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
32 #include "radv_private.h"
34 #include "util/disk_cache.h"
35 #include "util/strtod.h"
36 #include "util/vk_util.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"
48 radv_device_get_cache_uuid(enum radeon_family family
, void *uuid
)
50 uint32_t mesa_timestamp
, llvm_timestamp
;
52 memset(uuid
, 0, VK_UUID_SIZE
);
53 if (!disk_cache_get_function_timestamp(radv_device_get_cache_uuid
, &mesa_timestamp
) ||
54 !disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
, &llvm_timestamp
))
57 memcpy(uuid
, &mesa_timestamp
, 4);
58 memcpy((char*)uuid
+ 4, &llvm_timestamp
, 4);
59 memcpy((char*)uuid
+ 8, &f
, 2);
60 snprintf((char*)uuid
+ 10, VK_UUID_SIZE
- 10, "radv");
64 static const VkExtensionProperties instance_extensions
[] = {
66 .extensionName
= VK_KHR_SURFACE_EXTENSION_NAME
,
69 #ifdef VK_USE_PLATFORM_XCB_KHR
71 .extensionName
= VK_KHR_XCB_SURFACE_EXTENSION_NAME
,
75 #ifdef VK_USE_PLATFORM_XLIB_KHR
77 .extensionName
= VK_KHR_XLIB_SURFACE_EXTENSION_NAME
,
81 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
83 .extensionName
= VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME
,
89 static const VkExtensionProperties common_device_extensions
[] = {
91 .extensionName
= VK_KHR_MAINTENANCE1_EXTENSION_NAME
,
95 .extensionName
= VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME
,
99 .extensionName
= VK_KHR_SWAPCHAIN_EXTENSION_NAME
,
103 .extensionName
= VK_AMD_DRAW_INDIRECT_COUNT_EXTENSION_NAME
,
107 .extensionName
= VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME
,
111 .extensionName
= VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME
,
115 .extensionName
= VK_NV_DEDICATED_ALLOCATION_EXTENSION_NAME
,
121 radv_extensions_register(struct radv_instance
*instance
,
122 struct radv_extensions
*extensions
,
123 const VkExtensionProperties
*new_ext
,
127 VkExtensionProperties
*new_ptr
;
129 assert(new_ext
&& num_ext
> 0);
132 return VK_ERROR_INITIALIZATION_FAILED
;
134 new_size
= (extensions
->num_ext
+ num_ext
) * sizeof(VkExtensionProperties
);
135 new_ptr
= vk_realloc(&instance
->alloc
, extensions
->ext_array
,
136 new_size
, 8, VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
138 /* Old array continues to be valid, update nothing */
140 return VK_ERROR_OUT_OF_HOST_MEMORY
;
142 memcpy(&new_ptr
[extensions
->num_ext
], new_ext
,
143 num_ext
* sizeof(VkExtensionProperties
));
144 extensions
->ext_array
= new_ptr
;
145 extensions
->num_ext
+= num_ext
;
151 radv_extensions_finish(struct radv_instance
*instance
,
152 struct radv_extensions
*extensions
)
157 radv_loge("Attemted to free invalid extension struct\n");
159 if (extensions
->ext_array
)
160 vk_free(&instance
->alloc
, extensions
->ext_array
);
164 is_extension_enabled(const VkExtensionProperties
*extensions
,
168 assert(extensions
&& name
);
170 for (uint32_t i
= 0; i
< num_ext
; i
++) {
171 if (strcmp(name
, extensions
[i
].extensionName
) == 0)
179 radv_physical_device_init(struct radv_physical_device
*device
,
180 struct radv_instance
*instance
,
184 drmVersionPtr version
;
187 fd
= open(path
, O_RDWR
| O_CLOEXEC
);
189 return VK_ERROR_INCOMPATIBLE_DRIVER
;
191 version
= drmGetVersion(fd
);
194 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
195 "failed to get version %s: %m", path
);
198 if (strcmp(version
->name
, "amdgpu")) {
199 drmFreeVersion(version
);
201 return VK_ERROR_INCOMPATIBLE_DRIVER
;
203 drmFreeVersion(version
);
205 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
206 device
->instance
= instance
;
207 assert(strlen(path
) < ARRAY_SIZE(device
->path
));
208 strncpy(device
->path
, path
, ARRAY_SIZE(device
->path
));
210 device
->ws
= radv_amdgpu_winsys_create(fd
, instance
->debug_flags
);
212 result
= VK_ERROR_INCOMPATIBLE_DRIVER
;
216 device
->local_fd
= fd
;
217 device
->ws
->query_info(device
->ws
, &device
->rad_info
);
218 result
= radv_init_wsi(device
);
219 if (result
!= VK_SUCCESS
) {
220 device
->ws
->destroy(device
->ws
);
224 if (radv_device_get_cache_uuid(device
->rad_info
.family
, device
->uuid
)) {
225 radv_finish_wsi(device
);
226 device
->ws
->destroy(device
->ws
);
227 result
= vk_errorf(VK_ERROR_INITIALIZATION_FAILED
,
228 "cannot generate UUID");
232 result
= radv_extensions_register(instance
,
234 common_device_extensions
,
235 ARRAY_SIZE(common_device_extensions
));
236 if (result
!= VK_SUCCESS
)
239 fprintf(stderr
, "WARNING: radv is not a conformant vulkan implementation, testing use only.\n");
240 device
->name
= device
->rad_info
.name
;
250 radv_physical_device_finish(struct radv_physical_device
*device
)
252 radv_extensions_finish(device
->instance
, &device
->extensions
);
253 radv_finish_wsi(device
);
254 device
->ws
->destroy(device
->ws
);
255 close(device
->local_fd
);
260 default_alloc_func(void *pUserData
, size_t size
, size_t align
,
261 VkSystemAllocationScope allocationScope
)
267 default_realloc_func(void *pUserData
, void *pOriginal
, size_t size
,
268 size_t align
, VkSystemAllocationScope allocationScope
)
270 return realloc(pOriginal
, size
);
274 default_free_func(void *pUserData
, void *pMemory
)
279 static const VkAllocationCallbacks default_alloc
= {
281 .pfnAllocation
= default_alloc_func
,
282 .pfnReallocation
= default_realloc_func
,
283 .pfnFree
= default_free_func
,
286 static const struct debug_control radv_debug_options
[] = {
287 {"nofastclears", RADV_DEBUG_NO_FAST_CLEARS
},
288 {"nodcc", RADV_DEBUG_NO_DCC
},
289 {"shaders", RADV_DEBUG_DUMP_SHADERS
},
290 {"nocache", RADV_DEBUG_NO_CACHE
},
291 {"shaderstats", RADV_DEBUG_DUMP_SHADER_STATS
},
292 {"nohiz", RADV_DEBUG_NO_HIZ
},
293 {"nocompute", RADV_DEBUG_NO_COMPUTE_QUEUE
},
294 {"unsafemath", RADV_DEBUG_UNSAFE_MATH
},
295 {"allbos", RADV_DEBUG_ALL_BOS
},
296 {"noibs", RADV_DEBUG_NO_IBS
},
300 VkResult
radv_CreateInstance(
301 const VkInstanceCreateInfo
* pCreateInfo
,
302 const VkAllocationCallbacks
* pAllocator
,
303 VkInstance
* pInstance
)
305 struct radv_instance
*instance
;
307 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO
);
309 uint32_t client_version
;
310 if (pCreateInfo
->pApplicationInfo
&&
311 pCreateInfo
->pApplicationInfo
->apiVersion
!= 0) {
312 client_version
= pCreateInfo
->pApplicationInfo
->apiVersion
;
314 client_version
= VK_MAKE_VERSION(1, 0, 0);
317 if (VK_MAKE_VERSION(1, 0, 0) > client_version
||
318 client_version
> VK_MAKE_VERSION(1, 0, 0xfff)) {
319 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER
,
320 "Client requested version %d.%d.%d",
321 VK_VERSION_MAJOR(client_version
),
322 VK_VERSION_MINOR(client_version
),
323 VK_VERSION_PATCH(client_version
));
326 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
327 if (!is_extension_enabled(instance_extensions
,
328 ARRAY_SIZE(instance_extensions
),
329 pCreateInfo
->ppEnabledExtensionNames
[i
]))
330 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
333 instance
= vk_alloc2(&default_alloc
, pAllocator
, sizeof(*instance
), 8,
334 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE
);
336 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
338 memset(instance
, 0, sizeof(*instance
));
340 instance
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
343 instance
->alloc
= *pAllocator
;
345 instance
->alloc
= default_alloc
;
347 instance
->apiVersion
= client_version
;
348 instance
->physicalDeviceCount
= -1;
352 VG(VALGRIND_CREATE_MEMPOOL(instance
, 0, false));
354 instance
->debug_flags
= parse_debug_string(getenv("RADV_DEBUG"),
357 *pInstance
= radv_instance_to_handle(instance
);
362 void radv_DestroyInstance(
363 VkInstance _instance
,
364 const VkAllocationCallbacks
* pAllocator
)
366 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
371 for (int i
= 0; i
< instance
->physicalDeviceCount
; ++i
) {
372 radv_physical_device_finish(instance
->physicalDevices
+ i
);
375 VG(VALGRIND_DESTROY_MEMPOOL(instance
));
379 vk_free(&instance
->alloc
, instance
);
382 VkResult
radv_EnumeratePhysicalDevices(
383 VkInstance _instance
,
384 uint32_t* pPhysicalDeviceCount
,
385 VkPhysicalDevice
* pPhysicalDevices
)
387 RADV_FROM_HANDLE(radv_instance
, instance
, _instance
);
390 if (instance
->physicalDeviceCount
< 0) {
392 instance
->physicalDeviceCount
= 0;
393 for (unsigned i
= 0; i
< RADV_MAX_DRM_DEVICES
; i
++) {
394 snprintf(path
, sizeof(path
), "/dev/dri/renderD%d", 128 + i
);
395 result
= radv_physical_device_init(instance
->physicalDevices
+
396 instance
->physicalDeviceCount
,
398 if (result
== VK_SUCCESS
)
399 ++instance
->physicalDeviceCount
;
400 else if (result
!= VK_ERROR_INCOMPATIBLE_DRIVER
)
405 if (!pPhysicalDevices
) {
406 *pPhysicalDeviceCount
= instance
->physicalDeviceCount
;
408 *pPhysicalDeviceCount
= MIN2(*pPhysicalDeviceCount
, instance
->physicalDeviceCount
);
409 for (unsigned i
= 0; i
< *pPhysicalDeviceCount
; ++i
)
410 pPhysicalDevices
[i
] = radv_physical_device_to_handle(instance
->physicalDevices
+ i
);
413 return *pPhysicalDeviceCount
< instance
->physicalDeviceCount
? VK_INCOMPLETE
417 void radv_GetPhysicalDeviceFeatures(
418 VkPhysicalDevice physicalDevice
,
419 VkPhysicalDeviceFeatures
* pFeatures
)
421 // RADV_FROM_HANDLE(radv_physical_device, pdevice, physicalDevice);
423 memset(pFeatures
, 0, sizeof(*pFeatures
));
425 *pFeatures
= (VkPhysicalDeviceFeatures
) {
426 .robustBufferAccess
= true,
427 .fullDrawIndexUint32
= true,
428 .imageCubeArray
= true,
429 .independentBlend
= true,
430 .geometryShader
= true,
431 .tessellationShader
