anv: Make entrypoint resolution take a gen_device_info
[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(NULL, 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 ANV_FROM_HANDLE(anv_device, device, _device);
709 return anv_lookup_entrypoint(&device->info, pName);
710 }
711
712 static VkResult
713 anv_queue_init(struct anv_device *device, struct anv_queue *queue)
714 {
715 queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
716 queue->device = device;
717 queue->pool = &device->surface_state_pool;
718
719 return VK_SUCCESS;
720 }
721
722 static void
723 anv_queue_finish(struct anv_queue *queue)
724 {
725 }
726
727 static struct anv_state
728 anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, const void *p)
729 {
730 struct anv_state state;
731
732 state = anv_state_pool_alloc(pool, size, align);
733 memcpy(state.map, p, size);
734
735 if (!pool->block_pool->device->info.has_llc)
736 anv_state_clflush(state);
737
738 return state;
739 }
740
741 struct gen8_border_color {
742 union {
743 float float32[4];
744 uint32_t uint32[4];
745 };
746 /* Pad out to 64 bytes */
747 uint32_t _pad[12];
748 };
749
750 static void
751 anv_device_init_border_colors(struct anv_device *device)
752 {
753 static const struct gen8_border_color border_colors[] = {
754 [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } },
755 [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } },
756 [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } },
757 [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } },
758 [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } },
759 [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } },
760 };
761
762 device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool,
763 sizeof(border_colors), 64,
764 border_colors);
765 }
766
767 VkResult
768 anv_device_submit_simple_batch(struct anv_device *device,
769 struct anv_batch *batch)
770 {
771 struct drm_i915_gem_execbuffer2 execbuf;
772 struct drm_i915_gem_exec_object2 exec2_objects[1];
773 struct anv_bo bo;
774 VkResult result = VK_SUCCESS;
775 uint32_t size;
776 int64_t timeout;
777 int ret;
778
779 /* Kernel driver requires 8 byte aligned batch length */
780 size = align_u32(batch->next - batch->start, 8);
781 result = anv_bo_pool_alloc(&device->batch_bo_pool, &bo, size);
782 if (result != VK_SUCCESS)
783 return result;
784
785 memcpy(bo.map, batch->start, size);
786 if (!device->info.has_llc)
787 anv_clflush_range(bo.map, size);
788
789 exec2_objects[0].handle = bo.gem_handle;
790 exec2_objects[0].relocation_count = 0;
791 exec2_objects[0].relocs_ptr = 0;
792 exec2_objects[0].alignment = 0;
793 exec2_objects[0].offset = bo.offset;
794 exec2_objects[0].flags = 0;
795 exec2_objects[0].rsvd1 = 0;
796 exec2_objects[0].rsvd2 = 0;
797
798 execbuf.buffers_ptr = (uintptr_t) exec2_objects;
799 execbuf.buffer_count = 1;
800 execbuf.batch_start_offset = 0;
801 execbuf.batch_len = size;
802 execbuf.cliprects_ptr = 0;
803 execbuf.num_cliprects = 0;
804 execbuf.DR1 = 0;
805 execbuf.DR4 = 0;
806
807 execbuf.flags =
808 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
809 execbuf.rsvd1 = device->context_id;
810 execbuf.rsvd2 = 0;
811
812 ret = anv_gem_execbuffer(device, &execbuf);
813 if (ret != 0) {
814 /* We don't know the real error. */
815 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
816 goto fail;
817 }
818
819 timeout = INT64_MAX;
820 ret = anv_gem_wait(device, bo.gem_handle, &timeout);
821 if (ret != 0) {
822 /* We don't know the real error. */
823 result = vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY, "execbuf2 failed: %m");
824 goto fail;
825 }
826
827 fail:
828 anv_bo_pool_free(&device->batch_bo_pool, &bo);
829
830 return result;
831 }
832
833 VkResult anv_CreateDevice(
834 VkPhysicalDevice physicalDevice,
835 const VkDeviceCreateInfo* pCreateInfo,
836 const VkAllocationCallbacks* pAllocator,
837 VkDevice* pDevice)
838 {
839 ANV_FROM_HANDLE(anv_physical_device, physical_device, physicalDevice);
840 VkResult result;
841 struct anv_device *device;
842
843 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO);
844
845 for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
846 bool found = false;
847 for (uint32_t j = 0; j < ARRAY_SIZE(device_extensions); j++) {
848 if (strcmp(pCreateInfo->ppEnabledExtensionNames[i],
849 device_extensions[j].extensionName) == 0) {
850 found = true;
851 break;
852 }
853 }
854 if (!found)
855 return vk_error(VK_ERROR_EXTENSION_NOT_PRESENT);
856 }
857
858 device = anv_alloc2(&physical_device->instance->alloc, pAllocator,
859 sizeof(*device), 8,
860 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
861 if (!device)
862 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
863
864 device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
865 device->instance = physical_device->instance;
866 device->chipset_id = physical_device->chipset_id;
867
868 if (pAllocator)
869 device->alloc = *pAllocator;
870 else
871 device->alloc = physical_device->instance->alloc;
872
873 /* XXX(chadv): Can we dup() physicalDevice->fd here? */
874 device->fd = open(physical_device->path, O_RDWR | O_CLOEXEC);
875 if (device->fd == -1) {
876 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
877 goto fail_device;
878 }
879
880 device->context_id = anv_gem_create_context(device);
881 if (device->context_id == -1) {
882 result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
883 goto fail_fd;
884 }
885
886 device->info = physical_device->info;
887 device->isl_dev = physical_device->isl_dev;
888
889 /* On Broadwell and later, we can use batch chaining to more efficiently
890 * implement growing command buffers. Prior to Haswell, the kernel
891 * command parser gets in the way and we have to fall back to growing
892 * the batch.
