GPU (Linux): improve detection performance of Apple Silicon

This commit is contained in:
Carter Li
2024-04-18 16:41:19 +08:00
committed by 李通洲
parent 41df915f29
commit f8cbe4c008
4 changed files with 192 additions and 118 deletions
+21
View File
@@ -18,3 +18,24 @@ const char* ffDetectCPU(const FFCPUOptions* options, FFCPUResult* cpu)
ffStrbufTrimRight(&cpu->name, ' '); //If we removed the @ in previous step there was most likely a space before it
return NULL;
}
const char* ffCPUAppleCodeToName(uint32_t code)
{
// https://github.com/AsahiLinux/docs/wiki/Codenames
switch (code)
{
case 8103: return "Apple M1";
case 6000: return "Apple M1 Pro";
case 6001: return "Apple M1 Max";
case 6002: return "Apple M1 Ultra";
case 8112: return "Apple M2";
case 6020: return "Apple M2 Pro";
case 6021: return "Apple M2 Max";
case 6022: return "Apple M2 Ultra";
case 8122: return "Apple M3";
case 6030: return "Apple M3 Pro";
case 6031:
case 6034: return "Apple M3 Max";
default: return "Apple Silicon";
}
}
+1
View File
@@ -22,3 +22,4 @@ typedef struct FFCPUResult
const char* ffCPUDetectByCpuid(FFCPUResult* cpu);
const char* ffDetectCPU(const FFCPUOptions* options, FFCPUResult* cpu);
const char* ffCPUAppleCodeToName(uint32_t code);
+2 -17
View File
@@ -210,23 +210,8 @@ void detectAsahi(FFCPUResult* cpu)
char* modelName = memchr(content, '\0', (size_t) length) + 1;
if (modelName - content < length && ffStrStartsWith(modelName, "apple,t"))
{
// https://github.com/AsahiLinux/docs/wiki/Codenames
switch (strtoul(modelName + strlen("apple,t"), NULL, 10))
{
case 8103: ffStrbufSetStatic(&cpu->name, "Apple M1"); break;
case 6000: ffStrbufSetStatic(&cpu->name, "Apple M1 Pro"); break;
case 6001: ffStrbufSetStatic(&cpu->name, "Apple M1 Max"); break;
case 6002: ffStrbufSetStatic(&cpu->name, "Apple M1 Ultra"); break;
case 8112: ffStrbufSetStatic(&cpu->name, "Apple M2"); break;
case 6020: ffStrbufSetStatic(&cpu->name, "Apple M2 Pro"); break;
case 6021: ffStrbufSetStatic(&cpu->name, "Apple M2 Max"); break;
case 6022: ffStrbufSetStatic(&cpu->name, "Apple M2 Ultra"); break;
case 8122: ffStrbufSetStatic(&cpu->name, "Apple M3"); break;
case 6030: ffStrbufSetStatic(&cpu->name, "Apple M3 Pro"); break;
case 6031:
case 6034: ffStrbufSetStatic(&cpu->name, "Apple M3 Max"); break;
default: ffStrbufSetStatic(&cpu->name, "Apple Silicon"); break;
}
uint32_t deviceId = (uint32_t) strtoul(modelName + strlen("apple,t"), NULL, 10);
ffStrbufSetStatic(&cpu->name, ffCPUAppleCodeToName(deviceId));
}
}
}
+168 -101
View File
@@ -1,6 +1,7 @@
#include "detection/gpu/gpu.h"
#include "detection/vulkan/vulkan.h"
#include "detection/temps/temps_linux.h"
#include "detection/cpu/cpu.h"
#include "common/io/io.h"
#include "common/properties.h"
#include "util/stringUtils.h"
@@ -57,6 +58,7 @@ static void pciDetectDriver(FFGPUResult* gpu, FFstrbuf* pciDir, FFstrbuf* buffer
static void pciDetectVmem(FFGPUResult* gpu, FFstrbuf* pciDir, FFstrbuf* buffer)
{
// Works for AMD GPUs
// https://www.kernel.org/doc/html/v5.10/gpu/amdgpu.html#mem-info-vis-vram-total
ffStrbufAppendS(pciDir, "/mem_info_vis_vram_total");
