1 /*
2 * Copyright © 2015 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 (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #ifdef MAJOR_IN_MKDEV
28 #include <sys/mkdev.h>
29 #endif
30 #ifdef MAJOR_IN_SYSMACROS
31 #include <sys/sysmacros.h>
32 #endif
33 #include <sys/mman.h>
34 #include <sys/stat.h>
35 #include <unistd.h>
36 #include <fcntl.h>
37 #include "drm-uapi/drm_fourcc.h"
38 #include "drm-uapi/drm.h"
39 #include <xf86drm.h>
40
41 #include "anv_private.h"
42 #include "anv_measure.h"
43 #include "util/debug.h"
44 #include "util/build_id.h"
45 #include "util/disk_cache.h"
46 #include "util/mesa-sha1.h"
47 #include "util/os_file.h"
48 #include "util/os_misc.h"
49 #include "util/u_atomic.h"
50 #include "util/u_string.h"
51 #include "util/driconf.h"
52 #include "git_sha1.h"
53 #include "vk_util.h"
54 #include "vk_deferred_operation.h"
55 #include "common/intel_aux_map.h"
56 #include "common/intel_defines.h"
57 #include "common/intel_uuid.h"
58 #include "perf/intel_perf.h"
59
60 #include "genxml/gen7_pack.h"
61
62 static const driOptionDescription anv_dri_options[] = {
63 DRI_CONF_SECTION_PERFORMANCE
64 DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
65 DRI_CONF_VK_X11_STRICT_IMAGE_COUNT(false)
66 DRI_CONF_VK_XWAYLAND_WAIT_READY(true)
67 DRI_CONF_SECTION_END
68
69 DRI_CONF_SECTION_DEBUG
70 DRI_CONF_ALWAYS_FLUSH_CACHE(false)
71 DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST(false)
72 DRI_CONF_SECTION_END
73 };
74
75 /* This is probably far to big but it reflects the max size used for messages
76 * in OpenGLs KHR_debug.
77 */
78 #define MAX_DEBUG_MESSAGE_LENGTH 4096
79
80 /* Render engine timestamp register */
81 #define TIMESTAMP 0x2358
82
83 /* The "RAW" clocks on Linux are called "FAST" on FreeBSD */
84 #if !defined(CLOCK_MONOTONIC_RAW) && defined(CLOCK_MONOTONIC_FAST)
85 #define CLOCK_MONOTONIC_RAW CLOCK_MONOTONIC_FAST
86 #endif
87
88 static void
compiler_debug_log(void * data,UNUSED unsigned * id,const char * fmt,...)89 compiler_debug_log(void *data, UNUSED unsigned *id, const char *fmt, ...)
90 {
91 char str[MAX_DEBUG_MESSAGE_LENGTH];
92 struct anv_device *device = (struct anv_device *)data;
93 struct anv_instance *instance = device->physical->instance;
94
95 va_list args;
96 va_start(args, fmt);
97 (void) vsnprintf(str, MAX_DEBUG_MESSAGE_LENGTH, fmt, args);
98 va_end(args);
99
100 vk_logd(VK_LOG_NO_OBJS(&instance->vk), "%s", str);
101 }
102
103 static void
compiler_perf_log(UNUSED void * data,UNUSED unsigned * id,const char * fmt,...)104 compiler_perf_log(UNUSED void *data, UNUSED unsigned *id, const char *fmt, ...)
105 {
106 va_list args;
107 va_start(args, fmt);
108
109 if (INTEL_DEBUG(DEBUG_PERF))
110 mesa_logd_v(fmt, args);
111
112 va_end(args);
113 }
114
115 #if defined(VK_USE_PLATFORM_WAYLAND_KHR) || \
116 defined(VK_USE_PLATFORM_XCB_KHR) || \
117 defined(VK_USE_PLATFORM_XLIB_KHR) || \
118 defined(VK_USE_PLATFORM_DISPLAY_KHR)
119 #define ANV_USE_WSI_PLATFORM
120 #endif
121
122 #ifdef ANDROID
123 #define ANV_API_VERSION VK_MAKE_VERSION(1, 1, VK_HEADER_VERSION)
124 #else
125 #define ANV_API_VERSION VK_MAKE_VERSION(1, 2, VK_HEADER_VERSION)
126 #endif
127
anv_EnumerateInstanceVersion(uint32_t * pApiVersion)128 VkResult anv_EnumerateInstanceVersion(
129 uint32_t* pApiVersion)
130 {
131 *pApiVersion = ANV_API_VERSION;
132 return VK_SUCCESS;
133 }
134
135 static const struct vk_instance_extension_table instance_extensions = {
136 .KHR_device_group_creation = true,
137 .KHR_external_fence_capabilities = true,
138 .KHR_external_memory_capabilities = true,
139 .KHR_external_semaphore_capabilities = true,
140 .KHR_get_physical_device_properties2 = true,
141 .EXT_debug_report = true,
142
143 #ifdef ANV_USE_WSI_PLATFORM
144 .KHR_get_surface_capabilities2 = true,
145 .KHR_surface = true,
146 .KHR_surface_protected_capabilities = true,
147 #endif
148 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
149 .KHR_wayland_surface = true,
150 #endif
151 #ifdef VK_USE_PLATFORM_XCB_KHR
152 .KHR_xcb_surface = true,
153 #endif
154 #ifdef VK_USE_PLATFORM_XLIB_KHR
155 .KHR_xlib_surface = true,
156 #endif
157 #ifdef VK_USE_PLATFORM_XLIB_XRANDR_EXT
158 .EXT_acquire_xlib_display = true,
159 #endif
160 #ifdef VK_USE_PLATFORM_DISPLAY_KHR
161 .KHR_display = true,
162 .KHR_get_display_properties2 = true,
163 .EXT_direct_mode_display = true,
164 .EXT_display_surface_counter = true,
165 .EXT_acquire_drm_display = true,
166 #endif
167 };
168
169 static void
get_device_extensions(const struct anv_physical_device * device,struct vk_device_extension_table * ext)170 get_device_extensions(const struct anv_physical_device *device,
171 struct vk_device_extension_table *ext)
172 {
173 *ext = (struct vk_device_extension_table) {
174 .KHR_8bit_storage = device->info.ver >= 8,
175 .KHR_16bit_storage = device->info.ver >= 8,
176 .KHR_bind_memory2 = true,
177 .KHR_buffer_device_address = device->has_a64_buffer_access,
178 .KHR_copy_commands2 = true,
179 .KHR_create_renderpass2 = true,
180 .KHR_dedicated_allocation = true,
181 .KHR_deferred_host_operations = true,
182 .KHR_depth_stencil_resolve = true,
183 .KHR_descriptor_update_template = true,
184 .KHR_device_group = true,
185 .KHR_draw_indirect_count = true,
186 .KHR_driver_properties = true,
187 .KHR_external_fence = device->has_syncobj_wait,
188 .KHR_external_fence_fd = device->has_syncobj_wait,
189 .KHR_external_memory = true,
190 .KHR_external_memory_fd = true,
191 .KHR_external_semaphore = true,
192 .KHR_external_semaphore_fd = true,
193 .KHR_format_feature_flags2 = true,
194 .KHR_fragment_shading_rate = device->info.ver >= 11,
195 .KHR_get_memory_requirements2 = true,
196 .KHR_image_format_list = true,
197 .KHR_imageless_framebuffer = true,
198 #ifdef ANV_USE_WSI_PLATFORM
199 .KHR_incremental_present = true,
200 #endif
201 .KHR_maintenance1 = true,
202 .KHR_maintenance2 = true,
203 .KHR_maintenance3 = true,
204 .KHR_maintenance4 = true,
205 .KHR_multiview = true,
206 .KHR_performance_query =
207 device->use_softpin && device->perf &&
208 (device->perf->i915_perf_version >= 3 ||
209 INTEL_DEBUG(DEBUG_NO_OACONFIG)) &&
210 device->use_call_secondary,
211 .KHR_pipeline_executable_properties = true,
212 .KHR_push_descriptor = true,
213 .KHR_relaxed_block_layout = true,
214 .KHR_sampler_mirror_clamp_to_edge = true,
215 .KHR_sampler_ycbcr_conversion = true,
216 .KHR_separate_depth_stencil_layouts = true,
217 .KHR_shader_atomic_int64 = device->info.ver >= 9 &&
218 device->use_softpin,
219 .KHR_shader_clock = true,
220 .KHR_shader_draw_parameters = true,
221 .KHR_shader_float16_int8 = device->info.ver >= 8,
222 .KHR_shader_float_controls = device->info.ver >= 8,
223 .KHR_shader_integer_dot_product = true,
224 .KHR_shader_non_semantic_info = true,
225 .KHR_shader_subgroup_extended_types = device->info.ver >= 8,
226 .KHR_shader_subgroup_uniform_control_flow = true,
227 .KHR_shader_terminate_invocation = true,
228 .KHR_spirv_1_4 = true,
229 .KHR_storage_buffer_storage_class = true,
230 #ifdef ANV_USE_WSI_PLATFORM
231 .KHR_swapchain = true,
232 .KHR_swapchain_mutable_format = true,
233 #endif
234 .KHR_synchronization2 = true,
235 .KHR_timeline_semaphore = true,
236 .KHR_uniform_buffer_standard_layout = true,
237 .KHR_variable_pointers = true,
238 .KHR_vulkan_memory_model = true,
239 .KHR_workgroup_memory_explicit_layout = true,
240 .KHR_zero_initialize_workgroup_memory = true,
241 .EXT_4444_formats = true,
242 .EXT_buffer_device_address = device->has_a64_buffer_access,
243 .EXT_calibrated_timestamps = device->has_reg_timestamp,
244 .EXT_color_write_enable = true,
245 .EXT_conditional_rendering = device->info.verx10 >= 75,
246 .EXT_conservative_rasterization = device->info.ver >= 9,
247 .EXT_custom_border_color = device->info.ver >= 8,
248 .EXT_depth_clip_enable = true,
249 .EXT_descriptor_indexing = device->has_a64_buffer_access &&
250 device->has_bindless_images,
251 #ifdef VK_USE_PLATFORM_DISPLAY_KHR
252 .EXT_display_control = true,
253 #endif
254 .EXT_extended_dynamic_state = true,
255 .EXT_extended_dynamic_state2 = true,
256 .EXT_external_memory_dma_buf = true,
257 .EXT_external_memory_host = true,
258 .EXT_fragment_shader_interlock = device->info.ver >= 9,
259 .EXT_global_priority = device->has_context_priority,
260 .EXT_host_query_reset = true,
261 .EXT_image_robustness = true,
262 .EXT_image_drm_format_modifier = true,
263 .EXT_index_type_uint8 = true,
264 .EXT_inline_uniform_block = true,
265 .EXT_line_rasterization = true,
266 .EXT_memory_budget = device->sys.available,
267 .EXT_pci_bus_info = true,
268 .EXT_physical_device_drm = true,
269 .EXT_pipeline_creation_cache_control = true,
270 .EXT_pipeline_creation_feedback = true,
271 .EXT_post_depth_coverage = device->info.ver >= 9,
272 .EXT_primitive_topology_list_restart = true,
273 .EXT_private_data = true,
274 .EXT_provoking_vertex = true,
275 .EXT_queue_family_foreign = true,
276 .EXT_robustness2 = true,
277 .EXT_sample_locations = true,
278 .EXT_sampler_filter_minmax = device->info.ver >= 9,
279 .EXT_scalar_block_layout = true,
280 .EXT_separate_stencil_usage = true,
281 .EXT_shader_atomic_float = true,
282 .EXT_shader_atomic_float2 = device->info.ver >= 9,
283 .EXT_shader_demote_to_helper_invocation = true,
284 .EXT_shader_stencil_export = device->info.ver >= 9,
285 .EXT_shader_subgroup_ballot = true,
286 .EXT_shader_subgroup_vote = true,
287 .EXT_shader_viewport_index_layer = true,
288 .EXT_subgroup_size_control = true,
289 .EXT_texel_buffer_alignment = true,
290 .EXT_transform_feedback = true,
291 .EXT_vertex_attribute_divisor = true,
292 .EXT_ycbcr_image_arrays = true,
293 #ifdef ANDROID
294 .ANDROID_external_memory_android_hardware_buffer = true,
295 .ANDROID_native_buffer = true,
296 #endif
297 .GOOGLE_decorate_string = true,
298 .GOOGLE_hlsl_functionality1 = true,
299 .GOOGLE_user_type = true,
300 .INTEL_performance_query = device->perf &&
301 device->perf->i915_perf_version >= 3,
302 .INTEL_shader_integer_functions2 = device->info.ver >= 8,
303 .EXT_multi_draw = true,
304 .NV_compute_shader_derivatives = true,
305 };
306 }
307
308 static uint64_t
anv_compute_sys_heap_size(struct anv_physical_device * device,uint64_t total_ram)309 anv_compute_sys_heap_size(struct anv_physical_device *device,
310 uint64_t total_ram)
311 {
312 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
313 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
314 */
315 uint64_t available_ram;
316 if (total_ram <= 4ull * 1024ull * 1024ull * 1024ull)
317 available_ram = total_ram / 2;
318 else
319 available_ram = total_ram * 3 / 4;
320
321 /* We also want to leave some padding for things we allocate in the driver,
322 * so don't go over 3/4 of the GTT either.
323 */
324 available_ram = MIN2(available_ram, device->gtt_size * 3 / 4);
325
326 if (available_ram > (2ull << 30) && !device->supports_48bit_addresses) {
327 /* When running with an overridden PCI ID, we may get a GTT size from
328 * the kernel that is greater than 2 GiB but the execbuf check for 48bit
329 * address support can still fail. Just clamp the address space size to
330 * 2 GiB if we don't have 48-bit support.
331 */
332 mesa_logw("%s:%d: The kernel reported a GTT size larger than 2 GiB but "
333 "not support for 48-bit addresses",
334 __FILE__, __LINE__);
335 available_ram = 2ull << 30;
336 }
337
338 return available_ram;
339 }
340
341 static VkResult MUST_CHECK
anv_gather_meminfo(struct anv_physical_device * device,int fd,bool update)342 anv_gather_meminfo(struct anv_physical_device *device, int fd, bool update)
343 {
344 char sys_mem_regions[sizeof(struct drm_i915_query_memory_regions) +
345 sizeof(struct drm_i915_memory_region_info)];
346
347 struct drm_i915_query_memory_regions *mem_regions =
348 intel_i915_query_alloc(fd, DRM_I915_QUERY_MEMORY_REGIONS);
349 if (mem_regions == NULL) {
350 if (device->info.has_local_mem) {
351 return vk_errorf(device, VK_ERROR_INCOMPATIBLE_DRIVER,
352 "failed to memory regions: %m");
353 }
354
355 uint64_t total_phys;
356 if (!os_get_total_physical_memory(&total_phys)) {
357 return vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
358 "failed to get total physical memory: %m");
359 }
360
361 uint64_t available;
362 if (!os_get_available_system_memory(&available))
363 available = 0; /* Silently disable VK_EXT_memory_budget */
364
365 /* The kernel query failed. Fake it using OS memory queries. This
366 * should be roughly the same for integrated GPUs.
367 */
368 mem_regions = (void *)sys_mem_regions;
369 mem_regions->num_regions = 1;
370 mem_regions->regions[0] = (struct drm_i915_memory_region_info) {
371 .region.memory_class = I915_MEMORY_CLASS_SYSTEM,
372 .probed_size = total_phys,
373 .unallocated_size = available,
374 };
375 }
376
377 for(int i = 0; i < mem_regions->num_regions; i++) {
378 struct drm_i915_memory_region_info *info = &mem_regions->regions[i];
379
380 struct anv_memregion *region;
381 switch (info->region.memory_class) {
382 case I915_MEMORY_CLASS_SYSTEM:
383 region = &device->sys;
384 break;
385 case I915_MEMORY_CLASS_DEVICE:
386 region = &device->vram;
387 break;
388 default:
389 /* We don't know what kind of memory this is */
390 continue;
391 }
392
393 uint64_t size = info->probed_size;
394 if (info->region.memory_class == I915_MEMORY_CLASS_SYSTEM)
395 size = anv_compute_sys_heap_size(device, size);
396
397 uint64_t available = MIN2(size, info->unallocated_size);
398
399 if (update) {
400 assert(region->region.memory_class == info->region.memory_class);
401 assert(region->region.memory_instance == info->region.memory_instance);
402 assert(region->size == size);
403 } else {
404 region->region = info->region;
405 region->size = size;
406 }
407 region->available = available;
408 }
409
410 if (mem_regions != (void *)sys_mem_regions)
411 free(mem_regions);
412
413 return VK_SUCCESS;
414 }
415
416 static VkResult MUST_CHECK
anv_init_meminfo(struct anv_physical_device * device,int fd)417 anv_init_meminfo(struct anv_physical_device *device, int fd)
418 {
419 return anv_gather_meminfo(device, fd, false);
420 }
421
422 static void
anv_update_meminfo(struct anv_physical_device * device,int fd)423 anv_update_meminfo(struct anv_physical_device *device, int fd)
424 {
425 ASSERTED VkResult result = anv_gather_meminfo(device, fd, true);
426 assert(result == VK_SUCCESS);
427 }
428
429
430 static VkResult
anv_physical_device_init_heaps(struct anv_physical_device * device,int fd)431 anv_physical_device_init_heaps(struct anv_physical_device *device, int fd)
432 {
433 if (anv_gem_get_context_param(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE,
434 &device->gtt_size) == -1) {
435 /* If, for whatever reason, we can't actually get the GTT size from the
436 * kernel (too old?) fall back to the aperture size.
437 */
438 anv_perf_warn(VK_LOG_NO_OBJS(&device->instance->vk),
439 "Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
440
441 if (intel_get_aperture_size(fd, &device->gtt_size) == -1) {
442 return vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
443 "failed to get aperture size: %m");
444 }
445 }
446
447 /* We only allow 48-bit addresses with softpin because knowing the actual
448 * address is required for the vertex cache flush workaround.
449 */
450 device->supports_48bit_addresses = (device->info.ver >= 8) &&
451 device->gtt_size > (4ULL << 30 /* GiB */);
452
453 VkResult result = anv_init_meminfo(device, fd);
454 if (result != VK_SUCCESS)
455 return result;
456
457 assert(device->sys.size != 0);
458
459 if (device->vram.size > 0) {
460 /* We can create 2 different heaps when we have local memory support,
461 * first heap with local memory size and second with system memory size.
462 */
463 device->memory.heap_count = 2;
464 device->memory.heaps[0] = (struct anv_memory_heap) {
465 .size = device->vram.size,
466 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
467 .is_local_mem = true,
468 };
469 device->memory.heaps[1] = (struct anv_memory_heap) {
470 .size = device->sys.size,
471 .flags = 0,
472 .is_local_mem = false,
473 };
474
475 device->memory.type_count = 3;
476 device->memory.types[0] = (struct anv_memory_type) {
477 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
478 .heapIndex = 0,
479 };
480 device->memory.types[1] = (struct anv_memory_type) {
481 .propertyFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
482 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
483 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
484 .heapIndex = 1,
485 };
486 device->memory.types[2] = (struct anv_memory_type) {
487 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
488 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
489 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
490 .heapIndex = 0,
491 };
492 } else if (device->info.has_llc) {
493 device->memory.heap_count = 1;
494 device->memory.heaps[0] = (struct anv_memory_heap) {
495 .size = device->sys.size,
496 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
497 .is_local_mem = false,
498 };
499
500 /* Big core GPUs share LLC with the CPU and thus one memory type can be
501 * both cached and coherent at the same time.
502 */
503 device->memory.type_count = 1;
504 device->memory.types[0] = (struct anv_memory_type) {
505 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
506 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
507 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
508 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
509 .heapIndex = 0,
510 };
511 } else {
512 device->memory.heap_count = 1;
513 device->memory.heaps[0] = (struct anv_memory_heap) {
514 .size = device->sys.size,
515 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
516 .is_local_mem = false,
517 };
518
519 /* The spec requires that we expose a host-visible, coherent memory
520 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
521 * to give the application a choice between cached, but not coherent and
522 * coherent but uncached (WC though).
523 */
524 device->memory.type_count = 2;
525 device->memory.types[0] = (struct anv_memory_type) {
526 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
527 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
528 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
529 .heapIndex = 0,
530 };
531 device->memory.types[1] = (struct anv_memory_type) {
532 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
533 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
534 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
535 .heapIndex = 0,
536 };
537 }
538
539 device->memory.need_clflush = false;
540 for (unsigned i = 0; i < device->memory.type_count; i++) {
541 VkMemoryPropertyFlags props = device->memory.types[i].propertyFlags;
542 if ((props & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) &&
543 !(props & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
544 device->memory.need_clflush = true;
545 }
546
547 return VK_SUCCESS;
548 }
549
550 static VkResult
anv_physical_device_init_uuids(struct anv_physical_device * device)551 anv_physical_device_init_uuids(struct anv_physical_device *device)
552 {
553 const struct build_id_note *note =
554 build_id_find_nhdr_for_addr(anv_physical_device_init_uuids);
555 if (!note) {
556 return vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
557 "Failed to find build-id");
558 }
559
560 unsigned build_id_len = build_id_length(note);
561 if (build_id_len < 20) {
562 return vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
563 "build-id too short. It needs to be a SHA");
564 }
565
566 memcpy(device->driver_build_sha1, build_id_data(note), 20);
567
568 struct mesa_sha1 sha1_ctx;
569 uint8_t sha1[20];
570 STATIC_ASSERT(VK_UUID_SIZE <= sizeof(sha1));
571
572 /* The pipeline cache UUID is used for determining when a pipeline cache is
573 * invalid. It needs both a driver build and the PCI ID of the device.
574 */
575 _mesa_sha1_init(&sha1_ctx);
576 _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len);
577 _mesa_sha1_update(&sha1_ctx, &device->info.chipset_id,
578 sizeof(device->info.chipset_id));
579 _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless,
580 sizeof(device->always_use_bindless));
581 _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access,
582 sizeof(device->has_a64_buffer_access));
583 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images,
584 sizeof(device->has_bindless_images));
585 _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers,
586 sizeof(device->has_bindless_samplers));
587 _mesa_sha1_final(&sha1_ctx, sha1);
588 memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE);
589
590 intel_uuid_compute_driver_id(device->driver_uuid, &device->info, VK_UUID_SIZE);
591 intel_uuid_compute_device_id(device->device_uuid, &device->isl_dev, VK_UUID_SIZE);
592
593 return VK_SUCCESS;
594 }
595
596 static void
anv_physical_device_init_disk_cache(struct anv_physical_device * device)597 anv_physical_device_init_disk_cache(struct anv_physical_device *device)
598 {
599 #ifdef ENABLE_SHADER_CACHE
600 char renderer[10];
601 ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x",
602 device->info.chipset_id);
603 assert(len == sizeof(renderer) - 2);
604
605 char timestamp[41];
606 _mesa_sha1_format(timestamp, device->driver_build_sha1);
607
608 const uint64_t driver_flags =
609 brw_get_compiler_config_value(device->compiler);
610 device->disk_cache = disk_cache_create(renderer, timestamp, driver_flags);
611 #else
612 device->disk_cache = NULL;
613 #endif
614 }
615
616 static void
anv_physical_device_free_disk_cache(struct anv_physical_device * device)617 anv_physical_device_free_disk_cache(struct anv_physical_device *device)
618 {
619 #ifdef ENABLE_SHADER_CACHE
620 if (device->disk_cache)
621 disk_cache_destroy(device->disk_cache);
622 #else
623 assert(device->disk_cache == NULL);
624 #endif
625 }
626
627 /* The ANV_QUEUE_OVERRIDE environment variable is a comma separated list of
628 * queue overrides.
629 *
630 * To override the number queues:
631 * * "gc" is for graphics queues with compute support
632 * * "g" is for graphics queues with no compute support
633 * * "c" is for compute queues with no graphics support
634 *
635 * For example, ANV_QUEUE_OVERRIDE=gc=2,c=1 would override the number of
636 * advertised queues to be 2 queues with graphics+compute support, and 1 queue
637 * with compute-only support.
