regression: compute example doesn't work on wayland

[sigmaSd]

context: denoland/deno#24857

This used to work with old versions of wgpu (the one thats used by deno 1.43.0 ) but I didn't search for the version yet

Problem:

Running the compute example fails with

No numbers were provided, defaulting to [1, 2, 3, 4]
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
thread 'main' panicked at /home/mrcool/.cargo/registry/src/index.crates.io-6f17d22bba15001f/wgpu-22.1.0/src/backend/wgpu_core.rs:3411:5:
wgpu error: Validation Error

Caused by:
 In Device::create_compute_pipeline
   Internal error: The selected version doesn't support Features(DYNAMIC_ARRAY_SIZE)


note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

the deafult value in fedora for XDG_RUNTIME_DIR is /run/user/1000

if I run with that variable unset or changed to another location it works

XDG_RUNTIME_DIR="/tmp" cargo r

No numbers were provided, defaulting to [1, 2, 3, 4]
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
Steps: [0, 1, 7, 2]

[sigmaSd]

platform
fedora 40 Linux fedora 6.9.11-200.fc40.x86_64 #1 SMP PREEMPT_DYNAMIC Thu Jul 25 18:17:34 UTC 2024 x86_64 GNU/Linux
Intel(R) Core(TM) i3-3110M CPU @ 2.40GHz

[sigmaSd]

also forcing the hello-compute example to run on xorg make it work correctly

WAYLAND_DISPLAY="" cargo r

[sigmaSd]

It seems like changing XDG_RUNTIME_DIR is equivalent to running on xorg from the point of view of applications , so the problem lies in the relation with wayland

Its weird that the rendering examples (like hello-triangle, and the others) where not affected by this, but the compute exampe (which didn't occur to me it would need to know about the compositor) is affeced

[sigmaSd]

0.18.0 works

0.19.0 gives an error

No numbers were provided, defaulting to [1, 2, 3, 4]
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
thread 'main' panicked at src/main.rs:57:10:
called `Result::unwrap()` on an `Err` value: RequestDeviceError { inner: Core(LimitsExceeded(FailedLimit { name: "max_compute_workgroups_per_dimension", requested: 65535, allowed: 0 })) }
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

0.20.0 gives a different error

No numbers were provided, defaulting to [1, 2, 3, 4]
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
thread 'main' panicked at /home/mrcool/.cargo/registry/src/index.crates.io-6f17d22bba15001f/wgpu-22.1.0/src/backend/wgpu_core.rs:3411:5:
wgpu error: Validation Error

Caused by:
 In Device::create_compute_pipeline
   Internal error: The selected version doesn't support Features(DYNAMIC_ARRAY_SIZE)


note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

[sigmaSd]

Maybe it helps, with a minimal wayland server it works correctly
server.c

#include <stdio.h>
#include <stdlib.h>
#include <wayland-server.h>

// Forward declarations
static void bind_compositor(struct wl_client *client, void *data, uint32_t version, uint32_t id);
static void bind_shell(struct wl_client *client, void *data, uint32_t version, uint32_t id);

int main(int argc, char *argv[])
{
    struct wl_display *display = wl_display_create();
    if (!display) {
        printf("Failed to create Wayland display\n");
        return 1;
    }

    const char *socket = wl_display_add_socket_auto(display);
    if (!socket) {
        printf("Failed to add Wayland socket\n");
        wl_display_destroy(display);
        return 1;
    }

    fprintf(stderr, "Running Wayland server on %s\n", socket);

    // Create a global for the compositor interface
    struct wl_global *compositor_global = wl_global_create(display, &wl_compositor_interface, 4, NULL, bind_compositor);
    if (!compositor_global) {
        printf("Failed to create compositor global\n");
        wl_display_destroy(display);
        return 1;
    }

    // Create a global for the shell interface
    struct wl_global *shell_global = wl_global_create(display, &wl_shell_interface, 1, NULL, bind_shell);
    if (!shell_global) {
        printf("Failed to create shell global\n");
        wl_display_destroy(display);
        return 1;
    }

