Adding check for OS support to AVX

base/cpu.cc

 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "base/cpu.h"
 
 #include <string.h>
 
 #include <algorithm>
 
@@ skipping 14 matching lines @@
     stepping_(0),
     ext_model_(0),
     ext_family_(0),
     has_mmx_(false),
     has_sse_(false),
     has_sse2_(false),
     has_sse3_(false),
     has_ssse3_(false),
     has_sse41_(false),
     has_sse42_(false),
+    has_avx_(false),
     cpu_vendor_("unknown") {
   Initialize();
 }
 
 #if defined(ARCH_CPU_X86_FAMILY)
 #ifndef _MSC_VER
 
 #if defined(__pic__) && defined(__i386__)
 
 void __cpuid(int cpu_info[4], int info_type) {
@@ skipping 28 matching lines @@
 
 void __cpuidex(int cpu_info[4], int info_type, int info_index) {
   __asm__ volatile (
     "cpuid \n\t"
     : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
     : "a"(info_type), "c"(info_index)
   );
 }
 
 #endif
+
+unsigned long long _xgetbv(unsigned int xcr) {
+  unsigned int eax, edx;
+  __asm__ volatile (
+    ".byte 0x0f, 0x01, 0xd0"
+    : "=a"(eax), "=d"(edx)
+    : "c" (xcr)
+  );
+  return static_cast<unsigned long long>(eax) ^
+         static_cast<unsigned long long>(edx) << 32;
+}
 #endif  // _MSC_VER
 #endif  // ARCH_CPU_X86_FAMILY
 
 void CPU::Initialize() {
 #if defined(ARCH_CPU_X86_FAMILY)
   int cpu_info[4] = {-1};
   char cpu_string[48];
 
   // __cpuid with an InfoType argument of 0 returns the number of
   // valid Ids in CPUInfo[0] and the CPU identification string in
@@ skipping 17 matching lines @@
     type_ = (cpu_info[0] >> 12) & 0x3;
     ext_model_ = (cpu_info[0] >> 16) & 0xf;
     ext_family_ = (cpu_info[0] >> 20) & 0xff;
     has_mmx_ = (cpu_info[3] & 0x00800000) != 0;
     has_sse_ = (cpu_info[3] & 0x02000000) != 0;
     has_sse2_ = (cpu_info[3] & 0x04000000) != 0;
     has_sse3_ = (cpu_info[2] & 0x00000001) != 0;
     has_ssse3_ = (cpu_info[2] & 0x00000200) != 0;
     has_sse41_ = (cpu_info[2] & 0x00080000) != 0;
     has_sse42_ = (cpu_info[2] & 0x00100000) != 0;
-    has_avx_ = (cpu_info[2] & 0x10000000) != 0;
+    has_avx_ = (cpu_info[2] & 0x10000000) != 0 && (_xgetbv(0) & 6) == 6;

[DaleCurtis, 2013/03/07 23:37:15]

I think this needs #include <immintrin.h> on Windows and that only works on
64-bit according to this bug:

https://code.google.com/p/libyuv/issues/detail?id=106

Perhaps that's fixed now? Have you done a Windows build with this?

[apatrick_chromium, 2013/03/08 00:39:41]

This says that it need Visual Studio 2010 SP1, which is a Chromium build
prerequisite.

http://insufficientlycomplicated.wordpress.com/2011/11/07/detecting-intel-adv...

It also suggests the need to check for XSAVE/XRSTOR support, bit 27 of
cpu_info[2].

[whunt, 2013/03/08 18:30:32]

XSAVE/XRSTOR are required for SSE to work at all and are supported by all OSes
we support.

On 2013/03/08 00:39:41, apatrick_chromium wrote:

[whunt, 2013/03/08 18:30:32]

<intrin.h> should include <immintrin.h>
I've only run the bots with it which I believe are 32 bit (I'm not aware of a
64bit windows build).

On 2013/03/07 23:37:15, DaleCurtis wrote:

   }
 
   // Get the brand string of the cpu.
   __cpuid(cpu_info, 0x80000000);
   const int parameter_end = 0x80000004;
 
   if (cpu_info[0] >= parameter_end) {
     char* cpu_string_ptr = cpu_string;
 
     for (int parameter = 0x80000002; parameter <= parameter_end &&
@@ skipping 12 matching lines @@
   if (has_sse42()) return SSE42;
   if (has_sse41()) return SSE41;
   if (has_ssse3()) return SSSE3;
   if (has_sse3()) return SSE3;
   if (has_sse2()) return SSE2;
   if (has_sse()) return SSE;
   return PENTIUM;
 }
 
 }  // namespace base