893 */
894 device->can_chain_batches = device->info.gen >= 8;
895
896 device->robust_buffer_access = pCreateInfo->pEnabledFeatures &&
897 pCreateInfo->pEnabledFeatures->robustBufferAccess;
898
899 pthread_mutex_init(&device->mutex, NULL);
900
901 anv_bo_pool_init(&device->batch_bo_pool, device);
902
903 anv_block_pool_init(&device->dynamic_state_block_pool, device, 16384);
904
905 anv_state_pool_init(&device->dynamic_state_pool,
906 &device->dynamic_state_block_pool);
907
908 anv_block_pool_init(&device->instruction_block_pool, device, 128 * 1024);
909 anv_state_pool_init(&device->instruction_state_pool,
910 &device->instruction_block_pool);
911
912 anv_block_pool_init(&device->surface_state_block_pool, device, 4096);
913
914 anv_state_pool_init(&device->surface_state_pool,
915 &device->surface_state_block_pool);
916
917 anv_bo_init_new(&device->workaround_bo, device, 1024);
918
919 anv_scratch_pool_init(device, &device->scratch_pool);
920
921 anv_queue_init(device, &device->queue);
922
923 switch (device->info.gen) {
924 case 7:
925 if (!device->info.is_haswell)
926 result = gen7_init_device_state(device);
927 else
928 result = gen75_init_device_state(device);
929 break;
930 case 8:
931 result = gen8_init_device_state(device);
932 break;
933 case 9:
934 result = gen9_init_device_state(device);
935 break;
936 default:
937 /* Shouldn't get here as we don't create physical devices for any other
938 * gens. */
939 unreachable("unhandled gen");
940 }
941 if (result != VK_SUCCESS)
942 goto fail_fd;
943
944 anv_device_init_blorp(device);
945
946 anv_device_init_border_colors(device);
947
948 *pDevice = anv_device_to_handle(device);
949
950 return VK_SUCCESS;
951
952 fail_fd:
953 close(device->fd);
954 fail_device:
955 anv_free(&device->alloc, device);
956
957 return result;
958 }
959
960 void anv_DestroyDevice(
961 VkDevice _device,
962 const VkAllocationCallbacks* pAllocator)
963 {
964 ANV_FROM_HANDLE(anv_device, device, _device);
965
966 anv_queue_finish(&device->queue);
967
968 anv_device_finish_blorp(device);
969
970 #ifdef HAVE_VALGRIND
971 /* We only need to free these to prevent valgrind errors. The backing
972 * BO will go away in a couple of lines so we don't actually leak.
