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
30 #include "anv_private.h"
31 #include "anv_timestamp.h"
32 #include "util/strtod.h"
33 #include "util/debug.h"
35 #include "genxml/gen7_pack.h"
37 struct anv_dispatch_table dtable
;
40 compiler_debug_log(void *data
, const char *fmt
, ...)
44 compiler_perf_log(void *data
, const char *fmt
, ...)
49 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
50 vfprintf(stderr
, fmt
, args
);
56 anv_physical_device_init(struct anv_physical_device
*device
,
57 struct anv_instance
*instance
,
63 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
65 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
67 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
68 device
->instance
= instance
;
70 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
71 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
73 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
74 if (!device
->chipset_id
) {
75 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
79 device
->name
= gen_get_device_name(device
->chipset_id
);
80 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
81 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
85 if (device
->info
.is_haswell
) {
86 fprintf(stderr
, "WARNING: Haswell Vulkan support is incomplete\n");
87 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
88 fprintf(stderr
, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
89 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
90 fprintf(stderr
, "WARNING: Bay Trail Vulkan support is incomplete\n");
91 } else if (device
->info
.gen
>= 8) {
92 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
93 * supported as anything */
95 result
= vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
96 "Vulkan not yet supported on %s", device
->name
);
100 device
->cmd_parser_version
= -1;
101 if (device
->info
.gen
== 7) {
102 device
->cmd_parser_version
=
103 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
104 if (device
->cmd_parser_version
== -1) {
105 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
106 "failed to get command parser version");
111 if (anv_gem_get_aperture(fd
, &device
->aperture_size
) == -1) {
112 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
113 "failed to get aperture size: %m");
117 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
118 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
119 "kernel missing gem wait");
123 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
124 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
125 "kernel missing execbuf2");
129 if (!device
->info
.has_llc
&&
130 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
131 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
132 "kernel missing wc mmap");
136 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
138 /* GENs prior to 8 do not support EU/Subslice info */
139 if (device
->info
.gen
>= 8) {
140 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
141 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
143 /* Without this information, we cannot get the right Braswell
144 * brandstrings, and we have to use conservative numbers for GPGPU on
145 * many platforms, but otherwise, things will just work.
147 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
148 fprintf(stderr
, "WARNING: Kernel 4.1 required to properly"
149 " query GPU properties.\n");
151 } else if (device
->info
.gen
== 7) {
152 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
155 if (device
->info
.is_cherryview
&&
156 device
->subslice_total
> 0 && device
->eu_total
> 0) {
157 /* Logical CS threads = EUs per subslice * 7 threads per EU */
158 uint32_t max_cs_threads
= device
->eu_total
/ device
->subslice_total
* 7;
160 /* Fuse configurations may give more threads than expected, never less. */
161 if (max_cs_threads
> device
->info
.max_cs_threads
)
162 device
->info
.max_cs_threads
= max_cs_threads
;
167 brw_process_intel_debug_variable();
169 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
170 if (device
->compiler
== NULL
) {
171 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
174 device
->compiler
->shader_debug_log
= compiler_debug_log
;
175 device
->compiler
->shader_perf_log
= compiler_perf_log
;
177 result
= anv_init_wsi(device
);
178 if (result
!= VK_SUCCESS
)
181 /* XXX: Actually detect bit6 swizzling */
182 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
192 anv_physical_device_finish(struct anv_physical_device
*device
)
194 anv_finish_wsi(device
);
195 ralloc_free(device
->compiler
);
198 static const VkExtensionProperties global_extensions
[] = {
200 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
203 #ifdef VK_USE_PLATFORM_XCB_KHR
205 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
209 #ifdef VK_USE_PLATFORM_XLIB_KHR
211 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
215 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
217 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
223 static const VkExtensionProperties device_extensions
[] = {
225 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
231 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
232 VkSystemAllocationScope allocationScope
)
238 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
239 size_t align
, VkSystemAllocationScope allocationScope
)
241 return realloc(pOriginal
, size
);
245 default_free_func(void *pUserData
, void *pMemory
)
250 static const VkAllocationCallbacks default_alloc
= {
252 .pfnAllocation
= default_alloc_func
,
253 .pfnReallocation
= default_realloc_func
,
254 .pfnFree
= default_free_func
,
257 VkResult
anv_CreateInstance(
258 const VkInstanceCreateInfo
* pCreateInfo
,
259 const VkAllocationCallbacks
* pAllocator
,
260 VkInstance
* pInstance
)
262 struct anv_instance
*instance
;
264 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
266 uint32_t client_version
;
267 if (pCreateInfo
->pApplicationInfo
&&
268 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
269 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
271 client_version
= VK_MAKE_VERSION(1, 0, 0);
274 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
275 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
276 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
277 "Client requested version %d.%d.%d",
278 VK_VERSION_MAJOR(client_version
),
279 VK_VERSION_MINOR(client_version
),
280 VK_VERSION_PATCH(client_version
));
283 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
285 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
286 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
287 global_extensions
[j
].extensionName
) == 0) {
293 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
296 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
297 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
299 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
301 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
304 instance
->alloc
= *pAllocator
;
306 instance
->alloc
= default_alloc
;
308 instance
->apiVersion
= client_version
;
309 instance
->physicalDeviceCount
= -1;
313 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
315 *pInstance
= anv_instance_to_handle(instance
);
320 void anv_DestroyInstance(
321 VkInstance _instance
,
322 const VkAllocationCallbacks
* pAllocator
)
324 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
326 if (instance
->physicalDeviceCount
> 0) {
327 /* We support at most one physical device. */
328 assert(instance
->physicalDeviceCount
== 1);
329 anv_physical_device_finish(&instance
->physicalDevice
);
332 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
336 vk_free(&instance
->alloc
, instance
);
339 VkResult
anv_EnumeratePhysicalDevices(
340 VkInstance _instance
,
341 uint32_t* pPhysicalDeviceCount
,
342 VkPhysicalDevice
* pPhysicalDevices
)
344 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
347 if (instance
->physicalDeviceCount
< 0) {
349 for (unsigned i
= 0; i
< 8; i
++) {
350 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
351 result
= anv_physical_device_init(&instance
->physicalDevice
,
353 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
357 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
358 instance
->physicalDeviceCount
= 0;
359 } else if (result
== VK_SUCCESS
) {
360 instance
->physicalDeviceCount
= 1;
366 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
367 * otherwise it's an inout parameter.
