2 * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
11 #include "./vpx_dsp_rtcd.h"
12 #include "vpx_dsp/fwd_txfm.h"
14 void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride) {
15 // The 2D transform is done with two passes which are actually pretty
16 // similar. In the first one, we transform the columns and transpose
17 // the results. In the second one, we transform the rows. To achieve that,
18 // as the first pass results are transposed, we transpose the columns (that
19 // is the transposed rows) and transpose the results (so that it goes back
20 // in normal/row positions).
22 // We need an intermediate buffer between passes.
23 tran_low_t intermediate[4 * 4];
24 const int16_t *in_pass0 = input;
25 const tran_low_t *in = NULL;
26 tran_low_t *out = intermediate;
27 // Do the two transform/transpose passes
28 for (pass = 0; pass < 2; ++pass) {
29 tran_high_t input[4]; // canbe16
30 tran_high_t step[4]; // canbe16
31 tran_high_t temp1, temp2; // needs32
33 for (i = 0; i < 4; ++i) {
36 input[0] = in_pass0[0 * stride] * 16;
37 input[1] = in_pass0[1 * stride] * 16;
38 input[2] = in_pass0[2 * stride] * 16;
39 input[3] = in_pass0[3 * stride] * 16;
40 if (i == 0 && input[0]) {
50 step[0] = input[0] + input[3];
51 step[1] = input[1] + input[2];
52 step[2] = input[1] - input[2];
53 step[3] = input[0] - input[3];
54 temp1 = (step[0] + step[1]) * cospi_16_64;
55 temp2 = (step[0] - step[1]) * cospi_16_64;
56 out[0] = (tran_low_t)fdct_round_shift(temp1);
57 out[2] = (tran_low_t)fdct_round_shift(temp2);
58 temp1 = step[2] * cospi_24_64 + step[3] * cospi_8_64;
59 temp2 = -step[2] * cospi_8_64 + step[3] * cospi_24_64;
60 out[1] = (tran_low_t)fdct_round_shift(temp1);
61 out[3] = (tran_low_t)fdct_round_shift(temp2);
62 // Do next column (which is a transposed row in second/horizontal pass)
67 // Setup in/out for next pass.
74 for (i = 0; i < 4; ++i) {
75 for (j = 0; j < 4; ++j)
76 output[j + i * 4] = (output[j + i * 4] + 1) >> 2;
81 void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride) {
84 for (r = 0; r < 4; ++r)
85 for (c = 0; c < 4; ++c)
86 sum += input[r * stride + c];
91 void vpx_fdct8x8_c(const int16_t *input, tran_low_t *final_output, int stride) {
93 tran_low_t intermediate[64];
95 tran_low_t *output = intermediate;
96 const tran_low_t *in = NULL;
99 for (pass = 0; pass < 2; ++pass) {
100 tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; // canbe16
101 tran_high_t t0, t1, t2, t3; // needs32
102 tran_high_t x0, x1, x2, x3; // canbe16
105 for (i = 0; i < 8; i++) {
108 s0 = (input[0 * stride] + input[7 * stride]) * 4;
109 s1 = (input[1 * stride] + input[6 * stride]) * 4;
110 s2 = (input[2 * stride] + input[5 * stride]) * 4;
111 s3 = (input[3 * stride] + input[4 * stride]) * 4;
112 s4 = (input[3 * stride] - input[4 * stride]) * 4;
113 s5 = (input[2 * stride] - input[5 * stride]) * 4;
114 s6 = (input[1 * stride] - input[6 * stride]) * 4;
115 s7 = (input[0 * stride] - input[7 * stride]) * 4;
118 s0 = in[0 * 8] + in[7 * 8];
119 s1 = in[1 * 8] + in[6 * 8];
120 s2 = in[2 * 8] + in[5 * 8];
121 s3 = in[3 * 8] + in[4 * 8];
122 s4 = in[3 * 8] - in[4 * 8];
123 s5 = in[2 * 8] - in[5 * 8];
124 s6 = in[1 * 8] - in[6 * 8];
125 s7 = in[0 * 8] - in[7 * 8];
129 // fdct4(step, step);
134 t0 = (x0 + x1) * cospi_16_64;
135 t1 = (x0 - x1) * cospi_16_64;
136 t2 = x2 * cospi_24_64 + x3 * cospi_8_64;
137 t3 = -x2 * cospi_8_64 + x3 * cospi_24_64;
138 output[0] = (tran_low_t)fdct_round_shift(t0);
139 output[2] = (tran_low_t)fdct_round_shift(t2);
140 output[4] = (tran_low_t)fdct_round_shift(t1);
141 output[6] = (tran_low_t)fdct_round_shift(t3);
