GPU: better nvidia GPU temp detection

src/3rdparty/nvml/repo.json

@@ -1,6 +1,6 @@
 {
-    "home": "https://developer.nvidia.com/gpu-deployment-kit",
+    "home": "https://github.com/NVIDIA/nvidia-settings/blob/main/src/nvml.h",
     "license": "Embed in source",
-    "version": "v9",
+    "version": "545.23.06",
     "author": "NVIDIA Corporation"
 }

src/detection/gpu/gpu_linux.c

@@ -243,7 +243,11 @@ static void pciHandleDevice(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
     #endif
 
     if (gpu->temperature != gpu->temperature && gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA)
-        ffDetectNvidiaGpuTemp(&gpu->temperature, device->device_id);
+    {
+        char pciDeviceId[32];
+        snprintf(pciDeviceId, sizeof(pciDeviceId) - 1, "%04x:%02x:%02x.%d", device->domain, device->bus, device->dev, device->func);
+        ffDetectNvidiaGpuTemp(&gpu->temperature, pciDeviceId, 0, 0);
+    }
 }
 
 jmp_buf pciInitJmpBuf;

src/detection/gpu/gpu_windows.c

@@ -75,13 +75,6 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
                     gpu->dedicated.total = dedicatedVideoMemory + dedicatedSystemMemory;
                     gpu->shared.total = sharedSystemMemory;
                 }
-
-                if (options->temp && gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA)
-                {
-                    uint32_t deviceId;
-                    if(ffRegReadUint(hDirectxKey, L"DeviceId", &deviceId, NULL))
-                        ffDetectNvidiaGpuTemp(&gpu->temperature, deviceId);
-                }
             }
             else if (!ffRegReadUint64(hKey, L"HardwareInformation.qwMemorySize", &gpu->dedicated.total, NULL))
             {
@@ -91,6 +84,15 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
                 gpu->type = gpu->dedicated.total > 1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
             }
         }
+
+        if (options->temp && gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA)
+        {
+            uint32_t vendorId, deviceId, subSystemId;
+            // See: https://download.nvidia.com/XFree86/Linux-x86_64/545.23.06/README/supportedchips.html
+            // displayDevice.DeviceID = MatchingDeviceId "PCI\\VEN_10DE&DEV_2782&SUBSYS_513417AA&REV_A1"
+            if (swscanf(displayDevice.DeviceID, L"PCI\\VEN_%x&DEV_%x&SUBSYS_%x", &vendorId, &deviceId, &subSystemId) == 3)
+                ffDetectNvidiaGpuTemp(&gpu->temperature, NULL, (deviceId << 16) | vendorId, subSystemId);
+        }
     }
 
     return NULL;

src/detection/temps/temps_nvidia.c

@@ -6,13 +6,17 @@
 struct FFNvmlData {
     FF_LIBRARY_SYMBOL(nvmlDeviceGetCount_v2)
     FF_LIBRARY_SYMBOL(nvmlDeviceGetHandleByIndex_v2)
+    FF_LIBRARY_SYMBOL(nvmlDeviceGetHandleByPciBusId_v2)
     FF_LIBRARY_SYMBOL(nvmlDeviceGetPciInfo_v3)
     FF_LIBRARY_SYMBOL(nvmlDeviceGetTemperature)
 
     bool inited;
 } nvmlData;
 