369 * The Vulkan spec (git aaed022) says:
371 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
372 * that is initialized with the number of devices the application is
373 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
374 * an array of at least this many VkPhysicalDevice handles [...].
376 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
377 * overwrites the contents of the variable pointed to by
378 * pPhysicalDeviceCount with the number of physical devices in in the
379 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
380 * pPhysicalDeviceCount with the number of physical handles written to
383 if (!pPhysicalDevices
) {
384 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
385 } else if (*pPhysicalDeviceCount
>= 1) {
386 pPhysicalDevices
[0] = anv_physical_device_to_handle(&instance
->physicalDevice
);
387 *pPhysicalDeviceCount
= 1;
388 } else if (*pPhysicalDeviceCount
< instance
->physicalDeviceCount
) {
389 return VK_INCOMPLETE
;
391 *pPhysicalDeviceCount
= 0;
397 void anv_GetPhysicalDeviceFeatures(
398 VkPhysicalDevice physicalDevice
,
399 VkPhysicalDeviceFeatures
* pFeatures
)
401 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
403 *pFeatures
= (VkPhysicalDeviceFeatures
) {
404 .robustBufferAccess
= true,
405 .fullDrawIndexUint32
= true,
406 .imageCubeArray
= false,
407 .independentBlend
= true,
408 .geometryShader
= true,
409 .tessellationShader
= false,
410 .sampleRateShading
= true,
411 .dualSrcBlend
= true,
413 .multiDrawIndirect
= false,
414 .drawIndirectFirstInstance
= false,
416 .depthBiasClamp
= false,
417 .fillModeNonSolid
= true,
418 .depthBounds
= false,
422 .multiViewport
= true,
423 .samplerAnisotropy
= true,
424 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
425 pdevice
->info
.is_baytrail
,
426 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
427 .textureCompressionBC
= true,
428 .occlusionQueryPrecise
= true,
429 .pipelineStatisticsQuery
= false,
430 .fragmentStoresAndAtomics
= true,
431 .shaderTessellationAndGeometryPointSize
= true,
432 .shaderImageGatherExtended
= false,
433 .shaderStorageImageExtendedFormats
= false,
434 .shaderStorageImageMultisample
= false,
435 .shaderUniformBufferArrayDynamicIndexing
= true,
436 .shaderSampledImageArrayDynamicIndexing
= true,
437 .shaderStorageBufferArrayDynamicIndexing
= true,
438 .shaderStorageImageArrayDynamicIndexing
= true,
439 .shaderStorageImageReadWithoutFormat
= false,
440 .shaderStorageImageWriteWithoutFormat
= true,
441 .shaderClipDistance
= false,
442 .shaderCullDistance
= false,
443 .shaderFloat64
= false,
444 .shaderInt64
= false,
445 .shaderInt16
= false,
447 .variableMultisampleRate
= false,
448 .inheritedQueries
= false,
451 /* We can't do image stores in vec4 shaders */
452 pFeatures
->vertexPipelineStoresAndAtomics
=
453 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
454 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
458 anv_device_get_cache_uuid(void *uuid
)
460 memset(uuid
, 0, VK_UUID_SIZE
);
461 snprintf(uuid
, VK_UUID_SIZE
, "anv-%s", ANV_TIMESTAMP
);
464 void anv_GetPhysicalDeviceProperties(
465 VkPhysicalDevice physicalDevice
,
466 VkPhysicalDeviceProperties
* pProperties
)
468 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
469 const struct gen_device_info
*devinfo
= &pdevice
->info
;
471 const float time_stamp_base
= devinfo
->gen
>= 9 ? 83.333 : 80.0;
473 /* See assertions made when programming the buffer surface state. */
474 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
475 (1ul << 30) : (1ul << 27);
477 VkSampleCountFlags sample_counts
=
478 isl_device_get_sample_counts(&pdevice
->isl_dev
);
480 VkPhysicalDeviceLimits limits
= {
481 .maxImageDimension1D
= (1 << 14),
482 .maxImageDimension2D
= (1 << 14),
483 .maxImageDimension3D
= (1 << 11),
484 .maxImageDimensionCube
= (1 << 14),
485 .maxImageArrayLayers
= (1 << 11),
486 .maxTexelBufferElements
= 128 * 1024 * 1024,
487 .maxUniformBufferRange
= (1ul << 27),
488 .maxStorageBufferRange
= max_raw_buffer_sz
,
489 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
490 .maxMemoryAllocationCount
= UINT32_MAX
,
491 .maxSamplerAllocationCount
= 64 * 1024,
492 .bufferImageGranularity
= 64, /* A cache line */
493 .sparseAddressSpaceSize
= 0,
494 .maxBoundDescriptorSets
= MAX_SETS
,
495 .maxPerStageDescriptorSamplers
= 64,
496 .maxPerStageDescriptorUniformBuffers
= 64,
497 .maxPerStageDescriptorStorageBuffers
= 64,
498 .maxPerStageDescriptorSampledImages
= 64,
499 .maxPerStageDescriptorStorageImages
= 64,
500 .maxPerStageDescriptorInputAttachments
= 64,
501 .maxPerStageResources
= 128,
502 .maxDescriptorSetSamplers
= 256,
503 .maxDescriptorSetUniformBuffers
= 256,
504 .maxDescriptorSetUniformBuffersDynamic
= 256,
505 .maxDescriptorSetStorageBuffers
= 256,
506 .maxDescriptorSetStorageBuffersDynamic
