transforms.py 40.1 KB
Newer Older
1 2 3 4 5
# Lossy compression algorithms very often make use of DCT or DFT calculations,
# or variations of these calculations. This file is intended to be a short
# reference about classical DCT and DFT algorithms.


6 7 8
from random import random
from math import pi, sin, cos, sqrt
from cmath import exp
9 10

def exp_j (alpha):
11
    return exp (alpha * 1j)
12 13

def conjugate (c):
14 15
    c = c + 0j
    return c.real - 1j * c.imag
16 17

def vector (N):
18
    return [0j] * N
19 20


21 22
# Let us start withthe canonical definition of the unscaled DFT algorithm :
# (I can not draw sigmas in a text file so I'll use python code instead)  :)
23 24 25 26 27 28 29 30

def W (k, N):
    return exp_j ((-2*pi*k)/N)

def unscaled_DFT (N, input, output):
    for o in range(N):		# o is output index
	output[o] = 0
	for i in range(N):
31
	    output[o] = output[o] + input[i] * W (i*o, N)
32 33

# This algorithm takes complex input and output. There are N*N complex
34
# multiplications and N*(N-1) complex additions.
35 36


37
# Of course this algorithm is an extremely naive implementation and there are
38
# some ways to use the trigonometric properties of the coefficients to find
39 40 41
# some decompositions that can accelerate the calculation by several orders
# of magnitude... This is a well known and studied problem. One of the
# available explanations of this process is at this url :
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
# www.cmlab.csie.ntu.edu.tw/cml/dsp/training/coding/transform/fft.html


# Let's start with the radix-2 decimation-in-time algorithm :

def unscaled_DFT_radix2_time (N, input, output):
    even_input = vector(N/2)
    odd_input = vector(N/2)
    even_output = vector(N/2)
    odd_output = vector(N/2)

    for i in range(N/2):
	even_input[i] = input[2*i]
	odd_input[i] = input[2*i+1]

    unscaled_DFT (N/2, even_input, even_output)
    unscaled_DFT (N/2, odd_input, odd_output)

    for i in range(N/2):
61
	odd_output[i] = odd_output[i] * W (i, N)
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79

    for i in range(N/2):
	output[i] = even_output[i] + odd_output[i]
	output[i+N/2] = even_output[i] - odd_output[i]

# This algorithm takes complex input and output.

# We divide the DFT calculation into 2 DFT calculations of size N/2
# We then do N/2 complex multiplies followed by N complex additions.
# (actually W(0,N) = 1 and W(N/4,N) = -j so we can skip a few of these complex
# multiplies... we will skip 1 for i=0 and 1 for i=N/4. Also for i=N/8 and for
# i=3*N/8 the W(i,N) values can be special-cased to implement the complex
# multiplication using only 2 real additions and 2 real multiplies)

# Also note that all the basic stages of this DFT algorithm are easily
# reversible, so we can calculate the IDFT with the same complexity.


80
# A variant of this is the radix-2 decimation-in-frequency algorithm :
81 82 83 84 85 86 87 88 89 90 91 92

def unscaled_DFT_radix2_freq (N, input, output):
    even_input = vector(N/2)
    odd_input = vector(N/2)
    even_output = vector(N/2)
    odd_output = vector(N/2)

    for i in range(N/2):
	even_input[i] = input[i] + input[i+N/2]
	odd_input[i] = input[i] - input[i+N/2]

    for i in range(N/2):
93
	odd_input[i] = odd_input[i] * W (i, N)
94 95 96 97 98 99 100 101 102 103 104 105

    unscaled_DFT (N/2, even_input, even_output)
    unscaled_DFT (N/2, odd_input, odd_output)

    for i in range(N/2):
	output[2*i] = even_output[i]
	output[2*i+1] = odd_output[i]

# Note that the decimation-in-time and the decimation-in-frequency varients
# have exactly the same complexity, they only do the operations in a different
# order.

106
# Actually, if you look at the decimation-in-time variant of the DFT, and
107 108 109 110 111 112
# reverse it to calculate an IDFT, you get something that is extremely close
# to the decimation-in-frequency DFT algorithm...


# The radix-4 algorithms are slightly more efficient : they take into account
# the fact that with complex numbers, multiplications by j and -j are also
113 114
# "free"... i.e. when you code them using real numbers, they translate into
# a few data moves but no real operation.
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143

# Let's start with the radix-4 decimation-in-time algorithm :

def unscaled_DFT_radix4_time (N, input, output):
    input_0 = vector(N/4)
    input_1 = vector(N/4)
    input_2 = vector(N/4)
    input_3 = vector(N/4)
    output_0 = vector(N/4)
    output_1 = vector(N/4)
    output_2 = vector(N/4)
    output_3 = vector(N/4)
    tmp_0 = vector(N/4)
    tmp_1 = vector(N/4)
    tmp_2 = vector(N/4)
    tmp_3 = vector(N/4)

    for i in range(N/4):
	input_0[i] = input[4*i]
	input_1[i] = input[4*i+1]
	input_2[i] = input[4*i+2]
	input_3[i] = input[4*i+3]

    unscaled_DFT (N/4, input_0, output_0)
    unscaled_DFT (N/4, input_1, output_1)
    unscaled_DFT (N/4, input_2, output_2)
    unscaled_DFT (N/4, input_3, output_3)

    for i in range(N/4):
144 145 146
	output_1[i] = output_1[i] * W (i, N)
	output_2[i] = output_2[i] * W (2*i, N)
	output_3[i] = output_3[i] * W (3*i, N)
147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206

    for i in range(N/4):
	tmp_0[i] = output_0[i] + output_2[i]
	tmp_1[i] = output_0[i] - output_2[i]
	tmp_2[i] = output_1[i] + output_3[i]
	tmp_3[i] = output_1[i] - output_3[i]

    for i in range(N/4):
	output[i] = tmp_0[i] + tmp_2[i]
	output[i+N/4] = tmp_1[i] - 1j * tmp_3[i]
	output[i+N/2] = tmp_0[i] - tmp_2[i]
	output[i+3*N/4] = tmp_1[i] + 1j * tmp_3[i]

# This algorithm takes complex input and output.

# We divide the DFT calculation into 4 DFT calculations of size N/4
# We then do 3*N/4 complex multiplies followed by 2*N complex additions.
# (actually W(0,N) = 1 and W(N/4,N) = -j so we can skip a few of these complex
# multiplies... we will skip 3 for i=0 and 1 for i=N/8. Also for i=N/8
# the remaining W(i,N) and W(3*i,N) multiplies can be implemented using only
# 2 real additions and 2 real multiplies. For i=N/16 and i=3*N/16 we can also
# optimise the W(2*i/N) multiply this way.