638 *
639 * ANV_QUEUE_OVERRIDE=c=1 would override the number of advertised queues to
640 * include 1 queue with compute-only support, but it will not change the
641 * number of graphics+compute queues.
642 *
643 * ANV_QUEUE_OVERRIDE=gc=0,c=1 would override the number of advertised queues
644 * to include 1 queue with compute-only support, and it would override the
645 * number of graphics+compute queues to be 0.
646 */
647 static void
anv_override_engine_counts(int * gc_count,int * g_count,int * c_count)648 anv_override_engine_counts(int *gc_count, int *g_count, int *c_count)
649 {
650 int gc_override = -1;
651 int g_override = -1;
652 int c_override = -1;
653 char *env = getenv("ANV_QUEUE_OVERRIDE");
654
655 if (env == NULL)
656 return;
657
658 env = strdup(env);
659 char *save = NULL;
660 char *next = strtok_r(env, ",", &save);
661 while (next != NULL) {
662 if (strncmp(next, "gc=", 3) == 0) {
663 gc_override = strtol(next + 3, NULL, 0);
664 } else if (strncmp(next, "g=", 2) == 0) {
665 g_override = strtol(next + 2, NULL, 0);
666 } else if (strncmp(next, "c=", 2) == 0) {
667 c_override = strtol(next + 2, NULL, 0);
668 } else {
669 mesa_logw("Ignoring unsupported ANV_QUEUE_OVERRIDE token: %s", next);
670 }
671 next = strtok_r(NULL, ",", &save);
672 }
673 free(env);
674 if (gc_override >= 0)
675 *gc_count = gc_override;
676 if (g_override >= 0)
677 *g_count = g_override;
678 if (*g_count > 0 && *gc_count <= 0 && (gc_override >= 0 || g_override >= 0))
679 mesa_logw("ANV_QUEUE_OVERRIDE: gc=0 with g > 0 violates the "
680 "Vulkan specification");
681 if (c_override >= 0)
682 *c_count = c_override;
683 }
684
685 static void
anv_physical_device_init_queue_families(struct anv_physical_device * pdevice)686 anv_physical_device_init_queue_families(struct anv_physical_device *pdevice)
687 {
688 uint32_t family_count = 0;
689
690 if (pdevice->engine_info) {
691 int gc_count =
692 anv_gem_count_engines(pdevice->engine_info, I915_ENGINE_CLASS_RENDER);
693 int g_count = 0;
694 int c_count = 0;
695
696 anv_override_engine_counts(&gc_count, &g_count, &c_count);
697
698 if (gc_count > 0) {
699 pdevice->queue.families[family_count++] = (struct anv_queue_family) {
700 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
701 VK_QUEUE_COMPUTE_BIT |
702 VK_QUEUE_TRANSFER_BIT,
703 .queueCount = gc_count,
704 .engine_class = I915_ENGINE_CLASS_RENDER,
705 };
706 }
707 if (g_count > 0) {
708 pdevice->queue.families[family_count++] = (struct anv_queue_family) {
709 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
710 VK_QUEUE_TRANSFER_BIT,
711 .queueCount = g_count,
712 .engine_class = I915_ENGINE_CLASS_RENDER,
713 };
714 }
715 if (c_count > 0) {
716 pdevice->queue.families[family_count++] = (struct anv_queue_family) {
717 .queueFlags = VK_QUEUE_COMPUTE_BIT |
718 VK_QUEUE_TRANSFER_BIT,
719 .queueCount = c_count,
720 .engine_class = I915_ENGINE_CLASS_RENDER,
721 };
722 }
723 /* Increase count below when other families are added as a reminder to
724 * increase the ANV_MAX_QUEUE_FAMILIES value.
725 */
726 STATIC_ASSERT(ANV_MAX_QUEUE_FAMILIES >= 3);
727 } else {
728 /* Default to a single render queue */
729 pdevice->queue.families[family_count++] = (struct anv_queue_family) {
730 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
731 VK_QUEUE_COMPUTE_BIT |
732 VK_QUEUE_TRANSFER_BIT,
733 .queueCount = 1,
734 .engine_class = I915_ENGINE_CLASS_RENDER,
735 };
736 family_count = 1;
737 }
738 assert(family_count <= ANV_MAX_QUEUE_FAMILIES);
739 pdevice->queue.family_count = family_count;
740 }
741
742 static VkResult
anv_physical_device_try_create(struct anv_instance * instance,drmDevicePtr drm_device,struct anv_physical_device ** device_out)743 anv_physical_device_try_create(struct anv_instance *instance,
744 drmDevicePtr drm_device,
745 struct anv_physical_device **device_out)
746 {
747 const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
748 const char *path = drm_device->nodes[DRM_NODE_RENDER];
749 VkResult result;
750 int fd;
751 int master_fd = -1;
752
753 brw_process_intel_debug_variable();
754
755 fd = open(path, O_RDWR | O_CLOEXEC);
756 if (fd < 0) {
757 if (errno == ENOMEM) {
758 return vk_errorf(instance, VK_ERROR_OUT_OF_HOST_MEMORY,
759 "Unable to open device %s: out of memory", path);
760 }
761 return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
762 "Unable to open device %s: %m", path);
763 }
764
765 struct intel_device_info devinfo;
766 if (!intel_get_device_info_from_fd(fd, &devinfo)) {
767 result = vk_error(instance, VK_ERROR_INCOMPATIBLE_DRIVER);
768 goto fail_fd;
769 }
770
771 bool is_alpha = true;
772 if (devinfo.is_haswell) {
773 mesa_logw("Haswell Vulkan support is incomplete");
774 } else if (devinfo.ver == 7 && !devinfo.is_baytrail) {
775 mesa_logw("Ivy Bridge Vulkan support is incomplete");
776 } else if (devinfo.ver == 7 && devinfo.is_baytrail) {
777 mesa_logw("Bay Trail Vulkan support is incomplete");
778 } else if (devinfo.ver >= 8 && devinfo.ver <= 12) {
779 /* Gfx8-12 fully supported */
780 is_alpha = false;
781 } else {
782 result = vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
783 "Vulkan not yet supported on %s", devinfo.name);
784 goto fail_fd;
785 }
786
787 struct anv_physical_device *device =
788 vk_zalloc(&instance->vk.alloc, sizeof(*device), 8,
789 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
790 if (device == NULL) {
791 result = vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);
792 goto fail_fd;
793 }
794
795 struct vk_physical_device_dispatch_table dispatch_table;
796 vk_physical_device_dispatch_table_from_entrypoints(
797 &dispatch_table, &anv_physical_device_entrypoints, true);
798 vk_physical_device_dispatch_table_from_entrypoints(
799 &dispatch_table, &wsi_physical_device_entrypoints, false);
800
801 result = vk_physical_device_init(&device->vk, &instance->vk,
802 NULL, /* We set up extensions later */
803 &dispatch_table);
804 if (result != VK_SUCCESS) {
805 vk_error(instance, result);
806 goto fail_alloc;
807 }
808 device->instance = instance;
809
810 assert(strlen(path) < ARRAY_SIZE(device->path));
811 snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
812
813 device->info = devinfo;
814 device->is_alpha = is_alpha;
815
816 device->pci_info.domain = drm_device->businfo.pci->domain;
817 device->pci_info.bus = drm_device->businfo.pci->bus;
818 device->pci_info.device = drm_device->businfo.pci->dev;
819 device->pci_info.function = drm_device->businfo.pci->func;
820
821 device->cmd_parser_version = -1;
822 if (device->info.ver == 7) {
823 device->cmd_parser_version =
824 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
825 if (device->cmd_parser_version == -1) {
826 result = vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
827 "failed to get command parser version");
828 goto fail_base;
829 }
830 }
831
832 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
833 result = vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
834 "kernel missing gem wait");
835 goto fail_base;
836 }
837
838 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
839 result = vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
840 "kernel missing execbuf2");
841 goto fail_base;
842 }
843
844 if (!device->info.has_llc &&
845 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
846 result = vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
847 "kernel missing wc mmap");
848 goto fail_base;
849 }
850
851 if (device->info.ver >= 8 && !device->info.is_cherryview &&
852 !anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN)) {
853 result = vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
854 "kernel missing softpin");
855 goto fail_alloc;
856 }
857
858 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE_ARRAY)) {
859 result = vk_errorf(device, VK_ERROR_INITIALIZATION_FAILED,
860 "kernel missing syncobj support");
861 goto fail_base;
862 }
863
864 device->has_exec_async = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_ASYNC);
865 device->has_exec_capture = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_CAPTURE);
866 device->has_exec_fence = anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_FENCE);
867 device->has_syncobj_wait = anv_gem_supports_syncobj_wait(fd);
868 device->has_syncobj_wait_available =
869 anv_gem_get_drm_cap(fd, DRM_CAP_SYNCOBJ_TIMELINE) != 0;
870
871 device->has_context_priority = anv_gem_has_context_priority(fd);
872
873 /* Initialize memory regions struct to 0. */
874 memset(&device->vram, 0, sizeof(device->vram));
875 memset(&device->sys, 0, sizeof(device->sys));
876
877 result = anv_physical_device_init_heaps(device, fd);
878 if (result != VK_SUCCESS)
879 goto fail_base;
880
881 device->use_softpin = device->info.ver >= 8 &&
882 !device->info.is_cherryview;
883 assert(device->use_softpin == device->supports_48bit_addresses);
884
885 device->has_context_isolation =
886 anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION);
887
888 device->has_exec_timeline =
889 anv_gem_get_param(fd, I915_PARAM_HAS_EXEC_TIMELINE_FENCES);
890 if (env_var_as_boolean("ANV_QUEUE_THREAD_DISABLE", false))
891 device->has_exec_timeline = false;
892
893 device->has_thread_submit =
894 device->has_syncobj_wait_available && device->has_exec_timeline;
895
896 device->always_use_bindless =
897 env_var_as_boolean("ANV_ALWAYS_BINDLESS", false);
898
899 device->use_call_secondary =
900 device->use_softpin &&
901 !env_var_as_boolean("ANV_DISABLE_SECONDARY_CMD_BUFFER_CALLS", false);
902
903 /* We first got the A64 messages on broadwell and we can only use them if
904 * we can pass addresses directly into the shader which requires softpin.
905 */
906 device->has_a64_buffer_access = device->info.ver >= 8 &&
907 device->use_softpin;
908
909 /* We first get bindless image access on Skylake.
910 */
911 device->has_bindless_images = device->info.ver >= 9;
912
913 /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms)
914 * because it's just a matter of setting the sampler address in the sample
915 * message header. However, we've not bothered to wire it up for vec4 so
916 * we leave it disabled on gfx7.
917 */
918 device->has_bindless_samplers = device->info.ver >= 8;
919
920 device->has_implicit_ccs = device->info.has_aux_map;
921
922 /* Check if we can read the GPU timestamp register from the CPU */
923 uint64_t u64_ignore;
924 device->has_reg_timestamp = anv_gem_reg_read(fd, TIMESTAMP | I915_REG_READ_8B_WA,
925 &u64_ignore) == 0;
926
927 device->always_flush_cache = INTEL_DEBUG(DEBUG_SYNC) ||
928 driQueryOptionb(&instance->dri_options, "always_flush_cache");
929
930 device->has_mmap_offset =
931 anv_gem_get_param(fd, I915_PARAM_MMAP_GTT_VERSION) >= 4;
932
933 device->has_userptr_probe =
934 anv_gem_get_param(fd, I915_PARAM_HAS_USERPTR_PROBE);
935
936 device->compiler = brw_compiler_create(NULL, &device->info);
937 if (device->compiler == NULL) {
938 result = vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);
939 goto fail_base;
940 }
941 device->compiler->shader_debug_log = compiler_debug_log;
942 device->compiler->shader_perf_log = compiler_perf_log;
943 device->compiler->supports_pull_constants = false;
944 device->compiler->constant_buffer_0_is_relative =
945 device->info.ver < 8 || !device->has_context_isolation;
946 device->compiler->supports_shader_constants = true;
947 device->compiler->compact_params = false;
948 device->compiler->indirect_ubos_use_sampler = device->info.ver < 12;
949
950 /* Broadwell PRM says:
951 *
952 * "Before Gfx8, there was a historical configuration control field to
953 * swizzle address bit[6] for in X/Y tiling modes. This was set in three
954 * different places: TILECTL[1:0], ARB_MODE[5:4], and
955 * DISP_ARB_CTL[14:13].
956 *
957 * For Gfx8 and subsequent generations, the swizzle fields are all
958 * reserved, and the CPU's memory controller performs all address
959 * swizzling modifications."
960 */
961 bool swizzled =
962 device->info.ver < 8 && anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
963
964 isl_device_init(&device->isl_dev, &device->info, swizzled);
965
966 result = anv_physical_device_init_uuids(device);
967 if (result != VK_SUCCESS)
968 goto fail_compiler;
969
970 anv_physical_device_init_disk_cache(device);
971
972 if (instance->vk.enabled_extensions.KHR_display) {
973 master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
974 if (master_fd >= 0) {
975 /* prod the device with a GETPARAM call which will fail if
976 * we don't have permission to even render on this device
977 */
978 if (anv_gem_get_param(master_fd, I915_PARAM_CHIPSET_ID) == 0) {
979 close(master_fd);
980 master_fd = -1;
981 }
982 }
983 }
984 device->master_fd = master_fd;
985
986 device->engine_info = anv_gem_get_engine_info(fd);
987 anv_physical_device_init_queue_families(device);
988
989 result = anv_init_wsi(device);
990 if (result != VK_SUCCESS)
991 goto fail_engine_info;
992
993 anv_physical_device_init_perf(device, fd);
994
995 anv_measure_device_init(device);
996
997 get_device_extensions(device, &device->vk.supported_extensions);
998
999 device->local_fd = fd;
1000
1001 anv_genX(&device->info, init_physical_device_state)(device);
1002
1003 *device_out = device;
1004
1005 struct stat st;
1006
1007 if (stat(primary_path, &st) == 0) {
1008 device->has_master = true;
1009 device->master_major = major(st.st_rdev);
1010 device->master_minor = minor(st.st_rdev);
1011 } else {
1012 device->has_master = false;
1013 device->master_major = 0;
1014 device->master_minor = 0;
1015 }
1016
1017 if (stat(path, &st) == 0) {
1018 device->has_local = true;
1019 device->local_major = major(st.st_rdev);
1020 device->local_minor = minor(st.st_rdev);
1021 } else {
1022 device->has_local = false;
1023 device->local_major = 0;
1024 device->local_minor = 0;
1025 }
1026
1027 return VK_SUCCESS;
1028
1029 fail_engine_info:
1030 free(device->engine_info);
1031 anv_physical_device_free_disk_cache(device);
1032 fail_compiler:
1033 ralloc_free(device->compiler);
1034 fail_base:
1035 vk_physical_device_finish(&device->vk);
1036 fail_alloc:
1037 vk_free(&instance->vk.alloc, device);
1038 fail_fd:
1039 close(fd);
1040 if (master_fd != -1)
1041 close(master_fd);
1042 return result;
1043 }
1044
1045 static void
anv_physical_device_destroy(struct anv_physical_device * device)1046 anv_physical_device_destroy(struct anv_physical_device *device)
1047 {
1048 anv_finish_wsi(device);
1049 anv_measure_device_destroy(device);
1050 free(device->engine_info);
1051 anv_physical_device_free_disk_cache(device);
1052 ralloc_free(device->compiler);
1053 ralloc_free(device->perf);
1054 close(device->local_fd);
1055 if (device->master_fd >= 0)
1056 close(device->master_fd);
1057 vk_physical_device_finish(&device->vk);
1058 vk_free(&device->instance->vk.alloc, device);
1059 }
1060
anv_EnumerateInstanceExtensionProperties(const char * pLayerName,uint32_t * pPropertyCount,VkExtensionProperties * pProperties)1061 VkResult anv_EnumerateInstanceExtensionProperties(
1062 const char* pLayerName,
1063 uint32_t* pPropertyCount,
1064 VkExtensionProperties* pProperties)
1065 {
1066 if (pLayerName)
1067 return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
1068
1069 return vk_enumerate_instance_extension_properties(
1070 &instance_extensions, pPropertyCount, pProperties);
1071 }
1072
1073 static void
anv_init_dri_options(struct anv_instance * instance)1074 anv_init_dri_options(struct anv_instance *instance)
1075 {
1076 driParseOptionInfo(&instance->available_dri_options, anv_dri_options,
1077 ARRAY_SIZE(anv_dri_options));
1078 driParseConfigFiles(&instance->dri_options,
1079 &instance->available_dri_options, 0, "anv", NULL, NULL,
1080 instance->vk.app_info.app_name,
1081 instance->vk.app_info.app_version,
1082 instance->vk.app_info.engine_name,
1083 instance->vk.app_info.engine_version);
1084 }
1085
anv_CreateInstance(const VkInstanceCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkInstance * pInstance)1086 VkResult anv_CreateInstance(
1087 const VkInstanceCreateInfo* pCreateInfo,
1088 const VkAllocationCallbacks* pAllocator,
1089 VkInstance* pInstance)
1090 {
1091 struct anv_instance *instance;
1092 VkResult result;
1093
1094 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
1095
1096 if (pAllocator == NULL)
1097 pAllocator = vk_default_allocator();
1098
1099 instance = vk_alloc(pAllocator, sizeof(*instance), 8,
1100 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
1101 if (!instance)
1102 return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
1103
1104 struct vk_instance_dispatch_table dispatch_table;
1105 vk_instance_dispatch_table_from_entrypoints(
1106 &dispatch_table, &anv_instance_entrypoints, true);
1107 vk_instance_dispatch_table_from_entrypoints(
1108 &dispatch_table, &wsi_instance_entrypoints, false);
1109
1110 result = vk_instance_init(&instance->vk, &instance_extensions,
1111 &dispatch_table, pCreateInfo, pAllocator);
1112 if (result != VK_SUCCESS) {
1113 vk_free(pAllocator, instance);
1114 return vk_error(NULL, result);
1115 }
1116
1117 instance->physical_devices_enumerated = false;
1118 list_inithead(&instance->physical_devices);
1119
1120 instance->pipeline_cache_enabled =
1121 env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
1122
1123 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
1124
1125 anv_init_dri_options(instance);
1126
1127 *pInstance = anv_instance_to_handle(instance);
1128
1129 return VK_SUCCESS;
1130 }
1131
anv_DestroyInstance(VkInstance _instance,const VkAllocationCallbacks * pAllocator)1132 void anv_DestroyInstance(
1133 VkInstance _instance,
1134 const VkAllocationCallbacks* pAllocator)
1135 {
1136 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1137
1138 if (!instance)
1139 return;
1140
1141 list_for_each_entry_safe(struct anv_physical_device, pdevice,
1142 &instance->physical_devices, link)
1143 anv_physical_device_destroy(pdevice);
1144
1145 VG(VALGRIND_DESTROY_MEMPOOL(instance));
1146
1147 driDestroyOptionCache(&instance->dri_options);
1148 driDestroyOptionInfo(&instance->available_dri_options);
1149
1150 vk_instance_finish(&instance->vk);
1151 vk_free(&instance->vk.alloc, instance);
1152 }
1153
1154 static VkResult
anv_enumerate_physical_devices(struct anv_instance * instance)1155 anv_enumerate_physical_devices(struct anv_instance *instance)
1156 {
1157 if (instance->physical_devices_enumerated)
1158 return VK_SUCCESS;
1159
1160 instance->physical_devices_enumerated = true;
1161
1162 /* TODO: Check for more devices ? */
1163 drmDevicePtr devices[8];
1164 int max_devices;
1165
1166 max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
1167 if (max_devices < 1)
1168 return VK_SUCCESS;
1169
1170 VkResult result = VK_SUCCESS;
1171 for (unsigned i = 0; i < (unsigned)max_devices; i++) {
1172 if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
1173 devices[i]->bustype == DRM_BUS_PCI &&
1174 devices[i]->deviceinfo.pci->vendor_id == 0x8086) {
1175
1176 struct anv_physical_device *pdevice;
1177 result = anv_physical_device_try_create(instance, devices[i],
1178 &pdevice);
1179 /* Incompatible DRM device, skip. */
1180 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
1181 result = VK_SUCCESS;
1182 continue;
1183 }
1184
1185 /* Error creating the physical device, report the error. */
1186 if (result != VK_SUCCESS)
1187 break;
1188
1189 list_addtail(&pdevice->link, &instance->physical_devices);
1190 }
1191 }
1192 drmFreeDevices(devices, max_devices);
1193
1194 /* If we successfully enumerated any devices, call it success */
1195 return result;
1196 }
1197
anv_EnumeratePhysicalDevices(VkInstance _instance,uint32_t * pPhysicalDeviceCount,VkPhysicalDevice * pPhysicalDevices)1198 VkResult anv_EnumeratePhysicalDevices(
1199 VkInstance _instance,
1200 uint32_t* pPhysicalDeviceCount,
1201 VkPhysicalDevice* pPhysicalDevices)
1202 {
1203 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1204 VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
1205
1206 VkResult result = anv_enumerate_physical_devices(instance);
1207 if (result != VK_SUCCESS)
1208 return result;
1209
1210 list_for_each_entry(struct anv_physical_device, pdevice,
1211 &instance->physical_devices, link) {
1212 vk_outarray_append(&out, i) {
1213 *i = anv_physical_device_to_handle(pdevice);
1214 }
1215 }
1216
1217 return vk_outarray_status(&out);
1218 }
1219
anv_EnumeratePhysicalDeviceGroups(VkInstance _instance,uint32_t * pPhysicalDeviceGroupCount,VkPhysicalDeviceGroupProperties * pPhysicalDeviceGroupProperties)1220 VkResult anv_EnumeratePhysicalDeviceGroups(
1221 VkInstance _instance,
1222 uint32_t* pPhysicalDeviceGroupCount,
1223 VkPhysicalDeviceGroupProperties* pPhysicalDeviceGroupProperties)
1224 {
1225 ANV_FROM_HANDLE(anv_instance, instance, _instance);
1226 VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
1227 pPhysicalDeviceGroupCount);
1228
1229 VkResult result = anv_enumerate_physical_devices(instance);
1230 if (result != VK_SUCCESS)
1231 return result;
1232
1233 list_for_each_entry(struct anv_physical_device, pdevice,
1234 &instance->physical_devices, link) {
1235 vk_outarray_append(&out, p) {
1236 p->physicalDeviceCount = 1;
1237 memset(p->physicalDevices, 0, sizeof(p->physicalDevices));
1238 p->physicalDevices[0] = anv_physical_device_to_handle(pdevice);
1239 p->subsetAllocation = false;
1240
1241 vk_foreach_struct(ext, p->pNext)
1242 anv_debug_ignored_stype(ext->sType);
1243 }
1244 }
1245
1246 return vk_outarray_status(&out);
1247 }
1248
anv_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,VkPhysicalDeviceFeatures * pFeatures)1249 void anv_GetPhysicalDeviceFeatures(
1250 VkPhysicalDevice physicalDevice,
1251 VkPhysicalDeviceFeatures* pFeatures)
1252 {
1253 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1254
1255 /* Just pick one; they're all the same */
1256 const bool has_astc_ldr =
1257 isl_format_supports_sampling(&pdevice->info,
1258 ISL_FORMAT_ASTC_LDR_2D_4X4_FLT16);
1259
1260 *pFeatures = (VkPhysicalDeviceFeatures) {
1261 .robustBufferAccess = true,
1262 .fullDrawIndexUint32 = true,
1263 .imageCubeArray = true,
1264 .independentBlend = true,
1265 .geometryShader = true,
1266 .tessellationShader = true,
1267 .sampleRateShading = true,
1268 .dualSrcBlend = true,
1269 .logicOp = true,
1270 .multiDrawIndirect = true,
1271 .drawIndirectFirstInstance = true,
1272 .depthClamp = true,
1273 .depthBiasClamp = true,
1274 .fillModeNonSolid = true,
1275 .depthBounds = pdevice->info.ver >= 12,
1276 .wideLines = true,
1277 .largePoints = true,
1278 .alphaToOne = true,
1279 .multiViewport = true,
1280 .samplerAnisotropy = true,
1281 .textureCompressionETC2 = pdevice->info.ver >= 8 ||
1282 pdevice->info.is_baytrail,
1283 .textureCompressionASTC_LDR = has_astc_ldr,
1284 .textureCompressionBC = true,
1285 .occlusionQueryPrecise = true,
1286 .pipelineStatisticsQuery = true,
1287 .fragmentStoresAndAtomics = true,
1288 .shaderTessellationAndGeometryPointSize = true,
1289 .shaderImageGatherExtended = true,
1290 .shaderStorageImageExtendedFormats = true,
1291 .shaderStorageImageMultisample = false,
1292 .shaderStorageImageReadWithoutFormat = false,
1293 .shaderStorageImageWriteWithoutFormat = true,
1294 .shaderUniformBufferArrayDynamicIndexing = true,
1295 .shaderSampledImageArrayDynamicIndexing = true,
1296 .shaderStorageBufferArrayDynamicIndexing = true,
1297 .shaderStorageImageArrayDynamicIndexing = true,
1298 .shaderClipDistance = true,
1299 .shaderCullDistance = true,
1300 .shaderFloat64 = pdevice->info.ver >= 8 &&
1301 pdevice->info.has_64bit_float,
1302 .shaderInt64 = pdevice->info.ver >= 8,
1303 .shaderInt16 = pdevice->info.ver >= 8,
1304 .shaderResourceMinLod = pdevice->info.ver >= 9,
1305 .variableMultisampleRate = true,
1306 .inheritedQueries = true,
1307 };
1308
1309 /* We can't do image stores in vec4 shaders */
1310 pFeatures->vertexPipelineStoresAndAtomics =
1311 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
1312 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
1313
1314 struct vk_app_info *app_info = &pdevice->instance->vk.app_info;
1315
1316 /* The new DOOM and Wolfenstein games require depthBounds without
1317 * checking for it. They seem to run fine without it so just claim it's
1318 * there and accept the consequences.
