#include "drm-uapi/drm_fourcc.h"
#include "anv_private.h"
-#include "util/strtod.h"
#include "util/debug.h"
#include "util/build_id.h"
#include "util/disk_cache.h"
#include "util/os_file.h"
#include "util/u_atomic.h"
#include "util/u_string.h"
+#include "util/xmlpool.h"
#include "git_sha1.h"
#include "vk_util.h"
+#include "common/gen_aux_map.h"
#include "common/gen_defines.h"
#include "compiler/glsl_types.h"
#include "genxml/gen7_pack.h"
+static const char anv_dri_options_xml[] =
+DRI_CONF_BEGIN
+ DRI_CONF_SECTION_PERFORMANCE
+ DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
+ DRI_CONF_VK_X11_STRICT_IMAGE_COUNT("false")
+ DRI_CONF_SECTION_END
+DRI_CONF_END;
+
/* This is probably far to big but it reflects the max size used for messages
* in OpenGLs KHR_debug.
*/
char str[MAX_DEBUG_MESSAGE_LENGTH];
struct anv_device *device = (struct anv_device *)data;
- if (list_empty(&device->instance->debug_report_callbacks.callbacks))
+ if (list_is_empty(&device->instance->debug_report_callbacks.callbacks))
return;
va_list args;
assert(strlen(path) < ARRAY_SIZE(device->path));
snprintf(device->path, ARRAY_SIZE(device->path), "%s", path);
- device->no_hw = getenv("INTEL_NO_HW") != NULL;
+ if (!gen_get_device_info_from_fd(fd, &device->info)) {
+ result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
+ goto fail;
+ }
+ device->chipset_id = device->info.chipset_id;
+ device->no_hw = device->info.no_hw;
- const int pci_id_override = gen_get_pci_device_id_override();
- if (pci_id_override < 0) {
- device->chipset_id = anv_gem_get_param(fd, I915_PARAM_CHIPSET_ID);
- if (!device->chipset_id) {
- result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
- goto fail;
- }
- } else {
- device->chipset_id = pci_id_override;
+ if (getenv("INTEL_NO_HW") != NULL)
device->no_hw = true;
- }
device->pci_info.domain = drm_device->businfo.pci->domain;
device->pci_info.bus = drm_device->businfo.pci->bus;
device->pci_info.function = drm_device->businfo.pci->func;
device->name = gen_get_device_name(device->chipset_id);
- if (!gen_get_device_info(device->chipset_id, &device->info)) {
- result = vk_error(VK_ERROR_INCOMPATIBLE_DRIVER);
- goto fail;
- }
if (device->info.is_haswell) {
intel_logw("Haswell Vulkan support is incomplete");
intel_logw("Bay Trail Vulkan support is incomplete");
} else if (device->info.gen >= 8 && device->info.gen <= 11) {
/* Gen8-11 fully supported */
+ } else if (device->info.gen == 12) {
+ intel_logw("Vulkan is not yet fully supported on gen12");
} else {
result = vk_errorf(device->instance, device,
VK_ERROR_INCOMPATIBLE_DRIVER,
goto fail;
}
+ device->perf = anv_get_perf(&device->info, fd);
+
anv_physical_device_get_supported_extensions(device,
&device->supported_extensions);
anv_finish_wsi(device);
anv_physical_device_free_disk_cache(device);
ralloc_free(device->compiler);
+ ralloc_free(device->perf);
close(device->local_fd);
if (device->master_fd >= 0)
close(device->master_fd);
}
}
+ struct anv_physical_device *pdevice = &instance->physicalDevice;
+ for (unsigned i = 0; i < ARRAY_SIZE(pdevice->dispatch.entrypoints); i++) {
+ /* Vulkan requires that entrypoints for extensions which have not been
+ * enabled must not be advertised.
