X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fintel%2Fvulkan%2Fanv_device.c;h=50d910780cbac3dda369b9fad1fc07fbd9699068;hb=0616b7ac90cf4f86bb409d34101e3a3cceac8cbe;hp=3173bd715a853651a0eb09777e9cbaf8346b0b74;hpb=47fc359822494935852de1e70e4d840b2fe6a25c;p=mesa.git diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c index 3173bd715a8..50d910780cb 100644 --- a/src/intel/vulkan/anv_device.c +++ b/src/intel/vulkan/anv_device.c @@ -37,7 +37,10 @@ #include "util/build_id.h" #include "util/disk_cache.h" #include "util/mesa-sha1.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_defines.h" @@ -45,6 +48,14 @@ #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. */ @@ -150,6 +161,8 @@ anv_physical_device_init_heaps(struct anv_physical_device *device, int fd) */ device->memory.heap_count = 1; device->memory.heaps[0] = (struct anv_memory_heap) { + .vma_start = LOW_HEAP_MIN_ADDRESS, + .vma_size = LOW_HEAP_SIZE, .size = heap_size, .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT, .supports_48bit_addresses = false, @@ -169,11 +182,19 @@ anv_physical_device_init_heaps(struct anv_physical_device *device, int fd) device->memory.heap_count = 2; device->memory.heaps[0] = (struct anv_memory_heap) { + .vma_start = HIGH_HEAP_MIN_ADDRESS, + /* Leave the last 4GiB out of the high vma range, so that no state + * base address + size can overflow 48 bits. For more information see + * the comment about Wa32bitGeneralStateOffset in anv_allocator.c + */ + .vma_size = gtt_size - (1ull << 32) - HIGH_HEAP_MIN_ADDRESS, .size = heap_size_48bit, .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT, .supports_48bit_addresses = true, }; device->memory.heaps[1] = (struct anv_memory_heap) { + .vma_start = LOW_HEAP_MIN_ADDRESS, + .vma_size = LOW_HEAP_SIZE, .size = heap_size_32bit, .flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT, .supports_48bit_addresses = false, @@ -266,6 +287,14 @@ anv_physical_device_init_uuids(struct anv_physical_device *device) _mesa_sha1_update(&sha1_ctx, build_id_data(note), build_id_len); _mesa_sha1_update(&sha1_ctx, &device->chipset_id, sizeof(device->chipset_id)); + _mesa_sha1_update(&sha1_ctx, &device->always_use_bindless, + sizeof(device->always_use_bindless)); + _mesa_sha1_update(&sha1_ctx, &device->has_a64_buffer_access, + sizeof(device->has_a64_buffer_access)); + _mesa_sha1_update(&sha1_ctx, &device->has_bindless_images, + sizeof(device->has_bindless_images)); + _mesa_sha1_update(&sha1_ctx, &device->has_bindless_samplers, + sizeof(device->has_bindless_samplers)); _mesa_sha1_final(&sha1_ctx, sha1); memcpy(device->pipeline_cache_uuid, sha1, VK_UUID_SIZE); @@ -298,7 +327,7 @@ anv_physical_device_init_disk_cache(struct anv_physical_device *device) { #ifdef ENABLE_SHADER_CACHE char renderer[10]; - MAYBE_UNUSED int len = snprintf(renderer, sizeof(renderer), "anv_%04x", + ASSERTED int len = snprintf(renderer, sizeof(renderer), "anv_%04x", device->chipset_id); assert(len == sizeof(renderer) - 2); @@ -324,6 +353,29 @@ anv_physical_device_free_disk_cache(struct anv_physical_device *device) #endif } +static uint64_t +get_available_system_memory() +{ + char *meminfo = os_read_file("/proc/meminfo"); + if (!meminfo) + return 0; + + char *str = strstr(meminfo, "MemAvailable:"); + if (!str) { + free(meminfo); + return 0; + } + + uint64_t kb_mem_available; + if (sscanf(str, "MemAvailable: %" PRIx64, &kb_mem_available) == 1) { + free(meminfo); + return kb_mem_available << 10; + } + + free(meminfo); + return 0; +} + static VkResult anv_physical_device_init(struct anv_physical_device *device, struct anv_instance *instance, @@ -347,19 +399,15 @@ anv_physical_device_init(struct anv_physical_device *device, 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; @@ -367,10 +415,6 @@ anv_physical_device_init(struct anv_physical_device *device, 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"); @@ -378,10 +422,10 @@ anv_physical_device_init(struct anv_physical_device *device, intel_logw("Ivy Bridge Vulkan support is incomplete"); } else if (device->info.gen == 7 && device->info.is_baytrail) { intel_logw("Bay Trail Vulkan support is incomplete"); - } else if (device->info.gen >= 8 && device->info.gen <= 10) { - /* Gen8-10 fully supported */ - } else if (device->info.gen == 11) { - intel_logw("Vulkan is not yet fully supported on gen11."); + } 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, @@ -441,6 +485,30 @@ anv_physical_device_init(struct anv_physical_device *device, device->has_context_isolation = anv_gem_get_param(fd, I915_PARAM_HAS_CONTEXT_ISOLATION); + device->always_use_bindless = + env_var_as_boolean("ANV_ALWAYS_BINDLESS", false); + + /* We first got the A64 messages on broadwell and we can only use them if + * we can pass addresses directly into the shader which requires softpin. + */ + device->has_a64_buffer_access = device->info.gen >= 8 && + device->use_softpin; + + /* We first get bindless image access on Skylake and we can only really do + * it if we don't have any relocations so we need softpin. + */ + device->has_bindless_images = device->info.gen >= 9 && + device->use_softpin; + + /* We've had bindless samplers since Ivy Bridge (forever in Vulkan terms) + * because it's just a matter of