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
33 #include "anv_private.h"
34 #include "util/strtod.h"
35 #include "util/debug.h"
37 #include "genxml/gen7_pack.h"
39 struct anv_dispatch_table dtable
;
42 compiler_debug_log(void *data
, const char *fmt
, ...)
46 compiler_perf_log(void *data
, const char *fmt
, ...)
51 if (unlikely(INTEL_DEBUG
& DEBUG_PERF
))
52 vfprintf(stderr
, fmt
, args
);
58 anv_get_function_timestamp(void *ptr
, uint32_t* timestamp
)
62 if (!dladdr(ptr
, &info
) || !info
.dli_fname
)
65 if (stat(info
.dli_fname
, &st
))
68 *timestamp
= st
.st_mtim
.tv_sec
;
73 anv_device_get_cache_uuid(void *uuid
)
77 memset(uuid
, 0, VK_UUID_SIZE
);
78 if (anv_get_function_timestamp(anv_device_get_cache_uuid
, ×tamp
))
81 snprintf(uuid
, VK_UUID_SIZE
, "anv-%d", timestamp
);
86 anv_physical_device_init(struct anv_physical_device
*device
,
87 struct anv_instance
*instance
,
93 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
95 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
97 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
98 device
->instance
= instance
;
100 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
101 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
103 device
->chipset_id
= anv_gem_get_param(fd
, I915_PARAM_CHIPSET_ID
);
104 if (!device
->chipset_id
) {
105 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
109 device
->name
= gen_get_device_name(device
->chipset_id
);
110 if (!gen_get_device_info(device
->chipset_id
, &device
->info
)) {
111 result
= vk_error(VK_ERROR_INCOMPATIBLE_DRIVER
);
115 if (device
->info
.is_haswell
) {
116 fprintf(stderr
, "WARNING: Haswell Vulkan support is incomplete\n");
117 } else if (device
->info
.gen
== 7 && !device
->info
.is_baytrail
) {
118 fprintf(stderr
, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
119 } else if (device
->info
.gen
== 7 && device
->info
.is_baytrail
) {
120 fprintf(stderr
, "WARNING: Bay Trail Vulkan support is incomplete\n");
121 } else if (device
->info
.gen
>= 8) {
122 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
123 * supported as anything */
125 result
= vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
126 "Vulkan not yet supported on %s", device
->name
);
130 device
->cmd_parser_version
= -1;
131 if (device
->info
.gen
== 7) {
132 device
->cmd_parser_version
=
133 anv_gem_get_param(fd
, I915_PARAM_CMD_PARSER_VERSION
);
134 if (device
->cmd_parser_version
== -1) {
135 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
136 "failed to get command parser version");
141 if (anv_gem_get_aperture(fd
, &device
->aperture_size
) == -1) {
142 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
143 "failed to get aperture size: %m");
147 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_WAIT_TIMEOUT
)) {
148 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
149 "kernel missing gem wait");
153 if (!anv_gem_get_param(fd
, I915_PARAM_HAS_EXECBUF2
)) {
154 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
155 "kernel missing execbuf2");
159 if (!device
->info
.has_llc
&&
160 anv_gem_get_param(fd
, I915_PARAM_MMAP_VERSION
) < 1) {
161 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
162 "kernel missing wc mmap");
166 if (!anv_device_get_cache_uuid(device
->uuid
)) {
167 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
168 "cannot generate UUID");
171 bool swizzled
= anv_gem_get_bit6_swizzle(fd
, I915_TILING_X
);
173 /* GENs prior to 8 do not support EU/Subslice info */
174 if (device
->info
.gen
>= 8) {
175 device
->subslice_total
= anv_gem_get_param(fd
, I915_PARAM_SUBSLICE_TOTAL
);
176 device
->eu_total
= anv_gem_get_param(fd
, I915_PARAM_EU_TOTAL
);
178 /* Without this information, we cannot get the right Braswell
179 * brandstrings, and we have to use conservative numbers for GPGPU on
180 * many platforms, but otherwise, things will just work.
182 if (device
->subslice_total
< 1 || device
->eu_total
< 1) {
183 fprintf(stderr
, "WARNING: Kernel 4.1 required to properly"
184 " query GPU properties.\n");
186 } else if (device
->info
.gen
== 7) {
187 device
->subslice_total
= 1 << (device
->info
.gt
- 1);
190 if (device
->info
.is_cherryview
&&
191 device
->subslice_total
> 0 && device
->eu_total
> 0) {
192 /* Logical CS threads = EUs per subslice * 7 threads per EU */
193 uint32_t max_cs_threads
= device
->eu_total
/ device
->subslice_total
* 7;
195 /* Fuse configurations may give more threads than expected, never less. */
196 if (max_cs_threads
> device
->info
.max_cs_threads
)
197 device
->info
.max_cs_threads
= max_cs_threads
;
200 brw_process_intel_debug_variable();
202 device
->compiler
= brw_compiler_create(NULL
, &device
->info
);
203 if (device
->compiler
== NULL
) {
204 result
= vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
207 device
->compiler
->shader_debug_log
= compiler_debug_log
;
208 device
->compiler
->shader_perf_log
= compiler_perf_log
;
210 result
= anv_init_wsi(device
);
211 if (result
!= VK_SUCCESS
) {
212 ralloc_free(device
->compiler
);
216 isl_device_init(&device
->isl_dev
, &device
->info
, swizzled
);
227 anv_physical_device_finish(struct anv_physical_device
*device
)
229 anv_finish_wsi(device
);
230 ralloc_free(device
->compiler
);
233 static const VkExtensionProperties global_extensions
[] = {
235 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
238 #ifdef VK_USE_PLATFORM_XCB_KHR
240 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
244 #ifdef VK_USE_PLATFORM_XLIB_KHR
246 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
250 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
252 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
258 static const VkExtensionProperties device_extensions
[] = {
260 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
266 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
267 VkSystemAllocationScope allocationScope
)
273 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
274 size_t align
, VkSystemAllocationScope allocationScope
)
276 return realloc(pOriginal
, size
);
280 default_free_func(void *pUserData
, void *pMemory
)
285 static const VkAllocationCallbacks default_alloc
= {
287 .pfnAllocation
= default_alloc_func
,
288 .pfnReallocation
= default_realloc_func
,
