// igneum-gpu-telemetry: power, temperature, fan and clocks of every AMD GPU, one line per card per sample.
// 5 October 2026, after the project lead watched a 9070 XT at 90% usage with its fans barely turning and the app could not say
// what it drew (the app's draw, temperature and MH per watt line came from nvidia-smi only).
//
//   igneum-gpu-telemetry [-l SECONDS]      one sample (default), or one every SECONDS until stdin closes or SIGTERM
//
// Windows: ADLX (the AMD Device Library eXtra, amdadlx64.dll, shipped with Adrenalin; vendor/adlx is the SDK clone,
// MIT) for the metrics, keyed by the card's PCI bus from SetupAPI (the display class, matched by the same name ADLX
// reports). Without ADLX (no AMD driver, an old one, or the DLL missing) only the utilisation is read, from the
// GPU Engine performance counters through PDH, keyed by the adapter LUID that Windows uses there.
// Linux: the amdgpu sysfs (/sys/class/drm/card*/device: hwmon power1_average, temp1_input, fan1_input, pwm1,
// pp_dpm_mclk, gpu_busy_percent), keyed by the PCI address of the device link.
//
// Line format (space separated, every field present, a value the source cannot give prints as -):
//   amd <ordinal> bus <pci bus or address> kind integrated|discrete name "<name>" watts <W> temp_c <C> fan_rpm <rpm>
//   fan_pct <%> mclk_mhz <MHz> gclk_mhz <MHz> util_pct <%> source adlx|sysfs|perfcounter
// then one `end <ms>` line per sample. The app (engine.rs amd_telemetry_line) parses it; parsers are unit-tested
// against lines captured on PC 1.
#define _CRT_SECURE_NO_WARNINGS
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <signal.h>

static volatile int gStop = 0;
static void onSignal(int s) { (void)s; gStop = 1; }

typedef struct {
    char bus[64];
    char kind[16];
    char name[128];
    double watts, tempC, fanRpm, fanPct, mclk, gclk, util;   /* -1 = not available */
    const char* source;
} Sample;

static void sampleInit(Sample* s) { memset(s, 0, sizeof(*s)); strcpy(s->bus, "-"); strcpy(s->kind, "-"); strcpy(s->name, "-"); s->watts = s->tempC = s->fanRpm = s->fanPct = s->mclk = s->gclk = s->util = -1.0; s->source = "-"; }
static void printNum(double v, const char* fmt) { if (v < 0) printf(" -"); else printf(fmt, v); }
static void printSample(int ordinal, const Sample* s) {
    printf("amd %d bus %s kind %s name \"%s\" watts", ordinal, s->bus, s->kind, s->name);
    printNum(s->watts, " %.1f"); printf(" temp_c"); printNum(s->tempC, " %.1f"); printf(" fan_rpm"); printNum(s->fanRpm, " %.0f");
    printf(" fan_pct"); printNum(s->fanPct, " %.0f"); printf(" mclk_mhz"); printNum(s->mclk, " %.0f"); printf(" gclk_mhz"); printNum(s->gclk, " %.0f");
    printf(" util_pct"); printNum(s->util, " %.0f"); printf(" source %s\n", s->source);
}

#ifdef _WIN32
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#include <setupapi.h>
#include <pdh.h>
#include "../vendor/adlx/SDK/ADLXHelper/Windows/C/ADLXHelper.h"
#include "../vendor/adlx/SDK/Include/IPerformanceMonitoring.h"

/* The SDK declares these three and leaves them to the platform file of each sample. */
adlx_handle ADLX_CDECL_CALL adlx_load_library(const TCHAR* filename) { return (adlx_handle)LoadLibrary(filename); }
int ADLX_CDECL_CALL adlx_free_library(adlx_handle module) { return FreeLibrary((HMODULE)module) ? 1 : 0; }
void* ADLX_CDECL_CALL adlx_get_proc_address(adlx_handle module, const char* procName) { return (void*)GetProcAddress((HMODULE)module, procName); }

static double nowMs(void) { LARGE_INTEGER f, c; QueryPerformanceFrequency(&f); QueryPerformanceCounter(&c); return (double)c.QuadPart * 1000.0 / (double)f.QuadPart; }