1319 */
1320 if (app_info->engine_name && strcmp(app_info->engine_name, "idTech") == 0)
1321 pFeatures->depthBounds = true;
1322 }
1323
1324 static void
anv_get_physical_device_features_1_1(struct anv_physical_device * pdevice,VkPhysicalDeviceVulkan11Features * f)1325 anv_get_physical_device_features_1_1(struct anv_physical_device *pdevice,
1326 VkPhysicalDeviceVulkan11Features *f)
1327 {
1328 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES);
1329
1330 f->storageBuffer16BitAccess = pdevice->info.ver >= 8;
1331 f->uniformAndStorageBuffer16BitAccess = pdevice->info.ver >= 8;
1332 f->storagePushConstant16 = pdevice->info.ver >= 8;
1333 f->storageInputOutput16 = false;
1334 f->multiview = true;
1335 f->multiviewGeometryShader = true;
1336 f->multiviewTessellationShader = true;
1337 f->variablePointersStorageBuffer = true;
1338 f->variablePointers = true;
1339 f->protectedMemory = false;
1340 f->samplerYcbcrConversion = true;
1341 f->shaderDrawParameters = true;
1342 }
1343
1344 static void
anv_get_physical_device_features_1_2(struct anv_physical_device * pdevice,VkPhysicalDeviceVulkan12Features * f)1345 anv_get_physical_device_features_1_2(struct anv_physical_device *pdevice,
1346 VkPhysicalDeviceVulkan12Features *f)
1347 {
1348 assert(f->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES);
1349
1350 f->samplerMirrorClampToEdge = true;
1351 f->drawIndirectCount = true;
1352 f->storageBuffer8BitAccess = pdevice->info.ver >= 8;
1353 f->uniformAndStorageBuffer8BitAccess = pdevice->info.ver >= 8;
1354 f->storagePushConstant8 = pdevice->info.ver >= 8;
1355 f->shaderBufferInt64Atomics = pdevice->info.ver >= 9 &&
1356 pdevice->use_softpin;
1357 f->shaderSharedInt64Atomics = false;
1358 f->shaderFloat16 = pdevice->info.ver >= 8;
1359 f->shaderInt8 = pdevice->info.ver >= 8;
1360
1361 bool descIndexing = pdevice->has_a64_buffer_access &&
1362 pdevice->has_bindless_images;
1363 f->descriptorIndexing = descIndexing;
1364 f->shaderInputAttachmentArrayDynamicIndexing = false;
1365 f->shaderUniformTexelBufferArrayDynamicIndexing = descIndexing;
1366 f->shaderStorageTexelBufferArrayDynamicIndexing = descIndexing;
1367 f->shaderUniformBufferArrayNonUniformIndexing = descIndexing;
1368 f->shaderSampledImageArrayNonUniformIndexing = descIndexing;
1369 f->shaderStorageBufferArrayNonUniformIndexing = descIndexing;
1370 f->shaderStorageImageArrayNonUniformIndexing = descIndexing;
1371 f->shaderInputAttachmentArrayNonUniformIndexing = false;
1372 f->shaderUniformTexelBufferArrayNonUniformIndexing = descIndexing;
1373 f->shaderStorageTexelBufferArrayNonUniformIndexing = descIndexing;
1374 f->descriptorBindingUniformBufferUpdateAfterBind = descIndexing;
1375 f->descriptorBindingSampledImageUpdateAfterBind = descIndexing;
1376 f->descriptorBindingStorageImageUpdateAfterBind = descIndexing;
1377 f->descriptorBindingStorageBufferUpdateAfterBind = descIndexing;
1378 f->descriptorBindingUniformTexelBufferUpdateAfterBind = descIndexing;
1379 f->descriptorBindingStorageTexelBufferUpdateAfterBind = descIndexing;
1380 f->descriptorBindingUpdateUnusedWhilePending = descIndexing;
1381 f->descriptorBindingPartiallyBound = descIndexing;
1382 f->descriptorBindingVariableDescriptorCount = descIndexing;
1383 f->runtimeDescriptorArray = descIndexing;
1384
1385 f->samplerFilterMinmax = pdevice->info.ver >= 9;
1386 f->scalarBlockLayout = true;
1387 f->imagelessFramebuffer = true;
1388 f->uniformBufferStandardLayout = true;
1389 f->shaderSubgroupExtendedTypes = true;
1390 f->separateDepthStencilLayouts = true;
1391 f->hostQueryReset = true;
1392 f->timelineSemaphore = true;
1393 f->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1394 f->bufferDeviceAddressCaptureReplay = pdevice->has_a64_buffer_access;
1395 f->bufferDeviceAddressMultiDevice = false;
1396 f->vulkanMemoryModel = true;
1397 f->vulkanMemoryModelDeviceScope = true;
1398 f->vulkanMemoryModelAvailabilityVisibilityChains = true;
1399 f->shaderOutputViewportIndex = true;
1400 f->shaderOutputLayer = true;
1401 f->subgroupBroadcastDynamicId = true;
1402 }
1403
anv_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,VkPhysicalDeviceFeatures2 * pFeatures)1404 void anv_GetPhysicalDeviceFeatures2(
1405 VkPhysicalDevice physicalDevice,
1406 VkPhysicalDeviceFeatures2* pFeatures)
1407 {
1408 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1409 anv_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
1410
1411 VkPhysicalDeviceVulkan11Features core_1_1 = {
1412 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
1413 };
1414 anv_get_physical_device_features_1_1(pdevice, &core_1_1);
1415
1416 VkPhysicalDeviceVulkan12Features core_1_2 = {
1417 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
1418 };
1419 anv_get_physical_device_features_1_2(pdevice, &core_1_2);
1420
1421 vk_foreach_struct(ext, pFeatures->pNext) {
1422 if (vk_get_physical_device_core_1_1_feature_ext(ext, &core_1_1))
1423 continue;
1424 if (vk_get_physical_device_core_1_2_feature_ext(ext, &core_1_2))
1425 continue;
1426
1427 switch (ext->sType) {
1428 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_4444_FORMATS_FEATURES_EXT: {
1429 VkPhysicalDevice4444FormatsFeaturesEXT *features =
1430 (VkPhysicalDevice4444FormatsFeaturesEXT *)ext;
1431 features->formatA4R4G4B4 = true;
1432 features->formatA4B4G4R4 = false;
1433 break;
1434 }
1435
1436
1437 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ACCELERATION_STRUCTURE_FEATURES_KHR: {
1438 VkPhysicalDeviceAccelerationStructureFeaturesKHR *features = (void *)ext;
1439 features->accelerationStructure = false;
1440 features->accelerationStructureCaptureReplay = false;
1441 features->accelerationStructureIndirectBuild = false;
1442 features->accelerationStructureHostCommands = false;
1443 features->descriptorBindingAccelerationStructureUpdateAfterBind = true;
1444 break;
1445 }
1446
1447 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: {
1448 VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext;
1449 features->bufferDeviceAddress = pdevice->has_a64_buffer_access;
1450 features->bufferDeviceAddressCaptureReplay = false;
1451 features->bufferDeviceAddressMultiDevice = false;
1452 break;
1453 }
1454
1455
1456 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COLOR_WRITE_ENABLE_FEATURES_EXT: {
1457 VkPhysicalDeviceColorWriteEnableFeaturesEXT *features =
1458 (VkPhysicalDeviceColorWriteEnableFeaturesEXT *)ext;
1459 features->colorWriteEnable = true;
1460 break;
1461 }
1462
1463 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: {
1464 VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features =
1465 (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext;
1466 features->computeDerivativeGroupQuads = true;
1467 features->computeDerivativeGroupLinear = true;
1468 break;
1469 }
1470
1471 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
1472 VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
1473 (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext;
1474 features->conditionalRendering = pdevice->info.verx10 >= 75;
1475 features->inheritedConditionalRendering = pdevice->info.verx10 >= 75;
1476 break;
1477 }
1478
1479 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT: {
1480 VkPhysicalDeviceCustomBorderColorFeaturesEXT *features =
1481 (VkPhysicalDeviceCustomBorderColorFeaturesEXT *)ext;
1482 features->customBorderColors = pdevice->info.ver >= 8;
1483 features->customBorderColorWithoutFormat = pdevice->info.ver >= 8;
1484 break;
1485 }
1486
1487 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
1488 VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
1489 (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
1490 features->depthClipEnable = true;
1491 break;
1492 }
1493
1494 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: {
1495 VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features =
1496 (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext;
1497 features->fragmentShaderSampleInterlock = pdevice->info.ver >= 9;
1498 features->fragmentShaderPixelInterlock = pdevice->info.ver >= 9;
1499 features->fragmentShaderShadingRateInterlock = false;
1500 break;
1501 }
1502
1503 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_FEATURES_KHR: {
1504 VkPhysicalDeviceFragmentShadingRateFeaturesKHR *features =
1505 (VkPhysicalDeviceFragmentShadingRateFeaturesKHR *)ext;
1506 features->attachmentFragmentShadingRate = false;
1507 features->pipelineFragmentShadingRate = true;
1508 features->primitiveFragmentShadingRate = false;
1509 break;
1510 }
1511
1512 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_ROBUSTNESS_FEATURES_EXT: {
1513 VkPhysicalDeviceImageRobustnessFeaturesEXT *features =
1514 (VkPhysicalDeviceImageRobustnessFeaturesEXT *)ext;
1515 features->robustImageAccess = true;
1516 break;
1517 }
1518
1519 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
1520 VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
1521 (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
1522 features->indexTypeUint8 = true;
1523 break;
1524 }
1525
1526 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
1527 VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
1528 (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
1529 features->inlineUniformBlock = true;
1530 features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
1531 break;
1532 }
1533
1534 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
1535 VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
1536 (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
1537 features->rectangularLines = true;
1538 features->bresenhamLines = true;
1539 /* Support for Smooth lines with MSAA was removed on gfx11. From the
1540 * BSpec section "Multisample ModesState" table for "AA Line Support
1541 * Requirements":
1542 *
1543 * GFX10:BUG:######## NUM_MULTISAMPLES == 1
1544 *
1545 * Fortunately, this isn't a case most people care about.
1546 */
1547 features->smoothLines = pdevice->info.ver < 10;
1548 features->stippledRectangularLines = false;
1549 features->stippledBresenhamLines = true;
1550 features->stippledSmoothLines = false;
1551 break;
1552 }
1553
1554 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES_KHR: {
1555 VkPhysicalDeviceMaintenance4FeaturesKHR *features =
1556 (VkPhysicalDeviceMaintenance4FeaturesKHR *)ext;
1557 features->maintenance4 = true;
1558 break;
1559 }
1560
1561 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PERFORMANCE_QUERY_FEATURES_KHR: {
1562 VkPhysicalDevicePerformanceQueryFeaturesKHR *feature =
1563 (VkPhysicalDevicePerformanceQueryFeaturesKHR *)ext;
1564 feature->performanceCounterQueryPools = true;
1565 /* HW only supports a single configuration at a time. */
1566 feature->performanceCounterMultipleQueryPools = false;
1567 break;
1568 }
1569
1570 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_CREATION_CACHE_CONTROL_FEATURES_EXT: {
1571 VkPhysicalDevicePipelineCreationCacheControlFeaturesEXT *features =
1572 (VkPhysicalDevicePipelineCreationCacheControlFeaturesEXT *)ext;
1573 features->pipelineCreationCacheControl = true;
1574 break;
1575 }
1576
1577 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
1578 VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
1579 (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
1580 features->pipelineExecutableInfo = true;
1581 break;
1582 }
1583
1584 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIVATE_DATA_FEATURES_EXT: {
1585 VkPhysicalDevicePrivateDataFeaturesEXT *features = (void *)ext;
1586 features->privateData = true;
1587 break;
1588 }
1589
1590 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT: {
1591 VkPhysicalDeviceProvokingVertexFeaturesEXT *features =
1592 (VkPhysicalDeviceProvokingVertexFeaturesEXT *)ext;
1593 features->provokingVertexLast = true;
1594 features->transformFeedbackPreservesProvokingVertex = true;
1595 break;
1596 }
1597
1598 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT: {
1599 VkPhysicalDeviceRobustness2FeaturesEXT *features = (void *)ext;
1600 features->robustBufferAccess2 = true;
1601 features->robustImageAccess2 = true;
1602 features->nullDescriptor = true;
1603 break;
1604 }
1605
1606 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_FEATURES_EXT: {
1607 VkPhysicalDeviceShaderAtomicFloatFeaturesEXT *features = (void *)ext;
1608 features->shaderBufferFloat32Atomics = true;
1609 features->shaderBufferFloat32AtomicAdd = pdevice->info.has_lsc;
1610 features->shaderBufferFloat64Atomics = pdevice->info.has_lsc;
1611 features->shaderBufferFloat64AtomicAdd = false;
1612 features->shaderSharedFloat32Atomics = true;
1613 features->shaderSharedFloat32AtomicAdd = false;
1614 features->shaderSharedFloat64Atomics = false;
1615 features->shaderSharedFloat64AtomicAdd = false;
1616 features->shaderImageFloat32Atomics = true;
1617 features->shaderImageFloat32AtomicAdd = false;
1618 features->sparseImageFloat32Atomics = false;
1619 features->sparseImageFloat32AtomicAdd = false;
1620 break;
1621 }
1622
1623 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_FLOAT_2_FEATURES_EXT: {
1624 VkPhysicalDeviceShaderAtomicFloat2FeaturesEXT *features = (void *)ext;
1625 features->shaderBufferFloat16Atomics = false;
1626 features->shaderBufferFloat16AtomicAdd = false;
1627 features->shaderBufferFloat16AtomicMinMax = false;
1628 features->shaderBufferFloat32AtomicMinMax = pdevice->info.ver >= 9;
1629 features->shaderBufferFloat64AtomicMinMax = pdevice->info.has_lsc;
1630 features->shaderSharedFloat16Atomics = false;
1631 features->shaderSharedFloat16AtomicAdd = false;
1632 features->shaderSharedFloat16AtomicMinMax = false;
1633 features->shaderSharedFloat32AtomicMinMax = pdevice->info.ver >= 9;
1634 features->shaderSharedFloat64AtomicMinMax = false;
1635 features->shaderImageFloat32AtomicMinMax = false;
1636 features->sparseImageFloat32AtomicMinMax = false;
1637 break;
1638 }
1639
1640 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: {
1641 VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext;
1642 features->shaderDemoteToHelperInvocation = true;
1643 break;
1644 }
1645
1646 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
1647 VkPhysicalDeviceShaderClockFeaturesKHR *features =
1648 (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
1649 features->shaderSubgroupClock = true;
1650 features->shaderDeviceClock = false;
1651 break;
1652 }
1653
1654 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_FUNCTIONS_2_FEATURES_INTEL: {
1655 VkPhysicalDeviceShaderIntegerFunctions2FeaturesINTEL *features =
1656 (VkPhysicalDeviceShaderIntegerFunctions2FeaturesINTEL *)ext;
1657 features->shaderIntegerFunctions2 = true;
1658 break;
1659 }
1660
1661 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_DOT_PRODUCT_FEATURES_KHR: {
1662 VkPhysicalDeviceShaderIntegerDotProductFeaturesKHR *features =
1663 (VkPhysicalDeviceShaderIntegerDotProductFeaturesKHR *)ext;
1664 features->shaderIntegerDotProduct = true;
1665 break;
1666 };
1667
1668 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_UNIFORM_CONTROL_FLOW_FEATURES_KHR: {
1669 VkPhysicalDeviceShaderSubgroupUniformControlFlowFeaturesKHR *features =
1670 (VkPhysicalDeviceShaderSubgroupUniformControlFlowFeaturesKHR *)ext;
1671 features->shaderSubgroupUniformControlFlow = true;
1672 break;
1673 }
1674
1675 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_TERMINATE_INVOCATION_FEATURES_KHR: {
1676 VkPhysicalDeviceShaderTerminateInvocationFeaturesKHR *features =
1677 (VkPhysicalDeviceShaderTerminateInvocationFeaturesKHR *)ext;
1678 features->shaderTerminateInvocation = true;
1679 break;
1680 }
1681
1682 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
1683 VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
1684 (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
1685 features->subgroupSizeControl = true;
1686 features->computeFullSubgroups = true;
1687 break;
1688 }
1689
1690 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SYNCHRONIZATION_2_FEATURES_KHR: {
1691 VkPhysicalDeviceSynchronization2FeaturesKHR *features =
1692 (VkPhysicalDeviceSynchronization2FeaturesKHR *)ext;
1693 features->synchronization2 = true;
1694 break;
1695 }
1696
1697 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
1698 VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
1699 (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
1700 features->texelBufferAlignment = true;
1701 break;
1702 }
1703
1704 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
1705 VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
1706 (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext;
1707 features->transformFeedback = true;
1708 features->geometryStreams = true;
1709 break;
1710 }
1711
1712 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
1713 VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
1714 (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
1715 features->vertexAttributeInstanceRateDivisor = true;
1716 features->vertexAttributeInstanceRateZeroDivisor = true;
1717 break;
1718 }
1719
1720 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR: {
1721 VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR *features =
1722 (VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR *)ext;
1723 features->workgroupMemoryExplicitLayout = true;
1724 features->workgroupMemoryExplicitLayoutScalarBlockLayout = true;
1725 features->workgroupMemoryExplicitLayout8BitAccess = true;
1726 features->workgroupMemoryExplicitLayout16BitAccess = true;
1727 break;
1728 }
1729
1730 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
1731 VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
1732 (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
1733 features->ycbcrImageArrays = true;
1734 break;
1735 }
1736
1737 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT: {
1738 VkPhysicalDeviceExtendedDynamicStateFeaturesEXT *features =
1739 (VkPhysicalDeviceExtendedDynamicStateFeaturesEXT *)ext;
1740 features->extendedDynamicState = true;
1741 break;
1742 }
1743
1744 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_2_FEATURES_EXT: {
1745 VkPhysicalDeviceExtendedDynamicState2FeaturesEXT *features =
1746 (VkPhysicalDeviceExtendedDynamicState2FeaturesEXT *)ext;
1747 features->extendedDynamicState2 = true;
1748 features->extendedDynamicState2LogicOp = true;
1749 features->extendedDynamicState2PatchControlPoints = false;
1750 break;
1751 }
1752
1753 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ZERO_INITIALIZE_WORKGROUP_MEMORY_FEATURES_KHR: {
1754 VkPhysicalDeviceZeroInitializeWorkgroupMemoryFeaturesKHR *features =
1755 (VkPhysicalDeviceZeroInitializeWorkgroupMemoryFeaturesKHR *)ext;
1756 features->shaderZeroInitializeWorkgroupMemory = true;
1757 break;
1758 }
1759
1760 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_FEATURES_EXT: {
1761 VkPhysicalDeviceMultiDrawFeaturesEXT *features = (VkPhysicalDeviceMultiDrawFeaturesEXT *)ext;
1762 features->multiDraw = true;
1763 break;
1764 }
1765
1766 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIMITIVE_TOPOLOGY_LIST_RESTART_FEATURES_EXT: {
1767 VkPhysicalDevicePrimitiveTopologyListRestartFeaturesEXT *features =
1768 (VkPhysicalDevicePrimitiveTopologyListRestartFeaturesEXT *)ext;
1769 features->primitiveTopologyListRestart = true;
1770 features->primitiveTopologyPatchListRestart = true;
1771 break;
1772 }
1773
1774 default:
1775 anv_debug_ignored_stype(ext->sType);
1776 break;
1777 }
1778 }
1779
1780 }
1781
1782 #define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64
1783
1784 #define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64
1785 #define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256
1786
1787 #define MAX_CUSTOM_BORDER_COLORS 4096
1788
anv_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,VkPhysicalDeviceProperties * pProperties)1789 void anv_GetPhysicalDeviceProperties(
1790 VkPhysicalDevice physicalDevice,
1791 VkPhysicalDeviceProperties* pProperties)
1792 {
1793 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
1794 const struct intel_device_info *devinfo = &pdevice->info;
1795
1796 const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64;
1797 const uint32_t max_textures =
1798 pdevice->has_bindless_images ? UINT16_MAX : 128;
1799 const uint32_t max_samplers =
1800 pdevice->has_bindless_samplers ? UINT16_MAX :
1801 (devinfo->verx10 >= 75) ? 128 : 16;
1802 const uint32_t max_images =
1803 pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
1804
1805 /* If we can use bindless for everything, claim a high per-stage limit,
1806 * otherwise use the binding table size, minus the slots reserved for
1807 * render targets and one slot for the descriptor buffer. */
1808 const uint32_t max_per_stage =
1809 pdevice->has_bindless_images && pdevice->has_a64_buffer_access
1810 ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
1811
1812 const uint32_t max_workgroup_size = 32 * devinfo->max_cs_workgroup_threads;
1813
1814 VkSampleCountFlags sample_counts =
1815 isl_device_get_sample_counts(&pdevice->isl_dev);
1816
1817
1818 VkPhysicalDeviceLimits limits = {
1819 .maxImageDimension1D = (1 << 14),
1820 .maxImageDimension2D = (1 << 14),
1821 .maxImageDimension3D = (1 << 11),
1822 .maxImageDimensionCube = (1 << 14),
1823 .maxImageArrayLayers = (1 << 11),
1824 .maxTexelBufferElements = 128 * 1024 * 1024,
1825 .maxUniformBufferRange = (1ul << 27),
1826 .maxStorageBufferRange = pdevice->isl_dev.max_buffer_size,
1827 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
1828 .maxMemoryAllocationCount = UINT32_MAX,
1829 .maxSamplerAllocationCount = 64 * 1024,
1830 .bufferImageGranularity = 64, /* A cache line */
1831 .sparseAddressSpaceSize = 0,
1832 .maxBoundDescriptorSets = MAX_SETS,
1833 .maxPerStageDescriptorSamplers = max_samplers,
1834 .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS,
1835 .maxPerStageDescriptorStorageBuffers = max_ssbos,
1836 .maxPerStageDescriptorSampledImages = max_textures,
1837 .maxPerStageDescriptorStorageImages = max_images,
1838 .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS,
1839 .maxPerStageResources = max_per_stage,
1840 .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */
1841 .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */
1842 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1843 .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */
1844 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
1845 .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */
1846 .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */
1847 .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS,
1848 .maxVertexInputAttributes = MAX_VBS,
1849 .maxVertexInputBindings = MAX_VBS,
1850 .maxVertexInputAttributeOffset = 2047,
1851 .maxVertexInputBindingStride = 2048,
1852 .maxVertexOutputComponents = 128,
1853 .maxTessellationGenerationLevel = 64,
1854 .maxTessellationPatchSize = 32,
1855 .maxTessellationControlPerVertexInputComponents = 128,
1856 .maxTessellationControlPerVertexOutputComponents = 128,
1857 .maxTessellationControlPerPatchOutputComponents = 128,
1858 .maxTessellationControlTotalOutputComponents = 2048,
1859 .maxTessellationEvaluationInputComponents = 128,
1860 .maxTessellationEvaluationOutputComponents = 128,
1861 .maxGeometryShaderInvocations = 32,
1862 .maxGeometryInputComponents = devinfo->ver >= 8 ? 128 : 64,
1863 .maxGeometryOutputComponents = 128,
1864 .maxGeometryOutputVertices = 256,
1865 .maxGeometryTotalOutputComponents = 1024,
1866 .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */
1867 .maxFragmentOutputAttachments = 8,
1868 .maxFragmentDualSrcAttachments = 1,
1869 .maxFragmentCombinedOutputResources = 8,
1870 .maxComputeSharedMemorySize = 64 * 1024,
1871 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
1872 .maxComputeWorkGroupInvocations = max_workgroup_size,
1873 .maxComputeWorkGroupSize = {
1874 max_workgroup_size,
1875 max_workgroup_size,
1876 max_workgroup_size,
1877 },
1878 .subPixelPrecisionBits = 8,
1879 .subTexelPrecisionBits = 8,
1880 .mipmapPrecisionBits = 8,
1881 .maxDrawIndexedIndexValue = UINT32_MAX,
1882 .maxDrawIndirectCount = UINT32_MAX,
1883 .maxSamplerLodBias = 16,
1884 .maxSamplerAnisotropy = 16,
1885 .maxViewports = MAX_VIEWPORTS,
1886 .maxViewportDimensions = { (1 << 14), (1 << 14) },
1887 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
1888 .viewportSubPixelBits = 13, /* We take a float? */
1889 .minMemoryMapAlignment = 4096, /* A page */
1890 /* The dataport requires texel alignment so we need to assume a worst
1891 * case of R32G32B32A32 which is 16 bytes.
