iris: Enable HIZ_CCS sampling
[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 |=
482 1 << (has_ccs ? ISL_AUX_USAGE_HIZ_CCS : ISL_AUX_USAGE_HIZ);
483 } else if (has_ccs) {
484 if (want_ccs_e_for_format(devinfo, res->surf.format))
485 res->aux.possible_usages |= 1 << ISL_AUX_USAGE_CCS_E;
486
487 if (isl_format_supports_ccs_d(devinfo, res->surf.format))
488 res->aux.possible_usages |= 1 << ISL_AUX_USAGE_CCS_D;
489 }
490
491 res->aux.usage = util_last_bit(res->aux.possible_usages) - 1;
492
493 res->aux.sampler_usages = res->aux.possible_usages;
494
495 /* We don't always support sampling with hiz. But when we do, it must be
496 * single sampled.
497 */
498 if (!devinfo->has_sample_with_hiz || res->surf.samples > 1)
499 res->aux.sampler_usages &= ~(1 << ISL_AUX_USAGE_HIZ);
500
501 /* We don't always support sampling with HIZ_CCS. But when we do, treat it
502 * as CCS_E.*/
503 res->aux.sampler_usages &= ~(1 << ISL_AUX_USAGE_HIZ_CCS);
504 if (isl_surf_supports_hiz_ccs_wt(devinfo, &res->surf, res->aux.usage))
505 res->aux.sampler_usages |= 1 << ISL_AUX_USAGE_CCS_E;
506
507 enum isl_aux_state initial_state;
508 *aux_size_B = 0;
509 *alloc_flags = 0;
510 assert(!res->aux.bo);
511
512 switch (res->aux.usage) {
513 case ISL_AUX_USAGE_NONE:
514 /* Having no aux buffer is only okay if there's no modifier with aux. */
515 return !res->mod_info || res->mod_info->aux_usage == ISL_AUX_USAGE_NONE;
516 case ISL_AUX_USAGE_HIZ:
517 case ISL_AUX_USAGE_HIZ_CCS:
518 initial_state = ISL_AUX_STATE_AUX_INVALID;
519 break;
520 case ISL_AUX_USAGE_MCS:
521 /* The Ivybridge PRM, Vol 2 Part 1 p326 says:
522 *
523 * "When MCS buffer is enabled and bound to MSRT, it is required
524 * that it is cleared prior to any rendering."
525 *
526 * Since we only use the MCS buffer for rendering, we just clear it
527 * immediately on allocation. The clear value for MCS buffers is all
528 * 1's, so we simply memset it to 0xff.
529 */
530 initial_state = ISL_AUX_STATE_CLEAR;
531 break;
532 case ISL_AUX_USAGE_CCS_D:
533 case ISL_AUX_USAGE_CCS_E:
534 /* When CCS_E is used, we need to ensure that the CCS starts off in
535 * a valid state. From the Sky Lake PRM, "MCS Buffer for Render
536 * Target(s)":
537 *
538 * "If Software wants to enable Color Compression without Fast
539 * clear, Software needs to initialize MCS with zeros."
540 *
541 * A CCS value of 0 indicates that the corresponding block is in the
542 * pass-through state which is what we want.
543 *
544 * For CCS_D, do the same thing. On Gen9+, this avoids having any
545 * undefined bits in the aux buffer.
546 */
547 if (imported)
548 initial_state =
549 isl_drm_modifier_get_default_aux_state(res->mod_info->modifier);
550 else
551 initial_state = ISL_AUX_STATE_PASS_THROUGH;
552 *alloc_flags |= BO_ALLOC_ZEROED;
553 break;
554 }
555
556 if (!res->aux.state) {
557 /* Create the aux_state for the auxiliary buffer. */
558 res->aux.state = create_aux_state_map(res, initial_state);
559 if (!res->aux.state)
560 return false;
561 }
562
563 uint64_t size = res->aux.surf.size_B;
564
565 /* Allocate space in the buffer for storing the CCS. */
566 if (res->aux.extra_aux.surf.size_B > 0) {
567 res->aux.extra_aux.offset =
568 ALIGN(size, res->aux.extra_aux.surf.alignment_B);
569 size = res->aux.extra_aux.offset + res->aux.extra_aux.surf.size_B;
570 }
571
572 /* Allocate space in the buffer for storing the clear color. On modern
573 * platforms (gen > 9), we can read it directly from such buffer.
574 *
575 * On gen <= 9, we are going to store the clear color on the buffer
576 * anyways, and copy it back to the surface state during state emission.
577 */
578 res->aux.clear_color_offset = size;
579 size += iris_get_aux_clear_color_state_size(screen);
580 *aux_size_B = size;
581
582 if (isl_aux_usage_has_hiz(res->aux.usage)) {
583 for (unsigned level = 0; level < res->surf.levels; ++level) {
584 uint32_t width = u_minify(res->surf.phys_level0_sa.width, level);
585 uint32_t height = u_minify(res->surf.phys_level0_sa.height, level);
586
587 /* Disable HiZ for LOD > 0 unless the width/height are 8x4 aligned.
588 * For LOD == 0, we can grow the dimensions to make it work.
589 */
590 if (level == 0 || ((width & 7) == 0 && (height & 3) == 0))
591 res->aux.has_hiz |= 1 << level;
592 }
593 }
594
595 return true;
596 }
597
598 /**
599 * Initialize the aux buffer contents.
600 */
601 static bool
602 iris_resource_init_aux_buf(struct iris_resource *res, uint32_t alloc_flags,
603 unsigned clear_color_state_size)
604 {
605 if (!(alloc_flags & BO_ALLOC_ZEROED)) {
606 void *map = iris_bo_map(NULL, res->aux.bo, MAP_WRITE | MAP_RAW);
607
608 if (!map) {
609 iris_resource_disable_aux(res);
610 return false;
611 }
612
613 if (iris_resource_get_aux_state(res, 0, 0) != ISL_AUX_STATE_AUX_INVALID) {
614 uint8_t memset_value = res->aux.usage == ISL_AUX_USAGE_MCS ? 0xFF : 0;
615 memset((char*)map + res->aux.offset, memset_value,
616 res->aux.surf.size_B);
617 }
618
619 /* Resolved is usually a safe state for CCS_E. */
620 memset((char*)map + res->aux.extra_aux.offset, 0,
621 res->aux.extra_aux.surf.size_B);
622
623 /* Zero the indirect clear color to match ::fast_clear_color. */
624 memset((char *)map + res->aux.clear_color_offset, 0,
625 clear_color_state_size);
626
627 iris_bo_unmap(res->aux.bo);
628 }
629
630 if (res->aux.extra_aux.surf.size_B > 0) {
631 res->aux.extra_aux.bo = res->aux.bo;
632 iris_bo_reference(res->aux.extra_aux.bo);
633 }
634
635 if (clear_color_state_size > 0) {
636 res->aux.clear_color_bo = res->aux.bo;
637 iris_bo_reference(res->aux.clear_color_bo);
638 }
639
640 return true;
641 }
642
643 /**
644 * Allocate the initial aux surface for a resource based on aux.usage
645 */
646 static bool
647 iris_resource_alloc_separate_aux(struct iris_screen *screen,
648 struct iris_resource *res)
649 {
650 uint32_t alloc_flags;
651 uint64_t size;
652 if (!iris_resource_configure_aux(screen, res, false, &size, &alloc_flags))
653 return false;
654
655 if (size == 0)
656 return true;
657
658 /* Allocate the auxiliary buffer. ISL has stricter set of alignment rules
659 * the drm allocator. Therefore, one can pass the ISL dimensions in terms
660 * of bytes instead of trying to recalculate based on different format
661 * block sizes.
