Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
3769 | Serge | 1 | /* |
2 | * Copyright © 2010 Intel Corporation |
||
3 | * |
||
4 | * Permission is hereby granted, free of charge, to any person obtaining a |
||
5 | * copy of this software and associated documentation files (the |
||
6 | * "Software"), to deal in the Software without restriction, including |
||
7 | * without limitation the rights to use, copy, modify, merge, publish, |
||
8 | * distribute, sub license, and/or sell copies of the Software, and to |
||
9 | * permit persons to whom the Software is furnished to do so, subject to |
||
10 | * the following conditions: |
||
11 | * |
||
12 | * The above copyright notice and this permission notice (including the |
||
13 | * next paragraph) shall be included in all copies or substantial portions |
||
14 | * of the Software. |
||
15 | * |
||
16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS |
||
17 | * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF |
||
18 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. |
||
19 | * IN NO EVENT SHALL PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR |
||
20 | * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, |
||
21 | * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE |
||
22 | * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. |
||
23 | * |
||
24 | * Authors: |
||
25 | * Xiang Haihao |
||
26 | * |
||
27 | */ |
||
28 | |||
29 | #include |
||
30 | #include |
||
31 | #include |
||
32 | #include |
||
33 | |||
34 | #include "intel_batchbuffer.h" |
||
35 | #include "intel_driver.h" |
||
36 | #include "i965_defines.h" |
||
37 | #include "i965_structs.h" |
||
38 | #include "i965_drv_video.h" |
||
39 | #include "gen75_vpp_vebox.h" |
||
40 | #include "i965_post_processing.h" |
||
41 | #include "i965_render.h" |
||
42 | |||
43 | #define HAS_PP(ctx) (IS_IRONLAKE((ctx)->intel.device_id) || \ |
||
44 | IS_GEN6((ctx)->intel.device_id) || \ |
||
45 | IS_GEN7((ctx)->intel.device_id)) |
||
46 | |||
47 | #define SURFACE_STATE_PADDED_SIZE_0_I965 ALIGN(sizeof(struct i965_surface_state), 32) |
||
48 | #define SURFACE_STATE_PADDED_SIZE_1_I965 ALIGN(sizeof(struct i965_surface_state2), 32) |
||
49 | #define SURFACE_STATE_PADDED_SIZE_I965 MAX(SURFACE_STATE_PADDED_SIZE_0_I965, SURFACE_STATE_PADDED_SIZE_1_I965) |
||
50 | |||
51 | #define SURFACE_STATE_PADDED_SIZE_0_GEN7 ALIGN(sizeof(struct gen7_surface_state), 32) |
||
52 | #define SURFACE_STATE_PADDED_SIZE_1_GEN7 ALIGN(sizeof(struct gen7_surface_state2), 32) |
||
53 | #define SURFACE_STATE_PADDED_SIZE_GEN7 MAX(SURFACE_STATE_PADDED_SIZE_0_GEN7, SURFACE_STATE_PADDED_SIZE_1_GEN7) |
||
54 | |||
55 | #define SURFACE_STATE_PADDED_SIZE MAX(SURFACE_STATE_PADDED_SIZE_I965, SURFACE_STATE_PADDED_SIZE_GEN7) |
||
56 | #define SURFACE_STATE_OFFSET(index) (SURFACE_STATE_PADDED_SIZE * index) |
||
57 | #define BINDING_TABLE_OFFSET SURFACE_STATE_OFFSET(MAX_PP_SURFACES) |
||
58 | |||
59 | #define GPU_ASM_BLOCK_WIDTH 16 |
||
60 | #define GPU_ASM_BLOCK_HEIGHT 8 |
||
61 | #define GPU_ASM_X_OFFSET_ALIGNMENT 4 |
||
62 | |||
63 | static const uint32_t pp_null_gen5[][4] = { |
||
64 | #include "shaders/post_processing/gen5_6/null.g4b.gen5" |
||
65 | }; |
||
66 | |||
67 | static const uint32_t pp_nv12_load_save_nv12_gen5[][4] = { |
||
68 | #include "shaders/post_processing/gen5_6/nv12_load_save_nv12.g4b.gen5" |
||
69 | }; |
||
70 | |||
71 | static const uint32_t pp_nv12_load_save_pl3_gen5[][4] = { |
||
72 | #include "shaders/post_processing/gen5_6/nv12_load_save_pl3.g4b.gen5" |
||
73 | }; |
||
74 | |||
75 | static const uint32_t pp_pl3_load_save_nv12_gen5[][4] = { |
||
76 | #include "shaders/post_processing/gen5_6/pl3_load_save_nv12.g4b.gen5" |
||
77 | }; |
||
78 | |||
79 | static const uint32_t pp_pl3_load_save_pl3_gen5[][4] = { |
||
80 | #include "shaders/post_processing/gen5_6/pl3_load_save_pl3.g4b.gen5" |
||
81 | }; |
||
82 | |||
83 | static const uint32_t pp_nv12_scaling_gen5[][4] = { |
||
84 | #include "shaders/post_processing/gen5_6/nv12_scaling_nv12.g4b.gen5" |
||
85 | }; |
||
86 | |||
87 | static const uint32_t pp_nv12_avs_gen5[][4] = { |
||
88 | #include "shaders/post_processing/gen5_6/nv12_avs_nv12.g4b.gen5" |
||
89 | }; |
||
90 | |||
91 | static const uint32_t pp_nv12_dndi_gen5[][4] = { |
||
92 | #include "shaders/post_processing/gen5_6/nv12_dndi_nv12.g4b.gen5" |
||
93 | }; |
||
94 | |||
95 | static const uint32_t pp_nv12_dn_gen5[][4] = { |
||
96 | #include "shaders/post_processing/gen5_6/nv12_dn_nv12.g4b.gen5" |
||
97 | }; |
||
98 | |||
99 | static const uint32_t pp_nv12_load_save_pa_gen5[][4] = { |
||
100 | #include "shaders/post_processing/gen5_6/nv12_load_save_pa.g4b.gen5" |
||
101 | }; |
||
102 | |||
103 | static const uint32_t pp_pl3_load_save_pa_gen5[][4] = { |
||
104 | #include "shaders/post_processing/gen5_6/pl3_load_save_pa.g4b.gen5" |
||
105 | }; |
||
106 | |||
107 | static const uint32_t pp_pa_load_save_nv12_gen5[][4] = { |
||
108 | #include "shaders/post_processing/gen5_6/pa_load_save_nv12.g4b.gen5" |
||
109 | }; |
||
110 | |||
111 | static VAStatus pp_null_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
112 | const struct i965_surface *src_surface, |
||
113 | const VARectangle *src_rect, |
||
114 | struct i965_surface *dst_surface, |
||
115 | const VARectangle *dst_rect, |
||
116 | void *filter_param); |
||
117 | static VAStatus pp_nv12_avs_initialize_nlas(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
118 | const struct i965_surface *src_surface, |
||
119 | const VARectangle *src_rect, |
||
120 | struct i965_surface *dst_surface, |
||
121 | const VARectangle *dst_rect, |
||
122 | void *filter_param); |
||
123 | static VAStatus pp_nv12_scaling_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
124 | const struct i965_surface *src_surface, |
||
125 | const VARectangle *src_rect, |
||
126 | struct i965_surface *dst_surface, |
||
127 | const VARectangle *dst_rect, |
||
128 | void *filter_param); |
||
129 | static VAStatus gen6_nv12_scaling_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
130 | const struct i965_surface *src_surface, |
||
131 | const VARectangle *src_rect, |
||
132 | struct i965_surface *dst_surface, |
||
133 | const VARectangle *dst_rect, |
||
134 | void *filter_param); |
||
135 | static VAStatus pp_plx_load_save_plx_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
136 | const struct i965_surface *src_surface, |
||
137 | const VARectangle *src_rect, |
||
138 | struct i965_surface *dst_surface, |
||
139 | const VARectangle *dst_rect, |
||
140 | void *filter_param); |
||
141 | static VAStatus pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
142 | const struct i965_surface *src_surface, |
||
143 | const VARectangle *src_rect, |
||
144 | struct i965_surface *dst_surface, |
||
145 | const VARectangle *dst_rect, |
||
146 | void *filter_param); |
||
147 | |||
148 | static VAStatus pp_nv12_dn_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
149 | const struct i965_surface *src_surface, |
||
150 | const VARectangle *src_rect, |
||
151 | struct i965_surface *dst_surface, |
||
152 | const VARectangle *dst_rect, |
||
153 | void *filter_param); |
||
154 | |||
155 | static struct pp_module pp_modules_gen5[] = { |
||
156 | { |
||
157 | { |
||
158 | "NULL module (for testing)", |
||
159 | PP_NULL, |
||
160 | pp_null_gen5, |
||
161 | sizeof(pp_null_gen5), |
||
162 | NULL, |
||
163 | }, |
||
164 | |||
165 | pp_null_initialize, |
||
166 | }, |
||
167 | |||
168 | { |
||
169 | { |
||
170 | "NV12_NV12", |
||
171 | PP_NV12_LOAD_SAVE_N12, |
||
172 | pp_nv12_load_save_nv12_gen5, |
||
173 | sizeof(pp_nv12_load_save_nv12_gen5), |
||
174 | NULL, |
||
175 | }, |
||
176 | |||
177 | pp_plx_load_save_plx_initialize, |
||
178 | }, |
||
179 | |||
180 | { |
||
181 | { |
||
182 | "NV12_PL3", |
||
183 | PP_NV12_LOAD_SAVE_PL3, |
||
184 | pp_nv12_load_save_pl3_gen5, |
||
185 | sizeof(pp_nv12_load_save_pl3_gen5), |
||
186 | NULL, |
||
187 | }, |
||
188 | |||
189 | pp_plx_load_save_plx_initialize, |
||
190 | }, |
||
191 | |||
192 | { |
||
193 | { |
||
194 | "PL3_NV12", |
||
195 | PP_PL3_LOAD_SAVE_N12, |
||
196 | pp_pl3_load_save_nv12_gen5, |
||
197 | sizeof(pp_pl3_load_save_nv12_gen5), |
||
198 | NULL, |
||
199 | }, |
||
200 | |||
201 | pp_plx_load_save_plx_initialize, |
||
202 | }, |
||
203 | |||
204 | { |
||
205 | { |
||
206 | "PL3_PL3", |
||
207 | PP_PL3_LOAD_SAVE_PL3, |
||
208 | pp_pl3_load_save_pl3_gen5, |
||
209 | sizeof(pp_pl3_load_save_pl3_gen5), |
||
210 | NULL, |
||
211 | }, |
||
212 | |||
213 | pp_plx_load_save_plx_initialize |
||
214 | }, |
||
215 | |||
216 | { |
||
217 | { |
||
218 | "NV12 Scaling module", |
||
219 | PP_NV12_SCALING, |
||
220 | pp_nv12_scaling_gen5, |
||
221 | sizeof(pp_nv12_scaling_gen5), |
||
222 | NULL, |
||
223 | }, |
||
224 | |||
225 | pp_nv12_scaling_initialize, |
||
226 | }, |
||
227 | |||
228 | { |
||
229 | { |
||
230 | "NV12 AVS module", |
||
231 | PP_NV12_AVS, |
||
232 | pp_nv12_avs_gen5, |
||
233 | sizeof(pp_nv12_avs_gen5), |
||
234 | NULL, |
||
235 | }, |
||
236 | |||
237 | pp_nv12_avs_initialize_nlas, |
||
238 | }, |
||
239 | |||
240 | { |
||
241 | { |
||
242 | "NV12 DNDI module", |
||
243 | PP_NV12_DNDI, |
||
244 | pp_nv12_dndi_gen5, |
||
245 | sizeof(pp_nv12_dndi_gen5), |
||
246 | NULL, |
||
247 | }, |
||
248 | |||
249 | pp_nv12_dndi_initialize, |
||
250 | }, |
||
251 | |||
252 | { |
||
253 | { |
||
254 | "NV12 DN module", |
||
255 | PP_NV12_DN, |
||
256 | pp_nv12_dn_gen5, |
||
257 | sizeof(pp_nv12_dn_gen5), |
||
258 | NULL, |
||
259 | }, |
||
260 | |||
261 | pp_nv12_dn_initialize, |
||
262 | }, |
||
263 | |||
264 | { |
||
265 | { |
||
266 | "NV12_PA module", |
||
267 | PP_NV12_LOAD_SAVE_PA, |
||
268 | pp_nv12_load_save_pa_gen5, |
||
269 | sizeof(pp_nv12_load_save_pa_gen5), |
||
270 | NULL, |
||
271 | }, |
||
272 | |||
273 | pp_plx_load_save_plx_initialize, |
||
274 | }, |
||
275 | { |
||
276 | { |
||
277 | "PL3_PA module", |
||
278 | PP_PL3_LOAD_SAVE_PA, |
||
279 | pp_pl3_load_save_pa_gen5, |
||
280 | sizeof(pp_pl3_load_save_pa_gen5), |
||
281 | NULL, |
||
282 | }, |
||
283 | |||
284 | pp_plx_load_save_plx_initialize, |
||
285 | }, |
||
286 | { |
||
287 | { |
||
288 | "PA_NV12 module", |
||
289 | PP_PA_LOAD_SAVE_NV12, |
||
290 | pp_pa_load_save_nv12_gen5, |
||
291 | sizeof(pp_pa_load_save_nv12_gen5), |
||
292 | NULL, |
||
293 | }, |
||
294 | |||
295 | pp_plx_load_save_plx_initialize, |
||
296 | }, |
||
297 | |||
298 | }; |
||
299 | |||
300 | static const uint32_t pp_null_gen6[][4] = { |
||
301 | #include "shaders/post_processing/gen5_6/null.g6b" |
||
302 | }; |
||
303 | |||
304 | static const uint32_t pp_nv12_load_save_nv12_gen6[][4] = { |
||
305 | #include "shaders/post_processing/gen5_6/nv12_load_save_nv12.g6b" |
||
306 | }; |
||
307 | |||
308 | static const uint32_t pp_nv12_load_save_pl3_gen6[][4] = { |
||
309 | #include "shaders/post_processing/gen5_6/nv12_load_save_pl3.g6b" |
||
310 | }; |
||
311 | |||
312 | static const uint32_t pp_pl3_load_save_nv12_gen6[][4] = { |
||
313 | #include "shaders/post_processing/gen5_6/pl3_load_save_nv12.g6b" |
||
314 | }; |
||
315 | |||
316 | static const uint32_t pp_pl3_load_save_pl3_gen6[][4] = { |
||
317 | #include "shaders/post_processing/gen5_6/pl3_load_save_pl3.g6b" |
||
318 | }; |
||
319 | |||
320 | static const uint32_t pp_nv12_scaling_gen6[][4] = { |
||
321 | #include "shaders/post_processing/gen5_6/nv12_avs_nv12.g6b" |
||
322 | }; |
||
323 | |||
324 | static const uint32_t pp_nv12_avs_gen6[][4] = { |
||
325 | #include "shaders/post_processing/gen5_6/nv12_avs_nv12.g6b" |
||
326 | }; |
||
327 | |||
328 | static const uint32_t pp_nv12_dndi_gen6[][4] = { |
||
329 | #include "shaders/post_processing/gen5_6/nv12_dndi_nv12.g6b" |
||
330 | }; |
||
331 | |||
332 | static const uint32_t pp_nv12_dn_gen6[][4] = { |
||
333 | #include "shaders/post_processing/gen5_6/nv12_dn_nv12.g6b" |
||
334 | }; |
||
335 | |||
336 | static const uint32_t pp_nv12_load_save_pa_gen6[][4] = { |
||
337 | #include "shaders/post_processing/gen5_6/nv12_load_save_pa.g6b" |
||
338 | }; |
||
339 | |||
340 | static const uint32_t pp_pa_load_save_nv12_gen6[][4] = { |
||
341 | #include "shaders/post_processing/gen5_6/pa_load_save_nv12.g6b" |
||
342 | }; |
||
343 | |||
344 | static const uint32_t pp_pl3_load_save_pa_gen6[][4] = { |
||
345 | #include "shaders/post_processing/gen5_6/pl3_load_save_pa.g6b" |
||
346 | }; |
||
347 | |||
348 | static struct pp_module pp_modules_gen6[] = { |
||
349 | { |
||
350 | { |
||
351 | "NULL module (for testing)", |
||
352 | PP_NULL, |
||
353 | pp_null_gen6, |
||
354 | sizeof(pp_null_gen6), |
||
355 | NULL, |
||
356 | }, |
||
357 | |||
358 | pp_null_initialize, |
||
359 | }, |
||
360 | |||
361 | { |
||
362 | { |
||
363 | "NV12_NV12", |
||
364 | PP_NV12_LOAD_SAVE_N12, |
||
365 | pp_nv12_load_save_nv12_gen6, |
||
366 | sizeof(pp_nv12_load_save_nv12_gen6), |
||
367 | NULL, |
||
368 | }, |
||
369 | |||
370 | pp_plx_load_save_plx_initialize, |
||
371 | }, |
||
372 | |||
373 | { |
||
374 | { |
||
375 | "NV12_PL3", |
||
376 | PP_NV12_LOAD_SAVE_PL3, |
||
377 | pp_nv12_load_save_pl3_gen6, |
||
378 | sizeof(pp_nv12_load_save_pl3_gen6), |
||
379 | NULL, |
||
380 | }, |
||
381 | |||
382 | pp_plx_load_save_plx_initialize, |
||
383 | }, |
||
384 | |||
385 | { |
||
386 | { |
||
387 | "PL3_NV12", |
||
388 | PP_PL3_LOAD_SAVE_N12, |
||
389 | pp_pl3_load_save_nv12_gen6, |
||
390 | sizeof(pp_pl3_load_save_nv12_gen6), |
||
391 | NULL, |
||
392 | }, |
||
393 | |||
394 | pp_plx_load_save_plx_initialize, |
||
395 | }, |
||
396 | |||
397 | { |
||
398 | { |
||
399 | "PL3_PL3", |
||
400 | PP_PL3_LOAD_SAVE_N12, |
||
401 | pp_pl3_load_save_pl3_gen6, |
||
402 | sizeof(pp_pl3_load_save_pl3_gen6), |
||
403 | NULL, |
||
404 | }, |
||
405 | |||
406 | pp_plx_load_save_plx_initialize, |
||
407 | }, |
||
408 | |||
409 | { |
||
410 | { |
||
411 | "NV12 Scaling module", |
||
412 | PP_NV12_SCALING, |
||
413 | pp_nv12_scaling_gen6, |
||
414 | sizeof(pp_nv12_scaling_gen6), |
||
415 | NULL, |
||
416 | }, |
||
417 | |||
418 | gen6_nv12_scaling_initialize, |
||
419 | }, |
||
420 | |||
421 | { |
||
422 | { |
||
423 | "NV12 AVS module", |
||
424 | PP_NV12_AVS, |
||
425 | pp_nv12_avs_gen6, |
||
426 | sizeof(pp_nv12_avs_gen6), |
||
427 | NULL, |
||
428 | }, |
||
429 | |||
430 | pp_nv12_avs_initialize_nlas, |
||
431 | }, |
||
432 | |||
433 | { |
||
434 | { |
||
435 | "NV12 DNDI module", |
||
436 | PP_NV12_DNDI, |
||
437 | pp_nv12_dndi_gen6, |
||
438 | sizeof(pp_nv12_dndi_gen6), |
||
439 | NULL, |
||
440 | }, |
||
441 | |||
442 | pp_nv12_dndi_initialize, |
||
443 | }, |
||
444 | |||
445 | { |
||
446 | { |
||
447 | "NV12 DN module", |
||
448 | PP_NV12_DN, |
||
449 | pp_nv12_dn_gen6, |
||
450 | sizeof(pp_nv12_dn_gen6), |
||
451 | NULL, |
||
452 | }, |
||
453 | |||
454 | pp_nv12_dn_initialize, |
||
455 | }, |
||
456 | { |
||
457 | { |
||
458 | "NV12_PA module", |
||
459 | PP_NV12_LOAD_SAVE_PA, |
||
460 | pp_nv12_load_save_pa_gen6, |
||
461 | sizeof(pp_nv12_load_save_pa_gen6), |
||
462 | NULL, |
||
463 | }, |
||
464 | |||
465 | pp_plx_load_save_plx_initialize, |
||
466 | }, |
||
467 | { |
||
468 | { |
||
469 | "PL3_PA module", |
||
470 | PP_PL3_LOAD_SAVE_PA, |
||
471 | pp_pl3_load_save_pa_gen6, |
||
472 | sizeof(pp_pl3_load_save_pa_gen6), |
||
473 | NULL, |
||
474 | }, |
||
475 | |||
476 | pp_plx_load_save_plx_initialize, |
||
477 | }, |
||
478 | { |
||
479 | { |
||
480 | "PA_NV12 module", |
||
481 | PP_PA_LOAD_SAVE_NV12, |
||
482 | pp_pa_load_save_nv12_gen6, |
||
483 | sizeof(pp_pa_load_save_nv12_gen6), |
||
484 | NULL, |
||
485 | }, |
||
486 | |||
487 | pp_plx_load_save_plx_initialize, |
||
488 | }, |
||
489 | |||
490 | }; |
||
491 | |||
492 | static const uint32_t pp_null_gen7[][4] = { |
||
493 | }; |
||
494 | |||
495 | static const uint32_t pp_nv12_load_save_nv12_gen7[][4] = { |
||
496 | #include "shaders/post_processing/gen7/pl2_to_pl2.g7b" |
||
497 | }; |
||
498 | |||
499 | static const uint32_t pp_nv12_load_save_pl3_gen7[][4] = { |
||
500 | #include "shaders/post_processing/gen7/pl2_to_pl3.g7b" |
||
501 | }; |
||
502 | |||
503 | static const uint32_t pp_pl3_load_save_nv12_gen7[][4] = { |
||
504 | #include "shaders/post_processing/gen7/pl3_to_pl2.g7b" |
||
505 | }; |
||
506 | |||
507 | static const uint32_t pp_pl3_load_save_pl3_gen7[][4] = { |
||
508 | #include "shaders/post_processing/gen7/pl3_to_pl3.g7b" |
||
509 | }; |
||
510 | |||
511 | static const uint32_t pp_nv12_scaling_gen7[][4] = { |
||
512 | #include "shaders/post_processing/gen7/avs.g7b" |
||
513 | }; |
||
514 | |||
515 | static const uint32_t pp_nv12_avs_gen7[][4] = { |
||
516 | #include "shaders/post_processing/gen7/avs.g7b" |
||
517 | }; |
||
518 | |||
519 | static const uint32_t pp_nv12_dndi_gen7[][4] = { |
||
520 | #include "shaders/post_processing/gen7/dndi.g7b" |
||
521 | }; |
||
522 | |||
523 | static const uint32_t pp_nv12_dn_gen7[][4] = { |
||
524 | #include "shaders/post_processing/gen7/nv12_dn_nv12.g7b" |
||
525 | }; |
||
526 | |||
527 | static const uint32_t pp_nv12_load_save_pa_gen7[][4] = { |
||
528 | #include "shaders/post_processing/gen7/pl2_to_pa.g7b" |
||
529 | }; |
||
530 | |||
531 | static const uint32_t pp_pl3_load_save_pa_gen7[][4] = { |
||
532 | #include "shaders/post_processing/gen7/pl3_to_pa.g7b" |
||
533 | }; |
||
534 | |||
535 | static const uint32_t pp_pa_load_save_nv12_gen7[][4] = { |
||
536 | #include "shaders/post_processing/gen7/pa_to_pl2.g7b" |
||
537 | }; |
||
538 | |||
539 | static VAStatus gen7_pp_plx_avs_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
540 | const struct i965_surface *src_surface, |
||
541 | const VARectangle *src_rect, |
||
542 | struct i965_surface *dst_surface, |
||
543 | const VARectangle *dst_rect, |
||
544 | void *filter_param); |
||
545 | static VAStatus gen7_pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
546 | const struct i965_surface *src_surface, |
||
547 | const VARectangle *src_rect, |
||
548 | struct i965_surface *dst_surface, |
||
549 | const VARectangle *dst_rect, |
||
550 | void *filter_param); |
||
551 | static VAStatus gen7_pp_nv12_dn_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
552 | const struct i965_surface *src_surface, |
||
553 | const VARectangle *src_rect, |
||
554 | struct i965_surface *dst_surface, |
||
555 | const VARectangle *dst_rect, |
||
556 | void *filter_param); |
||
557 | |||
558 | static struct pp_module pp_modules_gen7[] = { |
||
559 | { |
||
560 | { |
||
561 | "NULL module (for testing)", |
||
562 | PP_NULL, |
||
563 | pp_null_gen7, |
||
564 | sizeof(pp_null_gen7), |
||
565 | NULL, |
||
566 | }, |
||
567 | |||
568 | pp_null_initialize, |
||
569 | }, |
||
570 | |||
571 | { |
||
572 | { |
||
573 | "NV12_NV12", |
||
574 | PP_NV12_LOAD_SAVE_N12, |
||
575 | pp_nv12_load_save_nv12_gen7, |
||
576 | sizeof(pp_nv12_load_save_nv12_gen7), |
||
577 | NULL, |
||
578 | }, |
||
579 | |||
580 | gen7_pp_plx_avs_initialize, |
||
581 | }, |
||
582 | |||
583 | { |
||
584 | { |
||
585 | "NV12_PL3", |
||
586 | PP_NV12_LOAD_SAVE_PL3, |
||
587 | pp_nv12_load_save_pl3_gen7, |
||
588 | sizeof(pp_nv12_load_save_pl3_gen7), |
||
589 | NULL, |
||
590 | }, |
||
591 | |||
592 | gen7_pp_plx_avs_initialize, |
||
593 | }, |
||
594 | |||
595 | { |
||
596 | { |
||
597 | "PL3_NV12", |
||
598 | PP_PL3_LOAD_SAVE_N12, |
||
599 | pp_pl3_load_save_nv12_gen7, |
||
600 | sizeof(pp_pl3_load_save_nv12_gen7), |
||
601 | NULL, |
||
602 | }, |
||
603 | |||
604 | gen7_pp_plx_avs_initialize, |
||
605 | }, |
||
606 | |||
607 | { |
||
608 | { |
||
609 | "PL3_PL3", |
||
610 | PP_PL3_LOAD_SAVE_N12, |
||
611 | pp_pl3_load_save_pl3_gen7, |
||
612 | sizeof(pp_pl3_load_save_pl3_gen7), |
||
613 | NULL, |
||
614 | }, |
||
615 | |||
616 | gen7_pp_plx_avs_initialize, |
||
617 | }, |
||
618 | |||
619 | { |
||
620 | { |
||
621 | "NV12 Scaling module", |
||
622 | PP_NV12_SCALING, |
||
623 | pp_nv12_scaling_gen7, |
||
624 | sizeof(pp_nv12_scaling_gen7), |
||
625 | NULL, |
||
626 | }, |
||
627 | |||
628 | gen7_pp_plx_avs_initialize, |
||
629 | }, |
||
630 | |||
631 | { |
||
632 | { |
||
633 | "NV12 AVS module", |
||
634 | PP_NV12_AVS, |
||
635 | pp_nv12_avs_gen7, |
||
636 | sizeof(pp_nv12_avs_gen7), |
||
637 | NULL, |
||
638 | }, |
||
639 | |||
640 | gen7_pp_plx_avs_initialize, |
||
641 | }, |
||
642 | |||
643 | { |
||
644 | { |
||
645 | "NV12 DNDI module", |
||
646 | PP_NV12_DNDI, |
||
647 | pp_nv12_dndi_gen7, |
||
648 | sizeof(pp_nv12_dndi_gen7), |
||
649 | NULL, |
||
650 | }, |
||
651 | |||
652 | gen7_pp_nv12_dndi_initialize, |
||
653 | }, |
||
654 | |||
655 | { |
||
656 | { |
||
657 | "NV12 DN module", |
||
658 | PP_NV12_DN, |
||
659 | pp_nv12_dn_gen7, |
||
660 | sizeof(pp_nv12_dn_gen7), |
||
661 | NULL, |
||
662 | }, |
||
663 | |||
664 | gen7_pp_nv12_dn_initialize, |
||
665 | }, |
||
666 | { |
||
667 | { |
||
668 | "NV12_PA module", |
||
669 | PP_NV12_LOAD_SAVE_PA, |
||
670 | pp_nv12_load_save_pa_gen7, |
||
671 | sizeof(pp_nv12_load_save_pa_gen7), |
||
672 | NULL, |
||
673 | }, |
||
674 | |||
675 | gen7_pp_plx_avs_initialize, |
||
676 | }, |
||
677 | { |
||
678 | { |
||
679 | "PL3_PA module", |
||
680 | PP_PL3_LOAD_SAVE_PA, |
||
681 | pp_pl3_load_save_pa_gen7, |
||
682 | sizeof(pp_pl3_load_save_pa_gen7), |
||
683 | NULL, |
||
684 | }, |
||
685 | |||
686 | gen7_pp_plx_avs_initialize, |
||
687 | }, |
||
688 | |||
689 | { |
||
690 | { |
||
691 | "PA_NV12 module", |
||
692 | PP_PA_LOAD_SAVE_NV12, |
||
693 | pp_pa_load_save_nv12_gen7, |
||
694 | sizeof(pp_pa_load_save_nv12_gen7), |
||
695 | NULL, |
||
696 | }, |
||
697 | |||
698 | gen7_pp_plx_avs_initialize, |
||
699 | }, |
||
700 | }; |
||
701 | |||
702 | static const uint32_t pp_null_gen75[][4] = { |
||
703 | }; |
||
704 | |||
705 | static const uint32_t pp_nv12_load_save_nv12_gen75[][4] = { |
||
706 | #include "shaders/post_processing/gen7/pl2_to_pl2.g75b" |
||
707 | }; |
||
708 | |||
709 | static const uint32_t pp_nv12_load_save_pl3_gen75[][4] = { |
||
710 | #include "shaders/post_processing/gen7/pl2_to_pl3.g75b" |
||
711 | }; |
||
712 | |||
713 | static const uint32_t pp_pl3_load_save_nv12_gen75[][4] = { |
||
714 | #include "shaders/post_processing/gen7/pl3_to_pl2.g75b" |
||
715 | }; |
||
716 | |||
717 | static const uint32_t pp_pl3_load_save_pl3_gen75[][4] = { |
||
718 | #include "shaders/post_processing/gen7/pl3_to_pl3.g75b" |
||
719 | }; |
||
720 | |||
721 | static const uint32_t pp_nv12_scaling_gen75[][4] = { |
||
722 | #include "shaders/post_processing/gen7/avs.g75b" |
||
723 | }; |
||
724 | |||
725 | static const uint32_t pp_nv12_avs_gen75[][4] = { |
||
726 | #include "shaders/post_processing/gen7/avs.g75b" |
||
727 | }; |
||
728 | |||
729 | static const uint32_t pp_nv12_dndi_gen75[][4] = { |
||
730 | // #include "shaders/post_processing/gen7/dndi.g75b" |
||
731 | }; |
||
732 | |||
733 | static const uint32_t pp_nv12_dn_gen75[][4] = { |
||
734 | // #include "shaders/post_processing/gen7/nv12_dn_nv12.g75b" |
||
735 | }; |
||
736 | static const uint32_t pp_nv12_load_save_pa_gen75[][4] = { |
||
737 | #include "shaders/post_processing/gen7/pl2_to_pa.g75b" |
||
738 | }; |
||
739 | |||
740 | static const uint32_t pp_pl3_load_save_pa_gen75[][4] = { |
||
741 | #include "shaders/post_processing/gen7/pl3_to_pa.g75b" |
||
742 | }; |
||
743 | |||
744 | static const uint32_t pp_pa_load_save_nv12_gen75[][4] = { |
||
745 | #include "shaders/post_processing/gen7/pa_to_pl2.g75b" |
||
746 | }; |
||
747 | |||
748 | static struct pp_module pp_modules_gen75[] = { |
||
749 | { |
||
750 | { |
||
751 | "NULL module (for testing)", |