973 */
974 anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
975 #endif
976
977 anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
978 anv_gem_close(device, device->workaround_bo.gem_handle);
979
980 anv_bo_pool_finish(&device->batch_bo_pool);
981 anv_state_pool_finish(&device->dynamic_state_pool);
982 anv_block_pool_finish(&device->dynamic_state_block_pool);
983 anv_state_pool_finish(&device->instruction_state_pool);
984 anv_block_pool_finish(&device->instruction_block_pool);
985 anv_state_pool_finish(&device->surface_state_pool);
986 anv_block_pool_finish(&device->surface_state_block_pool);
987 anv_scratch_pool_finish(device, &device->scratch_pool);
988
989 close(device->fd);
990
991 pthread_mutex_destroy(&device->mutex);
992
993 anv_free(&device->alloc, device);
994 }
995
996 VkResult anv_EnumerateInstanceExtensionProperties(
997 const char* pLayerName,
998 uint32_t* pPropertyCount,
999 VkExtensionProperties* pProperties)
1000 {
1001 if (pProperties == NULL) {
1002 *pPropertyCount = ARRAY_SIZE(global_extensions);
1003 return VK_SUCCESS;
1004 }
1005
1006 assert(*pPropertyCount >= ARRAY_SIZE(global_extensions));
1007
1008 *pPropertyCount = ARRAY_SIZE(global_extensions);
1009 memcpy(pProperties, global_extensions, sizeof(global_extensions));
1010
1011 return VK_SUCCESS;
1012 }
1013
1014 VkResult anv_EnumerateDeviceExtensionProperties(
1015 VkPhysicalDevice physicalDevice,
1016 const char* pLayerName,
1017 uint32_t* pPropertyCount,
1018 VkExtensionProperties* pProperties)
1019 {
1020 if (pProperties == NULL) {
1021 *pPropertyCount = ARRAY_SIZE(device_extensions);
1022 return VK_SUCCESS;
1023 }
1024
1025 assert(*pPropertyCount >= ARRAY_SIZE(device_extensions));
1026
1027 *pPropertyCount = ARRAY_SIZE(device_extensions);
1028 memcpy(pProperties, device_extensions, sizeof(device_extensions));
1029
1030 return VK_SUCCESS;
1031 }
1032
1033 VkResult anv_EnumerateInstanceLayerProperties(
1034 uint32_t* pPropertyCount,
1035 VkLayerProperties* pProperties)
1036 {
1037 if (pProperties == NULL) {
1038 *pPropertyCount = 0;
1039 return VK_SUCCESS;
1040 }
1041
1042 /* None supported at this time */
1043 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1044 }
1045
1046 VkResult anv_EnumerateDeviceLayerProperties(
1047 VkPhysicalDevice physicalDevice,
1048 uint32_t* pPropertyCount,
1049 VkLayerProperties* pProperties)
1050 {
1051 if (pProperties == NULL) {
1052 *pPropertyCount = 0;
1053 return VK_SUCCESS;
1054 }
1055
1056 /* None supported at this time */
1057 return vk_error(VK_ERROR_LAYER_NOT_PRESENT);
1058 }
1059
1060 void anv_GetDeviceQueue(
1061 VkDevice _device,
1062 uint32_t queueNodeIndex,
1063 uint32_t queueIndex,
1064 VkQueue* pQueue)
1065 {
1066 ANV_FROM_HANDLE(anv_device, device, _device);
1067
1068 assert(queueIndex == 0);
1069
1070 *pQueue = anv_queue_to_handle(&device->queue);
1071 }
1072
1073 VkResult anv_QueueSubmit(
1074 VkQueue _queue,
1075 uint32_t submitCount,
1076 const VkSubmitInfo* pSubmits,
1077 VkFence _fence)
1078 {
1079 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1080 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1081 struct anv_device *device = queue->device;
1082 int ret;
1083
1084 for (uint32_t i = 0; i < submitCount; i++) {
1085 for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
1086 ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer,
1087 pSubmits[i].pCommandBuffers[j]);
1088 assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
1089
1090 ret = anv_gem_execbuffer(device, &cmd_buffer->execbuf2.execbuf);
1091 if (ret != 0) {
1092 /* We don't know the real error. */
1093 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1094 "execbuf2 failed: %m");
1095 }
1096
1097 for (uint32_t k = 0; k < cmd_buffer->execbuf2.bo_count; k++)
1098 cmd_buffer->execbuf2.bos[k]->offset = cmd_buffer->execbuf2.objects[k].offset;
1099 }
1100 }
1101
1102 if (fence) {
1103 ret = anv_gem_execbuffer(device, &fence->execbuf);
1104 if (ret != 0) {
1105 /* We don't know the real error. */
1106 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1107 "execbuf2 failed: %m");
1108 }
1109 }
1110
1111 return VK_SUCCESS;
1112 }
1113
1114 VkResult anv_QueueWaitIdle(
1115 VkQueue _queue)
1116 {
1117 ANV_FROM_HANDLE(anv_queue, queue, _queue);
1118