144 t0 = (s6 - s5) * cospi_16_64;
145 t1 = (s6 + s5) * cospi_16_64;
146 t2 = fdct_round_shift(t0);
147 t3 = fdct_round_shift(t1);
156 t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
157 t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
158 t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
159 t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
160 output[1] = (tran_low_t)fdct_round_shift(t0);
161 output[3] = (tran_low_t)fdct_round_shift(t2);
162 output[5] = (tran_low_t)fdct_round_shift(t1);
163 output[7] = (tran_low_t)fdct_round_shift(t3);
167 output = final_output;
171 for (i = 0; i < 8; ++i) {
172 for (j = 0; j < 8; ++j)
173 final_output[j + i * 8] /= 2;
177 void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride) {
180 for (r = 0; r < 8; ++r)
181 for (c = 0; c < 8; ++c)
182 sum += input[r * stride + c];
187 void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride) {
188 // The 2D transform is done with two passes which are actually pretty
189 // similar. In the first one, we transform the columns and transpose
190 // the results. In the second one, we transform the rows. To achieve that,
191 // as the first pass results are transposed, we transpose the columns (that
192 // is the transposed rows) and transpose the results (so that it goes back
193 // in normal/row positions).
195 // We need an intermediate buffer between passes.
196 tran_low_t intermediate[256];
197 const int16_t *in_pass0 = input;
198 const tran_low_t *in = NULL;
199 tran_low_t *out = intermediate;
200 // Do the two transform/transpose passes
201 for (pass = 0; pass < 2; ++pass) {
202 tran_high_t step1[8]; // canbe16
203 tran_high_t step2[8]; // canbe16
204 tran_high_t step3[8]; // canbe16
205 tran_high_t input[8]; // canbe16
206 tran_high_t temp1, temp2; // needs32
208 for (i = 0; i < 16; i++) {
210 // Calculate input for the first 8 results.
211 input[0] = (in_pass0[0 * stride] + in_pass0[15 * stride]) * 4;
212 input[1] = (in_pass0[1 * stride] + in_pass0[14 * stride]) * 4;
213 input[2] = (in_pass0[2 * stride] + in_pass0[13 * stride]) * 4;
214 input[3] = (in_pass0[3 * stride] + in_pass0[12 * stride]) * 4;
215 input[4] = (in_pass0[4 * stride] + in_pass0[11 * stride]) * 4;
216 input[5] = (in_pass0[5 * stride] + in_pass0[10 * stride]) * 4;
217 input[6] = (in_pass0[6 * stride] + in_pass0[ 9 * stride]) * 4;
218 input[7] = (in_pass0[7 * stride] + in_pass0[ 8 * stride]) * 4;
219 // Calculate input for the next 8 results.
220 step1[0] = (in_pass0[7 * stride] - in_pass0[ 8 * stride]) * 4;
221 step1[1] = (in_pass0[6 * stride] - in_pass0[ 9 * stride]) * 4;
222 step1[2] = (in_pass0[5 * stride] - in_pass0[10 * stride]) * 4;
223 step1[3] = (in_pass0[4 * stride] - in_pass0[11 * stride]) * 4;
224 step1[4] = (in_pass0[3 * stride] - in_pass0[12 * stride]) * 4;
225 step1[5] = (in_pass0[2 * stride] - in_pass0[13 * stride]) * 4;
226 step1[6] = (in_pass0[1 * stride] - in_pass0[14 * stride]) * 4;
227 step1[7] = (in_pass0[0 * stride] - in_pass0[15 * stride]) * 4;
229 // Calculate input for the first 8 results.
230 input[0] = ((in[0 * 16] + 1) >> 2) + ((in[15 * 16] + 1) >> 2);
231 input[1] = ((in[1 * 16] + 1) >> 2) + ((in[14 * 16] + 1) >> 2);
232 input[2] = ((in[2 * 16] + 1) >> 2) + ((in[13 * 16] + 1) >> 2);
233 input[3] = ((in[3 * 16] + 1) >> 2) + ((in[12 * 16] + 1) >> 2);
234 input[4] = ((in[4 * 16] + 1) >> 2) + ((in[11 * 16] + 1) >> 2);
235 input[5] = ((in[5 * 16] + 1) >> 2) + ((in[10 * 16] + 1) >> 2);
236 input[6] = ((in[6 * 16] + 1) >> 2) + ((in[ 9 * 16] + 1) >> 2);
237 input[7] = ((in[7 * 16] + 1) >> 2) + ((in[ 8 * 16] + 1) >> 2);
238 // Calculate input for the next 8 results.