||
752 | PP_NULL, |
||
753 | pp_null_gen75, |
||
754 | sizeof(pp_null_gen75), |
||
755 | NULL, |
||
756 | }, |
||
757 | |||
758 | pp_null_initialize, |
||
759 | }, |
||
760 | |||
761 | { |
||
762 | { |
||
763 | "NV12_NV12", |
||
764 | PP_NV12_LOAD_SAVE_N12, |
||
765 | pp_nv12_load_save_nv12_gen75, |
||
766 | sizeof(pp_nv12_load_save_nv12_gen75), |
||
767 | NULL, |
||
768 | }, |
||
769 | |||
770 | gen7_pp_plx_avs_initialize, |
||
771 | }, |
||
772 | |||
773 | { |
||
774 | { |
||
775 | "NV12_PL3", |
||
776 | PP_NV12_LOAD_SAVE_PL3, |
||
777 | pp_nv12_load_save_pl3_gen75, |
||
778 | sizeof(pp_nv12_load_save_pl3_gen75), |
||
779 | NULL, |
||
780 | }, |
||
781 | |||
782 | gen7_pp_plx_avs_initialize, |
||
783 | }, |
||
784 | |||
785 | { |
||
786 | { |
||
787 | "PL3_NV12", |
||
788 | PP_PL3_LOAD_SAVE_N12, |
||
789 | pp_pl3_load_save_nv12_gen75, |
||
790 | sizeof(pp_pl3_load_save_nv12_gen75), |
||
791 | NULL, |
||
792 | }, |
||
793 | |||
794 | gen7_pp_plx_avs_initialize, |
||
795 | }, |
||
796 | |||
797 | { |
||
798 | { |
||
799 | "PL3_PL3", |
||
800 | PP_PL3_LOAD_SAVE_N12, |
||
801 | pp_pl3_load_save_pl3_gen75, |
||
802 | sizeof(pp_pl3_load_save_pl3_gen75), |
||
803 | NULL, |
||
804 | }, |
||
805 | |||
806 | gen7_pp_plx_avs_initialize, |
||
807 | }, |
||
808 | |||
809 | { |
||
810 | { |
||
811 | "NV12 Scaling module", |
||
812 | PP_NV12_SCALING, |
||
813 | pp_nv12_scaling_gen75, |
||
814 | sizeof(pp_nv12_scaling_gen75), |
||
815 | NULL, |
||
816 | }, |
||
817 | |||
818 | gen7_pp_plx_avs_initialize, |
||
819 | }, |
||
820 | |||
821 | { |
||
822 | { |
||
823 | "NV12 AVS module", |
||
824 | PP_NV12_AVS, |
||
825 | pp_nv12_avs_gen75, |
||
826 | sizeof(pp_nv12_avs_gen75), |
||
827 | NULL, |
||
828 | }, |
||
829 | |||
830 | gen7_pp_plx_avs_initialize, |
||
831 | }, |
||
832 | |||
833 | { |
||
834 | { |
||
835 | "NV12 DNDI module", |
||
836 | PP_NV12_DNDI, |
||
837 | pp_nv12_dndi_gen75, |
||
838 | sizeof(pp_nv12_dndi_gen75), |
||
839 | NULL, |
||
840 | }, |
||
841 | |||
842 | gen7_pp_nv12_dndi_initialize, |
||
843 | }, |
||
844 | |||
845 | { |
||
846 | { |
||
847 | "NV12 DN module", |
||
848 | PP_NV12_DN, |
||
849 | pp_nv12_dn_gen75, |
||
850 | sizeof(pp_nv12_dn_gen75), |
||
851 | NULL, |
||
852 | }, |
||
853 | |||
854 | gen7_pp_nv12_dn_initialize, |
||
855 | }, |
||
856 | { |
||
857 | { |
||
858 | "NV12_PA module", |
||
859 | PP_NV12_LOAD_SAVE_PA, |
||
860 | pp_nv12_load_save_pa_gen75, |
||
861 | sizeof(pp_nv12_load_save_pa_gen75), |
||
862 | NULL, |
||
863 | }, |
||
864 | |||
865 | gen7_pp_plx_avs_initialize, |
||
866 | }, |
||
867 | |||
868 | { |
||
869 | { |
||
870 | "PL3_PA module", |
||
871 | PP_PL3_LOAD_SAVE_PA, |
||
872 | pp_pl3_load_save_pa_gen75, |
||
873 | sizeof(pp_pl3_load_save_pa_gen75), |
||
874 | NULL, |
||
875 | }, |
||
876 | |||
877 | gen7_pp_plx_avs_initialize, |
||
878 | }, |
||
879 | |||
880 | { |
||
881 | { |
||
882 | "PA_NV12 module", |
||
883 | PP_PA_LOAD_SAVE_NV12, |
||
884 | pp_pa_load_save_nv12_gen75, |
||
885 | sizeof(pp_pa_load_save_nv12_gen75), |
||
886 | NULL, |
||
887 | }, |
||
888 | |||
889 | gen7_pp_plx_avs_initialize, |
||
890 | }, |
||
891 | |||
892 | }; |
||
893 | |||
894 | static int |
||
895 | pp_get_surface_fourcc(VADriverContextP ctx, const struct i965_surface *surface) |
||
896 | { |
||
897 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
898 | int fourcc; |
||
899 | |||
900 | if (surface->type == I965_SURFACE_TYPE_IMAGE) { |
||
901 | struct object_image *obj_image = IMAGE(surface->id); |
||
902 | fourcc = obj_image->image.format.fourcc; |
||
903 | } else { |
||
904 | struct object_surface *obj_surface = SURFACE(surface->id); |
||
905 | fourcc = obj_surface->fourcc; |
||
906 | } |
||
907 | |||
908 | return fourcc; |
||
909 | } |
||
910 | |||
911 | static void |
||
912 | pp_set_surface_tiling(struct i965_surface_state *ss, unsigned int tiling) |
||
913 | { |
||
914 | switch (tiling) { |
||
915 | case I915_TILING_NONE: |
||
916 | ss->ss3.tiled_surface = 0; |
||
917 | ss->ss3.tile_walk = 0; |
||
918 | break; |
||
919 | case I915_TILING_X: |
||
920 | ss->ss3.tiled_surface = 1; |
||
921 | ss->ss3.tile_walk = I965_TILEWALK_XMAJOR; |
||
922 | break; |
||
923 | case I915_TILING_Y: |
||
924 | ss->ss3.tiled_surface = 1; |
||
925 | ss->ss3.tile_walk = I965_TILEWALK_YMAJOR; |
||
926 | break; |
||
927 | } |
||
928 | } |
||
929 | |||
930 | static void |
||
931 | pp_set_surface2_tiling(struct i965_surface_state2 *ss, unsigned int tiling) |
||
932 | { |
||
933 | switch (tiling) { |
||
934 | case I915_TILING_NONE: |
||
935 | ss->ss2.tiled_surface = 0; |
||
936 | ss->ss2.tile_walk = 0; |
||
937 | break; |
||
938 | case I915_TILING_X: |
||
939 | ss->ss2.tiled_surface = 1; |
||
940 | ss->ss2.tile_walk = I965_TILEWALK_XMAJOR; |
||
941 | break; |
||
942 | case I915_TILING_Y: |
||
943 | ss->ss2.tiled_surface = 1; |
||
944 | ss->ss2.tile_walk = I965_TILEWALK_YMAJOR; |
||
945 | break; |
||
946 | } |
||
947 | } |
||
948 | |||
949 | static void |
||
950 | gen7_pp_set_surface_tiling(struct gen7_surface_state *ss, unsigned int tiling) |
||
951 | { |
||
952 | switch (tiling) { |
||
953 | case I915_TILING_NONE: |
||
954 | ss->ss0.tiled_surface = 0; |
||
955 | ss->ss0.tile_walk = 0; |
||
956 | break; |
||
957 | case I915_TILING_X: |
||
958 | ss->ss0.tiled_surface = 1; |
||
959 | ss->ss0.tile_walk = I965_TILEWALK_XMAJOR; |
||
960 | break; |
||
961 | case I915_TILING_Y: |
||
962 | ss->ss0.tiled_surface = 1; |
||
963 | ss->ss0.tile_walk = I965_TILEWALK_YMAJOR; |
||
964 | break; |
||
965 | } |
||
966 | } |
||
967 | |||
968 | static void |
||
969 | gen7_pp_set_surface2_tiling(struct gen7_surface_state2 *ss, unsigned int tiling) |
||
970 | { |
||
971 | switch (tiling) { |
||
972 | case I915_TILING_NONE: |
||
973 | ss->ss2.tiled_surface = 0; |
||
974 | ss->ss2.tile_walk = 0; |
||
975 | break; |
||
976 | case I915_TILING_X: |
||
977 | ss->ss2.tiled_surface = 1; |
||
978 | ss->ss2.tile_walk = I965_TILEWALK_XMAJOR; |
||
979 | break; |
||
980 | case I915_TILING_Y: |
||
981 | ss->ss2.tiled_surface = 1; |
||
982 | ss->ss2.tile_walk = I965_TILEWALK_YMAJOR; |
||
983 | break; |
||
984 | } |
||
985 | } |
||
986 | |||
987 | |||
988 | static void |
||
989 | ironlake_pp_interface_descriptor_table(struct i965_post_processing_context *pp_context) |
||
990 | { |
||
991 | struct i965_interface_descriptor *desc; |
||
992 | dri_bo *bo; |
||
993 | int pp_index = pp_context->current_pp; |
||
994 | |||
995 | bo = pp_context->idrt.bo; |
||
996 | dri_bo_map(bo, 1); |
||
997 | assert(bo->virtual); |
||
998 | desc = bo->virtual; |
||
999 | memset(desc, 0, sizeof(*desc)); |
||
1000 | desc->desc0.grf_reg_blocks = 10; |
||
1001 | desc->desc0.kernel_start_pointer = pp_context->pp_modules[pp_index].kernel.bo->offset >> 6; /* reloc */ |
||
1002 | desc->desc1.const_urb_entry_read_offset = 0; |
||
1003 | desc->desc1.const_urb_entry_read_len = 4; /* grf 1-4 */ |
||
1004 | desc->desc2.sampler_state_pointer = pp_context->sampler_state_table.bo->offset >> 5; |
||
1005 | desc->desc2.sampler_count = 0; |
||
1006 | desc->desc3.binding_table_entry_count = 0; |
||
1007 | desc->desc3.binding_table_pointer = (BINDING_TABLE_OFFSET >> 5); |
||
1008 | |||
1009 | dri_bo_emit_reloc(bo, |
||
1010 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
1011 | desc->desc0.grf_reg_blocks, |
||
1012 | offsetof(struct i965_interface_descriptor, desc0), |
||
1013 | pp_context->pp_modules[pp_index].kernel.bo); |
||
1014 | |||
1015 | dri_bo_emit_reloc(bo, |
||
1016 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
1017 | desc->desc2.sampler_count << 2, |
||
1018 | offsetof(struct i965_interface_descriptor, desc2), |
||
1019 | pp_context->sampler_state_table.bo); |
||
1020 | |||
1021 | dri_bo_unmap(bo); |
||
1022 | pp_context->idrt.num_interface_descriptors++; |
||
1023 | } |
||
1024 | |||
1025 | static void |
||
1026 | ironlake_pp_vfe_state(struct i965_post_processing_context *pp_context) |
||
1027 | { |
||
1028 | struct i965_vfe_state *vfe_state; |
||
1029 | dri_bo *bo; |
||
1030 | |||
1031 | bo = pp_context->vfe_state.bo; |
||
1032 | dri_bo_map(bo, 1); |
||
1033 | assert(bo->virtual); |
||
1034 | vfe_state = bo->virtual; |
||
1035 | memset(vfe_state, 0, sizeof(*vfe_state)); |
||
1036 | vfe_state->vfe1.max_threads = pp_context->urb.num_vfe_entries - 1; |
||
1037 | vfe_state->vfe1.urb_entry_alloc_size = pp_context->urb.size_vfe_entry - 1; |
||
1038 | vfe_state->vfe1.num_urb_entries = pp_context->urb.num_vfe_entries; |
||
1039 | vfe_state->vfe1.vfe_mode = VFE_GENERIC_MODE; |
||
1040 | vfe_state->vfe1.children_present = 0; |
||
1041 | vfe_state->vfe2.interface_descriptor_base = |
||
1042 | pp_context->idrt.bo->offset >> 4; /* reloc */ |
||
1043 | dri_bo_emit_reloc(bo, |
||
1044 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
1045 | 0, |
||
1046 | offsetof(struct i965_vfe_state, vfe2), |
||
1047 | pp_context->idrt.bo); |
||
1048 | dri_bo_unmap(bo); |
||
1049 | } |
||
1050 | |||
1051 | static void |
||
1052 | ironlake_pp_upload_constants(struct i965_post_processing_context *pp_context) |
||
1053 | { |
||
1054 | unsigned char *constant_buffer; |
||
1055 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
1056 | |||
1057 | assert(sizeof(*pp_static_parameter) == 128); |
||
1058 | dri_bo_map(pp_context->curbe.bo, 1); |
||
1059 | assert(pp_context->curbe.bo->virtual); |
||
1060 | constant_buffer = pp_context->curbe.bo->virtual; |
||
1061 | memcpy(constant_buffer, pp_static_parameter, sizeof(*pp_static_parameter)); |
||
1062 | dri_bo_unmap(pp_context->curbe.bo); |
||
1063 | } |
||
1064 | |||
1065 | static void |
||
1066 | ironlake_pp_states_setup(VADriverContextP ctx, |
||
1067 | struct i965_post_processing_context *pp_context) |
||
1068 | { |
||
1069 | ironlake_pp_interface_descriptor_table(pp_context); |
||
1070 | ironlake_pp_vfe_state(pp_context); |
||
1071 | ironlake_pp_upload_constants(pp_context); |
||
1072 | } |
||
1073 | |||
1074 | static void |
||
1075 | ironlake_pp_pipeline_select(VADriverContextP ctx, |
||
1076 | struct i965_post_processing_context *pp_context) |
||
1077 | { |
||
1078 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1079 | |||
1080 | BEGIN_BATCH(batch, 1); |
||
1081 | OUT_BATCH(batch, CMD_PIPELINE_SELECT | PIPELINE_SELECT_MEDIA); |
||
1082 | ADVANCE_BATCH(batch); |
||
1083 | } |
||
1084 | |||
1085 | static void |
||
1086 | ironlake_pp_urb_layout(VADriverContextP ctx, |
||
1087 | struct i965_post_processing_context *pp_context) |
||
1088 | { |
||
1089 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1090 | unsigned int vfe_fence, cs_fence; |
||
1091 | |||
1092 | vfe_fence = pp_context->urb.cs_start; |
||
1093 | cs_fence = pp_context->urb.size; |
||
1094 | |||
1095 | BEGIN_BATCH(batch, 3); |
||
1096 | OUT_BATCH(batch, CMD_URB_FENCE | UF0_VFE_REALLOC | UF0_CS_REALLOC | 1); |
||
1097 | OUT_BATCH(batch, 0); |
||
1098 | OUT_BATCH(batch, |
||
1099 | (vfe_fence << UF2_VFE_FENCE_SHIFT) | /* VFE_SIZE */ |
||
1100 | (cs_fence << UF2_CS_FENCE_SHIFT)); /* CS_SIZE */ |
||
1101 | ADVANCE_BATCH(batch); |
||
1102 | } |
||
1103 | |||
1104 | static void |
||
1105 | ironlake_pp_state_base_address(VADriverContextP ctx, |
||
1106 | struct i965_post_processing_context *pp_context) |
||
1107 | { |
||
1108 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1109 | |||
1110 | BEGIN_BATCH(batch, 8); |
||
1111 | OUT_BATCH(batch, CMD_STATE_BASE_ADDRESS | 6); |
||
1112 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
1113 | OUT_RELOC(batch, pp_context->surface_state_binding_table.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, BASE_ADDRESS_MODIFY); |
||
1114 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
1115 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
1116 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
1117 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
1118 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
1119 | ADVANCE_BATCH(batch); |
||
1120 | } |
||
1121 | |||
1122 | static void |
||
1123 | ironlake_pp_state_pointers(VADriverContextP ctx, |
||
1124 | struct i965_post_processing_context *pp_context) |
||
1125 | { |
||
1126 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1127 | |||
1128 | BEGIN_BATCH(batch, 3); |
||
1129 | OUT_BATCH(batch, CMD_MEDIA_STATE_POINTERS | 1); |
||
1130 | OUT_BATCH(batch, 0); |
||
1131 | OUT_RELOC(batch, pp_context->vfe_state.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0); |
||
1132 | ADVANCE_BATCH(batch); |
||
1133 | } |
||
1134 | |||
1135 | static void |
||
1136 | ironlake_pp_cs_urb_layout(VADriverContextP ctx, |
||
1137 | struct i965_post_processing_context *pp_context) |
||
1138 | { |
||
1139 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1140 | |||
1141 | BEGIN_BATCH(batch, 2); |
||
1142 | OUT_BATCH(batch, CMD_CS_URB_STATE | 0); |
||
1143 | OUT_BATCH(batch, |
||
1144 | ((pp_context->urb.size_cs_entry - 1) << 4) | /* URB Entry Allocation Size */ |
||
1145 | (pp_context->urb.num_cs_entries << 0)); /* Number of URB Entries */ |
||
1146 | ADVANCE_BATCH(batch); |
||
1147 | } |
||
1148 | |||
1149 | static void |
||
1150 | ironlake_pp_constant_buffer(VADriverContextP ctx, |
||
1151 | struct i965_post_processing_context *pp_context) |
||
1152 | { |
||
1153 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1154 | |||
1155 | BEGIN_BATCH(batch, 2); |
||
1156 | OUT_BATCH(batch, CMD_CONSTANT_BUFFER | (1 << 8) | (2 - 2)); |
||
1157 | OUT_RELOC(batch, pp_context->curbe.bo, |
||
1158 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
1159 | pp_context->urb.size_cs_entry - 1); |
||
1160 | ADVANCE_BATCH(batch); |
||
1161 | } |
||
1162 | |||
1163 | static void |
||
1164 | ironlake_pp_object_walker(VADriverContextP ctx, |
||
1165 | struct i965_post_processing_context *pp_context) |
||
1166 | { |
||
1167 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1168 | int x, x_steps, y, y_steps; |
||
1169 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
1170 | |||
1171 | x_steps = pp_context->pp_x_steps(&pp_context->private_context); |
||
1172 | y_steps = pp_context->pp_y_steps(&pp_context->private_context); |
||
1173 | |||
1174 | for (y = 0; y < y_steps; y++) { |
||
1175 | for (x = 0; x < x_steps; x++) { |
||
1176 | if (!pp_context->pp_set_block_parameter(pp_context, x, y)) { |
||
1177 | BEGIN_BATCH(batch, 20); |
||
1178 | OUT_BATCH(batch, CMD_MEDIA_OBJECT | 18); |
||
1179 | OUT_BATCH(batch, 0); |
||
1180 | OUT_BATCH(batch, 0); /* no indirect data */ |
||
1181 | OUT_BATCH(batch, 0); |
||
1182 | |||
1183 | /* inline data grf 5-6 */ |
||
1184 | assert(sizeof(*pp_inline_parameter) == 64); |
||
1185 | intel_batchbuffer_data(batch, pp_inline_parameter, sizeof(*pp_inline_parameter)); |
||
1186 | |||
1187 | ADVANCE_BATCH(batch); |
||
1188 | } |
||
1189 | } |
||
1190 | } |
||
1191 | } |
||
1192 | |||
1193 | static void |
||
1194 | ironlake_pp_pipeline_setup(VADriverContextP ctx, |
||
1195 | struct i965_post_processing_context *pp_context) |
||
1196 | { |
||
1197 | struct intel_batchbuffer *batch = pp_context->batch; |
||
1198 | |||
1199 | intel_batchbuffer_start_atomic(batch, 0x1000); |
||
1200 | intel_batchbuffer_emit_mi_flush(batch); |
||
1201 | ironlake_pp_pipeline_select(ctx, pp_context); |
||
1202 | ironlake_pp_state_base_address(ctx, pp_context); |
||
1203 | ironlake_pp_state_pointers(ctx, pp_context); |
||
1204 | ironlake_pp_urb_layout(ctx, pp_context); |
||
1205 | ironlake_pp_cs_urb_layout(ctx, pp_context); |
||
1206 | ironlake_pp_constant_buffer(ctx, pp_context); |
||
1207 | ironlake_pp_object_walker(ctx, pp_context); |
||
1208 | intel_batchbuffer_end_atomic(batch); |
||
1209 | } |
||
1210 | |||
1211 | // update u/v offset when the surface format are packed yuv |
||
1212 | static void i965_update_src_surface_uv_offset( |
||
1213 | VADriverContextP ctx, |
||
1214 | struct i965_post_processing_context *pp_context, |
||
1215 | const struct i965_surface *surface) |
||
1216 | { |
||
1217 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
1218 | int fourcc = pp_get_surface_fourcc(ctx, surface); |
||
1219 | |||
1220 | if (fourcc == VA_FOURCC('Y', 'U', 'Y', '2')) { |
||
1221 | pp_static_parameter->grf1.source_packed_u_offset = 1; |
||
1222 | pp_static_parameter->grf1.source_packed_v_offset = 3; |
||
1223 | } |
||
1224 | else if (fourcc == VA_FOURCC('U', 'Y', 'V', 'Y')) { |
||
1225 | pp_static_parameter->grf1.source_packed_y_offset = 1; |
||
1226 | pp_static_parameter->grf1.source_packed_v_offset = 2; |
||
1227 | } |
||
1228 | |||
1229 | } |
||
1230 | |||
1231 | static void i965_update_dst_surface_uv_offset( |
||
1232 | VADriverContextP ctx, |
||
1233 | struct i965_post_processing_context *pp_context, |
||
1234 | const struct i965_surface *surface) |
||
1235 | { |
||
1236 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
1237 | int fourcc = pp_get_surface_fourcc(ctx, surface); |
||
1238 | |||
1239 | if (fourcc == VA_FOURCC('Y', 'U', 'Y', '2')) { |
||
1240 | pp_static_parameter->grf1.r1_2.load_and_save.destination_packed_u_offset = 1; |
||
1241 | pp_static_parameter->grf1.r1_2.load_and_save.destination_packed_v_offset = 3; |
||
1242 | } |
||
1243 | else if (fourcc == VA_FOURCC('U', 'Y', 'V', 'Y')) { |
||
1244 | pp_static_parameter->grf1.r1_2.load_and_save.destination_packed_y_offset = 1; |
||
1245 | pp_static_parameter->grf1.r1_2.load_and_save.destination_packed_v_offset = 2; |
||
1246 | } |
||
1247 | |||
1248 | } |
||
1249 | |||
1250 | static void |
||
1251 | i965_pp_set_surface_state(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1252 | dri_bo *surf_bo, unsigned long surf_bo_offset, |
||
1253 | int width, int height, int pitch, int format, |
||
1254 | int index, int is_target) |
||
1255 | { |
||
1256 | struct i965_surface_state *ss; |
||
1257 | dri_bo *ss_bo; |
||
1258 | unsigned int tiling; |
||
1259 | unsigned int swizzle; |
||
1260 | |||
1261 | dri_bo_get_tiling(surf_bo, &tiling, &swizzle); |
||
1262 | ss_bo = pp_context->surface_state_binding_table.bo; |
||
1263 | assert(ss_bo); |
||
1264 | |||
1265 | dri_bo_map(ss_bo, True); |
||
1266 | assert(ss_bo->virtual); |
||
1267 | ss = (struct i965_surface_state *)((char *)ss_bo->virtual + SURFACE_STATE_OFFSET(index)); |
||
1268 | memset(ss, 0, sizeof(*ss)); |
||
1269 | ss->ss0.surface_type = I965_SURFACE_2D; |
||
1270 | ss->ss0.surface_format = format; |
||
1271 | ss->ss1.base_addr = surf_bo->offset + surf_bo_offset; |
||
1272 | ss->ss2.width = width - 1; |
||
1273 | ss->ss2.height = height - 1; |
||
1274 | ss->ss3.pitch = pitch - 1; |
||
1275 | pp_set_surface_tiling(ss, tiling); |
||
1276 | dri_bo_emit_reloc(ss_bo, |
||
1277 | I915_GEM_DOMAIN_RENDER, is_target ? I915_GEM_DOMAIN_RENDER : 0, |
||
1278 | surf_bo_offset, |
||
1279 | SURFACE_STATE_OFFSET(index) + offsetof(struct i965_surface_state, ss1), |
||
1280 | surf_bo); |
||
1281 | ((unsigned int *)((char *)ss_bo->virtual + BINDING_TABLE_OFFSET))[index] = SURFACE_STATE_OFFSET(index); |
||
1282 | dri_bo_unmap(ss_bo); |
||
1283 | } |
||
1284 | |||
1285 | static void |
||
1286 | i965_pp_set_surface2_state(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1287 | dri_bo *surf_bo, unsigned long surf_bo_offset, |
||
1288 | int width, int height, int wpitch, |
||
1289 | int xoffset, int yoffset, |
||
1290 | int format, int interleave_chroma, |
||
1291 | int index) |
||
1292 | { |
||
1293 | struct i965_surface_state2 *ss2; |
||
1294 | dri_bo *ss2_bo; |
||
1295 | unsigned int tiling; |
||
1296 | unsigned int swizzle; |
||
1297 | |||
1298 | dri_bo_get_tiling(surf_bo, &tiling, &swizzle); |
||
1299 | ss2_bo = pp_context->surface_state_binding_table.bo; |
||
1300 | assert(ss2_bo); |
||
1301 | |||
1302 | dri_bo_map(ss2_bo, True); |
||
1303 | assert(ss2_bo->virtual); |
||
1304 | ss2 = (struct i965_surface_state2 *)((char *)ss2_bo->virtual + SURFACE_STATE_OFFSET(index)); |
||
1305 | memset(ss2, 0, sizeof(*ss2)); |
||
1306 | ss2->ss0.surface_base_address = surf_bo->offset + surf_bo_offset; |
||
1307 | ss2->ss1.cbcr_pixel_offset_v_direction = 0; |
||
1308 | ss2->ss1.width = width - 1; |
||
1309 | ss2->ss1.height = height - 1; |
||
1310 | ss2->ss2.pitch = wpitch - 1; |
||
1311 | ss2->ss2.interleave_chroma = interleave_chroma; |
||
1312 | ss2->ss2.surface_format = format; |
||
1313 | ss2->ss3.x_offset_for_cb = xoffset; |
||
1314 | ss2->ss3.y_offset_for_cb = yoffset; |
||
1315 | pp_set_surface2_tiling(ss2, tiling); |
||
1316 | dri_bo_emit_reloc(ss2_bo, |
||
1317 | I915_GEM_DOMAIN_RENDER, 0, |
||
1318 | surf_bo_offset, |
||
1319 | SURFACE_STATE_OFFSET(index) + offsetof(struct i965_surface_state2, ss0), |
||
1320 | surf_bo); |
||
1321 | ((unsigned int *)((char *)ss2_bo->virtual + BINDING_TABLE_OFFSET))[index] = SURFACE_STATE_OFFSET(index); |
||
1322 | dri_bo_unmap(ss2_bo); |
||
1323 | } |
||
1324 | |||
1325 | static void |
||
1326 | gen7_pp_set_surface_state(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1327 | dri_bo *surf_bo, unsigned long surf_bo_offset, |
||
1328 | int width, int height, int pitch, int format, |
||
1329 | int index, int is_target) |
||
1330 | { |
||
1331 | struct i965_driver_data * const i965 = i965_driver_data(ctx); |
||
1332 | struct gen7_surface_state *ss; |
||
1333 | dri_bo *ss_bo; |
||
1334 | unsigned int tiling; |
||
1335 | unsigned int swizzle; |
||
1336 | |||
1337 | dri_bo_get_tiling(surf_bo, &tiling, &swizzle); |
||
1338 | ss_bo = pp_context->surface_state_binding_table.bo; |
||
1339 | assert(ss_bo); |
||
1340 | |||
1341 | dri_bo_map(ss_bo, True); |
||
1342 | assert(ss_bo->virtual); |
||
1343 | ss = (struct gen7_surface_state *)((char *)ss_bo->virtual + SURFACE_STATE_OFFSET(index)); |
||
1344 | memset(ss, 0, sizeof(*ss)); |
||
1345 | ss->ss0.surface_type = I965_SURFACE_2D; |
||
1346 | ss->ss0.surface_format = format; |
||
1347 | ss->ss1.base_addr = surf_bo->offset + surf_bo_offset; |
||
1348 | ss->ss2.width = width - 1; |
||
1349 | ss->ss2.height = height - 1; |
||
1350 | ss->ss3.pitch = pitch - 1; |
||
1351 | gen7_pp_set_surface_tiling(ss, tiling); |
||
1352 | if (IS_HASWELL(i965->intel.device_id)) |
||
1353 | gen7_render_set_surface_scs(ss); |
||
1354 | dri_bo_emit_reloc(ss_bo, |
||
1355 | I915_GEM_DOMAIN_RENDER, is_target ? I915_GEM_DOMAIN_RENDER : 0, |
||
1356 | surf_bo_offset, |
||
1357 | SURFACE_STATE_OFFSET(index) + offsetof(struct gen7_surface_state, ss1), |
||
1358 | surf_bo); |
||
1359 | ((unsigned int *)((char *)ss_bo->virtual + BINDING_TABLE_OFFSET))[index] = SURFACE_STATE_OFFSET(index); |
||
1360 | dri_bo_unmap(ss_bo); |
||
1361 | } |
||
1362 | |||
1363 | static void |
||
1364 | gen7_pp_set_surface2_state(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1365 | dri_bo *surf_bo, unsigned long surf_bo_offset, |
||
1366 | int width, int height, int wpitch, |
||
1367 | int xoffset, int yoffset, |
||
1368 | int format, int interleave_chroma, |
||
1369 | int index) |
||
1370 | { |
||
1371 | struct gen7_surface_state2 *ss2; |
||
1372 | dri_bo *ss2_bo; |
||
1373 | unsigned int tiling; |
||
1374 | unsigned int swizzle; |
||
1375 | |||
1376 | dri_bo_get_tiling(surf_bo, &tiling, &swizzle); |
||
1377 | ss2_bo = pp_context->surface_state_binding_table.bo; |
||
1378 | assert(ss2_bo); |
||
1379 | |||
1380 | dri_bo_map(ss2_bo, True); |
||
1381 | assert(ss2_bo->virtual); |
||
1382 | ss2 = (struct gen7_surface_state2 *)((char *)ss2_bo->virtual + SURFACE_STATE_OFFSET(index)); |
||
1383 | memset(ss2, 0, sizeof(*ss2)); |
||
1384 | ss2->ss0.surface_base_address = surf_bo->offset + surf_bo_offset; |
||
1385 | ss2->ss1.cbcr_pixel_offset_v_direction = 0; |
||
1386 | ss2->ss1.width = width - 1; |
