2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
34 #include "radv_private.h"
35 #include "util/strtod.h"
39 #include <amdgpu_drm.h>
40 #include "amdgpu_id.h"
41 #include "winsys/amdgpu/radv_amdgpu_winsys_public.h"
42 #include "ac_llvm_util.h"
43 #include "vk_format.h"
45 #include "util/debug.h"
46 struct radv_dispatch_table dtable
;
49 radv_get_function_timestamp(void *ptr
, uint32_t* timestamp
)
53 if (!dladdr(ptr
, &info
) || !info
.dli_fname
) {
56 if (stat(info
.dli_fname
, &st
)) {
59 *timestamp
= st
.st_mtim
.tv_sec
;
64 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
66 uint32_t mesa_timestamp
, llvm_timestamp
;
68 memset(uuid
, 0, VK_UUID_SIZE
);
69 if (radv_get_function_timestamp(radv_device_get_cache_uuid
, &mesa_timestamp
) ||
70 radv_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
, &llvm_timestamp
))
73 memcpy(uuid
, &mesa_timestamp
, 4);
74 memcpy((char*)uuid
+ 4, &llvm_timestamp
, 4);
75 memcpy((char*)uuid
+ 8, &f
, 2);
76 snprintf((char*)uuid
+ 10, VK_UUID_SIZE
- 10, "radv");
81 radv_physical_device_init(struct radv_physical_device
*device
,
82 struct radv_instance
*instance
,
86 drmVersionPtr version
;
89 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
91 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
92 "failed to open %s: %m", path
);
94 version
= drmGetVersion(fd
);
97 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
98 "failed to get version %s: %m", path
);
101 if (strcmp(version
->name
, "amdgpu")) {
102 drmFreeVersion(version
);
104 return VK_ERROR_INCOMPATIBLE_DRIVER
;
106 drmFreeVersion(version
);
108 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
109 device
->instance
= instance
;
110 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
111 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
113 device
->ws
= radv_amdgpu_winsys_create(fd
);
115 result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
118 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
119 result
= radv_init_wsi(device
);
120 if (result
!= VK_SUCCESS
) {
121 device
->ws
->destroy(device
->ws
);
125 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->uuid
)) {
126 radv_finish_wsi(device
);
127 device
->ws
->destroy(device
->ws
);
131 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
132 device
->name
= device
->rad_info
.name
;
141 radv_physical_device_finish(struct radv_physical_device
*device
)
143 radv_finish_wsi(device
);
144 device
->ws
->destroy(device
->ws
);
147 static const VkExtensionProperties global_extensions
[] = {
149 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
152 #ifdef VK_USE_PLATFORM_XCB_KHR
154 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
158 #ifdef VK_USE_PLATFORM_XLIB_KHR
160 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
164 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
166 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
172 static const VkExtensionProperties device_extensions
[] = {
174 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
178 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
182 .extensionName
= VK_AMD_NEGATIVE_VIEWPORT_HEIGHT_EXTENSION_NAME
,
188 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
189 VkSystemAllocationScope allocationScope
)
195 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
196 size_t align
, VkSystemAllocationScope allocationScope
)
198 return realloc(pOriginal
, size
);
202 default_free_func(void *pUserData
, void *pMemory
)
207 static const VkAllocationCallbacks default_alloc
= {
209 .pfnAllocation
= default_alloc_func
,
210 .pfnReallocation
= default_realloc_func
,
211 .pfnFree
= default_free_func
,
214 VkResult
radv_CreateInstance(
215 const VkInstanceCreateInfo
* pCreateInfo
,
216 const VkAllocationCallbacks
* pAllocator
,
217 VkInstance
* pInstance
)
219 struct radv_instance
*instance
;
221 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
223 uint32_t client_version
;
224 if (pCreateInfo
->pApplicationInfo
&&
225 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
226 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
228 client_version
= VK_MAKE_VERSION(1, 0, 0);
231 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
232 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
233 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
234 "Client requested version %d.%d.%d",
235 VK_VERSION_MAJOR(client_version
),
236 VK_VERSION_MINOR(client_version
),
237 VK_VERSION_PATCH(client_version
));
240 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
242 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
243 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
244 global_extensions
[j
].extensionName
) == 0) {
250 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
253 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
254 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
256 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
258 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
261 instance
->alloc
= *pAllocator
;
263 instance
->alloc
= default_alloc
;
265 instance
->apiVersion
= client_version
;
266 instance
->physicalDeviceCount
= -1;
270 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
272 *pInstance
= radv_instance_to_handle(instance
);
277 void radv_DestroyInstance(
278 VkInstance _instance
,
279 const VkAllocationCallbacks
* pAllocator
)
281 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
283 if (instance
->physicalDeviceCount
> 0) {
284 /* We support at most one physical device. */
285 assert(instance
->physicalDeviceCount
== 1);
286 radv_physical_device_finish(&instance
->physicalDevice
);
289 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
293 vk_free(&instance
->alloc
, instance
);
296 VkResult
radv_EnumeratePhysicalDevices(
297 VkInstance _instance
,
298 uint32_t* pPhysicalDeviceCount
,
299 VkPhysicalDevice
* pPhysicalDevices
)
301 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
304 if (instance
->physicalDeviceCount
< 0) {
306 for (unsigned i
= 0; i
< 8; i
++) {
307 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
308 result
= radv_physical_device_init(&instance
->physicalDevice
,
310 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
314 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
315 instance
->physicalDeviceCount
= 0;
316 } else if (result
== VK_SUCCESS
) {
317 instance
->physicalDeviceCount
= 1;
323 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
324 * otherwise it's an inout parameter.