289 .pfnFree
= default_free_func
,
292 VkResult
anv_CreateInstance(
293 const VkInstanceCreateInfo
* pCreateInfo
,
294 const VkAllocationCallbacks
* pAllocator
,
295 VkInstance
* pInstance
)
297 struct anv_instance
*instance
;
299 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
301 uint32_t client_version
;
302 if (pCreateInfo
->pApplicationInfo
&&
303 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
304 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
306 client_version
= VK_MAKE_VERSION(1, 0, 0);
309 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
310 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
311 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
312 "Client requested version %d.%d.%d",
313 VK_VERSION_MAJOR(client_version
),
314 VK_VERSION_MINOR(client_version
),
315 VK_VERSION_PATCH(client_version
));
318 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
320 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
321 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
322 global_extensions
[j
].extensionName
) == 0) {
328 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
331 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
332 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
334 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
336 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
339 instance
->alloc
= *pAllocator
;
341 instance
->alloc
= default_alloc
;
343 instance
->apiVersion
= client_version
;
344 instance
->physicalDeviceCount
= -1;
348 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
350 *pInstance
= anv_instance_to_handle(instance
);
355 void anv_DestroyInstance(
356 VkInstance _instance
,
357 const VkAllocationCallbacks
* pAllocator
)
359 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
361 if (instance
->physicalDeviceCount
> 0) {
362 /* We support at most one physical device. */
363 assert(instance
->physicalDeviceCount
== 1);
364 anv_physical_device_finish(&instance
->physicalDevice
);
367 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
371 vk_free(&instance
->alloc
, instance
);
374 VkResult
anv_EnumeratePhysicalDevices(
375 VkInstance _instance
,
376 uint32_t* pPhysicalDeviceCount
,
377 VkPhysicalDevice
* pPhysicalDevices
)
379 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
382 if (instance
->physicalDeviceCount
< 0) {
384 for (unsigned i
= 0; i
< 8; i
++) {
385 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
386 result
= anv_physical_device_init(&instance
->physicalDevice
,
388 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
392 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
393 instance
->physicalDeviceCount
= 0;
394 } else if (result
== VK_SUCCESS
) {
395 instance
->physicalDeviceCount
= 1;
401 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
402 * otherwise it's an inout parameter.
404 * The Vulkan spec (git aaed022) says:
406 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
407 * that is initialized with the number of devices the application is
408 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
409 * an array of at least this many VkPhysicalDevice handles [...].
411 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
412 * overwrites the contents of the variable pointed to by
413 * pPhysicalDeviceCount with the number of physical devices in in the
414 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
415 * pPhysicalDeviceCount with the number of physical handles written to
418 if (!pPhysicalDevices
) {
419 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
420 } else if (*pPhysicalDeviceCount
>= 1) {
421 pPhysicalDevices
[0] = anv_physical_device_to_handle(&instance
->physicalDevice
);
422 *pPhysicalDeviceCount
= 1;
423 } else if (*pPhysicalDeviceCount
< instance
->physicalDeviceCount
) {
424 return VK_INCOMPLETE
;
426 *pPhysicalDeviceCount
= 0;
432 void anv_GetPhysicalDeviceFeatures(
433 VkPhysicalDevice physicalDevice
,
434 VkPhysicalDeviceFeatures
* pFeatures
)
436 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
438 *pFeatures
= (VkPhysicalDeviceFeatures
) {
439 .robustBufferAccess
= true,
440 .fullDrawIndexUint32
= true,
441 .imageCubeArray
= false,
442 .independentBlend
= true,
443 .geometryShader
= true,
444 .tessellationShader
= false,
445 .sampleRateShading
= true,
446 .dualSrcBlend
= true,
448 .multiDrawIndirect
= false,
449 .drawIndirectFirstInstance
= false,
451 .depthBiasClamp
= false,
452 .fillModeNonSolid
= true,
453 .depthBounds
= false,
457 .multiViewport
= true,
458 .samplerAnisotropy
= true,
459 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
460 pdevice
->info
.is_baytrail
,
461 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
462 .textureCompressionBC
= true,
463 .occlusionQueryPrecise
= true,
464 .pipelineStatisticsQuery
= false,
465 .fragmentStoresAndAtomics
= true,
466 .shaderTessellationAndGeometryPointSize
= true,
467 .shaderImageGatherExtended
= false,
468 .shaderStorageImageExtendedFormats
= false,
469 .shaderStorageImageMultisample
= false,
470 .shaderUniformBufferArrayDynamicIndexing
= true,
471 .shaderSampledImageArrayDynamicIndexing
= true,
472 .shaderStorageBufferArrayDynamicIndexing
= true,
473 .shaderStorageImageArrayDynamicIndexing
= true,
474 .shaderStorageImageReadWithoutFormat
= false,
475 .shaderStorageImageWriteWithoutFormat
= true,
476 .shaderClipDistance
= true,
477 .shaderCullDistance
= true,
478 .shaderFloat64
= false,
479 .shaderInt64
= false,
480 .shaderInt16
= false,
482 .variableMultisampleRate
= false,
483 .inheritedQueries
= false,
486 /* We can't do image stores in vec4 shaders */
487 pFeatures
->vertexPipelineStoresAndAtomics
=
488 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
489 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
492 void anv_GetPhysicalDeviceProperties(
493 VkPhysicalDevice physicalDevice
,
494 VkPhysicalDeviceProperties
* pProperties
)
496 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
497 const struct gen_device_info
*devinfo
= &pdevice
->info
;
499 const float time_stamp_base
= devinfo
->gen
>= 9 ? 83.333 : 80.0;
501 /* See assertions made when programming the buffer surface state. */
502 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
