diff --git a/lib/BUILD.bazel b/lib/BUILD.bazel
index 4977d9bb2..f0239af29 100644
--- a/lib/BUILD.bazel
+++ b/lib/BUILD.bazel
@@ -30,6 +30,17 @@ cc_library(
 	copts = ccv_default_copts()
 )
 
+cc_library(
+	name = "metal_cpp_hdrs",
+	srcs = [
+		"nnc/mfa/3rdparty/metal-cpp/Dispatch.cpp",
+	],
+	hdrs = [
+		"nnc/mfa/3rdparty/metal-cpp/Dispatch.hpp",
+		"nnc/mfa/3rdparty/metal-cpp/Metal.hpp",
+	],
+)
+
 cc_library(
 	name = "siphash",
 	srcs = [
@@ -359,6 +370,17 @@ cuda_library(
 	]
 )
 
+cc_library(
+	name = "nnc_mfa_compat",
+	srcs = glob(["nnc/mfa/**/*.cpp"]),
+	hdrs = glob(["nnc/mfa/**/*.hpp"]),
+	copts = ccv_default_copts(),
+	deps = [
+		":metal_cpp_hdrs",
+		":nnc_headers"
+	]
+)
+
 objc_library(
 	name = "nnc_mps_compat",
 	non_arc_srcs = [
@@ -370,8 +392,9 @@ objc_library(
 	copts = ccv_default_copts(),
 	sdk_frameworks = ["Metal", "MetalPerformanceShaders", "MetalPerformanceShadersGraph"],
 	deps = [
+		":nnc_mfa_compat",
 		":nnc_headers",
-		":SFMT_hdrs",
+		":SFMT_hdrs"
 	]
 )
 
@@ -417,6 +440,7 @@ objc_library(
 	copts = ccv_default_copts(),
 	deps = [
 		":nnc_headers",
+		":nnc_mfa_compat",
 		":nnc_mps_compat",
 	]
 )
@@ -516,6 +540,7 @@ cc_library(
 		"//conditions:default": []
 	}) + select({
 		"//config:have_mps": [
+			":nnc_mfa_compat",
 			":nnc_mps_compat",
 			":cmd_mps"
 		],
diff --git a/lib/config.mk.in b/lib/config.mk.in
index 4387fe2c4..a8865c184 100644
--- a/lib/config.mk.in
+++ b/lib/config.mk.in
@@ -4,6 +4,7 @@ NVCC := @NVCC@
 CUDA_SRCS := @CUDA_SRCS@
 CUDA_COMPAT_LIB := @CUDA_COMPAT_LIB@
 CUDA_CMD_LIB := @CUDA_CMD_LIB@
+MFA_COMPAT_LIB := @MFA_COMPAT_LIB@
 MPS_COMPAT_LIB := @MPS_COMPAT_LIB@
 MPS_CMD_LIB := @MPS_CMD_LIB@
 DEFINE_MACROS := @DEFINE_MACROS@
diff --git a/lib/configure b/lib/configure
index a180ffca1..0e64d04e1 100755
--- a/lib/configure
+++ b/lib/configure
@@ -1,9 +1,10 @@
 #! /bin/sh
 # Guess values for system-dependent variables and create Makefiles.
-# Generated by GNU Autoconf 2.69 for libccv 0.7.
+# Generated by GNU Autoconf 2.71 for libccv 0.7.
 #
 #
-# Copyright (C) 1992-1996, 1998-2012 Free Software Foundation, Inc.
+# Copyright (C) 1992-1996, 1998-2017, 2020-2021 Free Software Foundation,
+# Inc.
 #
 #
 # This configure script is free software; the Free Software Foundation
@@ -14,14 +15,16 @@
 
 # Be more Bourne compatible
 DUALCASE=1; export DUALCASE # for MKS sh
-if test -n "${ZSH_VERSION+set}" && (emulate sh) >/dev/null 2>&1; then :
+as_nop=:
+if test ${ZSH_VERSION+y} && (emulate sh) >/dev/null 2>&1
+then :
   emulate sh
   NULLCMD=:
   # Pre-4.2 versions of Zsh do word splitting on ${1+"$@"}, which
   # is contrary to our usage.  Disable this feature.
   alias -g '${1+"$@"}'='"$@"'
   setopt NO_GLOB_SUBST
-else
+else $as_nop
   case `(set -o) 2>/dev/null` in #(
   *posix*) :
     set -o posix ;; #(
@@ -31,46 +34,46 @@ esac
 fi
 
 
+
+# Reset variables that may have inherited troublesome values from
+# the environment.
+
+# IFS needs to be set, to space, tab, and newline, in precisely that order.
+# (If _AS_PATH_WALK were called with IFS unset, it would have the
+# side effect of setting IFS to empty, thus disabling word splitting.)
+# Quoting is to prevent editors from complaining about space-tab.
 as_nl='
 '
 export as_nl
-# Printing a long string crashes Solaris 7 /usr/bin/printf.
-as_echo='\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\'
-as_echo=$as_echo$as_echo$as_echo$as_echo$as_echo
-as_echo=$as_echo$as_echo$as_echo$as_echo$as_echo$as_echo
-# Prefer a ksh shell builtin over an external printf program on Solaris,
-# but without wasting forks for bash or zsh.
-if test -z "$BASH_VERSION$ZSH_VERSION" \
-    && (test "X`print -r -- $as_echo`" = "X$as_echo") 2>/dev/null; then
-  as_echo='print -r --'
-  as_echo_n='print -rn --'
-elif (test "X`printf %s $as_echo`" = "X$as_echo") 2>/dev/null; then
-  as_echo='printf %s\n'
-  as_echo_n='printf %s'
-else
-  if test "X`(/usr/ucb/echo -n -n $as_echo) 2>/dev/null`" = "X-n $as_echo"; then
-    as_echo_body='eval /usr/ucb/echo -n "$1$as_nl"'
-    as_echo_n='/usr/ucb/echo -n'
-  else
-    as_echo_body='eval expr "X$1" : "X\\(.*\\)"'
-    as_echo_n_body='eval
-      arg=$1;
-      case $arg in #(
-      *"$as_nl"*)
-	expr "X$arg" : "X\\(.*\\)$as_nl";
-	arg=`expr "X$arg" : ".*$as_nl\\(.*\\)"`;;
-      esac;
-      expr "X$arg" : "X\\(.*\\)" | tr -d "$as_nl"
-    '
-    export as_echo_n_body
-    as_echo_n='sh -c $as_echo_n_body as_echo'
-  fi
-  export as_echo_body
-  as_echo='sh -c $as_echo_body as_echo'
-fi
+IFS=" ""	$as_nl"
+
+PS1='$ '
+PS2='> '
+PS4='+ '
+
+# Ensure predictable behavior from utilities with locale-dependent output.
+LC_ALL=C
+export LC_ALL
+LANGUAGE=C
+export LANGUAGE
+
+# We cannot yet rely on "unset" to work, but we need these variables
+# to be unset--not just set to an empty or harmless value--now, to
+# avoid bugs in old shells (e.g. pre-3.0 UWIN ksh).  This construct
+# also avoids known problems related to "unset" and subshell syntax
+# in other old shells (e.g. bash 2.01 and pdksh 5.2.14).
+for as_var in BASH_ENV ENV MAIL MAILPATH CDPATH
+do eval test \${$as_var+y} \
+  && ( (unset $as_var) || exit 1) >/dev/null 2>&1 && unset $as_var || :
+done
+
+# Ensure that fds 0, 1, and 2 are open.
+if (exec 3>&0) 2>/dev/null; then :; else exec 0</dev/null; fi
+if (exec 3>&1) 2>/dev/null; then :; else exec 1>/dev/null; fi
+if (exec 3>&2)            ; then :; else exec 2>/dev/null; fi
 
 # The user is always right.
-if test "${PATH_SEPARATOR+set}" != set; then
+if ${PATH_SEPARATOR+false} :; then
   PATH_SEPARATOR=:
   (PATH='/bin;/bin'; FPATH=$PATH; sh -c :) >/dev/null 2>&1 && {
     (PATH='/bin:/bin'; FPATH=$PATH; sh -c :) >/dev/null 2>&1 ||
@@ -79,13 +82,6 @@ if test "${PATH_SEPARATOR+set}" != set; then
 fi
 
 
-# IFS
-# We need space, tab and new line, in precisely that order.  Quoting is
-# there to prevent editors from complaining about space-tab.
-# (If _AS_PATH_WALK were called with IFS unset, it would disable word
-# splitting by setting IFS to empty value.)
-IFS=" ""	$as_nl"
-
 # Find who we are.  Look in the path if we contain no directory separator.
 as_myself=
 case $0 in #((
@@ -94,8 +90,12 @@ case $0 in #((
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
-    test -r "$as_dir/$0" && as_myself=$as_dir/$0 && break
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
+    test -r "$as_dir$0" && as_myself=$as_dir$0 && break
   done
 IFS=$as_save_IFS
 
@@ -107,30 +107,10 @@ if test "x$as_myself" = x; then
   as_myself=$0
 fi
 if test ! -f "$as_myself"; then
-  $as_echo "$as_myself: error: cannot find myself; rerun with an absolute file name" >&2
+  printf "%s\n" "$as_myself: error: cannot find myself; rerun with an absolute file name" >&2
   exit 1
 fi
 
-# Unset variables that we do not need and which cause bugs (e.g. in
-# pre-3.0 UWIN ksh).  But do not cause bugs in bash 2.01; the "|| exit 1"
-# suppresses any "Segmentation fault" message there.  '((' could
-# trigger a bug in pdksh 5.2.14.
-for as_var in BASH_ENV ENV MAIL MAILPATH
-do eval test x\${$as_var+set} = xset \
-  && ( (unset $as_var) || exit 1) >/dev/null 2>&1 && unset $as_var || :
-done
-PS1='$ '
-PS2='> '
-PS4='+ '
-
-# NLS nuisances.
-LC_ALL=C
-export LC_ALL
-LANGUAGE=C
-export LANGUAGE
-
-# CDPATH.
-(unset CDPATH) >/dev/null 2>&1 && unset CDPATH
 
 # Use a proper internal environment variable to ensure we don't fall
   # into an infinite loop, continuously re-executing ourselves.
@@ -152,20 +132,22 @@ esac
 exec $CONFIG_SHELL $as_opts "$as_myself" ${1+"$@"}
 # Admittedly, this is quite paranoid, since all the known shells bail
 # out after a failed `exec'.
-$as_echo "$0: could not re-execute with $CONFIG_SHELL" >&2
-as_fn_exit 255
+printf "%s\n" "$0: could not re-execute with $CONFIG_SHELL" >&2
+exit 255
   fi
   # We don't want this to propagate to other subprocesses.
           { _as_can_reexec=; unset _as_can_reexec;}
 if test "x$CONFIG_SHELL" = x; then
-  as_bourne_compatible="if test -n \"\${ZSH_VERSION+set}\" && (emulate sh) >/dev/null 2>&1; then :
+  as_bourne_compatible="as_nop=:
+if test \${ZSH_VERSION+y} && (emulate sh) >/dev/null 2>&1
+then :
   emulate sh
   NULLCMD=:
   # Pre-4.2 versions of Zsh do word splitting on \${1+\"\$@\"}, which
   # is contrary to our usage.  Disable this feature.
   alias -g '\${1+\"\$@\"}'='\"\$@\"'
   setopt NO_GLOB_SUBST
-else
+else \$as_nop
   case \`(set -o) 2>/dev/null\` in #(
   *posix*) :
     set -o posix ;; #(
@@ -185,41 +167,52 @@ as_fn_success || { exitcode=1; echo as_fn_success failed.; }
 as_fn_failure && { exitcode=1; echo as_fn_failure succeeded.; }
 as_fn_ret_success || { exitcode=1; echo as_fn_ret_success failed.; }
 as_fn_ret_failure && { exitcode=1; echo as_fn_ret_failure succeeded.; }
-if ( set x; as_fn_ret_success y && test x = \"\$1\" ); then :
+if ( set x; as_fn_ret_success y && test x = \"\$1\" )
+then :
 
-else
+else \$as_nop
   exitcode=1; echo positional parameters were not saved.
 fi
 test x\$exitcode = x0 || exit 1
+blah=\$(echo \$(echo blah))
+test x\"\$blah\" = xblah || exit 1
 test -x / || exit 1"
   as_suggested="  as_lineno_1=";as_suggested=$as_suggested$LINENO;as_suggested=$as_suggested" as_lineno_1a=\$LINENO
   as_lineno_2=";as_suggested=$as_suggested$LINENO;as_suggested=$as_suggested" as_lineno_2a=\$LINENO
   eval 'test \"x\$as_lineno_1'\$as_run'\" != \"x\$as_lineno_2'\$as_run'\" &&
   test \"x\`expr \$as_lineno_1'\$as_run' + 1\`\" = \"x\$as_lineno_2'\$as_run'\"' || exit 1"
-  if (eval "$as_required") 2>/dev/null; then :
+  if (eval "$as_required") 2>/dev/null
+then :
   as_have_required=yes
-else
+else $as_nop
   as_have_required=no
 fi
-  if test x$as_have_required = xyes && (eval "$as_suggested") 2>/dev/null; then :
+  if test x$as_have_required = xyes && (eval "$as_suggested") 2>/dev/null
+then :
 
-else
+else $as_nop
   as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 as_found=false
 for as_dir in /bin$PATH_SEPARATOR/usr/bin$PATH_SEPARATOR$PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
   as_found=:
   case $as_dir in #(
 	 /*)
 	   for as_base in sh bash ksh sh5; do
 	     # Try only shells that exist, to save several forks.
-	     as_shell=$as_dir/$as_base
+	     as_shell=$as_dir$as_base
 	     if { test -f "$as_shell" || test -f "$as_shell.exe"; } &&
-		    { $as_echo "$as_bourne_compatible""$as_required" | as_run=a "$as_shell"; } 2>/dev/null; then :
+		    as_run=a "$as_shell" -c "$as_bourne_compatible""$as_required" 2>/dev/null
+then :
   CONFIG_SHELL=$as_shell as_have_required=yes
-		   if { $as_echo "$as_bourne_compatible""$as_suggested" | as_run=a "$as_shell"; } 2>/dev/null; then :
+		   if as_run=a "$as_shell" -c "$as_bourne_compatible""$as_suggested" 2>/dev/null
+then :
   break 2
 fi
 fi
@@ -227,14 +220,21 @@ fi
        esac
   as_found=false
 done
-$as_found || { if { test -f "$SHELL" || test -f "$SHELL.exe"; } &&
-	      { $as_echo "$as_bourne_compatible""$as_required" | as_run=a "$SHELL"; } 2>/dev/null; then :
-  CONFIG_SHELL=$SHELL as_have_required=yes
-fi; }
 IFS=$as_save_IFS
+if $as_found
+then :
+
+else $as_nop
+  if { test -f "$SHELL" || test -f "$SHELL.exe"; } &&
+	      as_run=a "$SHELL" -c "$as_bourne_compatible""$as_required" 2>/dev/null
+then :
+  CONFIG_SHELL=$SHELL as_have_required=yes
+fi
+fi
 
 
-      if test "x$CONFIG_SHELL" != x; then :
+      if test "x$CONFIG_SHELL" != x
+then :
   export CONFIG_SHELL
              # We cannot yet assume a decent shell, so we have to provide a
 # neutralization value for shells without unset; and this also
@@ -252,18 +252,19 @@ esac
 exec $CONFIG_SHELL $as_opts "$as_myself" ${1+"$@"}
 # Admittedly, this is quite paranoid, since all the known shells bail
 # out after a failed `exec'.
-$as_echo "$0: could not re-execute with $CONFIG_SHELL" >&2
+printf "%s\n" "$0: could not re-execute with $CONFIG_SHELL" >&2
 exit 255
 fi
 
-    if test x$as_have_required = xno; then :
-  $as_echo "$0: This script requires a shell more modern than all"
-  $as_echo "$0: the shells that I found on your system."
-  if test x${ZSH_VERSION+set} = xset ; then
-    $as_echo "$0: In particular, zsh $ZSH_VERSION has bugs and should"
-    $as_echo "$0: be upgraded to zsh 4.3.4 or later."
+    if test x$as_have_required = xno
+then :
+  printf "%s\n" "$0: This script requires a shell more modern than all"
+  printf "%s\n" "$0: the shells that I found on your system."
+  if test ${ZSH_VERSION+y} ; then
+    printf "%s\n" "$0: In particular, zsh $ZSH_VERSION has bugs and should"
+    printf "%s\n" "$0: be upgraded to zsh 4.3.4 or later."
   else
-    $as_echo "$0: Please tell bug-autoconf@gnu.org about your system,
+    printf "%s\n" "$0: Please tell bug-autoconf@gnu.org about your system,
 $0: including any error possibly output before this
 $0: message. Then install a modern shell, or manually run
 $0: the script under such a shell if you do have one."
@@ -290,6 +291,7 @@ as_fn_unset ()
 }
 as_unset=as_fn_unset
 
+
 # as_fn_set_status STATUS
 # -----------------------
 # Set $? to STATUS, without forking.
@@ -307,6 +309,14 @@ as_fn_exit ()
   as_fn_set_status $1
   exit $1
 } # as_fn_exit
+# as_fn_nop
+# ---------
+# Do nothing but, unlike ":", preserve the value of $?.
+as_fn_nop ()
+{
+  return $?
+}
+as_nop=as_fn_nop
 
 # as_fn_mkdir_p
 # -------------
@@ -321,7 +331,7 @@ as_fn_mkdir_p ()
     as_dirs=
     while :; do
       case $as_dir in #(
-      *\'*) as_qdir=`$as_echo "$as_dir" | sed "s/'/'\\\\\\\\''/g"`;; #'(
+      *\'*) as_qdir=`printf "%s\n" "$as_dir" | sed "s/'/'\\\\\\\\''/g"`;; #'(
       *) as_qdir=$as_dir;;
       esac
       as_dirs="'$as_qdir' $as_dirs"
@@ -330,7 +340,7 @@ $as_expr X"$as_dir" : 'X\(.*[^/]\)//*[^/][^/]*/*$' \| \
 	 X"$as_dir" : 'X\(//\)[^/]' \| \
 	 X"$as_dir" : 'X\(//\)$' \| \
 	 X"$as_dir" : 'X\(/\)' \| . 2>/dev/null ||
-$as_echo X"$as_dir" |
+printf "%s\n" X"$as_dir" |
     sed '/^X\(.*[^/]\)\/\/*[^/][^/]*\/*$/{
 	    s//\1/
 	    q
@@ -369,12 +379,13 @@ as_fn_executable_p ()
 # advantage of any shell optimizations that allow amortized linear growth over
 # repeated appends, instead of the typical quadratic growth present in naive
 # implementations.
-if (eval "as_var=1; as_var+=2; test x\$as_var = x12") 2>/dev/null; then :
+if (eval "as_var=1; as_var+=2; test x\$as_var = x12") 2>/dev/null
+then :
   eval 'as_fn_append ()
   {
     eval $1+=\$2
   }'
-else
+else $as_nop
   as_fn_append ()
   {
     eval $1=\$$1\$2
@@ -386,18 +397,27 @@ fi # as_fn_append
 # Perform arithmetic evaluation on the ARGs, and store the result in the
 # global $as_val. Take advantage of shells that can avoid forks. The arguments
 # must be portable across $(()) and expr.
-if (eval "test \$(( 1 + 1 )) = 2") 2>/dev/null; then :
+if (eval "test \$(( 1 + 1 )) = 2") 2>/dev/null
+then :
   eval 'as_fn_arith ()
   {
     as_val=$(( $* ))
   }'
-else
+else $as_nop
   as_fn_arith ()
   {
     as_val=`expr "$@" || test $? -eq 1`
   }
 fi # as_fn_arith
 
+# as_fn_nop
+# ---------
+# Do nothing but, unlike ":", preserve the value of $?.
+as_fn_nop ()
+{
+  return $?
+}
+as_nop=as_fn_nop
 
 # as_fn_error STATUS ERROR [LINENO LOG_FD]
 # ----------------------------------------
@@ -409,9 +429,9 @@ as_fn_error ()
   as_status=$1; test $as_status -eq 0 && as_status=1
   if test "$4"; then
     as_lineno=${as_lineno-"$3"} as_lineno_stack=as_lineno_stack=$as_lineno_stack
-    $as_echo "$as_me:${as_lineno-$LINENO}: error: $2" >&$4
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: $2" >&$4
   fi
-  $as_echo "$as_me: error: $2" >&2
+  printf "%s\n" "$as_me: error: $2" >&2
   as_fn_exit $as_status
 } # as_fn_error
 
@@ -438,7 +458,7 @@ as_me=`$as_basename -- "$0" ||
 $as_expr X/"$0" : '.*/\([^/][^/]*\)/*$' \| \
 	 X"$0" : 'X\(//\)$' \| \
 	 X"$0" : 'X\(/\)' \| . 2>/dev/null ||
-$as_echo X/"$0" |
+printf "%s\n" X/"$0" |
     sed '/^.*\/\([^/][^/]*\)\/*$/{
 	    s//\1/
 	    q
@@ -482,7 +502,7 @@ as_cr_alnum=$as_cr_Letters$as_cr_digits
       s/-\n.*//
     ' >$as_me.lineno &&
   chmod +x "$as_me.lineno" ||
-    { $as_echo "$as_me: error: cannot create $as_me.lineno; rerun with a POSIX shell" >&2; as_fn_exit 1; }
+    { printf "%s\n" "$as_me: error: cannot create $as_me.lineno; rerun with a POSIX shell" >&2; as_fn_exit 1; }
 
   # If we had to re-execute with $CONFIG_SHELL, we're ensured to have
   # already done that, so ensure we don't try to do so again and fall
@@ -496,6 +516,10 @@ as_cr_alnum=$as_cr_Letters$as_cr_digits
   exit
 }
 
+
+# Determine whether it's possible to make 'echo' print without a newline.
+# These variables are no longer used directly by Autoconf, but are AC_SUBSTed
+# for compatibility with existing Makefiles.
 ECHO_C= ECHO_N= ECHO_T=
 case `echo -n x` in #(((((
 -n*)
@@ -509,6 +533,13 @@ case `echo -n x` in #(((((
   ECHO_N='-n';;
 esac
 
+# For backward compatibility with old third-party macros, we provide
+# the shell variables $as_echo and $as_echo_n.  New code should use
+# AS_ECHO(["message"]) and AS_ECHO_N(["message"]), respectively.
+as_echo='printf %s\n'
+as_echo_n='printf %s'
+
+
 rm -f conf$$ conf$$.exe conf$$.file
 if test -d conf$$.dir; then
   rm -f conf$$.dir/conf$$.file
@@ -596,6 +627,7 @@ NVCC
 CC
 MPS_CMD_LIB
 MPS_COMPAT_LIB
+MFA_COMPAT_LIB
 CUDA_CMD_LIB
 CUDA_COMPAT_LIB
 CUDA_SRCS
@@ -729,8 +761,6 @@ do
   *)    ac_optarg=yes ;;
   esac
 
-  # Accept the important Cygnus configure options, so we can diagnose typos.
-
   case $ac_dashdash$ac_option in
   --)
     ac_dashdash=yes ;;
@@ -771,9 +801,9 @@ do
     ac_useropt=`expr "x$ac_option" : 'x-*disable-\(.*\)'`
     # Reject names that are not valid shell variable names.
     expr "x$ac_useropt" : ".*[^-+._$as_cr_alnum]" >/dev/null &&
-      as_fn_error $? "invalid feature name: $ac_useropt"
+      as_fn_error $? "invalid feature name: \`$ac_useropt'"
     ac_useropt_orig=$ac_useropt
-    ac_useropt=`$as_echo "$ac_useropt" | sed 's/[-+.]/_/g'`
+    ac_useropt=`printf "%s\n" "$ac_useropt" | sed 's/[-+.]/_/g'`
     case $ac_user_opts in
       *"
 "enable_$ac_useropt"
@@ -797,9 +827,9 @@ do
     ac_useropt=`expr "x$ac_option" : 'x-*enable-\([^=]*\)'`
     # Reject names that are not valid shell variable names.
     expr "x$ac_useropt" : ".*[^-+._$as_cr_alnum]" >/dev/null &&
-      as_fn_error $? "invalid feature name: $ac_useropt"
+      as_fn_error $? "invalid feature name: \`$ac_useropt'"
     ac_useropt_orig=$ac_useropt
-    ac_useropt=`$as_echo "$ac_useropt" | sed 's/[-+.]/_/g'`
+    ac_useropt=`printf "%s\n" "$ac_useropt" | sed 's/[-+.]/_/g'`
     case $ac_user_opts in
       *"
 "enable_$ac_useropt"
@@ -1010,9 +1040,9 @@ do
     ac_useropt=`expr "x$ac_option" : 'x-*with-\([^=]*\)'`
     # Reject names that are not valid shell variable names.
     expr "x$ac_useropt" : ".*[^-+._$as_cr_alnum]" >/dev/null &&
-      as_fn_error $? "invalid package name: $ac_useropt"
+      as_fn_error $? "invalid package name: \`$ac_useropt'"
     ac_useropt_orig=$ac_useropt
-    ac_useropt=`$as_echo "$ac_useropt" | sed 's/[-+.]/_/g'`
+    ac_useropt=`printf "%s\n" "$ac_useropt" | sed 's/[-+.]/_/g'`
     case $ac_user_opts in
       *"
 "with_$ac_useropt"
@@ -1026,9 +1056,9 @@ do
     ac_useropt=`expr "x$ac_option" : 'x-*without-\(.*\)'`
     # Reject names that are not valid shell variable names.
     expr "x$ac_useropt" : ".*[^-+._$as_cr_alnum]" >/dev/null &&
-      as_fn_error $? "invalid package name: $ac_useropt"
+      as_fn_error $? "invalid package name: \`$ac_useropt'"
     ac_useropt_orig=$ac_useropt
-    ac_useropt=`$as_echo "$ac_useropt" | sed 's/[-+.]/_/g'`
+    ac_useropt=`printf "%s\n" "$ac_useropt" | sed 's/[-+.]/_/g'`
     case $ac_user_opts in
       *"
 "with_$ac_useropt"
@@ -1072,9 +1102,9 @@ Try \`$0 --help' for more information"
 
   *)
     # FIXME: should be removed in autoconf 3.0.
-    $as_echo "$as_me: WARNING: you should use --build, --host, --target" >&2
+    printf "%s\n" "$as_me: WARNING: you should use --build, --host, --target" >&2
     expr "x$ac_option" : ".*[^-._$as_cr_alnum]" >/dev/null &&
-      $as_echo "$as_me: WARNING: invalid host type: $ac_option" >&2
+      printf "%s\n" "$as_me: WARNING: invalid host type: $ac_option" >&2
     : "${build_alias=$ac_option} ${host_alias=$ac_option} ${target_alias=$ac_option}"
     ;;
 
@@ -1090,7 +1120,7 @@ if test -n "$ac_unrecognized_opts"; then
   case $enable_option_checking in
     no) ;;
     fatal) as_fn_error $? "unrecognized options: $ac_unrecognized_opts" ;;
-    *)     $as_echo "$as_me: WARNING: unrecognized options: $ac_unrecognized_opts" >&2 ;;
+    *)     printf "%s\n" "$as_me: WARNING: unrecognized options: $ac_unrecognized_opts" >&2 ;;
   esac
 fi
 
@@ -1154,7 +1184,7 @@ $as_expr X"$as_myself" : 'X\(.*[^/]\)//*[^/][^/]*/*$' \| \
 	 X"$as_myself" : 'X\(//\)[^/]' \| \
 	 X"$as_myself" : 'X\(//\)$' \| \
 	 X"$as_myself" : 'X\(/\)' \| . 2>/dev/null ||
-$as_echo X"$as_myself" |
+printf "%s\n" X"$as_myself" |
     sed '/^X\(.*[^/]\)\/\/*[^/][^/]*\/*$/{
 	    s//\1/
 	    q
@@ -1322,9 +1352,9 @@ if test "$ac_init_help" = "recursive"; then
 case "$ac_dir" in
 .) ac_dir_suffix= ac_top_builddir_sub=. ac_top_build_prefix= ;;
 *)
-  ac_dir_suffix=/`$as_echo "$ac_dir" | sed 's|^\.[\\/]||'`
+  ac_dir_suffix=/`printf "%s\n" "$ac_dir" | sed 's|^\.[\\/]||'`
   # A ".." for each directory in $ac_dir_suffix.
-  ac_top_builddir_sub=`$as_echo "$ac_dir_suffix" | sed 's|/[^\\/]*|/..|g;s|/||'`
+  ac_top_builddir_sub=`printf "%s\n" "$ac_dir_suffix" | sed 's|/[^\\/]*|/..|g;s|/||'`
   case $ac_top_builddir_sub in
   "") ac_top_builddir_sub=. ac_top_build_prefix= ;;
   *)  ac_top_build_prefix=$ac_top_builddir_sub/ ;;
@@ -1352,7 +1382,8 @@ esac
 ac_abs_srcdir=$ac_abs_top_srcdir$ac_dir_suffix
 
     cd "$ac_dir" || { ac_status=$?; continue; }
-    # Check for guested configure.
+    # Check for configure.gnu first; this name is used for a wrapper for
+    # Metaconfig's "Configure" on case-insensitive file systems.
     if test -f "$ac_srcdir/configure.gnu"; then
       echo &&
       $SHELL "$ac_srcdir/configure.gnu" --help=recursive
@@ -1360,7 +1391,7 @@ ac_abs_srcdir=$ac_abs_top_srcdir$ac_dir_suffix
       echo &&
       $SHELL "$ac_srcdir/configure" --help=recursive
     else
-      $as_echo "$as_me: WARNING: no configuration information is in $ac_dir" >&2
+      printf "%s\n" "$as_me: WARNING: no configuration information is in $ac_dir" >&2
     fi || ac_status=$?
     cd "$ac_pwd" || { ac_status=$?; break; }
   done
@@ -1370,9 +1401,9 @@ test -n "$ac_init_help" && exit $ac_status
 if $ac_init_version; then
   cat <<\_ACEOF
 libccv configure 0.7
-generated by GNU Autoconf 2.69
+generated by GNU Autoconf 2.71
 
-Copyright (C) 2012 Free Software Foundation, Inc.
+Copyright (C) 2021 Free Software Foundation, Inc.
 This configure script is free software; the Free Software Foundation
 gives unlimited permission to copy, distribute and modify it.
 _ACEOF
@@ -1389,14 +1420,14 @@ fi
 ac_fn_c_try_compile ()
 {
   as_lineno=${as_lineno-"$1"} as_lineno_stack=as_lineno_stack=$as_lineno_stack
-  rm -f conftest.$ac_objext
+  rm -f conftest.$ac_objext conftest.beam
   if { { ac_try="$ac_compile"
 case "(($ac_try" in
   *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;;
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_compile") 2>conftest.err
   ac_status=$?
   if test -s conftest.err; then
@@ -1404,14 +1435,15 @@ $as_echo "$ac_try_echo"; } >&5
     cat conftest.er1 >&5
     mv -f conftest.er1 conftest.err
   fi
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
   test $ac_status = 0; } && {
 	 test -z "$ac_c_werror_flag" ||
 	 test ! -s conftest.err
-       } && test -s conftest.$ac_objext; then :
+       } && test -s conftest.$ac_objext
+then :
   ac_retval=0
-else
-  $as_echo "$as_me: failed program was:" >&5
+else $as_nop
+  printf "%s\n" "$as_me: failed program was:" >&5
 sed 's/^/| /' conftest.$ac_ext >&5
 
 	ac_retval=1
@@ -1427,14 +1459,14 @@ fi
 ac_fn_c_try_link ()
 {
   as_lineno=${as_lineno-"$1"} as_lineno_stack=as_lineno_stack=$as_lineno_stack
-  rm -f conftest.$ac_objext conftest$ac_exeext
+  rm -f conftest.$ac_objext conftest.beam conftest$ac_exeext
   if { { ac_try="$ac_link"
 case "(($ac_try" in
   *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;;
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_link") 2>conftest.err
   ac_status=$?
   if test -s conftest.err; then
@@ -1442,17 +1474,18 @@ $as_echo "$ac_try_echo"; } >&5
     cat conftest.er1 >&5
     mv -f conftest.er1 conftest.err
   fi
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
   test $ac_status = 0; } && {
 	 test -z "$ac_c_werror_flag" ||
 	 test ! -s conftest.err
        } && test -s conftest$ac_exeext && {
 	 test "$cross_compiling" = yes ||
 	 test -x conftest$ac_exeext
-       }; then :
+       }
+then :
   ac_retval=0
-else
-  $as_echo "$as_me: failed program was:" >&5
+else $as_nop
+  printf "%s\n" "$as_me: failed program was:" >&5
 sed 's/^/| /' conftest.$ac_ext >&5
 
 	ac_retval=1
@@ -1479,7 +1512,7 @@ case "(($ac_try" in
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_cpp conftest.$ac_ext") 2>conftest.err
   ac_status=$?
   if test -s conftest.err; then
@@ -1487,14 +1520,15 @@ $as_echo "$ac_try_echo"; } >&5
     cat conftest.er1 >&5
     mv -f conftest.er1 conftest.err
   fi
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
   test $ac_status = 0; } > conftest.i && {
 	 test -z "$ac_c_preproc_warn_flag$ac_c_werror_flag" ||
 	 test ! -s conftest.err
-       }; then :
+       }
+then :
   ac_retval=0
-else
-  $as_echo "$as_me: failed program was:" >&5
+else $as_nop
+  printf "%s\n" "$as_me: failed program was:" >&5
 sed 's/^/| /' conftest.$ac_ext >&5
 
     ac_retval=1
@@ -1503,14 +1537,34 @@ fi
   as_fn_set_status $ac_retval
 
 } # ac_fn_c_try_cpp
+ac_configure_args_raw=
+for ac_arg
+do
+  case $ac_arg in
+  *\'*)
+    ac_arg=`printf "%s\n" "$ac_arg" | sed "s/'/'\\\\\\\\''/g"` ;;
+  esac
+  as_fn_append ac_configure_args_raw " '$ac_arg'"
+done
+
+case $ac_configure_args_raw in
+  *$as_nl*)
+    ac_safe_unquote= ;;
+  *)
+    ac_unsafe_z='|&;<>()$`\\"*?[ ''	' # This string ends in space, tab.
+    ac_unsafe_a="$ac_unsafe_z#~"
+    ac_safe_unquote="s/ '\\([^$ac_unsafe_a][^$ac_unsafe_z]*\\)'/ \\1/g"
+    ac_configure_args_raw=`      printf "%s\n" "$ac_configure_args_raw" | sed "$ac_safe_unquote"`;;
+esac
+
 cat >config.log <<_ACEOF
 This file contains any messages produced by compilers while
 running configure, to aid debugging if configure makes a mistake.
 
