anv: Get rid of meta
[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_error(VK_ERROR_INCOMPATIBLE_DRIVER);
66
67 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
68 device->instance = instance;
69
70 assert(strlen(path) < ARRAY_SIZE(device->path));
71 strncpy(device->path, path, ARRAY_SIZE(device->path));
72
73 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
74 if (!device->chipset_id) {
75 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
76 goto fail;
77 }
78
79 device->name = gen_get_device_name(device->chipset_id);
80 if (!gen_get_device_info(device->chipset_id, &device->info)) {
81 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
82 goto fail;
83 }
84
85 if (device->info.is_haswell) {
86 fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
87 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
88 fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
89 } else if (device->info.gen == 7 && device->info.is_baytrail) {
90 fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
91 } else if (device->info.gen >= 8) {
92 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
93 * supported as anything */
94 } else {
95 result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
96 "Vulkan not yet supported on %s", device->name);
97 goto fail;
98 }
99
100 device->cmd_parser_version = -1;
101 if (device->info.gen == 7) {
102 device->cmd_parser_version =
103 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
104 if (device->cmd_parser_version == -1) {
105 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
106 "failed to get command parser version");
107 goto fail;
108 }
109 }
110
111 if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
112 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
113 "failed to get aperture size: %m");
114 goto fail;
115 }
116
117 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
118 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
119 "kernel missing gem wait");
120 goto fail;
121 }
122
123 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
124 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
125 "kernel missing execbuf2");
126 goto fail;
127 }
128
129 if (!device->info.has_llc &&
130 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
131 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
132 "kernel missing wc mmap");
133 goto fail;
134 }
135
136 bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
137
138 /* GENs prior to 8 do not support EU/Subslice info */
139 if (device->info.gen >= 8) {
140 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
141 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
142
143 /* Without this information, we cannot get the right Braswell
144 * brandstrings, and we have to use conservative numbers for GPGPU on
145 * many platforms, but otherwise, things will just work.
146 */
147 if (device->subslice_total < 1 || device->eu_total < 1) {
148 fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
149 " query GPU properties.\n");
150 }
151 } else if (device->info.gen == 7) {
152 device->subslice_total = 1 << (device->info.gt - 1);
153 }
154
155 if (device->info.is_cherryview &&
156 device->subslice_total > 0 && device->eu_total > 0) {
157 /* Logical CS threads = EUs per subslice * 7 threads per EU */
158 uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
159
160 /* Fuse configurations may give more threads than expected, never less. */
161 if (max_cs_threads > device->info.max_cs_threads)
162 device->info.max_cs_threads = max_cs_threads;
163 }
164
165 close(fd);
166
167 brw_process_intel_debug_variable();
168
169 device->compiler = brw_compiler_create(NULL, &device->info);
170 if (device->compiler == NULL) {
171 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
172 goto fail;
173 }
174 device->compiler->shader_debug_log = compiler_debug_log;
175 device->compiler->shader_perf_log = compiler_perf_log;
176
177 result = anv_init_wsi(device);
178 if (result != VK_SUCCESS)
179 goto fail;
180
181 /* XXX: Actually detect bit6 swizzling */
182 isl_device_init(&device->isl_dev, &device->info, swizzled);
183
184 return VK_SUCCESS;
185
186 fail:
187 close(fd);
188 return result;
189 }
190
191 static void
192 anv_physical_device_finish(struct anv_physical_device *device)
193 {
194 anv_finish_wsi(device);
195 ralloc_free(device->compiler);
196 }
197
198 static const VkExtensionProperties global_extensions[] = {
199 {
200 .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
201 .specVersion = 25,
202 },
203 #ifdef VK_USE_PLATFORM_XCB_KHR
204 {
205 .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
206 .specVersion = 5,
207 },
208 #endif
209 #ifdef VK_USE_PLATFORM_XLIB_KHR
210 {
211 .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
212 .specVersion = 5,
213 },
214 #endif
215 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
216 {
217 .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
218 .specVersion = 4,
219 },
220 #endif
221 };
222
223 static const VkExtensionProperties device_extensions[] = {
224 {
225 .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
226 .specVersion = 67,
227 },
228 };
229
230 static void *
231 default_alloc_func(void *pUserData, size_t size, size_t align,
232 VkSystemAllocationScope allocationScope)
233 {
234 return malloc(size);
235 }
236
237 static void *
238 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
239 size_t align, VkSystemAllocationScope allocationScope)
240 {
241 return realloc(pOriginal, size);
242 }
243
244 static void
245 default_free_func(void *pUserData, void *pMemory)
246 {
247 free(pMemory);
248 }
249
250 static const VkAllocationCallbacks default_alloc = {
251 .pUserData = NULL,
252 .pfnAllocation = default_alloc_func,
253 .pfnReallocation = default_realloc_func,
254 .pfnFree = default_free_func,
255 };
256
257 VkResult anv_CreateInstance(
258 const VkInstanceCreateInfo* pCreateInfo,
259 const VkAllocationCallbacks* pAllocator,
260 VkInstance* pInstance)
261 {
262 struct anv_instance *instance;
263
264 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
265
266 uint32_t client_version;
267 if (pCreateInfo->pApplicationInfo &&
268 pCreateInfo->pApplicationInfo->apiVersion != 0) {
269 client_version = pCreateInfo->pApplicationInfo->apiVersion;
270 } else {
271 client_version = VK_MAKE_VERSION(1, 0, 0);
272 }
273
274 if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
275 client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
276 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
277 "Client requested version %d.%d.%d",
278 VK_VERSION_MAJOR(client_version),
279 VK_VERSION_MINOR(client_version),
280 VK_VERSION_PATCH(client_version));
281 }
282
283 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
284 bool found = false;
285 for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
286 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
287 global_extensions[j].extensionName) == 0) {
288 found = true;
289 break;
290 }
291 }
292 if (!found)
293 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
294 }
295
296 instance = anv_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
297 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
298 if (!instance)
299 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
300
301 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
302
303 if (pAllocator)
304 instance->alloc = *pAllocator;
305 else
306 instance->alloc = default_alloc;
307
308 instance->apiVersion = client_version;
309 instance->physicalDeviceCount = -1;
310
311 _mesa_locale_init();
312
313 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
314
315 *pInstance = anv_instance_to_handle(instance);
316
317 return VK_SUCCESS;
318 }
319
320 void anv_DestroyInstance(
321 VkInstance _instance,
322 const VkAllocationCallbacks* pAllocator)
323 {
324 ANV_FROM_HANDLE(anv_instance, instance, _instance);
325
326 if (instance->physicalDeviceCount > 0) {
327 /* We support at most one physical device. */
328 assert(instance->physicalDeviceCount == 1);
329 anv_physical_device_finish(&instance->physicalDevice);
330 }
331
332 VG(VALGRIND_DESTROY_MEMPOOL(instance));
333
334 _mesa_locale_fini();
335
336 anv_free(&instance->alloc, instance);
337 }
338
339 VkResult anv_EnumeratePhysicalDevices(
340 VkInstance _instance,
341 uint32_t* pPhysicalDeviceCount,
342 VkPhysicalDevice* pPhysicalDevices)
343 {
344 ANV_FROM_HANDLE(anv_instance, instance, _instance);
345 VkResult result;
346
347 if (instance->physicalDeviceCount < 0) {
348 char path[20];
349 for (unsigned i = 0; i < 8; i++) {
350 snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
351 result = anv_physical_device_init(&instance->physicalDevice,
352 instance, path);
353 if (result == VK_SUCCESS)
354 break;
355 }
356
357 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
358 instance->physicalDeviceCount = 0;
359 } else if (result == VK_SUCCESS) {
360 instance->physicalDeviceCount = 1;
361 } else {
362 return result;
363 }
364 }
365
366 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
367 * otherwise it's an inout parameter.
