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_device_get_cache_uuid(void *uuid
, uint16_t pci_id
)
103 const struct build_id_note
*note
= build_id_find_nhdr("libvulkan_intel.so");
107 unsigned build_id_len
= build_id_length(note
);
108 if (build_id_len
< 20) /* It should be a SHA-1 */
111 struct mesa_sha1 sha1_ctx
;
113 STATIC_ASSERT(VK_UUID_SIZE
<= sizeof(sha1
));
115 _mesa_sha1_init(&sha1_ctx
);
116 _mesa_sha1_update(&sha1_ctx
, build_id_data(note
), build_id_len
);
117 _mesa_sha1_update(&sha1_ctx
, &pci_id
, sizeof(pci_id
));
118 _mesa_sha1_final(&sha1_ctx
, sha1
);
120 memcpy(uuid
, sha1
, VK_UUID_SIZE
);
125 anv_physical_device_init(struct anv_physical_device
*device
,
126 struct anv_instance
*instance
,
132 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
134 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
136 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
137 device
->instance
= instance
;
139 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
140 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
142 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
143 if (!device
->chipset_id
) {
144 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
148 device
->name
= gen_get_device_name(device
->chipset_id
);
149 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
150 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
154 if (device
->info
.is_haswell
) {
155 fprintf(stderr
, "WARNING: Haswell Vulkan support is incomplete\n");
156 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
157 fprintf(stderr
, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
158 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
159 fprintf(stderr
, "WARNING: Bay Trail Vulkan support is incomplete\n");
160 } else if (device
->info
.gen
>= 8) {
161 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
162 * supported as anything */
164 result
= vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
165 "Vulkan not yet supported on %s", device
->name
);
169 device
->cmd_parser_version
= -1;
170 if (device
->info
.gen
== 7) {
171 device
->cmd_parser_version
=
172 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
173 if (device
->cmd_parser_version
== -1) {
174 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
175 "failed to get command parser version");
180 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
181 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
182 "kernel missing gem wait");
186 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
187 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
188 "kernel missing execbuf2");
192 if (!device
->info
.has_llc
&&
193 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
194 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
195 "kernel missing wc mmap");
199 device
->supports_48bit_addresses
= anv_gem_supports_48b_addresses(fd
);
201 result
= anv_compute_heap_size(fd
, &device
->heap_size
);
202 if (result
!= VK_SUCCESS
)
205 if (!anv_device_get_cache_uuid(device
->uuid
, device
->chipset_id
)) {
206 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
207 "cannot generate UUID");
210 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
212 /* GENs prior to 8 do not support EU/Subslice info */
213 if (device
->info
.gen
>= 8) {
214 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
215 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
217 /* Without this information, we cannot get the right Braswell
218 * brandstrings, and we have to use conservative numbers for GPGPU on
219 * many platforms, but otherwise, things will just work.
221 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
222 fprintf(stderr
, "WARNING: Kernel 4.1 required to properly"
223 " query GPU properties.\n");
225 } else if (device
->info
.gen
== 7) {
226 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
229 if (device
->info
.is_cherryview
&&
230 device
->subslice_total
> 0 && device
->eu_total
> 0) {
231 /* Logical CS threads = EUs per subslice * 7 threads per EU */
232 uint32_t max_cs_threads
= device
->eu_total
/ device
->subslice_total
* 7;
234 /* Fuse configurations may give more threads than expected, never less. */
235 if (max_cs_threads
> device
->info
.max_cs_threads
)
236 device
->info
.max_cs_threads
= max_cs_threads
;
239 brw_process_intel_debug_variable();
241 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
242 if (device
->compiler
== NULL
) {
243 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
246 device
->compiler
->shader_debug_log
= compiler_debug_log
;
247 device
->compiler
->shader_perf_log
= compiler_perf_log
;
249 result
= anv_init_wsi(device
);
250 if (result
!= VK_SUCCESS
) {
251 ralloc_free(device
->compiler
);
255 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
257 device
->local_fd
= fd
;
266 anv_physical_device_finish(struct anv_physical_device
*device
)
268 anv_finish_wsi(device
);
269 ralloc_free(device
->compiler
);
270 close(device
->local_fd
);
273 static const VkExtensionProperties global_extensions
[] = {
275 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
278 #ifdef VK_USE_PLATFORM_XCB_KHR
280 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
284 #ifdef VK_USE_PLATFORM_XLIB_KHR
286 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
290 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
292 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
297 .extensionName
= VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME
,
302 static const VkExtensionProperties device_extensions
[] = {
304 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
308 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
312 .extensionName
= VK_KHR_MAINTENANCE1_EXTENSION_NAME
,
316 .extensionName
= VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME
,
320 .extensionName
= VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME
,
324 .extensionName
= VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME
,
328 .extensionName
= VK_KHR_INCREMENTAL_PRESENT_EXTENSION_NAME
,
334 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
335 VkSystemAllocationScope allocationScope
)
341 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
342 size_t align
, VkSystemAllocationScope allocationScope
)
344 return realloc(pOriginal
, size
);
348 default_free_func(void *pUserData
, void *pMemory
)
353 static const VkAllocationCallbacks default_alloc
= {
355 .pfnAllocation
= default_alloc_func
,
356 .pfnReallocation
= default_realloc_func
,
357 .pfnFree
= default_free_func
,
360 VkResult
anv_CreateInstance(
361 const VkInstanceCreateInfo
* pCreateInfo
,
362 const VkAllocationCallbacks
* pAllocator
,
363 VkInstance
* pInstance
)
365 struct anv_instance
*instance
;
367 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
369 uint32_t client_version
;
370 if (pCreateInfo
->pApplicationInfo
&&
371 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
