iris: Create an unusable secondary aux surface
[mesa.git] / src / gallium / drivers / iris / iris_resource.c
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 iris_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/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 "iris_batch.h"
47 #include "iris_context.h"
48 #include "iris_resource.h"
49 #include "iris_screen.h"
50 #include "intel/common/gen_aux_map.h"
51 #include "intel/dev/gen_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
73 modifier_is_supported(const struct gen_device_info *devinfo,
74 enum pipe_format pfmt, uint64_t modifier)
75 {
76 /* XXX: do something real */
77 switch (modifier) {
78 case I915_FORMAT_MOD_Y_TILED_CCS: {
79 if (unlikely(INTEL_DEBUG & DEBUG_NO_RBC))
80 return false;
81
82 enum isl_format rt_format =
83 iris_format_for_usage(devinfo, pfmt,
84 ISL_SURF_USAGE_RENDER_TARGET_BIT).fmt;
85
86 enum isl_format linear_format = isl_format_srgb_to_linear(rt_format);
87
88 if (!isl_format_supports_ccs_e(devinfo, linear_format))
89 return false;
90
91 return devinfo->gen >= 9 && devinfo->gen <= 11;
92 }
93 case I915_FORMAT_MOD_Y_TILED:
94 case I915_FORMAT_MOD_X_TILED:
95 case DRM_FORMAT_MOD_LINEAR:
96 return true;
97 case DRM_FORMAT_MOD_INVALID:
98 default:
99 return false;
100 }
101 }
102
103 static uint64_t
104 select_best_modifier(struct gen_device_info *devinfo, enum pipe_format pfmt,
105 const uint64_t *modifiers,
106 int count)
107 {
108 enum modifier_priority prio = MODIFIER_PRIORITY_INVALID;
109
110 for (int i = 0; i < count; i++) {
111 if (!modifier_is_supported(devinfo, pfmt, modifiers[i]))
112 continue;
113
114 switch (modifiers[i]) {
115 case I915_FORMAT_MOD_Y_TILED_CCS:
116 prio = MAX2(prio, MODIFIER_PRIORITY_Y_CCS);
117 break;
118 case I915_FORMAT_MOD_Y_TILED:
119 prio = MAX2(prio, MODIFIER_PRIORITY_Y);
120 break;
121 case I915_FORMAT_MOD_X_TILED:
122 prio = MAX2(prio, MODIFIER_PRIORITY_X);
123 break;
124 case DRM_FORMAT_MOD_LINEAR:
125 prio = MAX2(prio, MODIFIER_PRIORITY_LINEAR);
126 break;
127 case DRM_FORMAT_MOD_INVALID:
128 default:
129 break;
130 }
131 }
132
133 return priority_to_modifier[prio];
134 }
135
136 enum isl_surf_dim
137 target_to_isl_surf_dim(enum pipe_texture_target target)
138 {
139 switch (target) {
140 case PIPE_BUFFER:
141 case PIPE_TEXTURE_1D:
142 case PIPE_TEXTURE_1D_ARRAY:
143 return ISL_SURF_DIM_1D;
144 case PIPE_TEXTURE_2D:
145 case PIPE_TEXTURE_CUBE:
146 case PIPE_TEXTURE_RECT:
147 case PIPE_TEXTURE_2D_ARRAY:
148 case PIPE_TEXTURE_CUBE_ARRAY:
149 return ISL_SURF_DIM_2D;
150 case PIPE_TEXTURE_3D:
151 return ISL_SURF_DIM_3D;
152 case PIPE_MAX_TEXTURE_TYPES:
153 break;
154 }
155 unreachable("invalid texture type");
156 }
157
158 static void
159 iris_query_dmabuf_modifiers(struct pipe_screen *pscreen,
160 enum pipe_format pfmt,
161 int max,
162 uint64_t *modifiers,
163 unsigned int *external_only,
164 int *count)
165 {
166 struct iris_screen *screen = (void *) pscreen;
167 const struct gen_device_info *devinfo = &screen->devinfo;
168
169 uint64_t all_modifiers[] = {
170 DRM_FORMAT_MOD_LINEAR,
171 I915_FORMAT_MOD_X_TILED,
172 I915_FORMAT_MOD_Y_TILED,
173 I915_FORMAT_MOD_Y_TILED_CCS,
174 };
175
176 int supported_mods = 0;
177
178 for (int i = 0; i < ARRAY_SIZE(all_modifiers); i++) {
179 if (!modifier_is_supported(devinfo, pfmt, all_modifiers[i]))
180 continue;
181
182 if (supported_mods < max) {
183 if (modifiers)
184 modifiers[supported_mods] = all_modifiers[i];
185
186 if (external_only)
187 external_only[supported_mods] = util_format_is_yuv(pfmt);
188 }
189
190 supported_mods++;
191 }
192
193 *count = supported_mods;
194 }
195
196 static isl_surf_usage_flags_t
197 pipe_bind_to_isl_usage(unsigned bindings)
198 {
199 isl_surf_usage_flags_t usage = 0;
200
201 if (bindings & PIPE_BIND_RENDER_TARGET)
202 usage |= ISL_SURF_USAGE_RENDER_TARGET_BIT;
203
204 if (bindings & PIPE_BIND_SAMPLER_VIEW)
205 usage |= ISL_SURF_USAGE_TEXTURE_BIT;
206
207 if (bindings & (PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SHADER_BUFFER))
208 usage |= ISL_SURF_USAGE_STORAGE_BIT;
209
210 if (bindings & PIPE_BIND_DISPLAY_TARGET)
211 usage |= ISL_SURF_USAGE_DISPLAY_BIT;
212
213 return usage;
214 }
215
216 struct pipe_resource *
217 iris_resource_get_separate_stencil(struct pipe_resource *p_res)
218 {
219 /* For packed depth-stencil, we treat depth as the primary resource
220 * and store S8 as the "second plane" resource.
221 */
222 if (p_res->next && p_res->next->format == PIPE_FORMAT_S8_UINT)
223 return p_res->next;
224
225 return NULL;
226
227 }
228
229 static void
230 iris_resource_set_separate_stencil(struct pipe_resource *p_res,
231 struct pipe_resource *stencil)
232 {
233 assert(util_format_has_depth(util_format_description(p_res->format)));
234 pipe_resource_reference(&p_res->next, stencil);
235 }
236
237 void
238 iris_get_depth_stencil_resources(struct pipe_resource *res,
239 struct iris_resource **out_z,
240 struct iris_resource **out_s)
241 {
242 if (!res) {
243 *out_z = NULL;
244 *out_s = NULL;
245 return;
246 }
247
248 if (res->format != PIPE_FORMAT_S8_UINT) {
249 *out_z = (void *) res;
250 *out_s = (void *) iris_resource_get_separate_stencil(res);
251 } else {
252 *out_z = NULL;
253 *out_s = (void *) res;
254 }
255 }
256
257 enum isl_dim_layout
258 iris_get_isl_dim_layout(const struct gen_device_info *devinfo,
259 enum isl_tiling tiling,
260 enum pipe_texture_target target)
261 {
262 switch (target) {
263 case PIPE_TEXTURE_1D:
264 case PIPE_TEXTURE_1D_ARRAY:
265 return (devinfo->gen >= 9 && tiling == ISL_TILING_LINEAR ?
266 ISL_DIM_LAYOUT_GEN9_1D : ISL_DIM_LAYOUT_GEN4_2D);
267
268 case PIPE_TEXTURE_2D:
269 case PIPE_TEXTURE_2D_ARRAY:
270 case PIPE_TEXTURE_RECT:
271 case PIPE_TEXTURE_CUBE:
272 case PIPE_TEXTURE_CUBE_ARRAY:
273 return ISL_DIM_LAYOUT_GEN4_2D;
274
275 case PIPE_TEXTURE_3D:
276 return (devinfo->gen >= 9 ?
277 ISL_DIM_LAYOUT_GEN4_2D : ISL_DIM_LAYOUT_GEN4_3D);
278
279 case PIPE_MAX_TEXTURE_TYPES:
280 case PIPE_BUFFER:
281 break;
282 }
283 unreachable("invalid texture type");
284 }
285
286 void
287 iris_resource_disable_aux(struct iris_resource *res)
288 {
289 iris_bo_unreference(res->aux.bo);
290 iris_bo_unreference(res->aux.extra_aux.bo);
291 iris_bo_unreference(res->aux.clear_color_bo);
292 free(res->aux.state);
293
294 res->aux.usage = ISL_AUX_USAGE_NONE;
295 res->aux.possible_usages = 1 << ISL_AUX_USAGE_NONE;
296 res->aux.sampler_usages = 1 << ISL_AUX_USAGE_NONE;
297 res->aux.has_hiz = 0;
298 res->aux.surf.size_B = 0;
299 res->aux.bo = NULL;
300 res->aux.extra_aux.surf.size_B = 0;
301 res->aux.extra_aux.bo = NULL;
302 res->aux.clear_color_bo = NULL;
303 res->aux.state = NULL;
304 }
305
306 static void
307 iris_resource_destroy(struct pipe_screen *screen,
308 struct pipe_resource *resource)
309 {
310 struct iris_resource *res = (struct iris_resource *)resource;
311
312 if (resource->target == PIPE_BUFFER)
313 util_range_destroy(&res->valid_buffer_range);
314
315 iris_resource_disable_aux(res);
316
317 iris_bo_unreference(res->bo);
318 free(res);
319 }
320
321 static struct iris_resource *
322 iris_alloc_resource(struct pipe_screen *pscreen,
323 const struct pipe_resource *templ)
324 {
325 struct iris_resource *res = calloc(1, sizeof(struct iris_resource));
326 if (!res)
327 return NULL;
328
329 res->base = *templ;
330 res->base.screen = pscreen;
331 pipe_reference_init(&res->base.reference, 1);
332
333 res->aux.possible_usages = 1 << ISL_AUX_USAGE_NONE;
334 res->aux.sampler_usages = 1 << ISL_AUX_USAGE_NONE;
335
336 if (templ->target == PIPE_BUFFER)
337 util_range_init(&res->valid_buffer_range);
338
339 return res;
340 }
341
342 unsigned
343 iris_get_num_logical_layers(const struct iris_resource *res, unsigned level)
344 {
345 if (res->surf.dim == ISL_SURF_DIM_3D)
346 return minify(res->surf.logical_level0_px.depth, level);
347 else
348 return res->surf.logical_level0_px.array_len;
349 }
350
351 static enum isl_aux_state **
352 create_aux_state_map(struct iris_resource *res, enum isl_aux_state initial)
353 {
354 uint32_t total_slices = 0;
355 for (uint32_t level = 0; level < res->surf.levels; level++)
356 total_slices += iris_get_num_logical_layers(res, level);
357
358 const size_t per_level_array_size =
359 res->surf.levels * sizeof(enum isl_aux_state *);
360
361 /* We're going to allocate a single chunk of data for both the per-level
362 * reference array and the arrays of aux_state. This makes cleanup
363 * significantly easier.
