1 /*
2 * Copyright © 2017 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 "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included
12 * in all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
15 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20 * DEALINGS IN THE SOFTWARE.
21 */
22
23 /**
24 * @file crocus_resource.c
25 *
26 * Resources are images, buffers, and other objects used by the GPU.
27 *
28 * XXX: explain resources
29 */
30
31 #include <stdio.h>
32 #include <errno.h>
33 #include "pipe/p_defines.h"
34 #include "pipe/p_state.h"
35 #include "pipe/p_context.h"
36 #include "pipe/p_screen.h"
37 #include "util/os_memory.h"
38 #include "util/u_cpu_detect.h"
39 #include "util/u_inlines.h"
40 #include "util/format/u_format.h"
41 #include "util/u_threaded_context.h"
42 #include "util/u_transfer.h"
43 #include "util/u_transfer_helper.h"
44 #include "util/u_upload_mgr.h"
45 #include "util/ralloc.h"
46 #include "util/u_memory.h"
47 #include "crocus_batch.h"
48 #include "crocus_context.h"
49 #include "crocus_resource.h"
50 #include "crocus_screen.h"
51 #include "intel/dev/intel_debug.h"
52 #include "isl/isl.h"
53 #include "drm-uapi/drm_fourcc.h"
54 #include "drm-uapi/i915_drm.h"
55
56 enum modifier_priority {
57 MODIFIER_PRIORITY_INVALID = 0,
58 MODIFIER_PRIORITY_LINEAR,
59 MODIFIER_PRIORITY_X,
60 MODIFIER_PRIORITY_Y,
61 MODIFIER_PRIORITY_Y_CCS,
62 };
63
64 static const uint64_t priority_to_modifier[] = {
65 [MODIFIER_PRIORITY_INVALID] = DRM_FORMAT_MOD_INVALID,
66 [MODIFIER_PRIORITY_LINEAR] = DRM_FORMAT_MOD_LINEAR,
67 [MODIFIER_PRIORITY_X] = I915_FORMAT_MOD_X_TILED,
68 [MODIFIER_PRIORITY_Y] = I915_FORMAT_MOD_Y_TILED,
69 [MODIFIER_PRIORITY_Y_CCS] = I915_FORMAT_MOD_Y_TILED_CCS,
70 };
71
72 static bool
modifier_is_supported(const struct intel_device_info * devinfo,enum pipe_format pfmt,unsigned bind,uint64_t modifier)73 modifier_is_supported(const struct intel_device_info *devinfo,
74 enum pipe_format pfmt, unsigned bind,
75 uint64_t modifier)
76 {
77 /* XXX: do something real */
78 switch (modifier) {
79 case I915_FORMAT_MOD_Y_TILED_CCS:
80 return false;
81 case I915_FORMAT_MOD_Y_TILED:
82 if (bind & PIPE_BIND_SCANOUT)
83 return false;
84 return devinfo->ver >= 6;
85 case I915_FORMAT_MOD_X_TILED:
86 case DRM_FORMAT_MOD_LINEAR:
87 return true;
88 case DRM_FORMAT_MOD_INVALID:
89 default:
90 return false;
91 }
92 }
93
94 static uint64_t
select_best_modifier(struct intel_device_info * devinfo,const struct pipe_resource * templ,const uint64_t * modifiers,int count)95 select_best_modifier(struct intel_device_info *devinfo,
96 const struct pipe_resource *templ,
97 const uint64_t *modifiers,
98 int count)
99 {
100 enum modifier_priority prio = MODIFIER_PRIORITY_INVALID;
101
102 for (int i = 0; i < count; i++) {
103 if (!modifier_is_supported(devinfo, templ->format, templ->bind,
104 modifiers[i]))
105 continue;
106
107 switch (modifiers[i]) {
108 case I915_FORMAT_MOD_Y_TILED_CCS:
109 prio = MAX2(prio, MODIFIER_PRIORITY_Y_CCS);
110 break;
111 case I915_FORMAT_MOD_Y_TILED:
112 prio = MAX2(prio, MODIFIER_PRIORITY_Y);
113 break;
114 case I915_FORMAT_MOD_X_TILED:
115 prio = MAX2(prio, MODIFIER_PRIORITY_X);
116 break;
117 case DRM_FORMAT_MOD_LINEAR:
118 prio = MAX2(prio, MODIFIER_PRIORITY_LINEAR);
119 break;
120 case DRM_FORMAT_MOD_INVALID:
121 default:
122 break;
123 }
124 }
125
126 return priority_to_modifier[prio];
127 }
128
129 static enum isl_surf_dim
crocus_target_to_isl_surf_dim(enum pipe_texture_target target)130 crocus_target_to_isl_surf_dim(enum pipe_texture_target target)
131 {
132 switch (target) {
133 case PIPE_BUFFER:
134 case PIPE_TEXTURE_1D:
135 case PIPE_TEXTURE_1D_ARRAY:
136 return ISL_SURF_DIM_1D;
137 case PIPE_TEXTURE_2D:
138 case PIPE_TEXTURE_CUBE:
139 case PIPE_TEXTURE_RECT:
140 case PIPE_TEXTURE_2D_ARRAY:
141 case PIPE_TEXTURE_CUBE_ARRAY:
142 return ISL_SURF_DIM_2D;
143 case PIPE_TEXTURE_3D:
144 return ISL_SURF_DIM_3D;
145 case PIPE_MAX_TEXTURE_TYPES:
146 break;
147 }
148 unreachable("invalid texture type");
149 }
150
151 static isl_surf_usage_flags_t
pipe_bind_to_isl_usage(unsigned bindings)152 pipe_bind_to_isl_usage(unsigned bindings)
153 {
154 isl_surf_usage_flags_t usage = 0;
155
156 if (bindings & PIPE_BIND_RENDER_TARGET)
157 usage |= ISL_SURF_USAGE_RENDER_TARGET_BIT;
158
159 if (bindings & PIPE_BIND_SAMPLER_VIEW)
160 usage |= ISL_SURF_USAGE_TEXTURE_BIT;
161
162 if (bindings & (PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SHADER_BUFFER))
163 usage |= ISL_SURF_USAGE_STORAGE_BIT;
164
165 if (bindings & PIPE_BIND_SCANOUT)
166 usage |= ISL_SURF_USAGE_DISPLAY_BIT;
167 return usage;
168 }
169
170 static bool
crocus_resource_configure_main(const struct crocus_screen * screen,struct crocus_resource * res,const struct pipe_resource * templ,uint64_t modifier,uint32_t row_pitch_B)171 crocus_resource_configure_main(const struct crocus_screen *screen,
172 struct crocus_resource *res,
173 const struct pipe_resource *templ,
174 uint64_t modifier, uint32_t row_pitch_B)
175 {
176 const struct intel_device_info *devinfo = &screen->devinfo;
177 const struct util_format_description *format_desc =
178 util_format_description(templ->format);
179 const bool has_depth = util_format_has_depth(format_desc);
180 isl_surf_usage_flags_t usage = pipe_bind_to_isl_usage(templ->bind);
181 isl_tiling_flags_t tiling_flags = ISL_TILING_ANY_MASK;
182
183 /* TODO: This used to be because there wasn't BLORP to handle Y-tiling. */
184 if (devinfo->ver < 6 && !util_format_is_depth_or_stencil(templ->format))
185 tiling_flags &= ~ISL_TILING_Y0_BIT;
186
187 if (modifier != DRM_FORMAT_MOD_INVALID) {
188 res->mod_info = isl_drm_modifier_get_info(modifier);
189
190 tiling_flags = 1 << res->mod_info->tiling;
191 } else {
192 if (templ->bind & PIPE_BIND_RENDER_TARGET && devinfo->ver < 6) {
193 modifier = I915_FORMAT_MOD_X_TILED;
194 res->mod_info = isl_drm_modifier_get_info(modifier);
195 tiling_flags = 1 << res->mod_info->tiling;
196 }
197 /* Use linear for staging buffers */
198 if (templ->usage == PIPE_USAGE_STAGING ||
199 templ->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR) )
200 tiling_flags = ISL_TILING_LINEAR_BIT;
201 else if (templ->bind & PIPE_BIND_SCANOUT)
202 tiling_flags = screen->devinfo.has_tiling_uapi ?
203 ISL_TILING_X_BIT : ISL_TILING_LINEAR_BIT;
204 }
205
206 if (templ->target == PIPE_TEXTURE_CUBE ||
207 templ->target == PIPE_TEXTURE_CUBE_ARRAY)
208 usage |= ISL_SURF_USAGE_CUBE_BIT;
209
210 if (templ->usage != PIPE_USAGE_STAGING) {
211 if (templ->format == PIPE_FORMAT_S8_UINT)
212 usage |= ISL_SURF_USAGE_STENCIL_BIT;
213 else if (has_depth) {
214 /* combined DS only on gen4/5 */
215 if (devinfo->ver < 6) {
216 if (templ->format == PIPE_FORMAT_Z24X8_UNORM ||
217 templ->format == PIPE_FORMAT_Z24_UNORM_S8_UINT ||
218 templ->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)
219 usage |= ISL_SURF_USAGE_STENCIL_BIT;
220 }
221 usage |= ISL_SURF_USAGE_DEPTH_BIT;
222 }
223
224 if (templ->format == PIPE_FORMAT_S8_UINT)
225 tiling_flags = ISL_TILING_W_BIT;
226 }
227
228 const enum isl_format format =
229 crocus_format_for_usage(&screen->devinfo, templ->format, usage).fmt;
230
231 if (row_pitch_B == 0 && templ->usage == PIPE_USAGE_STAGING &&
232 templ->target == PIPE_TEXTURE_2D &&
233 devinfo->ver < 6) {
234 /* align row pitch to 4 so we can keep using BLT engine */
235 row_pitch_B = util_format_get_stride(templ->format, templ->width0);
236 row_pitch_B = ALIGN(row_pitch_B, 4);
237 }
238
239 const struct isl_surf_init_info init_info = {
240 .dim = crocus_target_to_isl_surf_dim(templ->target),
241 .format = format,
242 .width = templ->width0,
243 .height = templ->height0,
244 .depth = templ->depth0,
245 .levels = templ->last_level + 1,
246 .array_len = templ->array_size,
247 .samples = MAX2(templ->nr_samples, 1),
248 .min_alignment_B = 0,
249 .row_pitch_B = row_pitch_B,
250 .usage = usage,
251 .tiling_flags = tiling_flags
252 };
253
254 if (!isl_surf_init_s(&screen->isl_dev, &res->surf, &init_info))
255 return false;
256
257 /*
258 * Don't create staging surfaces that will use > half the aperture
259 * since staging implies you are sending to another resource,
260 * which there is no way to fit both into aperture.
