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
,
184 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
185 VkSystemAllocationScope allocationScope
)
191 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
192 size_t align
, VkSystemAllocationScope allocationScope
)
194 return realloc(pOriginal
, size
);
198 default_free_func(void *pUserData
, void *pMemory
)
203 static const VkAllocationCallbacks default_alloc
= {
205 .pfnAllocation
= default_alloc_func
,
206 .pfnReallocation
= default_realloc_func
,
207 .pfnFree
= default_free_func
,
210 VkResult
radv_CreateInstance(
211 const VkInstanceCreateInfo
* pCreateInfo
,
212 const VkAllocationCallbacks
* pAllocator
,
213 VkInstance
* pInstance
)
215 struct radv_instance
*instance
;
217 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
219 uint32_t client_version
;
220 if (pCreateInfo
->pApplicationInfo
&&
221 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
222 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
224 client_version
= VK_MAKE_VERSION(1, 0, 0);
227 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
228 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
229 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
230 "Client requested version %d.%d.%d",
231 VK_VERSION_MAJOR(client_version
),
232 VK_VERSION_MINOR(client_version
),
233 VK_VERSION_PATCH(client_version
));
236 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
238 for (uint32_t j
= 0; j
< ARRAY_SIZE(global_extensions
); j
++) {
239 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
240 global_extensions
[j
].extensionName
) == 0) {
246 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
249 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
250 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
252 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
254 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
257 instance
->alloc
= *pAllocator
;
259 instance
->alloc
= default_alloc
;
261 instance
->apiVersion
= client_version
;
262 instance
->physicalDeviceCount
= -1;
266 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
268 *pInstance
= radv_instance_to_handle(instance
);
273 void radv_DestroyInstance(
274 VkInstance _instance
,
275 const VkAllocationCallbacks
* pAllocator
)
277 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
279 if (instance
->physicalDeviceCount
> 0) {
280 /* We support at most one physical device. */
281 assert(instance
->physicalDeviceCount
== 1);
282 radv_physical_device_finish(&instance
->physicalDevice
);
285 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
289 vk_free(&instance
->alloc
, instance
);
292 VkResult
radv_EnumeratePhysicalDevices(
293 VkInstance _instance
,
294 uint32_t* pPhysicalDeviceCount
,
295 VkPhysicalDevice
* pPhysicalDevices
)
297 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
300 if (instance
->physicalDeviceCount
< 0) {
302 for (unsigned i
= 0; i
< 8; i
++) {
303 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
304 result
= radv_physical_device_init(&instance
->physicalDevice
,
306 if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
310 if (result
== VK_ERROR_INCOMPATIBLE_DRIVER
) {
311 instance
->physicalDeviceCount
= 0;
312 } else if (result
== VK_SUCCESS
) {
313 instance
->physicalDeviceCount
= 1;
319 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
320 * otherwise it's an inout parameter.
322 * The Vulkan spec (git aaed022) says:
324 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
325 * that is initialized with the number of devices the application is
326 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
327 * an array of at least this many VkPhysicalDevice handles [...].
329 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
330 * overwrites the contents of the variable pointed to by
331 * pPhysicalDeviceCount with the number of physical devices in in the
332 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
333 * pPhysicalDeviceCount with the number of physical handles written to
336 if (!pPhysicalDevices
) {
337 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
338 } else if (*pPhysicalDeviceCount
>= 1) {
339 pPhysicalDevices
[0] = radv_physical_device_to_handle(&instance
->physicalDevice
);
340 *pPhysicalDeviceCount
= 1;
341 } else if (*pPhysicalDeviceCount
< instance
->physicalDeviceCount
) {
342 return VK_INCOMPLETE
;
344 *pPhysicalDeviceCount
= 0;
350 void radv_GetPhysicalDeviceFeatures(
351 VkPhysicalDevice physicalDevice
,
352 VkPhysicalDeviceFeatures
* pFeatures
)
354 // RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
356 memset(pFeatures
, 0, sizeof(*pFeatures
));
358 *pFeatures
= (VkPhysicalDeviceFeatures
) {
359 .robustBufferAccess
= true,
360 .fullDrawIndexUint32
= true,
361 .imageCubeArray
= true,
362 .independentBlend
= true,
363 .geometryShader
= false,
364 .tessellationShader
