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_SWAPCHAIN_EXTENSION_NAME
,
180 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
181 VkSystemAllocationScope allocationScope
)
187 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
188 size_t align
, VkSystemAllocationScope allocationScope
)
190 return realloc(pOriginal
, size
);
194 default_free_func(void *pUserData
, void *pMemory
)
199 static const VkAllocationCallbacks default_alloc
= {
201 .pfnAllocation
= default_alloc_func
,
202 .pfnReallocation
= default_realloc_func
,
203 .pfnFree
= default_free_func
,
206 VkResult
radv_CreateInstance(
207 const VkInstanceCreateInfo
* pCreateInfo
,
208 const VkAllocationCallbacks
* pAllocator
,
209 VkInstance
* pInstance
)
211 struct radv_instance
*instance
;
213 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
215 uint32_t client_version
;
216 if (pCreateInfo
->pApplicationInfo
&&
217 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
218 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
220 client_version
= VK_MAKE_VERSION(1, 0, 0);
223 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
224 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
225 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
226 "Client requested version %d.%d.%d",
227 VK_VERSION_MAJOR(client_version
),
228 VK_VERSION_MINOR(client_version
),
229 VK_VERSION_PATCH(client_version
));
232 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
234 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
235 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
236 global_extensions
[j
].extensionName
) == 0) {
242 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
245 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
246 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
248 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
250 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
253 instance
->alloc
= *pAllocator
;
255 instance
->alloc
= default_alloc
;
257 instance
->apiVersion
= client_version
;
258 instance
->physicalDeviceCount
= -1;
262 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
264 *pInstance
= radv_instance_to_handle(instance
);
269 void radv_DestroyInstance(
270 VkInstance _instance
,
271 const VkAllocationCallbacks
* pAllocator
)
273 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
275 if (instance
->physicalDeviceCount
> 0) {
276 /* We support at most one physical device. */
277 assert(instance
->physicalDeviceCount
== 1);
278 radv_physical_device_finish(&instance
->physicalDevice
);
281 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
285 vk_free(&instance
->alloc
, instance
);
288 VkResult
radv_EnumeratePhysicalDevices(
289 VkInstance _instance
,
290 uint32_t* pPhysicalDeviceCount
,
291 VkPhysicalDevice
* pPhysicalDevices
)
293 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
296 if (instance
->physicalDeviceCount
< 0) {
298 for (unsigned i
= 0; i
< 8; i
++) {
299 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
300 result
= radv_physical_device_init(&instance
->physicalDevice
,
302 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
306 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
307 instance
->physicalDeviceCount
= 0;
308 } else if (result
== VK_SUCCESS
) {
309 instance
->physicalDeviceCount
= 1;
315 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
316 * otherwise it's an inout parameter.
318 * The Vulkan spec (git aaed022) says:
320 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
321 * that is initialized with the number of devices the application is
322 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
323 * an array of at least this many VkPhysicalDevice handles [...].
325 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
326 * overwrites the contents of the variable pointed to by
327 * pPhysicalDeviceCount with the number of physical devices in in the
328 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
329 * pPhysicalDeviceCount with the number of physical handles written to
332 if (!pPhysicalDevices
) {
333 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
334 } else if (*pPhysicalDeviceCount
>= 1) {
335 pPhysicalDevices
[0] = radv_physical_device_to_handle(&instance
->physicalDevice
);
336 *pPhysicalDeviceCount
= 1;
337 } else if (*pPhysicalDeviceCount
< instance
->physicalDeviceCount
) {
338 return VK_INCOMPLETE
;
340 *pPhysicalDeviceCount
= 0;
346 void radv_GetPhysicalDeviceFeatures(
347 VkPhysicalDevice physicalDevice
,
348 VkPhysicalDeviceFeatures
* pFeatures
)
350 // RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
352 memset(pFeatures
, 0, sizeof(*pFeatures
));
354 *pFeatures
= (VkPhysicalDeviceFeatures
) {
355 .robustBufferAccess
= true,
356 .fullDrawIndexUint32
= true,
357 .imageCubeArray
= true,
358 .independentBlend
= true,
359 .geometryShader
= false,
360 .tessellationShader
