Adding ARM64 support and optimized Neon implementation (#270)

Add NEON intrinsics routines for lpc_compute_residual_from_qlp_coefficients
and lpc_compute_residual_from_qlp_coefficients_wide

CMakeLists.txt

@@ -104,6 +104,8 @@ enable_testing()
 check_include_file("byteswap.h" HAVE_BYTESWAP_H)
 check_include_file("inttypes.h" HAVE_INTTYPES_H)
 check_include_file("stdint.h" HAVE_STDINT_H)
+check_include_file("arm_neon.h" FLAC__HAS_NEONINTRIN)
+
 if(MSVC)
     check_include_file("intrin.h" FLAC__HAS_X86INTRIN)
 else()

cmake/CheckA64NEON.c.in

@@ -0,0 +1,6 @@
+#include <arm_neon.h>
+int main (void)
+{
+    float64x2_t tmp;
+    tmp = vdupq_n_f64(0.0f);
+}

cmake/CheckA64NEON.cmake

@@ -0,0 +1,14 @@
+macro(CHECK_A64NEON VARIABLE)
+    if(NOT DEFINED HAVE_${VARIABLE})
+        message(STATUS "Check whether A64 NEON can be used")
+        configure_file(${PROJECT_SOURCE_DIR}/cmake/CheckA64NEON.c.in ${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/CheckA64NEON.c @ONLY)
+        try_compile(HAVE_${VARIABLE} "${PROJECT_BINARY_DIR}"
+            "${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/CheckA64NEON.c")
+        if(HAVE_${VARIABLE})
+            message(STATUS "Check whether A64 NEON can be used - yes")
+            set(${VARIABLE} 1 CACHE INTERNAL "Result of CHECK_A64NEON" FORCE)
+        else ()
+            message(STATUS "Check whether A64 NEON can be used - no")
+        endif()
+    endif ()
+endmacro(CHECK_A64NEON)

cmake/CheckCPUArch.cmake

@@ -7,7 +7,7 @@ macro(_CHECK_CPU_ARCH ARCH ARCH_DEFINES VARIABLE)
             "${PROJECT_BINARY_DIR}/CMakeFiles/CMakeTmp/CheckCPUArch.c")
         if(HAVE_${VARIABLE})
             message(STATUS "Check CPU architecture is ${ARCH} - yes")
-            set(${VARIABLE} 1 CACHE INTERNAL "Result of CHECK_CPU_ARCH_X64" FORCE)
+            set(${VARIABLE} 1 CACHE INTERNAL "Result of CHECK_CPU_ARCH" FORCE)
         else ()
             message(STATUS "Check CPU architecture is ${ARCH} - no")
         endif()
@@ -25,3 +25,7 @@ endmacro(CHECK_CPU_ARCH_X86)
 macro(CHECK_CPU_ARCH_PPC64 VARIABLE)
     _CHECK_CPU_ARCH(ppc64 "defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__) ||defined(_ARCH_PPC64)" ${VARIABLE})
 endmacro(CHECK_CPU_ARCH_PPC64)
+
+macro(CHECK_CPU_ARCH_ARM64 VARIABLE)
+    _CHECK_CPU_ARCH(arm64 "defined(__aarch64__) || defined(__arm64__)" ${VARIABLE})
+endmacro(CHECK_CPU_ARCH_ARM64)

config.cmake.h.in

@@ -40,6 +40,9 @@
 /* Set to 1 if <x86intrin.h> is available. */
 #cmakedefine01 FLAC__HAS_X86INTRIN
 
+/* Set to 1 if <arm_neon.h> is available. */
+#cmakedefine01 FLAC__HAS_NEONINTRIN
+
 /* define if building for Darwin / MacOS X */
 #cmakedefine FLAC__SYS_DARWIN
 

configure.ac

@@ -58,7 +58,7 @@ AM_PROG_CC_C_O
 AC_C_INLINE
 AC_C_TYPEOF
 
-AC_CHECK_HEADERS([stdint.h inttypes.h byteswap.h sys/param.h sys/ioctl.h termios.h x86intrin.h cpuid.h])
+AC_CHECK_HEADERS([stdint.h inttypes.h byteswap.h sys/param.h sys/ioctl.h termios.h x86intrin.h cpuid.h arm_neon.h])
 
 XIPH_C_BSWAP32
 XIPH_C_BSWAP16
@@ -145,6 +145,12 @@ case "$host_cpu" in
 		AH_TEMPLATE(FLAC__CPU_PPC, [define if building for PowerPC])
 		asm_optimisation=$asm_opt
 		;;
+	arm64|aarch64)
+		cpu_arm64=true
+		AC_DEFINE(FLAC__CPU_ARM64)
+		AH_TEMPLATE(FLAC__CPU_ARM64, [define if building for ARM])
+		asm_optimisation=$asm_opt
+		;;
 	sparc)
 		cpu_sparc=true
 		AC_DEFINE(FLAC__CPU_SPARC)
@@ -156,6 +162,7 @@ AM_CONDITIONAL(FLAC__CPU_X86_64, test "x$cpu_x86_64" = xtrue)
 AM_CONDITIONAL(FLaC__CPU_IA32, test "x$cpu_ia32" = xtrue)
 AM_CONDITIONAL(FLaC__CPU_PPC, test "x$cpu_ppc" = xtrue)
 AM_CONDITIONAL(FLaC__CPU_PPC64, test "x$cpu_ppc64" = xtrue)
+AM_CONDITIONAL(FLAC__CPU_ARM64, test "x$cpu_arm64" = xtrue)
 AM_CONDITIONAL(FLaC__CPU_SPARC, test "x$cpu_sparc" = xtrue)
 
 if test "x$ac_cv_header_x86intrin_h" = xyes; then
@@ -164,6 +171,26 @@ else
 AC_DEFINE([FLAC__HAS_X86INTRIN], 0)
 fi
 
