• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3  *
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.
9  */
10 
11 #include <assert.h>
12 #include "./vpx_dsp_rtcd.h"
13 #include "vpx_dsp/fwd_txfm.h"
14 
vpx_fdct4x4_c(const int16_t * input,tran_low_t * output,int stride)15 void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride) {
16   // The 2D transform is done with two passes which are actually pretty
17   // similar. In the first one, we transform the columns and transpose
18   // the results. In the second one, we transform the rows. To achieve that,
19   // as the first pass results are transposed, we transpose the columns (that
20   // is the transposed rows) and transpose the results (so that it goes back
21   // in normal/row positions).
22   int pass;
23   // We need an intermediate buffer between passes.
24   tran_low_t intermediate[4 * 4];
25   const tran_low_t *in_low = 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 in_high[4];    // canbe16
30     tran_high_t step[4];       // canbe16
31     tran_high_t temp1, temp2;  // needs32
32     int i;
33     for (i = 0; i < 4; ++i) {
34       // Load inputs.
35       if (pass == 0) {
36         in_high[0] = input[0 * stride] * 16;
37         in_high[1] = input[1 * stride] * 16;
38         in_high[2] = input[2 * stride] * 16;
39         in_high[3] = input[3 * stride] * 16;
40         if (i == 0 && in_high[0]) {
41           ++in_high[0];
42         }
43       } else {
44         assert(in_low != NULL);
45         in_high[0] = in_low[0 * 4];
46         in_high[1] = in_low[1 * 4];
47         in_high[2] = in_low[2 * 4];
48         in_high[3] = in_low[3 * 4];
49         ++in_low;
50       }
51       // Transform.
52       step[0] = in_high[0] + in_high[3];
53       step[1] = in_high[1] + in_high[2];
54       step[2] = in_high[1] - in_high[2];
55       step[3] = in_high[0] - in_high[3];
56       temp1 = (step[0] + step[1]) * cospi_16_64;
57       temp2 = (step[0] - step[1]) * cospi_16_64;
58       out[0] = (tran_low_t)fdct_round_shift(temp1);
59       out[2] = (tran_low_t)fdct_round_shift(temp2);
60       temp1 = step[2] * cospi_24_64 + step[3] * cospi_8_64;
61       temp2 = -step[2] * cospi_8_64 + step[3] * cospi_24_64;
62       out[1] = (tran_low_t)fdct_round_shift(temp1);
63       out[3] = (tran_low_t)fdct_round_shift(temp2);
64       // Do next column (which is a transposed row in second/horizontal pass)
65       ++input;
66       out += 4;
67     }
68     // Setup in/out for next pass.
69     in_low = intermediate;
70     out = output;
71   }
72 
73   {
74     int i, j;
75     for (i = 0; i < 4; ++i) {
76       for (j = 0; j < 4; ++j) output[j + i * 4] = (output[j + i * 4] + 1) >> 2;
77     }
78   }
79 }
80 
vpx_fdct4x4_1_c(const int16_t * input,tran_low_t * output,int stride)81 void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride) {
82   int r, c;
83   tran_low_t sum = 0;
84   for (r = 0; r < 4; ++r)
85     for (c = 0; c < 4; ++c) sum += input[r * stride + c];
86 
87   output[0] = sum * 2;
88 }
89 
vpx_fdct8x8_c(const int16_t * input,tran_low_t * output,int stride)90 void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride) {
91   int i, j;
92   tran_low_t intermediate[64];
93   int pass;
94   tran_low_t *out = intermediate;
95   const tran_low_t *in = NULL;
96 
97   // Transform columns
98   for (pass = 0; pass < 2; ++pass) {
99     tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
100     tran_high_t t0, t1, t2, t3;                  // needs32
101     tran_high_t x0, x1, x2, x3;                  // canbe16
102 
103     for (i = 0; i < 8; i++) {
104       // stage 1
105       if (pass == 0) {
106         s0 = (input[0 * stride] + input[7 * stride]) * 4;
107         s1 = (input[1 * stride] + input[6 * stride]) * 4;
108         s2 = (input[2 * stride] + input[5 * stride]) * 4;
109         s3 = (input[3 * stride] + input[4 * stride]) * 4;
110         s4 = (input[3 * stride] - input[4 * stride]) * 4;
111         s5 = (input[2 * stride] - input[5 * stride]) * 4;
112         s6 = (input[1 * stride] - input[6 * stride]) * 4;
113         s7 = (input[0 * stride] - input[7 * stride]) * 4;
114         ++input;
115       } else {
116         s0 = in[0 * 8] + in[7 * 8];
117         s1 = in[1 * 8] + in[6 * 8];
118         s2 = in[2 * 8] + in[5 * 8];
119         s3 = in[3 * 8] + in[4 * 8];
120         s4 = in[3 * 8] - in[4 * 8];
121         s5 = in[2 * 8] - in[5 * 8];