= false,
361 .sampleRateShading
= false,
362 .dualSrcBlend
= true,
364 .multiDrawIndirect
= true,
365 .drawIndirectFirstInstance
= true,
367 .depthBiasClamp
= true,
368 .fillModeNonSolid
= true,
373 .multiViewport
= false,
374 .samplerAnisotropy
= false, /* FINISHME */
375 .textureCompressionETC2
= false,
376 .textureCompressionASTC_LDR
= false,
377 .textureCompressionBC
= true,
378 .occlusionQueryPrecise
= true,
379 .pipelineStatisticsQuery
= false,
380 .vertexPipelineStoresAndAtomics
= true,
381 .fragmentStoresAndAtomics
= true,
382 .shaderTessellationAndGeometryPointSize
= true,
383 .shaderImageGatherExtended
= false,
384 .shaderStorageImageExtendedFormats
= false,
385 .shaderStorageImageMultisample
= false,
386 .shaderUniformBufferArrayDynamicIndexing
= true,
387 .shaderSampledImageArrayDynamicIndexing
= true,
388 .shaderStorageBufferArrayDynamicIndexing
= true,
389 .shaderStorageImageArrayDynamicIndexing
= true,
390 .shaderStorageImageReadWithoutFormat
= false,
391 .shaderStorageImageWriteWithoutFormat
= true,
392 .shaderClipDistance
= true,
393 .shaderCullDistance
= true,
394 .shaderFloat64
= false,
395 .shaderInt64
= false,
396 .shaderInt16
= false,
398 .variableMultisampleRate
= false,
399 .inheritedQueries
= false,
403 void radv_GetPhysicalDeviceProperties(
404 VkPhysicalDevice physicalDevice
,
405 VkPhysicalDeviceProperties
* pProperties
)
407 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
408 VkSampleCountFlags sample_counts
= 0xf;
409 VkPhysicalDeviceLimits limits
= {
410 .maxImageDimension1D
= (1 << 14),
411 .maxImageDimension2D
= (1 << 14),
412 .maxImageDimension3D
= (1 << 11),
413 .maxImageDimensionCube
= (1 << 14),
414 .maxImageArrayLayers
= (1 << 11),
415 .maxTexelBufferElements
= 128 * 1024 * 1024,
416 .maxUniformBufferRange
= UINT32_MAX
,
417 .maxStorageBufferRange
= UINT32_MAX
,
418 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
419 .maxMemoryAllocationCount
= UINT32_MAX
,
420 .maxSamplerAllocationCount
= 64 * 1024,
421 .bufferImageGranularity
= 64, /* A cache line */
422 .sparseAddressSpaceSize
= 0,
423 .maxBoundDescriptorSets
= MAX_SETS
,
424 .maxPerStageDescriptorSamplers
= 64,
425 .maxPerStageDescriptorUniformBuffers
= 64,
426 .maxPerStageDescriptorStorageBuffers
= 64,
427 .maxPerStageDescriptorSampledImages
= 64,
428 .maxPerStageDescriptorStorageImages
= 64,
429 .maxPerStageDescriptorInputAttachments
= 64,
430 .maxPerStageResources
= 128,
431 .maxDescriptorSetSamplers
= 256,
432 .maxDescriptorSetUniformBuffers
= 256,
433 .maxDescriptorSetUniformBuffersDynamic
= 256,
434 .maxDescriptorSetStorageBuffers
= 256,
435 .maxDescriptorSetStorageBuffersDynamic
= 256,
436 .maxDescriptorSetSampledImages
= 256,
437 .maxDescriptorSetStorageImages
= 256,
438 .maxDescriptorSetInputAttachments
= 256,
439 .maxVertexInputAttributes
= 32,
440 .maxVertexInputBindings
= 32,
441 .maxVertexInputAttributeOffset
= 2047,
442 .maxVertexInputBindingStride
= 2048,
443 .maxVertexOutputComponents
= 128,
444 .maxTessellationGenerationLevel
= 0,
445 .maxTessellationPatchSize
= 0,
446 .maxTessellationControlPerVertexInputComponents
= 0,
447 .maxTessellationControlPerVertexOutputComponents
= 0,
448 .maxTessellationControlPerPatchOutputComponents
= 0,
449 .maxTessellationControlTotalOutputComponents
= 0,
450 .maxTessellationEvaluationInputComponents
= 0,
451 .maxTessellationEvaluationOutputComponents
= 0,
452 .maxGeometryShaderInvocations
= 32,
453 .maxGeometryInputComponents
= 64,
454 .maxGeometryOutputComponents
= 128,
455 .maxGeometryOutputVertices
= 256,
456 .maxGeometryTotalOutputComponents
= 1024,
457 .maxFragmentInputComponents
= 128,
458 .maxFragmentOutputAttachments
= 8,
459 .maxFragmentDualSrcAttachments
= 2,
460 .maxFragmentCombinedOutputResources
= 8,
461 .maxComputeSharedMemorySize
= 32768,
462 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
463 .maxComputeWorkGroupInvocations
= 16 * 1024,
464 .maxComputeWorkGroupSize
= {
465 16 * 1024/*devinfo->max_cs_threads*/,
469 .subPixelPrecisionBits
= 4 /* FIXME */,
470 .subTexelPrecisionBits
= 4 /* FIXME */,
471 .mipmapPrecisionBits
= 4 /* FIXME */,
472 .maxDrawIndexedIndexValue
= UINT32_MAX
,
473 .maxDrawIndirectCount
= UINT32_MAX
,
474 .maxSamplerLodBias
= 16,
475 .maxSamplerAnisotropy
= 16,
476 .maxViewports
= MAX_VIEWPORTS
,
477 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
478 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
479 .viewportSubPixelBits
= 13, /* We take a float? */
480 .minMemoryMapAlignment
= 4096, /* A page */
481 .minTexelBufferOffsetAlignment
= 1,
482 .minUniformBufferOffsetAlignment
= 4,
483 .minStorageBufferOffsetAlignment
= 4,
484 .minTexelOffset
= -8,
486 .minTexelGatherOffset
= -8,
487 .maxTexelGatherOffset
= 7,
488 .minInterpolationOffset
= 0, /* FIXME */
489 .maxInterpolationOffset
= 0, /* FIXME */
490 .subPixelInterpolationOffsetBits
= 0, /* FIXME */
491 .maxFramebufferWidth
