Merge remote-tracking branch 'mesa-public/master' into vulkan
[mesa.git] / src / vulkan / anv_device.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <unistd.h>
28 #include <fcntl.h>
29
30 #include "anv_private.h"
31 #include "mesa/main/git_sha1.h"
32 #include "util/strtod.h"
33 #include "util/debug.h"
34
35 #include "gen7_pack.h"
36
37 struct anv_dispatch_table dtable;
38
39 static void
40 compiler_debug_log(void *data, const char *fmt, ...)
41 { }
42
43 static void
44 compiler_perf_log(void *data, const char *fmt, ...)
45 {
46 va_list args;
47 va_start(args, fmt);
48
49 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
50 vfprintf(stderr, fmt, args);
51
52 va_end(args);
53 }
54
55 static VkResult
56 anv_physical_device_init(struct anv_physical_device *device,
57 struct anv_instance *instance,
58 const char *path)
59 {
60 VkResult result;
61 int fd;
62
63 fd = open(path, O_RDWR | O_CLOEXEC);
64 if (fd < 0)
65 return vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
66 "failed to open %s: %m", path);
67
68 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
69 device->instance = instance;
70 device->path = path;
71
72 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
73 if (!device->chipset_id) {
74 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
75 "failed to get chipset id: %m");
76 goto fail;
77 }
78
79 device->name = brw_get_device_name(device->chipset_id);
80 device->info = brw_get_device_info(device->chipset_id);
81 if (!device->info) {
82 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
83 "failed to get device info");
84 goto fail;
85 }
86
87 if (device->info->is_haswell) {
88 fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
89 } else if (device->info->gen == 7 && !device->info->is_baytrail) {
90 fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
91 } else if (device->info->gen == 7 && device->info->is_baytrail) {
92 fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
93 } else if (device->info->gen >= 8) {
94 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
95 * supported as anything */
96 } else {
97 result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
98 "Vulkan not yet supported on %s", device->name);
99 goto fail;
100 }
101
102 if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
103 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
104 "failed to get aperture size: %m");
105 goto fail;
106 }
107
108 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
109 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
110 "kernel missing gem wait");
111 goto fail;
112 }
113
114 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
115 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
116 "kernel missing execbuf2");
117 goto fail;
118 }
119
120 if (!device->info->has_llc &&
121 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
122 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
123 "kernel missing wc mmap");
124 goto fail;
125 }
126
127 bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
128
129 close(fd);
130
131 brw_process_intel_debug_variable();
132
133 device->compiler = brw_compiler_create(NULL, device->info);
134 if (device->compiler == NULL) {
135 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
136 goto fail;
137 }
138 device->compiler->shader_debug_log = compiler_debug_log;
139 device->compiler->shader_perf_log = compiler_perf_log;
140
141 /* XXX: Actually detect bit6 swizzling */
142 isl_device_init(&device->isl_dev, device->info, swizzled);
143
144 return VK_SUCCESS;
145
146 fail:
147 close(fd);
148 return result;
149 }
150
151 static void
152 anv_physical_device_finish(struct anv_physical_device *device)
153 {
154 ralloc_free(device->compiler);
155 }
156
157 static const VkExtensionProperties global_extensions[] = {
158 {
159 .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
160 .specVersion = 25,
161 },
162 {
163 .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
164 .specVersion = 5,
165 },
166 #ifdef HAVE_WAYLAND_PLATFORM
167 {
168 .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
169 .specVersion = 4,
170 },
171 #endif
172 };
173
174 static const VkExtensionProperties device_extensions[] = {
175 {
176 .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
177 .specVersion = 67,
178 },
179 };
180
181 static void *
182 default_alloc_func(void *pUserData, size_t size, size_t align,
183 VkSystemAllocationScope allocationScope)
184 {
185 return malloc(size);
186 }
187
188 static void *
189 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
190 size_t align, VkSystemAllocationScope allocationScope)
191 {
192 return realloc(pOriginal, size);
193 }
194
195 static void
196 default_free_func(void *pUserData, void *pMemory)
197 {
198 free(pMemory);
199 }
200
201 static const VkAllocationCallbacks default_alloc = {
202 .pUserData = NULL,
203 .pfnAllocation = default_alloc_func,
204 .pfnReallocation = default_realloc_func,
205 .pfnFree = default_free_func,
206 };
207
208 VkResult anv_CreateInstance(
209 const VkInstanceCreateInfo* pCreateInfo,
210 const VkAllocationCallbacks* pAllocator,
211 VkInstance* pInstance)
212 {
213 struct anv_instance *instance;
214
215 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
216
217 uint32_t client_version = pCreateInfo->pApplicationInfo->apiVersion;
218 if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
219 client_version > VK_MAKE_VERSION(1, 0, 2)) {
220 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
221 "Client requested version %d.%d.%d",
222 VK_VERSION_MAJOR(client_version),
223 VK_VERSION_MINOR(client_version),
224 VK_VERSION_PATCH(client_version));
225 }
226
227 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
228 bool found = false;
229 for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
230 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
231 global_extensions[j].extensionName) == 0) {
232 found = true;
233 break;
234 }
235 }
236 if (!found)
237 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
238 }
239
240 instance = anv_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
241 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
242 if (!instance)
243 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
244
245 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
246
247 if (pAllocator)
248 instance->alloc = *pAllocator;
249 else
250 instance->alloc = default_alloc;
251
252 instance->apiVersion = pCreateInfo->pApplicationInfo->apiVersion;
253 instance->physicalDeviceCount = -1;
254
255 _mesa_locale_init();
256
257 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
258
259 anv_init_wsi(instance);
260
261 *pInstance = anv_instance_to_handle(instance);
262
263 return VK_SUCCESS;
264 }
265
266 void anv_DestroyInstance(
267 VkInstance _instance,
268 const VkAllocationCallbacks* pAllocator)
269 {
270 ANV_FROM_HANDLE(anv_instance, instance, _instance);
271
272 if (instance->physicalDeviceCount > 0) {
273 /* We support at most one physical device. */
274 assert(instance->physicalDeviceCount == 1);
275 anv_physical_device_finish(&instance->physicalDevice);
276 }
277
278 anv_finish_wsi(instance);
279
280 VG(VALGRIND_DESTROY_MEMPOOL(instance));
281
282 _mesa_locale_fini();
283
284 anv_free(&instance->alloc, instance);
285 }
286
287 VkResult anv_EnumeratePhysicalDevices(
288 VkInstance _instance,
289 uint32_t* pPhysicalDeviceCount,
290 VkPhysicalDevice* pPhysicalDevices)
291 {
292 ANV_FROM_HANDLE(anv_instance, instance, _instance);
293 VkResult result;
294
295 if (instance->physicalDeviceCount < 0) {
296 result = anv_physical_device_init(&instance->physicalDevice,
297 instance, "/dev/dri/renderD128");
298 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
299 instance->physicalDeviceCount = 0;
300 } else if (result == VK_SUCCESS) {
301 instance->physicalDeviceCount = 1;
302 } else {
303 return result;
304 }
305 }
306
307 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
308 * otherwise it's an inout parameter.
