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
,
264 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
270 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
271 VkSystemAllocationScope allocationScope
)
277 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
278 size_t align
, VkSystemAllocationScope allocationScope
)
280 return realloc(pOriginal
, size
);
284 default_free_func(void *pUserData
, void *pMemory
)
289 static const VkAllocationCallbacks default_alloc
= {
291 .pfnAllocation
= default_alloc_func
,
292 .pfnReallocation
= default_realloc_func
,
293 .pfnFree
= default_free_func
,
296 VkResult
anv_CreateInstance(
297 const VkInstanceCreateInfo
* pCreateInfo
,
298 const VkAllocationCallbacks
* pAllocator
,
299 VkInstance
* pInstance
)
301 struct anv_instance
*instance
;
303 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
305 uint32_t client_version
;
306 if (pCreateInfo
->pApplicationInfo
&&
307 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
308 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
310 client_version
= VK_MAKE_VERSION(1, 0, 0);
313 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
314 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
315 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
316 "Client requested version %d.%d.%d",
317 VK_VERSION_MAJOR(client_version
),
318 VK_VERSION_MINOR(client_version
),
319 VK_VERSION_PATCH(client_version
));
322 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
324 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
325 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
326 global_extensions
[j
].extensionName
) == 0) {
332 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
335 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
336 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
338 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
340 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
343 instance
->alloc
= *pAllocator
;
345 instance
->alloc
= default_alloc
;
347 instance
->apiVersion
= client_version
;
348 instance
->physicalDeviceCount
= -1;
352 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
354 *pInstance
= anv_instance_to_handle(instance
);
359 void anv_DestroyInstance(
360 VkInstance _instance
,
361 const VkAllocationCallbacks
* pAllocator
)
363 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
365 if (instance
->physicalDeviceCount
> 0) {
366 /* We support at most one physical device. */
367 assert(instance
->physicalDeviceCount
== 1);
368 anv_physical_device_finish(&instance
->physicalDevice
);
371 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
375 vk_free(&instance
->alloc
, instance
);
378 VkResult
anv_EnumeratePhysicalDevices(
379 VkInstance _instance
,
380 uint32_t* pPhysicalDeviceCount
,
381 VkPhysicalDevice
* pPhysicalDevices
)
383 ANV_FROM_HANDLE(anv_instance
, instance
, _instance
);
386 if (instance
->physicalDeviceCount
< 0) {
388 for (unsigned i
= 0; i
< 8; i
++) {
389 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
390 result
= anv_physical_device_init(&instance
->physicalDevice
,
392 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
396 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
397 instance
->physicalDeviceCount
= 0;
398 } else if (result
== VK_SUCCESS
) {
399 instance
->physicalDeviceCount
= 1;
405 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
406 * otherwise it's an inout parameter.
408 * The Vulkan spec (git aaed022) says:
410 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
411 * that is initialized with the number of devices the application is
412 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
413 * an array of at least this many VkPhysicalDevice handles [...].
415 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
416 * overwrites the contents of the variable pointed to by
417 * pPhysicalDeviceCount with the number of physical devices in in the
418 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
419 * pPhysicalDeviceCount with the number of physical handles written to
422 if (!pPhysicalDevices
) {
423 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
424 } else if (*pPhysicalDeviceCount
>= 1) {
425 pPhysicalDevices
[0] = anv_physical_device_to_handle(&instance
->physicalDevice
);
426 *pPhysicalDeviceCount
= 1;
427 } else if (*pPhysicalDeviceCount
< instance
->physicalDeviceCount
) {
428 return VK_INCOMPLETE
;
430 *pPhysicalDeviceCount
= 0;
436 void anv_GetPhysicalDeviceFeatures(
437 VkPhysicalDevice physicalDevice
,
438 VkPhysicalDeviceFeatures
* pFeatures
)
440 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
442 *pFeatures
= (VkPhysicalDeviceFeatures
) {
443 .robustBufferAccess
= true,
444 .fullDrawIndexUint32
= true,
445 .imageCubeArray
= true,
446 .independentBlend
= true,
447 .geometryShader
= true,
448 .tessellationShader
= false,
449 .sampleRateShading
= true,
450 .dualSrcBlend
= true,
452 .multiDrawIndirect
= false,
453 .drawIndirectFirstInstance
= true,
455 .depthBiasClamp
= true,
456 .fillModeNonSolid
= true,
457 .depthBounds
= false,
461 .multiViewport
= true,
462 .samplerAnisotropy
= true,
463 .textureCompressionETC2
= pdevice
->info
.gen
>= 8 ||
464 pdevice
->info
.is_baytrail
,
465 .textureCompressionASTC_LDR
= pdevice
->info
.gen
>= 9, /* FINISHME CHV */
466 .textureCompressionBC
= true,
467 .occlusionQueryPrecise
= true,
468 .pipelineStatisticsQuery
= false,
469 .fragmentStoresAndAtomics
= true,
470 .shaderTessellationAndGeometryPointSize
= true,
471 .shaderImageGatherExtended
= true,
472 .shaderStorageImageExtendedFormats
= true,
473 .shaderStorageImageMultisample
= false,
474 .shaderStorageImageReadWithoutFormat
= false,
475 .shaderStorageImageWriteWithoutFormat
= false,
476 .shaderUniformBufferArrayDynamicIndexing
= true,
477 .shaderSampledImageArrayDynamicIndexing
= true,
478 .shaderStorageBufferArrayDynamicIndexing
= true,
479 .shaderStorageImageArrayDynamicIndexing
= true,
480 .shaderClipDistance
= true,
481 .shaderCullDistance
= true,
482 .shaderFloat64
= pdevice
->info
.gen
>= 8,
483 .shaderInt64
= false,
484 .shaderInt16
= false,
485 .shaderResourceMinLod
= false,
486 .variableMultisampleRate
= false,
487 .inheritedQueries
= false,
490 /* We can't do image stores in vec4 shaders */
491 pFeatures
->vertexPipelineStoresAndAtomics
=
492 pdevice
->compiler
->scalar_stage
[MESA_SHADER_VERTEX
] &&
493 pdevice
->compiler
->scalar_stage
[MESA_SHADER_GEOMETRY
];
496 void anv_GetPhysicalDeviceProperties(
497 VkPhysicalDevice physicalDevice
,
498 VkPhysicalDeviceProperties
* pProperties
)
500 ANV_FROM_HANDLE(anv_physical_device
, pdevice
, physicalDevice
);
501 const struct gen_device_info
*devinfo
= &pdevice
->info
;
503 const float time_stamp_base
= devinfo
->gen
>= 9 ? 83.333 : 80.0;
505 /* See assertions made when programming the buffer surface state. */
506 const uint32_t max_raw_buffer_sz
= devinfo
->gen
>= 7 ?
