projects / mesa.git / blob
? search:


f181eb78ddf3fc1f1a7897792ceb43cb3db0b608
[mesa.git] / src / intel / 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 "anv_timestamp.h"
32 #include "util/strtod.h"
33 #include "util/debug.h"
34
35 #include "genxml/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
71 assert(strlen(path) < ARRAY_SIZE(device->path));
72 strncpy(device->path, path, ARRAY_SIZE(device->path));
73
74 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
75 if (!device->chipset_id) {
76 result = VK_ERROR_INITIALIZATION_FAILED;
77 goto fail;
78 }
79
80 device->name = brw_get_device_name(device->chipset_id);
81 device->info = brw_get_device_info(device->chipset_id);
82 if (!device->info) {
83 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
84 "failed to get device info");
85 goto fail;
86 }
87
88 if (device->info->is_haswell) {
89 fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
90 } else if (device->info->gen == 7 && !device->info->is_baytrail) {
91 fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
92 } else if (device->info->gen == 7 && device->info->is_baytrail) {
93 fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
94 } else if (device->info->gen >= 8) {
95 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
96 * supported as anything */
97 } else {
98 result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
99 "Vulkan not yet supported on %s", device->name);
100 goto fail;
101 }
102
103 device->cmd_parser_version = -1;
104 if (device->info->gen == 7) {
105 device->cmd_parser_version =
106 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
107 if (device->cmd_parser_version == -1) {
108 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
109 "failed to get command parser version");
110 goto fail;
111 }
112 }
113
114 if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
115 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
116 "failed to get aperture size: %m");
117 goto fail;
118 }
119
120 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
121 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
122 "kernel missing gem wait");
123 goto fail;
124 }
125
126 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
127 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
128 "kernel missing execbuf2");
129 goto fail;
130 }
131
132 if (!device->info->has_llc &&
133 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
134 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
135 "kernel missing wc mmap");
136 goto fail;
137 }
138
139 bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
140
141 close(fd);
142
143 brw_process_intel_debug_variable();
144
145 device->compiler = brw_compiler_create(NULL, device->info);
146 if (device->compiler == NULL) {
147 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
148 goto fail;
149 }
150 device->compiler->shader_debug_log = compiler_debug_log;
151 device->compiler->shader_perf_log = compiler_perf_log;
152
153 result = anv_init_wsi(device);
154 if (result != VK_SUCCESS)
155 goto fail;
156
157 /* XXX: Actually detect bit6 swizzling */
158 isl_device_init(&device->isl_dev, device->info, swizzled);
159
160 return VK_SUCCESS;
161
162 fail:
163 close(fd);
164 return result;
165 }
166
167 static void
168 anv_physical_device_finish(struct anv_physical_device *device)
169 {
170 anv_finish_wsi(device);
171 ralloc_free(device->compiler);
172 }
173
174 static const VkExtensionProperties global_extensions[] = {
175 {
176 .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
177 .specVersion = 25,
178 },
179 #ifdef VK_USE_PLATFORM_XCB_KHR
180 {
181 .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
182 .specVersion = 5,
183 },
184 #endif
185 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
186 {
187 .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
188 .specVersion = 4,
189 },
190 #endif
191 };
192
193 static const VkExtensionProperties device_extensions[] = {
194 {
195 .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
196 .specVersion = 67,
197 },
198 };
199
200 static void *
201 default_alloc_func(void *pUserData, size_t size, size_t align,
202 VkSystemAllocationScope allocationScope)
203 {
204 return malloc(size);
205 }
206
207 static void *
208 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
209 size_t align, VkSystemAllocationScope allocationScope)
210 {
211 return realloc(pOriginal, size);
212 }
213
214 static void
215 default_free_func(void *pUserData, void *pMemory)
216 {
217 free(pMemory);
218 }
219
220 static const VkAllocationCallbacks default_alloc = {
221 .pUserData = NULL,
222 .pfnAllocation = default_alloc_func,
223 .pfnReallocation = default_realloc_func,
224 .pfnFree = default_free_func,
225 };
226
227 VkResult anv_CreateInstance(
228 const VkInstanceCreateInfo* pCreateInfo,
229 const VkAllocationCallbacks* pAllocator,
230 VkInstance* pInstance)
231 {
232 struct anv_instance *instance;
233
234 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
235
236 uint32_t client_version;
237 if (pCreateInfo->pApplicationInfo &&
238 pCreateInfo->pApplicationInfo->apiVersion != 0) {
239 client_version = pCreateInfo->pApplicationInfo->apiVersion;
240 } else {
241 client_version = VK_MAKE_VERSION(1, 0, 0);
242 }
243
244 if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
245 client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
246 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
247 "Client requested version %d.%d.%d",
248 VK_VERSION_MAJOR(client_version),
249 VK_VERSION_MINOR(client_version),
250 VK_VERSION_PATCH(client_version));
251 }
252
253 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
254 bool found = false;
255 for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
256 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
257 global_extensions[j].extensionName) == 0) {
258 found = true;
259 break;
260 }
261 }
262 if (!found)
263 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
264 }
265
266 instance = anv_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
267 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
268 if (!instance)
269 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
270
271 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
272
273 if (pAllocator)
274 instance->alloc = *pAllocator;
275 else
276 instance->alloc = default_alloc;
277
278 instance->apiVersion = client_version;
279 instance->physicalDeviceCount = -1;
280
281 _mesa_locale_init();
282
283 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
284
285 *pInstance = anv_instance_to_handle(instance);
286
287 return VK_SUCCESS;
288 }
289
290 void anv_DestroyInstance(
291 VkInstance _instance,
292 const VkAllocationCallbacks* pAllocator)
293 {
294 ANV_FROM_HANDLE(anv_instance, instance, _instance);
295
296 if (instance->physicalDeviceCount > 0) {
297 /* We support at most one physical device. */
298 assert(instance->physicalDeviceCount == 1);
299 anv_physical_device_finish(&instance->physicalDevice);
300 }
301
302 VG(VALGRIND_DESTROY_MEMPOOL(instance));
303
304 _mesa_locale_fini();
305
306 anv_free(&instance->alloc, instance);
307 }
308
309 VkResult anv_EnumeratePhysicalDevices(
310 VkInstance _instance,
311 uint32_t* pPhysicalDeviceCount,
312 VkPhysicalDevice* pPhysicalDevices)
313 {
314 ANV_FROM_HANDLE(anv_instance, instance, _instance);
315 VkResult result;
316
317 if (instance->physicalDeviceCount < 0) {
318 char path[20];
319 for (unsigned i = 0; i < 8; i++) {
320 snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
321 result = anv_physical_device_init(&instance->physicalDevice,
322 instance, path);
323 if (result == VK_SUCCESS)
324 break;
325 }
326
327 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
328 instance->physicalDeviceCount = 0;
329 } else if (result == VK_SUCCESS) {
330 instance->physicalDeviceCount = 1;
331 } else {
332 return result;
333 }
334 }
335
336 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
337 * otherwise it's an inout parameter.
