1/*
2 *  Copyright (c) 2012 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 <emmintrin.h>  // SSE2
12#include "vp9/common/vp9_idct.h"  // for cospi constants
13#include "vpx_ports/mem.h"
14
15#if FDCT32x32_HIGH_PRECISION
16static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
17  __m128i buf0, buf1;
18  buf0 = _mm_mul_epu32(a, b);
19  a = _mm_srli_epi64(a, 32);
20  b = _mm_srli_epi64(b, 32);
21  buf1 = _mm_mul_epu32(a, b);
22  return _mm_add_epi64(buf0, buf1);
23}
24
25static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) {
26  __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
27  __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
28  return _mm_unpacklo_epi64(buf0, buf1);
29}
30#endif
31
32void FDCT32x32_2D(const int16_t *input,
33                  int16_t *output_org, int stride) {
34  // Calculate pre-multiplied strides
35  const int str1 = stride;
36  const int str2 = 2 * stride;
37  const int str3 = 2 * stride + str1;
38  // We need an intermediate buffer between passes.
39  DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]);
40  // Constants
41  //    When we use them, in one case, they are all the same. In all others
42  //    it's a pair of them that we need to repeat four times. This is done
43  //    by constructing the 32 bit constant corresponding to that pair.
44  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(+cospi_16_64);
45  const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
46  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64,   cospi_24_64);
47  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
48  const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64,  cospi_8_64);
49  const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64,  cospi_20_64);
50  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64,  cospi_12_64);
51  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64,   cospi_28_64);
52  const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64,  cospi_4_64);
53  const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
54  const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
55  const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64,  cospi_2_64);
56  const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64,  cospi_18_64);
57  const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64,  cospi_10_64);
58  const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64,   cospi_26_64);
59  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64,  cospi_6_64);
60  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64,  cospi_22_64);
61  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64,  cospi_14_64);
62  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64,   cospi_30_64);
63  const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64,  cospi_1_64);
64  const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64,  cospi_17_64);
65  const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64,  cospi_9_64);
66  const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64,   cospi_25_64);
67  const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64,  cospi_7_64);
68  const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64,   cospi_23_64);
69  const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64,  cospi_15_64);
70  const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64,   cospi_31_64);
71  const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64,  cospi_5_64);
72  const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64,  cospi_21_64);
73  const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64,  cospi_13_64);
74  const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64,   cospi_29_64);
75  const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64,  cospi_3_64);
76  const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64,  cospi_19_64);
77  const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64,  cospi_11_64);
78  const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64,   cospi_27_64);
79  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
80  const __m128i kZero = _mm_set1_epi16(0);
81  const __m128i kOne  = _mm_set1_epi16(1);
82  // Do the two transform/transpose passes
83  int pass;
84  for (pass = 0; pass < 2; ++pass) {
85    // We process eight columns (transposed rows in second pass) at a time.
86    int column_start;
87    for (column_start = 0; column_start < 32; column_start += 8) {
88      __m128i step1[32];
89      __m128i step2[32];
90      __m128i step3[32];
91      __m128i out[32];
92      // Stage 1
93      // Note: even though all the loads below are aligned, using the aligned
94      //       intrinsic make the code slightly slower.
95      if (0 == pass) {
96        const int16_t *in  = &input[column_start];
97        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
98        // Note: the next four blocks could be in a loop. That would help the
99        //       instruction cache but is actually slower.
100        {
101          const int16_t *ina =  in +  0 * str1;
102          const int16_t *inb =  in + 31 * str1;
103          __m128i *step1a = &step1[ 0];
104          __m128i *step1b = &step1[31];
105          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
106          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
107          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
108          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
109          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
110          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
111          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
112          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
113          step1a[ 0] = _mm_add_epi16(ina0, inb0);
114          step1a[ 1] = _mm_add_epi16(ina1, inb1);
115          step1a[ 2] = _mm_add_epi16(ina2, inb2);
116          step1a[ 3] = _mm_add_epi16(ina3, inb3);
117          step1b[-3] = _mm_sub_epi16(ina3, inb3);
118          step1b[-2] = _mm_sub_epi16(ina2, inb2);
119          step1b[-1] = _mm_sub_epi16(ina1, inb1);
120          step1b[-0] = _mm_sub_epi16(ina0, inb0);
121          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
122          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
123          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
124          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
125          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
126          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
127          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
128          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
129        }
130        {
131          const int16_t *ina =  in +  4 * str1;
132          const int16_t *inb =  in + 27 * str1;
133          __m128i *step1a = &step1[ 4];
134          __m128i *step1b = &step1[27];
135          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
136          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
137          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
138          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
139          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
140          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
141          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
142          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
143          step1a[ 0] = _mm_add_epi16(ina0, inb0);
144          step1a[ 1] = _mm_add_epi16(ina1, inb1);
145          step1a[ 2] = _mm_add_epi16(ina2, inb2);
146          step1a[ 3] = _mm_add_epi16(ina3, inb3);
147          step1b[-3] = _mm_sub_epi16(ina3, inb3);
148          step1b[-2] = _mm_sub_epi16(ina2, inb2);
149          step1b[-1] = _mm_sub_epi16(ina1, inb1);
150          step1b[-0] = _mm_sub_epi16(ina0, inb0);
151          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
152          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
153          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
154          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
155          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
156          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
157          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
158          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
159        }
160        {
161          const int16_t *ina =  in +  8 * str1;
162          const int16_t *inb =  in + 23 * str1;
163          __m128i *step1a = &step1[ 8];
164          __m128i *step1b = &step1[23];
165          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
166          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
167          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
168          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
169          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
170          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
171          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
172          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
173          step1a[ 0] = _mm_add_epi16(ina0, inb0);
174          step1a[ 1] = _mm_add_epi16(ina1, inb1);
175          step1a[ 2] = _mm_add_epi16(ina2, inb2);
176          step1a[ 3] = _mm_add_epi16(ina3, inb3);
177          step1b[-3] = _mm_sub_epi16(ina3, inb3);
178          step1b[-2] = _mm_sub_epi16(ina2, inb2);
179          step1b[-1] = _mm_sub_epi16(ina1, inb1);
180          step1b[-0] = _mm_sub_epi16(ina0, inb0);
181          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
182          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
183          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
184          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
185          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
186          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
187          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
188          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
189        }
190        {
191          const int16_t *ina =  in + 12 * str1;
192          const int16_t *inb =  in + 19 * str1;
193          __m128i *step1a = &step1[12];
194          __m128i *step1b = &step1[19];
195          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
196          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
197          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
198          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
199          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
200          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
201          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
202          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
203          step1a[ 0] = _mm_add_epi16(ina0, inb0);
204          step1a[ 1] = _mm_add_epi16(ina1, inb1);
205          step1a[ 2] = _mm_add_epi16(ina2, inb2);
206          step1a[ 3] = _mm_add_epi16(ina3, inb3);
207          step1b[-3] = _mm_sub_epi16(ina3, inb3);
208          step1b[-2] = _mm_sub_epi16(ina2, inb2);
209          step1b[-1] = _mm_sub_epi16(ina1, inb1);
210          step1b[-0] = _mm_sub_epi16(ina0, inb0);
211          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
212          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
213          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
214          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
215          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
216          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
217          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
218          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
219        }
220      } else {
221        int16_t *in = &intermediate[column_start];
222        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
223        // Note: using the same approach as above to have common offset is
224        //       counter-productive as all offsets can be calculated at compile
225        //       time.
226        // Note: the next four blocks could be in a loop. That would help the
227        //       instruction cache but is actually slower.
228        {
229          __m128i in00  = _mm_loadu_si128((const __m128i *)(in +  0 * 32));
230          __m128i in01  = _mm_loadu_si128((const __m128i *)(in +  1 * 32));
231          __m128i in02  = _mm_loadu_si128((const __m128i *)(in +  2 * 32));
232          __m128i in03  = _mm_loadu_si128((const __m128i *)(in +  3 * 32));
233          __m128i in28  = _mm_loadu_si128((const __m128i *)(in + 28 * 32));
234          __m128i in29  = _mm_loadu_si128((const __m128i *)(in + 29 * 32));
235          __m128i in30  = _mm_loadu_si128((const __m128i *)(in + 30 * 32));
236          __m128i in31  = _mm_loadu_si128((const __m128i *)(in + 31 * 32));
237          step1[ 0] = _mm_add_epi16(in00, in31);
238          step1[ 1] = _mm_add_epi16(in01, in30);
239          step1[ 2] = _mm_add_epi16(in02, in29);
240          step1[ 3] = _mm_add_epi16(in03, in28);
241          step1[28] = _mm_sub_epi16(in03, in28);
242          step1[29] = _mm_sub_epi16(in02, in29);
243          step1[30] = _mm_sub_epi16(in01, in30);
244          step1[31] = _mm_sub_epi16(in00, in31);
245        }
246        {
247          __m128i in04  = _mm_loadu_si128((const __m128i *)(in +  4 * 32));
248          __m128i in05  = _mm_loadu_si128((const __m128i *)(in +  5 * 32));
249          __m128i in06  = _mm_loadu_si128((const __m128i *)(in +  6 * 32));
250          __m128i in07  = _mm_loadu_si128((const __m128i *)(in +  7 * 32));
251          __m128i in24  = _mm_loadu_si128((const __m128i *)(in + 24 * 32));
252          __m128i in25  = _mm_loadu_si128((const __m128i *)(in + 25 * 32));
253          __m128i in26  = _mm_loadu_si128((const __m128i *)(in + 26 * 32));
254          __m128i in27  = _mm_loadu_si128((const __m128i *)(in + 27 * 32));
255          step1[ 4] = _mm_add_epi16(in04, in27);
256          step1[ 5] = _mm_add_epi16(in05, in26);
257          step1[ 6] = _mm_add_epi16(in06, in25);
258          step1[ 7] = _mm_add_epi16(in07, in24);
259          step1[24] = _mm_sub_epi16(in07, in24);
260          step1[25] = _mm_sub_epi16(in06, in25);
261          step1[26] = _mm_sub_epi16(in05, in26);
262          step1[27] = _mm_sub_epi16(in04, in27);
263        }
264        {
265          __m128i in08  = _mm_loadu_si128((const __m128i *)(in +  8 * 32));
266          __m128i in09  = _mm_loadu_si128((const __m128i *)(in +  9 * 32));
267          __m128i in10  = _mm_loadu_si128((const __m128i *)(in + 10 * 32));
268          __m128i in11  = _mm_loadu_si128((const __m128i *)(in + 11 * 32));
269          __m128i in20  = _mm_loadu_si128((const __m128i *)(in + 20 * 32));
270          __m128i in21  = _mm_loadu_si128((const __m128i *)(in + 21 * 32));
271          __m128i in22  = _mm_loadu_si128((const __m128i *)(in + 22 * 32));
272          __m128i in23  = _mm_loadu_si128((const __m128i *)(in + 23 * 32));
273          step1[ 8] = _mm_add_epi16(in08, in23);
274          step1[ 9] = _mm_add_epi16(in09, in22);
275          step1[10] = _mm_add_epi16(in10, in21);
276          step1[11] = _mm_add_epi16(in11, in20);
277          step1[20] = _mm_sub_epi16(in11, in20);
278          step1[21] = _mm_sub_epi16(in10, in21);
279          step1[22] = _mm_sub_epi16(in09, in22);
280          step1[23] = _mm_sub_epi16(in08, in23);
281        }
282        {
283          __m128i in12  = _mm_loadu_si128((const __m128i *)(in + 12 * 32));
284          __m128i in13  = _mm_loadu_si128((const __m128i *)(in + 13 * 32));
285          __m128i in14  = _mm_loadu_si128((const __m128i *)(in + 14 * 32));
286          __m128i in15  = _mm_loadu_si128((const __m128i *)(in + 15 * 32));
287          __m128i in16  = _mm_loadu_si128((const __m128i *)(in + 16 * 32));
288          __m128i in17  = _mm_loadu_si128((const __m128i *)(in + 17 * 32));
289          __m128i in18  = _mm_loadu_si128((const __m128i *)(in + 18 * 32));
290          __m128i in19  = _mm_loadu_si128((const __m128i *)(in + 19 * 32));
291          step1[12] = _mm_add_epi16(in12, in19);
292          step1[13] = _mm_add_epi16(in13, in18);
293          step1[14] = _mm_add_epi16(in14, in17);
294          step1[15] = _mm_add_epi16(in15, in16);
295          step1[16] = _mm_sub_epi16(in15, in16);
296          step1[17] = _mm_sub_epi16(in14, in17);
297          step1[18] = _mm_sub_epi16(in13, in18);
298          step1[19] = _mm_sub_epi16(in12, in19);
299        }
300      }
301      // Stage 2
302      {
303        step2[ 0] = _mm_add_epi16(step1[0], step1[15]);
304        step2[ 1] = _mm_add_epi16(step1[1], step1[14]);
305        step2[ 2] = _mm_add_epi16(step1[2], step1[13]);
306        step2[ 3] = _mm_add_epi16(step1[3], step1[12]);
307        step2[ 4] = _mm_add_epi16(step1[4], step1[11]);
308        step2[ 5] = _mm_add_epi16(step1[5], step1[10]);
309        step2[ 6] = _mm_add_epi16(step1[6], step1[ 9]);
310        step2[ 7] = _mm_add_epi16(step1[7], step1[ 8]);
311        step2[ 8] = _mm_sub_epi16(step1[7], step1[ 8]);
312        step2[ 9] = _mm_sub_epi16(step1[6], step1[ 9]);
313        step2[10] = _mm_sub_epi16(step1[5], step1[10]);
314        step2[11] = _mm_sub_epi16(step1[4], step1[11]);
315        step2[12] = _mm_sub_epi16(step1[3], step1[12]);
316        step2[13] = _mm_sub_epi16(step1[2], step1[13]);
317        step2[14] = _mm_sub_epi16(step1[1], step1[14]);
318        step2[15] = _mm_sub_epi16(step1[0], step1[15]);
319      }
320      {
321        const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
322        const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
323        const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
324        const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
325        const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
326        const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
327        const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
328        const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
329        const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
330        const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
331        const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
332        const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
333        const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
334        const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
335        const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
336        const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
337        const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
338        const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
339        const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
340        const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
341        const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
342        const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
343        const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
344        const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
345        // dct_const_round_shift
346        const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
347        const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
348        const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
349        const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
350        const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
351        const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
352        const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
353        const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
354        const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
355        const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
356        const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
357        const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
358        const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
359        const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
360        const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
361        const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
362        const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
363        const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
364        const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
365        const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
366        const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
367        const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
368        const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
369        const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
370        const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
371        const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
372        const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
373        const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
374        const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
375        const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
376        const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
377        const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
378        // Combine
379        step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
380        step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
381        step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
382        step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
383        step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
384        step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
385        step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
386        step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
387      }
388
389#if !FDCT32x32_HIGH_PRECISION
390      // dump the magnitude by half, hence the intermediate values are within
391      // the range of 16 bits.