368 *
369 * The Vulkan spec (git aaed022) says:
370 *
371 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
372 * that is initialized with the number of devices the application is
373 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
374 * an array of at least this many VkPhysicalDevice handles [...].
375 *
376 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
377 * overwrites the contents of the variable pointed to by
378 * pPhysicalDeviceCount with the number of physical devices in in the
379 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
380 * pPhysicalDeviceCount with the number of physical handles written to
381 * pPhysicalDevices.
382 */
383 if (!pPhysicalDevices) {
384 *pPhysicalDeviceCount = instance->physicalDeviceCount;
385 } else if (*pPhysicalDeviceCount >= 1) {
386 pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
387 *pPhysicalDeviceCount = 1;
388 } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
389 return VK_INCOMPLETE;
390 } else {
391 *pPhysicalDeviceCount = 0;
392 }
393
394 return VK_SUCCESS;
395 }
396
397 void anv_GetPhysicalDeviceFeatures(
398 VkPhysicalDevice physicalDevice,
399 VkPhysicalDeviceFeatures* pFeatures)
400 {
401 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
402
403 *pFeatures = (VkPhysicalDeviceFeatures) {
404 .robustBufferAccess = true,
405 .fullDrawIndexUint32 = true,
406 .imageCubeArray = false,
407 .independentBlend = true,
408 .geometryShader = true,
409 .tessellationShader = false,
410 .sampleRateShading = true,
411 .dualSrcBlend = true,
412 .logicOp = true,
413 .multiDrawIndirect = false,
414 .drawIndirectFirstInstance = false,
415 .depthClamp = true,
416 .depthBiasClamp = false,
417 .fillModeNonSolid = true,
418 .depthBounds = false,
419 .wideLines = true,
420 .largePoints = true,
421 .alphaToOne = true,
422 .multiViewport = true,
423 .samplerAnisotropy = true,
424 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
425 pdevice->info.is_baytrail,
426 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
427 .textureCompressionBC = true,
428 .occlusionQueryPrecise = true,
429 .pipelineStatisticsQuery = false,
430 .fragmentStoresAndAtomics = true,
431 .shaderTessellationAndGeometryPointSize = true,
432 .shaderImageGatherExtended = false,
433 .shaderStorageImageExtendedFormats = false,
434 .shaderStorageImageMultisample = false,
435 .shaderUniformBufferArrayDynamicIndexing = true,
436 .shaderSampledImageArrayDynamicIndexing = true,
437 .shaderStorageBufferArrayDynamicIndexing = true,
438 .shaderStorageImageArrayDynamicIndexing = true,
439 .shaderStorageImageReadWithoutFormat = false,
440 .shaderStorageImageWriteWithoutFormat = true,
441 .shaderClipDistance = false,
442 .shaderCullDistance = false,
443 .shaderFloat64 = false,
444 .shaderInt64 = false,
445 .shaderInt16 = false,
446 .alphaToOne = true,
447 .variableMultisampleRate = false,
448 .inheritedQueries = false,
449 };
450
451 /* We can't do image stores in vec4 shaders */
452 pFeatures->vertexPipelineStoresAndAtomics =
453 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
454 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
455 }
456
457 void
458 anv_device_get_cache_uuid(void *uuid)
459 {
460 memset(uuid, 0, VK_UUID_SIZE);
461 snprintf(uuid, VK_UUID_SIZE, "anv-%s", ANV_TIMESTAMP);
462 }
463
464 void anv_GetPhysicalDeviceProperties(
465 VkPhysicalDevice physicalDevice,
466 VkPhysicalDeviceProperties* pProperties)
467 {
468 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
469 const struct gen_device_info *devinfo = &pdevice->info;
470
471 const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
472
473 /* See assertions made when programming the buffer surface state. */
474 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
475 (1ul << 30) : (1ul << 27);
476
477 VkSampleCountFlags sample_counts =
478 isl_device_get_sample_counts(&pdevice->isl_dev);
479
480 VkPhysicalDeviceLimits limits = {
481 .maxImageDimension1D = (1 << 14),
482 .maxImageDimension2D = (1 << 14),
483 .maxImageDimension3D = (1 << 11),
484 .maxImageDimensionCube = (1 << 14),
485 .maxImageArrayLayers = (1 << 11),
486 .maxTexelBufferElements = 128 * 1024 * 1024,
487 .maxUniformBufferRange = (1ul << 27),
488 .maxStorageBufferRange = max_raw_buffer_sz,
489 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
490 .maxMemoryAllocationCount = UINT32_MAX,
491 .maxSamplerAllocationCount = 64 * 1024,
492 .bufferImageGranularity = 64, /* A cache line */
493 .sparseAddressSpaceSize = 0,
494 .maxBoundDescriptorSets = MAX_SETS,
495 .maxPerStageDescriptorSamplers = 64,
496 .maxPerStageDescriptorUniformBuffers = 64,
497 .maxPerStageDescriptorStorageBuffers = 64,
498 .maxPerStageDescriptorSampledImages = 64,
499 .maxPerStageDescriptorStorageImages = 64,
500 .maxPerStageDescriptorInputAttachments = 64,
501 .maxPerStageResources = 128,
502 .maxDescriptorSetSamplers = 256,
503 .maxDescriptorSetUniformBuffers = 256,
504 .maxDescriptorSetUniformBuffersDynamic = 256,
505 .maxDescriptorSetStorageBuffers = 256,
506 .maxDescriptorSetStorageBuffersDynamic = 256,
507 .maxDescriptorSetSampledImages = 256,
508 .maxDescriptorSetStorageImages = 256,
509 .maxDescriptorSetInputAttachments = 256,
510 .maxVertexInputAttributes = 32,
511 .maxVertexInputBindings = 32,
512 .maxVertexInputAttributeOffset = 2047,
513 .maxVertexInputBindingStride = 2048,
514 .maxVertexOutputComponents = 128,
515 .maxTessellationGenerationLevel = 0,
516 .maxTessellationPatchSize = 0,
