From 63a0faa1d92837537c633fde523a837baab07e2a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E6=9D=8E=E9=80=9A=E6=B4=B2?= <zhangsongcui3371@sina.com>
Date: Wed, 17 Jul 2024 15:51:15 +0800
Subject: [PATCH] GPU (macOS): support GPU frequency detection for Apple
 Silicon

Also refactor CPU freq detection code
---
 src/detection/cpu/cpu_apple.c | 19 +++++++++-------
 src/detection/gpu/gpu_apple.c | 42 +++++++++++++++++++++++++++++++++++
 2 files changed, 53 insertions(+), 8 deletions(-)

diff --git a/src/detection/cpu/cpu_apple.c b/src/detection/cpu/cpu_apple.c
index 20ef4a023..ad6852732 100644
--- a/src/detection/cpu/cpu_apple.c
+++ b/src/detection/cpu/cpu_apple.c
@@ -2,7 +2,6 @@
 #include "common/sysctl.h"
 #include "detection/temps/temps_apple.h"
 #include "util/stringUtils.h"
-#include "util/apple/cf_helpers.h"
 
 static double detectCpuTemp(const FFstrbuf* cpuName)
 {
@@ -44,15 +43,19 @@ static const char* detectFrequency(FFCPUResult* cpu)
     if (!IOObjectConformsTo(entryDevice, "AppleARMIODevice"))
         return "\"pmgr\" should conform to \"AppleARMIODevice\"";
 
-    FF_CFTYPE_AUTO_RELEASE CFDataRef pFreqProperty = (CFDataRef) IORegistryEntryCreateCFProperty(entryDevice, CFSTR("voltage-states5-sram"), kCFAllocatorDefault, kNilOptions);
-    if (CFGetTypeID(pFreqProperty) != CFDataGetTypeID())
+    FF_CFTYPE_AUTO_RELEASE CFDataRef freqProperty = (CFDataRef) IORegistryEntryCreateCFProperty(entryDevice, CFSTR("voltage-states5-sram"), kCFAllocatorDefault, kNilOptions);
+    if (CFGetTypeID(freqProperty) != CFDataGetTypeID())
         return "\"voltage-states5-sram\" in \"pmgr\" is not found";
 
-    // voltage-states5-sram stores supported frequencies of pcores from the lowest to the highest
-    CFIndex pCoreFreqLength = CFDataGetLength(pFreqProperty);
-    uint32_t* pStart = (uint32_t*) CFDataGetBytePtr(pFreqProperty);
-    uint32_t pMax = 0;
-    for (CFIndex i = 0; i < pCoreFreqLength && pStart[i] > 0; i += 4)
+    // voltage-states5-sram stores supported <frequency / voltage> pairs of pcores from the lowest to the highest
+    // voltage-states1-sram stores ecores'
+    CFIndex propLength = CFDataGetLength(freqProperty);
+    if (propLength == 0 || propLength % (CFIndex) sizeof(uint32_t) * 2 != 0)
+        return "Invalid \"voltage-states5-sram\" length";
+
+    uint32_t* pStart = (uint32_t*) CFDataGetBytePtr(freqProperty);
+    uint32_t pMax = *pStart;
+    for (CFIndex i = 2; i < propLength / (CFIndex) sizeof(uint32_t) && pStart[i] > 0; i += 2 /* skip voltage */)
         pMax = pMax > pStart[i] ? pMax : pStart[i];
 
     if (pMax > 0)
diff --git a/src/detection/gpu/gpu_apple.c b/src/detection/gpu/gpu_apple.c
index 6bc0ee48e..c63974d1f 100644
--- a/src/detection/gpu/gpu_apple.c
+++ b/src/detection/gpu/gpu_apple.c
@@ -36,6 +36,43 @@ static double detectGpuTemp(const FFstrbuf* gpuName)
     return result;
 }
 
+#ifdef __aarch64__
+#include "util/apple/cf_helpers.h"
+
+#include <IOKit/IOKitLib.h>
+
+static const char* detectFrequency(FFGPUResult* gpu)
+{
+    // https://github.com/giampaolo/psutil/pull/2222/files
+
+    FF_IOOBJECT_AUTO_RELEASE io_registry_entry_t entryDevice = IOServiceGetMatchingService(MACH_PORT_NULL, IOServiceNameMatching("pmgr"));
+    if (!entryDevice)
+        return "IOServiceGetMatchingServices() failed";
+
+    if (!IOObjectConformsTo(entryDevice, "AppleARMIODevice"))
+        return "\"pmgr\" should conform to \"AppleARMIODevice\"";
+
+    FF_CFTYPE_AUTO_RELEASE CFDataRef freqProperty = (CFDataRef) IORegistryEntryCreateCFProperty(entryDevice, CFSTR("voltage-states9-sram"), kCFAllocatorDefault, kNilOptions);
+    if (CFGetTypeID(freqProperty) != CFDataGetTypeID())
+        return "\"voltage-states9-sram\" in \"pmgr\" is not found";
+
+    // voltage-states5-sram stores supported <frequency / voltage> pairs of gpu from the lowest to the highest
+    CFIndex propLength = CFDataGetLength(freqProperty);
+    if (propLength == 0 || propLength % (CFIndex) sizeof(uint32_t) * 2 != 0)
+        return "Invalid \"voltage-states9-sram\" length";
+
+    uint32_t* pStart = (uint32_t*) CFDataGetBytePtr(freqProperty);
+    uint32_t pMax = *pStart;
+    for (CFIndex i = 2; i < propLength / (CFIndex) sizeof(uint32_t) && pStart[i] > 0; i += 2 /* skip voltage */)
+        pMax = pMax > pStart[i] ? pMax : pStart[i];
+
+    if (pMax > 0)
+        gpu->frequency = pMax / 1000 / 1000;
+
+    return NULL;
+}
+#endif
+
 const char* ffDetectGPUImpl(const FFGPUOptions* options, FFlist* gpus)
 {
     FF_IOOBJECT_AUTO_RELEASE io_iterator_t iterator = IO_OBJECT_NULL;
@@ -94,6 +131,11 @@ const char* ffDetectGPUImpl(const FFGPUOptions* options, FFlist* gpus)
                 gpu->type = FF_GPU_TYPE_INTEGRATED;
             else if (vendorStr == FF_GPU_VENDOR_NAME_NVIDIA || vendorStr == FF_GPU_VENDOR_NAME_AMD)
                 gpu->type = FF_GPU_TYPE_DISCRETE;
+
+            #ifdef __aarch64__
+            if (vendorStr == FF_GPU_VENDOR_NAME_APPLE)
+                detectFrequency(gpu);
+            #endif
         }
 
         gpu->temperature = options->temp ? detectGpuTemp(&gpu->name) : FF_GPU_TEMP_UNSET;