338 *
339 * The Vulkan spec (git aaed022) says:
340 *
341 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
342 * that is initialized with the number of devices the application is
343 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
344 * an array of at least this many VkPhysicalDevice handles [...].
345 *
346 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
347 * overwrites the contents of the variable pointed to by
348 * pPhysicalDeviceCount with the number of physical devices in in the
349 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
350 * pPhysicalDeviceCount with the number of physical handles written to
351 * pPhysicalDevices.
352 */
353 if (!pPhysicalDevices) {
354 *pPhysicalDeviceCount = instance->physicalDeviceCount;
355 } else if (*pPhysicalDeviceCount >= 1) {
356 pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
357 *pPhysicalDeviceCount = 1;
358 } else {
359 *pPhysicalDeviceCount = 0;
360 }
361
362 return VK_SUCCESS;
363 }
364
365 void anv_GetPhysicalDeviceFeatures(
366 VkPhysicalDevice physicalDevice,
367 VkPhysicalDeviceFeatures* pFeatures)
368 {
369 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
370
371 *pFeatures = (VkPhysicalDeviceFeatures) {
372 .robustBufferAccess = true,
373 .fullDrawIndexUint32 = true,
374 .imageCubeArray = false,
375 .independentBlend = pdevice->info->gen >= 8,
376 .geometryShader = true,
377 .tessellationShader = false,
378 .sampleRateShading = false,
379 .dualSrcBlend = true,
380 .logicOp = true,
381 .multiDrawIndirect = false,
382 .drawIndirectFirstInstance = false,
383 .depthClamp = true,
384 .depthBiasClamp = false,
385 .fillModeNonSolid = true,
386 .depthBounds = false,
387 .wideLines = true,
388 .largePoints = true,
389 .alphaToOne = true,
390 .multiViewport = true,
391 .samplerAnisotropy = false, /* FINISHME */
392 .textureCompressionETC2 = pdevice->info->gen >= 8 ||
393 pdevice->info->is_baytrail,
394 .textureCompressionASTC_LDR = pdevice->info->gen >= 9, /* FINISHME CHV */
395 .textureCompressionBC = true,
396 .occlusionQueryPrecise = true,
397 .pipelineStatisticsQuery = false,
398 .fragmentStoresAndAtomics = true,
399 .shaderTessellationAndGeometryPointSize = true,
400 .shaderImageGatherExtended = false,
401 .shaderStorageImageExtendedFormats = false,
402 .shaderStorageImageMultisample = false,
403 .shaderUniformBufferArrayDynamicIndexing = true,
404 .shaderSampledImageArrayDynamicIndexing = true,
405 .shaderStorageBufferArrayDynamicIndexing = true,
406 .shaderStorageImageArrayDynamicIndexing = true,
407 .shaderStorageImageReadWithoutFormat = false,
408 .shaderStorageImageWriteWithoutFormat = true,
409 .shaderClipDistance = false,
410 .shaderCullDistance = false,
411 .shaderFloat64 = false,
412 .shaderInt64 = false,
413 .shaderInt16 = false,
414 .alphaToOne = true,
415 .variableMultisampleRate = false,
416 .inheritedQueries = false,
417 };
418
419 /* We can't do image stores in vec4 shaders */
420 pFeatures->vertexPipelineStoresAndAtomics =
421 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
422 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
423 }
424
425 void
426 anv_device_get_cache_uuid(void *uuid)
427 {
428 memset(uuid, 0, VK_UUID_SIZE);
429 snprintf(uuid, VK_UUID_SIZE, "anv-%s", ANV_TIMESTAMP);
430 }
431
432 void anv_GetPhysicalDeviceProperties(
433 VkPhysicalDevice physicalDevice,
434 VkPhysicalDeviceProperties* pProperties)
435 {
436 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
437 const struct brw_device_info *devinfo = pdevice->info;
438
439 const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
440
441 /* See assertions made when programming the buffer surface state. */
442 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
443 (1ul << 30) : (1ul << 27);
444
445 VkSampleCountFlags sample_counts =
446 isl_device_get_sample_counts(&pdevice->isl_dev);
447
448 VkPhysicalDeviceLimits limits = {
449 .maxImageDimension1D = (1 << 14),
450 .maxImageDimension2D = (1 << 14),
451 .maxImageDimension3D = (1 << 11),
452 .maxImageDimensionCube = (1 << 14),
453 .maxImageArrayLayers = (1 << 11),
454 .maxTexelBufferElements = 128 * 1024 * 1024,
455 .maxUniformBufferRange = (1ul << 27),
456 .maxStorageBufferRange = max_raw_buffer_sz,
457 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
458 .maxMemoryAllocationCount = UINT32_MAX,
459 .maxSamplerAllocationCount = 64 * 1024,
460 .bufferImageGranularity = 64, /* A cache line */
461 .sparseAddressSpaceSize = 0,
462 .maxBoundDescriptorSets = MAX_SETS,
463 .maxPerStageDescriptorSamplers = 64,
464 .maxPerStageDescriptorUniformBuffers = 64,
465 .maxPerStageDescriptorStorageBuffers = 64,
466 .maxPerStageDescriptorSampledImages = 64,
467 .maxPerStageDescriptorStorageImages = 64,
468 .maxPerStageDescriptorInputAttachments = 64,
469 .maxPerStageResources = 128,
470 .maxDescriptorSetSamplers = 256,
471 .maxDescriptorSetUniformBuffers = 256,
472 .maxDescriptorSetUniformBuffersDynamic = 256,
473 .maxDescriptorSetStorageBuffers = 256,
474 .maxDescriptorSetStorageBuffersDynamic = 256,
475 .maxDescriptorSetSampledImages = 256,
476 .maxDescriptorSetStorageImages = 256,
477 .maxDescriptorSetInputAttachments = 256,
478 .maxVertexInputAttributes = 32,
479 .maxVertexInputBindings = 32,
480 .maxVertexInputAttributeOffset = 2047,
481 .maxVertexInputBindingStride = 2048,
482 .maxVertexOutputComponents = 128,
483 .maxTessellationGenerationLevel = 0,
484 .maxTessellationPatchSize = 0,
485 .maxTessellationControlPerVertexInputComponents = 0,
486 .maxTessellationControlPerVertexOutputComponents = 0,
487 .maxTessellationControlPerPatchOutputComponents = 0,
488 .maxTessellationControlTotalOutputComponents = 0,
489 .maxTessellationEvaluationInputComponents = 0,
490 .maxTessellationEvaluationOutputComponents = 0,
491 .maxGeometryShaderInvocations = 32,
492 .maxGeometryInputComponents = 64,
493 .maxGeometryOutputComponents = 128,
494 .maxGeometryOutputVertices = 256,
495 .maxGeometryTotalOutputComponents = 1024,
496 .maxFragmentInputComponents = 128,