1119 return anv_DeviceWaitIdle(anv_device_to_handle(queue->device));
1120 }
1121
1122 VkResult anv_DeviceWaitIdle(
1123 VkDevice _device)
1124 {
1125 ANV_FROM_HANDLE(anv_device, device, _device);
1126 struct anv_batch batch;
1127
1128 uint32_t cmds[8];
1129 batch.start = batch.next = cmds;
1130 batch.end = (void *) cmds + sizeof(cmds);
1131
1132 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1133 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1134
1135 return anv_device_submit_simple_batch(device, &batch);
1136 }
1137
1138 VkResult
1139 anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size)
1140 {
1141 bo->gem_handle = anv_gem_create(device, size);
1142 if (!bo->gem_handle)
1143 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY);
1144
1145 bo->map = NULL;
1146 bo->index = 0;
1147 bo->offset = 0;
1148 bo->size = size;
1149 bo->is_winsys_bo = false;
1150
1151 return VK_SUCCESS;
1152 }
1153
1154 VkResult anv_AllocateMemory(
1155 VkDevice _device,
1156 const VkMemoryAllocateInfo* pAllocateInfo,
1157 const VkAllocationCallbacks* pAllocator,
1158 VkDeviceMemory* pMem)
1159 {
1160 ANV_FROM_HANDLE(anv_device, device, _device);
1161 struct anv_device_memory *mem;
1162 VkResult result;
1163
1164 assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
1165
1166 if (pAllocateInfo->allocationSize == 0) {
1167 /* Apparently, this is allowed */
1168 *pMem = VK_NULL_HANDLE;
1169 return VK_SUCCESS;
1170 }
1171
1172 /* We support exactly one memory heap. */
1173 assert(pAllocateInfo->memoryTypeIndex == 0 ||
1174 (!device->info.has_llc && pAllocateInfo->memoryTypeIndex < 2));
1175
1176 /* FINISHME: Fail if allocation request exceeds heap size. */
1177
1178 mem = anv_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
1179 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1180 if (mem == NULL)
1181 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1182
1183 /* The kernel is going to give us whole pages anyway */
1184 uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
1185
1186 result = anv_bo_init_new(&mem->bo, device, alloc_size);
1187 if (result != VK_SUCCESS)
1188 goto fail;
1189
1190 mem->type_index = pAllocateInfo->memoryTypeIndex;
1191
1192 *pMem = anv_device_memory_to_handle(mem);
1193
1194 return VK_SUCCESS;
1195
1196 fail:
1197 anv_free2(&device->alloc, pAllocator, mem);
1198
1199 return result;
1200 }
1201
1202 void anv_FreeMemory(
1203 VkDevice _device,
1204 VkDeviceMemory _mem,
1205 const VkAllocationCallbacks* pAllocator)
1206 {
1207 ANV_FROM_HANDLE(anv_device, device, _device);
1208 ANV_FROM_HANDLE(anv_device_memory, mem, _mem);
1209
1210 if (mem == NULL)
1211 return;
1212
1213 if (mem->bo.map)
1214 anv_gem_munmap(mem->bo.map, mem->bo.size);
1215
1216 if (mem->bo.gem_handle != 0)
1217 anv_gem_close(device, mem->bo.gem_handle);
1218
1219 anv_free2(&device->alloc, pAllocator, mem);
1220 }
1221
1222 VkResult anv_MapMemory(
1223 VkDevice _device,
1224 VkDeviceMemory _memory,
1225 VkDeviceSize offset,
1226 VkDeviceSize size,
1227 VkMemoryMapFlags flags,
1228 void** ppData)
1229 {
1230 ANV_FROM_HANDLE(anv_device, device, _device);
1231 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1232
1233 if (mem == NULL) {
1234 *ppData = NULL;
1235 return VK_SUCCESS;
1236 }
1237
1238 if (size == VK_WHOLE_SIZE)
1239 size = mem->bo.size - offset;
1240
1241 /* FIXME: Is this supposed to be thread safe? Since vkUnmapMemory() only
1242 * takes a VkDeviceMemory pointer, it seems like only one map of the memory
1243 * at a time is valid. We could just mmap up front and return an offset
1244 * pointer here, but that may exhaust virtual memory on 32 bit
1245 * userspace. */
1246
1247 uint32_t gem_flags = 0;
1248 if (!device->info.has_llc && mem->type_index == 0)
1249 gem_flags |= I915_MMAP_WC;
1250
1251 /* GEM will fail to map if the offset isn't 4k-aligned. Round down. */
1252 uint64_t map_offset = offset & ~4095ull;
1253 assert(offset >= map_offset);
1254 uint64_t map_size = (offset + size) - map_offset;
1255
1256 /* Let's map whole pages */
1257 map_size = align_u64(map_size, 4096);
1258
1259 mem->map = anv_gem_mmap(device, mem->bo.gem_handle,
1260 map_offset, map_size, gem_flags);
1261 mem->map_size = map_size;
1262
1263 *ppData = mem->map + (offset - map_offset);