= false,
432 .sampleRateShading
= false,
433 .dualSrcBlend
= true,
435 .multiDrawIndirect
= true,
436 .drawIndirectFirstInstance
= true,
438 .depthBiasClamp
= true,
439 .fillModeNonSolid
= true,
444 .multiViewport
= true,
445 .samplerAnisotropy
= true,
446 .textureCompressionETC2
= false,
447 .textureCompressionASTC_LDR
= false,
448 .textureCompressionBC
= true,
449 .occlusionQueryPrecise
= true,
450 .pipelineStatisticsQuery
= false,
451 .vertexPipelineStoresAndAtomics
= true,
452 .fragmentStoresAndAtomics
= true,
453 .shaderTessellationAndGeometryPointSize
= true,
454 .shaderImageGatherExtended
= true,
455 .shaderStorageImageExtendedFormats
= true,
456 .shaderStorageImageMultisample
= false,
457 .shaderUniformBufferArrayDynamicIndexing
= true,
458 .shaderSampledImageArrayDynamicIndexing
= true,
459 .shaderStorageBufferArrayDynamicIndexing
= true,
460 .shaderStorageImageArrayDynamicIndexing
= true,
461 .shaderStorageImageReadWithoutFormat
= true,
462 .shaderStorageImageWriteWithoutFormat
= true,
463 .shaderClipDistance
= true,
464 .shaderCullDistance
= true,
465 .shaderFloat64
= true,
466 .shaderInt64
= false,
467 .shaderInt16
= false,
468 .variableMultisampleRate
= false,
469 .inheritedQueries
= false,
473 void radv_GetPhysicalDeviceFeatures2KHR(
474 VkPhysicalDevice physicalDevice
,
475 VkPhysicalDeviceFeatures2KHR
*pFeatures
)
477 return radv_GetPhysicalDeviceFeatures(physicalDevice
, &pFeatures
->features
);
480 static uint32_t radv_get_driver_version()
482 const char *minor_string
= strchr(VERSION
, '.');
483 const char *patch_string
= minor_string
? strchr(minor_string
+ 1, ','): NULL
;
484 int major
= atoi(VERSION
);
485 int minor
= minor_string
? atoi(minor_string
+ 1) : 0;
486 int patch
= patch_string
? atoi(patch_string
+ 1) : 0;
487 if (strstr(VERSION
, "devel")) {
498 uint32_t version
= VK_MAKE_VERSION(major
, minor
, patch
);
502 void radv_GetPhysicalDeviceProperties(
503 VkPhysicalDevice physicalDevice
,
504 VkPhysicalDeviceProperties
* pProperties
)
506 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
507 VkSampleCountFlags sample_counts
= 0xf;
508 VkPhysicalDeviceLimits limits
= {
509 .maxImageDimension1D
= (1 << 14),
510 .maxImageDimension2D
= (1 << 14),
511 .maxImageDimension3D
= (1 << 11),
512 .maxImageDimensionCube
= (1 << 14),
513 .maxImageArrayLayers
= (1 << 11),
514 .maxTexelBufferElements
= 128 * 1024 * 1024,
515 .maxUniformBufferRange
= UINT32_MAX
,
516 .maxStorageBufferRange
= UINT32_MAX
,
517 .maxPushConstantsSize
= MAX_PUSH_CONSTANTS_SIZE
,
518 .maxMemoryAllocationCount
= UINT32_MAX
,
519 .maxSamplerAllocationCount
= 64 * 1024,
520 .bufferImageGranularity
= 64, /* A cache line */
521 .sparseAddressSpaceSize
= 0,
522 .maxBoundDescriptorSets
= MAX_SETS
,
523 .maxPerStageDescriptorSamplers
= 64,
524 .maxPerStageDescriptorUniformBuffers
= 64,
525 .maxPerStageDescriptorStorageBuffers
= 64,
526 .maxPerStageDescriptorSampledImages
= 64,
527 .maxPerStageDescriptorStorageImages
= 64,
528 .maxPerStageDescriptorInputAttachments
= 64,
529 .maxPerStageResources
= 128,
530 .maxDescriptorSetSamplers
= 256,
531 .maxDescriptorSetUniformBuffers
= 256,
532 .maxDescriptorSetUniformBuffersDynamic
= 256,
533 .maxDescriptorSetStorageBuffers
= 256,
534 .maxDescriptorSetStorageBuffersDynamic
= 256,
535 .maxDescriptorSetSampledImages
= 256,
536 .maxDescriptorSetStorageImages
= 256,
537 .maxDescriptorSetInputAttachments
= 256,
538 .maxVertexInputAttributes
= 32,
539 .maxVertexInputBindings
= 32,
540 .maxVertexInputAttributeOffset
= 2047,
541 .maxVertexInputBindingStride
= 2048,
542 .maxVertexOutputComponents
= 128,
543 .maxTessellationGenerationLevel
= 0,
544 .maxTessellationPatchSize
= 0,
545 .maxTessellationControlPerVertexInputComponents
= 0,
546 .maxTessellationControlPerVertexOutputComponents
= 0,
547 .maxTessellationControlPerPatchOutputComponents
= 0,
548 .maxTessellationControlTotalOutputComponents
= 0,
549 .maxTessellationEvaluationInputComponents
= 0,
550 .maxTessellationEvaluationOutputComponents
= 0,
551 .maxGeometryShaderInvocations
= 32,
552 .maxGeometryInputComponents
= 64,
553 .maxGeometryOutputComponents
= 128,
554 .maxGeometryOutputVertices
= 256,
555 .maxGeometryTotalOutputComponents
= 1024,
556 .maxFragmentInputComponents
= 128,
557 .maxFragmentOutputAttachments
= 8,
558 .maxFragmentDualSrcAttachments
= 1,
559 .maxFragmentCombinedOutputResources
= 8,
560 .maxComputeSharedMemorySize
= 32768,
561 .maxComputeWorkGroupCount
= { 65535, 65535, 65535 },
562 .maxComputeWorkGroupInvocations
= 2048,
563 .maxComputeWorkGroupSize
= {
568 .subPixelPrecisionBits
= 4 /* FIXME */,
569 .subTexelPrecisionBits
= 4 /* FIXME */,
570 .mipmapPrecisionBits
= 4 /* FIXME */,
571 .maxDrawIndexedIndexValue
= UINT32_MAX
,
572 .maxDrawIndirectCount
= UINT32_MAX
,
573 .maxSamplerLodBias
= 16,
574 .maxSamplerAnisotropy
= 16,
575 .maxViewports
= MAX_VIEWPORTS
,
576 .maxViewportDimensions
= { (1 << 14), (1 << 14) },
577 .viewportBoundsRange
= { INT16_MIN
, INT16_MAX
},
578 .viewportSubPixelBits
= 13, /* We take a float? */
579 .minMemoryMapAlignment
= 4096, /* A page */
580 .minTexelBufferOffsetAlignment
= 1,
581 .minUniformBufferOffsetAlignment
= 4,
582 .minStorageBufferOffsetAlignment
= 4,
583 .minTexelOffset
= -32,
584 .maxTexelOffset
= 31,
585 .minTexelGatherOffset
= -32,
586 .maxTexelGatherOffset
= 31,
587 .minInterpolationOffset
= -2,
588 .maxInterpolationOffset
= 2,
589 .subPixelInterpolationOffsetBits
= 8,
590 .maxFramebufferWidth
= (1 << 14),
591 .maxFramebufferHeight
= (1 << 14),
592 .maxFramebufferLayers
= (1 << 10),
593 .framebufferColorSampleCounts
= sample_counts
,
594 .framebufferDepthSampleCounts
= sample_counts
,
595 .framebufferStencilSampleCounts
= sample_counts
,
596 .framebufferNoAttachmentsSampleCounts
= sample_counts
,
597 .maxColorAttachments
= MAX_RTS
,
598 .sampledImageColorSampleCounts
= sample_counts
,
599 .sampledImageIntegerSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
600 .sampledImageDepthSampleCounts
= sample_counts
,
601 .sampledImageStencilSampleCounts
= sample_counts
,
602 .storageImageSampleCounts
= VK_SAMPLE_COUNT_1_BIT
,
603 .maxSampleMaskWords
= 1,
604 .timestampComputeAndGraphics
= false,
605 .timestampPeriod
= 100000.0 / pdevice
->rad_info
.clock_crystal_freq
,
606 .maxClipDistances
= 8,
607 .maxCullDistances
= 8,
608 .maxCombinedClipAndCullDistances
= 8,
609 .discreteQueuePriorities
= 1,
610 .pointSizeRange
= { 0.125, 255.875 },
611 .lineWidthRange
= { 0.0, 7.9921875 },
612 .pointSizeGranularity
= (1.0 / 8.0),
613 .lineWidthGranularity
= (1.0 / 128.0),
614 .strictLines
= false, /* FINISHME */
615 .standardSampleLocations
= true,
616 .optimalBufferCopyOffsetAlignment
= 128,
617 .optimalBufferCopyRowPitchAlignment
= 128,
618 .nonCoherentAtomSize
= 64,
621 *pProperties
= (VkPhysicalDeviceProperties
) {
622 .apiVersion
= VK_MAKE_VERSION(1, 0, 42),
623 .driverVersion
= radv_get_driver_version(),
625 .deviceID
= pdevice
->rad_info
.pci_id
,
626 .deviceType
= VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
,
628 .sparseProperties
= {0}, /* Broadwell doesn't do sparse. */
631 strcpy(pProperties
->deviceName
, pdevice
->name
);
632 memcpy(pProperties
->pipelineCacheUUID
, pdevice
->uuid
, VK_UUID_SIZE
);
635 void radv_GetPhysicalDeviceProperties2KHR(
636 VkPhysicalDevice physicalDevice
,
637 VkPhysicalDeviceProperties2KHR
*pProperties
)
639 return radv_GetPhysicalDeviceProperties(physicalDevice
, &pProperties
->properties
);
642 static void radv_get_physical_device_queue_family_properties(
643 struct radv_physical_device
* pdevice
,
645 VkQueueFamilyProperties
** pQueueFamilyProperties
)
647 int num_queue_families
= 1;
649 if (pdevice
->rad_info
.compute_rings
> 0 &&
650 pdevice
->rad_info
.chip_class
>= CIK
&&
651 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
))
652 num_queue_families
++;
654 if (pQueueFamilyProperties
== NULL
) {
655 *pCount
= num_queue_families
;
664 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
665 .queueFlags
= VK_QUEUE_GRAPHICS_BIT
|
666 VK_QUEUE_COMPUTE_BIT
|
667 VK_QUEUE_TRANSFER_BIT
,
669 .timestampValidBits
= 64,
670 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
675 if (pdevice
->rad_info
.compute_rings
> 0 &&
676 pdevice
->rad_info
.chip_class
>= CIK
&&
677 !(pdevice
->instance
->debug_flags
& RADV_DEBUG_NO_COMPUTE_QUEUE
)) {
679 *pQueueFamilyProperties
[idx
] = (VkQueueFamilyProperties
) {
680 .queueFlags
= VK_QUEUE_COMPUTE_BIT
| VK_QUEUE_TRANSFER_BIT
,
681 .queueCount
= pdevice
->rad_info
.compute_rings
,
682 .timestampValidBits
= 64,
683 .minImageTransferGranularity
= (VkExtent3D
) { 1, 1, 1 },
691 void radv_GetPhysicalDeviceQueueFamilyProperties(
692 VkPhysicalDevice physicalDevice
,
694 VkQueueFamilyProperties
* pQueueFamilyProperties
)
696 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
697 if (!pQueueFamilyProperties
) {
698 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
701 VkQueueFamilyProperties
*properties
[] = {
702 pQueueFamilyProperties
+ 0,
703 pQueueFamilyProperties
+ 1,
704 pQueueFamilyProperties
+ 2,
706 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
707 assert(*pCount
<= 3);
710 void radv_GetPhysicalDeviceQueueFamilyProperties2KHR(
711 VkPhysicalDevice physicalDevice
,
713 VkQueueFamilyProperties2KHR
*pQueueFamilyProperties
)
715 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
716 if (!pQueueFamilyProperties
) {
717 return radv_get_physical_device_queue_family_properties(pdevice
, pCount
, NULL
);
