iris: Bail resource creation upon aux creation error
[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.clear_color_bo);
291 free(res->aux.state);
292
293 res->aux.usage = ISL_AUX_USAGE_NONE;
294 res->aux.possible_usages = 1 << ISL_AUX_USAGE_NONE;
295 res->aux.sampler_usages = 1 << ISL_AUX_USAGE_NONE;
296 res->aux.has_hiz = 0;
297 res->aux.surf.size_B = 0;
298 res->aux.bo = NULL;
299 res->aux.extra_aux.surf.size_B = 0;
300 res->aux.clear_color_bo = NULL;
301 res->aux.state = NULL;
302 }
303
304 static void
305 iris_resource_destroy(struct pipe_screen *screen,
306 struct pipe_resource *resource)
307 {
308 struct iris_resource *res = (struct iris_resource *)resource;
309
310 if (resource->target == PIPE_BUFFER)
311 util_range_destroy(&res->valid_buffer_range);
312
313 iris_resource_disable_aux(res);
314
315 iris_bo_unreference(res->bo);
316 free(res);
317 }
318
319 static struct iris_resource *
320 iris_alloc_resource(struct pipe_screen *pscreen,
321 const struct pipe_resource *templ)
322 {
323 struct iris_resource *res = calloc(1, sizeof(struct iris_resource));
324 if (!res)
325 return NULL;
326
327 res->base = *templ;
328 res->base.screen = pscreen;
329 pipe_reference_init(&res->base.reference, 1);
330
331 res->aux.possible_usages = 1 << ISL_AUX_USAGE_NONE;
332 res->aux.sampler_usages = 1 << ISL_AUX_USAGE_NONE;
333
334 if (templ->target == PIPE_BUFFER)
335 util_range_init(&res->valid_buffer_range);
336
337 return res;
338 }
339
340 unsigned
341 iris_get_num_logical_layers(const struct iris_resource *res, unsigned level)
342 {
343 if (res->surf.dim == ISL_SURF_DIM_3D)
344 return minify(res->surf.logical_level0_px.depth, level);
345 else
346 return res->surf.logical_level0_px.array_len;
347 }
348
349 static enum isl_aux_state **
350 create_aux_state_map(struct iris_resource *res, enum isl_aux_state initial)
351 {
352 uint32_t total_slices = 0;
353 for (uint32_t level = 0; level < res->surf.levels; level++)
354 total_slices += iris_get_num_logical_layers(res, level);
355
356 const size_t per_level_array_size =
357 res->surf.levels * sizeof(enum isl_aux_state *);
358
359 /* We're going to allocate a single chunk of data for both the per-level
360 * reference array and the arrays of aux_state. This makes cleanup
361 * significantly easier.
362 */
363 const size_t total_size =
364 per_level_array_size + total_slices * sizeof(enum isl_aux_state);
365
366 void *data = malloc(total_size);
367 if (!data)
368 return NULL;
369
370 enum isl_aux_state **per_level_arr = data;
371 enum isl_aux_state *s = data + per_level_array_size;
372 for (uint32_t level = 0; level < res->surf.levels; level++) {
373 per_level_arr[level] = s;
374 const unsigned level_layers = iris_get_num_logical_layers(res, level);
375 for (uint32_t a = 0; a < level_layers; a++)
376 *(s++) = initial;
377 }
378 assert((void *)s == data + total_size);
379
380 return per_level_arr;
381 }
382
383 static unsigned
384 iris_get_aux_clear_color_state_size(struct iris_screen *screen)
385 {
386 const struct gen_device_info *devinfo = &screen->devinfo;
387 return devinfo->gen >= 10 ? screen->isl_dev.ss.clear_color_state_size : 0;
388 }
389
390 static void
391 map_aux_addresses(struct iris_screen *screen, struct iris_resource *res)
392 {
393 const struct gen_device_info *devinfo = &screen->devinfo;
394 if (devinfo->gen >= 12 && isl_aux_usage_has_ccs(res->aux.usage)) {
395 void *aux_map_ctx = iris_bufmgr_get_aux_map_context(screen->bufmgr);
396 assert(aux_map_ctx);
397 const unsigned aux_offset = res->aux.extra_aux.surf.size_B > 0 ?
398 res->aux.extra_aux.offset : res->aux.offset;
399 gen_aux_map_add_image(aux_map_ctx, &res->surf, res->bo->gtt_offset,
400 res->aux.bo->gtt_offset + aux_offset);
401 res->bo->aux_map_address = res->aux.bo->gtt_offset;
402 }
403 }
404
405 static bool
406 want_ccs_e_for_format(const struct gen_device_info *devinfo,
407 enum isl_format format)
408 {
409 if (!isl_format_supports_ccs_e(devinfo, format))
410 return false;
411
412 const struct isl_format_layout *fmtl = isl_format_get_layout(format);
413
414 /* CCS_E seems to significantly hurt performance with 32-bit floating
415 * point formats. For example, Paraview's "Wavelet Volume" case uses
416 * both R32_FLOAT and R32G32B32A32_FLOAT, and enabling CCS_E for those
417 * formats causes a 62% FPS drop.
418 *
419 * However, many benchmarks seem to use 16-bit float with no issues.
420 */
421 if (fmtl->channels.r.bits == 32 && fmtl->channels.r.type == ISL_SFLOAT)
422 return false;
423
424 return true;
425 }
426
427 /**
428 * Configure aux for the resource, but don't allocate it. For images which
429 * might be shared with modifiers, we must allocate the image and aux data in
430 * a single bo.
431 *
432 * Returns false on unexpected error (e.g. allocation failed, or invalid
433 * configuration result).
434 */
435 static bool
436 iris_resource_configure_aux(struct iris_screen *screen,
437 struct iris_resource *res, bool imported,
438 uint64_t *aux_size_B,
439 uint32_t *alloc_flags)
440 {
441 const struct gen_device_info *devinfo = &screen->devinfo;
442
443 /* Try to create the auxiliary surfaces allowed by the modifier or by
444 * the user if no modifier is specified.
445 */
446 assert(!res->mod_info || res->mod_info->aux_usage == ISL_AUX_USAGE_NONE ||
447 res->mod_info->aux_usage == ISL_AUX_USAGE_CCS_E);
448
449 const bool has_mcs = !res->mod_info &&
450 isl_surf_get_mcs_surf(&screen->isl_dev, &res->surf, &res->aux.surf);
451
452 const bool has_hiz = !res->mod_info && !(INTEL_DEBUG & DEBUG_NO_HIZ) &&
453 isl_surf_get_hiz_surf(&screen->isl_dev, &res->surf, &res->aux.surf);
454
455 const bool has_ccs =
456 ((!res->mod_info && !(INTEL_DEBUG & DEBUG_NO_RBC)) ||
457 (res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE)) &&
458 isl_surf_get_ccs_surf(&screen->isl_dev, &res->surf, &res->aux.surf,
459 &res->aux.extra_aux.surf, 0);
460
461 /* Having both HIZ and MCS is impossible. */
462 assert(!has_mcs || !has_hiz);
463
464 /* Ensure aux surface creation for MCS_CCS and HIZ_CCS is correct. */
465 if (has_ccs && (has_mcs || has_hiz)) {
466 assert(res->aux.extra_aux.surf.size_B > 0 &&
467 res->aux.extra_aux.surf.usage & ISL_SURF_USAGE_CCS_BIT);
468 assert(res->aux.surf.size_B > 0 &&
469 res->aux.surf.usage &
470 (ISL_SURF_USAGE_HIZ_BIT | ISL_SURF_USAGE_MCS_BIT));
471 }
472
473 if (res->mod_info && has_ccs) {
474 /* Only allow a CCS modifier if the aux was created successfully. */
475 res->aux.possible_usages |= 1 << res->mod_info->aux_usage;
476 } else if (has_mcs) {
477 res->aux.possible_usages |=
478 1 << (has_ccs ? ISL_AUX_USAGE_MCS_CCS : ISL_AUX_USAGE_MCS);
479 } else if (has_hiz) {
480 res->aux.possible_usages |=
481 1 << (has_ccs ? ISL_AUX_USAGE_HIZ_CCS : 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 /* We don't always support sampling with HIZ_CCS. But when we do, treat it
501 * as CCS_E.*/
502 res->aux.sampler_usages &= ~(1 << ISL_AUX_USAGE_HIZ_CCS);
503 if (isl_surf_supports_hiz_ccs_wt(devinfo, &res->surf, res->aux.usage))
504 res->aux.sampler_usages |= 1 << ISL_AUX_USAGE_CCS_E;
505
506 enum isl_aux_state initial_state;
507 *aux_size_B = 0;
508 *alloc_flags = 0;
509 assert(!res->aux.bo);
510
511 switch (res->aux.usage) {
512 case ISL_AUX_USAGE_NONE:
513 /* Having no aux buffer is only okay if there's no modifier with aux. */
514 return !res->mod_info || res->mod_info->aux_usage == ISL_AUX_USAGE_NONE;
515 case ISL_AUX_USAGE_HIZ:
516 case ISL_AUX_USAGE_HIZ_CCS:
517 initial_state = ISL_AUX_STATE_AUX_INVALID;
518 break;
519 case ISL_AUX_USAGE_MCS:
520 case ISL_AUX_USAGE_MCS_CCS:
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 * Returns false on unexpected error (e.g. mapping a BO failed).