= 256,
507 .maxDescriptorSetSampledImages
= 256,
508 .maxDescriptorSetStorageImages
= 256,
509 .maxDescriptorSetInputAttachments
= 256,
510 .maxVertexInputAttributes
= 32,
511 .maxVertexInputBindings
= 32,
512 .maxVertexInputAttributeOffset
= 2047,
513 .maxVertexInputBindingStride
= 2048,
514 .maxVertexOutputComponents
= 128,
515 .maxTessellationGenerationLevel
= 0,
516 .maxTessellationPatchSize
= 0,
517 .maxTessellationControlPerVertexInputComponents
= 0,
518 .maxTessellationControlPerVertexOutputComponents
= 0,
519 .maxTessellationControlPerPatchOutputComponents
= 0,
520 .maxTessellationControlTotalOutputComponents
= 0,
521 .maxTessellationEvaluationInputComponents
= 0,
522 .maxTessellationEvaluationOutputComponents
= 0,
523 .maxGeometryShaderInvocations
= 32,
524 .maxGeometryInputComponents
= 64,
525 .maxGeometryOutputComponents
= 128,
526 .maxGeometryOutputVertices
= 256,
527 .maxGeometryTotalOutputComponents
= 1024,
528 .maxFragmentInputComponents
= 128,
529 .maxFragmentOutputAttachments
= 8,
530 .maxFragmentDualSrcAttachments
= 2,
531 .maxFragmentCombinedOutputResources
= 8,
532 .maxComputeSharedMemorySize
= 32768,
533 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
534 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
535 .maxComputeWorkGroupSize
= {
536 16 * devinfo
->max_cs_threads
,
537 16 * devinfo
->max_cs_threads
,
538 16 * devinfo
->max_cs_threads
,
540 .subPixelPrecisionBits
= 4 /* FIXME */,
541 .subTexelPrecisionBits
= 4 /* FIXME */,
542 .mipmapPrecisionBits
= 4 /* FIXME */,
543 .maxDrawIndexedIndexValue
= UINT32_MAX
,
544 .maxDrawIndirectCount
= UINT32_MAX
,
545 .maxSamplerLodBias
= 16,
546 .maxSamplerAnisotropy
= 16,
547 .maxViewports
= MAX_VIEWPORTS
,
548 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
549 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
550 .viewportSubPixelBits
= 13, /* We take a float? */
551 .minMemoryMapAlignment
= 4096, /* A page */
552 .minTexelBufferOffsetAlignment
= 1,
553 .minUniformBufferOffsetAlignment
= 1,
554 .minStorageBufferOffsetAlignment
= 1,
555 .minTexelOffset
= -8,
557 .minTexelGatherOffset
= -8,
558 .maxTexelGatherOffset
= 7,
559 .minInterpolationOffset
= -0.5,
560 .maxInterpolationOffset
= 0.4375,
561 .subPixelInterpolationOffsetBits
= 4,
562 .maxFramebufferWidth
= (1 << 14),
563 .maxFramebufferHeight
= (1 << 14),
564 .maxFramebufferLayers
= (1 << 10),
565 .framebufferColorSampleCounts
= sample_counts
,
566 .framebufferDepthSampleCounts
= sample_counts
,
567 .framebufferStencilSampleCounts
= sample_counts
,
568 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
569 .maxColorAttachments
= MAX_RTS
,
570 .sampledImageColorSampleCounts
= sample_counts
,
571 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
572 .sampledImageDepthSampleCounts
= sample_counts
,
573 .sampledImageStencilSampleCounts
= sample_counts
,
574 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
575 .maxSampleMaskWords
= 1,
576 .timestampComputeAndGraphics
= false,
577 .timestampPeriod
= time_stamp_base
,
578 .maxClipDistances
= 0 /* FIXME */,
579 .maxCullDistances
= 0 /* FIXME */,
580 .maxCombinedClipAndCullDistances
= 0 /* FIXME */,
581 .discreteQueuePriorities
= 1,
582 .pointSizeRange
= { 0.125, 255.875 },
583 .lineWidthRange
= { 0.0, 7.9921875 },
584 .pointSizeGranularity
= (1.0 / 8.0),
585 .lineWidthGranularity
= (1.0 / 128.0),
586 .strictLines
= false, /* FINISHME */
587 .standardSampleLocations
= true,
588 .optimalBufferCopyOffsetAlignment
= 128,
589 .optimalBufferCopyRowPitchAlignment
= 128,
590 .nonCoherentAtomSize
= 64,
593 *pProperties
= (VkPhysicalDeviceProperties
) {
594 .apiVersion
= VK_MAKE_VERSION(1, 0, 5),
597 .deviceID
= pdevice
->chipset_id
,
598 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
600 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
603 strcpy(pProperties
->deviceName
, pdevice
->name
);
604 anv_device_get_cache_uuid(pProperties
->pipelineCacheUUID
);
607 void anv_GetPhysicalDeviceQueueFamilyProperties(
608 VkPhysicalDevice physicalDevice
,
610 VkQueueFamilyProperties
* pQueueFamilyProperties
)
612 if (pQueueFamilyProperties
== NULL
) {
617 assert(*pCount
>= 1);
619 *pQueueFamilyProperties
= (VkQueueFamilyProperties
) {
620 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
621 VK_QUEUE_COMPUTE_BIT
|
622 VK_QUEUE_TRANSFER_BIT
,
624 .timestampValidBits
= 36, /* XXX: Real value here */
625 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
629 void anv_GetPhysicalDeviceMemoryProperties(
630 VkPhysicalDevice physicalDevice
,
631 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
633 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
634 VkDeviceSize heap_size
;
636 /* Reserve some wiggle room for the driver by exposing only 75% of the
637 * aperture to the heap.