+neon=no
+if test "x$ac_cv_header_arm_neon_h" = xyes; then
+AC_DEFINE([FLAC__HAS_NEONINTRIN], 1, [Set to 1 if <arm_neon.h> is available.])
+neon=yes
+  AC_MSG_CHECKING([whether arm_neon.h has A64 functions])
+  AC_COMPILE_IFELSE(
+	[AC_LANG_PROGRAM([[#include <arm_neon.h>]],
+	[[float64x2_t sum5; sum5 = vdupq_n_f64(0.0f);]])],
+	[AC_MSG_RESULT([yes])
+	 has_a64neon=yes],
+	[AC_MSG_RESULT([no])])
+  if test "x$has_a64neon" = xyes; then
+  AC_DEFINE([FLAC__HAS_A64NEONINTRIN], 1, [Set to 1 if <arm_neon.h> has A64 instructions.])
+  else
+  AC_DEFINE([FLAC__HAS_A64NEONINTRIN], 0)
+  fi
+else
+AC_DEFINE([FLAC__HAS_NEONINTRIN], 0)
+fi
+
 if test x"$cpu_ppc64" = xtrue ; then
 
 AC_C_ATTRIBUTE([target("cpu=power8")],
@@ -664,6 +691,7 @@ if test x$ac_cv_c_compiler_gnu = xyes ; then
 fi
 	echo "    Compiler is Clang : ....................... ${xiph_cv_c_compiler_clang}"
 	echo "    SSE optimizations : ....................... ${sse_os}"
+	echo "    Neon optimizations : ...................... ${neon}"
 	echo "    Asm optimizations : ....................... ${asm_optimisation}"
 	echo "    Ogg/FLAC support : ........................ ${have_ogg}"
 	echo "    Stack protector  : ........................ ${enable_stack_smash_protection}"

src/libFLAC/CMakeLists.txt

@@ -10,6 +10,7 @@ include(CheckCSourceCompiles)
 include(CheckCPUArch)
 include(CheckAttribute)
 include(CheckVSX)
+include(CheckA64NEON)
 
 check_cpu_arch_x64(FLAC__CPU_X86_64)
 if(NOT FLAC__CPU_X86_64)
@@ -26,9 +27,15 @@ else()
         check_attribute_power8(FLAC__HAS_TARGET_POWER8)
         check_attribute_power9(FLAC__HAS_TARGET_POWER9)
         check_vsx(FLAC__USE_VSX)
+    else()
+        check_cpu_arch_arm64(FLAC__CPU_ARM64)
+        if(FLAC__CPU_ARM64)
+             check_a64neon(FLAC__HAS_A64NEONINTRIN)
+        endif()
     endif()
 endif()
 
+
 include(CheckLanguage)
 check_language(ASM_NASM)
 if(CMAKE_ASM_NASM_COMPILER)
@@ -68,6 +75,7 @@ add_library(FLAC
     float.c
     format.c
     lpc.c
+    lpc_intrin_neon.c
     lpc_intrin_sse2.c
     lpc_intrin_sse41.c
     lpc_intrin_avx2.c

src/libFLAC/Makefile.am

@@ -117,6 +117,7 @@ libFLAC_sources = \
 	lpc_intrin_sse41.c \
 	lpc_intrin_avx2.c \
 	lpc_intrin_vsx.c \
+	lpc_intrin_neon.c \
 	md5.c \
 	memory.c \
 	metadata_iterators.c \

src/libFLAC/Makefile.lite

@@ -90,6 +90,7 @@ SRCS_C = \
 	lpc_intrin_sse2.c \
 	lpc_intrin_sse41.c \
 	lpc_intrin_avx2.c \
+	lpc_intrin_neon.c \
 	md5.c \
 	memory.c \
 	metadata_iterators.c \

src/libFLAC/include/private/lpc.h

@@ -151,7 +151,13 @@ int FLAC__lpc_quantize_coefficients(const FLAC__real lp_coeff[], uint32_t order,
  */
 void FLAC__lpc_compute_residual_from_qlp_coefficients(const FLAC__int32 *data, uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 residual[]);
 void FLAC__lpc_compute_residual_from_qlp_coefficients_wide(const FLAC__int32 *data, uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 residual[]);
+
 #ifndef FLAC__NO_ASM
+#   ifdef FLAC__CPU_ARM64
+void FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_neon(const FLAC__int32 *data, uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 residual[]);
+void FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_neon(const FLAC__int32 *data, uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 residual[]);
+#   endif
+
 #  ifdef FLAC__CPU_IA32
 #    ifdef FLAC__HAS_NASM
 void FLAC__lpc_compute_residual_from_qlp_coefficients_asm_ia32(const FLAC__int32 *data, uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 residual[]);

src/libFLAC/stream_encoder.c

@@ -1003,8 +1003,16 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_(
 #    endif
 #   endif /* FLAC__HAS_X86INTRIN */
 #  endif /* FLAC__CPU_... */
+
+	#if defined FLAC__CPU_ARM64
+    encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_16bit = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_neon;
+    encoder->private_->local_lpc_compute_residual_from_qlp_coefficients       = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_neon;
+    encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_neon;
+	# endif
+
 	}
 # endif /* !FLAC__NO_ASM */
+
 #endif /* !FLAC__INTEGER_ONLY_LIBRARY */
 #if !defined FLAC__NO_ASM && FLAC__HAS_X86INTRIN
 	if(encoder->private_->cpuinfo.use_asm) {