720 VkQueueFamilyProperties
*properties
[] = {
721 &pQueueFamilyProperties
[0].queueFamilyProperties
,
722 &pQueueFamilyProperties
[1].queueFamilyProperties
,
723 &pQueueFamilyProperties
[2].queueFamilyProperties
,
725 radv_get_physical_device_queue_family_properties(pdevice
, pCount
, properties
);
726 assert(*pCount
<= 3);
729 void radv_GetPhysicalDeviceMemoryProperties(
730 VkPhysicalDevice physicalDevice
,
731 VkPhysicalDeviceMemoryProperties
*pMemoryProperties
)
733 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
735 STATIC_ASSERT(RADV_MEM_TYPE_COUNT
<= VK_MAX_MEMORY_TYPES
);
737 pMemoryProperties
->memoryTypeCount
= RADV_MEM_TYPE_COUNT
;
738 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_VRAM
] = (VkMemoryType
) {
739 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
,
740 .heapIndex
= RADV_MEM_HEAP_VRAM
,
742 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_GTT_WRITE_COMBINE
] = (VkMemoryType
) {
743 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
744 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
745 .heapIndex
= RADV_MEM_HEAP_GTT
,
747 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_VRAM_CPU_ACCESS
] = (VkMemoryType
) {
748 .propertyFlags
= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT
|
749 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
750 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
,
751 .heapIndex
= RADV_MEM_HEAP_VRAM_CPU_ACCESS
,
753 pMemoryProperties
->memoryTypes
[RADV_MEM_TYPE_GTT_CACHED
] = (VkMemoryType
) {
754 .propertyFlags
= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT
|
755 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT
|
756 VK_MEMORY_PROPERTY_HOST_CACHED_BIT
,
757 .heapIndex
= RADV_MEM_HEAP_GTT
,
760 STATIC_ASSERT(RADV_MEM_HEAP_COUNT
<= VK_MAX_MEMORY_HEAPS
);
762 pMemoryProperties
->memoryHeapCount
= RADV_MEM_HEAP_COUNT
;
763 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_VRAM
] = (VkMemoryHeap
) {
764 .size
= physical_device
->rad_info
.vram_size
-
765 physical_device
->rad_info
.visible_vram_size
,
766 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
768 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_VRAM_CPU_ACCESS
] = (VkMemoryHeap
) {
769 .size
= physical_device
->rad_info
.visible_vram_size
,
770 .flags
= VK_MEMORY_HEAP_DEVICE_LOCAL_BIT
,
772 pMemoryProperties
->memoryHeaps
[RADV_MEM_HEAP_GTT
] = (VkMemoryHeap
) {
773 .size
= physical_device
->rad_info
.gart_size
,
778 void radv_GetPhysicalDeviceMemoryProperties2KHR(
779 VkPhysicalDevice physicalDevice
,
780 VkPhysicalDeviceMemoryProperties2KHR
*pMemoryProperties
)
782 return radv_GetPhysicalDeviceMemoryProperties(physicalDevice
,
783 &pMemoryProperties
->memoryProperties
);
787 radv_queue_init(struct radv_device
*device
, struct radv_queue
*queue
,
788 int queue_family_index
, int idx
)
790 queue
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
791 queue
->device
= device
;
792 queue
->queue_family_index
= queue_family_index
;
793 queue
->queue_idx
= idx
;
795 queue
->hw_ctx
= device
->ws
->ctx_create(device
->ws
);
797 return VK_ERROR_OUT_OF_HOST_MEMORY
;
803 radv_queue_finish(struct radv_queue
*queue
)
806 queue
->device
->ws
->ctx_destroy(queue
->hw_ctx
);
808 if (queue
->initial_preamble_cs
)
809 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
810 if (queue
->continue_preamble_cs
)
811 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
812 if (queue
->descriptor_bo
)
813 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
814 if (queue
->scratch_bo
)
815 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
816 if (queue
->esgs_ring_bo
)
817 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
818 if (queue
->gsvs_ring_bo
)
819 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
820 if (queue
->compute_scratch_bo
)
821 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
825 radv_device_init_gs_info(struct radv_device
*device
)
827 switch (device
->physical_device
->rad_info
.family
) {
836 device
->gs_table_depth
= 16;
847 device
->gs_table_depth
= 32;
850 unreachable("unknown GPU");
854 VkResult
radv_CreateDevice(
855 VkPhysicalDevice physicalDevice
,
856 const VkDeviceCreateInfo
* pCreateInfo
,
857 const VkAllocationCallbacks
* pAllocator
,
860 RADV_FROM_HANDLE(radv_physical_device
, physical_device
, physicalDevice
);
862 struct radv_device
*device
;
864 for (uint32_t i
= 0; i
< pCreateInfo
->enabledExtensionCount
; i
++) {
865 if (!is_extension_enabled(physical_device
->extensions
.ext_array
,
866 physical_device
->extensions
.num_ext
,
867 pCreateInfo
->ppEnabledExtensionNames
[i
]))
868 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT
);
871 device
= vk_alloc2(&physical_device
->instance
->alloc
, pAllocator
,
873 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
875 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
877 memset(device
, 0, sizeof(*device
));
879 device
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
880 device
->instance
= physical_device
->instance
;
881 device
->physical_device
= physical_device
;
883 device
->debug_flags
= device
->instance
->debug_flags
;
885 device
->ws
= physical_device
->ws
;
887 device
->alloc
= *pAllocator
;
889 device
->alloc
= physical_device
->instance
->alloc
;
891 for (unsigned i
= 0; i
< pCreateInfo
->queueCreateInfoCount
; i
++) {
892 const VkDeviceQueueCreateInfo
*queue_create
= &pCreateInfo
->pQueueCreateInfos
[i
];
893 uint32_t qfi
= queue_create
->queueFamilyIndex
;
895 device
->queues
[qfi
] = vk_alloc(&device
->alloc
,
896 queue_create
->queueCount
* sizeof(struct radv_queue
), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE
);
897 if (!device
->queues
[qfi
]) {
898 result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
902 memset(device
->queues
[qfi
], 0, queue_create
->queueCount
* sizeof(struct radv_queue
));
904 device
->queue_count
[qfi
] = queue_create
->queueCount
;
906 for (unsigned q
= 0; q
< queue_create
->queueCount
; q
++) {
907 result
= radv_queue_init(device
, &device
->queues
[qfi
][q
], qfi
, q
);
908 if (result
!= VK_SUCCESS
)
913 #if HAVE_LLVM < 0x0400
914 device
->llvm_supports_spill
= false;
916 device
->llvm_supports_spill
= true;
919 /* The maximum number of scratch waves. Scratch space isn't divided
920 * evenly between CUs. The number is only a function of the number of CUs.
921 * We can decrease the constant to decrease the scratch buffer size.
923 * sctx->scratch_waves must be >= the maximum posible size of
924 * 1 threadgroup, so that the hw doesn't hang from being unable
927 * The recommended value is 4 per CU at most. Higher numbers don't
928 * bring much benefit, but they still occupy chip resources (think
929 * async compute). I've seen ~2% performance difference between 4 and 32.
931 uint32_t max_threads_per_block
= 2048;
932 device
->scratch_waves
= MAX2(32 * physical_device
->rad_info
.num_good_compute_units
,
933 max_threads_per_block
/ 64);
935 radv_device_init_gs_info(device
);
937 result
= radv_device_init_meta(device
);
938 if (result
!= VK_SUCCESS
)
941 radv_device_init_msaa(device
);
943 for (int family
= 0; family
< RADV_MAX_QUEUE_FAMILIES
; ++family
) {
944 device
->empty_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
946 case RADV_QUEUE_GENERAL
:
947 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
948 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_LOAD_ENABLE(1));
949 radeon_emit(device
->empty_cs
[family
], CONTEXT_CONTROL_SHADOW_ENABLE(1));
951 case RADV_QUEUE_COMPUTE
:
952 radeon_emit(device
->empty_cs
[family
], PKT3(PKT3_NOP
, 0, 0));
953 radeon_emit(device
->empty_cs
[family
], 0);
956 device
->ws
->cs_finalize(device
->empty_cs
[family
]);
958 device
->flush_cs
[family
] = device
->ws
->cs_create(device
->ws
, family
);
960 case RADV_QUEUE_GENERAL
:
961 case RADV_QUEUE_COMPUTE
:
962 si_cs_emit_cache_flush(device
->flush_cs
[family
],
963 device
->physical_device
->rad_info
.chip_class
,
964 family
== RADV_QUEUE_COMPUTE
&& device
->physical_device
->rad_info
.chip_class
>= CIK
,
965 RADV_CMD_FLAG_INV_ICACHE
|
966 RADV_CMD_FLAG_INV_SMEM_L1
|
967 RADV_CMD_FLAG_INV_VMEM_L1
|
968 RADV_CMD_FLAG_INV_GLOBAL_L2
);
971 device
->ws
->cs_finalize(device
->flush_cs
[family
]);
974 if (getenv("RADV_TRACE_FILE")) {
975 device
->trace_bo
= device
->ws
->buffer_create(device
->ws
, 4096, 8,
976 RADEON_DOMAIN_VRAM
, RADEON_FLAG_CPU_ACCESS
);
977 if (!device
->trace_bo
)
980 device
->trace_id_ptr
= device
->ws
->buffer_map(device
->trace_bo
);
981 if (!device
->trace_id_ptr
)
985 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
986 cik_create_gfx_config(device
);
988 *pDevice
= radv_device_to_handle(device
);
992 if (device
->trace_bo
)
993 device
->ws
->buffer_destroy(device
->trace_bo
);
995 if (device
->gfx_init
)
996 device
->ws
->buffer_destroy(device
->gfx_init
);
998 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
999 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1000 radv_queue_finish(&device
->queues
[i
][q
]);
1001 if (device
->queue_count
[i
])
1002 vk_free(&device
->alloc
, device
->queues
[i
]);
1005 vk_free(&device
->alloc
, device
);
1009 void radv_DestroyDevice(
1011 const VkAllocationCallbacks
* pAllocator
)
1013 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1018 if (device
->trace_bo
)
1019 device
->ws
->buffer_destroy(device
->trace_bo
);
1021 if (device
->gfx_init
)
1022 device
->ws
->buffer_destroy(device
->gfx_init
);