122         s6 = in[1 * 8] - in[6 * 8];
123         s7 = in[0 * 8] - in[7 * 8];
124         ++in;
125       }
126 
127       // fdct4(step, step);
128       x0 = s0 + s3;
129       x1 = s1 + s2;
130       x2 = s1 - s2;
131       x3 = s0 - s3;
132       t0 = (x0 + x1) * cospi_16_64;
133       t1 = (x0 - x1) * cospi_16_64;
134       t2 = x2 * cospi_24_64 + x3 * cospi_8_64;
135       t3 = -x2 * cospi_8_64 + x3 * cospi_24_64;
136       out[0] = (tran_low_t)fdct_round_shift(t0);
137       out[2] = (tran_low_t)fdct_round_shift(t2);
138       out[4] = (tran_low_t)fdct_round_shift(t1);
139       out[6] = (tran_low_t)fdct_round_shift(t3);
140 
141       // Stage 2
142       t0 = (s6 - s5) * cospi_16_64;
143       t1 = (s6 + s5) * cospi_16_64;
144       t2 = fdct_round_shift(t0);
145       t3 = fdct_round_shift(t1);
146 
147       // Stage 3
148       x0 = s4 + t2;
149       x1 = s4 - t2;
150       x2 = s7 - t3;
151       x3 = s7 + t3;
152 
153       // Stage 4
154       t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
155       t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
156       t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
157       t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
158       out[1] = (tran_low_t)fdct_round_shift(t0);
159       out[3] = (tran_low_t)fdct_round_shift(t2);
160       out[5] = (tran_low_t)fdct_round_shift(t1);
161       out[7] = (tran_low_t)fdct_round_shift(t3);
162       out += 8;
163     }
164     in = intermediate;
165     out = output;
166   }
167 
168   // Rows
169   for (i = 0; i < 8; ++i) {
170     for (j = 0; j < 8; ++j) output[j + i * 8] /= 2;
171   }
172 }
173 
vpx_fdct8x8_1_c(const int16_t * input,tran_low_t * output,int stride)174 void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride) {
175   int r, c;
176   tran_low_t sum = 0;
177   for (r = 0; r < 8; ++r)
178     for (c = 0; c < 8; ++c) sum += input[r * stride + c];
179 
180   output[0] = sum;
181 }
182 
vpx_fdct16x16_c(const int16_t * input,tran_low_t * output,int stride)183 void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride) {
184   // The 2D transform is done with two passes which are actually pretty
185   // similar. In the first one, we transform the columns and transpose
186   // the results. In the second one, we transform the rows. To achieve that,
187   // as the first pass results are transposed, we transpose the columns (that
188   // is the transposed rows) and transpose the results (so that it goes back
189   // in normal/row positions).
190   int pass;
191   // We need an intermediate buffer between passes.
192   tran_low_t intermediate[256];
193   const tran_low_t *in_low = NULL;
194   tran_low_t *out = intermediate;
195   // Do the two transform/transpose passes
196   for (pass = 0; pass < 2; ++pass) {
197     tran_high_t step1[8];      // canbe16
198     tran_high_t step2[8];      // canbe16
199     tran_high_t step3[8];      // canbe16
200     tran_high_t in_high[8];    // canbe16
201     tran_high_t temp1, temp2;  // needs32
202     int i;
203     for (i = 0; i < 16; i++) {
204       if (0 == pass) {
205         // Calculate input for the first 8 results.
206         in_high[0] = (input[0 * stride] + input[15 * stride]) * 4;
207         in_high[1] = (input[1 * stride] + input[14 * stride]) * 4;
208         in_high[2] = (input[2 * stride] + input[13 * stride]) * 4;
209         in_high[3] = (input[3 * stride] + input[12 * stride]) * 4;
210         in_high[4] = (input[4 * stride] + input[11 * stride]) * 4;
211         in_high[5] = (input[5 * stride] + input[10 * stride]) * 4;
212         in_high[6] = (input[6 * stride] + input[9 * stride]) * 4;
213         in_high[7] = (input[7 * stride] + input[8 * stride]) * 4;
214         // Calculate input for the next 8 results.
215         step1[0] = (input[7 * stride] - input[8 * stride]) * 4;
216         step1[1] = (input[6 * stride] - input[9 * stride]) * 4;
217         step1[2] = (input[5 * stride] - input[10 * stride]) * 4;
218         step1[3] = (input[4 * stride] - input[11 * stride]) * 4;
219         step1[4] = (input[3 * stride] - input[12 * stride]) * 4;
220         step1[5] = (input[2 * stride] - input[13 * stride]) * 4;
221         step1[6] = (input[1 * stride] - input[14 * stride]) * 4;
222         step1[7] = (input[0 * stride] - input[15 * stride]) * 4;
223       } else {
224         // Calculate input for the first 8 results.
