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"
38 #include "util/vk_util.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_uuids(struct anv_physical_device
*device
)
103 const struct build_id_note
*note
= build_id_find_nhdr("libvulkan_intel.so");
105 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
106 "Failed to find build-id");
109 unsigned build_id_len
= build_id_length(note
);
110 if (build_id_len
< 20) {
111 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
112 "build-id too short. It needs to be a SHA");
115 struct mesa_sha1 sha1_ctx
;
117 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
119 /* The pipeline cache UUID is used for determining when a pipeline cache is
120 * invalid. It needs both a driver build and the PCI ID of the device.
122 _mesa_sha1_init(&sha1_ctx
);
123 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
124 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
125 sizeof(device
->chipset_id
));
126 _mesa_sha1_final(&sha1_ctx
, sha1
);
127 memcpy(device
->pipeline_cache_uuid
, sha1
, VK_UUID_SIZE
);
129 /* The driver UUID is used for determining sharability of images and memory
130 * between two Vulkan instances in separate processes. People who want to
131 * share memory need to also check the device UUID (below) so all this
132 * needs to be is the build-id.
134 memcpy(device
->driver_uuid
, build_id_data(note
), VK_UUID_SIZE
);
136 /* The device UUID uniquely identifies the given device within the machine.
137 * Since we never have more than one device, this doesn't need to be a real
138 * UUID. However, on the off-chance that someone tries to use this to
139 * cache pre-tiled images or something of the like, we use the PCI ID and
140 * some bits of ISL info to ensure that this is safe.
142 _mesa_sha1_init(&sha1_ctx
);
143 _mesa_sha1_update(&sha1_ctx
, &device
->chipset_id
,
144 sizeof(device
->chipset_id
));
145 _mesa_sha1_update(&sha1_ctx
, &device
->isl_dev
.has_bit6_swizzling
,
146 sizeof(device
->isl_dev
.has_bit6_swizzling
));
147 _mesa_sha1_final(&sha1_ctx
, sha1
);
148 memcpy(device
->device_uuid
, sha1
, VK_UUID_SIZE
);
154 anv_physical_device_init(struct anv_physical_device
*device
,
155 struct anv_instance
*instance
,
161 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
163 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
165 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
166 device
->instance
= instance
;
168 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
169 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
171 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
172 if (!device
->chipset_id
) {
173 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
177 device
->name
= gen_get_device_name(device
->chipset_id
);
178 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
179 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
183 if (device
->info
.is_haswell
) {
184 fprintf(stderr
, "WARNING: Haswell Vulkan support is incomplete\n");
185 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
186 fprintf(stderr
, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
187 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
188 fprintf(stderr
, "WARNING: Bay Trail Vulkan support is incomplete\n");
189 } else if (device
->info
.gen
>= 8) {
190 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
191 * supported as anything */
193 result
= vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
194 "Vulkan not yet supported on %s", device
->name
);
198 device
->cmd_parser_version
= -1;
199 if (device
->info
.gen
== 7) {
200 device
->cmd_parser_version
=
201 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
202 if (device
->cmd_parser_version
== -1) {
203 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
204 "failed to get command parser version");
209 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
210 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
211 "kernel missing gem wait");
215 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
216 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
217 "kernel missing execbuf2");
221 if (!device
->info
.has_llc
&&
222 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
223 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
224 "kernel missing wc mmap");
228 device
->supports_48bit_addresses
= anv_gem_supports_48b_addresses(fd
);
230 result
= anv_compute_heap_size(fd
, &device
->heap_size
);
231 if (result
!= VK_SUCCESS
)
234 device
->has_exec_async
= anv_gem_get_param(fd
, I915_PARAM_HAS_EXEC_ASYNC
);
236 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
238 /* GENs prior to 8 do not support EU/Subslice info */
239 if (device
->info
.gen
>= 8) {
240 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
241 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
243 /* Without this information, we cannot get the right Braswell
244 * brandstrings, and we have to use conservative numbers for GPGPU on
245 * many platforms, but otherwise, things will just work.
247 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
248 fprintf(stderr
, "WARNING: Kernel 4.1 required to properly"
249 " query GPU properties.\n");
251 } else if (device
->info
.gen
== 7) {
252 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
255 if (device
->info
.is_cherryview
&&
256 device
->subslice_total
> 0 && device
->eu_total
> 0) {
257 /* Logical CS threads = EUs per subslice * 7 threads per EU */
258 uint32_t max_cs_threads
= device
->eu_total
/ device
->subslice_total
* 7;
260 /* Fuse configurations may give more threads than expected, never less. */
261 if (max_cs_threads
> device
->info
.max_cs_threads
)
262 device
->info
.max_cs_threads
= max_cs_threads
;
265 brw_process_intel_debug_variable();
267 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
268 if (device
->compiler
== NULL
) {
269 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
272 device
->compiler
->shader_debug_log
= compiler_debug_log
;
273 device
->compiler
->shader_perf_log
= compiler_perf_log
;
275 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
277 result
= anv_physical_device_init_uuids(device
);
278 if (result
!= VK_SUCCESS
)
281 result
= anv_init_wsi(device
);
282 if (result
!= VK_SUCCESS
) {
283 ralloc_free(device
->compiler
);
287 device
->local_fd
= fd
;
296 anv_physical_device_finish(struct anv_physical_device
*device
)
298 anv_finish_wsi(device
);
299 ralloc_free(device
->compiler
);
300 close(device
->local_fd
);
303 static const VkExtensionProperties global_extensions
[] = {
305 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
308 #ifdef VK_USE_PLATFORM_XCB_KHR
310 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
314 #ifdef VK_USE_PLATFORM_XLIB_KHR
316 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
320 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
322 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
327 .extensionName
= VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME
,
331 .extensionName
= VK_KHX_EXTERNAL_MEMORY_CAPABILITIES_EXTENSION_NAME
,
336 static const VkExtensionProperties device_extensions
[] = {
338 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
342 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
346 .extensionName
= VK_KHR_MAINTENANCE1_EXTENSION_NAME
,
350 .extensionName
= VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME
,
354 .extensionName
= VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME
,
358 .extensionName
= VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME
,
362 .extensionName
= VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME
,
366 .extensionName
= VK_KHX_EXTERNAL_MEMORY_EXTENSION_NAME
,
372 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
373 VkSystemAllocationScope allocationScope
)
379 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
380 size_t align
, VkSystemAllocationScope allocationScope
)
382 return realloc(pOriginal
, size
);
386 default_free_func(void *pUserData
, void *pMemory
)
391 static const VkAllocationCallbacks default_alloc
= {
393 .pfnAllocation
= default_alloc_func
,
394 .pfnReallocation
= default_realloc_func
,
395 .pfnFree
= default_free_func
,
398 VkResult
anv_CreateInstance(
399 const VkInstanceCreateInfo
* pCreateInfo
,
400 const VkAllocationCallbacks
* pAllocator
,
401 VkInstance
* pInstance
)
403 struct anv_instance
*instance
;
405 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
407 uint32_t client_version
;
408 if (pCreateInfo
->pApplicationInfo
&&
409 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
410 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
412 client_version
= VK_MAKE_VERSION(1, 0, 0);
415 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
416 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
417 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
418 "Client requested version %d.%d.%d",
419 VK_VERSION_MAJOR(client_version
),
420 VK_VERSION_MINOR(client_version
),
421 VK_VERSION_PATCH(client_version
));
424 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
426 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
427 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
428 global_extensions
[j
].extensionName
) == 0) {
434 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
437 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
438 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
440 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
442 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
445 instance
->alloc
= *pAllocator
;
447 instance
->alloc
= default_alloc
;
449 instance
->apiVersion
= client_version
;
450 instance
->physicalDeviceCount
= -1;
454 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
456 *pInstance
= anv_instance_to_handle(instance
);
461 void anv_DestroyInstance(
462 VkInstance _instance
,
463 const VkAllocationCallbacks
* pAllocator
)
465 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
470 if (instance
->physicalDeviceCount
> 0) {
471 /* We support at most one physical device. */
472 assert(instance
->physicalDeviceCount
== 1);
473 anv_physical_device_finish(&instance
->physicalDevice
);
476 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
480 vk_free(&instance
->alloc
, instance
);
484 anv_enumerate_devices(struct anv_instance
*instance
)
486 /* TODO: Check for more devices ? */
487 drmDevicePtr devices
[8];
488 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
491 instance
->physicalDeviceCount
= 0;
493 max_devices
= drmGetDevices2(0, devices
, sizeof(devices
));
495 return VK_ERROR_INCOMPATIBLE_DRIVER
;
497 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
498 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
499 devices
[i
]->bustype
== DRM_BUS_PCI
&&
500 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
502 result
= anv_physical_device_init(&instance
->physicalDevice
,
504 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
505 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
510 if (result
== VK_SUCCESS
)
511 instance
->physicalDeviceCount
= 1;
517 VkResult
anv_EnumeratePhysicalDevices(
518 VkInstance _instance
,
519 uint32_t* pPhysicalDeviceCount
,
520 VkPhysicalDevice
* pPhysicalDevices
)
522 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
523 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
526 if (instance
->physicalDeviceCount
< 0) {
527 result
= anv_enumerate_devices(instance
);
528 if (result
!= VK_SUCCESS
&&
529 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
533 if (instance
->physicalDeviceCount
> 0) {
534 assert(instance
->physicalDeviceCount
== 1);
535 vk_outarray_append(&out
, i
) {
536 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
540 return vk_outarray_status(&out
);
543 void anv_GetPhysicalDeviceFeatures(
544 VkPhysicalDevice physicalDevice
,
545 VkPhysicalDeviceFeatures
* pFeatures
)
547 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
549 *pFeatures
= (VkPhysicalDeviceFeatures
) {
550 .robustBufferAccess
= true,
551 .fullDrawIndexUint32
= true,
552 .imageCubeArray
= true,
553 .independentBlend
= true,
554 .geometryShader
= true,
555 .tessellationShader
= true,
556 .sampleRateShading
= true,
557 .dualSrcBlend
= true,
559 .multiDrawIndirect
= false,
560 .drawIndirectFirstInstance
= true,
562 .depthBiasClamp
= true,
563 .fillModeNonSolid
= true,
564 .depthBounds
= false,
568 .multiViewport
= true,
569 .samplerAnisotropy
= true,
570 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
571 pdevice
->info
.is_baytrail
,
572 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
573 .textureCompressionBC
= true,
574 .occlusionQueryPrecise
= true,
575 .pipelineStatisticsQuery
= true,
576 .fragmentStoresAndAtomics
= true,
577 .shaderTessellationAndGeometryPointSize
= true,
578 .shaderImageGatherExtended
= true,
579 .shaderStorageImageExtendedFormats
= true,
580 .shaderStorageImageMultisample
= false,
581 .shaderStorageImageReadWithoutFormat
= false,
582 .shaderStorageImageWriteWithoutFormat
= true,
583 .shaderUniformBufferArrayDynamicIndexing
= true,
584 .shaderSampledImageArrayDynamicIndexing
= true,
585 .shaderStorageBufferArrayDynamicIndexing
= true,
586 .shaderStorageImageArrayDynamicIndexing
= true,
587 .shaderClipDistance
= true,
588 .shaderCullDistance
= true,
589 .shaderFloat64
= pdevice
->info
.gen
>= 8,
590 .shaderInt64
= pdevice
->info
.gen
>= 8,
591 .shaderInt16
= false,
592 .shaderResourceMinLod
= false,
593 .variableMultisampleRate
= false,
594 .inheritedQueries
= true,
597 /* We can't do image stores in vec4 shaders */
598 pFeatures
->vertexPipelineStoresAndAtomics
=
599 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
600 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
603 void anv_GetPhysicalDeviceFeatures2KHR(
604 VkPhysicalDevice physicalDevice
,
605 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
607 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
609 vk_foreach_struct(ext
, pFeatures
->pNext
) {
610 switch (ext
->sType
) {
612 anv_debug_ignored_stype(ext
->sType
);
618 void anv_GetPhysicalDeviceProperties(
619 VkPhysicalDevice physicalDevice
,
620 VkPhysicalDeviceProperties
* pProperties
)
622 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
623 const struct gen_device_info
*devinfo
= &pdevice
->info
;
625 /* See assertions made when programming the buffer surface state. */
626 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
627 (1ul << 30) : (1ul << 27);
629 VkSampleCountFlags sample_counts
=
630 isl_device_get_sample_counts(&pdevice
->isl_dev
);
632 VkPhysicalDeviceLimits limits
= {
633 .maxImageDimension1D
= (1 << 14),
634 .maxImageDimension2D
= (1 << 14),
635 .maxImageDimension3D
= (1 << 11),
636 .maxImageDimensionCube
= (1 << 14),
637 .maxImageArrayLayers
= (1 << 11),
638 .maxTexelBufferElements
= 128 * 1024 * 1024,
639 .maxUniformBufferRange
= (1ul << 27),
640 .maxStorageBufferRange
= max_raw_buffer_sz
,
641 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
642 .maxMemoryAllocationCount
= UINT32_MAX
,
643 .maxSamplerAllocationCount
= 64 * 1024,
644 .bufferImageGranularity
= 64, /* A cache line */
