GPU (Windows): use setupapi to find GPU devices

Ref: https://github.com/gpustack/gpustack/pull/97#issuecomment-2264678404
This commit is contained in:
李通洲
2024-08-02 21:52:24 +08:00
parent a856350416
commit 0354fd5d9d
+91 -106
View File
@@ -3,48 +3,59 @@
#include "util/windows/unicode.h"
#include "util/windows/registry.h"
#include <inttypes.h>
#define INITGUID
#include <windows.h>
#include <setupapi.h>
#include <devguid.h>
static int isGpuNameEqual(const FFGPUResult* gpu, const FFstrbuf* name)
static inline void wrapSetupDiDestroyDeviceInfoList(HDEVINFO* hdev)
{
return ffStrbufEqual(&gpu->name, name);
if(*hdev)
SetupDiDestroyDeviceInfoList(*hdev);
}
#define FF_EMPTY_GUID_STR L"{00000000-0000-0000-0000-000000000000}"
#define FF_GUID_STRLEN (sizeof(FF_EMPTY_GUID_STR) / sizeof(wchar_t) - 1)
const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist* gpus)
{
DISPLAY_DEVICEW displayDevice = { .cb = sizeof(displayDevice) };
wchar_t regDirectxKey[MAX_PATH] = L"SOFTWARE\\Microsoft\\DirectX\\{";
const uint32_t regDirectxKeyPrefixLength = (uint32_t) wcslen(regDirectxKey);
wchar_t regControlVideoKey[MAX_PATH] = L"SYSTEM\\CurrentControlSet\\Control\\Video\\{";
const uint32_t regControlVideoKeyPrefixLength = (uint32_t) wcslen(regControlVideoKey);
const uint32_t deviceKeyPrefixLength = strlen("\\Registry\\Machine\\") + regControlVideoKeyPrefixLength;
wchar_t regPciKey[MAX_PATH] = L"SYSTEM\\CurrentControlSet\\Enum\\";
const uint32_t regPciKeyPrefixLength = (uint32_t) wcslen(regPciKey);
HDEVINFO hdev __attribute__((__cleanup__(wrapSetupDiDestroyDeviceInfoList))) =
SetupDiGetClassDevsW(&GUID_DEVCLASS_DISPLAY, NULL, NULL, DIGCF_PRESENT);
for (DWORD i = 0; EnumDisplayDevicesW(NULL, i, &displayDevice, 0); ++i)
if(hdev == INVALID_HANDLE_VALUE)
return "SetupDiGetClassDevsW(&GUID_DEVCLASS_DISPLAY) failed";
wchar_t regDirectxKey[] = L"SOFTWARE\\Microsoft\\DirectX\\" FF_EMPTY_GUID_STR;
const uint32_t regDirectxKeyPrefixLength = (uint32_t) strlen("SOFTWARE\\Microsoft\\DirectX\\");
wchar_t regControlVideoKey[] = L"SYSTEM\\CurrentControlSet\\Control\\Video\\" FF_EMPTY_GUID_STR L"\\0000";
const uint32_t regControlVideoKeyPrefixLength = (uint32_t) strlen("SYSTEM\\CurrentControlSet\\Control\\Video\\");
SP_DEVINFO_DATA did = { .cbSize = sizeof(did) };
for (DWORD idev = 0; SetupDiEnumDeviceInfo(hdev, idev, &did); ++idev)
{
if (displayDevice.StateFlags & DISPLAY_DEVICE_MIRRORING_DRIVER) continue;
wchar_t hardwareId[256];
if (!SetupDiGetDeviceRegistryPropertyW(hdev, &did, SPDRP_HARDWAREID, NULL, (PBYTE) hardwareId, sizeof(hardwareId), NULL))
continue;
const uint32_t deviceKeyLength = (uint32_t) wcslen(displayDevice.DeviceKey);
if (__builtin_expect(deviceKeyLength == 100, true))
wchar_t videoId[39];
{
if (wmemcmp(&displayDevice.DeviceKey[deviceKeyLength - 4], L"0000", 4) != 0) continue;
}
else
{
// DeviceKey can be empty. See #484
FF_STRBUF_AUTO_DESTROY gpuName = ffStrbufCreateWS(displayDevice.DeviceString);
