Tracking issue for the in-emulator profiling instrumentation added to this fork to find performance bottlenecks (goal: BOTW 4K60 on an Intel Panther Lake B390 iGPU, Mesa ANV, Linux).
What was added
- Stats framework (
LattePerformanceMonitor.h/.cpp): 15 nesting-safe TSC timers + 18 per-frame counters under performanceMonitor.bottleneck, with LATTE_PERF_SCOPE / LATTE_PERF_COUNT / LATTE_PERF_ADD macros. Commit 715c8fa.
- Latte core call sites (command processor idle spin, guest fence waits, draw-sequence-end causes, index cache hit/miss, buffer sync, shader/FBO/texture timers, texture reload counts, emulated vsync lateness). Commit 110c070.
- Vulkan call sites + GPU timestamps (first/fast draw counters, uniform/pipeline/descriptor/renderpass/submit timers, async-compile skips, forced submits, GPU-wait buckets, and per-command-buffer GPU busy time via TOP/BOTTOM_OF_PIPE timestamp query pairs read back on fence retirement). Commit fdbe9e9.
- Overlay panel: 'Bottleneck stats' section in the debug overlay (per-frame values incl. GPU busy ms).
- CSV dump: set
CEMU_PERFSTATS_CSV=/path/file.csv to log one row per frame for offline analysis.
Toggle / overhead
Collection is gated by a single runtime flag: enabled while Options > General settings > Graphics > Overlay > Debug is checked (or while a CSV dump is active). When off, every call site costs one predictable branch. When on: two TSC reads per timed scope plus 2 GPU timestamps per command buffer.
How to read it
GPU busy vs frame time -> GPU-bound vs CPU-bound.SeqEnd tex/ctxReg -> how often draw sequences break (validates the dirty-flag optimization thesis).IdxCache hit/miss -> index cache effectiveness.Waits: gpu/idleSpin/guestFence -> stall buckets on the GPU emulation thread.
Optimizations driven by this data are tracked separately.
Tracking issue for the in-emulator profiling instrumentation added to this fork to find performance bottlenecks (goal: BOTW 4K60 on an Intel Panther Lake B390 iGPU, Mesa ANV, Linux).
What was added
LattePerformanceMonitor.h/.cpp): 15 nesting-safe TSC timers + 18 per-frame counters underperformanceMonitor.bottleneck, withLATTE_PERF_SCOPE/LATTE_PERF_COUNT/LATTE_PERF_ADDmacros. Commit 715c8fa.CEMU_PERFSTATS_CSV=/path/file.csvto log one row per frame for offline analysis.Toggle / overhead
Collection is gated by a single runtime flag: enabled while Options > General settings > Graphics > Overlay > Debug is checked (or while a CSV dump is active). When off, every call site costs one predictable branch. When on: two TSC reads per timed scope plus 2 GPU timestamps per command buffer.
How to read it
GPU busyvs frame time -> GPU-bound vs CPU-bound.SeqEnd tex/ctxReg-> how often draw sequences break (validates the dirty-flag optimization thesis).IdxCache hit/miss-> index cache effectiveness.Waits: gpu/idleSpin/guestFence-> stall buckets on the GPU emulation thread.Optimizations driven by this data are tracked separately.