225         assert(in_low != NULL);
226         in_high[0] = ((in_low[0 * 16] + 1) >> 2) + ((in_low[15 * 16] + 1) >> 2);
227         in_high[1] = ((in_low[1 * 16] + 1) >> 2) + ((in_low[14 * 16] + 1) >> 2);
228         in_high[2] = ((in_low[2 * 16] + 1) >> 2) + ((in_low[13 * 16] + 1) >> 2);
229         in_high[3] = ((in_low[3 * 16] + 1) >> 2) + ((in_low[12 * 16] + 1) >> 2);
230         in_high[4] = ((in_low[4 * 16] + 1) >> 2) + ((in_low[11 * 16] + 1) >> 2);
231         in_high[5] = ((in_low[5 * 16] + 1) >> 2) + ((in_low[10 * 16] + 1) >> 2);
232         in_high[6] = ((in_low[6 * 16] + 1) >> 2) + ((in_low[9 * 16] + 1) >> 2);
233         in_high[7] = ((in_low[7 * 16] + 1) >> 2) + ((in_low[8 * 16] + 1) >> 2);
234         // Calculate input for the next 8 results.
235         step1[0] = ((in_low[7 * 16] + 1) >> 2) - ((in_low[8 * 16] + 1) >> 2);
236         step1[1] = ((in_low[6 * 16] + 1) >> 2) - ((in_low[9 * 16] + 1) >> 2);
237         step1[2] = ((in_low[5 * 16] + 1) >> 2) - ((in_low[10 * 16] + 1) >> 2);
238         step1[3] = ((in_low[4 * 16] + 1) >> 2) - ((in_low[11 * 16] + 1) >> 2);
239         step1[4] = ((in_low[3 * 16] + 1) >> 2) - ((in_low[12 * 16] + 1) >> 2);
240         step1[5] = ((in_low[2 * 16] + 1) >> 2) - ((in_low[13 * 16] + 1) >> 2);
241         step1[6] = ((in_low[1 * 16] + 1) >> 2) - ((in_low[14 * 16] + 1) >> 2);
242         step1[7] = ((in_low[0 * 16] + 1) >> 2) - ((in_low[15 * 16] + 1) >> 2);
243         in_low++;
244       }
245       // Work on the first eight values; fdct8(input, even_results);
246       {
247         tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
248         tran_high_t t0, t1, t2, t3;                  // needs32
249         tran_high_t x0, x1, x2, x3;                  // canbe16
250 
251         // stage 1
252         s0 = in_high[0] + in_high[7];
253         s1 = in_high[1] + in_high[6];
254         s2 = in_high[2] + in_high[5];
255         s3 = in_high[3] + in_high[4];
256         s4 = in_high[3] - in_high[4];
257         s5 = in_high[2] - in_high[5];
258         s6 = in_high[1] - in_high[6];
259         s7 = in_high[0] - in_high[7];
260 
261         // fdct4(step, step);
262         x0 = s0 + s3;
263         x1 = s1 + s2;
264         x2 = s1 - s2;
265         x3 = s0 - s3;
266         t0 = (x0 + x1) * cospi_16_64;
267         t1 = (x0 - x1) * cospi_16_64;
268         t2 = x3 * cospi_8_64 + x2 * cospi_24_64;
269         t3 = x3 * cospi_24_64 - x2 * cospi_8_64;
270         out[0] = (tran_low_t)fdct_round_shift(t0);
271         out[4] = (tran_low_t)fdct_round_shift(t2);
272         out[8] = (tran_low_t)fdct_round_shift(t1);
273         out[12] = (tran_low_t)fdct_round_shift(t3);
274 
275         // Stage 2
276         t0 = (s6 - s5) * cospi_16_64;
277         t1 = (s6 + s5) * cospi_16_64;
278         t2 = fdct_round_shift(t0);
279         t3 = fdct_round_shift(t1);
280 
281         // Stage 3
282         x0 = s4 + t2;
283         x1 = s4 - t2;
284         x2 = s7 - t3;
285         x3 = s7 + t3;
286 
287         // Stage 4
288         t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
289         t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
290         t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
291         t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
292         out[2] = (tran_low_t)fdct_round_shift(t0);
293         out[6] = (tran_low_t)fdct_round_shift(t2);
294         out[10] = (tran_low_t)fdct_round_shift(t1);
295         out[14] = (tran_low_t)fdct_round_shift(t3);
296       }
297       // Work on the next eight values; step1 -> odd_results
298       {
299         // step 2
300         temp1 = (step1[5] - step1[2]) * cospi_16_64;
301         temp2 = (step1[4] - step1[3]) * cospi_16_64;
302         step2[2] = fdct_round_shift(temp1);
303         step2[3] = fdct_round_shift(temp2);
304         temp1 = (step1[4] + step1[3]) * cospi_16_64;
305         temp2 = (step1[5] + step1[2]) * cospi_16_64;
306         step2[4] = fdct_round_shift(temp1);
307         step2[5] = fdct_round_shift(temp2);
308         // step 3
309         step3[0] = step1[0] + step2[3];
310         step3[1] = step1[1] + step2[2];
311         step3[2] = step1[1] - step2[2];
312         step3[3] = step1[0] - step2[3];
313         step3[4] = step1[7] - step2[4];
314         step3[5] = step1[6] - step2[5];
315         step3[6] = step1[6] + step2[5];
316         step3[7] = step1[7] + step2[4];
317         // step 4
318         temp1 = step3[1] * -cospi_8_64 + step3[6] * cospi_24_64;
319         temp2 = step3[2] * cospi_24_64 + step3[5] * cospi_8_64;
320         step2[1] = fdct_round_shift(temp1);