326 * The Vulkan spec (git aaed022) says:
328 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
329 * that is initialized with the number of devices the application is
330 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
331 * an array of at least this many VkPhysicalDevice handles [...].
333 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
334 * overwrites the contents of the variable pointed to by
335 * pPhysicalDeviceCount with the number of physical devices in in the
336 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
337 * pPhysicalDeviceCount with the number of physical handles written to
340 if (!pPhysicalDevices
) {
341 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
342 } else if (*pPhysicalDeviceCount
>= 1) {
343 pPhysicalDevices
[0] = radv_physical_device_to_handle(&instance
->physicalDevice
);
344 *pPhysicalDeviceCount
= 1;
345 } else if (*pPhysicalDeviceCount
< instance
->physicalDeviceCount
) {
346 return VK_INCOMPLETE
;
348 *pPhysicalDeviceCount
= 0;
354 void radv_GetPhysicalDeviceFeatures(
355 VkPhysicalDevice physicalDevice
,
356 VkPhysicalDeviceFeatures
* pFeatures
)
358 // RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
360 memset(pFeatures
, 0, sizeof(*pFeatures
));
362 *pFeatures
= (VkPhysicalDeviceFeatures
) {
363 .robustBufferAccess
= true,
364 .fullDrawIndexUint32
= true,
365 .imageCubeArray
= true,
366 .independentBlend
= true,
367 .geometryShader
= false,
368 .tessellationShader
= false,
369 .sampleRateShading
= false,
370 .dualSrcBlend
= true,
372 .multiDrawIndirect
= true,
373 .drawIndirectFirstInstance
= true,
375 .depthBiasClamp
= true,
376 .fillModeNonSolid
= true,
381 .multiViewport
= false,
382 .samplerAnisotropy
= true,
383 .textureCompressionETC2
= false,
384 .textureCompressionASTC_LDR
= false,
385 .textureCompressionBC
= true,
386 .occlusionQueryPrecise
= true,
387 .pipelineStatisticsQuery
= false,
388 .vertexPipelineStoresAndAtomics
= true,
389 .fragmentStoresAndAtomics
= true,
390 .shaderTessellationAndGeometryPointSize
= true,
391 .shaderImageGatherExtended
= false,
392 .shaderStorageImageExtendedFormats
= false,
393 .shaderStorageImageMultisample
= false,
394 .shaderUniformBufferArrayDynamicIndexing
= true,
395 .shaderSampledImageArrayDynamicIndexing
= true,
396 .shaderStorageBufferArrayDynamicIndexing
= true,
397 .shaderStorageImageArrayDynamicIndexing
= true,
398 .shaderStorageImageReadWithoutFormat
= false,
399 .shaderStorageImageWriteWithoutFormat
= true,
400 .shaderClipDistance
= true,
401 .shaderCullDistance
= true,
402 .shaderFloat64
= false,
403 .shaderInt64
= false,
404 .shaderInt16
= false,
406 .variableMultisampleRate
= false,
407 .inheritedQueries
= false,
411 void radv_GetPhysicalDeviceProperties(
412 VkPhysicalDevice physicalDevice
,
413 VkPhysicalDeviceProperties
* pProperties
)
415 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
416 VkSampleCountFlags sample_counts
= 0xf;
417 VkPhysicalDeviceLimits limits
= {
418 .maxImageDimension1D
= (1 << 14),
419 .maxImageDimension2D
= (1 << 14),
420 .maxImageDimension3D
= (1 << 11),
421 .maxImageDimensionCube
= (1 << 14),
422 .maxImageArrayLayers
= (1 << 11),
423 .maxTexelBufferElements
= 128 * 1024 * 1024,
424 .maxUniformBufferRange
= UINT32_MAX
,
425 .maxStorageBufferRange
= UINT32_MAX
,
426 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
427 .maxMemoryAllocationCount
= UINT32_MAX
,
428 .maxSamplerAllocationCount
= 64 * 1024,
429 .bufferImageGranularity
= 64, /* A cache line */
430 .sparseAddressSpaceSize
= 0,
431 .maxBoundDescriptorSets
= MAX_SETS
,
432 .maxPerStageDescriptorSamplers
= 64,
433 .maxPerStageDescriptorUniformBuffers
= 64,
434 .maxPerStageDescriptorStorageBuffers
= 64,
435 .maxPerStageDescriptorSampledImages
= 64,
436 .maxPerStageDescriptorStorageImages
= 64,
437 .maxPerStageDescriptorInputAttachments
= 64,
438 .maxPerStageResources
= 128,
439 .maxDescriptorSetSamplers
= 256,
440 .maxDescriptorSetUniformBuffers
= 256,
441 .maxDescriptorSetUniformBuffersDynamic
= 256,
442 .maxDescriptorSetStorageBuffers
= 256,
443 .maxDescriptorSetStorageBuffersDynamic
= 256,
444 .maxDescriptorSetSampledImages
= 256,
445 .maxDescriptorSetStorageImages
= 256,
446 .maxDescriptorSetInputAttachments
= 256,
447 .maxVertexInputAttributes
= 32,
448 .maxVertexInputBindings
= 32,
449 .maxVertexInputAttributeOffset
= 2047,
450 .maxVertexInputBindingStride
= 2048,
451 .maxVertexOutputComponents
= 128,
452 .maxTessellationGenerationLevel
= 0,
453 .maxTessellationPatchSize
= 0,
454 .maxTessellationControlPerVertexInputComponents
= 0,
455 .maxTessellationControlPerVertexOutputComponents
= 0,
456 .maxTessellationControlPerPatchOutputComponents
= 0,
457 .maxTessellationControlTotalOutputComponents
= 0,
458 .maxTessellationEvaluationInputComponents
= 0,
459 .maxTessellationEvaluationOutputComponents
= 0,
460 .maxGeometryShaderInvocations
= 32,
461 .maxGeometryInputComponents
= 64,
462 .maxGeometryOutputComponents
= 128,
463 .maxGeometryOutputVertices
= 256,
464 .maxGeometryTotalOutputComponents
= 1024,
465 .maxFragmentInputComponents
= 128,
466 .maxFragmentOutputAttachments
= 8,
467 .maxFragmentDualSrcAttachments
= 2,
468 .maxFragmentCombinedOutputResources
= 8,
469 .maxComputeSharedMemorySize
= 32768,
470 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
471 .maxComputeWorkGroupInvocations
= 16 * 1024,
472 .maxComputeWorkGroupSize