372 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
374 client_version
= VK_MAKE_VERSION(1, 0, 0);
377 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
378 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
379 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
380 "Client requested version %d.%d.%d",
381 VK_VERSION_MAJOR(client_version
),
382 VK_VERSION_MINOR(client_version
),
383 VK_VERSION_PATCH(client_version
));
386 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
388 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
389 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
390 global_extensions
[j
].extensionName
) == 0) {
396 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
399 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
400 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
402 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
404 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
407 instance
->alloc
= *pAllocator
;
409 instance
->alloc
= default_alloc
;
411 instance
->apiVersion
= client_version
;
412 instance
->physicalDeviceCount
= -1;
416 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
418 *pInstance
= anv_instance_to_handle(instance
);
423 void anv_DestroyInstance(
424 VkInstance _instance
,
425 const VkAllocationCallbacks
* pAllocator
)
427 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
432 if (instance
->physicalDeviceCount
> 0) {
433 /* We support at most one physical device. */
434 assert(instance
->physicalDeviceCount
== 1);
435 anv_physical_device_finish(&instance
->physicalDevice
);
438 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
442 vk_free(&instance
->alloc
, instance
);
446 anv_enumerate_devices(struct anv_instance
*instance
)
448 /* TODO: Check for more devices ? */
449 drmDevicePtr devices
[8];
450 VkResult result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
453 instance
->physicalDeviceCount
= 0;
455 max_devices
= drmGetDevices2(0, devices
, sizeof(devices
));
457 return VK_ERROR_INCOMPATIBLE_DRIVER
;
459 for (unsigned i
= 0; i
< (unsigned)max_devices
; i
++) {
460 if (devices
[i
]->available_nodes
& 1 << DRM_NODE_RENDER
&&
461 devices
[i
]->bustype
== DRM_BUS_PCI
&&
462 devices
[i
]->deviceinfo
.pci
->vendor_id
== 0x8086) {
464 result
= anv_physical_device_init(&instance
->physicalDevice
,
466 devices
[i
]->nodes
[DRM_NODE_RENDER
]);
467 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
472 if (result
== VK_SUCCESS
)
473 instance
->physicalDeviceCount
= 1;
479 VkResult
anv_EnumeratePhysicalDevices(
480 VkInstance _instance
,
481 uint32_t* pPhysicalDeviceCount
,
482 VkPhysicalDevice
* pPhysicalDevices
)
484 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
485 VK_OUTARRAY_MAKE(out
, pPhysicalDevices
, pPhysicalDeviceCount
);
488 if (instance
->physicalDeviceCount
< 0) {
489 result
= anv_enumerate_devices(instance
);
490 if (result
!= VK_SUCCESS
&&
491 result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
495 if (instance
->physicalDeviceCount
> 0) {
496 assert(instance
->physicalDeviceCount
== 1);
497 vk_outarray_append(&out
, i
) {
498 *i
= anv_physical_device_to_handle(&instance
->physicalDevice
);
502 return vk_outarray_status(&out
);
505 void anv_GetPhysicalDeviceFeatures(
506 VkPhysicalDevice physicalDevice
,
507 VkPhysicalDeviceFeatures
* pFeatures
)
509 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
511 *pFeatures
= (VkPhysicalDeviceFeatures
) {
512 .robustBufferAccess
= true,
513 .fullDrawIndexUint32
= true,
514 .imageCubeArray
= true,
515 .independentBlend
= true,
516 .geometryShader
= true,
517 .tessellationShader
= true,
518 .sampleRateShading
= true,
519 .dualSrcBlend
= true,
521 .multiDrawIndirect
= false,
522 .drawIndirectFirstInstance
= true,
524 .depthBiasClamp
= true,
525 .fillModeNonSolid
= true,
526 .depthBounds
= false,
530 .multiViewport
= true,
531 .samplerAnisotropy
= true,
532 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
533 pdevice
->info
.is_baytrail
,
534 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
535 .textureCompressionBC
= true,
536 .occlusionQueryPrecise
= true,
537 .pipelineStatisticsQuery
= true,
538 .fragmentStoresAndAtomics
= true,
539 .shaderTessellationAndGeometryPointSize
= true,
540 .shaderImageGatherExtended
= true,
541 .shaderStorageImageExtendedFormats
= true,
542 .shaderStorageImageMultisample
= false,
543 .shaderStorageImageReadWithoutFormat
= false,
544 .shaderStorageImageWriteWithoutFormat
= true,
545 .shaderUniformBufferArrayDynamicIndexing
= true,
546 .shaderSampledImageArrayDynamicIndexing
= true,
547 .shaderStorageBufferArrayDynamicIndexing
= true,
548 .shaderStorageImageArrayDynamicIndexing
= true,
549 .shaderClipDistance
= true,
550 .shaderCullDistance
= true,
551 .shaderFloat64
= pdevice
->info
.gen
>= 8,
552 .shaderInt64
= pdevice
->info
.gen
>= 8,
553 .shaderInt16
= false,
554 .shaderResourceMinLod
= false,
555 .variableMultisampleRate
= false,
556 .inheritedQueries
= true,
559 /* We can't do image stores in vec4 shaders */
560 pFeatures
->vertexPipelineStoresAndAtomics
=
561 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
562 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
565 void anv_GetPhysicalDeviceFeatures2KHR(
566 VkPhysicalDevice physicalDevice
,
567 VkPhysicalDeviceFeatures2KHR
* pFeatures
)
569 anv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
571 vk_foreach_struct(ext
, pFeatures
->pNext
) {
572 switch (ext
->sType
) {
574 anv_debug_ignored_stype(ext
->sType
);
580 void anv_GetPhysicalDeviceProperties(
581 VkPhysicalDevice physicalDevice
,
582 VkPhysicalDeviceProperties
* pProperties
)
584 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
585 const struct gen_device_info
*devinfo
= &pdevice
->info
;
587 /* See assertions made when programming the buffer surface state. */
588 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
589 (1ul << 30) : (1ul << 27);
591 VkSampleCountFlags sample_counts
=
592 isl_device_get_sample_counts(&pdevice
->isl_dev
);
594 VkPhysicalDeviceLimits limits
= {
595 .maxImageDimension1D
= (1 << 14),
596 .maxImageDimension2D
= (1 << 14),
597 .maxImageDimension3D
= (1 << 11),
598 .maxImageDimensionCube
= (1 << 14),
599 .maxImageArrayLayers
= (1 << 11),
600 .maxTexelBufferElements
= 128 * 1024 * 1024,
601 .maxUniformBufferRange
= (1ul << 27),
602 .maxStorageBufferRange
= max_raw_buffer_sz
,
603 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
604 .maxMemoryAllocationCount
= UINT32_MAX
,
605 .maxSamplerAllocationCount
= 64 * 1024,
606 .bufferImageGranularity
= 64, /* A cache line */
607 .sparseAddressSpaceSize
= 0,
608 .maxBoundDescriptorSets
= MAX_SETS
,
609 .maxPerStageDescriptorSamplers
= 64,
610 .maxPerStageDescriptorUniformBuffers
= 64,
611 .maxPerStageDescriptorStorageBuffers
= 64,
612 .maxPerStageDescriptorSampledImages
= 64,
613 .maxPerStageDescriptorStorageImages
