Line data Source code
1 : /*
2 : * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3 : *
4 : * Use of this source code is governed by a BSD-style license
5 : * that can be found in the LICENSE file in the root of the source
6 : * tree. An additional intellectual property rights grant can be found
7 : * in the file PATENTS. All contributing project authors may
8 : * be found in the AUTHORS file in the root of the source tree.
9 : */
10 :
11 : #include "./vpx_dsp_rtcd.h"
12 : #include "vpx_dsp/x86/inv_txfm_sse2.h"
13 : #include "vpx_dsp/x86/txfm_common_sse2.h"
14 :
15 : #define RECON_AND_STORE4X4(dest, in_x) \
16 : { \
17 : __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \
18 : d0 = _mm_unpacklo_epi8(d0, zero); \
19 : d0 = _mm_add_epi16(in_x, d0); \
20 : d0 = _mm_packus_epi16(d0, d0); \
21 : *(int *)(dest) = _mm_cvtsi128_si32(d0); \
22 : }
23 :
24 0 : void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest,
25 : int stride) {
26 0 : const __m128i zero = _mm_setzero_si128();
27 0 : const __m128i eight = _mm_set1_epi16(8);
28 0 : const __m128i cst = _mm_setr_epi16(
29 0 : (int16_t)cospi_16_64, (int16_t)cospi_16_64, (int16_t)cospi_16_64,
30 0 : (int16_t)-cospi_16_64, (int16_t)cospi_24_64, (int16_t)-cospi_8_64,
31 0 : (int16_t)cospi_8_64, (int16_t)cospi_24_64);
32 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
33 : __m128i input0, input1, input2, input3;
34 :
35 : // Rows
36 0 : input0 = load_input_data(input);
37 0 : input2 = load_input_data(input + 8);
38 :
39 : // Construct i3, i1, i3, i1, i2, i0, i2, i0
40 0 : input0 = _mm_shufflelo_epi16(input0, 0xd8);
41 0 : input0 = _mm_shufflehi_epi16(input0, 0xd8);
42 0 : input2 = _mm_shufflelo_epi16(input2, 0xd8);
43 0 : input2 = _mm_shufflehi_epi16(input2, 0xd8);
44 :
45 0 : input1 = _mm_unpackhi_epi32(input0, input0);
46 0 : input0 = _mm_unpacklo_epi32(input0, input0);
47 0 : input3 = _mm_unpackhi_epi32(input2, input2);
48 0 : input2 = _mm_unpacklo_epi32(input2, input2);
49 :
50 : // Stage 1
51 0 : input0 = _mm_madd_epi16(input0, cst);
52 0 : input1 = _mm_madd_epi16(input1, cst);
53 0 : input2 = _mm_madd_epi16(input2, cst);
54 0 : input3 = _mm_madd_epi16(input3, cst);
55 :
56 0 : input0 = _mm_add_epi32(input0, rounding);
57 0 : input1 = _mm_add_epi32(input1, rounding);
58 0 : input2 = _mm_add_epi32(input2, rounding);
59 0 : input3 = _mm_add_epi32(input3, rounding);
60 :
61 0 : input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
62 0 : input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
63 0 : input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
64 0 : input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
65 :
66 : // Stage 2
67 0 : input0 = _mm_packs_epi32(input0, input1);
68 0 : input1 = _mm_packs_epi32(input2, input3);
69 :
70 : // Transpose
71 0 : input2 = _mm_unpacklo_epi16(input0, input1);
72 0 : input3 = _mm_unpackhi_epi16(input0, input1);
73 0 : input0 = _mm_unpacklo_epi32(input2, input3);
74 0 : input1 = _mm_unpackhi_epi32(input2, input3);
75 :
76 : // Switch column2, column 3, and then, we got:
77 : // input2: column1, column 0; input3: column2, column 3.
78 0 : input1 = _mm_shuffle_epi32(input1, 0x4e);
79 0 : input2 = _mm_add_epi16(input0, input1);
80 0 : input3 = _mm_sub_epi16(input0, input1);
81 :
82 : // Columns
83 : // Construct i3, i1, i3, i1, i2, i0, i2, i0
84 0 : input0 = _mm_unpacklo_epi32(input2, input2);
85 0 : input1 = _mm_unpackhi_epi32(input2, input2);
86 0 : input2 = _mm_unpackhi_epi32(input3, input3);
87 0 : input3 = _mm_unpacklo_epi32(input3, input3);
88 :
89 : // Stage 1
90 0 : input0 = _mm_madd_epi16(input0, cst);
91 0 : input1 = _mm_madd_epi16(input1, cst);
92 0 : input2 = _mm_madd_epi16(input2, cst);
93 0 : input3 = _mm_madd_epi16(input3, cst);
94 :
95 0 : input0 = _mm_add_epi32(input0, rounding);
96 0 : input1 = _mm_add_epi32(input1, rounding);
97 0 : input2 = _mm_add_epi32(input2, rounding);
98 0 : input3 = _mm_add_epi32(input3, rounding);
99 :
100 0 : input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
101 0 : input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
102 0 : input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
103 0 : input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
104 :
105 : // Stage 2
106 0 : input0 = _mm_packs_epi32(input0, input2);
107 0 : input1 = _mm_packs_epi32(input1, input3);
108 :
109 : // Transpose
110 0 : input2 = _mm_unpacklo_epi16(input0, input1);
111 0 : input3 = _mm_unpackhi_epi16(input0, input1);
112 0 : input0 = _mm_unpacklo_epi32(input2, input3);
113 0 : input1 = _mm_unpackhi_epi32(input2, input3);
114 :
115 : // Switch column2, column 3, and then, we got:
116 : // input2: column1, column 0; input3: column2, column 3.
117 0 : input1 = _mm_shuffle_epi32(input1, 0x4e);
118 0 : input2 = _mm_add_epi16(input0, input1);
119 0 : input3 = _mm_sub_epi16(input0, input1);
120 :
121 : // Final round and shift
122 0 : input2 = _mm_add_epi16(input2, eight);
123 0 : input3 = _mm_add_epi16(input3, eight);
124 :
125 0 : input2 = _mm_srai_epi16(input2, 4);
126 0 : input3 = _mm_srai_epi16(input3, 4);
127 :
128 : // Reconstruction and Store
129 : {
130 0 : __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
131 0 : __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
132 0 : d0 = _mm_unpacklo_epi32(d0,
133 0 : _mm_cvtsi32_si128(*(const int *)(dest + stride)));
134 0 : d2 = _mm_unpacklo_epi32(
135 0 : _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)), d2);
136 0 : d0 = _mm_unpacklo_epi8(d0, zero);
137 0 : d2 = _mm_unpacklo_epi8(d2, zero);
138 0 : d0 = _mm_add_epi16(d0, input2);
139 0 : d2 = _mm_add_epi16(d2, input3);
140 0 : d0 = _mm_packus_epi16(d0, d2);
141 : // store input0
142 0 : *(int *)dest = _mm_cvtsi128_si32(d0);
143 : // store input1
144 0 : d0 = _mm_srli_si128(d0, 4);
145 0 : *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
146 : // store input2
147 0 : d0 = _mm_srli_si128(d0, 4);
148 0 : *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
149 : // store input3
150 0 : d0 = _mm_srli_si128(d0, 4);
151 0 : *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
152 : }
153 0 : }
154 :
155 0 : void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest,
156 : int stride) {
157 : __m128i dc_value;
158 0 : const __m128i zero = _mm_setzero_si128();
159 : int a;
160 :
161 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
162 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
163 0 : a = ROUND_POWER_OF_TWO(a, 4);
164 :
165 0 : dc_value = _mm_set1_epi16(a);
166 :
167 0 : RECON_AND_STORE4X4(dest + 0 * stride, dc_value);
168 0 : RECON_AND_STORE4X4(dest + 1 * stride, dc_value);
169 0 : RECON_AND_STORE4X4(dest + 2 * stride, dc_value);
170 0 : RECON_AND_STORE4X4(dest + 3 * stride, dc_value);
171 0 : }
172 :
173 0 : static INLINE void transpose_4x4(__m128i *res) {
174 0 : const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
175 0 : const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
176 :
177 0 : res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1);
178 0 : res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1);
179 0 : }
180 :
181 0 : void idct4_sse2(__m128i *in) {
182 0 : const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
183 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
184 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
185 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
186 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
187 : __m128i u[8], v[8];
188 :
189 0 : transpose_4x4(in);
190 : // stage 1
191 0 : u[0] = _mm_unpacklo_epi16(in[0], in[1]);
192 0 : u[1] = _mm_unpackhi_epi16(in[0], in[1]);
193 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
194 0 : v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
195 0 : v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
196 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
197 :
198 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
199 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
200 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
201 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
202 :
203 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
204 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
205 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
206 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
207 :
208 0 : u[0] = _mm_packs_epi32(v[0], v[1]);
209 0 : u[1] = _mm_packs_epi32(v[3], v[2]);
210 :
211 : // stage 2
212 0 : in[0] = _mm_add_epi16(u[0], u[1]);
213 0 : in[1] = _mm_sub_epi16(u[0], u[1]);
214 0 : in[1] = _mm_shuffle_epi32(in[1], 0x4E);
215 0 : }
216 :
217 0 : void iadst4_sse2(__m128i *in) {
218 0 : const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
219 0 : const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
220 0 : const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
221 0 : const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
222 0 : const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
223 0 : const __m128i kZero = _mm_set1_epi16(0);
224 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
225 : __m128i u[8], v[8], in7;
226 :
227 0 : transpose_4x4(in);
228 0 : in7 = _mm_srli_si128(in[1], 8);
229 0 : in7 = _mm_add_epi16(in7, in[0]);
230 0 : in7 = _mm_sub_epi16(in7, in[1]);
231 :
232 0 : u[0] = _mm_unpacklo_epi16(in[0], in[1]);
233 0 : u[1] = _mm_unpackhi_epi16(in[0], in[1]);
234 0 : u[2] = _mm_unpacklo_epi16(in7, kZero);
235 0 : u[3] = _mm_unpackhi_epi16(in[0], kZero);
236 :
237 0 : v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04); // s0 + s3
238 0 : v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02); // s2 + s5
239 0 : v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x2
240 0 : v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01); // s1 - s4
241 0 : v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04); // s2 - s6
242 0 : v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s2
243 :
244 0 : u[0] = _mm_add_epi32(v[0], v[1]);
245 0 : u[1] = _mm_add_epi32(v[3], v[4]);
246 0 : u[2] = v[2];
247 0 : u[3] = _mm_add_epi32(u[0], u[1]);
248 0 : u[4] = _mm_slli_epi32(v[5], 2);
249 0 : u[5] = _mm_add_epi32(u[3], v[5]);
250 0 : u[6] = _mm_sub_epi32(u[5], u[4]);
251 :
252 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
253 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
254 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
255 0 : v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
256 :
257 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
258 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
259 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
260 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
261 :
262 0 : in[0] = _mm_packs_epi32(u[0], u[1]);
263 0 : in[1] = _mm_packs_epi32(u[2], u[3]);
264 0 : }
265 :
266 : #define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
267 : out2, out3, out4, out5, out6, out7) \
268 : { \
269 : const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
270 : const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
271 : const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \
272 : const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \
273 : const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \
274 : const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \
275 : const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5); \
276 : const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \
277 : \
278 : const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
279 : const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); \
280 : const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
281 : const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); \
282 : const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
283 : const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); \
284 : const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
285 : const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); \
286 : \
287 : out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
288 : out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
289 : out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
290 : out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
291 : out4 = _mm_unpacklo_epi64(tr1_1, tr1_5); \
292 : out5 = _mm_unpackhi_epi64(tr1_1, tr1_5); \
293 : out6 = _mm_unpacklo_epi64(tr1_3, tr1_7); \
294 : out7 = _mm_unpackhi_epi64(tr1_3, tr1_7); \
295 : }
296 :
297 : #define TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, out0, out1, out2, out3) \
298 : { \
299 : const __m128i tr0_0 = _mm_unpackhi_epi16(tmp0, tmp1); \
300 : const __m128i tr0_1 = _mm_unpacklo_epi16(tmp1, tmp0); \
301 : const __m128i tr0_4 = _mm_unpacklo_epi16(tmp2, tmp3); \
302 : const __m128i tr0_5 = _mm_unpackhi_epi16(tmp3, tmp2); \
303 : \
304 : const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
305 : const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
306 : const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
307 : const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
308 : \
309 : out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
310 : out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
311 : out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
312 : out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
313 : }
314 :
315 : #define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \
316 : { \
317 : const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
318 : const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
319 : out0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
320 : out1 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
321 : }
322 :
323 : // Define Macro for multiplying elements by constants and adding them together.
324 : #define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, cst0, cst1, cst2, cst3, \
325 : res0, res1, res2, res3) \
326 : { \
327 : tmp0 = _mm_madd_epi16(lo_0, cst0); \
328 : tmp1 = _mm_madd_epi16(hi_0, cst0); \
329 : tmp2 = _mm_madd_epi16(lo_0, cst1); \
330 : tmp3 = _mm_madd_epi16(hi_0, cst1); \
331 : tmp4 = _mm_madd_epi16(lo_1, cst2); \
332 : tmp5 = _mm_madd_epi16(hi_1, cst2); \
333 : tmp6 = _mm_madd_epi16(lo_1, cst3); \
334 : tmp7 = _mm_madd_epi16(hi_1, cst3); \
335 : \
336 : tmp0 = _mm_add_epi32(tmp0, rounding); \
337 : tmp1 = _mm_add_epi32(tmp1, rounding); \
338 : tmp2 = _mm_add_epi32(tmp2, rounding); \
339 : tmp3 = _mm_add_epi32(tmp3, rounding); \
340 : tmp4 = _mm_add_epi32(tmp4, rounding); \
341 : tmp5 = _mm_add_epi32(tmp5, rounding); \
342 : tmp6 = _mm_add_epi32(tmp6, rounding); \
343 : tmp7 = _mm_add_epi32(tmp7, rounding); \
344 : \
345 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
346 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
347 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
348 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
349 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); \
350 : tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); \
351 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); \
352 : tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); \
353 : \
354 : res0 = _mm_packs_epi32(tmp0, tmp1); \
355 : res1 = _mm_packs_epi32(tmp2, tmp3); \
356 : res2 = _mm_packs_epi32(tmp4, tmp5); \
357 : res3 = _mm_packs_epi32(tmp6, tmp7); \
358 : }
359 :
360 : #define MULTIPLICATION_AND_ADD_2(lo_0, hi_0, cst0, cst1, res0, res1) \
361 : { \
362 : tmp0 = _mm_madd_epi16(lo_0, cst0); \
363 : tmp1 = _mm_madd_epi16(hi_0, cst0); \
364 : tmp2 = _mm_madd_epi16(lo_0, cst1); \
365 : tmp3 = _mm_madd_epi16(hi_0, cst1); \
366 : \
367 : tmp0 = _mm_add_epi32(tmp0, rounding); \
368 : tmp1 = _mm_add_epi32(tmp1, rounding); \
369 : tmp2 = _mm_add_epi32(tmp2, rounding); \
370 : tmp3 = _mm_add_epi32(tmp3, rounding); \
371 : \
372 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
373 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
374 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
375 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
376 : \
377 : res0 = _mm_packs_epi32(tmp0, tmp1); \
378 : res1 = _mm_packs_epi32(tmp2, tmp3); \
379 : }
380 :
381 : #define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3, \
382 : out4, out5, out6, out7) \
383 : { \
384 : /* Stage1 */ \
385 : { \
386 : const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \
387 : const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \
388 : const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); \
389 : const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); \
390 : \
391 : MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, stg1_1, \
392 : stg1_2, stg1_3, stp1_4, stp1_7, stp1_5, stp1_6) \
393 : } \
394 : \
395 : /* Stage2 */ \
396 : { \
397 : const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4); \
398 : const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4); \
399 : const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); \
400 : const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); \
401 : \
402 : MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, stg2_1, \
403 : stg2_2, stg2_3, stp2_0, stp2_1, stp2_2, stp2_3) \
404 : \
405 : stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
406 : stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
407 : stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
408 : stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
409 : } \
410 : \
411 : /* Stage3 */ \
412 : { \
413 : const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
414 : const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
415 : \
416 : stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
417 : stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
418 : stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
419 : stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
420 : \
421 : tmp0 = _mm_madd_epi16(lo_56, stg2_1); \
422 : tmp1 = _mm_madd_epi16(hi_56, stg2_1); \
423 : tmp2 = _mm_madd_epi16(lo_56, stg2_0); \
424 : tmp3 = _mm_madd_epi16(hi_56, stg2_0); \
425 : \
426 : tmp0 = _mm_add_epi32(tmp0, rounding); \
427 : tmp1 = _mm_add_epi32(tmp1, rounding); \
428 : tmp2 = _mm_add_epi32(tmp2, rounding); \
429 : tmp3 = _mm_add_epi32(tmp3, rounding); \
430 : \
431 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
432 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
433 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
434 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
435 : \
436 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
437 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
438 : } \
439 : \
440 : /* Stage4 */ \
441 : out0 = _mm_add_epi16(stp1_0, stp2_7); \
442 : out1 = _mm_add_epi16(stp1_1, stp1_6); \
443 : out2 = _mm_add_epi16(stp1_2, stp1_5); \
444 : out3 = _mm_add_epi16(stp1_3, stp2_4); \
445 : out4 = _mm_sub_epi16(stp1_3, stp2_4); \
446 : out5 = _mm_sub_epi16(stp1_2, stp1_5); \
447 : out6 = _mm_sub_epi16(stp1_1, stp1_6); \
448 : out7 = _mm_sub_epi16(stp1_0, stp2_7); \
449 : }
450 :
451 0 : void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest,
452 : int stride) {
453 0 : const __m128i zero = _mm_setzero_si128();
454 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
455 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 4);
456 0 : const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
457 0 : const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
458 0 : const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
459 0 : const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
460 0 : const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
461 0 : const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
462 0 : const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
463 0 : const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
464 :
465 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
466 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
467 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
468 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
469 : int i;
470 :
471 : // Load input data.
