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 <dlfcn.h>
25 #include <assert.h>
26 #include <stdbool.h>
27 #include <string.h>
28 #include <sys/mman.h>
29 #include <sys/stat.h>
30 #include <unistd.h>
31 #include <fcntl.h>
32
33 #include "anv_private.h"
34 #include "util/strtod.h"
35 #include "util/debug.h"
36
37 #include "genxml/gen7_pack.h"
38
39 struct anv_dispatch_table dtable;
40
41 static void
compiler_debug_log(void * data,const char * fmt,...)42 compiler_debug_log(void *data, const char *fmt, ...)
43 { }
44
45 static void
compiler_perf_log(void * data,const char * fmt,...)46 compiler_perf_log(void *data, const char *fmt, ...)
47 {
48 va_list args;
49 va_start(args, fmt);
50
51 if (unlikely(INTEL_DEBUG & DEBUG_PERF))
52 vfprintf(stderr, fmt, args);
53
54 va_end(args);
55 }
56
57 static bool
anv_get_function_timestamp(void * ptr,uint32_t * timestamp)58 anv_get_function_timestamp(void *ptr, uint32_t* timestamp)
59 {
60 Dl_info info;
61 struct stat st;
62 if (!dladdr(ptr, &info) || !info.dli_fname)
63 return false;
64
65 if (stat(info.dli_fname, &st))
66 return false;
67
68 *timestamp = st.st_mtim.tv_sec;
69 return true;
70 }
71
72 static bool
anv_device_get_cache_uuid(void * uuid)73 anv_device_get_cache_uuid(void *uuid)
74 {
75 uint32_t timestamp;
76
77 memset(uuid, 0, VK_UUID_SIZE);
78 if (!anv_get_function_timestamp(anv_device_get_cache_uuid, ×tamp))
79 return false;
80
81 snprintf(uuid, VK_UUID_SIZE, "anv-%d", timestamp);
82 return true;
83 }
84
85 static VkResult
anv_physical_device_init(struct anv_physical_device * device,struct anv_instance * instance,const char * path)86 anv_physical_device_init(struct anv_physical_device *device,
87 struct anv_instance *instance,
88 const char *path)
89 {
90 VkResult result;
91 int fd;
92
93 fd = open(path, O_RDWR | O_CLOEXEC);
94 if (fd < 0)
95 return vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
96
97 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
98 device->instance = instance;
99
100 assert(strlen(path) < ARRAY_SIZE(device->path));
101 strncpy(device->path, path, ARRAY_SIZE(device->path));
102
103 device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
104 if (!device->chipset_id) {
105 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
106 goto fail;
107 }
108
109 device->name = gen_get_device_name(device->chipset_id);
110 if (!gen_get_device_info(device->chipset_id, &device->info)) {
111 result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
112 goto fail;
113 }
114
115 if (device->info.is_haswell) {
116 fprintf(stderr, "WARNING: Haswell Vulkan support is incomplete\n");
117 } else if (device->info.gen == 7 && !device->info.is_baytrail) {
118 fprintf(stderr, "WARNING: Ivy Bridge Vulkan support is incomplete\n");
119 } else if (device->info.gen == 7 && device->info.is_baytrail) {
120 fprintf(stderr, "WARNING: Bay Trail Vulkan support is incomplete\n");
121 } else if (device->info.gen >= 8) {
122 /* Broadwell, Cherryview, Skylake, Broxton, Kabylake is as fully
123 * supported as anything */
124 } else {
125 result = vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
126 "Vulkan not yet supported on %s", device->name);
127 goto fail;
128 }
129
130 device->cmd_parser_version = -1;
131 if (device->info.gen == 7) {
132 device->cmd_parser_version =
133 anv_gem_get_param(fd, I915_PARAM_CMD_PARSER_VERSION);
134 if (device->cmd_parser_version == -1) {
135 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
136 "failed to get command parser version");
137 goto fail;
138 }
139 }
140
141 if (anv_gem_get_aperture(fd, &device->aperture_size) == -1) {
142 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
143 "failed to get aperture size: %m");
144 goto fail;
145 }
146
147 if (!anv_gem_get_param(fd, I915_PARAM_HAS_WAIT_TIMEOUT)) {
148 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
149 "kernel missing gem wait");
150 goto fail;
151 }
152
153 if (!anv_gem_get_param(fd, I915_PARAM_HAS_EXECBUF2)) {
154 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
155 "kernel missing execbuf2");
156 goto fail;
157 }
158
159 if (!device->info.has_llc &&
160 anv_gem_get_param(fd, I915_PARAM_MMAP_VERSION) < 1) {
161 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
162 "kernel missing wc mmap");
163 goto fail;
164 }
165
166 if (!anv_device_get_cache_uuid(device->uuid)) {
167 result = vk_errorf(VK_ERROR_INITIALIZATION_FAILED,
168 "cannot generate UUID");
169 goto fail;
170 }
171 bool swizzled = anv_gem_get_bit6_swizzle(fd, I915_TILING_X);
172
173 /* GENs prior to 8 do not support EU/Subslice info */
174 if (device->info.gen >= 8) {
175 device->subslice_total = anv_gem_get_param(fd, I915_PARAM_SUBSLICE_TOTAL);
176 device->eu_total = anv_gem_get_param(fd, I915_PARAM_EU_TOTAL);
177
178 /* Without this information, we cannot get the right Braswell
179 * brandstrings, and we have to use conservative numbers for GPGPU on
180 * many platforms, but otherwise, things will just work.
181 */
182 if (device->subslice_total < 1 || device->eu_total < 1) {
183 fprintf(stderr, "WARNING: Kernel 4.1 required to properly"
184 " query GPU properties.\n");
185 }
186 } else if (device->info.gen == 7) {
187 device->subslice_total = 1 << (device->info.gt - 1);
188 }
189
190 if (device->info.is_cherryview &&
191 device->subslice_total > 0 && device->eu_total > 0) {
192 /* Logical CS threads = EUs per subslice * 7 threads per EU */
193 uint32_t max_cs_threads = device->eu_total / device->subslice_total * 7;
194
195 /* Fuse configurations may give more threads than expected, never less. */
196 if (max_cs_threads > device->info.max_cs_threads)
197 device->info.max_cs_threads = max_cs_threads;
198 }
199
200 brw_process_intel_debug_variable();
201
202 device->compiler = brw_compiler_create(NULL, &device->info);
203 if (device->compiler == NULL) {
204 result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
205 goto fail;
206 }
207 device->compiler->shader_debug_log = compiler_debug_log;
208 device->compiler->shader_perf_log = compiler_perf_log;
209
210 result = anv_init_wsi(device);
211 if (result != VK_SUCCESS) {
212 ralloc_free(device->compiler);
213 goto fail;
214 }
215
216 isl_device_init(&device->isl_dev, &device->info, swizzled);
217
218 close(fd);
219 return VK_SUCCESS;
220
221 fail:
222 close(fd);
223 return result;
224 }
225
226 static void
anv_physical_device_finish(struct anv_physical_device * device)227 anv_physical_device_finish(struct anv_physical_device *device)
228 {
229 anv_finish_wsi(device);
230 ralloc_free(device->compiler);
231 }
232
233 static const VkExtensionProperties global_extensions[] = {
234 {
235 .extensionName = VK_KHR_SURFACE_EXTENSION_NAME,
236 .specVersion = 25,
237 },
238 #ifdef VK_USE_PLATFORM_XCB_KHR
239 {
240 .extensionName = VK_KHR_XCB_SURFACE_EXTENSION_NAME,
241 .specVersion = 6,
242 },
243 #endif
244 #ifdef VK_USE_PLATFORM_XLIB_KHR
245 {
246 .extensionName = VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
247 .specVersion = 6,
248 },
249 #endif
250 #ifdef VK_USE_PLATFORM_WAYLAND_KHR
251 {
252 .extensionName = VK_KHR_WAYLAND_SURFACE_EXTENSION_NAME,
253 .specVersion = 5,
254 },
255 #endif
256 };
257
258 static const VkExtensionProperties device_extensions[] = {
259 {
260 .extensionName = VK_KHR_SWAPCHAIN_EXTENSION_NAME,
261 .specVersion = 68,
262 },
263 {
264 .extensionName = VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
265 .specVersion = 1,
266 }
267 };
268
269 static void *
default_alloc_func(void * pUserData,size_t size,size_t align,VkSystemAllocationScope allocationScope)270 default_alloc_func(void *pUserData, size_t size, size_t align,
271 VkSystemAllocationScope allocationScope)
272 {
273 return malloc(size);
274 }
275
276 static void *
default_realloc_func(void * pUserData,void * pOriginal,size_t size,size_t align,VkSystemAllocationScope allocationScope)277 default_realloc_func(void *pUserData, void *pOriginal, size_t size,
278 size_t align, VkSystemAllocationScope allocationScope)
279 {
