roulette

;# Funktionen
.globl _asm_matrQuad
.globl _asm_matrTransp

.code 32

;# _asm_matrQuad(float *input, float *result);
_asm_matrQuad:
    push {r4-r7, sp, lr}                ;#Save to the stack

    mov r4, r1                              ;#Save pointer of r1 to r4
    sub sp, #256                            ;#Make space for the result matrix
    mov r1, sp
    bl _asm_matrTransp

    ;#set reg to default values
    ;#r5 = cache for pointer to a row
    mov r6, #0                              ;#row counter in table 1
    mov r7, #0                              ;#row counter in table 2

  iterate:
    ;#load the left table row
    mov r5, r6, LSL #5
    vldm r5, {q2, q3}                       ;#Load the current row in table 1 into q2 and q3
    mov r5, r7, LSL #5
    vldm r5, {q4, q5}                       ;#Load the current transposed column in table 2 into q4 and q5

    vmul.f32 q0, q2, q4                     ;#multiply q2 and q4
    vmla.f32 q0, q3, q5                     ;#add q2*q4 to q3*q5

    vadd.f32 d0, d1
    vadd.f32 s0, s1


    pop {r4-r7, sp, pc}             ;#Load the old variables from the stack
    BX lr


;#void _asm_matrTransp(float *input, float *result);
_asm_matrTransp:
    push {r4-r8, lr};               ;#Save to the stack

    mov r4, r0                      ;# Pointer to first matrix element of the current column in r4
    mov r5, #0                      ;# r5 current row
    mov r8, #0                      ;# r8 current column
    mov r6, #0                      ;# r6 is the current offset of the target result
  iterate:
    ldr r7, [r4, r5, LSL #5]        ;# r7 contains the current load element
    str r7, [r1, r6, LSL #2]        ;# Save the value of r7 into the result matrix

    add r6, #1                      ;#go to next result cell
    add r5, #1                      ;#go to the next row in the input matrix
    cmp r5, #8                      ;# row == 8 then next column
    bne iterate
    ;# If we are here we go to the next row
    mov r5, #0                      ;# row = 0
    add r4, #4                      ;# column pointer to the next float
    add r8, #1                      ;# column++
    cmp r8, #8
    bne iterate                     ;# if we have done all 8 columns its over

    pop {r4-r8, pc};                ;#Load the old variables from the stack