Line data Source code
1 : /*
2 : * Copyright (c) 2012 The WebM project authors. All Rights Reserved.
3 : *
4 : * Use of this source code is governed by a BSD-style license
5 : * that can be found in the LICENSE file in the root of the source
6 : * tree. An additional intellectual property rights grant can be found
7 : * in the file PATENTS. All contributing project authors may
8 : * be found in the AUTHORS file in the root of the source tree.
9 : */
10 :
11 : #include <assert.h>
12 : #include <emmintrin.h> // SSE2
13 :
14 : #include "./vp9_rtcd.h"
15 : #include "./vpx_dsp_rtcd.h"
16 : #include "vpx_dsp/txfm_common.h"
17 : #include "vpx_dsp/x86/fwd_txfm_sse2.h"
18 : #include "vpx_dsp/x86/txfm_common_sse2.h"
19 : #include "vpx_ports/mem.h"
20 :
21 0 : static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
22 : int stride) {
23 0 : const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
24 0 : const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
25 : __m128i mask;
26 :
27 0 : in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
28 0 : in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
29 0 : in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
30 0 : in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
31 :
32 0 : in[0] = _mm_slli_epi16(in[0], 4);
33 0 : in[1] = _mm_slli_epi16(in[1], 4);
34 0 : in[2] = _mm_slli_epi16(in[2], 4);
35 0 : in[3] = _mm_slli_epi16(in[3], 4);
36 :
37 0 : mask = _mm_cmpeq_epi16(in[0], k__nonzero_bias_a);
38 0 : in[0] = _mm_add_epi16(in[0], mask);
39 0 : in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b);
40 0 : }
41 :
42 0 : static INLINE void write_buffer_4x4(tran_low_t *output, __m128i *res) {
43 0 : const __m128i kOne = _mm_set1_epi16(1);
44 0 : __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]);
45 0 : __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]);
46 0 : __m128i out01 = _mm_add_epi16(in01, kOne);
47 0 : __m128i out23 = _mm_add_epi16(in23, kOne);
48 0 : out01 = _mm_srai_epi16(out01, 2);
49 0 : out23 = _mm_srai_epi16(out23, 2);
50 0 : store_output(&out01, (output + 0 * 8));
51 0 : store_output(&out23, (output + 1 * 8));
52 0 : }
53 :
54 0 : static INLINE void transpose_4x4(__m128i *res) {
55 : // Combine and transpose
56 : // 00 01 02 03 20 21 22 23
57 : // 10 11 12 13 30 31 32 33
58 0 : const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
59 0 : const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
60 :
61 : // 00 10 01 11 02 12 03 13
62 : // 20 30 21 31 22 32 23 33
63 0 : res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1);
64 0 : res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1);
65 :
66 : // 00 10 20 30 01 11 21 31
67 : // 02 12 22 32 03 13 23 33
68 : // only use the first 4 16-bit integers
69 0 : res[1] = _mm_unpackhi_epi64(res[0], res[0]);
70 0 : res[3] = _mm_unpackhi_epi64(res[2], res[2]);
71 0 : }
72 :
73 0 : static void fdct4_sse2(__m128i *in) {
74 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
75 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
76 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
77 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
78 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
79 :
80 : __m128i u[4], v[4];
81 0 : u[0] = _mm_unpacklo_epi16(in[0], in[1]);
82 0 : u[1] = _mm_unpacklo_epi16(in[3], in[2]);
83 :
84 0 : v[0] = _mm_add_epi16(u[0], u[1]);
85 0 : v[1] = _mm_sub_epi16(u[0], u[1]);
86 :
87 0 : u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16); // 0
88 0 : u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16); // 2
89 0 : u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24); // 1
90 0 : u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08); // 3
91 :
92 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
93 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
94 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
95 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
96 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
97 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
98 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
99 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
100 :
101 0 : in[0] = _mm_packs_epi32(u[0], u[1]);
102 0 : in[1] = _mm_packs_epi32(u[2], u[3]);
103 0 : transpose_4x4(in);
104 0 : }
105 :
106 0 : static void fadst4_sse2(__m128i *in) {
107 0 : const __m128i k__sinpi_p01_p02 = pair_set_epi16(sinpi_1_9, sinpi_2_9);
108 0 : const __m128i k__sinpi_p04_m01 = pair_set_epi16(sinpi_4_9, -sinpi_1_9);
109 0 : const __m128i k__sinpi_p03_p04 = pair_set_epi16(sinpi_3_9, sinpi_4_9);
110 0 : const __m128i k__sinpi_m03_p02 = pair_set_epi16(-sinpi_3_9, sinpi_2_9);
111 0 : const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
112 0 : const __m128i kZero = _mm_set1_epi16(0);
113 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
114 : __m128i u[8], v[8];
115 0 : __m128i in7 = _mm_add_epi16(in[0], in[1]);
116 :
117 0 : u[0] = _mm_unpacklo_epi16(in[0], in[1]);
118 0 : u[1] = _mm_unpacklo_epi16(in[2], in[3]);
119 0 : u[2] = _mm_unpacklo_epi16(in7, kZero);
120 0 : u[3] = _mm_unpacklo_epi16(in[2], kZero);
121 0 : u[4] = _mm_unpacklo_epi16(in[3], kZero);
122 :
123 0 : v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02); // s0 + s2
124 0 : v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04); // s4 + s5
125 0 : v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x1
126 0 : v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01); // s1 - s3
127 0 : v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02); // -s4 + s6
128 0 : v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s4
129 0 : v[6] = _mm_madd_epi16(u[4], k__sinpi_p03_p03);
130 :
131 0 : u[0] = _mm_add_epi32(v[0], v[1]);
132 0 : u[1] = _mm_sub_epi32(v[2], v[6]);
133 0 : u[2] = _mm_add_epi32(v[3], v[4]);
134 0 : u[3] = _mm_sub_epi32(u[2], u[0]);
135 0 : u[4] = _mm_slli_epi32(v[5], 2);
136 0 : u[5] = _mm_sub_epi32(u[4], v[5]);
137 0 : u[6] = _mm_add_epi32(u[3], u[5]);
138 :
139 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
140 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
141 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
142 0 : v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
143 :
144 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
145 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
146 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
147 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
148 :
149 0 : in[0] = _mm_packs_epi32(u[0], u[2]);
150 0 : in[1] = _mm_packs_epi32(u[1], u[3]);
151 0 : transpose_4x4(in);
152 0 : }
153 :
154 0 : void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride,
155 : int tx_type) {
156 : __m128i in[4];
157 :
158 0 : switch (tx_type) {
159 0 : case DCT_DCT: vpx_fdct4x4_sse2(input, output, stride); break;
160 : case ADST_DCT:
161 0 : load_buffer_4x4(input, in, stride);
162 0 : fadst4_sse2(in);
163 0 : fdct4_sse2(in);
164 0 : write_buffer_4x4(output, in);
165 0 : break;
166 : case DCT_ADST:
167 0 : load_buffer_4x4(input, in, stride);
168 0 : fdct4_sse2(in);
169 0 : fadst4_sse2(in);
170 0 : write_buffer_4x4(output, in);
171 0 : break;
172 : case ADST_ADST:
173 0 : load_buffer_4x4(input, in, stride);
174 0 : fadst4_sse2(in);
175 0 : fadst4_sse2(in);
176 0 : write_buffer_4x4(output, in);
177 0 : break;
178 0 : default: assert(0); break;
179 : }
180 0 : }
181 :
182 0 : void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride,
183 : int16_t *coeff_ptr, intptr_t n_coeffs,
184 : int skip_block, const int16_t *zbin_ptr,
185 : const int16_t *round_ptr, const int16_t *quant_ptr,
186 : const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
187 : int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
188 : uint16_t *eob_ptr, const int16_t *scan_ptr,
189 : const int16_t *iscan_ptr) {
190 : __m128i zero;
191 : int pass;
192 : // Constants
193 : // When we use them, in one case, they are all the same. In all others
194 : // it's a pair of them that we need to repeat four times. This is done
195 : // by constructing the 32 bit constant corresponding to that pair.
196 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
197 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
198 0 : const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
199 0 : const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
200 0 : const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
201 0 : const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
202 0 : const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
203 0 : const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
204 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
205 : // Load input
206 0 : __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
207 0 : __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
208 0 : __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
209 0 : __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
210 0 : __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
211 0 : __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
212 0 : __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
213 0 : __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
214 : __m128i *in[8];
215 0 : int index = 0;
216 :
217 : (void)scan_ptr;
218 : (void)zbin_ptr;
219 : (void)quant_shift_ptr;
220 : (void)coeff_ptr;
221 :
222 : // Pre-condition input (shift by two)
223 0 : in0 = _mm_slli_epi16(in0, 2);
224 0 : in1 = _mm_slli_epi16(in1, 2);
225 0 : in2 = _mm_slli_epi16(in2, 2);
226 0 : in3 = _mm_slli_epi16(in3, 2);
227 0 : in4 = _mm_slli_epi16(in4, 2);
228 0 : in5 = _mm_slli_epi16(in5, 2);
229 0 : in6 = _mm_slli_epi16(in6, 2);
230 0 : in7 = _mm_slli_epi16(in7, 2);
231 :
232 0 : in[0] = &in0;
233 0 : in[1] = &in1;
234 0 : in[2] = &in2;
235 0 : in[3] = &in3;
236 0 : in[4] = &in4;
237 0 : in[5] = &in5;
238 0 : in[6] = &in6;
239 0 : in[7] = &in7;
240 :
241 : // We do two passes, first the columns, then the rows. The results of the
242 : // first pass are transposed so that the same column code can be reused. The
243 : // results of the second pass are also transposed so that the rows (processed
244 : // as columns) are put back in row positions.
245 0 : for (pass = 0; pass < 2; pass++) {
246 : // To store results of each pass before the transpose.