392      if (1 == pass) {
393        __m128i s3_00_0 = _mm_cmplt_epi16(step2[ 0], kZero);
394        __m128i s3_01_0 = _mm_cmplt_epi16(step2[ 1], kZero);
395        __m128i s3_02_0 = _mm_cmplt_epi16(step2[ 2], kZero);
396        __m128i s3_03_0 = _mm_cmplt_epi16(step2[ 3], kZero);
397        __m128i s3_04_0 = _mm_cmplt_epi16(step2[ 4], kZero);
398        __m128i s3_05_0 = _mm_cmplt_epi16(step2[ 5], kZero);
399        __m128i s3_06_0 = _mm_cmplt_epi16(step2[ 6], kZero);
400        __m128i s3_07_0 = _mm_cmplt_epi16(step2[ 7], kZero);
401        __m128i s2_08_0 = _mm_cmplt_epi16(step2[ 8], kZero);
402        __m128i s2_09_0 = _mm_cmplt_epi16(step2[ 9], kZero);
403        __m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero);
404        __m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero);
405        __m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero);
406        __m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero);
407        __m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero);
408        __m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero);
409        __m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero);
410        __m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero);
411        __m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero);
412        __m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero);
413        __m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero);
414        __m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero);
415        __m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero);
416        __m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero);
417        __m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero);
418        __m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero);
419        __m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero);
420        __m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero);
421        __m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero);
422        __m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero);
423        __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero);
424        __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero);
425
426        step2[ 0] = _mm_sub_epi16(step2[ 0], s3_00_0);
427        step2[ 1] = _mm_sub_epi16(step2[ 1], s3_01_0);
428        step2[ 2] = _mm_sub_epi16(step2[ 2], s3_02_0);
429        step2[ 3] = _mm_sub_epi16(step2[ 3], s3_03_0);
430        step2[ 4] = _mm_sub_epi16(step2[ 4], s3_04_0);
431        step2[ 5] = _mm_sub_epi16(step2[ 5], s3_05_0);
432        step2[ 6] = _mm_sub_epi16(step2[ 6], s3_06_0);
433        step2[ 7] = _mm_sub_epi16(step2[ 7], s3_07_0);
434        step2[ 8] = _mm_sub_epi16(step2[ 8], s2_08_0);
435        step2[ 9] = _mm_sub_epi16(step2[ 9], s2_09_0);
436        step2[10] = _mm_sub_epi16(step2[10], s3_10_0);
437        step2[11] = _mm_sub_epi16(step2[11], s3_11_0);
438        step2[12] = _mm_sub_epi16(step2[12], s3_12_0);
439        step2[13] = _mm_sub_epi16(step2[13], s3_13_0);
440        step2[14] = _mm_sub_epi16(step2[14], s2_14_0);
441        step2[15] = _mm_sub_epi16(step2[15], s2_15_0);
442        step1[16] = _mm_sub_epi16(step1[16], s3_16_0);
443        step1[17] = _mm_sub_epi16(step1[17], s3_17_0);
444        step1[18] = _mm_sub_epi16(step1[18], s3_18_0);
445        step1[19] = _mm_sub_epi16(step1[19], s3_19_0);
446        step2[20] = _mm_sub_epi16(step2[20], s3_20_0);
447        step2[21] = _mm_sub_epi16(step2[21], s3_21_0);
448        step2[22] = _mm_sub_epi16(step2[22], s3_22_0);
449        step2[23] = _mm_sub_epi16(step2[23], s3_23_0);
450        step2[24] = _mm_sub_epi16(step2[24], s3_24_0);
451        step2[25] = _mm_sub_epi16(step2[25], s3_25_0);
452        step2[26] = _mm_sub_epi16(step2[26], s3_26_0);
453        step2[27] = _mm_sub_epi16(step2[27], s3_27_0);
454        step1[28] = _mm_sub_epi16(step1[28], s3_28_0);
455        step1[29] = _mm_sub_epi16(step1[29], s3_29_0);
456        step1[30] = _mm_sub_epi16(step1[30], s3_30_0);
457        step1[31] = _mm_sub_epi16(step1[31], s3_31_0);
458
459        step2[ 0] = _mm_add_epi16(step2[ 0], kOne);
460        step2[ 1] = _mm_add_epi16(step2[ 1], kOne);
461        step2[ 2] = _mm_add_epi16(step2[ 2], kOne);
462        step2[ 3] = _mm_add_epi16(step2[ 3], kOne);
463        step2[ 4] = _mm_add_epi16(step2[ 4], kOne);
464        step2[ 5] = _mm_add_epi16(step2[ 5], kOne);
465        step2[ 6] = _mm_add_epi16(step2[ 6], kOne);
466        step2[ 7] = _mm_add_epi16(step2[ 7], kOne);
467        step2[ 8] = _mm_add_epi16(step2[ 8], kOne);
468        step2[ 9] = _mm_add_epi16(step2[ 9], kOne);
469        step2[10] = _mm_add_epi16(step2[10], kOne);
470        step2[11] = _mm_add_epi16(step2[11], kOne);
471        step2[12] = _mm_add_epi16(step2[12], kOne);
472        step2[13] = _mm_add_epi16(step2[13], kOne);
473        step2[14] = _mm_add_epi16(step2[14], kOne);
474        step2[15] = _mm_add_epi16(step2[15], kOne);
475        step1[16] = _mm_add_epi16(step1[16], kOne);
476        step1[17] = _mm_add_epi16(step1[17], kOne);
477        step1[18] = _mm_add_epi16(step1[18], kOne);
478        step1[19] = _mm_add_epi16(step1[19], kOne);
479        step2[20] = _mm_add_epi16(step2[20], kOne);
480        step2[21] = _mm_add_epi16(step2[21], kOne);
481        step2[22] = _mm_add_epi16(step2[22], kOne);
482        step2[23] = _mm_add_epi16(step2[23], kOne);
483        step2[24] = _mm_add_epi16(step2[24], kOne);
484        step2[25] = _mm_add_epi16(step2[25], kOne);
485        step2[26] = _mm_add_epi16(step2[26], kOne);
486        step2[27] = _mm_add_epi16(step2[27], kOne);
487        step1[28] = _mm_add_epi16(step1[28], kOne);
488        step1[29] = _mm_add_epi16(step1[29], kOne);
489        step1[30] = _mm_add_epi16(step1[30], kOne);
490        step1[31] = _mm_add_epi16(step1[31], kOne);
491
492        step2[ 0] = _mm_srai_epi16(step2[ 0], 2);
493        step2[ 1] = _mm_srai_epi16(step2[ 1], 2);
494        step2[ 2] = _mm_srai_epi16(step2[ 2], 2);
495        step2[ 3] = _mm_srai_epi16(step2[ 3], 2);
496        step2[ 4] = _mm_srai_epi16(step2[ 4], 2);
497        step2[ 5] = _mm_srai_epi16(step2[ 5], 2);
498        step2[ 6] = _mm_srai_epi16(step2[ 6], 2);
499        step2[ 7] = _mm_srai_epi16(step2[ 7], 2);
500        step2[ 8] = _mm_srai_epi16(step2[ 8], 2);
501        step2[ 9] = _mm_srai_epi16(step2[ 9], 2);
502        step2[10] = _mm_srai_epi16(step2[10], 2);
503        step2[11] = _mm_srai_epi16(step2[11], 2);
504        step2[12] = _mm_srai_epi16(step2[12], 2);
505        step2[13] = _mm_srai_epi16(step2[13], 2);
506        step2[14] = _mm_srai_epi16(step2[14], 2);
507        step2[15] = _mm_srai_epi16(step2[15], 2);
508        step1[16] = _mm_srai_epi16(step1[16], 2);
509        step1[17] = _mm_srai_epi16(step1[17], 2);
510        step1[18] = _mm_srai_epi16(step1[18], 2);
511        step1[19] = _mm_srai_epi16(step1[19], 2);
512        step2[20] = _mm_srai_epi16(step2[20], 2);
513        step2[21] = _mm_srai_epi16(step2[21], 2);
514        step2[22] = _mm_srai_epi16(step2[22], 2);
515        step2[23] = _mm_srai_epi16(step2[23], 2);
516        step2[24] = _mm_srai_epi16(step2[24], 2);
517        step2[25] = _mm_srai_epi16(step2[25], 2);
518        step2[26] = _mm_srai_epi16(step2[26], 2);
519        step2[27] = _mm_srai_epi16(step2[27], 2);
520        step1[28] = _mm_srai_epi16(step1[28], 2);
521        step1[29] = _mm_srai_epi16(step1[29], 2);
522        step1[30] = _mm_srai_epi16(step1[30], 2);
523        step1[31] = _mm_srai_epi16(step1[31], 2);
524      }
525#endif
526
527#if FDCT32x32_HIGH_PRECISION
528      if (pass == 0) {
529#endif
530      // Stage 3
531      {
532        step3[0] = _mm_add_epi16(step2[(8 - 1)], step2[0]);
533        step3[1] = _mm_add_epi16(step2[(8 - 2)], step2[1]);
534        step3[2] = _mm_add_epi16(step2[(8 - 3)], step2[2]);
535        step3[3] = _mm_add_epi16(step2[(8 - 4)], step2[3]);
536        step3[4] = _mm_sub_epi16(step2[(8 - 5)], step2[4]);
537        step3[5] = _mm_sub_epi16(step2[(8 - 6)], step2[5]);
538        step3[6] = _mm_sub_epi16(step2[(8 - 7)], step2[6]);
539        step3[7] = _mm_sub_epi16(step2[(8 - 8)], step2[7]);
540      }
541      {
542        const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
543        const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
544        const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
545        const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
546        const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
547        const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
548        const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
549        const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
550        const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
551        const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
552        const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
553        const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
554        // dct_const_round_shift
555        const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
556        const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
557        const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
558        const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
559        const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
560        const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
561        const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
562        const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
563        const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
564        const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
565        const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
566        const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
567        const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
568        const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
569        const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
570        const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
571        // Combine
572        step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
573        step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
574        step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
575        step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
576      }
577      {
578        step3[16] = _mm_add_epi16(step2[23], step1[16]);
579        step3[17] = _mm_add_epi16(step2[22], step1[17]);
580        step3[18] = _mm_add_epi16(step2[21], step1[18]);
581        step3[19] = _mm_add_epi16(step2[20], step1[19]);
582        step3[20] = _mm_sub_epi16(step1[19], step2[20]);
583        step3[21] = _mm_sub_epi16(step1[18], step2[21]);
584        step3[22] = _mm_sub_epi16(step1[17], step2[22]);
585        step3[23] = _mm_sub_epi16(step1[16], step2[23]);
586        step3[24] = _mm_sub_epi16(step1[31], step2[24]);
587        step3[25] = _mm_sub_epi16(step1[30], step2[25]);
588        step3[26] = _mm_sub_epi16(step1[29], step2[26]);
589        step3[27] = _mm_sub_epi16(step1[28], step2[27]);
590        step3[28] = _mm_add_epi16(step2[27], step1[28]);
591        step3[29] = _mm_add_epi16(step2[26], step1[29]);
592        step3[30] = _mm_add_epi16(step2[25], step1[30]);
593        step3[31] = _mm_add_epi16(step2[24], step1[31]);
594      }
595
596      // Stage 4
597      {
598        step1[ 0] = _mm_add_epi16(step3[ 3], step3[ 0]);
599        step1[ 1] = _mm_add_epi16(step3[ 2], step3[ 1]);
600        step1[ 2] = _mm_sub_epi16(step3[ 1], step3[ 2]);
601        step1[ 3] = _mm_sub_epi16(step3[ 0], step3[ 3]);
602        step1[ 8] = _mm_add_epi16(step3[11], step2[ 8]);
603        step1[ 9] = _mm_add_epi16(step3[10], step2[ 9]);
604        step1[10] = _mm_sub_epi16(step2[ 9], step3[10]);
605        step1[11] = _mm_sub_epi16(step2[ 8], step3[11]);
606        step1[12] = _mm_sub_epi16(step2[15], step3[12]);
607        step1[13] = _mm_sub_epi16(step2[14], step3[13]);
608        step1[14] = _mm_add_epi16(step3[13], step2[14]);
609        step1[15] = _mm_add_epi16(step3[12], step2[15]);
610      }
611      {
612        const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
613        const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
614        const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
615        const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
616        const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
617        const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
618        // dct_const_round_shift
619        const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
620        const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
621        const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
622        const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
623        const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
624        const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
625        const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
626        const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
627        // Combine
628        step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
629        step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
630      }
631      {
632        const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
633        const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
634        const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
635        const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
636        const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
637        const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
638        const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
639        const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
640        const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
641        const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
642        const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
643        const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
644        const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
645        const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
646        const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
647        const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
648        const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
649        const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
650        const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
651        const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
652        const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
653        const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
654        const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
655        const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
656        // dct_const_round_shift
657        const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
658        const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
659        const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
660        const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
661        const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
662        const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
663        const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
664        const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
665        const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
666        const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
667        const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
668        const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
669        const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
670        const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
671        const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
672        const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
673        const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
674        const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
675        const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
676        const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
677        const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
678        const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
679        const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
680        const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
681        const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
682        const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
683        const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
684        const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
685        const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
686        const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
687        const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
688        const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
689        // Combine
690        step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
691        step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
692        step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
693        step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
694        step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
695        step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
696        step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
697        step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
698      }
699      // Stage 5
700      {
701        step2[4] = _mm_add_epi16(step1[5], step3[4]);
702        step2[5] = _mm_sub_epi16(step3[4], step1[5]);
703        step2[6] = _mm_sub_epi16(step3[7], step1[6]);
704        step2[7] = _mm_add_epi16(step1[6], step3[7]);
705      }
706      {
707        const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
708        const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
709        const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
710        const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
711        const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
712        const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
713        const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
714        const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
715        const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
716        const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
717        const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
718        const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
719        // dct_const_round_shift
720        const __m128i out_00_4 = _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
721        const __m128i out_00_5 = _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
722        const __m128i out_16_4 = _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
723        const __m128i out_16_5 = _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
724        const __m128i out_08_4 = _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
725        const __m128i out_08_5 = _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
726        const __m128i out_24_4 = _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
727        const __m128i out_24_5 = _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
728        const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
729        const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
730        const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
731        const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
732        const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
733        const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
734        const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
735        const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
736        // Combine
737        out[ 0] = _mm_packs_epi32(out_00_6, out_00_7);
738        out[16] = _mm_packs_epi32(out_16_6, out_16_7);
739        out[ 8] = _mm_packs_epi32(out_08_6, out_08_7);
740        out[24] = _mm_packs_epi32(out_24_6, out_24_7);
741      }
742      {
743        const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[ 9], step1[14]);
744        const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[ 9], step1[14]);
745        const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
746        const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
747        const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
748        const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
749        const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
750        const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
751        const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
752        const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
753        const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
754        const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
755        // dct_const_round_shift
756        const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
757        const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
758        const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
759        const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
760        const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
761        const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
762        const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
763        const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
764        const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
765        const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
766        const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
767        const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
768        const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
769        const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
770        const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
771        const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
772        // Combine
773        step2[ 9] = _mm_packs_epi32(s2_09_6, s2_09_7);
774        step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
775        step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
776        step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
777      }
778      {
779        step2[16] = _mm_add_epi16(step1[19], step3[16]);
780        step2[17] = _mm_add_epi16(step1[18], step3[17]);
781        step2[18] = _mm_sub_epi16(step3[17], step1[18]);
782        step2[19] = _mm_sub_epi16(step3[16], step1[19]);
783        step2[20] = _mm_sub_epi16(step3[23], step1[20]);
784        step2[21] = _mm_sub_epi16(step3[22], step1[21]);
785        step2[22] = _mm_add_epi16(step1[21], step3[22]);
786        step2[23] = _mm_add_epi16(step1[20], step3[23]);
787        step2[24] = _mm_add_epi16(step1[27], step3[24]);
788        step2[25] = _mm_add_epi16(step1[26], step3[25]);
789        step2[26] = _mm_sub_epi16(step3[25], step1[26]);
790        step2[27] = _mm_sub_epi16(step3[24], step1[27]);
791        step2[28] = _mm_sub_epi16(step3[31], step1[28]);
792        step2[29] = _mm_sub_epi16(step3[30], step1[29]);
793        step2[30] = _mm_add_epi16(step1[29], step3[30]);
794        step2[31] = _mm_add_epi16(step1[28], step3[31]);
795      }
796      // Stage 6
