1 /**************************************************************************
2 *
3 * Copyright 2015 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include <stdio.h>
29
30 #include "pipe/p_video_codec.h"
31
32 #include "util/u_video.h"
33 #include "util/u_memory.h"
34
35 #include "vl/vl_video_buffer.h"
36
37 #include "radeonsi/si_pipe.h"
38 #include "radeon_video.h"
39 #include "radeon_vce.h"
40
41 static const unsigned profiles[7] = { 66, 77, 88, 100, 110, 122, 244 };
42
get_rate_control_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)43 static void get_rate_control_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
44 {
45 enc->enc_pic.rc.rc_method = pic->rate_ctrl.rate_ctrl_method;
46 enc->enc_pic.rc.target_bitrate = pic->rate_ctrl.target_bitrate;
47 enc->enc_pic.rc.peak_bitrate = pic->rate_ctrl.peak_bitrate;
48 enc->enc_pic.rc.quant_i_frames = pic->quant_i_frames;
49 enc->enc_pic.rc.quant_p_frames = pic->quant_p_frames;
50 enc->enc_pic.rc.quant_b_frames = pic->quant_b_frames;
51 enc->enc_pic.rc.gop_size = pic->gop_size;
52 enc->enc_pic.rc.frame_rate_num = pic->rate_ctrl.frame_rate_num;
53 enc->enc_pic.rc.frame_rate_den = pic->rate_ctrl.frame_rate_den;
54 enc->enc_pic.rc.max_qp = 51;
55 enc->enc_pic.rc.vbv_buffer_size = pic->rate_ctrl.vbv_buffer_size;
56 enc->enc_pic.rc.vbv_buf_lv = pic->rate_ctrl.vbv_buf_lv;
57 enc->enc_pic.rc.fill_data_enable = pic->rate_ctrl.fill_data_enable;
58 enc->enc_pic.rc.enforce_hrd = pic->rate_ctrl.enforce_hrd;
59 enc->enc_pic.rc.target_bits_picture = pic->rate_ctrl.target_bits_picture;
60 enc->enc_pic.rc.peak_bits_picture_integer = pic->rate_ctrl.peak_bits_picture_integer;
61 enc->enc_pic.rc.peak_bits_picture_fraction = pic->rate_ctrl.peak_bits_picture_fraction;
62 }
63
get_motion_estimation_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)64 static void get_motion_estimation_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
65 {
66 enc->enc_pic.me.motion_est_quarter_pixel = pic->motion_est.motion_est_quarter_pixel;
67 enc->enc_pic.me.enc_disable_sub_mode = pic->motion_est.enc_disable_sub_mode;
68 enc->enc_pic.me.lsmvert = pic->motion_est.lsmvert;
69 enc->enc_pic.me.enc_en_ime_overw_dis_subm = pic->motion_est.enc_en_ime_overw_dis_subm;
70 enc->enc_pic.me.enc_ime_overw_dis_subm_no = pic->motion_est.enc_ime_overw_dis_subm_no;
71 enc->enc_pic.me.enc_ime2_search_range_x = pic->motion_est.enc_ime2_search_range_x;
72 enc->enc_pic.me.enc_ime2_search_range_y = pic->motion_est.enc_ime2_search_range_y;
73 enc->enc_pic.me.enc_ime_decimation_search = 0x00000001;
74 enc->enc_pic.me.motion_est_half_pixel = 0x00000001;
75 enc->enc_pic.me.enc_search_range_x = 0x00000010;
76 enc->enc_pic.me.enc_search_range_y = 0x00000010;
77 enc->enc_pic.me.enc_search1_range_x = 0x00000010;
78 enc->enc_pic.me.enc_search1_range_y = 0x00000010;
79 }
80
get_pic_control_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)81 static void get_pic_control_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
82 {
83 unsigned encNumMBsPerSlice;
84 encNumMBsPerSlice = align(enc->base.width, 16) / 16;
85 encNumMBsPerSlice *= align(enc->base.height, 16) / 16;
86 enc->enc_pic.pc.enc_crop_right_offset = (align(enc->base.width, 16) - enc->base.width) >> 1;
87 enc->enc_pic.pc.enc_crop_bottom_offset = (align(enc->base.height, 16) - enc->base.height) >> 1;
88 enc->enc_pic.pc.enc_num_mbs_per_slice = encNumMBsPerSlice;
89 enc->enc_pic.pc.enc_b_pic_pattern = MAX2(enc->base.max_references, 1) - 1;
90 enc->enc_pic.pc.enc_number_of_reference_frames = MIN2(enc->base.max_references, 2);