247 : __m128i res0, res1, res2, res3, res4, res5, res6, res7;
248 : // Add/subtract
249 0 : const __m128i q0 = _mm_add_epi16(in0, in7);
250 0 : const __m128i q1 = _mm_add_epi16(in1, in6);
251 0 : const __m128i q2 = _mm_add_epi16(in2, in5);
252 0 : const __m128i q3 = _mm_add_epi16(in3, in4);
253 0 : const __m128i q4 = _mm_sub_epi16(in3, in4);
254 0 : const __m128i q5 = _mm_sub_epi16(in2, in5);
255 0 : const __m128i q6 = _mm_sub_epi16(in1, in6);
256 0 : const __m128i q7 = _mm_sub_epi16(in0, in7);
257 : // Work on first four results
258 : {
259 : // Add/subtract
260 0 : const __m128i r0 = _mm_add_epi16(q0, q3);
261 0 : const __m128i r1 = _mm_add_epi16(q1, q2);
262 0 : const __m128i r2 = _mm_sub_epi16(q1, q2);
263 0 : const __m128i r3 = _mm_sub_epi16(q0, q3);
264 : // Interleave to do the multiply by constants which gets us into 32bits
265 0 : const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
266 0 : const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
267 0 : const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
268 0 : const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
269 0 : const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
270 0 : const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
271 0 : const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
272 0 : const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
273 0 : const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
274 0 : const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
275 0 : const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
276 0 : const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
277 : // dct_const_round_shift
278 0 : const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
279 0 : const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
280 0 : const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
281 0 : const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
282 0 : const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
283 0 : const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
284 0 : const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
285 0 : const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
286 0 : const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
287 0 : const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
288 0 : const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
289 0 : const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
290 0 : const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
291 0 : const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
292 0 : const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
293 0 : const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
294 : // Combine
295 0 : res0 = _mm_packs_epi32(w0, w1);
296 0 : res4 = _mm_packs_epi32(w2, w3);
297 0 : res2 = _mm_packs_epi32(w4, w5);
298 0 : res6 = _mm_packs_epi32(w6, w7);
299 : }
300 : // Work on next four results
301 : {
302 : // Interleave to do the multiply by constants which gets us into 32bits
303 0 : const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
304 0 : const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
305 0 : const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
306 0 : const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
307 0 : const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
308 0 : const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
309 : // dct_const_round_shift
310 0 : const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
311 0 : const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
312 0 : const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
313 0 : const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
314 0 : const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
315 0 : const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
316 0 : const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
317 0 : const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
318 : // Combine
319 0 : const __m128i r0 = _mm_packs_epi32(s0, s1);
320 0 : const __m128i r1 = _mm_packs_epi32(s2, s3);
321 : // Add/subtract
322 0 : const __m128i x0 = _mm_add_epi16(q4, r0);
323 0 : const __m128i x1 = _mm_sub_epi16(q4, r0);
324 0 : const __m128i x2 = _mm_sub_epi16(q7, r1);
325 0 : const __m128i x3 = _mm_add_epi16(q7, r1);
326 : // Interleave to do the multiply by constants which gets us into 32bits
327 0 : const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
328 0 : const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
329 0 : const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
330 0 : const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
331 0 : const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
332 0 : const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
333 0 : const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
334 0 : const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
335 0 : const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
336 0 : const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
337 0 : const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
338 0 : const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
339 : // dct_const_round_shift
340 0 : const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
341 0 : const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
342 0 : const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
343 0 : const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
344 0 : const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
345 0 : const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
346 0 : const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
347 0 : const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
348 0 : const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
349 0 : const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
350 0 : const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
351 0 : const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
352 0 : const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
353 0 : const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
354 0 : const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
355 0 : const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
356 : // Combine
357 0 : res1 = _mm_packs_epi32(w0, w1);
358 0 : res7 = _mm_packs_epi32(w2, w3);
359 0 : res5 = _mm_packs_epi32(w4, w5);
360 0 : res3 = _mm_packs_epi32(w6, w7);
361 : }
362 : // Transpose the 8x8.
363 : {
364 : // 00 01 02 03 04 05 06 07
365 : // 10 11 12 13 14 15 16 17
366 : // 20 21 22 23 24 25 26 27
367 : // 30 31 32 33 34 35 36 37
368 : // 40 41 42 43 44 45 46 47
369 : // 50 51 52 53 54 55 56 57
370 : // 60 61 62 63 64 65 66 67
371 : // 70 71 72 73 74 75 76 77
372 0 : const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
373 0 : const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
374 0 : const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
375 0 : const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
376 0 : const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
377 0 : const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
378 0 : const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
379 0 : const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
380 : // 00 10 01 11 02 12 03 13
381 : // 20 30 21 31 22 32 23 33
382 : // 04 14 05 15 06 16 07 17
383 : // 24 34 25 35 26 36 27 37
384 : // 40 50 41 51 42 52 43 53
385 : // 60 70 61 71 62 72 63 73
386 : // 54 54 55 55 56 56 57 57
387 : // 64 74 65 75 66 76 67 77
388 0 : const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
389 0 : const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
390 0 : const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
391 0 : const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
392 0 : const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
393 0 : const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
394 0 : const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
395 0 : const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
396 : // 00 10 20 30 01 11 21 31
397 : // 40 50 60 70 41 51 61 71
398 : // 02 12 22 32 03 13 23 33
399 : // 42 52 62 72 43 53 63 73
400 : // 04 14 24 34 05 15 21 36
401 : // 44 54 64 74 45 55 61 76
402 : // 06 16 26 36 07 17 27 37
403 : // 46 56 66 76 47 57 67 77
404 0 : in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
405 0 : in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
406 0 : in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
407 0 : in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
408 0 : in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
409 0 : in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
410 0 : in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
411 0 : in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
412 : // 00 10 20 30 40 50 60 70
413 : // 01 11 21 31 41 51 61 71
414 : // 02 12 22 32 42 52 62 72
415 : // 03 13 23 33 43 53 63 73
416 : // 04 14 24 34 44 54 64 74
417 : // 05 15 25 35 45 55 65 75
418 : // 06 16 26 36 46 56 66 76
419 : // 07 17 27 37 47 57 67 77
420 : }
421 : }
422 : // Post-condition output and store it
423 : {
424 : // Post-condition (division by two)
425 : // division of two 16 bits signed numbers using shifts
426 : // n / 2 = (n - (n >> 15)) >> 1
427 0 : const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
428 0 : const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
429 0 : const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
430 0 : const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
431 0 : const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
432 0 : const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
433 0 : const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
434 0 : const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
435 0 : in0 = _mm_sub_epi16(in0, sign_in0);
436 0 : in1 = _mm_sub_epi16(in1, sign_in1);
437 0 : in2 = _mm_sub_epi16(in2, sign_in2);
438 0 : in3 = _mm_sub_epi16(in3, sign_in3);
439 0 : in4 = _mm_sub_epi16(in4, sign_in4);
440 0 : in5 = _mm_sub_epi16(in5, sign_in5);
441 0 : in6 = _mm_sub_epi16(in6, sign_in6);
442 0 : in7 = _mm_sub_epi16(in7, sign_in7);
443 0 : in0 = _mm_srai_epi16(in0, 1);
444 0 : in1 = _mm_srai_epi16(in1, 1);
445 0 : in2 = _mm_srai_epi16(in2, 1);
446 0 : in3 = _mm_srai_epi16(in3, 1);
447 0 : in4 = _mm_srai_epi16(in4, 1);
448 0 : in5 = _mm_srai_epi16(in5, 1);
449 0 : in6 = _mm_srai_epi16(in6, 1);
450 0 : in7 = _mm_srai_epi16(in7, 1);
451 : }
452 :
453 0 : iscan_ptr += n_coeffs;
454 0 : qcoeff_ptr += n_coeffs;
455 0 : dqcoeff_ptr += n_coeffs;
456 0 : n_coeffs = -n_coeffs;
457 0 : zero = _mm_setzero_si128();
458 :
459 0 : if (!skip_block) {
460 : __m128i eob;
461 : __m128i round, quant, dequant;
462 : {
463 : __m128i coeff0, coeff1;
464 :
465 : // Setup global values
466 : {
467 0 : round = _mm_load_si128((const __m128i *)round_ptr);
468 0 : quant = _mm_load_si128((const __m128i *)quant_ptr);
469 0 : dequant = _mm_load_si128((const __m128i *)dequant_ptr);
470 : }
471 :
472 : {
473 : __m128i coeff0_sign, coeff1_sign;
474 : __m128i qcoeff0, qcoeff1;
475 : __m128i qtmp0, qtmp1;
476 : // Do DC and first 15 AC
477 0 : coeff0 = *in[0];
478 0 : coeff1 = *in[1];
479 :
480 : // Poor man's sign extract
481 0 : coeff0_sign = _mm_srai_epi16(coeff0, 15);
482 0 : coeff1_sign = _mm_srai_epi16(coeff1, 15);
483 0 : qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
484 0 : qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
485 0 : qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
486 0 : qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
487 :
488 0 : qcoeff0 = _mm_adds_epi16(qcoeff0, round);
489 0 : round = _mm_unpackhi_epi64(round, round);
490 0 : qcoeff1 = _mm_adds_epi16(qcoeff1, round);
491 0 : qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
492 0 : quant = _mm_unpackhi_epi64(quant, quant);
493 0 : qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
494 :
495 : // Reinsert signs
496 0 : qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
497 0 : qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
498 0 : qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
499 0 : qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
500 :
501 0 : _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
502 0 : _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
503 :
504 0 : coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
505 0 : dequant = _mm_unpackhi_epi64(dequant, dequant);
506 0 : coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
507 :
508 0 : _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
509 0 : _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
510 : }
511 :
512 : {
513 : // Scan for eob
514 : __m128i zero_coeff0, zero_coeff1;
515 : __m128i nzero_coeff0, nzero_coeff1;
516 : __m128i iscan0, iscan1;
517 : __m128i eob1;
518 0 : zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
519 0 : zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
520 0 : nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
521 0 : nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
522 0 : iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
523 0 : iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
524 : // Add one to convert from indices to counts