645 .sparseAddressSpaceSize
= 0,
646 .maxBoundDescriptorSets
= MAX_SETS
,
647 .maxPerStageDescriptorSamplers
= 64,
648 .maxPerStageDescriptorUniformBuffers
= 64,
649 .maxPerStageDescriptorStorageBuffers
= 64,
650 .maxPerStageDescriptorSampledImages
= 64,
651 .maxPerStageDescriptorStorageImages
= 64,
652 .maxPerStageDescriptorInputAttachments
= 64,
653 .maxPerStageResources
= 128,
654 .maxDescriptorSetSamplers
= 256,
655 .maxDescriptorSetUniformBuffers
= 256,
656 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
657 .maxDescriptorSetStorageBuffers
= 256,
658 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
659 .maxDescriptorSetSampledImages
= 256,
660 .maxDescriptorSetStorageImages
= 256,
661 .maxDescriptorSetInputAttachments
= 256,
662 .maxVertexInputAttributes
= MAX_VBS
,
663 .maxVertexInputBindings
= MAX_VBS
,
664 .maxVertexInputAttributeOffset
= 2047,
665 .maxVertexInputBindingStride
= 2048,
666 .maxVertexOutputComponents
= 128,
667 .maxTessellationGenerationLevel
= 64,
668 .maxTessellationPatchSize
= 32,
669 .maxTessellationControlPerVertexInputComponents
= 128,
670 .maxTessellationControlPerVertexOutputComponents
= 128,
671 .maxTessellationControlPerPatchOutputComponents
= 128,
672 .maxTessellationControlTotalOutputComponents
= 2048,
673 .maxTessellationEvaluationInputComponents
= 128,
674 .maxTessellationEvaluationOutputComponents
= 128,
675 .maxGeometryShaderInvocations
= 32,
676 .maxGeometryInputComponents
= 64,
677 .maxGeometryOutputComponents
= 128,
678 .maxGeometryOutputVertices
= 256,
679 .maxGeometryTotalOutputComponents
= 1024,
680 .maxFragmentInputComponents
= 128,
681 .maxFragmentOutputAttachments
= 8,
682 .maxFragmentDualSrcAttachments
= 1,
683 .maxFragmentCombinedOutputResources
= 8,
684 .maxComputeSharedMemorySize
= 32768,
685 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
686 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
687 .maxComputeWorkGroupSize
= {
688 16 * devinfo
->max_cs_threads
,
689 16 * devinfo
->max_cs_threads
,
690 16 * devinfo
->max_cs_threads
,
692 .subPixelPrecisionBits
= 4 /* FIXME */,
693 .subTexelPrecisionBits
= 4 /* FIXME */,
694 .mipmapPrecisionBits
= 4 /* FIXME */,
695 .maxDrawIndexedIndexValue
= UINT32_MAX
,
696 .maxDrawIndirectCount
= UINT32_MAX
,
697 .maxSamplerLodBias
= 16,
698 .maxSamplerAnisotropy
= 16,
699 .maxViewports
= MAX_VIEWPORTS
,
700 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
701 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
702 .viewportSubPixelBits
= 13, /* We take a float? */
703 .minMemoryMapAlignment
= 4096, /* A page */
704 .minTexelBufferOffsetAlignment
= 1,
705 .minUniformBufferOffsetAlignment
= 16,
706 .minStorageBufferOffsetAlignment
= 4,
707 .minTexelOffset
= -8,
709 .minTexelGatherOffset
= -32,
710 .maxTexelGatherOffset
= 31,
711 .minInterpolationOffset
= -0.5,
712 .maxInterpolationOffset
= 0.4375,
713 .subPixelInterpolationOffsetBits
= 4,
714 .maxFramebufferWidth
= (1 << 14),
715 .maxFramebufferHeight
= (1 << 14),
716 .maxFramebufferLayers
= (1 << 11),
717 .framebufferColorSampleCounts
= sample_counts
,
718 .framebufferDepthSampleCounts
= sample_counts
,
719 .framebufferStencilSampleCounts
= sample_counts
,
720 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
721 .maxColorAttachments
= MAX_RTS
,
722 .sampledImageColorSampleCounts
= sample_counts
,
723 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
724 .sampledImageDepthSampleCounts
= sample_counts
,
725 .sampledImageStencilSampleCounts
= sample_counts
,
726 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
727 .maxSampleMaskWords
= 1,
728 .timestampComputeAndGraphics
= false,
729 .timestampPeriod
= devinfo
->timebase_scale
,
730 .maxClipDistances
= 8,
731 .maxCullDistances
= 8,
732 .maxCombinedClipAndCullDistances
= 8,
733 .discreteQueuePriorities
= 1,
734 .pointSizeRange
= { 0.125, 255.875 },
735 .lineWidthRange
= { 0.0, 7.9921875 },
736 .pointSizeGranularity
= (1.0 / 8.0),
737 .lineWidthGranularity
= (1.0 / 128.0),
738 .strictLines
= false, /* FINISHME */
739 .standardSampleLocations
= true,
740 .optimalBufferCopyOffsetAlignment
= 128,
741 .optimalBufferCopyRowPitchAlignment
= 128,
742 .nonCoherentAtomSize
= 64,
745 *pProperties
= (VkPhysicalDeviceProperties
) {
746 .apiVersion
= VK_MAKE_VERSION(1, 0, 42),
749 .deviceID
= pdevice
->chipset_id
,
750 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
752 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
755 strcpy(pProperties
->deviceName
, pdevice
->name
);
756 memcpy(pProperties
->pipelineCacheUUID
,
757 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
760 void anv_GetPhysicalDeviceProperties2KHR(
761 VkPhysicalDevice physicalDevice
,
762 VkPhysicalDeviceProperties2KHR
* pProperties
)
764 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
766 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
768 vk_foreach_struct(ext
, pProperties
->pNext
) {
769 switch (ext
->sType
) {
770 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
771 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
772 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
774 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
778 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHX
: {
779 VkPhysicalDeviceIDPropertiesKHX
*id_props
=
780 (VkPhysicalDeviceIDPropertiesKHX
*)ext
;
781 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
782 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
783 /* The LUID is for Windows. */
784 id_props
->deviceLUIDValid
= false;
789 anv_debug_ignored_stype(ext
->sType
);
795 /* We support exactly one queue family. */
796 static const VkQueueFamilyProperties
797 anv_queue_family_properties
= {
798 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
799 VK_QUEUE_COMPUTE_BIT
|
800 VK_QUEUE_TRANSFER_BIT
,
802 .timestampValidBits
= 36, /* XXX: Real value here */
803 .minImageTransferGranularity
= { 1, 1, 1 },
806 void anv_GetPhysicalDeviceQueueFamilyProperties(
807 VkPhysicalDevice physicalDevice
,
809 VkQueueFamilyProperties
* pQueueFamilyProperties
)
811 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
813 vk_outarray_append(&out
, p
) {
814 *p
= anv_queue_family_properties
;
818 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
819 VkPhysicalDevice physicalDevice
,
820 uint32_t* pQueueFamilyPropertyCount
,
821 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
824 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
826 vk_outarray_append(&out
, p
) {
827 p
->queueFamilyProperties
= anv_queue_family_properties
;
829 vk_foreach_struct(s
, p
->pNext
) {
830 anv_debug_ignored_stype(s
->sType
);
835 void anv_GetPhysicalDeviceMemoryProperties(
836 VkPhysicalDevice physicalDevice
,
837 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
839 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
841 if (physical_device
->info
.has_llc
) {
842 /* Big core GPUs share LLC with the CPU and thus one memory type can be
843 * both cached and coherent at the same time.
845 pMemoryProperties
->memoryTypeCount
= 1;
846 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
847 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
848 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
849 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
850 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
854 /* The spec requires that we expose a host-visible, coherent memory
855 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
856 * to give the application a choice between cached, but not coherent and
857 * coherent but uncached (WC though).