662 */
663 res->aux.bo = iris_bo_alloc_tiled(screen->bufmgr, "aux buffer", size, 4096,
664 IRIS_MEMZONE_OTHER,
665 isl_tiling_to_i915_tiling(res->aux.surf.tiling),
666 res->aux.surf.row_pitch_B, alloc_flags);
667 if (!res->aux.bo) {
668 return false;
669 }
670
671 if (!iris_resource_init_aux_buf(res, alloc_flags,
672 iris_get_aux_clear_color_state_size(screen)))
673 return false;
674
675 map_aux_addresses(screen, res);
676
677 return true;
678 }
679
680 void
681 iris_resource_finish_aux_import(struct pipe_screen *pscreen,
682 struct iris_resource *res)
683 {
684 struct iris_screen *screen = (struct iris_screen *)pscreen;
685 assert(iris_resource_unfinished_aux_import(res));
686 assert(!res->mod_info->supports_clear_color);
687
688 struct iris_resource *aux_res = (void *) res->base.next;
689 assert(aux_res->aux.surf.row_pitch_B && aux_res->aux.offset &&
690 aux_res->aux.bo);
691
692 assert(res->bo == aux_res->aux.bo);
693 iris_bo_reference(aux_res->aux.bo);
694 res->aux.bo = aux_res->aux.bo;
695
696 res->aux.offset = aux_res->aux.offset;
697
698 assert(res->bo->size >= (res->aux.offset + res->aux.surf.size_B));
699 assert(res->aux.clear_color_bo == NULL);
700 res->aux.clear_color_offset = 0;
701
702 assert(aux_res->aux.surf.row_pitch_B == res->aux.surf.row_pitch_B);
703
704 unsigned clear_color_state_size =
705 iris_get_aux_clear_color_state_size(screen);
706
707 if (clear_color_state_size > 0) {
708 res->aux.clear_color_bo =
709 iris_bo_alloc(screen->bufmgr, "clear color buffer",
710 clear_color_state_size, IRIS_MEMZONE_OTHER);
711 res->aux.clear_color_offset = 0;
712 }
713
714 iris_resource_destroy(&screen->base, res->base.next);
715 res->base.next = NULL;
716 }
717
718 static struct pipe_resource *
719 iris_resource_create_for_buffer(struct pipe_screen *pscreen,
720 const struct pipe_resource *templ)
721 {
722 struct iris_screen *screen = (struct iris_screen *)pscreen;
723 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
724
725 assert(templ->target == PIPE_BUFFER);
726 assert(templ->height0 <= 1);
727 assert(templ->depth0 <= 1);
728 assert(templ->format == PIPE_FORMAT_NONE ||
729 util_format_get_blocksize(templ->format) == 1);
730
731 res->internal_format = templ->format;
732 res->surf.tiling = ISL_TILING_LINEAR;
733
734 enum iris_memory_zone memzone = IRIS_MEMZONE_OTHER;
735 const char *name = templ->target == PIPE_BUFFER ? "buffer" : "miptree";
736 if (templ->flags & IRIS_RESOURCE_FLAG_SHADER_MEMZONE) {
737 memzone = IRIS_MEMZONE_SHADER;
738 name = "shader kernels";
739 } else if (templ->flags & IRIS_RESOURCE_FLAG_SURFACE_MEMZONE) {
740 memzone = IRIS_MEMZONE_SURFACE;
741 name = "surface state";
742 } else if (templ->flags & IRIS_RESOURCE_FLAG_DYNAMIC_MEMZONE) {
743 memzone = IRIS_MEMZONE_DYNAMIC;
744 name = "dynamic state";
745 }
746
747 res->bo = iris_bo_alloc(screen->bufmgr, name, templ->width0, memzone);
748 if (!res->bo) {
749 iris_resource_destroy(pscreen, &res->base);
750 return NULL;
751 }
752
753 return &res->base;
754 }
755
756 static struct pipe_resource *
757 iris_resource_create_with_modifiers(struct pipe_screen *pscreen,
758 const struct pipe_resource *templ,
759 const uint64_t *modifiers,
760 int modifiers_count)
761 {
762 struct iris_screen *screen = (struct iris_screen *)pscreen;
763 struct gen_device_info *devinfo = &screen->devinfo;
764 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
765
766 if (!res)
767 return NULL;
768
769 const struct util_format_description *format_desc =
770 util_format_description(templ->format);
771 const bool has_depth = util_format_has_depth(format_desc);
772 uint64_t modifier =
773 select_best_modifier(devinfo, templ->format, modifiers, modifiers_count);
774
775 isl_tiling_flags_t tiling_flags = ISL_TILING_ANY_MASK;
776
777 if (modifier != DRM_FORMAT_MOD_INVALID) {
778 res->mod_info = isl_drm_modifier_get_info(modifier);
779
780 tiling_flags = 1 << res->mod_info->tiling;
781 } else {
782 if (modifiers_count > 0) {
783 fprintf(stderr, "Unsupported modifier, resource creation failed.\n");
784 return NULL;
785 }
786
787 /* Use linear for staging buffers */
788 if (templ->usage == PIPE_USAGE_STAGING ||
789 templ->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR) )
790 tiling_flags = ISL_TILING_LINEAR_BIT;
791 }
792
793 isl_surf_usage_flags_t usage = pipe_bind_to_isl_usage(templ->bind);
794
795 if (templ->target == PIPE_TEXTURE_CUBE ||
796 templ->target == PIPE_TEXTURE_CUBE_ARRAY)
797 usage |= ISL_SURF_USAGE_CUBE_BIT;
798
799 if (templ->usage != PIPE_USAGE_STAGING) {
800 if (templ->format == PIPE_FORMAT_S8_UINT)
801 usage |= ISL_SURF_USAGE_STENCIL_BIT;
802 else if (has_depth)
803 usage |= ISL_SURF_USAGE_DEPTH_BIT;
804 }
805
806 enum pipe_format pfmt = templ->format;
807 res->internal_format = pfmt;
808
809 /* Should be handled by u_transfer_helper */
810 assert(!util_format_is_depth_and_stencil(pfmt));
811
812 struct iris_format_info fmt = iris_format_for_usage(devinfo, pfmt, usage);
813 assert(fmt.fmt != ISL_FORMAT_UNSUPPORTED);
814
815 UNUSED const bool isl_surf_created_successfully =
816 isl_surf_init(&screen->isl_dev, &res->surf,
817 .dim = target_to_isl_surf_dim(templ->target),
818 .format = fmt.fmt,
819 .width = templ->width0,
820 .height = templ->height0,
821 .depth = templ->depth0,
822 .levels = templ->last_level + 1,
823 .array_len = templ->array_size,
824 .samples = MAX2(templ->nr_samples, 1),
825 .min_alignment_B = 0,
826 .row_pitch_B = 0,
827 .usage = usage,
828 .tiling_flags = tiling_flags);
829 assert(isl_surf_created_successfully);
830
831 const char *name = "miptree";
832 enum iris_memory_zone memzone = IRIS_MEMZONE_OTHER;
833
834 unsigned int flags = 0;
835 if (templ->usage == PIPE_USAGE_STAGING)
836 flags |= BO_ALLOC_COHERENT;
837
838 /* These are for u_upload_mgr buffers only */
839 assert(!(templ->flags & (IRIS_RESOURCE_FLAG_SHADER_MEMZONE |
840 IRIS_RESOURCE_FLAG_SURFACE_MEMZONE |
841 IRIS_RESOURCE_FLAG_DYNAMIC_MEMZONE)));
842
843 uint32_t aux_preferred_alloc_flags;
844 uint64_t aux_size = 0;
845 bool aux_enabled =
846 iris_resource_configure_aux(screen, res, false, &aux_size,
847 &aux_preferred_alloc_flags);
848 aux_enabled = aux_enabled && res->aux.surf.size_B > 0;
849 const bool separate_aux = aux_enabled && !res->mod_info;
850 uint64_t aux_offset;
851 uint64_t bo_size;
852
853 if (aux_enabled && !separate_aux) {
854 /* Allocate aux data with main surface. This is required for modifiers
855 * with aux data (ccs).
