5.txt

rollback_kernel_1 (32 x 32 x 1, warp i):
  v[o][j][i] = 0.5 * myVarY[i][j] * myDyy[i][0..2] `dot` myResult[o][i][j-1..j+1]
  u[o][j][i] = dtInv * myResult[o][i][j] +
    0.5 * 0.5 * myVarX[i][j] * myDxx[0..2][i] `dot` myResult[o][i-1..i+1][j] +
    v[o][j][i]
  a[j][i] =       - 0.5 * 0.5 * myVarX[j][i] * myDxx[0][i]
  b[j][i] = dtInv - 0.5 * 0.5 * myVarX[j][i] * myDxx[1][i]
  c[j][i] =       - 0.5 * 0.5 * myVarX[j][i] * myDxx[2][i]

tridag_kernel_0 (1 x 32 x 1, warp j):
  yy[j][i] = -a[j][i] * c[j][i-1]

tridag_kernel_1 (1 x 32 x 1, warp j):
  yy[j][0] = 1.0 / b[j][0]
  for i = 1..N:
    yy[j][i] = 1.0 / (b[j][i] - yy[j][i] * yy[j][i-1])

tridag_kernel_2 (64 x 1 x 1, warp i):
  a[j][i] = 1.0 / (c[j][i-1] * yy[j][i-1] - b[j][i] / a[j][i])
  b[j][i] = - c[j][i] * yy[j][i]
  u[o][j][i] = u[o][j][i] * yy[j][i]

tridag_kernel_3 (1 x 32 x 1, warp j):
  loop i = 1..N:
    u[o][j][i] += a[j][i] * u[o][j][i-1]
  loop i = N-1..0:
    u[o][j][i] += b[j][i] * u[o][j][i+1]

rollback_kernel_4 (1 x 32 x 1, warp j):
  a[i][j] =       - 0.5 * 0.5 * myVarY[i][j] * myDyy[0][j]
  b[i][j] = dtInv - 0.5 * 0.5 * myVarY[i][j] * myDyy[1][j]
  c[i][j] =       - 0.5 * 0.5 * myVarY[i][j] * myDyy[2][j]
  y[o][i][j] = dtInv * u[o][j][i] - 0.5 * v[o][j][i]

tridag_kernel_4 (1 x 32 x 1, warp j):
  yy[i][j] = -a[i][j] * c[i][j-1]

tridag_kernel_5 (32 x 1 x 1, warp i):
  yy[i][0] = 1.0 / b[i][0]
  for j = 1..N:
    yy[i][j] = 1.0 / (b[i][j] - yy[i][j] * yy[i][j-1])

tridag_kernel_6 (1 x 32 x 1, warp j):
  a[i][j] = 1.0 / (c[i][j-1] * yy[i][j-1] - b[i][j] / a[i][j])
  b[i][j] = - c[i][j] * yy[i][j]
  myResults[o][i][j] =   y[o][i][j] * yy[i][j]

tridag_kernel_7 (32 x 1 x 1, warp i):
  loop j = 1..N:
    myResults[o][i][j] += a[i][j] * myResults[o][i][j-1]
  loop j = N-1..0:
    myResults[o][i][j] += b[i][j] * myResults[o][i][j+1]