2 * Copyright © 2015 Intel Corporation
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:
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
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
28 #include <sys/sysinfo.h>
33 #include "anv_private.h"
34 #include "util/strtod.h"
35 #include "util/debug.h"
36 #include "util/build_id.h"
37 #include "util/mesa-sha1.h"
40 #include "genxml/gen7_pack.h"
43 compiler_debug_log(void *data
, const char *fmt
, ...)
47 compiler_perf_log(void *data
, const char *fmt
, ...)
52 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
53 vfprintf(stderr
, fmt
, args
);
59 anv_compute_heap_size(int fd
, uint64_t *heap_size
)
62 if (anv_gem_get_context_param(fd
, 0, I915_CONTEXT_PARAM_GTT_SIZE
,
64 /* If, for whatever reason, we can't actually get the GTT size from the
65 * kernel (too old?) fall back to the aperture size.
67 anv_perf_warn("Failed to get I915_CONTEXT_PARAM_GTT_SIZE: %m");
69 if (anv_gem_get_aperture(fd
, >t_size
) == -1) {
70 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
71 "failed to get aperture size: %m");
75 /* Query the total ram from the system */
79 uint64_t total_ram
= (uint64_t)info
.totalram
* (uint64_t)info
.mem_unit
;
81 /* We don't want to burn too much ram with the GPU. If the user has 4GiB
82 * or less, we use at most half. If they have more than 4GiB, we use 3/4.
84 uint64_t available_ram
;
85 if (total_ram
<= 4ull * 1024ull * 1024ull * 1024ull)
86 available_ram
= total_ram
/ 2;
88 available_ram
= total_ram
* 3 / 4;
90 /* We also want to leave some padding for things we allocate in the driver,
91 * so don't go over 3/4 of the GTT either.
93 uint64_t available_gtt
= gtt_size
* 3 / 4;
95 *heap_size
= MIN2(available_ram
, available_gtt
);
101 anv_physical_device_init_heaps(struct anv_physical_device
*device
, int fd
)
103 /* The kernel query only tells us whether or not the kernel supports the
104 * EXEC_OBJECT_SUPPORTS_48B_ADDRESS flag and not whether or not the
105 * hardware has actual 48bit address support.
107 device
->supports_48bit_addresses
=
108 (device
->info
.gen
>= 8) && anv_gem_supports_48b_addresses(fd
);
111 VkResult result
= anv_compute_heap_size(fd
, &heap_size
);
112 if (result
!= VK_SUCCESS
)
115 if (heap_size
<= 3ull * (1ull << 30)) {
116 /* In this case, everything fits nicely into the 32-bit address space,
117 * so there's no need for supporting 48bit addresses on client-allocated
120 device
->memory
.heap_count
= 1;
121 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
123 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
124 .supports_48bit_addresses
= false,
127 /* Not everything will fit nicely into a 32-bit address space. In this
128 * case we need a 64-bit heap. Advertise a small 32-bit heap and a
129 * larger 48-bit heap. If we're in this case, then we have a total heap
130 * size larger than 3GiB which most likely means they have 8 GiB of
131 * video memory and so carving off 1 GiB for the 32-bit heap should be
134 const uint64_t heap_size_32bit
= 1ull << 30;
135 const uint64_t heap_size_48bit
= heap_size
- heap_size_32bit
;
137 assert(device
->supports_48bit_addresses
);
139 device
->memory
.heap_count
= 2;
140 device
->memory
.heaps
[0] = (struct anv_memory_heap
) {
141 .size
= heap_size_48bit
,
142 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
143 .supports_48bit_addresses
= true,
145 device
->memory
.heaps
[1] = (struct anv_memory_heap
) {
146 .size
= heap_size_32bit
,
147 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
148 .supports_48bit_addresses
= false,
152 uint32_t type_count
= 0;
153 for (uint32_t heap
= 0; heap
< device
->memory
.heap_count
; heap
++) {
154 uint32_t valid_buffer_usage
= ~0;
156 /* There appears to be a hardware issue in the VF cache where it only
157 * considers the bottom 32 bits of memory addresses. If you happen to
158 * have two vertex buffers which get placed exactly 4 GiB apart and use
159 * them in back-to-back draw calls, you can get collisions. In order to
160 * solve this problem, we require vertex and index buffers be bound to
161 * memory allocated out of the 32-bit heap.
163 if (device
->memory
.heaps
[heap
].supports_48bit_addresses
) {
164 valid_buffer_usage
&= ~(VK_BUFFER_USAGE_INDEX_BUFFER_BIT
|
165 VK_BUFFER_USAGE_VERTEX_BUFFER_BIT
);
168 if (device
->info
.has_llc
) {
169 /* Big core GPUs share LLC with the CPU and thus one memory type can be
170 * both cached and coherent at the same time.
172 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
173 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
174 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
175 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
176 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
178 .valid_buffer_usage
= valid_buffer_usage
,
181 /* The spec requires that we expose a host-visible, coherent memory
182 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
183 * to give the application a choice between cached, but not coherent and
184 * coherent but uncached (WC though).
186 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
187 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
188 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
189 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
191 .valid_buffer_usage
= valid_buffer_usage
,
193 device
->memory
.types
[type_count
++] = (struct anv_memory_type
) {
194 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
195 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
196 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
198 .valid_buffer_usage
= valid_buffer_usage
,
202 device
->memory
.type_count
= type_count
;
208 anv_physical_device_init_uuids(struct anv_physical_device
*device
)
210 const struct build_id_note
*note
= build_id_find_nhdr("libvulkan_intel.so");
212 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
213 "Failed to find build-id");
216 unsigned build_id_len
= build_id_length(note
);
217 if (build_id_len
< 20) {
218 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
219 "build-id too short. It needs to be a SHA");
222 struct mesa_sha1 sha1_ctx
;
224 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
226 /* The pipeline cache UUID is used for determining when a pipeline cache is
227 * invalid. It needs both a driver build and the PCI ID of the device.
229 _mesa_sha1_init(&sha1_ctx
);
230 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
231 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
232 sizeof(device
->chipset_id
));
233 _mesa_sha1_final(&sha1_ctx
, sha1
);
234 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
236 /* The driver UUID is used for determining sharability of images and memory
237 * between two Vulkan instances in separate processes. People who want to
238 * share memory need to also check the device UUID (below) so all this
239 * needs to be is the build-id.
241 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
243 /* The device UUID uniquely identifies the given device within the machine.
244 * Since we never have more than one device, this doesn't need to be a real
245 * UUID. However, on the off-chance that someone tries to use this to
246 * cache pre-tiled images or something of the like, we use the PCI ID and
247 * some bits of ISL info to ensure that this is safe.
