[Bug][VMI Layout] group4 ui16 -> ui32 width-changing vinterpret_cast fails in packed UE8M0 exponent packing #1586

Summary

在 packed UE8M0 scale-factor packing 中，vcmax(group=4) 的结果是 grouped 4xui16，随后需要做 width-changing vinterpret_cast：

4xui16  ->  2xui32

目的是把两个 8-bit exponent 拼成一个 16-bit packed word，再 vcvt 回 4xui16 写入 SF UB。

当前 PTOAS 报 layout contract 失败：width-changing bitcast 需要 source 为 contiguous，但 source 从 group reduce 继承的是：

num_groups = 4, slots = 1

ensure_layout 没有 grouped -> contiguous 的 materialization 规则，因此编译失败。

Minimal PTODSL shape

mask128 = pto.vmi.create_mask(128, size=128)
mask4 = pto.vmi.create_mask(1, size=4, group=4)

scale4 = pto.vmi.vcmax(amax128, mask128, group=4)

# same-width reinterpret: group layout is preserved
bits = pto.vmi.vinterpret_cast(scale4, to_dtype=pto.ui32)

# group layout is carried through the elementwise/integer conversion
exp16 = pto.vmi.vcvt(
    pto.vmi.vshrs(bits, 23, mask4),
    to_dtype=pto.ui16,
    saturate="NOSAT",
)

# width-changing reinterpret: 4xui16 -> 2xui32
packed32 = pto.vmi.vshrs(
    pto.vmi.vinterpret_cast(exp16, to_dtype=pto.ui32),
    8,
    pto.vmi.create_mask(2, size=2),
)
packed16 = pto.vmi.vcvt(packed32, to_dtype=pto.ui16, saturate="NOSAT")

The failing operation is the second vinterpret_cast:

%packed32 = pto.vmi.vinterpret_cast %exp16
    : !pto.vmi.vreg<4xui16, #pto.vmi.layout<num_groups = 4, slots = 1>>
   -> !pto.vmi.vreg<2xui32>

Actual error

VMI-LAYOUT-CONTRACT: pto.vmi.bitcast operand #0 has type
!pto.vmi.vreg<4xui16, #pto.vmi.layout<num_groups = 4, slots = 1>>
but requires
!pto.vmi.vreg<4xui16, #pto.vmi.layout<contiguous>>;
pto.vmi.ensure_layout has no registered materialization support:
source/result layouts do not match a supported ensure_layout table row

Environment

• PTOAS 0.66 (v0.66-8-gb729c955c)
• TileKernels per_block_cast VMI path
• Reduce/bitcast shape: 128xf32 -> vcmax(group=4) -> 4xf32 -> ... -> 4xui16 -> 2xui32

Expected behavior

Allow a grouped 4xui16 value to feed a width-changing vinterpret_cast to 2xui32, either by:

1. adding a legal num_groups=4, slots=1 -> contiguous materialization path, or
2. recognizing this specific width-changing bitcast as layout-preserving where physically valid.

The former is the direct fix; the latter may be a useful optimization if the grouped physical layout already matches the required byte layout.

Workaround

Avoiding the width-changing vinterpret_cast and packing the two exponents entirely in the original 16-bit domain. This is not implemented in the kernel because it changes the VMI dataflow and may affect the desired ASC-aligned form.

Related issues

• [Bug][VMI Layout] ensure_layout 规则表缺行：gs(8) group-slot 结果无法过宽度变更 vcvt；128-lane f32 carrier 无法做 group reduce（附最小复现） #1472: group-slot result × width-changing conversion; same class of missing ensure_layout materialization.
• [VMI] Add contiguous-source lowering for f32 group_reduce_maxf(group=4) #1541: contiguous-source lowering for f32 group_reduce_maxf(group=4); would avoid producing the grouped source in the first place.
• [VMI] group_reduce_maxf group=4 needs consumer-aware lowering: packed vector vs direct point store #1559: consumer-aware lowering for group_reduce_maxf(group=4).