91 enc->enc_pic.pc.enc_max_num_ref_frames = enc->base.max_references + 1;
92 enc->enc_pic.pc.enc_num_default_active_ref_l0 = 0x00000001;
93 enc->enc_pic.pc.enc_num_default_active_ref_l1 = 0x00000001;
94 enc->enc_pic.pc.enc_cabac_enable = pic->pic_ctrl.enc_cabac_enable;
95 enc->enc_pic.pc.enc_constraint_set_flags = pic->pic_ctrl.enc_constraint_set_flags;
96 enc->enc_pic.pc.enc_num_default_active_ref_l0 = 0x00000001;
97 enc->enc_pic.pc.enc_num_default_active_ref_l1 = 0x00000001;
98 }
99
get_task_info_param(struct rvce_encoder * enc)100 static void get_task_info_param(struct rvce_encoder *enc)
101 {
102 enc->enc_pic.ti.offset_of_next_task_info = 0xffffffff;
103 }
104
get_feedback_buffer_param(struct rvce_encoder * enc)105 static void get_feedback_buffer_param(struct rvce_encoder *enc)
106 {
107 enc->enc_pic.fb.feedback_ring_size = 0x00000001;
108 }
109
get_config_ext_param(struct rvce_encoder * enc)110 static void get_config_ext_param(struct rvce_encoder *enc)
111 {
112 enc->enc_pic.ce.enc_enable_perf_logging = 0x00000003;
113 }
114
get_vui_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)115 static void get_vui_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
116 {
117 enc->enc_pic.enable_vui = pic->enable_vui;
118 enc->enc_pic.vui.video_format = 0x00000005;
119 enc->enc_pic.vui.color_prim = 0x00000002;
120 enc->enc_pic.vui.transfer_char = 0x00000002;
121 enc->enc_pic.vui.matrix_coef = 0x00000002;
122 enc->enc_pic.vui.timing_info_present_flag = 0x00000001;
123 enc->enc_pic.vui.num_units_in_tick = pic->rate_ctrl.frame_rate_den;
124 enc->enc_pic.vui.time_scale = pic->rate_ctrl.frame_rate_num * 2;
125 enc->enc_pic.vui.fixed_frame_rate_flag = 0x00000001;
126 enc->enc_pic.vui.bit_rate_scale = 0x00000004;
127 enc->enc_pic.vui.cpb_size_scale = 0x00000006;
128 enc->enc_pic.vui.initial_cpb_removal_delay_length_minus1 = 0x00000017;
129 enc->enc_pic.vui.cpb_removal_delay_length_minus1 = 0x00000017;
130 enc->enc_pic.vui.dpb_output_delay_length_minus1 = 0x00000017;
131 enc->enc_pic.vui.time_offset_length = 0x00000018;
132 enc->enc_pic.vui.motion_vectors_over_pic_boundaries_flag = 0x00000001;
133 enc->enc_pic.vui.max_bytes_per_pic_denom = 0x00000002;
134 enc->enc_pic.vui.max_bits_per_mb_denom = 0x00000001;
135 enc->enc_pic.vui.log2_max_mv_length_hori = 0x00000010;
136 enc->enc_pic.vui.log2_max_mv_length_vert = 0x00000010;
137 enc->enc_pic.vui.num_reorder_frames = 0x00000003;
138 enc->enc_pic.vui.max_dec_frame_buffering = 0x00000003;
139 }
140
si_vce_52_get_param(struct rvce_encoder * enc,struct pipe_h264_enc_picture_desc * pic)141 void si_vce_52_get_param(struct rvce_encoder *enc, struct pipe_h264_enc_picture_desc *pic)
142 {
143 get_rate_control_param(enc, pic);
144 get_motion_estimation_param(enc, pic);
145 get_pic_control_param(enc, pic);
146 get_task_info_param(enc);
147 get_feedback_buffer_param(enc);
148 get_vui_param(enc, pic);
149 get_config_ext_param(enc);
150
151 enc->enc_pic.picture_type = pic->picture_type;
152 enc->enc_pic.frame_num = pic->frame_num;
153 enc->enc_pic.frame_num_cnt = pic->frame_num_cnt;
154 enc->enc_pic.p_remain = pic->p_remain;
155 enc->enc_pic.i_remain = pic->i_remain;
156 enc->enc_pic.gop_cnt = pic->gop_cnt;
157 enc->enc_pic.pic_order_cnt = pic->pic_order_cnt;
158 enc->enc_pic.ref_idx_l0 = pic->ref_idx_l0;
159 enc->enc_pic.ref_idx_l1 = pic->ref_idx_l1;
160 enc->enc_pic.not_referenced = pic->not_referenced;
161 if (enc->dual_inst)
162 enc->enc_pic.addrmode_arraymode_disrdo_distwoinstants = 0x00000201;
163 else
164 enc->enc_pic.addrmode_arraymode_disrdo_distwoinstants = 0x01000201;