setting the sampler address in the sample + * message header. However, we've not bothered to wire it up for vec4 so + * we leave it disabled on gen7. + */ + device->has_bindless_samplers = device->info.gen >= 8; + + device->has_mem_available = get_available_system_memory() != 0; + /* Starting with Gen10, the timestamp frequency of the command streamer may * vary from one part to another. We can query the value from the kernel. */ @@ -707,9 +775,16 @@ VkResult anv_CreateInstance( 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; @@ -737,9 +812,12 @@ void anv_DestroyInstance( vk_debug_report_instance_destroy(&instance->debug_report_callbacks); - _mesa_glsl_release_types(); + glsl_type_singleton_decref(); _mesa_locale_fini(); + driDestroyOptionCache(&instance->dri_options); + driDestroyOptionInfo(&instance->available_dri_options); + vk_free(&instance->alloc, instance); } @@ -835,7 +913,7 @@ VkResult anv_EnumeratePhysicalDeviceGroups( memset(p->physicalDevices, 0, sizeof(p->physicalDevices)); p->physicalDevices[0] = anv_physical_device_to_handle(&instance->physicalDevice); - p->subsetAllocation = VK_FALSE; + p->subsetAllocation = false; vk_foreach_struct(ext, p->pNext) anv_debug_ignored_stype(ext->sType); @@ -927,8 +1005,6 @@ void anv_GetPhysicalDeviceFeatures2( case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR: { VkPhysicalDevice8BitStorageFeaturesKHR *features = (VkPhysicalDevice8BitStorageFeaturesKHR *)ext; - ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice); - features->storageBuffer8BitAccess = pdevice->info.gen >= 8; features->uniformAndStorageBuffer8BitAccess = pdevice->info.gen >= 8; features->storagePushConstant8 = pdevice->info.gen >= 8; @@ -945,20 +1021,25 @@ void anv_GetPhysicalDeviceFeatures2( break; } - case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_ADDRESS_FEATURES_EXT: { - VkPhysicalDeviceBufferAddressFeaturesEXT *features = (void *)ext; - features->bufferDeviceAddress = pdevice->use_softpin && - pdevice->info.gen >= 8; + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BUFFER_DEVICE_ADDRESS_FEATURES_EXT: { + VkPhysicalDeviceBufferDeviceAddressFeaturesEXT *features = (void *)ext; + features->bufferDeviceAddress = pdevice->has_a64_buffer_access; features->bufferDeviceAddressCaptureReplay = false; features->bufferDeviceAddressMultiDevice = false; break; } + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_SHADER_DERIVATIVES_FEATURES_NV: { + VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *features = + (VkPhysicalDeviceComputeShaderDerivativesFeaturesNV *)ext; + features->computeDerivativeGroupQuads = true; + features->computeDerivativeGroupLinear = true; + break; + } + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: { VkPhysicalDeviceConditionalRenderingFeaturesEXT *features = (VkPhysicalDeviceConditionalRenderingFeaturesEXT*)ext; - ANV_FROM_HANDLE(anv_physical_device, pdevice, physicalDevice); - features->conditionalRendering = pdevice->info.gen >= 8 || pdevice->info.is_haswell; features->inheritedConditionalRendering = pdevice->info.gen >= 8 || @@ -969,7 +1050,63 @@ void anv_GetPhysicalDeviceFeatures2( case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: { VkPhysicalDeviceDepthClipEnableFeaturesEXT *features = (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext; - features->depthClipEnable = VK_TRUE; + features->depthClipEnable = true; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR: { + VkPhysicalDeviceFloat16Int8FeaturesKHR *features = (void *)ext; + features->shaderFloat16 = pdevice->info.gen >= 8; + features->shaderInt8 = pdevice->info.gen >= 8; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADER_INTERLOCK_FEATURES_EXT: { + VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *features = + (VkPhysicalDeviceFragmentShaderInterlockFeaturesEXT *)ext; + features->fragmentShaderSampleInterlock = pdevice->info.gen >= 9; + features->fragmentShaderPixelInterlock = pdevice->info.gen >= 9; + features->fragmentShaderShadingRateInterlock = false; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT: { + VkPhysicalDeviceHostQueryResetFeaturesEXT *features = + (VkPhysicalDeviceHostQueryResetFeaturesEXT *)ext; + features->hostQueryReset = true; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: { + VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features = + (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *)ext; + features->shaderInputAttachmentArrayDynamicIndexing = false; + features->shaderUniformTexelBufferArrayDynamicIndexing = true; + features->shaderStorageTexelBufferArrayDynamicIndexing = true; + features->shaderUniformBufferArrayNonUniformIndexing = false; + features->shaderSampledImageArrayNonUniformIndexing = true; + features->shaderStorageBufferArrayNonUniformIndexing = true; + features->shaderStorageImageArrayNonUniformIndexing = true; + features->shaderInputAttachmentArrayNonUniformIndexing = false; + features->shaderUniformTexelBufferArrayNonUniformIndexing = true; + features->shaderStorageTexelBufferArrayNonUniformIndexing = true; + features->descriptorBindingUniformBufferUpdateAfterBind = false; + features->descriptorBindingSampledImageUpdateAfterBind = true; + features->descriptorBindingStorageImageUpdateAfterBind = true; + features->descriptorBindingStorageBufferUpdateAfterBind = true; + features->descriptorBindingUniformTexelBufferUpdateAfterBind = true; + features->descriptorBindingStorageTexelBufferUpdateAfterBind = true; + features->descriptorBindingUpdateUnusedWhilePending = true; + features->descriptorBindingPartiallyBound = true; + features->descriptorBindingVariableDescriptorCount = false; + features->runtimeDescriptorArray = true; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: { + VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features = + (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext; + features->indexTypeUint8 = true; break; } @@ -977,7 +1114,19 @@ void anv_GetPhysicalDeviceFeatures2( VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features = (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext; features->inlineUniformBlock = true; - features->descriptorBindingInlineUniformBlockUpdateAfterBind = false; + features->descriptorBindingInlineUniformBlockUpdateAfterBind = 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; } @@ -990,9 +1139,23 @@ void anv_GetPhysicalDeviceFeatures2( break; } + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGELESS_FRAMEBUFFER_FEATURES_KHR: { + VkPhysicalDeviceImagelessFramebufferFeaturesKHR *features = + (VkPhysicalDeviceImagelessFramebufferFeaturesKHR *)ext; + features->imagelessFramebuffer = true; + 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 = VK_FALSE; + features->protectedMemory = false; break; } @@ -1010,14 +1173,43 @@ void anv_GetPhysicalDeviceFeatures2( break; } - case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETER_FEATURES: { - VkPhysicalDeviceShaderDrawParameterFeatures *features = (void *)ext; + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR: { + VkPhysicalDeviceShaderAtomicInt64FeaturesKHR *features = (void *)ext; + features->shaderBufferInt64Atomics = + pdevice->info.gen >= 9 && pdevice->use_softpin; + features->shaderSharedInt64Atomics = VK_FALSE; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT: { + VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT *features = (void *)ext; + features->shaderDemoteToHelperInvocation = true; + 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_VARIABLE_POINTER_FEATURES: { - VkPhysicalDeviceVariablePointerFeatures *features = (void *)ext; + 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; + features->texelBufferAlignment = true; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: { + VkPhysicalDeviceVariablePointersFeatures *features = (void *)ext; features->variablePointersStorageBuffer = true; features->variablePointers = true; break; @@ -1026,23 +1218,30 @@ void anv_GetPhysicalDeviceFeatures2( case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: { VkPhysicalDeviceTransformFeedbackFeaturesEXT *features = (VkPhysicalDeviceTransformFeedbackFeaturesEXT *)ext; - features->transformFeedback = VK_TRUE; - features->geometryStreams = VK_TRUE; + features->transformFeedback = true; + features->geometryStreams = true; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR: { + VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *features = + (VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR *)ext; + features->uniformBufferStandardLayout = true; break; } case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: { VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features = (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext; - features->vertexAttributeInstanceRateDivisor = VK_TRUE; - features->vertexAttributeInstanceRateZeroDivisor = VK_TRUE; + features->vertexAttributeInstanceRateDivisor = true; + features->vertexAttributeInstanceRateZeroDivisor = true; break; } case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_YCBCR_IMAGE_ARRAYS_FEATURES_EXT: { VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *features = (VkPhysicalDeviceYcbcrImageArraysFeaturesEXT *)ext; - features->ycbcrImageArrays = VK_TRUE; + features->ycbcrImageArrays = true; break; } @@ -1053,6 +1252,11 @@ void anv_GetPhysicalDeviceFeatures2( } } +#define MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS 64 + +#define MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS 64 +#define MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS 256 + void anv_GetPhysicalDeviceProperties( VkPhysicalDevice physicalDevice, VkPhysicalDeviceProperties* pProperties) @@ -1064,10 +1268,21 @@ void anv_GetPhysicalDeviceProperties( const uint32_t max_raw_buffer_sz = devinfo->gen >= 7 ? (1ul << 30) : (1ul << 27); - const uint32_t max_samplers = (devinfo->gen >= 8 || devinfo->is_haswell) ? - 128 : 16; + const uint32_t max_ssbos = pdevice->has_a64_buffer_access ? UINT16_MAX : 64; + const uint32_t max_textures = + pdevice->has_bindless_images ? UINT16_MAX : 128; + const uint32_t max_samplers = + pdevice->has_bindless_samplers ? UINT16_MAX : + (devinfo->gen >= 8 || devinfo->is_haswell) ? 