859 pMemoryProperties
->memoryTypeCount
= 2;
860 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
861 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
862 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
863 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
866 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
867 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
868 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
869 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
874 pMemoryProperties
->memoryHeapCount
= 1;
875 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
876 .size
= physical_device
->heap_size
,
877 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
881 void anv_GetPhysicalDeviceMemoryProperties2KHR(
882 VkPhysicalDevice physicalDevice
,
883 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
885 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
886 &pMemoryProperties
->memoryProperties
);
888 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
889 switch (ext
->sType
) {
891 anv_debug_ignored_stype(ext
->sType
);
897 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
901 return anv_lookup_entrypoint(NULL
, pName
);
904 /* With version 1+ of the loader interface the ICD should expose
905 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
908 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
913 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
917 return anv_GetInstanceProcAddr(instance
, pName
);
920 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
924 ANV_FROM_HANDLE(anv_device
, device
, _device
);
925 return anv_lookup_entrypoint(&device
->info
, pName
);
929 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
931 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
932 queue
->device
= device
;
933 queue
->pool
= &device
->surface_state_pool
;
937 anv_queue_finish(struct anv_queue
*queue
)
941 static struct anv_state
942 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
944 struct anv_state state
;
946 state
= anv_state_pool_alloc(pool
, size
, align
);
947 memcpy(state
.map
, p
, size
);
949 anv_state_flush(pool
->block_pool
->device
, state
);
954 struct gen8_border_color
{
959 /* Pad out to 64 bytes */
964 anv_device_init_border_colors(struct anv_device
*device
)
966 static const struct gen8_border_color border_colors
[] = {
967 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
968 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
969 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
970 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
971 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
972 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
975 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
976 sizeof(border_colors
), 64,
981 anv_device_submit_simple_batch(struct anv_device
*device
,
982 struct anv_batch
*batch
)
984 struct drm_i915_gem_execbuffer2 execbuf
;
985 struct drm_i915_gem_exec_object2 exec2_objects
[1];
986 struct anv_bo bo
, *exec_bos
[1];
987 VkResult result
= VK_SUCCESS
;
990 /* Kernel driver requires 8 byte aligned batch length */
991 size
= align_u32(batch
->next
- batch
->start
, 8);
992 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &bo
, size
);
993 if (result
!= VK_SUCCESS
)
996 memcpy(bo
.map
, batch
->start
, size
);
997 if (!device
->info
.has_llc
)
998 anv_flush_range(bo
.map
, size
);
1001 exec2_objects
[0].handle
= bo
.gem_handle
;
1002 exec2_objects
[0].relocation_count
= 0;
1003 exec2_objects
[0].relocs_ptr
= 0;
1004 exec2_objects
[0].alignment
= 0;
1005 exec2_objects
[0].offset
= bo
.offset
;
1006 exec2_objects
[0].flags
= 0;
1007 exec2_objects
[0].rsvd1
= 0;
1008 exec2_objects
[0].rsvd2
= 0;
1010 execbuf
.buffers_ptr
= (uintptr_t) exec2_objects
;
1011 execbuf
.buffer_count
= 1;
1012 execbuf
.batch_start_offset
= 0;
1013 execbuf
.batch_len
= size
;
1014 execbuf
.cliprects_ptr
= 0;
1015 execbuf
.num_cliprects
= 0;
1020 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
1021 execbuf
.rsvd1
= device
->context_id
;
1024 result
= anv_device_execbuf(device
, &execbuf
, exec_bos
);
1025 if (result
!= VK_SUCCESS
)
1028 result
= anv_device_wait(device
, &bo
, INT64_MAX
);
1031 anv_bo_pool_free(&device
->batch_bo_pool
, &bo
);
1036 VkResult
anv_CreateDevice(
1037 VkPhysicalDevice physicalDevice
,
1038 const VkDeviceCreateInfo
* pCreateInfo
,
1039 const VkAllocationCallbacks
* pAllocator
,
1042 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1044 struct anv_device
*device
;
1046 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1048 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1050 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
1051 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1052 device_extensions
[j
].extensionName
) == 0) {
1058 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1061 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1063 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1065 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1067 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1068 device
->instance
= physical_device
->instance
;
1069 device
->chipset_id
= physical_device
->chipset_id
;
1070 device
->lost
= false;
1073 device
->alloc
= *pAllocator
;
1075 device
->alloc
= physical_device
->instance
->alloc
;
1077 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1078 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1079 if (device
->fd
== -1) {
1080 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1084 device
->context_id
= anv_gem_create_context(device
);
1085 if (device
->context_id
== -1) {
1086 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1090 device
->info
= physical_device
->info
;
1091 device
->isl_dev
= physical_device
->isl_dev
;
1093 /* On Broadwell and later, we can use batch chaining to more efficiently
1094 * implement growing command buffers. Prior to Haswell, the kernel
1095 * command parser gets in the way and we have to fall back to growing
1098 device
->can_chain_batches
= device
->info
.gen
>= 8;
1100 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1101 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1103 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1104 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1105 goto fail_context_id
;
1108 pthread_condattr_t condattr
;
1109 if (pthread_condattr_init(&condattr
) != 0) {
1110 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1113 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1114 pthread_condattr_destroy(&condattr
);
1115 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1118 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1119 pthread_condattr_destroy(&condattr
);
1120 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1123 pthread_condattr_destroy(&condattr
);
1125 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
1127 result
= anv_block_pool_init(&device
->dynamic_state_block_pool
, device
,
1129 if (result
!= VK_SUCCESS
)
1130 goto fail_batch_bo_pool
;
1132 anv_state_pool_init(&device
->dynamic_state_pool
,
1133 &device
->dynamic_state_block_pool
);
1135 result
= anv_block_pool_init(&device
->instruction_block_pool
, device
,
1137 if (result
!= VK_SUCCESS
)
1138 goto fail_dynamic_state_pool
;
1140 anv_state_pool_init(&device
->instruction_state_pool
,
1141 &device
->instruction_block_pool
);
1143 result
= anv_block_pool_init(&device
->surface_state_block_pool
, device
,
1145 if (result
!= VK_SUCCESS
)
1146 goto fail_instruction_state_pool
;
1148 anv_state_pool_init(&device
->surface_state_pool
,
1149 &device
->surface_state_block_pool
);
1151 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1152 if (result
!= VK_SUCCESS
)
1153 goto fail_surface_state_pool
;
1155 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1157 anv_queue_init(device
, &device
->queue
);
1159 switch (device
->info
.gen
) {
1161 if (!device
->info
.is_haswell
)
1162 result
= gen7_init_device_state(device
);
1164 result
= gen75_init_device_state(device
);
1167 result
= gen8_init_device_state(device
);
1170 result
= gen9_init_device_state(device
);
1173 /* Shouldn't get here as we don't create physical devices for any other
1175 unreachable("unhandled gen");
1177 if (result
!= VK_SUCCESS
)
1178 goto fail_workaround_bo
;
1180 anv_device_init_blorp(device
);
1182 anv_device_init_border_colors(device
);
1184 *pDevice
= anv_device_to_handle(device
);
1189 anv_queue_finish(&device
->queue
);
1190 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1191 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1192 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1193 fail_surface_state_pool
:
1194 anv_state_pool_finish(&device
->surface_state_pool
);
1195 anv_block_pool_finish(&device
->surface_state_block_pool
);
1196 fail_instruction_state_pool
:
1197 anv_state_pool_finish(&device
->instruction_state_pool
);
1198 anv_block_pool_finish(&device
->instruction_block_pool
);
1199 fail_dynamic_state_pool
:
1200 anv_state_pool_finish(&device
->dynamic_state_pool
);
1201 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1203 anv_bo_pool_finish(&device
->batch_bo_pool
);
1204 pthread_cond_destroy(&device
->queue_submit
);
1206 pthread_mutex_destroy(&device
->mutex
);
1208 anv_gem_destroy_context(device
, device
->context_id
);
1212 vk_free(&device
->alloc
, device
);
1217 void anv_DestroyDevice(
1219 const VkAllocationCallbacks
* pAllocator
)
1221 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1226 anv_device_finish_blorp(device
);
1228 anv_queue_finish(&device
->queue
);
1230 #ifdef HAVE_VALGRIND
1231 /* We only need to free these to prevent valgrind errors. The backing
1232 * BO will go away in a couple of lines so we don't actually leak.
