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
,
370 .extensionName
= VK_KHX_EXTERNAL_MEMORY_FD_EXTENSION_NAME
,
376 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
377 VkSystemAllocationScope allocationScope
)
383 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
384 size_t align
, VkSystemAllocationScope allocationScope
)
386 return realloc(pOriginal
, size
);
390 default_free_func(void *pUserData
, void *pMemory
)
395 static const VkAllocationCallbacks default_alloc
= {
397 .pfnAllocation
= default_alloc_func
,
398 .pfnReallocation
= default_realloc_func
,
399 .pfnFree
= default_free_func
,
402 VkResult
anv_CreateInstance(
403 const VkInstanceCreateInfo
* pCreateInfo
,
404 const VkAllocationCallbacks
* pAllocator
,
405 VkInstance
* pInstance
)
407 struct anv_instance
*instance
;
409 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
411 uint32_t client_version
;
412 if (pCreateInfo
->pApplicationInfo
&&
413 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
414 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
416 client_version
= VK_MAKE_VERSION(1, 0, 0);
419 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
420 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
421 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
422 "Client requested version %d.%d.%d",
423 VK_VERSION_MAJOR(client_version
),
424 VK_VERSION_MINOR(client_version
),
425 VK_VERSION_PATCH(client_version
));
428 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
430 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
431 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
432 global_extensions
[j
].extensionName
) == 0) {
438 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
441 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
442 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
444 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
446 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
449 instance
->alloc
= *pAllocator
;
451 instance
->alloc
= default_alloc
;
453 instance
->apiVersion
= client_version
;
454 instance
->physicalDeviceCount
= -1;
458 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
460 *pInstance
= anv_instance_to_handle(instance
);
465 void anv_DestroyInstance(
466 VkInstance _instance
,
467 const VkAllocationCallbacks
* pAllocator
)
469 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
474 if (instance
->physicalDeviceCount
> 0) {
475 /* We support at most one physical device. */
476 assert(instance
->physicalDeviceCount
== 1);
477 anv_physical_device_finish(&instance
->physicalDevice
);
480 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
484 vk_free(&instance
->alloc
, instance
);
488 anv_enumerate_devices(struct anv_instance
*instance
)
490 /* TODO: Check for more devices ? */
491 drmDevicePtr devices
[8];
492 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
495 instance
->physicalDeviceCount
= 0;
497 max_devices
= drmGetDevices2(0, devices
, sizeof(devices
));
499 return VK_ERROR_INCOMPATIBLE_DRIVER
;
501 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
502 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
503 devices
[i
]->bustype
== DRM_BUS_PCI
&&
504 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
506 result
= anv_physical_device_init(&instance
->physicalDevice
,
508 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
509 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
514 if (result
== VK_SUCCESS
)
515 instance
->physicalDeviceCount
= 1;
521 VkResult
anv_EnumeratePhysicalDevices(
522 VkInstance _instance
,
523 uint32_t* pPhysicalDeviceCount
,
524 VkPhysicalDevice
* pPhysicalDevices
)
526 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
527 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
530 if (instance
->physicalDeviceCount
< 0) {
531 result
= anv_enumerate_devices(instance
);
532 if (result
!= VK_SUCCESS
&&
533 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
537 if (instance
->physicalDeviceCount
> 0) {
538 assert(instance
->physicalDeviceCount
== 1);
539 vk_outarray_append(&out
, i
) {
540 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
544 return vk_outarray_status(&out
);
547 void anv_GetPhysicalDeviceFeatures(
548 VkPhysicalDevice physicalDevice
,
549 VkPhysicalDeviceFeatures
* pFeatures
)
551 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
553 *pFeatures
= (VkPhysicalDeviceFeatures
) {
554 .robustBufferAccess
= true,
555 .fullDrawIndexUint32
= true,
556 .imageCubeArray
= true,
557 .independentBlend
= true,
558 .geometryShader
= true,
559 .tessellationShader
= true,
560 .sampleRateShading
= true,
561 .dualSrcBlend
= true,
563 .multiDrawIndirect
= false,
564 .drawIndirectFirstInstance
= true,
566 .depthBiasClamp
= true,
567 .fillModeNonSolid
= true,
568 .depthBounds
= false,
572 .multiViewport
= true,
573 .samplerAnisotropy
= true,
574 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
575 pdevice
->info
.is_baytrail
,
576 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
577 .textureCompressionBC
= true,
578 .occlusionQueryPrecise
= true,
579 .pipelineStatisticsQuery
= true,
580 .fragmentStoresAndAtomics
= true,
581 .shaderTessellationAndGeometryPointSize
= true,
582 .shaderImageGatherExtended
= true,
583 .shaderStorageImageExtendedFormats
= true,
584 .shaderStorageImageMultisample
= false,
585 .shaderStorageImageReadWithoutFormat
= false,
586 .shaderStorageImageWriteWithoutFormat
= true,
587 .shaderUniformBufferArrayDynamicIndexing
= true,
588 .shaderSampledImageArrayDynamicIndexing
= true,
589 .shaderStorageBufferArrayDynamicIndexing
= true,
590 .shaderStorageImageArrayDynamicIndexing
= true,
591 .shaderClipDistance
= true,
592 .shaderCullDistance
= true,
593 .shaderFloat64
= pdevice
->info
.gen
>= 8,
594 .shaderInt64
= pdevice
->info
.gen
>= 8,
595 .shaderInt16
= false,
596 .shaderResourceMinLod
= false,
597 .variableMultisampleRate
= false,
598 .inheritedQueries
= true,
601 /* We can't do image stores in vec4 shaders */
602 pFeatures
->vertexPipelineStoresAndAtomics
=
603 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
604 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
607 void anv_GetPhysicalDeviceFeatures2KHR(
608 VkPhysicalDevice physicalDevice
,
609 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
611 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
613 vk_foreach_struct(ext
, pFeatures
->pNext
) {
614 switch (ext
->sType
) {
616 anv_debug_ignored_stype(ext
->sType
);
622 void anv_GetPhysicalDeviceProperties(
623 VkPhysicalDevice physicalDevice
,
624 VkPhysicalDeviceProperties
* pProperties
)
626 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
627 const struct gen_device_info
*devinfo
= &pdevice
->info
;
629 /* See assertions made when programming the buffer surface state. */
630 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
631 (1ul << 30) : (1ul << 27);
633 VkSampleCountFlags sample_counts
=
634 isl_device_get_sample_counts(&pdevice
->isl_dev
);
636 VkPhysicalDeviceLimits limits
= {
637 .maxImageDimension1D
= (1 << 14),
638 .maxImageDimension2D
= (1 << 14),
639 .maxImageDimension3D
= (1 << 11),
640 .maxImageDimensionCube
= (1 << 14),
641 .maxImageArrayLayers
= (1 << 11),
642 .maxTexelBufferElements
= 128 * 1024 * 1024,
643 .maxUniformBufferRange