# If we wanted to do the same decomposition with one radix-2 decomposition
# of size N and 2 radix-2 decompositions of size N/2, the total cost would be
# N complex multiplies and 2*N complex additions. Thus we see that the
# decomposition of one DFT calculation of size N into 4 calculations of size
# N/4 using the radix-4 algorithm instead of the radix-2 algorithm saved N/4
# complex multiplies...


# The radix-4 decimation-in-frequency algorithm is similar :

def unscaled_DFT_radix4_freq (N, input, output):
    input_0 = vector(N/4)
    input_1 = vector(N/4)
    input_2 = vector(N/4)
    input_3 = vector(N/4)
    output_0 = vector(N/4)
    output_1 = vector(N/4)
    output_2 = vector(N/4)
    output_3 = vector(N/4)
    tmp_0 = vector(N/4)
    tmp_1 = vector(N/4)
    tmp_2 = vector(N/4)
    tmp_3 = vector(N/4)

    for i in range(N/4):
	tmp_0[i] = input[i] + input[i+N/2]
	tmp_1[i] = input[i+N/4] + input[i+3*N/4]
	tmp_2[i] = input[i] - input[i+N/2]
	tmp_3[i] = input[i+N/4] - input[i+3*N/4]

    for i in range(N/4):
	input_0[i] = tmp_0[i] + tmp_1[i]
	input_1[i] = tmp_2[i] - 1j * tmp_3[i]
	input_2[i] = tmp_0[i] - tmp_1[i]
	input_3[i] = tmp_2[i] + 1j * tmp_3[i]

    for i in range(N/4):
207 208 209
	input_1[i] = input_1[i] * W (i, N)
	input_2[i] = input_2[i] * W (2*i, N)
	input_3[i] = input_3[i] * W (3*i, N)
210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250

    unscaled_DFT (N/4, input_0, output_0)
    unscaled_DFT (N/4, input_1, output_1)
    unscaled_DFT (N/4, input_2, output_2)
    unscaled_DFT (N/4, input_3, output_3)

    for i in range(N/4):
	output[4*i] = output_0[i]
	output[4*i+1] = output_1[i]
	output[4*i+2] = output_2[i]
	output[4*i+3] = output_3[i]

# Once again the complexity is exactly the same as for the radix-4
# decimation-in-time DFT algorithm, only the order of the operations is
# different.


# Now let us reorder the radix-4 algorithms in a different way :

#def unscaled_DFT_radix4_time (N, input, output):
#   input_0 = vector(N/4)
#   input_1 = vector(N/4)
#   input_2 = vector(N/4)
#   input_3 = vector(N/4)
#   output_0 = vector(N/4)
#   output_1 = vector(N/4)
#   output_2 = vector(N/4)
#   output_3 = vector(N/4)
#   tmp_0 = vector(N/4)
#   tmp_1 = vector(N/4)
#   tmp_2 = vector(N/4)
#   tmp_3 = vector(N/4)
#
#   for i in range(N/4):
#	input_0[i] = input[4*i]
#	input_2[i] = input[4*i+2]
#
#   unscaled_DFT (N/4, input_0, output_0)
#   unscaled_DFT (N/4, input_2, output_2)
#
#   for i in range(N/4):
251
#	output_2[i] = output_2[i] * W (2*i, N)
252 253 254 255 256 257 258 259 260 261 262 263 264
#
#   for i in range(N/4):
#	tmp_0[i] = output_0[i] + output_2[i]
#	tmp_1[i] = output_0[i] - output_2[i]
#
#   for i in range(N/4):
#	input_1[i] = input[4*i+1]
#	input_3[i] = input[4*i+3]
#
#   unscaled_DFT (N/4, input_1, output_1)
#   unscaled_DFT (N/4, input_3, output_3)
#
#   for i in range(N/4):
265 266
#	output_1[i] = output_1[i] * W (i, N)
#	output_3[i] = output_3[i] * W (3*i, N)
267 268 269 270 271 272 273 274 275 276 277
#
#   for i in range(N/4):
#	tmp_2[i] = output_1[i] + output_3[i]
#	tmp_3[i] = output_1[i] - output_3[i]
#
#   for i in range(N/4):
#	output[i] = tmp_0[i] + tmp_2[i]
#	output[i+N/4] = tmp_1[i] - 1j * tmp_3[i]
#	output[i+N/2] = tmp_0[i] - tmp_2[i]
#	output[i+3*N/4] = tmp_1[i] + 1j * tmp_3[i]

278
# We didn't do anything here, only reorder the operations. But now, look at the
279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311
# first part of this function, up to the calculations of tmp0 and tmp1 : this
# is extremely similar to the radix-2 decimation-in-time algorithm ! or more
# precisely, it IS the radix-2 decimation-in-time algorithm, with size N/2,
# applied on a vector representing the even input coefficients, and giving
# an output vector that is the concatenation of tmp0 and tmp1.
# This is important to notice, because this means we can now choose to
# calculate tmp0 and tmp1 using any DFT algorithm that we want, and we know
# that some of them are more efficient than radix-2...