239 step1[0] = ((in[7 * 16] + 1) >> 2) - ((in[ 8 * 16] + 1) >> 2);
240 step1[1] = ((in[6 * 16] + 1) >> 2) - ((in[ 9 * 16] + 1) >> 2);
241 step1[2] = ((in[5 * 16] + 1) >> 2) - ((in[10 * 16] + 1) >> 2);
242 step1[3] = ((in[4 * 16] + 1) >> 2) - ((in[11 * 16] + 1) >> 2);
243 step1[4] = ((in[3 * 16] + 1) >> 2) - ((in[12 * 16] + 1) >> 2);
244 step1[5] = ((in[2 * 16] + 1) >> 2) - ((in[13 * 16] + 1) >> 2);
245 step1[6] = ((in[1 * 16] + 1) >> 2) - ((in[14 * 16] + 1) >> 2);
246 step1[7] = ((in[0 * 16] + 1) >> 2) - ((in[15 * 16] + 1) >> 2);
248 // Work on the first eight values; fdct8(input, even_results);
250 tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; // canbe16
251 tran_high_t t0, t1, t2, t3; // needs32
252 tran_high_t x0, x1, x2, x3; // canbe16
255 s0 = input[0] + input[7];
256 s1 = input[1] + input[6];
257 s2 = input[2] + input[5];
258 s3 = input[3] + input[4];
259 s4 = input[3] - input[4];
260 s5 = input[2] - input[5];
261 s6 = input[1] - input[6];
262 s7 = input[0] - input[7];
264 // fdct4(step, step);
269 t0 = (x0 + x1) * cospi_16_64;
270 t1 = (x0 - x1) * cospi_16_64;
271 t2 = x3 * cospi_8_64 + x2 * cospi_24_64;
272 t3 = x3 * cospi_24_64 - x2 * cospi_8_64;
273 out[0] = (tran_low_t)fdct_round_shift(t0);
274 out[4] = (tran_low_t)fdct_round_shift(t2);
275 out[8] = (tran_low_t)fdct_round_shift(t1);
276 out[12] = (tran_low_t)fdct_round_shift(t3);
279 t0 = (s6 - s5) * cospi_16_64;
280 t1 = (s6 + s5) * cospi_16_64;
281 t2 = fdct_round_shift(t0);
282 t3 = fdct_round_shift(t1);
291 t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
292 t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
293 t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
294 t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
295 out[2] = (tran_low_t)fdct_round_shift(t0);
296 out[6] = (tran_low_t)fdct_round_shift(t2);
297 out[10] = (tran_low_t)fdct_round_shift(t1);
298 out[14] = (tran_low_t)fdct_round_shift(t3);
300 // Work on the next eight values; step1 -> odd_results
303 temp1 = (step1[5] - step1[2]) * cospi_16_64;
304 temp2 = (step1[4] - step1[3]) * cospi_16_64;
305 step2[2] = fdct_round_shift(temp1);
306 step2[3] = fdct_round_shift(temp2);
307 temp1 = (step1[4] + step1[3]) * cospi_16_64;
308 temp2 = (step1[5] + step1[2]) * cospi_16_64;
309 step2[4] = fdct_round_shift(temp1);
310 step2[5] = fdct_round_shift(temp2);
312 step3[0] = step1[0] + step2[3];
313 step3[1] = step1[1] + step2[2];
314 step3[2] = step1[1] - step2[2];
315 step3[3] = step1[0] - step2[3];
316 step3[4] = step1[7] - step2[4];
317 step3[5] = step1[6] - step2[5];
318 step3[6] = step1[6] + step2[5];
319 step3[7] = step1[7] + step2[4];
321 temp1 = step3[1] * -cospi_8_64 + step3[6] * cospi_24_64;
322 temp2 = step3[2] * cospi_24_64 + step3[5] * cospi_8_64;
323 step2[1] = fdct_round_shift(temp1);
324 step2[2] = fdct_round_shift(temp2);
325 temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
326 temp2 = step3[1] * cospi_24_64 + step3[6] * cospi_8_64;
327 step2[5] = fdct_round_shift(temp1);
328 step2[6] = fdct_round_shift(temp2);
330 step1[0] = step3[0] + step2[1];
331 step1[1] = step3[0] - step2[1];
332 step1[2] = step3[3] + step2[2];
333 step1[3] = step3[3] - step2[2];
334 step1[4] = step3[4] - step2[5];
335 step1[5] = step3[4] + step2[5];
336 step1[6] = step3[7] - step2[6];
337 step1[7] = step3[7] + step2[6];