856 */
857 aux_offset = ALIGN(res->surf.size_B, res->aux.surf.alignment_B);
858 bo_size = aux_offset + aux_size;
859 } else {
860 aux_offset = 0;
861 bo_size = res->surf.size_B;
862 }
863
864 uint32_t alignment = MAX2(4096, res->surf.alignment_B);
865 res->bo = iris_bo_alloc_tiled(screen->bufmgr, name, bo_size, alignment,
866 memzone,
867 isl_tiling_to_i915_tiling(res->surf.tiling),
868 res->surf.row_pitch_B, flags);
869
870 if (!res->bo)
871 goto fail;
872
873 if (aux_enabled) {
874 if (separate_aux) {
875 if (!iris_resource_alloc_separate_aux(screen, res))
876 aux_enabled = false;
877 } else {
878 res->aux.bo = res->bo;
879 iris_bo_reference(res->aux.bo);
880 res->aux.offset += aux_offset;
881 unsigned clear_color_state_size =
882 iris_get_aux_clear_color_state_size(screen);
883 if (clear_color_state_size > 0)
884 res->aux.clear_color_offset += aux_offset;
885 if (!iris_resource_init_aux_buf(res, flags, clear_color_state_size))
886 aux_enabled = false;
887 map_aux_addresses(screen, res);
888 }
889 }
890
891 if (!aux_enabled)
892 iris_resource_disable_aux(res);
893
894 return &res->base;
895
896 fail:
897 fprintf(stderr, "XXX: resource creation failed\n");
898 iris_resource_destroy(pscreen, &res->base);
899 return NULL;
900
901 }
902
903 static struct pipe_resource *
904 iris_resource_create(struct pipe_screen *pscreen,
905 const struct pipe_resource *templ)
906 {
907 if (templ->target == PIPE_BUFFER)
908 return iris_resource_create_for_buffer(pscreen, templ);
909 else
910 return iris_resource_create_with_modifiers(pscreen, templ, NULL, 0);
911 }
912
913 static uint64_t
914 tiling_to_modifier(uint32_t tiling)
915 {
916 static const uint64_t map[] = {
917 [I915_TILING_NONE] = DRM_FORMAT_MOD_LINEAR,
918 [I915_TILING_X] = I915_FORMAT_MOD_X_TILED,
919 [I915_TILING_Y] = I915_FORMAT_MOD_Y_TILED,
920 };
921
922 assert(tiling < ARRAY_SIZE(map));
923
924 return map[tiling];
925 }
926
927 static struct pipe_resource *
928 iris_resource_from_user_memory(struct pipe_screen *pscreen,
929 const struct pipe_resource *templ,
930 void *user_memory)
931 {
932 struct iris_screen *screen = (struct iris_screen *)pscreen;
933 struct iris_bufmgr *bufmgr = screen->bufmgr;
934 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
935 if (!res)
936 return NULL;
937
938 assert(templ->target == PIPE_BUFFER);
939
940 res->internal_format = templ->format;
941 res->bo = iris_bo_create_userptr(bufmgr, "user",
942 user_memory, templ->width0,
943 IRIS_MEMZONE_OTHER);
944 if (!res->bo) {
945 free(res);
946 return NULL;
947 }
948
949 util_range_add(&res->base, &res->valid_buffer_range, 0, templ->width0);
950
951 return &res->base;
952 }
953
954 static struct pipe_resource *
955 iris_resource_from_handle(struct pipe_screen *pscreen,
956 const struct pipe_resource *templ,
957 struct winsys_handle *whandle,
958 unsigned usage)
959 {
960 struct iris_screen *screen = (struct iris_screen *)pscreen;
961 struct gen_device_info *devinfo = &screen->devinfo;
962 struct iris_bufmgr *bufmgr = screen->bufmgr;
963 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
964 if (!res)
965 return NULL;
966
967 switch (whandle->type) {
968 case WINSYS_HANDLE_TYPE_FD:
969 res->bo = iris_bo_import_dmabuf(bufmgr, whandle->handle);
970 break;
971 case WINSYS_HANDLE_TYPE_SHARED:
972 res->bo = iris_bo_gem_create_from_name(bufmgr, "winsys image",
973 whandle->handle);
974 break;
975 default:
976 unreachable("invalid winsys handle type");
977 }
978 if (!res->bo)
979 return NULL;
980
981 res->offset = whandle->offset;
982
983 uint64_t modifier = whandle->modifier;
984 if (modifier == DRM_FORMAT_MOD_INVALID) {
985 modifier = tiling_to_modifier(res->bo->tiling_mode);
986 }
987 res->mod_info = isl_drm_modifier_get_info(modifier);
988 assert(res->mod_info);
989
990 isl_surf_usage_flags_t isl_usage = pipe_bind_to_isl_usage(templ->bind);
991
992 const struct iris_format_info fmt =
993 iris_format_for_usage(devinfo, templ->format, isl_usage);
994 res->internal_format = templ->format;
995
996 if (templ->target == PIPE_BUFFER) {
997 res->surf.tiling = ISL_TILING_LINEAR;
998 } else {
999 if (whandle->modifier == DRM_FORMAT_MOD_INVALID || whandle->plane == 0) {
1000 UNUSED const bool isl_surf_created_successfully =
1001 isl_surf_init(&screen->isl_dev, &res->surf,
1002 .dim = target_to_isl_surf_dim(templ->target),
1003 .format = fmt.fmt,
1004 .width = templ->width0,
1005 .height = templ->height0,
1006 .depth = templ->depth0,
1007 .levels = templ->last_level + 1,
1008 .array_len = templ->array_size,
1009 .samples = MAX2(templ->nr_samples, 1),
1010 .min_alignment_B = 0,
1011 .row_pitch_B = whandle->stride,
1012 .usage = isl_usage,
1013 .tiling_flags = 1 << res->mod_info->tiling);
1014 assert(isl_surf_created_successfully);
1015 assert(res->bo->tiling_mode ==
1016 isl_tiling_to_i915_tiling(res->surf.tiling));
1017
1018 // XXX: create_ccs_buf_for_image?
1019 if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
1020 if (!iris_resource_alloc_separate_aux(screen, res))
1021 goto fail;
1022 } else {
1023 if (res->mod_info->aux_usage != ISL_AUX_USAGE_NONE) {
1024 uint32_t alloc_flags;
1025 uint64_t size;
1026 bool ok = iris_resource_configure_aux(screen, res, true, &size,
1027 &alloc_flags);
1028 assert(ok);
1029 /* The gallium dri layer will create a separate plane resource
1030 * for the aux image. iris_resource_finish_aux_import will
1031 * merge the separate aux parameters back into a single
1032 * iris_resource.
1033 */
1034 }
1035 }
1036 } else {
1037 /* Save modifier import information to reconstruct later. After
1038 * import, this will be available under a second image accessible
1039 * from the main image with res->base.next. See
1040 * iris_resource_finish_aux_import.