||
1387 | ss2->ss1.height = height - 1; |
||
1388 | ss2->ss2.pitch = wpitch - 1; |
||
1389 | ss2->ss2.interleave_chroma = interleave_chroma; |
||
1390 | ss2->ss2.surface_format = format; |
||
1391 | ss2->ss3.x_offset_for_cb = xoffset; |
||
1392 | ss2->ss3.y_offset_for_cb = yoffset; |
||
1393 | gen7_pp_set_surface2_tiling(ss2, tiling); |
||
1394 | dri_bo_emit_reloc(ss2_bo, |
||
1395 | I915_GEM_DOMAIN_RENDER, 0, |
||
1396 | surf_bo_offset, |
||
1397 | SURFACE_STATE_OFFSET(index) + offsetof(struct gen7_surface_state2, ss0), |
||
1398 | surf_bo); |
||
1399 | ((unsigned int *)((char *)ss2_bo->virtual + BINDING_TABLE_OFFSET))[index] = SURFACE_STATE_OFFSET(index); |
||
1400 | dri_bo_unmap(ss2_bo); |
||
1401 | } |
||
1402 | |||
1403 | |||
1404 | static void |
||
1405 | pp_set_media_rw_message_surface(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1406 | const struct i965_surface *surface, |
||
1407 | int base_index, int is_target, |
||
1408 | int *width, int *height, int *pitch, int *offset) |
||
1409 | { |
||
1410 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
1411 | struct object_surface *obj_surface; |
||
1412 | struct object_image *obj_image; |
||
1413 | dri_bo *bo; |
||
1414 | int fourcc = pp_get_surface_fourcc(ctx, surface); |
||
1415 | const int Y = 0; |
||
1416 | const int U = fourcc == VA_FOURCC('Y', 'V', '1', '2') ? 2 : 1; |
||
1417 | const int V = fourcc == VA_FOURCC('Y', 'V', '1', '2') ? 1 : 2; |
||
1418 | const int UV = 1; |
||
1419 | int interleaved_uv = fourcc == VA_FOURCC('N', 'V', '1', '2'); |
||
1420 | int packed_yuv = (fourcc == VA_FOURCC('Y', 'U', 'Y', '2') || fourcc == VA_FOURCC('U', 'Y', 'V', 'Y')); |
||
1421 | |||
1422 | int full_packed_format = (fourcc == VA_FOURCC('R', 'G', 'B', 'A') || |
||
1423 | fourcc == VA_FOURCC('R', 'G', 'B', 'X') || |
||
1424 | fourcc == VA_FOURCC('B', 'G', 'R', 'A') || |
||
1425 | fourcc == VA_FOURCC('B', 'G', 'R', 'X')); |
||
1426 | int scale_factor_of_1st_plane_width_in_byte = 1; |
||
1427 | |||
1428 | |||
1429 | if (surface->type == I965_SURFACE_TYPE_SURFACE) { |
||
1430 | obj_surface = SURFACE(surface->id); |
||
1431 | bo = obj_surface->bo; |
||
1432 | width[0] = obj_surface->orig_width; |
||
1433 | height[0] = obj_surface->orig_height; |
||
1434 | pitch[0] = obj_surface->width; |
||
1435 | offset[0] = 0; |
||
1436 | |||
1437 | if (full_packed_format) { |
||
1438 | scale_factor_of_1st_plane_width_in_byte = 4; |
||
1439 | pitch[0] = obj_surface->width * 4; |
||
1440 | } |
||
1441 | else if (packed_yuv ) { |
||
1442 | scale_factor_of_1st_plane_width_in_byte = 2; |
||
1443 | pitch[0] = obj_surface->width * 2; |
||
1444 | } |
||
1445 | else if (interleaved_uv) { |
||
1446 | width[1] = obj_surface->orig_width; |
||
1447 | height[1] = obj_surface->orig_height / 2; |
||
1448 | pitch[1] = obj_surface->width; |
||
1449 | offset[1] = offset[0] + obj_surface->width * obj_surface->height; |
||
1450 | } else { |
||
1451 | width[1] = obj_surface->orig_width / 2; |
||
1452 | height[1] = obj_surface->orig_height / 2; |
||
1453 | pitch[1] = obj_surface->width / 2; |
||
1454 | offset[1] = offset[0] + obj_surface->width * obj_surface->height; |
||
1455 | width[2] = obj_surface->orig_width / 2; |
||
1456 | height[2] = obj_surface->orig_height / 2; |
||
1457 | pitch[2] = obj_surface->width / 2; |
||
1458 | offset[2] = offset[1] + (obj_surface->width / 2) * (obj_surface->height / 2); |
||
1459 | } |
||
1460 | } else { |
||
1461 | obj_image = IMAGE(surface->id); |
||
1462 | bo = obj_image->bo; |
||
1463 | width[0] = obj_image->image.width; |
||
1464 | height[0] = obj_image->image.height; |
||
1465 | pitch[0] = obj_image->image.pitches[0]; |
||
1466 | offset[0] = obj_image->image.offsets[0]; |
||
1467 | |||
1468 | if (full_packed_format) { |
||
1469 | scale_factor_of_1st_plane_width_in_byte = 4; |
||
1470 | } |
||
1471 | else if (packed_yuv ) { |
||
1472 | scale_factor_of_1st_plane_width_in_byte = 2; |
||
1473 | } |
||
1474 | else if (interleaved_uv) { |
||
1475 | width[1] = obj_image->image.width; |
||
1476 | height[1] = obj_image->image.height / 2; |
||
1477 | pitch[1] = obj_image->image.pitches[1]; |
||
1478 | offset[1] = obj_image->image.offsets[1]; |
||
1479 | } else { |
||
1480 | width[1] = obj_image->image.width / 2; |
||
1481 | height[1] = obj_image->image.height / 2; |
||
1482 | pitch[1] = obj_image->image.pitches[1]; |
||
1483 | offset[1] = obj_image->image.offsets[1]; |
||
1484 | width[2] = obj_image->image.width / 2; |
||
1485 | height[2] = obj_image->image.height / 2; |
||
1486 | pitch[2] = obj_image->image.pitches[2]; |
||
1487 | offset[2] = obj_image->image.offsets[2]; |
||
1488 | } |
||
1489 | } |
||
1490 | |||
1491 | /* Y surface */ |
||
1492 | i965_pp_set_surface_state(ctx, pp_context, |
||
1493 | bo, offset[Y], |
||
1494 | width[Y] *scale_factor_of_1st_plane_width_in_byte / 4, height[Y], pitch[Y], I965_SURFACEFORMAT_R8_UNORM, |
||
1495 | base_index, is_target); |
||
1496 | |||
1497 | if (!packed_yuv && !full_packed_format) { |
||
1498 | if (interleaved_uv) { |
||
1499 | i965_pp_set_surface_state(ctx, pp_context, |
||
1500 | bo, offset[UV], |
||
1501 | width[UV] / 4, height[UV], pitch[UV], I965_SURFACEFORMAT_R8_UNORM, |
||
1502 | base_index + 1, is_target); |
||
1503 | } else { |
||
1504 | /* U surface */ |
||
1505 | i965_pp_set_surface_state(ctx, pp_context, |
||
1506 | bo, offset[U], |
||
1507 | width[U] / 4, height[U], pitch[U], I965_SURFACEFORMAT_R8_UNORM, |
||
1508 | base_index + 1, is_target); |
||
1509 | |||
1510 | /* V surface */ |
||
1511 | i965_pp_set_surface_state(ctx, pp_context, |
||
1512 | bo, offset[V], |
||
1513 | width[V] / 4, height[V], pitch[V], I965_SURFACEFORMAT_R8_UNORM, |
||
1514 | base_index + 2, is_target); |
||
1515 | } |
||
1516 | } |
||
1517 | |||
1518 | } |
||
1519 | |||
1520 | static void |
||
1521 | gen7_pp_set_media_rw_message_surface(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1522 | const struct i965_surface *surface, |
||
1523 | int base_index, int is_target, |
||
1524 | int *width, int *height, int *pitch, int *offset) |
||
1525 | { |
||
1526 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
1527 | struct object_surface *obj_surface; |
||
1528 | struct object_image *obj_image; |
||
1529 | dri_bo *bo; |
||
1530 | int fourcc = pp_get_surface_fourcc(ctx, surface); |
||
1531 | const int U = (fourcc == VA_FOURCC('Y', 'V', '1', '2') || |
||
1532 | fourcc == VA_FOURCC('I', 'M', 'C', '1')) ? 2 : 1; |
||
1533 | const int V = (fourcc == VA_FOURCC('Y', 'V', '1', '2') || |
||
1534 | fourcc == VA_FOURCC('I', 'M', 'C', '1')) ? 1 : 2; |
||
1535 | int interleaved_uv = fourcc == VA_FOURCC('N', 'V', '1', '2'); |
||
1536 | int packed_yuv = (fourcc == VA_FOURCC('Y', 'U', 'Y', '2') || fourcc == VA_FOURCC('U', 'Y', 'V', 'Y')); |
||
1537 | |||
1538 | if (surface->type == I965_SURFACE_TYPE_SURFACE) { |
||
1539 | obj_surface = SURFACE(surface->id); |
||
1540 | bo = obj_surface->bo; |
||
1541 | width[0] = obj_surface->orig_width; |
||
1542 | height[0] = obj_surface->orig_height; |
||
1543 | pitch[0] = obj_surface->width; |
||
1544 | offset[0] = 0; |
||
1545 | |||
1546 | if (packed_yuv) { |
||
1547 | if (is_target) |
||
1548 | width[0] = obj_surface->orig_width * 2; /* surface format is R8, so double the width */ |
||
1549 | else |
||
1550 | width[0] = obj_surface->orig_width; /* surface foramt is YCBCR, width is specified in units of pixels */ |
||
1551 | |||
1552 | pitch[0] = obj_surface->width * 2; |
||
1553 | } |
||
1554 | |||
1555 | width[1] = obj_surface->cb_cr_width; |
||
1556 | height[1] = obj_surface->cb_cr_height; |
||
1557 | pitch[1] = obj_surface->cb_cr_pitch; |
||
1558 | offset[1] = obj_surface->y_cb_offset * obj_surface->width; |
||
1559 | |||
1560 | width[2] = obj_surface->cb_cr_width; |
||
1561 | height[2] = obj_surface->cb_cr_height; |
||
1562 | pitch[2] = obj_surface->cb_cr_pitch; |
||
1563 | offset[2] = obj_surface->y_cr_offset * obj_surface->width; |
||
1564 | } else { |
||
1565 | obj_image = IMAGE(surface->id); |
||
1566 | bo = obj_image->bo; |
||
1567 | width[0] = obj_image->image.width; |
||
1568 | height[0] = obj_image->image.height; |
||
1569 | pitch[0] = obj_image->image.pitches[0]; |
||
1570 | offset[0] = obj_image->image.offsets[0]; |
||
1571 | |||
1572 | if (packed_yuv) { |
||
1573 | if (is_target) |
||
1574 | width[0] = obj_image->image.width * 2; /* surface format is R8, so double the width */ |
||
1575 | else |
||
1576 | width[0] = obj_image->image.width; /* surface foramt is YCBCR, width is specified in units of pixels */ |
||
1577 | } else if (interleaved_uv) { |
||
1578 | width[1] = obj_image->image.width / 2; |
||
1579 | height[1] = obj_image->image.height / 2; |
||
1580 | pitch[1] = obj_image->image.pitches[1]; |
||
1581 | offset[1] = obj_image->image.offsets[1]; |
||
1582 | } else { |
||
1583 | width[1] = obj_image->image.width / 2; |
||
1584 | height[1] = obj_image->image.height / 2; |
||
1585 | pitch[1] = obj_image->image.pitches[U]; |
||
1586 | offset[1] = obj_image->image.offsets[U]; |
||
1587 | width[2] = obj_image->image.width / 2; |
||
1588 | height[2] = obj_image->image.height / 2; |
||
1589 | pitch[2] = obj_image->image.pitches[V]; |
||
1590 | offset[2] = obj_image->image.offsets[V]; |
||
1591 | } |
||
1592 | } |
||
1593 | |||
1594 | if (is_target) { |
||
1595 | gen7_pp_set_surface_state(ctx, pp_context, |
||
1596 | bo, 0, |
||
1597 | width[0] / 4, height[0], pitch[0], |
||
1598 | I965_SURFACEFORMAT_R8_SINT, |
||
1599 | base_index, 1); |
||
1600 | |||
1601 | if (!packed_yuv) { |
||
1602 | if (interleaved_uv) { |
||
1603 | gen7_pp_set_surface_state(ctx, pp_context, |
||
1604 | bo, offset[1], |
||
1605 | width[1] / 2, height[1], pitch[1], |
||
1606 | I965_SURFACEFORMAT_R8G8_SINT, |
||
1607 | base_index + 1, 1); |
||
1608 | } else { |
||
1609 | gen7_pp_set_surface_state(ctx, pp_context, |
||
1610 | bo, offset[1], |
||
1611 | width[1] / 4, height[1], pitch[1], |
||
1612 | I965_SURFACEFORMAT_R8_SINT, |
||
1613 | base_index + 1, 1); |
||
1614 | gen7_pp_set_surface_state(ctx, pp_context, |
||
1615 | bo, offset[2], |
||
1616 | width[2] / 4, height[2], pitch[2], |
||
1617 | I965_SURFACEFORMAT_R8_SINT, |
||
1618 | base_index + 2, 1); |
||
1619 | } |
||
1620 | } |
||
1621 | } else { |
||
1622 | int format0 = SURFACE_FORMAT_Y8_UNORM; |
||
1623 | |||
1624 | switch (fourcc) { |
||
1625 | case VA_FOURCC('Y', 'U', 'Y', '2'): |
||
1626 | format0 = SURFACE_FORMAT_YCRCB_NORMAL; |
||
1627 | break; |
||
1628 | |||
1629 | case VA_FOURCC('U', 'Y', 'V', 'Y'): |
||
1630 | format0 = SURFACE_FORMAT_YCRCB_SWAPY; |
||
1631 | break; |
||
1632 | |||
1633 | default: |
||
1634 | break; |
||
1635 | } |
||
1636 | |||
1637 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
1638 | bo, offset[0], |
||
1639 | width[0], height[0], pitch[0], |
||
1640 | 0, 0, |
||
1641 | format0, 0, |
||
1642 | base_index); |
||
1643 | |||
1644 | if (!packed_yuv) { |
||
1645 | if (interleaved_uv) { |
||
1646 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
1647 | bo, offset[1], |
||
1648 | width[1], height[1], pitch[1], |
||
1649 | 0, 0, |
||
1650 | SURFACE_FORMAT_R8B8_UNORM, 0, |
||
1651 | base_index + 1); |
||
1652 | } else { |
||
1653 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
1654 | bo, offset[1], |
||
1655 | width[1], height[1], pitch[1], |
||
1656 | 0, 0, |
||
1657 | SURFACE_FORMAT_R8_UNORM, 0, |
||
1658 | base_index + 1); |
||
1659 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
1660 | bo, offset[2], |
||
1661 | width[2], height[2], pitch[2], |
||
1662 | 0, 0, |
||
1663 | SURFACE_FORMAT_R8_UNORM, 0, |
||
1664 | base_index + 2); |
||
1665 | } |
||
1666 | } |
||
1667 | } |
||
1668 | } |
||
1669 | |||
1670 | static int |
||
1671 | pp_null_x_steps(void *private_context) |
||
1672 | { |
||
1673 | return 1; |
||
1674 | } |
||
1675 | |||
1676 | static int |
||
1677 | pp_null_y_steps(void *private_context) |
||
1678 | { |
||
1679 | return 1; |
||
1680 | } |
||
1681 | |||
1682 | static int |
||
1683 | pp_null_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
1684 | { |
||
1685 | return 0; |
||
1686 | } |
||
1687 | |||
1688 | static VAStatus |
||
1689 | pp_null_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1690 | const struct i965_surface *src_surface, |
||
1691 | const VARectangle *src_rect, |
||
1692 | struct i965_surface *dst_surface, |
||
1693 | const VARectangle *dst_rect, |
||
1694 | void *filter_param) |
||
1695 | { |
||
1696 | /* private function & data */ |
||
1697 | pp_context->pp_x_steps = pp_null_x_steps; |
||
1698 | pp_context->pp_y_steps = pp_null_y_steps; |
||
1699 | pp_context->pp_set_block_parameter = pp_null_set_block_parameter; |
||
1700 | |||
1701 | dst_surface->flags = src_surface->flags; |
||
1702 | |||
1703 | return VA_STATUS_SUCCESS; |
||
1704 | } |
||
1705 | |||
1706 | static int |
||
1707 | pp_load_save_x_steps(void *private_context) |
||
1708 | { |
||
1709 | return 1; |
||
1710 | } |
||
1711 | |||
1712 | static int |
||
1713 | pp_load_save_y_steps(void *private_context) |
||
1714 | { |
||
1715 | struct pp_load_save_context *pp_load_save_context = private_context; |
||
1716 | |||
1717 | return pp_load_save_context->dest_h / 8; |
||
1718 | } |
||
1719 | |||
1720 | static int |
||
1721 | pp_load_save_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
1722 | { |
||
1723 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
1724 | struct pp_load_save_context *pp_load_save_context = (struct pp_load_save_context *)&pp_context->private_context; |
||
1725 | |||
1726 | pp_inline_parameter->grf5.destination_block_horizontal_origin = x * 16 + pp_load_save_context->dest_x; |
||
1727 | pp_inline_parameter->grf5.destination_block_vertical_origin = y * 8 + pp_load_save_context->dest_y; |
||
1728 | |||
1729 | return 0; |
||
1730 | } |
||
1731 | |||
1732 | static void calculate_boundary_block_mask(struct i965_post_processing_context *pp_context, const VARectangle *dst_rect) |
||
1733 | { |
||
1734 | int i; |
||
1735 | /* x offset of dest surface must be dword aligned. |
||
1736 | * so we have to extend dst surface on left edge, and mask out pixels not interested |
||
1737 | */ |
||
1738 | if (dst_rect->x%GPU_ASM_X_OFFSET_ALIGNMENT) { |
||
1739 | pp_context->block_horizontal_mask_left = 0; |
||
1740 | for (i=dst_rect->x%GPU_ASM_X_OFFSET_ALIGNMENT; i |
||
1741 | { |
||
1742 | pp_context->block_horizontal_mask_left |= 1< |
||
1743 | } |
||
1744 | } |
||
1745 | else { |
||
1746 | pp_context->block_horizontal_mask_left = 0xffff; |
||
1747 | } |
||
1748 | |||
1749 | int dst_width_adjust = dst_rect->width + dst_rect->x%GPU_ASM_X_OFFSET_ALIGNMENT; |
||
1750 | if (dst_width_adjust%GPU_ASM_BLOCK_WIDTH){ |
||
1751 | pp_context->block_horizontal_mask_right = (1 << (dst_width_adjust%GPU_ASM_BLOCK_WIDTH)) - 1; |
||
1752 | } |
||
1753 | else { |
||
1754 | pp_context->block_horizontal_mask_right = 0xffff; |
||
1755 | } |
||
1756 | |||
1757 | if (dst_rect->height%GPU_ASM_BLOCK_HEIGHT){ |
||
1758 | pp_context->block_vertical_mask_bottom = (1 << (dst_rect->height%GPU_ASM_BLOCK_HEIGHT)) - 1; |
||
1759 | } |
||
1760 | else { |
||
1761 | pp_context->block_vertical_mask_bottom = 0xff; |
||
1762 | } |
||
1763 | |||
1764 | } |
||
1765 | static VAStatus |
||
1766 | pp_plx_load_save_plx_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1767 | const struct i965_surface *src_surface, |
||
1768 | const VARectangle *src_rect, |
||
1769 | struct i965_surface *dst_surface, |
||
1770 | const VARectangle *dst_rect, |
||
1771 | void *filter_param) |
||
1772 | { |
||
1773 | struct pp_load_save_context *pp_load_save_context = (struct pp_load_save_context *)&pp_context->private_context; |
||
1774 | int width[3], height[3], pitch[3], offset[3]; |
||
1775 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
1776 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
1777 | |||
1778 | /* source surface */ |
||
1779 | pp_set_media_rw_message_surface(ctx, pp_context, src_surface, 1, 0, |
||
1780 | width, height, pitch, offset); |
||
1781 | |||
1782 | /* destination surface */ |
||
1783 | pp_set_media_rw_message_surface(ctx, pp_context, dst_surface, 7, 1, |
||
1784 | width, height, pitch, offset); |
||
1785 | |||
1786 | /* private function & data */ |
||
1787 | pp_context->pp_x_steps = pp_load_save_x_steps; |
||
1788 | pp_context->pp_y_steps = pp_load_save_y_steps; |
||
1789 | pp_context->pp_set_block_parameter = pp_load_save_set_block_parameter; |
||
1790 | |||
1791 | int dst_left_edge_extend = dst_rect->x%GPU_ASM_X_OFFSET_ALIGNMENT;; |
||
1792 | pp_load_save_context->dest_x = dst_rect->x - dst_left_edge_extend; |
||
1793 | pp_load_save_context->dest_y = dst_rect->y; |
||
1794 | pp_load_save_context->dest_h = ALIGN(dst_rect->height, 8); |
||
1795 | pp_load_save_context->dest_w = ALIGN(dst_rect->width+dst_left_edge_extend, 16); |
||
1796 | |||
1797 | pp_inline_parameter->grf5.block_count_x = pp_load_save_context->dest_w / 16; /* 1 x N */ |
||
1798 | pp_inline_parameter->grf5.number_blocks = pp_load_save_context->dest_w / 16; |
||
1799 | |||
1800 | pp_static_parameter->grf3.horizontal_origin_offset = src_rect->x; |
||
1801 | pp_static_parameter->grf3.vertical_origin_offset = src_rect->y; |
||
1802 | |||
1803 | // update u/v offset for packed yuv |
||
1804 | i965_update_src_surface_uv_offset (ctx, pp_context, src_surface); |
||
1805 | i965_update_dst_surface_uv_offset (ctx, pp_context, dst_surface); |
||
1806 | |||
1807 | dst_surface->flags = src_surface->flags; |
||
1808 | |||
1809 | return VA_STATUS_SUCCESS; |
||
1810 | } |
||
1811 | |||
1812 | static int |
||
1813 | pp_scaling_x_steps(void *private_context) |
||
1814 | { |
||
1815 | return 1; |
||
1816 | } |
||
1817 | |||
1818 | static int |
||
1819 | pp_scaling_y_steps(void *private_context) |
||
1820 | { |
||
1821 | struct pp_scaling_context *pp_scaling_context = private_context; |
||
1822 | |||
1823 | return pp_scaling_context->dest_h / 8; |
||
1824 | } |
||
1825 | |||
1826 | static int |
||
1827 | pp_scaling_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
1828 | { |
||
1829 | struct pp_scaling_context *pp_scaling_context = (struct pp_scaling_context *)&pp_context->private_context; |
||
1830 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
1831 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
1832 | float src_x_steping = pp_inline_parameter->grf5.normalized_video_x_scaling_step; |
||
1833 | float src_y_steping = pp_static_parameter->grf1.r1_6.normalized_video_y_scaling_step; |
||
1834 | |||
1835 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin = src_x_steping * x * 16 + pp_scaling_context->src_normalized_x; |
||
1836 | pp_inline_parameter->grf5.source_surface_block_normalized_vertical_origin = src_y_steping * y * 8 + pp_scaling_context->src_normalized_y; |
||
1837 | pp_inline_parameter->grf5.destination_block_horizontal_origin = x * 16 + pp_scaling_context->dest_x; |
||
1838 | pp_inline_parameter->grf5.destination_block_vertical_origin = y * 8 + pp_scaling_context->dest_y; |
||
1839 | |||
1840 | return 0; |
||
1841 | } |
||
1842 | |||
1843 | static VAStatus |
||
1844 | pp_nv12_scaling_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
1845 | const struct i965_surface *src_surface, |
||
1846 | const VARectangle *src_rect, |
||
1847 | struct i965_surface *dst_surface, |
||
1848 | const VARectangle *dst_rect, |
||
1849 | void *filter_param) |
||
1850 | { |
||
1851 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
1852 | struct pp_scaling_context *pp_scaling_context = (struct pp_scaling_context *)&pp_context->private_context; |
||
1853 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
1854 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
1855 | struct object_surface *obj_surface; |
||
1856 | struct i965_sampler_state *sampler_state; |
||
1857 | int in_w, in_h, in_wpitch, in_hpitch; |
||
1858 | int out_w, out_h, out_wpitch, out_hpitch; |
||
1859 | |||
1860 | /* source surface */ |
||
1861 | obj_surface = SURFACE(src_surface->id); |
||
1862 | in_w = obj_surface->orig_width; |
||
1863 | in_h = obj_surface->orig_height; |
||
1864 | in_wpitch = obj_surface->width; |
||
1865 | in_hpitch = obj_surface->height; |
||
1866 | |||
1867 | /* source Y surface index 1 */ |
||
1868 | i965_pp_set_surface_state(ctx, pp_context, |
||
1869 | obj_surface->bo, 0, |
||
1870 | in_w, in_h, in_wpitch, I965_SURFACEFORMAT_R8_UNORM, |
||
1871 | 1, 0); |
||
1872 | |||
1873 | /* source UV surface index 2 */ |
||
1874 | i965_pp_set_surface_state(ctx, pp_context, |
||
1875 | obj_surface->bo, in_wpitch * in_hpitch, |
||
1876 | in_w / 2, in_h / 2, in_wpitch, I965_SURFACEFORMAT_R8G8_UNORM, |
||
1877 | 2, 0); |
||
1878 | |||
1879 | /* destination surface */ |
||
1880 | obj_surface = SURFACE(dst_surface->id); |
||
1881 | out_w = obj_surface->orig_width; |
||
1882 | out_h = obj_surface->orig_height; |
||
1883 | out_wpitch = obj_surface->width; |
||
1884 | out_hpitch = obj_surface->height; |
||
1885 | |||
1886 | /* destination Y surface index 7 */ |
||
1887 | i965_pp_set_surface_state(ctx, pp_context, |
||
1888 | obj_surface->bo, 0, |
||
1889 | out_w / 4, out_h, out_wpitch, I965_SURFACEFORMAT_R8_UNORM, |
||
1890 | 7, 1); |
||
1891 | |||
1892 | /* destination UV surface index 8 */ |
||
1893 | i965_pp_set_surface_state(ctx, pp_context, |
||
1894 | obj_surface->bo, out_wpitch * out_hpitch, |
||
1895 | out_w / 4, out_h / 2, out_wpitch, I965_SURFACEFORMAT_R8G8_UNORM, |
||
1896 | 8, 1); |
||
1897 | |||
1898 | /* sampler state */ |
||
1899 | dri_bo_map(pp_context->sampler_state_table.bo, True); |
||
1900 | assert(pp_context->sampler_state_table.bo->virtual); |
||
1901 | sampler_state = pp_context->sampler_state_table.bo->virtual; |
||
1902 | |||
1903 | /* SIMD16 Y index 1 */ |
||
1904 | sampler_state[1].ss0.min_filter = I965_MAPFILTER_LINEAR; |
||
1905 | sampler_state[1].ss0.mag_filter = I965_MAPFILTER_LINEAR; |
||
1906 | sampler_state[1].ss1.r_wrap_mode = I965_TEXCOORDMODE_CLAMP; |
||
1907 | sampler_state[1].ss1.s_wrap_mode = I965_TEXCOORDMODE_CLAMP; |
||
1908 | sampler_state[1].ss1.t_wrap_mode = I965_TEXCOORDMODE_CLAMP; |
||
1909 | |||
1910 | /* SIMD16 UV index 2 */ |
||
1911 | sampler_state[2].ss0.min_filter = I965_MAPFILTER_LINEAR; |
||
1912 | sampler_state[2].ss0.mag_filter = I965_MAPFILTER_LINEAR; |
||
1913 | sampler_state[2].ss1.r_wrap_mode = I965_TEXCOORDMODE_CLAMP; |
||
1914 | sampler_state[2].ss1.s_wrap_mode = I965_TEXCOORDMODE_CLAMP; |
||
1915 | sampler_state[2].ss1.t_wrap_mode = I965_TEXCOORDMODE_CLAMP; |
||
1916 | |||
1917 | dri_bo_unmap(pp_context->sampler_state_table.bo); |
||
1918 | |||
1919 | /* private function & data */ |
||
1920 | pp_context->pp_x_steps = pp_scaling_x_steps; |
||
1921 | pp_context->pp_y_steps = pp_scaling_y_steps; |
||
1922 | pp_context->pp_set_block_parameter = pp_scaling_set_block_parameter; |
||
1923 | |||
1924 | int dst_left_edge_extend = dst_rect->x%GPU_ASM_X_OFFSET_ALIGNMENT; |
||
1925 | float src_left_edge_extend = (float)dst_left_edge_extend*src_rect->width/dst_rect->width; |
||
1926 | pp_scaling_context->dest_x = dst_rect->x - dst_left_edge_extend; |
||
1927 | pp_scaling_context->dest_y = dst_rect->y; |
||
1928 | pp_scaling_context->dest_w = ALIGN(dst_rect->width + dst_left_edge_extend, 16); |
||
1929 | pp_scaling_context->dest_h = ALIGN(dst_rect->height, 8); |
||
1930 | pp_scaling_context->src_normalized_x = (float)(src_rect->x - src_left_edge_extend)/ in_w; |
||
1931 | pp_scaling_context->src_normalized_y = (float)src_rect->y / in_h; |
||
1932 | |||