+ */
+ if (!anv_physical_device_entrypoint_is_enabled(i, instance->app_info.api_version,
+ &instance->enabled_extensions)) {
+ pdevice->dispatch.entrypoints[i] = NULL;
+ } else {
+ pdevice->dispatch.entrypoints[i] =
+ anv_physical_device_dispatch_table.entrypoints[i];
+ }
+ }
+
for (unsigned i = 0; i < ARRAY_SIZE(instance->device_dispatch.entrypoints); i++) {
/* Vulkan requires that entrypoints for extensions which have not been
* enabled must not be advertised.
instance->pipeline_cache_enabled =
env_var_as_boolean("ANV_ENABLE_PIPELINE_CACHE", true);
- _mesa_locale_init();
glsl_type_singleton_init_or_ref();
VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
+ driParseOptionInfo(&instance->available_dri_options, anv_dri_options_xml);
+ driParseConfigFiles(&instance->dri_options, &instance->available_dri_options,
+ 0, "anv", NULL,
+ instance->app_info.engine_name,
+ instance->app_info.engine_version);
+
*pInstance = anv_instance_to_handle(instance);
return VK_SUCCESS;
vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
glsl_type_singleton_decref();
- _mesa_locale_fini();
+
+ driDestroyOptionCache(&instance->dri_options);
+ driDestroyOptionInfo(&instance->available_dri_options);
vk_free(&instance->alloc, instance);
}
.depthClamp = true,
.depthBiasClamp = true,
.fillModeNonSolid = true,
- .depthBounds = false,
+ .depthBounds = pdevice->info.gen >= 12,
.wideLines = true,
.largePoints = true,
.alphaToOne = true,
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
+ VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
+ (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
+ features->rectangularLines = true;
+ features->bresenhamLines = true;
+ features->smoothLines = true;
+ features->stippledRectangularLines = false;
+ features->stippledBresenhamLines = true;
+ features->stippledSmoothLines = false;
+ break;
+ }
+
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
VkPhysicalDeviceMultiviewFeatures *features =
(VkPhysicalDeviceMultiviewFeatures *)ext;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
+ VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
+ (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
+ features->pipelineExecutableInfo = true;
+ break;
+ }
+
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
VkPhysicalDeviceProtectedMemoryFeatures *features = (void *)ext;
features->protectedMemory = false;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_CLOCK_FEATURES_KHR: {
+ VkPhysicalDeviceShaderClockFeaturesKHR *features =
+ (VkPhysicalDeviceShaderClockFeaturesKHR *)ext;
+ features->shaderSubgroupClock = true;
+ features->shaderDeviceClock = false;
+ break;
+ }
+
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
VkPhysicalDeviceShaderDrawParametersFeatures *features = (void *)ext;
features->shaderDrawParameters = true;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_SUBGROUP_EXTENDED_TYPES_FEATURES_KHR: {
+ VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *features =
+ (VkPhysicalDeviceShaderSubgroupExtendedTypesFeaturesKHR *)ext;
+ features->shaderSubgroupExtendedTypes = true;
+ break;
+ }
+
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT: {
+ VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *features =
+ (VkPhysicalDeviceSubgroupSizeControlFeaturesEXT *)ext;
+ features->subgroupSizeControl = true;
+ features->computeFullSubgroups = true;
+ break;
+ }
+
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
(VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_MEMORY_MODEL_FEATURES_KHR: {
+ VkPhysicalDeviceVulkanMemoryModelFeaturesKHR *features = (void *)ext;
+ features->vulkanMemoryModel = true;
+ features->vulkanMemoryModelDeviceScope = true;
+ features->vulkanMemoryModelAvailabilityVisibilityChains = true;
+ break;
+ }
+
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: {
VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features =
(VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext;
const uint32_t max_images =
pdevice->has_bindless_images ? UINT16_MAX : MAX_IMAGES;
- /* The moment we have anything bindless, claim a high per-stage limit */
+ /* If we can use bindless for everything, claim a high per-stage limit,
+ * otherwise use the binding table size, minus the slots reserved for
+ * render targets and one slot for the descriptor buffer. */
const uint32_t max_per_stage =
- pdevice->has_a64_buffer_access ? UINT32_MAX :
- MAX_BINDING_TABLE_SIZE - MAX_RTS;
+ pdevice->has_bindless_images && pdevice->has_a64_buffer_access
+ ? UINT32_MAX : MAX_BINDING_TABLE_SIZE - MAX_RTS - 1;
+
+ const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
VkSampleCountFlags sample_counts =
isl_device_get_sample_counts(&pdevice->isl_dev);
.maxFragmentCombinedOutputResources = 8,
.maxComputeSharedMemorySize = 64 * 1024,
.maxComputeWorkGroupCount = { 65535, 65535, 65535 },
- .maxComputeWorkGroupInvocations = 32 * devinfo->max_cs_threads,
+ .maxComputeWorkGroupInvocations = max_workgroup_size,
.maxComputeWorkGroupSize = {
- 16 * devinfo->max_cs_threads,
- 16 * devinfo->max_cs_threads,
- 16 * devinfo->max_cs_threads,
+ max_workgroup_size,
+ max_workgroup_size,
+ max_workgroup_size,
},
.subPixelPrecisionBits = 8,
.subTexelPrecisionBits = 8,
.maxCombinedClipAndCullDistances = 8,
.discreteQueuePriorities = 2,
.pointSizeRange = { 0.125, 255.875 },
- .lineWidthRange = { 0.0, 7.9921875 },
+ .lineWidthRange = {
+ 0.0,
+ (devinfo->gen >= 9 || devinfo->is_cherryview) ?