339 temp1 = step1[0] * cospi_30_64 + step1[7] * cospi_2_64;
340 temp2 = step1[1] * cospi_14_64 + step1[6] * cospi_18_64;
341 out[1] = (tran_low_t)fdct_round_shift(temp1);
342 out[9] = (tran_low_t)fdct_round_shift(temp2);
343 temp1 = step1[2] * cospi_22_64 + step1[5] * cospi_10_64;
344 temp2 = step1[3] * cospi_6_64 + step1[4] * cospi_26_64;
345 out[5] = (tran_low_t)fdct_round_shift(temp1);
346 out[13] = (tran_low_t)fdct_round_shift(temp2);
347 temp1 = step1[3] * -cospi_26_64 + step1[4] * cospi_6_64;
348 temp2 = step1[2] * -cospi_10_64 + step1[5] * cospi_22_64;
349 out[3] = (tran_low_t)fdct_round_shift(temp1);
350 out[11] = (tran_low_t)fdct_round_shift(temp2);
351 temp1 = step1[1] * -cospi_18_64 + step1[6] * cospi_14_64;
352 temp2 = step1[0] * -cospi_2_64 + step1[7] * cospi_30_64;
353 out[7] = (tran_low_t)fdct_round_shift(temp1);
354 out[15] = (tran_low_t)fdct_round_shift(temp2);
356 // Do next column (which is a transposed row in second/horizontal pass)
361 // Setup in/out for next pass.
367 void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride) {
370 for (r = 0; r < 16; ++r)
371 for (c = 0; c < 16; ++c)
372 sum += input[r * stride + c];
374 output[0] = (tran_low_t)(sum >> 1);
377 static INLINE tran_high_t dct_32_round(tran_high_t input) {
378 tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
379 // TODO(debargha, peter.derivaz): Find new bounds for this assert,
380 // and make the bounds consts.
381 // assert(-131072 <= rv && rv <= 131071);
385 static INLINE tran_high_t half_round_shift(tran_high_t input) {
386 tran_high_t rv = (input + 1 + (input < 0)) >> 2;
390 void vpx_fdct32(const tran_high_t *input, tran_high_t *output, int round) {
391 tran_high_t step[32];
393 step[0] = input[0] + input[(32 - 1)];
394 step[1] = input[1] + input[(32 - 2)];
395 step[2] = input[2] + input[(32 - 3)];
396 step[3] = input[3] + input[(32 - 4)];
397 step[4] = input[4] + input[(32 - 5)];
398 step[5] = input[5] + input[(32 - 6)];
399 step[6] = input[6] + input[(32 - 7)];
400 step[7] = input[7] + input[(32 - 8)];
401 step[8] = input[8] + input[(32 - 9)];
402 step[9] = input[9] + input[(32 - 10)];
403 step[10] = input[10] + input[(32 - 11)];
404 step[11] = input[11] + input[(32 - 12)];
405 step[12] = input[12] + input[(32 - 13)];
406 step[13] = input[13] + input[(32 - 14)];
407 step[14] = input[14] + input[(32 - 15)];
408 step[15] = input[15] + input[(32 - 16)];
409 step[16] = -input[16] + input[(32 - 17)];
410 step[17] = -input[17] + input[(32 - 18)];
411 step[18] = -input[18] + input[(32 - 19)];
412 step[19] = -input[19] + input[(32 - 20)];
413 step[20] = -input[20] + input[(32 - 21)];
414 step[21] = -input[21] + input[(32 - 22)];
415 step[22] = -input[22] + input[(32 - 23)];
416 step[23] = -input[23] + input[(32 - 24)];
417 step[24] = -input[24] + input[(32 - 25)];
418 step[25] = -input[25] + input[(32 - 26)];
419 step[26] = -input[26] + input[(32 - 27)];
420 step[27] = -input[27] + input[(32 - 28)];
421 step[28] = -input[28] + input[(32 - 29)];
422 step[29] = -input[29] + input[(32 - 30)];
423 step[30] = -input[30] + input[(32 - 31)];
424 step[31] = -input[31] + input[(32 - 32)];
427 output[0] = step[0] + step[16 - 1];
428 output[1] = step[1] + step[16 - 2];
429 output[2] = step[2] + step[16 - 3];
430 output[3] = step[3] + step[16 - 4];