= {
473 16 * 1024/*devinfo->max_cs_threads*/,
477 .subPixelPrecisionBits
= 4 /* FIXME */,
478 .subTexelPrecisionBits
= 4 /* FIXME */,
479 .mipmapPrecisionBits
= 4 /* FIXME */,
480 .maxDrawIndexedIndexValue
= UINT32_MAX
,
481 .maxDrawIndirectCount
= UINT32_MAX
,
482 .maxSamplerLodBias
= 16,
483 .maxSamplerAnisotropy
= 16,
484 .maxViewports
= MAX_VIEWPORTS
,
485 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
486 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
487 .viewportSubPixelBits
= 13, /* We take a float? */
488 .minMemoryMapAlignment
= 4096, /* A page */
489 .minTexelBufferOffsetAlignment
= 1,
490 .minUniformBufferOffsetAlignment
= 4,
491 .minStorageBufferOffsetAlignment
= 4,
492 .minTexelOffset
= -8,
494 .minTexelGatherOffset
= -8,
495 .maxTexelGatherOffset
= 7,
496 .minInterpolationOffset
= 0, /* FIXME */
497 .maxInterpolationOffset
= 0, /* FIXME */
498 .subPixelInterpolationOffsetBits
= 0, /* FIXME */
499 .maxFramebufferWidth
= (1 << 14),
500 .maxFramebufferHeight
= (1 << 14),
501 .maxFramebufferLayers
= (1 << 10),
502 .framebufferColorSampleCounts
= sample_counts
,
503 .framebufferDepthSampleCounts
= sample_counts
,
504 .framebufferStencilSampleCounts
= sample_counts
,
505 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
506 .maxColorAttachments
= MAX_RTS
,
507 .sampledImageColorSampleCounts
= sample_counts
,
508 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
509 .sampledImageDepthSampleCounts
= sample_counts
,
510 .sampledImageStencilSampleCounts
= sample_counts
,
511 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
512 .maxSampleMaskWords
= 1,
513 .timestampComputeAndGraphics
= false,
514 .timestampPeriod
= 100000.0 / pdevice
->rad_info
.clock_crystal_freq
,
515 .maxClipDistances
= 8,
516 .maxCullDistances
= 8,
517 .maxCombinedClipAndCullDistances
= 8,
518 .discreteQueuePriorities
= 1,
519 .pointSizeRange
= { 0.125, 255.875 },
520 .lineWidthRange
= { 0.0, 7.9921875 },
521 .pointSizeGranularity
= (1.0 / 8.0),
522 .lineWidthGranularity
= (1.0 / 128.0),
523 .strictLines
= false, /* FINISHME */
524 .standardSampleLocations
= true,
525 .optimalBufferCopyOffsetAlignment
= 128,
526 .optimalBufferCopyRowPitchAlignment
= 128,
527 .nonCoherentAtomSize
= 64,
530 *pProperties
= (VkPhysicalDeviceProperties
) {
531 .apiVersion
= VK_MAKE_VERSION(1, 0, 5),
534 .deviceID
= pdevice
->rad_info
.pci_id
,
535 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
,
537 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
540 strcpy(pProperties
->deviceName
, pdevice
->name
);
541 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
544 void radv_GetPhysicalDeviceQueueFamilyProperties(
545 VkPhysicalDevice physicalDevice
,
547 VkQueueFamilyProperties
* pQueueFamilyProperties
)
549 if (pQueueFamilyProperties
== NULL
) {
553 assert(*pCount
>= 1);
555 *pQueueFamilyProperties
= (VkQueueFamilyProperties
) {
556 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
557 VK_QUEUE_COMPUTE_BIT
|
558 VK_QUEUE_TRANSFER_BIT
,
560 .timestampValidBits
= 64,
561 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
565 void radv_GetPhysicalDeviceMemoryProperties(
566 VkPhysicalDevice physicalDevice
,
567 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
569 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
571 pMemoryProperties
->memoryTypeCount
= 4;
572 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
573 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
576 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
577 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
578 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
581 pMemoryProperties
->memoryTypes
[2] = (VkMemoryType
) {
582 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
583 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
584 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
587 pMemoryProperties
->memoryTypes
[3] = (VkMemoryType
) {
588 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
589 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
590 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
594 pMemoryProperties
->memoryHeapCount
= 3;
595 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
596 .size
= physical_device
->rad_info
.vram_size
-
597 physical_device
->rad_info
.visible_vram_size
,
598 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
600 pMemoryProperties
->memoryHeaps
[1] = (VkMemoryHeap
) {
601 .size
= physical_device
->rad_info
.visible_vram_size
,
602 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
604 pMemoryProperties
->memoryHeaps
[2] = (VkMemoryHeap
) {
605 .size
= physical_device
->rad_info
.gart_size
,
611 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
)
613 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
614 queue
->device
= device
;
620 radv_queue_finish(struct radv_queue
*queue
)
624 VkResult
radv_CreateDevice(
625 VkPhysicalDevice physicalDevice
,
626 const VkDeviceCreateInfo
* pCreateInfo
,
627 const VkAllocationCallbacks
* pAllocator
,
630 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
632 struct radv_device
*device
;
634 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
636 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
637 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
638 device_extensions
[j
].extensionName
) == 0) {
644 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