249 _mesa_sha1_init(&sha1_ctx
);
250 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
251 sizeof(device
->chipset_id
));
252 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
253 sizeof(device
->isl_dev
.has_bit6_swizzling
));
254 _mesa_sha1_final(&sha1_ctx
, sha1
);
255 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
261 anv_physical_device_init(struct anv_physical_device
*device
,
262 struct anv_instance
*instance
,
268 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
270 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
272 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
273 device
->instance
= instance
;
275 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
276 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
278 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
279 if (!device
->chipset_id
) {
280 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
284 device
->name
= gen_get_device_name(device
->chipset_id
);
285 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
286 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
290 if (device
->info
.is_haswell
) {
291 fprintf(stderr
, "WARNING: Haswell Vulkan support is incomplete\n");
292 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
293 fprintf(stderr
, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
294 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
295 fprintf(stderr
, "WARNING: Bay Trail Vulkan support is incomplete\n");
296 } else if (device
->info
.gen
>= 8 && device
->info
.gen
<= 9) {
297 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
298 * supported as anything */
300 result
= vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
301 "Vulkan not yet supported on %s", device
->name
);
305 device
->cmd_parser_version
= -1;
306 if (device
->info
.gen
== 7) {
307 device
->cmd_parser_version
=
308 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
309 if (device
->cmd_parser_version
== -1) {
310 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
311 "failed to get command parser version");
316 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
317 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
318 "kernel missing gem wait");
322 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
323 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
324 "kernel missing execbuf2");
328 if (!device
->info
.has_llc
&&
329 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
330 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
331 "kernel missing wc mmap");
335 result
= anv_physical_device_init_heaps(device
, fd
);
336 if (result
!= VK_SUCCESS
)
339 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
341 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
343 /* GENs prior to 8 do not support EU/Subslice info */
344 if (device
->info
.gen
>= 8) {
345 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
346 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
348 /* Without this information, we cannot get the right Braswell
349 * brandstrings, and we have to use conservative numbers for GPGPU on
350 * many platforms, but otherwise, things will just work.
352 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
353 fprintf(stderr
, "WARNING: Kernel 4.1 required to properly"
354 " query GPU properties.\n");
356 } else if (device
->info
.gen
== 7) {
357 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
360 if (device
->info
.is_cherryview
&&
361 device
->subslice_total
> 0 && device
->eu_total
> 0) {
362 /* Logical CS threads = EUs per subslice * num threads per EU */
363 uint32_t max_cs_threads
=
364 device
->eu_total
/ device
->subslice_total
* device
->info
.num_thread_per_eu
;
366 /* Fuse configurations may give more threads than expected, never less. */
367 if (max_cs_threads
> device
->info
.max_cs_threads
)
368 device
->info
.max_cs_threads
= max_cs_threads
;
371 brw_process_intel_debug_variable();
373 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
374 if (device
->compiler
== NULL
) {
375 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
378 device
->compiler
->shader_debug_log
= compiler_debug_log
;
379 device
->compiler
->shader_perf_log
= compiler_perf_log
;
381 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
383 result
= anv_physical_device_init_uuids(device
);
384 if (result
!= VK_SUCCESS
)
387 result
= anv_init_wsi(device
);
388 if (result
!= VK_SUCCESS
) {
389 ralloc_free(device
->compiler
);
393 device
->local_fd
= fd
;
402 anv_physical_device_finish(struct anv_physical_device
*device
)
404 anv_finish_wsi(device
);
405 ralloc_free(device
->compiler
);
406 close(device
->local_fd
);
409 static const VkExtensionProperties global_extensions
[] = {
411 .extensionName
= VK_KHR_EXTERNAL_MEMORY_CAPABILITIES_EXTENSION_NAME
,
415 .extensionName
= VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME
,
419 .extensionName
= VK_KHR_GET_SURFACE_CAPABILITIES_2_EXTENSION_NAME
,
423 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
426 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
428 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
432 #ifdef VK_USE_PLATFORM_XCB_KHR
434 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
438 #ifdef VK_USE_PLATFORM_XLIB_KHR
440 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
446 static const VkExtensionProperties device_extensions
[] = {
448 .extensionName
= VK_KHR_DEDICATED_ALLOCATION_EXTENSION_NAME
,
452 .extensionName
= VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME
,
456 .extensionName
= VK_KHR_EXTERNAL_MEMORY_EXTENSION_NAME
,
460 .extensionName
= VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME
,
464 .extensionName
= VK_KHR_GET_MEMORY_REQUIREMENTS_2_EXTENSION_NAME
,
468 .extensionName
= VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME
,
472 .extensionName
= VK_KHR_MAINTENANCE1_EXTENSION_NAME
,
476 .extensionName
= VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME
,
480 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
484 .extensionName
= VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME
,
488 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
492 .extensionName
= VK_KHX_MULTIVIEW_EXTENSION_NAME
,
498 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
499 VkSystemAllocationScope allocationScope
)
505 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
506 size_t align
, VkSystemAllocationScope allocationScope
)
508 return realloc(pOriginal
, size
);
512 default_free_func(void *pUserData
, void *pMemory
)
517 static const VkAllocationCallbacks default_alloc
= {
519 .pfnAllocation
= default_alloc_func
,
520 .pfnReallocation
= default_realloc_func
,
521 .pfnFree
= default_free_func
,
524 VkResult
anv_CreateInstance(
525 const VkInstanceCreateInfo
* pCreateInfo
,
526 const VkAllocationCallbacks
* pAllocator
,
527 VkInstance
* pInstance
)
529 struct anv_instance
*instance
;
531 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
533 uint32_t client_version
;
534 if (pCreateInfo
->pApplicationInfo
&&
535 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
536 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
538 client_version
= VK_MAKE_VERSION(1, 0, 0);
541 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
542 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
543 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
544 "Client requested version %d.%d.%d",
545 VK_VERSION_MAJOR(client_version
),
546 VK_VERSION_MINOR(client_version
),
547 VK_VERSION_PATCH(client_version