472 0 : in0 = load_input_data(input);
473 0 : in1 = load_input_data(input + 8 * 1);
474 0 : in2 = load_input_data(input + 8 * 2);
475 0 : in3 = load_input_data(input + 8 * 3);
476 0 : in4 = load_input_data(input + 8 * 4);
477 0 : in5 = load_input_data(input + 8 * 5);
478 0 : in6 = load_input_data(input + 8 * 6);
479 0 : in7 = load_input_data(input + 8 * 7);
480 :
481 : // 2-D
482 0 : for (i = 0; i < 2; i++) {
483 : // 8x8 Transpose is copied from vpx_fdct8x8_sse2()
484 0 : TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
485 : in4, in5, in6, in7);
486 :
487 : // 4-stage 1D idct8x8
488 0 : IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4, in5,
489 : in6, in7);
490 : }
491 :
492 : // Final rounding and shift
493 0 : in0 = _mm_adds_epi16(in0, final_rounding);
494 0 : in1 = _mm_adds_epi16(in1, final_rounding);
495 0 : in2 = _mm_adds_epi16(in2, final_rounding);
496 0 : in3 = _mm_adds_epi16(in3, final_rounding);
497 0 : in4 = _mm_adds_epi16(in4, final_rounding);
498 0 : in5 = _mm_adds_epi16(in5, final_rounding);
499 0 : in6 = _mm_adds_epi16(in6, final_rounding);
500 0 : in7 = _mm_adds_epi16(in7, final_rounding);
501 :
502 0 : in0 = _mm_srai_epi16(in0, 5);
503 0 : in1 = _mm_srai_epi16(in1, 5);
504 0 : in2 = _mm_srai_epi16(in2, 5);
505 0 : in3 = _mm_srai_epi16(in3, 5);
506 0 : in4 = _mm_srai_epi16(in4, 5);
507 0 : in5 = _mm_srai_epi16(in5, 5);
508 0 : in6 = _mm_srai_epi16(in6, 5);
509 0 : in7 = _mm_srai_epi16(in7, 5);
510 :
511 0 : RECON_AND_STORE(dest + 0 * stride, in0);
512 0 : RECON_AND_STORE(dest + 1 * stride, in1);
513 0 : RECON_AND_STORE(dest + 2 * stride, in2);
514 0 : RECON_AND_STORE(dest + 3 * stride, in3);
515 0 : RECON_AND_STORE(dest + 4 * stride, in4);
516 0 : RECON_AND_STORE(dest + 5 * stride, in5);
517 0 : RECON_AND_STORE(dest + 6 * stride, in6);
518 0 : RECON_AND_STORE(dest + 7 * stride, in7);
519 0 : }
520 :
521 0 : void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest,
522 : int stride) {
523 : __m128i dc_value;
524 0 : const __m128i zero = _mm_setzero_si128();
525 : int a;
526 :
527 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
528 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
529 0 : a = ROUND_POWER_OF_TWO(a, 5);
530 :
531 0 : dc_value = _mm_set1_epi16(a);
532 :
533 0 : RECON_AND_STORE(dest + 0 * stride, dc_value);
534 0 : RECON_AND_STORE(dest + 1 * stride, dc_value);
535 0 : RECON_AND_STORE(dest + 2 * stride, dc_value);
536 0 : RECON_AND_STORE(dest + 3 * stride, dc_value);
537 0 : RECON_AND_STORE(dest + 4 * stride, dc_value);
538 0 : RECON_AND_STORE(dest + 5 * stride, dc_value);
539 0 : RECON_AND_STORE(dest + 6 * stride, dc_value);
540 0 : RECON_AND_STORE(dest + 7 * stride, dc_value);
541 0 : }
542 :
543 0 : void idct8_sse2(__m128i *in) {
544 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
545 0 : const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
546 0 : const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
547 0 : const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
548 0 : const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
549 0 : const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
550 0 : const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
551 0 : const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
552 0 : const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
553 :
554 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
555 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
556 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
557 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
558 :
559 : // 8x8 Transpose is copied from vpx_fdct8x8_sse2()
560 0 : TRANSPOSE_8X8(in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7], in0,
561 : in1, in2, in3, in4, in5, in6, in7);
562 :
563 : // 4-stage 1D idct8x8
564 0 : IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in[0], in[1], in[2], in[3],
565 : in[4], in[5], in[6], in[7]);
566 0 : }
567 :
568 0 : void iadst8_sse2(__m128i *in) {
569 0 : const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
570 0 : const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
571 0 : const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
572 0 : const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
573 0 : const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
574 0 : const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
575 0 : const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
576 0 : const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
577 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
578 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
579 0 : const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
580 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
581 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
582 0 : const __m128i k__const_0 = _mm_set1_epi16(0);
583 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
584 :
585 : __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
586 : __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
587 : __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
588 : __m128i s0, s1, s2, s3, s4, s5, s6, s7;
589 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
590 :
591 : // transpose
592 0 : array_transpose_8x8(in, in);
593 :
594 : // properly aligned for butterfly input
595 0 : in0 = in[7];
596 0 : in1 = in[0];
597 0 : in2 = in[5];
598 0 : in3 = in[2];
599 0 : in4 = in[3];
600 0 : in5 = in[4];
601 0 : in6 = in[1];
602 0 : in7 = in[6];
603 :
604 : // column transformation
605 : // stage 1
606 : // interleave and multiply/add into 32-bit integer
607 0 : s0 = _mm_unpacklo_epi16(in0, in1);
608 0 : s1 = _mm_unpackhi_epi16(in0, in1);
609 0 : s2 = _mm_unpacklo_epi16(in2, in3);
610 0 : s3 = _mm_unpackhi_epi16(in2, in3);
611 0 : s4 = _mm_unpacklo_epi16(in4, in5);
612 0 : s5 = _mm_unpackhi_epi16(in4, in5);
613 0 : s6 = _mm_unpacklo_epi16(in6, in7);
614 0 : s7 = _mm_unpackhi_epi16(in6, in7);
615 :
616 0 : u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
617 0 : u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
618 0 : u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
619 0 : u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
620 0 : u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
621 0 : u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
622 0 : u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
623 0 : u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
624 0 : u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
625 0 : u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
626 0 : u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
627 0 : u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
628 0 : u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
629 0 : u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
630 0 : u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
631 0 : u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
632 :
633 : // addition
634 0 : w0 = _mm_add_epi32(u0, u8);
635 0 : w1 = _mm_add_epi32(u1, u9);
636 0 : w2 = _mm_add_epi32(u2, u10);
637 0 : w3 = _mm_add_epi32(u3, u11);
638 0 : w4 = _mm_add_epi32(u4, u12);
639 0 : w5 = _mm_add_epi32(u5, u13);
640 0 : w6 = _mm_add_epi32(u6, u14);
641 0 : w7 = _mm_add_epi32(u7, u15);
642 0 : w8 = _mm_sub_epi32(u0, u8);
643 0 : w9 = _mm_sub_epi32(u1, u9);
644 0 : w10 = _mm_sub_epi32(u2, u10);
645 0 : w11 = _mm_sub_epi32(u3, u11);
646 0 : w12 = _mm_sub_epi32(u4, u12);
647 0 : w13 = _mm_sub_epi32(u5, u13);
648 0 : w14 = _mm_sub_epi32(u6, u14);
649 0 : w15 = _mm_sub_epi32(u7, u15);
650 :
651 : // shift and rounding
652 0 : v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
653 0 : v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
654 0 : v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
655 0 : v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
656 0 : v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
657 0 : v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
658 0 : v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
659 0 : v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
660 0 : v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
661 0 : v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
662 0 : v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
663 0 : v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
664 0 : v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
665 0 : v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
666 0 : v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
667 0 : v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
668 :
669 0 : u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
670 0 : u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
671 0 : u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
672 0 : u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
673 0 : u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
674 0 : u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
675 0 : u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
676 0 : u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
677 0 : u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
678 0 : u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
679 0 : u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
680 0 : u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
681 0 : u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
682 0 : u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
683 0 : u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
684 0 : u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
685 :
686 : // back to 16-bit and pack 8 integers into __m128i
687 0 : in[0] = _mm_packs_epi32(u0, u1);
688 0 : in[1] = _mm_packs_epi32(u2, u3);
689 0 : in[2] = _mm_packs_epi32(u4, u5);
690 0 : in[3] = _mm_packs_epi32(u6, u7);
691 0 : in[4] = _mm_packs_epi32(u8, u9);
692 0 : in[5] = _mm_packs_epi32(u10, u11);
693 0 : in[6] = _mm_packs_epi32(u12, u13);
694 0 : in[7] = _mm_packs_epi32(u14, u15);
695 :
696 : // stage 2
697 0 : s0 = _mm_add_epi16(in[0], in[2]);
698 0 : s1 = _mm_add_epi16(in[1], in[3]);
699 0 : s2 = _mm_sub_epi16(in[0], in[2]);
700 0 : s3 = _mm_sub_epi16(in[1], in[3]);
701 0 : u0 = _mm_unpacklo_epi16(in[4], in[5]);
702 0 : u1 = _mm_unpackhi_epi16(in[4], in[5]);
703 0 : u2 = _mm_unpacklo_epi16(in[6], in[7]);
704 0 : u3 = _mm_unpackhi_epi16(in[6], in[7]);
705 :
706 0 : v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
707 0 : v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
708 0 : v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
709 0 : v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
710 0 : v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
711 0 : v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
712 0 : v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
713 0 : v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
714 :
715 0 : w0 = _mm_add_epi32(v0, v4);
716 0 : w1 = _mm_add_epi32(v1, v5);
717 0 : w2 = _mm_add_epi32(v2, v6);
718 0 : w3 = _mm_add_epi32(v3, v7);
719 0 : w4 = _mm_sub_epi32(v0, v4);
720 0 : w5 = _mm_sub_epi32(v1, v5);
721 0 : w6 = _mm_sub_epi32(v2, v6);
722 0 : w7 = _mm_sub_epi32(v3, v7);
723 :
724 0 : v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
725 0 : v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
726 0 : v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
727 0 : v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
728 0 : v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
729 0 : v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
730 0 : v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
731 0 : v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
732 :
733 0 : u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
734 0 : u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
735 0 : u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
736 0 : u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
737 0 : u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
738 0 : u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
739 0 : u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
740 0 : u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
741 :
742 : // back to 16-bit intergers
743 0 : s4 = _mm_packs_epi32(u0, u1);
744 0 : s5 = _mm_packs_epi32(u2, u3);
745 0 : s6 = _mm_packs_epi32(u4, u5);
746 0 : s7 = _mm_packs_epi32(u6, u7);
747 :
748 : // stage 3
749 0 : u0 = _mm_unpacklo_epi16(s2, s3);
750 0 : u1 = _mm_unpackhi_epi16(s2, s3);
751 0 : u2 = _mm_unpacklo_epi16(s6, s7);
752 0 : u3 = _mm_unpackhi_epi16(s6, s7);
753 :
754 0 : v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
755 0 : v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
756 0 : v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
757 0 : v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
758 0 : v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
759 0 : v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
760 0 : v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
761 0 : v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
762 :
763 0 : u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
764 0 : u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
765 0 : u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
766 0 : u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
767 0 : u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
768 0 : u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
769 0 : u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
770 0 : u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
771 :
772 0 : v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
773 0 : v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
774 0 : v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
775 0 : v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
776 0 : v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
777 0 : v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
778 0 : v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
779 0 : v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
780 :
781 0 : s2 = _mm_packs_epi32(v0, v1);
782 0 : s3 = _mm_packs_epi32(v2, v3);
783 0 : s6 = _mm_packs_epi32(v4, v5);
784 0 : s7 = _mm_packs_epi32(v6, v7);
785 :
786 0 : in[0] = s0;
787 0 : in[1] = _mm_sub_epi16(k__const_0, s4);
788 0 : in[2] = s6;
789 0 : in[3] = _mm_sub_epi16(k__const_0, s2);
790 0 : in[4] = s3;
791 0 : in[5] = _mm_sub_epi16(k__const_0, s7);
792 0 : in[6] = s5;
793 0 : in[7] = _mm_sub_epi16(k__const_0, s1);
794 0 : }
795 :
796 0 : void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest,
797 : int stride) {
798 0 : const __m128i zero = _mm_setzero_si128();
799 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
800 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 4);
801 0 : const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
802 0 : const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
803 0 : const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
804 0 : const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
805 0 : const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
806 0 : const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
807 0 : const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
808 0 : const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
809 0 : const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
810 :
811 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
812 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
813 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
814 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
815 :
816 : // Rows. Load 4-row input data.
817 0 : in0 = load_input_data(input);
818 0 : in1 = load_input_data(input + 8 * 1);
819 0 : in2 = load_input_data(input + 8 * 2);
820 0 : in3 = load_input_data(input + 8 * 3);
821 :
822 : // 8x4 Transpose
823 0 : TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1);
824 : // Stage1
825 : {
826 0 : const __m128i lo_17 = _mm_unpackhi_epi16(in0, zero);
827 0 : const __m128i lo_35 = _mm_unpackhi_epi16(in1, zero);
828 :
829 0 : tmp0 = _mm_madd_epi16(lo_17, stg1_0);
830 0 : tmp2 = _mm_madd_epi16(lo_17, stg1_1);
831 0 : tmp4 = _mm_madd_epi16(lo_35, stg1_2);
832 0 : tmp6 = _mm_madd_epi16(lo_35, stg1_3);
833 :
834 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
835 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
836 0 : tmp4 = _mm_add_epi32(tmp4, rounding);
837 0 : tmp6 = _mm_add_epi32(tmp6, rounding);
838 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
839 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
840 0 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
841 0 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
842 :
843 0 : stp1_4 = _mm_packs_epi32(tmp0, tmp2);
844 0 : stp1_5 = _mm_packs_epi32(tmp4, tmp6);
845 : }
846 :
847 : // Stage2
848 : {
849 0 : const __m128i lo_04 = _mm_unpacklo_epi16(in0, zero);
850 0 : const __m128i lo_26 = _mm_unpacklo_epi16(in1, zero);
851 :
852 0 : tmp0 = _mm_madd_epi16(lo_04, stg2_0);
853 0 : tmp2 = _mm_madd_epi16(lo_04, stg2_1);
854 0 : tmp4 = _mm_madd_epi16(lo_26, stg2_2);
855 0 : tmp6 = _mm_madd_epi16(lo_26, stg2_3);
856 :
857 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
858 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
859 0 : tmp4 = _mm_add_epi32(tmp4, rounding);
860 0 : tmp6 = _mm_add_epi32(tmp6, rounding);
861 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
862 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
863 0 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
864 0 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
865 :
866 0 : stp2_0 = _mm_packs_epi32(tmp0, tmp2);
867 0 : stp2_2 = _mm_packs_epi32(tmp6, tmp4);
868 :
869 0 : tmp0 = _mm_add_epi16(stp1_4, stp1_5);
870 0 : tmp1 = _mm_sub_epi16(stp1_4, stp1_5);
871 :
872 0 : stp2_4 = tmp0;
873 0 : stp2_5 = _mm_unpacklo_epi64(tmp1, zero);
874 0 : stp2_6 = _mm_unpackhi_epi64(tmp1, zero);
875 : }
876 :
877 : // Stage3
878 : {
879 0 : const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6);
880 :
881 0 : tmp4 = _mm_add_epi16(stp2_0, stp2_2);
882 0 : tmp6 = _mm_sub_epi16(stp2_0, stp2_2);
883 :
884 0 : stp1_2 = _mm_unpackhi_epi64(tmp6, tmp4);
885 0 : stp1_3 = _mm_unpacklo_epi64(tmp6, tmp4);
886 :
887 0 : tmp0 = _mm_madd_epi16(lo_56, stg3_0);
888 0 : tmp2 = _mm_madd_epi16(lo_56, stg2_0); // stg3_1 = stg2_0
889 :
890 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
891 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
892 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
893 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
894 :
895 0 : stp1_5 = _mm_packs_epi32(tmp0, tmp2);
896 : }
897 :
898 : // Stage4
899 0 : tmp0 = _mm_add_epi16(stp1_3, stp2_4);
900 0 : tmp1 = _mm_add_epi16(stp1_2, stp1_5);
901 0 : tmp2 = _mm_sub_epi16(stp1_3, stp2_4);
902 0 : tmp3 = _mm_sub_epi16(stp1_2, stp1_5);
903 :
904 0 : TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3)
905 :
906 0 : IDCT8(in0, in1, in2, in3, zero, zero, zero, zero, in0, in1, in2, in3, in4,
907 : in5, in6, in7);
908 : // Final rounding and shift
909 0 : in0 = _mm_adds_epi16(in0, final_rounding);
910 0 : in1 = _mm_adds_epi16(in1, final_rounding);
911 0 : in2 = _mm_adds_epi16(in2, final_rounding);
912 0 : in3 = _mm_adds_epi16(in3, final_rounding);
913 0 : in4 = _mm_adds_epi16(in4, final_rounding);
914 0 : in5 = _mm_adds_epi16(in5, final_rounding);
915 0 : in6 = _mm_adds_epi16(in6, final_rounding);
916 0 : in7 = _mm_adds_epi16(in7, final_rounding);
917 :
918 0 : in0 = _mm_srai_epi16(in0, 5);
919 0 : in1 = _mm_srai_epi16(in1, 5);
920 0 : in2 = _mm_srai_epi16(in2, 5);
921 0 : in3 = _mm_srai_epi16(in3, 5);
922 0 : in4 = _mm_srai_epi16(in4, 5);
923 0 : in5 = _mm_srai_epi16(in5, 5);
924 0 : in6 = _mm_srai_epi16(in6, 5);
925 0 : in7 = _mm_srai_epi16(in7, 5);
926 :
927 0 : RECON_AND_STORE(dest + 0 * stride, in0);
928 0 : RECON_AND_STORE(dest + 1 * stride, in1);
929 0 : RECON_AND_STORE(dest + 2 * stride, in2);
930 0 : RECON_AND_STORE(dest + 3 * stride, in3);
931 0 : RECON_AND_STORE(dest + 4 * stride, in4);
932 0 : RECON_AND_STORE(dest + 5 * stride, in5);
933 0 : RECON_AND_STORE(dest + 6 * stride, in6);
934 0 : RECON_AND_STORE(dest + 7 * stride, in7);
935 0 : }
936 :
937 : #define IDCT16 \
938 : /* Stage2 */ \
939 : { \
940 : const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], in[15]); \
941 : const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], in[15]); \
942 : const __m128i lo_9_7 = _mm_unpacklo_epi16(in[9], in[7]); \
943 : const __m128i hi_9_7 = _mm_unpackhi_epi16(in[9], in[7]); \
944 : const __m128i lo_5_11 = _mm_unpacklo_epi16(in[5], in[11]); \
945 : const __m128i hi_5_11 = _mm_unpackhi_epi16(in[5], in[11]); \
946 : const __m128i lo_13_3 = _mm_unpacklo_epi16(in[13], in[3]); \
947 : const __m128i hi_13_3 = _mm_unpackhi_epi16(in[13], in[3]); \
948 : \
949 : MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, stg2_0, stg2_1, \
950 : stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, stp2_14) \
951 : \
952 : MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, stg2_4, stg2_5, \
953 : stg2_6, stg2_7, stp2_10, stp2_13, stp2_11, stp2_12) \
954 : } \
955 : \
956 : /* Stage3 */ \
957 : { \
958 : const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], in[14]); \
959 : const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], in[14]); \
960 : const __m128i lo_10_6 = _mm_unpacklo_epi16(in[10], in[6]); \
961 : const __m128i hi_10_6 = _mm_unpackhi_epi16(in[10], in[6]); \
962 : \
963 : MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, stg3_0, stg3_1, \
964 : stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, stp1_6) \
965 : \
966 : stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9); \
967 : stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \
968 : stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
969 : stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
970 : \
971 : stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); \
972 : stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
973 : stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
974 : stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
975 : } \
976 : \
977 : /* Stage4 */ \
978 : { \
979 : const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], in[8]); \
980 : const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], in[8]); \
981 : const __m128i lo_4_12 = _mm_unpacklo_epi16(in[4], in[12]); \
982 : const __m128i hi_4_12 = _mm_unpackhi_epi16(in[4], in[12]); \
983 : \
984 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
985 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
986 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
987 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
988 : \
989 : MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, stg4_0, stg4_1, \
990 : stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3) \
991 : \
992 : stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
993 : stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
994 : stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
995 : stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
996 : \
997 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
998 : stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10, \
999 : stp2_13) \
1000 : } \
1001 : \
1002 : /* Stage5 */ \
1003 : { \
1004 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
1005 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
1006 : \
1007 : stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
1008 : stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
1009 : stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
1010 : stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
1011 : \
1012 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
1013 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
1014 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
1015 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
1016 : \
1017 : tmp0 = _mm_add_epi32(tmp0, rounding); \
1018 : tmp1 = _mm_add_epi32(tmp1, rounding); \
1019 : tmp2 = _mm_add_epi32(tmp2, rounding); \
1020 : tmp3 = _mm_add_epi32(tmp3, rounding); \
1021 : \
1022 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
1023 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
1024 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
1025 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
1026 : \
1027 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
1028 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
1029 : \
1030 : stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); \
1031 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
1032 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
1033 : stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
1034 : \
1035 : stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
1036 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
1037 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
1038 : stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
1039 : } \
1040 : \
1041 : /* Stage6 */ \
1042 : { \
1043 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
1044 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
1045 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
1046 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
1047 : \
1048 : stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
1049 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
1050 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
1051 : stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
1052 : stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
1053 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
1054 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
1055 : stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
1056 : \
1057 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
1058 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
1059 : stp2_12) \
1060 : }
1061 :
1062 : #define IDCT16_10 \
1063 : /* Stage2 */ \
1064 : { \
1065 : const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], zero); \
1066 : const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], zero); \
1067 : const __m128i lo_13_3 = _mm_unpacklo_epi16(zero, in[3]); \
1068 : const __m128i hi_13_3 = _mm_unpackhi_epi16(zero, in[3]); \
1069 : \
1070 : MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_13_3, hi_13_3, stg2_0, stg2_1, \
1071 : stg2_6, stg2_7, stp1_8_0, stp1_15, stp1_11, \
1072 : stp1_12_0) \
1073 : } \
1074 : \
1075 : /* Stage3 */ \
1076 : { \
1077 : const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], zero); \
1078 : const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], zero); \
1079 : \
1080 : MULTIPLICATION_AND_ADD_2(lo_2_14, hi_2_14, stg3_0, stg3_1, stp2_4, stp2_7) \
1081 : \
1082 : stp1_9 = stp1_8_0; \
1083 : stp1_10 = stp1_11; \
1084 : \
1085 : stp1_13 = stp1_12_0; \
1086 : stp1_14 = stp1_15; \
1087 : } \
1088 : \
1089 : /* Stage4 */ \
1090 : { \
1091 : const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero); \
1092 : const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], zero); \
1093 : \
1094 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
1095 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
1096 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
1097 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
1098 : \
1099 : MULTIPLICATION_AND_ADD_2(lo_0_8, hi_0_8, stg4_0, stg4_1, stp1_0, stp1_1) \
1100 : stp2_5 = stp2_4; \
1101 : stp2_6 = stp2_7; \
1102 : \
1103 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
1104 : stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10, \
1105 : stp2_13) \
1106 : } \
1107 : \
1108 : /* Stage5 */ \
1109 : { \
1110 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
1111 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
1112 : \
1113 : stp1_2 = stp1_1; \
1114 : stp1_3 = stp1_0; \
1115 : \
1116 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
1117 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
1118 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
1119 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
1120 : \
1121 : tmp0 = _mm_add_epi32(tmp0, rounding); \
1122 : tmp1 = _mm_add_epi32(tmp1, rounding); \
1123 : tmp2 = _mm_add_epi32(tmp2, rounding); \
1124 : tmp3 = _mm_add_epi32(tmp3, rounding); \
1125 : \
1126 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
1127 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
1128 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
1129 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
1130 : \
1131 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
1132 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
1133 : \
1134 : stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); \
1135 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
1136 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
1137 : stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
1138 : \
1139 : stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
1140 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
1141 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
1142 : stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
1143 : } \
1144 : \
1145 : /* Stage6 */ \
1146 : { \
1147 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
1148 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
1149 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
1150 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
1151 : \
1152 : stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
1153 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
1154 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
1155 : stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
1156 : stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
1157 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
1158 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
1159 : stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
1160 : \
1161 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
1162 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
1163 : stp2_12) \
1164 : }
1165 :
1166 0 : void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
1167 : int stride) {
1168 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
1169 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
1170 0 : const __m128i zero = _mm_setzero_si128();
1171 :
1172 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
1173 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
1174 0 : const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
1175 0 : const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
1176 0 : const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
1177 0 : const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
1178 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
1179 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
1180 :
1181 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
1182 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
1183 0 : const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
1184 0 : const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
1185 :
1186 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
1187 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1188 0 : const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
1189 0 : const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
1190 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
1191 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
1192 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
1193 0 : const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
1194 :
1195 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1196 :
1197 : __m128i in[16], l[16], r[16], *curr1;
1198 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
1199 : stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
1200 : stp1_8_0, stp1_12_0;
1201 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
1202 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15;
1203 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
1204 : int i;
1205 :
1206 0 : curr1 = l;
1207 0 : for (i = 0; i < 2; i++) {
1208 : // 1-D idct
1209 :
1210 : // Load input data.
