enc->enc_pic.eo.l0_picture_order_count);
360 RVCE_CS(enc->enc_pic.eo.l0_luma_offset);
361 RVCE_CS(enc->enc_pic.eo.l0_chroma_offset);
362
363 // encReferencePictureL1[0]
364 RVCE_CS(0x00000000); // pictureStructure
365 if (enc->enc_pic.picture_type == PIPE_H2645_ENC_PICTURE_TYPE_B) {
366 struct rvce_cpb_slot *l1 = si_l1_slot(enc);
367 si_vce_frame_offset(enc, l1, &luma_offset, &chroma_offset);
368 RVCE_CS(l1->picture_type);
369 RVCE_CS(l1->frame_num);
370 RVCE_CS(l1->pic_order_cnt);
371 RVCE_CS(luma_offset);
372 RVCE_CS(chroma_offset);
373 } else {
374 enc->enc_pic.eo.l1_enc_pic_type = 0x00000000;
375 enc->enc_pic.eo.l1_frame_number = 0x00000000;
376 enc->enc_pic.eo.l1_picture_order_count = 0x00000000;
377 enc->enc_pic.eo.l1_luma_offset = 0xffffffff;
378 enc->enc_pic.eo.l1_chroma_offset = 0xffffffff;
379 RVCE_CS(enc->enc_pic.eo.l1_enc_pic_type);
380 RVCE_CS(enc->enc_pic.eo.l1_frame_number);
381 RVCE_CS(enc->enc_pic.eo.l1_picture_order_count);
382 RVCE_CS(enc->enc_pic.eo.l1_luma_offset);
383 RVCE_CS(enc->enc_pic.eo.l1_chroma_offset);
384 }
385
386 si_vce_frame_offset(enc, si_current_slot(enc), &luma_offset, &chroma_offset);
387 RVCE_CS(luma_offset);
388 RVCE_CS(chroma_offset);
389 RVCE_CS(enc->enc_pic.eo.enc_coloc_buffer_offset);
390 RVCE_CS(enc->enc_pic.eo.enc_reconstructed_ref_base_picture_luma_offset);
391 RVCE_CS(enc->enc_pic.eo.enc_reconstructed_ref_base_picture_chroma_offset);
392 RVCE_CS(enc->enc_pic.eo.enc_reference_ref_base_picture_luma_offset);
393 RVCE_CS(enc->enc_pic.eo.enc_reference_ref_base_picture_chroma_offset);
394 RVCE_CS(enc->enc_pic.frame_num_cnt - 1);
395 RVCE_CS(enc->enc_pic.frame_num);
396 RVCE_CS(enc->enc_pic.pic_order_cnt);
397 RVCE_CS(enc->enc_pic.i_remain);
398 RVCE_CS(enc->enc_pic.p_remain);
399 RVCE_CS(enc->enc_pic.eo.num_b_pic_remain_in_rcgop);
400 RVCE_CS(enc->enc_pic.eo.num_ir_pic_remain_in_rcgop);
401 RVCE_CS(enc->enc_pic.eo.enable_intra_refresh);
402
403 RVCE_CS(enc->enc_pic.eo.aq_variance_en);
404 RVCE_CS(enc->enc_pic.eo.aq_block_size);
405 RVCE_CS(enc->enc_pic.eo.aq_mb_variance_sel);
406 RVCE_CS(enc->enc_pic.eo.aq_frame_variance_sel);
407 RVCE_CS(enc->enc_pic.eo.aq_param_a);
408 RVCE_CS(enc->enc_pic.eo.aq_param_b);
409 RVCE_CS(enc->enc_pic.eo.aq_param_c);
410 RVCE_CS(enc->enc_pic.eo.aq_param_d);
411 RVCE_CS(enc->enc_pic.eo.aq_param_e);
412
413 RVCE_CS(enc->enc_pic.eo.context_in_sfb);
414 RVCE_END();
415 }
416
rate_control(struct rvce_encoder * enc)417 static void rate_control(struct rvce_encoder *enc)
418 {
419 RVCE_BEGIN(0x04000005); // rate control
420 RVCE_CS(enc->enc_pic.rc.rc_method);
421 RVCE_CS(enc->enc_pic.rc.target_bitrate);
422 RVCE_CS(enc->enc_pic.rc.peak_bitrate);