= (1 << 14),
492 .maxFramebufferHeight
= (1 << 14),
493 .maxFramebufferLayers
= (1 << 10),
494 .framebufferColorSampleCounts
= sample_counts
,
495 .framebufferDepthSampleCounts
= sample_counts
,
496 .framebufferStencilSampleCounts
= sample_counts
,
497 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
498 .maxColorAttachments
= MAX_RTS
,
499 .sampledImageColorSampleCounts
= sample_counts
,
500 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
501 .sampledImageDepthSampleCounts
= sample_counts
,
502 .sampledImageStencilSampleCounts
= sample_counts
,
503 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
504 .maxSampleMaskWords
= 1,
505 .timestampComputeAndGraphics
= false,
506 .timestampPeriod
= 100000.0 / pdevice
->rad_info
.clock_crystal_freq
,
507 .maxClipDistances
= 8,
508 .maxCullDistances
= 8,
509 .maxCombinedClipAndCullDistances
= 8,
510 .discreteQueuePriorities
= 1,
511 .pointSizeRange
= { 0.125, 255.875 },
512 .lineWidthRange
= { 0.0, 7.9921875 },
513 .pointSizeGranularity
= (1.0 / 8.0),
514 .lineWidthGranularity
= (1.0 / 128.0),
515 .strictLines
= false, /* FINISHME */
516 .standardSampleLocations
= true,
517 .optimalBufferCopyOffsetAlignment
= 128,
518 .optimalBufferCopyRowPitchAlignment
= 128,
519 .nonCoherentAtomSize
= 64,
522 *pProperties
= (VkPhysicalDeviceProperties
) {
523 .apiVersion
= VK_MAKE_VERSION(1, 0, 5),
526 .deviceID
= pdevice
->rad_info
.pci_id
,
527 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
,
529 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
532 strcpy(pProperties
->deviceName
, pdevice
->name
);
533 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
536 void radv_GetPhysicalDeviceQueueFamilyProperties(
537 VkPhysicalDevice physicalDevice
,
539 VkQueueFamilyProperties
* pQueueFamilyProperties
)
541 if (pQueueFamilyProperties
== NULL
) {
545 assert(*pCount
>= 1);
547 *pQueueFamilyProperties
= (VkQueueFamilyProperties
) {
548 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
549 VK_QUEUE_COMPUTE_BIT
|
550 VK_QUEUE_TRANSFER_BIT
,
552 .timestampValidBits
= 64,
553 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
557 void radv_GetPhysicalDeviceMemoryProperties(
558 VkPhysicalDevice physicalDevice
,
559 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
561 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
563 pMemoryProperties
->memoryTypeCount
= 4;
564 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
565 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
568 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
569 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
570 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
573 pMemoryProperties
->memoryTypes
[2] = (VkMemoryType
) {
574 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
575 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
576 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
579 pMemoryProperties
->memoryTypes
[3] = (VkMemoryType
) {
580 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
581 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
582 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
586 pMemoryProperties
->memoryHeapCount
= 3;
587 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
588 .size
= physical_device
->rad_info
.vram_size
-
589 physical_device
->rad_info
.visible_vram_size
,
590 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
592 pMemoryProperties
->memoryHeaps
[1] = (VkMemoryHeap
) {
593 .size
= physical_device
->rad_info
.visible_vram_size
,
594 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
596 pMemoryProperties
->memoryHeaps
[2] = (VkMemoryHeap
) {
597 .size
= physical_device
->rad_info
.gart_size
,
603 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
)
605 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
606 queue
->device
= device
;
612 radv_queue_finish(struct radv_queue
*queue
)
616 VkResult
radv_CreateDevice(
617 VkPhysicalDevice physicalDevice
,
618 const VkDeviceCreateInfo
* pCreateInfo
,
619 const VkAllocationCallbacks
* pAllocator
,
622 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
624 struct radv_device
*device
;
626 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
628 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
629 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
630 device_extensions
[j
].extensionName
) == 0) {
636 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
639 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
641 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
643 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
645 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
646 device
->instance
= physical_device
->instance
;
648 device
->ws
= physical_device
->ws
;
650 device
->alloc
= *pAllocator
;
652 device
->alloc
= physical_device
->instance
->alloc
;
654 device
->hw_ctx
= device
->ws