# This leads us directly to the split-radix decimation-in-time algorithm :

def unscaled_DFT_split_radix_time (N, input, output):
    even_input = vector(N/2)
    input_1 = vector(N/4)
    input_3 = vector(N/4)
    even_output = vector(N/2)
    output_1 = vector(N/4)
    output_3 = vector(N/4)
    tmp_0 = vector(N/4)
    tmp_1 = vector(N/4)

    for i in range(N/2):
	even_input[i] = input[2*i]

    for i in range(N/4):
	input_1[i] = input[4*i+1]
	input_3[i] = input[4*i+3]

    unscaled_DFT (N/2, even_input, even_output)
    unscaled_DFT (N/4, input_1, output_1)
    unscaled_DFT (N/4, input_3, output_3)

    for i in range(N/4):
312 313
	output_1[i] = output_1[i] * W (i, N)
	output_3[i] = output_3[i] * W (3*i, N)
314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356

    for i in range(N/4):
	tmp_0[i] = output_1[i] + output_3[i]
	tmp_1[i] = output_1[i] - output_3[i]

    for i in range(N/4):
	output[i] = even_output[i] + tmp_0[i]
	output[i+N/4] = even_output[i+N/4] - 1j * tmp_1[i]
	output[i+N/2] = even_output[i] - tmp_0[i]
	output[i+3*N/4] = even_output[i+N/4] + 1j * tmp_1[i]

# This function performs one DFT of size N/2 and two of size N/4, followed by
# N/2 complex multiplies and 3*N/2 complex additions.
# (actually W(0,N) = 1 and W(N/4,N) = -j so we can skip a few of these complex
# multiplies... we will skip 2 for i=0. Also for i=N/8 the W(i,N) and W(3*i,N)
# multiplies can be implemented using only 2 real additions and 2 real
# multiplies)


# We can similarly define the split-radix decimation-in-frequency DFT :

def unscaled_DFT_split_radix_freq (N, input, output):
    even_input = vector(N/2)
    input_1 = vector(N/4)
    input_3 = vector(N/4)
    even_output = vector(N/2)
    output_1 = vector(N/4)
    output_3 = vector(N/4)
    tmp_0 = vector(N/4)
    tmp_1 = vector(N/4)

    for i in range(N/2):
	even_input[i] = input[i] + input[i+N/2]

    for i in range(N/4):
	tmp_0[i] = input[i] - input[i+N/2]
	tmp_1[i] = input[i+N/4] - input[i+3*N/4]

    for i in range(N/4):
	input_1[i] = tmp_0[i] - 1j * tmp_1[i]
	input_3[i] = tmp_0[i] + 1j * tmp_1[i]

    for i in range(N/4):
357 358
	input_1[i] = input_1[i] * W (i, N)
	input_3[i] = input_3[i] * W (3*i, N)
359 360 361 362 363 364 365 366 367 368 369 370

    unscaled_DFT (N/2, even_input, even_output)
    unscaled_DFT (N/4, input_1, output_1)
    unscaled_DFT (N/4, input_3, output_3)

    for i in range(N/2):
	output[2*i] = even_output[i]

    for i in range(N/4):
	output[4*i+1] = output_1[i]
	output[4*i+3] = output_3[i]

371
# The complexity is again the same as for the decimation-in-time variant.
372 373 374 375 376 377 378 379


# Now let us now summarize our various algorithms for DFT decomposition :

# radix-2 : DFT(N) -> 2*DFT(N/2) using N/2 multiplies and N additions
# radix-4 : DFT(N) -> 4*DFT(N/2) using 3*N/4 multiplies and 2*N additions
# split-radix : DFT(N) -> DFT(N/2) + 2*DFT(N/4) using N/2 muls and 3*N/2 adds

380 381 382 383
# (we are always speaking of complex multiplies and complex additions... a
# complex addition is implemented with 2 real additions, and a complex
# multiply is implemented with either 2 adds and 4 muls or 3 adds and 3 muls,
# so we will keep a separate count of these)
384 385 386 387 388 389 390

# If we want to take into account the special values of W(i,N), we can remove
# a few complex multiplies. Supposing N>=16 we can remove :
# radix-2 : remove 2 complex multiplies, simplify 2 others
# radix-4 : remove 4 complex multiplies, simplify 4 others
# split-radix : remove 2 complex multiplies, simplify 2 others

391
# This gives the following table for the complexity of a complex DFT :
392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
#	N	real additions	real multiplies	complex multiplies
#       1               0               0               0
#       2               4               0               0
#       4              16               0               0
#       8              52               4               0
#      16             136               8               4
#      32             340              20              16
#      64             808              40              52
#     128            1876              84             144
#     256            4264             168             372
#     512            9556             340             912
#    1024           21160             680            2164
#    2048           46420            1364            5008
#    4096          101032            2728           11380
#    8192          218452            5460           25488
#   16384          469672           10920           56436
#   32768         1004884           21844          123792
#   65536         2140840           43688          269428

411 412 413
# If we chose to implement complex multiplies with 3 real muls + 3 real adds,
# then these results are consistent with the table at the end of the
# www.cmlab.csie.ntu.edu.tw DFT tutorial that I mentionned earlier.
414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467


# Now another important case for the DFT is the one where the inputs are
# real numbers instead of complex ones. We have to find ways to optimize for
# this important case.

