d6f8dda823
On Apple GPUs timestamps go only at pass boundaries, and the frame was bracketed by two one-thread compute passes. Two things made it read about 0.01 ms: - Metal resolves counter samples without waiting for the passes that write them. Resolved in the frame's own command buffer, the end timestamp was mostly the one an earlier frame left in the shared query. Each readback slot now has its own queries, and with markers a slot is resolved in a later frame, once on_submitted_work_done says the GPU finished the frame it timed. - A compute pass depends on nothing, so the end marker ran alongside the frame's render passes. The end is now a render pass that loads the frame's target, shading one pixel with every write masked: it runs after every pass that drew into the target. The start stays a compute pass (cheaper than a load and store of the target). M4 Pro, render distance 32, offscreen: GPU p50 0.002 ms before, now 1.95 ms against a 1.95 ms frame (village) and 4.46 ms against 4.51 ms (mountain), where the frame is GPU-bound. Vulkan and DirectX 12, which write timestamps between commands, are unchanged but for the per-slot queries. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>