+ 2047.9921875 : 7.9921875,
+ },
.pointSizeGranularity = (1.0 / 8.0),
.lineWidthGranularity = (1.0 / 128.0),
- .strictLines = false, /* FINISHME */
+ .strictLines = false,
.standardSampleLocations = true,
.optimalBufferCopyOffsetAlignment = 128,
.optimalBufferCopyRowPitchAlignment = 128,
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
+ VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
+ (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
+ /* In the Skylake PRM Vol. 7, subsection titled "GIQ (Diamond)
+ * Sampling Rules - Legacy Mode", it says the following:
+ *
+ * "Note that the device divides a pixel into a 16x16 array of
+ * subpixels, referenced by their upper left corners."
+ *
+ * This is the only known reference in the PRMs to the subpixel
+ * precision of line rasterization and a "16x16 array of subpixels"
+ * implies 4 subpixel precision bits. Empirical testing has shown
+ * that 4 subpixel precision bits applies to all line rasterization
+ * types.
+ */
+ props->lineSubPixelPrecisionBits = 4;
+ break;
+ }
+
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
VkPhysicalDeviceMaintenance3Properties *props =
(VkPhysicalDeviceMaintenance3Properties *)ext;
break;
}
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wswitch"
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENTATION_PROPERTIES_ANDROID: {
+ VkPhysicalDevicePresentationPropertiesANDROID *props =
+ (VkPhysicalDevicePresentationPropertiesANDROID *)ext;
+ props->sharedImage = VK_FALSE;
+ break;
+ }
+#pragma GCC diagnostic pop
+
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_PROPERTIES: {
VkPhysicalDeviceProtectedMemoryProperties *props =
(VkPhysicalDeviceProtectedMemoryProperties *)ext;
props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT;
break;
}
+ case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR : {
+ VkPhysicalDeviceFloatControlsPropertiesKHR *properties = (void *)ext;
+ properties->denormBehaviorIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL_KHR;
+ properties->roundingModeIndependence = VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_NONE_KHR;
+
+ /* Broadwell does not support HF denorms and there are restrictions
+ * other gens. According to Kabylake's PRM:
+ *
+ * "math - Extended Math Function
+ * [...]
+ * Restriction : Half-float denorms are always retained."