/* The PCI bus of every display-class device, by its name (SetupAPI; the names are the ones ADLX reports). */
typedef struct { char name[128]; int bus; } BusEntry;
static int listBuses(BusEntry* out, int cap) {
    static const GUID DISPLAY = { 0x4d36e968, 0xe325, 0x11ce, { 0xbf, 0xc1, 0x08, 0x00, 0x2b, 0xe1, 0x03, 0x18 } };
    HDEVINFO set = SetupDiGetClassDevsA(&DISPLAY, NULL, NULL, DIGCF_PRESENT);
    SP_DEVINFO_DATA d;
    DWORD i;
    int n = 0;
    if (set == INVALID_HANDLE_VALUE) return 0;
    d.cbSize = sizeof(d);
    for (i = 0; SetupDiEnumDeviceInfo(set, i, &d) && n < cap; ++i) {
        char name[128] = { 0 };
        DWORD bus = 0, type = 0, got = 0;
        if (!SetupDiGetDeviceRegistryPropertyA(set, &d, SPDRP_DEVICEDESC, &type, (BYTE*)name, sizeof(name) - 1, &got)) continue;
        if (!SetupDiGetDeviceRegistryPropertyA(set, &d, SPDRP_BUSNUMBER, &type, (BYTE*)&bus, sizeof(bus), &got)) continue;
        snprintf(out[n].name, sizeof(out[n].name), "%s", name); out[n].bus = (int)bus; ++n;
    }
    SetupDiDestroyDeviceInfoList(set);
    return n;
}
static int busOf(const BusEntry* b, int n, const char* name, int* taken) {
    int i;
    for (i = 0; i < n; ++i) if (!taken[i] && strcmp(b[i].name, name) == 0) { taken[i] = 1; return b[i].bus; }
    return -1;
}

