Line data Source code
1 : /*
2 : * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 : *
4 : * This source code is subject to the terms of the BSD 2 Clause License and
5 : * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 : * was not distributed with this source code in the LICENSE file, you can
7 : * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 : * Media Patent License 1.0 was not distributed with this source code in the
9 : * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 : */
11 :
12 : #include "./aom_dsp_rtcd.h"
13 : #include "aom_dsp/x86/inv_txfm_sse2.h"
14 : #include "aom_dsp/x86/txfm_common_sse2.h"
15 :
16 : #define RECON_AND_STORE4X4(dest, in_x) \
17 : { \
18 : __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \
19 : d0 = _mm_unpacklo_epi8(d0, zero); \
20 : d0 = _mm_add_epi16(in_x, d0); \
21 : d0 = _mm_packus_epi16(d0, d0); \
22 : *(int *)(dest) = _mm_cvtsi128_si32(d0); \
23 : }
24 :
25 0 : void aom_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest,
26 : int stride) {
27 0 : const __m128i zero = _mm_setzero_si128();
28 0 : const __m128i eight = _mm_set1_epi16(8);
29 0 : const __m128i cst = _mm_setr_epi16(
30 0 : (int16_t)cospi_16_64, (int16_t)cospi_16_64, (int16_t)cospi_16_64,
31 0 : (int16_t)-cospi_16_64, (int16_t)cospi_24_64, (int16_t)-cospi_8_64,
32 0 : (int16_t)cospi_8_64, (int16_t)cospi_24_64);
33 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
34 : __m128i input0, input1, input2, input3;
35 :
36 : // Rows
37 0 : input0 = load_input_data(input);
38 0 : input2 = load_input_data(input + 8);
39 :
40 : // Construct i3, i1, i3, i1, i2, i0, i2, i0
41 0 : input0 = _mm_shufflelo_epi16(input0, 0xd8);
42 0 : input0 = _mm_shufflehi_epi16(input0, 0xd8);
43 0 : input2 = _mm_shufflelo_epi16(input2, 0xd8);
44 0 : input2 = _mm_shufflehi_epi16(input2, 0xd8);
45 :
46 0 : input1 = _mm_unpackhi_epi32(input0, input0);
47 0 : input0 = _mm_unpacklo_epi32(input0, input0);
48 0 : input3 = _mm_unpackhi_epi32(input2, input2);
49 0 : input2 = _mm_unpacklo_epi32(input2, input2);
50 :
51 : // Stage 1
52 0 : input0 = _mm_madd_epi16(input0, cst);
53 0 : input1 = _mm_madd_epi16(input1, cst);
54 0 : input2 = _mm_madd_epi16(input2, cst);
55 0 : input3 = _mm_madd_epi16(input3, cst);
56 :
57 0 : input0 = _mm_add_epi32(input0, rounding);
58 0 : input1 = _mm_add_epi32(input1, rounding);
59 0 : input2 = _mm_add_epi32(input2, rounding);
60 0 : input3 = _mm_add_epi32(input3, rounding);
61 :
62 0 : input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
63 0 : input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
64 0 : input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
65 0 : input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
66 :
67 : // Stage 2
68 0 : input0 = _mm_packs_epi32(input0, input1);
69 0 : input1 = _mm_packs_epi32(input2, input3);
70 :
71 : // Transpose
72 0 : input2 = _mm_unpacklo_epi16(input0, input1);
73 0 : input3 = _mm_unpackhi_epi16(input0, input1);
74 0 : input0 = _mm_unpacklo_epi32(input2, input3);
75 0 : input1 = _mm_unpackhi_epi32(input2, input3);
76 :
77 : // Switch column2, column 3, and then, we got:
78 : // input2: column1, column 0; input3: column2, column 3.
79 0 : input1 = _mm_shuffle_epi32(input1, 0x4e);
80 0 : input2 = _mm_add_epi16(input0, input1);
81 0 : input3 = _mm_sub_epi16(input0, input1);
82 :
83 : // Columns
84 : // Construct i3, i1, i3, i1, i2, i0, i2, i0
85 0 : input0 = _mm_unpacklo_epi32(input2, input2);
86 0 : input1 = _mm_unpackhi_epi32(input2, input2);
87 0 : input2 = _mm_unpackhi_epi32(input3, input3);
88 0 : input3 = _mm_unpacklo_epi32(input3, input3);
89 :
90 : // Stage 1
91 0 : input0 = _mm_madd_epi16(input0, cst);
92 0 : input1 = _mm_madd_epi16(input1, cst);
93 0 : input2 = _mm_madd_epi16(input2, cst);
94 0 : input3 = _mm_madd_epi16(input3, cst);
95 :
96 0 : input0 = _mm_add_epi32(input0, rounding);
97 0 : input1 = _mm_add_epi32(input1, rounding);
98 0 : input2 = _mm_add_epi32(input2, rounding);
99 0 : input3 = _mm_add_epi32(input3, rounding);
100 :
101 0 : input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
102 0 : input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
103 0 : input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
104 0 : input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
105 :
106 : // Stage 2
107 0 : input0 = _mm_packs_epi32(input0, input2);
108 0 : input1 = _mm_packs_epi32(input1, input3);
109 :
110 : // Transpose
111 0 : input2 = _mm_unpacklo_epi16(input0, input1);
112 0 : input3 = _mm_unpackhi_epi16(input0, input1);
113 0 : input0 = _mm_unpacklo_epi32(input2, input3);
114 0 : input1 = _mm_unpackhi_epi32(input2, input3);
115 :
116 : // Switch column2, column 3, and then, we got:
117 : // input2: column1, column 0; input3: column2, column 3.
118 0 : input1 = _mm_shuffle_epi32(input1, 0x4e);
119 0 : input2 = _mm_add_epi16(input0, input1);
120 0 : input3 = _mm_sub_epi16(input0, input1);
121 :
122 : // Final round and shift
123 0 : input2 = _mm_add_epi16(input2, eight);
124 0 : input3 = _mm_add_epi16(input3, eight);
125 :
126 0 : input2 = _mm_srai_epi16(input2, 4);
127 0 : input3 = _mm_srai_epi16(input3, 4);
128 :
129 : // Reconstruction and Store
130 : {
131 0 : __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
132 0 : __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
133 0 : d0 = _mm_unpacklo_epi32(d0,
134 0 : _mm_cvtsi32_si128(*(const int *)(dest + stride)));
135 0 : d2 = _mm_unpacklo_epi32(
136 0 : _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)), d2);
137 0 : d0 = _mm_unpacklo_epi8(d0, zero);
138 0 : d2 = _mm_unpacklo_epi8(d2, zero);
139 0 : d0 = _mm_add_epi16(d0, input2);
140 0 : d2 = _mm_add_epi16(d2, input3);
141 0 : d0 = _mm_packus_epi16(d0, d2);
142 : // store input0
143 0 : *(int *)dest = _mm_cvtsi128_si32(d0);
144 : // store input1
145 0 : d0 = _mm_srli_si128(d0, 4);
146 0 : *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
147 : // store input2
148 0 : d0 = _mm_srli_si128(d0, 4);
149 0 : *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
150 : // store input3
151 0 : d0 = _mm_srli_si128(d0, 4);
152 0 : *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
153 : }
154 0 : }
155 :
156 0 : void aom_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest,
157 : int stride) {
158 : __m128i dc_value;
159 0 : const __m128i zero = _mm_setzero_si128();
160 : int a;
161 :
162 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
163 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
164 0 : a = ROUND_POWER_OF_TWO(a, 4);
165 :
166 0 : if (a == 0) return;
167 :
168 0 : dc_value = _mm_set1_epi16(a);
169 :
170 0 : RECON_AND_STORE4X4(dest + 0 * stride, dc_value);
171 0 : RECON_AND_STORE4X4(dest + 1 * stride, dc_value);
172 0 : RECON_AND_STORE4X4(dest + 2 * stride, dc_value);
173 0 : RECON_AND_STORE4X4(dest + 3 * stride, dc_value);
174 : }
175 :
176 0 : void aom_idct4_sse2(__m128i *in) {
177 0 : const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
178 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
179 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
180 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
181 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
182 : __m128i u[8], v[8];
183 :
184 0 : array_transpose_4x4(in);
185 : // stage 1
186 0 : u[0] = _mm_unpacklo_epi16(in[0], in[1]);
187 0 : u[1] = _mm_unpackhi_epi16(in[0], in[1]);
188 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
189 0 : v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
190 0 : v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
191 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
192 :
193 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
194 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
195 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
196 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
197 :
198 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
199 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
200 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
201 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
202 :
203 0 : u[0] = _mm_packs_epi32(v[0], v[1]);
204 0 : u[1] = _mm_packs_epi32(v[3], v[2]);
205 :
206 : // stage 2
207 0 : in[0] = _mm_add_epi16(u[0], u[1]);
208 0 : in[1] = _mm_sub_epi16(u[0], u[1]);
209 0 : in[1] = _mm_shuffle_epi32(in[1], 0x4E);
210 0 : }
211 :
212 0 : void aom_iadst4_sse2(__m128i *in) {
213 0 : const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
214 0 : const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
215 0 : const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
216 0 : const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
217 0 : const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
218 0 : const __m128i kZero = _mm_set1_epi16(0);
219 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
220 : __m128i u[8], v[8], in7;
221 :
222 0 : array_transpose_4x4(in);
223 0 : in7 = _mm_srli_si128(in[1], 8);
224 0 : in7 = _mm_add_epi16(in7, in[0]);
225 0 : in7 = _mm_sub_epi16(in7, in[1]);
226 :
227 0 : u[0] = _mm_unpacklo_epi16(in[0], in[1]);
228 0 : u[1] = _mm_unpackhi_epi16(in[0], in[1]);
229 0 : u[2] = _mm_unpacklo_epi16(in7, kZero);
230 0 : u[3] = _mm_unpackhi_epi16(in[0], kZero);
231 :
232 0 : v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04); // s0 + s3
233 0 : v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02); // s2 + s5
234 0 : v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x2
235 0 : v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01); // s1 - s4
236 0 : v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04); // s2 - s6
237 0 : v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s2
238 :
239 0 : u[0] = _mm_add_epi32(v[0], v[1]);
240 0 : u[1] = _mm_add_epi32(v[3], v[4]);
241 0 : u[2] = v[2];
242 0 : u[3] = _mm_add_epi32(u[0], u[1]);
243 0 : u[4] = _mm_slli_epi32(v[5], 2);
244 0 : u[5] = _mm_add_epi32(u[3], v[5]);
245 0 : u[6] = _mm_sub_epi32(u[5], u[4]);
246 :
247 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
248 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
249 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
250 0 : v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
251 :
252 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
253 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
254 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
255 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
256 :
257 0 : in[0] = _mm_packs_epi32(u[0], u[1]);
258 0 : in[1] = _mm_packs_epi32(u[2], u[3]);
259 0 : }
260 :
261 : // Define Macro for multiplying elements by constants and adding them together.
262 : #define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, cst0, cst1, cst2, cst3, \
263 : res0, res1, res2, res3) \
264 : { \
265 : tmp0 = _mm_madd_epi16(lo_0, cst0); \
266 : tmp1 = _mm_madd_epi16(hi_0, cst0); \
267 : tmp2 = _mm_madd_epi16(lo_0, cst1); \
268 : tmp3 = _mm_madd_epi16(hi_0, cst1); \
269 : tmp4 = _mm_madd_epi16(lo_1, cst2); \
270 : tmp5 = _mm_madd_epi16(hi_1, cst2); \
271 : tmp6 = _mm_madd_epi16(lo_1, cst3); \
272 : tmp7 = _mm_madd_epi16(hi_1, cst3); \
273 : \
274 : tmp0 = _mm_add_epi32(tmp0, rounding); \
275 : tmp1 = _mm_add_epi32(tmp1, rounding); \
276 : tmp2 = _mm_add_epi32(tmp2, rounding); \
277 : tmp3 = _mm_add_epi32(tmp3, rounding); \
278 : tmp4 = _mm_add_epi32(tmp4, rounding); \
279 : tmp5 = _mm_add_epi32(tmp5, rounding); \
280 : tmp6 = _mm_add_epi32(tmp6, rounding); \
281 : tmp7 = _mm_add_epi32(tmp7, rounding); \
282 : \
283 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
284 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
285 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
286 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
287 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); \
288 : tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); \
289 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); \
290 : tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); \
291 : \
292 : res0 = _mm_packs_epi32(tmp0, tmp1); \
293 : res1 = _mm_packs_epi32(tmp2, tmp3); \
294 : res2 = _mm_packs_epi32(tmp4, tmp5); \
295 : res3 = _mm_packs_epi32(tmp6, tmp7); \
296 : }
297 :
298 : #define MULTIPLICATION_AND_ADD_2(lo_0, hi_0, cst0, cst1, res0, res1) \
299 : { \
300 : tmp0 = _mm_madd_epi16(lo_0, cst0); \
301 : tmp1 = _mm_madd_epi16(hi_0, cst0); \
302 : tmp2 = _mm_madd_epi16(lo_0, cst1); \
303 : tmp3 = _mm_madd_epi16(hi_0, cst1); \
304 : \
305 : tmp0 = _mm_add_epi32(tmp0, rounding); \
306 : tmp1 = _mm_add_epi32(tmp1, rounding); \
307 : tmp2 = _mm_add_epi32(tmp2, rounding); \
308 : tmp3 = _mm_add_epi32(tmp3, rounding); \
309 : \
310 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
311 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
312 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
313 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
314 : \
315 : res0 = _mm_packs_epi32(tmp0, tmp1); \
316 : res1 = _mm_packs_epi32(tmp2, tmp3); \
317 : }
318 :
319 : #define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3, \
320 : out4, out5, out6, out7) \
321 : { \
322 : /* Stage1 */ \
323 : { \
324 : const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \
325 : const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \
326 : const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); \
327 : const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); \
328 : \
329 : MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, stg1_1, \
330 : stg1_2, stg1_3, stp1_4, stp1_7, stp1_5, stp1_6) \
331 : } \
332 : \
333 : /* Stage2 */ \
334 : { \
335 : const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4); \
336 : const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4); \
337 : const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); \
338 : const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); \
339 : \
340 : MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, stg2_1, \
341 : stg2_2, stg2_3, stp2_0, stp2_1, stp2_2, stp2_3) \
342 : \
343 : stp2_4 = _mm_adds_epi16(stp1_4, stp1_5); \
344 : stp2_5 = _mm_subs_epi16(stp1_4, stp1_5); \
345 : stp2_6 = _mm_subs_epi16(stp1_7, stp1_6); \
346 : stp2_7 = _mm_adds_epi16(stp1_7, stp1_6); \
347 : } \
348 : \
349 : /* Stage3 */ \
350 : { \
351 : const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
352 : const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
353 : \
354 : stp1_0 = _mm_adds_epi16(stp2_0, stp2_3); \
355 : stp1_1 = _mm_adds_epi16(stp2_1, stp2_2); \
356 : stp1_2 = _mm_subs_epi16(stp2_1, stp2_2); \
357 : stp1_3 = _mm_subs_epi16(stp2_0, stp2_3); \
358 : \
359 : tmp0 = _mm_madd_epi16(lo_56, stg2_1); \
360 : tmp1 = _mm_madd_epi16(hi_56, stg2_1); \
361 : tmp2 = _mm_madd_epi16(lo_56, stg2_0); \
362 : tmp3 = _mm_madd_epi16(hi_56, stg2_0); \
363 : \
364 : tmp0 = _mm_add_epi32(tmp0, rounding); \
365 : tmp1 = _mm_add_epi32(tmp1, rounding); \
366 : tmp2 = _mm_add_epi32(tmp2, rounding); \
367 : tmp3 = _mm_add_epi32(tmp3, rounding); \
368 : \
369 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
370 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
371 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
372 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
373 : \
374 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
375 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
376 : } \
377 : \
378 : /* Stage4 */ \
379 : out0 = _mm_adds_epi16(stp1_0, stp2_7); \
380 : out1 = _mm_adds_epi16(stp1_1, stp1_6); \
381 : out2 = _mm_adds_epi16(stp1_2, stp1_5); \
382 : out3 = _mm_adds_epi16(stp1_3, stp2_4); \
383 : out4 = _mm_subs_epi16(stp1_3, stp2_4); \
384 : out5 = _mm_subs_epi16(stp1_2, stp1_5); \
385 : out6 = _mm_subs_epi16(stp1_1, stp1_6); \
386 : out7 = _mm_subs_epi16(stp1_0, stp2_7); \
387 : }
388 :
389 0 : void aom_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest,
390 : int stride) {
391 0 : const __m128i zero = _mm_setzero_si128();
392 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
393 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 4);
394 0 : const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
395 0 : const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
396 0 : const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
397 0 : const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
398 0 : const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
399 0 : const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
400 0 : const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
401 0 : const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
402 :
403 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
404 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
405 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
406 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
407 : int i;
408 :
409 : // Load input data.