602 */
603 static bool
604 iris_resource_init_aux_buf(struct iris_resource *res, uint32_t alloc_flags,
605 unsigned clear_color_state_size)
606 {
607 if (!(alloc_flags & BO_ALLOC_ZEROED)) {
608 void *map = iris_bo_map(NULL, res->aux.bo, MAP_WRITE | MAP_RAW);
609
610 if (!map)
611 return false;
612
613 if (iris_resource_get_aux_state(res, 0, 0) != ISL_AUX_STATE_AUX_INVALID) {
614 uint8_t memset_value = isl_aux_usage_has_mcs(res->aux.usage) ? 0xFF : 0;
615 memset((char*)map + res->aux.offset, memset_value,
616 res->aux.surf.size_B);
617 }
618
619 /* Bspec section titled : MCS/CCS Buffers for Render Target(s) states:
620 * - If Software wants to enable Color Compression without Fast clear,
621 * Software needs to initialize MCS with zeros.
622 * - Lossless compression and CCS initialized to all F (using HW Fast
623 * Clear or SW direct Clear)
624 *
625 * We think, the first bullet point above is referring to CCS aux
626 * surface. Since we initialize the MCS in the clear state, we also
627 * initialize the CCS in the clear state (via SW direct clear) to keep
628 * the two in sync.
629 */
630 memset((char*)map + res->aux.extra_aux.offset,
631 isl_aux_usage_has_mcs(res->aux.usage) ? 0xFF : 0,
632 res->aux.extra_aux.surf.size_B);
633
634 /* Zero the indirect clear color to match ::fast_clear_color. */
635 memset((char *)map + res->aux.clear_color_offset, 0,
636 clear_color_state_size);
637
638 iris_bo_unmap(res->aux.bo);
639 }
640
641 if (clear_color_state_size > 0) {
642 res->aux.clear_color_bo = res->aux.bo;
643 iris_bo_reference(res->aux.clear_color_bo);
644 }
645
646 return true;
647 }
648
649 /**
650 * Allocate the initial aux surface for a resource based on aux.usage
651 *
652 * Returns false on unexpected error (e.g. allocation failed, or invalid
653 * configuration result).
654 */
655 static bool
656 iris_resource_alloc_separate_aux(struct iris_screen *screen,
657 struct iris_resource *res)
658 {
659 uint32_t alloc_flags;
660 uint64_t size;
661 if (!iris_resource_configure_aux(screen, res, false, &size, &alloc_flags))
662 return false;
663
664 if (size == 0)
665 return true;
666
667 /* Allocate the auxiliary buffer. ISL has stricter set of alignment rules
668 * the drm allocator. Therefore, one can pass the ISL dimensions in terms
669 * of bytes instead of trying to recalculate based on different format
670 * block sizes.
671 */
672 res->aux.bo = iris_bo_alloc_tiled(screen->bufmgr, "aux buffer", size, 4096,
673 IRIS_MEMZONE_OTHER,
674 isl_tiling_to_i915_tiling(res->aux.surf.tiling),
675 res->aux.surf.row_pitch_B, alloc_flags);
676 if (!res->aux.bo) {
677 return false;
678 }
679
680 if (!iris_resource_init_aux_buf(res, alloc_flags,
681 iris_get_aux_clear_color_state_size(screen)))
682 return false;
683
684 map_aux_addresses(screen, res);
685
686 return true;
687 }
688
689 void
690 iris_resource_finish_aux_import(struct pipe_screen *pscreen,
691 struct iris_resource *res)
692 {
693 struct iris_screen *screen = (struct iris_screen *)pscreen;
694 assert(iris_resource_unfinished_aux_import(res));
695 assert(!res->mod_info->supports_clear_color);
696
697 struct iris_resource *aux_res = (void *) res->base.next;
698 assert(aux_res->aux.surf.row_pitch_B && aux_res->aux.offset &&
699 aux_res->aux.bo);
700
701 assert(res->bo == aux_res->aux.bo);
702 iris_bo_reference(aux_res->aux.bo);
703 res->aux.bo = aux_res->aux.bo;
704
705 res->aux.offset = aux_res->aux.offset;
706
707 assert(res->bo->size >= (res->aux.offset + res->aux.surf.size_B));
708 assert(res->aux.clear_color_bo == NULL);
709 res->aux.clear_color_offset = 0;
710
711 assert(aux_res->aux.surf.row_pitch_B == res->aux.surf.row_pitch_B);
712
713 unsigned clear_color_state_size =
714 iris_get_aux_clear_color_state_size(screen);
715
716 if (clear_color_state_size > 0) {
717 res->aux.clear_color_bo =
718 iris_bo_alloc(screen->bufmgr, "clear color buffer",
719 clear_color_state_size, IRIS_MEMZONE_OTHER);
720 res->aux.clear_color_offset = 0;
721 }
722
723 iris_resource_destroy(&screen->base, res->base.next);
724 res->base.next = NULL;
725 }
726
727 static struct pipe_resource *
728 iris_resource_create_for_buffer(struct pipe_screen *pscreen,
729 const struct pipe_resource *templ)
730 {
731 struct iris_screen *screen = (struct iris_screen *)pscreen;
732 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
733
734 assert(templ->target == PIPE_BUFFER);
735 assert(templ->height0 <= 1);
736 assert(templ->depth0 <= 1);
737 assert(templ->format == PIPE_FORMAT_NONE ||
738 util_format_get_blocksize(templ->format) == 1);
739
740 res->internal_format = templ->format;
741 res->surf.tiling = ISL_TILING_LINEAR;
742
743 enum iris_memory_zone memzone = IRIS_MEMZONE_OTHER;
744 const char *name = templ->target == PIPE_BUFFER ? "buffer" : "miptree";
745 if (templ->flags & IRIS_RESOURCE_FLAG_SHADER_MEMZONE) {
746 memzone = IRIS_MEMZONE_SHADER;
747 name = "shader kernels";
748 } else if (templ->flags & IRIS_RESOURCE_FLAG_SURFACE_MEMZONE) {
749 memzone = IRIS_MEMZONE_SURFACE;
750 name = "surface state";
751 } else if (templ->flags & IRIS_RESOURCE_FLAG_DYNAMIC_MEMZONE) {
752 memzone = IRIS_MEMZONE_DYNAMIC;
753 name = "dynamic state";
754 }
755
756 res->bo = iris_bo_alloc(screen->bufmgr, name, templ->width0, memzone);
757 if (!res->bo) {
758 iris_resource_destroy(pscreen, &res->base);
759 return NULL;
760 }
761
762 return &res->base;
763 }
764
765 static struct pipe_resource *
766 iris_resource_create_with_modifiers(struct pipe_screen *pscreen,
767 const struct pipe_resource *templ,
768 const uint64_t *modifiers,
769 int modifiers_count)
770 {
771 struct iris_screen *screen = (struct iris_screen *)pscreen;
772 struct gen_device_info *devinfo = &screen->devinfo;
773 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
774
775 if (!res)
776 return NULL;
777
778 const struct util_format_description *format_desc =
779 util_format_description(templ->format);
780 const bool has_depth = util_format_has_depth(format_desc);
781 uint64_t modifier =
782 select_best_modifier(devinfo, templ->format, modifiers, modifiers_count);
783
784 isl_tiling_flags_t tiling_flags = ISL_TILING_ANY_MASK;
785
786 if (modifier != DRM_FORMAT_MOD_INVALID) {
787 res->mod_info = isl_drm_modifier_get_info(modifier);
788
789 tiling_flags = 1 << res->mod_info->tiling;
790 } else {
791 if (modifiers_count > 0) {
792 fprintf(stderr, "Unsupported modifier, resource creation failed.\n");
793 goto fail;
794 }
795
796 /* Use linear for staging buffers */
797 if (templ->usage == PIPE_USAGE_STAGING ||
798 templ->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR) )
799 tiling_flags = ISL_TILING_LINEAR_BIT;
800 }
801
802 isl_surf_usage_flags_t usage = pipe_bind_to_isl_usage(templ->bind);
803
804 if (templ->target == PIPE_TEXTURE_CUBE ||
805 templ->target == PIPE_TEXTURE_CUBE_ARRAY)