525 0 : iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
526 0 : iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
527 0 : eob = _mm_and_si128(iscan0, nzero_coeff0);
528 0 : eob1 = _mm_and_si128(iscan1, nzero_coeff1);
529 0 : eob = _mm_max_epi16(eob, eob1);
530 : }
531 0 : n_coeffs += 8 * 2;
532 : }
533 :
534 : // AC only loop
535 0 : index = 2;
536 0 : while (n_coeffs < 0) {
537 : __m128i coeff0, coeff1;
538 : {
539 : __m128i coeff0_sign, coeff1_sign;
540 : __m128i qcoeff0, qcoeff1;
541 : __m128i qtmp0, qtmp1;
542 :
543 0 : assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
544 0 : coeff0 = *in[index];
545 0 : coeff1 = *in[index + 1];
546 :
547 : // Poor man's sign extract
548 0 : coeff0_sign = _mm_srai_epi16(coeff0, 15);
549 0 : coeff1_sign = _mm_srai_epi16(coeff1, 15);
550 0 : qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
551 0 : qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
552 0 : qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
553 0 : qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
554 :
555 0 : qcoeff0 = _mm_adds_epi16(qcoeff0, round);
556 0 : qcoeff1 = _mm_adds_epi16(qcoeff1, round);
557 0 : qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
558 0 : qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
559 :
560 : // Reinsert signs
561 0 : qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
562 0 : qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
563 0 : qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
564 0 : qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
565 :
566 0 : _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
567 0 : _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
568 :
569 0 : coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
570 0 : coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
571 :
572 0 : _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
573 0 : _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
574 : }
575 :
576 : {
577 : // Scan for eob
578 : __m128i zero_coeff0, zero_coeff1;
579 : __m128i nzero_coeff0, nzero_coeff1;
580 : __m128i iscan0, iscan1;
581 : __m128i eob0, eob1;
582 0 : zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
583 0 : zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
584 0 : nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
585 0 : nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
586 0 : iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
587 0 : iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
588 : // Add one to convert from indices to counts
589 0 : iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
590 0 : iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
591 0 : eob0 = _mm_and_si128(iscan0, nzero_coeff0);
592 0 : eob1 = _mm_and_si128(iscan1, nzero_coeff1);
593 0 : eob0 = _mm_max_epi16(eob0, eob1);
594 0 : eob = _mm_max_epi16(eob, eob0);
595 : }
596 0 : n_coeffs += 8 * 2;
597 0 : index += 2;
598 : }
599 :
600 : // Accumulate EOB
601 : {
602 : __m128i eob_shuffled;
603 0 : eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
604 0 : eob = _mm_max_epi16(eob, eob_shuffled);
605 0 : eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
606 0 : eob = _mm_max_epi16(eob, eob_shuffled);
607 0 : eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
608 0 : eob = _mm_max_epi16(eob, eob_shuffled);
609 0 : *eob_ptr = _mm_extract_epi16(eob, 1);
610 : }
611 : } else {
612 : do {
613 0 : _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
614 0 : _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
615 0 : _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
616 0 : _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
617 0 : n_coeffs += 8 * 2;
618 0 : } while (n_coeffs < 0);
619 0 : *eob_ptr = 0;
620 : }
621 0 : }
622 :
623 : // load 8x8 array
624 0 : static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in,
625 : int stride) {
626 0 : in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
627 0 : in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
628 0 : in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
629 0 : in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
630 0 : in[4] = _mm_load_si128((const __m128i *)(input + 4 * stride));
631 0 : in[5] = _mm_load_si128((const __m128i *)(input + 5 * stride));
632 0 : in[6] = _mm_load_si128((const __m128i *)(input + 6 * stride));
633 0 : in[7] = _mm_load_si128((const __m128i *)(input + 7 * stride));
634 :
635 0 : in[0] = _mm_slli_epi16(in[0], 2);
636 0 : in[1] = _mm_slli_epi16(in[1], 2);
637 0 : in[2] = _mm_slli_epi16(in[2], 2);
638 0 : in[3] = _mm_slli_epi16(in[3], 2);
639 0 : in[4] = _mm_slli_epi16(in[4], 2);
640 0 : in[5] = _mm_slli_epi16(in[5], 2);
641 0 : in[6] = _mm_slli_epi16(in[6], 2);
642 0 : in[7] = _mm_slli_epi16(in[7], 2);
643 0 : }
644 :
645 : // right shift and rounding
646 0 : static INLINE void right_shift_8x8(__m128i *res, const int bit) {
647 0 : __m128i sign0 = _mm_srai_epi16(res[0], 15);
648 0 : __m128i sign1 = _mm_srai_epi16(res[1], 15);
649 0 : __m128i sign2 = _mm_srai_epi16(res[2], 15);
650 0 : __m128i sign3 = _mm_srai_epi16(res[3], 15);
651 0 : __m128i sign4 = _mm_srai_epi16(res[4], 15);
652 0 : __m128i sign5 = _mm_srai_epi16(res[5], 15);
653 0 : __m128i sign6 = _mm_srai_epi16(res[6], 15);
654 0 : __m128i sign7 = _mm_srai_epi16(res[7], 15);
655 :
656 0 : if (bit == 2) {
657 0 : const __m128i const_rounding = _mm_set1_epi16(1);
658 0 : res[0] = _mm_add_epi16(res[0], const_rounding);
659 0 : res[1] = _mm_add_epi16(res[1], const_rounding);
660 0 : res[2] = _mm_add_epi16(res[2], const_rounding);
661 0 : res[3] = _mm_add_epi16(res[3], const_rounding);
662 0 : res[4] = _mm_add_epi16(res[4], const_rounding);
663 0 : res[5] = _mm_add_epi16(res[5], const_rounding);
664 0 : res[6] = _mm_add_epi16(res[6], const_rounding);
665 0 : res[7] = _mm_add_epi16(res[7], const_rounding);
666 : }
667 :
668 0 : res[0] = _mm_sub_epi16(res[0], sign0);
669 0 : res[1] = _mm_sub_epi16(res[1], sign1);
670 0 : res[2] = _mm_sub_epi16(res[2], sign2);
671 0 : res[3] = _mm_sub_epi16(res[3], sign3);
672 0 : res[4] = _mm_sub_epi16(res[4], sign4);
673 0 : res[5] = _mm_sub_epi16(res[5], sign5);
674 0 : res[6] = _mm_sub_epi16(res[6], sign6);
675 0 : res[7] = _mm_sub_epi16(res[7], sign7);
676 :
677 0 : if (bit == 1) {
678 0 : res[0] = _mm_srai_epi16(res[0], 1);
679 0 : res[1] = _mm_srai_epi16(res[1], 1);
680 0 : res[2] = _mm_srai_epi16(res[2], 1);
681 0 : res[3] = _mm_srai_epi16(res[3], 1);
682 0 : res[4] = _mm_srai_epi16(res[4], 1);
683 0 : res[5] = _mm_srai_epi16(res[5], 1);
684 0 : res[6] = _mm_srai_epi16(res[6], 1);
685 0 : res[7] = _mm_srai_epi16(res[7], 1);
686 : } else {
687 0 : res[0] = _mm_srai_epi16(res[0], 2);
688 0 : res[1] = _mm_srai_epi16(res[1], 2);
689 0 : res[2] = _mm_srai_epi16(res[2], 2);
690 0 : res[3] = _mm_srai_epi16(res[3], 2);
691 0 : res[4] = _mm_srai_epi16(res[4], 2);
692 0 : res[5] = _mm_srai_epi16(res[5], 2);
693 0 : res[6] = _mm_srai_epi16(res[6], 2);
694 0 : res[7] = _mm_srai_epi16(res[7], 2);
695 : }
696 0 : }
697 :
698 : // write 8x8 array
699 0 : static INLINE void write_buffer_8x8(tran_low_t *output, __m128i *res,
700 : int stride) {
701 0 : store_output(&res[0], (output + 0 * stride));
702 0 : store_output(&res[1], (output + 1 * stride));
703 0 : store_output(&res[2], (output + 2 * stride));
704 0 : store_output(&res[3], (output + 3 * stride));
705 0 : store_output(&res[4], (output + 4 * stride));
706 0 : store_output(&res[5], (output + 5 * stride));
707 0 : store_output(&res[6], (output + 6 * stride));
708 0 : store_output(&res[7], (output + 7 * stride));
709 0 : }
710 :
711 : // perform in-place transpose
712 0 : static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
713 0 : const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
714 0 : const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
715 0 : const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
716 0 : const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
717 0 : const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
718 0 : const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
719 0 : const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
720 0 : const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
721 : // 00 10 01 11 02 12 03 13
722 : // 20 30 21 31 22 32 23 33
723 : // 04 14 05 15 06 16 07 17
724 : // 24 34 25 35 26 36 27 37
725 : // 40 50 41 51 42 52 43 53
726 : // 60 70 61 71 62 72 63 73
727 : // 44 54 45 55 46 56 47 57
728 : // 64 74 65 75 66 76 67 77
729 0 : const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
730 0 : const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
731 0 : const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
732 0 : const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
733 0 : const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
734 0 : const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
735 0 : const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
736 0 : const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
737 : // 00 10 20 30 01 11 21 31
738 : // 40 50 60 70 41 51 61 71
739 : // 02 12 22 32 03 13 23 33
740 : // 42 52 62 72 43 53 63 73
741 : // 04 14 24 34 05 15 25 35
742 : // 44 54 64 74 45 55 65 75
743 : // 06 16 26 36 07 17 27 37
744 : // 46 56 66 76 47 57 67 77
745 0 : res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
746 0 : res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
747 0 : res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
748 0 : res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
749 0 : res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
750 0 : res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
751 0 : res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
752 0 : res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
753 : // 00 10 20 30 40 50 60 70
754 : // 01 11 21 31 41 51 61 71
755 : // 02 12 22 32 42 52 62 72
756 : // 03 13 23 33 43 53 63 73
757 : // 04 14 24 34 44 54 64 74
758 : // 05 15 25 35 45 55 65 75
759 : // 06 16 26 36 46 56 66 76
760 : // 07 17 27 37 47 57 67 77
761 0 : }
762 :
763 0 : static void fdct8_sse2(__m128i *in) {
764 : // constants
765 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
766 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
767 0 : const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
768 0 : const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
769 0 : const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
770 0 : const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
771 0 : const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
772 0 : const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
773 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
774 : __m128i u0, u1, u2, u3, u4, u5, u6, u7;
775 : __m128i v0, v1, v2, v3, v4, v5, v6, v7;
776 : __m128i s0, s1, s2, s3, s4, s5, s6, s7;
777 :
778 : // stage 1
779 0 : s0 = _mm_add_epi16(in[0], in[7]);
780 0 : s1 = _mm_add_epi16(in[1], in[6]);
781 0 : s2 = _mm_add_epi16(in[2], in[5]);
782 0 : s3 = _mm_add_epi16(in[3], in[4]);
783 0 : s4 = _mm_sub_epi16(in[3], in[4]);
784 0 : s5 = _mm_sub_epi16(in[2], in[5]);
785 0 : s6 = _mm_sub_epi16(in[1], in[6]);
786 0 : s7 = _mm_sub_epi16(in[0], in[7]);
787 :
788 0 : u0 = _mm_add_epi16(s0, s3);
789 0 : u1 = _mm_add_epi16(s1, s2);
790 0 : u2 = _mm_sub_epi16(s1, s2);
791 0 : u3 = _mm_sub_epi16(s0, s3);
792 : // interleave and perform butterfly multiplication/addition
793 0 : v0 = _mm_unpacklo_epi16(u0, u1);
794 0 : v1 = _mm_unpackhi_epi16(u0, u1);
795 0 : v2 = _mm_unpacklo_epi16(u2, u3);
796 0 : v3 = _mm_unpackhi_epi16(u2, u3);
797 :
798 0 : u0 = _mm_madd_epi16(v0, k__cospi_p16_p16);
799 0 : u1 = _mm_madd_epi16(v1, k__cospi_p16_p16);
800 0 : u2 = _mm_madd_epi16(v0, k__cospi_p16_m16);
801 0 : u3 = _mm_madd_epi16(v1, k__cospi_p16_m16);
802 0 : u4 = _mm_madd_epi16(v2, k__cospi_p24_p08);
803 0 : u5 = _mm_madd_epi16(v3, k__cospi_p24_p08);
804 0 : u6 = _mm_madd_epi16(v2, k__cospi_m08_p24);
805 0 : u7 = _mm_madd_epi16(v3, k__cospi_m08_p24);
806 :
807 : // shift and rounding
808 0 : v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
809 0 : v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
810 0 : v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
811 0 : v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
812 0 : v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
813 0 : v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
814 0 : v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
815 0 : v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
816 :
817 0 : u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
818 0 : u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
819 0 : u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
820 0 : u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
821 0 : u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
822 0 : u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
823 0 : u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
824 0 : u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
825 :
826 0 : in[0] = _mm_packs_epi32(u0, u1);
827 0 : in[2] = _mm_packs_epi32(u4, u5);
828 0 : in[4] = _mm_packs_epi32(u2, u3);
829 0 : in[6] = _mm_packs_epi32(u6, u7);
830 :
831 : // stage 2
832 : // interleave and perform butterfly multiplication/addition
833 0 : u0 = _mm_unpacklo_epi16(s6, s5);
834 0 : u1 = _mm_unpackhi_epi16(s6, s5);