280 return realloc(pOriginal, size);
281 }
282
283 static void
default_free_func(void * pUserData,void * pMemory)284 default_free_func(void *pUserData, void *pMemory)
285 {
286 free(pMemory);
287 }
288
289 static const VkAllocationCallbacks default_alloc = {
290 .pUserData = NULL,
291 .pfnAllocation = default_alloc_func,
292 .pfnReallocation = default_realloc_func,
293 .pfnFree = default_free_func,
294 };
295
anv_CreateInstance(const VkInstanceCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkInstance * pInstance)296 VkResult anv_CreateInstance(
297 const VkInstanceCreateInfo* pCreateInfo,
298 const VkAllocationCallbacks* pAllocator,
299 VkInstance* pInstance)
300 {
301 struct anv_instance *instance;
302
303 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
304
305 uint32_t client_version;
306 if (pCreateInfo->pApplicationInfo &&
307 pCreateInfo->pApplicationInfo->apiVersion != 0) {
308 client_version = pCreateInfo->pApplicationInfo->apiVersion;
309 } else {
310 client_version = VK_MAKE_VERSION(1, 0, 0);
311 }
312
313 if (VK_MAKE_VERSION(1, 0, 0) > client_version ||
314 client_version > VK_MAKE_VERSION(1, 0, 0xfff)) {
315 return vk_errorf(VK_ERROR_INCOMPATIBLE_DRIVER,
316 "Client requested version %d.%d.%d",
317 VK_VERSION_MAJOR(client_version),
318 VK_VERSION_MINOR(client_version),
319 VK_VERSION_PATCH(client_version));
320 }
321
322 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
323 bool found = false;
324 for (uint32_t j = 0; j < ARRAY_SIZE(global_extensions); j++) {
325 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
326 global_extensions[j].extensionName) == 0) {
327 found = true;
328 break;
329 }
330 }
331 if (!found)
332 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
333 }
334
335 instance = vk_alloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
336 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
337 if (!instance)
338 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
339
340 instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
341
342 if (pAllocator)
343 instance->alloc = *pAllocator;
344 else
345 instance->alloc = default_alloc;
346
347 instance->apiVersion = client_version;
348 instance->physicalDeviceCount = -1;
349
350 _mesa_locale_init();
351
352 VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
353
354 *pInstance = anv_instance_to_handle(instance);
355
356 return VK_SUCCESS;
357 }
358
anv_DestroyInstance(VkInstance _instance,const VkAllocationCallbacks * pAllocator)359 void anv_DestroyInstance(
360 VkInstance _instance,
361 const VkAllocationCallbacks* pAllocator)
362 {
363 ANV_FROM_HANDLE(anv_instance, instance, _instance);
364
365 if (!instance)
366 return;
367
368 if (instance->physicalDeviceCount > 0) {
369 /* We support at most one physical device. */
370 assert(instance->physicalDeviceCount == 1);
371 anv_physical_device_finish(&instance->physicalDevice);
372 }
373
374 VG(VALGRIND_DESTROY_MEMPOOL(instance));
375
376 _mesa_locale_fini();
377
378 vk_free(&instance->alloc, instance);
379 }
380
anv_EnumeratePhysicalDevices(VkInstance _instance,uint32_t * pPhysicalDeviceCount,VkPhysicalDevice * pPhysicalDevices)381 VkResult anv_EnumeratePhysicalDevices(
382 VkInstance _instance,
383 uint32_t* pPhysicalDeviceCount,
384 VkPhysicalDevice* pPhysicalDevices)
385 {
386 ANV_FROM_HANDLE(anv_instance, instance, _instance);
387 VkResult result;
388
389 if (instance->physicalDeviceCount < 0) {
390 char path[20];
391 for (unsigned i = 0; i < 8; i++) {
392 snprintf(path, sizeof(path), "/dev/dri/renderD%d", 128 + i);
393 result = anv_physical_device_init(&instance->physicalDevice,
394 instance, path);
395 if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
396 break;
397 }
398
399 if (result == VK_ERROR_INCOMPATIBLE_DRIVER) {
400 instance->physicalDeviceCount = 0;
401 } else if (result == VK_SUCCESS) {
402 instance->physicalDeviceCount = 1;
403 } else {
404 return result;
405 }
406 }
407
408 /* pPhysicalDeviceCount is an out parameter if pPhysicalDevices is NULL;
409 * otherwise it's an inout parameter.
410 *
411 * The Vulkan spec (git aaed022) says:
412 *
413 * pPhysicalDeviceCount is a pointer to an unsigned integer variable
414 * that is initialized with the number of devices the application is
415 * prepared to receive handles to. pname:pPhysicalDevices is pointer to
416 * an array of at least this many VkPhysicalDevice handles [...].
417 *
418 * Upon success, if pPhysicalDevices is NULL, vkEnumeratePhysicalDevices
419 * overwrites the contents of the variable pointed to by
420 * pPhysicalDeviceCount with the number of physical devices in in the
421 * instance; otherwise, vkEnumeratePhysicalDevices overwrites
422 * pPhysicalDeviceCount with the number of physical handles written to
423 * pPhysicalDevices.
424 */
425 if (!pPhysicalDevices) {
426 *pPhysicalDeviceCount = instance->physicalDeviceCount;
427 } else if (*pPhysicalDeviceCount >= 1) {
428 pPhysicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice);
429 *pPhysicalDeviceCount = 1;
430 } else if (*pPhysicalDeviceCount < instance->physicalDeviceCount) {
431 return VK_INCOMPLETE;
432 } else {
433 *pPhysicalDeviceCount = 0;
434 }
435
436 return VK_SUCCESS;
437 }
438
anv_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,VkPhysicalDeviceFeatures * pFeatures)439 void anv_GetPhysicalDeviceFeatures(
440 VkPhysicalDevice physicalDevice,
441 VkPhysicalDeviceFeatures* pFeatures)
442 {
443 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
444
445 *pFeatures = (VkPhysicalDeviceFeatures) {
446 .robustBufferAccess = true,
447 .fullDrawIndexUint32 = true,
448 .imageCubeArray = true,
449 .independentBlend = true,
450 .geometryShader = true,
451 .tessellationShader = true,
452 .sampleRateShading = true,
453 .dualSrcBlend = true,
454 .logicOp = true,
455 .multiDrawIndirect = false,
456 .drawIndirectFirstInstance = true,
457 .depthClamp = true,
458 .depthBiasClamp = true,
459 .fillModeNonSolid = true,
460 .depthBounds = false,
461 .wideLines = true,
462 .largePoints = true,
463 .alphaToOne = true,
464 .multiViewport = true,
465 .samplerAnisotropy = true,
466 .textureCompressionETC2 = pdevice->info.gen >= 8 ||
467 pdevice->info.is_baytrail,
468 .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /* FINISHME CHV */
469 .textureCompressionBC = true,
470 .occlusionQueryPrecise = true,
471 .pipelineStatisticsQuery = false,
472 .fragmentStoresAndAtomics = true,
473 .shaderTessellationAndGeometryPointSize = true,
474 .shaderImageGatherExtended = true,
475 .shaderStorageImageExtendedFormats = true,
476 .shaderStorageImageMultisample = false,
477 .shaderStorageImageReadWithoutFormat = false,
478 .shaderStorageImageWriteWithoutFormat = false,
479 .shaderUniformBufferArrayDynamicIndexing = true,
480 .shaderSampledImageArrayDynamicIndexing = true,
481 .shaderStorageBufferArrayDynamicIndexing = true,
482 .shaderStorageImageArrayDynamicIndexing = true,
483 .shaderClipDistance = true,
484 .shaderCullDistance = true,
485 .shaderFloat64 = pdevice->info.gen >= 8,
486 .shaderInt64 = false,
487 .shaderInt16 = false,
488 .shaderResourceMinLod = false,
489 .variableMultisampleRate = false,
490 .inheritedQueries = false,
491 };
492
493 /* We can't do image stores in vec4 shaders */
494 pFeatures->vertexPipelineStoresAndAtomics =
495 pdevice->compiler->scalar_stage[MESA_SHADER_VERTEX] &&
496 pdevice->compiler->scalar_stage[MESA_SHADER_GEOMETRY];
497 }
498
anv_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,VkPhysicalDeviceProperties * pProperties)499 void anv_GetPhysicalDeviceProperties(
500 VkPhysicalDevice physicalDevice,
501 VkPhysicalDeviceProperties* pProperties)
502 {
503 ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice);
504 const struct gen_device_info *devinfo = &pdevice->info;
505
506 const float time_stamp_base = devinfo->gen >= 9 ? 83.333 : 80.0;
507
508 /* See assertions made when programming the buffer surface state. */
509 const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ?