497 .maxFragmentOutputAttachments = 8,
498 .maxFragmentDualSrcAttachments = 2,
499 .maxFragmentCombinedOutputResources = 8,
500 .maxComputeSharedMemorySize = 32768,
501 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
502 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
503 .maxComputeWorkGroupSize = {
504 16 * devinfo->max_cs_threads,
505 16 * devinfo->max_cs_threads,
506 16 * devinfo->max_cs_threads,
507 },
508 .subPixelPrecisionBits = 4 /* FIXME */,
509 .subTexelPrecisionBits = 4 /* FIXME */,
510 .mipmapPrecisionBits = 4 /* FIXME */,
511 .maxDrawIndexedIndexValue = UINT32_MAX,
512 .maxDrawIndirectCount = UINT32_MAX,
513 .maxSamplerLodBias = 16,
514 .maxSamplerAnisotropy = 16,
515 .maxViewports = MAX_VIEWPORTS,
516 .maxViewportDimensions = { (1 << 14), (1 << 14) },
517 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
518 .viewportSubPixelBits = 13, /* We take a float? */
519 .minMemoryMapAlignment = 4096, /* A page */
520 .minTexelBufferOffsetAlignment = 1,
521 .minUniformBufferOffsetAlignment = 1,
522 .minStorageBufferOffsetAlignment = 1,
523 .minTexelOffset = -8,
524 .maxTexelOffset = 7,
525 .minTexelGatherOffset = -8,
526 .maxTexelGatherOffset = 7,
527 .minInterpolationOffset = 0, /* FIXME */
528 .maxInterpolationOffset = 0, /* FIXME */
529 .subPixelInterpolationOffsetBits = 0, /* FIXME */
530 .maxFramebufferWidth = (1 << 14),
531 .maxFramebufferHeight = (1 << 14),
532 .maxFramebufferLayers = (1 << 10),
533 .framebufferColorSampleCounts = sample_counts,
534 .framebufferDepthSampleCounts = sample_counts,
535 .framebufferStencilSampleCounts = sample_counts,
536 .framebufferNoAttachmentsSampleCounts = sample_counts,
537 .maxColorAttachments = MAX_RTS,
538 .sampledImageColorSampleCounts = sample_counts,
539 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
540 .sampledImageDepthSampleCounts = sample_counts,
541 .sampledImageStencilSampleCounts = sample_counts,
542 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
543 .maxSampleMaskWords = 1,
544 .timestampComputeAndGraphics = false,
545 .timestampPeriod = time_stamp_base / (1000 * 1000 * 1000),
546 .maxClipDistances = 0 /* FIXME */,
547 .maxCullDistances = 0 /* FIXME */,
548 .maxCombinedClipAndCullDistances = 0 /* FIXME */,
549 .discreteQueuePriorities = 1,
550 .pointSizeRange = { 0.125, 255.875 },
551 .lineWidthRange = { 0.0, 7.9921875 },
552 .pointSizeGranularity = (1.0 / 8.0),
553 .lineWidthGranularity = (1.0 / 128.0),
554 .strictLines = false, /* FINISHME */
555 .standardSampleLocations = true,
556 .optimalBufferCopyOffsetAlignment = 128,
557 .optimalBufferCopyRowPitchAlignment = 128,
558 .nonCoherentAtomSize = 64,
559 };
560
561 *pProperties = (VkPhysicalDeviceProperties) {
562 .apiVersion = VK_MAKE_VERSION(1, 0, 5),
563 .driverVersion = 1,
564 .vendorID = 0x8086,
565 .deviceID = pdevice->chipset_id,
566 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
567 .limits = limits,
568 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
569 };
570
571 strcpy(pProperties->deviceName, pdevice->name);
572 anv_device_get_cache_uuid(pProperties->pipelineCacheUUID);
573 }
574
575 void anv_GetPhysicalDeviceQueueFamilyProperties(
576 VkPhysicalDevice physicalDevice,
577 uint32_t* pCount,
578 VkQueueFamilyProperties* pQueueFamilyProperties)
579 {
580 if (pQueueFamilyProperties == NULL) {
581 *pCount = 1;
582 return;
583 }
584
585 assert(*pCount >= 1);
586
587 *pQueueFamilyProperties = (VkQueueFamilyProperties) {
588 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
589 VK_QUEUE_COMPUTE_BIT |
590 VK_QUEUE_TRANSFER_BIT,
591 .queueCount = 1,
592 .timestampValidBits = 36, /* XXX: Real value here */
593 .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
594 };
595 }
596
597 void anv_GetPhysicalDeviceMemoryProperties(
598 VkPhysicalDevice physicalDevice,
599 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
600 {
601 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
602 VkDeviceSize heap_size;
603
604 /* Reserve some wiggle room for the driver by exposing only 75% of the
605 * aperture to the heap.
606 */
607 heap_size = 3 * physical_device->aperture_size / 4;
608
609 if (physical_device->info->has_llc) {
610 /* Big core GPUs share LLC with the CPU and thus one memory type can be
611 * both cached and coherent at the same time.
612 */
613 pMemoryProperties->memoryTypeCount = 1;
614 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
615 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
616 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
617 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
618 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
619 .heapIndex = 0,
620 };
621 } else {
622 /* The spec requires that we expose a host-visible, coherent memory
623 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
624 * to give the application a choice between cached, but not coherent and
625 * coherent but uncached (WC though).
626 */
627 pMemoryProperties->memoryTypeCount = 2;
628 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
629 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
630 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
631 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
632 .heapIndex = 0,
633 };
634 pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
635 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
636 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
637 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
638 .heapIndex = 0,
639 };
640 }
641
642 pMemoryProperties->memoryHeapCount = 1;
643 pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
644 .size = heap_size,
645 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
646 };
647 }
648
649 PFN_vkVoidFunction anv_GetInstanceProcAddr(
650 VkInstance instance,
651 const char* pName)
652 {
653 return anv_lookup_entrypoint(pName);
654 }
655
656 /* The loader wants us to expose a second GetInstanceProcAddr function
657 * to work around certain LD_PRELOAD issues seen in apps.