835 0 : v0 = _mm_madd_epi16(u0, k__cospi_p16_m16);
836 0 : v1 = _mm_madd_epi16(u1, k__cospi_p16_m16);
837 0 : v2 = _mm_madd_epi16(u0, k__cospi_p16_p16);
838 0 : v3 = _mm_madd_epi16(u1, k__cospi_p16_p16);
839 :
840 : // shift and rounding
841 0 : u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
842 0 : u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
843 0 : u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
844 0 : u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
845 :
846 0 : v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
847 0 : v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
848 0 : v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
849 0 : v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
850 :
851 0 : u0 = _mm_packs_epi32(v0, v1);
852 0 : u1 = _mm_packs_epi32(v2, v3);
853 :
854 : // stage 3
855 0 : s0 = _mm_add_epi16(s4, u0);
856 0 : s1 = _mm_sub_epi16(s4, u0);
857 0 : s2 = _mm_sub_epi16(s7, u1);
858 0 : s3 = _mm_add_epi16(s7, u1);
859 :
860 : // stage 4
861 0 : u0 = _mm_unpacklo_epi16(s0, s3);
862 0 : u1 = _mm_unpackhi_epi16(s0, s3);
863 0 : u2 = _mm_unpacklo_epi16(s1, s2);
864 0 : u3 = _mm_unpackhi_epi16(s1, s2);
865 :
866 0 : v0 = _mm_madd_epi16(u0, k__cospi_p28_p04);
867 0 : v1 = _mm_madd_epi16(u1, k__cospi_p28_p04);
868 0 : v2 = _mm_madd_epi16(u2, k__cospi_p12_p20);
869 0 : v3 = _mm_madd_epi16(u3, k__cospi_p12_p20);
870 0 : v4 = _mm_madd_epi16(u2, k__cospi_m20_p12);
871 0 : v5 = _mm_madd_epi16(u3, k__cospi_m20_p12);
872 0 : v6 = _mm_madd_epi16(u0, k__cospi_m04_p28);
873 0 : v7 = _mm_madd_epi16(u1, k__cospi_m04_p28);
874 :
875 : // shift and rounding
876 0 : u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
877 0 : u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
878 0 : u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
879 0 : u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
880 0 : u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
881 0 : u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
882 0 : u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
883 0 : u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
884 :
885 0 : v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
886 0 : v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
887 0 : v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
888 0 : v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
889 0 : v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
890 0 : v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
891 0 : v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
892 0 : v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
893 :
894 0 : in[1] = _mm_packs_epi32(v0, v1);
895 0 : in[3] = _mm_packs_epi32(v4, v5);
896 0 : in[5] = _mm_packs_epi32(v2, v3);
897 0 : in[7] = _mm_packs_epi32(v6, v7);
898 :
899 : // transpose
900 0 : array_transpose_8x8(in, in);
901 0 : }
902 :
903 0 : static void fadst8_sse2(__m128i *in) {
904 : // Constants
905 0 : const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
906 0 : const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
907 0 : const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
908 0 : const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
909 0 : const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
910 0 : const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
911 0 : const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
912 0 : const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
913 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
914 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
915 0 : const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
916 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
917 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
918 0 : const __m128i k__const_0 = _mm_set1_epi16(0);
919 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
920 :
921 : __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
922 : __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
923 : __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
924 : __m128i s0, s1, s2, s3, s4, s5, s6, s7;
925 : __m128i in0, in1, in2, in3, in4, in5, in6, in7;
926 :
927 : // properly aligned for butterfly input
928 0 : in0 = in[7];
929 0 : in1 = in[0];
930 0 : in2 = in[5];
931 0 : in3 = in[2];
932 0 : in4 = in[3];
933 0 : in5 = in[4];
934 0 : in6 = in[1];
935 0 : in7 = in[6];
936 :
937 : // column transformation
938 : // stage 1
939 : // interleave and multiply/add into 32-bit integer
940 0 : s0 = _mm_unpacklo_epi16(in0, in1);
941 0 : s1 = _mm_unpackhi_epi16(in0, in1);
942 0 : s2 = _mm_unpacklo_epi16(in2, in3);
943 0 : s3 = _mm_unpackhi_epi16(in2, in3);
944 0 : s4 = _mm_unpacklo_epi16(in4, in5);
945 0 : s5 = _mm_unpackhi_epi16(in4, in5);
946 0 : s6 = _mm_unpacklo_epi16(in6, in7);
947 0 : s7 = _mm_unpackhi_epi16(in6, in7);
948 :
949 0 : u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
950 0 : u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
951 0 : u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
952 0 : u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
953 0 : u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
954 0 : u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
955 0 : u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
956 0 : u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
957 0 : u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
958 0 : u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
959 0 : u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
960 0 : u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
961 0 : u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
962 0 : u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
963 0 : u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
964 0 : u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
965 :
966 : // addition
967 0 : w0 = _mm_add_epi32(u0, u8);
968 0 : w1 = _mm_add_epi32(u1, u9);
969 0 : w2 = _mm_add_epi32(u2, u10);
970 0 : w3 = _mm_add_epi32(u3, u11);
971 0 : w4 = _mm_add_epi32(u4, u12);
972 0 : w5 = _mm_add_epi32(u5, u13);
973 0 : w6 = _mm_add_epi32(u6, u14);
974 0 : w7 = _mm_add_epi32(u7, u15);
975 0 : w8 = _mm_sub_epi32(u0, u8);
976 0 : w9 = _mm_sub_epi32(u1, u9);
977 0 : w10 = _mm_sub_epi32(u2, u10);
978 0 : w11 = _mm_sub_epi32(u3, u11);
979 0 : w12 = _mm_sub_epi32(u4, u12);
980 0 : w13 = _mm_sub_epi32(u5, u13);
981 0 : w14 = _mm_sub_epi32(u6, u14);
982 0 : w15 = _mm_sub_epi32(u7, u15);
983 :
984 : // shift and rounding
985 0 : v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
986 0 : v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
987 0 : v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
988 0 : v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
989 0 : v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
990 0 : v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
991 0 : v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
992 0 : v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
993 0 : v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
994 0 : v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
995 0 : v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
996 0 : v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
997 0 : v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
998 0 : v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
999 0 : v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
1000 0 : v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
1001 :
1002 0 : u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
1003 0 : u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
1004 0 : u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
1005 0 : u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
1006 0 : u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
1007 0 : u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
1008 0 : u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
1009 0 : u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
1010 0 : u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
1011 0 : u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
1012 0 : u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
1013 0 : u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
1014 0 : u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
1015 0 : u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
1016 0 : u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
1017 0 : u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
1018 :
1019 : // back to 16-bit and pack 8 integers into __m128i
1020 0 : in[0] = _mm_packs_epi32(u0, u1);
1021 0 : in[1] = _mm_packs_epi32(u2, u3);
1022 0 : in[2] = _mm_packs_epi32(u4, u5);
1023 0 : in[3] = _mm_packs_epi32(u6, u7);
1024 0 : in[4] = _mm_packs_epi32(u8, u9);
1025 0 : in[5] = _mm_packs_epi32(u10, u11);
1026 0 : in[6] = _mm_packs_epi32(u12, u13);
1027 0 : in[7] = _mm_packs_epi32(u14, u15);
1028 :
1029 : // stage 2
1030 0 : s0 = _mm_add_epi16(in[0], in[2]);
1031 0 : s1 = _mm_add_epi16(in[1], in[3]);
1032 0 : s2 = _mm_sub_epi16(in[0], in[2]);
1033 0 : s3 = _mm_sub_epi16(in[1], in[3]);
1034 0 : u0 = _mm_unpacklo_epi16(in[4], in[5]);
1035 0 : u1 = _mm_unpackhi_epi16(in[4], in[5]);
1036 0 : u2 = _mm_unpacklo_epi16(in[6], in[7]);
1037 0 : u3 = _mm_unpackhi_epi16(in[6], in[7]);
1038 :
1039 0 : v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
1040 0 : v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
1041 0 : v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
1042 0 : v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
1043 0 : v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
1044 0 : v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
1045 0 : v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
1046 0 : v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
1047 :
1048 0 : w0 = _mm_add_epi32(v0, v4);
1049 0 : w1 = _mm_add_epi32(v1, v5);
1050 0 : w2 = _mm_add_epi32(v2, v6);
1051 0 : w3 = _mm_add_epi32(v3, v7);
1052 0 : w4 = _mm_sub_epi32(v0, v4);
1053 0 : w5 = _mm_sub_epi32(v1, v5);
1054 0 : w6 = _mm_sub_epi32(v2, v6);
1055 0 : w7 = _mm_sub_epi32(v3, v7);
1056 :
1057 0 : v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
1058 0 : v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
1059 0 : v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
1060 0 : v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
1061 0 : v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
1062 0 : v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
1063 0 : v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
1064 0 : v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
1065 :
1066 0 : u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
1067 0 : u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
1068 0 : u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
1069 0 : u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
1070 0 : u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
1071 0 : u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
1072 0 : u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
1073 0 : u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
1074 :
1075 : // back to 16-bit intergers
1076 0 : s4 = _mm_packs_epi32(u0, u1);
1077 0 : s5 = _mm_packs_epi32(u2, u3);
1078 0 : s6 = _mm_packs_epi32(u4, u5);
1079 0 : s7 = _mm_packs_epi32(u6, u7);
1080 :
1081 : // stage 3
1082 0 : u0 = _mm_unpacklo_epi16(s2, s3);
1083 0 : u1 = _mm_unpackhi_epi16(s2, s3);
1084 0 : u2 = _mm_unpacklo_epi16(s6, s7);
1085 0 : u3 = _mm_unpackhi_epi16(s6, s7);
1086 :
1087 0 : v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
1088 0 : v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
1089 0 : v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
1090 0 : v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
1091 0 : v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
1092 0 : v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
1093 0 : v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
1094 0 : v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
1095 :
1096 0 : u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
1097 0 : u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
1098 0 : u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
1099 0 : u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
1100 0 : u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
1101 0 : u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
1102 0 : u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
1103 0 : u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
1104 :
1105 0 : v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
1106 0 : v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
1107 0 : v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
1108 0 : v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
1109 0 : v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
1110 0 : v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
1111 0 : v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
1112 0 : v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
1113 :
1114 0 : s2 = _mm_packs_epi32(v0, v1);
1115 0 : s3 = _mm_packs_epi32(v2, v3);
1116 0 : s6 = _mm_packs_epi32(v4, v5);
1117 0 : s7 = _mm_packs_epi32(v6, v7);
1118 :
1119 : // FIXME(jingning): do subtract using bit inversion?