));
550 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
552 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
553 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
554 global_extensions
[j
].extensionName
) == 0) {
560 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
563 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
564 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
566 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
568 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
571 instance
->alloc
= *pAllocator
;
573 instance
->alloc
= default_alloc
;
575 instance
->apiVersion
= client_version
;
576 instance
->physicalDeviceCount
= -1;
580 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
582 *pInstance
= anv_instance_to_handle(instance
);
587 void anv_DestroyInstance(
588 VkInstance _instance
,
589 const VkAllocationCallbacks
* pAllocator
)
591 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
596 if (instance
->physicalDeviceCount
> 0) {
597 /* We support at most one physical device. */
598 assert(instance
->physicalDeviceCount
== 1);
599 anv_physical_device_finish(&instance
->physicalDevice
);
602 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
606 vk_free(&instance
->alloc
, instance
);
610 anv_enumerate_devices(struct anv_instance
*instance
)
612 /* TODO: Check for more devices ? */
613 drmDevicePtr devices
[8];
614 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
617 instance
->physicalDeviceCount
= 0;
619 max_devices
= drmGetDevices2(0, devices
, ARRAY_SIZE(devices
));
621 return VK_ERROR_INCOMPATIBLE_DRIVER
;
623 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
624 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
625 devices
[i
]->bustype
== DRM_BUS_PCI
&&
626 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
628 result
= anv_physical_device_init(&instance
->physicalDevice
,
630 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
631 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
635 drmFreeDevices(devices
, max_devices
);
637 if (result
== VK_SUCCESS
)
638 instance
->physicalDeviceCount
= 1;
644 VkResult
anv_EnumeratePhysicalDevices(
645 VkInstance _instance
,
646 uint32_t* pPhysicalDeviceCount
,
647 VkPhysicalDevice
* pPhysicalDevices
)
649 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
650 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
653 if (instance
->physicalDeviceCount
< 0) {
654 result
= anv_enumerate_devices(instance
);
655 if (result
!= VK_SUCCESS
&&
656 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
660 if (instance
->physicalDeviceCount
> 0) {
661 assert(instance
->physicalDeviceCount
== 1);
662 vk_outarray_append(&out
, i
) {
663 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
667 return vk_outarray_status(&out
);
670 void anv_GetPhysicalDeviceFeatures(
671 VkPhysicalDevice physicalDevice
,
672 VkPhysicalDeviceFeatures
* pFeatures
)
674 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
676 *pFeatures
= (VkPhysicalDeviceFeatures
) {
677 .robustBufferAccess
= true,
678 .fullDrawIndexUint32
= true,
679 .imageCubeArray
= true,
680 .independentBlend
= true,
681 .geometryShader
= true,
682 .tessellationShader
= true,
683 .sampleRateShading
= true,
684 .dualSrcBlend
= true,
686 .multiDrawIndirect
= true,
687 .drawIndirectFirstInstance
= true,
689 .depthBiasClamp
= true,
690 .fillModeNonSolid
= true,
691 .depthBounds
= false,
695 .multiViewport
= true,
696 .samplerAnisotropy
= true,
697 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
698 pdevice
->info
.is_baytrail
,
699 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
700 .textureCompressionBC
= true,
701 .occlusionQueryPrecise
= true,
702 .pipelineStatisticsQuery
= true,
703 .fragmentStoresAndAtomics
= true,
704 .shaderTessellationAndGeometryPointSize
= true,
705 .shaderImageGatherExtended
= true,
706 .shaderStorageImageExtendedFormats
= true,
707 .shaderStorageImageMultisample
= false,
708 .shaderStorageImageReadWithoutFormat
= false,
709 .shaderStorageImageWriteWithoutFormat
= true,
710 .shaderUniformBufferArrayDynamicIndexing
= true,
711 .shaderSampledImageArrayDynamicIndexing
= true,
712 .shaderStorageBufferArrayDynamicIndexing
= true,
713 .shaderStorageImageArrayDynamicIndexing
= true,
714 .shaderClipDistance
= true,
715 .shaderCullDistance
= true,
716 .shaderFloat64
= pdevice
->info
.gen
>= 8,
717 .shaderInt64
= pdevice
->info
.gen
>= 8,
718 .shaderInt16
= false,
719 .shaderResourceMinLod
= false,
720 .variableMultisampleRate
= false,
721 .inheritedQueries
= true,
724 /* We can't do image stores in vec4 shaders */
725 pFeatures
->vertexPipelineStoresAndAtomics
=
726 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
727 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
730 void anv_GetPhysicalDeviceFeatures2KHR(
731 VkPhysicalDevice physicalDevice
,
732 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
734 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
736 vk_foreach_struct(ext
, pFeatures
->pNext
) {
737 switch (ext
->sType
) {
738 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES_KHX
: {
739 VkPhysicalDeviceMultiviewFeaturesKHX
*features
=
740 (VkPhysicalDeviceMultiviewFeaturesKHX
*)ext
;
741 features
->multiview
= true;
742 features
->multiviewGeometryShader
= true;
743 features
->multiviewTessellationShader
= true;
748 anv_debug_ignored_stype(ext
->sType
);
754 void anv_GetPhysicalDeviceProperties(
755 VkPhysicalDevice physicalDevice
,
756 VkPhysicalDeviceProperties
* pProperties
)
758 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
759 const struct gen_device_info
*devinfo
= &pdevice
->info
;
761 /* See assertions made when programming the buffer surface state. */
762 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
763 (1ul << 30) : (1ul << 27);
765 const uint32_t max_samplers
= (devinfo
->gen
>= 8 || devinfo
->is_haswell
) ?
768 VkSampleCountFlags sample_counts
=
769 isl_device_get_sample_counts(&pdevice
->isl_dev
);
771 VkPhysicalDeviceLimits limits
= {
772 .maxImageDimension1D
= (1 << 14),
773 .maxImageDimension2D
= (1 << 14),
774 .maxImageDimension3D
= (1 << 11),
775 .maxImageDimensionCube
= (1 << 14),
776 .maxImageArrayLayers
= (1 << 11),
777 .maxTexelBufferElements
= 128 * 1024 * 1024,
778 .maxUniformBufferRange
= (1ul << 27),
779 .maxStorageBufferRange
= max_raw_buffer_sz
,
780 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
781 .maxMemoryAllocationCount
= UINT32_MAX
,
782 .maxSamplerAllocationCount
= 64 * 1024,
783 .bufferImageGranularity
= 64, /* A cache line */
784 .sparseAddressSpaceSize
= 0,
785 .maxBoundDescriptorSets
= MAX_SETS
,
786 .maxPerStageDescriptorSamplers
= max_samplers
,
787 .maxPerStageDescriptorUniformBuffers
= 64,
788 .maxPerStageDescriptorStorageBuffers
= 64,
789 .maxPerStageDescriptorSampledImages
= max_samplers
,
790 .maxPerStageDescriptorStorageImages
= 64,
791 .maxPerStageDescriptorInputAttachments
= 64,
792 .maxPerStageResources
= 250,
793 .maxDescriptorSetSamplers
= 256,
794 .maxDescriptorSetUniformBuffers
= 256,
795 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
796 .maxDescriptorSetStorageBuffers
= 256,
797 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
798 .maxDescriptorSetSampledImages
= 256,
799 .maxDescriptorSetStorageImages
= 256,
800 .maxDescriptorSetInputAttachments
= 256,
801 .maxVertexInputAttributes
= MAX_VBS
,
802 .maxVertexInputBindings
= MAX_VBS
,
803 .maxVertexInputAttributeOffset
= 2047,
804 .maxVertexInputBindingStride
= 2048,
