Software information

WRC Generations (Epic, also on Steam as 1953520), default settings. The game freezes at the "KT Engine" splash with a GPU hang on Intel. NVIDIA and AMD are not affected.

System information

• GPU: Intel Core Ultra 7 165H, Meteor Lake iGPU (8086:7dd5)
• Driver: Mesa 26.2.5 (ANV), Linux 7.2.10, hangs on both i915 and xe
• Wine version: GE-Proton10-34 (Proton-8.0 hangs too)
• VKD3D-Proton version: the one bundled with GE-Proton10-34

Cause

The hang is caused by compute shader 73dfbd1401ac5447, a numthreads(64,1,1) prefix sum dispatched as 1x1x1 groups.

Roughly:

for (chunk = 0; chunk < (count + 63) / 64; chunk++) {
    idx = chunk * 64 + lid;
    if (idx <= count) {
        GroupMemoryBarrierWithGroupSync();
        // load into groupshared
        GroupMemoryBarrierWithGroupSync();
        // scan, write out
    }
}

Both barriers are inside a non-uniform branch. On the last partial chunk, ANV runs the group as four SIMD16 threads. A thread with no lanes in range skips the barrier and the group never completes.

How I traced it:

• In the i915 error state, ACTHD sits right after the COMPUTE_WALKER for this dispatch.
• The kernel bytes match the INTEL_DEBUG=cs,hex output with src_hash 0x73dfbd1401ac5447.

Workaround

I patched the SPIR-V so that:

• every lane executes both barriers;
• out-of-range lanes store 0 to groupshared;
• only in-range lanes do the final stores.

With that patch loaded through VKD3D_SHADER_OVERRIDE, the game runs on both i915 and xe.

Patch, diff, and the scripts I used: https://github.com/4ni1ak/wrc-generations-intel-gpu-hang-fix

Maybe this could become a per-app shader workaround, or a generic fixup for barriers in divergent control flow? I can provide a GPU error state dump or logs if useful.