1041 */
1042 res->aux.surf.row_pitch_B = whandle->stride;
1043 res->aux.offset = whandle->offset;
1044 res->aux.bo = res->bo;
1045 res->bo = NULL;
1046 }
1047 }
1048
1049 return &res->base;
1050
1051 fail:
1052 iris_resource_destroy(pscreen, &res->base);
1053 return NULL;
1054 }
1055
1056 static void
1057 iris_flush_resource(struct pipe_context *ctx, struct pipe_resource *resource)
1058 {
1059 struct iris_context *ice = (struct iris_context *)ctx;
1060 struct iris_batch *render_batch = &ice->batches[IRIS_BATCH_RENDER];
1061 struct iris_resource *res = (void *) resource;
1062 const struct isl_drm_modifier_info *mod = res->mod_info;
1063
1064 iris_resource_prepare_access(ice, render_batch, res,
1065 0, INTEL_REMAINING_LEVELS,
1066 0, INTEL_REMAINING_LAYERS,
1067 mod ? mod->aux_usage : ISL_AUX_USAGE_NONE,
1068 mod ? mod->supports_clear_color : false);
1069 }
1070
1071 static void
1072 iris_resource_disable_aux_on_first_query(struct pipe_resource *resource,
1073 unsigned usage)
1074 {
1075 struct iris_resource *res = (struct iris_resource *)resource;
1076 bool mod_with_aux =
1077 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1078
1079 /* Disable aux usage if explicit flush not set and this is the first time
1080 * we are dealing with this resource and the resource was not created with
1081 * a modifier with aux.
1082 */
1083 if (!mod_with_aux &&
1084 (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) && res->aux.usage != 0) &&
1085 p_atomic_read(&resource->reference.count) == 1) {
1086 iris_resource_disable_aux(res);
1087 }
1088 }
1089
1090 static bool
1091 iris_resource_get_param(struct pipe_screen *screen,
1092 struct pipe_context *context,
1093 struct pipe_resource *resource,
1094 unsigned plane,
1095 unsigned layer,
1096 enum pipe_resource_param param,
1097 unsigned handle_usage,
1098 uint64_t *value)
1099 {
1100 struct iris_resource *res = (struct iris_resource *)resource;
1101 bool mod_with_aux =
1102 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1103 bool wants_aux = mod_with_aux && plane > 0;
1104 bool result;
1105 unsigned handle;
1106
1107 if (iris_resource_unfinished_aux_import(res))
1108 iris_resource_finish_aux_import(screen, res);
1109
1110 struct iris_bo *bo = wants_aux ? res->aux.bo : res->bo;
1111
1112 iris_resource_disable_aux_on_first_query(resource, handle_usage);
1113
1114 switch (param) {
1115 case PIPE_RESOURCE_PARAM_NPLANES:
1116 if (mod_with_aux) {
1117 *value = 2;
1118 } else {
1119 unsigned count = 0;
1120 for (struct pipe_resource *cur = resource; cur; cur = cur->next)
1121 count++;
1122 *value = count;
1123 }
1124 return true;
1125 case PIPE_RESOURCE_PARAM_STRIDE:
1126 *value = wants_aux ? res->aux.surf.row_pitch_B : res->surf.row_pitch_B;
1127 return true;
1128 case PIPE_RESOURCE_PARAM_OFFSET:
1129 *value = wants_aux ? res->aux.offset : 0;
1130 return true;
1131 case PIPE_RESOURCE_PARAM_MODIFIER:
1132 *value = res->mod_info ? res->mod_info->modifier :
1133 tiling_to_modifier(res->bo->tiling_mode);
1134 return true;
1135 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_SHARED:
1136 result = iris_bo_flink(bo, &handle) == 0;
1137 if (result)
1138 *value = handle;
1139 return result;
1140 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_KMS:
1141 *value = iris_bo_export_gem_handle(bo);
1142 return true;
1143 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_FD:
1144 result = iris_bo_export_dmabuf(bo, (int *) &handle) == 0;
1145 if (result)
1146 *value = handle;
1147 return result;
1148 default:
1149 return false;
1150 }
1151 }
1152
1153 static bool
1154 iris_resource_get_handle(struct pipe_screen *pscreen,
1155 struct pipe_context *ctx,
1156 struct pipe_resource *resource,
1157 struct winsys_handle *whandle,
1158 unsigned usage)
1159 {
1160 struct iris_resource *res = (struct iris_resource *)resource;
1161 bool mod_with_aux =
1162 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1163
1164 iris_resource_disable_aux_on_first_query(resource, usage);
1165
1166 struct iris_bo *bo;
1167 if (mod_with_aux && whandle->plane > 0) {
1168 assert(res->aux.bo);
1169 bo = res->aux.bo;
1170 whandle->stride = res->aux.surf.row_pitch_B;
1171 whandle->offset = res->aux.offset;
1172 } else {
1173 /* If this is a buffer, stride should be 0 - no need to special case */
1174 whandle->stride = res->surf.row_pitch_B;
1175 bo = res->bo;
1176 }
1177 whandle->modifier =
1178 res->mod_info ? res->mod_info->modifier
1179 : tiling_to_modifier(res->bo->tiling_mode);
1180
1181 #ifndef NDEBUG
1182 enum isl_aux_usage allowed_usage =
1183 res->mod_info ? res->mod_info->aux_usage : ISL_AUX_USAGE_NONE;
1184
1185 if (res->aux.usage != allowed_usage) {
1186 enum isl_aux_state aux_state = iris_resource_get_aux_state(res, 0, 0);
1187 assert(aux_state == ISL_AUX_STATE_RESOLVED ||
1188 aux_state == ISL_AUX_STATE_PASS_THROUGH);
1189 }
1190 #endif
1191
1192 switch (whandle->type) {
1193 case WINSYS_HANDLE_TYPE_SHARED:
1194 return iris_bo_flink(bo, &whandle->handle) == 0;
1195 case WINSYS_HANDLE_TYPE_KMS:
1196 whandle->handle = iris_bo_export_gem_handle(bo);
1197 return true;
1198 case WINSYS_HANDLE_TYPE_FD:
1199 return iris_bo_export_dmabuf(bo, (int *) &whandle->handle) == 0;
1200 }
1201
1202 return false;
1203 }
1204
1205 static bool
1206 resource_is_busy(struct iris_context *ice,
1207 struct iris_resource *res)
1208 {
1209 bool busy = iris_bo_busy(res->bo);
1210
1211 for (int i = 0; i < IRIS_BATCH_COUNT; i++)
1212 busy |= iris_batch_references(&ice->batches[i], res->bo);
1213
1214 return busy;
1215 }
1216
1217 static void
1218 iris_invalidate_resource(struct pipe_context *ctx,
1219 struct pipe_resource *resource)
1220 {
1221 struct iris_screen *screen = (void *) ctx->screen;
1222 struct iris_context *ice = (void *) ctx;
1223 struct iris_resource *res = (void *) resource;
1224
1225 if (resource->target != PIPE_BUFFER)
1226 return;
1227
1228 if (!resource_is_busy(ice, res)) {
1229 /* The resource is idle, so just mark that it contains no data and
1230 * keep using the same underlying buffer object.
1231 */
1232 util_range_set_empty(&res->valid_buffer_range);
1233 return;
1234 }
1235
1236 /* Otherwise, try and replace the backing storage with a new BO. */
1237
1238 /* We can't reallocate memory we didn't allocate in the first place. */
1239 if (res->bo->userptr)
1240 return;
1241
1242 // XXX: We should support this.
1243 if (res->bind_history & PIPE_BIND_STREAM_OUTPUT)
1244 return;
1245
1246 struct iris_bo *old_bo = res->bo;
1247 struct iris_bo *new_bo =
1248 iris_bo_alloc(screen->bufmgr, res->bo->name, resource->width0,
1249 iris_memzone_for_address(old_bo->gtt_offset));
1250 if (!new_bo)
1251 return;
1252
1253 /* Swap out the backing storage */
1254 res->bo = new_bo;
1255
1256 /* Rebind the buffer, replacing any state referring to the old BO's
1257 * address, and marking state dirty so it's reemitted.
