SAD, sum of absolute differences in ASM (NEON instr set)

    .cpu cortex-a8
    .fpu neon

    .text
    .align  2
    .global _Z8rect_sadPKhS0_jjj
    .type   _Z8rect_sadPKhS0_jjj, %function

_Z8rect_sadPKhS0_jjj:
    .fnstart
    src_line1_r0 .req  r0
    targ_line1_r1 .req r1
    rect_w_r2 .req r2
    rect_h_r3 .req r3
    stride_r4 .req  r4

    gap_r6 .req r6
    end_r7 .req r7
    line_end_r8 .req r8

.LFB0:
    @ args = 4, pretend = 0, frame = 8
    @ frame_needed = 0, uses_anonymous_args = 0
    @ link register save eliminated.
    @ r0 = stride
    @ r1 = targ_line
    @ r2 = rect_w
    @ r3 = rect_h
    @ [sp,#40] = stride

    stmfd   sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12}
    .save {r4, r5, r6, r7, r8, r9, r10, r11, r12}

.LCFI0:
    .pad #8
    sub sp, sp, #8  @,,

    ldr stride_r4, [sp, #40]                                @ stride, stride

    pld [src_line1_r0]
    pld [targ_line1_r1]

    veor.u8 q0, q0, q0                                      @ q0 = #0
    veor.u8 q2, q2, q2                                      @ q2 = #0

    sub gap_r6, stride_r4, rect_w_r2
    mla end_r7, rect_h_r3, stride_r4, src_line1_r0          @ end = rect_h * stride + src_line

.rect_sad_vert_loop:

    pld [src_line1_r0, #1024]
    pld [targ_line1_r1, #1024]

    add line_end_r8, src_line1_r0, rect_w_r2                    @ line_end = src_line + rect_w
    veor.u8 q1, q1, q1                                      @ q1 = #0

.rect_sad_horz_loop:
    vld1.8          {d4}, [src_line1_r0]!                       @ d4 = src_line[0..7]
    vld1.8          {d5}, [targ_line1_r1]!                      @ d5 = targ_line[0..7]

    vabal.u8        q1, d4, d5                              @ q1 += |d4 - d5|

    cmp src_line1_r0, line_end_r8                           @ if(src_line < line_end)
    blt .rect_sad_horz_loop

.rect_sad_vert_loop_end:
    vpadalq.u16 q0, q1                                      @ q0 = sum(q1)

    add src_line1_r0, src_line1_r0, gap_r6      @ src_line1 += gap_r6
    add targ_line1_r1, targ_line1_r1, gap_r6    @ targ_line1 += gap_r6

    cmp src_line1_r0, end_r7
    blt .rect_sad_vert_loop

.rect_sad_ret:
    vpadd.u32 d0, d0, d1                        @ d0 = d0 + d1
    vmov r0, r1, d0                             @ r0 = d0[0], r1 = d0[1]
    add  r0, r0, r1                             @ r0 = r0 + r1

    add sp, sp, #8              @,,
    ldmfd   sp!, {r4, r5, r6, r7, r8, r9, r10, r11, r12}
    bx  lr
.LFE0:
    .fnend
    .size   _Z8rect_sadPKhS0_jjj, .-_Z8rect_sadPKhS0_jjj
    .ident  "GCC: (GNU) 4.4.3"
    .section    .note.GNU-stack,"",%progbits