431 output[4] = step[4] + step[16 - 5];
432 output[5] = step[5] + step[16 - 6];
433 output[6] = step[6] + step[16 - 7];
434 output[7] = step[7] + step[16 - 8];
435 output[8] = -step[8] + step[16 - 9];
436 output[9] = -step[9] + step[16 - 10];
437 output[10] = -step[10] + step[16 - 11];
438 output[11] = -step[11] + step[16 - 12];
439 output[12] = -step[12] + step[16 - 13];
440 output[13] = -step[13] + step[16 - 14];
441 output[14] = -step[14] + step[16 - 15];
442 output[15] = -step[15] + step[16 - 16];
444 output[16] = step[16];
445 output[17] = step[17];
446 output[18] = step[18];
447 output[19] = step[19];
449 output[20] = dct_32_round((-step[20] + step[27]) * cospi_16_64);
450 output[21] = dct_32_round((-step[21] + step[26]) * cospi_16_64);
451 output[22] = dct_32_round((-step[22] + step[25]) * cospi_16_64);
452 output[23] = dct_32_round((-step[23] + step[24]) * cospi_16_64);
454 output[24] = dct_32_round((step[24] + step[23]) * cospi_16_64);
455 output[25] = dct_32_round((step[25] + step[22]) * cospi_16_64);
456 output[26] = dct_32_round((step[26] + step[21]) * cospi_16_64);
457 output[27] = dct_32_round((step[27] + step[20]) * cospi_16_64);
459 output[28] = step[28];
460 output[29] = step[29];
461 output[30] = step[30];
462 output[31] = step[31];
464 // dump the magnitude by 4, hence the intermediate values are within
465 // the range of 16 bits.
467 output[0] = half_round_shift(output[0]);
468 output[1] = half_round_shift(output[1]);
469 output[2] = half_round_shift(output[2]);
470 output[3] = half_round_shift(output[3]);
471 output[4] = half_round_shift(output[4]);
472 output[5] = half_round_shift(output[5]);
473 output[6] = half_round_shift(output[6]);
474 output[7] = half_round_shift(output[7]);
475 output[8] = half_round_shift(output[8]);
476 output[9] = half_round_shift(output[9]);
477 output[10] = half_round_shift(output[10]);
478 output[11] = half_round_shift(output[11]);
479 output[12] = half_round_shift(output[12]);
480 output[13] = half_round_shift(output[13]);
481 output[14] = half_round_shift(output[14]);
482 output[15] = half_round_shift(output[15]);
484 output[16] = half_round_shift(output[16]);
485 output[17] = half_round_shift(output[17]);
486 output[18] = half_round_shift(output[18]);
487 output[19] = half_round_shift(output[19]);
488 output[20] = half_round_shift(output[20]);
489 output[21] = half_round_shift(output[21]);
490 output[22] = half_round_shift(output[22]);
491 output[23] = half_round_shift(output[23]);
492 output[24] = half_round_shift(output[24]);
493 output[25] = half_round_shift(output[25]);
494 output[26] = half_round_shift(output[26]);
495 output[27] = half_round_shift(output[27]);
496 output[28] = half_round_shift(output[28]);
497 output[29] = half_round_shift(output[29]);
498 output[30] = half_round_shift(output[30]);
499 output[31] = half_round_shift(output[31]);
503 step[0] = output[0] + output[(8 - 1)];
504 step[1] = output[1] + output[(8 - 2)];
505 step[2] = output[2] + output[(8 - 3)];
506 step[3] = output[3] + output[(8 - 4)];
507 step[4] = -output[4] + output[(8 - 5)];
508 step[5] = -output[5] + output[(8 - 6)];
509 step[6] = -output[6] + output[(8 - 7)];
510 step[7] = -output[7] + output[(8 - 8)];
513 step[10] = dct_32_round((-output[10] + output[13]) * cospi_16_64);
514 step[11] = dct_32_round((-output[11] + output[12]) * cospi_16_64);
