diff --git a/Makefile.am b/Makefile.am index 6d1982842d..ace180b046 100644 --- a/Makefile.am +++ b/Makefile.am @@ -180,7 +180,6 @@ libs/libyuv/source/compare.cc \ libs/libyuv/source/compare_common.cc \ libs/libyuv/source/compare_gcc.cc \ libs/libyuv/source/compare_hvx.cc \ -libs/libyuv/source/compare_neon64.cc \ libs/libyuv/source/compare_neon.cc \ libs/libyuv/source/compare_wasm.cc \ libs/libyuv/source/compare_win.cc \ @@ -203,11 +202,8 @@ libs/libyuv/source/rotate_gcc.cc \ libs/libyuv/source/rotate_hvx.cc \ libs/libyuv/source/rotate_lasx.cc \ libs/libyuv/source/rotate_lsx.cc \ -libs/libyuv/source/rotate_neon64.cc \ libs/libyuv/source/rotate_neon.cc \ libs/libyuv/source/rotate_rvv.cc \ -libs/libyuv/source/rotate_sme.cc \ -libs/libyuv/source/rotate_sve.cc \ libs/libyuv/source/rotate_wasm.cc \ libs/libyuv/source/rotate_win.cc \ libs/libyuv/source/row_any.cc \ @@ -216,11 +212,8 @@ libs/libyuv/source/row_gcc.cc \ libs/libyuv/source/row_hvx.cc \ libs/libyuv/source/row_lasx.cc \ libs/libyuv/source/row_lsx.cc \ -libs/libyuv/source/row_neon64.cc \ libs/libyuv/source/row_neon.cc \ libs/libyuv/source/row_rvv.cc \ -libs/libyuv/source/row_sme.cc \ -libs/libyuv/source/row_sve.cc \ libs/libyuv/source/row_wasm.cc \ libs/libyuv/source/row_win.cc \ libs/libyuv/source/scale_any.cc \ @@ -230,11 +223,9 @@ libs/libyuv/source/scale_common.cc \ libs/libyuv/source/scale_gcc.cc \ libs/libyuv/source/scale_hvx.cc \ libs/libyuv/source/scale_lsx.cc \ -libs/libyuv/source/scale_neon64.cc \ libs/libyuv/source/scale_neon.cc \ libs/libyuv/source/scale_rgb.cc \ libs/libyuv/source/scale_rvv.cc \ -libs/libyuv/source/scale_sme.cc \ libs/libyuv/source/scale_uv.cc \ libs/libyuv/source/scale_wasm.cc \ libs/libyuv/source/scale_win.cc \ @@ -242,6 +233,46 @@ libs/libyuv/source/video_common.cc libfreeswitch_libyuv_la_CPPFLAGS = -O2 -fomit-frame-pointer -Ilibs/libyuv/include +libfreeswitch_libyuv_la_LIBADD = + +# libyuv's AArch64 dot-product (udot), i8mm (usdot), SVE and SME kernels emit +# instructions outside the baseline armv8-a target and are selected at runtime +# by libyuv's own CPU detection. Build them as separate objects with the +# required -march flags so the rest of the library stays portable to any +# armv8-a CPU. This mirrors libs/libyuv/CMakeLists.txt. +if LIBYUV_AARCH64 +noinst_LTLIBRARIES += libfreeswitch_libyuv_neon64.la +libfreeswitch_libyuv_neon64_la_SOURCES = \ +libs/libyuv/source/compare_neon64.cc \ +libs/libyuv/source/rotate_neon64.cc \ +libs/libyuv/source/row_neon64.cc \ +libs/libyuv/source/scale_neon64.cc +libfreeswitch_libyuv_neon64_la_CPPFLAGS = $(libfreeswitch_libyuv_la_CPPFLAGS) +libfreeswitch_libyuv_neon64_la_CXXFLAGS = -march=armv8.2-a+dotprod+i8mm +libfreeswitch_libyuv_la_LIBADD += libfreeswitch_libyuv_neon64.la + +noinst_LTLIBRARIES += libfreeswitch_libyuv_sve.la +libfreeswitch_libyuv_sve_la_SOURCES = \ +libs/libyuv/source/rotate_sve.cc \ +libs/libyuv/source/row_sve.cc +libfreeswitch_libyuv_sve_la_CPPFLAGS = $(libfreeswitch_libyuv_la_CPPFLAGS) +libfreeswitch_libyuv_sve_la_CXXFLAGS = -march=armv8.5-a+i8mm+sve2 +libfreeswitch_libyuv_la_LIBADD += libfreeswitch_libyuv_sve.la + +if LIBYUV_AARCH64_SME +noinst_LTLIBRARIES += libfreeswitch_libyuv_sme.la +libfreeswitch_libyuv_sme_la_SOURCES = \ +libs/libyuv/source/rotate_sme.cc \ +libs/libyuv/source/row_sme.cc \ +libs/libyuv/source/scale_sme.cc +libfreeswitch_libyuv_sme_la_CPPFLAGS = $(libfreeswitch_libyuv_la_CPPFLAGS) +libfreeswitch_libyuv_sme_la_CXXFLAGS = -march=armv8.5-a+i8mm+sme +libfreeswitch_libyuv_la_LIBADD += libfreeswitch_libyuv_sme.la +else +libfreeswitch_libyuv_la_CPPFLAGS += -DLIBYUV_DISABLE_SME +endif +endif + CORE_LIBS+=libfreeswitch_libyuv.la endif diff --git a/configure.ac b/configure.ac index 9507ab6fe1..08be69e1ff 100644 --- a/configure.ac +++ b/configure.ac @@ -549,6 +549,28 @@ AC_ARG_ENABLE(libyuv, AM_CONDITIONAL([ENABLE_LIBYUV],[test "${enable_libyuv}" = "yes"]) +# libyuv's AArch64 SIMD kernels (dot-product, i8mm, SVE, SME) use instructions +# outside the baseline armv8-a target, so they must be compiled with per-file +# -march flags (see Makefile.am). These conditionals mirror the logic in +# libs/libyuv/CMakeLists.txt. +AM_CONDITIONAL([LIBYUV_AARCH64], + [test "x$host_cpu" = "xaarch64" || test "x$host_cpu" = "xarm64"]) + +libyuv_have_sme=no +AS_IF([test "x$host_cpu" = "xaarch64" || test "x$host_cpu" = "xarm64"], [ + AC_LANG_PUSH([C]) + libyuv_save_CFLAGS="$CFLAGS" + CFLAGS="$CFLAGS -march=armv8.5-a+i8mm+sme" + AC_MSG_CHECKING([whether $CC supports AArch64 SME kernels]) + AC_COMPILE_IFELSE([AC_LANG_SOURCE([[__arm_locally_streaming void func(void) { } +int main(void) { return 0; }]])], + [libyuv_have_sme=yes]) + AC_MSG_RESULT([$libyuv_have_sme]) + CFLAGS="$libyuv_save_CFLAGS" + AC_LANG_POP([C]) +]) +AM_CONDITIONAL([LIBYUV_AARCH64_SME], [test "x$libyuv_have_sme" = "xyes"]) + AC_ARG_ENABLE(libvpx, [AC_HELP_STRING([--disable-libvpx],[build without libvpx])],[enable_libvpx="$enableval"],[enable_libvpx="yes"])