= 64,
614 .maxPerStageDescriptorInputAttachments
= 64,
615 .maxPerStageResources
= 128,
616 .maxDescriptorSetSamplers
= 256,
617 .maxDescriptorSetUniformBuffers
= 256,
618 .maxDescriptorSetUniformBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
619 .maxDescriptorSetStorageBuffers
= 256,
620 .maxDescriptorSetStorageBuffersDynamic
= MAX_DYNAMIC_BUFFERS
/ 2,
621 .maxDescriptorSetSampledImages
= 256,
622 .maxDescriptorSetStorageImages
= 256,
623 .maxDescriptorSetInputAttachments
= 256,
624 .maxVertexInputAttributes
= MAX_VBS
,
625 .maxVertexInputBindings
= MAX_VBS
,
626 .maxVertexInputAttributeOffset
= 2047,
627 .maxVertexInputBindingStride
= 2048,
628 .maxVertexOutputComponents
= 128,
629 .maxTessellationGenerationLevel
= 64,
630 .maxTessellationPatchSize
= 32,
631 .maxTessellationControlPerVertexInputComponents
= 128,
632 .maxTessellationControlPerVertexOutputComponents
= 128,
633 .maxTessellationControlPerPatchOutputComponents
= 128,
634 .maxTessellationControlTotalOutputComponents
= 2048,
635 .maxTessellationEvaluationInputComponents
= 128,
636 .maxTessellationEvaluationOutputComponents
= 128,
637 .maxGeometryShaderInvocations
= 32,
638 .maxGeometryInputComponents
= 64,
639 .maxGeometryOutputComponents
= 128,
640 .maxGeometryOutputVertices
= 256,
641 .maxGeometryTotalOutputComponents
= 1024,
642 .maxFragmentInputComponents
= 128,
643 .maxFragmentOutputAttachments
= 8,
644 .maxFragmentDualSrcAttachments
= 1,
645 .maxFragmentCombinedOutputResources
= 8,
646 .maxComputeSharedMemorySize
= 32768,
647 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
648 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
649 .maxComputeWorkGroupSize
= {
650 16 * devinfo
->max_cs_threads
,
651 16 * devinfo
->max_cs_threads
,
652 16 * devinfo
->max_cs_threads
,
654 .subPixelPrecisionBits
= 4 /* FIXME */,
655 .subTexelPrecisionBits
= 4 /* FIXME */,
656 .mipmapPrecisionBits
= 4 /* FIXME */,
657 .maxDrawIndexedIndexValue
= UINT32_MAX
,
658 .maxDrawIndirectCount
= UINT32_MAX
,
659 .maxSamplerLodBias
= 16,
660 .maxSamplerAnisotropy
= 16,
661 .maxViewports
= MAX_VIEWPORTS
,
662 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
663 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
664 .viewportSubPixelBits
= 13, /* We take a float? */
665 .minMemoryMapAlignment
= 4096, /* A page */
666 .minTexelBufferOffsetAlignment
= 1,
667 .minUniformBufferOffsetAlignment
= 16,
668 .minStorageBufferOffsetAlignment
= 4,
669 .minTexelOffset
= -8,
671 .minTexelGatherOffset
= -32,
672 .maxTexelGatherOffset
= 31,
673 .minInterpolationOffset
= -0.5,
674 .maxInterpolationOffset
= 0.4375,
675 .subPixelInterpolationOffsetBits
= 4,
676 .maxFramebufferWidth
= (1 << 14),
677 .maxFramebufferHeight
= (1 << 14),
678 .maxFramebufferLayers
= (1 << 11),
679 .framebufferColorSampleCounts
= sample_counts
,
680 .framebufferDepthSampleCounts
= sample_counts
,
681 .framebufferStencilSampleCounts
= sample_counts
,
682 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
683 .maxColorAttachments
= MAX_RTS
,
684 .sampledImageColorSampleCounts
= sample_counts
,
685 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
686 .sampledImageDepthSampleCounts
= sample_counts
,
687 .sampledImageStencilSampleCounts
= sample_counts
,
688 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
689 .maxSampleMaskWords
= 1,
690 .timestampComputeAndGraphics
= false,
691 .timestampPeriod
= devinfo
->timebase_scale
,
692 .maxClipDistances
= 8,
693 .maxCullDistances
= 8,
694 .maxCombinedClipAndCullDistances
= 8,
695 .discreteQueuePriorities
= 1,
696 .pointSizeRange
= { 0.125, 255.875 },
697 .lineWidthRange
= { 0.0, 7.9921875 },
698 .pointSizeGranularity
= (1.0 / 8.0),
699 .lineWidthGranularity
= (1.0 / 128.0),
700 .strictLines
= false, /* FINISHME */
701 .standardSampleLocations
= true,
702 .optimalBufferCopyOffsetAlignment
= 128,
703 .optimalBufferCopyRowPitchAlignment
= 128,
704 .nonCoherentAtomSize
= 64,
707 *pProperties
= (VkPhysicalDeviceProperties
) {
708 .apiVersion
= VK_MAKE_VERSION(1, 0, 42),
711 .deviceID
= pdevice
->chipset_id
,
712 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
714 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
717 strcpy(pProperties
->deviceName
, pdevice
->name
);
718 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
721 void anv_GetPhysicalDeviceProperties2KHR(
722 VkPhysicalDevice physicalDevice
,
723 VkPhysicalDeviceProperties2KHR
* pProperties
)
725 anv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
727 vk_foreach_struct(ext
, pProperties
->pNext
) {
728 switch (ext
->sType
) {
729 case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR
: {
730 VkPhysicalDevicePushDescriptorPropertiesKHR
*properties
=
731 (VkPhysicalDevicePushDescriptorPropertiesKHR
*) ext
;
733 properties
->maxPushDescriptors
= MAX_PUSH_DESCRIPTORS
;
738 anv_debug_ignored_stype(ext
->sType
);
744 /* We support exactly one queue family. */
745 static const VkQueueFamilyProperties
746 anv_queue_family_properties
= {
747 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
748 VK_QUEUE_COMPUTE_BIT
|
749 VK_QUEUE_TRANSFER_BIT
,
751 .timestampValidBits
= 36, /* XXX: Real value here */
752 .minImageTransferGranularity
= { 1, 1, 1 },
755 void anv_GetPhysicalDeviceQueueFamilyProperties(
756 VkPhysicalDevice physicalDevice
,
758 VkQueueFamilyProperties
* pQueueFamilyProperties
)
760 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pCount
);
762 vk_outarray_append(&out
, p
) {
763 *p
= anv_queue_family_properties
;
767 void anv_GetPhysicalDeviceQueueFamilyProperties2KHR(
768 VkPhysicalDevice physicalDevice
,
769 uint32_t* pQueueFamilyPropertyCount
,
770 VkQueueFamilyProperties2KHR
* pQueueFamilyProperties
)
773 VK_OUTARRAY_MAKE(out
, pQueueFamilyProperties
, pQueueFamilyPropertyCount
);
775 vk_outarray_append(&out
, p
) {
776 p
->queueFamilyProperties
= anv_queue_family_properties
;
778 vk_foreach_struct(s
, p
->pNext
) {
779 anv_debug_ignored_stype(s
->sType
);
784 void anv_GetPhysicalDeviceMemoryProperties(
785 VkPhysicalDevice physicalDevice
,
786 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
788 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
790 if (physical_device
->info
.has_llc
) {
791 /* Big core GPUs share LLC with the CPU and thus one memory type can be
792 * both cached and coherent at the same time.
794 pMemoryProperties
->memoryTypeCount
= 1;
795 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
796 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
797 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
798 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
799 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
803 /* The spec requires that we expose a host-visible, coherent memory
804 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
805 * to give the application a choice between cached, but not coherent and
806 * coherent but uncached (WC though).