797      {
798        const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
799        const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
800        const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
801        const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
802        const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
803        const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
804        const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
805        const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
806        const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
807        const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
808        const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
809        const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
810        const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
811        const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
812        const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
813        const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
814        // dct_const_round_shift
815        const __m128i out_04_4 = _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
816        const __m128i out_04_5 = _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
817        const __m128i out_20_4 = _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
818        const __m128i out_20_5 = _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
819        const __m128i out_12_4 = _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
820        const __m128i out_12_5 = _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
821        const __m128i out_28_4 = _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
822        const __m128i out_28_5 = _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
823        const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
824        const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
825        const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
826        const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
827        const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
828        const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
829        const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
830        const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
831        // Combine
832        out[ 4] = _mm_packs_epi32(out_04_6, out_04_7);
833        out[20] = _mm_packs_epi32(out_20_6, out_20_7);
834        out[12] = _mm_packs_epi32(out_12_6, out_12_7);
835        out[28] = _mm_packs_epi32(out_28_6, out_28_7);
836      }
837      {
838        step3[ 8] = _mm_add_epi16(step2[ 9], step1[ 8]);
839        step3[ 9] = _mm_sub_epi16(step1[ 8], step2[ 9]);
840        step3[10] = _mm_sub_epi16(step1[11], step2[10]);
841        step3[11] = _mm_add_epi16(step2[10], step1[11]);
842        step3[12] = _mm_add_epi16(step2[13], step1[12]);
843        step3[13] = _mm_sub_epi16(step1[12], step2[13]);
844        step3[14] = _mm_sub_epi16(step1[15], step2[14]);
845        step3[15] = _mm_add_epi16(step2[14], step1[15]);
846      }
847      {
848        const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
849        const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
850        const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
851        const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
852        const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
853        const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
854        const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
855        const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
856        const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
857        const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
858        const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
859        const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
860        const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
861        const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
862        const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
863        const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
864        const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
865        const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
866        const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
867        const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
868        const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
869        const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
870        const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
871        const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
872        // dct_const_round_shift
873        const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
874        const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
875        const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
876        const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
877        const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
878        const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
879        const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
880        const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
881        const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
882        const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
883        const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
884        const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
885        const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
886        const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
887        const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
888        const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
889        const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
890        const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
891        const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
892        const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
893        const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
894        const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
895        const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
896        const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
897        const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
898        const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
899        const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
900        const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
901        const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
902        const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
903        const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
904        const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
905        // Combine
906        step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
907        step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
908        step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
909        step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
910        // Combine
911        step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
912        step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
913        step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
914        step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
915      }
916      // Stage 7
917      {
918        const __m128i out_02_0 = _mm_unpacklo_epi16(step3[ 8], step3[15]);
919        const __m128i out_02_1 = _mm_unpackhi_epi16(step3[ 8], step3[15]);
920        const __m128i out_18_0 = _mm_unpacklo_epi16(step3[ 9], step3[14]);
921        const __m128i out_18_1 = _mm_unpackhi_epi16(step3[ 9], step3[14]);
922        const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
923        const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
924        const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
925        const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
926        const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
927        const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
928        const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
929        const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
930        const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
931        const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
932        const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
933        const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
934        const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
935        const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
936        const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
937        const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
938        const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
939        const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
940        const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
941        const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
942        // dct_const_round_shift
943        const __m128i out_02_4 = _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
944        const __m128i out_02_5 = _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
945        const __m128i out_18_4 = _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
946        const __m128i out_18_5 = _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
947        const __m128i out_10_4 = _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
948        const __m128i out_10_5 = _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
949        const __m128i out_26_4 = _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
950        const __m128i out_26_5 = _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
951        const __m128i out_06_4 = _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
952        const __m128i out_06_5 = _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
953        const __m128i out_22_4 = _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
954        const __m128i out_22_5 = _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
955        const __m128i out_14_4 = _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
956        const __m128i out_14_5 = _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
957        const __m128i out_30_4 = _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
958        const __m128i out_30_5 = _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
959        const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
960        const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
961        const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
962        const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
963        const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
964        const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
965        const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
966        const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
967        const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
968        const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
969        const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
970        const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
971        const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
972        const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
973        const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
974        const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
975        // Combine
976        out[ 2] = _mm_packs_epi32(out_02_6, out_02_7);
977        out[18] = _mm_packs_epi32(out_18_6, out_18_7);
978        out[10] = _mm_packs_epi32(out_10_6, out_10_7);
979        out[26] = _mm_packs_epi32(out_26_6, out_26_7);
980        out[ 6] = _mm_packs_epi32(out_06_6, out_06_7);
981        out[22] = _mm_packs_epi32(out_22_6, out_22_7);
982        out[14] = _mm_packs_epi32(out_14_6, out_14_7);
983        out[30] = _mm_packs_epi32(out_30_6, out_30_7);
984      }
985      {
986        step1[16] = _mm_add_epi16(step3[17], step2[16]);
987        step1[17] = _mm_sub_epi16(step2[16], step3[17]);
988        step1[18] = _mm_sub_epi16(step2[19], step3[18]);
989        step1[19] = _mm_add_epi16(step3[18], step2[19]);
990        step1[20] = _mm_add_epi16(step3[21], step2[20]);
991        step1[21] = _mm_sub_epi16(step2[20], step3[21]);
992        step1[22] = _mm_sub_epi16(step2[23], step3[22]);
993        step1[23] = _mm_add_epi16(step3[22], step2[23]);
994        step1[24] = _mm_add_epi16(step3[25], step2[24]);
995        step1[25] = _mm_sub_epi16(step2[24], step3[25]);
996        step1[26] = _mm_sub_epi16(step2[27], step3[26]);
997        step1[27] = _mm_add_epi16(step3[26], step2[27]);
998        step1[28] = _mm_add_epi16(step3[29], step2[28]);
999        step1[29] = _mm_sub_epi16(step2[28], step3[29]);
1000        step1[30] = _mm_sub_epi16(step2[31], step3[30]);
1001        step1[31] = _mm_add_epi16(step3[30], step2[31]);
1002      }
1003      // Final stage --- outputs indices are bit-reversed.
1004      {
1005        const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
1006        const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
1007        const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
1008        const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
1009        const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
1010        const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
1011        const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
1012        const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
1013        const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
1014        const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
1015        const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
1016        const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
1017        const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
1018        const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
1019        const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
1020        const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
1021        const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
1022        const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
1023        const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
1024        const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
1025        const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
1026        const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
1027        const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
1028        const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
1029        // dct_const_round_shift
1030        const __m128i out_01_4 = _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
1031        const __m128i out_01_5 = _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
1032        const __m128i out_17_4 = _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
1033        const __m128i out_17_5 = _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
1034        const __m128i out_09_4 = _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
1035        const __m128i out_09_5 = _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
1036        const __m128i out_25_4 = _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
1037        const __m128i out_25_5 = _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
1038        const __m128i out_07_4 = _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
1039        const __m128i out_07_5 = _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
1040        const __m128i out_23_4 = _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
1041        const __m128i out_23_5 = _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
1042        const __m128i out_15_4 = _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
1043        const __m128i out_15_5 = _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
1044        const __m128i out_31_4 = _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
1045        const __m128i out_31_5 = _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
1046        const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
1047        const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
1048        const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
1049        const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
1050        const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
1051        const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
1052        const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
1053        const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
1054        const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
1055        const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
1056        const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
1057        const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
1058        const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
1059        const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
1060        const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
1061        const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
1062        // Combine
1063        out[ 1] = _mm_packs_epi32(out_01_6, out_01_7);
1064        out[17] = _mm_packs_epi32(out_17_6, out_17_7);
1065        out[ 9] = _mm_packs_epi32(out_09_6, out_09_7);
1066        out[25] = _mm_packs_epi32(out_25_6, out_25_7);
1067        out[ 7] = _mm_packs_epi32(out_07_6, out_07_7);
1068        out[23] = _mm_packs_epi32(out_23_6, out_23_7);
1069        out[15] = _mm_packs_epi32(out_15_6, out_15_7);
1070        out[31] = _mm_packs_epi32(out_31_6, out_31_7);
1071      }
1072      {
1073        const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
1074        const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
1075        const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
1076        const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
1077        const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
1078        const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
1079        const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
1080        const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
1081        const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
1082        const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
1083        const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
1084        const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
1085        const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
1086        const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
1087        const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
1088        const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
1089        const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
1090        const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
1091        const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
1092        const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
1093        const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
1094        const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
1095        const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
1096        const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
1097        // dct_const_round_shift
1098        const __m128i out_05_4 = _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
1099        const __m128i out_05_5 = _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
1100        const __m128i out_21_4 = _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
1101        const __m128i out_21_5 = _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
1102        const __m128i out_13_4 = _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
1103        const __m128i out_13_5 = _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
1104        const __m128i out_29_4 = _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
1105        const __m128i out_29_5 = _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
1106        const __m128i out_03_4 = _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
1107        const __m128i out_03_5 = _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
1108        const __m128i out_19_4 = _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
1109        const __m128i out_19_5 = _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
1110        const __m128i out_11_4 = _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
1111        const __m128i out_11_5 = _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
1112        const __m128i out_27_4 = _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
1113        const __m128i out_27_5 = _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
1114        const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
1115        const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
1116        const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
1117        const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
1118        const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
1119        const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
1120        const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
1121        const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
1122        const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
1123        const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
1124        const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
1125        const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
1126        const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
1127        const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
1128        const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
1129        const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
1130        // Combine
1131        out[ 5] = _mm_packs_epi32(out_05_6, out_05_7);
1132        out[21] = _mm_packs_epi32(out_21_6, out_21_7);
1133        out[13] = _mm_packs_epi32(out_13_6, out_13_7);
1134        out[29] = _mm_packs_epi32(out_29_6, out_29_7);
1135        out[ 3] = _mm_packs_epi32(out_03_6, out_03_7);
1136        out[19] = _mm_packs_epi32(out_19_6, out_19_7);
1137        out[11] = _mm_packs_epi32(out_11_6, out_11_7);
1138        out[27] = _mm_packs_epi32(out_27_6, out_27_7);
1139      }
1140#if FDCT32x32_HIGH_PRECISION
1141      } else {
1142        __m128i lstep1[64], lstep2[64], lstep3[64];
1143        __m128i u[32], v[32], sign[16];
1144        const __m128i K32One = _mm_set_epi32(1, 1, 1, 1);
1145        // start using 32-bit operations
1146        // stage 3
1147        {
1148          // expanding to 32-bit length priori to addition operations
1149          lstep2[ 0] = _mm_unpacklo_epi16(step2[ 0], kZero);
1150          lstep2[ 1] = _mm_unpackhi_epi16(step2[ 0], kZero);
1151          lstep2[ 2] = _mm_unpacklo_epi16(step2[ 1], kZero);
1152          lstep2[ 3] = _mm_unpackhi_epi16(step2[ 1], kZero);
1153          lstep2[ 4] = _mm_unpacklo_epi16(step2[ 2], kZero);