1024 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1025 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++)
1026 radv_queue_finish(&device
->queues
[i
][q
]);
1027 if (device
->queue_count
[i
])
1028 vk_free(&device
->alloc
, device
->queues
[i
]);
1029 if (device
->empty_cs
[i
])
1030 device
->ws
->cs_destroy(device
->empty_cs
[i
]);
1031 if (device
->flush_cs
[i
])
1032 device
->ws
->cs_destroy(device
->flush_cs
[i
]);
1034 radv_device_finish_meta(device
);
1036 vk_free(&device
->alloc
, device
);
1039 VkResult
radv_EnumerateInstanceExtensionProperties(
1040 const char* pLayerName
,
1041 uint32_t* pPropertyCount
,
1042 VkExtensionProperties
* pProperties
)
1044 if (pProperties
== NULL
) {
1045 *pPropertyCount
= ARRAY_SIZE(instance_extensions
);
1049 *pPropertyCount
= MIN2(*pPropertyCount
, ARRAY_SIZE(instance_extensions
));
1050 typed_memcpy(pProperties
, instance_extensions
, *pPropertyCount
);
1052 if (*pPropertyCount
< ARRAY_SIZE(instance_extensions
))
1053 return VK_INCOMPLETE
;
1058 VkResult
radv_EnumerateDeviceExtensionProperties(
1059 VkPhysicalDevice physicalDevice
,
1060 const char* pLayerName
,
1061 uint32_t* pPropertyCount
,
1062 VkExtensionProperties
* pProperties
)
1064 RADV_FROM_HANDLE(radv_physical_device
, pdevice
, physicalDevice
);
1066 if (pProperties
== NULL
) {
1067 *pPropertyCount
= pdevice
->extensions
.num_ext
;
1071 *pPropertyCount
= MIN2(*pPropertyCount
, pdevice
->extensions
.num_ext
);
1072 typed_memcpy(pProperties
, pdevice
->extensions
.ext_array
, *pPropertyCount
);
1074 if (*pPropertyCount
< pdevice
->extensions
.num_ext
)
1075 return VK_INCOMPLETE
;
1080 VkResult
radv_EnumerateInstanceLayerProperties(
1081 uint32_t* pPropertyCount
,
1082 VkLayerProperties
* pProperties
)
1084 if (pProperties
== NULL
) {
1085 *pPropertyCount
= 0;
1089 /* None supported at this time */
1090 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1093 VkResult
radv_EnumerateDeviceLayerProperties(
1094 VkPhysicalDevice physicalDevice
,
1095 uint32_t* pPropertyCount
,
1096 VkLayerProperties
* pProperties
)
1098 if (pProperties
== NULL
) {
1099 *pPropertyCount
= 0;
1103 /* None supported at this time */
1104 return vk_error(VK_ERROR_LAYER_NOT_PRESENT
);
1107 void radv_GetDeviceQueue(
1109 uint32_t queueFamilyIndex
,
1110 uint32_t queueIndex
,
1113 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1115 *pQueue
= radv_queue_to_handle(&device
->queues
[queueFamilyIndex
][queueIndex
]);
1118 static void radv_dump_trace(struct radv_device
*device
,
1119 struct radeon_winsys_cs
*cs
)
1121 const char *filename
= getenv("RADV_TRACE_FILE");
1122 FILE *f
= fopen(filename
, "w");
1124 fprintf(stderr
, "Failed to write trace dump to %s\n", filename
);
1128 fprintf(f
, "Trace ID: %x\n", *device
->trace_id_ptr
);
1129 device
->ws
->cs_dump(cs
, f
, *device
->trace_id_ptr
);
1134 fill_geom_rings(struct radv_queue
*queue
,
1136 uint32_t esgs_ring_size
,
1137 struct radeon_winsys_bo
*esgs_ring_bo
,
1138 uint32_t gsvs_ring_size
,
1139 struct radeon_winsys_bo
*gsvs_ring_bo
)
1141 uint64_t esgs_va
= 0, gsvs_va
= 0;
1142 uint32_t *desc
= &map
[4];
1145 esgs_va
= queue
->device
->ws
->buffer_get_va(esgs_ring_bo
);
1147 gsvs_va
= queue
->device
->ws
->buffer_get_va(gsvs_ring_bo
);
1149 /* stride 0, num records - size, add tid, swizzle, elsize4,
1152 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32) |
1153 S_008F04_STRIDE(0) |
1154 S_008F04_SWIZZLE_ENABLE(true);
1155 desc
[2] = esgs_ring_size
;
1156 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1157 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1158 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1159 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1160 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1161 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1162 S_008F0C_ELEMENT_SIZE(1) |
1163 S_008F0C_INDEX_STRIDE(3) |
1164 S_008F0C_ADD_TID_ENABLE(true);
1167 /* GS entry for ES->GS ring */
1168 /* stride 0, num records - size, elsize0,
1171 desc
[1] = S_008F04_BASE_ADDRESS_HI(esgs_va
>> 32)|
1172 S_008F04_STRIDE(0) |
1173 S_008F04_SWIZZLE_ENABLE(false);
1174 desc
[2] = esgs_ring_size
;
1175 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1176 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1177 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1178 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1179 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1180 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1181 S_008F0C_ELEMENT_SIZE(0) |
1182 S_008F0C_INDEX_STRIDE(0) |
1183 S_008F0C_ADD_TID_ENABLE(false);
1186 /* VS entry for GS->VS ring */
1187 /* stride 0, num records - size, elsize0,
1190 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1191 S_008F04_STRIDE(0) |
1192 S_008F04_SWIZZLE_ENABLE(false);
1193 desc
[2] = gsvs_ring_size
;
1194 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1195 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1196 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1197 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1198 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1199 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1200 S_008F0C_ELEMENT_SIZE(0) |
1201 S_008F0C_INDEX_STRIDE(0) |
1202 S_008F0C_ADD_TID_ENABLE(false);
1205 /* stride gsvs_itemsize, num records 64
1206 elsize 4, index stride 16 */
1207 /* shader will patch stride and desc[2] */
1209 desc
[1] = S_008F04_BASE_ADDRESS_HI(gsvs_va
>> 32)|
1210 S_008F04_STRIDE(0) |
1211 S_008F04_SWIZZLE_ENABLE(true);
1213 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1214 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1215 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1216 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1217 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1218 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1219 S_008F0C_ELEMENT_SIZE(1) |
1220 S_008F0C_INDEX_STRIDE(1) |
1221 S_008F0C_ADD_TID_ENABLE(true);
1225 radv_get_preamble_cs(struct radv_queue
*queue
,
1226 uint32_t scratch_size
,
1227 uint32_t compute_scratch_size
,
1228 uint32_t esgs_ring_size
,
1229 uint32_t gsvs_ring_size
,
1230 struct radeon_winsys_cs
**initial_preamble_cs
,
1231 struct radeon_winsys_cs
**continue_preamble_cs
)
1233 struct radeon_winsys_bo
*scratch_bo
= NULL
;
1234 struct radeon_winsys_bo
*descriptor_bo
= NULL
;
1235 struct radeon_winsys_bo
*compute_scratch_bo
= NULL
;
1236 struct radeon_winsys_bo
*esgs_ring_bo
= NULL
;
1237 struct radeon_winsys_bo
*gsvs_ring_bo
= NULL
;
1238 struct radeon_winsys_cs
*dest_cs
[2] = {0};
1240 if (scratch_size
<= queue
->scratch_size
&&
1241 compute_scratch_size
<= queue
->compute_scratch_size
&&
1242 esgs_ring_size
<= queue
->esgs_ring_size
&&
1243 gsvs_ring_size
<= queue
->gsvs_ring_size
&&
1244 queue
->initial_preamble_cs
) {
1245 *initial_preamble_cs
= queue
->initial_preamble_cs
;
1246 *continue_preamble_cs
= queue
->continue_preamble_cs
;
1247 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
1248 *continue_preamble_cs
= NULL
;
1252 if (scratch_size
> queue
->scratch_size
) {
1253 scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1257 RADEON_FLAG_NO_CPU_ACCESS
);
1261 scratch_bo
= queue
->scratch_bo
;
1263 if (compute_scratch_size
> queue
->compute_scratch_size
) {
1264 compute_scratch_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1265 compute_scratch_size
,
1268 RADEON_FLAG_NO_CPU_ACCESS
);
1269 if (!compute_scratch_bo
)
1273 compute_scratch_bo
= queue
->compute_scratch_bo
;
1275 if (esgs_ring_size
> queue
->esgs_ring_size
) {
1276 esgs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1280 RADEON_FLAG_NO_CPU_ACCESS
);
1284 esgs_ring_bo
= queue
->esgs_ring_bo
;
1285 esgs_ring_size
= queue
->esgs_ring_size
;
1288 if (gsvs_ring_size
> queue
->gsvs_ring_size
) {
1289 gsvs_ring_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1293 RADEON_FLAG_NO_CPU_ACCESS
);
1297 gsvs_ring_bo
= queue
->gsvs_ring_bo
;
1298 gsvs_ring_size
= queue
->gsvs_ring_size
;
1301 if (scratch_bo
!= queue
->scratch_bo
||
1302 esgs_ring_bo
!= queue
->esgs_ring_bo
||
1303 gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
1305 if (gsvs_ring_bo
|| esgs_ring_bo
)
1306 size
= 80; /* 2 dword + 2 padding + 4 dword * 4 */
1307 else if (scratch_bo
)
1308 size
= 8; /* 2 dword */
1310 descriptor_bo
= queue
->device
->ws
->buffer_create(queue
->device
->ws
,
1314 RADEON_FLAG_CPU_ACCESS
);
1318 descriptor_bo
= queue
->descriptor_bo
;
1320 for(int i
= 0; i
< 2; ++i
) {
1321 struct radeon_winsys_cs
*cs
= NULL
;
1322 cs
= queue
->device
->ws
->cs_create(queue
->device
->ws
,
1323 queue
->queue_family_index
? RING_COMPUTE
: RING_GFX
);
1330 queue
->device
->ws
->cs_add_buffer(cs
, scratch_bo
, 8);
1333 queue
->device
->ws
->cs_add_buffer(cs
, esgs_ring_bo
, 8);
1336 queue
->device
->ws
->cs_add_buffer(cs
, gsvs_ring_bo
, 8);
1339 queue
->device
->ws
->cs_add_buffer(cs
, descriptor_bo
, 8);
1341 if (descriptor_bo
!= queue
->descriptor_bo
) {
1342 uint32_t *map
= (uint32_t*)queue
->device
->ws
->buffer_map(descriptor_bo
);
1345 uint64_t scratch_va
= queue
->device
->ws
->buffer_get_va(scratch_bo
);
1346 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
1347 S_008F04_SWIZZLE_ENABLE(1);
1348 map
[0] = scratch_va
;
1352 if (esgs_ring_bo
|| gsvs_ring_bo
)
1353 fill_geom_rings(queue
, map
, esgs_ring_size
, esgs_ring_bo
, gsvs_ring_size
, gsvs_ring_bo
);
1355 queue
->device
->ws
->buffer_unmap(descriptor_bo
);
1358 if (esgs_ring_bo