1258 */
1259 ice->vtbl.rebind_buffer(ice, res, old_bo->gtt_offset);
1260
1261 util_range_set_empty(&res->valid_buffer_range);
1262
1263 iris_bo_unreference(old_bo);
1264 }
1265
1266 static void
1267 iris_flush_staging_region(struct pipe_transfer *xfer,
1268 const struct pipe_box *flush_box)
1269 {
1270 if (!(xfer->usage & PIPE_TRANSFER_WRITE))
1271 return;
1272
1273 struct iris_transfer *map = (void *) xfer;
1274
1275 struct pipe_box src_box = *flush_box;
1276
1277 /* Account for extra alignment padding in staging buffer */
1278 if (xfer->resource->target == PIPE_BUFFER)
1279 src_box.x += xfer->box.x % IRIS_MAP_BUFFER_ALIGNMENT;
1280
1281 struct pipe_box dst_box = (struct pipe_box) {
1282 .x = xfer->box.x + flush_box->x,
1283 .y = xfer->box.y + flush_box->y,
1284 .z = xfer->box.z + flush_box->z,
1285 .width = flush_box->width,
1286 .height = flush_box->height,
1287 .depth = flush_box->depth,
1288 };
1289
1290 iris_copy_region(map->blorp, map->batch, xfer->resource, xfer->level,
1291 dst_box.x, dst_box.y, dst_box.z, map->staging, 0,
1292 &src_box);
1293 }
1294
1295 static void
1296 iris_unmap_copy_region(struct iris_transfer *map)
1297 {
1298 iris_resource_destroy(map->staging->screen, map->staging);
1299
1300 map->ptr = NULL;
1301 }
1302
1303 static void
1304 iris_map_copy_region(struct iris_transfer *map)
1305 {
1306 struct pipe_screen *pscreen = &map->batch->screen->base;
1307 struct pipe_transfer *xfer = &map->base;
1308 struct pipe_box *box = &xfer->box;
1309 struct iris_resource *res = (void *) xfer->resource;
1310
1311 unsigned extra = xfer->resource->target == PIPE_BUFFER ?
1312 box->x % IRIS_MAP_BUFFER_ALIGNMENT : 0;
1313
1314 struct pipe_resource templ = (struct pipe_resource) {
1315 .usage = PIPE_USAGE_STAGING,
1316 .width0 = box->width + extra,
1317 .height0 = box->height,
1318 .depth0 = 1,
1319 .nr_samples = xfer->resource->nr_samples,
1320 .nr_storage_samples = xfer->resource->nr_storage_samples,
1321 .array_size = box->depth,
1322 .format = res->internal_format,
1323 };
1324
1325 if (xfer->resource->target == PIPE_BUFFER)
1326 templ.target = PIPE_BUFFER;
1327 else if (templ.array_size > 1)
1328 templ.target = PIPE_TEXTURE_2D_ARRAY;
1329 else
1330 templ.target = PIPE_TEXTURE_2D;
1331
1332 map->staging = iris_resource_create(pscreen, &templ);
1333 assert(map->staging);
1334
1335 if (templ.target != PIPE_BUFFER) {
1336 struct isl_surf *surf = &((struct iris_resource *) map->staging)->surf;
1337 xfer->stride = isl_surf_get_row_pitch_B(surf);
1338 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1339 }
1340
1341 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1342 iris_copy_region(map->blorp, map->batch, map->staging, 0, extra, 0, 0,
1343 xfer->resource, xfer->level, box);
1344 /* Ensure writes to the staging BO land before we map it below. */
1345 iris_emit_pipe_control_flush(map->batch,
1346 "transfer read: flush before mapping",
1347 PIPE_CONTROL_RENDER_TARGET_FLUSH |
1348 PIPE_CONTROL_CS_STALL);
1349 }
1350
1351 struct iris_bo *staging_bo = iris_resource_bo(map->staging);
1352
1353 if (iris_batch_references(map->batch, staging_bo))
1354 iris_batch_flush(map->batch);
1355
1356 map->ptr =
1357 iris_bo_map(map->dbg, staging_bo, xfer->usage & MAP_FLAGS) + extra;
1358
1359 map->unmap = iris_unmap_copy_region;
1360 }
1361
1362 static void
1363 get_image_offset_el(const struct isl_surf *surf, unsigned level, unsigned z,
1364 unsigned *out_x0_el, unsigned *out_y0_el)
1365 {
1366 if (surf->dim == ISL_SURF_DIM_3D) {
1367 isl_surf_get_image_offset_el(surf, level, 0, z, out_x0_el, out_y0_el);
1368 } else {
1369 isl_surf_get_image_offset_el(surf, level, z, 0, out_x0_el, out_y0_el);
1370 }
1371 }
1372
1373 /**
1374 * This function computes the tile_w (in bytes) and tile_h (in rows) of
1375 * different tiling patterns.
1376 */
1377 static void
1378 iris_resource_get_tile_dims(enum isl_tiling tiling, uint32_t cpp,
1379 uint32_t *tile_w, uint32_t *tile_h)
1380 {
1381 switch (tiling) {
1382 case ISL_TILING_X:
1383 *tile_w = 512;
1384 *tile_h = 8;
1385 break;
1386 case ISL_TILING_Y0:
1387 *tile_w = 128;
1388 *tile_h = 32;
1389 break;
1390 case ISL_TILING_LINEAR:
1391 *tile_w = cpp;
1392 *tile_h = 1;
1393 break;
1394 default:
1395 unreachable("not reached");
1396 }
1397
1398 }
1399
1400 /**
1401 * This function computes masks that may be used to select the bits of the X
1402 * and Y coordinates that indicate the offset within a tile. If the BO is
1403 * untiled, the masks are set to 0.
1404 */
1405 static void
1406 iris_resource_get_tile_masks(enum isl_tiling tiling, uint32_t cpp,
1407 uint32_t *mask_x, uint32_t *mask_y)
1408 {
1409 uint32_t tile_w_bytes, tile_h;
1410
1411 iris_resource_get_tile_dims(tiling, cpp, &tile_w_bytes, &tile_h);
1412
1413 *mask_x = tile_w_bytes / cpp - 1;
1414 *mask_y = tile_h - 1;
1415 }
1416
1417 /**
1418 * Compute the offset (in bytes) from the start of the BO to the given x
1419 * and y coordinate. For tiled BOs, caller must ensure that x and y are
1420 * multiples of the tile size.
1421 */
1422 static uint32_t
1423 iris_resource_get_aligned_offset(const struct iris_resource *res,
1424 uint32_t x, uint32_t y)
1425 {
1426 const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
1427 unsigned cpp = fmtl->bpb / 8;
1428 uint32_t pitch = res->surf.row_pitch_B;
1429
1430 switch (res->surf.tiling) {
1431 default:
1432 unreachable("not reached");
1433 case ISL_TILING_LINEAR:
1434 return y * pitch + x * cpp;
1435 case ISL_TILING_X:
1436 assert((x % (512 / cpp)) == 0);
1437 assert((y % 8) == 0);
1438 return y * pitch + x / (512 / cpp) * 4096;
1439 case ISL_TILING_Y0:
1440 assert((x % (128 / cpp)) == 0);
1441 assert((y % 32) == 0);
1442 return y * pitch + x / (128 / cpp) * 4096;
1443 }
1444 }
1445
1446 /**
1447 * Rendering with tiled buffers requires that the base address of the buffer
1448 * be aligned to a page boundary. For renderbuffers, and sometimes with
1449 * textures, we may want the surface to point at a texture image level that
1450 * isn't at a page boundary.
1451 *
1452 * This function returns an appropriately-aligned base offset
1453 * according to the tiling restrictions, plus any required x/y offset
1454 * from there.
1455 */
1456 uint32_t
1457 iris_resource_get_tile_offsets(const struct iris_resource *res,
1458 uint32_t level, uint32_t z,
1459 uint32_t *tile_x, uint32_t *tile_y)
1460 {
1461 uint32_t x, y;
1462 uint32_t mask_x, mask_y;
1463
1464 const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
1465 const unsigned cpp = fmtl->bpb / 8;
1466
1467 iris_resource_get_tile_masks(res->surf.tiling, cpp, &mask_x, &mask_y);
1468 get_image_offset_el(&res->surf, level, z, &x, &y);
1469
1470 *tile_x = x & mask_x;
1471 *tile_y = y & mask_y;
1472
1473 return iris_resource_get_aligned_offset(res, x & ~mask_x, y & ~mask_y);
1474 }
1475
1476 /**
1477 * Get pointer offset into stencil buffer.