261 */
262 if (templ->usage == PIPE_USAGE_STAGING)
263 if (res->surf.size_B > screen->aperture_threshold / 2)
264 return false;
265
266 res->internal_format = templ->format;
267
268 return true;
269 }
270
271 static void
crocus_query_dmabuf_modifiers(struct pipe_screen * pscreen,enum pipe_format pfmt,int max,uint64_t * modifiers,unsigned int * external_only,int * count)272 crocus_query_dmabuf_modifiers(struct pipe_screen *pscreen,
273 enum pipe_format pfmt,
274 int max,
275 uint64_t *modifiers,
276 unsigned int *external_only,
277 int *count)
278 {
279 struct crocus_screen *screen = (void *) pscreen;
280 const struct intel_device_info *devinfo = &screen->devinfo;
281
282 uint64_t all_modifiers[] = {
283 DRM_FORMAT_MOD_LINEAR,
284 I915_FORMAT_MOD_X_TILED,
285 I915_FORMAT_MOD_Y_TILED,
286 I915_FORMAT_MOD_Y_TILED_CCS,
287 };
288
289 int supported_mods = 0;
290
291 for (int i = 0; i < ARRAY_SIZE(all_modifiers); i++) {
292 if (!modifier_is_supported(devinfo, pfmt, 0, all_modifiers[i]))
293 continue;
294
295 if (supported_mods < max) {
296 if (modifiers)
297 modifiers[supported_mods] = all_modifiers[i];
298
299 if (external_only)
300 external_only[supported_mods] = util_format_is_yuv(pfmt);
301 }
302
303 supported_mods++;
304 }
305
306 *count = supported_mods;
307 }
308
309 static struct pipe_resource *
crocus_resource_get_separate_stencil(struct pipe_resource * p_res)310 crocus_resource_get_separate_stencil(struct pipe_resource *p_res)
311 {
312 return _crocus_resource_get_separate_stencil(p_res);
313 }
314
315 static void
crocus_resource_set_separate_stencil(struct pipe_resource * p_res,struct pipe_resource * stencil)316 crocus_resource_set_separate_stencil(struct pipe_resource *p_res,
317 struct pipe_resource *stencil)
318 {
319 assert(util_format_has_depth(util_format_description(p_res->format)));
320 pipe_resource_reference(&p_res->next, stencil);
321 }
322
323 void
crocus_resource_disable_aux(struct crocus_resource * res)324 crocus_resource_disable_aux(struct crocus_resource *res)
325 {
326 crocus_bo_unreference(res->aux.bo);
327 free(res->aux.state);
328
329 res->aux.usage = ISL_AUX_USAGE_NONE;
330 res->aux.has_hiz = 0;
331 res->aux.surf.size_B = 0;
332 res->aux.surf.levels = 0;
333 res->aux.bo = NULL;
334 res->aux.state = NULL;
335 }
336
337 static void
crocus_resource_destroy(struct pipe_screen * screen,struct pipe_resource * resource)338 crocus_resource_destroy(struct pipe_screen *screen,
339 struct pipe_resource *resource)
340 {
341 struct crocus_resource *res = (struct crocus_resource *)resource;
342
343 if (resource->target == PIPE_BUFFER)
344 util_range_destroy(&res->valid_buffer_range);
345
346 if (res->shadow)
347 pipe_resource_reference((struct pipe_resource **)&res->shadow, NULL);
348 crocus_resource_disable_aux(res);
349
350 threaded_resource_deinit(resource);
351 crocus_bo_unreference(res->bo);
352 crocus_pscreen_unref(res->orig_screen);
353 free(res);
354 }
355
356 static struct crocus_resource *
crocus_alloc_resource(struct pipe_screen * pscreen,const struct pipe_resource * templ)357 crocus_alloc_resource(struct pipe_screen *pscreen,
358 const struct pipe_resource *templ)
359 {
360 struct crocus_resource *res = calloc(1, sizeof(struct crocus_resource));
361 if (!res)
362 return NULL;
363
364 res->base.b = *templ;
365 res->base.b.screen = pscreen;
366 res->orig_screen = crocus_pscreen_ref(pscreen);
367 pipe_reference_init(&res->base.b.reference, 1);
368 threaded_resource_init(&res->base.b, false);
369
370 if (templ->target == PIPE_BUFFER)
371 util_range_init(&res->valid_buffer_range);
372
373 return res;
374 }
375
376 unsigned
crocus_get_num_logical_layers(const struct crocus_resource * res,unsigned level)377 crocus_get_num_logical_layers(const struct crocus_resource *res, unsigned level)
378 {
379 if (res->surf.dim == ISL_SURF_DIM_3D)
380 return u_minify(res->surf.logical_level0_px.depth, level);
381 else
382 return res->surf.logical_level0_px.array_len;
383 }
384
385 static enum isl_aux_state **
create_aux_state_map(struct crocus_resource * res,enum isl_aux_state initial)386 create_aux_state_map(struct crocus_resource *res, enum isl_aux_state initial)
387 {
388 assert(res->aux.state == NULL);
389
390 uint32_t total_slices = 0;
391 for (uint32_t level = 0; level < res->surf.levels; level++)
392 total_slices += crocus_get_num_logical_layers(res, level);
393
394 const size_t per_level_array_size =
395 res->surf.levels * sizeof(enum isl_aux_state *);
396
397 /* We're going to allocate a single chunk of data for both the per-level
398 * reference array and the arrays of aux_state. This makes cleanup
399 * significantly easier.
400 */
401 const size_t total_size =
402 per_level_array_size + total_slices * sizeof(enum isl_aux_state);
403
404 void *data = malloc(total_size);
405 if (!data)
406 return NULL;
407
408 enum isl_aux_state **per_level_arr = data;
409 enum isl_aux_state *s = data + per_level_array_size;
410 for (uint32_t level = 0; level < res->surf.levels; level++) {
411 per_level_arr[level] = s;
412 const unsigned level_layers = crocus_get_num_logical_layers(res, level);
413 for (uint32_t a = 0; a < level_layers; a++)
414 *(s++) = initial;
415 }
416 assert((void *)s == data + total_size);
417
418 return per_level_arr;
419 }
420
421 /**
422 * Configure aux for the resource, but don't allocate it. For images which
423 * might be shared with modifiers, we must allocate the image and aux data in
424 * a single bo.
425 *
426 * Returns false on unexpected error (e.g. allocation failed, or invalid
427 * configuration result).
428 */
429 static bool
crocus_resource_configure_aux(struct crocus_screen * screen,struct crocus_resource * res,bool imported,uint64_t * aux_size_B,uint32_t * alloc_flags)430 crocus_resource_configure_aux(struct crocus_screen *screen,
431 struct crocus_resource *res, bool imported,
432 uint64_t *aux_size_B,
433 uint32_t *alloc_flags)
434 {
435 const struct intel_device_info *devinfo = &screen->devinfo;
436
437 /* Try to create the auxiliary surfaces allowed by the modifier or by
438 * the user if no modifier is specified.
439 */
440 assert(!res->mod_info || res->mod_info->aux_usage == ISL_AUX_USAGE_NONE);
441
442 const bool has_mcs = devinfo->ver >= 7 && !res->mod_info &&
443 isl_surf_get_mcs_surf(&screen->isl_dev, &res->surf, &res->aux.surf);
444
445 const bool has_hiz = devinfo->ver >= 6 && !res->mod_info &&
446 !INTEL_DEBUG(DEBUG_NO_HIZ) &&
447 isl_surf_get_hiz_surf(&screen->isl_dev, &res->surf, &res->aux.surf);
448
449 const bool has_ccs =
450 ((devinfo->ver >= 7 && !res->mod_info && !INTEL_DEBUG(DEBUG_NO_CCS)) ||
451 (res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE)) &&
452 isl_surf_get_ccs_surf(&screen->isl_dev, &res->surf, NULL,
453 &res->aux.surf, 0);
454
455 /* Having more than one type of compression is impossible */
456 assert(has_ccs + has_mcs + has_hiz <= 1);
457
458 if (res->mod_info && has_ccs) {
459 res->aux.usage = res->mod_info->aux_usage;
460 } else if (has_mcs) {
461 res->aux.usage = ISL_AUX_USAGE_MCS;
462 } else if (has_hiz) {
463 res->aux.usage = ISL_AUX_USAGE_HIZ;
464 } else if (has_ccs) {
465 if (isl_format_supports_ccs_d(devinfo, res->surf.format))
466 res->aux.usage = ISL_AUX_USAGE_CCS_D;
467 }
468
469 enum isl_aux_state initial_state = ISL_AUX_STATE_AUX_INVALID;
470 *aux_size_B = 0;
471 *alloc_flags = 0;
472 assert(!res->aux.bo);
473
474 switch (res->aux.usage) {
475 case ISL_AUX_USAGE_NONE:
476 /* Having no aux buffer is only okay if there's no modifier with aux. */
477 res->aux.surf.levels = 0;
478 return !res->mod_info || res->mod_info->aux_usage == ISL_AUX_USAGE_NONE;
479 case ISL_AUX_USAGE_HIZ:
480 initial_state = ISL_AUX_STATE_AUX_INVALID;
481 break;
482 case ISL_AUX_USAGE_MCS:
483 /* The Ivybridge PRM, Vol 2 Part 1 p326 says:
484 *
485 * "When MCS buffer is enabled and bound to MSRT, it is required
486 * that it is cleared prior to any rendering."
487 *
488 * Since we only use the MCS buffer for rendering, we just clear it
489 * immediately on allocation. The clear value for MCS buffers is all
490 * 1's, so we simply memset it to 0xff.
491 */
492 initial_state = ISL_AUX_STATE_CLEAR;
493 break;
494 case ISL_AUX_USAGE_CCS_D:
495 /* When CCS_E is used, we need to ensure that the CCS starts off in
496 * a valid state. From the Sky Lake PRM, "MCS Buffer for Render
497 * Target(s)":
498 *
499 * "If Software wants to enable Color Compression without Fast
500 * clear, Software needs to initialize MCS with zeros."
501 *
502 * A CCS value of 0 indicates that the corresponding block is in the
503 * pass-through state which is what we want.
504 *
505 * For CCS_D, do the same thing. On Gen9+, this avoids having any
506 * undefined bits in the aux buffer.