= false,
365 .sampleRateShading
= false,
366 .dualSrcBlend
= true,
368 .multiDrawIndirect
= true,
369 .drawIndirectFirstInstance
= true,
371 .depthBiasClamp
= true,
372 .fillModeNonSolid
= true,
377 .multiViewport
= false,
378 .samplerAnisotropy
= true,
379 .textureCompressionETC2
= false,
380 .textureCompressionASTC_LDR
= false,
381 .textureCompressionBC
= true,
382 .occlusionQueryPrecise
= true,
383 .pipelineStatisticsQuery
= false,
384 .vertexPipelineStoresAndAtomics
= true,
385 .fragmentStoresAndAtomics
= true,
386 .shaderTessellationAndGeometryPointSize
= true,
387 .shaderImageGatherExtended
= false,
388 .shaderStorageImageExtendedFormats
= false,
389 .shaderStorageImageMultisample
= false,
390 .shaderUniformBufferArrayDynamicIndexing
= true,
391 .shaderSampledImageArrayDynamicIndexing
= true,
392 .shaderStorageBufferArrayDynamicIndexing
= true,
393 .shaderStorageImageArrayDynamicIndexing
= true,
394 .shaderStorageImageReadWithoutFormat
= false,
395 .shaderStorageImageWriteWithoutFormat
= true,
396 .shaderClipDistance
= true,
397 .shaderCullDistance
= true,
398 .shaderFloat64
= false,
399 .shaderInt64
= false,
400 .shaderInt16
= false,
402 .variableMultisampleRate
= false,
403 .inheritedQueries
= false,
407 void radv_GetPhysicalDeviceProperties(
408 VkPhysicalDevice physicalDevice
,
409 VkPhysicalDeviceProperties
* pProperties
)
411 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
412 VkSampleCountFlags sample_counts
= 0xf;
413 VkPhysicalDeviceLimits limits
= {
414 .maxImageDimension1D
= (1 << 14),
415 .maxImageDimension2D
= (1 << 14),
416 .maxImageDimension3D
= (1 << 11),
417 .maxImageDimensionCube
= (1 << 14),
418 .maxImageArrayLayers
= (1 << 11),
419 .maxTexelBufferElements
= 128 * 1024 * 1024,
420 .maxUniformBufferRange
= UINT32_MAX
,
421 .maxStorageBufferRange
= UINT32_MAX
,
422 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
423 .maxMemoryAllocationCount
= UINT32_MAX
,
424 .maxSamplerAllocationCount
= 64 * 1024,
425 .bufferImageGranularity
= 64, /* A cache line */
426 .sparseAddressSpaceSize
= 0,
427 .maxBoundDescriptorSets
= MAX_SETS
,
428 .maxPerStageDescriptorSamplers
= 64,
429 .maxPerStageDescriptorUniformBuffers
= 64,
430 .maxPerStageDescriptorStorageBuffers
= 64,
431 .maxPerStageDescriptorSampledImages
= 64,
432 .maxPerStageDescriptorStorageImages
= 64,
433 .maxPerStageDescriptorInputAttachments
= 64,
434 .maxPerStageResources
= 128,
435 .maxDescriptorSetSamplers
= 256,
436 .maxDescriptorSetUniformBuffers
= 256,
437 .maxDescriptorSetUniformBuffersDynamic
= 256,
438 .maxDescriptorSetStorageBuffers
= 256,
439 .maxDescriptorSetStorageBuffersDynamic
= 256,
440 .maxDescriptorSetSampledImages
= 256,
441 .maxDescriptorSetStorageImages
= 256,
442 .maxDescriptorSetInputAttachments
= 256,
443 .maxVertexInputAttributes
= 32,
444 .maxVertexInputBindings
= 32,
445 .maxVertexInputAttributeOffset
= 2047,
446 .maxVertexInputBindingStride
= 2048,
447 .maxVertexOutputComponents
= 128,
448 .maxTessellationGenerationLevel
= 0,
449 .maxTessellationPatchSize
= 0,
450 .maxTessellationControlPerVertexInputComponents
= 0,
451 .maxTessellationControlPerVertexOutputComponents
= 0,
452 .maxTessellationControlPerPatchOutputComponents
= 0,
453 .maxTessellationControlTotalOutputComponents
= 0,
454 .maxTessellationEvaluationInputComponents
= 0,
455 .maxTessellationEvaluationOutputComponents
= 0,
456 .maxGeometryShaderInvocations
= 32,
457 .maxGeometryInputComponents
= 64,
458 .maxGeometryOutputComponents
= 128,
459 .maxGeometryOutputVertices
= 256,
460 .maxGeometryTotalOutputComponents
= 1024,
461 .maxFragmentInputComponents
= 128,
462 .maxFragmentOutputAttachments
= 8,
463 .maxFragmentDualSrcAttachments
= 2,
464 .maxFragmentCombinedOutputResources
= 8,
465 .maxComputeSharedMemorySize
= 32768,
466 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
467 .maxComputeWorkGroupInvocations
= 16 * 1024,
468 .maxComputeWorkGroupSize
= {
469 16 * 1024/*devinfo->max_cs_threads*/,
473 .subPixelPrecisionBits
= 4 /* FIXME */,
474 .subTexelPrecisionBits
= 4 /* FIXME */,
475 .mipmapPrecisionBits
= 4 /* FIXME */,
476 .maxDrawIndexedIndexValue
= UINT32_MAX
,
477 .maxDrawIndirectCount
= UINT32_MAX
,
478 .maxSamplerLodBias
= 16,
479 .maxSamplerAnisotropy
= 16,
480 .maxViewports
= MAX_VIEWPORTS
,
481 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
482 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
483 .viewportSubPixelBits
= 13, /* We take a float? */
484 .minMemoryMapAlignment
= 4096, /* A page */
485 .minTexelBufferOffsetAlignment
= 1,
486 .minUniformBufferOffsetAlignment
= 4,
487 .minStorageBufferOffsetAlignment
= 4,
488 .minTexelOffset
= -8,
490 .minTexelGatherOffset
= -8,
491 .maxTexelGatherOffset
= 7,
492 .minInterpolationOffset
= 0, /* FIXME */
493 .maxInterpolationOffset
= 0, /* FIXME */
494 .subPixelInterpolationOffsetBits
= 0, /* FIXME */
495 .maxFramebufferWidth
= (1 << 14),