658 */
659 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
660 VkInstance instance,
661 const char* pName);
662
663 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
664 VkInstance instance,
665 const char* pName)
666 {
667 return anv_GetInstanceProcAddr(instance, pName);
668 }
669
670 PFN_vkVoidFunction anv_GetDeviceProcAddr(
671 VkDevice device,
672 const char* pName)
673 {
674 return anv_lookup_entrypoint(pName);
675 }
676
677 static VkResult
678 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
679 {
680 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
681 queue->device = device;
682 queue->pool = &device->surface_state_pool;
683
684 return VK_SUCCESS;
685 }
686
687 static void
688 anv_queue_finish(struct anv_queue *queue)
689 {
690 }
691
692 static struct anv_state
693 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
694 {
695 struct anv_state state;
696
697 state = anv_state_pool_alloc(pool, size, align);
698 memcpy(state.map, p, size);
699
700 if (!pool->block_pool->device->info.has_llc)
701 anv_state_clflush(state);
702
703 return state;
704 }
705
706 struct gen8_border_color {
707 union {
708 float float32[4];
709 uint32_t uint32[4];
710 };
711 /* Pad out to 64 bytes */
712 uint32_t _pad[12];
713 };
714
715 static void
716 anv_device_init_border_colors(struct anv_device *device)
717 {
718 static const struct gen8_border_color border_colors[] = {
719 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
720 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
721 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
722 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
723 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
724 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
725 };
726
727 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
728 sizeof(border_colors), 64,
729 border_colors);
730 }
731
732 VkResult
733 anv_device_submit_simple_batch(struct anv_device *device,
734 struct anv_batch *batch)
735 {
736 struct drm_i915_gem_execbuffer2 execbuf;
737 struct drm_i915_gem_exec_object2 exec2_objects[1];
738 struct anv_bo bo;
739 VkResult result = VK_SUCCESS;
740 uint32_t size;
741 int64_t timeout;
742 int ret;
743
744 /* Kernel driver requires 8 byte aligned batch length */
745 size = align_u32(batch->next - batch->start, 8);
746 result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
747 if (result != VK_SUCCESS)
748 return result;
749
750 memcpy(bo.map, batch->start, size);
751 if (!device->info.has_llc)
752 anv_clflush_range(bo.map, size);
753
754 exec2_objects[0].handle = bo.gem_handle;
755 exec2_objects[0].relocation_count = 0;
756 exec2_objects[0].relocs_ptr = 0;
757 exec2_objects[0].alignment = 0;
758 exec2_objects[0].offset = bo.offset;
759 exec2_objects[0].flags = 0;
760 exec2_objects[0].rsvd1 = 0;
761 exec2_objects[0].rsvd2 = 0;
762
763 execbuf.buffers_ptr = (uintptr_t) exec2_objects;
764 execbuf.buffer_count = 1;
765 execbuf.batch_start_offset = 0;
766 execbuf.batch_len = size;
767 execbuf.cliprects_ptr = 0;
768 execbuf.num_cliprects = 0;
769 execbuf.DR1 = 0;
770 execbuf.DR4 = 0;
771
772 execbuf.flags =
773 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
774 execbuf.rsvd1 = device->context_id;
775 execbuf.rsvd2 = 0;
776
777 ret = anv_gem_execbuffer(device, &execbuf);
778 if (ret != 0) {
779 /* We don't know the real error. */
780 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
781 goto fail;
782 }
783
784 timeout = INT64_MAX;
785 ret = anv_gem_wait(device, bo.gem_handle, &timeout);
786 if (ret != 0) {
787 /* We don't know the real error. */
788 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
789 goto fail;
790 }
791
792 fail:
793 anv_bo_pool_free(&device->batch_bo_pool, &bo);
794
795 return result;
796 }
797
798 VkResult anv_CreateDevice(
799 VkPhysicalDevice physicalDevice,
800 const VkDeviceCreateInfo* pCreateInfo,
801 const VkAllocationCallbacks* pAllocator,
802 VkDevice* pDevice)
803 {
804 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
805 VkResult result;
806 struct anv_device *device;
807
808 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
809
810 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
811 bool found = false;
812 for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
813 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
814 device_extensions[j].extensionName) == 0) {
815 found = true;
816 break;
817 }
818 }
819 if (!found)
820 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
821 }
822
823 anv_set_dispatch_devinfo(physical_device->info);
824
825 device = anv_alloc2(&physical_device->instance->alloc, pAllocator,
826 sizeof(*device), 8,
827 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
828 if (!device)
829 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
830
831 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
832 device->instance = physical_device->instance;
833 device->chipset_id = physical_device->chipset_id;
834
835 if (pAllocator)
836 device->alloc = *pAllocator;
837 else
838 device->alloc = physical_device->instance->alloc;
839
840 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
841 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
842 if (device->fd == -1) {
843 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
844 goto fail_device;
845 }
846
847 device->context_id = anv_gem_create_context(device);
848 if (device->context_id == -1) {
849 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
850 goto fail_fd;
851 }
852
853 device->info = *physical_device->info;
854 device->isl_dev = physical_device->isl_dev;
855
856 /* On Broadwell and later, we can use batch chaining to more efficiently
857 * implement growing command buffers. Prior to Haswell, the kernel
858 * command parser gets in the way and we have to fall back to growing
859 * the batch.
860 */
861 device->can_chain_batches = device->info.gen >= 8;
862
863 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
864 pCreateInfo->pEnabledFeatures->robustBufferAccess;
865
866 pthread_mutex_init(&device->mutex, NULL);
867
868 anv_bo_pool_init(&device->batch_bo_pool, device);
869
870 anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
871
872 anv_state_pool_init(&device->dynamic_state_pool,
873 &device->dynamic_state_block_pool);
874
875 anv_block_pool_init(&device->instruction_block_pool, device, 128 * 1024);
876 anv_pipeline_cache_init(&device->default_pipeline_cache, device);
877
878 anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
879
880 anv_state_pool_init(&device->surface_state_pool,
881 &device->surface_state_block_pool);
882
883 anv_bo_init_new(&device->workaround_bo, device, 1024);
884
885 anv_scratch_pool_init(device, &device->scratch_pool);
886
887 anv_queue_init(device, &device->queue);
888
889 switch (device->info.gen) {
890 case 7:
891 if (!device->info.is_haswell)
892 result = gen7_init_device_state(device);
893 else
894 result = gen75_init_device_state(device);
895 break;
896 case 8:
897 result = gen8_init_device_state(device);
898 break;
899 case 9:
900 result = gen9_init_device_state(device);
901 break;
902 default:
903 /* Shouldn't get here as we don't create physical devices for any other
904 * gens. */
905 unreachable("unhandled gen");
906 }
907 if (result != VK_SUCCESS)
908 goto fail_fd;
909
910 result = anv_device_init_meta(device);
911 if (result != VK_SUCCESS)
912 goto fail_fd;
913
914 anv_device_init_border_colors(device);
915
916 *pDevice = anv_device_to_handle(device);
917
918 return VK_SUCCESS;
919
920 fail_fd:
921 close(device->fd);
922 fail_device:
923 anv_free(&device->alloc, device);
924
925 return result;
926 }
927
928 void anv_DestroyDevice(
929 VkDevice _device,
930 const VkAllocationCallbacks* pAllocator)
931 {
932 ANV_FROM_HANDLE(anv_device, device, _device);
933
934 anv_queue_finish(&device->queue);
935
936 anv_device_finish_meta(device);
937
938 #ifdef HAVE_VALGRIND
939 /* We only need to free these to prevent valgrind errors. The backing
940 * BO will go away in a couple of lines so we don't actually leak.