364 */
365 const size_t total_size =
366 per_level_array_size + total_slices * sizeof(enum isl_aux_state);
367
368 void *data = malloc(total_size);
369 if (!data)
370 return NULL;
371
372 enum isl_aux_state **per_level_arr = data;
373 enum isl_aux_state *s = data + per_level_array_size;
374 for (uint32_t level = 0; level < res->surf.levels; level++) {
375 per_level_arr[level] = s;
376 const unsigned level_layers = iris_get_num_logical_layers(res, level);
377 for (uint32_t a = 0; a < level_layers; a++)
378 *(s++) = initial;
379 }
380 assert((void *)s == data + total_size);
381
382 return per_level_arr;
383 }
384
385 static unsigned
386 iris_get_aux_clear_color_state_size(struct iris_screen *screen)
387 {
388 const struct gen_device_info *devinfo = &screen->devinfo;
389 return devinfo->gen >= 10 ? screen->isl_dev.ss.clear_color_state_size : 0;
390 }
391
392 static void
393 map_aux_addresses(struct iris_screen *screen, struct iris_resource *res)
394 {
395 const struct gen_device_info *devinfo = &screen->devinfo;
396 if (devinfo->gen >= 12 && isl_aux_usage_has_ccs(res->aux.usage)) {
397 void *aux_map_ctx = iris_bufmgr_get_aux_map_context(screen->bufmgr);
398 assert(aux_map_ctx);
399 const bool has_extra_ccs = res->aux.extra_aux.surf.size_B > 0;
400 struct iris_bo *aux_bo = has_extra_ccs ?
401 res->aux.extra_aux.bo : res->aux.bo;
402 const unsigned aux_offset = has_extra_ccs ?
403 res->aux.extra_aux.offset : res->aux.offset;
404 gen_aux_map_add_image(aux_map_ctx, &res->surf, res->bo->gtt_offset,
405 aux_bo->gtt_offset + aux_offset);
406 res->bo->aux_map_address = aux_bo->gtt_offset;
407 }
408 }
409
410 static bool
411 want_ccs_e_for_format(const struct gen_device_info *devinfo,
412 enum isl_format format)
413 {
414 if (!isl_format_supports_ccs_e(devinfo, format))
415 return false;
416
417 const struct isl_format_layout *fmtl = isl_format_get_layout(format);
418
419 /* CCS_E seems to significantly hurt performance with 32-bit floating
420 * point formats. For example, Paraview's "Wavelet Volume" case uses
421 * both R32_FLOAT and R32G32B32A32_FLOAT, and enabling CCS_E for those
422 * formats causes a 62% FPS drop.
423 *
424 * However, many benchmarks seem to use 16-bit float with no issues.
425 */
426 if (fmtl->channels.r.bits == 32 && fmtl->channels.r.type == ISL_SFLOAT)
427 return false;
428
429 return true;
430 }
431
432 /**
433 * Configure aux for the resource, but don't allocate it. For images which
434 * might be shared with modifiers, we must allocate the image and aux data in
435 * a single bo.
436 */
437 static bool
438 iris_resource_configure_aux(struct iris_screen *screen,
439 struct iris_resource *res, bool imported,
440 uint64_t *aux_size_B,
441 uint32_t *alloc_flags)
442 {
443 const struct gen_device_info *devinfo = &screen->devinfo;
444
445 /* Try to create the auxiliary surfaces allowed by the modifier or by
446 * the user if no modifier is specified.
447 */
448 assert(!res->mod_info || res->mod_info->aux_usage == ISL_AUX_USAGE_NONE ||
449 res->mod_info->aux_usage == ISL_AUX_USAGE_CCS_E);
450
451 const bool has_mcs = !res->mod_info &&
452 isl_surf_get_mcs_surf(&screen->isl_dev, &res->surf, &res->aux.surf);
453
454 const bool has_hiz = !res->mod_info && !(INTEL_DEBUG & DEBUG_NO_HIZ) &&
455 isl_surf_get_hiz_surf(&screen->isl_dev, &res->surf, &res->aux.surf);
456
457 const bool has_ccs =
458 ((!res->mod_info && !(INTEL_DEBUG & DEBUG_NO_RBC)) ||
459 (res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE)) &&
460 isl_surf_get_ccs_surf(&screen->isl_dev, &res->surf, &res->aux.surf,
461 &res->aux.extra_aux.surf, 0);
462
463 /* Having both HIZ and MCS is impossible. */
464 assert(!has_mcs || !has_hiz);
465
466 /* Ensure aux surface creation for MCS_CCS and HIZ_CCS is correct. */
467 if (has_ccs && (has_mcs || has_hiz)) {
468 assert(res->aux.extra_aux.surf.size_B > 0 &&
469 res->aux.extra_aux.surf.usage & ISL_SURF_USAGE_CCS_BIT);
470 assert(res->aux.surf.size_B > 0 &&
471 res->aux.surf.usage &
472 (ISL_SURF_USAGE_HIZ_BIT | ISL_SURF_USAGE_MCS_BIT));
473 }
474
475 if (res->mod_info && has_ccs) {
476 /* Only allow a CCS modifier if the aux was created successfully. */
477 res->aux.possible_usages |= 1 << res->mod_info->aux_usage;
478 } else if (has_mcs) {
479 res->aux.possible_usages |= 1 << ISL_AUX_USAGE_MCS;
480 } else if (has_hiz) {
481 res->aux.possible_usages |= 1 << ISL_AUX_USAGE_HIZ;
482 } else if (has_ccs) {
483 if (want_ccs_e_for_format(devinfo, res->surf.format))
484 res->aux.possible_usages |= 1 << ISL_AUX_USAGE_CCS_E;
485
486 if (isl_format_supports_ccs_d(devinfo, res->surf.format))
487 res->aux.possible_usages |= 1 << ISL_AUX_USAGE_CCS_D;
488 }
489
490 res->aux.usage = util_last_bit(res->aux.possible_usages) - 1;
491
492 res->aux.sampler_usages = res->aux.possible_usages;
493
494 /* We don't always support sampling with hiz. But when we do, it must be
495 * single sampled.
496 */
497 if (!devinfo->has_sample_with_hiz || res->surf.samples > 1)
498 res->aux.sampler_usages &= ~(1 << ISL_AUX_USAGE_HIZ);
499
500 enum isl_aux_state initial_state;
501 *aux_size_B = 0;
502 *alloc_flags = 0;
503 assert(!res->aux.bo);
504
505 switch (res->aux.usage) {
506 case ISL_AUX_USAGE_NONE:
507 /* Having no aux buffer is only okay if there's no modifier with aux. */
508 return !res->mod_info || res->mod_info->aux_usage == ISL_AUX_USAGE_NONE;
509 case ISL_AUX_USAGE_HIZ:
510 initial_state = ISL_AUX_STATE_AUX_INVALID;
511 break;
512 case ISL_AUX_USAGE_MCS:
513 /* The Ivybridge PRM, Vol 2 Part 1 p326 says:
514 *
515 * "When MCS buffer is enabled and bound to MSRT, it is required
516 * that it is cleared prior to any rendering."
517 *
518 * Since we only use the MCS buffer for rendering, we just clear it
519 * immediately on allocation. The clear value for MCS buffers is all
520 * 1's, so we simply memset it to 0xff.
521 */
522 initial_state = ISL_AUX_STATE_CLEAR;
523 break;
524 case ISL_AUX_USAGE_CCS_D:
525 case ISL_AUX_USAGE_CCS_E:
526 /* When CCS_E is used, we need to ensure that the CCS starts off in
527 * a valid state. From the Sky Lake PRM, "MCS Buffer for Render
528 * Target(s)":
529 *
530 * "If Software wants to enable Color Compression without Fast
531 * clear, Software needs to initialize MCS with zeros."
532 *
533 * A CCS value of 0 indicates that the corresponding block is in the
534 * pass-through state which is what we want.
535 *
536 * For CCS_D, do the same thing. On Gen9+, this avoids having any
537 * undefined bits in the aux buffer.