1933 | pp_static_parameter->grf1.r1_6.normalized_video_y_scaling_step = (float) src_rect->height / in_h / dst_rect->height; |
||
1934 | |||
1935 | pp_inline_parameter->grf5.normalized_video_x_scaling_step = (float) (src_rect->width + src_left_edge_extend)/ in_w / (dst_rect->width + dst_left_edge_extend); |
||
1936 | pp_inline_parameter->grf5.block_count_x = pp_scaling_context->dest_w / 16; /* 1 x N */ |
||
1937 | pp_inline_parameter->grf5.number_blocks = pp_scaling_context->dest_w / 16; |
||
1938 | |||
1939 | dst_surface->flags = src_surface->flags; |
||
1940 | |||
1941 | return VA_STATUS_SUCCESS; |
||
1942 | } |
||
1943 | |||
1944 | static int |
||
1945 | pp_avs_x_steps(void *private_context) |
||
1946 | { |
||
1947 | struct pp_avs_context *pp_avs_context = private_context; |
||
1948 | |||
1949 | return pp_avs_context->dest_w / 16; |
||
1950 | } |
||
1951 | |||
1952 | static int |
||
1953 | pp_avs_y_steps(void *private_context) |
||
1954 | { |
||
1955 | return 1; |
||
1956 | } |
||
1957 | |||
1958 | static int |
||
1959 | pp_avs_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
1960 | { |
||
1961 | struct pp_avs_context *pp_avs_context = (struct pp_avs_context *)&pp_context->private_context; |
||
1962 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
1963 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
1964 | float src_x_steping, src_y_steping, video_step_delta; |
||
1965 | int tmp_w = ALIGN(pp_avs_context->dest_h * pp_avs_context->src_w / pp_avs_context->src_h, 16); |
||
1966 | |||
1967 | if (pp_static_parameter->grf4.r4_2.avs.nlas == 0) { |
||
1968 | src_x_steping = pp_inline_parameter->grf5.normalized_video_x_scaling_step; |
||
1969 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin = src_x_steping * x * 16 + pp_avs_context->src_normalized_x; |
||
1970 | } else if (tmp_w >= pp_avs_context->dest_w) { |
||
1971 | pp_inline_parameter->grf5.normalized_video_x_scaling_step = 1.0 / tmp_w; |
||
1972 | pp_inline_parameter->grf6.video_step_delta = 0; |
||
1973 | |||
1974 | if (x == 0) { |
||
1975 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin = (float)(tmp_w - pp_avs_context->dest_w) / tmp_w / 2 + |
||
1976 | pp_avs_context->src_normalized_x; |
||
1977 | } else { |
||
1978 | src_x_steping = pp_inline_parameter->grf5.normalized_video_x_scaling_step; |
||
1979 | video_step_delta = pp_inline_parameter->grf6.video_step_delta; |
||
1980 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin += src_x_steping * 16 + |
||
1981 | 16 * 15 * video_step_delta / 2; |
||
1982 | } |
||
1983 | } else { |
||
1984 | int n0, n1, n2, nls_left, nls_right; |
||
1985 | int factor_a = 5, factor_b = 4; |
||
1986 | float f; |
||
1987 | |||
1988 | n0 = (pp_avs_context->dest_w - tmp_w) / (16 * 2); |
||
1989 | n1 = (pp_avs_context->dest_w - tmp_w) / 16 - n0; |
||
1990 | n2 = tmp_w / (16 * factor_a); |
||
1991 | nls_left = n0 + n2; |
||
1992 | nls_right = n1 + n2; |
||
1993 | f = (float) n2 * 16 / tmp_w; |
||
1994 | |||
1995 | if (n0 < 5) { |
||
1996 | pp_inline_parameter->grf6.video_step_delta = 0.0; |
||
1997 | |||
1998 | if (x == 0) { |
||
1999 | pp_inline_parameter->grf5.normalized_video_x_scaling_step = 1.0 / pp_avs_context->dest_w; |
||
2000 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin = pp_avs_context->src_normalized_x; |
||
2001 | } else { |
||
2002 | src_x_steping = pp_inline_parameter->grf5.normalized_video_x_scaling_step; |
||
2003 | video_step_delta = pp_inline_parameter->grf6.video_step_delta; |
||
2004 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin += src_x_steping * 16 + |
||
2005 | 16 * 15 * video_step_delta / 2; |
||
2006 | } |
||
2007 | } else { |
||
2008 | if (x < nls_left) { |
||
2009 | /* f = a * nls_left * 16 + b * nls_left * 16 * (nls_left * 16 - 1) / 2 */ |
||
2010 | float a = f / (nls_left * 16 * factor_b); |
||
2011 | float b = (f - nls_left * 16 * a) * 2 / (nls_left * 16 * (nls_left * 16 - 1)); |
||
2012 | |||
2013 | pp_inline_parameter->grf6.video_step_delta = b; |
||
2014 | |||
2015 | if (x == 0) { |
||
2016 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin = pp_avs_context->src_normalized_x; |
||
2017 | pp_inline_parameter->grf5.normalized_video_x_scaling_step = a; |
||
2018 | } else { |
||
2019 | src_x_steping = pp_inline_parameter->grf5.normalized_video_x_scaling_step; |
||
2020 | video_step_delta = pp_inline_parameter->grf6.video_step_delta; |
||
2021 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin += src_x_steping * 16 + |
||
2022 | 16 * 15 * video_step_delta / 2; |
||
2023 | pp_inline_parameter->grf5.normalized_video_x_scaling_step += 16 * b; |
||
2024 | } |
||
2025 | } else if (x < (pp_avs_context->dest_w / 16 - nls_right)) { |
||
2026 | /* scale the center linearly */ |
||
2027 | src_x_steping = pp_inline_parameter->grf5.normalized_video_x_scaling_step; |
||
2028 | video_step_delta = pp_inline_parameter->grf6.video_step_delta; |
||
2029 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin += src_x_steping * 16 + |
||
2030 | 16 * 15 * video_step_delta / 2; |
||
2031 | pp_inline_parameter->grf6.video_step_delta = 0.0; |
||
2032 | pp_inline_parameter->grf5.normalized_video_x_scaling_step = 1.0 / tmp_w; |
||
2033 | } else { |
||
2034 | float a = f / (nls_right * 16 * factor_b); |
||
2035 | float b = (f - nls_right * 16 * a) * 2 / (nls_right * 16 * (nls_right * 16 - 1)); |
||
2036 | |||
2037 | src_x_steping = pp_inline_parameter->grf5.normalized_video_x_scaling_step; |
||
2038 | video_step_delta = pp_inline_parameter->grf6.video_step_delta; |
||
2039 | pp_inline_parameter->grf5.r5_1.source_surface_block_normalized_horizontal_origin += src_x_steping * 16 + |
||
2040 | 16 * 15 * video_step_delta / 2; |
||
2041 | pp_inline_parameter->grf6.video_step_delta = -b; |
||
2042 | |||
2043 | if (x == (pp_avs_context->dest_w / 16 - nls_right)) |
||
2044 | pp_inline_parameter->grf5.normalized_video_x_scaling_step = a + (nls_right * 16 - 1) * b; |
||
2045 | else |
||
2046 | pp_inline_parameter->grf5.normalized_video_x_scaling_step -= b * 16; |
||
2047 | } |
||
2048 | } |
||
2049 | } |
||
2050 | |||
2051 | src_y_steping = pp_static_parameter->grf1.r1_6.normalized_video_y_scaling_step; |
||
2052 | pp_inline_parameter->grf5.source_surface_block_normalized_vertical_origin = src_y_steping * y * 8 + pp_avs_context->src_normalized_y; |
||
2053 | pp_inline_parameter->grf5.destination_block_horizontal_origin = x * 16 + pp_avs_context->dest_x; |
||
2054 | pp_inline_parameter->grf5.destination_block_vertical_origin = y * 8 + pp_avs_context->dest_y; |
||
2055 | |||
2056 | return 0; |
||
2057 | } |
||
2058 | |||
2059 | static VAStatus |
||
2060 | pp_nv12_avs_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
2061 | const struct i965_surface *src_surface, |
||
2062 | const VARectangle *src_rect, |
||
2063 | struct i965_surface *dst_surface, |
||
2064 | const VARectangle *dst_rect, |
||
2065 | void *filter_param, |
||
2066 | int nlas) |
||
2067 | { |
||
2068 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
2069 | struct pp_avs_context *pp_avs_context = (struct pp_avs_context *)&pp_context->private_context; |
||
2070 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
2071 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
2072 | struct object_surface *obj_surface; |
||
2073 | struct i965_sampler_8x8 *sampler_8x8; |
||
2074 | struct i965_sampler_8x8_state *sampler_8x8_state; |
||
2075 | int index; |
||
2076 | int in_w, in_h, in_wpitch, in_hpitch; |
||
2077 | int out_w, out_h, out_wpitch, out_hpitch; |
||
2078 | int i; |
||
2079 | |||
2080 | /* surface */ |
||
2081 | obj_surface = SURFACE(src_surface->id); |
||
2082 | in_w = obj_surface->orig_width; |
||
2083 | in_h = obj_surface->orig_height; |
||
2084 | in_wpitch = obj_surface->width; |
||
2085 | in_hpitch = obj_surface->height; |
||
2086 | |||
2087 | /* source Y surface index 1 */ |
||
2088 | i965_pp_set_surface2_state(ctx, pp_context, |
||
2089 | obj_surface->bo, 0, |
||
2090 | in_w, in_h, in_wpitch, |
||
2091 | 0, 0, |
||
2092 | SURFACE_FORMAT_Y8_UNORM, 0, |
||
2093 | 1); |
||
2094 | |||
2095 | /* source UV surface index 2 */ |
||
2096 | i965_pp_set_surface2_state(ctx, pp_context, |
||
2097 | obj_surface->bo, in_wpitch * in_hpitch, |
||
2098 | in_w / 2, in_h / 2, in_wpitch, |
||
2099 | 0, 0, |
||
2100 | SURFACE_FORMAT_R8B8_UNORM, 0, |
||
2101 | 2); |
||
2102 | |||
2103 | /* destination surface */ |
||
2104 | obj_surface = SURFACE(dst_surface->id); |
||
2105 | out_w = obj_surface->orig_width; |
||
2106 | out_h = obj_surface->orig_height; |
||
2107 | out_wpitch = obj_surface->width; |
||
2108 | out_hpitch = obj_surface->height; |
||
2109 | assert(out_w <= out_wpitch && out_h <= out_hpitch); |
||
2110 | |||
2111 | /* destination Y surface index 7 */ |
||
2112 | i965_pp_set_surface_state(ctx, pp_context, |
||
2113 | obj_surface->bo, 0, |
||
2114 | out_w / 4, out_h, out_wpitch, I965_SURFACEFORMAT_R8_UNORM, |
||
2115 | 7, 1); |
||
2116 | |||
2117 | /* destination UV surface index 8 */ |
||
2118 | i965_pp_set_surface_state(ctx, pp_context, |
||
2119 | obj_surface->bo, out_wpitch * out_hpitch, |
||
2120 | out_w / 4, out_h / 2, out_wpitch, I965_SURFACEFORMAT_R8G8_UNORM, |
||
2121 | 8, 1); |
||
2122 | |||
2123 | /* sampler 8x8 state */ |
||
2124 | dri_bo_map(pp_context->sampler_state_table.bo_8x8, True); |
||
2125 | assert(pp_context->sampler_state_table.bo_8x8->virtual); |
||
2126 | assert(sizeof(*sampler_8x8_state) == sizeof(int) * 138); |
||
2127 | sampler_8x8_state = pp_context->sampler_state_table.bo_8x8->virtual; |
||
2128 | memset(sampler_8x8_state, 0, sizeof(*sampler_8x8_state)); |
||
2129 | |||
2130 | for (i = 0; i < 17; i++) { |
||
2131 | /* for Y channel, currently ignore */ |
||
2132 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c0 = 0x00; |
||
2133 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c1 = 0x00; |
||
2134 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c2 = 0x08; |
||
2135 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c3 = 0x18; |
||
2136 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c4 = 0x18; |
||
2137 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c5 = 0x08; |
||
2138 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c6 = 0x00; |
||
2139 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c7 = 0x00; |
||
2140 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c0 = 0x00; |
||
2141 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c1 = 0x00; |
||
2142 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c2 = 0x10; |
||
2143 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c3 = 0x10; |
||
2144 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c4 = 0x10; |
||
2145 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c5 = 0x10; |
||
2146 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c6 = 0x00; |
||
2147 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c7 = 0x00; |
||
2148 | /* for U/V channel, 0.25 */ |
||
2149 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c0 = 0x0; |
||
2150 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c1 = 0x0; |
||
2151 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c2 = 0x10; |
||
2152 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c3 = 0x10; |
||
2153 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c4 = 0x10; |
||
2154 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c5 = 0x10; |
||
2155 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c6 = 0x0; |
||
2156 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c7 = 0x0; |
||
2157 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c0 = 0x0; |
||
2158 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c1 = 0x0; |
||
2159 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c2 = 0x10; |
||
2160 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c3 = 0x10; |
||
2161 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c4 = 0x10; |
||
2162 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c5 = 0x10; |
||
2163 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c6 = 0x0; |
||
2164 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c7 = 0x0; |
||
2165 | } |
||
2166 | |||
2167 | sampler_8x8_state->dw136.default_sharpness_level = 0; |
||
2168 | sampler_8x8_state->dw137.adaptive_filter_for_all_channel = 1; |
||
2169 | sampler_8x8_state->dw137.bypass_y_adaptive_filtering = 1; |
||
2170 | sampler_8x8_state->dw137.bypass_x_adaptive_filtering = 1; |
||
2171 | dri_bo_unmap(pp_context->sampler_state_table.bo_8x8); |
||
2172 | |||
2173 | /* sampler 8x8 */ |
||
2174 | dri_bo_map(pp_context->sampler_state_table.bo, True); |
||
2175 | assert(pp_context->sampler_state_table.bo->virtual); |
||
2176 | assert(sizeof(*sampler_8x8) == sizeof(int) * 16); |
||
2177 | sampler_8x8 = pp_context->sampler_state_table.bo->virtual; |
||
2178 | |||
2179 | /* sample_8x8 Y index 1 */ |
||
2180 | index = 1; |
||
2181 | memset(&sampler_8x8[index], 0, sizeof(*sampler_8x8)); |
||
2182 | sampler_8x8[index].dw0.avs_filter_type = AVS_FILTER_ADAPTIVE_8_TAP; |
||
2183 | sampler_8x8[index].dw0.ief_bypass = 1; |
||
2184 | sampler_8x8[index].dw0.ief_filter_type = IEF_FILTER_DETAIL; |
||
2185 | sampler_8x8[index].dw0.ief_filter_size = IEF_FILTER_SIZE_5X5; |
||
2186 | sampler_8x8[index].dw1.sampler_8x8_state_pointer = pp_context->sampler_state_table.bo_8x8->offset >> 5; |
||
2187 | sampler_8x8[index].dw2.global_noise_estimation = 22; |
||
2188 | sampler_8x8[index].dw2.strong_edge_threshold = 8; |
||
2189 | sampler_8x8[index].dw2.weak_edge_threshold = 1; |
||
2190 | sampler_8x8[index].dw3.strong_edge_weight = 7; |
||
2191 | sampler_8x8[index].dw3.regular_weight = 2; |
||
2192 | sampler_8x8[index].dw3.non_edge_weight = 0; |
||
2193 | sampler_8x8[index].dw3.gain_factor = 40; |
||
2194 | sampler_8x8[index].dw4.steepness_boost = 0; |
||
2195 | sampler_8x8[index].dw4.steepness_threshold = 0; |
||
2196 | sampler_8x8[index].dw4.mr_boost = 0; |
||
2197 | sampler_8x8[index].dw4.mr_threshold = 5; |
||
2198 | sampler_8x8[index].dw5.pwl1_point_1 = 4; |
||
2199 | sampler_8x8[index].dw5.pwl1_point_2 = 12; |
||
2200 | sampler_8x8[index].dw5.pwl1_point_3 = 16; |
||
2201 | sampler_8x8[index].dw5.pwl1_point_4 = 26; |
||
2202 | sampler_8x8[index].dw6.pwl1_point_5 = 40; |
||
2203 | sampler_8x8[index].dw6.pwl1_point_6 = 160; |
||
2204 | sampler_8x8[index].dw6.pwl1_r3_bias_0 = 127; |
||
2205 | sampler_8x8[index].dw6.pwl1_r3_bias_1 = 98; |
||
2206 | sampler_8x8[index].dw7.pwl1_r3_bias_2 = 88; |
||
2207 | sampler_8x8[index].dw7.pwl1_r3_bias_3 = 64; |
||
2208 | sampler_8x8[index].dw7.pwl1_r3_bias_4 = 44; |
||
2209 | sampler_8x8[index].dw7.pwl1_r3_bias_5 = 0; |
||
2210 | sampler_8x8[index].dw8.pwl1_r3_bias_6 = 0; |
||
2211 | sampler_8x8[index].dw8.pwl1_r5_bias_0 = 3; |
||
2212 | sampler_8x8[index].dw8.pwl1_r5_bias_1 = 32; |
||
2213 | sampler_8x8[index].dw8.pwl1_r5_bias_2 = 32; |
||
2214 | sampler_8x8[index].dw9.pwl1_r5_bias_3 = 58; |
||
2215 | sampler_8x8[index].dw9.pwl1_r5_bias_4 = 100; |
||
2216 | sampler_8x8[index].dw9.pwl1_r5_bias_5 = 108; |
||
2217 | sampler_8x8[index].dw9.pwl1_r5_bias_6 = 88; |
||
2218 | sampler_8x8[index].dw10.pwl1_r3_slope_0 = -116; |
||
2219 | sampler_8x8[index].dw10.pwl1_r3_slope_1 = -20; |
||
2220 | sampler_8x8[index].dw10.pwl1_r3_slope_2 = -96; |
||
2221 | sampler_8x8[index].dw10.pwl1_r3_slope_3 = -32; |
||
2222 | sampler_8x8[index].dw11.pwl1_r3_slope_4 = -50; |
||
2223 | sampler_8x8[index].dw11.pwl1_r3_slope_5 = 0; |
||
2224 | sampler_8x8[index].dw11.pwl1_r3_slope_6 = 0; |
||
2225 | sampler_8x8[index].dw11.pwl1_r5_slope_0 = 116; |
||
2226 | sampler_8x8[index].dw12.pwl1_r5_slope_1 = 0; |
||
2227 | sampler_8x8[index].dw12.pwl1_r5_slope_2 = 114; |
||
2228 | sampler_8x8[index].dw12.pwl1_r5_slope_3 = 67; |
||
2229 | sampler_8x8[index].dw12.pwl1_r5_slope_4 = 9; |
||
2230 | sampler_8x8[index].dw13.pwl1_r5_slope_5 = -3; |
||
2231 | sampler_8x8[index].dw13.pwl1_r5_slope_6 = -15; |
||
2232 | sampler_8x8[index].dw13.limiter_boost = 0; |
||
2233 | sampler_8x8[index].dw13.minimum_limiter = 10; |
||
2234 | sampler_8x8[index].dw13.maximum_limiter = 11; |
||
2235 | sampler_8x8[index].dw14.clip_limiter = 130; |
||
2236 | dri_bo_emit_reloc(pp_context->sampler_state_table.bo, |
||
2237 | I915_GEM_DOMAIN_RENDER, |
||
2238 | 0, |
||
2239 | 0, |
||
2240 | sizeof(*sampler_8x8) * index + offsetof(struct i965_sampler_8x8, dw1), |
||
2241 | pp_context->sampler_state_table.bo_8x8); |
||
2242 | |||
2243 | /* sample_8x8 UV index 2 */ |
||
2244 | index = 2; |
||
2245 | memset(&sampler_8x8[index], 0, sizeof(*sampler_8x8)); |
||
2246 | sampler_8x8[index].dw0.avs_filter_type = AVS_FILTER_ADAPTIVE_8_TAP; |
||
2247 | sampler_8x8[index].dw0.ief_bypass = 1; |
||
2248 | sampler_8x8[index].dw0.ief_filter_type = IEF_FILTER_DETAIL; |
||
2249 | sampler_8x8[index].dw0.ief_filter_size = IEF_FILTER_SIZE_5X5; |
||
2250 | sampler_8x8[index].dw1.sampler_8x8_state_pointer = pp_context->sampler_state_table.bo_8x8->offset >> 5; |
||
2251 | sampler_8x8[index].dw2.global_noise_estimation = 22; |
||
2252 | sampler_8x8[index].dw2.strong_edge_threshold = 8; |
||
2253 | sampler_8x8[index].dw2.weak_edge_threshold = 1; |
||
2254 | sampler_8x8[index].dw3.strong_edge_weight = 7; |
||
2255 | sampler_8x8[index].dw3.regular_weight = 2; |
||
2256 | sampler_8x8[index].dw3.non_edge_weight = 0; |
||
2257 | sampler_8x8[index].dw3.gain_factor = 40; |
||
2258 | sampler_8x8[index].dw4.steepness_boost = 0; |
||
2259 | sampler_8x8[index].dw4.steepness_threshold = 0; |
||
2260 | sampler_8x8[index].dw4.mr_boost = 0; |
||
2261 | sampler_8x8[index].dw4.mr_threshold = 5; |
||
2262 | sampler_8x8[index].dw5.pwl1_point_1 = 4; |
||
2263 | sampler_8x8[index].dw5.pwl1_point_2 = 12; |
||
2264 | sampler_8x8[index].dw5.pwl1_point_3 = 16; |
||
2265 | sampler_8x8[index].dw5.pwl1_point_4 = 26; |
||
2266 | sampler_8x8[index].dw6.pwl1_point_5 = 40; |
||
2267 | sampler_8x8[index].dw6.pwl1_point_6 = 160; |
||
2268 | sampler_8x8[index].dw6.pwl1_r3_bias_0 = 127; |
||
2269 | sampler_8x8[index].dw6.pwl1_r3_bias_1 = 98; |
||
2270 | sampler_8x8[index].dw7.pwl1_r3_bias_2 = 88; |
||
2271 | sampler_8x8[index].dw7.pwl1_r3_bias_3 = 64; |
||
2272 | sampler_8x8[index].dw7.pwl1_r3_bias_4 = 44; |
||
2273 | sampler_8x8[index].dw7.pwl1_r3_bias_5 = 0; |
||
2274 | sampler_8x8[index].dw8.pwl1_r3_bias_6 = 0; |
||
2275 | sampler_8x8[index].dw8.pwl1_r5_bias_0 = 3; |
||
2276 | sampler_8x8[index].dw8.pwl1_r5_bias_1 = 32; |
||
2277 | sampler_8x8[index].dw8.pwl1_r5_bias_2 = 32; |
||
2278 | sampler_8x8[index].dw9.pwl1_r5_bias_3 = 58; |
||
2279 | sampler_8x8[index].dw9.pwl1_r5_bias_4 = 100; |
||
2280 | sampler_8x8[index].dw9.pwl1_r5_bias_5 = 108; |
||
2281 | sampler_8x8[index].dw9.pwl1_r5_bias_6 = 88; |
||
2282 | sampler_8x8[index].dw10.pwl1_r3_slope_0 = -116; |
||
2283 | sampler_8x8[index].dw10.pwl1_r3_slope_1 = -20; |
||
2284 | sampler_8x8[index].dw10.pwl1_r3_slope_2 = -96; |
||
2285 | sampler_8x8[index].dw10.pwl1_r3_slope_3 = -32; |
||
2286 | sampler_8x8[index].dw11.pwl1_r3_slope_4 = -50; |
||
2287 | sampler_8x8[index].dw11.pwl1_r3_slope_5 = 0; |
||
2288 | sampler_8x8[index].dw11.pwl1_r3_slope_6 = 0; |
||
2289 | sampler_8x8[index].dw11.pwl1_r5_slope_0 = 116; |
||
2290 | sampler_8x8[index].dw12.pwl1_r5_slope_1 = 0; |
||
2291 | sampler_8x8[index].dw12.pwl1_r5_slope_2 = 114; |
||
2292 | sampler_8x8[index].dw12.pwl1_r5_slope_3 = 67; |
||
2293 | sampler_8x8[index].dw12.pwl1_r5_slope_4 = 9; |
||
2294 | sampler_8x8[index].dw13.pwl1_r5_slope_5 = -3; |
||
2295 | sampler_8x8[index].dw13.pwl1_r5_slope_6 = -15; |
||
2296 | sampler_8x8[index].dw13.limiter_boost = 0; |
||
2297 | sampler_8x8[index].dw13.minimum_limiter = 10; |
||
2298 | sampler_8x8[index].dw13.maximum_limiter = 11; |
||
2299 | sampler_8x8[index].dw14.clip_limiter = 130; |
||
2300 | dri_bo_emit_reloc(pp_context->sampler_state_table.bo, |
||
2301 | I915_GEM_DOMAIN_RENDER, |
||
2302 | 0, |
||
2303 | 0, |
||
2304 | sizeof(*sampler_8x8) * index + offsetof(struct i965_sampler_8x8, dw1), |
||
2305 | pp_context->sampler_state_table.bo_8x8); |
||
2306 | |||
2307 | dri_bo_unmap(pp_context->sampler_state_table.bo); |
||
2308 | |||
2309 | /* private function & data */ |
||
2310 | pp_context->pp_x_steps = pp_avs_x_steps; |
||
2311 | pp_context->pp_y_steps = pp_avs_y_steps; |
||
2312 | pp_context->pp_set_block_parameter = pp_avs_set_block_parameter; |
||
2313 | |||
2314 | int dst_left_edge_extend = dst_rect->x%GPU_ASM_X_OFFSET_ALIGNMENT; |
||
2315 | float src_left_edge_extend = (float)dst_left_edge_extend*src_rect->width/dst_rect->width; |
||
2316 | pp_avs_context->dest_x = dst_rect->x - dst_left_edge_extend; |
||
2317 | pp_avs_context->dest_y = dst_rect->y; |
||
2318 | pp_avs_context->dest_w = ALIGN(dst_rect->width + dst_left_edge_extend, 16); |
||
2319 | pp_avs_context->dest_h = ALIGN(dst_rect->height, 8); |
||
2320 | pp_avs_context->src_normalized_x = (float)(src_rect->x - src_left_edge_extend)/ in_w; |
||
2321 | pp_avs_context->src_normalized_y = (float)src_rect->y / in_h; |
||
2322 | pp_avs_context->src_w = src_rect->width + src_left_edge_extend; |
||
2323 | pp_avs_context->src_h = src_rect->height; |
||
2324 | |||
2325 | pp_static_parameter->grf4.r4_2.avs.nlas = nlas; |
||
2326 | pp_static_parameter->grf1.r1_6.normalized_video_y_scaling_step = (float) src_rect->height / in_h / dst_rect->height; |
||
2327 | |||
2328 | pp_inline_parameter->grf5.normalized_video_x_scaling_step = (float) (src_rect->width + src_left_edge_extend)/ in_w / (dst_rect->width + dst_left_edge_extend); |
||
2329 | pp_inline_parameter->grf5.block_count_x = 1; /* M x 1 */ |
||
2330 | pp_inline_parameter->grf5.number_blocks = pp_avs_context->dest_h / 8; |
||
2331 | pp_inline_parameter->grf6.video_step_delta = 0.0; |
||
2332 | |||
2333 | dst_surface->flags = src_surface->flags; |
||
2334 | |||
2335 | return VA_STATUS_SUCCESS; |
||
2336 | } |
||
2337 | |||
2338 | static VAStatus |
||
2339 | pp_nv12_avs_initialize_nlas(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
2340 | const struct i965_surface *src_surface, |
||
2341 | const VARectangle *src_rect, |
||
2342 | struct i965_surface *dst_surface, |
||
2343 | const VARectangle *dst_rect, |
||
2344 | void *filter_param) |
||
2345 | { |
||
2346 | return pp_nv12_avs_initialize(ctx, pp_context, |
||
2347 | src_surface, |
||
2348 | src_rect, |
||
2349 | dst_surface, |
||
2350 | dst_rect, |
||
2351 | filter_param, |
||
2352 | 1); |
||
2353 | } |
||
2354 | |||
2355 | static VAStatus |
||
2356 | gen6_nv12_scaling_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
2357 | const struct i965_surface *src_surface, |
||
2358 | const VARectangle *src_rect, |
||
2359 | struct i965_surface *dst_surface, |
||
2360 | const VARectangle *dst_rect, |
||
2361 | void *filter_param) |
||
2362 | { |
||
2363 | return pp_nv12_avs_initialize(ctx, pp_context, |
||
2364 | src_surface, |
||
2365 | src_rect, |
||
2366 | dst_surface, |
||
2367 | dst_rect, |
||
2368 | filter_param, |
||
2369 | 0); |
||
2370 | } |
||
2371 | |||
2372 | static int |
||
2373 | gen7_pp_avs_x_steps(void *private_context) |
||
2374 | { |
||
2375 | struct pp_avs_context *pp_avs_context = private_context; |
||
2376 | |||
2377 | return pp_avs_context->dest_w / 16; |
||
2378 | } |
||
2379 | |||
2380 | static int |
||
2381 | gen7_pp_avs_y_steps(void *private_context) |
||
2382 | { |
||
2383 | struct pp_avs_context *pp_avs_context = private_context; |
||
2384 | |||
2385 | return pp_avs_context->dest_h / 16; |
||
2386 | } |
||
2387 | |||
2388 | static int |
||
2389 | gen7_pp_avs_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
