Line data Source code
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 :
13 : #include "vpx_dsp/fwd_txfm.h"
14 : #include "vpx_dsp/txfm_common.h"
15 : #include "vpx_dsp/x86/txfm_common_sse2.h"
16 :
17 : // TODO(jingning) The high bit-depth version needs re-work for performance.
18 : // The current SSE2 implementation also causes cross reference to the static
19 : // functions in the C implementation file.
20 : #if DCT_HIGH_BIT_DEPTH
21 : #define ADD_EPI16 _mm_adds_epi16
22 : #define SUB_EPI16 _mm_subs_epi16
23 : #if FDCT32x32_HIGH_PRECISION
24 : void vpx_fdct32x32_rows_c(const int16_t *intermediate, tran_low_t *out) {
25 : int i, j;
26 : for (i = 0; i < 32; ++i) {
27 : tran_high_t temp_in[32], temp_out[32];
28 : for (j = 0; j < 32; ++j) temp_in[j] = intermediate[j * 32 + i];
29 : vpx_fdct32(temp_in, temp_out, 0);
30 : for (j = 0; j < 32; ++j)
31 : out[j + i * 32] =
32 : (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
33 : }
34 : }
35 : #define HIGH_FDCT32x32_2D_C vpx_highbd_fdct32x32_c
36 : #define HIGH_FDCT32x32_2D_ROWS_C vpx_fdct32x32_rows_c
37 : #else
38 : void vpx_fdct32x32_rd_rows_c(const int16_t *intermediate, tran_low_t *out) {
39 : int i, j;
40 : for (i = 0; i < 32; ++i) {
41 : tran_high_t temp_in[32], temp_out[32];
42 : for (j = 0; j < 32; ++j) temp_in[j] = intermediate[j * 32 + i];
43 : vpx_fdct32(temp_in, temp_out, 1);
44 : for (j = 0; j < 32; ++j) out[j + i * 32] = (tran_low_t)temp_out[j];
45 : }
46 : }
47 : #define HIGH_FDCT32x32_2D_C vpx_highbd_fdct32x32_rd_c
48 : #define HIGH_FDCT32x32_2D_ROWS_C vpx_fdct32x32_rd_rows_c
49 : #endif // FDCT32x32_HIGH_PRECISION
50 : #else
51 : #define ADD_EPI16 _mm_add_epi16
52 : #define SUB_EPI16 _mm_sub_epi16
53 : #endif // DCT_HIGH_BIT_DEPTH
54 :
55 0 : void FDCT32x32_2D(const int16_t *input, tran_low_t *output_org, int stride) {
56 : // Calculate pre-multiplied strides
57 0 : const int str1 = stride;
58 0 : const int str2 = 2 * stride;
59 0 : const int str3 = 2 * stride + str1;
60 : // We need an intermediate buffer between passes.
61 : DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]);
62 : // Constants
63 : // When we use them, in one case, they are all the same. In all others
64 : // it's a pair of them that we need to repeat four times. This is done
65 : // by constructing the 32 bit constant corresponding to that pair.
66 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
67 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
68 0 : const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
69 0 : const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
70 0 : const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64, cospi_8_64);
71 0 : const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64, cospi_20_64);
72 0 : const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
73 0 : const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
74 0 : const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64, cospi_4_64);
75 0 : const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
76 0 : const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
77 0 : const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64, cospi_2_64);
78 0 : const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64, cospi_18_64);
79 0 : const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64, cospi_10_64);
80 0 : const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64, cospi_26_64);
81 0 : const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
82 0 : const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
83 0 : const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
84 0 : const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
85 0 : const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64, cospi_1_64);
86 0 : const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64, cospi_17_64);
87 0 : const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64, cospi_9_64);
88 0 : const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64, cospi_25_64);
89 0 : const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64, cospi_7_64);
90 0 : const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64, cospi_23_64);
91 0 : const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64, cospi_15_64);
92 0 : const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64, cospi_31_64);
93 0 : const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64, cospi_5_64);
94 0 : const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64, cospi_21_64);
95 0 : const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64, cospi_13_64);
96 0 : const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64, cospi_29_64);
97 0 : const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64, cospi_3_64);
98 0 : const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64, cospi_19_64);
99 0 : const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64, cospi_11_64);
100 0 : const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64, cospi_27_64);
101 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
102 0 : const __m128i kZero = _mm_set1_epi16(0);
103 0 : const __m128i kOne = _mm_set1_epi16(1);
104 : // Do the two transform/transpose passes
105 : int pass;
106 : #if DCT_HIGH_BIT_DEPTH
107 : int overflow;
108 : #endif
109 0 : for (pass = 0; pass < 2; ++pass) {
110 : // We process eight columns (transposed rows in second pass) at a time.
111 : int column_start;
112 0 : for (column_start = 0; column_start < 32; column_start += 8) {
113 : __m128i step1[32];
114 : __m128i step2[32];
115 : __m128i step3[32];
116 : __m128i out[32];
117 : // Stage 1
118 : // Note: even though all the loads below are aligned, using the aligned
119 : // intrinsic make the code slightly slower.
120 0 : if (0 == pass) {
121 0 : const int16_t *in = &input[column_start];
122 : // step1[i] = (in[ 0 * stride] + in[(32 - 1) * stride]) << 2;
123 : // Note: the next four blocks could be in a loop. That would help the
124 : // instruction cache but is actually slower.
125 : {
126 0 : const int16_t *ina = in + 0 * str1;
127 0 : const int16_t *inb = in + 31 * str1;
128 0 : __m128i *step1a = &step1[0];
129 0 : __m128i *step1b = &step1[31];
130 0 : const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
131 0 : const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
132 0 : const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
133 0 : const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
134 0 : const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
135 0 : const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
136 0 : const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
137 0 : const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
138 0 : step1a[0] = _mm_add_epi16(ina0, inb0);
139 0 : step1a[1] = _mm_add_epi16(ina1, inb1);
140 0 : step1a[2] = _mm_add_epi16(ina2, inb2);
141 0 : step1a[3] = _mm_add_epi16(ina3, inb3);
142 0 : step1b[-3] = _mm_sub_epi16(ina3, inb3);
143 0 : step1b[-2] = _mm_sub_epi16(ina2, inb2);
144 0 : step1b[-1] = _mm_sub_epi16(ina1, inb1);
145 0 : step1b[-0] = _mm_sub_epi16(ina0, inb0);
146 0 : step1a[0] = _mm_slli_epi16(step1a[0], 2);
147 0 : step1a[1] = _mm_slli_epi16(step1a[1], 2);
148 0 : step1a[2] = _mm_slli_epi16(step1a[2], 2);
149 0 : step1a[3] = _mm_slli_epi16(step1a[3], 2);
150 0 : step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
151 0 : step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
152 0 : step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
153 0 : step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
154 : }
155 : {
156 0 : const int16_t *ina = in + 4 * str1;
157 0 : const int16_t *inb = in + 27 * str1;
158 0 : __m128i *step1a = &step1[4];
159 0 : __m128i *step1b = &step1[27];
160 0 : const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
161 0 : const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
162 0 : const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
163 0 : const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
164 0 : const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
165 0 : const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
166 0 : const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
167 0 : const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
168 0 : step1a[0] = _mm_add_epi16(ina0, inb0);
169 0 : step1a[1] = _mm_add_epi16(ina1, inb1);
170 0 : step1a[2] = _mm_add_epi16(ina2, inb2);
171 0 : step1a[3] = _mm_add_epi16(ina3, inb3);
172 0 : step1b[-3] = _mm_sub_epi16(ina3, inb3);
173 0 : step1b[-2] = _mm_sub_epi16(ina2, inb2);
174 0 : step1b[-1] = _mm_sub_epi16(ina1, inb1);
175 0 : step1b[-0] = _mm_sub_epi16(ina0, inb0);
176 0 : step1a[0] = _mm_slli_epi16(step1a[0], 2);
177 0 : step1a[1] = _mm_slli_epi16(step1a[1], 2);
178 0 : step1a[2] = _mm_slli_epi16(step1a[2], 2);
179 0 : step1a[3] = _mm_slli_epi16(step1a[3], 2);
180 0 : step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
181 0 : step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
182 0 : step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
183 0 : step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
184 : }
185 : {
186 0 : const int16_t *ina = in + 8 * str1;
187 0 : const int16_t *inb = in + 23 * str1;
188 0 : __m128i *step1a = &step1[8];
189 0 : __m128i *step1b = &step1[23];
190 0 : const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
191 0 : const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
192 0 : const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
193 0 : const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
194 0 : const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
195 0 : const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
196 0 : const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
197 0 : const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
198 0 : step1a[0] = _mm_add_epi16(ina0, inb0);
199 0 : step1a[1] = _mm_add_epi16(ina1, inb1);
200 0 : step1a[2] = _mm_add_epi16(ina2, inb2);
201 0 : step1a[3] = _mm_add_epi16(ina3, inb3);
202 0 : step1b[-3] = _mm_sub_epi16(ina3, inb3);
203 0 : step1b[-2] = _mm_sub_epi16(ina2, inb2);
204 0 : step1b[-1] = _mm_sub_epi16(ina1, inb1);
205 0 : step1b[-0] = _mm_sub_epi16(ina0, inb0);
206 0 : step1a[0] = _mm_slli_epi16(step1a[0], 2);
207 0 : step1a[1] = _mm_slli_epi16(step1a[1], 2);
208 0 : step1a[2] = _mm_slli_epi16(step1a[2], 2);
209 0 : step1a[3] = _mm_slli_epi16(step1a[3], 2);
210 0 : step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
211 0 : step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
212 0 : step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
213 0 : step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
214 : }
215 : {
216 0 : const int16_t *ina = in + 12 * str1;
217 0 : const int16_t *inb = in + 19 * str1;
218 0 : __m128i *step1a = &step1[12];
219 0 : __m128i *step1b = &step1[19];
220 0 : const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
221 0 : const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
222 0 : const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
223 0 : const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
224 0 : const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
225 0 : const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
226 0 : const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
227 0 : const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
228 0 : step1a[0] = _mm_add_epi16(ina0, inb0);
229 0 : step1a[1] = _mm_add_epi16(ina1, inb1);
230 0 : step1a[2] = _mm_add_epi16(ina2, inb2);
231 0 : step1a[3] = _mm_add_epi16(ina3, inb3);
232 0 : step1b[-3] = _mm_sub_epi16(ina3, inb3);
233 0 : step1b[-2] = _mm_sub_epi16(ina2, inb2);
234 0 : step1b[-1] = _mm_sub_epi16(ina1, inb1);
235 0 : step1b[-0] = _mm_sub_epi16(ina0, inb0);
236 0 : step1a[0] = _mm_slli_epi16(step1a[0], 2);
237 0 : step1a[1] = _mm_slli_epi16(step1a[1], 2);
238 0 : step1a[2] = _mm_slli_epi16(step1a[2], 2);
239 0 : step1a[3] = _mm_slli_epi16(step1a[3], 2);
240 0 : step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
241 0 : step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
242 0 : step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
243 0 : step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
244 : }
245 : } else {
246 0 : int16_t *in = &intermediate[column_start];
247 : // step1[i] = in[ 0 * 32] + in[(32 - 1) * 32];
248 : // Note: using the same approach as above to have common offset is
249 : // counter-productive as all offsets can be calculated at compile
250 : // time.
251 : // Note: the next four blocks could be in a loop. That would help the
252 : // instruction cache but is actually slower.
253 : {
254 0 : __m128i in00 = _mm_loadu_si128((const __m128i *)(in + 0 * 32));
255 0 : __m128i in01 = _mm_loadu_si128((const __m128i *)(in + 1 * 32));
256 0 : __m128i in02 = _mm_loadu_si128((const __m128i *)(in + 2 * 32));
257 0 : __m128i in03 = _mm_loadu_si128((const __m128i *)(in + 3 * 32));
258 0 : __m128i in28 = _mm_loadu_si128((const __m128i *)(in + 28 * 32));
259 0 : __m128i in29 = _mm_loadu_si128((const __m128i *)(in + 29 * 32));
260 0 : __m128i in30 = _mm_loadu_si128((const __m128i *)(in + 30 * 32));
261 0 : __m128i in31 = _mm_loadu_si128((const __m128i *)(in + 31 * 32));
262 0 : step1[0] = ADD_EPI16(in00, in31);
263 0 : step1[1] = ADD_EPI16(in01, in30);
264 0 : step1[2] = ADD_EPI16(in02, in29);
265 0 : step1[3] = ADD_EPI16(in03, in28);
266 0 : step1[28] = SUB_EPI16(in03, in28);
267 0 : step1[29] = SUB_EPI16(in02, in29);
268 0 : step1[30] = SUB_EPI16(in01, in30);
269 0 : step1[31] = SUB_EPI16(in00, in31);
270 : #if DCT_HIGH_BIT_DEPTH
271 : overflow = check_epi16_overflow_x8(&step1[0], &step1[1], &step1[2],
272 : &step1[3], &step1[28], &step1[29],
273 : &step1[30], &step1[31]);
274 : if (overflow) {
275 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
276 : return;
277 : }
278 : #endif // DCT_HIGH_BIT_DEPTH
279 : }
280 : {
281 0 : __m128i in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 32));
282 0 : __m128i in05 = _mm_loadu_si128((const __m128i *)(in + 5 * 32));
283 0 : __m128i in06 = _mm_loadu_si128((const __m128i *)(in + 6 * 32));
284 0 : __m128i in07 = _mm_loadu_si128((const __m128i *)(in + 7 * 32));
285 0 : __m128i in24 = _mm_loadu_si128((const __m128i *)(in + 24 * 32));
286 0 : __m128i in25 = _mm_loadu_si128((const __m128i *)(in + 25 * 32));
287 0 : __m128i in26 = _mm_loadu_si128((const __m128i *)(in + 26 * 32));
288 0 : __m128i in27 = _mm_loadu_si128((const __m128i *)(in + 27 * 32));
289 0 : step1[4] = ADD_EPI16(in04, in27);
290 0 : step1[5] = ADD_EPI16(in05, in26);
291 0 : step1[6] = ADD_EPI16(in06, in25);
292 0 : step1[7] = ADD_EPI16(in07, in24);
293 0 : step1[24] = SUB_EPI16(in07, in24);
294 0 : step1[25] = SUB_EPI16(in06, in25);
295 0 : step1[26] = SUB_EPI16(in05, in26);
296 0 : step1[27] = SUB_EPI16(in04, in27);
297 : #if DCT_HIGH_BIT_DEPTH
298 : overflow = check_epi16_overflow_x8(&step1[4], &step1[5], &step1[6],
299 : &step1[7], &step1[24], &step1[25],
300 : &step1[26], &step1[27]);
301 : if (overflow) {
302 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
303 : return;
304 : }
305 : #endif // DCT_HIGH_BIT_DEPTH
306 : }
307 : {
308 0 : __m128i in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 32));
309 0 : __m128i in09 = _mm_loadu_si128((const __m128i *)(in + 9 * 32));
310 0 : __m128i in10 = _mm_loadu_si128((const __m128i *)(in + 10 * 32));
311 0 : __m128i in11 = _mm_loadu_si128((const __m128i *)(in + 11 * 32));
312 0 : __m128i in20 = _mm_loadu_si128((const __m128i *)(in + 20 * 32));
313 0 : __m128i in21 = _mm_loadu_si128((const __m128i *)(in + 21 * 32));
314 0 : __m128i in22 = _mm_loadu_si128((const __m128i *)(in + 22 * 32));
315 0 : __m128i in23 = _mm_loadu_si128((const __m128i *)(in + 23 * 32));
316 0 : step1[8] = ADD_EPI16(in08, in23);
317 0 : step1[9] = ADD_EPI16(in09, in22);
318 0 : step1[10] = ADD_EPI16(in10, in21);
319 0 : step1[11] = ADD_EPI16(in11, in20);
320 0 : step1[20] = SUB_EPI16(in11, in20);
321 0 : step1[21] = SUB_EPI16(in10, in21);
322 0 : step1[22] = SUB_EPI16(in09, in22);
323 0 : step1[23] = SUB_EPI16(in08, in23);
324 : #if DCT_HIGH_BIT_DEPTH
325 : overflow = check_epi16_overflow_x8(&step1[8], &step1[9], &step1[10],
326 : &step1[11], &step1[20], &step1[21],
327 : &step1[22], &step1[23]);
328 : if (overflow) {
329 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
330 : return;
331 : }
332 : #endif // DCT_HIGH_BIT_DEPTH
333 : }
334 : {
335 0 : __m128i in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 32));
336 0 : __m128i in13 = _mm_loadu_si128((const __m128i *)(in + 13 * 32));
337 0 : __m128i in14 = _mm_loadu_si128((const __m128i *)(in + 14 * 32));
338 0 : __m128i in15 = _mm_loadu_si128((const __m128i *)(in + 15 * 32));
339 0 : __m128i in16 = _mm_loadu_si128((const __m128i *)(in + 16 * 32));
340 0 : __m128i in17 = _mm_loadu_si128((const __m128i *)(in + 17 * 32));
341 0 : __m128i in18 = _mm_loadu_si128((const __m128i *)(in + 18 * 32));
342 0 : __m128i in19 = _mm_loadu_si128((const __m128i *)(in + 19 * 32));
343 0 : step1[12] = ADD_EPI16(in12, in19);
344 0 : step1[13] = ADD_EPI16(in13, in18);
345 0 : step1[14] = ADD_EPI16(in14, in17);
346 0 : step1[15] = ADD_EPI16(in15, in16);
347 0 : step1[16] = SUB_EPI16(in15, in16);
348 0 : step1[17] = SUB_EPI16(in14, in17);
349 0 : step1[18] = SUB_EPI16(in13, in18);
350 0 : step1[19] = SUB_EPI16(in12, in19);
351 : #if DCT_HIGH_BIT_DEPTH
352 : overflow = check_epi16_overflow_x8(&step1[12], &step1[13], &step1[14],
353 : &step1[15], &step1[16], &step1[17],
354 : &step1[18], &step1[19]);
355 : if (overflow) {
356 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
357 : return;
358 : }
359 : #endif // DCT_HIGH_BIT_DEPTH
360 : }
361 : }
362 : // Stage 2
363 : {
364 0 : step2[0] = ADD_EPI16(step1[0], step1[15]);
365 0 : step2[1] = ADD_EPI16(step1[1], step1[14]);
366 0 : step2[2] = ADD_EPI16(step1[2], step1[13]);
367 0 : step2[3] = ADD_EPI16(step1[3], step1[12]);
368 0 : step2[4] = ADD_EPI16(step1[4], step1[11]);
369 0 : step2[5] = ADD_EPI16(step1[5], step1[10]);
370 0 : step2[6] = ADD_EPI16(step1[6], step1[9]);
371 0 : step2[7] = ADD_EPI16(step1[7], step1[8]);
372 0 : step2[8] = SUB_EPI16(step1[7], step1[8]);
373 0 : step2[9] = SUB_EPI16(step1[6], step1[9]);
374 0 : step2[10] = SUB_EPI16(step1[5], step1[10]);
375 0 : step2[11] = SUB_EPI16(step1[4], step1[11]);
376 0 : step2[12] = SUB_EPI16(step1[3], step1[12]);
377 0 : step2[13] = SUB_EPI16(step1[2], step1[13]);
378 0 : step2[14] = SUB_EPI16(step1[1], step1[14]);
379 0 : step2[15] = SUB_EPI16(step1[0], step1[15]);
380 : #if DCT_HIGH_BIT_DEPTH
381 : overflow = check_epi16_overflow_x16(
382 : &step2[0], &step2[1], &step2[2], &step2[3], &step2[4], &step2[5],
383 : &step2[6], &step2[7], &step2[8], &step2[9], &step2[10], &step2[11],
384 : &step2[12], &step2[13], &step2[14], &step2[15]);
385 : if (overflow) {
386 : if (pass == 0)
387 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
388 : else
389 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
390 : return;
391 : }
392 : #endif // DCT_HIGH_BIT_DEPTH
393 : }
394 : {
395 0 : const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
396 0 : const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
397 0 : const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
398 0 : const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
399 0 : const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
400 0 : const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
401 0 : const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
402 0 : const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
403 0 : const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
404 0 : const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
405 0 : const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
406 0 : const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
407 0 : const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
408 0 : const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
409 0 : const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
410 0 : const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
411 0 : const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
412 0 : const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
413 0 : const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
414 0 : const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
415 0 : const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
416 0 : const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
417 0 : const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
418 0 : const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
419 : // dct_const_round_shift
420 0 : const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
421 0 : const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
422 0 : const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
423 0 : const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
424 0 : const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
425 0 : const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
426 0 : const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
427 0 : const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
428 0 : const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
429 0 : const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
430 0 : const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
431 0 : const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
432 0 : const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
433 0 : const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
434 0 : const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
435 0 : const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
436 0 : const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
437 0 : const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
438 0 : const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
439 0 : const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
440 0 : const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
441 0 : const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
442 0 : const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
443 0 : const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
444 0 : const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
445 0 : const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
446 0 : const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
447 0 : const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
448 0 : const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
449 0 : const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
450 0 : const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
451 0 : const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
452 : // Combine
453 0 : step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
454 0 : step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
455 0 : step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
456 0 : step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
457 0 : step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
458 0 : step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
459 0 : step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
460 0 : step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
461 : #if DCT_HIGH_BIT_DEPTH
462 : overflow = check_epi16_overflow_x8(&step2[20], &step2[21], &step2[22],
463 : &step2[23], &step2[24], &step2[25],
464 : &step2[26], &step2[27]);
465 : if (overflow) {
466 : if (pass == 0)
467 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
468 : else
469 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
470 : return;
471 : }
472 : #endif // DCT_HIGH_BIT_DEPTH
473 : }
474 :
475 : #if !FDCT32x32_HIGH_PRECISION
476 : // dump the magnitude by half, hence the intermediate values are within
477 : // the range of 16 bits.
