From 200f8b1770c83b755918573dda578be1b2d9dddd Mon Sep 17 00:00:00 2001
From: quanghanh <quanghanh@visenze.com>
Date: Fri, 21 Jul 2023 09:59:38 +0000
Subject: [PATCH] feat: build vlfeat on arm64

---
 Makefile                  |  7 +++++++
 docsrc/install-shell.html |  5 +++++
 make/dll.mak              |  6 ++++++
 make/matlab.mak           |  4 ++++
 vl/host.c                 |  8 ++++++--
 vl/host.h                 | 14 ++++++++++++++
 vl/kmeans.c               |  2 +-
 7 files changed, 43 insertions(+), 3 deletions(-)

diff --git a/Makefile b/Makefile
index 9f0655a2..f0ac3b9e 100644
--- a/Makefile
+++ b/Makefile
@@ -130,6 +130,7 @@ Linux_i386_ARCH := glnx86
 Linux_i686_ARCH := glnx86
 Linux_unknown_ARCH := glnx86
 Linux_x86_64_ARCH := glnxa64
+Linux_aarch64_ARCH := armv8
 
 UNAME := $(shell uname -sm)
 ARCH ?= $($(shell echo "$(UNAME)" | tr \  _)_ARCH)
@@ -232,6 +233,12 @@ STD_CFLAGS += -m$(march) -D_GNU_SOURCE -fno-stack-protector
 STD_LDFLAGS += -m$(march) -Wl,--rpath,\$$ORIGIN/ -Wl,--as-needed
 endif
 
+# Linux ARM64
+# Linux
+ifeq "$(ARCH)" "$(filter $(ARCH),armv8)"
+DISABLE_SSE2:=yes
+endif
+
 # Convert back DISALBE_*="no" flags to be empty
 ifeq "$(DISABLE_SSE2)" "no"
 override DISABLE_SSE2:=
diff --git a/docsrc/install-shell.html b/docsrc/install-shell.html
index c7695c77..409a403b 100644
--- a/docsrc/install-shell.html
+++ b/docsrc/install-shell.html
@@ -69,6 +69,11 @@
 <td><code>glnxa64</code></td>
 <td><code>VLFEATROOT/bin/glnxa64/sift</code></td>
 </tr>
+<tr>
+<td>Linux ARM64</td>
+<td><code>armv8</code></td>
+<td><code>VLFEATROOT/bin/armv8/sift</code></td>
+</tr>
 </table>
 </div>
 
diff --git a/make/dll.mak b/make/dll.mak
index 22c30396..2ea561fe 100644
--- a/make/dll.mak
+++ b/make/dll.mak
@@ -76,6 +76,12 @@ DLL_SUFFIX := so
 DLL_LDFLAGS += -m64
 endif
 
+# Linux-arm64
+ifeq ($(ARCH),armv8)
+DLL_SUFFIX := so
+DLL_LDFLAGS += -march=armv8-a
+endif
+
 # --------------------------------------------------------------------
 #                                                                Build
 # --------------------------------------------------------------------
diff --git a/make/matlab.mak b/make/matlab.mak
index 4dde699e..086a3f84 100644
--- a/make/matlab.mak
+++ b/make/matlab.mak
@@ -59,6 +59,10 @@ ifeq ($(ARCH),glnxa64)
 MEX_SUFFIX := mexa64
 endif
 
+ifeq ($(ARCH),armv8)
+MEX_SUFFIX := mexa64
+endif
+
 MEX_BINDIR := toolbox/mex/$(MEX_SUFFIX)
 
 # generate the mex-dir target
diff --git a/vl/host.c b/vl/host.c
index ef97cdca..92efc62b 100644
--- a/vl/host.c
+++ b/vl/host.c
@@ -403,7 +403,9 @@ it supports POSIX threads.
 #undef HAS_CPUID
 #endif
 
-#if defined(HAS_CPUID) & defined(VL_COMPILER_MSC)
+#if defined(HAS_CPUID)
+
+#if defined(VL_COMPILER_MSC)
 #include <intrin.h>
 VL_INLINE void
 _vl_cpuid (vl_int32* info, int function)
@@ -412,7 +414,7 @@ _vl_cpuid (vl_int32* info, int function)
 }
 #endif
 
-#if defined(HAS_CPUID) & defined(VL_COMPILER_GNUC)
+#if defined(VL_COMPILER_GNUC)
 VL_INLINE void
 _vl_cpuid (vl_int32* info, int function)
 {
@@ -488,6 +490,8 @@ _vl_x86cpu_info_to_string_copy (VlX86CpuInfo const *self)
   return string ;
 }
 
+#endif // defined(HAS_CPUID)
+
 /** ------------------------------------------------------------------
  ** @brief Human readable static library configuration
  ** @return a new string with the static configuration.
diff --git a/vl/host.h b/vl/host.h
index 293fe1f5..73ddba6d 100644
--- a/vl/host.h
+++ b/vl/host.h
@@ -266,6 +266,17 @@ defined(__DOXYGEN__)
     defined(__DOXYGEN__)
 #define VL_ARCH_IA64
 #endif
+
+#if defined(_M_ARM64)     || \
+    defined(__aarch64__)  || \
+    defined(__ARM64_ARCH_8__) || \
+    defined(__ARM_ARCH_8A__)  || \
+    defined(__arm64__)  || \
+    defined(__arm64)    || \
+    defined(__arm64__)
+#define VL_ARCH_ARM64
+#endif
+
 /** @} */
 
 /** @name Identifying the host data model
@@ -299,6 +310,7 @@ defined(__DOXYGEN__)
     defined(VL_ARCH_IX86)      || \
     defined(VL_ARCH_IA64)      || \
     defined(VL_ARCH_X64)       || \
+    defined(VL_ARCH_ARM64)     || \
     defined(__DOXYGEN__)
 #define VL_ARCH_LITTLE_ENDIAN
 #endif
@@ -553,6 +565,7 @@ VL_INLINE void vl_swap_host_big_endianness_2 (void *dst, void* src) ;
 /** ------------------------------------------------------------------
  ** @name Obtaining host info at run time
  ** @{ */
+#if defined(VL_ARCH_IX86) || defined(VL_ARCH_IA64) || defined(VL_ARCH_X64)
 
 typedef struct _VlX86CpuInfo
 {
@@ -572,6 +585,7 @@ typedef struct _VlX86CpuInfo
 void _vl_x86cpu_info_init (VlX86CpuInfo *self) ;
 char * _vl_x86cpu_info_to_string_copy (VlX86CpuInfo const *self) ;
 
+#endif 
 /** @} */
 
 /** ------------------------------------------------------------------
diff --git a/vl/kmeans.c b/vl/kmeans.c
index c250ca89..b7c93088 100644
--- a/vl/kmeans.c
+++ b/vl/kmeans.c
@@ -669,7 +669,7 @@ VL_XCAT(_vl_kmeans_quantize_, SFX)
 #endif
 
 #ifdef _OPENMP
-#pragma omp parallel default(none) \
+#pragma omp parallel \
             shared(self, distances, assignments, numData, distFn, data) \
             num_threads(vl_get_max_threads())
 #endif