806 usage |= ISL_SURF_USAGE_CUBE_BIT;
807
808 if (templ->usage != PIPE_USAGE_STAGING) {
809 if (templ->format == PIPE_FORMAT_S8_UINT)
810 usage |= ISL_SURF_USAGE_STENCIL_BIT;
811 else if (has_depth)
812 usage |= ISL_SURF_USAGE_DEPTH_BIT;
813 }
814
815 enum pipe_format pfmt = templ->format;
816 res->internal_format = pfmt;
817
818 /* Should be handled by u_transfer_helper */
819 assert(!util_format_is_depth_and_stencil(pfmt));
820
821 struct iris_format_info fmt = iris_format_for_usage(devinfo, pfmt, usage);
822 assert(fmt.fmt != ISL_FORMAT_UNSUPPORTED);
823
824 UNUSED const bool isl_surf_created_successfully =
825 isl_surf_init(&screen->isl_dev, &res->surf,
826 .dim = target_to_isl_surf_dim(templ->target),
827 .format = fmt.fmt,
828 .width = templ->width0,
829 .height = templ->height0,
830 .depth = templ->depth0,
831 .levels = templ->last_level + 1,
832 .array_len = templ->array_size,
833 .samples = MAX2(templ->nr_samples, 1),
834 .min_alignment_B = 0,
835 .row_pitch_B = 0,
836 .usage = usage,
837 .tiling_flags = tiling_flags);
838 assert(isl_surf_created_successfully);
839
840 const char *name = "miptree";
841 enum iris_memory_zone memzone = IRIS_MEMZONE_OTHER;
842
843 unsigned int flags = 0;
844 if (templ->usage == PIPE_USAGE_STAGING)
845 flags |= BO_ALLOC_COHERENT;
846
847 /* These are for u_upload_mgr buffers only */
848 assert(!(templ->flags & (IRIS_RESOURCE_FLAG_SHADER_MEMZONE |
849 IRIS_RESOURCE_FLAG_SURFACE_MEMZONE |
850 IRIS_RESOURCE_FLAG_DYNAMIC_MEMZONE)));
851
852 uint32_t aux_preferred_alloc_flags;
853 uint64_t aux_size = 0;
854 if (!iris_resource_configure_aux(screen, res, false, &aux_size,
855 &aux_preferred_alloc_flags)) {
856 goto fail;
857 }
858
859 const bool aux_enabled = res->aux.surf.size_B > 0;
860 const bool separate_aux = aux_enabled && !res->mod_info;
861 uint64_t aux_offset;
862 uint64_t bo_size;
863
864 if (aux_enabled && !separate_aux) {
865 /* Allocate aux data with main surface. This is required for modifiers
866 * with aux data (ccs).
867 */
868 aux_offset = ALIGN(res->surf.size_B, res->aux.surf.alignment_B);
869 bo_size = aux_offset + aux_size;
870 } else {
871 aux_offset = 0;
872 bo_size = res->surf.size_B;
873 }
874
875 uint32_t alignment = MAX2(4096, res->surf.alignment_B);
876 res->bo = iris_bo_alloc_tiled(screen->bufmgr, name, bo_size, alignment,
877 memzone,
878 isl_tiling_to_i915_tiling(res->surf.tiling),
879 res->surf.row_pitch_B, flags);
880
881 if (!res->bo)
882 goto fail;
883
884 if (aux_enabled) {
885 if (separate_aux) {
886 if (!iris_resource_alloc_separate_aux(screen, res))
887 goto fail;
888 } else {
889 res->aux.bo = res->bo;
890 iris_bo_reference(res->aux.bo);
891 res->aux.offset += aux_offset;
892 unsigned clear_color_state_size =
893 iris_get_aux_clear_color_state_size(screen);
894 if (clear_color_state_size > 0)
895 res->aux.clear_color_offset += aux_offset;
896 if (!iris_resource_init_aux_buf(res, flags, clear_color_state_size))
897 goto fail;
898 map_aux_addresses(screen, res);
899 }
900 }
901
902 return &res->base;
903
904 fail:
905 fprintf(stderr, "XXX: resource creation failed\n");
906 iris_resource_destroy(pscreen, &res->base);
907 return NULL;
908
909 }
910
911 static struct pipe_resource *
912 iris_resource_create(struct pipe_screen *pscreen,
913 const struct pipe_resource *templ)
914 {
915 if (templ->target == PIPE_BUFFER)
916 return iris_resource_create_for_buffer(pscreen, templ);
917 else
918 return iris_resource_create_with_modifiers(pscreen, templ, NULL, 0);
919 }
920
921 static uint64_t
922 tiling_to_modifier(uint32_t tiling)
923 {
924 static const uint64_t map[] = {
925 [I915_TILING_NONE] = DRM_FORMAT_MOD_LINEAR,
926 [I915_TILING_X] = I915_FORMAT_MOD_X_TILED,
927 [I915_TILING_Y] = I915_FORMAT_MOD_Y_TILED,
928 };
929
930 assert(tiling < ARRAY_SIZE(map));
931
932 return map[tiling];
933 }
934
935 static struct pipe_resource *
936 iris_resource_from_user_memory(struct pipe_screen *pscreen,
937 const struct pipe_resource *templ,
938 void *user_memory)
939 {
940 struct iris_screen *screen = (struct iris_screen *)pscreen;
941 struct iris_bufmgr *bufmgr = screen->bufmgr;
942 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
943 if (!res)
944 return NULL;
945
946 assert(templ->target == PIPE_BUFFER);
947
948 res->internal_format = templ->format;
949 res->bo = iris_bo_create_userptr(bufmgr, "user",
950 user_memory, templ->width0,
951 IRIS_MEMZONE_OTHER);
952 if (!res->bo) {
953 free(res);
954 return NULL;
955 }
956
957 util_range_add(&res->base, &res->valid_buffer_range, 0, templ->width0);
958
959 return &res->base;
960 }
961
962 static struct pipe_resource *
963 iris_resource_from_handle(struct pipe_screen *pscreen,
964 const struct pipe_resource *templ,
965 struct winsys_handle *whandle,
966 unsigned usage)
967 {
968 struct iris_screen *screen = (struct iris_screen *)pscreen;
969 struct gen_device_info *devinfo = &screen->devinfo;
970 struct iris_bufmgr *bufmgr = screen->bufmgr;
971 struct iris_resource *res = iris_alloc_resource(pscreen, templ);
972 if (!res)
973 return NULL;
974
975 switch (whandle->type) {
976 case WINSYS_HANDLE_TYPE_FD:
977 res->bo = iris_bo_import_dmabuf(bufmgr, whandle->handle);
978 break;
979 case WINSYS_HANDLE_TYPE_SHARED:
980 res->bo = iris_bo_gem_create_from_name(bufmgr, "winsys image",
981 whandle->handle);
982 break;
983 default:
984 unreachable("invalid winsys handle type");
985 }
986 if (!res->bo)
987 return NULL;
988
989 res->offset = whandle->offset;
990
991 uint64_t modifier = whandle->modifier;
992 if (modifier == DRM_FORMAT_MOD_INVALID) {
993 modifier = tiling_to_modifier(res->bo->tiling_mode);
994 }
995 res->mod_info = isl_drm_modifier_get_info(modifier);
996 assert(res->mod_info);
997
998 isl_surf_usage_flags_t isl_usage = pipe_bind_to_isl_usage(templ->bind);
999
1000 const struct iris_format_info fmt =
1001 iris_format_for_usage(devinfo, templ->format, isl_usage);
1002 res->internal_format = templ->format;
1003
1004 if (templ->target == PIPE_BUFFER) {
1005 res->surf.tiling = ISL_TILING_LINEAR;
1006 } else {
1007 if (whandle->modifier == DRM_FORMAT_MOD_INVALID || whandle->plane == 0) {
1008 UNUSED const bool isl_surf_created_successfully =
1009 isl_surf_init(&screen->isl_dev, &res->surf,
1010 .dim = target_to_isl_surf_dim(templ->target),
1011 .format = fmt.fmt,
1012 .width = templ->width0,
1013 .height = templ->height0,
1014 .depth = templ->depth0,
1015 .levels = templ->last_level + 1,
1016 .array_len = templ->array_size,
1017 .samples = MAX2(templ->nr_samples, 1),
1018 .min_alignment_B = 0,
1019 .row_pitch_B = whandle->stride,
1020 .usage = isl_usage,
1021 .tiling_flags = 1 << res->mod_info->tiling);
1022 assert(isl_surf_created_successfully);
1023 assert(res->bo->tiling_mode ==
1024 isl_tiling_to_i915_tiling(res->surf.tiling));
1025
1026 // XXX: create_ccs_buf_for_image?