647 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
649 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
651 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
653 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
654 device
->instance
= physical_device
->instance
;
656 device
->ws
= physical_device
->ws
;
658 device
->alloc
= *pAllocator
;
660 device
->alloc
= physical_device
->instance
->alloc
;
662 device
->hw_ctx
= device
->ws
->ctx_create(device
->ws
);
663 if (!device
->hw_ctx
) {
664 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
668 radv_queue_init(device
, &device
->queue
);
670 result
= radv_device_init_meta(device
);
671 if (result
!= VK_SUCCESS
) {
672 device
->ws
->ctx_destroy(device
->hw_ctx
);
675 device
->allow_fast_clears
= env_var_as_boolean("RADV_FAST_CLEARS", false);
676 device
->allow_dcc
= !env_var_as_boolean("RADV_DCC_DISABLE", false);
677 device
->shader_stats_dump
= env_var_as_boolean("RADV_SHADER_STATS", false);
679 if (device
->allow_fast_clears
&& device
->allow_dcc
)
680 radv_finishme("DCC fast clears have not been tested\n");
682 radv_device_init_msaa(device
);
683 device
->empty_cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
684 radeon_emit(device
->empty_cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
685 radeon_emit(device
->empty_cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
686 radeon_emit(device
->empty_cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
687 device
->ws
->cs_finalize(device
->empty_cs
);
688 *pDevice
= radv_device_to_handle(device
);
691 vk_free(&device
->alloc
, device
);
695 void radv_DestroyDevice(
697 const VkAllocationCallbacks
* pAllocator
)
699 RADV_FROM_HANDLE(radv_device
, device
, _device
);
701 device
->ws
->ctx_destroy(device
->hw_ctx
);
702 radv_queue_finish(&device
->queue
);
703 radv_device_finish_meta(device
);
705 vk_free(&device
->alloc
, device
);
708 VkResult
radv_EnumerateInstanceExtensionProperties(
709 const char* pLayerName
,
710 uint32_t* pPropertyCount
,
711 VkExtensionProperties
* pProperties
)
714 if (pProperties
== NULL
) {
715 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
719 for (i
= 0; i
< *pPropertyCount
; i
++)
720 memcpy(&pProperties
[i
], &global_extensions
[i
], sizeof(VkExtensionProperties
));
723 if (i
< ARRAY_SIZE(global_extensions
))
724 return VK_INCOMPLETE
;
729 VkResult
radv_EnumerateDeviceExtensionProperties(
730 VkPhysicalDevice physicalDevice
,
731 const char* pLayerName
,
732 uint32_t* pPropertyCount
,
733 VkExtensionProperties
* pProperties
)
737 if (pProperties
== NULL
) {
738 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
742 for (i
= 0; i
< *pPropertyCount
; i
++)
743 memcpy(&pProperties
[i
], &device_extensions
[i
], sizeof(VkExtensionProperties
));
746 if (i
< ARRAY_SIZE(device_extensions
))
747 return VK_INCOMPLETE
;
751 VkResult
radv_EnumerateInstanceLayerProperties(
752 uint32_t* pPropertyCount
,
753 VkLayerProperties
* pProperties
)
755 if (pProperties
== NULL
) {
760 /* None supported at this time */
761 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
764 VkResult
radv_EnumerateDeviceLayerProperties(
765 VkPhysicalDevice physicalDevice
,
766 uint32_t* pPropertyCount
,
767 VkLayerProperties
* pProperties
)
769 if (pProperties
== NULL
) {
774 /* None supported at this time */
775 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
778 void radv_GetDeviceQueue(
780 uint32_t queueNodeIndex
,
784 RADV_FROM_HANDLE(radv_device
, device
, _device
);
786 assert(queueIndex
== 0);
788 *pQueue
= radv_queue_to_handle(&device
->queue
);
791 VkResult
radv_QueueSubmit(
793 uint32_t submitCount
,
794 const VkSubmitInfo
* pSubmits
,
797 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
798 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
799 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
800 struct radeon_winsys_ctx
*ctx
= queue
->device
->hw_ctx
;
803 for (uint32_t i
= 0; i
< submitCount
; i
++) {
804 struct radeon_winsys_cs
**cs_array
;
805 bool can_patch
= true;
807 if (!pSubmits
[i
].commandBufferCount
)
810 cs_array
= malloc(sizeof(struct radeon_winsys_cs
*) *
811 pSubmits
[i
].commandBufferCount
);
813 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
814 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
815 pSubmits
[i
].pCommandBuffers
[j
]);
816 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
818 cs_array
[j
] = cmd_buffer
->cs
;
819 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
822 ret
= queue
->device
->ws
->cs_submit(ctx
, cs_array
,
823 pSubmits
[i
].commandBufferCount
,
824 can_patch
, base_fence
);
826 radv_loge("failed to submit CS %d\n", i
);
832 ret
= queue
->device
->ws
->cs_submit(ctx
, &queue
->device
->empty_cs
,
833 1, false, base_fence
);
835 fence
->submitted
= true;
841 VkResult
radv_QueueWaitIdle(
844 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
846 queue
->device
->ws
->ctx_wait_idle(queue
->device
->hw_ctx
);
850 VkResult
radv_DeviceWaitIdle(
853 RADV_FROM_HANDLE(radv_device
, device
, _device
);
855 device
->ws
->ctx_wait_idle(device
->hw_ctx
);
859 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
863 return radv_lookup_entrypoint(pName
);
866 /* The loader wants us to expose a second GetInstanceProcAddr function
867 * to work around certain LD_PRELOAD issues seen in apps.
