stack32 out_stack
int32 out
stack32 u_stack
int32 u
int32 eax
int32 ebx
int32 esi
int32 edi
int32 ebp
stack32 ebp_stack
float80 d0
float80 d1
float80 d2
float80 d3
float80 d4
float80 e0
float80 e1
float80 e2
float80 e3
float80 e4
float80 f0
float80 f1
float80 f2
float80 f3
float80 f4
float80 g0
float80 g1
float80 g2
float80 g3
float80 g4
float80 h0
float80 h1
float80 h2
float80 h3
float80 h4
float80 i0
float80 i1
float80 i2
float80 i3
float80 i4
float80 j0
float80 j1
float80 j2
float80 j3
float80 j4
float80 k0
float80 k1
float80 k2
float80 k3
float80 k4

enter gaudry_pm_hadamard
input out_stack
input u_stack

caller eax
caller ebx
caller esi
caller edi
caller ebp

  u = u_stack
  out = out_stack

  f0 = *(float64 *) (u + 0)
  d0 = *(float64 *) (u + 40)
  g0 = *(float64 *) (u + 80)
  e0 = *(float64 *) (u + 120)
  d0 += f0
  e0 += g0
  f0 -= *(float64 *) (u + 40)
  g0 -= *(float64 *) (u + 120)
    f1 = *(float64 *) (u + 8)
    d1 = *(float64 *) (u + 48)
    g1 = *(float64 *) (u + 88)
    e1 = *(float64 *) (u + 128)
    d1 += f1
    e1 += g1
    f1 -= *(float64 *) (u + 48)
    g1 -= *(float64 *) (u + 128)
  *(float64 *) (out + 0) = d0
  *(float64 *) (out + 40) = e0
  *(float64 *) (out + 80) = f0
  *(float64 *) (out + 120) = g0
    *(float64 *) (out + 8) = d1
    *(float64 *) (out + 48) = e1
    *(float64 *) (out + 88) = f1
    *(float64 *) (out + 128) = g1

      f2 = *(float64 *) (u + 16)
      d2 = *(float64 *) (u + 56)
      g2 = *(float64 *) (u + 96)
      e2 = *(float64 *) (u + 136)
      d2 += f2
      e2 += g2
      f2 -= *(float64 *) (u + 56)
      g2 -= *(float64 *) (u + 136)
        f3 = *(float64 *) (u + 24)
        d3 = *(float64 *) (u + 64)
        g3 = *(float64 *) (u + 104)
        e3 = *(float64 *) (u + 144)
        d3 += f3
        e3 += g3
        f3 -= *(float64 *) (u + 64)
        g3 -= *(float64 *) (u + 144)
      *(float64 *) (out + 16) = d2
      *(float64 *) (out + 56) = e2
      *(float64 *) (out + 96) = f2
      *(float64 *) (out + 136) = g2
        *(float64 *) (out + 24) = d3
        *(float64 *) (out + 64) = e3
        *(float64 *) (out + 104) = f3
        *(float64 *) (out + 144) = g3

          f4 = *(float64 *) (u + 32)
          d4 = *(float64 *) (u + 72)
          g4 = *(float64 *) (u + 112)
          e4 = *(float64 *) (u + 152)
          d4 += f4
          e4 += g4
          f4 -= *(float64 *) (u + 72)
          g4 -= *(float64 *) (u + 152)

  j0 = *(float64 *) (out + 0)
  h0 = *(float64 *) (out + 40)
  k0 = *(float64 *) (out + 80)
  i0 = *(float64 *) (out + 120)
  h0 += j0
  i0 += k0
  j0 -= *(float64 *) (out + 40)
  k0 -= *(float64 *) (out + 120)
          *(float64 *) (out + 32) = d4
          *(float64 *) (out + 72) = e4
          *(float64 *) (out + 112) = f4
          *(float64 *) (out + 152) = g4
  *(float64 *) (out + 0) = h0
  *(float64 *) (out + 40) = j0
  *(float64 *) (out + 80) = i0
  *(float64 *) (out + 120) = k0

    j1 = *(float64 *) (out + 8)
    h1 = *(float64 *) (out + 48)
    k1 = *(float64 *) (out + 88)
    i1 = *(float64 *) (out + 128)
    h1 += j1
    i1 += k1
    j1 -= *(float64 *) (out + 48)
    k1 -= *(float64 *) (out + 128)

      j2 = *(float64 *) (out + 16)
      h2 = *(float64 *) (out + 56)
      k2 = *(float64 *) (out + 96)
      i2 = *(float64 *) (out + 136)
      h2 += j2
      i2 += k2
      j2 -= *(float64 *) (out + 56)
      k2 -= *(float64 *) (out + 136)
    *(float64 *) (out + 8) = h1
    *(float64 *) (out + 48) = j1
    *(float64 *) (out + 88) = i1
    *(float64 *) (out + 128) = k1
      *(float64 *) (out + 16) = h2
      *(float64 *) (out + 56) = j2
      *(float64 *) (out + 96) = i2
      *(float64 *) (out + 136) = k2

        j3 = *(float64 *) (out + 24)
        h3 = *(float64 *) (out + 64)
        k3 = *(float64 *) (out + 104)
        i3 = *(float64 *) (out + 144)
        h3 += j3
        i3 += k3
        j3 -= *(float64 *) (out + 64)
        k3 -= *(float64 *) (out + 144)

          j4 = *(float64 *) (out + 32)
          h4 = *(float64 *) (out + 72)
          k4 = *(float64 *) (out + 112)
          i4 = *(float64 *) (out + 152)
          h4 += j4
          i4 += k4
          j4 -= *(float64 *) (out + 72)
          k4 -= *(float64 *) (out + 152)
        *(float64 *) (out + 24) = h3
        *(float64 *) (out + 64) = j3
        *(float64 *) (out + 104) = i3
        *(float64 *) (out + 144) = k3
          *(float64 *) (out + 32) = h4
          *(float64 *) (out + 72) = j4
          *(float64 *) (out + 112) = i4
          *(float64 *) (out + 152) = k4


leave