538 */
539 if (imported)
540 initial_state =
541 isl_drm_modifier_get_default_aux_state(res->mod_info->modifier);
542 else
543 initial_state = ISL_AUX_STATE_PASS_THROUGH;
544 *alloc_flags |= BO_ALLOC_ZEROED;
545 break;
546 }
547
548 if (!res->aux.state) {
549 /* Create the aux_state for the auxiliary buffer. */
550 res->aux.state = create_aux_state_map(res, initial_state);
551 if (!res->aux.state)
552 return false;
553 }
554
555 uint64_t size = res->aux.surf.size_B;
556
557 /* Allocate space in the buffer for storing the CCS. */
558 if (res->aux.extra_aux.surf.size_B > 0) {
559 res->aux.extra_aux.offset =
560 ALIGN(size, res->aux.extra_aux.surf.alignment_B);
561 size = res->aux.extra_aux.offset + res->aux.extra_aux.surf.size_B;
562 }
563
564 /* Allocate space in the buffer for storing the clear color. On modern
565 * platforms (gen > 9), we can read it directly from such buffer.
566 *
567 * On gen <= 9, we are going to store the clear color on the buffer
568 * anyways, and copy it back to the surface state during state emission.
569 */
570 res->aux.clear_color_offset = size;
571 size += iris_get_aux_clear_color_state_size(screen);
572 *aux_size_B = size;
573
574 if (res->aux.usage == ISL_AUX_USAGE_HIZ) {
575 for (unsigned level = 0; level < res->surf.levels; ++level) {
576 uint32_t width = u_minify(res->surf.phys_level0_sa.width, level);
577 uint32_t height = u_minify(res->surf.phys_level0_sa.height, level);
578
579 /* Disable HiZ for LOD > 0 unless the width/height are 8x4 aligned.
580 * For LOD == 0, we can grow the dimensions to make it work.
581 */
582 if (level == 0 || ((width & 7) == 0 && (height & 3) == 0))
583 res->aux.has_hiz |= 1 << level;
584 }
585 }
586
587 return true;
588 }
589
590 /**
591 * Initialize the aux buffer contents.
592 */
593 static bool
594 iris_resource_init_aux_buf(struct iris_resource *res, uint32_t alloc_flags,
595 unsigned clear_color_state_size)
596 {
597 if (!(alloc_flags & BO_ALLOC_ZEROED)) {
598 void *map = iris_bo_map(NULL, res->aux.bo, MAP_WRITE | MAP_RAW);
599
600 if (!map) {
601 iris_resource_disable_aux(res);
602 return false;
603 }
604
605 if (iris_resource_get_aux_state(res, 0, 0) != ISL_AUX_STATE_AUX_INVALID) {
606 uint8_t memset_value = res->aux.usage == ISL_AUX_USAGE_MCS ? 0xFF : 0;
607 memset((char*)map + res->aux.offset, memset_value,
608 res->aux.surf.size_B);
609 }
610
611 /* Resolved is usually a safe state for CCS_E. */
612 memset((char*)map + res->aux.extra_aux.offset, 0,
613 res->aux.extra_aux.surf.size_B);
614
615 /* Zero the indirect clear color to match ::fast_clear_color. */
616 memset((char *)map + res->aux.clear_color_offset, 0,
617 clear_color_state_size);
618
619 iris_bo_unmap(res->aux.bo);
620 }
621
622 if (res->aux.extra_aux.surf.size_B > 0) {
623 res->aux.extra_aux.bo = res->aux.bo;
624 iris_bo_reference(res->aux.extra_aux.bo);
625 }
626
627 if (clear_color_state_size > 0) {
628 res->aux.clear_color_bo = res->aux.bo;
629 iris_bo_reference(res->aux.clear_color_bo);
630 }
631
632 return true;
633 }
634
635 /**
636 * Allocate the initial aux surface for a resource based on aux.usage
637 */
638 static bool
639 iris_resource_alloc_separate_aux(struct iris_screen *screen,
640 struct iris_resource *res)
641 {
642 uint32_t alloc_flags;
643 uint64_t size;
644 if (!iris_resource_configure_aux(screen, res, false, &size, &alloc_flags))
645 return false;
646
647 if (size == 0)
648 return true;
649
650 /* Allocate the auxiliary buffer. ISL has stricter set of alignment rules
651 * the drm allocator. Therefore, one can pass the ISL dimensions in terms
652 * of bytes instead of trying to recalculate based on different format
653 * block sizes.
654 */
655 res->aux.bo = iris_bo_alloc_tiled(screen->bufmgr, "aux buffer", size, 4096,
656 IRIS_MEMZONE_OTHER,
657 isl_tiling_to_i915_tiling(res->aux.surf.tiling),
658 res->aux.surf.row_pitch_B, alloc_flags);
659 if (!res->aux.bo) {
660 return false;
661 }
662
663 if (!iris_resource_init_aux_buf(res, alloc_flags,
664 iris_get_aux_clear_color_state_size(screen)))
665 return false;
666
667 map_aux_addresses(screen, res);
668
669 return true;
670 }
671
672 void
673 iris_resource_finish_aux_import(struct pipe_screen *pscreen,
674 struct iris_resource *res)
675 {
676 struct iris_screen *screen = (struct iris_screen *)pscreen;
677 assert(iris_resource_unfinished_aux_import(res));
678 assert(!res->mod_info->supports_clear_color);
679
680 struct iris_resource *aux_res = (void *) res->base.next;
681 assert(aux_res->aux.surf.row_pitch_B && aux_res->aux.offset &&
682 aux_res->aux.bo);
683
684 assert(res->bo == aux_res->aux.bo);
685 iris_bo_reference(aux_res->aux.bo);
686 res->aux.bo = aux_res->aux.bo;
687
688 res->aux.offset = aux_res->aux.offset;
689
690 assert(res->bo->size >= (res->aux.offset + res->aux.surf.size_B));
691 assert(res->aux.clear_color_bo == NULL);
692 res->aux.clear_color_offset = 0;
693
694 assert(aux_res->aux.surf.row_pitch_B == res->aux.surf.row_pitch_B);
695
696 unsigned clear_color_state_size =
697 iris_get_aux_clear_color_state_size(screen);
698
699 if (clear_color_state_size > 0) {
700 res->aux.clear_color_bo =
701 iris_bo_alloc(screen->bufmgr, "clear color buffer",
702 clear_color_state_size, IRIS_MEMZONE_OTHER);
703 res->aux.clear_color_offset = 0;
704 }
705
706 iris_resource_destroy(&screen->base, res->base.next);
707 res->base.next = NULL;
708 }
709
710 static struct pipe_resource *
711 iris_resource_create_for_buffer(struct pipe_screen *pscreen,
712 const struct pipe_resource *templ)
713 {
714 struct iris_screen *screen = (struct iris_screen *)pscreen;
715 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
716
717 assert(templ->target == PIPE_BUFFER);
718 assert(templ->height0 <= 1);
719 assert(templ->depth0 <= 1);
720 assert(templ->format == PIPE_FORMAT_NONE ||
721 util_format_get_blocksize(templ->format) == 1);
722
723 res->internal_format = templ->format;
724 res->surf.tiling = ISL_TILING_LINEAR;
725
726 enum iris_memory_zone memzone = IRIS_MEMZONE_OTHER;
727 const char *name = templ->target == PIPE_BUFFER ? "buffer" : "miptree";
728 if (templ->flags & IRIS_RESOURCE_FLAG_SHADER_MEMZONE) {
729 memzone = IRIS_MEMZONE_SHADER;
730 name = "shader kernels";
731 } else if (templ->flags & IRIS_RESOURCE_FLAG_SURFACE_MEMZONE) {
732 memzone = IRIS_MEMZONE_SURFACE;
733 name = "surface state";
734 } else if (templ->flags & IRIS_RESOURCE_FLAG_DYNAMIC_MEMZONE) {
735 memzone = IRIS_MEMZONE_DYNAMIC;
736 name = "dynamic state";
737 }
738
739 res->bo = iris_bo_alloc(screen->bufmgr, name, templ->width0, memzone);
740 if (!res->bo) {
741 iris_resource_destroy(pscreen, &res->base);
742 return NULL;
743 }
744
745 return &res->base;
746 }
747
748 static struct pipe_resource *
749 iris_resource_create_with_modifiers(struct pipe_screen *pscreen,
750 const struct pipe_resource *templ,
751 const uint64_t *modifiers,
752 int modifiers_count)
753 {
754 struct iris_screen *screen = (struct iris_screen *)pscreen;
755 struct gen_device_info *devinfo = &screen->devinfo;
756 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
757
758 if (!res)
759 return NULL;
760
761 const struct util_format_description *format_desc =
762 util_format_description(templ->format);
763 const bool has_depth = util_format_has_depth(format_desc);
764 uint64_t modifier =
765 select_best_modifier(devinfo, templ->format, modifiers, modifiers_count);
766
767 isl_tiling_flags_t tiling_flags = ISL_TILING_ANY_MASK;
768
769 if (modifier != DRM_FORMAT_MOD_INVALID) {
770 res->mod_info = isl_drm_modifier_get_info(modifier);
771
772 tiling_flags = 1 << res->mod_info->tiling;
773 } else {
774 if (modifiers_count > 0) {
775 fprintf(stderr, "Unsupported modifier, resource creation failed.\n");
776 return NULL;
777 }
778
779 /* Use linear for staging buffers */
780 if (templ->usage == PIPE_USAGE_STAGING ||
781 templ->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR) )
782 tiling_flags = ISL_TILING_LINEAR_BIT;
783 }
784
785 isl_surf_usage_flags_t usage = pipe_bind_to_isl_usage(templ->bind);
786
787 if (templ->target == PIPE_TEXTURE_CUBE ||
788 templ->target == PIPE_TEXTURE_CUBE_ARRAY)
789 usage |= ISL_SURF_USAGE_CUBE_BIT;
790
791 if (templ->usage != PIPE_USAGE_STAGING) {
792 if (templ->format == PIPE_FORMAT_S8_UINT)
793 usage |= ISL_SURF_USAGE_STENCIL_BIT;
794 else if (has_depth)
795 usage |= ISL_SURF_USAGE_DEPTH_BIT;
796 }
797
798 enum pipe_format pfmt = templ->format;
799 res->internal_format = pfmt;
800
801 /* Should be handled by u_transfer_helper */
802 assert(!util_format_is_depth_and_stencil(pfmt));
803
804 struct iris_format_info fmt = iris_format_for_usage(devinfo, pfmt, usage);
805 assert(fmt.fmt != ISL_FORMAT_UNSUPPORTED);
806
807 UNUSED const bool isl_surf_created_successfully =
808 isl_surf_init(&screen->isl_dev, &res->surf,
809 .dim = target_to_isl_surf_dim(templ->target),
810 .format = fmt.fmt,
811 .width = templ->width0,
812 .height = templ->height0,
813 .depth = templ->depth0,
814 .levels = templ->last_level + 1,
815 .array_len = templ->array_size,
816 .samples = MAX2(templ->nr_samples, 1),
817 .min_alignment_B = 0,
818 .row_pitch_B = 0,
819 .usage = usage,
820 .tiling_flags = tiling_flags);
821 assert(isl_surf_created_successfully);
822
823 const char *name = "miptree";
824 enum iris_memory_zone memzone = IRIS_MEMZONE_OTHER;
825
826 unsigned int flags = 0;
827 if (templ->usage == PIPE_USAGE_STAGING)
828 flags |= BO_ALLOC_COHERENT;
829
830 /* These are for u_upload_mgr buffers only */
831 assert(!(templ->flags & (IRIS_RESOURCE_FLAG_SHADER_MEMZONE |
832 IRIS_RESOURCE_FLAG_SURFACE_MEMZONE |
833 IRIS_RESOURCE_FLAG_DYNAMIC_MEMZONE)));
834
835 uint32_t aux_preferred_alloc_flags;
836 uint64_t aux_size = 0;
837 bool aux_enabled =
838 iris_resource_configure_aux(screen, res, false, &aux_size,
839 &aux_preferred_alloc_flags);
840 aux_enabled = aux_enabled && res->aux.surf.size_B > 0;
841 const bool separate_aux = aux_enabled && !res->mod_info;
842 uint64_t aux_offset;
843 uint64_t bo_size;
844
845 if (aux_enabled && !separate_aux) {
846 /* Allocate aux data with main surface. This is required for modifiers
847 * with aux data (ccs).