->ctx_create(device
->ws
);
655 if (!device
->hw_ctx
) {
656 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
660 radv_queue_init(device
, &device
->queue
);
662 result
= radv_device_init_meta(device
);
663 if (result
!= VK_SUCCESS
) {
664 device
->ws
->ctx_destroy(device
->hw_ctx
);
667 device
->allow_fast_clears
= env_var_as_boolean("RADV_FAST_CLEARS", false);
668 device
->allow_dcc
= !env_var_as_boolean("RADV_DCC_DISABLE", false);
669 device
->shader_stats_dump
= env_var_as_boolean("RADV_SHADER_STATS", false);
671 if (device
->allow_fast_clears
&& device
->allow_dcc
)
672 radv_finishme("DCC fast clears have not been tested\n");
674 radv_device_init_msaa(device
);
675 device
->empty_cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
676 radeon_emit(device
->empty_cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
677 radeon_emit(device
->empty_cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
678 radeon_emit(device
->empty_cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
679 device
->ws
->cs_finalize(device
->empty_cs
);
680 *pDevice
= radv_device_to_handle(device
);
683 vk_free(&device
->alloc
, device
);
687 void radv_DestroyDevice(
689 const VkAllocationCallbacks
* pAllocator
)
691 RADV_FROM_HANDLE(radv_device
, device
, _device
);
693 device
->ws
->ctx_destroy(device
->hw_ctx
);
694 radv_queue_finish(&device
->queue
);
695 radv_device_finish_meta(device
);
697 vk_free(&device
->alloc
, device
);
700 VkResult
radv_EnumerateInstanceExtensionProperties(
701 const char* pLayerName
,
702 uint32_t* pPropertyCount
,
703 VkExtensionProperties
* pProperties
)
706 if (pProperties
== NULL
) {
707 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
711 for (i
= 0; i
< *pPropertyCount
; i
++)
712 memcpy(&pProperties
[i
], &global_extensions
[i
], sizeof(VkExtensionProperties
));
715 if (i
< ARRAY_SIZE(global_extensions
))
716 return VK_INCOMPLETE
;
721 VkResult
radv_EnumerateDeviceExtensionProperties(
722 VkPhysicalDevice physicalDevice
,
723 const char* pLayerName
,
724 uint32_t* pPropertyCount
,
725 VkExtensionProperties
* pProperties
)
729 if (pProperties
== NULL
) {
730 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
734 for (i
= 0; i
< *pPropertyCount
; i
++)
735 memcpy(&pProperties
[i
], &device_extensions
[i
], sizeof(VkExtensionProperties
));
738 if (i
< ARRAY_SIZE(device_extensions
))
739 return VK_INCOMPLETE
;
743 VkResult
radv_EnumerateInstanceLayerProperties(
744 uint32_t* pPropertyCount
,
745 VkLayerProperties
* pProperties
)
747 if (pProperties
== NULL
) {
752 /* None supported at this time */
753 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
756 VkResult
radv_EnumerateDeviceLayerProperties(
757 VkPhysicalDevice physicalDevice
,
758 uint32_t* pPropertyCount
,
759 VkLayerProperties
* pProperties
)
761 if (pProperties
== NULL
) {
766 /* None supported at this time */
767 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
770 void radv_GetDeviceQueue(
772 uint32_t queueNodeIndex
,
776 RADV_FROM_HANDLE(radv_device
, device
, _device
);
778 assert(queueIndex
== 0);
780 *pQueue
= radv_queue_to_handle(&device
->queue
);
783 VkResult
radv_QueueSubmit(
785 uint32_t submitCount
,
786 const VkSubmitInfo
* pSubmits
,
789 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
790 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
791 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
792 struct radeon_winsys_ctx
*ctx
= queue
->device
->hw_ctx
;
795 for (uint32_t i
= 0; i
< submitCount
; i
++) {
796 struct radeon_winsys_cs
**cs_array
;
797 bool can_patch
= true;
799 if (!pSubmits
[i
].commandBufferCount
)
802 cs_array
= malloc(sizeof(struct radeon_winsys_cs
*) *
803 pSubmits
[i
].commandBufferCount
);
805 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
806 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
807 pSubmits
[i
].pCommandBuffers
[j
]);
808 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
810 cs_array
[j
] = cmd_buffer
->cs
;
811 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
814 ret
= queue
->device
->ws
->cs_submit(ctx
, cs_array
,
815 pSubmits
[i
].commandBufferCount
,
816 can_patch
, base_fence
);
818 radv_loge("failed to submit CS %d\n", i
);
824 ret
= queue
->device
->ws
->cs_submit(ctx
, &queue
->device
->empty_cs
,
825 1, false, base_fence
);
827 fence
->submitted
= true;
833 VkResult
radv_QueueWaitIdle(
836 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
838 queue
->device
->ws
->ctx_wait_idle(queue
->device
->hw_ctx
);
842 VkResult
radv_DeviceWaitIdle(
845 RADV_FROM_HANDLE(radv_device
, device
, _device
);
847 device
->ws
->ctx_wait_idle(device
->hw_ctx
);
851 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
855 return radv_lookup_entrypoint(pName
);
858 /* The loader wants us to expose a second GetInstanceProcAddr function
859 * to work around certain LD_PRELOAD issues seen in apps.