309 *
310 * The Vulkan spec (git aaed022) says:
311 *
312 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
313 * that is initialized with the number of devices the application is
314 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
315 * an array of at least this many VkPhysicalDevice handles [...].
316 *
317 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
318 * overwrites the contents of the variable pointed to by
319 * pPhysicalDeviceCount with the number of physical devices in in the
320 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
321 * pPhysicalDeviceCount with the number of physical handles written to
322 * pPhysicalDevices.
323 */
324 if (!pPhysicalDevices) {
325 *pPhysicalDeviceCount = instance->physicalDeviceCount;
326 } else if (*pPhysicalDeviceCount >= 1) {
327 pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
328 *pPhysicalDeviceCount = 1;
329 } else {
330 *pPhysicalDeviceCount = 0;
331 }
332
333 return VK_SUCCESS;
334 }
335
336 void anv_GetPhysicalDeviceFeatures(
337 VkPhysicalDevice physicalDevice,
338 VkPhysicalDeviceFeatures* pFeatures)
339 {
340 anv_finishme("Get correct values for PhysicalDeviceFeatures");
341
342 *pFeatures = (VkPhysicalDeviceFeatures) {
343 .robustBufferAccess = true,
344 .fullDrawIndexUint32 = false,
345 .imageCubeArray = false,
346 .independentBlend = false,
347 .geometryShader = true,
348 .tessellationShader = false,
349 .sampleRateShading = false,
350 .dualSrcBlend = true,
351 .logicOp = true,
352 .multiDrawIndirect = false,
353 .drawIndirectFirstInstance = false,
354 .depthClamp = false,
355 .depthBiasClamp = false,
356 .fillModeNonSolid = true,
357 .depthBounds = false,
358 .wideLines = true,
359 .largePoints = true,
360 .alphaToOne = true,
361 .multiViewport = true,
362 .samplerAnisotropy = false, /* FINISHME */
363 .textureCompressionETC2 = true,
364 .textureCompressionASTC_LDR = true,
365 .textureCompressionBC = true,
366 .occlusionQueryPrecise = false, /* FINISHME */
367 .pipelineStatisticsQuery = true,
368 .vertexPipelineStoresAndAtomics = false,
369 .fragmentStoresAndAtomics = true,
370 .shaderTessellationAndGeometryPointSize = true,
371 .shaderImageGatherExtended = true,
372 .shaderStorageImageExtendedFormats = false,
373 .shaderStorageImageMultisample = false,
374 .shaderUniformBufferArrayDynamicIndexing = true,
375 .shaderSampledImageArrayDynamicIndexing = false,
376 .shaderStorageBufferArrayDynamicIndexing = false,
377 .shaderStorageImageArrayDynamicIndexing = false,
378 .shaderStorageImageReadWithoutFormat = false,
379 .shaderStorageImageWriteWithoutFormat = true,
380 .shaderClipDistance = false,
381 .shaderCullDistance = false,
382 .shaderFloat64 = false,
383 .shaderInt64 = false,
384 .shaderInt16 = false,
385 .alphaToOne = true,
386 .variableMultisampleRate = false,
387 .inheritedQueries = false,
388 };
389 }
390
391 void anv_GetPhysicalDeviceProperties(
392 VkPhysicalDevice physicalDevice,
393 VkPhysicalDeviceProperties* pProperties)
394 {
395 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
396 const struct brw_device_info *devinfo = pdevice->info;
397
398 anv_finishme("Get correct values for VkPhysicalDeviceLimits");
399
400 const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
401
402 VkSampleCountFlags sample_counts =
403 isl_device_get_sample_counts(&pdevice->isl_dev);
404
405 VkPhysicalDeviceLimits limits = {
406 .maxImageDimension1D = (1 << 14),
407 .maxImageDimension2D = (1 << 14),
408 .maxImageDimension3D = (1 << 10),
409 .maxImageDimensionCube = (1 << 14),
410 .maxImageArrayLayers = (1 << 10),
411 .maxTexelBufferElements = 128 * 1024 * 1024,
412 .maxUniformBufferRange = UINT32_MAX,
413 .maxStorageBufferRange = UINT32_MAX,
414 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
415 .maxMemoryAllocationCount = UINT32_MAX,
416 .maxSamplerAllocationCount = 64 * 1024,
417 .bufferImageGranularity = 64, /* A cache line */
418 .sparseAddressSpaceSize = 0,
419 .maxBoundDescriptorSets = MAX_SETS,
420 .maxPerStageDescriptorSamplers = 64,
421 .maxPerStageDescriptorUniformBuffers = 64,
422 .maxPerStageDescriptorStorageBuffers = 64,
423 .maxPerStageDescriptorSampledImages = 64,
424 .maxPerStageDescriptorStorageImages = 64,
425 .maxPerStageDescriptorInputAttachments = 64,
426 .maxPerStageResources = 128,
427 .maxDescriptorSetSamplers = 256,
428 .maxDescriptorSetUniformBuffers = 256,
429 .maxDescriptorSetUniformBuffersDynamic = 256,
430 .maxDescriptorSetStorageBuffers = 256,
431 .maxDescriptorSetStorageBuffersDynamic = 256,
432 .maxDescriptorSetSampledImages = 256,
433 .maxDescriptorSetStorageImages = 256,
434 .maxDescriptorSetInputAttachments = 256,
435 .maxVertexInputAttributes = 32,
436 .maxVertexInputBindings = 32,
437 .maxVertexInputAttributeOffset = 2047,
438 .maxVertexInputBindingStride = 2048,
439 .maxVertexOutputComponents = 128,
440 .maxTessellationGenerationLevel = 0,
441 .maxTessellationPatchSize = 0,
442 .maxTessellationControlPerVertexInputComponents = 0,
443 .maxTessellationControlPerVertexOutputComponents = 0,
444 .maxTessellationControlPerPatchOutputComponents = 0,
445 .maxTessellationControlTotalOutputComponents = 0,
446 .maxTessellationEvaluationInputComponents = 0,
447 .maxTessellationEvaluationOutputComponents = 0,