507 (1ul << 30) : (1ul << 27);
509 VkSampleCountFlags sample_counts
=
510 isl_device_get_sample_counts(&pdevice
->isl_dev
);
512 VkPhysicalDeviceLimits limits
= {
513 .maxImageDimension1D
= (1 << 14),
514 .maxImageDimension2D
= (1 << 14),
515 .maxImageDimension3D
= (1 << 11),
516 .maxImageDimensionCube
= (1 << 14),
517 .maxImageArrayLayers
= (1 << 11),
518 .maxTexelBufferElements
= 128 * 1024 * 1024,
519 .maxUniformBufferRange
= (1ul << 27),
520 .maxStorageBufferRange
= max_raw_buffer_sz
,
521 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
522 .maxMemoryAllocationCount
= UINT32_MAX
,
523 .maxSamplerAllocationCount
= 64 * 1024,
524 .bufferImageGranularity
= 64, /* A cache line */
525 .sparseAddressSpaceSize
= 0,
526 .maxBoundDescriptorSets
= MAX_SETS
,
527 .maxPerStageDescriptorSamplers
= 64,
528 .maxPerStageDescriptorUniformBuffers
= 64,
529 .maxPerStageDescriptorStorageBuffers
= 64,
530 .maxPerStageDescriptorSampledImages
= 64,
531 .maxPerStageDescriptorStorageImages
= 64,
532 .maxPerStageDescriptorInputAttachments
= 64,
533 .maxPerStageResources
= 128,
534 .maxDescriptorSetSamplers
= 256,
535 .maxDescriptorSetUniformBuffers
= 256,
536 .maxDescriptorSetUniformBuffersDynamic
= 256,
537 .maxDescriptorSetStorageBuffers
= 256,
538 .maxDescriptorSetStorageBuffersDynamic
= 256,
539 .maxDescriptorSetSampledImages
= 256,
540 .maxDescriptorSetStorageImages
= 256,
541 .maxDescriptorSetInputAttachments
= 256,
542 .maxVertexInputAttributes
= 32,
543 .maxVertexInputBindings
= 32,
544 .maxVertexInputAttributeOffset
= 2047,
545 .maxVertexInputBindingStride
= 2048,
546 .maxVertexOutputComponents
= 128,
547 .maxTessellationGenerationLevel
= 64,
548 .maxTessellationPatchSize
= 32,
549 .maxTessellationControlPerVertexInputComponents
= 128,
550 .maxTessellationControlPerVertexOutputComponents
= 128,
551 .maxTessellationControlPerPatchOutputComponents
= 128,
552 .maxTessellationControlTotalOutputComponents
= 2048,
553 .maxTessellationEvaluationInputComponents
= 128,
554 .maxTessellationEvaluationOutputComponents
= 128,
555 .maxGeometryShaderInvocations
= 32,
556 .maxGeometryInputComponents
= 64,
557 .maxGeometryOutputComponents
= 128,
558 .maxGeometryOutputVertices
= 256,
559 .maxGeometryTotalOutputComponents
= 1024,
560 .maxFragmentInputComponents
= 128,
561 .maxFragmentOutputAttachments
= 8,
562 .maxFragmentDualSrcAttachments
= 1,
563 .maxFragmentCombinedOutputResources
= 8,
564 .maxComputeSharedMemorySize
= 32768,
565 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
566 .maxComputeWorkGroupInvocations
= 16 * devinfo
->max_cs_threads
,
567 .maxComputeWorkGroupSize
= {
568 16 * devinfo
->max_cs_threads
,
569 16 * devinfo
->max_cs_threads
,
570 16 * devinfo
->max_cs_threads
,
572 .subPixelPrecisionBits
= 4 /* FIXME */,
573 .subTexelPrecisionBits
= 4 /* FIXME */,
574 .mipmapPrecisionBits
= 4 /* FIXME */,
575 .maxDrawIndexedIndexValue
= UINT32_MAX
,
576 .maxDrawIndirectCount
= UINT32_MAX
,
577 .maxSamplerLodBias
= 16,
578 .maxSamplerAnisotropy
= 16,
579 .maxViewports
= MAX_VIEWPORTS
,
580 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
581 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
582 .viewportSubPixelBits
= 13, /* We take a float? */
583 .minMemoryMapAlignment
= 4096, /* A page */
584 .minTexelBufferOffsetAlignment
= 1,
585 .minUniformBufferOffsetAlignment
= 16,
586 .minStorageBufferOffsetAlignment
= 4,
587 .minTexelOffset
= -8,
589 .minTexelGatherOffset
= -32,
590 .maxTexelGatherOffset
= 31,
591 .minInterpolationOffset
= -0.5,
592 .maxInterpolationOffset
= 0.4375,
593 .subPixelInterpolationOffsetBits
= 4,
594 .maxFramebufferWidth
= (1 << 14),
595 .maxFramebufferHeight
= (1 << 14),
596 .maxFramebufferLayers
= (1 << 11),
597 .framebufferColorSampleCounts
= sample_counts
,
598 .framebufferDepthSampleCounts
= sample_counts
,
599 .framebufferStencilSampleCounts
= sample_counts
,
600 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
601 .maxColorAttachments
= MAX_RTS
,
602 .sampledImageColorSampleCounts
= sample_counts
,
603 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
604 .sampledImageDepthSampleCounts
= sample_counts
,
605 .sampledImageStencilSampleCounts
= sample_counts
,
606 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
607 .maxSampleMaskWords
= 1,
608 .timestampComputeAndGraphics
= false,
609 .timestampPeriod
= time_stamp_base
,
610 .maxClipDistances
= 8,
611 .maxCullDistances
= 8,
612 .maxCombinedClipAndCullDistances
= 8,
613 .discreteQueuePriorities
= 1,
614 .pointSizeRange
= { 0.125, 255.875 },
615 .lineWidthRange
= { 0.0, 7.9921875 },
616 .pointSizeGranularity
= (1.0 / 8.0),
617 .lineWidthGranularity
= (1.0 / 128.0),
618 .strictLines
= false, /* FINISHME */
619 .standardSampleLocations
= true,
620 .optimalBufferCopyOffsetAlignment
= 128,
621 .optimalBufferCopyRowPitchAlignment
= 128,
622 .nonCoherentAtomSize
= 64,
625 *pProperties
= (VkPhysicalDeviceProperties
) {
626 .apiVersion
= VK_MAKE_VERSION(1, 0, 5),
629 .deviceID
= pdevice
->chipset_id
,
630 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU
,
632 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
635 strcpy(pProperties
->deviceName
, pdevice
->name
);
636 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
639 void anv_GetPhysicalDeviceQueueFamilyProperties(
640 VkPhysicalDevice physicalDevice
,
642 VkQueueFamilyProperties
* pQueueFamilyProperties
)
644 if (pQueueFamilyProperties
== NULL
) {
649 /* The spec implicitly allows the incoming count to be 0. From the Vulkan
650 * 1.0.38 spec, Section 4.1 Physical Devices:
652 * If the value referenced by pQueueFamilyPropertyCount is not 0 [then
658 *pQueueFamilyProperties
= (VkQueueFamilyProperties
) {
659 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
660 VK_QUEUE_COMPUTE_BIT
|
661 VK_QUEUE_TRANSFER_BIT
,
663 .timestampValidBits
= 36, /* XXX: Real value here */
664 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
670 void anv_GetPhysicalDeviceMemoryProperties(
671 VkPhysicalDevice physicalDevice
,
672 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
674 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
675 VkDeviceSize heap_size
;
677 /* Reserve some wiggle room for the driver by exposing only 75% of the
678 * aperture to the heap.