410 0 : in0 = load_input_data(input);
411 0 : in1 = load_input_data(input + 8 * 1);
412 0 : in2 = load_input_data(input + 8 * 2);
413 0 : in3 = load_input_data(input + 8 * 3);
414 0 : in4 = load_input_data(input + 8 * 4);
415 0 : in5 = load_input_data(input + 8 * 5);
416 0 : in6 = load_input_data(input + 8 * 6);
417 0 : in7 = load_input_data(input + 8 * 7);
418 :
419 : // 2-D
420 0 : for (i = 0; i < 2; i++) {
421 : // 8x8 Transpose is copied from aom_fdct8x8_sse2()
422 0 : TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
423 : in4, in5, in6, in7);
424 :
425 : // 4-stage 1D idct8x8
426 0 : IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4, in5,
427 : in6, in7);
428 : }
429 :
430 : // Final rounding and shift
431 0 : in0 = _mm_adds_epi16(in0, final_rounding);
432 0 : in1 = _mm_adds_epi16(in1, final_rounding);
433 0 : in2 = _mm_adds_epi16(in2, final_rounding);
434 0 : in3 = _mm_adds_epi16(in3, final_rounding);
435 0 : in4 = _mm_adds_epi16(in4, final_rounding);
436 0 : in5 = _mm_adds_epi16(in5, final_rounding);
437 0 : in6 = _mm_adds_epi16(in6, final_rounding);
438 0 : in7 = _mm_adds_epi16(in7, final_rounding);
439 :
440 0 : in0 = _mm_srai_epi16(in0, 5);
441 0 : in1 = _mm_srai_epi16(in1, 5);
442 0 : in2 = _mm_srai_epi16(in2, 5);
443 0 : in3 = _mm_srai_epi16(in3, 5);
444 0 : in4 = _mm_srai_epi16(in4, 5);
445 0 : in5 = _mm_srai_epi16(in5, 5);
446 0 : in6 = _mm_srai_epi16(in6, 5);
447 0 : in7 = _mm_srai_epi16(in7, 5);
448 :
449 0 : RECON_AND_STORE(dest + 0 * stride, in0);
450 0 : RECON_AND_STORE(dest + 1 * stride, in1);
451 0 : RECON_AND_STORE(dest + 2 * stride, in2);
452 0 : RECON_AND_STORE(dest + 3 * stride, in3);
453 0 : RECON_AND_STORE(dest + 4 * stride, in4);
454 0 : RECON_AND_STORE(dest + 5 * stride, in5);
455 0 : RECON_AND_STORE(dest + 6 * stride, in6);
456 0 : RECON_AND_STORE(dest + 7 * stride, in7);
457 0 : }
458 :
459 0 : void aom_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest,
460 : int stride) {
461 : __m128i dc_value;
462 0 : const __m128i zero = _mm_setzero_si128();
463 : int a;
464 :
465 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
466 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
467 0 : a = ROUND_POWER_OF_TWO(a, 5);
468 :
469 0 : if (a == 0) return;
470 :
471 0 : dc_value = _mm_set1_epi16(a);
472 :
473 0 : RECON_AND_STORE(dest + 0 * stride, dc_value);
474 0 : RECON_AND_STORE(dest + 1 * stride, dc_value);
475 0 : RECON_AND_STORE(dest + 2 * stride, dc_value);
476 0 : RECON_AND_STORE(dest + 3 * stride, dc_value);
477 0 : RECON_AND_STORE(dest + 4 * stride, dc_value);
478 0 : RECON_AND_STORE(dest + 5 * stride, dc_value);
479 0 : RECON_AND_STORE(dest + 6 * stride, dc_value);
480 0 : RECON_AND_STORE(dest + 7 * stride, dc_value);
481 : }
482 :
483 0 : void aom_idct8_sse2(__m128i *in) {
484 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
485 0 : const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
486 0 : const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
487 0 : const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
488 0 : const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
489 0 : const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
490 0 : const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
491 0 : const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
492 0 : const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
493 :
494 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
495 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
496 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
497 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
498 :
499 : // 8x8 Transpose is copied from aom_fdct8x8_sse2()
500 0 : TRANSPOSE_8X8(in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7], in0,
501 : in1, in2, in3, in4, in5, in6, in7);
502 :
503 : // 4-stage 1D idct8x8
504 0 : IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in[0], in[1], in[2], in[3],
505 : in[4], in[5], in[6], in[7]);
506 0 : }
507 :
508 0 : void aom_iadst8_sse2(__m128i *in) {
509 0 : const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
510 0 : const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
511 0 : const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
512 0 : const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
513 0 : const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
514 0 : const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
515 0 : const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
516 0 : const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
517 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
518 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
519 0 : const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
520 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
521 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
522 0 : const __m128i k__const_0 = _mm_set1_epi16(0);
523 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
524 :
525 : __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
526 : __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
527 : __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
528 : __m128i s0, s1, s2, s3, s4, s5, s6, s7;
529 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
530 :
531 : // transpose
532 0 : array_transpose_8x8(in, in);
533 :
534 : // properly aligned for butterfly input
535 0 : in0 = in[7];
536 0 : in1 = in[0];
537 0 : in2 = in[5];
538 0 : in3 = in[2];
539 0 : in4 = in[3];
540 0 : in5 = in[4];
541 0 : in6 = in[1];
542 0 : in7 = in[6];
543 :
544 : // column transformation
545 : // stage 1
546 : // interleave and multiply/add into 32-bit integer
547 0 : s0 = _mm_unpacklo_epi16(in0, in1);
548 0 : s1 = _mm_unpackhi_epi16(in0, in1);
549 0 : s2 = _mm_unpacklo_epi16(in2, in3);
550 0 : s3 = _mm_unpackhi_epi16(in2, in3);
551 0 : s4 = _mm_unpacklo_epi16(in4, in5);
552 0 : s5 = _mm_unpackhi_epi16(in4, in5);
553 0 : s6 = _mm_unpacklo_epi16(in6, in7);
554 0 : s7 = _mm_unpackhi_epi16(in6, in7);
555 :
556 0 : u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
557 0 : u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
558 0 : u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
559 0 : u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
560 0 : u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
561 0 : u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
562 0 : u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
563 0 : u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
564 0 : u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
565 0 : u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
566 0 : u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
567 0 : u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
568 0 : u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
569 0 : u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
570 0 : u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
571 0 : u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
572 :
573 : // addition
574 0 : w0 = _mm_add_epi32(u0, u8);
575 0 : w1 = _mm_add_epi32(u1, u9);
576 0 : w2 = _mm_add_epi32(u2, u10);
577 0 : w3 = _mm_add_epi32(u3, u11);
578 0 : w4 = _mm_add_epi32(u4, u12);
579 0 : w5 = _mm_add_epi32(u5, u13);
580 0 : w6 = _mm_add_epi32(u6, u14);
581 0 : w7 = _mm_add_epi32(u7, u15);
582 0 : w8 = _mm_sub_epi32(u0, u8);
583 0 : w9 = _mm_sub_epi32(u1, u9);
584 0 : w10 = _mm_sub_epi32(u2, u10);
585 0 : w11 = _mm_sub_epi32(u3, u11);
586 0 : w12 = _mm_sub_epi32(u4, u12);
587 0 : w13 = _mm_sub_epi32(u5, u13);
588 0 : w14 = _mm_sub_epi32(u6, u14);
589 0 : w15 = _mm_sub_epi32(u7, u15);
590 :
591 : // shift and rounding
592 0 : v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
593 0 : v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
594 0 : v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
595 0 : v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
596 0 : v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
597 0 : v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
598 0 : v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
599 0 : v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
600 0 : v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
601 0 : v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
602 0 : v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
603 0 : v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
604 0 : v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
605 0 : v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
606 0 : v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
607 0 : v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
608 :
609 0 : u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
610 0 : u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
611 0 : u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
612 0 : u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
613 0 : u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
614 0 : u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
615 0 : u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
616 0 : u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
617 0 : u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
618 0 : u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
619 0 : u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
620 0 : u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
621 0 : u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
622 0 : u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
623 0 : u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
624 0 : u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
625 :
626 : // back to 16-bit and pack 8 integers into __m128i
627 0 : in[0] = _mm_packs_epi32(u0, u1);
628 0 : in[1] = _mm_packs_epi32(u2, u3);
629 0 : in[2] = _mm_packs_epi32(u4, u5);
630 0 : in[3] = _mm_packs_epi32(u6, u7);
631 0 : in[4] = _mm_packs_epi32(u8, u9);
632 0 : in[5] = _mm_packs_epi32(u10, u11);
633 0 : in[6] = _mm_packs_epi32(u12, u13);
634 0 : in[7] = _mm_packs_epi32(u14, u15);
635 :
636 : // stage 2
637 0 : s0 = _mm_add_epi16(in[0], in[2]);
638 0 : s1 = _mm_add_epi16(in[1], in[3]);
639 0 : s2 = _mm_sub_epi16(in[0], in[2]);
640 0 : s3 = _mm_sub_epi16(in[1], in[3]);
641 0 : u0 = _mm_unpacklo_epi16(in[4], in[5]);
642 0 : u1 = _mm_unpackhi_epi16(in[4], in[5]);
643 0 : u2 = _mm_unpacklo_epi16(in[6], in[7]);
644 0 : u3 = _mm_unpackhi_epi16(in[6], in[7]);
645 :
646 0 : v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
647 0 : v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
648 0 : v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
649 0 : v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
650 0 : v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
651 0 : v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
652 0 : v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
653 0 : v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
654 :
655 0 : w0 = _mm_add_epi32(v0, v4);
656 0 : w1 = _mm_add_epi32(v1, v5);
657 0 : w2 = _mm_add_epi32(v2, v6);
658 0 : w3 = _mm_add_epi32(v3, v7);
659 0 : w4 = _mm_sub_epi32(v0, v4);
660 0 : w5 = _mm_sub_epi32(v1, v5);
661 0 : w6 = _mm_sub_epi32(v2, v6);
662 0 : w7 = _mm_sub_epi32(v3, v7);
663 :
664 0 : v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
665 0 : v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
666 0 : v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
667 0 : v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
668 0 : v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
669 0 : v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
670 0 : v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
671 0 : v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
672 :
673 0 : u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
674 0 : u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
675 0 : u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
676 0 : u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
677 0 : u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
678 0 : u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
679 0 : u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
680 0 : u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
681 :
682 : // back to 16-bit intergers
683 0 : s4 = _mm_packs_epi32(u0, u1);
684 0 : s5 = _mm_packs_epi32(u2, u3);
685 0 : s6 = _mm_packs_epi32(u4, u5);
686 0 : s7 = _mm_packs_epi32(u6, u7);
687 :
688 : // stage 3
689 0 : u0 = _mm_unpacklo_epi16(s2, s3);
690 0 : u1 = _mm_unpackhi_epi16(s2, s3);
691 0 : u2 = _mm_unpacklo_epi16(s6, s7);
692 0 : u3 = _mm_unpackhi_epi16(s6, s7);
693 :
694 0 : v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
695 0 : v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
696 0 : v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
697 0 : v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
698 0 : v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
699 0 : v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
700 0 : v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
701 0 : v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
702 :
703 0 : u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
704 0 : u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
705 0 : u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
706 0 : u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
707 0 : u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
708 0 : u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
709 0 : u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
710 0 : u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
711 :
712 0 : v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
713 0 : v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
714 0 : v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
715 0 : v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
716 0 : v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
717 0 : v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
718 0 : v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
719 0 : v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
720 :
721 0 : s2 = _mm_packs_epi32(v0, v1);
722 0 : s3 = _mm_packs_epi32(v2, v3);
723 0 : s6 = _mm_packs_epi32(v4, v5);
724 0 : s7 = _mm_packs_epi32(v6, v7);
725 :
726 0 : in[0] = s0;
727 0 : in[1] = _mm_sub_epi16(k__const_0, s4);
728 0 : in[2] = s6;
729 0 : in[3] = _mm_sub_epi16(k__const_0, s2);
730 0 : in[4] = s3;
731 0 : in[5] = _mm_sub_epi16(k__const_0, s7);
732 0 : in[6] = s5;
733 0 : in[7] = _mm_sub_epi16(k__const_0, s1);
734 0 : }
735 :
736 0 : void aom_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest,
737 : int stride) {
738 0 : const __m128i zero = _mm_setzero_si128();
739 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
740 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 4);
741 0 : const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
742 0 : const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
743 0 : const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
744 0 : const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
745 0 : const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
746 0 : const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
747 0 : const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
748 0 : const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
749 0 : const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
750 :
751 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
752 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
753 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
754 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
755 :
756 : // Rows. Load 4-row input data.
757 0 : in0 = load_input_data(input);
758 0 : in1 = load_input_data(input + 8 * 1);
759 0 : in2 = load_input_data(input + 8 * 2);
760 0 : in3 = load_input_data(input + 8 * 3);
761 :
762 : // 8x4 Transpose
763 0 : TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1);
764 : // Stage1
765 : {
766 0 : const __m128i lo_17 = _mm_unpackhi_epi16(in0, zero);
767 0 : const __m128i lo_35 = _mm_unpackhi_epi16(in1, zero);
768 :
769 0 : tmp0 = _mm_madd_epi16(lo_17, stg1_0);
770 0 : tmp2 = _mm_madd_epi16(lo_17, stg1_1);
771 0 : tmp4 = _mm_madd_epi16(lo_35, stg1_2);
772 0 : tmp6 = _mm_madd_epi16(lo_35, stg1_3);
773 :
774 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
775 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
776 0 : tmp4 = _mm_add_epi32(tmp4, rounding);
777 0 : tmp6 = _mm_add_epi32(tmp6, rounding);
778 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
779 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
780 0 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
781 0 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
782 :
783 0 : stp1_4 = _mm_packs_epi32(tmp0, tmp2);
784 0 : stp1_5 = _mm_packs_epi32(tmp4, tmp6);
785 : }
786 :
787 : // Stage2
788 : {
789 0 : const __m128i lo_04 = _mm_unpacklo_epi16(in0, zero);
790 0 : const __m128i lo_26 = _mm_unpacklo_epi16(in1, zero);
791 :
792 0 : tmp0 = _mm_madd_epi16(lo_04, stg2_0);
793 0 : tmp2 = _mm_madd_epi16(lo_04, stg2_1);
794 0 : tmp4 = _mm_madd_epi16(lo_26, stg2_2);
795 0 : tmp6 = _mm_madd_epi16(lo_26, stg2_3);
796 :
797 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
798 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
799 0 : tmp4 = _mm_add_epi32(tmp4, rounding);
800 0 : tmp6 = _mm_add_epi32(tmp6, rounding);
801 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
802 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
803 0 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
804 0 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
805 :
806 0 : stp2_0 = _mm_packs_epi32(tmp0, tmp2);
807 0 : stp2_2 = _mm_packs_epi32(tmp6, tmp4);
808 :
809 0 : tmp0 = _mm_adds_epi16(stp1_4, stp1_5);
810 0 : tmp1 = _mm_subs_epi16(stp1_4, stp1_5);
811 :
812 0 : stp2_4 = tmp0;
813 0 : stp2_5 = _mm_unpacklo_epi64(tmp1, zero);
814 0 : stp2_6 = _mm_unpackhi_epi64(tmp1, zero);
815 : }
816 :
817 : // Stage3
818 : {
819 0 : const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6);
820 :
821 0 : tmp4 = _mm_adds_epi16(stp2_0, stp2_2);
822 0 : tmp6 = _mm_subs_epi16(stp2_0, stp2_2);
823 :
824 0 : stp1_2 = _mm_unpackhi_epi64(tmp6, tmp4);
825 0 : stp1_3 = _mm_unpacklo_epi64(tmp6, tmp4);
826 :
827 0 : tmp0 = _mm_madd_epi16(lo_56, stg3_0);
828 0 : tmp2 = _mm_madd_epi16(lo_56, stg2_0); // stg3_1 = stg2_0
829 :
830 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
831 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
832 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
833 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
834 :
835 0 : stp1_5 = _mm_packs_epi32(tmp0, tmp2);
836 : }
837 :
838 : // Stage4
839 0 : tmp0 = _mm_adds_epi16(stp1_3, stp2_4);
840 0 : tmp1 = _mm_adds_epi16(stp1_2, stp1_5);
841 0 : tmp2 = _mm_subs_epi16(stp1_3, stp2_4);
842 0 : tmp3 = _mm_subs_epi16(stp1_2, stp1_5);
843 :
844 0 : TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3)
845 :
846 0 : IDCT8(in0, in1, in2, in3, zero, zero, zero, zero, in0, in1, in2, in3, in4,
847 : in5, in6, in7);
848 : // Final rounding and shift
849 0 : in0 = _mm_adds_epi16(in0, final_rounding);
850 0 : in1 = _mm_adds_epi16(in1, final_rounding);
851 0 : in2 = _mm_adds_epi16(in2, final_rounding);
852 0 : in3 = _mm_adds_epi16(in3, final_rounding);
853 0 : in4 = _mm_adds_epi16(in4, final_rounding);
854 0 : in5 = _mm_adds_epi16(in5, final_rounding);
855 0 : in6 = _mm_adds_epi16(in6, final_rounding);
856 0 : in7 = _mm_adds_epi16(in7, final_rounding);
857 :
858 0 : in0 = _mm_srai_epi16(in0, 5);
859 0 : in1 = _mm_srai_epi16(in1, 5);
860 0 : in2 = _mm_srai_epi16(in2, 5);
861 0 : in3 = _mm_srai_epi16(in3, 5);
862 0 : in4 = _mm_srai_epi16(in4, 5);
863 0 : in5 = _mm_srai_epi16(in5, 5);
864 0 : in6 = _mm_srai_epi16(in6, 5);
865 0 : in7 = _mm_srai_epi16(in7, 5);
866 :
867 0 : RECON_AND_STORE(dest + 0 * stride, in0);
868 0 : RECON_AND_STORE(dest + 1 * stride, in1);
869 0 : RECON_AND_STORE(dest + 2 * stride, in2);
870 0 : RECON_AND_STORE(dest + 3 * stride, in3);
871 0 : RECON_AND_STORE(dest + 4 * stride, in4);
872 0 : RECON_AND_STORE(dest + 5 * stride, in5);
873 0 : RECON_AND_STORE(dest + 6 * stride, in6);
874 0 : RECON_AND_STORE(dest + 7 * stride, in7);
875 0 : }
876 :
877 : #define IDCT16 \
878 : /* Stage2 */ \
879 : { \
880 : const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], in[15]); \
881 : const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], in[15]); \
882 : const __m128i lo_9_7 = _mm_unpacklo_epi16(in[9], in[7]); \
883 : const __m128i hi_9_7 = _mm_unpackhi_epi16(in[9], in[7]); \
884 : const __m128i lo_5_11 = _mm_unpacklo_epi16(in[5], in[11]); \
885 : const __m128i hi_5_11 = _mm_unpackhi_epi16(in[5], in[11]); \
886 : const __m128i lo_13_3 = _mm_unpacklo_epi16(in[13], in[3]); \
887 : const __m128i hi_13_3 = _mm_unpackhi_epi16(in[13], in[3]); \
888 : \
889 : MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, stg2_0, stg2_1, \
890 : stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, stp2_14) \
891 : \
892 : MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, stg2_4, stg2_5, \