639 heap_size
= 3 * physical_device
->aperture_size
/ 4;
641 if (physical_device
->info
.has_llc
) {
642 /* Big core GPUs share LLC with the CPU and thus one memory type can be
643 * both cached and coherent at the same time.
645 pMemoryProperties
->memoryTypeCount
= 1;
646 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
647 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
648 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
649 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
650 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
654 /* The spec requires that we expose a host-visible, coherent memory
655 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
656 * to give the application a choice between cached, but not coherent and
657 * coherent but uncached (WC though).
659 pMemoryProperties
->memoryTypeCount
= 2;
660 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
661 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
662 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
663 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
666 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
667 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
668 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
669 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
674 pMemoryProperties
->memoryHeapCount
= 1;
675 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
677 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
681 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
685 return anv_lookup_entrypoint(NULL
, pName
);
688 /* With version 1+ of the loader interface the ICD should expose
689 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
692 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
697 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
701 return anv_GetInstanceProcAddr(instance
, pName
);
704 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
708 ANV_FROM_HANDLE(anv_device
, device
, _device
);
709 return anv_lookup_entrypoint(&device
->info
, pName
);
713 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
715 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
716 queue
->device
= device
;
717 queue
->pool
= &device
->surface_state_pool
;
723 anv_queue_finish(struct anv_queue
*queue
)
727 static struct anv_state
728 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
730 struct anv_state state
;
732 state
= anv_state_pool_alloc(pool
, size
, align
);
733 memcpy(state
.map
, p
, size
);
735 if (!pool
->block_pool
->device
->info
.has_llc
)
736 anv_state_clflush(state
);
741 struct gen8_border_color
{
746 /* Pad out to 64 bytes */
751 anv_device_init_border_colors(struct anv_device
*device
)
753 static const struct gen8_border_color border_colors
[] = {
754 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
755 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
756 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
757 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
758 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
759 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
762 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
763 sizeof(border_colors
), 64,
768 anv_device_submit_simple_batch(struct anv_device
*device
,
769 struct anv_batch
*batch
)
771 struct drm_i915_gem_execbuffer2 execbuf
;
772 struct drm_i915_gem_exec_object2 exec2_objects
[1];
773 struct anv_bo bo
, *exec_bos
[1];
774 VkResult result
= VK_SUCCESS
;
779 /* Kernel driver requires 8 byte aligned batch length */
780 size
= align_u32(batch
->next
- batch
->start
, 8);
781 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &bo
, size
);
782 if (result
!= VK_SUCCESS
)
785 memcpy(bo
.map
, batch
->start
, size
);
786 if (!device
->info
.has_llc
)
787 anv_clflush_range(bo
.map
, size
);
790 exec2_objects
[0].handle
= bo
.gem_handle
;
791 exec2_objects
[0].relocation_count
= 0;
792 exec2_objects
[0].relocs_ptr
= 0;
793 exec2_objects
[0].alignment
= 0;
794 exec2_objects
[0].offset
= bo
.offset
;
795 exec2_objects
[0].flags
= 0;
796 exec2_objects
[0].rsvd1
= 0;
797 exec2_objects
[0].rsvd2
= 0;
799 execbuf
.buffers_ptr
= (uintptr_t) exec2_objects
;
800 execbuf
.buffer_count
= 1;
801 execbuf
.batch_start_offset
= 0;
802 execbuf
.batch_len
= size
;
803 execbuf
.cliprects_ptr
= 0;
804 execbuf
.num_cliprects
= 0;
809 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
810 execbuf
.rsvd1
= device
->context_id
;
813 result
= anv_device_execbuf(device
, &execbuf
, exec_bos
);
814 if (result
!= VK_SUCCESS
)
818 ret
= anv_gem_wait(device
, bo
.gem_handle
, &timeout
);
820 /* We don't know the real error. */
821 result
= vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
826 anv_bo_pool_free(&device
->batch_bo_pool
, &bo
);
831 VkResult
anv_CreateDevice(
832 VkPhysicalDevice physicalDevice
,
833 const VkDeviceCreateInfo
* pCreateInfo
,
834 const VkAllocationCallbacks
* pAllocator
,
837 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
839 struct anv_device
*device
;
841 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
843 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
845 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
846 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
847 device_extensions
[j
].extensionName
) == 0) {
853 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
856 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
858 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
860 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
862 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
863 device
->instance
= physical_device
->instance
;
864 device
->chipset_id
= physical_device
->chipset_id
;
867 device
->alloc
= *pAllocator
;
869 device
->alloc
= physical_device
->instance
->alloc
;
871 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
872 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
873 if (device
->fd
== -1) {
874 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
878 device
->context_id
= anv_gem_create_context(device
);
879 if (device
->context_id
== -1) {
880 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
884 device
->info
= physical_device
->info
;