uint64_t size = 0;
@@ -80,6 +82,21 @@ static void pciDetectVmem(FFGPUResult* gpu, FFstrbuf* pciDir, FFstrbuf* buffer)
}
}
static void pciDetectVfreq(FFGPUResult* gpu, FFstrbuf* pciDir, FFstrbuf* buffer)
{
// Works for Intel GPUs
// https://patchwork.kernel.org/project/intel-gfx/patch/1422039866-11572-3-git-send-email-ville.syrjala@linux.intel.com/
ffStrbufSetNS(buffer, ffStrbufLastIndexC(pciDir, '/'), pciDir->chars);
ffStrbufAppendS(buffer, "/gt_cur_freq_mhz");
char str[16];
ssize_t len = ffReadFileData(buffer->chars, sizeof(str) - 1, str);
if (len > 1)
{
str[len] = '\0';
gpu->frequency = (double) strtoul(str, NULL, 10) / 1000.0;
}
}
static bool loadPciIds(FFstrbuf* pciids)
{
#ifdef FF_CUSTOM_PCI_IDS_PATH
@@ -99,121 +116,171 @@ static bool loadPciIds(FFstrbuf* pciids)
return false;
}
static const char* pciDetectGPUs(const FFGPUOptions* options, FFlist* gpus)
static const char* detectPci(const FFGPUOptions* options, FFlist* gpus, FFstrbuf* buffer, FFstrbuf* drmDir)
{
//https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-bus-pci
const char* pciDirPath = "/sys/bus/pci/devices/";
const uint32_t drmDirPathLength = drmDir->length;
uint32_t vendorId, deviceId, subVendorId, subDeviceId;
uint8_t classId, subclassId;
if (sscanf(buffer->chars + strlen("pci:"), "v%8" SCNx32 "d%8" SCNx32 "sv%8" SCNx32 "sd%8" SCNx32 "bc%2" SCNx8 "sc%2" SCNx8, &vendorId, &deviceId, &subVendorId, &subDeviceId, &classId, &subclassId) != 6)
return "Invalid modalias string";
FF_AUTO_CLOSE_DIR DIR* dirp = opendir(pciDirPath);
if(dirp == NULL)
return "Failed to open `/sys/bus/pci/devices/`";
if (classId != 0x03 /*PCI_BASE_CLASS_DISPLAY*/)
return "Should not happen";
FF_STRBUF_AUTO_DESTROY pciDir = ffStrbufCreateA(64);
ffStrbufAppendS(&pciDir, pciDirPath);
char pciPath[PATH_MAX];
ssize_t pathLength = readlink(drmDir->chars, pciPath, sizeof(pciPath) - 1);
if (pathLength <= 0)
return "Unable to get PCI device path";
pciPath[pathLength] = '\0';
const char* pPciPath = strrchr(pciPath, '/');
if (pPciPath)
pPciPath++;
else
pPciPath = pciPath;
const uint32_t pciBaseDirLength = pciDir.length;
uint32_t pciDomain, pciBus, pciDevice, pciFunc;
if (sscanf(pPciPath, "%" SCNx32 ":%" SCNx32 ":%" SCNx32 ".%" SCNx32, &pciDomain, &pciBus, &pciDevice, &pciFunc) != 4)
return "Invalid PCI device path";
FFGPUResult* gpu = (FFGPUResult*)ffListAdd(gpus);
ffStrbufInitStatic(&gpu->vendor, ffGetGPUVendorString((uint16_t) vendorId));
ffStrbufInit(&gpu->name);
ffStrbufInit(&gpu->driver);
ffStrbufInit(&gpu->platformApi);
gpu->temperature = FF_GPU_TEMP_UNSET;
gpu->coreCount = FF_GPU_CORE_COUNT_UNSET;
gpu->type = FF_GPU_TYPE_UNKNOWN;
gpu->dedicated.total = gpu->dedicated.used = gpu->shared.total = gpu->shared.used = FF_GPU_VMEM_SIZE_UNSET;
gpu->deviceId = ((uint64_t) pciDomain << 6) | ((uint64_t) pciBus << 4) | (deviceId << 2) | pciFunc;
gpu->frequency = FF_GPU_FREQUENCY_UNSET;
if (gpu->vendor.chars == FF_GPU_VENDOR_NAME_AMD)
{
ffStrbufAppendS(drmDir, "/revision");
if (ffReadFileBuffer(drmDir->chars, buffer))
{
char* pend;
uint64_t revision = strtoul(buffer->chars, &pend, 16);
if (pend != buffer->chars)
{
