dw_downsizer: Emit only the sub-transfers TKEEP marks valid - #8

The bug

axi_stream_dw_downsizer emits TotalSubTransfers sub-transfers for every input beat regardless of TKEEP, so the lanes a partial beat marks invalid are pushed downstream as real data.

Surfaced as pulp-platform/pulp-ethernet#5: a 52-byte Ethernet payload produced an 84-byte frame instead of 78. The excess of exactly six is the invalid lanes of the final 64-bit beat (8 − 2 valid bytes). pulp-ethernet instantiates this module to go 64→8 for the MAC (eth_top.sv:123) and does not connect TKEEP/TSTRB onward (framing_top.sv:214-218), so the null lanes reach the wire.

Reproduced in Verilator against the unmodified module — a single beat with TKEEP=0x03, TLAST=1:

sub-beats emitted      : 8
of which keep=1        : 2
spurious bytes on wire : 6

The change

The last sub-transfer of a beat is now the index of its highest asserted TKEEP bit, latched when the beat is accepted.

It cannot be derived from tkeep_received_q — that register is shifted right on every sub-transfer, so the target index would shrink as the counter climbs and a full beat would terminate early. I made exactly that mistake first; a full 0xFF beat emitted 5 bytes instead of 8 and hung the testbench. Hence the latch, and hence the full-beat check below.

Two edge cases:

• A beat whose only valid lane is 0 has no DataOut phase — that byte is itself the last sub-transfer and must carry TLAST.
• A beat with TKEEP all-zero still yields one sub-transfer, so it is consumed rather than stalling the stream.

Scope: trailing invalid lanes are dropped. Non-contiguous null lanes in the middle of a beat are not compacted — that needs real compaction logic and a wider behavioural decision, so I left it out rather than half-doing it.

Also in this commit

tstrb_received_d = tstrb_received_d >> StrbWidthOut read the next-state variable where both neighbouring lines use _q. It happens to produce the right value today only because of the default assignment above it. Changed to _q for consistency.

Testing

New self-contained testbench, test/tb_axi_stream_dw_downsizer_tkeep.sv, registered in Bender.yml. It states exact byte counts rather than using the randomized drivers, so the checks read as the specification.

Four of the ten checks fail against the unfixed RTL, so they are load-bearing. The full-beat case passes both before and after — that is the one showing fully valid beats are unchanged.

Verified with Verilator 5.006. I note that CI runs license, commit and verible lint but no simulation, so these results are from a local run rather than from the pipeline — worth an independent check on your side. I did not modify the existing tb_axi_stream_dw_downsizer, which uses the randomized driver framework and cannot easily express a specific TKEEP pattern.

Unrelated observation

pulp-ethernet pins this package at 472751f550e3918215603e21734fe0ece3c66f79 in its Bender.lock, and that revision is not reachable in this repository — I fetched all refs. So that consumer currently pins a commit nobody can fetch, and this fix will not reach it until the pin moves. Flagging it here since it affects whether the fix actually lands for the reporter; happy to raise it separately.