2390 | { |
||
2391 | struct pp_avs_context *pp_avs_context = (struct pp_avs_context *)&pp_context->private_context; |
||
2392 | struct gen7_pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
2393 | |||
2394 | pp_inline_parameter->grf7.destination_block_horizontal_origin = x * 16 + pp_avs_context->dest_x; |
||
2395 | pp_inline_parameter->grf7.destination_block_vertical_origin = y * 16 + pp_avs_context->dest_y; |
||
2396 | pp_inline_parameter->grf7.constant_0 = 0xffffffff; |
||
2397 | pp_inline_parameter->grf7.sampler_load_main_video_x_scaling_step = 1.0 / pp_avs_context->src_w; |
||
2398 | |||
2399 | return 0; |
||
2400 | } |
||
2401 | |||
2402 | static void gen7_update_src_surface_uv_offset(VADriverContextP ctx, |
||
2403 | struct i965_post_processing_context *pp_context, |
||
2404 | const struct i965_surface *surface) |
||
2405 | { |
||
2406 | struct gen7_pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
2407 | int fourcc = pp_get_surface_fourcc(ctx, surface); |
||
2408 | |||
2409 | if (fourcc == VA_FOURCC('Y', 'U', 'Y', '2')) { |
||
2410 | pp_static_parameter->grf2.di_destination_packed_y_component_offset = 0; |
||
2411 | pp_static_parameter->grf2.di_destination_packed_u_component_offset = 1; |
||
2412 | pp_static_parameter->grf2.di_destination_packed_v_component_offset = 3; |
||
2413 | } else if (fourcc == VA_FOURCC('U', 'Y', 'V', 'Y')) { |
||
2414 | pp_static_parameter->grf2.di_destination_packed_y_component_offset = 1; |
||
2415 | pp_static_parameter->grf2.di_destination_packed_u_component_offset = 0; |
||
2416 | pp_static_parameter->grf2.di_destination_packed_v_component_offset = 2; |
||
2417 | } |
||
2418 | } |
||
2419 | |||
2420 | static VAStatus |
||
2421 | gen7_pp_plx_avs_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
2422 | const struct i965_surface *src_surface, |
||
2423 | const VARectangle *src_rect, |
||
2424 | struct i965_surface *dst_surface, |
||
2425 | const VARectangle *dst_rect, |
||
2426 | void *filter_param) |
||
2427 | { |
||
2428 | struct pp_avs_context *pp_avs_context = (struct pp_avs_context *)&pp_context->private_context; |
||
2429 | struct gen7_pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
2430 | struct gen7_sampler_8x8 *sampler_8x8; |
||
2431 | struct i965_sampler_8x8_state *sampler_8x8_state; |
||
2432 | int index, i; |
||
2433 | int width[3], height[3], pitch[3], offset[3]; |
||
2434 | int src_height; |
||
2435 | |||
2436 | /* source surface */ |
||
2437 | gen7_pp_set_media_rw_message_surface(ctx, pp_context, src_surface, 0, 0, |
||
2438 | width, height, pitch, offset); |
||
2439 | src_height = height[0]; |
||
2440 | |||
2441 | /* destination surface */ |
||
2442 | gen7_pp_set_media_rw_message_surface(ctx, pp_context, dst_surface, 24, 1, |
||
2443 | width, height, pitch, offset); |
||
2444 | |||
2445 | /* sampler 8x8 state */ |
||
2446 | dri_bo_map(pp_context->sampler_state_table.bo_8x8, True); |
||
2447 | assert(pp_context->sampler_state_table.bo_8x8->virtual); |
||
2448 | assert(sizeof(*sampler_8x8_state) == sizeof(int) * 138); |
||
2449 | sampler_8x8_state = pp_context->sampler_state_table.bo_8x8->virtual; |
||
2450 | memset(sampler_8x8_state, 0, sizeof(*sampler_8x8_state)); |
||
2451 | |||
2452 | for (i = 0; i < 17; i++) { |
||
2453 | /* for Y channel, currently ignore */ |
||
2454 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c0 = 0x0; |
||
2455 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c1 = 0x0; |
||
2456 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c2 = 0x0; |
||
2457 | sampler_8x8_state->coefficients[i].dw0.table_0x_filter_c3 = 0x0; |
||
2458 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c4 = 0x0; |
||
2459 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c5 = 0x0; |
||
2460 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c6 = 0x0; |
||
2461 | sampler_8x8_state->coefficients[i].dw1.table_0x_filter_c7 = 0x0; |
||
2462 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c0 = 0x0; |
||
2463 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c1 = 0x0; |
||
2464 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c2 = 0x0; |
||
2465 | sampler_8x8_state->coefficients[i].dw2.table_0y_filter_c3 = 0x0; |
||
2466 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c4 = 0x0; |
||
2467 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c5 = 0x0; |
||
2468 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c6 = 0x0; |
||
2469 | sampler_8x8_state->coefficients[i].dw3.table_0y_filter_c7 = 0x0; |
||
2470 | /* for U/V channel, 0.25 */ |
||
2471 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c0 = 0x0; |
||
2472 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c1 = 0x0; |
||
2473 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c2 = 0x10; |
||
2474 | sampler_8x8_state->coefficients[i].dw4.table_1x_filter_c3 = 0x10; |
||
2475 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c4 = 0x10; |
||
2476 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c5 = 0x10; |
||
2477 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c6 = 0x0; |
||
2478 | sampler_8x8_state->coefficients[i].dw5.table_1x_filter_c7 = 0x0; |
||
2479 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c0 = 0x0; |
||
2480 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c1 = 0x0; |
||
2481 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c2 = 0x10; |
||
2482 | sampler_8x8_state->coefficients[i].dw6.table_1y_filter_c3 = 0x10; |
||
2483 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c4 = 0x10; |
||
2484 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c5 = 0x10; |
||
2485 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c6 = 0x0; |
||
2486 | sampler_8x8_state->coefficients[i].dw7.table_1y_filter_c7 = 0x0; |
||
2487 | } |
||
2488 | |||
2489 | sampler_8x8_state->dw136.default_sharpness_level = 0; |
||
2490 | sampler_8x8_state->dw137.adaptive_filter_for_all_channel = 1; |
||
2491 | sampler_8x8_state->dw137.bypass_y_adaptive_filtering = 1; |
||
2492 | sampler_8x8_state->dw137.bypass_x_adaptive_filtering = 1; |
||
2493 | dri_bo_unmap(pp_context->sampler_state_table.bo_8x8); |
||
2494 | |||
2495 | /* sampler 8x8 */ |
||
2496 | dri_bo_map(pp_context->sampler_state_table.bo, True); |
||
2497 | assert(pp_context->sampler_state_table.bo->virtual); |
||
2498 | assert(sizeof(*sampler_8x8) == sizeof(int) * 4); |
||
2499 | sampler_8x8 = pp_context->sampler_state_table.bo->virtual; |
||
2500 | |||
2501 | /* sample_8x8 Y index 4 */ |
||
2502 | index = 4; |
||
2503 | memset(&sampler_8x8[index], 0, sizeof(*sampler_8x8)); |
||
2504 | sampler_8x8[index].dw0.global_noise_estimation = 255; |
||
2505 | sampler_8x8[index].dw0.ief_bypass = 1; |
||
2506 | |||
2507 | sampler_8x8[index].dw1.sampler_8x8_state_pointer = pp_context->sampler_state_table.bo_8x8->offset >> 5; |
||
2508 | |||
2509 | sampler_8x8[index].dw2.weak_edge_threshold = 1; |
||
2510 | sampler_8x8[index].dw2.strong_edge_threshold = 8; |
||
2511 | sampler_8x8[index].dw2.r5x_coefficient = 9; |
||
2512 | sampler_8x8[index].dw2.r5cx_coefficient = 8; |
||
2513 | sampler_8x8[index].dw2.r5c_coefficient = 3; |
||
2514 | |||
2515 | sampler_8x8[index].dw3.r3x_coefficient = 27; |
||
2516 | sampler_8x8[index].dw3.r3c_coefficient = 5; |
||
2517 | sampler_8x8[index].dw3.gain_factor = 40; |
||
2518 | sampler_8x8[index].dw3.non_edge_weight = 1; |
||
2519 | sampler_8x8[index].dw3.regular_weight = 2; |
||
2520 | sampler_8x8[index].dw3.strong_edge_weight = 7; |
||
2521 | sampler_8x8[index].dw3.ief4_smooth_enable = 0; |
||
2522 | |||
2523 | dri_bo_emit_reloc(pp_context->sampler_state_table.bo, |
||
2524 | I915_GEM_DOMAIN_RENDER, |
||
2525 | 0, |
||
2526 | 0, |
||
2527 | sizeof(*sampler_8x8) * index + offsetof(struct i965_sampler_8x8, dw1), |
||
2528 | pp_context->sampler_state_table.bo_8x8); |
||
2529 | |||
2530 | /* sample_8x8 UV index 8 */ |
||
2531 | index = 8; |
||
2532 | memset(&sampler_8x8[index], 0, sizeof(*sampler_8x8)); |
||
2533 | sampler_8x8[index].dw0.disable_8x8_filter = 0; |
||
2534 | sampler_8x8[index].dw0.global_noise_estimation = 255; |
||
2535 | sampler_8x8[index].dw0.ief_bypass = 1; |
||
2536 | sampler_8x8[index].dw1.sampler_8x8_state_pointer = pp_context->sampler_state_table.bo_8x8->offset >> 5; |
||
2537 | sampler_8x8[index].dw2.weak_edge_threshold = 1; |
||
2538 | sampler_8x8[index].dw2.strong_edge_threshold = 8; |
||
2539 | sampler_8x8[index].dw2.r5x_coefficient = 9; |
||
2540 | sampler_8x8[index].dw2.r5cx_coefficient = 8; |
||
2541 | sampler_8x8[index].dw2.r5c_coefficient = 3; |
||
2542 | sampler_8x8[index].dw3.r3x_coefficient = 27; |
||
2543 | sampler_8x8[index].dw3.r3c_coefficient = 5; |
||
2544 | sampler_8x8[index].dw3.gain_factor = 40; |
||
2545 | sampler_8x8[index].dw3.non_edge_weight = 1; |
||
2546 | sampler_8x8[index].dw3.regular_weight = 2; |
||
2547 | sampler_8x8[index].dw3.strong_edge_weight = 7; |
||
2548 | sampler_8x8[index].dw3.ief4_smooth_enable = 0; |
||
2549 | |||
2550 | dri_bo_emit_reloc(pp_context->sampler_state_table.bo, |
||
2551 | I915_GEM_DOMAIN_RENDER, |
||
2552 | 0, |
||
2553 | 0, |
||
2554 | sizeof(*sampler_8x8) * index + offsetof(struct i965_sampler_8x8, dw1), |
||
2555 | pp_context->sampler_state_table.bo_8x8); |
||
2556 | |||
2557 | /* sampler_8x8 V, index 12 */ |
||
2558 | index = 12; |
||
2559 | memset(&sampler_8x8[index], 0, sizeof(*sampler_8x8)); |
||
2560 | sampler_8x8[index].dw0.disable_8x8_filter = 0; |
||
2561 | sampler_8x8[index].dw0.global_noise_estimation = 255; |
||
2562 | sampler_8x8[index].dw0.ief_bypass = 1; |
||
2563 | sampler_8x8[index].dw1.sampler_8x8_state_pointer = pp_context->sampler_state_table.bo_8x8->offset >> 5; |
||
2564 | sampler_8x8[index].dw2.weak_edge_threshold = 1; |
||
2565 | sampler_8x8[index].dw2.strong_edge_threshold = 8; |
||
2566 | sampler_8x8[index].dw2.r5x_coefficient = 9; |
||
2567 | sampler_8x8[index].dw2.r5cx_coefficient = 8; |
||
2568 | sampler_8x8[index].dw2.r5c_coefficient = 3; |
||
2569 | sampler_8x8[index].dw3.r3x_coefficient = 27; |
||
2570 | sampler_8x8[index].dw3.r3c_coefficient = 5; |
||
2571 | sampler_8x8[index].dw3.gain_factor = 40; |
||
2572 | sampler_8x8[index].dw3.non_edge_weight = 1; |
||
2573 | sampler_8x8[index].dw3.regular_weight = 2; |
||
2574 | sampler_8x8[index].dw3.strong_edge_weight = 7; |
||
2575 | sampler_8x8[index].dw3.ief4_smooth_enable = 0; |
||
2576 | |||
2577 | dri_bo_emit_reloc(pp_context->sampler_state_table.bo, |
||
2578 | I915_GEM_DOMAIN_RENDER, |
||
2579 | 0, |
||
2580 | 0, |
||
2581 | sizeof(*sampler_8x8) * index + offsetof(struct i965_sampler_8x8, dw1), |
||
2582 | pp_context->sampler_state_table.bo_8x8); |
||
2583 | |||
2584 | dri_bo_unmap(pp_context->sampler_state_table.bo); |
||
2585 | |||
2586 | /* private function & data */ |
||
2587 | pp_context->pp_x_steps = gen7_pp_avs_x_steps; |
||
2588 | pp_context->pp_y_steps = gen7_pp_avs_y_steps; |
||
2589 | pp_context->pp_set_block_parameter = gen7_pp_avs_set_block_parameter; |
||
2590 | |||
2591 | pp_avs_context->dest_x = dst_rect->x; |
||
2592 | pp_avs_context->dest_y = dst_rect->y; |
||
2593 | pp_avs_context->dest_w = ALIGN(dst_rect->width, 16); |
||
2594 | pp_avs_context->dest_h = ALIGN(dst_rect->height, 16); |
||
2595 | pp_avs_context->src_w = src_rect->width; |
||
2596 | pp_avs_context->src_h = src_rect->height; |
||
2597 | |||
2598 | int dw = (pp_avs_context->src_w - 1) / 16 + 1; |
||
2599 | dw = MAX(dw, pp_avs_context->dest_w); |
||
2600 | |||
2601 | pp_static_parameter->grf1.pointer_to_inline_parameter = 7; |
||
2602 | pp_static_parameter->grf2.avs_wa_enable = 1; /* must be set for GEN7 */ |
||
2603 | pp_static_parameter->grf2.avs_wa_width = dw; |
||
2604 | pp_static_parameter->grf2.avs_wa_one_div_256_width = (float) 1.0 / (256 * dw); |
||
2605 | pp_static_parameter->grf2.avs_wa_five_div_256_width = (float) 5.0 / (256 * dw); |
||
2606 | |||
2607 | pp_static_parameter->grf3.sampler_load_horizontal_scaling_step_ratio = (float) pp_avs_context->src_w / pp_avs_context->dest_w; |
||
2608 | pp_static_parameter->grf4.sampler_load_vertical_scaling_step = (float) src_rect->height / src_height / pp_avs_context->dest_h; |
||
2609 | pp_static_parameter->grf5.sampler_load_vertical_frame_origin = -(float)pp_avs_context->dest_y / pp_avs_context->dest_h; |
||
2610 | pp_static_parameter->grf6.sampler_load_horizontal_frame_origin = -(float)pp_avs_context->dest_x / pp_avs_context->dest_w; |
||
2611 | |||
2612 | gen7_update_src_surface_uv_offset(ctx, pp_context, dst_surface); |
||
2613 | |||
2614 | dst_surface->flags = src_surface->flags; |
||
2615 | |||
2616 | return VA_STATUS_SUCCESS; |
||
2617 | } |
||
2618 | |||
2619 | |||
2620 | static int |
||
2621 | pp_dndi_x_steps(void *private_context) |
||
2622 | { |
||
2623 | return 1; |
||
2624 | } |
||
2625 | |||
2626 | static int |
||
2627 | pp_dndi_y_steps(void *private_context) |
||
2628 | { |
||
2629 | struct pp_dndi_context *pp_dndi_context = private_context; |
||
2630 | |||
2631 | return pp_dndi_context->dest_h / 4; |
||
2632 | } |
||
2633 | |||
2634 | static int |
||
2635 | pp_dndi_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
2636 | { |
||
2637 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
2638 | |||
2639 | pp_inline_parameter->grf5.destination_block_horizontal_origin = x * 16; |
||
2640 | pp_inline_parameter->grf5.destination_block_vertical_origin = y * 4; |
||
2641 | |||
2642 | return 0; |
||
2643 | } |
||
2644 | |||
2645 | static |
||
2646 | VAStatus pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
2647 | const struct i965_surface *src_surface, |
||
2648 | const VARectangle *src_rect, |
||
2649 | struct i965_surface *dst_surface, |
||
2650 | const VARectangle *dst_rect, |
||
2651 | void *filter_param) |
||
2652 | { |
||
2653 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
2654 | struct pp_dndi_context *pp_dndi_context = (struct pp_dndi_context *)&pp_context->private_context; |
||
2655 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
2656 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
2657 | struct object_surface *obj_surface; |
||
2658 | struct i965_sampler_dndi *sampler_dndi; |
||
2659 | int index; |
||
2660 | int w, h; |
||
2661 | int orig_w, orig_h; |
||
2662 | int dndi_top_first = 1; |
||
2663 | |||
2664 | if (src_surface->flags == I965_SURFACE_FLAG_FRAME) |
||
2665 | return VA_STATUS_ERROR_FLAG_NOT_SUPPORTED; |
||
2666 | |||
2667 | if (src_surface->flags == I965_SURFACE_FLAG_TOP_FIELD_FIRST) |
||
2668 | dndi_top_first = 1; |
||
2669 | else |
||
2670 | dndi_top_first = 0; |
||
2671 | |||
2672 | /* surface */ |
||
2673 | obj_surface = SURFACE(src_surface->id); |
||
2674 | orig_w = obj_surface->orig_width; |
||
2675 | orig_h = obj_surface->orig_height; |
||
2676 | w = obj_surface->width; |
||
2677 | h = obj_surface->height; |
||
2678 | |||
2679 | if (pp_context->stmm.bo == NULL) { |
||
2680 | pp_context->stmm.bo = dri_bo_alloc(i965->intel.bufmgr, |
||
2681 | "STMM surface", |
||
2682 | w * h, |
||
2683 | 4096); |
||
2684 | assert(pp_context->stmm.bo); |
||
2685 | } |
||
2686 | |||
2687 | /* source UV surface index 2 */ |
||
2688 | i965_pp_set_surface_state(ctx, pp_context, |
||
2689 | obj_surface->bo, w * h, |
||
2690 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
2691 | 2, 0); |
||
2692 | |||
2693 | /* source YUV surface index 4 */ |
||
2694 | i965_pp_set_surface2_state(ctx, pp_context, |
||
2695 | obj_surface->bo, 0, |
||
2696 | orig_w, orig_h, w, |
||
2697 | 0, h, |
||
2698 | SURFACE_FORMAT_PLANAR_420_8, 1, |
||
2699 | 4); |
||
2700 | |||
2701 | /* source STMM surface index 20 */ |
||
2702 | i965_pp_set_surface_state(ctx, pp_context, |
||
2703 | pp_context->stmm.bo, 0, |
||
2704 | orig_w, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
2705 | 20, 1); |
||
2706 | |||
2707 | /* destination surface */ |
||
2708 | obj_surface = SURFACE(dst_surface->id); |
||
2709 | orig_w = obj_surface->orig_width; |
||
2710 | orig_h = obj_surface->orig_height; |
||
2711 | w = obj_surface->width; |
||
2712 | h = obj_surface->height; |
||
2713 | |||
2714 | /* destination Y surface index 7 */ |
||
2715 | i965_pp_set_surface_state(ctx, pp_context, |
||
2716 | obj_surface->bo, 0, |
||
2717 | orig_w / 4, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
2718 | 7, 1); |
||
2719 | |||
2720 | /* destination UV surface index 8 */ |
||
2721 | i965_pp_set_surface_state(ctx, pp_context, |
||
2722 | obj_surface->bo, w * h, |
||
2723 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
2724 | 8, 1); |
||
2725 | /* sampler dndi */ |
||
2726 | dri_bo_map(pp_context->sampler_state_table.bo, True); |
||
2727 | assert(pp_context->sampler_state_table.bo->virtual); |
||
2728 | assert(sizeof(*sampler_dndi) == sizeof(int) * 8); |
||
2729 | sampler_dndi = pp_context->sampler_state_table.bo->virtual; |
||
2730 | |||
2731 | /* sample dndi index 1 */ |
||
2732 | index = 0; |
||
2733 | sampler_dndi[index].dw0.denoise_asd_threshold = 0; |
||
2734 | sampler_dndi[index].dw0.denoise_history_delta = 8; // 0-15, default is 8 |
||
2735 | sampler_dndi[index].dw0.denoise_maximum_history = 128; // 128-240 |
||
2736 | sampler_dndi[index].dw0.denoise_stad_threshold = 0; |
||
2737 | |||
2738 | sampler_dndi[index].dw1.denoise_threshold_for_sum_of_complexity_measure = 64; |
||
2739 | sampler_dndi[index].dw1.denoise_moving_pixel_threshold = 4; |
||
2740 | sampler_dndi[index].dw1.stmm_c2 = 1; |
||
2741 | sampler_dndi[index].dw1.low_temporal_difference_threshold = 8; |
||
2742 | sampler_dndi[index].dw1.temporal_difference_threshold = 16; |
||
2743 | |||
2744 | sampler_dndi[index].dw2.block_noise_estimate_noise_threshold = 15; // 0-31 |
||
2745 | sampler_dndi[index].dw2.block_noise_estimate_edge_threshold = 7; // 0-15 |
||
2746 | sampler_dndi[index].dw2.denoise_edge_threshold = 7; // 0-15 |
||
2747 | sampler_dndi[index].dw2.good_neighbor_threshold = 4; // 0-63 |
||
2748 | |||
2749 | sampler_dndi[index].dw3.maximum_stmm = 128; |
||
2750 | sampler_dndi[index].dw3.multipler_for_vecm = 2; |
||
2751 | sampler_dndi[index].dw3.blending_constant_across_time_for_small_values_of_stmm = 0; |
||
2752 | sampler_dndi[index].dw3.blending_constant_across_time_for_large_values_of_stmm = 64; |
||
2753 | sampler_dndi[index].dw3.stmm_blending_constant_select = 0; |
||
2754 | |||
2755 | sampler_dndi[index].dw4.sdi_delta = 8; |
||
2756 | sampler_dndi[index].dw4.sdi_threshold = 128; |
||
2757 | sampler_dndi[index].dw4.stmm_output_shift = 7; // stmm_max - stmm_min = 2 ^ stmm_output_shift |
||
2758 | sampler_dndi[index].dw4.stmm_shift_up = 0; |
||
2759 | sampler_dndi[index].dw4.stmm_shift_down = 0; |
||
2760 | sampler_dndi[index].dw4.minimum_stmm = 0; |
||
2761 | |||
2762 | sampler_dndi[index].dw5.fmd_temporal_difference_threshold = 8; |
||
2763 | sampler_dndi[index].dw5.sdi_fallback_mode_2_constant = 32; |
||
2764 | sampler_dndi[index].dw5.sdi_fallback_mode_1_t2_constant = 64; |
||
2765 | sampler_dndi[index].dw5.sdi_fallback_mode_1_t1_constant = 32; |
||
2766 | |||
2767 | sampler_dndi[index].dw6.dn_enable = 1; |
||
2768 | sampler_dndi[index].dw6.di_enable = 1; |
||
2769 | sampler_dndi[index].dw6.di_partial = 0; |
||
2770 | sampler_dndi[index].dw6.dndi_top_first = dndi_top_first; |
||
2771 | sampler_dndi[index].dw6.dndi_stream_id = 0; |
||
2772 | sampler_dndi[index].dw6.dndi_first_frame = 1; |
||
2773 | sampler_dndi[index].dw6.progressive_dn = 0; |
||
2774 | sampler_dndi[index].dw6.fmd_tear_threshold = 63; |
||
2775 | sampler_dndi[index].dw6.fmd2_vertical_difference_threshold = 32; |
||
2776 | sampler_dndi[index].dw6.fmd1_vertical_difference_threshold = 32; |
||
2777 | |||
2778 | sampler_dndi[index].dw7.fmd_for_1st_field_of_current_frame = 0; |
||
2779 | sampler_dndi[index].dw7.fmd_for_2nd_field_of_previous_frame = 0; |
||
2780 | sampler_dndi[index].dw7.vdi_walker_enable = 0; |
||
2781 | sampler_dndi[index].dw7.column_width_minus1 = 0; |
||
2782 | |||
2783 | dri_bo_unmap(pp_context->sampler_state_table.bo); |
||
2784 | |||
2785 | /* private function & data */ |
||
2786 | pp_context->pp_x_steps = pp_dndi_x_steps; |
||
2787 | pp_context->pp_y_steps = pp_dndi_y_steps; |
||
2788 | pp_context->pp_set_block_parameter = pp_dndi_set_block_parameter; |
||
2789 | |||
2790 | pp_static_parameter->grf1.statistics_surface_picth = w / 2; |
||
2791 | pp_static_parameter->grf1.r1_6.di.top_field_first = dndi_top_first; |
||
2792 | pp_static_parameter->grf4.r4_2.di.motion_history_coefficient_m2 = 0; |
||
2793 | pp_static_parameter->grf4.r4_2.di.motion_history_coefficient_m1 = 0; |
||
2794 | |||
2795 | pp_inline_parameter->grf5.block_count_x = w / 16; /* 1 x N */ |
||
2796 | pp_inline_parameter->grf5.number_blocks = w / 16; |
||
2797 | pp_inline_parameter->grf5.block_vertical_mask = 0xff; |
||
2798 | pp_inline_parameter->grf5.block_horizontal_mask = 0xffff; |
||
2799 | |||
2800 | pp_dndi_context->dest_w = w; |
||
2801 | pp_dndi_context->dest_h = h; |
||
2802 | |||
2803 | dst_surface->flags = I965_SURFACE_FLAG_FRAME; |
||
2804 | |||
2805 | return VA_STATUS_SUCCESS; |
||
2806 | } |
||
2807 | |||
2808 | static int |
||
2809 | pp_dn_x_steps(void *private_context) |
||
2810 | { |
||
2811 | return 1; |
||
2812 | } |
||
2813 | |||
2814 | static int |
||
2815 | pp_dn_y_steps(void *private_context) |
||
2816 | { |
||
2817 | struct pp_dn_context *pp_dn_context = private_context; |
||
2818 | |||
2819 | return pp_dn_context->dest_h / 8; |
||
2820 | } |
||
2821 | |||
2822 | static int |
||
2823 | pp_dn_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
2824 | { |
||
2825 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
2826 | |||
2827 | pp_inline_parameter->grf5.destination_block_horizontal_origin = x * 16; |
||
2828 | pp_inline_parameter->grf5.destination_block_vertical_origin = y * 8; |
||
2829 | |||
2830 | return 0; |
||
2831 | } |
||
2832 | |||
2833 | static |
||
2834 | VAStatus pp_nv12_dn_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
2835 | const struct i965_surface *src_surface, |
||
2836 | const VARectangle *src_rect, |
||
2837 | struct i965_surface *dst_surface, |
||
2838 | const VARectangle *dst_rect, |
||
2839 | void *filter_param) |
||
2840 | { |
||
2841 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
2842 | struct pp_dn_context *pp_dn_context = (struct pp_dn_context *)&pp_context->private_context; |
||
2843 | struct object_surface *obj_surface; |
||
2844 | struct i965_sampler_dndi *sampler_dndi; |
||
2845 | struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
2846 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
2847 | |||
2848 | int index; |
||
2849 | int w, h; |
||
2850 | int orig_w, orig_h; |
||
2851 | int dn_strength = 15; |
||
2852 | int dndi_top_first = 1; |
||
2853 | int dn_progressive = 0; |
||
2854 | |||
2855 | if (src_surface->flags == I965_SURFACE_FLAG_FRAME) { |
||
2856 | dndi_top_first = 1; |
||
2857 | dn_progressive = 1; |
||
2858 | } else if (src_surface->flags == I965_SURFACE_FLAG_TOP_FIELD_FIRST) { |
||
2859 | dndi_top_first = 1; |
||
2860 | dn_progressive = 0; |
||
2861 | } else { |
||
2862 | dndi_top_first = 0; |
||
2863 | dn_progressive = 0; |
||
2864 | } |
||
2865 | |||
2866 | /* surface */ |
||
2867 | obj_surface = SURFACE(src_surface->id); |
||
2868 | orig_w = obj_surface->orig_width; |
||
2869 | orig_h = obj_surface->orig_height; |
||
2870 | w = obj_surface->width; |
||
2871 | h = obj_surface->height; |
||
2872 | |||
2873 | if (pp_context->stmm.bo == NULL) { |
||
2874 | pp_context->stmm.bo = dri_bo_alloc(i965->intel.bufmgr, |
||
2875 | "STMM surface", |
||
2876 | w * h, |