517 .maxTessellationControlPerVertexInputComponents = 0,
518 .maxTessellationControlPerVertexOutputComponents = 0,
519 .maxTessellationControlPerPatchOutputComponents = 0,
520 .maxTessellationControlTotalOutputComponents = 0,
521 .maxTessellationEvaluationInputComponents = 0,
522 .maxTessellationEvaluationOutputComponents = 0,
523 .maxGeometryShaderInvocations = 32,
524 .maxGeometryInputComponents = 64,
525 .maxGeometryOutputComponents = 128,
526 .maxGeometryOutputVertices = 256,
527 .maxGeometryTotalOutputComponents = 1024,
528 .maxFragmentInputComponents = 128,
529 .maxFragmentOutputAttachments = 8,
530 .maxFragmentDualSrcAttachments = 2,
531 .maxFragmentCombinedOutputResources = 8,
532 .maxComputeSharedMemorySize = 32768,
533 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
534 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
535 .maxComputeWorkGroupSize = {
536 16 * devinfo->max_cs_threads,
537 16 * devinfo->max_cs_threads,
538 16 * devinfo->max_cs_threads,
539 },
540 .subPixelPrecisionBits = 4 /* FIXME */,
541 .subTexelPrecisionBits = 4 /* FIXME */,
542 .mipmapPrecisionBits = 4 /* FIXME */,
543 .maxDrawIndexedIndexValue = UINT32_MAX,
544 .maxDrawIndirectCount = UINT32_MAX,
545 .maxSamplerLodBias = 16,
546 .maxSamplerAnisotropy = 16,
547 .maxViewports = MAX_VIEWPORTS,
548 .maxViewportDimensions = { (1 << 14), (1 << 14) },
549 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
550 .viewportSubPixelBits = 13, /* We take a float? */
551 .minMemoryMapAlignment = 4096, /* A page */
552 .minTexelBufferOffsetAlignment = 1,
553 .minUniformBufferOffsetAlignment = 1,
554 .minStorageBufferOffsetAlignment = 1,
555 .minTexelOffset = -8,
556 .maxTexelOffset = 7,
557 .minTexelGatherOffset = -8,
558 .maxTexelGatherOffset = 7,
559 .minInterpolationOffset = -0.5,
560 .maxInterpolationOffset = 0.4375,
561 .subPixelInterpolationOffsetBits = 4,
562 .maxFramebufferWidth = (1 << 14),
563 .maxFramebufferHeight = (1 << 14),
564 .maxFramebufferLayers = (1 << 10),
565 .framebufferColorSampleCounts = sample_counts,
566 .framebufferDepthSampleCounts = sample_counts,
567 .framebufferStencilSampleCounts = sample_counts,
568 .framebufferNoAttachmentsSampleCounts = sample_counts,
569 .maxColorAttachments = MAX_RTS,
570 .sampledImageColorSampleCounts = sample_counts,
571 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
572 .sampledImageDepthSampleCounts = sample_counts,
573 .sampledImageStencilSampleCounts = sample_counts,
574 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
575 .maxSampleMaskWords = 1,
576 .timestampComputeAndGraphics = false,
577 .timestampPeriod = time_stamp_base,
578 .maxClipDistances = 0 /* FIXME */,
579 .maxCullDistances = 0 /* FIXME */,
580 .maxCombinedClipAndCullDistances = 0 /* FIXME */,
581 .discreteQueuePriorities = 1,
582 .pointSizeRange = { 0.125, 255.875 },
583 .lineWidthRange = { 0.0, 7.9921875 },
584 .pointSizeGranularity = (1.0 / 8.0),
585 .lineWidthGranularity = (1.0 / 128.0),
586 .strictLines = false, /* FINISHME */
587 .standardSampleLocations = true,
588 .optimalBufferCopyOffsetAlignment = 128,
589 .optimalBufferCopyRowPitchAlignment = 128,
590 .nonCoherentAtomSize = 64,
591 };
592
593 *pProperties = (VkPhysicalDeviceProperties) {
594 .apiVersion = VK_MAKE_VERSION(1, 0, 5),
595 .driverVersion = 1,
596 .vendorID = 0x8086,
597 .deviceID = pdevice->chipset_id,
598 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
599 .limits = limits,
600 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
601 };
602
603 strcpy(pProperties->deviceName, pdevice->name);
604 anv_device_get_cache_uuid(pProperties->pipelineCacheUUID);
605 }
606
607 void anv_GetPhysicalDeviceQueueFamilyProperties(
608 VkPhysicalDevice physicalDevice,
609 uint32_t* pCount,
610 VkQueueFamilyProperties* pQueueFamilyProperties)
611 {
612 if (pQueueFamilyProperties == NULL) {
613 *pCount = 1;
614 return;
615 }
616
617 assert(*pCount >= 1);
618
619 *pQueueFamilyProperties = (VkQueueFamilyProperties) {
620 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
621 VK_QUEUE_COMPUTE_BIT |
622 VK_QUEUE_TRANSFER_BIT,
623 .queueCount = 1,
624 .timestampValidBits = 36, /* XXX: Real value here */
625 .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
626 };
627 }
628
629 void anv_GetPhysicalDeviceMemoryProperties(
630 VkPhysicalDevice physicalDevice,
631 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
632 {
633 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
634 VkDeviceSize heap_size;
635
636 /* Reserve some wiggle room for the driver by exposing only 75% of the
637 * aperture to the heap.
638 */
639 heap_size = 3 * physical_device->aperture_size / 4;
640
641 if (physical_device->info.has_llc) {
642 /* Big core GPUs share LLC with the CPU and thus one memory type can be
643 * both cached and coherent at the same time.
644 */
645 pMemoryProperties->memoryTypeCount = 1;
646 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
647 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
648 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
649 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
650 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
651 .heapIndex = 0,
652 };
653 } else {
654 /* The spec requires that we expose a host-visible, coherent memory
655 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
656 * to give the application a choice between cached, but not coherent and
657 * coherent but uncached (WC though).