448 .maxGeometryShaderInvocations = 32,
449 .maxGeometryInputComponents = 64,
450 .maxGeometryOutputComponents = 128,
451 .maxGeometryOutputVertices = 256,
452 .maxGeometryTotalOutputComponents = 1024,
453 .maxFragmentInputComponents = 128,
454 .maxFragmentOutputAttachments = 8,
455 .maxFragmentDualSrcAttachments = 2,
456 .maxFragmentCombinedOutputResources = 8,
457 .maxComputeSharedMemorySize = 32768,
458 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
459 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
460 .maxComputeWorkGroupSize = {
461 16 * devinfo->max_cs_threads,
462 16 * devinfo->max_cs_threads,
463 16 * devinfo->max_cs_threads,
464 },
465 .subPixelPrecisionBits = 4 /* FIXME */,
466 .subTexelPrecisionBits = 4 /* FIXME */,
467 .mipmapPrecisionBits = 4 /* FIXME */,
468 .maxDrawIndexedIndexValue = UINT32_MAX,
469 .maxDrawIndirectCount = UINT32_MAX,
470 .maxSamplerLodBias = 16,
471 .maxSamplerAnisotropy = 16,
472 .maxViewports = MAX_VIEWPORTS,
473 .maxViewportDimensions = { (1 << 14), (1 << 14) },
474 .viewportBoundsRange = { -16384.0, 16384.0 },
475 .viewportSubPixelBits = 13, /* We take a float? */
476 .minMemoryMapAlignment = 4096, /* A page */
477 .minTexelBufferOffsetAlignment = 1,
478 .minUniformBufferOffsetAlignment = 1,
479 .minStorageBufferOffsetAlignment = 1,
480 .minTexelOffset = -8,
481 .maxTexelOffset = 7,
482 .minTexelGatherOffset = -8,
483 .maxTexelGatherOffset = 7,
484 .minInterpolationOffset = 0, /* FIXME */
485 .maxInterpolationOffset = 0, /* FIXME */
486 .subPixelInterpolationOffsetBits = 0, /* FIXME */
487 .maxFramebufferWidth = (1 << 14),
488 .maxFramebufferHeight = (1 << 14),
489 .maxFramebufferLayers = (1 << 10),
490 .framebufferColorSampleCounts = sample_counts,
491 .framebufferDepthSampleCounts = sample_counts,
492 .framebufferStencilSampleCounts = sample_counts,
493 .framebufferNoAttachmentsSampleCounts = sample_counts,
494 .maxColorAttachments = MAX_RTS,
495 .sampledImageColorSampleCounts = sample_counts,
496 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
497 .sampledImageDepthSampleCounts = sample_counts,
498 .sampledImageStencilSampleCounts = sample_counts,
499 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
500 .maxSampleMaskWords = 1,
501 .timestampComputeAndGraphics = false,
502 .timestampPeriod = time_stamp_base / (1000 * 1000 * 1000),
503 .maxClipDistances = 0 /* FIXME */,
504 .maxCullDistances = 0 /* FIXME */,
505 .maxCombinedClipAndCullDistances = 0 /* FIXME */,
506 .discreteQueuePriorities = 1,
507 .pointSizeRange = { 0.125, 255.875 },
508 .lineWidthRange = { 0.0, 7.9921875 },
509 .pointSizeGranularity = (1.0 / 8.0),
510 .lineWidthGranularity = (1.0 / 128.0),
511 .strictLines = false, /* FINISHME */
512 .standardSampleLocations = true,
513 .optimalBufferCopyOffsetAlignment = 128,
514 .optimalBufferCopyRowPitchAlignment = 128,
515 .nonCoherentAtomSize = 64,
516 };
517
518 *pProperties = (VkPhysicalDeviceProperties) {
519 .apiVersion = VK_MAKE_VERSION(1, 0, 2),
520 .driverVersion = 1,
521 .vendorID = 0x8086,
522 .deviceID = pdevice->chipset_id,
523 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
524 .limits = limits,
525 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
526 };
527
528 strcpy(pProperties->deviceName, pdevice->name);
529 snprintf((char *)pProperties->pipelineCacheUUID, VK_UUID_SIZE,
530 "anv-%s", MESA_GIT_SHA1 + 4);
531 }
532
533 void anv_GetPhysicalDeviceQueueFamilyProperties(
534 VkPhysicalDevice physicalDevice,
535 uint32_t* pCount,
536 VkQueueFamilyProperties* pQueueFamilyProperties)
537 {
538 if (pQueueFamilyProperties == NULL) {
539 *pCount = 1;
540 return;
541 }
542
543 assert(*pCount >= 1);
544
545 *pQueueFamilyProperties = (VkQueueFamilyProperties) {
546 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
547 VK_QUEUE_COMPUTE_BIT |
548 VK_QUEUE_TRANSFER_BIT,
549 .queueCount = 1,
550 .timestampValidBits = 36, /* XXX: Real value here */
551 .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
552 };
553 }
554
555 void anv_GetPhysicalDeviceMemoryProperties(
556 VkPhysicalDevice physicalDevice,
557 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
558 {
559 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
560 VkDeviceSize heap_size;
561
562 /* Reserve some wiggle room for the driver by exposing only 75% of the
563 * aperture to the heap.
564 */
565 heap_size = 3 * physical_device->aperture_size / 4;
566
567 if (physical_device->info->has_llc) {
568 /* Big core GPUs share LLC with the CPU and thus one memory type can be
569 * both cached and coherent at the same time.
570 */
571 pMemoryProperties->memoryTypeCount = 1;
572 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
573 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
574 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
575 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
576 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
577 .heapIndex = 0,
578 };
579 } else {
580 /* The spec requires that we expose a host-visible, coherent memory
581 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
582 * to give the application a choice between cached, but not coherent and
583 * coherent but uncached (WC though).
