cpu: add AVX512 support

3 files changed, 6 insertions, 1 deletions

diff --git a/cpu.c b/cpu.c
index 3dd2ae8..a7b7231 100644
--- a/cpu.c
+++ b/cpu.c
@@ -76,6 +76,8 @@ static int get_cpu_flags_x86(void)
             if (ebx & 0x00000100)
                 rval |= MG2DI_CPU_FLAG_BMI2;
         }
+        if ((ebx & 0xD0030000) == 0xD0030000)
+            rval |= MG2DI_CPU_FLAG_AVX512;
     }
 
     mg2di_cpu_cpuid(0x80000000, &max_ext_level, &ebx, &ecx, &edx);
diff --git a/cpu.h b/cpu.h
index e0c501d..a757003 100644
--- a/cpu.h
+++ b/cpu.h
@@ -43,6 +43,7 @@
 #define MG2DI_CPU_FLAG_FMA3        0x10000 ///< Haswell FMA3 functions
 #define MG2DI_CPU_FLAG_BMI1        0x20000 ///< Bit Manipulation Instruction Set 1
 #define MG2DI_CPU_FLAG_BMI2        0x40000 ///< Bit Manipulation Instruction Set 2
+#define MG2DI_CPU_FLAG_AVX512      (1<<16)
 
 int mg2di_cpu_flags_get(void);
 
diff --git a/util.asm b/util.asm
index 846ea3b..864080f 100644
--- a/util.asm
+++ b/util.asm
@@ -29,7 +29,9 @@
         %error Invalid selector %3
     %endif
 
-    %if mmsize == 32 && cpuflag(avx2)
+    %if mmsize == 64 && cpuflag(avx512)
+        vbroadcastsd %1, %2
+    %elif mmsize == 32 && cpuflag(avx2)
         vpermpd %1, %2, _IMM8SPLAT2B(%3)
     %elif mmsize == 32 && cpuflag(avx)
         shufpd %1, %2, %2, _IMM4SPLAT1B(%3 & 1)