893 : stg2_6, stg2_7, stp2_10, stp2_13, stp2_11, stp2_12) \
894 : } \
895 : \
896 : /* Stage3 */ \
897 : { \
898 : const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], in[14]); \
899 : const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], in[14]); \
900 : const __m128i lo_10_6 = _mm_unpacklo_epi16(in[10], in[6]); \
901 : const __m128i hi_10_6 = _mm_unpackhi_epi16(in[10], in[6]); \
902 : \
903 : MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, stg3_0, stg3_1, \
904 : stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, stp1_6) \
905 : \
906 : stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9); \
907 : stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \
908 : stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
909 : stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
910 : \
911 : stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); \
912 : stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
913 : stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
914 : stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
915 : } \
916 : \
917 : /* Stage4 */ \
918 : { \
919 : const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], in[8]); \
920 : const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], in[8]); \
921 : const __m128i lo_4_12 = _mm_unpacklo_epi16(in[4], in[12]); \
922 : const __m128i hi_4_12 = _mm_unpackhi_epi16(in[4], in[12]); \
923 : \
924 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
925 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
926 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
927 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
928 : \
929 : MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, stg4_0, stg4_1, \
930 : stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3) \
931 : \
932 : stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
933 : stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
934 : stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
935 : stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
936 : \
937 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
938 : stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10, \
939 : stp2_13) \
940 : } \
941 : \
942 : /* Stage5 */ \
943 : { \
944 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
945 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
946 : \
947 : stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
948 : stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
949 : stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
950 : stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
951 : \
952 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
953 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
954 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
955 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
956 : \
957 : tmp0 = _mm_add_epi32(tmp0, rounding); \
958 : tmp1 = _mm_add_epi32(tmp1, rounding); \
959 : tmp2 = _mm_add_epi32(tmp2, rounding); \
960 : tmp3 = _mm_add_epi32(tmp3, rounding); \
961 : \
962 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
963 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
964 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
965 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
966 : \
967 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
968 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
969 : \
970 : stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); \
971 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
972 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
973 : stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
974 : \
975 : stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
976 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
977 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
978 : stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
979 : } \
980 : \
981 : /* Stage6 */ \
982 : { \
983 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
984 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
985 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
986 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
987 : \
988 : stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
989 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
990 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
991 : stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
992 : stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
993 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
994 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
995 : stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
996 : \
997 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
998 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
999 : stp2_12) \
1000 : }
1001 :
1002 : #define IDCT16_10 \
1003 : /* Stage2 */ \
1004 : { \
1005 : const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], zero); \
1006 : const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], zero); \
1007 : const __m128i lo_13_3 = _mm_unpacklo_epi16(zero, in[3]); \
1008 : const __m128i hi_13_3 = _mm_unpackhi_epi16(zero, in[3]); \
1009 : \
1010 : MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_13_3, hi_13_3, stg2_0, stg2_1, \
1011 : stg2_6, stg2_7, stp1_8_0, stp1_15, stp1_11, \
1012 : stp1_12_0) \
1013 : } \
1014 : \
1015 : /* Stage3 */ \
1016 : { \
1017 : const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], zero); \
1018 : const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], zero); \
1019 : \
1020 : MULTIPLICATION_AND_ADD_2(lo_2_14, hi_2_14, stg3_0, stg3_1, stp2_4, stp2_7) \
1021 : \
1022 : stp1_9 = stp1_8_0; \
1023 : stp1_10 = stp1_11; \
1024 : \
1025 : stp1_13 = stp1_12_0; \
1026 : stp1_14 = stp1_15; \
1027 : } \
1028 : \
1029 : /* Stage4 */ \
1030 : { \
1031 : const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero); \
1032 : const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], zero); \
1033 : \
1034 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
1035 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
1036 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
1037 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
1038 : \
1039 : MULTIPLICATION_AND_ADD_2(lo_0_8, hi_0_8, stg4_0, stg4_1, stp1_0, stp1_1) \
1040 : stp2_5 = stp2_4; \
1041 : stp2_6 = stp2_7; \
1042 : \
1043 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
1044 : stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10, \
1045 : stp2_13) \
1046 : } \
1047 : \
1048 : /* Stage5 */ \
1049 : { \
1050 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
1051 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
1052 : \
1053 : stp1_2 = stp1_1; \
1054 : stp1_3 = stp1_0; \
1055 : \
1056 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
1057 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
1058 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
1059 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
1060 : \
1061 : tmp0 = _mm_add_epi32(tmp0, rounding); \
1062 : tmp1 = _mm_add_epi32(tmp1, rounding); \
1063 : tmp2 = _mm_add_epi32(tmp2, rounding); \
1064 : tmp3 = _mm_add_epi32(tmp3, rounding); \
1065 : \
1066 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
1067 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
1068 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
1069 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
1070 : \
1071 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
1072 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
1073 : \
1074 : stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); \
1075 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
1076 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
1077 : stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
1078 : \
1079 : stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
1080 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
1081 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
1082 : stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
1083 : } \
1084 : \
1085 : /* Stage6 */ \
1086 : { \
1087 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
1088 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
1089 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
1090 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
1091 : \
1092 : stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
1093 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
1094 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
1095 : stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
1096 : stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
1097 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
1098 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
1099 : stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
1100 : \
1101 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
1102 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
1103 : stp2_12) \
1104 : }
1105 :
1106 0 : void aom_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
1107 : int stride) {
1108 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
1109 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
1110 0 : const __m128i zero = _mm_setzero_si128();
1111 :
1112 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
1113 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
1114 0 : const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
1115 0 : const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
1116 0 : const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
1117 0 : const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
1118 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
1119 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
1120 :
1121 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
1122 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
1123 0 : const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
1124 0 : const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
1125 :
1126 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
1127 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1128 0 : const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
1129 0 : const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
1130 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
1131 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
1132 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
1133 0 : const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
1134 :
1135 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1136 :
1137 : __m128i in[16], l[16], r[16], *curr1;
1138 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
1139 : stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
1140 : stp1_8_0, stp1_12_0;
1141 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
1142 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15;
1143 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1144 : int i;
1145 :
1146 0 : curr1 = l;
1147 0 : for (i = 0; i < 2; i++) {
1148 : // 1-D idct
1149 :
1150 : // Load input data.
1151 0 : in[0] = load_input_data(input);
1152 0 : in[8] = load_input_data(input + 8 * 1);
1153 0 : in[1] = load_input_data(input + 8 * 2);
1154 0 : in[9] = load_input_data(input + 8 * 3);
1155 0 : in[2] = load_input_data(input + 8 * 4);
1156 0 : in[10] = load_input_data(input + 8 * 5);
1157 0 : in[3] = load_input_data(input + 8 * 6);
1158 0 : in[11] = load_input_data(input + 8 * 7);
1159 0 : in[4] = load_input_data(input + 8 * 8);
1160 0 : in[12] = load_input_data(input + 8 * 9);
1161 0 : in[5] = load_input_data(input + 8 * 10);
1162 0 : in[13] = load_input_data(input + 8 * 11);
1163 0 : in[6] = load_input_data(input + 8 * 12);
1164 0 : in[14] = load_input_data(input + 8 * 13);
1165 0 : in[7] = load_input_data(input + 8 * 14);
1166 0 : in[15] = load_input_data(input + 8 * 15);
1167 :
1168 0 : array_transpose_8x8(in, in);
1169 0 : array_transpose_8x8(in + 8, in + 8);
1170 :
1171 0 : IDCT16
1172 :
1173 : // Stage7
1174 0 : curr1[0] = _mm_add_epi16(stp2_0, stp1_15);
1175 0 : curr1[1] = _mm_add_epi16(stp2_1, stp1_14);
1176 0 : curr1[2] = _mm_add_epi16(stp2_2, stp2_13);
1177 0 : curr1[3] = _mm_add_epi16(stp2_3, stp2_12);
1178 0 : curr1[4] = _mm_add_epi16(stp2_4, stp2_11);
1179 0 : curr1[5] = _mm_add_epi16(stp2_5, stp2_10);
1180 0 : curr1[6] = _mm_add_epi16(stp2_6, stp1_9);
1181 0 : curr1[7] = _mm_add_epi16(stp2_7, stp1_8);
1182 0 : curr1[8] = _mm_sub_epi16(stp2_7, stp1_8);
1183 0 : curr1[9] = _mm_sub_epi16(stp2_6, stp1_9);
1184 0 : curr1[10] = _mm_sub_epi16(stp2_5, stp2_10);
1185 0 : curr1[11] = _mm_sub_epi16(stp2_4, stp2_11);
1186 0 : curr1[12] = _mm_sub_epi16(stp2_3, stp2_12);
1187 0 : curr1[13] = _mm_sub_epi16(stp2_2, stp2_13);
1188 0 : curr1[14] = _mm_sub_epi16(stp2_1, stp1_14);
1189 0 : curr1[15] = _mm_sub_epi16(stp2_0, stp1_15);
1190 :
1191 0 : curr1 = r;
1192 0 : input += 128;
1193 : }
1194 0 : for (i = 0; i < 2; i++) {
1195 : int j;
1196 : // 1-D idct
1197 0 : array_transpose_8x8(l + i * 8, in);
1198 0 : array_transpose_8x8(r + i * 8, in + 8);
1199 :
1200 0 : IDCT16
1201 :
1202 : // 2-D
1203 0 : in[0] = _mm_add_epi16(stp2_0, stp1_15);
1204 0 : in[1] = _mm_add_epi16(stp2_1, stp1_14);
1205 0 : in[2] = _mm_add_epi16(stp2_2, stp2_13);
1206 0 : in[3] = _mm_add_epi16(stp2_3, stp2_12);
1207 0 : in[4] = _mm_add_epi16(stp2_4, stp2_11);
1208 0 : in[5] = _mm_add_epi16(stp2_5, stp2_10);
1209 0 : in[6] = _mm_add_epi16(stp2_6, stp1_9);
1210 0 : in[7] = _mm_add_epi16(stp2_7, stp1_8);
1211 0 : in[8] = _mm_sub_epi16(stp2_7, stp1_8);
1212 0 : in[9] = _mm_sub_epi16(stp2_6, stp1_9);
1213 0 : in[10] = _mm_sub_epi16(stp2_5, stp2_10);
1214 0 : in[11] = _mm_sub_epi16(stp2_4, stp2_11);
1215 0 : in[12] = _mm_sub_epi16(stp2_3, stp2_12);
1216 0 : in[13] = _mm_sub_epi16(stp2_2, stp2_13);
1217 0 : in[14] = _mm_sub_epi16(stp2_1, stp1_14);
1218 0 : in[15] = _mm_sub_epi16(stp2_0, stp1_15);
1219 :
1220 0 : for (j = 0; j < 16; ++j) {
1221 : // Final rounding and shift
1222 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
1223 0 : in[j] = _mm_srai_epi16(in[j], 6);
1224 0 : RECON_AND_STORE(dest + j * stride, in[j]);
1225 : }
1226 :
1227 0 : dest += 8;
1228 : }
1229 0 : }
1230 :
1231 0 : void aom_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest,
1232 : int stride) {
1233 : __m128i dc_value;
1234 0 : const __m128i zero = _mm_setzero_si128();
1235 : int a, i;
1236 :
1237 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
1238 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
1239 0 : a = ROUND_POWER_OF_TWO(a, 6);
1240 :
1241 0 : if (a == 0) return;
1242 :
1243 0 : dc_value = _mm_set1_epi16(a);
1244 :
1245 0 : for (i = 0; i < 16; ++i) {
1246 0 : RECON_AND_STORE(dest + 0, dc_value);
1247 0 : RECON_AND_STORE(dest + 8, dc_value);
1248 0 : dest += stride;
1249 : }
1250 : }
1251 :
1252 0 : void iadst16_8col(__m128i *in) {
1253 : // perform 16x16 1-D ADST for 8 columns
1254 : __m128i s[16], x[16], u[32], v[32];
1255 0 : const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
1256 0 : const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
1257 0 : const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
1258 0 : const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
1259 0 : const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
1260 0 : const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
1261 0 : const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
1262 0 : const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
1263 0 : const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
1264 0 : const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
1265 0 : const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
1266 0 : const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
1267 0 : const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
1268 0 : const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
1269 0 : const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
1270 0 : const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
1271 0 : const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
1272 0 : const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
1273 0 : const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
1274 0 : const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
1275 0 : const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
1276 0 : const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
1277 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
1278 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
1279 0 : const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
1280 0 : const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
1281 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
1282 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1283 0 : const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1284 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
1285 0 : const __m128i kZero = _mm_set1_epi16(0);
1286 :
1287 0 : u[0] = _mm_unpacklo_epi16(in[15], in[0]);
1288 0 : u[1] = _mm_unpackhi_epi16(in[15], in[0]);
1289 0 : u[2] = _mm_unpacklo_epi16(in[13], in[2]);
1290 0 : u[3] = _mm_unpackhi_epi16(in[13], in[2]);
1291 0 : u[4] = _mm_unpacklo_epi16(in[11], in[4]);
1292 0 : u[5] = _mm_unpackhi_epi16(in[11], in[4]);
1293 0 : u[6] = _mm_unpacklo_epi16(in[9], in[6]);
1294 0 : u[7] = _mm_unpackhi_epi16(in[9], in[6]);
1295 0 : u[8] = _mm_unpacklo_epi16(in[7], in[8]);
1296 0 : u[9] = _mm_unpackhi_epi16(in[7], in[8]);
1297 0 : u[10] = _mm_unpacklo_epi16(in[5], in[10]);
1298 0 : u[11] = _mm_unpackhi_epi16(in[5], in[10]);
1299 0 : u[12] = _mm_unpacklo_epi16(in[3], in[12]);
1300 0 : u[13] = _mm_unpackhi_epi16(in[3], in[12]);
1301 0 : u[14] = _mm_unpacklo_epi16(in[1], in[14]);
1302 0 : u[15] = _mm_unpackhi_epi16(in[1], in[14]);
1303 :
1304 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
1305 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
1306 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
1307 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
1308 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
1309 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
1310 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
1311 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
1312 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
1313 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
1314 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
1315 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
1316 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
1317 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
1318 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
1319 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
1320 0 : v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
1321 0 : v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
1322 0 : v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
1323 0 : v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
1324 0 : v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
1325 0 : v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
1326 0 : v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
1327 0 : v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
1328 0 : v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
1329 0 : v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
1330 0 : v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
1331 0 : v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
1332 0 : v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
1333 0 : v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
1334 0 : v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
1335 0 : v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
1336 :
1337 0 : u[0] = _mm_add_epi32(v[0], v[16]);
1338 0 : u[1] = _mm_add_epi32(v[1], v[17]);
1339 0 : u[2] = _mm_add_epi32(v[2], v[18]);
1340 0 : u[3] = _mm_add_epi32(v[3], v[19]);
1341 0 : u[4] = _mm_add_epi32(v[4], v[20]);
1342 0 : u[5] = _mm_add_epi32(v[5], v[21]);
1343 0 : u[6] = _mm_add_epi32(v[6], v[22]);
1344 0 : u[7] = _mm_add_epi32(v[7], v[23]);
1345 0 : u[8] = _mm_add_epi32(v[8], v[24]);
1346 0 : u[9] = _mm_add_epi32(v[9], v[25]);
1347 0 : u[10] = _mm_add_epi32(v[10], v[26]);
1348 0 : u[11] = _mm_add_epi32(v[11], v[27]);
1349 0 : u[12] = _mm_add_epi32(v[12], v[28]);
1350 0 : u[13] = _mm_add_epi32(v[13], v[29]);
1351 0 : u[14] = _mm_add_epi32(v[14], v[30]);
1352 0 : u[15] = _mm_add_epi32(v[15], v[31]);
1353 0 : u[16] = _mm_sub_epi32(v[0], v[16]);
1354 0 : u[17] = _mm_sub_epi32(v[1], v[17]);
1355 0 : u[18] = _mm_sub_epi32(v[2], v[18]);
1356 0 : u[19] = _mm_sub_epi32(v[3], v[19]);
1357 0 : u[20] = _mm_sub_epi32(v[4], v[20]);
1358 0 : u[21] = _mm_sub_epi32(v[5], v[21]);
1359 0 : u[22] = _mm_sub_epi32(v[6], v[22]);
1360 0 : u[23] = _mm_sub_epi32(v[7], v[23]);
1361 0 : u[24] = _mm_sub_epi32(v[8], v[24]);
1362 0 : u[25] = _mm_sub_epi32(v[9], v[25]);
1363 0 : u[26] = _mm_sub_epi32(v[10], v[26]);
1364 0 : u[27] = _mm_sub_epi32(v[11], v[27]);
1365 0 : u[28] = _mm_sub_epi32(v[12], v[28]);
1366 0 : u[29] = _mm_sub_epi32(v[13], v[29]);
1367 0 : u[30] = _mm_sub_epi32(v[14], v[30]);