if (ffListContains(gpus, &gpuName, (void*) isGpuNameEqual)) continue;
FF_HKEY_AUTO_DESTROY hKey = SetupDiOpenDevRegKey(hdev, &did, DICS_FLAG_GLOBAL, 0, DIREG_DEV, KEY_QUERY_VALUE);
if (!hKey) continue;
DWORD videoIdLength = sizeof(videoId);
if (RegGetValueW(hKey, NULL, L"VideoID", RRF_RT_REG_SZ, NULL, videoId, &videoIdLength) != ERROR_SUCCESS || videoIdLength != (FF_GUID_STRLEN + 1) * sizeof(wchar_t))
continue;
}
// 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"
wmemcpy(regControlVideoKey + regControlVideoKeyPrefixLength, videoId, FF_GUID_STRLEN);
FF_HKEY_AUTO_DESTROY hKey = NULL;
if (!ffRegOpenKeyForRead(HKEY_LOCAL_MACHINE, regControlVideoKey, &hKey, NULL)) continue;
unsigned vendorId = 0, deviceId = 0, subSystemId = 0, revId = 0;
swscanf(displayDevice.DeviceID, L"PCI\\VEN_%x&DEV_%x&SUBSYS_%x&REV_%x", &vendorId, &deviceId, &subSystemId, &revId);
swscanf(hardwareId, L"PCI\\VEN_%x&DEV_%x&SUBSYS_%x&REV_%x", &vendorId, &deviceId, &subSystemId, &revId);
FFGPUResult* gpu = (FFGPUResult*)ffListAdd(gpus);
ffStrbufInitStatic(&gpu->vendor, ffGetGPUVendorString(vendorId));
ffStrbufInitWS(&gpu->name, displayDevice.DeviceString);
ffStrbufInit(&gpu->vendor);
ffStrbufInit(&gpu->name);
ffStrbufInit(&gpu->driver);
ffStrbufInitStatic(&gpu->platformApi, "Direct3D");
gpu->temperature = FF_GPU_TEMP_UNSET;
@@ -55,56 +66,55 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
gpu->deviceId = 0;
gpu->frequency = FF_GPU_FREQUENCY_UNSET;
if (deviceKeyLength == 100 && displayDevice.DeviceKey[deviceKeyPrefixLength - 1] == '{')
if (!ffRegReadStrbuf(hKey, L"DriverDesc", &gpu->name, NULL)) // faster, for some reason
{
wmemcpy(regControlVideoKey + regControlVideoKeyPrefixLength, displayDevice.DeviceKey + deviceKeyPrefixLength, strlen("00000000-0000-0000-0000-000000000000}\\0000"));
FF_HKEY_AUTO_DESTROY hKey = NULL;
if (!ffRegOpenKeyForRead(HKEY_LOCAL_MACHINE, regControlVideoKey, &hKey, NULL)) continue;
wchar_t name[256];
if (SetupDiGetDeviceRegistryPropertyW(hdev, &did, SPDRP_DEVICEDESC, NULL, (PBYTE) name, sizeof(name), NULL))
ffStrbufSetWS(&gpu->name, name);
}
ffRegReadStrbuf(hKey, L"DriverVersion", &gpu->driver, NULL);
ffRegReadStrbuf(hKey, L"DriverVersion", &gpu->driver, NULL);
wmemcpy(regDirectxKey + regDirectxKeyPrefixLength, videoId, FF_GUID_STRLEN);
FF_HKEY_AUTO_DESTROY hDirectxKey = NULL;
if (ffRegOpenKeyForRead(HKEY_LOCAL_MACHINE, regDirectxKey, &hDirectxKey, NULL))
{
uint64_t dedicatedVideoMemory = 0;
if(ffRegReadUint64(hDirectxKey, L"DedicatedVideoMemory", &dedicatedVideoMemory, NULL))
gpu->type = dedicatedVideoMemory >= 1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
wmemcpy(regDirectxKey + regDirectxKeyPrefixLength, displayDevice.DeviceKey + deviceKeyPrefixLength, strlen("00000000-0000-0000-0000-000000000000}"));
FF_HKEY_AUTO_DESTROY hDirectxKey = NULL;
if (ffRegOpenKeyForRead(HKEY_LOCAL_MACHINE, regDirectxKey, &hDirectxKey, NULL))
uint64_t dedicatedSystemMemory, sharedSystemMemory;