1120 0 : in[0] = s0;
1121 0 : in[1] = _mm_sub_epi16(k__const_0, s4);
1122 0 : in[2] = s6;
1123 0 : in[3] = _mm_sub_epi16(k__const_0, s2);
1124 0 : in[4] = s3;
1125 0 : in[5] = _mm_sub_epi16(k__const_0, s7);
1126 0 : in[6] = s5;
1127 0 : in[7] = _mm_sub_epi16(k__const_0, s1);
1128 :
1129 : // transpose
1130 0 : array_transpose_8x8(in, in);
1131 0 : }
1132 :
1133 0 : void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride,
1134 : int tx_type) {
1135 : __m128i in[8];
1136 :
1137 0 : switch (tx_type) {
1138 0 : case DCT_DCT: vpx_fdct8x8_sse2(input, output, stride); break;
1139 : case ADST_DCT:
1140 0 : load_buffer_8x8(input, in, stride);
1141 0 : fadst8_sse2(in);
1142 0 : fdct8_sse2(in);
1143 0 : right_shift_8x8(in, 1);
1144 0 : write_buffer_8x8(output, in, 8);
1145 0 : break;
1146 : case DCT_ADST:
1147 0 : load_buffer_8x8(input, in, stride);
1148 0 : fdct8_sse2(in);
1149 0 : fadst8_sse2(in);
1150 0 : right_shift_8x8(in, 1);
1151 0 : write_buffer_8x8(output, in, 8);
1152 0 : break;
1153 : case ADST_ADST:
1154 0 : load_buffer_8x8(input, in, stride);
1155 0 : fadst8_sse2(in);
1156 0 : fadst8_sse2(in);
1157 0 : right_shift_8x8(in, 1);
1158 0 : write_buffer_8x8(output, in, 8);
1159 0 : break;
1160 0 : default: assert(0); break;
1161 : }
1162 0 : }
1163 :
1164 0 : static INLINE void load_buffer_16x16(const int16_t *input, __m128i *in0,
1165 : __m128i *in1, int stride) {
1166 : // load first 8 columns
1167 0 : load_buffer_8x8(input, in0, stride);
1168 0 : load_buffer_8x8(input + 8 * stride, in0 + 8, stride);
1169 :
1170 0 : input += 8;
1171 : // load second 8 columns
1172 0 : load_buffer_8x8(input, in1, stride);
1173 0 : load_buffer_8x8(input + 8 * stride, in1 + 8, stride);
1174 0 : }
1175 :
1176 0 : static INLINE void write_buffer_16x16(tran_low_t *output, __m128i *in0,
1177 : __m128i *in1, int stride) {
1178 : // write first 8 columns
1179 0 : write_buffer_8x8(output, in0, stride);
1180 0 : write_buffer_8x8(output + 8 * stride, in0 + 8, stride);
1181 : // write second 8 columns
1182 0 : output += 8;
1183 0 : write_buffer_8x8(output, in1, stride);
1184 0 : write_buffer_8x8(output + 8 * stride, in1 + 8, stride);
1185 0 : }
1186 :
1187 0 : static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
1188 : __m128i tbuf[8];
1189 0 : array_transpose_8x8(res0, res0);
1190 0 : array_transpose_8x8(res1, tbuf);
1191 0 : array_transpose_8x8(res0 + 8, res1);
1192 0 : array_transpose_8x8(res1 + 8, res1 + 8);
1193 :
1194 0 : res0[8] = tbuf[0];
1195 0 : res0[9] = tbuf[1];
1196 0 : res0[10] = tbuf[2];
1197 0 : res0[11] = tbuf[3];
1198 0 : res0[12] = tbuf[4];
1199 0 : res0[13] = tbuf[5];
1200 0 : res0[14] = tbuf[6];
1201 0 : res0[15] = tbuf[7];
1202 0 : }
1203 :
1204 0 : static INLINE void right_shift_16x16(__m128i *res0, __m128i *res1) {
1205 : // perform rounding operations
1206 0 : right_shift_8x8(res0, 2);
1207 0 : right_shift_8x8(res0 + 8, 2);
1208 0 : right_shift_8x8(res1, 2);
1209 0 : right_shift_8x8(res1 + 8, 2);
1210 0 : }
1211 :
1212 0 : static void fdct16_8col(__m128i *in) {
1213 : // perform 16x16 1-D DCT for 8 columns
1214 : __m128i i[8], s[8], p[8], t[8], u[16], v[16];
1215 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
1216 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1217 0 : const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1218 0 : const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
1219 0 : const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
1220 0 : const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
1221 0 : const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
1222 0 : const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
1223 0 : const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
1224 0 : const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
1225 0 : const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
1226 0 : const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
1227 0 : const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
1228 0 : const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
1229 0 : const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
1230 0 : const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
1231 0 : const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
1232 0 : const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
1233 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
1234 :
1235 : // stage 1
1236 0 : i[0] = _mm_add_epi16(in[0], in[15]);
1237 0 : i[1] = _mm_add_epi16(in[1], in[14]);
1238 0 : i[2] = _mm_add_epi16(in[2], in[13]);
1239 0 : i[3] = _mm_add_epi16(in[3], in[12]);
1240 0 : i[4] = _mm_add_epi16(in[4], in[11]);
1241 0 : i[5] = _mm_add_epi16(in[5], in[10]);
1242 0 : i[6] = _mm_add_epi16(in[6], in[9]);
1243 0 : i[7] = _mm_add_epi16(in[7], in[8]);
1244 :
1245 0 : s[0] = _mm_sub_epi16(in[7], in[8]);
1246 0 : s[1] = _mm_sub_epi16(in[6], in[9]);
1247 0 : s[2] = _mm_sub_epi16(in[5], in[10]);
1248 0 : s[3] = _mm_sub_epi16(in[4], in[11]);
1249 0 : s[4] = _mm_sub_epi16(in[3], in[12]);
1250 0 : s[5] = _mm_sub_epi16(in[2], in[13]);
1251 0 : s[6] = _mm_sub_epi16(in[1], in[14]);
1252 0 : s[7] = _mm_sub_epi16(in[0], in[15]);
1253 :
1254 0 : p[0] = _mm_add_epi16(i[0], i[7]);
1255 0 : p[1] = _mm_add_epi16(i[1], i[6]);
1256 0 : p[2] = _mm_add_epi16(i[2], i[5]);
1257 0 : p[3] = _mm_add_epi16(i[3], i[4]);
1258 0 : p[4] = _mm_sub_epi16(i[3], i[4]);
1259 0 : p[5] = _mm_sub_epi16(i[2], i[5]);
1260 0 : p[6] = _mm_sub_epi16(i[1], i[6]);
1261 0 : p[7] = _mm_sub_epi16(i[0], i[7]);
1262 :
1263 0 : u[0] = _mm_add_epi16(p[0], p[3]);
1264 0 : u[1] = _mm_add_epi16(p[1], p[2]);
1265 0 : u[2] = _mm_sub_epi16(p[1], p[2]);
1266 0 : u[3] = _mm_sub_epi16(p[0], p[3]);
1267 :
1268 0 : v[0] = _mm_unpacklo_epi16(u[0], u[1]);
1269 0 : v[1] = _mm_unpackhi_epi16(u[0], u[1]);
1270 0 : v[2] = _mm_unpacklo_epi16(u[2], u[3]);
1271 0 : v[3] = _mm_unpackhi_epi16(u[2], u[3]);
1272 :
1273 0 : u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);
1274 0 : u[1] = _mm_madd_epi16(v[1], k__cospi_p16_p16);
1275 0 : u[2] = _mm_madd_epi16(v[0], k__cospi_p16_m16);
1276 0 : u[3] = _mm_madd_epi16(v[1], k__cospi_p16_m16);
1277 0 : u[4] = _mm_madd_epi16(v[2], k__cospi_p24_p08);
1278 0 : u[5] = _mm_madd_epi16(v[3], k__cospi_p24_p08);
1279 0 : u[6] = _mm_madd_epi16(v[2], k__cospi_m08_p24);
1280 0 : u[7] = _mm_madd_epi16(v[3], k__cospi_m08_p24);
1281 :
1282 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1283 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1284 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1285 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1286 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1287 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1288 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1289 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1290 :
1291 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1292 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1293 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1294 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1295 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1296 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1297 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1298 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1299 :
1300 0 : in[0] = _mm_packs_epi32(u[0], u[1]);
1301 0 : in[4] = _mm_packs_epi32(u[4], u[5]);
1302 0 : in[8] = _mm_packs_epi32(u[2], u[3]);
1303 0 : in[12] = _mm_packs_epi32(u[6], u[7]);
1304 :
1305 0 : u[0] = _mm_unpacklo_epi16(p[5], p[6]);
1306 0 : u[1] = _mm_unpackhi_epi16(p[5], p[6]);
1307 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
1308 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
1309 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
1310 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
1311 :
1312 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1313 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1314 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1315 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1316 :
1317 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1318 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1319 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1320 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1321 :
1322 0 : u[0] = _mm_packs_epi32(v[0], v[1]);
1323 0 : u[1] = _mm_packs_epi32(v[2], v[3]);
1324 :
1325 0 : t[0] = _mm_add_epi16(p[4], u[0]);
1326 0 : t[1] = _mm_sub_epi16(p[4], u[0]);
1327 0 : t[2] = _mm_sub_epi16(p[7], u[1]);
1328 0 : t[3] = _mm_add_epi16(p[7], u[1]);
1329 :
1330 0 : u[0] = _mm_unpacklo_epi16(t[0], t[3]);
1331 0 : u[1] = _mm_unpackhi_epi16(t[0], t[3]);
1332 0 : u[2] = _mm_unpacklo_epi16(t[1], t[2]);
1333 0 : u[3] = _mm_unpackhi_epi16(t[1], t[2]);
1334 :
1335 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p28_p04);
1336 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p28_p04);
1337 0 : v[2] = _mm_madd_epi16(u[2], k__cospi_p12_p20);
1338 0 : v[3] = _mm_madd_epi16(u[3], k__cospi_p12_p20);
1339 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_m20_p12);
1340 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_m20_p12);
1341 0 : v[6] = _mm_madd_epi16(u[0], k__cospi_m04_p28);
1342 0 : v[7] = _mm_madd_epi16(u[1], k__cospi_m04_p28);
1343 :
1344 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1345 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1346 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1347 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1348 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1349 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1350 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1351 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1352 :
1353 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1354 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1355 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1356 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1357 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1358 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1359 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1360 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1361 :
1362 0 : in[2] = _mm_packs_epi32(v[0], v[1]);
1363 0 : in[6] = _mm_packs_epi32(v[4], v[5]);
1364 0 : in[10] = _mm_packs_epi32(v[2], v[3]);
1365 0 : in[14] = _mm_packs_epi32(v[6], v[7]);
1366 :
1367 : // stage 2
1368 0 : u[0] = _mm_unpacklo_epi16(s[2], s[5]);
1369 0 : u[1] = _mm_unpackhi_epi16(s[2], s[5]);
1370 0 : u[2] = _mm_unpacklo_epi16(s[3], s[4]);
1371 0 : u[3] = _mm_unpackhi_epi16(s[3], s[4]);
1372 :
1373 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
1374 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
1375 0 : v[2] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
1376 0 : v[3] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
1377 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
1378 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
1379 0 : v[6] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
1380 0 : v[7] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
1381 :