507 */
508 if (imported)
509 initial_state =
510 isl_drm_modifier_get_default_aux_state(res->mod_info->modifier);
511 else
512 initial_state = ISL_AUX_STATE_PASS_THROUGH;
513 *alloc_flags |= BO_ALLOC_ZEROED;
514 break;
515 default:
516 unreachable("non-crocus aux");
517 }
518
519 /* Create the aux_state for the auxiliary buffer. */
520 res->aux.state = create_aux_state_map(res, initial_state);
521 if (!res->aux.state)
522 return false;
523
524 /* Increase the aux offset if the main and aux surfaces will share a BO. */
525 res->aux.offset =
526 !res->mod_info || res->mod_info->aux_usage == res->aux.usage ?
527 ALIGN(res->surf.size_B, res->aux.surf.alignment_B) : 0;
528 uint64_t size = res->aux.surf.size_B;
529
530 /* Allocate space in the buffer for storing the clear color. On modern
531 * platforms (gen > 9), we can read it directly from such buffer.
532 *
533 * On gen <= 9, we are going to store the clear color on the buffer
534 * anyways, and copy it back to the surface state during state emission.
535 *
536 * Also add some padding to make sure the fast clear color state buffer
537 * starts at a 4K alignment. We believe that 256B might be enough, but due
538 * to lack of testing we will leave this as 4K for now.
539 */
540 size = ALIGN(size, 4096);
541 *aux_size_B = size;
542
543 if (isl_aux_usage_has_hiz(res->aux.usage)) {
544 for (unsigned level = 0; level < res->surf.levels; ++level) {
545 uint32_t width = u_minify(res->surf.phys_level0_sa.width, level);
546 uint32_t height = u_minify(res->surf.phys_level0_sa.height, level);
547
548 /* Disable HiZ for LOD > 0 unless the width/height are 8x4 aligned.
549 * For LOD == 0, we can grow the dimensions to make it work.
550 */
551 if (devinfo->verx10 < 75 ||
552 (level == 0 || ((width & 7) == 0 && (height & 3) == 0)))
553 res->aux.has_hiz |= 1 << level;
554 }
555 }
556
557 return true;
558 }
559
560 /**
561 * Initialize the aux buffer contents.
562 *
563 * Returns false on unexpected error (e.g. mapping a BO failed).
564 */
565 static bool
crocus_resource_init_aux_buf(struct crocus_resource * res,uint32_t alloc_flags)566 crocus_resource_init_aux_buf(struct crocus_resource *res, uint32_t alloc_flags)
567 {
568 if (!(alloc_flags & BO_ALLOC_ZEROED)) {
569 void *map = crocus_bo_map(NULL, res->aux.bo, MAP_WRITE | MAP_RAW);
570
571 if (!map)
572 return false;
573
574 if (crocus_resource_get_aux_state(res, 0, 0) != ISL_AUX_STATE_AUX_INVALID) {
575 uint8_t memset_value = isl_aux_usage_has_mcs(res->aux.usage) ? 0xFF : 0;
576 memset((char*)map + res->aux.offset, memset_value,
577 res->aux.surf.size_B);
578 }
579
580 crocus_bo_unmap(res->aux.bo);
581 }
582
583 return true;
584 }
585
586 /**
587 * Allocate the initial aux surface for a resource based on aux.usage
588 *
589 * Returns false on unexpected error (e.g. allocation failed, or invalid
590 * configuration result).
591 */
592 static bool
crocus_resource_alloc_separate_aux(struct crocus_screen * screen,struct crocus_resource * res)593 crocus_resource_alloc_separate_aux(struct crocus_screen *screen,
594 struct crocus_resource *res)
595 {
596 uint32_t alloc_flags;
597 uint64_t size;
598 if (!crocus_resource_configure_aux(screen, res, false, &size, &alloc_flags))
599 return false;
600
601 if (size == 0)
602 return true;
603
604 /* Allocate the auxiliary buffer. ISL has stricter set of alignment rules
605 * the drm allocator. Therefore, one can pass the ISL dimensions in terms
606 * of bytes instead of trying to recalculate based on different format
607 * block sizes.
608 */
609 res->aux.bo = crocus_bo_alloc_tiled(screen->bufmgr, "aux buffer", size, 4096,
610 isl_tiling_to_i915_tiling(res->aux.surf.tiling),
611 res->aux.surf.row_pitch_B, alloc_flags);
612 if (!res->aux.bo) {
613 return false;
614 }
615
616 if (!crocus_resource_init_aux_buf(res, alloc_flags))
617 return false;
618
619 return true;
620 }
621
622 void
crocus_resource_finish_aux_import(struct pipe_screen * pscreen,struct crocus_resource * res)623 crocus_resource_finish_aux_import(struct pipe_screen *pscreen,
624 struct crocus_resource *res)
625 {
626 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
627 assert(crocus_resource_unfinished_aux_import(res));
628 assert(!res->mod_info->supports_clear_color);
629
630 struct crocus_resource *aux_res = (void *) res->base.b.next;
631 assert(aux_res->aux.surf.row_pitch_B && aux_res->aux.offset &&
632 aux_res->aux.bo);
633
634 assert(res->bo == aux_res->aux.bo);
635 crocus_bo_reference(aux_res->aux.bo);
636 res->aux.bo = aux_res->aux.bo;
637
638 res->aux.offset = aux_res->aux.offset;
639
640 assert(res->bo->size >= (res->aux.offset + res->aux.surf.size_B));
641 assert(aux_res->aux.surf.row_pitch_B == res->aux.surf.row_pitch_B);
642
643 crocus_resource_destroy(&screen->base, res->base.b.next);
644 res->base.b.next = NULL;
645 }
646
647 static struct pipe_resource *
crocus_resource_create_for_buffer(struct pipe_screen * pscreen,const struct pipe_resource * templ)648 crocus_resource_create_for_buffer(struct pipe_screen *pscreen,
649 const struct pipe_resource *templ)
650 {
651 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
652 struct crocus_resource *res = crocus_alloc_resource(pscreen, templ);
653
654 assert(templ->target == PIPE_BUFFER);
655 assert(templ->height0 <= 1);
656 assert(templ->depth0 <= 1);
657 assert(templ->format == PIPE_FORMAT_NONE ||
658 util_format_get_blocksize(templ->format) == 1);
659
660 res->internal_format = templ->format;
661 res->surf.tiling = ISL_TILING_LINEAR;
662
663 const char *name = templ->target == PIPE_BUFFER ? "buffer" : "miptree";
664
665 res->bo = crocus_bo_alloc(screen->bufmgr, name, templ->width0);
666 if (!res->bo) {
667 crocus_resource_destroy(pscreen, &res->base.b);
668 return NULL;
669 }
670
671 return &res->base.b;
672 }
673
674 static struct pipe_resource *
crocus_resource_create_with_modifiers(struct pipe_screen * pscreen,const struct pipe_resource * templ,const uint64_t * modifiers,int modifiers_count)675 crocus_resource_create_with_modifiers(struct pipe_screen *pscreen,
676 const struct pipe_resource *templ,
677 const uint64_t *modifiers,
678 int modifiers_count)
679 {
680 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
681 struct intel_device_info *devinfo = &screen->devinfo;
682 struct crocus_resource *res = crocus_alloc_resource(pscreen, templ);
683
684 if (!res)
685 return NULL;
686
687 uint64_t modifier =
688 select_best_modifier(devinfo, templ, modifiers, modifiers_count);
689
690 if (modifier == DRM_FORMAT_MOD_INVALID && modifiers_count > 0) {
691 fprintf(stderr, "Unsupported modifier, resource creation failed.\n");
692 goto fail;
693 }
694
695 if (templ->usage == PIPE_USAGE_STAGING &&
696 templ->bind == PIPE_BIND_DEPTH_STENCIL &&
697 devinfo->ver < 6)
698 return NULL;
699
700 const bool isl_surf_created_successfully =
701 crocus_resource_configure_main(screen, res, templ, modifier, 0);
702 if (!isl_surf_created_successfully)
703 return NULL;
704
705 const char *name = "miptree";
706
707 unsigned int flags = 0;
708 if (templ->usage == PIPE_USAGE_STAGING)
709 flags |= BO_ALLOC_COHERENT;
710
711 /* Scanout buffers need to be WC. */
712 if (templ->bind & PIPE_BIND_SCANOUT)
713 flags |= BO_ALLOC_SCANOUT;
714
715 uint64_t aux_size = 0;
716 uint32_t aux_preferred_alloc_flags;
717
718 if (!crocus_resource_configure_aux(screen, res, false, &aux_size,
719 &aux_preferred_alloc_flags)) {
720 goto fail;
721 }
722
723 /* Modifiers require the aux data to be in the same buffer as the main
724 * surface, but we combine them even when a modifiers is not being used.