1234 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1237 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1239 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1240 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1242 anv_state_pool_finish(&device
->surface_state_pool
);
1243 anv_block_pool_finish(&device
->surface_state_block_pool
);
1244 anv_state_pool_finish(&device
->instruction_state_pool
);
1245 anv_block_pool_finish(&device
->instruction_block_pool
);
1246 anv_state_pool_finish(&device
->dynamic_state_pool
);
1247 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1249 anv_bo_pool_finish(&device
->batch_bo_pool
);
1251 pthread_cond_destroy(&device
->queue_submit
);
1252 pthread_mutex_destroy(&device
->mutex
);
1254 anv_gem_destroy_context(device
, device
->context_id
);
1258 vk_free(&device
->alloc
, device
);
1261 VkResult
anv_EnumerateInstanceExtensionProperties(
1262 const char* pLayerName
,
1263 uint32_t* pPropertyCount
,
1264 VkExtensionProperties
* pProperties
)
1266 if (pProperties
== NULL
) {
1267 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1271 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(global_extensions
));
1272 typed_memcpy(pProperties
, global_extensions
, *pPropertyCount
);
1274 if (*pPropertyCount
< ARRAY_SIZE(global_extensions
))
1275 return VK_INCOMPLETE
;
1280 VkResult
anv_EnumerateDeviceExtensionProperties(
1281 VkPhysicalDevice physicalDevice
,
1282 const char* pLayerName
,
1283 uint32_t* pPropertyCount
,
1284 VkExtensionProperties
* pProperties
)
1286 if (pProperties
== NULL
) {
1287 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1291 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(device_extensions
));
1292 typed_memcpy(pProperties
, device_extensions
, *pPropertyCount
);
1294 if (*pPropertyCount
< ARRAY_SIZE(device_extensions
))
1295 return VK_INCOMPLETE
;
1300 VkResult
anv_EnumerateInstanceLayerProperties(
1301 uint32_t* pPropertyCount
,
1302 VkLayerProperties
* pProperties
)
1304 if (pProperties
== NULL
) {
1305 *pPropertyCount
= 0;
1309 /* None supported at this time */
1310 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1313 VkResult
anv_EnumerateDeviceLayerProperties(
1314 VkPhysicalDevice physicalDevice
,
1315 uint32_t* pPropertyCount
,
1316 VkLayerProperties
* pProperties
)
1318 if (pProperties
== NULL
) {
1319 *pPropertyCount
= 0;
1323 /* None supported at this time */
1324 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1327 void anv_GetDeviceQueue(
1329 uint32_t queueNodeIndex
,
1330 uint32_t queueIndex
,
1333 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1335 assert(queueIndex
== 0);
1337 *pQueue
= anv_queue_to_handle(&device
->queue
);
1341 anv_device_execbuf(struct anv_device
*device
,
1342 struct drm_i915_gem_execbuffer2
*execbuf
,
1343 struct anv_bo
**execbuf_bos
)
1345 int ret
= anv_gem_execbuffer(device
, execbuf
);
1347 /* We don't know the real error. */
1348 device
->lost
= true;
1349 return vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
1352 struct drm_i915_gem_exec_object2
*objects
=
1353 (void *)(uintptr_t)execbuf
->buffers_ptr
;
1354 for (uint32_t k
= 0; k
< execbuf
->buffer_count
; k
++)
1355 execbuf_bos
[k
]->offset
= objects
[k
].offset
;
1361 anv_device_query_status(struct anv_device
*device
)
1363 /* This isn't likely as most of the callers of this function already check
1364 * for it. However, it doesn't hurt to check and it potentially lets us
1367 if (unlikely(device
->lost
))
1368 return VK_ERROR_DEVICE_LOST
;
1370 uint32_t active
, pending
;
1371 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1373 /* We don't know the real error. */
1374 device
->lost
= true;
1375 return vk_errorf(VK_ERROR_DEVICE_LOST
, "get_reset_stats failed: %m");
1379 device
->lost
= true;
1380 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1381 "GPU hung on one of our command buffers");
1382 } else if (pending
) {
1383 device
->lost
= true;
1384 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1385 "GPU hung with commands in-flight");
1392 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1394 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1395 * Other usages of the BO (such as on different hardware) will not be
1396 * flagged as "busy" by this ioctl. Use with care.
1398 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1400 return VK_NOT_READY
;
1401 } else if (ret
== -1) {
1402 /* We don't know the real error. */
1403 device
->lost
= true;
1404 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1407 /* Query for device status after the busy call. If the BO we're checking
1408 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1409 * client because it clearly doesn't have valid data. Yes, this most
1410 * likely means an ioctl, but we just did an ioctl to query the busy status
1411 * so it's no great loss.
1413 return anv_device_query_status(device
);
1417 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1420 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1421 if (ret
== -1 && errno
== ETIME
) {
1423 } else if (ret
== -1) {
1424 /* We don't know the real error. */
1425 device
->lost
= true;
1426 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1429 /* Query for device status after the wait. If the BO we're waiting on got
1430 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1431 * because it clearly doesn't have valid data. Yes, this most likely means
1432 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1434 return anv_device_query_status(device
);
1437 VkResult
anv_QueueSubmit(
1439 uint32_t submitCount
,
1440 const VkSubmitInfo
* pSubmits
,
1443 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1444 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1445 struct anv_device
*device
= queue
->device
;
1447 /* Query for device status prior to submitting. Technically, we don't need
1448 * to do this. However, if we have a client that's submitting piles of
1449 * garbage, we would rather break as early as possible to keep the GPU
1450 * hanging contained. If we don't check here, we'll either be waiting for
1451 * the kernel to kick us or we'll have to wait until the client waits on a
1452 * fence before we actually know whether or not we've hung.
1454 VkResult result
= anv_device_query_status(device
);
1455 if (result
!= VK_SUCCESS
)
1458 /* We lock around QueueSubmit for three main reasons:
1460 * 1) When a block pool is resized, we create a new gem handle with a
1461 * different size and, in the case of surface states, possibly a
1462 * different center offset but we re-use the same anv_bo struct when
1463 * we do so. If this happens in the middle of setting up an execbuf,
1464 * we could end up with our list of BOs out of sync with our list of
1467 * 2) The algorithm we use for building the list of unique buffers isn't
1468 * thread-safe. While the client is supposed to syncronize around
1469 * QueueSubmit, this would be extremely difficult to debug if it ever
1470 * came up in the wild due to a broken app. It's better to play it
1471 * safe and just lock around QueueSubmit.
1473 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
1474 * userspace. Due to the fact that the surface state buffer is shared
1475 * between batches, we can't afford to have that happen from multiple
1476 * threads at the same time. Even though the user is supposed to
1477 * ensure this doesn't happen, we play it safe as in (2) above.
1479 * Since the only other things that ever take the device lock such as block
1480 * pool resize only rarely happen, this will almost never be contended so
1481 * taking a lock isn't really an expensive operation in this case.
1483 pthread_mutex_lock(&device
->mutex
);
1485 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1486 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1487 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
,
1488 pSubmits
[i
].pCommandBuffers
[j
]);
1489 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1490 assert(!anv_batch_has_error(&cmd_buffer
->batch
));
1492 result
= anv_cmd_buffer_execbuf(device
, cmd_buffer
);
1493 if (result
!= VK_SUCCESS
)
1499 struct anv_bo
*fence_bo
= &fence
->bo
;
1500 result
= anv_device_execbuf(device
, &fence
->execbuf
, &fence_bo
);
1501 if (result
!= VK_SUCCESS
)
1504 /* Update the fence and wake up any waiters */
1505 assert(fence
->state
== ANV_FENCE_STATE_RESET
);
1506 fence
->state
= ANV_FENCE_STATE_SUBMITTED
;
1507 pthread_cond_broadcast(&device
->queue_submit
);
1511 if (result
!= VK_SUCCESS
) {
1512 /* In the case that something has gone wrong we may end up with an
1513 * inconsistent state from which it may not be trivial to recover.
1514 * For example, we might have computed address relocations and
1515 * any future attempt to re-submit this job will need to know about
1516 * this and avoid computing relocation addresses again.