423 RVCE_CS(enc->enc_pic.rc.frame_rate_num);
424 RVCE_CS(enc->enc_pic.rc.gop_size);
425 RVCE_CS(enc->enc_pic.rc.quant_i_frames);
426 RVCE_CS(enc->enc_pic.rc.quant_p_frames);
427 RVCE_CS(enc->enc_pic.rc.quant_b_frames);
428 RVCE_CS(enc->enc_pic.rc.vbv_buffer_size);
429 RVCE_CS(enc->enc_pic.rc.frame_rate_den);
430 RVCE_CS(enc->enc_pic.rc.vbv_buf_lv);
431 RVCE_CS(enc->enc_pic.rc.max_au_size);
432 RVCE_CS(enc->enc_pic.rc.qp_initial_mode);
433 RVCE_CS(enc->enc_pic.rc.target_bits_picture);
434 RVCE_CS(enc->enc_pic.rc.peak_bits_picture_integer);
435 RVCE_CS(enc->enc_pic.rc.peak_bits_picture_fraction);
436 RVCE_CS(enc->enc_pic.rc.min_qp);
437 RVCE_CS(enc->enc_pic.rc.max_qp);
438 RVCE_CS(enc->enc_pic.rc.skip_frame_enable);
439 RVCE_CS(enc->enc_pic.rc.fill_data_enable);
440 RVCE_CS(enc->enc_pic.rc.enforce_hrd);
441 RVCE_CS(enc->enc_pic.rc.b_pics_delta_qp);
442 RVCE_CS(enc->enc_pic.rc.ref_b_pics_delta_qp);
443 RVCE_CS(enc->enc_pic.rc.rc_reinit_disable);
444 RVCE_CS(enc->enc_pic.rc.enc_lcvbr_init_qp_flag);
445 RVCE_CS(enc->enc_pic.rc.lcvbrsatd_based_nonlinear_bit_budget_flag);
446 RVCE_END();
447 }
448
config(struct rvce_encoder * enc)449 static void config(struct rvce_encoder *enc)
450 {
451 enc->task_info(enc, 0x00000002, 0, 0xffffffff, 0);
452 enc->rate_control(enc);
453 enc->config_extension(enc);
454 enc->motion_estimation(enc);
455 enc->rdo(enc);
456 if (enc->use_vui)
457 enc->vui(enc);
458 enc->pic_control(enc);
459 }
460
config_extension(struct rvce_encoder * enc)461 static void config_extension(struct rvce_encoder *enc)
462 {
463 RVCE_BEGIN(0x04000001); // config extension
464 RVCE_CS(enc->enc_pic.ce.enc_enable_perf_logging);
465 RVCE_END();
466 }
467
feedback(struct rvce_encoder * enc)468 static void feedback(struct rvce_encoder *enc)
469 {
470 RVCE_BEGIN(0x05000005); // feedback buffer
471 RVCE_WRITE(enc->fb->res->buf, enc->fb->res->domains, 0x0); // feedbackRingAddressHi/Lo
472 RVCE_CS(enc->enc_pic.fb.feedback_ring_size);
473 RVCE_END();
474 }
475
destroy(struct rvce_encoder * enc)476 static void destroy(struct rvce_encoder *enc)
477 {
478 enc->task_info(enc, 0x00000001, 0, 0, 0);
479
480 feedback(enc);
481
482 RVCE_BEGIN(0x02000001); // destroy
483 RVCE_END();
484 }
485
motion_estimation(struct rvce_encoder * enc)486 static void motion_estimation(struct rvce_encoder *enc)
487 {
488 RVCE_BEGIN(0x04000007); // motion estimation
489 RVCE_CS(enc->enc_pic.me.enc_ime_decimation_search);
490 RVCE_CS(enc->enc_pic.me.motion_est_half_pixel);
491 RVCE_CS(enc->enc_pic.me.motion_est_quarter_pixel);
492 RVCE_CS(enc->enc_pic.me.disable_favor_pmv_point);
493 RVCE_CS(enc->enc_pic.me.force_zero_point_center);