503 (1ul << 30) : (1ul << 27);
505 VkSampleCountFlags sample_counts
=
506 isl_device_get_sample_counts(&pdevice
->isl_dev
);
508 VkPhysicalDeviceLimits limits
= {
509 .maxImageDimension1D
= (1 << 14),
510 .maxImageDimension2D
= (1 << 14),
511 .maxImageDimension3D
= (1 << 11),
512 .maxImageDimensionCube
= (1 << 14),
513 .maxImageArrayLayers
= (1 << 11),
514 .maxTexelBufferElements
= 128 * 1024 * 1024,
515 .maxUniformBufferRange
= (1ul << 27),
516 .maxStorageBufferRange
= max_raw_buffer_sz
,
517 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
518 .maxMemoryAllocationCount
= UINT32_MAX
,
519 .maxSamplerAllocationCount
= 64 * 1024,
520 .bufferImageGranularity
= 64, /* A cache line */
521 .sparseAddressSpaceSize
= 0,
522 .maxBoundDescriptorSets
= MAX_SETS
,
523 .maxPerStageDescriptorSamplers
= 64,
524 .maxPerStageDescriptorUniformBuffers
= 64,
525 .maxPerStageDescriptorStorageBuffers
= 64,
526 .maxPerStageDescriptorSampledImages
= 64,
527 .maxPerStageDescriptorStorageImages
= 64,
528 .maxPerStageDescriptorInputAttachments
= 64,
529 .maxPerStageResources
= 128,
530 .maxDescriptorSetSamplers
= 256,
531 .maxDescriptorSetUniformBuffers
= 256,
532 .maxDescriptorSetUniformBuffersDynamic
= 256,
533 .maxDescriptorSetStorageBuffers
= 256,
534 .maxDescriptorSetStorageBuffersDynamic
= 256,
535 .maxDescriptorSetSampledImages
= 256,
536 .maxDescriptorSetStorageImages
= 256,
537 .maxDescriptorSetInputAttachments
= 256,
538 .maxVertexInputAttributes
= 32,
539 .maxVertexInputBindings
= 32,
540 .maxVertexInputAttributeOffset
= 2047,
541 .maxVertexInputBindingStride
= 2048,
542 .maxVertexOutputComponents
= 128,
543 .maxTessellationGenerationLevel
= 0,
544 .maxTessellationPatchSize
= 0,
545 .maxTessellationControlPerVertexInputComponents
= 0,
546 .maxTessellationControlPerVertexOutputComponents
= 0,
547 .maxTessellationControlPerPatchOutputComponents
= 0,
548 .maxTessellationControlTotalOutputComponents
= 0,
549 .maxTessellationEvaluationInputComponents
= 0,
550 .maxTessellationEvaluationOutputComponents
= 0,
551 .maxGeometryShaderInvocations
= 32,
552 .maxGeometryInputComponents
= 64,
553 .maxGeometryOutputComponents
= 128,
554 .maxGeometryOutputVertices
= 256,
555 .maxGeometryTotalOutputComponents
= 1024,
556 .maxFragmentInputComponents
= 128,
557 .maxFragmentOutputAttachments
= 8,
558 .maxFragmentDualSrcAttachments
= 2,
559 .maxFragmentCombinedOutputResources
= 8,
560 .maxComputeSharedMemorySize
= 32768,
561 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
562 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
563 .maxComputeWorkGroupSize
= {
564 16 * devinfo
->max_cs_threads
,
565 16 * devinfo
->max_cs_threads
,
566 16 * devinfo
->max_cs_threads
,
568 .subPixelPrecisionBits
= 4 /* FIXME */,
569 .subTexelPrecisionBits
= 4 /* FIXME */,
570 .mipmapPrecisionBits
= 4 /* FIXME */,
571 .maxDrawIndexedIndexValue
= UINT32_MAX
,
572 .maxDrawIndirectCount
= UINT32_MAX
,
573 .maxSamplerLodBias
= 16,
574 .maxSamplerAnisotropy
= 16,
575 .maxViewports
= MAX_VIEWPORTS
,
576 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
577 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
578 .viewportSubPixelBits
= 13, /* We take a float? */
579 .minMemoryMapAlignment
= 4096, /* A page */
580 .minTexelBufferOffsetAlignment
= 1,
581 .minUniformBufferOffsetAlignment
= 1,
582 .minStorageBufferOffsetAlignment
= 1,
583 .minTexelOffset
= -8,
585 .minTexelGatherOffset
= -8,
586 .maxTexelGatherOffset
= 7,
587 .minInterpolationOffset
= -0.5,
588 .maxInterpolationOffset
= 0.4375,
589 .subPixelInterpolationOffsetBits
= 4,
590 .maxFramebufferWidth
= (1 << 14),
591 .maxFramebufferHeight
= (1 << 14),
592 .maxFramebufferLayers
= (1 << 10),
593 .framebufferColorSampleCounts
= sample_counts
,
594 .framebufferDepthSampleCounts
= sample_counts
,
595 .framebufferStencilSampleCounts
= sample_counts
,
596 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
597 .maxColorAttachments
= MAX_RTS
,
598 .sampledImageColorSampleCounts
= sample_counts
,
599 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
600 .sampledImageDepthSampleCounts
= sample_counts
,
601 .sampledImageStencilSampleCounts
= sample_counts
,
602 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
603 .maxSampleMaskWords
= 1,
604 .timestampComputeAndGraphics
= false,
605 .timestampPeriod
= time_stamp_base
,
606 .maxClipDistances
= 8,
607 .maxCullDistances
= 8,
608 .maxCombinedClipAndCullDistances
= 8,
609 .discreteQueuePriorities
= 1,
610 .pointSizeRange
= { 0.125, 255.875 },
611 .lineWidthRange
= { 0.0, 7.9921875 },
612 .pointSizeGranularity
= (1.0 / 8.0),
613 .lineWidthGranularity
= (1.0 / 128.0),
614 .strictLines
= false, /* FINISHME */
615 .standardSampleLocations
= true,
616 .optimalBufferCopyOffsetAlignment
= 128,
617 .optimalBufferCopyRowPitchAlignment
= 128,
618 .nonCoherentAtomSize
= 64,
621 *pProperties
= (VkPhysicalDeviceProperties
) {
622 .apiVersion
= VK_MAKE_VERSION(1, 0, 5),
625 .deviceID
= pdevice
->chipset_id
,
626 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
628 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
631 strcpy(pProperties
->deviceName
, pdevice
->name
);
632 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
635 void anv_GetPhysicalDeviceQueueFamilyProperties(
636 VkPhysicalDevice physicalDevice
,
638 VkQueueFamilyProperties
* pQueueFamilyProperties
)
640 if (pQueueFamilyProperties
== NULL
) {
645 assert(*pCount
>= 1);
647 *pQueueFamilyProperties
= (VkQueueFamilyProperties
) {
648 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
649 VK_QUEUE_COMPUTE_BIT
|
650 VK_QUEUE_TRANSFER_BIT
,
652 .timestampValidBits
= 36, /* XXX: Real value here */
653 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
657 void anv_GetPhysicalDeviceMemoryProperties(
658 VkPhysicalDevice physicalDevice
,
659 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
661 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
662 VkDeviceSize heap_size
;
664 /* Reserve some wiggle room for the driver by exposing only 75% of the
665 * aperture to the heap.
667 heap_size
= 3 * physical_device
->aperture_size
/ 4;
669 if (physical_device
->info
.has_llc
) {
670 /* Big core GPUs share LLC with the CPU and thus one memory type can be
671 * both cached and coherent at the same time.