658 */
659 pMemoryProperties->memoryTypeCount = 2;
660 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
661 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
662 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
663 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
664 .heapIndex = 0,
665 };
666 pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
667 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
668 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
669 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
670 .heapIndex = 0,
671 };
672 }
673
674 pMemoryProperties->memoryHeapCount = 1;
675 pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
676 .size = heap_size,
677 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
678 };
679 }
680
681 PFN_vkVoidFunction anv_GetInstanceProcAddr(
682 VkInstance instance,
683 const char* pName)
684 {
685 return anv_lookup_entrypoint(pName);
686 }
687
688 /* With version 1+ of the loader interface the ICD should expose
689 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
690 */
691 PUBLIC
692 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
693 VkInstance instance,
694 const char* pName);
695
696 PUBLIC
697 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
698 VkInstance instance,
699 const char* pName)
700 {
701 return anv_GetInstanceProcAddr(instance, pName);
702 }
703
704 PFN_vkVoidFunction anv_GetDeviceProcAddr(
705 VkDevice device,
706 const char* pName)
707 {
708 return anv_lookup_entrypoint(pName);
709 }
710
711 static VkResult
712 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
713 {
714 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
715 queue->device = device;
716 queue->pool = &device->surface_state_pool;
717
718 return VK_SUCCESS;
719 }
720
721 static void
722 anv_queue_finish(struct anv_queue *queue)
723 {
724 }
725
726 static struct anv_state
727 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
728 {
729 struct anv_state state;
730
731 state = anv_state_pool_alloc(pool, size, align);
732 memcpy(state.map, p, size);
733
734 if (!pool->block_pool->device->info.has_llc)
735 anv_state_clflush(state);
736
737 return state;
738 }
739
740 struct gen8_border_color {
741 union {
742 float float32[4];
743 uint32_t uint32[4];
744 };
745 /* Pad out to 64 bytes */
746 uint32_t _pad[12];
747 };
748
749 static void
750 anv_device_init_border_colors(struct anv_device *device)
751 {
752 static const struct gen8_border_color border_colors[] = {
753 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
754 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
755 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
756 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
757 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
758 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
759 };
760
761 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
762 sizeof(border_colors), 64,
763 border_colors);
764 }
765
766 VkResult
767 anv_device_submit_simple_batch(struct anv_device *device,
768 struct anv_batch *batch)
769 {
770 struct drm_i915_gem_execbuffer2 execbuf;
771 struct drm_i915_gem_exec_object2 exec2_objects[1];
772 struct anv_bo bo;
773 VkResult result = VK_SUCCESS;
774 uint32_t size;
775 int64_t timeout;
776 int ret;
777
778 /* Kernel driver requires 8 byte aligned batch length */
779 size = align_u32(batch->next - batch->start, 8);
780 result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
781 if (result != VK_SUCCESS)
782 return result;
783
784 memcpy(bo.map, batch->start, size);
785 if (!device->info.has_llc)
786 anv_clflush_range(bo.map, size);
787
788 exec2_objects[0].handle = bo.gem_handle;
789 exec2_objects[0].relocation_count = 0;
790 exec2_objects[0].relocs_ptr = 0;
791 exec2_objects[0].alignment = 0;
792 exec2_objects[0].offset = bo.offset;
793 exec2_objects[0].flags = 0;
794 exec2_objects[0].rsvd1 = 0;
795 exec2_objects[0].rsvd2 = 0;
796
797 execbuf.buffers_ptr = (uintptr_t) exec2_objects;
798 execbuf.buffer_count = 1;
799 execbuf.batch_start_offset = 0;
800 execbuf.batch_len = size;
801 execbuf.cliprects_ptr = 0;
802 execbuf.num_cliprects = 0;
803 execbuf.DR1 = 0;
804 execbuf.DR4 = 0;
805
806 execbuf.flags =
807 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
808 execbuf.rsvd1 = device->context_id;
809 execbuf.rsvd2 = 0;
810
811 ret = anv_gem_execbuffer(device, &execbuf);
812 if (ret != 0) {
813 /* We don't know the real error. */
814 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
815 goto fail;
816 }
817
818 timeout = INT64_MAX;
819 ret = anv_gem_wait(device, bo.gem_handle, &timeout);
820 if (ret != 0) {
821 /* We don't know the real error. */
822 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
823 goto fail;
824 }
825
826 fail:
827 anv_bo_pool_free(&device->batch_bo_pool, &bo);
828
829 return result;
830 }
831
832 VkResult anv_CreateDevice(
833 VkPhysicalDevice physicalDevice,
834 const VkDeviceCreateInfo* pCreateInfo,
835 const VkAllocationCallbacks* pAllocator,
836 VkDevice* pDevice)
837 {
838 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
839 VkResult result;
840 struct anv_device *device;
841
842 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
843
844 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
845 bool found = false;
846 for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
847 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
848 device_extensions[j].extensionName) == 0) {
849 found = true;
850 break;
851 }
852 }
853 if (!found)
854 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
855 }
856
857 anv_set_dispatch_devinfo(&physical_device->info);
858
859 device = anv_alloc2(&physical_device->instance->alloc, pAllocator,
860 sizeof(*device), 8,
861 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
862 if (!device)
863 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
864
865 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
866 device->instance = physical_device->instance;
867 device->chipset_id = physical_device->chipset_id;
868
869 if (pAllocator)
870 device->alloc = *pAllocator;
871 else
872 device->alloc = physical_device->instance->alloc;
873
874 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
875 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
876 if (device->fd == -1) {
877 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
878 goto fail_device;
879 }
880
881 device->context_id = anv_gem_create_context(device);
882 if (device->context_id == -1) {
883 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
884 goto fail_fd;
885 }
886
887 device->info = physical_device->info;
888 device->isl_dev = physical_device->isl_dev;
889
890 /* On Broadwell and later, we can use batch chaining to more efficiently
891 * implement growing command buffers. Prior to Haswell, the kernel
892 * command parser gets in the way and we have to fall back to growing
893 * the batch.
894 */
895 device->can_chain_batches = device->info.gen >= 8;
896
897 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
898 pCreateInfo->pEnabledFeatures->robustBufferAccess;
899
900 pthread_mutex_init(&device->mutex, NULL);
901
902 anv_bo_pool_init(&device->batch_bo_pool, device);
903
904 anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
905
906 anv_state_pool_init(&device->dynamic_state_pool,
907 &device->dynamic_state_block_pool);
908
909 anv_block_pool_init(&device->instruction_block_pool, device, 128 * 1024);
910 anv_state_pool_init(&device->instruction_state_pool,
911 &device->instruction_block_pool);
912
913 anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
914
915 anv_state_pool_init(&device->surface_state_pool,
916 &device->surface_state_block_pool);
917
918 anv_bo_init_new(&device->workaround_bo, device, 1024);
919
920 anv_scratch_pool_init(device, &device->scratch_pool);
921
922 anv_queue_init(device, &device->queue);
923
924 switch (device->info.gen) {
925 case 7:
926 if (!device->info.is_haswell)
927 result = gen7_init_device_state(device);
928 else
929 result = gen75_init_device_state(device);
930 break;
931 case 8:
932 result = gen8_init_device_state(device);
933 break;
934 case 9:
935 result = gen9_init_device_state(device);
936 break;
937 default:
938 /* Shouldn't get here as we don't create physical devices for any other
939 * gens. */
940 unreachable("unhandled gen");
941 }
942 if (result != VK_SUCCESS)
943 goto fail_fd;
944
945 anv_device_init_blorp(device);
946
947 anv_device_init_border_colors(device);
948
949 *pDevice = anv_device_to_handle(device);
950
951 return VK_SUCCESS;
952
953 fail_fd:
954 close(device->fd);
955 fail_device:
956 anv_free(&device->alloc, device);
957
958 return result;
959 }
960
961 void anv_DestroyDevice(
962 VkDevice _device,
963 const VkAllocationCallbacks* pAllocator)
964 {
965 ANV_FROM_HANDLE(anv_device, device, _device);
966
967 anv_queue_finish(&device->queue);
968
969 anv_device_finish_blorp(device);
970
971 #ifdef HAVE_VALGRIND
972 /* We only need to free these to prevent valgrind errors. The backing
973 * BO will go away in a couple of lines so we don't actually leak.