680 heap_size
= 3 * physical_device
->aperture_size
/ 4;
682 if (physical_device
->info
.has_llc
) {
683 /* Big core GPUs share LLC with the CPU and thus one memory type can be
684 * both cached and coherent at the same time.
686 pMemoryProperties
->memoryTypeCount
= 1;
687 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
688 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
689 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
690 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
691 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
695 /* The spec requires that we expose a host-visible, coherent memory
696 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
697 * to give the application a choice between cached, but not coherent and
698 * coherent but uncached (WC though).
700 pMemoryProperties
->memoryTypeCount
= 2;
701 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
702 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
703 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
704 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
707 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
708 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
709 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
710 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
715 pMemoryProperties
->memoryHeapCount
= 1;
716 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
718 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
722 PFN_vkVoidFunction
anv_GetInstanceProcAddr(
726 return anv_lookup_entrypoint(NULL
, pName
);
729 /* With version 1+ of the loader interface the ICD should expose
730 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
733 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
738 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
742 return anv_GetInstanceProcAddr(instance
, pName
);
745 PFN_vkVoidFunction
anv_GetDeviceProcAddr(
749 ANV_FROM_HANDLE(anv_device
, device
, _device
);
750 return anv_lookup_entrypoint(&device
->info
, pName
);
754 anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
)
756 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
757 queue
->device
= device
;
758 queue
->pool
= &device
->surface_state_pool
;
762 anv_queue_finish(struct anv_queue
*queue
)
766 static struct anv_state
767 anv_state_pool_emit_data(struct anv_state_pool
*pool
, size_t size
, size_t align
, const void *p
)
769 struct anv_state state
;
771 state
= anv_state_pool_alloc(pool
, size
, align
);
772 memcpy(state
.map
, p
, size
);
774 if (!pool
->block_pool
->device
->info
.has_llc
)
775 anv_state_clflush(state
);
780 struct gen8_border_color
{
785 /* Pad out to 64 bytes */
790 anv_device_init_border_colors(struct anv_device
*device
)
792 static const struct gen8_border_color border_colors
[] = {
793 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 0.0 } },
794 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
] = { .float32
= { 0.0, 0.0, 0.0, 1.0 } },
795 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
] = { .float32
= { 1.0, 1.0, 1.0, 1.0 } },
796 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
] = { .uint32
= { 0, 0, 0, 0 } },
797 [VK_BORDER_COLOR_INT_OPAQUE_BLACK
] = { .uint32
= { 0, 0, 0, 1 } },
798 [VK_BORDER_COLOR_INT_OPAQUE_WHITE
] = { .uint32
= { 1, 1, 1, 1 } },
801 device
->border_colors
= anv_state_pool_emit_data(&device
->dynamic_state_pool
,
802 sizeof(border_colors
), 64,
807 anv_device_submit_simple_batch(struct anv_device
*device
,
808 struct anv_batch
*batch
)
810 struct drm_i915_gem_execbuffer2 execbuf
;
811 struct drm_i915_gem_exec_object2 exec2_objects
[1];
812 struct anv_bo bo
, *exec_bos
[1];
813 VkResult result
= VK_SUCCESS
;
818 /* Kernel driver requires 8 byte aligned batch length */
819 size
= align_u32(batch
->next
- batch
->start
, 8);
820 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &bo
, size
);
821 if (result
!= VK_SUCCESS
)
824 memcpy(bo
.map
, batch
->start
, size
);
825 if (!device
->info
.has_llc
)
826 anv_clflush_range(bo
.map
, size
);
829 exec2_objects
[0].handle
= bo
.gem_handle
;
830 exec2_objects
[0].relocation_count
= 0;
831 exec2_objects
[0].relocs_ptr
= 0;
832 exec2_objects
[0].alignment
= 0;
833 exec2_objects
[0].offset
= bo
.offset
;
834 exec2_objects
[0].flags
= 0;
835 exec2_objects
[0].rsvd1
= 0;
836 exec2_objects
[0].rsvd2
= 0;
838 execbuf
.buffers_ptr
= (uintptr_t) exec2_objects
;
839 execbuf
.buffer_count
= 1;
840 execbuf
.batch_start_offset
= 0;
841 execbuf
.batch_len
= size
;
842 execbuf
.cliprects_ptr
= 0;
843 execbuf
.num_cliprects
= 0;
848 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
849 execbuf
.rsvd1
= device
->context_id
;
852 result
= anv_device_execbuf(device
, &execbuf
, exec_bos
);
853 if (result
!= VK_SUCCESS
)
857 ret
= anv_gem_wait(device
, bo
.gem_handle
, &timeout
);
859 /* We don't know the real error. */
860 result
= vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
865 anv_bo_pool_free(&device
->batch_bo_pool
, &bo
);
870 VkResult
anv_CreateDevice(
871 VkPhysicalDevice physicalDevice
,
872 const VkDeviceCreateInfo
* pCreateInfo
,
873 const VkAllocationCallbacks
* pAllocator
,
876 ANV_FROM_HANDLE(anv_physical_device
, physical_device
, physicalDevice
);
878 struct anv_device
*device
;
880 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO
);
882 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
884 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
885 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
886 device_extensions
[j
].extensionName
) == 0) {
892 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