char query[32];
snprintf(query, sizeof(query), "%X,\t%X,", (unsigned) deviceId, (unsigned) revision);
ffParsePropFileData("libdrm/amdgpu.ids", query, &gpu->name);
}
}
ffStrbufSubstrBefore(drmDir, drmDirPathLength);
}
if (gpu->name.length == 0)
{
static FFstrbuf pciids;
if (pciids.chars == NULL)
{
ffStrbufInit(&pciids);
loadPciIds(&pciids);
}
if (pciids.length)
ffGPUParsePciIds(&pciids, subclassId, (uint16_t) vendorId, (uint16_t) deviceId, gpu);
}
pciDetectVmem(gpu, drmDir, buffer);
ffStrbufSubstrBefore(drmDir, drmDirPathLength);
pciDetectVfreq(gpu, drmDir, buffer);
ffStrbufSubstrBefore(drmDir, drmDirPathLength);
pciDetectDriver(gpu, drmDir, buffer);
ffStrbufSubstrBefore(drmDir, drmDirPathLength);
#ifdef FF_USE_PROPRIETARY_GPU_DRIVER_API
if (gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA && (options->temp || options->driverSpecific))
{
ffDetectNvidiaGpuInfo(&(FFGpuDriverCondition) {
.type = FF_GPU_DRIVER_CONDITION_TYPE_BUS_ID,
.pciBusId = {
.domain = pciDomain,
.bus = pciBus,
.device = pciDevice,
.func = pciFunc,
},
}, (FFGpuDriverResult) {
.temp = options->temp ? &gpu->temperature : NULL,
.memory = options->driverSpecific ? &gpu->dedicated : NULL,
.coreCount = options->driverSpecific ? (uint32_t*) &gpu->coreCount : NULL,
.type = &gpu->type,
.frequency = &gpu->frequency,
}, "libnvidia-ml.so");
if (gpu->dedicated.total != FF_GPU_VMEM_SIZE_UNSET)
gpu->type = gpu->dedicated.total > (uint64_t)1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
}
#endif // FF_USE_PROPRIETARY_GPU_DRIVER_API
#ifdef __linux__
if(options->temp && gpu->temperature != gpu->temperature)
pciDetectTemp(gpu, ((uint32_t) classId << 8) + subclassId);
#endif
return NULL;
}
FF_MAYBE_UNUSED static const char* detectAsahi(FFlist* gpus, FFstrbuf* buffer, FFstrbuf* drmDir)
{
uint32_t index = ffStrbufFirstIndexS(buffer, "apple,agx-t");
if (index == buffer->length) return "display-subsystem?";
index += strlen("apple,agx-t");
FFGPUResult* gpu = (FFGPUResult*)ffListAdd(gpus);
gpu->deviceId = strtoul(buffer->chars + index, NULL, 10);
ffStrbufInitStatic(&gpu->name, ffCPUAppleCodeToName(gpu->deviceId));
ffStrbufInitStatic(&gpu->vendor, FF_GPU_VENDOR_NAME_APPLE);
ffStrbufInit(&gpu->driver);
ffStrbufInit(&gpu->platformApi);
gpu->temperature = FF_GPU_TEMP_UNSET;
gpu->coreCount = FF_GPU_CORE_COUNT_UNSET;
gpu->type = FF_GPU_TYPE_INTEGRATED;
gpu->dedicated.total = gpu->dedicated.used = gpu->shared.total = gpu->shared.used = FF_GPU_VMEM_SIZE_UNSET;
gpu->frequency = FF_GPU_FREQUENCY_UNSET;
pciDetectDriver(gpu, drmDir, buffer);
return NULL;
}
static const char* drmDetectGPUs(const FFGPUOptions* options, FFlist* gpus)
{
FF_STRBUF_AUTO_DESTROY drmDir = ffStrbufCreateA(64);
ffStrbufAppendS(&drmDir, "/sys/class/drm/");
const uint32_t drmDirLength = drmDir.length;
FF_AUTO_CLOSE_DIR DIR* dir = opendir(drmDir.chars);
if(dir == NULL)
return "/sys/class/drm doesn't exist";
FF_STRBUF_AUTO_DESTROY buffer = ffStrbufCreate();
FF_STRBUF_AUTO_DESTROY pciids = ffStrbufCreate();
struct dirent* entry;
while((entry = readdir(dirp)) != NULL)
while ((entry = readdir(dir)) != NULL)
{
if(entry->d_name[0] == '.')