 It was created by libccv $as_me 0.7, which was
-generated by GNU Autoconf 2.69.  Invocation command line was
+generated by GNU Autoconf 2.71.  Invocation command line was
 
-  $ $0 $@
+  $ $0$ac_configure_args_raw
 
 _ACEOF
 exec 5>>config.log
@@ -1543,8 +1597,12 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
-    $as_echo "PATH: $as_dir"
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
+    printf "%s\n" "PATH: $as_dir"
   done
 IFS=$as_save_IFS
 
@@ -1579,7 +1637,7 @@ do
     | -silent | --silent | --silen | --sile | --sil)
       continue ;;
     *\'*)
-      ac_arg=`$as_echo "$ac_arg" | sed "s/'/'\\\\\\\\''/g"` ;;
+      ac_arg=`printf "%s\n" "$ac_arg" | sed "s/'/'\\\\\\\\''/g"` ;;
     esac
     case $ac_pass in
     1) as_fn_append ac_configure_args0 " '$ac_arg'" ;;
@@ -1614,11 +1672,13 @@ done
 # WARNING: Use '\'' to represent an apostrophe within the trap.
 # WARNING: Do not start the trap code with a newline, due to a FreeBSD 4.0 bug.
 trap 'exit_status=$?
+  # Sanitize IFS.
+  IFS=" ""	$as_nl"
   # Save into config.log some information that might help in debugging.
   {
     echo
 
-    $as_echo "## ---------------- ##
+    printf "%s\n" "## ---------------- ##
 ## Cache variables. ##
 ## ---------------- ##"
     echo
@@ -1629,8 +1689,8 @@ trap 'exit_status=$?
     case $ac_val in #(
     *${as_nl}*)
       case $ac_var in #(
-      *_cv_*) { $as_echo "$as_me:${as_lineno-$LINENO}: WARNING: cache variable $ac_var contains a newline" >&5
-$as_echo "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
+      *_cv_*) { printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: cache variable $ac_var contains a newline" >&5
+printf "%s\n" "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
       esac
       case $ac_var in #(
       _ | IFS | as_nl) ;; #(
@@ -1654,7 +1714,7 @@ $as_echo "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
 )
     echo
 
-    $as_echo "## ----------------- ##
+    printf "%s\n" "## ----------------- ##
 ## Output variables. ##
 ## ----------------- ##"
     echo
@@ -1662,14 +1722,14 @@ $as_echo "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
     do
       eval ac_val=\$$ac_var
       case $ac_val in
-      *\'\''*) ac_val=`$as_echo "$ac_val" | sed "s/'\''/'\''\\\\\\\\'\'''\''/g"`;;
+      *\'\''*) ac_val=`printf "%s\n" "$ac_val" | sed "s/'\''/'\''\\\\\\\\'\'''\''/g"`;;
       esac
-      $as_echo "$ac_var='\''$ac_val'\''"
+      printf "%s\n" "$ac_var='\''$ac_val'\''"
     done | sort
     echo
 
     if test -n "$ac_subst_files"; then
-      $as_echo "## ------------------- ##
+      printf "%s\n" "## ------------------- ##
 ## File substitutions. ##
 ## ------------------- ##"
       echo
@@ -1677,15 +1737,15 @@ $as_echo "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
       do
 	eval ac_val=\$$ac_var
 	case $ac_val in
-	*\'\''*) ac_val=`$as_echo "$ac_val" | sed "s/'\''/'\''\\\\\\\\'\'''\''/g"`;;
+	*\'\''*) ac_val=`printf "%s\n" "$ac_val" | sed "s/'\''/'\''\\\\\\\\'\'''\''/g"`;;
 	esac
-	$as_echo "$ac_var='\''$ac_val'\''"
+	printf "%s\n" "$ac_var='\''$ac_val'\''"
       done | sort
       echo
     fi
 
     if test -s confdefs.h; then
-      $as_echo "## ----------- ##
+      printf "%s\n" "## ----------- ##
 ## confdefs.h. ##
 ## ----------- ##"
       echo
@@ -1693,8 +1753,8 @@ $as_echo "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
       echo
     fi
     test "$ac_signal" != 0 &&
-      $as_echo "$as_me: caught signal $ac_signal"
-    $as_echo "$as_me: exit $exit_status"
+      printf "%s\n" "$as_me: caught signal $ac_signal"
+    printf "%s\n" "$as_me: exit $exit_status"
   } >&5
   rm -f core *.core core.conftest.* &&
     rm -f -r conftest* confdefs* conf$$* $ac_clean_files &&
@@ -1708,63 +1768,48 @@ ac_signal=0
 # confdefs.h avoids OS command line length limits that DEFS can exceed.
 rm -f -r conftest* confdefs.h
 
-$as_echo "/* confdefs.h */" > confdefs.h
+printf "%s\n" "/* confdefs.h */" > confdefs.h
 
 # Predefined preprocessor variables.
 
-cat >>confdefs.h <<_ACEOF
-#define PACKAGE_NAME "$PACKAGE_NAME"
-_ACEOF
+printf "%s\n" "#define PACKAGE_NAME \"$PACKAGE_NAME\"" >>confdefs.h
 
-cat >>confdefs.h <<_ACEOF
-#define PACKAGE_TARNAME "$PACKAGE_TARNAME"
-_ACEOF
+printf "%s\n" "#define PACKAGE_TARNAME \"$PACKAGE_TARNAME\"" >>confdefs.h
 
-cat >>confdefs.h <<_ACEOF
-#define PACKAGE_VERSION "$PACKAGE_VERSION"
-_ACEOF
+printf "%s\n" "#define PACKAGE_VERSION \"$PACKAGE_VERSION\"" >>confdefs.h
 
-cat >>confdefs.h <<_ACEOF
-#define PACKAGE_STRING "$PACKAGE_STRING"
-_ACEOF
+printf "%s\n" "#define PACKAGE_STRING \"$PACKAGE_STRING\"" >>confdefs.h
 
-cat >>confdefs.h <<_ACEOF
-#define PACKAGE_BUGREPORT "$PACKAGE_BUGREPORT"
-_ACEOF
+printf "%s\n" "#define PACKAGE_BUGREPORT \"$PACKAGE_BUGREPORT\"" >>confdefs.h
 
-cat >>confdefs.h <<_ACEOF
-#define PACKAGE_URL "$PACKAGE_URL"
-_ACEOF
+printf "%s\n" "#define PACKAGE_URL \"$PACKAGE_URL\"" >>confdefs.h
 
 
 # Let the site file select an alternate cache file if it wants to.
 # Prefer an explicitly selected file to automatically selected ones.
-ac_site_file1=NONE
-ac_site_file2=NONE
 if test -n "$CONFIG_SITE"; then
-  # We do not want a PATH search for config.site.
-  case $CONFIG_SITE in #((
-    -*)  ac_site_file1=./$CONFIG_SITE;;
-    */*) ac_site_file1=$CONFIG_SITE;;
-    *)   ac_site_file1=./$CONFIG_SITE;;
-  esac
+  ac_site_files="$CONFIG_SITE"
 elif test "x$prefix" != xNONE; then
-  ac_site_file1=$prefix/share/config.site
-  ac_site_file2=$prefix/etc/config.site
+  ac_site_files="$prefix/share/config.site $prefix/etc/config.site"
 else
-  ac_site_file1=$ac_default_prefix/share/config.site
-  ac_site_file2=$ac_default_prefix/etc/config.site
+  ac_site_files="$ac_default_prefix/share/config.site $ac_default_prefix/etc/config.site"
 fi
-for ac_site_file in "$ac_site_file1" "$ac_site_file2"
+
+for ac_site_file in $ac_site_files
 do
-  test "x$ac_site_file" = xNONE && continue
-  if test /dev/null != "$ac_site_file" && test -r "$ac_site_file"; then
-    { $as_echo "$as_me:${as_lineno-$LINENO}: loading site script $ac_site_file" >&5
-$as_echo "$as_me: loading site script $ac_site_file" >&6;}
+  case $ac_site_file in #(
+  */*) :
+     ;; #(
+  *) :
+    ac_site_file=./$ac_site_file ;;
+esac
+  if test -f "$ac_site_file" && test -r "$ac_site_file"; then
+    { printf "%s\n" "$as_me:${as_lineno-$LINENO}: loading site script $ac_site_file" >&5
+printf "%s\n" "$as_me: loading site script $ac_site_file" >&6;}
     sed 's/^/| /' "$ac_site_file" >&5
     . "$ac_site_file" \
-      || { { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
+      || { { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
 as_fn_error $? "failed to load site script $ac_site_file
 See \`config.log' for more details" "$LINENO" 5; }
   fi
@@ -1774,19 +1819,327 @@ if test -r "$cache_file"; then
   # Some versions of bash will fail to source /dev/null (special files
   # actually), so we avoid doing that.  DJGPP emulates it as a regular file.
   if test /dev/null != "$cache_file" && test -f "$cache_file"; then
-    { $as_echo "$as_me:${as_lineno-$LINENO}: loading cache $cache_file" >&5
-$as_echo "$as_me: loading cache $cache_file" >&6;}
+    { printf "%s\n" "$as_me:${as_lineno-$LINENO}: loading cache $cache_file" >&5
+printf "%s\n" "$as_me: loading cache $cache_file" >&6;}
     case $cache_file in
       [\\/]* | ?:[\\/]* ) . "$cache_file";;
       *)                      . "./$cache_file";;
     esac
   fi
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: creating cache $cache_file" >&5
-$as_echo "$as_me: creating cache $cache_file" >&6;}
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: creating cache $cache_file" >&5
+printf "%s\n" "$as_me: creating cache $cache_file" >&6;}
   >$cache_file
 fi
 
+# Test code for whether the C compiler supports C89 (global declarations)
+ac_c_conftest_c89_globals='
+/* Does the compiler advertise C89 conformance?
+   Do not test the value of __STDC__, because some compilers set it to 0
+   while being otherwise adequately conformant. */
+#if !defined __STDC__
+# error "Compiler does not advertise C89 conformance"
+#endif
+
+#include <stddef.h>
+#include <stdarg.h>
+struct stat;
+/* Most of the following tests are stolen from RCS 5.7 src/conf.sh.  */
+struct buf { int x; };
+struct buf * (*rcsopen) (struct buf *, struct stat *, int);
+static char *e (p, i)
+     char **p;
+     int i;
+{
+  return p[i];
+}
+static char *f (char * (*g) (char **, int), char **p, ...)
+{
+  char *s;
+  va_list v;
+  va_start (v,p);
+  s = g (p, va_arg (v,int));
+  va_end (v);
+  return s;
+}
+
+/* OSF 4.0 Compaq cc is some sort of almost-ANSI by default.  It has
+   function prototypes and stuff, but not \xHH hex character constants.
+   These do not provoke an error unfortunately, instead are silently treated
+   as an "x".  The following induces an error, until -std is added to get
+   proper ANSI mode.  Curiously \x00 != x always comes out true, for an
+   array size at least.  It is necessary to write \x00 == 0 to get something
+   that is true only with -std.  */
+int osf4_cc_array ['\''\x00'\'' == 0 ? 1 : -1];
+
+/* IBM C 6 for AIX is almost-ANSI by default, but it replaces macro parameters
+   inside strings and character constants.  */
+#define FOO(x) '\''x'\''
+int xlc6_cc_array[FOO(a) == '\''x'\'' ? 1 : -1];
+
+int test (int i, double x);
+struct s1 {int (*f) (int a);};
+struct s2 {int (*f) (double a);};
+int pairnames (int, char **, int *(*)(struct buf *, struct stat *, int),
+               int, int);'
+
+# Test code for whether the C compiler supports C89 (body of main).
+ac_c_conftest_c89_main='
+ok |= (argc == 0 || f (e, argv, 0) != argv[0] || f (e, argv, 1) != argv[1]);
+'
+
+# Test code for whether the C compiler supports C99 (global declarations)
+ac_c_conftest_c99_globals='
+// Does the compiler advertise C99 conformance?
+#if !defined __STDC_VERSION__ || __STDC_VERSION__ < 199901L
+# error "Compiler does not advertise C99 conformance"
+#endif
+
+#include <stdbool.h>
+extern int puts (const char *);
+extern int printf (const char *, ...);
+extern int dprintf (int, const char *, ...);
+extern void *malloc (size_t);
+
+// Check varargs macros.  These examples are taken from C99 6.10.3.5.
+// dprintf is used instead of fprintf to avoid needing to declare
+// FILE and stderr.
+#define debug(...) dprintf (2, __VA_ARGS__)
+#define showlist(...) puts (#__VA_ARGS__)
+#define report(test,...) ((test) ? puts (#test) : printf (__VA_ARGS__))
+static void
+test_varargs_macros (void)
+{
+  int x = 1234;
+  int y = 5678;
+  debug ("Flag");
+  debug ("X = %d\n", x);
+  showlist (The first, second, and third items.);
+  report (x>y, "x is %d but y is %d", x, y);
+}
+
+// Check long long types.
+#define BIG64 18446744073709551615ull
+#define BIG32 4294967295ul
+#define BIG_OK (BIG64 / BIG32 == 4294967297ull && BIG64 % BIG32 == 0)
+#if !BIG_OK
+  #error "your preprocessor is broken"
+#endif
+#if BIG_OK
+#else
+  #error "your preprocessor is broken"
+#endif
+static long long int bignum = -9223372036854775807LL;
+static unsigned long long int ubignum = BIG64;
+
+struct incomplete_array
+{
+  int datasize;
+  double data[];
+};
+
+struct named_init {
+  int number;
+  const wchar_t *name;
+  double average;
+};
+
+typedef const char *ccp;
+
+static inline int
+test_restrict (ccp restrict text)
+{
+  // See if C++-style comments work.
+  // Iterate through items via the restricted pointer.
+  // Also check for declarations in for loops.
+  for (unsigned int i = 0; *(text+i) != '\''\0'\''; ++i)
+    continue;
+  return 0;
+}
+
+// Check varargs and va_copy.
+static bool
+test_varargs (const char *format, ...)
+{
+  va_list args;
+  va_start (args, format);
+  va_list args_copy;
+  va_copy (args_copy, args);
+
+  const char *str = "";
+  int number = 0;
+  float fnumber = 0;
+
+  while (*format)
+    {
+      switch (*format++)
+	{
+	case '\''s'\'': // string
+	  str = va_arg (args_copy, const char *);
+	  break;
+	case '\''d'\'': // int
+	  number = va_arg (args_copy, int);
+	  break;
+	case '\''f'\'': // float
+	  fnumber = va_arg (args_copy, double);
+	  break;
+	default:
+	  break;
+	}
+    }
+  va_end (args_copy);
+  va_end (args);
+
+  return *str && number && fnumber;
+}
+'
+
+# Test code for whether the C compiler supports C99 (body of main).
+ac_c_conftest_c99_main='
+  // Check bool.
+  _Bool success = false;
+  success |= (argc != 0);
+
+  // Check restrict.
+  if (test_restrict ("String literal") == 0)
+    success = true;
+  char *restrict newvar = "Another string";
+
+  // Check varargs.
+  success &= test_varargs ("s, d'\'' f .", "string", 65, 34.234);
+  test_varargs_macros ();
+
+  // Check flexible array members.
+  struct incomplete_array *ia =
+    malloc (sizeof (struct incomplete_array) + (sizeof (double) * 10));
+  ia->datasize = 10;
+  for (int i = 0; i < ia->datasize; ++i)
+    ia->data[i] = i * 1.234;
+
+  // Check named initializers.
+  struct named_init ni = {
+    .number = 34,
+    .name = L"Test wide string",
+    .average = 543.34343,
+  };
+
+  ni.number = 58;
+
+  int dynamic_array[ni.number];
+  dynamic_array[0] = argv[0][0];
+  dynamic_array[ni.number - 1] = 543;
+
+  // work around unused variable warnings
+  ok |= (!success || bignum == 0LL || ubignum == 0uLL || newvar[0] == '\''x'\''
+	 || dynamic_array[ni.number - 1] != 543);
+'
+
+# Test code for whether the C compiler supports C11 (global declarations)
+ac_c_conftest_c11_globals='
+// Does the compiler advertise C11 conformance?
+#if !defined __STDC_VERSION__ || __STDC_VERSION__ < 201112L
+# error "Compiler does not advertise C11 conformance"
+#endif
+
+// Check _Alignas.
+char _Alignas (double) aligned_as_double;
+char _Alignas (0) no_special_alignment;
+extern char aligned_as_int;
+char _Alignas (0) _Alignas (int) aligned_as_int;
+
+// Check _Alignof.
+enum
+{
+  int_alignment = _Alignof (int),
+  int_array_alignment = _Alignof (int[100]),
+  char_alignment = _Alignof (char)
+};
+_Static_assert (0 < -_Alignof (int), "_Alignof is signed");
+
+// Check _Noreturn.
+int _Noreturn does_not_return (void) { for (;;) continue; }
+
+// Check _Static_assert.
+struct test_static_assert
+{
+  int x;
+  _Static_assert (sizeof (int) <= sizeof (long int),
+                  "_Static_assert does not work in struct");
+  long int y;
+};
+
+// Check UTF-8 literals.
+#define u8 syntax error!
+char const utf8_literal[] = u8"happens to be ASCII" "another string";
+
+// Check duplicate typedefs.
+typedef long *long_ptr;
+typedef long int *long_ptr;
+typedef long_ptr long_ptr;
+
+// Anonymous structures and unions -- taken from C11 6.7.2.1 Example 1.
+struct anonymous
+{
+  union {
+    struct { int i; int j; };
+    struct { int k; long int l; } w;
+  };
+  int m;
+} v1;
+'
+
+# Test code for whether the C compiler supports C11 (body of main).
+ac_c_conftest_c11_main='
+  _Static_assert ((offsetof (struct anonymous, i)
+		   == offsetof (struct anonymous, w.k)),
+		  "Anonymous union alignment botch");
+  v1.i = 2;
+  v1.w.k = 5;
+  ok |= v1.i != 5;
+'
+
+# Test code for whether the C compiler supports C11 (complete).
+ac_c_conftest_c11_program="${ac_c_conftest_c89_globals}
+${ac_c_conftest_c99_globals}
+${ac_c_conftest_c11_globals}
+
+int
+main (int argc, char **argv)
+{
+  int ok = 0;
+  ${ac_c_conftest_c89_main}
+  ${ac_c_conftest_c99_main}
+  ${ac_c_conftest_c11_main}
+  return ok;
+}
+"
+
+# Test code for whether the C compiler supports C99 (complete).
+ac_c_conftest_c99_program="${ac_c_conftest_c89_globals}
+${ac_c_conftest_c99_globals}
+
+int
+main (int argc, char **argv)
+{
+  int ok = 0;
+  ${ac_c_conftest_c89_main}
+  ${ac_c_conftest_c99_main}
+  return ok;
+}
+"
+
+# Test code for whether the C compiler supports C89 (complete).
+ac_c_conftest_c89_program="${ac_c_conftest_c89_globals}
+
+int
+main (int argc, char **argv)
+{
+  int ok = 0;
+  ${ac_c_conftest_c89_main}
+  return ok;
+}
+"
+
 # Check that the precious variables saved in the cache have kept the same
 # value.
 ac_cache_corrupted=false
@@ -1797,12 +2150,12 @@ for ac_var in $ac_precious_vars; do
   eval ac_new_val=\$ac_env_${ac_var}_value
   case $ac_old_set,$ac_new_set in
     set,)
-      { $as_echo "$as_me:${as_lineno-$LINENO}: error: \`$ac_var' was set to \`$ac_old_val' in the previous run" >&5
-$as_echo "$as_me: error: \`$ac_var' was set to \`$ac_old_val' in the previous run" >&2;}
+      { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: \`$ac_var' was set to \`$ac_old_val' in the previous run" >&5
+printf "%s\n" "$as_me: error: \`$ac_var' was set to \`$ac_old_val' in the previous run" >&2;}
       ac_cache_corrupted=: ;;
     ,set)
-      { $as_echo "$as_me:${as_lineno-$LINENO}: error: \`$ac_var' was not set in the previous run" >&5
-$as_echo "$as_me: error: \`$ac_var' was not set in the previous run" >&2;}
+      { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: \`$ac_var' was not set in the previous run" >&5
+printf "%s\n" "$as_me: error: \`$ac_var' was not set in the previous run" >&2;}
       ac_cache_corrupted=: ;;
     ,);;
     *)
@@ -1811,24 +2164,24 @@ $as_echo "$as_me: error: \`$ac_var' was not set in the previous run" >&2;}
 	ac_old_val_w=`echo x $ac_old_val`
 	ac_new_val_w=`echo x $ac_new_val`
 	if test "$ac_old_val_w" != "$ac_new_val_w"; then
-	  { $as_echo "$as_me:${as_lineno-$LINENO}: error: \`$ac_var' has changed since the previous run:" >&5
-$as_echo "$as_me: error: \`$ac_var' has changed since the previous run:" >&2;}
+	  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: \`$ac_var' has changed since the previous run:" >&5
+printf "%s\n" "$as_me: error: \`$ac_var' has changed since the previous run:" >&2;}
 	  ac_cache_corrupted=:
 	else
-	  { $as_echo "$as_me:${as_lineno-$LINENO}: warning: ignoring whitespace changes in \`$ac_var' since the previous run:" >&5
-$as_echo "$as_me: warning: ignoring whitespace changes in \`$ac_var' since the previous run:" >&2;}
+	  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: warning: ignoring whitespace changes in \`$ac_var' since the previous run:" >&5
+printf "%s\n" "$as_me: warning: ignoring whitespace changes in \`$ac_var' since the previous run:" >&2;}
 	  eval $ac_var=\$ac_old_val
 	fi
-	{ $as_echo "$as_me:${as_lineno-$LINENO}:   former value:  \`$ac_old_val'" >&5
-$as_echo "$as_me:   former value:  \`$ac_old_val'" >&2;}
-	{ $as_echo "$as_me:${as_lineno-$LINENO}:   current value: \`$ac_new_val'" >&5
-$as_echo "$as_me:   current value: \`$ac_new_val'" >&2;}
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}:   former value:  \`$ac_old_val'" >&5
+printf "%s\n" "$as_me:   former value:  \`$ac_old_val'" >&2;}
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}:   current value: \`$ac_new_val'" >&5
+printf "%s\n" "$as_me:   current value: \`$ac_new_val'" >&2;}
       fi;;
   esac
   # Pass precious variables to config.status.
   if test "$ac_new_set" = set; then
     case $ac_new_val in
-    *\'*) ac_arg=$ac_var=`$as_echo "$ac_new_val" | sed "s/'/'\\\\\\\\''/g"` ;;
+    *\'*) ac_arg=$ac_var=`printf "%s\n" "$ac_new_val" | sed "s/'/'\\\\\\\\''/g"` ;;
     *) ac_arg=$ac_var=$ac_new_val ;;
     esac
     case " $ac_configure_args " in
@@ -1838,11 +2191,12 @@ $as_echo "$as_me:   current value: \`$ac_new_val'" >&2;}
   fi
 done
 if $ac_cache_corrupted; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
-  { $as_echo "$as_me:${as_lineno-$LINENO}: error: changes in the environment can compromise the build" >&5
-$as_echo "$as_me: error: changes in the environment can compromise the build" >&2;}
-  as_fn_error $? "run \`make distclean' and/or \`rm $cache_file' and start over" "$LINENO" 5
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: changes in the environment can compromise the build" >&5
+printf "%s\n" "$as_me: error: changes in the environment can compromise the build" >&2;}
+  as_fn_error $? "run \`${MAKE-make} distclean' and/or \`rm $cache_file'
+	    and start over" "$LINENO" 5
 fi
 ## -------------------- ##
 ## Main body of script. ##
@@ -1868,6 +2222,8 @@ CUDA_COMPAT_LIB="gpu/libnnc-compat-cuda.o"
 
 CUDA_CMD_LIB="libnnc-cmd-cuda.o"
 
+MFA_COMPAT_LIB="mfa/libnnc-compat-mfa.o"
+
 MPS_COMPAT_LIB="mps/libnnc-compat-mps.o"
 
 MPS_CMD_LIB="libnnc-cmd-mps.o"
@@ -1876,11 +2232,12 @@ MPS_CMD_LIB="libnnc-cmd-mps.o"
 # check if clang exists
 # Extract the first word of "clang", so it can be a program name with args.
 set dummy clang; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_CC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$CC"; then
   ac_cv_prog_CC="$CC" # Let the user override the test.
 else
@@ -1888,11 +2245,15 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
     ac_cv_prog_CC="clang"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -1903,11 +2264,11 @@ fi
 fi
 CC=$ac_cv_prog_CC
 if test -n "$CC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
-$as_echo "$CC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
+printf "%s\n" "$CC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 
@@ -1915,11 +2276,12 @@ fi
 # check if nvcc exists
 # Extract the first word of "nvcc", so it can be a program name with args.
 set dummy nvcc; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_NVCC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_NVCC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$NVCC"; then
   ac_cv_prog_NVCC="$NVCC" # Let the user override the test.
 else
@@ -1927,11 +2289,15 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
     ac_cv_prog_NVCC="nvcc"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -1942,11 +2308,11 @@ fi
 fi
 NVCC=$ac_cv_prog_NVCC
 if test -n "$NVCC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $NVCC" >&5
-$as_echo "$NVCC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $NVCC" >&5
+printf "%s\n" "$NVCC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 
@@ -1954,21 +2320,31 @@ fi
 
 
 # check for ARM NEON support
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking neon" >&5
-$as_echo_n "checking neon... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking neon" >&5
+printf %s "checking neon... " >&6; }
 # Check whether --enable-neon was given.
-if test "${enable_neon+set}" = set; then :
+if test ${enable_neon+y}
+then :
   enableval=$enable_neon; neon_support=$enableval
-else
+else $as_nop
   neon_support="no"
 fi
 
 if test "$neon_support" = yes; then
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: yes" >&5
-$as_echo "yes" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: yes" >&5
+printf "%s\n" "yes" >&6; }
 	DEFINE_MACROS="$DEFINE_MACROS-D HAVE_NEON "
 
-	ac_ext=c
+
+
+
+
+
+
+
+
+
+ac_ext=c
 ac_cpp='$CPP $CPPFLAGS'
 ac_compile='$CC -c $CFLAGS $CPPFLAGS conftest.$ac_ext >&5'
 ac_link='$CC -o conftest$ac_exeext $CFLAGS $CPPFLAGS $LDFLAGS conftest.$ac_ext $LIBS >&5'
@@ -1976,11 +2352,12 @@ ac_compiler_gnu=$ac_cv_c_compiler_gnu
 if test -n "$ac_tool_prefix"; then
   # Extract the first word of "${ac_tool_prefix}gcc", so it can be a program name with args.
 set dummy ${ac_tool_prefix}gcc; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_CC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$CC"; then
   ac_cv_prog_CC="$CC" # Let the user override the test.
 else
@@ -1988,11 +2365,15 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
     ac_cv_prog_CC="${ac_tool_prefix}gcc"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -2003,11 +2384,11 @@ fi
 fi
 CC=$ac_cv_prog_CC
 if test -n "$CC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
-$as_echo "$CC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
+printf "%s\n" "$CC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 
@@ -2016,11 +2397,12 @@ if test -z "$ac_cv_prog_CC"; then
   ac_ct_CC=$CC
   # Extract the first word of "gcc", so it can be a program name with args.
 set dummy gcc; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_ac_ct_CC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_ac_ct_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$ac_ct_CC"; then
   ac_cv_prog_ac_ct_CC="$ac_ct_CC" # Let the user override the test.
 else
@@ -2028,11 +2410,15 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
     ac_cv_prog_ac_ct_CC="gcc"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -2043,11 +2429,11 @@ fi
 fi
 ac_ct_CC=$ac_cv_prog_ac_ct_CC
 if test -n "$ac_ct_CC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_ct_CC" >&5
-$as_echo "$ac_ct_CC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_ct_CC" >&5
+printf "%s\n" "$ac_ct_CC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
   if test "x$ac_ct_CC" = x; then
@@ -2055,8 +2441,8 @@ fi
   else
     case $cross_compiling:$ac_tool_warned in
 yes:)
-{ $as_echo "$as_me:${as_lineno-$LINENO}: WARNING: using cross tools not prefixed with host triplet" >&5
-$as_echo "$as_me: WARNING: using cross tools not prefixed with host triplet" >&2;}
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: using cross tools not prefixed with host triplet" >&5
+printf "%s\n" "$as_me: WARNING: using cross tools not prefixed with host triplet" >&2;}
 ac_tool_warned=yes ;;
 esac
     CC=$ac_ct_CC
@@ -2069,11 +2455,12 @@ if test -z "$CC"; then
           if test -n "$ac_tool_prefix"; then
     # Extract the first word of "${ac_tool_prefix}cc", so it can be a program name with args.
 set dummy ${ac_tool_prefix}cc; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_CC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$CC"; then
   ac_cv_prog_CC="$CC" # Let the user override the test.
 else
@@ -2081,11 +2468,15 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
     ac_cv_prog_CC="${ac_tool_prefix}cc"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -2096,11 +2487,11 @@ fi
 fi
 CC=$ac_cv_prog_CC
 if test -n "$CC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
-$as_echo "$CC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
+printf "%s\n" "$CC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 
@@ -2109,11 +2500,12 @@ fi
 if test -z "$CC"; then
   # Extract the first word of "cc", so it can be a program name with args.
 set dummy cc; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_CC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$CC"; then
   ac_cv_prog_CC="$CC" # Let the user override the test.
 else
@@ -2122,15 +2514,19 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
-    if test "$as_dir/$ac_word$ac_exec_ext" = "/usr/ucb/cc"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
+    if test "$as_dir$ac_word$ac_exec_ext" = "/usr/ucb/cc"; then
        ac_prog_rejected=yes
        continue
      fi
     ac_cv_prog_CC="cc"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -2146,18 +2542,18 @@ if test $ac_prog_rejected = yes; then
     # However, it has the same basename, so the bogon will be chosen
     # first if we set CC to just the basename; use the full file name.
     shift
-    ac_cv_prog_CC="$as_dir/$ac_word${1+' '}$@"
+    ac_cv_prog_CC="$as_dir$ac_word${1+' '}$@"
   fi
 fi
 fi
 fi
 CC=$ac_cv_prog_CC
 if test -n "$CC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
-$as_echo "$CC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
+printf "%s\n" "$CC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 
@@ -2168,11 +2564,12 @@ if test -z "$CC"; then
   do
     # Extract the first word of "$ac_tool_prefix$ac_prog", so it can be a program name with args.
 set dummy $ac_tool_prefix$ac_prog; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_CC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$CC"; then
   ac_cv_prog_CC="$CC" # Let the user override the test.
 else
@@ -2180,11 +2577,15 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
     ac_cv_prog_CC="$ac_tool_prefix$ac_prog"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -2195,11 +2596,11 @@ fi
 fi
 CC=$ac_cv_prog_CC
 if test -n "$CC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
-$as_echo "$CC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
+printf "%s\n" "$CC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 
@@ -2212,11 +2613,12 @@ if test -z "$CC"; then
 do
   # Extract the first word of "$ac_prog", so it can be a program name with args.
 set dummy $ac_prog; ac_word=$2
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
-$as_echo_n "checking for $ac_word... " >&6; }
-if ${ac_cv_prog_ac_ct_CC+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_ac_ct_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   if test -n "$ac_ct_CC"; then
   ac_cv_prog_ac_ct_CC="$ac_ct_CC" # Let the user override the test.
 else
@@ -2224,11 +2626,15 @@ as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
     for ac_exec_ext in '' $ac_executable_extensions; do
-  if as_fn_executable_p "$as_dir/$ac_word$ac_exec_ext"; then
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
     ac_cv_prog_ac_ct_CC="$ac_prog"
-    $as_echo "$as_me:${as_lineno-$LINENO}: found $as_dir/$ac_word$ac_exec_ext" >&5
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
     break 2
   fi
 done
@@ -2239,11 +2645,11 @@ fi
 fi
 ac_ct_CC=$ac_cv_prog_ac_ct_CC
 if test -n "$ac_ct_CC"; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_ct_CC" >&5
-$as_echo "$ac_ct_CC" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_ct_CC" >&5
+printf "%s\n" "$ac_ct_CC" >&6; }
 else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 
@@ -2255,34 +2661,138 @@ done
   else
     case $cross_compiling:$ac_tool_warned in
 yes:)
-{ $as_echo "$as_me:${as_lineno-$LINENO}: WARNING: using cross tools not prefixed with host triplet" >&5
-$as_echo "$as_me: WARNING: using cross tools not prefixed with host triplet" >&2;}
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: using cross tools not prefixed with host triplet" >&5
+printf "%s\n" "$as_me: WARNING: using cross tools not prefixed with host triplet" >&2;}
+ac_tool_warned=yes ;;
+esac
+    CC=$ac_ct_CC
+  fi
+fi
+
+fi
+if test -z "$CC"; then
+  if test -n "$ac_tool_prefix"; then
+  # Extract the first word of "${ac_tool_prefix}clang", so it can be a program name with args.
+set dummy ${ac_tool_prefix}clang; ac_word=$2
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+  if test -n "$CC"; then
+  ac_cv_prog_CC="$CC" # Let the user override the test.
+else
+as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
+for as_dir in $PATH
+do
+  IFS=$as_save_IFS
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
+    for ac_exec_ext in '' $ac_executable_extensions; do
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
+    ac_cv_prog_CC="${ac_tool_prefix}clang"
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
+    break 2
+  fi
+done
+  done
+IFS=$as_save_IFS
+
+fi
+fi
+CC=$ac_cv_prog_CC
+if test -n "$CC"; then
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $CC" >&5
+printf "%s\n" "$CC" >&6; }
+else
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
+fi
+
+
+fi
+if test -z "$ac_cv_prog_CC"; then
+  ac_ct_CC=$CC
+  # Extract the first word of "clang", so it can be a program name with args.
+set dummy clang; ac_word=$2
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $ac_word" >&5
+printf %s "checking for $ac_word... " >&6; }
+if test ${ac_cv_prog_ac_ct_CC+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+  if test -n "$ac_ct_CC"; then
+  ac_cv_prog_ac_ct_CC="$ac_ct_CC" # Let the user override the test.
+else
+as_save_IFS=$IFS; IFS=$PATH_SEPARATOR
+for as_dir in $PATH
+do
+  IFS=$as_save_IFS
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
+    for ac_exec_ext in '' $ac_executable_extensions; do
+  if as_fn_executable_p "$as_dir$ac_word$ac_exec_ext"; then
+    ac_cv_prog_ac_ct_CC="clang"
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: found $as_dir$ac_word$ac_exec_ext" >&5
+    break 2
+  fi
+done
+  done
+IFS=$as_save_IFS
+
+fi
+fi
+ac_ct_CC=$ac_cv_prog_ac_ct_CC
+if test -n "$ac_ct_CC"; then
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_ct_CC" >&5
+printf "%s\n" "$ac_ct_CC" >&6; }
+else
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
+fi
+
+  if test "x$ac_ct_CC" = x; then
+    CC=""
+  else
+    case $cross_compiling:$ac_tool_warned in
+yes:)
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: using cross tools not prefixed with host triplet" >&5
+printf "%s\n" "$as_me: WARNING: using cross tools not prefixed with host triplet" >&2;}
 ac_tool_warned=yes ;;
 esac
     CC=$ac_ct_CC
   fi
+else
+  CC="$ac_cv_prog_CC"
 fi
 
 fi
 
 
-test -z "$CC" && { { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
+test -z "$CC" && { { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
 as_fn_error $? "no acceptable C compiler found in \$PATH
 See \`config.log' for more details" "$LINENO" 5; }
 
