Merge pull request #6193 from secondlife/release/26.4

Release/26.4

1 files changed, 139 insertions, 19 deletions

diff --git a/indra/llcommon/llprocessor.cpp b/indra/llcommon/llprocessor.cpp
index 0778b123ea..456f2b02d4 100644
--- a/indra/llcommon/llprocessor.cpp
+++ b/indra/llcommon/llprocessor.cpp
@@ -123,6 +123,9 @@ namespace
         eSSE4_1_Features = 38,
         eSSE4_2_Features = 39,
         eSSE4a_Features = 40,
+        eAVX_Features = 41,
+        eAVX2_Features = 42,
+        eAVX512F_Features = 43,
     };
 
     const char* cpu_feature_names[] =
@@ -170,6 +173,9 @@ namespace
         "SSE4.1 Instructions",
         "SSE4.2 Instructions",
         "SSE4a Instructions",
+        "AVX Instructions",   // 41
+        "AVX2 Instructions",  // 42
+        "AVX-512F Instructions", // 43
     };
 
     std::string intel_CPUFamilyName(int composed_family)
@@ -283,6 +289,21 @@ public:
         return hasExtension(cpu_feature_names[eSSE4a_Features]);
     }
 
+    bool hasAVX() const
+    {
+        return hasExtension(cpu_feature_names[eAVX_Features]);
+    }
+
+    bool hasAVX2() const
+    {
+        return hasExtension(cpu_feature_names[eAVX2_Features]);
+    }
+
+    bool hasAVX512F() const
+    {
+        return hasExtension(cpu_feature_names[eAVX512F_Features]);
+    }
+
     bool hasAltivec() const
     {
         return hasExtension("Altivec");
@@ -572,6 +593,12 @@ private:
                     setExtension(cpu_feature_names[eSSE4_2_Features]);
                 }
 
+                // AVX: CPUID leaf 1, ECX bit 28, OSXSAVE + XCR0[1:2] enabled
+                if ((cpu_info[2] & 0x18000000) == 0x18000000 && ((_xgetbv(0) & 0x6) == 0x6))
+                {
+                    setExtension(cpu_feature_names[eAVX_Features]);
+                }
+
                 unsigned int feature_info = (unsigned int) cpu_info[3];
                 for(unsigned int index = 0, bit = 1; index < eSSE3_Features; ++index, bit <<= 1)
                 {
@@ -581,6 +608,27 @@ private:
                     }
                 }
             }
+            else if (i == 7)
+            {
+                // AVX2/AVX-512* require AVX to be usable (OSXSAVE + XCR0 state)
+                if (hasExtension(cpu_feature_names[eAVX_Features]))
+                {
+                    int cpu_info7[4] = { -1 };
+                    __cpuidex(cpu_info7, 7, 0);
+                    // AVX2: EBX bit 5
+                    if (cpu_info7[1] & 0x20)
+                    {
+                        setExtension(cpu_feature_names[eAVX2_Features]);
+                    }
+
+                    // AVX-512F: EBX bit 16; ZMM state in XCR0 (bits 5-7)
+                    const unsigned long long xcr0 = _xgetbv(0);
+                    if ((cpu_info7[1] & 0x10000) && ((xcr0 & 0xE6) == 0xE6))
+                    {
+                        setExtension(cpu_feature_names[eAVX512F_Features]);
+                    }
+                }
+            }
         }
 
         // Calling __cpuid with 0x80000000 as the InfoType argument
@@ -779,10 +827,47 @@ private:
             // Not supposed to happen?
             setExtension(cpu_feature_names[eSSE4a_Features]);
         }
+        if (cpu_features_str.find(" AVX1.0 ") != std::string::npos)
+        {
+            setExtension(cpu_feature_names[eAVX_Features]);
+        }
+
+        // AVX2 and AVX-512F are reported in machdep.cpu.leaf7_features on macOS
+        char cpu_leaf7_features[1024];
+        len = sizeof(cpu_leaf7_features);
+        memset(cpu_leaf7_features, 0, len);
+        sysctlbyname("machdep.cpu.leaf7_features", (void*)cpu_leaf7_features, &len, NULL, 0);
+
+        std::string cpu_leaf7_str(cpu_leaf7_features);
+        cpu_leaf7_str = " " + cpu_leaf7_str + " ";
+
+        if (cpu_leaf7_str.find(" AVX2 ") != std::string::npos)
+        {
+            setExtension(cpu_feature_names[eAVX2_Features]);
+        }
+
+        if (cpu_leaf7_str.find(" AVX512F ") != std::string::npos)
+        {
+            setExtension(cpu_feature_names[eAVX512F_Features]);
+        }
     }
 };
 
 #elif LL_LINUX
+
+// *NOTE:Mani - eww, macros! srry.
+#define LLPI_SET_INFO_STRING(llpi_id, cpuinfo_id) \
+        if (!cpuinfo[cpuinfo_id].empty()) \
+        { setInfo(llpi_id, cpuinfo[cpuinfo_id]);}
+
+#define LLPI_SET_INFO_INT(llpi_id, cpuinfo_id) \
+        {\
+            S32 result; \
+            if (!cpuinfo[cpuinfo_id].empty() \
+                && LLStringUtil::convertToS32(cpuinfo[cpuinfo_id], result)) \
+            { setInfo(llpi_id, result);} \
+        }
+
 const char CPUINFO_FILE[] = "/proc/cpuinfo";
 
 class LLProcessorInfoLinuxImpl : public LLProcessorInfoImpl
@@ -794,12 +879,36 @@ public:
     }
 
     virtual ~LLProcessorInfoLinuxImpl() {}
+
 private:
 