1382 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1383 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1384 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1385 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1386 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1387 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1388 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1389 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1390 :
1391 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1392 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1393 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1394 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1395 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1396 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1397 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1398 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1399 :
1400 0 : t[2] = _mm_packs_epi32(v[0], v[1]);
1401 0 : t[3] = _mm_packs_epi32(v[2], v[3]);
1402 0 : t[4] = _mm_packs_epi32(v[4], v[5]);
1403 0 : t[5] = _mm_packs_epi32(v[6], v[7]);
1404 :
1405 : // stage 3
1406 0 : p[0] = _mm_add_epi16(s[0], t[3]);
1407 0 : p[1] = _mm_add_epi16(s[1], t[2]);
1408 0 : p[2] = _mm_sub_epi16(s[1], t[2]);
1409 0 : p[3] = _mm_sub_epi16(s[0], t[3]);
1410 0 : p[4] = _mm_sub_epi16(s[7], t[4]);
1411 0 : p[5] = _mm_sub_epi16(s[6], t[5]);
1412 0 : p[6] = _mm_add_epi16(s[6], t[5]);
1413 0 : p[7] = _mm_add_epi16(s[7], t[4]);
1414 :
1415 : // stage 4
1416 0 : u[0] = _mm_unpacklo_epi16(p[1], p[6]);
1417 0 : u[1] = _mm_unpackhi_epi16(p[1], p[6]);
1418 0 : u[2] = _mm_unpacklo_epi16(p[2], p[5]);
1419 0 : u[3] = _mm_unpackhi_epi16(p[2], p[5]);
1420 :
1421 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
1422 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
1423 0 : v[2] = _mm_madd_epi16(u[2], k__cospi_p24_p08);
1424 0 : v[3] = _mm_madd_epi16(u[3], k__cospi_p24_p08);
1425 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p08_m24);
1426 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p08_m24);
1427 0 : v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08);
1428 0 : v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08);
1429 :
1430 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1431 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1432 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1433 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1434 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1435 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1436 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1437 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1438 :
1439 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1440 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1441 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1442 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1443 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1444 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1445 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1446 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1447 :
1448 0 : t[1] = _mm_packs_epi32(v[0], v[1]);
1449 0 : t[2] = _mm_packs_epi32(v[2], v[3]);
1450 0 : t[5] = _mm_packs_epi32(v[4], v[5]);
1451 0 : t[6] = _mm_packs_epi32(v[6], v[7]);
1452 :
1453 : // stage 5
1454 0 : s[0] = _mm_add_epi16(p[0], t[1]);
1455 0 : s[1] = _mm_sub_epi16(p[0], t[1]);
1456 0 : s[2] = _mm_add_epi16(p[3], t[2]);
1457 0 : s[3] = _mm_sub_epi16(p[3], t[2]);
1458 0 : s[4] = _mm_sub_epi16(p[4], t[5]);
1459 0 : s[5] = _mm_add_epi16(p[4], t[5]);
1460 0 : s[6] = _mm_sub_epi16(p[7], t[6]);
1461 0 : s[7] = _mm_add_epi16(p[7], t[6]);
1462 :
1463 : // stage 6
1464 0 : u[0] = _mm_unpacklo_epi16(s[0], s[7]);
1465 0 : u[1] = _mm_unpackhi_epi16(s[0], s[7]);
1466 0 : u[2] = _mm_unpacklo_epi16(s[1], s[6]);
1467 0 : u[3] = _mm_unpackhi_epi16(s[1], s[6]);
1468 0 : u[4] = _mm_unpacklo_epi16(s[2], s[5]);
1469 0 : u[5] = _mm_unpackhi_epi16(s[2], s[5]);
1470 0 : u[6] = _mm_unpacklo_epi16(s[3], s[4]);
1471 0 : u[7] = _mm_unpackhi_epi16(s[3], s[4]);
1472 :
1473 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p30_p02);
1474 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p30_p02);
1475 0 : v[2] = _mm_madd_epi16(u[2], k__cospi_p14_p18);
1476 0 : v[3] = _mm_madd_epi16(u[3], k__cospi_p14_p18);
1477 0 : v[4] = _mm_madd_epi16(u[4], k__cospi_p22_p10);
1478 0 : v[5] = _mm_madd_epi16(u[5], k__cospi_p22_p10);
1479 0 : v[6] = _mm_madd_epi16(u[6], k__cospi_p06_p26);
1480 0 : v[7] = _mm_madd_epi16(u[7], k__cospi_p06_p26);
1481 0 : v[8] = _mm_madd_epi16(u[6], k__cospi_m26_p06);
1482 0 : v[9] = _mm_madd_epi16(u[7], k__cospi_m26_p06);
1483 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_m10_p22);
1484 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_m10_p22);
1485 0 : v[12] = _mm_madd_epi16(u[2], k__cospi_m18_p14);
1486 0 : v[13] = _mm_madd_epi16(u[3], k__cospi_m18_p14);
1487 0 : v[14] = _mm_madd_epi16(u[0], k__cospi_m02_p30);
1488 0 : v[15] = _mm_madd_epi16(u[1], k__cospi_m02_p30);
1489 :
1490 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1491 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1492 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1493 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1494 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1495 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1496 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1497 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1498 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1499 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1500 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1501 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1502 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1503 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1504 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1505 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1506 :
1507 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1508 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1509 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1510 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1511 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1512 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1513 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1514 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1515 0 : v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1516 0 : v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1517 0 : v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1518 0 : v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1519 0 : v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1520 0 : v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1521 0 : v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1522 0 : v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1523 :
1524 0 : in[1] = _mm_packs_epi32(v[0], v[1]);
1525 0 : in[9] = _mm_packs_epi32(v[2], v[3]);
1526 0 : in[5] = _mm_packs_epi32(v[4], v[5]);
1527 0 : in[13] = _mm_packs_epi32(v[6], v[7]);
1528 0 : in[3] = _mm_packs_epi32(v[8], v[9]);
1529 0 : in[11] = _mm_packs_epi32(v[10], v[11]);
1530 0 : in[7] = _mm_packs_epi32(v[12], v[13]);
1531 0 : in[15] = _mm_packs_epi32(v[14], v[15]);
1532 0 : }
1533 :
1534 0 : static void fadst16_8col(__m128i *in) {
1535 : // perform 16x16 1-D ADST for 8 columns
1536 : __m128i s[16], x[16], u[32], v[32];
1537 0 : const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
1538 0 : const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
1539 0 : const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
1540 0 : const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
1541 0 : const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
1542 0 : const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
1543 0 : const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
1544 0 : const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
1545 0 : const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
1546 0 : const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
1547 0 : const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
1548 0 : const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
1549 0 : const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
1550 0 : const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
1551 0 : const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
1552 0 : const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
1553 0 : const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
1554 0 : const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
1555 0 : const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
1556 0 : const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
1557 0 : const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
1558 0 : const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
1559 0 : const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
1560 0 : const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
1561 0 : const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
1562 0 : const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
1563 0 : const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
1564 0 : const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
1565 0 : const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
1566 0 : const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
1567 0 : const __m128i kZero = _mm_set1_epi16(0);
1568 :
1569 0 : u[0] = _mm_unpacklo_epi16(in[15], in[0]);
1570 0 : u[1] = _mm_unpackhi_epi16(in[15], in[0]);
1571 0 : u[2] = _mm_unpacklo_epi16(in[13], in[2]);
1572 0 : u[3] = _mm_unpackhi_epi16(in[13], in[2]);
1573 0 : u[4] = _mm_unpacklo_epi16(in[11], in[4]);
1574 0 : u[5] = _mm_unpackhi_epi16(in[11], in[4]);
1575 0 : u[6] = _mm_unpacklo_epi16(in[9], in[6]);
1576 0 : u[7] = _mm_unpackhi_epi16(in[9], in[6]);
1577 0 : u[8] = _mm_unpacklo_epi16(in[7], in[8]);
1578 0 : u[9] = _mm_unpackhi_epi16(in[7], in[8]);
1579 0 : u[10] = _mm_unpacklo_epi16(in[5], in[10]);
1580 0 : u[11] = _mm_unpackhi_epi16(in[5], in[10]);
1581 0 : u[12] = _mm_unpacklo_epi16(in[3], in[12]);
1582 0 : u[13] = _mm_unpackhi_epi16(in[3], in[12]);
1583 0 : u[14] = _mm_unpacklo_epi16(in[1], in[14]);
1584 0 : u[15] = _mm_unpackhi_epi16(in[1], in[14]);
1585 :
1586 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
1587 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
1588 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
1589 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
1590 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
1591 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
1592 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
1593 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
1594 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
1595 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
1596 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
1597 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
1598 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
1599 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
1600 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