805 .maxVertexOutputComponents
= 128,
806 .maxTessellationGenerationLevel
= 64,
807 .maxTessellationPatchSize
= 32,
808 .maxTessellationControlPerVertexInputComponents
= 128,
809 .maxTessellationControlPerVertexOutputComponents
= 128,
810 .maxTessellationControlPerPatchOutputComponents
= 128,
811 .maxTessellationControlTotalOutputComponents
= 2048,
812 .maxTessellationEvaluationInputComponents
= 128,
813 .maxTessellationEvaluationOutputComponents
= 128,
814 .maxGeometryShaderInvocations
= 32,
815 .maxGeometryInputComponents
= 64,
816 .maxGeometryOutputComponents
= 128,
817 .maxGeometryOutputVertices
= 256,
818 .maxGeometryTotalOutputComponents
= 1024,
819 .maxFragmentInputComponents
= 128,
820 .maxFragmentOutputAttachments
= 8,
821 .maxFragmentDualSrcAttachments
= 1,
822 .maxFragmentCombinedOutputResources
= 8,
823 .maxComputeSharedMemorySize
= 32768,
824 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
825 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
826 .maxComputeWorkGroupSize
= {
827 16 * devinfo
->max_cs_threads
,
828 16 * devinfo
->max_cs_threads
,
829 16 * devinfo
->max_cs_threads
,
831 .subPixelPrecisionBits
= 4 /* FIXME */,
832 .subTexelPrecisionBits
= 4 /* FIXME */,
833 .mipmapPrecisionBits
= 4 /* FIXME */,
834 .maxDrawIndexedIndexValue
= UINT32_MAX
,
835 .maxDrawIndirectCount
= UINT32_MAX
,
836 .maxSamplerLodBias
= 16,
837 .maxSamplerAnisotropy
= 16,
838 .maxViewports
= MAX_VIEWPORTS
,
839 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
840 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
841 .viewportSubPixelBits
= 13, /* We take a float? */
842 .minMemoryMapAlignment
= 4096, /* A page */
843 .minTexelBufferOffsetAlignment
= 1,
844 .minUniformBufferOffsetAlignment
= 16,
845 .minStorageBufferOffsetAlignment
= 4,
846 .minTexelOffset
= -8,
848 .minTexelGatherOffset
= -32,
849 .maxTexelGatherOffset
= 31,
850 .minInterpolationOffset
= -0.5,
851 .maxInterpolationOffset
= 0.4375,
852 .subPixelInterpolationOffsetBits
= 4,
853 .maxFramebufferWidth
= (1 << 14),
854 .maxFramebufferHeight
= (1 << 14),
855 .maxFramebufferLayers
= (1 << 11),
856 .framebufferColorSampleCounts
= sample_counts
,
857 .framebufferDepthSampleCounts
= sample_counts
,
858 .framebufferStencilSampleCounts
= sample_counts
,
859 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
860 .maxColorAttachments
= MAX_RTS
,
861 .sampledImageColorSampleCounts
= sample_counts
,
862 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
863 .sampledImageDepthSampleCounts
= sample_counts
,
864 .sampledImageStencilSampleCounts
= sample_counts
,
865 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
866 .maxSampleMaskWords
= 1,
867 .timestampComputeAndGraphics
= false,
868 .timestampPeriod
= 1000000000.0 / devinfo
->timestamp_frequency
,
869 .maxClipDistances
= 8,
870 .maxCullDistances
= 8,
871 .maxCombinedClipAndCullDistances
= 8,
872 .discreteQueuePriorities
= 1,
873 .pointSizeRange
= { 0.125, 255.875 },
874 .lineWidthRange
= { 0.0, 7.9921875 },
875 .pointSizeGranularity
= (1.0 / 8.0),
876 .lineWidthGranularity
= (1.0 / 128.0),
877 .strictLines
= false, /* FINISHME */
878 .standardSampleLocations
= true,
879 .optimalBufferCopyOffsetAlignment
= 128,
880 .optimalBufferCopyRowPitchAlignment
= 128,
881 .nonCoherentAtomSize
= 64,
884 *pProperties
= (VkPhysicalDeviceProperties
) {
885 .apiVersion
= VK_MAKE_VERSION(1, 0, 54),
886 .driverVersion
= vk_get_driver_version(),
888 .deviceID
= pdevice
->chipset_id
,
889 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
891 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
894 snprintf(pProperties
->deviceName
, sizeof(pProperties
->deviceName
),
895 "%s", pdevice
->name
);
896 memcpy(pProperties
->pipelineCacheUUID
,
897 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
900 void anv_GetPhysicalDeviceProperties2KHR(
901 VkPhysicalDevice physicalDevice
,
902 VkPhysicalDeviceProperties2KHR
* pProperties
)
904 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
906 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
908 vk_foreach_struct(ext
, pProperties
->pNext
) {
909 switch (ext
->sType
) {
910 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
911 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
912 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
914 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
918 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHR
: {
919 VkPhysicalDeviceIDPropertiesKHR
*id_props
=
920 (VkPhysicalDeviceIDPropertiesKHR
*)ext
;
921 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
922 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
923 /* The LUID is for Windows. */
924 id_props
->deviceLUIDValid
= false;
928 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES_KHX
: {
929 VkPhysicalDeviceMultiviewPropertiesKHX
*properties
=
930 (VkPhysicalDeviceMultiviewPropertiesKHX
*)ext
;
931 properties
->maxMultiviewViewCount
= 16;
932 properties
->maxMultiviewInstanceIndex
= UINT32_MAX
/ 16;
937 anv_debug_ignored_stype(ext
->sType
);
943 /* We support exactly one queue family. */
944 static const VkQueueFamilyProperties
945 anv_queue_family_properties
= {
946 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
947 VK_QUEUE_COMPUTE_BIT
|
948 VK_QUEUE_TRANSFER_BIT
,
950 .timestampValidBits
= 36, /* XXX: Real value here */
951 .minImageTransferGranularity
= { 1, 1, 1 },
954 void anv_GetPhysicalDeviceQueueFamilyProperties(
955 VkPhysicalDevice physicalDevice
,
957 VkQueueFamilyProperties
* pQueueFamilyProperties
)
959 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
961 vk_outarray_append(&out
, p
) {
962 *p
= anv_queue_family_properties
;
966 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
967 VkPhysicalDevice physicalDevice
,
968 uint32_t* pQueueFamilyPropertyCount
,
969 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
972 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
974 vk_outarray_append(&out
, p
) {
975 p
->queueFamilyProperties
= anv_queue_family_properties
;
977 vk_foreach_struct(s
, p
->pNext
) {
978 anv_debug_ignored_stype(s
->sType
);
983 void anv_GetPhysicalDeviceMemoryProperties(
984 VkPhysicalDevice physicalDevice
,
985 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
987 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
989 pMemoryProperties
->memoryTypeCount
= physical_device
->memory
.type_count
;
990 for (uint32_t i
= 0; i
< physical_device
->memory
.type_count
; i
++) {
991 pMemoryProperties
->memoryTypes
[i
] = (VkMemoryType
) {
992 .propertyFlags
= physical_device
->memory
.types
[i
].propertyFlags
,
993 .heapIndex
= physical_device
->memory
.types
[i
].heapIndex
,
997 pMemoryProperties
->memoryHeapCount
= physical_device
->memory
.heap_count
;
998 for (uint32_t i
= 0; i
< physical_device
->memory
.heap_count
; i
++) {
999 pMemoryProperties
->memoryHeaps
[i
] = (VkMemoryHeap
) {
1000 .size
= physical_device
->memory
.heaps
[i
].size
,
1001 .flags
= physical_device
->memory
.heaps
[i
].flags
,
1006 void anv_GetPhysicalDeviceMemoryProperties2KHR(
1007 VkPhysicalDevice physicalDevice
,
1008 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
1010 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
1011 &pMemoryProperties
->memoryProperties
);
1013 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
1014 switch (ext
->sType
) {
1016 anv_debug_ignored_stype(ext
->sType
);
1022 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
1023 VkInstance instance
,
1026 return anv_lookup_entrypoint(NULL
, pName
);