1368 0 : u[31] = _mm_sub_epi32(v[15], v[31]);
1369 :
1370 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1371 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1372 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1373 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1374 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1375 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1376 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1377 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1378 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1379 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1380 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1381 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1382 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1383 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1384 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1385 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1386 0 : v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
1387 0 : v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
1388 0 : v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
1389 0 : v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
1390 0 : v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
1391 0 : v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
1392 0 : v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
1393 0 : v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
1394 0 : v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
1395 0 : v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
1396 0 : v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
1397 0 : v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
1398 0 : v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
1399 0 : v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
1400 0 : v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
1401 0 : v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
1402 :
1403 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1404 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1405 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1406 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1407 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1408 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1409 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1410 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1411 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
1412 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
1413 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1414 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1415 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1416 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1417 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1418 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1419 0 : u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
1420 0 : u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
1421 0 : u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
1422 0 : u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
1423 0 : u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
1424 0 : u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
1425 0 : u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
1426 0 : u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
1427 0 : u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
1428 0 : u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
1429 0 : u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
1430 0 : u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
1431 0 : u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
1432 0 : u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
1433 0 : u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
1434 0 : u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
1435 :
1436 0 : s[0] = _mm_packs_epi32(u[0], u[1]);
1437 0 : s[1] = _mm_packs_epi32(u[2], u[3]);
1438 0 : s[2] = _mm_packs_epi32(u[4], u[5]);
1439 0 : s[3] = _mm_packs_epi32(u[6], u[7]);
1440 0 : s[4] = _mm_packs_epi32(u[8], u[9]);
1441 0 : s[5] = _mm_packs_epi32(u[10], u[11]);
1442 0 : s[6] = _mm_packs_epi32(u[12], u[13]);
1443 0 : s[7] = _mm_packs_epi32(u[14], u[15]);
1444 0 : s[8] = _mm_packs_epi32(u[16], u[17]);
1445 0 : s[9] = _mm_packs_epi32(u[18], u[19]);
1446 0 : s[10] = _mm_packs_epi32(u[20], u[21]);
1447 0 : s[11] = _mm_packs_epi32(u[22], u[23]);
1448 0 : s[12] = _mm_packs_epi32(u[24], u[25]);
1449 0 : s[13] = _mm_packs_epi32(u[26], u[27]);
1450 0 : s[14] = _mm_packs_epi32(u[28], u[29]);
1451 0 : s[15] = _mm_packs_epi32(u[30], u[31]);
1452 :
1453 : // stage 2
1454 0 : u[0] = _mm_unpacklo_epi16(s[8], s[9]);
1455 0 : u[1] = _mm_unpackhi_epi16(s[8], s[9]);
1456 0 : u[2] = _mm_unpacklo_epi16(s[10], s[11]);
1457 0 : u[3] = _mm_unpackhi_epi16(s[10], s[11]);
1458 0 : u[4] = _mm_unpacklo_epi16(s[12], s[13]);
1459 0 : u[5] = _mm_unpackhi_epi16(s[12], s[13]);
1460 0 : u[6] = _mm_unpacklo_epi16(s[14], s[15]);
1461 0 : u[7] = _mm_unpackhi_epi16(s[14], s[15]);
1462 :
1463 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
1464 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
1465 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
1466 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
1467 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
1468 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
1469 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
1470 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
1471 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
1472 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
1473 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
1474 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
1475 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
1476 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
1477 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
1478 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
1479 :
1480 0 : u[0] = _mm_add_epi32(v[0], v[8]);
1481 0 : u[1] = _mm_add_epi32(v[1], v[9]);
1482 0 : u[2] = _mm_add_epi32(v[2], v[10]);
1483 0 : u[3] = _mm_add_epi32(v[3], v[11]);
1484 0 : u[4] = _mm_add_epi32(v[4], v[12]);
1485 0 : u[5] = _mm_add_epi32(v[5], v[13]);
1486 0 : u[6] = _mm_add_epi32(v[6], v[14]);
1487 0 : u[7] = _mm_add_epi32(v[7], v[15]);
1488 0 : u[8] = _mm_sub_epi32(v[0], v[8]);
1489 0 : u[9] = _mm_sub_epi32(v[1], v[9]);
1490 0 : u[10] = _mm_sub_epi32(v[2], v[10]);
1491 0 : u[11] = _mm_sub_epi32(v[3], v[11]);
1492 0 : u[12] = _mm_sub_epi32(v[4], v[12]);
1493 0 : u[13] = _mm_sub_epi32(v[5], v[13]);
1494 0 : u[14] = _mm_sub_epi32(v[6], v[14]);
1495 0 : u[15] = _mm_sub_epi32(v[7], v[15]);
1496 :
1497 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1498 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1499 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1500 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1501 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1502 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1503 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1504 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1505 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1506 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1507 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1508 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1509 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1510 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1511 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1512 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1513 :
1514 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1515 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1516 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1517 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1518 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1519 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1520 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1521 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1522 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
1523 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
1524 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1525 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1526 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1527 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1528 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1529 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1530 :
1531 0 : x[0] = _mm_add_epi16(s[0], s[4]);
1532 0 : x[1] = _mm_add_epi16(s[1], s[5]);
1533 0 : x[2] = _mm_add_epi16(s[2], s[6]);
1534 0 : x[3] = _mm_add_epi16(s[3], s[7]);
1535 0 : x[4] = _mm_sub_epi16(s[0], s[4]);
1536 0 : x[5] = _mm_sub_epi16(s[1], s[5]);
1537 0 : x[6] = _mm_sub_epi16(s[2], s[6]);
1538 0 : x[7] = _mm_sub_epi16(s[3], s[7]);
1539 0 : x[8] = _mm_packs_epi32(u[0], u[1]);
1540 0 : x[9] = _mm_packs_epi32(u[2], u[3]);
1541 0 : x[10] = _mm_packs_epi32(u[4], u[5]);
1542 0 : x[11] = _mm_packs_epi32(u[6], u[7]);
1543 0 : x[12] = _mm_packs_epi32(u[8], u[9]);
1544 0 : x[13] = _mm_packs_epi32(u[10], u[11]);
1545 0 : x[14] = _mm_packs_epi32(u[12], u[13]);
1546 0 : x[15] = _mm_packs_epi32(u[14], u[15]);
1547 :
1548 : // stage 3
1549 0 : u[0] = _mm_unpacklo_epi16(x[4], x[5]);
1550 0 : u[1] = _mm_unpackhi_epi16(x[4], x[5]);
1551 0 : u[2] = _mm_unpacklo_epi16(x[6], x[7]);
1552 0 : u[3] = _mm_unpackhi_epi16(x[6], x[7]);
1553 0 : u[4] = _mm_unpacklo_epi16(x[12], x[13]);
1554 0 : u[5] = _mm_unpackhi_epi16(x[12], x[13]);
1555 0 : u[6] = _mm_unpacklo_epi16(x[14], x[15]);
1556 0 : u[7] = _mm_unpackhi_epi16(x[14], x[15]);
1557 :
1558 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
1559 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
1560 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
1561 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
1562 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
1563 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
1564 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
1565 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
1566 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
1567 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
1568 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
1569 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
1570 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
1571 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
1572 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
1573 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
1574 :
1575 0 : u[0] = _mm_add_epi32(v[0], v[4]);
1576 0 : u[1] = _mm_add_epi32(v[1], v[5]);
1577 0 : u[2] = _mm_add_epi32(v[2], v[6]);
1578 0 : u[3] = _mm_add_epi32(v[3], v[7]);
1579 0 : u[4] = _mm_sub_epi32(v[0], v[4]);
1580 0 : u[5] = _mm_sub_epi32(v[1], v[5]);
1581 0 : u[6] = _mm_sub_epi32(v[2], v[6]);
1582 0 : u[7] = _mm_sub_epi32(v[3], v[7]);
1583 0 : u[8] = _mm_add_epi32(v[8], v[12]);
1584 0 : u[9] = _mm_add_epi32(v[9], v[13]);
1585 0 : u[10] = _mm_add_epi32(v[10], v[14]);
1586 0 : u[11] = _mm_add_epi32(v[11], v[15]);
1587 0 : u[12] = _mm_sub_epi32(v[8], v[12]);
1588 0 : u[13] = _mm_sub_epi32(v[9], v[13]);
1589 0 : u[14] = _mm_sub_epi32(v[10], v[14]);
1590 0 : u[15] = _mm_sub_epi32(v[11], v[15]);
1591 :
1592 0 : u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1593 0 : u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1594 0 : u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1595 0 : u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1596 0 : u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1597 0 : u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1598 0 : u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1599 0 : u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1600 0 : u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1601 0 : u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1602 0 : u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1603 0 : u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1604 0 : u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1605 0 : u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1606 0 : u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1607 0 : u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1608 :
1609 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1610 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1611 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1612 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1613 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1614 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1615 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1616 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1617 0 : v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1618 0 : v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1619 0 : v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1620 0 : v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1621 0 : v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1622 0 : v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1623 0 : v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1624 0 : v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1625 :
1626 0 : s[0] = _mm_add_epi16(x[0], x[2]);
1627 0 : s[1] = _mm_add_epi16(x[1], x[3]);
1628 0 : s[2] = _mm_sub_epi16(x[0], x[2]);
1629 0 : s[3] = _mm_sub_epi16(x[1], x[3]);
1630 0 : s[4] = _mm_packs_epi32(v[0], v[1]);
1631 0 : s[5] = _mm_packs_epi32(v[2], v[3]);
1632 0 : s[6] = _mm_packs_epi32(v[4], v[5]);
1633 0 : s[7] = _mm_packs_epi32(v[6], v[7]);
1634 0 : s[8] = _mm_add_epi16(x[8], x[10]);
1635 0 : s[9] = _mm_add_epi16(x[9], x[11]);
1636 0 : s[10] = _mm_sub_epi16(x[8], x[10]);
1637 0 : s[11] = _mm_sub_epi16(x[9], x[11]);
1638 0 : s[12] = _mm_packs_epi32(v[8], v[9]);
1639 0 : s[13] = _mm_packs_epi32(v[10], v[11]);
1640 0 : s[14] = _mm_packs_epi32(v[12], v[13]);
1641 0 : s[15] = _mm_packs_epi32(v[14], v[15]);
1642 :
1643 : // stage 4
1644 0 : u[0] = _mm_unpacklo_epi16(s[2], s[3]);
1645 0 : u[1] = _mm_unpackhi_epi16(s[2], s[3]);
1646 0 : u[2] = _mm_unpacklo_epi16(s[6], s[7]);
1647 0 : u[3] = _mm_unpackhi_epi16(s[6], s[7]);
1648 0 : u[4] = _mm_unpacklo_epi16(s[10], s[11]);
1649 0 : u[5] = _mm_unpackhi_epi16(s[10], s[11]);
1650 0 : u[6] = _mm_unpacklo_epi16(s[14], s[15]);
1651 0 : u[7] = _mm_unpackhi_epi16(s[14], s[15]);
1652 :
1653 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
1654 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
1655 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
1656 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
1657 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
1658 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
1659 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
1660 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
1661 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
1662 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
1663 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
1664 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
1665 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
1666 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
1667 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
1668 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
1669 :
1670 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1671 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1672 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1673 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1674 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1675 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1676 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1677 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1678 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1679 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1680 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1681 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1682 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1683 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1684 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1685 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1686 :
1687 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1688 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1689 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1690 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1691 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1692 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1693 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1694 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1695 0 : v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1696 0 : v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1697 0 : v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1698 0 : v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1699 0 : v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1700 0 : v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1701 0 : v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1702 0 : v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1703 :
1704 0 : in[0] = s[0];
1705 0 : in[1] = _mm_sub_epi16(kZero, s[8]);
1706 0 : in[2] = s[12];
1707 0 : in[3] = _mm_sub_epi16(kZero, s[4]);
1708 0 : in[4] = _mm_packs_epi32(v[4], v[5]);
1709 0 : in[5] = _mm_packs_epi32(v[12], v[13]);
1710 0 : in[6] = _mm_packs_epi32(v[8], v[9]);
1711 0 : in[7] = _mm_packs_epi32(v[0], v[1]);
1712 0 : in[8] = _mm_packs_epi32(v[2], v[3]);
1713 0 : in[9] = _mm_packs_epi32(v[10], v[11]);
1714 0 : in[10] = _mm_packs_epi32(v[14], v[15]);
1715 0 : in[11] = _mm_packs_epi32(v[6], v[7]);
1716 0 : in[12] = s[5];
1717 0 : in[13] = _mm_sub_epi16(kZero, s[13]);
1718 0 : in[14] = s[9];
1719 0 : in[15] = _mm_sub_epi16(kZero, s[1]);
1720 0 : }
1721 :
1722 0 : void idct16_8col(__m128i *in) {
1723 0 : const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
1724 0 : const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
1725 0 : const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
1726 0 : const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
1727 0 : const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
1728 0 : const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
1729 0 : const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
1730 0 : const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
1731 0 : const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
1732 0 : const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
1733 0 : const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
1734 0 : const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
1735 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
1736 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1737 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
1738 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
1739 0 : const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
1740 0 : const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
1741 0 : const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
1742 0 : const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1743 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
1744 : __m128i v[16], u[16], s[16], t[16];
1745 :
1746 : // stage 1
1747 0 : s[0] = in[0];
1748 0 : s[1] = in[8];
1749 0 : s[2] = in[4];
1750 0 : s[3] = in[12];
1751 0 : s[4] = in[2];
1752 0 : s[5] = in[10];
1753 0 : s[6] = in[6];
1754 0 : s[7] = in[14];
1755 0 : s[8] = in[1];
1756 0 : s[9] = in[9];
1757 0 : s[10] = in[5];
1758 0 : s[11] = in[13];
1759 0 : s[12] = in[3];
1760 0 : s[13] = in[11];
1761 0 : s[14] = in[7];
1762 0 : s[15] = in[15];
1763 :
1764 : // stage 2
1765 0 : u[0] = _mm_unpacklo_epi16(s[8], s[15]);
1766 0 : u[1] = _mm_unpackhi_epi16(s[8], s[15]);
1767 0 : u[2] = _mm_unpacklo_epi16(s[9], s[14]);
1768 0 : u[3] = _mm_unpackhi_epi16(s[9], s[14]);
1769 0 : u[4] = _mm_unpacklo_epi16(s[10], s[13]);
1770 0 : u[5] = _mm_unpackhi_epi16(s[10], s[13]);
1771 0 : u[6] = _mm_unpacklo_epi16(s[11], s[12]);
1772 0 : u[7] = _mm_unpackhi_epi16(s[11], s[12]);
1773 :
1774 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
1775 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
1776 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
1777 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
1778 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
1779 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
1780 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
1781 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
1782 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
1783 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10);
1784 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22);
1785 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22);
1786 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26);
