// GPU-hog forensics (M4 step 3 triage): when gpu=/vsw= balloons at full
// clock, the question is WHICH process is loading the GPU. Samples the
// per-process "GPU Engine" performance counters (two PDH collections a few
// hundred ms apart), sums utilization per pid across engines, and returns
// the top consumers as one printable line. Companion to clk= — clk= says
// contention-vs-P-state, this names the contender.

#pragma once

#include <string>

namespace sauna {

// Blocking (~sampleMs). Returns e.g.
//   "chrome.exe(1234) 41%  dwm.exe(888) 12%  spatial_light.exe(77) 9%"
// or "" when the counters are unavailable. Skips entries below 5%.
std::string TopGpuConsumers(int sampleMs = 300, int maxEntries = 4);

// CPU-side stall forensics (M4 step 3 follow-up: episodes with [gpu hogs]
// at 0% and 2 s tick stretches = kernel-level stall, not GPU work).
// One PDH pass sampling kernel pressure (_Total % DPC / % interrupt /
// % privileged time) plus the top per-process CPU consumers (per-core
// scale: 100 = one core). High DPC/interrupt = a driver is storming
// (LatencyMon territory); everything low while the machine still stalls
// points at SMI/firmware, which Windows cannot see. Blocking ~sampleMs.
// Returns e.g. "dpc=4.1% int=0.8% priv=22%  top: MsMpEng 180  dwm 95".
std::string CpuStallForensics(int sampleMs = 300, int maxEntries = 4);

}  // namespace sauna