= (1ul << 27),
644 .maxStorageBufferRange
= max_raw_buffer_sz
,
645 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
646 .maxMemoryAllocationCount
= UINT32_MAX
,
647 .maxSamplerAllocationCount
= 64 * 1024,
648 .bufferImageGranularity
= 64, /* A cache line */
649 .sparseAddressSpaceSize
= 0,
650 .maxBoundDescriptorSets
= MAX_SETS
,
651 .maxPerStageDescriptorSamplers
= 64,
652 .maxPerStageDescriptorUniformBuffers
= 64,
653 .maxPerStageDescriptorStorageBuffers
= 64,
654 .maxPerStageDescriptorSampledImages
= 64,
655 .maxPerStageDescriptorStorageImages
= 64,
656 .maxPerStageDescriptorInputAttachments
= 64,
657 .maxPerStageResources
= 128,
658 .maxDescriptorSetSamplers
= 256,
659 .maxDescriptorSetUniformBuffers
= 256,
660 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
661 .maxDescriptorSetStorageBuffers
= 256,
662 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
663 .maxDescriptorSetSampledImages
= 256,
664 .maxDescriptorSetStorageImages
= 256,
665 .maxDescriptorSetInputAttachments
= 256,
666 .maxVertexInputAttributes
= MAX_VBS
,
667 .maxVertexInputBindings
= MAX_VBS
,
668 .maxVertexInputAttributeOffset
= 2047,
669 .maxVertexInputBindingStride
= 2048,
670 .maxVertexOutputComponents
= 128,
671 .maxTessellationGenerationLevel
= 64,
672 .maxTessellationPatchSize
= 32,
673 .maxTessellationControlPerVertexInputComponents
= 128,
674 .maxTessellationControlPerVertexOutputComponents
= 128,
675 .maxTessellationControlPerPatchOutputComponents
= 128,
676 .maxTessellationControlTotalOutputComponents
= 2048,
677 .maxTessellationEvaluationInputComponents
= 128,
678 .maxTessellationEvaluationOutputComponents
= 128,
679 .maxGeometryShaderInvocations
= 32,
680 .maxGeometryInputComponents
= 64,
681 .maxGeometryOutputComponents
= 128,
682 .maxGeometryOutputVertices
= 256,
683 .maxGeometryTotalOutputComponents
= 1024,
684 .maxFragmentInputComponents
= 128,
685 .maxFragmentOutputAttachments
= 8,
686 .maxFragmentDualSrcAttachments
= 1,
687 .maxFragmentCombinedOutputResources
= 8,
688 .maxComputeSharedMemorySize
= 32768,
689 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
690 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
691 .maxComputeWorkGroupSize
= {
692 16 * devinfo
->max_cs_threads
,
693 16 * devinfo
->max_cs_threads
,
694 16 * devinfo
->max_cs_threads
,
696 .subPixelPrecisionBits
= 4 /* FIXME */,
697 .subTexelPrecisionBits
= 4 /* FIXME */,
698 .mipmapPrecisionBits
= 4 /* FIXME */,
699 .maxDrawIndexedIndexValue
= UINT32_MAX
,
700 .maxDrawIndirectCount
= UINT32_MAX
,
701 .maxSamplerLodBias
= 16,
702 .maxSamplerAnisotropy
= 16,
703 .maxViewports
= MAX_VIEWPORTS
,
704 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
705 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
706 .viewportSubPixelBits
= 13, /* We take a float? */
707 .minMemoryMapAlignment
= 4096, /* A page */
708 .minTexelBufferOffsetAlignment
= 1,
709 .minUniformBufferOffsetAlignment
= 16,
710 .minStorageBufferOffsetAlignment
= 4,
711 .minTexelOffset
= -8,
713 .minTexelGatherOffset
= -32,
714 .maxTexelGatherOffset
= 31,
715 .minInterpolationOffset
= -0.5,
716 .maxInterpolationOffset
= 0.4375,
717 .subPixelInterpolationOffsetBits
= 4,
718 .maxFramebufferWidth
= (1 << 14),
719 .maxFramebufferHeight
= (1 << 14),
720 .maxFramebufferLayers
= (1 << 11),
721 .framebufferColorSampleCounts
= sample_counts
,
722 .framebufferDepthSampleCounts
= sample_counts
,
723 .framebufferStencilSampleCounts
= sample_counts
,
724 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
725 .maxColorAttachments
= MAX_RTS
,
726 .sampledImageColorSampleCounts
= sample_counts
,
727 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
728 .sampledImageDepthSampleCounts
= sample_counts
,
729 .sampledImageStencilSampleCounts
= sample_counts
,
730 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
731 .maxSampleMaskWords
= 1,
732 .timestampComputeAndGraphics
= false,
733 .timestampPeriod
= devinfo
->timebase_scale
,
734 .maxClipDistances
= 8,
735 .maxCullDistances
= 8,
736 .maxCombinedClipAndCullDistances
= 8,
737 .discreteQueuePriorities
= 1,
738 .pointSizeRange
= { 0.125, 255.875 },
739 .lineWidthRange
= { 0.0, 7.9921875 },
740 .pointSizeGranularity
= (1.0 / 8.0),
741 .lineWidthGranularity
= (1.0 / 128.0),
742 .strictLines
= false, /* FINISHME */
743 .standardSampleLocations
= true,
744 .optimalBufferCopyOffsetAlignment
= 128,
745 .optimalBufferCopyRowPitchAlignment
= 128,
746 .nonCoherentAtomSize
= 64,
749 *pProperties
= (VkPhysicalDeviceProperties
) {
750 .apiVersion
= VK_MAKE_VERSION(1, 0, 42),
753 .deviceID
= pdevice
->chipset_id
,
754 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
756 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
759 strcpy(pProperties
->deviceName
, pdevice
->name
);
760 memcpy(pProperties
->pipelineCacheUUID
,
761 pdevice
->pipeline_cache_uuid
, VK_UUID_SIZE
);
764 void anv_GetPhysicalDeviceProperties2KHR(
765 VkPhysicalDevice physicalDevice
,
766 VkPhysicalDeviceProperties2KHR
* pProperties
)
768 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
770 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
772 vk_foreach_struct(ext
, pProperties
->pNext
) {
773 switch (ext
->sType
) {
774 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
775 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
776 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
778 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
782 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES_KHX
: {
783 VkPhysicalDeviceIDPropertiesKHX
*id_props
=
784 (VkPhysicalDeviceIDPropertiesKHX
*)ext
;
785 memcpy(id_props
->deviceUUID
, pdevice
->device_uuid
, VK_UUID_SIZE
);
786 memcpy(id_props
->driverUUID
, pdevice
->driver_uuid
, VK_UUID_SIZE
);
787 /* The LUID is for Windows. */
788 id_props
->deviceLUIDValid
= false;
793 anv_debug_ignored_stype(ext
->sType
);
799 /* We support exactly one queue family. */
800 static const VkQueueFamilyProperties
801 anv_queue_family_properties
= {
802 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
803 VK_QUEUE_COMPUTE_BIT
|
804 VK_QUEUE_TRANSFER_BIT
,
806 .timestampValidBits
= 36, /* XXX: Real value here */
807 .minImageTransferGranularity
= { 1, 1, 1 },
810 void anv_GetPhysicalDeviceQueueFamilyProperties(
811 VkPhysicalDevice physicalDevice
,
813 VkQueueFamilyProperties
* pQueueFamilyProperties
)
815 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
817 vk_outarray_append(&out
, p
) {
818 *p
= anv_queue_family_properties
;
822 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
823 VkPhysicalDevice physicalDevice
,
824 uint32_t* pQueueFamilyPropertyCount
,
825 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
828 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
830 vk_outarray_append(&out
, p
) {
831 p
->queueFamilyProperties
= anv_queue_family_properties
;
833 vk_foreach_struct(s
, p
->pNext
) {
834 anv_debug_ignored_stype(s
->sType
);
839 void anv_GetPhysicalDeviceMemoryProperties(
840 VkPhysicalDevice physicalDevice
,
841 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
843 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
845 if (physical_device
->info
.has_llc
) {
846 /* Big core GPUs share LLC with the CPU and thus one memory type can be
847 * both cached and coherent at the same time.