515 step[12] = dct_32_round((output[12] + output[11]) * cospi_16_64);
516 step[13] = dct_32_round((output[13] + output[10]) * cospi_16_64);
517 step[14] = output[14];
518 step[15] = output[15];
520 step[16] = output[16] + output[23];
521 step[17] = output[17] + output[22];
522 step[18] = output[18] + output[21];
523 step[19] = output[19] + output[20];
524 step[20] = -output[20] + output[19];
525 step[21] = -output[21] + output[18];
526 step[22] = -output[22] + output[17];
527 step[23] = -output[23] + output[16];
528 step[24] = -output[24] + output[31];
529 step[25] = -output[25] + output[30];
530 step[26] = -output[26] + output[29];
531 step[27] = -output[27] + output[28];
532 step[28] = output[28] + output[27];
533 step[29] = output[29] + output[26];
534 step[30] = output[30] + output[25];
535 step[31] = output[31] + output[24];
538 output[0] = step[0] + step[3];
539 output[1] = step[1] + step[2];
540 output[2] = -step[2] + step[1];
541 output[3] = -step[3] + step[0];
543 output[5] = dct_32_round((-step[5] + step[6]) * cospi_16_64);
544 output[6] = dct_32_round((step[6] + step[5]) * cospi_16_64);
546 output[8] = step[8] + step[11];
547 output[9] = step[9] + step[10];
548 output[10] = -step[10] + step[9];
549 output[11] = -step[11] + step[8];
550 output[12] = -step[12] + step[15];
551 output[13] = -step[13] + step[14];
552 output[14] = step[14] + step[13];
553 output[15] = step[15] + step[12];
555 output[16] = step[16];
556 output[17] = step[17];
557 output[18] = dct_32_round(step[18] * -cospi_8_64 + step[29] * cospi_24_64);
558 output[19] = dct_32_round(step[19] * -cospi_8_64 + step[28] * cospi_24_64);
559 output[20] = dct_32_round(step[20] * -cospi_24_64 + step[27] * -cospi_8_64);
560 output[21] = dct_32_round(step[21] * -cospi_24_64 + step[26] * -cospi_8_64);
561 output[22] = step[22];
562 output[23] = step[23];
563 output[24] = step[24];
564 output[25] = step[25];
565 output[26] = dct_32_round(step[26] * cospi_24_64 + step[21] * -cospi_8_64);
566 output[27] = dct_32_round(step[27] * cospi_24_64 + step[20] * -cospi_8_64);
567 output[28] = dct_32_round(step[28] * cospi_8_64 + step[19] * cospi_24_64);
568 output[29] = dct_32_round(step[29] * cospi_8_64 + step[18] * cospi_24_64);
569 output[30] = step[30];
570 output[31] = step[31];
573 step[0] = dct_32_round((output[0] + output[1]) * cospi_16_64);
574 step[1] = dct_32_round((-output[1] + output[0]) * cospi_16_64);
575 step[2] = dct_32_round(output[2] * cospi_24_64 + output[3] * cospi_8_64);
576 step[3] = dct_32_round(output[3] * cospi_24_64 - output[2] * cospi_8_64);
577 step[4] = output[4] + output[5];
578 step[5] = -output[5] + output[4];
579 step[6] = -output[6] + output[7];
580 step[7] = output[7] + output[6];
582 step[9] = dct_32_round(output[9] * -cospi_8_64 + output[14] * cospi_24_64);
583 step[10] = dct_32_round(output[10] * -cospi_24_64 + output[13] * -cospi_8_64);
584 step[11] = output[11];
585 step[12] = output[12];
586 step[13] = dct_32_round(output[13] * cospi_24_64 + output[10] * -cospi_8_64);
587 step[14] = dct_32_round(output[14] * cospi_8_64 + output[9] * cospi_24_64);
588 step[15] = output[15];
590 step[16] = output[16] + output[19];
591 step[17] = output[17] + output[18];
592 step[18] = -output[18] + output[17];
593 step[19] = -output[19] + output[16];