1264
1265 return VK_SUCCESS;
1266 }
1267
1268 void anv_UnmapMemory(
1269 VkDevice _device,
1270 VkDeviceMemory _memory)
1271 {
1272 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1273
1274 if (mem == NULL)
1275 return;
1276
1277 anv_gem_munmap(mem->map, mem->map_size);
1278 }
1279
1280 static void
1281 clflush_mapped_ranges(struct anv_device *device,
1282 uint32_t count,
1283 const VkMappedMemoryRange *ranges)
1284 {
1285 for (uint32_t i = 0; i < count; i++) {
1286 ANV_FROM_HANDLE(anv_device_memory, mem, ranges[i].memory);
1287 void *p = mem->map + (ranges[i].offset & ~CACHELINE_MASK);
1288 void *end;
1289
1290 if (ranges[i].offset + ranges[i].size > mem->map_size)
1291 end = mem->map + mem->map_size;
1292 else
1293 end = mem->map + ranges[i].offset + ranges[i].size;
1294
1295 while (p < end) {
1296 __builtin_ia32_clflush(p);
1297 p += CACHELINE_SIZE;
1298 }
1299 }
1300 }
1301
1302 VkResult anv_FlushMappedMemoryRanges(
1303 VkDevice _device,
1304 uint32_t memoryRangeCount,
1305 const VkMappedMemoryRange* pMemoryRanges)
1306 {
1307 ANV_FROM_HANDLE(anv_device, device, _device);
1308
1309 if (device->info.has_llc)
1310 return VK_SUCCESS;
1311
1312 /* Make sure the writes we're flushing have landed. */
1313 __builtin_ia32_mfence();
1314
1315 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1316
1317 return VK_SUCCESS;
1318 }
1319
1320 VkResult anv_InvalidateMappedMemoryRanges(
1321 VkDevice _device,
1322 uint32_t memoryRangeCount,
1323 const VkMappedMemoryRange* pMemoryRanges)
1324 {
1325 ANV_FROM_HANDLE(anv_device, device, _device);
1326
1327 if (device->info.has_llc)
1328 return VK_SUCCESS;
1329
1330 clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
1331
1332 /* Make sure no reads get moved up above the invalidate. */
1333 __builtin_ia32_mfence();
1334
1335 return VK_SUCCESS;
1336 }
1337
1338 void anv_GetBufferMemoryRequirements(
1339 VkDevice device,
1340 VkBuffer _buffer,
1341 VkMemoryRequirements* pMemoryRequirements)
1342 {
1343 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1344
1345 /* The Vulkan spec (git aaed022) says:
1346 *
1347 * memoryTypeBits is a bitfield and contains one bit set for every
1348 * supported memory type for the resource. The bit `1<<i` is set if and
1349 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1350 * structure for the physical device is supported.
1351 *
1352 * We support exactly one memory type.
1353 */
1354 pMemoryRequirements->memoryTypeBits = 1;
1355
1356 pMemoryRequirements->size = buffer->size;
1357 pMemoryRequirements->alignment = 16;
1358 }
1359
1360 void anv_GetImageMemoryRequirements(
1361 VkDevice device,
1362 VkImage _image,
1363 VkMemoryRequirements* pMemoryRequirements)
1364 {
1365 ANV_FROM_HANDLE(anv_image, image, _image);
1366
1367 /* The Vulkan spec (git aaed022) says:
1368 *
1369 * memoryTypeBits is a bitfield and contains one bit set for every
1370 * supported memory type for the resource. The bit `1<<i` is set if and
1371 * only if the memory type `i` in the VkPhysicalDeviceMemoryProperties
1372 * structure for the physical device is supported.
1373 *
1374 * We support exactly one memory type.
1375 */
1376 pMemoryRequirements->memoryTypeBits = 1;
1377
1378 pMemoryRequirements->size = image->size;
1379 pMemoryRequirements->alignment = image->alignment;
1380 }
1381
1382 void anv_GetImageSparseMemoryRequirements(
1383 VkDevice device,
1384 VkImage image,
1385 uint32_t* pSparseMemoryRequirementCount,
1386 VkSparseImageMemoryRequirements* pSparseMemoryRequirements)
1387 {
1388 stub();
1389 }
1390
1391 void anv_GetDeviceMemoryCommitment(
1392 VkDevice device,
1393 VkDeviceMemory memory,
1394 VkDeviceSize* pCommittedMemoryInBytes)
1395 {
1396 *pCommittedMemoryInBytes = 0;
1397 }
1398
1399 VkResult anv_BindBufferMemory(
1400 VkDevice device,
1401 VkBuffer _buffer,
1402 VkDeviceMemory _memory,
1403 VkDeviceSize memoryOffset)
1404 {
1405 ANV_FROM_HANDLE(anv_device_memory, mem, _memory);
1406 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1407
1408 if (mem) {
1409 buffer->bo = &mem->bo;
1410 buffer->offset = memoryOffset;
1411 } else {
1412 buffer->bo = NULL;
1413 buffer->offset = 0;
1414 }
1415
1416 return VK_SUCCESS;
1417 }
1418
1419 VkResult anv_QueueBindSparse(
1420 VkQueue queue,