862 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
867 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
871 return radv_GetInstanceProcAddr(instance
, pName
);
874 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
878 return radv_lookup_entrypoint(pName
);
881 VkResult
radv_AllocateMemory(
883 const VkMemoryAllocateInfo
* pAllocateInfo
,
884 const VkAllocationCallbacks
* pAllocator
,
885 VkDeviceMemory
* pMem
)
887 RADV_FROM_HANDLE(radv_device
, device
, _device
);
888 struct radv_device_memory
*mem
;
890 enum radeon_bo_domain domain
;
892 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
894 if (pAllocateInfo
->allocationSize
== 0) {
895 /* Apparently, this is allowed */
896 *pMem
= VK_NULL_HANDLE
;
900 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
901 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
903 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
905 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
906 if (pAllocateInfo
->memoryTypeIndex
== 1 || pAllocateInfo
->memoryTypeIndex
== 3)
907 domain
= RADEON_DOMAIN_GTT
;
909 domain
= RADEON_DOMAIN_VRAM
;
911 if (pAllocateInfo
->memoryTypeIndex
== 0)
912 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
914 flags
|= RADEON_FLAG_CPU_ACCESS
;
916 if (pAllocateInfo
->memoryTypeIndex
== 1)
917 flags
|= RADEON_FLAG_GTT_WC
;
919 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, 32768,
923 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
926 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
928 *pMem
= radv_device_memory_to_handle(mem
);
933 vk_free2(&device
->alloc
, pAllocator
, mem
);
938 void radv_FreeMemory(
941 const VkAllocationCallbacks
* pAllocator
)
943 RADV_FROM_HANDLE(radv_device
, device
, _device
);
944 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
949 device
->ws
->buffer_destroy(mem
->bo
);
952 vk_free2(&device
->alloc
, pAllocator
, mem
);
955 VkResult
radv_MapMemory(
957 VkDeviceMemory _memory
,
960 VkMemoryMapFlags flags
,
963 RADV_FROM_HANDLE(radv_device
, device
, _device
);
964 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
971 *ppData
= device
->ws
->buffer_map(mem
->bo
);
977 return VK_ERROR_MEMORY_MAP_FAILED
;
980 void radv_UnmapMemory(
982 VkDeviceMemory _memory
)
984 RADV_FROM_HANDLE(radv_device
, device
, _device
);
985 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
990 device
->ws
->buffer_unmap(mem
->bo
);
993 VkResult
radv_FlushMappedMemoryRanges(
995 uint32_t memoryRangeCount
,
996 const VkMappedMemoryRange
* pMemoryRanges
)
1001 VkResult
radv_InvalidateMappedMemoryRanges(
1003 uint32_t memoryRangeCount
,
1004 const VkMappedMemoryRange
* pMemoryRanges
)
1009 void radv_GetBufferMemoryRequirements(
1012 VkMemoryRequirements
* pMemoryRequirements
)
1014 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1016 /* The Vulkan spec (git aaed022) says:
1018 * memoryTypeBits is a bitfield and contains one bit set for every
1019 * supported memory type for the resource. The bit `1<<i` is set if and
1020 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1021 * structure for the physical device is supported.
1023 * We support exactly one memory type.
1025 pMemoryRequirements
->memoryTypeBits
= 0x7;
1027 pMemoryRequirements
->size
= buffer
->size
;
1028 pMemoryRequirements
->alignment
= 16;
1031 void radv_GetImageMemoryRequirements(
1034 VkMemoryRequirements
* pMemoryRequirements
)
1036 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1038 /* The Vulkan spec (git aaed022) says:
1040 * memoryTypeBits is a bitfield and contains one bit set for every
1041 * supported memory type for the resource. The bit `1<<i` is set if and
1042 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1043 * structure for the physical device is supported.
1045 * We support exactly one memory type.
1047 pMemoryRequirements
->memoryTypeBits
= 0x7;
1049 pMemoryRequirements
->size
= image
->size
;
1050 pMemoryRequirements
->alignment
= image
->alignment
;
1053 void radv_GetImageSparseMemoryRequirements(
1056 uint32_t* pSparseMemoryRequirementCount
,
1057 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1062 void radv_GetDeviceMemoryCommitment(
1064 VkDeviceMemory memory
,
1065 VkDeviceSize
* pCommittedMemoryInBytes
)
1067 *pCommittedMemoryInBytes
= 0;
1070 VkResult
radv_BindBufferMemory(
1073 VkDeviceMemory _memory
,
1074 VkDeviceSize memoryOffset
)
1076 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1077 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1080 buffer
->bo
= mem
->bo
;
1081 buffer
->offset
= memoryOffset
;
1090 VkResult
radv_BindImageMemory(
1093 VkDeviceMemory _memory
,
1094 VkDeviceSize memoryOffset
)
1096 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1097 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1100 image
->bo
= mem
->bo
;
1101 image
->offset
= memoryOffset
;
1110 VkResult
radv_QueueBindSparse(
1112 uint32_t bindInfoCount
,
1113 const VkBindSparseInfo
* pBindInfo
,
1116 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1119 VkResult
radv_CreateFence(
1121 const VkFenceCreateInfo
* pCreateInfo
,
1122 const VkAllocationCallbacks
* pAllocator
,
1125 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1126 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
1128 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1131 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1133 memset(fence
, 0, sizeof(*fence
));
1134 fence
->submitted
= false;
1135 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
1136 fence
->fence
= device
->ws
->create_fence();
1139 *pFence
= radv_fence_to_handle(fence
);
1144 void radv_DestroyFence(
1147 const VkAllocationCallbacks
* pAllocator
)
1149 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1150 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1154 device
->ws
->destroy_fence(fence
->fence
);
1155 vk_free2(&device
->alloc
, pAllocator
, fence
);
1158 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
1160 uint64_t current_time
;
1163 clock_gettime(CLOCK_MONOTONIC
, &tv
);
1164 current_time
= tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
1166 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
1168 return current_time
+ timeout
;
1171 VkResult
radv_WaitForFences(
1173 uint32_t fenceCount
,
1174 const VkFence
* pFences
,
1178 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1179 timeout
= radv_get_absolute_timeout(timeout
);
1181 if (!waitAll
&& fenceCount
> 1) {
1182 fprintf(stderr
, "radv: WaitForFences without waitAll not implemented yet\n");
1185 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
1186 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1187 bool expired
= false;
1189 if (fence
->signalled
)
1192 if (!fence
->submitted
)
1195 expired
= device
->ws
->fence_wait(device
->ws
, fence
->fence
, true, timeout
);
1199 fence
->signalled
= true;
1205 VkResult
radv_ResetFences(VkDevice device
,
1206 uint32_t fenceCount
,
1207 const VkFence
*pFences
)
1209 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
1210 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1211 fence
->submitted
= fence
->signalled
= false;
1217 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
1219 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1220 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1222 if (fence
->signalled
)
1224 if (!fence
->submitted
)
1225 return VK_NOT_READY
;
1227 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
1228 return VK_NOT_READY
;
1234 // Queue semaphore functions
1236 VkResult
radv_CreateSemaphore(
1238 const VkSemaphoreCreateInfo
* pCreateInfo
,
1239 const VkAllocationCallbacks
* pAllocator
,
1240 VkSemaphore
* pSemaphore
)
1242 /* The DRM execbuffer ioctl always execute in-oder, even between different
1243 * rings. As such, there's nothing to do for the user space semaphore.