1892 */
1893 .minTexelBufferOffsetAlignment = 16,
1894 .minUniformBufferOffsetAlignment = ANV_UBO_ALIGNMENT,
1895 .minStorageBufferOffsetAlignment = ANV_SSBO_ALIGNMENT,
1896 .minTexelOffset = -8,
1897 .maxTexelOffset = 7,
1898 .minTexelGatherOffset = -32,
1899 .maxTexelGatherOffset = 31,
1900 .minInterpolationOffset = -0.5,
1901 .maxInterpolationOffset = 0.4375,
1902 .subPixelInterpolationOffsetBits = 4,
1903 .maxFramebufferWidth = (1 << 14),
1904 .maxFramebufferHeight = (1 << 14),
1905 .maxFramebufferLayers = (1 << 11),
1906 .framebufferColorSampleCounts = sample_counts,
1907 .framebufferDepthSampleCounts = sample_counts,
1908 .framebufferStencilSampleCounts = sample_counts,
1909 .framebufferNoAttachmentsSampleCounts = sample_counts,
1910 .maxColorAttachments = MAX_RTS,
1911 .sampledImageColorSampleCounts = sample_counts,
1912 .sampledImageIntegerSampleCounts = sample_counts,
1913 .sampledImageDepthSampleCounts = sample_counts,
1914 .sampledImageStencilSampleCounts = sample_counts,
1915 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
1916 .maxSampleMaskWords = 1,
1917 .timestampComputeAndGraphics = true,
1918 .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency,
1919 .maxClipDistances = 8,
1920 .maxCullDistances = 8,
1921 .maxCombinedClipAndCullDistances = 8,
1922 .discreteQueuePriorities = 2,
1923 .pointSizeRange = { 0.125, 255.875 },
1924 /* While SKL and up support much wider lines than we are setting here,
1925 * in practice we run into conformance issues if we go past this limit.
1926 * Since the Windows driver does the same, it's probably fair to assume
1927 * that no one needs more than this.
1928 */
1929 .lineWidthRange = { 0.0, 7.9921875 },
1930 .pointSizeGranularity = (1.0 / 8.0),
1931 .lineWidthGranularity = (1.0 / 128.0),
1932 .strictLines = false,
1933 .standardSampleLocations = true,
1934 .optimalBufferCopyOffsetAlignment = 128,
1935 .optimalBufferCopyRowPitchAlignment = 128,
1936 .nonCoherentAtomSize = 64,
1937 };
1938
1939 *pProperties = (VkPhysicalDeviceProperties) {
1940 .apiVersion = ANV_API_VERSION,
1941 .driverVersion = vk_get_driver_version(),
1942 .vendorID = 0x8086,
1943 .deviceID = pdevice->info.chipset_id,
1944 .deviceType = pdevice->info.has_local_mem ?
1945 VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU :
1946 VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
1947 .limits = limits,
1948 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
1949 };
1950
1951 snprintf(pProperties->deviceName, sizeof(pProperties->deviceName),
1952 "%s", pdevice->info.name);
1953 memcpy(pProperties->pipelineCacheUUID,
1954 pdevice->pipeline_cache_uuid, VK_UUID_SIZE);
1955 }
1956
1957 static void
anv_get_physical_device_properties_1_1(struct anv_physical_device * pdevice,VkPhysicalDeviceVulkan11Properties * p)1958 anv_get_physical_device_properties_1_1(struct anv_physical_device *pdevice,
1959 VkPhysicalDeviceVulkan11Properties *p)
1960 {
1961 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
1962
1963 memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
1964 memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
1965 memset(p->deviceLUID, 0, VK_LUID_SIZE);
1966 p->deviceNodeMask = 0;
1967 p->deviceLUIDValid = false;
1968
1969 p->subgroupSize = BRW_SUBGROUP_SIZE;
1970 VkShaderStageFlags scalar_stages = 0;
1971 for (unsigned stage = 0; stage < MESA_SHADER_STAGES; stage++) {
1972 if (pdevice->compiler->scalar_stage[stage])
1973 scalar_stages |= mesa_to_vk_shader_stage(stage);
1974 }
1975 if (pdevice->vk.supported_extensions.KHR_ray_tracing_pipeline) {
1976 scalar_stages |= MESA_SHADER_RAYGEN |
1977 MESA_SHADER_ANY_HIT |
1978 MESA_SHADER_CLOSEST_HIT |
1979 MESA_SHADER_MISS |
1980 MESA_SHADER_INTERSECTION |
1981 MESA_SHADER_CALLABLE;
1982 }
1983 p->subgroupSupportedStages = scalar_stages;
1984 p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
1985 VK_SUBGROUP_FEATURE_VOTE_BIT |
1986 VK_SUBGROUP_FEATURE_BALLOT_BIT |
1987 VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
1988 VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
1989 VK_SUBGROUP_FEATURE_QUAD_BIT;
1990 if (pdevice->info.ver >= 8) {
1991 /* TODO: There's no technical reason why these can't be made to
1992 * work on gfx7 but they don't at the moment so it's best to leave
1993 * the feature disabled than enabled and broken.
1994 */
1995 p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_ARITHMETIC_BIT |
1996 VK_SUBGROUP_FEATURE_CLUSTERED_BIT;
1997 }
1998 p->subgroupQuadOperationsInAllStages = pdevice->info.ver >= 8;
1999
2000 p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY;
2001 p->maxMultiviewViewCount = 16;
2002 p->maxMultiviewInstanceIndex = UINT32_MAX / 16;
2003 p->protectedNoFault = false;
2004 /* This value doesn't matter for us today as our per-stage descriptors are
2005 * the real limit.
2006 */
2007 p->maxPerSetDescriptors = 1024;
2008 p->maxMemoryAllocationSize = MAX_MEMORY_ALLOCATION_SIZE;
2009 }
2010
2011 static void
anv_get_physical_device_properties_1_2(struct anv_physical_device * pdevice,VkPhysicalDeviceVulkan12Properties * p)2012 anv_get_physical_device_properties_1_2(struct anv_physical_device *pdevice,
2013 VkPhysicalDeviceVulkan12Properties *p)
2014 {
2015 assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
2016
2017 p->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR;
2018 memset(p->driverName, 0, sizeof(p->driverName));
2019 snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR,
2020 "Intel open-source Mesa driver");
2021 memset(p->driverInfo, 0, sizeof(p->driverInfo));
2022 snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR,
2023 "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
2024
2025 /* Don't advertise conformance with a particular version if the hardware's
2026 * support is incomplete/alpha.
2027 */
2028 if (pdevice->is_alpha) {
2029 p->conformanceVersion = (VkConformanceVersionKHR) {
2030 .major = 0,
2031 .minor = 0,
2032 .subminor = 0,
2033 .patch = 0,
2034 };
2035 }
2036 else {
2037 p->conformanceVersion = (VkConformanceVersionKHR) {
2038 .major = 1,
2039 .minor = 2,
2040 .subminor = 0,
2041 .patch = 0,
2042 };
2043 }
2044
2045 p->denormBehaviorIndependence =
2046 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
2047 p->roundingModeIndependence =
2048 VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
2049
2050 /* Broadwell does not support HF denorms and there are restrictions
2051 * other gens. According to Kabylake's PRM:
2052 *
2053 * "math - Extended Math Function
2054 * [...]
2055 * Restriction : Half-float denorms are always retained."
2056 */
2057 p->shaderDenormFlushToZeroFloat16 = false;
2058 p->shaderDenormPreserveFloat16 = pdevice->info.ver > 8;
2059 p->shaderRoundingModeRTEFloat16 = true;
2060 p->shaderRoundingModeRTZFloat16 = true;
2061 p->shaderSignedZeroInfNanPreserveFloat16 = true;
2062
2063 p->shaderDenormFlushToZeroFloat32 = true;
2064 p->shaderDenormPreserveFloat32 = true;
2065 p->shaderRoundingModeRTEFloat32 = true;
2066 p->shaderRoundingModeRTZFloat32 = true;
2067 p->shaderSignedZeroInfNanPreserveFloat32 = true;
2068
2069 p->shaderDenormFlushToZeroFloat64 = true;
2070 p->shaderDenormPreserveFloat64 = true;
2071 p->shaderRoundingModeRTEFloat64 = true;
2072 p->shaderRoundingModeRTZFloat64 = true;
2073 p->shaderSignedZeroInfNanPreserveFloat64 = true;
2074
2075 /* It's a bit hard to exactly map our implementation to the limits
2076 * described by Vulkan. The bindless surface handle in the extended
2077 * message descriptors is 20 bits and it's an index into the table of
2078 * RENDER_SURFACE_STATE structs that starts at bindless surface base
2079 * address. This means that we can have at must 1M surface states
2080 * allocated at any given time. Since most image views take two
2081 * descriptors, this means we have a limit of about 500K image views.
2082 *
2083 * However, since we allocate surface states at vkCreateImageView time,
2084 * this means our limit is actually something on the order of 500K image
2085 * views allocated at any time. The actual limit describe by Vulkan, on
2086 * the other hand, is a limit of how many you can have in a descriptor set.
2087 * Assuming anyone using 1M descriptors will be using the same image view
2088 * twice a bunch of times (or a bunch of null descriptors), we can safely
2089 * advertise a larger limit here.
2090 */
2091 const unsigned max_bindless_views = 1 << 20;
2092 p->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views;
2093 p->shaderUniformBufferArrayNonUniformIndexingNative = false;
2094 p->shaderSampledImageArrayNonUniformIndexingNative = false;
2095 p->shaderStorageBufferArrayNonUniformIndexingNative = true;
2096 p->shaderStorageImageArrayNonUniformIndexingNative = false;
2097 p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
2098 p->robustBufferAccessUpdateAfterBind = true;
2099 p->quadDivergentImplicitLod = false;
2100 p->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views;
2101 p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
2102 p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX;
2103 p->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views;
2104 p->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views;
2105 p->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS;
2106 p->maxPerStageUpdateAfterBindResources = UINT32_MAX;
2107 p->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views;
2108 p->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS;
2109 p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
2110 p->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX;
2111 p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2;
2112 p->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views;
2113 p->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views;
2114 p->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS;
2115
2116 /* We support all of the depth resolve modes */
2117 p->supportedDepthResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR |
2118 VK_RESOLVE_MODE_AVERAGE_BIT_KHR |
2119 VK_RESOLVE_MODE_MIN_BIT_KHR |
2120 VK_RESOLVE_MODE_MAX_BIT_KHR;
2121 /* Average doesn't make sense for stencil so we don't support that */
2122 p->supportedStencilResolveModes = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR;
2123 if (pdevice->info.ver >= 8) {
2124 /* The advanced stencil resolve modes currently require stencil
2125 * sampling be supported by the hardware.
2126 */
2127 p->supportedStencilResolveModes |= VK_RESOLVE_MODE_MIN_BIT_KHR |
2128 VK_RESOLVE_MODE_MAX_BIT_KHR;
2129 }
2130 p->independentResolveNone = true;
2131 p->independentResolve = true;
2132
2133 p->filterMinmaxSingleComponentFormats = pdevice->info.ver >= 9;
2134 p->filterMinmaxImageComponentMapping = pdevice->info.ver >= 9;
2135
2136 p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
2137
2138 p->framebufferIntegerColorSampleCounts =
2139 isl_device_get_sample_counts(&pdevice->isl_dev);
2140 }
2141
anv_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,VkPhysicalDeviceProperties2 * pProperties)2142 void anv_GetPhysicalDeviceProperties2(
2143 VkPhysicalDevice physicalDevice,
2144 VkPhysicalDeviceProperties2* pProperties)
2145 {
2146 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
2147
2148 anv_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
2149
2150 VkPhysicalDeviceVulkan11Properties core_1_1 = {
2151 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
2152 };
2153 anv_get_physical_device_properties_1_1(pdevice, &core_1_1);
2154
2155 VkPhysicalDeviceVulkan12Properties core_1_2 = {
2156 .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
2157 };
2158 anv_get_physical_device_properties_1_2(pdevice, &core_1_2);
2159
2160 vk_foreach_struct(ext, pProperties->pNext) {
2161 if (vk_get_physical_device_core_1_1_property_ext(ext, &core_1_1))
2162 continue;
2163 if (vk_get_physical_device_core_1_2_property_ext(ext, &core_1_2))
2164 continue;
2165
2166 switch (ext->sType) {
2167 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ACCELERATION_STRUCTURE_PROPERTIES_KHR: {
2168 VkPhysicalDeviceAccelerationStructurePropertiesKHR *props = (void *)ext;
2169 props->maxGeometryCount = (1u << 24) - 1;
2170 props->maxInstanceCount = (1u << 24) - 1;
2171 props->maxPrimitiveCount = (1u << 29) - 1;
2172 props->maxPerStageDescriptorAccelerationStructures = UINT16_MAX;
2173 props->maxPerStageDescriptorUpdateAfterBindAccelerationStructures = UINT16_MAX;
2174 props->maxDescriptorSetAccelerationStructures = UINT16_MAX;
2175 props->maxDescriptorSetUpdateAfterBindAccelerationStructures = UINT16_MAX;
2176 props->minAccelerationStructureScratchOffsetAlignment = 64;
2177 break;
2178 }
2179
2180 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONSERVATIVE_RASTERIZATION_PROPERTIES_EXT: {
2181 /* TODO: Real limits */
2182 VkPhysicalDeviceConservativeRasterizationPropertiesEXT *properties =
2183 (VkPhysicalDeviceConservativeRasterizationPropertiesEXT *)ext;
2184 /* There's nothing in the public docs about this value as far as I
2185 * can tell. However, this is the value the Windows driver reports
2186 * and there's a comment on a rejected HW feature in the internal
2187 * docs that says:
2188 *
2189 * "This is similar to conservative rasterization, except the
2190 * primitive area is not extended by 1/512 and..."
2191 *
2192 * That's a bit of an obtuse reference but it's the best we've got
2193 * for now.
2194 */
2195 properties->primitiveOverestimationSize = 1.0f / 512.0f;
2196 properties->maxExtraPrimitiveOverestimationSize = 0.0f;
2197 properties->extraPrimitiveOverestimationSizeGranularity = 0.0f;
2198 properties->primitiveUnderestimation = false;
2199 properties->conservativePointAndLineRasterization = false;
2200 properties->degenerateTrianglesRasterized = true;
2201 properties->degenerateLinesRasterized = false;
2202 properties->fullyCoveredFragmentShaderInputVariable = false;
2203 properties->conservativeRasterizationPostDepthCoverage = true;
2204 break;
2205 }
2206
2207 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_PROPERTIES_EXT: {
2208 VkPhysicalDeviceCustomBorderColorPropertiesEXT *properties =
2209 (VkPhysicalDeviceCustomBorderColorPropertiesEXT *)ext;
2210 properties->maxCustomBorderColorSamplers = MAX_CUSTOM_BORDER_COLORS;
2211 break;
2212 }
2213
2214
2215 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_PROPERTIES_KHR: {
2216 VkPhysicalDeviceFragmentShadingRatePropertiesKHR *props =
2217 (VkPhysicalDeviceFragmentShadingRatePropertiesKHR *)ext;
2218 /* Those must be 0 if attachmentFragmentShadingRate is not
2219 * supported.
2220 */
2221 props->minFragmentShadingRateAttachmentTexelSize = (VkExtent2D) { 0, 0 };
2222 props->maxFragmentShadingRateAttachmentTexelSize = (VkExtent2D) { 0, 0 };
2223 props->maxFragmentShadingRateAttachmentTexelSizeAspectRatio = 0;
2224
2225 props->primitiveFragmentShadingRateWithMultipleViewports = false;
2226 props->layeredShadingRateAttachments = false;
2227 props->fragmentShadingRateNonTrivialCombinerOps = false;
2228 props->maxFragmentSize = (VkExtent2D) { 4, 4 };
2229 props->maxFragmentSizeAspectRatio = 4;
2230 props->maxFragmentShadingRateCoverageSamples = 4 * 4 * 16;
2231 props->maxFragmentShadingRateRasterizationSamples = VK_SAMPLE_COUNT_16_BIT;
2232 props->fragmentShadingRateWithShaderDepthStencilWrites = false;
2233 props->fragmentShadingRateWithSampleMask = true;
2234 props->fragmentShadingRateWithShaderSampleMask = false;
2235 props->fragmentShadingRateWithConservativeRasterization = true;
2236 props->fragmentShadingRateWithFragmentShaderInterlock = true;
2237 props->fragmentShadingRateWithCustomSampleLocations = true;
2238 props->fragmentShadingRateStrictMultiplyCombiner = false;
2239 break;
2240 }
2241
2242 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRM_PROPERTIES_EXT: {
2243 VkPhysicalDeviceDrmPropertiesEXT *props =
2244 (VkPhysicalDeviceDrmPropertiesEXT *)ext;
2245
2246 props->hasPrimary = pdevice->has_master;
2247 props->primaryMajor = pdevice->master_major;
2248 props->primaryMinor = pdevice->master_minor;
2249
2250 props->hasRender = pdevice->has_local;
2251 props->renderMajor = pdevice->local_major;
2252 props->renderMinor = pdevice->local_minor;
2253
2254 break;
2255 }
2256
2257 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_MEMORY_HOST_PROPERTIES_EXT: {
2258 VkPhysicalDeviceExternalMemoryHostPropertiesEXT *props =
2259 (VkPhysicalDeviceExternalMemoryHostPropertiesEXT *) ext;
2260 /* Userptr needs page aligned memory. */
2261 props->minImportedHostPointerAlignment = 4096;
2262 break;
2263 }
2264
2265 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_PROPERTIES_EXT: {
2266 VkPhysicalDeviceInlineUniformBlockPropertiesEXT *props =
2267 (VkPhysicalDeviceInlineUniformBlockPropertiesEXT *)ext;
2268 props->maxInlineUniformBlockSize = MAX_INLINE_UNIFORM_BLOCK_SIZE;
2269 props->maxPerStageDescriptorInlineUniformBlocks =
2270 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
2271 props->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks =
2272 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
2273 props->maxDescriptorSetInlineUniformBlocks =
2274 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
2275 props->maxDescriptorSetUpdateAfterBindInlineUniformBlocks =
2276 MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS;
2277 break;
2278 }
2279
2280 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
2281 VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
2282 (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
2283 /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
2284 * Sampling Rules - Legacy Mode", it says the following:
2285 *
2286 * "Note that the device divides a pixel into a 16x16 array of
2287 * subpixels, referenced by their upper left corners."
2288 *
2289 * This is the only known reference in the PRMs to the subpixel
2290 * precision of line rasterization and a "16x16 array of subpixels"
2291 * implies 4 subpixel precision bits. Empirical testing has shown
2292 * that 4 subpixel precision bits applies to all line rasterization
2293 * types.
2294 */
2295 props->lineSubPixelPrecisionBits = 4;
2296 break;
2297 }
2298
2299 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_PROPERTIES_KHR: {
2300 VkPhysicalDeviceMaintenance4PropertiesKHR *properties =
2301 (VkPhysicalDeviceMaintenance4PropertiesKHR *)ext;
2302 properties->maxBufferSize = pdevice->isl_dev.max_buffer_size;
2303 break;
2304 }
2305
2306 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PCI_BUS_INFO_PROPERTIES_EXT: {
2307 VkPhysicalDevicePCIBusInfoPropertiesEXT *properties =
2308 (VkPhysicalDevicePCIBusInfoPropertiesEXT *)ext;
2309 properties->pciDomain = pdevice->pci_info.domain;
2310 properties->pciBus = pdevice->pci_info.bus;
2311 properties->pciDevice = pdevice->pci_info.device;
2312 properties->pciFunction = pdevice->pci_info.function;
2313 break;
2314 }
2315
2316 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PERFORMANCE_QUERY_PROPERTIES_KHR: {
2317 VkPhysicalDevicePerformanceQueryPropertiesKHR *properties =
2318 (VkPhysicalDevicePerformanceQueryPropertiesKHR *)ext;
2319 /* We could support this by spawning a shader to do the equation
2320 * normalization.
