Fixed AVX undefined symbols on AARCH64 and removed AVX from ehs::Util::Compare.

CMakeLists.txt

@@ -217,6 +217,7 @@ if (IS_OS_WINDOWS)
         src/system/CPU_MSVC_AMD64.asm src/HRNG_MSVC.asm src/Math_MSVC_AMD64.asm
         src/io/Directory_W32.cpp include/ehs/io/Directory_W32.h
         include/ehs/io/socket/ICMP_W32.h src/io/socket/ICMP_W32.cpp
+        src/system/AVX2_MSVC_AMD64.asm src/system/AVX512_MSVC_AMD64.asm
     )
 
 elseif (IS_OS_LINUX)
@@ -236,6 +237,7 @@ elseif (IS_OS_LINUX)
         src/io/Directory_LNX.cpp include/ehs/io/Directory_LNX.h
         src/io/Usb_LNX.cpp include/ehs/io/Usb_LNX.h
         include/ehs/io/socket/ICMP_LNX.h src/io/socket/ICMP_LNX.cpp
+        src/system/AVX2_GCC_AMD64.asm src/system/AVX512_GCC_AMD64.asm
     )
 
     #set(LINUX_WINDOW_SYSTEM "Wayland" CACHE STRING "Linux Window System")
@@ -254,9 +256,9 @@ elseif (IS_OS_LINUX)
     endif ()
 
     if (IS_ARCH_AMD64)
-        list(APPEND EHS_SOURCES src/system/CPU_GCC_AMD64.asm src/HRNG_GCC.asm src/Math_GCC_AMD64.asm src/system/AVX2_GCC_AMD64.asm src/system/AVX512_GCC_AMD64.asm)
+        list(APPEND EHS_SOURCES src/system/CPU_GCC_AMD64.asm src/HRNG_GCC.asm src/Math_GCC_AMD64.asm)
     elseif (IS_ARCH_ARM64)
-        list(APPEND EHS_SOURCES src/system/CPU_ARM64.cpp src/HRNG_ARM64.cpp src/Math_GCC_ARM64.s)
+        list(APPEND EHS_SOURCES src/system/CPU_ARM64.cpp src/HRNG_ARM64.cpp src/Math_GCC_ARM64.s src/system/AVX2_AARCH64.cpp src/system/AVX512_AARCH64.cpp)
     endif ()
 endif()
 

include/ehs/system/AVX512.h

@@ -33,4 +33,4 @@ namespace ehs
         /// @note The parameters "a", and "b" must have alignas(32).
         static bool CompareAligned(const SInt_64 *a, const SInt_64 *b);
     };
-}
+}

src/Util.cpp

@@ -15,21 +15,7 @@ namespace ehs
 
 		while (i < size)
 		{
-			if (CPU::hasAVX512F && remainder >= 64)
-			{
-				if (!AVX512::CompareUnaligned((UInt_64*)&aBytes[i], (UInt_64*)&bBytes[i]))
-					return false;
-
-				i += 64;
-			}
-			else if (CPU::hasAVX2 && remainder >= 32)
-			{
-				if (!AVX2::CompareUnaligned((UInt_64*)&aBytes[i], (UInt_64*)&bBytes[i]))
-					return false;
-
-				i += 32;
-			}
-			else if (remainder >= sizeof(UInt_64))
+			if (remainder >= sizeof(UInt_64))
 			{
 				if (*(UInt_64*)&aBytes[i] != *(UInt_64*)&bBytes[i])
 					return false;

src/system/AVX2_AARCH64.cpp

@@ -0,0 +1,24 @@
+#include "ehs/system/AVX2.h"
+
+namespace ehs
+{
+    bool AVX2::CompareUnaligned(const UInt_64* a, const UInt_64* b)
+    {
+        return false;
+    }
+
+    bool AVX2::CompareUnaligned(const SInt_64* a, const SInt_64* b)
+    {
+        return false;
+    }
+
+    bool AVX2::CompareAligned(const UInt_64* a, const UInt_64* b)
+    {
+        return false;
+    }
+
+    bool AVX2::CompareAligned(const SInt_64* a, const SInt_64* b)
+    {
+        return false;
+    }
+}