725 */
726 const uint64_t bo_size =
727 MAX2(res->surf.size_B, res->aux.offset + aux_size);
728 uint32_t alignment = MAX2(4096, res->surf.alignment_B);
729 res->bo = crocus_bo_alloc_tiled(screen->bufmgr, name, bo_size, alignment,
730 isl_tiling_to_i915_tiling(res->surf.tiling),
731 res->surf.row_pitch_B, flags);
732
733 if (!res->bo)
734 goto fail;
735
736 if (aux_size > 0) {
737 res->aux.bo = res->bo;
738 crocus_bo_reference(res->aux.bo);
739 if (!crocus_resource_init_aux_buf(res, flags))
740 goto fail;
741 }
742
743 if (templ->format == PIPE_FORMAT_S8_UINT && !(templ->usage == PIPE_USAGE_STAGING) &&
744 devinfo->ver == 7 && (templ->bind & PIPE_BIND_SAMPLER_VIEW)) {
745 struct pipe_resource templ_shadow = (struct pipe_resource) {
746 .usage = 0,
747 .bind = PIPE_BIND_SAMPLER_VIEW,
748 .width0 = res->base.b.width0,
749 .height0 = res->base.b.height0,
750 .depth0 = res->base.b.depth0,
751 .last_level = res->base.b.last_level,
752 .nr_samples = res->base.b.nr_samples,
753 .nr_storage_samples = res->base.b.nr_storage_samples,
754 .array_size = res->base.b.array_size,
755 .format = PIPE_FORMAT_R8_UINT,
756 .target = res->base.b.target,
757 };
758 res->shadow = (struct crocus_resource *)screen->base.resource_create(&screen->base, &templ_shadow);
759 assert(res->shadow);
760 }
761
762 return &res->base.b;
763
764 fail:
765 fprintf(stderr, "XXX: resource creation failed\n");
766 crocus_resource_destroy(pscreen, &res->base.b);
767 return NULL;
768
769 }
770
771 static struct pipe_resource *
crocus_resource_create(struct pipe_screen * pscreen,const struct pipe_resource * templ)772 crocus_resource_create(struct pipe_screen *pscreen,
773 const struct pipe_resource *templ)
774 {
775 if (templ->target == PIPE_BUFFER)
776 return crocus_resource_create_for_buffer(pscreen, templ);
777 else
778 return crocus_resource_create_with_modifiers(pscreen, templ, NULL, 0);
779 }
780
781 static uint64_t
tiling_to_modifier(uint32_t tiling)782 tiling_to_modifier(uint32_t tiling)
783 {
784 static const uint64_t map[] = {
785 [I915_TILING_NONE] = DRM_FORMAT_MOD_LINEAR,
786 [I915_TILING_X] = I915_FORMAT_MOD_X_TILED,
787 [I915_TILING_Y] = I915_FORMAT_MOD_Y_TILED,
788 };
789
790 assert(tiling < ARRAY_SIZE(map));
791
792 return map[tiling];
793 }
794
795 static struct pipe_resource *
crocus_resource_from_user_memory(struct pipe_screen * pscreen,const struct pipe_resource * templ,void * user_memory)796 crocus_resource_from_user_memory(struct pipe_screen *pscreen,
797 const struct pipe_resource *templ,
798 void *user_memory)
799 {
800 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
801 struct crocus_bufmgr *bufmgr = screen->bufmgr;
802 struct crocus_resource *res = crocus_alloc_resource(pscreen, templ);
803 if (!res)
804 return NULL;
805
806 assert(templ->target == PIPE_BUFFER);
807
808 res->internal_format = templ->format;
809 res->bo = crocus_bo_create_userptr(bufmgr, "user",
810 user_memory, templ->width0);
811 if (!res->bo) {
812 free(res);
813 return NULL;
814 }
815
816 util_range_add(&res->base.b, &res->valid_buffer_range, 0, templ->width0);
817
818 return &res->base.b;
819 }
820
821 static struct pipe_resource *
crocus_resource_from_handle(struct pipe_screen * pscreen,const struct pipe_resource * templ,struct winsys_handle * whandle,unsigned usage)822 crocus_resource_from_handle(struct pipe_screen *pscreen,
823 const struct pipe_resource *templ,
824 struct winsys_handle *whandle,
825 unsigned usage)
826 {
827 assert(templ->target != PIPE_BUFFER);
828
829 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
830 struct crocus_bufmgr *bufmgr = screen->bufmgr;
831 struct crocus_resource *res = crocus_alloc_resource(pscreen, templ);
832
833 if (!res)
834 return NULL;
835
836 switch (whandle->type) {
837 case WINSYS_HANDLE_TYPE_FD:
838 res->bo = crocus_bo_import_dmabuf(bufmgr, whandle->handle,
839 whandle->modifier);
840 break;
841 case WINSYS_HANDLE_TYPE_SHARED:
842 res->bo = crocus_bo_gem_create_from_name(bufmgr, "winsys image",
843 whandle->handle);
844 break;
845 default:
846 unreachable("invalid winsys handle type");
847 }
848 if (!res->bo)
849 return NULL;
850
851 res->offset = whandle->offset;
852 res->external_format = whandle->format;
853
854 if (whandle->plane < util_format_get_num_planes(whandle->format)) {
855 const uint64_t modifier =
856 whandle->modifier != DRM_FORMAT_MOD_INVALID ?
857 whandle->modifier : tiling_to_modifier(res->bo->tiling_mode);
858
859 UNUSED const bool isl_surf_created_successfully =
860 crocus_resource_configure_main(screen, res, templ, modifier,
861 whandle->stride);
862 assert(isl_surf_created_successfully);
863 assert(res->bo->tiling_mode ==
864 isl_tiling_to_i915_tiling(res->surf.tiling));
865
866 // XXX: create_ccs_buf_for_image?
867 if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
868 if (!crocus_resource_alloc_separate_aux(screen, res))
869 goto fail;
870 } else {
871 if (res->mod_info->aux_usage != ISL_AUX_USAGE_NONE) {
872 uint32_t alloc_flags;
873 uint64_t size;
874 UNUSED bool ok = crocus_resource_configure_aux(screen, res, true, &size,
875 &alloc_flags);
876 assert(ok);
877 /* The gallium dri layer will create a separate plane resource
878 * for the aux image. crocus_resource_finish_aux_import will
879 * merge the separate aux parameters back into a single
880 * crocus_resource.
881 */
882 }
883 }
884 } else {
885 /* Save modifier import information to reconstruct later. After
886 * import, this will be available under a second image accessible
887 * from the main image with res->base.next. See
888 * crocus_resource_finish_aux_import.
889 */
890 res->aux.surf.row_pitch_B = whandle->stride;
891 res->aux.offset = whandle->offset;
892 res->aux.bo = res->bo;
893 res->bo = NULL;
894 }
895
896 return &res->base.b;
897
898 fail:
899 crocus_resource_destroy(pscreen, &res->base.b);
900 return NULL;
901 }
902
903 static struct pipe_resource *
crocus_resource_from_memobj(struct pipe_screen * pscreen,const struct pipe_resource * templ,struct pipe_memory_object * pmemobj,uint64_t offset)904 crocus_resource_from_memobj(struct pipe_screen *pscreen,
905 const struct pipe_resource *templ,
906 struct pipe_memory_object *pmemobj,
907 uint64_t offset)
908 {
909 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
910 struct crocus_memory_object *memobj = (struct crocus_memory_object *)pmemobj;
911 struct crocus_resource *res = crocus_alloc_resource(pscreen, templ);
912
913 if (!res)
914 return NULL;
915
916 /* Disable Depth, and combined Depth+Stencil for now. */
917 if (util_format_has_depth(util_format_description(templ->format)))
918 return NULL;
919
920 if (templ->flags & PIPE_RESOURCE_FLAG_TEXTURING_MORE_LIKELY) {
921 UNUSED const bool isl_surf_created_successfully =
922 crocus_resource_configure_main(screen, res, templ, DRM_FORMAT_MOD_INVALID, 0);
923 assert(isl_surf_created_successfully);
924 }
925
926 res->bo = memobj->bo;
927 res->offset = offset;
928 res->external_format = memobj->format;
929
930 crocus_bo_reference(memobj->bo);
931
932 return &res->base.b;
933 }
934
935 static void
crocus_flush_resource(struct pipe_context * ctx,struct pipe_resource * resource)936 crocus_flush_resource(struct pipe_context *ctx, struct pipe_resource *resource)
937 {
938 struct crocus_context *ice = (struct crocus_context *)ctx;
939 struct crocus_resource *res = (void *) resource;
940 const struct isl_drm_modifier_info *mod = res->mod_info;
941
942 crocus_resource_prepare_access(ice, res,
943 0, INTEL_REMAINING_LEVELS,
944 0, INTEL_REMAINING_LAYERS,
945 mod ? mod->aux_usage : ISL_AUX_USAGE_NONE,
946 mod ? mod->supports_clear_color : false);
947 }
948
949 static void
crocus_resource_disable_aux_on_first_query(struct pipe_resource * resource,unsigned usage)950 crocus_resource_disable_aux_on_first_query(struct pipe_resource *resource,
951 unsigned usage)
952 {
953 struct crocus_resource *res = (struct crocus_resource *)resource;
954 bool mod_with_aux =
955 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
956
957 /* Disable aux usage if explicit flush not set and this is the first time
958 * we are dealing with this resource and the resource was not created with
959 * a modifier with aux.
960 */
961 if (!mod_with_aux &&
962 (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) && res->aux.usage != 0) &&
963 p_atomic_read(&resource->reference.count) == 1) {
964 crocus_resource_disable_aux(res);
965 }
966 }
967
968 static bool
crocus_resource_get_param(struct pipe_screen * pscreen,struct pipe_context * context,struct pipe_resource * resource,unsigned plane,unsigned layer,unsigned level,enum pipe_resource_param param,unsigned handle_usage,uint64_t * value)969 crocus_resource_get_param(struct pipe_screen *pscreen,
970 struct pipe_context *context,
971 struct pipe_resource *resource,
972 unsigned plane,
973 unsigned layer,
974 unsigned level,
975 enum pipe_resource_param param,
976 unsigned handle_usage,
977 uint64_t *value)
978 {
979 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
980 struct crocus_resource *res = (struct crocus_resource *)resource;
981 bool mod_with_aux =
982 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
983 bool wants_aux = mod_with_aux && plane > 0;
984 bool result;
985 unsigned handle;
986
987 if (crocus_resource_unfinished_aux_import(res))
988 crocus_resource_finish_aux_import(pscreen, res);
989
990 struct crocus_bo *bo = wants_aux ? res->aux.bo : res->bo;
991
992 crocus_resource_disable_aux_on_first_query(resource, handle_usage);
993
994 switch (param) {
995 case PIPE_RESOURCE_PARAM_NPLANES:
996 if (mod_with_aux) {
997 *value = util_format_get_num_planes(res->external_format);
998 } else {
999 unsigned count = 0;
1000 for (struct pipe_resource *cur = resource; cur; cur = cur->next)
1001 count++;
1002 *value = count;
1003 }
1004 return true;
1005 case PIPE_RESOURCE_PARAM_STRIDE:
1006 *value = wants_aux ? res->aux.surf.row_pitch_B : res->surf.row_pitch_B;
1007 return true;
1008 case PIPE_RESOURCE_PARAM_OFFSET:
1009 *value = wants_aux ? res->aux.offset : 0;
1010 return true;
1011 case PIPE_RESOURCE_PARAM_MODIFIER:
1012 *value = res->mod_info ? res->mod_info->modifier :
1013 tiling_to_modifier(isl_tiling_to_i915_tiling(res->surf.tiling));
1014 return true;
1015 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_SHARED:
1016 result = crocus_bo_flink(bo, &handle) == 0;
1017 if (result)
1018 *value = handle;
1019 return result;
1020 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_KMS: {
1021 /* Because we share the same drm file across multiple crocus_screen, when
1022 * we export a GEM handle we must make sure it is valid in the DRM file
1023 * descriptor the caller is using (this is the FD given at screen
1024 * creation).
