/*=== W[ 0] ===*/
  ldr r14, T05_ADDR
  ldr r12, T07_ADDR
  ldr r11, T10_ADDR
  ldr r10, T13_ADDR
  ldr r9, T14_ADDR
  sub r3, r9, r12
/*(-- should do +10, +13, +1tr, +5 --)*/
  add r5, r11, r10
  add r5, r3
  add r5, r14
  S32_0 r8 r5
  ldr r5, [r1, #1*4]
  add r8, r5
  str r8, [r0, #0*4]
/*=== W[ 3] ===*/
  ldr r14, T00_ADDR
  ldr r9, T01_ADDR
  ldr r8, T08_ADDR
  ldr r7, T10_ADDR
  ldr r6, T13_ADDR
  add r2, r9, r11
/*(-- should do +0, +13, +8, -2tr --)*/
  add r5, r14, r10
  add r5, r8
  sub r5, r2
  S32_3 r9 r5
  ldr r5, [r1, #4*4]
  add r9, r5
  str r9, [r0, #3*4]
/*=== W[ 6] ===*/
  ldr r12, T00_ADDR
  ldr r11, T03_ADDR
  ldr r9, T04_ADDR
  ldr r8, T11_ADDR
  ldr r7, T13_ADDR
/*(-- should do +13, +4, -0, -11, -3 --)*/
  add r5, r10, r9
  sub r5, r14
  sub r5, r8
  sub r5, r11
  S32_1 r8 r5
  ldr r5, [r1, #7*4]
  add r8, r5
  str r8, [r0, #6*4]
/*=== W[ 9] ===*/
  ldr r10, T00_ADDR
  ldr r9, T03_ADDR
  ldr r8, T06_ADDR
  sub r4, r8, r11
/*(-- should do +0, +0tr, +1tr --)*/
  add r5, r14, r4
  add r5, r3
  S32_4 r5
  ldr r11, [r1, #10*4]
  add r5, r11
  str r5, [r0, #9*4]
/*=== W[12] ===*/
  ldr r11, T09_ADDR
/*(-- should do +2tr, -0tr, -9 --)*/
  sub r5, r2, r4
  sub r5, r11
  S32_2 r7 r5
  ldr r5, [r1, #13*4]
  add r7, r5
  str r7, [r0, #12*4]
/*=== W[15] ===*/
  ldr r14, T04_ADDR
  ldr r12, T06_ADDR
  ldr r10, T09_ADDR
  ldr r9, T12_ADDR
  ldr r7, T13_ADDR
  sub r4, r9, r11
  sub r3, r7, r12
/*(-- should do +0tr, +1tr, -4 --)*/
  add r5, r4, r3
  sub r5, r14
  S32_0 r7 r5
  ldr r5, [r1, #0*4]
  add r7, r5
  str r7, [r0, #15*4]
/*=== W[ 2] ===*/
  ldr r14, T00_ADDR
  ldr r7, T07_ADDR
  ldr r6, T15_ADDR
/*(-- should do +0, +15, +7, -0tr --)*/
  add r5, r14, r6
  add r5, r7
  sub r5, r4
  S32_2 r7 r5
  ldr r5, [r1, #3*4]
  add r7, r5
  str r7, [r0, #2*4]
/*=== W[ 5] ===*/
  ldr r14, T02_ADDR
  ldr r12, T03_ADDR
  ldr r11, T10_ADDR
  ldr r8, T12_ADDR
  ldr r7, T15_ADDR
  sub r4, r7, r14
  sub r2, r12, r9
/*(-- should do +0tr, +10, +2tr --)*/
  add r5, r4, r11
  add r5, r2
  S32_0 r12 r5
  ldr r5, [r1, #6*4]
  add r12, r5
  str r12, [r0, #5*4]
/*=== W[ 8] ===*/
  ldr r12, T05_ADDR
/*(-- should do +1tr, -0tr, -5 --)*/
  sub r5, r3, r4
  sub r5, r12
  S32_3 r11 r5
  ldr r5, [r1, #9*4]
  add r11, r5
  str r11, [r0, #8*4]
/*=== W[11] ===*/
  ldr r11, T00_ADDR
  ldr r9, T02_ADDR
  ldr r8, T05_ADDR
  ldr r7, T08_ADDR
  ldr r6, T09_ADDR
  sub r4, r7, r12
/*(-- should do +0tr, +9, -0, -2 --)*/
  add r5, r4, r10
  sub r5, r11
  sub r5, r14
  S32_1 r11 r5
  ldr r5, [r1, #12*4]
  add r11, r5
  str r11, [r0, #11*4]
/*=== W[14] ===*/
  ldr r11, T11_ADDR
/*(-- should do +0tr, +2tr, -11 --)*/
  add r5, r4, r2
  sub r5, r11
  S32_4 r5
  ldr r7, [r1, #15*4]
  add r5, r7
  str r5, [r0, #14*4]
/*=== W[ 1] ===*/
  ldr r14, T06_ADDR
  ldr r12, T08_ADDR
  ldr r10, T11_ADDR
  ldr r8, T14_ADDR
  ldr r7, T15_ADDR
  add r4, r11, r8
  add r3, r12, r7
/*(-- should do +0tr, +6, -1tr --)*/
  add r5, r4, r14
  sub r5, r3
  S32_1 r14 r5
  ldr r5, [r1, #2*4]
  add r14, r5
  str r14, [r0, #1*4]
/*=== W[ 4] ===*/
  ldr r14, T01_ADDR
  ldr r12, T02_ADDR
  ldr r7, T09_ADDR
/*(-- should do +1, +2, +9, -0tr --)*/
  add r5, r14, r12
  add r5, r7
  sub r5, r4
  S32_4 r5
  ldr r8, [r1, #5*4]
  add r5, r8
  str r5, [r0, #4*4]
/*=== W[ 7] ===*/
  ldr r12, T01_ADDR
  ldr r11, T04_ADDR
  ldr r8, T05_ADDR
  ldr r7, T12_ADDR
  ldr r6, T14_ADDR
/*(-- should do +1, -12, -14, -4, -5 --)*/
  sub r5, r14, r7
  sub r5, r6
  sub r5, r11
  sub r5, r8
  S32_2 r14 r5
  ldr r5, [r1, #8*4]
  add r14, r5
  str r14, [r0, #7*4]
/*=== W[10] ===*/
  ldr r14, T01_ADDR
  ldr r8, T04_ADDR
  ldr r7, T07_ADDR
  add r4, r11, r7
/*(-- should do +1tr, -0tr, -1 --)*/
  sub r5, r3, r4
  sub r5, r14
  S32_0 r14 r5
  ldr r5, [r1, #11*4]
  add r14, r5
  str r14, [r0, #10*4]
/*=== W[13] ===*/
  ldr r14, T02_ADDR
  ldr r12, T10_ADDR
  ldr r11, T11_ADDR
/*(-- should do +0tr, +10, +11, +2 --)*/
  add r5, r4, r12
  add r5, r11
  add r5, r14
  S32_3 r14 r5
  ldr r5, [r1, #14*4]
  add r14, r5
  str r14, [r0, #13*4]