1421 uint32_t bindInfoCount,
1422 const VkBindSparseInfo* pBindInfo,
1423 VkFence fence)
1424 {
1425 stub_return(VK_ERROR_INCOMPATIBLE_DRIVER);
1426 }
1427
1428 VkResult anv_CreateFence(
1429 VkDevice _device,
1430 const VkFenceCreateInfo* pCreateInfo,
1431 const VkAllocationCallbacks* pAllocator,
1432 VkFence* pFence)
1433 {
1434 ANV_FROM_HANDLE(anv_device, device, _device);
1435 struct anv_bo fence_bo;
1436 struct anv_fence *fence;
1437 struct anv_batch batch;
1438 VkResult result;
1439
1440 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FENCE_CREATE_INFO);
1441
1442 result = anv_bo_pool_alloc(&device->batch_bo_pool, &fence_bo, 4096);
1443 if (result != VK_SUCCESS)
1444 return result;
1445
1446 /* Fences are small. Just store the CPU data structure in the BO. */
1447 fence = fence_bo.map;
1448 fence->bo = fence_bo;
1449
1450 /* Place the batch after the CPU data but on its own cache line. */
1451 const uint32_t batch_offset = align_u32(sizeof(*fence), CACHELINE_SIZE);
1452 batch.next = batch.start = fence->bo.map + batch_offset;
1453 batch.end = fence->bo.map + fence->bo.size;
1454 anv_batch_emit(&batch, GEN7_MI_BATCH_BUFFER_END, bbe);
1455 anv_batch_emit(&batch, GEN7_MI_NOOP, noop);
1456
1457 if (!device->info.has_llc) {
1458 assert(((uintptr_t) batch.start & CACHELINE_MASK) == 0);
1459 assert(batch.next - batch.start <= CACHELINE_SIZE);
1460 __builtin_ia32_mfence();
1461 __builtin_ia32_clflush(batch.start);
1462 }
1463
1464 fence->exec2_objects[0].handle = fence->bo.gem_handle;
1465 fence->exec2_objects[0].relocation_count = 0;
1466 fence->exec2_objects[0].relocs_ptr = 0;
1467 fence->exec2_objects[0].alignment = 0;
1468 fence->exec2_objects[0].offset = fence->bo.offset;
1469 fence->exec2_objects[0].flags = 0;
1470 fence->exec2_objects[0].rsvd1 = 0;
1471 fence->exec2_objects[0].rsvd2 = 0;
1472
1473 fence->execbuf.buffers_ptr = (uintptr_t) fence->exec2_objects;
1474 fence->execbuf.buffer_count = 1;
1475 fence->execbuf.batch_start_offset = batch.start - fence->bo.map;
1476 fence->execbuf.batch_len = batch.next - batch.start;
1477 fence->execbuf.cliprects_ptr = 0;
1478 fence->execbuf.num_cliprects = 0;
1479 fence->execbuf.DR1 = 0;
1480 fence->execbuf.DR4 = 0;
1481
1482 fence->execbuf.flags =
1483 I915_EXEC_HANDLE_LUT | I915_EXEC_NO_RELOC | I915_EXEC_RENDER;
1484 fence->execbuf.rsvd1 = device->context_id;
1485 fence->execbuf.rsvd2 = 0;
1486
1487 fence->ready = false;
1488
1489 *pFence = anv_fence_to_handle(fence);
1490
1491 return VK_SUCCESS;
1492 }
1493
1494 void anv_DestroyFence(
1495 VkDevice _device,
1496 VkFence _fence,
1497 const VkAllocationCallbacks* pAllocator)
1498 {
1499 ANV_FROM_HANDLE(anv_device, device, _device);
1500 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1501
1502 assert(fence->bo.map == fence);
1503 anv_bo_pool_free(&device->batch_bo_pool, &fence->bo);
1504 }
1505
1506 VkResult anv_ResetFences(
1507 VkDevice _device,
1508 uint32_t fenceCount,
1509 const VkFence* pFences)
1510 {
1511 for (uint32_t i = 0; i < fenceCount; i++) {
1512 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1513 fence->ready = false;
1514 }
1515
1516 return VK_SUCCESS;
1517 }
1518
1519 VkResult anv_GetFenceStatus(
1520 VkDevice _device,
1521 VkFence _fence)
1522 {
1523 ANV_FROM_HANDLE(anv_device, device, _device);
1524 ANV_FROM_HANDLE(anv_fence, fence, _fence);
1525 int64_t t = 0;
1526 int ret;
1527
1528 if (fence->ready)
1529 return VK_SUCCESS;
1530
1531 ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1532 if (ret == 0) {
1533 fence->ready = true;
1534 return VK_SUCCESS;
1535 }
1536
1537 return VK_NOT_READY;
1538 }
1539
1540 VkResult anv_WaitForFences(
1541 VkDevice _device,
1542 uint32_t fenceCount,
1543 const VkFence* pFences,
1544 VkBool32 waitAll,
1545 uint64_t timeout)
1546 {
1547 ANV_FROM_HANDLE(anv_device, device, _device);
1548
1549 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
1550 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
1551 * for a couple of kernel releases. Since there's no way to know
1552 * whether or not the kernel we're using is one of the broken ones, the
1553 * best we can do is to clamp the timeout to INT64_MAX. This limits the
1554 * maximum timeout from 584 years to 292 years - likely not a big deal.