-const char* ffDetectNvidiaGpuTemp(double* temp, uint32_t deviceId)
+// Use pciBusId if not NULL; use pciDeviceId and pciSubSystemId otherwise
+// pciBusId = "domain:bus:device.function"
+// pciDeviceId = (deviceId << 16) | vendorId
+const char* ffDetectNvidiaGpuTemp(double* temp, const char* pciBusId, uint32_t pciDeviceId, uint32_t pciSubSystemId)
 {
     if (!nvmlData.inited)
     {
@@ -29,6 +33,7 @@ const char* ffDetectNvidiaGpuTemp(double* temp, uint32_t deviceId)
         FF_LIBRARY_LOAD_SYMBOL_MESSAGE(libnvml, nvmlShutdown)
         FF_LIBRARY_LOAD_SYMBOL_VAR_MESSAGE(libnvml, nvmlData, nvmlDeviceGetCount_v2)
         FF_LIBRARY_LOAD_SYMBOL_VAR_MESSAGE(libnvml, nvmlData, nvmlDeviceGetHandleByIndex_v2)
+        FF_LIBRARY_LOAD_SYMBOL_VAR_MESSAGE(libnvml, nvmlData, nvmlDeviceGetHandleByPciBusId_v2)
         FF_LIBRARY_LOAD_SYMBOL_VAR_MESSAGE(libnvml, nvmlData, nvmlDeviceGetPciInfo_v3)
         FF_LIBRARY_LOAD_SYMBOL_VAR_MESSAGE(libnvml, nvmlData, nvmlDeviceGetTemperature)
 
@@ -44,30 +49,40 @@ const char* ffDetectNvidiaGpuTemp(double* temp, uint32_t deviceId)
     if (nvmlData.ffnvmlDeviceGetTemperature == NULL)
         return "loading nvml library failed";
 
-    uint32_t count;
-    if (nvmlData.ffnvmlDeviceGetCount_v2(&count) != NVML_SUCCESS)
-        return "nvmlDeviceGetCount_v2() failed";
-
-    for (uint32_t i = 0; i < count; i++)
+    nvmlDevice_t device = NULL;
+    if (pciBusId)
     {
-        nvmlDevice_t device;
-        if (nvmlData.ffnvmlDeviceGetHandleByIndex_v2(i, &device) != NVML_SUCCESS)
-            continue;
+        nvmlReturn_t ret = nvmlData.ffnvmlDeviceGetHandleByPciBusId_v2(pciBusId, &device);
+        if (ret != NVML_SUCCESS)
+            return "nvmlDeviceGetHandleByPciBusId_v2() failed";
+    }
+    else
+    {
+        uint32_t count;
+        if (nvmlData.ffnvmlDeviceGetCount_v2(&count) != NVML_SUCCESS)
+            return "nvmlDeviceGetCount_v2() failed";
 
-        nvmlPciInfo_t pciInfo;
-        if (nvmlData.ffnvmlDeviceGetPciInfo_v3(device, &pciInfo) != NVML_SUCCESS)
-            continue;
+        for (uint32_t i = 0; i < count; i++, device = NULL)
+        {
+            if (nvmlData.ffnvmlDeviceGetHandleByIndex_v2(i, &device) != NVML_SUCCESS)
+                continue;
 
-        if (pciInfo.pciDeviceId >> 16 != deviceId)
-            continue;
+            nvmlPciInfo_t pciInfo;
+            if (nvmlData.ffnvmlDeviceGetPciInfo_v3(device, &pciInfo) != NVML_SUCCESS)
+                continue;
 
-        uint32_t value;
-        if (nvmlData.ffnvmlDeviceGetTemperature(device, NVML_TEMPERATURE_GPU, &value) != NVML_SUCCESS)
-            return "nvmlDeviceGetTemperature(device, NVML_TEMPERATURE_GPU, &value) failed";
+            if (pciInfo.pciDeviceId != pciDeviceId || pciInfo.pciSubSystemId != pciSubSystemId)
+                continue;
 
-        *temp = value;
-        return NULL;
+            break;
+        }
+        if (!device) return "Device not found";
     }
 
-    return "Device not found";
+    uint32_t value;
+    if (nvmlData.ffnvmlDeviceGetTemperature(device, NVML_TEMPERATURE_GPU, &value) != NVML_SUCCESS)
+        return "nvmlDeviceGetTemperature(device, NVML_TEMPERATURE_GPU, &value) failed";
+
+    *temp = value;
+    return NULL;
 }

src/detection/temps/temps_nvidia.h

@@ -2,4 +2,4 @@
 
 #include <stdint.h>
 
-const char* ffDetectNvidiaGpuTemp(double* temp, uint32_t deviceId);
+const char* ffDetectNvidiaGpuTemp(double* temp, const char* pciBusId, uint32_t pciDeviceId, uint32_t pciSubSystemId);