849 pMemoryProperties
->memoryTypeCount
= 1;
850 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
851 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
852 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
853 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
854 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
858 /* The spec requires that we expose a host-visible, coherent memory
859 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
860 * to give the application a choice between cached, but not coherent and
861 * coherent but uncached (WC though).
863 pMemoryProperties
->memoryTypeCount
= 2;
864 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
865 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
866 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
867 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
870 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
871 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
872 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
873 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
878 pMemoryProperties
->memoryHeapCount
= 1;
879 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
880 .size
= physical_device
->heap_size
,
881 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
885 void anv_GetPhysicalDeviceMemoryProperties2KHR(
886 VkPhysicalDevice physicalDevice
,
887 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
889 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
890 &pMemoryProperties
->memoryProperties
);
892 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
893 switch (ext
->sType
) {
895 anv_debug_ignored_stype(ext
->sType
);
901 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
905 return anv_lookup_entrypoint(NULL
, pName
);
908 /* With version 1+ of the loader interface the ICD should expose
909 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
912 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
917 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
921 return anv_GetInstanceProcAddr(instance
, pName
);
924 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
928 ANV_FROM_HANDLE(anv_device
, device
, _device
);
929 return anv_lookup_entrypoint(&device
->info
, pName
);
933 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
935 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
936 queue
->device
= device
;
937 queue
->pool
= &device
->surface_state_pool
;
941 anv_queue_finish(struct anv_queue
*queue
)
945 static struct anv_state
946 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
948 struct anv_state state
;
950 state
= anv_state_pool_alloc(pool
, size
, align
);
951 memcpy(state
.map
, p
, size
);
953 anv_state_flush(pool
->block_pool
->device
, state
);
958 struct gen8_border_color
{
963 /* Pad out to 64 bytes */
968 anv_device_init_border_colors(struct anv_device
*device
)
970 static const struct gen8_border_color border_colors
[] = {
971 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
972 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
973 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
974 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
975 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
976 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
979 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
980 sizeof(border_colors
), 64,
985 anv_device_submit_simple_batch(struct anv_device
*device
,
986 struct anv_batch
*batch
)
988 struct drm_i915_gem_execbuffer2 execbuf
;
989 struct drm_i915_gem_exec_object2 exec2_objects
[1];
990 struct anv_bo bo
, *exec_bos
[1];
991 VkResult result
= VK_SUCCESS
;
994 /* Kernel driver requires 8 byte aligned batch length */
995 size
= align_u32(batch
->next
- batch
->start
, 8);
996 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &bo
, size
);
997 if (result
!= VK_SUCCESS
)
1000 memcpy(bo
.map
, batch
->start
, size
);
1001 if (!device
->info
.has_llc
)
1002 anv_flush_range(bo
.map
, size
);
1005 exec2_objects
[0].handle
= bo
.gem_handle
;
1006 exec2_objects
[0].relocation_count
= 0;
1007 exec2_objects
[0].relocs_ptr
= 0;
1008 exec2_objects
[0].alignment
= 0;
1009 exec2_objects
[0].offset
= bo
.offset
;
1010 exec2_objects
[0].flags
= 0;
1011 exec2_objects
[0].rsvd1
= 0;
1012 exec2_objects
[0].rsvd2
= 0;
1014 execbuf
.buffers_ptr
= (uintptr_t) exec2_objects
;
1015 execbuf
.buffer_count
= 1;
1016 execbuf
.batch_start_offset
= 0;
1017 execbuf
.batch_len
= size
;
1018 execbuf
.cliprects_ptr
= 0;
1019 execbuf
.num_cliprects
= 0;
1024 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
1025 execbuf
.rsvd1
= device
->context_id
;
1028 result
= anv_device_execbuf(device
, &execbuf
, exec_bos
);
1029 if (result
!= VK_SUCCESS
)
1032 result
= anv_device_wait(device
, &bo
, INT64_MAX
);
1035 anv_bo_pool_free(&device
->batch_bo_pool
, &bo
);
1040 VkResult
anv_CreateDevice(
1041 VkPhysicalDevice physicalDevice
,
1042 const VkDeviceCreateInfo
* pCreateInfo
,
1043 const VkAllocationCallbacks
* pAllocator
,
1046 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
1048 struct anv_device
*device
;
1050 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
1052 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
1054 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
1055 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1056 device_extensions
[j
].extensionName
) == 0) {
1062 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1065 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1067 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1069 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1071 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1072 device
->instance
= physical_device
->instance
;
1073 device
->chipset_id
= physical_device
->chipset_id
;
1074 device
->lost
= false;
1077 device
->alloc
= *pAllocator
;
1079 device
->alloc
= physical_device
->instance
->alloc
;
1081 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1082 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1083 if (device
->fd
== -1) {
1084 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1088 device
->context_id
= anv_gem_create_context(device
);
1089 if (device
->context_id
== -1) {
1090 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1094 device
->info
= physical_device
->info
;
1095 device
->isl_dev
= physical_device
->isl_dev
;
1097 /* On Broadwell and later, we can use batch chaining to more efficiently
1098 * implement growing command buffers. Prior to Haswell, the kernel
1099 * command parser gets in the way and we have to fall back to growing
1102 device
->can_chain_batches
= device
->info
.gen
>= 8;
1104 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1105 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1107 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1108 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1109 goto fail_context_id
;
1112 pthread_condattr_t condattr
;
1113 if (pthread_condattr_init(&condattr
) != 0) {
1114 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1117 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1118 pthread_condattr_destroy(&condattr
);
1119 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1122 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1123 pthread_condattr_destroy(&condattr
);
1124 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1127 pthread_condattr_destroy(&condattr