1154          lstep2[ 5] = _mm_unpackhi_epi16(step2[ 2], kZero);
1155          lstep2[ 6] = _mm_unpacklo_epi16(step2[ 3], kZero);
1156          lstep2[ 7] = _mm_unpackhi_epi16(step2[ 3], kZero);
1157          lstep2[ 8] = _mm_unpacklo_epi16(step2[ 4], kZero);
1158          lstep2[ 9] = _mm_unpackhi_epi16(step2[ 4], kZero);
1159          lstep2[10] = _mm_unpacklo_epi16(step2[ 5], kZero);
1160          lstep2[11] = _mm_unpackhi_epi16(step2[ 5], kZero);
1161          lstep2[12] = _mm_unpacklo_epi16(step2[ 6], kZero);
1162          lstep2[13] = _mm_unpackhi_epi16(step2[ 6], kZero);
1163          lstep2[14] = _mm_unpacklo_epi16(step2[ 7], kZero);
1164          lstep2[15] = _mm_unpackhi_epi16(step2[ 7], kZero);
1165          lstep2[ 0] = _mm_madd_epi16(lstep2[ 0], kOne);
1166          lstep2[ 1] = _mm_madd_epi16(lstep2[ 1], kOne);
1167          lstep2[ 2] = _mm_madd_epi16(lstep2[ 2], kOne);
1168          lstep2[ 3] = _mm_madd_epi16(lstep2[ 3], kOne);
1169          lstep2[ 4] = _mm_madd_epi16(lstep2[ 4], kOne);
1170          lstep2[ 5] = _mm_madd_epi16(lstep2[ 5], kOne);
1171          lstep2[ 6] = _mm_madd_epi16(lstep2[ 6], kOne);
1172          lstep2[ 7] = _mm_madd_epi16(lstep2[ 7], kOne);
1173          lstep2[ 8] = _mm_madd_epi16(lstep2[ 8], kOne);
1174          lstep2[ 9] = _mm_madd_epi16(lstep2[ 9], kOne);
1175          lstep2[10] = _mm_madd_epi16(lstep2[10], kOne);
1176          lstep2[11] = _mm_madd_epi16(lstep2[11], kOne);
1177          lstep2[12] = _mm_madd_epi16(lstep2[12], kOne);
1178          lstep2[13] = _mm_madd_epi16(lstep2[13], kOne);
1179          lstep2[14] = _mm_madd_epi16(lstep2[14], kOne);
1180          lstep2[15] = _mm_madd_epi16(lstep2[15], kOne);
1181
1182          lstep3[ 0] = _mm_add_epi32(lstep2[14], lstep2[ 0]);
1183          lstep3[ 1] = _mm_add_epi32(lstep2[15], lstep2[ 1]);
1184          lstep3[ 2] = _mm_add_epi32(lstep2[12], lstep2[ 2]);
1185          lstep3[ 3] = _mm_add_epi32(lstep2[13], lstep2[ 3]);
1186          lstep3[ 4] = _mm_add_epi32(lstep2[10], lstep2[ 4]);
1187          lstep3[ 5] = _mm_add_epi32(lstep2[11], lstep2[ 5]);
1188          lstep3[ 6] = _mm_add_epi32(lstep2[ 8], lstep2[ 6]);
1189          lstep3[ 7] = _mm_add_epi32(lstep2[ 9], lstep2[ 7]);
1190          lstep3[ 8] = _mm_sub_epi32(lstep2[ 6], lstep2[ 8]);
1191          lstep3[ 9] = _mm_sub_epi32(lstep2[ 7], lstep2[ 9]);
1192          lstep3[10] = _mm_sub_epi32(lstep2[ 4], lstep2[10]);
1193          lstep3[11] = _mm_sub_epi32(lstep2[ 5], lstep2[11]);
1194          lstep3[12] = _mm_sub_epi32(lstep2[ 2], lstep2[12]);
1195          lstep3[13] = _mm_sub_epi32(lstep2[ 3], lstep2[13]);
1196          lstep3[14] = _mm_sub_epi32(lstep2[ 0], lstep2[14]);
1197          lstep3[15] = _mm_sub_epi32(lstep2[ 1], lstep2[15]);
1198        }
1199        {
1200          const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
1201          const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
1202          const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
1203          const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
1204          const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
1205          const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
1206          const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
1207          const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
1208          const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
1209          const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
1210          const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
1211          const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
1212          // dct_const_round_shift
1213          const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
1214          const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
1215          const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
1216          const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
1217          const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
1218          const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
1219          const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
1220          const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
1221          lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
1222          lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
1223          lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
1224          lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
1225          lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
1226          lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
1227          lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
1228          lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
1229        }
1230        {
1231          lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero);
1232          lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero);
1233          lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero);
1234          lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero);
1235          lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero);
1236          lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero);
1237          lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero);
1238          lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero);
1239          lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero);
1240          lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero);
1241          lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero);
1242          lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero);
1243          lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero);
1244          lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero);
1245          lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero);
1246          lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero);
1247          lstep2[40] = _mm_madd_epi16(lstep2[40], kOne);
1248          lstep2[41] = _mm_madd_epi16(lstep2[41], kOne);
1249          lstep2[42] = _mm_madd_epi16(lstep2[42], kOne);
1250          lstep2[43] = _mm_madd_epi16(lstep2[43], kOne);
1251          lstep2[44] = _mm_madd_epi16(lstep2[44], kOne);
1252          lstep2[45] = _mm_madd_epi16(lstep2[45], kOne);
1253          lstep2[46] = _mm_madd_epi16(lstep2[46], kOne);
1254          lstep2[47] = _mm_madd_epi16(lstep2[47], kOne);
1255          lstep2[48] = _mm_madd_epi16(lstep2[48], kOne);
1256          lstep2[49] = _mm_madd_epi16(lstep2[49], kOne);
1257          lstep2[50] = _mm_madd_epi16(lstep2[50], kOne);
1258          lstep2[51] = _mm_madd_epi16(lstep2[51], kOne);
1259          lstep2[52] = _mm_madd_epi16(lstep2[52], kOne);
1260          lstep2[53] = _mm_madd_epi16(lstep2[53], kOne);
1261          lstep2[54] = _mm_madd_epi16(lstep2[54], kOne);
1262          lstep2[55] = _mm_madd_epi16(lstep2[55], kOne);
1263
1264          lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero);
1265          lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero);
1266          lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero);
1267          lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero);
1268          lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero);
1269          lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero);
1270          lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero);
1271          lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero);
1272          lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero);
1273          lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero);
1274          lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero);
1275          lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero);
1276          lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero);
1277          lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero);
1278          lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero);
1279          lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero);
1280          lstep1[32] = _mm_madd_epi16(lstep1[32], kOne);
1281          lstep1[33] = _mm_madd_epi16(lstep1[33], kOne);
1282          lstep1[34] = _mm_madd_epi16(lstep1[34], kOne);
1283          lstep1[35] = _mm_madd_epi16(lstep1[35], kOne);
1284          lstep1[36] = _mm_madd_epi16(lstep1[36], kOne);
1285          lstep1[37] = _mm_madd_epi16(lstep1[37], kOne);
1286          lstep1[38] = _mm_madd_epi16(lstep1[38], kOne);
1287          lstep1[39] = _mm_madd_epi16(lstep1[39], kOne);
1288          lstep1[56] = _mm_madd_epi16(lstep1[56], kOne);
1289          lstep1[57] = _mm_madd_epi16(lstep1[57], kOne);
1290          lstep1[58] = _mm_madd_epi16(lstep1[58], kOne);
1291          lstep1[59] = _mm_madd_epi16(lstep1[59], kOne);
1292          lstep1[60] = _mm_madd_epi16(lstep1[60], kOne);
1293          lstep1[61] = _mm_madd_epi16(lstep1[61], kOne);
1294          lstep1[62] = _mm_madd_epi16(lstep1[62], kOne);
1295          lstep1[63] = _mm_madd_epi16(lstep1[63], kOne);
1296
1297          lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]);
1298          lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]);
1299
1300          lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]);
1301          lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]);
1302          lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]);
1303          lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]);
1304          lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]);
1305          lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]);
1306          lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]);
1307          lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]);
1308          lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]);
1309          lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]);
1310          lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]);
1311          lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]);
1312          lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]);
1313          lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]);
1314          lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]);
1315          lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]);
1316          lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]);
1317          lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]);
1318          lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]);
1319          lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]);
1320          lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]);
1321          lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]);
1322          lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]);
1323          lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]);
1324          lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]);
1325          lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]);
1326          lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]);
1327          lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]);
1328          lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]);
1329          lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]);
1330        }
1331
1332        // stage 4
1333        {
1334          // expanding to 32-bit length priori to addition operations
1335          lstep2[16] = _mm_unpacklo_epi16(step2[ 8], kZero);
1336          lstep2[17] = _mm_unpackhi_epi16(step2[ 8], kZero);
1337          lstep2[18] = _mm_unpacklo_epi16(step2[ 9], kZero);
1338          lstep2[19] = _mm_unpackhi_epi16(step2[ 9], kZero);
1339          lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero);
1340          lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero);
1341          lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero);
1342          lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero);
1343          lstep2[16] = _mm_madd_epi16(lstep2[16], kOne);
1344          lstep2[17] = _mm_madd_epi16(lstep2[17], kOne);
1345          lstep2[18] = _mm_madd_epi16(lstep2[18], kOne);
1346          lstep2[19] = _mm_madd_epi16(lstep2[19], kOne);
1347          lstep2[28] = _mm_madd_epi16(lstep2[28], kOne);
1348          lstep2[29] = _mm_madd_epi16(lstep2[29], kOne);
1349          lstep2[30] = _mm_madd_epi16(lstep2[30], kOne);
1350          lstep2[31] = _mm_madd_epi16(lstep2[31], kOne);
1351
1352          lstep1[ 0] = _mm_add_epi32(lstep3[ 6], lstep3[ 0]);
1353          lstep1[ 1] = _mm_add_epi32(lstep3[ 7], lstep3[ 1]);
1354          lstep1[ 2] = _mm_add_epi32(lstep3[ 4], lstep3[ 2]);
1355          lstep1[ 3] = _mm_add_epi32(lstep3[ 5], lstep3[ 3]);
1356          lstep1[ 4] = _mm_sub_epi32(lstep3[ 2], lstep3[ 4]);
1357          lstep1[ 5] = _mm_sub_epi32(lstep3[ 3], lstep3[ 5]);
1358          lstep1[ 6] = _mm_sub_epi32(lstep3[ 0], lstep3[ 6]);
1359          lstep1[ 7] = _mm_sub_epi32(lstep3[ 1], lstep3[ 7]);
1360          lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]);
1361          lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]);
1362          lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]);
1363          lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]);
1364          lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]);
1365          lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]);
1366          lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]);
1367          lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]);
1368          lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]);
1369          lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]);
1370          lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]);
1371          lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]);
1372          lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]);
1373          lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]);
1374          lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]);
1375          lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]);
1376        }
1377        {
1378        // to be continued...
1379        //
1380        const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
1381        const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
1382
1383        u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]);
1384        u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]);
1385        u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]);
1386        u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]);
1387
1388        // TODO(jingning): manually inline k_madd_epi32_ to further hide
1389        // instruction latency.
1390        v[ 0] = k_madd_epi32(u[0], k32_p16_m16);
1391        v[ 1] = k_madd_epi32(u[1], k32_p16_m16);
1392        v[ 2] = k_madd_epi32(u[2], k32_p16_m16);
1393        v[ 3] = k_madd_epi32(u[3], k32_p16_m16);
1394        v[ 4] = k_madd_epi32(u[0], k32_p16_p16);
1395        v[ 5] = k_madd_epi32(u[1], k32_p16_p16);
1396        v[ 6] = k_madd_epi32(u[2], k32_p16_p16);
1397        v[ 7] = k_madd_epi32(u[3], k32_p16_p16);
1398
1399        u[0] = k_packs_epi64(v[0], v[1]);
1400        u[1] = k_packs_epi64(v[2], v[3]);
1401        u[2] = k_packs_epi64(v[4], v[5]);
1402        u[3] = k_packs_epi64(v[6], v[7]);
1403
1404        v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1405        v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1406        v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1407        v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1408
1409        lstep1[10] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1410        lstep1[11] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1411        lstep1[12] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1412        lstep1[13] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1413        }
1414        {
1415          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
1416          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
1417          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
1418
1419          u[ 0] = _mm_unpacklo_epi32(lstep3[36], lstep3[58]);
1420          u[ 1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]);
1421          u[ 2] = _mm_unpacklo_epi32(lstep3[37], lstep3[59]);
1422          u[ 3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]);
1423          u[ 4] = _mm_unpacklo_epi32(lstep3[38], lstep3[56]);
1424          u[ 5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]);
1425          u[ 6] = _mm_unpacklo_epi32(lstep3[39], lstep3[57]);
1426          u[ 7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]);
1427          u[ 8] = _mm_unpacklo_epi32(lstep3[40], lstep3[54]);
1428          u[ 9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]);
1429          u[10] = _mm_unpacklo_epi32(lstep3[41], lstep3[55]);
1430          u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]);
1431          u[12] = _mm_unpacklo_epi32(lstep3[42], lstep3[52]);
1432          u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]);
1433          u[14] = _mm_unpacklo_epi32(lstep3[43], lstep3[53]);
1434          u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]);
1435
1436          v[ 0] = k_madd_epi32(u[ 0], k32_m08_p24);
1437          v[ 1] = k_madd_epi32(u[ 1], k32_m08_p24);
1438          v[ 2] = k_madd_epi32(u[ 2], k32_m08_p24);
1439          v[ 3] = k_madd_epi32(u[ 3], k32_m08_p24);
1440          v[ 4] = k_madd_epi32(u[ 4], k32_m08_p24);
1441          v[ 5] = k_madd_epi32(u[ 5], k32_m08_p24);
1442          v[ 6] = k_madd_epi32(u[ 6], k32_m08_p24);
1443          v[ 7] = k_madd_epi32(u[ 7], k32_m08_p24);
1444          v[ 8] = k_madd_epi32(u[ 8], k32_m24_m08);
1445          v[ 9] = k_madd_epi32(u[ 9], k32_m24_m08);
1446          v[10] = k_madd_epi32(u[10], k32_m24_m08);
1447          v[11] = k_madd_epi32(u[11], k32_m24_m08);
1448          v[12] = k_madd_epi32(u[12], k32_m24_m08);
1449          v[13] = k_madd_epi32(u[13], k32_m24_m08);
1450          v[14] = k_madd_epi32(u[14], k32_m24_m08);
1451          v[15] = k_madd_epi32(u[15], k32_m24_m08);
1452          v[16] = k_madd_epi32(u[12], k32_m08_p24);
1453          v[17] = k_madd_epi32(u[13], k32_m08_p24);
1454          v[18] = k_madd_epi32(u[14], k32_m08_p24);
1455          v[19] = k_madd_epi32(u[15], k32_m08_p24);
1456          v[20] = k_madd_epi32(u[ 8], k32_m08_p24);
1457          v[21] = k_madd_epi32(u[ 9], k32_m08_p24);
1458          v[22] = k_madd_epi32(u[10], k32_m08_p24);
1459          v[23] = k_madd_epi32(u[11], k32_m08_p24);
1460          v[24] = k_madd_epi32(u[ 4], k32_p24_p08);
1461          v[25] = k_madd_epi32(u[ 5], k32_p24_p08);
1462          v[26] = k_madd_epi32(u[ 6], k32_p24_p08);
1463          v[27] = k_madd_epi32(u[ 7], k32_p24_p08);
1464          v[28] = k_madd_epi32(u[ 0], k32_p24_p08);
1465          v[29] = k_madd_epi32(u[ 1], k32_p24_p08);
1466          v[30] = k_madd_epi32(u[ 2], k32_p24_p08);
1467          v[31] = k_madd_epi32(u[ 3], k32_p24_p08);
1468
1469          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
1470          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
1471          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
1472          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
1473          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
1474          u[ 5] = k_packs_epi64(v[10], v[11]);
1475          u[ 6] = k_packs_epi64(v[12], v[13]);
1476          u[ 7] = k_packs_epi64(v[14], v[15]);
1477          u[ 8] = k_packs_epi64(v[16], v[17]);
1478          u[ 9] = k_packs_epi64(v[18], v[19]);
1479          u[10] = k_packs_epi64(v[20], v[21]);
1480          u[11] = k_packs_epi64(v[22], v[23]);
1481          u[12] = k_packs_epi64(v[24], v[25]);
1482          u[13] = k_packs_epi64(v[26], v[27]);
1483          u[14] = k_packs_epi64(v[28], v[29]);
1484          u[15] = k_packs_epi64(v[30], v[31]);
1485
1486          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
1487          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
1488          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
1489          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
1490          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
1491          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
1492          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
1493          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
1494          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
1495          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
1496          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1497          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1498          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1499          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1500          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1501          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1502
1503          lstep1[36] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
1504          lstep1[37] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
1505          lstep1[38] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
1506          lstep1[39] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
1507          lstep1[40] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
1508          lstep1[41] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
1509          lstep1[42] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
1510          lstep1[43] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
1511          lstep1[52] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
1512          lstep1[53] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
1513          lstep1[54] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1514          lstep1[55] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1515          lstep1[56] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1516          lstep1[57] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1517          lstep1[58] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1518          lstep1[59] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1519        }
1520        // stage 5
1521        {
1522          lstep2[ 8] = _mm_add_epi32(lstep1[10], lstep3[ 8]);
1523          lstep2[ 9] = _mm_add_epi32(lstep1[11], lstep3[ 9]);
1524          lstep2[10] = _mm_sub_epi32(lstep3[ 8], lstep1[10]);
1525          lstep2[11] = _mm_sub_epi32(lstep3[ 9], lstep1[11]);
1526          lstep2[12] = _mm_sub_epi32(lstep3[14], lstep1[12]);
1527          lstep2[13] = _mm_sub_epi32(lstep3[15], lstep1[13]);
1528          lstep2[14] = _mm_add_epi32(lstep1[12], lstep3[14]);
1529          lstep2[15] = _mm_add_epi32(lstep1[13], lstep3[15]);
1530        }
1531        {
1532          const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
1533          const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
1534          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
1535          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
1536
1537          u[0] = _mm_unpacklo_epi32(lstep1[0], lstep1[2]);
1538          u[1] = _mm_unpackhi_epi32(lstep1[0], lstep1[2]);
1539          u[2] = _mm_unpacklo_epi32(lstep1[1], lstep1[3]);
1540          u[3] = _mm_unpackhi_epi32(lstep1[1], lstep1[3]);
1541          u[4] = _mm_unpacklo_epi32(lstep1[4], lstep1[6]);
1542          u[5] = _mm_unpackhi_epi32(lstep1[4], lstep1[6]);
1543          u[6] = _mm_unpacklo_epi32(lstep1[5], lstep1[7]);
1544          u[7] = _mm_unpackhi_epi32(lstep1[5], lstep1[7]);
1545
1546          // TODO(jingning): manually inline k_madd_epi32_ to further hide
1547          // instruction latency.