584 */
585 pMemoryProperties->memoryTypeCount = 2;
586 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
587 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
588 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
589 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
590 .heapIndex = 0,
591 };
592 pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
593 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
594 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
595 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
596 .heapIndex = 0,
597 };
598 }
599
600 pMemoryProperties->memoryHeapCount = 1;
601 pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
602 .size = heap_size,
603 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
604 };
605 }
606
607 PFN_vkVoidFunction anv_GetInstanceProcAddr(
608 VkInstance instance,
609 const char* pName)
610 {
611 return anv_lookup_entrypoint(pName);
612 }
613
614 PFN_vkVoidFunction anv_GetDeviceProcAddr(
615 VkDevice device,
616 const char* pName)
617 {
618 return anv_lookup_entrypoint(pName);
619 }
620
621 static VkResult
622 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
623 {
624 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
625 queue->device = device;
626 queue->pool = &device->surface_state_pool;
627
628 return VK_SUCCESS;
629 }
630
631 static void
632 anv_queue_finish(struct anv_queue *queue)
633 {
634 }
635
636 static struct anv_state
637 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
638 {
639 struct anv_state state;
640
641 state = anv_state_pool_alloc(pool, size, align);
642 memcpy(state.map, p, size);
643
644 if (!pool->block_pool->device->info.has_llc)
645 anv_state_clflush(state);
646
647 return state;
648 }
649
650 struct gen8_border_color {
651 union {
652 float float32[4];
653 uint32_t uint32[4];
654 };
655 /* Pad out to 64 bytes */
656 uint32_t _pad[12];
657 };
658
659 static void
660 anv_device_init_border_colors(struct anv_device *device)
661 {
662 static const struct gen8_border_color border_colors[] = {
663 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
664 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
665 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
666 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
667 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
668 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
669 };
670
671 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
672 sizeof(border_colors), 64,
673 border_colors);
674 }
675
676 VkResult anv_CreateDevice(
677 VkPhysicalDevice physicalDevice,
678 const VkDeviceCreateInfo* pCreateInfo,
679 const VkAllocationCallbacks* pAllocator,
680 VkDevice* pDevice)
681 {
682 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
683 VkResult result;
684 struct anv_device *device;
685
686 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
687
688 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
689 bool found = false;
690 for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
691 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
692 device_extensions[j].extensionName) == 0) {
693 found = true;
694 break;
695 }
696 }
697 if (!found)
698 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
699 }
700
701 anv_set_dispatch_devinfo(physical_device->info);
702
703 device = anv_alloc2(&physical_device->instance->alloc, pAllocator,
704 sizeof(*device), 8,
705 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
706 if (!device)
707 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
708
709 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
710 device->instance = physical_device->instance;
711
712 if (pAllocator)
713 device->alloc = *pAllocator;
714 else
715 device->alloc = physical_device->instance->alloc;
716
717 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
718 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
719 if (device->fd == -1) {
720 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
721 goto fail_device;
722 }
723
724 device->context_id = anv_gem_create_context(device);
725 if (device->context_id == -1) {
726 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
727 goto fail_fd;
728 }
729
730 device->info = *physical_device->info;
731 device->isl_dev = physical_device->isl_dev;
732
733 pthread_mutex_init(&device->mutex, NULL);
734
735 anv_bo_pool_init(&device->batch_bo_pool, device, ANV_CMD_BUFFER_BATCH_SIZE);
736
737 anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
738
739 anv_state_pool_init(&device->dynamic_state_pool,
740 &device->dynamic_state_block_pool);
741
742 anv_block_pool_init(&device->instruction_block_pool, device, 128 * 1024);
743 anv_pipeline_cache_init(&device->default_pipeline_cache, device);
744
745 anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
746
747 anv_state_pool_init(&device->surface_state_pool,
748 &device->surface_state_block_pool);
749
750 anv_bo_init_new(&device->workaround_bo, device, 1024);
751
752 anv_block_pool_init(&device->scratch_block_pool, device, 0x10000);
753
754 anv_queue_init(device, &device->queue);
755
756 result = anv_device_init_meta(device);
757 if (result != VK_SUCCESS)
758 goto fail_fd;
759
760 anv_device_init_border_colors(device);
761
762 *pDevice = anv_device_to_handle(device);
763
764 return VK_SUCCESS;
765
766 fail_fd:
767 close(device->fd);
768 fail_device:
769 anv_free(&device->alloc, device);
770
771 return result;
772 }
773
774 void anv_DestroyDevice(
775 VkDevice _device,
776 const VkAllocationCallbacks* pAllocator)
777 {
778 ANV_FROM_HANDLE(anv_device, device, _device);
779
780 anv_queue_finish(&device->queue);
781
782 anv_device_finish_meta(device);
783
784 #ifdef HAVE_VALGRIND
785 /* We only need to free these to prevent valgrind errors. The backing
786 * BO will go away in a couple of lines so we don't actually leak.
787 */
788 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
789 #endif
790
791 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
792 anv_gem_close(device, device->workaround_bo.gem_handle);
793
794 anv_bo_pool_finish(&device->batch_bo_pool);
795 anv_state_pool_finish(&device->dynamic_state_pool);
796 anv_block_pool_finish(&device->dynamic_state_block_pool);
797 anv_block_pool_finish(&device->instruction_block_pool);
798 anv_state_pool_finish(&device->surface_state_pool);
799 anv_block_pool_finish(&device->surface_state_block_pool);