# If the DFT inputs are real-valued, then the DFT outputs have nice properties
# too : output[0] and output[N/2] will be real numbers, and output[N-i] will
# be the conjugate of output[i] for i in 0...N/2-1

# Likewise, if the DFT inputs are purely imaginary numbers, then the DFT
# outputs will have special properties too : output[0] and output[N/2] will be
# purely imaginary, and output[N-i] will be the opposite of the conjugate of
# output[i] for i in 0...N/2-1

# We can use these properties to calculate two real-valued DFT at once :

def two_real_unscaled_DFT (N, input1, input2, output1, output2):
    input = vector(N)
    output = vector(N)

    for i in range(N):
	input[i] = input1[i] + 1j * input2[i]

    unscaled_DFT (N, input, output)

    output1[0] = output[0].real + 0j
    output2[0] = output[0].imag + 0j

    for i in range(N/2)[1:]:
	output1[i] = 0.5 * (output[i] + conjugate(output[N-i]))
	output2[i] = -0.5j * (output[i] - conjugate(output[N-i]))

	output1[N-i] = conjugate(output1[i])
	output2[N-i] = conjugate(output2[i])

    output1[N/2] = output[N/2].real + 0j
    output2[N/2] = output[N/2].imag + 0j

# This routine does a total of N-2 complex additions and N-2 complex
# multiplies by 0.5

# This routine can also be inverted to calculate the IDFT of two vectors at
# once if we know that the outputs will be real-valued.


# If we have only one real-valued DFT calculation to do, we can still cut this
# calculation in several parts using one of the decimate-in-time methods
# (so that the different parts are still real-valued)

# As with complex DFT calculations, the best method is to use a split radix.
# There are a lot of symetries in the DFT outputs that we can exploit to
# reduce the number of operations...

468
def real_unscaled_DFT_split_radix_time_1 (N, input, output):
469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503
    even_input = vector(N/2)
    even_output = vector(N/2)
    input_1 = vector(N/4)
    output_1 = vector(N/4)
    input_3 = vector(N/4)
    output_3 = vector(N/4)
    tmp_0 = vector(N/4)
    tmp_1 = vector(N/4)

    for i in range(N/2):
	even_input[i] = input[2*i]

    for i in range(N/4):
	input_1[i] = input[4*i+1]
	input_3[i] = input[4*i+3]

    unscaled_DFT (N/2, even_input, even_output)
    # this is again a real DFT !
    # we will only use even_output[i] for i in 0 ... N/4 included. we know that
    # even_output[N/2-i] is the conjugate of even_output[i]... also we know
    # that even_output[0] and even_output[N/4] are purely real.

    unscaled_DFT (N/4, input_1, output_1)
    unscaled_DFT (N/4, input_3, output_3)
    # these are real DFTs too... with symetries in the outputs... once again

    tmp_0[0] = output_1[0] + output_3[0]	# real numbers
    tmp_1[0] = output_1[0] - output_3[0]	# real numbers

    tmp__0 = (output_1[N/8] + output_3[N/8]) * sqrt(0.5)	# real numbers
    tmp__1 = (output_1[N/8] - output_3[N/8]) * sqrt(0.5)	# real numbers
    tmp_0[N/8] = tmp__1 - 1j * tmp__0		# real + 1j * real
    tmp_1[N/8] = tmp__0 - 1j * tmp__1		# real + 1j * real

    for i in range(N/8)[1:]:
504 505
	output_1[i] = output_1[i] * W (i, N)
	output_3[i] = output_3[i] * W (3*i, N)
506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532

	tmp_0[i] = output_1[i] + output_3[i]
	tmp_1[i] = output_1[i] - output_3[i]

	tmp_0[N/4-i] = -1j * conjugate(tmp_1[i])
	tmp_1[N/4-i] = -1j * conjugate(tmp_0[i])

    output[0] = even_output[0] + tmp_0[0]		# real numbers
    output[N/4] = even_output[N/4] - 1j * tmp_1[0]	# real + 1j * real
    output[N/2] = even_output[0] - tmp_0[0]		# real numbers
    output[3*N/4] = even_output[N/4] + 1j * tmp_1[0]	# real + 1j * real

    for i in range(N/4)[1:]:
	output[i] = even_output[i] + tmp_0[i]
	output[i+N/4] = conjugate(even_output[N/4-i]) - 1j * tmp_1[i]

	output[N-i] = conjugate(output[i])
	output[3*N/4-i] = conjugate(output[i+N/4])

# This function performs one real DFT of size N/2 and two real DFT of size
# N/4, followed by 6 real additions, 2 real multiplies, 3*N/4-4 complex
# additions and N/4-2 complex multiplies.


# We can also try to combine the two real DFT of size N/4 into a single complex
# DFT :

533
def real_unscaled_DFT_split_radix_time_2 (N, input, output):
534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571
    even_input = vector(N/2)
    even_output = vector(N/2)
    odd_input = vector(N/4)
    odd_output = vector(N/4)
    tmp_0 = vector(N/4)
    tmp_1 = vector(N/4)

    for i in range(N/2):
	even_input[i] = input[2*i]

    for i in range(N/4):
	odd_input[i] = input[4*i+1] + 1j * input[4*i+3]

    unscaled_DFT (N/2, even_input, even_output)
    # this is again a real DFT !
    # we will only use even_output[i] for i in 0 ... N/4 included. we know that
    # even_output[N/2-i] is the conjugate of even_output[i]... also we know
    # that even_output[0] and even_output[N/4] are purely real.

    unscaled_DFT (N/4, odd_input, odd_output)
    # but this one is a complex DFT so no special properties here

    output_1 = odd_output[0].real
    output_3 = odd_output[0].imag
    tmp_0[0] = output_1 + output_3	# real numbers
    tmp_1[0] = output_1 - output_3	# real numbers

    output_1 = odd_output[N/8].real
    output_3 = odd_output[N/8].imag
    tmp__0 = (output_1 + output_3) * sqrt(0.5)	# real numbers
    tmp__1 = (output_1 - output_3) * sqrt(0.5)	# real numbers
    tmp_0[N/8] = tmp__1 - 1j * tmp__0		# real + 1j * real
    tmp_1[N/8] = tmp__0 - 1j * tmp__1		# real + 1j * real

    for i in range(N/8)[1:]:
	output_1 = odd_output[i] + conjugate(odd_output[N/4-i])
	output_3 = odd_output[i] - conjugate(odd_output[N/4-i])