1518 * To avoid this sort of issues, we assume that if something was
1519 * wrong during submission we must already be in a really bad situation
1520 * anyway (such us being out of memory) and return
1521 * VK_ERROR_DEVICE_LOST to ensure that clients do not attempt to
1522 * submit the same job again to this device.
1524 result
= VK_ERROR_DEVICE_LOST
;
1525 device
->lost
= true;
1527 /* If we return VK_ERROR_DEVICE LOST here, we need to ensure that
1528 * vkWaitForFences() and vkGetFenceStatus() return a valid result
1529 * (VK_SUCCESS or VK_ERROR_DEVICE_LOST) in a finite amount of time.
1530 * Setting the fence status to SIGNALED ensures this will happen in
1534 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1537 pthread_mutex_unlock(&device
->mutex
);
1542 VkResult
anv_QueueWaitIdle(
1545 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1547 return anv_DeviceWaitIdle(anv_device_to_handle(queue
->device
));
1550 VkResult
anv_DeviceWaitIdle(
1553 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1554 if (unlikely(device
->lost
))
1555 return VK_ERROR_DEVICE_LOST
;
1557 struct anv_batch batch
;
1560 batch
.start
= batch
.next
= cmds
;
1561 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1563 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1564 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1566 return anv_device_submit_simple_batch(device
, &batch
);
1570 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1572 uint32_t gem_handle
= anv_gem_create(device
, size
);
1574 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1576 anv_bo_init(bo
, gem_handle
, size
);
1578 if (device
->instance
->physicalDevice
.supports_48bit_addresses
)
1579 bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1581 if (device
->instance
->physicalDevice
.has_exec_async
)
1582 bo
->flags
|= EXEC_OBJECT_ASYNC
;
1587 VkResult
anv_AllocateMemory(
1589 const VkMemoryAllocateInfo
* pAllocateInfo
,
1590 const VkAllocationCallbacks
* pAllocator
,
1591 VkDeviceMemory
* pMem
)
1593 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1594 struct anv_device_memory
*mem
;
1597 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1599 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1600 assert(pAllocateInfo
->allocationSize
> 0);
1602 /* We support exactly one memory heap. */
1603 assert(pAllocateInfo
->memoryTypeIndex
== 0 ||
1604 (!device
->info
.has_llc
&& pAllocateInfo
->memoryTypeIndex
< 2));
1606 /* The kernel relocation API has a limitation of a 32-bit delta value
1607 * applied to the address before it is written which, in spite of it being
1608 * unsigned, is treated as signed . Because of the way that this maps to
1609 * the Vulkan API, we cannot handle an offset into a buffer that does not
1610 * fit into a signed 32 bits. The only mechanism we have for dealing with
1611 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1612 * of 2GB each. The Vulkan spec allows us to do this:
1614 * "Some platforms may have a limit on the maximum size of a single
1615 * allocation. For example, certain systems may fail to create
1616 * allocations with a size greater than or equal to 4GB. Such a limit is
1617 * implementation-dependent, and if such a failure occurs then the error
1618 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1620 * We don't use vk_error here because it's not an error so much as an
1621 * indication to the application that the allocation is too large.
1623 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1624 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1626 /* FINISHME: Fail if allocation request exceeds heap size. */
1628 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1629 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1631 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1633 /* The kernel is going to give us whole pages anyway */
1634 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
1636 result
= anv_bo_init_new(&mem
->bo
, device
, alloc_size
);
1637 if (result
!= VK_SUCCESS
)
1640 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1645 *pMem
= anv_device_memory_to_handle(mem
);
1650 vk_free2(&device
->alloc
, pAllocator
, mem
);
1655 void anv_FreeMemory(
1657 VkDeviceMemory _mem
,
1658 const VkAllocationCallbacks
* pAllocator
)
1660 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1661 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1667 anv_UnmapMemory(_device
, _mem
);
1670 anv_gem_munmap(mem
->bo
.map
, mem
->bo
.size
);
1672 if (mem
->bo
.gem_handle
!= 0)
1673 anv_gem_close(device
, mem
->bo
.gem_handle
);
1675 vk_free2(&device
->alloc
, pAllocator
, mem
);
1678 VkResult
anv_MapMemory(
1680 VkDeviceMemory _memory
,
1681 VkDeviceSize offset
,
1683 VkMemoryMapFlags flags
,
1686 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1687 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1694 if (size
== VK_WHOLE_SIZE
)
1695 size
= mem
->bo
.size
- offset
;
1697 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1699 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1700 * assert(size != 0);
1701 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1702 * equal to the size of the memory minus offset
1705 assert(offset
+ size
<= mem
->bo
.size
);
1707 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1708 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1709 * at a time is valid. We could just mmap up front and return an offset
1710 * pointer here, but that may exhaust virtual memory on 32 bit
1713 uint32_t gem_flags
= 0;
1714 if (!device
->info
.has_llc
&& mem
->type_index
== 0)
1715 gem_flags
|= I915_MMAP_WC
;
1717 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1718 uint64_t map_offset
= offset
& ~4095ull;
1719 assert(offset
>= map_offset
);
1720 uint64_t map_size
= (offset
+ size
) - map_offset
;
1722 /* Let's map whole pages */
1723 map_size
= align_u64(map_size
, 4096);
1725 void *map
= anv_gem_mmap(device
, mem
->bo
.gem_handle
,
1726 map_offset
, map_size
, gem_flags
);
1727 if (map
== MAP_FAILED
)
1728 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1731 mem
->map_size
= map_size
;
1733 *ppData
= mem
->map
+ (offset
- map_offset
);
1738 void anv_UnmapMemory(
1740 VkDeviceMemory _memory
)
1742 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1747 anv_gem_munmap(mem
->map
, mem
->map_size
);
1754 clflush_mapped_ranges(struct anv_device
*device
,
1756 const VkMappedMemoryRange
*ranges
)
1758 for (uint32_t i
= 0; i
< count
; i
++) {
1759 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1760 if (ranges
[i
].offset
>= mem
->map_size
)
1763 anv_clflush_range(mem
->map
+ ranges
[i
].offset
,
1764 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1768 VkResult
anv_FlushMappedMemoryRanges(
1770 uint32_t memoryRangeCount
,
1771 const VkMappedMemoryRange
* pMemoryRanges
)
1773 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1775 if (device
->info
.has_llc
)
1778 /* Make sure the writes we're flushing have landed. */
1779 __builtin_ia32_mfence();
1781 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1786 VkResult
anv_InvalidateMappedMemoryRanges(
1788 uint32_t memoryRangeCount
,
1789 const VkMappedMemoryRange
* pMemoryRanges
)
1791 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1793 if (device
->info
.has_llc
)
1796 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1798 /* Make sure no reads get moved up above the invalidate. */
1799 __builtin_ia32_mfence();
1804 void anv_GetBufferMemoryRequirements(
1807 VkMemoryRequirements
* pMemoryRequirements
)
1809 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1810 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1812 /* The Vulkan spec (git aaed022) says:
1814 * memoryTypeBits is a bitfield and contains one bit set for every
1815 * supported memory type for the resource. The bit `1<<i` is set if and
1816 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1817 * structure for the physical device is supported.
1819 * We support exactly one memory type on LLC, two on non-LLC.
1821 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1823 pMemoryRequirements
->size
= buffer
->size
;
1824 pMemoryRequirements
->alignment
= 16;
1827 void anv_GetImageMemoryRequirements(
1830 VkMemoryRequirements
* pMemoryRequirements
)
1832 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1833 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1835 /* The Vulkan spec (git aaed022) says:
1837 * memoryTypeBits is a bitfield and contains one bit set for every
1838 * supported memory type for the resource. The bit `1<<i` is set if and
1839 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1840 * structure for the physical device is supported.