974 */
975 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
976 #endif
977
978 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
979 anv_gem_close(device, device->workaround_bo.gem_handle);
980
981 anv_bo_pool_finish(&device->batch_bo_pool);
982 anv_state_pool_finish(&device->dynamic_state_pool);
983 anv_block_pool_finish(&device->dynamic_state_block_pool);
984 anv_state_pool_finish(&device->instruction_state_pool);
985 anv_block_pool_finish(&device->instruction_block_pool);
986 anv_state_pool_finish(&device->surface_state_pool);
987 anv_block_pool_finish(&device->surface_state_block_pool);
988 anv_scratch_pool_finish(device, &device->scratch_pool);
989
990 close(device->fd);
991
992 pthread_mutex_destroy(&device->mutex);
993
994 anv_free(&device->alloc, device);
995 }
996
997 VkResult anv_EnumerateInstanceExtensionProperties(
998 const char* pLayerName,
999 uint32_t* pPropertyCount,
1000 VkExtensionProperties* pProperties)
1001 {
1002 if (pProperties == NULL) {
1003 *pPropertyCount = ARRAY_SIZE(global_extensions);
1004 return VK_SUCCESS;
1005 }
1006
1007 assert(*pPropertyCount >= ARRAY_SIZE(global_extensions));
1008
1009 *pPropertyCount = ARRAY_SIZE(global_extensions);
1010 memcpy(pProperties, global_extensions, sizeof(global_extensions));
1011
1012 return VK_SUCCESS;
1013 }
1014
1015 VkResult anv_EnumerateDeviceExtensionProperties(
1016 VkPhysicalDevice physicalDevice,
1017 const char* pLayerName,
1018 uint32_t* pPropertyCount,
1019 VkExtensionProperties* pProperties)
1020 {
1021 if (pProperties == NULL) {
1022 *pPropertyCount = ARRAY_SIZE(device_extensions);
1023 return VK_SUCCESS;
1024 }
1025
1026 assert(*pPropertyCount >= ARRAY_SIZE(device_extensions));
1027
1028 *pPropertyCount = ARRAY_SIZE(device_extensions);
1029 memcpy(pProperties, device_extensions, sizeof(device_extensions));
1030
1031 return VK_SUCCESS;
1032 }
1033
1034 VkResult anv_EnumerateInstanceLayerProperties(
1035 uint32_t* pPropertyCount,
1036 VkLayerProperties* pProperties)
1037 {
1038 if (pProperties == NULL) {
1039 *pPropertyCount = 0;
1040 return VK_SUCCESS;
1041 }
1042
1043 /* None supported at this time */
1044 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1045 }
1046
1047 VkResult anv_EnumerateDeviceLayerProperties(
1048 VkPhysicalDevice physicalDevice,
1049 uint32_t* pPropertyCount,
1050 VkLayerProperties* pProperties)
1051 {
1052 if (pProperties == NULL) {
1053 *pPropertyCount = 0;
1054 return VK_SUCCESS;
1055 }
1056
1057 /* None supported at this time */
1058 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1059 }
1060
1061 void anv_GetDeviceQueue(
1062 VkDevice _device,
1063 uint32_t queueNodeIndex,
1064 uint32_t queueIndex,
1065 VkQueue* pQueue)
1066 {
1067 ANV_FROM_HANDLE(anv_device, device, _device);
1068
1069 assert(queueIndex == 0);
1070
1071 *pQueue = anv_queue_to_handle(&device->queue);
1072 }
1073
1074 VkResult anv_QueueSubmit(
1075 VkQueue _queue,
1076 uint32_t submitCount,
1077 const VkSubmitInfo* pSubmits,
1078 VkFence _fence)
1079 {
1080 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1081 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1082 struct anv_device *device = queue->device;
1083 int ret;
1084
1085 for (uint32_t i = 0; i < submitCount; i++) {
1086 for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1087 ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1088 pSubmits[i].pCommandBuffers[j]);
1089 assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1090
1091 ret = anv_gem_execbuffer(device, &cmd_buffer->execbuf2.execbuf);
1092 if (ret != 0) {
1093 /* We don't know the real error. */
1094 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1095 "execbuf2 failed: %m");
1096 }
1097
1098 for (uint32_t k = 0; k < cmd_buffer->execbuf2.bo_count; k++)
1099 cmd_buffer->execbuf2.bos[k]->offset = cmd_buffer->execbuf2.objects[k].offset;
1100 }
1101 }
1102
1103 if (fence) {
1104 ret = anv_gem_execbuffer(device, &fence->execbuf);
1105 if (ret != 0) {
1106 /* We don't know the real error. */
1107 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1108 "execbuf2 failed: %m");
1109 }
1110 }
1111
1112 return VK_SUCCESS;
1113 }
1114
1115 VkResult anv_QueueWaitIdle(
1116 VkQueue _queue)
1117 {
1118 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1119
1120 return ANV_CALL(DeviceWaitIdle)(anv_device_to_handle(queue->device));
1121 }
1122
1123 VkResult anv_DeviceWaitIdle(
1124 VkDevice _device)
1125 {
1126 ANV_FROM_HANDLE(anv_device, device, _device);
1127 struct anv_batch batch;
1128
1129 uint32_t cmds[8];
1130 batch.start = batch.next = cmds;
1131 batch.end = (void *) cmds + sizeof(cmds);
1132
1133 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1134 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1135
1136 return anv_device_submit_simple_batch(device, &batch);
1137 }
1138
1139 VkResult
1140 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1141 {
1142 bo->gem_handle = anv_gem_create(device, size);
1143 if (!bo->gem_handle)
1144 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1145
1146 bo->map = NULL;
1147 bo->index = 0;
1148 bo->offset = 0;
1149 bo->size = size;
1150 bo->is_winsys_bo = false;
1151
1152 return VK_SUCCESS;
1153 }
1154
1155 VkResult anv_AllocateMemory(
1156 VkDevice _device,
1157 const VkMemoryAllocateInfo* pAllocateInfo,
1158 const VkAllocationCallbacks* pAllocator,
1159 VkDeviceMemory* pMem)
1160 {
1161 ANV_FROM_HANDLE(anv_device, device, _device);
1162 struct anv_device_memory *mem;
1163 VkResult result;
1164
1165 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1166
1167 if (pAllocateInfo->allocationSize == 0) {
1168 /* Apparently, this is allowed */
1169 *pMem = VK_NULL_HANDLE;
1170 return VK_SUCCESS;
1171 }
1172
1173 /* We support exactly one memory heap. */