+    F64 getCPUMaxMHZ()
+    {
+        // Nicky: We just look into cpu0. In theory we could iterate over all cores
+        // "/sys/devices/system/cpu/cpu*/cpufreq/cpuinfo_max_freq"
+        // But those should not fluctuate that much?
+        llifstream fIn{ "/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq" };
+
+        if( !fIn.is_open() )
+            return 0.0;
+
+        std::string strLine;
+        fIn >> strLine;
+        if( strLine.empty() )
+            return 0.0l;
+
+        F64 mhz {};
+        if( !LLStringUtil::convertToF64(strLine, mhz ) )
+            return 0.0;
+
+        mhz = mhz / 1000.0;
+        return mhz;
+    }
+
     void get_proc_cpuinfo()
     {
         std::map< std::string, std::string > cpuinfo;
-        LLFILE* cpuinfo_fp = LLFile::fopen(CPUINFO_FILE, "rb");
+        LLFILE* cpuinfo_fp = LLFile::fopen(CPUINFO_FILE, LLFILE_MODE("rb"));
         if(cpuinfo_fp)
         {
             char line[MAX_STRING];
@@ -830,24 +939,17 @@ private:
         }
 # if LL_X86
 
-// *NOTE:Mani - eww, macros! srry.
-#define LLPI_SET_INFO_STRING(llpi_id, cpuinfo_id) \
-        if (!cpuinfo[cpuinfo_id].empty()) \
-        { setInfo(llpi_id, cpuinfo[cpuinfo_id]);}
-
-#define LLPI_SET_INFO_INT(llpi_id, cpuinfo_id) \
-        {\
-            S32 result; \
-            if (!cpuinfo[cpuinfo_id].empty() \
-                && LLStringUtil::convertToS32(cpuinfo[cpuinfo_id], result)) \
-            { setInfo(llpi_id, result);} \
+        F64 mhzFromSys = getCPUMaxMHZ();
+        F64 mhzFromProc {};
+        if( !LLStringUtil::convertToF64(cpuinfo["cpu mhz"], mhzFromProc ) )
+            mhzFromProc = 0.0;
+        if (mhzFromSys > 1.0 && mhzFromSys > mhzFromProc )
+        {
+            setInfo( eFrequency, mhzFromSys );
         }
-
-        F64 mhz;
-        if (LLStringUtil::convertToF64(cpuinfo["cpu mhz"], mhz)
-            && 200.0 < mhz && mhz < 10000.0)
+        else if (  200.0 < mhzFromProc && mhzFromProc < 10000.0)
         {
-            setInfo(eFrequency,(F64)(mhz));
+            setInfo(eFrequency,(F64)(mhzFromProc));
         }
 
         LLPI_SET_INFO_STRING(eBrandName, "model name");
@@ -857,7 +959,7 @@ private:
         LLPI_SET_INFO_INT(eModel, "model");
 
 
-        S32 family;
+        S32 family{};
         if (!cpuinfo["cpu family"].empty()
             && LLStringUtil::convertToS32(cpuinfo["cpu family"], family))
         {
@@ -915,6 +1017,21 @@ private:
             setExtension(cpu_feature_names[eSSE4a_Features]);
         }
 
+        if (flags.find(" avx ") != std::string::npos)
+        {
+            setExtension(cpu_feature_names[eAVX_Features]);
+        }
+
+        if (flags.find(" avx2 ") != std::string::npos)
+        {
+            setExtension(cpu_feature_names[eAVX2_Features]);
+        }
+
+        if (flags.find(" avx512f ") != std::string::npos)
+        {
+            setExtension(cpu_feature_names[eAVX512F_Features]);
+        }
+
 # endif // LL_X86
     }
 
@@ -923,7 +1040,7 @@ private:
         std::ostringstream s;
 
         // *NOTE:Mani - This is for linux only.
-        LLFILE* cpuinfo = LLFile::fopen(CPUINFO_FILE, "rb");
+        LLFILE* cpuinfo = LLFile::fopen(CPUINFO_FILE, LLFILE_MODE("rb"));
         if(cpuinfo)
         {
             char line[MAX_STRING];
@@ -979,6 +1096,9 @@ bool LLProcessorInfo::hasSSE3S() const { return mImpl->hasSSE3S(); }
 bool LLProcessorInfo::hasSSE41() const { return mImpl->hasSSE41(); }
 bool LLProcessorInfo::hasSSE42() const { return mImpl->hasSSE42(); }
 bool LLProcessorInfo::hasSSE4a() const { return mImpl->hasSSE4a(); }
+bool LLProcessorInfo::hasAVX() const { return mImpl->hasAVX(); }
+bool LLProcessorInfo::hasAVX2() const { return mImpl->hasAVX2(); }
+bool LLProcessorInfo::hasAVX512F() const { return mImpl->hasAVX512F(); }
 bool LLProcessorInfo::hasAltivec() const { return mImpl->hasAltivec(); }
 std::string LLProcessorInfo::getCPUFamilyName() const { return mImpl->getCPUFamilyName(); }
 std::string LLProcessorInfo::getCPUBrandName() const { return mImpl->getCPUBrandName(); }