Workarounds for icc, still need operator() to change

3 files changed, 27 insertions, 6 deletions

diff --git a/compile_test_avx512vnni.cc b/compile_test_avx512vnni.cc
index fc1c3dd..611cc53 100644
--- a/compile_test_avx512vnni.cc
+++ b/compile_test_avx512vnni.cc
@@ -1,6 +1,11 @@
 #include <immintrin.h>
 
-__attribute__ ((target ("avx512f,avx512bw,avx512dq,avx512vnni"))) bool Foo() {
+#ifdef __INTEL_COMPILER
+__attribute__ ((target ("avx512f")))
+#else
+__attribute__ ((target ("avx512f,avx512bw,avx512dq,avx512vnni")))
+#endif
+bool Foo() {
   // AVX512F
   __m512i value = _mm512_set1_epi32(1);
   // AVX512BW
@@ -14,5 +19,11 @@ __attribute__ ((target ("avx512f,avx512bw,avx512dq,avx512vnni"))) bool Foo() {
 }
 
 int main() {
-  return Foo() && __builtin_cpu_supports("avx512vnni");
+  return Foo() &&
+#ifdef __INTEL_COMPILER
+    _may_i_use_cpu_feature(_FEATURE_AVX512_VNNI)
+#else
+    __builtin_cpu_supports("avx512vnni")
+#endif
+    ;
 }
diff --git a/intgemm.h b/intgemm.h
index 6f6839c..8b933f2 100644
--- a/intgemm.h
+++ b/intgemm.h
@@ -130,14 +130,25 @@ typedef Unsupported_8bit AVX512VNNI_8bit;
  * unsupported otherwise
  */
 template <class T> T ChooseCPU(T avx512vnni, T avx512, T avx2, T ssse3, T sse2, T unsupported) {
-  // TODO: don't catch Knights processors here!
 #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
-  if (__builtin_cpu_supports("avx512vnni")) {
+  if (
+#ifdef __INTEL_COMPILER
+      _may_i_use_cpu_feature(_FEATURE_AVX512_VNNI)
+#else
+      __builtin_cpu_supports("avx512vnni")
+#endif
+      ) {
     return avx512vnni;
   }
 #endif
 #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512
-  if (__builtin_cpu_supports("avx512f")) {
+  if (
+#ifdef __INTEL_COMPILER
+      __builtin_cpu_supports("avx512bw")
+#else
+      _may_i_use_cpu_feature(_FEATURE_AVX512BW)
+#endif
+      ) {
     return avx512;
   }
 #endif
diff --git a/types.h b/types.h
index fb010b0..d901e18 100644
--- a/types.h
+++ b/types.h
@@ -10,7 +10,6 @@
 #define INTGEMM_AVX512F __attribute__ ((target ("avx512f")))
 #define INTGEMM_AVX512BW __attribute__ ((target ("avx512f")))
 #define INTGEMM_AVX512DQ __attribute__ ((target ("avx512f")))
-// TODO is this right?
 #define INTGEMM_AVX512VNNI __attribute__ ((target ("avx512f")))
 #else
 #define INTGEMM_AVX512F __attribute__ ((target ("avx512f")))