321         step2[2] = fdct_round_shift(temp2);
322         temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
323         temp2 = step3[1] * cospi_24_64 + step3[6] * cospi_8_64;
324         step2[5] = fdct_round_shift(temp1);
325         step2[6] = fdct_round_shift(temp2);
326         // step 5
327         step1[0] = step3[0] + step2[1];
328         step1[1] = step3[0] - step2[1];
329         step1[2] = step3[3] + step2[2];
330         step1[3] = step3[3] - step2[2];
331         step1[4] = step3[4] - step2[5];
332         step1[5] = step3[4] + step2[5];
333         step1[6] = step3[7] - step2[6];
334         step1[7] = step3[7] + step2[6];
335         // step 6
336         temp1 = step1[0] * cospi_30_64 + step1[7] * cospi_2_64;
337         temp2 = step1[1] * cospi_14_64 + step1[6] * cospi_18_64;
338         out[1] = (tran_low_t)fdct_round_shift(temp1);
339         out[9] = (tran_low_t)fdct_round_shift(temp2);
340         temp1 = step1[2] * cospi_22_64 + step1[5] * cospi_10_64;
341         temp2 = step1[3] * cospi_6_64 + step1[4] * cospi_26_64;
342         out[5] = (tran_low_t)fdct_round_shift(temp1);
343         out[13] = (tran_low_t)fdct_round_shift(temp2);
344         temp1 = step1[3] * -cospi_26_64 + step1[4] * cospi_6_64;
345         temp2 = step1[2] * -cospi_10_64 + step1[5] * cospi_22_64;
346         out[3] = (tran_low_t)fdct_round_shift(temp1);
347         out[11] = (tran_low_t)fdct_round_shift(temp2);
348         temp1 = step1[1] * -cospi_18_64 + step1[6] * cospi_14_64;
349         temp2 = step1[0] * -cospi_2_64 + step1[7] * cospi_30_64;
350         out[7] = (tran_low_t)fdct_round_shift(temp1);
351         out[15] = (tran_low_t)fdct_round_shift(temp2);
352       }
353       // Do next column (which is a transposed row in second/horizontal pass)
354       input++;
355       out += 16;
356     }
357     // Setup in/out for next pass.
358     in_low = intermediate;
359     out = output;
360   }
361 }
362 
vpx_fdct16x16_1_c(const int16_t * input,tran_low_t * output,int stride)363 void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride) {
364   int r, c;
365   int sum = 0;
366   for (r = 0; r < 16; ++r)
367     for (c = 0; c < 16; ++c) sum += input[r * stride + c];
368 
369   output[0] = (tran_low_t)(sum >> 1);
370 }
371 
dct_32_round(tran_high_t input)372 static INLINE tran_high_t dct_32_round(tran_high_t input) {
373   tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
374   // TODO(debargha, peter.derivaz): Find new bounds for this assert,
375   // and make the bounds consts.
376   // assert(-131072 <= rv && rv <= 131071);
377   return rv;
378 }
379 
half_round_shift(tran_high_t input)380 static INLINE tran_high_t half_round_shift(tran_high_t input) {
381   tran_high_t rv = (input + 1 + (input < 0)) >> 2;
382   return rv;
383 }
384 
vpx_fdct32(const tran_high_t * input,tran_high_t * output,int round)385 void vpx_fdct32(const tran_high_t *input, tran_high_t *output, int round) {
386   tran_high_t step[32];
387   // Stage 1
388   step[0] = input[0] + input[(32 - 1)];
389   step[1] = input[1] + input[(32 - 2)];
390   step[2] = input[2] + input[(32 - 3)];
391   step[3] = input[3] + input[(32 - 4)];
392   step[4] = input[4] + input[(32 - 5)];
393   step[5] = input[5] + input[(32 - 6)];
394   step[6] = input[6] + input[(32 - 7)];
395   step[7] = input[7] + input[(32 - 8)];
396   step[8] = input[8] + input[(32 - 9)];
397   step[9] = input[9] + input[(32 - 10)];
398   step[10] = input[10] + input[(32 - 11)];
399   step[11] = input[11] + input[(32 - 12)];
400   step[12] = input[12] + input[(32 - 13)];
401   step[13] = input[13] + input[(32 - 14)];
402   step[14] = input[14] + input[(32 - 15)];
403   step[15] = input[15] + input[(32 - 16)];
404   step[16] = -input[16] + input[(32 - 17)];
405   step[17] = -input[17] + input[(32 - 18)];
406   step[18] = -input[18] + input[(32 - 19)];
407   step[19] = -input[19] + input[(32 - 20)];
408   step[20] = -input[20] + input[(32 - 21)];
409   step[21] = -input[21] + input[(32 - 22)];
410   step[22] = -input[22] + input[(32 - 23)];
411   step[23] = -input[23] + input[(32 - 24)];
412   step[24] = -input[24] + input[(32 - 25)];
413   step[25] = -input[25] + input[(32 - 26)];
414   step[26] = -input[26] + input[(32 - 27)];
415   step[27] = -input[27] + input[(32 - 28)];
416   step[28] = -input[28] + input[(32 - 29)];
417   step[29] = -input[29] + input[(32 - 30)];
418   step[30] = -input[30] + input[(32 - 31)];
419   step[31] = -input[31] + input[(32 - 32)];
420 
421   // Stage 2
422   output[0] = step[0] + step[16 - 1];
423   output[1] = step[1] + step[16 - 2];
424   output[2] = step[2] + step[16 - 3];