572 573
	output_1 = output_1 * 0.5 * W (i, N)
	output_3 = output_3 * -0.5j * W (3*i, N)
574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603

	tmp_0[i] = output_1 + output_3
	tmp_1[i] = output_1 - output_3

	tmp_0[N/4-i] = -1j * conjugate(tmp_1[i])
	tmp_1[N/4-i] = -1j * conjugate(tmp_0[i])

    output[0] = even_output[0] + tmp_0[0]		# real numbers
    output[N/4] = even_output[N/4] - 1j * tmp_1[0]	# real + 1j * real
    output[N/2] = even_output[0] - tmp_0[0]		# real numbers
    output[3*N/4] = even_output[N/4] + 1j * tmp_1[0]	# real + 1j * real

    for i in range(N/4)[1:]:
	output[i] = even_output[i] + tmp_0[i]
	output[i+N/4] = conjugate(even_output[N/4-i]) - 1j * tmp_1[i]

	output[N-i] = conjugate(output[i])
	output[3*N/4-i] = conjugate(output[i+N/4])

# This function performs one real DFT of size N/2 and one complex DFT of size
# N/4, followed by 6 real additions, 2 real multiplies, N-6 complex additions
# and N/4-2 complex multiplies.

# After comparing the performance, it turns out that for real-valued DFT, the
# version of the algorithm that subdivides the calculation into one real
# DFT of size N/2 and two real DFT of size N/4 is the most efficient one.
# The other version gives exactly the same number of multiplies and a few more
# real additions.


604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656
# Now we can also try the decimate-in-frequency method for a real-valued DFT.
# Using the split-radix algorithm, and by taking into account the symetries of
# the outputs :

def real_unscaled_DFT_split_radix_freq (N, input, output):
    even_input = vector(N/2)
    input_1 = vector(N/4)
    even_output = vector(N/2)
    output_1 = vector(N/4)
    tmp_0 = vector(N/4)
    tmp_1 = vector(N/4)

    for i in range(N/2):
	even_input[i] = input[i] + input[i+N/2]

    for i in range(N/4):
	tmp_0[i] = input[i] - input[i+N/2]
	tmp_1[i] = input[i+N/4] - input[i+3*N/4]

    for i in range(N/4):
	input_1[i] = tmp_0[i] - 1j * tmp_1[i]

    for i in range(N/4):
	input_1[i] = input_1[i] * W (i, N)

    unscaled_DFT (N/2, even_input, even_output)
    # This is still a real-valued DFT

    unscaled_DFT (N/4, input_1, output_1)
    # But that one is a complex-valued DFT

    for i in range(N/2):
	output[2*i] = even_output[i]

    for i in range(N/4):
	output[4*i+1] = output_1[i]
	output[N-1-4*i] = conjugate(output_1[i])

# I think this implementation is much more elegant than the decimate-in-time
# version ! It looks very much like the complex-valued version, all we had to
# do was remove one of the complex-valued internal DFT calls because we could
# deduce the outputs by using the symetries of the problem.

# As for performance, we did N real additions, N/4 complex multiplies (a bit
# less actually, because W(0,N) = 1 and W(N/8,N) is a "simple" multiply), then
# one real DFT of size N/2 and one complex DFT of size N/4.

# It turns out that even if the methods are so different, the number of
# operations is exactly the same as for the best of the two decimation-in-time
# methods that we tried.


# This gives us the following performance for real-valued DFT :
657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675
#	N	real additions	real multiplies	complex multiplies
#       2               2               0               0
#       4               6               0               0
#       8              20               2               0
#      16              54               4               2
#      32             140              10               8
#      64             342              20              26
#     128             812              42              72
#     256            1878              84             186
#     512            4268             170             456
#    1024            9558             340            1082
#    2048           21164             682            2504
#    4096           46422            1364            5690
#    8192          101036            2730           12744
#   16384          218454            5460           28218
#   32768          469676           10922           61896
#   65536         1004886           21844          134714


676
# As an example, this is an implementation of the real-valued DFT8 :
677 678

def DFT8 (input, output):
679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705
    even_0 = input[0] + input[4]
    even_1 = input[1] + input[5]
    even_2 = input[2] + input[6]
    even_3 = input[3] + input[7]

    tmp_0 = even_0 + even_2
    tmp_1 = even_0 - even_2
    tmp_2 = even_1 + even_3
    tmp_3 = even_1 - even_3

    output[0] = tmp_0 + tmp_2
    output[2] = tmp_1 - 1j * tmp_3
    output[4] = tmp_0 - tmp_2

    odd_0_r = input[0] - input[4]
    odd_0_i = input[2] - input[6]

    tmp_0 = input[1] - input[5]
    tmp_1 = input[3] - input[7]
    odd_1_r = (tmp_0 - tmp_1) * sqrt(0.5)
    odd_1_i = (tmp_0 + tmp_1) * sqrt(0.5)

    output[1] = (odd_0_r + odd_1_r) - 1j * (odd_0_i + odd_1_i)
    output[5] = (odd_0_r - odd_1_r) - 1j * (odd_0_i - odd_1_i)

    output[3] = conjugate(output[5])
    output[6] = conjugate(output[2])
706 707 708 709 710 711 712 713 714 715 716
    output[7] = conjugate(output[1])


# Also a basic implementation of the real-valued DFT4 :

def DFT4 (input, output):
    tmp_0 = input[0] + input[2]
    tmp_1 = input[0] - input[2]
    tmp_2 = input[1] + input[3]
    tmp_3 = input[1] - input[3]

717 718 719 720
    output[0] = tmp_0 + tmp_2
    output[1] = tmp_1 - 1j * tmp_3
    output[2] = tmp_0 - tmp_2
    output[3] = tmp_1 + 1j * tmp_3
721 722