2321 */
2322 properties->allowCommandBufferQueryCopies = false;
2323 break;
2324 }
2325
2326 #pragma GCC diagnostic push
2327 #pragma GCC diagnostic ignored "-Wswitch"
2328 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
2329 VkPhysicalDevicePresentationPropertiesANDROID *props =
2330 (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
2331 props->sharedImage = VK_FALSE;
2332 break;
2333 }
2334 #pragma GCC diagnostic pop
2335
2336 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_PROPERTIES_EXT: {
2337 VkPhysicalDeviceProvokingVertexPropertiesEXT *properties =
2338 (VkPhysicalDeviceProvokingVertexPropertiesEXT *)ext;
2339 properties->provokingVertexModePerPipeline = true;
2340 properties->transformFeedbackPreservesTriangleFanProvokingVertex = false;
2341 break;
2342 }
2343
2344 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
2345 VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
2346 (VkPhysicalDevicePushDescriptorPropertiesKHR *) ext;
2347 properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
2348 break;
2349 }
2350
2351 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_PROPERTIES_EXT: {
2352 VkPhysicalDeviceRobustness2PropertiesEXT *properties = (void *)ext;
2353 properties->robustStorageBufferAccessSizeAlignment =
2354 ANV_SSBO_BOUNDS_CHECK_ALIGNMENT;
2355 properties->robustUniformBufferAccessSizeAlignment =
2356 ANV_UBO_ALIGNMENT;
2357 break;
2358 }
2359
2360 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_DOT_PRODUCT_PROPERTIES_KHR: {
2361 VkPhysicalDeviceShaderIntegerDotProductPropertiesKHR *props =
2362 (VkPhysicalDeviceShaderIntegerDotProductPropertiesKHR *)ext;
2363
2364 props->integerDotProduct8BitUnsignedAccelerated = false;
2365 props->integerDotProduct8BitSignedAccelerated = false;
2366 props->integerDotProduct8BitMixedSignednessAccelerated = false;
2367 props->integerDotProduct4x8BitPackedUnsignedAccelerated = pdevice->info.ver >= 12;
2368 props->integerDotProduct4x8BitPackedSignedAccelerated = pdevice->info.ver >= 12;
2369 props->integerDotProduct4x8BitPackedMixedSignednessAccelerated = pdevice->info.ver >= 12;
2370 props->integerDotProduct16BitUnsignedAccelerated = false;
2371 props->integerDotProduct16BitSignedAccelerated = false;
2372 props->integerDotProduct16BitMixedSignednessAccelerated = false;
2373 props->integerDotProduct32BitUnsignedAccelerated = false;
2374 props->integerDotProduct32BitSignedAccelerated = false;
2375 props->integerDotProduct32BitMixedSignednessAccelerated = false;
2376 props->integerDotProduct64BitUnsignedAccelerated = false;
2377 props->integerDotProduct64BitSignedAccelerated = false;
2378 props->integerDotProduct64BitMixedSignednessAccelerated = false;
2379 props->integerDotProductAccumulatingSaturating8BitUnsignedAccelerated = false;
2380 props->integerDotProductAccumulatingSaturating8BitSignedAccelerated = false;
2381 props->integerDotProductAccumulatingSaturating8BitMixedSignednessAccelerated = false;
2382 props->integerDotProductAccumulatingSaturating4x8BitPackedUnsignedAccelerated = pdevice->info.ver >= 12;
2383 props->integerDotProductAccumulatingSaturating4x8BitPackedSignedAccelerated = pdevice->info.ver >= 12;
2384 props->integerDotProductAccumulatingSaturating4x8BitPackedMixedSignednessAccelerated = pdevice->info.ver >= 12;
2385 props->integerDotProductAccumulatingSaturating16BitUnsignedAccelerated = false;
2386 props->integerDotProductAccumulatingSaturating16BitSignedAccelerated = false;
2387 props->integerDotProductAccumulatingSaturating16BitMixedSignednessAccelerated = false;
2388 props->integerDotProductAccumulatingSaturating32BitUnsignedAccelerated = false;
2389 props->integerDotProductAccumulatingSaturating32BitSignedAccelerated = false;
2390 props->integerDotProductAccumulatingSaturating32BitMixedSignednessAccelerated = false;
2391 props->integerDotProductAccumulatingSaturating64BitUnsignedAccelerated = false;
2392 props->integerDotProductAccumulatingSaturating64BitSignedAccelerated = false;
2393 props->integerDotProductAccumulatingSaturating64BitMixedSignednessAccelerated = false;
2394
2395 break;
2396 }
2397
2398 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: {
2399 VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props =
2400 (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext;
2401 STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32);
2402 props->minSubgroupSize = 8;
2403 props->maxSubgroupSize = 32;
2404 props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_workgroup_threads;
2405 props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
2406 break;
2407 }
2408
2409 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT: {
2410 VkPhysicalDeviceSampleLocationsPropertiesEXT *props =
2411 (VkPhysicalDeviceSampleLocationsPropertiesEXT *)ext;
2412
2413 props->sampleLocationSampleCounts =
2414 isl_device_get_sample_counts(&pdevice->isl_dev);
2415
2416 /* See also anv_GetPhysicalDeviceMultisamplePropertiesEXT */
2417 props->maxSampleLocationGridSize.width = 1;
2418 props->maxSampleLocationGridSize.height = 1;
2419
2420 props->sampleLocationCoordinateRange[0] = 0;
2421 props->sampleLocationCoordinateRange[1] = 0.9375;
2422 props->sampleLocationSubPixelBits = 4;
2423
2424 props->variableSampleLocations = true;
2425 break;
2426 }
2427
2428 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
2429 VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
2430 (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext;
2431
2432 /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface
2433 * Base Address:
2434 *
2435 * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field
2436 * specifies the base address of the first element of the surface,
2437 * computed in software by adding the surface base address to the
2438 * byte offset of the element in the buffer. The base address must
2439 * be aligned to element size."
2440 *
2441 * The typed dataport messages require that things be texel aligned.
2442 * Otherwise, we may just load/store the wrong data or, in the worst
2443 * case, there may be hangs.
2444 */
2445 props->storageTexelBufferOffsetAlignmentBytes = 16;
2446 props->storageTexelBufferOffsetSingleTexelAlignment = true;
2447
2448 /* The sampler, however, is much more forgiving and it can handle
2449 * arbitrary byte alignment for linear and buffer surfaces. It's
2450 * hard to find a good PRM citation for this but years of empirical
2451 * experience demonstrate that this is true.
2452 */
2453 props->uniformTexelBufferOffsetAlignmentBytes = 1;
2454 props->uniformTexelBufferOffsetSingleTexelAlignment = false;
2455 break;
2456 }
2457
2458 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
2459 VkPhysicalDeviceTransformFeedbackPropertiesEXT *props =
2460 (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
2461
2462 props->maxTransformFeedbackStreams = MAX_XFB_STREAMS;
2463 props->maxTransformFeedbackBuffers = MAX_XFB_BUFFERS;
2464 props->maxTransformFeedbackBufferSize = (1ull << 32);
2465 props->maxTransformFeedbackStreamDataSize = 128 * 4;
2466 props->maxTransformFeedbackBufferDataSize = 128 * 4;
2467 props->maxTransformFeedbackBufferDataStride = 2048;
2468 props->transformFeedbackQueries = true;
2469 props->transformFeedbackStreamsLinesTriangles = false;
2470 props->transformFeedbackRasterizationStreamSelect = false;
2471 /* This requires MI_MATH */
2472 props->transformFeedbackDraw = pdevice->info.verx10 >= 75;
2473 break;
2474 }
2475
2476 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
2477 VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
2478 (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
2479 /* We have to restrict this a bit for multiview */
2480 props->maxVertexAttribDivisor = UINT32_MAX / 16;
2481 break;
2482 }
2483
2484 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_PROPERTIES_EXT: {
2485 VkPhysicalDeviceMultiDrawPropertiesEXT *props = (VkPhysicalDeviceMultiDrawPropertiesEXT *)ext;
2486 props->maxMultiDrawCount = 2048;
2487 break;
2488 }
2489
2490 default:
2491 anv_debug_ignored_stype(ext->sType);
2492 break;
2493 }
2494 }
2495 }
2496
2497 static const VkQueueFamilyProperties
2498 anv_queue_family_properties_template = {
2499 .timestampValidBits = 36, /* XXX: Real value here */
2500 .minImageTransferGranularity = { 1, 1, 1 },
2501 };
2502
anv_GetPhysicalDeviceQueueFamilyProperties(VkPhysicalDevice physicalDevice,uint32_t * pCount,VkQueueFamilyProperties * pQueueFamilyProperties)2503 void anv_GetPhysicalDeviceQueueFamilyProperties(
2504 VkPhysicalDevice physicalDevice,
2505 uint32_t* pCount,
2506 VkQueueFamilyProperties* pQueueFamilyProperties)
2507 {
2508 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
2509 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pCount);
2510
2511 for (uint32_t i = 0; i < pdevice->queue.family_count; i++) {
2512 struct anv_queue_family *queue_family = &pdevice->queue.families[i];
2513 vk_outarray_append(&out, p) {
2514 *p = anv_queue_family_properties_template;
2515 p->queueFlags = queue_family->queueFlags;
2516 p->queueCount = queue_family->queueCount;
2517 }
2518 }
2519 }
2520
anv_GetPhysicalDeviceQueueFamilyProperties2(VkPhysicalDevice physicalDevice,uint32_t * pQueueFamilyPropertyCount,VkQueueFamilyProperties2 * pQueueFamilyProperties)2521 void anv_GetPhysicalDeviceQueueFamilyProperties2(
2522 VkPhysicalDevice physicalDevice,
2523 uint32_t* pQueueFamilyPropertyCount,
2524 VkQueueFamilyProperties2* pQueueFamilyProperties)
2525 {
2526 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
2527 VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
2528
2529 for (uint32_t i = 0; i < pdevice->queue.family_count; i++) {
2530 struct anv_queue_family *queue_family = &pdevice->queue.families[i];
2531 vk_outarray_append(&out, p) {
2532 p->queueFamilyProperties = anv_queue_family_properties_template;
2533 p->queueFamilyProperties.queueFlags = queue_family->queueFlags;
2534 p->queueFamilyProperties.queueCount = queue_family->queueCount;
2535
2536 vk_foreach_struct(s, p->pNext) {
2537 anv_debug_ignored_stype(s->sType);
2538 }
2539 }
2540 }
2541 }
2542
anv_GetPhysicalDeviceMemoryProperties(VkPhysicalDevice physicalDevice,VkPhysicalDeviceMemoryProperties * pMemoryProperties)2543 void anv_GetPhysicalDeviceMemoryProperties(
2544 VkPhysicalDevice physicalDevice,
2545 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
2546 {
2547 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2548
2549 pMemoryProperties->memoryTypeCount = physical_device->memory.type_count;
2550 for (uint32_t i = 0; i < physical_device->memory.type_count; i++) {
2551 pMemoryProperties->memoryTypes[i] = (VkMemoryType) {
2552 .propertyFlags = physical_device->memory.types[i].propertyFlags,
2553 .heapIndex = physical_device->memory.types[i].heapIndex,
2554 };
2555 }
2556
2557 pMemoryProperties->memoryHeapCount = physical_device->memory.heap_count;
2558 for (uint32_t i = 0; i < physical_device->memory.heap_count; i++) {
2559 pMemoryProperties->memoryHeaps[i] = (VkMemoryHeap) {
2560 .size = physical_device->memory.heaps[i].size,
2561 .flags = physical_device->memory.heaps[i].flags,
2562 };
2563 }
2564 }
2565
2566 static void
anv_get_memory_budget(VkPhysicalDevice physicalDevice,VkPhysicalDeviceMemoryBudgetPropertiesEXT * memoryBudget)2567 anv_get_memory_budget(VkPhysicalDevice physicalDevice,
2568 VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget)
2569 {
2570 ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice);
2571
2572 anv_update_meminfo(device, device->local_fd);
2573
2574 VkDeviceSize total_sys_heaps_size = 0, total_vram_heaps_size = 0;
2575 for (size_t i = 0; i < device->memory.heap_count; i++) {
2576 if (device->memory.heaps[i].is_local_mem) {
2577 total_vram_heaps_size += device->memory.heaps[i].size;
2578 } else {
2579 total_sys_heaps_size += device->memory.heaps[i].size;
2580 }
2581 }
2582
2583 for (size_t i = 0; i < device->memory.heap_count; i++) {
2584 VkDeviceSize heap_size = device->memory.heaps[i].size;
2585 VkDeviceSize heap_used = device->memory.heaps[i].used;
2586 VkDeviceSize heap_budget, total_heaps_size;
2587 uint64_t mem_available = 0;
2588
2589 if (device->memory.heaps[i].is_local_mem) {
2590 total_heaps_size = total_vram_heaps_size;
2591 mem_available = device->vram.available;
2592 } else {
2593 total_heaps_size = total_sys_heaps_size;
2594 mem_available = device->sys.available;
2595 }
2596
2597 double heap_proportion = (double) heap_size / total_heaps_size;
2598 VkDeviceSize available_prop = mem_available * heap_proportion;
2599
2600 /*
2601 * Let's not incite the app to starve the system: report at most 90% of
2602 * the available heap memory.
2603 */
2604 uint64_t heap_available = available_prop * 9 / 10;
2605 heap_budget = MIN2(heap_size, heap_used + heap_available);
2606
2607 /*
2608 * Round down to the nearest MB
2609 */
2610 heap_budget &= ~((1ull << 20) - 1);
2611
2612 /*
2613 * The heapBudget value must be non-zero for array elements less than
2614 * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget
2615 * value must be less than or equal to VkMemoryHeap::size for each heap.
2616 */
2617 assert(0 < heap_budget && heap_budget <= heap_size);
2618
2619 memoryBudget->heapUsage[i] = heap_used;
2620 memoryBudget->heapBudget[i] = heap_budget;
2621 }
2622
2623 /* The heapBudget and heapUsage values must be zero for array elements
2624 * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
2625 */
2626 for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) {
2627 memoryBudget->heapBudget[i] = 0;
2628 memoryBudget->heapUsage[i] = 0;
2629 }
2630 }
2631
anv_GetPhysicalDeviceMemoryProperties2(VkPhysicalDevice physicalDevice,VkPhysicalDeviceMemoryProperties2 * pMemoryProperties)2632 void anv_GetPhysicalDeviceMemoryProperties2(
2633 VkPhysicalDevice physicalDevice,
2634 VkPhysicalDeviceMemoryProperties2* pMemoryProperties)
2635 {
2636 anv_GetPhysicalDeviceMemoryProperties(physicalDevice,
2637 &pMemoryProperties->memoryProperties);
2638
2639 vk_foreach_struct(ext, pMemoryProperties->pNext) {
2640 switch (ext->sType) {
2641 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT:
2642 anv_get_memory_budget(physicalDevice, (void*)ext);
2643 break;
2644 default:
2645 anv_debug_ignored_stype(ext->sType);
2646 break;
2647 }
2648 }
2649 }
2650
2651 void
anv_GetDeviceGroupPeerMemoryFeatures(VkDevice device,uint32_t heapIndex,uint32_t localDeviceIndex,uint32_t remoteDeviceIndex,VkPeerMemoryFeatureFlags * pPeerMemoryFeatures)2652 anv_GetDeviceGroupPeerMemoryFeatures(
2653 VkDevice device,
2654 uint32_t heapIndex,
2655 uint32_t localDeviceIndex,
2656 uint32_t remoteDeviceIndex,
2657 VkPeerMemoryFeatureFlags* pPeerMemoryFeatures)
2658 {
2659 assert(localDeviceIndex == 0 && remoteDeviceIndex == 0);
2660 *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
2661 VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
2662 VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
2663 VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
2664 }
2665
anv_GetInstanceProcAddr(VkInstance _instance,const char * pName)2666 PFN_vkVoidFunction anv_GetInstanceProcAddr(
2667 VkInstance _instance,
2668 const char* pName)
2669 {
2670 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2671 return vk_instance_get_proc_addr(&instance->vk,
2672 &anv_instance_entrypoints,
2673 pName);
2674 }
2675
2676 /* With version 1+ of the loader interface the ICD should expose
2677 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
2678 */
2679 PUBLIC
2680 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2681 VkInstance instance,
2682 const char* pName);
2683
2684 PUBLIC
vk_icdGetInstanceProcAddr(VkInstance instance,const char * pName)2685 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
2686 VkInstance instance,
2687 const char* pName)
2688 {
2689 return anv_GetInstanceProcAddr(instance, pName);
2690 }
2691
2692 /* With version 4+ of the loader interface the ICD should expose
2693 * vk_icdGetPhysicalDeviceProcAddr()
2694 */
2695 PUBLIC
2696 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
2697 VkInstance _instance,
2698 const char* pName);
2699
vk_icdGetPhysicalDeviceProcAddr(VkInstance _instance,const char * pName)2700 PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
2701 VkInstance _instance,
2702 const char* pName)
2703 {
2704 ANV_FROM_HANDLE(anv_instance, instance, _instance);
2705 return vk_instance_get_physical_device_proc_addr(&instance->vk, pName);
2706 }
2707
2708 static struct anv_state
anv_state_pool_emit_data(struct anv_state_pool * pool,size_t size,size_t align,const void * p)2709 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
2710 {
2711 struct anv_state state;
2712
2713 state = anv_state_pool_alloc(pool, size, align);
2714 memcpy(state.map, p, size);
2715
2716 return state;
2717 }
2718
2719 static void
anv_device_init_border_colors(struct anv_device * device)2720 anv_device_init_border_colors(struct anv_device *device)
2721 {
2722 if (device->info.is_haswell) {
2723 static const struct hsw_border_color border_colors[] = {
2724 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2725 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2726 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2727 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2728 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2729 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2730 };
2731
2732 device->border_colors =
2733 anv_state_pool_emit_data(&device->dynamic_state_pool,
2734 sizeof(border_colors), 512, border_colors);
2735 } else {
2736 static const struct gfx8_border_color border_colors[] = {
2737 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
2738 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
2739 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
2740 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
2741 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
2742 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
2743 };
2744
2745 device->border_colors =
2746 anv_state_pool_emit_data(&device->dynamic_state_pool,
2747 sizeof(border_colors), 64, border_colors);
2748 }
2749 }
2750
2751 static VkResult
anv_device_init_trivial_batch(struct anv_device * device)2752 anv_device_init_trivial_batch(struct anv_device *device)
2753 {
2754 VkResult result = anv_device_alloc_bo(device, "trivial-batch", 4096,
2755 ANV_BO_ALLOC_MAPPED,
2756 0 /* explicit_address */,
2757 &device->trivial_batch_bo);
2758 if (result != VK_SUCCESS)
2759 return result;
2760
2761 struct anv_batch batch = {
2762 .start = device->trivial_batch_bo->map,
2763 .next = device->trivial_batch_bo->map,
2764 .end = device->trivial_batch_bo->map + 4096,
2765 };
2766
2767 anv_batch_emit(&batch, GFX7_MI_BATCH_BUFFER_END, bbe);
2768 anv_batch_emit(&batch, GFX7_MI_NOOP, noop);
2769
2770 if (!device->info.has_llc)
2771 intel_clflush_range(batch.start, batch.next - batch.start);
2772
2773 return VK_SUCCESS;
2774 }
2775
2776 static int
vk_priority_to_gen(int priority)2777 vk_priority_to_gen(int priority)
2778 {
2779 switch (priority) {
2780 case VK_QUEUE_GLOBAL_PRIORITY_LOW_EXT:
2781 return INTEL_CONTEXT_LOW_PRIORITY;
2782 case VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT:
2783 return INTEL_CONTEXT_MEDIUM_PRIORITY;
2784 case VK_QUEUE_GLOBAL_PRIORITY_HIGH_EXT:
2785 return INTEL_CONTEXT_HIGH_PRIORITY;
2786 case VK_QUEUE_GLOBAL_PRIORITY_REALTIME_EXT:
2787 return INTEL_CONTEXT_REALTIME_PRIORITY;
2788 default:
2789 unreachable("Invalid priority");
2790 }
2791 }
2792
2793 static bool
get_bo_from_pool(struct intel_batch_decode_bo * ret,struct anv_block_pool * pool,uint64_t address)2794 get_bo_from_pool(struct intel_batch_decode_bo *ret,
2795 struct anv_block_pool *pool,
2796 uint64_t address)
2797 {
2798 anv_block_pool_foreach_bo(bo, pool) {
2799 uint64_t bo_address = intel_48b_address(bo->offset);
2800 if (address >= bo_address && address < (bo_address + bo->size)) {
2801 *ret = (struct intel_batch_decode_bo) {
2802 .addr = bo_address,
2803 .size = bo->size,
2804 .map = bo->map,
2805 };
2806 return true;
2807 }
2808 }
2809 return false;
2810 }
2811
2812 /* Finding a buffer for batch decoding */
2813 static struct intel_batch_decode_bo
decode_get_bo(void * v_batch,bool ppgtt,uint64_t address)2814 decode_get_bo(void *v_batch, bool ppgtt, uint64_t address)
2815 {
2816 struct anv_device *device = v_batch;
2817 struct intel_batch_decode_bo ret_bo = {};
2818
2819 assert(ppgtt);
2820
2821 if (get_bo_from_pool(&ret_bo, &device->dynamic_state_pool.block_pool, address))
2822 return ret_bo;
2823 if (get_bo_from_pool(&ret_bo, &device->instruction_state_pool.block_pool, address))
2824 return ret_bo;
2825 if (get_bo_from_pool(&ret_bo, &device->binding_table_pool.block_pool, address))
2826 return ret_bo;
2827 if (get_bo_from_pool(&ret_bo, &device->surface_state_pool.block_pool, address))
2828 return ret_bo;
2829
2830 if (!device->cmd_buffer_being_decoded)
2831 return (struct intel_batch_decode_bo) { };
2832
2833 struct anv_batch_bo **bo;
2834
2835 u_vector_foreach(bo, &device->cmd_buffer_being_decoded->seen_bbos) {
2836 /* The decoder zeroes out the top 16 bits, so we need to as well */
2837 uint64_t bo_address = (*bo)->bo->offset & (~0ull >> 16);
2838
2839 if (address >= bo_address && address < bo_address + (*bo)->bo->size) {
2840 return (struct intel_batch_decode_bo) {
2841 .addr = bo_address,
2842 .size = (*bo)->bo->size,
2843 .map = (*bo)->bo->map,
2844 };
2845 }
2846 }
2847
2848 return (struct intel_batch_decode_bo) { };
2849 }
2850
2851 struct intel_aux_map_buffer {
2852 struct intel_buffer base;
2853 struct anv_state state;
2854 };
2855
2856 static struct intel_buffer *
intel_aux_map_buffer_alloc(void * driver_ctx,uint32_t size)2857 intel_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
2858 {
2859 struct intel_aux_map_buffer *buf = malloc(sizeof(struct intel_aux_map_buffer));
2860 if (!buf)
2861 return NULL;
2862
2863 struct anv_device *device = (struct anv_device*)driver_ctx;
2864 assert(device->physical->supports_48bit_addresses &&
2865 device->physical->use_softpin);
2866
2867 struct anv_state_pool *pool = &device->dynamic_state_pool;
2868 buf->state = anv_state_pool_alloc(pool, size, size);
2869
2870 buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
2871 buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
2872 buf->base.map = buf->state.map;
2873 buf->base.driver_bo = &buf->state;
2874 return &buf->base;
2875 }
2876
2877 static void
intel_aux_map_buffer_free(void * driver_ctx,struct intel_buffer * buffer)2878 intel_aux_map_buffer_free(void *driver_ctx, struct intel_buffer *buffer)
2879 {
2880 struct intel_aux_map_buffer *buf = (struct intel_aux_map_buffer*)buffer;
2881 struct anv_device *device = (struct anv_device*)driver_ctx;
2882 struct anv_state_pool *pool = &device->dynamic_state_pool;
2883 anv_state_pool_free(pool, buf->state);
2884 free(buf);
2885 }
2886
2887 static struct intel_mapped_pinned_buffer_alloc aux_map_allocator = {
2888 .alloc = intel_aux_map_buffer_alloc,
2889 .free = intel_aux_map_buffer_free,
2890 };
2891
anv_CreateDevice(VkPhysicalDevice physicalDevice,const VkDeviceCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkDevice * pDevice)2892 VkResult anv_CreateDevice(
2893 VkPhysicalDevice physicalDevice,
2894 const VkDeviceCreateInfo* pCreateInfo,
2895 const VkAllocationCallbacks* pAllocator,
2896 VkDevice* pDevice)
2897 {
2898 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
2899 VkResult result;
2900 struct anv_device *device;
2901
2902 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
2903
2904 /* Check enabled features */
2905 bool robust_buffer_access = false;
2906 if (pCreateInfo->pEnabledFeatures) {
2907 if (pCreateInfo->pEnabledFeatures->robustBufferAccess)
2908 robust_buffer_access = true;
2909 }
2910
2911 vk_foreach_struct_const(ext, pCreateInfo->pNext) {
2912 switch (ext->sType) {
2913 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2: {
2914 const VkPhysicalDeviceFeatures2 *features = (const void *)ext;
2915 if (features->features.robustBufferAccess)
2916 robust_buffer_access = true;
2917 break;
2918 }
2919
2920 default:
2921 /* Don't warn */
2922 break;
2923 }
2924 }
2925
2926 /* Check requested queues and fail if we are requested to create any
2927 * queues with flags we don't support.