1548          v[ 0] = k_madd_epi32(u[0], k32_p16_p16);
1549          v[ 1] = k_madd_epi32(u[1], k32_p16_p16);
1550          v[ 2] = k_madd_epi32(u[2], k32_p16_p16);
1551          v[ 3] = k_madd_epi32(u[3], k32_p16_p16);
1552          v[ 4] = k_madd_epi32(u[0], k32_p16_m16);
1553          v[ 5] = k_madd_epi32(u[1], k32_p16_m16);
1554          v[ 6] = k_madd_epi32(u[2], k32_p16_m16);
1555          v[ 7] = k_madd_epi32(u[3], k32_p16_m16);
1556          v[ 8] = k_madd_epi32(u[4], k32_p24_p08);
1557          v[ 9] = k_madd_epi32(u[5], k32_p24_p08);
1558          v[10] = k_madd_epi32(u[6], k32_p24_p08);
1559          v[11] = k_madd_epi32(u[7], k32_p24_p08);
1560          v[12] = k_madd_epi32(u[4], k32_m08_p24);
1561          v[13] = k_madd_epi32(u[5], k32_m08_p24);
1562          v[14] = k_madd_epi32(u[6], k32_m08_p24);
1563          v[15] = k_madd_epi32(u[7], k32_m08_p24);
1564
1565          u[0] = k_packs_epi64(v[0], v[1]);
1566          u[1] = k_packs_epi64(v[2], v[3]);
1567          u[2] = k_packs_epi64(v[4], v[5]);
1568          u[3] = k_packs_epi64(v[6], v[7]);
1569          u[4] = k_packs_epi64(v[8], v[9]);
1570          u[5] = k_packs_epi64(v[10], v[11]);
1571          u[6] = k_packs_epi64(v[12], v[13]);
1572          u[7] = k_packs_epi64(v[14], v[15]);
1573
1574          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1575          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1576          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1577          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1578          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1579          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1580          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1581          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1582
1583          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1584          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1585          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1586          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1587          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1588          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1589          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1590          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1591
1592          sign[0] = _mm_cmplt_epi32(u[0], kZero);
1593          sign[1] = _mm_cmplt_epi32(u[1], kZero);
1594          sign[2] = _mm_cmplt_epi32(u[2], kZero);
1595          sign[3] = _mm_cmplt_epi32(u[3], kZero);
1596          sign[4] = _mm_cmplt_epi32(u[4], kZero);
1597          sign[5] = _mm_cmplt_epi32(u[5], kZero);
1598          sign[6] = _mm_cmplt_epi32(u[6], kZero);
1599          sign[7] = _mm_cmplt_epi32(u[7], kZero);
1600
1601          u[0] = _mm_sub_epi32(u[0], sign[0]);
1602          u[1] = _mm_sub_epi32(u[1], sign[1]);
1603          u[2] = _mm_sub_epi32(u[2], sign[2]);
1604          u[3] = _mm_sub_epi32(u[3], sign[3]);
1605          u[4] = _mm_sub_epi32(u[4], sign[4]);
1606          u[5] = _mm_sub_epi32(u[5], sign[5]);
1607          u[6] = _mm_sub_epi32(u[6], sign[6]);
1608          u[7] = _mm_sub_epi32(u[7], sign[7]);
1609
1610          u[0] = _mm_add_epi32(u[0], K32One);
1611          u[1] = _mm_add_epi32(u[1], K32One);
1612          u[2] = _mm_add_epi32(u[2], K32One);
1613          u[3] = _mm_add_epi32(u[3], K32One);
1614          u[4] = _mm_add_epi32(u[4], K32One);
1615          u[5] = _mm_add_epi32(u[5], K32One);
1616          u[6] = _mm_add_epi32(u[6], K32One);
1617          u[7] = _mm_add_epi32(u[7], K32One);
1618
1619          u[0] = _mm_srai_epi32(u[0], 2);
1620          u[1] = _mm_srai_epi32(u[1], 2);
1621          u[2] = _mm_srai_epi32(u[2], 2);
1622          u[3] = _mm_srai_epi32(u[3], 2);
1623          u[4] = _mm_srai_epi32(u[4], 2);
1624          u[5] = _mm_srai_epi32(u[5], 2);
1625          u[6] = _mm_srai_epi32(u[6], 2);
1626          u[7] = _mm_srai_epi32(u[7], 2);
1627
1628          // Combine
1629          out[ 0] = _mm_packs_epi32(u[0], u[1]);
1630          out[16] = _mm_packs_epi32(u[2], u[3]);
1631          out[ 8] = _mm_packs_epi32(u[4], u[5]);
1632          out[24] = _mm_packs_epi32(u[6], u[7]);
1633        }
1634        {
1635          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
1636          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
1637          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
1638
1639          u[0] = _mm_unpacklo_epi32(lstep1[18], lstep1[28]);
1640          u[1] = _mm_unpackhi_epi32(lstep1[18], lstep1[28]);
1641          u[2] = _mm_unpacklo_epi32(lstep1[19], lstep1[29]);
1642          u[3] = _mm_unpackhi_epi32(lstep1[19], lstep1[29]);
1643          u[4] = _mm_unpacklo_epi32(lstep1[20], lstep1[26]);
1644          u[5] = _mm_unpackhi_epi32(lstep1[20], lstep1[26]);
1645          u[6] = _mm_unpacklo_epi32(lstep1[21], lstep1[27]);
1646          u[7] = _mm_unpackhi_epi32(lstep1[21], lstep1[27]);
1647
1648          v[0] = k_madd_epi32(u[0], k32_m08_p24);
1649          v[1] = k_madd_epi32(u[1], k32_m08_p24);
1650          v[2] = k_madd_epi32(u[2], k32_m08_p24);
1651          v[3] = k_madd_epi32(u[3], k32_m08_p24);
1652          v[4] = k_madd_epi32(u[4], k32_m24_m08);
1653          v[5] = k_madd_epi32(u[5], k32_m24_m08);
1654          v[6] = k_madd_epi32(u[6], k32_m24_m08);
1655          v[7] = k_madd_epi32(u[7], k32_m24_m08);
1656          v[ 8] = k_madd_epi32(u[4], k32_m08_p24);
1657          v[ 9] = k_madd_epi32(u[5], k32_m08_p24);
1658          v[10] = k_madd_epi32(u[6], k32_m08_p24);
1659          v[11] = k_madd_epi32(u[7], k32_m08_p24);
1660          v[12] = k_madd_epi32(u[0], k32_p24_p08);
1661          v[13] = k_madd_epi32(u[1], k32_p24_p08);
1662          v[14] = k_madd_epi32(u[2], k32_p24_p08);
1663          v[15] = k_madd_epi32(u[3], k32_p24_p08);
1664
1665          u[0] = k_packs_epi64(v[0], v[1]);
1666          u[1] = k_packs_epi64(v[2], v[3]);
1667          u[2] = k_packs_epi64(v[4], v[5]);
1668          u[3] = k_packs_epi64(v[6], v[7]);
1669          u[4] = k_packs_epi64(v[8], v[9]);
1670          u[5] = k_packs_epi64(v[10], v[11]);
1671          u[6] = k_packs_epi64(v[12], v[13]);
1672          u[7] = k_packs_epi64(v[14], v[15]);
1673
1674          u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1675          u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1676          u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1677          u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1678          u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1679          u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1680          u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1681          u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1682
1683          lstep2[18] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1684          lstep2[19] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1685          lstep2[20] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1686          lstep2[21] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1687          lstep2[26] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1688          lstep2[27] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1689          lstep2[28] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1690          lstep2[29] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1691        }
1692        {
1693          lstep2[32] = _mm_add_epi32(lstep1[38], lstep3[32]);
1694          lstep2[33] = _mm_add_epi32(lstep1[39], lstep3[33]);
1695          lstep2[34] = _mm_add_epi32(lstep1[36], lstep3[34]);
1696          lstep2[35] = _mm_add_epi32(lstep1[37], lstep3[35]);
1697          lstep2[36] = _mm_sub_epi32(lstep3[34], lstep1[36]);
1698          lstep2[37] = _mm_sub_epi32(lstep3[35], lstep1[37]);
1699          lstep2[38] = _mm_sub_epi32(lstep3[32], lstep1[38]);
1700          lstep2[39] = _mm_sub_epi32(lstep3[33], lstep1[39]);
1701          lstep2[40] = _mm_sub_epi32(lstep3[46], lstep1[40]);
1702          lstep2[41] = _mm_sub_epi32(lstep3[47], lstep1[41]);
1703          lstep2[42] = _mm_sub_epi32(lstep3[44], lstep1[42]);
1704          lstep2[43] = _mm_sub_epi32(lstep3[45], lstep1[43]);
1705          lstep2[44] = _mm_add_epi32(lstep1[42], lstep3[44]);
1706          lstep2[45] = _mm_add_epi32(lstep1[43], lstep3[45]);
1707          lstep2[46] = _mm_add_epi32(lstep1[40], lstep3[46]);
1708          lstep2[47] = _mm_add_epi32(lstep1[41], lstep3[47]);
1709          lstep2[48] = _mm_add_epi32(lstep1[54], lstep3[48]);
1710          lstep2[49] = _mm_add_epi32(lstep1[55], lstep3[49]);
1711          lstep2[50] = _mm_add_epi32(lstep1[52], lstep3[50]);
1712          lstep2[51] = _mm_add_epi32(lstep1[53], lstep3[51]);
1713          lstep2[52] = _mm_sub_epi32(lstep3[50], lstep1[52]);
1714          lstep2[53] = _mm_sub_epi32(lstep3[51], lstep1[53]);
1715          lstep2[54] = _mm_sub_epi32(lstep3[48], lstep1[54]);
1716          lstep2[55] = _mm_sub_epi32(lstep3[49], lstep1[55]);
1717          lstep2[56] = _mm_sub_epi32(lstep3[62], lstep1[56]);
1718          lstep2[57] = _mm_sub_epi32(lstep3[63], lstep1[57]);
1719          lstep2[58] = _mm_sub_epi32(lstep3[60], lstep1[58]);
1720          lstep2[59] = _mm_sub_epi32(lstep3[61], lstep1[59]);
1721          lstep2[60] = _mm_add_epi32(lstep1[58], lstep3[60]);
1722          lstep2[61] = _mm_add_epi32(lstep1[59], lstep3[61]);
1723          lstep2[62] = _mm_add_epi32(lstep1[56], lstep3[62]);
1724          lstep2[63] = _mm_add_epi32(lstep1[57], lstep3[63]);
1725        }
1726        // stage 6
1727        {
1728          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
1729          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
1730          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
1731          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
1732
1733          u[0] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
1734          u[1] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
1735          u[2] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
1736          u[3] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
1737          u[4] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
1738          u[5] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
1739          u[6] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
1740          u[7] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
1741          u[8] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
1742          u[9] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
1743          u[10] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
1744          u[11] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
1745          u[12] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
1746          u[13] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
1747          u[14] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
1748          u[15] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
1749
1750          v[0] = k_madd_epi32(u[0], k32_p28_p04);
1751          v[1] = k_madd_epi32(u[1], k32_p28_p04);
1752          v[2] = k_madd_epi32(u[2], k32_p28_p04);
1753          v[3] = k_madd_epi32(u[3], k32_p28_p04);
1754          v[4] = k_madd_epi32(u[4], k32_p12_p20);
1755          v[5] = k_madd_epi32(u[5], k32_p12_p20);
1756          v[6] = k_madd_epi32(u[6], k32_p12_p20);
1757          v[7] = k_madd_epi32(u[7], k32_p12_p20);
1758          v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
1759          v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
1760          v[10] = k_madd_epi32(u[10], k32_m20_p12);
1761          v[11] = k_madd_epi32(u[11], k32_m20_p12);
1762          v[12] = k_madd_epi32(u[12], k32_m04_p28);
1763          v[13] = k_madd_epi32(u[13], k32_m04_p28);
1764          v[14] = k_madd_epi32(u[14], k32_m04_p28);
1765          v[15] = k_madd_epi32(u[15], k32_m04_p28);
1766
1767          u[0] = k_packs_epi64(v[0], v[1]);
1768          u[1] = k_packs_epi64(v[2], v[3]);
1769          u[2] = k_packs_epi64(v[4], v[5]);
1770          u[3] = k_packs_epi64(v[6], v[7]);
1771          u[4] = k_packs_epi64(v[8], v[9]);
1772          u[5] = k_packs_epi64(v[10], v[11]);
1773          u[6] = k_packs_epi64(v[12], v[13]);
1774          u[7] = k_packs_epi64(v[14], v[15]);
1775
1776          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1777          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1778          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1779          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1780          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1781          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1782          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1783          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1784
1785          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1786          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1787          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1788          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1789          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1790          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1791          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1792          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1793
1794          sign[0] = _mm_cmplt_epi32(u[0], kZero);
1795          sign[1] = _mm_cmplt_epi32(u[1], kZero);
1796          sign[2] = _mm_cmplt_epi32(u[2], kZero);
1797          sign[3] = _mm_cmplt_epi32(u[3], kZero);
1798          sign[4] = _mm_cmplt_epi32(u[4], kZero);
1799          sign[5] = _mm_cmplt_epi32(u[5], kZero);
1800          sign[6] = _mm_cmplt_epi32(u[6], kZero);
1801          sign[7] = _mm_cmplt_epi32(u[7], kZero);
1802
1803          u[0] = _mm_sub_epi32(u[0], sign[0]);
1804          u[1] = _mm_sub_epi32(u[1], sign[1]);
1805          u[2] = _mm_sub_epi32(u[2], sign[2]);
1806          u[3] = _mm_sub_epi32(u[3], sign[3]);
1807          u[4] = _mm_sub_epi32(u[4], sign[4]);
1808          u[5] = _mm_sub_epi32(u[5], sign[5]);
1809          u[6] = _mm_sub_epi32(u[6], sign[6]);
1810          u[7] = _mm_sub_epi32(u[7], sign[7]);
1811
1812          u[0] = _mm_add_epi32(u[0], K32One);
1813          u[1] = _mm_add_epi32(u[1], K32One);
1814          u[2] = _mm_add_epi32(u[2], K32One);
1815          u[3] = _mm_add_epi32(u[3], K32One);
1816          u[4] = _mm_add_epi32(u[4], K32One);
1817          u[5] = _mm_add_epi32(u[5], K32One);
1818          u[6] = _mm_add_epi32(u[6], K32One);
1819          u[7] = _mm_add_epi32(u[7], K32One);
1820
1821          u[0] = _mm_srai_epi32(u[0], 2);
1822          u[1] = _mm_srai_epi32(u[1], 2);
1823          u[2] = _mm_srai_epi32(u[2], 2);
1824          u[3] = _mm_srai_epi32(u[3], 2);
1825          u[4] = _mm_srai_epi32(u[4], 2);
1826          u[5] = _mm_srai_epi32(u[5], 2);
1827          u[6] = _mm_srai_epi32(u[6], 2);
1828          u[7] = _mm_srai_epi32(u[7], 2);
1829
1830          out[ 4] = _mm_packs_epi32(u[0], u[1]);
1831          out[20] = _mm_packs_epi32(u[2], u[3]);
1832          out[12] = _mm_packs_epi32(u[4], u[5]);
1833          out[28] = _mm_packs_epi32(u[6], u[7]);
1834        }
1835        {
1836          lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]);
1837          lstep3[17] = _mm_add_epi32(lstep2[19], lstep1[17]);
1838          lstep3[18] = _mm_sub_epi32(lstep1[16], lstep2[18]);
1839          lstep3[19] = _mm_sub_epi32(lstep1[17], lstep2[19]);
1840          lstep3[20] = _mm_sub_epi32(lstep1[22], lstep2[20]);
1841          lstep3[21] = _mm_sub_epi32(lstep1[23], lstep2[21]);
1842          lstep3[22] = _mm_add_epi32(lstep2[20], lstep1[22]);
1843          lstep3[23] = _mm_add_epi32(lstep2[21], lstep1[23]);
1844          lstep3[24] = _mm_add_epi32(lstep2[26], lstep1[24]);
1845          lstep3[25] = _mm_add_epi32(lstep2[27], lstep1[25]);