 # Provide some information about the compiler.
-$as_echo "$as_me:${as_lineno-$LINENO}: checking for C compiler version" >&5
+printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for C compiler version" >&5
 set X $ac_compile
 ac_compiler=$2
-for ac_option in --version -v -V -qversion; do
+for ac_option in --version -v -V -qversion -version; do
   { { ac_try="$ac_compiler $ac_option >&5"
 case "(($ac_try" in
   *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;;
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_compiler $ac_option >&5") 2>conftest.err
   ac_status=$?
   if test -s conftest.err; then
@@ -2292,7 +2802,7 @@ $as_echo "$ac_try_echo"; } >&5
     cat conftest.er1 >&5
   fi
   rm -f conftest.er1 conftest.err
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
   test $ac_status = 0; }
 done
 
@@ -2300,7 +2810,7 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 int
-main ()
+main (void)
 {
 
   ;
@@ -2312,9 +2822,9 @@ ac_clean_files="$ac_clean_files a.out a.out.dSYM a.exe b.out"
 # Try to create an executable without -o first, disregard a.out.
 # It will help us diagnose broken compilers, and finding out an intuition
 # of exeext.
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking whether the C compiler works" >&5
-$as_echo_n "checking whether the C compiler works... " >&6; }
-ac_link_default=`$as_echo "$ac_link" | sed 's/ -o *conftest[^ ]*//'`
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking whether the C compiler works" >&5
+printf %s "checking whether the C compiler works... " >&6; }
+ac_link_default=`printf "%s\n" "$ac_link" | sed 's/ -o *conftest[^ ]*//'`
 
 # The possible output files:
 ac_files="a.out conftest.exe conftest a.exe a_out.exe b.out conftest.*"
@@ -2335,11 +2845,12 @@ case "(($ac_try" in
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_link_default") 2>&5
   ac_status=$?
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
-  test $ac_status = 0; }; then :
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  test $ac_status = 0; }
+then :
   # Autoconf-2.13 could set the ac_cv_exeext variable to `no'.
 # So ignore a value of `no', otherwise this would lead to `EXEEXT = no'
 # in a Makefile.  We should not override ac_cv_exeext if it was cached,
@@ -2356,7 +2867,7 @@ do
 	# certainly right.
 	break;;
     *.* )
-	if test "${ac_cv_exeext+set}" = set && test "$ac_cv_exeext" != no;
+	if test ${ac_cv_exeext+y} && test "$ac_cv_exeext" != no;
 	then :; else
 	   ac_cv_exeext=`expr "$ac_file" : '[^.]*\(\..*\)'`
 	fi
@@ -2372,44 +2883,46 @@ do
 done
 test "$ac_cv_exeext" = no && ac_cv_exeext=
 
-else
+else $as_nop
   ac_file=''
 fi
-if test -z "$ac_file"; then :
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
-$as_echo "$as_me: failed program was:" >&5
+if test -z "$ac_file"
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
+printf "%s\n" "$as_me: failed program was:" >&5
 sed 's/^/| /' conftest.$ac_ext >&5
 
-{ { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
+{ { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
 as_fn_error 77 "C compiler cannot create executables
 See \`config.log' for more details" "$LINENO" 5; }
-else
-  { $as_echo "$as_me:${as_lineno-$LINENO}: result: yes" >&5
-$as_echo "yes" >&6; }
-fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for C compiler default output file name" >&5
-$as_echo_n "checking for C compiler default output file name... " >&6; }
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_file" >&5
-$as_echo "$ac_file" >&6; }
+else $as_nop
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: yes" >&5
+printf "%s\n" "yes" >&6; }
+fi
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for C compiler default output file name" >&5
+printf %s "checking for C compiler default output file name... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_file" >&5
+printf "%s\n" "$ac_file" >&6; }
 ac_exeext=$ac_cv_exeext
 
 rm -f -r a.out a.out.dSYM a.exe conftest$ac_cv_exeext b.out
 ac_clean_files=$ac_clean_files_save
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for suffix of executables" >&5
-$as_echo_n "checking for suffix of executables... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for suffix of executables" >&5
+printf %s "checking for suffix of executables... " >&6; }
 if { { ac_try="$ac_link"
 case "(($ac_try" in
   *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;;
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_link") 2>&5
   ac_status=$?
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
-  test $ac_status = 0; }; then :
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  test $ac_status = 0; }
+then :
   # If both `conftest.exe' and `conftest' are `present' (well, observable)
 # catch `conftest.exe'.  For instance with Cygwin, `ls conftest' will
 # work properly (i.e., refer to `conftest.exe'), while it won't with
@@ -2423,15 +2936,15 @@ for ac_file in conftest.exe conftest conftest.*; do
     * ) break;;
   esac
 done
-else
-  { { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
+else $as_nop
+  { { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
 as_fn_error $? "cannot compute suffix of executables: cannot compile and link
 See \`config.log' for more details" "$LINENO" 5; }
 fi
 rm -f conftest conftest$ac_cv_exeext
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_exeext" >&5
-$as_echo "$ac_cv_exeext" >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_exeext" >&5
+printf "%s\n" "$ac_cv_exeext" >&6; }
 
 rm -f conftest.$ac_ext
 EXEEXT=$ac_cv_exeext
@@ -2440,7 +2953,7 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 #include <stdio.h>
 int
-main ()
+main (void)
 {
 FILE *f = fopen ("conftest.out", "w");
  return ferror (f) || fclose (f) != 0;
@@ -2452,8 +2965,8 @@ _ACEOF
 ac_clean_files="$ac_clean_files conftest.out"
 # Check that the compiler produces executables we can run.  If not, either
 # the compiler is broken, or we cross compile.
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking whether we are cross compiling" >&5
-$as_echo_n "checking whether we are cross compiling... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking whether we are cross compiling" >&5
+printf %s "checking whether we are cross compiling... " >&6; }
 if test "$cross_compiling" != yes; then
   { { ac_try="$ac_link"
 case "(($ac_try" in
@@ -2461,10 +2974,10 @@ case "(($ac_try" in
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_link") 2>&5
   ac_status=$?
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
   test $ac_status = 0; }
   if { ac_try='./conftest$ac_cv_exeext'
   { { case "(($ac_try" in
@@ -2472,39 +2985,40 @@ $as_echo "$ac_try_echo"; } >&5
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_try") 2>&5
   ac_status=$?
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
   test $ac_status = 0; }; }; then
     cross_compiling=no
   else
     if test "$cross_compiling" = maybe; then
 	cross_compiling=yes
     else
-	{ { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
-as_fn_error $? "cannot run C compiled programs.
+	{ { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
+as_fn_error 77 "cannot run C compiled programs.
 If you meant to cross compile, use \`--host'.
 See \`config.log' for more details" "$LINENO" 5; }
     fi
   fi
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $cross_compiling" >&5
-$as_echo "$cross_compiling" >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $cross_compiling" >&5
+printf "%s\n" "$cross_compiling" >&6; }
 
 rm -f conftest.$ac_ext conftest$ac_cv_exeext conftest.out
 ac_clean_files=$ac_clean_files_save
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for suffix of object files" >&5
-$as_echo_n "checking for suffix of object files... " >&6; }
-if ${ac_cv_objext+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for suffix of object files" >&5
+printf %s "checking for suffix of object files... " >&6; }
+if test ${ac_cv_objext+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 int
-main ()
+main (void)
 {
 
   ;
@@ -2518,11 +3032,12 @@ case "(($ac_try" in
   *) ac_try_echo=$ac_try;;
 esac
 eval ac_try_echo="\"\$as_me:${as_lineno-$LINENO}: $ac_try_echo\""
-$as_echo "$ac_try_echo"; } >&5
+printf "%s\n" "$ac_try_echo"; } >&5
   (eval "$ac_compile") 2>&5
   ac_status=$?
-  $as_echo "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
-  test $ac_status = 0; }; then :
+  printf "%s\n" "$as_me:${as_lineno-$LINENO}: \$? = $ac_status" >&5
+  test $ac_status = 0; }
+then :
   for ac_file in conftest.o conftest.obj conftest.*; do
   test -f "$ac_file" || continue;
   case $ac_file in
@@ -2531,31 +3046,32 @@ $as_echo "$ac_try_echo"; } >&5
        break;;
   esac
 done
-else
-  $as_echo "$as_me: failed program was:" >&5
+else $as_nop
+  printf "%s\n" "$as_me: failed program was:" >&5
 sed 's/^/| /' conftest.$ac_ext >&5
 
-{ { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
+{ { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
 as_fn_error $? "cannot compute suffix of object files: cannot compile
 See \`config.log' for more details" "$LINENO" 5; }
 fi
 rm -f conftest.$ac_cv_objext conftest.$ac_ext
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_objext" >&5
-$as_echo "$ac_cv_objext" >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_objext" >&5
+printf "%s\n" "$ac_cv_objext" >&6; }
 OBJEXT=$ac_cv_objext
 ac_objext=$OBJEXT
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking whether we are using the GNU C compiler" >&5
-$as_echo_n "checking whether we are using the GNU C compiler... " >&6; }
-if ${ac_cv_c_compiler_gnu+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking whether the compiler supports GNU C" >&5
+printf %s "checking whether the compiler supports GNU C... " >&6; }
+if test ${ac_cv_c_compiler_gnu+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 int
-main ()
+main (void)
 {
 #ifndef __GNUC__
        choke me
@@ -2565,29 +3081,33 @@ main ()
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_compile "$LINENO"; then :
+if ac_fn_c_try_compile "$LINENO"
+then :
   ac_compiler_gnu=yes
-else
+else $as_nop
   ac_compiler_gnu=no
 fi
-rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
+rm -f core conftest.err conftest.$ac_objext conftest.beam conftest.$ac_ext
 ac_cv_c_compiler_gnu=$ac_compiler_gnu
 
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_c_compiler_gnu" >&5
-$as_echo "$ac_cv_c_compiler_gnu" >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_c_compiler_gnu" >&5
+printf "%s\n" "$ac_cv_c_compiler_gnu" >&6; }
+ac_compiler_gnu=$ac_cv_c_compiler_gnu
+
 if test $ac_compiler_gnu = yes; then
   GCC=yes
 else
   GCC=
 fi
-ac_test_CFLAGS=${CFLAGS+set}
+ac_test_CFLAGS=${CFLAGS+y}
 ac_save_CFLAGS=$CFLAGS
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking whether $CC accepts -g" >&5
-$as_echo_n "checking whether $CC accepts -g... " >&6; }
-if ${ac_cv_prog_cc_g+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking whether $CC accepts -g" >&5
+printf %s "checking whether $CC accepts -g... " >&6; }
+if test ${ac_cv_prog_cc_g+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_save_c_werror_flag=$ac_c_werror_flag
    ac_c_werror_flag=yes
    ac_cv_prog_cc_g=no
@@ -2596,57 +3116,60 @@ else
 /* end confdefs.h.  */
 
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_compile "$LINENO"; then :
+if ac_fn_c_try_compile "$LINENO"
+then :
   ac_cv_prog_cc_g=yes
-else
+else $as_nop
   CFLAGS=""
       cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_compile "$LINENO"; then :
+if ac_fn_c_try_compile "$LINENO"
+then :
 
-else
+else $as_nop
   ac_c_werror_flag=$ac_save_c_werror_flag
 	 CFLAGS="-g"
 	 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_compile "$LINENO"; then :
+if ac_fn_c_try_compile "$LINENO"
+then :
   ac_cv_prog_cc_g=yes
 fi
-rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
+rm -f core conftest.err conftest.$ac_objext conftest.beam conftest.$ac_ext
 fi
-rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
+rm -f core conftest.err conftest.$ac_objext conftest.beam conftest.$ac_ext
 fi
-rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
+rm -f core conftest.err conftest.$ac_objext conftest.beam conftest.$ac_ext
    ac_c_werror_flag=$ac_save_c_werror_flag
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_cc_g" >&5
-$as_echo "$ac_cv_prog_cc_g" >&6; }
-if test "$ac_test_CFLAGS" = set; then
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_cc_g" >&5
+printf "%s\n" "$ac_cv_prog_cc_g" >&6; }
+if test $ac_test_CFLAGS; then
   CFLAGS=$ac_save_CFLAGS
 elif test $ac_cv_prog_cc_g = yes; then
   if test "$GCC" = yes; then
@@ -2661,94 +3184,144 @@ else
     CFLAGS=
   fi
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $CC option to accept ISO C89" >&5
-$as_echo_n "checking for $CC option to accept ISO C89... " >&6; }
-if ${ac_cv_prog_cc_c89+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
-  ac_cv_prog_cc_c89=no
+ac_prog_cc_stdc=no
+if test x$ac_prog_cc_stdc = xno
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $CC option to enable C11 features" >&5
+printf %s "checking for $CC option to enable C11 features... " >&6; }
+if test ${ac_cv_prog_cc_c11+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+  ac_cv_prog_cc_c11=no
 ac_save_CC=$CC
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
-#include <stdarg.h>
-#include <stdio.h>
-struct stat;
-/* Most of the following tests are stolen from RCS 5.7's src/conf.sh.  */
-struct buf { int x; };
-FILE * (*rcsopen) (struct buf *, struct stat *, int);
-static char *e (p, i)
-     char **p;
-     int i;
-{
-  return p[i];
-}
-static char *f (char * (*g) (char **, int), char **p, ...)
-{
-  char *s;
-  va_list v;
-  va_start (v,p);
-  s = g (p, va_arg (v,int));
-  va_end (v);
-  return s;
-}
-
-/* OSF 4.0 Compaq cc is some sort of almost-ANSI by default.  It has
-   function prototypes and stuff, but not '\xHH' hex character constants.
-   These don't provoke an error unfortunately, instead are silently treated
-   as 'x'.  The following induces an error, until -std is added to get
-   proper ANSI mode.  Curiously '\x00'!='x' always comes out true, for an
-   array size at least.  It's necessary to write '\x00'==0 to get something
-   that's true only with -std.  */
-int osf4_cc_array ['\x00' == 0 ? 1 : -1];
+$ac_c_conftest_c11_program
+_ACEOF
+for ac_arg in '' -std=gnu11
+do
+  CC="$ac_save_CC $ac_arg"
+  if ac_fn_c_try_compile "$LINENO"
+then :
+  ac_cv_prog_cc_c11=$ac_arg
+fi
+rm -f core conftest.err conftest.$ac_objext conftest.beam
+  test "x$ac_cv_prog_cc_c11" != "xno" && break
+done
+rm -f conftest.$ac_ext
+CC=$ac_save_CC
+fi
 
-/* IBM C 6 for AIX is almost-ANSI by default, but it replaces macro parameters
-   inside strings and character constants.  */
-#define FOO(x) 'x'
-int xlc6_cc_array[FOO(a) == 'x' ? 1 : -1];
+if test "x$ac_cv_prog_cc_c11" = xno
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: unsupported" >&5
+printf "%s\n" "unsupported" >&6; }
+else $as_nop
+  if test "x$ac_cv_prog_cc_c11" = x
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: none needed" >&5
+printf "%s\n" "none needed" >&6; }
+else $as_nop
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_cc_c11" >&5
+printf "%s\n" "$ac_cv_prog_cc_c11" >&6; }
+     CC="$CC $ac_cv_prog_cc_c11"
+fi
+  ac_cv_prog_cc_stdc=$ac_cv_prog_cc_c11
+  ac_prog_cc_stdc=c11
+fi
+fi
+if test x$ac_prog_cc_stdc = xno
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $CC option to enable C99 features" >&5
+printf %s "checking for $CC option to enable C99 features... " >&6; }
+if test ${ac_cv_prog_cc_c99+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+  ac_cv_prog_cc_c99=no
+ac_save_CC=$CC
+cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+$ac_c_conftest_c99_program
+_ACEOF
+for ac_arg in '' -std=gnu99 -std=c99 -c99 -qlanglvl=extc1x -qlanglvl=extc99 -AC99 -D_STDC_C99=
+do
+  CC="$ac_save_CC $ac_arg"
+  if ac_fn_c_try_compile "$LINENO"
+then :
+  ac_cv_prog_cc_c99=$ac_arg
+fi
+rm -f core conftest.err conftest.$ac_objext conftest.beam
+  test "x$ac_cv_prog_cc_c99" != "xno" && break
+done
+rm -f conftest.$ac_ext
+CC=$ac_save_CC
+fi
 
-int test (int i, double x);
-struct s1 {int (*f) (int a);};
-struct s2 {int (*f) (double a);};
-int pairnames (int, char **, FILE *(*)(struct buf *, struct stat *, int), int, int);
-int argc;
-char **argv;
-int
-main ()
-{
-return f (e, argv, 0) != argv[0]  ||  f (e, argv, 1) != argv[1];
-  ;
-  return 0;
-}
+if test "x$ac_cv_prog_cc_c99" = xno
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: unsupported" >&5
+printf "%s\n" "unsupported" >&6; }
+else $as_nop
+  if test "x$ac_cv_prog_cc_c99" = x
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: none needed" >&5
+printf "%s\n" "none needed" >&6; }
+else $as_nop
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_cc_c99" >&5
+printf "%s\n" "$ac_cv_prog_cc_c99" >&6; }
+     CC="$CC $ac_cv_prog_cc_c99"
+fi
+  ac_cv_prog_cc_stdc=$ac_cv_prog_cc_c99
+  ac_prog_cc_stdc=c99
+fi
+fi
+if test x$ac_prog_cc_stdc = xno
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $CC option to enable C89 features" >&5
+printf %s "checking for $CC option to enable C89 features... " >&6; }
+if test ${ac_cv_prog_cc_c89+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+  ac_cv_prog_cc_c89=no
+ac_save_CC=$CC
+cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+$ac_c_conftest_c89_program
 _ACEOF
-for ac_arg in '' -qlanglvl=extc89 -qlanglvl=ansi -std \
-	-Ae "-Aa -D_HPUX_SOURCE" "-Xc -D__EXTENSIONS__"
+for ac_arg in '' -qlanglvl=extc89 -qlanglvl=ansi -std -Ae "-Aa -D_HPUX_SOURCE" "-Xc -D__EXTENSIONS__"
 do
   CC="$ac_save_CC $ac_arg"
-  if ac_fn_c_try_compile "$LINENO"; then :
+  if ac_fn_c_try_compile "$LINENO"
+then :
   ac_cv_prog_cc_c89=$ac_arg
 fi
-rm -f core conftest.err conftest.$ac_objext
+rm -f core conftest.err conftest.$ac_objext conftest.beam
   test "x$ac_cv_prog_cc_c89" != "xno" && break
 done
 rm -f conftest.$ac_ext
 CC=$ac_save_CC
-
 fi
-# AC_CACHE_VAL
-case "x$ac_cv_prog_cc_c89" in
-  x)
-    { $as_echo "$as_me:${as_lineno-$LINENO}: result: none needed" >&5
-$as_echo "none needed" >&6; } ;;
-  xno)
-    { $as_echo "$as_me:${as_lineno-$LINENO}: result: unsupported" >&5
-$as_echo "unsupported" >&6; } ;;
-  *)
-    CC="$CC $ac_cv_prog_cc_c89"
-    { $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_cc_c89" >&5
-$as_echo "$ac_cv_prog_cc_c89" >&6; } ;;
-esac
-if test "x$ac_cv_prog_cc_c89" != xno; then :
 
+if test "x$ac_cv_prog_cc_c89" = xno
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: unsupported" >&5
+printf "%s\n" "unsupported" >&6; }
+else $as_nop
+  if test "x$ac_cv_prog_cc_c89" = x
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: none needed" >&5
+printf "%s\n" "none needed" >&6; }
+else $as_nop
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_cc_c89" >&5
+printf "%s\n" "$ac_cv_prog_cc_c89" >&6; }
+     CC="$CC $ac_cv_prog_cc_c89"
+fi
+  ac_cv_prog_cc_stdc=$ac_cv_prog_cc_c89
+  ac_prog_cc_stdc=c89
+fi
 fi
 
 ac_ext=c
@@ -2758,11 +3331,12 @@ ac_link='$CC -o conftest$ac_exeext $CFLAGS $CPPFLAGS $LDFLAGS conftest.$ac_ext $
 ac_compiler_gnu=$ac_cv_c_compiler_gnu
 
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking whether C compiler accepts -mfpu=neon" >&5
-$as_echo_n "checking whether C compiler accepts -mfpu=neon... " >&6; }
-if ${ax_cv_check_cflags__Werror__mfpu_neon+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking whether C compiler accepts -mfpu=neon" >&5
+printf %s "checking whether C compiler accepts -mfpu=neon... " >&6; }
+if test ${ax_cv_check_cflags__Werror__mfpu_neon+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS -Werror -mfpu=neon"
@@ -2770,64 +3344,68 @@ else
 /* end confdefs.h.  */
 
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_compile "$LINENO"; then :
+if ac_fn_c_try_compile "$LINENO"
+then :
   ax_cv_check_cflags__Werror__mfpu_neon=yes
-else
+else $as_nop
   ax_cv_check_cflags__Werror__mfpu_neon=no
 fi
-rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext
+rm -f core conftest.err conftest.$ac_objext conftest.beam conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags__Werror__mfpu_neon" >&5
-$as_echo "$ax_cv_check_cflags__Werror__mfpu_neon" >&6; }
-	if test "x$ax_cv_check_cflags__Werror__mfpu_neon" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags__Werror__mfpu_neon" >&5
+printf "%s\n" "$ax_cv_check_cflags__Werror__mfpu_neon" >&6; }
+	if test "x$ax_cv_check_cflags__Werror__mfpu_neon" = xyes
+then :
   MKCFLAGS="$MKCFLAGS-mfpu=neon -mfloat-abi=hard "
 
-else
+else $as_nop
   :
 fi
 
 else
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking arch" >&5
-$as_echo_n "checking arch... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking arch" >&5
+printf %s "checking arch... " >&6; }
 
 # Check whether --with-arch was given.
-if test "${with_arch+set}" = set; then :
+if test ${with_arch+y}
+then :
   withval=$with_arch; arch_native=$withval
-else
+else $as_nop
   arch_native="native"
 fi
 
 if test "$arch_native" = native; then
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: native" >&5
-$as_echo "native" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: native" >&5
+printf "%s\n" "native" >&6; }
 else
 	MKCFLAGS="$MKCFLAGS-march=$arch_native -mtune=$arch_native "
 
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $arch_native" >&5
-$as_echo "$arch_native" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $arch_native" >&5
+printf "%s\n" "$arch_native" >&6; }
 fi
 
 # check for blas
 blas_ok=no
 # check for ATLAS library
 if test "$blas_ok" = no; then
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for ATL_xerbla in -latlas" >&5
-$as_echo_n "checking for ATL_xerbla in -latlas... " >&6; }
-if ${ac_cv_lib_atlas_ATL_xerbla+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for ATL_xerbla in -latlas" >&5
+printf %s "checking for ATL_xerbla in -latlas... " >&6; }
+if test ${ac_cv_lib_atlas_ATL_xerbla+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-latlas  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -2836,35 +3414,35 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char ATL_xerbla ();
 int
-main ()
+main (void)
 {
 return ATL_xerbla ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_atlas_ATL_xerbla=yes
-else
+else $as_nop
   ac_cv_lib_atlas_ATL_xerbla=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_atlas_ATL_xerbla" >&5
-$as_echo "$ac_cv_lib_atlas_ATL_xerbla" >&6; }
-if test "x$ac_cv_lib_atlas_ATL_xerbla" = xyes; then :
-  { $as_echo "$as_me:${as_lineno-$LINENO}: checking for cblas_dgemm in -lcblas" >&5
-$as_echo_n "checking for cblas_dgemm in -lcblas... " >&6; }
-if ${ac_cv_lib_cblas_cblas_dgemm+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_atlas_ATL_xerbla" >&5
+printf "%s\n" "$ac_cv_lib_atlas_ATL_xerbla" >&6; }
+if test "x$ac_cv_lib_atlas_ATL_xerbla" = xyes
+then :
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for cblas_dgemm in -lcblas" >&5
+printf %s "checking for cblas_dgemm in -lcblas... " >&6; }
+if test ${ac_cv_lib_cblas_cblas_dgemm+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-lcblas -latlas $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -2873,30 +3451,29 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char cblas_dgemm ();
 int
-main ()
+main (void)
 {
 return cblas_dgemm ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_cblas_cblas_dgemm=yes
-else
+else $as_nop
   ac_cv_lib_cblas_cblas_dgemm=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_cblas_cblas_dgemm" >&5
-$as_echo "$ac_cv_lib_cblas_cblas_dgemm" >&6; }
-if test "x$ac_cv_lib_cblas_cblas_dgemm" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_cblas_cblas_dgemm" >&5
+printf "%s\n" "$ac_cv_lib_cblas_cblas_dgemm" >&6; }
+if test "x$ac_cv_lib_cblas_cblas_dgemm" = xyes
+then :
   blas_ok=yes
 		BLAS_LIBS="-lcblas -latlas"
 fi
@@ -2906,11 +3483,12 @@ fi
 fi
 # check for Generic BLAS library
 if test "$blas_ok" = no; then
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for sgemm in -lblas" >&5
-$as_echo_n "checking for sgemm in -lblas... " >&6; }
-if ${ac_cv_lib_blas_sgemm+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for sgemm in -lblas" >&5
+printf %s "checking for sgemm in -lblas... " >&6; }
+if test ${ac_cv_lib_blas_sgemm+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-lblas  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -2919,41 +3497,41 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char sgemm ();
 int
-main ()
+main (void)
 {
 return sgemm ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_blas_sgemm=yes
-else
+else $as_nop
   ac_cv_lib_blas_sgemm=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_blas_sgemm" >&5
-$as_echo "$ac_cv_lib_blas_sgemm" >&6; }
-if test "x$ac_cv_lib_blas_sgemm" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_blas_sgemm" >&5
+printf "%s\n" "$ac_cv_lib_blas_sgemm" >&6; }
+if test "x$ac_cv_lib_blas_sgemm" = xyes
+then :
   blas_ok=yes; BLAS_LIBS="-lblas"
 fi
 
 fi
 # check for OpenBLAS library
 if test "$blas_ok" = no; then
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for cblas_sgemm in -lopenblas" >&5
-$as_echo_n "checking for cblas_sgemm in -lopenblas... " >&6; }
-if ${ac_cv_lib_openblas_cblas_sgemm+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for cblas_sgemm in -lopenblas" >&5
+printf %s "checking for cblas_sgemm in -lopenblas... " >&6; }
+if test ${ac_cv_lib_openblas_cblas_sgemm+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-lopenblas  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -2962,30 +3540,29 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char cblas_sgemm ();
 int
-main ()
+main (void)
 {
 return cblas_sgemm ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_openblas_cblas_sgemm=yes
-else
+else $as_nop
   ac_cv_lib_openblas_cblas_sgemm=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_openblas_cblas_sgemm" >&5
-$as_echo "$ac_cv_lib_openblas_cblas_sgemm" >&6; }
-if test "x$ac_cv_lib_openblas_cblas_sgemm" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_openblas_cblas_sgemm" >&5
+printf "%s\n" "$ac_cv_lib_openblas_cblas_sgemm" >&6; }
+if test "x$ac_cv_lib_openblas_cblas_sgemm" = xyes
+then :
   blas_ok=yes; BLAS_LIBS="-lopenblas"
 fi
 
@@ -2997,11 +3574,12 @@ if test "$blas_ok" = yes; then
 
 fi
 
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for clock_gettime in -lrt" >&5
-$as_echo_n "checking for clock_gettime in -lrt... " >&6; }
-if ${ac_cv_lib_rt_clock_gettime+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for clock_gettime in -lrt" >&5
+printf %s "checking for clock_gettime in -lrt... " >&6; }
+if test ${ac_cv_lib_rt_clock_gettime+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-lrt  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -3010,30 +3588,29 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char clock_gettime ();
 int
-main ()
+main (void)
 {
 return clock_gettime ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_rt_clock_gettime=yes
-else
+else $as_nop
   ac_cv_lib_rt_clock_gettime=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_rt_clock_gettime" >&5
-$as_echo "$ac_cv_lib_rt_clock_gettime" >&6; }
-if test "x$ac_cv_lib_rt_clock_gettime" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_rt_clock_gettime" >&5
+printf "%s\n" "$ac_cv_lib_rt_clock_gettime" >&6; }
+if test "x$ac_cv_lib_rt_clock_gettime" = xyes
+then :
   MKLDFLAGS="$MKLDFLAGS-lrt "
 
 fi
@@ -3045,40 +3622,36 @@ ac_cpp='$CPP $CPPFLAGS'
 ac_compile='$CC -c $CFLAGS $CPPFLAGS conftest.$ac_ext >&5'
 ac_link='$CC -o conftest$ac_exeext $CFLAGS $CPPFLAGS $LDFLAGS conftest.$ac_ext $LIBS >&5'
 ac_compiler_gnu=$ac_cv_c_compiler_gnu
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking how to run the C preprocessor" >&5
-$as_echo_n "checking how to run the C preprocessor... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking how to run the C preprocessor" >&5
+printf %s "checking how to run the C preprocessor... " >&6; }
 # On Suns, sometimes $CPP names a directory.
 if test -n "$CPP" && test -d "$CPP"; then
   CPP=
 fi
 if test -z "$CPP"; then
-  if ${ac_cv_prog_CPP+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
-      # Double quotes because CPP needs to be expanded
-    for CPP in "$CC -E" "$CC -E -traditional-cpp" "/lib/cpp"
+  if test ${ac_cv_prog_CPP+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+      # Double quotes because $CC needs to be expanded
+    for CPP in "$CC -E" "$CC -E -traditional-cpp" cpp /lib/cpp
     do
       ac_preproc_ok=false
 for ac_c_preproc_warn_flag in '' yes
 do
   # Use a header file that comes with gcc, so configuring glibc
   # with a fresh cross-compiler works.
-  # Prefer <limits.h> to <assert.h> if __STDC__ is defined, since
-  # <limits.h> exists even on freestanding compilers.
   # On the NeXT, cc -E runs the code through the compiler's parser,
   # not just through cpp. "Syntax error" is here to catch this case.
   cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
-#ifdef __STDC__
-# include <limits.h>
-#else
-# include <assert.h>
-#endif
+#include <limits.h>
 		     Syntax error
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
 
-else
+else $as_nop
   # Broken: fails on valid input.
 continue
 fi
@@ -3090,10 +3663,11 @@ rm -f conftest.err conftest.i conftest.$ac_ext
 /* end confdefs.h.  */
 #include <ac_nonexistent.h>
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   # Broken: success on invalid input.
 continue
-else
+else $as_nop
   # Passes both tests.
 ac_preproc_ok=:
 break
@@ -3103,7 +3677,8 @@ rm -f conftest.err conftest.i conftest.$ac_ext
 done
 # Because of `break', _AC_PREPROC_IFELSE's cleaning code was skipped.
 rm -f conftest.i conftest.err conftest.$ac_ext
-if $ac_preproc_ok; then :
+if $ac_preproc_ok
+then :
   break
 fi
 
@@ -3115,29 +3690,24 @@ fi
 else
   ac_cv_prog_CPP=$CPP
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $CPP" >&5
-$as_echo "$CPP" >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $CPP" >&5
+printf "%s\n" "$CPP" >&6; }
 ac_preproc_ok=false
 for ac_c_preproc_warn_flag in '' yes
 do
   # Use a header file that comes with gcc, so configuring glibc
   # with a fresh cross-compiler works.
-  # Prefer <limits.h> to <assert.h> if __STDC__ is defined, since
-  # <limits.h> exists even on freestanding compilers.
   # On the NeXT, cc -E runs the code through the compiler's parser,
   # not just through cpp. "Syntax error" is here to catch this case.
   cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
-#ifdef __STDC__
-# include <limits.h>
-#else
-# include <assert.h>
-#endif
+#include <limits.h>
 		     Syntax error
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
 