478 0 : if (1 == pass) {
479 0 : __m128i s3_00_0 = _mm_cmplt_epi16(step2[0], kZero);
480 0 : __m128i s3_01_0 = _mm_cmplt_epi16(step2[1], kZero);
481 0 : __m128i s3_02_0 = _mm_cmplt_epi16(step2[2], kZero);
482 0 : __m128i s3_03_0 = _mm_cmplt_epi16(step2[3], kZero);
483 0 : __m128i s3_04_0 = _mm_cmplt_epi16(step2[4], kZero);
484 0 : __m128i s3_05_0 = _mm_cmplt_epi16(step2[5], kZero);
485 0 : __m128i s3_06_0 = _mm_cmplt_epi16(step2[6], kZero);
486 0 : __m128i s3_07_0 = _mm_cmplt_epi16(step2[7], kZero);
487 0 : __m128i s2_08_0 = _mm_cmplt_epi16(step2[8], kZero);
488 0 : __m128i s2_09_0 = _mm_cmplt_epi16(step2[9], kZero);
489 0 : __m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero);
490 0 : __m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero);
491 0 : __m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero);
492 0 : __m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero);
493 0 : __m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero);
494 0 : __m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero);
495 0 : __m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero);
496 0 : __m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero);
497 0 : __m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero);
498 0 : __m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero);
499 0 : __m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero);
500 0 : __m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero);
501 0 : __m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero);
502 0 : __m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero);
503 0 : __m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero);
504 0 : __m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero);
505 0 : __m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero);
506 0 : __m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero);
507 0 : __m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero);
508 0 : __m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero);
509 0 : __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero);
510 0 : __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero);
511 :
512 0 : step2[0] = SUB_EPI16(step2[0], s3_00_0);
513 0 : step2[1] = SUB_EPI16(step2[1], s3_01_0);
514 0 : step2[2] = SUB_EPI16(step2[2], s3_02_0);
515 0 : step2[3] = SUB_EPI16(step2[3], s3_03_0);
516 0 : step2[4] = SUB_EPI16(step2[4], s3_04_0);
517 0 : step2[5] = SUB_EPI16(step2[5], s3_05_0);
518 0 : step2[6] = SUB_EPI16(step2[6], s3_06_0);
519 0 : step2[7] = SUB_EPI16(step2[7], s3_07_0);
520 0 : step2[8] = SUB_EPI16(step2[8], s2_08_0);
521 0 : step2[9] = SUB_EPI16(step2[9], s2_09_0);
522 0 : step2[10] = SUB_EPI16(step2[10], s3_10_0);
523 0 : step2[11] = SUB_EPI16(step2[11], s3_11_0);
524 0 : step2[12] = SUB_EPI16(step2[12], s3_12_0);
525 0 : step2[13] = SUB_EPI16(step2[13], s3_13_0);
526 0 : step2[14] = SUB_EPI16(step2[14], s2_14_0);
527 0 : step2[15] = SUB_EPI16(step2[15], s2_15_0);
528 0 : step1[16] = SUB_EPI16(step1[16], s3_16_0);
529 0 : step1[17] = SUB_EPI16(step1[17], s3_17_0);
530 0 : step1[18] = SUB_EPI16(step1[18], s3_18_0);
531 0 : step1[19] = SUB_EPI16(step1[19], s3_19_0);
532 0 : step2[20] = SUB_EPI16(step2[20], s3_20_0);
533 0 : step2[21] = SUB_EPI16(step2[21], s3_21_0);
534 0 : step2[22] = SUB_EPI16(step2[22], s3_22_0);
535 0 : step2[23] = SUB_EPI16(step2[23], s3_23_0);
536 0 : step2[24] = SUB_EPI16(step2[24], s3_24_0);
537 0 : step2[25] = SUB_EPI16(step2[25], s3_25_0);
538 0 : step2[26] = SUB_EPI16(step2[26], s3_26_0);
539 0 : step2[27] = SUB_EPI16(step2[27], s3_27_0);
540 0 : step1[28] = SUB_EPI16(step1[28], s3_28_0);
541 0 : step1[29] = SUB_EPI16(step1[29], s3_29_0);
542 0 : step1[30] = SUB_EPI16(step1[30], s3_30_0);
543 0 : step1[31] = SUB_EPI16(step1[31], s3_31_0);
544 : #if DCT_HIGH_BIT_DEPTH
545 : overflow = check_epi16_overflow_x32(
546 : &step2[0], &step2[1], &step2[2], &step2[3], &step2[4], &step2[5],
547 : &step2[6], &step2[7], &step2[8], &step2[9], &step2[10], &step2[11],
548 : &step2[12], &step2[13], &step2[14], &step2[15], &step1[16],
549 : &step1[17], &step1[18], &step1[19], &step2[20], &step2[21],
550 : &step2[22], &step2[23], &step2[24], &step2[25], &step2[26],
551 : &step2[27], &step1[28], &step1[29], &step1[30], &step1[31]);
552 : if (overflow) {
553 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
554 : return;
555 : }
556 : #endif // DCT_HIGH_BIT_DEPTH
557 0 : step2[0] = _mm_add_epi16(step2[0], kOne);
558 0 : step2[1] = _mm_add_epi16(step2[1], kOne);
559 0 : step2[2] = _mm_add_epi16(step2[2], kOne);
560 0 : step2[3] = _mm_add_epi16(step2[3], kOne);
561 0 : step2[4] = _mm_add_epi16(step2[4], kOne);
562 0 : step2[5] = _mm_add_epi16(step2[5], kOne);
563 0 : step2[6] = _mm_add_epi16(step2[6], kOne);
564 0 : step2[7] = _mm_add_epi16(step2[7], kOne);
565 0 : step2[8] = _mm_add_epi16(step2[8], kOne);
566 0 : step2[9] = _mm_add_epi16(step2[9], kOne);
567 0 : step2[10] = _mm_add_epi16(step2[10], kOne);
568 0 : step2[11] = _mm_add_epi16(step2[11], kOne);
569 0 : step2[12] = _mm_add_epi16(step2[12], kOne);
570 0 : step2[13] = _mm_add_epi16(step2[13], kOne);
571 0 : step2[14] = _mm_add_epi16(step2[14], kOne);
572 0 : step2[15] = _mm_add_epi16(step2[15], kOne);
573 0 : step1[16] = _mm_add_epi16(step1[16], kOne);
574 0 : step1[17] = _mm_add_epi16(step1[17], kOne);
575 0 : step1[18] = _mm_add_epi16(step1[18], kOne);
576 0 : step1[19] = _mm_add_epi16(step1[19], kOne);
577 0 : step2[20] = _mm_add_epi16(step2[20], kOne);
578 0 : step2[21] = _mm_add_epi16(step2[21], kOne);
579 0 : step2[22] = _mm_add_epi16(step2[22], kOne);
580 0 : step2[23] = _mm_add_epi16(step2[23], kOne);
581 0 : step2[24] = _mm_add_epi16(step2[24], kOne);
582 0 : step2[25] = _mm_add_epi16(step2[25], kOne);
583 0 : step2[26] = _mm_add_epi16(step2[26], kOne);
584 0 : step2[27] = _mm_add_epi16(step2[27], kOne);
585 0 : step1[28] = _mm_add_epi16(step1[28], kOne);
586 0 : step1[29] = _mm_add_epi16(step1[29], kOne);
587 0 : step1[30] = _mm_add_epi16(step1[30], kOne);
588 0 : step1[31] = _mm_add_epi16(step1[31], kOne);
589 :
590 0 : step2[0] = _mm_srai_epi16(step2[0], 2);
591 0 : step2[1] = _mm_srai_epi16(step2[1], 2);
592 0 : step2[2] = _mm_srai_epi16(step2[2], 2);
593 0 : step2[3] = _mm_srai_epi16(step2[3], 2);
594 0 : step2[4] = _mm_srai_epi16(step2[4], 2);
595 0 : step2[5] = _mm_srai_epi16(step2[5], 2);
596 0 : step2[6] = _mm_srai_epi16(step2[6], 2);
597 0 : step2[7] = _mm_srai_epi16(step2[7], 2);
598 0 : step2[8] = _mm_srai_epi16(step2[8], 2);
599 0 : step2[9] = _mm_srai_epi16(step2[9], 2);
600 0 : step2[10] = _mm_srai_epi16(step2[10], 2);
601 0 : step2[11] = _mm_srai_epi16(step2[11], 2);
602 0 : step2[12] = _mm_srai_epi16(step2[12], 2);
603 0 : step2[13] = _mm_srai_epi16(step2[13], 2);
604 0 : step2[14] = _mm_srai_epi16(step2[14], 2);
605 0 : step2[15] = _mm_srai_epi16(step2[15], 2);
606 0 : step1[16] = _mm_srai_epi16(step1[16], 2);
607 0 : step1[17] = _mm_srai_epi16(step1[17], 2);
608 0 : step1[18] = _mm_srai_epi16(step1[18], 2);
609 0 : step1[19] = _mm_srai_epi16(step1[19], 2);
610 0 : step2[20] = _mm_srai_epi16(step2[20], 2);
611 0 : step2[21] = _mm_srai_epi16(step2[21], 2);
612 0 : step2[22] = _mm_srai_epi16(step2[22], 2);
613 0 : step2[23] = _mm_srai_epi16(step2[23], 2);
614 0 : step2[24] = _mm_srai_epi16(step2[24], 2);
615 0 : step2[25] = _mm_srai_epi16(step2[25], 2);
616 0 : step2[26] = _mm_srai_epi16(step2[26], 2);
617 0 : step2[27] = _mm_srai_epi16(step2[27], 2);
618 0 : step1[28] = _mm_srai_epi16(step1[28], 2);
619 0 : step1[29] = _mm_srai_epi16(step1[29], 2);
620 0 : step1[30] = _mm_srai_epi16(step1[30], 2);
621 0 : step1[31] = _mm_srai_epi16(step1[31], 2);
622 : }
623 : #endif // !FDCT32x32_HIGH_PRECISION
624 :
625 : #if FDCT32x32_HIGH_PRECISION
626 0 : if (pass == 0) {
627 : #endif
628 : // Stage 3
629 : {
630 0 : step3[0] = ADD_EPI16(step2[(8 - 1)], step2[0]);
631 0 : step3[1] = ADD_EPI16(step2[(8 - 2)], step2[1]);
632 0 : step3[2] = ADD_EPI16(step2[(8 - 3)], step2[2]);
633 0 : step3[3] = ADD_EPI16(step2[(8 - 4)], step2[3]);
634 0 : step3[4] = SUB_EPI16(step2[(8 - 5)], step2[4]);
635 0 : step3[5] = SUB_EPI16(step2[(8 - 6)], step2[5]);
636 0 : step3[6] = SUB_EPI16(step2[(8 - 7)], step2[6]);
637 0 : step3[7] = SUB_EPI16(step2[(8 - 8)], step2[7]);
638 : #if DCT_HIGH_BIT_DEPTH
639 : overflow = check_epi16_overflow_x8(&step3[0], &step3[1], &step3[2],
640 : &step3[3], &step3[4], &step3[5],
641 : &step3[6], &step3[7]);
642 : if (overflow) {
643 : if (pass == 0)
644 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
645 : else
646 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
647 : return;
648 : }
649 : #endif // DCT_HIGH_BIT_DEPTH
650 : }
651 : {
652 0 : const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
653 0 : const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
654 0 : const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
655 0 : const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
656 0 : const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
657 0 : const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
658 0 : const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
659 0 : const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
660 0 : const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
661 0 : const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
662 0 : const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
663 0 : const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
664 : // dct_const_round_shift
665 0 : const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
666 0 : const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
667 0 : const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
668 0 : const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
669 0 : const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
670 0 : const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
671 0 : const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
672 0 : const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
673 0 : const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
674 0 : const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
675 0 : const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
676 0 : const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
677 0 : const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
678 0 : const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
679 0 : const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
680 0 : const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
681 : // Combine
682 0 : step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
683 0 : step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
684 0 : step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
685 0 : step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
686 : #if DCT_HIGH_BIT_DEPTH
687 : overflow = check_epi16_overflow_x4(&step3[10], &step3[11], &step3[12],
688 : &step3[13]);
689 : if (overflow) {
690 : if (pass == 0)
691 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
692 : else
693 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
694 : return;
695 : }
696 : #endif // DCT_HIGH_BIT_DEPTH
697 : }
698 : {
699 0 : step3[16] = ADD_EPI16(step2[23], step1[16]);
700 0 : step3[17] = ADD_EPI16(step2[22], step1[17]);
701 0 : step3[18] = ADD_EPI16(step2[21], step1[18]);
702 0 : step3[19] = ADD_EPI16(step2[20], step1[19]);
703 0 : step3[20] = SUB_EPI16(step1[19], step2[20]);
704 0 : step3[21] = SUB_EPI16(step1[18], step2[21]);
705 0 : step3[22] = SUB_EPI16(step1[17], step2[22]);
706 0 : step3[23] = SUB_EPI16(step1[16], step2[23]);
707 0 : step3[24] = SUB_EPI16(step1[31], step2[24]);
708 0 : step3[25] = SUB_EPI16(step1[30], step2[25]);
709 0 : step3[26] = SUB_EPI16(step1[29], step2[26]);
710 0 : step3[27] = SUB_EPI16(step1[28], step2[27]);
711 0 : step3[28] = ADD_EPI16(step2[27], step1[28]);
712 0 : step3[29] = ADD_EPI16(step2[26], step1[29]);
713 0 : step3[30] = ADD_EPI16(step2[25], step1[30]);
714 0 : step3[31] = ADD_EPI16(step2[24], step1[31]);
715 : #if DCT_HIGH_BIT_DEPTH
716 : overflow = check_epi16_overflow_x16(
717 : &step3[16], &step3[17], &step3[18], &step3[19], &step3[20],
718 : &step3[21], &step3[22], &step3[23], &step3[24], &step3[25],
719 : &step3[26], &step3[27], &step3[28], &step3[29], &step3[30],
720 : &step3[31]);
721 : if (overflow) {
722 : if (pass == 0)
723 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
724 : else
725 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
726 : return;
727 : }
728 : #endif // DCT_HIGH_BIT_DEPTH
729 : }
730 :
731 : // Stage 4
732 : {
733 0 : step1[0] = ADD_EPI16(step3[3], step3[0]);
734 0 : step1[1] = ADD_EPI16(step3[2], step3[1]);
735 0 : step1[2] = SUB_EPI16(step3[1], step3[2]);
736 0 : step1[3] = SUB_EPI16(step3[0], step3[3]);
737 0 : step1[8] = ADD_EPI16(step3[11], step2[8]);
738 0 : step1[9] = ADD_EPI16(step3[10], step2[9]);
739 0 : step1[10] = SUB_EPI16(step2[9], step3[10]);
740 0 : step1[11] = SUB_EPI16(step2[8], step3[11]);
741 0 : step1[12] = SUB_EPI16(step2[15], step3[12]);
742 0 : step1[13] = SUB_EPI16(step2[14], step3[13]);
743 0 : step1[14] = ADD_EPI16(step3[13], step2[14]);
744 0 : step1[15] = ADD_EPI16(step3[12], step2[15]);
745 : #if DCT_HIGH_BIT_DEPTH
746 : overflow = check_epi16_overflow_x16(
747 : &step1[0], &step1[1], &step1[2], &step1[3], &step1[4], &step1[5],
748 : &step1[6], &step1[7], &step1[8], &step1[9], &step1[10],
749 : &step1[11], &step1[12], &step1[13], &step1[14], &step1[15]);
750 : if (overflow) {
751 : if (pass == 0)
752 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
753 : else
754 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
755 : return;
756 : }
757 : #endif // DCT_HIGH_BIT_DEPTH
758 : }
759 : {
760 0 : const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
761 0 : const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
762 0 : const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
763 0 : const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
764 0 : const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
765 0 : const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
766 : // dct_const_round_shift
767 0 : const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
768 0 : const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
769 0 : const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
770 0 : const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
771 0 : const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
772 0 : const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
773 0 : const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
774 0 : const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
775 : // Combine
776 0 : step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
777 0 : step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
778 : #if DCT_HIGH_BIT_DEPTH
779 : overflow = check_epi16_overflow_x2(&step1[5], &step1[6]);
780 : if (overflow) {
781 : if (pass == 0)
782 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
783 : else
784 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
785 : return;
786 : }
787 : #endif // DCT_HIGH_BIT_DEPTH
788 : }
789 : {
790 0 : const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
791 0 : const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
792 0 : const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
793 0 : const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
794 0 : const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
795 0 : const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
796 0 : const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
797 0 : const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
798 0 : const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
799 0 : const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
800 0 : const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
801 0 : const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
802 0 : const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
803 0 : const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
804 0 : const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
805 0 : const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
806 0 : const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
807 0 : const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
808 0 : const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
809 0 : const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
810 0 : const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
811 0 : const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
812 0 : const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
813 0 : const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
814 : // dct_const_round_shift
815 0 : const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
816 0 : const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
817 0 : const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
818 0 : const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
819 0 : const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
820 0 : const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
821 0 : const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
822 0 : const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
823 0 : const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
824 0 : const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
825 0 : const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
826 0 : const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
827 0 : const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
828 0 : const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
829 0 : const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
830 0 : const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
831 0 : const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
832 0 : const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
833 0 : const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
834 0 : const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
835 0 : const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
836 0 : const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
837 0 : const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
838 0 : const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
839 0 : const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
840 0 : const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
841 0 : const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
842 0 : const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
843 0 : const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
844 0 : const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
845 0 : const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
846 0 : const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
847 : // Combine
848 0 : step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
849 0 : step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
850 0 : step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
851 0 : step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
852 0 : step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
853 0 : step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
854 0 : step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
855 0 : step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
856 : #if DCT_HIGH_BIT_DEPTH
857 : overflow = check_epi16_overflow_x8(&step1[18], &step1[19], &step1[20],
858 : &step1[21], &step1[26], &step1[27],
859 : &step1[28], &step1[29]);
860 : if (overflow) {
861 : if (pass == 0)
862 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
863 : else
864 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
865 : return;
866 : }
867 : #endif // DCT_HIGH_BIT_DEPTH
868 : }
869 : // Stage 5
870 : {
871 0 : step2[4] = ADD_EPI16(step1[5], step3[4]);
872 0 : step2[5] = SUB_EPI16(step3[4], step1[5]);
873 0 : step2[6] = SUB_EPI16(step3[7], step1[6]);
874 0 : step2[7] = ADD_EPI16(step1[6], step3[7]);
875 : #if DCT_HIGH_BIT_DEPTH
876 : overflow = check_epi16_overflow_x4(&step2[4], &step2[5], &step2[6],
877 : &step2[7]);
878 : if (overflow) {
879 : if (pass == 0)
880 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
881 : else
882 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
883 : return;
884 : }
885 : #endif // DCT_HIGH_BIT_DEPTH
886 : }
887 : {
888 0 : const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
889 0 : const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
890 0 : const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
891 0 : const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
892 0 : const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
893 0 : const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
894 0 : const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
895 0 : const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
896 0 : const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
897 0 : const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
898 0 : const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
899 0 : const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
