vulkan: add vk_x11_strict_image_count option
[mesa.git] / src / intel / vulkan / anv_device.c
index bfba11fe8c92f9c8dab471c6305f5ee63ba32c2b..50d910780cbac3dda369b9fad1fc07fbd9699068 100644 (file)
@@ -40,6 +40,7 @@
 #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"
 
 #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.
  */
@@ -318,7 +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);
 
@@ -390,19 +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;
@@ -410,10 +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");
@@ -423,6 +424,8 @@ anv_physical_device_init(struct anv_physical_device *device,
       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,
@@ -776,6 +779,12 @@ VkResult anv_CreateInstance(
 
    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;
@@ -806,6 +815,9 @@ void anv_DestroyInstance(
    glsl_type_singleton_decref();
    _mesa_locale_fini();
 
+   driDestroyOptionCache(&instance->dri_options);
+   driDestroyOptionInfo(&instance->available_dri_options);
+
    vk_free(&instance->alloc, instance);
 }
 
@@ -1091,6 +1103,13 @@ void anv_GetPhysicalDeviceFeatures2(
          break;
       }
 
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
+         VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
+            (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
+         features->indexTypeUint8 = true;
+         break;
+      }
+
       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INLINE_UNIFORM_BLOCK_FEATURES_EXT: {
          VkPhysicalDeviceInlineUniformBlockFeaturesEXT *features =
             (VkPhysicalDeviceInlineUniformBlockFeaturesEXT *)ext;
@@ -1099,6 +1118,18 @@ void anv_GetPhysicalDeviceFeatures2(
          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;
@@ -1108,6 +1139,20 @@ 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 = false;
@@ -1148,6 +1193,14 @@ void anv_GetPhysicalDeviceFeatures2(
          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;
@@ -1229,6 +1282,8 @@ void anv_GetPhysicalDeviceProperties(
       pdevice->has_a64_buffer_access ? UINT32_MAX :
                                        MAX_BINDING_TABLE_SIZE - MAX_RTS;
 
+   const uint32_t max_workgroup_size = 32 * devinfo->max_cs_threads;
+
    VkSampleCountFlags sample_counts =
       isl_device_get_sample_counts(&pdevice->isl_dev);
 
@@ -1287,11 +1342,11 @@ void anv_GetPhysicalDeviceProperties(
       .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,
@@ -1333,17 +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,
@@ -1452,11 +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,
@@ -1500,6 +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;
@@ -1536,6 +1614,16 @@ void anv_GetPhysicalDeviceProperties2(
          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;
@@ -1573,13 +1661,31 @@ 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 = 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;
       }
 
@@ -2041,6 +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;
@@ -2402,7 +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;
 
@@ -2440,6 +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. */
@@ -2512,6 +2624,7 @@ void anv_DestroyDevice(
     * 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);
@@ -2763,7 +2876,7 @@ 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 {
-      MAYBE_UNUSED const struct anv_physical_device *physical_device =
+      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 +
@@ -3708,17 +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;