);
1129 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
1131 result
= anv_bo_cache_init(&device
->bo_cache
);
1132 if (result
!= VK_SUCCESS
)
1133 goto fail_batch_bo_pool
;
1135 result
= anv_block_pool_init(&device
->dynamic_state_block_pool
, device
,
1137 if (result
!= VK_SUCCESS
)
1140 anv_state_pool_init(&device
->dynamic_state_pool
,
1141 &device
->dynamic_state_block_pool
);
1143 result
= anv_block_pool_init(&device
->instruction_block_pool
, device
,
1145 if (result
!= VK_SUCCESS
)
1146 goto fail_dynamic_state_pool
;
1148 anv_state_pool_init(&device
->instruction_state_pool
,
1149 &device
->instruction_block_pool
);
1151 result
= anv_block_pool_init(&device
->surface_state_block_pool
, device
,
1153 if (result
!= VK_SUCCESS
)
1154 goto fail_instruction_state_pool
;
1156 anv_state_pool_init(&device
->surface_state_pool
,
1157 &device
->surface_state_block_pool
);
1159 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1160 if (result
!= VK_SUCCESS
)
1161 goto fail_surface_state_pool
;
1163 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1165 anv_queue_init(device
, &device
->queue
);
1167 switch (device
->info
.gen
) {
1169 if (!device
->info
.is_haswell
)
1170 result
= gen7_init_device_state(device
);
1172 result
= gen75_init_device_state(device
);
1175 result
= gen8_init_device_state(device
);
1178 result
= gen9_init_device_state(device
);
1181 /* Shouldn't get here as we don't create physical devices for any other
1183 unreachable("unhandled gen");
1185 if (result
!= VK_SUCCESS
)
1186 goto fail_workaround_bo
;
1188 anv_device_init_blorp(device
);
1190 anv_device_init_border_colors(device
);
1192 *pDevice
= anv_device_to_handle(device
);
1197 anv_queue_finish(&device
->queue
);
1198 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1199 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1200 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1201 fail_surface_state_pool
:
1202 anv_state_pool_finish(&device
->surface_state_pool
);
1203 anv_block_pool_finish(&device
->surface_state_block_pool
);
1204 fail_instruction_state_pool
:
1205 anv_state_pool_finish(&device
->instruction_state_pool
);
1206 anv_block_pool_finish(&device
->instruction_block_pool
);
1207 fail_dynamic_state_pool
:
1208 anv_state_pool_finish(&device
->dynamic_state_pool
);
1209 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1211 anv_bo_cache_finish(&device
->bo_cache
);
1213 anv_bo_pool_finish(&device
->batch_bo_pool
);
1214 pthread_cond_destroy(&device
->queue_submit
);
1216 pthread_mutex_destroy(&device
->mutex
);
1218 anv_gem_destroy_context(device
, device
->context_id
);
1222 vk_free(&device
->alloc
, device
);
1227 void anv_DestroyDevice(
1229 const VkAllocationCallbacks
* pAllocator
)
1231 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1236 anv_device_finish_blorp(device
);
1238 anv_queue_finish(&device
->queue
);
1240 #ifdef HAVE_VALGRIND
1241 /* We only need to free these to prevent valgrind errors. The backing
1242 * BO will go away in a couple of lines so we don't actually leak.
1244 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1247 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1249 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1250 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1252 anv_state_pool_finish(&device
->surface_state_pool
);
1253 anv_block_pool_finish(&device
->surface_state_block_pool
);
1254 anv_state_pool_finish(&device
->instruction_state_pool
);
1255 anv_block_pool_finish(&device
->instruction_block_pool
);
1256 anv_state_pool_finish(&device
->dynamic_state_pool
);
1257 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1259 anv_bo_cache_finish(&device
->bo_cache
);
1261 anv_bo_pool_finish(&device
->batch_bo_pool
);
1263 pthread_cond_destroy(&device
->queue_submit
);
1264 pthread_mutex_destroy(&device
->mutex
);
1266 anv_gem_destroy_context(device
, device
->context_id
);
1270 vk_free(&device
->alloc
, device
);
1273 VkResult
anv_EnumerateInstanceExtensionProperties(
1274 const char* pLayerName
,
1275 uint32_t* pPropertyCount
,
1276 VkExtensionProperties
* pProperties
)
1278 if (pProperties
== NULL
) {
1279 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1283 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(global_extensions
));
1284 typed_memcpy(pProperties
, global_extensions
, *pPropertyCount
);
1286 if (*pPropertyCount
< ARRAY_SIZE(global_extensions
))
1287 return VK_INCOMPLETE
;
1292 VkResult
anv_EnumerateDeviceExtensionProperties(
1293 VkPhysicalDevice physicalDevice
,
1294 const char* pLayerName
,
1295 uint32_t* pPropertyCount
,
1296 VkExtensionProperties
* pProperties
)
1298 if (pProperties
== NULL
) {
1299 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1303 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(device_extensions
));
1304 typed_memcpy(pProperties
, device_extensions
, *pPropertyCount
);
1306 if (*pPropertyCount
< ARRAY_SIZE(device_extensions
))
1307 return VK_INCOMPLETE
;
1312 VkResult
anv_EnumerateInstanceLayerProperties(
1313 uint32_t* pPropertyCount
,
1314 VkLayerProperties
* pProperties
)
1316 if (pProperties
== NULL
) {
1317 *pPropertyCount
= 0;
1321 /* None supported at this time */
1322 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1325 VkResult
anv_EnumerateDeviceLayerProperties(
1326 VkPhysicalDevice physicalDevice
,
1327 uint32_t* pPropertyCount
,
1328 VkLayerProperties
* pProperties
)
1330 if (pProperties
== NULL
) {
1331 *pPropertyCount
= 0;
1335 /* None supported at this time */
1336 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1339 void anv_GetDeviceQueue(
1341 uint32_t queueNodeIndex
,
1342 uint32_t queueIndex
,
1345 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1347 assert(queueIndex
== 0);
1349 *pQueue
= anv_queue_to_handle(&device
->queue
);
1353 anv_device_execbuf(struct anv_device
*device
,
1354 struct drm_i915_gem_execbuffer2
*execbuf
,
1355 struct anv_bo
**execbuf_bos
)
1357 int ret
= anv_gem_execbuffer(device
, execbuf
);
1359 /* We don't know the real error. */
1360 device
->lost
= true;
1361 return vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
1364 struct drm_i915_gem_exec_object2
*objects
=
1365 (void *)(uintptr_t)execbuf
->buffers_ptr
;
1366 for (uint32_t k
= 0; k
< execbuf
->buffer_count
; k
++)
1367 execbuf_bos
[k
]->offset
= objects
[k
].offset
;
1373 anv_device_query_status(struct anv_device
*device
)
1375 /* This isn't likely as most of the callers of this function already check
1376 * for it. However, it doesn't hurt to check and it potentially lets us
1379 if (unlikely(device
->lost
))
1380 return VK_ERROR_DEVICE_LOST
;
1382 uint32_t active
, pending
;
1383 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1385 /* We don't know the real error. */
1386 device
->lost
= true;
1387 return vk_errorf(VK_ERROR_DEVICE_LOST
, "get_reset_stats failed: %m");
1391 device
->lost
= true;
1392 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1393 "GPU hung on one of our command buffers");
1394 } else if (pending
) {
1395 device
->lost
= true;
1396 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1397 "GPU hung with commands in-flight");
1404 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1406 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1407 * Other usages of the BO (such as on different hardware) will not be
1408 * flagged as "busy" by this ioctl. Use with care.