800 anv_block_pool_finish(&device->scratch_block_pool);
801
802 close(device->fd);
803
804 pthread_mutex_destroy(&device->mutex);
805
806 anv_free(&device->alloc, device);
807 }
808
809 VkResult anv_EnumerateInstanceExtensionProperties(
810 const char* pLayerName,
811 uint32_t* pPropertyCount,
812 VkExtensionProperties* pProperties)
813 {
814 if (pProperties == NULL) {
815 *pPropertyCount = ARRAY_SIZE(global_extensions);
816 return VK_SUCCESS;
817 }
818
819 assert(*pPropertyCount >= ARRAY_SIZE(global_extensions));
820
821 *pPropertyCount = ARRAY_SIZE(global_extensions);
822 memcpy(pProperties, global_extensions, sizeof(global_extensions));
823
824 return VK_SUCCESS;
825 }
826
827 VkResult anv_EnumerateDeviceExtensionProperties(
828 VkPhysicalDevice physicalDevice,
829 const char* pLayerName,
830 uint32_t* pPropertyCount,
831 VkExtensionProperties* pProperties)
832 {
833 if (pProperties == NULL) {
834 *pPropertyCount = ARRAY_SIZE(device_extensions);
835 return VK_SUCCESS;
836 }
837
838 assert(*pPropertyCount >= ARRAY_SIZE(device_extensions));
839
840 *pPropertyCount = ARRAY_SIZE(device_extensions);
841 memcpy(pProperties, device_extensions, sizeof(device_extensions));
842
843 return VK_SUCCESS;
844 }
845
846 VkResult anv_EnumerateInstanceLayerProperties(
847 uint32_t* pPropertyCount,
848 VkLayerProperties* pProperties)
849 {
850 if (pProperties == NULL) {
851 *pPropertyCount = 0;
852 return VK_SUCCESS;
853 }
854
855 /* None supported at this time */
856 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
857 }
858
859 VkResult anv_EnumerateDeviceLayerProperties(
860 VkPhysicalDevice physicalDevice,
861 uint32_t* pPropertyCount,
862 VkLayerProperties* pProperties)
863 {
864 if (pProperties == NULL) {
865 *pPropertyCount = 0;
866 return VK_SUCCESS;
867 }
868
869 /* None supported at this time */
870 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
871 }
872
873 void anv_GetDeviceQueue(
874 VkDevice _device,
875 uint32_t queueNodeIndex,
876 uint32_t queueIndex,
877 VkQueue* pQueue)
878 {
879 ANV_FROM_HANDLE(anv_device, device, _device);
880
881 assert(queueIndex == 0);
882
883 *pQueue = anv_queue_to_handle(&device->queue);
884 }
885
886 VkResult anv_QueueSubmit(
887 VkQueue _queue,
888 uint32_t submitCount,
889 const VkSubmitInfo* pSubmits,
890 VkFence _fence)
891 {
892 ANV_FROM_HANDLE(anv_queue, queue, _queue);
893 ANV_FROM_HANDLE(anv_fence, fence, _fence);
894 struct anv_device *device = queue->device;
895 int ret;
896
897 for (uint32_t i = 0; i < submitCount; i++) {
898 for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
899 ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
900 pSubmits[i].pCommandBuffers[j]);
901 assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
902
903 ret = anv_gem_execbuffer(device, &cmd_buffer->execbuf2.execbuf);
904 if (ret != 0) {
905 /* We don't know the real error. */
906 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
907 "execbuf2 failed: %m");
908 }
909
910 if (fence) {
911 ret = anv_gem_execbuffer(device, &fence->execbuf);
912 if (ret != 0) {
913 /* We don't know the real error. */
914 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
915 "execbuf2 failed: %m");
916 }
917 }
918
919 for (uint32_t k = 0; k < cmd_buffer->execbuf2.bo_count; k++)
920 cmd_buffer->execbuf2.bos[k]->offset = cmd_buffer->execbuf2.objects[k].offset;
921 }
922 }
923
924 return VK_SUCCESS;
925 }
926
927 VkResult anv_QueueWaitIdle(
928 VkQueue _queue)
929 {
930 ANV_FROM_HANDLE(anv_queue, queue, _queue);
931
932 return ANV_CALL(DeviceWaitIdle)(anv_device_to_handle(queue->device));
933 }
934
935 VkResult anv_DeviceWaitIdle(
936 VkDevice _device)
937 {
938 ANV_FROM_HANDLE(anv_device, device, _device);
939 struct anv_state state;
940 struct anv_batch batch;
941 struct drm_i915_gem_execbuffer2 execbuf;
942 struct drm_i915_gem_exec_object2 exec2_objects[1];
943 struct anv_bo *bo = NULL;
944 VkResult result;
945 int64_t timeout;
946 int ret;
947
948 state = anv_state_pool_alloc(&device->dynamic_state_pool, 32, 32);
949 bo = &device->dynamic_state_pool.block_pool->bo;
950 batch.start = batch.next = state.map;
951 batch.end = state.map + 32;
952 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END);
953 anv_batch_emit(&batch, GEN7_MI_NOOP);
954
955 if (!device->info.has_llc)
956 anv_state_clflush(state);
957
958 exec2_objects[0].handle = bo->gem_handle;
959 exec2_objects[0].relocation_count = 0;
960 exec2_objects[0].relocs_ptr = 0;
961 exec2_objects[0].alignment = 0;
962 exec2_objects[0].offset = bo->offset;
963 exec2_objects[0].flags = 0;
964 exec2_objects[0].rsvd1 = 0;
965 exec2_objects[0].rsvd2 = 0;
966
967 execbuf.buffers_ptr = (uintptr_t) exec2_objects;
968 execbuf.buffer_count = 1;
969 execbuf.batch_start_offset = state.offset;
970 execbuf.batch_len = batch.next - state.map;
971 execbuf.cliprects_ptr = 0;
972 execbuf.num_cliprects = 0;
973 execbuf.DR1 = 0;
974 execbuf.DR4 = 0;
975
976 execbuf.flags =
977 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
978 execbuf.rsvd1 = device->context_id;
979 execbuf.rsvd2 = 0;
980
981 ret = anv_gem_execbuffer(device, &execbuf);
982 if (ret != 0) {
983 /* We don't know the real error. */
984 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
985 goto fail;
986 }
987
988 timeout = INT64_MAX;
989 ret = anv_gem_wait(device, bo->gem_handle, &timeout);
990 if (ret != 0) {
991 /* We don't know the real error. */
992 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
993 goto fail;
994 }
995
996 anv_state_pool_free(&device->dynamic_state_pool, state);
997
998 return VK_SUCCESS;
999
1000 fail:
1001 anv_state_pool_free(&device->dynamic_state_pool, state);
1002
1003 return result;
1004 }
1005
1006 VkResult
1007 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1008 {
1009 bo->gem_handle = anv_gem_create(device, size);
1010 if (!bo->gem_handle)
1011 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1012
1013 bo->map = NULL;
1014 bo->index = 0;
1015 bo->offset = 0;
1016 bo->size = size;
1017
1018 return VK_SUCCESS;
1019 }
1020
1021 VkResult anv_AllocateMemory(
1022 VkDevice _device,
1023 const VkMemoryAllocateInfo* pAllocateInfo,
1024 const VkAllocationCallbacks* pAllocator,
1025 VkDeviceMemory* pMem)
1026 {
1027 ANV_FROM_HANDLE(anv_device, device, _device);
1028 struct anv_device_memory *mem;
1029 VkResult result;
1030
1031 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1032
1033 if (pAllocateInfo->allocationSize == 0) {