1846          lstep3[26] = _mm_sub_epi32(lstep1[24], lstep2[26]);
1847          lstep3[27] = _mm_sub_epi32(lstep1[25], lstep2[27]);
1848          lstep3[28] = _mm_sub_epi32(lstep1[30], lstep2[28]);
1849          lstep3[29] = _mm_sub_epi32(lstep1[31], lstep2[29]);
1850          lstep3[30] = _mm_add_epi32(lstep2[28], lstep1[30]);
1851          lstep3[31] = _mm_add_epi32(lstep2[29], lstep1[31]);
1852        }
1853        {
1854          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
1855          const __m128i k32_m28_m04 = pair_set_epi32(-cospi_28_64, -cospi_4_64);
1856          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
1857          const __m128i k32_m12_m20 = pair_set_epi32(-cospi_12_64,
1858                                                     -cospi_20_64);
1859          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
1860          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
1861
1862          u[ 0] = _mm_unpacklo_epi32(lstep2[34], lstep2[60]);
1863          u[ 1] = _mm_unpackhi_epi32(lstep2[34], lstep2[60]);
1864          u[ 2] = _mm_unpacklo_epi32(lstep2[35], lstep2[61]);
1865          u[ 3] = _mm_unpackhi_epi32(lstep2[35], lstep2[61]);
1866          u[ 4] = _mm_unpacklo_epi32(lstep2[36], lstep2[58]);
1867          u[ 5] = _mm_unpackhi_epi32(lstep2[36], lstep2[58]);
1868          u[ 6] = _mm_unpacklo_epi32(lstep2[37], lstep2[59]);
1869          u[ 7] = _mm_unpackhi_epi32(lstep2[37], lstep2[59]);
1870          u[ 8] = _mm_unpacklo_epi32(lstep2[42], lstep2[52]);
1871          u[ 9] = _mm_unpackhi_epi32(lstep2[42], lstep2[52]);
1872          u[10] = _mm_unpacklo_epi32(lstep2[43], lstep2[53]);
1873          u[11] = _mm_unpackhi_epi32(lstep2[43], lstep2[53]);
1874          u[12] = _mm_unpacklo_epi32(lstep2[44], lstep2[50]);
1875          u[13] = _mm_unpackhi_epi32(lstep2[44], lstep2[50]);
1876          u[14] = _mm_unpacklo_epi32(lstep2[45], lstep2[51]);
1877          u[15] = _mm_unpackhi_epi32(lstep2[45], lstep2[51]);
1878
1879          v[ 0] = k_madd_epi32(u[ 0], k32_m04_p28);
1880          v[ 1] = k_madd_epi32(u[ 1], k32_m04_p28);
1881          v[ 2] = k_madd_epi32(u[ 2], k32_m04_p28);
1882          v[ 3] = k_madd_epi32(u[ 3], k32_m04_p28);
1883          v[ 4] = k_madd_epi32(u[ 4], k32_m28_m04);
1884          v[ 5] = k_madd_epi32(u[ 5], k32_m28_m04);
1885          v[ 6] = k_madd_epi32(u[ 6], k32_m28_m04);
1886          v[ 7] = k_madd_epi32(u[ 7], k32_m28_m04);
1887          v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
1888          v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
1889          v[10] = k_madd_epi32(u[10], k32_m20_p12);
1890          v[11] = k_madd_epi32(u[11], k32_m20_p12);
1891          v[12] = k_madd_epi32(u[12], k32_m12_m20);
1892          v[13] = k_madd_epi32(u[13], k32_m12_m20);
1893          v[14] = k_madd_epi32(u[14], k32_m12_m20);
1894          v[15] = k_madd_epi32(u[15], k32_m12_m20);
1895          v[16] = k_madd_epi32(u[12], k32_m20_p12);
1896          v[17] = k_madd_epi32(u[13], k32_m20_p12);
1897          v[18] = k_madd_epi32(u[14], k32_m20_p12);
1898          v[19] = k_madd_epi32(u[15], k32_m20_p12);
1899          v[20] = k_madd_epi32(u[ 8], k32_p12_p20);
1900          v[21] = k_madd_epi32(u[ 9], k32_p12_p20);
1901          v[22] = k_madd_epi32(u[10], k32_p12_p20);
1902          v[23] = k_madd_epi32(u[11], k32_p12_p20);
1903          v[24] = k_madd_epi32(u[ 4], k32_m04_p28);
1904          v[25] = k_madd_epi32(u[ 5], k32_m04_p28);
1905          v[26] = k_madd_epi32(u[ 6], k32_m04_p28);
1906          v[27] = k_madd_epi32(u[ 7], k32_m04_p28);
1907          v[28] = k_madd_epi32(u[ 0], k32_p28_p04);
1908          v[29] = k_madd_epi32(u[ 1], k32_p28_p04);
1909          v[30] = k_madd_epi32(u[ 2], k32_p28_p04);
1910          v[31] = k_madd_epi32(u[ 3], k32_p28_p04);
1911
1912          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
1913          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
1914          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
1915          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
1916          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
1917          u[ 5] = k_packs_epi64(v[10], v[11]);
1918          u[ 6] = k_packs_epi64(v[12], v[13]);
1919          u[ 7] = k_packs_epi64(v[14], v[15]);
1920          u[ 8] = k_packs_epi64(v[16], v[17]);
1921          u[ 9] = k_packs_epi64(v[18], v[19]);
1922          u[10] = k_packs_epi64(v[20], v[21]);
1923          u[11] = k_packs_epi64(v[22], v[23]);
1924          u[12] = k_packs_epi64(v[24], v[25]);
1925          u[13] = k_packs_epi64(v[26], v[27]);
1926          u[14] = k_packs_epi64(v[28], v[29]);
1927          u[15] = k_packs_epi64(v[30], v[31]);
1928
1929          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
1930          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
1931          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
1932          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
1933          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
1934          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
1935          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
1936          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
1937          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
1938          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
1939          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1940          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1941          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1942          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1943          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1944          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1945
1946          lstep3[34] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
1947          lstep3[35] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
1948          lstep3[36] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
1949          lstep3[37] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
1950          lstep3[42] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
1951          lstep3[43] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
1952          lstep3[44] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
1953          lstep3[45] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
1954          lstep3[50] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
1955          lstep3[51] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
1956          lstep3[52] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1957          lstep3[53] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1958          lstep3[58] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1959          lstep3[59] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1960          lstep3[60] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1961          lstep3[61] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1962        }
1963        // stage 7
1964        {
1965          const __m128i k32_p30_p02 = pair_set_epi32(cospi_30_64, cospi_2_64);
1966          const __m128i k32_p14_p18 = pair_set_epi32(cospi_14_64, cospi_18_64);
1967          const __m128i k32_p22_p10 = pair_set_epi32(cospi_22_64, cospi_10_64);
1968          const __m128i k32_p06_p26 = pair_set_epi32(cospi_6_64,  cospi_26_64);
1969          const __m128i k32_m26_p06 = pair_set_epi32(-cospi_26_64, cospi_6_64);
1970          const __m128i k32_m10_p22 = pair_set_epi32(-cospi_10_64, cospi_22_64);
1971          const __m128i k32_m18_p14 = pair_set_epi32(-cospi_18_64, cospi_14_64);
1972          const __m128i k32_m02_p30 = pair_set_epi32(-cospi_2_64, cospi_30_64);
1973
1974          u[ 0] = _mm_unpacklo_epi32(lstep3[16], lstep3[30]);
1975          u[ 1] = _mm_unpackhi_epi32(lstep3[16], lstep3[30]);
1976          u[ 2] = _mm_unpacklo_epi32(lstep3[17], lstep3[31]);
1977          u[ 3] = _mm_unpackhi_epi32(lstep3[17], lstep3[31]);
1978          u[ 4] = _mm_unpacklo_epi32(lstep3[18], lstep3[28]);
1979          u[ 5] = _mm_unpackhi_epi32(lstep3[18], lstep3[28]);
1980          u[ 6] = _mm_unpacklo_epi32(lstep3[19], lstep3[29]);
1981          u[ 7] = _mm_unpackhi_epi32(lstep3[19], lstep3[29]);
1982          u[ 8] = _mm_unpacklo_epi32(lstep3[20], lstep3[26]);
1983          u[ 9] = _mm_unpackhi_epi32(lstep3[20], lstep3[26]);
1984          u[10] = _mm_unpacklo_epi32(lstep3[21], lstep3[27]);
1985          u[11] = _mm_unpackhi_epi32(lstep3[21], lstep3[27]);
1986          u[12] = _mm_unpacklo_epi32(lstep3[22], lstep3[24]);
1987          u[13] = _mm_unpackhi_epi32(lstep3[22], lstep3[24]);
1988          u[14] = _mm_unpacklo_epi32(lstep3[23], lstep3[25]);
1989          u[15] = _mm_unpackhi_epi32(lstep3[23], lstep3[25]);
1990
1991          v[ 0] = k_madd_epi32(u[ 0], k32_p30_p02);
1992          v[ 1] = k_madd_epi32(u[ 1], k32_p30_p02);
1993          v[ 2] = k_madd_epi32(u[ 2], k32_p30_p02);
1994          v[ 3] = k_madd_epi32(u[ 3], k32_p30_p02);
1995          v[ 4] = k_madd_epi32(u[ 4], k32_p14_p18);
1996          v[ 5] = k_madd_epi32(u[ 5], k32_p14_p18);
1997          v[ 6] = k_madd_epi32(u[ 6], k32_p14_p18);
1998          v[ 7] = k_madd_epi32(u[ 7], k32_p14_p18);
1999          v[ 8] = k_madd_epi32(u[ 8], k32_p22_p10);
2000          v[ 9] = k_madd_epi32(u[ 9], k32_p22_p10);
2001          v[10] = k_madd_epi32(u[10], k32_p22_p10);
2002          v[11] = k_madd_epi32(u[11], k32_p22_p10);
2003          v[12] = k_madd_epi32(u[12], k32_p06_p26);
2004          v[13] = k_madd_epi32(u[13], k32_p06_p26);
2005          v[14] = k_madd_epi32(u[14], k32_p06_p26);
2006          v[15] = k_madd_epi32(u[15], k32_p06_p26);
2007          v[16] = k_madd_epi32(u[12], k32_m26_p06);
2008          v[17] = k_madd_epi32(u[13], k32_m26_p06);
2009          v[18] = k_madd_epi32(u[14], k32_m26_p06);
2010          v[19] = k_madd_epi32(u[15], k32_m26_p06);
2011          v[20] = k_madd_epi32(u[ 8], k32_m10_p22);
2012          v[21] = k_madd_epi32(u[ 9], k32_m10_p22);
2013          v[22] = k_madd_epi32(u[10], k32_m10_p22);
2014          v[23] = k_madd_epi32(u[11], k32_m10_p22);
2015          v[24] = k_madd_epi32(u[ 4], k32_m18_p14);
2016          v[25] = k_madd_epi32(u[ 5], k32_m18_p14);
2017          v[26] = k_madd_epi32(u[ 6], k32_m18_p14);
2018          v[27] = k_madd_epi32(u[ 7], k32_m18_p14);
2019          v[28] = k_madd_epi32(u[ 0], k32_m02_p30);
2020          v[29] = k_madd_epi32(u[ 1], k32_m02_p30);
2021          v[30] = k_madd_epi32(u[ 2], k32_m02_p30);
2022          v[31] = k_madd_epi32(u[ 3], k32_m02_p30);
2023
2024          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
2025          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
2026          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
2027          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
2028          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
2029          u[ 5] = k_packs_epi64(v[10], v[11]);
2030          u[ 6] = k_packs_epi64(v[12], v[13]);
2031          u[ 7] = k_packs_epi64(v[14], v[15]);
2032          u[ 8] = k_packs_epi64(v[16], v[17]);
2033          u[ 9] = k_packs_epi64(v[18], v[19]);
2034          u[10] = k_packs_epi64(v[20], v[21]);
2035          u[11] = k_packs_epi64(v[22], v[23]);
2036          u[12] = k_packs_epi64(v[24], v[25]);
2037          u[13] = k_packs_epi64(v[26], v[27]);
2038          u[14] = k_packs_epi64(v[28], v[29]);
2039          u[15] = k_packs_epi64(v[30], v[31]);
2040
2041          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
2042          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
2043          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
2044          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
2045          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
2046          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
2047          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
2048          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
2049          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
2050          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
2051          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2052          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2053          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2054          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2055          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2056          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2057
2058          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
2059          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
2060          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
2061          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
2062          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
2063          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
2064          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
2065          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
2066          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
2067          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
2068          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
2069          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
2070          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
2071          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
2072          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
2073          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
2074
2075          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
2076          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
2077          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
2078          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
2079          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
2080          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
2081          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
2082          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
2083          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
2084          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
2085          v[10] = _mm_cmplt_epi32(u[10], kZero);
2086          v[11] = _mm_cmplt_epi32(u[11], kZero);
2087          v[12] = _mm_cmplt_epi32(u[12], kZero);
2088          v[13] = _mm_cmplt_epi32(u[13], kZero);
2089          v[14] = _mm_cmplt_epi32(u[14], kZero);
2090          v[15] = _mm_cmplt_epi32(u[15], kZero);
2091
2092          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
2093          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
2094          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
2095          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
2096          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
2097          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
2098          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
2099          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
2100          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
2101          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
2102          u[10] = _mm_sub_epi32(u[10], v[10]);
2103          u[11] = _mm_sub_epi32(u[11], v[11]);
2104          u[12] = _mm_sub_epi32(u[12], v[12]);
2105          u[13] = _mm_sub_epi32(u[13], v[13]);
2106          u[14] = _mm_sub_epi32(u[14], v[14]);
2107          u[15] = _mm_sub_epi32(u[15], v[15]);
2108
2109          v[ 0] = _mm_add_epi32(u[ 0], K32One);
2110          v[ 1] = _mm_add_epi32(u[ 1], K32One);
2111          v[ 2] = _mm_add_epi32(u[ 2], K32One);
2112          v[ 3] = _mm_add_epi32(u[ 3], K32One);