/* ADLX: one sample of every GPU. Returns the number of lines printed, -1 when ADLX is not usable (reason printed). */
static IADLXSystem* gSys = NULL;
static IADLXPerformanceMonitoringServices* gPerf = NULL;
static int adlxOpen(void) {
    ADLX_RESULT r = ADLXHelper_Initialize();
    if (!ADLX_SUCCEEDED(r)) { printf("info adlx: ADLXHelper_Initialize returned %d (no AMD driver with ADLX; amdadlx64.dll missing or too old)\n", (int)r); return 0; }
    gSys = ADLXHelper_GetSystemServices();
    if (!gSys) { printf("info adlx: no system services\n"); return 0; }
    r = gSys->pVtbl->GetPerformanceMonitoringServices(gSys, &gPerf);
    if (!ADLX_SUCCEEDED(r) || !gPerf) { printf("info adlx: GetPerformanceMonitoringServices returned %d\n", (int)r); return 0; }
    return 1;
}
static int adlxSample(const BusEntry* buses, int nBuses) {
    IADLXGPUList* gpus = NULL;
    adlx_uint it;
    int ordinal = 0;
    int taken[32] = { 0 };
    ADLX_RESULT r = gSys->pVtbl->GetGPUs(gSys, &gpus);
    if (!ADLX_SUCCEEDED(r) || !gpus) { printf("info adlx: GetGPUs returned %d\n", (int)r); return 0; }
    for (it = gpus->pVtbl->Begin(gpus); it != gpus->pVtbl->End(gpus); ++it) {
        IADLXGPU* gpu = NULL;
        IADLXGPUMetrics* m = NULL;
        Sample s;
        const char* name = NULL;
        ADLX_GPU_TYPE type = GPUTYPE_UNDEFINED;
        adlx_double dv = 0; adlx_int iv = 0;
        if (!ADLX_SUCCEEDED(gpus->pVtbl->At_GPUList(gpus, it, &gpu)) || !gpu) continue;
        sampleInit(&s);
        s.source = "adlx";
        if (ADLX_SUCCEEDED(gpu->pVtbl->Name(gpu, &name)) && name) snprintf(s.name, sizeof(s.name), "%s", name);
        if (ADLX_SUCCEEDED(gpu->pVtbl->Type(gpu, &type))) strcpy(s.kind, type == GPUTYPE_INTEGRATED ? "integrated" : type == GPUTYPE_DISCRETE ? "discrete" : "-");
        { int b = busOf(buses, nBuses, s.name, taken); if (b >= 0) snprintf(s.bus, sizeof(s.bus), "%d", b); }
        r = gPerf->pVtbl->GetCurrentGPUMetrics(gPerf, gpu, &m);
        if (ADLX_SUCCEEDED(r) && m) {
            if (ADLX_SUCCEEDED(m->pVtbl->GPUPower(m, &dv))) s.watts = dv;
            if (s.watts < 0 && ADLX_SUCCEEDED(m->pVtbl->GPUTotalBoardPower(m, &dv))) s.watts = dv;
            if (ADLX_SUCCEEDED(m->pVtbl->GPUTemperature(m, &dv))) s.tempC = dv;
            if (ADLX_SUCCEEDED(m->pVtbl->GPUFanSpeed(m, &iv))) s.fanRpm = iv;
            if (ADLX_SUCCEEDED(m->pVtbl->GPUVRAMClockSpeed(m, &iv))) s.mclk = iv;
            if (ADLX_SUCCEEDED(m->pVtbl->GPUClockSpeed(m, &iv))) s.gclk = iv;
            if (ADLX_SUCCEEDED(m->pVtbl->GPUUsage(m, &dv))) s.util = dv;
            m->pVtbl->Release(m);
        } else {
            printf("info adlx: GetCurrentGPUMetrics for \"%s\" returned %d\n", s.name, (int)r);
        }
        /* fan percent: ADLX gives rpm only here; the tuning interface has the range, the app shows rpm when pct is - */
        printSample(ordinal++, &s);
        gpu->pVtbl->Release(gpu);
    }
    gpus->pVtbl->Release(gpus);
    return ordinal;
}

/* PDH fallback: GPU engine utilisation per adapter LUID, summed over the engines (no power, no temperature). */
static int pdhSample(void) {
    PDH_HQUERY q = NULL;
    PDH_HCOUNTER c = NULL;
    DWORD size = 0, count = 0, i;
    PDH_FMT_COUNTERVALUE_ITEM_A* items;
    int ordinal = 0;
    if (PdhOpenQueryA(NULL, 0, &q) != ERROR_SUCCESS) { printf("info perfcounter: PdhOpenQuery failed\n"); return 0; }
    if (PdhAddEnglishCounterA(q, "\\GPU Engine(*)\\Utilization Percentage", 0, &c) != ERROR_SUCCESS) { printf("info perfcounter: no GPU Engine counters\n"); PdhCloseQuery(q); return 0; }
    PdhCollectQueryData(q); Sleep(1000); PdhCollectQueryData(q);
    PdhGetFormattedCounterArrayA(c, PDH_FMT_DOUBLE, &size, &count, NULL);
    items = (PDH_FMT_COUNTERVALUE_ITEM_A*)malloc(size ? size : 1);
    if (PdhGetFormattedCounterArrayA(c, PDH_FMT_DOUBLE, &size, &count, items) == ERROR_SUCCESS) {
        /* instance names: pid_1234_luid_0x00000000_0x0000D4E3_phys_0_eng_0_engtype_3D; sum per luid */
        char luids[16][40]; double sums[16]; int n = 0, k;
        for (i = 0; i < count; ++i) {
            const char* p = strstr(items[i].szName, "luid_");
            char luid[40];
            if (!p) continue;
            snprintf(luid, sizeof(luid), "%.39s", p); { char* e = strstr(luid, "_phys"); if (e) *e = 0; }
            for (k = 0; k < n; ++k) if (strcmp(luids[k], luid) == 0) break;
            if (k == n && n < 16) { strcpy(luids[n], luid); sums[n] = 0; ++n; }
            if (k < 16) sums[k] += items[i].FmtValue.doubleValue;
        }
        for (k = 0; k < n; ++k) {
            Sample s; sampleInit(&s); s.source = "perfcounter";
            snprintf(s.bus, sizeof(s.bus), "%s", luids[k]);
            s.util = sums[k] > 100.0 ? 100.0 : sums[k];
            printSample(ordinal++, &s);
        }
    }
    free(items);
    PdhCloseQuery(q);
    return ordinal;
}