673 pMemoryProperties
->memoryTypeCount
= 1;
674 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
675 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
676 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
677 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
678 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
682 /* The spec requires that we expose a host-visible, coherent memory
683 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
684 * to give the application a choice between cached, but not coherent and
685 * coherent but uncached (WC though).
687 pMemoryProperties
->memoryTypeCount
= 2;
688 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
689 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
690 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
691 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
694 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
695 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
696 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
697 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
702 pMemoryProperties
->memoryHeapCount
= 1;
703 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
705 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
709 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
713 return anv_lookup_entrypoint(NULL
, pName
);
716 /* With version 1+ of the loader interface the ICD should expose
717 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
720 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
725 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
729 return anv_GetInstanceProcAddr(instance
, pName
);
732 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
736 ANV_FROM_HANDLE(anv_device
, device
, _device
);
737 return anv_lookup_entrypoint(&device
->info
, pName
);
741 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
743 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
744 queue
->device
= device
;
745 queue
->pool
= &device
->surface_state_pool
;
751 anv_queue_finish(struct anv_queue
*queue
)
755 static struct anv_state
756 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
758 struct anv_state state
;
760 state
= anv_state_pool_alloc(pool
, size
, align
);
761 memcpy(state
.map
, p
, size
);
763 if (!pool
->block_pool
->device
->info
.has_llc
)
764 anv_state_clflush(state
);
769 struct gen8_border_color
{
774 /* Pad out to 64 bytes */
779 anv_device_init_border_colors(struct anv_device
*device
)
781 static const struct gen8_border_color border_colors
[] = {
782 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
783 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
784 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
785 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
786 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
787 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
790 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
791 sizeof(border_colors
), 64,
796 anv_device_submit_simple_batch(struct anv_device
*device
,
797 struct anv_batch
*batch
)
799 struct drm_i915_gem_execbuffer2 execbuf
;
800 struct drm_i915_gem_exec_object2 exec2_objects
[1];
801 struct anv_bo bo
, *exec_bos
[1];
802 VkResult result
= VK_SUCCESS
;
807 /* Kernel driver requires 8 byte aligned batch length */
808 size
= align_u32(batch
->next
- batch
->start
, 8);
809 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &bo
, size
);
810 if (result
!= VK_SUCCESS
)
813 memcpy(bo
.map
, batch
->start
, size
);
814 if (!device
->info
.has_llc
)
815 anv_clflush_range(bo
.map
, size
);
818 exec2_objects
[0].handle
= bo
.gem_handle
;
819 exec2_objects
[0].relocation_count
= 0;
820 exec2_objects
[0].relocs_ptr
= 0;
821 exec2_objects
[0].alignment
= 0;
822 exec2_objects
[0].offset
= bo
.offset
;
823 exec2_objects
[0].flags
= 0;
824 exec2_objects
[0].rsvd1
= 0;
825 exec2_objects
[0].rsvd2
= 0;
827 execbuf
.buffers_ptr
= (uintptr_t) exec2_objects
;
828 execbuf
.buffer_count
= 1;
829 execbuf
.batch_start_offset
= 0;
830 execbuf
.batch_len
= size
;
831 execbuf
.cliprects_ptr
= 0;
832 execbuf
.num_cliprects
= 0;
837 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
838 execbuf
.rsvd1
= device
->context_id
;
841 result
= anv_device_execbuf(device
, &execbuf
, exec_bos
);
842 if (result
!= VK_SUCCESS
)
846 ret
= anv_gem_wait(device
, bo
.gem_handle
, &timeout
);
848 /* We don't know the real error. */
849 result
= vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
854 anv_bo_pool_free(&device
->batch_bo_pool
, &bo
);
859 VkResult
anv_CreateDevice(
860 VkPhysicalDevice physicalDevice
,
861 const VkDeviceCreateInfo
* pCreateInfo
,
862 const VkAllocationCallbacks
* pAllocator
,
865 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
867 struct anv_device
*device
;
869 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
871 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
873 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
874 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
875 device_extensions
[j
].extensionName
) == 0) {
881 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
884 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
886 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
888 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
890 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
891 device
->instance
= physical_device
->instance
;
892 device
->chipset_id
= physical_device
->chipset_id
;
895 device
->alloc
= *pAllocator
;
897 device
->alloc
= physical_device
->instance
->alloc
;
899 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
900 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
901 if (device
->fd
== -1) {
902 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
906 device
->context_id
= anv_gem_create_context(device
);
907 if (device
->context_id
== -1) {
908 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
912 device
->info
= physical_device
->info
;
913 device
->isl_dev
= physical_device
->isl_dev
;
915 /* On Broadwell and later, we can use batch chaining to more efficiently
916 * implement growing command buffers. Prior to Haswell, the kernel
917 * command parser gets in the way and we have to fall back to growing
920 device
->can_chain_batches
= device
->info
.gen
>= 8;
922 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
923 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
925 pthread_mutex_init(&device
->mutex
, NULL
);
927 pthread_condattr_t condattr
;
928 pthread_condattr_init(&condattr
);
929 pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
);
930 pthread_cond_init(&device
->queue_submit
, NULL
);
931 pthread_condattr_destroy(&condattr
);
933 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
935 anv_block_pool_init(&device
->dynamic_state_block_pool
, device
, 16384);
937 anv_state_pool_init(&device
->dynamic_state_pool
,
938 &device
->dynamic_state_block_pool
);
940 anv_block_pool_init(&device
->instruction_block_pool
, device
, 128 * 1024);
941 anv_state_pool_init(&device
->instruction_state_pool
,
942 &device
->instruction_block_pool
);
944 anv_block_pool_init(&device
->surface_state_block_pool
, device
, 4096);
946 anv_state_pool_init(&device
->surface_state_pool
,
947 &device
->surface_state_block_pool
);
949 anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
951 anv_scratch_pool_init(device
, &device
->scratch_pool
);
953 anv_queue_init(device
, &device
->queue
);
955 switch (device
->info
.gen
) {
957 if (!device
->info
.is_haswell
)
958 result
= gen7_init_device_state(device
);
960 result
= gen75_init_device_state(device
);
963 result
= gen8_init_device_state(device
);
966 result
= gen9_init_device_state(device
);
969 /* Shouldn't get here as we don't create physical devices for any other
971 unreachable("unhandled gen");
973 if (result
!= VK_SUCCESS
)
976 anv_device_init_blorp(device
);
978 anv_device_init_border_colors(device
);
980 *pDevice
= anv_device_to_handle(device
);
987 vk_free(&device
->alloc
, device
);
992 void anv_DestroyDevice(
994 const VkAllocationCallbacks
* pAllocator
)
996 ANV_FROM_HANDLE(anv_device
, device
, _device
);
998 anv_queue_finish(&device
->queue
);
1000 anv_device_finish_blorp(device
);
1002 #ifdef HAVE_VALGRIND
1003 /* We only need to free these to prevent valgrind errors. The backing
1004 * BO will go away in a couple of lines so we don't actually leak.