510 (1ul << 30) : (1ul << 27);
511
512 VkSampleCountFlags sample_counts =
513 isl_device_get_sample_counts(&pdevice->isl_dev);
514
515 VkPhysicalDeviceLimits limits = {
516 .maxImageDimension1D = (1 << 14),
517 .maxImageDimension2D = (1 << 14),
518 .maxImageDimension3D = (1 << 11),
519 .maxImageDimensionCube = (1 << 14),
520 .maxImageArrayLayers = (1 << 11),
521 .maxTexelBufferElements = 128 * 1024 * 1024,
522 .maxUniformBufferRange = (1ul << 27),
523 .maxStorageBufferRange = max_raw_buffer_sz,
524 .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
525 .maxMemoryAllocationCount = UINT32_MAX,
526 .maxSamplerAllocationCount = 64 * 1024,
527 .bufferImageGranularity = 64, /* A cache line */
528 .sparseAddressSpaceSize = 0,
529 .maxBoundDescriptorSets = MAX_SETS,
530 .maxPerStageDescriptorSamplers = 64,
531 .maxPerStageDescriptorUniformBuffers = 64,
532 .maxPerStageDescriptorStorageBuffers = 64,
533 .maxPerStageDescriptorSampledImages = 64,
534 .maxPerStageDescriptorStorageImages = 64,
535 .maxPerStageDescriptorInputAttachments = 64,
536 .maxPerStageResources = 128,
537 .maxDescriptorSetSamplers = 256,
538 .maxDescriptorSetUniformBuffers = 256,
539 .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
540 .maxDescriptorSetStorageBuffers = 256,
541 .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2,
542 .maxDescriptorSetSampledImages = 256,
543 .maxDescriptorSetStorageImages = 256,
544 .maxDescriptorSetInputAttachments = 256,
545 .maxVertexInputAttributes = 32,
546 .maxVertexInputBindings = 32,
547 .maxVertexInputAttributeOffset = 2047,
548 .maxVertexInputBindingStride = 2048,
549 .maxVertexOutputComponents = 128,
550 .maxTessellationGenerationLevel = 64,
551 .maxTessellationPatchSize = 32,
552 .maxTessellationControlPerVertexInputComponents = 128,
553 .maxTessellationControlPerVertexOutputComponents = 128,
554 .maxTessellationControlPerPatchOutputComponents = 128,
555 .maxTessellationControlTotalOutputComponents = 2048,
556 .maxTessellationEvaluationInputComponents = 128,
557 .maxTessellationEvaluationOutputComponents = 128,
558 .maxGeometryShaderInvocations = 32,
559 .maxGeometryInputComponents = 64,
560 .maxGeometryOutputComponents = 128,
561 .maxGeometryOutputVertices = 256,
562 .maxGeometryTotalOutputComponents = 1024,
563 .maxFragmentInputComponents = 128,
564 .maxFragmentOutputAttachments = 8,
565 .maxFragmentDualSrcAttachments = 1,
566 .maxFragmentCombinedOutputResources = 8,
567 .maxComputeSharedMemorySize = 32768,
568 .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
569 .maxComputeWorkGroupInvocations = 16 * devinfo->max_cs_threads,
570 .maxComputeWorkGroupSize = {
571 16 * devinfo->max_cs_threads,
572 16 * devinfo->max_cs_threads,
573 16 * devinfo->max_cs_threads,
574 },
575 .subPixelPrecisionBits = 4 /* FIXME */,
576 .subTexelPrecisionBits = 4 /* FIXME */,
577 .mipmapPrecisionBits = 4 /* FIXME */,
578 .maxDrawIndexedIndexValue = UINT32_MAX,
579 .maxDrawIndirectCount = UINT32_MAX,
580 .maxSamplerLodBias = 16,
581 .maxSamplerAnisotropy = 16,
582 .maxViewports = MAX_VIEWPORTS,
583 .maxViewportDimensions = { (1 << 14), (1 << 14) },
584 .viewportBoundsRange = { INT16_MIN, INT16_MAX },
585 .viewportSubPixelBits = 13, /* We take a float? */
586 .minMemoryMapAlignment = 4096, /* A page */
587 .minTexelBufferOffsetAlignment = 1,
588 .minUniformBufferOffsetAlignment = 16,
589 .minStorageBufferOffsetAlignment = 4,
590 .minTexelOffset = -8,
591 .maxTexelOffset = 7,
592 .minTexelGatherOffset = -32,
593 .maxTexelGatherOffset = 31,
594 .minInterpolationOffset = -0.5,
595 .maxInterpolationOffset = 0.4375,
596 .subPixelInterpolationOffsetBits = 4,
597 .maxFramebufferWidth = (1 << 14),
598 .maxFramebufferHeight = (1 << 14),
599 .maxFramebufferLayers = (1 << 11),
600 .framebufferColorSampleCounts = sample_counts,
601 .framebufferDepthSampleCounts = sample_counts,
602 .framebufferStencilSampleCounts = sample_counts,
603 .framebufferNoAttachmentsSampleCounts = sample_counts,
604 .maxColorAttachments = MAX_RTS,
605 .sampledImageColorSampleCounts = sample_counts,
606 .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
607 .sampledImageDepthSampleCounts = sample_counts,
608 .sampledImageStencilSampleCounts = sample_counts,
609 .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
610 .maxSampleMaskWords = 1,
611 .timestampComputeAndGraphics = false,
612 .timestampPeriod = time_stamp_base,
613 .maxClipDistances = 8,
614 .maxCullDistances = 8,
615 .maxCombinedClipAndCullDistances = 8,
616 .discreteQueuePriorities = 1,
617 .pointSizeRange = { 0.125, 255.875 },
618 .lineWidthRange = { 0.0, 7.9921875 },
619 .pointSizeGranularity = (1.0 / 8.0),
620 .lineWidthGranularity = (1.0 / 128.0),
621 .strictLines = false, /* FINISHME */
622 .standardSampleLocations = true,
623 .optimalBufferCopyOffsetAlignment = 128,
624 .optimalBufferCopyRowPitchAlignment = 128,
625 .nonCoherentAtomSize = 64,
626 };
627
628 *pProperties = (VkPhysicalDeviceProperties) {
629 .apiVersion = VK_MAKE_VERSION(1, 0, 5),
630 .driverVersion = 1,
631 .vendorID = 0x8086,
632 .deviceID = pdevice->chipset_id,
633 .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
634 .limits = limits,
635 .sparseProperties = {0}, /* Broadwell doesn't do sparse. */
636 };
637
638 strcpy(pProperties->deviceName, pdevice->name);
639 memcpy(pProperties->pipelineCacheUUID, pdevice->uuid, VK_UUID_SIZE);
640 }
641
anv_GetPhysicalDeviceQueueFamilyProperties(VkPhysicalDevice physicalDevice,uint32_t * pCount,VkQueueFamilyProperties * pQueueFamilyProperties)642 void anv_GetPhysicalDeviceQueueFamilyProperties(
643 VkPhysicalDevice physicalDevice,
644 uint32_t* pCount,
645 VkQueueFamilyProperties* pQueueFamilyProperties)
646 {
647 if (pQueueFamilyProperties == NULL) {
648 *pCount = 1;
649 return;
650 }
651
652 /* The spec implicitly allows the incoming count to be 0. From the Vulkan
653 * 1.0.38 spec, Section 4.1 Physical Devices:
654 *
655 * If the value referenced by pQueueFamilyPropertyCount is not 0 [then
656 * do stuff].
657 */
658 if (*pCount == 0)
659 return;
660
661 *pQueueFamilyProperties = (VkQueueFamilyProperties) {
662 .queueFlags = VK_QUEUE_GRAPHICS_BIT |
663 VK_QUEUE_COMPUTE_BIT |
664 VK_QUEUE_TRANSFER_BIT,
665 .queueCount = 1,
666 .timestampValidBits = 36, /* XXX: Real value here */
667 .minImageTransferGranularity = (VkExtent3D) { 1, 1, 1 },
668 };
669
670 *pCount = 1;
671 }
672
anv_GetPhysicalDeviceMemoryProperties(VkPhysicalDevice physicalDevice,VkPhysicalDeviceMemoryProperties * pMemoryProperties)673 void anv_GetPhysicalDeviceMemoryProperties(
674 VkPhysicalDevice physicalDevice,
675 VkPhysicalDeviceMemoryProperties* pMemoryProperties)
676 {
677 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
678 VkDeviceSize heap_size;
679
680 /* Reserve some wiggle room for the driver by exposing only 75% of the
681 * aperture to the heap.
682 */
683 heap_size = 3 * physical_device->aperture_size / 4;
684
685 if (physical_device->info.has_llc) {
686 /* Big core GPUs share LLC with the CPU and thus one memory type can be
687 * both cached and coherent at the same time.
688 */
689 pMemoryProperties->memoryTypeCount = 1;
690 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
691 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
692 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
693 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
694 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
695 .heapIndex = 0,
696 };
697 } else {
698 /* The spec requires that we expose a host-visible, coherent memory
699 * type, but Atom GPUs don't share LLC. Thus we offer two memory types
700 * to give the application a choice between cached, but not coherent and
701 * coherent but uncached (WC though).
702 */
703 pMemoryProperties->memoryTypeCount = 2;
704 pMemoryProperties->memoryTypes[0] = (VkMemoryType) {
705 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
706 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
707 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
708 .heapIndex = 0,
709 };
710 pMemoryProperties->memoryTypes[1] = (VkMemoryType) {
711 .propertyFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
712 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
713 VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
714 .heapIndex = 0,
715 };
716 }
717
718 pMemoryProperties->memoryHeapCount = 1;
719 pMemoryProperties->memoryHeaps[0] = (VkMemoryHeap) {
720 .size = heap_size,
721 .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT,
722 };
723 }
724
anv_GetInstanceProcAddr(VkInstance instance,const char * pName)725 PFN_vkVoidFunction anv_GetInstanceProcAddr(
726 VkInstance instance,
727 const char* pName)
728 {
729 return anv_lookup_entrypoint(NULL, pName);
730 }
731
732 /* With version 1+ of the loader interface the ICD should expose
733 * vk_icdGetInstanceProcAddr to work around certain LD_PRELOAD issues seen in apps.