723 724 725 726 727 728 729
# A similar idea might be used to calculate only the real part of the output
# of a complex DFT : we take an DFT algorithm for real inputs and complex
# outputs and we simply reverse it. The resulting algorithm will only work
# with inputs that satisfy the conjugaison rule (input[i] is the conjugate of
# input[N-i]) so we can do a first pass to modify the input so that it follows
# this rule. An example implementation is as follows (adapted from the
# unscaled_DFT_split_radix_time algorithm) :
730

731
def complex2real_unscaled_DFT_split_radix_time (N, input, output):
732
    even_input = vector(N/2)
733
    input_1 = vector(N/4)
734
    even_output = vector(N/2)
735
    output_1 = vector(N/4)
736 737 738 739

    for i in range(N/2):
	even_input[i] = input[2*i]

740 741 742
    for i in range(N/4):
	input_1[i] = input[4*i+1] + conjugate(input[N-1-4*i])

743
    unscaled_DFT (N/2, even_input, even_output)
744
    unscaled_DFT (N/4, input_1, output_1)
745

746 747
    for i in range(N/4):
	output_1[i] = output_1[i] * W (i, N)
748

749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786
    for i in range(N/4):
	output[i] = even_output[i] + output_1[i].real
	output[i+N/4] = even_output[i+N/4] + output_1[i].imag
	output[i+N/2] = even_output[i] - output_1[i].real
	output[i+3*N/4] = even_output[i+N/4] - output_1[i].imag

# This algorithm does N/4 complex additions, N/4-1 complex multiplies
# (including one "simple" multiply for i=N/8), N real additions, one
# "complex-to-real" DFT of size N/2, and one complex DFT of size N/4.
# Also, in the complex DFT of size N/4, we do not care about the imaginary
# part of output_1[0], which in practice allows us to save one real addition.

# This gives us the following performance for complex DFT with real outputs :
#	N	real additions	real multiplies	complex multiplies
#       1               0               0               0
#       2               2               0               0
#       4               8               0               0
#       8              25               2               0
#      16              66               4               2
#      32             167              10               8
#      64             400              20              26
#     128             933              42              72
#     256            2126              84             186
#     512            4771             170             456
#    1024           10572             340            1082
#    2048           23201             682            2504
#    4096           50506            1364            5690
#    8192          109215            2730           12744
#   16384          234824            5460           28218
#   32768          502429           10922           61896
#   65536         1070406           21844          134714


# Now let's talk about the DCT algorithm. The canonical definition for it is
# as follows :

def C (k, N):
    return cos ((k*pi)/(2*N))
787

788 789 790 791 792
def unscaled_DCT (N, input, output):
    for o in range(N):		# o is output index
	output[o] = 0
	for i in range(N):	# i is input index
	    output[o] = output[o] + input[i] * C ((2*i+1)*o, N)
793

794
# This trivial algorithm uses N*N multiplications and N*(N-1) additions.
795 796


797 798
# One possible decomposition on this calculus is to use the fact that C (i, N)
# and C (2*N-i, N) are opposed. This can lead to this decomposition :
799

800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833
#def unscaled_DCT (N, input, output):
#   even_input = vector (N)
#   odd_input = vector (N)
#   even_output = vector (N)
#   odd_output = vector (N)
#
#   for i in range(N/2):
#	even_input[i] = input[i] + input[N-1-i]
#	odd_input[i] = input[i] - input[N-1-i]
#
#   unscaled_DCT (N, even_input, even_output)
#   unscaled_DCT (N, odd_input, odd_output)
#
#   for i in range(N/2):
#	output[2*i] = even_output[2*i]
#	output[2*i+1] = odd_output[2*i+1]

# Now the even part can easily be calculated : by looking at the C(k,N)
# formula, we see that the even part is actually an unscaled DCT of size N/2.
# The odd part looks like a DCT of size N/2, but the coefficients are
# actually C ((2*i+1)*(2*o+1), 2*N) instead of C ((2*i+1)*o, N).

# We use a trigonometric relation here :
# 2 * C ((a+b)/2, N) * C ((a-b)/2, N) = C (a, N) + C (b, N)
# Thus with a = (2*i+1)*o and b = (2*i+1)*(o+1) :
# 2 * C((2*i+1)*(2*o+1),2N) * C(2*i+1,2N) = C((2*i+1)*o,N) + C((2*i+1)*(o+1),N)

# This leads us to the Lee DCT algorithm :

def unscaled_DCT_Lee (N, input, output):
    even_input = vector(N/2)
    odd_input = vector(N/2)
    even_output = vector(N/2)
    odd_output = vector(N/2)
834

835 836 837
    for i in range(N/2):
	even_input[i] = input[i] + input[N-1-i]
	odd_input[i] = input[i] - input[N-1-i]
838

839 840
    for i in range(N/2):
	odd_input[i] = odd_input[i] * (0.5 / C (2*i+1, N))
841

842 843
    unscaled_DCT (N/2, even_input, even_output)
    unscaled_DCT (N/2, odd_input, odd_output)
844

845 846
    for i in range(N/2-1):
	odd_output[i] = odd_output[i] + odd_output[i+1]
847 848

    for i in range(N/2):
849 850
	output[2*i] = even_output[i]
	output[2*i+1] = odd_output[i];
851

852
# Notes about this algorithm :
853

854 855
# The algorithm can be easily inverted to calculate the IDCT instead :
# each of the basic stages are separately inversible...
856

857 858 859 860 861 862 863 864
# This function does N adds, then N/2 muls, then 2 recursive calls with
# size N/2, then N/2-1 adds again. If we apply it recursively, the total
# number of operations will be N*log2(N)/2 multiplies and N*(3*log2(N)/2-1) + 1
# additions. So this is much faster than the canonical algorithm.