848 */
849 aux_offset = ALIGN(res->surf.size_B, res->aux.surf.alignment_B);
850 bo_size = aux_offset + aux_size;
851 } else {
852 aux_offset = 0;
853 bo_size = res->surf.size_B;
854 }
855
856 uint32_t alignment = MAX2(4096, res->surf.alignment_B);
857 res->bo = iris_bo_alloc_tiled(screen->bufmgr, name, bo_size, alignment,
858 memzone,
859 isl_tiling_to_i915_tiling(res->surf.tiling),
860 res->surf.row_pitch_B, flags);
861
862 if (!res->bo)
863 goto fail;
864
865 if (aux_enabled) {
866 if (separate_aux) {
867 if (!iris_resource_alloc_separate_aux(screen, res))
868 aux_enabled = false;
869 } else {
870 res->aux.bo = res->bo;
871 iris_bo_reference(res->aux.bo);
872 res->aux.offset += aux_offset;
873 unsigned clear_color_state_size =
874 iris_get_aux_clear_color_state_size(screen);
875 if (clear_color_state_size > 0)
876 res->aux.clear_color_offset += aux_offset;
877 if (!iris_resource_init_aux_buf(res, flags, clear_color_state_size))
878 aux_enabled = false;
879 map_aux_addresses(screen, res);
880 }
881 }
882
883 if (!aux_enabled)
884 iris_resource_disable_aux(res);
885
886 return &res->base;
887
888 fail:
889 fprintf(stderr, "XXX: resource creation failed\n");
890 iris_resource_destroy(pscreen, &res->base);
891 return NULL;
892
893 }
894
895 static struct pipe_resource *
896 iris_resource_create(struct pipe_screen *pscreen,
897 const struct pipe_resource *templ)
898 {
899 if (templ->target == PIPE_BUFFER)
900 return iris_resource_create_for_buffer(pscreen, templ);
901 else
902 return iris_resource_create_with_modifiers(pscreen, templ, NULL, 0);
903 }
904
905 static uint64_t
906 tiling_to_modifier(uint32_t tiling)
907 {
908 static const uint64_t map[] = {
909 [I915_TILING_NONE] = DRM_FORMAT_MOD_LINEAR,
910 [I915_TILING_X] = I915_FORMAT_MOD_X_TILED,
911 [I915_TILING_Y] = I915_FORMAT_MOD_Y_TILED,
912 };
913
914 assert(tiling < ARRAY_SIZE(map));
915
916 return map[tiling];
917 }
918
919 static struct pipe_resource *
920 iris_resource_from_user_memory(struct pipe_screen *pscreen,
921 const struct pipe_resource *templ,
922 void *user_memory)
923 {
924 struct iris_screen *screen = (struct iris_screen *)pscreen;
925 struct iris_bufmgr *bufmgr = screen->bufmgr;
926 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
927 if (!res)
928 return NULL;
929
930 assert(templ->target == PIPE_BUFFER);
931
932 res->internal_format = templ->format;
933 res->bo = iris_bo_create_userptr(bufmgr, "user",
934 user_memory, templ->width0,
935 IRIS_MEMZONE_OTHER);
936 if (!res->bo) {
937 free(res);
938 return NULL;
939 }
940
941 util_range_add(&res->base, &res->valid_buffer_range, 0, templ->width0);
942
943 return &res->base;
944 }
945
946 static struct pipe_resource *
947 iris_resource_from_handle(struct pipe_screen *pscreen,
948 const struct pipe_resource *templ,
949 struct winsys_handle *whandle,
950 unsigned usage)
951 {
952 struct iris_screen *screen = (struct iris_screen *)pscreen;
953 struct gen_device_info *devinfo = &screen->devinfo;
954 struct iris_bufmgr *bufmgr = screen->bufmgr;
955 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
956 if (!res)
957 return NULL;
958
959 switch (whandle->type) {
960 case WINSYS_HANDLE_TYPE_FD:
961 res->bo = iris_bo_import_dmabuf(bufmgr, whandle->handle);
962 break;
963 case WINSYS_HANDLE_TYPE_SHARED:
964 res->bo = iris_bo_gem_create_from_name(bufmgr, "winsys image",
965 whandle->handle);
966 break;
967 default:
968 unreachable("invalid winsys handle type");
969 }
970 if (!res->bo)
971 return NULL;
972
973 res->offset = whandle->offset;
974
975 uint64_t modifier = whandle->modifier;
976 if (modifier == DRM_FORMAT_MOD_INVALID) {
977 modifier = tiling_to_modifier(res->bo->tiling_mode);
978 }
979 res->mod_info = isl_drm_modifier_get_info(modifier);
980 assert(res->mod_info);
981
982 isl_surf_usage_flags_t isl_usage = pipe_bind_to_isl_usage(templ->bind);
983
984 const struct iris_format_info fmt =
985 iris_format_for_usage(devinfo, templ->format, isl_usage);
986 res->internal_format = templ->format;
987
988 if (templ->target == PIPE_BUFFER) {
989 res->surf.tiling = ISL_TILING_LINEAR;
990 } else {
991 if (whandle->modifier == DRM_FORMAT_MOD_INVALID || whandle->plane == 0) {
992 UNUSED const bool isl_surf_created_successfully =
993 isl_surf_init(&screen->isl_dev, &res->surf,
994 .dim = target_to_isl_surf_dim(templ->target),
995 .format = fmt.fmt,
996 .width = templ->width0,
997 .height = templ->height0,
998 .depth = templ->depth0,
999 .levels = templ->last_level + 1,
1000 .array_len = templ->array_size,
1001 .samples = MAX2(templ->nr_samples, 1),
1002 .min_alignment_B = 0,
1003 .row_pitch_B = whandle->stride,
1004 .usage = isl_usage,
1005 .tiling_flags = 1 << res->mod_info->tiling);
1006 assert(isl_surf_created_successfully);
1007 assert(res->bo->tiling_mode ==
1008 isl_tiling_to_i915_tiling(res->surf.tiling));
1009
1010 // XXX: create_ccs_buf_for_image?
1011 if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
1012 if (!iris_resource_alloc_separate_aux(screen, res))
1013 goto fail;
1014 } else {
1015 if (res->mod_info->aux_usage != ISL_AUX_USAGE_NONE) {
1016 uint32_t alloc_flags;
1017 uint64_t size;
1018 bool ok = iris_resource_configure_aux(screen, res, true, &size,
1019 &alloc_flags);
1020 assert(ok);
1021 /* The gallium dri layer will create a separate plane resource
1022 * for the aux image. iris_resource_finish_aux_import will
1023 * merge the separate aux parameters back into a single
1024 * iris_resource.
1025 */
1026 }
1027 }
1028 } else {
1029 /* Save modifier import information to reconstruct later. After
1030 * import, this will be available under a second image accessible
1031 * from the main image with res->base.next. See
1032 * iris_resource_finish_aux_import.