870 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
875 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
879 return radv_GetInstanceProcAddr(instance
, pName
);
882 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
886 return radv_lookup_entrypoint(pName
);
889 VkResult
radv_AllocateMemory(
891 const VkMemoryAllocateInfo
* pAllocateInfo
,
892 const VkAllocationCallbacks
* pAllocator
,
893 VkDeviceMemory
* pMem
)
895 RADV_FROM_HANDLE(radv_device
, device
, _device
);
896 struct radv_device_memory
*mem
;
898 enum radeon_bo_domain domain
;
900 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
902 if (pAllocateInfo
->allocationSize
== 0) {
903 /* Apparently, this is allowed */
904 *pMem
= VK_NULL_HANDLE
;
908 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
909 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
911 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
913 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
914 if (pAllocateInfo
->memoryTypeIndex
== 1 || pAllocateInfo
->memoryTypeIndex
== 3)
915 domain
= RADEON_DOMAIN_GTT
;
917 domain
= RADEON_DOMAIN_VRAM
;
919 if (pAllocateInfo
->memoryTypeIndex
== 0)
920 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
922 flags
|= RADEON_FLAG_CPU_ACCESS
;
924 if (pAllocateInfo
->memoryTypeIndex
== 1)
925 flags
|= RADEON_FLAG_GTT_WC
;
927 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, 32768,
931 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
934 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
936 *pMem
= radv_device_memory_to_handle(mem
);
941 vk_free2(&device
->alloc
, pAllocator
, mem
);
946 void radv_FreeMemory(
949 const VkAllocationCallbacks
* pAllocator
)
951 RADV_FROM_HANDLE(radv_device
, device
, _device
);
952 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
957 device
->ws
->buffer_destroy(mem
->bo
);
960 vk_free2(&device
->alloc
, pAllocator
, mem
);
963 VkResult
radv_MapMemory(
965 VkDeviceMemory _memory
,
968 VkMemoryMapFlags flags
,
971 RADV_FROM_HANDLE(radv_device
, device
, _device
);
972 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
979 *ppData
= device
->ws
->buffer_map(mem
->bo
);
985 return VK_ERROR_MEMORY_MAP_FAILED
;
988 void radv_UnmapMemory(
990 VkDeviceMemory _memory
)
992 RADV_FROM_HANDLE(radv_device
, device
, _device
);
993 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
998 device
->ws
->buffer_unmap(mem
->bo
);
1001 VkResult
radv_FlushMappedMemoryRanges(
1003 uint32_t memoryRangeCount
,
1004 const VkMappedMemoryRange
* pMemoryRanges
)
1009 VkResult
radv_InvalidateMappedMemoryRanges(
1011 uint32_t memoryRangeCount
,
1012 const VkMappedMemoryRange
* pMemoryRanges
)
1017 void radv_GetBufferMemoryRequirements(
1020 VkMemoryRequirements
* pMemoryRequirements
)
1022 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1024 /* The Vulkan spec (git aaed022) says:
1026 * memoryTypeBits is a bitfield and contains one bit set for every
1027 * supported memory type for the resource. The bit `1<<i` is set if and
1028 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1029 * structure for the physical device is supported.
1031 * We support exactly one memory type.
1033 pMemoryRequirements
->memoryTypeBits
= 0x7;
1035 pMemoryRequirements
->size
= buffer
->size
;
1036 pMemoryRequirements
->alignment
= 16;
1039 void radv_GetImageMemoryRequirements(
1042 VkMemoryRequirements
* pMemoryRequirements
)
1044 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1046 /* The Vulkan spec (git aaed022) says:
1048 * memoryTypeBits is a bitfield and contains one bit set for every
1049 * supported memory type for the resource. The bit `1<<i` is set if and
1050 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1051 * structure for the physical device is supported.
1053 * We support exactly one memory type.