496 .maxFramebufferHeight
= (1 << 14),
497 .maxFramebufferLayers
= (1 << 10),
498 .framebufferColorSampleCounts
= sample_counts
,
499 .framebufferDepthSampleCounts
= sample_counts
,
500 .framebufferStencilSampleCounts
= sample_counts
,
501 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
502 .maxColorAttachments
= MAX_RTS
,
503 .sampledImageColorSampleCounts
= sample_counts
,
504 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
505 .sampledImageDepthSampleCounts
= sample_counts
,
506 .sampledImageStencilSampleCounts
= sample_counts
,
507 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
508 .maxSampleMaskWords
= 1,
509 .timestampComputeAndGraphics
= false,
510 .timestampPeriod
= 100000.0 / pdevice
->rad_info
.clock_crystal_freq
,
511 .maxClipDistances
= 8,
512 .maxCullDistances
= 8,
513 .maxCombinedClipAndCullDistances
= 8,
514 .discreteQueuePriorities
= 1,
515 .pointSizeRange
= { 0.125, 255.875 },
516 .lineWidthRange
= { 0.0, 7.9921875 },
517 .pointSizeGranularity
= (1.0 / 8.0),
518 .lineWidthGranularity
= (1.0 / 128.0),
519 .strictLines
= false, /* FINISHME */
520 .standardSampleLocations
= true,
521 .optimalBufferCopyOffsetAlignment
= 128,
522 .optimalBufferCopyRowPitchAlignment
= 128,
523 .nonCoherentAtomSize
= 64,
526 *pProperties
= (VkPhysicalDeviceProperties
) {
527 .apiVersion
= VK_MAKE_VERSION(1, 0, 5),
530 .deviceID
= pdevice
->rad_info
.pci_id
,
531 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
,
533 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
536 strcpy(pProperties
->deviceName
, pdevice
->name
);
537 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
540 void radv_GetPhysicalDeviceQueueFamilyProperties(
541 VkPhysicalDevice physicalDevice
,
543 VkQueueFamilyProperties
* pQueueFamilyProperties
)
545 if (pQueueFamilyProperties
== NULL
) {
549 assert(*pCount
>= 1);
551 *pQueueFamilyProperties
= (VkQueueFamilyProperties
) {
552 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
553 VK_QUEUE_COMPUTE_BIT
|
554 VK_QUEUE_TRANSFER_BIT
,
556 .timestampValidBits
= 64,
557 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
561 void radv_GetPhysicalDeviceMemoryProperties(
562 VkPhysicalDevice physicalDevice
,
563 VkPhysicalDeviceMemoryProperties
* pMemoryProperties
)
565 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
567 pMemoryProperties
->memoryTypeCount
= 4;
568 pMemoryProperties
->memoryTypes
[0] = (VkMemoryType
) {
569 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
572 pMemoryProperties
->memoryTypes
[1] = (VkMemoryType
) {
573 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
574 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
577 pMemoryProperties
->memoryTypes
[2] = (VkMemoryType
) {
578 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
579 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
580 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
583 pMemoryProperties
->memoryTypes
[3] = (VkMemoryType
) {
584 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
585 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
586 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
590 pMemoryProperties
->memoryHeapCount
= 3;
591 pMemoryProperties
->memoryHeaps
[0] = (VkMemoryHeap
) {
592 .size
= physical_device
->rad_info
.vram_size
-
593 physical_device
->rad_info
.visible_vram_size
,
594 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
596 pMemoryProperties
->memoryHeaps
[1] = (VkMemoryHeap
) {
597 .size
= physical_device
->rad_info
.visible_vram_size
,
598 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
600 pMemoryProperties
->memoryHeaps
[2] = (VkMemoryHeap
) {
601 .size
= physical_device
->rad_info
.gart_size
,
607 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
)
609 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
610 queue
->device
= device
;
616 radv_queue_finish(struct radv_queue
*queue
)
620 VkResult
radv_CreateDevice(
621 VkPhysicalDevice physicalDevice
,
622 const VkDeviceCreateInfo
* pCreateInfo
,
623 const VkAllocationCallbacks
* pAllocator
,
626 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
628 struct radv_device
*device
;
630 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
632 for (uint32_t j
= 0; j
< ARRAY_SIZE(device_extensions
); j
++) {
633 if (strcmp(pCreateInfo
->ppEnabledExtensionNames
[i
],
634 device_extensions
[j
].extensionName
) == 0) {
640 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
643 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
645 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
647 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
649 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
650 device
->instance
= physical_device
->instance
;
652 device
->ws
= physical_device
->ws
;
654 device
->alloc
= *pAllocator
;
656 device
->alloc
= physical_device
->instance
->alloc
;
658 device
->hw_ctx