if(ffRegReadUint64(hDirectxKey, L"DedicatedSystemMemory", &dedicatedSystemMemory, NULL) &&
ffRegReadUint64(hDirectxKey, L"SharedSystemMemory", &sharedSystemMemory, NULL))
{
uint64_t dedicatedVideoMemory = 0;
if(ffRegReadUint64(hDirectxKey, L"DedicatedVideoMemory", &dedicatedVideoMemory, NULL))
gpu->type = dedicatedVideoMemory >= 1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
uint64_t dedicatedSystemMemory, sharedSystemMemory;
if(ffRegReadUint64(hDirectxKey, L"DedicatedSystemMemory", &dedicatedSystemMemory, NULL) &&
ffRegReadUint64(hDirectxKey, L"SharedSystemMemory", &sharedSystemMemory, NULL))
{
gpu->dedicated.total = dedicatedVideoMemory + dedicatedSystemMemory;
gpu->shared.total = sharedSystemMemory;
}
ffRegReadUint64(hDirectxKey, L"AdapterLuid", &gpu->deviceId, NULL);
uint32_t featureLevel = 0;
if(ffRegReadUint(hDirectxKey, L"MaxD3D12FeatureLevel", &featureLevel, NULL) && featureLevel)
ffStrbufSetF(&gpu->platformApi, "Direct3D 12.%u", (featureLevel & 0x0F00) >> 8);
else if(ffRegReadUint(hDirectxKey, L"MaxD3D11FeatureLevel", &featureLevel, NULL) && featureLevel)
ffStrbufSetF(&gpu->platformApi, "Direct3D 11.%u", (featureLevel & 0x0F00) >> 8);
}
else if (!ffRegReadUint64(hKey, L"HardwareInformation.qwMemorySize", &gpu->dedicated.total, NULL))
{
uint32_t vmem = 0;
if (ffRegReadUint(hKey, L"HardwareInformation.MemorySize", &vmem, NULL))
gpu->dedicated.total = vmem;
gpu->type = gpu->dedicated.total > 1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
gpu->dedicated.total = dedicatedVideoMemory + dedicatedSystemMemory;
gpu->shared.total = sharedSystemMemory;
}
if (gpu->vendor.length == 0)
{
ffRegReadStrbuf(hKey, L"ProviderName", &gpu->vendor, NULL);
if (ffStrbufContainS(&gpu->vendor, "Intel"))
ffStrbufSetStatic(&gpu->vendor, FF_GPU_VENDOR_NAME_INTEL);
else if (ffStrbufContainS(&gpu->vendor, "NVIDIA"))
ffStrbufSetStatic(&gpu->vendor, FF_GPU_VENDOR_NAME_NVIDIA);
else if (ffStrbufContainS(&gpu->vendor, "AMD") || ffStrbufContainS(&gpu->vendor, "ATI"))
ffStrbufSetStatic(&gpu->vendor, FF_GPU_VENDOR_NAME_AMD);
}
ffRegReadUint64(hDirectxKey, L"AdapterLuid", &gpu->deviceId, NULL);
uint32_t featureLevel = 0;
if(ffRegReadUint(hDirectxKey, L"MaxD3D12FeatureLevel", &featureLevel, NULL) && featureLevel)
ffStrbufSetF(&gpu->platformApi, "Direct3D 12.%u", (featureLevel & 0x0F00) >> 8);
else if(ffRegReadUint(hDirectxKey, L"MaxD3D11FeatureLevel", &featureLevel, NULL) && featureLevel)
ffStrbufSetF(&gpu->platformApi, "Direct3D 11.%u", (featureLevel & 0x0F00) >> 8);
}
else if (!ffRegReadUint64(hKey, L"HardwareInformation.qwMemorySize", &gpu->dedicated.total, NULL))
{
uint32_t vmem = 0;
if (ffRegReadUint(hKey, L"HardwareInformation.MemorySize", &vmem, NULL))
gpu->dedicated.total = vmem;
gpu->type = gpu->dedicated.total > 1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
}
if (gpu->vendor.length == 0)
{
ffRegReadStrbuf(hKey, L"ProviderName", &gpu->vendor, NULL);
if (ffStrbufContainS(&gpu->vendor, "Intel"))
ffStrbufSetStatic(&gpu->vendor, FF_GPU_VENDOR_NAME_INTEL);