425   output[3] = step[3] + step[16 - 4];
426   output[4] = step[4] + step[16 - 5];
427   output[5] = step[5] + step[16 - 6];
428   output[6] = step[6] + step[16 - 7];
429   output[7] = step[7] + step[16 - 8];
430   output[8] = -step[8] + step[16 - 9];
431   output[9] = -step[9] + step[16 - 10];
432   output[10] = -step[10] + step[16 - 11];
433   output[11] = -step[11] + step[16 - 12];
434   output[12] = -step[12] + step[16 - 13];
435   output[13] = -step[13] + step[16 - 14];
436   output[14] = -step[14] + step[16 - 15];
437   output[15] = -step[15] + step[16 - 16];
438 
439   output[16] = step[16];
440   output[17] = step[17];
441   output[18] = step[18];
442   output[19] = step[19];
443 
444   output[20] = dct_32_round((-step[20] + step[27]) * cospi_16_64);
445   output[21] = dct_32_round((-step[21] + step[26]) * cospi_16_64);
446   output[22] = dct_32_round((-step[22] + step[25]) * cospi_16_64);
447   output[23] = dct_32_round((-step[23] + step[24]) * cospi_16_64);
448 
449   output[24] = dct_32_round((step[24] + step[23]) * cospi_16_64);
450   output[25] = dct_32_round((step[25] + step[22]) * cospi_16_64);
451   output[26] = dct_32_round((step[26] + step[21]) * cospi_16_64);
452   output[27] = dct_32_round((step[27] + step[20]) * cospi_16_64);
453 
454   output[28] = step[28];
455   output[29] = step[29];
456   output[30] = step[30];
457   output[31] = step[31];
458 
459   // dump the magnitude by 4, hence the intermediate values are within
460   // the range of 16 bits.
461   if (round) {
462     output[0] = half_round_shift(output[0]);
463     output[1] = half_round_shift(output[1]);
464     output[2] = half_round_shift(output[2]);
465     output[3] = half_round_shift(output[3]);
466     output[4] = half_round_shift(output[4]);
467     output[5] = half_round_shift(output[5]);
468     output[6] = half_round_shift(output[6]);
469     output[7] = half_round_shift(output[7]);
470     output[8] = half_round_shift(output[8]);
471     output[9] = half_round_shift(output[9]);
472     output[10] = half_round_shift(output[10]);
473     output[11] = half_round_shift(output[11]);
474     output[12] = half_round_shift(output[12]);
475     output[13] = half_round_shift(output[13]);
476     output[14] = half_round_shift(output[14]);
477     output[15] = half_round_shift(output[15]);
478 
479     output[16] = half_round_shift(output[16]);
480     output[17] = half_round_shift(output[17]);
481     output[18] = half_round_shift(output[18]);
482     output[19] = half_round_shift(output[19]);
483     output[20] = half_round_shift(output[20]);
484     output[21] = half_round_shift(output[21]);
485     output[22] = half_round_shift(output[22]);
486     output[23] = half_round_shift(output[23]);
487     output[24] = half_round_shift(output[24]);
488     output[25] = half_round_shift(output[25]);
489     output[26] = half_round_shift(output[26]);
490     output[27] = half_round_shift(output[27]);
491     output[28] = half_round_shift(output[28]);
492     output[29] = half_round_shift(output[29]);
493     output[30] = half_round_shift(output[30]);
494     output[31] = half_round_shift(output[31]);
495   }
496 
497   // Stage 3
498   step[0] = output[0] + output[(8 - 1)];
499   step[1] = output[1] + output[(8 - 2)];
500   step[2] = output[2] + output[(8 - 3)];
501   step[3] = output[3] + output[(8 - 4)];
502   step[4] = -output[4] + output[(8 - 5)];
503   step[5] = -output[5] + output[(8 - 6)];
504   step[6] = -output[6] + output[(8 - 7)];
505   step[7] = -output[7] + output[(8 - 8)];
506   step[8] = output[8];
507   step[9] = output[9];
508   step[10] = dct_32_round((-output[10] + output[13]) * cospi_16_64);
509   step[11] = dct_32_round((-output[11] + output[12]) * cospi_16_64);
510   step[12] = dct_32_round((output[12] + output[11]) * cospi_16_64);
511   step[13] = dct_32_round((output[13] + output[10]) * cospi_16_64);
512   step[14] = output[14];
513   step[15] = output[15];
514 
515   step[16] = output[16] + output[23];
516   step[17] = output[17] + output[22];
517   step[18] = output[18] + output[21];
518   step[19] = output[19] + output[20];
519   step[20] = -output[20] + output[19];
520   step[21] = -output[21] + output[18];
521   step[22] = -output[22] + output[17];
522   step[23] = -output[23] + output[16];
523   step[24] = -output[24] + output[31];
524   step[25] = -output[25] + output[30];
525   step[26] = -output[26] + output[29];
526   step[27] = -output[27] + output[28];
527   step[28] = output[28] + output[27];
528   step[29] = output[29] + output[26];
529   step[30] = output[30] + output[25];
530   step[31] = output[31] + output[24];
531 
532   // Stage 4