1033 */
1034 res->aux.surf.row_pitch_B = whandle->stride;
1035 res->aux.offset = whandle->offset;
1036 res->aux.bo = res->bo;
1037 res->bo = NULL;
1038 }
1039 }
1040
1041 return &res->base;
1042
1043 fail:
1044 iris_resource_destroy(pscreen, &res->base);
1045 return NULL;
1046 }
1047
1048 static void
1049 iris_flush_resource(struct pipe_context *ctx, struct pipe_resource *resource)
1050 {
1051 struct iris_context *ice = (struct iris_context *)ctx;
1052 struct iris_batch *render_batch = &ice->batches[IRIS_BATCH_RENDER];
1053 struct iris_resource *res = (void *) resource;
1054 const struct isl_drm_modifier_info *mod = res->mod_info;
1055
1056 iris_resource_prepare_access(ice, render_batch, res,
1057 0, INTEL_REMAINING_LEVELS,
1058 0, INTEL_REMAINING_LAYERS,
1059 mod ? mod->aux_usage : ISL_AUX_USAGE_NONE,
1060 mod ? mod->supports_clear_color : false);
1061 }
1062
1063 static void
1064 iris_resource_disable_aux_on_first_query(struct pipe_resource *resource,
1065 unsigned usage)
1066 {
1067 struct iris_resource *res = (struct iris_resource *)resource;
1068 bool mod_with_aux =
1069 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1070
1071 /* Disable aux usage if explicit flush not set and this is the first time
1072 * we are dealing with this resource and the resource was not created with
1073 * a modifier with aux.
1074 */
1075 if (!mod_with_aux &&
1076 (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) && res->aux.usage != 0) &&
1077 p_atomic_read(&resource->reference.count) == 1) {
1078 iris_resource_disable_aux(res);
1079 }
1080 }
1081
1082 static bool
1083 iris_resource_get_param(struct pipe_screen *screen,
1084 struct pipe_context *context,
1085 struct pipe_resource *resource,
1086 unsigned plane,
1087 unsigned layer,
1088 enum pipe_resource_param param,
1089 unsigned handle_usage,
1090 uint64_t *value)
1091 {
1092 struct iris_resource *res = (struct iris_resource *)resource;
1093 bool mod_with_aux =
1094 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1095 bool wants_aux = mod_with_aux && plane > 0;
1096 bool result;
1097 unsigned handle;
1098
1099 if (iris_resource_unfinished_aux_import(res))
1100 iris_resource_finish_aux_import(screen, res);
1101
1102 struct iris_bo *bo = wants_aux ? res->aux.bo : res->bo;
1103
1104 iris_resource_disable_aux_on_first_query(resource, handle_usage);
1105
1106 switch (param) {
1107 case PIPE_RESOURCE_PARAM_NPLANES:
1108 if (mod_with_aux) {
1109 *value = 2;
1110 } else {
1111 unsigned count = 0;
1112 for (struct pipe_resource *cur = resource; cur; cur = cur->next)
1113 count++;
1114 *value = count;
1115 }
1116 return true;
1117 case PIPE_RESOURCE_PARAM_STRIDE:
1118 *value = wants_aux ? res->aux.surf.row_pitch_B : res->surf.row_pitch_B;
1119 return true;
1120 case PIPE_RESOURCE_PARAM_OFFSET:
1121 *value = wants_aux ? res->aux.offset : 0;
1122 return true;
1123 case PIPE_RESOURCE_PARAM_MODIFIER:
1124 *value = res->mod_info ? res->mod_info->modifier :
1125 tiling_to_modifier(res->bo->tiling_mode);
1126 return true;
1127 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_SHARED:
1128 result = iris_bo_flink(bo, &handle) == 0;
1129 if (result)
1130 *value = handle;
1131 return result;
1132 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_KMS:
1133 *value = iris_bo_export_gem_handle(bo);
1134 return true;
1135 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_FD:
1136 result = iris_bo_export_dmabuf(bo, (int *) &handle) == 0;
1137 if (result)
1138 *value = handle;
1139 return result;
1140 default:
1141 return false;
1142 }
1143 }
1144
1145 static bool
1146 iris_resource_get_handle(struct pipe_screen *pscreen,
1147 struct pipe_context *ctx,
1148 struct pipe_resource *resource,
1149 struct winsys_handle *whandle,
1150 unsigned usage)
1151 {
1152 struct iris_resource *res = (struct iris_resource *)resource;
1153 bool mod_with_aux =
1154 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1155
1156 iris_resource_disable_aux_on_first_query(resource, usage);
1157
1158 struct iris_bo *bo;
1159 if (mod_with_aux && whandle->plane > 0) {
1160 assert(res->aux.bo);
1161 bo = res->aux.bo;
1162 whandle->stride = res->aux.surf.row_pitch_B;
1163 whandle->offset = res->aux.offset;
1164 } else {
1165 /* If this is a buffer, stride should be 0 - no need to special case */
1166 whandle->stride = res->surf.row_pitch_B;
1167 bo = res->bo;
1168 }
1169 whandle->modifier =
1170 res->mod_info ? res->mod_info->modifier
1171 : tiling_to_modifier(res->bo->tiling_mode);
1172
1173 #ifndef NDEBUG
1174 enum isl_aux_usage allowed_usage =
1175 res->mod_info ? res->mod_info->aux_usage : ISL_AUX_USAGE_NONE;
1176
1177 if (res->aux.usage != allowed_usage) {
1178 enum isl_aux_state aux_state = iris_resource_get_aux_state(res, 0, 0);
1179 assert(aux_state == ISL_AUX_STATE_RESOLVED ||
1180 aux_state == ISL_AUX_STATE_PASS_THROUGH);
1181 }
1182 #endif
1183
1184 switch (whandle->type) {
1185 case WINSYS_HANDLE_TYPE_SHARED:
1186 return iris_bo_flink(bo, &whandle->handle) == 0;
1187 case WINSYS_HANDLE_TYPE_KMS:
1188 whandle->handle = iris_bo_export_gem_handle(bo);
1189 return true;
1190 case WINSYS_HANDLE_TYPE_FD:
1191 return iris_bo_export_dmabuf(bo, (int *) &whandle->handle) == 0;
1192 }
1193
1194 return false;
1195 }
1196
1197 static bool
1198 resource_is_busy(struct iris_context *ice,
1199 struct iris_resource *res)
1200 {
1201 bool busy = iris_bo_busy(res->bo);
1202
1203 for (int i = 0; i < IRIS_BATCH_COUNT; i++)
1204 busy |= iris_batch_references(&ice->batches[i], res->bo);
1205
1206 return busy;
1207 }
1208
1209 static void
1210 iris_invalidate_resource(struct pipe_context *ctx,
1211 struct pipe_resource *resource)
1212 {
1213 struct iris_screen *screen = (void *) ctx->screen;
1214 struct iris_context *ice = (void *) ctx;
1215 struct iris_resource *res = (void *) resource;
1216
1217 if (resource->target != PIPE_BUFFER)
1218 return;
1219
1220 if (!resource_is_busy(ice, res)) {
1221 /* The resource is idle, so just mark that it contains no data and
1222 * keep using the same underlying buffer object.
1223 */
1224 util_range_set_empty(&res->valid_buffer_range);
1225 return;
1226 }
1227
1228 /* Otherwise, try and replace the backing storage with a new BO. */
1229
1230 /* We can't reallocate memory we didn't allocate in the first place. */
1231 if (res->bo->userptr)
1232 return;
1233
1234 // XXX: We should support this.
1235 if (res->bind_history & PIPE_BIND_STREAM_OUTPUT)
1236 return;
1237
1238 struct iris_bo *old_bo = res->bo;
1239 struct iris_bo *new_bo =
1240 iris_bo_alloc(screen->bufmgr, res->bo->name, resource->width0,
1241 iris_memzone_for_address(old_bo->gtt_offset));
1242 if (!new_bo)
1243 return;
1244
1245 /* Swap out the backing storage */
1246 res->bo = new_bo;
1247
1248 /* Rebind the buffer, replacing any state referring to the old BO's
1249 * address, and marking state dirty so it's reemitted.
1250 */
1251 ice->vtbl.rebind_buffer(ice, res, old_bo->gtt_offset);
1252
1253 util_range_set_empty(&res->valid_buffer_range);
1254
1255 iris_bo_unreference(old_bo);
1256 }
1257
1258 static void
1259 iris_flush_staging_region(struct pipe_transfer *xfer,
1260 const struct pipe_box *flush_box)
1261 {
1262 if (!(xfer->usage & PIPE_TRANSFER_WRITE))
1263 return;
1264
1265 struct iris_transfer *map = (void *) xfer;
1266
1267 struct pipe_box src_box = *flush_box;
1268
1269 /* Account for extra alignment padding in staging buffer */
1270 if (xfer->resource->target == PIPE_BUFFER)
1271 src_box.x += xfer->box.x % IRIS_MAP_BUFFER_ALIGNMENT;
1272
1273 struct pipe_box dst_box = (struct pipe_box) {
1274 .x = xfer->box.x + flush_box->x,
1275 .y = xfer->box.y + flush_box->y,
1276 .z = xfer->box.z + flush_box->z,
1277 .width = flush_box->width,
1278 .height = flush_box->height,
1279 .depth = flush_box->depth,
1280 };
1281
1282 iris_copy_region(map->blorp, map->batch, xfer->resource, xfer->level,
1283 dst_box.x, dst_box.y, dst_box.z, map->staging, 0,
1284 &src_box);
1285 }
1286
1287 static void
1288 iris_unmap_copy_region(struct iris_transfer *map)
1289 {
1290 iris_resource_destroy(map->staging->screen, map->staging);
1291
1292 map->ptr = NULL;
1293 }
1294
1295 static void
1296 iris_map_copy_region(struct iris_transfer *map)
1297 {
1298 struct pipe_screen *pscreen = &map->batch->screen->base;
1299 struct pipe_transfer *xfer = &map->base;
1300 struct pipe_box *box = &xfer->box;
1301 struct iris_resource *res = (void *) xfer->resource;
1302
1303 unsigned extra = xfer->resource->target == PIPE_BUFFER ?