594 step[20] = -output[20] + output[23];
595 step[21] = -output[21] + output[22];
596 step[22] = output[22] + output[21];
597 step[23] = output[23] + output[20];
598 step[24] = output[24] + output[27];
599 step[25] = output[25] + output[26];
600 step[26] = -output[26] + output[25];
601 step[27] = -output[27] + output[24];
602 step[28] = -output[28] + output[31];
603 step[29] = -output[29] + output[30];
604 step[30] = output[30] + output[29];
605 step[31] = output[31] + output[28];
612 output[4] = dct_32_round(step[4] * cospi_28_64 + step[7] * cospi_4_64);
613 output[5] = dct_32_round(step[5] * cospi_12_64 + step[6] * cospi_20_64);
614 output[6] = dct_32_round(step[6] * cospi_12_64 + step[5] * -cospi_20_64);
615 output[7] = dct_32_round(step[7] * cospi_28_64 + step[4] * -cospi_4_64);
616 output[8] = step[8] + step[9];
617 output[9] = -step[9] + step[8];
618 output[10] = -step[10] + step[11];
619 output[11] = step[11] + step[10];
620 output[12] = step[12] + step[13];
621 output[13] = -step[13] + step[12];
622 output[14] = -step[14] + step[15];
623 output[15] = step[15] + step[14];
625 output[16] = step[16];
626 output[17] = dct_32_round(step[17] * -cospi_4_64 + step[30] * cospi_28_64);
627 output[18] = dct_32_round(step[18] * -cospi_28_64 + step[29] * -cospi_4_64);
628 output[19] = step[19];
629 output[20] = step[20];
630 output[21] = dct_32_round(step[21] * -cospi_20_64 + step[26] * cospi_12_64);
631 output[22] = dct_32_round(step[22] * -cospi_12_64 + step[25] * -cospi_20_64);
632 output[23] = step[23];
633 output[24] = step[24];
634 output[25] = dct_32_round(step[25] * cospi_12_64 + step[22] * -cospi_20_64);
635 output[26] = dct_32_round(step[26] * cospi_20_64 + step[21] * cospi_12_64);
636 output[27] = step[27];
637 output[28] = step[28];
638 output[29] = dct_32_round(step[29] * cospi_28_64 + step[18] * -cospi_4_64);
639 output[30] = dct_32_round(step[30] * cospi_4_64 + step[17] * cospi_28_64);
640 output[31] = step[31];
651 step[8] = dct_32_round(output[8] * cospi_30_64 + output[15] * cospi_2_64);
652 step[9] = dct_32_round(output[9] * cospi_14_64 + output[14] * cospi_18_64);
653 step[10] = dct_32_round(output[10] * cospi_22_64 + output[13] * cospi_10_64);
654 step[11] = dct_32_round(output[11] * cospi_6_64 + output[12] * cospi_26_64);
655 step[12] = dct_32_round(output[12] * cospi_6_64 + output[11] * -cospi_26_64);
656 step[13] = dct_32_round(output[13] * cospi_22_64 + output[10] * -cospi_10_64);
657 step[14] = dct_32_round(output[14] * cospi_14_64 + output[9] * -cospi_18_64);
658 step[15] = dct_32_round(output[15] * cospi_30_64 + output[8] * -cospi_2_64);
660 step[16] = output[16] + output[17];
661 step[17] = -output[17] + output[16];
662 step[18] = -output[18] + output[19];
663 step[19] = output[19] + output[18];
664 step[20] = output[20] + output[21];
665 step[21] = -output[21] + output[20];
666 step[22] = -output[22] + output[23];
667 step[23] = output[23] + output[22];
668 step[24] = output[24] + output[25];
669 step[25] = -output[25] + output[24];
670 step[26] = -output[26] + output[27];
671 step[27] = output[27] + output[26];
672 step[28] = output[28] + output[29];
673 step[29] = -output[29] + output[28];
674 step[30] = -output[30] + output[31];
675 step[31] = output[31] + output[30];
677 // Final stage --- outputs indices are bit-reversed.