1034 /* Apparently, this is allowed */
1035 *pMem = VK_NULL_HANDLE;
1036 return VK_SUCCESS;
1037 }
1038
1039 /* We support exactly one memory heap. */
1040 assert(pAllocateInfo->memoryTypeIndex == 0 ||
1041 (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1042
1043 /* FINISHME: Fail if allocation request exceeds heap size. */
1044
1045 mem = anv_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1046 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1047 if (mem == NULL)
1048 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1049
1050 /* The kernel is going to give us whole pages anyway */
1051 uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1052
1053 result = anv_bo_init_new(&mem->bo, device, alloc_size);
1054 if (result != VK_SUCCESS)
1055 goto fail;
1056
1057 mem->type_index = pAllocateInfo->memoryTypeIndex;
1058
1059 *pMem = anv_device_memory_to_handle(mem);
1060
1061 return VK_SUCCESS;
1062
1063 fail:
1064 anv_free2(&device->alloc, pAllocator, mem);
1065
1066 return result;
1067 }
1068
1069 void anv_FreeMemory(
1070 VkDevice _device,
1071 VkDeviceMemory _mem,
1072 const VkAllocationCallbacks* pAllocator)
1073 {
1074 ANV_FROM_HANDLE(anv_device, device, _device);
1075 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1076
1077 if (mem == NULL)
1078 return;
1079
1080 if (mem->bo.map)
1081 anv_gem_munmap(mem->bo.map, mem->bo.size);
1082
1083 if (mem->bo.gem_handle != 0)
1084 anv_gem_close(device, mem->bo.gem_handle);
1085
1086 anv_free2(&device->alloc, pAllocator, mem);
1087 }
1088
1089 VkResult anv_MapMemory(
1090 VkDevice _device,
1091 VkDeviceMemory _memory,
1092 VkDeviceSize offset,
1093 VkDeviceSize size,
1094 VkMemoryMapFlags flags,
1095 void** ppData)
1096 {
1097 ANV_FROM_HANDLE(anv_device, device, _device);
1098 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1099
1100 if (mem == NULL) {
1101 *ppData = NULL;
1102 return VK_SUCCESS;
1103 }
1104
1105 if (size == VK_WHOLE_SIZE)
1106 size = mem->bo.size - offset;
1107
1108 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1109 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1110 * at a time is valid. We could just mmap up front and return an offset
1111 * pointer here, but that may exhaust virtual memory on 32 bit
1112 * userspace. */
1113
1114 uint32_t gem_flags = 0;
1115 if (!device->info.has_llc && mem->type_index == 0)
1116 gem_flags |= I915_MMAP_WC;
1117
1118 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1119 uint64_t map_offset = offset & ~4095ull;
1120 assert(offset >= map_offset);
1121 uint64_t map_size = (offset + size) - map_offset;
1122
1123 /* Let's map whole pages */
1124 map_size = align_u64(map_size, 4096);
1125
1126 mem->map = anv_gem_mmap(device, mem->bo.gem_handle,
1127 map_offset, map_size, gem_flags);
1128 mem->map_size = map_size;
1129
1130 *ppData = mem->map + (offset - map_offset);
1131
1132 return VK_SUCCESS;
1133 }
1134
1135 void anv_UnmapMemory(
1136 VkDevice _device,
1137 VkDeviceMemory _memory)
1138 {
1139 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1140
1141 if (mem == NULL)
1142 return;
1143
1144 anv_gem_munmap(mem->map, mem->map_size);
1145 }
1146
1147 static void
1148 clflush_mapped_ranges(struct anv_device *device,
1149 uint32_t count,
1150 const VkMappedMemoryRange *ranges)
1151 {
1152 for (uint32_t i = 0; i < count; i++) {
1153 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1154 void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1155 void *end;
1156
1157 if (ranges[i].offset + ranges[i].size > mem->map_size)
1158 end = mem->map + mem->map_size;
1159 else
1160 end = mem->map + ranges[i].offset + ranges[i].size;
1161
1162 while (p < end) {
1163 __builtin_ia32_clflush(p);
1164 p += CACHELINE_SIZE;
1165 }
1166 }
1167 }
1168
1169 VkResult anv_FlushMappedMemoryRanges(
1170 VkDevice _device,
1171 uint32_t memoryRangeCount,
1172 const VkMappedMemoryRange* pMemoryRanges)
1173 {
1174 ANV_FROM_HANDLE(anv_device, device, _device);
1175
1176 if (device->info.has_llc)
1177 return VK_SUCCESS;
1178
1179 /* Make sure the writes we're flushing have landed. */
1180 __builtin_ia32_mfence();
1181
1182 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1183
1184 return VK_SUCCESS;
1185 }
1186
1187 VkResult anv_InvalidateMappedMemoryRanges(
1188 VkDevice _device,
1189 uint32_t memoryRangeCount,
1190 const VkMappedMemoryRange* pMemoryRanges)
1191 {
1192 ANV_FROM_HANDLE(anv_device, device, _device);
1193
1194 if (device->info.has_llc)
1195 return VK_SUCCESS;
1196
1197 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1198
1199 /* Make sure no reads get moved up above the invalidate. */
1200 __builtin_ia32_mfence();
1201
1202 return VK_SUCCESS;
1203 }
1204
1205 void anv_GetBufferMemoryRequirements(
1206 VkDevice device,
1207 VkBuffer _buffer,
1208 VkMemoryRequirements* pMemoryRequirements)
1209 {
1210 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1211
1212 /* The Vulkan spec (git aaed022) says:
1213 *
1214 * memoryTypeBits is a bitfield and contains one bit set for every
1215 * supported memory type for the resource. The bit `1<<i` is set if and
1216 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1217 * structure for the physical device is supported.
1218 *
1219 * We support exactly one memory type.
1220 */
1221 pMemoryRequirements->memoryTypeBits = 1;
1222
1223 pMemoryRequirements->size = buffer->size;
1224 pMemoryRequirements->alignment = 16;
1225 }
1226
1227 void anv_GetImageMemoryRequirements(
1228 VkDevice device,
1229 VkImage _image,
1230 VkMemoryRequirements* pMemoryRequirements)
1231 {
1232 ANV_FROM_HANDLE(anv_image, image, _image);
1233
1234 /* The Vulkan spec (git aaed022) says:
1235 *
1236 * memoryTypeBits is a bitfield and contains one bit set for every
1237 * supported memory type for the resource. The bit `1<<i` is set if and
1238 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1239 * structure for the physical device is supported.
1240 *
1241 * We support exactly one memory type.
1242 */
1243 pMemoryRequirements->memoryTypeBits = 1;
1244
1245 pMemoryRequirements->size = image->size;
1246 pMemoryRequirements->alignment = image->alignment;
1247 }
1248
1249 void anv_GetImageSparseMemoryRequirements(
1250 VkDevice device,
1251 VkImage image,
1252 uint32_t* pSparseMemoryRequirementCount,
1253 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
1254 {
1255 stub();
1256 }
1257
1258 void anv_GetDeviceMemoryCommitment(