if (!ffStrStartsWith(entry->d_name, "card") ||
strchr(entry->d_name + 4, '-') != NULL)
continue;
ffStrbufSubstrBefore(&pciDir, pciBaseDirLength);
ffStrbufAppendS(&pciDir, entry->d_name);
ffStrbufAppendS(&drmDir, entry->d_name);
const uint32_t pciDevDirLength = pciDir.length;
ffStrbufAppendS(&pciDir, "/modalias");
if (!ffReadFileBuffer(pciDir.chars, &buffer))
ffStrbufAppendS(&drmDir, "/device/modalias");
if (!ffReadFileBuffer(drmDir.chars, &buffer))
continue;
ffStrbufSubstrBefore(&pciDir, pciDevDirLength);
ffStrbufSubstrBefore(&drmDir, drmDir.length - (uint32_t) strlen("/modalias"));
uint32_t vendorId, deviceId, subVendorId, subDeviceId;
uint8_t classId, subclassId;
if (sscanf(buffer.chars, "pci:v%8" SCNx32 "d%8" SCNx32 "sv%8" SCNx32 "sd%8" SCNx32 "bc%2" SCNx8 "sc%2" SCNx8, &vendorId, &deviceId, &subVendorId, &subDeviceId, &classId, &subclassId) != 6)
continue;
if (classId != 0x03 /*PCI_BASE_CLASS_DISPLAY*/)
continue;
uint32_t pciDomain, pciBus, pciDevice, pciFunc;
if (sscanf(entry->d_name, "%" SCNx32 ":%" SCNx32 ":%" SCNx32 ".%" SCNx32, &pciDomain, &pciBus, &pciDevice, &pciFunc) != 4)
continue;
FFGPUResult* gpu = (FFGPUResult*)ffListAdd(gpus);
ffStrbufInitStatic(&gpu->vendor, ffGetGPUVendorString((uint16_t) vendorId));
ffStrbufInit(&gpu->name);
ffStrbufInit(&gpu->driver);
ffStrbufInit(&gpu->platformApi);
gpu->temperature = FF_GPU_TEMP_UNSET;
gpu->coreCount = FF_GPU_CORE_COUNT_UNSET;
gpu->type = FF_GPU_TYPE_UNKNOWN;
gpu->dedicated.total = gpu->dedicated.used = gpu->shared.total = gpu->shared.used = FF_GPU_VMEM_SIZE_UNSET;
gpu->deviceId = ((uint64_t) pciDomain << 6) | ((uint64_t) pciBus << 4) | (deviceId << 2) | pciFunc;
gpu->frequency = FF_GPU_FREQUENCY_UNSET;
if (gpu->vendor.chars == FF_GPU_VENDOR_NAME_AMD)
{
ffStrbufAppendS(&pciDir, "/revision");
if (ffReadFileBuffer(pciDir.chars, &buffer))
{
char* pend;
uint64_t revision = strtoul(buffer.chars, &pend, 16);
if (pend != buffer.chars)
{
char query[32];
snprintf(query, sizeof(query), "%X,\t%X,", (unsigned) deviceId, (unsigned) revision);
ffParsePropFileData("libdrm/amdgpu.ids", query, &gpu->name);
}
}
ffStrbufSubstrBefore(&pciDir, pciDevDirLength);
}
if (gpu->name.length == 0)
{
if (!pciids.length)
loadPciIds(&pciids);
ffGPUParsePciIds(&pciids, subclassId, (uint16_t) vendorId, (uint16_t) deviceId, gpu);
}
pciDetectVmem(gpu, &pciDir, &buffer);
ffStrbufSubstrBefore(&pciDir, pciDevDirLength);
pciDetectDriver(gpu, &pciDir, &buffer);
ffStrbufSubstrBefore(&pciDir, pciDevDirLength);
#ifdef FF_USE_PROPRIETARY_GPU_DRIVER_API
if (gpu->vendor.chars == FF_GPU_VENDOR_NAME_NVIDIA && (options->temp || options->driverSpecific))
{
ffDetectNvidiaGpuInfo(&(FFGpuDriverCondition) {
.type = FF_GPU_DRIVER_CONDITION_TYPE_BUS_ID,
.pciBusId = {
.domain = pciDomain,
.bus = pciBus,
.device = pciDevice,
.func = pciFunc,
},
}, (FFGpuDriverResult) {
.temp = options->temp ? &gpu->temperature : NULL,
.memory = options->driverSpecific ? &gpu->dedicated : NULL,
.coreCount = options->driverSpecific ? (uint32_t*) &gpu->coreCount : NULL,
.type = &gpu->type,
.frequency = &gpu->frequency,
}, "libnvidia-ml.so");
if (gpu->dedicated.total != FF_GPU_VMEM_SIZE_UNSET)
gpu->type = gpu->dedicated.total > (uint64_t)1024 * 1024 * 1024 ? FF_GPU_TYPE_DISCRETE : FF_GPU_TYPE_INTEGRATED;
}
#endif // FF_USE_PROPRIETARY_GPU_DRIVER_API
#ifdef __linux__
if(options->temp && gpu->temperature != gpu->temperature)
pciDetectTemp(gpu, ((uint32_t) classId << 8) + subclassId);
if (ffStrbufStartsWithS(&buffer, "pci:"))
detectPci(options, gpus, &buffer, &drmDir);
#ifdef __aarch64__
else if (ffStrbufStartsWithS(&buffer, "of:"))
detectAsahi(gpus, &buffer, &drmDir);
#endif
ffStrbufSubstrBefore(&drmDir, drmDirLength);
}
return NULL;
@@ -227,5 +294,5 @@ const char* ffDetectGPUImpl(const FFGPUOptions* options, FFlist* gpus)
return NULL;
#endif
return pciDetectGPUs(options, gpus);
return drmDetectGPUs(options, gpus);
}