1027 if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
1028 if (!iris_resource_alloc_separate_aux(screen, res))
1029 goto fail;
1030 } else {
1031 if (res->mod_info->aux_usage != ISL_AUX_USAGE_NONE) {
1032 uint32_t alloc_flags;
1033 uint64_t size;
1034 bool ok = iris_resource_configure_aux(screen, res, true, &size,
1035 &alloc_flags);
1036 assert(ok);
1037 /* The gallium dri layer will create a separate plane resource
1038 * for the aux image. iris_resource_finish_aux_import will
1039 * merge the separate aux parameters back into a single
1040 * iris_resource.
1041 */
1042 }
1043 }
1044 } else {
1045 /* Save modifier import information to reconstruct later. After
1046 * import, this will be available under a second image accessible
1047 * from the main image with res->base.next. See
1048 * iris_resource_finish_aux_import.
1049 */
1050 res->aux.surf.row_pitch_B = whandle->stride;
1051 res->aux.offset = whandle->offset;
1052 res->aux.bo = res->bo;
1053 res->bo = NULL;
1054 }
1055 }
1056
1057 return &res->base;
1058
1059 fail:
1060 iris_resource_destroy(pscreen, &res->base);
1061 return NULL;
1062 }
1063
1064 static void
1065 iris_flush_resource(struct pipe_context *ctx, struct pipe_resource *resource)
1066 {
1067 struct iris_context *ice = (struct iris_context *)ctx;
1068 struct iris_batch *render_batch = &ice->batches[IRIS_BATCH_RENDER];
1069 struct iris_resource *res = (void *) resource;
1070 const struct isl_drm_modifier_info *mod = res->mod_info;
1071
1072 iris_resource_prepare_access(ice, render_batch, res,
1073 0, INTEL_REMAINING_LEVELS,
1074 0, INTEL_REMAINING_LAYERS,
1075 mod ? mod->aux_usage : ISL_AUX_USAGE_NONE,
1076 mod ? mod->supports_clear_color : false);
1077 }
1078
1079 static void
1080 iris_resource_disable_aux_on_first_query(struct pipe_resource *resource,
1081 unsigned usage)
1082 {
1083 struct iris_resource *res = (struct iris_resource *)resource;
1084 bool mod_with_aux =
1085 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1086
1087 /* Disable aux usage if explicit flush not set and this is the first time
1088 * we are dealing with this resource and the resource was not created with
1089 * a modifier with aux.
1090 */
1091 if (!mod_with_aux &&
1092 (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) && res->aux.usage != 0) &&
1093 p_atomic_read(&resource->reference.count) == 1) {
1094 iris_resource_disable_aux(res);
1095 }
1096 }
1097
1098 static bool
1099 iris_resource_get_param(struct pipe_screen *screen,
1100 struct pipe_context *context,
1101 struct pipe_resource *resource,
1102 unsigned plane,
1103 unsigned layer,
1104 enum pipe_resource_param param,
1105 unsigned handle_usage,
1106 uint64_t *value)
1107 {
1108 struct iris_resource *res = (struct iris_resource *)resource;
1109 bool mod_with_aux =
1110 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1111 bool wants_aux = mod_with_aux && plane > 0;
1112 bool result;
1113 unsigned handle;
1114
1115 if (iris_resource_unfinished_aux_import(res))
1116 iris_resource_finish_aux_import(screen, res);
1117
1118 struct iris_bo *bo = wants_aux ? res->aux.bo : res->bo;
1119
1120 iris_resource_disable_aux_on_first_query(resource, handle_usage);
1121
1122 switch (param) {
1123 case PIPE_RESOURCE_PARAM_NPLANES:
1124 if (mod_with_aux) {
1125 *value = 2;
1126 } else {
1127 unsigned count = 0;
1128 for (struct pipe_resource *cur = resource; cur; cur = cur->next)
1129 count++;
1130 *value = count;
1131 }
1132 return true;
1133 case PIPE_RESOURCE_PARAM_STRIDE:
1134 *value = wants_aux ? res->aux.surf.row_pitch_B : res->surf.row_pitch_B;
1135 return true;
1136 case PIPE_RESOURCE_PARAM_OFFSET:
1137 *value = wants_aux ? res->aux.offset : 0;
1138 return true;
1139 case PIPE_RESOURCE_PARAM_MODIFIER:
1140 *value = res->mod_info ? res->mod_info->modifier :
1141 tiling_to_modifier(res->bo->tiling_mode);
1142 return true;
1143 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_SHARED:
1144 result = iris_bo_flink(bo, &handle) == 0;
1145 if (result)
1146 *value = handle;
1147 return result;
1148 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_KMS:
1149 *value = iris_bo_export_gem_handle(bo);
1150 return true;
1151 case PIPE_RESOURCE_PARAM_HANDLE_TYPE_FD:
1152 result = iris_bo_export_dmabuf(bo, (int *) &handle) == 0;
1153 if (result)
1154 *value = handle;
1155 return result;
1156 default:
1157 return false;
1158 }
1159 }
1160
1161 static bool
1162 iris_resource_get_handle(struct pipe_screen *pscreen,
1163 struct pipe_context *ctx,
1164 struct pipe_resource *resource,
1165 struct winsys_handle *whandle,
1166 unsigned usage)
1167 {
1168 struct iris_resource *res = (struct iris_resource *)resource;
1169 bool mod_with_aux =
1170 res->mod_info && res->mod_info->aux_usage != ISL_AUX_USAGE_NONE;
1171
1172 iris_resource_disable_aux_on_first_query(resource, usage);
1173
1174 struct iris_bo *bo;
1175 if (mod_with_aux && whandle->plane > 0) {
1176 assert(res->aux.bo);
1177 bo = res->aux.bo;
1178 whandle->stride = res->aux.surf.row_pitch_B;
1179 whandle->offset = res->aux.offset;
1180 } else {
1181 /* If this is a buffer, stride should be 0 - no need to special case */
1182 whandle->stride = res->surf.row_pitch_B;
1183 bo = res->bo;
1184 }
1185 whandle->modifier =
1186 res->mod_info ? res->mod_info->modifier
1187 : tiling_to_modifier(res->bo->tiling_mode);
1188
1189 #ifndef NDEBUG
1190 enum isl_aux_usage allowed_usage =
1191 res->mod_info ? res->mod_info->aux_usage : ISL_AUX_USAGE_NONE;
1192
1193 if (res->aux.usage != allowed_usage) {
1194 enum isl_aux_state aux_state = iris_resource_get_aux_state(res, 0, 0);
1195 assert(aux_state == ISL_AUX_STATE_RESOLVED ||
1196 aux_state == ISL_AUX_STATE_PASS_THROUGH);
1197 }
1198 #endif
1199
1200 switch (whandle->type) {
1201 case WINSYS_HANDLE_TYPE_SHARED:
1202 return iris_bo_flink(bo, &whandle->handle) == 0;
1203 case WINSYS_HANDLE_TYPE_KMS:
1204 whandle->handle = iris_bo_export_gem_handle(bo);
1205 return true;
1206 case WINSYS_HANDLE_TYPE_FD:
1207 return iris_bo_export_dmabuf(bo, (int *) &whandle->handle) == 0;
1208 }
1209
1210 return false;
1211 }
1212
1213 static bool
1214 resource_is_busy(struct iris_context *ice,
1215 struct iris_resource *res)
1216 {
1217 bool busy = iris_bo_busy(res->bo);
1218
1219 for (int i = 0; i < IRIS_BATCH_COUNT; i++)
1220 busy |= iris_batch_references(&ice->batches[i], res->bo);
1221
1222 return busy;
1223 }
1224
1225 static void
1226 iris_invalidate_resource(struct pipe_context *ctx,
1227 struct pipe_resource *resource)
1228 {
1229 struct iris_screen *screen = (void *) ctx->screen;
1230 struct iris_context *ice = (void *) ctx;
1231 struct iris_resource *res = (void *) resource;
1232
1233 if (resource->target != PIPE_BUFFER)
1234 return;
1235
1236 if (!resource_is_busy(ice, res)) {
1237 /* The resource is idle, so just mark that it contains no data and
1238 * keep using the same underlying buffer object.