1246 *pSemaphore
= (VkSemaphore
)1;
1251 void radv_DestroySemaphore(
1253 VkSemaphore semaphore
,
1254 const VkAllocationCallbacks
* pAllocator
)
1258 VkResult
radv_CreateEvent(
1260 const VkEventCreateInfo
* pCreateInfo
,
1261 const VkAllocationCallbacks
* pAllocator
,
1264 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1265 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
1267 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1270 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1272 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
1274 RADEON_FLAG_CPU_ACCESS
);
1276 vk_free2(&device
->alloc
, pAllocator
, event
);
1277 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1280 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
1282 *pEvent
= radv_event_to_handle(event
);
1287 void radv_DestroyEvent(
1290 const VkAllocationCallbacks
* pAllocator
)
1292 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1293 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1297 device
->ws
->buffer_destroy(event
->bo
);
1298 vk_free2(&device
->alloc
, pAllocator
, event
);
1301 VkResult
radv_GetEventStatus(
1305 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1307 if (*event
->map
== 1)
1308 return VK_EVENT_SET
;
1309 return VK_EVENT_RESET
;
1312 VkResult
radv_SetEvent(
1316 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1322 VkResult
radv_ResetEvent(
1326 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1332 VkResult
radv_CreateBuffer(
1334 const VkBufferCreateInfo
* pCreateInfo
,
1335 const VkAllocationCallbacks
* pAllocator
,
1338 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1339 struct radv_buffer
*buffer
;
1341 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
1343 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
1344 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1346 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1348 buffer
->size
= pCreateInfo
->size
;
1349 buffer
->usage
= pCreateInfo
->usage
;
1353 *pBuffer
= radv_buffer_to_handle(buffer
);
1358 void radv_DestroyBuffer(
1361 const VkAllocationCallbacks
* pAllocator
)
1363 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1364 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1369 vk_free2(&device
->alloc
, pAllocator
, buffer
);
1372 static inline unsigned
1373 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
1376 return image
->surface
.stencil_tiling_index
[level
];
1378 return image
->surface
.tiling_index
[level
];
1382 radv_initialise_color_surface(struct radv_device
*device
,
1383 struct radv_color_buffer_info
*cb
,
1384 struct radv_image_view
*iview
)
1386 const struct vk_format_description
*desc
;
1387 unsigned ntype
, format
, swap
, endian
;
1388 unsigned blend_clamp
= 0, blend_bypass
= 0;
1389 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
1391 const struct radeon_surf
*surf
= &iview
->image
->surface
;
1392 const struct radeon_surf_level
*level_info
= &surf
->level
[iview
->base_mip
];
1394 desc
= vk_format_description(iview
->vk_format
);
1396 memset(cb
, 0, sizeof(*cb
));
1398 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1399 va
+= level_info
->offset
;
1400 cb
->cb_color_base
= va
>> 8;
1402 /* CMASK variables */
1403 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1404 va
+= iview
->image
->cmask
.offset
;
1405 cb
->cb_color_cmask
= va
>> 8;
1406 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
1408 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1409 va
+= iview
->image
->dcc_offset
;
1410 cb
->cb_dcc_base
= va
>> 8;
1412 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
1413 S_028C6C_SLICE_MAX(iview
->base_layer
+ iview
->extent
.depth
- 1);
1415 cb
->micro_tile_mode
= iview
->image
->surface
.micro_tile_mode
;
1416 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
1417 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
1418 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
1420 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
1421 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
1423 /* Intensity is implemented as Red, so treat it that way. */
1424 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
) |
1425 S_028C74_TILE_MODE_INDEX(tile_mode_index
);
1427 if (iview
->image
->samples
> 1) {
1428 unsigned log_samples
= util_logbase2(iview
->image
->samples
);
1430 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
1431 S_028C74_NUM_FRAGMENTS(log_samples
);
1434 if (iview
->image
->fmask
.size
) {
1435 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
1436 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
)
1437 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
1438 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
1439 cb
->cb_color_fmask
= va
>> 8;
1440 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
1442 /* This must be set for fast clear to work without FMASK. */
1443 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
)
1444 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
1445 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
1446 cb
->cb_color_fmask
= cb
->cb_color_base
;
1447 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
1450 ntype
= radv_translate_color_numformat(iview
->vk_format
,
1452 vk_format_get_first_non_void_channel(iview
->vk_format
));
1453 format
= radv_translate_colorformat(iview
->vk_format
);
1454 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
1455 radv_finishme("Illegal color\n");
1456 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
1457 endian
= radv_colorformat_endian_swap(format
);
1459 /* blend clamp should be set for all NORM/SRGB types */
1460 if (ntype
== V_028C70_NUMBER_UNORM
||
1461 ntype
== V_028C70_NUMBER_SNORM
||
1462 ntype
== V_028C70_NUMBER_SRGB
)