1029 /* With version 1+ of the loader interface the ICD should expose
1030 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
1033 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1034 VkInstance instance
,
1038 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1039 VkInstance instance
,
1042 return anv_GetInstanceProcAddr(instance
, pName
);
1045 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
1049 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1050 return anv_lookup_entrypoint(&device
->info
, pName
);
1054 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
1056 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1057 queue
->device
= device
;
1058 queue
->pool
= &device
->surface_state_pool
;
1062 anv_queue_finish(struct anv_queue
*queue
)
1066 static struct anv_state
1067 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
1069 struct anv_state state
;
1071 state
= anv_state_pool_alloc(pool
, size
, align
);
1072 memcpy(state
.map
, p
, size
);
1074 anv_state_flush(pool
->block_pool
.device
, state
);
1079 struct gen8_border_color
{
1084 /* Pad out to 64 bytes */
1089 anv_device_init_border_colors(struct anv_device
*device
)
1091 static const struct gen8_border_color border_colors
[] = {
1092 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
1093 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
1094 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
1095 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
1096 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
1097 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
1100 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
1101 sizeof(border_colors
), 64,
1105 VkResult
anv_CreateDevice(
1106 VkPhysicalDevice physicalDevice
,
1107 const VkDeviceCreateInfo
* pCreateInfo
,
1108 const VkAllocationCallbacks
* pAllocator
,
1111 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1113 struct anv_device
*device
;
1115 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1117 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1119 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
1120 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1121 device_extensions
[j
].extensionName
) == 0) {
1127 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1130 /* Check enabled features */
1131 if (pCreateInfo
->pEnabledFeatures
) {
1132 VkPhysicalDeviceFeatures supported_features
;
1133 anv_GetPhysicalDeviceFeatures(physicalDevice
, &supported_features
);
1134 VkBool32
*supported_feature
= (VkBool32
*)&supported_features
;
1135 VkBool32
*enabled_feature
= (VkBool32
*)pCreateInfo
->pEnabledFeatures
;
1136 unsigned num_features
= sizeof(VkPhysicalDeviceFeatures
) / sizeof(VkBool32
);
1137 for (uint32_t i
= 0; i
< num_features
; i
++) {
1138 if (enabled_feature
[i
] && !supported_feature
[i
])
1139 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1143 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1145 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1147 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1149 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1150 device
->instance
= physical_device
->instance
;
1151 device
->chipset_id
= physical_device
->chipset_id
;
1152 device
->lost
= false;
1155 device
->alloc
= *pAllocator
;
1157 device
->alloc
= physical_device
->instance
->alloc
;
1159 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1160 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1161 if (device
->fd
== -1) {
1162 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1166 device
->context_id
= anv_gem_create_context(device
);
1167 if (device
->context_id
== -1) {
1168 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1172 device
->info
= physical_device
->info
;
1173 device
->isl_dev
= physical_device
->isl_dev
;
1175 /* On Broadwell and later, we can use batch chaining to more efficiently
1176 * implement growing command buffers. Prior to Haswell, the kernel
1177 * command parser gets in the way and we have to fall back to growing
1180 device
->can_chain_batches
= device
->info
.gen
>= 8;
1182 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1183 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1185 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1186 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1187 goto fail_context_id
;
1190 pthread_condattr_t condattr
;
1191 if (pthread_condattr_init(&condattr
) != 0) {
1192 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1195 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1196 pthread_condattr_destroy(&condattr
);
1197 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1200 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1201 pthread_condattr_destroy(&condattr
);
1202 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1205 pthread_condattr_destroy(&condattr
);
1207 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
1209 result
= anv_bo_cache_init(&device
->bo_cache
);
1210 if (result
!= VK_SUCCESS
)
1211 goto fail_batch_bo_pool
;
1213 result
= anv_state_pool_init(&device
->dynamic_state_pool
, device
, 16384);
1214 if (result
!= VK_SUCCESS
)
1217 result
= anv_state_pool_init(&device
->instruction_state_pool
, device
, 16384);
1218 if (result
!= VK_SUCCESS
)
1219 goto fail_dynamic_state_pool
;
1221 result
= anv_state_pool_init(&device
->surface_state_pool
, device
, 4096);
1222 if (result
!= VK_SUCCESS
)
1223 goto fail_instruction_state_pool
;
1225 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1226 if (result
!= VK_SUCCESS
)
1227 goto fail_surface_state_pool
;
1229 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1231 anv_queue_init(device
, &device
->queue
);
1233 switch (device
->info
.gen
) {
1235 if (!device
->info
.is_haswell
)
1236 result
= gen7_init_device_state(device
);
1238 result
= gen75_init_device_state(device
);
1241 result
= gen8_init_device_state(device
);
1244 result
= gen9_init_device_state(device
);
1247 result
= gen10_init_device_state(device
);
1250 /* Shouldn't get here as we don't create physical devices for any other
1252 unreachable("unhandled gen");
1254 if (result
!= VK_SUCCESS
)
1255 goto fail_workaround_bo
;
1257 anv_device_init_blorp(device
);
1259 anv_device_init_border_colors(device
);
1261 *pDevice
= anv_device_to_handle(device
);
1266 anv_queue_finish(&device
->queue
);
1267 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1268 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1269 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1270 fail_surface_state_pool
:
1271 anv_state_pool_finish(&device
->surface_state_pool
);
1272 fail_instruction_state_pool
:
1273 anv_state_pool_finish(&device
->instruction_state_pool
);
1274 fail_dynamic_state_pool
:
1275 anv_state_pool_finish(&device
->dynamic_state_pool
);
1277 anv_bo_cache_finish(&device
->bo_cache
);
1279 anv_bo_pool_finish(&device
->batch_bo_pool
);
1280 pthread_cond_destroy(&device
->queue_submit
);
1282 pthread_mutex_destroy(&device
->mutex
);
1284 anv_gem_destroy_context(device
, device
->context_id
);
1288 vk_free(&device
->alloc
, device
);
1293 void anv_DestroyDevice(
1295 const VkAllocationCallbacks
* pAllocator
)
1297 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1302 anv_device_finish_blorp(device
);
1304 anv_queue_finish(&device
->queue
);
1306 #ifdef HAVE_VALGRIND
1307 /* We only need to free these to prevent valgrind errors. The backing
1308 * BO will go away in a couple of lines so we don't actually leak.
