intapply_gaussian:
    push {r4-r10, fp, lr}
    add fp, sp, #32
    // sub SP, SP, #4
    mov r4, #0 //i
    mov r7, r0
    mov r8, r1
    mov r9, r2
    mov r10, r3
forExt:
    CMP r4, r10
    BGE endForExt
    MOV r5, #0 // j
forInt:
    CMP r5, r9
    BGE endForInt
    MOV r0, #3
    MUL r6, r4, r9 // i * width
    ADD r6, r6, r5 // i * width + j
    // MUL r6, r0, r6 // each pixel occupies 3 bytes
    ADD r6, r6, r8 // addr(im2) + i * width + j
    MOV r0, r7
    MOV r1, r9
    MOV r2, r10
    MOV r3, r4
    PUSH {r5}
    BL gaussian
    STR r0, [r6]
    POP {r5}
    ADD r5, r5, #1
    B
forIntendForInt:
    ADD r4, r4, #1
    B
forExtendForExt:
    sub sp, fp, #32
    pop {r4-r10,fp,lr}
    bx lr
.end