494 RVCE_CS(enc->enc_pic.me.lsmvert);
495 RVCE_CS(enc->enc_pic.me.enc_search_range_x);
496 RVCE_CS(enc->enc_pic.me.enc_search_range_y);
497 RVCE_CS(enc->enc_pic.me.enc_search1_range_x);
498 RVCE_CS(enc->enc_pic.me.enc_search1_range_y);
499 RVCE_CS(enc->enc_pic.me.disable_16x16_frame1);
500 RVCE_CS(enc->enc_pic.me.disable_satd);
501 RVCE_CS(enc->enc_pic.me.enable_amd);
502 RVCE_CS(enc->enc_pic.me.enc_disable_sub_mode);
503 RVCE_CS(enc->enc_pic.me.enc_ime_skip_x);
504 RVCE_CS(enc->enc_pic.me.enc_ime_skip_y);
505 RVCE_CS(enc->enc_pic.me.enc_en_ime_overw_dis_subm);
506 RVCE_CS(enc->enc_pic.me.enc_ime_overw_dis_subm_no);
507 RVCE_CS(enc->enc_pic.me.enc_ime2_search_range_x);
508 RVCE_CS(enc->enc_pic.me.enc_ime2_search_range_y);
509 RVCE_CS(enc->enc_pic.me.parallel_mode_speedup_enable);
510 RVCE_CS(enc->enc_pic.me.fme0_enc_disable_sub_mode);
511 RVCE_CS(enc->enc_pic.me.fme1_enc_disable_sub_mode);
512 RVCE_CS(enc->enc_pic.me.ime_sw_speedup_enable);
513 RVCE_END();
514 }
515
pic_control(struct rvce_encoder * enc)516 static void pic_control(struct rvce_encoder *enc)
517 {
518 RVCE_BEGIN(0x04000002); // pic control
519 RVCE_CS(enc->enc_pic.pc.enc_use_constrained_intra_pred);
520 RVCE_CS(enc->enc_pic.pc.enc_cabac_enable);
521 RVCE_CS(enc->enc_pic.pc.enc_cabac_idc);
522 RVCE_CS(enc->enc_pic.pc.enc_loop_filter_disable);
523 RVCE_CS(enc->enc_pic.pc.enc_lf_beta_offset);
524 RVCE_CS(enc->enc_pic.pc.enc_lf_alpha_c0_offset);
525 RVCE_CS(enc->enc_pic.pc.enc_crop_left_offset);
526 RVCE_CS(enc->enc_pic.pc.enc_crop_right_offset);
527 RVCE_CS(enc->enc_pic.pc.enc_crop_top_offset);
528 RVCE_CS(enc->enc_pic.pc.enc_crop_bottom_offset);
529 RVCE_CS(enc->enc_pic.pc.enc_num_mbs_per_slice);
530 RVCE_CS(enc->enc_pic.pc.enc_intra_refresh_num_mbs_per_slot);
531 RVCE_CS(enc->enc_pic.pc.enc_force_intra_refresh);
532 RVCE_CS(enc->enc_pic.pc.enc_force_imb_period);
533 RVCE_CS(enc->enc_pic.pc.enc_pic_order_cnt_type);
534 RVCE_CS(enc->enc_pic.pc.log2_max_pic_order_cnt_lsb_minus4);
535 RVCE_CS(enc->enc_pic.pc.enc_sps_id);
536 RVCE_CS(enc->enc_pic.pc.enc_pps_id);
537 RVCE_CS(enc->enc_pic.pc.enc_constraint_set_flags);
538 RVCE_CS(enc->enc_pic.pc.enc_b_pic_pattern);
539 RVCE_CS(enc->enc_pic.pc.weight_pred_mode_b_picture);
540 RVCE_CS(enc->enc_pic.pc.enc_number_of_reference_frames);
541 RVCE_CS(enc->enc_pic.pc.enc_max_num_ref_frames);
542 RVCE_CS(enc->enc_pic.pc.enc_num_default_active_ref_l0);
543 RVCE_CS(enc->enc_pic.pc.enc_num_default_active_ref_l1);
544 RVCE_CS(enc->enc_pic.pc.enc_slice_mode);
545 RVCE_CS(enc->enc_pic.pc.enc_max_slice_size);
546 RVCE_END();
547 }
548
rdo(struct rvce_encoder * enc)549 static void rdo(struct rvce_encoder *enc)
