GPU: support memory, core count detection

src/detection/gpu/gpu_linux.c

@@ -1,6 +1,6 @@
 #include "detection/gpu/gpu.h"
+#include "detection/gpu/gpu_nvidia.h"
 #include "detection/vulkan/vulkan.h"
-#include "detection/temps/temps_nvidia.h"
 
 #ifdef FF_HAVE_LIBPCI
 #include "common/io/io.h"
@@ -247,11 +247,18 @@ static void pciHandleDevice(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
         pciDetectTemp(gpu, device);
     #endif
 
-    if (gpu->temperature != gpu->temperature && gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA)
+    if (gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA)
     {
         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);
+
+        ffDetectNvidiaGpuInfo((FFGpuNvidiaCondition) { .pciBusId = pciDeviceId }, (FFGpuNvidiaResult) {
+            .temp = gpu->temperature != gpu->temperature ? &gpu->temperature : NULL,
+            .memory = &gpu->dedicated,
+            .coreCount = (uint32_t*) &gpu->coreCount,
+        });
+
+        gpu->type = gpu->dedicated.total > 1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
     }
 }
 

src/detection/gpu/gpu_nvidia.c

@@ -0,0 +1,104 @@
+#include "gpu_nvidia.h"
+
+#include "3rdparty/nvml/nvml.h"
+#include "common/library.h"
+
+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)
+    FF_LIBRARY_SYMBOL(nvmlDeviceGetMemoryInfo_v2)
+    FF_LIBRARY_SYMBOL(nvmlDeviceGetNumGpuCores)
+
+    bool inited;
+} nvmlData;
+
+const char* ffDetectNvidiaGpuInfo(FFGpuNvidiaCondition cond, FFGpuNvidiaResult result)
+{
+    if (!nvmlData.inited)
+    {
+        nvmlData.inited = true;
+        FF_LIBRARY_LOAD(libnvml, NULL, "dlopen nvml failed",
+#ifdef _WIN32
+            "nvml"
+#else
+            "libnvidia-ml"
+#endif
+            FF_LIBRARY_EXTENSION, -1
+        );
+        FF_LIBRARY_LOAD_SYMBOL_MESSAGE(libnvml, nvmlInit_v2)
+        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)
+        FF_LIBRARY_LOAD_SYMBOL_VAR_MESSAGE(libnvml, nvmlData, nvmlDeviceGetMemoryInfo_v2)
+        FF_LIBRARY_LOAD_SYMBOL_VAR_MESSAGE(libnvml, nvmlData, nvmlDeviceGetNumGpuCores)
+
+        if (ffnvmlInit_v2() != NVML_SUCCESS)
+        {
+            nvmlData.ffnvmlDeviceGetNumGpuCores = NULL;
+            return "nvmlInit_v2() failed";
+        }
+        atexit((void*) ffnvmlShutdown);
+        libnvml = NULL; // don't close nvml
+    }
+
+    if (nvmlData.ffnvmlDeviceGetNumGpuCores == NULL)
+        return "loading nvml library failed";
+
+    nvmlDevice_t device = NULL;
+    if (cond.pciBusId)
+    {
+        nvmlReturn_t ret = nvmlData.ffnvmlDeviceGetHandleByPciBusId_v2(cond.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";
+
+        for (uint32_t i = 0; i < count; i++, device = NULL)
+        {
+            if (nvmlData.ffnvmlDeviceGetHandleByIndex_v2(i, &device) != NVML_SUCCESS)
+                continue;
+
+            nvmlPciInfo_t pciInfo;
+            if (nvmlData.ffnvmlDeviceGetPciInfo_v3(device, &pciInfo) != NVML_SUCCESS)
+                continue;
+
+            if (pciInfo.pciDeviceId != cond.pciDeviceId || pciInfo.pciSubSystemId != cond.pciSubSystemId)
+                continue;
+
+            break;
+        }
+        if (!device) return "Device not found";
+    }
+
+    if (result.temp)
+    {
+        uint32_t value;
+        if (nvmlData.ffnvmlDeviceGetTemperature(device, NVML_TEMPERATURE_GPU, &value) == NVML_SUCCESS)
+            *result.temp = value;
+    }
+
+    if (result.memory)
+    {
+        nvmlMemory_v2_t memory = { .version = nvmlMemory_v2 };
+        if (nvmlData.ffnvmlDeviceGetMemoryInfo_v2(device, &memory) == NVML_SUCCESS)
+        {
+            result.memory->total = memory.used + memory.free;
+            result.memory->used = memory.used;
+        }
+    }
+
+    if (result.coreCount)
+        nvmlData.ffnvmlDeviceGetNumGpuCores(device, result.coreCount);
+
+    return NULL;
+}

src/detection/gpu/gpu_nvidia.h

@@ -0,0 +1,23 @@
+#pragma once
+
+#include "gpu.h"
+
+// Use pciBusId if not NULL; use pciDeviceId and pciSubSystemId otherwise
+// pciBusId = "domain:bus:device.function"
+// pciDeviceId = (deviceId << 16) | vendorId
+typedef struct FFGpuNvidiaCondition
+{
+    const char* pciBusId;
+    uint32_t pciDeviceId;
+    uint32_t pciSubSystemId;
+} FFGpuNvidiaCondition;
+
+// detect x if not NULL
+typedef struct FFGpuNvidiaResult
+{
+    double* temp;
+    FFGPUMemory* memory;
+    uint32_t* coreCount;
+} FFGpuNvidiaResult;
+
+const char* ffDetectNvidiaGpuInfo(FFGpuNvidiaCondition cond, FFGpuNvidiaResult result);

src/detection/gpu/gpu_windows.c

@@ -1,5 +1,5 @@
 #include "gpu.h"
-#include "detection/temps/temps_nvidia.h"
+#include "detection/gpu/gpu_nvidia.h"
 #include "util/windows/unicode.h"
 #include "util/windows/registry.h"
 
@@ -85,13 +85,22 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
             }
         }
 
-        if (options->temp && gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA)
+        if (gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA && (options->temp || instance.config.general.allowSlowOperations))
         {
             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);
+            {
+                ffDetectNvidiaGpuInfo((FFGpuNvidiaCondition) {
+                    .pciDeviceId = (deviceId << 16) | vendorId,
+                    .pciSubSystemId = subSystemId,
+                }, (FFGpuNvidiaResult) {
+                    .temp = options->temp ? &gpu->temperature : NULL,
+                    .memory = &gpu->dedicated,
+                    .coreCount = (uint32_t*) &gpu->coreCount,
+                });
+            }
         }
     }