1787 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26);
1788 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06);
1789 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06);
1790 :
1791 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1792 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1793 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1794 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1795 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1796 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1797 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1798 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1799 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1800 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1801 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1802 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1803 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1804 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1805 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1806 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1807 :
1808 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1809 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1810 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1811 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1812 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1813 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1814 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1815 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1816 0 : u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1817 0 : u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1818 0 : u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1819 0 : u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1820 0 : u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1821 0 : u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1822 0 : u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1823 0 : u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1824 :
1825 0 : s[8] = _mm_packs_epi32(u[0], u[1]);
1826 0 : s[15] = _mm_packs_epi32(u[2], u[3]);
1827 0 : s[9] = _mm_packs_epi32(u[4], u[5]);
1828 0 : s[14] = _mm_packs_epi32(u[6], u[7]);
1829 0 : s[10] = _mm_packs_epi32(u[8], u[9]);
1830 0 : s[13] = _mm_packs_epi32(u[10], u[11]);
1831 0 : s[11] = _mm_packs_epi32(u[12], u[13]);
1832 0 : s[12] = _mm_packs_epi32(u[14], u[15]);
1833 :
1834 : // stage 3
1835 0 : t[0] = s[0];
1836 0 : t[1] = s[1];
1837 0 : t[2] = s[2];
1838 0 : t[3] = s[3];
1839 0 : u[0] = _mm_unpacklo_epi16(s[4], s[7]);
1840 0 : u[1] = _mm_unpackhi_epi16(s[4], s[7]);
1841 0 : u[2] = _mm_unpacklo_epi16(s[5], s[6]);
1842 0 : u[3] = _mm_unpackhi_epi16(s[5], s[6]);
1843 :
1844 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
1845 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
1846 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
1847 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
1848 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
1849 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
1850 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
1851 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
1852 :
1853 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1854 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1855 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1856 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1857 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1858 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1859 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1860 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1861 :
1862 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1863 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1864 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1865 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1866 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1867 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1868 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1869 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1870 :
1871 0 : t[4] = _mm_packs_epi32(u[0], u[1]);
1872 0 : t[7] = _mm_packs_epi32(u[2], u[3]);
1873 0 : t[5] = _mm_packs_epi32(u[4], u[5]);
1874 0 : t[6] = _mm_packs_epi32(u[6], u[7]);
1875 0 : t[8] = _mm_add_epi16(s[8], s[9]);
1876 0 : t[9] = _mm_sub_epi16(s[8], s[9]);
1877 0 : t[10] = _mm_sub_epi16(s[11], s[10]);
1878 0 : t[11] = _mm_add_epi16(s[10], s[11]);
1879 0 : t[12] = _mm_add_epi16(s[12], s[13]);
1880 0 : t[13] = _mm_sub_epi16(s[12], s[13]);
1881 0 : t[14] = _mm_sub_epi16(s[15], s[14]);
1882 0 : t[15] = _mm_add_epi16(s[14], s[15]);
1883 :
1884 : // stage 4
1885 0 : u[0] = _mm_unpacklo_epi16(t[0], t[1]);
1886 0 : u[1] = _mm_unpackhi_epi16(t[0], t[1]);
1887 0 : u[2] = _mm_unpacklo_epi16(t[2], t[3]);
1888 0 : u[3] = _mm_unpackhi_epi16(t[2], t[3]);
1889 0 : u[4] = _mm_unpacklo_epi16(t[9], t[14]);
1890 0 : u[5] = _mm_unpackhi_epi16(t[9], t[14]);
1891 0 : u[6] = _mm_unpacklo_epi16(t[10], t[13]);
1892 0 : u[7] = _mm_unpackhi_epi16(t[10], t[13]);
1893 :
1894 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
1895 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
1896 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
1897 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
1898 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08);
1899 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08);
1900 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
1901 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
1902 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24);
1903 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24);
1904 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08);
1905 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08);
1906 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08);
1907 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08);
1908 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24);
1909 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24);
1910 :
1911 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1912 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1913 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1914 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1915 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1916 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1917 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1918 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1919 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1920 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1921 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1922 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1923 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1924 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1925 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1926 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1927 :
1928 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1929 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1930 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1931 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1932 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1933 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1934 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1935 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1936 0 : u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1937 0 : u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1938 0 : u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1939 0 : u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1940 0 : u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1941 0 : u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1942 0 : u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1943 0 : u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1944 :
1945 0 : s[0] = _mm_packs_epi32(u[0], u[1]);
1946 0 : s[1] = _mm_packs_epi32(u[2], u[3]);
1947 0 : s[2] = _mm_packs_epi32(u[4], u[5]);
1948 0 : s[3] = _mm_packs_epi32(u[6], u[7]);
1949 0 : s[4] = _mm_add_epi16(t[4], t[5]);
1950 0 : s[5] = _mm_sub_epi16(t[4], t[5]);
1951 0 : s[6] = _mm_sub_epi16(t[7], t[6]);
1952 0 : s[7] = _mm_add_epi16(t[6], t[7]);
1953 0 : s[8] = t[8];
1954 0 : s[15] = t[15];
1955 0 : s[9] = _mm_packs_epi32(u[8], u[9]);
1956 0 : s[14] = _mm_packs_epi32(u[10], u[11]);
1957 0 : s[10] = _mm_packs_epi32(u[12], u[13]);
1958 0 : s[13] = _mm_packs_epi32(u[14], u[15]);
1959 0 : s[11] = t[11];
1960 0 : s[12] = t[12];
1961 :
1962 : // stage 5
1963 0 : t[0] = _mm_add_epi16(s[0], s[3]);
1964 0 : t[1] = _mm_add_epi16(s[1], s[2]);
1965 0 : t[2] = _mm_sub_epi16(s[1], s[2]);
1966 0 : t[3] = _mm_sub_epi16(s[0], s[3]);
1967 0 : t[4] = s[4];
1968 0 : t[7] = s[7];
1969 :
1970 0 : u[0] = _mm_unpacklo_epi16(s[5], s[6]);
1971 0 : u[1] = _mm_unpackhi_epi16(s[5], s[6]);
1972 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
1973 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
1974 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
1975 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
1976 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1977 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1978 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1979 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1980 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1981 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1982 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1983 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1984 0 : t[5] = _mm_packs_epi32(u[0], u[1]);
1985 0 : t[6] = _mm_packs_epi32(u[2], u[3]);
1986 :
1987 0 : t[8] = _mm_add_epi16(s[8], s[11]);
1988 0 : t[9] = _mm_add_epi16(s[9], s[10]);
1989 0 : t[10] = _mm_sub_epi16(s[9], s[10]);
1990 0 : t[11] = _mm_sub_epi16(s[8], s[11]);
1991 0 : t[12] = _mm_sub_epi16(s[15], s[12]);
1992 0 : t[13] = _mm_sub_epi16(s[14], s[13]);
1993 0 : t[14] = _mm_add_epi16(s[13], s[14]);
1994 0 : t[15] = _mm_add_epi16(s[12], s[15]);
1995 :
1996 : // stage 6
1997 0 : s[0] = _mm_add_epi16(t[0], t[7]);
1998 0 : s[1] = _mm_add_epi16(t[1], t[6]);
1999 0 : s[2] = _mm_add_epi16(t[2], t[5]);
2000 0 : s[3] = _mm_add_epi16(t[3], t[4]);
2001 0 : s[4] = _mm_sub_epi16(t[3], t[4]);
2002 0 : s[5] = _mm_sub_epi16(t[2], t[5]);
2003 0 : s[6] = _mm_sub_epi16(t[1], t[6]);
2004 0 : s[7] = _mm_sub_epi16(t[0], t[7]);
2005 0 : s[8] = t[8];
2006 0 : s[9] = t[9];
2007 :
2008 0 : u[0] = _mm_unpacklo_epi16(t[10], t[13]);
2009 0 : u[1] = _mm_unpackhi_epi16(t[10], t[13]);
2010 0 : u[2] = _mm_unpacklo_epi16(t[11], t[12]);
2011 0 : u[3] = _mm_unpackhi_epi16(t[11], t[12]);
2012 :
2013 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
2014 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
2015 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
2016 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
2017 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
2018 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
2019 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
2020 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
2021 :
2022 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
2023 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
2024 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
2025 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
2026 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
2027 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
2028 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
2029 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
2030 :
2031 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
2032 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
2033 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
2034 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
2035 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
2036 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
2037 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
2038 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
2039 :
2040 0 : s[10] = _mm_packs_epi32(u[0], u[1]);
2041 0 : s[13] = _mm_packs_epi32(u[2], u[3]);
2042 0 : s[11] = _mm_packs_epi32(u[4], u[5]);
2043 0 : s[12] = _mm_packs_epi32(u[6], u[7]);
2044 0 : s[14] = t[14];
2045 0 : s[15] = t[15];
2046 :
2047 : // stage 7
2048 0 : in[0] = _mm_add_epi16(s[0], s[15]);
2049 0 : in[1] = _mm_add_epi16(s[1], s[14]);
2050 0 : in[2] = _mm_add_epi16(s[2], s[13]);
2051 0 : in[3] = _mm_add_epi16(s[3], s[12]);
2052 0 : in[4] = _mm_add_epi16(s[4], s[11]);
2053 0 : in[5] = _mm_add_epi16(s[5], s[10]);
2054 0 : in[6] = _mm_add_epi16(s[6], s[9]);
2055 0 : in[7] = _mm_add_epi16(s[7], s[8]);
2056 0 : in[8] = _mm_sub_epi16(s[7], s[8]);
2057 0 : in[9] = _mm_sub_epi16(s[6], s[9]);
2058 0 : in[10] = _mm_sub_epi16(s[5], s[10]);
2059 0 : in[11] = _mm_sub_epi16(s[4], s[11]);
2060 0 : in[12] = _mm_sub_epi16(s[3], s[12]);
2061 0 : in[13] = _mm_sub_epi16(s[2], s[13]);
2062 0 : in[14] = _mm_sub_epi16(s[1], s[14]);
2063 0 : in[15] = _mm_sub_epi16(s[0], s[15]);
2064 0 : }
2065 :
2066 0 : void aom_idct16_sse2(__m128i *in0, __m128i *in1) {
2067 0 : array_transpose_16x16(in0, in1);
2068 0 : idct16_8col(in0);
2069 0 : idct16_8col(in1);
2070 0 : }
2071 :
2072 0 : void aom_iadst16_sse2(__m128i *in0, __m128i *in1) {
2073 0 : array_transpose_16x16(in0, in1);
2074 0 : iadst16_8col(in0);
2075 0 : iadst16_8col(in1);
2076 0 : }
2077 :
2078 0 : void aom_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest,
2079 : int stride) {
2080 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
2081 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
2082 0 : const __m128i zero = _mm_setzero_si128();
2083 :
2084 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
2085 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
2086 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
2087 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
2088 :
2089 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
2090 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
2091 :
2092 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
2093 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
2094 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
2095 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
2096 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
2097 0 : const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
2098 :
2099 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
2100 : __m128i in[16], l[16];
2101 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_8,
2102 : stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, stp1_8_0,
2103 : stp1_12_0;
2104 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
2105 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14;
2106 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
2107 : int i;
2108 : // First 1-D inverse DCT
2109 : // Load input data.
2110 0 : in[0] = load_input_data(input);
2111 0 : in[1] = load_input_data(input + 8 * 2);
2112 0 : in[2] = load_input_data(input + 8 * 4);
2113 0 : in[3] = load_input_data(input + 8 * 6);
2114 :
2115 0 : TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]);
2116 :
2117 : // Stage2
2118 : {
2119 0 : const __m128i lo_1_15 = _mm_unpackhi_epi16(in[0], zero);
2120 0 : const __m128i lo_13_3 = _mm_unpackhi_epi16(zero, in[1]);
2121 :
2122 0 : tmp0 = _mm_madd_epi16(lo_1_15, stg2_0);
2123 0 : tmp2 = _mm_madd_epi16(lo_1_15, stg2_1);
2124 0 : tmp5 = _mm_madd_epi16(lo_13_3, stg2_6);
2125 0 : tmp7 = _mm_madd_epi16(lo_13_3, stg2_7);
2126 :
2127 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2128 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2129 0 : tmp5 = _mm_add_epi32(tmp5, rounding);
2130 0 : tmp7 = _mm_add_epi32(tmp7, rounding);
2131 :
2132 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2133 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2134 0 : tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
2135 0 : tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
2136 :
2137 0 : stp2_8 = _mm_packs_epi32(tmp0, tmp2);
2138 0 : stp2_11 = _mm_packs_epi32(tmp5, tmp7);
2139 : }
2140 :
2141 : // Stage3
2142 : {
2143 0 : const __m128i lo_2_14 = _mm_unpacklo_epi16(in[1], zero);
2144 :
2145 0 : tmp0 = _mm_madd_epi16(lo_2_14, stg3_0);
2146 0 : tmp2 = _mm_madd_epi16(lo_2_14, stg3_1);
2147 :
2148 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2149 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2150 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2151 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2152 :
2153 0 : stp1_13 = _mm_unpackhi_epi64(stp2_11, zero);
2154 0 : stp1_14 = _mm_unpackhi_epi64(stp2_8, zero);
2155 :
2156 0 : stp1_4 = _mm_packs_epi32(tmp0, tmp2);
2157 : }
2158 :
2159 : // Stage4
2160 : {
2161 0 : const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero);
2162 0 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14);
2163 0 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13);
2164 :
2165 0 : tmp0 = _mm_madd_epi16(lo_0_8, stg4_0);
2166 0 : tmp2 = _mm_madd_epi16(lo_0_8, stg4_1);
2167 0 : tmp1 = _mm_madd_epi16(lo_9_14, stg4_4);
2168 0 : tmp3 = _mm_madd_epi16(lo_9_14, stg4_5);
2169 0 : tmp5 = _mm_madd_epi16(lo_10_13, stg4_6);
2170 0 : tmp7 = _mm_madd_epi16(lo_10_13, stg4_7);
2171 :
2172 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2173 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2174 0 : tmp1 = _mm_add_epi32(tmp1, rounding);
2175 0 : tmp3 = _mm_add_epi32(tmp3, rounding);
2176 0 : tmp5 = _mm_add_epi32(tmp5, rounding);
2177 0 : tmp7 = _mm_add_epi32(tmp7, rounding);
2178 :
2179 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2180 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2181 0 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
2182 0 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
2183 0 : tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
2184 0 : tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
2185 :
2186 0 : stp1_0 = _mm_packs_epi32(tmp0, tmp0);
2187 0 : stp1_1 = _mm_packs_epi32(tmp2, tmp2);
2188 0 : stp2_9 = _mm_packs_epi32(tmp1, tmp3);
2189 0 : stp2_10 = _mm_packs_epi32(tmp5, tmp7);
2190 :
2191 0 : stp2_6 = _mm_unpackhi_epi64(stp1_4, zero);
2192 : }
2193 :
2194 : // Stage5 and Stage6
2195 : {
2196 0 : tmp0 = _mm_add_epi16(stp2_8, stp2_11);
2197 0 : tmp1 = _mm_sub_epi16(stp2_8, stp2_11);
2198 0 : tmp2 = _mm_add_epi16(stp2_9, stp2_10);
2199 0 : tmp3 = _mm_sub_epi16(stp2_9, stp2_10);
2200 :
2201 0 : stp1_9 = _mm_unpacklo_epi64(tmp2, zero);
2202 0 : stp1_10 = _mm_unpacklo_epi64(tmp3, zero);
2203 0 : stp1_8 = _mm_unpacklo_epi64(tmp0, zero);
2204 0 : stp1_11 = _mm_unpacklo_epi64(tmp1, zero);
2205 :
2206 0 : stp1_13 = _mm_unpackhi_epi64(tmp3, zero);
2207 0 : stp1_14 = _mm_unpackhi_epi64(tmp2, zero);
2208 0 : stp1_12 = _mm_unpackhi_epi64(tmp1, zero);
2209 0 : stp1_15 = _mm_unpackhi_epi64(tmp0, zero);
2210 : }
2211 :
2212 : // Stage6
2213 : {
2214 0 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp1_4);
2215 0 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);
2216 0 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);
2217 :
2218 0 : tmp1 = _mm_madd_epi16(lo_6_5, stg4_1);
2219 0 : tmp3 = _mm_madd_epi16(lo_6_5, stg4_0);
2220 0 : tmp0 = _mm_madd_epi16(lo_10_13, stg6_0);
2221 0 : tmp2 = _mm_madd_epi16(lo_10_13, stg4_0);
2222 0 : tmp4 = _mm_madd_epi16(lo_11_12, stg6_0);
2223 0 : tmp6 = _mm_madd_epi16(lo_11_12, stg4_0);
2224 :
2225 0 : tmp1 = _mm_add_epi32(tmp1, rounding);
2226 0 : tmp3 = _mm_add_epi32(tmp3, rounding);
2227 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2228 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2229 0 : tmp4 = _mm_add_epi32(tmp4, rounding);
2230 0 : tmp6 = _mm_add_epi32(tmp6, rounding);
2231 :
2232 0 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
2233 0 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
2234 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2235 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2236 0 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
2237 0 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
2238 :
2239 0 : stp1_6 = _mm_packs_epi32(tmp3, tmp1);
2240 :
2241 0 : stp2_10 = _mm_packs_epi32(tmp0, zero);
2242 0 : stp2_13 = _mm_packs_epi32(tmp2, zero);
2243 0 : stp2_11 = _mm_packs_epi32(tmp4, zero);
2244 0 : stp2_12 = _mm_packs_epi32(tmp6, zero);
2245 :
2246 0 : tmp0 = _mm_add_epi16(stp1_0, stp1_4);
2247 0 : tmp1 = _mm_sub_epi16(stp1_0, stp1_4);
2248 0 : tmp2 = _mm_add_epi16(stp1_1, stp1_6);
2249 0 : tmp3 = _mm_sub_epi16(stp1_1, stp1_6);
2250 :
2251 0 : stp2_0 = _mm_unpackhi_epi64(tmp0, zero);
2252 0 : stp2_1 = _mm_unpacklo_epi64(tmp2, zero);
2253 0 : stp2_2 = _mm_unpackhi_epi64(tmp2, zero);
2254 0 : stp2_3 = _mm_unpacklo_epi64(tmp0, zero);
2255 0 : stp2_4 = _mm_unpacklo_epi64(tmp1, zero);
2256 0 : stp2_5 = _mm_unpackhi_epi64(tmp3, zero);
2257 0 : stp2_6 = _mm_unpacklo_epi64(tmp3, zero);
2258 0 : stp2_7 = _mm_unpackhi_epi64(tmp1, zero);
2259 : }
2260 :
2261 : // Stage7. Left 8x16 only.