885 device
->isl_dev
= physical_device
->isl_dev
;
887 /* On Broadwell and later, we can use batch chaining to more efficiently
888 * implement growing command buffers. Prior to Haswell, the kernel
889 * command parser gets in the way and we have to fall back to growing
892 device
->can_chain_batches
= device
->info
.gen
>= 8;
894 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
895 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
897 pthread_mutex_init(&device
->mutex
, NULL
);
899 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
901 anv_block_pool_init(&device
->dynamic_state_block_pool
, device
, 16384);
903 anv_state_pool_init(&device
->dynamic_state_pool
,
904 &device
->dynamic_state_block_pool
);
906 anv_block_pool_init(&device
->instruction_block_pool
, device
, 128 * 1024);
907 anv_state_pool_init(&device
->instruction_state_pool
,
908 &device
->instruction_block_pool
);
910 anv_block_pool_init(&device
->surface_state_block_pool
, device
, 4096);
912 anv_state_pool_init(&device
->surface_state_pool
,
913 &device
->surface_state_block_pool
);
915 anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
917 anv_scratch_pool_init(device
, &device
->scratch_pool
);
919 anv_queue_init(device
, &device
->queue
);
921 switch (device
->info
.gen
) {
923 if (!device
->info
.is_haswell
)
924 result
= gen7_init_device_state(device
);
926 result
= gen75_init_device_state(device
);
929 result
= gen8_init_device_state(device
);
932 result
= gen9_init_device_state(device
);
935 /* Shouldn't get here as we don't create physical devices for any other
937 unreachable("unhandled gen");
939 if (result
!= VK_SUCCESS
)
942 anv_device_init_blorp(device
);
944 anv_device_init_border_colors(device
);
946 *pDevice
= anv_device_to_handle(device
);
953 vk_free(&device
->alloc
, device
);
958 void anv_DestroyDevice(
960 const VkAllocationCallbacks
* pAllocator
)
962 ANV_FROM_HANDLE(anv_device
, device
, _device
);
964 anv_queue_finish(&device
->queue
);
966 anv_device_finish_blorp(device
);
969 /* We only need to free these to prevent valgrind errors. The backing
970 * BO will go away in a couple of lines so we don't actually leak.
972 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
975 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
976 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
978 anv_bo_pool_finish(&device
->batch_bo_pool
);
979 anv_state_pool_finish(&device
->dynamic_state_pool
);
980 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
981 anv_state_pool_finish(&device
->instruction_state_pool
);
982 anv_block_pool_finish(&device
->instruction_block_pool
);
983 anv_state_pool_finish(&device
->surface_state_pool
);
984 anv_block_pool_finish(&device
->surface_state_block_pool
);
985 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
989 pthread_mutex_destroy(&device
->mutex
);
991 vk_free(&device
->alloc
, device
);
994 VkResult
anv_EnumerateInstanceExtensionProperties(
995 const char* pLayerName
,
996 uint32_t* pPropertyCount
,
997 VkExtensionProperties
* pProperties
)
999 if (pProperties
== NULL
) {
1000 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1004 assert(*pPropertyCount
>= ARRAY_SIZE(global_extensions
));
1006 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1007 memcpy(pProperties
, global_extensions
, sizeof(global_extensions
));
1012 VkResult
anv_EnumerateDeviceExtensionProperties(
1013 VkPhysicalDevice physicalDevice
,
1014 const char* pLayerName
,
1015 uint32_t* pPropertyCount
,
1016 VkExtensionProperties
* pProperties
)
1018 if (pProperties
== NULL
) {
1019 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1023 assert(*pPropertyCount
>= ARRAY_SIZE(device_extensions
));
1025 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1026 memcpy(pProperties
, device_extensions
, sizeof(device_extensions
));
1031 VkResult
anv_EnumerateInstanceLayerProperties(
1032 uint32_t* pPropertyCount
,
1033 VkLayerProperties
* pProperties
)
1035 if (pProperties
== NULL
) {
1036 *pPropertyCount
= 0;
1040 /* None supported at this time */
1041 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1044 VkResult
anv_EnumerateDeviceLayerProperties(
1045 VkPhysicalDevice physicalDevice
,
1046 uint32_t* pPropertyCount
,
1047 VkLayerProperties
* pProperties
)
1049 if (pProperties
== NULL
) {
1050 *pPropertyCount
= 0;
1054 /* None supported at this time */
1055 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1058 void anv_GetDeviceQueue(
1060 uint32_t queueNodeIndex
,
1061 uint32_t queueIndex
,
1064 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1066 assert(queueIndex
== 0);
1068 *pQueue
= anv_queue_to_handle(&device
->queue
);
1072 anv_device_execbuf(struct anv_device
*device
,
1073 struct drm_i915_gem_execbuffer2
*execbuf
,
1074 struct anv_bo
**execbuf_bos
)
1076 int ret
= anv_gem_execbuffer(device
, execbuf
);
1078 /* We don't know the real error. */
1079 return vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
1082 struct drm_i915_gem_exec_object2
*objects
= (void *)execbuf
->buffers_ptr
;
1083 for (uint32_t k
= 0; k
< execbuf
->buffer_count
; k
++)
1084 execbuf_bos
[k
]->offset
= objects
[k
].offset
;
1089 VkResult
anv_QueueSubmit(
1091 uint32_t submitCount
,
1092 const VkSubmitInfo
* pSubmits
,
1095 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1096 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1097 struct anv_device
*device
= queue
->device
;
1098 VkResult result
= VK_SUCCESS
;
1100 /* We lock around QueueSubmit for three main reasons:
1102 * 1) When a block pool is resized, we create a new gem handle with a
1103 * different size and, in the case of surface states, possibly a
1104 * different center offset but we re-use the same anv_bo struct when
1105 * we do so. If this happens in the middle of setting up an execbuf,
1106 * we could end up with our list of BOs out of sync with our list of
1109 * 2) The algorithm we use for building the list of unique buffers isn't
1110 * thread-safe. While the client is supposed to syncronize around
1111 * QueueSubmit, this would be extremely difficult to debug if it ever
1112 * came up in the wild due to a broken app. It's better to play it
1113 * safe and just lock around QueueSubmit.
