[x264-devel] [PATCH] Updated configure to detect MIPS platform and set flags. Added define for MIPS MSA. MSA is SIMD architecture of MIPS. Flags are added to enable MSA.

[BugMaster]

On Tue, 24 Mar 2015 14:31:05 +0530, Kaustubh Raste wrote:
> ---
>  configure | 46 ++++++++++++++++++++++++++++++++++++++++++++++
>  x264.h    |  3 +++
>  2 files changed, 49 insertions(+)

> diff --git a/configure b/configure
> index 15b1c91..1999ad6 100755
> --- a/configure
> +++ b/configure
> @@ -735,6 +735,7 @@ case $host_cpu in
>          ;;
>      mips|mipsel|mips64|mips64el)
>          ARCH="MIPS"
> +        AS="${AS-${cross_prefix}gcc}"

AS="${AS-${CC}}"

>          ;;
>      arm*)
>          ARCH="ARM"
> @@ -860,6 +861,51 @@ if [ $asm = auto -a \( $ARCH = ARM -o $ARCH = AARCH64 \) ] ; then
>      as_check ".func test${NL}.endfunc" && define HAVE_AS_FUNC 1
>  fi
>  
> +if [ $asm = auto -a $ARCH = MIPS ] ; then
> +    # set flags so msa is built by default
> +    echo $CFLAGS | grep -Eq '(-mcpu|-march|-arch)' || CFLAGS="$CFLAGS -mfp64 -mmsa -flax-vector-conversions"

How broad is support of MSA among MIPS cpus to make default? Same for
the -mfp64. It may be ok if we really would have optimizations that
need (same as we target ARMv7 by default) but not without real
optimizations.

https://gcc.gnu.org/onlinedocs/gcc/C-Dialect-Options.html says that
-flax-vector-conversions should not be used for new code.

> +
> +    if cc_check '' '' '__asm__("move.v $w5, $w11");' ; then define HAVE_MSA
> +        CFLAGS="$CFLAGS -msched-weight -funroll-loops -mload-store-pairs"
> +        LDFLAGS="$LDFLAGS -mmsa -mfp64"
> +
> +        if cc_check '' '' '__asm__("ulw  $6,  4($9)");' ; then
> +            case $host_cpu in
> +                mips|mipsel)
> +                    CFLAGS="$CFLAGS -mips32r5 -mtune=p5600"
> +                    ASFLAGS="$ASFLAGS -mips32r5 -mtune=p5600"
> +                    LDFLAGS="$LDFLAGS -mips32r5 -mtune=p5600"
> +                   ;;
> +                mips64|mips64el)
> +                    CFLAGS="$CFLAGS -mips64r5 -mabi=64"
> +                    ASFLAGS="$ASFLAGS -mips64r5 -mabi=64"
> +                    LDFLAGS="$LDFLAGS -mips64r5 -mabi=64"
> +                    ;;
> +            esac
> +        else
> +            case $host_cpu in
> +                mips|mipsel)
> +                    CFLAGS="$CFLAGS -mips32r6"
> +                    ASFLAGS="$ASFLAGS -mips32r6"
> +                    LDFLAGS="$LDFLAGS -mips32r6"
> +                    ;;
> +                mips64|mips64el)
> +                    CFLAGS="$CFLAGS -mips64r6 -mabi=64 -mtune=i6400"
> +                    ASFLAGS="$ASFLAGS -mips64r6 -mabi=64 -mtune=i6400"
> +                    LDFLAGS="$LDFLAGS -mips64r6 -mabi=64 -mtune=i6400"
> +                    ;;
> +            esac
> +        fi
> +    else
> +        CFLAGS="$CFLAGS -mips32r2"
> +    fi
> +
> +    if cc_check '' '' '__asm__("madd.d $f0, $f2, $f4, $f6");' ; then
> +        CFLAGS="$CFLAGS -mhard-float"
> +        ASFLAGS="$ASFLAGS -mhard-float"
> +    fi
> +fi
> +

Detection of supported CPU features by checking compiler support looks
wrong. A specially if we take cross compiling into account.

>  [ $asm = no ] && AS=""
>  [ "x$AS" = x ] && asm="no" || asm="yes"
>  
> diff --git a/x264.h b/x264.h
> index 7cc5fcc..8194246 100644
> --- a/x264.h
> +++ b/x264.h
> @@ -158,6 +158,9 @@ typedef struct
>  #define X264_CPU_FAST_NEON_MRC   0x0000004  /* Transfer from NEON
> to ARM register is fast (Cortex-A9) */
>  #define X264_CPU_ARMV8           0x0000008
>  
> +/* MIPS */
> +#define X264_CPU_MSA             0x0000001  /* MIPS MSA */

This define is nowhere set or used so this chunk is meaningless at
least without real code to use it.

> +
>  /* Analyse flags */
>  #define X264_ANALYSE_I4x4       0x0001  /* Analyse i4x4 */
>  #define X264_ANALYSE_I8x8       0x0002  /* Analyse i8x8 (requires 8x8 transform) */