1006 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1009 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1010 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1012 anv_bo_pool_finish(&device
->batch_bo_pool
);
1013 anv_state_pool_finish(&device
->dynamic_state_pool
);
1014 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1015 anv_state_pool_finish(&device
->instruction_state_pool
);
1016 anv_block_pool_finish(&device
->instruction_block_pool
);
1017 anv_state_pool_finish(&device
->surface_state_pool
);
1018 anv_block_pool_finish(&device
->surface_state_block_pool
);
1019 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1023 pthread_mutex_destroy(&device
->mutex
);
1025 vk_free(&device
->alloc
, device
);
1028 VkResult
anv_EnumerateInstanceExtensionProperties(
1029 const char* pLayerName
,
1030 uint32_t* pPropertyCount
,
1031 VkExtensionProperties
* pProperties
)
1033 if (pProperties
== NULL
) {
1034 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1038 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(global_extensions
));
1039 typed_memcpy(pProperties
, global_extensions
, *pPropertyCount
);
1041 if (*pPropertyCount
< ARRAY_SIZE(global_extensions
))
1042 return VK_INCOMPLETE
;
1047 VkResult
anv_EnumerateDeviceExtensionProperties(
1048 VkPhysicalDevice physicalDevice
,
1049 const char* pLayerName
,
1050 uint32_t* pPropertyCount
,
1051 VkExtensionProperties
* pProperties
)
1053 if (pProperties
== NULL
) {
1054 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1058 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(device_extensions
));
1059 typed_memcpy(pProperties
, device_extensions
, *pPropertyCount
);
1061 if (*pPropertyCount
< ARRAY_SIZE(device_extensions
))
1062 return VK_INCOMPLETE
;
1067 VkResult
anv_EnumerateInstanceLayerProperties(
1068 uint32_t* pPropertyCount
,
1069 VkLayerProperties
* pProperties
)
1071 if (pProperties
== NULL
) {
1072 *pPropertyCount
= 0;
1076 /* None supported at this time */
1077 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1080 VkResult
anv_EnumerateDeviceLayerProperties(
1081 VkPhysicalDevice physicalDevice
,
1082 uint32_t* pPropertyCount
,
1083 VkLayerProperties
* pProperties
)
1085 if (pProperties
== NULL
) {
1086 *pPropertyCount
= 0;
1090 /* None supported at this time */
1091 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1094 void anv_GetDeviceQueue(
1096 uint32_t queueNodeIndex
,
1097 uint32_t queueIndex
,
1100 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1102 assert(queueIndex
== 0);
1104 *pQueue
= anv_queue_to_handle(&device
->queue
);
1108 anv_device_execbuf(struct anv_device
*device
,
1109 struct drm_i915_gem_execbuffer2
*execbuf
,
1110 struct anv_bo
**execbuf_bos
)
1112 int ret
= anv_gem_execbuffer(device
, execbuf
);
1114 /* We don't know the real error. */
1115 return vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
1118 struct drm_i915_gem_exec_object2
*objects
=
1119 (void *)(uintptr_t)execbuf
->buffers_ptr
;
1120 for (uint32_t k
= 0; k
< execbuf
->buffer_count
; k
++)
1121 execbuf_bos
[k
]->offset
= objects
[k
].offset
;
1126 VkResult
anv_QueueSubmit(
1128 uint32_t submitCount
,
1129 const VkSubmitInfo
* pSubmits
,
1132 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1133 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1134 struct anv_device
*device
= queue
->device
;
1135 VkResult result
= VK_SUCCESS
;
1137 /* We lock around QueueSubmit for three main reasons:
1139 * 1) When a block pool is resized, we create a new gem handle with a
1140 * different size and, in the case of surface states, possibly a
1141 * different center offset but we re-use the same anv_bo struct when
1142 * we do so. If this happens in the middle of setting up an execbuf,
1143 * we could end up with our list of BOs out of sync with our list of
1146 * 2) The algorithm we use for building the list of unique buffers isn't
1147 * thread-safe. While the client is supposed to syncronize around
1148 * QueueSubmit, this would be extremely difficult to debug if it ever
1149 * came up in the wild due to a broken app. It's better to play it
1150 * safe and just lock around QueueSubmit.
1152 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
1153 * userspace. Due to the fact that the surface state buffer is shared
1154 * between batches, we can't afford to have that happen from multiple
1155 * threads at the same time. Even though the user is supposed to
1156 * ensure this doesn't happen, we play it safe as in (2) above.
1158 * Since the only other things that ever take the device lock such as block
1159 * pool resize only rarely happen, this will almost never be contended so
1160 * taking a lock isn't really an expensive operation in this case.
