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