941 */
942 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
943 #endif
944
945 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
946 anv_gem_close(device, device->workaround_bo.gem_handle);
947
948 anv_bo_pool_finish(&device->batch_bo_pool);
949 anv_state_pool_finish(&device->dynamic_state_pool);
950 anv_block_pool_finish(&device->dynamic_state_block_pool);
951 anv_block_pool_finish(&device->instruction_block_pool);
952 anv_state_pool_finish(&device->surface_state_pool);
953 anv_block_pool_finish(&device->surface_state_block_pool);
954 anv_scratch_pool_finish(device, &device->scratch_pool);
955
956 close(device->fd);
957
958 pthread_mutex_destroy(&device->mutex);
959
960 anv_free(&device->alloc, device);
961 }
962
963 VkResult anv_EnumerateInstanceExtensionProperties(
964 const char* pLayerName,
965 uint32_t* pPropertyCount,
966 VkExtensionProperties* pProperties)
967 {
968 if (pProperties == NULL) {
969 *pPropertyCount = ARRAY_SIZE(global_extensions);
970 return VK_SUCCESS;
971 }
972
973 assert(*pPropertyCount >= ARRAY_SIZE(global_extensions));
974
975 *pPropertyCount = ARRAY_SIZE(global_extensions);
976 memcpy(pProperties, global_extensions, sizeof(global_extensions));
977
978 return VK_SUCCESS;
979 }
980
981 VkResult anv_EnumerateDeviceExtensionProperties(
982 VkPhysicalDevice physicalDevice,
983 const char* pLayerName,
984 uint32_t* pPropertyCount,
985 VkExtensionProperties* pProperties)
986 {
987 if (pProperties == NULL) {
988 *pPropertyCount = ARRAY_SIZE(device_extensions);
989 return VK_SUCCESS;
990 }
991
992 assert(*pPropertyCount >= ARRAY_SIZE(device_extensions));
993
994 *pPropertyCount = ARRAY_SIZE(device_extensions);
995 memcpy(pProperties, device_extensions, sizeof(device_extensions));
996
997 return VK_SUCCESS;
998 }
999
1000 VkResult anv_EnumerateInstanceLayerProperties(
1001 uint32_t* pPropertyCount,
1002 VkLayerProperties* pProperties)
1003 {
1004 if (pProperties == NULL) {
1005 *pPropertyCount = 0;
1006 return VK_SUCCESS;
1007 }
1008
1009 /* None supported at this time */
1010 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1011 }
1012
1013 VkResult anv_EnumerateDeviceLayerProperties(
1014 VkPhysicalDevice physicalDevice,
1015 uint32_t* pPropertyCount,
1016 VkLayerProperties* pProperties)
1017 {
1018 if (pProperties == NULL) {
1019 *pPropertyCount = 0;
1020 return VK_SUCCESS;
1021 }
1022
1023 /* None supported at this time */
1024 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1025 }
1026
1027 void anv_GetDeviceQueue(
1028 VkDevice _device,
1029 uint32_t queueNodeIndex,
1030 uint32_t queueIndex,
1031 VkQueue* pQueue)
1032 {
1033 ANV_FROM_HANDLE(anv_device, device, _device);
1034
1035 assert(queueIndex == 0);
1036
1037 *pQueue = anv_queue_to_handle(&device->queue);
1038 }
1039
1040 VkResult anv_QueueSubmit(
1041 VkQueue _queue,
1042 uint32_t submitCount,
1043 const VkSubmitInfo* pSubmits,
1044 VkFence _fence)
1045 {
1046 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1047 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1048 struct anv_device *device = queue->device;
1049 int ret;
1050
1051 for (uint32_t i = 0; i < submitCount; i++) {
1052 for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1053 ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1054 pSubmits[i].pCommandBuffers[j]);
1055 assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1056
1057 ret = anv_gem_execbuffer(device, &cmd_buffer->execbuf2.execbuf);
1058 if (ret != 0) {
1059 /* We don't know the real error. */
1060 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1061 "execbuf2 failed: %m");
1062 }
1063
1064 for (uint32_t k = 0; k < cmd_buffer->execbuf2.bo_count; k++)
1065 cmd_buffer->execbuf2.bos[k]->offset = cmd_buffer->execbuf2.objects[k].offset;
1066 }
1067 }
1068
1069 if (fence) {
1070 ret = anv_gem_execbuffer(device, &fence->execbuf);
1071 if (ret != 0) {
1072 /* We don't know the real error. */
1073 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1074 "execbuf2 failed: %m");
1075 }
1076 }
1077
1078 return VK_SUCCESS;
1079 }
1080
1081 VkResult anv_QueueWaitIdle(
1082 VkQueue _queue)
1083 {
1084 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1085
1086 return ANV_CALL(DeviceWaitIdle)(anv_device_to_handle(queue->device));
1087 }
1088
1089 VkResult anv_DeviceWaitIdle(
1090 VkDevice _device)
1091 {
1092 ANV_FROM_HANDLE(anv_device, device, _device);
1093 struct anv_batch batch;
1094
1095 uint32_t cmds[8];
1096 batch.start = batch.next = cmds;
1097 batch.end = (void *) cmds + sizeof(cmds);
1098
1099 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1100 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1101
1102 return anv_device_submit_simple_batch(device, &batch);
1103 }
1104
1105 VkResult
1106 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1107 {
1108 bo->gem_handle = anv_gem_create(device, size);
1109 if (!bo->gem_handle)
1110 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1111
1112 bo->map = NULL;
1113 bo->index = 0;
1114 bo->offset = 0;
1115 bo->size = size;
1116 bo->is_winsys_bo = false;
1117
1118 return VK_SUCCESS;
1119 }
1120
1121 VkResult anv_AllocateMemory(
1122 VkDevice _device,
1123 const VkMemoryAllocateInfo* pAllocateInfo,
1124 const VkAllocationCallbacks* pAllocator,
1125 VkDeviceMemory* pMem)
1126 {
1127 ANV_FROM_HANDLE(anv_device, device, _device);
1128 struct anv_device_memory *mem;
1129 VkResult result;
1130
1131 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1132
1133 if (pAllocateInfo->allocationSize == 0) {
1134 /* Apparently, this is allowed */
1135 *pMem = VK_NULL_HANDLE;
1136 return VK_SUCCESS;
1137 }
1138
1139 /* We support exactly one memory heap. */
1140 assert(pAllocateInfo->memoryTypeIndex == 0 ||