1304 box->x % IRIS_MAP_BUFFER_ALIGNMENT : 0;
1305
1306 struct pipe_resource templ = (struct pipe_resource) {
1307 .usage = PIPE_USAGE_STAGING,
1308 .width0 = box->width + extra,
1309 .height0 = box->height,
1310 .depth0 = 1,
1311 .nr_samples = xfer->resource->nr_samples,
1312 .nr_storage_samples = xfer->resource->nr_storage_samples,
1313 .array_size = box->depth,
1314 .format = res->internal_format,
1315 };
1316
1317 if (xfer->resource->target == PIPE_BUFFER)
1318 templ.target = PIPE_BUFFER;
1319 else if (templ.array_size > 1)
1320 templ.target = PIPE_TEXTURE_2D_ARRAY;
1321 else
1322 templ.target = PIPE_TEXTURE_2D;
1323
1324 map->staging = iris_resource_create(pscreen, &templ);
1325 assert(map->staging);
1326
1327 if (templ.target != PIPE_BUFFER) {
1328 struct isl_surf *surf = &((struct iris_resource *) map->staging)->surf;
1329 xfer->stride = isl_surf_get_row_pitch_B(surf);
1330 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1331 }
1332
1333 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1334 iris_copy_region(map->blorp, map->batch, map->staging, 0, extra, 0, 0,
1335 xfer->resource, xfer->level, box);
1336 /* Ensure writes to the staging BO land before we map it below. */
1337 iris_emit_pipe_control_flush(map->batch,
1338 "transfer read: flush before mapping",
1339 PIPE_CONTROL_RENDER_TARGET_FLUSH |
1340 PIPE_CONTROL_CS_STALL);
1341 }
1342
1343 struct iris_bo *staging_bo = iris_resource_bo(map->staging);
1344
1345 if (iris_batch_references(map->batch, staging_bo))
1346 iris_batch_flush(map->batch);
1347
1348 map->ptr =
1349 iris_bo_map(map->dbg, staging_bo, xfer->usage & MAP_FLAGS) + extra;
1350
1351 map->unmap = iris_unmap_copy_region;
1352 }
1353
1354 static void
1355 get_image_offset_el(const struct isl_surf *surf, unsigned level, unsigned z,
1356 unsigned *out_x0_el, unsigned *out_y0_el)
1357 {
1358 if (surf->dim == ISL_SURF_DIM_3D) {
1359 isl_surf_get_image_offset_el(surf, level, 0, z, out_x0_el, out_y0_el);
1360 } else {
1361 isl_surf_get_image_offset_el(surf, level, z, 0, out_x0_el, out_y0_el);
1362 }
1363 }
1364
1365 /**
1366 * This function computes the tile_w (in bytes) and tile_h (in rows) of
1367 * different tiling patterns.
1368 */
1369 static void
1370 iris_resource_get_tile_dims(enum isl_tiling tiling, uint32_t cpp,
1371 uint32_t *tile_w, uint32_t *tile_h)
1372 {
1373 switch (tiling) {
1374 case ISL_TILING_X:
1375 *tile_w = 512;
1376 *tile_h = 8;
1377 break;
1378 case ISL_TILING_Y0:
1379 *tile_w = 128;
1380 *tile_h = 32;
1381 break;
1382 case ISL_TILING_LINEAR:
1383 *tile_w = cpp;
1384 *tile_h = 1;
1385 break;
1386 default:
1387 unreachable("not reached");
1388 }
1389
1390 }
1391
1392 /**
1393 * This function computes masks that may be used to select the bits of the X
1394 * and Y coordinates that indicate the offset within a tile. If the BO is
1395 * untiled, the masks are set to 0.
1396 */
1397 static void
1398 iris_resource_get_tile_masks(enum isl_tiling tiling, uint32_t cpp,
1399 uint32_t *mask_x, uint32_t *mask_y)
1400 {
1401 uint32_t tile_w_bytes, tile_h;
1402
1403 iris_resource_get_tile_dims(tiling, cpp, &tile_w_bytes, &tile_h);
1404
1405 *mask_x = tile_w_bytes / cpp - 1;
1406 *mask_y = tile_h - 1;
1407 }
1408
1409 /**
1410 * Compute the offset (in bytes) from the start of the BO to the given x
1411 * and y coordinate. For tiled BOs, caller must ensure that x and y are
1412 * multiples of the tile size.
1413 */
1414 static uint32_t
1415 iris_resource_get_aligned_offset(const struct iris_resource *res,
1416 uint32_t x, uint32_t y)
1417 {
1418 const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
1419 unsigned cpp = fmtl->bpb / 8;
1420 uint32_t pitch = res->surf.row_pitch_B;
1421
1422 switch (res->surf.tiling) {
1423 default:
1424 unreachable("not reached");
1425 case ISL_TILING_LINEAR:
1426 return y * pitch + x * cpp;
1427 case ISL_TILING_X:
1428 assert((x % (512 / cpp)) == 0);
1429 assert((y % 8) == 0);
1430 return y * pitch + x / (512 / cpp) * 4096;
1431 case ISL_TILING_Y0:
1432 assert((x % (128 / cpp)) == 0);
1433 assert((y % 32) == 0);
1434 return y * pitch + x / (128 / cpp) * 4096;
1435 }
1436 }
1437
1438 /**
1439 * Rendering with tiled buffers requires that the base address of the buffer
1440 * be aligned to a page boundary. For renderbuffers, and sometimes with
1441 * textures, we may want the surface to point at a texture image level that
1442 * isn't at a page boundary.
1443 *
1444 * This function returns an appropriately-aligned base offset
1445 * according to the tiling restrictions, plus any required x/y offset
1446 * from there.
1447 */
1448 uint32_t
1449 iris_resource_get_tile_offsets(const struct iris_resource *res,
1450 uint32_t level, uint32_t z,
1451 uint32_t *tile_x, uint32_t *tile_y)
1452 {
1453 uint32_t x, y;
1454 uint32_t mask_x, mask_y;
1455
1456 const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
1457 const unsigned cpp = fmtl->bpb / 8;
1458
1459 iris_resource_get_tile_masks(res->surf.tiling, cpp, &mask_x, &mask_y);
1460 get_image_offset_el(&res->surf, level, z, &x, &y);
1461
1462 *tile_x = x & mask_x;
1463 *tile_y = y & mask_y;
1464
1465 return iris_resource_get_aligned_offset(res, x & ~mask_x, y & ~mask_y);
1466 }
1467
1468 /**
1469 * Get pointer offset into stencil buffer.
1470 *
1471 * The stencil buffer is W tiled. Since the GTT is incapable of W fencing, we
1472 * must decode the tile's layout in software.
1473 *
1474 * See
1475 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.2.1 W-Major Tile
1476 * Format.
1477 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.3 Tiling Algorithm
1478 *
1479 * Even though the returned offset is always positive, the return type is
1480 * signed due to
1481 * commit e8b1c6d6f55f5be3bef25084fdd8b6127517e137
1482 * mesa: Fix return type of _mesa_get_format_bytes() (#37351)
1483 */
1484 static intptr_t
1485 s8_offset(uint32_t stride, uint32_t x, uint32_t y)
1486 {
1487 uint32_t tile_size = 4096;
1488 uint32_t tile_width = 64;
1489 uint32_t tile_height = 64;
1490 uint32_t row_size = 64 * stride / 2; /* Two rows are interleaved. */
1491
1492 uint32_t tile_x = x / tile_width;
1493 uint32_t tile_y = y / tile_height;
1494
1495 /* The byte's address relative to the tile's base addres. */
1496 uint32_t byte_x = x % tile_width;
1497 uint32_t byte_y = y % tile_height;
1498
1499 uintptr_t u = tile_y * row_size
1500 + tile_x * tile_size
1501 + 512 * (byte_x / 8)
1502 + 64 * (byte_y / 8)
1503 + 32 * ((byte_y / 4) % 2)
1504 + 16 * ((byte_x / 4) % 2)
1505 + 8 * ((byte_y / 2) % 2)
1506 + 4 * ((byte_x / 2) % 2)
1507 + 2 * (byte_y % 2)
1508 + 1 * (byte_x % 2);
1509
1510 return u;
1511 }
1512
1513 static void
1514 iris_unmap_s8(struct iris_transfer *map)
1515 {
1516 struct pipe_transfer *xfer = &map->base;
1517 const struct pipe_box *box = &xfer->box;
1518 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1519 struct isl_surf *surf = &res->surf;
1520
1521 if (xfer->usage & PIPE_TRANSFER_WRITE) {
1522 uint8_t *untiled_s8_map = map->ptr;
1523 uint8_t *tiled_s8_map =
1524 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1525
1526 for (int s = 0; s < box->depth; s++) {
1527 unsigned x0_el, y0_el;
1528 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1529
1530 for (uint32_t y = 0; y < box->height; y++) {
1531 for (uint32_t x = 0; x < box->width; x++) {
1532 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1533 x0_el + box->x + x,
1534 y0_el + box->y + y);
1535 tiled_s8_map[offset] =
1536 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x];
1537 }
1538 }
1539 }
1540 }
1541
1542 free(map->buffer);
1543 }
1544
1545 static void
1546 iris_map_s8(struct iris_transfer *map)
1547 {
1548 struct pipe_transfer *xfer = &map->base;
1549 const struct pipe_box *box = &xfer->box;
1550 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1551 struct isl_surf *surf = &res->surf;
1552
1553 xfer->stride = surf->row_pitch_B;
1554 xfer->layer_stride = xfer->stride * box->height;
1555
1556 /* The tiling and detiling functions require that the linear buffer has
1557 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1558 * over-allocate the linear buffer to get the proper alignment.
1559 */
1560 map->buffer = map->ptr = malloc(xfer->layer_stride * box->depth);
1561 assert(map->buffer);
1562
1563 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
1564 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
1565 * invalidate is set, since we'll be writing the whole rectangle from our
1566 * temporary buffer back out.