808 pMemoryProperties
->memoryTypeCount
= 2;
809 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
810 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
811 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
812 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
815 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
816 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
817 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
818 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
823 pMemoryProperties
->memoryHeapCount
= 1;
824 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
825 .size
= physical_device
->heap_size
,
826 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
830 void anv_GetPhysicalDeviceMemoryProperties2KHR(
831 VkPhysicalDevice physicalDevice
,
832 VkPhysicalDeviceMemoryProperties2KHR
* pMemoryProperties
)
834 anv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
835 &pMemoryProperties
->memoryProperties
);
837 vk_foreach_struct(ext
, pMemoryProperties
->pNext
) {
838 switch (ext
->sType
) {
840 anv_debug_ignored_stype(ext
->sType
);
846 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
850 return anv_lookup_entrypoint(NULL
, pName
);
853 /* With version 1+ of the loader interface the ICD should expose
854 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
857 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
862 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
866 return anv_GetInstanceProcAddr(instance
, pName
);
869 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
873 ANV_FROM_HANDLE(anv_device
, device
, _device
);
874 return anv_lookup_entrypoint(&device
->info
, pName
);
878 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
880 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
881 queue
->device
= device
;
882 queue
->pool
= &device
->surface_state_pool
;
886 anv_queue_finish(struct anv_queue
*queue
)
890 static struct anv_state
891 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
893 struct anv_state state
;
895 state
= anv_state_pool_alloc(pool
, size
, align
);
896 memcpy(state
.map
, p
, size
);
898 anv_state_flush(pool
->block_pool
->device
, state
);
903 struct gen8_border_color
{
908 /* Pad out to 64 bytes */
913 anv_device_init_border_colors(struct anv_device
*device
)
915 static const struct gen8_border_color border_colors
[] = {
916 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
917 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
918 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
919 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
920 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
921 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
924 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
925 sizeof(border_colors
), 64,
930 anv_device_submit_simple_batch(struct anv_device
*device
,
931 struct anv_batch
*batch
)
933 struct drm_i915_gem_execbuffer2 execbuf
;
934 struct drm_i915_gem_exec_object2 exec2_objects
[1];
935 struct anv_bo bo
, *exec_bos
[1];
936 VkResult result
= VK_SUCCESS
;
939 /* Kernel driver requires 8 byte aligned batch length */
940 size
= align_u32(batch
->next
- batch
->start
, 8);
941 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &bo
, size
);
942 if (result
!= VK_SUCCESS
)
945 memcpy(bo
.map
, batch
->start
, size
);
946 if (!device
->info
.has_llc
)
947 anv_flush_range(bo
.map
, size
);
950 exec2_objects
[0].handle
= bo
.gem_handle
;
951 exec2_objects
[0].relocation_count
= 0;
952 exec2_objects
[0].relocs_ptr
= 0;
953 exec2_objects
[0].alignment
= 0;
954 exec2_objects
[0].offset
= bo
.offset
;
955 exec2_objects
[0].flags
= 0;
956 exec2_objects
[0].rsvd1
= 0;
957 exec2_objects
[0].rsvd2
= 0;
959 execbuf
.buffers_ptr
= (uintptr_t) exec2_objects
;
960 execbuf
.buffer_count
= 1;
961 execbuf
.batch_start_offset
= 0;
962 execbuf
.batch_len
= size
;
963 execbuf
.cliprects_ptr
= 0;
964 execbuf
.num_cliprects
= 0;
969 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
970 execbuf
.rsvd1
= device
->context_id
;
973 result
= anv_device_execbuf(device
, &execbuf
, exec_bos
);
974 if (result
!= VK_SUCCESS
)
977 result
= anv_device_wait(device
, &bo
, INT64_MAX
);
980 anv_bo_pool_free(&device
->batch_bo_pool
, &bo
);
985 VkResult
anv_CreateDevice(
986 VkPhysicalDevice physicalDevice
,
987 const VkDeviceCreateInfo
* pCreateInfo
,
988 const VkAllocationCallbacks
* pAllocator
,
991 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
993 struct anv_device
*device
;
995 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
997 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
999 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
1000 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
1001 device_extensions
[j
].extensionName
) == 0) {
1007 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
1010 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
1012 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
1014 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1016 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1017 device
->instance
= physical_device
->instance
;
1018 device
->chipset_id
= physical_device
->chipset_id
;
1019 device
->lost
= false;
1022 device
->alloc
= *pAllocator
;
1024 device
->alloc
= physical_device
->instance
->alloc
;
1026 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
1027 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
1028 if (device
->fd
== -1) {
1029 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1033 device
->context_id
= anv_gem_create_context(device
);
1034 if (device
->context_id
== -1) {
1035 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1039 device
->info
= physical_device
->info
;
1040 device
->isl_dev
= physical_device
->isl_dev
;
1042 /* On Broadwell and later, we can use batch chaining to more efficiently
1043 * implement growing command buffers. Prior to Haswell, the kernel
1044 * command parser gets in the way and we have to fall back to growing
1047 device
->can_chain_batches
= device
->info
.gen
>= 8;
1049 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
1050 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
1052 if (pthread_mutex_init(&device
->mutex
, NULL
) != 0) {
1053 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1054 goto fail_context_id
;
1057 pthread_condattr_t condattr
;
1058 if (pthread_condattr_init(&condattr
) != 0) {
1059 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1062 if (pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
) != 0) {
1063 pthread_condattr_destroy(&condattr
);
1064 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1067 if (pthread_cond_init(&device
->queue_submit
, NULL
) != 0) {
1068 pthread_condattr_destroy(&condattr
);
1069 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
1072 pthread_condattr_destroy(&condattr
);
1074 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
1076 result
= anv_block_pool_init(&device
->dynamic_state_block_pool
, device
,
1078 if (result
!= VK_SUCCESS
)
1079 goto fail_batch_bo_pool
;
1081 anv_state_pool_init(&device
->dynamic_state_pool
,
1082 &device
->dynamic_state_block_pool
);
1084 result
= anv_block_pool_init(&device
->instruction_block_pool
, device
,
1086 if (result
!= VK_SUCCESS
)
1087 goto fail_dynamic_state_pool
;
1089 anv_state_pool_init(&device
->instruction_state_pool
,
1090 &device
->instruction_block_pool
);
1092 result
= anv_block_pool_init(&device
->surface_state_block_pool
, device
,
1094 if (result
!= VK_SUCCESS
)
1095 goto fail_instruction_state_pool
;
1097 anv_state_pool_init(&device
->surface_state_pool
,
1098 &device
->surface_state_block_pool
);
1100 result
= anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
1101 if (result
!= VK_SUCCESS
)
1102 goto fail_surface_state_pool
;
1104 anv_scratch_pool_init(device
, &device
->scratch_pool
);
1106 anv_queue_init(device
, &device
->queue
);
1108 switch (device
->info
.gen
) {
1110 if (!device
->info
.is_haswell
)
1111 result
= gen7_init_device_state(device
);
1113 result
= gen75_init_device_state(device
);
1116 result
= gen8_init_device_state(device
);
1119 result
= gen9_init_device_state(device
);
1122 /* Shouldn't get here as we don't create physical devices for any other
1124 unreachable("unhandled gen");
1126 if (result
!= VK_SUCCESS
)
1127 goto fail_workaround_bo
;
1129 anv_device_init_blorp(device
);
1131 anv_device_init_border_colors(device
);
1133 *pDevice
= anv_device_to_handle(device
);
1138 anv_queue_finish(&device
->queue
);
1139 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1140 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1141 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1142 fail_surface_state_pool
:
1143 anv_state_pool_finish(&device
->surface_state_pool
);
1144 anv_block_pool_finish(&device
->surface_state_block_pool
);
1145 fail_instruction_state_pool
:
1146 anv_state_pool_finish(&device
->instruction_state_pool
);
1147 anv_block_pool_finish(&device
->instruction_block_pool
);
1148 fail_dynamic_state_pool
:
1149 anv_state_pool_finish(&device
->dynamic_state_pool
);
1150 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1152 anv_bo_pool_finish(&device
->batch_bo_pool
);
1153 pthread_cond_destroy(&device
->queue_submit
);
1155 pthread_mutex_destroy(&device
->mutex
);
1157 anv_gem_destroy_context(device
, device
->context_id
);
1161 vk_free(&device
->alloc
, device
);
1166 void anv_DestroyDevice(
1168 const VkAllocationCallbacks
* pAllocator
)
1170 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1175 anv_device_finish_blorp(device
);
1177 anv_queue_finish(&device
->queue
);
1179 #ifdef HAVE_VALGRIND
1180 /* We only need to free these to prevent valgrind errors. The backing
1181 * BO will go away in a couple of lines so we don't actually leak.