1115 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
1116 * userspace. Due to the fact that the surface state buffer is shared
1117 * between batches, we can't afford to have that happen from multiple
1118 * threads at the same time. Even though the user is supposed to
1119 * ensure this doesn't happen, we play it safe as in (2) above.
1121 * Since the only other things that ever take the device lock such as block
1122 * pool resize only rarely happen, this will almost never be contended so
1123 * taking a lock isn't really an expensive operation in this case.
1125 pthread_mutex_lock(&device
->mutex
);
1127 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1128 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1129 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
,
1130 pSubmits
[i
].pCommandBuffers
[j
]);
1131 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1133 result
= anv_cmd_buffer_execbuf(device
, cmd_buffer
);
1134 if (result
!= VK_SUCCESS
)
1140 struct anv_bo
*fence_bo
= &fence
->bo
;
1141 result
= anv_device_execbuf(device
, &fence
->execbuf
, &fence_bo
);
1142 if (result
!= VK_SUCCESS
)
1147 pthread_mutex_unlock(&device
->mutex
);
1152 VkResult
anv_QueueWaitIdle(
1155 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1157 return anv_DeviceWaitIdle(anv_device_to_handle(queue
->device
));
1160 VkResult
anv_DeviceWaitIdle(
1163 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1164 struct anv_batch batch
;
1167 batch
.start
= batch
.next
= cmds
;
1168 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1170 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1171 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1173 return anv_device_submit_simple_batch(device
, &batch
);
1177 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1179 uint32_t gem_handle
= anv_gem_create(device
, size
);
1181 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1183 anv_bo_init(bo
, gem_handle
, size
);
1188 VkResult
anv_AllocateMemory(
1190 const VkMemoryAllocateInfo
* pAllocateInfo
,
1191 const VkAllocationCallbacks
* pAllocator
,
1192 VkDeviceMemory
* pMem
)
1194 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1195 struct anv_device_memory
*mem
;
1198 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1200 if (pAllocateInfo
->allocationSize
== 0) {
1201 /* Apparently, this is allowed */
1202 *pMem
= VK_NULL_HANDLE
;
1206 /* We support exactly one memory heap. */
1207 assert(pAllocateInfo
->memoryTypeIndex
== 0 ||
1208 (!device
->info
.has_llc
&& pAllocateInfo
->memoryTypeIndex
< 2));
1210 /* FINISHME: Fail if allocation request exceeds heap size. */
1212 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1213 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1215 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1217 /* The kernel is going to give us whole pages anyway */
1218 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
1220 result
= anv_bo_init_new(&mem
->bo
, device
, alloc_size
);
1221 if (result
!= VK_SUCCESS
)
1224 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1226 *pMem
= anv_device_memory_to_handle(mem
);
1231 vk_free2(&device
->alloc
, pAllocator
, mem
);
1236 void anv_FreeMemory(
1238 VkDeviceMemory _mem
,
1239 const VkAllocationCallbacks
* pAllocator
)
1241 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1242 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1248 anv_gem_munmap(mem
->bo
.map
, mem
->bo
.size
);
1250 if (mem
->bo
.gem_handle
!= 0)
1251 anv_gem_close(device
, mem
->bo
.gem_handle
);
1253 vk_free2(&device
->alloc
, pAllocator
, mem
);
1256 VkResult
anv_MapMemory(
1258 VkDeviceMemory _memory
,
1259 VkDeviceSize offset
,
1261 VkMemoryMapFlags flags
,
1264 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1265 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1272 if (size
== VK_WHOLE_SIZE
)
1273 size
= mem
->bo
.size
- offset
;
1275 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1276 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1277 * at a time is valid. We could just mmap up front and return an offset
1278 * pointer here, but that may exhaust virtual memory on 32 bit
1281 uint32_t gem_flags
= 0;
1282 if (!device
->info
.has_llc
&& mem
->type_index
== 0)
1283 gem_flags
|= I915_MMAP_WC
;
1285 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1286 uint64_t map_offset
= offset
& ~4095ull;
1287 assert(offset
>= map_offset
);
1288 uint64_t map_size
= (offset
+ size
) - map_offset
;
1290 /* Let's map whole pages */
1291 map_size
= align_u64(map_size
, 4096);
1293 mem
->map
= anv_gem_mmap(device
, mem
->bo
.gem_handle
,
1294 map_offset
, map_size
, gem_flags
);
1295 mem
->map_size
= map_size
;
1297 *ppData
= mem
->map
+ (offset
- map_offset
);
1302 void anv_UnmapMemory(
1304 VkDeviceMemory _memory
)
1306 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1311 anv_gem_munmap(mem
->map
, mem
->map_size
);
1315 clflush_mapped_ranges(struct anv_device
*device
,
1317 const VkMappedMemoryRange
*ranges
)
1319 for (uint32_t i
= 0; i
< count
; i
++) {
1320 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1321 void *p
= mem
->map
+ (ranges
[i
].offset
& ~CACHELINE_MASK
);
1324 if (ranges
[i
].offset
+ ranges
[i
].size
> mem
->map_size
)
1325 end
= mem
->map
+ mem
->map_size
;
1327 end
= mem
->map
+ ranges
[i
].offset
+ ranges
[i
].size
;
1330 __builtin_ia32_clflush(p
);
1331 p
+= CACHELINE_SIZE
;
1336 VkResult
anv_FlushMappedMemoryRanges(
1338 uint32_t memoryRangeCount
,
1339 const VkMappedMemoryRange
* pMemoryRanges
)
1341 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1343 if (device
->info
.has_llc
)
1346 /* Make sure the writes we're flushing have landed. */
1347 __builtin_ia32_mfence();
1349 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1354 VkResult
anv_InvalidateMappedMemoryRanges(
1356 uint32_t memoryRangeCount
,
1357 const VkMappedMemoryRange
* pMemoryRanges
)
1359 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1361 if (device
->info
.has_llc
)
1364 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1366 /* Make sure no reads get moved up above the invalidate. */
1367 __builtin_ia32_mfence();
1372 void anv_GetBufferMemoryRequirements(
1375 VkMemoryRequirements
* pMemoryRequirements
)
1377 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1379 /* The Vulkan spec (git aaed022) says:
1381 * memoryTypeBits is a bitfield and contains one bit set for every
1382 * supported memory type for the resource. The bit `1<<i` is set if and
1383 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1384 * structure for the physical device is supported.