1410 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1412 return VK_NOT_READY
;
1413 } else if (ret
== -1) {
1414 /* We don't know the real error. */
1415 device
->lost
= true;
1416 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1419 /* Query for device status after the busy call. If the BO we're checking
1420 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1421 * client because it clearly doesn't have valid data. Yes, this most
1422 * likely means an ioctl, but we just did an ioctl to query the busy status
1423 * so it's no great loss.
1425 return anv_device_query_status(device
);
1429 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1432 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1433 if (ret
== -1 && errno
== ETIME
) {
1435 } else if (ret
== -1) {
1436 /* We don't know the real error. */
1437 device
->lost
= true;
1438 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1441 /* Query for device status after the wait. If the BO we're waiting on got
1442 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1443 * because it clearly doesn't have valid data. Yes, this most likely means
1444 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1446 return anv_device_query_status(device
);
1449 VkResult
anv_QueueSubmit(
1451 uint32_t submitCount
,
1452 const VkSubmitInfo
* pSubmits
,
1455 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1456 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1457 struct anv_device
*device
= queue
->device
;
1459 /* Query for device status prior to submitting. Technically, we don't need
1460 * to do this. However, if we have a client that's submitting piles of
1461 * garbage, we would rather break as early as possible to keep the GPU
1462 * hanging contained. If we don't check here, we'll either be waiting for
1463 * the kernel to kick us or we'll have to wait until the client waits on a
1464 * fence before we actually know whether or not we've hung.
1466 VkResult result
= anv_device_query_status(device
);
1467 if (result
!= VK_SUCCESS
)
1470 /* We lock around QueueSubmit for three main reasons:
1472 * 1) When a block pool is resized, we create a new gem handle with a
1473 * different size and, in the case of surface states, possibly a
1474 * different center offset but we re-use the same anv_bo struct when
1475 * we do so. If this happens in the middle of setting up an execbuf,
1476 * we could end up with our list of BOs out of sync with our list of
1479 * 2) The algorithm we use for building the list of unique buffers isn't
1480 * thread-safe. While the client is supposed to syncronize around
1481 * QueueSubmit, this would be extremely difficult to debug if it ever
1482 * came up in the wild due to a broken app. It's better to play it
1483 * safe and just lock around QueueSubmit.
1485 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
1486 * userspace. Due to the fact that the surface state buffer is shared
1487 * between batches, we can't afford to have that happen from multiple
1488 * threads at the same time. Even though the user is supposed to
1489 * ensure this doesn't happen, we play it safe as in (2) above.
1491 * Since the only other things that ever take the device lock such as block
1492 * pool resize only rarely happen, this will almost never be contended so
1493 * taking a lock isn't really an expensive operation in this case.
1495 pthread_mutex_lock(&device
->mutex
);
1497 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1498 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1499 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
,
1500 pSubmits
[i
].pCommandBuffers
[j
]);
1501 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1502 assert(!anv_batch_has_error(&cmd_buffer
->batch
));
1504 result
= anv_cmd_buffer_execbuf(device
, cmd_buffer
);
1505 if (result
!= VK_SUCCESS
)
1511 struct anv_bo
*fence_bo
= &fence
->bo
;
1512 result
= anv_device_execbuf(device
, &fence
->execbuf
, &fence_bo
);
1513 if (result
!= VK_SUCCESS
)
1516 /* Update the fence and wake up any waiters */
1517 assert(fence
->state
== ANV_FENCE_STATE_RESET
);
1518 fence
->state
= ANV_FENCE_STATE_SUBMITTED
;
1519 pthread_cond_broadcast(&device
->queue_submit
);
1523 if (result
!= VK_SUCCESS
) {
1524 /* In the case that something has gone wrong we may end up with an
1525 * inconsistent state from which it may not be trivial to recover.
1526 * For example, we might have computed address relocations and
1527 * any future attempt to re-submit this job will need to know about
1528 * this and avoid computing relocation addresses again.
1530 * To avoid this sort of issues, we assume that if something was
1531 * wrong during submission we must already be in a really bad situation
1532 * anyway (such us being out of memory) and return
1533 * VK_ERROR_DEVICE_LOST to ensure that clients do not attempt to
1534 * submit the same job again to this device.
1536 result
= VK_ERROR_DEVICE_LOST
;
1537 device
->lost
= true;
1539 /* If we return VK_ERROR_DEVICE LOST here, we need to ensure that
1540 * vkWaitForFences() and vkGetFenceStatus() return a valid result
1541 * (VK_SUCCESS or VK_ERROR_DEVICE_LOST) in a finite amount of time.
1542 * Setting the fence status to SIGNALED ensures this will happen in
1546 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1549 pthread_mutex_unlock(&device
->mutex
);
1554 VkResult
anv_QueueWaitIdle(
1557 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1559 return anv_DeviceWaitIdle(anv_device_to_handle(queue
->device
));
1562 VkResult
anv_DeviceWaitIdle(
1565 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1566 if (unlikely(device
->lost
))
1567 return VK_ERROR_DEVICE_LOST
;
1569 struct anv_batch batch
;
1572 batch
.start
= batch
.next
= cmds
;
1573 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1575 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1576 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1578 return anv_device_submit_simple_batch(device
, &batch
);
1582 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1584 uint32_t gem_handle
= anv_gem_create(device
, size
);
1586 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1588 anv_bo_init(bo
, gem_handle
, size
);
1590 if (device
->instance
->physicalDevice
.supports_48bit_addresses
)
1591 bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1593 if (device
->instance
->physicalDevice
.has_exec_async
)
1594 bo
->flags
|= EXEC_OBJECT_ASYNC
;
1599 VkResult
anv_AllocateMemory(
1601 const VkMemoryAllocateInfo
* pAllocateInfo
,
1602 const VkAllocationCallbacks
* pAllocator
,
1603 VkDeviceMemory
* pMem
)
1605 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1606 struct anv_device_memory
*mem
;
1607 VkResult result
= VK_SUCCESS
;
1609 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1611 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1612 assert(pAllocateInfo
->allocationSize
> 0);
1614 /* We support exactly one memory heap. */
1615 assert(pAllocateInfo
->memoryTypeIndex
== 0 ||
1616 (!device
->info
.has_llc
&& pAllocateInfo
->memoryTypeIndex
< 2));
1618 /* The kernel relocation API has a limitation of a 32-bit delta value
1619 * applied to the address before it is written which, in spite of it being
1620 * unsigned, is treated as signed . Because of the way that this maps to
1621 * the Vulkan API, we cannot handle an offset into a buffer that does not
1622 * fit into a signed 32 bits. The only mechanism we have for dealing with
1623 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1624 * of 2GB each. The Vulkan spec allows us to do this:
1626 * "Some platforms may have a limit on the maximum size of a single
1627 * allocation. For example, certain systems may fail to create
1628 * allocations with a size greater than or equal to 4GB. Such a limit is
1629 * implementation-dependent, and if such a failure occurs then the error
1630 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1632 * We don't use vk_error here because it's not an error so much as an
1633 * indication to the application that the allocation is too large.