734 */
735 PUBLIC
736 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
737 VkInstance instance,
738 const char* pName);
739
740 PUBLIC
vk_icdGetInstanceProcAddr(VkInstance instance,const char * pName)741 VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetInstanceProcAddr(
742 VkInstance instance,
743 const char* pName)
744 {
745 return anv_GetInstanceProcAddr(instance, pName);
746 }
747
anv_GetDeviceProcAddr(VkDevice _device,const char * pName)748 PFN_vkVoidFunction anv_GetDeviceProcAddr(
749 VkDevice _device,
750 const char* pName)
751 {
752 ANV_FROM_HANDLE(anv_device, device, _device);
753 return anv_lookup_entrypoint(&device->info, pName);
754 }
755
756 static void
anv_queue_init(struct anv_device * device,struct anv_queue * queue)757 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
758 {
759 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
760 queue->device = device;
761 queue->pool = &device->surface_state_pool;
762 }
763
764 static void
anv_queue_finish(struct anv_queue * queue)765 anv_queue_finish(struct anv_queue *queue)
766 {
767 }
768
769 static struct anv_state
anv_state_pool_emit_data(struct anv_state_pool * pool,size_t size,size_t align,const void * p)770 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
771 {
772 struct anv_state state;
773
774 state = anv_state_pool_alloc(pool, size, align);
775 memcpy(state.map, p, size);
776
777 if (!pool->block_pool->device->info.has_llc)
778 anv_state_clflush(state);
779
780 return state;
781 }
782
783 struct gen8_border_color {
784 union {
785 float float32[4];
786 uint32_t uint32[4];
787 };
788 /* Pad out to 64 bytes */
789 uint32_t _pad[12];
790 };
791
792 static void
anv_device_init_border_colors(struct anv_device * device)793 anv_device_init_border_colors(struct anv_device *device)
794 {
795 static const struct gen8_border_color border_colors[] = {
796 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
797 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
798 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
799 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
800 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
801 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
802 };
803
804 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
805 sizeof(border_colors), 64,
806 border_colors);
807 }
808
809 VkResult
anv_device_submit_simple_batch(struct anv_device * device,struct anv_batch * batch)810 anv_device_submit_simple_batch(struct anv_device *device,
811 struct anv_batch *batch)
812 {
813 struct drm_i915_gem_execbuffer2 execbuf;
814 struct drm_i915_gem_exec_object2 exec2_objects[1];
815 struct anv_bo bo, *exec_bos[1];
816 VkResult result = VK_SUCCESS;
817 uint32_t size;
818 int64_t timeout;
819 int ret;
820
821 /* Kernel driver requires 8 byte aligned batch length */
822 size = align_u32(batch->next - batch->start, 8);
823 result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
824 if (result != VK_SUCCESS)
825 return result;
826
827 memcpy(bo.map, batch->start, size);
828 if (!device->info.has_llc)
829 anv_clflush_range(bo.map, size);
830
831 exec_bos[0] = &bo;
832 exec2_objects[0].handle = bo.gem_handle;
833 exec2_objects[0].relocation_count = 0;
834 exec2_objects[0].relocs_ptr = 0;
835 exec2_objects[0].alignment = 0;
836 exec2_objects[0].offset = bo.offset;
837 exec2_objects[0].flags = 0;
838 exec2_objects[0].rsvd1 = 0;
839 exec2_objects[0].rsvd2 = 0;
840
841 execbuf.buffers_ptr = (uintptr_t) exec2_objects;
842 execbuf.buffer_count = 1;
843 execbuf.batch_start_offset = 0;
844 execbuf.batch_len = size;
845 execbuf.cliprects_ptr = 0;
846 execbuf.num_cliprects = 0;
847 execbuf.DR1 = 0;
848 execbuf.DR4 = 0;
849
850 execbuf.flags =
851 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
852 execbuf.rsvd1 = device->context_id;
853 execbuf.rsvd2 = 0;
854
855 result = anv_device_execbuf(device, &execbuf, exec_bos);
856 if (result != VK_SUCCESS)
857 goto fail;
858
859 timeout = INT64_MAX;
860 ret = anv_gem_wait(device, bo.gem_handle, &timeout);
861 if (ret != 0) {
862 /* We don't know the real error. */
863 result = vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
864 goto fail;
865 }
866
867 fail:
868 anv_bo_pool_free(&device->batch_bo_pool, &bo);
869
870 return result;
871 }
872
anv_CreateDevice(VkPhysicalDevice physicalDevice,const VkDeviceCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkDevice * pDevice)873 VkResult anv_CreateDevice(
874 VkPhysicalDevice physicalDevice,
875 const VkDeviceCreateInfo* pCreateInfo,
876 const VkAllocationCallbacks* pAllocator,
877 VkDevice* pDevice)
878 {
879 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
880 VkResult result;
881 struct anv_device *device;
882
883 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
884
885 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
886 bool found = false;
887 for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
888 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
889 device_extensions[j].extensionName) == 0) {
890 found = true;
891 break;
892 }
893 }
894 if (!found)
895 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
896 }
897
898 device = vk_alloc2(&physical_device->instance->alloc, pAllocator,
899 sizeof(*device), 8,
900 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
901 if (!device)
902 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
903
904 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
905 device->instance = physical_device->instance;
906 device->chipset_id = physical_device->chipset_id;
907
908 if (pAllocator)
909 device->alloc = *pAllocator;
910 else
911 device->alloc = physical_device->instance->alloc;
912
913 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
914 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
915 if (device->fd == -1) {
916 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
917 goto fail_device;
918 }
919
920 device->context_id = anv_gem_create_context(device);
921 if (device->context_id == -1) {
922 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
923 goto fail_fd;
924 }
925
926 device->info = physical_device->info;
927 device->isl_dev = physical_device->isl_dev;
928
929 /* On Broadwell and later, we can use batch chaining to more efficiently
930 * implement growing command buffers. Prior to Haswell, the kernel
931 * command parser gets in the way and we have to fall back to growing
932 * the batch.
933 */
934 device->can_chain_batches = device->info.gen >= 8;
935
936 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
937 pCreateInfo->pEnabledFeatures->robustBufferAccess;
938
939 pthread_mutex_init(&device->mutex, NULL);
940
941 pthread_condattr_t condattr;
942 pthread_condattr_init(&condattr);
943 pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC);
944 pthread_cond_init(&device->queue_submit, NULL);
945 pthread_condattr_destroy(&condattr);
946
947 anv_bo_pool_init(&device->batch_bo_pool, device);
948
949 anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
950
951 anv_state_pool_init(&device->dynamic_state_pool,
952 &device->dynamic_state_block_pool);
953
954 anv_block_pool_init(&device->instruction_block_pool, device, 1024 * 1024);
955 anv_state_pool_init(&device->instruction_state_pool,
956 &device->instruction_block_pool);
957
958 anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
959
960 anv_state_pool_init(&device->surface_state_pool,
961 &device->surface_state_block_pool);
962
963 anv_bo_init_new(&device->workaround_bo, device, 1024);
964
965 anv_scratch_pool_init(device, &device->scratch_pool);
966
967 anv_queue_init(device, &device->queue);
968
969 switch (device->info.gen) {
970 case 7:
971 if (!device->info.is_haswell)
972 result = gen7_init_device_state(device);
973 else
974 result = gen75_init_device_state(device);
975 break;
976 case 8:
977 result = gen8_init_device_state(device);
978 break;
979 case 9:
980 result = gen9_init_device_state(device);
981 break;
982 default:
983 /* Shouldn't get here as we don't create physical devices for any other
984 * gens. */
985 unreachable("unhandled gen");
986 }
987 if (result != VK_SUCCESS)
988 goto fail_fd;
989
990 anv_device_init_blorp(device);
991
992 anv_device_init_border_colors(device);
993
994 *pDevice = anv_device_to_handle(device);
995
996 return VK_SUCCESS;
997
998 fail_fd:
999 close(device->fd);
1000 fail_device:
1001 vk_free(&device->alloc, device);
1002
1003 return result;
1004 }
1005
anv_DestroyDevice(VkDevice _device,const VkAllocationCallbacks * pAllocator)1006 void anv_DestroyDevice(
1007 VkDevice _device,
1008 const VkAllocationCallbacks* pAllocator)
1009 {
1010 ANV_FROM_HANDLE(anv_device, device, _device);
1011
1012 if (!device)
1013 return;
1014
1015 anv_device_finish_blorp(device);
1016
1017 anv_queue_finish(&device->queue);
1018
1019 #ifdef HAVE_VALGRIND
1020 /* We only need to free these to prevent valgrind errors. The backing
1021 * BO will go away in a couple of lines so we don't actually leak.