1211 0 : in[0] = load_input_data(input);
1212 0 : in[8] = load_input_data(input + 8 * 1);
1213 0 : in[1] = load_input_data(input + 8 * 2);
1214 0 : in[9] = load_input_data(input + 8 * 3);
1215 0 : in[2] = load_input_data(input + 8 * 4);
1216 0 : in[10] = load_input_data(input + 8 * 5);
1217 0 : in[3] = load_input_data(input + 8 * 6);
1218 0 : in[11] = load_input_data(input + 8 * 7);
1219 0 : in[4] = load_input_data(input + 8 * 8);
1220 0 : in[12] = load_input_data(input + 8 * 9);
1221 0 : in[5] = load_input_data(input + 8 * 10);
1222 0 : in[13] = load_input_data(input + 8 * 11);
1223 0 : in[6] = load_input_data(input + 8 * 12);
1224 0 : in[14] = load_input_data(input + 8 * 13);
1225 0 : in[7] = load_input_data(input + 8 * 14);
1226 0 : in[15] = load_input_data(input + 8 * 15);
1227 :
1228 0 : array_transpose_8x8(in, in);
1229 0 : array_transpose_8x8(in + 8, in + 8);
1230 :
1231 0 : IDCT16
1232 :
1233 : // Stage7
1234 0 : curr1[0] = _mm_add_epi16(stp2_0, stp1_15);
1235 0 : curr1[1] = _mm_add_epi16(stp2_1, stp1_14);
1236 0 : curr1[2] = _mm_add_epi16(stp2_2, stp2_13);
1237 0 : curr1[3] = _mm_add_epi16(stp2_3, stp2_12);
1238 0 : curr1[4] = _mm_add_epi16(stp2_4, stp2_11);
1239 0 : curr1[5] = _mm_add_epi16(stp2_5, stp2_10);
1240 0 : curr1[6] = _mm_add_epi16(stp2_6, stp1_9);
1241 0 : curr1[7] = _mm_add_epi16(stp2_7, stp1_8);
1242 0 : curr1[8] = _mm_sub_epi16(stp2_7, stp1_8);
1243 0 : curr1[9] = _mm_sub_epi16(stp2_6, stp1_9);
1244 0 : curr1[10] = _mm_sub_epi16(stp2_5, stp2_10);
1245 0 : curr1[11] = _mm_sub_epi16(stp2_4, stp2_11);
1246 0 : curr1[12] = _mm_sub_epi16(stp2_3, stp2_12);
1247 0 : curr1[13] = _mm_sub_epi16(stp2_2, stp2_13);
1248 0 : curr1[14] = _mm_sub_epi16(stp2_1, stp1_14);
1249 0 : curr1[15] = _mm_sub_epi16(stp2_0, stp1_15);
1250 :
1251 0 : curr1 = r;
1252 0 : input += 128;
1253 : }
1254 0 : for (i = 0; i < 2; i++) {
1255 : int j;
1256 : // 1-D idct
1257 0 : array_transpose_8x8(l + i * 8, in);
1258 0 : array_transpose_8x8(r + i * 8, in + 8);
1259 :
1260 0 : IDCT16
1261 :
1262 : // 2-D
1263 0 : in[0] = _mm_add_epi16(stp2_0, stp1_15);
1264 0 : in[1] = _mm_add_epi16(stp2_1, stp1_14);
1265 0 : in[2] = _mm_add_epi16(stp2_2, stp2_13);
1266 0 : in[3] = _mm_add_epi16(stp2_3, stp2_12);
1267 0 : in[4] = _mm_add_epi16(stp2_4, stp2_11);
1268 0 : in[5] = _mm_add_epi16(stp2_5, stp2_10);
1269 0 : in[6] = _mm_add_epi16(stp2_6, stp1_9);
1270 0 : in[7] = _mm_add_epi16(stp2_7, stp1_8);
1271 0 : in[8] = _mm_sub_epi16(stp2_7, stp1_8);
1272 0 : in[9] = _mm_sub_epi16(stp2_6, stp1_9);
1273 0 : in[10] = _mm_sub_epi16(stp2_5, stp2_10);
1274 0 : in[11] = _mm_sub_epi16(stp2_4, stp2_11);
1275 0 : in[12] = _mm_sub_epi16(stp2_3, stp2_12);
1276 0 : in[13] = _mm_sub_epi16(stp2_2, stp2_13);
1277 0 : in[14] = _mm_sub_epi16(stp2_1, stp1_14);
1278 0 : in[15] = _mm_sub_epi16(stp2_0, stp1_15);
1279 :
1280 0 : for (j = 0; j < 16; ++j) {
1281 : // Final rounding and shift
1282 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
1283 0 : in[j] = _mm_srai_epi16(in[j], 6);
1284 0 : RECON_AND_STORE(dest + j * stride, in[j]);
1285 : }
1286 :
1287 0 : dest += 8;
1288 : }
1289 0 : }
1290 :
1291 0 : void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest,
1292 : int stride) {
1293 : __m128i dc_value;
1294 0 : const __m128i zero = _mm_setzero_si128();
1295 : int a, i;
1296 :
1297 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
1298 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
1299 0 : a = ROUND_POWER_OF_TWO(a, 6);
1300 :
1301 0 : dc_value = _mm_set1_epi16(a);
1302 :
1303 0 : for (i = 0; i < 16; ++i) {
1304 0 : RECON_AND_STORE(dest + 0, dc_value);
1305 0 : RECON_AND_STORE(dest + 8, dc_value);
1306 0 : dest += stride;
1307 : }
1308 0 : }
1309 :
1310 0 : static void iadst16_8col(__m128i *in) {
1311 : // perform 16x16 1-D ADST for 8 columns
1312 : __m128i s[16], x[16], u[32], v[32];
1313 0 : const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
1314 0 : const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
1315 0 : const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
1316 0 : const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
1317 0 : const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
1318 0 : const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
1319 0 : const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
1320 0 : const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
1321 0 : const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
1322 0 : const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
1323 0 : const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
1324 0 : const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
1325 0 : const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
1326 0 : const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
1327 0 : const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
1328 0 : const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
1329 0 : const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
1330 0 : const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
1331 0 : const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
1332 0 : const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
1333 0 : const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
1334 0 : const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
1335 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
1336 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
1337 0 : const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
1338 0 : const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
1339 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
1340 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1341 0 : const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1342 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
1343 0 : const __m128i kZero = _mm_set1_epi16(0);
1344 :
1345 0 : u[0] = _mm_unpacklo_epi16(in[15], in[0]);
1346 0 : u[1] = _mm_unpackhi_epi16(in[15], in[0]);
1347 0 : u[2] = _mm_unpacklo_epi16(in[13], in[2]);
1348 0 : u[3] = _mm_unpackhi_epi16(in[13], in[2]);
1349 0 : u[4] = _mm_unpacklo_epi16(in[11], in[4]);
1350 0 : u[5] = _mm_unpackhi_epi16(in[11], in[4]);
1351 0 : u[6] = _mm_unpacklo_epi16(in[9], in[6]);
1352 0 : u[7] = _mm_unpackhi_epi16(in[9], in[6]);
1353 0 : u[8] = _mm_unpacklo_epi16(in[7], in[8]);
1354 0 : u[9] = _mm_unpackhi_epi16(in[7], in[8]);
1355 0 : u[10] = _mm_unpacklo_epi16(in[5], in[10]);
1356 0 : u[11] = _mm_unpackhi_epi16(in[5], in[10]);
1357 0 : u[12] = _mm_unpacklo_epi16(in[3], in[12]);
1358 0 : u[13] = _mm_unpackhi_epi16(in[3], in[12]);
1359 0 : u[14] = _mm_unpacklo_epi16(in[1], in[14]);
1360 0 : u[15] = _mm_unpackhi_epi16(in[1], in[14]);
1361 :
1362 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
1363 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
1364 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
1365 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
1366 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
1367 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
1368 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
1369 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
1370 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
1371 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
1372 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
1373 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
1374 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
1375 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
1376 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
1377 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
1378 0 : v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
1379 0 : v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
1380 0 : v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
1381 0 : v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
1382 0 : v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
1383 0 : v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
1384 0 : v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
1385 0 : v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
1386 0 : v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
1387 0 : v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
1388 0 : v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
1389 0 : v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
1390 0 : v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
1391 0 : v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
1392 0 : v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
1393 0 : v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
1394 :
1395 0 : u[0] = _mm_add_epi32(v[0], v[16]);
1396 0 : u[1] = _mm_add_epi32(v[1], v[17]);
1397 0 : u[2] = _mm_add_epi32(v[2], v[18]);
1398 0 : u[3] = _mm_add_epi32(v[3], v[19]);
1399 0 : u[4] = _mm_add_epi32(v[4], v[20]);
1400 0 : u[5] = _mm_add_epi32(v[5], v[21]);
1401 0 : u[6] = _mm_add_epi32(v[6], v[22]);
1402 0 : u[7] = _mm_add_epi32(v[7], v[23]);
1403 0 : u[8] = _mm_add_epi32(v[8], v[24]);
1404 0 : u[9] = _mm_add_epi32(v[9], v[25]);
1405 0 : u[10] = _mm_add_epi32(v[10], v[26]);
1406 0 : u[11] = _mm_add_epi32(v[11], v[27]);
1407 0 : u[12] = _mm_add_epi32(v[12], v[28]);
1408 0 : u[13] = _mm_add_epi32(v[13], v[29]);
1409 0 : u[14] = _mm_add_epi32(v[14], v[30]);
1410 0 : u[15] = _mm_add_epi32(v[15], v[31]);
1411 0 : u[16] = _mm_sub_epi32(v[0], v[16]);
1412 0 : u[17] = _mm_sub_epi32(v[1], v[17]);
1413 0 : u[18] = _mm_sub_epi32(v[2], v[18]);
1414 0 : u[19] = _mm_sub_epi32(v[3], v[19]);
1415 0 : u[20] = _mm_sub_epi32(v[4], v[20]);
1416 0 : u[21] = _mm_sub_epi32(v[5], v[21]);
1417 0 : u[22] = _mm_sub_epi32(v[6], v[22]);
1418 0 : u[23] = _mm_sub_epi32(v[7], v[23]);
1419 0 : u[24] = _mm_sub_epi32(v[8], v[24]);
1420 0 : u[25] = _mm_sub_epi32(v[9], v[25]);
1421 0 : u[26] = _mm_sub_epi32(v[10], v[26]);
1422 0 : u[27] = _mm_sub_epi32(v[11], v[27]);
1423 0 : u[28] = _mm_sub_epi32(v[12], v[28]);
1424 0 : u[29] = _mm_sub_epi32(v[13], v[29]);
1425 0 : u[30] = _mm_sub_epi32(v[14], v[30]);
1426 0 : u[31] = _mm_sub_epi32(v[15], v[31]);
1427 :
1428 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1429 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1430 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1431 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1432 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1433 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1434 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1435 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1436 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1437 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1438 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1439 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1440 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1441 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1442 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1443 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1444 0 : v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
1445 0 : v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
1446 0 : v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
1447 0 : v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
1448 0 : v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
1449 0 : v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
1450 0 : v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
1451 0 : v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
1452 0 : v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
1453 0 : v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
1454 0 : v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
1455 0 : v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
1456 0 : v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
1457 0 : v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
1458 0 : v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
1459 0 : v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
1460 :
1461 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1462 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1463 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1464 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1465 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1466 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1467 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1468 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1469 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
1470 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
1471 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1472 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1473 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1474 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1475 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1476 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1477 0 : u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
1478 0 : u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
1479 0 : u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
1480 0 : u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
1481 0 : u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
1482 0 : u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
1483 0 : u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
1484 0 : u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
1485 0 : u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
1486 0 : u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
1487 0 : u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
1488 0 : u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
1489 0 : u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
1490 0 : u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
1491 0 : u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
1492 0 : u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
1493 :
1494 0 : s[0] = _mm_packs_epi32(u[0], u[1]);
1495 0 : s[1] = _mm_packs_epi32(u[2], u[3]);
1496 0 : s[2] = _mm_packs_epi32(u[4], u[5]);
1497 0 : s[3] = _mm_packs_epi32(u[6], u[7]);
1498 0 : s[4] = _mm_packs_epi32(u[8], u[9]);
1499 0 : s[5] = _mm_packs_epi32(u[10], u[11]);
1500 0 : s[6] = _mm_packs_epi32(u[12], u[13]);
1501 0 : s[7] = _mm_packs_epi32(u[14], u[15]);
1502 0 : s[8] = _mm_packs_epi32(u[16], u[17]);
1503 0 : s[9] = _mm_packs_epi32(u[18], u[19]);
1504 0 : s[10] = _mm_packs_epi32(u[20], u[21]);
1505 0 : s[11] = _mm_packs_epi32(u[22], u[23]);
1506 0 : s[12] = _mm_packs_epi32(u[24], u[25]);
1507 0 : s[13] = _mm_packs_epi32(u[26], u[27]);
1508 0 : s[14] = _mm_packs_epi32(u[28], u[29]);
1509 0 : s[15] = _mm_packs_epi32(u[30], u[31]);
1510 :
1511 : // stage 2
1512 0 : u[0] = _mm_unpacklo_epi16(s[8], s[9]);
1513 0 : u[1] = _mm_unpackhi_epi16(s[8], s[9]);
1514 0 : u[2] = _mm_unpacklo_epi16(s[10], s[11]);
1515 0 : u[3] = _mm_unpackhi_epi16(s[10], s[11]);
1516 0 : u[4] = _mm_unpacklo_epi16(s[12], s[13]);
1517 0 : u[5] = _mm_unpackhi_epi16(s[12], s[13]);
1518 0 : u[6] = _mm_unpacklo_epi16(s[14], s[15]);
1519 0 : u[7] = _mm_unpackhi_epi16(s[14], s[15]);
1520 :
1521 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
1522 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
1523 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
1524 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
1525 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
1526 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
1527 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
1528 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
1529 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
1530 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
1531 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
1532 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
1533 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
1534 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
1535 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
1536 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
1537 :
1538 0 : u[0] = _mm_add_epi32(v[0], v[8]);
1539 0 : u[1] = _mm_add_epi32(v[1], v[9]);
1540 0 : u[2] = _mm_add_epi32(v[2], v[10]);
1541 0 : u[3] = _mm_add_epi32(v[3], v[11]);
1542 0 : u[4] = _mm_add_epi32(v[4], v[12]);
1543 0 : u[5] = _mm_add_epi32(v[5], v[13]);
1544 0 : u[6] = _mm_add_epi32(v[6], v[14]);
1545 0 : u[7] = _mm_add_epi32(v[7], v[15]);
1546 0 : u[8] = _mm_sub_epi32(v[0], v[8]);
1547 0 : u[9] = _mm_sub_epi32(v[1], v[9]);
1548 0 : u[10] = _mm_sub_epi32(v[2], v[10]);
1549 0 : u[11] = _mm_sub_epi32(v[3], v[11]);
1550 0 : u[12] = _mm_sub_epi32(v[4], v[12]);
1551 0 : u[13] = _mm_sub_epi32(v[5], v[13]);
1552 0 : u[14] = _mm_sub_epi32(v[6], v[14]);
1553 0 : u[15] = _mm_sub_epi32(v[7], v[15]);
1554 :
1555 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1556 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1557 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1558 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1559 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1560 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1561 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1562 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1563 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1564 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1565 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1566 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1567 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1568 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1569 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1570 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1571 :
1572 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1573 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1574 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1575 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1576 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1577 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1578 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1579 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1580 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
1581 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
1582 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1583 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1584 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1585 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1586 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1587 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1588 :
1589 0 : x[0] = _mm_add_epi16(s[0], s[4]);
1590 0 : x[1] = _mm_add_epi16(s[1], s[5]);
1591 0 : x[2] = _mm_add_epi16(s[2], s[6]);
1592 0 : x[3] = _mm_add_epi16(s[3], s[7]);
1593 0 : x[4] = _mm_sub_epi16(s[0], s[4]);
1594 0 : x[5] = _mm_sub_epi16(s[1], s[5]);
1595 0 : x[6] = _mm_sub_epi16(s[2], s[6]);
1596 0 : x[7] = _mm_sub_epi16(s[3], s[7]);
1597 0 : x[8] = _mm_packs_epi32(u[0], u[1]);
1598 0 : x[9] = _mm_packs_epi32(u[2], u[3]);
1599 0 : x[10] = _mm_packs_epi32(u[4], u[5]);
1600 0 : x[11] = _mm_packs_epi32(u[6], u[7]);
1601 0 : x[12] = _mm_packs_epi32(u[8], u[9]);
1602 0 : x[13] = _mm_packs_epi32(u[10], u[11]);
1603 0 : x[14] = _mm_packs_epi32(u[12], u[13]);
1604 0 : x[15] = _mm_packs_epi32(u[14], u[15]);
1605 :
1606 : // stage 3
1607 0 : u[0] = _mm_unpacklo_epi16(x[4], x[5]);
1608 0 : u[1] = _mm_unpackhi_epi16(x[4], x[5]);
1609 0 : u[2] = _mm_unpacklo_epi16(x[6], x[7]);
1610 0 : u[3] = _mm_unpackhi_epi16(x[6], x[7]);
1611 0 : u[4] = _mm_unpacklo_epi16(x[12], x[13]);
1612 0 : u[5] = _mm_unpackhi_epi16(x[12], x[13]);
1613 0 : u[6] = _mm_unpacklo_epi16(x[14], x[15]);
1614 0 : u[7] = _mm_unpackhi_epi16(x[14], x[15]);
1615 :
1616 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
1617 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
1618 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
1619 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
1620 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
1621 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
1622 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
1623 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
1624 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
1625 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
1626 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
1627 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
1628 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
1629 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
1630 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
1631 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
1632 :
1633 0 : u[0] = _mm_add_epi32(v[0], v[4]);
1634 0 : u[1] = _mm_add_epi32(v[1], v[5]);
1635 0 : u[2] = _mm_add_epi32(v[2], v[6]);
1636 0 : u[3] = _mm_add_epi32(v[3], v[7]);
1637 0 : u[4] = _mm_sub_epi32(v[0], v[4]);
1638 0 : u[5] = _mm_sub_epi32(v[1], v[5]);
1639 0 : u[6] = _mm_sub_epi32(v[2], v[6]);
1640 0 : u[7] = _mm_sub_epi32(v[3], v[7]);
1641 0 : u[8] = _mm_add_epi32(v[8], v[12]);
1642 0 : u[9] = _mm_add_epi32(v[9], v[13]);
1643 0 : u[10] = _mm_add_epi32(v[10], v[14]);
1644 0 : u[11] = _mm_add_epi32(v[11], v[15]);
1645 0 : u[12] = _mm_sub_epi32(v[8], v[12]);
1646 0 : u[13] = _mm_sub_epi32(v[9], v[13]);
1647 0 : u[14] = _mm_sub_epi32(v[10], v[14]);
1648 0 : u[15] = _mm_sub_epi32(v[11], v[15]);
1649 :
1650 0 : u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1651 0 : u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1652 0 : u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1653 0 : u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1654 0 : u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1655 0 : u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1656 0 : u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1657 0 : u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1658 0 : u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1659 0 : u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1660 0 : u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1661 0 : u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1662 0 : u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1663 0 : u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1664 0 : u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1665 0 : u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1666 :