900 : // dct_const_round_shift
901 0 : const __m128i out_00_4 =
902 : _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
903 0 : const __m128i out_00_5 =
904 : _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
905 0 : const __m128i out_16_4 =
906 : _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
907 0 : const __m128i out_16_5 =
908 : _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
909 0 : const __m128i out_08_4 =
910 : _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
911 0 : const __m128i out_08_5 =
912 : _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
913 0 : const __m128i out_24_4 =
914 : _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
915 0 : const __m128i out_24_5 =
916 : _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
917 0 : const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
918 0 : const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
919 0 : const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
920 0 : const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
921 0 : const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
922 0 : const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
923 0 : const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
924 0 : const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
925 : // Combine
926 0 : out[0] = _mm_packs_epi32(out_00_6, out_00_7);
927 0 : out[16] = _mm_packs_epi32(out_16_6, out_16_7);
928 0 : out[8] = _mm_packs_epi32(out_08_6, out_08_7);
929 0 : out[24] = _mm_packs_epi32(out_24_6, out_24_7);
930 : #if DCT_HIGH_BIT_DEPTH
931 : overflow =
932 : check_epi16_overflow_x4(&out[0], &out[16], &out[8], &out[24]);
933 : if (overflow) {
934 : if (pass == 0)
935 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
936 : else
937 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
938 : return;
939 : }
940 : #endif // DCT_HIGH_BIT_DEPTH
941 : }
942 : {
943 0 : const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[9], step1[14]);
944 0 : const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[9], step1[14]);
945 0 : const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
946 0 : const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
947 0 : const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
948 0 : const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
949 0 : const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
950 0 : const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
951 0 : const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
952 0 : const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
953 0 : const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
954 0 : const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
955 : // dct_const_round_shift
956 0 : const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
957 0 : const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
958 0 : const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
959 0 : const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
960 0 : const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
961 0 : const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
962 0 : const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
963 0 : const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
964 0 : const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
965 0 : const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
966 0 : const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
967 0 : const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
968 0 : const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
969 0 : const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
970 0 : const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
971 0 : const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
972 : // Combine
973 0 : step2[9] = _mm_packs_epi32(s2_09_6, s2_09_7);
974 0 : step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
975 0 : step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
976 0 : step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
977 : #if DCT_HIGH_BIT_DEPTH
978 : overflow = check_epi16_overflow_x4(&step2[9], &step2[10], &step2[13],
979 : &step2[14]);
980 : if (overflow) {
981 : if (pass == 0)
982 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
983 : else
984 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
985 : return;
986 : }
987 : #endif // DCT_HIGH_BIT_DEPTH
988 : }
989 : {
990 0 : step2[16] = ADD_EPI16(step1[19], step3[16]);
991 0 : step2[17] = ADD_EPI16(step1[18], step3[17]);
992 0 : step2[18] = SUB_EPI16(step3[17], step1[18]);
993 0 : step2[19] = SUB_EPI16(step3[16], step1[19]);
994 0 : step2[20] = SUB_EPI16(step3[23], step1[20]);
995 0 : step2[21] = SUB_EPI16(step3[22], step1[21]);
996 0 : step2[22] = ADD_EPI16(step1[21], step3[22]);
997 0 : step2[23] = ADD_EPI16(step1[20], step3[23]);
998 0 : step2[24] = ADD_EPI16(step1[27], step3[24]);
999 0 : step2[25] = ADD_EPI16(step1[26], step3[25]);
1000 0 : step2[26] = SUB_EPI16(step3[25], step1[26]);
1001 0 : step2[27] = SUB_EPI16(step3[24], step1[27]);
1002 0 : step2[28] = SUB_EPI16(step3[31], step1[28]);
1003 0 : step2[29] = SUB_EPI16(step3[30], step1[29]);
1004 0 : step2[30] = ADD_EPI16(step1[29], step3[30]);
1005 0 : step2[31] = ADD_EPI16(step1[28], step3[31]);
1006 : #if DCT_HIGH_BIT_DEPTH
1007 : overflow = check_epi16_overflow_x16(
1008 : &step2[16], &step2[17], &step2[18], &step2[19], &step2[20],
1009 : &step2[21], &step2[22], &step2[23], &step2[24], &step2[25],
1010 : &step2[26], &step2[27], &step2[28], &step2[29], &step2[30],
1011 : &step2[31]);
1012 : if (overflow) {
1013 : if (pass == 0)
1014 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1015 : else
1016 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1017 : return;
1018 : }
1019 : #endif // DCT_HIGH_BIT_DEPTH
1020 : }
1021 : // Stage 6
1022 : {
1023 0 : const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
1024 0 : const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
1025 0 : const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
1026 0 : const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
1027 0 : const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
1028 0 : const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
1029 0 : const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
1030 0 : const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
1031 0 : const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
1032 0 : const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
1033 0 : const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
1034 0 : const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
1035 0 : const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
1036 0 : const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
1037 0 : const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
1038 0 : const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
1039 : // dct_const_round_shift
1040 0 : const __m128i out_04_4 =
1041 : _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
1042 0 : const __m128i out_04_5 =
1043 : _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
1044 0 : const __m128i out_20_4 =
1045 : _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
1046 0 : const __m128i out_20_5 =
1047 : _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
1048 0 : const __m128i out_12_4 =
1049 : _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
1050 0 : const __m128i out_12_5 =
1051 : _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
1052 0 : const __m128i out_28_4 =
1053 : _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
1054 0 : const __m128i out_28_5 =
1055 : _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
1056 0 : const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
1057 0 : const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
1058 0 : const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
1059 0 : const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
1060 0 : const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
1061 0 : const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
1062 0 : const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
1063 0 : const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
1064 : // Combine
1065 0 : out[4] = _mm_packs_epi32(out_04_6, out_04_7);
1066 0 : out[20] = _mm_packs_epi32(out_20_6, out_20_7);
1067 0 : out[12] = _mm_packs_epi32(out_12_6, out_12_7);
1068 0 : out[28] = _mm_packs_epi32(out_28_6, out_28_7);
1069 : #if DCT_HIGH_BIT_DEPTH
1070 : overflow =
1071 : check_epi16_overflow_x4(&out[4], &out[20], &out[12], &out[28]);
1072 : if (overflow) {
1073 : if (pass == 0)
1074 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1075 : else
1076 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1077 : return;
1078 : }
1079 : #endif // DCT_HIGH_BIT_DEPTH
1080 : }
1081 : {
1082 0 : step3[8] = ADD_EPI16(step2[9], step1[8]);
1083 0 : step3[9] = SUB_EPI16(step1[8], step2[9]);
1084 0 : step3[10] = SUB_EPI16(step1[11], step2[10]);
1085 0 : step3[11] = ADD_EPI16(step2[10], step1[11]);
1086 0 : step3[12] = ADD_EPI16(step2[13], step1[12]);
1087 0 : step3[13] = SUB_EPI16(step1[12], step2[13]);
1088 0 : step3[14] = SUB_EPI16(step1[15], step2[14]);
1089 0 : step3[15] = ADD_EPI16(step2[14], step1[15]);
1090 : #if DCT_HIGH_BIT_DEPTH
1091 : overflow = check_epi16_overflow_x8(&step3[8], &step3[9], &step3[10],
1092 : &step3[11], &step3[12], &step3[13],
1093 : &step3[14], &step3[15]);
1094 : if (overflow) {
1095 : if (pass == 0)
1096 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1097 : else
1098 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1099 : return;
1100 : }
1101 : #endif // DCT_HIGH_BIT_DEPTH
1102 : }
1103 : {
1104 0 : const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
1105 0 : const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
1106 0 : const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
1107 0 : const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
1108 0 : const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
1109 0 : const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
1110 0 : const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
1111 0 : const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
1112 0 : const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
1113 0 : const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
1114 0 : const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
1115 0 : const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
1116 0 : const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
1117 0 : const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
1118 0 : const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
1119 0 : const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
1120 0 : const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
1121 0 : const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
1122 0 : const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
1123 0 : const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
1124 0 : const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
1125 0 : const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
1126 0 : const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
1127 0 : const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
1128 : // dct_const_round_shift
1129 0 : const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
1130 0 : const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
1131 0 : const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
1132 0 : const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
1133 0 : const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
1134 0 : const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
1135 0 : const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
1136 0 : const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
1137 0 : const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
1138 0 : const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
1139 0 : const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
1140 0 : const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
1141 0 : const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
1142 0 : const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
1143 0 : const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
1144 0 : const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
1145 0 : const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
1146 0 : const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
1147 0 : const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
1148 0 : const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
1149 0 : const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
1150 0 : const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
1151 0 : const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
1152 0 : const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
1153 0 : const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
1154 0 : const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
1155 0 : const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
1156 0 : const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
1157 0 : const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
1158 0 : const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
1159 0 : const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
1160 0 : const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
1161 : // Combine
1162 0 : step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
1163 0 : step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
1164 0 : step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
1165 0 : step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
1166 : // Combine
1167 0 : step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
1168 0 : step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
1169 0 : step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
1170 0 : step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
1171 : #if DCT_HIGH_BIT_DEPTH
1172 : overflow = check_epi16_overflow_x8(&step3[17], &step3[18], &step3[21],
1173 : &step3[22], &step3[25], &step3[26],
1174 : &step3[29], &step3[30]);
1175 : if (overflow) {
1176 : if (pass == 0)
1177 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1178 : else
1179 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1180 : return;
1181 : }
1182 : #endif // DCT_HIGH_BIT_DEPTH
1183 : }
1184 : // Stage 7
1185 : {
1186 0 : const __m128i out_02_0 = _mm_unpacklo_epi16(step3[8], step3[15]);
1187 0 : const __m128i out_02_1 = _mm_unpackhi_epi16(step3[8], step3[15]);
1188 0 : const __m128i out_18_0 = _mm_unpacklo_epi16(step3[9], step3[14]);
1189 0 : const __m128i out_18_1 = _mm_unpackhi_epi16(step3[9], step3[14]);
1190 0 : const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
1191 0 : const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
1192 0 : const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
1193 0 : const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
1194 0 : const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
1195 0 : const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
1196 0 : const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
1197 0 : const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
1198 0 : const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
1199 0 : const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
1200 0 : const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
1201 0 : const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
1202 0 : const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
1203 0 : const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
1204 0 : const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
1205 0 : const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
1206 0 : const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
1207 0 : const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
1208 0 : const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
1209 0 : const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
1210 : // dct_const_round_shift
1211 0 : const __m128i out_02_4 =
1212 : _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
1213 0 : const __m128i out_02_5 =
1214 : _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
1215 0 : const __m128i out_18_4 =
1216 : _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
1217 0 : const __m128i out_18_5 =
1218 : _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
1219 0 : const __m128i out_10_4 =
1220 : _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
1221 0 : const __m128i out_10_5 =
1222 : _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
1223 0 : const __m128i out_26_4 =
1224 : _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
1225 0 : const __m128i out_26_5 =
1226 : _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
1227 0 : const __m128i out_06_4 =
1228 : _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
1229 0 : const __m128i out_06_5 =
1230 : _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
1231 0 : const __m128i out_22_4 =
1232 : _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
1233 0 : const __m128i out_22_5 =
1234 : _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
1235 0 : const __m128i out_14_4 =
1236 : _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
1237 0 : const __m128i out_14_5 =
1238 : _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
1239 0 : const __m128i out_30_4 =
1240 : _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
1241 0 : const __m128i out_30_5 =
1242 : _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
1243 0 : const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
1244 0 : const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
1245 0 : const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
1246 0 : const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
1247 0 : const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
1248 0 : const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
1249 0 : const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
1250 0 : const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
1251 0 : const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
1252 0 : const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
1253 0 : const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
1254 0 : const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
1255 0 : const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
1256 0 : const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
1257 0 : const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
1258 0 : const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
1259 : // Combine
1260 0 : out[2] = _mm_packs_epi32(out_02_6, out_02_7);
1261 0 : out[18] = _mm_packs_epi32(out_18_6, out_18_7);
1262 0 : out[10] = _mm_packs_epi32(out_10_6, out_10_7);
1263 0 : out[26] = _mm_packs_epi32(out_26_6, out_26_7);
1264 0 : out[6] = _mm_packs_epi32(out_06_6, out_06_7);
1265 0 : out[22] = _mm_packs_epi32(out_22_6, out_22_7);
1266 0 : out[14] = _mm_packs_epi32(out_14_6, out_14_7);
1267 0 : out[30] = _mm_packs_epi32(out_30_6, out_30_7);
1268 : #if DCT_HIGH_BIT_DEPTH
1269 : overflow =
1270 : check_epi16_overflow_x8(&out[2], &out[18], &out[10], &out[26],
1271 : &out[6], &out[22], &out[14], &out[30]);
1272 : if (overflow) {
1273 : if (pass == 0)
1274 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1275 : else
1276 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1277 : return;
1278 : }
1279 : #endif // DCT_HIGH_BIT_DEPTH
1280 : }
1281 : {
1282 0 : step1[16] = ADD_EPI16(step3[17], step2[16]);
1283 0 : step1[17] = SUB_EPI16(step2[16], step3[17]);
1284 0 : step1[18] = SUB_EPI16(step2[19], step3[18]);
1285 0 : step1[19] = ADD_EPI16(step3[18], step2[19]);
1286 0 : step1[20] = ADD_EPI16(step3[21], step2[20]);
1287 0 : step1[21] = SUB_EPI16(step2[20], step3[21]);
1288 0 : step1[22] = SUB_EPI16(step2[23], step3[22]);
1289 0 : step1[23] = ADD_EPI16(step3[22], step2[23]);
1290 0 : step1[24] = ADD_EPI16(step3[25], step2[24]);
1291 0 : step1[25] = SUB_EPI16(step2[24], step3[25]);
1292 0 : step1[26] = SUB_EPI16(step2[27], step3[26]);
1293 0 : step1[27] = ADD_EPI16(step3[26], step2[27]);
1294 0 : step1[28] = ADD_EPI16(step3[29], step2[28]);
1295 0 : step1[29] = SUB_EPI16(step2[28], step3[29]);
1296 0 : step1[30] = SUB_EPI16(step2[31], step3[30]);
1297 0 : step1[31] = ADD_EPI16(step3[30], step2[31]);
1298 : #if DCT_HIGH_BIT_DEPTH
1299 : overflow = check_epi16_overflow_x16(
1300 : &step1[16], &step1[17], &step1[18], &step1[19], &step1[20],
1301 : &step1[21], &step1[22], &step1[23], &step1[24], &step1[25],
1302 : &step1[26], &step1[27], &step1[28], &step1[29], &step1[30],
1303 : &step1[31]);
1304 : if (overflow) {
1305 : if (pass == 0)
1306 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1307 : else
1308 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1309 : return;
1310 : }
1311 : #endif // DCT_HIGH_BIT_DEPTH
1312 : }
1313 : // Final stage --- outputs indices are bit-reversed.