679 output[16] = step[1];
681 output[24] = step[3];
683 output[20] = step[5];
684 output[12] = step[6];
685 output[28] = step[7];
687 output[18] = step[9];
688 output[10] = step[10];
689 output[26] = step[11];
690 output[6] = step[12];
691 output[22] = step[13];
692 output[14] = step[14];
693 output[30] = step[15];
695 output[1] = dct_32_round(step[16] * cospi_31_64 + step[31] * cospi_1_64);
696 output[17] = dct_32_round(step[17] * cospi_15_64 + step[30] * cospi_17_64);
697 output[9] = dct_32_round(step[18] * cospi_23_64 + step[29] * cospi_9_64);
698 output[25] = dct_32_round(step[19] * cospi_7_64 + step[28] * cospi_25_64);
699 output[5] = dct_32_round(step[20] * cospi_27_64 + step[27] * cospi_5_64);
700 output[21] = dct_32_round(step[21] * cospi_11_64 + step[26] * cospi_21_64);
701 output[13] = dct_32_round(step[22] * cospi_19_64 + step[25] * cospi_13_64);
702 output[29] = dct_32_round(step[23] * cospi_3_64 + step[24] * cospi_29_64);
703 output[3] = dct_32_round(step[24] * cospi_3_64 + step[23] * -cospi_29_64);
704 output[19] = dct_32_round(step[25] * cospi_19_64 + step[22] * -cospi_13_64);
705 output[11] = dct_32_round(step[26] * cospi_11_64 + step[21] * -cospi_21_64);
706 output[27] = dct_32_round(step[27] * cospi_27_64 + step[20] * -cospi_5_64);
707 output[7] = dct_32_round(step[28] * cospi_7_64 + step[19] * -cospi_25_64);
708 output[23] = dct_32_round(step[29] * cospi_23_64 + step[18] * -cospi_9_64);
709 output[15] = dct_32_round(step[30] * cospi_15_64 + step[17] * -cospi_17_64);
710 output[31] = dct_32_round(step[31] * cospi_31_64 + step[16] * -cospi_1_64);
713 void vpx_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
715 tran_high_t output[32 * 32];
718 for (i = 0; i < 32; ++i) {
719 tran_high_t temp_in[32], temp_out[32];
720 for (j = 0; j < 32; ++j)
721 temp_in[j] = input[j * stride + i] * 4;
722 vpx_fdct32(temp_in, temp_out, 0);
723 for (j = 0; j < 32; ++j)
724 output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
728 for (i = 0; i < 32; ++i) {
729 tran_high_t temp_in[32], temp_out[32];
730 for (j = 0; j < 32; ++j)
731 temp_in[j] = output[j + i * 32];
732 vpx_fdct32(temp_in, temp_out, 0);
733 for (j = 0; j < 32; ++j)
735 (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
739 // Note that although we use dct_32_round in dct32 computation flow,
740 // this 2d fdct32x32 for rate-distortion optimization loop is operating
741 // within 16 bits precision.
742 void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *out, int stride) {
744 tran_high_t output[32 * 32];
747 for (i = 0; i < 32; ++i) {
748 tran_high_t temp_in[32], temp_out[32];
749 for (j = 0; j < 32; ++j)
750 temp_in[j] = input[j * stride + i] * 4;
751 vpx_fdct32(temp_in, temp_out, 0);
752 for (j = 0; j < 32; ++j)
753 // TODO(cd): see quality impact of only doing
754 // output[j * 32 + i] = (temp_out[j] + 1) >> 2;
755 // PS: also change code in vpx_dsp/x86/vpx_dct_sse2.c
756 output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
760 for (i = 0; i < 32; ++i) {
761 tran_high_t temp_in[32], temp_out[32];
762 for (j = 0; j < 32; ++j)
763 temp_in[j] = output[j + i * 32];
764 vpx_fdct32(temp_in, temp_out, 1);
765 for (j = 0; j < 32; ++j)
766 out[j + i * 32] = (tran_low_t)temp_out[j];
770 void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride) {
773 for (r = 0; r < 32; ++r)
774 for (c = 0; c < 32; ++c)
775 sum += input[r * stride + c];
777 output[0] = (tran_low_t)(sum >> 3);
780 #if CONFIG_VP9_HIGHBITDEPTH
781 void vpx_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output,
783 vpx_fdct4x4_c(input, output, stride);
786 void vpx_highbd_fdct8x8_c(const int16_t *input, tran_low_t *final_output,
788 vpx_fdct8x8_c(input, final_output, stride);
791 void vpx_highbd_fdct8x8_1_c(const int16_t *input, tran_low_t *final_output,
793 vpx_fdct8x8_1_c(input, final_output, stride);
796 void vpx_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output,
798 vpx_fdct16x16_c(input, output, stride);
801 void vpx_highbd_fdct16x16_1_c(const int16_t *input, tran_low_t *output,
803 vpx_fdct16x16_1_c(input, output, stride);
806 void vpx_highbd_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
807 vpx_fdct32x32_c(input, out, stride);
810 void vpx_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *out,
812 vpx_fdct32x32_rd_c(input, out, stride);
815 void vpx_highbd_fdct32x32_1_c(const int16_t *input, tran_low_t *out,
817 vpx_fdct32x32_1_c(input, out, stride);
819 #endif // CONFIG_VP9_HIGHBITDEPTH