165 enc->enc_pic.is_idr = (pic->picture_type == PIPE_H264_ENC_PICTURE_TYPE_IDR);
166 }
167
create(struct rvce_encoder * enc)168 static void create(struct rvce_encoder *enc)
169 {
170 struct si_screen *sscreen = (struct si_screen *)enc->screen;
171 enc->task_info(enc, 0x00000000, 0, 0, 0);
172
173 RVCE_BEGIN(0x01000001); // create cmd
174 RVCE_CS(enc->enc_pic.ec.enc_use_circular_buffer);
175 RVCE_CS(profiles[enc->base.profile -
176 PIPE_VIDEO_PROFILE_MPEG4_AVC_BASELINE]); // encProfile
177 RVCE_CS(enc->base.level); // encLevel
178 RVCE_CS(enc->enc_pic.ec.enc_pic_struct_restriction);
179 RVCE_CS(enc->base.width); // encImageWidth
180 RVCE_CS(enc->base.height); // encImageHeight
181
182 if (sscreen->info.chip_class < GFX9) {
183 RVCE_CS(enc->luma->u.legacy.level[0].nblk_x * enc->luma->bpe); // encRefPicLumaPitch
184 RVCE_CS(enc->chroma->u.legacy.level[0].nblk_x * enc->chroma->bpe); // encRefPicChromaPitch
185 RVCE_CS(align(enc->luma->u.legacy.level[0].nblk_y, 16) / 8); // encRefYHeightInQw
186 } else {
187 RVCE_CS(enc->luma->u.gfx9.surf_pitch * enc->luma->bpe); // encRefPicLumaPitch
188 RVCE_CS(enc->chroma->u.gfx9.surf_pitch * enc->chroma->bpe); // encRefPicChromaPitch
189 RVCE_CS(align(enc->luma->u.gfx9.surf_height, 16) / 8); // encRefYHeightInQw
190 }
191
192 RVCE_CS(enc->enc_pic.addrmode_arraymode_disrdo_distwoinstants);
193
194 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_context_buffer_offset);
195 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_input_luma_buffer_offset);
196 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_input_chroma_buffer_offset);
197 RVCE_CS(enc->enc_pic.ec.enc_pre_encode_mode_chromaflag_vbaqmode_scenechangesensitivity);
198 RVCE_END();
199 }
200
encode(struct rvce_encoder * enc)201 static void encode(struct rvce_encoder *enc)
202 {
203 struct si_screen *sscreen = (struct si_screen *)enc->screen;
204 signed luma_offset, chroma_offset, bs_offset;
205 unsigned dep, bs_idx = enc->bs_idx++;
206 int i;
207
208 if (enc->dual_inst) {
209 if (bs_idx == 0)
210 dep = 1;
211 else if (enc->enc_pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_IDR)
212 dep = 0;
213 else
214 dep = 2;
215 } else
216 dep = 0;
217
218 enc->task_info(enc, 0x00000003, dep, 0, bs_idx);
219
220 RVCE_BEGIN(0x05000001); // context buffer
221 RVCE_READWRITE(enc->cpb.res->buf, enc->cpb.res->domains, 0); // encodeContextAddressHi/Lo
222 RVCE_END();
223
224 bs_offset = -(signed)(bs_idx * enc->bs_size);
225
226 RVCE_BEGIN(0x05000004); // video bitstream buffer
227 RVCE_WRITE(enc->bs_handle, RADEON_DOMAIN_GTT, bs_offset); // videoBitstreamRingAddressHi/Lo
228 RVCE_CS(enc->bs_size); // videoBitstreamRingSize
229 RVCE_END();
230
231 if (enc->dual_pipe) {
232 unsigned aux_offset = enc->cpb.res->buf->size -
233 RVCE_MAX_AUX_BUFFER_NUM * RVCE_MAX_BITSTREAM_OUTPUT_ROW_SIZE * 2;
234 RVCE_BEGIN(0x05000002); // auxiliary buffer
235 for (i = 0; i < 8; ++i) {
236 RVCE_CS(aux_offset);
237 aux_offset += RVCE_MAX_BITSTREAM_OUTPUT_ROW_SIZE;
238 }
239 for (i = 0; i < 8; ++i)
240 RVCE_CS(RVCE_MAX_BITSTREAM_OUTPUT_ROW_SIZE);
241 RVCE_END();
242 }
243
244 RVCE_BEGIN(0x03000001); // encode
245 RVCE_CS(enc->enc_pic.frame_num ? 0x0 : 0x11); // insertHeaders
246 RVCE_CS(enc->enc_pic.eo.picture_structure);
247 RVCE_CS(enc->bs_size); // allowedMaxBitstreamSize
248 RVCE_CS(enc->enc_pic.eo.force_refresh_map);
249 RVCE_CS(enc->enc_pic.eo.insert_aud);
250 RVCE_CS(enc->enc_pic.eo.end_of_sequence);
251 RVCE_CS(enc->enc_pic.eo.end_of_stream);
252
253 if (sscreen->info.chip_class < GFX9) {
254 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