2113          v[ 4] = _mm_add_epi32(u[ 4], K32One);
2114          v[ 5] = _mm_add_epi32(u[ 5], K32One);
2115          v[ 6] = _mm_add_epi32(u[ 6], K32One);
2116          v[ 7] = _mm_add_epi32(u[ 7], K32One);
2117          v[ 8] = _mm_add_epi32(u[ 8], K32One);
2118          v[ 9] = _mm_add_epi32(u[ 9], K32One);
2119          v[10] = _mm_add_epi32(u[10], K32One);
2120          v[11] = _mm_add_epi32(u[11], K32One);
2121          v[12] = _mm_add_epi32(u[12], K32One);
2122          v[13] = _mm_add_epi32(u[13], K32One);
2123          v[14] = _mm_add_epi32(u[14], K32One);
2124          v[15] = _mm_add_epi32(u[15], K32One);
2125
2126          u[ 0] = _mm_srai_epi32(v[ 0], 2);
2127          u[ 1] = _mm_srai_epi32(v[ 1], 2);
2128          u[ 2] = _mm_srai_epi32(v[ 2], 2);
2129          u[ 3] = _mm_srai_epi32(v[ 3], 2);
2130          u[ 4] = _mm_srai_epi32(v[ 4], 2);
2131          u[ 5] = _mm_srai_epi32(v[ 5], 2);
2132          u[ 6] = _mm_srai_epi32(v[ 6], 2);
2133          u[ 7] = _mm_srai_epi32(v[ 7], 2);
2134          u[ 8] = _mm_srai_epi32(v[ 8], 2);
2135          u[ 9] = _mm_srai_epi32(v[ 9], 2);
2136          u[10] = _mm_srai_epi32(v[10], 2);
2137          u[11] = _mm_srai_epi32(v[11], 2);
2138          u[12] = _mm_srai_epi32(v[12], 2);
2139          u[13] = _mm_srai_epi32(v[13], 2);
2140          u[14] = _mm_srai_epi32(v[14], 2);
2141          u[15] = _mm_srai_epi32(v[15], 2);
2142
2143          out[ 2] = _mm_packs_epi32(u[0], u[1]);
2144          out[18] = _mm_packs_epi32(u[2], u[3]);
2145          out[10] = _mm_packs_epi32(u[4], u[5]);
2146          out[26] = _mm_packs_epi32(u[6], u[7]);
2147          out[ 6] = _mm_packs_epi32(u[8], u[9]);
2148          out[22] = _mm_packs_epi32(u[10], u[11]);
2149          out[14] = _mm_packs_epi32(u[12], u[13]);
2150          out[30] = _mm_packs_epi32(u[14], u[15]);
2151        }
2152        {
2153          lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]);
2154          lstep1[33] = _mm_add_epi32(lstep3[35], lstep2[33]);
2155          lstep1[34] = _mm_sub_epi32(lstep2[32], lstep3[34]);
2156          lstep1[35] = _mm_sub_epi32(lstep2[33], lstep3[35]);
2157          lstep1[36] = _mm_sub_epi32(lstep2[38], lstep3[36]);
2158          lstep1[37] = _mm_sub_epi32(lstep2[39], lstep3[37]);
2159          lstep1[38] = _mm_add_epi32(lstep3[36], lstep2[38]);
2160          lstep1[39] = _mm_add_epi32(lstep3[37], lstep2[39]);
2161          lstep1[40] = _mm_add_epi32(lstep3[42], lstep2[40]);
2162          lstep1[41] = _mm_add_epi32(lstep3[43], lstep2[41]);
2163          lstep1[42] = _mm_sub_epi32(lstep2[40], lstep3[42]);
2164          lstep1[43] = _mm_sub_epi32(lstep2[41], lstep3[43]);
2165          lstep1[44] = _mm_sub_epi32(lstep2[46], lstep3[44]);
2166          lstep1[45] = _mm_sub_epi32(lstep2[47], lstep3[45]);
2167          lstep1[46] = _mm_add_epi32(lstep3[44], lstep2[46]);
2168          lstep1[47] = _mm_add_epi32(lstep3[45], lstep2[47]);
2169          lstep1[48] = _mm_add_epi32(lstep3[50], lstep2[48]);
2170          lstep1[49] = _mm_add_epi32(lstep3[51], lstep2[49]);
2171          lstep1[50] = _mm_sub_epi32(lstep2[48], lstep3[50]);
2172          lstep1[51] = _mm_sub_epi32(lstep2[49], lstep3[51]);
2173          lstep1[52] = _mm_sub_epi32(lstep2[54], lstep3[52]);
2174          lstep1[53] = _mm_sub_epi32(lstep2[55], lstep3[53]);
2175          lstep1[54] = _mm_add_epi32(lstep3[52], lstep2[54]);
2176          lstep1[55] = _mm_add_epi32(lstep3[53], lstep2[55]);
2177          lstep1[56] = _mm_add_epi32(lstep3[58], lstep2[56]);
2178          lstep1[57] = _mm_add_epi32(lstep3[59], lstep2[57]);
2179          lstep1[58] = _mm_sub_epi32(lstep2[56], lstep3[58]);
2180          lstep1[59] = _mm_sub_epi32(lstep2[57], lstep3[59]);
2181          lstep1[60] = _mm_sub_epi32(lstep2[62], lstep3[60]);
2182          lstep1[61] = _mm_sub_epi32(lstep2[63], lstep3[61]);
2183          lstep1[62] = _mm_add_epi32(lstep3[60], lstep2[62]);
2184          lstep1[63] = _mm_add_epi32(lstep3[61], lstep2[63]);
2185        }
2186        // stage 8
2187        {
2188          const __m128i k32_p31_p01 = pair_set_epi32(cospi_31_64, cospi_1_64);
2189          const __m128i k32_p15_p17 = pair_set_epi32(cospi_15_64, cospi_17_64);
2190          const __m128i k32_p23_p09 = pair_set_epi32(cospi_23_64, cospi_9_64);
2191          const __m128i k32_p07_p25 = pair_set_epi32(cospi_7_64, cospi_25_64);
2192          const __m128i k32_m25_p07 = pair_set_epi32(-cospi_25_64, cospi_7_64);
2193          const __m128i k32_m09_p23 = pair_set_epi32(-cospi_9_64, cospi_23_64);
2194          const __m128i k32_m17_p15 = pair_set_epi32(-cospi_17_64, cospi_15_64);
2195          const __m128i k32_m01_p31 = pair_set_epi32(-cospi_1_64, cospi_31_64);
2196
2197          u[ 0] = _mm_unpacklo_epi32(lstep1[32], lstep1[62]);
2198          u[ 1] = _mm_unpackhi_epi32(lstep1[32], lstep1[62]);
2199          u[ 2] = _mm_unpacklo_epi32(lstep1[33], lstep1[63]);
2200          u[ 3] = _mm_unpackhi_epi32(lstep1[33], lstep1[63]);
2201          u[ 4] = _mm_unpacklo_epi32(lstep1[34], lstep1[60]);
2202          u[ 5] = _mm_unpackhi_epi32(lstep1[34], lstep1[60]);
2203          u[ 6] = _mm_unpacklo_epi32(lstep1[35], lstep1[61]);
2204          u[ 7] = _mm_unpackhi_epi32(lstep1[35], lstep1[61]);
2205          u[ 8] = _mm_unpacklo_epi32(lstep1[36], lstep1[58]);
2206          u[ 9] = _mm_unpackhi_epi32(lstep1[36], lstep1[58]);
2207          u[10] = _mm_unpacklo_epi32(lstep1[37], lstep1[59]);
2208          u[11] = _mm_unpackhi_epi32(lstep1[37], lstep1[59]);
2209          u[12] = _mm_unpacklo_epi32(lstep1[38], lstep1[56]);
2210          u[13] = _mm_unpackhi_epi32(lstep1[38], lstep1[56]);
2211          u[14] = _mm_unpacklo_epi32(lstep1[39], lstep1[57]);
2212          u[15] = _mm_unpackhi_epi32(lstep1[39], lstep1[57]);
2213
2214          v[ 0] = k_madd_epi32(u[ 0], k32_p31_p01);
2215          v[ 1] = k_madd_epi32(u[ 1], k32_p31_p01);
2216          v[ 2] = k_madd_epi32(u[ 2], k32_p31_p01);
2217          v[ 3] = k_madd_epi32(u[ 3], k32_p31_p01);
2218          v[ 4] = k_madd_epi32(u[ 4], k32_p15_p17);
2219          v[ 5] = k_madd_epi32(u[ 5], k32_p15_p17);
2220          v[ 6] = k_madd_epi32(u[ 6], k32_p15_p17);
2221          v[ 7] = k_madd_epi32(u[ 7], k32_p15_p17);
2222          v[ 8] = k_madd_epi32(u[ 8], k32_p23_p09);
2223          v[ 9] = k_madd_epi32(u[ 9], k32_p23_p09);
2224          v[10] = k_madd_epi32(u[10], k32_p23_p09);
2225          v[11] = k_madd_epi32(u[11], k32_p23_p09);
2226          v[12] = k_madd_epi32(u[12], k32_p07_p25);
2227          v[13] = k_madd_epi32(u[13], k32_p07_p25);
2228          v[14] = k_madd_epi32(u[14], k32_p07_p25);
2229          v[15] = k_madd_epi32(u[15], k32_p07_p25);
2230          v[16] = k_madd_epi32(u[12], k32_m25_p07);
2231          v[17] = k_madd_epi32(u[13], k32_m25_p07);
2232          v[18] = k_madd_epi32(u[14], k32_m25_p07);
2233          v[19] = k_madd_epi32(u[15], k32_m25_p07);
2234          v[20] = k_madd_epi32(u[ 8], k32_m09_p23);
2235          v[21] = k_madd_epi32(u[ 9], k32_m09_p23);
2236          v[22] = k_madd_epi32(u[10], k32_m09_p23);
2237          v[23] = k_madd_epi32(u[11], k32_m09_p23);
2238          v[24] = k_madd_epi32(u[ 4], k32_m17_p15);
2239          v[25] = k_madd_epi32(u[ 5], k32_m17_p15);
2240          v[26] = k_madd_epi32(u[ 6], k32_m17_p15);
2241          v[27] = k_madd_epi32(u[ 7], k32_m17_p15);
2242          v[28] = k_madd_epi32(u[ 0], k32_m01_p31);
2243          v[29] = k_madd_epi32(u[ 1], k32_m01_p31);
2244          v[30] = k_madd_epi32(u[ 2], k32_m01_p31);
2245          v[31] = k_madd_epi32(u[ 3], k32_m01_p31);
2246
2247          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
2248          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
2249          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
2250          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
2251          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
2252          u[ 5] = k_packs_epi64(v[10], v[11]);
2253          u[ 6] = k_packs_epi64(v[12], v[13]);
2254          u[ 7] = k_packs_epi64(v[14], v[15]);
2255          u[ 8] = k_packs_epi64(v[16], v[17]);
2256          u[ 9] = k_packs_epi64(v[18], v[19]);
2257          u[10] = k_packs_epi64(v[20], v[21]);
2258          u[11] = k_packs_epi64(v[22], v[23]);
2259          u[12] = k_packs_epi64(v[24], v[25]);
2260          u[13] = k_packs_epi64(v[26], v[27]);
2261          u[14] = k_packs_epi64(v[28], v[29]);
2262          u[15] = k_packs_epi64(v[30], v[31]);
2263
2264          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
2265          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
2266          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
2267          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
2268          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
2269          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
2270          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
2271          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
2272          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
2273          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
2274          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2275          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2276          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2277          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2278          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2279          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2280
2281          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
2282          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
2283          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
2284          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
2285          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
2286          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
2287          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
2288          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
2289          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
2290          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
2291          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
2292          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
2293          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
2294          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
2295          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
2296          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
2297
2298          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
2299          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
2300          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
2301          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
2302          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
2303          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
2304          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
2305          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
2306          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
2307          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
2308          v[10] = _mm_cmplt_epi32(u[10], kZero);
2309          v[11] = _mm_cmplt_epi32(u[11], kZero);
2310          v[12] = _mm_cmplt_epi32(u[12], kZero);
2311          v[13] = _mm_cmplt_epi32(u[13], kZero);
2312          v[14] = _mm_cmplt_epi32(u[14], kZero);
2313          v[15] = _mm_cmplt_epi32(u[15], kZero);
2314
2315          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
2316          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
2317          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
2318          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
2319          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
2320          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
2321          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
2322          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
2323          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
2324          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
2325          u[10] = _mm_sub_epi32(u[10], v[10]);
2326          u[11] = _mm_sub_epi32(u[11], v[11]);
2327          u[12] = _mm_sub_epi32(u[12], v[12]);
2328          u[13] = _mm_sub_epi32(u[13], v[13]);
2329          u[14] = _mm_sub_epi32(u[14], v[14]);
2330          u[15] = _mm_sub_epi32(u[15], v[15]);
2331
2332          v[0] = _mm_add_epi32(u[0], K32One);
2333          v[1] = _mm_add_epi32(u[1], K32One);
2334          v[2] = _mm_add_epi32(u[2], K32One);
2335          v[3] = _mm_add_epi32(u[3], K32One);
2336          v[4] = _mm_add_epi32(u[4], K32One);
2337          v[5] = _mm_add_epi32(u[5], K32One);
2338          v[6] = _mm_add_epi32(u[6], K32One);
2339          v[7] = _mm_add_epi32(u[7], K32One);
2340          v[8] = _mm_add_epi32(u[8], K32One);
2341          v[9] = _mm_add_epi32(u[9], K32One);
2342          v[10] = _mm_add_epi32(u[10], K32One);
2343          v[11] = _mm_add_epi32(u[11], K32One);
2344          v[12] = _mm_add_epi32(u[12], K32One);
2345          v[13] = _mm_add_epi32(u[13], K32One);
2346          v[14] = _mm_add_epi32(u[14], K32One);
2347          v[15] = _mm_add_epi32(u[15], K32One);
2348
2349          u[0] = _mm_srai_epi32(v[0], 2);
2350          u[1] = _mm_srai_epi32(v[1], 2);
2351          u[2] = _mm_srai_epi32(v[2], 2);
2352          u[3] = _mm_srai_epi32(v[3], 2);
2353          u[4] = _mm_srai_epi32(v[4], 2);
2354          u[5] = _mm_srai_epi32(v[5], 2);
2355          u[6] = _mm_srai_epi32(v[6], 2);
2356          u[7] = _mm_srai_epi32(v[7], 2);
2357          u[8] = _mm_srai_epi32(v[8], 2);
2358          u[9] = _mm_srai_epi32(v[9], 2);
2359          u[10] = _mm_srai_epi32(v[10], 2);
2360          u[11] = _mm_srai_epi32(v[11], 2);
2361          u[12] = _mm_srai_epi32(v[12], 2);
2362          u[13] = _mm_srai_epi32(v[13], 2);
2363          u[14] = _mm_srai_epi32(v[14], 2);
2364          u[15] = _mm_srai_epi32(v[15], 2);
2365
2366          out[ 1] = _mm_packs_epi32(u[0], u[1]);
2367          out[17] = _mm_packs_epi32(u[2], u[3]);
2368          out[ 9] = _mm_packs_epi32(u[4], u[5]);
2369          out[25] = _mm_packs_epi32(u[6], u[7]);
2370          out[ 7] = _mm_packs_epi32(u[8], u[9]);
2371          out[23] = _mm_packs_epi32(u[10], u[11]);
2372          out[15] = _mm_packs_epi32(u[12], u[13]);
2373          out[31] = _mm_packs_epi32(u[14], u[15]);
2374        }
2375        {
2376          const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64);
2377          const __m128i k32_p11_p21 = pair_set_epi32(cospi_11_64, cospi_21_64);
2378          const __m128i k32_p19_p13 = pair_set_epi32(cospi_19_64, cospi_13_64);
2379          const __m128i k32_p03_p29 = pair_set_epi32(cospi_3_64, cospi_29_64);
2380          const __m128i k32_m29_p03 = pair_set_epi32(-cospi_29_64, cospi_3_64);
2381          const __m128i k32_m13_p19 = pair_set_epi32(-cospi_13_64, cospi_19_64);
2382          const __m128i k32_m21_p11 = pair_set_epi32(-cospi_21_64, cospi_11_64);
2383          const __m128i k32_m05_p27 = pair_set_epi32(-cospi_5_64, cospi_27_64);
2384
2385          u[ 0] = _mm_unpacklo_epi32(lstep1[40], lstep1[54]);
2386          u[ 1] = _mm_unpackhi_epi32(lstep1[40], lstep1[54]);
2387          u[ 2] = _mm_unpacklo_epi32(lstep1[41], lstep1[55]);
2388          u[ 3] = _mm_unpackhi_epi32(lstep1[41], lstep1[55]);
2389          u[ 4] = _mm_unpacklo_epi32(lstep1[42], lstep1[52]);
2390          u[ 5] = _mm_unpackhi_epi32(lstep1[42], lstep1[52]);
2391          u[ 6] = _mm_unpacklo_epi32(lstep1[43], lstep1[53]);
2392          u[ 7] = _mm_unpackhi_epi32(lstep1[43], lstep1[53]);
2393          u[ 8] = _mm_unpacklo_epi32(lstep1[44], lstep1[50]);