||
2877 | 4096); |
||
2878 | assert(pp_context->stmm.bo); |
||
2879 | } |
||
2880 | |||
2881 | /* source UV surface index 2 */ |
||
2882 | i965_pp_set_surface_state(ctx, pp_context, |
||
2883 | obj_surface->bo, w * h, |
||
2884 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
2885 | 2, 0); |
||
2886 | |||
2887 | /* source YUV surface index 4 */ |
||
2888 | i965_pp_set_surface2_state(ctx, pp_context, |
||
2889 | obj_surface->bo, 0, |
||
2890 | orig_w, orig_h, w, |
||
2891 | 0, h, |
||
2892 | SURFACE_FORMAT_PLANAR_420_8, 1, |
||
2893 | 4); |
||
2894 | |||
2895 | /* source STMM surface index 20 */ |
||
2896 | i965_pp_set_surface_state(ctx, pp_context, |
||
2897 | pp_context->stmm.bo, 0, |
||
2898 | orig_w, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
2899 | 20, 1); |
||
2900 | |||
2901 | /* destination surface */ |
||
2902 | obj_surface = SURFACE(dst_surface->id); |
||
2903 | orig_w = obj_surface->orig_width; |
||
2904 | orig_h = obj_surface->orig_height; |
||
2905 | w = obj_surface->width; |
||
2906 | h = obj_surface->height; |
||
2907 | |||
2908 | /* destination Y surface index 7 */ |
||
2909 | i965_pp_set_surface_state(ctx, pp_context, |
||
2910 | obj_surface->bo, 0, |
||
2911 | orig_w / 4, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
2912 | 7, 1); |
||
2913 | |||
2914 | /* destination UV surface index 8 */ |
||
2915 | i965_pp_set_surface_state(ctx, pp_context, |
||
2916 | obj_surface->bo, w * h, |
||
2917 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
2918 | 8, 1); |
||
2919 | /* sampler dn */ |
||
2920 | dri_bo_map(pp_context->sampler_state_table.bo, True); |
||
2921 | assert(pp_context->sampler_state_table.bo->virtual); |
||
2922 | assert(sizeof(*sampler_dndi) == sizeof(int) * 8); |
||
2923 | sampler_dndi = pp_context->sampler_state_table.bo->virtual; |
||
2924 | |||
2925 | /* sample dndi index 1 */ |
||
2926 | index = 0; |
||
2927 | sampler_dndi[index].dw0.denoise_asd_threshold = 0; |
||
2928 | sampler_dndi[index].dw0.denoise_history_delta = 8; // 0-15, default is 8 |
||
2929 | sampler_dndi[index].dw0.denoise_maximum_history = 128; // 128-240 |
||
2930 | sampler_dndi[index].dw0.denoise_stad_threshold = 0; |
||
2931 | |||
2932 | sampler_dndi[index].dw1.denoise_threshold_for_sum_of_complexity_measure = 64; |
||
2933 | sampler_dndi[index].dw1.denoise_moving_pixel_threshold = 0; |
||
2934 | sampler_dndi[index].dw1.stmm_c2 = 0; |
||
2935 | sampler_dndi[index].dw1.low_temporal_difference_threshold = 8; |
||
2936 | sampler_dndi[index].dw1.temporal_difference_threshold = 16; |
||
2937 | |||
2938 | sampler_dndi[index].dw2.block_noise_estimate_noise_threshold = dn_strength; // 0-31 |
||
2939 | sampler_dndi[index].dw2.block_noise_estimate_edge_threshold = 7; // 0-15 |
||
2940 | sampler_dndi[index].dw2.denoise_edge_threshold = 7; // 0-15 |
||
2941 | sampler_dndi[index].dw2.good_neighbor_threshold = 7; // 0-63 |
||
2942 | |||
2943 | sampler_dndi[index].dw3.maximum_stmm = 128; |
||
2944 | sampler_dndi[index].dw3.multipler_for_vecm = 2; |
||
2945 | sampler_dndi[index].dw3.blending_constant_across_time_for_small_values_of_stmm = 0; |
||
2946 | sampler_dndi[index].dw3.blending_constant_across_time_for_large_values_of_stmm = 64; |
||
2947 | sampler_dndi[index].dw3.stmm_blending_constant_select = 0; |
||
2948 | |||
2949 | sampler_dndi[index].dw4.sdi_delta = 8; |
||
2950 | sampler_dndi[index].dw4.sdi_threshold = 128; |
||
2951 | sampler_dndi[index].dw4.stmm_output_shift = 7; // stmm_max - stmm_min = 2 ^ stmm_output_shift |
||
2952 | sampler_dndi[index].dw4.stmm_shift_up = 0; |
||
2953 | sampler_dndi[index].dw4.stmm_shift_down = 0; |
||
2954 | sampler_dndi[index].dw4.minimum_stmm = 0; |
||
2955 | |||
2956 | sampler_dndi[index].dw5.fmd_temporal_difference_threshold = 0; |
||
2957 | sampler_dndi[index].dw5.sdi_fallback_mode_2_constant = 0; |
||
2958 | sampler_dndi[index].dw5.sdi_fallback_mode_1_t2_constant = 0; |
||
2959 | sampler_dndi[index].dw5.sdi_fallback_mode_1_t1_constant = 0; |
||
2960 | |||
2961 | sampler_dndi[index].dw6.dn_enable = 1; |
||
2962 | sampler_dndi[index].dw6.di_enable = 0; |
||
2963 | sampler_dndi[index].dw6.di_partial = 0; |
||
2964 | sampler_dndi[index].dw6.dndi_top_first = dndi_top_first; |
||
2965 | sampler_dndi[index].dw6.dndi_stream_id = 1; |
||
2966 | sampler_dndi[index].dw6.dndi_first_frame = 1; |
||
2967 | sampler_dndi[index].dw6.progressive_dn = dn_progressive; |
||
2968 | sampler_dndi[index].dw6.fmd_tear_threshold = 32; |
||
2969 | sampler_dndi[index].dw6.fmd2_vertical_difference_threshold = 32; |
||
2970 | sampler_dndi[index].dw6.fmd1_vertical_difference_threshold = 32; |
||
2971 | |||
2972 | sampler_dndi[index].dw7.fmd_for_1st_field_of_current_frame = 2; |
||
2973 | sampler_dndi[index].dw7.fmd_for_2nd_field_of_previous_frame = 1; |
||
2974 | sampler_dndi[index].dw7.vdi_walker_enable = 0; |
||
2975 | sampler_dndi[index].dw7.column_width_minus1 = w / 16; |
||
2976 | |||
2977 | dri_bo_unmap(pp_context->sampler_state_table.bo); |
||
2978 | |||
2979 | /* private function & data */ |
||
2980 | pp_context->pp_x_steps = pp_dn_x_steps; |
||
2981 | pp_context->pp_y_steps = pp_dn_y_steps; |
||
2982 | pp_context->pp_set_block_parameter = pp_dn_set_block_parameter; |
||
2983 | |||
2984 | pp_static_parameter->grf1.statistics_surface_picth = w / 2; |
||
2985 | pp_static_parameter->grf1.r1_6.di.top_field_first = 0; |
||
2986 | pp_static_parameter->grf4.r4_2.di.motion_history_coefficient_m2 = 64; |
||
2987 | pp_static_parameter->grf4.r4_2.di.motion_history_coefficient_m1 = 192; |
||
2988 | |||
2989 | pp_inline_parameter->grf5.block_count_x = w / 16; /* 1 x N */ |
||
2990 | pp_inline_parameter->grf5.number_blocks = w / 16; |
||
2991 | pp_inline_parameter->grf5.block_vertical_mask = 0xff; |
||
2992 | pp_inline_parameter->grf5.block_horizontal_mask = 0xffff; |
||
2993 | |||
2994 | pp_dn_context->dest_w = w; |
||
2995 | pp_dn_context->dest_h = h; |
||
2996 | |||
2997 | return VA_STATUS_SUCCESS; |
||
2998 | } |
||
2999 | |||
3000 | static int |
||
3001 | gen7_pp_dndi_x_steps(void *private_context) |
||
3002 | { |
||
3003 | struct pp_dndi_context *pp_dndi_context = private_context; |
||
3004 | |||
3005 | return pp_dndi_context->dest_w / 16; |
||
3006 | } |
||
3007 | |||
3008 | static int |
||
3009 | gen7_pp_dndi_y_steps(void *private_context) |
||
3010 | { |
||
3011 | struct pp_dndi_context *pp_dndi_context = private_context; |
||
3012 | |||
3013 | return pp_dndi_context->dest_h / 4; |
||
3014 | } |
||
3015 | |||
3016 | static int |
||
3017 | gen7_pp_dndi_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
3018 | { |
||
3019 | struct gen7_pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
3020 | |||
3021 | pp_inline_parameter->grf7.destination_block_horizontal_origin = x * 16; |
||
3022 | pp_inline_parameter->grf7.destination_block_vertical_origin = y * 4; |
||
3023 | |||
3024 | return 0; |
||
3025 | } |
||
3026 | |||
3027 | static VAStatus |
||
3028 | gen7_pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
3029 | const struct i965_surface *src_surface, |
||
3030 | const VARectangle *src_rect, |
||
3031 | struct i965_surface *dst_surface, |
||
3032 | const VARectangle *dst_rect, |
||
3033 | void *filter_param) |
||
3034 | { |
||
3035 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
3036 | struct pp_dndi_context *pp_dndi_context = (struct pp_dndi_context *)&pp_context->private_context; |
||
3037 | struct gen7_pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
3038 | struct object_surface *obj_surface; |
||
3039 | struct gen7_sampler_dndi *sampler_dndi; |
||
3040 | int index; |
||
3041 | int w, h; |
||
3042 | int orig_w, orig_h; |
||
3043 | int dndi_top_first = 1; |
||
3044 | |||
3045 | if (src_surface->flags == I965_SURFACE_FLAG_FRAME) |
||
3046 | return VA_STATUS_ERROR_FLAG_NOT_SUPPORTED; |
||
3047 | |||
3048 | if (src_surface->flags == I965_SURFACE_FLAG_TOP_FIELD_FIRST) |
||
3049 | dndi_top_first = 1; |
||
3050 | else |
||
3051 | dndi_top_first = 0; |
||
3052 | |||
3053 | /* surface */ |
||
3054 | obj_surface = SURFACE(src_surface->id); |
||
3055 | orig_w = obj_surface->orig_width; |
||
3056 | orig_h = obj_surface->orig_height; |
||
3057 | w = obj_surface->width; |
||
3058 | h = obj_surface->height; |
||
3059 | |||
3060 | if (pp_context->stmm.bo == NULL) { |
||
3061 | pp_context->stmm.bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3062 | "STMM surface", |
||
3063 | w * h, |
||
3064 | 4096); |
||
3065 | assert(pp_context->stmm.bo); |
||
3066 | } |
||
3067 | |||
3068 | /* source UV surface index 1 */ |
||
3069 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3070 | obj_surface->bo, w * h, |
||
3071 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
3072 | 1, 0); |
||
3073 | |||
3074 | /* source YUV surface index 3 */ |
||
3075 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
3076 | obj_surface->bo, 0, |
||
3077 | orig_w, orig_h, w, |
||
3078 | 0, h, |
||
3079 | SURFACE_FORMAT_PLANAR_420_8, 1, |
||
3080 | 3); |
||
3081 | |||
3082 | /* source (temporal reference) YUV surface index 4 */ |
||
3083 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
3084 | obj_surface->bo, 0, |
||
3085 | orig_w, orig_h, w, |
||
3086 | 0, h, |
||
3087 | SURFACE_FORMAT_PLANAR_420_8, 1, |
||
3088 | 4); |
||
3089 | |||
3090 | /* STMM / History Statistics input surface, index 5 */ |
||
3091 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3092 | pp_context->stmm.bo, 0, |
||
3093 | orig_w, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
3094 | 5, 1); |
||
3095 | |||
3096 | /* destination surface */ |
||
3097 | obj_surface = SURFACE(dst_surface->id); |
||
3098 | orig_w = obj_surface->orig_width; |
||
3099 | orig_h = obj_surface->orig_height; |
||
3100 | w = obj_surface->width; |
||
3101 | h = obj_surface->height; |
||
3102 | |||
3103 | /* destination(Previous frame) Y surface index 27 */ |
||
3104 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3105 | obj_surface->bo, 0, |
||
3106 | orig_w / 4, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
3107 | 27, 1); |
||
3108 | |||
3109 | /* destination(Previous frame) UV surface index 28 */ |
||
3110 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3111 | obj_surface->bo, w * h, |
||
3112 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
3113 | 28, 1); |
||
3114 | |||
3115 | /* destination(Current frame) Y surface index 30 */ |
||
3116 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3117 | obj_surface->bo, 0, |
||
3118 | orig_w / 4, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
3119 | 30, 1); |
||
3120 | |||
3121 | /* destination(Current frame) UV surface index 31 */ |
||
3122 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3123 | obj_surface->bo, w * h, |
||
3124 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
3125 | 31, 1); |
||
3126 | |||
3127 | /* STMM output surface, index 33 */ |
||
3128 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3129 | pp_context->stmm.bo, 0, |
||
3130 | orig_w, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
3131 | 33, 1); |
||
3132 | |||
3133 | |||
3134 | /* sampler dndi */ |
||
3135 | dri_bo_map(pp_context->sampler_state_table.bo, True); |
||
3136 | assert(pp_context->sampler_state_table.bo->virtual); |
||
3137 | assert(sizeof(*sampler_dndi) == sizeof(int) * 8); |
||
3138 | sampler_dndi = pp_context->sampler_state_table.bo->virtual; |
||
3139 | |||
3140 | /* sample dndi index 0 */ |
||
3141 | index = 0; |
||
3142 | sampler_dndi[index].dw0.denoise_asd_threshold = 0; |
||
3143 | sampler_dndi[index].dw0.dnmh_delt = 8; |
||
3144 | sampler_dndi[index].dw0.vdi_walker_y_stride = 0; |
||
3145 | sampler_dndi[index].dw0.vdi_walker_frame_sharing_enable = 0; |
||
3146 | sampler_dndi[index].dw0.denoise_maximum_history = 128; // 128-240 |
||
3147 | sampler_dndi[index].dw0.denoise_stad_threshold = 0; |
||
3148 | |||
3149 | sampler_dndi[index].dw1.denoise_threshold_for_sum_of_complexity_measure = 64; |
||
3150 | sampler_dndi[index].dw1.denoise_moving_pixel_threshold = 0; |
||
3151 | sampler_dndi[index].dw1.stmm_c2 = 0; |
||
3152 | sampler_dndi[index].dw1.low_temporal_difference_threshold = 8; |
||
3153 | sampler_dndi[index].dw1.temporal_difference_threshold = 16; |
||
3154 | |||
3155 | sampler_dndi[index].dw2.block_noise_estimate_noise_threshold = 15; // 0-31 |
||
3156 | sampler_dndi[index].dw2.bne_edge_th = 1; |
||
3157 | sampler_dndi[index].dw2.smooth_mv_th = 0; |
||
3158 | sampler_dndi[index].dw2.sad_tight_th = 5; |
||
3159 | sampler_dndi[index].dw2.cat_slope_minus1 = 9; |
||
3160 | sampler_dndi[index].dw2.good_neighbor_th = 4; |
||
3161 | |||
3162 | sampler_dndi[index].dw3.maximum_stmm = 128; |
||
3163 | sampler_dndi[index].dw3.multipler_for_vecm = 2; |
||
3164 | sampler_dndi[index].dw3.blending_constant_across_time_for_small_values_of_stmm = 0; |
||
3165 | sampler_dndi[index].dw3.blending_constant_across_time_for_large_values_of_stmm = 64; |
||
3166 | sampler_dndi[index].dw3.stmm_blending_constant_select = 0; |
||
3167 | |||
3168 | sampler_dndi[index].dw4.sdi_delta = 8; |
||
3169 | sampler_dndi[index].dw4.sdi_threshold = 128; |
||
3170 | sampler_dndi[index].dw4.stmm_output_shift = 7; // stmm_max - stmm_min = 2 ^ stmm_output_shift |
||
3171 | sampler_dndi[index].dw4.stmm_shift_up = 0; |
||
3172 | sampler_dndi[index].dw4.stmm_shift_down = 0; |
||
3173 | sampler_dndi[index].dw4.minimum_stmm = 0; |
||
3174 | |||
3175 | sampler_dndi[index].dw5.fmd_temporal_difference_threshold = 0; |
||
3176 | sampler_dndi[index].dw5.sdi_fallback_mode_2_constant = 0; |
||
3177 | sampler_dndi[index].dw5.sdi_fallback_mode_1_t2_constant = 0; |
||
3178 | sampler_dndi[index].dw5.sdi_fallback_mode_1_t1_constant = 0; |
||
3179 | |||
3180 | sampler_dndi[index].dw6.dn_enable = 0; |
||
3181 | sampler_dndi[index].dw6.di_enable = 1; |
||
3182 | sampler_dndi[index].dw6.di_partial = 0; |
||
3183 | sampler_dndi[index].dw6.dndi_top_first = dndi_top_first; |
||
3184 | sampler_dndi[index].dw6.dndi_stream_id = 1; |
||
3185 | sampler_dndi[index].dw6.dndi_first_frame = 1; |
||
3186 | sampler_dndi[index].dw6.progressive_dn = 0; |
||
3187 | sampler_dndi[index].dw6.mcdi_enable = 0; |
||
3188 | sampler_dndi[index].dw6.fmd_tear_threshold = 32; |
||
3189 | sampler_dndi[index].dw6.cat_th1 = 0; |
||
3190 | sampler_dndi[index].dw6.fmd2_vertical_difference_threshold = 32; |
||
3191 | sampler_dndi[index].dw6.fmd1_vertical_difference_threshold = 32; |
||
3192 | |||
3193 | sampler_dndi[index].dw7.sad_tha = 5; |
||
3194 | sampler_dndi[index].dw7.sad_thb = 10; |
||
3195 | sampler_dndi[index].dw7.fmd_for_1st_field_of_current_frame = 0; |
||
3196 | sampler_dndi[index].dw7.mc_pixel_consistency_th = 25; |
||
3197 | sampler_dndi[index].dw7.fmd_for_2nd_field_of_previous_frame = 0; |
||
3198 | sampler_dndi[index].dw7.vdi_walker_enable = 0; |
||
3199 | sampler_dndi[index].dw7.neighborpixel_th = 10; |
||
3200 | sampler_dndi[index].dw7.column_width_minus1 = w / 16; |
||
3201 | |||
3202 | dri_bo_unmap(pp_context->sampler_state_table.bo); |
||
3203 | |||
3204 | /* private function & data */ |
||
3205 | pp_context->pp_x_steps = gen7_pp_dndi_x_steps; |
||
3206 | pp_context->pp_y_steps = gen7_pp_dndi_y_steps; |
||
3207 | pp_context->pp_set_block_parameter = gen7_pp_dndi_set_block_parameter; |
||
3208 | |||
3209 | pp_static_parameter->grf1.di_statistics_surface_pitch_div2 = w / 2; |
||
3210 | pp_static_parameter->grf1.di_statistics_surface_height_div4 = h / 4; |
||
3211 | pp_static_parameter->grf1.di_top_field_first = 0; |
||
3212 | pp_static_parameter->grf1.pointer_to_inline_parameter = 7; |
||
3213 | |||
3214 | pp_static_parameter->grf2.di_destination_packed_y_component_offset = 0; |
||
3215 | pp_static_parameter->grf2.di_destination_packed_u_component_offset = 1; |
||
3216 | pp_static_parameter->grf2.di_destination_packed_v_component_offset = 3; |
||
3217 | |||
3218 | pp_static_parameter->grf4.di_hoffset_svf_from_dvf = 0; |
||
3219 | pp_static_parameter->grf4.di_voffset_svf_from_dvf = 0; |
||
3220 | |||
3221 | pp_dndi_context->dest_w = w; |
||
3222 | pp_dndi_context->dest_h = h; |
||
3223 | |||
3224 | dst_surface->flags = I965_SURFACE_FLAG_FRAME; |
||
3225 | |||
3226 | return VA_STATUS_SUCCESS; |
||
3227 | } |
||
3228 | |||
3229 | static int |
||
3230 | gen7_pp_dn_x_steps(void *private_context) |
||
3231 | { |
||
3232 | struct pp_dn_context *pp_dn_context = private_context; |
||
3233 | |||
3234 | return pp_dn_context->dest_w / 16; |
||
3235 | } |
||
3236 | |||
3237 | static int |
||
3238 | gen7_pp_dn_y_steps(void *private_context) |
||
3239 | { |
||
3240 | struct pp_dn_context *pp_dn_context = private_context; |
||
3241 | |||
3242 | return pp_dn_context->dest_h / 4; |
||
3243 | } |
||
3244 | |||
3245 | static int |
||
3246 | gen7_pp_dn_set_block_parameter(struct i965_post_processing_context *pp_context, int x, int y) |
||
3247 | { |
||
3248 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
3249 | |||
3250 | pp_inline_parameter->grf5.destination_block_horizontal_origin = x * 16; |
||
3251 | pp_inline_parameter->grf5.destination_block_vertical_origin = y * 4; |
||
3252 | |||
3253 | return 0; |
||
3254 | } |
||
3255 | |||
3256 | static VAStatus |
||
3257 | gen7_pp_nv12_dn_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context, |
||
3258 | const struct i965_surface *src_surface, |
||
3259 | const VARectangle *src_rect, |
||
3260 | struct i965_surface *dst_surface, |
||
3261 | const VARectangle *dst_rect, |
||
3262 | void *filter_param) |
||
3263 | { |
||
3264 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
3265 | struct pp_dn_context *pp_dn_context = (struct pp_dn_context *)&pp_context->private_context; |
||
3266 | struct gen7_pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter; |
||
3267 | struct object_surface *obj_surface; |
||
3268 | struct gen7_sampler_dndi *sampler_dn; |
||
3269 | |||
3270 | int index; |
||
3271 | int w, h; |
||
3272 | int orig_w, orig_h; |
||
3273 | int dn_strength = 15; |
||
3274 | int dndi_top_first = 1; |
||
3275 | int dn_progressive = 0; |
||
3276 | |||
3277 | if (src_surface->flags == I965_SURFACE_FLAG_FRAME) { |
||
3278 | dndi_top_first = 1; |
||
3279 | dn_progressive = 1; |
||
3280 | } else if (src_surface->flags == I965_SURFACE_FLAG_TOP_FIELD_FIRST) { |
||
3281 | dndi_top_first = 1; |
||
3282 | dn_progressive = 0; |
||
3283 | } else { |
||
3284 | dndi_top_first = 0; |
||
3285 | dn_progressive = 0; |
||
3286 | } |
||
3287 | |||
3288 | /* surface */ |
||
3289 | obj_surface = SURFACE(src_surface->id); |
||
3290 | orig_w = obj_surface->orig_width; |
||
3291 | orig_h = obj_surface->orig_height; |
||
3292 | w = obj_surface->width; |
||
3293 | h = obj_surface->height; |
||
3294 | |||
3295 | if (pp_context->stmm.bo == NULL) { |
||
3296 | pp_context->stmm.bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3297 | "STMM surface", |
||
3298 | w * h, |
||
3299 | 4096); |
||
3300 | assert(pp_context->stmm.bo); |
||
3301 | } |
||
3302 | |||
3303 | /* source UV surface index 1 */ |
||
3304 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3305 | obj_surface->bo, w * h, |
||
3306 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
3307 | 1, 0); |
||
3308 | |||
3309 | /* source YUV surface index 3 */ |
||
3310 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
3311 | obj_surface->bo, 0, |
||
3312 | orig_w, orig_h, w, |
||
3313 | 0, h, |
||
3314 | SURFACE_FORMAT_PLANAR_420_8, 1, |
||
3315 | 3); |
||
3316 | |||
3317 | /* source (temporal reference) YUV surface index 4 */ |
||
3318 | gen7_pp_set_surface2_state(ctx, pp_context, |
||
3319 | obj_surface->bo, 0, |
||
3320 | orig_w, orig_h, w, |
||
3321 | 0, h, |
||
3322 | SURFACE_FORMAT_PLANAR_420_8, 1, |
||
3323 | 4); |
||
3324 | |||
3325 | /* STMM / History Statistics input surface, index 5 */ |
||
3326 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3327 | pp_context->stmm.bo, 0, |
||
3328 | orig_w, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
3329 | 5, 1); |
||
3330 | |||
3331 | /* destination surface */ |
||
3332 | obj_surface = SURFACE(dst_surface->id); |
||
3333 | orig_w = obj_surface->orig_width; |
||
3334 | orig_h = obj_surface->orig_height; |
||
3335 | w = obj_surface->width; |
||
3336 | h = obj_surface->height; |
||
3337 | |||
3338 | /* destination Y surface index 24 */ |
||
3339 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3340 | obj_surface->bo, 0, |
||
3341 | orig_w / 4, orig_h, w, I965_SURFACEFORMAT_R8_UNORM, |
||
3342 | 24, 1); |
||
3343 | |||
3344 | /* destination UV surface index 25 */ |
||
3345 | gen7_pp_set_surface_state(ctx, pp_context, |
||
3346 | obj_surface->bo, w * h, |
||
3347 | orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM, |
||
3348 | 25, 1); |
||
3349 | |||
3350 | /* sampler dn */ |
||
3351 | dri_bo_map(pp_context->sampler_state_table.bo, True); |
||
3352 | assert(pp_context->sampler_state_table.bo->virtual); |
||
3353 | assert(sizeof(*sampler_dn) == sizeof(int) * 8); |
||
3354 | sampler_dn = pp_context->sampler_state_table.bo->virtual; |
||
3355 | |||
3356 | /* sample dn index 1 */ |
||
3357 | index = 0; |
||
3358 | sampler_dn[index].dw0.denoise_asd_threshold = 0; |
||
3359 | sampler_dn[index].dw0.dnmh_delt = 8; |
||
3360 | sampler_dn[index].dw0.vdi_walker_y_stride = 0; |
||
3361 | sampler_dn[index].dw0.vdi_walker_frame_sharing_enable = 0; |
||
3362 | sampler_dn[index].dw0.denoise_maximum_history = 128; // 128-240 |
||
3363 | sampler_dn[index].dw0.denoise_stad_threshold = 0; |
||
3364 | |||
3365 | sampler_dn[index].dw1.denoise_threshold_for_sum_of_complexity_measure = 64; |
||
3366 | sampler_dn[index].dw1.denoise_moving_pixel_threshold = 0; |
||
3367 | sampler_dn[index].dw1.stmm_c2 = 0; |
||
3368 | sampler_dn[index].dw1.low_temporal_difference_threshold = 8; |
||
3369 | sampler_dn[index].dw1.temporal_difference_threshold = 16; |
||
3370 | |||
3371 | sampler_dn[index].dw2.block_noise_estimate_noise_threshold = dn_strength; // 0-31 |
||
3372 | sampler_dn[index].dw2.bne_edge_th = 1; |
||
3373 | sampler_dn[index].dw2.smooth_mv_th = 0; |
||
3374 | sampler_dn[index].dw2.sad_tight_th = 5; |
||
3375 | sampler_dn[index].dw2.cat_slope_minus1 = 9; |
||
3376 | sampler_dn[index].dw2.good_neighbor_th = 4; |
||
3377 | |||
3378 | sampler_dn[index].dw3.maximum_stmm = 128; |
||
3379 | sampler_dn[index].dw3.multipler_for_vecm = 2; |
||
3380 | sampler_dn[index].dw3.blending_constant_across_time_for_small_values_of_stmm = 0; |
||
3381 | sampler_dn[index].dw3.blending_constant_across_time_for_large_values_of_stmm = 64; |
||
3382 | sampler_dn[index].dw3.stmm_blending_constant_select = 0; |
||
3383 | |||
3384 | sampler_dn[index].dw4.sdi_delta = 8; |
||
3385 | sampler_dn[index].dw4.sdi_threshold = 128; |
||
3386 | sampler_dn[index].dw4.stmm_output_shift = 7; // stmm_max - stmm_min = 2 ^ stmm_output_shift |
||
3387 | sampler_dn[index].dw4.stmm_shift_up = 0; |
||
3388 | sampler_dn[index].dw4.stmm_shift_down = 0; |
||
3389 | sampler_dn[index].dw4.minimum_stmm = 0; |
||
3390 | |||
3391 | sampler_dn[index].dw5.fmd_temporal_difference_threshold = 0; |
||
3392 | sampler_dn[index].dw5.sdi_fallback_mode_2_constant = 0; |
||
3393 | sampler_dn[index].dw5.sdi_fallback_mode_1_t2_constant = 0; |
||
3394 | sampler_dn[index].dw5.sdi_fallback_mode_1_t1_constant = 0; |
||
3395 | |||
3396 | sampler_dn[index].dw6.dn_enable = 1; |
||
3397 | sampler_dn[index].dw6.di_enable = 0; |
||
3398 | sampler_dn[index].dw6.di_partial = 0; |
||
3399 | sampler_dn[index].dw6.dndi_top_first = dndi_top_first; |
||
3400 | sampler_dn[index].dw6.dndi_stream_id = 1; |