= device
->ws
->ctx_create(device
->ws
);
659 if (!device
->hw_ctx
) {
660 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
664 radv_queue_init(device
, &device
->queue
);
666 result
= radv_device_init_meta(device
);
667 if (result
!= VK_SUCCESS
) {
668 device
->ws
->ctx_destroy(device
->hw_ctx
);
671 device
->allow_fast_clears
= env_var_as_boolean("RADV_FAST_CLEARS", false);
672 device
->allow_dcc
= !env_var_as_boolean("RADV_DCC_DISABLE", false);
673 device
->shader_stats_dump
= env_var_as_boolean("RADV_SHADER_STATS", false);
675 if (device
->allow_fast_clears
&& device
->allow_dcc
)
676 radv_finishme("DCC fast clears have not been tested\n");
678 radv_device_init_msaa(device
);
679 device
->empty_cs
= device
->ws
->cs_create(device
->ws
, RING_GFX
);
680 radeon_emit(device
->empty_cs
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
681 radeon_emit(device
->empty_cs
, CONTEXT_CONTROL_LOAD_ENABLE(1));
682 radeon_emit(device
->empty_cs
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
683 device
->ws
->cs_finalize(device
->empty_cs
);
684 *pDevice
= radv_device_to_handle(device
);
687 vk_free(&device
->alloc
, device
);
691 void radv_DestroyDevice(
693 const VkAllocationCallbacks
* pAllocator
)
695 RADV_FROM_HANDLE(radv_device
, device
, _device
);
697 device
->ws
->ctx_destroy(device
->hw_ctx
);
698 radv_queue_finish(&device
->queue
);
699 radv_device_finish_meta(device
);
701 vk_free(&device
->alloc
, device
);
704 VkResult
radv_EnumerateInstanceExtensionProperties(
705 const char* pLayerName
,
706 uint32_t* pPropertyCount
,
707 VkExtensionProperties
* pProperties
)
710 if (pProperties
== NULL
) {
711 *pPropertyCount
= ARRAY_SIZE(global_extensions
);
715 for (i
= 0; i
< *pPropertyCount
; i
++)
716 memcpy(&pProperties
[i
], &global_extensions
[i
], sizeof(VkExtensionProperties
));
719 if (i
< ARRAY_SIZE(global_extensions
))
720 return VK_INCOMPLETE
;
725 VkResult
radv_EnumerateDeviceExtensionProperties(
726 VkPhysicalDevice physicalDevice
,
727 const char* pLayerName
,
728 uint32_t* pPropertyCount
,
729 VkExtensionProperties
* pProperties
)
733 if (pProperties
== NULL
) {
734 *pPropertyCount
= ARRAY_SIZE(device_extensions
);
738 for (i
= 0; i
< *pPropertyCount
; i
++)
739 memcpy(&pProperties
[i
], &device_extensions
[i
], sizeof(VkExtensionProperties
));
742 if (i
< ARRAY_SIZE(device_extensions
))
743 return VK_INCOMPLETE
;
747 VkResult
radv_EnumerateInstanceLayerProperties(
748 uint32_t* pPropertyCount
,
749 VkLayerProperties
* pProperties
)
751 if (pProperties
== NULL
) {
756 /* None supported at this time */
757 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
760 VkResult
radv_EnumerateDeviceLayerProperties(
761 VkPhysicalDevice physicalDevice
,
762 uint32_t* pPropertyCount
,
763 VkLayerProperties
* pProperties
)
765 if (pProperties
== NULL
) {
770 /* None supported at this time */
771 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
774 void radv_GetDeviceQueue(
776 uint32_t queueNodeIndex
,
780 RADV_FROM_HANDLE(radv_device
, device
, _device
);
782 assert(queueIndex
== 0);
784 *pQueue
= radv_queue_to_handle(&device
->queue
);
787 VkResult
radv_QueueSubmit(
789 uint32_t submitCount
,
790 const VkSubmitInfo
* pSubmits
,
793 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
794 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
795 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
796 struct radeon_winsys_ctx
*ctx
= queue
->device
->hw_ctx
;
799 for (uint32_t i
= 0; i
< submitCount
; i
++) {
800 struct radeon_winsys_cs
**cs_array
;
801 bool can_patch
= true;
803 if (!pSubmits
[i
].commandBufferCount
)
806 cs_array
= malloc(sizeof(struct radeon_winsys_cs
*) *
807 pSubmits
[i
].commandBufferCount
);
809 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
810 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
811 pSubmits
[i
].pCommandBuffers
[j
]);
812 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
814 cs_array
[j
] = cmd_buffer
->cs
;
815 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
818 ret
= queue
->device
->ws
->cs_submit(ctx
, cs_array
,
819 pSubmits
[i
].commandBufferCount
,
820 can_patch
, base_fence
);
822 radv_loge("failed to submit CS %d\n", i
);
828 ret
= queue
->device
->ws
->cs_submit(ctx
, &queue
->device
->empty_cs
,
829 1, false, base_fence
);
831 fence
->submitted
= true;
837 VkResult
radv_QueueWaitIdle(
840 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
842 queue
->device
->ws
->ctx_wait_idle(queue
->device
->hw_ctx
);
846 VkResult
radv_DeviceWaitIdle(
849 RADV_FROM_HANDLE(radv_device
, device
, _device
);
851 device
->ws
->ctx_wait_idle(device
->hw_ctx
);
855 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
859 return radv_lookup_entrypoint(pName
);
862 /* The loader wants us to expose a second GetInstanceProcAddr function
863 * to work around certain LD_PRELOAD issues seen in apps.