1667 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1668 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1669 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1670 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1671 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1672 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1673 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1674 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1675 0 : v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1676 0 : v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1677 0 : v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1678 0 : v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1679 0 : v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1680 0 : v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1681 0 : v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1682 0 : v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1683 :
1684 0 : s[0] = _mm_add_epi16(x[0], x[2]);
1685 0 : s[1] = _mm_add_epi16(x[1], x[3]);
1686 0 : s[2] = _mm_sub_epi16(x[0], x[2]);
1687 0 : s[3] = _mm_sub_epi16(x[1], x[3]);
1688 0 : s[4] = _mm_packs_epi32(v[0], v[1]);
1689 0 : s[5] = _mm_packs_epi32(v[2], v[3]);
1690 0 : s[6] = _mm_packs_epi32(v[4], v[5]);
1691 0 : s[7] = _mm_packs_epi32(v[6], v[7]);
1692 0 : s[8] = _mm_add_epi16(x[8], x[10]);
1693 0 : s[9] = _mm_add_epi16(x[9], x[11]);
1694 0 : s[10] = _mm_sub_epi16(x[8], x[10]);
1695 0 : s[11] = _mm_sub_epi16(x[9], x[11]);
1696 0 : s[12] = _mm_packs_epi32(v[8], v[9]);
1697 0 : s[13] = _mm_packs_epi32(v[10], v[11]);
1698 0 : s[14] = _mm_packs_epi32(v[12], v[13]);
1699 0 : s[15] = _mm_packs_epi32(v[14], v[15]);
1700 :
1701 : // stage 4
1702 0 : u[0] = _mm_unpacklo_epi16(s[2], s[3]);
1703 0 : u[1] = _mm_unpackhi_epi16(s[2], s[3]);
1704 0 : u[2] = _mm_unpacklo_epi16(s[6], s[7]);
1705 0 : u[3] = _mm_unpackhi_epi16(s[6], s[7]);
1706 0 : u[4] = _mm_unpacklo_epi16(s[10], s[11]);
1707 0 : u[5] = _mm_unpackhi_epi16(s[10], s[11]);
1708 0 : u[6] = _mm_unpacklo_epi16(s[14], s[15]);
1709 0 : u[7] = _mm_unpackhi_epi16(s[14], s[15]);
1710 :
1711 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
1712 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
1713 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
1714 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
1715 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
1716 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
1717 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
1718 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
1719 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
1720 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
1721 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
1722 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
1723 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
1724 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
1725 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
1726 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
1727 :
1728 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1729 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1730 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1731 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1732 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1733 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1734 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1735 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1736 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1737 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1738 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1739 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1740 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1741 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1742 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1743 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1744 :
1745 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1746 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1747 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1748 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1749 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1750 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1751 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1752 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1753 0 : v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1754 0 : v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1755 0 : v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1756 0 : v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1757 0 : v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1758 0 : v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1759 0 : v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1760 0 : v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1761 :
1762 0 : in[0] = s[0];
1763 0 : in[1] = _mm_sub_epi16(kZero, s[8]);
1764 0 : in[2] = s[12];
1765 0 : in[3] = _mm_sub_epi16(kZero, s[4]);
1766 0 : in[4] = _mm_packs_epi32(v[4], v[5]);
1767 0 : in[5] = _mm_packs_epi32(v[12], v[13]);
1768 0 : in[6] = _mm_packs_epi32(v[8], v[9]);
1769 0 : in[7] = _mm_packs_epi32(v[0], v[1]);
1770 0 : in[8] = _mm_packs_epi32(v[2], v[3]);
1771 0 : in[9] = _mm_packs_epi32(v[10], v[11]);
1772 0 : in[10] = _mm_packs_epi32(v[14], v[15]);
1773 0 : in[11] = _mm_packs_epi32(v[6], v[7]);
1774 0 : in[12] = s[5];
1775 0 : in[13] = _mm_sub_epi16(kZero, s[13]);
1776 0 : in[14] = s[9];
1777 0 : in[15] = _mm_sub_epi16(kZero, s[1]);
1778 0 : }
1779 :
1780 0 : static void idct16_8col(__m128i *in) {
1781 0 : const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
1782 0 : const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
1783 0 : const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
1784 0 : const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
1785 0 : const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
1786 0 : const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
1787 0 : const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
1788 0 : const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
1789 0 : const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
1790 0 : const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
1791 0 : const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
1792 0 : const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
1793 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
1794 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1795 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
1796 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
1797 0 : const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
1798 0 : const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
1799 0 : const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
1800 0 : const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1801 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
1802 : __m128i v[16], u[16], s[16], t[16];
1803 :
1804 : // stage 1
1805 0 : s[0] = in[0];
1806 0 : s[1] = in[8];
1807 0 : s[2] = in[4];
1808 0 : s[3] = in[12];
1809 0 : s[4] = in[2];
1810 0 : s[5] = in[10];
1811 0 : s[6] = in[6];
1812 0 : s[7] = in[14];
1813 0 : s[8] = in[1];
1814 0 : s[9] = in[9];
1815 0 : s[10] = in[5];
1816 0 : s[11] = in[13];
1817 0 : s[12] = in[3];
1818 0 : s[13] = in[11];
1819 0 : s[14] = in[7];
1820 0 : s[15] = in[15];
1821 :
1822 : // stage 2
1823 0 : u[0] = _mm_unpacklo_epi16(s[8], s[15]);
1824 0 : u[1] = _mm_unpackhi_epi16(s[8], s[15]);
1825 0 : u[2] = _mm_unpacklo_epi16(s[9], s[14]);
1826 0 : u[3] = _mm_unpackhi_epi16(s[9], s[14]);
1827 0 : u[4] = _mm_unpacklo_epi16(s[10], s[13]);
1828 0 : u[5] = _mm_unpackhi_epi16(s[10], s[13]);
1829 0 : u[6] = _mm_unpacklo_epi16(s[11], s[12]);
1830 0 : u[7] = _mm_unpackhi_epi16(s[11], s[12]);
1831 :
1832 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
1833 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
1834 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
1835 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
1836 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
1837 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
1838 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
1839 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
1840 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
1841 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10);
1842 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22);
1843 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22);
1844 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26);
1845 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26);
1846 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06);
1847 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06);
1848 :
1849 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1850 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1851 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1852 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1853 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1854 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1855 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1856 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1857 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1858 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1859 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1860 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1861 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1862 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1863 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1864 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1865 :
1866 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1867 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1868 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1869 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1870 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1871 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1872 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1873 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1874 0 : u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1875 0 : u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1876 0 : u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1877 0 : u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1878 0 : u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1879 0 : u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1880 0 : u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1881 0 : u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1882 :
1883 0 : s[8] = _mm_packs_epi32(u[0], u[1]);
1884 0 : s[15] = _mm_packs_epi32(u[2], u[3]);
1885 0 : s[9] = _mm_packs_epi32(u[4], u[5]);
1886 0 : s[14] = _mm_packs_epi32(u[6], u[7]);
1887 0 : s[10] = _mm_packs_epi32(u[8], u[9]);
1888 0 : s[13] = _mm_packs_epi32(u[10], u[11]);
1889 0 : s[11] = _mm_packs_epi32(u[12], u[13]);
1890 0 : s[12] = _mm_packs_epi32(u[14], u[15]);
1891 :
1892 : // stage 3
1893 0 : t[0] = s[0];
1894 0 : t[1] = s[1];
1895 0 : t[2] = s[2];
1896 0 : t[3] = s[3];
1897 0 : u[0] = _mm_unpacklo_epi16(s[4], s[7]);
1898 0 : u[1] = _mm_unpackhi_epi16(s[4], s[7]);
1899 0 : u[2] = _mm_unpacklo_epi16(s[5], s[6]);
1900 0 : u[3] = _mm_unpackhi_epi16(s[5], s[6]);
1901 :
1902 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
1903 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
1904 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
1905 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
1906 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
1907 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
1908 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
1909 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
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 :
1920 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1921 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1922 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1923 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1924 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1925 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1926 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1927 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1928 :
1929 0 : t[4] = _mm_packs_epi32(u[0], u[1]);
1930 0 : t[7] = _mm_packs_epi32(u[2], u[3]);
1931 0 : t[5] = _mm_packs_epi32(u[4], u[5]);
1932 0 : t[6] = _mm_packs_epi32(u[6], u[7]);
1933 0 : t[8] = _mm_add_epi16(s[8], s[9]);
1934 0 : t[9] = _mm_sub_epi16(s[8], s[9]);
1935 0 : t[10] = _mm_sub_epi16(s[11], s[10]);
1936 0 : t[11] = _mm_add_epi16(s[10], s[11]);
1937 0 : t[12] = _mm_add_epi16(s[12], s[13]);
1938 0 : t[13] = _mm_sub_epi16(s[12], s[13]);
1939 0 : t[14] = _mm_sub_epi16(s[15], s[14]);
1940 0 : t[15] = _mm_add_epi16(s[14], s[15]);
1941 :
1942 : // stage 4
1943 0 : u[0] = _mm_unpacklo_epi16(t[0], t[1]);
1944 0 : u[1] = _mm_unpackhi_epi16(t[0], t[1]);
1945 0 : u[2] = _mm_unpacklo_epi16(t[2], t[3]);
1946 0 : u[3] = _mm_unpackhi_epi16(t[2], t[3]);
1947 0 : u[4] = _mm_unpacklo_epi16(t[9], t[14]);
1948 0 : u[5] = _mm_unpackhi_epi16(t[9], t[14]);
1949 0 : u[6] = _mm_unpacklo_epi16(t[10], t[13]);
1950 0 : u[7] = _mm_unpackhi_epi16(t[10], t[13]);
1951 :
1952 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
1953 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
1954 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
1955 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
1956 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08);
1957 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08);
1958 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
1959 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
1960 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24);
1961 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24);
1962 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08);
1963 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08);
1964 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08);
1965 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08);
1966 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24);
1967 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24);
1968 :
1969 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1970 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1971 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1972 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1973 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1974 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1975 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1976 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1977 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1978 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1979 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1980 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1981 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1982 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1983 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1984 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1985 :
1986 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1987 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1988 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1989 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1990 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1991 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1992 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1993 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1994 0 : u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1995 0 : u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1996 0 : u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1997 0 : u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1998 0 : u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1999 0 : u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
2000 0 : u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
2001 0 : u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
2002 :
2003 0 : s[0] = _mm_packs_epi32(u[0], u[1]);
2004 0 : s[1] = _mm_packs_epi32(u[2], u[3]);
2005 0 : s[2] = _mm_packs_epi32(u[4], u[5]);
2006 0 : s[3] = _mm_packs_epi32(u[6], u[7]);
2007 0 : s[4] = _mm_add_epi16(t[4], t[5]);
2008 0 : s[5] = _mm_sub_epi16(t[4], t[5]);
2009 0 : s[6] = _mm_sub_epi16(t[7], t[6]);
2010 0 : s[7] = _mm_add_epi16(t[6], t[7]);
2011 0 : s[8] = t[8];
2012 0 : s[15] = t[15];
2013 0 : s[9] = _mm_packs_epi32(u[8], u[9]);
2014 0 : s[14] = _mm_packs_epi32(u[10], u[11]);
2015 0 : s[10] = _mm_packs_epi32(u[12], u[13]);
2016 0 : s[13] = _mm_packs_epi32(u[14], u[15]);
2017 0 : s[11] = t[11];
2018 0 : s[12] = t[12];
2019 :
2020 : // stage 5
2021 0 : t[0] = _mm_add_epi16(s[0], s[3]);
2022 0 : t[1] = _mm_add_epi16(s[1], s[2]);
2023 0 : t[2] = _mm_sub_epi16(s[1], s[2]);
2024 0 : t[3] = _mm_sub_epi16(s[0], s[3]);
2025 0 : t[4] = s[4];
2026 0 : t[7] = s[7];
2027 :
2028 0 : u[0] = _mm_unpacklo_epi16(s[5], s[6]);
2029 0 : u[1] = _mm_unpackhi_epi16(s[5], s[6]);
2030 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
2031 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
2032 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
2033 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
2034 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
2035 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
2036 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
2037 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
2038 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
2039 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
2040 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
2041 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
2042 0 : t[5] = _mm_packs_epi32(u[0], u[1]);
2043 0 : t[6] = _mm_packs_epi32(u[2], u[3]);
2044 :
2045 0 : t[8] = _mm_add_epi16(s[8], s[11]);
2046 0 : t[9] = _mm_add_epi16(s[9], s[10]);
2047 0 : t[10] = _mm_sub_epi16(s[9], s[10]);
2048 0 : t[11] = _mm_sub_epi16(s[8], s[11]);
2049 0 : t[12] = _mm_sub_epi16(s[15], s[12]);
2050 0 : t[13] = _mm_sub_epi16(s[14], s[13]);
2051 0 : t[14] = _mm_add_epi16(s[13], s[14]);
2052 0 : t[15] = _mm_add_epi16(s[12], s[15]);
2053 :
2054 : // stage 6
2055 0 : s[0] = _mm_add_epi16(t[0], t[7]);
2056 0 : s[1] = _mm_add_epi16(t[1], t[6]);
2057 0 : s[2] = _mm_add_epi16(t[2], t[5]);
2058 0 : s[3] = _mm_add_epi16(t[3], t[4]);
2059 0 : s[4] = _mm_sub_epi16(t[3], t[4]);
2060 0 : s[5] = _mm_sub_epi16(t[2], t[5]);
2061 0 : s[6] = _mm_sub_epi16(t[1], t[6]);
2062 0 : s[7] = _mm_sub_epi16(t[0], t[7]);
2063 0 : s[8] = t[8];
2064 0 : s[9] = t[9];
2065 :
2066 0 : u[0] = _mm_unpacklo_epi16(t[10], t[13]);
2067 0 : u[1] = _mm_unpackhi_epi16(t[10], t[13]);
2068 0 : u[2] = _mm_unpacklo_epi16(t[11], t[12]);
2069 0 : u[3] = _mm_unpackhi_epi16(t[11], t[12]);
2070 :
2071 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
2072 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
2073 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
2074 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
2075 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
2076 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
2077 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
2078 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
2079 :
2080 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
2081 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
2082 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
2083 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
2084 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
2085 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
2086 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
2087 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
2088 :
2089 0 : u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
2090 0 : u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
2091 0 : u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
2092 0 : u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
2093 0 : u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
2094 0 : u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
2095 0 : u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
2096 0 : u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
2097 :
2098 0 : s[10] = _mm_packs_epi32(u[0], u[1]);
2099 0 : s[13] = _mm_packs_epi32(u[2], u[3]);
2100 0 : s[11] = _mm_packs_epi32(u[4], u[5]);
2101 0 : s[12] = _mm_packs_epi32(u[6], u[7]);
2102 0 : s[14] = t[14];
2103 0 : s[15] = t[15];
2104 :
2105 : // stage 7
2106 0 : in[0] = _mm_add_epi16(s[0], s[15]);
2107 0 : in[1] = _mm_add_epi16(s[1], s[14]);
2108 0 : in[2] = _mm_add_epi16(s[2], s[13]);
2109 0 : in[3] = _mm_add_epi16(s[3], s[12]);
2110 0 : in[4] = _mm_add_epi16(s[4], s[11]);
2111 0 : in[5] = _mm_add_epi16(s[5], s[10]);
2112 0 : in[6] = _mm_add_epi16(s[6], s[9]);
2113 0 : in[7] = _mm_add_epi16(s[7], s[8]);
2114 0 : in[8] = _mm_sub_epi16(s[7], s[8]);
2115 0 : in[9] = _mm_sub_epi16(s[6], s[9]);
2116 0 : in[10] = _mm_sub_epi16(s[5], s[10]);
2117 0 : in[11] = _mm_sub_epi16(s[4], s[11]);
2118 0 : in[12] = _mm_sub_epi16(s[3], s[12]);
2119 0 : in[13] = _mm_sub_epi16(s[2], s[13]);
2120 0 : in[14] = _mm_sub_epi16(s[1], s[14]);
2121 0 : in[15] = _mm_sub_epi16(s[0], s[15]);
2122 0 : }
2123 :
2124 0 : void idct16_sse2(__m128i *in0, __m128i *in1) {
2125 0 : array_transpose_16x16(in0, in1);
2126 0 : idct16_8col(in0);
2127 0 : idct16_8col(in1);
2128 0 : }
2129 :
2130 0 : void iadst16_sse2(__m128i *in0, __m128i *in1) {
2131 0 : array_transpose_16x16(in0, in1);
2132 0 : iadst16_8col(in0);
2133 0 : iadst16_8col(in1);
2134 0 : }
2135 :
2136 0 : void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest,
2137 : int stride) {
2138 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
2139 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
2140 0 : const __m128i zero = _mm_setzero_si128();
2141 :
2142 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
2143 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
2144 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
2145 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
2146 :
2147 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
2148 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
2149 :
2150 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
2151 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
2152 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
2153 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
2154 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
2155 0 : const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
2156 :
2157 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
2158 : __m128i in[16], l[16];
2159 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_8,
2160 : stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, stp1_8_0,
2161 : stp1_12_0;
2162 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
2163 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14;
2164 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
2165 : int i;
2166 : // First 1-D inverse DCT
2167 : // Load input data.