1314 : {
1315 0 : const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
1316 0 : const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
1317 0 : const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
1318 0 : const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
1319 0 : const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
1320 0 : const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
1321 0 : const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
1322 0 : const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
1323 0 : const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
1324 0 : const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
1325 0 : const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
1326 0 : const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
1327 0 : const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
1328 0 : const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
1329 0 : const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
1330 0 : const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
1331 0 : const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
1332 0 : const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
1333 0 : const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
1334 0 : const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
1335 0 : const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
1336 0 : const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
1337 0 : const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
1338 0 : const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
1339 : // dct_const_round_shift
1340 0 : const __m128i out_01_4 =
1341 : _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
1342 0 : const __m128i out_01_5 =
1343 : _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
1344 0 : const __m128i out_17_4 =
1345 : _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
1346 0 : const __m128i out_17_5 =
1347 : _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
1348 0 : const __m128i out_09_4 =
1349 : _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
1350 0 : const __m128i out_09_5 =
1351 : _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
1352 0 : const __m128i out_25_4 =
1353 : _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
1354 0 : const __m128i out_25_5 =
1355 : _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
1356 0 : const __m128i out_07_4 =
1357 : _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
1358 0 : const __m128i out_07_5 =
1359 : _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
1360 0 : const __m128i out_23_4 =
1361 : _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
1362 0 : const __m128i out_23_5 =
1363 : _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
1364 0 : const __m128i out_15_4 =
1365 : _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
1366 0 : const __m128i out_15_5 =
1367 : _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
1368 0 : const __m128i out_31_4 =
1369 : _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
1370 0 : const __m128i out_31_5 =
1371 : _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
1372 0 : const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
1373 0 : const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
1374 0 : const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
1375 0 : const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
1376 0 : const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
1377 0 : const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
1378 0 : const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
1379 0 : const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
1380 0 : const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
1381 0 : const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
1382 0 : const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
1383 0 : const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
1384 0 : const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
1385 0 : const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
1386 0 : const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
1387 0 : const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
1388 : // Combine
1389 0 : out[1] = _mm_packs_epi32(out_01_6, out_01_7);
1390 0 : out[17] = _mm_packs_epi32(out_17_6, out_17_7);
1391 0 : out[9] = _mm_packs_epi32(out_09_6, out_09_7);
1392 0 : out[25] = _mm_packs_epi32(out_25_6, out_25_7);
1393 0 : out[7] = _mm_packs_epi32(out_07_6, out_07_7);
1394 0 : out[23] = _mm_packs_epi32(out_23_6, out_23_7);
1395 0 : out[15] = _mm_packs_epi32(out_15_6, out_15_7);
1396 0 : out[31] = _mm_packs_epi32(out_31_6, out_31_7);
1397 : #if DCT_HIGH_BIT_DEPTH
1398 : overflow =
1399 : check_epi16_overflow_x8(&out[1], &out[17], &out[9], &out[25],
1400 : &out[7], &out[23], &out[15], &out[31]);
1401 : if (overflow) {
1402 : if (pass == 0)
1403 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1404 : else
1405 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1406 : return;
1407 : }
1408 : #endif // DCT_HIGH_BIT_DEPTH
1409 : }
1410 : {
1411 0 : const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
1412 0 : const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
1413 0 : const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
1414 0 : const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
1415 0 : const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
1416 0 : const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
1417 0 : const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
1418 0 : const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
1419 0 : const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
1420 0 : const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
1421 0 : const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
1422 0 : const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
1423 0 : const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
1424 0 : const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
1425 0 : const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
1426 0 : const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
1427 0 : const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
1428 0 : const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
1429 0 : const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
1430 0 : const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
1431 0 : const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
1432 0 : const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
1433 0 : const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
1434 0 : const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
1435 : // dct_const_round_shift
1436 0 : const __m128i out_05_4 =
1437 : _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
1438 0 : const __m128i out_05_5 =
1439 : _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
1440 0 : const __m128i out_21_4 =
1441 : _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
1442 0 : const __m128i out_21_5 =
1443 : _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
1444 0 : const __m128i out_13_4 =
1445 : _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
1446 0 : const __m128i out_13_5 =
1447 : _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
1448 0 : const __m128i out_29_4 =
1449 : _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
1450 0 : const __m128i out_29_5 =
1451 : _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
1452 0 : const __m128i out_03_4 =
1453 : _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
1454 0 : const __m128i out_03_5 =
1455 : _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
1456 0 : const __m128i out_19_4 =
1457 : _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
1458 0 : const __m128i out_19_5 =
1459 : _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
1460 0 : const __m128i out_11_4 =
1461 : _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
1462 0 : const __m128i out_11_5 =
1463 : _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
1464 0 : const __m128i out_27_4 =
1465 : _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
1466 0 : const __m128i out_27_5 =
1467 : _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
1468 0 : const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
1469 0 : const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
1470 0 : const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
1471 0 : const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
1472 0 : const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
1473 0 : const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
1474 0 : const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
1475 0 : const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
1476 0 : const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
1477 0 : const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
1478 0 : const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
1479 0 : const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
1480 0 : const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
1481 0 : const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
1482 0 : const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
1483 0 : const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
1484 : // Combine
1485 0 : out[5] = _mm_packs_epi32(out_05_6, out_05_7);
1486 0 : out[21] = _mm_packs_epi32(out_21_6, out_21_7);
1487 0 : out[13] = _mm_packs_epi32(out_13_6, out_13_7);
1488 0 : out[29] = _mm_packs_epi32(out_29_6, out_29_7);
1489 0 : out[3] = _mm_packs_epi32(out_03_6, out_03_7);
1490 0 : out[19] = _mm_packs_epi32(out_19_6, out_19_7);
1491 0 : out[11] = _mm_packs_epi32(out_11_6, out_11_7);
1492 0 : out[27] = _mm_packs_epi32(out_27_6, out_27_7);
1493 : #if DCT_HIGH_BIT_DEPTH
1494 : overflow =
1495 : check_epi16_overflow_x8(&out[5], &out[21], &out[13], &out[29],
1496 : &out[3], &out[19], &out[11], &out[27]);
1497 : if (overflow) {
1498 : if (pass == 0)
1499 : HIGH_FDCT32x32_2D_C(input, output_org, stride);
1500 : else
1501 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1502 : return;
1503 : }
1504 : #endif // DCT_HIGH_BIT_DEPTH
1505 : }
1506 : #if FDCT32x32_HIGH_PRECISION
1507 : } else {
1508 : __m128i lstep1[64], lstep2[64], lstep3[64];
1509 : __m128i u[32], v[32], sign[16];
1510 0 : const __m128i K32One = _mm_set_epi32(1, 1, 1, 1);
1511 : // start using 32-bit operations
1512 : // stage 3
1513 : {
1514 : // expanding to 32-bit length priori to addition operations
1515 0 : lstep2[0] = _mm_unpacklo_epi16(step2[0], kZero);
1516 0 : lstep2[1] = _mm_unpackhi_epi16(step2[0], kZero);
1517 0 : lstep2[2] = _mm_unpacklo_epi16(step2[1], kZero);
1518 0 : lstep2[3] = _mm_unpackhi_epi16(step2[1], kZero);
1519 0 : lstep2[4] = _mm_unpacklo_epi16(step2[2], kZero);
1520 0 : lstep2[5] = _mm_unpackhi_epi16(step2[2], kZero);
1521 0 : lstep2[6] = _mm_unpacklo_epi16(step2[3], kZero);
1522 0 : lstep2[7] = _mm_unpackhi_epi16(step2[3], kZero);
1523 0 : lstep2[8] = _mm_unpacklo_epi16(step2[4], kZero);
1524 0 : lstep2[9] = _mm_unpackhi_epi16(step2[4], kZero);
1525 0 : lstep2[10] = _mm_unpacklo_epi16(step2[5], kZero);
1526 0 : lstep2[11] = _mm_unpackhi_epi16(step2[5], kZero);
1527 0 : lstep2[12] = _mm_unpacklo_epi16(step2[6], kZero);
1528 0 : lstep2[13] = _mm_unpackhi_epi16(step2[6], kZero);
1529 0 : lstep2[14] = _mm_unpacklo_epi16(step2[7], kZero);
1530 0 : lstep2[15] = _mm_unpackhi_epi16(step2[7], kZero);
1531 0 : lstep2[0] = _mm_madd_epi16(lstep2[0], kOne);
1532 0 : lstep2[1] = _mm_madd_epi16(lstep2[1], kOne);
1533 0 : lstep2[2] = _mm_madd_epi16(lstep2[2], kOne);
1534 0 : lstep2[3] = _mm_madd_epi16(lstep2[3], kOne);
1535 0 : lstep2[4] = _mm_madd_epi16(lstep2[4], kOne);
1536 0 : lstep2[5] = _mm_madd_epi16(lstep2[5], kOne);
1537 0 : lstep2[6] = _mm_madd_epi16(lstep2[6], kOne);
1538 0 : lstep2[7] = _mm_madd_epi16(lstep2[7], kOne);
1539 0 : lstep2[8] = _mm_madd_epi16(lstep2[8], kOne);
1540 0 : lstep2[9] = _mm_madd_epi16(lstep2[9], kOne);
1541 0 : lstep2[10] = _mm_madd_epi16(lstep2[10], kOne);
1542 0 : lstep2[11] = _mm_madd_epi16(lstep2[11], kOne);
1543 0 : lstep2[12] = _mm_madd_epi16(lstep2[12], kOne);
1544 0 : lstep2[13] = _mm_madd_epi16(lstep2[13], kOne);
1545 0 : lstep2[14] = _mm_madd_epi16(lstep2[14], kOne);
1546 0 : lstep2[15] = _mm_madd_epi16(lstep2[15], kOne);
1547 :
1548 0 : lstep3[0] = _mm_add_epi32(lstep2[14], lstep2[0]);
1549 0 : lstep3[1] = _mm_add_epi32(lstep2[15], lstep2[1]);
1550 0 : lstep3[2] = _mm_add_epi32(lstep2[12], lstep2[2]);
1551 0 : lstep3[3] = _mm_add_epi32(lstep2[13], lstep2[3]);
1552 0 : lstep3[4] = _mm_add_epi32(lstep2[10], lstep2[4]);
1553 0 : lstep3[5] = _mm_add_epi32(lstep2[11], lstep2[5]);
1554 0 : lstep3[6] = _mm_add_epi32(lstep2[8], lstep2[6]);
1555 0 : lstep3[7] = _mm_add_epi32(lstep2[9], lstep2[7]);
1556 0 : lstep3[8] = _mm_sub_epi32(lstep2[6], lstep2[8]);
1557 0 : lstep3[9] = _mm_sub_epi32(lstep2[7], lstep2[9]);
1558 0 : lstep3[10] = _mm_sub_epi32(lstep2[4], lstep2[10]);
1559 0 : lstep3[11] = _mm_sub_epi32(lstep2[5], lstep2[11]);
1560 0 : lstep3[12] = _mm_sub_epi32(lstep2[2], lstep2[12]);
1561 0 : lstep3[13] = _mm_sub_epi32(lstep2[3], lstep2[13]);
1562 0 : lstep3[14] = _mm_sub_epi32(lstep2[0], lstep2[14]);
1563 0 : lstep3[15] = _mm_sub_epi32(lstep2[1], lstep2[15]);
1564 : }
1565 : {
1566 0 : const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
1567 0 : const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
1568 0 : const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
1569 0 : const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
1570 0 : const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
1571 0 : const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
1572 0 : const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
1573 0 : const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
1574 0 : const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
1575 0 : const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
1576 0 : const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
1577 0 : const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
1578 : // dct_const_round_shift
1579 0 : const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
1580 0 : const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
1581 0 : const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
1582 0 : const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
1583 0 : const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
1584 0 : const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
1585 0 : const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
1586 0 : const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
1587 0 : lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
1588 0 : lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
1589 0 : lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
1590 0 : lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
1591 0 : lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
1592 0 : lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
1593 0 : lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
1594 0 : lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
1595 : }
1596 : {
1597 0 : lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero);
1598 0 : lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero);
1599 0 : lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero);
1600 0 : lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero);
1601 0 : lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero);
1602 0 : lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero);
1603 0 : lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero);
1604 0 : lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero);
1605 0 : lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero);
1606 0 : lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero);
1607 0 : lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero);
1608 0 : lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero);
1609 0 : lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero);
1610 0 : lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero);
1611 0 : lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero);
1612 0 : lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero);
1613 0 : lstep2[40] = _mm_madd_epi16(lstep2[40], kOne);
1614 0 : lstep2[41] = _mm_madd_epi16(lstep2[41], kOne);
1615 0 : lstep2[42] = _mm_madd_epi16(lstep2[42], kOne);
1616 0 : lstep2[43] = _mm_madd_epi16(lstep2[43], kOne);
1617 0 : lstep2[44] = _mm_madd_epi16(lstep2[44], kOne);
1618 0 : lstep2[45] = _mm_madd_epi16(lstep2[45], kOne);
1619 0 : lstep2[46] = _mm_madd_epi16(lstep2[46], kOne);
1620 0 : lstep2[47] = _mm_madd_epi16(lstep2[47], kOne);
1621 0 : lstep2[48] = _mm_madd_epi16(lstep2[48], kOne);
1622 0 : lstep2[49] = _mm_madd_epi16(lstep2[49], kOne);
1623 0 : lstep2[50] = _mm_madd_epi16(lstep2[50], kOne);
1624 0 : lstep2[51] = _mm_madd_epi16(lstep2[51], kOne);
1625 0 : lstep2[52] = _mm_madd_epi16(lstep2[52], kOne);
1626 0 : lstep2[53] = _mm_madd_epi16(lstep2[53], kOne);
1627 0 : lstep2[54] = _mm_madd_epi16(lstep2[54], kOne);
1628 0 : lstep2[55] = _mm_madd_epi16(lstep2[55], kOne);
1629 :
1630 0 : lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero);
1631 0 : lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero);
1632 0 : lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero);
1633 0 : lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero);
1634 0 : lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero);
1635 0 : lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero);
1636 0 : lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero);
1637 0 : lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero);
1638 0 : lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero);
1639 0 : lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero);
1640 0 : lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero);
1641 0 : lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero);
1642 0 : lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero);
1643 0 : lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero);
1644 0 : lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero);
1645 0 : lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero);
1646 0 : lstep1[32] = _mm_madd_epi16(lstep1[32], kOne);
1647 0 : lstep1[33] = _mm_madd_epi16(lstep1[33], kOne);
1648 0 : lstep1[34] = _mm_madd_epi16(lstep1[34], kOne);
1649 0 : lstep1[35] = _mm_madd_epi16(lstep1[35], kOne);
1650 0 : lstep1[36] = _mm_madd_epi16(lstep1[36], kOne);
1651 0 : lstep1[37] = _mm_madd_epi16(lstep1[37], kOne);
1652 0 : lstep1[38] = _mm_madd_epi16(lstep1[38], kOne);
1653 0 : lstep1[39] = _mm_madd_epi16(lstep1[39], kOne);
1654 0 : lstep1[56] = _mm_madd_epi16(lstep1[56], kOne);
1655 0 : lstep1[57] = _mm_madd_epi16(lstep1[57], kOne);
1656 0 : lstep1[58] = _mm_madd_epi16(lstep1[58], kOne);
1657 0 : lstep1[59] = _mm_madd_epi16(lstep1[59], kOne);
1658 0 : lstep1[60] = _mm_madd_epi16(lstep1[60], kOne);
1659 0 : lstep1[61] = _mm_madd_epi16(lstep1[61], kOne);
1660 0 : lstep1[62] = _mm_madd_epi16(lstep1[62], kOne);
1661 0 : lstep1[63] = _mm_madd_epi16(lstep1[63], kOne);
1662 :
1663 0 : lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]);
1664 0 : lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]);
1665 :
1666 0 : lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]);
1667 0 : lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]);
1668 0 : lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]);
1669 0 : lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]);
1670 0 : lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]);
1671 0 : lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]);
1672 0 : lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]);
1673 0 : lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]);
1674 0 : lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]);
1675 0 : lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]);
1676 0 : lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]);
1677 0 : lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]);
1678 0 : lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]);
1679 0 : lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]);
1680 0 : lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]);
1681 0 : lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]);
1682 0 : lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]);
1683 0 : lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]);
1684 0 : lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]);
1685 0 : lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]);
1686 0 : lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]);
1687 0 : lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]);
1688 0 : lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]);
1689 0 : lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]);
1690 0 : lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]);
1691 0 : lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]);
1692 0 : lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]);
1693 0 : lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]);
1694 0 : lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]);
1695 0 : lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]);
1696 : }
1697 :
1698 : // stage 4
1699 : {
1700 : // expanding to 32-bit length priori to addition operations
1701 0 : lstep2[16] = _mm_unpacklo_epi16(step2[8], kZero);
1702 0 : lstep2[17] = _mm_unpackhi_epi16(step2[8], kZero);
1703 0 : lstep2[18] = _mm_unpacklo_epi16(step2[9], kZero);
1704 0 : lstep2[19] = _mm_unpackhi_epi16(step2[9], kZero);
1705 0 : lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero);
1706 0 : lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero);
1707 0 : lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero);
1708 0 : lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero);
1709 0 : lstep2[16] = _mm_madd_epi16(lstep2[16], kOne);
1710 0 : lstep2[17] = _mm_madd_epi16(lstep2[17], kOne);
1711 0 : lstep2[18] = _mm_madd_epi16(lstep2[18], kOne);
1712 0 : lstep2[19] = _mm_madd_epi16(lstep2[19], kOne);
1713 0 : lstep2[28] = _mm_madd_epi16(lstep2[28], kOne);
1714 0 : lstep2[29] = _mm_madd_epi16(lstep2[29], kOne);
1715 0 : lstep2[30] = _mm_madd_epi16(lstep2[30], kOne);
1716 0 : lstep2[31] = _mm_madd_epi16(lstep2[31], kOne);
1717 :
1718 0 : lstep1[0] = _mm_add_epi32(lstep3[6], lstep3[0]);
1719 0 : lstep1[1] = _mm_add_epi32(lstep3[7], lstep3[1]);
1720 0 : lstep1[2] = _mm_add_epi32(lstep3[4], lstep3[2]);
1721 0 : lstep1[3] = _mm_add_epi32(lstep3[5], lstep3[3]);
1722 0 : lstep1[4] = _mm_sub_epi32(lstep3[2], lstep3[4]);
1723 0 : lstep1[5] = _mm_sub_epi32(lstep3[3], lstep3[5]);
1724 0 : lstep1[6] = _mm_sub_epi32(lstep3[0], lstep3[6]);
1725 0 : lstep1[7] = _mm_sub_epi32(lstep3[1], lstep3[7]);
1726 0 : lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]);
1727 0 : lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]);
1728 0 : lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]);
1729 0 : lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]);
1730 0 : lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]);
1731 0 : lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]);
1732 0 : lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]);
1733 0 : lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]);
1734 0 : lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]);
1735 0 : lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]);
1736 0 : lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]);
1737 0 : lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]);
1738 0 : lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]);
1739 0 : lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]);
1740 0 : lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]);
1741 0 : lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]);
1742 : }
1743 : {
1744 : // to be continued...
1745 : //
1746 0 : const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
1747 0 : const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
1748 :
1749 0 : u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]);
1750 0 : u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]);
1751 0 : u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]);
1752 0 : u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]);
1753 :
1754 : // TODO(jingning): manually inline k_madd_epi32_ to further hide
1755 : // instruction latency.
1756 0 : v[0] = k_madd_epi32(u[0], k32_p16_m16);
1757 0 : v[1] = k_madd_epi32(u[1], k32_p16_m16);
1758 0 : v[2] = k_madd_epi32(u[2], k32_p16_m16);
1759 0 : v[3] = k_madd_epi32(u[3], k32_p16_m16);
1760 0 : v[4] = k_madd_epi32(u[0], k32_p16_p16);
1761 0 : v[5] = k_madd_epi32(u[1], k32_p16_p16);
1762 0 : v[6] = k_madd_epi32(u[2], k32_p16_p16);
1763 0 : v[7] = k_madd_epi32(u[3], k32_p16_p16);
1764 : #if DCT_HIGH_BIT_DEPTH
1765 : overflow = k_check_epi32_overflow_8(&v[0], &v[1], &v[2], &v[3], &v[4],
1766 : &v[5], &v[6], &v[7], &kZero);
1767 : if (overflow) {
1768 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1769 : return;
1770 : }
1771 : #endif // DCT_HIGH_BIT_DEPTH
1772 0 : u[0] = k_packs_epi64(v[0], v[1]);
1773 0 : u[1] = k_packs_epi64(v[2], v[3]);
1774 0 : u[2] = k_packs_epi64(v[4], v[5]);
1775 0 : u[3] = k_packs_epi64(v[6], v[7]);
1776 :
1777 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1778 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1779 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1780 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1781 :
1782 0 : lstep1[10] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1783 0 : lstep1[11] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1784 0 : lstep1[12] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1785 0 : lstep1[13] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1786 : }
1787 : {
1788 0 : const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
1789 0 : const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
1790 0 : const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
1791 :
1792 0 : u[0] = _mm_unpacklo_epi32(lstep3[36], lstep3[58]);
1793 0 : u[1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]);
1794 0 : u[2] = _mm_unpacklo_epi32(lstep3[37], lstep3[59]);
1795 0 : u[3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]);
1796 0 : u[4] = _mm_unpacklo_epi32(lstep3[38], lstep3[56]);
1797 0 : u[5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]);
1798 0 : u[6] = _mm_unpacklo_epi32(lstep3[39], lstep3[57]);
1799 0 : u[7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]);
1800 0 : u[8] = _mm_unpacklo_epi32(lstep3[40], lstep3[54]);
1801 0 : u[9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]);
1802 0 : u[10] = _mm_unpacklo_epi32(lstep3[41], lstep3[55]);
1803 0 : u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]);
1804 0 : u[12] = _mm_unpacklo_epi32(lstep3[42], lstep3[52]);
1805 0 : u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]);
1806 0 : u[14] = _mm_unpacklo_epi32(lstep3[43], lstep3[53]);
1807 0 : u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]);
1808 :
1809 0 : v[0] = k_madd_epi32(u[0], k32_m08_p24);
1810 0 : v[1] = k_madd_epi32(u[1], k32_m08_p24);
1811 0 : v[2] = k_madd_epi32(u[2], k32_m08_p24);
1812 0 : v[3] = k_madd_epi32(u[3], k32_m08_p24);
1813 0 : v[4] = k_madd_epi32(u[4], k32_m08_p24);
1814 0 : v[5] = k_madd_epi32(u[5], k32_m08_p24);
1815 0 : v[6] = k_madd_epi32(u[6], k32_m08_p24);
1816 0 : v[7] = k_madd_epi32(u[7], k32_m08_p24);
1817 0 : v[8] = k_madd_epi32(u[8], k32_m24_m08);
1818 0 : v[9] = k_madd_epi32(u[9], k32_m24_m08);
1819 0 : v[10] = k_madd_epi32(u[10], k32_m24_m08);
1820 0 : v[11] = k_madd_epi32(u[11], k32_m24_m08);
1821 0 : v[12] = k_madd_epi32(u[12], k32_m24_m08);
1822 0 : v[13] = k_madd_epi32(u[13], k32_m24_m08);
1823 0 : v[14] = k_madd_epi32(u[14], k32_m24_m08);
1824 0 : v[15] = k_madd_epi32(u[15], k32_m24_m08);
1825 0 : v[16] = k_madd_epi32(u[12], k32_m08_p24);
1826 0 : v[17] = k_madd_epi32(u[13], k32_m08_p24);
1827 0 : v[18] = k_madd_epi32(u[14], k32_m08_p24);
1828 0 : v[19] = k_madd_epi32(u[15], k32_m08_p24);
1829 0 : v[20] = k_madd_epi32(u[8], k32_m08_p24);
1830 0 : v[21] = k_madd_epi32(u[9], k32_m08_p24);
1831 0 : v[22] = k_madd_epi32(u[10], k32_m08_p24);
1832 0 : v[23] = k_madd_epi32(u[11], k32_m08_p24);
1833 0 : v[24] = k_madd_epi32(u[4], k32_p24_p08);
1834 0 : v[25] = k_madd_epi32(u[5], k32_p24_p08);
1835 0 : v[26] = k_madd_epi32(u[6], k32_p24_p08);
1836 0 : v[27] = k_madd_epi32(u[7], k32_p24_p08);
1837 0 : v[28] = k_madd_epi32(u[0], k32_p24_p08);
1838 0 : v[29] = k_madd_epi32(u[1], k32_p24_p08);
1839 0 : v[30] = k_madd_epi32(u[2], k32_p24_p08);
1840 0 : v[31] = k_madd_epi32(u[3], k32_p24_p08);
1841 :
1842 : #if DCT_HIGH_BIT_DEPTH
1843 : overflow = k_check_epi32_overflow_32(
1844 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
1845 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
1846 : &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
1847 : &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
1848 : if (overflow) {
1849 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1850 : return;
1851 : }
1852 : #endif // DCT_HIGH_BIT_DEPTH
1853 0 : u[0] = k_packs_epi64(v[0], v[1]);
1854 0 : u[1] = k_packs_epi64(v[2], v[3]);
1855 0 : u[2] = k_packs_epi64(v[4], v[5]);
1856 0 : u[3] = k_packs_epi64(v[6], v[7]);
1857 0 : u[4] = k_packs_epi64(v[8], v[9]);
1858 0 : u[5] = k_packs_epi64(v[10], v[11]);
1859 0 : u[6] = k_packs_epi64(v[12], v[13]);
1860 0 : u[7] = k_packs_epi64(v[14], v[15]);
1861 0 : u[8] = k_packs_epi64(v[16], v[17]);
1862 0 : u[9] = k_packs_epi64(v[18], v[19]);
1863 0 : u[10] = k_packs_epi64(v[20], v[21]);
1864 0 : u[11] = k_packs_epi64(v[22], v[23]);
1865 0 : u[12] = k_packs_epi64(v[24], v[25]);
1866 0 : u[13] = k_packs_epi64(v[26], v[27]);
1867 0 : u[14] = k_packs_epi64(v[28], v[29]);
1868 0 : u[15] = k_packs_epi64(v[30], v[31]);
1869 :
1870 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1871 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1872 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1873 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1874 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1875 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1876 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1877 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1878 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1879 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1880 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1881 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1882 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1883 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1884 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1885 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1886 :
1887 0 : lstep1[36] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1888 0 : lstep1[37] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1889 0 : lstep1[38] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1890 0 : lstep1[39] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1891 0 : lstep1[40] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1892 0 : lstep1[41] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1893 0 : lstep1[42] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1894 0 : lstep1[43] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1895 0 : lstep1[52] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
1896 0 : lstep1[53] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
1897 0 : lstep1[54] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1898 0 : lstep1[55] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1899 0 : lstep1[56] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1900 0 : lstep1[57] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1901 0 : lstep1[58] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1902 0 : lstep1[59] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1903 : }
1904 : // stage 5
1905 : {
1906 0 : lstep2[8] = _mm_add_epi32(lstep1[10], lstep3[8]);
1907 0 : lstep2[9] = _mm_add_epi32(lstep1[11], lstep3[9]);
1908 0 : lstep2[10] = _mm_sub_epi32(lstep3[8], lstep1[10]);
1909 0 : lstep2[11] = _mm_sub_epi32(lstep3[9], lstep1[11]);
1910 0 : lstep2[12] = _mm_sub_epi32(lstep3[14], lstep1[12]);
1911 0 : lstep2[13] = _mm_sub_epi32(lstep3[15], lstep1[13]);
1912 0 : lstep2[14] = _mm_add_epi32(lstep1[12], lstep3[14]);
1913 0 : lstep2[15] = _mm_add_epi32(lstep1[13], lstep3[15]);
1914 : }
1915 : {
1916 0 : const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
1917 0 : const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
1918 0 : const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
1919 0 : const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
1920 :
1921 0 : u[0] = _mm_unpacklo_epi32(lstep1[0], lstep1[2]);
1922 0 : u[1] = _mm_unpackhi_epi32(lstep1[0], lstep1[2]);
1923 0 : u[2] = _mm_unpacklo_epi32(lstep1[1], lstep1[3]);
1924 0 : u[3] = _mm_unpackhi_epi32(lstep1[1], lstep1[3]);
1925 0 : u[4] = _mm_unpacklo_epi32(lstep1[4], lstep1[6]);
1926 0 : u[5] = _mm_unpackhi_epi32(lstep1[4], lstep1[6]);
1927 0 : u[6] = _mm_unpacklo_epi32(lstep1[5], lstep1[7]);
1928 0 : u[7] = _mm_unpackhi_epi32(lstep1[5], lstep1[7]);
1929 :
1930 : // TODO(jingning): manually inline k_madd_epi32_ to further hide
1931 : // instruction latency.