1025 */
1026 uint32_t handle;
1027 if (crocus_bo_export_gem_handle_for_device(bo, screen->winsys_fd, &handle))
1028 return false;
1029 *value = handle;
1030 return true;
1031 }
1032 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_FD:
1033 result = crocus_bo_export_dmabuf(bo, (int *) &handle) == 0;
1034 if (result)
1035 *value = handle;
1036 return result;
1037 default:
1038 return false;
1039 }
1040 }
1041
1042 static bool
crocus_resource_get_handle(struct pipe_screen * pscreen,struct pipe_context * ctx,struct pipe_resource * resource,struct winsys_handle * whandle,unsigned usage)1043 crocus_resource_get_handle(struct pipe_screen *pscreen,
1044 struct pipe_context *ctx,
1045 struct pipe_resource *resource,
1046 struct winsys_handle *whandle,
1047 unsigned usage)
1048 {
1049 struct crocus_screen *screen = (struct crocus_screen *) pscreen;
1050 struct crocus_resource *res = (struct crocus_resource *)resource;
1051 bool mod_with_aux =
1052 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1053
1054 crocus_resource_disable_aux_on_first_query(resource, usage);
1055
1056 struct crocus_bo *bo;
1057 if (mod_with_aux && whandle->plane > 0) {
1058 assert(res->aux.bo);
1059 bo = res->aux.bo;
1060 whandle->stride = res->aux.surf.row_pitch_B;
1061 whandle->offset = res->aux.offset;
1062 } else {
1063 /* If this is a buffer, stride should be 0 - no need to special case */
1064 whandle->stride = res->surf.row_pitch_B;
1065 bo = res->bo;
1066 }
1067 whandle->format = res->external_format;
1068 whandle->modifier =
1069 res->mod_info ? res->mod_info->modifier
1070 : tiling_to_modifier(res->bo->tiling_mode);
1071
1072 #ifndef NDEBUG
1073 enum isl_aux_usage allowed_usage =
1074 res->mod_info ? res->mod_info->aux_usage : ISL_AUX_USAGE_NONE;
1075
1076 if (res->aux.usage != allowed_usage) {
1077 enum isl_aux_state aux_state = crocus_resource_get_aux_state(res, 0, 0);
1078 assert(aux_state == ISL_AUX_STATE_RESOLVED ||
1079 aux_state == ISL_AUX_STATE_PASS_THROUGH);
1080 }
1081 #endif
1082
1083 switch (whandle->type) {
1084 case WINSYS_HANDLE_TYPE_SHARED:
1085 return crocus_bo_flink(bo, &whandle->handle) == 0;
1086 case WINSYS_HANDLE_TYPE_KMS: {
1087 /* Because we share the same drm file across multiple crocus_screen, when
1088 * we export a GEM handle we must make sure it is valid in the DRM file
1089 * descriptor the caller is using (this is the FD given at screen
1090 * creation).
1091 */
1092 uint32_t handle;
1093 if (crocus_bo_export_gem_handle_for_device(bo, screen->winsys_fd, &handle))
1094 return false;
1095 whandle->handle = handle;
1096 return true;
1097 }
1098 case WINSYS_HANDLE_TYPE_FD:
1099 return crocus_bo_export_dmabuf(bo, (int *) &whandle->handle) == 0;
1100 }
1101
1102 return false;
1103 }
1104
1105 static bool
resource_is_busy(struct crocus_context * ice,struct crocus_resource * res)1106 resource_is_busy(struct crocus_context *ice,
1107 struct crocus_resource *res)
1108 {
1109 bool busy = crocus_bo_busy(res->bo);
1110
1111 for (int i = 0; i < ice->batch_count; i++)
1112 busy |= crocus_batch_references(&ice->batches[i], res->bo);
1113
1114 return busy;
1115 }
1116
1117 void
crocus_replace_buffer_storage(struct pipe_context * ctx,struct pipe_resource * p_dst,struct pipe_resource * p_src,unsigned num_rebinds,uint32_t rebind_mask,uint32_t delete_buffer_id)1118 crocus_replace_buffer_storage(struct pipe_context *ctx,
1119 struct pipe_resource *p_dst,
1120 struct pipe_resource *p_src,
1121 unsigned num_rebinds,
1122 uint32_t rebind_mask,
1123 uint32_t delete_buffer_id)
1124 {
1125 struct crocus_screen *screen = (void *) ctx->screen;
1126 struct crocus_context *ice = (void *) ctx;
1127 struct crocus_resource *dst = (void *) p_dst;
1128 struct crocus_resource *src = (void *) p_src;
1129
1130 assert(memcmp(&dst->surf, &src->surf, sizeof(dst->surf)) == 0);
1131
1132 struct crocus_bo *old_bo = dst->bo;
1133
1134 /* Swap out the backing storage */
1135 crocus_bo_reference(src->bo);
1136 dst->bo = src->bo;
1137
1138 /* Rebind the buffer, replacing any state referring to the old BO's
1139 * address, and marking state dirty so it's reemitted.
1140 */
1141 screen->vtbl.rebind_buffer(ice, dst);
1142
1143 crocus_bo_unreference(old_bo);
1144 }
1145
1146 static void
crocus_invalidate_resource(struct pipe_context * ctx,struct pipe_resource * resource)1147 crocus_invalidate_resource(struct pipe_context *ctx,
1148 struct pipe_resource *resource)
1149 {
1150 struct crocus_screen *screen = (void *) ctx->screen;
1151 struct crocus_context *ice = (void *) ctx;
1152 struct crocus_resource *res = (void *) resource;
1153
1154 if (resource->target != PIPE_BUFFER)
1155 return;
1156
1157 /* If it's already invalidated, don't bother doing anything. */
1158 if (res->valid_buffer_range.start > res->valid_buffer_range.end)
1159 return;
1160
1161 if (!resource_is_busy(ice, res)) {
1162 /* The resource is idle, so just mark that it contains no data and
1163 * keep using the same underlying buffer object.
1164 */
1165 util_range_set_empty(&res->valid_buffer_range);
1166 return;
1167 }
1168
1169 /* Otherwise, try and replace the backing storage with a new BO. */
1170
1171 /* We can't reallocate memory we didn't allocate in the first place. */
1172 if (res->bo->userptr)
1173 return;
1174
1175 struct crocus_bo *old_bo = res->bo;
1176 struct crocus_bo *new_bo =
1177 crocus_bo_alloc(screen->bufmgr, res->bo->name, resource->width0);
1178
1179 if (!new_bo)
1180 return;
1181
1182 /* Swap out the backing storage */
1183 res->bo = new_bo;
1184
1185 /* Rebind the buffer, replacing any state referring to the old BO's
1186 * address, and marking state dirty so it's reemitted.
1187 */
1188 screen->vtbl.rebind_buffer(ice, res);
1189
1190 util_range_set_empty(&res->valid_buffer_range);
1191
1192 crocus_bo_unreference(old_bo);
1193 }
1194
1195 static void
crocus_flush_staging_region(struct pipe_transfer * xfer,const struct pipe_box * flush_box)1196 crocus_flush_staging_region(struct pipe_transfer *xfer,
1197 const struct pipe_box *flush_box)
1198 {
1199 if (!(xfer->usage & PIPE_MAP_WRITE))
1200 return;
1201
1202 struct crocus_transfer *map = (void *) xfer;
1203
1204 struct pipe_box src_box = *flush_box;
1205
1206 /* Account for extra alignment padding in staging buffer */
1207 if (xfer->resource->target == PIPE_BUFFER)
1208 src_box.x += xfer->box.x % CROCUS_MAP_BUFFER_ALIGNMENT;
1209
1210 struct pipe_box dst_box = (struct pipe_box) {
1211 .x = xfer->box.x + flush_box->x,
1212 .y = xfer->box.y + flush_box->y,
1213 .z = xfer->box.z + flush_box->z,
1214 .width = flush_box->width,
1215 .height = flush_box->height,
1216 .depth = flush_box->depth,
1217 };
1218
1219 crocus_copy_region(map->blorp, map->batch, xfer->resource, xfer->level,
1220 dst_box.x, dst_box.y, dst_box.z, map->staging, 0,
1221 &src_box);
1222 }
1223
1224 static void
crocus_unmap_copy_region(struct crocus_transfer * map)1225 crocus_unmap_copy_region(struct crocus_transfer *map)
1226 {
1227 crocus_resource_destroy(map->staging->screen, map->staging);
1228
1229 map->ptr = NULL;
1230 }
1231
1232 static void
crocus_map_copy_region(struct crocus_transfer * map)1233 crocus_map_copy_region(struct crocus_transfer *map)
1234 {
1235 struct pipe_screen *pscreen = &map->batch->screen->base;
1236 struct pipe_transfer *xfer = &map->base.b;
1237 struct pipe_box *box = &xfer->box;
1238 struct crocus_resource *res = (void *) xfer->resource;
1239
1240 unsigned extra = xfer->resource->target == PIPE_BUFFER ?
1241 box->x % CROCUS_MAP_BUFFER_ALIGNMENT : 0;
1242
1243 struct pipe_resource templ = (struct pipe_resource) {
1244 .usage = PIPE_USAGE_STAGING,
1245 .width0 = box->width + extra,
1246 .height0 = box->height,
1247 .depth0 = 1,
1248 .nr_samples = xfer->resource->nr_samples,
1249 .nr_storage_samples = xfer->resource->nr_storage_samples,
1250 .array_size = box->depth,
1251 .format = res->internal_format,
1252 };
1253
1254 if (xfer->resource->target == PIPE_BUFFER)
1255 templ.target = PIPE_BUFFER;
1256 else if (templ.array_size > 1)
1257 templ.target = PIPE_TEXTURE_2D_ARRAY;
1258 else
1259 templ.target = PIPE_TEXTURE_2D;
1260
1261 map->staging = crocus_resource_create(pscreen, &templ);
1262 assert(map->staging);
1263
1264 if (templ.target != PIPE_BUFFER) {
1265 struct isl_surf *surf = &((struct crocus_resource *) map->staging)->surf;
1266 xfer->stride = isl_surf_get_row_pitch_B(surf);
1267 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1268 }
1269
1270 if (!(xfer->usage & PIPE_MAP_DISCARD_RANGE)) {
1271 crocus_copy_region(map->blorp, map->batch, map->staging, 0, extra, 0, 0,
1272 xfer->resource, xfer->level, box);
1273 /* Ensure writes to the staging BO land before we map it below. */
1274 crocus_emit_pipe_control_flush(map->batch,
1275 "transfer read: flush before mapping",
1276 PIPE_CONTROL_RENDER_TARGET_FLUSH |
1277 PIPE_CONTROL_CS_STALL);
1278 }
1279
1280 struct crocus_bo *staging_bo = crocus_resource_bo(map->staging);
1281
1282 if (crocus_batch_references(map->batch, staging_bo))
1283 crocus_batch_flush(map->batch);
1284
1285 map->ptr =
1286 crocus_bo_map(map->dbg, staging_bo, xfer->usage & MAP_FLAGS) + extra;
1287
1288 map->unmap = crocus_unmap_copy_region;
1289 }
1290
1291 static void
get_image_offset_el(const struct isl_surf * surf,unsigned level,unsigned z,unsigned * out_x0_el,unsigned * out_y0_el)1292 get_image_offset_el(const struct isl_surf *surf, unsigned level, unsigned z,
1293 unsigned *out_x0_el, unsigned *out_y0_el)
1294 {
1295 ASSERTED uint32_t z0_el, a0_el;
1296 if (surf->dim == ISL_SURF_DIM_3D) {
1297 isl_surf_get_image_offset_el(surf, level, 0, z,
1298 out_x0_el, out_y0_el, &z0_el, &a0_el);
1299 } else {
1300 isl_surf_get_image_offset_el(surf, level, z, 0,
1301 out_x0_el, out_y0_el, &z0_el, &a0_el);
1302 }
1303 assert(z0_el == 0 && a0_el == 0);
1304 }
1305
1306 void
crocus_resource_get_image_offset(struct crocus_resource * res,uint32_t level,uint32_t z,uint32_t * x,uint32_t * y)1307 crocus_resource_get_image_offset(struct crocus_resource *res,
1308 uint32_t level, uint32_t z,
1309 uint32_t *x, uint32_t *y)
1310 {
1311 get_image_offset_el(&res->surf, level, z, x, y);
1312 }
1313
1314 /**
1315 * Get pointer offset into stencil buffer.