895 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
897 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
899 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
901 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
902 device
->instance
= physical_device
->instance
;
903 device
->chipset_id
= physical_device
->chipset_id
;
906 device
->alloc
= *pAllocator
;
908 device
->alloc
= physical_device
->instance
->alloc
;
910 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
911 device
->fd
= open(physical_device
->path
, O_RDWR
| O_CLOEXEC
);
912 if (device
->fd
== -1) {
913 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
917 device
->context_id
= anv_gem_create_context(device
);
918 if (device
->context_id
== -1) {
919 result
= vk_error(VK_ERROR_INITIALIZATION_FAILED
);
923 device
->info
= physical_device
->info
;
924 device
->isl_dev
= physical_device
->isl_dev
;
926 /* On Broadwell and later, we can use batch chaining to more efficiently
927 * implement growing command buffers. Prior to Haswell, the kernel
928 * command parser gets in the way and we have to fall back to growing
931 device
->can_chain_batches
= device
->info
.gen
>= 8;
933 device
->robust_buffer_access
= pCreateInfo
->pEnabledFeatures
&&
934 pCreateInfo
->pEnabledFeatures
->robustBufferAccess
;
936 pthread_mutex_init(&device
->mutex
, NULL
);
938 pthread_condattr_t condattr
;
939 pthread_condattr_init(&condattr
);
940 pthread_condattr_setclock(&condattr
, CLOCK_MONOTONIC
);
941 pthread_cond_init(&device
->queue_submit
, NULL
);
942 pthread_condattr_destroy(&condattr
);
944 anv_bo_pool_init(&device
->batch_bo_pool
, device
);
946 anv_block_pool_init(&device
->dynamic_state_block_pool
, device
, 16384);
948 anv_state_pool_init(&device
->dynamic_state_pool
,
949 &device
->dynamic_state_block_pool
);
951 anv_block_pool_init(&device
->instruction_block_pool
, device
, 128 * 1024);
952 anv_state_pool_init(&device
->instruction_state_pool
,
953 &device
->instruction_block_pool
);
955 anv_block_pool_init(&device
->surface_state_block_pool
, device
, 4096);
957 anv_state_pool_init(&device
->surface_state_pool
,
958 &device
->surface_state_block_pool
);
960 anv_bo_init_new(&device
->workaround_bo
, device
, 1024);
962 anv_scratch_pool_init(device
, &device
->scratch_pool
);
964 anv_queue_init(device
, &device
->queue
);
966 switch (device
->info
.gen
) {
968 if (!device
->info
.is_haswell
)
969 result
= gen7_init_device_state(device
);
971 result
= gen75_init_device_state(device
);
974 result
= gen8_init_device_state(device
);
977 result
= gen9_init_device_state(device
);
980 /* Shouldn't get here as we don't create physical devices for any other
982 unreachable("unhandled gen");
984 if (result
!= VK_SUCCESS
)
987 anv_device_init_blorp(device
);
989 anv_device_init_border_colors(device
);
991 *pDevice
= anv_device_to_handle(device
);
998 vk_free(&device
->alloc
, device
);
1003 void anv_DestroyDevice(
1005 const VkAllocationCallbacks
* pAllocator
)
1007 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1009 anv_device_finish_blorp(device
);
1011 anv_queue_finish(&device
->queue
);
1013 #ifdef HAVE_VALGRIND
1014 /* We only need to free these to prevent valgrind errors. The backing
1015 * BO will go away in a couple of lines so we don't actually leak.
1017 anv_state_pool_free(&device
->dynamic_state_pool
, device
->border_colors
);
1020 anv_scratch_pool_finish(device
, &device
->scratch_pool
);
1022 anv_gem_munmap(device
->workaround_bo
.map
, device
->workaround_bo
.size
);
1023 anv_gem_close(device
, device
->workaround_bo
.gem_handle
);
1025 anv_state_pool_finish(&device
->surface_state_pool
);
1026 anv_block_pool_finish(&device
->surface_state_block_pool
);
1027 anv_state_pool_finish(&device
->instruction_state_pool
);
1028 anv_block_pool_finish(&device
->instruction_block_pool
);
1029 anv_state_pool_finish(&device
->dynamic_state_pool
);
1030 anv_block_pool_finish(&device
->dynamic_state_block_pool
);
1032 anv_bo_pool_finish(&device
->batch_bo_pool
);
1034 pthread_cond_destroy(&device
->queue_submit
);
1035 pthread_mutex_destroy(&device
->mutex
);
1037 anv_gem_destroy_context(device
, device
->context_id
);
1041 vk_free(&device
->alloc
, device
);
1044 VkResult
anv_EnumerateInstanceExtensionProperties(
1045 const char* pLayerName
,
1046 uint32_t* pPropertyCount
,
1047 VkExtensionProperties
* pProperties
)
1049 if (pProperties
== NULL
) {
1050 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
1054 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(global_extensions
));
1055 typed_memcpy(pProperties
, global_extensions
, *pPropertyCount
);
1057 if (*pPropertyCount
< ARRAY_SIZE(global_extensions
))
1058 return VK_INCOMPLETE
;
1063 VkResult
anv_EnumerateDeviceExtensionProperties(
1064 VkPhysicalDevice physicalDevice
,
1065 const char* pLayerName
,
1066 uint32_t* pPropertyCount
,
1067 VkExtensionProperties
* pProperties
)
1069 if (pProperties
== NULL
) {
1070 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
1074 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(device_extensions
));
1075 typed_memcpy(pProperties
, device_extensions
, *pPropertyCount
);
1077 if (*pPropertyCount
< ARRAY_SIZE(device_extensions
))
1078 return VK_INCOMPLETE
;
1083 VkResult
anv_EnumerateInstanceLayerProperties(
1084 uint32_t* pPropertyCount
,
1085 VkLayerProperties
* pProperties
)
1087 if (pProperties
== NULL
) {
1088 *pPropertyCount
= 0;
1092 /* None supported at this time */
1093 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1096 VkResult
anv_EnumerateDeviceLayerProperties(
1097 VkPhysicalDevice physicalDevice
,
1098 uint32_t* pPropertyCount
,