|| gsvs_ring_bo
) {
1359 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1360 radeon_emit(cs
, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH
) | EVENT_INDEX(4));
1361 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1362 radeon_emit(cs
, EVENT_TYPE(V_028A90_VGT_FLUSH
) | EVENT_INDEX(0));
1364 if (queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1365 radeon_set_uconfig_reg_seq(cs
, R_030900_VGT_ESGS_RING_SIZE
, 2);
1366 radeon_emit(cs
, esgs_ring_size
>> 8);
1367 radeon_emit(cs
, gsvs_ring_size
>> 8);
1369 radeon_set_config_reg_seq(cs
, R_0088C8_VGT_ESGS_RING_SIZE
, 2);
1370 radeon_emit(cs
, esgs_ring_size
>> 8);
1371 radeon_emit(cs
, gsvs_ring_size
>> 8);
1375 if (descriptor_bo
) {
1376 uint32_t regs
[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0
,
1377 R_00B130_SPI_SHADER_USER_DATA_VS_0
,
1378 R_00B230_SPI_SHADER_USER_DATA_GS_0
,
1379 R_00B330_SPI_SHADER_USER_DATA_ES_0
,
1380 R_00B430_SPI_SHADER_USER_DATA_HS_0
,
1381 R_00B530_SPI_SHADER_USER_DATA_LS_0
};
1383 uint64_t va
= queue
->device
->ws
->buffer_get_va(descriptor_bo
);
1385 for (int i
= 0; i
< ARRAY_SIZE(regs
); ++i
) {
1386 radeon_set_sh_reg_seq(cs
, regs
[i
], 2);
1387 radeon_emit(cs
, va
);
1388 radeon_emit(cs
, va
>> 32);
1392 if (compute_scratch_bo
) {
1393 uint64_t scratch_va
= queue
->device
->ws
->buffer_get_va(compute_scratch_bo
);
1394 uint32_t rsrc1
= S_008F04_BASE_ADDRESS_HI(scratch_va
>> 32) |
1395 S_008F04_SWIZZLE_ENABLE(1);
1397 queue
->device
->ws
->cs_add_buffer(cs
, compute_scratch_bo
, 8);
1399 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
, 2);
1400 radeon_emit(cs
, scratch_va
);
1401 radeon_emit(cs
, rsrc1
);
1405 si_cs_emit_cache_flush(cs
,
1406 queue
->device
->physical_device
->rad_info
.chip_class
,
1407 queue
->queue_family_index
== RING_COMPUTE
&&
1408 queue
->device
->physical_device
->rad_info
.chip_class
>= CIK
,
1409 RADV_CMD_FLAG_INV_ICACHE
|
1410 RADV_CMD_FLAG_INV_SMEM_L1
|
1411 RADV_CMD_FLAG_INV_VMEM_L1
|
1412 RADV_CMD_FLAG_INV_GLOBAL_L2
);
1415 if (!queue
->device
->ws
->cs_finalize(cs
))
1419 if (queue
->initial_preamble_cs
)
1420 queue
->device
->ws
->cs_destroy(queue
->initial_preamble_cs
);
1422 if (queue
->continue_preamble_cs
)
1423 queue
->device
->ws
->cs_destroy(queue
->continue_preamble_cs
);
1425 queue
->initial_preamble_cs
= dest_cs
[0];
1426 queue
->continue_preamble_cs
= dest_cs
[1];
1428 if (scratch_bo
!= queue
->scratch_bo
) {
1429 if (queue
->scratch_bo
)
1430 queue
->device
->ws
->buffer_destroy(queue
->scratch_bo
);
1431 queue
->scratch_bo
= scratch_bo
;
1432 queue
->scratch_size
= scratch_size
;
1435 if (compute_scratch_bo
!= queue
->compute_scratch_bo
) {
1436 if (queue
->compute_scratch_bo
)
1437 queue
->device
->ws
->buffer_destroy(queue
->compute_scratch_bo
);
1438 queue
->compute_scratch_bo
= compute_scratch_bo
;
1439 queue
->compute_scratch_size
= compute_scratch_size
;
1442 if (esgs_ring_bo
!= queue
->esgs_ring_bo
) {
1443 if (queue
->esgs_ring_bo
)
1444 queue
->device
->ws
->buffer_destroy(queue
->esgs_ring_bo
);
1445 queue
->esgs_ring_bo
= esgs_ring_bo
;
1446 queue
->esgs_ring_size
= esgs_ring_size
;
1449 if (gsvs_ring_bo
!= queue
->gsvs_ring_bo
) {
1450 if (queue
->gsvs_ring_bo
)
1451 queue
->device
->ws
->buffer_destroy(queue
->gsvs_ring_bo
);
1452 queue
->gsvs_ring_bo
= gsvs_ring_bo
;
1453 queue
->gsvs_ring_size
= gsvs_ring_size
;
1456 if (descriptor_bo
!= queue
->descriptor_bo
) {
1457 if (queue
->descriptor_bo
)
1458 queue
->device
->ws
->buffer_destroy(queue
->descriptor_bo
);
1460 queue
->descriptor_bo
= descriptor_bo
;
1463 *initial_preamble_cs
= queue
->initial_preamble_cs
;
1464 *continue_preamble_cs
= queue
->continue_preamble_cs
;
1465 if (!scratch_size
&& !compute_scratch_size
&& !esgs_ring_size
&& !gsvs_ring_size
)
1466 *continue_preamble_cs
= NULL
;
1469 for (int i
= 0; i
< ARRAY_SIZE(dest_cs
); ++i
)
1471 queue
->device
->ws
->cs_destroy(dest_cs
[i
]);
1472 if (descriptor_bo
&& descriptor_bo
!= queue
->descriptor_bo
)
1473 queue
->device
->ws
->buffer_destroy(descriptor_bo
);
1474 if (scratch_bo
&& scratch_bo
!= queue
->scratch_bo
)
1475 queue
->device
->ws
->buffer_destroy(scratch_bo
);
1476 if (compute_scratch_bo
&& compute_scratch_bo
!= queue
->compute_scratch_bo
)
1477 queue
->device
->ws
->buffer_destroy(compute_scratch_bo
);
1478 if (esgs_ring_bo
&& esgs_ring_bo
!= queue
->esgs_ring_bo
)
1479 queue
->device
->ws
->buffer_destroy(esgs_ring_bo
);
1480 if (gsvs_ring_bo
&& gsvs_ring_bo
!= queue
->gsvs_ring_bo
)
1481 queue
->device
->ws
->buffer_destroy(gsvs_ring_bo
);
1482 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1485 VkResult
radv_QueueSubmit(
1487 uint32_t submitCount
,
1488 const VkSubmitInfo
* pSubmits
,
1491 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
1492 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1493 struct radeon_winsys_fence
*base_fence
= fence
? fence
->fence
: NULL
;
1494 struct radeon_winsys_ctx
*ctx
= queue
->hw_ctx
;
1496 uint32_t max_cs_submission
= queue
->device
->trace_bo
? 1 : UINT32_MAX
;
1497 uint32_t scratch_size
= 0;
1498 uint32_t compute_scratch_size
= 0;
1499 uint32_t esgs_ring_size
= 0, gsvs_ring_size
= 0;
1500 struct radeon_winsys_cs
*initial_preamble_cs
= NULL
, *continue_preamble_cs
= NULL
;
1502 bool fence_emitted
= false;
1504 /* Do this first so failing to allocate scratch buffers can't result in
1505 * partially executed submissions. */
1506 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1507 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1508 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
1509 pSubmits
[i
].pCommandBuffers
[j
]);
1511 scratch_size
= MAX2(scratch_size
, cmd_buffer
->scratch_size_needed
);
1512 compute_scratch_size
= MAX2(compute_scratch_size
,
1513 cmd_buffer
->compute_scratch_size_needed
);
1514 esgs_ring_size
= MAX2(esgs_ring_size
, cmd_buffer
->esgs_ring_size_needed
);
1515 gsvs_ring_size
= MAX2(gsvs_ring_size
, cmd_buffer
->gsvs_ring_size_needed
);
1519 result
= radv_get_preamble_cs(queue
, scratch_size
, compute_scratch_size
,
1520 esgs_ring_size
, gsvs_ring_size
,
1521 &initial_preamble_cs
, &continue_preamble_cs
);
1522 if (result
!= VK_SUCCESS
)
1525 for (uint32_t i
= 0; i
< submitCount
; i
++) {
1526 struct radeon_winsys_cs
**cs_array
;
1527 bool has_flush
= !submitCount
;
1528 bool can_patch
= !has_flush
;
1531 if (!pSubmits
[i
].commandBufferCount
) {
1532 if (pSubmits
[i
].waitSemaphoreCount
|| pSubmits
[i
].signalSemaphoreCount
) {
1533 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
1534 &queue
->device
->empty_cs
[queue
->queue_family_index
],
1536 (struct radeon_winsys_sem
**)pSubmits
[i
].pWaitSemaphores
,
1537 pSubmits
[i
].waitSemaphoreCount
,
1538 (struct radeon_winsys_sem
**)pSubmits
[i
].pSignalSemaphores
,
1539 pSubmits
[i
].signalSemaphoreCount
,
1542 radv_loge("failed to submit CS %d\n", i
);
1545 fence_emitted
= true;
1550 cs_array
= malloc(sizeof(struct radeon_winsys_cs
*) *
1551 (pSubmits
[i
].commandBufferCount
+ has_flush
));
1554 cs_array
[0] = queue
->device
->flush_cs
[queue
->queue_family_index
];
1556 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
1557 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
,
1558 pSubmits
[i
].pCommandBuffers
[j
]);
1559 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1561 cs_array
[j
+ has_flush
] = cmd_buffer
->cs
;
1562 if ((cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT
))
1566 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
+ has_flush
; j
+= advance
) {
1567 advance
= MIN2(max_cs_submission
,
1568 pSubmits
[i
].commandBufferCount
+ has_flush
- j
);
1570 bool e
= j
+ advance
== pSubmits
[i
].commandBufferCount
+ has_flush
;
1572 if (queue
->device
->trace_bo
)
1573 *queue
->device
->trace_id_ptr
= 0;
1575 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
, cs_array
+ j
,
1576 advance
, initial_preamble_cs
, continue_preamble_cs
,
1577 (struct radeon_winsys_sem
**)pSubmits
[i
].pWaitSemaphores
,
1578 b
? pSubmits
[i
].waitSemaphoreCount
: 0,
1579 (struct radeon_winsys_sem
**)pSubmits
[i
].pSignalSemaphores
,
1580 e
? pSubmits
[i
].signalSemaphoreCount
: 0,
1581 can_patch
, base_fence
);
1584 radv_loge("failed to submit CS %d\n", i
);
1587 fence_emitted
= true;
1588 if (queue
->device
->trace_bo
) {
1589 bool success
= queue
->device
->ws
->ctx_wait_idle(
1591 radv_queue_family_to_ring(
1592 queue
->queue_family_index
),
1595 if (!success
) { /* Hang */
1596 radv_dump_trace(queue
->device
, cs_array
[j
]);
1606 ret
= queue
->device
->ws
->cs_submit(ctx
, queue
->queue_idx
,
1607 &queue
->device
->empty_cs
[queue
->queue_family_index
],
1608 1, NULL
, NULL
, NULL
, 0, NULL
, 0,
1611 fence
->submitted
= true;
1617 VkResult
radv_QueueWaitIdle(
1620 RADV_FROM_HANDLE(radv_queue
, queue
, _queue
);
1622 queue
->device
->ws
->ctx_wait_idle(queue
->hw_ctx
,
1623 radv_queue_family_to_ring(queue
->queue_family_index
),
1628 VkResult
radv_DeviceWaitIdle(
1631 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1633 for (unsigned i
= 0; i
< RADV_MAX_QUEUE_FAMILIES
; i
++) {
1634 for (unsigned q
= 0; q
< device
->queue_count
[i
]; q
++) {
1635 radv_QueueWaitIdle(radv_queue_to_handle(&device
->queues
[i
][q
]));
1641 PFN_vkVoidFunction
radv_GetInstanceProcAddr(
1642 VkInstance instance
,
1645 return radv_lookup_entrypoint(pName
);
1648 /* The loader wants us to expose a second GetInstanceProcAddr function
1649 * to work around certain LD_PRELOAD issues seen in apps.