1842 * We support exactly one memory type on LLC, two on non-LLC.
1844 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1846 pMemoryRequirements
->size
= image
->size
;
1847 pMemoryRequirements
->alignment
= image
->alignment
;
1850 void anv_GetImageSparseMemoryRequirements(
1853 uint32_t* pSparseMemoryRequirementCount
,
1854 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1856 *pSparseMemoryRequirementCount
= 0;
1859 void anv_GetDeviceMemoryCommitment(
1861 VkDeviceMemory memory
,
1862 VkDeviceSize
* pCommittedMemoryInBytes
)
1864 *pCommittedMemoryInBytes
= 0;
1867 VkResult
anv_BindBufferMemory(
1870 VkDeviceMemory _memory
,
1871 VkDeviceSize memoryOffset
)
1873 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1874 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1877 buffer
->bo
= &mem
->bo
;
1878 buffer
->offset
= memoryOffset
;
1887 VkResult
anv_QueueBindSparse(
1889 uint32_t bindInfoCount
,
1890 const VkBindSparseInfo
* pBindInfo
,
1893 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1894 if (unlikely(queue
->device
->lost
))
1895 return VK_ERROR_DEVICE_LOST
;
1897 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1900 VkResult
anv_CreateFence(
1902 const VkFenceCreateInfo
* pCreateInfo
,
1903 const VkAllocationCallbacks
* pAllocator
,
1906 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1907 struct anv_bo fence_bo
;
1908 struct anv_fence
*fence
;
1909 struct anv_batch batch
;
1912 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_CREATE_INFO
);
1914 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &fence_bo
, 4096);
1915 if (result
!= VK_SUCCESS
)
1918 /* Fences are small. Just store the CPU data structure in the BO. */
1919 fence
= fence_bo
.map
;
1920 fence
->bo
= fence_bo
;
1922 /* Place the batch after the CPU data but on its own cache line. */
1923 const uint32_t batch_offset
= align_u32(sizeof(*fence
), CACHELINE_SIZE
);
1924 batch
.next
= batch
.start
= fence
->bo
.map
+ batch_offset
;
1925 batch
.end
= fence
->bo
.map
+ fence
->bo
.size
;
1926 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1927 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1929 if (!device
->info
.has_llc
) {
1930 assert(((uintptr_t) batch
.start
& CACHELINE_MASK
) == 0);
1931 assert(batch
.next
- batch
.start
<= CACHELINE_SIZE
);
1932 __builtin_ia32_mfence();
1933 __builtin_ia32_clflush(batch
.start
);
1936 fence
->exec2_objects
[0].handle
= fence
->bo
.gem_handle
;
1937 fence
->exec2_objects
[0].relocation_count
= 0;
1938 fence
->exec2_objects
[0].relocs_ptr
= 0;
1939 fence
->exec2_objects
[0].alignment
= 0;
1940 fence
->exec2_objects
[0].offset
= fence
->bo
.offset
;
1941 fence
->exec2_objects
[0].flags
= 0;
1942 fence
->exec2_objects
[0].rsvd1
= 0;
1943 fence
->exec2_objects
[0].rsvd2
= 0;
1945 fence
->execbuf
.buffers_ptr
= (uintptr_t) fence
->exec2_objects
;
1946 fence
->execbuf
.buffer_count
= 1;
1947 fence
->execbuf
.batch_start_offset
= batch
.start
- fence
->bo
.map
;
1948 fence
->execbuf
.batch_len
= batch
.next
- batch
.start
;
1949 fence
->execbuf
.cliprects_ptr
= 0;
1950 fence
->execbuf
.num_cliprects
= 0;
1951 fence
->execbuf
.DR1
= 0;
1952 fence
->execbuf
.DR4
= 0;
1954 fence
->execbuf
.flags
=
1955 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
1956 fence
->execbuf
.rsvd1
= device
->context_id
;
1957 fence
->execbuf
.rsvd2
= 0;
1959 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
1960 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1962 fence
->state
= ANV_FENCE_STATE_RESET
;
1965 *pFence
= anv_fence_to_handle(fence
);
1970 void anv_DestroyFence(
1973 const VkAllocationCallbacks
* pAllocator
)
1975 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1976 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1981 assert(fence
->bo
.map
== fence
);
1982 anv_bo_pool_free(&device
->batch_bo_pool
, &fence
->bo
);
1985 VkResult
anv_ResetFences(
1987 uint32_t fenceCount
,
1988 const VkFence
* pFences
)
1990 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1991 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1992 fence
->state
= ANV_FENCE_STATE_RESET
;
1998 VkResult
anv_GetFenceStatus(
2002 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2003 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
2005 if (unlikely(device
->lost
))
2006 return VK_ERROR_DEVICE_LOST
;
2008 switch (fence
->state
) {
2009 case ANV_FENCE_STATE_RESET
:
2010 /* If it hasn't even been sent off to the GPU yet, it's not ready */
2011 return VK_NOT_READY
;
2013 case ANV_FENCE_STATE_SIGNALED
:
2014 /* It's been signaled, return success */
2017 case ANV_FENCE_STATE_SUBMITTED
: {
2018 VkResult result
= anv_device_bo_busy(device
, &fence
->bo
);
2019 if (result
== VK_SUCCESS
) {
2020 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
2027 unreachable("Invalid fence status");
2031 #define NSEC_PER_SEC 1000000000
2032 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
2034 VkResult
anv_WaitForFences(
2036 uint32_t fenceCount
,
2037 const VkFence
* pFences
,
2041 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2044 if (unlikely(device
->lost
))
2045 return VK_ERROR_DEVICE_LOST
;
2047 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
2048 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
2049 * for a couple of kernel releases. Since there's no way to know
2050 * whether or not the kernel we're using is one of the broken ones, the
2051 * best we can do is to clamp the timeout to INT64_MAX. This limits the
2052 * maximum timeout from 584 years to 292 years - likely not a big deal.
2054 int64_t timeout
= MIN2(_timeout
, INT64_MAX
);
2056 VkResult result
= VK_SUCCESS
;
2057 uint32_t pending_fences
= fenceCount
;
2058 while (pending_fences
) {
2060 bool signaled_fences
= false;
2061 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
2062 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
2063 switch (fence
->state
) {
2064 case ANV_FENCE_STATE_RESET
:
2065 /* This fence hasn't been submitted yet, we'll catch it the next
2066 * time around. Yes, this may mean we dead-loop but, short of
2067 * lots of locking and a condition variable, there's not much that
2068 * we can do about that.
2073 case ANV_FENCE_STATE_SIGNALED
:
2074 /* This fence is not pending. If waitAll isn't set, we can return
2075 * early. Otherwise, we have to keep going.
2078 result
= VK_SUCCESS
;
2083 case ANV_FENCE_STATE_SUBMITTED
:
2084 /* These are the fences we really care about. Go ahead and wait
2085 * on it until we hit a timeout.
2087 result
= anv_device_wait(device
, &fence
->bo
, timeout
);
2090 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
2091 signaled_fences
= true;
2105 if (pending_fences
&& !signaled_fences
) {
2106 /* If we've hit this then someone decided to vkWaitForFences before
2107 * they've actually submitted any of them to a queue. This is a
2108 * fairly pessimal case, so it's ok to lock here and use a standard
2109 * pthreads condition variable.