1099 VkLayerProperties
* pProperties
)
1101 if (pProperties
== NULL
) {
1102 *pPropertyCount
= 0;
1106 /* None supported at this time */
1107 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1110 void anv_GetDeviceQueue(
1112 uint32_t queueNodeIndex
,
1113 uint32_t queueIndex
,
1116 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1118 assert(queueIndex
== 0);
1120 *pQueue
= anv_queue_to_handle(&device
->queue
);
1124 anv_device_execbuf(struct anv_device
*device
,
1125 struct drm_i915_gem_execbuffer2
*execbuf
,
1126 struct anv_bo
**execbuf_bos
)
1128 int ret
= anv_gem_execbuffer(device
, execbuf
);
1130 /* We don't know the real error. */
1131 return vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
1134 struct drm_i915_gem_exec_object2
*objects
=
1135 (void *)(uintptr_t)execbuf
->buffers_ptr
;
1136 for (uint32_t k
= 0; k
< execbuf
->buffer_count
; k
++)
1137 execbuf_bos
[k
]->offset
= objects
[k
].offset
;
1142 VkResult
anv_QueueSubmit(
1144 uint32_t submitCount
,
1145 const VkSubmitInfo
* pSubmits
,
1148 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1149 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1150 struct anv_device
*device
= queue
->device
;
1151 VkResult result
= VK_SUCCESS
;
1153 /* We lock around QueueSubmit for three main reasons:
1155 * 1) When a block pool is resized, we create a new gem handle with a
1156 * different size and, in the case of surface states, possibly a
1157 * different center offset but we re-use the same anv_bo struct when
1158 * we do so. If this happens in the middle of setting up an execbuf,
1159 * we could end up with our list of BOs out of sync with our list of
1162 * 2) The algorithm we use for building the list of unique buffers isn't
1163 * thread-safe. While the client is supposed to syncronize around
1164 * QueueSubmit, this would be extremely difficult to debug if it ever
1165 * came up in the wild due to a broken app. It's better to play it
1166 * safe and just lock around QueueSubmit.
1168 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
1169 * userspace. Due to the fact that the surface state buffer is shared
1170 * between batches, we can't afford to have that happen from multiple
1171 * threads at the same time. Even though the user is supposed to
1172 * ensure this doesn't happen, we play it safe as in (2) above.
1174 * Since the only other things that ever take the device lock such as block
1175 * pool resize only rarely happen, this will almost never be contended so
1176 * taking a lock isn't really an expensive operation in this case.
1178 pthread_mutex_lock(&device
->mutex
);
1180 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1181 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1182 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
,
1183 pSubmits
[i
].pCommandBuffers
[j
]);
1184 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1186 result
= anv_cmd_buffer_execbuf(device
, cmd_buffer
);
1187 if (result
!= VK_SUCCESS
)
1193 struct anv_bo
*fence_bo
= &fence
->bo
;
1194 result
= anv_device_execbuf(device
, &fence
->execbuf
, &fence_bo
);
1195 if (result
!= VK_SUCCESS
)
1198 /* Update the fence and wake up any waiters */
1199 assert(fence
->state
== ANV_FENCE_STATE_RESET
);
1200 fence
->state
= ANV_FENCE_STATE_SUBMITTED
;
1201 pthread_cond_broadcast(&device
->queue_submit
);
1205 pthread_mutex_unlock(&device
->mutex
);
1210 VkResult
anv_QueueWaitIdle(
1213 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
1215 return anv_DeviceWaitIdle(anv_device_to_handle(queue
->device
));
1218 VkResult
anv_DeviceWaitIdle(
1221 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1222 struct anv_batch batch
;
1225 batch
.start
= batch
.next
= cmds
;
1226 batch
.end
= (void *) cmds
+ sizeof(cmds
);
1228 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1229 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1231 return anv_device_submit_simple_batch(device
, &batch
);
1235 anv_bo_init_new(struct anv_bo
*bo
, struct anv_device
*device
, uint64_t size
)
1237 uint32_t gem_handle
= anv_gem_create(device
, size
);
1239 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
1241 anv_bo_init(bo
, gem_handle
, size
);
1246 VkResult
anv_AllocateMemory(
1248 const VkMemoryAllocateInfo
* pAllocateInfo
,
1249 const VkAllocationCallbacks
* pAllocator
,
1250 VkDeviceMemory
* pMem
)
1252 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1253 struct anv_device_memory
*mem
;
1256 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1258 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1259 assert(pAllocateInfo
->allocationSize
> 0);
1261 /* We support exactly one memory heap. */
1262 assert(pAllocateInfo
->memoryTypeIndex
== 0 ||
1263 (!device
->info
.has_llc
&& pAllocateInfo
->memoryTypeIndex
< 2));
1265 /* FINISHME: Fail if allocation request exceeds heap size. */
1267 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1268 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1270 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1272 /* The kernel is going to give us whole pages anyway */
1273 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
1275 result
= anv_bo_init_new(&mem
->bo
, device
, alloc_size
);
1276 if (result
!= VK_SUCCESS
)
1279 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1284 *pMem
= anv_device_memory_to_handle(mem
);
1289 vk_free2(&device
->alloc
, pAllocator
, mem
);
1294 void anv_FreeMemory(
1296 VkDeviceMemory _mem
,
1297 const VkAllocationCallbacks
* pAllocator
)
1299 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1300 ANV_FROM_HANDLE(anv_device_memory
, mem
, _mem
);
1306 anv_UnmapMemory(_device