1652 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1653 VkInstance instance
,
1657 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
vk_icdGetInstanceProcAddr(
1658 VkInstance instance
,
1661 return radv_GetInstanceProcAddr(instance
, pName
);
1664 PFN_vkVoidFunction
radv_GetDeviceProcAddr(
1668 return radv_lookup_entrypoint(pName
);
1671 bool radv_get_memory_fd(struct radv_device
*device
,
1672 struct radv_device_memory
*memory
,
1675 struct radeon_bo_metadata metadata
;
1677 if (memory
->image
) {
1678 radv_init_metadata(device
, memory
->image
, &metadata
);
1679 device
->ws
->buffer_set_metadata(memory
->bo
, &metadata
);
1682 return device
->ws
->buffer_get_fd(device
->ws
, memory
->bo
,
1686 VkResult
radv_AllocateMemory(
1688 const VkMemoryAllocateInfo
* pAllocateInfo
,
1689 const VkAllocationCallbacks
* pAllocator
,
1690 VkDeviceMemory
* pMem
)
1692 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1693 struct radv_device_memory
*mem
;
1695 enum radeon_bo_domain domain
;
1697 const VkDedicatedAllocationMemoryAllocateInfoNV
*dedicate_info
= NULL
;
1698 assert(pAllocateInfo
->sType
== VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO
);
1700 if (pAllocateInfo
->allocationSize
== 0) {
1701 /* Apparently, this is allowed */
1702 *pMem
= VK_NULL_HANDLE
;
1706 vk_foreach_struct(ext
, pAllocateInfo
->pNext
) {
1707 switch (ext
->sType
) {
1708 case VK_STRUCTURE_TYPE_DEDICATED_ALLOCATION_MEMORY_ALLOCATE_INFO_NV
:
1709 dedicate_info
= (const VkDedicatedAllocationMemoryAllocateInfoNV
*)ext
;
1716 mem
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*mem
), 8,
1717 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1719 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
1721 if (dedicate_info
) {
1722 mem
->image
= radv_image_from_handle(dedicate_info
->image
);
1723 mem
->buffer
= radv_buffer_from_handle(dedicate_info
->buffer
);
1729 uint64_t alloc_size
= align_u64(pAllocateInfo
->allocationSize
, 4096);
1730 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
||
1731 pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_CACHED
)
1732 domain
= RADEON_DOMAIN_GTT
;
1734 domain
= RADEON_DOMAIN_VRAM
;
1736 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_VRAM
)
1737 flags
|= RADEON_FLAG_NO_CPU_ACCESS
;
1739 flags
|= RADEON_FLAG_CPU_ACCESS
;
1741 if (pAllocateInfo
->memoryTypeIndex
== RADV_MEM_TYPE_GTT_WRITE_COMBINE
)
1742 flags
|= RADEON_FLAG_GTT_WC
;
1744 mem
->bo
= device
->ws
->buffer_create(device
->ws
, alloc_size
, 65536,
1748 result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
1751 mem
->type_index
= pAllocateInfo
->memoryTypeIndex
;
1753 *pMem
= radv_device_memory_to_handle(mem
);
1758 vk_free2(&device
->alloc
, pAllocator
, mem
);
1763 void radv_FreeMemory(
1765 VkDeviceMemory _mem
,
1766 const VkAllocationCallbacks
* pAllocator
)
1768 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1769 RADV_FROM_HANDLE(radv_device_memory
, mem
, _mem
);
1774 device
->ws
->buffer_destroy(mem
->bo
);
1777 vk_free2(&device
->alloc
, pAllocator
, mem
);
1780 VkResult
radv_MapMemory(
1782 VkDeviceMemory _memory
,
1783 VkDeviceSize offset
,
1785 VkMemoryMapFlags flags
,
1788 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1789 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1796 *ppData
= device
->ws
->buffer_map(mem
->bo
);
1802 return VK_ERROR_MEMORY_MAP_FAILED
;
1805 void radv_UnmapMemory(
1807 VkDeviceMemory _memory
)
1809 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1810 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1815 device
->ws
->buffer_unmap(mem
->bo
);
1818 VkResult
radv_FlushMappedMemoryRanges(
1820 uint32_t memoryRangeCount
,
1821 const VkMappedMemoryRange
* pMemoryRanges
)
1826 VkResult
radv_InvalidateMappedMemoryRanges(
1828 uint32_t memoryRangeCount
,
1829 const VkMappedMemoryRange
* pMemoryRanges
)
1834 void radv_GetBufferMemoryRequirements(
1837 VkMemoryRequirements
* pMemoryRequirements
)
1839 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1841 pMemoryRequirements
->memoryTypeBits
= (1u << RADV_MEM_TYPE_COUNT
) - 1;
1843 pMemoryRequirements
->size
= buffer
->size
;
1844 pMemoryRequirements
->alignment
= 16;
1847 void radv_GetImageMemoryRequirements(
1850 VkMemoryRequirements
* pMemoryRequirements
)
1852 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1854 pMemoryRequirements
->memoryTypeBits
= (1u << RADV_MEM_TYPE_COUNT
) - 1;
1856 pMemoryRequirements
->size
= image
->size
;
1857 pMemoryRequirements
->alignment
= image
->alignment
;
1860 void radv_GetImageSparseMemoryRequirements(
1863 uint32_t* pSparseMemoryRequirementCount
,
1864 VkSparseImageMemoryRequirements
* pSparseMemoryRequirements
)
1869 void radv_GetDeviceMemoryCommitment(
1871 VkDeviceMemory memory
,
1872 VkDeviceSize
* pCommittedMemoryInBytes
)
1874 *pCommittedMemoryInBytes
= 0;
1877 VkResult
radv_BindBufferMemory(
1880 VkDeviceMemory _memory
,
1881 VkDeviceSize memoryOffset
)
1883 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1884 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
1887 buffer
->bo
= mem
->bo
;
1888 buffer
->offset
= memoryOffset
;
1897 VkResult
radv_BindImageMemory(
1900 VkDeviceMemory _memory
,
1901 VkDeviceSize memoryOffset
)
1903 RADV_FROM_HANDLE(radv_device_memory
, mem
, _memory
);
1904 RADV_FROM_HANDLE(radv_image
, image
, _image
);
1907 image
->bo
= mem
->bo
;
1908 image
->offset
= memoryOffset
;
1917 VkResult
radv_QueueBindSparse(
1919 uint32_t bindInfoCount
,
1920 const VkBindSparseInfo
* pBindInfo
,
1923 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER
);
1926 VkResult
radv_CreateFence(
1928 const VkFenceCreateInfo
* pCreateInfo
,
1929 const VkAllocationCallbacks
* pAllocator
,
1932 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1933 struct radv_fence
*fence
= vk_alloc2(&device
->alloc
, pAllocator
,
1935 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1938 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1940 memset(fence
, 0, sizeof(*fence
));
1941 fence
->submitted
= false;
1942 fence
->signalled
= !!(pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
);
1943 fence
->fence
= device
->ws
->create_fence();
1944 if (!fence
->fence
) {
1945 vk_free2(&device
->alloc
, pAllocator
, fence
);
1946 return VK_ERROR_OUT_OF_HOST_MEMORY
;
1949 *pFence
= radv_fence_to_handle(fence
);
1954 void radv_DestroyFence(
1957 const VkAllocationCallbacks
* pAllocator
)
1959 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1960 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
1964 device
->ws
->destroy_fence(fence
->fence
);
1965 vk_free2(&device
->alloc
, pAllocator
, fence
);
1968 static uint64_t radv_get_absolute_timeout(uint64_t timeout
)
1970 uint64_t current_time
;
1973 clock_gettime(CLOCK_MONOTONIC
, &tv
);
1974 current_time
= tv
.tv_nsec
+ tv
.tv_sec
*1000000000ull;
1976 timeout
= MIN2(UINT64_MAX
- current_time
, timeout
);
1978 return current_time
+ timeout
;
1981 VkResult
radv_WaitForFences(
1983 uint32_t fenceCount
,
1984 const VkFence
* pFences
,
1988 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1989 timeout
= radv_get_absolute_timeout(timeout
);
1991 if (!waitAll
&& fenceCount
> 1) {
1992 fprintf(stderr
, "radv: WaitForFences without waitAll not implemented yet\n");
1995 for (uint32_t i
= 0; i
< fenceCount
; ++i
) {
1996 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
1997 bool expired
= false;
1999 if (fence
->signalled
)
2002 if (!fence
->submitted
)
2005 expired
= device
->ws
->fence_wait(device
->ws
, fence
->fence
, true, timeout
);
2009 fence
->signalled
= true;
2015 VkResult
radv_ResetFences(VkDevice device
,
2016 uint32_t fenceCount
,
2017 const VkFence
*pFences
)
2019 for (unsigned i
= 0; i
< fenceCount
; ++i
) {
2020 RADV_FROM_HANDLE(radv_fence
, fence
, pFences
[i
]);
2021 fence
->submitted
= fence
->signalled
= false;
2027 VkResult
radv_GetFenceStatus(VkDevice _device
, VkFence _fence
)
2029 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2030 RADV_FROM_HANDLE(radv_fence
, fence
, _fence
);
2032 if (fence
->signalled
)
2034 if (!fence
->submitted
)
2035 return VK_NOT_READY
;
2037 if (!device
->ws
->fence_wait(device
->ws
, fence
->fence
, false, 0))
2038 return VK_NOT_READY
;
2044 // Queue semaphore functions
2046 VkResult
radv_CreateSemaphore(
2048 const VkSemaphoreCreateInfo
* pCreateInfo
,
2049 const VkAllocationCallbacks
* pAllocator
,
2050 VkSemaphore
* pSemaphore
)
2052 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2053 struct radeon_winsys_sem
*sem
;
2055 sem
= device
->ws
->create_sem(device
->ws
);
2057 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2059 *pSemaphore
= radeon_winsys_sem_to_handle(sem
);
2063 void radv_DestroySemaphore(
2065 VkSemaphore _semaphore
,
2066 const VkAllocationCallbacks
* pAllocator
)
2068 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2069 RADV_FROM_HANDLE(radeon_winsys_sem
, sem
, _semaphore
);
2073 device
->ws
->destroy_sem(sem
);
2076 VkResult
radv_CreateEvent(
2078 const VkEventCreateInfo
* pCreateInfo
,
2079 const VkAllocationCallbacks
* pAllocator
,
2082 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2083 struct radv_event
*event
= vk_alloc2(&device
->alloc
, pAllocator
,
2085 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2088 return VK_ERROR_OUT_OF_HOST_MEMORY
;
2090 event
->bo
= device
->ws
->buffer_create(device
->ws
, 8, 8,
2092 RADEON_FLAG_CPU_ACCESS
);
2094 vk_free2(&device
->alloc
, pAllocator
, event
);
2095 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2098 event
->map
= (uint64_t*)device
->ws
->buffer_map(event
->bo
);
2100 *pEvent
= radv_event_to_handle(event
);
2105 void radv_DestroyEvent(
2108 const VkAllocationCallbacks
* pAllocator
)
2110 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2111 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2115 device
->ws
->buffer_destroy(event
->bo
);
2116 vk_free2(&device
->alloc
, pAllocator
, event
);
2119 VkResult
radv_GetEventStatus(
2123 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2125 if (*event
->map
== 1)
2126 return VK_EVENT_SET
;
2127 return VK_EVENT_RESET
;
2130 VkResult
radv_SetEvent(
2134 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2140 VkResult
radv_ResetEvent(
2144 RADV_FROM_HANDLE(radv_event
, event
, _event
);
2150 VkResult
radv_CreateBuffer(
2152 const VkBufferCreateInfo
* pCreateInfo
,
2153 const VkAllocationCallbacks
* pAllocator
,
2156 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2157 struct radv_buffer
*buffer
;
2159 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO
);
2161 buffer
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*buffer
), 8,
2162 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2164 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2166 buffer