+ */
+ properties->shaderDenormFlushToZeroFloat16 = false;
+ properties->shaderDenormPreserveFloat16 = pdevice->info.gen > 8;
+ properties->shaderRoundingModeRTEFloat16 = true;
+ properties->shaderRoundingModeRTZFloat16 = true;
+ properties->shaderSignedZeroInfNanPreserveFloat16 = true;
+
+ properties->shaderDenormFlushToZeroFloat32 = true;
+ properties->shaderDenormPreserveFloat32 = true;
+ properties->shaderRoundingModeRTEFloat32 = true;
+ properties->shaderRoundingModeRTZFloat32 = true;
+ properties->shaderSignedZeroInfNanPreserveFloat32 = true;
+
+ properties->shaderDenormFlushToZeroFloat64 = true;
+ properties->shaderDenormPreserveFloat64 = true;
+ properties->shaderRoundingModeRTEFloat64 = true;
+ properties->shaderRoundingModeRTZFloat64 = true;
+ properties->shaderSignedZeroInfNanPreserveFloat64 = true;
+ break;
+ }
case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: {
VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props =
if (idx >= 0)
return instance->dispatch.entrypoints[idx];
+ idx = anv_get_physical_device_entrypoint_index(pName);
+ if (idx >= 0)
+ return instance->physicalDevice.dispatch.entrypoints[idx];
+
idx = anv_get_device_entrypoint_index(pName);
if (idx >= 0)
return instance->device_dispatch.entrypoints[idx];
return device->dispatch.entrypoints[idx];
}
+/* With version 4+ of the loader interface the ICD should expose
+ * vk_icdGetPhysicalDeviceProcAddr()
+ */
+PUBLIC
+VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vk_icdGetPhysicalDeviceProcAddr(
+ VkInstance _instance,
+ const char* pName);
+
+PFN_vkVoidFunction vk_icdGetPhysicalDeviceProcAddr(
+ VkInstance _instance,
+ const char* pName)
+{
+ ANV_FROM_HANDLE(anv_instance, instance, _instance);
+
+ if (!pName || !instance)
+ return NULL;
+
+ int idx = anv_get_physical_device_entrypoint_index(pName);
+ if (idx < 0)
+ return NULL;
+
+ return instance->physicalDevice.dispatch.entrypoints[idx];
+}
+
+
VkResult
anv_CreateDebugReportCallbackEXT(VkInstance _instance,
const VkDebugReportCallbackCreateInfoEXT* pCreateInfo,
{
const struct anv_device_dispatch_table *genX_table;
switch (device->info.gen) {
+ case 12:
+ genX_table = &gen12_device_dispatch_table;
+ break;
case 11:
genX_table = &gen11_device_dispatch_table;
break;
return (struct gen_batch_decode_bo) { };
}
+struct gen_aux_map_buffer {
+ struct gen_buffer base;
+ struct anv_state state;
+};
+
+static struct gen_buffer *
+gen_aux_map_buffer_alloc(void *driver_ctx, uint32_t size)
+{
+ struct gen_aux_map_buffer *buf = malloc(sizeof(struct gen_aux_map_buffer));
+ if (!buf)
+ return NULL;
+
+ struct anv_device *device = (struct anv_device*)driver_ctx;
+ assert(device->instance->physicalDevice.supports_48bit_addresses &&
+ device->instance->physicalDevice.use_softpin);
+
+ struct anv_state_pool *pool = &device->dynamic_state_pool;
+ buf->state = anv_state_pool_alloc(pool, size, size);
+
+ buf->base.gpu = pool->block_pool.bo->offset + buf->state.offset;
+ buf->base.gpu_end = buf->base.gpu + buf->state.alloc_size;
+ buf->base.map = buf->state.map;
+ buf->base.driver_bo = &buf->state;
+ return &buf->base;
+}
+
+static void
+gen_aux_map_buffer_free(void *driver_ctx, struct gen_buffer *buffer)
+{
+ struct gen_aux_map_buffer *buf = (struct gen_aux_map_buffer*)buffer;
+ struct anv_device *device = (struct anv_device*)driver_ctx;
+ struct anv_state_pool *pool = &device->dynamic_state_pool;
+ anv_state_pool_free(pool, buf->state);
+ free(buf);
+}
+
+static struct gen_mapped_pinned_buffer_alloc aux_map_allocator = {
+ .alloc = gen_aux_map_buffer_alloc,
+ .free = gen_aux_map_buffer_free,
+};
+
VkResult anv_CreateDevice(
VkPhysicalDevice physicalDevice,
const VkDeviceCreateInfo* pCreateInfo,
if (physical_device->use_softpin) {
if (pthread_mutex_init(&device->vma_mutex, NULL) != 0) {