1932 0 : v[0] = k_madd_epi32(u[0], k32_p16_p16);
1933 0 : v[1] = k_madd_epi32(u[1], k32_p16_p16);
1934 0 : v[2] = k_madd_epi32(u[2], k32_p16_p16);
1935 0 : v[3] = k_madd_epi32(u[3], k32_p16_p16);
1936 0 : v[4] = k_madd_epi32(u[0], k32_p16_m16);
1937 0 : v[5] = k_madd_epi32(u[1], k32_p16_m16);
1938 0 : v[6] = k_madd_epi32(u[2], k32_p16_m16);
1939 0 : v[7] = k_madd_epi32(u[3], k32_p16_m16);
1940 0 : v[8] = k_madd_epi32(u[4], k32_p24_p08);
1941 0 : v[9] = k_madd_epi32(u[5], k32_p24_p08);
1942 0 : v[10] = k_madd_epi32(u[6], k32_p24_p08);
1943 0 : v[11] = k_madd_epi32(u[7], k32_p24_p08);
1944 0 : v[12] = k_madd_epi32(u[4], k32_m08_p24);
1945 0 : v[13] = k_madd_epi32(u[5], k32_m08_p24);
1946 0 : v[14] = k_madd_epi32(u[6], k32_m08_p24);
1947 0 : v[15] = k_madd_epi32(u[7], k32_m08_p24);
1948 :
1949 : #if DCT_HIGH_BIT_DEPTH
1950 : overflow = k_check_epi32_overflow_16(
1951 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
1952 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero);
1953 : if (overflow) {
1954 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
1955 : return;
1956 : }
1957 : #endif // DCT_HIGH_BIT_DEPTH
1958 0 : u[0] = k_packs_epi64(v[0], v[1]);
1959 0 : u[1] = k_packs_epi64(v[2], v[3]);
1960 0 : u[2] = k_packs_epi64(v[4], v[5]);
1961 0 : u[3] = k_packs_epi64(v[6], v[7]);
1962 0 : u[4] = k_packs_epi64(v[8], v[9]);
1963 0 : u[5] = k_packs_epi64(v[10], v[11]);
1964 0 : u[6] = k_packs_epi64(v[12], v[13]);
1965 0 : u[7] = k_packs_epi64(v[14], v[15]);
1966 :
1967 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1968 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1969 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1970 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1971 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1972 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1973 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1974 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1975 :
1976 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1977 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1978 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1979 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1980 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1981 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1982 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1983 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1984 :
1985 0 : sign[0] = _mm_cmplt_epi32(u[0], kZero);
1986 0 : sign[1] = _mm_cmplt_epi32(u[1], kZero);
1987 0 : sign[2] = _mm_cmplt_epi32(u[2], kZero);
1988 0 : sign[3] = _mm_cmplt_epi32(u[3], kZero);
1989 0 : sign[4] = _mm_cmplt_epi32(u[4], kZero);
1990 0 : sign[5] = _mm_cmplt_epi32(u[5], kZero);
1991 0 : sign[6] = _mm_cmplt_epi32(u[6], kZero);
1992 0 : sign[7] = _mm_cmplt_epi32(u[7], kZero);
1993 :
1994 0 : u[0] = _mm_sub_epi32(u[0], sign[0]);
1995 0 : u[1] = _mm_sub_epi32(u[1], sign[1]);
1996 0 : u[2] = _mm_sub_epi32(u[2], sign[2]);
1997 0 : u[3] = _mm_sub_epi32(u[3], sign[3]);
1998 0 : u[4] = _mm_sub_epi32(u[4], sign[4]);
1999 0 : u[5] = _mm_sub_epi32(u[5], sign[5]);
2000 0 : u[6] = _mm_sub_epi32(u[6], sign[6]);
2001 0 : u[7] = _mm_sub_epi32(u[7], sign[7]);
2002 :
2003 0 : u[0] = _mm_add_epi32(u[0], K32One);
2004 0 : u[1] = _mm_add_epi32(u[1], K32One);
2005 0 : u[2] = _mm_add_epi32(u[2], K32One);
2006 0 : u[3] = _mm_add_epi32(u[3], K32One);
2007 0 : u[4] = _mm_add_epi32(u[4], K32One);
2008 0 : u[5] = _mm_add_epi32(u[5], K32One);
2009 0 : u[6] = _mm_add_epi32(u[6], K32One);
2010 0 : u[7] = _mm_add_epi32(u[7], K32One);
2011 :
2012 0 : u[0] = _mm_srai_epi32(u[0], 2);
2013 0 : u[1] = _mm_srai_epi32(u[1], 2);
2014 0 : u[2] = _mm_srai_epi32(u[2], 2);
2015 0 : u[3] = _mm_srai_epi32(u[3], 2);
2016 0 : u[4] = _mm_srai_epi32(u[4], 2);
2017 0 : u[5] = _mm_srai_epi32(u[5], 2);
2018 0 : u[6] = _mm_srai_epi32(u[6], 2);
2019 0 : u[7] = _mm_srai_epi32(u[7], 2);
2020 :
2021 : // Combine
2022 0 : out[0] = _mm_packs_epi32(u[0], u[1]);
2023 0 : out[16] = _mm_packs_epi32(u[2], u[3]);
2024 0 : out[8] = _mm_packs_epi32(u[4], u[5]);
2025 0 : out[24] = _mm_packs_epi32(u[6], u[7]);
2026 : #if DCT_HIGH_BIT_DEPTH
2027 : overflow =
2028 : check_epi16_overflow_x4(&out[0], &out[16], &out[8], &out[24]);
2029 : if (overflow) {
2030 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2031 : return;
2032 : }
2033 : #endif // DCT_HIGH_BIT_DEPTH
2034 : }
2035 : {
2036 0 : const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
2037 0 : const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
2038 0 : const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
2039 :
2040 0 : u[0] = _mm_unpacklo_epi32(lstep1[18], lstep1[28]);
2041 0 : u[1] = _mm_unpackhi_epi32(lstep1[18], lstep1[28]);
2042 0 : u[2] = _mm_unpacklo_epi32(lstep1[19], lstep1[29]);
2043 0 : u[3] = _mm_unpackhi_epi32(lstep1[19], lstep1[29]);
2044 0 : u[4] = _mm_unpacklo_epi32(lstep1[20], lstep1[26]);
2045 0 : u[5] = _mm_unpackhi_epi32(lstep1[20], lstep1[26]);
2046 0 : u[6] = _mm_unpacklo_epi32(lstep1[21], lstep1[27]);
2047 0 : u[7] = _mm_unpackhi_epi32(lstep1[21], lstep1[27]);
2048 :
2049 0 : v[0] = k_madd_epi32(u[0], k32_m08_p24);
2050 0 : v[1] = k_madd_epi32(u[1], k32_m08_p24);
2051 0 : v[2] = k_madd_epi32(u[2], k32_m08_p24);
2052 0 : v[3] = k_madd_epi32(u[3], k32_m08_p24);
2053 0 : v[4] = k_madd_epi32(u[4], k32_m24_m08);
2054 0 : v[5] = k_madd_epi32(u[5], k32_m24_m08);
2055 0 : v[6] = k_madd_epi32(u[6], k32_m24_m08);
2056 0 : v[7] = k_madd_epi32(u[7], k32_m24_m08);
2057 0 : v[8] = k_madd_epi32(u[4], k32_m08_p24);
2058 0 : v[9] = k_madd_epi32(u[5], k32_m08_p24);
2059 0 : v[10] = k_madd_epi32(u[6], k32_m08_p24);
2060 0 : v[11] = k_madd_epi32(u[7], k32_m08_p24);
2061 0 : v[12] = k_madd_epi32(u[0], k32_p24_p08);
2062 0 : v[13] = k_madd_epi32(u[1], k32_p24_p08);
2063 0 : v[14] = k_madd_epi32(u[2], k32_p24_p08);
2064 0 : v[15] = k_madd_epi32(u[3], k32_p24_p08);
2065 :
2066 : #if DCT_HIGH_BIT_DEPTH
2067 : overflow = k_check_epi32_overflow_16(
2068 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
2069 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero);
2070 : if (overflow) {
2071 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2072 : return;
2073 : }
2074 : #endif // DCT_HIGH_BIT_DEPTH
2075 0 : u[0] = k_packs_epi64(v[0], v[1]);
2076 0 : u[1] = k_packs_epi64(v[2], v[3]);
2077 0 : u[2] = k_packs_epi64(v[4], v[5]);
2078 0 : u[3] = k_packs_epi64(v[6], v[7]);
2079 0 : u[4] = k_packs_epi64(v[8], v[9]);
2080 0 : u[5] = k_packs_epi64(v[10], v[11]);
2081 0 : u[6] = k_packs_epi64(v[12], v[13]);
2082 0 : u[7] = k_packs_epi64(v[14], v[15]);
2083 :
2084 0 : u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2085 0 : u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2086 0 : u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2087 0 : u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2088 0 : u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2089 0 : u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2090 0 : u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2091 0 : u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2092 :
2093 0 : lstep2[18] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
2094 0 : lstep2[19] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
2095 0 : lstep2[20] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
2096 0 : lstep2[21] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
2097 0 : lstep2[26] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
2098 0 : lstep2[27] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
2099 0 : lstep2[28] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
2100 0 : lstep2[29] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
2101 : }
2102 : {
2103 0 : lstep2[32] = _mm_add_epi32(lstep1[38], lstep3[32]);
2104 0 : lstep2[33] = _mm_add_epi32(lstep1[39], lstep3[33]);
2105 0 : lstep2[34] = _mm_add_epi32(lstep1[36], lstep3[34]);
2106 0 : lstep2[35] = _mm_add_epi32(lstep1[37], lstep3[35]);
2107 0 : lstep2[36] = _mm_sub_epi32(lstep3[34], lstep1[36]);
2108 0 : lstep2[37] = _mm_sub_epi32(lstep3[35], lstep1[37]);
2109 0 : lstep2[38] = _mm_sub_epi32(lstep3[32], lstep1[38]);
2110 0 : lstep2[39] = _mm_sub_epi32(lstep3[33], lstep1[39]);
2111 0 : lstep2[40] = _mm_sub_epi32(lstep3[46], lstep1[40]);
2112 0 : lstep2[41] = _mm_sub_epi32(lstep3[47], lstep1[41]);
2113 0 : lstep2[42] = _mm_sub_epi32(lstep3[44], lstep1[42]);
2114 0 : lstep2[43] = _mm_sub_epi32(lstep3[45], lstep1[43]);
2115 0 : lstep2[44] = _mm_add_epi32(lstep1[42], lstep3[44]);
2116 0 : lstep2[45] = _mm_add_epi32(lstep1[43], lstep3[45]);
2117 0 : lstep2[46] = _mm_add_epi32(lstep1[40], lstep3[46]);
2118 0 : lstep2[47] = _mm_add_epi32(lstep1[41], lstep3[47]);
2119 0 : lstep2[48] = _mm_add_epi32(lstep1[54], lstep3[48]);
2120 0 : lstep2[49] = _mm_add_epi32(lstep1[55], lstep3[49]);
2121 0 : lstep2[50] = _mm_add_epi32(lstep1[52], lstep3[50]);
2122 0 : lstep2[51] = _mm_add_epi32(lstep1[53], lstep3[51]);
2123 0 : lstep2[52] = _mm_sub_epi32(lstep3[50], lstep1[52]);
2124 0 : lstep2[53] = _mm_sub_epi32(lstep3[51], lstep1[53]);
2125 0 : lstep2[54] = _mm_sub_epi32(lstep3[48], lstep1[54]);
2126 0 : lstep2[55] = _mm_sub_epi32(lstep3[49], lstep1[55]);
2127 0 : lstep2[56] = _mm_sub_epi32(lstep3[62], lstep1[56]);
2128 0 : lstep2[57] = _mm_sub_epi32(lstep3[63], lstep1[57]);
2129 0 : lstep2[58] = _mm_sub_epi32(lstep3[60], lstep1[58]);
2130 0 : lstep2[59] = _mm_sub_epi32(lstep3[61], lstep1[59]);
2131 0 : lstep2[60] = _mm_add_epi32(lstep1[58], lstep3[60]);
2132 0 : lstep2[61] = _mm_add_epi32(lstep1[59], lstep3[61]);
2133 0 : lstep2[62] = _mm_add_epi32(lstep1[56], lstep3[62]);
2134 0 : lstep2[63] = _mm_add_epi32(lstep1[57], lstep3[63]);
2135 : }
2136 : // stage 6
2137 : {
2138 0 : const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
2139 0 : const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
2140 0 : const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
2141 0 : const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
2142 :
2143 0 : u[0] = _mm_unpacklo_epi32(lstep2[8], lstep2[14]);
2144 0 : u[1] = _mm_unpackhi_epi32(lstep2[8], lstep2[14]);
2145 0 : u[2] = _mm_unpacklo_epi32(lstep2[9], lstep2[15]);
2146 0 : u[3] = _mm_unpackhi_epi32(lstep2[9], lstep2[15]);
2147 0 : u[4] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
2148 0 : u[5] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
2149 0 : u[6] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
2150 0 : u[7] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
2151 0 : u[8] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
2152 0 : u[9] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
2153 0 : u[10] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
2154 0 : u[11] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
2155 0 : u[12] = _mm_unpacklo_epi32(lstep2[8], lstep2[14]);
2156 0 : u[13] = _mm_unpackhi_epi32(lstep2[8], lstep2[14]);
2157 0 : u[14] = _mm_unpacklo_epi32(lstep2[9], lstep2[15]);
2158 0 : u[15] = _mm_unpackhi_epi32(lstep2[9], lstep2[15]);
2159 :
2160 0 : v[0] = k_madd_epi32(u[0], k32_p28_p04);
2161 0 : v[1] = k_madd_epi32(u[1], k32_p28_p04);
2162 0 : v[2] = k_madd_epi32(u[2], k32_p28_p04);
2163 0 : v[3] = k_madd_epi32(u[3], k32_p28_p04);
2164 0 : v[4] = k_madd_epi32(u[4], k32_p12_p20);
2165 0 : v[5] = k_madd_epi32(u[5], k32_p12_p20);
2166 0 : v[6] = k_madd_epi32(u[6], k32_p12_p20);
2167 0 : v[7] = k_madd_epi32(u[7], k32_p12_p20);
2168 0 : v[8] = k_madd_epi32(u[8], k32_m20_p12);
2169 0 : v[9] = k_madd_epi32(u[9], k32_m20_p12);
2170 0 : v[10] = k_madd_epi32(u[10], k32_m20_p12);
2171 0 : v[11] = k_madd_epi32(u[11], k32_m20_p12);
2172 0 : v[12] = k_madd_epi32(u[12], k32_m04_p28);
2173 0 : v[13] = k_madd_epi32(u[13], k32_m04_p28);
2174 0 : v[14] = k_madd_epi32(u[14], k32_m04_p28);
2175 0 : v[15] = k_madd_epi32(u[15], k32_m04_p28);
2176 :
2177 : #if DCT_HIGH_BIT_DEPTH
2178 : overflow = k_check_epi32_overflow_16(
2179 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
2180 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero);
2181 : if (overflow) {
2182 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2183 : return;
2184 : }
2185 : #endif // DCT_HIGH_BIT_DEPTH
2186 0 : u[0] = k_packs_epi64(v[0], v[1]);
2187 0 : u[1] = k_packs_epi64(v[2], v[3]);
2188 0 : u[2] = k_packs_epi64(v[4], v[5]);
2189 0 : u[3] = k_packs_epi64(v[6], v[7]);
2190 0 : u[4] = k_packs_epi64(v[8], v[9]);
2191 0 : u[5] = k_packs_epi64(v[10], v[11]);
2192 0 : u[6] = k_packs_epi64(v[12], v[13]);
2193 0 : u[7] = k_packs_epi64(v[14], v[15]);
2194 :
2195 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2196 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2197 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2198 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2199 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2200 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2201 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2202 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2203 :
2204 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
2205 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
2206 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
2207 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
2208 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
2209 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
2210 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
2211 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
2212 :
2213 0 : sign[0] = _mm_cmplt_epi32(u[0], kZero);
2214 0 : sign[1] = _mm_cmplt_epi32(u[1], kZero);
2215 0 : sign[2] = _mm_cmplt_epi32(u[2], kZero);
2216 0 : sign[3] = _mm_cmplt_epi32(u[3], kZero);
2217 0 : sign[4] = _mm_cmplt_epi32(u[4], kZero);
2218 0 : sign[5] = _mm_cmplt_epi32(u[5], kZero);
2219 0 : sign[6] = _mm_cmplt_epi32(u[6], kZero);
2220 0 : sign[7] = _mm_cmplt_epi32(u[7], kZero);
2221 :
2222 0 : u[0] = _mm_sub_epi32(u[0], sign[0]);
2223 0 : u[1] = _mm_sub_epi32(u[1], sign[1]);
2224 0 : u[2] = _mm_sub_epi32(u[2], sign[2]);
2225 0 : u[3] = _mm_sub_epi32(u[3], sign[3]);
2226 0 : u[4] = _mm_sub_epi32(u[4], sign[4]);
2227 0 : u[5] = _mm_sub_epi32(u[5], sign[5]);
2228 0 : u[6] = _mm_sub_epi32(u[6], sign[6]);
2229 0 : u[7] = _mm_sub_epi32(u[7], sign[7]);
2230 :
2231 0 : u[0] = _mm_add_epi32(u[0], K32One);
2232 0 : u[1] = _mm_add_epi32(u[1], K32One);
2233 0 : u[2] = _mm_add_epi32(u[2], K32One);
2234 0 : u[3] = _mm_add_epi32(u[3], K32One);
2235 0 : u[4] = _mm_add_epi32(u[4], K32One);
2236 0 : u[5] = _mm_add_epi32(u[5], K32One);
2237 0 : u[6] = _mm_add_epi32(u[6], K32One);
2238 0 : u[7] = _mm_add_epi32(u[7], K32One);
2239 :
2240 0 : u[0] = _mm_srai_epi32(u[0], 2);
2241 0 : u[1] = _mm_srai_epi32(u[1], 2);
2242 0 : u[2] = _mm_srai_epi32(u[2], 2);
2243 0 : u[3] = _mm_srai_epi32(u[3], 2);
2244 0 : u[4] = _mm_srai_epi32(u[4], 2);
2245 0 : u[5] = _mm_srai_epi32(u[5], 2);
2246 0 : u[6] = _mm_srai_epi32(u[6], 2);
2247 0 : u[7] = _mm_srai_epi32(u[7], 2);
2248 :
2249 0 : out[4] = _mm_packs_epi32(u[0], u[1]);
2250 0 : out[20] = _mm_packs_epi32(u[2], u[3]);
2251 0 : out[12] = _mm_packs_epi32(u[4], u[5]);
2252 0 : out[28] = _mm_packs_epi32(u[6], u[7]);
2253 : #if DCT_HIGH_BIT_DEPTH
2254 : overflow =
2255 : check_epi16_overflow_x4(&out[4], &out[20], &out[12], &out[28]);