1141 (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1142
1143 /* FINISHME: Fail if allocation request exceeds heap size. */
1144
1145 mem = anv_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1146 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1147 if (mem == NULL)
1148 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1149
1150 /* The kernel is going to give us whole pages anyway */
1151 uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1152
1153 result = anv_bo_init_new(&mem->bo, device, alloc_size);
1154 if (result != VK_SUCCESS)
1155 goto fail;
1156
1157 mem->type_index = pAllocateInfo->memoryTypeIndex;
1158
1159 *pMem = anv_device_memory_to_handle(mem);
1160
1161 return VK_SUCCESS;
1162
1163 fail:
1164 anv_free2(&device->alloc, pAllocator, mem);
1165
1166 return result;
1167 }
1168
1169 void anv_FreeMemory(
1170 VkDevice _device,
1171 VkDeviceMemory _mem,
1172 const VkAllocationCallbacks* pAllocator)
1173 {
1174 ANV_FROM_HANDLE(anv_device, device, _device);
1175 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1176
1177 if (mem == NULL)
1178 return;
1179
1180 if (mem->bo.map)
1181 anv_gem_munmap(mem->bo.map, mem->bo.size);
1182
1183 if (mem->bo.gem_handle != 0)
1184 anv_gem_close(device, mem->bo.gem_handle);
1185
1186 anv_free2(&device->alloc, pAllocator, mem);
1187 }
1188
1189 VkResult anv_MapMemory(
1190 VkDevice _device,
1191 VkDeviceMemory _memory,
1192 VkDeviceSize offset,
1193 VkDeviceSize size,
1194 VkMemoryMapFlags flags,
1195 void** ppData)
1196 {
1197 ANV_FROM_HANDLE(anv_device, device, _device);
1198 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1199
1200 if (mem == NULL) {
1201 *ppData = NULL;
1202 return VK_SUCCESS;
1203 }
1204
1205 if (size == VK_WHOLE_SIZE)
1206 size = mem->bo.size - offset;
1207
1208 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1209 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1210 * at a time is valid. We could just mmap up front and return an offset
1211 * pointer here, but that may exhaust virtual memory on 32 bit
1212 * userspace. */
1213
1214 uint32_t gem_flags = 0;
1215 if (!device->info.has_llc && mem->type_index == 0)
1216 gem_flags |= I915_MMAP_WC;
1217
1218 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1219 uint64_t map_offset = offset & ~4095ull;
1220 assert(offset >= map_offset);
1221 uint64_t map_size = (offset + size) - map_offset;
1222
1223 /* Let's map whole pages */
1224 map_size = align_u64(map_size, 4096);
1225
1226 mem->map = anv_gem_mmap(device, mem->bo.gem_handle,
1227 map_offset, map_size, gem_flags);
1228 mem->map_size = map_size;
1229
1230 *ppData = mem->map + (offset - map_offset);
1231
1232 return VK_SUCCESS;
1233 }
1234
1235 void anv_UnmapMemory(
1236 VkDevice _device,
1237 VkDeviceMemory _memory)
1238 {
1239 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1240
1241 if (mem == NULL)
1242 return;
1243
1244 anv_gem_munmap(mem->map, mem->map_size);
1245 }
1246
1247 static void
1248 clflush_mapped_ranges(struct anv_device *device,
1249 uint32_t count,
1250 const VkMappedMemoryRange *ranges)
1251 {
1252 for (uint32_t i = 0; i < count; i++) {
1253 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1254 void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1255 void *end;
1256
1257 if (ranges[i].offset + ranges[i].size > mem->map_size)
1258 end = mem->map + mem->map_size;
1259 else
1260 end = mem->map + ranges[i].offset + ranges[i].size;
1261
1262 while (p < end) {
1263 __builtin_ia32_clflush(p);
1264 p += CACHELINE_SIZE;
1265 }
1266 }
1267 }
1268
1269 VkResult anv_FlushMappedMemoryRanges(
1270 VkDevice _device,
1271 uint32_t memoryRangeCount,
1272 const VkMappedMemoryRange* pMemoryRanges)
1273 {
1274 ANV_FROM_HANDLE(anv_device, device, _device);
1275
1276 if (device->info.has_llc)
1277 return VK_SUCCESS;
1278
1279 /* Make sure the writes we're flushing have landed. */
1280 __builtin_ia32_mfence();
1281
1282 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1283
1284 return VK_SUCCESS;
1285 }
1286
1287 VkResult anv_InvalidateMappedMemoryRanges(
1288 VkDevice _device,
1289 uint32_t memoryRangeCount,
1290 const VkMappedMemoryRange* pMemoryRanges)
1291 {
1292 ANV_FROM_HANDLE(anv_device, device, _device);
1293
1294 if (device->info.has_llc)
1295 return VK_SUCCESS;
1296
1297 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1298
1299 /* Make sure no reads get moved up above the invalidate. */
1300 __builtin_ia32_mfence();
1301
1302 return VK_SUCCESS;
1303 }
1304
1305 void anv_GetBufferMemoryRequirements(
1306 VkDevice device,
1307 VkBuffer _buffer,
1308 VkMemoryRequirements* pMemoryRequirements)
1309 {
1310 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1311
1312 /* The Vulkan spec (git aaed022) says:
1313 *
1314 * memoryTypeBits is a bitfield and contains one bit set for every
1315 * supported memory type for the resource. The bit `1<<i` is set if and
1316 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1317 * structure for the physical device is supported.
1318 *
1319 * We support exactly one memory type.
1320 */
1321 pMemoryRequirements->memoryTypeBits = 1;
1322
1323 pMemoryRequirements->size = buffer->size;
1324 pMemoryRequirements->alignment = 16;
1325 }
1326
1327 void anv_GetImageMemoryRequirements(
1328 VkDevice device,
1329 VkImage _image,
1330 VkMemoryRequirements* pMemoryRequirements)
1331 {
1332 ANV_FROM_HANDLE(anv_image, image, _image);
1333
1334 /* The Vulkan spec (git aaed022) says:
1335 *
1336 * memoryTypeBits is a bitfield and contains one bit set for every
1337 * supported memory type for the resource. The bit `1<<i` is set if and
1338 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1339 * structure for the physical device is supported.
1340 *
1341 * We support exactly one memory type.