1022 */
1023 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
1024 #endif
1025
1026 anv_scratch_pool_finish(device, &device->scratch_pool);
1027
1028 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
1029 anv_gem_close(device, device->workaround_bo.gem_handle);
1030
1031 anv_state_pool_finish(&device->surface_state_pool);
1032 anv_block_pool_finish(&device->surface_state_block_pool);
1033 anv_state_pool_finish(&device->instruction_state_pool);
1034 anv_block_pool_finish(&device->instruction_block_pool);
1035 anv_state_pool_finish(&device->dynamic_state_pool);
1036 anv_block_pool_finish(&device->dynamic_state_block_pool);
1037
1038 anv_bo_pool_finish(&device->batch_bo_pool);
1039
1040 pthread_cond_destroy(&device->queue_submit);
1041 pthread_mutex_destroy(&device->mutex);
1042
1043 anv_gem_destroy_context(device, device->context_id);
1044
1045 close(device->fd);
1046
1047 vk_free(&device->alloc, device);
1048 }
1049
anv_EnumerateInstanceExtensionProperties(const char * pLayerName,uint32_t * pPropertyCount,VkExtensionProperties * pProperties)1050 VkResult anv_EnumerateInstanceExtensionProperties(
1051 const char* pLayerName,
1052 uint32_t* pPropertyCount,
1053 VkExtensionProperties* pProperties)
1054 {
1055 if (pProperties == NULL) {
1056 *pPropertyCount = ARRAY_SIZE(global_extensions);
1057 return VK_SUCCESS;
1058 }
1059
1060 *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(global_extensions));
1061 typed_memcpy(pProperties, global_extensions, *pPropertyCount);
1062
1063 if (*pPropertyCount < ARRAY_SIZE(global_extensions))
1064 return VK_INCOMPLETE;
1065
1066 return VK_SUCCESS;
1067 }
1068
anv_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,const char * pLayerName,uint32_t * pPropertyCount,VkExtensionProperties * pProperties)1069 VkResult anv_EnumerateDeviceExtensionProperties(
1070 VkPhysicalDevice physicalDevice,
1071 const char* pLayerName,
1072 uint32_t* pPropertyCount,
1073 VkExtensionProperties* pProperties)
1074 {
1075 if (pProperties == NULL) {
1076 *pPropertyCount = ARRAY_SIZE(device_extensions);
1077 return VK_SUCCESS;
1078 }
1079
1080 *pPropertyCount = MIN2(*pPropertyCount, ARRAY_SIZE(device_extensions));
1081 typed_memcpy(pProperties, device_extensions, *pPropertyCount);
1082
1083 if (*pPropertyCount < ARRAY_SIZE(device_extensions))
1084 return VK_INCOMPLETE;
1085
1086 return VK_SUCCESS;
1087 }
1088
anv_EnumerateInstanceLayerProperties(uint32_t * pPropertyCount,VkLayerProperties * pProperties)1089 VkResult anv_EnumerateInstanceLayerProperties(
1090 uint32_t* pPropertyCount,
1091 VkLayerProperties* pProperties)
1092 {
1093 if (pProperties == NULL) {
1094 *pPropertyCount = 0;
1095 return VK_SUCCESS;
1096 }
1097
1098 /* None supported at this time */
1099 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1100 }
1101
anv_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,uint32_t * pPropertyCount,VkLayerProperties * pProperties)1102 VkResult anv_EnumerateDeviceLayerProperties(
1103 VkPhysicalDevice physicalDevice,
1104 uint32_t* pPropertyCount,
1105 VkLayerProperties* pProperties)
1106 {
1107 if (pProperties == NULL) {
1108 *pPropertyCount = 0;
1109 return VK_SUCCESS;
1110 }
1111
1112 /* None supported at this time */
1113 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1114 }
1115
anv_GetDeviceQueue(VkDevice _device,uint32_t queueNodeIndex,uint32_t queueIndex,VkQueue * pQueue)1116 void anv_GetDeviceQueue(
1117 VkDevice _device,
1118 uint32_t queueNodeIndex,
1119 uint32_t queueIndex,
1120 VkQueue* pQueue)
1121 {
1122 ANV_FROM_HANDLE(anv_device, device, _device);
1123
1124 assert(queueIndex == 0);
1125
1126 *pQueue = anv_queue_to_handle(&device->queue);
1127 }
1128
1129 VkResult
anv_device_execbuf(struct anv_device * device,struct drm_i915_gem_execbuffer2 * execbuf,struct anv_bo ** execbuf_bos)1130 anv_device_execbuf(struct anv_device *device,
1131 struct drm_i915_gem_execbuffer2 *execbuf,
1132 struct anv_bo **execbuf_bos)
1133 {
1134 int ret = anv_gem_execbuffer(device, execbuf);
1135 if (ret != 0) {
1136 /* We don't know the real error. */
1137 return vk_errorf(VK_ERROR_DEVICE_LOST, "execbuf2 failed: %m");
1138 }
1139
1140 struct drm_i915_gem_exec_object2 *objects =
1141 (void *)(uintptr_t)execbuf->buffers_ptr;
1142 for (uint32_t k = 0; k < execbuf->buffer_count; k++)
1143 execbuf_bos[k]->offset = objects[k].offset;
1144
1145 return VK_SUCCESS;
1146 }
1147
anv_QueueSubmit(VkQueue _queue,uint32_t submitCount,const VkSubmitInfo * pSubmits,VkFence _fence)1148 VkResult anv_QueueSubmit(
1149 VkQueue _queue,
1150 uint32_t submitCount,
1151 const VkSubmitInfo* pSubmits,
1152 VkFence _fence)
1153 {
1154 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1155 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1156 struct anv_device *device = queue->device;
1157 VkResult result = VK_SUCCESS;
1158
1159 /* We lock around QueueSubmit for three main reasons:
1160 *
1161 * 1) When a block pool is resized, we create a new gem handle with a
1162 * different size and, in the case of surface states, possibly a
1163 * different center offset but we re-use the same anv_bo struct when
1164 * we do so. If this happens in the middle of setting up an execbuf,
1165 * we could end up with our list of BOs out of sync with our list of
1166 * gem handles.
1167 *
1168 * 2) The algorithm we use for building the list of unique buffers isn't
1169 * thread-safe. While the client is supposed to syncronize around
1170 * QueueSubmit, this would be extremely difficult to debug if it ever
1171 * came up in the wild due to a broken app. It's better to play it
1172 * safe and just lock around QueueSubmit.
1173 *
1174 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
1175 * userspace. Due to the fact that the surface state buffer is shared
1176 * between batches, we can't afford to have that happen from multiple
1177 * threads at the same time. Even though the user is supposed to
1178 * ensure this doesn't happen, we play it safe as in (2) above.
1179 *
1180 * Since the only other things that ever take the device lock such as block
1181 * pool resize only rarely happen, this will almost never be contended so
1182 * taking a lock isn't really an expensive operation in this case.
1183 */
1184 pthread_mutex_lock(&device->mutex);
1185
1186 for (uint32_t i = 0; i < submitCount; i++) {
1187 for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1188 ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1189 pSubmits[i].pCommandBuffers[j]);
1190 assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1191
1192 result = anv_cmd_buffer_execbuf(device, cmd_buffer);
1193 if (result != VK_SUCCESS)
1194 goto out;
1195 }
1196 }
1197
1198 if (fence) {
1199 struct anv_bo *fence_bo = &fence->bo;
1200 result = anv_device_execbuf(device, &fence->execbuf, &fence_bo);
1201 if (result != VK_SUCCESS)
1202 goto out;
1203
1204 /* Update the fence and wake up any waiters */
1205 assert(fence->state == ANV_FENCE_STATE_RESET);
1206 fence->state = ANV_FENCE_STATE_SUBMITTED;
1207 pthread_cond_broadcast(&device->queue_submit);
1208 }
1209
1210 out:
1211 pthread_mutex_unlock(&device->mutex);
1212
1213 return result;
1214 }
1215
anv_QueueWaitIdle(VkQueue _queue)1216 VkResult anv_QueueWaitIdle(
1217 VkQueue _queue)
1218 {
1219 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1220
1221 return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1222 }
1223
anv_DeviceWaitIdle(VkDevice _device)1224 VkResult anv_DeviceWaitIdle(
1225 VkDevice _device)
1226 {
1227 ANV_FROM_HANDLE(anv_device, device, _device);
1228 struct anv_batch batch;
1229
1230 uint32_t cmds[8];
1231 batch.start = batch.next = cmds;
1232 batch.end = (void *) cmds + sizeof(cmds);
1233
1234 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1235 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1236
1237 return anv_device_submit_simple_batch(device, &batch);
1238 }
1239
1240 VkResult
anv_bo_init_new(struct anv_bo * bo,struct anv_device * device,uint64_t size)1241 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1242 {
1243 uint32_t gem_handle = anv_gem_create(device, size);
1244 if (!gem_handle)
1245 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1246
1247 anv_bo_init(bo, gem_handle, size);
1248
1249 return VK_SUCCESS;
1250 }
1251
anv_AllocateMemory(VkDevice _device,const VkMemoryAllocateInfo * pAllocateInfo,const VkAllocationCallbacks * pAllocator,VkDeviceMemory * pMem)1252 VkResult anv_AllocateMemory(
1253 VkDevice _device,
1254 const VkMemoryAllocateInfo* pAllocateInfo,
1255 const VkAllocationCallbacks* pAllocator,
1256 VkDeviceMemory* pMem)
1257 {
1258 ANV_FROM_HANDLE(anv_device, device, _device);
1259 struct anv_device_memory *mem;
1260 VkResult result;
1261
1262 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1263
1264 /* The Vulkan 1.0.33 spec says "allocationSize must be greater than 0". */
1265 assert(pAllocateInfo->allocationSize > 0);
1266
1267 /* We support exactly one memory heap. */
1268 assert(pAllocateInfo->memoryTypeIndex == 0 ||
1269 (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1270
1271 /* FINISHME: Fail if allocation request exceeds heap size. */
1272