255 enc->luma->u.legacy.level[0].offset); // inputPictureLumaAddressHi/Lo
256 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
257 enc->chroma->u.legacy.level[0].offset); // inputPictureChromaAddressHi/Lo
258 RVCE_CS(align(enc->luma->u.legacy.level[0].nblk_y, 16)); // encInputFrameYPitch
259 RVCE_CS(enc->luma->u.legacy.level[0].nblk_x * enc->luma->bpe); // encInputPicLumaPitch
260 RVCE_CS(enc->chroma->u.legacy.level[0].nblk_x * enc->chroma->bpe); // encInputPicChromaPitch
261 } else {
262 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
263 enc->luma->u.gfx9.surf_offset); // inputPictureLumaAddressHi/Lo
264 RVCE_READ(enc->handle, RADEON_DOMAIN_VRAM,
265 enc->chroma->u.gfx9.surf_offset); // inputPictureChromaAddressHi/Lo
266 RVCE_CS(align(enc->luma->u.gfx9.surf_height, 16)); // encInputFrameYPitch
267 RVCE_CS(enc->luma->u.gfx9.surf_pitch * enc->luma->bpe); // encInputPicLumaPitch
268 RVCE_CS(enc->chroma->u.gfx9.surf_pitch * enc->chroma->bpe); // encInputPicChromaPitch
269 }
270
271 if (enc->dual_pipe)
272 enc->enc_pic.eo.enc_input_pic_addr_array_disable2pipe_disablemboffload = 0x00000000;
273 else
274 enc->enc_pic.eo.enc_input_pic_addr_array_disable2pipe_disablemboffload = 0x00010000;
275 RVCE_CS(enc->enc_pic.eo.enc_input_pic_addr_array_disable2pipe_disablemboffload);
276 RVCE_CS(enc->enc_pic.eo.enc_input_pic_tile_config);
277 RVCE_CS(enc->enc_pic.picture_type); // encPicType
278 RVCE_CS(enc->enc_pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_IDR); // encIdrFlag
279 if ((enc->enc_pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_IDR) && (enc->enc_pic.eo.enc_idr_pic_id !=0))
280 enc->enc_pic.eo.enc_idr_pic_id = enc->enc_pic.idr_pic_id - 1;
281 else
282 enc->enc_pic.eo.enc_idr_pic_id = 0x00000000;
283 RVCE_CS(enc->enc_pic.eo.enc_idr_pic_id);
284 RVCE_CS(enc->enc_pic.eo.enc_mgs_key_pic);
285 RVCE_CS(!enc->enc_pic.not_referenced);
286 RVCE_CS(enc->enc_pic.eo.enc_temporal_layer_index);
287 RVCE_CS(enc->enc_pic.eo.num_ref_idx_active_override_flag);
288 RVCE_CS(enc->enc_pic.eo.num_ref_idx_l0_active_minus1);
289 RVCE_CS(enc->enc_pic.eo.num_ref_idx_l1_active_minus1);
290
291 i = enc->enc_pic.frame_num - enc->enc_pic.ref_idx_l0;
292 if (i > 1 && enc->enc_pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_P) {
293 enc->enc_pic.eo.enc_ref_list_modification_op = 0x00000001;
294 enc->enc_pic.eo.enc_ref_list_modification_num = i - 1;
295 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_op);
296 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_num);
297 } else {
298 enc->enc_pic.eo.enc_ref_list_modification_op = 0x00000000;
299 enc->enc_pic.eo.enc_ref_list_modification_num = 0x00000000;
300 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_op);
301 RVCE_CS(enc->enc_pic.eo.enc_ref_list_modification_num);
302 }
303
304 for (i = 0; i < 3; ++i) {
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 for (i = 0; i < 4; ++i) {
311 RVCE_CS(enc->enc_pic.eo.enc_decoded_picture_marking_op);
312 RVCE_CS(enc->enc_pic.eo.enc_decoded_picture_marking_num);
313 RVCE_CS(enc->enc_pic.eo.enc_decoded_picture_marking_idx);
314 RVCE_CS(enc->enc_pic.eo.enc_decoded_ref_base_picture_marking_op);
315 RVCE_CS(enc->enc_pic.eo.enc_decoded_ref_base_picture_marking_num);
316 }
317
318 // encReferencePictureL0[0]
319 RVCE_CS(0x00000000); // pictureStructure
320 if(enc->enc_pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_P ||
321 enc->enc_pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_B) {
322 struct rvce_cpb_slot *l0 = si_l0_slot(enc);
323 si_vce_frame_offset(enc, l0, &luma_offset, &chroma_offset);
324 RVCE_CS(l0->picture_type);