1342 */
1343 pMemoryRequirements->memoryTypeBits = 1;
1344
1345 pMemoryRequirements->size = image->size;
1346 pMemoryRequirements->alignment = image->alignment;
1347 }
1348
1349 void anv_GetImageSparseMemoryRequirements(
1350 VkDevice device,
1351 VkImage image,
1352 uint32_t* pSparseMemoryRequirementCount,
1353 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
1354 {
1355 stub();
1356 }
1357
1358 void anv_GetDeviceMemoryCommitment(
1359 VkDevice device,
1360 VkDeviceMemory memory,
1361 VkDeviceSize* pCommittedMemoryInBytes)
1362 {
1363 *pCommittedMemoryInBytes = 0;
1364 }
1365
1366 VkResult anv_BindBufferMemory(
1367 VkDevice device,
1368 VkBuffer _buffer,
1369 VkDeviceMemory _memory,
1370 VkDeviceSize memoryOffset)
1371 {
1372 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1373 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1374
1375 if (mem) {
1376 buffer->bo = &mem->bo;
1377 buffer->offset = memoryOffset;
1378 } else {
1379 buffer->bo = NULL;
1380 buffer->offset = 0;
1381 }
1382
1383 return VK_SUCCESS;
1384 }
1385
1386 VkResult anv_BindImageMemory(
1387 VkDevice device,
1388 VkImage _image,
1389 VkDeviceMemory _memory,
1390 VkDeviceSize memoryOffset)
1391 {
1392 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1393 ANV_FROM_HANDLE(anv_image, image, _image);
1394
1395 if (mem) {
1396 image->bo = &mem->bo;
1397 image->offset = memoryOffset;
1398 } else {
1399 image->bo = NULL;
1400 image->offset = 0;
1401 }
1402
1403 return VK_SUCCESS;
1404 }
1405
1406 VkResult anv_QueueBindSparse(
1407 VkQueue queue,
1408 uint32_t bindInfoCount,
1409 const VkBindSparseInfo* pBindInfo,
1410 VkFence fence)
1411 {
1412 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1413 }
1414
1415 VkResult anv_CreateFence(
1416 VkDevice _device,
1417 const VkFenceCreateInfo* pCreateInfo,
1418 const VkAllocationCallbacks* pAllocator,
1419 VkFence* pFence)
1420 {
1421 ANV_FROM_HANDLE(anv_device, device, _device);
1422 struct anv_bo fence_bo;
1423 struct anv_fence *fence;
1424 struct anv_batch batch;
1425 VkResult result;
1426
1427 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1428
1429 result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1430 if (result != VK_SUCCESS)
1431 return result;
1432
1433 /* Fences are small. Just store the CPU data structure in the BO. */
1434 fence = fence_bo.map;
1435 fence->bo = fence_bo;
1436
1437 /* Place the batch after the CPU data but on its own cache line. */
1438 const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1439 batch.next = batch.start = fence->bo.map + batch_offset;
1440 batch.end = fence->bo.map + fence->bo.size;
1441 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1442 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1443
1444 if (!device->info.has_llc) {
1445 assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1446 assert(batch.next - batch.start <= CACHELINE_SIZE);
1447 __builtin_ia32_mfence();
1448 __builtin_ia32_clflush(batch.start);
1449 }
1450
1451 fence->exec2_objects[0].handle = fence->bo.gem_handle;
1452 fence->exec2_objects[0].relocation_count = 0;
1453 fence->exec2_objects[0].relocs_ptr = 0;
1454 fence->exec2_objects[0].alignment = 0;
1455 fence->exec2_objects[0].offset = fence->bo.offset;
1456 fence->exec2_objects[0].flags = 0;
1457 fence->exec2_objects[0].rsvd1 = 0;
1458 fence->exec2_objects[0].rsvd2 = 0;
1459
1460 fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1461 fence->execbuf.buffer_count = 1;
1462 fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1463 fence->execbuf.batch_len = batch.next - batch.start;
1464 fence->execbuf.cliprects_ptr = 0;
1465 fence->execbuf.num_cliprects = 0;
1466 fence->execbuf.DR1 = 0;
1467 fence->execbuf.DR4 = 0;
1468
1469 fence->execbuf.flags =
1470 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1471 fence->execbuf.rsvd1 = device->context_id;
1472 fence->execbuf.rsvd2 = 0;
1473
1474 fence->ready = false;
1475
1476 *pFence = anv_fence_to_handle(fence);
1477
1478 return VK_SUCCESS;
1479 }
1480
1481 void anv_DestroyFence(
1482 VkDevice _device,
1483 VkFence _fence,
1484 const VkAllocationCallbacks* pAllocator)
1485 {
1486 ANV_FROM_HANDLE(anv_device, device, _device);
1487 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1488
1489 assert(fence->bo.map == fence);
1490 anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1491 }
1492
1493 VkResult anv_ResetFences(
1494 VkDevice _device,
1495 uint32_t fenceCount,
1496 const VkFence* pFences)
1497 {
1498 for (uint32_t i = 0; i < fenceCount; i++) {
1499 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1500 fence->ready = false;
1501 }
1502
1503 return VK_SUCCESS;
1504 }
1505
1506 VkResult anv_GetFenceStatus(
1507 VkDevice _device,
1508 VkFence _fence)
1509 {
1510 ANV_FROM_HANDLE(anv_device, device, _device);
1511 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1512 int64_t t = 0;
1513 int ret;
1514
1515 if (fence->ready)
1516 return VK_SUCCESS;
1517
1518 ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1519 if (ret == 0) {
1520 fence->ready = true;
1521 return VK_SUCCESS;
1522 }
1523
1524 return VK_NOT_READY;
1525 }
1526
1527 VkResult anv_WaitForFences(
1528 VkDevice _device,
1529 uint32_t fenceCount,
1530 const VkFence* pFences,
1531 VkBool32 waitAll,
1532 uint64_t timeout)
1533 {
1534 ANV_FROM_HANDLE(anv_device, device, _device);
1535
1536 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1537 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1538 * for a couple of kernel releases. Since there's no way to know
1539 * whether or not the kernel we're using is one of the broken ones, the
1540 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1541 * maximum timeout from 584 years to 292 years - likely not a big deal.
1542 */
1543 if (timeout > INT64_MAX)
1544 timeout = INT64_MAX;
1545
1546 int64_t t = timeout;
1547
1548 /* FIXME: handle !waitAll */
1549
1550 for (uint32_t i = 0; i < fenceCount; i++) {
1551 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1552 int ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1553 if (ret == -1 && errno == ETIME) {
1554 return VK_TIMEOUT;
1555 } else if (ret == -1) {
1556 /* We don't know the real error. */
1557 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1558 "gem wait failed: %m");
1559 }
1560 }
1561
1562 return VK_SUCCESS;
1563 }
1564
1565 // Queue semaphore functions
1566
1567 VkResult anv_CreateSemaphore(
1568 VkDevice device,
1569 const VkSemaphoreCreateInfo* pCreateInfo,
1570 const VkAllocationCallbacks* pAllocator,
1571 VkSemaphore* pSemaphore)
1572 {
1573 /* The DRM execbuffer ioctl always execute in-oder, even between different
1574 * rings. As such, there's nothing to do for the user space semaphore.