# Some of the multiplication coefficients 0.5/cos(...) can get quite large.
# This means that a small error in the input will give a large error on the
# output... For a DCT of size N the biggest coefficient will be at i=N/2-1
865
# and it will be slightly more than N/pi which can be large for large N's.
866 867 868 869 870 871 872 873 874 875 876 877 878 879

# In the IDCT however, the multiplication coefficients for the reverse
# transformation are of the form 2*cos(...) so they can not get big and there
# is no accuracy problem.

# You can find another description of this algorithm at
# http://www.intel.com/drg/mmx/appnotes/ap533.htm



# Another idea is to observe that the DCT calculation can be made to look like
# the DFT calculation : C (k, N) is the real part of W (k, 4*N) or W (-k, 4*N).
# We can use this idea translate the DCT algorithm into a call to the DFT
# algorithm :
880

881 882 883
def unscaled_DCT_DFT (N, input, output):
    DFT_input = vector (4*N)
    DFT_output = vector (4*N)
884

885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948
    for i in range(N):
	DFT_input[2*i+1] = input[i]
	#DFT_input[4*N-2*i-1] = input[i]	# We could use this instead

    unscaled_DFT (4*N, DFT_input, DFT_output)

    for i in range(N):
	output[i] = DFT_output[i].real


# We can then use our knowledge of the DFT calculation to optimize for this
# particular case. For example using the radix-2 decimation-in-time method :

#def unscaled_DCT_DFT (N, input, output):
#   DFT_input = vector (2*N)
#   DFT_output = vector (2*N)
#
#   for i in range(N):
#	DFT_input[i] = input[i]
#	#DFT_input[2*N-1-i] = input[i]	# We could use this instead
#
#   unscaled_DFT (2*N, DFT_input, DFT_output)
#
#   for i in range(N):
#       DFT_output[i] = DFT_output[i] * W (i, 4*N)
#
#   for i in range(N):
#	output[i] = DFT_output[i].real

# This leads us to the AAN implementation of the DCT algorithm : if we set
# both DFT_input[i] and DFT_input[2*N-1-i] to input[i], then the imaginary
# parts of W(2*i+1) and W(-2*i-1) will compensate, and output_DFT[i] will
# already be a real after the multiplication by W(i,4*N). Which means that
# before the multiplication, it is the product of a real number and W(-i,4*N).
# This leads to the following code, called the AAN algorithm :

def unscaled_DCT_AAN (N, input, output):
    DFT_input = vector (2*N)
    DFT_output = vector (2*N)

    for i in range(N):
	DFT_input[i] = input[i]
	DFT_input[2*N-1-i] = input[i]

    symetrical_unscaled_DFT (2*N, DFT_input, DFT_output)

    for i in range(N):
	output[i] = DFT_output[i].real * (0.5 / C (i, N))

# Notes about the AAN algorithm :

# The cost of this function is N real multiplies and a DFT of size 2*N. The
# DFT to calculate has special properties : the inputs are real and symmetric.
# Also, we only need to calculate the real parts of the N first DFT outputs.
# We can try to take advantage of all that.

# We can invert this algorithm to calculate the IDCT. The final multiply
# stage is trivially invertible. The DFT stage is invertible too, but we have
# to take into account the special properties of this particular DFT for that.

# Once again we have to take care of numerical precision for the DFT : the
# output coefficients can get large, so that a small error in the input will
# give a large error on the output... For a DCT of size N the biggest
# coefficient will be at i=N/2-1 and it will be slightly more than N/pi
949

950 951 952 953 954 955 956 957 958 959 960
# You can find another description of this algorithm at this url :
# www.cmlab.csie.ntu.edu.tw/cml/dsp/training/coding/transform/fastdct.html
# (It is the same server where we already found a description of the fast DFT)


# To optimize the DFT calculation, we can take a lot of specific things into
# account : the input is real and symetric, and we only care about the real
# part of the output. Also, we only care about the N first output coefficients,
# but that one does not save operations actually, because the other
# coefficients are the conjugates of the ones we look anyway.

961
# One useful way to use the symmetry of the input is to use the radix-2
962
# decimation-in-frequency algorithm. We can write a version of
963
# unscaled_DFT_radix2_freq for the case where the input is symmetrical :
964
# we have removed a few additions in the first stages because even_input
965
# is symmetrical and odd_input is antisymetrical. Also, we have modified the
966 967 968
# odd_input vector so that the second half of it is set to zero and the real
# part of the DFT output is not modified. After that modification, the second
# part of the odd_input was null so we used the radix-2 decimation-in-frequency
969
# again on the odd DFT. Also odd_output is symmetrical because input is real...
970 971 972 973 974 975 976

def symetrical_unscaled_DFT (N, input, output):
    even_input = vector(N/2)
    odd_tmp = vector(N/2)
    odd_input = vector(N/2)
    even_output = vector(N/2)
    odd_output = vector(N/2)
977

978 979
    for i in range(N/4):
	even_input[N/2-i-1] = even_input[i] = input[i] + input[N/2-1-i]
980

981 982 983 984 985 986 987 988
    for i in range(N/4):
	odd_tmp[i] = input[i] - input[N/2-1-i]

    odd_input[0] = odd_tmp[0]
    for i in range(N/4)[1:]:
	odd_input[i] = (odd_tmp[i] + odd_tmp[i-1]) * W (i, N)

    unscaled_DFT (N/2, even_input, even_output)
989
    # symmetrical real inputs, real outputs
990 991 992 993 994 995 996 997 998 999 1000

    unscaled_DFT (N/4, odd_input, odd_output)
    # complex inputs, real outputs

    for i in range(N/2):
	output[2*i] = even_output[i]

    for i in range(N/4):
	output[N-1-4*i] = output[4*i+1] = odd_output[i]

# This procedure takes 3*N/4-1 real additions and N/2-3 real multiplies,
1001
# followed by another symmetrical real DFT of size N/2 and a "complex to real"
1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049
# DFT of size N/4.