2168 0 : in[0] = load_input_data(input);
2169 0 : in[1] = load_input_data(input + 8 * 2);
2170 0 : in[2] = load_input_data(input + 8 * 4);
2171 0 : in[3] = load_input_data(input + 8 * 6);
2172 :
2173 0 : TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]);
2174 :
2175 : // Stage2
2176 : {
2177 0 : const __m128i lo_1_15 = _mm_unpackhi_epi16(in[0], zero);
2178 0 : const __m128i lo_13_3 = _mm_unpackhi_epi16(zero, in[1]);
2179 :
2180 0 : tmp0 = _mm_madd_epi16(lo_1_15, stg2_0);
2181 0 : tmp2 = _mm_madd_epi16(lo_1_15, stg2_1);
2182 0 : tmp5 = _mm_madd_epi16(lo_13_3, stg2_6);
2183 0 : tmp7 = _mm_madd_epi16(lo_13_3, stg2_7);
2184 :
2185 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2186 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2187 0 : tmp5 = _mm_add_epi32(tmp5, rounding);
2188 0 : tmp7 = _mm_add_epi32(tmp7, rounding);
2189 :
2190 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2191 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2192 0 : tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
2193 0 : tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
2194 :
2195 0 : stp2_8 = _mm_packs_epi32(tmp0, tmp2);
2196 0 : stp2_11 = _mm_packs_epi32(tmp5, tmp7);
2197 : }
2198 :
2199 : // Stage3
2200 : {
2201 0 : const __m128i lo_2_14 = _mm_unpacklo_epi16(in[1], zero);
2202 :
2203 0 : tmp0 = _mm_madd_epi16(lo_2_14, stg3_0);
2204 0 : tmp2 = _mm_madd_epi16(lo_2_14, stg3_1);
2205 :
2206 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2207 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2208 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2209 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2210 :
2211 0 : stp1_13 = _mm_unpackhi_epi64(stp2_11, zero);
2212 0 : stp1_14 = _mm_unpackhi_epi64(stp2_8, zero);
2213 :
2214 0 : stp1_4 = _mm_packs_epi32(tmp0, tmp2);
2215 : }
2216 :
2217 : // Stage4
2218 : {
2219 0 : const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero);
2220 0 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14);
2221 0 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13);
2222 :
2223 0 : tmp0 = _mm_madd_epi16(lo_0_8, stg4_0);
2224 0 : tmp2 = _mm_madd_epi16(lo_0_8, stg4_1);
2225 0 : tmp1 = _mm_madd_epi16(lo_9_14, stg4_4);
2226 0 : tmp3 = _mm_madd_epi16(lo_9_14, stg4_5);
2227 0 : tmp5 = _mm_madd_epi16(lo_10_13, stg4_6);
2228 0 : tmp7 = _mm_madd_epi16(lo_10_13, stg4_7);
2229 :
2230 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2231 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2232 0 : tmp1 = _mm_add_epi32(tmp1, rounding);
2233 0 : tmp3 = _mm_add_epi32(tmp3, rounding);
2234 0 : tmp5 = _mm_add_epi32(tmp5, rounding);
2235 0 : tmp7 = _mm_add_epi32(tmp7, rounding);
2236 :
2237 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2238 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2239 0 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
2240 0 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
2241 0 : tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
2242 0 : tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
2243 :
2244 0 : stp1_0 = _mm_packs_epi32(tmp0, tmp0);
2245 0 : stp1_1 = _mm_packs_epi32(tmp2, tmp2);
2246 0 : stp2_9 = _mm_packs_epi32(tmp1, tmp3);
2247 0 : stp2_10 = _mm_packs_epi32(tmp5, tmp7);
2248 :
2249 0 : stp2_6 = _mm_unpackhi_epi64(stp1_4, zero);
2250 : }
2251 :
2252 : // Stage5 and Stage6
2253 : {
2254 0 : tmp0 = _mm_add_epi16(stp2_8, stp2_11);
2255 0 : tmp1 = _mm_sub_epi16(stp2_8, stp2_11);
2256 0 : tmp2 = _mm_add_epi16(stp2_9, stp2_10);
2257 0 : tmp3 = _mm_sub_epi16(stp2_9, stp2_10);
2258 :
2259 0 : stp1_9 = _mm_unpacklo_epi64(tmp2, zero);
2260 0 : stp1_10 = _mm_unpacklo_epi64(tmp3, zero);
2261 0 : stp1_8 = _mm_unpacklo_epi64(tmp0, zero);
2262 0 : stp1_11 = _mm_unpacklo_epi64(tmp1, zero);
2263 :
2264 0 : stp1_13 = _mm_unpackhi_epi64(tmp3, zero);
2265 0 : stp1_14 = _mm_unpackhi_epi64(tmp2, zero);
2266 0 : stp1_12 = _mm_unpackhi_epi64(tmp1, zero);
2267 0 : stp1_15 = _mm_unpackhi_epi64(tmp0, zero);
2268 : }
2269 :
2270 : // Stage6
2271 : {
2272 0 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp1_4);
2273 0 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);
2274 0 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);
2275 :
2276 0 : tmp1 = _mm_madd_epi16(lo_6_5, stg4_1);
2277 0 : tmp3 = _mm_madd_epi16(lo_6_5, stg4_0);
2278 0 : tmp0 = _mm_madd_epi16(lo_10_13, stg6_0);
2279 0 : tmp2 = _mm_madd_epi16(lo_10_13, stg4_0);
2280 0 : tmp4 = _mm_madd_epi16(lo_11_12, stg6_0);
2281 0 : tmp6 = _mm_madd_epi16(lo_11_12, stg4_0);
2282 :
2283 0 : tmp1 = _mm_add_epi32(tmp1, rounding);
2284 0 : tmp3 = _mm_add_epi32(tmp3, rounding);
2285 0 : tmp0 = _mm_add_epi32(tmp0, rounding);
2286 0 : tmp2 = _mm_add_epi32(tmp2, rounding);
2287 0 : tmp4 = _mm_add_epi32(tmp4, rounding);
2288 0 : tmp6 = _mm_add_epi32(tmp6, rounding);
2289 :
2290 0 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
2291 0 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
2292 0 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
2293 0 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
2294 0 : tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
2295 0 : tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
2296 :
2297 0 : stp1_6 = _mm_packs_epi32(tmp3, tmp1);
2298 :
2299 0 : stp2_10 = _mm_packs_epi32(tmp0, zero);
2300 0 : stp2_13 = _mm_packs_epi32(tmp2, zero);
2301 0 : stp2_11 = _mm_packs_epi32(tmp4, zero);
2302 0 : stp2_12 = _mm_packs_epi32(tmp6, zero);
2303 :
2304 0 : tmp0 = _mm_add_epi16(stp1_0, stp1_4);
2305 0 : tmp1 = _mm_sub_epi16(stp1_0, stp1_4);
2306 0 : tmp2 = _mm_add_epi16(stp1_1, stp1_6);
2307 0 : tmp3 = _mm_sub_epi16(stp1_1, stp1_6);
2308 :
2309 0 : stp2_0 = _mm_unpackhi_epi64(tmp0, zero);
2310 0 : stp2_1 = _mm_unpacklo_epi64(tmp2, zero);
2311 0 : stp2_2 = _mm_unpackhi_epi64(tmp2, zero);
2312 0 : stp2_3 = _mm_unpacklo_epi64(tmp0, zero);
2313 0 : stp2_4 = _mm_unpacklo_epi64(tmp1, zero);
2314 0 : stp2_5 = _mm_unpackhi_epi64(tmp3, zero);
2315 0 : stp2_6 = _mm_unpacklo_epi64(tmp3, zero);
2316 0 : stp2_7 = _mm_unpackhi_epi64(tmp1, zero);
2317 : }
2318 :
2319 : // Stage7. Left 8x16 only.
2320 0 : l[0] = _mm_add_epi16(stp2_0, stp1_15);
2321 0 : l[1] = _mm_add_epi16(stp2_1, stp1_14);
2322 0 : l[2] = _mm_add_epi16(stp2_2, stp2_13);
2323 0 : l[3] = _mm_add_epi16(stp2_3, stp2_12);
2324 0 : l[4] = _mm_add_epi16(stp2_4, stp2_11);
2325 0 : l[5] = _mm_add_epi16(stp2_5, stp2_10);
2326 0 : l[6] = _mm_add_epi16(stp2_6, stp1_9);
2327 0 : l[7] = _mm_add_epi16(stp2_7, stp1_8);
2328 0 : l[8] = _mm_sub_epi16(stp2_7, stp1_8);
2329 0 : l[9] = _mm_sub_epi16(stp2_6, stp1_9);
2330 0 : l[10] = _mm_sub_epi16(stp2_5, stp2_10);
2331 0 : l[11] = _mm_sub_epi16(stp2_4, stp2_11);
2332 0 : l[12] = _mm_sub_epi16(stp2_3, stp2_12);
2333 0 : l[13] = _mm_sub_epi16(stp2_2, stp2_13);
2334 0 : l[14] = _mm_sub_epi16(stp2_1, stp1_14);
2335 0 : l[15] = _mm_sub_epi16(stp2_0, stp1_15);
2336 :
2337 : // Second 1-D inverse transform, performed per 8x16 block
2338 0 : for (i = 0; i < 2; i++) {
2339 : int j;
2340 0 : array_transpose_4X8(l + 8 * i, in);
2341 :
2342 0 : IDCT16_10
2343 :
2344 : // Stage7
2345 0 : in[0] = _mm_add_epi16(stp2_0, stp1_15);
2346 0 : in[1] = _mm_add_epi16(stp2_1, stp1_14);
2347 0 : in[2] = _mm_add_epi16(stp2_2, stp2_13);
2348 0 : in[3] = _mm_add_epi16(stp2_3, stp2_12);
2349 0 : in[4] = _mm_add_epi16(stp2_4, stp2_11);
2350 0 : in[5] = _mm_add_epi16(stp2_5, stp2_10);
2351 0 : in[6] = _mm_add_epi16(stp2_6, stp1_9);
2352 0 : in[7] = _mm_add_epi16(stp2_7, stp1_8);
2353 0 : in[8] = _mm_sub_epi16(stp2_7, stp1_8);
2354 0 : in[9] = _mm_sub_epi16(stp2_6, stp1_9);
2355 0 : in[10] = _mm_sub_epi16(stp2_5, stp2_10);
2356 0 : in[11] = _mm_sub_epi16(stp2_4, stp2_11);
2357 0 : in[12] = _mm_sub_epi16(stp2_3, stp2_12);
2358 0 : in[13] = _mm_sub_epi16(stp2_2, stp2_13);
2359 0 : in[14] = _mm_sub_epi16(stp2_1, stp1_14);
2360 0 : in[15] = _mm_sub_epi16(stp2_0, stp1_15);
2361 :
2362 0 : for (j = 0; j < 16; ++j) {
2363 : // Final rounding and shift
2364 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
2365 0 : in[j] = _mm_srai_epi16(in[j], 6);
2366 0 : RECON_AND_STORE(dest + j * stride, in[j]);
2367 : }
2368 :
2369 0 : dest += 8;
2370 : }
2371 0 : }
2372 :
2373 : #define LOAD_DQCOEFF(reg, input) \
2374 : { \
2375 : reg = load_input_data(input); \
2376 : input += 8; \
2377 : }
2378 :
2379 : #define IDCT32_34 \
2380 : /* Stage1 */ \
2381 : { \
2382 : const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], zero); \
2383 : const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], zero); \
2384 : \
2385 : const __m128i lo_25_7 = _mm_unpacklo_epi16(zero, in[7]); \
2386 : const __m128i hi_25_7 = _mm_unpackhi_epi16(zero, in[7]); \
2387 : \
2388 : const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], zero); \
2389 : const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], zero); \
2390 : \
2391 : const __m128i lo_29_3 = _mm_unpacklo_epi16(zero, in[3]); \
2392 : const __m128i hi_29_3 = _mm_unpackhi_epi16(zero, in[3]); \
2393 : \
2394 : MULTIPLICATION_AND_ADD_2(lo_1_31, hi_1_31, stg1_0, stg1_1, stp1_16, \
2395 : stp1_31); \
2396 : MULTIPLICATION_AND_ADD_2(lo_25_7, hi_25_7, stg1_6, stg1_7, stp1_19, \
2397 : stp1_28); \
2398 : MULTIPLICATION_AND_ADD_2(lo_5_27, hi_5_27, stg1_8, stg1_9, stp1_20, \
2399 : stp1_27); \
2400 : MULTIPLICATION_AND_ADD_2(lo_29_3, hi_29_3, stg1_14, stg1_15, stp1_23, \
2401 : stp1_24); \
2402 : } \
2403 : \
2404 : /* Stage2 */ \
2405 : { \
2406 : const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], zero); \
2407 : const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], zero); \
2408 : \
2409 : const __m128i lo_26_6 = _mm_unpacklo_epi16(zero, in[6]); \
2410 : const __m128i hi_26_6 = _mm_unpackhi_epi16(zero, in[6]); \
2411 : \
2412 : MULTIPLICATION_AND_ADD_2(lo_2_30, hi_2_30, stg2_0, stg2_1, stp2_8, \
2413 : stp2_15); \
2414 : MULTIPLICATION_AND_ADD_2(lo_26_6, hi_26_6, stg2_6, stg2_7, stp2_11, \
2415 : stp2_12); \
2416 : \
2417 : stp2_16 = stp1_16; \
2418 : stp2_19 = stp1_19; \
2419 : \
2420 : stp2_20 = stp1_20; \
2421 : stp2_23 = stp1_23; \
2422 : \
2423 : stp2_24 = stp1_24; \
2424 : stp2_27 = stp1_27; \
2425 : \
2426 : stp2_28 = stp1_28; \
2427 : stp2_31 = stp1_31; \
2428 : } \
2429 : \
2430 : /* Stage3 */ \
2431 : { \
2432 : const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], zero); \
2433 : const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], zero); \
2434 : \
2435 : const __m128i lo_17_30 = _mm_unpacklo_epi16(stp1_16, stp1_31); \
2436 : const __m128i hi_17_30 = _mm_unpackhi_epi16(stp1_16, stp1_31); \
2437 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp1_19, stp1_28); \
2438 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp1_19, stp1_28); \
2439 : \
2440 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp1_20, stp1_27); \
2441 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp1_20, stp1_27); \
2442 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp1_23, stp1_24); \
2443 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp1_23, stp2_24); \
2444 : \
2445 : MULTIPLICATION_AND_ADD_2(lo_4_28, hi_4_28, stg3_0, stg3_1, stp1_4, \
2446 : stp1_7); \
2447 : \
2448 : stp1_8 = stp2_8; \
2449 : stp1_11 = stp2_11; \
2450 : stp1_12 = stp2_12; \
2451 : stp1_15 = stp2_15; \
2452 : \
2453 : MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
2454 : stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18, \
2455 : stp1_29) \
2456 : MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
2457 : stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \
2458 : stp1_25) \
2459 : \
2460 : stp1_16 = stp2_16; \
2461 : stp1_31 = stp2_31; \
2462 : stp1_19 = stp2_19; \
2463 : stp1_20 = stp2_20; \
2464 : stp1_23 = stp2_23; \
2465 : stp1_24 = stp2_24; \
2466 : stp1_27 = stp2_27; \
2467 : stp1_28 = stp2_28; \
2468 : } \
2469 : \
2470 : /* Stage4 */ \
2471 : { \
2472 : const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], zero); \
2473 : const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], zero); \
2474 : \
2475 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp2_15); \
2476 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp2_8, stp2_15); \
2477 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp2_12); \
2478 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp2_11, stp2_12); \
2479 : \
2480 : MULTIPLICATION_AND_ADD_2(lo_0_16, hi_0_16, stg4_0, stg4_1, stp2_0, \
2481 : stp2_1); \
2482 : \
2483 : stp2_4 = stp1_4; \
2484 : stp2_5 = stp1_4; \
2485 : stp2_6 = stp1_7; \
2486 : stp2_7 = stp1_7; \
2487 : \
2488 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
2489 : stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10, \
2490 : stp2_13) \
2491 : \
2492 : stp2_8 = stp1_8; \
2493 : stp2_15 = stp1_15; \
2494 : stp2_11 = stp1_11; \
2495 : stp2_12 = stp1_12; \
2496 : \
2497 : stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
2498 : stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
2499 : stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
2500 : stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
2501 : stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
2502 : stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
2503 : stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
2504 : stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
2505 : \
2506 : stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
2507 : stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
2508 : stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
2509 : stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
2510 : stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
2511 : stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
2512 : stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
2513 : stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
2514 : } \
2515 : \
2516 : /* Stage5 */ \
2517 : { \
2518 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
2519 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
2520 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
2521 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
2522 : \
2523 : const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
2524 : const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
2525 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2526 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2527 : \
2528 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2529 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2530 : \
2531 : stp1_0 = stp2_0; \
2532 : stp1_1 = stp2_1; \
2533 : stp1_2 = stp2_1; \
2534 : stp1_3 = stp2_0; \
2535 : \
2536 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
2537 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
2538 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
2539 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
2540 : \
2541 : tmp0 = _mm_add_epi32(tmp0, rounding); \
2542 : tmp1 = _mm_add_epi32(tmp1, rounding); \
2543 : tmp2 = _mm_add_epi32(tmp2, rounding); \
2544 : tmp3 = _mm_add_epi32(tmp3, rounding); \
2545 : \
2546 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
2547 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
2548 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
2549 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
2550 : \
2551 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
2552 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
2553 : \
2554 : stp1_4 = stp2_4; \
2555 : stp1_7 = stp2_7; \
2556 : \
2557 : stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
2558 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
2559 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
2560 : stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
2561 : stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
2562 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
2563 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
2564 : stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
2565 : \
2566 : stp1_16 = stp2_16; \
2567 : stp1_17 = stp2_17; \
2568 : \
2569 : MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
2570 : stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19, \
2571 : stp1_28) \
2572 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
2573 : stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21, \
2574 : stp1_26) \
2575 : \
2576 : stp1_22 = stp2_22; \
2577 : stp1_23 = stp2_23; \
2578 : stp1_24 = stp2_24; \
2579 : stp1_25 = stp2_25; \
2580 : stp1_30 = stp2_30; \
2581 : stp1_31 = stp2_31; \
2582 : } \
2583 : \
2584 : /* Stage6 */ \
2585 : { \
2586 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
2587 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
2588 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
2589 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
2590 : \
2591 : stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
2592 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
2593 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
2594 : stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
2595 : stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
2596 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
2597 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
2598 : stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
2599 : \
2600 : stp2_8 = stp1_8; \
2601 : stp2_9 = stp1_9; \
2602 : stp2_14 = stp1_14; \
2603 : stp2_15 = stp1_15; \
2604 : \
2605 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
2606 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
2607 : stp2_12) \
2608 : \
2609 : stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
2610 : stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
2611 : stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
2612 : stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
2613 : stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
2614 : stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
2615 : stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
2616 : stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
2617 : \
2618 : stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
2619 : stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
2620 : stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
2621 : stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
2622 : stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
2623 : stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
2624 : stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
2625 : stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
2626 : } \
2627 : \
2628 : /* Stage7 */ \
2629 : { \
2630 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2631 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2632 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2633 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2634 : \
2635 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
2636 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
2637 : const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
2638 : const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
2639 : \
2640 : stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
2641 : stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
2642 : stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
2643 : stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
2644 : stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
2645 : stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
2646 : stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
2647 : stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
2648 : stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
2649 : stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
2650 : stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
2651 : stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
2652 : stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
2653 : stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
2654 : stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
2655 : stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
2656 : \
2657 : stp1_16 = stp2_16; \
2658 : stp1_17 = stp2_17; \
2659 : stp1_18 = stp2_18; \
2660 : stp1_19 = stp2_19; \
2661 : \
2662 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
2663 : stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21, \
2664 : stp1_26) \
2665 : MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
2666 : stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23, \
