hal/vulkan: With internal_error_panic, panic on Vulkan validation errors (gfx-rs#9879)

Author: andyleiserson

Cargo.lock

@@ -24,6 +24,7 @@ log.workspace = true
 pico-args.workspace = true
 tokio = { workspace = true, features = ["full"] }
 termcolor.workspace = true
+wgpu-hal = { workspace = true, features = ["internal_error_panic"] }
 
 [dev-dependencies]
 tempfile.workspace = true

cts_runner/Cargo.toml

@@ -186,10 +186,11 @@ portable-atomic = ["dep:portable-atomic", "dep:portable-atomic-util"]
 # Panic when running into a device lost error (for debugging purposes).
 # Only affects the d3d12 and vulkan backends.
 device_lost_panic = []
-# Panic when running into an internal error other than out-of-memory and device lost
-# (for debugging purposes).
-#
-# Only affects the d3d12 and vulkan backends.
+# Panic on unexpected errors (does _not_ include out-of-memory errors or device
+# loss) from platform APIs, for debugging purposes. Such errors likely indicate
+# bug(s) in wgpu, platform components, or drivers. Despite the name, this is not
+# related to the spec's `GPUInternalError`. Only affects the d3d12 and vulkan
+# backends.
 internal_error_panic = []
 # Tracks validation errors in a `VALIDATION_CANARY` static.
 validation_canary = ["dep:parking_lot"]

wgpu-hal/Cargo.toml

@@ -158,9 +158,80 @@ unsafe extern "system" fn debug_utils_messenger_callback(
         crate::VALIDATION_CANARY.add(message.to_string());
     }
 
+    #[cfg(all(debug_assertions, feature = "internal_error_panic"))]
+    if level == log::Level::Error
+        && message_type.contains(vk::DebugUtilsMessageTypeFlagsEXT::VALIDATION)
+        && !error_is_waived(cd.message_id_number)
+        && !cts_error_is_waived(cd.message_id_number)
+    {
+        use alloc::string::ToString as _;
+        panic!("{}", message.to_string());
+    }
+
     vk::FALSE
 }
 
+/// Validation errors known to fire, not just in the CTS.
+///
+/// These never panic.
+#[cfg(feature = "internal_error_panic")]
+fn error_is_waived(message_id_number: i32) -> bool {
+    const WAIVED_MESSAGE_IDS: &[i32] = &[
+        // SYNC-HAZARD-WRITE-AFTER-WRITE
+        // e.g. webgpu:api,operation,memory_sync,texture,readonly_depth_stencil:sampling_while_testing:*
+        // https://github.com/gfx-rs/wgpu/issues/5231
+        // https://github.com/gfx-rs/wgpu/issues/8705
+        0x5c0ec5d6_u32 as i32,
+    ];
+
+    WAIVED_MESSAGE_IDS.contains(&message_id_number)
+}
+
+/// Validation errors known to fire when running the CTS.
+///
+/// These waivers are keyed off the `WGPU_CTS_XTASK` environment variable, which
+/// is set in `xtask/src/cts.rs`.
+#[cfg(feature = "internal_error_panic")]
+fn cts_error_is_waived(message_id_number: i32) -> bool {
+    use std::sync::LazyLock;
+
+    static WGPU_CTS_XTASK: LazyLock<bool> =
+        LazyLock::new(|| std::env::var_os("WGPU_CTS_XTASK").is_some());
+
+    if !*WGPU_CTS_XTASK {
+        return false;
+    }
+
+    const WAIVED_MESSAGE_IDS: &[i32] = &[
+        // VUID-vkCmdPushConstants-offset-01795
+        // e.g. webgpu:api,operation,command_buffer,programmable,immediate:*
+        0x27bc88c6_u32 as i32,
+        // VUID-SampleMask-SampleMask-04359
+        // e.g. webgpu:api,validation,render_pipeline,inter_stage:max_variables_count,*
+        0x34d444b2_u32 as i32,
+        // VUID-vkCmdCopyImage-srcImage-01728
+        // e.g. webgpu:api,validation,encoding,cmds,copyTextureToTexture:*
+        0x6b654496_u32 as i32,
+        // VUID-StandaloneSpirv-OpImageQuerySizeLod-04659
+        // e.g. webgpu:shader,execution,expression,call,builtin,textureNumLayers:*
+        0x82396078_u32 as i32,
+        // VUID-RuntimeSpirv-Location-06272
+        // e.g. webgpu:api,validation,render_pipeline,inter_stage:max_variables_count,*
+        0xa3614f8b_u32 as i32,
+        // VUID-VkViewport-width-01770
+        // e.g. webgpu:api,validation,encoding,cmds,render,dynamic_state:*
+        0xa4164ba5_u32 as i32,
+        // VUID-VkImageViewCreateInfo-image-04441
+        // e.g. webgpu:api,validation,createView:texture_view_usage:*
+        0xb75da543_u32 as i32,
+        // VUID-VkBufferCreateInfo-None-09500
+        // e.g. webgpu:api,validation,buffer,create:usage,*
+        0xf6d454db_u32 as i32,
+    ];
+
+    WAIVED_MESSAGE_IDS.contains(&message_id_number)
+}
+
 impl super::DebugUtilsCreateInfo {
     fn to_vk_create_info(&self) -> vk::DebugUtilsMessengerCreateInfoEXT<'_> {
         let user_data_ptr: *const super::DebugUtilsMessengerUserData = &*self.callback_data;

wgpu-hal/src/vulkan/instance.rs

@@ -72,6 +72,10 @@ pub fn run_cts(
     let llvm_cov = args.contains("--llvm-cov");
     let release = args.contains("--release");
 
+    // This is used in the Vulkan hal to waive pre-existing validation
+    // errors in the CTS, until they can be fixed.
+    shell.set_var("WGPU_CTS_XTASK", "1");
+
     let output_filter = args
         .opt_value_from_str::<_, String>("--print-output-when")?
         .map(|f| {