Fix CPU detection and aligned alloc on mac

 Committer: Nikolay Bogoychev <nheart@gmail.com>

2 files changed, 5 insertions, 1 deletions

diff --git a/aligned.h b/aligned.h
index 8a7d6cc..c07c23e 100644
--- a/aligned.h
+++ b/aligned.h
@@ -1,5 +1,6 @@
 #pragma once
 #include <cstdlib>
+#include <algorithm>
 
 // 64-byte aligned simple vector.
 
@@ -8,7 +9,8 @@ namespace intgemm {
 template <class T> class AlignedVector {
   public:
     explicit AlignedVector(std::size_t size)
-      : mem_(static_cast<T*>(aligned_alloc(64, size * sizeof(T)))), size_(size) {}
+      : mem_(static_cast<T*>(aligned_alloc(64, std::max((size_t)64, size * sizeof(T))))), size_(size) {
+    }
 
     AlignedVector(const AlignedVector&) = delete;
     AlignedVector& operator=(const AlignedVector&) = delete;
diff --git a/intgemm.h b/intgemm.h
index 8940085..5e5f556 100644
--- a/intgemm.h
+++ b/intgemm.h
@@ -124,6 +124,7 @@ typedef Unsupported_8bit AVX512VNNI_8bit;
 #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
 // gcc 5.4.0 bizarrely supports avx512bw targets but not __builtin_cpu_supports("avx512bw").  So implement it manually.
 inline bool CheckAVX512BW() {
+  __builtin_cpu_init ();
 #ifdef __INTEL_COMPILER
   return _may_i_use_cpu_feature(_FEATURE_AVX512BW)
 #elif __GNUC__
@@ -153,6 +154,7 @@ inline bool CheckAVX512BW() {
  * unsupported otherwise
  */
 template <class T> T ChooseCPU(T avx512vnni, T avx512bw, T avx2, T ssse3, T sse2, T unsupported) {
+  __builtin_cpu_init ();
 #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
   if (
 #ifdef __INTEL_COMPILER