2667 : stp1_24) \
2668 : \
2669 : stp1_28 = stp2_28; \
2670 : stp1_29 = stp2_29; \
2671 : stp1_30 = stp2_30; \
2672 : stp1_31 = stp2_31; \
2673 : }
2674 :
2675 : #define IDCT32 \
2676 : /* Stage1 */ \
2677 : { \
2678 : const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], in[31]); \
2679 : const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], in[31]); \
2680 : const __m128i lo_17_15 = _mm_unpacklo_epi16(in[17], in[15]); \
2681 : const __m128i hi_17_15 = _mm_unpackhi_epi16(in[17], in[15]); \
2682 : \
2683 : const __m128i lo_9_23 = _mm_unpacklo_epi16(in[9], in[23]); \
2684 : const __m128i hi_9_23 = _mm_unpackhi_epi16(in[9], in[23]); \
2685 : const __m128i lo_25_7 = _mm_unpacklo_epi16(in[25], in[7]); \
2686 : const __m128i hi_25_7 = _mm_unpackhi_epi16(in[25], in[7]); \
2687 : \
2688 : const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], in[27]); \
2689 : const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], in[27]); \
2690 : const __m128i lo_21_11 = _mm_unpacklo_epi16(in[21], in[11]); \
2691 : const __m128i hi_21_11 = _mm_unpackhi_epi16(in[21], in[11]); \
2692 : \
2693 : const __m128i lo_13_19 = _mm_unpacklo_epi16(in[13], in[19]); \
2694 : const __m128i hi_13_19 = _mm_unpackhi_epi16(in[13], in[19]); \
2695 : const __m128i lo_29_3 = _mm_unpacklo_epi16(in[29], in[3]); \
2696 : const __m128i hi_29_3 = _mm_unpackhi_epi16(in[29], in[3]); \
2697 : \
2698 : MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0, \
2699 : stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, stp1_17, \
2700 : stp1_30) \
2701 : MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, stg1_5, \
2702 : stg1_6, stg1_7, stp1_18, stp1_29, stp1_19, stp1_28) \
2703 : MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8, \
2704 : stg1_9, stg1_10, stg1_11, stp1_20, stp1_27, \
2705 : stp1_21, stp1_26) \
2706 : MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12, \
2707 : stg1_13, stg1_14, stg1_15, stp1_22, stp1_25, \
2708 : stp1_23, stp1_24) \
2709 : } \
2710 : \
2711 : /* Stage2 */ \
2712 : { \
2713 : const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], in[30]); \
2714 : const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], in[30]); \
2715 : const __m128i lo_18_14 = _mm_unpacklo_epi16(in[18], in[14]); \
2716 : const __m128i hi_18_14 = _mm_unpackhi_epi16(in[18], in[14]); \
2717 : \
2718 : const __m128i lo_10_22 = _mm_unpacklo_epi16(in[10], in[22]); \
2719 : const __m128i hi_10_22 = _mm_unpackhi_epi16(in[10], in[22]); \
2720 : const __m128i lo_26_6 = _mm_unpacklo_epi16(in[26], in[6]); \
2721 : const __m128i hi_26_6 = _mm_unpackhi_epi16(in[26], in[6]); \
2722 : \
2723 : MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0, \
2724 : stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, \
2725 : stp2_14) \
2726 : MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4, \
2727 : stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, stp2_11, \
2728 : stp2_12) \
2729 : \
2730 : stp2_16 = _mm_add_epi16(stp1_16, stp1_17); \
2731 : stp2_17 = _mm_sub_epi16(stp1_16, stp1_17); \
2732 : stp2_18 = _mm_sub_epi16(stp1_19, stp1_18); \
2733 : stp2_19 = _mm_add_epi16(stp1_19, stp1_18); \
2734 : \
2735 : stp2_20 = _mm_add_epi16(stp1_20, stp1_21); \
2736 : stp2_21 = _mm_sub_epi16(stp1_20, stp1_21); \
2737 : stp2_22 = _mm_sub_epi16(stp1_23, stp1_22); \
2738 : stp2_23 = _mm_add_epi16(stp1_23, stp1_22); \
2739 : \
2740 : stp2_24 = _mm_add_epi16(stp1_24, stp1_25); \
2741 : stp2_25 = _mm_sub_epi16(stp1_24, stp1_25); \
2742 : stp2_26 = _mm_sub_epi16(stp1_27, stp1_26); \
2743 : stp2_27 = _mm_add_epi16(stp1_27, stp1_26); \
2744 : \
2745 : stp2_28 = _mm_add_epi16(stp1_28, stp1_29); \
2746 : stp2_29 = _mm_sub_epi16(stp1_28, stp1_29); \
2747 : stp2_30 = _mm_sub_epi16(stp1_31, stp1_30); \
2748 : stp2_31 = _mm_add_epi16(stp1_31, stp1_30); \
2749 : } \
2750 : \
2751 : /* Stage3 */ \
2752 : { \
2753 : const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], in[28]); \
2754 : const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], in[28]); \
2755 : const __m128i lo_20_12 = _mm_unpacklo_epi16(in[20], in[12]); \
2756 : const __m128i hi_20_12 = _mm_unpackhi_epi16(in[20], in[12]); \
2757 : \
2758 : const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30); \
2759 : const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30); \
2760 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
2761 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
2762 : \
2763 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2764 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2765 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
2766 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
2767 : \
2768 : MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0, \
2769 : stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, \
2770 : stp1_6) \
2771 : \
2772 : stp1_8 = _mm_add_epi16(stp2_8, stp2_9); \
2773 : stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \
2774 : stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
2775 : stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
2776 : stp1_12 = _mm_add_epi16(stp2_12, stp2_13); \
2777 : stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
2778 : stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
2779 : stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
2780 : \
2781 : MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
2782 : stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18, \
2783 : stp1_29) \
2784 : MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
2785 : stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \
2786 : stp1_25) \
2787 : \
2788 : stp1_16 = stp2_16; \
2789 : stp1_31 = stp2_31; \
2790 : stp1_19 = stp2_19; \
2791 : stp1_20 = stp2_20; \
2792 : stp1_23 = stp2_23; \
2793 : stp1_24 = stp2_24; \
2794 : stp1_27 = stp2_27; \
2795 : stp1_28 = stp2_28; \
2796 : } \
2797 : \
2798 : /* Stage4 */ \
2799 : { \
2800 : const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], in[16]); \
2801 : const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], in[16]); \
2802 : const __m128i lo_8_24 = _mm_unpacklo_epi16(in[8], in[24]); \
2803 : const __m128i hi_8_24 = _mm_unpackhi_epi16(in[8], in[24]); \
2804 : \
2805 : const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
2806 : const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
2807 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
2808 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
2809 : \
2810 : MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, stg4_1, \
2811 : stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3) \
2812 : \
2813 : stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
2814 : stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
2815 : stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
2816 : stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
2817 : \
2818 : MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
2819 : stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10, \
2820 : stp2_13) \
2821 : \
2822 : stp2_8 = stp1_8; \
2823 : stp2_15 = stp1_15; \
2824 : stp2_11 = stp1_11; \
2825 : stp2_12 = stp1_12; \
2826 : \
2827 : stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
2828 : stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
2829 : stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
2830 : stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
2831 : stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
2832 : stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
2833 : stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
2834 : stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
2835 : \
2836 : stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
2837 : stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
2838 : stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
2839 : stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
2840 : stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
2841 : stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
2842 : stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
2843 : stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
2844 : } \
2845 : \
2846 : /* Stage5 */ \
2847 : { \
2848 : const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
2849 : const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
2850 : const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
2851 : const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
2852 : \
2853 : const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
2854 : const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
2855 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2856 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2857 : \
2858 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2859 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2860 : \
2861 : stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
2862 : stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
2863 : stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
2864 : stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
2865 : \
2866 : tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
2867 : tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
2868 : tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
2869 : tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
2870 : \
2871 : tmp0 = _mm_add_epi32(tmp0, rounding); \
2872 : tmp1 = _mm_add_epi32(tmp1, rounding); \
2873 : tmp2 = _mm_add_epi32(tmp2, rounding); \
2874 : tmp3 = _mm_add_epi32(tmp3, rounding); \
2875 : \
2876 : tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
2877 : tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
2878 : tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
2879 : tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
2880 : \
2881 : stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
2882 : stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
2883 : \
2884 : stp1_4 = stp2_4; \
2885 : stp1_7 = stp2_7; \
2886 : \
2887 : stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
2888 : stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
2889 : stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
2890 : stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
2891 : stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
2892 : stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
2893 : stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
2894 : stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
2895 : \
2896 : stp1_16 = stp2_16; \
2897 : stp1_17 = stp2_17; \
2898 : \
2899 : MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
2900 : stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19, \
2901 : stp1_28) \
2902 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
2903 : stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21, \
2904 : stp1_26) \
2905 : \
2906 : stp1_22 = stp2_22; \
2907 : stp1_23 = stp2_23; \
2908 : stp1_24 = stp2_24; \
2909 : stp1_25 = stp2_25; \
2910 : stp1_30 = stp2_30; \
2911 : stp1_31 = stp2_31; \
2912 : } \
2913 : \
2914 : /* Stage6 */ \
2915 : { \
2916 : const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
2917 : const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
2918 : const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
2919 : const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
2920 : \
2921 : stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
2922 : stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
2923 : stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
2924 : stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
2925 : stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
2926 : stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
2927 : stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
2928 : stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
2929 : \
2930 : stp2_8 = stp1_8; \
2931 : stp2_9 = stp1_9; \
2932 : stp2_14 = stp1_14; \
2933 : stp2_15 = stp1_15; \
2934 : \
2935 : MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \
2936 : stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \
2937 : stp2_12) \
2938 : \
2939 : stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
2940 : stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
2941 : stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
2942 : stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
2943 : stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
2944 : stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
2945 : stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
2946 : stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
2947 : \
2948 : stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
2949 : stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
2950 : stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
2951 : stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
2952 : stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
2953 : stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
2954 : stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
2955 : stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
2956 : } \
2957 : \
2958 : /* Stage7 */ \
2959 : { \
2960 : const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
2961 : const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
2962 : const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
2963 : const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
2964 : \
2965 : const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
2966 : const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
2967 : const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
2968 : const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
2969 : \
2970 : stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
2971 : stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
2972 : stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
2973 : stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
2974 : stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
2975 : stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
2976 : stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
2977 : stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
2978 : stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
2979 : stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
2980 : stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
2981 : stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
2982 : stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
2983 : stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
2984 : stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
2985 : stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
2986 : \
2987 : stp1_16 = stp2_16; \
2988 : stp1_17 = stp2_17; \
2989 : stp1_18 = stp2_18; \
2990 : stp1_19 = stp2_19; \
2991 : \
2992 : MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
2993 : stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21, \
2994 : stp1_26) \
2995 : MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
2996 : stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23, \
2997 : stp1_24) \
2998 : \
2999 : stp1_28 = stp2_28; \
3000 : stp1_29 = stp2_29; \
3001 : stp1_30 = stp2_30; \
3002 : stp1_31 = stp2_31; \
3003 : }
3004 :
3005 : // Only upper-left 8x8 has non-zero coeff
3006 0 : void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest,
3007 : int stride) {
3008 0 : const __m128i zero = _mm_setzero_si128();
3009 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
3010 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
3011 :
3012 : // idct constants for each stage
3013 0 : const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
3014 0 : const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
3015 0 : const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
3016 0 : const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
3017 0 : const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
3018 0 : const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
3019 0 : const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
3020 0 : const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
3021 :
3022 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
3023 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
3024 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
3025 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
3026 :
3027 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
3028 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
3029 0 : const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
3030 0 : const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
3031 0 : const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
3032 0 : const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
3033 0 : const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
3034 0 : const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
3035 :
3036 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
3037 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
3038 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
3039 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
3040 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
3041 :
3042 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
3043 :
3044 : __m128i in[32], col[32];
3045 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
3046 : stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
3047 : stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
3048 : stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
3049 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
3050 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
3051 : stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
3052 : stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
3053 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
3054 : int i;
3055 :
3056 : // Load input data. Only need to load the top left 8x8 block.
3057 0 : in[0] = load_input_data(input);
3058 0 : in[1] = load_input_data(input + 32);
3059 0 : in[2] = load_input_data(input + 64);
3060 0 : in[3] = load_input_data(input + 96);
3061 0 : in[4] = load_input_data(input + 128);
3062 0 : in[5] = load_input_data(input + 160);
3063 0 : in[6] = load_input_data(input + 192);
3064 0 : in[7] = load_input_data(input + 224);
3065 :
3066 0 : array_transpose_8x8(in, in);
3067 0 : IDCT32_34
3068 :
3069 : // 1_D: Store 32 intermediate results for each 8x32 block.
3070 0 : col[0] = _mm_add_epi16(stp1_0, stp1_31);
3071 0 : col[1] = _mm_add_epi16(stp1_1, stp1_30);
3072 0 : col[2] = _mm_add_epi16(stp1_2, stp1_29);
3073 0 : col[3] = _mm_add_epi16(stp1_3, stp1_28);
3074 0 : col[4] = _mm_add_epi16(stp1_4, stp1_27);
3075 0 : col[5] = _mm_add_epi16(stp1_5, stp1_26);
3076 0 : col[6] = _mm_add_epi16(stp1_6, stp1_25);
3077 0 : col[7] = _mm_add_epi16(stp1_7, stp1_24);
3078 0 : col[8] = _mm_add_epi16(stp1_8, stp1_23);
3079 0 : col[9] = _mm_add_epi16(stp1_9, stp1_22);
3080 0 : col[10] = _mm_add_epi16(stp1_10, stp1_21);
3081 0 : col[11] = _mm_add_epi16(stp1_11, stp1_20);
3082 0 : col[12] = _mm_add_epi16(stp1_12, stp1_19);
3083 0 : col[13] = _mm_add_epi16(stp1_13, stp1_18);
3084 0 : col[14] = _mm_add_epi16(stp1_14, stp1_17);
3085 0 : col[15] = _mm_add_epi16(stp1_15, stp1_16);
3086 0 : col[16] = _mm_sub_epi16(stp1_15, stp1_16);
3087 0 : col[17] = _mm_sub_epi16(stp1_14, stp1_17);
3088 0 : col[18] = _mm_sub_epi16(stp1_13, stp1_18);
3089 0 : col[19] = _mm_sub_epi16(stp1_12, stp1_19);
3090 0 : col[20] = _mm_sub_epi16(stp1_11, stp1_20);
3091 0 : col[21] = _mm_sub_epi16(stp1_10, stp1_21);
3092 0 : col[22] = _mm_sub_epi16(stp1_9, stp1_22);
3093 0 : col[23] = _mm_sub_epi16(stp1_8, stp1_23);
3094 0 : col[24] = _mm_sub_epi16(stp1_7, stp1_24);
3095 0 : col[25] = _mm_sub_epi16(stp1_6, stp1_25);
3096 0 : col[26] = _mm_sub_epi16(stp1_5, stp1_26);
3097 0 : col[27] = _mm_sub_epi16(stp1_4, stp1_27);
3098 0 : col[28] = _mm_sub_epi16(stp1_3, stp1_28);
3099 0 : col[29] = _mm_sub_epi16(stp1_2, stp1_29);
3100 0 : col[30] = _mm_sub_epi16(stp1_1, stp1_30);
3101 0 : col[31] = _mm_sub_epi16(stp1_0, stp1_31);
3102 0 : for (i = 0; i < 4; i++) {
3103 : int j;
3104 : // Transpose 32x8 block to 8x32 block
3105 0 : array_transpose_8x8(col + i * 8, in);
3106 0 : IDCT32_34
3107 :
3108 : // 2_D: Calculate the results and store them to destination.