1575 */
1576
1577 *pSemaphore = (VkSemaphore)1;
1578
1579 return VK_SUCCESS;
1580 }
1581
1582 void anv_DestroySemaphore(
1583 VkDevice device,
1584 VkSemaphore semaphore,
1585 const VkAllocationCallbacks* pAllocator)
1586 {
1587 }
1588
1589 // Event functions
1590
1591 VkResult anv_CreateEvent(
1592 VkDevice _device,
1593 const VkEventCreateInfo* pCreateInfo,
1594 const VkAllocationCallbacks* pAllocator,
1595 VkEvent* pEvent)
1596 {
1597 ANV_FROM_HANDLE(anv_device, device, _device);
1598 struct anv_state state;
1599 struct anv_event *event;
1600
1601 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1602
1603 state = anv_state_pool_alloc(&device->dynamic_state_pool,
1604 sizeof(*event), 8);
1605 event = state.map;
1606 event->state = state;
1607 event->semaphore = VK_EVENT_RESET;
1608
1609 if (!device->info.has_llc) {
1610 /* Make sure the writes we're flushing have landed. */
1611 __builtin_ia32_mfence();
1612 __builtin_ia32_clflush(event);
1613 }
1614
1615 *pEvent = anv_event_to_handle(event);
1616
1617 return VK_SUCCESS;
1618 }
1619
1620 void anv_DestroyEvent(
1621 VkDevice _device,
1622 VkEvent _event,
1623 const VkAllocationCallbacks* pAllocator)
1624 {
1625 ANV_FROM_HANDLE(anv_device, device, _device);
1626 ANV_FROM_HANDLE(anv_event, event, _event);
1627
1628 anv_state_pool_free(&device->dynamic_state_pool, event->state);
1629 }
1630
1631 VkResult anv_GetEventStatus(
1632 VkDevice _device,
1633 VkEvent _event)
1634 {
1635 ANV_FROM_HANDLE(anv_device, device, _device);
1636 ANV_FROM_HANDLE(anv_event, event, _event);
1637
1638 if (!device->info.has_llc) {
1639 /* Invalidate read cache before reading event written by GPU. */
1640 __builtin_ia32_clflush(event);
1641 __builtin_ia32_mfence();
1642
1643 }
1644
1645 return event->semaphore;
1646 }
1647
1648 VkResult anv_SetEvent(
1649 VkDevice _device,
1650 VkEvent _event)
1651 {
1652 ANV_FROM_HANDLE(anv_device, device, _device);
1653 ANV_FROM_HANDLE(anv_event, event, _event);
1654
1655 event->semaphore = VK_EVENT_SET;
1656
1657 if (!device->info.has_llc) {
1658 /* Make sure the writes we're flushing have landed. */
1659 __builtin_ia32_mfence();
1660 __builtin_ia32_clflush(event);
1661 }
1662
1663 return VK_SUCCESS;
1664 }
1665
1666 VkResult anv_ResetEvent(
1667 VkDevice _device,
1668 VkEvent _event)
1669 {
1670 ANV_FROM_HANDLE(anv_device, device, _device);
1671 ANV_FROM_HANDLE(anv_event, event, _event);
1672
1673 event->semaphore = VK_EVENT_RESET;
1674
1675 if (!device->info.has_llc) {
1676 /* Make sure the writes we're flushing have landed. */
1677 __builtin_ia32_mfence();
1678 __builtin_ia32_clflush(event);
1679 }
1680
1681 return VK_SUCCESS;
1682 }
1683
1684 // Buffer functions
1685
1686 VkResult anv_CreateBuffer(
1687 VkDevice _device,
1688 const VkBufferCreateInfo* pCreateInfo,
1689 const VkAllocationCallbacks* pAllocator,
1690 VkBuffer* pBuffer)
1691 {
1692 ANV_FROM_HANDLE(anv_device, device, _device);
1693 struct anv_buffer *buffer;
1694
1695 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1696
1697 buffer = anv_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1698 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1699 if (buffer == NULL)
1700 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1701
1702 buffer->size = pCreateInfo->size;
1703 buffer->usage = pCreateInfo->usage;
1704 buffer->bo = NULL;
1705 buffer->offset = 0;
1706
1707 *pBuffer = anv_buffer_to_handle(buffer);
1708
1709 return VK_SUCCESS;
1710 }
1711
1712 void anv_DestroyBuffer(
1713 VkDevice _device,
1714 VkBuffer _buffer,
1715 const VkAllocationCallbacks* pAllocator)
1716 {
1717 ANV_FROM_HANDLE(anv_device, device, _device);
1718 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1719
1720 anv_free2(&device->alloc, pAllocator, buffer);
1721 }
1722
1723 void
1724 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1725 enum isl_format format,
1726 uint32_t offset, uint32_t range, uint32_t stride)
1727 {
1728 isl_buffer_fill_state(&device->isl_dev, state.map,
1729 .address = offset,
1730 .mocs = device->default_mocs,
1731 .size = range,
1732 .format = format,
1733 .stride = stride);
1734
1735 if (!device->info.has_llc)
1736 anv_state_clflush(state);
1737 }
1738
1739 void anv_DestroySampler(
1740 VkDevice _device,
1741 VkSampler _sampler,
1742 const VkAllocationCallbacks* pAllocator)
1743 {
1744 ANV_FROM_HANDLE(anv_device, device, _device);
1745 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
1746
1747 anv_free2(&device->alloc, pAllocator, sampler);
1748 }
1749
1750 VkResult anv_CreateFramebuffer(
1751 VkDevice _device,
1752 const VkFramebufferCreateInfo* pCreateInfo,
1753 const VkAllocationCallbacks* pAllocator,
1754 VkFramebuffer* pFramebuffer)
1755 {
1756 ANV_FROM_HANDLE(anv_device, device, _device);
1757 struct anv_framebuffer *framebuffer;
1758
1759 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1760
1761 size_t size = sizeof(*framebuffer) +
1762 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
1763 framebuffer = anv_alloc2(&device->alloc, pAllocator, size, 8,
1764 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1765 if (framebuffer == NULL)
1766 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1767
1768 framebuffer->attachment_count = pCreateInfo->attachmentCount;
1769 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1770 VkImageView _iview = pCreateInfo->pAttachments[i];
1771 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
1772 }
1773
1774 framebuffer->width = pCreateInfo->width;
1775 framebuffer->height = pCreateInfo->height;
1776 framebuffer->layers = pCreateInfo->layers;
1777
1778 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
1779
1780 return VK_SUCCESS;
1781 }
1782
1783 void anv_DestroyFramebuffer(
1784 VkDevice _device,
1785 VkFramebuffer _fb,
1786 const VkAllocationCallbacks* pAllocator)
1787 {
1788 ANV_FROM_HANDLE(anv_device, device, _device);
1789 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
1790
1791 anv_free2(&device->alloc, pAllocator, fb);
1792 }
1793
1794 void vkCmdDbgMarkerBegin(
1795 VkCommandBuffer commandBuffer,
1796 const char* pMarker)
1797 __attribute__ ((visibility ("default")));
1798
1799 void vkCmdDbgMarkerEnd(
1800 VkCommandBuffer commandBuffer)
1801 __attribute__ ((visibility ("default")));
1802
1803 void vkCmdDbgMarkerBegin(
1804 VkCommandBuffer commandBuffer,
1805 const char* pMarker)
1806 {
1807 }
1808
1809 void vkCmdDbgMarkerEnd(
1810 VkCommandBuffer commandBuffer)
1811 {
1812 }