Files
llama-swap/internal/perf/pdh_windows.go
T
George 0ab214d1c8 perf: add vendor-agnostic GPU monitoring for Windows (experimental) (#779)
Add GPU monitoring support for AMD and Intel GPUs on Windows using
D3DKMT (DirectX) and PDH performance counters.

- Add PDH-based GPU utilization via \GPU Engine(*)\Utilization
Percentage counter, summing all engine types per adapter (3D, Compute,
Copy, Video).
- Add D3DKMT bindings for adapter enumeration, memory segments, and
adapter perf data.
- Use PDH as primary utilization source (works on all vendors), with
D3DKMT RunningTime as fallback for systems without PDH counters.
- Prefer nvidia-smi when available, fall back to D3DKMT + PDH for
AMD/Intel.
- Backend priority: nvidia-smi -> D3DKMT + PDH -> ErrNoGpuTool.

Verified on AMD 7900XTX GPU with llama.cpp Vulkan & ROCm backend: GPU
utilization correctly shows ~99% during inference, ~0-2% when idle.

---

LLM disclosure: GLM 5.1 & Kimi K2.6 have been used extensively during
exploration and coding to the point that the LLM's wrote over 3/4 of the
code, and I have done additional verification myself.
As such, it should be considered experimental.
Additional verification is needed.

I have tested it on my 7900XTX system with Windows 11, and it works
correctly, but as I only have this one rig, I cannot verify it
everywhere.
2026-06-16 21:49:09 -07:00

160 lines
3.9 KiB
Go

//go:build windows
package perf
import (
"fmt"
"strconv"
"strings"
"unsafe"
"golang.org/x/sys/windows"
)
var (
pdhDLL = windows.NewLazySystemDLL("pdh.dll")
procPdhOpenQuery = pdhDLL.NewProc("PdhOpenQueryW")
procPdhAddEnglishCounter = pdhDLL.NewProc("PdhAddEnglishCounterW")
procPdhCollectQueryData = pdhDLL.NewProc("PdhCollectQueryData")
procPdhGetFormattedCounterArray = pdhDLL.NewProc("PdhGetFormattedCounterArrayW")
procPdhCloseQuery = pdhDLL.NewProc("PdhCloseQuery")
)
const (
pdhFmtDouble = 0x00000200
pdhMoreData = 0x800007D2
pdhNoData = 0x800007D5
)
type pdhCounterValue struct {
CStatus uint32
DblVal float64
}
type pdhCounterValueItem struct {
SzName *uint16
FmtValue pdhCounterValue
}
func init() {
var item pdhCounterValueItem
if unsafe.Sizeof(item) != 24 {
panic(fmt.Sprintf("pdhCounterValueItem size %d != expected 24 on x64", unsafe.Sizeof(item)))
}
}
type pdhGpuUtil struct {
query uintptr
counter uintptr
}
// initPdhGpuUtil creates a PDH query for the GPU Engine utilization counter.
// Returns nil with an error if PDH or the counter is unavailable.
func initPdhGpuUtil() (*pdhGpuUtil, error) {
var query uintptr
if ret, _, _ := procPdhOpenQuery.Call(0, 0, uintptr(unsafe.Pointer(&query))); ret != 0 {
return nil, fmt.Errorf("PdhOpenQuery: 0x%x", ret)
}
path, _ := windows.UTF16PtrFromString(`\GPU Engine(*)\Utilization Percentage`)
var counter uintptr
if ret, _, _ := procPdhAddEnglishCounter.Call(
query, uintptr(unsafe.Pointer(path)), 0, uintptr(unsafe.Pointer(&counter)),
); ret != 0 {
procPdhCloseQuery.Call(query)
return nil, fmt.Errorf("PdhAddEnglishCounter(GPU Engine): 0x%x", ret)
}
procPdhCollectQueryData.Call(query)
return &pdhGpuUtil{query: query, counter: counter}, nil
}
// close releases the PDH query handle.
func (p *pdhGpuUtil) close() {
if p.query != 0 {
procPdhCloseQuery.Call(p.query)
p.query = 0
}
}
// collect reads the PDH counter and returns a map of adapter LUID to
// aggregated GPU utilization percentage, summed across all engine instances
// per adapter and clamped to 100%.
func (p *pdhGpuUtil) collect() map[LUID]float64 {
ret, _, _ := procPdhCollectQueryData.Call(p.query)
if ret != 0 && ret != pdhNoData {
return nil
}
var bufSize uint32
var itemCount uint32
ret, _, _ = procPdhGetFormattedCounterArray.Call(
p.counter, pdhFmtDouble,
uintptr(unsafe.Pointer(&bufSize)),
uintptr(unsafe.Pointer(&itemCount)),
0,
)
if ret != pdhMoreData || itemCount == 0 {
return nil
}
buf := make([]byte, bufSize)
ret, _, _ = procPdhGetFormattedCounterArray.Call(
p.counter, pdhFmtDouble,
uintptr(unsafe.Pointer(&bufSize)),
uintptr(unsafe.Pointer(&itemCount)),
uintptr(unsafe.Pointer(&buf[0])),
)
if ret != 0 {
return nil
}
itemSize := uint32(unsafe.Sizeof(pdhCounterValueItem{}))
result := make(map[LUID]float64)
for i := uint32(0); i < itemCount; i++ {
item := (*pdhCounterValueItem)(unsafe.Pointer(&buf[i*itemSize]))
if item.FmtValue.CStatus != 0 {
continue
}
luid, ok := parsePdhLuid(windows.UTF16PtrToString(item.SzName))
if !ok {
continue
}
result[luid] += item.FmtValue.DblVal
}
for luid := range result {
if result[luid] > 100.0 {
result[luid] = 100.0
}
}
return result
}
// parsePdhLuid extracts the adapter LUID (high and low parts) from a PDH
// GPU Engine instance name (e.g. "pid_1234_luid_0x00000000_0x000148BF_phys_0_eng_2_engtype_Compute").
func parsePdhLuid(name string) (LUID, bool) {
idx := strings.Index(name, "luid_0x")
if idx < 0 {
return LUID{}, false
}
rest := name[idx+7:]
parts := strings.SplitN(rest, "_", 4)
if len(parts) < 3 {
return LUID{}, false
}
hp, err := strconv.ParseUint(parts[0], 16, 32)
if err != nil {
return LUID{}, false
}
lpStr := strings.TrimPrefix(parts[1], "0x")
lp, err := strconv.ParseUint(lpStr, 16, 32)
if err != nil {
return LUID{}, false
}
return LUID{LowPart: uint32(lp), HighPart: int32(hp)}, true
}