2928 */
2929 assert(pCreateInfo->queueCreateInfoCount > 0);
2930 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2931 if (pCreateInfo->pQueueCreateInfos[i].flags != 0)
2932 return vk_error(physical_device, VK_ERROR_INITIALIZATION_FAILED);
2933 }
2934
2935 /* Check if client specified queue priority. */
2936 const VkDeviceQueueGlobalPriorityCreateInfoEXT *queue_priority =
2937 vk_find_struct_const(pCreateInfo->pQueueCreateInfos[0].pNext,
2938 DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
2939
2940 VkQueueGlobalPriorityEXT priority =
2941 queue_priority ? queue_priority->globalPriority :
2942 VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT;
2943
2944 device = vk_zalloc2(&physical_device->instance->vk.alloc, pAllocator,
2945 sizeof(*device), 8,
2946 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
2947 if (!device)
2948 return vk_error(physical_device, VK_ERROR_OUT_OF_HOST_MEMORY);
2949
2950 struct vk_device_dispatch_table dispatch_table;
2951 vk_device_dispatch_table_from_entrypoints(&dispatch_table,
2952 anv_genX(&physical_device->info, device_entrypoints), true);
2953 vk_device_dispatch_table_from_entrypoints(&dispatch_table,
2954 &anv_device_entrypoints, false);
2955 vk_device_dispatch_table_from_entrypoints(&dispatch_table,
2956 &wsi_device_entrypoints, false);
2957
2958 result = vk_device_init(&device->vk, &physical_device->vk,
2959 &dispatch_table, pCreateInfo, pAllocator);
2960 if (result != VK_SUCCESS)
2961 goto fail_alloc;
2962
2963 if (INTEL_DEBUG(DEBUG_BATCH)) {
2964 const unsigned decode_flags =
2965 INTEL_BATCH_DECODE_FULL |
2966 (INTEL_DEBUG(DEBUG_COLOR) ? INTEL_BATCH_DECODE_IN_COLOR : 0) |
2967 INTEL_BATCH_DECODE_OFFSETS |
2968 INTEL_BATCH_DECODE_FLOATS;
2969
2970 intel_batch_decode_ctx_init(&device->decoder_ctx,
2971 &physical_device->info,
2972 stderr, decode_flags, NULL,
2973 decode_get_bo, NULL, device);
2974 }
2975
2976 device->physical = physical_device;
2977 device->_lost = false;
2978
2979 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
2980 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
2981 if (device->fd == -1) {
2982 result = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
2983 goto fail_device;
2984 }
2985
2986 uint32_t num_queues = 0;
2987 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++)
2988 num_queues += pCreateInfo->pQueueCreateInfos[i].queueCount;
2989
2990 if (device->physical->engine_info) {
2991 /* The kernel API supports at most 64 engines */
2992 assert(num_queues <= 64);
2993 uint16_t engine_classes[64];
2994 int engine_count = 0;
2995 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
2996 const VkDeviceQueueCreateInfo *queueCreateInfo =
2997 &pCreateInfo->pQueueCreateInfos[i];
2998
2999 assert(queueCreateInfo->queueFamilyIndex <
3000 physical_device->queue.family_count);
3001 struct anv_queue_family *queue_family =
3002 &physical_device->queue.families[queueCreateInfo->queueFamilyIndex];
3003
3004 for (uint32_t j = 0; j < queueCreateInfo->queueCount; j++)
3005 engine_classes[engine_count++] = queue_family->engine_class;
3006 }
3007 device->context_id =
3008 anv_gem_create_context_engines(device,
3009 physical_device->engine_info,
3010 engine_count, engine_classes);
3011 } else {
3012 assert(num_queues == 1);
3013 device->context_id = anv_gem_create_context(device);
3014 }
3015 if (device->context_id == -1) {
3016 result = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
3017 goto fail_fd;
3018 }
3019
3020 /* Here we tell the kernel not to attempt to recover our context but
3021 * immediately (on the next batchbuffer submission) report that the
3022 * context is lost, and we will do the recovery ourselves. In the case
3023 * of Vulkan, recovery means throwing VK_ERROR_DEVICE_LOST and letting
3024 * the client clean up the pieces.
3025 */
3026 anv_gem_set_context_param(device->fd, device->context_id,
3027 I915_CONTEXT_PARAM_RECOVERABLE, false);
3028
3029 device->has_thread_submit = physical_device->has_thread_submit;
3030
3031 device->queues =
3032 vk_zalloc(&device->vk.alloc, num_queues * sizeof(*device->queues), 8,
3033 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
3034 if (device->queues == NULL) {
3035 result = vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
3036 goto fail_context_id;
3037 }
3038
3039 device->queue_count = 0;
3040 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
3041 const VkDeviceQueueCreateInfo *queueCreateInfo =
3042 &pCreateInfo->pQueueCreateInfos[i];
3043
3044 for (uint32_t j = 0; j < queueCreateInfo->queueCount; j++) {
3045 /* When using legacy contexts, we use I915_EXEC_RENDER but, with
3046 * engine-based contexts, the bottom 6 bits of exec_flags are used
3047 * for the engine ID.
3048 */
3049 uint32_t exec_flags = device->physical->engine_info ?
3050 device->queue_count : I915_EXEC_RENDER;
3051
3052 result = anv_queue_init(device, &device->queues[device->queue_count],
3053 exec_flags, queueCreateInfo, j);
3054 if (result != VK_SUCCESS)
3055 goto fail_queues;
3056
3057 device->queue_count++;
3058 }
3059 }
3060
3061 if (physical_device->use_softpin) {
3062 if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
3063 result = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
3064 goto fail_queues;
3065 }
3066
3067 /* keep the page with address zero out of the allocator */
3068 util_vma_heap_init(&device->vma_lo,
3069 LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE);
3070
3071 util_vma_heap_init(&device->vma_cva, CLIENT_VISIBLE_HEAP_MIN_ADDRESS,
3072 CLIENT_VISIBLE_HEAP_SIZE);
3073
3074 /* Leave the last 4GiB out of the high vma range, so that no state
3075 * base address + size can overflow 48 bits. For more information see
3076 * the comment about Wa32bitGeneralStateOffset in anv_allocator.c
3077 */
3078 util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS,
3079 physical_device->gtt_size - (1ull << 32) -
3080 HIGH_HEAP_MIN_ADDRESS);
3081 }
3082
3083 list_inithead(&device->memory_objects);
3084
3085 /* As per spec, the driver implementation may deny requests to acquire
3086 * a priority above the default priority (MEDIUM) if the caller does not
3087 * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT
3088 * is returned.
3089 */
3090 if (physical_device->has_context_priority) {
3091 int err = anv_gem_set_context_param(device->fd, device->context_id,
3092 I915_CONTEXT_PARAM_PRIORITY,
3093 vk_priority_to_gen(priority));
3094 if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
3095 result = vk_error(device, VK_ERROR_NOT_PERMITTED_EXT);
3096 goto fail_vmas;
3097 }
3098 }
3099
3100 device->info = physical_device->info;
3101 device->isl_dev = physical_device->isl_dev;
3102
3103 /* On Broadwell and later, we can use batch chaining to more efficiently
3104 * implement growing command buffers. Prior to Haswell, the kernel
3105 * command parser gets in the way and we have to fall back to growing
3106 * the batch.
3107 */
3108 device->can_chain_batches = device->info.ver >= 8;
3109
3110 device->robust_buffer_access = robust_buffer_access;
3111
3112 if (pthread_mutex_init(&device->mutex, NULL) != 0) {
3113 result = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
3114 goto fail_queues;
3115 }
3116
3117 pthread_condattr_t condattr;
3118 if (pthread_condattr_init(&condattr) != 0) {
3119 result = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
3120 goto fail_mutex;
3121 }
3122 if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
3123 pthread_condattr_destroy(&condattr);
3124 result = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
3125 goto fail_mutex;
3126 }
3127 if (pthread_cond_init(&device->queue_submit, &condattr) != 0) {
3128 pthread_condattr_destroy(&condattr);
3129 result = vk_error(device, VK_ERROR_INITIALIZATION_FAILED);
3130 goto fail_mutex;
3131 }
3132 pthread_condattr_destroy(&condattr);
3133
3134 result = anv_bo_cache_init(&device->bo_cache, device);
3135 if (result != VK_SUCCESS)
3136 goto fail_queue_cond;
3137
3138 anv_bo_pool_init(&device->batch_bo_pool, device, "batch");
3139
3140 /* Because scratch is also relative to General State Base Address, we leave
3141 * the base address 0 and start the pool memory at an offset. This way we
3142 * get the correct offsets in the anv_states that get allocated from it.
3143 */
3144 result = anv_state_pool_init(&device->general_state_pool, device,
3145 "general pool",
3146 0, GENERAL_STATE_POOL_MIN_ADDRESS, 16384);
3147 if (result != VK_SUCCESS)
3148 goto fail_batch_bo_pool;
3149
3150 result = anv_state_pool_init(&device->dynamic_state_pool, device,
3151 "dynamic pool",
3152 DYNAMIC_STATE_POOL_MIN_ADDRESS, 0, 16384);
3153 if (result != VK_SUCCESS)
3154 goto fail_general_state_pool;
3155
3156 if (device->info.ver >= 8) {
3157 /* The border color pointer is limited to 24 bits, so we need to make
3158 * sure that any such color used at any point in the program doesn't
3159 * exceed that limit.
3160 * We achieve that by reserving all the custom border colors we support
3161 * right off the bat, so they are close to the base address.
3162 */
3163 anv_state_reserved_pool_init(&device->custom_border_colors,
3164 &device->dynamic_state_pool,
3165 MAX_CUSTOM_BORDER_COLORS,
3166 sizeof(struct gfx8_border_color), 64);
3167 }
3168
3169 result = anv_state_pool_init(&device->instruction_state_pool, device,
3170 "instruction pool",
3171 INSTRUCTION_STATE_POOL_MIN_ADDRESS, 0, 16384);
3172 if (result != VK_SUCCESS)
3173 goto fail_dynamic_state_pool;
3174
3175 result = anv_state_pool_init(&device->surface_state_pool, device,
3176 "surface state pool",
3177 SURFACE_STATE_POOL_MIN_ADDRESS, 0, 4096);
3178 if (result != VK_SUCCESS)
3179 goto fail_instruction_state_pool;
3180
3181 if (physical_device->use_softpin) {
3182 int64_t bt_pool_offset = (int64_t)BINDING_TABLE_POOL_MIN_ADDRESS -
3183 (int64_t)SURFACE_STATE_POOL_MIN_ADDRESS;
3184 assert(INT32_MIN < bt_pool_offset && bt_pool_offset < 0);
3185 result = anv_state_pool_init(&device->binding_table_pool, device,
3186 "binding table pool",
3187 SURFACE_STATE_POOL_MIN_ADDRESS,
3188 bt_pool_offset, 4096);
3189 if (result != VK_SUCCESS)
3190 goto fail_surface_state_pool;
3191 }
3192
3193 if (device->info.has_aux_map) {
3194 device->aux_map_ctx = intel_aux_map_init(device, &aux_map_allocator,
3195 &physical_device->info);
3196 if (!device->aux_map_ctx)
3197 goto fail_binding_table_pool;
3198 }
3199
3200 result = anv_device_alloc_bo(device, "workaround", 4096,
3201 ANV_BO_ALLOC_CAPTURE |
3202 ANV_BO_ALLOC_MAPPED |
3203 ANV_BO_ALLOC_LOCAL_MEM,
3204 0 /* explicit_address */,
3205 &device->workaround_bo);
3206 if (result != VK_SUCCESS)
3207 goto fail_surface_aux_map_pool;
3208
3209 device->workaround_address = (struct anv_address) {
3210 .bo = device->workaround_bo,
3211 .offset = align_u32(
3212 intel_debug_write_identifiers(device->workaround_bo->map,
3213 device->workaround_bo->size,
3214 "Anv") + 8, 8),
3215 };
3216
3217 device->debug_frame_desc =
3218 intel_debug_get_identifier_block(device->workaround_bo->map,
3219 device->workaround_bo->size,
3220 INTEL_DEBUG_BLOCK_TYPE_FRAME);
3221
3222 result = anv_device_init_trivial_batch(device);
3223 if (result != VK_SUCCESS)
3224 goto fail_workaround_bo;
3225
3226 /* Allocate a null surface state at surface state offset 0. This makes
3227 * NULL descriptor handling trivial because we can just memset structures
3228 * to zero and they have a valid descriptor.
3229 */
3230 device->null_surface_state =
3231 anv_state_pool_alloc(&device->surface_state_pool,
3232 device->isl_dev.ss.size,
3233 device->isl_dev.ss.align);
3234 isl_null_fill_state(&device->isl_dev, device->null_surface_state.map,
3235 .size = isl_extent3d(1, 1, 1) /* This shouldn't matter */);
3236 assert(device->null_surface_state.offset == 0);
3237
3238 anv_scratch_pool_init(device, &device->scratch_pool);
3239
3240 /* TODO(RT): Do we want some sort of data structure for this? */
3241 memset(device->rt_scratch_bos, 0, sizeof(device->rt_scratch_bos));
3242
3243 result = anv_genX(&device->info, init_device_state)(device);
3244 if (result != VK_SUCCESS)
3245 goto fail_trivial_batch_bo_and_scratch_pool;
3246
3247 anv_pipeline_cache_init(&device->default_pipeline_cache, device,
3248 true /* cache_enabled */, false /* external_sync */);
3249
3250 result = anv_device_init_rt_shaders(device);
3251 if (result != VK_SUCCESS)
3252 goto fail_rt_trampoline;
3253
3254 anv_device_init_blorp(device);
3255
3256 anv_device_init_border_colors(device);
3257
3258 anv_device_perf_init(device);
3259
3260 *pDevice = anv_device_to_handle(device);
3261
3262 return VK_SUCCESS;
3263
3264 fail_rt_trampoline:
3265 anv_pipeline_cache_finish(&device->default_pipeline_cache);
3266 fail_trivial_batch_bo_and_scratch_pool:
3267 anv_scratch_pool_finish(device, &device->scratch_pool);
3268 anv_device_release_bo(device, device->trivial_batch_bo);
3269 fail_workaround_bo:
3270 anv_device_release_bo(device, device->workaround_bo);
3271 fail_surface_aux_map_pool:
3272 if (device->info.has_aux_map) {
3273 intel_aux_map_finish(device->aux_map_ctx);
3274 device->aux_map_ctx = NULL;
3275 }
3276 fail_binding_table_pool:
3277 if (physical_device->use_softpin)
3278 anv_state_pool_finish(&device->binding_table_pool);
3279 fail_surface_state_pool:
3280 anv_state_pool_finish(&device->surface_state_pool);
3281 fail_instruction_state_pool:
3282 anv_state_pool_finish(&device->instruction_state_pool);
3283 fail_dynamic_state_pool:
3284 if (device->info.ver >= 8)
3285 anv_state_reserved_pool_finish(&device->custom_border_colors);
3286 anv_state_pool_finish(&device->dynamic_state_pool);
3287 fail_general_state_pool:
3288 anv_state_pool_finish(&device->general_state_pool);
3289 fail_batch_bo_pool:
3290 anv_bo_pool_finish(&device->batch_bo_pool);
3291 anv_bo_cache_finish(&device->bo_cache);
3292 fail_queue_cond:
3293 pthread_cond_destroy(&device->queue_submit);
3294 fail_mutex:
3295 pthread_mutex_destroy(&device->mutex);
3296 fail_vmas:
3297 if (physical_device->use_softpin) {
3298 util_vma_heap_finish(&device->vma_hi);
3299 util_vma_heap_finish(&device->vma_cva);
3300 util_vma_heap_finish(&device->vma_lo);
3301 }
3302 fail_queues:
3303 for (uint32_t i = 0; i < device->queue_count; i++)
3304 anv_queue_finish(&device->queues[i]);
3305 vk_free(&device->vk.alloc, device->queues);
3306 fail_context_id:
3307 anv_gem_destroy_context(device, device->context_id);
3308 fail_fd:
3309 close(device->fd);
3310 fail_device:
3311 vk_device_finish(&device->vk);
3312 fail_alloc:
3313 vk_free(&device->vk.alloc, device);
3314
3315 return result;
3316 }
3317
anv_DestroyDevice(VkDevice _device,const VkAllocationCallbacks * pAllocator)3318 void anv_DestroyDevice(
3319 VkDevice _device,
3320 const VkAllocationCallbacks* pAllocator)
3321 {
3322 ANV_FROM_HANDLE(anv_device, device, _device);
3323
3324 if (!device)
3325 return;
3326
3327 anv_device_finish_blorp(device);
3328
3329 anv_device_finish_rt_shaders(device);
3330
3331 anv_pipeline_cache_finish(&device->default_pipeline_cache);
3332
3333 #ifdef HAVE_VALGRIND
3334 /* We only need to free these to prevent valgrind errors. The backing
3335 * BO will go away in a couple of lines so we don't actually leak.
3336 */
3337 if (device->info.ver >= 8)
3338 anv_state_reserved_pool_finish(&device->custom_border_colors);
3339 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
3340 anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
3341 #endif
3342
3343 for (unsigned i = 0; i < ARRAY_SIZE(device->rt_scratch_bos); i++) {
3344 if (device->rt_scratch_bos[i] != NULL)
3345 anv_device_release_bo(device, device->rt_scratch_bos[i]);
3346 }
3347
3348 anv_scratch_pool_finish(device, &device->scratch_pool);
3349
3350 anv_device_release_bo(device, device->workaround_bo);
3351 anv_device_release_bo(device, device->trivial_batch_bo);
3352
3353 if (device->info.has_aux_map) {
3354 intel_aux_map_finish(device->aux_map_ctx);
3355 device->aux_map_ctx = NULL;
3356 }
3357
3358 if (device->physical->use_softpin)
3359 anv_state_pool_finish(&device->binding_table_pool);
3360 anv_state_pool_finish(&device->surface_state_pool);
3361 anv_state_pool_finish(&device->instruction_state_pool);
3362 anv_state_pool_finish(&device->dynamic_state_pool);
3363 anv_state_pool_finish(&device->general_state_pool);
3364
3365 anv_bo_pool_finish(&device->batch_bo_pool);
3366
3367 anv_bo_cache_finish(&device->bo_cache);
3368
3369 if (device->physical->use_softpin) {
3370 util_vma_heap_finish(&device->vma_hi);
3371 util_vma_heap_finish(&device->vma_cva);
3372 util_vma_heap_finish(&device->vma_lo);
3373 }
3374
3375 pthread_cond_destroy(&device->queue_submit);
3376 pthread_mutex_destroy(&device->mutex);
3377
3378 for (uint32_t i = 0; i < device->queue_count; i++)
3379 anv_queue_finish(&device->queues[i]);
3380 vk_free(&device->vk.alloc, device->queues);
3381
3382 anv_gem_destroy_context(device, device->context_id);
3383
3384 if (INTEL_DEBUG(DEBUG_BATCH))
3385 intel_batch_decode_ctx_finish(&device->decoder_ctx);
3386
3387 close(device->fd);
3388
3389 vk_device_finish(&device->vk);
3390 vk_free(&device->vk.alloc, device);
3391 }
3392
anv_EnumerateInstanceLayerProperties(uint32_t * pPropertyCount,VkLayerProperties * pProperties)3393 VkResult anv_EnumerateInstanceLayerProperties(
3394 uint32_t* pPropertyCount,
3395 VkLayerProperties* pProperties)
3396 {
3397 if (pProperties == NULL) {
3398 *pPropertyCount = 0;
3399 return VK_SUCCESS;
3400 }
3401
3402 /* None supported at this time */
3403 return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
3404 }
3405
3406 void
_anv_device_report_lost(struct anv_device * device)3407 _anv_device_report_lost(struct anv_device *device)
3408 {
3409 assert(p_atomic_read(&device->_lost) > 0);
3410
3411 device->lost_reported = true;
3412
3413 for (uint32_t i = 0; i < device->queue_count; i++) {
3414 struct anv_queue *queue = &device->queues[i];
3415 if (queue->lost) {
3416 __vk_errorf(queue, VK_ERROR_DEVICE_LOST,
3417 queue->error_file, queue->error_line,
3418 "%s", queue->error_msg);
3419 }
3420 }
3421 }
3422
3423 VkResult
_anv_device_set_lost(struct anv_device * device,const char * file,int line,const char * msg,...)3424 _anv_device_set_lost(struct anv_device *device,
3425 const char *file, int line,
3426 const char *msg, ...)
3427 {
3428 VkResult err;
3429 va_list ap;
3430
3431 if (p_atomic_read(&device->_lost) > 0)
3432 return VK_ERROR_DEVICE_LOST;
3433
3434 p_atomic_inc(&device->_lost);
3435 device->lost_reported = true;
3436
3437 va_start(ap, msg);
3438 err = __vk_errorv(device, VK_ERROR_DEVICE_LOST, file, line, msg, ap);
3439 va_end(ap);
3440
3441 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3442 abort();
3443
3444 return err;
3445 }
3446
3447 VkResult
_anv_queue_set_lost(struct anv_queue * queue,const char * file,int line,const char * msg,...)3448 _anv_queue_set_lost(struct anv_queue *queue,
3449 const char *file, int line,
3450 const char *msg, ...)
3451 {
3452 va_list ap;
3453
3454 if (queue->lost)
3455 return VK_ERROR_DEVICE_LOST;
3456
3457 queue->lost = true;
3458
3459 queue->error_file = file;
3460 queue->error_line = line;
3461 va_start(ap, msg);
3462 vsnprintf(queue->error_msg, sizeof(queue->error_msg),
3463 msg, ap);
3464 va_end(ap);
3465
3466 p_atomic_inc(&queue->device->_lost);
3467
3468 if (env_var_as_boolean("ANV_ABORT_ON_DEVICE_LOSS", false))
3469 abort();
3470
3471 return VK_ERROR_DEVICE_LOST;
3472 }
3473
3474 VkResult
anv_device_query_status(struct anv_device * device)3475 anv_device_query_status(struct anv_device *device)
3476 {
3477 /* This isn't likely as most of the callers of this function already check
3478 * for it. However, it doesn't hurt to check and it potentially lets us
3479 * avoid an ioctl.
3480 */
3481 if (anv_device_is_lost(device))
3482 return VK_ERROR_DEVICE_LOST;
3483
3484 uint32_t active, pending;
3485 int ret = anv_gem_context_get_reset_stats(device->fd, device->context_id,
3486 &active, &pending);
3487 if (ret == -1) {
3488 /* We don't know the real error. */
3489 return anv_device_set_lost(device, "get_reset_stats failed: %m");
3490 }
3491
3492 if (active) {
3493 return anv_device_set_lost(device, "GPU hung on one of our command buffers");
3494 } else if (pending) {
3495 return anv_device_set_lost(device, "GPU hung with commands in-flight");
3496 }
3497
3498 return VK_SUCCESS;
3499 }
3500
3501 VkResult
anv_device_bo_busy(struct anv_device * device,struct anv_bo * bo)3502 anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo)
3503 {
3504 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
3505 * Other usages of the BO (such as on different hardware) will not be
3506 * flagged as "busy" by this ioctl. Use with care.
3507 */
3508 int ret = anv_gem_busy(device, bo->gem_handle);
3509 if (ret == 1) {
3510 return VK_NOT_READY;
3511 } else if (ret == -1) {
3512 /* We don't know the real error. */
3513 return anv_device_set_lost(device, "gem wait failed: %m");
3514 }
3515
3516 /* Query for device status after the busy call. If the BO we're checking
3517 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
3518 * client because it clearly doesn't have valid data. Yes, this most
3519 * likely means an ioctl, but we just did an ioctl to query the busy status
3520 * so it's no great loss.
3521 */
3522 return anv_device_query_status(device);
3523 }
3524
3525 VkResult
anv_device_wait(struct anv_device * device,struct anv_bo * bo,int64_t timeout)3526 anv_device_wait(struct anv_device *device, struct anv_bo *bo,
3527 int64_t timeout)
3528 {
3529 int ret = anv_gem_wait(device, bo->gem_handle, &timeout);
3530 if (ret == -1 && errno == ETIME) {
3531 return VK_TIMEOUT;
3532 } else if (ret == -1) {
3533 /* We don't know the real error. */
3534 return anv_device_set_lost(device, "gem wait failed: %m");
3535 }
3536
3537 /* Query for device status after the wait. If the BO we're waiting on got
3538 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
3539 * because it clearly doesn't have valid data. Yes, this most likely means
3540 * an ioctl, but we just did an ioctl to wait so it's no great loss.