1386 * We support exactly one memory type.
1388 pMemoryRequirements
->memoryTypeBits
= 1;
1390 pMemoryRequirements
->size
= buffer
->size
;
1391 pMemoryRequirements
->alignment
= 16;
1394 void anv_GetImageMemoryRequirements(
1397 VkMemoryRequirements
* pMemoryRequirements
)
1399 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1401 /* The Vulkan spec (git aaed022) says:
1403 * memoryTypeBits is a bitfield and contains one bit set for every
1404 * supported memory type for the resource. The bit `1<<i` is set if and
1405 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1406 * structure for the physical device is supported.
1408 * We support exactly one memory type.
1410 pMemoryRequirements
->memoryTypeBits
= 1;
1412 pMemoryRequirements
->size
= image
->size
;
1413 pMemoryRequirements
->alignment
= image
->alignment
;
1416 void anv_GetImageSparseMemoryRequirements(
1419 uint32_t* pSparseMemoryRequirementCount
,
1420 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1425 void anv_GetDeviceMemoryCommitment(
1427 VkDeviceMemory memory
,
1428 VkDeviceSize
* pCommittedMemoryInBytes
)
1430 *pCommittedMemoryInBytes
= 0;
1433 VkResult
anv_BindBufferMemory(
1436 VkDeviceMemory _memory
,
1437 VkDeviceSize memoryOffset
)
1439 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1440 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1443 buffer
->bo
= &mem
->bo
;
1444 buffer
->offset
= memoryOffset
;
1453 VkResult
anv_QueueBindSparse(
1455 uint32_t bindInfoCount
,
1456 const VkBindSparseInfo
* pBindInfo
,
1459 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1462 VkResult
anv_CreateFence(
1464 const VkFenceCreateInfo
* pCreateInfo
,
1465 const VkAllocationCallbacks
* pAllocator
,
1468 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1469 struct anv_bo fence_bo
;
1470 struct anv_fence
*fence
;
1471 struct anv_batch batch
;
1474 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_CREATE_INFO
);
1476 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &fence_bo
, 4096);
1477 if (result
!= VK_SUCCESS
)
1480 /* Fences are small. Just store the CPU data structure in the BO. */
1481 fence
= fence_bo
.map
;
1482 fence
->bo
= fence_bo
;
1484 /* Place the batch after the CPU data but on its own cache line. */
1485 const uint32_t batch_offset
= align_u32(sizeof(*fence
), CACHELINE_SIZE
);
1486 batch
.next
= batch
.start
= fence
->bo
.map
+ batch_offset
;
1487 batch
.end
= fence
->bo
.map
+ fence
->bo
.size
;
1488 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1489 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1491 if (!device
->info
.has_llc
) {
1492 assert(((uintptr_t) batch
.start
& CACHELINE_MASK
) == 0);
1493 assert(batch
.next
- batch
.start
<= CACHELINE_SIZE
);
1494 __builtin_ia32_mfence();
1495 __builtin_ia32_clflush(batch
.start
);
1498 fence
->exec2_objects
[0].handle
= fence
->bo
.gem_handle
;
1499 fence
->exec2_objects
[0].relocation_count
= 0;
1500 fence
->exec2_objects
[0].relocs_ptr
= 0;
1501 fence
->exec2_objects
[0].alignment
= 0;
1502 fence
->exec2_objects
[0].offset
= fence
->bo
.offset
;
1503 fence
->exec2_objects
[0].flags
= 0;
1504 fence
->exec2_objects
[0].rsvd1
= 0;
1505 fence
->exec2_objects
[0].rsvd2
= 0;
1507 fence
->execbuf
.buffers_ptr
= (uintptr_t) fence
->exec2_objects
;
1508 fence
->execbuf
.buffer_count
= 1;
1509 fence
->execbuf
.batch_start_offset
= batch
.start
- fence
->bo
.map
;
1510 fence
->execbuf
.batch_len
= batch
.next
- batch
.start
;
1511 fence
->execbuf
.cliprects_ptr
= 0;
1512 fence
->execbuf
.num_cliprects
= 0;
1513 fence
->execbuf
.DR1
= 0;
1514 fence
->execbuf
.DR4
= 0;
1516 fence
->execbuf
.flags
=
1517 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
1518 fence
->execbuf
.rsvd1
= device
->context_id
;
1519 fence
->execbuf
.rsvd2
= 0;
1521 fence
->ready
= false;
1523 *pFence
= anv_fence_to_handle(fence
);
1528 void anv_DestroyFence(
1531 const VkAllocationCallbacks
* pAllocator
)
1533 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1534 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1536 assert(fence
->bo
.map
== fence
);
1537 anv_bo_pool_free(&device
->batch_bo_pool
, &fence
->bo
);
1540 VkResult
anv_ResetFences(
1542 uint32_t fenceCount
,
1543 const VkFence
* pFences
)
1545 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1546 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1547 fence
->ready
= false;
1553 VkResult
anv_GetFenceStatus(
1557 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1558 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1565 ret
= anv_gem_wait(device
, fence
->bo
.gem_handle
, &t
);
1567 fence
->ready
= true;
1571 return VK_NOT_READY
;
1574 VkResult
anv_WaitForFences(
1576 uint32_t fenceCount
,
1577 const VkFence
* pFences
,
1581 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1583 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1584 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1585 * for a couple of kernel releases. Since there's no way to know
1586 * whether or not the kernel we're using is one of the broken ones, the
1587 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1588 * maximum timeout from 584 years to 292 years - likely not a big deal.