2262 0 : l[0] = _mm_add_epi16(stp2_0, stp1_15);
2263 0 : l[1] = _mm_add_epi16(stp2_1, stp1_14);
2264 0 : l[2] = _mm_add_epi16(stp2_2, stp2_13);
2265 0 : l[3] = _mm_add_epi16(stp2_3, stp2_12);
2266 0 : l[4] = _mm_add_epi16(stp2_4, stp2_11);
2267 0 : l[5] = _mm_add_epi16(stp2_5, stp2_10);
2268 0 : l[6] = _mm_add_epi16(stp2_6, stp1_9);
2269 0 : l[7] = _mm_add_epi16(stp2_7, stp1_8);
2270 0 : l[8] = _mm_sub_epi16(stp2_7, stp1_8);
2271 0 : l[9] = _mm_sub_epi16(stp2_6, stp1_9);
2272 0 : l[10] = _mm_sub_epi16(stp2_5, stp2_10);
2273 0 : l[11] = _mm_sub_epi16(stp2_4, stp2_11);
2274 0 : l[12] = _mm_sub_epi16(stp2_3, stp2_12);
2275 0 : l[13] = _mm_sub_epi16(stp2_2, stp2_13);
2276 0 : l[14] = _mm_sub_epi16(stp2_1, stp1_14);
2277 0 : l[15] = _mm_sub_epi16(stp2_0, stp1_15);
2278 :
2279 : // Second 1-D inverse transform, performed per 8x16 block
2280 0 : for (i = 0; i < 2; i++) {
2281 : int j;
2282 0 : array_transpose_4X8(l + 8 * i, in);
2283 :
2284 0 : IDCT16_10
2285 :
2286 : // Stage7
2287 0 : in[0] = _mm_add_epi16(stp2_0, stp1_15);
2288 0 : in[1] = _mm_add_epi16(stp2_1, stp1_14);
2289 0 : in[2] = _mm_add_epi16(stp2_2, stp2_13);
2290 0 : in[3] = _mm_add_epi16(stp2_3, stp2_12);
2291 0 : in[4] = _mm_add_epi16(stp2_4, stp2_11);
2292 0 : in[5] = _mm_add_epi16(stp2_5, stp2_10);
2293 0 : in[6] = _mm_add_epi16(stp2_6, stp1_9);
2294 0 : in[7] = _mm_add_epi16(stp2_7, stp1_8);
2295 0 : in[8] = _mm_sub_epi16(stp2_7, stp1_8);
2296 0 : in[9] = _mm_sub_epi16(stp2_6, stp1_9);
2297 0 : in[10] = _mm_sub_epi16(stp2_5, stp2_10);
2298 0 : in[11] = _mm_sub_epi16(stp2_4, stp2_11);
2299 0 : in[12] = _mm_sub_epi16(stp2_3, stp2_12);
2300 0 : in[13] = _mm_sub_epi16(stp2_2, stp2_13);
2301 0 : in[14] = _mm_sub_epi16(stp2_1, stp1_14);
2302 0 : in[15] = _mm_sub_epi16(stp2_0, stp1_15);
2303 :
2304 0 : for (j = 0; j < 16; ++j) {
2305 : // Final rounding and shift
2306 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
2307 0 : in[j] = _mm_srai_epi16(in[j], 6);
2308 0 : RECON_AND_STORE(dest + j * stride, in[j]);
2309 : }
2310 :
2311 0 : dest += 8;
2312 : }
2313 0 : }
2314 :
2315 : #define LOAD_DQCOEFF(reg, input) \
2316 : { \
2317 : reg = load_input_data(input); \
2318 : input += 8; \
2319 : }
2320 :
2321 : #define IDCT32_34 \
2322 : /* Stage1 */ \
2323 : { \
2324 : const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], zero); \
2325 : const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], zero); \
2326 : \
2327 : const __m128i lo_25_7 = _mm_unpacklo_epi16(zero, in[7]); \
2328 : const __m128i hi_25_7 = _mm_unpackhi_epi16(zero, in[7]); \
2329 : \
2330 : const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], zero); \
2331 : const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], zero); \
2332 : \
2333 : const __m128i lo_29_3 = _mm_unpacklo_epi16(zero, in[3]); \
2334 : const __m128i hi_29_3 = _mm_unpackhi_epi16(zero, in[3]); \
2335 : \
2336 : MULTIPLICATION_AND_ADD_2(lo_1_31, hi_1_31, stg1_0, stg1_1, stp1_16, \
2337 : stp1_31); \
2338 : MULTIPLICATION_AND_ADD_2(lo_25_7, hi_25_7, stg1_6, stg1_7, stp1_19, \
2339 : stp1_28); \
2340 : MULTIPLICATION_AND_ADD_2(lo_5_27, hi_5_27, stg1_8, stg1_9, stp1_20, \
2341 : stp1_27); \
2342 : MULTIPLICATION_AND_ADD_2(lo_29_3, hi_29_3, stg1_14, stg1_15, stp1_23, \
2343 : stp1_24); \
2344 : } \
2345 : \
2346 : /* Stage2 */ \
2347 : { \
2348 : const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], zero); \
2349 : const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], zero); \
2350 : \
2351 : const __m128i lo_26_6 = _mm_unpacklo_epi16(zero, in[6]); \
2352 : const __m128i hi_26_6 = _mm_unpackhi_epi16(zero, in[6]); \
2353 : \
2354 : MULTIPLICATION_AND_ADD_2(lo_2_30, hi_2_30, stg2_0, stg2_1, stp2_8, \
2355 : stp2_15); \
2356 : MULTIPLICATION_AND_ADD_2(lo_26_6, hi_26_6, stg2_6, stg2_7, stp2_11, \
2357 : stp2_12); \
2358 : \
2359 : stp2_16 = stp1_16; \
2360 : stp2_19 = stp1_19; \
2361 : \
2362 : stp2_20 = stp1_20; \
2363 : stp2_23 = stp1_23; \
2364 : \
2365 : stp2_24 = stp1_24; \
2366 : stp2_27 = stp1_27; \
2367 : \
2368 : stp2_28 = stp1_28; \
2369 : stp2_31 = stp1_31; \
2370 : } \
2371 : \
2372 : /* Stage3 */ \
2373 : { \
2374 : const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], zero); \
2375 : const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], zero); \
2376 : \
2377 : const __m128i lo_17_30 = _mm_unpacklo_epi16(stp1_16, stp1_31); \
2378 : const __m128i hi_17_30 = _mm_unpackhi_epi16(stp1_16, stp1_31); \
2379 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp1_19, stp1_28); \
2380 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp1_19, stp1_28); \
2381 : \
2382 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp1_20, stp1_27); \
2383 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp1_20, stp1_27); \
2384 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp1_23, stp1_24); \
2385 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp1_23, stp2_24); \
2386 : \
2387 : MULTIPLICATION_AND_ADD_2(lo_4_28, hi_4_28, stg3_0, stg3_1, stp1_4, \
2388 : stp1_7); \
2389 : \
2390 : stp1_8 = stp2_8; \
2391 : stp1_11 = stp2_11; \
2392 : stp1_12 = stp2_12; \
2393 : stp1_15 = stp2_15; \
2394 : \
2395 : MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
2396 : stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18, \
2397 : stp1_29) \
2398 : MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
2399 : stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \
2400 : stp1_25) \
2401 : \
2402 : stp1_16 = stp2_16; \
2403 : stp1_31 = stp2_31; \
2404 : stp1_19 = stp2_19; \
2405 : stp1_20 = stp2_20; \
2406 : stp1_23 = stp2_23; \
2407 : stp1_24 = stp2_24; \
2408 : stp1_27 = stp2_27; \
2409 : stp1_28 = stp2_28; \
2410 : } \
2411 : \
2412 : /* Stage4 */ \
2413 : { \
2414 : const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], zero); \
2415 : const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], zero); \
2416 : \
2417 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp2_15); \
2418 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp2_8, stp2_15); \
2419 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp2_12); \
2420 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp2_11, stp2_12); \
2421 : \
2422 : MULTIPLICATION_AND_ADD_2(lo_0_16, hi_0_16, stg4_0, stg4_1, stp2_0, \
2423 : stp2_1); \
2424 : \
2425 : stp2_4 = stp1_4; \
2426 : stp2_5 = stp1_4; \
2427 : stp2_6 = stp1_7; \
2428 : stp2_7 = stp1_7; \
2429 : \
2430 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
2431 : stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10, \
2432 : stp2_13) \
2433 : \
2434 : stp2_8 = stp1_8; \
2435 : stp2_15 = stp1_15; \
2436 : stp2_11 = stp1_11; \
2437 : stp2_12 = stp1_12; \
2438 : \
2439 : stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
2440 : stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
2441 : stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
2442 : stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
2443 : stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
2444 : stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
2445 : stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
2446 : stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
2447 : \
2448 : stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
2449 : stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
2450 : stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
2451 : stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
2452 : stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
2453 : stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
2454 : stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
2455 : stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
2456 : } \
2457 : \
2458 : /* Stage5 */ \
2459 : { \
2460 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
2461 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
2462 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
2463 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
2464 : \
2465 : const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
2466 : const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
2467 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2468 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2469 : \
2470 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2471 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2472 : \
2473 : stp1_0 = stp2_0; \
2474 : stp1_1 = stp2_1; \
2475 : stp1_2 = stp2_1; \
2476 : stp1_3 = stp2_0; \
2477 : \
2478 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
2479 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
2480 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
2481 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
2482 : \
2483 : tmp0 = _mm_add_epi32(tmp0, rounding); \
2484 : tmp1 = _mm_add_epi32(tmp1, rounding); \
2485 : tmp2 = _mm_add_epi32(tmp2, rounding); \
2486 : tmp3 = _mm_add_epi32(tmp3, rounding); \
2487 : \
2488 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
2489 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
2490 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
2491 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
2492 : \
2493 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
2494 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
2495 : \
2496 : stp1_4 = stp2_4; \
2497 : stp1_7 = stp2_7; \
2498 : \
2499 : stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
2500 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
2501 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
2502 : stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
2503 : stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
2504 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
2505 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
2506 : stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
2507 : \
2508 : stp1_16 = stp2_16; \
2509 : stp1_17 = stp2_17; \
2510 : \
2511 : MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
2512 : stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19, \
2513 : stp1_28) \
2514 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
2515 : stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21, \
2516 : stp1_26) \
2517 : \
2518 : stp1_22 = stp2_22; \
2519 : stp1_23 = stp2_23; \
2520 : stp1_24 = stp2_24; \
2521 : stp1_25 = stp2_25; \
2522 : stp1_30 = stp2_30; \
2523 : stp1_31 = stp2_31; \
2524 : } \
2525 : \
2526 : /* Stage6 */ \
2527 : { \
2528 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
2529 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
2530 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
2531 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
2532 : \
2533 : stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
2534 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
2535 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
2536 : stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
2537 : stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
2538 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
2539 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
2540 : stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
2541 : \
2542 : stp2_8 = stp1_8; \
2543 : stp2_9 = stp1_9; \
2544 : stp2_14 = stp1_14; \
2545 : stp2_15 = stp1_15; \
2546 : \
2547 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
2548 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
2549 : stp2_12) \
2550 : \
2551 : stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
2552 : stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
2553 : stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
2554 : stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
2555 : stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
2556 : stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
2557 : stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
2558 : stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
2559 : \
2560 : stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
2561 : stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
2562 : stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
2563 : stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
2564 : stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
2565 : stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
2566 : stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
2567 : stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
2568 : } \
2569 : \
2570 : /* Stage7 */ \
2571 : { \
2572 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2573 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2574 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2575 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2576 : \
2577 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
2578 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
2579 : const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
2580 : const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
2581 : \
2582 : stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
2583 : stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
2584 : stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
2585 : stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
2586 : stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
2587 : stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
2588 : stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
2589 : stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
2590 : stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
2591 : stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
2592 : stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
2593 : stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
2594 : stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
2595 : stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
2596 : stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
2597 : stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
2598 : \
2599 : stp1_16 = stp2_16; \
2600 : stp1_17 = stp2_17; \
2601 : stp1_18 = stp2_18; \
2602 : stp1_19 = stp2_19; \
2603 : \
2604 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
2605 : stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21, \
2606 : stp1_26) \
2607 : MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
2608 : stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23, \
2609 : stp1_24) \
2610 : \
2611 : stp1_28 = stp2_28; \
2612 : stp1_29 = stp2_29; \
2613 : stp1_30 = stp2_30; \
2614 : stp1_31 = stp2_31; \
2615 : }
2616 :
2617 : #define IDCT32(in0, in1) \
2618 : /* Stage1 */ \
2619 : { \
2620 : const __m128i lo_1_31 = _mm_unpacklo_epi16((in0)[1], (in1)[15]); \
2621 : const __m128i hi_1_31 = _mm_unpackhi_epi16((in0)[1], (in1)[15]); \
2622 : const __m128i lo_17_15 = _mm_unpacklo_epi16((in1)[1], (in0)[15]); \
2623 : const __m128i hi_17_15 = _mm_unpackhi_epi16((in1)[1], (in0)[15]); \
2624 : \
2625 : const __m128i lo_9_23 = _mm_unpacklo_epi16((in0)[9], (in1)[7]); \
2626 : const __m128i hi_9_23 = _mm_unpackhi_epi16((in0)[9], (in1)[7]); \
2627 : const __m128i lo_25_7 = _mm_unpacklo_epi16((in1)[9], (in0)[7]); \
2628 : const __m128i hi_25_7 = _mm_unpackhi_epi16((in1)[9], (in0)[7]); \
2629 : \
2630 : const __m128i lo_5_27 = _mm_unpacklo_epi16((in0)[5], (in1)[11]); \
2631 : const __m128i hi_5_27 = _mm_unpackhi_epi16((in0)[5], (in1)[11]); \
2632 : const __m128i lo_21_11 = _mm_unpacklo_epi16((in1)[5], (in0)[11]); \
2633 : const __m128i hi_21_11 = _mm_unpackhi_epi16((in1)[5], (in0)[11]); \
2634 : \
2635 : const __m128i lo_13_19 = _mm_unpacklo_epi16((in0)[13], (in1)[3]); \
2636 : const __m128i hi_13_19 = _mm_unpackhi_epi16((in0)[13], (in1)[3]); \
2637 : const __m128i lo_29_3 = _mm_unpacklo_epi16((in1)[13], (in0)[3]); \
2638 : const __m128i hi_29_3 = _mm_unpackhi_epi16((in1)[13], (in0)[3]); \
2639 : \
2640 : MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0, \
2641 : stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, stp1_17, \
2642 : stp1_30) \
2643 : MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, stg1_5, \
2644 : stg1_6, stg1_7, stp1_18, stp1_29, stp1_19, stp1_28) \
2645 : MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8, \
2646 : stg1_9, stg1_10, stg1_11, stp1_20, stp1_27, \
2647 : stp1_21, stp1_26) \
2648 : MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12, \
2649 : stg1_13, stg1_14, stg1_15, stp1_22, stp1_25, \
2650 : stp1_23, stp1_24) \
2651 : } \
2652 : \
2653 : /* Stage2 */ \
2654 : { \
2655 : const __m128i lo_2_30 = _mm_unpacklo_epi16((in0)[2], (in1)[14]); \
2656 : const __m128i hi_2_30 = _mm_unpackhi_epi16((in0)[2], (in1)[14]); \
2657 : const __m128i lo_18_14 = _mm_unpacklo_epi16((in1)[2], (in0)[14]); \
2658 : const __m128i hi_18_14 = _mm_unpackhi_epi16((in1)[2], (in0)[14]); \
2659 : \
2660 : const __m128i lo_10_22 = _mm_unpacklo_epi16((in0)[10], (in1)[6]); \
2661 : const __m128i hi_10_22 = _mm_unpackhi_epi16((in0)[10], (in1)[6]); \
2662 : const __m128i lo_26_6 = _mm_unpacklo_epi16((in1)[10], (in0)[6]); \
2663 : const __m128i hi_26_6 = _mm_unpackhi_epi16((in1)[10], (in0)[6]); \
2664 : \
2665 : MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0, \
2666 : stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, \
2667 : stp2_14) \
2668 : MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4, \
2669 : stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, stp2_11, \
2670 : stp2_12) \
2671 : \
2672 : stp2_16 = _mm_add_epi16(stp1_16, stp1_17); \
2673 : stp2_17 = _mm_sub_epi16(stp1_16, stp1_17); \
2674 : stp2_18 = _mm_sub_epi16(stp1_19, stp1_18); \
2675 : stp2_19 = _mm_add_epi16(stp1_19, stp1_18); \
2676 : \
2677 : stp2_20 = _mm_add_epi16(stp1_20, stp1_21); \
2678 : stp2_21 = _mm_sub_epi16(stp1_20, stp1_21); \
2679 : stp2_22 = _mm_sub_epi16(stp1_23, stp1_22); \
2680 : stp2_23 = _mm_add_epi16(stp1_23, stp1_22); \
2681 : \
2682 : stp2_24 = _mm_add_epi16(stp1_24, stp1_25); \
2683 : stp2_25 = _mm_sub_epi16(stp1_24, stp1_25); \
2684 : stp2_26 = _mm_sub_epi16(stp1_27, stp1_26); \
2685 : stp2_27 = _mm_add_epi16(stp1_27, stp1_26); \
2686 : \
2687 : stp2_28 = _mm_add_epi16(stp1_28, stp1_29); \
2688 : stp2_29 = _mm_sub_epi16(stp1_28, stp1_29); \
2689 : stp2_30 = _mm_sub_epi16(stp1_31, stp1_30); \
2690 : stp2_31 = _mm_add_epi16(stp1_31, stp1_30); \
2691 : } \
2692 : \
2693 : /* Stage3 */ \
2694 : { \
2695 : const __m128i lo_4_28 = _mm_unpacklo_epi16((in0)[4], (in1)[12]); \
2696 : const __m128i hi_4_28 = _mm_unpackhi_epi16((in0)[4], (in1)[12]); \
2697 : const __m128i lo_20_12 = _mm_unpacklo_epi16((in1)[4], (in0)[12]); \
2698 : const __m128i hi_20_12 = _mm_unpackhi_epi16((in1)[4], (in0)[12]); \
2699 : \
2700 : const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30); \
2701 : const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30); \
2702 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
2703 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
2704 : \
2705 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2706 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2707 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
2708 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
2709 : \
2710 : MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0, \
2711 : stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, \
2712 : stp1_6) \
2713 : \
2714 : stp1_8 = _mm_add_epi16(stp2_8, stp2_9); \
2715 : stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \
2716 : stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
2717 : stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
2718 : stp1_12 = _mm_add_epi16(stp2_12, stp2_13); \
2719 : stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
2720 : stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
2721 : stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
2722 : \
2723 : MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
2724 : stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18, \
2725 : stp1_29) \
2726 : MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
2727 : stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \
2728 : stp1_25) \
2729 : \
2730 : stp1_16 = stp2_16; \
2731 : stp1_31 = stp2_31; \
2732 : stp1_19 = stp2_19; \
2733 : stp1_20 = stp2_20; \
2734 : stp1_23 = stp2_23; \
2735 : stp1_24 = stp2_24; \
2736 : stp1_27 = stp2_27; \
2737 : stp1_28 = stp2_28; \
2738 : } \
2739 : \
2740 : /* Stage4 */ \
2741 : { \
2742 : const __m128i lo_0_16 = _mm_unpacklo_epi16((in0)[0], (in1)[0]); \
2743 : const __m128i hi_0_16 = _mm_unpackhi_epi16((in0)[0], (in1)[0]); \
2744 : const __m128i lo_8_24 = _mm_unpacklo_epi16((in0)[8], (in1)[8]); \
2745 : const __m128i hi_8_24 = _mm_unpackhi_epi16((in0)[8], (in1)[8]); \
2746 : \
2747 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
2748 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
2749 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
2750 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
2751 : \
2752 : MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, stg4_1, \
2753 : stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3) \
2754 : \
2755 : stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
2756 : stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
2757 : stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
2758 : stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
2759 : \
2760 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
2761 : stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10, \
2762 : stp2_13) \
2763 : \
2764 : stp2_8 = stp1_8; \
2765 : stp2_15 = stp1_15; \
2766 : stp2_11 = stp1_11; \
2767 : stp2_12 = stp1_12; \
2768 : \
2769 : stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
2770 : stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
2771 : stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
2772 : stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
2773 : stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
2774 : stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
2775 : stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
2776 : stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
2777 : \
2778 : stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
2779 : stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
2780 : stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
2781 : stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
2782 : stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
2783 : stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
2784 : stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
2785 : stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
2786 : } \
2787 : \
2788 : /* Stage5 */ \
2789 : { \
2790 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
2791 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
2792 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
2793 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
2794 : \
2795 : const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
2796 : const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
2797 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2798 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2799 : \
2800 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2801 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2802 : \
2803 : stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
2804 : stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
2805 : stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
2806 : stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
2807 : \
2808 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
2809 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
2810 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
2811 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
2812 : \
2813 : tmp0 = _mm_add_epi32(tmp0, rounding); \
2814 : tmp1 = _mm_add_epi32(tmp1, rounding); \
2815 : tmp2 = _mm_add_epi32(tmp2, rounding); \
2816 : tmp3 = _mm_add_epi32(tmp3, rounding); \
2817 : \
2818 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
2819 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
2820 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
2821 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
2822 : \
2823 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
2824 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
2825 : \
2826 : stp1_4 = stp2_4; \
2827 : stp1_7 = stp2_7; \
2828 : \
2829 : stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
2830 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
2831 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
2832 : stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
2833 : stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
2834 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
2835 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
2836 : stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
2837 : \
2838 : stp1_16 = stp2_16; \
2839 : stp1_17 = stp2_17; \
2840 : \
2841 : MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
2842 : stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19, \
2843 : stp1_28) \
2844 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
2845 : stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21, \
2846 : stp1_26) \
2847 : \
2848 : stp1_22 = stp2_22; \
2849 : stp1_23 = stp2_23; \
2850 : stp1_24 = stp2_24; \
2851 : stp1_25 = stp2_25; \
2852 : stp1_30 = stp2_30; \
2853 : stp1_31 = stp2_31; \
2854 : } \
2855 : \
2856 : /* Stage6 */ \
2857 : { \
2858 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
2859 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
2860 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
2861 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
2862 : \
2863 : stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
2864 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
2865 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
2866 : stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
2867 : stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
2868 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
2869 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
2870 : stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
2871 : \
2872 : stp2_8 = stp1_8; \
2873 : stp2_9 = stp1_9; \
2874 : stp2_14 = stp1_14; \
2875 : stp2_15 = stp1_15; \
2876 : \
2877 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
2878 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
2879 : stp2_12) \
2880 : \
2881 : stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
2882 : stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
2883 : stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
2884 : stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
2885 : stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
2886 : stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
2887 : stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
2888 : stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
2889 : \
2890 : stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
2891 : stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
2892 : stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
2893 : stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
2894 : stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
2895 : stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
2896 : stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
2897 : stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
2898 : } \
2899 : \
2900 : /* Stage7 */ \
2901 : { \
2902 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2903 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2904 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2905 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2906 : \
2907 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
2908 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
2909 : const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
2910 : const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
2911 : \
2912 : stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
2913 : stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
2914 : stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
2915 : stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
2916 : stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
2917 : stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
2918 : stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
2919 : stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
2920 : stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
2921 : stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
2922 : stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
2923 : stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
2924 : stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
2925 : stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
2926 : stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
2927 : stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
2928 : \
2929 : stp1_16 = stp2_16; \
2930 : stp1_17 = stp2_17; \
2931 : stp1_18 = stp2_18; \
2932 : stp1_19 = stp2_19; \
2933 : \
2934 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
2935 : stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21, \
2936 : stp1_26) \
2937 : MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
2938 : stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23, \
2939 : stp1_24) \
2940 : \
2941 : stp1_28 = stp2_28; \
2942 : stp1_29 = stp2_29; \
2943 : stp1_30 = stp2_30; \
2944 : stp1_31 = stp2_31; \
2945 : }
2946 :
2947 : // Only upper-left 8x8 has non-zero coeff
2948 0 : void aom_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest,
2949 : int stride) {
2950 0 : const __m128i zero = _mm_setzero_si128();
2951 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
2952 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
2953 :
2954 : // idct constants for each stage
2955 0 : const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
2956 0 : const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
2957 0 : const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
2958 0 : const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
2959 0 : const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
2960 0 : const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
2961 0 : const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
2962 0 : const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
2963 :
2964 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
2965 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
2966 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
2967 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
2968 :
2969 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
2970 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
2971 0 : const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
2972 0 : const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
2973 0 : const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
2974 0 : const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
2975 0 : const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
2976 0 : const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
2977 :
2978 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
2979 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
2980 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
2981 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
2982 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
2983 :
2984 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
2985 :
2986 : __m128i in[32], col[32];
2987 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
2988 : stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
2989 : stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
2990 : stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
2991 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
2992 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
2993 : stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
2994 : stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
2995 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
2996 : int i;
2997 :
2998 : // Load input data. Only need to load the top left 8x8 block.