550 {
551 RVCE_BEGIN(0x04000008); // rdo
552 RVCE_CS(enc->enc_pic.rdo.enc_disable_tbe_pred_i_frame);
553 RVCE_CS(enc->enc_pic.rdo.enc_disable_tbe_pred_p_frame);
554 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_y);
555 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_uv);
556 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_y);
557 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_uv);
558 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_y_1);
559 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_uv_1);
560 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_y_1);
561 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_uv_1);
562 RVCE_CS(enc->enc_pic.rdo.enc_16x16_cost_adj);
563 RVCE_CS(enc->enc_pic.rdo.enc_skip_cost_adj);
564 RVCE_CS(enc->enc_pic.rdo.enc_force_16x16_skip);
565 RVCE_CS(enc->enc_pic.rdo.enc_disable_threshold_calc_a);
566 RVCE_CS(enc->enc_pic.rdo.enc_luma_coeff_cost);
567 RVCE_CS(enc->enc_pic.rdo.enc_luma_mb_coeff_cost);
568 RVCE_CS(enc->enc_pic.rdo.enc_chroma_coeff_cost);
569 RVCE_END();
570 }
571
session(struct rvce_encoder * enc)572 static void session(struct rvce_encoder *enc)
573 {
574 RVCE_BEGIN(0x00000001); // session cmd
575 RVCE_CS(enc->stream_handle);
576 RVCE_END();
577 }
578
task_info(struct rvce_encoder * enc,uint32_t op,uint32_t dep,uint32_t fb_idx,uint32_t ring_idx)579 static void task_info(struct rvce_encoder *enc, uint32_t op, uint32_t dep, uint32_t fb_idx,
580 uint32_t ring_idx)
581 {
582 RVCE_BEGIN(0x00000002); // task info
583 if (op == 0x3) {
584 if (enc->task_info_idx) {
585 uint32_t offs = enc->cs.current.cdw - enc->task_info_idx + 3;
586 // Update offsetOfNextTaskInfo
587 enc->cs.current.buf[enc->task_info_idx] = offs;
588 }
589 enc->task_info_idx = enc->cs.current.cdw;
590 }
591 enc->enc_pic.ti.task_operation = op;
592 enc->enc_pic.ti.reference_picture_dependency = dep;
593 enc->enc_pic.ti.feedback_index = fb_idx;
594 enc->enc_pic.ti.video_bitstream_ring_index = ring_idx;
595 RVCE_CS(enc->enc_pic.ti.offset_of_next_task_info);
596 RVCE_CS(enc->enc_pic.ti.task_operation);
597 RVCE_CS(enc->enc_pic.ti.reference_picture_dependency);
598 RVCE_CS(enc->enc_pic.ti.collocate_flag_dependency);
599 RVCE_CS(enc->enc_pic.ti.feedback_index);
600 RVCE_CS(enc->enc_pic.ti.video_bitstream_ring_index);
601 RVCE_END();
602 }
603
vui(struct rvce_encoder * enc)604 static void vui(struct rvce_encoder *enc)
605 {
606 int i;
607
608 if (!enc->enc_pic.enable_vui)
609 return;
610
611 RVCE_BEGIN(0x04000009); // vui
612 RVCE_CS(enc->enc_pic.vui.aspect_ratio_info_present_flag);
613 RVCE_CS(enc->enc_pic.vui.aspect_ratio_idc);
614 RVCE_CS(enc->enc_pic.vui.sar_width);