2256 : if (overflow) {
2257 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2258 : return;
2259 : }
2260 : #endif // DCT_HIGH_BIT_DEPTH
2261 : }
2262 : {
2263 0 : lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]);
2264 0 : lstep3[17] = _mm_add_epi32(lstep2[19], lstep1[17]);
2265 0 : lstep3[18] = _mm_sub_epi32(lstep1[16], lstep2[18]);
2266 0 : lstep3[19] = _mm_sub_epi32(lstep1[17], lstep2[19]);
2267 0 : lstep3[20] = _mm_sub_epi32(lstep1[22], lstep2[20]);
2268 0 : lstep3[21] = _mm_sub_epi32(lstep1[23], lstep2[21]);
2269 0 : lstep3[22] = _mm_add_epi32(lstep2[20], lstep1[22]);
2270 0 : lstep3[23] = _mm_add_epi32(lstep2[21], lstep1[23]);
2271 0 : lstep3[24] = _mm_add_epi32(lstep2[26], lstep1[24]);
2272 0 : lstep3[25] = _mm_add_epi32(lstep2[27], lstep1[25]);
2273 0 : lstep3[26] = _mm_sub_epi32(lstep1[24], lstep2[26]);
2274 0 : lstep3[27] = _mm_sub_epi32(lstep1[25], lstep2[27]);
2275 0 : lstep3[28] = _mm_sub_epi32(lstep1[30], lstep2[28]);
2276 0 : lstep3[29] = _mm_sub_epi32(lstep1[31], lstep2[29]);
2277 0 : lstep3[30] = _mm_add_epi32(lstep2[28], lstep1[30]);
2278 0 : lstep3[31] = _mm_add_epi32(lstep2[29], lstep1[31]);
2279 : }
2280 : {
2281 0 : const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
2282 0 : const __m128i k32_m28_m04 = pair_set_epi32(-cospi_28_64, -cospi_4_64);
2283 0 : const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
2284 0 : const __m128i k32_m12_m20 =
2285 0 : pair_set_epi32(-cospi_12_64, -cospi_20_64);
2286 0 : const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
2287 0 : const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
2288 :
2289 0 : u[0] = _mm_unpacklo_epi32(lstep2[34], lstep2[60]);
2290 0 : u[1] = _mm_unpackhi_epi32(lstep2[34], lstep2[60]);
2291 0 : u[2] = _mm_unpacklo_epi32(lstep2[35], lstep2[61]);
2292 0 : u[3] = _mm_unpackhi_epi32(lstep2[35], lstep2[61]);
2293 0 : u[4] = _mm_unpacklo_epi32(lstep2[36], lstep2[58]);
2294 0 : u[5] = _mm_unpackhi_epi32(lstep2[36], lstep2[58]);
2295 0 : u[6] = _mm_unpacklo_epi32(lstep2[37], lstep2[59]);
2296 0 : u[7] = _mm_unpackhi_epi32(lstep2[37], lstep2[59]);
2297 0 : u[8] = _mm_unpacklo_epi32(lstep2[42], lstep2[52]);
2298 0 : u[9] = _mm_unpackhi_epi32(lstep2[42], lstep2[52]);
2299 0 : u[10] = _mm_unpacklo_epi32(lstep2[43], lstep2[53]);
2300 0 : u[11] = _mm_unpackhi_epi32(lstep2[43], lstep2[53]);
2301 0 : u[12] = _mm_unpacklo_epi32(lstep2[44], lstep2[50]);
2302 0 : u[13] = _mm_unpackhi_epi32(lstep2[44], lstep2[50]);
2303 0 : u[14] = _mm_unpacklo_epi32(lstep2[45], lstep2[51]);
2304 0 : u[15] = _mm_unpackhi_epi32(lstep2[45], lstep2[51]);
2305 :
2306 0 : v[0] = k_madd_epi32(u[0], k32_m04_p28);
2307 0 : v[1] = k_madd_epi32(u[1], k32_m04_p28);
2308 0 : v[2] = k_madd_epi32(u[2], k32_m04_p28);
2309 0 : v[3] = k_madd_epi32(u[3], k32_m04_p28);
2310 0 : v[4] = k_madd_epi32(u[4], k32_m28_m04);
2311 0 : v[5] = k_madd_epi32(u[5], k32_m28_m04);
2312 0 : v[6] = k_madd_epi32(u[6], k32_m28_m04);
2313 0 : v[7] = k_madd_epi32(u[7], k32_m28_m04);
2314 0 : v[8] = k_madd_epi32(u[8], k32_m20_p12);
2315 0 : v[9] = k_madd_epi32(u[9], k32_m20_p12);
2316 0 : v[10] = k_madd_epi32(u[10], k32_m20_p12);
2317 0 : v[11] = k_madd_epi32(u[11], k32_m20_p12);
2318 0 : v[12] = k_madd_epi32(u[12], k32_m12_m20);
2319 0 : v[13] = k_madd_epi32(u[13], k32_m12_m20);
2320 0 : v[14] = k_madd_epi32(u[14], k32_m12_m20);
2321 0 : v[15] = k_madd_epi32(u[15], k32_m12_m20);
2322 0 : v[16] = k_madd_epi32(u[12], k32_m20_p12);
2323 0 : v[17] = k_madd_epi32(u[13], k32_m20_p12);
2324 0 : v[18] = k_madd_epi32(u[14], k32_m20_p12);
2325 0 : v[19] = k_madd_epi32(u[15], k32_m20_p12);
2326 0 : v[20] = k_madd_epi32(u[8], k32_p12_p20);
2327 0 : v[21] = k_madd_epi32(u[9], k32_p12_p20);
2328 0 : v[22] = k_madd_epi32(u[10], k32_p12_p20);
2329 0 : v[23] = k_madd_epi32(u[11], k32_p12_p20);
2330 0 : v[24] = k_madd_epi32(u[4], k32_m04_p28);
2331 0 : v[25] = k_madd_epi32(u[5], k32_m04_p28);
2332 0 : v[26] = k_madd_epi32(u[6], k32_m04_p28);
2333 0 : v[27] = k_madd_epi32(u[7], k32_m04_p28);
2334 0 : v[28] = k_madd_epi32(u[0], k32_p28_p04);
2335 0 : v[29] = k_madd_epi32(u[1], k32_p28_p04);
2336 0 : v[30] = k_madd_epi32(u[2], k32_p28_p04);
2337 0 : v[31] = k_madd_epi32(u[3], k32_p28_p04);
2338 :
2339 : #if DCT_HIGH_BIT_DEPTH
2340 : overflow = k_check_epi32_overflow_32(
2341 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
2342 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
2343 : &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
2344 : &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
2345 : if (overflow) {
2346 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2347 : return;
2348 : }
2349 : #endif // DCT_HIGH_BIT_DEPTH
2350 0 : u[0] = k_packs_epi64(v[0], v[1]);
2351 0 : u[1] = k_packs_epi64(v[2], v[3]);
2352 0 : u[2] = k_packs_epi64(v[4], v[5]);
2353 0 : u[3] = k_packs_epi64(v[6], v[7]);
2354 0 : u[4] = k_packs_epi64(v[8], v[9]);
2355 0 : u[5] = k_packs_epi64(v[10], v[11]);
2356 0 : u[6] = k_packs_epi64(v[12], v[13]);
2357 0 : u[7] = k_packs_epi64(v[14], v[15]);
2358 0 : u[8] = k_packs_epi64(v[16], v[17]);
2359 0 : u[9] = k_packs_epi64(v[18], v[19]);
2360 0 : u[10] = k_packs_epi64(v[20], v[21]);
2361 0 : u[11] = k_packs_epi64(v[22], v[23]);
2362 0 : u[12] = k_packs_epi64(v[24], v[25]);
2363 0 : u[13] = k_packs_epi64(v[26], v[27]);
2364 0 : u[14] = k_packs_epi64(v[28], v[29]);
2365 0 : u[15] = k_packs_epi64(v[30], v[31]);
2366 :
2367 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2368 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2369 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2370 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2371 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2372 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2373 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2374 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2375 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
2376 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
2377 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2378 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2379 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2380 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2381 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2382 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2383 :
2384 0 : lstep3[34] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
2385 0 : lstep3[35] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
2386 0 : lstep3[36] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
2387 0 : lstep3[37] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
2388 0 : lstep3[42] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
2389 0 : lstep3[43] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
2390 0 : lstep3[44] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
2391 0 : lstep3[45] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
2392 0 : lstep3[50] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
2393 0 : lstep3[51] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
2394 0 : lstep3[52] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
2395 0 : lstep3[53] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
2396 0 : lstep3[58] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
2397 0 : lstep3[59] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
2398 0 : lstep3[60] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
2399 0 : lstep3[61] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
2400 : }
2401 : // stage 7
2402 : {
2403 0 : const __m128i k32_p30_p02 = pair_set_epi32(cospi_30_64, cospi_2_64);
2404 0 : const __m128i k32_p14_p18 = pair_set_epi32(cospi_14_64, cospi_18_64);
2405 0 : const __m128i k32_p22_p10 = pair_set_epi32(cospi_22_64, cospi_10_64);
2406 0 : const __m128i k32_p06_p26 = pair_set_epi32(cospi_6_64, cospi_26_64);
2407 0 : const __m128i k32_m26_p06 = pair_set_epi32(-cospi_26_64, cospi_6_64);
2408 0 : const __m128i k32_m10_p22 = pair_set_epi32(-cospi_10_64, cospi_22_64);
2409 0 : const __m128i k32_m18_p14 = pair_set_epi32(-cospi_18_64, cospi_14_64);
2410 0 : const __m128i k32_m02_p30 = pair_set_epi32(-cospi_2_64, cospi_30_64);
2411 :
2412 0 : u[0] = _mm_unpacklo_epi32(lstep3[16], lstep3[30]);
2413 0 : u[1] = _mm_unpackhi_epi32(lstep3[16], lstep3[30]);
2414 0 : u[2] = _mm_unpacklo_epi32(lstep3[17], lstep3[31]);
2415 0 : u[3] = _mm_unpackhi_epi32(lstep3[17], lstep3[31]);
2416 0 : u[4] = _mm_unpacklo_epi32(lstep3[18], lstep3[28]);
2417 0 : u[5] = _mm_unpackhi_epi32(lstep3[18], lstep3[28]);
2418 0 : u[6] = _mm_unpacklo_epi32(lstep3[19], lstep3[29]);
2419 0 : u[7] = _mm_unpackhi_epi32(lstep3[19], lstep3[29]);
2420 0 : u[8] = _mm_unpacklo_epi32(lstep3[20], lstep3[26]);
2421 0 : u[9] = _mm_unpackhi_epi32(lstep3[20], lstep3[26]);
2422 0 : u[10] = _mm_unpacklo_epi32(lstep3[21], lstep3[27]);
2423 0 : u[11] = _mm_unpackhi_epi32(lstep3[21], lstep3[27]);
2424 0 : u[12] = _mm_unpacklo_epi32(lstep3[22], lstep3[24]);
2425 0 : u[13] = _mm_unpackhi_epi32(lstep3[22], lstep3[24]);
2426 0 : u[14] = _mm_unpacklo_epi32(lstep3[23], lstep3[25]);
2427 0 : u[15] = _mm_unpackhi_epi32(lstep3[23], lstep3[25]);
2428 :
2429 0 : v[0] = k_madd_epi32(u[0], k32_p30_p02);
2430 0 : v[1] = k_madd_epi32(u[1], k32_p30_p02);
2431 0 : v[2] = k_madd_epi32(u[2], k32_p30_p02);
2432 0 : v[3] = k_madd_epi32(u[3], k32_p30_p02);
2433 0 : v[4] = k_madd_epi32(u[4], k32_p14_p18);
2434 0 : v[5] = k_madd_epi32(u[5], k32_p14_p18);
2435 0 : v[6] = k_madd_epi32(u[6], k32_p14_p18);
2436 0 : v[7] = k_madd_epi32(u[7], k32_p14_p18);
2437 0 : v[8] = k_madd_epi32(u[8], k32_p22_p10);
2438 0 : v[9] = k_madd_epi32(u[9], k32_p22_p10);
2439 0 : v[10] = k_madd_epi32(u[10], k32_p22_p10);
2440 0 : v[11] = k_madd_epi32(u[11], k32_p22_p10);
2441 0 : v[12] = k_madd_epi32(u[12], k32_p06_p26);
2442 0 : v[13] = k_madd_epi32(u[13], k32_p06_p26);
2443 0 : v[14] = k_madd_epi32(u[14], k32_p06_p26);
2444 0 : v[15] = k_madd_epi32(u[15], k32_p06_p26);
2445 0 : v[16] = k_madd_epi32(u[12], k32_m26_p06);
2446 0 : v[17] = k_madd_epi32(u[13], k32_m26_p06);
2447 0 : v[18] = k_madd_epi32(u[14], k32_m26_p06);
2448 0 : v[19] = k_madd_epi32(u[15], k32_m26_p06);
2449 0 : v[20] = k_madd_epi32(u[8], k32_m10_p22);
2450 0 : v[21] = k_madd_epi32(u[9], k32_m10_p22);
2451 0 : v[22] = k_madd_epi32(u[10], k32_m10_p22);
2452 0 : v[23] = k_madd_epi32(u[11], k32_m10_p22);
2453 0 : v[24] = k_madd_epi32(u[4], k32_m18_p14);
2454 0 : v[25] = k_madd_epi32(u[5], k32_m18_p14);
2455 0 : v[26] = k_madd_epi32(u[6], k32_m18_p14);
2456 0 : v[27] = k_madd_epi32(u[7], k32_m18_p14);
2457 0 : v[28] = k_madd_epi32(u[0], k32_m02_p30);
2458 0 : v[29] = k_madd_epi32(u[1], k32_m02_p30);
2459 0 : v[30] = k_madd_epi32(u[2], k32_m02_p30);
2460 0 : v[31] = k_madd_epi32(u[3], k32_m02_p30);
2461 :
2462 : #if DCT_HIGH_BIT_DEPTH
2463 : overflow = k_check_epi32_overflow_32(
2464 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
2465 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
2466 : &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
2467 : &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
2468 : if (overflow) {
2469 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2470 : return;
2471 : }
2472 : #endif // DCT_HIGH_BIT_DEPTH
2473 0 : u[0] = k_packs_epi64(v[0], v[1]);
2474 0 : u[1] = k_packs_epi64(v[2], v[3]);
2475 0 : u[2] = k_packs_epi64(v[4], v[5]);
2476 0 : u[3] = k_packs_epi64(v[6], v[7]);
2477 0 : u[4] = k_packs_epi64(v[8], v[9]);
2478 0 : u[5] = k_packs_epi64(v[10], v[11]);
2479 0 : u[6] = k_packs_epi64(v[12], v[13]);
2480 0 : u[7] = k_packs_epi64(v[14], v[15]);
2481 0 : u[8] = k_packs_epi64(v[16], v[17]);
2482 0 : u[9] = k_packs_epi64(v[18], v[19]);
2483 0 : u[10] = k_packs_epi64(v[20], v[21]);
2484 0 : u[11] = k_packs_epi64(v[22], v[23]);
2485 0 : u[12] = k_packs_epi64(v[24], v[25]);
2486 0 : u[13] = k_packs_epi64(v[26], v[27]);
2487 0 : u[14] = k_packs_epi64(v[28], v[29]);
2488 0 : u[15] = k_packs_epi64(v[30], v[31]);
2489 :
2490 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2491 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2492 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2493 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2494 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2495 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2496 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2497 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2498 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
2499 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
2500 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2501 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2502 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2503 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2504 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2505 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2506 :
2507 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
2508 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
2509 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
2510 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
2511 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
2512 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
2513 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
2514 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
2515 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
2516 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
2517 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
2518 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
2519 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
2520 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
2521 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
2522 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
2523 :
2524 0 : v[0] = _mm_cmplt_epi32(u[0], kZero);
2525 0 : v[1] = _mm_cmplt_epi32(u[1], kZero);
2526 0 : v[2] = _mm_cmplt_epi32(u[2], kZero);
2527 0 : v[3] = _mm_cmplt_epi32(u[3], kZero);
2528 0 : v[4] = _mm_cmplt_epi32(u[4], kZero);
2529 0 : v[5] = _mm_cmplt_epi32(u[5], kZero);
2530 0 : v[6] = _mm_cmplt_epi32(u[6], kZero);
2531 0 : v[7] = _mm_cmplt_epi32(u[7], kZero);
2532 0 : v[8] = _mm_cmplt_epi32(u[8], kZero);
2533 0 : v[9] = _mm_cmplt_epi32(u[9], kZero);
2534 0 : v[10] = _mm_cmplt_epi32(u[10], kZero);
2535 0 : v[11] = _mm_cmplt_epi32(u[11], kZero);
2536 0 : v[12] = _mm_cmplt_epi32(u[12], kZero);
2537 0 : v[13] = _mm_cmplt_epi32(u[13], kZero);
2538 0 : v[14] = _mm_cmplt_epi32(u[14], kZero);
2539 0 : v[15] = _mm_cmplt_epi32(u[15], kZero);
2540 :
2541 0 : u[0] = _mm_sub_epi32(u[0], v[0]);
2542 0 : u[1] = _mm_sub_epi32(u[1], v[1]);
2543 0 : u[2] = _mm_sub_epi32(u[2], v[2]);
2544 0 : u[3] = _mm_sub_epi32(u[3], v[3]);
2545 0 : u[4] = _mm_sub_epi32(u[4], v[4]);
2546 0 : u[5] = _mm_sub_epi32(u[5], v[5]);
2547 0 : u[6] = _mm_sub_epi32(u[6], v[6]);
2548 0 : u[7] = _mm_sub_epi32(u[7], v[7]);
2549 0 : u[8] = _mm_sub_epi32(u[8], v[8]);
2550 0 : u[9] = _mm_sub_epi32(u[9], v[9]);
2551 0 : u[10] = _mm_sub_epi32(u[10], v[10]);
2552 0 : u[11] = _mm_sub_epi32(u[11], v[11]);
2553 0 : u[12] = _mm_sub_epi32(u[12], v[12]);
2554 0 : u[13] = _mm_sub_epi32(u[13], v[13]);