2999 0 : in[0] = load_input_data(input);
3000 0 : in[1] = load_input_data(input + 32);
3001 0 : in[2] = load_input_data(input + 64);
3002 0 : in[3] = load_input_data(input + 96);
3003 0 : in[4] = load_input_data(input + 128);
3004 0 : in[5] = load_input_data(input + 160);
3005 0 : in[6] = load_input_data(input + 192);
3006 0 : in[7] = load_input_data(input + 224);
3007 :
3008 0 : for (i = 8; i < 32; ++i) {
3009 0 : in[i] = _mm_setzero_si128();
3010 : }
3011 :
3012 0 : array_transpose_8x8(in, in);
3013 : // TODO(hkuang): Following transposes are unnecessary. But remove them will
3014 : // lead to performance drop on some devices.
3015 0 : array_transpose_8x8(in + 8, in + 8);
3016 0 : array_transpose_8x8(in + 16, in + 16);
3017 0 : array_transpose_8x8(in + 24, in + 24);
3018 :
3019 0 : IDCT32_34
3020 :
3021 : // 1_D: Store 32 intermediate results for each 8x32 block.
3022 0 : col[0] = _mm_add_epi16(stp1_0, stp1_31);
3023 0 : col[1] = _mm_add_epi16(stp1_1, stp1_30);
3024 0 : col[2] = _mm_add_epi16(stp1_2, stp1_29);
3025 0 : col[3] = _mm_add_epi16(stp1_3, stp1_28);
3026 0 : col[4] = _mm_add_epi16(stp1_4, stp1_27);
3027 0 : col[5] = _mm_add_epi16(stp1_5, stp1_26);
3028 0 : col[6] = _mm_add_epi16(stp1_6, stp1_25);
3029 0 : col[7] = _mm_add_epi16(stp1_7, stp1_24);
3030 0 : col[8] = _mm_add_epi16(stp1_8, stp1_23);
3031 0 : col[9] = _mm_add_epi16(stp1_9, stp1_22);
3032 0 : col[10] = _mm_add_epi16(stp1_10, stp1_21);
3033 0 : col[11] = _mm_add_epi16(stp1_11, stp1_20);
3034 0 : col[12] = _mm_add_epi16(stp1_12, stp1_19);
3035 0 : col[13] = _mm_add_epi16(stp1_13, stp1_18);
3036 0 : col[14] = _mm_add_epi16(stp1_14, stp1_17);
3037 0 : col[15] = _mm_add_epi16(stp1_15, stp1_16);
3038 0 : col[16] = _mm_sub_epi16(stp1_15, stp1_16);
3039 0 : col[17] = _mm_sub_epi16(stp1_14, stp1_17);
3040 0 : col[18] = _mm_sub_epi16(stp1_13, stp1_18);
3041 0 : col[19] = _mm_sub_epi16(stp1_12, stp1_19);
3042 0 : col[20] = _mm_sub_epi16(stp1_11, stp1_20);
3043 0 : col[21] = _mm_sub_epi16(stp1_10, stp1_21);
3044 0 : col[22] = _mm_sub_epi16(stp1_9, stp1_22);
3045 0 : col[23] = _mm_sub_epi16(stp1_8, stp1_23);
3046 0 : col[24] = _mm_sub_epi16(stp1_7, stp1_24);
3047 0 : col[25] = _mm_sub_epi16(stp1_6, stp1_25);
3048 0 : col[26] = _mm_sub_epi16(stp1_5, stp1_26);
3049 0 : col[27] = _mm_sub_epi16(stp1_4, stp1_27);
3050 0 : col[28] = _mm_sub_epi16(stp1_3, stp1_28);
3051 0 : col[29] = _mm_sub_epi16(stp1_2, stp1_29);
3052 0 : col[30] = _mm_sub_epi16(stp1_1, stp1_30);
3053 0 : col[31] = _mm_sub_epi16(stp1_0, stp1_31);
3054 0 : for (i = 0; i < 4; i++) {
3055 : int j;
3056 : // Transpose 32x8 block to 8x32 block
3057 0 : array_transpose_8x8(col + i * 8, in);
3058 0 : IDCT32_34
3059 :
3060 : // 2_D: Calculate the results and store them to destination.
3061 0 : in[0] = _mm_add_epi16(stp1_0, stp1_31);
3062 0 : in[1] = _mm_add_epi16(stp1_1, stp1_30);
3063 0 : in[2] = _mm_add_epi16(stp1_2, stp1_29);
3064 0 : in[3] = _mm_add_epi16(stp1_3, stp1_28);
3065 0 : in[4] = _mm_add_epi16(stp1_4, stp1_27);
3066 0 : in[5] = _mm_add_epi16(stp1_5, stp1_26);
3067 0 : in[6] = _mm_add_epi16(stp1_6, stp1_25);
3068 0 : in[7] = _mm_add_epi16(stp1_7, stp1_24);
3069 0 : in[8] = _mm_add_epi16(stp1_8, stp1_23);
3070 0 : in[9] = _mm_add_epi16(stp1_9, stp1_22);
3071 0 : in[10] = _mm_add_epi16(stp1_10, stp1_21);
3072 0 : in[11] = _mm_add_epi16(stp1_11, stp1_20);
3073 0 : in[12] = _mm_add_epi16(stp1_12, stp1_19);
3074 0 : in[13] = _mm_add_epi16(stp1_13, stp1_18);
3075 0 : in[14] = _mm_add_epi16(stp1_14, stp1_17);
3076 0 : in[15] = _mm_add_epi16(stp1_15, stp1_16);
3077 0 : in[16] = _mm_sub_epi16(stp1_15, stp1_16);
3078 0 : in[17] = _mm_sub_epi16(stp1_14, stp1_17);
3079 0 : in[18] = _mm_sub_epi16(stp1_13, stp1_18);
3080 0 : in[19] = _mm_sub_epi16(stp1_12, stp1_19);
3081 0 : in[20] = _mm_sub_epi16(stp1_11, stp1_20);
3082 0 : in[21] = _mm_sub_epi16(stp1_10, stp1_21);
3083 0 : in[22] = _mm_sub_epi16(stp1_9, stp1_22);
3084 0 : in[23] = _mm_sub_epi16(stp1_8, stp1_23);
3085 0 : in[24] = _mm_sub_epi16(stp1_7, stp1_24);
3086 0 : in[25] = _mm_sub_epi16(stp1_6, stp1_25);
3087 0 : in[26] = _mm_sub_epi16(stp1_5, stp1_26);
3088 0 : in[27] = _mm_sub_epi16(stp1_4, stp1_27);
3089 0 : in[28] = _mm_sub_epi16(stp1_3, stp1_28);
3090 0 : in[29] = _mm_sub_epi16(stp1_2, stp1_29);
3091 0 : in[30] = _mm_sub_epi16(stp1_1, stp1_30);
3092 0 : in[31] = _mm_sub_epi16(stp1_0, stp1_31);
3093 :
3094 0 : for (j = 0; j < 32; ++j) {
3095 : // Final rounding and shift
3096 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
3097 0 : in[j] = _mm_srai_epi16(in[j], 6);
3098 0 : RECON_AND_STORE(dest + j * stride, in[j]);
3099 : }
3100 :
3101 0 : dest += 8;
3102 : }
3103 0 : }
3104 :
3105 0 : void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest,
3106 : int stride) {
3107 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
3108 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
3109 0 : const __m128i zero = _mm_setzero_si128();
3110 :
3111 : // idct constants for each stage
3112 0 : const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
3113 0 : const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
3114 0 : const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
3115 0 : const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
3116 0 : const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
3117 0 : const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
3118 0 : const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
3119 0 : const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
3120 0 : const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
3121 0 : const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
3122 0 : const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
3123 0 : const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
3124 0 : const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
3125 0 : const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
3126 0 : const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
3127 0 : const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
3128 :
3129 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
3130 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
3131 0 : const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
3132 0 : const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
3133 0 : const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
3134 0 : const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
3135 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
3136 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
3137 :
3138 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
3139 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
3140 0 : const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
3141 0 : const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
3142 0 : const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
3143 0 : const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
3144 0 : const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
3145 0 : const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
3146 0 : const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
3147 0 : const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
3148 :
3149 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
3150 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
3151 0 : const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
3152 0 : const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
3153 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
3154 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
3155 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
3156 :
3157 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
3158 :
3159 : __m128i in[32], col[128], zero_idx[16];
3160 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
3161 : stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
3162 : stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
3163 : stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
3164 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
3165 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
3166 : stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
3167 : stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
3168 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
3169 : int i, j, i32;
3170 :
3171 0 : for (i = 0; i < 4; i++) {
3172 0 : i32 = (i << 5);
3173 : // First 1-D idct
3174 : // Load input data.
3175 0 : LOAD_DQCOEFF(in[0], input);
3176 0 : LOAD_DQCOEFF(in[8], input);
3177 0 : LOAD_DQCOEFF(in[16], input);
3178 0 : LOAD_DQCOEFF(in[24], input);
3179 0 : LOAD_DQCOEFF(in[1], input);
3180 0 : LOAD_DQCOEFF(in[9], input);
3181 0 : LOAD_DQCOEFF(in[17], input);
3182 0 : LOAD_DQCOEFF(in[25], input);
3183 0 : LOAD_DQCOEFF(in[2], input);
3184 0 : LOAD_DQCOEFF(in[10], input);
3185 0 : LOAD_DQCOEFF(in[18], input);
3186 0 : LOAD_DQCOEFF(in[26], input);
3187 0 : LOAD_DQCOEFF(in[3], input);
3188 0 : LOAD_DQCOEFF(in[11], input);
3189 0 : LOAD_DQCOEFF(in[19], input);
3190 0 : LOAD_DQCOEFF(in[27], input);
3191 :
3192 0 : LOAD_DQCOEFF(in[4], input);
3193 0 : LOAD_DQCOEFF(in[12], input);
3194 0 : LOAD_DQCOEFF(in[20], input);
3195 0 : LOAD_DQCOEFF(in[28], input);
3196 0 : LOAD_DQCOEFF(in[5], input);
3197 0 : LOAD_DQCOEFF(in[13], input);
3198 0 : LOAD_DQCOEFF(in[21], input);
3199 0 : LOAD_DQCOEFF(in[29], input);
3200 0 : LOAD_DQCOEFF(in[6], input);
3201 0 : LOAD_DQCOEFF(in[14], input);
3202 0 : LOAD_DQCOEFF(in[22], input);
3203 0 : LOAD_DQCOEFF(in[30], input);
3204 0 : LOAD_DQCOEFF(in[7], input);
3205 0 : LOAD_DQCOEFF(in[15], input);
3206 0 : LOAD_DQCOEFF(in[23], input);
3207 0 : LOAD_DQCOEFF(in[31], input);
3208 :
3209 : // checking if all entries are zero
3210 0 : zero_idx[0] = _mm_or_si128(in[0], in[1]);
3211 0 : zero_idx[1] = _mm_or_si128(in[2], in[3]);
3212 0 : zero_idx[2] = _mm_or_si128(in[4], in[5]);
3213 0 : zero_idx[3] = _mm_or_si128(in[6], in[7]);
3214 0 : zero_idx[4] = _mm_or_si128(in[8], in[9]);
3215 0 : zero_idx[5] = _mm_or_si128(in[10], in[11]);
3216 0 : zero_idx[6] = _mm_or_si128(in[12], in[13]);
3217 0 : zero_idx[7] = _mm_or_si128(in[14], in[15]);
3218 0 : zero_idx[8] = _mm_or_si128(in[16], in[17]);
3219 0 : zero_idx[9] = _mm_or_si128(in[18], in[19]);
3220 0 : zero_idx[10] = _mm_or_si128(in[20], in[21]);
3221 0 : zero_idx[11] = _mm_or_si128(in[22], in[23]);
3222 0 : zero_idx[12] = _mm_or_si128(in[24], in[25]);
3223 0 : zero_idx[13] = _mm_or_si128(in[26], in[27]);
3224 0 : zero_idx[14] = _mm_or_si128(in[28], in[29]);
3225 0 : zero_idx[15] = _mm_or_si128(in[30], in[31]);
3226 :
3227 0 : zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]);
3228 0 : zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]);
3229 0 : zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]);
3230 0 : zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]);
3231 0 : zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]);
3232 0 : zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]);
3233 0 : zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]);
3234 0 : zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]);
3235 :
3236 0 : zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]);
3237 0 : zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]);
3238 0 : zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]);
3239 0 : zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]);
3240 0 : zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]);
3241 0 : zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]);
3242 0 : zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]);
3243 :
3244 0 : if (_mm_movemask_epi8(_mm_cmpeq_epi32(zero_idx[14], zero)) == 0xFFFF) {
3245 0 : col[i32 + 0] = _mm_setzero_si128();
3246 0 : col[i32 + 1] = _mm_setzero_si128();
3247 0 : col[i32 + 2] = _mm_setzero_si128();
3248 0 : col[i32 + 3] = _mm_setzero_si128();
3249 0 : col[i32 + 4] = _mm_setzero_si128();
3250 0 : col[i32 + 5] = _mm_setzero_si128();
3251 0 : col[i32 + 6] = _mm_setzero_si128();
3252 0 : col[i32 + 7] = _mm_setzero_si128();
3253 0 : col[i32 + 8] = _mm_setzero_si128();
3254 0 : col[i32 + 9] = _mm_setzero_si128();
3255 0 : col[i32 + 10] = _mm_setzero_si128();
3256 0 : col[i32 + 11] = _mm_setzero_si128();
3257 0 : col[i32 + 12] = _mm_setzero_si128();
3258 0 : col[i32 + 13] = _mm_setzero_si128();
3259 0 : col[i32 + 14] = _mm_setzero_si128();
3260 0 : col[i32 + 15] = _mm_setzero_si128();
3261 0 : col[i32 + 16] = _mm_setzero_si128();
3262 0 : col[i32 + 17] = _mm_setzero_si128();
3263 0 : col[i32 + 18] = _mm_setzero_si128();
3264 0 : col[i32 + 19] = _mm_setzero_si128();
3265 0 : col[i32 + 20] = _mm_setzero_si128();
3266 0 : col[i32 + 21] = _mm_setzero_si128();
3267 0 : col[i32 + 22] = _mm_setzero_si128();
3268 0 : col[i32 + 23] = _mm_setzero_si128();
3269 0 : col[i32 + 24] = _mm_setzero_si128();
3270 0 : col[i32 + 25] = _mm_setzero_si128();
3271 0 : col[i32 + 26] = _mm_setzero_si128();
3272 0 : col[i32 + 27] = _mm_setzero_si128();
3273 0 : col[i32 + 28] = _mm_setzero_si128();
3274 0 : col[i32 + 29] = _mm_setzero_si128();
3275 0 : col[i32 + 30] = _mm_setzero_si128();
3276 0 : col[i32 + 31] = _mm_setzero_si128();
3277 0 : continue;
3278 : }
3279 :
3280 : // Transpose 32x8 block to 8x32 block
3281 0 : array_transpose_8x8(in, in);
3282 0 : array_transpose_8x8(in + 8, in + 8);
3283 0 : array_transpose_8x8(in + 16, in + 16);
3284 0 : array_transpose_8x8(in + 24, in + 24);
3285 :
3286 0 : IDCT32(in, in + 16)
3287 :
3288 : // 1_D: Store 32 intermediate results for each 8x32 block.
3289 0 : col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31);
3290 0 : col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30);
3291 0 : col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29);
3292 0 : col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28);
3293 0 : col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27);
3294 0 : col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26);
3295 0 : col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25);
3296 0 : col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24);
3297 0 : col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23);
3298 0 : col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22);
3299 0 : col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21);
3300 0 : col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20);
3301 0 : col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19);
3302 0 : col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18);
3303 0 : col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17);
3304 0 : col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16);
3305 0 : col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16);
3306 0 : col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17);
3307 0 : col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18);
3308 0 : col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19);
3309 0 : col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20);
3310 0 : col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21);
3311 0 : col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22);
3312 0 : col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23);
3313 0 : col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24);
3314 0 : col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25);
3315 0 : col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26);
3316 0 : col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27);
3317 0 : col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28);
3318 0 : col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29);
3319 0 : col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
3320 0 : col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
3321 : }
3322 0 : for (i = 0; i < 4; i++) {
3323 : // Second 1-D idct
3324 0 : j = i << 3;
3325 :
3326 : // Transpose 32x8 block to 8x32 block
3327 0 : array_transpose_8x8(col + j, in);
3328 0 : array_transpose_8x8(col + j + 32, in + 8);
3329 0 : array_transpose_8x8(col + j + 64, in + 16);
3330 0 : array_transpose_8x8(col + j + 96, in + 24);
3331 :
3332 0 : IDCT32(in, in + 16)
3333 :
3334 : // 2_D: Calculate the results and store them to destination.