1310 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1313 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1315 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1316 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1318 anv_state_pool_finish(&device
->surface_state_pool
);
1319 anv_state_pool_finish(&device
->instruction_state_pool
);
1320 anv_state_pool_finish(&device
->dynamic_state_pool
);
1322 anv_bo_cache_finish(&device
->bo_cache
);
1324 anv_bo_pool_finish(&device
->batch_bo_pool
);
1326 pthread_cond_destroy(&device
->queue_submit
);
1327 pthread_mutex_destroy(&device
->mutex
);
1329 anv_gem_destroy_context(device
, device
->context_id
);
1333 vk_free(&device
->alloc
, device
);
1336 VkResult
anv_EnumerateInstanceExtensionProperties(
1337 const char* pLayerName
,
1338 uint32_t* pPropertyCount
,
1339 VkExtensionProperties
* pProperties
)
1341 if (pProperties
== NULL
) {
1342 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1346 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(global_extensions
));
1347 typed_memcpy(pProperties
, global_extensions
, *pPropertyCount
);
1349 if (*pPropertyCount
< ARRAY_SIZE(global_extensions
))
1350 return VK_INCOMPLETE
;
1355 VkResult
anv_EnumerateDeviceExtensionProperties(
1356 VkPhysicalDevice physicalDevice
,
1357 const char* pLayerName
,
1358 uint32_t* pPropertyCount
,
1359 VkExtensionProperties
* pProperties
)
1361 if (pProperties
== NULL
) {
1362 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1366 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(device_extensions
));
1367 typed_memcpy(pProperties
, device_extensions
, *pPropertyCount
);
1369 if (*pPropertyCount
< ARRAY_SIZE(device_extensions
))
1370 return VK_INCOMPLETE
;
1375 VkResult
anv_EnumerateInstanceLayerProperties(
1376 uint32_t* pPropertyCount
,
1377 VkLayerProperties
* pProperties
)
1379 if (pProperties
== NULL
) {
1380 *pPropertyCount
= 0;
1384 /* None supported at this time */
1385 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1388 VkResult
anv_EnumerateDeviceLayerProperties(
1389 VkPhysicalDevice physicalDevice
,
1390 uint32_t* pPropertyCount
,
1391 VkLayerProperties
* pProperties
)
1393 if (pProperties
== NULL
) {
1394 *pPropertyCount
= 0;
1398 /* None supported at this time */
1399 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1402 void anv_GetDeviceQueue(
1404 uint32_t queueNodeIndex
,
1405 uint32_t queueIndex
,
1408 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1410 assert(queueIndex
== 0);
1412 *pQueue
= anv_queue_to_handle(&device
->queue
);
1416 anv_device_query_status(struct anv_device
*device
)
1418 /* This isn't likely as most of the callers of this function already check
1419 * for it. However, it doesn't hurt to check and it potentially lets us
1422 if (unlikely(device
->lost
))
1423 return VK_ERROR_DEVICE_LOST
;
1425 uint32_t active
, pending
;
1426 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1428 /* We don't know the real error. */
1429 device
->lost
= true;
1430 return vk_errorf(VK_ERROR_DEVICE_LOST
, "get_reset_stats failed: %m");
1434 device
->lost
= true;
1435 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1436 "GPU hung on one of our command buffers");
1437 } else if (pending
) {
1438 device
->lost
= true;
1439 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1440 "GPU hung with commands in-flight");
1447 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1449 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1450 * Other usages of the BO (such as on different hardware) will not be
1451 * flagged as "busy" by this ioctl. Use with care.
1453 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1455 return VK_NOT_READY
;
1456 } else if (ret
== -1) {
1457 /* We don't know the real error. */
1458 device
->lost
= true;
1459 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1462 /* Query for device status after the busy call. If the BO we're checking
1463 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1464 * client because it clearly doesn't have valid data. Yes, this most
1465 * likely means an ioctl, but we just did an ioctl to query the busy status
1466 * so it's no great loss.
1468 return anv_device_query_status(device
);
1472 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1475 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1476 if (ret
== -1 && errno
== ETIME
) {
1478 } else if (ret
== -1) {
1479 /* We don't know the real error. */
1480 device
->lost
= true;
1481 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1484 /* Query for device status after the wait. If the BO we're waiting on got
1485 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1486 * because it clearly doesn't have valid data. Yes, this most likely means
1487 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1489 return anv_device_query_status(device
);
1492 VkResult
anv_DeviceWaitIdle(
1495 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1496 if (unlikely(device
->lost
))
1497 return VK_ERROR_DEVICE_LOST
;
1499 struct anv_batch batch
;
1502 batch
.start
= batch
.next
= cmds
;
1503 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1505 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1506 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1508 return anv_device_submit_simple_batch(device
, &batch
);
1512 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1514 uint32_t gem_handle
= anv_gem_create(device
, size
);
1516 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1518 anv_bo_init(bo
, gem_handle
, size
);
1523 VkResult
anv_AllocateMemory(
1525 const VkMemoryAllocateInfo
* pAllocateInfo
,
1526 const VkAllocationCallbacks
* pAllocator
,
1527 VkDeviceMemory
* pMem
)
1529 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1530 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1531 struct anv_device_memory
*mem
;
1532 VkResult result
= VK_SUCCESS
;
1534 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1536 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1537 assert(pAllocateInfo
->allocationSize
> 0);
1539 /* The kernel relocation API has a limitation of a 32-bit delta value
1540 * applied to the address before it is written which, in spite of it being
1541 * unsigned, is treated as signed . Because of the way that this maps to
1542 * the Vulkan API, we cannot handle an offset into a buffer that does not
1543 * fit into a signed 32 bits. The only mechanism we have for dealing with
1544 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1545 * of 2GB each. The Vulkan spec allows us to do this:
1547 * "Some platforms may have a limit on the maximum size of a single
1548 * allocation. For example, certain systems may fail to create
1549 * allocations with a size greater than or equal to 4GB. Such a limit is
1550 * implementation-dependent, and if such a failure occurs then the error
1551 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1553 * We don't use vk_error here because it's not an error so much as an
1554 * indication to the application that the allocation is too large.
1556 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1557 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1559 /* FINISHME: Fail if allocation request exceeds heap size. */
1561 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1562 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1564 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1566 assert(pAllocateInfo
->memoryTypeIndex
< pdevice
->memory
.type_count
);
1567 mem
->type
= &pdevice
->memory
.types
[pAllocateInfo
->memoryTypeIndex
];
1571 const VkImportMemoryFdInfoKHR
*fd_info
=
1572 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHR
);
1574 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1577 if (fd_info
&& fd_info
->handleType
) {
1578 /* At the moment, we only support the OPAQUE_FD memory type which is
1579 * just a GEM buffer.