    // Start the Wayland server
    wl_display_init_shm(display); // Initialize shared memory support

    wl_display_run(display);
    wl_display_destroy(display);
    return 0;
}

// Compositor bind function
static void bind_compositor(struct wl_client *client, void *data, uint32_t version, uint32_t id)
{
    struct wl_resource *resource = wl_resource_create(client, &wl_compositor_interface, version, id);
    if (!resource) {
        wl_client_post_no_memory(client);
        return;
    }
}

// Shell bind function
static void bind_shell(struct wl_client *client, void *data, uint32_t version, uint32_t id)
{
    struct wl_resource *resource = wl_resource_create(client, &wl_shell_interface, version, id);
    if (!resource) {
        wl_client_post_no_memory(client);
        return;
    }
}

the servers I tried are mutter and weston, both error the same way

[teoxoy]

This is probably a result of #4718.
It worked in 0.18 because that PR is part of 0.19+.

The validation errors you are getting are expected as the OpenGL backend/driver are more limited.

[teoxoy]

Does changing XDG_RUNTIME_DIR select a different OpenGL driver? That would explain why it works or why it doesn't.
It's not something we have control over though.

[sigmaSd]

Im using the vulkan backend,

Wayland clients looks for Wayland sockets in XDG_RUNTIME_DIR so when it's changed the apps fallback to xorg, this have no relation to opengl or backbends as far as I can tell

[teoxoy]

Do you have multiple Vulkan drivers installed? Could you include the AdapterInfo when testing on 0.18 and on 0.19+ also with the different XDG_RUNTIME_DIR settings?

[sigmaSd]

in version 0.19.0

wayland

[src/main.rs:44:5] &adapter.get_info() = AdapterInfo {
    name: "Mesa Intel(R) HD Graphics 4000 (IVB GT2)",
    vendor: 32902,
    device: 0,
    device_type: IntegratedGpu,
    driver: "",
    driver_info: "",
    backend: Gl,
}

xorg

No numbers were provided, defaulting to [1, 2, 3, 4]
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
[src/main.rs:44:5] &adapter.get_info() = AdapterInfo {
    name: "llvmpipe (LLVM 18.1.6, 256 bits)",
    vendor: 65541,
    device: 0,
    device_type: Cpu,
    driver: "llvmpipe",
    driver_info: "Mesa 24.1.5 (LLVM 18.1.6)",
    backend: Vulkan,
}
Steps: [0, 1, 7, 2]

so it seems like when I use wayland it indeed use opengl backend,

and xorg it uses vulkan backend but with the cpu implementation

so I guess I have a coupe of remarks/questions left

• I still have this message printed MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete when running on wayland that's why I thought its using the vulkan backend
• I wanted to try the vulkan backend in wayland but WGPU_BACKEND=vulkan doesn't work, that's probably on purpose because the cpu is not supported but is it possible like xorg to use llvmpipe ?

So indeed its because of the old cpu, though I think it would be great if the error message said your cpu is not supported

[sigmaSd]

also I want to try vulkan in wayland but this doesn't work (it still chooses gl)

WGPU_BACKEND=vulkan WGPU_ALLOW_UNDERLYING_NONCOMPLIANT_ADAPTER=1 $CARGO_TARGET_DIR/debug/a
No numbers were provided, defaulting to [1, 2, 3, 4]
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
[src/main.rs:44:5] &adapter.get_info() = AdapterInfo {
    name: "Mesa Intel(R) HD Graphics 4000 (IVB GT2)",
    vendor: 32902,
    device: 0,
    device_type: IntegratedGpu,
    driver: "",
    driver_info: "",
    backend: Gl,
}
thread 'main' panicked at src/main.rs:58:10:
called `Result::unwrap()` on an `Err` value: RequestDeviceError { inner: Core(LimitsExceeded(FailedLimit { name: "max_compute_workgroups_per_dimension", requested: 65535, allowed: 0 })) }
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

[teoxoy]

I think the "MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete" message is printed by the driver itself which happens before we filter out the driver.

So indeed its because of the old cpu, though I think it would be great if the error message said your cpu is not supported

I think we do show a warning, make sure to set RUST_LOG to warn or lower.

also I want to try vulkan in wayland but this doesn't work (it still chooses gl)

The env var won't work unless you hook it up.

You should set the InstanceFlags::ALLOW_UNDERLYING_NONCOMPLIANT_ADAPTER flag instead.