added neon support for popcount (#861)

added neon support for popcount

include/libtorrent/aux_/cpuid.hpp

@@ -40,6 +40,7 @@ namespace libtorrent { namespace aux
 	// initialized by static initializers (in cpuid.cpp)
 	TORRENT_EXTRA_EXPORT extern bool sse42_support;
 	TORRENT_EXTRA_EXPORT extern bool mmx_support;
+	TORRENT_EXTRA_EXPORT extern bool arm_neon_support;
 } }
 
 #endif // TORRENT_CPUID_HPP_INCLUDED

include/libtorrent/config.hpp

@@ -538,5 +538,10 @@ POSSIBILITY OF SUCH DAMAGE.
 #	define TORRENT_HAS_BUILTIN_CLZ 0
 #endif // TORRENT_HAS_BUILTIN_CLZ
 
-#endif // TORRENT_CONFIG_HPP_INCLUDED
+#if TORRENT_HAS_ARM && defined __ARM_NEON
+#	define TORRENT_HAS_ARM_NEON 1
+#else
+#	define TORRENT_HAS_ARM_NEON 0
+#endif // TORRENT_HAS_ARM_NEON
 
+#endif // TORRENT_CONFIG_HPP_INCLUDED

src/bitfield.cpp

@@ -37,6 +37,10 @@ POSSIBILITY OF SUCH DAMAGE.
 #include <intrin.h>
 #endif
 
+#if TORRENT_HAS_ARM_NEON
+#include <arm_neon.h>
+#endif
+
 namespace libtorrent
 {
 	bool bitfield::all_set() const
@@ -79,6 +83,25 @@ namespace libtorrent
 		}
 #endif // TORRENT_HAS_SSE
 
+#if TORRENT_HAS_ARM_NEON
+		if (aux::arm_neon_support)
+		{
+			for (int i = 0; i < words; ++i)
+			{
+				uint8x8_t const in_val = vld1_u8((unsigned char *) &m_buf[i]);
+				uint8x8_t const cnt8x8_val = vcnt_u8(in_val);
+				uint16x4_t const cnt16x4_val = vpaddl_u8(cnt8x8_val);
+				uint32x2_t const cnt32x2_val = vpaddl_u16(cnt16x4_val);
+				uint32_t cnt;
+				vst1_u32(&cnt, cnt32x2_val);
+
+				ret += cnt;
+			}
+
+			return ret;
+		}
+#endif // TORRENT_HAS_ARM_NEON
+
 		for (int i = 0; i < words; ++i)
 		{
 			std::uint32_t v = m_buf[i];

src/cpuid.cpp

@@ -44,6 +44,10 @@ POSSIBILITY OF SUCH DAMAGE.
 #include <cstring> // for std::memset
 #endif
 
+#if TORRENT_HAS_ARM
+#include <sys/auxv.h>
+#endif
+
 namespace libtorrent { namespace aux
 {
 	namespace {
@@ -87,8 +91,24 @@ namespace libtorrent { namespace aux
 #endif
 	}
 
+	bool supports_arm_neon()
+	{
+#if TORRENT_HAS_ARM_NEON
+#if defined __arm__
+		//return (getauxval(AT_HWCAP) & HWCAP_NEON);
+		return (getauxval(16) & (1 << 12));
+#elif defined __aarch64__
+		//return (getauxval(AT_HWCAP) & HWCAP_ASIMD);
+		return (getauxval(16) & (1 << 1));
+#endif // TORRENT_HAS_ARM
+#else
+		return false;
+#endif
+	}
+
 	} // anonymous namespace
 
 	bool sse42_support = supports_sse42();
 	bool mmx_support = supports_mmx();
+	bool arm_neon_support = supports_arm_neon();
 } }

test/Jamfile

@@ -125,7 +125,6 @@ test-suite libtorrent :
 		test_socket_io.cpp
 #		test_random.cpp
 		test_gzip.cpp
-		test_bitfield.cpp
 		test_part_file.cpp
 		test_peer_list.cpp
 		test_torrent_info.cpp
@@ -160,6 +159,7 @@ test-suite libtorrent :
 		]
 
 	[ run test_sha1_hash.cpp ]
+	[ run test_bitfield.cpp ]
 	[ run test_receive_buffer.cpp ]
 	[ run test_alert_manager.cpp ]
 	[ run test_direct_dht.cpp ]
@@ -271,6 +271,7 @@ explicit win-tests ;
 
 alias arm-tests :
     test_sha1_hash
+    test_bitfield
     ;
 
 explicit arm-tests ;

test/test_bitfield.cpp

@@ -32,6 +32,7 @@ POSSIBILITY OF SUCH DAMAGE.
 
 #include "test.hpp"
 #include "libtorrent/bitfield.hpp"
+#include "libtorrent/aux_/cpuid.hpp"
 #include <stdlib.h>
 
 using namespace libtorrent;
@@ -187,4 +188,10 @@ TORRENT_TEST(bitfield)
 		TEST_EQUAL(test1.count(), 32 + 8 + 2);
 		TEST_EQUAL(test1.all_set(), true);
 	}
+
+#if TORRENT_HAS_ARM
+	TORRENT_ASSERT(aux::arm_neon_support);
+#else
+	TORRENT_ASSERT(!aux::arm_neon_support);
+#endif
 }