3541 */
3542 return anv_device_query_status(device);
3543 }
3544
3545 uint64_t
anv_vma_alloc(struct anv_device * device,uint64_t size,uint64_t align,enum anv_bo_alloc_flags alloc_flags,uint64_t client_address)3546 anv_vma_alloc(struct anv_device *device,
3547 uint64_t size, uint64_t align,
3548 enum anv_bo_alloc_flags alloc_flags,
3549 uint64_t client_address)
3550 {
3551 pthread_mutex_lock(&device->vma_mutex);
3552
3553 uint64_t addr = 0;
3554
3555 if (alloc_flags & ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS) {
3556 if (client_address) {
3557 if (util_vma_heap_alloc_addr(&device->vma_cva,
3558 client_address, size)) {
3559 addr = client_address;
3560 }
3561 } else {
3562 addr = util_vma_heap_alloc(&device->vma_cva, size, align);
3563 }
3564 /* We don't want to fall back to other heaps */
3565 goto done;
3566 }
3567
3568 assert(client_address == 0);
3569
3570 if (!(alloc_flags & ANV_BO_ALLOC_32BIT_ADDRESS))
3571 addr = util_vma_heap_alloc(&device->vma_hi, size, align);
3572
3573 if (addr == 0)
3574 addr = util_vma_heap_alloc(&device->vma_lo, size, align);
3575
3576 done:
3577 pthread_mutex_unlock(&device->vma_mutex);
3578
3579 assert(addr == intel_48b_address(addr));
3580 return intel_canonical_address(addr);
3581 }
3582
3583 void
anv_vma_free(struct anv_device * device,uint64_t address,uint64_t size)3584 anv_vma_free(struct anv_device *device,
3585 uint64_t address, uint64_t size)
3586 {
3587 const uint64_t addr_48b = intel_48b_address(address);
3588
3589 pthread_mutex_lock(&device->vma_mutex);
3590
3591 if (addr_48b >= LOW_HEAP_MIN_ADDRESS &&
3592 addr_48b <= LOW_HEAP_MAX_ADDRESS) {
3593 util_vma_heap_free(&device->vma_lo, addr_48b, size);
3594 } else if (addr_48b >= CLIENT_VISIBLE_HEAP_MIN_ADDRESS &&
3595 addr_48b <= CLIENT_VISIBLE_HEAP_MAX_ADDRESS) {
3596 util_vma_heap_free(&device->vma_cva, addr_48b, size);
3597 } else {
3598 assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS);
3599 util_vma_heap_free(&device->vma_hi, addr_48b, size);
3600 }
3601
3602 pthread_mutex_unlock(&device->vma_mutex);
3603 }
3604
anv_AllocateMemory(VkDevice _device,const VkMemoryAllocateInfo * pAllocateInfo,const VkAllocationCallbacks * pAllocator,VkDeviceMemory * pMem)3605 VkResult anv_AllocateMemory(
3606 VkDevice _device,
3607 const VkMemoryAllocateInfo* pAllocateInfo,
3608 const VkAllocationCallbacks* pAllocator,
3609 VkDeviceMemory* pMem)
3610 {
3611 ANV_FROM_HANDLE(anv_device, device, _device);
3612 struct anv_physical_device *pdevice = device->physical;
3613 struct anv_device_memory *mem;
3614 VkResult result = VK_SUCCESS;
3615
3616 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
3617
3618 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
3619 assert(pAllocateInfo->allocationSize > 0);
3620
3621 VkDeviceSize aligned_alloc_size =
3622 align_u64(pAllocateInfo->allocationSize, 4096);
3623
3624 if (aligned_alloc_size > MAX_MEMORY_ALLOCATION_SIZE)
3625 return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
3626
3627 assert(pAllocateInfo->memoryTypeIndex < pdevice->memory.type_count);
3628 struct anv_memory_type *mem_type =
3629 &pdevice->memory.types[pAllocateInfo->memoryTypeIndex];
3630 assert(mem_type->heapIndex < pdevice->memory.heap_count);
3631 struct anv_memory_heap *mem_heap =
3632 &pdevice->memory.heaps[mem_type->heapIndex];
3633
3634 uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
3635 if (mem_heap_used + aligned_alloc_size > mem_heap->size)
3636 return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
3637
3638 mem = vk_object_alloc(&device->vk, pAllocator, sizeof(*mem),
3639 VK_OBJECT_TYPE_DEVICE_MEMORY);
3640 if (mem == NULL)
3641 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
3642
3643 mem->type = mem_type;
3644 mem->map = NULL;
3645 mem->map_size = 0;
3646 mem->ahw = NULL;
3647 mem->host_ptr = NULL;
3648
3649 enum anv_bo_alloc_flags alloc_flags = 0;
3650
3651 const VkExportMemoryAllocateInfo *export_info = NULL;
3652 const VkImportAndroidHardwareBufferInfoANDROID *ahw_import_info = NULL;
3653 const VkImportMemoryFdInfoKHR *fd_info = NULL;
3654 const VkImportMemoryHostPointerInfoEXT *host_ptr_info = NULL;
3655 const VkMemoryDedicatedAllocateInfo *dedicated_info = NULL;
3656 VkMemoryAllocateFlags vk_flags = 0;
3657 uint64_t client_address = 0;
3658
3659 vk_foreach_struct_const(ext, pAllocateInfo->pNext) {
3660 switch (ext->sType) {
3661 case VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO:
3662 export_info = (void *)ext;
3663 break;
3664
3665 case VK_STRUCTURE_TYPE_IMPORT_ANDROID_HARDWARE_BUFFER_INFO_ANDROID:
3666 ahw_import_info = (void *)ext;
3667 break;
3668
3669 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR:
3670 fd_info = (void *)ext;
3671 break;
3672
3673 case VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT:
3674 host_ptr_info = (void *)ext;
3675 break;
3676
3677 case VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_FLAGS_INFO: {
3678 const VkMemoryAllocateFlagsInfo *flags_info = (void *)ext;
3679 vk_flags = flags_info->flags;
3680 break;
3681 }
3682
3683 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO:
3684 dedicated_info = (void *)ext;
3685 break;
3686
3687 case VK_STRUCTURE_TYPE_MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO_KHR: {
3688 const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *addr_info =
3689 (const VkMemoryOpaqueCaptureAddressAllocateInfoKHR *)ext;
3690 client_address = addr_info->opaqueCaptureAddress;
3691 break;
3692 }
3693
3694 default:
3695 anv_debug_ignored_stype(ext->sType);
3696 break;
3697 }
3698 }
3699
3700 /* By default, we want all VkDeviceMemory objects to support CCS */
3701 if (device->physical->has_implicit_ccs)
3702 alloc_flags |= ANV_BO_ALLOC_IMPLICIT_CCS;
3703
3704 if (vk_flags & VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_BIT_KHR)
3705 alloc_flags |= ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS;
3706
3707 if ((export_info && export_info->handleTypes) ||
3708 (fd_info && fd_info->handleType) ||
3709 (host_ptr_info && host_ptr_info->handleType)) {
3710 /* Anything imported or exported is EXTERNAL */
3711 alloc_flags |= ANV_BO_ALLOC_EXTERNAL;
3712
3713 /* We can't have implicit CCS on external memory with an AUX-table.
3714 * Doing so would require us to sync the aux tables across processes
3715 * which is impractical.
3716 */
3717 if (device->info.has_aux_map)
3718 alloc_flags &= ~ANV_BO_ALLOC_IMPLICIT_CCS;
3719 }
3720
3721 /* Check if we need to support Android HW buffer export. If so,
3722 * create AHardwareBuffer and import memory from it.
3723 */
3724 bool android_export = false;
3725 if (export_info && export_info->handleTypes &
3726 VK_EXTERNAL_MEMORY_HANDLE_TYPE_ANDROID_HARDWARE_BUFFER_BIT_ANDROID)
3727 android_export = true;
3728
3729 if (ahw_import_info) {
3730 result = anv_import_ahw_memory(_device, mem, ahw_import_info);
3731 if (result != VK_SUCCESS)
3732 goto fail;
3733
3734 goto success;
3735 } else if (android_export) {
3736 result = anv_create_ahw_memory(_device, mem, pAllocateInfo);
3737 if (result != VK_SUCCESS)
3738 goto fail;
3739
3740 goto success;
3741 }
3742
3743 /* The Vulkan spec permits handleType to be 0, in which case the struct is
3744 * ignored.
3745 */
3746 if (fd_info && fd_info->handleType) {
3747 /* At the moment, we support only the below handle types. */
3748 assert(fd_info->handleType ==
3749 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3750 fd_info->handleType ==
3751 VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3752
3753 result = anv_device_import_bo(device, fd_info->fd, alloc_flags,
3754 client_address, &mem->bo);
3755 if (result != VK_SUCCESS)
3756 goto fail;
3757
3758 /* For security purposes, we reject importing the bo if it's smaller
3759 * than the requested allocation size. This prevents a malicious client
3760 * from passing a buffer to a trusted client, lying about the size, and
3761 * telling the trusted client to try and texture from an image that goes
3762 * out-of-bounds. This sort of thing could lead to GPU hangs or worse
3763 * in the trusted client. The trusted client can protect itself against
3764 * this sort of attack but only if it can trust the buffer size.
3765 */
3766 if (mem->bo->size < aligned_alloc_size) {
3767 result = vk_errorf(device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
3768 "aligned allocationSize too large for "
3769 "VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT: "
3770 "%"PRIu64"B > %"PRIu64"B",
3771 aligned_alloc_size, mem->bo->size);
3772 anv_device_release_bo(device, mem->bo);
3773 goto fail;
3774 }
3775
3776 /* From the Vulkan spec:
3777 *
3778 * "Importing memory from a file descriptor transfers ownership of
3779 * the file descriptor from the application to the Vulkan
3780 * implementation. The application must not perform any operations on
3781 * the file descriptor after a successful import."
3782 *
3783 * If the import fails, we leave the file descriptor open.
3784 */
3785 close(fd_info->fd);
3786 goto success;
3787 }
3788
3789 if (host_ptr_info && host_ptr_info->handleType) {
3790 if (host_ptr_info->handleType ==
3791 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_MAPPED_FOREIGN_MEMORY_BIT_EXT) {
3792 result = vk_error(device, VK_ERROR_INVALID_EXTERNAL_HANDLE);
3793 goto fail;
3794 }
3795
3796 assert(host_ptr_info->handleType ==
3797 VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT);
3798
3799 result = anv_device_import_bo_from_host_ptr(device,
3800 host_ptr_info->pHostPointer,
3801 pAllocateInfo->allocationSize,
3802 alloc_flags,
3803 client_address,
3804 &mem->bo);
3805 if (result != VK_SUCCESS)
3806 goto fail;
3807
3808 mem->host_ptr = host_ptr_info->pHostPointer;
3809 goto success;
3810 }
3811
3812 /* Set ALLOC_LOCAL_MEM flag if heap has device local bit set and requested
3813 * memory property flag has DEVICE_LOCAL_BIT set.
3814 */
3815 if (mem_type->propertyFlags & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT)
3816 alloc_flags |= ANV_BO_ALLOC_LOCAL_MEM;
3817
3818 /* Regular allocate (not importing memory). */
3819
3820 result = anv_device_alloc_bo(device, "user", pAllocateInfo->allocationSize,
3821 alloc_flags, client_address, &mem->bo);
3822 if (result != VK_SUCCESS)
3823 goto fail;
3824
3825 if (dedicated_info && dedicated_info->image != VK_NULL_HANDLE) {
3826 ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
3827
3828 /* Some legacy (non-modifiers) consumers need the tiling to be set on
3829 * the BO. In this case, we have a dedicated allocation.
3830 */
3831 if (image->vk.wsi_legacy_scanout) {
3832 const uint32_t i915_tiling =
3833 isl_tiling_to_i915_tiling(image->planes[0].primary_surface.isl.tiling);
3834 int ret = anv_gem_set_tiling(device, mem->bo->gem_handle,
3835 image->planes[0].primary_surface.isl.row_pitch_B,
3836 i915_tiling);
3837 if (ret) {
3838 anv_device_release_bo(device, mem->bo);
3839 result = vk_errorf(device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3840 "failed to set BO tiling: %m");
3841 goto fail;
3842 }
3843 }
3844 }
3845
3846 success:
3847 mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
3848 if (mem_heap_used > mem_heap->size) {
3849 p_atomic_add(&mem_heap->used, -mem->bo->size);
3850 anv_device_release_bo(device, mem->bo);
3851 result = vk_errorf(device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
3852 "Out of heap memory");
3853 goto fail;
3854 }
3855
3856 pthread_mutex_lock(&device->mutex);
3857 list_addtail(&mem->link, &device->memory_objects);
3858 pthread_mutex_unlock(&device->mutex);
3859
3860 *pMem = anv_device_memory_to_handle(mem);
3861
3862 return VK_SUCCESS;
3863
3864 fail:
3865 vk_object_free(&device->vk, pAllocator, mem);
3866
3867 return result;
3868 }
3869
anv_GetMemoryFdKHR(VkDevice device_h,const VkMemoryGetFdInfoKHR * pGetFdInfo,int * pFd)3870 VkResult anv_GetMemoryFdKHR(
3871 VkDevice device_h,
3872 const VkMemoryGetFdInfoKHR* pGetFdInfo,
3873 int* pFd)
3874 {
3875 ANV_FROM_HANDLE(anv_device, dev, device_h);
3876 ANV_FROM_HANDLE(anv_device_memory, mem, pGetFdInfo->memory);
3877
3878 assert(pGetFdInfo->sType == VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR);
3879
3880 assert(pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT ||
3881 pGetFdInfo->handleType == VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
3882
3883 return anv_device_export_bo(dev, mem->bo, pFd);
3884 }
3885
anv_GetMemoryFdPropertiesKHR(VkDevice _device,VkExternalMemoryHandleTypeFlagBits handleType,int fd,VkMemoryFdPropertiesKHR * pMemoryFdProperties)3886 VkResult anv_GetMemoryFdPropertiesKHR(
3887 VkDevice _device,
3888 VkExternalMemoryHandleTypeFlagBits handleType,
3889 int fd,
3890 VkMemoryFdPropertiesKHR* pMemoryFdProperties)
3891 {
3892 ANV_FROM_HANDLE(anv_device, device, _device);
3893
3894 switch (handleType) {
3895 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT:
3896 /* dma-buf can be imported as any memory type */
3897 pMemoryFdProperties->memoryTypeBits =
3898 (1 << device->physical->memory.type_count) - 1;
3899 return VK_SUCCESS;
3900
3901 default:
3902 /* The valid usage section for this function says:
3903 *
3904 * "handleType must not be one of the handle types defined as
3905 * opaque."
3906 *
3907 * So opaque handle types fall into the default "unsupported" case.
3908 */
3909 return vk_error(device, VK_ERROR_INVALID_EXTERNAL_HANDLE);
3910 }
3911 }
3912
anv_GetMemoryHostPointerPropertiesEXT(VkDevice _device,VkExternalMemoryHandleTypeFlagBits handleType,const void * pHostPointer,VkMemoryHostPointerPropertiesEXT * pMemoryHostPointerProperties)3913 VkResult anv_GetMemoryHostPointerPropertiesEXT(
3914 VkDevice _device,
3915 VkExternalMemoryHandleTypeFlagBits handleType,
3916 const void* pHostPointer,
3917 VkMemoryHostPointerPropertiesEXT* pMemoryHostPointerProperties)
3918 {
3919 ANV_FROM_HANDLE(anv_device, device, _device);
3920
3921 assert(pMemoryHostPointerProperties->sType ==
3922 VK_STRUCTURE_TYPE_MEMORY_HOST_POINTER_PROPERTIES_EXT);
3923
3924 switch (handleType) {
3925 case VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT:
3926 /* Host memory can be imported as any memory type. */
3927 pMemoryHostPointerProperties->memoryTypeBits =
3928 (1ull << device->physical->memory.type_count) - 1;
3929
3930 return VK_SUCCESS;
3931
3932 default:
3933 return VK_ERROR_INVALID_EXTERNAL_HANDLE;
3934 }
3935 }
3936
anv_FreeMemory(VkDevice _device,VkDeviceMemory _mem,const VkAllocationCallbacks * pAllocator)3937 void anv_FreeMemory(
3938 VkDevice _device,
3939 VkDeviceMemory _mem,
3940 const VkAllocationCallbacks* pAllocator)
3941 {
3942 ANV_FROM_HANDLE(anv_device, device, _device);
3943 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
3944
3945 if (mem == NULL)
3946 return;
3947
3948 pthread_mutex_lock(&device->mutex);
3949 list_del(&mem->link);
3950 pthread_mutex_unlock(&device->mutex);
3951
3952 if (mem->map)
3953 anv_UnmapMemory(_device, _mem);
3954
3955 p_atomic_add(&device->physical->memory.heaps[mem->type->heapIndex].used,
3956 -mem->bo->size);
3957
3958 anv_device_release_bo(device, mem->bo);
3959
3960 #if defined(ANDROID) && ANDROID_API_LEVEL >= 26
3961 if (mem->ahw)
3962 AHardwareBuffer_release(mem->ahw);
3963 #endif
3964
3965 vk_object_free(&device->vk, pAllocator, mem);
3966 }
3967
anv_MapMemory(VkDevice _device,VkDeviceMemory _memory,VkDeviceSize offset,VkDeviceSize size,VkMemoryMapFlags flags,void ** ppData)3968 VkResult anv_MapMemory(
3969 VkDevice _device,
3970 VkDeviceMemory _memory,
3971 VkDeviceSize offset,
3972 VkDeviceSize size,
3973 VkMemoryMapFlags flags,
3974 void** ppData)
3975 {
3976 ANV_FROM_HANDLE(anv_device, device, _device);
3977 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
3978
3979 if (mem == NULL) {
3980 *ppData = NULL;
3981 return VK_SUCCESS;
3982 }
3983
3984 if (mem->host_ptr) {
3985 *ppData = mem->host_ptr + offset;
3986 return VK_SUCCESS;
3987 }
3988
3989 if (size == VK_WHOLE_SIZE)
3990 size = mem->bo->size - offset;
3991
3992 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
3993 *
3994 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
3995 * assert(size != 0);
3996 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
3997 * equal to the size of the memory minus offset
3998 */
3999 assert(size > 0);
4000 assert(offset + size <= mem->bo->size);
4001
4002 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
4003 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
4004 * at a time is valid. We could just mmap up front and return an offset
4005 * pointer here, but that may exhaust virtual memory on 32 bit
4006 * userspace. */
4007
4008 uint32_t gem_flags = 0;
4009
4010 if (!device->info.has_llc &&
4011 (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT))
4012 gem_flags |= I915_MMAP_WC;
4013
4014 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
4015 uint64_t map_offset;
4016 if (!device->physical->has_mmap_offset)
4017 map_offset = offset & ~4095ull;
4018 else
4019 map_offset = 0;
4020 assert(offset >= map_offset);
4021 uint64_t map_size = (offset + size) - map_offset;
4022
4023 /* Let's map whole pages */
4024 map_size = align_u64(map_size, 4096);
4025
4026 void *map = anv_gem_mmap(device, mem->bo->gem_handle,
4027 map_offset, map_size, gem_flags);
4028 if (map == MAP_FAILED)
4029 return vk_error(device, VK_ERROR_MEMORY_MAP_FAILED);
4030
4031 mem->map = map;
4032 mem->map_size = map_size;
4033 mem->map_delta = (offset - map_offset);
4034
4035 *ppData = mem->map + mem->map_delta;
4036
4037 return VK_SUCCESS;
4038 }
4039
anv_UnmapMemory(VkDevice _device,VkDeviceMemory _memory)4040 void anv_UnmapMemory(
4041 VkDevice _device,
4042 VkDeviceMemory _memory)
4043 {
4044 ANV_FROM_HANDLE(anv_device, device, _device);
4045 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
4046
4047 if (mem == NULL || mem->host_ptr)
4048 return;
4049
4050 anv_gem_munmap(device, mem->map, mem->map_size);
4051
4052 mem->map = NULL;
4053 mem->map_size = 0;
4054 mem->map_delta = 0;
4055 }
4056
4057 static void
clflush_mapped_ranges(struct anv_device * device,uint32_t count,const VkMappedMemoryRange * ranges)4058 clflush_mapped_ranges(struct anv_device *device,
4059 uint32_t count,
4060 const VkMappedMemoryRange *ranges)
4061 {
4062 for (uint32_t i = 0; i < count; i++) {
4063 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
4064 uint64_t map_offset = ranges[i].offset + mem->map_delta;
4065 if (map_offset >= mem->map_size)
4066 continue;
4067
4068 if (mem->type->propertyFlags & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT)
4069 continue;
4070
4071 intel_clflush_range(mem->map + map_offset,
4072 MIN2(ranges[i].size, mem->map_size - map_offset));
4073 }
4074 }
4075
anv_FlushMappedMemoryRanges(VkDevice _device,uint32_t memoryRangeCount,const VkMappedMemoryRange * pMemoryRanges)4076 VkResult anv_FlushMappedMemoryRanges(
4077 VkDevice _device,
4078 uint32_t memoryRangeCount,
4079 const VkMappedMemoryRange* pMemoryRanges)
4080 {
4081 ANV_FROM_HANDLE(anv_device, device, _device);
4082
4083 if (!device->physical->memory.need_clflush)
4084 return VK_SUCCESS;
4085
4086 /* Make sure the writes we're flushing have landed. */
4087 __builtin_ia32_mfence();
4088
4089 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
4090
4091 return VK_SUCCESS;
4092 }
4093
anv_InvalidateMappedMemoryRanges(VkDevice _device,uint32_t memoryRangeCount,const VkMappedMemoryRange * pMemoryRanges)4094 VkResult anv_InvalidateMappedMemoryRanges(
4095 VkDevice _device,
4096 uint32_t memoryRangeCount,
4097 const VkMappedMemoryRange* pMemoryRanges)
4098 {
4099 ANV_FROM_HANDLE(anv_device, device, _device);
4100
4101 if (!device->physical->memory.need_clflush)
4102 return VK_SUCCESS;
4103
4104 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
4105
4106 /* Make sure no reads get moved up above the invalidate. */
4107 __builtin_ia32_mfence();
4108
4109 return VK_SUCCESS;
4110 }
4111
anv_GetDeviceMemoryCommitment(VkDevice device,VkDeviceMemory memory,VkDeviceSize * pCommittedMemoryInBytes)4112 void anv_GetDeviceMemoryCommitment(
4113 VkDevice device,
4114 VkDeviceMemory memory,
4115 VkDeviceSize* pCommittedMemoryInBytes)
4116 {
4117 *pCommittedMemoryInBytes = 0;
4118 }
4119
4120 static void
anv_bind_buffer_memory(const VkBindBufferMemoryInfo * pBindInfo)4121 anv_bind_buffer_memory(const VkBindBufferMemoryInfo *pBindInfo)
4122 {
4123 ANV_FROM_HANDLE(anv_device_memory, mem, pBindInfo->memory);
4124 ANV_FROM_HANDLE(anv_buffer, buffer, pBindInfo->buffer);
4125
4126 assert(pBindInfo->sType == VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO);
4127
4128 if (mem) {
4129 assert(pBindInfo->memoryOffset < mem->bo->size);
4130 assert(mem->bo->size - pBindInfo->memoryOffset >= buffer->size);
4131 buffer->address = (struct anv_address) {
4132 .bo = mem->bo,
4133 .offset = pBindInfo->memoryOffset,
4134 };
4135 } else {
4136 buffer->address = ANV_NULL_ADDRESS;
4137 }
4138 }
4139
anv_BindBufferMemory2(VkDevice device,uint32_t bindInfoCount,const VkBindBufferMemoryInfo * pBindInfos)4140 VkResult anv_BindBufferMemory2(
4141 VkDevice device,
4142 uint32_t bindInfoCount,
4143 const VkBindBufferMemoryInfo* pBindInfos)
4144 {
4145 for (uint32_t i = 0; i < bindInfoCount; i++)
4146 anv_bind_buffer_memory(&pBindInfos[i]);
4147
4148 return VK_SUCCESS;
4149 }
4150
anv_QueueBindSparse(VkQueue _queue,uint32_t bindInfoCount,const VkBindSparseInfo * pBindInfo,VkFence fence)4151 VkResult anv_QueueBindSparse(
4152 VkQueue _queue,
4153 uint32_t bindInfoCount,
4154 const VkBindSparseInfo* pBindInfo,
4155 VkFence fence)
4156 {
4157 ANV_FROM_HANDLE(anv_queue, queue, _queue);
4158 if (anv_device_is_lost(queue->device))
4159 return VK_ERROR_DEVICE_LOST;
4160
4161 return vk_error(queue, VK_ERROR_FEATURE_NOT_PRESENT);
4162 }
4163
4164 // Event functions
4165
anv_CreateEvent(VkDevice _device,const VkEventCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkEvent * pEvent)4166 VkResult anv_CreateEvent(
4167 VkDevice _device,
4168 const VkEventCreateInfo* pCreateInfo,
4169 const VkAllocationCallbacks* pAllocator,
4170 VkEvent* pEvent)
4171 {
4172 ANV_FROM_HANDLE(anv_device, device, _device);
4173 struct anv_event *event;
4174
4175 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
4176
4177 event = vk_object_alloc(&device->vk, pAllocator, sizeof(*event),
4178 VK_OBJECT_TYPE_EVENT);
4179 if (event == NULL)
4180 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
4181
4182 event->state = anv_state_pool_alloc(&device->dynamic_state_pool,
4183 sizeof(uint64_t), 8);
4184 *(uint64_t *)event->state.map = VK_EVENT_RESET;
4185
4186 *pEvent = anv_event_to_handle(event);
4187
4188 return VK_SUCCESS;
4189 }
4190
anv_DestroyEvent(VkDevice _device,VkEvent _event,const VkAllocationCallbacks * pAllocator)4191 void anv_DestroyEvent(
4192 VkDevice _device,
4193 VkEvent _event,
4194 const VkAllocationCallbacks* pAllocator)
4195 {
4196 ANV_FROM_HANDLE(anv_device, device, _device);
4197 ANV_FROM_HANDLE(anv_event, event, _event);
4198
4199 if (!event)
4200 return;
4201
4202 anv_state_pool_free(&device->dynamic_state_pool, event->state);
4203
4204 vk_object_free(&device->vk, pAllocator, event);
4205 }
4206
anv_GetEventStatus(VkDevice _device,VkEvent _event)4207 VkResult anv_GetEventStatus(
4208 VkDevice _device,
4209 VkEvent _event)
4210 {
4211 ANV_FROM_HANDLE(anv_device, device, _device);
4212 ANV_FROM_HANDLE(anv_event, event, _event);
4213
4214 if (anv_device_is_lost(device))
4215 return VK_ERROR_DEVICE_LOST;
4216
4217 return *(uint64_t *)event->state.map;
4218 }
4219
anv_SetEvent(VkDevice _device,VkEvent _event)4220 VkResult anv_SetEvent(
4221 VkDevice _device,
4222 VkEvent _event)
4223 {
4224 ANV_FROM_HANDLE(anv_event, event, _event);
4225
4226 *(uint64_t *)event->state.map = VK_EVENT_SET;
4227
4228 return VK_SUCCESS;
4229 }
4230
anv_ResetEvent(VkDevice _device,VkEvent _event)4231 VkResult anv_ResetEvent(
4232 VkDevice _device,
4233 VkEvent _event)
4234 {
4235 ANV_FROM_HANDLE(anv_event, event, _event);
4236
4237 *(uint64_t *)event->state.map = VK_EVENT_RESET;
4238
4239 return VK_SUCCESS;
4240 }
4241
4242 // Buffer functions
4243
4244 static void
anv_get_buffer_memory_requirements(struct anv_device * device,VkDeviceSize size,VkBufferUsageFlags usage,VkMemoryRequirements2 * pMemoryRequirements)4245 anv_get_buffer_memory_requirements(struct anv_device *device,
4246 VkDeviceSize size,
4247 VkBufferUsageFlags usage,
4248 VkMemoryRequirements2* pMemoryRequirements)
4249 {
4250 /* The Vulkan spec (git aaed022) says:
4251 *
4252 * memoryTypeBits is a bitfield and contains one bit set for every
4253 * supported memory type for the resource. The bit `1<<i` is set if and
4254 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
4255 * structure for the physical device is supported.