128 : 16; + 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 */ + const uint32_t max_per_stage = + pdevice->has_a64_buffer_access ? UINT32_MAX : + MAX_BINDING_TABLE_SIZE - MAX_RTS; - const uint32_t max_images = devinfo->gen < 9 ? MAX_GEN8_IMAGES : MAX_IMAGES; + const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads; VkSampleCountFlags sample_counts = isl_device_get_sample_counts(&pdevice->isl_dev); @@ -1089,20 +1304,20 @@ void anv_GetPhysicalDeviceProperties( .sparseAddressSpaceSize = 0, .maxBoundDescriptorSets = MAX_SETS, .maxPerStageDescriptorSamplers = max_samplers, - .maxPerStageDescriptorUniformBuffers = 64, - .maxPerStageDescriptorStorageBuffers = 64, - .maxPerStageDescriptorSampledImages = max_samplers, + .maxPerStageDescriptorUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, + .maxPerStageDescriptorStorageBuffers = max_ssbos, + .maxPerStageDescriptorSampledImages = max_textures, .maxPerStageDescriptorStorageImages = max_images, - .maxPerStageDescriptorInputAttachments = 64, - .maxPerStageResources = 250, + .maxPerStageDescriptorInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS, + .maxPerStageResources = max_per_stage, .maxDescriptorSetSamplers = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSamplers */ - .maxDescriptorSetUniformBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorUniformBuffers */ + .maxDescriptorSetUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS, /* number of stages * maxPerStageDescriptorUniformBuffers */ .maxDescriptorSetUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2, - .maxDescriptorSetStorageBuffers = 6 * 64, /* number of stages * maxPerStageDescriptorStorageBuffers */ + .maxDescriptorSetStorageBuffers = 6 * max_ssbos, /* number of stages * maxPerStageDescriptorStorageBuffers */ .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2, - .maxDescriptorSetSampledImages = 6 * max_samplers, /* number of stages * maxPerStageDescriptorSampledImages */ + .maxDescriptorSetSampledImages = 6 * max_textures, /* number of stages * maxPerStageDescriptorSampledImages */ .maxDescriptorSetStorageImages = 6 * max_images, /* number of stages * maxPerStageDescriptorStorageImages */ - .maxDescriptorSetInputAttachments = 256, + .maxDescriptorSetInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS, .maxVertexInputAttributes = MAX_VBS, .maxVertexInputBindings = MAX_VBS, .maxVertexInputAttributeOffset = 2047, @@ -1121,21 +1336,21 @@ void anv_GetPhysicalDeviceProperties( .maxGeometryOutputComponents = 128, .maxGeometryOutputVertices = 256, .maxGeometryTotalOutputComponents = 1024, - .maxFragmentInputComponents = 112, /* 128 components - (POS, PSIZ, CLIP_DIST0, CLIP_DIST1) */ + .maxFragmentInputComponents = 116, /* 128 components - (PSIZ, CLIP_DIST0, CLIP_DIST1) */ .maxFragmentOutputAttachments = 8, .maxFragmentDualSrcAttachments = 1, .maxFragmentCombinedOutputResources = 8, - .maxComputeSharedMemorySize = 32768, + .maxComputeSharedMemorySize = 64 * 1024, .maxComputeWorkGroupCount = { 65535, 65535, 65535 }, - .maxComputeWorkGroupInvocations = 16 * 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 = 4 /* FIXME */, - .mipmapPrecisionBits = 4 /* FIXME */, + .subTexelPrecisionBits = 8, + .mipmapPrecisionBits = 8, .maxDrawIndexedIndexValue = UINT32_MAX, .maxDrawIndirectCount = UINT32_MAX, .maxSamplerLodBias = 16, @@ -1145,7 +1360,10 @@ void anv_GetPhysicalDeviceProperties( .viewportBoundsRange = { INT16_MIN, INT16_MAX }, .viewportSubPixelBits = 13, /* We take a float? */ .minMemoryMapAlignment = 4096, /* A page */ - .minTexelBufferOffsetAlignment = 1, + /* The dataport requires texel alignment so we need to assume a worst + * case of R32G32B32A32 which is 16 bytes. + */ + .minTexelBufferOffsetAlignment = 16, /* We need 16 for UBO block reads to work and 32 for push UBOs */ .minUniformBufferOffsetAlignment = 32, .minStorageBufferOffsetAlignment = 4, @@ -1170,17 +1388,21 @@ void anv_GetPhysicalDeviceProperties( .sampledImageStencilSampleCounts = sample_counts, .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT, .maxSampleMaskWords = 1, - .timestampComputeAndGraphics = false, + .timestampComputeAndGraphics = true, .timestampPeriod = 1000000000.0 / devinfo->timestamp_frequency, .maxClipDistances = 8, .maxCullDistances = 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, @@ -1236,8 +1458,51 @@ void anv_GetPhysicalDeviceProperties2( VK_RESOLVE_MODE_MAX_BIT_KHR; } - props->independentResolveNone = VK_TRUE; - props->independentResolve = VK_TRUE; + props->independentResolveNone = true; + props->independentResolve = true; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES_EXT: { + VkPhysicalDeviceDescriptorIndexingPropertiesEXT *props = + (VkPhysicalDeviceDescriptorIndexingPropertiesEXT *)ext; + + /* It's a bit hard to exactly map our implementation to the limits + * described here. The bindless surface handle in the extended + * message descriptors is 20 bits and it's an index into the table of + * RENDER_SURFACE_STATE structs that starts at bindless surface base + * address. Given that most things consume two surface states per + * view (general/sampled for textures and write-only/read-write for + * images), we claim 2^19 things. + * + * For SSBOs, we just use A64 messages so there is no real limit + * there beyond the limit on the total size of a descriptor set. + */ + const unsigned max_bindless_views = 1 << 19; + + props->maxUpdateAfterBindDescriptorsInAllPools = max_bindless_views; + props->shaderUniformBufferArrayNonUniformIndexingNative = false; + props->shaderSampledImageArrayNonUniformIndexingNative = false; + props->shaderStorageBufferArrayNonUniformIndexingNative = true; + props->shaderStorageImageArrayNonUniformIndexingNative = false; + props->shaderInputAttachmentArrayNonUniformIndexingNative = false; + props->robustBufferAccessUpdateAfterBind = true; + props->quadDivergentImplicitLod = false; + props->maxPerStageDescriptorUpdateAfterBindSamplers = max_bindless_views; + props->maxPerStageDescriptorUpdateAfterBindUniformBuffers = MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS; + props->maxPerStageDescriptorUpdateAfterBindStorageBuffers = UINT32_MAX; + props->maxPerStageDescriptorUpdateAfterBindSampledImages = max_bindless_views; + props->maxPerStageDescriptorUpdateAfterBindStorageImages = max_bindless_views; + props->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_PER_STAGE_DESCRIPTOR_INPUT_ATTACHMENTS; + props->maxPerStageUpdateAfterBindResources = UINT32_MAX; + props->maxDescriptorSetUpdateAfterBindSamplers = max_bindless_views; + props->maxDescriptorSetUpdateAfterBindUniformBuffers = 6 * MAX_PER_STAGE_DESCRIPTOR_UNIFORM_BUFFERS; + props->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2; + props->maxDescriptorSetUpdateAfterBindStorageBuffers = UINT32_MAX; + props->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_BUFFERS / 2; + props->maxDescriptorSetUpdateAfterBindSampledImages = max_bindless_views; + props->maxDescriptorSetUpdateAfterBindStorageImages = max_bindless_views; + props->maxDescriptorSetUpdateAfterBindInputAttachments = MAX_DESCRIPTOR_SET_INPUT_ATTACHMENTS; break; } @@ -1246,11 +1511,11 @@ void anv_GetPhysicalDeviceProperties2( (VkPhysicalDeviceDriverPropertiesKHR *) ext; driver_props->driverID = VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA_KHR; - util_snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR, - "Intel open-source Mesa driver"); + snprintf(driver_props->driverName, VK_MAX_DRIVER_NAME_SIZE_KHR, + "Intel open-source Mesa driver"); - util_snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR, - "Mesa " PACKAGE_VERSION MESA_GIT_SHA1); + snprintf(driver_props->driverInfo, VK_MAX_DRIVER_INFO_SIZE_KHR, + "Mesa " PACKAGE_VERSION MESA_GIT_SHA1); driver_props->conformanceVersion = (VkConformanceVersionKHR) { .major = 1, @@ -1294,6 +1559,25 @@ void anv_GetPhysicalDeviceProperties2( 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; @@ -1326,11 +1610,20 @@ void anv_GetPhysicalDeviceProperties2( case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: { VkPhysicalDevicePointClippingProperties *properties = (VkPhysicalDevicePointClippingProperties *) ext; - properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES; - anv_finishme("Implement pop-free point clipping"); + properties->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_USER_CLIP_PLANES_ONLY; 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; @@ -1368,13 +1661,61 @@ void anv_GetPhysicalDeviceProperties2( properties->supportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT | VK_SUBGROUP_FEATURE_VOTE_BIT | - VK_SUBGROUP_FEATURE_ARITHMETIC_BIT | VK_SUBGROUP_FEATURE_BALLOT_BIT | VK_SUBGROUP_FEATURE_SHUFFLE_BIT | VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT | - VK_SUBGROUP_FEATURE_CLUSTERED_BIT | VK_SUBGROUP_FEATURE_QUAD_BIT; - properties->quadOperationsInAllStages = VK_TRUE; + if (pdevice->info.gen >= 8) { + /* TODO: There's no technical reason why these can't be made to + * work on gen7 but they don't at the moment so it's best to leave + * the feature disabled than enabled and broken. + */ + properties->supportedOperations |= + VK_SUBGROUP_FEATURE_ARITHMETIC_BIT | + VK_SUBGROUP_FEATURE_CLUSTERED_BIT; + } + properties->quadOperationsInAllStages = pdevice->info.gen >= 8; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT: { + VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *props = + (VkPhysicalDeviceSubgroupSizeControlPropertiesEXT *)ext; + STATIC_ASSERT(8 <= BRW_SUBGROUP_SIZE && BRW_SUBGROUP_SIZE <= 32); + props->minSubgroupSize = 8; + props->maxSubgroupSize = 32; + props->maxComputeWorkgroupSubgroups = pdevice->info.max_cs_threads; + props->requiredSubgroupSizeStages = VK_SHADER_STAGE_COMPUTE_BIT; + break; + } + + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_PROPERTIES_EXT: { + VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *props = + (VkPhysicalDeviceTexelBufferAlignmentPropertiesEXT *)ext; + + /* From the SKL PRM Vol. 2d, docs for RENDER_SURFACE_STATE::Surface + * Base Address: + * + * "For SURFTYPE_BUFFER non-rendertarget surfaces, this field + * specifies the base address of the first element of the surface, + * computed in software by adding the surface base address to the + * byte offset of the element in the buffer. The base address must + * be aligned to element size." + * + * The typed dataport messages require that things be texel aligned. + * Otherwise, we may just load/store the wrong data or, in the worst + * case, there may be hangs. + */ + props->storageTexelBufferOffsetAlignmentBytes = 16; + props->storageTexelBufferOffsetSingleTexelAlignment = true; + + /* The sampler, however, is much more forgiving and it can handle + * arbitrary byte alignment for linear and buffer