1316 *
1317 * The stencil buffer is W tiled. Since the GTT is incapable of W fencing, we
1318 * must decode the tile's layout in software.
1319 *
1320 * See
1321 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.2.1 W-Major Tile
1322 * Format.
1323 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.3 Tiling Algorithm
1324 *
1325 * Even though the returned offset is always positive, the return type is
1326 * signed due to
1327 * commit e8b1c6d6f55f5be3bef25084fdd8b6127517e137
1328 * mesa: Fix return type of _mesa_get_format_bytes() (#37351)
1329 */
1330 static intptr_t
s8_offset(uint32_t stride,uint32_t x,uint32_t y,bool swizzled)1331 s8_offset(uint32_t stride, uint32_t x, uint32_t y, bool swizzled)
1332 {
1333 uint32_t tile_size = 4096;
1334 uint32_t tile_width = 64;
1335 uint32_t tile_height = 64;
1336 uint32_t row_size = 64 * stride / 2; /* Two rows are interleaved. */
1337
1338 uint32_t tile_x = x / tile_width;
1339 uint32_t tile_y = y / tile_height;
1340
1341 /* The byte's address relative to the tile's base addres. */
1342 uint32_t byte_x = x % tile_width;
1343 uint32_t byte_y = y % tile_height;
1344
1345 uintptr_t u = tile_y * row_size
1346 + tile_x * tile_size
1347 + 512 * (byte_x / 8)
1348 + 64 * (byte_y / 8)
1349 + 32 * ((byte_y / 4) % 2)
1350 + 16 * ((byte_x / 4) % 2)
1351 + 8 * ((byte_y / 2) % 2)
1352 + 4 * ((byte_x / 2) % 2)
1353 + 2 * (byte_y % 2)
1354 + 1 * (byte_x % 2);
1355
1356 if (swizzled) {
1357 /* adjust for bit6 swizzling */
1358 if (((byte_x / 8) % 2) == 1) {
1359 if (((byte_y / 8) % 2) == 0) {
1360 u += 64;
1361 } else {
1362 u -= 64;
1363 }
1364 }
1365 }
1366
1367 return u;
1368 }
1369
1370 static void
crocus_unmap_s8(struct crocus_transfer * map)1371 crocus_unmap_s8(struct crocus_transfer *map)
1372 {
1373 struct pipe_transfer *xfer = &map->base.b;
1374 const struct pipe_box *box = &xfer->box;
1375 struct crocus_resource *res = (struct crocus_resource *) xfer->resource;
1376 struct isl_surf *surf = &res->surf;
1377
1378 if (xfer->usage & PIPE_MAP_WRITE) {
1379 uint8_t *untiled_s8_map = map->ptr;
1380 uint8_t *tiled_s8_map =
1381 crocus_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1382
1383 for (int s = 0; s < box->depth; s++) {
1384 unsigned x0_el, y0_el;
1385 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1386
1387 for (uint32_t y = 0; y < box->height; y++) {
1388 for (uint32_t x = 0; x < box->width; x++) {
1389 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1390 x0_el + box->x + x,
1391 y0_el + box->y + y,
1392 map->has_swizzling);
1393 tiled_s8_map[offset] =
1394 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x];
1395 }
1396 }
1397 }
1398 }
1399
1400 free(map->buffer);
1401 }
1402
1403 static void
crocus_map_s8(struct crocus_transfer * map)1404 crocus_map_s8(struct crocus_transfer *map)
1405 {
1406 struct pipe_transfer *xfer = &map->base.b;
1407 const struct pipe_box *box = &xfer->box;
1408 struct crocus_resource *res = (struct crocus_resource *) xfer->resource;
1409 struct isl_surf *surf = &res->surf;
1410
1411 xfer->stride = surf->row_pitch_B;
1412 xfer->layer_stride = xfer->stride * box->height;
1413
1414 /* The tiling and detiling functions require that the linear buffer has
1415 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1416 * over-allocate the linear buffer to get the proper alignment.
1417 */
1418 map->buffer = map->ptr = malloc(xfer->layer_stride * box->depth);
1419 assert(map->buffer);
1420
1421 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
1422 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
1423 * invalidate is set, since we'll be writing the whole rectangle from our
1424 * temporary buffer back out.
1425 */
1426 if (!(xfer->usage & PIPE_MAP_DISCARD_RANGE)) {
1427 uint8_t *untiled_s8_map = map->ptr;
1428 uint8_t *tiled_s8_map =
1429 crocus_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1430
1431 for (int s = 0; s < box->depth; s++) {
1432 unsigned x0_el, y0_el;
1433 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1434
1435 for (uint32_t y = 0; y < box->height; y++) {
1436 for (uint32_t x = 0; x < box->width; x++) {
1437 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1438 x0_el + box->x + x,
1439 y0_el + box->y + y,
1440 map->has_swizzling);
1441 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x] =
1442 tiled_s8_map[offset];
1443 }
1444 }
1445 }
1446 }
1447
1448 map->unmap = crocus_unmap_s8;
1449 }
1450
1451 /* Compute extent parameters for use with tiled_memcpy functions.
1452 * xs are in units of bytes and ys are in units of strides.
1453 */
1454 static inline void
tile_extents(const struct isl_surf * surf,const struct pipe_box * box,unsigned level,int z,unsigned * x1_B,unsigned * x2_B,unsigned * y1_el,unsigned * y2_el)1455 tile_extents(const struct isl_surf *surf,
1456 const struct pipe_box *box,
1457 unsigned level, int z,
1458 unsigned *x1_B, unsigned *x2_B,
1459 unsigned *y1_el, unsigned *y2_el)
1460 {
1461 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1462 const unsigned cpp = fmtl->bpb / 8;
1463
1464 assert(box->x % fmtl->bw == 0);
1465 assert(box->y % fmtl->bh == 0);
1466
1467 unsigned x0_el, y0_el;
1468 get_image_offset_el(surf, level, box->z + z, &x0_el, &y0_el);
1469
1470 *x1_B = (box->x / fmtl->bw + x0_el) * cpp;
1471 *y1_el = box->y / fmtl->bh + y0_el;
1472 *x2_B = (DIV_ROUND_UP(box->x + box->width, fmtl->bw) + x0_el) * cpp;
1473 *y2_el = DIV_ROUND_UP(box->y + box->height, fmtl->bh) + y0_el;
1474 }
1475
1476 static void
crocus_unmap_tiled_memcpy(struct crocus_transfer * map)1477 crocus_unmap_tiled_memcpy(struct crocus_transfer *map)
1478 {
1479 struct pipe_transfer *xfer = &map->base.b;
1480 const struct pipe_box *box = &xfer->box;
1481 struct crocus_resource *res = (struct crocus_resource *) xfer->resource;
1482 struct isl_surf *surf = &res->surf;
1483
1484 if (xfer->usage & PIPE_MAP_WRITE) {
1485 char *dst =
1486 crocus_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1487
1488 for (int s = 0; s < box->depth; s++) {
1489 unsigned x1, x2, y1, y2;
1490 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1491
1492 void *ptr = map->ptr + s * xfer->layer_stride;
1493
1494 isl_memcpy_linear_to_tiled(x1, x2, y1, y2, dst, ptr,
1495 surf->row_pitch_B, xfer->stride,
1496 map->has_swizzling,
1497 surf->tiling, ISL_MEMCPY);
1498 }
1499 }
1500 os_free_aligned(map->buffer);
1501 map->buffer = map->ptr = NULL;
1502 }
1503
1504 static void
crocus_map_tiled_memcpy(struct crocus_transfer * map)1505 crocus_map_tiled_memcpy(struct crocus_transfer *map)
1506 {
1507 struct pipe_transfer *xfer = &map->base.b;
1508 const struct pipe_box *box = &xfer->box;
1509 struct crocus_resource *res = (struct crocus_resource *) xfer->resource;
1510 struct isl_surf *surf = &res->surf;
1511
1512 xfer->stride = ALIGN(surf->row_pitch_B, 16);
1513 xfer->layer_stride = xfer->stride * box->height;
1514
1515 unsigned x1, x2, y1, y2;
1516 tile_extents(surf, box, xfer->level, 0, &x1, &x2, &y1, &y2);
1517
1518 /* The tiling and detiling functions require that the linear buffer has
1519 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1520 * over-allocate the linear buffer to get the proper alignment.