1259 VkDevice device,
1260 VkDeviceMemory memory,
1261 VkDeviceSize* pCommittedMemoryInBytes)
1262 {
1263 *pCommittedMemoryInBytes = 0;
1264 }
1265
1266 VkResult anv_BindBufferMemory(
1267 VkDevice device,
1268 VkBuffer _buffer,
1269 VkDeviceMemory _memory,
1270 VkDeviceSize memoryOffset)
1271 {
1272 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1273 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1274
1275 if (mem) {
1276 buffer->bo = &mem->bo;
1277 buffer->offset = memoryOffset;
1278 } else {
1279 buffer->bo = NULL;
1280 buffer->offset = 0;
1281 }
1282
1283 return VK_SUCCESS;
1284 }
1285
1286 VkResult anv_BindImageMemory(
1287 VkDevice device,
1288 VkImage _image,
1289 VkDeviceMemory _memory,
1290 VkDeviceSize memoryOffset)
1291 {
1292 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1293 ANV_FROM_HANDLE(anv_image, image, _image);
1294
1295 if (mem) {
1296 image->bo = &mem->bo;
1297 image->offset = memoryOffset;
1298 } else {
1299 image->bo = NULL;
1300 image->offset = 0;
1301 }
1302
1303 return VK_SUCCESS;
1304 }
1305
1306 VkResult anv_QueueBindSparse(
1307 VkQueue queue,
1308 uint32_t bindInfoCount,
1309 const VkBindSparseInfo* pBindInfo,
1310 VkFence fence)
1311 {
1312 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1313 }
1314
1315 VkResult anv_CreateFence(
1316 VkDevice _device,
1317 const VkFenceCreateInfo* pCreateInfo,
1318 const VkAllocationCallbacks* pAllocator,
1319 VkFence* pFence)
1320 {
1321 ANV_FROM_HANDLE(anv_device, device, _device);
1322 struct anv_fence *fence;
1323 struct anv_batch batch;
1324 VkResult result;
1325
1326 const uint32_t fence_size = 128;
1327
1328 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1329
1330 fence = anv_alloc2(&device->alloc, pAllocator, sizeof(*fence), 8,
1331 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1332 if (fence == NULL)
1333 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1334
1335 result = anv_bo_init_new(&fence->bo, device, fence_size);
1336 if (result != VK_SUCCESS)
1337 goto fail;
1338
1339 fence->bo.map =
1340 anv_gem_mmap(device, fence->bo.gem_handle, 0, fence->bo.size, 0);
1341 batch.next = batch.start = fence->bo.map;
1342 batch.end = fence->bo.map + fence->bo.size;
1343 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END);
1344 anv_batch_emit(&batch, GEN7_MI_NOOP);
1345
1346 if (!device->info.has_llc) {
1347 assert(((uintptr_t) fence->bo.map & CACHELINE_MASK) == 0);
1348 assert(batch.next - fence->bo.map <= CACHELINE_SIZE);
1349 __builtin_ia32_mfence();
1350 __builtin_ia32_clflush(fence->bo.map);
1351 }
1352
1353 fence->exec2_objects[0].handle = fence->bo.gem_handle;
1354 fence->exec2_objects[0].relocation_count = 0;
1355 fence->exec2_objects[0].relocs_ptr = 0;
1356 fence->exec2_objects[0].alignment = 0;
1357 fence->exec2_objects[0].offset = fence->bo.offset;
1358 fence->exec2_objects[0].flags = 0;
1359 fence->exec2_objects[0].rsvd1 = 0;
1360 fence->exec2_objects[0].rsvd2 = 0;
1361
1362 fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1363 fence->execbuf.buffer_count = 1;
1364 fence->execbuf.batch_start_offset = 0;
1365 fence->execbuf.batch_len = batch.next - fence->bo.map;
1366 fence->execbuf.cliprects_ptr = 0;
1367 fence->execbuf.num_cliprects = 0;
1368 fence->execbuf.DR1 = 0;
1369 fence->execbuf.DR4 = 0;
1370
1371 fence->execbuf.flags =
1372 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1373 fence->execbuf.rsvd1 = device->context_id;
1374 fence->execbuf.rsvd2 = 0;
1375
1376 fence->ready = false;
1377
1378 *pFence = anv_fence_to_handle(fence);
1379
1380 return VK_SUCCESS;
1381
1382 fail:
1383 anv_free2(&device->alloc, pAllocator, fence);
1384
1385 return result;
1386 }
1387
1388 void anv_DestroyFence(
1389 VkDevice _device,
1390 VkFence _fence,
1391 const VkAllocationCallbacks* pAllocator)
1392 {
1393 ANV_FROM_HANDLE(anv_device, device, _device);
1394 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1395
1396 anv_gem_munmap(fence->bo.map, fence->bo.size);
1397 anv_gem_close(device, fence->bo.gem_handle);
1398 anv_free2(&device->alloc, pAllocator, fence);
1399 }
1400
1401 VkResult anv_ResetFences(
1402 VkDevice _device,
1403 uint32_t fenceCount,
1404 const VkFence* pFences)
1405 {
1406 for (uint32_t i = 0; i < fenceCount; i++) {
1407 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1408 fence->ready = false;
1409 }
1410
1411 return VK_SUCCESS;
1412 }
1413
1414 VkResult anv_GetFenceStatus(
1415 VkDevice _device,
1416 VkFence _fence)
1417 {
1418 ANV_FROM_HANDLE(anv_device, device, _device);
1419 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1420 int64_t t = 0;
1421 int ret;
1422
1423 if (fence->ready)
1424 return VK_SUCCESS;
1425
1426 ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1427 if (ret == 0) {
1428 fence->ready = true;
1429 return VK_SUCCESS;
1430 }
1431
1432 return VK_NOT_READY;
1433 }
1434
1435 VkResult anv_WaitForFences(
1436 VkDevice _device,
1437 uint32_t fenceCount,
1438 const VkFence* pFences,
1439 VkBool32 waitAll,
1440 uint64_t timeout)
1441 {
1442 ANV_FROM_HANDLE(anv_device, device, _device);
1443
1444 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1445 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1446 * for a couple of kernel releases. Since there's no way to know
1447 * whether or not the kernel we're using is one of the broken ones, the
1448 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1449 * maximum timeout from 584 years to 292 years - likely not a big deal.
1450 */
1451 if (timeout > INT64_MAX)
1452 timeout = INT64_MAX;
1453
1454 int64_t t = timeout;
1455
1456 /* FIXME: handle !waitAll */
1457
1458 for (uint32_t i = 0; i < fenceCount; i++) {
1459 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1460 int ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1461 if (ret == -1 && errno == ETIME) {
1462 return VK_TIMEOUT;
1463 } else if (ret == -1) {
1464 /* We don't know the real error. */
1465 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1466 "gem wait failed: %m");
1467 }
1468 }
1469
1470 return VK_SUCCESS;
1471 }
1472
1473 // Queue semaphore functions
1474
1475 VkResult anv_CreateSemaphore(
1476 VkDevice device,
1477 const VkSemaphoreCreateInfo* pCreateInfo,
1478 const VkAllocationCallbacks* pAllocator,
1479 VkSemaphore* pSemaphore)
1480 {
1481 /* The DRM execbuffer ioctl always execute in-oder, even between different
1482 * rings. As such, there's nothing to do for the user space semaphore.