1239 */
1240 util_range_set_empty(&res->valid_buffer_range);
1241 return;
1242 }
1243
1244 /* Otherwise, try and replace the backing storage with a new BO. */
1245
1246 /* We can't reallocate memory we didn't allocate in the first place. */
1247 if (res->bo->userptr)
1248 return;
1249
1250 // XXX: We should support this.
1251 if (res->bind_history & PIPE_BIND_STREAM_OUTPUT)
1252 return;
1253
1254 struct iris_bo *old_bo = res->bo;
1255 struct iris_bo *new_bo =
1256 iris_bo_alloc(screen->bufmgr, res->bo->name, resource->width0,
1257 iris_memzone_for_address(old_bo->gtt_offset));
1258 if (!new_bo)
1259 return;
1260
1261 /* Swap out the backing storage */
1262 res->bo = new_bo;
1263
1264 /* Rebind the buffer, replacing any state referring to the old BO's
1265 * address, and marking state dirty so it's reemitted.
1266 */
1267 ice->vtbl.rebind_buffer(ice, res, old_bo->gtt_offset);
1268
1269 util_range_set_empty(&res->valid_buffer_range);
1270
1271 iris_bo_unreference(old_bo);
1272 }
1273
1274 static void
1275 iris_flush_staging_region(struct pipe_transfer *xfer,
1276 const struct pipe_box *flush_box)
1277 {
1278 if (!(xfer->usage & PIPE_TRANSFER_WRITE))
1279 return;
1280
1281 struct iris_transfer *map = (void *) xfer;
1282
1283 struct pipe_box src_box = *flush_box;
1284
1285 /* Account for extra alignment padding in staging buffer */
1286 if (xfer->resource->target == PIPE_BUFFER)
1287 src_box.x += xfer->box.x % IRIS_MAP_BUFFER_ALIGNMENT;
1288
1289 struct pipe_box dst_box = (struct pipe_box) {
1290 .x = xfer->box.x + flush_box->x,
1291 .y = xfer->box.y + flush_box->y,
1292 .z = xfer->box.z + flush_box->z,
1293 .width = flush_box->width,
1294 .height = flush_box->height,
1295 .depth = flush_box->depth,
1296 };
1297
1298 iris_copy_region(map->blorp, map->batch, xfer->resource, xfer->level,
1299 dst_box.x, dst_box.y, dst_box.z, map->staging, 0,
1300 &src_box);
1301 }
1302
1303 static void
1304 iris_unmap_copy_region(struct iris_transfer *map)
1305 {
1306 iris_resource_destroy(map->staging->screen, map->staging);
1307
1308 map->ptr = NULL;
1309 }
1310
1311 static void
1312 iris_map_copy_region(struct iris_transfer *map)
1313 {
1314 struct pipe_screen *pscreen = &map->batch->screen->base;
1315 struct pipe_transfer *xfer = &map->base;
1316 struct pipe_box *box = &xfer->box;
1317 struct iris_resource *res = (void *) xfer->resource;
1318
1319 unsigned extra = xfer->resource->target == PIPE_BUFFER ?
1320 box->x % IRIS_MAP_BUFFER_ALIGNMENT : 0;
1321
1322 struct pipe_resource templ = (struct pipe_resource) {
1323 .usage = PIPE_USAGE_STAGING,
1324 .width0 = box->width + extra,
1325 .height0 = box->height,
1326 .depth0 = 1,
1327 .nr_samples = xfer->resource->nr_samples,
1328 .nr_storage_samples = xfer->resource->nr_storage_samples,
1329 .array_size = box->depth,
1330 .format = res->internal_format,
1331 };
1332
1333 if (xfer->resource->target == PIPE_BUFFER)
1334 templ.target = PIPE_BUFFER;
1335 else if (templ.array_size > 1)
1336 templ.target = PIPE_TEXTURE_2D_ARRAY;
1337 else
1338 templ.target = PIPE_TEXTURE_2D;
1339
1340 map->staging = iris_resource_create(pscreen, &templ);
1341 assert(map->staging);
1342
1343 if (templ.target != PIPE_BUFFER) {
1344 struct isl_surf *surf = &((struct iris_resource *) map->staging)->surf;
1345 xfer->stride = isl_surf_get_row_pitch_B(surf);
1346 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1347 }
1348
1349 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1350 iris_copy_region(map->blorp, map->batch, map->staging, 0, extra, 0, 0,
1351 xfer->resource, xfer->level, box);
1352 /* Ensure writes to the staging BO land before we map it below. */
1353 iris_emit_pipe_control_flush(map->batch,
1354 "transfer read: flush before mapping",
1355 PIPE_CONTROL_RENDER_TARGET_FLUSH |
1356 PIPE_CONTROL_CS_STALL);
1357 }
1358
1359 struct iris_bo *staging_bo = iris_resource_bo(map->staging);
1360
1361 if (iris_batch_references(map->batch, staging_bo))
1362 iris_batch_flush(map->batch);
1363
1364 map->ptr =
1365 iris_bo_map(map->dbg, staging_bo, xfer->usage & MAP_FLAGS) + extra;
1366
1367 map->unmap = iris_unmap_copy_region;
1368 }
1369
1370 static void
1371 get_image_offset_el(const struct isl_surf *surf, unsigned level, unsigned z,
1372 unsigned *out_x0_el, unsigned *out_y0_el)
1373 {
1374 if (surf->dim == ISL_SURF_DIM_3D) {
1375 isl_surf_get_image_offset_el(surf, level, 0, z, out_x0_el, out_y0_el);
1376 } else {
1377 isl_surf_get_image_offset_el(surf, level, z, 0, out_x0_el, out_y0_el);
1378 }
1379 }
1380
1381 /**
1382 * This function computes the tile_w (in bytes) and tile_h (in rows) of
1383 * different tiling patterns.
1384 */
1385 static void
1386 iris_resource_get_tile_dims(enum isl_tiling tiling, uint32_t cpp,
1387 uint32_t *tile_w, uint32_t *tile_h)
1388 {
1389 switch (tiling) {
1390 case ISL_TILING_X:
1391 *tile_w = 512;
1392 *tile_h = 8;
1393 break;
1394 case ISL_TILING_Y0:
1395 *tile_w = 128;
1396 *tile_h = 32;
1397 break;
1398 case ISL_TILING_LINEAR:
1399 *tile_w = cpp;
1400 *tile_h = 1;
1401 break;
1402 default:
1403 unreachable("not reached");
1404 }
1405
1406 }
1407
1408 /**
1409 * This function computes masks that may be used to select the bits of the X
1410 * and Y coordinates that indicate the offset within a tile. If the BO is
1411 * untiled, the masks are set to 0.
1412 */
1413 static void
1414 iris_resource_get_tile_masks(enum isl_tiling tiling, uint32_t cpp,
1415 uint32_t *mask_x, uint32_t *mask_y)
1416 {
1417 uint32_t tile_w_bytes, tile_h;
1418
1419 iris_resource_get_tile_dims(tiling, cpp, &tile_w_bytes, &tile_h);
1420
1421 *mask_x = tile_w_bytes / cpp - 1;
1422 *mask_y = tile_h - 1;
1423 }
1424
1425 /**
1426 * Compute the offset (in bytes) from the start of the BO to the given x
1427 * and y coordinate. For tiled BOs, caller must ensure that x and y are
1428 * multiples of the tile size.
1429 */
1430 static uint32_t
1431 iris_resource_get_aligned_offset(const struct iris_resource *res,
1432 uint32_t x, uint32_t y)
1433 {
1434 const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
1435 unsigned cpp = fmtl->bpb / 8;
1436 uint32_t pitch = res->surf.row_pitch_B;
1437
1438 switch (res->surf.tiling) {
1439 default:
1440 unreachable("not reached");
1441 case ISL_TILING_LINEAR:
1442 return y * pitch + x * cpp;
1443 case ISL_TILING_X:
1444 assert((x % (512 / cpp)) == 0);
1445 assert((y % 8) == 0);
1446 return y * pitch + x / (512 / cpp) * 4096;
1447 case ISL_TILING_Y0:
1448 assert((x % (128 / cpp)) == 0);
1449 assert((y % 32) == 0);
1450 return y * pitch + x / (128 / cpp) * 4096;
1451 }
1452 }
1453
1454 /**
1455 * Rendering with tiled buffers requires that the base address of the buffer
1456 * be aligned to a page boundary. For renderbuffers, and sometimes with
1457 * textures, we may want the surface to point at a texture image level that
1458 * isn't at a page boundary.