1521 */
1522 map->buffer =
1523 os_malloc_aligned(xfer->layer_stride * box->depth, 16);
1524 assert(map->buffer);
1525 map->ptr = (char *)map->buffer + (x1 & 0xf);
1526
1527 if (!(xfer->usage & PIPE_MAP_DISCARD_RANGE)) {
1528 char *src =
1529 crocus_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1530
1531 for (int s = 0; s < box->depth; s++) {
1532 unsigned x1, x2, y1, y2;
1533 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1534
1535 /* Use 's' rather than 'box->z' to rebase the first slice to 0. */
1536 void *ptr = map->ptr + s * xfer->layer_stride;
1537
1538 isl_memcpy_tiled_to_linear(x1, x2, y1, y2, ptr, src, xfer->stride,
1539 surf->row_pitch_B,
1540 map->has_swizzling,
1541 surf->tiling,
1542 #if defined(USE_SSE41)
1543 util_get_cpu_caps()->has_sse4_1 ? ISL_MEMCPY_STREAMING_LOAD :
1544 #endif
1545 ISL_MEMCPY);
1546 }
1547 }
1548
1549 map->unmap = crocus_unmap_tiled_memcpy;
1550 }
1551
1552 static void
crocus_map_direct(struct crocus_transfer * map)1553 crocus_map_direct(struct crocus_transfer *map)
1554 {
1555 struct pipe_transfer *xfer = &map->base.b;
1556 struct pipe_box *box = &xfer->box;
1557 struct crocus_resource *res = (struct crocus_resource *) xfer->resource;
1558
1559 void *ptr = crocus_bo_map(map->dbg, res->bo, xfer->usage & MAP_FLAGS);
1560
1561 if (res->base.b.target == PIPE_BUFFER) {
1562 xfer->stride = 0;
1563 xfer->layer_stride = 0;
1564
1565 map->ptr = ptr + box->x;
1566 } else {
1567 struct isl_surf *surf = &res->surf;
1568 const struct isl_format_layout *fmtl =
1569 isl_format_get_layout(surf->format);
1570 const unsigned cpp = fmtl->bpb / 8;
1571 unsigned x0_el, y0_el;
1572
1573 assert(box->x % fmtl->bw == 0);
1574 assert(box->y % fmtl->bh == 0);
1575 get_image_offset_el(surf, xfer->level, box->z, &x0_el, &y0_el);
1576
1577 x0_el += box->x / fmtl->bw;
1578 y0_el += box->y / fmtl->bh;
1579
1580 xfer->stride = isl_surf_get_row_pitch_B(surf);
1581 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1582
1583 map->ptr = ptr + y0_el * xfer->stride + x0_el * cpp;
1584 }
1585 }
1586
1587 static bool
can_promote_to_async(const struct crocus_resource * res,const struct pipe_box * box,unsigned usage)1588 can_promote_to_async(const struct crocus_resource *res,
1589 const struct pipe_box *box,
1590 unsigned usage)
1591 {
1592 /* If we're writing to a section of the buffer that hasn't even been
1593 * initialized with useful data, then we can safely promote this write
1594 * to be unsynchronized. This helps the common pattern of appending data.
1595 */
1596 return res->base.b.target == PIPE_BUFFER && (usage & PIPE_MAP_WRITE) &&
1597 !(usage & TC_TRANSFER_MAP_NO_INFER_UNSYNCHRONIZED) &&
1598 !util_ranges_intersect(&res->valid_buffer_range, box->x,
1599 box->x + box->width);
1600 }
1601
1602 static void *
crocus_transfer_map(struct pipe_context * ctx,struct pipe_resource * resource,unsigned level,unsigned usage,const struct pipe_box * box,struct pipe_transfer ** ptransfer)1603 crocus_transfer_map(struct pipe_context *ctx,
1604 struct pipe_resource *resource,
1605 unsigned level,
1606 unsigned usage,
1607 const struct pipe_box *box,
1608 struct pipe_transfer **ptransfer)
1609 {
1610 struct crocus_context *ice = (struct crocus_context *)ctx;
1611 struct crocus_resource *res = (struct crocus_resource *)resource;
1612 struct isl_surf *surf = &res->surf;
1613 struct crocus_screen *screen = (struct crocus_screen *)ctx->screen;
1614
1615 if (usage & PIPE_MAP_DISCARD_WHOLE_RESOURCE) {
1616 /* Replace the backing storage with a fresh buffer for non-async maps */
1617 if (!(usage & (PIPE_MAP_UNSYNCHRONIZED |
1618 TC_TRANSFER_MAP_NO_INVALIDATE)))
1619 crocus_invalidate_resource(ctx, resource);
1620
1621 /* If we can discard the whole resource, we can discard the range. */
1622 usage |= PIPE_MAP_DISCARD_RANGE;
1623 }
1624
1625 if (!(usage & PIPE_MAP_UNSYNCHRONIZED) &&
1626 can_promote_to_async(res, box, usage)) {
1627 usage |= PIPE_MAP_UNSYNCHRONIZED;
1628 }
1629
1630 bool map_would_stall = false;
1631
1632 if (!(usage & PIPE_MAP_UNSYNCHRONIZED)) {
1633 map_would_stall = resource_is_busy(ice, res) ||
1634 crocus_has_invalid_primary(res, level, 1, box->z, box->depth);
1635
1636
1637 if (map_would_stall && (usage & PIPE_MAP_DONTBLOCK) &&
1638 (usage & PIPE_MAP_DIRECTLY))
1639 return NULL;
1640 }
1641
1642 if (surf->tiling != ISL_TILING_LINEAR &&
1643 (usage & PIPE_MAP_DIRECTLY))
1644 return NULL;
1645
1646 struct crocus_transfer *map;
1647 if (usage & TC_TRANSFER_MAP_THREADED_UNSYNC)
1648 map = slab_zalloc(&ice->transfer_pool_unsync);
1649 else
1650 map = slab_zalloc(&ice->transfer_pool);
1651
1652 struct pipe_transfer *xfer = &map->base.b;
1653
1654 if (!map)
1655 return NULL;
1656
1657 map->dbg = &ice->dbg;
1658
1659 map->has_swizzling = screen->devinfo.has_bit6_swizzle;
1660 pipe_resource_reference(&xfer->resource, resource);
1661 xfer->level = level;
1662 xfer->usage = usage;
1663 xfer->box = *box;
1664 *ptransfer = xfer;
1665
1666 map->dest_had_defined_contents =
1667 util_ranges_intersect(&res->valid_buffer_range, box->x,
1668 box->x + box->width);
1669
1670 if (usage & PIPE_MAP_WRITE)
1671 util_range_add(&res->base.b, &res->valid_buffer_range, box->x, box->x + box->width);
1672
1673 /* Avoid using GPU copies for persistent/coherent buffers, as the idea
1674 * there is to access them simultaneously on the CPU & GPU. This also
1675 * avoids trying to use GPU copies for our u_upload_mgr buffers which
1676 * contain state we're constructing for a GPU draw call, which would
1677 * kill us with infinite stack recursion.
1678 */
1679 bool no_gpu = usage & (PIPE_MAP_PERSISTENT |
1680 PIPE_MAP_COHERENT |
1681 PIPE_MAP_DIRECTLY);
1682
1683 /* GPU copies are not useful for buffer reads. Instead of stalling to
1684 * read from the original buffer, we'd simply copy it to a temporary...
1685 * then stall (a bit longer) to read from that buffer.
1686 *
1687 * Images are less clear-cut. Color resolves are destructive, removing
1688 * the underlying compression, so we'd rather blit the data to a linear
1689 * temporary and map that, to avoid the resolve. (It might be better to
1690 * a tiled temporary and use the tiled_memcpy paths...)
1691 */
1692 if (!(usage & PIPE_MAP_DISCARD_RANGE) &&
1693 !crocus_has_invalid_primary(res, level, 1, box->z, box->depth))
1694 no_gpu = true;
1695
1696 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1697 if (fmtl->txc == ISL_TXC_ASTC)
1698 no_gpu = true;
1699
1700 if (map_would_stall && !no_gpu) {
1701 /* If we need a synchronous mapping and the resource is busy, or needs
1702 * resolving, we copy to/from a linear temporary buffer using the GPU.
1703 */
1704 map->batch = &ice->batches[CROCUS_BATCH_RENDER];
1705 map->blorp = &ice->blorp;
1706 crocus_map_copy_region(map);
1707 } else {
1708 /* Otherwise we're free to map on the CPU. */
1709
1710 if (resource->target != PIPE_BUFFER) {
1711 crocus_resource_access_raw(ice, res,
1712 level, box->z, box->depth,
1713 usage & PIPE_MAP_WRITE);
1714 }
1715
1716 if (!(usage & PIPE_MAP_UNSYNCHRONIZED)) {
1717 for (int i = 0; i < ice->batch_count; i++) {
1718 if (crocus_batch_references(&ice->batches[i], res->bo))
1719 crocus_batch_flush(&ice->batches[i]);
1720 }
1721 }
1722
1723 if (surf->tiling == ISL_TILING_W) {
1724 /* TODO: Teach crocus_map_tiled_memcpy about W-tiling... */
1725 crocus_map_s8(map);
1726 } else if (surf->tiling != ISL_TILING_LINEAR && screen->devinfo.ver > 4) {
1727 crocus_map_tiled_memcpy(map);
1728 } else {
1729 crocus_map_direct(map);
1730 }
1731 }
1732
1733 return map->ptr;
1734 }
1735
1736 static void
crocus_transfer_flush_region(struct pipe_context * ctx,struct pipe_transfer * xfer,const struct pipe_box * box)1737 crocus_transfer_flush_region(struct pipe_context *ctx,
1738 struct pipe_transfer *xfer,
1739 const struct pipe_box *box)
1740 {
1741 struct crocus_context *ice = (struct crocus_context *)ctx;
1742 struct crocus_resource *res = (struct crocus_resource *) xfer->resource;
1743 struct crocus_transfer *map = (void *) xfer;
1744
1745 if (map->staging)
1746 crocus_flush_staging_region(xfer, box);
1747
1748 uint32_t history_flush = 0;
1749
1750 if (res->base.b.target == PIPE_BUFFER) {
1751 if (map->staging)
1752 history_flush |= PIPE_CONTROL_RENDER_TARGET_FLUSH;
1753
1754 if (map->dest_had_defined_contents)
1755 history_flush |= crocus_flush_bits_for_history(res);
1756
1757 util_range_add(&res->base.b, &res->valid_buffer_range, box->x, box->x + box->width);
1758 }
1759
1760 if (history_flush & ~PIPE_CONTROL_CS_STALL) {
1761 for (int i = 0; i < ice->batch_count; i++) {
1762 struct crocus_batch *batch = &ice->batches[i];
1763
1764 if (!batch->command.bo)
1765 continue;
1766 if (batch->contains_draw || batch->cache.render->entries) {
1767 crocus_batch_maybe_flush(batch, 24);
1768 crocus_emit_pipe_control_flush(batch,
1769 "cache history: transfer flush",
1770 history_flush);
1771 }
1772 }
1773 }
1774
1775 /* Make sure we flag constants dirty even if there's no need to emit
1776 * any PIPE_CONTROLs to a batch.