1183 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1186 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1188 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1189 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1191 anv_state_pool_finish(&device
->surface_state_pool
);
1192 anv_block_pool_finish(&device
->surface_state_block_pool
);
1193 anv_state_pool_finish(&device
->instruction_state_pool
);
1194 anv_block_pool_finish(&device
->instruction_block_pool
);
1195 anv_state_pool_finish(&device
->dynamic_state_pool
);
1196 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1198 anv_bo_pool_finish(&device
->batch_bo_pool
);
1200 pthread_cond_destroy(&device
->queue_submit
);
1201 pthread_mutex_destroy(&device
->mutex
);
1203 anv_gem_destroy_context(device
, device
->context_id
);
1207 vk_free(&device
->alloc
, device
);
1210 VkResult
anv_EnumerateInstanceExtensionProperties(
1211 const char* pLayerName
,
1212 uint32_t* pPropertyCount
,
1213 VkExtensionProperties
* pProperties
)
1215 if (pProperties
== NULL
) {
1216 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1220 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(global_extensions
));
1221 typed_memcpy(pProperties
, global_extensions
, *pPropertyCount
);
1223 if (*pPropertyCount
< ARRAY_SIZE(global_extensions
))
1224 return VK_INCOMPLETE
;
1229 VkResult
anv_EnumerateDeviceExtensionProperties(
1230 VkPhysicalDevice physicalDevice
,
1231 const char* pLayerName
,
1232 uint32_t* pPropertyCount
,
1233 VkExtensionProperties
* pProperties
)
1235 if (pProperties
== NULL
) {
1236 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1240 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(device_extensions
));
1241 typed_memcpy(pProperties
, device_extensions
, *pPropertyCount
);
1243 if (*pPropertyCount
< ARRAY_SIZE(device_extensions
))
1244 return VK_INCOMPLETE
;
1249 VkResult
anv_EnumerateInstanceLayerProperties(
1250 uint32_t* pPropertyCount
,
1251 VkLayerProperties
* pProperties
)
1253 if (pProperties
== NULL
) {
1254 *pPropertyCount
= 0;
1258 /* None supported at this time */
1259 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1262 VkResult
anv_EnumerateDeviceLayerProperties(
1263 VkPhysicalDevice physicalDevice
,
1264 uint32_t* pPropertyCount
,
1265 VkLayerProperties
* pProperties
)
1267 if (pProperties
== NULL
) {
1268 *pPropertyCount
= 0;
1272 /* None supported at this time */
1273 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1276 void anv_GetDeviceQueue(
1278 uint32_t queueNodeIndex
,
1279 uint32_t queueIndex
,
1282 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1284 assert(queueIndex
== 0);
1286 *pQueue
= anv_queue_to_handle(&device
->queue
);
1290 anv_device_execbuf(struct anv_device
*device
,
1291 struct drm_i915_gem_execbuffer2
*execbuf
,
1292 struct anv_bo
**execbuf_bos
)
1294 int ret
= anv_gem_execbuffer(device
, execbuf
);
1296 /* We don't know the real error. */
1297 device
->lost
= true;
1298 return vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
1301 struct drm_i915_gem_exec_object2
*objects
=
1302 (void *)(uintptr_t)execbuf
->buffers_ptr
;
1303 for (uint32_t k
= 0; k
< execbuf
->buffer_count
; k
++)
1304 execbuf_bos
[k
]->offset
= objects
[k
].offset
;
1310 anv_device_query_status(struct anv_device
*device
)
1312 /* This isn't likely as most of the callers of this function already check
1313 * for it. However, it doesn't hurt to check and it potentially lets us
1316 if (unlikely(device
->lost
))
1317 return VK_ERROR_DEVICE_LOST
;
1319 uint32_t active
, pending
;
1320 int ret
= anv_gem_gpu_get_reset_stats(device
, &active
, &pending
);
1322 /* We don't know the real error. */
1323 device
->lost
= true;
1324 return vk_errorf(VK_ERROR_DEVICE_LOST
, "get_reset_stats failed: %m");
1328 device
->lost
= true;
1329 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1330 "GPU hung on one of our command buffers");
1331 } else if (pending
) {
1332 device
->lost
= true;
1333 return vk_errorf(VK_ERROR_DEVICE_LOST
,
1334 "GPU hung with commands in-flight");
1341 anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
)
1343 /* Note: This only returns whether or not the BO is in use by an i915 GPU.
1344 * Other usages of the BO (such as on different hardware) will not be
1345 * flagged as "busy" by this ioctl. Use with care.
1347 int ret
= anv_gem_busy(device
, bo
->gem_handle
);
1349 return VK_NOT_READY
;
1350 } else if (ret
== -1) {
1351 /* We don't know the real error. */
1352 device
->lost
= true;
1353 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1356 /* Query for device status after the busy call. If the BO we're checking
1357 * got caught in a GPU hang we don't want to return VK_SUCCESS to the
1358 * client because it clearly doesn't have valid data. Yes, this most
1359 * likely means an ioctl, but we just did an ioctl to query the busy status
1360 * so it's no great loss.
1362 return anv_device_query_status(device
);
1366 anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1369 int ret
= anv_gem_wait(device
, bo
->gem_handle
, &timeout
);
1370 if (ret
== -1 && errno
== ETIME
) {
1372 } else if (ret
== -1) {
1373 /* We don't know the real error. */
1374 device
->lost
= true;
1375 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1378 /* Query for device status after the wait. If the BO we're waiting on got
1379 * caught in a GPU hang we don't want to return VK_SUCCESS to the client
1380 * because it clearly doesn't have valid data. Yes, this most likely means
1381 * an ioctl, but we just did an ioctl to wait so it's no great loss.
1383 return anv_device_query_status(device
);
1386 VkResult
anv_QueueSubmit(
1388 uint32_t submitCount
,
1389 const VkSubmitInfo
* pSubmits
,
1392 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1393 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1394 struct anv_device
*device
= queue
->device
;
1396 /* Query for device status prior to submitting. Technically, we don't need
1397 * to do this. However, if we have a client that's submitting piles of
1398 * garbage, we would rather break as early as possible to keep the GPU
1399 * hanging contained. If we don't check here, we'll either be waiting for
1400 * the kernel to kick us or we'll have to wait until the client waits on a
1401 * fence before we actually know whether or not we've hung.
1403 VkResult result
= anv_device_query_status(device
);
1404 if (result
!= VK_SUCCESS
)
1407 /* We lock around QueueSubmit for three main reasons:
1409 * 1) When a block pool is resized, we create a new gem handle with a
1410 * different size and, in the case of surface states, possibly a
1411 * different center offset but we re-use the same anv_bo struct when
1412 * we do so. If this happens in the middle of setting up an execbuf,
1413 * we could end up with our list of BOs out of sync with our list of
1416 * 2) The algorithm we use for building the list of unique buffers isn't
1417 * thread-safe. While the client is supposed to syncronize around
1418 * QueueSubmit, this would be extremely difficult to debug if it ever
1419 * came up in the wild due to a broken app. It's better to play it
1420 * safe and just lock around QueueSubmit.
1422 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
1423 * userspace. Due to the fact that the surface state buffer is shared
1424 * between batches, we can't afford to have that happen from multiple
1425 * threads at the same time. Even though the user is supposed to
1426 * ensure this doesn't happen, we play it safe as in (2) above.
1428 * Since the only other things that ever take the device lock such as block
1429 * pool resize only rarely happen, this will almost never be contended so
1430 * taking a lock isn't really an expensive operation in this case.
1432 pthread_mutex_lock(&device
->mutex
);
1434 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1435 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1436 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
,
1437 pSubmits
[i
].pCommandBuffers
[j
]);
1438 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1439 assert(!anv_batch_has_error(&cmd_buffer
->batch
));
1441 result
= anv_cmd_buffer_execbuf(device
, cmd_buffer
);
1442 if (result
!= VK_SUCCESS
)
1448 struct anv_bo
*fence_bo
= &fence
->bo
;
1449 result
= anv_device_execbuf(device
, &fence
->execbuf
, &fence_bo
);
1450 if (result
!= VK_SUCCESS
)
1453 /* Update the fence and wake up any waiters */
1454 assert(fence
->state
== ANV_FENCE_STATE_RESET
);
1455 fence
->state
= ANV_FENCE_STATE_SUBMITTED
;
1456 pthread_cond_broadcast(&device
->queue_submit
);
1460 if (result
!= VK_SUCCESS
) {
1461 /* In the case that something has gone wrong we may end up with an
1462 * inconsistent state from which it may not be trivial to recover.
1463 * For example, we might have computed address relocations and
1464 * any future attempt to re-submit this job will need to know about
1465 * this and avoid computing relocation addresses again.