1465 /* set blend bypass according to docs if SINT/UINT or
1466 8/24 COLOR variants */
1467 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
1468 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
1469 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
1474 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
1475 (format
== V_028C70_COLOR_8
||
1476 format
== V_028C70_COLOR_8_8
||
1477 format
== V_028C70_COLOR_8_8_8_8
))
1478 ->color_is_int8
= true;
1480 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
1481 S_028C70_COMP_SWAP(swap
) |
1482 S_028C70_BLEND_CLAMP(blend_clamp
) |
1483 S_028C70_BLEND_BYPASS(blend_bypass
) |
1484 S_028C70_SIMPLE_FLOAT(1) |
1485 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
1486 ntype
!= V_028C70_NUMBER_SNORM
&&
1487 ntype
!= V_028C70_NUMBER_SRGB
&&
1488 format
!= V_028C70_COLOR_8_24
&&
1489 format
!= V_028C70_COLOR_24_8
) |
1490 S_028C70_NUMBER_TYPE(ntype
) |
1491 S_028C70_ENDIAN(endian
);
1492 if (iview
->image
->samples
> 1)
1493 if (iview
->image
->fmask
.size
)
1494 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
1496 if (iview
->image
->cmask
.size
&& device
->allow_fast_clears
)
1497 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
1499 if (iview
->image
->surface
.dcc_size
&& level_info
->dcc_enabled
)
1500 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
1502 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= VI
) {
1503 unsigned max_uncompressed_block_size
= 2;
1504 if (iview
->image
->samples
> 1) {
1505 if (iview
->image
->surface
.bpe
== 1)
1506 max_uncompressed_block_size
= 0;
1507 else if (iview
->image
->surface
.bpe
== 2)
1508 max_uncompressed_block_size
= 1;
1511 cb
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
1512 S_028C78_INDEPENDENT_64B_BLOCKS(1);
1515 /* This must be set for fast clear to work without FMASK. */
1516 if (!iview
->image
->fmask
.size
&&
1517 device
->instance
->physicalDevice
.rad_info
.chip_class
== SI
) {
1518 unsigned bankh
= util_logbase2(iview
->image
->surface
.bankh
);
1519 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
1524 radv_initialise_ds_surface(struct radv_device
*device
,
1525 struct radv_ds_buffer_info
*ds
,
1526 struct radv_image_view
*iview
)
1528 unsigned level
= iview
->base_mip
;
1530 uint64_t va
, s_offs
, z_offs
;
1531 const struct radeon_surf_level
*level_info
= &iview
->image
->surface
.level
[level
];
1532 memset(ds
, 0, sizeof(*ds
));
1533 switch (iview
->vk_format
) {
1534 case VK_FORMAT_D24_UNORM_S8_UINT
:
1535 case VK_FORMAT_X8_D24_UNORM_PACK32
:
1536 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
1537 ds
->offset_scale
= 2.0f
;
1539 case VK_FORMAT_D16_UNORM
:
1540 case VK_FORMAT_D16_UNORM_S8_UINT
:
1541 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
1542 ds
->offset_scale
= 4.0f
;
1544 case VK_FORMAT_D32_SFLOAT
:
1545 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
1546 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
1547 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
1548 ds
->offset_scale
= 1.0f
;
1554 format
= radv_translate_dbformat(iview
->vk_format
);
1555 if (format
== V_028040_Z_INVALID
) {
1556 fprintf(stderr
, "Invalid DB format: %d, disabling DB.\n", iview
->vk_format
);
1559 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1560 s_offs
= z_offs
= va
;
1561 z_offs
+= iview
->image
->surface
.level
[level
].offset
;
1562 s_offs
+= iview
->image
->surface
.stencil_level
[level
].offset
;
1564 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
1565 S_028008_SLICE_MAX(iview
->base_layer
+ iview
->extent
.depth
- 1);
1566 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(1);
1567 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
1569 if (iview
->image
->samples
> 1)
1570 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->samples
));
1572 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
)
1573 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_8
);
1575 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_INVALID
);
1577 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
) {
1578 struct radeon_info
*info
= &device
->instance
->physicalDevice
.rad_info
;
1579 unsigned tiling_index
= iview
->image
->surface
.tiling_index
[level
];
1580 unsigned stencil_index
= iview
->image
->surface
.stencil_tiling_index
[level
];
1581 unsigned macro_index
= iview
->image
->surface
.macro_tile_index
;
1582 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
1583 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
1584 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
1586 ds
->db_depth_info
|=
1587 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
1588 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
1589 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
1590 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
1591 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
1592 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
1593 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
1594 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
1596 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
1597 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
1598 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
1599 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
1602 if (iview
->image
->htile
.size
&& !level
) {
1603 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
1604 S_028040_ALLOW_EXPCLEAR(1);
1606 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
) {
1607 /* Workaround: For a not yet understood reason, the
1608 * combination of MSAA, fast stencil clear and stencil
1609 * decompress messes with subsequent stencil buffer
1610 * uses. Problem was reproduced on Verde, Bonaire,
1611 * Tonga, and Carrizo.
1613 * Disabling EXPCLEAR works around the problem.