1478 *
1479 * The stencil buffer is W tiled. Since the GTT is incapable of W fencing, we
1480 * must decode the tile's layout in software.
1481 *
1482 * See
1483 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.2.1 W-Major Tile
1484 * Format.
1485 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.3 Tiling Algorithm
1486 *
1487 * Even though the returned offset is always positive, the return type is
1488 * signed due to
1489 * commit e8b1c6d6f55f5be3bef25084fdd8b6127517e137
1490 * mesa: Fix return type of _mesa_get_format_bytes() (#37351)
1491 */
1492 static intptr_t
1493 s8_offset(uint32_t stride, uint32_t x, uint32_t y)
1494 {
1495 uint32_t tile_size = 4096;
1496 uint32_t tile_width = 64;
1497 uint32_t tile_height = 64;
1498 uint32_t row_size = 64 * stride / 2; /* Two rows are interleaved. */
1499
1500 uint32_t tile_x = x / tile_width;
1501 uint32_t tile_y = y / tile_height;
1502
1503 /* The byte's address relative to the tile's base addres. */
1504 uint32_t byte_x = x % tile_width;
1505 uint32_t byte_y = y % tile_height;
1506
1507 uintptr_t u = tile_y * row_size
1508 + tile_x * tile_size
1509 + 512 * (byte_x / 8)
1510 + 64 * (byte_y / 8)
1511 + 32 * ((byte_y / 4) % 2)
1512 + 16 * ((byte_x / 4) % 2)
1513 + 8 * ((byte_y / 2) % 2)
1514 + 4 * ((byte_x / 2) % 2)
1515 + 2 * (byte_y % 2)
1516 + 1 * (byte_x % 2);
1517
1518 return u;
1519 }
1520
1521 static void
1522 iris_unmap_s8(struct iris_transfer *map)
1523 {
1524 struct pipe_transfer *xfer = &map->base;
1525 const struct pipe_box *box = &xfer->box;
1526 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1527 struct isl_surf *surf = &res->surf;
1528
1529 if (xfer->usage & PIPE_TRANSFER_WRITE) {
1530 uint8_t *untiled_s8_map = map->ptr;
1531 uint8_t *tiled_s8_map =
1532 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1533
1534 for (int s = 0; s < box->depth; s++) {
1535 unsigned x0_el, y0_el;
1536 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1537
1538 for (uint32_t y = 0; y < box->height; y++) {
1539 for (uint32_t x = 0; x < box->width; x++) {
1540 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1541 x0_el + box->x + x,
1542 y0_el + box->y + y);
1543 tiled_s8_map[offset] =
1544 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x];
1545 }
1546 }
1547 }
1548 }
1549
1550 free(map->buffer);
1551 }
1552
1553 static void
1554 iris_map_s8(struct iris_transfer *map)
1555 {
1556 struct pipe_transfer *xfer = &map->base;
1557 const struct pipe_box *box = &xfer->box;
1558 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1559 struct isl_surf *surf = &res->surf;
1560
1561 xfer->stride = surf->row_pitch_B;
1562 xfer->layer_stride = xfer->stride * box->height;
1563
1564 /* The tiling and detiling functions require that the linear buffer has
1565 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1566 * over-allocate the linear buffer to get the proper alignment.
1567 */
1568 map->buffer = map->ptr = malloc(xfer->layer_stride * box->depth);
1569 assert(map->buffer);
1570
1571 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
1572 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
1573 * invalidate is set, since we'll be writing the whole rectangle from our
1574 * temporary buffer back out.
1575 */
1576 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1577 uint8_t *untiled_s8_map = map->ptr;
1578 uint8_t *tiled_s8_map =
1579 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1580
1581 for (int s = 0; s < box->depth; s++) {
1582 unsigned x0_el, y0_el;
1583 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1584
1585 for (uint32_t y = 0; y < box->height; y++) {
1586 for (uint32_t x = 0; x < box->width; x++) {
1587 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1588 x0_el + box->x + x,
1589 y0_el + box->y + y);
1590 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x] =
1591 tiled_s8_map[offset];
1592 }
1593 }
1594 }
1595 }
1596
1597 map->unmap = iris_unmap_s8;
1598 }
1599
1600 /* Compute extent parameters for use with tiled_memcpy functions.
1601 * xs are in units of bytes and ys are in units of strides.
1602 */
1603 static inline void
1604 tile_extents(const struct isl_surf *surf,
1605 const struct pipe_box *box,
1606 unsigned level, int z,
1607 unsigned *x1_B, unsigned *x2_B,
1608 unsigned *y1_el, unsigned *y2_el)
1609 {
1610 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1611 const unsigned cpp = fmtl->bpb / 8;
1612
1613 assert(box->x % fmtl->bw == 0);
1614 assert(box->y % fmtl->bh == 0);
1615
1616 unsigned x0_el, y0_el;
1617 get_image_offset_el(surf, level, box->z + z, &x0_el, &y0_el);
1618
1619 *x1_B = (box->x / fmtl->bw + x0_el) * cpp;
1620 *y1_el = box->y / fmtl->bh + y0_el;
1621 *x2_B = (DIV_ROUND_UP(box->x + box->width, fmtl->bw) + x0_el) * cpp;
1622 *y2_el = DIV_ROUND_UP(box->y + box->height, fmtl->bh) + y0_el;
1623 }
1624
1625 static void
1626 iris_unmap_tiled_memcpy(struct iris_transfer *map)
1627 {
1628 struct pipe_transfer *xfer = &map->base;
1629 const struct pipe_box *box = &xfer->box;
1630 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1631 struct isl_surf *surf = &res->surf;
1632
1633 const bool has_swizzling = false;
1634
1635 if (xfer->usage & PIPE_TRANSFER_WRITE) {
1636 char *dst =
1637 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1638
1639 for (int s = 0; s < box->depth; s++) {
1640 unsigned x1, x2, y1, y2;
1641 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1642
1643 void *ptr = map->ptr + s * xfer->layer_stride;
1644
1645 isl_memcpy_linear_to_tiled(x1, x2, y1, y2, dst, ptr,
1646 surf->row_pitch_B, xfer->stride,
1647 has_swizzling, surf->tiling, ISL_MEMCPY);
1648 }
1649 }
1650 os_free_aligned(map->buffer);
1651 map->buffer = map->ptr = NULL;
1652 }
1653
1654 static void
1655 iris_map_tiled_memcpy(struct iris_transfer *map)
1656 {
1657 struct pipe_transfer *xfer = &map->base;
1658 const struct pipe_box *box = &xfer->box;
1659 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1660 struct isl_surf *surf = &res->surf;
1661
1662 xfer->stride = ALIGN(surf->row_pitch_B, 16);
1663 xfer->layer_stride = xfer->stride * box->height;
1664
1665 unsigned x1, x2, y1, y2;
1666 tile_extents(surf, box, xfer->level, 0, &x1, &x2, &y1, &y2);
1667
1668 /* The tiling and detiling functions require that the linear buffer has
1669 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1670 * over-allocate the linear buffer to get the proper alignment.