surfaces. It's + * hard to find a good PRM citation for this but years of empirical + * experience demonstrate that this is true. + */ + props->uniformTexelBufferOffsetAlignmentBytes = 1; + props->uniformTexelBufferOffsetSingleTexelAlignment = false; break; } @@ -1388,10 +1729,10 @@ void anv_GetPhysicalDeviceProperties2( props->maxTransformFeedbackStreamDataSize = 128 * 4; props->maxTransformFeedbackBufferDataSize = 128 * 4; props->maxTransformFeedbackBufferDataStride = 2048; - props->transformFeedbackQueries = VK_TRUE; - props->transformFeedbackStreamsLinesTriangles = VK_FALSE; - props->transformFeedbackRasterizationStreamSelect = VK_FALSE; - props->transformFeedbackDraw = VK_TRUE; + props->transformFeedbackQueries = true; + props->transformFeedbackStreamsLinesTriangles = false; + props->transformFeedbackRasterizationStreamSelect = false; + props->transformFeedbackDraw = true; break; } @@ -1473,6 +1814,58 @@ void anv_GetPhysicalDeviceMemoryProperties( } } +static void +anv_get_memory_budget(VkPhysicalDevice physicalDevice, + VkPhysicalDeviceMemoryBudgetPropertiesEXT *memoryBudget) +{ + ANV_FROM_HANDLE(anv_physical_device, device, physicalDevice); + uint64_t sys_available = get_available_system_memory(); + assert(sys_available > 0); + + VkDeviceSize total_heaps_size = 0; + for (size_t i = 0; i < device->memory.heap_count; i++) + total_heaps_size += device->memory.heaps[i].size; + + for (size_t i = 0; i < device->memory.heap_count; i++) { + VkDeviceSize heap_size = device->memory.heaps[i].size; + VkDeviceSize heap_used = device->memory.heaps[i].used; + VkDeviceSize heap_budget; + + double heap_proportion = (double) heap_size / total_heaps_size; + VkDeviceSize sys_available_prop = sys_available * heap_proportion; + + /* + * Let's not incite the app to starve the system: report at most 90% of + * available system memory. + */ + uint64_t heap_available = sys_available_prop * 9 / 10; + heap_budget = MIN2(heap_size, heap_used + heap_available); + + /* + * Round down to the nearest MB + */ + heap_budget &= ~((1ull << 20) - 1); + + /* + * The heapBudget value must be non-zero for array elements less than + * VkPhysicalDeviceMemoryProperties::memoryHeapCount. The heapBudget + * value must be less than or equal to VkMemoryHeap::size for each heap. + */ + assert(0 < heap_budget && heap_budget <= heap_size); + + memoryBudget->heapUsage[i] = heap_used; + memoryBudget->heapBudget[i] = heap_budget; + } + + /* The heapBudget and heapUsage values must be zero for array elements + * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount + */ + for (uint32_t i = device->memory.heap_count; i < VK_MAX_MEMORY_HEAPS; i++) { + memoryBudget->heapBudget[i] = 0; + memoryBudget->heapUsage[i] = 0; + } +} + void anv_GetPhysicalDeviceMemoryProperties2( VkPhysicalDevice physicalDevice, VkPhysicalDeviceMemoryProperties2* pMemoryProperties) @@ -1482,6 +1875,9 @@ void anv_GetPhysicalDeviceMemoryProperties2( vk_foreach_struct(ext, pMemoryProperties->pNext) { switch (ext->sType) { + case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT: + anv_get_memory_budget(physicalDevice, (void*)ext); + break; default: anv_debug_ignored_stype(ext->sType); break; @@ -1635,6 +2031,25 @@ anv_state_pool_emit_data(struct anv_state_pool *pool, size_t size, size_t align, return state; } +/* Haswell border color is a bit of a disaster. Float and unorm formats use a + * straightforward 32-bit float color in the first 64 bytes. Instead of using + * a nice float/integer union like Gen8+, Haswell specifies the integer border + * color as a separate entry /after/ the float color. The layout of this entry + * also depends on the format's bpp (with extra hacks for RG32), and overlaps. + * + * Since we don't know the format/bpp, we can't make any of the border colors + * containing '1' work for all formats, as it would be in the wrong place for + * some of them. We opt to make 32-bit integers work as this seems like the + * most common option. Fortunately, transparent black works regardless, as + * all zeroes is the same in every bit-size. + */ +struct hsw_border_color { + float float32[4]; + uint32_t _pad0[12]; + uint32_t uint32[4]; + uint32_t _pad1[108]; +}; + struct gen8_border_color { union { float float32[4]; @@ -1647,18 +2062,33 @@ struct gen8_border_color { static void anv_device_init_border_colors(struct anv_device *device) { - static const struct gen8_border_color border_colors[] = { - [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } }, - [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } }, - [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } }, - [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } }, - [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } }, - [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } }, - }; + if (device->info.is_haswell) { + static const struct hsw_border_color border_colors[] = { + [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } }, + [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } }, + [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } }, + [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } }, + [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } }, + [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } }, + }; - device->border_colors = anv_state_pool_emit_data(&device->dynamic_state_pool, - sizeof(border_colors), 64, - border_colors); + device->border_colors = + anv_state_pool_emit_data(&device->dynamic_state_pool, + sizeof(border_colors), 512, border_colors); + } else { + static const struct gen8_border_color border_colors[] = { + [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 0.0 } }, + [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = { .float32 = { 0.0, 0.0, 0.0, 1.0 } }, + [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = { .float32 = { 1.0, 1.0, 1.0, 1.0 } }, + [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = { .uint32 = { 0, 0, 0, 0 } }, + [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = { .uint32 = { 0, 0, 0, 1 } }, + [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = { .uint32 = { 1, 1, 1, 1 } }, + }; + + device->border_colors = + anv_state_pool_emit_data(&device->dynamic_state_pool, + sizeof(border_colors), 64, border_colors); + } } static void @@ -1717,6 +2147,9 @@ anv_device_init_dispatch(struct anv_device *device) { 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; @@ -1922,16 +2355,18 @@ VkResult anv_CreateDevice( if (!device) return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - const unsigned decode_flags = - GEN_BATCH_DECODE_FULL | - ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) | - GEN_BATCH_DECODE_OFFSETS | - GEN_BATCH_DECODE_FLOATS; + if (INTEL_DEBUG & DEBUG_BATCH) { + const unsigned decode_flags = + GEN_BATCH_DECODE_FULL | + ((INTEL_DEBUG & DEBUG_COLOR) ? GEN_BATCH_DECODE_IN_COLOR : 0) | + GEN_BATCH_DECODE_OFFSETS | + GEN_BATCH_DECODE_FLOATS; - gen_batch_decode_ctx_init(&device->decoder_ctx, - &physical_device->info, - stderr, decode_flags, NULL, - decode_get_bo, NULL, device); + gen_batch_decode_ctx_init(&device->decoder_ctx, + &physical_device->info, + stderr, decode_flags, NULL, + decode_get_bo, NULL, device); + } device->_loader_data.loaderMagic = ICD_LOADER_MAGIC; device->instance = physical_device->instance; @@ -1964,20 +2399,20 @@ VkResult anv_CreateDevice( } /* keep the page with address zero out of the allocator */ - util_vma_heap_init(&device->vma_lo, LOW_HEAP_MIN_ADDRESS, LOW_HEAP_SIZE); - device->vma_lo_available = - physical_device->memory.heaps[physical_device->memory.heap_count - 1].size; - - /* Leave the last 4GiB out of the high vma range, so that no state base - * address + size can overflow 48 bits. For more information see the - * comment about Wa32bitGeneralStateOffset in anv_allocator.c - */ - util_vma_heap_init(&device->vma_hi, HIGH_HEAP_MIN_ADDRESS, - HIGH_HEAP_SIZE); + struct anv_memory_heap *low_heap = + &physical_device->memory.heaps[physical_device->memory.heap_count - 1]; + util_vma_heap_init(&device->vma_lo, low_heap->vma_start, low_heap->vma_size); + device->vma_lo_available = low_heap->size; + + struct anv_memory_heap *high_heap = + &physical_device->memory.heaps[0]; + util_vma_heap_init(&device->vma_hi, high_heap->vma_start, high_heap->vma_size); device->vma_hi_available = physical_device->memory.heap_count == 1 ? 0 : - physical_device->memory.heaps[0].size; + high_heap->size; } + list_inithead(&device->memory_objects); + /* As per spec, the driver implementation may deny requests to acquire * a priority above the default priority (MEDIUM) if the caller does not * have sufficient privileges. In this scenario VK_ERROR_NOT_PERMITTED_EXT @@ -2024,7 +2459,7 @@ VkResult anv_CreateDevice( result = vk_error(VK_ERROR_INITIALIZATION_FAILED); goto fail_mutex; } - if (pthread_cond_init(&device->queue_submit, NULL) != 0) { + if (pthread_cond_init(&device->queue_submit, &condattr) != 0) { pthread_condattr_destroy(&condattr); result = vk_error(VK_ERROR_INITIALIZATION_FAILED); goto fail_mutex; @@ -2076,7 +2511,7 @@ VkResult anv_CreateDevice( goto fail_surface_state_pool; } - result = anv_bo_init_new(&device->workaround_bo, device, 1024); + result = anv_bo_init_new(&device->workaround_bo, device, 4096); if (result != VK_SUCCESS) goto fail_binding_table_pool; @@ -2091,9 +2526,6 @@ VkResult anv_CreateDevice( if (device->info.gen >= 10) anv_device_init_hiz_clear_value_bo(device); - if (physical_device->use_softpin) - device->pinned_buffers = _mesa_pointer_set_create(NULL); - anv_scratch_pool_init(device, &device->scratch_pool); anv_queue_init(device, &device->queue); @@ -2117,6 +2549,9 @@ VkResult anv_CreateDevice( 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. */ @@ -2184,14 +2619,12 @@ void anv_DestroyDevice( anv_queue_finish(&device->queue); - if (physical_device->use_softpin) - _mesa_set_destroy(device->pinned_buffers, NULL); - #ifdef HAVE_VALGRIND /* We only need to free these to prevent valgrind errors. The backing * 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); @@ -2220,7 +2653,8 @@ void anv_DestroyDevice( anv_gem_destroy_context(device, device->context_id); - gen_batch_decode_ctx_finish(&device->decoder_ctx); + if (INTEL_DEBUG & DEBUG_BATCH) + gen_batch_decode_ctx_finish(&device->decoder_ctx); close(device->fd); @@ -2442,8 +2876,11 @@ anv_vma_free(struct anv_device *device, struct anv_bo *bo) util_vma_heap_free(&device->vma_lo, addr_48b, bo->size); device->vma_lo_available += bo->size; } else { - assert(addr_48b >= HIGH_HEAP_MIN_ADDRESS && - addr_48b <= HIGH_HEAP_MAX_ADDRESS); + ASSERTED const struct anv_physical_device *physical_device = + &device->instance->physicalDevice; + assert(addr_48b >= physical_device->memory.heaps[0].vma_start && + addr_48b < (physical_device->memory.heaps[0].vma_start + + physical_device->memory.heaps[0].vma_size)); util_vma_heap_free(&device->vma_hi, addr_48b, bo->size); device->vma_hi_available += bo->size; } @@ -2668,8 +3105,15 @@ VkResult anv_AllocateMemory( } success: + pthread_mutex_lock(&device->mutex); + list_addtail(&mem->link, &device->memory_objects); + pthread_mutex_unlock(&device->mutex); + *pMem = anv_device_memory_to_handle(mem); + p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used, + mem->bo->size); + return VK_SUCCESS; fail: @@ -2755,16 +3199,24 @@ void anv_FreeMemory( { ANV_FROM_HANDLE(anv_device, device, _device); ANV_FROM_HANDLE(anv_device_memory, mem, _mem); + struct anv_physical_device *pdevice = &device->instance->physicalDevice; if (mem == NULL) return; + pthread_mutex_lock(&device->mutex); + list_del(&mem->link); + pthread_mutex_unlock(&device->mutex); + if (mem->map) anv_UnmapMemory(_device, _mem); + p_atomic_add(&pdevice->memory.heaps[mem->type->heapIndex].used, + -mem->bo->size); + anv_bo_cache_release(device, &device->bo_cache, mem->bo); -#ifdef ANDROID +#if defined(ANDROID) && ANDROID_API_LEVEL >= 26 if (mem->ahw) AHardwareBuffer_release(mem->ahw); #endif @@ -2963,8 +3415,8 @@ void anv_GetBufferMemoryRequirements2( switch (ext->sType) { case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: { VkMemoryDedicatedRequirements *requirements = (void *)ext; - requirements->prefersDedicatedAllocation = VK_FALSE; - requirements->requiresDedicatedAllocation = VK_FALSE; + requirements->prefersDedicatedAllocation = false; + requirements->requiresDedicatedAllocation = false; break; } @@ -3069,11 +3521,11 @@ void anv_GetImageMemoryRequirements2( * * See also anv_AllocateMemory. */ - requirements->prefersDedicatedAllocation = VK_TRUE; - requirements->requiresDedicatedAllocation = VK_TRUE; + requirements->prefersDedicatedAllocation = true; + requirements->requiresDedicatedAllocation = true; } else { - requirements->prefersDedicatedAllocation = VK_FALSE; - requirements->requiresDedicatedAllocation = VK_FALSE; + requirements->prefersDedicatedAllocation = false; + requirements->requiresDedicatedAllocation = false; } break; } @@ -3294,12 +3746,6 @@ VkResult anv_CreateBuffer( buffer->usage = pCreateInfo->usage; buffer->address = ANV_NULL_ADDRESS; - if (buffer->usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT) { - pthread_mutex_lock(&device->mutex); - _mesa_set_add(device->pinned_buffers, buffer); - pthread_mutex_unlock(&device->mutex); - } - *pBuffer = anv_buffer_to_handle(buffer); return VK_SUCCESS; @@ -3316,12 +3762,6 @@ void anv_DestroyBuffer( if (!buffer) return; - if (buffer->usage & VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT_EXT) { - pthread_mutex_lock(&device->mutex); - _mesa_set_remove_key(device->pinned_buffers, buffer); - pthread_mutex_unlock(&device->mutex); - } - vk_free2(&device->alloc, pAllocator, buffer); } @@ -3362,6 +3802,11 @@ void anv_DestroySampler( if (!sampler) return; + if (sampler->bindless_state.map) { + anv_state_pool_free(&device->dynamic_state_pool, + sampler->bindless_state); + } + vk_free2(&device->alloc, pAllocator, sampler); } @@ -3376,17 +3821,33 @@ VkResult anv_CreateFramebuffer( assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO); - size_t size = sizeof(*framebuffer) + - sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount; - framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8, - VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); - if (framebuffer == NULL) - return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + size_t size = sizeof(*framebuffer); + + /* VK_KHR_imageless_framebuffer extension says: + * + * If flags includes VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR, + * parameter pAttachments is ignored. + */ + if (!(pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR)) { + size += sizeof(struct anv_image_view *) * pCreateInfo->attachmentCount; + framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + if (framebuffer == NULL) + return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); + + for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) { + ANV_FROM_HANDLE(anv_image_view, iview, pCreateInfo->pAttachments[i]); + framebuffer->attachments[i] = iview; + } + framebuffer->attachment_count = pCreateInfo->attachmentCount; + } else { + assert(device->enabled_extensions.KHR_imageless_framebuffer); + framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8, + VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); + if (framebuffer == NULL) + return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); - framebuffer->attachment_count = pCreateInfo->attachmentCount; - for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) { - VkImageView _iview = pCreateInfo->pAttachments[i]; - framebuffer->attachments[i] = anv_image_view_from_handle(_iview); + framebuffer->attachment_count = 0; } framebuffer->width = pCreateInfo->width;