1174 assert(pAllocateInfo->memoryTypeIndex == 0 ||
1175 (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1176
1177 /* FINISHME: Fail if allocation request exceeds heap size. */
1178
1179 mem = anv_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1180 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1181 if (mem == NULL)
1182 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1183
1184 /* The kernel is going to give us whole pages anyway */
1185 uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1186
1187 result = anv_bo_init_new(&mem->bo, device, alloc_size);
1188 if (result != VK_SUCCESS)
1189 goto fail;
1190
1191 mem->type_index = pAllocateInfo->memoryTypeIndex;
1192
1193 *pMem = anv_device_memory_to_handle(mem);
1194
1195 return VK_SUCCESS;
1196
1197 fail:
1198 anv_free2(&device->alloc, pAllocator, mem);
1199
1200 return result;
1201 }
1202
1203 void anv_FreeMemory(
1204 VkDevice _device,
1205 VkDeviceMemory _mem,
1206 const VkAllocationCallbacks* pAllocator)
1207 {
1208 ANV_FROM_HANDLE(anv_device, device, _device);
1209 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1210
1211 if (mem == NULL)
1212 return;
1213
1214 if (mem->bo.map)
1215 anv_gem_munmap(mem->bo.map, mem->bo.size);
1216
1217 if (mem->bo.gem_handle != 0)
1218 anv_gem_close(device, mem->bo.gem_handle);
1219
1220 anv_free2(&device->alloc, pAllocator, mem);
1221 }
1222
1223 VkResult anv_MapMemory(
1224 VkDevice _device,
1225 VkDeviceMemory _memory,
1226 VkDeviceSize offset,
1227 VkDeviceSize size,
1228 VkMemoryMapFlags flags,
1229 void** ppData)
1230 {
1231 ANV_FROM_HANDLE(anv_device, device, _device);
1232 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1233
1234 if (mem == NULL) {
1235 *ppData = NULL;
1236 return VK_SUCCESS;
1237 }
1238
1239 if (size == VK_WHOLE_SIZE)
1240 size = mem->bo.size - offset;
1241
1242 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1243 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1244 * at a time is valid. We could just mmap up front and return an offset
1245 * pointer here, but that may exhaust virtual memory on 32 bit
1246 * userspace. */
1247
1248 uint32_t gem_flags = 0;
1249 if (!device->info.has_llc && mem->type_index == 0)
1250 gem_flags |= I915_MMAP_WC;
1251
1252 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1253 uint64_t map_offset = offset & ~4095ull;
1254 assert(offset >= map_offset);
1255 uint64_t map_size = (offset + size) - map_offset;
1256
1257 /* Let's map whole pages */
1258 map_size = align_u64(map_size, 4096);
1259
1260 mem->map = anv_gem_mmap(device, mem->bo.gem_handle,
1261 map_offset, map_size, gem_flags);
1262 mem->map_size = map_size;
1263
1264 *ppData = mem->map + (offset - map_offset);
1265
1266 return VK_SUCCESS;
1267 }
1268
1269 void anv_UnmapMemory(
1270 VkDevice _device,
1271 VkDeviceMemory _memory)
1272 {
1273 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1274
1275 if (mem == NULL)
1276 return;
1277
1278 anv_gem_munmap(mem->map, mem->map_size);
1279 }
1280
1281 static void
1282 clflush_mapped_ranges(struct anv_device *device,
1283 uint32_t count,
1284 const VkMappedMemoryRange *ranges)
1285 {
1286 for (uint32_t i = 0; i < count; i++) {
1287 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1288 void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1289 void *end;
1290
1291 if (ranges[i].offset + ranges[i].size > mem->map_size)
1292 end = mem->map + mem->map_size;
1293 else
1294 end = mem->map + ranges[i].offset + ranges[i].size;
1295
1296 while (p < end) {
1297 __builtin_ia32_clflush(p);
1298 p += CACHELINE_SIZE;
1299 }
1300 }
1301 }
1302
1303 VkResult anv_FlushMappedMemoryRanges(
1304 VkDevice _device,
1305 uint32_t memoryRangeCount,
1306 const VkMappedMemoryRange* pMemoryRanges)
1307 {
1308 ANV_FROM_HANDLE(anv_device, device, _device);
1309
1310 if (device->info.has_llc)
1311 return VK_SUCCESS;
1312
1313 /* Make sure the writes we're flushing have landed. */
1314 __builtin_ia32_mfence();
1315
1316 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1317
1318 return VK_SUCCESS;
1319 }
1320
1321 VkResult anv_InvalidateMappedMemoryRanges(
1322 VkDevice _device,
1323 uint32_t memoryRangeCount,
1324 const VkMappedMemoryRange* pMemoryRanges)
1325 {
1326 ANV_FROM_HANDLE(anv_device, device, _device);
1327
1328 if (device->info.has_llc)
1329 return VK_SUCCESS;
1330
1331 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1332
1333 /* Make sure no reads get moved up above the invalidate. */
1334 __builtin_ia32_mfence();
1335
1336 return VK_SUCCESS;
1337 }
1338
1339 void anv_GetBufferMemoryRequirements(
1340 VkDevice device,
1341 VkBuffer _buffer,
1342 VkMemoryRequirements* pMemoryRequirements)
1343 {
1344 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1345
1346 /* The Vulkan spec (git aaed022) says:
1347 *
1348 * memoryTypeBits is a bitfield and contains one bit set for every
1349 * supported memory type for the resource. The bit `1<<i` is set if and
1350 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1351 * structure for the physical device is supported.
1352 *
1353 * We support exactly one memory type.
1354 */
1355 pMemoryRequirements->memoryTypeBits = 1;
1356
1357 pMemoryRequirements->size = buffer->size;
1358 pMemoryRequirements->alignment = 16;
1359 }
1360
1361 void anv_GetImageMemoryRequirements(
1362 VkDevice device,
1363 VkImage _image,
1364 VkMemoryRequirements* pMemoryRequirements)
1365 {
1366 ANV_FROM_HANDLE(anv_image, image, _image);
1367
1368 /* The Vulkan spec (git aaed022) says:
1369 *
1370 * memoryTypeBits is a bitfield and contains one bit set for every
1371 * supported memory type for the resource. The bit `1<<i` is set if and
1372 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1373 * structure for the physical device is supported.