1635 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1636 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1638 /* FINISHME: Fail if allocation request exceeds heap size. */
1640 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1641 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1643 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1645 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1649 const VkImportMemoryFdInfoKHX
*fd_info
=
1650 vk_find_struct_const(pAllocateInfo
->pNext
, IMPORT_MEMORY_FD_INFO_KHX
);
1652 /* The Vulkan spec permits handleType to be 0, in which case the struct is
1655 if (fd_info
&& fd_info
->handleType
) {
1656 /* At the moment, we only support the OPAQUE_FD memory type which is
1657 * just a GEM buffer.
1659 assert(fd_info
->handleType
==
1660 VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHX
);
1662 result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1663 fd_info
->fd
, pAllocateInfo
->allocationSize
,
1665 if (result
!= VK_SUCCESS
)
1668 result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
1669 pAllocateInfo
->allocationSize
,
1671 if (result
!= VK_SUCCESS
)
1675 *pMem
= anv_device_memory_to_handle(mem
);
1680 vk_free2(&device
->alloc
, pAllocator
, mem
);
1685 VkResult
anv_GetMemoryFdKHX(
1687 VkDeviceMemory memory_h
,
1688 VkExternalMemoryHandleTypeFlagBitsKHX handleType
,
1691 ANV_FROM_HANDLE(anv_device
, dev
, device_h
);
1692 ANV_FROM_HANDLE(anv_device_memory
, mem
, memory_h
);
1694 /* We support only one handle type. */
1695 assert(handleType
== VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHX
);
1697 return anv_bo_cache_export(dev
, &dev
->bo_cache
, mem
->bo
, pFd
);
1700 VkResult
anv_GetMemoryFdPropertiesKHX(
1702 VkExternalMemoryHandleTypeFlagBitsKHX handleType
,
1704 VkMemoryFdPropertiesKHX
* pMemoryFdProperties
)
1706 /* The valid usage section for this function says:
1708 * "handleType must not be one of the handle types defined as opaque."
1710 * Since we only handle opaque handles for now, there are no FD properties.
1712 return VK_ERROR_INVALID_EXTERNAL_HANDLE_KHX
;
1715 void anv_FreeMemory(
1717 VkDeviceMemory _mem
,
1718 const VkAllocationCallbacks
* pAllocator
)
1720 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1721 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1727 anv_UnmapMemory(_device
, _mem
);
1729 anv_bo_cache_release(device
, &device
->bo_cache
, mem
->bo
);
1731 vk_free2(&device
->alloc
, pAllocator
, mem
);
1734 VkResult
anv_MapMemory(
1736 VkDeviceMemory _memory
,
1737 VkDeviceSize offset
,
1739 VkMemoryMapFlags flags
,
1742 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1743 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1750 if (size
== VK_WHOLE_SIZE
)
1751 size
= mem
->bo
->size
- offset
;
1753 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1755 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1756 * assert(size != 0);
1757 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1758 * equal to the size of the memory minus offset
1761 assert(offset
+ size
<= mem
->bo
->size
);
1763 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1764 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1765 * at a time is valid. We could just mmap up front and return an offset
1766 * pointer here, but that may exhaust virtual memory on 32 bit
1769 uint32_t gem_flags
= 0;
1770 if (!device
->info
.has_llc
&& mem
->type_index
== 0)
1771 gem_flags
|= I915_MMAP_WC
;
1773 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1774 uint64_t map_offset
= offset
& ~4095ull;
1775 assert(offset
>= map_offset
);
1776 uint64_t map_size
= (offset
+ size
) - map_offset
;
1778 /* Let's map whole pages */
1779 map_size
= align_u64(map_size
, 4096);
1781 void *map
= anv_gem_mmap(device
, mem
->bo
->gem_handle
,
1782 map_offset
, map_size
, gem_flags
);
1783 if (map
== MAP_FAILED
)
1784 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1787 mem
->map_size
= map_size
;
1789 *ppData
= mem
->map
+ (offset
- map_offset
);
1794 void anv_UnmapMemory(
1796 VkDeviceMemory _memory
)
1798 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1803 anv_gem_munmap(mem
->map
, mem
->map_size
);
1810 clflush_mapped_ranges(struct anv_device
*device
,
1812 const VkMappedMemoryRange
*ranges
)
1814 for (uint32_t i
= 0; i
< count
; i
++) {
1815 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1816 if (ranges
[i
].offset
>= mem
->map_size
)
1819 anv_clflush_range(mem
->map
+ ranges
[i
].offset
,
1820 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1824 VkResult
anv_FlushMappedMemoryRanges(
1826 uint32_t memoryRangeCount
,
1827 const VkMappedMemoryRange
* pMemoryRanges
)
1829 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1831 if (device
->info
.has_llc
)
1834 /* Make sure the writes we're flushing have landed. */
1835 __builtin_ia32_mfence();
1837 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1842 VkResult
anv_InvalidateMappedMemoryRanges(
1844 uint32_t memoryRangeCount
,
1845 const VkMappedMemoryRange
* pMemoryRanges
)
1847 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1849 if (device
->info
.has_llc
)
1852 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1854 /* Make sure no reads get moved up above the invalidate. */
1855 __builtin_ia32_mfence();
1860 void anv_GetBufferMemoryRequirements(
1863 VkMemoryRequirements
* pMemoryRequirements
)
1865 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1866 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1868 /* The Vulkan spec (git aaed022) says:
1870 * memoryTypeBits is a bitfield and contains one bit set for every
1871 * supported memory type for the resource. The bit `1<<i` is set if and
1872 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1873 * structure for the physical device is supported.
1875 * We support exactly one memory type on LLC, two on non-LLC.
1877 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1879 pMemoryRequirements
->size
= buffer
->size
;
1880 pMemoryRequirements
->alignment
= 16;
1883 void anv_GetImageMemoryRequirements(
1886 VkMemoryRequirements
* pMemoryRequirements
)
1888 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1889 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1891 /* The Vulkan spec (git aaed022) says:
1893 * memoryTypeBits is a bitfield and contains one bit set for every
1894 * supported memory type for the resource. The bit `1<<i` is set if and
1895 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1896 * structure for the physical device is supported.
1898 * We support exactly one memory type on LLC, two on non-LLC.