3109 0 : in[0] = _mm_add_epi16(stp1_0, stp1_31);
3110 0 : in[1] = _mm_add_epi16(stp1_1, stp1_30);
3111 0 : in[2] = _mm_add_epi16(stp1_2, stp1_29);
3112 0 : in[3] = _mm_add_epi16(stp1_3, stp1_28);
3113 0 : in[4] = _mm_add_epi16(stp1_4, stp1_27);
3114 0 : in[5] = _mm_add_epi16(stp1_5, stp1_26);
3115 0 : in[6] = _mm_add_epi16(stp1_6, stp1_25);
3116 0 : in[7] = _mm_add_epi16(stp1_7, stp1_24);
3117 0 : in[8] = _mm_add_epi16(stp1_8, stp1_23);
3118 0 : in[9] = _mm_add_epi16(stp1_9, stp1_22);
3119 0 : in[10] = _mm_add_epi16(stp1_10, stp1_21);
3120 0 : in[11] = _mm_add_epi16(stp1_11, stp1_20);
3121 0 : in[12] = _mm_add_epi16(stp1_12, stp1_19);
3122 0 : in[13] = _mm_add_epi16(stp1_13, stp1_18);
3123 0 : in[14] = _mm_add_epi16(stp1_14, stp1_17);
3124 0 : in[15] = _mm_add_epi16(stp1_15, stp1_16);
3125 0 : in[16] = _mm_sub_epi16(stp1_15, stp1_16);
3126 0 : in[17] = _mm_sub_epi16(stp1_14, stp1_17);
3127 0 : in[18] = _mm_sub_epi16(stp1_13, stp1_18);
3128 0 : in[19] = _mm_sub_epi16(stp1_12, stp1_19);
3129 0 : in[20] = _mm_sub_epi16(stp1_11, stp1_20);
3130 0 : in[21] = _mm_sub_epi16(stp1_10, stp1_21);
3131 0 : in[22] = _mm_sub_epi16(stp1_9, stp1_22);
3132 0 : in[23] = _mm_sub_epi16(stp1_8, stp1_23);
3133 0 : in[24] = _mm_sub_epi16(stp1_7, stp1_24);
3134 0 : in[25] = _mm_sub_epi16(stp1_6, stp1_25);
3135 0 : in[26] = _mm_sub_epi16(stp1_5, stp1_26);
3136 0 : in[27] = _mm_sub_epi16(stp1_4, stp1_27);
3137 0 : in[28] = _mm_sub_epi16(stp1_3, stp1_28);
3138 0 : in[29] = _mm_sub_epi16(stp1_2, stp1_29);
3139 0 : in[30] = _mm_sub_epi16(stp1_1, stp1_30);
3140 0 : in[31] = _mm_sub_epi16(stp1_0, stp1_31);
3141 :
3142 0 : for (j = 0; j < 32; ++j) {
3143 : // Final rounding and shift
3144 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
3145 0 : in[j] = _mm_srai_epi16(in[j], 6);
3146 0 : RECON_AND_STORE(dest + j * stride, in[j]);
3147 : }
3148 :
3149 0 : dest += 8;
3150 : }
3151 0 : }
3152 :
3153 0 : void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest,
3154 : int stride) {
3155 0 : const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
3156 0 : const __m128i final_rounding = _mm_set1_epi16(1 << 5);
3157 0 : const __m128i zero = _mm_setzero_si128();
3158 :
3159 : // idct constants for each stage
3160 0 : const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
3161 0 : const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
3162 0 : const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
3163 0 : const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
3164 0 : const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
3165 0 : const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
3166 0 : const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
3167 0 : const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
3168 0 : const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
3169 0 : const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
3170 0 : const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
3171 0 : const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
3172 0 : const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
3173 0 : const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
3174 0 : const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
3175 0 : const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
3176 :
3177 0 : const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
3178 0 : const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
3179 0 : const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
3180 0 : const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
3181 0 : const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
3182 0 : const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
3183 0 : const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
3184 0 : const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
3185 :
3186 0 : const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
3187 0 : const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
3188 0 : const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
3189 0 : const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
3190 0 : const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
3191 0 : const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
3192 0 : const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
3193 0 : const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
3194 0 : const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
3195 0 : const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
3196 :
3197 0 : const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
3198 0 : const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
3199 0 : const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
3200 0 : const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
3201 0 : const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
3202 0 : const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
3203 0 : const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
3204 :
3205 0 : const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
3206 :
3207 : __m128i in[32], col[128], zero_idx[16];
3208 : __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
3209 : stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
3210 : stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
3211 : stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
3212 : __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
3213 : stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
3214 : stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
3215 : stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
3216 : __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
3217 : int i, j, i32;
3218 :
3219 0 : for (i = 0; i < 4; i++) {
3220 0 : i32 = (i << 5);
3221 : // First 1-D idct
3222 : // Load input data.
3223 0 : LOAD_DQCOEFF(in[0], input);
3224 0 : LOAD_DQCOEFF(in[8], input);
3225 0 : LOAD_DQCOEFF(in[16], input);
3226 0 : LOAD_DQCOEFF(in[24], input);
3227 0 : LOAD_DQCOEFF(in[1], input);
3228 0 : LOAD_DQCOEFF(in[9], input);
3229 0 : LOAD_DQCOEFF(in[17], input);
3230 0 : LOAD_DQCOEFF(in[25], input);
3231 0 : LOAD_DQCOEFF(in[2], input);
3232 0 : LOAD_DQCOEFF(in[10], input);
3233 0 : LOAD_DQCOEFF(in[18], input);
3234 0 : LOAD_DQCOEFF(in[26], input);
3235 0 : LOAD_DQCOEFF(in[3], input);
3236 0 : LOAD_DQCOEFF(in[11], input);
3237 0 : LOAD_DQCOEFF(in[19], input);
3238 0 : LOAD_DQCOEFF(in[27], input);
3239 :
3240 0 : LOAD_DQCOEFF(in[4], input);
3241 0 : LOAD_DQCOEFF(in[12], input);
3242 0 : LOAD_DQCOEFF(in[20], input);
3243 0 : LOAD_DQCOEFF(in[28], input);
3244 0 : LOAD_DQCOEFF(in[5], input);
3245 0 : LOAD_DQCOEFF(in[13], input);
3246 0 : LOAD_DQCOEFF(in[21], input);
3247 0 : LOAD_DQCOEFF(in[29], input);
3248 0 : LOAD_DQCOEFF(in[6], input);
3249 0 : LOAD_DQCOEFF(in[14], input);
3250 0 : LOAD_DQCOEFF(in[22], input);
3251 0 : LOAD_DQCOEFF(in[30], input);
3252 0 : LOAD_DQCOEFF(in[7], input);
3253 0 : LOAD_DQCOEFF(in[15], input);
3254 0 : LOAD_DQCOEFF(in[23], input);
3255 0 : LOAD_DQCOEFF(in[31], input);
3256 :
3257 : // checking if all entries are zero
3258 0 : zero_idx[0] = _mm_or_si128(in[0], in[1]);
3259 0 : zero_idx[1] = _mm_or_si128(in[2], in[3]);
3260 0 : zero_idx[2] = _mm_or_si128(in[4], in[5]);
3261 0 : zero_idx[3] = _mm_or_si128(in[6], in[7]);
3262 0 : zero_idx[4] = _mm_or_si128(in[8], in[9]);
3263 0 : zero_idx[5] = _mm_or_si128(in[10], in[11]);
3264 0 : zero_idx[6] = _mm_or_si128(in[12], in[13]);
3265 0 : zero_idx[7] = _mm_or_si128(in[14], in[15]);
3266 0 : zero_idx[8] = _mm_or_si128(in[16], in[17]);
3267 0 : zero_idx[9] = _mm_or_si128(in[18], in[19]);
3268 0 : zero_idx[10] = _mm_or_si128(in[20], in[21]);
3269 0 : zero_idx[11] = _mm_or_si128(in[22], in[23]);
3270 0 : zero_idx[12] = _mm_or_si128(in[24], in[25]);
3271 0 : zero_idx[13] = _mm_or_si128(in[26], in[27]);
3272 0 : zero_idx[14] = _mm_or_si128(in[28], in[29]);
3273 0 : zero_idx[15] = _mm_or_si128(in[30], in[31]);
3274 :
3275 0 : zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]);
3276 0 : zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]);
3277 0 : zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]);
3278 0 : zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]);
3279 0 : zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]);
3280 0 : zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]);
3281 0 : zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]);
3282 0 : zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]);
3283 :
3284 0 : zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]);
3285 0 : zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]);
3286 0 : zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]);
3287 0 : zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]);
3288 0 : zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]);
3289 0 : zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]);
3290 0 : zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]);
3291 :
3292 0 : if (_mm_movemask_epi8(_mm_cmpeq_epi32(zero_idx[14], zero)) == 0xFFFF) {
3293 0 : col[i32 + 0] = _mm_setzero_si128();
3294 0 : col[i32 + 1] = _mm_setzero_si128();
3295 0 : col[i32 + 2] = _mm_setzero_si128();
3296 0 : col[i32 + 3] = _mm_setzero_si128();
3297 0 : col[i32 + 4] = _mm_setzero_si128();
3298 0 : col[i32 + 5] = _mm_setzero_si128();
3299 0 : col[i32 + 6] = _mm_setzero_si128();
3300 0 : col[i32 + 7] = _mm_setzero_si128();
3301 0 : col[i32 + 8] = _mm_setzero_si128();
3302 0 : col[i32 + 9] = _mm_setzero_si128();
3303 0 : col[i32 + 10] = _mm_setzero_si128();
3304 0 : col[i32 + 11] = _mm_setzero_si128();
3305 0 : col[i32 + 12] = _mm_setzero_si128();
3306 0 : col[i32 + 13] = _mm_setzero_si128();
3307 0 : col[i32 + 14] = _mm_setzero_si128();
3308 0 : col[i32 + 15] = _mm_setzero_si128();
3309 0 : col[i32 + 16] = _mm_setzero_si128();
3310 0 : col[i32 + 17] = _mm_setzero_si128();
3311 0 : col[i32 + 18] = _mm_setzero_si128();
3312 0 : col[i32 + 19] = _mm_setzero_si128();
3313 0 : col[i32 + 20] = _mm_setzero_si128();
3314 0 : col[i32 + 21] = _mm_setzero_si128();
3315 0 : col[i32 + 22] = _mm_setzero_si128();
3316 0 : col[i32 + 23] = _mm_setzero_si128();
3317 0 : col[i32 + 24] = _mm_setzero_si128();
3318 0 : col[i32 + 25] = _mm_setzero_si128();
3319 0 : col[i32 + 26] = _mm_setzero_si128();
3320 0 : col[i32 + 27] = _mm_setzero_si128();
3321 0 : col[i32 + 28] = _mm_setzero_si128();
3322 0 : col[i32 + 29] = _mm_setzero_si128();
3323 0 : col[i32 + 30] = _mm_setzero_si128();
3324 0 : col[i32 + 31] = _mm_setzero_si128();
3325 0 : continue;
3326 : }
3327 :
3328 : // Transpose 32x8 block to 8x32 block
3329 0 : array_transpose_8x8(in, in);
3330 0 : array_transpose_8x8(in + 8, in + 8);
3331 0 : array_transpose_8x8(in + 16, in + 16);
3332 0 : array_transpose_8x8(in + 24, in + 24);
3333 :
3334 0 : IDCT32
3335 :
3336 : // 1_D: Store 32 intermediate results for each 8x32 block.
3337 0 : col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31);
3338 0 : col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30);
3339 0 : col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29);
3340 0 : col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28);
3341 0 : col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27);
3342 0 : col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26);
3343 0 : col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25);
3344 0 : col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24);
3345 0 : col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23);
3346 0 : col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22);
3347 0 : col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21);
3348 0 : col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20);
3349 0 : col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19);
3350 0 : col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18);
3351 0 : col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17);
3352 0 : col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16);
3353 0 : col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16);
3354 0 : col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17);
3355 0 : col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18);
3356 0 : col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19);
3357 0 : col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20);
3358 0 : col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21);
3359 0 : col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22);
3360 0 : col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23);
3361 0 : col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24);
3362 0 : col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25);
3363 0 : col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26);
3364 0 : col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27);
3365 0 : col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28);
3366 0 : col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29);
3367 0 : col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
3368 0 : col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
3369 : }
3370 0 : for (i = 0; i < 4; i++) {
3371 : // Second 1-D idct
3372 0 : j = i << 3;
3373 :
3374 : // Transpose 32x8 block to 8x32 block
3375 0 : array_transpose_8x8(col + j, in);
3376 0 : array_transpose_8x8(col + j + 32, in + 8);
3377 0 : array_transpose_8x8(col + j + 64, in + 16);
3378 0 : array_transpose_8x8(col + j + 96, in + 24);
3379 :
3380 0 : IDCT32
3381 :
3382 : // 2_D: Calculate the results and store them to destination.
3383 0 : in[0] = _mm_add_epi16(stp1_0, stp1_31);
3384 0 : in[1] = _mm_add_epi16(stp1_1, stp1_30);
3385 0 : in[2] = _mm_add_epi16(stp1_2, stp1_29);
3386 0 : in[3] = _mm_add_epi16(stp1_3, stp1_28);
3387 0 : in[4] = _mm_add_epi16(stp1_4, stp1_27);
3388 0 : in[5] = _mm_add_epi16(stp1_5, stp1_26);
3389 0 : in[6] = _mm_add_epi16(stp1_6, stp1_25);
3390 0 : in[7] = _mm_add_epi16(stp1_7, stp1_24);
3391 0 : in[8] = _mm_add_epi16(stp1_8, stp1_23);
3392 0 : in[9] = _mm_add_epi16(stp1_9, stp1_22);
3393 0 : in[10] = _mm_add_epi16(stp1_10, stp1_21);
3394 0 : in[11] = _mm_add_epi16(stp1_11, stp1_20);
3395 0 : in[12] = _mm_add_epi16(stp1_12, stp1_19);
3396 0 : in[13] = _mm_add_epi16(stp1_13, stp1_18);
3397 0 : in[14] = _mm_add_epi16(stp1_14, stp1_17);
3398 0 : in[15] = _mm_add_epi16(stp1_15, stp1_16);
3399 0 : in[16] = _mm_sub_epi16(stp1_15, stp1_16);
3400 0 : in[17] = _mm_sub_epi16(stp1_14, stp1_17);
3401 0 : in[18] = _mm_sub_epi16(stp1_13, stp1_18);
3402 0 : in[19] = _mm_sub_epi16(stp1_12, stp1_19);
3403 0 : in[20] = _mm_sub_epi16(stp1_11, stp1_20);
3404 0 : in[21] = _mm_sub_epi16(stp1_10, stp1_21);
3405 0 : in[22] = _mm_sub_epi16(stp1_9, stp1_22);
3406 0 : in[23] = _mm_sub_epi16(stp1_8, stp1_23);
3407 0 : in[24] = _mm_sub_epi16(stp1_7, stp1_24);
3408 0 : in[25] = _mm_sub_epi16(stp1_6, stp1_25);
3409 0 : in[26] = _mm_sub_epi16(stp1_5, stp1_26);
3410 0 : in[27] = _mm_sub_epi16(stp1_4, stp1_27);
3411 0 : in[28] = _mm_sub_epi16(stp1_3, stp1_28);
3412 0 : in[29] = _mm_sub_epi16(stp1_2, stp1_29);
3413 0 : in[30] = _mm_sub_epi16(stp1_1, stp1_30);
3414 0 : in[31] = _mm_sub_epi16(stp1_0, stp1_31);
3415 :
3416 0 : for (j = 0; j < 32; ++j) {
3417 : // Final rounding and shift
3418 0 : in[j] = _mm_adds_epi16(in[j], final_rounding);
3419 0 : in[j] = _mm_srai_epi16(in[j], 6);
3420 0 : RECON_AND_STORE(dest + j * stride, in[j]);
3421 : }
3422 :
3423 0 : dest += 8;
3424 : }
3425 0 : }
3426 :
3427 0 : void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest,
3428 : int stride) {
3429 : __m128i dc_value;
3430 0 : const __m128i zero = _mm_setzero_si128();
3431 : int a, j;
3432 :
3433 0 : a = (int)dct_const_round_shift(input[0] * cospi_16_64);
3434 0 : a = (int)dct_const_round_shift(a * cospi_16_64);
3435 0 : a = ROUND_POWER_OF_TWO(a, 6);
3436 :
3437 0 : dc_value = _mm_set1_epi16(a);
3438 :
3439 0 : for (j = 0; j < 32; ++j) {
3440 0 : RECON_AND_STORE(dest + 0 + j * stride, dc_value);
3441 0 : RECON_AND_STORE(dest + 8 + j * stride, dc_value);
3442 0 : RECON_AND_STORE(dest + 16 + j * stride, dc_value);
3443 0 : RECON_AND_STORE(dest + 24 + j * stride, dc_value);
3444 : }
3445 0 : }
3446 :
3447 : #if CONFIG_VP9_HIGHBITDEPTH
3448 : static INLINE __m128i clamp_high_sse2(__m128i value, int bd) {
3449 : __m128i ubounded, retval;
3450 : const __m128i zero = _mm_set1_epi16(0);
3451 : const __m128i one = _mm_set1_epi16(1);
3452 : const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
3453 : ubounded = _mm_cmpgt_epi16(value, max);
3454 : retval = _mm_andnot_si128(ubounded, value);
3455 : ubounded = _mm_and_si128(ubounded, max);
3456 : retval = _mm_or_si128(retval, ubounded);
3457 : retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero));
3458 : return retval;
3459 : }
3460 :
3461 : void vpx_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8,
3462 : int stride, int bd) {
3463 : tran_low_t out[4 * 4];
3464 : tran_low_t *outptr = out;
3465 : int i, j;
3466 : __m128i inptr[4];
3467 : __m128i sign_bits[2];
3468 : __m128i temp_mm, min_input, max_input;
3469 : int test;
3470 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
3471 : int optimised_cols = 0;
3472 : const __m128i zero = _mm_set1_epi16(0);
3473 : const __m128i eight = _mm_set1_epi16(8);
3474 : const __m128i max = _mm_set1_epi16(12043);
3475 : const __m128i min = _mm_set1_epi16(-12043);
3476 : // Load input into __m128i
3477 : inptr[0] = _mm_loadu_si128((const __m128i *)input);
3478 : inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4));
3479 : inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8));
3480 : inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12));
3481 :
3482 : // Pack to 16 bits
3483 : inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]);
3484 : inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]);
3485 :
3486 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3487 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3488 : max_input = _mm_cmpgt_epi16(max_input, max);
3489 : min_input = _mm_cmplt_epi16(min_input, min);
3490 : temp_mm = _mm_or_si128(max_input, min_input);
3491 : test = _mm_movemask_epi8(temp_mm);
3492 :
3493 : if (!test) {
3494 : // Do the row transform
3495 : idct4_sse2(inptr);
3496 :
3497 : // Check the min & max values
3498 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3499 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3500 : max_input = _mm_cmpgt_epi16(max_input, max);
3501 : min_input = _mm_cmplt_epi16(min_input, min);
3502 : temp_mm = _mm_or_si128(max_input, min_input);
3503 : test = _mm_movemask_epi8(temp_mm);
3504 :
3505 : if (test) {
3506 : transpose_4x4(inptr);
3507 : sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero);
3508 : sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero);
3509 : inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]);
3510 : inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]);
3511 : inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]);
3512 : inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]);
3513 : _mm_storeu_si128((__m128i *)outptr, inptr[0]);
3514 : _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]);
3515 : _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]);
3516 : _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]);
3517 : } else {
3518 : // Set to use the optimised transform for the column
3519 : optimised_cols = 1;
3520 : }
3521 : } else {
3522 : // Run the un-optimised row transform
3523 : for (i = 0; i < 4; ++i) {
3524 : vpx_highbd_idct4_c(input, outptr, bd);
3525 : input += 4;
3526 : outptr += 4;
3527 : }
3528 : }
3529 :
