GPU (Windows): greatly improve performance of GPU temp detection

... when `--gpu-driver-specific` is not set. Supported on Windows 10 and later

Author: CarterLi

src/detection/gpu/gpu_windows.c

@@ -164,7 +164,7 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
         __typeof__(&ffDetectNvidiaGpuInfo) detectFn;
         const char* dllName;
 
-        if (getDriverSpecificDetectionFn(gpu->vendor.chars, &detectFn, &dllName) && (options->temp || options->driverSpecific))
+        if (options->driverSpecific && getDriverSpecificDetectionFn(gpu->vendor.chars, &detectFn, &dllName))
         {
             unsigned vendorId = 0, deviceId = 0, subSystemId = 0, revId = 0;
             bufferLen = sizeof(buffer);
@@ -242,6 +242,20 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
                                 gpu->type = FF_GPU_TYPE_INTEGRATED;
                         }
                     }
+
+                    if (options->temp && gpu->temperature != gpu->temperature)
+                    {
+                        D3DKMT_QUERYSTATISTICS queryStatistics = {
+                            .Type = D3DKMT_QUERYSTATISTICS_PHYSICAL_ADAPTER,
+                            .AdapterLuid = *(LUID*)&adapterLuid,
+                            .QueryPhysAdapter = { .PhysicalAdapterIndex = 0 },
+                        };
+                        if (NT_SUCCESS(D3DKMTQueryStatistics(&queryStatistics)))
+                            gpu->temperature = queryStatistics.QueryResult.PhysAdapterInformation.AdapterPerfData.Temperature / 10.0;
+                    }
+
+                    D3DKMT_CLOSEADAPTER closeAdapter = { .hAdapter = openAdapterFromLuid.hAdapter };
+                    (void) D3DKMTCloseAdapter(&closeAdapter);
                 }
             }
         }