3335 0 : in[0] = _mm_add_epi16(stp1_0, stp1_31);
3336 0 : in[1] = _mm_add_epi16(stp1_1, stp1_30);
3337 0 : in[2] = _mm_add_epi16(stp1_2, stp1_29);
3338 0 : in[3] = _mm_add_epi16(stp1_3, stp1_28);
3339 0 : in[4] = _mm_add_epi16(stp1_4, stp1_27);
3340 0 : in[5] = _mm_add_epi16(stp1_5, stp1_26);
3341 0 : in[6] = _mm_add_epi16(stp1_6, stp1_25);
3342 0 : in[7] = _mm_add_epi16(stp1_7, stp1_24);
3343 0 : in[8] = _mm_add_epi16(stp1_8, stp1_23);
3344 0 : in[9] = _mm_add_epi16(stp1_9, stp1_22);
3345 0 : in[10] = _mm_add_epi16(stp1_10, stp1_21);
3346 0 : in[11] = _mm_add_epi16(stp1_11, stp1_20);
3347 0 : in[12] = _mm_add_epi16(stp1_12, stp1_19);
3348 0 : in[13] = _mm_add_epi16(stp1_13, stp1_18);
3349 0 : in[14] = _mm_add_epi16(stp1_14, stp1_17);
3350 0 : in[15] = _mm_add_epi16(stp1_15, stp1_16);
3351 0 : in[16] = _mm_sub_epi16(stp1_15, stp1_16);
3352 0 : in[17] = _mm_sub_epi16(stp1_14, stp1_17);
3353 0 : in[18] = _mm_sub_epi16(stp1_13, stp1_18);
3354 0 : in[19] = _mm_sub_epi16(stp1_12, stp1_19);
3355 0 : in[20] = _mm_sub_epi16(stp1_11, stp1_20);
3356 0 : in[21] = _mm_sub_epi16(stp1_10, stp1_21);
3357 0 : in[22] = _mm_sub_epi16(stp1_9, stp1_22);
3358 0 : in[23] = _mm_sub_epi16(stp1_8, stp1_23);
3359 0 : in[24] = _mm_sub_epi16(stp1_7, stp1_24);
3360 0 : in[25] = _mm_sub_epi16(stp1_6, stp1_25);
3361 0 : in[26] = _mm_sub_epi16(stp1_5, stp1_26);
3362 0 : in[27] = _mm_sub_epi16(stp1_4, stp1_27);
3363 0 : in[28] = _mm_sub_epi16(stp1_3, stp1_28);
3364 0 : in[29] = _mm_sub_epi16(stp1_2, stp1_29);
3365 0 : in[30] = _mm_sub_epi16(stp1_1, stp1_30);
3366 0 : in[31] = _mm_sub_epi16(stp1_0, stp1_31);
3367 :
3368 0 : for (j = 0; j < 32; ++j) {
3369 : // Final rounding and shift
3370 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
3371 0 : in[j] = _mm_srai_epi16(in[j], 6);
3372 0 : RECON_AND_STORE(dest + j * stride, in[j]);
3373 : }
3374 :
3375 0 : dest += 8;
3376 : }
3377 0 : }
3378 :
3379 0 : void aom_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest,
3380 : int stride) {
3381 : __m128i dc_value;
3382 0 : const __m128i zero = _mm_setzero_si128();
3383 : int a, j;
3384 :
3385 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
3386 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
3387 0 : a = ROUND_POWER_OF_TWO(a, 6);
3388 :
3389 0 : if (a == 0) return;
3390 :
3391 0 : dc_value = _mm_set1_epi16(a);
3392 :
3393 0 : for (j = 0; j < 32; ++j) {
3394 0 : RECON_AND_STORE(dest + 0 + j * stride, dc_value);
3395 0 : RECON_AND_STORE(dest + 8 + j * stride, dc_value);
3396 0 : RECON_AND_STORE(dest + 16 + j * stride, dc_value);
3397 0 : RECON_AND_STORE(dest + 24 + j * stride, dc_value);
3398 : }
3399 : }
3400 :
3401 : // Apply a 32-element IDCT to 8 columns. This does not do any transposition
3402 : // of its input - the caller is expected to have done that.
3403 : // The input buffers are the top and bottom halves of an 8x32 block.
3404 0 : void idct32_8col(__m128i *in0, __m128i *in1) {
3405 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
3406 :
3407 : // idct constants for each stage
3408 0 : const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
3409 0 : const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
3410 0 : const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
3411 0 : const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
3412 0 : const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
3413 0 : const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
3414 0 : const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
3415 0 : const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
3416 0 : const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
3417 0 : const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
3418 0 : const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
3419 0 : const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
3420 0 : const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
3421 0 : const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
3422 0 : const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
3423 0 : const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
3424 :
3425 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
3426 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
3427 0 : const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
3428 0 : const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
3429 0 : const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
3430 0 : const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
3431 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
3432 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
3433 :
3434 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
3435 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
3436 0 : const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
3437 0 : const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
3438 0 : const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
3439 0 : const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
3440 0 : const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
3441 0 : const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
3442 0 : const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
3443 0 : const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
3444 :
3445 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
3446 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
3447 0 : const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
3448 0 : const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
3449 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
3450 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
3451 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
3452 :
3453 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
3454 :
3455 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
3456 : stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
3457 : stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
3458 : stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
3459 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
3460 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
3461 : stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
3462 : stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
3463 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
3464 :
3465 0 : IDCT32(in0, in1)
3466 :
3467 : // 2_D: Calculate the results and store them to destination.
3468 0 : in0[0] = _mm_add_epi16(stp1_0, stp1_31);
3469 0 : in0[1] = _mm_add_epi16(stp1_1, stp1_30);
3470 0 : in0[2] = _mm_add_epi16(stp1_2, stp1_29);
3471 0 : in0[3] = _mm_add_epi16(stp1_3, stp1_28);
3472 0 : in0[4] = _mm_add_epi16(stp1_4, stp1_27);
3473 0 : in0[5] = _mm_add_epi16(stp1_5, stp1_26);
3474 0 : in0[6] = _mm_add_epi16(stp1_6, stp1_25);
3475 0 : in0[7] = _mm_add_epi16(stp1_7, stp1_24);
3476 0 : in0[8] = _mm_add_epi16(stp1_8, stp1_23);
3477 0 : in0[9] = _mm_add_epi16(stp1_9, stp1_22);
3478 0 : in0[10] = _mm_add_epi16(stp1_10, stp1_21);
3479 0 : in0[11] = _mm_add_epi16(stp1_11, stp1_20);
3480 0 : in0[12] = _mm_add_epi16(stp1_12, stp1_19);
3481 0 : in0[13] = _mm_add_epi16(stp1_13, stp1_18);
3482 0 : in0[14] = _mm_add_epi16(stp1_14, stp1_17);
3483 0 : in0[15] = _mm_add_epi16(stp1_15, stp1_16);
3484 0 : in1[0] = _mm_sub_epi16(stp1_15, stp1_16);
3485 0 : in1[1] = _mm_sub_epi16(stp1_14, stp1_17);
3486 0 : in1[2] = _mm_sub_epi16(stp1_13, stp1_18);
3487 0 : in1[3] = _mm_sub_epi16(stp1_12, stp1_19);
3488 0 : in1[4] = _mm_sub_epi16(stp1_11, stp1_20);
3489 0 : in1[5] = _mm_sub_epi16(stp1_10, stp1_21);
3490 0 : in1[6] = _mm_sub_epi16(stp1_9, stp1_22);
3491 0 : in1[7] = _mm_sub_epi16(stp1_8, stp1_23);
3492 0 : in1[8] = _mm_sub_epi16(stp1_7, stp1_24);
3493 0 : in1[9] = _mm_sub_epi16(stp1_6, stp1_25);
3494 0 : in1[10] = _mm_sub_epi16(stp1_5, stp1_26);
3495 0 : in1[11] = _mm_sub_epi16(stp1_4, stp1_27);
3496 0 : in1[12] = _mm_sub_epi16(stp1_3, stp1_28);
3497 0 : in1[13] = _mm_sub_epi16(stp1_2, stp1_29);
3498 0 : in1[14] = _mm_sub_epi16(stp1_1, stp1_30);
3499 0 : in1[15] = _mm_sub_epi16(stp1_0, stp1_31);
3500 0 : }
3501 :
3502 : #if CONFIG_HIGHBITDEPTH
3503 0 : static INLINE __m128i clamp_high_sse2(__m128i value, int bd) {
3504 : __m128i ubounded, retval;
3505 0 : const __m128i zero = _mm_set1_epi16(0);
3506 0 : const __m128i one = _mm_set1_epi16(1);
3507 0 : const __m128i max = _mm_subs_epi16(_mm_slli_epi16(one, bd), one);
3508 0 : ubounded = _mm_cmpgt_epi16(value, max);
3509 0 : retval = _mm_andnot_si128(ubounded, value);
3510 0 : ubounded = _mm_and_si128(ubounded, max);
3511 0 : retval = _mm_or_si128(retval, ubounded);
3512 0 : retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero));
3513 0 : return retval;
3514 : }
3515 :
3516 0 : void aom_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8,
3517 : int stride, int bd) {
3518 : tran_low_t out[4 * 4];
3519 0 : tran_low_t *outptr = out;
3520 : int i, j;
3521 : __m128i inptr[4];
3522 : __m128i sign_bits[2];
3523 : __m128i temp_mm, min_input, max_input;
3524 : int test;
3525 0 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
3526 0 : int optimised_cols = 0;
3527 0 : const __m128i zero = _mm_set1_epi16(0);
3528 0 : const __m128i eight = _mm_set1_epi16(8);
3529 0 : const __m128i max = _mm_set1_epi16(12043);
3530 0 : const __m128i min = _mm_set1_epi16(-12043);
3531 : // Load input into __m128i
3532 0 : inptr[0] = _mm_loadu_si128((const __m128i *)input);
3533 0 : inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4));
3534 0 : inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8));
3535 0 : inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12));
3536 :
3537 : // Pack to 16 bits
3538 0 : inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]);
3539 0 : inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]);
3540 :
3541 0 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3542 0 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3543 0 : max_input = _mm_cmpgt_epi16(max_input, max);
3544 0 : min_input = _mm_cmplt_epi16(min_input, min);
3545 0 : temp_mm = _mm_or_si128(max_input, min_input);
3546 0 : test = _mm_movemask_epi8(temp_mm);
3547 :
3548 0 : if (!test) {
3549 : // Do the row transform
3550 0 : aom_idct4_sse2(inptr);
3551 :
3552 : // Check the min & max values
3553 0 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3554 0 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3555 0 : max_input = _mm_cmpgt_epi16(max_input, max);
3556 0 : min_input = _mm_cmplt_epi16(min_input, min);
3557 0 : temp_mm = _mm_or_si128(max_input, min_input);
3558 0 : test = _mm_movemask_epi8(temp_mm);
3559 :
3560 0 : if (test) {
3561 0 : array_transpose_4x4(inptr);
3562 0 : sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero);
3563 0 : sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero);
3564 0 : inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]);
3565 0 : inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]);
3566 0 : inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]);
3567 0 : inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]);
3568 0 : _mm_storeu_si128((__m128i *)outptr, inptr[0]);
3569 0 : _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]);
3570 0 : _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]);
3571 0 : _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]);
3572 : } else {
3573 : // Set to use the optimised transform for the column
3574 0 : optimised_cols = 1;
3575 : }
3576 : } else {
3577 : // Run the un-optimised row transform
3578 0 : for (i = 0; i < 4; ++i) {
3579 0 : aom_highbd_idct4_c(input, outptr, bd);
3580 0 : input += 4;
3581 0 : outptr += 4;
3582 : }
3583 : }
3584 :
3585 0 : if (optimised_cols) {
3586 0 : aom_idct4_sse2(inptr);
3587 :
3588 : // Final round and shift
3589 0 : inptr[0] = _mm_add_epi16(inptr[0], eight);
3590 0 : inptr[1] = _mm_add_epi16(inptr[1], eight);
3591 :
3592 0 : inptr[0] = _mm_srai_epi16(inptr[0], 4);
3593 0 : inptr[1] = _mm_srai_epi16(inptr[1], 4);
3594 :
3595 : // Reconstruction and Store
3596 : {
3597 0 : __m128i d0 = _mm_loadl_epi64((const __m128i *)dest);
3598 0 : __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2));
3599 0 : d0 = _mm_unpacklo_epi64(
3600 0 : d0, _mm_loadl_epi64((const __m128i *)(dest + stride)));
3601 0 : d2 = _mm_unpacklo_epi64(
3602 0 : d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3)));
3603 0 : d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd);
3604 0 : d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd);
3605 : // store input0
3606 : _mm_storel_epi64((__m128i *)dest, d0);
3607 : // store input1
3608 0 : d0 = _mm_srli_si128(d0, 8);
3609 0 : _mm_storel_epi64((__m128i *)(dest + stride), d0);
3610 : // store input2
3611 0 : _mm_storel_epi64((__m128i *)(dest + stride * 2), d2);
3612 : // store input3
3613 0 : d2 = _mm_srli_si128(d2, 8);
3614 0 : _mm_storel_epi64((__m128i *)(dest + stride * 3), d2);
3615 : }
3616 : } else {
3617 : // Run the un-optimised column transform
3618 : tran_low_t temp_in[4], temp_out[4];
3619 : // Columns
3620 0 : for (i = 0; i < 4; ++i) {
3621 0 : for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i];
3622 0 : aom_highbd_idct4_c(temp_in, temp_out, bd);
3623 0 : for (j = 0; j < 4; ++j) {
3624 0 : dest[j * stride + i] = highbd_clip_pixel_add(
3625 0 : dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
3626 : }
3627 : }
3628 : }
3629 0 : }
3630 :
3631 0 : void aom_highbd_idct8x8_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
3632 : int stride, int bd) {
3633 0 : tran_low_t out[8 * 8] = { 0 };
3634 0 : tran_low_t *outptr = out;
3635 : int i, j, test;
3636 : __m128i inptr[8];
3637 : __m128i min_input, max_input, temp1, temp2, sign_bits;
3638 0 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
3639 0 : const __m128i zero = _mm_set1_epi16(0);
3640 0 : const __m128i sixteen = _mm_set1_epi16(16);
3641 0 : const __m128i max = _mm_set1_epi16(6201);
3642 0 : const __m128i min = _mm_set1_epi16(-6201);
3643 0 : int optimised_cols = 0;
3644 :
3645 : // Load input into __m128i & pack to 16 bits
3646 0 : for (i = 0; i < 8; i++) {
3647 0 : temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
3648 0 : temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
3649 0 : inptr[i] = _mm_packs_epi32(temp1, temp2);
3650 : }
3651 :
3652 : // Find the min & max for the row transform
3653 : // only first 4 row has non-zero coefs
3654 0 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3655 0 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3656 0 : for (i = 2; i < 4; i++) {
3657 0 : max_input = _mm_max_epi16(max_input, inptr[i]);
3658 0 : min_input = _mm_min_epi16(min_input, inptr[i]);
3659 : }
3660 0 : max_input = _mm_cmpgt_epi16(max_input, max);
3661 0 : min_input = _mm_cmplt_epi16(min_input, min);
3662 0 : temp1 = _mm_or_si128(max_input, min_input);
3663 0 : test = _mm_movemask_epi8(temp1);
3664 :
3665 0 : if (!test) {
3666 : // Do the row transform
3667 0 : aom_idct8_sse2(inptr);
3668 :
3669 : // Find the min & max for the column transform
3670 : // N.B. Only first 4 cols contain non-zero coeffs
3671 0 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3672 0 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3673 0 : for (i = 2; i < 8; i++) {
3674 0 : max_input = _mm_max_epi16(max_input, inptr[i]);
3675 0 : min_input = _mm_min_epi16(min_input, inptr[i]);
3676 : }
3677 0 : max_input = _mm_cmpgt_epi16(max_input, max);
3678 0 : min_input = _mm_cmplt_epi16(min_input, min);
3679 0 : temp1 = _mm_or_si128(max_input, min_input);
3680 0 : test = _mm_movemask_epi8(temp1);
3681 :
3682 0 : if (test) {
3683 : // Use fact only first 4 rows contain non-zero coeffs
3684 0 : array_transpose_4X8(inptr, inptr);
3685 0 : for (i = 0; i < 4; i++) {
3686 0 : sign_bits = _mm_cmplt_epi16(inptr[i], zero);
3687 0 : temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
3688 0 : temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
3689 0 : _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
3690 0 : _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
3691 : }
3692 : } else {
3693 : // Set to use the optimised transform for the column
3694 0 : optimised_cols = 1;
3695 : }
3696 : } else {
3697 : // Run the un-optimised row transform
3698 0 : for (i = 0; i < 4; ++i) {
3699 0 : aom_highbd_idct8_c(input, outptr, bd);
3700 0 : input += 8;
3701 0 : outptr += 8;
3702 : }
3703 : }
3704 :
3705 0 : if (optimised_cols) {
3706 0 : aom_idct8_sse2(inptr);
3707 :
3708 : // Final round & shift and Reconstruction and Store
3709 : {
3710 : __m128i d[8];
3711 0 : for (i = 0; i < 8; i++) {
3712 0 : inptr[i] = _mm_add_epi16(inptr[i], sixteen);
3713 0 : d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i));
3714 0 : inptr[i] = _mm_srai_epi16(inptr[i], 5);
3715 0 : d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
3716 : // Store
3717 0 : _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]);
3718 : }
3719 : }
3720 : } else {
3721 : // Run the un-optimised column transform
3722 : tran_low_t temp_in[8], temp_out[8];
3723 0 : for (i = 0; i < 8; ++i) {
3724 0 : for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i];
3725 0 : aom_highbd_idct8_c(temp_in, temp_out, bd);
3726 0 : for (j = 0; j < 8; ++j) {
3727 0 : dest[j * stride + i] = highbd_clip_pixel_add(
3728 0 : dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
3729 : }
3730 : }
3731 : }
3732 0 : }
3733 :
3734 : #endif // CONFIG_HIGHBITDEPTH
|