# We thus get the following performance results :
#	N	real additions	real multiplies	complex multiplies
#       1               0               0               0
#       2               0               0               0
#       4               2               0               0
#       8               9               1               0
#      16              28               6               0
#      32              76              21               0
#      64             189              54               2
#     128             451             125              10
#     256            1042             270              36
#     512            2358             565             108
#    1024            5251            1158             294
#    2048           11557            2349             750
#    4096           25200            4734            1832
#    8192           54544            9509            4336
#   16384          117337           19062           10026
#   32768          251127           38173           22770
#   65536          535102           76398           50988


# We thus get a better performance with the AAN DCT algorithm than with the
# Lee DCT algorithm : we can do a DCT of size 32 with 189 additions, 54+32 real
# multiplies, and 2 complex multiplies. The Lee algorithm would have used 209
# additions and 80 multiplies. With the AAN algorithm, we also have the
# advantage that a big number of the multiplies are actually grouped at the
# output stage of the algorithm, so if we want to do a DCT followed by a
# quantization stage, we will be able to group the multiply of the output with
# the multiply of the quantization stage, thus saving 32 more operations. In
# the mpeg audio layer 1 or 2 processing, we can also group the multiply of the
# output with the multiply of the convolution stage...

# Another source code for the AAN algorithm (implemented on 8 points, and
# without all of the explanations) can be found at this URL :
# http://developer.intel.com/drg/pentiumII/appnotes/aan_org.c . This
# implementation uses 28 adds and 6+8 muls instead of 29 adds and 5+8 muls -
# the difference is that in the symetrical_unscaled_DFT procedure, they noticed
# how odd_input[i] and odd_input[N/4-i] will be combined at the start of the
# complex-to-real DFT and they took advantage of this to convert 2 real adds
# and 4 real muls into one complex multiply.


# TODO : write about multi-dimentional DCT


# TEST CODE
1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062

def dump (vector):
    str = ""
    for i in range(len(vector)):
	if i:
	    str = str + ", "
	vector[i] = vector[i] + 0j
	realstr = "%+.4f" % vector[i].real
	imagstr = "%+.4fj" % vector[i].imag
	if (realstr == "-0.0000"):
	    realstr = "+0.0000"
	if (imagstr == "-0.0000j"):
	    imagstr = "+0.0000j"
1063
	str = str + realstr #+ imagstr
1064 1065 1066 1067 1068 1069 1070 1071
    return "[%s]" % str

def test(N):
    input = vector(N)
    output = vector(N)
    verify = vector(N)

    for i in range(N):
1072
	input[i] = random() + 1j * random()
1073

1074 1075
    unscaled_DFT (N, input, output)
    unscaled_DFT (N, input, verify)
1076

1077 1078
    if (dump(output) != dump(verify)):
	print dump(output)
1079 1080
	print dump(verify)

1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172
#test (64)


# PERFORMANCE ANALYSIS CODE

def display (table):
    N = 1
    print "#\tN\treal additions\treal multiplies\tcomplex multiplies"
    while table.has_key(N):
	print "#%8d%16d%16d%16d" % (N, table[N][0], table[N][1], table[N][2])
	N = 2*N
    print

best_complex_DFT = {}

def complex_DFT (max_N):
    best_complex_DFT[1] = (0,0,0)
    best_complex_DFT[2] = (4,0,0)
    best_complex_DFT[4] = (16,0,0)
    N = 8
    while (N<=max_N):
	# best method = split radix
	best2 = best_complex_DFT[N/2]
	best4 = best_complex_DFT[N/4]
	best_complex_DFT[N] = (best2[0] + 2*best4[0] + 3*N + 4,
			       best2[1] + 2*best4[1] + 4,
			       best2[2] + 2*best4[2] + N/2 - 4)
	N = 2*N

best_real_DFT = {}

def real_DFT (max_N):
    best_real_DFT[1] = (0,0,0)
    best_real_DFT[2] = (2,0,0)
    best_real_DFT[4] = (6,0,0)
    N = 8
    while (N<=max_N):
	# best method = split radix decimate-in-frequency
	best2 = best_real_DFT[N/2]
	best4 = best_complex_DFT[N/4]
	best_real_DFT[N] = (best2[0] + best4[0] + N + 2,
			    best2[1] + best4[1] + 2,
			    best2[2] + best4[2] + N/4 - 2)
	N = 2*N

best_complex2real_DFT = {}

def complex2real_DFT (max_N):
    best_complex2real_DFT[1] = (0,0,0)
    best_complex2real_DFT[2] = (2,0,0)
    best_complex2real_DFT[4] = (8,0,0)
    N = 8
    while (N<=max_N):
	best2 = best_complex2real_DFT[N/2]
	best4 = best_complex_DFT[N/4]
	best_complex2real_DFT[N] = (best2[0] + best4[0] + 3*N/2 + 1,
				    best2[1] + best4[1] + 2,
				    best2[2] + best4[2] + N/4 - 2)
	N = 2*N

best_real_symetric_DFT = {}

def real_symetric_DFT (max_N):
    best_real_symetric_DFT[1] = (0,0,0)
    best_real_symetric_DFT[2] = (0,0,0)
    best_real_symetric_DFT[4] = (2,0,0)
    N = 8
    while (N<=max_N):
	best2 = best_real_symetric_DFT[N/2]
	best4 = best_complex2real_DFT[N/4]
	best_real_symetric_DFT[N] = (best2[0] + best4[0] + 3*N/4 - 1,
				     best2[1] + best4[1] + N/2 - 3,
				     best2[2] + best4[2])
	N = 2*N

complex_DFT (65536)
real_DFT (65536)
complex2real_DFT (65536)
real_symetric_DFT (65536)


print "complex DFT"
display (best_complex_DFT)

print "real DFT"
display (best_real_DFT)

print "complex2real DFT"
display (best_complex2real_DFT)

print "real symetric DFT"
display (best_real_symetric_DFT)