1601 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
1602 0 : v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
1603 0 : v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
1604 0 : v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
1605 0 : v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
1606 0 : v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
1607 0 : v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
1608 0 : v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
1609 0 : v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
1610 0 : v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
1611 0 : v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
1612 0 : v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
1613 0 : v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
1614 0 : v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
1615 0 : v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
1616 0 : v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
1617 0 : v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
1618 :
1619 0 : u[0] = _mm_add_epi32(v[0], v[16]);
1620 0 : u[1] = _mm_add_epi32(v[1], v[17]);
1621 0 : u[2] = _mm_add_epi32(v[2], v[18]);
1622 0 : u[3] = _mm_add_epi32(v[3], v[19]);
1623 0 : u[4] = _mm_add_epi32(v[4], v[20]);
1624 0 : u[5] = _mm_add_epi32(v[5], v[21]);
1625 0 : u[6] = _mm_add_epi32(v[6], v[22]);
1626 0 : u[7] = _mm_add_epi32(v[7], v[23]);
1627 0 : u[8] = _mm_add_epi32(v[8], v[24]);
1628 0 : u[9] = _mm_add_epi32(v[9], v[25]);
1629 0 : u[10] = _mm_add_epi32(v[10], v[26]);
1630 0 : u[11] = _mm_add_epi32(v[11], v[27]);
1631 0 : u[12] = _mm_add_epi32(v[12], v[28]);
1632 0 : u[13] = _mm_add_epi32(v[13], v[29]);
1633 0 : u[14] = _mm_add_epi32(v[14], v[30]);
1634 0 : u[15] = _mm_add_epi32(v[15], v[31]);
1635 0 : u[16] = _mm_sub_epi32(v[0], v[16]);
1636 0 : u[17] = _mm_sub_epi32(v[1], v[17]);
1637 0 : u[18] = _mm_sub_epi32(v[2], v[18]);
1638 0 : u[19] = _mm_sub_epi32(v[3], v[19]);
1639 0 : u[20] = _mm_sub_epi32(v[4], v[20]);
1640 0 : u[21] = _mm_sub_epi32(v[5], v[21]);
1641 0 : u[22] = _mm_sub_epi32(v[6], v[22]);
1642 0 : u[23] = _mm_sub_epi32(v[7], v[23]);
1643 0 : u[24] = _mm_sub_epi32(v[8], v[24]);
1644 0 : u[25] = _mm_sub_epi32(v[9], v[25]);
1645 0 : u[26] = _mm_sub_epi32(v[10], v[26]);
1646 0 : u[27] = _mm_sub_epi32(v[11], v[27]);
1647 0 : u[28] = _mm_sub_epi32(v[12], v[28]);
1648 0 : u[29] = _mm_sub_epi32(v[13], v[29]);
1649 0 : u[30] = _mm_sub_epi32(v[14], v[30]);
1650 0 : u[31] = _mm_sub_epi32(v[15], v[31]);
1651 :
1652 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1653 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1654 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1655 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1656 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1657 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1658 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1659 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1660 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1661 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1662 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1663 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1664 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1665 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1666 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1667 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1668 0 : v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
1669 0 : v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
1670 0 : v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
1671 0 : v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
1672 0 : v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
1673 0 : v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
1674 0 : v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
1675 0 : v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
1676 0 : v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
1677 0 : v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
1678 0 : v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
1679 0 : v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
1680 0 : v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
1681 0 : v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
1682 0 : v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
1683 0 : v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
1684 :
1685 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1686 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1687 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1688 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1689 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1690 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1691 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1692 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1693 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
1694 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
1695 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1696 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1697 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1698 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1699 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1700 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1701 0 : u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
1702 0 : u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
1703 0 : u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
1704 0 : u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
1705 0 : u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
1706 0 : u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
1707 0 : u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
1708 0 : u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
1709 0 : u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
1710 0 : u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
1711 0 : u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
1712 0 : u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
1713 0 : u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
1714 0 : u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
1715 0 : u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
1716 0 : u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
1717 :
1718 0 : s[0] = _mm_packs_epi32(u[0], u[1]);
1719 0 : s[1] = _mm_packs_epi32(u[2], u[3]);
1720 0 : s[2] = _mm_packs_epi32(u[4], u[5]);
1721 0 : s[3] = _mm_packs_epi32(u[6], u[7]);
1722 0 : s[4] = _mm_packs_epi32(u[8], u[9]);
1723 0 : s[5] = _mm_packs_epi32(u[10], u[11]);
1724 0 : s[6] = _mm_packs_epi32(u[12], u[13]);
1725 0 : s[7] = _mm_packs_epi32(u[14], u[15]);
1726 0 : s[8] = _mm_packs_epi32(u[16], u[17]);
1727 0 : s[9] = _mm_packs_epi32(u[18], u[19]);
1728 0 : s[10] = _mm_packs_epi32(u[20], u[21]);
1729 0 : s[11] = _mm_packs_epi32(u[22], u[23]);
1730 0 : s[12] = _mm_packs_epi32(u[24], u[25]);
1731 0 : s[13] = _mm_packs_epi32(u[26], u[27]);
1732 0 : s[14] = _mm_packs_epi32(u[28], u[29]);
1733 0 : s[15] = _mm_packs_epi32(u[30], u[31]);
1734 :
1735 : // stage 2
1736 0 : u[0] = _mm_unpacklo_epi16(s[8], s[9]);
1737 0 : u[1] = _mm_unpackhi_epi16(s[8], s[9]);
1738 0 : u[2] = _mm_unpacklo_epi16(s[10], s[11]);
1739 0 : u[3] = _mm_unpackhi_epi16(s[10], s[11]);
1740 0 : u[4] = _mm_unpacklo_epi16(s[12], s[13]);
1741 0 : u[5] = _mm_unpackhi_epi16(s[12], s[13]);
1742 0 : u[6] = _mm_unpacklo_epi16(s[14], s[15]);
1743 0 : u[7] = _mm_unpackhi_epi16(s[14], s[15]);
1744 :
1745 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
1746 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
1747 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
1748 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
1749 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
1750 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
1751 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
1752 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
1753 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
1754 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
1755 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
1756 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
1757 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
1758 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
1759 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
1760 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
1761 :
1762 0 : u[0] = _mm_add_epi32(v[0], v[8]);
1763 0 : u[1] = _mm_add_epi32(v[1], v[9]);
1764 0 : u[2] = _mm_add_epi32(v[2], v[10]);
1765 0 : u[3] = _mm_add_epi32(v[3], v[11]);
1766 0 : u[4] = _mm_add_epi32(v[4], v[12]);
1767 0 : u[5] = _mm_add_epi32(v[5], v[13]);
1768 0 : u[6] = _mm_add_epi32(v[6], v[14]);
1769 0 : u[7] = _mm_add_epi32(v[7], v[15]);
1770 0 : u[8] = _mm_sub_epi32(v[0], v[8]);
1771 0 : u[9] = _mm_sub_epi32(v[1], v[9]);
1772 0 : u[10] = _mm_sub_epi32(v[2], v[10]);
1773 0 : u[11] = _mm_sub_epi32(v[3], v[11]);
1774 0 : u[12] = _mm_sub_epi32(v[4], v[12]);
1775 0 : u[13] = _mm_sub_epi32(v[5], v[13]);
1776 0 : u[14] = _mm_sub_epi32(v[6], v[14]);
1777 0 : u[15] = _mm_sub_epi32(v[7], v[15]);
1778 :
1779 0 : v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1780 0 : v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1781 0 : v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1782 0 : v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1783 0 : v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1784 0 : v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1785 0 : v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1786 0 : v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1787 0 : v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1788 0 : v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1789 0 : v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1790 0 : v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1791 0 : v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1792 0 : v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1793 0 : v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1794 0 : v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1795 :
1796 0 : u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
1797 0 : u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
1798 0 : u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
1799 0 : u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
1800 0 : u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
1801 0 : u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
1802 0 : u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
1803 0 : u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
1804 0 : u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
1805 0 : u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
1806 0 : u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
1807 0 : u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
1808 0 : u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
1809 0 : u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