1567 */
1568 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1569 uint8_t *untiled_s8_map = map->ptr;
1570 uint8_t *tiled_s8_map =
1571 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1572
1573 for (int s = 0; s < box->depth; s++) {
1574 unsigned x0_el, y0_el;
1575 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1576
1577 for (uint32_t y = 0; y < box->height; y++) {
1578 for (uint32_t x = 0; x < box->width; x++) {
1579 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1580 x0_el + box->x + x,
1581 y0_el + box->y + y);
1582 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x] =
1583 tiled_s8_map[offset];
1584 }
1585 }
1586 }
1587 }
1588
1589 map->unmap = iris_unmap_s8;
1590 }
1591
1592 /* Compute extent parameters for use with tiled_memcpy functions.
1593 * xs are in units of bytes and ys are in units of strides.
1594 */
1595 static inline void
1596 tile_extents(const struct isl_surf *surf,
1597 const struct pipe_box *box,
1598 unsigned level, int z,
1599 unsigned *x1_B, unsigned *x2_B,
1600 unsigned *y1_el, unsigned *y2_el)
1601 {
1602 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1603 const unsigned cpp = fmtl->bpb / 8;
1604
1605 assert(box->x % fmtl->bw == 0);
1606 assert(box->y % fmtl->bh == 0);
1607
1608 unsigned x0_el, y0_el;
1609 get_image_offset_el(surf, level, box->z + z, &x0_el, &y0_el);
1610
1611 *x1_B = (box->x / fmtl->bw + x0_el) * cpp;
1612 *y1_el = box->y / fmtl->bh + y0_el;
1613 *x2_B = (DIV_ROUND_UP(box->x + box->width, fmtl->bw) + x0_el) * cpp;
1614 *y2_el = DIV_ROUND_UP(box->y + box->height, fmtl->bh) + y0_el;
1615 }
1616
1617 static void
1618 iris_unmap_tiled_memcpy(struct iris_transfer *map)
1619 {
1620 struct pipe_transfer *xfer = &map->base;
1621 const struct pipe_box *box = &xfer->box;
1622 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1623 struct isl_surf *surf = &res->surf;
1624
1625 const bool has_swizzling = false;
1626
1627 if (xfer->usage & PIPE_TRANSFER_WRITE) {
1628 char *dst =
1629 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1630
1631 for (int s = 0; s < box->depth; s++) {
1632 unsigned x1, x2, y1, y2;
1633 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1634
1635 void *ptr = map->ptr + s * xfer->layer_stride;
1636
1637 isl_memcpy_linear_to_tiled(x1, x2, y1, y2, dst, ptr,
1638 surf->row_pitch_B, xfer->stride,
1639 has_swizzling, surf->tiling, ISL_MEMCPY);
1640 }
1641 }
1642 os_free_aligned(map->buffer);
1643 map->buffer = map->ptr = NULL;
1644 }
1645
1646 static void
1647 iris_map_tiled_memcpy(struct iris_transfer *map)
1648 {
1649 struct pipe_transfer *xfer = &map->base;
1650 const struct pipe_box *box = &xfer->box;
1651 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1652 struct isl_surf *surf = &res->surf;
1653
1654 xfer->stride = ALIGN(surf->row_pitch_B, 16);
1655 xfer->layer_stride = xfer->stride * box->height;
1656
1657 unsigned x1, x2, y1, y2;
1658 tile_extents(surf, box, xfer->level, 0, &x1, &x2, &y1, &y2);
1659
1660 /* The tiling and detiling functions require that the linear buffer has
1661 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1662 * over-allocate the linear buffer to get the proper alignment.
1663 */
1664 map->buffer =
1665 os_malloc_aligned(xfer->layer_stride * box->depth, 16);
1666 assert(map->buffer);
1667 map->ptr = (char *)map->buffer + (x1 & 0xf);
1668
1669 const bool has_swizzling = false;
1670
1671 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1672 char *src =
1673 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1674
1675 for (int s = 0; s < box->depth; s++) {
1676 unsigned x1, x2, y1, y2;
1677 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1678
1679 /* Use 's' rather than 'box->z' to rebase the first slice to 0. */
1680 void *ptr = map->ptr + s * xfer->layer_stride;
1681
1682 isl_memcpy_tiled_to_linear(x1, x2, y1, y2, ptr, src, xfer->stride,
1683 surf->row_pitch_B, has_swizzling,
1684 surf->tiling, ISL_MEMCPY_STREAMING_LOAD);
1685 }
1686 }
1687
1688 map->unmap = iris_unmap_tiled_memcpy;
1689 }
1690
1691 static void
1692 iris_map_direct(struct iris_transfer *map)
1693 {
1694 struct pipe_transfer *xfer = &map->base;
1695 struct pipe_box *box = &xfer->box;
1696 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1697
1698 void *ptr = iris_bo_map(map->dbg, res->bo, xfer->usage & MAP_FLAGS);
1699
1700 if (res->base.target == PIPE_BUFFER) {
1701 xfer->stride = 0;
1702 xfer->layer_stride = 0;
1703
1704 map->ptr = ptr + box->x;
1705 } else {
1706 struct isl_surf *surf = &res->surf;
1707 const struct isl_format_layout *fmtl =
1708 isl_format_get_layout(surf->format);
1709 const unsigned cpp = fmtl->bpb / 8;
1710 unsigned x0_el, y0_el;
1711
1712 get_image_offset_el(surf, xfer->level, box->z, &x0_el, &y0_el);
1713
1714 xfer->stride = isl_surf_get_row_pitch_B(surf);
1715 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1716
1717 map->ptr = ptr + (y0_el + box->y) * xfer->stride + (x0_el + box->x) * cpp;
1718 }
1719 }
1720
1721 static bool
1722 can_promote_to_async(const struct iris_resource *res,
1723 const struct pipe_box *box,
1724 enum pipe_transfer_usage usage)
1725 {
1726 /* If we're writing to a section of the buffer that hasn't even been
1727 * initialized with useful data, then we can safely promote this write
1728 * to be unsynchronized. This helps the common pattern of appending data.
1729 */
1730 return res->base.target == PIPE_BUFFER && (usage & PIPE_TRANSFER_WRITE) &&
1731 !(usage & TC_TRANSFER_MAP_NO_INFER_UNSYNCHRONIZED) &&
1732 !util_ranges_intersect(&res->valid_buffer_range, box->x,
1733 box->x + box->width);
1734 }
1735
1736 static void *
1737 iris_transfer_map(struct pipe_context *ctx,
1738 struct pipe_resource *resource,
1739 unsigned level,
1740 enum pipe_transfer_usage usage,
1741 const struct pipe_box *box,
1742 struct pipe_transfer **ptransfer)
1743 {
1744 struct iris_context *ice = (struct iris_context *)ctx;
1745 struct iris_resource *res = (struct iris_resource *)resource;
1746 struct isl_surf *surf = &res->surf;
1747
1748 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
1749 /* Replace the backing storage with a fresh buffer for non-async maps */
1750 if (!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
1751 TC_TRANSFER_MAP_NO_INVALIDATE)))
1752 iris_invalidate_resource(ctx, resource);
1753
1754 /* If we can discard the whole resource, we can discard the range. */
1755 usage |= PIPE_TRANSFER_DISCARD_RANGE;
1756 }
1757
1758 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
1759 can_promote_to_async(res, box, usage)) {
1760 usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
1761 }
1762
1763 bool need_resolve = false;
1764 bool need_color_resolve = false;
1765
1766 if (resource->target != PIPE_BUFFER) {
1767 bool need_hiz_resolve = iris_resource_level_has_hiz(res, level);
1768
1769 need_color_resolve =
1770 (res->aux.usage == ISL_AUX_USAGE_CCS_D ||
1771 res->aux.usage == ISL_AUX_USAGE_CCS_E) &&
1772 iris_has_color_unresolved(res, level, 1, box->z, box->depth);
1773
1774 need_resolve = need_color_resolve || need_hiz_resolve;
1775 }
1776
1777 bool map_would_stall = false;
1778
1779 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
1780 map_would_stall = need_resolve || resource_is_busy(ice, res);
1781
1782 if (map_would_stall && (usage & PIPE_TRANSFER_DONTBLOCK) &&
1783 (usage & PIPE_TRANSFER_MAP_DIRECTLY))
1784 return NULL;
1785 }
1786
1787 if (surf->tiling != ISL_TILING_LINEAR &&
1788 (usage & PIPE_TRANSFER_MAP_DIRECTLY))
1789 return NULL;
1790
1791 struct iris_transfer *map = slab_alloc(&ice->transfer_pool);
1792 struct pipe_transfer *xfer = &map->base;
1793
1794 if (!map)
1795 return NULL;
1796
1797 memset(map, 0, sizeof(*map));
1798 map->dbg = &ice->dbg;
1799
1800 pipe_resource_reference(&xfer->resource, resource);
1801 xfer->level = level;
1802 xfer->usage = usage;
1803 xfer->box = *box;
1804 *ptransfer = xfer;
1805
1806 map->dest_had_defined_contents =
1807 util_ranges_intersect(&res->valid_buffer_range, box->x,
1808 box->x + box->width);
1809
1810 if (usage & PIPE_TRANSFER_WRITE)
1811 util_range_add(&res->base, &res->valid_buffer_range, box->x, box->x + box->width);
1812
1813 /* Avoid using GPU copies for persistent/coherent buffers, as the idea
1814 * there is to access them simultaneously on the CPU & GPU. This also
1815 * avoids trying to use GPU copies for our u_upload_mgr buffers which
1816 * contain state we're constructing for a GPU draw call, which would
1817 * kill us with infinite stack recursion.