-else
+else $as_nop
   # Broken: fails on valid input.
 continue
 fi
@@ -3149,10 +3719,11 @@ rm -f conftest.err conftest.i conftest.$ac_ext
 /* end confdefs.h.  */
 #include <ac_nonexistent.h>
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   # Broken: success on invalid input.
 continue
-else
+else $as_nop
   # Passes both tests.
 ac_preproc_ok=:
 break
@@ -3162,11 +3733,12 @@ rm -f conftest.err conftest.i conftest.$ac_ext
 done
 # Because of `break', _AC_PREPROC_IFELSE's cleaning code was skipped.
 rm -f conftest.i conftest.err conftest.$ac_ext
-if $ac_preproc_ok; then :
+if $ac_preproc_ok
+then :
 
-else
-  { { $as_echo "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
-$as_echo "$as_me: error: in \`$ac_pwd':" >&2;}
+else $as_nop
+  { { printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: in \`$ac_pwd':" >&5
+printf "%s\n" "$as_me: error: in \`$ac_pwd':" >&2;}
 as_fn_error $? "C preprocessor \"$CPP\" fails sanity check
 See \`config.log' for more details" "$LINENO" 5; }
 fi
@@ -3178,11 +3750,12 @@ ac_link='$CC -o conftest$ac_exeext $CFLAGS $CPPFLAGS $LDFLAGS conftest.$ac_ext $
 ac_compiler_gnu=$ac_cv_c_compiler_gnu
 
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking png.h presence" >&5
-$as_echo_n "checking png.h presence... " >&6; }
-if ${ax_cv_check_cflags_png_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking png.h presence" >&5
+printf %s "checking png.h presence... " >&6; }
+if test ${ax_cv_check_cflags_png_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS png.h"
@@ -3190,36 +3763,39 @@ else
 /* end confdefs.h.  */
 #include <png.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_png_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_png_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_png_h" >&5
-$as_echo "$ax_cv_check_cflags_png_h" >&6; }
-	if test "x$ax_cv_check_cflags_png_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_png_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_png_h" >&6; }
+	if test "x$ax_cv_check_cflags_png_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_LIBPNG "
  MKLDFLAGS="$MKLDFLAGS-lpng "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking jpeglib.h presence" >&5
-$as_echo_n "checking jpeglib.h presence... " >&6; }
-if ${ax_cv_check_cflags_jpeglib_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking jpeglib.h presence" >&5
+printf %s "checking jpeglib.h presence... " >&6; }
+if test ${ax_cv_check_cflags_jpeglib_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS jpeglib.h"
@@ -3227,47 +3803,51 @@ else
 /* end confdefs.h.  */
 #include <jpeglib.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_jpeglib_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_jpeglib_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_jpeglib_h" >&5
-$as_echo "$ax_cv_check_cflags_jpeglib_h" >&6; }
-	if test "x$ax_cv_check_cflags_jpeglib_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_jpeglib_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_jpeglib_h" >&6; }
+	if test "x$ax_cv_check_cflags_jpeglib_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_LIBJPEG "
  MKLDFLAGS="$MKLDFLAGS-ljpeg "
 
-else
+else $as_nop
   :
 fi
 
 
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking fftw3" >&5
-$as_echo_n "checking fftw3... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking fftw3" >&5
+printf %s "checking fftw3... " >&6; }
 # Check whether --enable-fftw3 was given.
-if test "${enable_fftw3+set}" = set; then :
+if test ${enable_fftw3+y}
+then :
   enableval=$enable_fftw3; fftw3_enable=$enableval
-else
+else $as_nop
   fftw3_enable="yes"
 fi
 
 if test "$fftw3_enable" != no; then
-	 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking fftw3.h presence" >&5
-$as_echo_n "checking fftw3.h presence... " >&6; }
-if ${ax_cv_check_cflags_fftw3_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking fftw3.h presence" >&5
+printf %s "checking fftw3.h presence... " >&6; }
+if test ${ax_cv_check_cflags_fftw3_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS fftw3.h"
@@ -3275,41 +3855,44 @@ else
 /* end confdefs.h.  */
 #include <fftw3.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_fftw3_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_fftw3_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_fftw3_h" >&5
-$as_echo "$ax_cv_check_cflags_fftw3_h" >&6; }
-	if test "x$ax_cv_check_cflags_fftw3_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_fftw3_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_fftw3_h" >&6; }
+	if test "x$ax_cv_check_cflags_fftw3_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_FFTW3 "
  MKLDFLAGS="$MKLDFLAGS-lfftw3 -lfftw3f -lpthread "
 
-else
+else $as_nop
   :
 fi
 
 else
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: disabled" >&5
-$as_echo "disabled" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: disabled" >&5
+printf "%s\n" "disabled" >&6; }
 fi
 # Check pthread
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking pthread.h presence" >&5
-$as_echo_n "checking pthread.h presence... " >&6; }
-if ${ax_cv_check_cflags_pthread_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking pthread.h presence" >&5
+printf %s "checking pthread.h presence... " >&6; }
+if test ${ax_cv_check_cflags_pthread_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS pthread.h"
@@ -3317,36 +3900,39 @@ else
 /* end confdefs.h.  */
 #include <pthread.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_pthread_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_pthread_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_pthread_h" >&5
-$as_echo "$ax_cv_check_cflags_pthread_h" >&6; }
-	if test "x$ax_cv_check_cflags_pthread_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_pthread_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_pthread_h" >&6; }
+	if test "x$ax_cv_check_cflags_pthread_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_PTHREAD "
  MKLDFLAGS="$MKLDFLAGS-lpthread "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking linear.h presence" >&5
-$as_echo_n "checking linear.h presence... " >&6; }
-if ${ax_cv_check_cflags_linear_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking linear.h presence" >&5
+printf %s "checking linear.h presence... " >&6; }
+if test ${ax_cv_check_cflags_linear_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS linear.h"
@@ -3354,36 +3940,39 @@ else
 /* end confdefs.h.  */
 #include <linear.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_linear_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_linear_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_linear_h" >&5
-$as_echo "$ax_cv_check_cflags_linear_h" >&6; }
-	if test "x$ax_cv_check_cflags_linear_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_linear_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_linear_h" >&6; }
+	if test "x$ax_cv_check_cflags_linear_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_LIBLINEAR "
  MKLDFLAGS="$MKLDFLAGS-llinear "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking tesseract/capi.h presence" >&5
-$as_echo_n "checking tesseract/capi.h presence... " >&6; }
-if ${ax_cv_check_cflags_tesseract_capi_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking tesseract/capi.h presence" >&5
+printf %s "checking tesseract/capi.h presence... " >&6; }
+if test ${ax_cv_check_cflags_tesseract_capi_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS tesseract/capi.h"
@@ -3391,36 +3980,39 @@ else
 /* end confdefs.h.  */
 #include <tesseract/capi.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_tesseract_capi_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_tesseract_capi_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_tesseract_capi_h" >&5
-$as_echo "$ax_cv_check_cflags_tesseract_capi_h" >&6; }
-	if test "x$ax_cv_check_cflags_tesseract_capi_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_tesseract_capi_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_tesseract_capi_h" >&6; }
+	if test "x$ax_cv_check_cflags_tesseract_capi_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_TESSERACT "
  MKLDFLAGS="$MKLDFLAGS-ltesseract "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking Accelerate/Accelerate.h presence" >&5
-$as_echo_n "checking Accelerate/Accelerate.h presence... " >&6; }
-if ${ax_cv_check_cflags_Accelerate_Accelerate_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking Accelerate/Accelerate.h presence" >&5
+printf %s "checking Accelerate/Accelerate.h presence... " >&6; }
+if test ${ax_cv_check_cflags_Accelerate_Accelerate_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS Accelerate/Accelerate.h"
@@ -3428,36 +4020,39 @@ else
 /* end confdefs.h.  */
 #include <Accelerate/Accelerate.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_Accelerate_Accelerate_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_Accelerate_Accelerate_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_Accelerate_Accelerate_h" >&5
-$as_echo "$ax_cv_check_cflags_Accelerate_Accelerate_h" >&6; }
-	if test "x$ax_cv_check_cflags_Accelerate_Accelerate_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_Accelerate_Accelerate_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_Accelerate_Accelerate_h" >&6; }
+	if test "x$ax_cv_check_cflags_Accelerate_Accelerate_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_ACCELERATE_FRAMEWORK "
  MKLDFLAGS="$MKLDFLAGS-framework Accelerate "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking libavcodec/avcodec.h presence" >&5
-$as_echo_n "checking libavcodec/avcodec.h presence... " >&6; }
-if ${ax_cv_check_cflags_libavcodec_avcodec_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking libavcodec/avcodec.h presence" >&5
+printf %s "checking libavcodec/avcodec.h presence... " >&6; }
+if test ${ax_cv_check_cflags_libavcodec_avcodec_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS libavcodec/avcodec.h"
@@ -3465,36 +4060,39 @@ else
 /* end confdefs.h.  */
 #include <libavcodec/avcodec.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_libavcodec_avcodec_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_libavcodec_avcodec_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libavcodec_avcodec_h" >&5
-$as_echo "$ax_cv_check_cflags_libavcodec_avcodec_h" >&6; }
-	if test "x$ax_cv_check_cflags_libavcodec_avcodec_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libavcodec_avcodec_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_libavcodec_avcodec_h" >&6; }
+	if test "x$ax_cv_check_cflags_libavcodec_avcodec_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_AVCODEC "
  MKLDFLAGS="$MKLDFLAGS-lavcodec "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking libavformat/avformat.h presence" >&5
-$as_echo_n "checking libavformat/avformat.h presence... " >&6; }
-if ${ax_cv_check_cflags_libavformat_avformat_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking libavformat/avformat.h presence" >&5
+printf %s "checking libavformat/avformat.h presence... " >&6; }
+if test ${ax_cv_check_cflags_libavformat_avformat_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS libavformat/avformat.h"
@@ -3502,36 +4100,39 @@ else
 /* end confdefs.h.  */
 #include <libavformat/avformat.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_libavformat_avformat_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_libavformat_avformat_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libavformat_avformat_h" >&5
-$as_echo "$ax_cv_check_cflags_libavformat_avformat_h" >&6; }
-	if test "x$ax_cv_check_cflags_libavformat_avformat_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libavformat_avformat_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_libavformat_avformat_h" >&6; }
+	if test "x$ax_cv_check_cflags_libavformat_avformat_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_AVFORMAT "
  MKLDFLAGS="$MKLDFLAGS-lavformat "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking libavutil/avutil.h presence" >&5
-$as_echo_n "checking libavutil/avutil.h presence... " >&6; }
-if ${ax_cv_check_cflags_libavutil_avutil_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking libavutil/avutil.h presence" >&5
+printf %s "checking libavutil/avutil.h presence... " >&6; }
+if test ${ax_cv_check_cflags_libavutil_avutil_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS libavutil/avutil.h"
@@ -3539,36 +4140,39 @@ else
 /* end confdefs.h.  */
 #include <libavutil/avutil.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_libavutil_avutil_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_libavutil_avutil_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libavutil_avutil_h" >&5
-$as_echo "$ax_cv_check_cflags_libavutil_avutil_h" >&6; }
-	if test "x$ax_cv_check_cflags_libavutil_avutil_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libavutil_avutil_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_libavutil_avutil_h" >&6; }
+	if test "x$ax_cv_check_cflags_libavutil_avutil_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_AVUTIL "
  MKLDFLAGS="$MKLDFLAGS-lavutil "
 
-else
+else $as_nop
   :
 fi
 
- 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking libswscale/swscale.h presence" >&5
-$as_echo_n "checking libswscale/swscale.h presence... " >&6; }
-if ${ax_cv_check_cflags_libswscale_swscale_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+ 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking libswscale/swscale.h presence" >&5
+printf %s "checking libswscale/swscale.h presence... " >&6; }
+if test ${ax_cv_check_cflags_libswscale_swscale_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS libswscale/swscale.h"
@@ -3576,95 +4180,108 @@ else
 /* end confdefs.h.  */
 #include <libswscale/swscale.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_libswscale_swscale_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_libswscale_swscale_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libswscale_swscale_h" >&5
-$as_echo "$ax_cv_check_cflags_libswscale_swscale_h" >&6; }
-	if test "x$ax_cv_check_cflags_libswscale_swscale_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_libswscale_swscale_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_libswscale_swscale_h" >&6; }
+	if test "x$ax_cv_check_cflags_libswscale_swscale_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_SWSCALE "
  MKLDFLAGS="$MKLDFLAGS-lswscale "
 
-else
+else $as_nop
   :
 fi
 
 # try to find CPU parallel libraries, OpenMP or libdispatch, we will prefer OpenMP when possible.
-
-  OPENMP_CFLAGS=
-  # Check whether --enable-openmp was given.
-if test "${enable_openmp+set}" = set; then :
+if test -e penmp || test -e mp; then
+  as_fn_error $? "AC_OPENMP clobbers files named 'mp' and 'penmp'. Aborting configure because one of these files already exists." "$LINENO" 5
+fi
+# Check whether --enable-openmp was given.
+if test ${enable_openmp+y}
+then :
   enableval=$enable_openmp;
 fi
 
+  OPENMP_CFLAGS=
   if test "$enable_openmp" != no; then
-    { $as_echo "$as_me:${as_lineno-$LINENO}: checking for $CC option to support OpenMP" >&5
-$as_echo_n "checking for $CC option to support OpenMP... " >&6; }
-if ${ac_cv_prog_c_openmp+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
-  cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+    { printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $CC option to support OpenMP" >&5
+printf %s "checking for $CC option to support OpenMP... " >&6; }
+if test ${ac_cv_prog_c_openmp+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+  ac_cv_prog_c_openmp='not found'
+                                                                        for ac_option in '' -fopenmp -xopenmp -openmp -mp -omp -qsmp=omp -homp \
+                       -Popenmp --openmp; do
+
+        ac_save_CFLAGS=$CFLAGS
+        CFLAGS="$CFLAGS $ac_option"
+        cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 #ifndef _OPENMP
- choke me
+#error "OpenMP not supported"
 #endif
 #include <omp.h>
-int main () { return omp_get_num_threads (); }
+int main (void) { return omp_get_num_threads (); }
 
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
-  ac_cv_prog_c_openmp='none needed'
-else
-  ac_cv_prog_c_openmp='unsupported'
-	  	  	  	  	  	  	                                	  	  	  	  	  	  for ac_option in -fopenmp -xopenmp -openmp -mp -omp -qsmp=omp -homp \
-                           -Popenmp --openmp; do
-	    ac_save_CFLAGS=$CFLAGS
-	    CFLAGS="$CFLAGS $ac_option"
-	    cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+if ac_fn_c_try_compile "$LINENO"
+then :
+  cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* end confdefs.h.  */
 
 #ifndef _OPENMP
- choke me
+#error "OpenMP not supported"
 #endif
 #include <omp.h>
-int main () { return omp_get_num_threads (); }
+int main (void) { return omp_get_num_threads (); }
 
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_prog_c_openmp=$ac_option
+else $as_nop
+  ac_cv_prog_c_openmp='unsupported'
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
-	    CFLAGS=$ac_save_CFLAGS
-	    if test "$ac_cv_prog_c_openmp" != unsupported; then
-	      break
-	    fi
-	  done
 fi
-rm -f core conftest.err conftest.$ac_objext \
-    conftest$ac_exeext conftest.$ac_ext
+rm -f core conftest.err conftest.$ac_objext conftest.beam conftest.$ac_ext
+        CFLAGS=$ac_save_CFLAGS
+
+        if test "$ac_cv_prog_c_openmp" != 'not found'; then
+          break
+        fi
+      done
+      if test "$ac_cv_prog_c_openmp" = 'not found'; then
+        ac_cv_prog_c_openmp='unsupported'
+      elif test "$ac_cv_prog_c_openmp" = ''; then
+        ac_cv_prog_c_openmp='none needed'
+      fi
+                        rm -f penmp mp
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_c_openmp" >&5
-$as_echo "$ac_cv_prog_c_openmp" >&6; }
-    case $ac_cv_prog_c_openmp in #(
-      "none needed" | unsupported)
-	;; #(
-      *)
-	OPENMP_CFLAGS=$ac_cv_prog_c_openmp ;;
-    esac
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_prog_c_openmp" >&5
+printf "%s\n" "$ac_cv_prog_c_openmp" >&6; }
+    if test "$ac_cv_prog_c_openmp" != 'unsupported' && \
+       test "$ac_cv_prog_c_openmp" != 'none needed'; then
+      OPENMP_CFLAGS="$ac_cv_prog_c_openmp"
+    fi
   fi
 
 
@@ -3678,11 +4295,12 @@ if test -n "${OPENMP_CFLAGS}"; then
 fi
 # only check dispatch if compiled with clang
 if test $CC = clang; then
-	 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking dispatch/dispatch.h presence" >&5
-$as_echo_n "checking dispatch/dispatch.h presence... " >&6; }
-if ${ax_cv_check_cflags_dispatch_dispatch_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking dispatch/dispatch.h presence" >&5
+printf %s "checking dispatch/dispatch.h presence... " >&6; }
+if test ${ax_cv_check_cflags_dispatch_dispatch_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS dispatch/dispatch.h"
@@ -3690,36 +4308,39 @@ else
 /* end confdefs.h.  */
 #include <dispatch/dispatch.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_dispatch_dispatch_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_dispatch_dispatch_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_dispatch_dispatch_h" >&5
-$as_echo "$ax_cv_check_cflags_dispatch_dispatch_h" >&6; }
-	if test "x$ax_cv_check_cflags_dispatch_dispatch_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_dispatch_dispatch_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_dispatch_dispatch_h" >&6; }
+	if test "x$ax_cv_check_cflags_dispatch_dispatch_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D USE_DISPATCH "
  MKCFLAGS="$MKCFLAGS-fblocks "
 
-else
+else $as_nop
   :
 fi
 
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for dispatch_apply in -ldispatch" >&5
-$as_echo_n "checking for dispatch_apply in -ldispatch... " >&6; }
-if ${ac_cv_lib_dispatch_dispatch_apply+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for dispatch_apply in -ldispatch" >&5
+printf %s "checking for dispatch_apply in -ldispatch... " >&6; }
+if test ${ac_cv_lib_dispatch_dispatch_apply+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-ldispatch  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -3728,30 +4349,29 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char dispatch_apply ();
 int
-main ()
+main (void)
 {
 return dispatch_apply ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_dispatch_dispatch_apply=yes
-else
+else $as_nop
   ac_cv_lib_dispatch_dispatch_apply=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_dispatch_dispatch_apply" >&5
-$as_echo "$ac_cv_lib_dispatch_dispatch_apply" >&6; }
-if test "x$ac_cv_lib_dispatch_dispatch_apply" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_dispatch_dispatch_apply" >&5
+printf "%s\n" "$ac_cv_lib_dispatch_dispatch_apply" >&6; }
+if test "x$ac_cv_lib_dispatch_dispatch_apply" = xyes
+then :
   MKLDFLAGS="$MKLDFLAGS-ldispatch -lBlocksRuntime "
 
 fi
@@ -3759,11 +4379,12 @@ fi
 fi
 # check for SSE2 support only we don't enable NEON explicitly
 if test "$neon_support" != yes; then
-	 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking xmmintrin.h presence" >&5
-$as_echo_n "checking xmmintrin.h presence... " >&6; }
-if ${ax_cv_check_cflags_xmmintrin_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking xmmintrin.h presence" >&5
+printf %s "checking xmmintrin.h presence... " >&6; }
+if test ${ax_cv_check_cflags_xmmintrin_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS xmmintrin.h"
@@ -3771,49 +4392,53 @@ else
 /* end confdefs.h.  */
 #include <xmmintrin.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_xmmintrin_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_xmmintrin_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_xmmintrin_h" >&5
-$as_echo "$ax_cv_check_cflags_xmmintrin_h" >&6; }
-	if test "x$ax_cv_check_cflags_xmmintrin_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_xmmintrin_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_xmmintrin_h" >&6; }
+	if test "x$ax_cv_check_cflags_xmmintrin_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_SSE2 "
  MKCFLAGS="$MKCFLAGS-msse2 "
 
-else
+else $as_nop
   :
 fi
 
 fi
 
 # check for gsl, and I need to first check these two before I can check gsl
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking gsl" >&5
-$as_echo_n "checking gsl... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking gsl" >&5
+printf %s "checking gsl... " >&6; }
 # Check whether --enable-gsl was given.
-if test "${enable_gsl+set}" = set; then :
+if test ${enable_gsl+y}
+then :
   enableval=$enable_gsl; gsl_enable=$enableval
-else
+else $as_nop
   gsl_enable="yes"
 fi
 
 if test "$gsl_enable" != no; then
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for cos in -lm" >&5
-$as_echo_n "checking for cos in -lm... " >&6; }
-if ${ac_cv_lib_m_cos+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for cos in -lm" >&5
+printf %s "checking for cos in -lm... " >&6; }
+if test ${ac_cv_lib_m_cos+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-lm  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -3822,43 +4447,41 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char cos ();
 int
-main ()
+main (void)
 {
 return cos ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_m_cos=yes
-else
+else $as_nop
   ac_cv_lib_m_cos=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_m_cos" >&5
-$as_echo "$ac_cv_lib_m_cos" >&6; }
-if test "x$ac_cv_lib_m_cos" = xyes; then :
-  cat >>confdefs.h <<_ACEOF
-#define HAVE_LIBM 1
-_ACEOF
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_m_cos" >&5
+printf "%s\n" "$ac_cv_lib_m_cos" >&6; }
+if test "x$ac_cv_lib_m_cos" = xyes
+then :
+  printf "%s\n" "#define HAVE_LIBM 1" >>confdefs.h
 
   LIBS="-lm $LIBS"
 
 fi
 
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for cblas_dgemm in -lgslcblas" >&5
-$as_echo_n "checking for cblas_dgemm in -lgslcblas... " >&6; }
-if ${ac_cv_lib_gslcblas_cblas_dgemm+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for cblas_dgemm in -lgslcblas" >&5
+printf %s "checking for cblas_dgemm in -lgslcblas... " >&6; }
+if test ${ac_cv_lib_gslcblas_cblas_dgemm+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-lgslcblas  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -3867,43 +4490,41 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char cblas_dgemm ();
 int
-main ()
+main (void)
 {
 return cblas_dgemm ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_gslcblas_cblas_dgemm=yes
-else
+else $as_nop
   ac_cv_lib_gslcblas_cblas_dgemm=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_gslcblas_cblas_dgemm" >&5
-$as_echo "$ac_cv_lib_gslcblas_cblas_dgemm" >&6; }
-if test "x$ac_cv_lib_gslcblas_cblas_dgemm" = xyes; then :
-  cat >>confdefs.h <<_ACEOF
-#define HAVE_LIBGSLCBLAS 1
-_ACEOF
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_gslcblas_cblas_dgemm" >&5
+printf "%s\n" "$ac_cv_lib_gslcblas_cblas_dgemm" >&6; }
+if test "x$ac_cv_lib_gslcblas_cblas_dgemm" = xyes
+then :
+  printf "%s\n" "#define HAVE_LIBGSLCBLAS 1" >>confdefs.h
 
   LIBS="-lgslcblas $LIBS"
 
 fi
 
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for gsl_blas_dgemm in -lgsl" >&5
-$as_echo_n "checking for gsl_blas_dgemm in -lgsl... " >&6; }
-if ${ac_cv_lib_gsl_gsl_blas_dgemm+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for gsl_blas_dgemm in -lgsl" >&5
+printf %s "checking for gsl_blas_dgemm in -lgsl... " >&6; }
+if test ${ac_cv_lib_gsl_gsl_blas_dgemm+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   ac_check_lib_save_LIBS=$LIBS
 LIBS="-lgsl  $LIBS"
 cat confdefs.h - <<_ACEOF >conftest.$ac_ext
@@ -3912,48 +4533,48 @@ cat confdefs.h - <<_ACEOF >conftest.$ac_ext
 /* Override any GCC internal prototype to avoid an error.
    Use char because int might match the return type of a GCC
    builtin and then its argument prototype would still apply.  */
-#ifdef __cplusplus
-extern "C"
-#endif
 char gsl_blas_dgemm ();
 int
-main ()
+main (void)
 {
 return gsl_blas_dgemm ();
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_link "$LINENO"; then :
+if ac_fn_c_try_link "$LINENO"
+then :
   ac_cv_lib_gsl_gsl_blas_dgemm=yes
-else
+else $as_nop
   ac_cv_lib_gsl_gsl_blas_dgemm=no
 fi
-rm -f core conftest.err conftest.$ac_objext \
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
     conftest$ac_exeext conftest.$ac_ext
 LIBS=$ac_check_lib_save_LIBS
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_gsl_gsl_blas_dgemm" >&5
-$as_echo "$ac_cv_lib_gsl_gsl_blas_dgemm" >&6; }
-if test "x$ac_cv_lib_gsl_gsl_blas_dgemm" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_gsl_gsl_blas_dgemm" >&5
+printf "%s\n" "$ac_cv_lib_gsl_gsl_blas_dgemm" >&6; }
+if test "x$ac_cv_lib_gsl_gsl_blas_dgemm" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_GSL "
  MKLDFLAGS="$MKLDFLAGS-lgsl -lgslcblas "
 
 fi
 
 else
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: disabled" >&5
-$as_echo "disabled" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: disabled" >&5
+printf "%s\n" "disabled" >&6; }
 fi
 
 # prepare for cuda
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking cuda" >&5
-$as_echo_n "checking cuda... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking cuda" >&5
+printf %s "checking cuda... " >&6; }
 
 # Check whether --with-cuda was given.
-if test "${with_cuda+set}" = set; then :
+if test ${with_cuda+y}
+then :
   withval=$with_cuda; cuda_prefix=$withval
-else
+else $as_nop
   cuda_prefix="/usr/local/cuda"
 fi
 
@@ -3968,19 +4589,20 @@ if [ -d "$cuda_prefix" ]; then
 	if [ -d "$cuda_prefix/lib64" ]; then
 		MKLDFLAGS="$MKLDFLAGS-L$cuda_prefix/lib64 "
 
-		{ $as_echo "$as_me:${as_lineno-$LINENO}: result: yes, x86_64" >&5
-$as_echo "yes, x86_64" >&6; }
+		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: yes, x86_64" >&5
+printf "%s\n" "yes, x86_64" >&6; }
 	else
 		MKLDFLAGS="$MKLDFLAGS-L$cuda_prefix/lib "
 
-		{ $as_echo "$as_me:${as_lineno-$LINENO}: result: yes, i386" >&5
-$as_echo "yes, i386" >&6; }
+		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: yes, i386" >&5
+printf "%s\n" "yes, i386" >&6; }
 	fi
-	 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking cudnn.h presence" >&5
-$as_echo_n "checking cudnn.h presence... " >&6; }
-if ${ax_cv_check_cflags_cudnn_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking cudnn.h presence" >&5
+printf %s "checking cudnn.h presence... " >&6; }
+if test ${ax_cv_check_cflags_cudnn_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS cudnn.h"
@@ -3988,36 +4610,39 @@ else
 /* end confdefs.h.  */
 #include <cudnn.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_cudnn_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_cudnn_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_cudnn_h" >&5
-$as_echo "$ax_cv_check_cflags_cudnn_h" >&6; }
-	if test "x$ax_cv_check_cflags_cudnn_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_cudnn_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_cudnn_h" >&6; }
+	if test "x$ax_cv_check_cflags_cudnn_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_CUDNN "
  MKLDFLAGS="$MKLDFLAGS-lcudnn "
 
-else
+else $as_nop
   :
 fi
 
-	 		{ $as_echo "$as_me:${as_lineno-$LINENO}: checking nccl.h presence" >&5
-$as_echo_n "checking nccl.h presence... " >&6; }
-if ${ax_cv_check_cflags_nccl_h+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking nccl.h presence" >&5
+printf %s "checking nccl.h presence... " >&6; }
+if test ${ax_cv_check_cflags_nccl_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
 
 		ax_check_save_flags=$CFLAGS
 		CFLAGS="$CFLAGS nccl.h"
@@ -4025,37 +4650,40 @@ else
 /* end confdefs.h.  */
 #include <nccl.h>
 int
-main ()
+main (void)
 {
 
   ;
   return 0;
 }
 _ACEOF
-if ac_fn_c_try_cpp "$LINENO"; then :
+if ac_fn_c_try_cpp "$LINENO"
+then :
   ax_cv_check_cflags_nccl_h=yes
-else
+else $as_nop
   ax_cv_check_cflags_nccl_h=no
 fi
 rm -f conftest.err conftest.i conftest.$ac_ext
 		CFLAGS=$ax_check_save_flags
 fi
-{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_nccl_h" >&5
-$as_echo "$ax_cv_check_cflags_nccl_h" >&6; }
-	if test "x$ax_cv_check_cflags_nccl_h" = xyes; then :
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_nccl_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_nccl_h" >&6; }
+	if test "x$ax_cv_check_cflags_nccl_h" = xyes
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_NCCL "
  MKLDFLAGS="$MKLDFLAGS-lnccl "
 
-else
+else $as_nop
   :
 fi
 
-	as_ac_File=`$as_echo "ac_cv_file_$cuda_prefix/include/cub/cub.cuh" | $as_tr_sh`
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking for $cuda_prefix/include/cub/cub.cuh" >&5
-$as_echo_n "checking for $cuda_prefix/include/cub/cub.cuh... " >&6; }
-if eval \${$as_ac_File+:} false; then :
-  $as_echo_n "(cached) " >&6
-else
+	as_ac_File=`printf "%s\n" "ac_cv_file_$cuda_prefix/include/cub/cub.cuh" | $as_tr_sh`
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for $cuda_prefix/include/cub/cub.cuh" >&5
+printf %s "checking for $cuda_prefix/include/cub/cub.cuh... " >&6; }
+if eval test \${$as_ac_File+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
   test "$cross_compiling" = yes &&
   as_fn_error $? "cannot check for file existence when cross compiling" "$LINENO" 5
 if test -r "$cuda_prefix/include/cub/cub.cuh"; then
@@ -4065,9 +4693,10 @@ else
 fi
 fi
 eval ac_res=\$$as_ac_File
-	       { $as_echo "$as_me:${as_lineno-$LINENO}: result: $ac_res" >&5
-$as_echo "$ac_res" >&6; }
-if eval test \"x\$"$as_ac_File"\" = x"yes"; then :
+	       { printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_res" >&5
+printf "%s\n" "$ac_res" >&6; }
+if eval test \"x\$"$as_ac_File"\" = x"yes"
+then :
   DEFINE_MACROS="$DEFINE_MACROS-D USE_SYSTEM_CUB "
 
 fi
@@ -4079,26 +4708,27 @@ else
 
 	CUDA_CMD_LIB=""
 
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
 fi
 
 # check for MPS support
-{ $as_echo "$as_me:${as_lineno-$LINENO}: checking Metal Performance Shaders" >&5
-$as_echo_n "checking Metal Performance Shaders... " >&6; }
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking Metal Performance Shaders" >&5
+printf %s "checking Metal Performance Shaders... " >&6; }
 # Check whether --enable-mps was given.
-if test "${enable_mps+set}" = set; then :
+if test ${enable_mps+y}
+then :
   enableval=$enable_mps; mps_support=$enableval
-else
+else $as_nop
   mps_support="no"
 fi
 
 if test "$mps_support" = yes; then
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: yes" >&5
-$as_echo "yes" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: yes" >&5
+printf "%s\n" "yes" >&6; }
 	DEFINE_MACROS="$DEFINE_MACROS-D HAVE_MPS "
 
-	MKLDFLAGS="$MKLDFLAGS-framework MetalPerformanceShaders -framework MetalPerformanceShadersGraph -framework Foundation -framework Metal "
+	MKLDFLAGS="$MKLDFLAGS-framework MetalPerformanceShaders -framework MetalPerformanceShadersGraph -framework Foundation -framework Metal -lc++ "
 
 	CUDA_SRCS=""
 
@@ -4107,8 +4737,10 @@ $as_echo "yes" >&6; }
 	CUDA_CMD_LIB=""
 
 else
-	{ $as_echo "$as_me:${as_lineno-$LINENO}: result: no" >&5
-$as_echo "no" >&6; }
+	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
+printf "%s\n" "no" >&6; }
+	MFA_COMPAT_LIB=""
+
 	MPS_COMPAT_LIB=""
 