result = vk_error(VK_ERROR_INITIALIZATION_FAILED);
- goto fail_fd;
+ goto fail_context_id;
}
/* keep the page with address zero out of the allocator */
vk_priority_to_gen(priority));
if (err != 0 && priority > VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_EXT) {
result = vk_error(VK_ERROR_NOT_PERMITTED_EXT);
- goto fail_fd;
+ goto fail_vmas;
}
}
goto fail_surface_state_pool;
}
+ if (device->info.gen >= 12) {
+ device->aux_map_ctx = gen_aux_map_init(device, &aux_map_allocator,
+ &physical_device->info);
+ if (!device->aux_map_ctx)
+ goto fail_binding_table_pool;
+ }
+
result = anv_bo_init_new(&device->workaround_bo, device, 4096);
if (result != VK_SUCCESS)
- goto fail_binding_table_pool;
+ goto fail_surface_aux_map_pool;
if (physical_device->use_softpin)
device->workaround_bo.flags |= EXEC_OBJECT_PINNED;
case 11:
result = gen11_init_device_state(device);
break;
+ case 12:
+ result = gen12_init_device_state(device);
+ break;
default:
/* Shouldn't get here as we don't create physical devices for any other
* gens. */
anv_device_init_border_colors(device);
+ anv_device_perf_init(device);
+
*pDevice = anv_device_to_handle(device);
return VK_SUCCESS;
anv_scratch_pool_finish(device, &device->scratch_pool);
anv_gem_munmap(device->workaround_bo.map, device->workaround_bo.size);
anv_gem_close(device, device->workaround_bo.gem_handle);
+ fail_surface_aux_map_pool:
+ if (device->info.gen >= 12) {
+ gen_aux_map_finish(device->aux_map_ctx);
+ device->aux_map_ctx = NULL;
+ }
fail_binding_table_pool:
if (physical_device->use_softpin)
anv_state_pool_finish(&device->binding_table_pool);
pthread_cond_destroy(&device->queue_submit);
fail_mutex:
pthread_mutex_destroy(&device->mutex);
+ fail_vmas:
+ if (physical_device->use_softpin) {
+ util_vma_heap_finish(&device->vma_hi);
+ util_vma_heap_finish(&device->vma_lo);
+ }
fail_context_id:
anv_gem_destroy_context(device, device->context_id);
fail_fd:
* BO will go away in a couple of lines so we don't actually leak.
*/
anv_state_pool_free(&device->dynamic_state_pool, device->border_colors);
+ anv_state_pool_free(&device->dynamic_state_pool, device->slice_hash);
#endif
anv_scratch_pool_finish(device, &device->scratch_pool);
if (device->info.gen >= 10)
anv_gem_close(device, device->hiz_clear_bo.gem_handle);
+ if (device->info.gen >= 12) {
+ gen_aux_map_finish(device->aux_map_ctx);
+ device->aux_map_ctx = NULL;
+ }
+
if (physical_device->use_softpin)
anv_state_pool_finish(&device->binding_table_pool);
anv_state_pool_finish(&device->surface_state_pool);
anv_bo_pool_finish(&device->batch_bo_pool);
+ if (physical_device->use_softpin) {
+ util_vma_heap_finish(&device->vma_hi);
+ util_vma_heap_finish(&device->vma_lo);
+ }
+
pthread_cond_destroy(&device->queue_submit);
pthread_mutex_destroy(&device->mutex);
VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
result = anv_bo_cache_import(device, &device->bo_cache, fd_info->fd,
- bo_flags | ANV_BO_EXTERNAL, &mem->bo);
+ bo_flags, &mem->bo);
if (result != VK_SUCCESS)
goto fail;
/* Regular allocate (not importing memory). */
- if (export_info && export_info->handleTypes)
- bo_flags |= ANV_BO_EXTERNAL;
-
+ bool is_external = export_info && export_info->handleTypes;
result = anv_bo_cache_alloc(device, &device->bo_cache,
- pAllocateInfo->allocationSize, bo_flags,
+ pAllocateInfo->allocationSize,
+ bo_flags, is_external,
&mem->bo);
if (result != VK_SUCCESS)
goto fail;
* - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
* vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
* because the loader no longer does so.
+ *
+ * - Loader interface v4 differs from v3 in:
+ * - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
*/
- *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
+ *pSupportedVersion = MIN2(*pSupportedVersion, 4u);
return VK_SUCCESS;
}
projects / mesa.git / blobdiff