?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Refactor vp9_idct.h file
[aom.git] / vpx_dsp / x86 / fwd_txfm_impl_sse2.h
blobabe06aae2103f704432794b3aaa45dfd07e24f78
1 /*
2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <emmintrin.h> // SSE2
12
13 #include "./vpx_dsp_rtcd.h"
14 #include "vp9/encoder/x86/vp9_dct_sse2.h"
15 #include "vpx_dsp/txfm_common.h"
16 #include "vpx_dsp/x86/txfm_common_sse2.h"
17 #include "vpx_ports/mem.h"
18
19 // TODO(jingning) The high bit-depth functions need rework for performance.
20 // After we properly fix the high bit-depth function implementations, this
21 // file's dependency should be substantially simplified.
22 #if DCT_HIGH_BIT_DEPTH
23 #define ADD_EPI16 _mm_adds_epi16
24 #define SUB_EPI16 _mm_subs_epi16
25
26 #else
27 #define ADD_EPI16 _mm_add_epi16
28 #define SUB_EPI16 _mm_sub_epi16
29 #endif
30
31 void FDCT4x4_2D(const int16_t *input, tran_low_t *output, int stride) {
32 // This 2D transform implements 4 vertical 1D transforms followed
33 // by 4 horizontal 1D transforms. The multiplies and adds are as given
34 // by Chen, Smith and Fralick ('77). The commands for moving the data
35 // around have been minimized by hand.
36 // For the purposes of the comments, the 16 inputs are referred to at i0
37 // through iF (in raster order), intermediate variables are a0, b0, c0
38 // through f, and correspond to the in-place computations mapped to input
39 // locations. The outputs, o0 through oF are labeled according to the
40 // output locations.
41
42 // Constants
43 // These are the coefficients used for the multiplies.
44 // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64),
45 // where cospi_N_64 = cos(N pi /64)
46 const __m128i k__cospi_A = octa_set_epi16(cospi_16_64, cospi_16_64,
47 cospi_16_64, cospi_16_64,
48 cospi_16_64, -cospi_16_64,
49 cospi_16_64, -cospi_16_64);
50 const __m128i k__cospi_B = octa_set_epi16(cospi_16_64, -cospi_16_64,
51 cospi_16_64, -cospi_16_64,
52 cospi_16_64, cospi_16_64,
53 cospi_16_64, cospi_16_64);
54 const __m128i k__cospi_C = octa_set_epi16(cospi_8_64, cospi_24_64,
55 cospi_8_64, cospi_24_64,
56 cospi_24_64, -cospi_8_64,
57 cospi_24_64, -cospi_8_64);
58 const __m128i k__cospi_D = octa_set_epi16(cospi_24_64, -cospi_8_64,
59 cospi_24_64, -cospi_8_64,
60 cospi_8_64, cospi_24_64,
61 cospi_8_64, cospi_24_64);
62 const __m128i k__cospi_E = octa_set_epi16(cospi_16_64, cospi_16_64,
63 cospi_16_64, cospi_16_64,
64 cospi_16_64, cospi_16_64,
65 cospi_16_64, cospi_16_64);
66 const __m128i k__cospi_F = octa_set_epi16(cospi_16_64, -cospi_16_64,
67 cospi_16_64, -cospi_16_64,
68 cospi_16_64, -cospi_16_64,
69 cospi_16_64, -cospi_16_64);
70 const __m128i k__cospi_G = octa_set_epi16(cospi_8_64, cospi_24_64,
71 cospi_8_64, cospi_24_64,
72 -cospi_8_64, -cospi_24_64,
73 -cospi_8_64, -cospi_24_64);
74 const __m128i k__cospi_H = octa_set_epi16(cospi_24_64, -cospi_8_64,
75 cospi_24_64, -cospi_8_64,
76 -cospi_24_64, cospi_8_64,
77 -cospi_24_64, cospi_8_64);
78
79 const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
80 // This second rounding constant saves doing some extra adds at the end
81 const __m128i k__DCT_CONST_ROUNDING2 = _mm_set1_epi32(DCT_CONST_ROUNDING
82 +(DCT_CONST_ROUNDING << 1));
83 const int DCT_CONST_BITS2 = DCT_CONST_BITS + 2;
84 const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
85 const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
86 __m128i in0, in1;
87 #if DCT_HIGH_BIT_DEPTH
88 __m128i cmp0, cmp1;
89 int test, overflow;
90 #endif
91
92 // Load inputs.
