update to Vc 0.7.3-dev

[u/mrichter/AliRoot.git] / Vc / src / support.cpp

diff --git a/Vc/src/support.cpp b/Vc/src/support.cpp
index c9aa20cc1f09be0fd8badd7c999f9d9893eb74e2..4d81664325f9f38037b873888c8a3484d5833312 100644 (file)
--- a/Vc/src/support.cpp
+++ b/Vc/src/support.cpp
@@ -19,22 +19,42 @@
 
 #include <Vc/global.h>
 #include <Vc/cpuid.h>
-#include "Vc/common/support.h"
+#include <Vc/support.h>
 
 #ifdef VC_MSVC
 #include <intrin.h>
 #endif
 
+#if defined(VC_GCC) && VC_GCC >= 0x40400
+#define VC_TARGET_NO_SIMD __attribute__((target("no-sse2,no-avx")))
+#else
+#define VC_TARGET_NO_SIMD
+#endif
+
+namespace AliRoot {
 namespace Vc
 {
 
-#ifdef VC_GCC
-    __attribute__((target("no-sse2,no-avx")))
+VC_TARGET_NO_SIMD
+static inline bool xgetbvCheck(unsigned int bits)
+{
+#if defined(VC_MSVC) && VC_MSVC >= 160040219 // MSVC 2010 SP1 introduced _xgetbv
+    unsigned long long xcrFeatureMask = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
+    return (xcrFeatureMask & bits) == bits;
+#elif defined(VC_GNU_ASM) && !defined(VC_NO_XGETBV)
+    unsigned int eax;
+    asm("xgetbv" : "=a"(eax) : "c"(0) : "edx");
+    return (eax & bits) == bits;
+#else
+    // can't check, but if OSXSAVE is true let's assume it'll work
+    return bits > 0; // ignore 'warning: unused parameter'
 #endif
+}
+
+VC_TARGET_NO_SIMD
 bool isImplementationSupported(Implementation impl)
 {
     CpuId::init();
-    // for AVX we need to check for OSXSAVE and AVX
 
     switch (impl) {
     case ScalarImpl:
@@ -49,28 +69,51 @@ bool isImplementationSupported(Implementation impl)
         return CpuId::hasSse41();
     case SSE42Impl:
         return CpuId::hasSse42();
-    case SSE4aImpl:
-        return CpuId::hasSse4a();
-    case XopImpl:
-        return CpuId::hasXop();
-    case Fma4Impl:
-        return CpuId::hasFma4();
     case AVXImpl:
-#if defined(VC_MSVC) && VC_MSVC >= 160040219 // MSVC 2010 SP1 introduced _xgetbv
-        unsigned long long xcrFeatureMask = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
-        return (xcrFeatureMask & 0x6) != 0;
-#elif !defined(VC_NO_XGETBV)
-        if (CpuId::hasOsxsave() && CpuId::hasAvx()) {
-            unsigned int eax;
-            asm("xgetbv" : "=a"(eax) : "c"(0) : "edx");
-            return (eax & 0x06) == 0x06;
-        }
-#endif
+        return CpuId::hasOsxsave() && CpuId::hasAvx() && xgetbvCheck(0x6);
+    case AVX2Impl:
+        return false;
+    case ImplementationMask:
         return false;
     }
     return false;
 }
 
+VC_TARGET_NO_SIMD
+Vc::Implementation bestImplementationSupported()
+{
+    CpuId::init();
+
+    if (!CpuId::hasSse2 ()) return Vc::ScalarImpl;
+    if (!CpuId::hasSse3 ()) return Vc::SSE2Impl;
+    if (!CpuId::hasSsse3()) return Vc::SSE3Impl;
+    if (!CpuId::hasSse41()) return Vc::SSSE3Impl;
+    if (!CpuId::hasSse42()) return Vc::SSE41Impl;
+    if (CpuId::hasAvx() && CpuId::hasOsxsave() && xgetbvCheck(0x6)) {
+        return Vc::AVXImpl;
+    }
+    return Vc::SSE42Impl;
+}
+
+VC_TARGET_NO_SIMD
+unsigned int extraInstructionsSupported()
+{
+    unsigned int flags = 0;
+    if (CpuId::hasF16c()) flags |= Vc::Float16cInstructions;
+    if (CpuId::hasFma4()) flags |= Vc::Fma4Instructions;
+    if (CpuId::hasXop ()) flags |= Vc::XopInstructions;
+    if (CpuId::hasPopcnt()) flags |= Vc::PopcntInstructions;
+    if (CpuId::hasSse4a()) flags |= Vc::Sse4aInstructions;
+    if (CpuId::hasFma ()) flags |= Vc::FmaInstructions;
+    //if (CpuId::hasPclmulqdq()) flags |= Vc::PclmulqdqInstructions;
+    //if (CpuId::hasAes()) flags |= Vc::AesInstructions;
+    //if (CpuId::hasRdrand()) flags |= Vc::RdrandInstructions;
+    return flags;
+}
+
 } // namespace Vc
+} // namespace AliRoot
+
+#undef VC_TARGET_NO_SIMD
 
 // vim: sw=4 sts=4 et tw=100