533   output[0] = step[0] + step[3];
534   output[1] = step[1] + step[2];
535   output[2] = -step[2] + step[1];
536   output[3] = -step[3] + step[0];
537   output[4] = step[4];
538   output[5] = dct_32_round((-step[5] + step[6]) * cospi_16_64);
539   output[6] = dct_32_round((step[6] + step[5]) * cospi_16_64);
540   output[7] = step[7];
541   output[8] = step[8] + step[11];
542   output[9] = step[9] + step[10];
543   output[10] = -step[10] + step[9];
544   output[11] = -step[11] + step[8];
545   output[12] = -step[12] + step[15];
546   output[13] = -step[13] + step[14];
547   output[14] = step[14] + step[13];
548   output[15] = step[15] + step[12];
549 
550   output[16] = step[16];
551   output[17] = step[17];
552   output[18] = dct_32_round(step[18] * -cospi_8_64 + step[29] * cospi_24_64);
553   output[19] = dct_32_round(step[19] * -cospi_8_64 + step[28] * cospi_24_64);
554   output[20] = dct_32_round(step[20] * -cospi_24_64 + step[27] * -cospi_8_64);
555   output[21] = dct_32_round(step[21] * -cospi_24_64 + step[26] * -cospi_8_64);
556   output[22] = step[22];
557   output[23] = step[23];
558   output[24] = step[24];
559   output[25] = step[25];
560   output[26] = dct_32_round(step[26] * cospi_24_64 + step[21] * -cospi_8_64);
561   output[27] = dct_32_round(step[27] * cospi_24_64 + step[20] * -cospi_8_64);
562   output[28] = dct_32_round(step[28] * cospi_8_64 + step[19] * cospi_24_64);
563   output[29] = dct_32_round(step[29] * cospi_8_64 + step[18] * cospi_24_64);
564   output[30] = step[30];
565   output[31] = step[31];
566 
567   // Stage 5
568   step[0] = dct_32_round((output[0] + output[1]) * cospi_16_64);
569   step[1] = dct_32_round((-output[1] + output[0]) * cospi_16_64);
570   step[2] = dct_32_round(output[2] * cospi_24_64 + output[3] * cospi_8_64);
571   step[3] = dct_32_round(output[3] * cospi_24_64 - output[2] * cospi_8_64);
572   step[4] = output[4] + output[5];
573   step[5] = -output[5] + output[4];
574   step[6] = -output[6] + output[7];
575   step[7] = output[7] + output[6];
576   step[8] = output[8];
577   step[9] = dct_32_round(output[9] * -cospi_8_64 + output[14] * cospi_24_64);
578   step[10] = dct_32_round(output[10] * -cospi_24_64 + output[13] * -cospi_8_64);
579   step[11] = output[11];
580   step[12] = output[12];
581   step[13] = dct_32_round(output[13] * cospi_24_64 + output[10] * -cospi_8_64);
582   step[14] = dct_32_round(output[14] * cospi_8_64 + output[9] * cospi_24_64);
583   step[15] = output[15];
584 
585   step[16] = output[16] + output[19];
586   step[17] = output[17] + output[18];
587   step[18] = -output[18] + output[17];
588   step[19] = -output[19] + output[16];
589   step[20] = -output[20] + output[23];
590   step[21] = -output[21] + output[22];
591   step[22] = output[22] + output[21];
592   step[23] = output[23] + output[20];
593   step[24] = output[24] + output[27];
594   step[25] = output[25] + output[26];
595   step[26] = -output[26] + output[25];
596   step[27] = -output[27] + output[24];
597   step[28] = -output[28] + output[31];
598   step[29] = -output[29] + output[30];
599   step[30] = output[30] + output[29];
600   step[31] = output[31] + output[28];
601 
602   // Stage 6
603   output[0] = step[0];
604   output[1] = step[1];
605   output[2] = step[2];
606   output[3] = step[3];
607   output[4] = dct_32_round(step[4] * cospi_28_64 + step[7] * cospi_4_64);
608   output[5] = dct_32_round(step[5] * cospi_12_64 + step[6] * cospi_20_64);
609   output[6] = dct_32_round(step[6] * cospi_12_64 + step[5] * -cospi_20_64);
610   output[7] = dct_32_round(step[7] * cospi_28_64 + step[4] * -cospi_4_64);
611   output[8] = step[8] + step[9];
612   output[9] = -step[9] + step[8];
613   output[10] = -step[10] + step[11];
614   output[11] = step[11] + step[10];
615   output[12] = step[12] + step[13];
616   output[13] = -step[13] + step[12];
617   output[14] = -step[14] + step[15];
618   output[15] = step[15] + step[14];
619 
620   output[16] = step[16];
621   output[17] = dct_32_round(step[17] * -cospi_4_64 + step[30] * cospi_28_64);
622   output[18] = dct_32_round(step[18] * -cospi_28_64 + step[29] * -cospi_4_64);
623   output[19] = step[19];
624   output[20] = step[20];
625   output[21] = dct_32_round(step[21] * -cospi_20_64 + step[26] * cospi_12_64);
626   output[22] = dct_32_round(step[22] * -cospi_12_64 + step[25] * -cospi_20_64);
627   output[23] = step[23];
628   output[24] = step[24];
629   output[25] = dct_32_round(step[25] * cospi_12_64 + step[22] * -cospi_20_64);
630   output[26] = dct_32_round(step[26] * cospi_20_64 + step[21] * cospi_12_64);
631   output[27] = step[27];
632   output[28] = step[28];