1459 *
1460 * This function returns an appropriately-aligned base offset
1461 * according to the tiling restrictions, plus any required x/y offset
1462 * from there.
1463 */
1464 uint32_t
1465 iris_resource_get_tile_offsets(const struct iris_resource *res,
1466 uint32_t level, uint32_t z,
1467 uint32_t *tile_x, uint32_t *tile_y)
1468 {
1469 uint32_t x, y;
1470 uint32_t mask_x, mask_y;
1471
1472 const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
1473 const unsigned cpp = fmtl->bpb / 8;
1474
1475 iris_resource_get_tile_masks(res->surf.tiling, cpp, &mask_x, &mask_y);
1476 get_image_offset_el(&res->surf, level, z, &x, &y);
1477
1478 *tile_x = x & mask_x;
1479 *tile_y = y & mask_y;
1480
1481 return iris_resource_get_aligned_offset(res, x & ~mask_x, y & ~mask_y);
1482 }
1483
1484 /**
1485 * Get pointer offset into stencil buffer.
1486 *
1487 * The stencil buffer is W tiled. Since the GTT is incapable of W fencing, we
1488 * must decode the tile's layout in software.
1489 *
1490 * See
1491 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.2.1 W-Major Tile
1492 * Format.
1493 * - PRM, 2011 Sandy Bridge, Volume 1, Part 2, Section 4.5.3 Tiling Algorithm
1494 *
1495 * Even though the returned offset is always positive, the return type is
1496 * signed due to
1497 * commit e8b1c6d6f55f5be3bef25084fdd8b6127517e137
1498 * mesa: Fix return type of _mesa_get_format_bytes() (#37351)
1499 */
1500 static intptr_t
1501 s8_offset(uint32_t stride, uint32_t x, uint32_t y)
1502 {
1503 uint32_t tile_size = 4096;
1504 uint32_t tile_width = 64;
1505 uint32_t tile_height = 64;
1506 uint32_t row_size = 64 * stride / 2; /* Two rows are interleaved. */
1507
1508 uint32_t tile_x = x / tile_width;
1509 uint32_t tile_y = y / tile_height;
1510
1511 /* The byte's address relative to the tile's base addres. */
1512 uint32_t byte_x = x % tile_width;
1513 uint32_t byte_y = y % tile_height;
1514
1515 uintptr_t u = tile_y * row_size
1516 + tile_x * tile_size
1517 + 512 * (byte_x / 8)
1518 + 64 * (byte_y / 8)
1519 + 32 * ((byte_y / 4) % 2)
1520 + 16 * ((byte_x / 4) % 2)
1521 + 8 * ((byte_y / 2) % 2)
1522 + 4 * ((byte_x / 2) % 2)
1523 + 2 * (byte_y % 2)
1524 + 1 * (byte_x % 2);
1525
1526 return u;
1527 }
1528
1529 static void
1530 iris_unmap_s8(struct iris_transfer *map)
1531 {
1532 struct pipe_transfer *xfer = &map->base;
1533 const struct pipe_box *box = &xfer->box;
1534 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1535 struct isl_surf *surf = &res->surf;
1536
1537 if (xfer->usage & PIPE_TRANSFER_WRITE) {
1538 uint8_t *untiled_s8_map = map->ptr;
1539 uint8_t *tiled_s8_map =
1540 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1541
1542 for (int s = 0; s < box->depth; s++) {
1543 unsigned x0_el, y0_el;
1544 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1545
1546 for (uint32_t y = 0; y < box->height; y++) {
1547 for (uint32_t x = 0; x < box->width; x++) {
1548 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1549 x0_el + box->x + x,
1550 y0_el + box->y + y);
1551 tiled_s8_map[offset] =
1552 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x];
1553 }
1554 }
1555 }
1556 }
1557
1558 free(map->buffer);
1559 }
1560
1561 static void
1562 iris_map_s8(struct iris_transfer *map)
1563 {
1564 struct pipe_transfer *xfer = &map->base;
1565 const struct pipe_box *box = &xfer->box;
1566 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1567 struct isl_surf *surf = &res->surf;
1568
1569 xfer->stride = surf->row_pitch_B;
1570 xfer->layer_stride = xfer->stride * box->height;
1571
1572 /* The tiling and detiling functions require that the linear buffer has
1573 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1574 * over-allocate the linear buffer to get the proper alignment.
1575 */
1576 map->buffer = map->ptr = malloc(xfer->layer_stride * box->depth);
1577 assert(map->buffer);
1578
1579 /* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
1580 * INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
1581 * invalidate is set, since we'll be writing the whole rectangle from our
1582 * temporary buffer back out.
1583 */
1584 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1585 uint8_t *untiled_s8_map = map->ptr;
1586 uint8_t *tiled_s8_map =
1587 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1588
1589 for (int s = 0; s < box->depth; s++) {
1590 unsigned x0_el, y0_el;
1591 get_image_offset_el(surf, xfer->level, box->z + s, &x0_el, &y0_el);
1592
1593 for (uint32_t y = 0; y < box->height; y++) {
1594 for (uint32_t x = 0; x < box->width; x++) {
1595 ptrdiff_t offset = s8_offset(surf->row_pitch_B,
1596 x0_el + box->x + x,
1597 y0_el + box->y + y);
1598 untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x] =
1599 tiled_s8_map[offset];
1600 }
1601 }
1602 }
1603 }
1604
1605 map->unmap = iris_unmap_s8;
1606 }
1607
1608 /* Compute extent parameters for use with tiled_memcpy functions.
1609 * xs are in units of bytes and ys are in units of strides.
1610 */
1611 static inline void
1612 tile_extents(const struct isl_surf *surf,
1613 const struct pipe_box *box,
1614 unsigned level, int z,
1615 unsigned *x1_B, unsigned *x2_B,
1616 unsigned *y1_el, unsigned *y2_el)
1617 {
1618 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1619 const unsigned cpp = fmtl->bpb / 8;
1620
1621 assert(box->x % fmtl->bw == 0);
1622 assert(box->y % fmtl->bh == 0);
1623
1624 unsigned x0_el, y0_el;
1625 get_image_offset_el(surf, level, box->z + z, &x0_el, &y0_el);
1626
1627 *x1_B = (box->x / fmtl->bw + x0_el) * cpp;
1628 *y1_el = box->y / fmtl->bh + y0_el;
1629 *x2_B = (DIV_ROUND_UP(box->x + box->width, fmtl->bw) + x0_el) * cpp;
1630 *y2_el = DIV_ROUND_UP(box->y + box->height, fmtl->bh) + y0_el;
1631 }
1632
1633 static void
1634 iris_unmap_tiled_memcpy(struct iris_transfer *map)
1635 {
1636 struct pipe_transfer *xfer = &map->base;
1637 const struct pipe_box *box = &xfer->box;
1638 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1639 struct isl_surf *surf = &res->surf;
1640
1641 const bool has_swizzling = false;
1642
1643 if (xfer->usage & PIPE_TRANSFER_WRITE) {
1644 char *dst =
1645 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1646
1647 for (int s = 0; s < box->depth; s++) {
1648 unsigned x1, x2, y1, y2;
1649 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1650
1651 void *ptr = map->ptr + s * xfer->layer_stride;
1652
1653 isl_memcpy_linear_to_tiled(x1, x2, y1, y2, dst, ptr,
1654 surf->row_pitch_B, xfer->stride,
1655 has_swizzling, surf->tiling, ISL_MEMCPY);
1656 }
1657 }
1658 os_free_aligned(map->buffer);
1659 map->buffer = map->ptr = NULL;
1660 }
1661
1662 static void
1663 iris_map_tiled_memcpy(struct iris_transfer *map)
1664 {
1665 struct pipe_transfer *xfer = &map->base;
1666 const struct pipe_box *box = &xfer->box;
1667 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1668 struct isl_surf *surf = &res->surf;
1669
1670 xfer->stride = ALIGN(surf->row_pitch_B, 16);
1671 xfer->layer_stride = xfer->stride * box->height;
1672
1673 unsigned x1, x2, y1, y2;
1674 tile_extents(surf, box, xfer->level, 0, &x1, &x2, &y1, &y2);
1675
1676 /* The tiling and detiling functions require that the linear buffer has
1677 * a 16-byte alignment (that is, its `x0` is 16-byte aligned). Here we
1678 * over-allocate the linear buffer to get the proper alignment.