615 RVCE_CS(enc->enc_pic.vui.sar_height);
616 RVCE_CS(enc->enc_pic.vui.overscan_info_present_flag);
617 RVCE_CS(enc->enc_pic.vui.overscan_Approp_flag);
618 RVCE_CS(enc->enc_pic.vui.video_signal_type_present_flag);
619 RVCE_CS(enc->enc_pic.vui.video_format);
620 RVCE_CS(enc->enc_pic.vui.video_full_range_flag);
621 RVCE_CS(enc->enc_pic.vui.color_description_present_flag);
622 RVCE_CS(enc->enc_pic.vui.color_prim);
623 RVCE_CS(enc->enc_pic.vui.transfer_char);
624 RVCE_CS(enc->enc_pic.vui.matrix_coef);
625 RVCE_CS(enc->enc_pic.vui.chroma_loc_info_present_flag);
626 RVCE_CS(enc->enc_pic.vui.chroma_loc_top);
627 RVCE_CS(enc->enc_pic.vui.chroma_loc_bottom);
628 RVCE_CS(enc->enc_pic.vui.timing_info_present_flag);
629 RVCE_CS(enc->enc_pic.vui.num_units_in_tick);
630 RVCE_CS(enc->enc_pic.vui.time_scale);
631 RVCE_CS(enc->enc_pic.vui.fixed_frame_rate_flag);
632 RVCE_CS(enc->enc_pic.vui.nal_hrd_parameters_present_flag);
633 RVCE_CS(enc->enc_pic.vui.cpb_cnt_minus1);
634 RVCE_CS(enc->enc_pic.vui.bit_rate_scale);
635 RVCE_CS(enc->enc_pic.vui.cpb_size_scale);
636 for (i = 0; i < 32; i++) {
637 RVCE_CS(enc->enc_pic.vui.bit_rate_value_minus);
638 RVCE_CS(enc->enc_pic.vui.cpb_size_value_minus);
639 RVCE_CS(enc->enc_pic.vui.cbr_flag);
640 }
641 RVCE_CS(enc->enc_pic.vui.initial_cpb_removal_delay_length_minus1);
642 RVCE_CS(enc->enc_pic.vui.cpb_removal_delay_length_minus1);
643 RVCE_CS(enc->enc_pic.vui.dpb_output_delay_length_minus1);
644 RVCE_CS(enc->enc_pic.vui.time_offset_length);
645 RVCE_CS(enc->enc_pic.vui.low_delay_hrd_flag);
646 RVCE_CS(enc->enc_pic.vui.pic_struct_present_flag);
647 RVCE_CS(enc->enc_pic.vui.bitstream_restriction_present_flag);
648 RVCE_CS(enc->enc_pic.vui.motion_vectors_over_pic_boundaries_flag);
649 RVCE_CS(enc->enc_pic.vui.max_bytes_per_pic_denom);
650 RVCE_CS(enc->enc_pic.vui.max_bits_per_mb_denom);
651 RVCE_CS(enc->enc_pic.vui.log2_max_mv_length_hori);
652 RVCE_CS(enc->enc_pic.vui.log2_max_mv_length_vert);
653 RVCE_CS(enc->enc_pic.vui.num_reorder_frames);
654 RVCE_CS(enc->enc_pic.vui.max_dec_frame_buffering);
655 RVCE_END();
656 }
657
si_vce_52_init(struct rvce_encoder * enc)658 void si_vce_52_init(struct rvce_encoder *enc)
659 {
660 enc->session = session;
661 enc->task_info = task_info;
662 enc->create = create;
663 enc->feedback = feedback;
664 enc->rate_control = rate_control;
665 enc->config_extension = config_extension;
666 enc->pic_control = pic_control;
667 enc->motion_estimation = motion_estimation;
668 enc->rdo = rdo;
669 enc->vui = vui;
670 enc->config = config;
671 enc->encode = encode;
672 enc->destroy = destroy;
673 enc->si_get_pic_param = si_vce_52_get_param;
674 }
675