int main(int argc, char** argv) {
    int every = 0, i, haveAdlx;
    BusEntry buses[32];
    int nBuses;
    for (i = 1; i < argc; ++i) if (strcmp(argv[i], "-l") == 0 && i + 1 < argc) every = atoi(argv[++i]);
    signal(SIGINT, onSignal); signal(SIGTERM, onSignal);
    setvbuf(stdout, NULL, _IOLBF, 0);
    nBuses = listBuses(buses, 32);
    for (i = 0; i < nBuses; ++i) printf("info display device \"%s\" bus %d\n", buses[i].name, buses[i].bus);
    haveAdlx = adlxOpen();
    do {
        double t0 = nowMs();
        int n = haveAdlx ? adlxSample(buses, nBuses) : pdhSample();
        printf("end %.1f ms %d card(s)\n", nowMs() - t0, n);
        fflush(stdout);   /* a redirected stdout is fully buffered on the Windows CRT whatever setvbuf asks (PC 1 lost 60 s of samples at the kill) */
        if (every > 0) Sleep((DWORD)every * 1000);
    } while (every > 0 && !gStop);
    if (haveAdlx) { if (gPerf) gPerf->pVtbl->Release(gPerf); ADLXHelper_Terminate(); }
    return 0;
}
#else
#include <dirent.h>
#include <unistd.h>
#include <time.h>
static double nowMs(void) { struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts); return ts.tv_sec * 1000.0 + ts.tv_nsec / 1e6; }
static int readText(const char* path, char* out, size_t cap) { FILE* f = fopen(path, "r"); size_t n; if (!f) return 0; n = fread(out, 1, cap - 1, f); fclose(f); out[n] = 0; return 1; }
static double readNumber(const char* path) { char b[64]; if (!readText(path, b, sizeof(b))) return -1.0; return atof(b); }
/* pp_dpm_mclk: lines "0: 96Mhz", "3: 1258Mhz *"; the starred line is the current state */
static double dpmCurrent(const char* text) {
    const char* p = text;
    while (p && *p) {
        const char* nl = strchr(p, '\n');
        size_t len = nl ? (size_t)(nl - p) : strlen(p);
        const char* star = memchr(p, '*', len);
        if (star) { const char* colon = memchr(p, ':', len); if (colon) return atof(colon + 1); }
        p = nl ? nl + 1 : NULL;
    }
    return -1.0;
}
static int sysfsSample(const char* root) {
    DIR* d = opendir(root);
    struct dirent* e;
    int ordinal = 0;
    if (!d) { printf("info sysfs: no %s\n", root); return 0; }
    while ((e = readdir(d)) != NULL) {
        char dev[512], path[640], text[4096], link[512];
        ssize_t ln;
        Sample s;
        DIR* hw; struct dirent* he;
        if (strncmp(e->d_name, "card", 4) != 0 || strchr(e->d_name + 4, '-')) continue;
        snprintf(dev, sizeof(dev), "%s/%s/device", root, e->d_name);
        snprintf(path, sizeof(path), "%s/vendor", dev);
        if (!readText(path, text, sizeof(text)) || strtol(text, NULL, 16) != 0x1002) continue;
        sampleInit(&s);
        s.source = "sysfs";
        ln = readlink(dev, link, sizeof(link) - 1);
        if (ln > 0) { link[ln] = 0; { const char* base = strrchr(link, '/'); snprintf(s.bus, sizeof(s.bus), "%.63s", base ? base + 1 : link); } }