1162 pthread_mutex_lock(&device
->mutex
);
1164 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1165 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1166 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
,
1167 pSubmits
[i
].pCommandBuffers
[j
]);
1168 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1170 result
= anv_cmd_buffer_execbuf(device
, cmd_buffer
);
1171 if (result
!= VK_SUCCESS
)
1177 struct anv_bo
*fence_bo
= &fence
->bo
;
1178 result
= anv_device_execbuf(device
, &fence
->execbuf
, &fence_bo
);
1179 if (result
!= VK_SUCCESS
)
1182 /* Update the fence and wake up any waiters */
1183 assert(fence
->state
== ANV_FENCE_STATE_RESET
);
1184 fence
->state
= ANV_FENCE_STATE_SUBMITTED
;
1185 pthread_cond_broadcast(&device
->queue_submit
);
1189 pthread_mutex_unlock(&device
->mutex
);
1194 VkResult
anv_QueueWaitIdle(
1197 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1199 return anv_DeviceWaitIdle(anv_device_to_handle(queue
->device
));
1202 VkResult
anv_DeviceWaitIdle(
1205 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1206 struct anv_batch batch
;
1209 batch
.start
= batch
.next
= cmds
;
1210 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1212 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1213 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1215 return anv_device_submit_simple_batch(device
, &batch
);
1219 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1221 uint32_t gem_handle
= anv_gem_create(device
, size
);
1223 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1225 anv_bo_init(bo
, gem_handle
, size
);
1230 VkResult
anv_AllocateMemory(
1232 const VkMemoryAllocateInfo
* pAllocateInfo
,
1233 const VkAllocationCallbacks
* pAllocator
,
1234 VkDeviceMemory
* pMem
)
1236 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1237 struct anv_device_memory
*mem
;
1240 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1242 if (pAllocateInfo
->allocationSize
== 0) {
1243 /* Apparently, this is allowed */
1244 *pMem
= VK_NULL_HANDLE
;
1248 /* We support exactly one memory heap. */
1249 assert(pAllocateInfo
->memoryTypeIndex
== 0 ||
1250 (!device
->info
.has_llc
&& pAllocateInfo
->memoryTypeIndex
< 2));
1252 /* FINISHME: Fail if allocation request exceeds heap size. */
1254 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1255 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1257 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1259 /* The kernel is going to give us whole pages anyway */
1260 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
1262 result
= anv_bo_init_new(&mem
->bo
, device
, alloc_size
);
1263 if (result
!= VK_SUCCESS
)
1266 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1271 *pMem
= anv_device_memory_to_handle(mem
);
1276 vk_free2(&device
->alloc
, pAllocator
, mem
);
1281 void anv_FreeMemory(
1283 VkDeviceMemory _mem
,
1284 const VkAllocationCallbacks
* pAllocator
)
1286 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1287 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1293 anv_UnmapMemory(_device
, _mem
);
1296 anv_gem_munmap(mem
->bo
.map
, mem
->bo
.size
);
1298 if (mem
->bo
.gem_handle
!= 0)
1299 anv_gem_close(device
, mem
->bo
.gem_handle
);
1301 vk_free2(&device
->alloc
, pAllocator
, mem
);
1304 VkResult
anv_MapMemory(
1306 VkDeviceMemory _memory
,
1307 VkDeviceSize offset
,
1309 VkMemoryMapFlags flags
,
1312 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1313 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1320 if (size
== VK_WHOLE_SIZE
)
1321 size
= mem
->bo
.size
- offset
;
1323 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1325 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1326 * assert(size != 0);
1327 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1328 * equal to the size of the memory minus offset
1331 assert(offset
+ size
<= mem
->bo
.size
);
1333 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1334 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1335 * at a time is valid. We could just mmap up front and return an offset
1336 * pointer here, but that may exhaust virtual memory on 32 bit
1339 uint32_t gem_flags
= 0;
1340 if (!device
->info
.has_llc
&& mem
->type_index
== 0)
1341 gem_flags
|= I915_MMAP_WC
;
1343 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1344 uint64_t map_offset
= offset
& ~4095ull;
1345 assert(offset
>= map_offset
);
1346 uint64_t map_size
= (offset
+ size
) - map_offset
;
1348 /* Let's map whole pages */
1349 map_size
= align_u64(map_size
, 4096);
1351 void *map
= anv_gem_mmap(device
, mem
->bo
.gem_handle
,
1352 map_offset
, map_size
, gem_flags
);
1353 if (map
== MAP_FAILED
)
1354 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1357 mem
->map_size
= map_size
;
1359 *ppData
= mem
->map
+ (offset
- map_offset
);
1364 void anv_UnmapMemory(
1366 VkDeviceMemory _memory
)
1368 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1373 anv_gem_munmap(mem
->map
, mem
->map_size
);
1380 clflush_mapped_ranges(struct anv_device
*device
,
1382 const VkMappedMemoryRange
*ranges
)
1384 for (uint32_t i
= 0; i
< count
; i
++) {
1385 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1386 void *p
= mem
->map
+ (ranges
[i
].offset
& ~CACHELINE_MASK
);
1389 if (ranges
[i
].offset
+ ranges
[i
].size
> mem
->map_size
)
1390 end
= mem
->map
+ mem
->map_size
;
1392 end
= mem
->map
+ ranges
[i
].offset
+ ranges
[i
].size
;
1395 __builtin_ia32_clflush(p
);
1396 p
+= CACHELINE_SIZE
;
1401 VkResult
anv_FlushMappedMemoryRanges(
1403 uint32_t memoryRangeCount
,
1404 const VkMappedMemoryRange
* pMemoryRanges
)
1406 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1408 if (device
->info
.has_llc
)
1411 /* Make sure the writes we're flushing have landed. */
1412 __builtin_ia32_mfence();
1414 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1419 VkResult
anv_InvalidateMappedMemoryRanges(
1421 uint32_t memoryRangeCount
,
1422 const VkMappedMemoryRange
* pMemoryRanges
)
1424 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1426 if (device
->info
.has_llc
)
1429 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1431 /* Make sure no reads get moved up above the invalidate. */
1432 __builtin_ia32_mfence();
1437 void anv_GetBufferMemoryRequirements(
1440 VkMemoryRequirements
* pMemoryRequirements
)
1442 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1444 /* The Vulkan spec (git aaed022) says:
1446 * memoryTypeBits is a bitfield and contains one bit set for every
1447 * supported memory type for the resource. The bit `1<<i` is set if and
1448 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1449 * structure for the physical device is supported.
1451 * We support exactly one memory type.
1453 pMemoryRequirements
->memoryTypeBits
= 1;
1455 pMemoryRequirements
->size
= buffer
->size
;
1456 pMemoryRequirements
->alignment
= 16;
1459 void anv_GetImageMemoryRequirements(
1462 VkMemoryRequirements
* pMemoryRequirements
)
1464 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1466 /* The Vulkan spec (git aaed022) says:
1468 * memoryTypeBits is a bitfield and contains one bit set for every
1469 * supported memory type for the resource. The bit `1<<i` is set if and
1470 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1471 * structure for the physical device is supported.
1473 * We support exactly one memory type.