1055 pMemoryRequirements
->memoryTypeBits
= 0x7;
1057 pMemoryRequirements
->size
= image
->size
;
1058 pMemoryRequirements
->alignment
= image
->alignment
;
1061 void radv_GetImageSparseMemoryRequirements(
1064 uint32_t* pSparseMemoryRequirementCount
,
1065 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1070 void radv_GetDeviceMemoryCommitment(
1072 VkDeviceMemory memory
,
1073 VkDeviceSize
* pCommittedMemoryInBytes
)
1075 *pCommittedMemoryInBytes
= 0;
1078 VkResult
radv_BindBufferMemory(
1081 VkDeviceMemory _memory
,
1082 VkDeviceSize memoryOffset
)
1084 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1085 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1088 buffer
->bo
= mem
->bo
;
1089 buffer
->offset
= memoryOffset
;
1098 VkResult
radv_BindImageMemory(
1101 VkDeviceMemory _memory
,
1102 VkDeviceSize memoryOffset
)
1104 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1105 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1108 image
->bo
= mem
->bo
;
1109 image
->offset
= memoryOffset
;
1118 VkResult
radv_QueueBindSparse(
1120 uint32_t bindInfoCount
,
1121 const VkBindSparseInfo
* pBindInfo
,
1124 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1127 VkResult
radv_CreateFence(
1129 const VkFenceCreateInfo
* pCreateInfo
,
1130 const VkAllocationCallbacks
* pAllocator
,
1133 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1134 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
1136 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1139 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1141 memset(fence
, 0, sizeof(*fence
));
1142 fence
->submitted
= false;
1143 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
1144 fence
->fence
= device
->ws
->create_fence();
1147 *pFence
= radv_fence_to_handle(fence
);
1152 void radv_DestroyFence(
1155 const VkAllocationCallbacks
* pAllocator
)
1157 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1158 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1162 device
->ws
->destroy_fence(fence
->fence
);
1163 vk_free2(&device
->alloc
, pAllocator
, fence
);
1166 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
1168 uint64_t current_time
;
1171 clock_gettime(CLOCK_MONOTONIC
, &tv
);
1172 current_time
= tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
1174 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
1176 return current_time
+ timeout
;
1179 VkResult
radv_WaitForFences(
1181 uint32_t fenceCount
,
1182 const VkFence
* pFences
,
1186 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1187 timeout
= radv_get_absolute_timeout(timeout
);
1189 if (!waitAll
&& fenceCount
> 1) {
1190 fprintf(stderr
, "radv: WaitForFences without waitAll not implemented yet\n");
1193 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
1194 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1195 bool expired
= false;
1197 if (fence
->signalled
)
1200 if (!fence
->submitted
)
1203 expired
= device
->ws
->fence_wait(device
->ws
, fence
->fence
, true, timeout
);
1207 fence
->signalled
= true;
1213 VkResult
radv_ResetFences(VkDevice device
,
1214 uint32_t fenceCount
,
1215 const VkFence
*pFences
)
1217 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
1218 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1219 fence
->submitted
= fence
->signalled
= false;
1225 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
1227 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1228 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1230 if (fence
->signalled
)
1232 if (!fence
->submitted
)
1233 return VK_NOT_READY
;
1235 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
1236 return VK_NOT_READY
;
1242 // Queue semaphore functions
1244 VkResult
radv_CreateSemaphore(
1246 const VkSemaphoreCreateInfo
* pCreateInfo
,
1247 const VkAllocationCallbacks
* pAllocator
,
1248 VkSemaphore
* pSemaphore
)
1250 /* The DRM execbuffer ioctl always execute in-oder, even between different
1251 * rings. As such, there's nothing to do for the user space semaphore.
1254 *pSemaphore
= (VkSemaphore
)1;
1259 void radv_DestroySemaphore(
1261 VkSemaphore semaphore
,
1262 const VkAllocationCallbacks
* pAllocator
)
1266 VkResult
radv_CreateEvent(
1268 const VkEventCreateInfo
* pCreateInfo
,
1269 const VkAllocationCallbacks
* pAllocator
,
1272 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1273 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
1275 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1278 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1280 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
1282 RADEON_FLAG_CPU_ACCESS
);
1284 vk_free2(&device
->alloc
, pAllocator
, event
);
1285 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1288 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
1290 *pEvent
= radv_event_to_handle(event
);
1295 void radv_DestroyEvent(
1298 const VkAllocationCallbacks
* pAllocator
)
1300 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1301 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1305 device
->ws
->buffer_destroy(event
->bo
);
1306 vk_free2(&device
->alloc
, pAllocator
, event
);
1309 VkResult
radv_GetEventStatus(
1313 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1315 if (*event
->map
== 1)
1316 return VK_EVENT_SET
;
1317 return VK_EVENT_RESET
;
1320 VkResult
radv_SetEvent(
1324 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1330 VkResult
radv_ResetEvent(
1334 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1340 VkResult
radv_CreateBuffer(
1342 const VkBufferCreateInfo
* pCreateInfo
,
1343 const VkAllocationCallbacks
* pAllocator
,
1346 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1347 struct radv_buffer
*buffer
;
1349 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
1351 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
1352 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1354 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1356 buffer
->size
= pCreateInfo
->size
;
1357 buffer
->usage
= pCreateInfo
->usage
;
1361 *pBuffer
= radv_buffer_to_handle(buffer
);
1366 void radv_DestroyBuffer(
1369 const VkAllocationCallbacks
* pAllocator
)
1371 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1372 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1377 vk_free2(&device
->alloc
, pAllocator
, buffer
);
1380 static inline unsigned
1381 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
1384 return image
->surface
.stencil_tiling_index
[level
];
1386 return image
->surface
.tiling_index
[level
];
1390 radv_initialise_color_surface(struct radv_device
*device
,
1391 struct radv_color_buffer_info
*cb
,
1392 struct radv_image_view
*iview
)
1394 const struct vk_format_description
*desc
;
1395 unsigned ntype
, format
, swap
, endian
;
1396 unsigned blend_clamp
= 0, blend_bypass
= 0;
1397 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
1399 const struct radeon_surf
*surf
= &iview
->image
->surface
;
1400 const struct radeon_surf_level
*level_info
= &surf
->level
[iview
->base_mip
];
1402 desc
= vk_format_description(iview
->vk_format
);
1404 memset(cb
, 0, sizeof(*cb
));
1406 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1407 va
+= level_info
->offset
;
1408 cb
->cb_color_base
= va
>> 8;
1410 /* CMASK variables */
1411 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1412 va
+= iview
->image
->cmask
.offset
;
1413 cb
->cb_color_cmask
= va
>> 8;
1414 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
1416 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1417 va
+= iview
->image
->dcc_offset
;
1418 cb
->cb_dcc_base
= va
>> 8;
1420 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
1421 S_028C6C_SLICE_MAX(iview
->base_layer
+ iview
->extent
.depth
- 1);
1423 cb
->micro_tile_mode
= iview
->image
->surface
.micro_tile_mode
