Exhaustive precision sweep for mul_mat_subgroup_matrix_VEC_F16xF32xF32 at (m, n, k) = (256, 256, 256). Reference output is spec.js_reference — a naive WebGPU matmul with f32 accumulator, fixed @workgroup_size(16, 16). Pass = bit-equal to the reference. Drift columns (max abs diff / mean abs diff / max relative diff) are reported for information only.
72 | valid (after valid_Args): 72idle