1483 */
1484
1485 *pSemaphore = (VkSemaphore)1;
1486
1487 return VK_SUCCESS;
1488 }
1489
1490 void anv_DestroySemaphore(
1491 VkDevice device,
1492 VkSemaphore semaphore,
1493 const VkAllocationCallbacks* pAllocator)
1494 {
1495 }
1496
1497 // Event functions
1498
1499 VkResult anv_CreateEvent(
1500 VkDevice _device,
1501 const VkEventCreateInfo* pCreateInfo,
1502 const VkAllocationCallbacks* pAllocator,
1503 VkEvent* pEvent)
1504 {
1505 ANV_FROM_HANDLE(anv_device, device, _device);
1506 struct anv_state state;
1507 struct anv_event *event;
1508
1509 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1510
1511 state = anv_state_pool_alloc(&device->dynamic_state_pool,
1512 sizeof(*event), 4);
1513 event = state.map;
1514 event->state = state;
1515 event->semaphore = VK_EVENT_RESET;
1516
1517 if (!device->info.has_llc) {
1518 /* Make sure the writes we're flushing have landed. */
1519 __builtin_ia32_mfence();
1520 __builtin_ia32_clflush(event);
1521 }
1522
1523 *pEvent = anv_event_to_handle(event);
1524
1525 return VK_SUCCESS;
1526 }
1527
1528 void anv_DestroyEvent(
1529 VkDevice _device,
1530 VkEvent _event,
1531 const VkAllocationCallbacks* pAllocator)
1532 {
1533 ANV_FROM_HANDLE(anv_device, device, _device);
1534 ANV_FROM_HANDLE(anv_event, event, _event);
1535
1536 anv_state_pool_free(&device->dynamic_state_pool, event->state);
1537 }
1538
1539 VkResult anv_GetEventStatus(
1540 VkDevice _device,
1541 VkEvent _event)
1542 {
1543 ANV_FROM_HANDLE(anv_device, device, _device);
1544 ANV_FROM_HANDLE(anv_event, event, _event);
1545
1546 if (!device->info.has_llc) {
1547 /* Invalidate read cache before reading event written by GPU. */
1548 __builtin_ia32_clflush(event);
1549 __builtin_ia32_mfence();
1550
1551 }
1552
1553 return event->semaphore;
1554 }
1555
1556 VkResult anv_SetEvent(
1557 VkDevice _device,
1558 VkEvent _event)
1559 {
1560 ANV_FROM_HANDLE(anv_device, device, _device);
1561 ANV_FROM_HANDLE(anv_event, event, _event);
1562
1563 event->semaphore = VK_EVENT_SET;
1564
1565 if (!device->info.has_llc) {
1566 /* Make sure the writes we're flushing have landed. */
1567 __builtin_ia32_mfence();
1568 __builtin_ia32_clflush(event);
1569 }
1570
1571 return VK_SUCCESS;
1572 }
1573
1574 VkResult anv_ResetEvent(
1575 VkDevice _device,
1576 VkEvent _event)
1577 {
1578 ANV_FROM_HANDLE(anv_device, device, _device);
1579 ANV_FROM_HANDLE(anv_event, event, _event);
1580
1581 event->semaphore = VK_EVENT_RESET;
1582
1583 if (!device->info.has_llc) {
1584 /* Make sure the writes we're flushing have landed. */
1585 __builtin_ia32_mfence();
1586 __builtin_ia32_clflush(event);
1587 }
1588
1589 return VK_SUCCESS;
1590 }
1591
1592 // Buffer functions
1593
1594 VkResult anv_CreateBuffer(
1595 VkDevice _device,
1596 const VkBufferCreateInfo* pCreateInfo,
1597 const VkAllocationCallbacks* pAllocator,
1598 VkBuffer* pBuffer)
1599 {
1600 ANV_FROM_HANDLE(anv_device, device, _device);
1601 struct anv_buffer *buffer;
1602
1603 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1604
1605 buffer = anv_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1606 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1607 if (buffer == NULL)
1608 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1609
1610 buffer->size = pCreateInfo->size;
1611 buffer->usage = pCreateInfo->usage;
1612 buffer->bo = NULL;
1613 buffer->offset = 0;
1614
1615 *pBuffer = anv_buffer_to_handle(buffer);
1616
1617 return VK_SUCCESS;
1618 }
1619
1620 void anv_DestroyBuffer(
1621 VkDevice _device,
1622 VkBuffer _buffer,
1623 const VkAllocationCallbacks* pAllocator)
1624 {
1625 ANV_FROM_HANDLE(anv_device, device, _device);
1626 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1627
1628 anv_free2(&device->alloc, pAllocator, buffer);
1629 }
1630
1631 void
1632 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1633 enum isl_format format,
1634 uint32_t offset, uint32_t range, uint32_t stride)
1635 {
1636 switch (device->info.gen) {
1637 case 7:
1638 if (device->info.is_haswell)
1639 gen75_fill_buffer_surface_state(state.map, format, offset, range,
1640 stride);
1641 else
1642 gen7_fill_buffer_surface_state(state.map, format, offset, range,
1643 stride);
1644 break;
1645 case 8:
1646 gen8_fill_buffer_surface_state(state.map, format, offset, range, stride);
1647 break;
1648 case 9:
1649 gen9_fill_buffer_surface_state(state.map, format, offset, range, stride);
1650 break;
1651 default:
1652 unreachable("unsupported gen\n");
1653 }
1654
1655 if (!device->info.has_llc)
1656 anv_state_clflush(state);
1657 }
1658
1659 void anv_DestroySampler(
1660 VkDevice _device,
1661 VkSampler _sampler,
1662 const VkAllocationCallbacks* pAllocator)
1663 {
1664 ANV_FROM_HANDLE(anv_device, device, _device);
1665 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
1666
1667 anv_free2(&device->alloc, pAllocator, sampler);
1668 }
1669
1670 VkResult anv_CreateFramebuffer(
1671 VkDevice _device,
1672 const VkFramebufferCreateInfo* pCreateInfo,
1673 const VkAllocationCallbacks* pAllocator,
1674 VkFramebuffer* pFramebuffer)
1675 {
1676 ANV_FROM_HANDLE(anv_device, device, _device);
1677 struct anv_framebuffer *framebuffer;
1678
1679 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1680
1681 size_t size = sizeof(*framebuffer) +
1682 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
1683 framebuffer = anv_alloc2(&device->alloc, pAllocator, size, 8,
1684 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1685 if (framebuffer == NULL)
1686 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1687
1688 framebuffer->attachment_count = pCreateInfo->attachmentCount;
1689 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1690 VkImageView _iview = pCreateInfo->pAttachments[i];
1691 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
1692 }
1693
1694 framebuffer->width = pCreateInfo->width;
1695 framebuffer->height = pCreateInfo->height;
1696 framebuffer->layers = pCreateInfo->layers;
1697
1698 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
1699
1700 return VK_SUCCESS;
1701 }
1702
1703 void anv_DestroyFramebuffer(
1704 VkDevice _device,
1705 VkFramebuffer _fb,
1706 const VkAllocationCallbacks* pAllocator)
1707 {
1708 ANV_FROM_HANDLE(anv_device, device, _device);
1709 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
1710
1711 anv_free2(&device->alloc, pAllocator, fb);
1712 }
1713
1714 void vkCmdDbgMarkerBegin(
1715 VkCommandBuffer commandBuffer,
1716 const char* pMarker)
1717 __attribute__ ((visibility ("default")));
1718
1719 void vkCmdDbgMarkerEnd(
1720 VkCommandBuffer commandBuffer)
1721 __attribute__ ((visibility ("default")));
1722
1723 void vkCmdDbgMarkerBegin(
1724 VkCommandBuffer commandBuffer,
1725 const char* pMarker)
1726 {
1727 }
1728
1729 void vkCmdDbgMarkerEnd(
1730 VkCommandBuffer commandBuffer)
1731 {
1732 }