||
3401 | sampler_dn[index].dw6.dndi_first_frame = 1; |
||
3402 | sampler_dn[index].dw6.progressive_dn = dn_progressive; |
||
3403 | sampler_dn[index].dw6.mcdi_enable = 0; |
||
3404 | sampler_dn[index].dw6.fmd_tear_threshold = 32; |
||
3405 | sampler_dn[index].dw6.cat_th1 = 0; |
||
3406 | sampler_dn[index].dw6.fmd2_vertical_difference_threshold = 32; |
||
3407 | sampler_dn[index].dw6.fmd1_vertical_difference_threshold = 32; |
||
3408 | |||
3409 | sampler_dn[index].dw7.sad_tha = 5; |
||
3410 | sampler_dn[index].dw7.sad_thb = 10; |
||
3411 | sampler_dn[index].dw7.fmd_for_1st_field_of_current_frame = 2; |
||
3412 | sampler_dn[index].dw7.mc_pixel_consistency_th = 25; |
||
3413 | sampler_dn[index].dw7.fmd_for_2nd_field_of_previous_frame = 1; |
||
3414 | sampler_dn[index].dw7.vdi_walker_enable = 0; |
||
3415 | sampler_dn[index].dw7.neighborpixel_th = 10; |
||
3416 | sampler_dn[index].dw7.column_width_minus1 = w / 16; |
||
3417 | |||
3418 | dri_bo_unmap(pp_context->sampler_state_table.bo); |
||
3419 | |||
3420 | /* private function & data */ |
||
3421 | pp_context->pp_x_steps = gen7_pp_dn_x_steps; |
||
3422 | pp_context->pp_y_steps = gen7_pp_dn_y_steps; |
||
3423 | pp_context->pp_set_block_parameter = gen7_pp_dn_set_block_parameter; |
||
3424 | |||
3425 | pp_static_parameter->grf1.di_statistics_surface_pitch_div2 = w / 2; |
||
3426 | pp_static_parameter->grf1.di_statistics_surface_height_div4 = h / 4; |
||
3427 | pp_static_parameter->grf1.di_top_field_first = 0; |
||
3428 | pp_static_parameter->grf1.pointer_to_inline_parameter = 7; |
||
3429 | |||
3430 | pp_static_parameter->grf2.di_destination_packed_y_component_offset = 0; |
||
3431 | pp_static_parameter->grf2.di_destination_packed_u_component_offset = 1; |
||
3432 | pp_static_parameter->grf2.di_destination_packed_v_component_offset = 3; |
||
3433 | |||
3434 | pp_static_parameter->grf4.di_hoffset_svf_from_dvf = 0; |
||
3435 | pp_static_parameter->grf4.di_voffset_svf_from_dvf = 0; |
||
3436 | |||
3437 | pp_dn_context->dest_w = w; |
||
3438 | pp_dn_context->dest_h = h; |
||
3439 | |||
3440 | dst_surface->flags = src_surface->flags; |
||
3441 | |||
3442 | return VA_STATUS_SUCCESS; |
||
3443 | } |
||
3444 | |||
3445 | static VAStatus |
||
3446 | ironlake_pp_initialize( |
||
3447 | VADriverContextP ctx, |
||
3448 | struct i965_post_processing_context *pp_context, |
||
3449 | const struct i965_surface *src_surface, |
||
3450 | const VARectangle *src_rect, |
||
3451 | struct i965_surface *dst_surface, |
||
3452 | const VARectangle *dst_rect, |
||
3453 | int pp_index, |
||
3454 | void *filter_param |
||
3455 | ) |
||
3456 | { |
||
3457 | VAStatus va_status; |
||
3458 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
3459 | struct pp_module *pp_module; |
||
3460 | dri_bo *bo; |
||
3461 | int static_param_size, inline_param_size; |
||
3462 | |||
3463 | dri_bo_unreference(pp_context->surface_state_binding_table.bo); |
||
3464 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3465 | "surface state & binding table", |
||
3466 | (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_PP_SURFACES, |
||
3467 | 4096); |
||
3468 | assert(bo); |
||
3469 | pp_context->surface_state_binding_table.bo = bo; |
||
3470 | |||
3471 | dri_bo_unreference(pp_context->curbe.bo); |
||
3472 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3473 | "constant buffer", |
||
3474 | 4096, |
||
3475 | 4096); |
||
3476 | assert(bo); |
||
3477 | pp_context->curbe.bo = bo; |
||
3478 | |||
3479 | dri_bo_unreference(pp_context->idrt.bo); |
||
3480 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3481 | "interface discriptor", |
||
3482 | sizeof(struct i965_interface_descriptor), |
||
3483 | 4096); |
||
3484 | assert(bo); |
||
3485 | pp_context->idrt.bo = bo; |
||
3486 | pp_context->idrt.num_interface_descriptors = 0; |
||
3487 | |||
3488 | dri_bo_unreference(pp_context->sampler_state_table.bo); |
||
3489 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3490 | "sampler state table", |
||
3491 | 4096, |
||
3492 | 4096); |
||
3493 | assert(bo); |
||
3494 | dri_bo_map(bo, True); |
||
3495 | memset(bo->virtual, 0, bo->size); |
||
3496 | dri_bo_unmap(bo); |
||
3497 | pp_context->sampler_state_table.bo = bo; |
||
3498 | |||
3499 | dri_bo_unreference(pp_context->sampler_state_table.bo_8x8); |
||
3500 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3501 | "sampler 8x8 state ", |
||
3502 | 4096, |
||
3503 | 4096); |
||
3504 | assert(bo); |
||
3505 | pp_context->sampler_state_table.bo_8x8 = bo; |
||
3506 | |||
3507 | dri_bo_unreference(pp_context->sampler_state_table.bo_8x8_uv); |
||
3508 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3509 | "sampler 8x8 state ", |
||
3510 | 4096, |
||
3511 | 4096); |
||
3512 | assert(bo); |
||
3513 | pp_context->sampler_state_table.bo_8x8_uv = bo; |
||
3514 | |||
3515 | dri_bo_unreference(pp_context->vfe_state.bo); |
||
3516 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3517 | "vfe state", |
||
3518 | sizeof(struct i965_vfe_state), |
||
3519 | 4096); |
||
3520 | assert(bo); |
||
3521 | pp_context->vfe_state.bo = bo; |
||
3522 | |||
3523 | static_param_size = sizeof(struct pp_static_parameter); |
||
3524 | inline_param_size = sizeof(struct pp_inline_parameter); |
||
3525 | |||
3526 | memset(pp_context->pp_static_parameter, 0, static_param_size); |
||
3527 | memset(pp_context->pp_inline_parameter, 0, inline_param_size); |
||
3528 | |||
3529 | assert(pp_index >= PP_NULL && pp_index < NUM_PP_MODULES); |
||
3530 | pp_context->current_pp = pp_index; |
||
3531 | pp_module = &pp_context->pp_modules[pp_index]; |
||
3532 | |||
3533 | if (pp_module->initialize) |
||
3534 | va_status = pp_module->initialize(ctx, pp_context, |
||
3535 | src_surface, |
||
3536 | src_rect, |
||
3537 | dst_surface, |
||
3538 | dst_rect, |
||
3539 | filter_param); |
||
3540 | else |
||
3541 | va_status = VA_STATUS_ERROR_UNIMPLEMENTED; |
||
3542 | |||
3543 | return va_status; |
||
3544 | } |
||
3545 | |||
3546 | static VAStatus |
||
3547 | ironlake_post_processing( |
||
3548 | VADriverContextP ctx, |
||
3549 | struct i965_post_processing_context *pp_context, |
||
3550 | const struct i965_surface *src_surface, |
||
3551 | const VARectangle *src_rect, |
||
3552 | struct i965_surface *dst_surface, |
||
3553 | const VARectangle *dst_rect, |
||
3554 | int pp_index, |
||
3555 | void *filter_param |
||
3556 | ) |
||
3557 | { |
||
3558 | VAStatus va_status; |
||
3559 | |||
3560 | va_status = ironlake_pp_initialize(ctx, pp_context, |
||
3561 | src_surface, |
||
3562 | src_rect, |
||
3563 | dst_surface, |
||
3564 | dst_rect, |
||
3565 | pp_index, |
||
3566 | filter_param); |
||
3567 | |||
3568 | if (va_status == VA_STATUS_SUCCESS) { |
||
3569 | ironlake_pp_states_setup(ctx, pp_context); |
||
3570 | ironlake_pp_pipeline_setup(ctx, pp_context); |
||
3571 | } |
||
3572 | |||
3573 | return va_status; |
||
3574 | } |
||
3575 | |||
3576 | static VAStatus |
||
3577 | gen6_pp_initialize( |
||
3578 | VADriverContextP ctx, |
||
3579 | struct i965_post_processing_context *pp_context, |
||
3580 | const struct i965_surface *src_surface, |
||
3581 | const VARectangle *src_rect, |
||
3582 | struct i965_surface *dst_surface, |
||
3583 | const VARectangle *dst_rect, |
||
3584 | int pp_index, |
||
3585 | void * filter_param |
||
3586 | ) |
||
3587 | { |
||
3588 | VAStatus va_status; |
||
3589 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
3590 | struct pp_module *pp_module; |
||
3591 | dri_bo *bo; |
||
3592 | int static_param_size, inline_param_size; |
||
3593 | |||
3594 | dri_bo_unreference(pp_context->surface_state_binding_table.bo); |
||
3595 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3596 | "surface state & binding table", |
||
3597 | (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_PP_SURFACES, |
||
3598 | 4096); |
||
3599 | assert(bo); |
||
3600 | pp_context->surface_state_binding_table.bo = bo; |
||
3601 | |||
3602 | dri_bo_unreference(pp_context->curbe.bo); |
||
3603 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3604 | "constant buffer", |
||
3605 | 4096, |
||
3606 | 4096); |
||
3607 | assert(bo); |
||
3608 | pp_context->curbe.bo = bo; |
||
3609 | |||
3610 | dri_bo_unreference(pp_context->idrt.bo); |
||
3611 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3612 | "interface discriptor", |
||
3613 | sizeof(struct gen6_interface_descriptor_data), |
||
3614 | 4096); |
||
3615 | assert(bo); |
||
3616 | pp_context->idrt.bo = bo; |
||
3617 | pp_context->idrt.num_interface_descriptors = 0; |
||
3618 | |||
3619 | dri_bo_unreference(pp_context->sampler_state_table.bo); |
||
3620 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3621 | "sampler state table", |
||
3622 | 4096, |
||
3623 | 4096); |
||
3624 | assert(bo); |
||
3625 | dri_bo_map(bo, True); |
||
3626 | memset(bo->virtual, 0, bo->size); |
||
3627 | dri_bo_unmap(bo); |
||
3628 | pp_context->sampler_state_table.bo = bo; |
||
3629 | |||
3630 | dri_bo_unreference(pp_context->sampler_state_table.bo_8x8); |
||
3631 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3632 | "sampler 8x8 state ", |
||
3633 | 4096, |
||
3634 | 4096); |
||
3635 | assert(bo); |
||
3636 | pp_context->sampler_state_table.bo_8x8 = bo; |
||
3637 | |||
3638 | dri_bo_unreference(pp_context->sampler_state_table.bo_8x8_uv); |
||
3639 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3640 | "sampler 8x8 state ", |
||
3641 | 4096, |
||
3642 | 4096); |
||
3643 | assert(bo); |
||
3644 | pp_context->sampler_state_table.bo_8x8_uv = bo; |
||
3645 | |||
3646 | dri_bo_unreference(pp_context->vfe_state.bo); |
||
3647 | bo = dri_bo_alloc(i965->intel.bufmgr, |
||
3648 | "vfe state", |
||
3649 | sizeof(struct i965_vfe_state), |
||
3650 | 4096); |
||
3651 | assert(bo); |
||
3652 | pp_context->vfe_state.bo = bo; |
||
3653 | |||
3654 | if (IS_GEN7(i965->intel.device_id)) { |
||
3655 | static_param_size = sizeof(struct gen7_pp_static_parameter); |
||
3656 | inline_param_size = sizeof(struct gen7_pp_inline_parameter); |
||
3657 | } else { |
||
3658 | static_param_size = sizeof(struct pp_static_parameter); |
||
3659 | inline_param_size = sizeof(struct pp_inline_parameter); |
||
3660 | } |
||
3661 | |||
3662 | memset(pp_context->pp_static_parameter, 0, static_param_size); |
||
3663 | memset(pp_context->pp_inline_parameter, 0, inline_param_size); |
||
3664 | |||
3665 | assert(pp_index >= PP_NULL && pp_index < NUM_PP_MODULES); |
||
3666 | pp_context->current_pp = pp_index; |
||
3667 | pp_module = &pp_context->pp_modules[pp_index]; |
||
3668 | |||
3669 | if (pp_module->initialize) |
||
3670 | va_status = pp_module->initialize(ctx, pp_context, |
||
3671 | src_surface, |
||
3672 | src_rect, |
||
3673 | dst_surface, |
||
3674 | dst_rect, |
||
3675 | filter_param); |
||
3676 | else |
||
3677 | va_status = VA_STATUS_ERROR_UNIMPLEMENTED; |
||
3678 | |||
3679 | calculate_boundary_block_mask(pp_context, dst_rect); |
||
3680 | |||
3681 | return va_status; |
||
3682 | } |
||
3683 | |||
3684 | static void |
||
3685 | gen6_pp_interface_descriptor_table(VADriverContextP ctx, |
||
3686 | struct i965_post_processing_context *pp_context) |
||
3687 | { |
||
3688 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
3689 | struct gen6_interface_descriptor_data *desc; |
||
3690 | dri_bo *bo; |
||
3691 | int pp_index = pp_context->current_pp; |
||
3692 | |||
3693 | bo = pp_context->idrt.bo; |
||
3694 | dri_bo_map(bo, True); |
||
3695 | assert(bo->virtual); |
||
3696 | desc = bo->virtual; |
||
3697 | memset(desc, 0, sizeof(*desc)); |
||
3698 | desc->desc0.kernel_start_pointer = |
||
3699 | pp_context->pp_modules[pp_index].kernel.bo->offset >> 6; /* reloc */ |
||
3700 | desc->desc1.single_program_flow = 1; |
||
3701 | desc->desc1.floating_point_mode = FLOATING_POINT_IEEE_754; |
||
3702 | desc->desc2.sampler_count = 1; /* 1 - 4 samplers used */ |
||
3703 | desc->desc2.sampler_state_pointer = |
||
3704 | pp_context->sampler_state_table.bo->offset >> 5; |
||
3705 | desc->desc3.binding_table_entry_count = 0; |
||
3706 | desc->desc3.binding_table_pointer = (BINDING_TABLE_OFFSET >> 5); |
||
3707 | desc->desc4.constant_urb_entry_read_offset = 0; |
||
3708 | |||
3709 | if (IS_GEN7(i965->intel.device_id)) |
||
3710 | desc->desc4.constant_urb_entry_read_length = 6; /* grf 1-6 */ |
||
3711 | else |
||
3712 | desc->desc4.constant_urb_entry_read_length = 4; /* grf 1-4 */ |
||
3713 | |||
3714 | dri_bo_emit_reloc(bo, |
||
3715 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
3716 | 0, |
||
3717 | offsetof(struct gen6_interface_descriptor_data, desc0), |
||
3718 | pp_context->pp_modules[pp_index].kernel.bo); |
||
3719 | |||
3720 | dri_bo_emit_reloc(bo, |
||
3721 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
3722 | desc->desc2.sampler_count << 2, |
||
3723 | offsetof(struct gen6_interface_descriptor_data, desc2), |
||
3724 | pp_context->sampler_state_table.bo); |
||
3725 | |||
3726 | dri_bo_unmap(bo); |
||
3727 | pp_context->idrt.num_interface_descriptors++; |
||
3728 | } |
||
3729 | |||
3730 | static void |
||
3731 | gen6_pp_upload_constants(VADriverContextP ctx, |
||
3732 | struct i965_post_processing_context *pp_context) |
||
3733 | { |
||
3734 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
3735 | unsigned char *constant_buffer; |
||
3736 | int param_size; |
||
3737 | |||
3738 | assert(sizeof(struct pp_static_parameter) == 128); |
||
3739 | assert(sizeof(struct gen7_pp_static_parameter) == 192); |
||
3740 | |||
3741 | if (IS_GEN7(i965->intel.device_id)) |
||
3742 | param_size = sizeof(struct gen7_pp_static_parameter); |
||
3743 | else |
||
3744 | param_size = sizeof(struct pp_static_parameter); |
||
3745 | |||
3746 | dri_bo_map(pp_context->curbe.bo, 1); |
||
3747 | assert(pp_context->curbe.bo->virtual); |
||
3748 | constant_buffer = pp_context->curbe.bo->virtual; |
||
3749 | memcpy(constant_buffer, pp_context->pp_static_parameter, param_size); |
||
3750 | dri_bo_unmap(pp_context->curbe.bo); |
||
3751 | } |
||
3752 | |||
3753 | static void |
||
3754 | gen6_pp_states_setup(VADriverContextP ctx, |
||
3755 | struct i965_post_processing_context *pp_context) |
||
3756 | { |
||
3757 | gen6_pp_interface_descriptor_table(ctx, pp_context); |
||
3758 | gen6_pp_upload_constants(ctx, pp_context); |
||
3759 | } |
||
3760 | |||
3761 | static void |
||
3762 | gen6_pp_pipeline_select(VADriverContextP ctx, |
||
3763 | struct i965_post_processing_context *pp_context) |
||
3764 | { |
||
3765 | struct intel_batchbuffer *batch = pp_context->batch; |
||
3766 | |||
3767 | BEGIN_BATCH(batch, 1); |
||
3768 | OUT_BATCH(batch, CMD_PIPELINE_SELECT | PIPELINE_SELECT_MEDIA); |
||
3769 | ADVANCE_BATCH(batch); |
||
3770 | } |
||
3771 | |||
3772 | static void |
||
3773 | gen6_pp_state_base_address(VADriverContextP ctx, |
||
3774 | struct i965_post_processing_context *pp_context) |
||
3775 | { |
||
3776 | struct intel_batchbuffer *batch = pp_context->batch; |
||
3777 | |||
3778 | BEGIN_BATCH(batch, 10); |
||
3779 | OUT_BATCH(batch, CMD_STATE_BASE_ADDRESS | (10 - 2)); |
||
3780 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3781 | OUT_RELOC(batch, pp_context->surface_state_binding_table.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, BASE_ADDRESS_MODIFY); /* Surface state base address */ |
||
3782 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3783 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3784 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3785 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3786 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3787 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3788 | OUT_BATCH(batch, 0 | BASE_ADDRESS_MODIFY); |
||
3789 | ADVANCE_BATCH(batch); |
||
3790 | } |
||
3791 | |||
3792 | static void |
||
3793 | gen6_pp_vfe_state(VADriverContextP ctx, |
||
3794 | struct i965_post_processing_context *pp_context) |
||
3795 | { |
||
3796 | struct intel_batchbuffer *batch = pp_context->batch; |
||
3797 | |||
3798 | BEGIN_BATCH(batch, 8); |
||
3799 | OUT_BATCH(batch, CMD_MEDIA_VFE_STATE | (8 - 2)); |
||
3800 | OUT_BATCH(batch, 0); |
||
3801 | OUT_BATCH(batch, |
||
3802 | (pp_context->urb.num_vfe_entries - 1) << 16 | |
||
3803 | pp_context->urb.num_vfe_entries << 8); |
||
3804 | OUT_BATCH(batch, 0); |
||
3805 | OUT_BATCH(batch, |
||
3806 | (pp_context->urb.size_vfe_entry * 2) << 16 | /* URB Entry Allocation Size, in 256 bits unit */ |
||
3807 | (pp_context->urb.size_cs_entry * pp_context->urb.num_cs_entries * 2)); /* CURBE Allocation Size, in 256 bits unit */ |
||
3808 | OUT_BATCH(batch, 0); |
||
3809 | OUT_BATCH(batch, 0); |
||
3810 | OUT_BATCH(batch, 0); |
||
3811 | ADVANCE_BATCH(batch); |
||
3812 | } |
||
3813 | |||
3814 | static void |
||
3815 | gen6_pp_curbe_load(VADriverContextP ctx, |
||
3816 | struct i965_post_processing_context *pp_context) |
||
3817 | { |
||
3818 | struct intel_batchbuffer *batch = pp_context->batch; |
||
3819 | |||
3820 | assert(pp_context->urb.size_cs_entry * pp_context->urb.num_cs_entries * 2 * 32 <= pp_context->curbe.bo->size); |
||
3821 | |||
3822 | BEGIN_BATCH(batch, 4); |
||
3823 | OUT_BATCH(batch, CMD_MEDIA_CURBE_LOAD | (4 - 2)); |
||
3824 | OUT_BATCH(batch, 0); |
||
3825 | OUT_BATCH(batch, |
||
3826 | pp_context->urb.size_cs_entry * pp_context->urb.num_cs_entries * 2 * 32); |
||
3827 | OUT_RELOC(batch, |
||
3828 | pp_context->curbe.bo, |
||
3829 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
3830 | 0); |
||
3831 | ADVANCE_BATCH(batch); |
||
3832 | } |
||
3833 | |||
3834 | static void |
||
3835 | gen6_interface_descriptor_load(VADriverContextP ctx, |
||
3836 | struct i965_post_processing_context *pp_context) |
||
3837 | { |
||
3838 | struct intel_batchbuffer *batch = pp_context->batch; |
||
3839 | |||
3840 | BEGIN_BATCH(batch, 4); |
||
3841 | OUT_BATCH(batch, CMD_MEDIA_INTERFACE_DESCRIPTOR_LOAD | (4 - 2)); |
||
3842 | OUT_BATCH(batch, 0); |
||
3843 | OUT_BATCH(batch, |
||
3844 | pp_context->idrt.num_interface_descriptors * sizeof(struct gen6_interface_descriptor_data)); |
||
3845 | OUT_RELOC(batch, |
||
3846 | pp_context->idrt.bo, |
||
3847 | I915_GEM_DOMAIN_INSTRUCTION, 0, |
||
3848 | 0); |
||
3849 | ADVANCE_BATCH(batch); |
||
3850 | } |
||
3851 | |||
3852 | static void update_block_mask_parameter(struct i965_post_processing_context *pp_context, int x, int y, int x_steps, int y_steps) |
||
3853 | { |
||
3854 | struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter; |
||
3855 | |||
3856 | pp_inline_parameter->grf5.block_vertical_mask = 0xff; |
||
3857 | pp_inline_parameter->grf6.block_vertical_mask_bottom = pp_context->block_vertical_mask_bottom; |
||
3858 | // for the first block, it always on the left edge. the second block will reload horizontal_mask from grf6.block_horizontal_mask_middle |
||
3859 | pp_inline_parameter->grf5.block_horizontal_mask = pp_context->block_horizontal_mask_left; |
||
3860 | pp_inline_parameter->grf6.block_horizontal_mask_middle = 0xffff; |
||
3861 | pp_inline_parameter->grf6.block_horizontal_mask_right = pp_context->block_horizontal_mask_right; |
||
3862 | |||
3863 | /* 1 x N */ |
||
3864 | if (x_steps == 1) { |
||
3865 | if (y == y_steps-1) { |
||
3866 | pp_inline_parameter->grf5.block_vertical_mask = pp_context->block_vertical_mask_bottom; |
||
3867 | } |
||
3868 | else { |
||
3869 | pp_inline_parameter->grf6.block_vertical_mask_bottom = 0xff; |
||
3870 | } |
||
3871 | } |
||
3872 | |||
3873 | /* M x 1 */ |
||
3874 | if (y_steps == 1) { |
||
3875 | if (x == 0) { // all blocks in this group are on the left edge |
||
3876 | pp_inline_parameter->grf6.block_horizontal_mask_middle = pp_context->block_horizontal_mask_left; |
||
3877 | pp_inline_parameter->grf6.block_horizontal_mask_right = pp_context->block_horizontal_mask_left; |
||
3878 | } |
||
3879 | else if (x == x_steps-1) { |
||
3880 | pp_inline_parameter->grf5.block_horizontal_mask = pp_context->block_horizontal_mask_right; |
||
3881 | pp_inline_parameter->grf6.block_horizontal_mask_middle = pp_context->block_horizontal_mask_right; |
||
3882 | } |
||
3883 | else { |
||
3884 | pp_inline_parameter->grf5.block_horizontal_mask = 0xffff; |
||
3885 | pp_inline_parameter->grf6.block_horizontal_mask_middle = 0xffff; |
||
3886 | pp_inline_parameter->grf6.block_horizontal_mask_right = 0xffff; |
||
3887 | } |
||
3888 | } |
||
3889 | |||
3890 | } |
||
3891 | |||
3892 | static void |
||
3893 | gen6_pp_object_walker(VADriverContextP ctx, |
||
3894 | struct i965_post_processing_context *pp_context) |
||
3895 | { |
||
3896 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
3897 | struct intel_batchbuffer *batch = pp_context->batch; |
||
3898 | int x, x_steps, y, y_steps; |
||
3899 | int param_size, command_length_in_dws; |
||
3900 | dri_bo *command_buffer; |
||
3901 | unsigned int *command_ptr; |
||
3902 | |||
3903 | if (IS_GEN7(i965->intel.device_id)) |
||
3904 | param_size = sizeof(struct gen7_pp_inline_parameter); |
||
3905 | else |
||
3906 | param_size = sizeof(struct pp_inline_parameter); |
||
3907 | |||
3908 | x_steps = pp_context->pp_x_steps(&pp_context->private_context); |
||
3909 | y_steps = pp_context->pp_y_steps(&pp_context->private_context); |
||
3910 | command_length_in_dws = 6 + (param_size >> 2); |
||
3911 | command_buffer = dri_bo_alloc(i965->intel.bufmgr, |
||
3912 | "command objects buffer", |
||
3913 | command_length_in_dws * 4 * x_steps * y_steps + 8, |
||
3914 | 4096); |
||
3915 | |||
3916 | dri_bo_map(command_buffer, 1); |
||
3917 | command_ptr = command_buffer->virtual; |
||
3918 | |||
3919 | for (y = 0; y < y_steps; y++) { |
||
3920 | for (x = 0; x < x_steps; x++) { |
||
3921 | if (!pp_context->pp_set_block_parameter(pp_context, x, y)) { |
||
3922 | // some common block parameter update goes here, apply to all pp functions |
||
3923 | if (IS_GEN6(i965->intel.device_id)) |
||
3924 | update_block_mask_parameter (pp_context, x, y, x_steps, y_steps); |
||
3925 | |||
3926 | *command_ptr++ = (CMD_MEDIA_OBJECT | (command_length_in_dws - 2)); |
||
3927 | *command_ptr++ = 0; |
||
3928 | *command_ptr++ = 0; |
||
3929 | *command_ptr++ = 0; |
||
3930 | *command_ptr++ = 0; |
||
3931 | *command_ptr++ = 0; |
||
3932 | memcpy(command_ptr, pp_context->pp_inline_parameter, param_size); |
||
3933 | command_ptr += (param_size >> 2); |
||
3934 | } |
||
3935 | } |
||
3936 | } |
||
3937 | |||
3938 | if (command_length_in_dws * x_steps * y_steps % 2 == 0) |
||
3939 | *command_ptr++ = 0; |
||
3940 | |||
3941 | *command_ptr = MI_BATCH_BUFFER_END; |
||
3942 | |||
3943 | dri_bo_unmap(command_buffer); |
||
3944 | |||
3945 | BEGIN_BATCH(batch, 2); |
||
3946 | OUT_BATCH(batch, MI_BATCH_BUFFER_START | (2 << 6)); |
||
3947 | OUT_RELOC(batch, command_buffer, |
||
3948 | I915_GEM_DOMAIN_COMMAND, 0, |
||
3949 | 0); |
||
3950 | ADVANCE_BATCH(batch); |
||
3951 | |||
3952 | dri_bo_unreference(command_buffer); |
||
3953 | |||
3954 | /* Have to execute the batch buffer here becuase MI_BATCH_BUFFER_END |
||
3955 | * will cause control to pass back to ring buffer |
||
3956 | */ |
||
3957 | intel_batchbuffer_end_atomic(batch); |
||
3958 | intel_batchbuffer_flush(batch); |
||
3959 | intel_batchbuffer_start_atomic(batch, 0x1000); |
||
3960 | } |
||
3961 | static void |
||
3962 | gen6_pp_pipeline_setup(VADriverContextP ctx, |
||
3963 | struct i965_post_processing_context *pp_context) |
||
3964 | { |
||
3965 | struct intel_batchbuffer *batch = pp_context->batch; |
||
3966 | |||
3967 | intel_batchbuffer_start_atomic(batch, 0x1000); |
||
3968 | intel_batchbuffer_emit_mi_flush(batch); |