 	MPS_CMD_LIB=""
@@ -4173,8 +4805,8 @@ _ACEOF
     case $ac_val in #(
     *${as_nl}*)
       case $ac_var in #(
-      *_cv_*) { $as_echo "$as_me:${as_lineno-$LINENO}: WARNING: cache variable $ac_var contains a newline" >&5
-$as_echo "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
+      *_cv_*) { printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: cache variable $ac_var contains a newline" >&5
+printf "%s\n" "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
       esac
       case $ac_var in #(
       _ | IFS | as_nl) ;; #(
@@ -4204,15 +4836,15 @@ $as_echo "$as_me: WARNING: cache variable $ac_var contains a newline" >&2;} ;;
      /^ac_cv_env_/b end
      t clear
      :clear
-     s/^\([^=]*\)=\(.*[{}].*\)$/test "${\1+set}" = set || &/
+     s/^\([^=]*\)=\(.*[{}].*\)$/test ${\1+y} || &/
      t end
      s/^\([^=]*\)=\(.*\)$/\1=${\1=\2}/
      :end' >>confcache
 if diff "$cache_file" confcache >/dev/null 2>&1; then :; else
   if test -w "$cache_file"; then
     if test "x$cache_file" != "x/dev/null"; then
-      { $as_echo "$as_me:${as_lineno-$LINENO}: updating cache $cache_file" >&5
-$as_echo "$as_me: updating cache $cache_file" >&6;}
+      { printf "%s\n" "$as_me:${as_lineno-$LINENO}: updating cache $cache_file" >&5
+printf "%s\n" "$as_me: updating cache $cache_file" >&6;}
       if test ! -f "$cache_file" || test -h "$cache_file"; then
 	cat confcache >"$cache_file"
       else
@@ -4226,8 +4858,8 @@ $as_echo "$as_me: updating cache $cache_file" >&6;}
       fi
     fi
   else
-    { $as_echo "$as_me:${as_lineno-$LINENO}: not updating unwritable cache $cache_file" >&5
-$as_echo "$as_me: not updating unwritable cache $cache_file" >&6;}
+    { printf "%s\n" "$as_me:${as_lineno-$LINENO}: not updating unwritable cache $cache_file" >&5
+printf "%s\n" "$as_me: not updating unwritable cache $cache_file" >&6;}
   fi
 fi
 rm -f confcache
@@ -4280,7 +4912,7 @@ U=
 for ac_i in : $LIBOBJS; do test "x$ac_i" = x: && continue
   # 1. Remove the extension, and $U if already installed.
   ac_script='s/\$U\././;s/\.o$//;s/\.obj$//'
-  ac_i=`$as_echo "$ac_i" | sed "$ac_script"`
+  ac_i=`printf "%s\n" "$ac_i" | sed "$ac_script"`
   # 2. Prepend LIBOBJDIR.  When used with automake>=1.10 LIBOBJDIR
   #    will be set to the directory where LIBOBJS objects are built.
   as_fn_append ac_libobjs " \${LIBOBJDIR}$ac_i\$U.$ac_objext"
@@ -4296,8 +4928,8 @@ LTLIBOBJS=$ac_ltlibobjs
 ac_write_fail=0
 ac_clean_files_save=$ac_clean_files
 ac_clean_files="$ac_clean_files $CONFIG_STATUS"
-{ $as_echo "$as_me:${as_lineno-$LINENO}: creating $CONFIG_STATUS" >&5
-$as_echo "$as_me: creating $CONFIG_STATUS" >&6;}
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: creating $CONFIG_STATUS" >&5
+printf "%s\n" "$as_me: creating $CONFIG_STATUS" >&6;}
 as_write_fail=0
 cat >$CONFIG_STATUS <<_ASEOF || as_write_fail=1
 #! $SHELL
@@ -4320,14 +4952,16 @@ cat >>$CONFIG_STATUS <<\_ASEOF || as_write_fail=1
 
 # Be more Bourne compatible
 DUALCASE=1; export DUALCASE # for MKS sh
-if test -n "${ZSH_VERSION+set}" && (emulate sh) >/dev/null 2>&1; then :
+as_nop=:
+if test ${ZSH_VERSION+y} && (emulate sh) >/dev/null 2>&1
+then :
   emulate sh
   NULLCMD=:
   # Pre-4.2 versions of Zsh do word splitting on ${1+"$@"}, which
   # is contrary to our usage.  Disable this feature.
   alias -g '${1+"$@"}'='"$@"'
   setopt NO_GLOB_SUBST
-else
+else $as_nop
   case `(set -o) 2>/dev/null` in #(
   *posix*) :
     set -o posix ;; #(
@@ -4337,46 +4971,46 @@ esac
 fi
 
 
+
+# Reset variables that may have inherited troublesome values from
+# the environment.
+
+# IFS needs to be set, to space, tab, and newline, in precisely that order.
+# (If _AS_PATH_WALK were called with IFS unset, it would have the
+# side effect of setting IFS to empty, thus disabling word splitting.)
+# Quoting is to prevent editors from complaining about space-tab.
 as_nl='
 '
 export as_nl
-# Printing a long string crashes Solaris 7 /usr/bin/printf.
-as_echo='\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\'
-as_echo=$as_echo$as_echo$as_echo$as_echo$as_echo
-as_echo=$as_echo$as_echo$as_echo$as_echo$as_echo$as_echo
-# Prefer a ksh shell builtin over an external printf program on Solaris,
-# but without wasting forks for bash or zsh.
-if test -z "$BASH_VERSION$ZSH_VERSION" \
-    && (test "X`print -r -- $as_echo`" = "X$as_echo") 2>/dev/null; then
-  as_echo='print -r --'
-  as_echo_n='print -rn --'
-elif (test "X`printf %s $as_echo`" = "X$as_echo") 2>/dev/null; then
-  as_echo='printf %s\n'
-  as_echo_n='printf %s'
-else
-  if test "X`(/usr/ucb/echo -n -n $as_echo) 2>/dev/null`" = "X-n $as_echo"; then
-    as_echo_body='eval /usr/ucb/echo -n "$1$as_nl"'
-    as_echo_n='/usr/ucb/echo -n'
-  else
-    as_echo_body='eval expr "X$1" : "X\\(.*\\)"'
-    as_echo_n_body='eval
-      arg=$1;
-      case $arg in #(
-      *"$as_nl"*)
-	expr "X$arg" : "X\\(.*\\)$as_nl";
-	arg=`expr "X$arg" : ".*$as_nl\\(.*\\)"`;;
-      esac;
-      expr "X$arg" : "X\\(.*\\)" | tr -d "$as_nl"
-    '
-    export as_echo_n_body
-    as_echo_n='sh -c $as_echo_n_body as_echo'
-  fi
-  export as_echo_body
-  as_echo='sh -c $as_echo_body as_echo'
-fi
+IFS=" ""	$as_nl"
+
+PS1='$ '
+PS2='> '
+PS4='+ '
+
+# Ensure predictable behavior from utilities with locale-dependent output.
+LC_ALL=C
+export LC_ALL
+LANGUAGE=C
+export LANGUAGE
+
+# We cannot yet rely on "unset" to work, but we need these variables
+# to be unset--not just set to an empty or harmless value--now, to
+# avoid bugs in old shells (e.g. pre-3.0 UWIN ksh).  This construct
+# also avoids known problems related to "unset" and subshell syntax
+# in other old shells (e.g. bash 2.01 and pdksh 5.2.14).
+for as_var in BASH_ENV ENV MAIL MAILPATH CDPATH
+do eval test \${$as_var+y} \
+  && ( (unset $as_var) || exit 1) >/dev/null 2>&1 && unset $as_var || :
+done
+
+# Ensure that fds 0, 1, and 2 are open.
+if (exec 3>&0) 2>/dev/null; then :; else exec 0</dev/null; fi
+if (exec 3>&1) 2>/dev/null; then :; else exec 1>/dev/null; fi
+if (exec 3>&2)            ; then :; else exec 2>/dev/null; fi
 
 # The user is always right.
-if test "${PATH_SEPARATOR+set}" != set; then
+if ${PATH_SEPARATOR+false} :; then
   PATH_SEPARATOR=:
   (PATH='/bin;/bin'; FPATH=$PATH; sh -c :) >/dev/null 2>&1 && {
     (PATH='/bin:/bin'; FPATH=$PATH; sh -c :) >/dev/null 2>&1 ||
@@ -4385,13 +5019,6 @@ if test "${PATH_SEPARATOR+set}" != set; then
 fi
 
 
-# IFS
-# We need space, tab and new line, in precisely that order.  Quoting is
-# there to prevent editors from complaining about space-tab.
-# (If _AS_PATH_WALK were called with IFS unset, it would disable word
-# splitting by setting IFS to empty value.)
-IFS=" ""	$as_nl"
-
 # Find who we are.  Look in the path if we contain no directory separator.
 as_myself=
 case $0 in #((
@@ -4400,8 +5027,12 @@ case $0 in #((
 for as_dir in $PATH
 do
   IFS=$as_save_IFS
-  test -z "$as_dir" && as_dir=.
-    test -r "$as_dir/$0" && as_myself=$as_dir/$0 && break
+  case $as_dir in #(((
+    '') as_dir=./ ;;
+    */) ;;
+    *) as_dir=$as_dir/ ;;
+  esac
+    test -r "$as_dir$0" && as_myself=$as_dir$0 && break
   done
 IFS=$as_save_IFS
 
@@ -4413,30 +5044,10 @@ if test "x$as_myself" = x; then
   as_myself=$0
 fi
 if test ! -f "$as_myself"; then
-  $as_echo "$as_myself: error: cannot find myself; rerun with an absolute file name" >&2
+  printf "%s\n" "$as_myself: error: cannot find myself; rerun with an absolute file name" >&2
   exit 1
 fi
 
-# Unset variables that we do not need and which cause bugs (e.g. in
-# pre-3.0 UWIN ksh).  But do not cause bugs in bash 2.01; the "|| exit 1"
-# suppresses any "Segmentation fault" message there.  '((' could
-# trigger a bug in pdksh 5.2.14.
-for as_var in BASH_ENV ENV MAIL MAILPATH
-do eval test x\${$as_var+set} = xset \
-  && ( (unset $as_var) || exit 1) >/dev/null 2>&1 && unset $as_var || :
-done
-PS1='$ '
-PS2='> '
-PS4='+ '
-
-# NLS nuisances.
-LC_ALL=C
-export LC_ALL
-LANGUAGE=C
-export LANGUAGE
-
-# CDPATH.
-(unset CDPATH) >/dev/null 2>&1 && unset CDPATH
 
 
 # as_fn_error STATUS ERROR [LINENO LOG_FD]
@@ -4449,13 +5060,14 @@ as_fn_error ()
   as_status=$1; test $as_status -eq 0 && as_status=1
   if test "$4"; then
     as_lineno=${as_lineno-"$3"} as_lineno_stack=as_lineno_stack=$as_lineno_stack
-    $as_echo "$as_me:${as_lineno-$LINENO}: error: $2" >&$4
+    printf "%s\n" "$as_me:${as_lineno-$LINENO}: error: $2" >&$4
   fi
-  $as_echo "$as_me: error: $2" >&2
+  printf "%s\n" "$as_me: error: $2" >&2
   as_fn_exit $as_status
 } # as_fn_error
 
 
+
 # as_fn_set_status STATUS
 # -----------------------
 # Set $? to STATUS, without forking.
@@ -4482,18 +5094,20 @@ as_fn_unset ()
   { eval $1=; unset $1;}
 }
 as_unset=as_fn_unset
+
 # as_fn_append VAR VALUE
 # ----------------------
 # Append the text in VALUE to the end of the definition contained in VAR. Take
 # advantage of any shell optimizations that allow amortized linear growth over
 # repeated appends, instead of the typical quadratic growth present in naive
 # implementations.
-if (eval "as_var=1; as_var+=2; test x\$as_var = x12") 2>/dev/null; then :
+if (eval "as_var=1; as_var+=2; test x\$as_var = x12") 2>/dev/null
+then :
   eval 'as_fn_append ()
   {
     eval $1+=\$2
   }'
-else
+else $as_nop
   as_fn_append ()
   {
     eval $1=\$$1\$2
@@ -4505,12 +5119,13 @@ fi # as_fn_append
 # Perform arithmetic evaluation on the ARGs, and store the result in the
 # global $as_val. Take advantage of shells that can avoid forks. The arguments
 # must be portable across $(()) and expr.
-if (eval "test \$(( 1 + 1 )) = 2") 2>/dev/null; then :
+if (eval "test \$(( 1 + 1 )) = 2") 2>/dev/null
+then :
   eval 'as_fn_arith ()
   {
     as_val=$(( $* ))
   }'
-else
+else $as_nop
   as_fn_arith ()
   {
     as_val=`expr "$@" || test $? -eq 1`
@@ -4541,7 +5156,7 @@ as_me=`$as_basename -- "$0" ||
 $as_expr X/"$0" : '.*/\([^/][^/]*\)/*$' \| \
 	 X"$0" : 'X\(//\)$' \| \
 	 X"$0" : 'X\(/\)' \| . 2>/dev/null ||
-$as_echo X/"$0" |
+printf "%s\n" X/"$0" |
     sed '/^.*\/\([^/][^/]*\)\/*$/{
 	    s//\1/
 	    q
@@ -4563,6 +5178,10 @@ as_cr_Letters=$as_cr_letters$as_cr_LETTERS
 as_cr_digits='0123456789'
 as_cr_alnum=$as_cr_Letters$as_cr_digits
 
+
+# Determine whether it's possible to make 'echo' print without a newline.
+# These variables are no longer used directly by Autoconf, but are AC_SUBSTed
+# for compatibility with existing Makefiles.
 ECHO_C= ECHO_N= ECHO_T=
 case `echo -n x` in #(((((
 -n*)
@@ -4576,6 +5195,12 @@ case `echo -n x` in #(((((
   ECHO_N='-n';;
 esac
 
+# For backward compatibility with old third-party macros, we provide
+# the shell variables $as_echo and $as_echo_n.  New code should use
+# AS_ECHO(["message"]) and AS_ECHO_N(["message"]), respectively.
+as_echo='printf %s\n'
+as_echo_n='printf %s'
+
 rm -f conf$$ conf$$.exe conf$$.file
 if test -d conf$$.dir; then
   rm -f conf$$.dir/conf$$.file
@@ -4617,7 +5242,7 @@ as_fn_mkdir_p ()
     as_dirs=
     while :; do
       case $as_dir in #(
-      *\'*) as_qdir=`$as_echo "$as_dir" | sed "s/'/'\\\\\\\\''/g"`;; #'(
+      *\'*) as_qdir=`printf "%s\n" "$as_dir" | sed "s/'/'\\\\\\\\''/g"`;; #'(
       *) as_qdir=$as_dir;;
       esac
       as_dirs="'$as_qdir' $as_dirs"
@@ -4626,7 +5251,7 @@ $as_expr X"$as_dir" : 'X\(.*[^/]\)//*[^/][^/]*/*$' \| \
 	 X"$as_dir" : 'X\(//\)[^/]' \| \
 	 X"$as_dir" : 'X\(//\)$' \| \
 	 X"$as_dir" : 'X\(/\)' \| . 2>/dev/null ||
-$as_echo X"$as_dir" |
+printf "%s\n" X"$as_dir" |
     sed '/^X\(.*[^/]\)\/\/*[^/][^/]*\/*$/{
 	    s//\1/
 	    q
@@ -4689,7 +5314,7 @@ cat >>$CONFIG_STATUS <<\_ACEOF || ac_write_fail=1
 # values after options handling.
 ac_log="
 This file was extended by libccv $as_me 0.7, which was
-generated by GNU Autoconf 2.69.  Invocation command line was
+generated by GNU Autoconf 2.71.  Invocation command line was
 
   CONFIG_FILES    = $CONFIG_FILES
   CONFIG_HEADERS  = $CONFIG_HEADERS
@@ -4738,14 +5363,16 @@ $config_files
 Report bugs to the package provider."
 
 _ACEOF
+ac_cs_config=`printf "%s\n" "$ac_configure_args" | sed "$ac_safe_unquote"`
+ac_cs_config_escaped=`printf "%s\n" "$ac_cs_config" | sed "s/^ //; s/'/'\\\\\\\\''/g"`
 cat >>$CONFIG_STATUS <<_ACEOF || ac_write_fail=1
-ac_cs_config="`$as_echo "$ac_configure_args" | sed 's/^ //; s/[\\""\`\$]/\\\\&/g'`"
+ac_cs_config='$ac_cs_config_escaped'
 ac_cs_version="\\
 libccv config.status 0.7
-configured by $0, generated by GNU Autoconf 2.69,
+configured by $0, generated by GNU Autoconf 2.71,
   with options \\"\$ac_cs_config\\"
 
-Copyright (C) 2012 Free Software Foundation, Inc.
+Copyright (C) 2021 Free Software Foundation, Inc.
 This config.status script is free software; the Free Software Foundation
 gives unlimited permission to copy, distribute and modify it."
 
@@ -4782,21 +5409,21 @@ do
   -recheck | --recheck | --rechec | --reche | --rech | --rec | --re | --r)
     ac_cs_recheck=: ;;
   --version | --versio | --versi | --vers | --ver | --ve | --v | -V )
-    $as_echo "$ac_cs_version"; exit ;;
+    printf "%s\n" "$ac_cs_version"; exit ;;
   --config | --confi | --conf | --con | --co | --c )
-    $as_echo "$ac_cs_config"; exit ;;
+    printf "%s\n" "$ac_cs_config"; exit ;;
   --debug | --debu | --deb | --de | --d | -d )
     debug=: ;;
   --file | --fil | --fi | --f )
     $ac_shift
     case $ac_optarg in
-    *\'*) ac_optarg=`$as_echo "$ac_optarg" | sed "s/'/'\\\\\\\\''/g"` ;;
+    *\'*) ac_optarg=`printf "%s\n" "$ac_optarg" | sed "s/'/'\\\\\\\\''/g"` ;;
     '') as_fn_error $? "missing file argument" ;;
     esac
     as_fn_append CONFIG_FILES " '$ac_optarg'"
     ac_need_defaults=false;;
   --he | --h |  --help | --hel | -h )
-    $as_echo "$ac_cs_usage"; exit ;;
+    printf "%s\n" "$ac_cs_usage"; exit ;;
   -q | -quiet | --quiet | --quie | --qui | --qu | --q \
   | -silent | --silent | --silen | --sile | --sil | --si | --s)
     ac_cs_silent=: ;;
@@ -4824,7 +5451,7 @@ cat >>$CONFIG_STATUS <<_ACEOF || ac_write_fail=1
 if \$ac_cs_recheck; then
   set X $SHELL '$0' $ac_configure_args \$ac_configure_extra_args --no-create --no-recursion
   shift
-  \$as_echo "running CONFIG_SHELL=$SHELL \$*" >&6
+  \printf "%s\n" "running CONFIG_SHELL=$SHELL \$*" >&6
   CONFIG_SHELL='$SHELL'
   export CONFIG_SHELL
   exec "\$@"
@@ -4838,7 +5465,7 @@ exec 5>>config.log
   sed 'h;s/./-/g;s/^.../## /;s/...$/ ##/;p;x;p;x' <<_ASBOX
 ## Running $as_me. ##
 _ASBOX
-  $as_echo "$ac_log"
+  printf "%s\n" "$ac_log"
 } >&5
 
 _ACEOF
@@ -4863,7 +5490,7 @@ done
 # We use the long form for the default assignment because of an extremely
 # bizarre bug on SunOS 4.1.3.
 if $ac_need_defaults; then
-  test "${CONFIG_FILES+set}" = set || CONFIG_FILES=$config_files
+  test ${CONFIG_FILES+y} || CONFIG_FILES=$config_files
 fi
 
 # Have a temporary directory for convenience.  Make it in the build tree
@@ -5091,7 +5718,7 @@ do
 	   esac ||
 	   as_fn_error 1 "cannot find input file: \`$ac_f'" "$LINENO" 5;;
       esac
-      case $ac_f in *\'*) ac_f=`$as_echo "$ac_f" | sed "s/'/'\\\\\\\\''/g"`;; esac
+      case $ac_f in *\'*) ac_f=`printf "%s\n" "$ac_f" | sed "s/'/'\\\\\\\\''/g"`;; esac
       as_fn_append ac_file_inputs " '$ac_f'"
     done
 
@@ -5099,17 +5726,17 @@ do
     # use $as_me), people would be surprised to read:
     #    /* config.h.  Generated by config.status.  */
     configure_input='Generated from '`
-	  $as_echo "$*" | sed 's|^[^:]*/||;s|:[^:]*/|, |g'
+	  printf "%s\n" "$*" | sed 's|^[^:]*/||;s|:[^:]*/|, |g'
 	`' by configure.'
     if test x"$ac_file" != x-; then
       configure_input="$ac_file.  $configure_input"
-      { $as_echo "$as_me:${as_lineno-$LINENO}: creating $ac_file" >&5
-$as_echo "$as_me: creating $ac_file" >&6;}
+      { printf "%s\n" "$as_me:${as_lineno-$LINENO}: creating $ac_file" >&5
+printf "%s\n" "$as_me: creating $ac_file" >&6;}
     fi
     # Neutralize special characters interpreted by sed in replacement strings.
     case $configure_input in #(
     *\&* | *\|* | *\\* )
-       ac_sed_conf_input=`$as_echo "$configure_input" |
+       ac_sed_conf_input=`printf "%s\n" "$configure_input" |
        sed 's/[\\\\&|]/\\\\&/g'`;; #(
     *) ac_sed_conf_input=$configure_input;;
     esac
@@ -5126,7 +5753,7 @@ $as_expr X"$ac_file" : 'X\(.*[^/]\)//*[^/][^/]*/*$' \| \
 	 X"$ac_file" : 'X\(//\)[^/]' \| \
 	 X"$ac_file" : 'X\(//\)$' \| \
 	 X"$ac_file" : 'X\(/\)' \| . 2>/dev/null ||
-$as_echo X"$ac_file" |
+printf "%s\n" X"$ac_file" |
     sed '/^X\(.*[^/]\)\/\/*[^/][^/]*\/*$/{
 	    s//\1/
 	    q
@@ -5150,9 +5777,9 @@ $as_echo X"$ac_file" |
 case "$ac_dir" in
 .) ac_dir_suffix= ac_top_builddir_sub=. ac_top_build_prefix= ;;
 *)
-  ac_dir_suffix=/`$as_echo "$ac_dir" | sed 's|^\.[\\/]||'`
+  ac_dir_suffix=/`printf "%s\n" "$ac_dir" | sed 's|^\.[\\/]||'`
   # A ".." for each directory in $ac_dir_suffix.
-  ac_top_builddir_sub=`$as_echo "$ac_dir_suffix" | sed 's|/[^\\/]*|/..|g;s|/||'`
+  ac_top_builddir_sub=`printf "%s\n" "$ac_dir_suffix" | sed 's|/[^\\/]*|/..|g;s|/||'`
   case $ac_top_builddir_sub in
   "") ac_top_builddir_sub=. ac_top_build_prefix= ;;
   *)  ac_top_build_prefix=$ac_top_builddir_sub/ ;;
@@ -5205,8 +5832,8 @@ ac_sed_dataroot='
 case `eval "sed -n \"\$ac_sed_dataroot\" $ac_file_inputs"` in
 *datarootdir*) ac_datarootdir_seen=yes;;
 *@datadir@*|*@docdir@*|*@infodir@*|*@localedir@*|*@mandir@*)
-  { $as_echo "$as_me:${as_lineno-$LINENO}: WARNING: $ac_file_inputs seems to ignore the --datarootdir setting" >&5
-$as_echo "$as_me: WARNING: $ac_file_inputs seems to ignore the --datarootdir setting" >&2;}
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: $ac_file_inputs seems to ignore the --datarootdir setting" >&5
+printf "%s\n" "$as_me: WARNING: $ac_file_inputs seems to ignore the --datarootdir setting" >&2;}
 _ACEOF
 cat >>$CONFIG_STATUS <<_ACEOF || ac_write_fail=1
   ac_datarootdir_hack='
@@ -5248,9 +5875,9 @@ test -z "$ac_datarootdir_hack$ac_datarootdir_seen" &&
   { ac_out=`sed -n '/\${datarootdir}/p' "$ac_tmp/out"`; test -n "$ac_out"; } &&
   { ac_out=`sed -n '/^[	 ]*datarootdir[	 ]*:*=/p' \
       "$ac_tmp/out"`; test -z "$ac_out"; } &&
-  { $as_echo "$as_me:${as_lineno-$LINENO}: WARNING: $ac_file contains a reference to the variable \`datarootdir'
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: $ac_file contains a reference to the variable \`datarootdir'
 which seems to be undefined.  Please make sure it is defined" >&5
-$as_echo "$as_me: WARNING: $ac_file contains a reference to the variable \`datarootdir'
+printf "%s\n" "$as_me: WARNING: $ac_file contains a reference to the variable \`datarootdir'
 which seems to be undefined.  Please make sure it is defined" >&2;}
 
   rm -f "$ac_tmp/stdin"
@@ -5297,7 +5924,8 @@ if test "$no_create" != yes; then
   $ac_cs_success || as_fn_exit 1
 fi
 if test -n "$ac_unrecognized_opts" && test "$enable_option_checking" != no; then
-  { $as_echo "$as_me:${as_lineno-$LINENO}: WARNING: unrecognized options: $ac_unrecognized_opts" >&5
-$as_echo "$as_me: WARNING: unrecognized options: $ac_unrecognized_opts" >&2;}
+  { printf "%s\n" "$as_me:${as_lineno-$LINENO}: WARNING: unrecognized options: $ac_unrecognized_opts" >&5
+printf "%s\n" "$as_me: WARNING: unrecognized options: $ac_unrecognized_opts" >&2;}
 fi
 
+
diff --git a/lib/configure.ac b/lib/configure.ac
index 6b16fb798..00450891f 100644
--- a/lib/configure.ac
+++ b/lib/configure.ac
@@ -6,6 +6,7 @@ AC_SUBST(MKLDFLAGS, ["-lm "])
 AC_SUBST(CUDA_SRCS, ["cuda/cwc_convnet.cu cuda/cwc_convnet_ext.c cuda/convnet/cwc_convnet_convolutional.cu cuda/convnet/cwc_convnet_rnorm.cu cuda/convnet/cwc_convnet_pool.cu cuda/convnet/cwc_convnet_full_connect.cu"])
 AC_SUBST(CUDA_COMPAT_LIB, ["gpu/libnnc-compat-cuda.o"])
 AC_SUBST(CUDA_CMD_LIB, ["libnnc-cmd-cuda.o"])
+AC_SUBST(MFA_COMPAT_LIB, ["mfa/libnnc-compat-mfa.o"])
 AC_SUBST(MPS_COMPAT_LIB, ["mps/libnnc-compat-mps.o"])
 AC_SUBST(MPS_CMD_LIB, ["libnnc-cmd-mps.o"])
 