1590 if (timeout
> INT64_MAX
)
1591 timeout
= INT64_MAX
;
1593 int64_t t
= timeout
;
1595 /* FIXME: handle !waitAll */
1597 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1598 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1599 int ret
= anv_gem_wait(device
, fence
->bo
.gem_handle
, &t
);
1600 if (ret
== -1 && errno
== ETIME
) {
1602 } else if (ret
== -1) {
1603 /* We don't know the real error. */
1604 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY
,
1605 "gem wait failed: %m");
1612 // Queue semaphore functions
1614 VkResult
anv_CreateSemaphore(
1616 const VkSemaphoreCreateInfo
* pCreateInfo
,
1617 const VkAllocationCallbacks
* pAllocator
,
1618 VkSemaphore
* pSemaphore
)
1620 /* The DRM execbuffer ioctl always execute in-oder, even between different
1621 * rings. As such, there's nothing to do for the user space semaphore.
1624 *pSemaphore
= (VkSemaphore
)1;
1629 void anv_DestroySemaphore(
1631 VkSemaphore semaphore
,
1632 const VkAllocationCallbacks
* pAllocator
)
1638 VkResult
anv_CreateEvent(
1640 const VkEventCreateInfo
* pCreateInfo
,
1641 const VkAllocationCallbacks
* pAllocator
,
1644 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1645 struct anv_state state
;
1646 struct anv_event
*event
;
1648 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
1650 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
1653 event
->state
= state
;
1654 event
->semaphore
= VK_EVENT_RESET
;
1656 if (!device
->info
.has_llc
) {
1657 /* Make sure the writes we're flushing have landed. */
1658 __builtin_ia32_mfence();
1659 __builtin_ia32_clflush(event
);
1662 *pEvent
= anv_event_to_handle(event
);
1667 void anv_DestroyEvent(
1670 const VkAllocationCallbacks
* pAllocator
)
1672 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1673 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1675 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
1678 VkResult
anv_GetEventStatus(
1682 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1683 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1685 if (!device
->info
.has_llc
) {
1686 /* Invalidate read cache before reading event written by GPU. */
1687 __builtin_ia32_clflush(event
);
1688 __builtin_ia32_mfence();
1692 return event
->semaphore
;
1695 VkResult
anv_SetEvent(
1699 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1700 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1702 event
->semaphore
= VK_EVENT_SET
;
1704 if (!device
->info
.has_llc
) {
1705 /* Make sure the writes we're flushing have landed. */
1706 __builtin_ia32_mfence();
1707 __builtin_ia32_clflush(event
);
1713 VkResult
anv_ResetEvent(
1717 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1718 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1720 event
->semaphore
= VK_EVENT_RESET
;
1722 if (!device
->info
.has_llc
) {
1723 /* Make sure the writes we're flushing have landed. */
1724 __builtin_ia32_mfence();
1725 __builtin_ia32_clflush(event
);
1733 VkResult
anv_CreateBuffer(
1735 const VkBufferCreateInfo
* pCreateInfo
,
1736 const VkAllocationCallbacks
* pAllocator
,
1739 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1740 struct anv_buffer
*buffer
;
1742 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
1744 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
1745 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1747 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1749 buffer
->size
= pCreateInfo
->size
;
1750 buffer
->usage
= pCreateInfo
->usage
;
1754 *pBuffer
= anv_buffer_to_handle(buffer
);
1759 void anv_DestroyBuffer(
1762 const VkAllocationCallbacks
* pAllocator
)
1764 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1765 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1767 vk_free2(&device
->alloc
, pAllocator
, buffer
);
1771 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
1772 enum isl_format format
,
1773 uint32_t offset
, uint32_t range
, uint32_t stride
)
1775 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
1777 .mocs
= device
->default_mocs
,
1782 if (!device
->info
.has_llc
)
1783 anv_state_clflush(state
);
1786 void anv_DestroySampler(
1789 const VkAllocationCallbacks
* pAllocator
)
1791 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1792 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
1794 vk_free2(&device
->alloc
, pAllocator
, sampler
);
1797 VkResult
anv_CreateFramebuffer(
1799 const VkFramebufferCreateInfo
* pCreateInfo
,
1800 const VkAllocationCallbacks
* pAllocator
,
1801 VkFramebuffer
* pFramebuffer
)
1803 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1804 struct anv_framebuffer
*framebuffer
;
1806 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
1808 size_t size
= sizeof(*framebuffer
) +
1809 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
1810 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
1811 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1812 if (framebuffer
== NULL
)
1813 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1815 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
1816 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
1817 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
1818 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
1821 framebuffer
->width
= pCreateInfo
->width
;
1822 framebuffer
->height
= pCreateInfo
->height
;
1823 framebuffer
->layers
= pCreateInfo
->layers
;
1825 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
1830 void anv_DestroyFramebuffer(
1833 const VkAllocationCallbacks
* pAllocator
)
1835 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1836 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
1838 vk_free2(&device
->alloc
, pAllocator
, fb
);