1273 mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1274 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1275 if (mem == NULL)
1276 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1277
1278 /* The kernel is going to give us whole pages anyway */
1279 uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1280
1281 result = anv_bo_init_new(&mem->bo, device, alloc_size);
1282 if (result != VK_SUCCESS)
1283 goto fail;
1284
1285 mem->type_index = pAllocateInfo->memoryTypeIndex;
1286
1287 mem->map = NULL;
1288 mem->map_size = 0;
1289
1290 *pMem = anv_device_memory_to_handle(mem);
1291
1292 return VK_SUCCESS;
1293
1294 fail:
1295 vk_free2(&device->alloc, pAllocator, mem);
1296
1297 return result;
1298 }
1299
anv_FreeMemory(VkDevice _device,VkDeviceMemory _mem,const VkAllocationCallbacks * pAllocator)1300 void anv_FreeMemory(
1301 VkDevice _device,
1302 VkDeviceMemory _mem,
1303 const VkAllocationCallbacks* pAllocator)
1304 {
1305 ANV_FROM_HANDLE(anv_device, device, _device);
1306 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1307
1308 if (mem == NULL)
1309 return;
1310
1311 if (mem->map)
1312 anv_UnmapMemory(_device, _mem);
1313
1314 if (mem->bo.map)
1315 anv_gem_munmap(mem->bo.map, mem->bo.size);
1316
1317 if (mem->bo.gem_handle != 0)
1318 anv_gem_close(device, mem->bo.gem_handle);
1319
1320 vk_free2(&device->alloc, pAllocator, mem);
1321 }
1322
anv_MapMemory(VkDevice _device,VkDeviceMemory _memory,VkDeviceSize offset,VkDeviceSize size,VkMemoryMapFlags flags,void ** ppData)1323 VkResult anv_MapMemory(
1324 VkDevice _device,
1325 VkDeviceMemory _memory,
1326 VkDeviceSize offset,
1327 VkDeviceSize size,
1328 VkMemoryMapFlags flags,
1329 void** ppData)
1330 {
1331 ANV_FROM_HANDLE(anv_device, device, _device);
1332 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1333
1334 if (mem == NULL) {
1335 *ppData = NULL;
1336 return VK_SUCCESS;
1337 }
1338
1339 if (size == VK_WHOLE_SIZE)
1340 size = mem->bo.size - offset;
1341
1342 /* From the Vulkan spec version 1.0.32 docs for MapMemory:
1343 *
1344 * * If size is not equal to VK_WHOLE_SIZE, size must be greater than 0
1345 * assert(size != 0);
1346 * * If size is not equal to VK_WHOLE_SIZE, size must be less than or
1347 * equal to the size of the memory minus offset
1348 */
1349 assert(size > 0);
1350 assert(offset + size <= mem->bo.size);
1351
1352 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1353 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1354 * at a time is valid. We could just mmap up front and return an offset
1355 * pointer here, but that may exhaust virtual memory on 32 bit
1356 * userspace. */
1357
1358 uint32_t gem_flags = 0;
1359 if (!device->info.has_llc && mem->type_index == 0)
1360 gem_flags |= I915_MMAP_WC;
1361
1362 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1363 uint64_t map_offset = offset & ~4095ull;
1364 assert(offset >= map_offset);
1365 uint64_t map_size = (offset + size) - map_offset;
1366
1367 /* Let's map whole pages */
1368 map_size = align_u64(map_size, 4096);
1369
1370 void *map = anv_gem_mmap(device, mem->bo.gem_handle,
1371 map_offset, map_size, gem_flags);
1372 if (map == MAP_FAILED)
1373 return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
1374
1375 mem->map = map;
1376 mem->map_size = map_size;
1377
1378 *ppData = mem->map + (offset - map_offset);
1379
1380 return VK_SUCCESS;
1381 }
1382
anv_UnmapMemory(VkDevice _device,VkDeviceMemory _memory)1383 void anv_UnmapMemory(
1384 VkDevice _device,
1385 VkDeviceMemory _memory)
1386 {
1387 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1388
1389 if (mem == NULL)
1390 return;
1391
1392 anv_gem_munmap(mem->map, mem->map_size);
1393
1394 mem->map = NULL;
1395 mem->map_size = 0;
1396 }
1397
1398 static void
clflush_mapped_ranges(struct anv_device * device,uint32_t count,const VkMappedMemoryRange * ranges)1399 clflush_mapped_ranges(struct anv_device *device,
1400 uint32_t count,
1401 const VkMappedMemoryRange *ranges)
1402 {
1403 for (uint32_t i = 0; i < count; i++) {
1404 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1405 void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1406 void *end;
1407
1408 if (ranges[i].offset + ranges[i].size > mem->map_size)
1409 end = mem->map + mem->map_size;
1410 else
1411 end = mem->map + ranges[i].offset + ranges[i].size;
1412
1413 while (p < end) {
1414 __builtin_ia32_clflush(p);
1415 p += CACHELINE_SIZE;
1416 }
1417 }
1418 }
1419
anv_FlushMappedMemoryRanges(VkDevice _device,uint32_t memoryRangeCount,const VkMappedMemoryRange * pMemoryRanges)1420 VkResult anv_FlushMappedMemoryRanges(
1421 VkDevice _device,
1422 uint32_t memoryRangeCount,
1423 const VkMappedMemoryRange* pMemoryRanges)
1424 {
1425 ANV_FROM_HANDLE(anv_device, device, _device);
1426
1427 if (device->info.has_llc)
1428 return VK_SUCCESS;
1429
1430 /* Make sure the writes we're flushing have landed. */
1431 __builtin_ia32_mfence();
1432
1433 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1434
1435 return VK_SUCCESS;
1436 }
1437
anv_InvalidateMappedMemoryRanges(VkDevice _device,uint32_t memoryRangeCount,const VkMappedMemoryRange * pMemoryRanges)1438 VkResult anv_InvalidateMappedMemoryRanges(
1439 VkDevice _device,
1440 uint32_t memoryRangeCount,
1441 const VkMappedMemoryRange* pMemoryRanges)
1442 {
1443 ANV_FROM_HANDLE(anv_device, device, _device);
1444
1445 if (device->info.has_llc)
1446 return VK_SUCCESS;
1447
1448 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1449
1450 /* Make sure no reads get moved up above the invalidate. */
1451 __builtin_ia32_mfence();
1452
1453 return VK_SUCCESS;
1454 }
1455
anv_GetBufferMemoryRequirements(VkDevice _device,VkBuffer _buffer,VkMemoryRequirements * pMemoryRequirements)1456 void anv_GetBufferMemoryRequirements(
1457 VkDevice _device,
1458 VkBuffer _buffer,
1459 VkMemoryRequirements* pMemoryRequirements)
1460 {
1461 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1462 ANV_FROM_HANDLE(anv_device, device, _device);
1463
1464 /* The Vulkan spec (git aaed022) says:
1465 *
1466 * memoryTypeBits is a bitfield and contains one bit set for every
1467 * supported memory type for the resource. The bit `1<<i` is set if and
1468 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1469 * structure for the physical device is supported.
1470 *
1471 * We support exactly one memory type on LLC, two on non-LLC.
1472 */
1473 pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1474
1475 pMemoryRequirements->size = buffer->size;
1476 pMemoryRequirements->alignment = 16;
1477 }
1478
anv_GetImageMemoryRequirements(VkDevice _device,VkImage _image,VkMemoryRequirements * pMemoryRequirements)1479 void anv_GetImageMemoryRequirements(
1480 VkDevice _device,
1481 VkImage _image,
1482 VkMemoryRequirements* pMemoryRequirements)
1483 {
1484 ANV_FROM_HANDLE(anv_image, image, _image);
1485 ANV_FROM_HANDLE(anv_device, device, _device);
1486
1487 /* The Vulkan spec (git aaed022) says:
1488 *
1489 * memoryTypeBits is a bitfield and contains one bit set for every
1490 * supported memory type for the resource. The bit `1<<i` is set if and
1491 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1492 * structure for the physical device is supported.
1493 *
1494 * We support exactly one memory type on LLC, two on non-LLC.
1495 */
1496 pMemoryRequirements->memoryTypeBits = device->info.has_llc ? 1 : 3;
1497
1498 pMemoryRequirements->size = image->size;
1499 pMemoryRequirements->alignment = image->alignment;
1500 }
1501
anv_GetImageSparseMemoryRequirements(VkDevice device,VkImage image,uint32_t * pSparseMemoryRequirementCount,VkSparseImageMemoryRequirements * pSparseMemoryRequirements)1502 void anv_GetImageSparseMemoryRequirements(
1503 VkDevice device,
1504 VkImage image,
1505 uint32_t* pSparseMemoryRequirementCount,
1506 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
1507 {
1508 stub();
1509 }
1510
anv_GetDeviceMemoryCommitment(VkDevice device,VkDeviceMemory memory,VkDeviceSize * pCommittedMemoryInBytes)1511 void anv_GetDeviceMemoryCommitment(
1512 VkDevice device,
1513 VkDeviceMemory memory,
1514 VkDeviceSize* pCommittedMemoryInBytes)
1515 {
1516 *pCommittedMemoryInBytes = 0;
1517 }
1518
anv_BindBufferMemory(VkDevice device,VkBuffer _buffer,VkDeviceMemory _memory,VkDeviceSize memoryOffset)1519 VkResult anv_BindBufferMemory(
1520 VkDevice device,
1521 VkBuffer _buffer,
1522 VkDeviceMemory _memory,
1523 VkDeviceSize memoryOffset)
1524 {
1525 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1526 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1527
1528 if (mem) {
1529 buffer->bo = &mem->bo;
1530 buffer->offset = memoryOffset;
1531 } else {
1532 buffer->bo = NULL;
1533 buffer->offset = 0;
1534 }
1535
1536 return VK_SUCCESS;
1537 }
1538
anv_QueueBindSparse(VkQueue queue,uint32_t bindInfoCount,const VkBindSparseInfo * pBindInfo,VkFence fence)1539 VkResult anv_QueueBindSparse(
1540 VkQueue queue,
1541 uint32_t bindInfoCount,
1542 const VkBindSparseInfo* pBindInfo,
1543 VkFence fence)
1544 {
1545 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1546 }
1547
anv_CreateFence(VkDevice _device,const VkFenceCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkFence * pFence)1548 VkResult anv_CreateFence(