@@ -190,12 +191,13 @@ AC_ARG_ENABLE(mps, [AS_HELP_STRING([--enable-mps], [Enable Metal support])], [mp
 if test "$mps_support" = yes; then
 	AC_MSG_RESULT(yes)
 	AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D HAVE_MPS "])
-	AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-framework MetalPerformanceShaders -framework MetalPerformanceShadersGraph -framework Foundation -framework Metal "])
+	AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-framework MetalPerformanceShaders -framework MetalPerformanceShadersGraph -framework Foundation -framework Metal -lc++ "])
 	AC_SUBST(CUDA_SRCS, [""])
 	AC_SUBST(CUDA_COMPAT_LIB, [""])
 	AC_SUBST(CUDA_CMD_LIB, [""])
 else
 	AC_MSG_RESULT(no)
+	AC_SUBST(MFA_COMPAT_LIB, [""])
 	AC_SUBST(MPS_COMPAT_LIB, [""])
 	AC_SUBST(MPS_CMD_LIB, [""])
 fi
diff --git a/lib/nnc/cmd/blas/mps/ccv_nnc_add_mps.m b/lib/nnc/cmd/blas/mps/ccv_nnc_add_mps.m
index 38ae0a87b..63d16a1a6 100644
--- a/lib/nnc/cmd/blas/mps/ccv_nnc_add_mps.m
+++ b/lib/nnc/cmd/blas/mps/ccv_nnc_add_mps.m
@@ -23,7 +23,7 @@ static int _ccv_nnc_add_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 	if (inputs[1] == 0)
 	{
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			if (p == 1)
 			{
 				MPSGraph* graph = [MPSGraph new];
@@ -52,13 +52,13 @@ static int _ccv_nnc_add_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 				MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 				ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &c, (int*[]){ c->info.dim }, (int*[]){ c->stride }, 1);
 			}
-			ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+			ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 		}
 		return CCV_NNC_EXEC_SUCCESS;
 	}
 	const ccv_nnc_tensor_view_t* const b = (const ccv_nnc_tensor_view_t*)inputs[1];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[2];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -89,7 +89,7 @@ static int _ccv_nnc_add_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 		MPSGraphTensorData* data_b = ccv_nnc_mps_graph_tensor_data(b, b->info.dim, b->stride);
 		MPSGraphTensorData* data[] = {data_a, data_b};
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]]], &c, (int*[]){ c->info.dim }, (int*[]){ c->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m b/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m
index ef7465c3a..bb2caa4d7 100644
--- a/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m
+++ b/lib/nnc/cmd/blas/mps/ccv_nnc_gemm_mps.m
@@ -114,77 +114,229 @@ static int _ccv_nnc_gemm_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 	if (w_batch_size == 1 && b_batch_size > 1)
 		w_batch_inc = 0;
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
-		// If all these conditions are met, let's use MPS directly.
-		if ((!CCV_IS_TENSOR_VIEW(a) || ccv_nnc_tensor_view_is_contiguous(adim, astride)) &&
+		const int is_contiguous =
+			(!CCV_IS_TENSOR_VIEW(a) || ccv_nnc_tensor_view_is_contiguous(adim, astride)) &&
 			(!CCV_IS_TENSOR_VIEW(w) || ccv_nnc_tensor_view_is_contiguous(w->info.dim, w->stride)) &&
-			(!CCV_IS_TENSOR_VIEW(b) || ccv_nnc_tensor_view_is_contiguous(b->info.dim, b->stride)) &&
-			a->info.datatype == w->info.datatype && a->info.datatype == b->info.datatype &&
-			a_batch_size == w_batch_size && a_batch_size == b_batch_size && !(ccv_nnc_flags() & CCV_NNC_DISABLE_MIXED_MPS_GEMM) &&
-			!bias)
-		{
-			id<MTLBuffer> a_buffer = mpgetbuffer((ccv_nnc_tensor_t*)a);
-			MPSMatrix* leftMatrix = [[MPSMatrix alloc] initWithBuffer:a_buffer offset:a->dataof descriptor:[MPSMatrixDescriptor matrixDescriptorWithRows:(is_transpose_a ? a_cols : a_rows) columns:(is_transpose_a ? a_rows : a_cols) matrices:b_batch_size rowBytes:CCV_GET_DATA_TYPE_SIZE(a->info.datatype) * (is_transpose_a ? a_cols_inc : a_rows_inc) matrixBytes:CCV_GET_DATA_TYPE_SIZE(a->info.datatype) * a_batch_inc dataType:ccv_nnc_mps_datatype(a->info.datatype)]];
-			id<MTLBuffer> w_buffer = mpgetbuffer((ccv_nnc_tensor_t*)w);
-			MPSMatrix* rightMatrix = [[MPSMatrix alloc] initWithBuffer:w_buffer offset:w->dataof descriptor:[MPSMatrixDescriptor matrixDescriptorWithRows:(is_transpose_w ? w_cols : w_rows) columns:(is_transpose_w ? w_rows : w_cols) matrices:b_batch_size rowBytes:CCV_GET_DATA_TYPE_SIZE(w->info.datatype) * (is_transpose_w ? w_cols_inc : w_rows_inc) matrixBytes:CCV_GET_DATA_TYPE_SIZE(w->info.datatype) * w_batch_inc dataType:ccv_nnc_mps_datatype(w->info.datatype)]];
-			id<MTLBuffer> b_buffer = mpgetbuffer((ccv_nnc_tensor_t*)b);
-			MPSMatrix* resultMatrix = [[MPSMatrix alloc] initWithBuffer:b_buffer offset:b->dataof descriptor:[MPSMatrixDescriptor matrixDescriptorWithRows:b_rows columns:b_cols matrices:b_batch_size rowBytes:CCV_GET_DATA_TYPE_SIZE(b->info.datatype) * b_rows_inc matrixBytes:CCV_GET_DATA_TYPE_SIZE(b->info.datatype) * b_batch_inc dataType:ccv_nnc_mps_datatype(b->info.datatype)]];
-			MPSMatrixMultiplication* matrixMultiplication = [[MPSMatrixMultiplication alloc] initWithDevice:ccv_nnc_default_device() transposeLeft:(is_transpose_a ? YES : NO) transposeRight:(is_transpose_w ? YES : NO) resultRows:b_rows resultColumns:b_cols interiorColumns:a_cols alpha:1 beta:0];
-			[leftMatrix synchronizeOnCommandBuffer:command_buffer];
-			[rightMatrix synchronizeOnCommandBuffer:command_buffer];
-			[matrixMultiplication encodeToCommandBuffer:command_buffer leftMatrix:leftMatrix rightMatrix:rightMatrix resultMatrix:resultMatrix];
-			[resultMatrix synchronizeOnCommandBuffer:command_buffer];
-			[matrixMultiplication release];
-			[leftMatrix release];
-			[rightMatrix release];
-			[resultMatrix release];
-			// TODO: Try to use MPSMatrixFullyConnected for with bias case.
+			(!CCV_IS_TENSOR_VIEW(b) || ccv_nnc_tensor_view_is_contiguous(b->info.dim, b->stride));
+
+		const int is_same_dtype =
+			(a->info.datatype == w->info.datatype) &&
+			(a->info.datatype == b->info.datatype);
+		
+		int is_supported_dtype = 0;
+		uint32_t mtl_data_type = UINT32_MAX;
+		switch (a->info.datatype) {
+			case CCV_16F: {
+				is_supported_dtype = 1;
+				mtl_data_type = 16;
+				break;
+			}
+			case CCV_32F: {
+				is_supported_dtype = 1;
+				mtl_data_type = 3;
+				break;
+			}
+			default: {
+				break;
+			}
+		}
+
+		const int is_same_batch =
+			(a_batch_size == w_batch_size) &&
+			(a_batch_size == b_batch_size);
+		
+		// NNC uses the convention B = A * W.
+		// MFA uses the convention C = A * B.
+		int is_batched = 0;
+		int is_mfa_compatible_batch = 0;
+		int A_batch_size = a_batch_size;
+		int B_batch_size = w_batch_size;
+		int C_batch_size = b_batch_size;
+		if (A_batch_size == 1 && B_batch_size == 1 && C_batch_size == 1) {
+			// Not batched.
+		} else if (A_batch_size <= 0 || B_batch_size <= 0 || C_batch_size <= 0) {
+			// Invalid batch size.
 		} else {
-			// Otherwise, use MPSGraph
-			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
-			// Key will be consumed by the next method, therefore, no need to free.
-			int indices[3];
-			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
-				MPSGraphTensor* mps_input_a;
-				MPSGraphTensor* mps_a = ccv_nnc_mps_graph_tensor_input(graph, a, adim_r, astride_r, &mps_input_a);
-				MPSGraphTensor* mps_input_w;
-				MPSGraphTensor* mps_w = ccv_nnc_mps_graph_tensor_input(graph, w, w->info.dim, w->stride, &mps_input_w);
-				MPSGraphShapedType* mps_a_shape = ccv_nnc_mps_graph_tensor_input_shape(a, adim_r, astride_r);
-				MPSGraphShapedType* mps_w_shape = ccv_nnc_mps_graph_tensor_input_shape(w, w->info.dim, w->stride);
-				if (is_transpose_a)
-					mps_a = [graph transposeTensor:mps_a dimension:-2 withDimension:-1 name:nil];
-				if (is_transpose_w)
-					mps_w = [graph transposeTensor:mps_w dimension:-2 withDimension:-1 name:nil];
-				MPSGraphTensor* mps_b = [graph matrixMultiplicationWithPrimaryTensor:mps_a secondaryTensor:mps_w name:nil];
-				[inputTensors addObject:mps_input_a];
-				[inputShapedTypes addObject:mps_a_shape];
-				[inputTensors addObject:mps_input_w];
-				[inputShapedTypes addObject:mps_w_shape];
-				if (bias)
+			is_batched = 1;
+			if (A_batch_size == C_batch_size) {
+				if (A_batch_size == B_batch_size) {
+					is_mfa_compatible_batch = 1;
+				} else if (B_batch_size == 1) {
+					is_mfa_compatible_batch = 1;
+				}
+			}
+		}
+
+		ccv_nnc_mfa_context_t* context = ccv_nnc_default_mfa_context();
+		const int is_mfa_supported =
+			ccv_nnc_mfa_context_supported(context) && is_contiguous && is_same_dtype && is_supported_dtype && (is_mfa_compatible_batch || !is_batched) && !bias;
+
+		if (METAL_LOG_LEVEL(context) >= 3)
+		{
+			if (is_mfa_supported)
+			{
+				ccv_nnc_mfa_log_message("Compatible GEMM found.");
+			} else {
+				ccv_nnc_mfa_log_message("Incompatible GEMM found. Incompatible because:");
+				if (!is_contiguous)
+				{
+					ccv_nnc_mfa_log_message("  Strided.");
+				}
+				if (!is_same_dtype)
+				{
+					ccv_nnc_mfa_log_message("  Mixed precision.");
+				}
+				if (!is_same_dtype)
 				{
-					MPSGraphTensor* mps_input_bias;
-					MPSGraphTensor* mps_bias = ccv_nnc_mps_graph_tensor_input(graph, bias, biasdim_r, biasstride_r, &mps_input_bias);
-					MPSGraphShapedType* mps_bias_shape = ccv_nnc_mps_graph_tensor_input_shape(bias, biasdim_r, biasstride_r);
-					// Add support broadcast directly.
-					mps_b = [graph additionWithPrimaryTensor:mps_b secondaryTensor:mps_bias name:nil];
-					[inputTensors addObject:mps_input_bias];
-					[inputShapedTypes addObject:mps_bias_shape];
+					ccv_nnc_mfa_log_message("  Unsupported data type.");
 				}
-				[resultTensors addObject:mps_b];
-			});
-			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, adim, astride);
-			MPSGraphTensorData* data_w = ccv_nnc_mps_graph_tensor_data(w, w->info.dim, w->stride);
-			if (bias)
+				if (!(is_mfa_compatible_batch || !is_batched))
+				{
+					ccv_nnc_mfa_log_message("  Unsupported batch.");
+				}
+				if (!(!bias))
+				{
+					ccv_nnc_mfa_log_message("  Requires fused activations.");
+				}
+			}
+		}
+
+		if (is_mfa_supported)
+		{
+			// On supported devices, use Metal directly.
+			ccv_nnc_mfa_gemm_params_t params = {
+				.data_type = mtl_data_type,
+				.M = (uint32_t)b_rows, // C_rows
+				.N = (uint32_t)b_cols, // C_cols
+				.K = (uint32_t)w_rows, // B_rows
+				.A_trans = (is_transpose_a ? 1 : 0),
+				.B_trans = (is_transpose_w ? 1 : 0),
+				.alpha = (float)1.0,
+				.beta = (float)0.0,
+				.batched = is_batched,
+				.fused_activation = 0,
+				
+				.batch_dims_a = { 0 },
+				.batch_dims_b = { 0 },
+			};
+			if (is_batched) {
+				// Create a null-terminated list of batch dimensions.
+				int A_batch_dim = a_nd - 2;
+				for (int i = 0; i < A_batch_dim; ++i) {
+					params.batch_dims_a[i] = adim[i];
+				}
+				if (A_batch_dim < CCV_NNC_MAX_DIM_ALLOC) {
+					params.batch_dims_a[A_batch_dim] = 0;
+				}
+				
+				int B_batch_dim = w_nd - 2;
+				for (int i = 0; i < B_batch_dim; ++i) {
+					params.batch_dims_b[i] = w->info.dim[i];
+				}
+				if (B_batch_dim < CCV_NNC_MAX_DIM_ALLOC) {
+					params.batch_dims_b[B_batch_dim] = 0;
+				}
+			}
+			ccv_nnc_mfa_sync_prepare_gemm(context, params);
+
+			// Creating a new command buffer has a >10 µs penalty CPU-side. Still
+			// faster the >50 µs penalty for MPSGraph (probably why
+			// MPSMatrixMultiplication is faster for GEMM).
+			mtl_command_batch_t* command_batch = ccv_nnc_stream_context_start_command_batch(stream_context);
+			mtl_buffer_t* tensors[4] = {
+				mpgetbuffer((ccv_nnc_tensor_t*)a), // A
+				mpgetbuffer((ccv_nnc_tensor_t*)w), // B
+				mpgetbuffer((ccv_nnc_tensor_t*)b), // C
+				NULL
+			};
+			size_t tensor_offsets[3] = {
+				a->dataof, // A offset
+				w->dataof, // B offset
+				b->dataof, // C offset
+			};
+			ccv_nnc_mfa_encode_gemm(context, params, command_batch, tensors, tensor_offsets);
+
+			// TODO: Add this diagnostic once we consistently capture >>1 commands/batch.
+//			if (METAL_LOG_LEVEL(context) >= 3) {
+//				if (command_batch->batched_command_count == 0) {
+//					ccv_nnc_mfa_log_message("Encoded 0 commands in the batch.");
+//				} else if (command_batch->batched_command_count == 1) {
+//					ccv_nnc_mfa_log_message("Encoded 1 command in the batch.");
+//				} else {
+//					ccv_nnc_mfa_log_message("Encoded >1 commands in the batch.");
+//				}
+//			}
+			ccv_nnc_stream_context_finish_command_batch(stream_context, command_batch);
+			// TODO: Try to use `fused_activation` for with bias case.
+		} else {
+			// Otherwise, incur the ~10-50 microsecond latency of MPS.
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
+			
+			// If all conditions are met, use MPSMatrixMultiplication.
+			if (is_contiguous && is_same_dtype && is_same_batch && !(ccv_nnc_flags() & CCV_NNC_DISABLE_MIXED_MPS_GEMM) && !bias)
 			{
-				MPSGraphTensorData* data_bias = ccv_nnc_mps_graph_tensor_data(bias, biasdim, biasstride);
-				MPSGraphTensorData* data[] = {data_a, data_w, data_bias};
-				ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]], data[indices[2]]], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
+				id<MTLBuffer> a_buffer = mpgetbuffer((ccv_nnc_tensor_t*)a);
+				MPSMatrix* leftMatrix = [[MPSMatrix alloc] initWithBuffer:a_buffer offset:a->dataof descriptor:[MPSMatrixDescriptor matrixDescriptorWithRows:(is_transpose_a ? a_cols : a_rows) columns:(is_transpose_a ? a_rows : a_cols) matrices:b_batch_size rowBytes:CCV_GET_DATA_TYPE_SIZE(a->info.datatype) * (is_transpose_a ? a_cols_inc : a_rows_inc) matrixBytes:CCV_GET_DATA_TYPE_SIZE(a->info.datatype) * a_batch_inc dataType:ccv_nnc_mps_datatype(a->info.datatype)]];
+				id<MTLBuffer> w_buffer = mpgetbuffer((ccv_nnc_tensor_t*)w);
+				MPSMatrix* rightMatrix = [[MPSMatrix alloc] initWithBuffer:w_buffer offset:w->dataof descriptor:[MPSMatrixDescriptor matrixDescriptorWithRows:(is_transpose_w ? w_cols : w_rows) columns:(is_transpose_w ? w_rows : w_cols) matrices:b_batch_size rowBytes:CCV_GET_DATA_TYPE_SIZE(w->info.datatype) * (is_transpose_w ? w_cols_inc : w_rows_inc) matrixBytes:CCV_GET_DATA_TYPE_SIZE(w->info.datatype) * w_batch_inc dataType:ccv_nnc_mps_datatype(w->info.datatype)]];
+				id<MTLBuffer> b_buffer = mpgetbuffer((ccv_nnc_tensor_t*)b);
+				MPSMatrix* resultMatrix = [[MPSMatrix alloc] initWithBuffer:b_buffer offset:b->dataof descriptor:[MPSMatrixDescriptor matrixDescriptorWithRows:b_rows columns:b_cols matrices:b_batch_size rowBytes:CCV_GET_DATA_TYPE_SIZE(b->info.datatype) * b_rows_inc matrixBytes:CCV_GET_DATA_TYPE_SIZE(b->info.datatype) * b_batch_inc dataType:ccv_nnc_mps_datatype(b->info.datatype)]];
+				MPSMatrixMultiplication* matrixMultiplication = [[MPSMatrixMultiplication alloc] initWithDevice:ccv_nnc_default_device() transposeLeft:(is_transpose_a ? YES : NO) transposeRight:(is_transpose_w ? YES : NO) resultRows:b_rows resultColumns:b_cols interiorColumns:a_cols alpha:1 beta:0];
+				[leftMatrix synchronizeOnCommandBuffer:command_buffer];
+				[rightMatrix synchronizeOnCommandBuffer:command_buffer];
+				[matrixMultiplication encodeToCommandBuffer:command_buffer leftMatrix:leftMatrix rightMatrix:rightMatrix resultMatrix:resultMatrix];
+				[resultMatrix synchronizeOnCommandBuffer:command_buffer];
+				[matrixMultiplication release];
+				[leftMatrix release];
+				[rightMatrix release];
+				[resultMatrix release];
 			} else {
-				MPSGraphTensorData* data[] = {data_a, data_w};
-				ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]]], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
+				// Otherwise, use MPSGraph.
+				ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
+				// Key will be consumed by the next method, therefore, no need to free.
+				int indices[3];
+				MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
+					MPSGraphTensor* mps_input_a;
+					MPSGraphTensor* mps_a = ccv_nnc_mps_graph_tensor_input(graph, a, adim_r, astride_r, &mps_input_a);
+					MPSGraphTensor* mps_input_w;
+					MPSGraphTensor* mps_w = ccv_nnc_mps_graph_tensor_input(graph, w, w->info.dim, w->stride, &mps_input_w);
+					MPSGraphShapedType* mps_a_shape = ccv_nnc_mps_graph_tensor_input_shape(a, adim_r, astride_r);
+					MPSGraphShapedType* mps_w_shape = ccv_nnc_mps_graph_tensor_input_shape(w, w->info.dim, w->stride);
+					if (is_transpose_a)
+						mps_a = [graph transposeTensor:mps_a dimension:-2 withDimension:-1 name:nil];
+					if (is_transpose_w)
+						mps_w = [graph transposeTensor:mps_w dimension:-2 withDimension:-1 name:nil];
+					MPSGraphTensor* mps_b = [graph matrixMultiplicationWithPrimaryTensor:mps_a secondaryTensor:mps_w name:nil];
+					[inputTensors addObject:mps_input_a];
+					[inputShapedTypes addObject:mps_a_shape];
+					[inputTensors addObject:mps_input_w];
+					[inputShapedTypes addObject:mps_w_shape];
+					if (bias)
+					{
+						MPSGraphTensor* mps_input_bias;
+						MPSGraphTensor* mps_bias = ccv_nnc_mps_graph_tensor_input(graph, bias, biasdim_r, biasstride_r, &mps_input_bias);
+						MPSGraphShapedType* mps_bias_shape = ccv_nnc_mps_graph_tensor_input_shape(bias, biasdim_r, biasstride_r);
+						// Add support broadcast directly.
+						mps_b = [graph additionWithPrimaryTensor:mps_b secondaryTensor:mps_bias name:nil];
+						[inputTensors addObject:mps_input_bias];
+						[inputShapedTypes addObject:mps_bias_shape];
+					}
+					[resultTensors addObject:mps_b];
+				});
+				MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, adim, astride);
+				MPSGraphTensorData* data_w = ccv_nnc_mps_graph_tensor_data(w, w->info.dim, w->stride);
+				if (bias)
+				{
+					MPSGraphTensorData* data_bias = ccv_nnc_mps_graph_tensor_data(bias, biasdim, biasstride);
+					MPSGraphTensorData* data[] = {data_a, data_w, data_bias};
+					ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]], data[indices[2]]], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
+				} else {
+					MPSGraphTensorData* data[] = {data_a, data_w};
+					ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]]], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
+				}
 			}
+			ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/blas/mps/ccv_nnc_mul_mps.m b/lib/nnc/cmd/blas/mps/ccv_nnc_mul_mps.m
index 681bc2612..ed068ef97 100644
--- a/lib/nnc/cmd/blas/mps/ccv_nnc_mul_mps.m
+++ b/lib/nnc/cmd/blas/mps/ccv_nnc_mul_mps.m
@@ -22,7 +22,7 @@ static int _ccv_nnc_mul_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 	if (inputs[1] == 0)
 	{
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			if (p == 1)
 			{
 				MPSGraph* graph = [MPSGraph new];
@@ -50,13 +50,13 @@ static int _ccv_nnc_mul_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 				MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 				ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &c, (int*[]){ c->info.dim }, (int*[]){ c->stride }, 1);
 			}
-			ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+			ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 		}
 		return CCV_NNC_EXEC_SUCCESS;
 	}
 	const ccv_nnc_tensor_view_t* const b = (const ccv_nnc_tensor_view_t*)inputs[1];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[2];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -82,7 +82,7 @@ static int _ccv_nnc_mul_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 		MPSGraphTensorData* data_b = ccv_nnc_mps_graph_tensor_data(b, b->info.dim, b->stride);
 		MPSGraphTensorData* data[] = {data_a, data_b};
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]]], &c, (int*[]){ c->info.dim }, (int*[]){ c->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
@@ -102,7 +102,7 @@ static int _ccv_nnc_scalar_mul_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const c = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		if (p == 1)
 		{
 			MPSGraph* graph = [MPSGraph new];
@@ -131,7 +131,7 @@ static int _ccv_nnc_scalar_mul_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &c, (int*[]){ c->info.dim }, (int*[]){ c->stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/convolution/mps/ccv_nnc_conv_mps.m b/lib/nnc/cmd/convolution/mps/ccv_nnc_conv_mps.m
index 6730b8191..0c1b56b4a 100644
--- a/lib/nnc/cmd/convolution/mps/ccv_nnc_conv_mps.m
+++ b/lib/nnc/cmd/convolution/mps/ccv_nnc_conv_mps.m
@@ -51,7 +51,7 @@ static int _ccv_nnc_conv_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 		}
 	}
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int* adim_r = adim;
 		int* astride_r = astride;
@@ -96,7 +96,7 @@ static int _ccv_nnc_conv_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 			MPSGraphTensorData* data[] = {data_a, data_w};
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]]], &b, (int*[]){ bdim }, (int*[]){ bstride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/ew/mps/ccv_nnc_ew_mps.m b/lib/nnc/cmd/ew/mps/ccv_nnc_ew_mps.m
index d5d5d886a..d2aded1d1 100644
--- a/lib/nnc/cmd/ew/mps/ccv_nnc_ew_mps.m
+++ b/lib/nnc/cmd/ew/mps/ccv_nnc_ew_mps.m
@@ -19,7 +19,7 @@ static int _ccv_nnc_ewsum_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 		}
 	ccv_nnc_tensor_view_t* const c = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int* indices = (int*)ccv_nnc_stream_context_get_workspace(stream_context, (sizeof(int) + sizeof(MPSGraphTensorData*)) * input_size, CCV_TENSOR_CPU_MEMORY);
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -58,7 +58,7 @@ static int _ccv_nnc_ewsum_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 			[feeds addObject:data[indices[z]]];
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, feeds, &c, (int*[]){ c->info.dim }, (int*[]){ c->stride }, 1);
 		[feeds release];
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
@@ -88,7 +88,7 @@ static int _ccv_nnc_ewdiv_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 		for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC && a->info.dim[i] > 0; i++)
 			{ assert(a->info.dim[i] == b->info.dim[i]); }
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[2];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -114,7 +114,7 @@ static int _ccv_nnc_ewdiv_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 		}
 	} else {
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -154,7 +154,7 @@ static int _ccv_nnc_ewexp_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 	for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC && a->info.dim[i] > 0; i++)
 		{ assert(a->info.dim[i] == c->info.dim[i]); }
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -193,7 +193,7 @@ static int _ccv_nnc_ewlog_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 	for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC && a->info.dim[i] > 0; i++)
 		{ assert(a->info.dim[i] == c->info.dim[i]); }
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -232,7 +232,7 @@ static int _ccv_nnc_ewsqrt_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hi
 	for (i = 0; i < CCV_NNC_MAX_DIM_ALLOC && a->info.dim[i] > 0; i++)
 		{ assert(a->info.dim[i] == c->info.dim[i]); }
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -276,7 +276,7 @@ static int _ccv_nnc_clamp_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 	if (isnan(minv))
 	{
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -296,7 +296,7 @@ static int _ccv_nnc_clamp_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 		}
 	} else if (isnan(maxv)) {
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -316,7 +316,7 @@ static int _ccv_nnc_clamp_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 		}
 	} else {
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
diff --git a/lib/nnc/cmd/gelu/mps/ccv_nnc_gelu_mps.m b/lib/nnc/cmd/gelu/mps/ccv_nnc_gelu_mps.m
index 955c933f7..73a22376b 100644
--- a/lib/nnc/cmd/gelu/mps/ccv_nnc_gelu_mps.m
+++ b/lib/nnc/cmd/gelu/mps/ccv_nnc_gelu_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_gelu_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -48,7 +48,7 @@ static int _ccv_nnc_gelu_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 		});
 		MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/index/mps/ccv_nnc_index_select_mps.m b/lib/nnc/cmd/index/mps/ccv_nnc_index_select_mps.m
index d0e9ce7e8..6793ab983 100644
--- a/lib/nnc/cmd/index/mps/ccv_nnc_index_select_mps.m
+++ b/lib/nnc/cmd/index/mps/ccv_nnc_index_select_mps.m
@@ -23,7 +23,7 @@ static int _ccv_nnc_index_select_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hin
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	assert(ccv_nnc_tensor_nd(b->info.dim) <= 2);
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int idx[2];
 		int indices_dim[CCV_NNC_MAX_DIM_ALLOC] = {0};
@@ -102,7 +102,7 @@ static int _ccv_nnc_index_select_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hin
 		MPSGraphTensorData* data_indices = ccv_nnc_mps_graph_tensor_data(indices, indices_dim, indices_stride);
 		MPSGraphTensorData* data[] = {data_a, data_indices};
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[idx[0]], data[idx[1]]], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/leaky_relu/mps/ccv_nnc_leaky_relu_mps.m b/lib/nnc/cmd/leaky_relu/mps/ccv_nnc_leaky_relu_mps.m
index cf00f4bdf..4bc621ef2 100644
--- a/lib/nnc/cmd/leaky_relu/mps/ccv_nnc_leaky_relu_mps.m
+++ b/lib/nnc/cmd/leaky_relu/mps/ccv_nnc_leaky_relu_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_leaky_relu_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		const double alpha = (double)cmd.info.leaky_relu.negative_slope;
@@ -27,7 +27,7 @@ static int _ccv_nnc_leaky_relu_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 		});
 		MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/norm/mps/ccv_nnc_group_norm_mps.m b/lib/nnc/cmd/norm/mps/ccv_nnc_group_norm_mps.m
index c27daa44d..0ad2b5f72 100644
--- a/lib/nnc/cmd/norm/mps/ccv_nnc_group_norm_mps.m
+++ b/lib/nnc/cmd/norm/mps/ccv_nnc_group_norm_mps.m
@@ -54,7 +54,7 @@ static int _ccv_nnc_group_norm_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 		&bt
 	}, 4);
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[3];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -140,7 +140,7 @@ static int _ccv_nnc_group_norm_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]], data[indices[2]]], (ccv_nnc_tensor_view_t* []){ (ccv_nnc_tensor_view_t*)outputs[0], &saved_meant, &saved_inv_stdt }, (int*[]){ outputs[0]->info.dim, saved_meant.info.dim, saved_inv_stdt.info.dim }, (int*[]){ ((ccv_nnc_tensor_view_t*)outputs[0])->stride, saved_meant.stride, saved_inv_stdt.stride }, 3);
 		else
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]], data[indices[2]]], (ccv_nnc_tensor_view_t* []){ &bt, &saved_meant, &saved_inv_stdt }, (int*[]){ bt.info.dim, saved_meant.info.dim, saved_inv_stdt.info.dim }, (int*[]){ bt.stride, saved_meant.stride, saved_inv_stdt.stride }, 3);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/norm/mps/ccv_nnc_layer_norm_mps.m b/lib/nnc/cmd/norm/mps/ccv_nnc_layer_norm_mps.m
index 489784b71..4967b49c8 100644
--- a/lib/nnc/cmd/norm/mps/ccv_nnc_layer_norm_mps.m
+++ b/lib/nnc/cmd/norm/mps/ccv_nnc_layer_norm_mps.m
@@ -22,7 +22,7 @@ static int _ccv_nnc_layer_norm_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 		&bt
 	}, 4);
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[3];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -79,7 +79,7 @@ static int _ccv_nnc_layer_norm_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 		MPSGraphTensorData* data_bias = ccv_nnc_mps_graph_tensor_data(&biast, biast.info.dim, biast.stride);
 		MPSGraphTensorData* data[] = {data_a, data_scale, data_bias};
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]], data[indices[2]]], (ccv_nnc_tensor_view_t* []){ &bt, &saved_meant, &saved_inv_stdt }, (int*[]){ bt.info.dim, saved_meant.info.dim, saved_inv_stdt.info.dim }, (int*[]){ bt.stride, saved_meant.stride, saved_inv_stdt.stride }, 3);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/pool/mps/ccv_nnc_avg_pool_mps.m b/lib/nnc/cmd/pool/mps/ccv_nnc_avg_pool_mps.m
index 4e8d69703..e3e655569 100644
--- a/lib/nnc/cmd/pool/mps/ccv_nnc_avg_pool_mps.m
+++ b/lib/nnc/cmd/pool/mps/ccv_nnc_avg_pool_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_avg_pool_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -27,7 +27,7 @@ static int _ccv_nnc_avg_pool_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 		});
 		MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/pool/mps/ccv_nnc_max_pool_mps.m b/lib/nnc/cmd/pool/mps/ccv_nnc_max_pool_mps.m
index 8f2f6beea..39eaf50f3 100644
--- a/lib/nnc/cmd/pool/mps/ccv_nnc_max_pool_mps.m
+++ b/lib/nnc/cmd/pool/mps/ccv_nnc_max_pool_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_max_pool_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -27,7 +27,7 @@ static int _ccv_nnc_max_pool_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 		});
 		MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/rand/mps/ccv_nnc_rand_normal_mps.m b/lib/nnc/cmd/rand/mps/ccv_nnc_rand_normal_mps.m
index cc6a6e46f..0e91a5e0d 100644
--- a/lib/nnc/cmd/rand/mps/ccv_nnc_rand_normal_mps.m
+++ b/lib/nnc/cmd/rand/mps/ccv_nnc_rand_normal_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_random_normal(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 	const float std = cmd.info.blas.a[0];
 	const float mean = cmd.info.blas.a[1];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		for (i = 0; i < output_size; i++)
 		{
 			ccv_nnc_tensor_view_t* const a = (ccv_nnc_tensor_view_t*)outputs[i];
@@ -46,7 +46,7 @@ static int _ccv_nnc_random_normal(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 			ccv_nnc_mps_graph_result(graph, command_buffer, @{}, mps_r[0], a, a->info.dim, a->stride);
 			[graph release];
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/rand/mps/ccv_nnc_rand_uniform_mps.m b/lib/nnc/cmd/rand/mps/ccv_nnc_rand_uniform_mps.m
index cc57d6f0f..a56eb744c 100644
--- a/lib/nnc/cmd/rand/mps/ccv_nnc_rand_uniform_mps.m
+++ b/lib/nnc/cmd/rand/mps/ccv_nnc_rand_uniform_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_random_uniform(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 	const float l = cmd.info.blas.a[0];
 	const float u = cmd.info.blas.a[1];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		for (i = 0; i < output_size; i++)
 		{
 			ccv_nnc_tensor_view_t* const a = (ccv_nnc_tensor_view_t*)outputs[i];
@@ -30,7 +30,7 @@ static int _ccv_nnc_random_uniform(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 			ccv_nnc_mps_graph_result(graph, command_buffer, @{}, mps_a, a, a->info.dim, a->stride);
 			[graph release];
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/reduce/mps/ccv_nnc_argmax_mps.m b/lib/nnc/cmd/reduce/mps/ccv_nnc_argmax_mps.m
index e9a38c36b..fcda2de1c 100644
--- a/lib/nnc/cmd/reduce/mps/ccv_nnc_argmax_mps.m
+++ b/lib/nnc/cmd/reduce/mps/ccv_nnc_argmax_mps.m
@@ -23,7 +23,7 @@ static int _ccv_nnc_argmax_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hi
 		noop = btv.info.dim[i] != atv.info.dim[i];
 	const int axis = cmd.info.reduce.axis[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		if (noop)
 		{
 			MPSGraph* graph = [MPSGraph new];
@@ -51,7 +51,7 @@ static int _ccv_nnc_argmax_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hi
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(&atv, atv.info.dim, atv.stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &tvs[1], (int*[]){ btv.info.dim }, (int*[]){ btv.stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/reduce/mps/ccv_nnc_argmin_mps.m b/lib/nnc/cmd/reduce/mps/ccv_nnc_argmin_mps.m
index 3188476ea..2737ec138 100644
--- a/lib/nnc/cmd/reduce/mps/ccv_nnc_argmin_mps.m
+++ b/lib/nnc/cmd/reduce/mps/ccv_nnc_argmin_mps.m
@@ -23,7 +23,7 @@ static int _ccv_nnc_argmin_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hi
 		noop = btv.info.dim[i] != atv.info.dim[i];
 	const int axis = cmd.info.reduce.axis[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		if (noop)
 		{
 			MPSGraph* graph = [MPSGraph new];
@@ -51,7 +51,7 @@ static int _ccv_nnc_argmin_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hi
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(&atv, atv.info.dim, atv.stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &tvs[1], (int*[]){ btv.info.dim }, (int*[]){ btv.stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_max_mps.m b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_max_mps.m
index a5266340d..7b317805e 100644
--- a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_max_mps.m
+++ b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_max_mps.m
@@ -21,7 +21,7 @@ static int _ccv_nnc_reduce_max_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 	for (i = 0; noop && i < a_nd; i++)
 		noop = btv.info.dim[i] != atv.info.dim[i];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		if (noop)
 		{
 			MPSGraph* graph = [MPSGraph new];
@@ -55,7 +55,7 @@ static int _ccv_nnc_reduce_max_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(&atv, atv.info.dim, atv.stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &tvs[1], (int*[]){ btv.info.dim }, (int*[]){ btv.stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_mean_mps.m b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_mean_mps.m
index 7ede989ba..6049e1739 100644
--- a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_mean_mps.m
+++ b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_mean_mps.m
@@ -21,7 +21,7 @@ static int _ccv_nnc_reduce_mean_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint
 	for (i = 0; noop && i < a_nd; i++)
 		noop = btv.info.dim[i] != atv.info.dim[i];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		if (noop)
 		{
 			MPSGraph* graph = [MPSGraph new];
@@ -55,7 +55,7 @@ static int _ccv_nnc_reduce_mean_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(&atv, atv.info.dim, atv.stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &tvs[1], (int*[]){ btv.info.dim }, (int*[]){ btv.stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_min_mps.m b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_min_mps.m
index 38d6fac6c..d36ec6a24 100644
--- a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_min_mps.m
+++ b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_min_mps.m
@@ -21,7 +21,7 @@ static int _ccv_nnc_reduce_min_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 	for (i = 0; noop && i < a_nd; i++)
 		noop = btv.info.dim[i] != atv.info.dim[i];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		if (noop)
 		{
 			MPSGraph* graph = [MPSGraph new];
@@ -55,7 +55,7 @@ static int _ccv_nnc_reduce_min_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(&atv, atv.info.dim, atv.stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &tvs[1], (int*[]){ btv.info.dim }, (int*[]){ btv.stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_sum_mps.m b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_sum_mps.m
index 077d7114e..ab85cb984 100644
--- a/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_sum_mps.m
+++ b/lib/nnc/cmd/reduce/mps/ccv_nnc_reduce_sum_mps.m
@@ -21,7 +21,7 @@ static int _ccv_nnc_reduce_sum_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 	for (i = 0; noop && i < a_nd; i++)
 		noop = btv.info.dim[i] != atv.info.dim[i];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		if (noop)
 		{
 			MPSGraph* graph = [MPSGraph new];
@@ -55,7 +55,7 @@ static int _ccv_nnc_reduce_sum_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(&atv, atv.info.dim, atv.stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &tvs[1], (int*[]){ btv.info.dim }, (int*[]){ btv.stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/relu/mps/ccv_nnc_relu_mps.m b/lib/nnc/cmd/relu/mps/ccv_nnc_relu_mps.m
index 56009ebe2..54ba2d557 100644
--- a/lib/nnc/cmd/relu/mps/ccv_nnc_relu_mps.m
+++ b/lib/nnc/cmd/relu/mps/ccv_nnc_relu_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_relu_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -26,7 +26,7 @@ static int _ccv_nnc_relu_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 		});
 		MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/sigmoid/mps/ccv_nnc_sigmoid_mps.m b/lib/nnc/cmd/sigmoid/mps/ccv_nnc_sigmoid_mps.m
index 58c689677..1bd6487ad 100644
--- a/lib/nnc/cmd/sigmoid/mps/ccv_nnc_sigmoid_mps.m
+++ b/lib/nnc/cmd/sigmoid/mps/ccv_nnc_sigmoid_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_sigmoid_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t h
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -26,7 +26,7 @@ static int _ccv_nnc_sigmoid_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t h
 		});
 		MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
@@ -39,7 +39,7 @@ static int _ccv_nnc_sigmoid_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t h
 	const ccv_nnc_tensor_view_t* const b = (const ccv_nnc_tensor_view_t*)inputs[2];
 	ccv_nnc_tensor_view_t* const h = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -65,7 +65,7 @@ static int _ccv_nnc_sigmoid_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t h
 		MPSGraphTensorData* data_b = ccv_nnc_mps_graph_tensor_data(b, b->info.dim, b->stride);
 		MPSGraphTensorData* data[] = {data_g, data_b};
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data[indices[0]], data[indices[1]]], &h, (int*[]){ h->info.dim }, (int*[]){ h->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/softmax/mps/ccv_nnc_softmax_mps.m b/lib/nnc/cmd/softmax/mps/ccv_nnc_softmax_mps.m
index 3249aec98..44595cb84 100644
--- a/lib/nnc/cmd/softmax/mps/ccv_nnc_softmax_mps.m
+++ b/lib/nnc/cmd/softmax/mps/ccv_nnc_softmax_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_softmax_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t h
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		const int a_nd = ccv_nnc_tensor_nd(a->info.dim);
 		const int b_nd = ccv_nnc_tensor_nd(b->info.dim);
 		if (a_nd <= 2 && b_nd <= 2 && !(ccv_nnc_flags() & CCV_NNC_DISABLE_MIXED_MPS_SOFTMAX)) // Simple case, we use MPS directly.
@@ -73,7 +73,7 @@ static int _ccv_nnc_softmax_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t h
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/swish/mps/ccv_nnc_swish_mps.m b/lib/nnc/cmd/swish/mps/ccv_nnc_swish_mps.m
index d9c5146a3..8834368fb 100644
--- a/lib/nnc/cmd/swish/mps/ccv_nnc_swish_mps.m
+++ b/lib/nnc/cmd/swish/mps/ccv_nnc_swish_mps.m
@@ -12,7 +12,7 @@ static int _ccv_nnc_swish_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 	const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[0];
 	ccv_nnc_tensor_view_t* const b = (ccv_nnc_tensor_view_t*)outputs[0];
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 		int indices[1];
 		MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -30,7 +30,7 @@ static int _ccv_nnc_swish_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hin
 		});
 		MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 		ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/cmd/upsample/mps/ccv_nnc_upsample_mps.m b/lib/nnc/cmd/upsample/mps/ccv_nnc_upsample_mps.m
index b9d5de984..b6f3c6985 100644
--- a/lib/nnc/cmd/upsample/mps/ccv_nnc_upsample_mps.m
+++ b/lib/nnc/cmd/upsample/mps/ccv_nnc_upsample_mps.m
@@ -39,7 +39,7 @@ static int _ccv_nnc_upsample_nearest_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc
 	if (a->info.format == CCV_TENSOR_FORMAT_NCHW)
 	{
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -53,12 +53,12 @@ static int _ccv_nnc_upsample_nearest_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc
 			});
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, adim, astride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ bdim }, (int*[]){ bstride }, 1);
-			ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+			ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 		}
 	} else {
 		assert(a->info.format == CCV_TENSOR_FORMAT_NHWC);
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -72,7 +72,7 @@ static int _ccv_nnc_upsample_nearest_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc
 			});
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, adim, astride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ bdim }, (int*[]){ bstride }, 1);
-			ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+			ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 		}
 	}
 	return CCV_NNC_EXEC_SUCCESS;
@@ -104,7 +104,7 @@ static int _ccv_nnc_upsample_bilinear_forw(const ccv_nnc_cmd_t cmd, const ccv_nn
 	if (a->info.format == CCV_TENSOR_FORMAT_NCHW)
 	{
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -118,12 +118,12 @@ static int _ccv_nnc_upsample_bilinear_forw(const ccv_nnc_cmd_t cmd, const ccv_nn
 			});
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, adim, astride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ bdim }, (int*[]){ bstride }, 1);
-			ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+			ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 		}
 	} else {
 		assert(a->info.format == CCV_TENSOR_FORMAT_NHWC);
 		@autoreleasepool {
-			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+			MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 			ccv_nnc_mps_graph_key_t key = ccv_nnc_mps_graph_key_new(cmd, hint, flags, inputs, input_size, outputs, output_size);
 			int indices[1];
 			MPSGraphExecutable* executable = ccv_nnc_mps_graph_executable_cache(key, indices, ^void (MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors) {
@@ -137,7 +137,7 @@ static int _ccv_nnc_upsample_bilinear_forw(const ccv_nnc_cmd_t cmd, const ccv_nn
 			});
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, adim, astride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ bdim }, (int*[]){ bstride }, 1);
-			ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+			ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 		}
 	}
 	return CCV_NNC_EXEC_SUCCESS;
diff --git a/lib/nnc/cmd/util/mps/ccv_nnc_util_mps.m b/lib/nnc/cmd/util/mps/ccv_nnc_util_mps.m
index 0bde6ed1c..dc89a0210 100644
--- a/lib/nnc/cmd/util/mps/ccv_nnc_util_mps.m
+++ b/lib/nnc/cmd/util/mps/ccv_nnc_util_mps.m
@@ -29,11 +29,11 @@ static int _ccv_nnc_data_transfer(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 			const off_t offset_b = mpgetoffset(b);
 			@autoreleasepool {
 				id<MTLBuffer> buffer_a = [ccv_nnc_default_device() newBufferWithBytesNoCopy:aligned_ptr length:aligned_size options:MTLResourceCPUCacheModeDefaultCache | MTLResourceStorageModeShared deallocator:nil];
-				id<MTLCommandBuffer> command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+				id<MTLCommandBuffer> command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 				id<MTLBlitCommandEncoder> encoder = [command_buffer blitCommandEncoder];
 				[encoder copyFromBuffer:buffer_a sourceOffset:offset_a toBuffer:buffer_b destinationOffset:offset_b size:size];
 				[encoder endEncoding];
-				ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+				ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 				[buffer_a release];
 			}
 		} else if (CCV_TENSOR_GET_MEMORY(a->info.type) == CCV_TENSOR_GPU_MEMORY && CCV_TENSOR_GET_MEMORY(b->info.type) == CCV_TENSOR_CPU_MEMORY) {
@@ -44,11 +44,11 @@ static int _ccv_nnc_data_transfer(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 			const size_t aligned_size = ((size + offset_b + vm_page_size - 1) & -vm_page_size);
 			@autoreleasepool {
 				id<MTLBuffer> buffer_b = [ccv_nnc_default_device() newBufferWithBytesNoCopy:aligned_ptr length:aligned_size options:MTLResourceCPUCacheModeDefaultCache | MTLResourceStorageModeShared deallocator:nil];
-				id<MTLCommandBuffer> command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+				id<MTLCommandBuffer> command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 				id<MTLBlitCommandEncoder> encoder = [command_buffer blitCommandEncoder];
 				[encoder copyFromBuffer:buffer_a sourceOffset:offset_a toBuffer:buffer_b destinationOffset:offset_b size:size];
 				[encoder endEncoding];
-				ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+				ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 				[buffer_b release];
 			}
 		} else if (CCV_TENSOR_GET_MEMORY(a->info.type) == CCV_TENSOR_CPU_MEMORY && CCV_TENSOR_GET_MEMORY(b->info.type) == CCV_TENSOR_CPU_MEMORY)
@@ -62,11 +62,11 @@ static int _ccv_nnc_data_transfer(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t
 			const off_t offset_a = mpgetoffset(a);
 			const off_t offset_b = mpgetoffset(b);
 			@autoreleasepool {
-				id<MTLCommandBuffer> command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+				id<MTLCommandBuffer> command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 				id<MTLBlitCommandEncoder> encoder = [command_buffer blitCommandEncoder];
 				[encoder copyFromBuffer:buffer_a sourceOffset:offset_a toBuffer:buffer_b destinationOffset:offset_b size:size];
 				[encoder endEncoding];
-				ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+				ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 			}
 		}
 	}
@@ -96,7 +96,7 @@ static int _ccv_nnc_transpose(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 	assert(output_size <= input_size);
 	int i;
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		for (i = 0; i < output_size; i++)
 		{
 			const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[i];
@@ -115,7 +115,7 @@ static int _ccv_nnc_transpose(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint
 			MPSGraphTensorData* data_a = ccv_nnc_mps_graph_tensor_data(a, a->info.dim, a->stride);
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[data_a], &b, (int*[]){ b->info.dim }, (int*[]){ b->stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
@@ -142,7 +142,7 @@ static int _ccv_nnc_set_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 {
 	int i, j;
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		for (i = 0; i < output_size; i++)
 		{
 			ccv_nnc_tensor_view_t* const a = (ccv_nnc_tensor_view_t*)outputs[i];
@@ -158,7 +158,7 @@ static int _ccv_nnc_set_forw(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 			[shape release];
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[], &a, (int*[]){ a->info.dim }, (int*[]){ a->stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
@@ -167,7 +167,7 @@ static int _ccv_nnc_set_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 {
 	int i, j;
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		for (i = 0; i < output_size; i++)
 		{
 			ccv_nnc_tensor_view_t* const a = (ccv_nnc_tensor_view_t*)outputs[i];
@@ -183,7 +183,7 @@ static int _ccv_nnc_set_back(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint,
 			[shape release];
 			ccv_nnc_mps_graph_executable_result(executable, command_buffer, @[], &a, (int*[]){ a->info.dim }, (int*[]){ a->stride }, 1);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
@@ -211,7 +211,7 @@ static int _ccv_nnc_format_transform(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint
 	assert(output_size <= input_size);
 	int i;
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		for (i = 0; i < output_size; i++)
 		{
 			const ccv_nnc_tensor_view_t* const a = (const ccv_nnc_tensor_view_t*)inputs[i];
@@ -311,7 +311,7 @@ static int _ccv_nnc_format_transform(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint
 				ccv_nnc_mps_export_data(data_a, command_buffer, &bt, bdim, bstride);
 			[graph release];
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
@@ -339,7 +339,7 @@ static int _ccv_nnc_datatype_conversion(const ccv_nnc_cmd_t cmd, const ccv_nnc_h
 	assert(output_size <= input_size);
 	int i;
 	@autoreleasepool {
-		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_get_command_buffer(stream_context);
+		MPSCommandBuffer* command_buffer = ccv_nnc_stream_context_start_mps_command_buffer(stream_context);
 		for (i = 0; i < output_size; i++)
 		{
 			const ccv_nnc_tensor_view_t* a = (ccv_nnc_tensor_view_t*)inputs[i];
@@ -362,7 +362,7 @@ static int _ccv_nnc_datatype_conversion(const ccv_nnc_cmd_t cmd, const ccv_nnc_h
 			} else
 				ccv_nnc_mps_export_data(data_a, command_buffer, b, b->info.dim, b->stride);
 		}
-		ccv_nnc_stream_context_commit_command_buffer(stream_context, command_buffer);
+		ccv_nnc_stream_context_finish_mps_command_buffer(stream_context, command_buffer);
 	}
 	return CCV_NNC_EXEC_SUCCESS;
 }
diff --git a/lib/nnc/makefile b/lib/nnc/makefile
index d886c32f3..b7a7a94ce 100644
--- a/lib/nnc/makefile
+++ b/lib/nnc/makefile
@@ -7,7 +7,7 @@ SRCS := ccv_nnc_cmd.c ccv_nnc_tensor.c ccv_nnc_tensor_io.c ccv_nnc_stream.c ccv_
 
