vkd3d: enable EXT_device_generated_commands

this adds new paths to use the EXT version of the functionality
when available instead of NV, enabling broader support across
drivers/hardware while providing the same functionality

Signed-off-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Co-authored-by: Hans-Kristian Arntzen <post@arntzen-software.no>

libs/vkd3d/device.c

@@ -90,6 +90,7 @@ static const struct vkd3d_optional_extension_info optional_device_extensions[] =
     VK_EXTENSION(EXT_CONSERVATIVE_RASTERIZATION, EXT_conservative_rasterization),
     VK_EXTENSION(EXT_CUSTOM_BORDER_COLOR, EXT_custom_border_color),
     VK_EXTENSION(EXT_DEPTH_CLIP_ENABLE, EXT_depth_clip_enable),
+    VK_EXTENSION(EXT_DEVICE_GENERATED_COMMANDS, EXT_device_generated_commands),
     VK_EXTENSION(EXT_IMAGE_VIEW_MIN_LOD, EXT_image_view_min_lod),
     VK_EXTENSION(EXT_ROBUSTNESS_2, EXT_robustness2),
     VK_EXTENSION(EXT_SHADER_STENCIL_EXPORT, EXT_shader_stencil_export),
@@ -1628,6 +1629,16 @@ static void vkd3d_physical_device_info_init(struct vkd3d_physical_device_info *i
         vk_prepend_struct(&info->features2, &info->depth_clip_features);
     }
 
+    if (vulkan_info->EXT_device_generated_commands)
+    {
+        info->device_generated_commands_features_ext.sType =
+                VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEVICE_GENERATED_COMMANDS_FEATURES_EXT;
+        info->device_generated_commands_properties_ext.sType =
+                VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEVICE_GENERATED_COMMANDS_PROPERTIES_EXT;
+        vk_prepend_struct(&info->features2, &info->device_generated_commands_features_ext);
+        vk_prepend_struct(&info->properties2, &info->device_generated_commands_properties_ext);
+    }
+
     if (vulkan_info->EXT_robustness2)
     {
         info->robustness2_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT;
@@ -3181,6 +3192,17 @@ static HRESULT d3d12_device_create_scratch_buffer(struct d3d12_device *device, e
         alloc_info.heap_flags = D3D12_HEAP_FLAG_ALLOW_ONLY_BUFFERS | D3D12_HEAP_FLAG_CREATE_NOT_ZEROED;
         alloc_info.flags = VKD3D_ALLOCATION_FLAG_GLOBAL_BUFFER | VKD3D_ALLOCATION_FLAG_INTERNAL_SCRATCH;
         alloc_info.vk_memory_priority = vkd3d_convert_to_vk_prio(D3D12_RESIDENCY_PRIORITY_NORMAL);
+        if (device->device_info.device_generated_commands_features_ext.deviceGeneratedCommands)
+        {
+            /* this flag cannot be used with the existing buffer heaps */
+            alloc_info.explicit_global_buffer_usage =
+                    VK_BUFFER_USAGE_2_PREPROCESS_BUFFER_BIT_EXT |
+                    VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT;
+
+            /* Need this to override any type flags provided by heap properties.
+             * BUFFER_USAGE_2_PREPROCESS_BUFFER implies 32-bit only types. */
+            alloc_info.flags |= VKD3D_ALLOCATION_FLAG_DEDICATED;
+        }
 
         if (FAILED(hr = vkd3d_allocate_memory(device, &device->memory_allocator,
                 &alloc_info, &scratch->allocation)))
@@ -8795,7 +8817,8 @@ static void vkd3d_scratch_pool_init(struct d3d12_device *device)
     }
 
     if ((vkd3d_config_flags & VKD3D_CONFIG_FLAG_REQUIRES_COMPUTE_INDIRECT_TEMPLATES) &&
-            device->device_info.device_generated_commands_compute_features_nv.deviceGeneratedCompute &&
+            (device->device_info.device_generated_commands_compute_features_nv.deviceGeneratedCompute ||
+             device->device_info.device_generated_commands_properties_ext.supportedIndirectCommandsShaderStages & VK_SHADER_STAGE_COMPUTE_BIT) &&
             device->device_info.vulkan_1_2_properties.driverID == VK_DRIVER_ID_NVIDIA_PROPRIETARY)
     {
         /* DGCC preprocess buffers are gigantic on NV. Starfield requires 27 MB for 4096 dispatches ... */

libs/vkd3d/heap.c

@@ -280,6 +280,7 @@ static HRESULT d3d12_heap_init(struct d3d12_heap *heap, struct d3d12_device *dev
     alloc_info.heap_desc = heap->desc;
     alloc_info.host_ptr = host_address;
     alloc_info.extra_allocation_flags = 0;
+    alloc_info.explicit_global_buffer_usage = 0;
 