, _mem
);
1309 anv_gem_munmap(mem
->bo
.map
, mem
->bo
.size
);
1311 if (mem
->bo
.gem_handle
!= 0)
1312 anv_gem_close(device
, mem
->bo
.gem_handle
);
1314 vk_free2(&device
->alloc
, pAllocator
, mem
);
1317 VkResult
anv_MapMemory(
1319 VkDeviceMemory _memory
,
1320 VkDeviceSize offset
,
1322 VkMemoryMapFlags flags
,
1325 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1326 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1333 if (size
== VK_WHOLE_SIZE
)
1334 size
= mem
->bo
.size
- offset
;
1336 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1338 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1339 * assert(size != 0);
1340 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1341 * equal to the size of the memory minus offset
1344 assert(offset
+ size
<= mem
->bo
.size
);
1346 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1347 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1348 * at a time is valid. We could just mmap up front and return an offset
1349 * pointer here, but that may exhaust virtual memory on 32 bit
1352 uint32_t gem_flags
= 0;
1353 if (!device
->info
.has_llc
&& mem
->type_index
== 0)
1354 gem_flags
|= I915_MMAP_WC
;
1356 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1357 uint64_t map_offset
= offset
& ~4095ull;
1358 assert(offset
>= map_offset
);
1359 uint64_t map_size
= (offset
+ size
) - map_offset
;
1361 /* Let's map whole pages */
1362 map_size
= align_u64(map_size
, 4096);
1364 void *map
= anv_gem_mmap(device
, mem
->bo
.gem_handle
,
1365 map_offset
, map_size
, gem_flags
);
1366 if (map
== MAP_FAILED
)
1367 return vk_error(VK_ERROR_MEMORY_MAP_FAILED
);
1370 mem
->map_size
= map_size
;
1372 *ppData
= mem
->map
+ (offset
- map_offset
);
1377 void anv_UnmapMemory(
1379 VkDeviceMemory _memory
)
1381 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1386 anv_gem_munmap(mem
->map
, mem
->map_size
);
1393 clflush_mapped_ranges(struct anv_device
*device
,
1395 const VkMappedMemoryRange
*ranges
)
1397 for (uint32_t i
= 0; i
< count
; i
++) {
1398 ANV_FROM_HANDLE(anv_device_memory
, mem
, ranges
[i
].memory
);
1399 void *p
= mem
->map
+ (ranges
[i
].offset
& ~CACHELINE_MASK
);
1402 if (ranges
[i
].offset
+ ranges
[i
].size
> mem
->map_size
)
1403 end
= mem
->map
+ mem
->map_size
;
1405 end
= mem
->map
+ ranges
[i
].offset
+ ranges
[i
].size
;
1408 __builtin_ia32_clflush(p
);
1409 p
+= CACHELINE_SIZE
;
1414 VkResult
anv_FlushMappedMemoryRanges(
1416 uint32_t memoryRangeCount
,
1417 const VkMappedMemoryRange
* pMemoryRanges
)
1419 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1421 if (device
->info
.has_llc
)
1424 /* Make sure the writes we're flushing have landed. */
1425 __builtin_ia32_mfence();
1427 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1432 VkResult
anv_InvalidateMappedMemoryRanges(
1434 uint32_t memoryRangeCount
,
1435 const VkMappedMemoryRange
* pMemoryRanges
)
1437 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1439 if (device
->info
.has_llc
)
1442 clflush_mapped_ranges(device
, memoryRangeCount
, pMemoryRanges
);
1444 /* Make sure no reads get moved up above the invalidate. */
1445 __builtin_ia32_mfence();
1450 void anv_GetBufferMemoryRequirements(
1453 VkMemoryRequirements
* pMemoryRequirements
)
1455 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1457 /* The Vulkan spec (git aaed022) says:
1459 * memoryTypeBits is a bitfield and contains one bit set for every
1460 * supported memory type for the resource. The bit `1<<i` is set if and
1461 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1462 * structure for the physical device is supported.
1464 * We support exactly one memory type.
1466 pMemoryRequirements
->memoryTypeBits
= 1;
1468 pMemoryRequirements
->size
= buffer
->size
;
1469 pMemoryRequirements
->alignment
= 16;
1472 void anv_GetImageMemoryRequirements(
1475 VkMemoryRequirements
* pMemoryRequirements
)
1477 ANV_FROM_HANDLE(anv_image
, image
, _image
);
1479 /* The Vulkan spec (git aaed022) says:
1481 * memoryTypeBits is a bitfield and contains one bit set for every
1482 * supported memory type for the resource. The bit `1<<i` is set if and
1483 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1484 * structure for the physical device is supported.
1486 * We support exactly one memory type.
1488 pMemoryRequirements
->memoryTypeBits
= 1;
1490 pMemoryRequirements
->size
= image
->size
;
1491 pMemoryRequirements
->alignment
= image
->alignment
;
1494 void anv_GetImageSparseMemoryRequirements(
1497 uint32_t* pSparseMemoryRequirementCount
,
1498 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1503 void anv_GetDeviceMemoryCommitment(
1505 VkDeviceMemory memory
,
1506 VkDeviceSize
* pCommittedMemoryInBytes
)
1508 *pCommittedMemoryInBytes
= 0;
1511 VkResult
anv_BindBufferMemory(
1514 VkDeviceMemory _memory
,
1515 VkDeviceSize memoryOffset
)
1517 ANV_FROM_HANDLE(anv_device_memory
, mem
, _memory
);
1518 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1521 buffer
->bo
= &mem
->bo
;
1522 buffer
->offset
= memoryOffset
;
1531 VkResult
anv_QueueBindSparse(
1533 uint32_t bindInfoCount
,
1534 const VkBindSparseInfo
* pBindInfo
,
1537 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1540 VkResult
anv_CreateFence(
1542 const VkFenceCreateInfo
* pCreateInfo
,
1543 const VkAllocationCallbacks
* pAllocator
,
1546 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1547 struct anv_bo fence_bo
;
1548 struct anv_fence
*fence
;
1549 struct anv_batch batch
;
1552 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_CREATE_INFO
);
1554 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &fence_bo
