AGC (Intel Graphics, Arrow Lake, Linux): four agc_driver tests fail: f32 denormal flush, f64 frexp mantissa, T# MIN_LOD clamp and mesh v_interp_mov #15817

What happens

Four agc_driver tests fail every run on the integrated GPU of a Core Ultra 9 285K (Arrow Lake-S, Mesa ANV, Linux). They pass on the RTX 5080 in the same machine, from the same build:

Title, test or build step

The four ctests above. No title.

AnyPS5 commit

main b11c7bf84f, nothing merged on top.

System

CachyOS, kernel 7.3.1-1-cachyos, GCC 16.2.1, Release. Intel(R) Graphics (ARL) (8086:7d67) of a Core Ultra 9 285K, Mesa 26.2.5-arch3.1 (ANV), Vulkan 1.4.354. The same machine has an NVIDIA GeForce RTX 5080 (driver 615.78.08).

Host report

No title was run. Device properties that bear on these tests, from vulkaninfo:

• shaderFloat64, minLod (VK_EXT_image_view_min_lod) and VK_EXT_mesh_shader are present; VK_KHR_fragment_shader_barycentric is not.
• Float controls: denormBehaviorIndependence ALL, roundingModeIndependence NONE. Denormal flush and preserve, RTE and RTZ, and signed zero/Inf/NaN preservation are supported for f32 and f64.
• Subgroup sizes 8 to 32 (32 by default).

Steps to reproduce

ANYPS5_GPU=Intel ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 \
  ctest --test-dir build -R '^(agc_driver_f32_denormal_flush|agc_driver_f64_rounding_denormal_modes|agc_driver_image_sample_lod_clamp|agc_driver_mesh_barycentric_emulation)$' -V

Each fails in under a second.

Log

[src/VulkanDevice.cpp:1471 VulkanDevice] Selected GPU name=Intel(R) Graphics (ARL) vendor=0x8086 device=0x7d67 subgroup=32
[float-controls] f32 denormal flush: hardware (0 of 5472 probe results differ from the emulated flush, 42 ms)
[float-controls] f32 denormal preserve: requested (0 of 256 probe results on normal values differ from the default mode by more than the Vulkan precision bounds, 3 ms)
AGC graphics: f32 denormal flush: fused lane 4 no float mode v_mul_f32 v12, v4, v5 is 0x00800000, expected 0x00000000

The other three print the same device lines before the errors in the table.

What was checked

• Each test failed 5 of 5 runs alone, with and without VK_LAYER_KHRONOS_validation (Vulkan-ValidationLayers 1.4.363). With the layer, none of them logs a VUID- line, and no run logs was not found but was requested.
• On the RTX 5080 in the same machine and build, all four pass 5 of 5, also with the layer.
• The whole GPU selection (^(agc|shader|recompiler|videoout|video_out|guest_video_out|vulkan)|gpu, 1,384 tests) on this GPU: 1,312 passed, 5 failed and 67 skipped. The fifth failure, agc_driver_scattered_runs, also fails on the RTX 5080 because this kernel's udmabuf size_limit_mb is 2147483647; it is unrelated to these four.
• Related tests that pass on this GPU: agc_driver_image_sample_lod_clamp_shader_lane (the same LOD clamp with APS5_NO_IMAGE_VIEW_MIN_LOD=1, so clamped in the shader), agc_driver_pixel_barycentric_emulation, agc_fragment_barycentric, agc_driver_mesh_provoking_vertex and agc_driver_mesh_provoking_layer, and 37 of the 39 tests whose names contain denorm, rounding or frexp.
• None of the four is listed as failing or skipped in the AMD GPU test run on RADV posted on fix(libc): retain ordinary short character scan fields #13119 (main 7feb37e, 3,693 of 3,743 tests run).
• Open fix(shader): preserve compute implicit LOD and quad inputs #10867 changes ImageSampleLodClamp.cpp. With it merged onto main 5b1d381d5a, agc_driver_image_sample_lod_clamp still fails 3 of 3 with the same message here. Open fix(shader): emit OpFmaKHR for f32 fma when VK_KHR_shader_fma is available #15343 emits OpFmaKHR only on RADV, by its description. I did not test open fix(shader): fall back to barycentrics for a pixel input read through both I/J pairs #15793, which changes the barycentric fallback for pixel inputs read through two I/J pairs.
• Closed reports and fixes nearby: AGC: agc_driver_mesh_barycentric_emulation fails since #9193 on devices with mesh shaders and VK_KHR_fragment_shader_barycentric #12121 and test(agc): disable barycentric built-ins in the mesh barycentric emulation test #12062 (the mesh barycentric emulation on devices with VK_KHR_fragment_shader_barycentric, which this device lacks), feat(agc): apply a T#'s MIN_LOD clamp in the shader without VK_EXT_image_view_min_lod #12662 (the MIN_LOD clamp in the shader without VK_EXT_image_view_min_lod, which this device has), fix(shader): flush f64 floor, ceil, fract and frexp denormal inputs #7036 (f64 frexp of denormal inputs) and test(shader): accept the split fma sum in F32DenormalFlush #3822 (the split fma sum in the f32 flush test).
• Not checked on another Intel GPU or Mesa version.

AI-assisted: yes (Claude Code)