;
1424 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
1425 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
1426 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
1428 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
1429 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
1431 /* Intensity is implemented as Red, so treat it that way. */
1432 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
) |
1433 S_028C74_TILE_MODE_INDEX(tile_mode_index
);
1435 if (iview
->image
->samples
> 1) {
1436 unsigned log_samples
= util_logbase2(iview
->image
->samples
);
1438 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
1439 S_028C74_NUM_FRAGMENTS(log_samples
);
1442 if (iview
->image
->fmask
.size
) {
1443 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
1444 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
)
1445 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
1446 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
1447 cb
->cb_color_fmask
= va
>> 8;
1448 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
1450 /* This must be set for fast clear to work without FMASK. */
1451 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
)
1452 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
1453 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
1454 cb
->cb_color_fmask
= cb
->cb_color_base
;
1455 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
1458 ntype
= radv_translate_color_numformat(iview
->vk_format
,
1460 vk_format_get_first_non_void_channel(iview
->vk_format
));
1461 format
= radv_translate_colorformat(iview
->vk_format
);
1462 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
1463 radv_finishme("Illegal color\n");
1464 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
1465 endian
= radv_colorformat_endian_swap(format
);
1467 /* blend clamp should be set for all NORM/SRGB types */
1468 if (ntype
== V_028C70_NUMBER_UNORM
||
1469 ntype
== V_028C70_NUMBER_SNORM
||
1470 ntype
== V_028C70_NUMBER_SRGB
)
1473 /* set blend bypass according to docs if SINT/UINT or
1474 8/24 COLOR variants */
1475 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
1476 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
1477 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
1482 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
1483 (format
== V_028C70_COLOR_8
||
1484 format
== V_028C70_COLOR_8_8
||
1485 format
== V_028C70_COLOR_8_8_8_8
))
1486 ->color_is_int8
= true;
1488 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
1489 S_028C70_COMP_SWAP(swap
) |
1490 S_028C70_BLEND_CLAMP(blend_clamp
) |
1491 S_028C70_BLEND_BYPASS(blend_bypass
) |
1492 S_028C70_SIMPLE_FLOAT(1) |
1493 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
1494 ntype
!= V_028C70_NUMBER_SNORM
&&
1495 ntype
!= V_028C70_NUMBER_SRGB
&&
1496 format
!= V_028C70_COLOR_8_24
&&
1497 format
!= V_028C70_COLOR_24_8
) |
1498 S_028C70_NUMBER_TYPE(ntype
) |
1499 S_028C70_ENDIAN(endian
);
1500 if (iview
->image
->samples
> 1)
1501 if (iview
->image
->fmask
.size
)
1502 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
1504 if (iview
->image
->cmask
.size
&& device
->allow_fast_clears
)
1505 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
1507 if (iview
->image
->surface
.dcc_size
&& level_info
->dcc_enabled
)
1508 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
1510 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= VI
) {
1511 unsigned max_uncompressed_block_size
= 2;
1512 if (iview
->image
->samples
> 1) {
1513 if (iview
->image
->surface
.bpe
== 1)
1514 max_uncompressed_block_size
= 0;
1515 else if (iview
->image
->surface
.bpe
== 2)
1516 max_uncompressed_block_size
= 1;
1519 cb
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
1520 S_028C78_INDEPENDENT_64B_BLOCKS(1);
1523 /* This must be set for fast clear to work without FMASK. */
1524 if (!iview
->image
->fmask
.size
&&
1525 device
->instance
->physicalDevice
.rad_info
.chip_class
== SI
) {
1526 unsigned bankh
= util_logbase2(iview
->image
->surface
.bankh
);
1527 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
1532 radv_initialise_ds_surface(struct radv_device
*device
,
1533 struct radv_ds_buffer_info
*ds
,
1534 struct radv_image_view
*iview
)
1536 unsigned level
= iview
->base_mip
;
1538 uint64_t va
, s_offs
, z_offs
;
1539 const struct radeon_surf_level
*level_info
= &iview
->image
->surface
.level
[level
];
1540 memset(ds
, 0, sizeof(*ds
));
1541 switch (iview
->vk_format
) {
1542 case VK_FORMAT_D24_UNORM_S8_UINT
:
1543 case VK_FORMAT_X8_D24_UNORM_PACK32
:
1544 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
1545 ds
->offset_scale
= 2.0f
;
1547 case VK_FORMAT_D16_UNORM
:
1548 case VK_FORMAT_D16_UNORM_S8_UINT
:
1549 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
1550 ds
->offset_scale
= 4.0f
;
1552 case VK_FORMAT_D32_SFLOAT
:
1553 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
1554 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
1555 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
1556 ds
->offset_scale
= 1.0f
;
1562 format
= radv_translate_dbformat(iview
->vk_format
);
1563 if (format
== V_028040_Z_INVALID
) {
1564 fprintf(stderr
, "Invalid DB format: %d, disabling DB.\n", iview
->vk_format
);
1567 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1568 s_offs
= z_offs
= va
;
1569 z_offs
+= iview
->image
->surface
.level
[level
].offset
;
1570 s_offs
+= iview
->image
->surface
.stencil_level
[level
].offset
;
1572 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
1573 S_028008_SLICE_MAX(iview
->base_layer
+ iview
->extent
.depth
- 1);
1574 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(1);
1575 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
1577 if (iview
->image
->samples
> 1)
1578 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->samples
));
1580 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
)
1581 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_8
);
1583 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_INVALID
);
1585 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
) {
1586 struct radeon_info
*info
= &device
->instance
->physicalDevice
.rad_info
;
1587 unsigned tiling_index
= iview
->image
->surface
.tiling_index
[level
];
1588 unsigned stencil_index
= iview
->image
->surface
.stencil_tiling_index
[level
];
1589 unsigned macro_index
= iview
->image
->surface
.macro_tile_index
;
1590 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
1591 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
1592 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
1594 ds
->db_depth_info
|=
1595 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
1596 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
1597 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
1598 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
1599 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
1600 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
1601 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
1602 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
1604 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
1605 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
1606 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
1607 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
1610 if (iview
->image
->htile
.size
&& !level
) {
1611 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
1612 S_028040_ALLOW_EXPCLEAR(1);
1614 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
) {
1615 /* Workaround: For a not yet understood reason, the
1616 * combination of MSAA, fast stencil clear and stencil
1617 * decompress messes with subsequent stencil buffer
1618 * uses. Problem was reproduced on Verde, Bonaire,
1619 * Tonga, and Carrizo.