2111 pthread_mutex_lock(&device
->mutex
);
2113 /* It's possible that some of the fences have changed state since the
2114 * last time we checked. Now that we have the lock, check for
2115 * pending fences again and don't wait if it's changed.
2117 uint32_t now_pending_fences
= 0;
2118 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
2119 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
2120 if (fence
->state
== ANV_FENCE_STATE_RESET
)
2121 now_pending_fences
++;
2123 assert(now_pending_fences
<= pending_fences
);
2125 if (now_pending_fences
== pending_fences
) {
2126 struct timespec before
;
2127 clock_gettime(CLOCK_MONOTONIC
, &before
);
2129 uint32_t abs_nsec
= before
.tv_nsec
+ timeout
% NSEC_PER_SEC
;
2130 uint64_t abs_sec
= before
.tv_sec
+ (abs_nsec
/ NSEC_PER_SEC
) +
2131 (timeout
/ NSEC_PER_SEC
);
2132 abs_nsec
%= NSEC_PER_SEC
;
2134 /* Avoid roll-over in tv_sec on 32-bit systems if the user
2135 * provided timeout is UINT64_MAX
2137 struct timespec abstime
;
2138 abstime
.tv_nsec
= abs_nsec
;
2139 abstime
.tv_sec
= MIN2(abs_sec
, INT_TYPE_MAX(abstime
.tv_sec
));
2141 ret
= pthread_cond_timedwait(&device
->queue_submit
,
2142 &device
->mutex
, &abstime
);
2143 assert(ret
!= EINVAL
);
2145 struct timespec after
;
2146 clock_gettime(CLOCK_MONOTONIC
, &after
);
2147 uint64_t time_elapsed
=
2148 ((uint64_t)after
.tv_sec
* NSEC_PER_SEC
+ after
.tv_nsec
) -
2149 ((uint64_t)before
.tv_sec
* NSEC_PER_SEC
+ before
.tv_nsec
);
2151 if (time_elapsed
>= timeout
) {
2152 pthread_mutex_unlock(&device
->mutex
);
2153 result
= VK_TIMEOUT
;
2157 timeout
-= time_elapsed
;
2160 pthread_mutex_unlock(&device
->mutex
);
2165 if (unlikely(device
->lost
))
2166 return VK_ERROR_DEVICE_LOST
;
2171 // Queue semaphore functions
2173 VkResult
anv_CreateSemaphore(
2175 const VkSemaphoreCreateInfo
* pCreateInfo
,
2176 const VkAllocationCallbacks
* pAllocator
,
2177 VkSemaphore
* pSemaphore
)
2179 /* The DRM execbuffer ioctl always execute in-oder, even between different
2180 * rings. As such, there's nothing to do for the user space semaphore.
2183 *pSemaphore
= (VkSemaphore
)1;
2188 void anv_DestroySemaphore(
2190 VkSemaphore semaphore
,
2191 const VkAllocationCallbacks
* pAllocator
)
2197 VkResult
anv_CreateEvent(
2199 const VkEventCreateInfo
* pCreateInfo
,
2200 const VkAllocationCallbacks
* pAllocator
,
2203 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2204 struct anv_state state
;
2205 struct anv_event
*event
;
2207 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2209 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2212 event
->state
= state
;
2213 event
->semaphore
= VK_EVENT_RESET
;
2215 if (!device
->info
.has_llc
) {
2216 /* Make sure the writes we're flushing have landed. */
2217 __builtin_ia32_mfence();
2218 __builtin_ia32_clflush(event
);
2221 *pEvent
= anv_event_to_handle(event
);
2226 void anv_DestroyEvent(
2229 const VkAllocationCallbacks
* pAllocator
)
2231 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2232 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2237 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2240 VkResult
anv_GetEventStatus(
2244 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2245 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2247 if (unlikely(device
->lost
))
2248 return VK_ERROR_DEVICE_LOST
;
2250 if (!device
->info
.has_llc
) {
2251 /* Invalidate read cache before reading event written by GPU. */
2252 __builtin_ia32_clflush(event
);
2253 __builtin_ia32_mfence();
2257 return event
->semaphore
;
2260 VkResult
anv_SetEvent(
2264 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2265 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2267 event
->semaphore
= VK_EVENT_SET
;
2269 if (!device
->info
.has_llc
) {
2270 /* Make sure the writes we're flushing have landed. */
2271 __builtin_ia32_mfence();
2272 __builtin_ia32_clflush(event
);
2278 VkResult
anv_ResetEvent(
2282 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2283 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2285 event
->semaphore
= VK_EVENT_RESET
;
2287 if (!device
->info
.has_llc
) {
2288 /* Make sure the writes we're flushing have landed. */
2289 __builtin_ia32_mfence();
2290 __builtin_ia32_clflush(event
);
2298 VkResult
anv_CreateBuffer(
2300 const VkBufferCreateInfo
* pCreateInfo
,
2301 const VkAllocationCallbacks
* pAllocator
,
2304 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2305 struct anv_buffer
*buffer
;
2307 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2309 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2310 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2312 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2314 buffer
->size
= pCreateInfo
->size
;
2315 buffer
->usage
= pCreateInfo
->usage
;
2319 *pBuffer
= anv_buffer_to_handle(buffer
);
2324 void anv_DestroyBuffer(
2327 const VkAllocationCallbacks
* pAllocator
)
2329 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2330 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2335 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2339 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2340 enum isl_format format
,
2341 uint32_t offset
, uint32_t range
, uint32_t stride
)
2343 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2345 .mocs
= device
->default_mocs
,
2350 anv_state_flush(device
, state
);
2353 void anv_DestroySampler(
2356 const VkAllocationCallbacks
* pAllocator
)
2358 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2359 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2364 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2367 VkResult
anv_CreateFramebuffer(
2369 const VkFramebufferCreateInfo
* pCreateInfo
,
2370 const VkAllocationCallbacks
* pAllocator
,
2371 VkFramebuffer
* pFramebuffer
)
2373 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2374 struct anv_framebuffer
*framebuffer
;
2376 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2378 size_t size
= sizeof(*framebuffer
) +
2379 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2380 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2381 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2382 if (framebuffer
== NULL
)
2383 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2385 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2386 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2387 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2388 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2391 framebuffer
->width
= pCreateInfo
->width
;
2392 framebuffer
->height
= pCreateInfo
->height
;
2393 framebuffer
->layers
= pCreateInfo
->layers
;
2395 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2400 void anv_DestroyFramebuffer(
2403 const VkAllocationCallbacks
* pAllocator
)
2405 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2406 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2411 vk_free2(&device
->alloc
, pAllocator
, fb
);
2414 /* vk_icd.h does not declare this function, so we declare it here to
2415 * suppress Wmissing-prototypes.
2417 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2418 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2420 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2421 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2423 /* For the full details on loader interface versioning, see
2424 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2425 * What follows is a condensed summary, to help you navigate the large and
2426 * confusing official doc.
2428 * - Loader interface v0 is incompatible with later versions. We don't
2431 * - In loader interface v1:
2432 * - The first ICD entrypoint called by the loader is
2433 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2435 * - The ICD must statically expose no other Vulkan symbol unless it is
2436 * linked with -Bsymbolic.
2437 * - Each dispatchable Vulkan handle created by the ICD must be
2438 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2439 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2440 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2441 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2442 * such loader-managed surfaces.
2444 * - Loader interface v2 differs from v1 in:
2445 * - The first ICD entrypoint called by the loader is
2446 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2447 * statically expose this entrypoint.
2449 * - Loader interface v3 differs from v2 in:
2450 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2451 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2452 * because the loader no longer does so.
2454 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);