->size
= pCreateInfo
->size
;
2167 buffer
->usage
= pCreateInfo
->usage
;
2171 *pBuffer
= radv_buffer_to_handle(buffer
);
2176 void radv_DestroyBuffer(
2179 const VkAllocationCallbacks
* pAllocator
)
2181 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2182 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
2187 vk_free2(&device
->alloc
, pAllocator
, buffer
);
2190 static inline unsigned
2191 si_tile_mode_index(const struct radv_image
*image
, unsigned level
, bool stencil
)
2194 return image
->surface
.stencil_tiling_index
[level
];
2196 return image
->surface
.tiling_index
[level
];
2199 static uint32_t radv_surface_layer_count(struct radv_image_view
*iview
)
2201 return iview
->type
== VK_IMAGE_VIEW_TYPE_3D
? iview
->extent
.depth
: iview
->layer_count
;
2205 radv_initialise_color_surface(struct radv_device
*device
,
2206 struct radv_color_buffer_info
*cb
,
2207 struct radv_image_view
*iview
)
2209 const struct vk_format_description
*desc
;
2210 unsigned ntype
, format
, swap
, endian
;
2211 unsigned blend_clamp
= 0, blend_bypass
= 0;
2212 unsigned pitch_tile_max
, slice_tile_max
, tile_mode_index
;
2214 const struct radeon_surf
*surf
= &iview
->image
->surface
;
2215 const struct radeon_surf_level
*level_info
= &surf
->level
[iview
->base_mip
];
2217 desc
= vk_format_description(iview
->vk_format
);
2219 memset(cb
, 0, sizeof(*cb
));
2221 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2222 va
+= level_info
->offset
;
2223 cb
->cb_color_base
= va
>> 8;
2225 /* CMASK variables */
2226 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2227 va
+= iview
->image
->cmask
.offset
;
2228 cb
->cb_color_cmask
= va
>> 8;
2229 cb
->cb_color_cmask_slice
= iview
->image
->cmask
.slice_tile_max
;
2231 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2232 va
+= iview
->image
->dcc_offset
;
2233 cb
->cb_dcc_base
= va
>> 8;
2235 uint32_t max_slice
= radv_surface_layer_count(iview
);
2236 cb
->cb_color_view
= S_028C6C_SLICE_START(iview
->base_layer
) |
2237 S_028C6C_SLICE_MAX(iview
->base_layer
+ max_slice
- 1);
2239 cb
->micro_tile_mode
= iview
->image
->surface
.micro_tile_mode
;
2240 pitch_tile_max
= level_info
->nblk_x
/ 8 - 1;
2241 slice_tile_max
= (level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1;
2242 tile_mode_index
= si_tile_mode_index(iview
->image
, iview
->base_mip
, false);
2244 cb
->cb_color_pitch
= S_028C64_TILE_MAX(pitch_tile_max
);
2245 cb
->cb_color_slice
= S_028C68_TILE_MAX(slice_tile_max
);
2247 /* Intensity is implemented as Red, so treat it that way. */
2248 cb
->cb_color_attrib
= S_028C74_FORCE_DST_ALPHA_1(desc
->swizzle
[3] == VK_SWIZZLE_1
) |
2249 S_028C74_TILE_MODE_INDEX(tile_mode_index
);
2251 if (iview
->image
->samples
> 1) {
2252 unsigned log_samples
= util_logbase2(iview
->image
->samples
);
2254 cb
->cb_color_attrib
|= S_028C74_NUM_SAMPLES(log_samples
) |
2255 S_028C74_NUM_FRAGMENTS(log_samples
);
2258 if (iview
->image
->fmask
.size
) {
2259 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+ iview
->image
->fmask
.offset
;
2260 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
2261 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(iview
->image
->fmask
.pitch_in_pixels
/ 8 - 1);
2262 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(iview
->image
->fmask
.tile_mode_index
);
2263 cb
->cb_color_fmask
= va
>> 8;
2264 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(iview
->image
->fmask
.slice_tile_max
);
2266 /* This must be set for fast clear to work without FMASK. */
2267 if (device
->physical_device
->rad_info
.chip_class
>= CIK
)
2268 cb
->cb_color_pitch
|= S_028C64_FMASK_TILE_MAX(pitch_tile_max
);
2269 cb
->cb_color_attrib
|= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index
);
2270 cb
->cb_color_fmask
= cb
->cb_color_base
;
2271 cb
->cb_color_fmask_slice
= S_028C88_TILE_MAX(slice_tile_max
);
2274 ntype
= radv_translate_color_numformat(iview
->vk_format
,
2276 vk_format_get_first_non_void_channel(iview
->vk_format
));
2277 format
= radv_translate_colorformat(iview
->vk_format
);
2278 if (format
== V_028C70_COLOR_INVALID
|| ntype
== ~0u)
2279 radv_finishme("Illegal color\n");
2280 swap
= radv_translate_colorswap(iview
->vk_format
, FALSE
);
2281 endian
= radv_colorformat_endian_swap(format
);
2283 /* blend clamp should be set for all NORM/SRGB types */
2284 if (ntype
== V_028C70_NUMBER_UNORM
||
2285 ntype
== V_028C70_NUMBER_SNORM
||
2286 ntype
== V_028C70_NUMBER_SRGB
)
2289 /* set blend bypass according to docs if SINT/UINT or
2290 8/24 COLOR variants */
2291 if (ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
||
2292 format
== V_028C70_COLOR_8_24
|| format
== V_028C70_COLOR_24_8
||
2293 format
== V_028C70_COLOR_X24_8_32_FLOAT
) {
2298 if ((ntype
== V_028C70_NUMBER_UINT
|| ntype
== V_028C70_NUMBER_SINT
) &&
2299 (format
== V_028C70_COLOR_8
||
2300 format
== V_028C70_COLOR_8_8
||
2301 format
== V_028C70_COLOR_8_8_8_8
))
2302 ->color_is_int8
= true;
2304 cb
->cb_color_info
= S_028C70_FORMAT(format
) |
2305 S_028C70_COMP_SWAP(swap
) |
2306 S_028C70_BLEND_CLAMP(blend_clamp
) |
2307 S_028C70_BLEND_BYPASS(blend_bypass
) |
2308 S_028C70_SIMPLE_FLOAT(1) |
2309 S_028C70_ROUND_MODE(ntype
!= V_028C70_NUMBER_UNORM
&&
2310 ntype
!= V_028C70_NUMBER_SNORM
&&
2311 ntype
!= V_028C70_NUMBER_SRGB
&&
2312 format
!= V_028C70_COLOR_8_24
&&
2313 format
!= V_028C70_COLOR_24_8
) |
2314 S_028C70_NUMBER_TYPE(ntype
) |
2315 S_028C70_ENDIAN(endian
);
2316 if (iview
->image
->samples
> 1)
2317 if (iview
->image
->fmask
.size
)
2318 cb
->cb_color_info
|= S_028C70_COMPRESSION(1);
2320 if (iview
->image
->cmask
.size
&&
2321 !(device
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
))
2322 cb
->cb_color_info
|= S_028C70_FAST_CLEAR(1);
2324 if (iview
->image
->surface
.dcc_size
&& level_info
->dcc_enabled
)
2325 cb
->cb_color_info
|= S_028C70_DCC_ENABLE(1);
2327 if (device
->physical_device
->rad_info
.chip_class
>= VI
) {
2328 unsigned max_uncompressed_block_size
= 2;
2329 if (iview
->image
->samples
> 1) {
2330 if (iview
->image
->surface
.bpe
== 1)
2331 max_uncompressed_block_size
= 0;
2332 else if (iview
->image
->surface
.bpe
== 2)
2333 max_uncompressed_block_size
= 1;
2336 cb
->cb_dcc_control
= S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size
) |
2337 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2340 /* This must be set for fast clear to work without FMASK. */
2341 if (!iview
->image
->fmask
.size
&&
2342 device
->physical_device
->rad_info
.chip_class
== SI
) {
2343 unsigned bankh
= util_logbase2(iview
->image
->surface
.bankh
);
2344 cb
->cb_color_attrib
|= S_028C74_FMASK_BANK_HEIGHT(bankh
);
2349 radv_initialise_ds_surface(struct radv_device
*device
,
2350 struct radv_ds_buffer_info
*ds
,
2351 struct radv_image_view
*iview
)
2353 unsigned level
= iview
->base_mip
;
2355 uint64_t va
, s_offs
, z_offs
;
2356 const struct radeon_surf_level
*level_info
= &iview
->image
->surface
.level
[level
];
2357 memset(ds
, 0, sizeof(*ds
));
2358 switch (iview
->vk_format
) {
2359 case VK_FORMAT_D24_UNORM_S8_UINT
:
2360 case VK_FORMAT_X8_D24_UNORM_PACK32
:
2361 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
2362 ds
->offset_scale
= 2.0f
;
2364 case VK_FORMAT_D16_UNORM
:
2365 case VK_FORMAT_D16_UNORM_S8_UINT
:
2366 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
2367 ds
->offset_scale
= 4.0f
;
2369 case VK_FORMAT_D32_SFLOAT
:
2370 case VK_FORMAT_D32_SFLOAT_S8_UINT
:
2371 ds
->pa_su_poly_offset_db_fmt_cntl
= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
2372 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
2373 ds
->offset_scale
= 1.0f
;
2379 format
= radv_translate_dbformat(iview
->vk_format
);
2380 if (format
== V_028040_Z_INVALID
) {
2381 fprintf(stderr
, "Invalid DB format: %d, disabling DB.\n", iview
->vk_format
);
2384 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
;
2385 s_offs
= z_offs
= va
;
2386 z_offs
+= iview
->image
->surface
.level
[level
].offset
;
2387 s_offs
+= iview
->image
->surface
.stencil_level
[level
].offset
;
2389 uint32_t max_slice
= radv_surface_layer_count(iview
);
2390 ds
->db_depth_view
= S_028008_SLICE_START(iview
->base_layer
) |
2391 S_028008_SLICE_MAX(iview
->base_layer
+ max_slice
- 1);
2392 ds
->db_depth_info
= S_02803C_ADDR5_SWIZZLE_MASK(1);
2393 ds
->db_z_info
= S_028040_FORMAT(format
) | S_028040_ZRANGE_PRECISION(1);
2395 if (iview
->image
->samples
> 1)
2396 ds
->db_z_info
|= S_028040_NUM_SAMPLES(util_logbase2(iview
->image
->samples
));
2398 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
)
2399 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_8
);
2401 ds
->db_stencil_info
= S_028044_FORMAT(V_028044_STENCIL_INVALID
);
2403 if (device
->physical_device
->rad_info
.chip_class
>= CIK
) {
2404 struct radeon_info
*info
= &device
->physical_device
->rad_info
;
2405 unsigned tiling_index
= iview
->image
->surface
.tiling_index
[level
];
2406 unsigned stencil_index
= iview
->image
->surface
.stencil_tiling_index
[level
];
2407 unsigned macro_index
= iview
->image
->surface
.macro_tile_index
;
2408 unsigned tile_mode
= info
->si_tile_mode_array
[tiling_index
];
2409 unsigned stencil_tile_mode
= info
->si_tile_mode_array
[stencil_index
];
2410 unsigned macro_mode
= info
->cik_macrotile_mode_array
[macro_index
];
2412 ds
->db_depth_info
|=
2413 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode
)) |
2414 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode
)) |
2415 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode
)) |
2416 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode
)) |
2417 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode
)) |
2418 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode
));
2419 ds
->db_z_info
|= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode
));
2420 ds
->db_stencil_info
|= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode
));
2422 unsigned tile_mode_index
= si_tile_mode_index(iview
->image
, level
, false);
2423 ds
->db_z_info
|= S_028040_TILE_MODE_INDEX(tile_mode_index
);
2424 tile_mode_index
= si_tile_mode_index(iview
->image
, level
, true);
2425 ds
->db_stencil_info
|= S_028044_TILE_MODE_INDEX(tile_mode_index
);
2428 if (iview
->image
->surface
.htile_size
&& !level
) {
2429 ds
->db_z_info
|= S_028040_TILE_SURFACE_ENABLE(1) |
2430 S_028040_ALLOW_EXPCLEAR(1);
2432 if (iview
->image
->surface
.flags
& RADEON_SURF_SBUFFER
) {
2433 /* Workaround: For a not yet understood reason, the
2434 * combination of MSAA, fast stencil clear and stencil
2435 * decompress messes with subsequent stencil buffer
2436 * uses. Problem was reproduced on Verde, Bonaire,
2437 * Tonga, and Carrizo.