1581 assert(fd_info
->handleType
==
1582 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
1584 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1585 fd_info
->fd
, pAllocateInfo
->allocationSize
,
1587 if (result
!= VK_SUCCESS
)
1590 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1591 pAllocateInfo
->allocationSize
,
1593 if (result
!= VK_SUCCESS
)
1597 assert(mem
->type
->heapIndex
< pdevice
->memory
.heap_count
);
1598 if (pdevice
->memory
.heaps
[mem
->type
->heapIndex
].supports_48bit_addresses
)
1599 mem
->bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1601 if (pdevice
->has_exec_async
)
1602 mem
->bo
->flags
|= EXEC_OBJECT_ASYNC
;
1604 *pMem
= anv_device_memory_to_handle(mem
);
1609 vk_free2(&device
->alloc
, pAllocator
, mem
);
1614 VkResult
anv_GetMemoryFdKHR(
1616 const VkMemoryGetFdInfoKHR
* pGetFdInfo
,
1619 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1620 ANV_FROM_HANDLE(anv_device_memory
, mem
, pGetFdInfo
->memory
);
1622 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR
);
1624 /* We support only one handle type. */
1625 assert(pGetFdInfo
->handleType
==
1626 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
1628 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1631 VkResult
anv_GetMemoryFdPropertiesKHR(
1633 VkExternalMemoryHandleTypeFlagBitsKHR handleType
,
1635 VkMemoryFdPropertiesKHR
* pMemoryFdProperties
)
1637 /* The valid usage section for this function says:
1639 * "handleType must not be one of the handle types defined as opaque."
1641 * Since we only handle opaque handles for now, there are no FD properties.
1643 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
;
1646 void anv_FreeMemory(
1648 VkDeviceMemory _mem
,
1649 const VkAllocationCallbacks
* pAllocator
)
1651 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1652 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1658 anv_UnmapMemory(_device
, _mem
);
1660 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1662 vk_free2(&device
->alloc
, pAllocator
, mem
);
1665 VkResult
anv_MapMemory(
1667 VkDeviceMemory _memory
,
1668 VkDeviceSize offset
,
1670 VkMemoryMapFlags flags
,
1673 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1674 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1681 if (size
== VK_WHOLE_SIZE
)
1682 size
= mem
->bo
->size
- offset
;
1684 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1686 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1687 * assert(size != 0);
1688 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1689 * equal to the size of the memory minus offset
1692 assert(offset
+ size
<= mem
->bo
->size
);
1694 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1695 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1696 * at a time is valid. We could just mmap up front and return an offset
1697 * pointer here, but that may exhaust virtual memory on 32 bit
1700 uint32_t gem_flags
= 0;
1702 if (!device
->info
.has_llc
&&
1703 (mem
->type
->propertyFlags
& VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
))
1704 gem_flags
|= I915_MMAP_WC
;
1706 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1707 uint64_t map_offset
= offset
& ~4095ull;
1708 assert(offset
>= map_offset
);
1709 uint64_t map_size
= (offset
+ size
) - map_offset
;
1711 /* Let's map whole pages */
1712 map_size
= align_u64(map_size
, 4096);
1714 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1715 map_offset
, map_size
, gem_flags
);
1716 if (map
== MAP_FAILED
)
1717 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1720 mem
->map_size
= map_size
;
1722 *ppData
= mem
->map
+ (offset
- map_offset
);
1727 void anv_UnmapMemory(
1729 VkDeviceMemory _memory
)
1731 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1736 anv_gem_munmap(mem
->map
, mem
->map_size
);
1743 clflush_mapped_ranges(struct anv_device
*device
,
1745 const VkMappedMemoryRange
*ranges
)
1747 for (uint32_t i
= 0; i
< count
; i
++) {
1748 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1749 if (ranges
[i
].offset
>= mem
->map_size
)
1752 gen_clflush_range(mem
->map
+ ranges
[i
].offset
,
1753 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1757 VkResult
anv_FlushMappedMemoryRanges(
1759 uint32_t memoryRangeCount
,
1760 const VkMappedMemoryRange
* pMemoryRanges
)
1762 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1764 if (device
->info
.has_llc
)
1767 /* Make sure the writes we're flushing have landed. */
1768 __builtin_ia32_mfence();
1770 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1775 VkResult
anv_InvalidateMappedMemoryRanges(
1777 uint32_t memoryRangeCount
,
1778 const VkMappedMemoryRange
* pMemoryRanges
)
1780 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1782 if (device
->info
.has_llc
)
1785 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1787 /* Make sure no reads get moved up above the invalidate. */
1788 __builtin_ia32_mfence();
1793 void anv_GetBufferMemoryRequirements(
1796 VkMemoryRequirements
* pMemoryRequirements
)
1798 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1799 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1800 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1802 /* The Vulkan spec (git aaed022) says:
1804 * memoryTypeBits is a bitfield and contains one bit set for every
1805 * supported memory type for the resource. The bit `1<<i` is set if and
1806 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1807 * structure for the physical device is supported.
1809 uint32_t memory_types
= 0;
1810 for (uint32_t i
= 0; i
< pdevice
->memory
.type_count
; i
++) {
1811 uint32_t valid_usage
= pdevice
->memory
.types
[i
].valid_buffer_usage
;
1812 if ((valid_usage
& buffer
->usage
) == buffer
->usage
)
1813 memory_types
|= (1u << i
);
1816 pMemoryRequirements
->size
= buffer
->size
;
1817 pMemoryRequirements
->alignment
= 16;
1818 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1821 void anv_GetBufferMemoryRequirements2KHR(
1823 const VkBufferMemoryRequirementsInfo2KHR
* pInfo
,
1824 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1826 anv_GetBufferMemoryRequirements(_device
, pInfo
->buffer
,
1827 &pMemoryRequirements
->memoryRequirements
);
1829 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1830 switch (ext
->sType
) {
1831 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1832 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1833 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
1834 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
1839 anv_debug_ignored_stype(ext
->sType
);
1845 void anv_GetImageMemoryRequirements(
1848 VkMemoryRequirements
* pMemoryRequirements
)
1850 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1851 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1852 struct anv_physical_device
*pdevice
= &device
->instance
->physicalDevice
;
1854 /* The Vulkan spec (git aaed022) says:
1856 * memoryTypeBits is a bitfield and contains one bit set for every
1857 * supported memory type for the resource. The bit `1<<i` is set if and
1858 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1859 * structure for the physical device is supported.
1861 * All types are currently supported for images.