1475 pMemoryRequirements
->memoryTypeBits
= 1;
1477 pMemoryRequirements
->size
= image
->size
;
1478 pMemoryRequirements
->alignment
= image
->alignment
;
1481 void anv_GetImageSparseMemoryRequirements(
1484 uint32_t* pSparseMemoryRequirementCount
,
1485 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1490 void anv_GetDeviceMemoryCommitment(
1492 VkDeviceMemory memory
,
1493 VkDeviceSize
* pCommittedMemoryInBytes
)
1495 *pCommittedMemoryInBytes
= 0;
1498 VkResult
anv_BindBufferMemory(
1501 VkDeviceMemory _memory
,
1502 VkDeviceSize memoryOffset
)
1504 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1505 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1508 buffer
->bo
= &mem
->bo
;
1509 buffer
->offset
= memoryOffset
;
1518 VkResult
anv_QueueBindSparse(
1520 uint32_t bindInfoCount
,
1521 const VkBindSparseInfo
* pBindInfo
,
1524 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1527 VkResult
anv_CreateFence(
1529 const VkFenceCreateInfo
* pCreateInfo
,
1530 const VkAllocationCallbacks
* pAllocator
,
1533 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1534 struct anv_bo fence_bo
;
1535 struct anv_fence
*fence
;
1536 struct anv_batch batch
;
1539 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_CREATE_INFO
);
1541 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &fence_bo
, 4096);
1542 if (result
!= VK_SUCCESS
)
1545 /* Fences are small. Just store the CPU data structure in the BO. */
1546 fence
= fence_bo
.map
;
1547 fence
->bo
= fence_bo
;
1549 /* Place the batch after the CPU data but on its own cache line. */
1550 const uint32_t batch_offset
= align_u32(sizeof(*fence
), CACHELINE_SIZE
);
1551 batch
.next
= batch
.start
= fence
->bo
.map
+ batch_offset
;
1552 batch
.end
= fence
->bo
.map
+ fence
->bo
.size
;
1553 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1554 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1556 if (!device
->info
.has_llc
) {
1557 assert(((uintptr_t) batch
.start
& CACHELINE_MASK
) == 0);
1558 assert(batch
.next
- batch
.start
<= CACHELINE_SIZE
);
1559 __builtin_ia32_mfence();
1560 __builtin_ia32_clflush(batch
.start
);
1563 fence
->exec2_objects
[0].handle
= fence
->bo
.gem_handle
;
1564 fence
->exec2_objects
[0].relocation_count
= 0;
1565 fence
->exec2_objects
[0].relocs_ptr
= 0;
1566 fence
->exec2_objects
[0].alignment
= 0;
1567 fence
->exec2_objects
[0].offset
= fence
->bo
.offset
;
1568 fence
->exec2_objects
[0].flags
= 0;
1569 fence
->exec2_objects
[0].rsvd1
= 0;
1570 fence
->exec2_objects
[0].rsvd2
= 0;
1572 fence
->execbuf
.buffers_ptr
= (uintptr_t) fence
->exec2_objects
;
1573 fence
->execbuf
.buffer_count
= 1;
1574 fence
->execbuf
.batch_start_offset
= batch
.start
- fence
->bo
.map
;
1575 fence
->execbuf
.batch_len
= batch
.next
- batch
.start
;
1576 fence
->execbuf
.cliprects_ptr
= 0;
1577 fence
->execbuf
.num_cliprects
= 0;
1578 fence
->execbuf
.DR1
= 0;
1579 fence
->execbuf
.DR4
= 0;
1581 fence
->execbuf
.flags
=
1582 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
1583 fence
->execbuf
.rsvd1
= device
->context_id
;
1584 fence
->execbuf
.rsvd2
= 0;
1586 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
1587 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1589 fence
->state
= ANV_FENCE_STATE_RESET
;
1592 *pFence
= anv_fence_to_handle(fence
);
1597 void anv_DestroyFence(
1600 const VkAllocationCallbacks
* pAllocator
)
1602 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1603 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1608 assert(fence
->bo
.map
== fence
);
1609 anv_bo_pool_free(&device
->batch_bo_pool
, &fence
->bo
);
1612 VkResult
anv_ResetFences(
1614 uint32_t fenceCount
,
1615 const VkFence
* pFences
)
1617 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1618 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1619 fence
->state
= ANV_FENCE_STATE_RESET
;
1625 VkResult
anv_GetFenceStatus(
1629 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1630 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1634 switch (fence
->state
) {
1635 case ANV_FENCE_STATE_RESET
:
1636 /* If it hasn't even been sent off to the GPU yet, it's not ready */
1637 return VK_NOT_READY
;
1639 case ANV_FENCE_STATE_SIGNALED
:
1640 /* It's been signaled, return success */
1643 case ANV_FENCE_STATE_SUBMITTED
:
1644 /* It's been submitted to the GPU but we don't know if it's done yet. */
1645 ret
= anv_gem_wait(device
, fence
->bo
.gem_handle
, &t
);
1647 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1650 return VK_NOT_READY
;
1653 unreachable("Invalid fence status");
1657 #define NSEC_PER_SEC 1000000000
1658 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1660 VkResult
anv_WaitForFences(
1662 uint32_t fenceCount
,
1663 const VkFence
* pFences
,
1667 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1670 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1671 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1672 * for a couple of kernel releases. Since there's no way to know
1673 * whether or not the kernel we're using is one of the broken ones, the
1674 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1675 * maximum timeout from 584 years to 292 years - likely not a big deal.
1677 int64_t timeout
= MIN2(_timeout
, INT64_MAX
);
1679 uint32_t pending_fences
= fenceCount
;
1680 while (pending_fences
) {
1682 bool signaled_fences
= false;
1683 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1684 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1685 switch (fence
->state
) {
1686 case ANV_FENCE_STATE_RESET
:
1687 /* This fence hasn't been submitted yet, we'll catch it the next
1688 * time around. Yes, this may mean we dead-loop but, short of
1689 * lots of locking and a condition variable, there's not much that
1690 * we can do about that.
1695 case ANV_FENCE_STATE_SIGNALED
:
1696 /* This fence is not pending. If waitAll isn't set, we can return
1697 * early. Otherwise, we have to keep going.
1703 case ANV_FENCE_STATE_SUBMITTED
:
1704 /* These are the fences we really care about. Go ahead and wait
1705 * on it until we hit a timeout.
1707 ret
= anv_gem_wait(device
, fence
->bo
.gem_handle
, &timeout
);
1708 if (ret
== -1 && errno
== ETIME
) {
1710 } else if (ret
== -1) {
1711 /* We don't know the real error. */
1712 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1714 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1715 signaled_fences
= true;
1723 if (pending_fences
&& !signaled_fences
) {
1724 /* If we've hit this then someone decided to vkWaitForFences before
1725 * they've actually submitted any of them to a queue. This is a
1726 * fairly pessimal case, so it's ok to lock here and use a standard
1727 * pthreads condition variable.