4256 */
4257 uint32_t memory_types = (1ull << device->physical->memory.type_count) - 1;
4258
4259 /* Base alignment requirement of a cache line */
4260 uint32_t alignment = 16;
4261
4262 if (usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT)
4263 alignment = MAX2(alignment, ANV_UBO_ALIGNMENT);
4264
4265 pMemoryRequirements->memoryRequirements.size = size;
4266 pMemoryRequirements->memoryRequirements.alignment = alignment;
4267
4268 /* Storage and Uniform buffers should have their size aligned to
4269 * 32-bits to avoid boundary checks when last DWord is not complete.
4270 * This would ensure that not internal padding would be needed for
4271 * 16-bit types.
4272 */
4273 if (device->robust_buffer_access &&
4274 (usage & VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT ||
4275 usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT))
4276 pMemoryRequirements->memoryRequirements.size = align_u64(size, 4);
4277
4278 pMemoryRequirements->memoryRequirements.memoryTypeBits = memory_types;
4279
4280 vk_foreach_struct(ext, pMemoryRequirements->pNext) {
4281 switch (ext->sType) {
4282 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
4283 VkMemoryDedicatedRequirements *requirements = (void *)ext;
4284 requirements->prefersDedicatedAllocation = false;
4285 requirements->requiresDedicatedAllocation = false;
4286 break;
4287 }
4288
4289 default:
4290 anv_debug_ignored_stype(ext->sType);
4291 break;
4292 }
4293 }
4294 }
4295
anv_GetBufferMemoryRequirements2(VkDevice _device,const VkBufferMemoryRequirementsInfo2 * pInfo,VkMemoryRequirements2 * pMemoryRequirements)4296 void anv_GetBufferMemoryRequirements2(
4297 VkDevice _device,
4298 const VkBufferMemoryRequirementsInfo2* pInfo,
4299 VkMemoryRequirements2* pMemoryRequirements)
4300 {
4301 ANV_FROM_HANDLE(anv_device, device, _device);
4302 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
4303
4304 anv_get_buffer_memory_requirements(device,
4305 buffer->size,
4306 buffer->usage,
4307 pMemoryRequirements);
4308 }
4309
anv_GetDeviceBufferMemoryRequirementsKHR(VkDevice _device,const VkDeviceBufferMemoryRequirementsKHR * pInfo,VkMemoryRequirements2 * pMemoryRequirements)4310 void anv_GetDeviceBufferMemoryRequirementsKHR(
4311 VkDevice _device,
4312 const VkDeviceBufferMemoryRequirementsKHR* pInfo,
4313 VkMemoryRequirements2* pMemoryRequirements)
4314 {
4315 ANV_FROM_HANDLE(anv_device, device, _device);
4316
4317 anv_get_buffer_memory_requirements(device,
4318 pInfo->pCreateInfo->size,
4319 pInfo->pCreateInfo->usage,
4320 pMemoryRequirements);
4321 }
4322
anv_CreateBuffer(VkDevice _device,const VkBufferCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkBuffer * pBuffer)4323 VkResult anv_CreateBuffer(
4324 VkDevice _device,
4325 const VkBufferCreateInfo* pCreateInfo,
4326 const VkAllocationCallbacks* pAllocator,
4327 VkBuffer* pBuffer)
4328 {
4329 ANV_FROM_HANDLE(anv_device, device, _device);
4330 struct anv_buffer *buffer;
4331
4332 /* Don't allow creating buffers bigger than our address space. The real
4333 * issue here is that we may align up the buffer size and we don't want
4334 * doing so to cause roll-over. However, no one has any business
4335 * allocating a buffer larger than our GTT size.
4336 */
4337 if (pCreateInfo->size > device->physical->gtt_size)
4338 return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
4339
4340 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
4341
4342 buffer = vk_object_alloc(&device->vk, pAllocator, sizeof(*buffer),
4343 VK_OBJECT_TYPE_BUFFER);
4344 if (buffer == NULL)
4345 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
4346
4347 buffer->create_flags = pCreateInfo->flags;
4348 buffer->size = pCreateInfo->size;
4349 buffer->usage = pCreateInfo->usage;
4350 buffer->address = ANV_NULL_ADDRESS;
4351
4352 *pBuffer = anv_buffer_to_handle(buffer);
4353
4354 return VK_SUCCESS;
4355 }
4356
anv_DestroyBuffer(VkDevice _device,VkBuffer _buffer,const VkAllocationCallbacks * pAllocator)4357 void anv_DestroyBuffer(
4358 VkDevice _device,
4359 VkBuffer _buffer,
4360 const VkAllocationCallbacks* pAllocator)
4361 {
4362 ANV_FROM_HANDLE(anv_device, device, _device);
4363 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
4364
4365 if (!buffer)
4366 return;
4367
4368 vk_object_free(&device->vk, pAllocator, buffer);
4369 }
4370
anv_GetBufferDeviceAddress(VkDevice device,const VkBufferDeviceAddressInfoKHR * pInfo)4371 VkDeviceAddress anv_GetBufferDeviceAddress(
4372 VkDevice device,
4373 const VkBufferDeviceAddressInfoKHR* pInfo)
4374 {
4375 ANV_FROM_HANDLE(anv_buffer, buffer, pInfo->buffer);
4376
4377 assert(!anv_address_is_null(buffer->address));
4378 assert(buffer->address.bo->flags & EXEC_OBJECT_PINNED);
4379
4380 return anv_address_physical(buffer->address);
4381 }
4382
anv_GetBufferOpaqueCaptureAddress(VkDevice device,const VkBufferDeviceAddressInfoKHR * pInfo)4383 uint64_t anv_GetBufferOpaqueCaptureAddress(
4384 VkDevice device,
4385 const VkBufferDeviceAddressInfoKHR* pInfo)
4386 {
4387 return 0;
4388 }
4389
anv_GetDeviceMemoryOpaqueCaptureAddress(VkDevice device,const VkDeviceMemoryOpaqueCaptureAddressInfoKHR * pInfo)4390 uint64_t anv_GetDeviceMemoryOpaqueCaptureAddress(
4391 VkDevice device,
4392 const VkDeviceMemoryOpaqueCaptureAddressInfoKHR* pInfo)
4393 {
4394 ANV_FROM_HANDLE(anv_device_memory, memory, pInfo->memory);
4395
4396 assert(memory->bo->flags & EXEC_OBJECT_PINNED);
4397 assert(memory->bo->has_client_visible_address);
4398
4399 return intel_48b_address(memory->bo->offset);
4400 }
4401
4402 void
anv_fill_buffer_surface_state(struct anv_device * device,struct anv_state state,enum isl_format format,isl_surf_usage_flags_t usage,struct anv_address address,uint32_t range,uint32_t stride)4403 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
4404 enum isl_format format,
4405 isl_surf_usage_flags_t usage,
4406 struct anv_address address,
4407 uint32_t range, uint32_t stride)
4408 {
4409 isl_buffer_fill_state(&device->isl_dev, state.map,
4410 .address = anv_address_physical(address),
4411 .mocs = isl_mocs(&device->isl_dev, usage,
4412 address.bo && address.bo->is_external),
4413 .size_B = range,
4414 .format = format,
4415 .swizzle = ISL_SWIZZLE_IDENTITY,
4416 .stride_B = stride);
4417 }
4418
anv_DestroySampler(VkDevice _device,VkSampler _sampler,const VkAllocationCallbacks * pAllocator)4419 void anv_DestroySampler(
4420 VkDevice _device,
4421 VkSampler _sampler,
4422 const VkAllocationCallbacks* pAllocator)
4423 {
4424 ANV_FROM_HANDLE(anv_device, device, _device);
4425 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
4426
4427 if (!sampler)
4428 return;
4429
4430 if (sampler->bindless_state.map) {
4431 anv_state_pool_free(&device->dynamic_state_pool,
4432 sampler->bindless_state);
4433 }
4434
4435 if (sampler->custom_border_color.map) {
4436 anv_state_reserved_pool_free(&device->custom_border_colors,
4437 sampler->custom_border_color);
4438 }
4439
4440 vk_object_free(&device->vk, pAllocator, sampler);
4441 }
4442
anv_CreateFramebuffer(VkDevice _device,const VkFramebufferCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkFramebuffer * pFramebuffer)4443 VkResult anv_CreateFramebuffer(
4444 VkDevice _device,
4445 const VkFramebufferCreateInfo* pCreateInfo,
4446 const VkAllocationCallbacks* pAllocator,
4447 VkFramebuffer* pFramebuffer)
4448 {
4449 ANV_FROM_HANDLE(anv_device, device, _device);
4450 struct anv_framebuffer *framebuffer;
4451
4452 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
4453
4454 size_t size = sizeof(*framebuffer);
4455
4456 /* VK_KHR_imageless_framebuffer extension says:
4457 *
4458 * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR,
4459 * parameter pAttachments is ignored.
4460 */
4461 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR))
4462 size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
4463
4464 framebuffer = vk_object_alloc(&device->vk, pAllocator, size,
4465 VK_OBJECT_TYPE_FRAMEBUFFER);
4466 if (framebuffer == NULL)
4467 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
4468
4469 framebuffer->width = pCreateInfo->width;
4470 framebuffer->height = pCreateInfo->height;
4471 framebuffer->layers = pCreateInfo->layers;
4472
4473 if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) {
4474 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
4475 ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]);
4476 framebuffer->attachments[i] = iview;
4477 }
4478 framebuffer->attachment_count = pCreateInfo->attachmentCount;
4479 }
4480
4481 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
4482
4483 return VK_SUCCESS;
4484 }
4485
anv_DestroyFramebuffer(VkDevice _device,VkFramebuffer _fb,const VkAllocationCallbacks * pAllocator)4486 void anv_DestroyFramebuffer(
4487 VkDevice _device,
4488 VkFramebuffer _fb,
4489 const VkAllocationCallbacks* pAllocator)
4490 {
4491 ANV_FROM_HANDLE(anv_device, device, _device);
4492 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
4493
4494 if (!fb)
4495 return;
4496
4497 vk_object_free(&device->vk, pAllocator, fb);
4498 }
4499
4500 static const VkTimeDomainEXT anv_time_domains[] = {
4501 VK_TIME_DOMAIN_DEVICE_EXT,
4502 VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT,
4503 #ifdef CLOCK_MONOTONIC_RAW
4504 VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT,
4505 #endif
4506 };
4507
anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(VkPhysicalDevice physicalDevice,uint32_t * pTimeDomainCount,VkTimeDomainEXT * pTimeDomains)4508 VkResult anv_GetPhysicalDeviceCalibrateableTimeDomainsEXT(
4509 VkPhysicalDevice physicalDevice,
4510 uint32_t *pTimeDomainCount,
4511 VkTimeDomainEXT *pTimeDomains)
4512 {
4513 int d;
4514 VK_OUTARRAY_MAKE(out, pTimeDomains, pTimeDomainCount);
4515
4516 for (d = 0; d < ARRAY_SIZE(anv_time_domains); d++) {
4517 vk_outarray_append(&out, i) {
4518 *i = anv_time_domains[d];
4519 }
4520 }
4521
4522 return vk_outarray_status(&out);
4523 }
4524
4525 static uint64_t
anv_clock_gettime(clockid_t clock_id)4526 anv_clock_gettime(clockid_t clock_id)
4527 {
4528 struct timespec current;
4529 int ret;
4530
4531 ret = clock_gettime(clock_id, ¤t);
4532 #ifdef CLOCK_MONOTONIC_RAW
4533 if (ret < 0 && clock_id == CLOCK_MONOTONIC_RAW)
4534 ret = clock_gettime(CLOCK_MONOTONIC, ¤t);
4535 #endif
4536 if (ret < 0)
4537 return 0;
4538
4539 return (uint64_t) current.tv_sec * 1000000000ULL + current.tv_nsec;
4540 }
4541
anv_GetCalibratedTimestampsEXT(VkDevice _device,uint32_t timestampCount,const VkCalibratedTimestampInfoEXT * pTimestampInfos,uint64_t * pTimestamps,uint64_t * pMaxDeviation)4542 VkResult anv_GetCalibratedTimestampsEXT(
4543 VkDevice _device,
4544 uint32_t timestampCount,
4545 const VkCalibratedTimestampInfoEXT *pTimestampInfos,
4546 uint64_t *pTimestamps,
4547 uint64_t *pMaxDeviation)
4548 {
4549 ANV_FROM_HANDLE(anv_device, device, _device);
4550 uint64_t timestamp_frequency = device->info.timestamp_frequency;
4551 int ret;
4552 int d;
4553 uint64_t begin, end;
4554 uint64_t max_clock_period = 0;
4555
4556 #ifdef CLOCK_MONOTONIC_RAW
4557 begin = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4558 #else
4559 begin = anv_clock_gettime(CLOCK_MONOTONIC);
4560 #endif
4561
4562 for (d = 0; d < timestampCount; d++) {
4563 switch (pTimestampInfos[d].timeDomain) {
4564 case VK_TIME_DOMAIN_DEVICE_EXT:
4565 ret = anv_gem_reg_read(device->fd, TIMESTAMP | I915_REG_READ_8B_WA,
4566 &pTimestamps[d]);
4567
4568 if (ret != 0) {
4569 return anv_device_set_lost(device, "Failed to read the TIMESTAMP "
4570 "register: %m");
4571 }
4572 uint64_t device_period = DIV_ROUND_UP(1000000000, timestamp_frequency);
4573 max_clock_period = MAX2(max_clock_period, device_period);
4574 break;
4575 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_EXT:
4576 pTimestamps[d] = anv_clock_gettime(CLOCK_MONOTONIC);
4577 max_clock_period = MAX2(max_clock_period, 1);
4578 break;
4579
4580 #ifdef CLOCK_MONOTONIC_RAW
4581 case VK_TIME_DOMAIN_CLOCK_MONOTONIC_RAW_EXT:
4582 pTimestamps[d] = begin;
4583 break;
4584 #endif
4585 default:
4586 pTimestamps[d] = 0;
4587 break;
4588 }
4589 }
4590
4591 #ifdef CLOCK_MONOTONIC_RAW
4592 end = anv_clock_gettime(CLOCK_MONOTONIC_RAW);
4593 #else
4594 end = anv_clock_gettime(CLOCK_MONOTONIC);
4595 #endif
4596
4597 /*
4598 * The maximum deviation is the sum of the interval over which we
4599 * perform the sampling and the maximum period of any sampled
4600 * clock. That's because the maximum skew between any two sampled
4601 * clock edges is when the sampled clock with the largest period is
4602 * sampled at the end of that period but right at the beginning of the
4603 * sampling interval and some other clock is sampled right at the
4604 * begining of its sampling period and right at the end of the
4605 * sampling interval. Let's assume the GPU has the longest clock
4606 * period and that the application is sampling GPU and monotonic:
4607 *
4608 * s e
4609 * w x y z 0 1 2 3 4 5 6 7 8 9 a b c d e f
4610 * Raw -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4611 *
4612 * g
4613 * 0 1 2 3
4614 * GPU -----_____-----_____-----_____-----_____
4615 *
4616 * m
4617 * x y z 0 1 2 3 4 5 6 7 8 9 a b c
4618 * Monotonic -_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
4619 *
4620 * Interval <----------------->
4621 * Deviation <-------------------------->
4622 *
4623 * s = read(raw) 2
4624 * g = read(GPU) 1
4625 * m = read(monotonic) 2
4626 * e = read(raw) b
4627 *
4628 * We round the sample interval up by one tick to cover sampling error
4629 * in the interval clock
4630 */
4631
4632 uint64_t sample_interval = end - begin + 1;
4633
4634 *pMaxDeviation = sample_interval + max_clock_period;
4635
4636 return VK_SUCCESS;
4637 }
4638
anv_GetPhysicalDeviceMultisamplePropertiesEXT(VkPhysicalDevice physicalDevice,VkSampleCountFlagBits samples,VkMultisamplePropertiesEXT * pMultisampleProperties)4639 void anv_GetPhysicalDeviceMultisamplePropertiesEXT(
4640 VkPhysicalDevice physicalDevice,
4641 VkSampleCountFlagBits samples,
4642 VkMultisamplePropertiesEXT* pMultisampleProperties)
4643 {
4644 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
4645
4646 assert(pMultisampleProperties->sType ==
4647 VK_STRUCTURE_TYPE_MULTISAMPLE_PROPERTIES_EXT);
4648
4649 VkExtent2D grid_size;
4650 if (samples & isl_device_get_sample_counts(&physical_device->isl_dev)) {
4651 grid_size.width = 1;
4652 grid_size.height = 1;
4653 } else {
4654 grid_size.width = 0;
4655 grid_size.height = 0;
4656 }
4657 pMultisampleProperties->maxSampleLocationGridSize = grid_size;
4658
4659 vk_foreach_struct(ext, pMultisampleProperties->pNext)
4660 anv_debug_ignored_stype(ext->sType);
4661 }
4662
4663 /* vk_icd.h does not declare this function, so we declare it here to
4664 * suppress Wmissing-prototypes.
4665 */
4666 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
4667 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
4668
4669 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t * pSupportedVersion)4670 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
4671 {
4672 /* For the full details on loader interface versioning, see
4673 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
4674 * What follows is a condensed summary, to help you navigate the large and
4675 * confusing official doc.
4676 *
4677 * - Loader interface v0 is incompatible with later versions. We don't
4678 * support it.
4679 *
4680 * - In loader interface v1:
4681 * - The first ICD entrypoint called by the loader is
4682 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
4683 * entrypoint.
4684 * - The ICD must statically expose no other Vulkan symbol unless it is
4685 * linked with -Bsymbolic.
4686 * - Each dispatchable Vulkan handle created by the ICD must be
4687 * a pointer to a struct whose first member is VK_LOADER_DATA. The
4688 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
4689 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
4690 * vkDestroySurfaceKHR(). The ICD must be capable of working with
4691 * such loader-managed surfaces.
4692 *
4693 * - Loader interface v2 differs from v1 in:
4694 * - The first ICD entrypoint called by the loader is
4695 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
4696 * statically expose this entrypoint.
4697 *
4698 * - Loader interface v3 differs from v2 in:
4699 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
4700 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
4701 * because the loader no longer does so.
4702 *
4703 * - Loader interface v4 differs from v3 in:
4704 * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
4705 */
4706 *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
4707 return VK_SUCCESS;
4708 }
4709
anv_GetPhysicalDeviceFragmentShadingRatesKHR(VkPhysicalDevice physicalDevice,uint32_t * pFragmentShadingRateCount,VkPhysicalDeviceFragmentShadingRateKHR * pFragmentShadingRates)4710 VkResult anv_GetPhysicalDeviceFragmentShadingRatesKHR(
4711 VkPhysicalDevice physicalDevice,
4712 uint32_t* pFragmentShadingRateCount,
4713 VkPhysicalDeviceFragmentShadingRateKHR* pFragmentShadingRates)
4714 {
4715 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
4716 VK_OUTARRAY_MAKE(out, pFragmentShadingRates, pFragmentShadingRateCount);
4717
4718 #define append_rate(_samples, _width, _height) \
4719 do { \
4720 vk_outarray_append(&out, __r) { \
4721 __r->sampleCounts = _samples; \
4722 __r->fragmentSize = (VkExtent2D) { \
4723 .width = _width, \
4724 .height = _height, \
4725 }; \
4726 } \
4727 } while (0)
4728
4729 VkSampleCountFlags sample_counts =
4730 isl_device_get_sample_counts(&physical_device->isl_dev);
4731
4732 for (uint32_t x = 4; x >= 1; x /= 2) {
4733 for (uint32_t y = 4; y >= 1; y /= 2) {
4734 /* For size {1, 1}, the sample count must be ~0 */
4735 if (x == 1 && y == 1)
4736 append_rate(~0, x, y);
4737 else
4738 append_rate(sample_counts, x, y);
4739 }
4740 }
4741
4742 #undef append_rate
4743
4744 return vk_outarray_status(&out);
4745 }
4746