325 RVCE_CS(l0->frame_num);
326 RVCE_CS(l0->pic_order_cnt);
327 RVCE_CS(luma_offset);
328 RVCE_CS(chroma_offset);
329 } else {
330 enc->enc_pic.eo.l0_enc_pic_type = 0x00000000;
331 enc->enc_pic.eo.l0_frame_number = 0x00000000;
332 enc->enc_pic.eo.l0_picture_order_count = 0x00000000;
333 enc->enc_pic.eo.l0_luma_offset = 0xffffffff;
334 enc->enc_pic.eo.l0_chroma_offset = 0xffffffff;
335 RVCE_CS(enc->enc_pic.eo.l0_enc_pic_type);
336 RVCE_CS(enc->enc_pic.eo.l0_frame_number);
337 RVCE_CS(enc->enc_pic.eo.l0_picture_order_count);
338 RVCE_CS(enc->enc_pic.eo.l0_luma_offset);
339 RVCE_CS(enc->enc_pic.eo.l0_chroma_offset);
340 }
341
342 // encReferencePictureL0[1]
343 enc->enc_pic.eo.l0_picture_structure = 0x00000000;
344 enc->enc_pic.eo.l0_enc_pic_type = 0x00000000;
345 enc->enc_pic.eo.l0_frame_number = 0x00000000;
346 enc->enc_pic.eo.l0_picture_order_count = 0x00000000;
347 enc->enc_pic.eo.l0_luma_offset = 0xffffffff;
348 enc->enc_pic.eo.l0_chroma_offset = 0xffffffff;
349 RVCE_CS(enc->enc_pic.eo.l0_picture_structure);
350 RVCE_CS(enc->enc_pic.eo.l0_enc_pic_type);
351 RVCE_CS(enc->enc_pic.eo.l0_frame_number);
352 RVCE_CS(enc->enc_pic.eo.l0_picture_order_count);
353 RVCE_CS(enc->enc_pic.eo.l0_luma_offset);
354 RVCE_CS(enc->enc_pic.eo.l0_chroma_offset);
355
356 // encReferencePictureL1[0]
357 RVCE_CS(0x00000000); // pictureStructure
358 if(enc->enc_pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_B) {
359 struct rvce_cpb_slot *l1 = si_l1_slot(enc);
360 si_vce_frame_offset(enc, l1, &luma_offset, &chroma_offset);
361 RVCE_CS(l1->picture_type);
362 RVCE_CS(l1->frame_num);
363 RVCE_CS(l1->pic_order_cnt);
364 RVCE_CS(luma_offset);
365 RVCE_CS(chroma_offset);
366 } else {
367 enc->enc_pic.eo.l1_enc_pic_type = 0x00000000;
368 enc->enc_pic.eo.l1_frame_number = 0x00000000;
369 enc->enc_pic.eo.l1_picture_order_count = 0x00000000;
370 enc->enc_pic.eo.l1_luma_offset = 0xffffffff;
371 enc->enc_pic.eo.l1_chroma_offset = 0xffffffff;
372 RVCE_CS(enc->enc_pic.eo.l1_enc_pic_type);
373 RVCE_CS(enc->enc_pic.eo.l1_frame_number);
374 RVCE_CS(enc->enc_pic.eo.l1_picture_order_count);
375 RVCE_CS(enc->enc_pic.eo.l1_luma_offset);
376 RVCE_CS(enc->enc_pic.eo.l1_chroma_offset);
377 }
378
379 si_vce_frame_offset(enc, si_current_slot(enc), &luma_offset, &chroma_offset);
380 RVCE_CS(luma_offset);
381 RVCE_CS(chroma_offset);
382 RVCE_CS(enc->enc_pic.eo.enc_coloc_buffer_offset);
383 RVCE_CS(enc->enc_pic.eo.enc_reconstructed_ref_base_picture_luma_offset);
384 RVCE_CS(enc->enc_pic.eo.enc_reconstructed_ref_base_picture_chroma_offset);
385 RVCE_CS(enc->enc_pic.eo.enc_reference_ref_base_picture_luma_offset);
386 RVCE_CS(enc->enc_pic.eo.enc_reference_ref_base_picture_chroma_offset);
387 RVCE_CS(enc->enc_pic.frame_num_cnt-1);
388 RVCE_CS(enc->enc_pic.frame_num);
389 RVCE_CS(enc->enc_pic.pic_order_cnt);
390 RVCE_CS(enc->enc_pic.i_remain);
391 RVCE_CS(enc->enc_pic.p_remain);
392 RVCE_CS(enc->enc_pic.eo.num_b_pic_remain_in_rcgop);
393 RVCE_CS(enc->enc_pic.eo.num_ir_pic_remain_in_rcgop);
394 RVCE_CS(enc->enc_pic.eo.enable_intra_refresh);
395
396 RVCE_CS(enc->enc_pic.eo.aq_variance_en);
397 RVCE_CS(enc->enc_pic.eo.aq_block_size);
398 RVCE_CS(enc->enc_pic.eo.aq_mb_variance_sel);
399 RVCE_CS(enc->enc_pic.eo.aq_frame_variance_sel);
400 RVCE_CS(enc->enc_pic.eo.aq_param_a);
401 RVCE_CS(enc->enc_pic.eo.aq_param_b);
402 RVCE_CS(enc->enc_pic.eo.aq_param_c);
403 RVCE_CS(enc->enc_pic.eo.aq_param_d);
404 RVCE_CS(enc->enc_pic.eo.aq_param_e);
405
406 RVCE_CS(enc->enc_pic.eo.context_in_sfb);
407 RVCE_END();
408 }
409
rate_control(struct rvce_encoder * enc)410 static void rate_control(struct rvce_encoder *enc)