1777 */
1778 crocus_dirty_for_history(ice, res);
1779 }
1780
1781 static void
crocus_transfer_unmap(struct pipe_context * ctx,struct pipe_transfer * xfer)1782 crocus_transfer_unmap(struct pipe_context *ctx, struct pipe_transfer *xfer)
1783 {
1784 struct crocus_context *ice = (struct crocus_context *)ctx;
1785 struct crocus_transfer *map = (void *) xfer;
1786
1787 if (!(xfer->usage & (PIPE_MAP_FLUSH_EXPLICIT |
1788 PIPE_MAP_COHERENT))) {
1789 struct pipe_box flush_box = {
1790 .x = 0, .y = 0, .z = 0,
1791 .width = xfer->box.width,
1792 .height = xfer->box.height,
1793 .depth = xfer->box.depth,
1794 };
1795 crocus_transfer_flush_region(ctx, xfer, &flush_box);
1796 }
1797
1798 if (map->unmap)
1799 map->unmap(map);
1800
1801 pipe_resource_reference(&xfer->resource, NULL);
1802 /* transfer_unmap is always called from the driver thread, so we have to
1803 * use transfer_pool, not transfer_pool_unsync. Freeing an object into a
1804 * different pool is allowed, however.
1805 */
1806 slab_free(&ice->transfer_pool, map);
1807 }
1808
1809 /**
1810 * Mark state dirty that needs to be re-emitted when a resource is written.
1811 */
1812 void
crocus_dirty_for_history(struct crocus_context * ice,struct crocus_resource * res)1813 crocus_dirty_for_history(struct crocus_context *ice,
1814 struct crocus_resource *res)
1815 {
1816 uint64_t stage_dirty = 0ull;
1817
1818 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1819 stage_dirty |= ((uint64_t)res->bind_stages) << CROCUS_SHIFT_FOR_STAGE_DIRTY_CONSTANTS;
1820 }
1821
1822 ice->state.stage_dirty |= stage_dirty;
1823 }
1824
1825 /**
1826 * Produce a set of PIPE_CONTROL bits which ensure data written to a
1827 * resource becomes visible, and any stale read cache data is invalidated.
1828 */
1829 uint32_t
crocus_flush_bits_for_history(struct crocus_resource * res)1830 crocus_flush_bits_for_history(struct crocus_resource *res)
1831 {
1832 uint32_t flush = PIPE_CONTROL_CS_STALL;
1833
1834 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1835 flush |= PIPE_CONTROL_CONST_CACHE_INVALIDATE |
1836 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1837 }
1838
1839 if (res->bind_history & PIPE_BIND_SAMPLER_VIEW)
1840 flush |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1841
1842 if (res->bind_history & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
1843 flush |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1844
1845 if (res->bind_history & (PIPE_BIND_SHADER_BUFFER | PIPE_BIND_SHADER_IMAGE))
1846 flush |= PIPE_CONTROL_DATA_CACHE_FLUSH;
1847
1848 return flush;
1849 }
1850
1851 void
crocus_flush_and_dirty_for_history(struct crocus_context * ice,struct crocus_batch * batch,struct crocus_resource * res,uint32_t extra_flags,const char * reason)1852 crocus_flush_and_dirty_for_history(struct crocus_context *ice,
1853 struct crocus_batch *batch,
1854 struct crocus_resource *res,
1855 uint32_t extra_flags,
1856 const char *reason)
1857 {
1858 if (res->base.b.target != PIPE_BUFFER)
1859 return;
1860
1861 uint32_t flush = crocus_flush_bits_for_history(res) | extra_flags;
1862
1863 crocus_emit_pipe_control_flush(batch, reason, flush);
1864
1865 crocus_dirty_for_history(ice, res);
1866 }
1867
1868 bool
crocus_resource_set_clear_color(struct crocus_context * ice,struct crocus_resource * res,union isl_color_value color)1869 crocus_resource_set_clear_color(struct crocus_context *ice,
1870 struct crocus_resource *res,
1871 union isl_color_value color)
1872 {
1873 if (memcmp(&res->aux.clear_color, &color, sizeof(color)) != 0) {
1874 res->aux.clear_color = color;
1875 return true;
1876 }
1877
1878 return false;
1879 }
1880
1881 union isl_color_value
crocus_resource_get_clear_color(const struct crocus_resource * res)1882 crocus_resource_get_clear_color(const struct crocus_resource *res)
1883 {
1884 assert(res->aux.bo);
1885
1886 return res->aux.clear_color;
1887 }
1888
1889 static enum pipe_format
crocus_resource_get_internal_format(struct pipe_resource * p_res)1890 crocus_resource_get_internal_format(struct pipe_resource *p_res)
1891 {
1892 struct crocus_resource *res = (void *) p_res;
1893 return res->internal_format;
1894 }
1895
1896 static const struct u_transfer_vtbl transfer_vtbl = {
1897 .resource_create = crocus_resource_create,
1898 .resource_destroy = crocus_resource_destroy,
1899 .transfer_map = crocus_transfer_map,
1900 .transfer_unmap = crocus_transfer_unmap,
1901 .transfer_flush_region = crocus_transfer_flush_region,
1902 .get_internal_format = crocus_resource_get_internal_format,
1903 .set_stencil = crocus_resource_set_separate_stencil,
1904 .get_stencil = crocus_resource_get_separate_stencil,
1905 };
1906
1907 static bool
crocus_is_dmabuf_modifier_supported(struct pipe_screen * pscreen,uint64_t modifier,enum pipe_format pfmt,bool * external_only)1908 crocus_is_dmabuf_modifier_supported(struct pipe_screen *pscreen,
1909 uint64_t modifier, enum pipe_format pfmt,
1910 bool *external_only)
1911 {
1912 struct crocus_screen *screen = (void *) pscreen;
1913 const struct intel_device_info *devinfo = &screen->devinfo;
1914
1915 if (modifier_is_supported(devinfo, pfmt, 0, modifier)) {
1916 if (external_only)
1917 *external_only = false;
1918
1919 return true;
1920 }
1921
1922 return false;
1923 }
1924
1925 static unsigned int
crocus_get_dmabuf_modifier_planes(struct pipe_screen * pscreen,uint64_t modifier,enum pipe_format format)1926 crocus_get_dmabuf_modifier_planes(struct pipe_screen *pscreen, uint64_t modifier,
1927 enum pipe_format format)
1928 {
1929 return util_format_get_num_planes(format);
1930 }
1931
1932 static struct pipe_memory_object *
crocus_memobj_create_from_handle(struct pipe_screen * pscreen,struct winsys_handle * whandle,bool dedicated)1933 crocus_memobj_create_from_handle(struct pipe_screen *pscreen,
1934 struct winsys_handle *whandle,
1935 bool dedicated)
1936 {
1937 struct crocus_screen *screen = (struct crocus_screen *)pscreen;
1938 struct crocus_memory_object *memobj = CALLOC_STRUCT(crocus_memory_object);
1939 struct crocus_bo *bo;
1940 const struct isl_drm_modifier_info *mod_inf;
1941
1942 if (!memobj)
1943 return NULL;
1944
1945 switch (whandle->type) {
1946 case WINSYS_HANDLE_TYPE_SHARED:
1947 bo = crocus_bo_gem_create_from_name(screen->bufmgr, "winsys image",
1948 whandle->handle);
1949 break;
1950 case WINSYS_HANDLE_TYPE_FD:
1951 mod_inf = isl_drm_modifier_get_info(whandle->modifier);
1952 if (mod_inf) {
1953 bo = crocus_bo_import_dmabuf(screen->bufmgr, whandle->handle,
1954 whandle->modifier);
1955 } else {
1956 /* If we can't get information about the tiling from the
1957 * kernel we ignore it. We are going to set it when we
1958 * create the resource.
1959 */
1960 bo = crocus_bo_import_dmabuf_no_mods(screen->bufmgr,
1961 whandle->handle);
1962 }
1963
1964 break;
1965 default:
1966 unreachable("invalid winsys handle type");
1967 }
1968
1969 if (!bo) {
1970 free(memobj);
1971 return NULL;
1972 }
1973
1974 memobj->b.dedicated = dedicated;
1975 memobj->bo = bo;
1976 memobj->format = whandle->format;
1977 memobj->stride = whandle->stride;
1978
1979 return &memobj->b;
1980 }
1981
1982 static void
crocus_memobj_destroy(struct pipe_screen * pscreen,struct pipe_memory_object * pmemobj)1983 crocus_memobj_destroy(struct pipe_screen *pscreen,
1984 struct pipe_memory_object *pmemobj)
1985 {
1986 struct crocus_memory_object *memobj = (struct crocus_memory_object *)pmemobj;
1987
1988 crocus_bo_unreference(memobj->bo);
1989 free(memobj);
1990 }
1991
1992 void
crocus_init_screen_resource_functions(struct pipe_screen * pscreen)1993 crocus_init_screen_resource_functions(struct pipe_screen *pscreen)
1994 {
1995 struct crocus_screen *screen = (void *) pscreen;
1996 pscreen->query_dmabuf_modifiers = crocus_query_dmabuf_modifiers;
1997 pscreen->is_dmabuf_modifier_supported = crocus_is_dmabuf_modifier_supported;
1998 pscreen->get_dmabuf_modifier_planes = crocus_get_dmabuf_modifier_planes;
1999 pscreen->resource_create_with_modifiers =
2000 crocus_resource_create_with_modifiers;
2001 pscreen->resource_create = u_transfer_helper_resource_create;
2002 pscreen->resource_from_user_memory = crocus_resource_from_user_memory;
2003 pscreen->resource_from_handle = crocus_resource_from_handle;
2004 pscreen->resource_from_memobj = crocus_resource_from_memobj;
2005 pscreen->resource_get_handle = crocus_resource_get_handle;
2006 pscreen->resource_get_param = crocus_resource_get_param;
2007 pscreen->resource_destroy = u_transfer_helper_resource_destroy;
2008 pscreen->memobj_create_from_handle = crocus_memobj_create_from_handle;
2009 pscreen->memobj_destroy = crocus_memobj_destroy;
2010 pscreen->transfer_helper =
2011 u_transfer_helper_create(&transfer_vtbl, screen->devinfo.ver >= 6,
2012 screen->devinfo.ver >= 6, false, true, false);
2013 }
2014
2015 void
crocus_init_resource_functions(struct pipe_context * ctx)2016 crocus_init_resource_functions(struct pipe_context *ctx)
2017 {
2018 ctx->flush_resource = crocus_flush_resource;
2019 ctx->invalidate_resource = crocus_invalidate_resource;
2020 ctx->buffer_map = u_transfer_helper_transfer_map;
2021 ctx->texture_map = u_transfer_helper_transfer_map;
2022 ctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
2023 ctx->buffer_unmap = u_transfer_helper_transfer_unmap;
2024 ctx->texture_unmap = u_transfer_helper_transfer_unmap;
2025 ctx->buffer_subdata = u_default_buffer_subdata;
2026 ctx->texture_subdata = u_default_texture_subdata;
2027 }
2028