Three ways to make a draw cheaper, none kept

What changed

The layer's source ends the session where it started. Replay (the macOS side that turns recorded D3D12 commands into Metal calls) costs about 520 ns of CPU per draw in bgfx's drawstress example (about 61,000 draws a frame). I skipped each part of the draw path in turn and measured what went away (see the cost chart). No part dominates, and most of the time goes into Metal calls, which run under Rosetta (Apple's x86-to-ARM translator) just as the layer does.

I tried three changes (see the diagram):

1. Cached upload memory. A spike (a small standalone program) showed that an atomic operation right after a store to write-combined memory (memory the CPU writes around its caches) costs 81 to 92 ns, against 7 ns otherwise. Replay writes each draw's root arguments to such a buffer, and atomics follow. Making the buffer cached changed nothing measurable, so I didn't keep it.
2. One argument-table bind per encoder. The Metal 4 header says the table's contents are snapshotted at each draw, so binding it once should be enough. In practice ModelViewer's GPU time fell from 12.3 to 1.8 ms: most draws had silently vanished. The reason is still unknown.
3. Zero less of each draw record. Each recorded draw is a 1848-byte struct, cleared in full before it is filled in. Clearing only the fields read without a count looked like a 6% cut in recording time. Back-to-back A/B runs said otherwise (see the A/B chart).

The interesting part

The first measurement of change 3 showed bench-draws recording at 1.09 ms against 1.18 ms. An hour later, the same code measured 1.30 ms, and the old code 1.20 ms. On one machine, one run of a benchmark can drift by 15% within the hour, so only runs made minutes apart count as a comparison. The review also found a real hole: ExecuteIndirect could extend the vertex-buffer count past the part that was copied, which left slots uninitialised. So the change was slower on one benchmark, only slightly faster on the other, and less safe. I reverted it.

The perf harness caught change 2, as bench-room's GPU timestamps stopped advancing. I did not run the conformance tests on it.

Numbers

• Spike: atomic after a write-combined store 81.3 and 92.0 ns, 7.0 ns otherwise; reads from write-combined memory 69 ns per cache line.
• A/B for change 3: bench-draws record 1.20 to 1.29 ms, drawstress frame 48.6 to 47.3 ms, drawstress unix CPU 32.7 to 32.9 ms.
• Full test run: 837 pass, 28 fail, 76 skip of 941, the same as the previous entry.