, 4096);
1555 if (result
!= VK_SUCCESS
)
1558 /* Fences are small. Just store the CPU data structure in the BO. */
1559 fence
= fence_bo
.map
;
1560 fence
->bo
= fence_bo
;
1562 /* Place the batch after the CPU data but on its own cache line. */
1563 const uint32_t batch_offset
= align_u32(sizeof(*fence
), CACHELINE_SIZE
);
1564 batch
.next
= batch
.start
= fence
->bo
.map
+ batch_offset
;
1565 batch
.end
= fence
->bo
.map
+ fence
->bo
.size
;
1566 anv_batch_emit(&batch
, GEN7_MI_BATCH_BUFFER_END
, bbe
);
1567 anv_batch_emit(&batch
, GEN7_MI_NOOP
, noop
);
1569 if (!device
->info
.has_llc
) {
1570 assert(((uintptr_t) batch
.start
& CACHELINE_MASK
) == 0);
1571 assert(batch
.next
- batch
.start
<= CACHELINE_SIZE
);
1572 __builtin_ia32_mfence();
1573 __builtin_ia32_clflush(batch
.start
);
1576 fence
->exec2_objects
[0].handle
= fence
->bo
.gem_handle
;
1577 fence
->exec2_objects
[0].relocation_count
= 0;
1578 fence
->exec2_objects
[0].relocs_ptr
= 0;
1579 fence
->exec2_objects
[0].alignment
= 0;
1580 fence
->exec2_objects
[0].offset
= fence
->bo
.offset
;
1581 fence
->exec2_objects
[0].flags
= 0;
1582 fence
->exec2_objects
[0].rsvd1
= 0;
1583 fence
->exec2_objects
[0].rsvd2
= 0;
1585 fence
->execbuf
.buffers_ptr
= (uintptr_t) fence
->exec2_objects
;
1586 fence
->execbuf
.buffer_count
= 1;
1587 fence
->execbuf
.batch_start_offset
= batch
.start
- fence
->bo
.map
;
1588 fence
->execbuf
.batch_len
= batch
.next
- batch
.start
;
1589 fence
->execbuf
.cliprects_ptr
= 0;
1590 fence
->execbuf
.num_cliprects
= 0;
1591 fence
->execbuf
.DR1
= 0;
1592 fence
->execbuf
.DR4
= 0;
1594 fence
->execbuf
.flags
=
1595 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
1596 fence
->execbuf
.rsvd1
= device
->context_id
;
1597 fence
->execbuf
.rsvd2
= 0;
1599 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
1600 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1602 fence
->state
= ANV_FENCE_STATE_RESET
;
1605 *pFence
= anv_fence_to_handle(fence
);
1610 void anv_DestroyFence(
1613 const VkAllocationCallbacks
* pAllocator
)
1615 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1616 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1621 assert(fence
->bo
.map
== fence
);
1622 anv_bo_pool_free(&device
->batch_bo_pool
, &fence
->bo
);
1625 VkResult
anv_ResetFences(
1627 uint32_t fenceCount
,
1628 const VkFence
* pFences
)
1630 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1631 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1632 fence
->state
= ANV_FENCE_STATE_RESET
;
1638 VkResult
anv_GetFenceStatus(
1642 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1643 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
1647 switch (fence
->state
) {
1648 case ANV_FENCE_STATE_RESET
:
1649 /* If it hasn't even been sent off to the GPU yet, it's not ready */
1650 return VK_NOT_READY
;
1652 case ANV_FENCE_STATE_SIGNALED
:
1653 /* It's been signaled, return success */
1656 case ANV_FENCE_STATE_SUBMITTED
:
1657 /* It's been submitted to the GPU but we don't know if it's done yet. */
1658 ret
= anv_gem_wait(device
, fence
->bo
.gem_handle
, &t
);
1660 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1663 return VK_NOT_READY
;
1666 unreachable("Invalid fence status");
1670 #define NSEC_PER_SEC 1000000000
1671 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1673 VkResult
anv_WaitForFences(
1675 uint32_t fenceCount
,
1676 const VkFence
* pFences
,
1680 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1683 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1684 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1685 * for a couple of kernel releases. Since there's no way to know
1686 * whether or not the kernel we're using is one of the broken ones, the
1687 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1688 * maximum timeout from 584 years to 292 years - likely not a big deal.
1690 int64_t timeout
= MIN2(_timeout
, INT64_MAX
);
1692 uint32_t pending_fences
= fenceCount
;
1693 while (pending_fences
) {
1695 bool signaled_fences
= false;
1696 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1697 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1698 switch (fence
->state
) {
1699 case ANV_FENCE_STATE_RESET
:
1700 /* This fence hasn't been submitted yet, we'll catch it the next
1701 * time around. Yes, this may mean we dead-loop but, short of
1702 * lots of locking and a condition variable, there's not much that
1703 * we can do about that.
1708 case ANV_FENCE_STATE_SIGNALED
:
1709 /* This fence is not pending. If waitAll isn't set, we can return
1710 * early. Otherwise, we have to keep going.
1716 case ANV_FENCE_STATE_SUBMITTED
:
1717 /* These are the fences we really care about. Go ahead and wait
1718 * on it until we hit a timeout.
1720 ret
= anv_gem_wait(device
, fence
->bo
.gem_handle
, &timeout
);
1721 if (ret
== -1 && errno
== ETIME
) {
1723 } else if (ret
== -1) {
1724 /* We don't know the real error. */
1725 return vk_errorf(VK_ERROR_DEVICE_LOST
, "gem wait failed: %m");
1727 fence
->state
= ANV_FENCE_STATE_SIGNALED
;
1728 signaled_fences
= true;
1736 if (pending_fences
&& !signaled_fences
) {
1737 /* If we've hit this then someone decided to vkWaitForFences before
1738 * they've actually submitted any of them to a queue. This is a
1739 * fairly pessimal case, so it's ok to lock here and use a standard
1740 * pthreads condition variable.
1742 pthread_mutex_lock(&device
->mutex
);
1744 /* It's possible that some of the fences have changed state since the
1745 * last time we checked. Now that we have the lock, check for
1746 * pending fences again and don't wait if it's changed.