411 {
412 RVCE_BEGIN(0x04000005); // rate control
413 RVCE_CS(enc->enc_pic.rc.rc_method);
414 RVCE_CS(enc->enc_pic.rc.target_bitrate);
415 RVCE_CS(enc->enc_pic.rc.peak_bitrate);
416 RVCE_CS(enc->enc_pic.rc.frame_rate_num);
417 RVCE_CS(enc->enc_pic.rc.gop_size);
418 RVCE_CS(enc->enc_pic.rc.quant_i_frames);
419 RVCE_CS(enc->enc_pic.rc.quant_p_frames);
420 RVCE_CS(enc->enc_pic.rc.quant_b_frames);
421 RVCE_CS(enc->enc_pic.rc.vbv_buffer_size);
422 RVCE_CS(enc->enc_pic.rc.frame_rate_den);
423 RVCE_CS(enc->enc_pic.rc.vbv_buf_lv);
424 RVCE_CS(enc->enc_pic.rc.max_au_size);
425 RVCE_CS(enc->enc_pic.rc.qp_initial_mode);
426 RVCE_CS(enc->enc_pic.rc.target_bits_picture);
427 RVCE_CS(enc->enc_pic.rc.peak_bits_picture_integer);
428 RVCE_CS(enc->enc_pic.rc.peak_bits_picture_fraction);
429 RVCE_CS(enc->enc_pic.rc.min_qp);
430 RVCE_CS(enc->enc_pic.rc.max_qp);
431 RVCE_CS(enc->enc_pic.rc.skip_frame_enable);
432 RVCE_CS(enc->enc_pic.rc.fill_data_enable);
433 RVCE_CS(enc->enc_pic.rc.enforce_hrd);
434 RVCE_CS(enc->enc_pic.rc.b_pics_delta_qp);
435 RVCE_CS(enc->enc_pic.rc.ref_b_pics_delta_qp);
436 RVCE_CS(enc->enc_pic.rc.rc_reinit_disable);
437 RVCE_CS(enc->enc_pic.rc.enc_lcvbr_init_qp_flag);
438 RVCE_CS(enc->enc_pic.rc.lcvbrsatd_based_nonlinear_bit_budget_flag);
439 RVCE_END();
440 }
441
config(struct rvce_encoder * enc)442 static void config(struct rvce_encoder *enc)
443 {
444 enc->task_info(enc, 0x00000002, 0, 0xffffffff, 0);
445 enc->rate_control(enc);
446 enc->config_extension(enc);
447 enc->motion_estimation(enc);
448 enc->rdo(enc);
449 if (enc->use_vui)
450 enc->vui(enc);
451 enc->pic_control(enc);
452 }
453
config_extension(struct rvce_encoder * enc)454 static void config_extension(struct rvce_encoder *enc)
455 {
456 RVCE_BEGIN(0x04000001); // config extension
457 RVCE_CS(enc->enc_pic.ce.enc_enable_perf_logging);
458 RVCE_END();
459 }
460
feedback(struct rvce_encoder * enc)461 static void feedback(struct rvce_encoder *enc)
462 {
463 RVCE_BEGIN(0x05000005); // feedback buffer
464 RVCE_WRITE(enc->fb->res->buf, enc->fb->res->domains, 0x0); // feedbackRingAddressHi/Lo
465 RVCE_CS(enc->enc_pic.fb.feedback_ring_size);
466 RVCE_END();
467 }
468
destroy(struct rvce_encoder * enc)469 static void destroy(struct rvce_encoder *enc)
470 {
471 enc->task_info(enc, 0x00000001, 0, 0, 0);
472
473 feedback(enc);
474
475 RVCE_BEGIN(0x02000001); // destroy
476 RVCE_END();
477 }
478
motion_estimation(struct rvce_encoder * enc)479 static void motion_estimation(struct rvce_encoder *enc)
480 {
481 RVCE_BEGIN(0x04000007); // motion estimation
482 RVCE_CS(enc->enc_pic.me.enc_ime_decimation_search);
483 RVCE_CS(enc->enc_pic.me.motion_est_half_pixel);
484 RVCE_CS(enc->enc_pic.me.motion_est_quarter_pixel);
485 RVCE_CS(enc->enc_pic.me.disable_favor_pmv_point);
486 RVCE_CS(enc->enc_pic.me.force_zero_point_center);
487 RVCE_CS(enc->enc_pic.me.lsmvert);
488 RVCE_CS(enc->enc_pic.me.enc_search_range_x);
489 RVCE_CS(enc->enc_pic.me.enc_search_range_y);
490 RVCE_CS(enc->enc_pic.me.enc_search1_range_x);
491 RVCE_CS(enc->enc_pic.me.enc_search1_range_y);
492 RVCE_CS(enc->enc_pic.me.disable_16x16_frame1);
493 RVCE_CS(enc->enc_pic.me.disable_satd);
494 RVCE_CS(enc->enc_pic.me.enable_amd);
495 RVCE_CS(enc->enc_pic.me.enc_disable_sub_mode);
496 RVCE_CS(enc->enc_pic.me.enc_ime_skip_x);
497 RVCE_CS(enc->enc_pic.me.enc_ime_skip_y);
498 RVCE_CS(enc->enc_pic.me.enc_en_ime_overw_dis_subm);
499 RVCE_CS(enc->enc_pic.me.enc_ime_overw_dis_subm_no);
500 RVCE_CS(enc->enc_pic.me.enc_ime2_search_range_x);
501 RVCE_CS(enc->enc_pic.me.enc_ime2_search_range_y);