633   output[29] = dct_32_round(step[29] * cospi_28_64 + step[18] * -cospi_4_64);
634   output[30] = dct_32_round(step[30] * cospi_4_64 + step[17] * cospi_28_64);
635   output[31] = step[31];
636 
637   // Stage 7
638   step[0] = output[0];
639   step[1] = output[1];
640   step[2] = output[2];
641   step[3] = output[3];
642   step[4] = output[4];
643   step[5] = output[5];
644   step[6] = output[6];
645   step[7] = output[7];
646   step[8] = dct_32_round(output[8] * cospi_30_64 + output[15] * cospi_2_64);
647   step[9] = dct_32_round(output[9] * cospi_14_64 + output[14] * cospi_18_64);
648   step[10] = dct_32_round(output[10] * cospi_22_64 + output[13] * cospi_10_64);
649   step[11] = dct_32_round(output[11] * cospi_6_64 + output[12] * cospi_26_64);
650   step[12] = dct_32_round(output[12] * cospi_6_64 + output[11] * -cospi_26_64);
651   step[13] = dct_32_round(output[13] * cospi_22_64 + output[10] * -cospi_10_64);
652   step[14] = dct_32_round(output[14] * cospi_14_64 + output[9] * -cospi_18_64);
653   step[15] = dct_32_round(output[15] * cospi_30_64 + output[8] * -cospi_2_64);
654 
655   step[16] = output[16] + output[17];
656   step[17] = -output[17] + output[16];
657   step[18] = -output[18] + output[19];
658   step[19] = output[19] + output[18];
659   step[20] = output[20] + output[21];
660   step[21] = -output[21] + output[20];
661   step[22] = -output[22] + output[23];
662   step[23] = output[23] + output[22];
663   step[24] = output[24] + output[25];
664   step[25] = -output[25] + output[24];
665   step[26] = -output[26] + output[27];
666   step[27] = output[27] + output[26];
667   step[28] = output[28] + output[29];
668   step[29] = -output[29] + output[28];
669   step[30] = -output[30] + output[31];
670   step[31] = output[31] + output[30];
671 
672   // Final stage --- outputs indices are bit-reversed.
673   output[0] = step[0];
674   output[16] = step[1];
675   output[8] = step[2];
676   output[24] = step[3];
677   output[4] = step[4];
678   output[20] = step[5];
679   output[12] = step[6];
680   output[28] = step[7];
681   output[2] = step[8];
682   output[18] = step[9];
683   output[10] = step[10];
684   output[26] = step[11];
685   output[6] = step[12];
686   output[22] = step[13];
687   output[14] = step[14];
688   output[30] = step[15];
689 
690   output[1] = dct_32_round(step[16] * cospi_31_64 + step[31] * cospi_1_64);
691   output[17] = dct_32_round(step[17] * cospi_15_64 + step[30] * cospi_17_64);
692   output[9] = dct_32_round(step[18] * cospi_23_64 + step[29] * cospi_9_64);
693   output[25] = dct_32_round(step[19] * cospi_7_64 + step[28] * cospi_25_64);
694   output[5] = dct_32_round(step[20] * cospi_27_64 + step[27] * cospi_5_64);
695   output[21] = dct_32_round(step[21] * cospi_11_64 + step[26] * cospi_21_64);
696   output[13] = dct_32_round(step[22] * cospi_19_64 + step[25] * cospi_13_64);
697   output[29] = dct_32_round(step[23] * cospi_3_64 + step[24] * cospi_29_64);
698   output[3] = dct_32_round(step[24] * cospi_3_64 + step[23] * -cospi_29_64);
699   output[19] = dct_32_round(step[25] * cospi_19_64 + step[22] * -cospi_13_64);
700   output[11] = dct_32_round(step[26] * cospi_11_64 + step[21] * -cospi_21_64);
701   output[27] = dct_32_round(step[27] * cospi_27_64 + step[20] * -cospi_5_64);
702   output[7] = dct_32_round(step[28] * cospi_7_64 + step[19] * -cospi_25_64);
703   output[23] = dct_32_round(step[29] * cospi_23_64 + step[18] * -cospi_9_64);
704   output[15] = dct_32_round(step[30] * cospi_15_64 + step[17] * -cospi_17_64);
705   output[31] = dct_32_round(step[31] * cospi_31_64 + step[16] * -cospi_1_64);
706 }
707 
vpx_fdct32x32_c(const int16_t * input,tran_low_t * output,int stride)708 void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride) {
709   int i, j;
710   tran_high_t out[32 * 32];
711 
712   // Columns
713   for (i = 0; i < 32; ++i) {
714     tran_high_t temp_in[32], temp_out[32];
715     for (j = 0; j < 32; ++j) temp_in[j] = input[j * stride + i] * 4;
716     vpx_fdct32(temp_in, temp_out, 0);
717     for (j = 0; j < 32; ++j)
718       out[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
719   }
720 
721   // Rows
722   for (i = 0; i < 32; ++i) {
723     tran_high_t temp_in[32], temp_out[32];
724     for (j = 0; j < 32; ++j) temp_in[j] = out[j + i * 32];
725     vpx_fdct32(temp_in, temp_out, 0);
726     for (j = 0; j < 32; ++j)
727       output[j + i * 32] =
728           (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
729   }
730 }
731 
732 // Note that although we use dct_32_round in dct32 computation flow,
733 // this 2d fdct32x32 for rate-distortion optimization loop is operating
734 // within 16 bits precision.