src/util/windows/nt.h

@@ -3,6 +3,8 @@
 #include <winnt.h>
 #include <winternl.h>
 
+#define D3DKMT_ALIGN64 __attribute__((aligned(8)))
+
 typedef struct _PROCESSOR_POWER_INFORMATION {
   ULONG Number;
   ULONG MaxMhz;
@@ -48,20 +50,20 @@ typedef struct _D3DKMT_ADAPTERTYPE
     {
         struct
         {
-            UINT   RenderSupported              :  1;
+            UINT   RenderSupported              :  1; // WDDM 1.2, Windows 8
             UINT   DisplaySupported             :  1;
             UINT   SoftwareDevice               :  1;
             UINT   PostDevice                   :  1;
-            UINT   HybridDiscrete               :  1;
+            UINT   HybridDiscrete               :  1; // WDDM 1.3, Windows 8.1
             UINT   HybridIntegrated             :  1;
             UINT   IndirectDisplayDevice        :  1;
-            UINT   Paravirtualized              :  1;
+            UINT   Paravirtualized              :  1; // WDDM 2.3, Windows 10 Fall Creators Update (version 1709)
             UINT   ACGSupported                 :  1;
             UINT   SupportSetTimingsFromVidPn   :  1;
             UINT   Detachable                   :  1;
-            UINT   ComputeOnly                  :  1;
+            UINT   ComputeOnly                  :  1; // WDDM 2.6, Windows 10 May 2019 Update (Version 1903)
             UINT   Prototype                    :  1;
-            UINT   RuntimePowerManagement       :  1;
+            UINT   RuntimePowerManagement       :  1; // WDDM 2.9, Windows 10 Insider Preview "Iron"
             UINT   Reserved                     : 18;
         };
         UINT Value;
@@ -81,3 +83,67 @@ typedef struct _D3DKMT_QUERYADAPTERINFO
 } D3DKMT_QUERYADAPTERINFO;
 EXTERN_C _Check_return_ NTSTATUS APIENTRY D3DKMTQueryAdapterInfo(_Inout_ CONST D3DKMT_QUERYADAPTERINFO*);
 #endif
+
+typedef enum _D3DKMT_QUERYSTATISTICS_TYPE
+{
+    D3DKMT_QUERYSTATISTICS_PHYSICAL_ADAPTER       = 10, // WDDM 2.4, Windows 10 April 2018 Update (version 1803)
+} D3DKMT_QUERYSTATISTICS_TYPE;
+typedef struct _D3DKMT_QUERYSTATISTICS_QUERY_PHYSICAL_ADAPTER
+{
+    ULONG PhysicalAdapterIndex;
+} D3DKMT_QUERYSTATISTICS_QUERY_PHYSICAL_ADAPTER;
+typedef struct _D3DKMT_ADAPTER_PERFDATA
+{
+    UINT32          PhysicalAdapterIndex;   // in: The physical adapter index, in an LDA chain
+    D3DKMT_ALIGN64 ULONGLONG MemoryFrequency;        // out: Clock frequency of the memory in hertz
+    D3DKMT_ALIGN64 ULONGLONG MaxMemoryFrequency;     // out: Max memory clock frequency
+    D3DKMT_ALIGN64 ULONGLONG MaxMemoryFrequencyOC;   // out: Clock frequency of the memory while overclocked in hertz.
+    D3DKMT_ALIGN64 ULONGLONG MemoryBandwidth;        // out: Amount of memory transferred in bytes
+    D3DKMT_ALIGN64 ULONGLONG PCIEBandwidth;          // out: Amount of memory transferred over PCI-E in bytes
+    ULONG           FanRPM;                 // out: Fan rpm
+    ULONG           Power;                  // out: Power draw of the adapter in tenths of a percentage
+    ULONG           Temperature;            // out: Temperature in deci-Celsius 1 = 0.1C
+    UCHAR           PowerStateOverride;     // out: Overrides dxgkrnls power view of linked adapters.
+} D3DKMT_ADAPTER_PERFDATA;
+typedef struct _D3DKMT_ADAPTER_PERFDATACAPS
+{
+    UINT32      PhysicalAdapterIndex;   // in: The physical adapter index, in an LDA chain
+    D3DKMT_ALIGN64 ULONGLONG MaxMemoryBandwidth;     // out: Max memory bandwidth in bytes for 1 second
+    D3DKMT_ALIGN64 ULONGLONG MaxPCIEBandwidth;       // out: Max pcie bandwidth in bytes for 1 second
+    ULONG       MaxFanRPM;              // out: Max fan rpm
+    ULONG       TemperatureMax;         // out: Max temperature before damage levels
+    ULONG       TemperatureWarning;     // out: The temperature level where throttling begins.
+} D3DKMT_ADAPTER_PERFDATACAPS;
+
+#define DXGK_MAX_GPUVERSION_NAME_LENGTH 32
+typedef struct _D3DKMT_GPUVERSION
+{
+    UINT32          PhysicalAdapterIndex;                             // in: The physical adapter index, in an LDA chain
+    WCHAR           BiosVersion[DXGK_MAX_GPUVERSION_NAME_LENGTH];     //out: The gpu bios version
+    WCHAR           GpuArchitecture[DXGK_MAX_GPUVERSION_NAME_LENGTH]; //out: The gpu architectures name.
+} D3DKMT_GPUVERSION;
+typedef struct _D3DKMT_QUERYSTATISTICS_PHYSICAL_ADAPTER_INFORMATION
+{
+    D3DKMT_ADAPTER_PERFDATA      AdapterPerfData;
+    D3DKMT_ADAPTER_PERFDATACAPS  AdapterPerfDataCaps;
+    D3DKMT_GPUVERSION            GpuVersion;
+} D3DKMT_QUERYSTATISTICS_PHYSICAL_ADAPTER_INFORMATION;
+typedef union _D3DKMT_QUERYSTATISTICS_RESULT
+{
+    D3DKMT_QUERYSTATISTICS_PHYSICAL_ADAPTER_INFORMATION PhysAdapterInformation; // in: id of physical adapter to get statistics for
+    uint8_t Padding[776];
+} D3DKMT_QUERYSTATISTICS_RESULT;
+typedef struct _D3DKMT_QUERYSTATISTICS
+{
+    D3DKMT_QUERYSTATISTICS_TYPE   Type;        // in: type of data requested
+    LUID                          AdapterLuid; // in: adapter to get export / statistics from
+    HANDLE*                       hProcess;    // in: process to get statistics for, if required for this query type
+    D3DKMT_QUERYSTATISTICS_RESULT QueryResult; // out: requested data
+
+    union
+    {
+        D3DKMT_QUERYSTATISTICS_QUERY_PHYSICAL_ADAPTER QueryPhysAdapter;
+    };
+} D3DKMT_QUERYSTATISTICS;
+static_assert(sizeof(D3DKMT_QUERYSTATISTICS) == 0x328, "D3DKMT_QUERYSTATISTICS structure size mismatch");
+EXTERN_C _Check_return_ NTSTATUS APIENTRY D3DKMTQueryStatistics(_In_ CONST D3DKMT_QUERYSTATISTICS*);