1555 */
1556 if (timeout > INT64_MAX)
1557 timeout = INT64_MAX;
1558
1559 int64_t t = timeout;
1560
1561 /* FIXME: handle !waitAll */
1562
1563 for (uint32_t i = 0; i < fenceCount; i++) {
1564 ANV_FROM_HANDLE(anv_fence, fence, pFences[i]);
1565 int ret = anv_gem_wait(device, fence->bo.gem_handle, &t);
1566 if (ret == -1 && errno == ETIME) {
1567 return VK_TIMEOUT;
1568 } else if (ret == -1) {
1569 /* We don't know the real error. */
1570 return vk_errorf(VK_ERROR_OUT_OF_DEVICE_MEMORY,
1571 "gem wait failed: %m");
1572 }
1573 }
1574
1575 return VK_SUCCESS;
1576 }
1577
1578 // Queue semaphore functions
1579
1580 VkResult anv_CreateSemaphore(
1581 VkDevice device,
1582 const VkSemaphoreCreateInfo* pCreateInfo,
1583 const VkAllocationCallbacks* pAllocator,
1584 VkSemaphore* pSemaphore)
1585 {
1586 /* The DRM execbuffer ioctl always execute in-oder, even between different
1587 * rings. As such, there's nothing to do for the user space semaphore.
1588 */
1589
1590 *pSemaphore = (VkSemaphore)1;
1591
1592 return VK_SUCCESS;
1593 }
1594
1595 void anv_DestroySemaphore(
1596 VkDevice device,
1597 VkSemaphore semaphore,
1598 const VkAllocationCallbacks* pAllocator)
1599 {
1600 }
1601
1602 // Event functions
1603
1604 VkResult anv_CreateEvent(
1605 VkDevice _device,
1606 const VkEventCreateInfo* pCreateInfo,
1607 const VkAllocationCallbacks* pAllocator,
1608 VkEvent* pEvent)
1609 {
1610 ANV_FROM_HANDLE(anv_device, device, _device);
1611 struct anv_state state;
1612 struct anv_event *event;
1613
1614 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_EVENT_CREATE_INFO);
1615
1616 state = anv_state_pool_alloc(&device->dynamic_state_pool,
1617 sizeof(*event), 8);
1618 event = state.map;
1619 event->state = state;
1620 event->semaphore = VK_EVENT_RESET;
1621
1622 if (!device->info.has_llc) {
1623 /* Make sure the writes we're flushing have landed. */
1624 __builtin_ia32_mfence();
1625 __builtin_ia32_clflush(event);
1626 }
1627
1628 *pEvent = anv_event_to_handle(event);
1629
1630 return VK_SUCCESS;
1631 }
1632
1633 void anv_DestroyEvent(
1634 VkDevice _device,
1635 VkEvent _event,
1636 const VkAllocationCallbacks* pAllocator)
1637 {
1638 ANV_FROM_HANDLE(anv_device, device, _device);
1639 ANV_FROM_HANDLE(anv_event, event, _event);
1640
1641 anv_state_pool_free(&device->dynamic_state_pool, event->state);
1642 }
1643
1644 VkResult anv_GetEventStatus(
1645 VkDevice _device,
1646 VkEvent _event)
1647 {
1648 ANV_FROM_HANDLE(anv_device, device, _device);
1649 ANV_FROM_HANDLE(anv_event, event, _event);
1650
1651 if (!device->info.has_llc) {
1652 /* Invalidate read cache before reading event written by GPU. */
1653 __builtin_ia32_clflush(event);
1654 __builtin_ia32_mfence();
1655
1656 }
1657
1658 return event->semaphore;
1659 }
1660
1661 VkResult anv_SetEvent(
1662 VkDevice _device,
1663 VkEvent _event)
1664 {
1665 ANV_FROM_HANDLE(anv_device, device, _device);
1666 ANV_FROM_HANDLE(anv_event, event, _event);
1667
1668 event->semaphore = VK_EVENT_SET;
1669
1670 if (!device->info.has_llc) {
1671 /* Make sure the writes we're flushing have landed. */
1672 __builtin_ia32_mfence();
1673 __builtin_ia32_clflush(event);
1674 }
1675
1676 return VK_SUCCESS;
1677 }
1678
1679 VkResult anv_ResetEvent(
1680 VkDevice _device,
1681 VkEvent _event)
1682 {
1683 ANV_FROM_HANDLE(anv_device, device, _device);
1684 ANV_FROM_HANDLE(anv_event, event, _event);
1685
1686 event->semaphore = VK_EVENT_RESET;
1687
1688 if (!device->info.has_llc) {
1689 /* Make sure the writes we're flushing have landed. */