2439 * Disabling EXPCLEAR works around the problem.
2441 * Check piglit's arb_texture_multisample-stencil-clear
2442 * test if you want to try changing this.
2444 if (iview
->image
->samples
<= 1)
2445 ds
->db_stencil_info
|= S_028044_ALLOW_EXPCLEAR(1);
2447 /* Use all of the htile_buffer for depth if there's no stencil. */
2448 ds
->db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
2450 va
= device
->ws
->buffer_get_va(iview
->bo
) + iview
->image
->offset
+
2451 iview
->image
->htile_offset
;
2452 ds
->db_htile_data_base
= va
>> 8;
2453 ds
->db_htile_surface
= S_028ABC_FULL_CACHE(1);
2455 ds
->db_htile_data_base
= 0;
2456 ds
->db_htile_surface
= 0;
2459 ds
->db_z_read_base
= ds
->db_z_write_base
= z_offs
>> 8;
2460 ds
->db_stencil_read_base
= ds
->db_stencil_write_base
= s_offs
>> 8;
2462 ds
->db_depth_size
= S_028058_PITCH_TILE_MAX((level_info
->nblk_x
/ 8) - 1) |
2463 S_028058_HEIGHT_TILE_MAX((level_info
->nblk_y
/ 8) - 1);
2464 ds
->db_depth_slice
= S_02805C_SLICE_TILE_MAX((level_info
->nblk_x
* level_info
->nblk_y
) / 64 - 1);
2467 VkResult
radv_CreateFramebuffer(
2469 const VkFramebufferCreateInfo
* pCreateInfo
,
2470 const VkAllocationCallbacks
* pAllocator
,
2471 VkFramebuffer
* pFramebuffer
)
2473 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2474 struct radv_framebuffer
*framebuffer
;
2476 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
);
2478 size_t size
= sizeof(*framebuffer
) +
2479 sizeof(struct radv_attachment_info
) * pCreateInfo
->attachmentCount
;
2480 framebuffer
= vk_alloc2(&device
->alloc
, pAllocator
, size
, 8,
2481 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2482 if (framebuffer
== NULL
)
2483 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2485 framebuffer
->attachment_count
= pCreateInfo
->attachmentCount
;
2486 framebuffer
->width
= pCreateInfo
->width
;
2487 framebuffer
->height
= pCreateInfo
->height
;
2488 framebuffer
->layers
= pCreateInfo
->layers
;
2489 for (uint32_t i
= 0; i
< pCreateInfo
->attachmentCount
; i
++) {
2490 VkImageView _iview
= pCreateInfo
->pAttachments
[i
];
2491 struct radv_image_view
*iview
= radv_image_view_from_handle(_iview
);
2492 framebuffer
->attachments
[i
].attachment
= iview
;
2493 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
2494 radv_initialise_color_surface(device
, &framebuffer
->attachments
[i
].cb
, iview
);
2495 } else if (iview
->aspect_mask
& (VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
)) {
2496 radv_initialise_ds_surface(device
, &framebuffer
->attachments
[i
].ds
, iview
);
2498 framebuffer
->width
= MIN2(framebuffer
->width
, iview
->extent
.width
);
2499 framebuffer
->height
= MIN2(framebuffer
->height
, iview
->extent
.height
);
2500 framebuffer
->layers
= MIN2(framebuffer
->layers
, radv_surface_layer_count(iview
));
2503 *pFramebuffer
= radv_framebuffer_to_handle(framebuffer
);
2507 void radv_DestroyFramebuffer(
2510 const VkAllocationCallbacks
* pAllocator
)
2512 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2513 RADV_FROM_HANDLE(radv_framebuffer
, fb
, _fb
);
2517 vk_free2(&device
->alloc
, pAllocator
, fb
);
2520 static unsigned radv_tex_wrap(VkSamplerAddressMode address_mode
)
2522 switch (address_mode
) {
2523 case VK_SAMPLER_ADDRESS_MODE_REPEAT
:
2524 return V_008F30_SQ_TEX_WRAP
;
2525 case VK_SAMPLER_ADDRESS_MODE_MIRRORED_REPEAT
:
2526 return V_008F30_SQ_TEX_MIRROR
;
2527 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE
:
2528 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL
;
2529 case VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_BORDER
:
2530 return V_008F30_SQ_TEX_CLAMP_BORDER
;
2531 case VK_SAMPLER_ADDRESS_MODE_MIRROR_CLAMP_TO_EDGE
:
2532 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL
;
2534 unreachable("illegal tex wrap mode");
2540 radv_tex_compare(VkCompareOp op
)
2543 case VK_COMPARE_OP_NEVER
:
2544 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER
;
2545 case VK_COMPARE_OP_LESS
:
2546 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS
;
2547 case VK_COMPARE_OP_EQUAL
:
2548 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL
;
2549 case VK_COMPARE_OP_LESS_OR_EQUAL
:
2550 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL
;
2551 case VK_COMPARE_OP_GREATER
:
2552 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER
;
2553 case VK_COMPARE_OP_NOT_EQUAL
:
2554 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL
;
2555 case VK_COMPARE_OP_GREATER_OR_EQUAL
:
2556 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL
;
2557 case VK_COMPARE_OP_ALWAYS
:
2558 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS
;
2560 unreachable("illegal compare mode");
2566 radv_tex_filter(VkFilter filter
, unsigned max_ansio
)
2569 case VK_FILTER_NEAREST
:
2570 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
:
2571 V_008F38_SQ_TEX_XY_FILTER_POINT
);
2572 case VK_FILTER_LINEAR
:
2573 return (max_ansio
> 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
:
2574 V_008F38_SQ_TEX_XY_FILTER_BILINEAR
);
2575 case VK_FILTER_CUBIC_IMG
:
2577 fprintf(stderr
, "illegal texture filter");
2583 radv_tex_mipfilter(VkSamplerMipmapMode mode
)
2586 case VK_SAMPLER_MIPMAP_MODE_NEAREST
:
2587 return V_008F38_SQ_TEX_Z_FILTER_POINT
;
2588 case VK_SAMPLER_MIPMAP_MODE_LINEAR
:
2589 return V_008F38_SQ_TEX_Z_FILTER_LINEAR
;
2591 return V_008F38_SQ_TEX_Z_FILTER_NONE
;
2596 radv_tex_bordercolor(VkBorderColor bcolor
)
2599 case VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK
:
2600 case VK_BORDER_COLOR_INT_TRANSPARENT_BLACK
:
2601 return V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK
;
2602 case VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK
:
2603 case VK_BORDER_COLOR_INT_OPAQUE_BLACK
:
2604 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK
;
2605 case VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE
:
2606 case VK_BORDER_COLOR_INT_OPAQUE_WHITE
:
2607 return V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE
;
2615 radv_tex_aniso_filter(unsigned filter
)
2629 radv_init_sampler(struct radv_device
*device
,
2630 struct radv_sampler
*sampler
,
2631 const VkSamplerCreateInfo
*pCreateInfo
)
2633 uint32_t max_aniso
= pCreateInfo
->anisotropyEnable
&& pCreateInfo
->maxAnisotropy
> 1.0 ?
2634 (uint32_t) pCreateInfo
->maxAnisotropy
: 0;
2635 uint32_t max_aniso_ratio
= radv_tex_aniso_filter(max_aniso
);
2636 bool is_vi
= (device
->physical_device
->rad_info
.chip_class
>= VI
);
2638 sampler
->state
[0] = (S_008F30_CLAMP_X(radv_tex_wrap(pCreateInfo
->addressModeU
)) |
2639 S_008F30_CLAMP_Y(radv_tex_wrap(pCreateInfo
->addressModeV
)) |
2640 S_008F30_CLAMP_Z(radv_tex_wrap(pCreateInfo
->addressModeW
)) |
2641 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio
) |
2642 S_008F30_DEPTH_COMPARE_FUNC(radv_tex_compare(pCreateInfo
->compareOp
)) |
2643 S_008F30_FORCE_UNNORMALIZED(pCreateInfo
->unnormalizedCoordinates
? 1 : 0) |
2644 S_008F30_ANISO_THRESHOLD(max_aniso_ratio
>> 1) |
2645 S_008F30_ANISO_BIAS(max_aniso_ratio
) |
2646 S_008F30_DISABLE_CUBE_WRAP(0) |
2647 S_008F30_COMPAT_MODE(is_vi
));
2648 sampler
->state
[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(pCreateInfo
->minLod
, 0, 15), 8)) |
2649 S_008F34_MAX_LOD(S_FIXED(CLAMP(pCreateInfo
->maxLod
, 0, 15), 8)) |
2650 S_008F34_PERF_MIP(max_aniso_ratio
? max_aniso_ratio
+ 6 : 0));
2651 sampler
->state
[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(pCreateInfo
->mipLodBias
, -16, 16), 8)) |
2652 S_008F38_XY_MAG_FILTER(radv_tex_filter(pCreateInfo
->magFilter
, max_aniso
)) |
2653 S_008F38_XY_MIN_FILTER(radv_tex_filter(pCreateInfo
->minFilter
, max_aniso
)) |
2654 S_008F38_MIP_FILTER(radv_tex_mipfilter(pCreateInfo
->mipmapMode
)) |
2655 S_008F38_MIP_POINT_PRECLAMP(0) |
2656 S_008F38_DISABLE_LSB_CEIL(1) |
2657 S_008F38_FILTER_PREC_FIX(1) |
2658 S_008F38_ANISO_OVERRIDE(is_vi
));
2659 sampler
->state
[3] = (S_008F3C_BORDER_COLOR_PTR(0) |
2660 S_008F3C_BORDER_COLOR_TYPE(radv_tex_bordercolor(pCreateInfo
->borderColor
)));
2663 VkResult
radv_CreateSampler(
2665 const VkSamplerCreateInfo
* pCreateInfo
,
2666 const VkAllocationCallbacks
* pAllocator
,
2667 VkSampler
* pSampler
)
2669 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2670 struct radv_sampler
*sampler
;
2672 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO
);
2674 sampler
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*sampler
), 8,
2675 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2677 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2679 radv_init_sampler(device
, sampler
, pCreateInfo
);
2680 *pSampler
= radv_sampler_to_handle(sampler
);
2685 void radv_DestroySampler(
2688 const VkAllocationCallbacks
* pAllocator
)
2690 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2691 RADV_FROM_HANDLE(radv_sampler
, sampler
, _sampler
);
2695 vk_free2(&device
->alloc
, pAllocator
, sampler
);
2699 /* vk_icd.h does not declare this function, so we declare it here to
2700 * suppress Wmissing-prototypes.
2702 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2703 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
);
2705 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2706 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion
)
2708 /* For the full details on loader interface versioning, see
2709 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2710 * What follows is a condensed summary, to help you navigate the large and
2711 * confusing official doc.
2713 * - Loader interface v0 is incompatible with later versions. We don't
2716 * - In loader interface v1:
2717 * - The first ICD entrypoint called by the loader is
2718 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2720 * - The ICD must statically expose no other Vulkan symbol unless it is
2721 * linked with -Bsymbolic.
2722 * - Each dispatchable Vulkan handle created by the ICD must be
2723 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2724 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2725 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2726 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2727 * such loader-managed surfaces.
2729 * - Loader interface v2 differs from v1 in:
2730 * - The first ICD entrypoint called by the loader is
2731 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2732 * statically expose this entrypoint.
2734 * - Loader interface v3 differs from v2 in:
2735 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2736 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2737 * because the loader no longer does so.
2739 *pSupportedVersion
= MIN2(*pSupportedVersion
, 3u);