93 in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
94 in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
95 in1 = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *)
96 (input + 2 * stride)));
97 in0 = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *)
98 (input + 3 * stride)));
99 // in0 = [i0 i1 i2 i3 iC iD iE iF]
100 // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
101 #if DCT_HIGH_BIT_DEPTH
102 // Check inputs small enough to use optimised code
103 cmp0 = _mm_xor_si128(_mm_cmpgt_epi16(in0, _mm_set1_epi16(0x3ff)),
104 _mm_cmplt_epi16(in0, _mm_set1_epi16(0xfc00)));
105 cmp1 = _mm_xor_si128(_mm_cmpgt_epi16(in1, _mm_set1_epi16(0x3ff)),
106 _mm_cmplt_epi16(in1, _mm_set1_epi16(0xfc00)));
107 test = _mm_movemask_epi8(_mm_or_si128(cmp0, cmp1));
108 if (test) {
109 vp9_highbd_fdct4x4_c(input, output, stride);
110 return;
111 }
112 #endif // DCT_HIGH_BIT_DEPTH
113
114 // multiply by 16 to give some extra precision
115 in0 = _mm_slli_epi16(in0, 4);
116 in1 = _mm_slli_epi16(in1, 4);
117 // if (i == 0 && input[0]) input[0] += 1;
118 // add 1 to the upper left pixel if it is non-zero, which helps reduce
119 // the round-trip error
120 {
121 // The mask will only contain whether the first value is zero, all
122 // other comparison will fail as something shifted by 4 (above << 4)
123 // can never be equal to one. To increment in the non-zero case, we
124 // add the mask and one for the first element:
125 // - if zero, mask = -1, v = v - 1 + 1 = v
126 // - if non-zero, mask = 0, v = v + 0 + 1 = v + 1
127 __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a);
128 in0 = _mm_add_epi16(in0, mask);
129 in0 = _mm_add_epi16(in0, k__nonzero_bias_b);
130 }
131 // There are 4 total stages, alternating between an add/subtract stage
132 // followed by an multiply-and-add stage.
133 {
134 // Stage 1: Add/subtract
135
136 // in0 = [i0 i1 i2 i3 iC iD iE iF]
137 // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
138 const __m128i r0 = _mm_unpacklo_epi16(in0, in1);
139 const __m128i r1 = _mm_unpackhi_epi16(in0, in1);
140 // r0 = [i0 i4 i1 i5 i2 i6 i3 i7]
141 // r1 = [iC i8 iD i9 iE iA iF iB]
142 const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4);
143 const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4);
144 // r2 = [i0 i4 i1 i5 i3 i7 i2 i6]
145 // r3 = [iC i8 iD i9 iF iB iE iA]
146
147 const __m128i t0 = _mm_add_epi16(r2, r3);
148 const __m128i t1 = _mm_sub_epi16(r2, r3);
149 // t0 = [a0 a4 a1 a5 a3 a7 a2 a6]
150 // t1 = [aC a8 aD a9 aF aB aE aA]
151
152 // Stage 2: multiply by constants (which gets us into 32 bits).
153 // The constants needed here are:
154 // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16]
155 // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16]
156 // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08]
157 // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24]
158 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
159 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
160 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
161 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
162 // Then add and right-shift to get back to 16-bit range
163 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
164 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
165 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
166 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
167 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
168 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
169 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
170 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
171 // w0 = [b0 b1 b7 b6]
172 // w1 = [b8 b9 bF bE]
173 // w2 = [b4 b5 b3 b2]
174 // w3 = [bC bD bB bA]
175 const __m128i x0 = _mm_packs_epi32(w0, w1);
176 const __m128i x1 = _mm_packs_epi32(w2, w3);
177 #if DCT_HIGH_BIT_DEPTH
178 overflow = check_epi16_overflow_x2(&x0, &x1);
179 if (overflow) {
180 vp9_highbd_fdct4x4_c(input, output, stride);
181 return;
182 }
183 #endif // DCT_HIGH_BIT_DEPTH
184 // x0 = [b0 b1 b7 b6 b8 b9 bF bE]
185 // x1 = [b4 b5 b3 b2 bC bD bB bA]
186 in0 = _mm_shuffle_epi32(x0, 0xD8);
187 in1 = _mm_shuffle_epi32(x1, 0x8D);
188 // in0 = [b0 b1 b8 b9 b7 b6 bF bE]
189 // in1 = [b3 b2 bB bA b4 b5 bC bD]
190 }
191 {
192 // vertical DCTs finished. Now we do the horizontal DCTs.
193 // Stage 3: Add/subtract
194
195 const __m128i t0 = ADD_EPI16(in0, in1);
196 const __m128i t1 = SUB_EPI16(in0, in1);
197 // t0 = [c0 c1 c8 c9 c4 c5 cC cD]
198 // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE]
199 #if DCT_HIGH_BIT_DEPTH
200 overflow = check_epi16_overflow_x2(&t0, &t1);
201 if (overflow) {
202 vp9_highbd_fdct4x4_c(input, output, stride);
203 return;
204 }
205 #endif // DCT_HIGH_BIT_DEPTH
206
207 // Stage 4: multiply by constants (which gets us into 32 bits).