502 RVCE_CS(enc->enc_pic.me.parallel_mode_speedup_enable);
503 RVCE_CS(enc->enc_pic.me.fme0_enc_disable_sub_mode);
504 RVCE_CS(enc->enc_pic.me.fme1_enc_disable_sub_mode);
505 RVCE_CS(enc->enc_pic.me.ime_sw_speedup_enable);
506 RVCE_END();
507 }
508
pic_control(struct rvce_encoder * enc)509 static void pic_control(struct rvce_encoder *enc)
510 {
511 RVCE_BEGIN(0x04000002); // pic control
512 RVCE_CS(enc->enc_pic.pc.enc_use_constrained_intra_pred);
513 RVCE_CS(enc->enc_pic.pc.enc_cabac_enable);
514 RVCE_CS(enc->enc_pic.pc.enc_cabac_idc);
515 RVCE_CS(enc->enc_pic.pc.enc_loop_filter_disable);
516 RVCE_CS(enc->enc_pic.pc.enc_lf_beta_offset);
517 RVCE_CS(enc->enc_pic.pc.enc_lf_alpha_c0_offset);
518 RVCE_CS(enc->enc_pic.pc.enc_crop_left_offset);
519 RVCE_CS(enc->enc_pic.pc.enc_crop_right_offset);
520 RVCE_CS(enc->enc_pic.pc.enc_crop_top_offset);
521 RVCE_CS(enc->enc_pic.pc.enc_crop_bottom_offset);
522 RVCE_CS(enc->enc_pic.pc.enc_num_mbs_per_slice);
523 RVCE_CS(enc->enc_pic.pc.enc_intra_refresh_num_mbs_per_slot);
524 RVCE_CS(enc->enc_pic.pc.enc_force_intra_refresh);
525 RVCE_CS(enc->enc_pic.pc.enc_force_imb_period);
526 RVCE_CS(enc->enc_pic.pc.enc_pic_order_cnt_type);
527 RVCE_CS(enc->enc_pic.pc.log2_max_pic_order_cnt_lsb_minus4);
528 RVCE_CS(enc->enc_pic.pc.enc_sps_id);
529 RVCE_CS(enc->enc_pic.pc.enc_pps_id);
530 RVCE_CS(enc->enc_pic.pc.enc_constraint_set_flags);
531 RVCE_CS(enc->enc_pic.pc.enc_b_pic_pattern);
532 RVCE_CS(enc->enc_pic.pc.weight_pred_mode_b_picture);
533 RVCE_CS(enc->enc_pic.pc.enc_number_of_reference_frames);
534 RVCE_CS(enc->enc_pic.pc.enc_max_num_ref_frames);
535 RVCE_CS(enc->enc_pic.pc.enc_num_default_active_ref_l0);
536 RVCE_CS(enc->enc_pic.pc.enc_num_default_active_ref_l1);
537 RVCE_CS(enc->enc_pic.pc.enc_slice_mode);
538 RVCE_CS(enc->enc_pic.pc.enc_max_slice_size);
539 RVCE_END();
540 }
541
rdo(struct rvce_encoder * enc)542 static void rdo(struct rvce_encoder *enc)
543 {
544 RVCE_BEGIN(0x04000008); // rdo
545 RVCE_CS(enc->enc_pic.rdo.enc_disable_tbe_pred_i_frame);
546 RVCE_CS(enc->enc_pic.rdo.enc_disable_tbe_pred_p_frame);
547 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_y);
548 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_uv);
549 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_y);
550 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_uv);
551 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_y_1);
552 RVCE_CS(enc->enc_pic.rdo.use_fme_interpol_uv_1);
553 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_y_1);
554 RVCE_CS(enc->enc_pic.rdo.use_fme_intrapol_uv_1);
555 RVCE_CS(enc->enc_pic.rdo.enc_16x16_cost_adj);
556 RVCE_CS(enc->enc_pic.rdo.enc_skip_cost_adj);
557 RVCE_CS(enc->enc_pic.rdo.enc_force_16x16_skip);
558 RVCE_CS(enc->enc_pic.rdo.enc_disable_threshold_calc_a);
559 RVCE_CS(enc->enc_pic.rdo.enc_luma_coeff_cost);
560 RVCE_CS(enc->enc_pic.rdo.enc_luma_mb_coeff_cost);
561 RVCE_CS(enc->enc_pic.rdo.enc_chroma_coeff_cost);
562 RVCE_END();
563 }
564
session(struct rvce_encoder * enc)565 static void session(struct rvce_encoder *enc)
566 {
567 RVCE_BEGIN(0x00000001); // session cmd
568 RVCE_CS(enc->stream_handle);
569 RVCE_END();
570 }
571
task_info(struct rvce_encoder * enc,uint32_t op,uint32_t dep,uint32_t fb_idx,uint32_t ring_idx)572 static void task_info(struct rvce_encoder *enc, uint32_t op,
573 uint32_t dep, uint32_t fb_idx, uint32_t ring_idx)
574 {
575 RVCE_BEGIN(0x00000002); // task info
576 if (op == 0x3) {
577 if (enc->task_info_idx) {
578 uint32_t offs = enc->cs->current.cdw - enc->task_info_idx + 3;
579 // Update offsetOfNextTaskInfo