1863 uint32_t memory_types
= (1ull << pdevice
->memory
.type_count
) - 1;
1865 pMemoryRequirements
->size
= image
->size
;
1866 pMemoryRequirements
->alignment
= image
->alignment
;
1867 pMemoryRequirements
->memoryTypeBits
= memory_types
;
1870 void anv_GetImageMemoryRequirements2KHR(
1872 const VkImageMemoryRequirementsInfo2KHR
* pInfo
,
1873 VkMemoryRequirements2KHR
* pMemoryRequirements
)
1875 anv_GetImageMemoryRequirements(_device
, pInfo
->image
,
1876 &pMemoryRequirements
->memoryRequirements
);
1878 vk_foreach_struct(ext
, pMemoryRequirements
->pNext
) {
1879 switch (ext
->sType
) {
1880 case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS_KHR
: {
1881 VkMemoryDedicatedRequirementsKHR
*requirements
= (void *)ext
;
1882 requirements
->prefersDedicatedAllocation
= VK_FALSE
;
1883 requirements
->requiresDedicatedAllocation
= VK_FALSE
;
1888 anv_debug_ignored_stype(ext
->sType
);
1894 void anv_GetImageSparseMemoryRequirements(
1897 uint32_t* pSparseMemoryRequirementCount
,
1898 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1900 *pSparseMemoryRequirementCount
= 0;
1903 void anv_GetImageSparseMemoryRequirements2KHR(
1905 const VkImageSparseMemoryRequirementsInfo2KHR
* pInfo
,
1906 uint32_t* pSparseMemoryRequirementCount
,
1907 VkSparseImageMemoryRequirements2KHR
* pSparseMemoryRequirements
)
1909 *pSparseMemoryRequirementCount
= 0;
1912 void anv_GetDeviceMemoryCommitment(
1914 VkDeviceMemory memory
,
1915 VkDeviceSize
* pCommittedMemoryInBytes
)
1917 *pCommittedMemoryInBytes
= 0;
1920 VkResult
anv_BindBufferMemory(
1923 VkDeviceMemory _memory
,
1924 VkDeviceSize memoryOffset
)
1926 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1927 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1930 assert((buffer
->usage
& mem
->type
->valid_buffer_usage
) == buffer
->usage
);
1931 buffer
->bo
= mem
->bo
;
1932 buffer
->offset
= memoryOffset
;
1941 VkResult
anv_QueueBindSparse(
1943 uint32_t bindInfoCount
,
1944 const VkBindSparseInfo
* pBindInfo
,
1947 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1948 if (unlikely(queue
->device
->lost
))
1949 return VK_ERROR_DEVICE_LOST
;
1951 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1956 VkResult
anv_CreateEvent(
1958 const VkEventCreateInfo
* pCreateInfo
,
1959 const VkAllocationCallbacks
* pAllocator
,
1962 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1963 struct anv_state state
;
1964 struct anv_event
*event
;
1966 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
1968 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
1971 event
->state
= state
;
1972 event
->semaphore
= VK_EVENT_RESET
;
1974 if (!device
->info
.has_llc
) {
1975 /* Make sure the writes we're flushing have landed. */
1976 __builtin_ia32_mfence();
1977 __builtin_ia32_clflush(event
);
1980 *pEvent
= anv_event_to_handle(event
);
1985 void anv_DestroyEvent(
1988 const VkAllocationCallbacks
* pAllocator
)
1990 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1991 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1996 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
1999 VkResult
anv_GetEventStatus(
2003 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2004 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2006 if (unlikely(device
->lost
))
2007 return VK_ERROR_DEVICE_LOST
;
2009 if (!device
->info
.has_llc
) {
2010 /* Invalidate read cache before reading event written by GPU. */
2011 __builtin_ia32_clflush(event
);
2012 __builtin_ia32_mfence();
2016 return event
->semaphore
;
2019 VkResult
anv_SetEvent(
2023 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2024 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2026 event
->semaphore
= VK_EVENT_SET
;
2028 if (!device
->info
.has_llc
) {
2029 /* Make sure the writes we're flushing have landed. */
2030 __builtin_ia32_mfence();
2031 __builtin_ia32_clflush(event
);
2037 VkResult
anv_ResetEvent(
2041 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2042 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2044 event
->semaphore
= VK_EVENT_RESET
;
2046 if (!device
->info
.has_llc
) {
2047 /* Make sure the writes we're flushing have landed. */
2048 __builtin_ia32_mfence();
2049 __builtin_ia32_clflush(event
);
2057 VkResult
anv_CreateBuffer(
2059 const VkBufferCreateInfo
* pCreateInfo
,
2060 const VkAllocationCallbacks
* pAllocator
,
2063 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2064 struct anv_buffer
*buffer
;
2066 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2068 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2069 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2071 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2073 buffer
->size
= pCreateInfo
->size
;
2074 buffer
->usage
= pCreateInfo
->usage
;
2078 *pBuffer
= anv_buffer_to_handle(buffer
);
2083 void anv_DestroyBuffer(
2086 const VkAllocationCallbacks
* pAllocator
)
2088 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2089 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2094 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2098 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2099 enum isl_format format
,
2100 uint32_t offset
, uint32_t range
, uint32_t stride
)
2102 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2104 .mocs
= device
->default_mocs
,
2109 anv_state_flush(device
, state
);
2112 void anv_DestroySampler(
2115 const VkAllocationCallbacks
* pAllocator
)
2117 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2118 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2123 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2126 VkResult
anv_CreateFramebuffer(
2128 const VkFramebufferCreateInfo
* pCreateInfo
,
2129 const VkAllocationCallbacks
* pAllocator
,
2130 VkFramebuffer
* pFramebuffer
)
2132 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2133 struct anv_framebuffer
*framebuffer
;
2135 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2137 size_t size
= sizeof(*framebuffer
) +
2138 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2139 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2140 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2141 if (framebuffer
== NULL
)
2142 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2144 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2145 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2146 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2147 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2150 framebuffer
->width
= pCreateInfo
->width
;
2151 framebuffer
->height
= pCreateInfo
->height
;
2152 framebuffer
->layers
= pCreateInfo
->layers
;
2154 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2159 void anv_DestroyFramebuffer(
2162 const VkAllocationCallbacks
* pAllocator
)
2164 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2165 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2170 vk_free2(&device
->alloc
, pAllocator
, fb
);
2173 /* vk_icd.h does not declare this function, so we declare it here to
2174 * suppress Wmissing-prototypes.
2176 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2177 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2179 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2180 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2182 /* For the full details on loader interface versioning, see
2183 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2184 * What follows is a condensed summary, to help you navigate the large and
2185 * confusing official doc.
2187 * - Loader interface v0 is incompatible with later versions. We don't
2190 * - In loader interface v1:
2191 * - The first ICD entrypoint called by the loader is
2192 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2194 * - The ICD must statically expose no other Vulkan symbol unless it is
2195 * linked with -Bsymbolic.
2196 * - Each dispatchable Vulkan handle created by the ICD must be
2197 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2198 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2199 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2200 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2201 * such loader-managed surfaces.
2203 * - Loader interface v2 differs from v1 in:
2204 * - The first ICD entrypoint called by the loader is
2205 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2206 * statically expose this entrypoint.
2208 * - Loader interface v3 differs from v2 in:
2209 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2210 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2211 * because the loader no longer does so.
2213 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);