866 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
871 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
875 return radv_GetInstanceProcAddr(instance
, pName
);
878 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
882 return radv_lookup_entrypoint(pName
);
885 VkResult
radv_AllocateMemory(
887 const VkMemoryAllocateInfo
* pAllocateInfo
,
888 const VkAllocationCallbacks
* pAllocator
,
889 VkDeviceMemory
* pMem
)
891 RADV_FROM_HANDLE(radv_device
, device
, _device
);
892 struct radv_device_memory
*mem
;
894 enum radeon_bo_domain domain
;
896 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
898 if (pAllocateInfo
->allocationSize
== 0) {
899 /* Apparently, this is allowed */
900 *pMem
= VK_NULL_HANDLE
;
904 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
905 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
907 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
909 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
910 if (pAllocateInfo
->memoryTypeIndex
== 1 || pAllocateInfo
->memoryTypeIndex
== 3)
911 domain
= RADEON_DOMAIN_GTT
;
913 domain
= RADEON_DOMAIN_VRAM
;
915 if (pAllocateInfo
->memoryTypeIndex
== 0)
916 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
918 flags
|= RADEON_FLAG_CPU_ACCESS
;
920 if (pAllocateInfo
->memoryTypeIndex
== 1)
921 flags
|= RADEON_FLAG_GTT_WC
;
923 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, 32768,
927 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
930 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
932 *pMem
= radv_device_memory_to_handle(mem
);
937 vk_free2(&device
->alloc
, pAllocator
, mem
);
942 void radv_FreeMemory(
945 const VkAllocationCallbacks
* pAllocator
)
947 RADV_FROM_HANDLE(radv_device
, device
, _device
);
948 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
953 device
->ws
->buffer_destroy(mem
->bo
);
956 vk_free2(&device
->alloc
, pAllocator
, mem
);
959 VkResult
radv_MapMemory(
961 VkDeviceMemory _memory
,
964 VkMemoryMapFlags flags
,
967 RADV_FROM_HANDLE(radv_device
, device
, _device
);
968 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
975 *ppData
= device
->ws
->buffer_map(mem
->bo
);
981 return VK_ERROR_MEMORY_MAP_FAILED
;
984 void radv_UnmapMemory(
986 VkDeviceMemory _memory
)
988 RADV_FROM_HANDLE(radv_device
, device
, _device
);
989 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
994 device
->ws
->buffer_unmap(mem
->bo
);
997 VkResult
radv_FlushMappedMemoryRanges(
999 uint32_t memoryRangeCount
,
1000 const VkMappedMemoryRange
* pMemoryRanges
)
1005 VkResult
radv_InvalidateMappedMemoryRanges(
1007 uint32_t memoryRangeCount
,
1008 const VkMappedMemoryRange
* pMemoryRanges
)
1013 void radv_GetBufferMemoryRequirements(
1016 VkMemoryRequirements
* pMemoryRequirements
)
1018 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1020 /* The Vulkan spec (git aaed022) says:
1022 * memoryTypeBits is a bitfield and contains one bit set for every
1023 * supported memory type for the resource. The bit `1<<i` is set if and
1024 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1025 * structure for the physical device is supported.
1027 * We support exactly one memory type.
1029 pMemoryRequirements
->memoryTypeBits
= 0x7;
1031 pMemoryRequirements
->size
= buffer
->size
;
1032 pMemoryRequirements
->alignment
= 16;
1035 void radv_GetImageMemoryRequirements(
1038 VkMemoryRequirements
* pMemoryRequirements
)
1040 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1042 /* The Vulkan spec (git aaed022) says:
1044 * memoryTypeBits is a bitfield and contains one bit set for every
1045 * supported memory type for the resource. The bit `1<<i` is set if and
1046 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1047 * structure for the physical device is supported.
1049 * We support exactly one memory type.
1051 pMemoryRequirements
->memoryTypeBits
= 0x7;
1053 pMemoryRequirements
->size
= image
->size
;
1054 pMemoryRequirements
->alignment
= image
->alignment
;
1057 void radv_GetImageSparseMemoryRequirements(
1060 uint32_t* pSparseMemoryRequirementCount
,
1061 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1066 void radv_GetDeviceMemoryCommitment(
1068 VkDeviceMemory memory
,
1069 VkDeviceSize
* pCommittedMemoryInBytes
)
1071 *pCommittedMemoryInBytes
= 0;
1074 VkResult
radv_BindBufferMemory(
1077 VkDeviceMemory _memory
,
1078 VkDeviceSize memoryOffset
)
1080 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1081 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1084 buffer
->bo
= mem
->bo
;
1085 buffer
->offset
= memoryOffset
;
1094 VkResult
radv_BindImageMemory(
1097 VkDeviceMemory _memory
,
1098 VkDeviceSize memoryOffset
)
1100 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1101 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1104 image
->bo
= mem
->bo
;
1105 image
->offset
= memoryOffset
;
1114 VkResult
radv_QueueBindSparse(
1116 uint32_t bindInfoCount
,
1117 const VkBindSparseInfo
* pBindInfo
,
1120 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1123 VkResult
radv_CreateFence(
1125 const VkFenceCreateInfo
* pCreateInfo
,
1126 const VkAllocationCallbacks
* pAllocator
,
1129 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1130 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
1132 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1135 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1137 memset(fence
, 0, sizeof(*fence
));
1138 fence
->submitted
= false;
1139 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
1140 fence
->fence
= device
->ws
->create_fence();
1143 *pFence
= radv_fence_to_handle(fence
);
1148 void radv_DestroyFence(
1151 const VkAllocationCallbacks
* pAllocator
)
1153 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1154 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1158 device
->ws
->destroy_fence(fence
->fence
);
1159 vk_free2(&device
->alloc
, pAllocator
, fence
);
1162 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
1164 uint64_t current_time
;
1167 clock_gettime(CLOCK_MONOTONIC
, &tv
);
1168 current_time
= tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
1170 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
1172 return current_time
+ timeout
;
1175 VkResult
radv_WaitForFences(
1177 uint32_t fenceCount
,
1178 const VkFence
* pFences
,
1182 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1183 timeout
= radv_get_absolute_timeout(timeout
);
1185 if (!waitAll
&& fenceCount
> 1) {
1186 fprintf(stderr
, "radv: WaitForFences without waitAll not implemented yet\n");
1189 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
1190 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1191 bool expired
= false;
1193 if (fence
->signalled
)
1196 if (!fence
->submitted
)
1199 expired
= device
->ws
->fence_wait(device
->ws
, fence
->fence
, true, timeout
);
1203 fence
->signalled
= true;
1209 VkResult
radv_ResetFences(VkDevice device
,
1210 uint32_t fenceCount
,
1211 const VkFence
*pFences
)
1213 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
1214 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1215 fence
->submitted
= fence
->signalled
= false;
1221 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
1223 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1224 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1226 if (fence
->signalled
)
1228 if (!fence
->submitted
)
1229 return VK_NOT_READY
;
1231 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
1232 return VK_NOT_READY
;
1238 // Queue semaphore functions
1240 VkResult
radv_CreateSemaphore(
1242 const VkSemaphoreCreateInfo
* pCreateInfo
,
1243 const VkAllocationCallbacks
* pAllocator
,
1244 VkSemaphore
* pSemaphore
)
1246 /* The DRM execbuffer ioctl always execute in-oder, even between different
1247 * rings. As such, there's nothing to do for the user space semaphore.