1900 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1902 pMemoryRequirements
->size
= image
->size
;
1903 pMemoryRequirements
->alignment
= image
->alignment
;
1906 void anv_GetImageSparseMemoryRequirements(
1909 uint32_t* pSparseMemoryRequirementCount
,
1910 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1912 *pSparseMemoryRequirementCount
= 0;
1915 void anv_GetDeviceMemoryCommitment(
1917 VkDeviceMemory memory
,
1918 VkDeviceSize
* pCommittedMemoryInBytes
)
1920 *pCommittedMemoryInBytes
= 0;
1923 VkResult
anv_BindBufferMemory(
1926 VkDeviceMemory _memory
,
1927 VkDeviceSize memoryOffset
)
1929 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1930 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1933 buffer
->bo
= mem
->bo
;
1934 buffer
->offset
= memoryOffset
;
1943 VkResult
anv_QueueBindSparse(
1945 uint32_t bindInfoCount
,
1946 const VkBindSparseInfo
* pBindInfo
,
1949 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1950 if (unlikely(queue
->device
->lost
))
1951 return VK_ERROR_DEVICE_LOST
;
1953 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1956 VkResult
anv_CreateFence(
1958 const VkFenceCreateInfo
* pCreateInfo
,
1959 const VkAllocationCallbacks
* pAllocator
,
1962 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1963 struct anv_bo fence_bo
;
1964 struct anv_fence
*fence
;
1965 struct anv_batch batch
;
1968 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_CREATE_INFO
);
1970 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &fence_bo
, 4096);
1971 if (result
!= VK_SUCCESS
)
1974 /* Fences are small. Just store the CPU data structure in the BO. */
1975 fence
= fence_bo
.map
;
1976 fence
->bo
= fence_bo
;
1978 /* Place the batch after the CPU data but on its own cache line. */
1979 const uint32_t batch_offset
= align_u32(sizeof(*fence
), CACHELINE_SIZE
);
1980 batch
.next
= batch
.start
= fence
->bo
.map
+ batch_offset
;
1981 batch
.end
= fence
->bo
.map
+ fence
->bo
.size
;
1982 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1983 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1985 if (!device
->info
.has_llc
) {
1986 assert(((uintptr_t) batch
.start
& CACHELINE_MASK
) == 0);
1987 assert(batch
.next
- batch
.start
<= CACHELINE_SIZE
);
1988 __builtin_ia32_mfence();
1989 __builtin_ia32_clflush(batch
.start
);
1992 fence
->exec2_objects
[0].handle
= fence
->bo
.gem_handle
;
1993 fence
->exec2_objects
[0].relocation_count
= 0;
1994 fence
->exec2_objects
[0].relocs_ptr
= 0;
1995 fence
->exec2_objects
[0].alignment
= 0;
1996 fence
->exec2_objects
[0].offset
= fence
->bo
.offset
;
1997 fence
->exec2_objects
[0].flags
= 0;
1998 fence
->exec2_objects
[0].rsvd1
= 0;
1999 fence
->exec2_objects
[0].rsvd2
= 0;
2001 fence
->execbuf
.buffers_ptr
= (uintptr_t) fence
->exec2_objects
;
2002 fence
->execbuf
.buffer_count
= 1;
2003 fence
->execbuf
.batch_start_offset
= batch
.start
- fence
->bo
.map
;
2004 fence
->execbuf
.batch_len
= batch
.next
- batch
.start
;
2005 fence
->execbuf
.cliprects_ptr
= 0;
2006 fence
->execbuf
.num_cliprects
= 0;
2007 fence
->execbuf
.DR1
= 0;
2008 fence
->execbuf
.DR4
= 0;
2010 fence
->execbuf
.flags
=
2011 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
2012 fence
->execbuf
.rsvd1
= device
->context_id
;
2013 fence
->execbuf
.rsvd2
= 0;
2015 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
2016 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
2018 fence
->state
= ANV_FENCE_STATE_RESET
;
2021 *pFence
= anv_fence_to_handle(fence
);
2026 void anv_DestroyFence(
2029 const VkAllocationCallbacks
* pAllocator
)
2031 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2032 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
2037 assert(fence
->bo
.map
== fence
);
2038 anv_bo_pool_free(&device
->batch_bo_pool
, &fence
->bo
);
2041 VkResult
anv_ResetFences(
2043 uint32_t fenceCount
,
2044 const VkFence
* pFences
)
2046 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
2047 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
2048 fence
->state
= ANV_FENCE_STATE_RESET
;
2054 VkResult
anv_GetFenceStatus(
2058 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2059 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
2061 if (unlikely(device
->lost
))
2062 return VK_ERROR_DEVICE_LOST
;
2064 switch (fence
->state
) {
2065 case ANV_FENCE_STATE_RESET
:
2066 /* If it hasn't even been sent off to the GPU yet, it's not ready */
2067 return VK_NOT_READY
;
2069 case ANV_FENCE_STATE_SIGNALED
:
2070 /* It's been signaled, return success */
2073 case ANV_FENCE_STATE_SUBMITTED
: {
2074 VkResult result
= anv_device_bo_busy(device
, &fence
->bo
);
2075 if (result
== VK_SUCCESS
) {
2076 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
2083 unreachable("Invalid fence status");
2087 #define NSEC_PER_SEC 1000000000
2088 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
2090 VkResult
anv_WaitForFences(
2092 uint32_t fenceCount
,
2093 const VkFence
* pFences
,
2097 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2100 if (unlikely(device
->lost
))
2101 return VK_ERROR_DEVICE_LOST
;
2103 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
2104 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
2105 * for a couple of kernel releases. Since there's no way to know
2106 * whether or not the kernel we're using is one of the broken ones, the
2107 * best we can do is to clamp the timeout to INT64_MAX. This limits the
2108 * maximum timeout from 584 years to 292 years - likely not a big deal.
2110 int64_t timeout
= MIN2(_timeout
, INT64_MAX
);
2112 VkResult result
= VK_SUCCESS
;
2113 uint32_t pending_fences
= fenceCount
;
2114 while (pending_fences
) {
2116 bool signaled_fences
= false;
2117 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
2118 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
2119 switch (fence
->state
) {
2120 case ANV_FENCE_STATE_RESET
:
2121 /* This fence hasn't been submitted yet, we'll catch it the next
2122 * time around. Yes, this may mean we dead-loop but, short of
2123 * lots of locking and a condition variable, there's not much that
2124 * we can do about that.
2129 case ANV_FENCE_STATE_SIGNALED
:
2130 /* This fence is not pending. If waitAll isn't set, we can return
2131 * early. Otherwise, we have to keep going.
2134 result
= VK_SUCCESS
;
2139 case ANV_FENCE_STATE_SUBMITTED
:
2140 /* These are the fences we really care about. Go ahead and wait
2141 * on it until we hit a timeout.
2143 result
= anv_device_wait(device
, &fence
->bo
, timeout
);
2146 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
2147 signaled_fences
= true;
2161 if (pending_fences
&& !signaled_fences
) {
2162 /* If we've hit this then someone decided to vkWaitForFences before
2163 * they've actually submitted any of them to a queue. This is a
2164 * fairly pessimal case, so it's ok to lock here and use a standard
2165 * pthreads condition variable.
2167 pthread_mutex_lock(&device
->mutex
);
2169 /* It's possible that some of the fences have changed state since the
2170 * last time we checked. Now that we have the lock, check for
2171 * pending fences again and don't wait if it's changed.