 SRC_OBJS := $(patsubst %.c,%.o,$(SRCS))
 
-.PHONY: release all lib clean cmd libnnc.o cmd/libnnc-cmd.o gpu/libnnc-compat-cuda.o mps/libnnc-compat-mps.o
+.PHONY: release all lib clean cmd libnnc.o cmd/libnnc-cmd.o gpu/libnnc-compat-cuda.o mfa/libnnc-compat-mfa.o mps/libnnc-compat-mps.o
 
 release: all
 
@@ -18,9 +18,9 @@ all: lib
 lib: libnnc.o
 
 clean:
-	rm -f *.o gpu/*.o && ${MAKE} -C ./gpu clean && ${MAKE} -C ./mps clean && ${MAKE} -C ./cmd clean
+	rm -f *.o gpu/*.o && ${MAKE} -C ./gpu clean && ${MAKE} -C ./mfa clean && ${MAKE} -C ./mps clean && ${MAKE} -C ./cmd clean
 
-libnnc.o: $(SRC_OBJS) cmd/libnnc-cmd.o $(CUDA_COMPAT_LIB) $(MPS_COMPAT_LIB)
+libnnc.o: $(SRC_OBJS) cmd/libnnc-cmd.o $(CUDA_COMPAT_LIB) $(MFA_COMPAT_LIB) $(MPS_COMPAT_LIB)
 	ld -r $^ -o $@
 
 %.o: %.c
@@ -38,6 +38,9 @@ cmd:
 gpu/libnnc-compat-cuda.o:
 	${MAKE} -C ./gpu
 
+mfa/libnnc-compat-mfa.o:
+	${MAKE} -C ./mfa
+
 mps/libnnc-compat-mps.o:
 	${MAKE} -C ./mps
 
diff --git a/lib/nnc/mfa/.ycm_extra_conf.py b/lib/nnc/mfa/.ycm_extra_conf.py
new file mode 100644
index 000000000..4829ad51e
--- /dev/null
+++ b/lib/nnc/mfa/.ycm_extra_conf.py
@@ -0,0 +1,54 @@
+import os
+import ycm_core
+from clang_helpers import PrepareClangFlags
+
+flags = [
+  '-ffast-math',
+  '-Wall',
+  '-msse2',
+  '-D HAVE_SSE2',
+  '-D HAVE_CBLAS',
+  '-D HAVE_MPS',
+  '-I',
+  '../..'
+]
+
+def DirectoryOfThisScript():
+  return os.path.dirname(os.path.abspath(__file__))
+
+def MakeRelativePathsInFlagsAbsolute(flags, working_directory):
+  if not working_directory:
+    return flags
+  new_flags = []
+  make_next_absolute = False
+  path_flags = ['-isystem', '-I', '-iquote', '--sysroot=']
+  for flag in flags:
+    new_flag = flag
+
+    if make_next_absolute:
+      make_next_absolute = False
+      if not flag.startswith('/'):
+        new_flag = os.path.join(working_directory, flag)
+
+    for path_flag in path_flags:
+      if flag == path_flag:
+        make_next_absolute = True
+        break
+
+      if flag.startswith(path_flag):
+        path = flag[len(path_flag):]
+        new_flag = path_flag + os.path.join(working_directory, path)
+        break
+
+    if new_flag:
+      new_flags.append(new_flag)
+  return new_flags
+
+def FlagsForFile(filename):
+  relative_to = DirectoryOfThisScript()
+  final_flags = MakeRelativePathsInFlagsAbsolute(flags, relative_to)
+  return {
+    'flags' : final_flags,
+    'do_cache' : True
+  }
+
diff --git a/lib/nnc/mfa/3rdparty/libmfaios16-0.2.metallib b/lib/nnc/mfa/3rdparty/libmfaios16-0.2.metallib
new file mode 100644
index 000000000..948c2df7a
Binary files /dev/null and b/lib/nnc/mfa/3rdparty/libmfaios16-0.2.metallib differ
diff --git a/lib/nnc/mfa/3rdparty/libmfamacos13-0.2.metallib b/lib/nnc/mfa/3rdparty/libmfamacos13-0.2.metallib
new file mode 100644
index 000000000..5bcba51cc
Binary files /dev/null and b/lib/nnc/mfa/3rdparty/libmfamacos13-0.2.metallib differ
diff --git a/lib/3rdparty/metal-cpp/Dispatch.cpp b/lib/nnc/mfa/3rdparty/metal-cpp/Dispatch.cpp
similarity index 100%
rename from lib/3rdparty/metal-cpp/Dispatch.cpp
rename to lib/nnc/mfa/3rdparty/metal-cpp/Dispatch.cpp
diff --git a/lib/3rdparty/metal-cpp/Dispatch.hpp b/lib/nnc/mfa/3rdparty/metal-cpp/Dispatch.hpp
similarity index 100%
rename from lib/3rdparty/metal-cpp/Dispatch.hpp
rename to lib/nnc/mfa/3rdparty/metal-cpp/Dispatch.hpp
diff --git a/lib/3rdparty/metal-cpp/Metal.hpp b/lib/nnc/mfa/3rdparty/metal-cpp/Metal.hpp
similarity index 100%
rename from lib/3rdparty/metal-cpp/Metal.hpp
rename to lib/nnc/mfa/3rdparty/metal-cpp/Metal.hpp
diff --git a/lib/nnc/mfa/Metal.cpp b/lib/nnc/mfa/Metal.cpp
new file mode 100644
index 000000000..11942a3e3
--- /dev/null
+++ b/lib/nnc/mfa/Metal.cpp
@@ -0,0 +1,5 @@
+#define NS_PRIVATE_IMPLEMENTATION
+#define CA_PRIVATE_IMPLEMENTATION
+#define MTL_PRIVATE_IMPLEMENTATION
+
+#include "nnc/mfa/3rdparty/metal-cpp/Metal.hpp"
diff --git a/lib/nnc/mfa/ccv_nnc_mfa.cpp b/lib/nnc/mfa/ccv_nnc_mfa.cpp
new file mode 100644
index 000000000..38de74ce7
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa.cpp
@@ -0,0 +1,170 @@
+#include "ccv_nnc_mfa.hpp"
+using namespace ccv::nnc;
+
+#include <iostream>
+
+// MARK: - C
+
+mfa::context* ccv_nnc_init_mfa_context(MTL::Device* device, const char* metallib_path) {
+  return new mfa::context(device, metallib_path);
+}
+
+void ccv_nnc_deinit_mfa_context(mfa::context* context) {
+  delete context;
+}
+
+uint8_t ccv_nnc_mfa_context_supported(mfa::context* context) {
+  return context->supported ? 1 : 0;
+}
+
+uint16_t ccv_nnc_mfa_context_log_level(mfa::context* context) {
+  return context->log_level;
+}
+
+void ccv_nnc_mfa_log_message(const char* message) {
+  std::cerr << METAL_LOG_HEADER << message << std::endl;
+}
+
+MTL::CommandBatch* ccv_nnc_start_command_batch(MTL::CommandQueue* command_queue) {
+  return new MTL::CommandBatch(command_queue);
+}
+
+void ccv_nnc_finish_command_batch(MTL::CommandBatch* command_batch) {
+  delete command_batch;
+}
+
+// MARK: - C++
+
+template <typename T, typename U>
+mfa::cache<T, U>::cache()
+{
+  map = {};
+}
+
+template <typename T, typename U>
+mfa::cache<T, U>::~cache()
+{
+  for (auto it = map.begin(); it != map.end(); ++it) {
+    delete it->second;
+  }
+}
+
+// This is a workaround. If we use a template member function directly, the
+// symbols won't link.
+template <typename T, typename U>
+inline void _mfa_cache_prepare(std::unordered_map<T, U*>* map, mfa::context* context, T hash, bool async)
+{
+  if (map->find(hash) == map->end()) {
+    if (METAL_LOG_LEVEL(context) >= 2) {
+      std::cout << METAL_LOG_HEADER << "PSO cache miss." << std::endl;
+      std::cout << METAL_LOG_HEADER << "  Creating new PSO asynchronously: " << async << std::endl;
+      std::cout << METAL_LOG_HEADER << "  Contents of map (before):" << std::endl;
+      for (auto it = map->begin(); it != map->end(); ++it) {
+        std::cout << METAL_LOG_HEADER << "    " << it->first << ": " << it->second << std::endl;
+      }
+    }
+    
+    auto* pipeline = new mfa::gemm::pipeline(context, hash, async);
+    (*map)[hash] = pipeline;
+    
+    if (METAL_LOG_LEVEL(context) >= 2) {
+      std::cout << METAL_LOG_HEADER << "  Contents of map (after):" << std::endl;
+      for (auto it = map->begin(); it != map->end(); ++it) {
+        std::cout << METAL_LOG_HEADER << "    " << it->first << ": " << it->second << std::endl;
+      }
+    }
+  }
+}
+
+template <>
+void mfa::cache<mfa::gemm::hash, mfa::gemm::pipeline>::prepare(mfa::context* context, mfa::gemm::hash hash, bool async)
+{
+  _mfa_cache_prepare(&map, context, hash, async);
+}
+
+mfa::context::context(MTL::Device* device, const char* metallib_path)
+{
+  auto* pool = NS::AutoreleasePool::alloc()->init();
+  
+  this->log_level = 0;
+#if CCV_METAL_LOGGING_ENABLE
+  const char* log_level_repr = getenv("CCV_METAL_LOG_LEVEL");
+  if (log_level_repr) {
+    int log_level_raw = atoi(log_level_repr);
+    std::cerr << METAL_LOG_HEADER << "Using log level: " << log_level_raw << std::endl;
+    CCV_NNC_MFA_PRECONDITION(log_level_raw >= 0 && log_level_raw <= 3)
+    
+    this->log_level = uint16_t(log_level_raw);
+  }
+#endif
+  
+  // Example: /usr/local/MetalFlashAttention/lib/libMetalFlashAttention.metallib
+  // We need to have two different variants based on the operating system. macOS
+  // will not accept a metallib compiled for iOS/tvOS/visionOS and vice versa.
+  if (!metallib_path) {
+    this->supported = false;
+    return;
+  }
+  if (METAL_LOG_LEVEL(this) >= 1) {
+    std::cerr << METAL_LOG_HEADER << "Started loading 'libMetalFlashAttention.metallib'." << std::endl;
+  }
+  
+  // Check whether the device architecture is supported.
+  this->supported = device->supportsFamily(MTL::GPUFamilyApple7);
+  if (!supported) {
+    if (METAL_LOG_LEVEL(this) >= 1) {
+      std::cerr << METAL_LOG_HEADER << "Device architecture not supported by Metal FlashAttention." << std::endl;
+    }
+    return;
+  }
+  
+  this->device = NS::RetainPtr(device);
+#if TARGET_OS_OSX
+  // This method is only available on macOS 13.3+. To make the code compatible
+  // with macOS 12, we need to call ObjC runtime functions that check whether
+  // the selector actually exists.
+  device->setShouldMaximizeConcurrentCompilation(true);
+#endif
+  
+  // Create a URL out of the path string.
+  auto c_path = metallib_path;
+  auto swift_path = NS::String::string(c_path, NS::UTF8StringEncoding);
+  auto url = NS::URL::fileURLWithPath(swift_path);
+  
+  // Attempt to load the library, otherwise crash with a detailed log message.
+  NS::Error* error;
+  this->library = NS::TransferPtr(device->newLibrary(url, &error));
+  CCV_NNC_MFA_CHECK_ERROR(error)
+  
+  // Notify that this finished successfully, and is not just stalling on one of
+  // the previous lines of code.
+  if (METAL_LOG_LEVEL(this) >= 1) {
+    std::cerr << METAL_LOG_HEADER << "Finished loading 'libMetalFlashAttention.metallib'." << std::endl;
+  }
+  
+  pool->drain();
+}
+
+MTL::CommandBatch::CommandBatch(MTL::CommandQueue* command_queue) {
+  command_buffer = command_queue->commandBuffer();
+  command_encoder = command_buffer->computeCommandEncoder();
+}
+
+MTL::ComputeCommandEncoder* MTL::CommandBatch::start_command(MTL::ComputePipelineState* pso) {
+  CCV_NNC_MFA_PRECONDITION(command_active == 0)
+  command_active = 1;
+  command_encoder->setComputePipelineState(pso);
+  return command_encoder;
+}
+
+void MTL::CommandBatch::finish_command(MTL::ComputeCommandEncoder* command_encoder) {
+  CCV_NNC_MFA_PRECONDITION(command_active == 1)
+  command_active = 0;
+  batched_command_count += 1;
+}
+
+MTL::CommandBatch::~CommandBatch() {
+  CCV_NNC_MFA_PRECONDITION(command_active == 0)
+  command_encoder->endEncoding();
+  command_buffer->commit();
+}
diff --git a/lib/nnc/mfa/ccv_nnc_mfa.hpp b/lib/nnc/mfa/ccv_nnc_mfa.hpp
new file mode 100644
index 000000000..423101b52
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa.hpp
@@ -0,0 +1,73 @@
+#ifndef GUARD_ccv_nnc_mfa_hpp
+#define GUARD_ccv_nnc_mfa_hpp
+
+#include "nnc/ccv_nnc.h"
+#include "ccv_nnc_mfa_defines.hpp"
+#include "ccv_nnc_mfa_gemm.hpp"
+
+#ifdef __cplusplus
+#include "nnc/mfa/3rdparty/metal-cpp/Dispatch.hpp"
+#include "nnc/mfa/3rdparty/metal-cpp/Metal.hpp"
+#include "ccv_nnc_mfa_error.hpp"
+#include <unordered_map>
+
+namespace ccv {
+namespace nnc {
+namespace mfa {
+
+class context;
+
+template <typename T, typename U>
+class cache {
+public:
+  std::unordered_map<T, U*> map;
+  
+  cache();
+  ~cache();
+  
+  void prepare(context* context, T hash, bool async);
+};
+
+class context {
+public:
+  bool supported;
+  uint16_t log_level;
+  
+  NS::SharedPtr<MTL::Device> device;
+  NS::SharedPtr<MTL::Library> library;
+  
+  context(MTL::Device* device, const char* metallib_path);
+  
+  // MFA keeps internal caches of pipeline state objects. If you're eagerly
+  // executing a command, call `sync_prepare_*` just before encoding it. This
+  // incurs non-negligible latency, which can be removed by compiling during
+  // graph compilation. Use `async_prepare_*` during graph compilation, which
+  // will transform the subsequent `sync_prepare_*` into a NOP. The async
+  // version has more latency but utilizes multicore CPU parallelism.
+  //
+  // After preparing the pipeline, call `encode_*`. Pass each tensor's backing
+  // `MTL::Buffer*` through a null-terminated list.
+  cache<gemm::hash, gemm::pipeline> gemm_cache;
+};
+
+} // namespace mfa
+} // namespace nnc
+} // namespace ccv
+
+extern "C" {
+#endif // __cplusplus
+
+ccv_nnc_mfa_context_t* ccv_nnc_init_mfa_context(mtl_device_t* context, const char* metallib_path);
+void ccv_nnc_deinit_mfa_context(ccv_nnc_mfa_context_t* context);
+uint8_t ccv_nnc_mfa_context_supported(ccv_nnc_mfa_context_t* context);
+uint16_t ccv_nnc_mfa_context_log_level(ccv_nnc_mfa_context_t* context);
+void ccv_nnc_mfa_log_message(const char* message);
+
+mtl_command_batch_t* ccv_nnc_start_command_batch(mtl_command_queue_t* command_queue);
+void ccv_nnc_finish_command_batch(mtl_command_batch_t* command_batch);
+
+#ifdef __cplusplus
+} // extern "C"
+#endif // __cplusplus
+
+#endif
diff --git a/lib/nnc/mfa/ccv_nnc_mfa_defines.hpp b/lib/nnc/mfa/ccv_nnc_mfa_defines.hpp
new file mode 100644
index 000000000..09c54a784
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa_defines.hpp
@@ -0,0 +1,91 @@
+#ifndef GUARD_ccv_nnc_mfa_defines_hpp
+#define GUARD_ccv_nnc_mfa_defines_hpp
+
+// MARK: - Types
+
+#ifdef __cplusplus
+#include "nnc/mfa/3rdparty/metal-cpp/Metal.hpp"
+namespace ccv {
+namespace nnc {
+namespace mfa {
+class context;
+} // namespace mfa
+} // namespace nnc
+} // namespace ccv
+
+typedef ccv::nnc::mfa::context ccv_nnc_mfa_context_t;
+typedef MTL::Buffer mtl_buffer_t;
+typedef MTL::CommandBuffer mtl_command_buffer_t;
+typedef MTL::ComputeCommandEncoder mtl_compute_command_encoder_t;
+typedef MTL::CommandQueue mtl_command_queue_t;
+typedef MTL::Device mtl_device_t;
+#else
+typedef void ccv_nnc_mfa_context_t;
+typedef void mtl_buffer_t;
+typedef void mtl_command_buffer_t;
+typedef void mtl_compute_command_encoder_t;
+typedef void mtl_command_queue_t;
+typedef void mtl_device_t;
+#endif // __cplusplus
+
+#ifdef __cplusplus
+namespace MTL {
+class CommandBatch {
+public:
+  MTL::CommandBuffer* command_buffer;
+  
+  // Although labeled `MTL::ComputeCommandEncoder`, this should be used for
+  // memcpy and memset as well. Here is a performant reference implementation
+  // using custom shaders to bypass the CPU-side latency of switching encoders:
+  // https://github.com/philipturner/metal-usm/tree/main/BlitEncoderAlternative
+  MTL::ComputeCommandEncoder* command_encoder;
+  
+  uint16_t batched_command_count = 0;
+  uint8_t command_active = 0;
+  
+  CommandBatch(MTL::CommandQueue* command_queue);
+  ~CommandBatch();
+  
+  MTL::ComputeCommandEncoder* start_command(MTL::ComputePipelineState* pso);
+  void finish_command(MTL::ComputeCommandEncoder* command_encoder);
+};
+} // namespace MTL
+
+typedef MTL::CommandBatch mtl_command_batch_t;
+#else // __cplusplus
+typedef struct {
+  mtl_command_buffer_t* command_buffer;
+  mtl_compute_command_encoder_t* command_encoder;
+  uint16_t batched_command_count;
+  uint8_t command_active;
+} MTLCommandBatch;
+
+typedef MTLCommandBatch mtl_command_batch_t;
+#endif // __cplusplus
+
+// MARK: - Diagnostics
+
+#ifndef CCV_METAL_LOGGING_ENABLE
+#define CCV_METAL_LOGGING_ENABLE 1
+#endif
+
+// 0 - crash reports
+// 1 - metallib initialization
+// 2 - PSO creation
+// 3 - command encoding
+
+#if CCV_METAL_LOGGING_ENABLE
+
+#ifdef __cplusplus
+#define METAL_LOG_LEVEL(CONTEXT) CONTEXT->log_level
+#else
+#define METAL_LOG_LEVEL(CONTEXT) ccv_nnc_mfa_context_log_level(CONTEXT)
+#endif // __cplusplus
+
+#else // CCV_NNC_METAL_LOGGING_ENABLE
+
+#define METAL_LOG_LEVEL(CONTEXT) 0
+
+#endif // CCV_NNC_METAL_LOGGING_ENABLE
+
+#endif
diff --git a/lib/nnc/mfa/ccv_nnc_mfa_error.cpp b/lib/nnc/mfa/ccv_nnc_mfa_error.cpp
new file mode 100644
index 000000000..00071191f
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa_error.cpp
@@ -0,0 +1,46 @@
+#include "ccv_nnc_mfa.hpp"
+using namespace ccv::nnc;
+
+#include <iostream>
+
+inline void log_source_location(int line, const char *file_name, const char *function_name) {
+  std::cerr << METAL_LOG_HEADER << "Encountered unexpected error in: " << function_name << std::endl;
+  std::cerr << "\e[0;1m" << file_name << ":" << line << ":\e[0m ";
+  std::cerr << "\e[0;31m" << "error:" << "\e[0m ";
+}
+
+void mfa::fatal_error(NS::Error* error, int line, const char *file_name, const char *function_name) {
+  auto description = error->localizedDescription();
+  auto recovery_suggestion = error->localizedRecoverySuggestion();
+  auto failure_reason = error->localizedFailureReason();
+  
+  log_source_location(line, file_name, function_name);
+  std::cerr << "\e[0;1m";
+  if (description) {
+    std::cerr << description->cString(NS::UTF8StringEncoding);
+  } else {
+    std::cerr << "[description not available]";
+  }
+  std::cerr << "\e[0m" << std::endl;
+  if (recovery_suggestion) {
+    std::cerr << METAL_LOG_HEADER << "Recovery suggestion: " << recovery_suggestion->cString(NS::UTF8StringEncoding) << std::endl;
+  }
+  if (failure_reason) {
+    std::cerr << METAL_LOG_HEADER << "Failure reason: " << failure_reason->cString(NS::UTF8StringEncoding) << std::endl;
+  }
+  std::cerr << METAL_LOG_HEADER << "Quitting now." << std::endl;
+  exit(-1);
+}
+
+void mfa::precondition_failure(const char *message, int line, const char *file_name, const char *function_name) {
+  log_source_location(line, file_name, function_name);
+  std::cerr << "\e[0;1m";
+  if (message) {
+    std::cerr << message;
+  } else {
+    std::cerr << "[precondition failure]";
+  }
+  std::cerr << "\e[0m" << std::endl;
+  std::cerr << METAL_LOG_HEADER << "Quitting now." << std::endl;
+  exit(-1);
+}
diff --git a/lib/nnc/mfa/ccv_nnc_mfa_error.hpp b/lib/nnc/mfa/ccv_nnc_mfa_error.hpp
new file mode 100644
index 000000000..3fc0238e9
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa_error.hpp
@@ -0,0 +1,30 @@
+#ifndef GUARD_ccv_nnc_mfa_error_hpp
+#define GUARD_ccv_nnc_mfa_error_hpp
+
+#include "nnc/mfa/3rdparty/metal-cpp/Metal.hpp"
+
+// `std::cout` and `CACurrentMediaTime()` for profiling.
+#include <iostream>
+#include <QuartzCore/QuartzCore.h>
+
+namespace ccv {
+namespace nnc {
+namespace mfa {
+
+#define METAL_LOG_HEADER "\e[0;36m[Metal]\e[0m "
+
+#define CCV_NNC_MFA_CHECK_ERROR(error) \
+if (error) { ccv::nnc::mfa::fatal_error(error, __LINE__, __FILE__, __FUNCTION__); } \
+
+void fatal_error(NS::Error* error, int line, const char *file_name, const char *function_name);
+
+#define CCV_NNC_MFA_PRECONDITION(expr) \
+if (!(expr)) { ccv::nnc::mfa::precondition_failure(nullptr, __LINE__, __FILE__, __FUNCTION__); } \
+
+void precondition_failure(const char *message, int line, const char *file_name, const char *function_name);
+
+} // namespace mfa
+} // namespace nnc
+} // namespace ccv
+
+#endif
diff --git a/lib/nnc/mfa/ccv_nnc_mfa_gemm.cpp b/lib/nnc/mfa/ccv_nnc_mfa_gemm.cpp
new file mode 100644
index 000000000..5467bcf18
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa_gemm.cpp
@@ -0,0 +1,384 @@
+#include "ccv_nnc_mfa.hpp"
+#include "ccv_nnc_mfa_hash.hpp"
+#include <simd/simd.h>
+using namespace ccv::nnc;
+
+#include <string>
+
+// MARK: - C
+
+void ccv_nnc_mfa_async_prepare_gemm(mfa::context* context, ccv_nnc_mfa_gemm_params_t params)
+{
+  context->gemm_cache.prepare(context, mfa::gemm::hash(params), true);
+}
+
+void ccv_nnc_mfa_sync_prepare_gemm(mfa::context* context, ccv_nnc_mfa_gemm_params_t params)
+{
+  context->gemm_cache.prepare(context, mfa::gemm::hash(params), false);
+}
+
+void ccv_nnc_mfa_encode_gemm(mfa::context* context, ccv_nnc_mfa_gemm_params_t params, MTL::CommandBatch* command_batch, MTL::Buffer** tensors, size_t* tensor_offsets)
+{
+  mfa::gemm::hash hash(params);
+  auto iterator = context->gemm_cache.map.find(hash);
+  if (iterator == context->gemm_cache.map.end()) {
+    mfa::precondition_failure("GEMM hash not cached.", __LINE__, __FILE__, __FUNCTION__);
+  }
+  
+  auto* pipeline = iterator->second;
+  pipeline->wait();
+  
+  auto* encoder = command_batch->start_command(pipeline->get_pso());
+  encoder->setThreadgroupMemoryLength(pipeline->get_threadgroup_memory_length(), 0);
+  
+  int num_tensors = 0;
+  while (tensors[num_tensors] != nullptr) {
+    num_tensors += 1;
+  }
+  CCV_NNC_MFA_PRECONDITION(num_tensors == 3)
+  for (int i = 0; i < num_tensors; ++i) {
+    if (i < 2) {
+      encoder->useResource(tensors[i], MTL::ResourceUsageRead);
+    } else if (i < 3) {
+      encoder->useResource(tensors[i], MTL::ResourceUsageWrite);
+    } else {
+      // This should never happen.
+      CCV_NNC_MFA_PRECONDITION(false);
+    }
+    encoder->setBuffer(tensors[i], tensor_offsets[i], i);
+  }
+  
+  uint32_t batch_size;
+  if (pipeline->get_batched()) {
+    uint16_t num_batch_dims_a = 0;
+    uint64_t batch_size_a = 1;
+    for (int i = 0; i < CCV_NNC_MAX_DIM_ALLOC; ++i) {
+      if (params.batch_dims_a[i] == 0) {
+        break;
+      }
+      num_batch_dims_a += 1;
+      batch_size_a *= params.batch_dims_a[i];
+    }
+    
+    uint16_t num_batch_dims_b = 0;
+    uint64_t batch_size_b = 1;
+    for (int i = 0; i < CCV_NNC_MAX_DIM_ALLOC; ++i) {
+      if (params.batch_dims_b[i] == 0) {
+        break;
+      }
+      num_batch_dims_b += 1;
+      batch_size_b *= params.batch_dims_b[i];
+    }
+    
+    bool same_batch_dims = true;
+    if (num_batch_dims_a != num_batch_dims_b) {
+      same_batch_dims = false;
+    } else if (batch_size_a != batch_size_b) {
+      same_batch_dims = false;
+    } else {
+      for (int i = 0; i < CCV_NNC_MAX_DIM_ALLOC; ++i) {
+        if (params.batch_dims_a[i] != params.batch_dims_b[i]) {
+          same_batch_dims = false;
+        }
+      }
+    }
+    
+    if (!same_batch_dims) {
+      CCV_NNC_MFA_PRECONDITION(batch_size_b == 1);
+    }
+    batch_size = batch_size_a;
+    
+    uint16_t element_size = 0;
+    switch (params.data_type) {
+      case MTL::DataTypeHalf: {
+        element_size = 2;
+        break;
+      }
+      case MTL::DataTypeFloat: {
+        element_size = 4;
+        break;
+      }
+      default:
+        CCV_NNC_MFA_PRECONDITION(false);
+        break;
+    }
+    uint64_t byte_stride_a = hash.M * hash.K * element_size;
+    uint64_t byte_stride_b = hash.K * hash.N * element_size;
+    uint64_t byte_stride_c = hash.M * hash.N * element_size;
+    if (batch_size_b == 1) {
+      byte_stride_b = 0;
+    }
+    
+    simd::ulong4 matrix_offsets[batch_size];
+    for (int i = 0; i < batch_size; ++i) {
+      matrix_offsets[i] = simd::ulong4 {
+        i * byte_stride_a,
+        i * byte_stride_b,
+        i * byte_stride_c,
+        0
+      };
+    }
+    encoder->setBytes(matrix_offsets, batch_size * 32, 10);
+  } else {
+    batch_size = 1;
+  }
+  
+  auto grid_size = pipeline->get_grid_size();
+  grid_size.depth = batch_size;
+  encoder->dispatchThreadgroups(grid_size, pipeline->get_group_size());
+  command_batch->finish_command(encoder);
+}
+
+// MARK: - C++
+
+mfa::gemm::hash::hash(ccv_nnc_mfa_gemm_params_t params) {
+  data_type = params.data_type;
+  M = params.M;
+  N = params.N;
+  K = params.K;
+  A_trans = params.A_trans;
+  B_trans = params.B_trans;
+  alpha = params.alpha;
+  beta = params.beta;
+  batched = params.batched;
+  fused_activation = params.fused_activation;
+}
+
+bool mfa::gemm::hash::operator==(const mfa::gemm::hash& hash) const {
+  return
+  (data_type == hash.data_type) &&
+  (M == hash.M) &&
+  (N == hash.N) &&
+  (K == hash.K) &&
+  (A_trans == hash.A_trans) &&
+  (B_trans == hash.B_trans) &&
+  (alpha == hash.alpha) &&
+  (beta == hash.beta) &&
+  (batched == hash.batched) &&
+  (fused_activation == hash.fused_activation);
+}
+
+mfa::gemm::pipeline::pipeline(mfa::context* context, mfa::gemm::hash hash, bool async) {
+  CCV_NNC_MFA_PRECONDITION((hash.data_type == MTL::DataTypeFloat) || (hash.data_type == MTL::DataTypeHalf))
+  CCV_NNC_MFA_PRECONDITION(hash.alpha == 1.0)
+  CCV_NNC_MFA_PRECONDITION(hash.beta == 0.0)
+  CCV_NNC_MFA_PRECONDITION(hash.fused_activation == false)
+  
+  auto* pool = NS::AutoreleasePool::alloc()->init();
+  
+  if (async) {
+    finished = false;
+    semaphore = new Dispatch::Semaphore(0);
+  } else {
+    finished = true;
+    semaphore = nullptr;
+  }
+  this->batched = hash.batched;
+  
+  auto constants = NS::TransferPtr(MTL::FunctionConstantValues::alloc()->init());
+  constants->setConstantValue(&hash.M, MTL::DataTypeUInt, NS::UInteger(0));
+  constants->setConstantValue(&hash.N, MTL::DataTypeUInt, 1);
+  constants->setConstantValue(&hash.K, MTL::DataTypeUInt, 2);
+  constants->setConstantValue(&hash.A_trans, MTL::DataTypeBool, 10);
+  constants->setConstantValue(&hash.B_trans, MTL::DataTypeBool, 11);
+  constants->setConstantValue(&hash.alpha, MTL::DataTypeFloat, 20);
+  constants->setConstantValue(&hash.beta, MTL::DataTypeFloat, 21);
+  constants->setConstantValue(&hash.batched, MTL::DataTypeBool, 100);
+  constants->setConstantValue(&hash.fused_activation, MTL::DataTypeBool, 101);
+  
+  // Eventually, this will incorporate the batch size.
+  // BxMxN > 1,000,000 -> 48x48, only if M >= 88 and N >= 88
+  // BxMxN > 4,000,000 -> 64x64, only if M >= 120 and N >= 120
+  uint64_t C_elements = uint64_t(hash.M) * uint64_t(hash.N);
+  if (batched) {
+    C_elements *= 2;
+  }
+  int is_half = (hash.data_type == MTL::DataTypeHalf); // SD v1 attention
+  int is_float = (hash.data_type == MTL::DataTypeFloat); // SD v2 attention
+  
+  uint16_t M_group = 32;
+  uint16_t N_group = 32;
+  uint16_t K_group = 32;
+  if (C_elements > 1000 * 1000) {
+    M_group = 48;
+    N_group = 48;
+  }
+  
+  // If K_simd is perfectly equal to matrix K, the compiler can elide a large
+  // amount of logic in the kernel.
+  if (hash.K >= 33 && hash.K <= 40) {
+    K_group = 40; // 1 * 40
+  } else if (is_half && hash.K >= 73 && hash.K <= 80) {
+    K_group = 40; // 2 * 40
+  } else if (C_elements > 1000 * 1000) {
+    if (hash.K <= 16) {
+      K_group = 16; // 1 * 16
+    } else if (hash.K <= 24) {
+      K_group = 24; // 1 * 24
+    } else if (hash.K <= 32) {
+      K_group = 32; // 1 * 32
+    } else if (hash.K <= 48) {
+      K_group = 24;
+    } else if (hash.K <= 64) {
+      K_group = 32;
+    } else if (is_float) {
+      K_group = 24;
+    }
+  }
+  
+  uint16_t M_splits = 2;
+  uint16_t N_splits = 2;
+  uint16_t K_splits = 1;
+  uint16_t M_simd = M_group / M_splits;
+  uint16_t N_simd = N_group / N_splits;
+  uint16_t K_simd = K_group / K_splits;
+  
+  constants->setConstantValue(&M_simd, MTL::DataTypeUShort, 200);
+  constants->setConstantValue(&N_simd, MTL::DataTypeUShort, 201);
+  constants->setConstantValue(&K_simd, MTL::DataTypeUShort, 202);
+  constants->setConstantValue(&M_splits, MTL::DataTypeUShort, 210);
+  constants->setConstantValue(&N_splits, MTL::DataTypeUShort, 211);
+  constants->setConstantValue(&K_splits, MTL::DataTypeUShort, 212);
+  
+  std::string cpp_name;
+  uint16_t data_type_size = UINT16_MAX;
+  switch (hash.data_type) {
+    case MTL::DataTypeHalf: {
+      cpp_name = "hgemm";
+      data_type_size = 2;
+      break;
+    }
+    case MTL::DataTypeFloat: {
+      cpp_name = "sgemm";
+      data_type_size = 4;
+      break;
+    }
+    default: {
+      CCV_NNC_MFA_PRECONDITION(false)
+      break;
+    }
+  }
+  auto* swift_name = NS::String::string(cpp_name.c_str(), NS::UTF8StringEncoding);
+  
+  uint16_t A_block_bytes = M_group * K_group * data_type_size;
+  uint16_t B_block_bytes = K_group * N_group * data_type_size;
+  uint16_t C_block_bytes = M_group * N_group * data_type_size;
+  threadgroup_memory_length = A_block_bytes + B_block_bytes;
+  
+  if ((hash.M % 8 > 0) && (hash.N % 8 > 0)) {
+    if (C_block_bytes > threadgroup_memory_length) {
+      threadgroup_memory_length = C_block_bytes;
+    }
+  }
+  
+  std::function<size_t(size_t, uint16_t)> ceil_divide = [](size_t original, uint16_t granularity) {
+    return (original + size_t(granularity) - 1) / size_t(granularity);
+  };
+  grid_size = MTL::Size(ceil_divide(hash.N, N_group), ceil_divide(hash.M, M_group), 1);
+  group_size = MTL::Size(128 * K_splits, 1, 1);
+  
+  NS::Error* error;
+  auto function = NS::TransferPtr(context->library->newFunction(swift_name, constants.get(), &error));
+  CCV_NNC_MFA_CHECK_ERROR(error)
+  
+  if (async) {
+    context->device->newComputePipelineState(function.get(), [=](MTL::ComputePipelineState* pipeline, NS::Error* error) {
+      CCV_NNC_MFA_CHECK_ERROR(error)
+      
+      pipeline->retain();
+      pso = pipeline;
+      semaphore->signal();
+    });
+  } else {
+    pso = context->device->newComputePipelineState(function.get(), &error);
+    CCV_NNC_MFA_CHECK_ERROR(error)
+  }
+  
+  pool->drain();
+}
+
+mfa::gemm::pipeline::~pipeline() {
+  if (semaphore) {
+    delete semaphore;
+  }
+  pso->release();
+}
+
+void mfa::gemm::pipeline::wait() {
+  if (!finished) {
+    semaphore->wait();
+    finished = true;
+  }
+}
+
+MTL::ComputePipelineState* mfa::gemm::pipeline::get_pso() const {
+  if (finished) {
+    return pso;
+  } else {
+    return nullptr;
+  }
+}
+
+bool mfa::gemm::pipeline::get_batched() const {
+  if (finished) {
+    return batched;
+  } else {
+    return false;
+  }
+}
+
+uint16_t mfa::gemm::pipeline::get_threadgroup_memory_length() const {
+  if (finished) {
+    return threadgroup_memory_length;
+  } else {
+    return UINT16_MAX;
+  }
+}
+
+MTL::Size mfa::gemm::pipeline::get_grid_size() const {
+  if (finished) {
+    return grid_size;
+  } else {
+    return MTL::Size(0, UINT64_MAX, UINT64_MAX);
+  }
+}
+
+MTL::Size mfa::gemm::pipeline::get_group_size() const {
+  if (finished) {
+    return group_size;
+  } else {
+    return MTL::Size(0, UINT64_MAX, UINT64_MAX);
+  }
+}
+
+std::ostream& operator<<(std::ostream& os, const mfa::gemm::hash& hash)
+{
+  os << "mfa::gemm::hash {";
+  os << " .data_type = " << hash.data_type << ',';
+  os << " .M = " << hash.M << ',';
+  os << " .N = " << hash.N << ',';
+  os << " .K = " << hash.K << ',';
+  os << " .A_trans = " << bool(hash.A_trans) << ',';
+  os << " .B_trans = " << bool(hash.B_trans) << ',';
+  os << " .alpha = " << double(hash.alpha) << ',';
+  os << " .beta = " << double(hash.beta) << ',';
+  os << " .batched = " << bool(hash.batched) << ',';
+  os << " .fused_activation = " << bool(hash.fused_activation);
+  os << "}";
+  return os;
+}
+
+std::size_t std::hash<mfa::gemm::hash>::operator()(const mfa::gemm::hash& hash) const noexcept {
+  std::size_t seed = 0;
+  mfa::hash::combine_64(seed, hash.data_type);
+  mfa::hash::combine_32(seed, hash.M);
+  mfa::hash::combine_32(seed, hash.N);
+  mfa::hash::combine_32(seed, hash.K);
+  mfa::hash::combine_32(seed, uint32_t(hash.A_trans));
+  mfa::hash::combine_32(seed, uint32_t(hash.B_trans));
+  mfa::hash::combine_32(seed, *reinterpret_cast<const uint32_t*>(&hash.alpha));
+  mfa::hash::combine_32(seed, *reinterpret_cast<const uint32_t*>(&hash.beta));
+  mfa::hash::combine_32(seed, uint32_t(hash.batched));
+  mfa::hash::combine_32(seed, uint32_t(hash.fused_activation));
+  return seed;
+}
diff --git a/lib/nnc/mfa/ccv_nnc_mfa_gemm.hpp b/lib/nnc/mfa/ccv_nnc_mfa_gemm.hpp
new file mode 100644
index 000000000..8bf66fe22
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa_gemm.hpp
@@ -0,0 +1,100 @@
+#ifndef GUARD_ccv_nnc_mfa_gemm_hpp
+#define GUARD_ccv_nnc_mfa_gemm_hpp
+
+typedef struct {
+  uint64_t data_type;
+  uint32_t M;
+  uint32_t N;
+  uint32_t K;
+  uint8_t A_trans;
+  uint8_t B_trans;
+  float alpha;
+  float beta;
+  uint8_t batched;
+  uint8_t fused_activation;
+  
+  // Fill these in the same order as the original shape, but null-terminated.
+  // Both arrays must have the same length.
+  uint32_t batch_dims_a[CCV_NNC_MAX_DIM_ALLOC];
+  uint32_t batch_dims_b[CCV_NNC_MAX_DIM_ALLOC];
+} ccv_nnc_mfa_gemm_params_t;
+
+#ifdef __cplusplus
+#include "nnc/mfa/3rdparty/metal-cpp/Dispatch.hpp"
+#include "nnc/mfa/3rdparty/metal-cpp/Metal.hpp"
+
+namespace ccv {
+namespace nnc {
+namespace mfa {
+namespace gemm {
+
+class hash {
+public:
+  uint64_t data_type;
+  uint32_t M;
+  uint32_t N;
+  uint32_t K;
+  uint8_t A_trans;
+  uint8_t B_trans;
+  float alpha;
+  float beta;
+  uint8_t batched;
+  uint8_t fused_activation;
+  
+  hash(ccv_nnc_mfa_gemm_params_t);
+  
+  bool operator==(const hash& rhs) const;
+};
+
+class pipeline {
+  bool finished;
+  Dispatch::Semaphore* semaphore;
+  
+  MTL::ComputePipelineState* pso;
+  
+  bool batched;
+  uint16_t threadgroup_memory_length;
+  MTL::Size grid_size;
+  MTL::Size group_size;
+  
+public:
+  pipeline(context* context, hash hash, bool async);
+  ~pipeline();
+  
+  // This is a potentially blocking function. Call it before accessing any of
+  // the property getters.
+  void wait();
+  
+  MTL::ComputePipelineState* get_pso() const;
+  
+  bool get_batched() const;
+  uint16_t get_threadgroup_memory_length() const;
+  MTL::Size get_grid_size() const;
+  MTL::Size get_group_size() const;
+};
+
+} // namespace gemm
+} // namespace mfa
+} // namespace nnc
+} // namespace ccv
+
+std::ostream& operator<<(std::ostream& os, const ccv::nnc::mfa::gemm::hash& hash);
+
+template<>
+struct std::hash<ccv::nnc::mfa::gemm::hash>
+{
+  std::size_t operator()(const ccv::nnc::mfa::gemm::hash& hash) const noexcept;
+};
+
+extern "C" {
+#endif // __cplusplus
+
+void ccv_nnc_mfa_async_prepare_gemm(ccv_nnc_mfa_context_t* context, ccv_nnc_mfa_gemm_params_t params);
+void ccv_nnc_mfa_sync_prepare_gemm(ccv_nnc_mfa_context_t* context, ccv_nnc_mfa_gemm_params_t params);
+void ccv_nnc_mfa_encode_gemm(ccv_nnc_mfa_context_t* context, ccv_nnc_mfa_gemm_params_t params, mtl_command_batch_t* command_batch, mtl_buffer_t** tensors, size_t* tensor_offsets);
+
+#ifdef __cplusplus
+} // extern "C"
+#endif // __cplusplus
+
+#endif
diff --git a/lib/nnc/mfa/ccv_nnc_mfa_hash.hpp b/lib/nnc/mfa/ccv_nnc_mfa_hash.hpp
new file mode 100644
index 000000000..60d58a056
--- /dev/null
+++ b/lib/nnc/mfa/ccv_nnc_mfa_hash.hpp
@@ -0,0 +1,58 @@
+#ifndef GUARD_ccv_nnc_mfa_hash_hpp
+#define GUARD_ccv_nnc_mfa_hash_hpp
+
+// Source:
+// https://stackoverflow.com/a/50978188
+
+namespace {
+template<typename T>
+T xorshift(const T& n,int i){
+  return n^(n>>i);
+}
+
+// a hash function with another name as to not confuse with std::hash
+uint32_t distribute_32(const uint32_t& n){
+  uint32_t p = 0x55555555ul; // pattern of alternating 0 and 1
+  uint32_t c = 3423571495ul; // random uneven integer constant;
+  return c*xorshift(p*xorshift(n,16),16);
+}
+
+// a hash function with another name as to not confuse with std::hash
+uint64_t distribute_64(const uint64_t& n){
+  uint64_t p = 0x5555555555555555ull; // pattern of alternating 0 and 1
+  uint64_t c = 17316035218449499591ull;// random uneven integer constant;
+  return c*xorshift(p*xorshift(n,32),32);
+}
+
+// if c++20 rotl is not available:
+template <typename T,typename S>
+typename std::enable_if<std::is_unsigned<T>::value,T>::type
+constexpr rotl(const T n, const S i){
+  const T m = (std::numeric_limits<T>::digits-1);
+  const T c = i&m;
+  return (n<<c)|(n>>((T(0)-c)&m)); // this is usually recognized by the compiler to mean rotation, also c++20 now gives us rotl directly
+}
+}
+
+namespace ccv {
+namespace nnc {
+namespace mfa {
+namespace hash {
+
+// call this function with the old seed and the new key to be hashed and combined into the new seed value, respectively the final hash
+inline size_t combine_32(std::size_t& seed, const uint32_t& v)
+{
+    return rotl(seed,std::numeric_limits<size_t>::digits/3) ^ distribute_32(v);
+}
+
+inline size_t combine_64(std::size_t& seed, const uint64_t& v)
+{
+    return rotl(seed,std::numeric_limits<size_t>::digits/3) ^ distribute_64(v);
+}
+
+} // namespace hash
+} // namespace mfa
+} // namespace nnc
+} // namespace ccv
+
+#endif
diff --git a/lib/nnc/mfa/makefile b/lib/nnc/mfa/makefile
new file mode 100644
index 000000000..864d18d31
--- /dev/null
+++ b/lib/nnc/mfa/makefile
@@ -0,0 +1,39 @@
+include ../../config.mk
+
+CFLAGS := -std=c++17 -O3 -Wall -I"../../" $(CFLAGS)
+
+SRCS := Metal.cpp ccv_nnc_mfa.cpp ccv_nnc_mfa_error.cpp ccv_nnc_mfa_gemm.cpp 3rdparty/metal-cpp/Dispatch.cpp
+
+SRC_OBJS := $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(SRCS)))
+
+.PHONY: release all lib clean
+
+release: all
+
+include ../../scheme.mk
+
+all: lib
+
+lib: libnnc-compat-mfa.o
+
+clean:
+	rm -f *.o
+
+libnnc-compat-mfa.o: $(SRC_OBJS)
+	ld -r $^ -o $@
+
+%.o: %.c
+	$(CC) $< -o $@ -c $(CFLAGS)
+
+%.o: %.cpp
+	$(CC) $< -o $@ -c $(CFLAGS)
+
+dep: .dep.mk
+.dep.mk: $(SRCS)
+	echo '' > .dep.mk
+	for SRC in $(patsubst %.c,,$^) ; do \
+		$(CC) $(CFLAGS) -M $$SRC | sed -e 's/^.*\://g' | (echo "$${SRC%%.*}.o: \\" && cat) >> .dep.mk ; \
+	done
+
+-include .dep.mk
+
diff --git a/lib/nnc/mps/ccv_nnc_mps.h b/lib/nnc/mps/ccv_nnc_mps.h
index 191d252e0..4ada91697 100644
--- a/lib/nnc/mps/ccv_nnc_mps.h
+++ b/lib/nnc/mps/ccv_nnc_mps.h
@@ -3,6 +3,7 @@
 