1250 *pSemaphore
= (VkSemaphore
)1;
1255 void radv_DestroySemaphore(
1257 VkSemaphore semaphore
,
1258 const VkAllocationCallbacks
* pAllocator
)
1262 VkResult
radv_CreateEvent(
1264 const VkEventCreateInfo
* pCreateInfo
,
1265 const VkAllocationCallbacks
* pAllocator
,
1268 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1269 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
1271 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1274 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1276 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
1278 RADEON_FLAG_CPU_ACCESS
);
1280 vk_free2(&device
->alloc
, pAllocator
, event
);
1281 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1284 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
1286 *pEvent
= radv_event_to_handle(event
);
1291 void radv_DestroyEvent(
1294 const VkAllocationCallbacks
* pAllocator
)
1296 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1297 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1301 device
->ws
->buffer_destroy(event
->bo
);
1302 vk_free2(&device
->alloc
, pAllocator
, event
);
1305 VkResult
radv_GetEventStatus(
1309 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1311 if (*event
->map
== 1)
1312 return VK_EVENT_SET
;
1313 return VK_EVENT_RESET
;
1316 VkResult
radv_SetEvent(
1320 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1326 VkResult
radv_ResetEvent(
1330 RADV_FROM_HANDLE(radv_event
, event
, _event
);
1336 VkResult
radv_CreateBuffer(
1338 const VkBufferCreateInfo
* pCreateInfo
,
1339 const VkAllocationCallbacks
* pAllocator
,
1342 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1343 struct radv_buffer
*buffer
;
1345 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
1347 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
1348 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1350 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1352 buffer
->size
= pCreateInfo
->size
;
1353 buffer
->usage
= pCreateInfo
->usage
;
1357 *pBuffer
= radv_buffer_to_handle(buffer
);
1362 void radv_DestroyBuffer(
1365 const VkAllocationCallbacks
* pAllocator
)
1367 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1368 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1373 vk_free2(&device
->alloc
, pAllocator
, buffer
);
1376 static inline unsigned
1377 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
1380 return image
->surface
.stencil_tiling_index
[level
];
1382 return image
->surface
.tiling_index
[level
];
1386 radv_initialise_color_surface(struct radv_device
*device
,
1387 struct radv_color_buffer_info
*cb
,
1388 struct radv_image_view
*iview
)
1390 const struct vk_format_description
*desc
;
1391 unsigned ntype
, format
, swap
, endian
;
1392 unsigned blend_clamp
= 0, blend_bypass
= 0;
1393 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
1395 const struct radeon_surf
*surf
= &iview
->image
->surface
;
1396 const struct radeon_surf_level
*level_info
= &surf
->level
[iview
->base_mip
];
1398 desc
= vk_format_description(iview
->vk_format
);
1400 memset(cb
, 0, sizeof(*cb
));
1402 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1403 va
+= level_info
->offset
;
1404 cb
->cb_color_base
= va
>> 8;
1406 /* CMASK variables */
1407 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1408 va
+= iview
->image
->cmask
.offset
;
1409 cb
->cb_color_cmask
= va
>> 8;
1410 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
1412 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1413 va
+= iview
->image
->dcc_offset
;
1414 cb
->cb_dcc_base
= va
>> 8;
1416 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
1417 S_028C6C_SLICE_MAX(iview
->base_layer
+ iview
->extent
.depth
- 1);
1419 cb
->micro_tile_mode
= iview
->image
->surface
.micro_tile_mode
;
1420 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
1421 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
1422 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
1424 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
1425 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
1427 /* Intensity is implemented as Red, so treat it that way. */
1428 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
) |
1429 S_028C74_TILE_MODE_INDEX(tile_mode_index
);
1431 if (iview
->image
->samples
> 1) {
1432 unsigned log_samples
= util_logbase2(iview
->image
->samples
);
1434 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
1435 S_028C74_NUM_FRAGMENTS(log_samples
);
1438 if (iview
->image
->fmask
.size
) {
1439 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
1440 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
)
1441 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
1442 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
1443 cb
->cb_color_fmask
= va
>> 8;
1444 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
1446 /* This must be set for fast clear to work without FMASK. */
1447 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
)
1448 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
1449 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
1450 cb
->cb_color_fmask
= cb
->cb_color_base
;
1451 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
1454 ntype
= radv_translate_color_numformat(iview
->vk_format
,
1456 vk_format_get_first_non_void_channel(iview
->vk_format
));
1457 format
= radv_translate_colorformat(iview
->vk_format
);
1458 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
1459 radv_finishme("Illegal color\n");
1460 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
1461 endian
= radv_colorformat_endian_swap(format
);
1463 /* blend clamp should be set for all NORM/SRGB types */
1464 if (ntype
== V_028C70_NUMBER_UNORM
||
1465 ntype
== V_028C70_NUMBER_SNORM
||
1466 ntype
== V_028C70_NUMBER_SRGB
)