else if (ffStrbufContainS(&gpu->vendor, "NVIDIA"))
ffStrbufSetStatic(&gpu->vendor, FF_GPU_VENDOR_NAME_NVIDIA);
else if (ffStrbufContainS(&gpu->vendor, "AMD") || ffStrbufContainS(&gpu->vendor, "ATI"))
ffStrbufSetStatic(&gpu->vendor, FF_GPU_VENDOR_NAME_AMD);
}
__typeof__(&ffDetectNvidiaGpuInfo) detectFn;
@@ -112,43 +122,18 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
if (getDriverSpecificDetectionFn(gpu->vendor.chars, &detectFn, &dllName) && (options->temp || options->driverSpecific))
{
if (vendorId && deviceId)
uint32_t pciBus, pciAddr;
if (SetupDiGetDeviceRegistryPropertyW(hdev, &did, SPDRP_BUSNUMBER, NULL, (PBYTE) &pciBus, sizeof(pciBus), NULL) &&
SetupDiGetDeviceRegistryPropertyW(hdev, &did, SPDRP_ADDRESS, NULL, (PBYTE) &pciAddr, sizeof(pciAddr), NULL))
{
int bus = -1, dev = -1, func = -1;
if (detectFn == ffDetectNvidiaGpuInfo || detectFn == ffDetectMthreadsGpuInfo)
{
// Find PCI id from video id
// HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Enum\PCI\DEVICE_ID\INSTANCE_ID\Device Parameters\VideoID
wcscpy(regPciKey + regPciKeyPrefixLength, displayDevice.DeviceID);
FF_HKEY_AUTO_DESTROY hKey = NULL;
if (!ffRegOpenKeyForRead(HKEY_LOCAL_MACHINE, regPciKey, &hKey, NULL)) continue;
for (uint32_t idx = 0; ; ++idx)
{
wchar_t instanceKey[256];
DWORD bufSizeInstanceKey = sizeof(instanceKey) / sizeof(*instanceKey);
if (RegEnumKeyExW(hKey, idx, instanceKey, &bufSizeInstanceKey, NULL, NULL, NULL, NULL) != ERROR_SUCCESS) break;
wcscpy(instanceKey + bufSizeInstanceKey, L"\\Device Parameters");
wchar_t videoId[256];
DWORD bufSizeVideoId = sizeof(videoId) / sizeof(*videoId);
if (RegGetValueW(hKey, instanceKey, L"VideoID", RRF_RT_REG_SZ, NULL, videoId, &bufSizeVideoId) != ERROR_SUCCESS) continue;
if (videoId[0] != L'{') continue;
if (wmemcmp(videoId + 1 /* Ignore {} */, displayDevice.DeviceKey + deviceKeyPrefixLength, strlen("00000000-0000-0000-0000-000000000000")) != 0) continue;
// We finally find it
instanceKey[bufSizeInstanceKey] = L'\0';
bufSizeVideoId = sizeof(videoId) / sizeof(*videoId);
if (RegGetValueW(hKey, instanceKey, L"LocationInformation", RRF_RT_REG_SZ, NULL, videoId, &bufSizeVideoId) != ERROR_SUCCESS) break;
swscanf(videoId, L"%*[^(](%d,%d,%d)", &bus, &dev, &func);
break;
}
}
uint32_t pciDev = (pciAddr >> 16) & 0xFFFF;
uint32_t pciFunc = pciAddr & 0xFFFF;
detectFn(
&(FFGpuDriverCondition) {
.type = FF_GPU_DRIVER_CONDITION_TYPE_DEVICE_ID
| FF_GPU_DRIVER_CONDITION_TYPE_LUID
| (bus >= 0 ? FF_GPU_DRIVER_CONDITION_TYPE_BUS_ID : 0),
| (gpu->deviceId > 0 ? FF_GPU_DRIVER_CONDITION_TYPE_LUID : 0)
| FF_GPU_DRIVER_CONDITION_TYPE_BUS_ID,
.pciDeviceId = {
.deviceId = deviceId,
.vendorId = vendorId,
@@ -157,9 +142,9 @@ const char* ffDetectGPUImpl(FF_MAYBE_UNUSED const FFGPUOptions* options, FFlist*
},
.pciBusId = {
.domain = 0,
.bus = (uint32_t) bus,
.device = (uint32_t) dev,
.func = (uint32_t) func,
.bus = pciBus,
.device = pciDev,
.func = pciFunc,
},
.luid = gpu->deviceId,
},