1729 pthread_mutex_lock(&device
->mutex
);
1731 /* It's possible that some of the fences have changed state since the
1732 * last time we checked. Now that we have the lock, check for
1733 * pending fences again and don't wait if it's changed.
1735 uint32_t now_pending_fences
= 0;
1736 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1737 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1738 if (fence
->state
== ANV_FENCE_STATE_RESET
)
1739 now_pending_fences
++;
1741 assert(now_pending_fences
<= pending_fences
);
1743 if (now_pending_fences
== pending_fences
) {
1744 struct timespec before
;
1745 clock_gettime(CLOCK_MONOTONIC
, &before
);
1747 uint32_t abs_nsec
= before
.tv_nsec
+ timeout
% NSEC_PER_SEC
;
1748 uint64_t abs_sec
= before
.tv_sec
+ (abs_nsec
/ NSEC_PER_SEC
) +
1749 (timeout
/ NSEC_PER_SEC
);
1750 abs_nsec
%= NSEC_PER_SEC
;
1752 /* Avoid roll-over in tv_sec on 32-bit systems if the user
1753 * provided timeout is UINT64_MAX
1755 struct timespec abstime
;
1756 abstime
.tv_nsec
= abs_nsec
;
1757 abstime
.tv_sec
= MIN2(abs_sec
, INT_TYPE_MAX(abstime
.tv_sec
));
1759 ret
= pthread_cond_timedwait(&device
->queue_submit
,
1760 &device
->mutex
, &abstime
);
1761 assert(ret
!= EINVAL
);
1763 struct timespec after
;
1764 clock_gettime(CLOCK_MONOTONIC
, &after
);
1765 uint64_t time_elapsed
=
1766 ((uint64_t)after
.tv_sec
* NSEC_PER_SEC
+ after
.tv_nsec
) -
1767 ((uint64_t)before
.tv_sec
* NSEC_PER_SEC
+ before
.tv_nsec
);
1769 if (time_elapsed
>= timeout
) {
1770 pthread_mutex_unlock(&device
->mutex
);
1774 timeout
-= time_elapsed
;
1777 pthread_mutex_unlock(&device
->mutex
);
1784 // Queue semaphore functions
1786 VkResult
anv_CreateSemaphore(
1788 const VkSemaphoreCreateInfo
* pCreateInfo
,
1789 const VkAllocationCallbacks
* pAllocator
,
1790 VkSemaphore
* pSemaphore
)
1792 /* The DRM execbuffer ioctl always execute in-oder, even between different
1793 * rings. As such, there's nothing to do for the user space semaphore.
1796 *pSemaphore
= (VkSemaphore
)1;
1801 void anv_DestroySemaphore(
1803 VkSemaphore semaphore
,
1804 const VkAllocationCallbacks
* pAllocator
)
1810 VkResult
anv_CreateEvent(
1812 const VkEventCreateInfo
* pCreateInfo
,
1813 const VkAllocationCallbacks
* pAllocator
,
1816 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1817 struct anv_state state
;
1818 struct anv_event
*event
;
1820 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
1822 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
1825 event
->state
= state
;
1826 event
->semaphore
= VK_EVENT_RESET
;
1828 if (!device
->info
.has_llc
) {
1829 /* Make sure the writes we're flushing have landed. */
1830 __builtin_ia32_mfence();
1831 __builtin_ia32_clflush(event
);
1834 *pEvent
= anv_event_to_handle(event
);
1839 void anv_DestroyEvent(
1842 const VkAllocationCallbacks
* pAllocator
)
1844 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1845 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1850 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
1853 VkResult
anv_GetEventStatus(
1857 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1858 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1860 if (!device
->info
.has_llc
) {
1861 /* Invalidate read cache before reading event written by GPU. */
1862 __builtin_ia32_clflush(event
);
1863 __builtin_ia32_mfence();
1867 return event
->semaphore
;
1870 VkResult
anv_SetEvent(
1874 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1875 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1877 event
->semaphore
= VK_EVENT_SET
;
1879 if (!device
->info
.has_llc
) {
1880 /* Make sure the writes we're flushing have landed. */
1881 __builtin_ia32_mfence();
1882 __builtin_ia32_clflush(event
);
1888 VkResult
anv_ResetEvent(
1892 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1893 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1895 event
->semaphore
= VK_EVENT_RESET
;
1897 if (!device
->info
.has_llc
) {
1898 /* Make sure the writes we're flushing have landed. */
1899 __builtin_ia32_mfence();
1900 __builtin_ia32_clflush(event
);
1908 VkResult
anv_CreateBuffer(
1910 const VkBufferCreateInfo
* pCreateInfo
,
1911 const VkAllocationCallbacks
* pAllocator
,
1914 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1915 struct anv_buffer
*buffer
;
1917 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
1919 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
1920 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1922 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1924 buffer
->size
= pCreateInfo
->size
;
1925 buffer
->usage
= pCreateInfo
->usage
;
1929 *pBuffer
= anv_buffer_to_handle(buffer
);
1934 void anv_DestroyBuffer(
1937 const VkAllocationCallbacks
* pAllocator
)
1939 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1940 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1945 vk_free2(&device
->alloc
, pAllocator
, buffer
);
1949 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
1950 enum isl_format format
,
1951 uint32_t offset
, uint32_t range
, uint32_t stride
)
1953 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
1955 .mocs
= device
->default_mocs
,
1960 if (!device
->info
.has_llc
)
1961 anv_state_clflush(state
);
1964 void anv_DestroySampler(
1967 const VkAllocationCallbacks
* pAllocator
)
1969 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1970 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
1975 vk_free2(&device
->alloc
, pAllocator
, sampler
);
1978 VkResult
anv_CreateFramebuffer(
1980 const VkFramebufferCreateInfo
* pCreateInfo
,
1981 const VkAllocationCallbacks
* pAllocator
,
1982 VkFramebuffer
* pFramebuffer
)
1984 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1985 struct anv_framebuffer
*framebuffer
;
1987 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
1989 size_t size
= sizeof(*framebuffer
) +
1990 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
1991 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
1992 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1993 if (framebuffer
== NULL
)
1994 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1996 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
1997 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
1998 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
1999 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2002 framebuffer
->width
= pCreateInfo
->width
;
2003 framebuffer
->height
= pCreateInfo
->height
;
2004 framebuffer
->layers
= pCreateInfo
->layers
;
2006 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2011 void anv_DestroyFramebuffer(
2014 const VkAllocationCallbacks
* pAllocator
)
2016 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2017 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2022 vk_free2(&device
->alloc
, pAllocator
, fb
);