1690 __builtin_ia32_mfence();
1691 __builtin_ia32_clflush(event);
1692 }
1693
1694 return VK_SUCCESS;
1695 }
1696
1697 // Buffer functions
1698
1699 VkResult anv_CreateBuffer(
1700 VkDevice _device,
1701 const VkBufferCreateInfo* pCreateInfo,
1702 const VkAllocationCallbacks* pAllocator,
1703 VkBuffer* pBuffer)
1704 {
1705 ANV_FROM_HANDLE(anv_device, device, _device);
1706 struct anv_buffer *buffer;
1707
1708 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
1709
1710 buffer = anv_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
1711 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1712 if (buffer == NULL)
1713 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1714
1715 buffer->size = pCreateInfo->size;
1716 buffer->usage = pCreateInfo->usage;
1717 buffer->bo = NULL;
1718 buffer->offset = 0;
1719
1720 *pBuffer = anv_buffer_to_handle(buffer);
1721
1722 return VK_SUCCESS;
1723 }
1724
1725 void anv_DestroyBuffer(
1726 VkDevice _device,
1727 VkBuffer _buffer,
1728 const VkAllocationCallbacks* pAllocator)
1729 {
1730 ANV_FROM_HANDLE(anv_device, device, _device);
1731 ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
1732
1733 anv_free2(&device->alloc, pAllocator, buffer);
1734 }
1735
1736 void
1737 anv_fill_buffer_surface_state(struct anv_device *device, struct anv_state state,
1738 enum isl_format format,
1739 uint32_t offset, uint32_t range, uint32_t stride)
1740 {
1741 isl_buffer_fill_state(&device->isl_dev, state.map,
1742 .address = offset,
1743 .mocs = device->default_mocs,
1744 .size = range,
1745 .format = format,
1746 .stride = stride);
1747
1748 if (!device->info.has_llc)
1749 anv_state_clflush(state);
1750 }
1751
1752 void anv_DestroySampler(
1753 VkDevice _device,
1754 VkSampler _sampler,
1755 const VkAllocationCallbacks* pAllocator)
1756 {
1757 ANV_FROM_HANDLE(anv_device, device, _device);
1758 ANV_FROM_HANDLE(anv_sampler, sampler, _sampler);
1759
1760 anv_free2(&device->alloc, pAllocator, sampler);
1761 }
1762
1763 VkResult anv_CreateFramebuffer(
1764 VkDevice _device,
1765 const VkFramebufferCreateInfo* pCreateInfo,
1766 const VkAllocationCallbacks* pAllocator,
1767 VkFramebuffer* pFramebuffer)
1768 {
1769 ANV_FROM_HANDLE(anv_device, device, _device);
1770 struct anv_framebuffer *framebuffer;
1771
1772 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
1773
1774 size_t size = sizeof(*framebuffer) +
1775 sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount;
1776 framebuffer = anv_alloc2(&device->alloc, pAllocator, size, 8,
1777 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1778 if (framebuffer == NULL)
1779 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
1780
1781 framebuffer->attachment_count = pCreateInfo->attachmentCount;
1782 for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
1783 VkImageView _iview = pCreateInfo->pAttachments[i];
1784 framebuffer->attachments[i] = anv_image_view_from_handle(_iview);
1785 }
1786
1787 framebuffer->width = pCreateInfo->width;
1788 framebuffer->height = pCreateInfo->height;
1789 framebuffer->layers = pCreateInfo->layers;
1790
1791 *pFramebuffer = anv_framebuffer_to_handle(framebuffer);
1792
1793 return VK_SUCCESS;
1794 }
1795
1796 void anv_DestroyFramebuffer(
1797 VkDevice _device,
1798 VkFramebuffer _fb,
1799 const VkAllocationCallbacks* pAllocator)
1800 {
1801 ANV_FROM_HANDLE(anv_device, device, _device);
1802 ANV_FROM_HANDLE(anv_framebuffer, fb, _fb);
1803
1804 anv_free2(&device->alloc, pAllocator, fb);
1805 }