 #include "nnc/ccv_nnc.h"
 #include "nnc/_ccv_nnc_stream.h"
+#include "nnc/mfa/ccv_nnc_mfa.hpp"
 
 void* mpheapalloc(int device, size_t size);
 void mpheapfree(int device, void* ptr);
@@ -61,8 +62,11 @@ typedef struct {
 off_t mpgetoffset(const ccv_nnc_tensor_t* const tensor);
 id<MTLBuffer> mpgetbuffer(const ccv_nnc_tensor_t* const tensor);
 id<MTLDevice> ccv_nnc_default_device(void);
-CCV_WARN_UNUSED(MPSCommandBuffer*) ccv_nnc_stream_context_get_command_buffer(ccv_nnc_stream_context_t* const stream_context);
-void ccv_nnc_stream_context_commit_command_buffer(ccv_nnc_stream_context_t* const stream_context, MPSCommandBuffer* command_buffer);
+ccv_nnc_mfa_context_t* ccv_nnc_default_mfa_context(void);
+CCV_WARN_UNUSED(MTLCommandBatch*) ccv_nnc_stream_context_start_command_batch(ccv_nnc_stream_context_t* const stream_context);
+CCV_WARN_UNUSED(MPSCommandBuffer*) ccv_nnc_stream_context_start_mps_command_buffer(ccv_nnc_stream_context_t* const stream_context);
+void ccv_nnc_stream_context_finish_command_batch(ccv_nnc_stream_context_t* const stream_context, MTLCommandBatch* command_batch);
+void ccv_nnc_stream_context_finish_mps_command_buffer(ccv_nnc_stream_context_t* const stream_context, MPSCommandBuffer* command_buffer);
 CCV_WARN_UNUSED(MPSGraphExecutable*) ccv_nnc_mps_graph_executable_cache(const ccv_nnc_mps_graph_key_t key, int* indices, void(NS_NOESCAPE ^block)(MPSGraph* graph, NSMutableArray<MPSGraphTensor*>* inputTensors, NSMutableArray<MPSGraphShapedType*>* inputShapedTypes, NSMutableArray<MPSGraphTensor*>* resultTensors));
 CCV_WARN_UNUSED(ccv_nnc_mps_graph_key_t) ccv_nnc_mps_graph_key_new(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const int flags, ccv_nnc_tensor_t* const* const inputs, const int input_size, ccv_nnc_tensor_t* const* const outputs, const int output_size);
 CCV_WARN_UNUSED(MPSDataType) ccv_nnc_mps_datatype(const int datatype); // Get the datatype corresponding to MPS datatype.
diff --git a/lib/nnc/mps/ccv_nnc_mps.m b/lib/nnc/mps/ccv_nnc_mps.m
index e5c29c311..10df9db84 100644
--- a/lib/nnc/mps/ccv_nnc_mps.m
+++ b/lib/nnc/mps/ccv_nnc_mps.m
@@ -6,6 +6,7 @@
 #include <string.h>
 #import <CoreFoundation/CoreFoundation.h>
 #import <Foundation/Foundation.h>
+#import <TargetConditionals.h>
 #import <MetalPerformanceShaders/MetalPerformanceShaders.h>
 #import <MetalPerformanceShadersGraph/MetalPerformanceShadersGraph.h>
 #import <objc/runtime.h>
@@ -24,6 +25,32 @@
 	return device;
 }
 
+@interface MTLFileBackedBuffer: NSObject
+@property (nonatomic, copy) NSString* path;
+@property (nonatomic, assign) NSUInteger size;
+@end
+
+ccv_nnc_mfa_context_t* ccv_nnc_default_mfa_context(void)
+{
+	static dispatch_once_t once;
+	static ccv_nnc_mfa_context_t* context;
+	dispatch_once(&once, ^{
+		const char* metallib_path = getenv("CCV_NNC_MFA_METALLIB_PATH");
+		if (metallib_path)
+			context = ccv_nnc_init_mfa_context((__bridge mtl_device_t*)ccv_nnc_default_device(), metallib_path);
+		else {
+			NSBundle* bundle = [NSBundle bundleForClass:[MTLFileBackedBuffer class]];
+#if TARGET_OS_IPHONE || TARGET_OS_MACCATALYST
+			NSString* path = [bundle pathForResource:@"libmfaios16-0.2" ofType:@"metallib"];
+#else
+			NSString* path = [bundle pathForResource:@"libmfamacos13-0.2" ofType:@"metallib"];
+#endif
+			context = ccv_nnc_init_mfa_context((__bridge mtl_device_t*)ccv_nnc_default_device(), path.UTF8String);
+		}
+	});
+	return context;
+}
+
 MPSGraphDevice* _ccv_nnc_default_mps_device(void)
 {
 	static dispatch_once_t once;
@@ -196,10 +223,6 @@ void mpobjfree(int device, void* ptr)
 	return buffer;
 }
 
-@interface MTLFileBackedBuffer: NSObject
-@property (nonatomic, copy) NSString* path;
-@property (nonatomic, assign) NSUInteger size;
-@end
 @implementation MTLFileBackedBuffer
 @end
 
@@ -517,6 +540,8 @@ ccv_nnc_mps_graph_key_t ccv_nnc_mps_graph_key_new(const ccv_nnc_cmd_t cmd, const
 // Stream context
 ccv_nnc_stream_context_t* ccv_nnc_init_stream_context(ccv_nnc_stream_context_t* const stream_context)
 {
+	// Initialize the MFA context.
+	ccv_nnc_default_mfa_context();
 	return stream_context;
 }
 
@@ -657,7 +682,12 @@ int ccv_nnc_gpu_device_count(void)
 	return 1;
 }
 
-MPSCommandBuffer* ccv_nnc_stream_context_get_command_buffer(ccv_nnc_stream_context_t* const stream_context)
+MTLCommandBatch* ccv_nnc_stream_context_start_command_batch(ccv_nnc_stream_context_t* const stream_context)
+{
+	return ccv_nnc_start_command_batch((__bridge mtl_command_queue_t*)_ccv_nnc_default_queue());
+}
+
+MPSCommandBuffer* ccv_nnc_stream_context_start_mps_command_buffer(ccv_nnc_stream_context_t* const stream_context)
 {
 	return [MPSCommandBuffer commandBufferFromCommandQueue:_ccv_nnc_default_queue()];
 }
@@ -667,14 +697,25 @@ void ccv_nnc_mps_unbounded_command_buffers(int state)
 	enable_unbounded_command_buffers = state;
 }
 
-void ccv_nnc_stream_context_commit_command_buffer(ccv_nnc_stream_context_t* const stream_context, MPSCommandBuffer* command_buffer)
+void ccv_nnc_stream_context_finish_command_buffer(ccv_nnc_stream_context_t* const stream_context, MPSCommandBuffer* mps_command_buffer, MTLCommandBatch* command_batch)
 {
+	id<MTLCommandBuffer> mtl_command_buffer;
+	if (mps_command_buffer != nil) {
+		mtl_command_buffer = mps_command_buffer.commandBuffer;
+	} else {
+		mtl_command_buffer = command_batch->command_buffer;
+	}
+	
 	int i;
 	const int buffer_size = enable_unbounded_command_buffers ? OLD_MAX_COMMAND_BUFFER_SIZE : OLD_LIMITED_COMMAND_BUFFER_SIZE;
 	if (!stream_context)
 	{
-		id<MTLCommandBuffer> committed_command_buffer = [command_buffer.commandBuffer retain];
-		[command_buffer commit];
+		id<MTLCommandBuffer> committed_command_buffer = [mtl_command_buffer retain];
+		if (mps_command_buffer != nil) {
+			[mps_command_buffer commit];
+		} else {
+			ccv_nnc_finish_command_batch(command_batch);
+		}
 		id<MTLCommandBuffer> last_buffer;
 		id<MTLCommandBuffer> old_buffers[buffer_size];
 		os_unfair_lock_lock(&queue_lock);
@@ -703,10 +744,53 @@ void ccv_nnc_stream_context_commit_command_buffer(ccv_nnc_stream_context_t* cons
 			old_last_command_buffers[i] = old_last_command_buffers[i + 1];
 		old_last_command_buffers[buffer_size - 1] = last_command_buffer;
 	} else
-		old_last_command_buffer = [command_buffer.commandBuffer retain];
-	last_command_buffer = [command_buffer.commandBuffer retain];
+		old_last_command_buffer = [mtl_command_buffer retain];
+	last_command_buffer = [mtl_command_buffer retain];
+	
+	// There is an opportunity to automatically batch MFA commands or custom
+	// shaders into a command batch. Instead of explicitly starting and
+	// finishing, have a background thread automatically commit it. To prevent
+	// committing from happenning in the middle of encoding, protect the command
+	// batch using `queue_lock` (unknown latency) or a pthread mutex lock (~200 ns
+	// latency).
+	//
+	// Every 50 microseconds, the background thread checks whether a command batch
+	// is active. If so, it commits all the commands. The main thread can also
+	// commit when `batched_command_count` exceeds a certain threshold (while it's
+	// still holding the lock). The best threshold is unknown, but 8 would be a
+	// reasonable first guess.
+	//
+	// When encountering an MPS command, you will have to abort the command batch
+	// (i.e. eagerly commit it with `batched_command_count=1`) because
+	// MPSCommandBuffer can't guarantee the command buffer stays the same. Even if
+	// it did, creating a separate `MTL::ComputeCommandEncoder` for each command
+	// is no better than creating a new `MTL::CommandBuffer`. Closing the
+	// `MTL::ComputeCommandEncoder` and creating a new `MTL::BlitCommandEncoder`
+	// for memory copies is equally as slow.
+	//
+	// Until ~50% of all operations transition from MPS -> custom shaders, this
+	// optimization is not worthwhile. It should only be employed on platforms
+	// where custom shaders are consistently faster than MPS (e.g. Apple 7+ with
+	// MFA GEMM). A good start would be creating custom shaders for all the
+	// elementwise operations in MPSGraph, and 4-byte aligned memcpy/memset. There
+	// are many places in ML models where elementwise activations follow GEMM. 90%
+	// of the time, they would be automatically batched within the 50-µs window,
+	// providing a 10x speedup for those layers.
+	//
+	// As a final optimization, delay the encoding of the GEMM. Fuse the GEMM +
+	// activation into one command through the `fused_activation` MFA function
+	// constant. This will require two separate GEMM variants to be ready, one
+	// with and one without `fused_activation` enabled. Same for the elementwise -
+	// one `MTLComputePipelineState` and another `MTLVisibleFunctionTable`. It
+	// might be wise to delay the creation of the fused variant, until you detect
+	// a specific GEMM shape received 2+ opportunities for fusion.
+	//
+	// Alternative implementation path: compile-time graph transformations that
+	// - Check whether MFA is supported
+	// - Fuse groups of consecutive MFA-compatible commands
 	os_unfair_lock_unlock(&queue_lock);
-	[command_buffer.commandBuffer addCompletedHandler:^(id<MTLCommandBuffer> buffer) {
+	
+	[mtl_command_buffer addCompletedHandler:^(id<MTLCommandBuffer> buffer) {
 		id<MTLCommandBuffer> found_buffer = nil;
 		os_unfair_lock_lock(&queue_lock);
 		if (buffer == last_command_buffer)
@@ -726,12 +810,26 @@ void ccv_nnc_stream_context_commit_command_buffer(ccv_nnc_stream_context_t* cons
 		os_unfair_lock_unlock(&queue_lock);
 		[found_buffer release];
 	}];
-	[command_buffer commit];
+	if (mps_command_buffer != nil) {
+		[mps_command_buffer commit];
+	} else {
+		ccv_nnc_finish_command_batch(command_batch);
+	}
 	// Wait if we need to bound how many in-flight command buffers there are. This helps memory usage.
 	[old_last_command_buffer waitUntilCompleted];
 	[old_last_command_buffer release];
 }
 
+void ccv_nnc_stream_context_finish_command_batch(ccv_nnc_stream_context_t* const stream_context, MTLCommandBatch* command_batch)
+{
+	ccv_nnc_stream_context_finish_command_buffer(stream_context, nil, command_batch);
+}
+
+void ccv_nnc_stream_context_finish_mps_command_buffer(ccv_nnc_stream_context_t* const stream_context, MPSCommandBuffer* command_buffer)
+{
+	ccv_nnc_stream_context_finish_command_buffer(stream_context, command_buffer, NULL);
+}
+
 MPSDataType ccv_nnc_mps_datatype(const int datatype)
 {
 	switch (datatype)
diff --git a/lib/nnc/mps/makefile b/lib/nnc/mps/makefile
index e90ad192f..cbba0a9c9 100644
--- a/lib/nnc/mps/makefile
+++ b/lib/nnc/mps/makefile
@@ -32,7 +32,7 @@ dep: .dep.mk
 .dep.mk: $(SRCS)
 	echo '' > .dep.mk
 	for SRC in $(patsubst %.c,,$^) ; do \
-		$(NVCC) $(NVFLAGS) -M $$SRC | sed -e 's/^.*\://g' | (echo "$${SRC%%.*}.o: \\" && cat) >> .dep.mk ; \
+		$(CC) $(CFLAGS) -M $$SRC | sed -e 's/^.*\://g' | (echo "$${SRC%%.*}.o: \\" && cat) >> .dep.mk ; \
 	done
 
 -include .dep.mk
diff --git a/test/int/nnc/mpsdnn.tests.c b/test/int/nnc/mpsdnn.tests.c
index 67e5ddf76..2dde7e26b 100644
--- a/test/int/nnc/mpsdnn.tests.c
+++ b/test/int/nnc/mpsdnn.tests.c
@@ -377,7 +377,6 @@ TEST_CASE("compare sigmoid with mps in half precision")
 
 TEST_CASE("compare sigmoid gradient with mps")
 {
-	ccv_cli_set_output_levels(ccv_cli_output_level_and_above(CCV_CLI_VERBOSE));
 	GUARD_ELSE_RETURN(ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_FORWARD, CCV_NNC_BACKEND_MPS) &&
 		ccv_nnc_cmd_ok(CCV_NNC_SIGMOID_BACKWARD, CCV_NNC_BACKEND_MPS));
 	ccv_nnc_symbolic_graph_t* const symbolic_graph = ccv_nnc_symbolic_graph_new();