580 enc->cs->current.buf[enc->task_info_idx] = offs;
581 }
582 enc->task_info_idx = enc->cs->current.cdw;
583 }
584 enc->enc_pic.ti.task_operation = op;
585 enc->enc_pic.ti.reference_picture_dependency = dep;
586 enc->enc_pic.ti.feedback_index = fb_idx;
587 enc->enc_pic.ti.video_bitstream_ring_index = ring_idx;
588 RVCE_CS(enc->enc_pic.ti.offset_of_next_task_info);
589 RVCE_CS(enc->enc_pic.ti.task_operation);
590 RVCE_CS(enc->enc_pic.ti.reference_picture_dependency);
591 RVCE_CS(enc->enc_pic.ti.collocate_flag_dependency);
592 RVCE_CS(enc->enc_pic.ti.feedback_index);
593 RVCE_CS(enc->enc_pic.ti.video_bitstream_ring_index);
594 RVCE_END();
595 }
596
vui(struct rvce_encoder * enc)597 static void vui(struct rvce_encoder *enc)
598 {
599 int i;
600
601 if (!enc->enc_pic.enable_vui)
602 return;
603
604 RVCE_BEGIN(0x04000009); // vui
605 RVCE_CS(enc->enc_pic.vui.aspect_ratio_info_present_flag);
606 RVCE_CS(enc->enc_pic.vui.aspect_ratio_idc);
607 RVCE_CS(enc->enc_pic.vui.sar_width);
608 RVCE_CS(enc->enc_pic.vui.sar_height);
609 RVCE_CS(enc->enc_pic.vui.overscan_info_present_flag);
610 RVCE_CS(enc->enc_pic.vui.overscan_Approp_flag);
611 RVCE_CS(enc->enc_pic.vui.video_signal_type_present_flag);
612 RVCE_CS(enc->enc_pic.vui.video_format);
613 RVCE_CS(enc->enc_pic.vui.video_full_range_flag);
614 RVCE_CS(enc->enc_pic.vui.color_description_present_flag);
615 RVCE_CS(enc->enc_pic.vui.color_prim);
616 RVCE_CS(enc->enc_pic.vui.transfer_char);
617 RVCE_CS(enc->enc_pic.vui.matrix_coef);
618 RVCE_CS(enc->enc_pic.vui.chroma_loc_info_present_flag);
619 RVCE_CS(enc->enc_pic.vui.chroma_loc_top);
620 RVCE_CS(enc->enc_pic.vui.chroma_loc_bottom);
621 RVCE_CS(enc->enc_pic.vui.timing_info_present_flag);
622 RVCE_CS(enc->enc_pic.vui.num_units_in_tick);
623 RVCE_CS(enc->enc_pic.vui.time_scale);
624 RVCE_CS(enc->enc_pic.vui.fixed_frame_rate_flag);
625 RVCE_CS(enc->enc_pic.vui.nal_hrd_parameters_present_flag);
626 RVCE_CS(enc->enc_pic.vui.cpb_cnt_minus1);
627 RVCE_CS(enc->enc_pic.vui.bit_rate_scale);
628 RVCE_CS(enc->enc_pic.vui.cpb_size_scale);
629 for (i = 0; i < 32; i++) {
630 RVCE_CS(enc->enc_pic.vui.bit_rate_value_minus);
631 RVCE_CS(enc->enc_pic.vui.cpb_size_value_minus);
632 RVCE_CS(enc->enc_pic.vui.cbr_flag);
633 }
634 RVCE_CS(enc->enc_pic.vui.initial_cpb_removal_delay_length_minus1);
635 RVCE_CS(enc->enc_pic.vui.cpb_removal_delay_length_minus1);
636 RVCE_CS(enc->enc_pic.vui.dpb_output_delay_length_minus1);
637 RVCE_CS(enc->enc_pic.vui.time_offset_length);
638 RVCE_CS(enc->enc_pic.vui.low_delay_hrd_flag);
639 RVCE_CS(enc->enc_pic.vui.pic_struct_present_flag);
640 RVCE_CS(enc->enc_pic.vui.bitstream_restriction_present_flag);
641 RVCE_CS(enc->enc_pic.vui.motion_vectors_over_pic_boundaries_flag);
642 RVCE_CS(enc->enc_pic.vui.max_bytes_per_pic_denom);
643 RVCE_CS(enc->enc_pic.vui.max_bits_per_mb_denom);
644 RVCE_CS(enc->enc_pic.vui.log2_max_mv_length_hori);
645 RVCE_CS(enc->enc_pic.vui.log2_max_mv_length_vert);
646 RVCE_CS(enc->enc_pic.vui.num_reorder_frames);
647 RVCE_CS(enc->enc_pic.vui.max_dec_frame_buffering);
648 RVCE_END();
649 }
650
si_vce_52_init(struct rvce_encoder * enc)651 void si_vce_52_init(struct rvce_encoder *enc)
652 {
653 enc->session = session;
654 enc->task_info = task_info;
655 enc->create = create;
656 enc->feedback = feedback;
657 enc->rate_control = rate_control;
658 enc->config_extension = config_extension;
659 enc->pic_control = pic_control;
660 enc->motion_estimation = motion_estimation;
661 enc->rdo = rdo;
662 enc->vui = vui;
663 enc->config = config;
664 enc->encode = encode;
665 enc->destroy = destroy;
666 }
667