2173 uint32_t now_pending_fences
= 0;
2174 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
2175 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
2176 if (fence
->state
== ANV_FENCE_STATE_RESET
)
2177 now_pending_fences
++;
2179 assert(now_pending_fences
<= pending_fences
);
2181 if (now_pending_fences
== pending_fences
) {
2182 struct timespec before
;
2183 clock_gettime(CLOCK_MONOTONIC
, &before
);
2185 uint32_t abs_nsec
= before
.tv_nsec
+ timeout
% NSEC_PER_SEC
;
2186 uint64_t abs_sec
= before
.tv_sec
+ (abs_nsec
/ NSEC_PER_SEC
) +
2187 (timeout
/ NSEC_PER_SEC
);
2188 abs_nsec
%= NSEC_PER_SEC
;
2190 /* Avoid roll-over in tv_sec on 32-bit systems if the user
2191 * provided timeout is UINT64_MAX
2193 struct timespec abstime
;
2194 abstime
.tv_nsec
= abs_nsec
;
2195 abstime
.tv_sec
= MIN2(abs_sec
, INT_TYPE_MAX(abstime
.tv_sec
));
2197 ret
= pthread_cond_timedwait(&device
->queue_submit
,
2198 &device
->mutex
, &abstime
);
2199 assert(ret
!= EINVAL
);
2201 struct timespec after
;
2202 clock_gettime(CLOCK_MONOTONIC
, &after
);
2203 uint64_t time_elapsed
=
2204 ((uint64_t)after
.tv_sec
* NSEC_PER_SEC
+ after
.tv_nsec
) -
2205 ((uint64_t)before
.tv_sec
* NSEC_PER_SEC
+ before
.tv_nsec
);
2207 if (time_elapsed
>= timeout
) {
2208 pthread_mutex_unlock(&device
->mutex
);
2209 result
= VK_TIMEOUT
;
2213 timeout
-= time_elapsed
;
2216 pthread_mutex_unlock(&device
->mutex
);
2221 if (unlikely(device
->lost
))
2222 return VK_ERROR_DEVICE_LOST
;
2227 // Queue semaphore functions
2229 VkResult
anv_CreateSemaphore(
2231 const VkSemaphoreCreateInfo
* pCreateInfo
,
2232 const VkAllocationCallbacks
* pAllocator
,
2233 VkSemaphore
* pSemaphore
)
2235 /* The DRM execbuffer ioctl always execute in-oder, even between different
2236 * rings. As such, there's nothing to do for the user space semaphore.
2239 *pSemaphore
= (VkSemaphore
)1;
2244 void anv_DestroySemaphore(
2246 VkSemaphore semaphore
,
2247 const VkAllocationCallbacks
* pAllocator
)
2253 VkResult
anv_CreateEvent(
2255 const VkEventCreateInfo
* pCreateInfo
,
2256 const VkAllocationCallbacks
* pAllocator
,
2259 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2260 struct anv_state state
;
2261 struct anv_event
*event
;
2263 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2265 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2268 event
->state
= state
;
2269 event
->semaphore
= VK_EVENT_RESET
;
2271 if (!device
->info
.has_llc
) {
2272 /* Make sure the writes we're flushing have landed. */
2273 __builtin_ia32_mfence();
2274 __builtin_ia32_clflush(event
);
2277 *pEvent
= anv_event_to_handle(event
);
2282 void anv_DestroyEvent(
2285 const VkAllocationCallbacks
* pAllocator
)
2287 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2288 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2293 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2296 VkResult
anv_GetEventStatus(
2300 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2301 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2303 if (unlikely(device
->lost
))
2304 return VK_ERROR_DEVICE_LOST
;
2306 if (!device
->info
.has_llc
) {
2307 /* Invalidate read cache before reading event written by GPU. */
2308 __builtin_ia32_clflush(event
);
2309 __builtin_ia32_mfence();
2313 return event
->semaphore
;
2316 VkResult
anv_SetEvent(
2320 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2321 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2323 event
->semaphore
= VK_EVENT_SET
;
2325 if (!device
->info
.has_llc
) {
2326 /* Make sure the writes we're flushing have landed. */
2327 __builtin_ia32_mfence();
2328 __builtin_ia32_clflush(event
);
2334 VkResult
anv_ResetEvent(
2338 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2339 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2341 event
->semaphore
= VK_EVENT_RESET
;
2343 if (!device
->info
.has_llc
) {
2344 /* Make sure the writes we're flushing have landed. */
2345 __builtin_ia32_mfence();
2346 __builtin_ia32_clflush(event
);
2354 VkResult
anv_CreateBuffer(
2356 const VkBufferCreateInfo
* pCreateInfo
,
2357 const VkAllocationCallbacks
* pAllocator
,
2360 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2361 struct anv_buffer
*buffer
;
2363 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2365 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2366 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2368 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2370 buffer
->size
= pCreateInfo
->size
;
2371 buffer
->usage
= pCreateInfo
->usage
;
2375 *pBuffer
= anv_buffer_to_handle(buffer
);
2380 void anv_DestroyBuffer(
2383 const VkAllocationCallbacks
* pAllocator
)
2385 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2386 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2391 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2395 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2396 enum isl_format format
,
2397 uint32_t offset
, uint32_t range
, uint32_t stride
)
2399 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2401 .mocs
= device
->default_mocs
,
2406 anv_state_flush(device
, state
);
2409 void anv_DestroySampler(
2412 const VkAllocationCallbacks
* pAllocator
)
2414 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2415 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2420 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2423 VkResult
anv_CreateFramebuffer(
2425 const VkFramebufferCreateInfo
* pCreateInfo
,
2426 const VkAllocationCallbacks
* pAllocator
,
2427 VkFramebuffer
* pFramebuffer
)
2429 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2430 struct anv_framebuffer
*framebuffer
;
2432 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2434 size_t size
= sizeof(*framebuffer
) +
2435 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2436 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2437 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2438 if (framebuffer
== NULL
)
2439 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2441 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2442 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2443 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2444 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2447 framebuffer
->width
= pCreateInfo
->width
;
2448 framebuffer
->height
= pCreateInfo
->height
;
2449 framebuffer
->layers
= pCreateInfo
->layers
;
2451 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2456 void anv_DestroyFramebuffer(
2459 const VkAllocationCallbacks
* pAllocator
)
2461 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2462 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2467 vk_free2(&device
->alloc
, pAllocator
, fb
);
2470 /* vk_icd.h does not declare this function, so we declare it here to
2471 * suppress Wmissing-prototypes.
2473 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2474 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2476 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2477 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2479 /* For the full details on loader interface versioning, see
2480 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2481 * What follows is a condensed summary, to help you navigate the large and
2482 * confusing official doc.
2484 * - Loader interface v0 is incompatible with later versions. We don't
2487 * - In loader interface v1:
2488 * - The first ICD entrypoint called by the loader is
2489 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2491 * - The ICD must statically expose no other Vulkan symbol unless it is
2492 * linked with -Bsymbolic.
2493 * - Each dispatchable Vulkan handle created by the ICD must be
2494 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2495 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2496 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2497 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2498 * such loader-managed surfaces.
2500 * - Loader interface v2 differs from v1 in:
2501 * - The first ICD entrypoint called by the loader is
2502 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2503 * statically expose this entrypoint.
2505 * - Loader interface v3 differs from v2 in:
2506 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2507 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2508 * because the loader no longer does so.
2510 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);