        snprintf(path, sizeof(path), "%s/product_name", dev);
        if (readText(path, text, sizeof(text))) { text[strcspn(text, "\n")] = 0; snprintf(s.name, sizeof(s.name), "%s", text); }
        else { snprintf(path, sizeof(path), "%s/device", dev); if (readText(path, text, sizeof(text))) { text[strcspn(text, "\n")] = 0; snprintf(s.name, sizeof(s.name), "amdgpu %s", text); } }
        snprintf(path, sizeof(path), "%s/boot_vga", dev);
        strcpy(s.kind, "discrete");
        snprintf(path, sizeof(path), "%s/hwmon", dev);
        hw = opendir(path);
        if (hw) {
            while ((he = readdir(hw)) != NULL) {
                char hp[900];
                if (strncmp(he->d_name, "hwmon", 5) != 0) continue;
                snprintf(hp, sizeof(hp), "%s/%s/power1_average", path, he->d_name); s.watts = readNumber(hp); if (s.watts < 0) { snprintf(hp, sizeof(hp), "%s/%s/power1_input", path, he->d_name); s.watts = readNumber(hp); } if (s.watts >= 0) s.watts /= 1e6;
                snprintf(hp, sizeof(hp), "%s/%s/temp1_input", path, he->d_name); s.tempC = readNumber(hp); if (s.tempC >= 0) s.tempC /= 1000.0;
                snprintf(hp, sizeof(hp), "%s/%s/fan1_input", path, he->d_name); s.fanRpm = readNumber(hp);
                { double pwm, pwmMax; snprintf(hp, sizeof(hp), "%s/%s/pwm1", path, he->d_name); pwm = readNumber(hp); snprintf(hp, sizeof(hp), "%s/%s/pwm1_max", path, he->d_name); pwmMax = readNumber(hp); if (pwm >= 0 && pwmMax > 0) s.fanPct = 100.0 * pwm / pwmMax; else if (pwm >= 0) s.fanPct = 100.0 * pwm / 255.0; }
                break;
            }
            closedir(hw);
        }
        snprintf(path, sizeof(path), "%s/pp_dpm_mclk", dev); if (readText(path, text, sizeof(text))) s.mclk = dpmCurrent(text);
        snprintf(path, sizeof(path), "%s/pp_dpm_sclk", dev); if (readText(path, text, sizeof(text))) s.gclk = dpmCurrent(text);
        snprintf(path, sizeof(path), "%s/gpu_busy_percent", dev); s.util = readNumber(path);
        printSample(ordinal++, &s);
    }
    closedir(d);
    return ordinal;
}
int main(int argc, char** argv) {
    int every = 0, i;
    const char* root = getenv("IGNEUM_DRM_ROOT") ? getenv("IGNEUM_DRM_ROOT") : "/sys/class/drm";   /* a fixture tree for tests */
    for (i = 1; i < argc; ++i) if (strcmp(argv[i], "-l") == 0 && i + 1 < argc) every = atoi(argv[++i]);
    signal(SIGINT, onSignal); signal(SIGTERM, onSignal);
    setvbuf(stdout, NULL, _IOLBF, 0);
    do {
        double t0 = nowMs();
        int n = sysfsSample(root);
        printf("end %.1f ms %d card(s)\n", nowMs() - t0, n);
        fflush(stdout);   /* a redirected stdout is fully buffered on the Windows CRT whatever setvbuf asks (PC 1 lost 60 s of samples at the kill) */
        if (every > 0) sleep((unsigned)every);
    } while (every > 0 && !gStop);
    return 0;
}
#endif