1467 * To avoid this sort of issues, we assume that if something was
1468 * wrong during submission we must already be in a really bad situation
1469 * anyway (such us being out of memory) and return
1470 * VK_ERROR_DEVICE_LOST to ensure that clients do not attempt to
1471 * submit the same job again to this device.
1473 result
= VK_ERROR_DEVICE_LOST
;
1474 device
->lost
= true;
1476 /* If we return VK_ERROR_DEVICE LOST here, we need to ensure that
1477 * vkWaitForFences() and vkGetFenceStatus() return a valid result
1478 * (VK_SUCCESS or VK_ERROR_DEVICE_LOST) in a finite amount of time.
1479 * Setting the fence status to SIGNALED ensures this will happen in
1483 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1486 pthread_mutex_unlock(&device
->mutex
);
1491 VkResult
anv_QueueWaitIdle(
1494 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1496 return anv_DeviceWaitIdle(anv_device_to_handle(queue
->device
));
1499 VkResult
anv_DeviceWaitIdle(
1502 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1503 if (unlikely(device
->lost
))
1504 return VK_ERROR_DEVICE_LOST
;
1506 struct anv_batch batch
;
1509 batch
.start
= batch
.next
= cmds
;
1510 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1512 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1513 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1515 return anv_device_submit_simple_batch(device
, &batch
);
1519 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1521 uint32_t gem_handle
= anv_gem_create(device
, size
);
1523 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1525 anv_bo_init(bo
, gem_handle
, size
);
1527 if (device
->instance
->physicalDevice
.supports_48bit_addresses
)
1528 bo
->flags
|= EXEC_OBJECT_SUPPORTS_48B_ADDRESS
;
1533 VkResult
anv_AllocateMemory(
1535 const VkMemoryAllocateInfo
* pAllocateInfo
,
1536 const VkAllocationCallbacks
* pAllocator
,
1537 VkDeviceMemory
* pMem
)
1539 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1540 struct anv_device_memory
*mem
;
1543 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1545 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1546 assert(pAllocateInfo
->allocationSize
> 0);
1548 /* We support exactly one memory heap. */
1549 assert(pAllocateInfo
->memoryTypeIndex
== 0 ||
1550 (!device
->info
.has_llc
&& pAllocateInfo
->memoryTypeIndex
< 2));
1552 /* The kernel relocation API has a limitation of a 32-bit delta value
1553 * applied to the address before it is written which, in spite of it being
1554 * unsigned, is treated as signed . Because of the way that this maps to
1555 * the Vulkan API, we cannot handle an offset into a buffer that does not
1556 * fit into a signed 32 bits. The only mechanism we have for dealing with
1557 * this at the moment is to limit all VkDeviceMemory objects to a maximum
1558 * of 2GB each. The Vulkan spec allows us to do this:
1560 * "Some platforms may have a limit on the maximum size of a single
1561 * allocation. For example, certain systems may fail to create
1562 * allocations with a size greater than or equal to 4GB. Such a limit is
1563 * implementation-dependent, and if such a failure occurs then the error
1564 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
1566 * We don't use vk_error here because it's not an error so much as an
1567 * indication to the application that the allocation is too large.
1569 if (pAllocateInfo
->allocationSize
> (1ull << 31))
1570 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1572 /* FINISHME: Fail if allocation request exceeds heap size. */
1574 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1575 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1577 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1579 /* The kernel is going to give us whole pages anyway */
1580 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
1582 result
= anv_bo_init_new(&mem
->bo
, device
, alloc_size
);
1583 if (result
!= VK_SUCCESS
)
1586 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1591 *pMem
= anv_device_memory_to_handle(mem
);
1596 vk_free2(&device
->alloc
, pAllocator
, mem
);
1601 void anv_FreeMemory(
1603 VkDeviceMemory _mem
,
1604 const VkAllocationCallbacks
* pAllocator
)
1606 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1607 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1613 anv_UnmapMemory(_device
, _mem
);
1616 anv_gem_munmap(mem
->bo
.map
, mem
->bo
.size
);
1618 if (mem
->bo
.gem_handle
!= 0)
1619 anv_gem_close(device
, mem
->bo
.gem_handle
);
1621 vk_free2(&device
->alloc
, pAllocator
, mem
);
1624 VkResult
anv_MapMemory(
1626 VkDeviceMemory _memory
,
1627 VkDeviceSize offset
,
1629 VkMemoryMapFlags flags
,
1632 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1633 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1640 if (size
== VK_WHOLE_SIZE
)
1641 size
= mem
->bo
.size
- offset
;
1643 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1645 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1646 * assert(size != 0);
1647 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1648 * equal to the size of the memory minus offset
1651 assert(offset
+ size
<= mem
->bo
.size
);
1653 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1654 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1655 * at a time is valid. We could just mmap up front and return an offset
1656 * pointer here, but that may exhaust virtual memory on 32 bit
1659 uint32_t gem_flags
= 0;
1660 if (!device
->info
.has_llc
&& mem
->type_index
== 0)
1661 gem_flags
|= I915_MMAP_WC
;
1663 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1664 uint64_t map_offset
= offset
& ~4095ull;
1665 assert(offset
>= map_offset
);
1666 uint64_t map_size
= (offset
+ size
) - map_offset
;
1668 /* Let's map whole pages */
1669 map_size
= align_u64(map_size
, 4096);
1671 void *map
= anv_gem_mmap(device
, mem
->bo
.gem_handle
,
1672 map_offset
, map_size
, gem_flags
);
1673 if (map
== MAP_FAILED
)
1674 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1677 mem
->map_size
= map_size
;
1679 *ppData
= mem
->map
+ (offset
- map_offset
);
1684 void anv_UnmapMemory(
1686 VkDeviceMemory _memory
)
1688 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1693 anv_gem_munmap(mem
->map
, mem
->map_size
);
1700 clflush_mapped_ranges(struct anv_device
*device
,
1702 const VkMappedMemoryRange
*ranges
)
1704 for (uint32_t i
= 0; i
< count
; i
++) {
1705 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1706 if (ranges
[i
].offset
>= mem
->map_size
)
1709 anv_clflush_range(mem
->map
+ ranges
[i
].offset
,
1710 MIN2(ranges
[i
].size
, mem
->map_size
- ranges
[i
].offset
));
1714 VkResult
anv_FlushMappedMemoryRanges(
1716 uint32_t memoryRangeCount
,
1717 const VkMappedMemoryRange
* pMemoryRanges
)
1719 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1721 if (device
->info
.has_llc
)
1724 /* Make sure the writes we're flushing have landed. */
1725 __builtin_ia32_mfence();
1727 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1732 VkResult
anv_InvalidateMappedMemoryRanges(
1734 uint32_t memoryRangeCount
,
1735 const VkMappedMemoryRange
* pMemoryRanges
)
1737 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1739 if (device
->info
.has_llc
)
1742 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1744 /* Make sure no reads get moved up above the invalidate. */
1745 __builtin_ia32_mfence();
1750 void anv_GetBufferMemoryRequirements(
1753 VkMemoryRequirements
* pMemoryRequirements
)
1755 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1756 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1758 /* The Vulkan spec (git aaed022) says:
1760 * memoryTypeBits is a bitfield and contains one bit set for every
1761 * supported memory type for the resource. The bit `1<<i` is set if and
1762 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1763 * structure for the physical device is supported.
1765 * We support exactly one memory type on LLC, two on non-LLC.