1549 VkDevice _device,
1550 const VkFenceCreateInfo* pCreateInfo,
1551 const VkAllocationCallbacks* pAllocator,
1552 VkFence* pFence)
1553 {
1554 ANV_FROM_HANDLE(anv_device, device, _device);
1555 struct anv_bo fence_bo;
1556 struct anv_fence *fence;
1557 struct anv_batch batch;
1558 VkResult result;
1559
1560 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1561
1562 result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1563 if (result != VK_SUCCESS)
1564 return result;
1565
1566 /* Fences are small. Just store the CPU data structure in the BO. */
1567 fence = fence_bo.map;
1568 fence->bo = fence_bo;
1569
1570 /* Place the batch after the CPU data but on its own cache line. */
1571 const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1572 batch.next = batch.start = fence->bo.map + batch_offset;
1573 batch.end = fence->bo.map + fence->bo.size;
1574 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1575 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1576
1577 if (!device->info.has_llc) {
1578 assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1579 assert(batch.next - batch.start <= CACHELINE_SIZE);
1580 __builtin_ia32_mfence();
1581 __builtin_ia32_clflush(batch.start);
1582 }
1583
1584 fence->exec2_objects[0].handle = fence->bo.gem_handle;
1585 fence->exec2_objects[0].relocation_count = 0;
1586 fence->exec2_objects[0].relocs_ptr = 0;
1587 fence->exec2_objects[0].alignment = 0;
1588 fence->exec2_objects[0].offset = fence->bo.offset;
1589 fence->exec2_objects[0].flags = 0;
1590 fence->exec2_objects[0].rsvd1 = 0;
1591 fence->exec2_objects[0].rsvd2 = 0;
1592
1593 fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1594 fence->execbuf.buffer_count = 1;
1595 fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1596 fence->execbuf.batch_len = batch.next - batch.start;
1597 fence->execbuf.cliprects_ptr = 0;
1598 fence->execbuf.num_cliprects = 0;
1599 fence->execbuf.DR1 = 0;
1600 fence->execbuf.DR4 = 0;
1601
1602 fence->execbuf.flags =
1603 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1604 fence->execbuf.rsvd1 = device->context_id;
1605 fence->execbuf.rsvd2 = 0;
1606
1607 if (pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT) {
1608 fence->state = ANV_FENCE_STATE_SIGNALED;
1609 } else {
1610 fence->state = ANV_FENCE_STATE_RESET;
1611 }
1612
1613 *pFence = anv_fence_to_handle(fence);
1614
1615 return VK_SUCCESS;
1616 }
1617
anv_DestroyFence(VkDevice _device,VkFence _fence,const VkAllocationCallbacks * pAllocator)1618 void anv_DestroyFence(
1619 VkDevice _device,
1620 VkFence _fence,
1621 const VkAllocationCallbacks* pAllocator)
1622 {
1623 ANV_FROM_HANDLE(anv_device, device, _device);
1624 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1625
1626 if (!fence)
1627 return;
1628
1629 assert(fence->bo.map == fence);
1630 anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1631 }
1632
anv_ResetFences(VkDevice _device,uint32_t fenceCount,const VkFence * pFences)1633 VkResult anv_ResetFences(
1634 VkDevice _device,
1635 uint32_t fenceCount,
1636 const VkFence* pFences)
1637 {
1638 for (uint32_t i = 0; i < fenceCount; i++) {
1639 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1640 fence->state = ANV_FENCE_STATE_RESET;
1641 }
1642
1643 return VK_SUCCESS;
1644 }
1645
anv_GetFenceStatus(VkDevice _device,VkFence _fence)1646 VkResult anv_GetFenceStatus(
1647 VkDevice _device,
1648 VkFence _fence)
1649 {
1650 ANV_FROM_HANDLE(anv_device, device, _device);
1651 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1652 int64_t t = 0;
1653 int ret;
1654
1655 switch (fence->state) {
1656 case ANV_FENCE_STATE_RESET:
1657 /* If it hasn't even been sent off to the GPU yet, it's not ready */
1658 return VK_NOT_READY;
1659
1660 case ANV_FENCE_STATE_SIGNALED:
1661 /* It's been signaled, return success */
1662 return VK_SUCCESS;
1663
1664 case ANV_FENCE_STATE_SUBMITTED:
1665 /* It's been submitted to the GPU but we don't know if it's done yet. */
1666 ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1667 if (ret == 0) {
1668 fence->state = ANV_FENCE_STATE_SIGNALED;
1669 return VK_SUCCESS;
1670 } else {
1671 return VK_NOT_READY;
1672 }
1673 default:
1674 unreachable("Invalid fence status");
1675 }
1676 }
1677
1678 #define NSEC_PER_SEC 1000000000
1679 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
1680
anv_WaitForFences(VkDevice _device,uint32_t fenceCount,const VkFence * pFences,VkBool32 waitAll,uint64_t _timeout)1681 VkResult anv_WaitForFences(
1682 VkDevice _device,
1683 uint32_t fenceCount,
1684 const VkFence* pFences,
1685 VkBool32 waitAll,
1686 uint64_t _timeout)
1687 {
1688 ANV_FROM_HANDLE(anv_device, device, _device);
1689 int ret;
1690
1691 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1692 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1693 * for a couple of kernel releases. Since there's no way to know
1694 * whether or not the kernel we're using is one of the broken ones, the
1695 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1696 * maximum timeout from 584 years to 292 years - likely not a big deal.
1697 */
1698 int64_t timeout = MIN2(_timeout, INT64_MAX);
1699
1700 uint32_t pending_fences = fenceCount;
1701 while (pending_fences) {
1702 pending_fences = 0;
1703 bool signaled_fences = false;
1704 for (uint32_t i = 0; i < fenceCount; i++) {
1705 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1706 switch (fence->state) {
1707 case ANV_FENCE_STATE_RESET:
1708 /* This fence hasn't been submitted yet, we'll catch it the next
1709 * time around. Yes, this may mean we dead-loop but, short of
1710 * lots of locking and a condition variable, there's not much that
1711 * we can do about that.
1712 */
1713 pending_fences++;
1714 continue;
1715
1716 case ANV_FENCE_STATE_SIGNALED:
1717 /* This fence is not pending. If waitAll isn't set, we can return
1718 * early. Otherwise, we have to keep going.
1719 */
1720 if (!waitAll)
1721 return VK_SUCCESS;
1722 continue;
1723
1724 case ANV_FENCE_STATE_SUBMITTED:
1725 /* These are the fences we really care about. Go ahead and wait
1726 * on it until we hit a timeout.
1727 */
1728 ret = anv_gem_wait(device, fence->bo.gem_handle, &timeout);
1729 if (ret == -1 && errno == ETIME) {
1730 return VK_TIMEOUT;
1731 } else if (ret == -1) {
1732 /* We don't know the real error. */
1733 return vk_errorf(VK_ERROR_DEVICE_LOST, "gem wait failed: %m");
1734 } else {
1735 fence->state = ANV_FENCE_STATE_SIGNALED;
1736 signaled_fences = true;
1737 if (!waitAll)
1738 return VK_SUCCESS;
1739 continue;
1740 }
1741 }
1742 }
1743
1744 if (pending_fences && !signaled_fences) {
1745 /* If we've hit this then someone decided to vkWaitForFences before
1746 * they've actually submitted any of them to a queue. This is a
1747 * fairly pessimal case, so it's ok to lock here and use a standard
1748 * pthreads condition variable.
1749 */
1750 pthread_mutex_lock(&device->mutex);
1751
1752 /* It's possible that some of the fences have changed state since the
1753 * last time we checked. Now that we have the lock, check for
1754 * pending fences again and don't wait if it's changed.
1755 */
1756 uint32_t now_pending_fences = 0;
1757 for (uint32_t i = 0; i < fenceCount; i++) {
1758 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1759 if (fence->state == ANV_FENCE_STATE_RESET)
1760 now_pending_fences++;
1761 }
1762 assert(now_pending_fences <= pending_fences);
1763
1764 if (now_pending_fences == pending_fences) {
1765 struct timespec before;
1766 clock_gettime(CLOCK_MONOTONIC, &before);
1767
1768 uint32_t abs_nsec = before.tv_nsec + timeout % NSEC_PER_SEC;
1769 uint64_t abs_sec = before.tv_sec + (abs_nsec / NSEC_PER_SEC) +
1770 (timeout / NSEC_PER_SEC);
1771 abs_nsec %= NSEC_PER_SEC;
1772
1773 /* Avoid roll-over in tv_sec on 32-bit systems if the user
1774 * provided timeout is UINT64_MAX
1775 */
1776 struct timespec abstime;
1777 abstime.tv_nsec = abs_nsec;
1778 abstime.tv_sec = MIN2(abs_sec, INT_TYPE_MAX(abstime.tv_sec));
1779
1780 ret = pthread_cond_timedwait(&device->queue_submit,
1781 &device->mutex, &abstime);
1782 assert(ret != EINVAL);
1783
1784 struct timespec after;
1785 clock_gettime(CLOCK_MONOTONIC, &after);
1786 uint64_t time_elapsed =
1787 ((uint64_t)after.tv_sec * NSEC_PER_SEC + after.tv_nsec) -
1788 ((uint64_t)before.tv_sec * NSEC_PER_SEC + before.tv_nsec);
1789
1790 if (time_elapsed >= timeout) {
1791 pthread_mutex_unlock(&device->mutex);
1792 return VK_TIMEOUT;
1793 }
1794
1795 timeout -= time_elapsed;
1796 }
1797
1798 pthread_mutex_unlock(&device->mutex);
1799 }
1800 }
1801
1802 return VK_SUCCESS;
1803 }
1804
1805 // Queue semaphore functions
1806
anv_CreateSemaphore(VkDevice device,const VkSemaphoreCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkSemaphore * pSemaphore)1807 VkResult anv_CreateSemaphore(
1808 VkDevice device,
1809 const VkSemaphoreCreateInfo* pCreateInfo,
1810 const VkAllocationCallbacks* pAllocator,
1811 VkSemaphore* pSemaphore)
1812 {
1813 /* The DRM execbuffer ioctl always execute in-oder, even between different
1814 * rings. As such, there's nothing to do for the user space semaphore.