1679 */
1680 map->buffer =
1681 os_malloc_aligned(xfer->layer_stride * box->depth, 16);
1682 assert(map->buffer);
1683 map->ptr = (char *)map->buffer + (x1 & 0xf);
1684
1685 const bool has_swizzling = false;
1686
1687 if (!(xfer->usage & PIPE_TRANSFER_DISCARD_RANGE)) {
1688 char *src =
1689 iris_bo_map(map->dbg, res->bo, (xfer->usage | MAP_RAW) & MAP_FLAGS);
1690
1691 for (int s = 0; s < box->depth; s++) {
1692 unsigned x1, x2, y1, y2;
1693 tile_extents(surf, box, xfer->level, s, &x1, &x2, &y1, &y2);
1694
1695 /* Use 's' rather than 'box->z' to rebase the first slice to 0. */
1696 void *ptr = map->ptr + s * xfer->layer_stride;
1697
1698 isl_memcpy_tiled_to_linear(x1, x2, y1, y2, ptr, src, xfer->stride,
1699 surf->row_pitch_B, has_swizzling,
1700 surf->tiling, ISL_MEMCPY_STREAMING_LOAD);
1701 }
1702 }
1703
1704 map->unmap = iris_unmap_tiled_memcpy;
1705 }
1706
1707 static void
1708 iris_map_direct(struct iris_transfer *map)
1709 {
1710 struct pipe_transfer *xfer = &map->base;
1711 struct pipe_box *box = &xfer->box;
1712 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1713
1714 void *ptr = iris_bo_map(map->dbg, res->bo, xfer->usage & MAP_FLAGS);
1715
1716 if (res->base.target == PIPE_BUFFER) {
1717 xfer->stride = 0;
1718 xfer->layer_stride = 0;
1719
1720 map->ptr = ptr + box->x;
1721 } else {
1722 struct isl_surf *surf = &res->surf;
1723 const struct isl_format_layout *fmtl =
1724 isl_format_get_layout(surf->format);
1725 const unsigned cpp = fmtl->bpb / 8;
1726 unsigned x0_el, y0_el;
1727
1728 get_image_offset_el(surf, xfer->level, box->z, &x0_el, &y0_el);
1729
1730 xfer->stride = isl_surf_get_row_pitch_B(surf);
1731 xfer->layer_stride = isl_surf_get_array_pitch(surf);
1732
1733 map->ptr = ptr + (y0_el + box->y) * xfer->stride + (x0_el + box->x) * cpp;
1734 }
1735 }
1736
1737 static bool
1738 can_promote_to_async(const struct iris_resource *res,
1739 const struct pipe_box *box,
1740 enum pipe_transfer_usage usage)
1741 {
1742 /* If we're writing to a section of the buffer that hasn't even been
1743 * initialized with useful data, then we can safely promote this write
1744 * to be unsynchronized. This helps the common pattern of appending data.
1745 */
1746 return res->base.target == PIPE_BUFFER && (usage & PIPE_TRANSFER_WRITE) &&
1747 !(usage & TC_TRANSFER_MAP_NO_INFER_UNSYNCHRONIZED) &&
1748 !util_ranges_intersect(&res->valid_buffer_range, box->x,
1749 box->x + box->width);
1750 }
1751
1752 static void *
1753 iris_transfer_map(struct pipe_context *ctx,
1754 struct pipe_resource *resource,
1755 unsigned level,
1756 enum pipe_transfer_usage usage,
1757 const struct pipe_box *box,
1758 struct pipe_transfer **ptransfer)
1759 {
1760 struct iris_context *ice = (struct iris_context *)ctx;
1761 struct iris_resource *res = (struct iris_resource *)resource;
1762 struct isl_surf *surf = &res->surf;
1763
1764 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
1765 /* Replace the backing storage with a fresh buffer for non-async maps */
1766 if (!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
1767 TC_TRANSFER_MAP_NO_INVALIDATE)))
1768 iris_invalidate_resource(ctx, resource);
1769
1770 /* If we can discard the whole resource, we can discard the range. */
1771 usage |= PIPE_TRANSFER_DISCARD_RANGE;
1772 }
1773
1774 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
1775 can_promote_to_async(res, box, usage)) {
1776 usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
1777 }
1778
1779 bool need_resolve = false;
1780 bool need_color_resolve = false;
1781
1782 if (resource->target != PIPE_BUFFER) {
1783 bool need_hiz_resolve = iris_resource_level_has_hiz(res, level);
1784
1785 need_color_resolve =
1786 (res->aux.usage == ISL_AUX_USAGE_CCS_D ||
1787 res->aux.usage == ISL_AUX_USAGE_CCS_E) &&
1788 iris_has_color_unresolved(res, level, 1, box->z, box->depth);
1789
1790 need_resolve = need_color_resolve || need_hiz_resolve;
1791 }
1792
1793 bool map_would_stall = false;
1794
1795 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
1796 map_would_stall = need_resolve || resource_is_busy(ice, res);
1797
1798 if (map_would_stall && (usage & PIPE_TRANSFER_DONTBLOCK) &&
1799 (usage & PIPE_TRANSFER_MAP_DIRECTLY))
1800 return NULL;
1801 }
1802
1803 if (surf->tiling != ISL_TILING_LINEAR &&
1804 (usage & PIPE_TRANSFER_MAP_DIRECTLY))
1805 return NULL;
1806
1807 struct iris_transfer *map = slab_alloc(&ice->transfer_pool);
1808 struct pipe_transfer *xfer = &map->base;
1809
1810 if (!map)
1811 return NULL;
1812
1813 memset(map, 0, sizeof(*map));
1814 map->dbg = &ice->dbg;
1815
1816 pipe_resource_reference(&xfer->resource, resource);
1817 xfer->level = level;
1818 xfer->usage = usage;
1819 xfer->box = *box;
1820 *ptransfer = xfer;
1821
1822 map->dest_had_defined_contents =
1823 util_ranges_intersect(&res->valid_buffer_range, box->x,
1824 box->x + box->width);
1825
1826 if (usage & PIPE_TRANSFER_WRITE)
1827 util_range_add(&res->base, &res->valid_buffer_range, box->x, box->x + box->width);
1828
1829 /* Avoid using GPU copies for persistent/coherent buffers, as the idea
1830 * there is to access them simultaneously on the CPU & GPU. This also
1831 * avoids trying to use GPU copies for our u_upload_mgr buffers which
1832 * contain state we're constructing for a GPU draw call, which would
1833 * kill us with infinite stack recursion.
1834 */
1835 bool no_gpu = usage & (PIPE_TRANSFER_PERSISTENT |
1836 PIPE_TRANSFER_COHERENT |
1837 PIPE_TRANSFER_MAP_DIRECTLY);
1838
1839 /* GPU copies are not useful for buffer reads. Instead of stalling to
1840 * read from the original buffer, we'd simply copy it to a temporary...
1841 * then stall (a bit longer) to read from that buffer.
1842 *
1843 * Images are less clear-cut. Color resolves are destructive, removing
1844 * the underlying compression, so we'd rather blit the data to a linear
1845 * temporary and map that, to avoid the resolve. (It might be better to
1846 * a tiled temporary and use the tiled_memcpy paths...)
1847 */
1848 if (!(usage & PIPE_TRANSFER_DISCARD_RANGE) && !need_color_resolve)
1849 no_gpu = true;
1850
1851 const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
1852 if (fmtl->txc == ISL_TXC_ASTC)
1853 no_gpu = true;
1854
1855 if ((map_would_stall || res->aux.usage == ISL_AUX_USAGE_CCS_E) && !no_gpu) {
1856 /* If we need a synchronous mapping and the resource is busy, or needs
1857 * resolving, we copy to/from a linear temporary buffer using the GPU.