1767 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1769 pMemoryRequirements
->size
= buffer
->size
;
1770 pMemoryRequirements
->alignment
= 16;
1773 void anv_GetImageMemoryRequirements(
1776 VkMemoryRequirements
* pMemoryRequirements
)
1778 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1779 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1781 /* The Vulkan spec (git aaed022) says:
1783 * memoryTypeBits is a bitfield and contains one bit set for every
1784 * supported memory type for the resource. The bit `1<<i` is set if and
1785 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1786 * structure for the physical device is supported.
1788 * We support exactly one memory type on LLC, two on non-LLC.
1790 pMemoryRequirements
->memoryTypeBits
= device
->info
.has_llc
? 1 : 3;
1792 pMemoryRequirements
->size
= image
->size
;
1793 pMemoryRequirements
->alignment
= image
->alignment
;
1796 void anv_GetImageSparseMemoryRequirements(
1799 uint32_t* pSparseMemoryRequirementCount
,
1800 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1802 *pSparseMemoryRequirementCount
= 0;
1805 void anv_GetDeviceMemoryCommitment(
1807 VkDeviceMemory memory
,
1808 VkDeviceSize
* pCommittedMemoryInBytes
)
1810 *pCommittedMemoryInBytes
= 0;
1813 VkResult
anv_BindBufferMemory(
1816 VkDeviceMemory _memory
,
1817 VkDeviceSize memoryOffset
)
1819 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1820 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1823 buffer
->bo
= &mem
->bo
;
1824 buffer
->offset
= memoryOffset
;
1833 VkResult
anv_QueueBindSparse(
1835 uint32_t bindInfoCount
,
1836 const VkBindSparseInfo
* pBindInfo
,
1839 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1840 if (unlikely(queue
->device
->lost
))
1841 return VK_ERROR_DEVICE_LOST
;
1843 return vk_error(VK_ERROR_FEATURE_NOT_PRESENT
);
1846 VkResult
anv_CreateFence(
1848 const VkFenceCreateInfo
* pCreateInfo
,
1849 const VkAllocationCallbacks
* pAllocator
,
1852 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1853 struct anv_bo fence_bo
;
1854 struct anv_fence
*fence
;
1855 struct anv_batch batch
;
1858 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_CREATE_INFO
);
1860 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &fence_bo
, 4096);
1861 if (result
!= VK_SUCCESS
)
1864 /* Fences are small. Just store the CPU data structure in the BO. */
1865 fence
= fence_bo
.map
;
1866 fence
->bo
= fence_bo
;
1868 /* Place the batch after the CPU data but on its own cache line. */
1869 const uint32_t batch_offset
= align_u32(sizeof(*fence
), CACHELINE_SIZE
);
1870 batch
.next
= batch
.start
= fence
->bo
.map
+ batch_offset
;
1871 batch
.end
= fence
->bo
.map
+ fence
->bo
.size
;
1872 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1873 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1875 if (!device
->info
.has_llc
) {
1876 assert(((uintptr_t) batch
.start
& CACHELINE_MASK
) == 0);
1877 assert(batch
.next
- batch
.start
<= CACHELINE_SIZE
);
1878 __builtin_ia32_mfence();
1879 __builtin_ia32_clflush(batch
.start
);
1882 fence
->exec2_objects
[0].handle
= fence
->bo
.gem_handle
;
1883 fence
->exec2_objects
[0].relocation_count
= 0;
1884 fence
->exec2_objects
[0].relocs_ptr
= 0;
1885 fence
->exec2_objects
[0].alignment
= 0;
1886 fence
->exec2_objects
[0].offset
= fence
->bo
.offset
;
1887 fence
->exec2_objects
[0].flags
= 0;
1888 fence
->exec2_objects
[0].rsvd1
= 0;
1889 fence
->exec2_objects
[0].rsvd2
= 0;
1891 fence
->execbuf
.buffers_ptr
= (uintptr_t) fence
->exec2_objects
;
1892 fence
->execbuf
.buffer_count
= 1;
1893 fence
->execbuf
.batch_start_offset
= batch
.start
- fence
->bo
.map
;
1894 fence
->execbuf
.batch_len
= batch
.next
- batch
.start
;
1895 fence
->execbuf
.cliprects_ptr
= 0;
1896 fence
->execbuf
.num_cliprects
= 0;
1897 fence
->execbuf
.DR1
= 0;
1898 fence
->execbuf
.DR4
= 0;
1900 fence
->execbuf
.flags
=
1901 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
1902 fence
->execbuf
.rsvd1
= device
->context_id
;
1903 fence
->execbuf
.rsvd2
= 0;
1905 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
1906 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1908 fence
->state
= ANV_FENCE_STATE_RESET
;
1911 *pFence
= anv_fence_to_handle(fence
);
1916 void anv_DestroyFence(
1919 const VkAllocationCallbacks
* pAllocator
)
1921 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1922 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1927 assert(fence
->bo
.map
== fence
);
1928 anv_bo_pool_free(&device
->batch_bo_pool
, &fence
->bo
);
1931 VkResult
anv_ResetFences(
1933 uint32_t fenceCount
,
1934 const VkFence
* pFences
)
1936 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1937 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1938 fence
->state
= ANV_FENCE_STATE_RESET
;
1944 VkResult
anv_GetFenceStatus(
1948 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1949 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1951 if (unlikely(device
->lost
))
1952 return VK_ERROR_DEVICE_LOST
;
1954 switch (fence
->state
) {
1955 case ANV_FENCE_STATE_RESET
:
1956 /* If it hasn't even been sent off to the GPU yet, it's not ready */
1957 return VK_NOT_READY
;
1959 case ANV_FENCE_STATE_SIGNALED
:
1960 /* It's been signaled, return success */
1963 case ANV_FENCE_STATE_SUBMITTED
: {
1964 VkResult result
= anv_device_bo_busy(device
, &fence
->bo
);
1965 if (result
== VK_SUCCESS
) {
1966 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1973 unreachable("Invalid fence status");
1977 #define NSEC_PER_SEC 1000000000
1978 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1980 VkResult
anv_WaitForFences(
1982 uint32_t fenceCount
,
1983 const VkFence
* pFences
,
1987 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1990 if (unlikely(device
->lost
))
1991 return VK_ERROR_DEVICE_LOST
;
1993 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1994 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1995 * for a couple of kernel releases. Since there's no way to know
1996 * whether or not the kernel we're using is one of the broken ones, the
1997 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1998 * maximum timeout from 584 years to 292 years - likely not a big deal.
2000 int64_t timeout
= MIN2(_timeout
, INT64_MAX
);
2002 VkResult result
= VK_SUCCESS
;
2003 uint32_t pending_fences
= fenceCount
;
2004 while (pending_fences
) {
2006 bool signaled_fences
= false;
2007 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
2008 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
2009 switch (fence
->state
) {
2010 case ANV_FENCE_STATE_RESET
:
2011 /* This fence hasn't been submitted yet, we'll catch it the next
2012 * time around. Yes, this may mean we dead-loop but, short of
2013 * lots of locking and a condition variable, there's not much that
2014 * we can do about that.
2019 case ANV_FENCE_STATE_SIGNALED
:
2020 /* This fence is not pending. If waitAll isn't set, we can return
2021 * early. Otherwise, we have to keep going.
2024 result
= VK_SUCCESS
;
2029 case ANV_FENCE_STATE_SUBMITTED
:
2030 /* These are the fences we really care about. Go ahead and wait
2031 * on it until we hit a timeout.
2033 result
= anv_device_wait(device
, &fence
->bo
, timeout
);
2036 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
2037 signaled_fences
= true;
2051 if (pending_fences
&& !signaled_fences
) {
2052 /* If we've hit this then someone decided to vkWaitForFences before
2053 * they've actually submitted any of them to a queue. This is a
2054 * fairly pessimal case, so it's ok to lock here and use a standard
2055 * pthreads condition variable.