1748 uint32_t now_pending_fences
= 0;
1749 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
1750 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
1751 if (fence
->state
== ANV_FENCE_STATE_RESET
)
1752 now_pending_fences
++;
1754 assert(now_pending_fences
<= pending_fences
);
1756 if (now_pending_fences
== pending_fences
) {
1757 struct timespec before
;
1758 clock_gettime(CLOCK_MONOTONIC
, &before
);
1760 uint32_t abs_nsec
= before
.tv_nsec
+ timeout
% NSEC_PER_SEC
;
1761 uint64_t abs_sec
= before
.tv_sec
+ (abs_nsec
/ NSEC_PER_SEC
) +
1762 (timeout
/ NSEC_PER_SEC
);
1763 abs_nsec
%= NSEC_PER_SEC
;
1765 /* Avoid roll-over in tv_sec on 32-bit systems if the user
1766 * provided timeout is UINT64_MAX
1768 struct timespec abstime
;
1769 abstime
.tv_nsec
= abs_nsec
;
1770 abstime
.tv_sec
= MIN2(abs_sec
, INT_TYPE_MAX(abstime
.tv_sec
));
1772 ret
= pthread_cond_timedwait(&device
->queue_submit
,
1773 &device
->mutex
, &abstime
);
1774 assert(ret
!= EINVAL
);
1776 struct timespec after
;
1777 clock_gettime(CLOCK_MONOTONIC
, &after
);
1778 uint64_t time_elapsed
=
1779 ((uint64_t)after
.tv_sec
* NSEC_PER_SEC
+ after
.tv_nsec
) -
1780 ((uint64_t)before
.tv_sec
* NSEC_PER_SEC
+ before
.tv_nsec
);
1782 if (time_elapsed
>= timeout
) {
1783 pthread_mutex_unlock(&device
->mutex
);
1787 timeout
-= time_elapsed
;
1790 pthread_mutex_unlock(&device
->mutex
);
1797 // Queue semaphore functions
1799 VkResult
anv_CreateSemaphore(
1801 const VkSemaphoreCreateInfo
* pCreateInfo
,
1802 const VkAllocationCallbacks
* pAllocator
,
1803 VkSemaphore
* pSemaphore
)
1805 /* The DRM execbuffer ioctl always execute in-oder, even between different
1806 * rings. As such, there's nothing to do for the user space semaphore.
1809 *pSemaphore
= (VkSemaphore
)1;
1814 void anv_DestroySemaphore(
1816 VkSemaphore semaphore
,
1817 const VkAllocationCallbacks
* pAllocator
)
1823 VkResult
anv_CreateEvent(
1825 const VkEventCreateInfo
* pCreateInfo
,
1826 const VkAllocationCallbacks
* pAllocator
,
1829 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1830 struct anv_state state
;
1831 struct anv_event
*event
;
1833 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_EVENT_CREATE_INFO
);
1835 state
= anv_state_pool_alloc(&device
->dynamic_state_pool
,
1838 event
->state
= state
;
1839 event
->semaphore
= VK_EVENT_RESET
;
1841 if (!device
->info
.has_llc
) {
1842 /* Make sure the writes we're flushing have landed. */
1843 __builtin_ia32_mfence();
1844 __builtin_ia32_clflush(event
);
1847 *pEvent
= anv_event_to_handle(event
);
1852 void anv_DestroyEvent(
1855 const VkAllocationCallbacks
* pAllocator
)
1857 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1858 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1863 anv_state_pool_free(&device
->dynamic_state_pool
, event
->state
);
1866 VkResult
anv_GetEventStatus(
1870 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1871 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1873 if (!device
->info
.has_llc
) {
1874 /* Invalidate read cache before reading event written by GPU. */
1875 __builtin_ia32_clflush(event
);
1876 __builtin_ia32_mfence();
1880 return event
->semaphore
;
1883 VkResult
anv_SetEvent(
1887 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1888 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1890 event
->semaphore
= VK_EVENT_SET
;
1892 if (!device
->info
.has_llc
) {
1893 /* Make sure the writes we're flushing have landed. */
1894 __builtin_ia32_mfence();
1895 __builtin_ia32_clflush(event
);
1901 VkResult
anv_ResetEvent(
1905 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1906 ANV_FROM_HANDLE(anv_event
, event
, _event
);
1908 event
->semaphore
= VK_EVENT_RESET
;
1910 if (!device
->info
.has_llc
) {
1911 /* Make sure the writes we're flushing have landed. */
1912 __builtin_ia32_mfence();
1913 __builtin_ia32_clflush(event
);
1921 VkResult
anv_CreateBuffer(
1923 const VkBufferCreateInfo
* pCreateInfo
,
1924 const VkAllocationCallbacks
* pAllocator
,
1927 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1928 struct anv_buffer
*buffer
;
1930 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
1932 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
1933 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1935 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1937 buffer
->size
= pCreateInfo
->size
;
1938 buffer
->usage
= pCreateInfo
->usage
;
1942 *pBuffer
= anv_buffer_to_handle(buffer
);
1947 void anv_DestroyBuffer(
1950 const VkAllocationCallbacks
* pAllocator
)
1952 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1953 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
1958 vk_free2(&device
->alloc
, pAllocator
, buffer
);
1962 anv_fill_buffer_surface_state(struct anv_device
*device
, struct anv_state state
,
1963 enum isl_format format
,
1964 uint32_t offset
, uint32_t range
, uint32_t stride
)
1966 isl_buffer_fill_state(&device
->isl_dev
, state
.map
,
1968 .mocs
= device
->default_mocs
,
1973 if (!device
->info
.has_llc
)
1974 anv_state_clflush(state
);
1977 void anv_DestroySampler(
1980 const VkAllocationCallbacks
* pAllocator
)
1982 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1983 ANV_FROM_HANDLE(anv_sampler
, sampler
, _sampler
);
1988 vk_free2(&device
->alloc
, pAllocator
, sampler
);
1991 VkResult
anv_CreateFramebuffer(
1993 const VkFramebufferCreateInfo
* pCreateInfo
,
1994 const VkAllocationCallbacks
* pAllocator
,
1995 VkFramebuffer
* pFramebuffer
)
1997 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1998 struct anv_framebuffer
*framebuffer
;
2000 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2002 size_t size
= sizeof(*framebuffer
) +
2003 sizeof(struct anv_image_view
*) * pCreateInfo
->attachmentCount
;
2004 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2005 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2006 if (framebuffer
== NULL
)
2007 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2009 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2010 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2011 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2012 framebuffer
->attachments
[i
] = anv_image_view_from_handle(_iview
);
2015 framebuffer
->width
= pCreateInfo
->width
;
2016 framebuffer
->height
= pCreateInfo
->height
;
2017 framebuffer
->layers
= pCreateInfo
->layers
;
2019 *pFramebuffer
= anv_framebuffer_to_handle(framebuffer
);
2024 void anv_DestroyFramebuffer(
2027 const VkAllocationCallbacks
* pAllocator
)
2029 ANV_FROM_HANDLE(anv_device
, device
, _device
);
2030 ANV_FROM_HANDLE(anv_framebuffer
, fb
, _fb
);
2035 vk_free2(&device
->alloc
, pAllocator
, fb
);