     if ((alloc_info.heap_desc.Flags & D3D12_HEAP_FLAG_DENY_BUFFERS) &&
             device->d3d12_caps.options.ResourceHeapTier >= D3D12_RESOURCE_HEAP_TIER_2)

libs/vkd3d/memory.c

@@ -1640,6 +1640,7 @@ static HRESULT vkd3d_memory_allocator_try_add_chunk(struct vkd3d_memory_allocato
     alloc_info.flags = VKD3D_ALLOCATION_FLAG_NO_FALLBACK;
     alloc_info.optional_memory_properties = optional_properties;
     alloc_info.vk_memory_priority = vkd3d_convert_to_vk_prio(D3D12_RESIDENCY_PRIORITY_NORMAL);
+    alloc_info.explicit_global_buffer_usage = 0;
 
     if (minimum_size < VKD3D_VA_BLOCK_SIZE)
         alloc_info.memory_requirements.size = VKD3D_MEMORY_CHUNK_SIZE;
@@ -1941,6 +1942,7 @@ HRESULT vkd3d_allocate_heap_memory(struct d3d12_device *device, struct vkd3d_mem
     alloc_info.heap_flags = info->heap_desc.Flags;
     alloc_info.host_ptr = info->host_ptr;
     alloc_info.vk_memory_priority = info->vk_memory_priority;
+    alloc_info.explicit_global_buffer_usage = info->explicit_global_buffer_usage;
 
     alloc_info.flags |= info->extra_allocation_flags;
 

libs/vkd3d/meta.c

@@ -1637,7 +1637,8 @@ static HRESULT vkd3d_predicate_ops_init(struct vkd3d_predicate_ops *meta_predica
         meta_predicate_ops->data_sizes[i] = spec_data[i].arg_count * sizeof(uint32_t);
     }
 
-    if (device->device_info.device_generated_commands_features_nv.deviceGeneratedCommands)
+    if (device->device_info.device_generated_commands_features_nv.deviceGeneratedCommands ||
+        device->device_info.device_generated_commands_features_ext.deviceGeneratedCommands)
     {
         uint32_t num_active_words;
         spec_info.mapEntryCount = 1;

libs/vkd3d/resource.c

@@ -3565,6 +3565,7 @@ HRESULT d3d12_resource_create_committed(struct d3d12_device *device, const D3D12
         memset(&allocate_info, 0, sizeof(allocate_info));
         allocate_info.memory_requirements = memory_requirements.memoryRequirements;
         allocate_info.heap_flags = heap_flags;
+        allocate_info.explicit_global_buffer_usage = 0;
 
         if (object->flags & VKD3D_RESOURCE_LINEAR_STAGING_COPY)
         {

libs/vkd3d/state.c

@@ -6649,6 +6649,13 @@ static bool vkd3d_bindless_supports_mutable_type(struct d3d12_device *device, ui
 static uint32_t vkd3d_bindless_state_get_bindless_flags(struct d3d12_device *device)
 {
     const struct vkd3d_physical_device_info *device_info = &device->device_info;
+    const VkShaderStageFlags minimum_dgcc_stages =
+            VK_SHADER_STAGE_VERTEX_BIT |
+            VK_SHADER_STAGE_FRAGMENT_BIT |
+            VK_SHADER_STAGE_COMPUTE_BIT;
+    const VkShaderStageFlags minimum_dgc_stages =
+            VK_SHADER_STAGE_VERTEX_BIT |
+            VK_SHADER_STAGE_FRAGMENT_BIT;
     uint32_t flags = 0;
 
     if (!d3d12_device_uses_descriptor_buffers(device))
@@ -6710,18 +6717,23 @@ static uint32_t vkd3d_bindless_state_get_bindless_flags(struct d3d12_device *dev
     }
 
     if ((vkd3d_config_flags & VKD3D_CONFIG_FLAG_REQUIRES_COMPUTE_INDIRECT_TEMPLATES) &&
-            !device->device_info.device_generated_commands_compute_features_nv.deviceGeneratedCompute)
+            !device->device_info.device_generated_commands_compute_features_nv.deviceGeneratedCompute &&
+            !(device->device_info.device_generated_commands_properties_ext.supportedIndirectCommandsShaderStages & VK_SHADER_STAGE_COMPUTE_BIT))
     {
         INFO("Forcing push UBO path for compute root parameters.\n");
         flags |= VKD3D_FORCE_COMPUTE_ROOT_PARAMETERS_PUSH_UBO;
     }
 