src/system/AVX2_MSVC_AMD64.asm

@@ -0,0 +1,49 @@
+global ?CompareUnaligned@AVX2@ehs@@SA_NPEBK0@Z
+global ?CompareUnaligned@AVX2@ehs@@SA_NPEBJ0@Z
+global ?CompareAligned@AVX2@ehs@@SA_NPEBK0@Z
+global ?CompareAligned@AVX2@ehs@@SA_NPEBJ0@Z
+
+section .text
+    ?CompareUnaligned@AVX2@ehs@@SA_NPEBK0@Z:
+        VMOVDQU YMM0, [RCX]
+        VMOVDQU YMM1, [RDX]
+        VPCMPEQQ YMM2, YMM0, YMM1
+
+        VPMOVMSKB EAX, YMM2
+        CMP EAX, 0xFFFFFFFF
+        SETE AL
+
+        RET
+
+    ?CompareUnaligned@AVX2@ehs@@SA_NPEBJ0@Z:
+        VMOVDQU YMM0, [RCX]
+        VMOVDQU YMM1, [RDX]
+        VPCMPEQQ YMM2, YMM0, YMM1
+
+        VPMOVMSKB EAX, YMM2
+        CMP EAX, 0xFFFFFFFF
+        SETE AL
+
+        RET
+
+    ?CompareAligned@AVX2@ehs@@SA_NPEBK0@Z:
+        VMOVDQA YMM0, [RCX]
+        VMOVDQA YMM1, [RDX]
+        VPCMPEQQ YMM2, YMM0, YMM1
+
+        VPMOVMSKB EAX, YMM2
+        CMP EAX, 0xFFFFFFFF
+        SETE AL
+
+        RET
+
+    ?CompareAligned@AVX2@ehs@@SA_NPEBJ0@Z:
+        VMOVDQA YMM0, [RCX]
+        VMOVDQA YMM1, [RDX]
+        VPCMPEQQ YMM2, YMM0, YMM1
+
+        VPMOVMSKB EAX, YMM2
+        CMP EAX, 0xFFFFFFFF
+        SETE AL
+
+        RET

src/system/AVX512_AARCH64.cpp

@@ -0,0 +1,24 @@
+#include "ehs/system/AVX512.h"
+
+namespace ehs
+{
+    bool AVX512::CompareUnaligned(const UInt_64* a, const UInt_64* b)
+    {
+        return false;
+    }
+
+    bool AVX512::CompareUnaligned(const SInt_64* a, const SInt_64* b)
+    {
+        return false;
+    }
+
+    bool AVX512::CompareAligned(const UInt_64* a, const UInt_64* b)
+    {
+        return false;
+    }
+
+    bool AVX512::CompareAligned(const SInt_64* a, const SInt_64* b)
+    {
+        return false;
+    }
+}

src/system/AVX512_MSVC_AMD64.asm

@@ -0,0 +1,49 @@
+global ?CompareUnaligned@AVX512@ehs@@SA_NPEBK0@Z
+global ?CompareUnaligned@AVX512@ehs@@SA_NPEBJ0@Z
+global ?CompareAligned@AVX512@ehs@@SA_NPEBK0@Z
+global ?CompareAligned@AVX512@ehs@@SA_NPEBJ0@Z
+
+section .text
+	?CompareUnaligned@AVX512@ehs@@SA_NPEBK0@Z:
+		VMOVDQU64 ZMM0, [RCX]
+		VMOVDQU64 ZMM1, [RDX]
+
+		VPCMPEQQ K1, ZMM0, ZMM1
+
+		KORTESTQ K1, K1
+		SETC AL
+
+		RET
+
+	?CompareUnaligned@AVX512@ehs@@SA_NPEBJ0@Z:
+		VMOVDQU64 ZMM0, [RCX]
+		VMOVDQU64 ZMM1, [RDX]
+
+		VPCMPEQQ K1, ZMM0, ZMM1
+
+		KORTESTQ K1, K1
+		SETC AL
+
+		RET
+
+	?CompareAligned@AVX512@ehs@@SA_NPEBK0@Z:
+		VMOVDQA64 ZMM0, [RCX]
+		VMOVDQA64 ZMM1, [RDX]
+
+		VPCMPEQQ K1, ZMM0, ZMM1
+
+		KORTESTQ K1, K1
+		SETC AL
+
+		RET
+
+	?CompareAligned@AVX512@ehs@@SA_NPEBJ0@Z:
+		VMOVDQA64 ZMM0, [RCX]
+		VMOVDQA64 ZMM1, [RDX]
+
+		VPCMPEQQ K1, ZMM0, ZMM1
+
+		KORTESTQ K1, K1
+		SETC AL
+
+		RET

src/system/CPU.cpp

@@ -744,7 +744,7 @@ namespace ehs
     bool CPU::RetrieveAVX512F()
     {
         #ifdef EHS_ARCH_X64
-	            return GetExtFeatureBits_1() & 0b00000000000000001000000000000000;
+	            return GetExtFeatureBits_1() & 0b00000000000000010000000000000000;
         #else
 	            return false;
         #endif