1671 */
1672 map->buffer =
1673 os_malloc_aligned(xfer->layer_stride * box->depth, 16);
1674 assert(map->buffer);
1675 map->ptr = (char *)map->buffer + (x1 & 0xf);
1676
1677 const bool has_swizzling = false;
1678
1679 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1680 char *src =
1681 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1682
1683 for (int s = 0; s < box->depth; s++) {
1684 unsigned x1, x2, y1, y2;
1685 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1686
1687 /* Use 's' rather than 'box->z' to rebase the first slice to 0. */
1688 void *ptr = map->ptr + s * xfer->layer_stride;
1689
1690 isl_memcpy_tiled_to_linear(x1, x2, y1, y2, ptr, src, xfer->stride,
1691 surf->row_pitch_B, has_swizzling,
1692 surf->tiling, ISL_MEMCPY_STREAMING_LOAD);
1693 }
1694 }
1695
1696 map->unmap = iris_unmap_tiled_memcpy;
1697 }
1698
1699 static void
1700 iris_map_direct(struct iris_transfer *map)
1701 {
1702 struct pipe_transfer *xfer = &map->base;
1703 struct pipe_box *box = &xfer->box;
1704 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1705
1706 void *ptr = iris_bo_map(map->dbg, res->bo, xfer->usage & MAP_FLAGS);
1707
1708 if (res->base.target == PIPE_BUFFER) {
1709 xfer->stride = 0;
1710 xfer->layer_stride = 0;
1711
1712 map->ptr = ptr + box->x;
1713 } else {
1714 struct isl_surf *surf = &res->surf;
1715 const struct isl_format_layout *fmtl =
1716 isl_format_get_layout(surf->format);
1717 const unsigned cpp = fmtl->bpb / 8;
1718 unsigned x0_el, y0_el;
1719
1720 get_image_offset_el(surf, xfer->level, box->z, &x0_el, &y0_el);
1721
1722 xfer->stride = isl_surf_get_row_pitch_B(surf);
1723 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1724
1725 map->ptr = ptr + (y0_el + box->y) * xfer->stride + (x0_el + box->x) * cpp;
1726 }
1727 }
1728
1729 static bool
1730 can_promote_to_async(const struct iris_resource *res,
1731 const struct pipe_box *box,
1732 enum pipe_transfer_usage usage)
1733 {
1734 /* If we're writing to a section of the buffer that hasn't even been
1735 * initialized with useful data, then we can safely promote this write
1736 * to be unsynchronized. This helps the common pattern of appending data.
1737 */
1738 return res->base.target == PIPE_BUFFER && (usage & PIPE_TRANSFER_WRITE) &&
1739 !(usage & TC_TRANSFER_MAP_NO_INFER_UNSYNCHRONIZED) &&
1740 !util_ranges_intersect(&res->valid_buffer_range, box->x,
1741 box->x + box->width);
1742 }
1743
1744 static void *
1745 iris_transfer_map(struct pipe_context *ctx,
1746 struct pipe_resource *resource,
1747 unsigned level,
1748 enum pipe_transfer_usage usage,
1749 const struct pipe_box *box,
1750 struct pipe_transfer **ptransfer)
1751 {
1752 struct iris_context *ice = (struct iris_context *)ctx;
1753 struct iris_resource *res = (struct iris_resource *)resource;
1754 struct isl_surf *surf = &res->surf;
1755
1756 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
1757 /* Replace the backing storage with a fresh buffer for non-async maps */
1758 if (!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
1759 TC_TRANSFER_MAP_NO_INVALIDATE)))
1760 iris_invalidate_resource(ctx, resource);
1761
1762 /* If we can discard the whole resource, we can discard the range. */
1763 usage |= PIPE_TRANSFER_DISCARD_RANGE;
1764 }
1765
1766 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
1767 can_promote_to_async(res, box, usage)) {
1768 usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
1769 }
1770
1771 bool need_resolve = false;
1772 bool need_color_resolve = false;
1773
1774 if (resource->target != PIPE_BUFFER) {
1775 bool need_hiz_resolve = iris_resource_level_has_hiz(res, level);
1776
1777 need_color_resolve =
1778 (res->aux.usage == ISL_AUX_USAGE_CCS_D ||
1779 res->aux.usage == ISL_AUX_USAGE_CCS_E) &&
1780 iris_has_color_unresolved(res, level, 1, box->z, box->depth);
1781
1782 need_resolve = need_color_resolve || need_hiz_resolve;
1783 }
1784
1785 bool map_would_stall = false;
1786
1787 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
1788 map_would_stall = need_resolve || resource_is_busy(ice, res);
1789
1790 if (map_would_stall && (usage & PIPE_TRANSFER_DONTBLOCK) &&
1791 (usage & PIPE_TRANSFER_MAP_DIRECTLY))
1792 return NULL;
1793 }
1794
1795 if (surf->tiling != ISL_TILING_LINEAR &&
1796 (usage & PIPE_TRANSFER_MAP_DIRECTLY))
1797 return NULL;
1798
1799 struct iris_transfer *map = slab_alloc(&ice->transfer_pool);
1800 struct pipe_transfer *xfer = &map->base;
1801
1802 if (!map)
1803 return NULL;
1804
1805 memset(map, 0, sizeof(*map));
1806 map->dbg = &ice->dbg;
1807
1808 pipe_resource_reference(&xfer->resource, resource);
1809 xfer->level = level;
1810 xfer->usage = usage;
1811 xfer->box = *box;
1812 *ptransfer = xfer;
1813
1814 map->dest_had_defined_contents =
1815 util_ranges_intersect(&res->valid_buffer_range, box->x,
1816 box->x + box->width);
1817
1818 if (usage & PIPE_TRANSFER_WRITE)
1819 util_range_add(&res->base, &res->valid_buffer_range, box->x, box->x + box->width);
1820
1821 /* Avoid using GPU copies for persistent/coherent buffers, as the idea
1822 * there is to access them simultaneously on the CPU & GPU. This also
1823 * avoids trying to use GPU copies for our u_upload_mgr buffers which
1824 * contain state we're constructing for a GPU draw call, which would
1825 * kill us with infinite stack recursion.
1826 */
1827 bool no_gpu = usage & (PIPE_TRANSFER_PERSISTENT |
1828 PIPE_TRANSFER_COHERENT |
1829 PIPE_TRANSFER_MAP_DIRECTLY);
1830
1831 /* GPU copies are not useful for buffer reads. Instead of stalling to
1832 * read from the original buffer, we'd simply copy it to a temporary...
1833 * then stall (a bit longer) to read from that buffer.
1834 *
1835 * Images are less clear-cut. Color resolves are destructive, removing
1836 * the underlying compression, so we'd rather blit the data to a linear
1837 * temporary and map that, to avoid the resolve. (It might be better to
1838 * a tiled temporary and use the tiled_memcpy paths...)
1839 */
1840 if (!(usage & PIPE_TRANSFER_DISCARD_RANGE) && !need_color_resolve)
1841 no_gpu = true;
1842
1843 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1844 if (fmtl->txc == ISL_TXC_ASTC)
1845 no_gpu = true;
1846
1847 if ((map_would_stall || res->aux.usage == ISL_AUX_USAGE_CCS_E) && !no_gpu) {
1848 /* If we need a synchronous mapping and the resource is busy, or needs
1849 * resolving, we copy to/from a linear temporary buffer using the GPU.