1374 *
1375 * We support exactly one memory type.
1376 */
1377 pMemoryRequirements->memoryTypeBits = 1;
1378
1379 pMemoryRequirements->size = image->size;
1380 pMemoryRequirements->alignment = image->alignment;
1381 }
1382
1383 void anv_GetImageSparseMemoryRequirements(
1384 VkDevice device,
1385 VkImage image,
1386 uint32_t* pSparseMemoryRequirementCount,
1387 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
1388 {
1389 stub();
1390 }
1391
1392 void anv_GetDeviceMemoryCommitment(
1393 VkDevice device,
1394 VkDeviceMemory memory,
1395 VkDeviceSize* pCommittedMemoryInBytes)
1396 {
1397 *pCommittedMemoryInBytes = 0;
1398 }
1399
1400 VkResult anv_BindBufferMemory(
1401 VkDevice device,
1402 VkBuffer _buffer,
1403 VkDeviceMemory _memory,
1404 VkDeviceSize memoryOffset)
1405 {
1406 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1407 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1408
1409 if (mem) {
1410 buffer->bo = &mem->bo;
1411 buffer->offset = memoryOffset;
1412 } else {
1413 buffer->bo = NULL;
1414 buffer->offset = 0;
1415 }
1416
1417 return VK_SUCCESS;
1418 }
1419
1420 VkResult anv_QueueBindSparse(
1421 VkQueue queue,
1422 uint32_t bindInfoCount,
1423 const VkBindSparseInfo* pBindInfo,
1424 VkFence fence)
1425 {
1426 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1427 }
1428
1429 VkResult anv_CreateFence(
1430 VkDevice _device,
1431 const VkFenceCreateInfo* pCreateInfo,
1432 const VkAllocationCallbacks* pAllocator,
1433 VkFence* pFence)
1434 {
1435 ANV_FROM_HANDLE(anv_device, device, _device);
1436 struct anv_bo fence_bo;
1437 struct anv_fence *fence;
1438 struct anv_batch batch;
1439 VkResult result;
1440
1441 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1442
1443 result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1444 if (result != VK_SUCCESS)
1445 return result;
1446
1447 /* Fences are small. Just store the CPU data structure in the BO. */
1448 fence = fence_bo.map;
1449 fence->bo = fence_bo;
1450
1451 /* Place the batch after the CPU data but on its own cache line. */
1452 const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1453 batch.next = batch.start = fence->bo.map + batch_offset;
1454 batch.end = fence->bo.map + fence->bo.size;
1455 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1456 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1457
1458 if (!device->info.has_llc) {
1459 assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1460 assert(batch.next - batch.start <= CACHELINE_SIZE);
1461 __builtin_ia32_mfence();
1462 __builtin_ia32_clflush(batch.start);
1463 }
1464
1465 fence->exec2_objects[0].handle = fence->bo.gem_handle;
1466 fence->exec2_objects[0].relocation_count = 0;
1467 fence->exec2_objects[0].relocs_ptr = 0;
1468 fence->exec2_objects[0].alignment = 0;
1469 fence->exec2_objects[0].offset = fence->bo.offset;
1470 fence->exec2_objects[0].flags = 0;
1471 fence->exec2_objects[0].rsvd1 = 0;
1472 fence->exec2_objects[0].rsvd2 = 0;
1473
1474 fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1475 fence->execbuf.buffer_count = 1;
1476 fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1477 fence->execbuf.batch_len = batch.next - batch.start;
1478 fence->execbuf.cliprects_ptr = 0;
1479 fence->execbuf.num_cliprects = 0;
1480 fence->execbuf.DR1 = 0;
1481 fence->execbuf.DR4 = 0;
1482
1483 fence->execbuf.flags =
1484 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1485 fence->execbuf.rsvd1 = device->context_id;
1486 fence->execbuf.rsvd2 = 0;
1487
1488 fence->ready = false;
1489
1490 *pFence = anv_fence_to_handle(fence);
1491
1492 return VK_SUCCESS;
1493 }
1494
1495 void anv_DestroyFence(
1496 VkDevice _device,
1497 VkFence _fence,
1498 const VkAllocationCallbacks* pAllocator)
1499 {
1500 ANV_FROM_HANDLE(anv_device, device, _device);
1501 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1502
1503 assert(fence->bo.map == fence);
1504 anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1505 }
1506
1507 VkResult anv_ResetFences(
1508 VkDevice _device,
1509 uint32_t fenceCount,
1510 const VkFence* pFences)
1511 {
1512 for (uint32_t i = 0; i < fenceCount; i++) {
1513 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1514 fence->ready = false;
1515 }
1516
1517 return VK_SUCCESS;
1518 }
1519
1520 VkResult anv_GetFenceStatus(
1521 VkDevice _device,
1522 VkFence _fence)
1523 {
1524 ANV_FROM_HANDLE(anv_device, device, _device);
1525 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1526 int64_t t = 0;
1527 int ret;
1528
1529 if (fence->ready)
1530 return VK_SUCCESS;
1531
1532 ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1533 if (ret == 0) {
1534 fence->ready = true;
1535 return VK_SUCCESS;
1536 }
1537
1538 return VK_NOT_READY;
1539 }
1540
1541 VkResult anv_WaitForFences(
1542 VkDevice _device,
1543 uint32_t fenceCount,
1544 const VkFence* pFences,
1545 VkBool32 waitAll,
1546 uint64_t timeout)
1547 {
1548 ANV_FROM_HANDLE(anv_device, device, _device);
1549
1550 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1551 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1552 * for a couple of kernel releases. Since there's no way to know
1553 * whether or not the kernel we're using is one of the broken ones, the
1554 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1555 * maximum timeout from 584 years to 292 years - likely not a big deal.
1556 */
1557 if (timeout > INT64_MAX)
1558 timeout = INT64_MAX;
1559
1560 int64_t t = timeout;
1561
1562 /* FIXME: handle !waitAll */
1563
1564 for (uint32_t i = 0; i < fenceCount; i++) {
1565 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1566 int ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1567 if (ret == -1 && errno == ETIME) {
1568 return VK_TIMEOUT;
1569 } else if (ret == -1) {
1570 /* We don't know the real error. */
1571 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1572 "gem wait failed: %m");
1573 }
1574 }
1575
1576 return VK_SUCCESS;
1577 }
1578
1579 // Queue semaphore functions
1580
1581 VkResult anv_CreateSemaphore(
1582 VkDevice device,
1583 const VkSemaphoreCreateInfo* pCreateInfo,
1584 const VkAllocationCallbacks* pAllocator,
1585 VkSemaphore* pSemaphore)
1586 {
1587 /* The DRM execbuffer ioctl always execute in-oder, even between different
1588 * rings. As such, there's nothing to do for the user space semaphore.