2394          u[ 9] = _mm_unpackhi_epi32(lstep1[44], lstep1[50]);
2395          u[10] = _mm_unpacklo_epi32(lstep1[45], lstep1[51]);
2396          u[11] = _mm_unpackhi_epi32(lstep1[45], lstep1[51]);
2397          u[12] = _mm_unpacklo_epi32(lstep1[46], lstep1[48]);
2398          u[13] = _mm_unpackhi_epi32(lstep1[46], lstep1[48]);
2399          u[14] = _mm_unpacklo_epi32(lstep1[47], lstep1[49]);
2400          u[15] = _mm_unpackhi_epi32(lstep1[47], lstep1[49]);
2401
2402          v[ 0] = k_madd_epi32(u[ 0], k32_p27_p05);
2403          v[ 1] = k_madd_epi32(u[ 1], k32_p27_p05);
2404          v[ 2] = k_madd_epi32(u[ 2], k32_p27_p05);
2405          v[ 3] = k_madd_epi32(u[ 3], k32_p27_p05);
2406          v[ 4] = k_madd_epi32(u[ 4], k32_p11_p21);
2407          v[ 5] = k_madd_epi32(u[ 5], k32_p11_p21);
2408          v[ 6] = k_madd_epi32(u[ 6], k32_p11_p21);
2409          v[ 7] = k_madd_epi32(u[ 7], k32_p11_p21);
2410          v[ 8] = k_madd_epi32(u[ 8], k32_p19_p13);
2411          v[ 9] = k_madd_epi32(u[ 9], k32_p19_p13);
2412          v[10] = k_madd_epi32(u[10], k32_p19_p13);
2413          v[11] = k_madd_epi32(u[11], k32_p19_p13);
2414          v[12] = k_madd_epi32(u[12], k32_p03_p29);
2415          v[13] = k_madd_epi32(u[13], k32_p03_p29);
2416          v[14] = k_madd_epi32(u[14], k32_p03_p29);
2417          v[15] = k_madd_epi32(u[15], k32_p03_p29);
2418          v[16] = k_madd_epi32(u[12], k32_m29_p03);
2419          v[17] = k_madd_epi32(u[13], k32_m29_p03);
2420          v[18] = k_madd_epi32(u[14], k32_m29_p03);
2421          v[19] = k_madd_epi32(u[15], k32_m29_p03);
2422          v[20] = k_madd_epi32(u[ 8], k32_m13_p19);
2423          v[21] = k_madd_epi32(u[ 9], k32_m13_p19);
2424          v[22] = k_madd_epi32(u[10], k32_m13_p19);
2425          v[23] = k_madd_epi32(u[11], k32_m13_p19);
2426          v[24] = k_madd_epi32(u[ 4], k32_m21_p11);
2427          v[25] = k_madd_epi32(u[ 5], k32_m21_p11);
2428          v[26] = k_madd_epi32(u[ 6], k32_m21_p11);
2429          v[27] = k_madd_epi32(u[ 7], k32_m21_p11);
2430          v[28] = k_madd_epi32(u[ 0], k32_m05_p27);
2431          v[29] = k_madd_epi32(u[ 1], k32_m05_p27);
2432          v[30] = k_madd_epi32(u[ 2], k32_m05_p27);
2433          v[31] = k_madd_epi32(u[ 3], k32_m05_p27);
2434
2435          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
2436          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
2437          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
2438          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
2439          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
2440          u[ 5] = k_packs_epi64(v[10], v[11]);
2441          u[ 6] = k_packs_epi64(v[12], v[13]);
2442          u[ 7] = k_packs_epi64(v[14], v[15]);
2443          u[ 8] = k_packs_epi64(v[16], v[17]);
2444          u[ 9] = k_packs_epi64(v[18], v[19]);
2445          u[10] = k_packs_epi64(v[20], v[21]);
2446          u[11] = k_packs_epi64(v[22], v[23]);
2447          u[12] = k_packs_epi64(v[24], v[25]);
2448          u[13] = k_packs_epi64(v[26], v[27]);
2449          u[14] = k_packs_epi64(v[28], v[29]);
2450          u[15] = k_packs_epi64(v[30], v[31]);
2451
2452          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
2453          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
2454          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
2455          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
2456          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
2457          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
2458          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
2459          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
2460          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
2461          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
2462          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2463          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2464          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2465          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2466          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2467          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2468
2469          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
2470          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
2471          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
2472          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
2473          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
2474          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
2475          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
2476          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
2477          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
2478          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
2479          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
2480          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
2481          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
2482          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
2483          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
2484          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
2485
2486          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
2487          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
2488          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
2489          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
2490          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
2491          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
2492          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
2493          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
2494          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
2495          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
2496          v[10] = _mm_cmplt_epi32(u[10], kZero);
2497          v[11] = _mm_cmplt_epi32(u[11], kZero);
2498          v[12] = _mm_cmplt_epi32(u[12], kZero);
2499          v[13] = _mm_cmplt_epi32(u[13], kZero);
2500          v[14] = _mm_cmplt_epi32(u[14], kZero);
2501          v[15] = _mm_cmplt_epi32(u[15], kZero);
2502
2503          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
2504          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
2505          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
2506          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
2507          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
2508          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
2509          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
2510          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
2511          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
2512          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
2513          u[10] = _mm_sub_epi32(u[10], v[10]);
2514          u[11] = _mm_sub_epi32(u[11], v[11]);
2515          u[12] = _mm_sub_epi32(u[12], v[12]);
2516          u[13] = _mm_sub_epi32(u[13], v[13]);
2517          u[14] = _mm_sub_epi32(u[14], v[14]);
2518          u[15] = _mm_sub_epi32(u[15], v[15]);
2519
2520          v[0] = _mm_add_epi32(u[0], K32One);
2521          v[1] = _mm_add_epi32(u[1], K32One);
2522          v[2] = _mm_add_epi32(u[2], K32One);
2523          v[3] = _mm_add_epi32(u[3], K32One);
2524          v[4] = _mm_add_epi32(u[4], K32One);
2525          v[5] = _mm_add_epi32(u[5], K32One);
2526          v[6] = _mm_add_epi32(u[6], K32One);
2527          v[7] = _mm_add_epi32(u[7], K32One);
2528          v[8] = _mm_add_epi32(u[8], K32One);
2529          v[9] = _mm_add_epi32(u[9], K32One);
2530          v[10] = _mm_add_epi32(u[10], K32One);
2531          v[11] = _mm_add_epi32(u[11], K32One);
2532          v[12] = _mm_add_epi32(u[12], K32One);
2533          v[13] = _mm_add_epi32(u[13], K32One);
2534          v[14] = _mm_add_epi32(u[14], K32One);
2535          v[15] = _mm_add_epi32(u[15], K32One);
2536
2537          u[0] = _mm_srai_epi32(v[0], 2);
2538          u[1] = _mm_srai_epi32(v[1], 2);
2539          u[2] = _mm_srai_epi32(v[2], 2);
2540          u[3] = _mm_srai_epi32(v[3], 2);
2541          u[4] = _mm_srai_epi32(v[4], 2);
2542          u[5] = _mm_srai_epi32(v[5], 2);
2543          u[6] = _mm_srai_epi32(v[6], 2);
2544          u[7] = _mm_srai_epi32(v[7], 2);
2545          u[8] = _mm_srai_epi32(v[8], 2);
2546          u[9] = _mm_srai_epi32(v[9], 2);
2547          u[10] = _mm_srai_epi32(v[10], 2);
2548          u[11] = _mm_srai_epi32(v[11], 2);
2549          u[12] = _mm_srai_epi32(v[12], 2);
2550          u[13] = _mm_srai_epi32(v[13], 2);
2551          u[14] = _mm_srai_epi32(v[14], 2);
2552          u[15] = _mm_srai_epi32(v[15], 2);
2553
2554          out[ 5] = _mm_packs_epi32(u[0], u[1]);
2555          out[21] = _mm_packs_epi32(u[2], u[3]);
2556          out[13] = _mm_packs_epi32(u[4], u[5]);
2557          out[29] = _mm_packs_epi32(u[6], u[7]);
2558          out[ 3] = _mm_packs_epi32(u[8], u[9]);
2559          out[19] = _mm_packs_epi32(u[10], u[11]);
2560          out[11] = _mm_packs_epi32(u[12], u[13]);
2561          out[27] = _mm_packs_epi32(u[14], u[15]);
2562        }
2563      }
2564#endif
2565      // Transpose the results, do it as four 8x8 transposes.
2566      {
2567        int transpose_block;
2568        int16_t *output;
2569        if (0 == pass) {
2570          output = &intermediate[column_start * 32];
2571        } else {
2572          output = &output_org[column_start * 32];
2573        }
2574        for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
2575          __m128i *this_out = &out[8 * transpose_block];
2576          // 00 01 02 03 04 05 06 07
2577          // 10 11 12 13 14 15 16 17
2578          // 20 21 22 23 24 25 26 27
2579          // 30 31 32 33 34 35 36 37
2580          // 40 41 42 43 44 45 46 47
2581          // 50 51 52 53 54 55 56 57
2582          // 60 61 62 63 64 65 66 67
2583          // 70 71 72 73 74 75 76 77
2584          const __m128i tr0_0 = _mm_unpacklo_epi16(this_out[0], this_out[1]);
2585          const __m128i tr0_1 = _mm_unpacklo_epi16(this_out[2], this_out[3]);
2586          const __m128i tr0_2 = _mm_unpackhi_epi16(this_out[0], this_out[1]);
2587          const __m128i tr0_3 = _mm_unpackhi_epi16(this_out[2], this_out[3]);
2588          const __m128i tr0_4 = _mm_unpacklo_epi16(this_out[4], this_out[5]);
2589          const __m128i tr0_5 = _mm_unpacklo_epi16(this_out[6], this_out[7]);
2590          const __m128i tr0_6 = _mm_unpackhi_epi16(this_out[4], this_out[5]);
2591          const __m128i tr0_7 = _mm_unpackhi_epi16(this_out[6], this_out[7]);
2592          // 00 10 01 11 02 12 03 13
2593          // 20 30 21 31 22 32 23 33
2594          // 04 14 05 15 06 16 07 17
2595          // 24 34 25 35 26 36 27 37
2596          // 40 50 41 51 42 52 43 53
2597          // 60 70 61 71 62 72 63 73
2598          // 54 54 55 55 56 56 57 57
2599          // 64 74 65 75 66 76 67 77
2600          const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
2601          const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
2602          const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
2603          const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
2604          const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
2605          const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
2606          const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
2607          const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
2608          // 00 10 20 30 01 11 21 31
2609          // 40 50 60 70 41 51 61 71
2610          // 02 12 22 32 03 13 23 33
2611          // 42 52 62 72 43 53 63 73
2612          // 04 14 24 34 05 15 21 36
2613          // 44 54 64 74 45 55 61 76
2614          // 06 16 26 36 07 17 27 37
2615          // 46 56 66 76 47 57 67 77
2616          __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
2617          __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
2618          __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
2619          __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
2620          __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
2621          __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
2622          __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
2623          __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
2624          // 00 10 20 30 40 50 60 70
2625          // 01 11 21 31 41 51 61 71
2626          // 02 12 22 32 42 52 62 72
2627          // 03 13 23 33 43 53 63 73
2628          // 04 14 24 34 44 54 64 74
2629          // 05 15 25 35 45 55 65 75
2630          // 06 16 26 36 46 56 66 76
2631          // 07 17 27 37 47 57 67 77
2632          if (0 == pass) {
2633            // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
2634            // TODO(cd): see quality impact of only doing
2635            //           output[j] = (output[j] + 1) >> 2;
2636            //           which would remove the code between here ...
2637            __m128i tr2_0_0 = _mm_cmpgt_epi16(tr2_0, kZero);
2638            __m128i tr2_1_0 = _mm_cmpgt_epi16(tr2_1, kZero);
2639            __m128i tr2_2_0 = _mm_cmpgt_epi16(tr2_2, kZero);
2640            __m128i tr2_3_0 = _mm_cmpgt_epi16(tr2_3, kZero);
2641            __m128i tr2_4_0 = _mm_cmpgt_epi16(tr2_4, kZero);
2642            __m128i tr2_5_0 = _mm_cmpgt_epi16(tr2_5, kZero);
2643            __m128i tr2_6_0 = _mm_cmpgt_epi16(tr2_6, kZero);
2644            __m128i tr2_7_0 = _mm_cmpgt_epi16(tr2_7, kZero);
2645            tr2_0 = _mm_sub_epi16(tr2_0, tr2_0_0);
2646            tr2_1 = _mm_sub_epi16(tr2_1, tr2_1_0);
2647            tr2_2 = _mm_sub_epi16(tr2_2, tr2_2_0);
2648            tr2_3 = _mm_sub_epi16(tr2_3, tr2_3_0);
2649            tr2_4 = _mm_sub_epi16(tr2_4, tr2_4_0);
2650            tr2_5 = _mm_sub_epi16(tr2_5, tr2_5_0);
2651            tr2_6 = _mm_sub_epi16(tr2_6, tr2_6_0);
2652            tr2_7 = _mm_sub_epi16(tr2_7, tr2_7_0);
2653            //           ... and here.
2654            //           PS: also change code in vp9/encoder/vp9_dct.c
2655            tr2_0 = _mm_add_epi16(tr2_0, kOne);
2656            tr2_1 = _mm_add_epi16(tr2_1, kOne);
2657            tr2_2 = _mm_add_epi16(tr2_2, kOne);
2658            tr2_3 = _mm_add_epi16(tr2_3, kOne);
2659            tr2_4 = _mm_add_epi16(tr2_4, kOne);
2660            tr2_5 = _mm_add_epi16(tr2_5, kOne);
2661            tr2_6 = _mm_add_epi16(tr2_6, kOne);
2662            tr2_7 = _mm_add_epi16(tr2_7, kOne);
2663            tr2_0 = _mm_srai_epi16(tr2_0, 2);
2664            tr2_1 = _mm_srai_epi16(tr2_1, 2);
2665            tr2_2 = _mm_srai_epi16(tr2_2, 2);
2666            tr2_3 = _mm_srai_epi16(tr2_3, 2);
2667            tr2_4 = _mm_srai_epi16(tr2_4, 2);
2668            tr2_5 = _mm_srai_epi16(tr2_5, 2);
2669            tr2_6 = _mm_srai_epi16(tr2_6, 2);
2670            tr2_7 = _mm_srai_epi16(tr2_7, 2);
2671          }
2672          // Note: even though all these stores are aligned, using the aligned
2673          //       intrinsic make the code slightly slower.
2674          _mm_storeu_si128((__m128i *)(output + 0 * 32), tr2_0);
2675          _mm_storeu_si128((__m128i *)(output + 1 * 32), tr2_1);
2676          _mm_storeu_si128((__m128i *)(output + 2 * 32), tr2_2);
2677          _mm_storeu_si128((__m128i *)(output + 3 * 32), tr2_3);
2678          _mm_storeu_si128((__m128i *)(output + 4 * 32), tr2_4);
2679          _mm_storeu_si128((__m128i *)(output + 5 * 32), tr2_5);
2680          _mm_storeu_si128((__m128i *)(output + 6 * 32), tr2_6);
2681          _mm_storeu_si128((__m128i *)(output + 7 * 32), tr2_7);
2682          // Process next 8x8
2683          output += 8;
2684        }
2685      }
2686    }
2687  }
2688}  // NOLINT
2689