1850 */
1851 map->batch = &ice->batches[IRIS_BATCH_RENDER];
1852 map->blorp = &ice->blorp;
1853 iris_map_copy_region(map);
1854 } else {
1855 /* Otherwise we're free to map on the CPU. */
1856
1857 if (need_resolve) {
1858 iris_resource_access_raw(ice, &ice->batches[IRIS_BATCH_RENDER], res,
1859 level, box->z, box->depth,
1860 usage & PIPE_TRANSFER_WRITE);
1861 }
1862
1863 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
1864 for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
1865 if (iris_batch_references(&ice->batches[i], res->bo))
1866 iris_batch_flush(&ice->batches[i]);
1867 }
1868 }
1869
1870 if (surf->tiling == ISL_TILING_W) {
1871 /* TODO: Teach iris_map_tiled_memcpy about W-tiling... */
1872 iris_map_s8(map);
1873 } else if (surf->tiling != ISL_TILING_LINEAR) {
1874 iris_map_tiled_memcpy(map);
1875 } else {
1876 iris_map_direct(map);
1877 }
1878 }
1879
1880 return map->ptr;
1881 }
1882
1883 static void
1884 iris_transfer_flush_region(struct pipe_context *ctx,
1885 struct pipe_transfer *xfer,
1886 const struct pipe_box *box)
1887 {
1888 struct iris_context *ice = (struct iris_context *)ctx;
1889 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1890 struct iris_transfer *map = (void *) xfer;
1891
1892 if (map->staging)
1893 iris_flush_staging_region(xfer, box);
1894
1895 uint32_t history_flush = 0;
1896
1897 if (res->base.target == PIPE_BUFFER) {
1898 if (map->staging)
1899 history_flush |= PIPE_CONTROL_RENDER_TARGET_FLUSH;
1900
1901 if (map->dest_had_defined_contents)
1902 history_flush |= iris_flush_bits_for_history(res);
1903
1904 util_range_add(&res->base, &res->valid_buffer_range, box->x, box->x + box->width);
1905 }
1906
1907 if (history_flush & ~PIPE_CONTROL_CS_STALL) {
1908 for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
1909 struct iris_batch *batch = &ice->batches[i];
1910 if (batch->contains_draw || batch->cache.render->entries) {
1911 iris_batch_maybe_flush(batch, 24);
1912 iris_emit_pipe_control_flush(batch,
1913 "cache history: transfer flush",
1914 history_flush);
1915 }
1916 }
1917 }
1918
1919 /* Make sure we flag constants dirty even if there's no need to emit
1920 * any PIPE_CONTROLs to a batch.
1921 */
1922 iris_dirty_for_history(ice, res);
1923 }
1924
1925 static void
1926 iris_transfer_unmap(struct pipe_context *ctx, struct pipe_transfer *xfer)
1927 {
1928 struct iris_context *ice = (struct iris_context *)ctx;
1929 struct iris_transfer *map = (void *) xfer;
1930
1931 if (!(xfer->usage & (PIPE_TRANSFER_FLUSH_EXPLICIT |
1932 PIPE_TRANSFER_COHERENT))) {
1933 struct pipe_box flush_box = {
1934 .x = 0, .y = 0, .z = 0,
1935 .width = xfer->box.width,
1936 .height = xfer->box.height,
1937 .depth = xfer->box.depth,
1938 };
1939 iris_transfer_flush_region(ctx, xfer, &flush_box);
1940 }
1941
1942 if (map->unmap)
1943 map->unmap(map);
1944
1945 pipe_resource_reference(&xfer->resource, NULL);
1946 slab_free(&ice->transfer_pool, map);
1947 }
1948
1949 /**
1950 * Mark state dirty that needs to be re-emitted when a resource is written.
1951 */
1952 void
1953 iris_dirty_for_history(struct iris_context *ice,
1954 struct iris_resource *res)
1955 {
1956 uint64_t dirty = 0ull;
1957
1958 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1959 dirty |= ((uint64_t)res->bind_stages) << IRIS_SHIFT_FOR_DIRTY_CONSTANTS;
1960 }
1961
1962 ice->state.dirty |= dirty;
1963 }
1964
1965 /**
1966 * Produce a set of PIPE_CONTROL bits which ensure data written to a
1967 * resource becomes visible, and any stale read cache data is invalidated.
1968 */
1969 uint32_t
1970 iris_flush_bits_for_history(struct iris_resource *res)
1971 {
1972 uint32_t flush = PIPE_CONTROL_CS_STALL;
1973
1974 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1975 flush |= PIPE_CONTROL_CONST_CACHE_INVALIDATE |
1976 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1977 }
1978
1979 if (res->bind_history & PIPE_BIND_SAMPLER_VIEW)
1980 flush |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1981
1982 if (res->bind_history & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
1983 flush |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1984
1985 if (res->bind_history & (PIPE_BIND_SHADER_BUFFER | PIPE_BIND_SHADER_IMAGE))
1986 flush |= PIPE_CONTROL_DATA_CACHE_FLUSH;
1987
1988 return flush;
1989 }
1990
1991 void
1992 iris_flush_and_dirty_for_history(struct iris_context *ice,
1993 struct iris_batch *batch,
1994 struct iris_resource *res,
1995 uint32_t extra_flags,
1996 const char *reason)
1997 {
1998 if (res->base.target != PIPE_BUFFER)
1999 return;
2000
2001 uint32_t flush = iris_flush_bits_for_history(res) | extra_flags;
2002
2003 iris_emit_pipe_control_flush(batch, reason, flush);
2004
2005 iris_dirty_for_history(ice, res);
2006 }
2007
2008 bool
2009 iris_resource_set_clear_color(struct iris_context *ice,
2010 struct iris_resource *res,
2011 union isl_color_value color)
2012 {
2013 if (memcmp(&res->aux.clear_color, &color, sizeof(color)) != 0) {
2014 res->aux.clear_color = color;
2015 return true;
2016 }
2017
2018 return false;
2019 }
2020
2021 union isl_color_value
2022 iris_resource_get_clear_color(const struct iris_resource *res,
2023 struct iris_bo **clear_color_bo,
2024 uint64_t *clear_color_offset)
2025 {
2026 assert(res->aux.bo);
2027
2028 if (clear_color_bo)
2029 *clear_color_bo = res->aux.clear_color_bo;
2030 if (clear_color_offset)
2031 *clear_color_offset = res->aux.clear_color_offset;
2032 return res->aux.clear_color;
2033 }
2034
2035 static enum pipe_format
2036 iris_resource_get_internal_format(struct pipe_resource *p_res)
2037 {
2038 struct iris_resource *res = (void *) p_res;
2039 return res->internal_format;
2040 }
2041
2042 static const struct u_transfer_vtbl transfer_vtbl = {
2043 .resource_create = iris_resource_create,
2044 .resource_destroy = iris_resource_destroy,
2045 .transfer_map = iris_transfer_map,
2046 .transfer_unmap = iris_transfer_unmap,
2047 .transfer_flush_region = iris_transfer_flush_region,
2048 .get_internal_format = iris_resource_get_internal_format,
2049 .set_stencil = iris_resource_set_separate_stencil,
2050 .get_stencil = iris_resource_get_separate_stencil,
2051 };
2052
2053 void
2054 iris_init_screen_resource_functions(struct pipe_screen *pscreen)
2055 {
2056 pscreen->query_dmabuf_modifiers = iris_query_dmabuf_modifiers;
2057 pscreen->resource_create_with_modifiers =
2058 iris_resource_create_with_modifiers;
2059 pscreen->resource_create = u_transfer_helper_resource_create;
2060 pscreen->resource_from_user_memory = iris_resource_from_user_memory;
2061 pscreen->resource_from_handle = iris_resource_from_handle;
2062 pscreen->resource_get_handle = iris_resource_get_handle;
2063 pscreen->resource_get_param = iris_resource_get_param;
2064 pscreen->resource_destroy = u_transfer_helper_resource_destroy;
2065 pscreen->transfer_helper =
2066 u_transfer_helper_create(&transfer_vtbl, true, true, false, true);
2067 }
2068
2069 void
2070 iris_init_resource_functions(struct pipe_context *ctx)
2071 {
2072 ctx->flush_resource = iris_flush_resource;
2073 ctx->invalidate_resource = iris_invalidate_resource;
2074 ctx->transfer_map = u_transfer_helper_transfer_map;
2075 ctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
2076 ctx->transfer_unmap = u_transfer_helper_transfer_unmap;
2077 ctx->buffer_subdata = u_default_buffer_subdata;
2078 ctx->texture_subdata = u_default_texture_subdata;
2079 }