vpx_fdct32x32_rd_c(const int16_t * input,tran_low_t * output,int stride)735 void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride) {
736   int i, j;
737   tran_high_t out[32 * 32];
738 
739   // Columns
740   for (i = 0; i < 32; ++i) {
741     tran_high_t temp_in[32], temp_out[32];
742     for (j = 0; j < 32; ++j) temp_in[j] = input[j * stride + i] * 4;
743     vpx_fdct32(temp_in, temp_out, 0);
744     for (j = 0; j < 32; ++j)
745       // TODO(cd): see quality impact of only doing
746       //           output[j * 32 + i] = (temp_out[j] + 1) >> 2;
747       //           PS: also change code in vpx_dsp/x86/vpx_dct_sse2.c
748       out[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
749   }
750 
751   // Rows
752   for (i = 0; i < 32; ++i) {
753     tran_high_t temp_in[32], temp_out[32];
754     for (j = 0; j < 32; ++j) temp_in[j] = out[j + i * 32];
755     vpx_fdct32(temp_in, temp_out, 1);
756     for (j = 0; j < 32; ++j) output[j + i * 32] = (tran_low_t)temp_out[j];
757   }
758 }
759 
vpx_fdct32x32_1_c(const int16_t * input,tran_low_t * output,int stride)760 void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride) {
761   int r, c;
762   int sum = 0;
763   for (r = 0; r < 32; ++r)
764     for (c = 0; c < 32; ++c) sum += input[r * stride + c];
765 
766   output[0] = (tran_low_t)(sum >> 3);
767 }
768 
769 #if CONFIG_VP9_HIGHBITDEPTH
vpx_highbd_fdct4x4_c(const int16_t * input,tran_low_t * output,int stride)770 void vpx_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output,
771                           int stride) {
772   vpx_fdct4x4_c(input, output, stride);
773 }
774 
vpx_highbd_fdct8x8_c(const int16_t * input,tran_low_t * output,int stride)775 void vpx_highbd_fdct8x8_c(const int16_t *input, tran_low_t *output,
776                           int stride) {
777   vpx_fdct8x8_c(input, output, stride);
778 }
779 
vpx_highbd_fdct8x8_1_c(const int16_t * input,tran_low_t * output,int stride)780 void vpx_highbd_fdct8x8_1_c(const int16_t *input, tran_low_t *output,
781                             int stride) {
782   vpx_fdct8x8_1_c(input, output, stride);
783 }
784 
vpx_highbd_fdct16x16_c(const int16_t * input,tran_low_t * output,int stride)785 void vpx_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output,
786                             int stride) {
787   vpx_fdct16x16_c(input, output, stride);
788 }
789 
vpx_highbd_fdct16x16_1_c(const int16_t * input,tran_low_t * output,int stride)790 void vpx_highbd_fdct16x16_1_c(const int16_t *input, tran_low_t *output,
791                               int stride) {
792   vpx_fdct16x16_1_c(input, output, stride);
793 }
794 
vpx_highbd_fdct32x32_c(const int16_t * input,tran_low_t * output,int stride)795 void vpx_highbd_fdct32x32_c(const int16_t *input, tran_low_t *output,
796                             int stride) {
797   vpx_fdct32x32_c(input, output, stride);
798 }
799 
vpx_highbd_fdct32x32_rd_c(const int16_t * input,tran_low_t * output,int stride)800 void vpx_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *output,
801                                int stride) {
802   vpx_fdct32x32_rd_c(input, output, stride);
803 }
804 
vpx_highbd_fdct32x32_1_c(const int16_t * input,tran_low_t * output,int stride)805 void vpx_highbd_fdct32x32_1_c(const int16_t *input, tran_low_t *output,
806                               int stride) {
807   vpx_fdct32x32_1_c(input, output, stride);
808 }
809 #endif  // CONFIG_VP9_HIGHBITDEPTH
810