/* fv0 fv4 fv8 fv12 */
        .global _sobel_y
_sobel_y:
__setup:
        mova _const_100f,r0
        mov r0,r3

        mov r4,r0
        mov r5,r7 /* r5 saved as r7 */

        fldi1 fr8    /* 1.0 */
        fldi1 fr9    /* 2.0 */
        fldi1 fr10   /* 1.0 */
        fldi0 fr11   /* 0.0 */
        fadd fr9,fr9

        fldi1 fr12
        fmov fr9,fr13
        fldi1 fr14
        fldi0 fr15
        fneg fr12
        fneg fr13
        fneg fr14

        /* offsets */
        mov #(1 * 4),r1
        mov #(2 * 4),r2
        mov.w _const_640,r4
        mov.w _const_644,r5
        mov.w _const_648,r6

        bra _loop
        nop

        .align 4
_const_100f:    .float 100
_const_640:     .word (640 * 4)
_const_644:     .word (641 * 4)
_const_648:     .word (642 * 4)

_loop:
        fmov.s @r0,fr0
        fmov.s @(r0,r1),fr1
        fmov.s @(r0,r2),fr2
        fldi0 fr3
        fipr fv8,fv0

        fmov.s @(r0,r4),fr4
        fmov.s @(r0,r5),fr5
        fmov.s @(r0,r6),fr6
        fldi0 fr7
        fipr fv12,fv4

        add #4,r0

        fadd fr3,fr7

        fmov.s r1,@-r7

        rts
        nop