1589 */
1590
1591 *pSemaphore = (VkSemaphore)1;
1592
1593 return VK_SUCCESS;
1594 }
1595
1596 void anv_DestroySemaphore(
1597 VkDevice device,
1598 VkSemaphore semaphore,
1599 const VkAllocationCallbacks* pAllocator)
1600 {
1601 }
1602
1603 // Event functions
1604
1605 VkResult anv_CreateEvent(
1606 VkDevice _device,
1607 const VkEventCreateInfo* pCreateInfo,
1608 const VkAllocationCallbacks* pAllocator,
1609 VkEvent* pEvent)
1610 {
1611 ANV_FROM_HANDLE(anv_device, device, _device);
1612 struct anv_state state;
1613 struct anv_event *event;
1614
1615 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1616
1617 state = anv_state_pool_alloc(&device->dynamic_state_pool,
1618 sizeof(*event), 8);
1619 event = state.map;
1620 event->state = state;
1621 event->semaphore = VK_EVENT_RESET;
1622
1623 if (!device->info.has_llc) {
1624 /* Make sure the writes we're flushing have landed. */
1625 __builtin_ia32_mfence();
1626 __builtin_ia32_clflush(event);
1627 }
1628
1629 *pEvent = anv_event_to_handle(event);
1630
1631 return VK_SUCCESS;
1632 }
1633
1634 void anv_DestroyEvent(
1635 VkDevice _device,
1636 VkEvent _event,
1637 const VkAllocationCallbacks* pAllocator)
1638 {
1639 ANV_FROM_HANDLE(anv_device, device, _device);
1640 ANV_FROM_HANDLE(anv_event, event, _event);
1641
1642 anv_state_pool_free(&device->dynamic_state_pool, event->state);
1643 }
1644
1645 VkResult anv_GetEventStatus(
1646 VkDevice _device,
1647 VkEvent _event)
1648 {
1649 ANV_FROM_HANDLE(anv_device, device, _device);
1650 ANV_FROM_HANDLE(anv_event, event, _event);
1651
1652 if (!device->info.has_llc) {
1653 /* Invalidate read cache before reading event written by GPU. */
1654 __builtin_ia32_clflush(event);
1655 __builtin_ia32_mfence();
1656
1657 }
1658
1659 return event->semaphore;
1660 }
1661
1662 VkResult anv_SetEvent(
1663 VkDevice _device,
1664 VkEvent _event)
1665 {
1666 ANV_FROM_HANDLE(anv_device, device, _device);
1667 ANV_FROM_HANDLE(anv_event, event, _event);
1668
1669 event->semaphore = VK_EVENT_SET;
1670
1671 if (!device->info.has_llc) {
1672 /* Make sure the writes we're flushing have landed. */
1673 __builtin_ia32_mfence();
1674 __builtin_ia32_clflush(event);
1675 }
1676
1677 return VK_SUCCESS;
1678 }
1679
1680 VkResult anv_ResetEvent(
1681 VkDevice _device,
1682 VkEvent _event)
1683 {
1684 ANV_FROM_HANDLE(anv_device, device, _device);
1685 ANV_FROM_HANDLE(anv_event, event, _event);
1686
1687 event->semaphore = VK_EVENT_RESET;
1688
1689 if (!device->info.has_llc) {
1690 /* Make sure the writes we're flushing have landed. */
1691 __builtin_ia32_mfence();
1692 __builtin_ia32_clflush(event);
1693 }
1694
1695 return VK_SUCCESS;
1696 }
1697
1698 // Buffer functions
1699
1700 VkResult anv_CreateBuffer(
1701 VkDevice _device,
1702 const VkBufferCreateInfo* pCreateInfo,
1703 const VkAllocationCallbacks* pAllocator,
1704 VkBuffer* pBuffer)
1705 {
1706 ANV_FROM_HANDLE(anv_device, device, _device);
1707 struct anv_buffer *buffer;
1708
1709 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1710
1711 buffer = anv_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1712 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1713 if (buffer == NULL)
1714 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1715
1716 buffer->size = pCreateInfo->size;
1717 buffer->usage = pCreateInfo->usage;
1718 buffer->bo = NULL;
1719 buffer->offset = 0;
1720
1721 *pBuffer = anv_buffer_to_handle(buffer);
1722
1723 return VK_SUCCESS;
1724 }
1725
1726 void anv_DestroyBuffer(
1727 VkDevice _device,
1728 VkBuffer _buffer,
1729 const VkAllocationCallbacks* pAllocator)
1730 {
1731 ANV_FROM_HANDLE(anv_device, device, _device);
1732 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1733
1734 anv_free2(&device->alloc, pAllocator, buffer);
1735 }
1736
1737 void
1738 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1739 enum isl_format format,
1740 uint32_t offset, uint32_t range, uint32_t stride)
1741 {
1742 isl_buffer_fill_state(&device->isl_dev, state.map,
1743 .address = offset,
1744 .mocs = device->default_mocs,
1745 .size = range,
1746 .format = format,
1747 .stride = stride);
1748
1749 if (!device->info.has_llc)
1750 anv_state_clflush(state);
1751 }
1752
1753 void anv_DestroySampler(
1754 VkDevice _device,
1755 VkSampler _sampler,
1756 const VkAllocationCallbacks* pAllocator)
1757 {
1758 ANV_FROM_HANDLE(anv_device, device, _device);
1759 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
1760
1761 anv_free2(&device->alloc, pAllocator, sampler);
1762 }
1763
1764 VkResult anv_CreateFramebuffer(
1765 VkDevice _device,
1766 const VkFramebufferCreateInfo* pCreateInfo,
1767 const VkAllocationCallbacks* pAllocator,
1768 VkFramebuffer* pFramebuffer)
1769 {
1770 ANV_FROM_HANDLE(anv_device, device, _device);
1771 struct anv_framebuffer *framebuffer;
1772
1773 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1774
1775 size_t size = sizeof(*framebuffer) +
1776 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
1777 framebuffer = anv_alloc2(&device->alloc, pAllocator, size, 8,
1778 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1779 if (framebuffer == NULL)
1780 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1781
1782 framebuffer->attachment_count = pCreateInfo->attachmentCount;
1783 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1784 VkImageView _iview = pCreateInfo->pAttachments[i];
1785 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
1786 }
1787
1788 framebuffer->width = pCreateInfo->width;
1789 framebuffer->height = pCreateInfo->height;
1790 framebuffer->layers = pCreateInfo->layers;
1791
1792 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
1793
1794 return VK_SUCCESS;
1795 }
1796
1797 void anv_DestroyFramebuffer(
1798 VkDevice _device,
1799 VkFramebuffer _fb,
1800 const VkAllocationCallbacks* pAllocator)
1801 {
1802 ANV_FROM_HANDLE(anv_device, device, _device);
1803 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
1804
1805 anv_free2(&device->alloc, pAllocator, fb);
1806 }