2555 0 : u[14] = _mm_sub_epi32(u[14], v[14]);
2556 0 : u[15] = _mm_sub_epi32(u[15], v[15]);
2557 :
2558 0 : v[0] = _mm_add_epi32(u[0], K32One);
2559 0 : v[1] = _mm_add_epi32(u[1], K32One);
2560 0 : v[2] = _mm_add_epi32(u[2], K32One);
2561 0 : v[3] = _mm_add_epi32(u[3], K32One);
2562 0 : v[4] = _mm_add_epi32(u[4], K32One);
2563 0 : v[5] = _mm_add_epi32(u[5], K32One);
2564 0 : v[6] = _mm_add_epi32(u[6], K32One);
2565 0 : v[7] = _mm_add_epi32(u[7], K32One);
2566 0 : v[8] = _mm_add_epi32(u[8], K32One);
2567 0 : v[9] = _mm_add_epi32(u[9], K32One);
2568 0 : v[10] = _mm_add_epi32(u[10], K32One);
2569 0 : v[11] = _mm_add_epi32(u[11], K32One);
2570 0 : v[12] = _mm_add_epi32(u[12], K32One);
2571 0 : v[13] = _mm_add_epi32(u[13], K32One);
2572 0 : v[14] = _mm_add_epi32(u[14], K32One);
2573 0 : v[15] = _mm_add_epi32(u[15], K32One);
2574 :
2575 0 : u[0] = _mm_srai_epi32(v[0], 2);
2576 0 : u[1] = _mm_srai_epi32(v[1], 2);
2577 0 : u[2] = _mm_srai_epi32(v[2], 2);
2578 0 : u[3] = _mm_srai_epi32(v[3], 2);
2579 0 : u[4] = _mm_srai_epi32(v[4], 2);
2580 0 : u[5] = _mm_srai_epi32(v[5], 2);
2581 0 : u[6] = _mm_srai_epi32(v[6], 2);
2582 0 : u[7] = _mm_srai_epi32(v[7], 2);
2583 0 : u[8] = _mm_srai_epi32(v[8], 2);
2584 0 : u[9] = _mm_srai_epi32(v[9], 2);
2585 0 : u[10] = _mm_srai_epi32(v[10], 2);
2586 0 : u[11] = _mm_srai_epi32(v[11], 2);
2587 0 : u[12] = _mm_srai_epi32(v[12], 2);
2588 0 : u[13] = _mm_srai_epi32(v[13], 2);
2589 0 : u[14] = _mm_srai_epi32(v[14], 2);
2590 0 : u[15] = _mm_srai_epi32(v[15], 2);
2591 :
2592 0 : out[2] = _mm_packs_epi32(u[0], u[1]);
2593 0 : out[18] = _mm_packs_epi32(u[2], u[3]);
2594 0 : out[10] = _mm_packs_epi32(u[4], u[5]);
2595 0 : out[26] = _mm_packs_epi32(u[6], u[7]);
2596 0 : out[6] = _mm_packs_epi32(u[8], u[9]);
2597 0 : out[22] = _mm_packs_epi32(u[10], u[11]);
2598 0 : out[14] = _mm_packs_epi32(u[12], u[13]);
2599 0 : out[30] = _mm_packs_epi32(u[14], u[15]);
2600 : #if DCT_HIGH_BIT_DEPTH
2601 : overflow =
2602 : check_epi16_overflow_x8(&out[2], &out[18], &out[10], &out[26],
2603 : &out[6], &out[22], &out[14], &out[30]);
2604 : if (overflow) {
2605 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2606 : return;
2607 : }
2608 : #endif // DCT_HIGH_BIT_DEPTH
2609 : }
2610 : {
2611 0 : lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]);
2612 0 : lstep1[33] = _mm_add_epi32(lstep3[35], lstep2[33]);
2613 0 : lstep1[34] = _mm_sub_epi32(lstep2[32], lstep3[34]);
2614 0 : lstep1[35] = _mm_sub_epi32(lstep2[33], lstep3[35]);
2615 0 : lstep1[36] = _mm_sub_epi32(lstep2[38], lstep3[36]);
2616 0 : lstep1[37] = _mm_sub_epi32(lstep2[39], lstep3[37]);
2617 0 : lstep1[38] = _mm_add_epi32(lstep3[36], lstep2[38]);
2618 0 : lstep1[39] = _mm_add_epi32(lstep3[37], lstep2[39]);
2619 0 : lstep1[40] = _mm_add_epi32(lstep3[42], lstep2[40]);
2620 0 : lstep1[41] = _mm_add_epi32(lstep3[43], lstep2[41]);
2621 0 : lstep1[42] = _mm_sub_epi32(lstep2[40], lstep3[42]);
2622 0 : lstep1[43] = _mm_sub_epi32(lstep2[41], lstep3[43]);
2623 0 : lstep1[44] = _mm_sub_epi32(lstep2[46], lstep3[44]);
2624 0 : lstep1[45] = _mm_sub_epi32(lstep2[47], lstep3[45]);
2625 0 : lstep1[46] = _mm_add_epi32(lstep3[44], lstep2[46]);
2626 0 : lstep1[47] = _mm_add_epi32(lstep3[45], lstep2[47]);
2627 0 : lstep1[48] = _mm_add_epi32(lstep3[50], lstep2[48]);
2628 0 : lstep1[49] = _mm_add_epi32(lstep3[51], lstep2[49]);
2629 0 : lstep1[50] = _mm_sub_epi32(lstep2[48], lstep3[50]);
2630 0 : lstep1[51] = _mm_sub_epi32(lstep2[49], lstep3[51]);
2631 0 : lstep1[52] = _mm_sub_epi32(lstep2[54], lstep3[52]);
2632 0 : lstep1[53] = _mm_sub_epi32(lstep2[55], lstep3[53]);
2633 0 : lstep1[54] = _mm_add_epi32(lstep3[52], lstep2[54]);
2634 0 : lstep1[55] = _mm_add_epi32(lstep3[53], lstep2[55]);
2635 0 : lstep1[56] = _mm_add_epi32(lstep3[58], lstep2[56]);
2636 0 : lstep1[57] = _mm_add_epi32(lstep3[59], lstep2[57]);
2637 0 : lstep1[58] = _mm_sub_epi32(lstep2[56], lstep3[58]);
2638 0 : lstep1[59] = _mm_sub_epi32(lstep2[57], lstep3[59]);
2639 0 : lstep1[60] = _mm_sub_epi32(lstep2[62], lstep3[60]);
2640 0 : lstep1[61] = _mm_sub_epi32(lstep2[63], lstep3[61]);
2641 0 : lstep1[62] = _mm_add_epi32(lstep3[60], lstep2[62]);
2642 0 : lstep1[63] = _mm_add_epi32(lstep3[61], lstep2[63]);
2643 : }
2644 : // stage 8
2645 : {
2646 0 : const __m128i k32_p31_p01 = pair_set_epi32(cospi_31_64, cospi_1_64);
2647 0 : const __m128i k32_p15_p17 = pair_set_epi32(cospi_15_64, cospi_17_64);
2648 0 : const __m128i k32_p23_p09 = pair_set_epi32(cospi_23_64, cospi_9_64);
2649 0 : const __m128i k32_p07_p25 = pair_set_epi32(cospi_7_64, cospi_25_64);
2650 0 : const __m128i k32_m25_p07 = pair_set_epi32(-cospi_25_64, cospi_7_64);
2651 0 : const __m128i k32_m09_p23 = pair_set_epi32(-cospi_9_64, cospi_23_64);
2652 0 : const __m128i k32_m17_p15 = pair_set_epi32(-cospi_17_64, cospi_15_64);
2653 0 : const __m128i k32_m01_p31 = pair_set_epi32(-cospi_1_64, cospi_31_64);
2654 :
2655 0 : u[0] = _mm_unpacklo_epi32(lstep1[32], lstep1[62]);
2656 0 : u[1] = _mm_unpackhi_epi32(lstep1[32], lstep1[62]);
2657 0 : u[2] = _mm_unpacklo_epi32(lstep1[33], lstep1[63]);
2658 0 : u[3] = _mm_unpackhi_epi32(lstep1[33], lstep1[63]);
2659 0 : u[4] = _mm_unpacklo_epi32(lstep1[34], lstep1[60]);
2660 0 : u[5] = _mm_unpackhi_epi32(lstep1[34], lstep1[60]);
2661 0 : u[6] = _mm_unpacklo_epi32(lstep1[35], lstep1[61]);
2662 0 : u[7] = _mm_unpackhi_epi32(lstep1[35], lstep1[61]);
2663 0 : u[8] = _mm_unpacklo_epi32(lstep1[36], lstep1[58]);
2664 0 : u[9] = _mm_unpackhi_epi32(lstep1[36], lstep1[58]);
2665 0 : u[10] = _mm_unpacklo_epi32(lstep1[37], lstep1[59]);
2666 0 : u[11] = _mm_unpackhi_epi32(lstep1[37], lstep1[59]);
2667 0 : u[12] = _mm_unpacklo_epi32(lstep1[38], lstep1[56]);
2668 0 : u[13] = _mm_unpackhi_epi32(lstep1[38], lstep1[56]);
2669 0 : u[14] = _mm_unpacklo_epi32(lstep1[39], lstep1[57]);
2670 0 : u[15] = _mm_unpackhi_epi32(lstep1[39], lstep1[57]);
2671 :
2672 0 : v[0] = k_madd_epi32(u[0], k32_p31_p01);
2673 0 : v[1] = k_madd_epi32(u[1], k32_p31_p01);
2674 0 : v[2] = k_madd_epi32(u[2], k32_p31_p01);
2675 0 : v[3] = k_madd_epi32(u[3], k32_p31_p01);
2676 0 : v[4] = k_madd_epi32(u[4], k32_p15_p17);
2677 0 : v[5] = k_madd_epi32(u[5], k32_p15_p17);
2678 0 : v[6] = k_madd_epi32(u[6], k32_p15_p17);
2679 0 : v[7] = k_madd_epi32(u[7], k32_p15_p17);
2680 0 : v[8] = k_madd_epi32(u[8], k32_p23_p09);
2681 0 : v[9] = k_madd_epi32(u[9], k32_p23_p09);
2682 0 : v[10] = k_madd_epi32(u[10], k32_p23_p09);
2683 0 : v[11] = k_madd_epi32(u[11], k32_p23_p09);
2684 0 : v[12] = k_madd_epi32(u[12], k32_p07_p25);
2685 0 : v[13] = k_madd_epi32(u[13], k32_p07_p25);
2686 0 : v[14] = k_madd_epi32(u[14], k32_p07_p25);
2687 0 : v[15] = k_madd_epi32(u[15], k32_p07_p25);
2688 0 : v[16] = k_madd_epi32(u[12], k32_m25_p07);
2689 0 : v[17] = k_madd_epi32(u[13], k32_m25_p07);
2690 0 : v[18] = k_madd_epi32(u[14], k32_m25_p07);
2691 0 : v[19] = k_madd_epi32(u[15], k32_m25_p07);
2692 0 : v[20] = k_madd_epi32(u[8], k32_m09_p23);
2693 0 : v[21] = k_madd_epi32(u[9], k32_m09_p23);
2694 0 : v[22] = k_madd_epi32(u[10], k32_m09_p23);
2695 0 : v[23] = k_madd_epi32(u[11], k32_m09_p23);
2696 0 : v[24] = k_madd_epi32(u[4], k32_m17_p15);
2697 0 : v[25] = k_madd_epi32(u[5], k32_m17_p15);
2698 0 : v[26] = k_madd_epi32(u[6], k32_m17_p15);
2699 0 : v[27] = k_madd_epi32(u[7], k32_m17_p15);
2700 0 : v[28] = k_madd_epi32(u[0], k32_m01_p31);
2701 0 : v[29] = k_madd_epi32(u[1], k32_m01_p31);
2702 0 : v[30] = k_madd_epi32(u[2], k32_m01_p31);
2703 0 : v[31] = k_madd_epi32(u[3], k32_m01_p31);
2704 :
2705 : #if DCT_HIGH_BIT_DEPTH
2706 : overflow = k_check_epi32_overflow_32(
2707 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
2708 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
2709 : &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
2710 : &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
2711 : if (overflow) {
2712 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2713 : return;
2714 : }
2715 : #endif // DCT_HIGH_BIT_DEPTH
2716 0 : u[0] = k_packs_epi64(v[0], v[1]);
2717 0 : u[1] = k_packs_epi64(v[2], v[3]);
2718 0 : u[2] = k_packs_epi64(v[4], v[5]);
2719 0 : u[3] = k_packs_epi64(v[6], v[7]);
2720 0 : u[4] = k_packs_epi64(v[8], v[9]);
2721 0 : u[5] = k_packs_epi64(v[10], v[11]);
2722 0 : u[6] = k_packs_epi64(v[12], v[13]);
2723 0 : u[7] = k_packs_epi64(v[14], v[15]);
2724 0 : u[8] = k_packs_epi64(v[16], v[17]);
2725 0 : u[9] = k_packs_epi64(v[18], v[19]);
2726 0 : u[10] = k_packs_epi64(v[20], v[21]);
2727 0 : u[11] = k_packs_epi64(v[22], v[23]);
2728 0 : u[12] = k_packs_epi64(v[24], v[25]);
2729 0 : u[13] = k_packs_epi64(v[26], v[27]);
2730 0 : u[14] = k_packs_epi64(v[28], v[29]);
2731 0 : u[15] = k_packs_epi64(v[30], v[31]);
2732 :
2733 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2734 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2735 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2736 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2737 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2738 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2739 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2740 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2741 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
2742 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
2743 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2744 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2745 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2746 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2747 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2748 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2749 :
2750 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
2751 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
2752 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
2753 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
2754 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
2755 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
2756 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
2757 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
2758 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
2759 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
2760 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
2761 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
2762 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
2763 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
2764 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
2765 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
2766 :
2767 0 : v[0] = _mm_cmplt_epi32(u[0], kZero);
2768 0 : v[1] = _mm_cmplt_epi32(u[1], kZero);
2769 0 : v[2] = _mm_cmplt_epi32(u[2], kZero);
2770 0 : v[3] = _mm_cmplt_epi32(u[3], kZero);
2771 0 : v[4] = _mm_cmplt_epi32(u[4], kZero);
2772 0 : v[5] = _mm_cmplt_epi32(u[5], kZero);
2773 0 : v[6] = _mm_cmplt_epi32(u[6], kZero);
2774 0 : v[7] = _mm_cmplt_epi32(u[7], kZero);
2775 0 : v[8] = _mm_cmplt_epi32(u[8], kZero);
2776 0 : v[9] = _mm_cmplt_epi32(u[9], kZero);
2777 0 : v[10] = _mm_cmplt_epi32(u[10], kZero);
2778 0 : v[11] = _mm_cmplt_epi32(u[11], kZero);
2779 0 : v[12] = _mm_cmplt_epi32(u[12], kZero);
2780 0 : v[13] = _mm_cmplt_epi32(u[13], kZero);
2781 0 : v[14] = _mm_cmplt_epi32(u[14], kZero);
2782 0 : v[15] = _mm_cmplt_epi32(u[15], kZero);
2783 :
2784 0 : u[0] = _mm_sub_epi32(u[0], v[0]);
2785 0 : u[1] = _mm_sub_epi32(u[1], v[1]);
2786 0 : u[2] = _mm_sub_epi32(u[2], v[2]);
2787 0 : u[3] = _mm_sub_epi32(u[3], v[3]);
2788 0 : u[4] = _mm_sub_epi32(u[4], v[4]);
2789 0 : u[5] = _mm_sub_epi32(u[5], v[5]);
2790 0 : u[6] = _mm_sub_epi32(u[6], v[6]);
2791 0 : u[7] = _mm_sub_epi32(u[7], v[7]);
2792 0 : u[8] = _mm_sub_epi32(u[8], v[8]);
2793 0 : u[9] = _mm_sub_epi32(u[9], v[9]);
2794 0 : u[10] = _mm_sub_epi32(u[10], v[10]);
2795 0 : u[11] = _mm_sub_epi32(u[11], v[11]);
2796 0 : u[12] = _mm_sub_epi32(u[12], v[12]);
2797 0 : u[13] = _mm_sub_epi32(u[13], v[13]);
2798 0 : u[14] = _mm_sub_epi32(u[14], v[14]);
2799 0 : u[15] = _mm_sub_epi32(u[15], v[15]);
2800 :
2801 0 : v[0] = _mm_add_epi32(u[0], K32One);
2802 0 : v[1] = _mm_add_epi32(u[1], K32One);
2803 0 : v[2] = _mm_add_epi32(u[2], K32One);
2804 0 : v[3] = _mm_add_epi32(u[3], K32One);
2805 0 : v[4] = _mm_add_epi32(u[4], K32One);
2806 0 : v[5] = _mm_add_epi32(u[5], K32One);
2807 0 : v[6] = _mm_add_epi32(u[6], K32One);
2808 0 : v[7] = _mm_add_epi32(u[7], K32One);
2809 0 : v[8] = _mm_add_epi32(u[8], K32One);
2810 0 : v[9] = _mm_add_epi32(u[9], K32One);
2811 0 : v[10] = _mm_add_epi32(u[10], K32One);
2812 0 : v[11] = _mm_add_epi32(u[11], K32One);
2813 0 : v[12] = _mm_add_epi32(u[12], K32One);
2814 0 : v[13] = _mm_add_epi32(u[13], K32One);
2815 0 : v[14] = _mm_add_epi32(u[14], K32One);
2816 0 : v[15] = _mm_add_epi32(u[15], K32One);
2817 :
2818 0 : u[0] = _mm_srai_epi32(v[0], 2);
2819 0 : u[1] = _mm_srai_epi32(v[1], 2);
2820 0 : u[2] = _mm_srai_epi32(v[2], 2);
2821 0 : u[3] = _mm_srai_epi32(v[3], 2);
2822 0 : u[4] = _mm_srai_epi32(v[4], 2);
2823 0 : u[5] = _mm_srai_epi32(v[5], 2);
2824 0 : u[6] = _mm_srai_epi32(v[6], 2);
2825 0 : u[7] = _mm_srai_epi32(v[7], 2);
2826 0 : u[8] = _mm_srai_epi32(v[8], 2);
2827 0 : u[9] = _mm_srai_epi32(v[9], 2);
2828 0 : u[10] = _mm_srai_epi32(v[10], 2);
2829 0 : u[11] = _mm_srai_epi32(v[11], 2);
2830 0 : u[12] = _mm_srai_epi32(v[12], 2);
2831 0 : u[13] = _mm_srai_epi32(v[13], 2);
2832 0 : u[14] = _mm_srai_epi32(v[14], 2);
2833 0 : u[15] = _mm_srai_epi32(v[15], 2);
2834 :
2835 0 : out[1] = _mm_packs_epi32(u[0], u[1]);
2836 0 : out[17] = _mm_packs_epi32(u[2], u[3]);
2837 0 : out[9] = _mm_packs_epi32(u[4], u[5]);
2838 0 : out[25] = _mm_packs_epi32(u[6], u[7]);
2839 0 : out[7] = _mm_packs_epi32(u[8], u[9]);
2840 0 : out[23] = _mm_packs_epi32(u[10], u[11]);
2841 0 : out[15] = _mm_packs_epi32(u[12], u[13]);
2842 0 : out[31] = _mm_packs_epi32(u[14], u[15]);
2843 : #if DCT_HIGH_BIT_DEPTH
2844 : overflow =
2845 : check_epi16_overflow_x8(&out[1], &out[17], &out[9], &out[25],
2846 : &out[7], &out[23], &out[15], &out[31]);
2847 : if (overflow) {
2848 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2849 : return;
2850 : }
2851 : #endif // DCT_HIGH_BIT_DEPTH
2852 : }
2853 : {
2854 0 : const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64);
2855 0 : const __m128i k32_p11_p21 = pair_set_epi32(cospi_11_64, cospi_21_64);
2856 0 : const __m128i k32_p19_p13 = pair_set_epi32(cospi_19_64, cospi_13_64);
2857 0 : const __m128i k32_p03_p29 = pair_set_epi32(cospi_3_64, cospi_29_64);
2858 0 : const __m128i k32_m29_p03 = pair_set_epi32(-cospi_29_64, cospi_3_64);
2859 0 : const __m128i k32_m13_p19 = pair_set_epi32(-cospi_13_64, cospi_19_64);
2860 0 : const __m128i k32_m21_p11 = pair_set_epi32(-cospi_21_64, cospi_11_64);
2861 0 : const __m128i k32_m05_p27 = pair_set_epi32(-cospi_5_64, cospi_27_64);
2862 :
2863 0 : u[0] = _mm_unpacklo_epi32(lstep1[40], lstep1[54]);
2864 0 : u[1] = _mm_unpackhi_epi32(lstep1[40], lstep1[54]);
2865 0 : u[2] = _mm_unpacklo_epi32(lstep1[41], lstep1[55]);
2866 0 : u[3] = _mm_unpackhi_epi32(lstep1[41], lstep1[55]);
2867 0 : u[4] = _mm_unpacklo_epi32(lstep1[42], lstep1[52]);
2868 0 : u[5] = _mm_unpackhi_epi32(lstep1[42], lstep1[52]);
2869 0 : u[6] = _mm_unpacklo_epi32(lstep1[43], lstep1[53]);
2870 0 : u[7] = _mm_unpackhi_epi32(lstep1[43], lstep1[53]);