||
3969 | gen6_pp_pipeline_select(ctx, pp_context); |
||
3970 | gen6_pp_state_base_address(ctx, pp_context); |
||
3971 | gen6_pp_vfe_state(ctx, pp_context); |
||
3972 | gen6_pp_curbe_load(ctx, pp_context); |
||
3973 | gen6_interface_descriptor_load(ctx, pp_context); |
||
3974 | gen6_pp_vfe_state(ctx, pp_context); |
||
3975 | gen6_pp_object_walker(ctx, pp_context); |
||
3976 | intel_batchbuffer_end_atomic(batch); |
||
3977 | } |
||
3978 | |||
3979 | static VAStatus |
||
3980 | gen6_post_processing( |
||
3981 | VADriverContextP ctx, |
||
3982 | struct i965_post_processing_context *pp_context, |
||
3983 | const struct i965_surface *src_surface, |
||
3984 | const VARectangle *src_rect, |
||
3985 | struct i965_surface *dst_surface, |
||
3986 | const VARectangle *dst_rect, |
||
3987 | int pp_index, |
||
3988 | void *filter_param |
||
3989 | ) |
||
3990 | { |
||
3991 | VAStatus va_status; |
||
3992 | |||
3993 | va_status = gen6_pp_initialize(ctx, pp_context, |
||
3994 | src_surface, |
||
3995 | src_rect, |
||
3996 | dst_surface, |
||
3997 | dst_rect, |
||
3998 | pp_index, |
||
3999 | filter_param); |
||
4000 | |||
4001 | if (va_status == VA_STATUS_SUCCESS) { |
||
4002 | gen6_pp_states_setup(ctx, pp_context); |
||
4003 | gen6_pp_pipeline_setup(ctx, pp_context); |
||
4004 | } |
||
4005 | |||
4006 | return va_status; |
||
4007 | } |
||
4008 | |||
4009 | static VAStatus |
||
4010 | gen75_post_processing( |
||
4011 | VADriverContextP ctx, |
||
4012 | struct i965_post_processing_context *pp_context, |
||
4013 | const struct i965_surface *src_surface, |
||
4014 | const VARectangle *src_rect, |
||
4015 | struct i965_surface *dst_surface, |
||
4016 | const VARectangle *dst_rect, |
||
4017 | int pp_index, |
||
4018 | void *filter_param |
||
4019 | ) |
||
4020 | { |
||
4021 | VAStatus va_status; |
||
4022 | struct intel_vebox_context * vebox_ctx = pp_context->pp_vebox_context; |
||
4023 | |||
4024 | assert(pp_index == PP_NV12_DNDI); |
||
4025 | |||
4026 | vebox_ctx->filters_mask = VPP_DNDI_DI; |
||
4027 | vebox_ctx->surface_input = src_surface->id; |
||
4028 | vebox_ctx->surface_output = dst_surface->id; |
||
4029 | |||
4030 | va_status = gen75_vebox_process_picture(ctx, vebox_ctx); |
||
4031 | |||
4032 | return va_status; |
||
4033 | } |
||
4034 | |||
4035 | static VAStatus |
||
4036 | i965_post_processing_internal( |
||
4037 | VADriverContextP ctx, |
||
4038 | struct i965_post_processing_context *pp_context, |
||
4039 | const struct i965_surface *src_surface, |
||
4040 | const VARectangle *src_rect, |
||
4041 | struct i965_surface *dst_surface, |
||
4042 | const VARectangle *dst_rect, |
||
4043 | int pp_index, |
||
4044 | void *filter_param |
||
4045 | ) |
||
4046 | { |
||
4047 | VAStatus va_status; |
||
4048 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4049 | |||
4050 | if(IS_HASWELL(i965->intel.device_id) && |
||
4051 | pp_index == PP_NV12_DNDI){ |
||
4052 | va_status = gen75_post_processing(ctx, pp_context, src_surface, src_rect, dst_surface, dst_rect, pp_index, filter_param); |
||
4053 | }else if (IS_GEN6(i965->intel.device_id) || |
||
4054 | IS_GEN7(i965->intel.device_id)){ |
||
4055 | va_status = gen6_post_processing(ctx, pp_context, src_surface, src_rect, dst_surface, dst_rect, pp_index, filter_param); |
||
4056 | }else{ |
||
4057 | va_status = ironlake_post_processing(ctx, pp_context, src_surface, src_rect, dst_surface, dst_rect, pp_index, filter_param); |
||
4058 | } |
||
4059 | |||
4060 | return va_status; |
||
4061 | } |
||
4062 | |||
4063 | VAStatus |
||
4064 | i965_DestroySurfaces(VADriverContextP ctx, |
||
4065 | VASurfaceID *surface_list, |
||
4066 | int num_surfaces); |
||
4067 | VAStatus |
||
4068 | i965_CreateSurfaces(VADriverContextP ctx, |
||
4069 | int width, |
||
4070 | int height, |
||
4071 | int format, |
||
4072 | int num_surfaces, |
||
4073 | VASurfaceID *surfaces); |
||
4074 | |||
4075 | static void |
||
4076 | rgb_to_yuv(unsigned int argb, |
||
4077 | unsigned char *y, |
||
4078 | unsigned char *u, |
||
4079 | unsigned char *v, |
||
4080 | unsigned char *a) |
||
4081 | { |
||
4082 | int r = ((argb >> 16) & 0xff); |
||
4083 | int g = ((argb >> 8) & 0xff); |
||
4084 | int b = ((argb >> 0) & 0xff); |
||
4085 | |||
4086 | *y = (257 * r + 504 * g + 98 * b) / 1000 + 16; |
||
4087 | *v = (439 * r - 368 * g - 71 * b) / 1000 + 128; |
||
4088 | *u = (-148 * r - 291 * g + 439 * b) / 1000 + 128; |
||
4089 | *a = ((argb >> 24) & 0xff); |
||
4090 | } |
||
4091 | |||
4092 | static void |
||
4093 | i965_vpp_clear_surface(VADriverContextP ctx, |
||
4094 | struct i965_post_processing_context *pp_context, |
||
4095 | VASurfaceID surface, |
||
4096 | unsigned int color) |
||
4097 | { |
||
4098 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4099 | struct intel_batchbuffer *batch = pp_context->batch; |
||
4100 | struct object_surface *obj_surface = SURFACE(surface); |
||
4101 | unsigned int blt_cmd, br13; |
||
4102 | unsigned int tiling = 0, swizzle = 0; |
||
4103 | int pitch; |
||
4104 | unsigned char y, u, v, a = 0; |
||
4105 | |||
4106 | /* Currently only support NV12 surface */ |
||
4107 | if (!obj_surface || obj_surface->fourcc != VA_FOURCC('N', 'V', '1', '2')) |
||
4108 | return; |
||
4109 | |||
4110 | rgb_to_yuv(color, &y, &u, &v, &a); |
||
4111 | |||
4112 | if (a == 0) |
||
4113 | return; |
||
4114 | |||
4115 | dri_bo_get_tiling(obj_surface->bo, &tiling, &swizzle); |
||
4116 | blt_cmd = XY_COLOR_BLT_CMD; |
||
4117 | pitch = obj_surface->width; |
||
4118 | |||
4119 | if (tiling != I915_TILING_NONE) { |
||
4120 | blt_cmd |= XY_COLOR_BLT_DST_TILED; |
||
4121 | pitch >>= 2; |
||
4122 | } |
||
4123 | |||
4124 | br13 = 0xf0 << 16; |
||
4125 | br13 |= BR13_8; |
||
4126 | br13 |= pitch; |
||
4127 | |||
4128 | if (IS_GEN6(i965->intel.device_id) || |
||
4129 | IS_GEN7(i965->intel.device_id)) { |
||
4130 | intel_batchbuffer_start_atomic_blt(batch, 48); |
||
4131 | BEGIN_BLT_BATCH(batch, 12); |
||
4132 | } else { |
||
4133 | intel_batchbuffer_start_atomic(batch, 48); |
||
4134 | BEGIN_BATCH(batch, 12); |
||
4135 | } |
||
4136 | |||
4137 | OUT_BATCH(batch, blt_cmd); |
||
4138 | OUT_BATCH(batch, br13); |
||
4139 | OUT_BATCH(batch, |
||
4140 | |||
4141 | 0); |
||
4142 | OUT_BATCH(batch, |
||
4143 | obj_surface->height << 16 | |
||
4144 | obj_surface->width); |
||
4145 | OUT_RELOC(batch, obj_surface->bo, |
||
4146 | I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER, |
||
4147 | 0); |
||
4148 | OUT_BATCH(batch, y); |
||
4149 | |||
4150 | br13 = 0xf0 << 16; |
||
4151 | br13 |= BR13_565; |
||
4152 | br13 |= pitch; |
||
4153 | |||
4154 | OUT_BATCH(batch, blt_cmd); |
||
4155 | OUT_BATCH(batch, br13); |
||
4156 | OUT_BATCH(batch, |
||
4157 | |||
4158 | 0); |
||
4159 | OUT_BATCH(batch, |
||
4160 | obj_surface->height / 2 << 16 | |
||
4161 | obj_surface->width / 2); |
||
4162 | OUT_RELOC(batch, obj_surface->bo, |
||
4163 | I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER, |
||
4164 | obj_surface->width * obj_surface->y_cb_offset); |
||
4165 | OUT_BATCH(batch, v << 8 | u); |
||
4166 | |||
4167 | ADVANCE_BATCH(batch); |
||
4168 | intel_batchbuffer_end_atomic(batch); |
||
4169 | } |
||
4170 | |||
4171 | VASurfaceID |
||
4172 | i965_post_processing( |
||
4173 | VADriverContextP ctx, |
||
4174 | VASurfaceID surface, |
||
4175 | const VARectangle *src_rect, |
||
4176 | const VARectangle *dst_rect, |
||
4177 | unsigned int flags, |
||
4178 | int *has_done_scaling |
||
4179 | ) |
||
4180 | { |
||
4181 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4182 | VASurfaceID in_surface_id = surface; |
||
4183 | VASurfaceID out_surface_id = VA_INVALID_ID; |
||
4184 | |||
4185 | if (HAS_PP(i965)) { |
||
4186 | /* Currently only support post processing for NV12 surface */ |
||
4187 | if (i965->render_state.interleaved_uv) { |
||
4188 | struct object_surface *obj_surface; |
||
4189 | VAStatus status; |
||
4190 | struct i965_surface src_surface; |
||
4191 | struct i965_surface dst_surface; |
||
4192 | |||
4193 | if (flags & I965_PP_FLAG_DEINTERLACING) { |
||
4194 | obj_surface = SURFACE(in_surface_id); |
||
4195 | status = i965_CreateSurfaces(ctx, |
||
4196 | obj_surface->orig_width, |
||
4197 | obj_surface->orig_height, |
||
4198 | VA_RT_FORMAT_YUV420, |
||
4199 | 1, |
||
4200 | &out_surface_id); |
||
4201 | assert(status == VA_STATUS_SUCCESS); |
||
4202 | obj_surface = SURFACE(out_surface_id); |
||
4203 | i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC('N','V','1','2'), SUBSAMPLE_YUV420); |
||
4204 | |||
4205 | i965_vpp_clear_surface(ctx, i965->pp_context, out_surface_id, 0); |
||
4206 | |||
4207 | src_surface.id = in_surface_id; |
||
4208 | src_surface.type = I965_SURFACE_TYPE_SURFACE; |
||
4209 | src_surface.flags = (flags & I965_PP_FLAG_TOP_FIELD) ? |
||
4210 | I965_SURFACE_FLAG_TOP_FIELD_FIRST : I965_SURFACE_FLAG_BOTTOM_FIELD_FIRST; |
||
4211 | dst_surface.id = out_surface_id; |
||
4212 | dst_surface.type = I965_SURFACE_TYPE_SURFACE; |
||
4213 | dst_surface.flags = I965_SURFACE_FLAG_FRAME; |
||
4214 | |||
4215 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4216 | &src_surface, |
||
4217 | src_rect, |
||
4218 | &dst_surface, |
||
4219 | dst_rect, |
||
4220 | PP_NV12_DNDI, |
||
4221 | NULL); |
||
4222 | printf("Deinterlace is executed here\n"); |
||
4223 | } |
||
4224 | |||
4225 | if (flags & I965_PP_FLAG_AVS) { |
||
4226 | struct i965_render_state *render_state = &i965->render_state; |
||
4227 | struct intel_region *dest_region = render_state->draw_region; |
||
4228 | |||
4229 | if (out_surface_id != VA_INVALID_ID) |
||
4230 | in_surface_id = out_surface_id; |
||
4231 | |||
4232 | status = i965_CreateSurfaces(ctx, |
||
4233 | dest_region->width, |
||
4234 | dest_region->height, |
||
4235 | VA_RT_FORMAT_YUV420, |
||
4236 | 1, |
||
4237 | &out_surface_id); |
||
4238 | assert(status == VA_STATUS_SUCCESS); |
||
4239 | obj_surface = SURFACE(out_surface_id); |
||
4240 | i965_check_alloc_surface_bo(ctx, obj_surface, 0, VA_FOURCC('N','V','1','2'), SUBSAMPLE_YUV420); |
||
4241 | i965_vpp_clear_surface(ctx, i965->pp_context, out_surface_id, 0); |
||
4242 | |||
4243 | src_surface.id = in_surface_id; |
||
4244 | src_surface.type = I965_SURFACE_TYPE_SURFACE; |
||
4245 | src_surface.flags = I965_SURFACE_FLAG_FRAME; |
||
4246 | dst_surface.id = out_surface_id; |
||
4247 | dst_surface.type = I965_SURFACE_TYPE_SURFACE; |
||
4248 | dst_surface.flags = I965_SURFACE_FLAG_FRAME; |
||
4249 | |||
4250 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4251 | &src_surface, |
||
4252 | src_rect, |
||
4253 | &dst_surface, |
||
4254 | dst_rect, |
||
4255 | PP_NV12_AVS, |
||
4256 | NULL); |
||
4257 | |||
4258 | if (in_surface_id != surface) |
||
4259 | i965_DestroySurfaces(ctx, &in_surface_id, 1); |
||
4260 | |||
4261 | *has_done_scaling = 1; |
||
4262 | } |
||
4263 | } |
||
4264 | } |
||
4265 | |||
4266 | return out_surface_id; |
||
4267 | } |
||
4268 | |||
4269 | static VAStatus |
||
4270 | i965_image_pl3_processing(VADriverContextP ctx, |
||
4271 | const struct i965_surface *src_surface, |
||
4272 | const VARectangle *src_rect, |
||
4273 | struct i965_surface *dst_surface, |
||
4274 | const VARectangle *dst_rect) |
||
4275 | { |
||
4276 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4277 | struct i965_post_processing_context *pp_context = i965->pp_context; |
||
4278 | int fourcc = pp_get_surface_fourcc(ctx, dst_surface); |
||
4279 | |||
4280 | if (fourcc == VA_FOURCC('N', 'V', '1', '2')) { |
||
4281 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4282 | src_surface, |
||
4283 | src_rect, |
||
4284 | dst_surface, |
||
4285 | dst_rect, |
||
4286 | PP_PL3_LOAD_SAVE_N12, |
||
4287 | NULL); |
||
4288 | } else if (fourcc == VA_FOURCC('I', 'M', 'C', '1') || |
||
4289 | fourcc == VA_FOURCC('I', 'M', 'C', '3') || |
||
4290 | fourcc == VA_FOURCC('Y', 'V', '1', '2') || |
||
4291 | fourcc == VA_FOURCC('I', '4', '2', '0') ) { |
||
4292 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4293 | src_surface, |
||
4294 | src_rect, |
||
4295 | dst_surface, |
||
4296 | dst_rect, |
||
4297 | PP_PL3_LOAD_SAVE_PL3, |
||
4298 | NULL); |
||
4299 | } else if (fourcc == VA_FOURCC('Y', 'U', 'Y', '2') || |
||
4300 | fourcc == VA_FOURCC('U', 'Y', 'V', 'Y')) { |
||
4301 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4302 | src_surface, |
||
4303 | src_rect, |
||
4304 | dst_surface, |
||
4305 | dst_rect, |
||
4306 | PP_PL3_LOAD_SAVE_PA, |
||
4307 | NULL); |
||
4308 | |||
4309 | } |
||
4310 | else { |
||
4311 | assert(0); |
||
4312 | } |
||
4313 | |||
4314 | intel_batchbuffer_flush(pp_context->batch); |
||
4315 | |||
4316 | return VA_STATUS_SUCCESS; |
||
4317 | } |
||
4318 | |||
4319 | static VAStatus |
||
4320 | i965_image_pl2_processing(VADriverContextP ctx, |
||
4321 | const struct i965_surface *src_surface, |
||
4322 | const VARectangle *src_rect, |
||
4323 | struct i965_surface *dst_surface, |
||
4324 | const VARectangle *dst_rect) |
||
4325 | { |
||
4326 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4327 | struct i965_post_processing_context *pp_context = i965->pp_context; |
||
4328 | int fourcc = pp_get_surface_fourcc(ctx, dst_surface); |
||
4329 | |||
4330 | if (fourcc == VA_FOURCC('N', 'V', '1', '2')) { |
||
4331 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4332 | src_surface, |
||
4333 | src_rect, |
||
4334 | dst_surface, |
||
4335 | dst_rect, |
||
4336 | PP_NV12_LOAD_SAVE_N12, |
||
4337 | NULL); |
||
4338 | } else if (fourcc == VA_FOURCC('I', 'M', 'C', '1') || |
||
4339 | fourcc == VA_FOURCC('I', 'M', 'C', '3') || |
||
4340 | fourcc == VA_FOURCC('Y', 'V', '1', '2') || |
||
4341 | fourcc == VA_FOURCC('I', '4', '2', '0') ) { |
||
4342 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4343 | src_surface, |
||
4344 | src_rect, |
||
4345 | dst_surface, |
||
4346 | dst_rect, |
||
4347 | PP_NV12_LOAD_SAVE_PL3, |
||
4348 | NULL); |
||
4349 | } else if (fourcc == VA_FOURCC('Y', 'U', 'Y', '2') || |
||
4350 | fourcc == VA_FOURCC('U', 'Y', 'V', 'Y')) { |
||
4351 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4352 | src_surface, |
||
4353 | src_rect, |
||
4354 | dst_surface, |
||
4355 | dst_rect, |
||
4356 | PP_NV12_LOAD_SAVE_PA, |
||
4357 | NULL); |
||
4358 | } |
||
4359 | |||
4360 | intel_batchbuffer_flush(pp_context->batch); |
||
4361 | |||
4362 | return VA_STATUS_SUCCESS; |
||
4363 | } |
||
4364 | |||
4365 | static VAStatus |
||
4366 | i965_image_pl1_processing(VADriverContextP ctx, |
||
4367 | const struct i965_surface *src_surface, |
||
4368 | const VARectangle *src_rect, |
||
4369 | struct i965_surface *dst_surface, |
||
4370 | const VARectangle *dst_rect) |
||
4371 | { |
||
4372 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4373 | struct i965_post_processing_context *pp_context = i965->pp_context; |
||
4374 | int fourcc = pp_get_surface_fourcc(ctx, dst_surface); |
||
4375 | |||
4376 | if (fourcc == VA_FOURCC('N', 'V', '1', '2')) { |
||
4377 | i965_post_processing_internal(ctx, i965->pp_context, |
||
4378 | src_surface, |
||
4379 | src_rect, |
||
4380 | dst_surface, |
||
4381 | dst_rect, |
||
4382 | PP_PA_LOAD_SAVE_NV12, |
||
4383 | NULL); |
||
4384 | } |
||
4385 | else { |
||
4386 | return VA_STATUS_ERROR_UNKNOWN; |
||
4387 | } |
||
4388 | |||
4389 | intel_batchbuffer_flush(pp_context->batch); |
||
4390 | |||
4391 | return VA_STATUS_SUCCESS; |
||
4392 | } |
||
4393 | |||
4394 | VAStatus |
||
4395 | i965_image_processing(VADriverContextP ctx, |
||
4396 | const struct i965_surface *src_surface, |
||
4397 | const VARectangle *src_rect, |
||
4398 | struct i965_surface *dst_surface, |
||
4399 | const VARectangle *dst_rect) |
||
4400 | { |
||
4401 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4402 | VAStatus status = VA_STATUS_ERROR_UNIMPLEMENTED; |
||
4403 | |||
4404 | if (HAS_PP(i965)) { |
||
4405 | int fourcc = pp_get_surface_fourcc(ctx, src_surface); |
||
4406 | |||
4407 | switch (fourcc) { |
||
4408 | case VA_FOURCC('Y', 'V', '1', '2'): |
||
4409 | case VA_FOURCC('I', '4', '2', '0'): |
||
4410 | case VA_FOURCC('I', 'M', 'C', '1'): |
||
4411 | case VA_FOURCC('I', 'M', 'C', '3'): |
||
4412 | status = i965_image_pl3_processing(ctx, |
||
4413 | src_surface, |
||
4414 | src_rect, |
||
4415 | dst_surface, |
||
4416 | dst_rect); |
||
4417 | break; |
||
4418 | |||
4419 | case VA_FOURCC('N', 'V', '1', '2'): |
||
4420 | status = i965_image_pl2_processing(ctx, |
||
4421 | src_surface, |
||
4422 | src_rect, |
||
4423 | dst_surface, |
||
4424 | dst_rect); |
||
4425 | break; |
||
4426 | case VA_FOURCC('Y', 'U', 'Y', '2'): |
||
4427 | case VA_FOURCC('U', 'Y', 'V', 'Y'): |
||
4428 | status = i965_image_pl1_processing(ctx, |
||
4429 | src_surface, |
||
4430 | src_rect, |
||
4431 | dst_surface, |
||
4432 | dst_rect); |
||
4433 | break; |
||
4434 | |||
4435 | default: |
||
4436 | status = VA_STATUS_ERROR_UNIMPLEMENTED; |
||
4437 | break; |
||
4438 | } |
||
4439 | } |
||
4440 | |||
4441 | return status; |
||
4442 | } |
||
4443 | |||
4444 | |||
4445 | static void |
||
4446 | i965_post_processing_context_finalize(struct i965_post_processing_context *pp_context) |
||
4447 | { |
||
4448 | int i; |
||
4449 | |||
4450 | dri_bo_unreference(pp_context->surface_state_binding_table.bo); |
||
4451 | pp_context->surface_state_binding_table.bo = NULL; |
||
4452 | |||
4453 | dri_bo_unreference(pp_context->curbe.bo); |
||
4454 | pp_context->curbe.bo = NULL; |
||
4455 | |||
4456 | dri_bo_unreference(pp_context->sampler_state_table.bo); |
||
4457 | pp_context->sampler_state_table.bo = NULL; |
||
4458 | |||
4459 | dri_bo_unreference(pp_context->sampler_state_table.bo_8x8); |
||
4460 | pp_context->sampler_state_table.bo_8x8 = NULL; |
||
4461 | |||
4462 | dri_bo_unreference(pp_context->sampler_state_table.bo_8x8_uv); |
||
4463 | pp_context->sampler_state_table.bo_8x8_uv = NULL; |
||
4464 | |||
4465 | dri_bo_unreference(pp_context->idrt.bo); |
||
4466 | pp_context->idrt.bo = NULL; |
||
4467 | pp_context->idrt.num_interface_descriptors = 0; |
||
4468 | |||
4469 | dri_bo_unreference(pp_context->vfe_state.bo); |
||
4470 | pp_context->vfe_state.bo = NULL; |
||
4471 | |||
4472 | dri_bo_unreference(pp_context->stmm.bo); |
||
4473 | pp_context->stmm.bo = NULL; |
||
4474 | |||
4475 | for (i = 0; i < NUM_PP_MODULES; i++) { |
||
4476 | struct pp_module *pp_module = &pp_context->pp_modules[i]; |
||
4477 | |||
4478 | dri_bo_unreference(pp_module->kernel.bo); |
||
4479 | pp_module->kernel.bo = NULL; |
||
4480 | } |
||
4481 | |||
4482 | free(pp_context->pp_static_parameter); |
||
4483 | free(pp_context->pp_inline_parameter); |
||
4484 | pp_context->pp_static_parameter = NULL; |
||
4485 | pp_context->pp_inline_parameter = NULL; |
||
4486 | } |
||
4487 | |||
4488 | Bool |
||
4489 | i965_post_processing_terminate(VADriverContextP ctx) |
||
4490 | { |
||
4491 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4492 | struct i965_post_processing_context *pp_context = i965->pp_context; |
||
4493 | |||
4494 | if (pp_context) { |
||
4495 | if(IS_HASWELL(i965->intel.device_id)){ |
||
4496 | gen75_vebox_context_destroy(ctx, pp_context->pp_vebox_context); |
||
4497 | } |
||
4498 | |||
4499 | i965_post_processing_context_finalize(pp_context); |
||
4500 | free(pp_context); |
||
4501 | } |
||
4502 | |||
4503 | i965->pp_context = NULL; |
||
4504 | |||
4505 | return True; |
||
4506 | } |
||
4507 | |||
4508 | static void |
||
4509 | i965_post_processing_context_init(VADriverContextP ctx, |
||
4510 | struct i965_post_processing_context *pp_context, |
||
4511 | struct intel_batchbuffer *batch) |
||
4512 | { |
||
4513 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4514 | int i; |
||
4515 | |||
4516 | printf("%s\n", __FUNCTION__); |
||
4517 | |||
4518 | pp_context->urb.size = URB_SIZE((&i965->intel)); |
||
4519 | pp_context->urb.num_vfe_entries = 32; |
||
4520 | pp_context->urb.size_vfe_entry = 1; /* in 512 bits unit */ |
||
4521 | pp_context->urb.num_cs_entries = 1; |
||
4522 | |||
4523 | if (IS_GEN7(i965->intel.device_id)) |
||
4524 | pp_context->urb.size_cs_entry = 4; /* in 512 bits unit */ |
||
4525 | else |
||
4526 | pp_context->urb.size_cs_entry = 2; |
||
4527 | |||
4528 | pp_context->urb.vfe_start = 0; |
||
4529 | pp_context->urb.cs_start = pp_context->urb.vfe_start + |
||
4530 | pp_context->urb.num_vfe_entries * pp_context->urb.size_vfe_entry; |
||
4531 | assert(pp_context->urb.cs_start + |
||
4532 | pp_context->urb.num_cs_entries * pp_context->urb.size_cs_entry <= URB_SIZE((&i965->intel))); |
||
4533 | |||
4534 | assert(NUM_PP_MODULES == ARRAY_ELEMS(pp_modules_gen5)); |
||
4535 | assert(NUM_PP_MODULES == ARRAY_ELEMS(pp_modules_gen6)); |
||
4536 | assert(NUM_PP_MODULES == ARRAY_ELEMS(pp_modules_gen7)); |
||
4537 | assert(NUM_PP_MODULES == ARRAY_ELEMS(pp_modules_gen75)); |
||
4538 | |||
4539 | if (IS_HASWELL(i965->intel.device_id)) |
||
4540 | memcpy(pp_context->pp_modules, pp_modules_gen75, sizeof(pp_context->pp_modules)); |
||
4541 | else if (IS_GEN7(i965->intel.device_id)) |
||
4542 | memcpy(pp_context->pp_modules, pp_modules_gen7, sizeof(pp_context->pp_modules)); |
||
4543 | else if (IS_GEN6(i965->intel.device_id)) |
||
4544 | memcpy(pp_context->pp_modules, pp_modules_gen6, sizeof(pp_context->pp_modules)); |
||
4545 | else if (IS_IRONLAKE(i965->intel.device_id)) |
||
4546 | memcpy(pp_context->pp_modules, pp_modules_gen5, sizeof(pp_context->pp_modules)); |
||
4547 | |||
4548 | for (i = 0; i < NUM_PP_MODULES; i++) { |
||
4549 | struct pp_module *pp_module = &pp_context->pp_modules[i]; |
||
4550 | dri_bo_unreference(pp_module->kernel.bo); |
||
4551 | if (pp_module->kernel.bin && pp_module->kernel.size) { |
||
4552 | pp_module->kernel.bo = dri_bo_alloc(i965->intel.bufmgr, |
||
4553 | pp_module->kernel.name, |
||
4554 | pp_module->kernel.size, |
||
4555 | 4096); |
||
4556 | assert(pp_module->kernel.bo); |
||
4557 | dri_bo_subdata(pp_module->kernel.bo, 0, pp_module->kernel.size, pp_module->kernel.bin); |
||
4558 | } else { |
||
4559 | pp_module->kernel.bo = NULL; |
||
4560 | } |
||
4561 | } |
||
4562 | |||
4563 | /* static & inline parameters */ |
||
4564 | if (IS_GEN7(i965->intel.device_id)) { |
||
4565 | pp_context->pp_static_parameter = calloc(sizeof(struct gen7_pp_static_parameter), 1); |
||
4566 | pp_context->pp_inline_parameter = calloc(sizeof(struct gen7_pp_inline_parameter), 1); |
||
4567 | } else { |
||
4568 | pp_context->pp_static_parameter = calloc(sizeof(struct pp_static_parameter), 1); |
||
4569 | pp_context->pp_inline_parameter = calloc(sizeof(struct pp_inline_parameter), 1); |
||
4570 | } |
||
4571 | |||
4572 | pp_context->batch = batch; |
||
4573 | } |
||
4574 | |||
4575 | Bool |
||
4576 | i965_post_processing_init(VADriverContextP ctx) |
||
4577 | { |
||
4578 | struct i965_driver_data *i965 = i965_driver_data(ctx); |
||
4579 | struct i965_post_processing_context *pp_context = i965->pp_context; |
||
4580 | |||
4581 | printf("%s\n", __FUNCTION__); |
||
4582 | |||
4583 | if (HAS_PP(i965)) { |
||
4584 | if (pp_context == NULL) { |
||
4585 | pp_context = calloc(1, sizeof(*pp_context)); |
||
4586 | i965_post_processing_context_init(ctx, pp_context, i965->batch); |
||
4587 | i965->pp_context = pp_context; |
||
4588 | |||
4589 | if(IS_HASWELL(i965->intel.device_id)){ |
||
4590 | pp_context->pp_vebox_context = gen75_vebox_context_init(ctx); |
||
4591 | } |
||
4592 | } |
||
4593 | } |
||
4594 | printf("%s done\n", __FUNCTION__); |
||
4595 | |||
4596 | return True; |
||
4597 | }>=>>><>><>><>><>><>><>><>><>>>=>><>><>><>><>>>>>=>=>>>>><>><> |
||
4598 | |||
4599 |