1818 */
1819 bool no_gpu = usage & (PIPE_TRANSFER_PERSISTENT |
1820 PIPE_TRANSFER_COHERENT |
1821 PIPE_TRANSFER_MAP_DIRECTLY);
1822
1823 /* GPU copies are not useful for buffer reads. Instead of stalling to
1824 * read from the original buffer, we'd simply copy it to a temporary...
1825 * then stall (a bit longer) to read from that buffer.
1826 *
1827 * Images are less clear-cut. Color resolves are destructive, removing
1828 * the underlying compression, so we'd rather blit the data to a linear
1829 * temporary and map that, to avoid the resolve. (It might be better to
1830 * a tiled temporary and use the tiled_memcpy paths...)
1831 */
1832 if (!(usage & PIPE_TRANSFER_DISCARD_RANGE) && !need_color_resolve)
1833 no_gpu = true;
1834
1835 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1836 if (fmtl->txc == ISL_TXC_ASTC)
1837 no_gpu = true;
1838
1839 if ((map_would_stall || res->aux.usage == ISL_AUX_USAGE_CCS_E) && !no_gpu) {
1840 /* If we need a synchronous mapping and the resource is busy, or needs
1841 * resolving, we copy to/from a linear temporary buffer using the GPU.
1842 */
1843 map->batch = &ice->batches[IRIS_BATCH_RENDER];
1844 map->blorp = &ice->blorp;
1845 iris_map_copy_region(map);
1846 } else {
1847 /* Otherwise we're free to map on the CPU. */
1848
1849 if (need_resolve) {
1850 iris_resource_access_raw(ice, &ice->batches[IRIS_BATCH_RENDER], res,
1851 level, box->z, box->depth,
1852 usage & PIPE_TRANSFER_WRITE);
1853 }
1854
1855 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
1856 for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
1857 if (iris_batch_references(&ice->batches[i], res->bo))
1858 iris_batch_flush(&ice->batches[i]);
1859 }
1860 }
1861
1862 if (surf->tiling == ISL_TILING_W) {
1863 /* TODO: Teach iris_map_tiled_memcpy about W-tiling... */
1864 iris_map_s8(map);
1865 } else if (surf->tiling != ISL_TILING_LINEAR) {
1866 iris_map_tiled_memcpy(map);
1867 } else {
1868 iris_map_direct(map);
1869 }
1870 }
1871
1872 return map->ptr;
1873 }
1874
1875 static void
1876 iris_transfer_flush_region(struct pipe_context *ctx,
1877 struct pipe_transfer *xfer,
1878 const struct pipe_box *box)
1879 {
1880 struct iris_context *ice = (struct iris_context *)ctx;
1881 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1882 struct iris_transfer *map = (void *) xfer;
1883
1884 if (map->staging)
1885 iris_flush_staging_region(xfer, box);
1886
1887 uint32_t history_flush = 0;
1888
1889 if (res->base.target == PIPE_BUFFER) {
1890 if (map->staging)
1891 history_flush |= PIPE_CONTROL_RENDER_TARGET_FLUSH;
1892
1893 if (map->dest_had_defined_contents)
1894 history_flush |= iris_flush_bits_for_history(res);
1895
1896 util_range_add(&res->base, &res->valid_buffer_range, box->x, box->x + box->width);
1897 }
1898
1899 if (history_flush & ~PIPE_CONTROL_CS_STALL) {
1900 for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
1901 struct iris_batch *batch = &ice->batches[i];
1902 if (batch->contains_draw || batch->cache.render->entries) {
1903 iris_batch_maybe_flush(batch, 24);
1904 iris_emit_pipe_control_flush(batch,
1905 "cache history: transfer flush",
1906 history_flush);
1907 }
1908 }
1909 }
1910
1911 /* Make sure we flag constants dirty even if there's no need to emit
1912 * any PIPE_CONTROLs to a batch.
1913 */
1914 iris_dirty_for_history(ice, res);
1915 }
1916
1917 static void
1918 iris_transfer_unmap(struct pipe_context *ctx, struct pipe_transfer *xfer)
1919 {
1920 struct iris_context *ice = (struct iris_context *)ctx;
1921 struct iris_transfer *map = (void *) xfer;
1922
1923 if (!(xfer->usage & (PIPE_TRANSFER_FLUSH_EXPLICIT |
1924 PIPE_TRANSFER_COHERENT))) {
1925 struct pipe_box flush_box = {
1926 .x = 0, .y = 0, .z = 0,
1927 .width = xfer->box.width,
1928 .height = xfer->box.height,
1929 .depth = xfer->box.depth,
1930 };
1931 iris_transfer_flush_region(ctx, xfer, &flush_box);
1932 }
1933
1934 if (map->unmap)
1935 map->unmap(map);
1936
1937 pipe_resource_reference(&xfer->resource, NULL);
1938 slab_free(&ice->transfer_pool, map);
1939 }
1940
1941 /**
1942 * Mark state dirty that needs to be re-emitted when a resource is written.
1943 */
1944 void
1945 iris_dirty_for_history(struct iris_context *ice,
1946 struct iris_resource *res)
1947 {
1948 uint64_t dirty = 0ull;
1949
1950 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1951 dirty |= ((uint64_t)res->bind_stages) << IRIS_SHIFT_FOR_DIRTY_CONSTANTS;
1952 }
1953
1954 ice->state.dirty |= dirty;
1955 }
1956
1957 /**
1958 * Produce a set of PIPE_CONTROL bits which ensure data written to a
1959 * resource becomes visible, and any stale read cache data is invalidated.
1960 */
1961 uint32_t
1962 iris_flush_bits_for_history(struct iris_resource *res)
1963 {
1964 uint32_t flush = PIPE_CONTROL_CS_STALL;
1965
1966 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1967 flush |= PIPE_CONTROL_CONST_CACHE_INVALIDATE |
1968 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1969 }
1970
1971 if (res->bind_history & PIPE_BIND_SAMPLER_VIEW)
1972 flush |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1973
1974 if (res->bind_history & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
1975 flush |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1976
1977 if (res->bind_history & (PIPE_BIND_SHADER_BUFFER | PIPE_BIND_SHADER_IMAGE))
1978 flush |= PIPE_CONTROL_DATA_CACHE_FLUSH;
1979
1980 return flush;
1981 }
1982
1983 void
1984 iris_flush_and_dirty_for_history(struct iris_context *ice,
1985 struct iris_batch *batch,
1986 struct iris_resource *res,
1987 uint32_t extra_flags,
1988 const char *reason)
1989 {
1990 if (res->base.target != PIPE_BUFFER)
1991 return;
1992
1993 uint32_t flush = iris_flush_bits_for_history(res) | extra_flags;
1994
1995 iris_emit_pipe_control_flush(batch, reason, flush);
1996
1997 iris_dirty_for_history(ice, res);
1998 }
1999
2000 bool
2001 iris_resource_set_clear_color(struct iris_context *ice,
2002 struct iris_resource *res,
2003 union isl_color_value color)
2004 {
2005 if (memcmp(&res->aux.clear_color, &color, sizeof(color)) != 0) {
2006 res->aux.clear_color = color;
2007 return true;
2008 }
2009
2010 return false;
2011 }
2012
2013 union isl_color_value
2014 iris_resource_get_clear_color(const struct iris_resource *res,
2015 struct iris_bo **clear_color_bo,
2016 uint64_t *clear_color_offset)
2017 {
2018 assert(res->aux.bo);
2019
2020 if (clear_color_bo)
2021 *clear_color_bo = res->aux.clear_color_bo;
2022 if (clear_color_offset)
2023 *clear_color_offset = res->aux.clear_color_offset;
2024 return res->aux.clear_color;
2025 }
2026
2027 static enum pipe_format
2028 iris_resource_get_internal_format(struct pipe_resource *p_res)
2029 {
2030 struct iris_resource *res = (void *) p_res;
2031 return res->internal_format;
2032 }
2033
2034 static const struct u_transfer_vtbl transfer_vtbl = {
2035 .resource_create = iris_resource_create,
2036 .resource_destroy = iris_resource_destroy,
2037 .transfer_map = iris_transfer_map,
2038 .transfer_unmap = iris_transfer_unmap,
2039 .transfer_flush_region = iris_transfer_flush_region,
2040 .get_internal_format = iris_resource_get_internal_format,
2041 .set_stencil = iris_resource_set_separate_stencil,
2042 .get_stencil = iris_resource_get_separate_stencil,
2043 };
2044
2045 void
2046 iris_init_screen_resource_functions(struct pipe_screen *pscreen)
2047 {
2048 pscreen->query_dmabuf_modifiers = iris_query_dmabuf_modifiers;
2049 pscreen->resource_create_with_modifiers =
2050 iris_resource_create_with_modifiers;
2051 pscreen->resource_create = u_transfer_helper_resource_create;
2052 pscreen->resource_from_user_memory = iris_resource_from_user_memory;
2053 pscreen->resource_from_handle = iris_resource_from_handle;
2054 pscreen->resource_get_handle = iris_resource_get_handle;
2055 pscreen->resource_get_param = iris_resource_get_param;
2056 pscreen->resource_destroy = u_transfer_helper_resource_destroy;
2057 pscreen->transfer_helper =
2058 u_transfer_helper_create(&transfer_vtbl, true, true, false, true);
2059 }
2060
2061 void
2062 iris_init_resource_functions(struct pipe_context *ctx)
2063 {
2064 ctx->flush_resource = iris_flush_resource;
2065 ctx->invalidate_resource = iris_invalidate_resource;
2066 ctx->transfer_map = u_transfer_helper_transfer_map;
2067 ctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
2068 ctx->transfer_unmap = u_transfer_helper_transfer_unmap;
2069 ctx->buffer_subdata = u_default_buffer_subdata;
2070 ctx->texture_subdata = u_default_texture_subdata;
2071 }