3530 : if (optimised_cols) {
3531 : idct4_sse2(inptr);
3532 :
3533 : // Final round and shift
3534 : inptr[0] = _mm_add_epi16(inptr[0], eight);
3535 : inptr[1] = _mm_add_epi16(inptr[1], eight);
3536 :
3537 : inptr[0] = _mm_srai_epi16(inptr[0], 4);
3538 : inptr[1] = _mm_srai_epi16(inptr[1], 4);
3539 :
3540 : // Reconstruction and Store
3541 : {
3542 : __m128i d0 = _mm_loadl_epi64((const __m128i *)dest);
3543 : __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2));
3544 : d0 = _mm_unpacklo_epi64(
3545 : d0, _mm_loadl_epi64((const __m128i *)(dest + stride)));
3546 : d2 = _mm_unpacklo_epi64(
3547 : d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3)));
3548 : d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd);
3549 : d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd);
3550 : // store input0
3551 : _mm_storel_epi64((__m128i *)dest, d0);
3552 : // store input1
3553 : d0 = _mm_srli_si128(d0, 8);
3554 : _mm_storel_epi64((__m128i *)(dest + stride), d0);
3555 : // store input2
3556 : _mm_storel_epi64((__m128i *)(dest + stride * 2), d2);
3557 : // store input3
3558 : d2 = _mm_srli_si128(d2, 8);
3559 : _mm_storel_epi64((__m128i *)(dest + stride * 3), d2);
3560 : }
3561 : } else {
3562 : // Run the un-optimised column transform
3563 : tran_low_t temp_in[4], temp_out[4];
3564 : // Columns
3565 : for (i = 0; i < 4; ++i) {
3566 : for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i];
3567 : vpx_highbd_idct4_c(temp_in, temp_out, bd);
3568 : for (j = 0; j < 4; ++j) {
3569 : dest[j * stride + i] = highbd_clip_pixel_add(
3570 : dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
3571 : }
3572 : }
3573 : }
3574 : }
3575 :
3576 : void vpx_highbd_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest8,
3577 : int stride, int bd) {
3578 : tran_low_t out[8 * 8];
3579 : tran_low_t *outptr = out;
3580 : int i, j, test;
3581 : __m128i inptr[8];
3582 : __m128i min_input, max_input, temp1, temp2, sign_bits;
3583 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
3584 : const __m128i zero = _mm_set1_epi16(0);
3585 : const __m128i sixteen = _mm_set1_epi16(16);
3586 : const __m128i max = _mm_set1_epi16(6201);
3587 : const __m128i min = _mm_set1_epi16(-6201);
3588 : int optimised_cols = 0;
3589 :
3590 : // Load input into __m128i & pack to 16 bits
3591 : for (i = 0; i < 8; i++) {
3592 : temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
3593 : temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
3594 : inptr[i] = _mm_packs_epi32(temp1, temp2);
3595 : }
3596 :
3597 : // Find the min & max for the row transform
3598 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3599 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3600 : for (i = 2; i < 8; i++) {
3601 : max_input = _mm_max_epi16(max_input, inptr[i]);
3602 : min_input = _mm_min_epi16(min_input, inptr[i]);
3603 : }
3604 : max_input = _mm_cmpgt_epi16(max_input, max);
3605 : min_input = _mm_cmplt_epi16(min_input, min);
3606 : temp1 = _mm_or_si128(max_input, min_input);
3607 : test = _mm_movemask_epi8(temp1);
3608 :
3609 : if (!test) {
3610 : // Do the row transform
3611 : idct8_sse2(inptr);
3612 :
3613 : // Find the min & max for the column transform
3614 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3615 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3616 : for (i = 2; i < 8; i++) {
3617 : max_input = _mm_max_epi16(max_input, inptr[i]);
3618 : min_input = _mm_min_epi16(min_input, inptr[i]);
3619 : }
3620 : max_input = _mm_cmpgt_epi16(max_input, max);
3621 : min_input = _mm_cmplt_epi16(min_input, min);
3622 : temp1 = _mm_or_si128(max_input, min_input);
3623 : test = _mm_movemask_epi8(temp1);
3624 :
3625 : if (test) {
3626 : array_transpose_8x8(inptr, inptr);
3627 : for (i = 0; i < 8; i++) {
3628 : sign_bits = _mm_cmplt_epi16(inptr[i], zero);
3629 : temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
3630 : temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
3631 : _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
3632 : _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
3633 : }
3634 : } else {
3635 : // Set to use the optimised transform for the column
3636 : optimised_cols = 1;
3637 : }
3638 : } else {
3639 : // Run the un-optimised row transform
3640 : for (i = 0; i < 8; ++i) {
3641 : vpx_highbd_idct8_c(input, outptr, bd);
3642 : input += 8;
3643 : outptr += 8;
3644 : }
3645 : }
3646 :
3647 : if (optimised_cols) {
3648 : idct8_sse2(inptr);
3649 :
3650 : // Final round & shift and Reconstruction and Store
3651 : {
3652 : __m128i d[8];
3653 : for (i = 0; i < 8; i++) {
3654 : inptr[i] = _mm_add_epi16(inptr[i], sixteen);
3655 : d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i));
3656 : inptr[i] = _mm_srai_epi16(inptr[i], 5);
3657 : d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
3658 : // Store
3659 : _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]);
3660 : }
3661 : }
3662 : } else {
3663 : // Run the un-optimised column transform
3664 : tran_low_t temp_in[8], temp_out[8];
3665 : for (i = 0; i < 8; ++i) {
3666 : for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i];
3667 : vpx_highbd_idct8_c(temp_in, temp_out, bd);
3668 : for (j = 0; j < 8; ++j) {
3669 : dest[j * stride + i] = highbd_clip_pixel_add(
3670 : dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
3671 : }
3672 : }
3673 : }
3674 : }
3675 :
3676 : void vpx_highbd_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest8,
3677 : int stride, int bd) {
3678 : tran_low_t out[8 * 8] = { 0 };
3679 : tran_low_t *outptr = out;
3680 : int i, j, test;
3681 : __m128i inptr[8];
3682 : __m128i min_input, max_input, temp1, temp2, sign_bits;
3683 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
3684 : const __m128i zero = _mm_set1_epi16(0);
3685 : const __m128i sixteen = _mm_set1_epi16(16);
3686 : const __m128i max = _mm_set1_epi16(6201);
3687 : const __m128i min = _mm_set1_epi16(-6201);
3688 : int optimised_cols = 0;
3689 :
3690 : // Load input into __m128i & pack to 16 bits
3691 : for (i = 0; i < 8; i++) {
3692 : temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
3693 : temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
3694 : inptr[i] = _mm_packs_epi32(temp1, temp2);
3695 : }
3696 :
3697 : // Find the min & max for the row transform
3698 : // only first 4 row has non-zero coefs
3699 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3700 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3701 : for (i = 2; i < 4; i++) {
3702 : max_input = _mm_max_epi16(max_input, inptr[i]);
3703 : min_input = _mm_min_epi16(min_input, inptr[i]);
3704 : }
3705 : max_input = _mm_cmpgt_epi16(max_input, max);
3706 : min_input = _mm_cmplt_epi16(min_input, min);
3707 : temp1 = _mm_or_si128(max_input, min_input);
3708 : test = _mm_movemask_epi8(temp1);
3709 :
3710 : if (!test) {
3711 : // Do the row transform
3712 : idct8_sse2(inptr);
3713 :
3714 : // Find the min & max for the column transform
3715 : // N.B. Only first 4 cols contain non-zero coeffs
3716 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3717 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3718 : for (i = 2; i < 8; i++) {
3719 : max_input = _mm_max_epi16(max_input, inptr[i]);
3720 : min_input = _mm_min_epi16(min_input, inptr[i]);
3721 : }
3722 : max_input = _mm_cmpgt_epi16(max_input, max);
3723 : min_input = _mm_cmplt_epi16(min_input, min);
3724 : temp1 = _mm_or_si128(max_input, min_input);
3725 : test = _mm_movemask_epi8(temp1);
3726 :
3727 : if (test) {
3728 : // Use fact only first 4 rows contain non-zero coeffs
3729 : array_transpose_4X8(inptr, inptr);
3730 : for (i = 0; i < 4; i++) {
3731 : sign_bits = _mm_cmplt_epi16(inptr[i], zero);
3732 : temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
3733 : temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
3734 : _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
3735 : _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
3736 : }
3737 : } else {
3738 : // Set to use the optimised transform for the column
3739 : optimised_cols = 1;
3740 : }
3741 : } else {
3742 : // Run the un-optimised row transform
3743 : for (i = 0; i < 4; ++i) {
3744 : vpx_highbd_idct8_c(input, outptr, bd);
3745 : input += 8;
3746 : outptr += 8;
3747 : }
3748 : }
3749 :
3750 : if (optimised_cols) {
3751 : idct8_sse2(inptr);
3752 :
3753 : // Final round & shift and Reconstruction and Store
3754 : {
3755 : __m128i d[8];
3756 : for (i = 0; i < 8; i++) {
3757 : inptr[i] = _mm_add_epi16(inptr[i], sixteen);
3758 : d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i));
3759 : inptr[i] = _mm_srai_epi16(inptr[i], 5);
3760 : d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
3761 : // Store
3762 : _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]);
3763 : }
3764 : }
3765 : } else {
3766 : // Run the un-optimised column transform
3767 : tran_low_t temp_in[8], temp_out[8];
3768 : for (i = 0; i < 8; ++i) {
3769 : for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i];
3770 : vpx_highbd_idct8_c(temp_in, temp_out, bd);
3771 : for (j = 0; j < 8; ++j) {
3772 : dest[j * stride + i] = highbd_clip_pixel_add(
3773 : dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
3774 : }
3775 : }
3776 : }
3777 : }
3778 :
3779 : void vpx_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest8,
3780 : int stride, int bd) {
3781 : tran_low_t out[16 * 16];
3782 : tran_low_t *outptr = out;
3783 : int i, j, test;
3784 : __m128i inptr[32];
3785 : __m128i min_input, max_input, temp1, temp2, sign_bits;
3786 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
3787 : const __m128i zero = _mm_set1_epi16(0);
3788 : const __m128i rounding = _mm_set1_epi16(32);
3789 : const __m128i max = _mm_set1_epi16(3155);
3790 : const __m128i min = _mm_set1_epi16(-3155);
3791 : int optimised_cols = 0;
3792 :
3793 : // Load input into __m128i & pack to 16 bits
3794 : for (i = 0; i < 16; i++) {
3795 : temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
3796 : temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
3797 : inptr[i] = _mm_packs_epi32(temp1, temp2);
3798 : temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
3799 : temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
3800 : inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
3801 : }
3802 :
3803 : // Find the min & max for the row transform
3804 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3805 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3806 : for (i = 2; i < 32; i++) {
3807 : max_input = _mm_max_epi16(max_input, inptr[i]);
3808 : min_input = _mm_min_epi16(min_input, inptr[i]);
3809 : }
3810 : max_input = _mm_cmpgt_epi16(max_input, max);
3811 : min_input = _mm_cmplt_epi16(min_input, min);
3812 : temp1 = _mm_or_si128(max_input, min_input);
3813 : test = _mm_movemask_epi8(temp1);
3814 :
3815 : if (!test) {
3816 : // Do the row transform
3817 : idct16_sse2(inptr, inptr + 16);
3818 :
3819 : // Find the min & max for the column transform
3820 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3821 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3822 : for (i = 2; i < 32; i++) {
3823 : max_input = _mm_max_epi16(max_input, inptr[i]);
3824 : min_input = _mm_min_epi16(min_input, inptr[i]);
3825 : }
3826 : max_input = _mm_cmpgt_epi16(max_input, max);
3827 : min_input = _mm_cmplt_epi16(min_input, min);
3828 : temp1 = _mm_or_si128(max_input, min_input);
3829 : test = _mm_movemask_epi8(temp1);
3830 :
3831 : if (test) {
3832 : array_transpose_16x16(inptr, inptr + 16);
3833 : for (i = 0; i < 16; i++) {
3834 : sign_bits = _mm_cmplt_epi16(inptr[i], zero);
3835 : temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
3836 : temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
3837 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
3838 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
3839 : sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
3840 : temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
3841 : temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
3842 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
3843 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
3844 : }
3845 : } else {
3846 : // Set to use the optimised transform for the column
3847 : optimised_cols = 1;
3848 : }
3849 : } else {
3850 : // Run the un-optimised row transform
3851 : for (i = 0; i < 16; ++i) {
3852 : vpx_highbd_idct16_c(input, outptr, bd);
3853 : input += 16;
3854 : outptr += 16;
3855 : }
3856 : }
3857 :
3858 : if (optimised_cols) {
3859 : idct16_sse2(inptr, inptr + 16);
3860 :
3861 : // Final round & shift and Reconstruction and Store
3862 : {
3863 : __m128i d[2];
3864 : for (i = 0; i < 16; i++) {
3865 : inptr[i] = _mm_add_epi16(inptr[i], rounding);
3866 : inptr[i + 16] = _mm_add_epi16(inptr[i + 16], rounding);
3867 : d[0] = _mm_loadu_si128((const __m128i *)(dest + stride * i));
3868 : d[1] = _mm_loadu_si128((const __m128i *)(dest + stride * i + 8));
3869 : inptr[i] = _mm_srai_epi16(inptr[i], 6);
3870 : inptr[i + 16] = _mm_srai_epi16(inptr[i + 16], 6);
3871 : d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i]), bd);
3872 : d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i + 16]), bd);
3873 : // Store
3874 : _mm_storeu_si128((__m128i *)(dest + stride * i), d[0]);
3875 : _mm_storeu_si128((__m128i *)(dest + stride * i + 8), d[1]);
3876 : }
3877 : }
3878 : } else {
3879 : // Run the un-optimised column transform
3880 : tran_low_t temp_in[16], temp_out[16];
3881 : for (i = 0; i < 16; ++i) {
3882 : for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i];
3883 : vpx_highbd_idct16_c(temp_in, temp_out, bd);
3884 : for (j = 0; j < 16; ++j) {
3885 : dest[j * stride + i] = highbd_clip_pixel_add(
3886 : dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
3887 : }
3888 : }
3889 : }
3890 : }
3891 :
3892 : void vpx_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
3893 : int stride, int bd) {
3894 : tran_low_t out[16 * 16] = { 0 };
3895 : tran_low_t *outptr = out;
3896 : int i, j, test;
3897 : __m128i inptr[32];
3898 : __m128i min_input, max_input, temp1, temp2, sign_bits;
3899 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
3900 : const __m128i zero = _mm_set1_epi16(0);
3901 : const __m128i rounding = _mm_set1_epi16(32);
3902 : const __m128i max = _mm_set1_epi16(3155);
3903 : const __m128i min = _mm_set1_epi16(-3155);
3904 : int optimised_cols = 0;
3905 :
3906 : // Load input into __m128i & pack to 16 bits
3907 : for (i = 0; i < 16; i++) {
3908 : temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
3909 : temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
3910 : inptr[i] = _mm_packs_epi32(temp1, temp2);
3911 : temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
3912 : temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
3913 : inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
3914 : }
3915 :
3916 : // Find the min & max for the row transform
3917 : // Since all non-zero dct coefficients are in upper-left 4x4 area,
3918 : // we only need to consider first 4 rows here.
3919 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3920 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3921 : for (i = 2; i < 4; i++) {
3922 : max_input = _mm_max_epi16(max_input, inptr[i]);
3923 : min_input = _mm_min_epi16(min_input, inptr[i]);
3924 : }
3925 : max_input = _mm_cmpgt_epi16(max_input, max);
3926 : min_input = _mm_cmplt_epi16(min_input, min);
3927 : temp1 = _mm_or_si128(max_input, min_input);
3928 : test = _mm_movemask_epi8(temp1);
3929 :
3930 : if (!test) {
3931 : // Do the row transform (N.B. This transposes inptr)
3932 : idct16_sse2(inptr, inptr + 16);
3933 :
3934 : // Find the min & max for the column transform
3935 : // N.B. Only first 4 cols contain non-zero coeffs
3936 : max_input = _mm_max_epi16(inptr[0], inptr[1]);
3937 : min_input = _mm_min_epi16(inptr[0], inptr[1]);
3938 : for (i = 2; i < 16; i++) {
3939 : max_input = _mm_max_epi16(max_input, inptr[i]);
3940 : min_input = _mm_min_epi16(min_input, inptr[i]);
3941 : }
3942 : max_input = _mm_cmpgt_epi16(max_input, max);
3943 : min_input = _mm_cmplt_epi16(min_input, min);
3944 : temp1 = _mm_or_si128(max_input, min_input);
3945 : test = _mm_movemask_epi8(temp1);
3946 :
3947 : if (test) {
3948 : // Use fact only first 4 rows contain non-zero coeffs
3949 : array_transpose_8x8(inptr, inptr);
3950 : array_transpose_8x8(inptr + 8, inptr + 16);
3951 : for (i = 0; i < 4; i++) {
3952 : sign_bits = _mm_cmplt_epi16(inptr[i], zero);
3953 : temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
3954 : temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
3955 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
3956 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
3957 : sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
3958 : temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
3959 : temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
3960 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
3961 : _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
3962 : }
3963 : } else {
3964 : // Set to use the optimised transform for the column
3965 : optimised_cols = 1;
3966 : }
3967 : } else {
3968 : // Run the un-optimised row transform
3969 : for (i = 0; i < 4; ++i) {
3970 : vpx_highbd_idct16_c(input, outptr, bd);
3971 : input += 16;
3972 : outptr += 16;
3973 : }
3974 : }
3975 :
3976 : if (optimised_cols) {
3977 : idct16_sse2(inptr, inptr + 16);
3978 :
3979 : // Final round & shift and Reconstruction and Store
3980 : {
3981 : __m128i d[2];
3982 : for (i = 0; i < 16; i++) {
3983 : inptr[i] = _mm_add_epi16(inptr[i], rounding);
3984 : inptr[i + 16] = _mm_add_epi16(inptr[i + 16], rounding);
3985 : d[0] = _mm_loadu_si128((const __m128i *)(dest + stride * i));
3986 : d[1] = _mm_loadu_si128((const __m128i *)(dest + stride * i + 8));
3987 : inptr[i] = _mm_srai_epi16(inptr[i], 6);
3988 : inptr[i + 16] = _mm_srai_epi16(inptr[i + 16], 6);
3989 : d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i]), bd);
3990 : d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i + 16]), bd);
3991 : // Store
3992 : _mm_storeu_si128((__m128i *)(dest + stride * i), d[0]);
3993 : _mm_storeu_si128((__m128i *)(dest + stride * i + 8), d[1]);
3994 : }
3995 : }
3996 : } else {
3997 : // Run the un-optimised column transform
3998 : tran_low_t temp_in[16], temp_out[16];
3999 : for (i = 0; i < 16; ++i) {
4000 : for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i];
4001 : vpx_highbd_idct16_c(temp_in, temp_out, bd);
4002 : for (j = 0; j < 16; ++j) {
4003 : dest[j * stride + i] = highbd_clip_pixel_add(
4004 : dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
4005 : }
4006 : }
4007 : }
4008 : }
4009 :
4010 : void vpx_highbd_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest8,
4011 : int stride, int bd) {
4012 : __m128i dc_value, d;
4013 : const __m128i zero = _mm_setzero_si128();
4014 : const __m128i one = _mm_set1_epi16(1);
4015 : const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
4016 : int a, i, j;
4017 : uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
4018 : tran_low_t out;
4019 :
4020 : out = dct_const_round_shift(input[0] * cospi_16_64);
4021 : out = dct_const_round_shift(out * cospi_16_64);
4022 : a = ROUND_POWER_OF_TWO(out, 6);
4023 :
4024 : d = _mm_set1_epi32(a);
4025 : dc_value = _mm_packs_epi32(d, d);
4026 : for (i = 0; i < 32; ++i) {
4027 : for (j = 0; j < 4; ++j) {
4028 : d = _mm_loadu_si128((const __m128i *)(&dest[j * 8]));
4029 : d = _mm_adds_epi16(d, dc_value);
4030 : d = _mm_max_epi16(d, zero);
4031 : d = _mm_min_epi16(d, max);
4032 : _mm_storeu_si128((__m128i *)(&dest[j * 8]), d);
4033 : }
4034 : dest += stride;
4035 : }
4036 : }
4037 : #endif // CONFIG_VP9_HIGHBITDEPTH
|