2871 0 : u[8] = _mm_unpacklo_epi32(lstep1[44], lstep1[50]);
2872 0 : u[9] = _mm_unpackhi_epi32(lstep1[44], lstep1[50]);
2873 0 : u[10] = _mm_unpacklo_epi32(lstep1[45], lstep1[51]);
2874 0 : u[11] = _mm_unpackhi_epi32(lstep1[45], lstep1[51]);
2875 0 : u[12] = _mm_unpacklo_epi32(lstep1[46], lstep1[48]);
2876 0 : u[13] = _mm_unpackhi_epi32(lstep1[46], lstep1[48]);
2877 0 : u[14] = _mm_unpacklo_epi32(lstep1[47], lstep1[49]);
2878 0 : u[15] = _mm_unpackhi_epi32(lstep1[47], lstep1[49]);
2879 :
2880 0 : v[0] = k_madd_epi32(u[0], k32_p27_p05);
2881 0 : v[1] = k_madd_epi32(u[1], k32_p27_p05);
2882 0 : v[2] = k_madd_epi32(u[2], k32_p27_p05);
2883 0 : v[3] = k_madd_epi32(u[3], k32_p27_p05);
2884 0 : v[4] = k_madd_epi32(u[4], k32_p11_p21);
2885 0 : v[5] = k_madd_epi32(u[5], k32_p11_p21);
2886 0 : v[6] = k_madd_epi32(u[6], k32_p11_p21);
2887 0 : v[7] = k_madd_epi32(u[7], k32_p11_p21);
2888 0 : v[8] = k_madd_epi32(u[8], k32_p19_p13);
2889 0 : v[9] = k_madd_epi32(u[9], k32_p19_p13);
2890 0 : v[10] = k_madd_epi32(u[10], k32_p19_p13);
2891 0 : v[11] = k_madd_epi32(u[11], k32_p19_p13);
2892 0 : v[12] = k_madd_epi32(u[12], k32_p03_p29);
2893 0 : v[13] = k_madd_epi32(u[13], k32_p03_p29);
2894 0 : v[14] = k_madd_epi32(u[14], k32_p03_p29);
2895 0 : v[15] = k_madd_epi32(u[15], k32_p03_p29);
2896 0 : v[16] = k_madd_epi32(u[12], k32_m29_p03);
2897 0 : v[17] = k_madd_epi32(u[13], k32_m29_p03);
2898 0 : v[18] = k_madd_epi32(u[14], k32_m29_p03);
2899 0 : v[19] = k_madd_epi32(u[15], k32_m29_p03);
2900 0 : v[20] = k_madd_epi32(u[8], k32_m13_p19);
2901 0 : v[21] = k_madd_epi32(u[9], k32_m13_p19);
2902 0 : v[22] = k_madd_epi32(u[10], k32_m13_p19);
2903 0 : v[23] = k_madd_epi32(u[11], k32_m13_p19);
2904 0 : v[24] = k_madd_epi32(u[4], k32_m21_p11);
2905 0 : v[25] = k_madd_epi32(u[5], k32_m21_p11);
2906 0 : v[26] = k_madd_epi32(u[6], k32_m21_p11);
2907 0 : v[27] = k_madd_epi32(u[7], k32_m21_p11);
2908 0 : v[28] = k_madd_epi32(u[0], k32_m05_p27);
2909 0 : v[29] = k_madd_epi32(u[1], k32_m05_p27);
2910 0 : v[30] = k_madd_epi32(u[2], k32_m05_p27);
2911 0 : v[31] = k_madd_epi32(u[3], k32_m05_p27);
2912 :
2913 : #if DCT_HIGH_BIT_DEPTH
2914 : overflow = k_check_epi32_overflow_32(
2915 : &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
2916 : &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
2917 : &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
2918 : &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
2919 : if (overflow) {
2920 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
2921 : return;
2922 : }
2923 : #endif // DCT_HIGH_BIT_DEPTH
2924 0 : u[0] = k_packs_epi64(v[0], v[1]);
2925 0 : u[1] = k_packs_epi64(v[2], v[3]);
2926 0 : u[2] = k_packs_epi64(v[4], v[5]);
2927 0 : u[3] = k_packs_epi64(v[6], v[7]);
2928 0 : u[4] = k_packs_epi64(v[8], v[9]);
2929 0 : u[5] = k_packs_epi64(v[10], v[11]);
2930 0 : u[6] = k_packs_epi64(v[12], v[13]);
2931 0 : u[7] = k_packs_epi64(v[14], v[15]);
2932 0 : u[8] = k_packs_epi64(v[16], v[17]);
2933 0 : u[9] = k_packs_epi64(v[18], v[19]);
2934 0 : u[10] = k_packs_epi64(v[20], v[21]);
2935 0 : u[11] = k_packs_epi64(v[22], v[23]);
2936 0 : u[12] = k_packs_epi64(v[24], v[25]);
2937 0 : u[13] = k_packs_epi64(v[26], v[27]);
2938 0 : u[14] = k_packs_epi64(v[28], v[29]);
2939 0 : u[15] = k_packs_epi64(v[30], v[31]);
2940 :
2941 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2942 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2943 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2944 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2945 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2946 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2947 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2948 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2949 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
2950 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
2951 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2952 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2953 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2954 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2955 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2956 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2957 :
2958 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
2959 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
2960 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
2961 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
2962 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
2963 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
2964 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
2965 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
2966 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
2967 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
2968 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
2969 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
2970 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
2971 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
2972 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
2973 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
2974 :
2975 0 : v[0] = _mm_cmplt_epi32(u[0], kZero);
2976 0 : v[1] = _mm_cmplt_epi32(u[1], kZero);
2977 0 : v[2] = _mm_cmplt_epi32(u[2], kZero);
2978 0 : v[3] = _mm_cmplt_epi32(u[3], kZero);
2979 0 : v[4] = _mm_cmplt_epi32(u[4], kZero);
2980 0 : v[5] = _mm_cmplt_epi32(u[5], kZero);
2981 0 : v[6] = _mm_cmplt_epi32(u[6], kZero);
2982 0 : v[7] = _mm_cmplt_epi32(u[7], kZero);
2983 0 : v[8] = _mm_cmplt_epi32(u[8], kZero);
2984 0 : v[9] = _mm_cmplt_epi32(u[9], kZero);
2985 0 : v[10] = _mm_cmplt_epi32(u[10], kZero);
2986 0 : v[11] = _mm_cmplt_epi32(u[11], kZero);
2987 0 : v[12] = _mm_cmplt_epi32(u[12], kZero);
2988 0 : v[13] = _mm_cmplt_epi32(u[13], kZero);
2989 0 : v[14] = _mm_cmplt_epi32(u[14], kZero);
2990 0 : v[15] = _mm_cmplt_epi32(u[15], kZero);
2991 :
2992 0 : u[0] = _mm_sub_epi32(u[0], v[0]);
2993 0 : u[1] = _mm_sub_epi32(u[1], v[1]);
2994 0 : u[2] = _mm_sub_epi32(u[2], v[2]);
2995 0 : u[3] = _mm_sub_epi32(u[3], v[3]);
2996 0 : u[4] = _mm_sub_epi32(u[4], v[4]);
2997 0 : u[5] = _mm_sub_epi32(u[5], v[5]);
2998 0 : u[6] = _mm_sub_epi32(u[6], v[6]);
2999 0 : u[7] = _mm_sub_epi32(u[7], v[7]);
3000 0 : u[8] = _mm_sub_epi32(u[8], v[8]);
3001 0 : u[9] = _mm_sub_epi32(u[9], v[9]);
3002 0 : u[10] = _mm_sub_epi32(u[10], v[10]);
3003 0 : u[11] = _mm_sub_epi32(u[11], v[11]);
3004 0 : u[12] = _mm_sub_epi32(u[12], v[12]);
3005 0 : u[13] = _mm_sub_epi32(u[13], v[13]);
3006 0 : u[14] = _mm_sub_epi32(u[14], v[14]);
3007 0 : u[15] = _mm_sub_epi32(u[15], v[15]);
3008 :
3009 0 : v[0] = _mm_add_epi32(u[0], K32One);
3010 0 : v[1] = _mm_add_epi32(u[1], K32One);
3011 0 : v[2] = _mm_add_epi32(u[2], K32One);
3012 0 : v[3] = _mm_add_epi32(u[3], K32One);
3013 0 : v[4] = _mm_add_epi32(u[4], K32One);
3014 0 : v[5] = _mm_add_epi32(u[5], K32One);
3015 0 : v[6] = _mm_add_epi32(u[6], K32One);
3016 0 : v[7] = _mm_add_epi32(u[7], K32One);
3017 0 : v[8] = _mm_add_epi32(u[8], K32One);
3018 0 : v[9] = _mm_add_epi32(u[9], K32One);
3019 0 : v[10] = _mm_add_epi32(u[10], K32One);
3020 0 : v[11] = _mm_add_epi32(u[11], K32One);
3021 0 : v[12] = _mm_add_epi32(u[12], K32One);
3022 0 : v[13] = _mm_add_epi32(u[13], K32One);
3023 0 : v[14] = _mm_add_epi32(u[14], K32One);
3024 0 : v[15] = _mm_add_epi32(u[15], K32One);
3025 :
3026 0 : u[0] = _mm_srai_epi32(v[0], 2);
3027 0 : u[1] = _mm_srai_epi32(v[1], 2);
3028 0 : u[2] = _mm_srai_epi32(v[2], 2);
3029 0 : u[3] = _mm_srai_epi32(v[3], 2);
3030 0 : u[4] = _mm_srai_epi32(v[4], 2);
3031 0 : u[5] = _mm_srai_epi32(v[5], 2);
3032 0 : u[6] = _mm_srai_epi32(v[6], 2);
3033 0 : u[7] = _mm_srai_epi32(v[7], 2);
3034 0 : u[8] = _mm_srai_epi32(v[8], 2);
3035 0 : u[9] = _mm_srai_epi32(v[9], 2);
3036 0 : u[10] = _mm_srai_epi32(v[10], 2);
3037 0 : u[11] = _mm_srai_epi32(v[11], 2);
3038 0 : u[12] = _mm_srai_epi32(v[12], 2);
3039 0 : u[13] = _mm_srai_epi32(v[13], 2);
3040 0 : u[14] = _mm_srai_epi32(v[14], 2);
3041 0 : u[15] = _mm_srai_epi32(v[15], 2);
3042 :
3043 0 : out[5] = _mm_packs_epi32(u[0], u[1]);
3044 0 : out[21] = _mm_packs_epi32(u[2], u[3]);
3045 0 : out[13] = _mm_packs_epi32(u[4], u[5]);
3046 0 : out[29] = _mm_packs_epi32(u[6], u[7]);
3047 0 : out[3] = _mm_packs_epi32(u[8], u[9]);
3048 0 : out[19] = _mm_packs_epi32(u[10], u[11]);
3049 0 : out[11] = _mm_packs_epi32(u[12], u[13]);
3050 0 : out[27] = _mm_packs_epi32(u[14], u[15]);
3051 : #if DCT_HIGH_BIT_DEPTH
3052 : overflow =
3053 : check_epi16_overflow_x8(&out[5], &out[21], &out[13], &out[29],
3054 : &out[3], &out[19], &out[11], &out[27]);
3055 : if (overflow) {
3056 : HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
3057 : return;
3058 : }
3059 : #endif // DCT_HIGH_BIT_DEPTH
3060 : }
3061 : }
3062 : #endif // FDCT32x32_HIGH_PRECISION
3063 : // Transpose the results, do it as four 8x8 transposes.
3064 : {
3065 : int transpose_block;
3066 0 : int16_t *output0 = &intermediate[column_start * 32];
3067 0 : tran_low_t *output1 = &output_org[column_start * 32];
3068 0 : for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
3069 0 : __m128i *this_out = &out[8 * transpose_block];
3070 : // 00 01 02 03 04 05 06 07
3071 : // 10 11 12 13 14 15 16 17
3072 : // 20 21 22 23 24 25 26 27
3073 : // 30 31 32 33 34 35 36 37
3074 : // 40 41 42 43 44 45 46 47
3075 : // 50 51 52 53 54 55 56 57
3076 : // 60 61 62 63 64 65 66 67
3077 : // 70 71 72 73 74 75 76 77
3078 0 : const __m128i tr0_0 = _mm_unpacklo_epi16(this_out[0], this_out[1]);
3079 0 : const __m128i tr0_1 = _mm_unpacklo_epi16(this_out[2], this_out[3]);
3080 0 : const __m128i tr0_2 = _mm_unpackhi_epi16(this_out[0], this_out[1]);
3081 0 : const __m128i tr0_3 = _mm_unpackhi_epi16(this_out[2], this_out[3]);
3082 0 : const __m128i tr0_4 = _mm_unpacklo_epi16(this_out[4], this_out[5]);
3083 0 : const __m128i tr0_5 = _mm_unpacklo_epi16(this_out[6], this_out[7]);
3084 0 : const __m128i tr0_6 = _mm_unpackhi_epi16(this_out[4], this_out[5]);
3085 0 : const __m128i tr0_7 = _mm_unpackhi_epi16(this_out[6], this_out[7]);
3086 : // 00 10 01 11 02 12 03 13
3087 : // 20 30 21 31 22 32 23 33
3088 : // 04 14 05 15 06 16 07 17
3089 : // 24 34 25 35 26 36 27 37
3090 : // 40 50 41 51 42 52 43 53
3091 : // 60 70 61 71 62 72 63 73
3092 : // 54 54 55 55 56 56 57 57
3093 : // 64 74 65 75 66 76 67 77
3094 0 : const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
3095 0 : const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
3096 0 : const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
3097 0 : const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
3098 0 : const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
3099 0 : const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
3100 0 : const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
3101 0 : const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
3102 : // 00 10 20 30 01 11 21 31
3103 : // 40 50 60 70 41 51 61 71
3104 : // 02 12 22 32 03 13 23 33
3105 : // 42 52 62 72 43 53 63 73
3106 : // 04 14 24 34 05 15 21 36
3107 : // 44 54 64 74 45 55 61 76
3108 : // 06 16 26 36 07 17 27 37
3109 : // 46 56 66 76 47 57 67 77
3110 0 : __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
3111 0 : __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
3112 0 : __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
3113 0 : __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
3114 0 : __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
3115 0 : __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
3116 0 : __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
3117 0 : __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
3118 : // 00 10 20 30 40 50 60 70
3119 : // 01 11 21 31 41 51 61 71
3120 : // 02 12 22 32 42 52 62 72
3121 : // 03 13 23 33 43 53 63 73
3122 : // 04 14 24 34 44 54 64 74
3123 : // 05 15 25 35 45 55 65 75
3124 : // 06 16 26 36 46 56 66 76
3125 : // 07 17 27 37 47 57 67 77
3126 0 : if (0 == pass) {
3127 : // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
3128 : // TODO(cd): see quality impact of only doing
3129 : // output[j] = (output[j] + 1) >> 2;
3130 : // which would remove the code between here ...
3131 0 : __m128i tr2_0_0 = _mm_cmpgt_epi16(tr2_0, kZero);
3132 0 : __m128i tr2_1_0 = _mm_cmpgt_epi16(tr2_1, kZero);
3133 0 : __m128i tr2_2_0 = _mm_cmpgt_epi16(tr2_2, kZero);
3134 0 : __m128i tr2_3_0 = _mm_cmpgt_epi16(tr2_3, kZero);
3135 0 : __m128i tr2_4_0 = _mm_cmpgt_epi16(tr2_4, kZero);
3136 0 : __m128i tr2_5_0 = _mm_cmpgt_epi16(tr2_5, kZero);
3137 0 : __m128i tr2_6_0 = _mm_cmpgt_epi16(tr2_6, kZero);
3138 0 : __m128i tr2_7_0 = _mm_cmpgt_epi16(tr2_7, kZero);
3139 0 : tr2_0 = _mm_sub_epi16(tr2_0, tr2_0_0);
3140 0 : tr2_1 = _mm_sub_epi16(tr2_1, tr2_1_0);
3141 0 : tr2_2 = _mm_sub_epi16(tr2_2, tr2_2_0);
3142 0 : tr2_3 = _mm_sub_epi16(tr2_3, tr2_3_0);
3143 0 : tr2_4 = _mm_sub_epi16(tr2_4, tr2_4_0);
3144 0 : tr2_5 = _mm_sub_epi16(tr2_5, tr2_5_0);
3145 0 : tr2_6 = _mm_sub_epi16(tr2_6, tr2_6_0);
3146 0 : tr2_7 = _mm_sub_epi16(tr2_7, tr2_7_0);
3147 : // ... and here.
3148 : // PS: also change code in vp9/encoder/vp9_dct.c
3149 0 : tr2_0 = _mm_add_epi16(tr2_0, kOne);
3150 0 : tr2_1 = _mm_add_epi16(tr2_1, kOne);
3151 0 : tr2_2 = _mm_add_epi16(tr2_2, kOne);
3152 0 : tr2_3 = _mm_add_epi16(tr2_3, kOne);
3153 0 : tr2_4 = _mm_add_epi16(tr2_4, kOne);
3154 0 : tr2_5 = _mm_add_epi16(tr2_5, kOne);
3155 0 : tr2_6 = _mm_add_epi16(tr2_6, kOne);
3156 0 : tr2_7 = _mm_add_epi16(tr2_7, kOne);
3157 0 : tr2_0 = _mm_srai_epi16(tr2_0, 2);
3158 0 : tr2_1 = _mm_srai_epi16(tr2_1, 2);
3159 0 : tr2_2 = _mm_srai_epi16(tr2_2, 2);
3160 0 : tr2_3 = _mm_srai_epi16(tr2_3, 2);
3161 0 : tr2_4 = _mm_srai_epi16(tr2_4, 2);
3162 0 : tr2_5 = _mm_srai_epi16(tr2_5, 2);
3163 0 : tr2_6 = _mm_srai_epi16(tr2_6, 2);
3164 0 : tr2_7 = _mm_srai_epi16(tr2_7, 2);
3165 : }
3166 : // Note: even though all these stores are aligned, using the aligned
3167 : // intrinsic make the code slightly slower.
3168 0 : if (pass == 0) {
3169 0 : _mm_storeu_si128((__m128i *)(output0 + 0 * 32), tr2_0);
3170 0 : _mm_storeu_si128((__m128i *)(output0 + 1 * 32), tr2_1);
3171 0 : _mm_storeu_si128((__m128i *)(output0 + 2 * 32), tr2_2);
3172 0 : _mm_storeu_si128((__m128i *)(output0 + 3 * 32), tr2_3);
3173 0 : _mm_storeu_si128((__m128i *)(output0 + 4 * 32), tr2_4);
3174 0 : _mm_storeu_si128((__m128i *)(output0 + 5 * 32), tr2_5);
3175 0 : _mm_storeu_si128((__m128i *)(output0 + 6 * 32), tr2_6);
3176 0 : _mm_storeu_si128((__m128i *)(output0 + 7 * 32), tr2_7);
3177 : // Process next 8x8
3178 0 : output0 += 8;
3179 : } else {
3180 0 : storeu_output(&tr2_0, (output1 + 0 * 32));
3181 0 : storeu_output(&tr2_1, (output1 + 1 * 32));
3182 0 : storeu_output(&tr2_2, (output1 + 2 * 32));
3183 0 : storeu_output(&tr2_3, (output1 + 3 * 32));
3184 0 : storeu_output(&tr2_4, (output1 + 4 * 32));
3185 0 : storeu_output(&tr2_5, (output1 + 5 * 32));
3186 0 : storeu_output(&tr2_6, (output1 + 6 * 32));
3187 0 : storeu_output(&tr2_7, (output1 + 7 * 32));
3188 : // Process next 8x8
3189 0 : output1 += 8;
3190 : }
3191 : }
3192 : }
3193 : }
3194 : }
3195 0 : } // NOLINT
3196 :
3197 : #undef ADD_EPI16
3198 : #undef SUB_EPI16
3199 : #undef HIGH_FDCT32x32_2D_C
3200 : #undef HIGH_FDCT32x32_2D_ROWS_C
|