1621 * Disabling EXPCLEAR works around the problem.
1623 * Check piglit's arb_texture_multisample-stencil-clear
1624 * test if you want to try changing this.
1626 if (iview
->image
->samples
<= 1)
1627 ds
->db_stencil_info
|= S_028044_ALLOW_EXPCLEAR(1);
1629 /* Use all of the htile_buffer for depth if there's no stencil. */
1630 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
1632 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+
1633 iview
->image
->htile
.offset
;
1634 ds
->db_htile_data_base
= va
>> 8;
1635 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
1637 ds
->db_htile_data_base
= 0;
1638 ds
->db_htile_surface
= 0;
1641 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
1642 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
1644 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
1645 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
1646 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
1649 VkResult
radv_CreateFramebuffer(
1651 const VkFramebufferCreateInfo
* pCreateInfo
,
1652 const VkAllocationCallbacks
* pAllocator
,
1653 VkFramebuffer
* pFramebuffer
)
1655 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1656 struct radv_framebuffer
*framebuffer
;
1658 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
1660 size_t size
= sizeof(*framebuffer
) +
1661 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
1662 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
1663 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1664 if (framebuffer
== NULL
)
1665 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1667 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
1668 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
1669 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
1670 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
1671 framebuffer
->attachments
[i
].attachment
= iview
;
1672 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
1673 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
1674 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
1675 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
1679 framebuffer
->width
= pCreateInfo
->width
;
1680 framebuffer
->height
= pCreateInfo
->height
;
1681 framebuffer
->layers
= pCreateInfo
->layers
;
1683 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
1687 void radv_DestroyFramebuffer(
1690 const VkAllocationCallbacks
* pAllocator
)
1692 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1693 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
1697 vk_free2(&device
->alloc
, pAllocator
, fb
);
1700 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
1702 switch (address_mode
) {
1703 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
1704 return V_008F30_SQ_TEX_WRAP
;
1705 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
1706 return V_008F30_SQ_TEX_MIRROR
;
1707 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
1708 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
1709 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
1710 return V_008F30_SQ_TEX_CLAMP_BORDER
;
1711 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
1712 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
1714 unreachable("illegal tex wrap mode");
1720 radv_tex_compare(VkCompareOp op
)
1723 case VK_COMPARE_OP_NEVER
:
1724 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
1725 case VK_COMPARE_OP_LESS
:
1726 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
1727 case VK_COMPARE_OP_EQUAL
:
1728 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
1729 case VK_COMPARE_OP_LESS_OR_EQUAL
:
1730 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
1731 case VK_COMPARE_OP_GREATER
:
1732 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
1733 case VK_COMPARE_OP_NOT_EQUAL
:
1734 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
1735 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
1736 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
1737 case VK_COMPARE_OP_ALWAYS
:
1738 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
1740 unreachable("illegal compare mode");
1746 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
1749 case VK_FILTER_NEAREST
:
1750 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
1751 V_008F38_SQ_TEX_XY_FILTER_POINT
);
1752 case VK_FILTER_LINEAR
:
1753 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
1754 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
1755 case VK_FILTER_CUBIC_IMG
:
1757 fprintf(stderr
, "illegal texture filter");
1763 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
1766 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
1767 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
1768 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
1769 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
1771 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
1776 radv_tex_bordercolor(VkBorderColor bcolor
)
1779 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
1780 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
1781 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
1782 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
1783 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
1784 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
1785 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
1786 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
1787 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
1795 radv_tex_aniso_filter(unsigned filter
)
1809 radv_init_sampler(struct radv_device
*device
,
1810 struct radv_sampler
*sampler
,
1811 const VkSamplerCreateInfo
*pCreateInfo
)
1813 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
1814 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
1815 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
1816 bool is_vi
= (device
->instance
->physicalDevice
.rad_info
.chip_class
>= VI
);
1818 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
1819 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
1820 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
1821 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
1822 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
1823 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
1824 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
1825 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
1826 S_008F30_DISABLE_CUBE_WRAP(0) |
1827 S_008F30_COMPAT_MODE(is_vi
));
1828 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
1829 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
1830 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
1831 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
1832 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
1833 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
1834 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
1835 S_008F38_MIP_POINT_PRECLAMP(1) |
1836 S_008F38_DISABLE_LSB_CEIL(1) |
1837 S_008F38_FILTER_PREC_FIX(1) |
1838 S_008F38_ANISO_OVERRIDE(is_vi
));
1839 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
1840 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
1843 VkResult
radv_CreateSampler(
1845 const VkSamplerCreateInfo
* pCreateInfo
,
1846 const VkAllocationCallbacks
* pAllocator
,
1847 VkSampler
* pSampler
)
1849 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1850 struct radv_sampler
*sampler
;
1852 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
1854 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
1855 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1857 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1859 radv_init_sampler(device
, sampler
, pCreateInfo
);
1860 *pSampler
= radv_sampler_to_handle(sampler
);
1865 void radv_DestroySampler(
1868 const VkAllocationCallbacks
* pAllocator
)
1870 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1871 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
1875 vk_free2(&device
->alloc
, pAllocator
, sampler
);