1815 */
1816
1817 *pSemaphore = (VkSemaphore)1;
1818
1819 return VK_SUCCESS;
1820 }
1821
anv_DestroySemaphore(VkDevice device,VkSemaphore semaphore,const VkAllocationCallbacks * pAllocator)1822 void anv_DestroySemaphore(
1823 VkDevice device,
1824 VkSemaphore semaphore,
1825 const VkAllocationCallbacks* pAllocator)
1826 {
1827 }
1828
1829 // Event functions
1830
anv_CreateEvent(VkDevice _device,const VkEventCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkEvent * pEvent)1831 VkResult anv_CreateEvent(
1832 VkDevice _device,
1833 const VkEventCreateInfo* pCreateInfo,
1834 const VkAllocationCallbacks* pAllocator,
1835 VkEvent* pEvent)
1836 {
1837 ANV_FROM_HANDLE(anv_device, device, _device);
1838 struct anv_state state;
1839 struct anv_event *event;
1840
1841 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1842
1843 state = anv_state_pool_alloc(&device->dynamic_state_pool,
1844 sizeof(*event), 8);
1845 event = state.map;
1846 event->state = state;
1847 event->semaphore = VK_EVENT_RESET;
1848
1849 if (!device->info.has_llc) {
1850 /* Make sure the writes we're flushing have landed. */
1851 __builtin_ia32_mfence();
1852 __builtin_ia32_clflush(event);
1853 }
1854
1855 *pEvent = anv_event_to_handle(event);
1856
1857 return VK_SUCCESS;
1858 }
1859
anv_DestroyEvent(VkDevice _device,VkEvent _event,const VkAllocationCallbacks * pAllocator)1860 void anv_DestroyEvent(
1861 VkDevice _device,
1862 VkEvent _event,
1863 const VkAllocationCallbacks* pAllocator)
1864 {
1865 ANV_FROM_HANDLE(anv_device, device, _device);
1866 ANV_FROM_HANDLE(anv_event, event, _event);
1867
1868 if (!event)
1869 return;
1870
1871 anv_state_pool_free(&device->dynamic_state_pool, event->state);
1872 }
1873
anv_GetEventStatus(VkDevice _device,VkEvent _event)1874 VkResult anv_GetEventStatus(
1875 VkDevice _device,
1876 VkEvent _event)
1877 {
1878 ANV_FROM_HANDLE(anv_device, device, _device);
1879 ANV_FROM_HANDLE(anv_event, event, _event);
1880
1881 if (!device->info.has_llc) {
1882 /* Invalidate read cache before reading event written by GPU. */
1883 __builtin_ia32_clflush(event);
1884 __builtin_ia32_mfence();
1885
1886 }
1887
1888 return event->semaphore;
1889 }
1890
anv_SetEvent(VkDevice _device,VkEvent _event)1891 VkResult anv_SetEvent(
1892 VkDevice _device,
1893 VkEvent _event)
1894 {
1895 ANV_FROM_HANDLE(anv_device, device, _device);
1896 ANV_FROM_HANDLE(anv_event, event, _event);
1897
1898 event->semaphore = VK_EVENT_SET;
1899
1900 if (!device->info.has_llc) {
1901 /* Make sure the writes we're flushing have landed. */
1902 __builtin_ia32_mfence();
1903 __builtin_ia32_clflush(event);
1904 }
1905
1906 return VK_SUCCESS;
1907 }
1908
anv_ResetEvent(VkDevice _device,VkEvent _event)1909 VkResult anv_ResetEvent(
1910 VkDevice _device,
1911 VkEvent _event)
1912 {
1913 ANV_FROM_HANDLE(anv_device, device, _device);
1914 ANV_FROM_HANDLE(anv_event, event, _event);
1915
1916 event->semaphore = VK_EVENT_RESET;
1917
1918 if (!device->info.has_llc) {
1919 /* Make sure the writes we're flushing have landed. */
1920 __builtin_ia32_mfence();
1921 __builtin_ia32_clflush(event);
1922 }
1923
1924 return VK_SUCCESS;
1925 }
1926
1927 // Buffer functions
1928
anv_CreateBuffer(VkDevice _device,const VkBufferCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkBuffer * pBuffer)1929 VkResult anv_CreateBuffer(
1930 VkDevice _device,
1931 const VkBufferCreateInfo* pCreateInfo,
1932 const VkAllocationCallbacks* pAllocator,
1933 VkBuffer* pBuffer)
1934 {
1935 ANV_FROM_HANDLE(anv_device, device, _device);
1936 struct anv_buffer *buffer;
1937
1938 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1939
1940 buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1941 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1942 if (buffer == NULL)
1943 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1944
1945 buffer->size = pCreateInfo->size;
1946 buffer->usage = pCreateInfo->usage;
1947 buffer->bo = NULL;
1948 buffer->offset = 0;
1949
1950 *pBuffer = anv_buffer_to_handle(buffer);
1951
1952 return VK_SUCCESS;
1953 }
1954
anv_DestroyBuffer(VkDevice _device,VkBuffer _buffer,const VkAllocationCallbacks * pAllocator)1955 void anv_DestroyBuffer(
1956 VkDevice _device,
1957 VkBuffer _buffer,
1958 const VkAllocationCallbacks* pAllocator)
1959 {
1960 ANV_FROM_HANDLE(anv_device, device, _device);
1961 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1962
1963 if (!buffer)
1964 return;
1965
1966 vk_free2(&device->alloc, pAllocator, buffer);
1967 }
1968
1969 void
anv_fill_buffer_surface_state(struct anv_device * device,struct anv_state state,enum isl_format format,uint32_t offset,uint32_t range,uint32_t stride)1970 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1971 enum isl_format format,
1972 uint32_t offset, uint32_t range, uint32_t stride)
1973 {
1974 isl_buffer_fill_state(&device->isl_dev, state.map,
1975 .address = offset,
1976 .mocs = device->default_mocs,
1977 .size = range,
1978 .format = format,
1979 .stride = stride);
1980
1981 if (!device->info.has_llc)
1982 anv_state_clflush(state);
1983 }
1984
anv_DestroySampler(VkDevice _device,VkSampler _sampler,const VkAllocationCallbacks * pAllocator)1985 void anv_DestroySampler(
1986 VkDevice _device,
1987 VkSampler _sampler,
1988 const VkAllocationCallbacks* pAllocator)
1989 {
1990 ANV_FROM_HANDLE(anv_device, device, _device);
1991 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
1992
1993 if (!sampler)
1994 return;
1995
1996 vk_free2(&device->alloc, pAllocator, sampler);
1997 }
1998
anv_CreateFramebuffer(VkDevice _device,const VkFramebufferCreateInfo * pCreateInfo,const VkAllocationCallbacks * pAllocator,VkFramebuffer * pFramebuffer)1999 VkResult anv_CreateFramebuffer(
2000 VkDevice _device,
2001 const VkFramebufferCreateInfo* pCreateInfo,
2002 const VkAllocationCallbacks* pAllocator,
2003 VkFramebuffer* pFramebuffer)
2004 {
2005 ANV_FROM_HANDLE(anv_device, device, _device);
2006 struct anv_framebuffer *framebuffer;
2007
2008 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
2009
2010 size_t size = sizeof(*framebuffer) +
2011 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
2012 framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
2013 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
2014 if (framebuffer == NULL)
2015 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
2016
2017 framebuffer->attachment_count = pCreateInfo->attachmentCount;
2018 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
2019 VkImageView _iview = pCreateInfo->pAttachments[i];
2020 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
2021 }
2022
2023 framebuffer->width = pCreateInfo->width;
2024 framebuffer->height = pCreateInfo->height;
2025 framebuffer->layers = pCreateInfo->layers;
2026
2027 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
2028
2029 return VK_SUCCESS;
2030 }
2031
anv_DestroyFramebuffer(VkDevice _device,VkFramebuffer _fb,const VkAllocationCallbacks * pAllocator)2032 void anv_DestroyFramebuffer(
2033 VkDevice _device,
2034 VkFramebuffer _fb,
2035 const VkAllocationCallbacks* pAllocator)
2036 {
2037 ANV_FROM_HANDLE(anv_device, device, _device);
2038 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
2039
2040 if (!fb)
2041 return;
2042
2043 vk_free2(&device->alloc, pAllocator, fb);
2044 }
2045
2046 /* vk_icd.h does not declare this function, so we declare it here to
2047 * suppress Wmissing-prototypes.
2048 */
2049 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
2050 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion);
2051
2052 PUBLIC VKAPI_ATTR VkResult VKAPI_CALL
vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t * pSupportedVersion)2053 vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t* pSupportedVersion)
2054 {
2055 /* For the full details on loader interface versioning, see
2056 * <https://github.com/KhronosGroup/Vulkan-LoaderAndValidationLayers/blob/master/loader/LoaderAndLayerInterface.md>.
2057 * What follows is a condensed summary, to help you navigate the large and
2058 * confusing official doc.
2059 *
2060 * - Loader interface v0 is incompatible with later versions. We don't
2061 * support it.
2062 *
2063 * - In loader interface v1:
2064 * - The first ICD entrypoint called by the loader is
2065 * vk_icdGetInstanceProcAddr(). The ICD must statically expose this
2066 * entrypoint.
2067 * - The ICD must statically expose no other Vulkan symbol unless it is
2068 * linked with -Bsymbolic.
2069 * - Each dispatchable Vulkan handle created by the ICD must be
2070 * a pointer to a struct whose first member is VK_LOADER_DATA. The
2071 * ICD must initialize VK_LOADER_DATA.loadMagic to ICD_LOADER_MAGIC.
2072 * - The loader implements vkCreate{PLATFORM}SurfaceKHR() and
2073 * vkDestroySurfaceKHR(). The ICD must be capable of working with
2074 * such loader-managed surfaces.
2075 *
2076 * - Loader interface v2 differs from v1 in:
2077 * - The first ICD entrypoint called by the loader is
2078 * vk_icdNegotiateLoaderICDInterfaceVersion(). The ICD must
2079 * statically expose this entrypoint.
2080 *
2081 * - Loader interface v3 differs from v2 in:
2082 * - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
2083 * vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
2084 * because the loader no longer does so.
2085 */
2086 *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
2087 return VK_SUCCESS;
2088 }
2089