1810 0 : u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
1811 0 : u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
1812 :
1813 0 : x[0] = _mm_add_epi16(s[0], s[4]);
1814 0 : x[1] = _mm_add_epi16(s[1], s[5]);
1815 0 : x[2] = _mm_add_epi16(s[2], s[6]);
1816 0 : x[3] = _mm_add_epi16(s[3], s[7]);
1817 0 : x[4] = _mm_sub_epi16(s[0], s[4]);
1818 0 : x[5] = _mm_sub_epi16(s[1], s[5]);
1819 0 : x[6] = _mm_sub_epi16(s[2], s[6]);
1820 0 : x[7] = _mm_sub_epi16(s[3], s[7]);
1821 0 : x[8] = _mm_packs_epi32(u[0], u[1]);
1822 0 : x[9] = _mm_packs_epi32(u[2], u[3]);
1823 0 : x[10] = _mm_packs_epi32(u[4], u[5]);
1824 0 : x[11] = _mm_packs_epi32(u[6], u[7]);
1825 0 : x[12] = _mm_packs_epi32(u[8], u[9]);
1826 0 : x[13] = _mm_packs_epi32(u[10], u[11]);
1827 0 : x[14] = _mm_packs_epi32(u[12], u[13]);
1828 0 : x[15] = _mm_packs_epi32(u[14], u[15]);
1829 :
1830 : // stage 3
1831 0 : u[0] = _mm_unpacklo_epi16(x[4], x[5]);
1832 0 : u[1] = _mm_unpackhi_epi16(x[4], x[5]);
1833 0 : u[2] = _mm_unpacklo_epi16(x[6], x[7]);
1834 0 : u[3] = _mm_unpackhi_epi16(x[6], x[7]);
1835 0 : u[4] = _mm_unpacklo_epi16(x[12], x[13]);
1836 0 : u[5] = _mm_unpackhi_epi16(x[12], x[13]);
1837 0 : u[6] = _mm_unpacklo_epi16(x[14], x[15]);
1838 0 : u[7] = _mm_unpackhi_epi16(x[14], x[15]);
1839 :
1840 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
1841 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
1842 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
1843 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
1844 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
1845 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
1846 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
1847 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
1848 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
1849 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
1850 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
1851 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
1852 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
1853 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
1854 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
1855 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
1856 :
1857 0 : u[0] = _mm_add_epi32(v[0], v[4]);
1858 0 : u[1] = _mm_add_epi32(v[1], v[5]);
1859 0 : u[2] = _mm_add_epi32(v[2], v[6]);
1860 0 : u[3] = _mm_add_epi32(v[3], v[7]);
1861 0 : u[4] = _mm_sub_epi32(v[0], v[4]);
1862 0 : u[5] = _mm_sub_epi32(v[1], v[5]);
1863 0 : u[6] = _mm_sub_epi32(v[2], v[6]);
1864 0 : u[7] = _mm_sub_epi32(v[3], v[7]);
1865 0 : u[8] = _mm_add_epi32(v[8], v[12]);
1866 0 : u[9] = _mm_add_epi32(v[9], v[13]);
1867 0 : u[10] = _mm_add_epi32(v[10], v[14]);
1868 0 : u[11] = _mm_add_epi32(v[11], v[15]);
1869 0 : u[12] = _mm_sub_epi32(v[8], v[12]);
1870 0 : u[13] = _mm_sub_epi32(v[9], v[13]);
1871 0 : u[14] = _mm_sub_epi32(v[10], v[14]);
1872 0 : u[15] = _mm_sub_epi32(v[11], v[15]);
1873 :
1874 0 : u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1875 0 : u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1876 0 : u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1877 0 : u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1878 0 : u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1879 0 : u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1880 0 : u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1881 0 : u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1882 0 : u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
1883 0 : u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
1884 0 : u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
1885 0 : u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
1886 0 : u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
1887 0 : u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
1888 0 : u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
1889 0 : u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
1890 :
1891 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1892 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1893 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1894 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1895 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1896 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1897 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1898 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1899 0 : v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1900 0 : v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1901 0 : v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1902 0 : v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1903 0 : v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1904 0 : v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1905 0 : v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1906 0 : v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1907 :
1908 0 : s[0] = _mm_add_epi16(x[0], x[2]);
1909 0 : s[1] = _mm_add_epi16(x[1], x[3]);
1910 0 : s[2] = _mm_sub_epi16(x[0], x[2]);
1911 0 : s[3] = _mm_sub_epi16(x[1], x[3]);
1912 0 : s[4] = _mm_packs_epi32(v[0], v[1]);
1913 0 : s[5] = _mm_packs_epi32(v[2], v[3]);
1914 0 : s[6] = _mm_packs_epi32(v[4], v[5]);
1915 0 : s[7] = _mm_packs_epi32(v[6], v[7]);
1916 0 : s[8] = _mm_add_epi16(x[8], x[10]);
1917 0 : s[9] = _mm_add_epi16(x[9], x[11]);
1918 0 : s[10] = _mm_sub_epi16(x[8], x[10]);
1919 0 : s[11] = _mm_sub_epi16(x[9], x[11]);
1920 0 : s[12] = _mm_packs_epi32(v[8], v[9]);
1921 0 : s[13] = _mm_packs_epi32(v[10], v[11]);
1922 0 : s[14] = _mm_packs_epi32(v[12], v[13]);
1923 0 : s[15] = _mm_packs_epi32(v[14], v[15]);
1924 :
1925 : // stage 4
1926 0 : u[0] = _mm_unpacklo_epi16(s[2], s[3]);
1927 0 : u[1] = _mm_unpackhi_epi16(s[2], s[3]);
1928 0 : u[2] = _mm_unpacklo_epi16(s[6], s[7]);
1929 0 : u[3] = _mm_unpackhi_epi16(s[6], s[7]);
1930 0 : u[4] = _mm_unpacklo_epi16(s[10], s[11]);
1931 0 : u[5] = _mm_unpackhi_epi16(s[10], s[11]);
1932 0 : u[6] = _mm_unpacklo_epi16(s[14], s[15]);
1933 0 : u[7] = _mm_unpackhi_epi16(s[14], s[15]);
1934 :
1935 0 : v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
1936 0 : v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
1937 0 : v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
1938 0 : v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
1939 0 : v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
1940 0 : v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
1941 0 : v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
1942 0 : v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
1943 0 : v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
1944 0 : v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
1945 0 : v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
1946 0 : v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
1947 0 : v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
1948 0 : v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
1949 0 : v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
1950 0 : v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
1951 :
1952 0 : u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1953 0 : u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1954 0 : u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1955 0 : u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1956 0 : u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1957 0 : u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1958 0 : u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1959 0 : u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1960 0 : u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
1961 0 : u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
1962 0 : u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
1963 0 : u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
1964 0 : u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
1965 0 : u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
1966 0 : u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
1967 0 : u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
1968 :
1969 0 : v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
1970 0 : v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
1971 0 : v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
1972 0 : v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
1973 0 : v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
1974 0 : v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
1975 0 : v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
1976 0 : v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
1977 0 : v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
1978 0 : v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
1979 0 : v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
1980 0 : v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
1981 0 : v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
1982 0 : v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
1983 0 : v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
1984 0 : v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
1985 :
1986 0 : in[0] = s[0];
1987 0 : in[1] = _mm_sub_epi16(kZero, s[8]);
1988 0 : in[2] = s[12];
1989 0 : in[3] = _mm_sub_epi16(kZero, s[4]);
1990 0 : in[4] = _mm_packs_epi32(v[4], v[5]);
1991 0 : in[5] = _mm_packs_epi32(v[12], v[13]);
1992 0 : in[6] = _mm_packs_epi32(v[8], v[9]);
1993 0 : in[7] = _mm_packs_epi32(v[0], v[1]);
1994 0 : in[8] = _mm_packs_epi32(v[2], v[3]);
1995 0 : in[9] = _mm_packs_epi32(v[10], v[11]);
1996 0 : in[10] = _mm_packs_epi32(v[14], v[15]);
1997 0 : in[11] = _mm_packs_epi32(v[6], v[7]);
1998 0 : in[12] = s[5];
1999 0 : in[13] = _mm_sub_epi16(kZero, s[13]);
2000 0 : in[14] = s[9];
2001 0 : in[15] = _mm_sub_epi16(kZero, s[1]);
2002 0 : }
2003 :
2004 0 : static void fdct16_sse2(__m128i *in0, __m128i *in1) {
2005 0 : fdct16_8col(in0);
2006 0 : fdct16_8col(in1);
2007 0 : array_transpose_16x16(in0, in1);
2008 0 : }
2009 :
2010 0 : static void fadst16_sse2(__m128i *in0, __m128i *in1) {
2011 0 : fadst16_8col(in0);
2012 0 : fadst16_8col(in1);
2013 0 : array_transpose_16x16(in0, in1);
2014 0 : }
2015 :
2016 0 : void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride,
2017 : int tx_type) {
2018 : __m128i in0[16], in1[16];
2019 :
2020 0 : switch (tx_type) {
2021 0 : case DCT_DCT: vpx_fdct16x16_sse2(input, output, stride); break;
2022 : case ADST_DCT:
2023 0 : load_buffer_16x16(input, in0, in1, stride);
2024 0 : fadst16_sse2(in0, in1);
2025 0 : right_shift_16x16(in0, in1);
2026 0 : fdct16_sse2(in0, in1);
2027 0 : write_buffer_16x16(output, in0, in1, 16);
2028 0 : break;
2029 : case DCT_ADST:
2030 0 : load_buffer_16x16(input, in0, in1, stride);
2031 0 : fdct16_sse2(in0, in1);
2032 0 : right_shift_16x16(in0, in1);
2033 0 : fadst16_sse2(in0, in1);
2034 0 : write_buffer_16x16(output, in0, in1, 16);
2035 0 : break;
2036 : case ADST_ADST:
2037 0 : load_buffer_16x16(input, in0, in1, stride);
2038 0 : fadst16_sse2(in0, in1);
2039 0 : right_shift_16x16(in0, in1);
2040 0 : fadst16_sse2(in0, in1);
2041 0 : write_buffer_16x16(output, in0, in1, 16);
2042 0 : break;
2043 0 : default: assert(0); break;
2044 : }
2045 0 : }
|