1469 /* set blend bypass according to docs if SINT/UINT or
1470 8/24 COLOR variants */
1471 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
1472 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
1473 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
1478 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
1479 (format
== V_028C70_COLOR_8
||
1480 format
== V_028C70_COLOR_8_8
||
1481 format
== V_028C70_COLOR_8_8_8_8
))
1482 ->color_is_int8
= true;
1484 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
1485 S_028C70_COMP_SWAP(swap
) |
1486 S_028C70_BLEND_CLAMP(blend_clamp
) |
1487 S_028C70_BLEND_BYPASS(blend_bypass
) |
1488 S_028C70_SIMPLE_FLOAT(1) |
1489 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
1490 ntype
!= V_028C70_NUMBER_SNORM
&&
1491 ntype
!= V_028C70_NUMBER_SRGB
&&
1492 format
!= V_028C70_COLOR_8_24
&&
1493 format
!= V_028C70_COLOR_24_8
) |
1494 S_028C70_NUMBER_TYPE(ntype
) |
1495 S_028C70_ENDIAN(endian
);
1496 if (iview
->image
->samples
> 1)
1497 if (iview
->image
->fmask
.size
)
1498 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
1500 if (iview
->image
->cmask
.size
&& device
->allow_fast_clears
)
1501 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
1503 if (iview
->image
->surface
.dcc_size
&& level_info
->dcc_enabled
)
1504 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
1506 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= VI
) {
1507 unsigned max_uncompressed_block_size
= 2;
1508 if (iview
->image
->samples
> 1) {
1509 if (iview
->image
->surface
.bpe
== 1)
1510 max_uncompressed_block_size
= 0;
1511 else if (iview
->image
->surface
.bpe
== 2)
1512 max_uncompressed_block_size
= 1;
1515 cb
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
1516 S_028C78_INDEPENDENT_64B_BLOCKS(1);
1519 /* This must be set for fast clear to work without FMASK. */
1520 if (!iview
->image
->fmask
.size
&&
1521 device
->instance
->physicalDevice
.rad_info
.chip_class
== SI
) {
1522 unsigned bankh
= util_logbase2(iview
->image
->surface
.bankh
);
1523 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
1528 radv_initialise_ds_surface(struct radv_device
*device
,
1529 struct radv_ds_buffer_info
*ds
,
1530 struct radv_image_view
*iview
)
1532 unsigned level
= iview
->base_mip
;
1534 uint64_t va
, s_offs
, z_offs
;
1535 const struct radeon_surf_level
*level_info
= &iview
->image
->surface
.level
[level
];
1536 memset(ds
, 0, sizeof(*ds
));
1537 switch (iview
->vk_format
) {
1538 case VK_FORMAT_D24_UNORM_S8_UINT
:
1539 case VK_FORMAT_X8_D24_UNORM_PACK32
:
1540 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
1541 ds
->offset_scale
= 2.0f
;
1543 case VK_FORMAT_D16_UNORM
:
1544 case VK_FORMAT_D16_UNORM_S8_UINT
:
1545 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
1546 ds
->offset_scale
= 4.0f
;
1548 case VK_FORMAT_D32_SFLOAT
:
1549 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
1550 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
1551 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
1552 ds
->offset_scale
= 1.0f
;
1558 format
= radv_translate_dbformat(iview
->vk_format
);
1559 if (format
== V_028040_Z_INVALID
) {
1560 fprintf(stderr
, "Invalid DB format: %d, disabling DB.\n", iview
->vk_format
);
1563 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
1564 s_offs
= z_offs
= va
;
1565 z_offs
+= iview
->image
->surface
.level
[level
].offset
;
1566 s_offs
+= iview
->image
->surface
.stencil_level
[level
].offset
;
1568 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
1569 S_028008_SLICE_MAX(iview
->base_layer
+ iview
->extent
.depth
- 1);
1570 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(1);
1571 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
1573 if (iview
->image
->samples
> 1)
1574 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->samples
));
1576 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
)
1577 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_8
);
1579 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_INVALID
);
1581 if (device
->instance
->physicalDevice
.rad_info
.chip_class
>= CIK
) {
1582 struct radeon_info
*info
= &device
->instance
->physicalDevice
.rad_info
;
1583 unsigned tiling_index
= iview
->image
->surface
.tiling_index
[level
];
1584 unsigned stencil_index
= iview
->image
->surface
.stencil_tiling_index
[level
];
1585 unsigned macro_index
= iview
->image
->surface
.macro_tile_index
;
1586 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
1587 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
1588 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
1590 ds
->db_depth_info
|=
1591 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
1592 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
1593 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
1594 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
1595 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
1596 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
1597 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
1598 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
1600 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
1601 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
1602 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
1603 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
1606 if (iview
->image
->htile
.size
&& !level
) {
1607 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
1608 S_028040_ALLOW_EXPCLEAR(1);
1610 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
) {
1611 /* Workaround: For a not yet understood reason, the
1612 * combination of MSAA, fast stencil clear and stencil
1613 * decompress messes with subsequent stencil buffer
1614 * uses. Problem was reproduced on Verde, Bonaire,
1615 * Tonga, and Carrizo.