2057 pthread_mutex_lock(&device
->mutex
);
2059 /* It's possible that some of the fences have changed state since the
2060 * last time we checked. Now that we have the lock, check for
2061 * pending fences again and don't wait if it's changed.
2063 uint32_t now_pending_fences
= 0;
2064 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
2065 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
2066 if (fence
->state
== ANV_FENCE_STATE_RESET
)
2067 now_pending_fences
++;
2069 assert(now_pending_fences
<= pending_fences
);
2071 if (now_pending_fences
== pending_fences
) {
2072 struct timespec before
;
2073 clock_gettime(CLOCK_MONOTONIC
, &before
);
2075 uint32_t abs_nsec
= before
.tv_nsec
+ timeout
% NSEC_PER_SEC
;
2076 uint64_t abs_sec
= before
.tv_sec
+ (abs_nsec
/ NSEC_PER_SEC
) +
2077 (timeout
/ NSEC_PER_SEC
);
2078 abs_nsec
%= NSEC_PER_SEC
;
2080 /* Avoid roll-over in tv_sec on 32-bit systems if the user
2081 * provided timeout is UINT64_MAX
2083 struct timespec abstime
;
2084 abstime
.tv_nsec
= abs_nsec
;
2085 abstime
.tv_sec
= MIN2(abs_sec
, INT_TYPE_MAX(abstime
.tv_sec
));
2087 ret
= pthread_cond_timedwait(&device
->queue_submit
,
2088 &device
->mutex
, &abstime
);
2089 assert(ret
!= EINVAL
);
2091 struct timespec after
;
2092 clock_gettime(CLOCK_MONOTONIC
, &after
);
2093 uint64_t time_elapsed
=
2094 ((uint64_t)after
.tv_sec
* NSEC_PER_SEC
+ after
.tv_nsec
) -
2095 ((uint64_t)before
.tv_sec
* NSEC_PER_SEC
+ before
.tv_nsec
);
2097 if (time_elapsed
>= timeout
) {
2098 pthread_mutex_unlock(&device
->mutex
);
2099 result
= VK_TIMEOUT
;
2103 timeout
-= time_elapsed
;
2106 pthread_mutex_unlock(&device
->mutex
);
2111 if (unlikely(device
->lost
))
2112 return VK_ERROR_DEVICE_LOST
;
2117 // Queue semaphore functions
2119 VkResult
anv_CreateSemaphore(
2121 const VkSemaphoreCreateInfo
* pCreateInfo
,
2122 const VkAllocationCallbacks
* pAllocator
,
2123 VkSemaphore
* pSemaphore
)
2125 /* The DRM execbuffer ioctl always execute in-oder, even between different
2126 * rings. As such, there's nothing to do for the user space semaphore.
2129 *pSemaphore
= (VkSemaphore
)1;
2134 void anv_DestroySemaphore(
2136 VkSemaphore semaphore
,
2137 const VkAllocationCallbacks
* pAllocator
)
2143 VkResult
anv_CreateEvent(
2145 const VkEventCreateInfo
* pCreateInfo
,
2146 const VkAllocationCallbacks
* pAllocator
,
2149 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2150 struct anv_state state
;
2151 struct anv_event
*event
;
2153 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
2155 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
2158 event
->state
= state
;
2159 event
->semaphore
= VK_EVENT_RESET
;
2161 if (!device
->info
.has_llc
) {
2162 /* Make sure the writes we're flushing have landed. */
2163 __builtin_ia32_mfence();
2164 __builtin_ia32_clflush(event
);
2167 *pEvent
= anv_event_to_handle(event
);
2172 void anv_DestroyEvent(
2175 const VkAllocationCallbacks
* pAllocator
)
2177 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2178 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2183 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
2186 VkResult
anv_GetEventStatus(
2190 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2191 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2193 if (unlikely(device
->lost
))
2194 return VK_ERROR_DEVICE_LOST
;
2196 if (!device
->info
.has_llc
) {
2197 /* Invalidate read cache before reading event written by GPU. */
2198 __builtin_ia32_clflush(event
);
2199 __builtin_ia32_mfence();
2203 return event
->semaphore
;
2206 VkResult
anv_SetEvent(
2210 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2211 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2213 event
->semaphore
= VK_EVENT_SET
;
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
);
2224 VkResult
anv_ResetEvent(
2228 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2229 ANV_FROM_HANDLE(anv_event
, event
, _event
);
2231 event
->semaphore
= VK_EVENT_RESET
;
2233 if (!device
->info
.has_llc
) {
2234 /* Make sure the writes we're flushing have landed. */
2235 __builtin_ia32_mfence();
2236 __builtin_ia32_clflush(event
);
2244 VkResult
anv_CreateBuffer(
2246 const VkBufferCreateInfo
* pCreateInfo
,
2247 const VkAllocationCallbacks
* pAllocator
,
2250 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2251 struct anv_buffer
*buffer
;
2253 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2255 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2256 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2258 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2260 buffer
->size
= pCreateInfo
->size
;
2261 buffer
->usage
= pCreateInfo
->usage
;
2265 *pBuffer
= anv_buffer_to_handle(buffer
);
2270 void anv_DestroyBuffer(
2273 const VkAllocationCallbacks
* pAllocator
)
2275 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2276 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
2281 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2285 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
2286 enum isl_format format
,
2287 uint32_t offset
, uint32_t range
, uint32_t stride
)
2289 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
2291 .mocs
= device
->default_mocs
,
2296 anv_state_flush(device
, state
);
2299 void anv_DestroySampler(
2302 const VkAllocationCallbacks
* pAllocator
)
2304 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2305 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
2310 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2313 VkResult
anv_CreateFramebuffer(
2315 const VkFramebufferCreateInfo
* pCreateInfo
,
2316 const VkAllocationCallbacks
* pAllocator
,
2317 VkFramebuffer
* pFramebuffer
)
2319 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2320 struct anv_framebuffer
*framebuffer
;
2322 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2324 size_t size
= sizeof(*framebuffer
) +
2325 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2326 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2327 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2328 if (framebuffer
== NULL
)
2329 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2331 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2332 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2333 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2334 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2337 framebuffer
->width
= pCreateInfo
->width
;
2338 framebuffer
->height
= pCreateInfo
->height
;
2339 framebuffer
->layers
= pCreateInfo
->layers
;
2341 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2346 void anv_DestroyFramebuffer(
2349 const VkAllocationCallbacks
* pAllocator
)
2351 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2352 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2357 vk_free2(&device
->alloc
, pAllocator
, fb
);
2360 /* vk_icd.h does not declare this function, so we declare it here to
2361 * suppress Wmissing-prototypes.
2363 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2364 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
);
2366 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2367 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion
)
2369 /* For the full details on loader interface versioning, see
2370 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2371 * What follows is a condensed summary, to help you navigate the large and
2372 * confusing official doc.
2374 * - Loader interface v0 is incompatible with later versions. We don't
2377 * - In loader interface v1:
2378 * - The first ICD entrypoint called by the loader is
2379 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2381 * - The ICD must statically expose no other Vulkan symbol unless it is
2382 * linked with -Bsymbolic.
2383 * - Each dispatchable Vulkan handle created by the ICD must be
2384 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2385 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2386 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2387 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2388 * such loader-managed surfaces.
2390 * - Loader interface v2 differs from v1 in:
2391 * - The first ICD entrypoint called by the loader is
2392 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2393 * statically expose this entrypoint.
2395 * - Loader interface v3 differs from v2 in:
2396 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2397 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2398 * because the loader no longer does so.
2400 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);