-    if (device->device_info.device_generated_commands_compute_features_nv.deviceGeneratedCompute &&
-            device->device_info.device_generated_commands_features_nv.deviceGeneratedCommands)
+    if ((device->device_info.device_generated_commands_compute_features_nv.deviceGeneratedCompute &&
+            device->device_info.device_generated_commands_features_nv.deviceGeneratedCommands) ||
+            (device->device_info.device_generated_commands_properties_ext.supportedIndirectCommandsShaderStages &
+                    minimum_dgcc_stages) == minimum_dgcc_stages)
     {
         INFO("Enabling fast paths for advanced ExecuteIndirect() graphics and compute.\n");
     }
-    else if (device->device_info.device_generated_commands_features_nv.deviceGeneratedCommands)
+    else if (device->device_info.device_generated_commands_features_nv.deviceGeneratedCommands ||
+            (device->device_info.device_generated_commands_properties_ext.supportedIndirectCommandsShaderStages &
+                    minimum_dgc_stages) == minimum_dgc_stages)
     {
         INFO("Enabling fast paths for advanced ExecuteIndirect() graphics.\n");
     }

libs/vkd3d/vkd3d_private.h

@@ -136,6 +136,7 @@ struct vkd3d_vulkan_info
     bool EXT_conservative_rasterization;
     bool EXT_custom_border_color;
     bool EXT_depth_clip_enable;
+    bool EXT_device_generated_commands;
     bool EXT_image_view_min_lod;
     bool EXT_robustness2;
     bool EXT_shader_stencil_export;
@@ -710,6 +711,7 @@ struct vkd3d_allocate_heap_memory_info
     void *host_ptr;
     uint32_t extra_allocation_flags;
     float vk_memory_priority;
+    VkBufferUsageFlags explicit_global_buffer_usage;
 };
 
 struct vkd3d_allocate_resource_memory_info
@@ -3309,8 +3311,10 @@ struct d3d12_command_signature
             VkBuffer buffer;
             VkDeviceAddress buffer_va;
             struct vkd3d_device_memory_allocation memory;
-            VkIndirectCommandsLayoutNV layout_implicit;
-            VkIndirectCommandsLayoutNV layout_preprocess;
+            VkIndirectCommandsLayoutNV layout_implicit_nv;
+            VkIndirectCommandsLayoutNV layout_preprocess_nv;
+            VkIndirectCommandsLayoutEXT layout_implicit_ext;
+            VkIndirectCommandsLayoutEXT layout_preprocess_ext;
             uint32_t stride;
             struct vkd3d_execute_indirect_info pipeline;
         } dgc;
@@ -4444,6 +4448,7 @@ struct vkd3d_physical_device_info
     VkPhysicalDeviceFragmentShadingRatePropertiesKHR fragment_shading_rate_properties;
     VkPhysicalDeviceConservativeRasterizationPropertiesEXT conservative_rasterization_properties;
     VkPhysicalDeviceDeviceGeneratedCommandsPropertiesNV device_generated_commands_properties_nv;
+    VkPhysicalDeviceDeviceGeneratedCommandsPropertiesEXT device_generated_commands_properties_ext;
     VkPhysicalDeviceMeshShaderPropertiesEXT mesh_shader_properties;
     VkPhysicalDeviceShaderModuleIdentifierPropertiesEXT shader_module_identifier_properties;
     VkPhysicalDeviceDescriptorBufferPropertiesEXT descriptor_buffer_properties;
@@ -4481,6 +4486,7 @@ struct vkd3d_physical_device_info
     VkPhysicalDeviceCoherentMemoryFeaturesAMD device_coherent_memory_features_amd;
     VkPhysicalDeviceRayTracingMaintenance1FeaturesKHR ray_tracing_maintenance1_features;
     VkPhysicalDeviceDeviceGeneratedCommandsFeaturesNV device_generated_commands_features_nv;
+    VkPhysicalDeviceDeviceGeneratedCommandsFeaturesEXT device_generated_commands_features_ext;
     VkPhysicalDeviceMeshShaderFeaturesEXT mesh_shader_features;
     VkPhysicalDeviceShaderModuleIdentifierFeaturesEXT shader_module_identifier_features;
     VkPhysicalDevicePresentIdFeaturesKHR present_id_features;

libs/vkd3d/vulkan_procs.h

@@ -338,6 +338,13 @@ VK_DEVICE_EXT_PFN(vkGetGeneratedCommandsMemoryRequirementsNV)
 VK_DEVICE_EXT_PFN(vkCmdExecuteGeneratedCommandsNV)
 VK_DEVICE_EXT_PFN(vkCmdPreprocessGeneratedCommandsNV)
 
+/* VK_EXT_device_generated_commands */
+VK_DEVICE_EXT_PFN(vkCreateIndirectCommandsLayoutEXT)
+VK_DEVICE_EXT_PFN(vkDestroyIndirectCommandsLayoutEXT)
+VK_DEVICE_EXT_PFN(vkGetGeneratedCommandsMemoryRequirementsEXT)
+VK_DEVICE_EXT_PFN(vkCmdExecuteGeneratedCommandsEXT)
+VK_DEVICE_EXT_PFN(vkCmdPreprocessGeneratedCommandsEXT)
+
 /* VK_EXT_shader_module_identifier */
 VK_DEVICE_EXT_PFN(vkGetShaderModuleIdentifierEXT)