1617 * Disabling EXPCLEAR works around the problem.
1619 * Check piglit's arb_texture_multisample-stencil-clear
1620 * test if you want to try changing this.
1622 if (iview
->image
->samples
<= 1)
1623 ds
->db_stencil_info
|= S_028044_ALLOW_EXPCLEAR(1);
1625 /* Use all of the htile_buffer for depth if there's no stencil. */
1626 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
1628 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+
1629 iview
->image
->htile
.offset
;
1630 ds
->db_htile_data_base
= va
>> 8;
1631 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
1633 ds
->db_htile_data_base
= 0;
1634 ds
->db_htile_surface
= 0;
1637 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
1638 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
1640 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
1641 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
1642 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
1645 VkResult
radv_CreateFramebuffer(
1647 const VkFramebufferCreateInfo
* pCreateInfo
,
1648 const VkAllocationCallbacks
* pAllocator
,
1649 VkFramebuffer
* pFramebuffer
)
1651 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1652 struct radv_framebuffer
*framebuffer
;
1654 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
1656 size_t size
= sizeof(*framebuffer
) +
1657 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
1658 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
1659 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1660 if (framebuffer
== NULL
)
1661 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1663 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
1664 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
1665 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
1666 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
1667 framebuffer
->attachments
[i
].attachment
= iview
;
1668 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
1669 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
1670 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
1671 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
1675 framebuffer
->width
= pCreateInfo
->width
;
1676 framebuffer
->height
= pCreateInfo
->height
;
1677 framebuffer
->layers
= pCreateInfo
->layers
;
1679 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
1683 void radv_DestroyFramebuffer(
1686 const VkAllocationCallbacks
* pAllocator
)
1688 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1689 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
1693 vk_free2(&device
->alloc
, pAllocator
, fb
);
1696 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
1698 switch (address_mode
) {
1699 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
1700 return V_008F30_SQ_TEX_WRAP
;
1701 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
1702 return V_008F30_SQ_TEX_MIRROR
;
1703 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
1704 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
1705 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
1706 return V_008F30_SQ_TEX_CLAMP_BORDER
;
1707 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
1708 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
1710 unreachable("illegal tex wrap mode");
1716 radv_tex_compare(VkCompareOp op
)
1719 case VK_COMPARE_OP_NEVER
:
1720 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
1721 case VK_COMPARE_OP_LESS
:
1722 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
1723 case VK_COMPARE_OP_EQUAL
:
1724 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
1725 case VK_COMPARE_OP_LESS_OR_EQUAL
:
1726 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
1727 case VK_COMPARE_OP_GREATER
:
1728 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
1729 case VK_COMPARE_OP_NOT_EQUAL
:
1730 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
1731 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
1732 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
1733 case VK_COMPARE_OP_ALWAYS
:
1734 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
1736 unreachable("illegal compare mode");
1742 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
1745 case VK_FILTER_NEAREST
:
1746 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
1747 V_008F38_SQ_TEX_XY_FILTER_POINT
);
1748 case VK_FILTER_LINEAR
:
1749 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
1750 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
1751 case VK_FILTER_CUBIC_IMG
:
1753 fprintf(stderr
, "illegal texture filter");
1759 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
1762 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
1763 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
1764 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
1765 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
1767 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
1772 radv_tex_bordercolor(VkBorderColor bcolor
)
1775 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
1776 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
1777 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
1778 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
1779 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
1780 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
1781 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
1782 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
1783 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
1791 radv_tex_aniso_filter(unsigned filter
)
1805 radv_init_sampler(struct radv_device
*device
,
1806 struct radv_sampler
*sampler
,
1807 const VkSamplerCreateInfo
*pCreateInfo
)
1809 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
1810 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
1811 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
1812 bool is_vi
= (device
->instance
->physicalDevice
.rad_info
.chip_class
>= VI
);
1814 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
1815 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
1816 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
1817 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
1818 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
1819 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
1820 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
1821 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
1822 S_008F30_DISABLE_CUBE_WRAP(0) |
1823 S_008F30_COMPAT_MODE(is_vi
));
1824 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
1825 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
1826 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
1827 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
1828 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
1829 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
1830 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
1831 S_008F38_MIP_POINT_PRECLAMP(1) |
1832 S_008F38_DISABLE_LSB_CEIL(1) |
1833 S_008F38_FILTER_PREC_FIX(1) |
1834 S_008F38_ANISO_OVERRIDE(is_vi
));
1835 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
1836 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
1839 VkResult
radv_CreateSampler(
1841 const VkSamplerCreateInfo
* pCreateInfo
,
1842 const VkAllocationCallbacks
* pAllocator
,
1843 VkSampler
* pSampler
)
1845 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1846 struct radv_sampler
*sampler
;
1848 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
1850 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
1851 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1853 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1855 radv_init_sampler(device
, sampler
, pCreateInfo
);
1856 *pSampler
= radv_sampler_to_handle(sampler
);
1861 void radv_DestroySampler(
1864 const VkAllocationCallbacks
* pAllocator
)
1866 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1867 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
1871 vk_free2(&device
->alloc
, pAllocator
, sampler
);