dct8x8

def  DCT8x8(data):
    tmp = np.zeros((64,),dtype=np.int32)
    result = np.zeros((64,),dtype=np.int32)
    for i in range(8):
        p0 = data[i * 8 + 0]
        p1 = data[i * 8 + 1]
        p2 = data[i * 8 + 2]
        p3 = data[i * 8 + 3]
        p4 = data[i * 8 + 4]
        p5 = data[i * 8 + 5]
        p6 = data[i * 8 + 6]
        p7 = data[i * 8 + 7]
        a0 = p0 + p7
        a1 = p1 + p6
        a2 = p2 + p5
        a3 = p3 + p4
        b0 = a0 + a3
        b1 = a1 + a2
        b2 = a0 - a3
        b3 = a1 - a2
        a0 = p0 - p7
        a1 = p1 - p6
        a2 = p2 - p5
        a3 = p3 - p4
        b4 = a1 + a2 + ((a0 >> 1) + a0)
        b5 = a0 - a3 - ((a2 >> 1) + a2)
        b6 = a0 + a3 - ((a1 >> 1) + a1)
        b7 = a1 - a2 + ((a3 >> 1) + a3)
        tmp[i * 8 + 0] = b0 + b1
        tmp[i * 8 + 1] = b4 + (b7 >> 2)
        tmp[i * 8 + 2] = b2 + (b3 >> 1)
        tmp[i * 8 + 3] = b5 + (b6 >> 2)
        tmp[i * 8 + 4] = b0 - b1
        tmp[i * 8 + 5] = b6 - (b5 >> 2)
        tmp[i * 8 + 6] = (b2 >> 1) - b3
        tmp[i * 8 + 7] = (b4 >> 2) - b7
    for i in range(8):
        p0 = tmp[0 * 8 + i]
        p1 = tmp[1 * 8 + i]
        p2 = tmp[2 * 8 + i]
        p3 = tmp[3 * 8 + i]
        p4 = tmp[4 * 8 + i]
        p5 = tmp[5 * 8 + i]
        p6 = tmp[6 * 8 + i]
        p7 = tmp[7 * 8 + i]
        a0 = p0 + p7
        a1 = p1 + p6
        a2 = p2 + p5
        a3 = p3 + p4
        b0 = a0 + a3
        b1 = a1 + a2
        b2 = a0 - a3
        b3 = a1 - a2
        a0 = p0 - p7
        a1 = p1 - p6
        a2 = p2 - p5
        a3 = p3 - p4
        b4 = a1 + a2 + ((a0 >> 1) + a0)
        b5 = a0 - a3 - ((a2 >> 1) + a2)
        b6 = a0 + a3 - ((a1 >> 1) + a1)
        b7 = a1 - a2 + ((a3 >> 1) + a3)
        result[0 * 8 + i] = b0 + b1
        result[1 * 8 + i] = b4 + (b7 >> 2)
        result[2 * 8 + i] = b2 + (b3 >> 1)
        result[3 * 8 + i] = b5 + (b6 >> 2)
        result[4 * 8 + i] = b0 - b1
        result[5 * 8 + i] = b6 - (b5 >> 2)
        result[6 * 8 + i] = (b2 >> 1) - b3
        result[7 * 8 + i] = (b4 >> 2) - b7
    return result

def IDCT8x8(data):
    tmp = np.zeros((64,), dtype=np.int32)
    result = np.zeros((64,), dtype=np.int32)
    for i in range(8):
        p0 = data[i * 8 + 0]
        p1 = data[i * 8 + 1]
        p2 = data[i * 8 + 2]
        p3 = data[i * 8 + 3]
        p4 = data[i * 8 + 4]
        p5 = data[i * 8 + 5]
        p6 = data[i * 8 + 6]
        p7 = data[i * 8 + 7]
        a0 = p0 + p4
        a1 = p0 - p4
        a2 = p6 - (p2 >> 1)
        a3 = p2 + (p6 >> 1)
        b0 = a0 + a3
        b2 = a1 - a2
        b4 = a1 + a2
        b6 = a0 - a3
        a0 = -p3 + p5 - p7 - (p7 >> 1)
        a1 = p1 + p7 - p3 - (p3 >> 1)
        a2 = -p1 + p7 + p5 + (p5 >> 1)
        a3 = p3 + p5 + p1 + (p1 >> 1)
        b1 = a0 + (a3 >> 2)
        b3 = a1 + (a2 >> 2)
        b5 = a2 - (a1 >> 2)
        b7 = a3 - (a0 >> 2)
        tmp[i * 8 + 0] = b0 + b7
        tmp[i * 8 + 1] = b2 - b5
        tmp[i * 8 + 2] = b4 + b3
        tmp[i * 8 + 3] = b6 + b1
        tmp[i * 8 + 4] = b6 - b1
        tmp[i * 8 + 5] = b4 - b3
        tmp[i * 8 + 6] = b2 + b5
        tmp[i * 8 + 7] = b0 - b7
    for i in range(8):
        p0 = tmp[0 * 8 + i]
        p1 = tmp[1 * 8 + i]
        p2 = tmp[2 * 8 + i]
        p3 = tmp[3 * 8 + i]
        p4 = tmp[4 * 8 + i]
        p5 = tmp[5 * 8 + i]
        p6 = tmp[6 * 8 + i]
        p7 = tmp[7 * 8 + i]
        a0 = p0 + p4
        a1 = p0 - p4
        a2 = p6 - (p2 >> 1)
        a3 = p2 + (p6 >> 1)
        b0 = a0 + a3
        b2 = a1 - a2
        b4 = a1 + a2
        b6 = a0 - a3
        a0 = -p3 + p5 - p7 - (p7 >> 1)
        a1 = p1 + p7 - p3 - (p3 >> 1)
        a2 = -p1 + p7 + p5 + (p5 >> 1)
        a3 = p3 + p5 + p1 + (p1 >> 1)
        b1 = a0 + (a3 >> 2)
        b7 = a3 - (a0 >> 2)
        b3 = a1 + (a2 >> 2)
        b5 = a2 - (a1 >> 2)
        result[0 * 8 + i] = ((b0 + b7) >> 6)
        result[1 * 8 + i] = ((b2 - b5) >> 6)
        result[2 * 8 + i] = ((b4 + b3) >> 6)
        result[3 * 8 + i] = ((b6 + b1) >> 6)
        result[4 * 8 + i] = ((b6 - b1) >> 6)
        result[5 * 8 + i] = ((b4 - b3) >> 6)
        result[6 * 8 + i] = ((b2 + b5) >> 6)
        result[7 * 8 + i] = ((b0 - b7) >> 6)
    result = result.clip(0,255)
    return result