1615 * Check piglit's arb_texture_multisample-stencil-clear
1616 * test if you want to try changing this.
1618 if (iview
->image
->samples
<= 1)
1619 ds
->db_stencil_info
|= S_028044_ALLOW_EXPCLEAR(1);
1621 /* Use all of the htile_buffer for depth if there's no stencil. */
1622 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
1624 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+
1625 iview
->image
->htile
.offset
;
1626 ds
->db_htile_data_base
= va
>> 8;
1627 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
1629 ds
->db_htile_data_base
= 0;
1630 ds
->db_htile_surface
= 0;
1633 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
1634 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
1636 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
1637 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
1638 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
1641 VkResult
radv_CreateFramebuffer(
1643 const VkFramebufferCreateInfo
* pCreateInfo
,
1644 const VkAllocationCallbacks
* pAllocator
,
1645 VkFramebuffer
* pFramebuffer
)
1647 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1648 struct radv_framebuffer
*framebuffer
;
1650 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
1652 size_t size
= sizeof(*framebuffer
) +
1653 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
1654 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
1655 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1656 if (framebuffer
== NULL
)
1657 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1659 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
1660 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
1661 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
1662 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
1663 framebuffer
->attachments
[i
].attachment
= iview
;
1664 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
1665 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
1666 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
1667 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
1671 framebuffer
->width
= pCreateInfo
->width
;
1672 framebuffer
->height
= pCreateInfo
->height
;
1673 framebuffer
->layers
= pCreateInfo
->layers
;
1675 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
1679 void radv_DestroyFramebuffer(
1682 const VkAllocationCallbacks
* pAllocator
)
1684 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1685 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
1689 vk_free2(&device
->alloc
, pAllocator
, fb
);
1692 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
1694 switch (address_mode
) {
1695 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
1696 return V_008F30_SQ_TEX_WRAP
;
1697 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
1698 return V_008F30_SQ_TEX_MIRROR
;
1699 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
1700 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
1701 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
1702 return V_008F30_SQ_TEX_CLAMP_BORDER
;
1703 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
1704 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
1706 unreachable("illegal tex wrap mode");
1712 radv_tex_compare(VkCompareOp op
)
1715 case VK_COMPARE_OP_NEVER
:
1716 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
1717 case VK_COMPARE_OP_LESS
:
1718 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
1719 case VK_COMPARE_OP_EQUAL
:
1720 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
1721 case VK_COMPARE_OP_LESS_OR_EQUAL
:
1722 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
1723 case VK_COMPARE_OP_GREATER
:
1724 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
1725 case VK_COMPARE_OP_NOT_EQUAL
:
1726 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
1727 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
1728 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
1729 case VK_COMPARE_OP_ALWAYS
:
1730 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
1732 unreachable("illegal compare mode");
1738 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
1741 case VK_FILTER_NEAREST
:
1742 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
1743 V_008F38_SQ_TEX_XY_FILTER_POINT
);
1744 case VK_FILTER_LINEAR
:
1745 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
1746 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
1747 case VK_FILTER_CUBIC_IMG
:
1749 fprintf(stderr
, "illegal texture filter");
1755 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
1758 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
1759 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
1760 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
1761 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
1763 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
1768 radv_tex_bordercolor(VkBorderColor bcolor
)
1771 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
1772 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
1773 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
1774 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
1775 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
1776 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
1777 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
1778 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
1779 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
1787 radv_tex_aniso_filter(unsigned filter
)
1801 radv_init_sampler(struct radv_device
*device
,
1802 struct radv_sampler
*sampler
,
1803 const VkSamplerCreateInfo
*pCreateInfo
)
1805 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
1806 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
1807 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
1809 is_vi
= (device
->instance
->physicalDevice
.rad_info
.chip_class
>= VI
);
1811 if (!is_vi
&& max_aniso
> 0) {
1812 radv_finishme("Anisotropic filtering must be disabled manually "
1813 "by the shader on SI-CI when BASE_LEVEL == LAST_LEVEL\n");
1814 max_aniso
= max_aniso_ratio
= 0;
1817 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
1818 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
1819 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
1820 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
1821 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
1822 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
1823 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
1824 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
1825 S_008F30_DISABLE_CUBE_WRAP(0) |
1826 S_008F30_COMPAT_MODE(is_vi
));
1827 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
1828 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
1829 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
1830 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
1831 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
1832 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
1833 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
1834 S_008F38_MIP_POINT_PRECLAMP(1) |
1835 S_008F38_DISABLE_LSB_CEIL(1) |
1836 S_008F38_FILTER_PREC_FIX(1) |
1837 S_008F38_ANISO_OVERRIDE(is_vi
));
1838 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
1839 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
1842 VkResult
radv_CreateSampler(
1844 const VkSamplerCreateInfo
* pCreateInfo
,
1845 const VkAllocationCallbacks
* pAllocator
,
1846 VkSampler
* pSampler
)
1848 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1849 struct radv_sampler
*sampler
;
1851 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
1853 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
1854 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1856 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1858 radv_init_sampler(device
, sampler
, pCreateInfo
);
1859 *pSampler
= radv_sampler_to_handle(sampler
);
1864 void radv_DestroySampler(
1867 const VkAllocationCallbacks
* pAllocator
)
1869 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1870 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
1874 vk_free2(&device
->alloc
, pAllocator
, sampler
);