1858 */
1859 map->batch = &ice->batches[IRIS_BATCH_RENDER];
1860 map->blorp = &ice->blorp;
1861 iris_map_copy_region(map);
1862 } else {
1863 /* Otherwise we're free to map on the CPU. */
1864
1865 if (need_resolve) {
1866 iris_resource_access_raw(ice, &ice->batches[IRIS_BATCH_RENDER], res,
1867 level, box->z, box->depth,
1868 usage & PIPE_TRANSFER_WRITE);
1869 }
1870
1871 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
1872 for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
1873 if (iris_batch_references(&ice->batches[i], res->bo))
1874 iris_batch_flush(&ice->batches[i]);
1875 }
1876 }
1877
1878 if (surf->tiling == ISL_TILING_W) {
1879 /* TODO: Teach iris_map_tiled_memcpy about W-tiling... */
1880 iris_map_s8(map);
1881 } else if (surf->tiling != ISL_TILING_LINEAR) {
1882 iris_map_tiled_memcpy(map);
1883 } else {
1884 iris_map_direct(map);
1885 }
1886 }
1887
1888 return map->ptr;
1889 }
1890
1891 static void
1892 iris_transfer_flush_region(struct pipe_context *ctx,
1893 struct pipe_transfer *xfer,
1894 const struct pipe_box *box)
1895 {
1896 struct iris_context *ice = (struct iris_context *)ctx;
1897 struct iris_resource *res = (struct iris_resource *) xfer->resource;
1898 struct iris_transfer *map = (void *) xfer;
1899
1900 if (map->staging)
1901 iris_flush_staging_region(xfer, box);
1902
1903 uint32_t history_flush = 0;
1904
1905 if (res->base.target == PIPE_BUFFER) {
1906 if (map->staging)
1907 history_flush |= PIPE_CONTROL_RENDER_TARGET_FLUSH;
1908
1909 if (map->dest_had_defined_contents)
1910 history_flush |= iris_flush_bits_for_history(res);
1911
1912 util_range_add(&res->base, &res->valid_buffer_range, box->x, box->x + box->width);
1913 }
1914
1915 if (history_flush & ~PIPE_CONTROL_CS_STALL) {
1916 for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
1917 struct iris_batch *batch = &ice->batches[i];
1918 if (batch->contains_draw || batch->cache.render->entries) {
1919 iris_batch_maybe_flush(batch, 24);
1920 iris_emit_pipe_control_flush(batch,
1921 "cache history: transfer flush",
1922 history_flush);
1923 }
1924 }
1925 }
1926
1927 /* Make sure we flag constants dirty even if there's no need to emit
1928 * any PIPE_CONTROLs to a batch.
1929 */
1930 iris_dirty_for_history(ice, res);
1931 }
1932
1933 static void
1934 iris_transfer_unmap(struct pipe_context *ctx, struct pipe_transfer *xfer)
1935 {
1936 struct iris_context *ice = (struct iris_context *)ctx;
1937 struct iris_transfer *map = (void *) xfer;
1938
1939 if (!(xfer->usage & (PIPE_TRANSFER_FLUSH_EXPLICIT |
1940 PIPE_TRANSFER_COHERENT))) {
1941 struct pipe_box flush_box = {
1942 .x = 0, .y = 0, .z = 0,
1943 .width = xfer->box.width,
1944 .height = xfer->box.height,
1945 .depth = xfer->box.depth,
1946 };
1947 iris_transfer_flush_region(ctx, xfer, &flush_box);
1948 }
1949
1950 if (map->unmap)
1951 map->unmap(map);
1952
1953 pipe_resource_reference(&xfer->resource, NULL);
1954 slab_free(&ice->transfer_pool, map);
1955 }
1956
1957 /**
1958 * Mark state dirty that needs to be re-emitted when a resource is written.
1959 */
1960 void
1961 iris_dirty_for_history(struct iris_context *ice,
1962 struct iris_resource *res)
1963 {
1964 uint64_t dirty = 0ull;
1965
1966 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1967 dirty |= ((uint64_t)res->bind_stages) << IRIS_SHIFT_FOR_DIRTY_CONSTANTS;
1968 }
1969
1970 ice->state.dirty |= dirty;
1971 }
1972
1973 /**
1974 * Produce a set of PIPE_CONTROL bits which ensure data written to a
1975 * resource becomes visible, and any stale read cache data is invalidated.
1976 */
1977 uint32_t
1978 iris_flush_bits_for_history(struct iris_resource *res)
1979 {
1980 uint32_t flush = PIPE_CONTROL_CS_STALL;
1981
1982 if (res->bind_history & PIPE_BIND_CONSTANT_BUFFER) {
1983 flush |= PIPE_CONTROL_CONST_CACHE_INVALIDATE |
1984 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1985 }
1986
1987 if (res->bind_history & PIPE_BIND_SAMPLER_VIEW)
1988 flush |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1989
1990 if (res->bind_history & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
1991 flush |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1992
1993 if (res->bind_history & (PIPE_BIND_SHADER_BUFFER | PIPE_BIND_SHADER_IMAGE))
1994 flush |= PIPE_CONTROL_DATA_CACHE_FLUSH;
1995
1996 return flush;
1997 }
1998
1999 void
2000 iris_flush_and_dirty_for_history(struct iris_context *ice,
2001 struct iris_batch *batch,
2002 struct iris_resource *res,
2003 uint32_t extra_flags,
2004 const char *reason)
2005 {
2006 if (res->base.target != PIPE_BUFFER)
2007 return;
2008
2009 uint32_t flush = iris_flush_bits_for_history(res) | extra_flags;
2010
2011 iris_emit_pipe_control_flush(batch, reason, flush);
2012
2013 iris_dirty_for_history(ice, res);
2014 }
2015
2016 bool
2017 iris_resource_set_clear_color(struct iris_context *ice,
2018 struct iris_resource *res,
2019 union isl_color_value color)
2020 {
2021 if (memcmp(&res->aux.clear_color, &color, sizeof(color)) != 0) {
2022 res->aux.clear_color = color;
2023 return true;
2024 }
2025
2026 return false;
2027 }
2028
2029 union isl_color_value
2030 iris_resource_get_clear_color(const struct iris_resource *res,
2031 struct iris_bo **clear_color_bo,
2032 uint64_t *clear_color_offset)
2033 {
2034 assert(res->aux.bo);
2035
2036 if (clear_color_bo)
2037 *clear_color_bo = res->aux.clear_color_bo;
2038 if (clear_color_offset)
2039 *clear_color_offset = res->aux.clear_color_offset;
2040 return res->aux.clear_color;
2041 }
2042
2043 static enum pipe_format
2044 iris_resource_get_internal_format(struct pipe_resource *p_res)
2045 {
2046 struct iris_resource *res = (void *) p_res;
2047 return res->internal_format;
2048 }
2049
2050 static const struct u_transfer_vtbl transfer_vtbl = {
2051 .resource_create = iris_resource_create,
2052 .resource_destroy = iris_resource_destroy,
2053 .transfer_map = iris_transfer_map,
2054 .transfer_unmap = iris_transfer_unmap,
2055 .transfer_flush_region = iris_transfer_flush_region,
2056 .get_internal_format = iris_resource_get_internal_format,
2057 .set_stencil = iris_resource_set_separate_stencil,
2058 .get_stencil = iris_resource_get_separate_stencil,
2059 };
2060
2061 void
2062 iris_init_screen_resource_functions(struct pipe_screen *pscreen)
2063 {
2064 pscreen->query_dmabuf_modifiers = iris_query_dmabuf_modifiers;
2065 pscreen->resource_create_with_modifiers =
2066 iris_resource_create_with_modifiers;
2067 pscreen->resource_create = u_transfer_helper_resource_create;
2068 pscreen->resource_from_user_memory = iris_resource_from_user_memory;
2069 pscreen->resource_from_handle = iris_resource_from_handle;
2070 pscreen->resource_get_handle = iris_resource_get_handle;
2071 pscreen->resource_get_param = iris_resource_get_param;
2072 pscreen->resource_destroy = u_transfer_helper_resource_destroy;
2073 pscreen->transfer_helper =
2074 u_transfer_helper_create(&transfer_vtbl, true, true, false, true);
2075 }
2076
2077 void
2078 iris_init_resource_functions(struct pipe_context *ctx)
2079 {
2080 ctx->flush_resource = iris_flush_resource;
2081 ctx->invalidate_resource = iris_invalidate_resource;
2082 ctx->transfer_map = u_transfer_helper_transfer_map;
2083 ctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
2084 ctx->transfer_unmap = u_transfer_helper_transfer_unmap;
2085 ctx->buffer_subdata = u_default_buffer_subdata;
2086 ctx->texture_subdata = u_default_texture_subdata;
2087 }