208 {
209 // The constants needed here are:
210 // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16]
211 // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16]
212 // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24]
213 // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08]
214 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
215 const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
216 const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
217 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
218 // Then add and right-shift to get back to 16-bit range
219 // but this combines the final right-shift as well to save operations
220 // This unusual rounding operations is to maintain bit-accurate
221 // compatibility with the c version of this function which has two
222 // rounding steps in a row.
223 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2);
224 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2);
225 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2);
226 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2);
227 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2);
228 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2);
229 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2);
230 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2);
231 // w0 = [o0 o4 o8 oC]
232 // w1 = [o2 o6 oA oE]
233 // w2 = [o1 o5 o9 oD]
234 // w3 = [o3 o7 oB oF]
235 // remember the o's are numbered according to the correct output location
236 const __m128i x0 = _mm_packs_epi32(w0, w1);
237 const __m128i x1 = _mm_packs_epi32(w2, w3);
238 #if DCT_HIGH_BIT_DEPTH
239 overflow = check_epi16_overflow_x2(&x0, &x1);
240 if (overflow) {
241 vp9_highbd_fdct4x4_c(input, output, stride);
242 return;
243 }
244 #endif // DCT_HIGH_BIT_DEPTH
245 {
246 // x0 = [o0 o4 o8 oC o2 o6 oA oE]
247 // x1 = [o1 o5 o9 oD o3 o7 oB oF]
248 const __m128i y0 = _mm_unpacklo_epi16(x0, x1);
249 const __m128i y1 = _mm_unpackhi_epi16(x0, x1);
250 // y0 = [o0 o1 o4 o5 o8 o9 oC oD]
251 // y1 = [o2 o3 o6 o7 oA oB oE oF]
252 in0 = _mm_unpacklo_epi32(y0, y1);
253 // in0 = [o0 o1 o2 o3 o4 o5 o6 o7]
254 in1 = _mm_unpackhi_epi32(y0, y1);
255 // in1 = [o8 o9 oA oB oC oD oE oF]
256 }
257 }
258 }
259 // Post-condition (v + 1) >> 2 is now incorporated into previous
260 // add and right-shift commands. Only 2 store instructions needed
261 // because we are using the fact that 1/3 are stored just after 0/2.
262 storeu_output(&in0, output + 0 * 4);
263 storeu_output(&in1, output + 2 * 4);
264 }
265
266
267 void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) {
268 int pass;
269 // Constants
270 // When we use them, in one case, they are all the same. In all others
271 // it's a pair of them that we need to repeat four times. This is done
272 // by constructing the 32 bit constant corresponding to that pair.
273 const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
274 const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
275 const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
276 const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
277 const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
278 const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
279 const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
280 const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
281 const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
282 #if DCT_HIGH_BIT_DEPTH
283 int overflow;
284 #endif
285 // Load input
286 __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
287 __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
288 __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
289 __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
290 __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
291 __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
292 __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
293 __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
294 // Pre-condition input (shift by two)
295 in0 = _mm_slli_epi16(in0, 2);
296 in1 = _mm_slli_epi16(in1, 2);
297 in2 = _mm_slli_epi16(in2, 2);
298 in3 = _mm_slli_epi16(in3, 2);
299 in4 = _mm_slli_epi16(in4, 2);
300 in5 = _mm_slli_epi16(in5, 2);
301 in6 = _mm_slli_epi16(in6, 2);
302 in7 = _mm_slli_epi16(in7, 2);
303
304 // We do two passes, first the columns, then the rows. The results of the
305 // first pass are transposed so that the same column code can be reused. The
306 // results of the second pass are also transposed so that the rows (processed
307 // as columns) are put back in row positions.
308 for (pass = 0; pass < 2; pass++) {
309 // To store results of each pass before the transpose.
310 __m128i res0, res1, res2, res3, res4, res5, res6, res7;
311 // Add/subtract
312 const __m128i q0 = ADD_EPI16(in0, in7);
313 const __m128i q1 = ADD_EPI16(in1, in6);
314 const __m128i q2 = ADD_EPI16(in2, in5);
315 const __m128i q3 = ADD_EPI16(in3, in4);
316 const __m128i q4 = SUB_EPI16(in3, in4);
317 const __m128i q5 = SUB_EPI16(in2, in5);
318 const __m128i q6 = SUB_EPI16(in1, in6);
319 const __m128i q7 = SUB_EPI16(in0, in7);
320 #if DCT_HIGH_BIT_DEPTH
321 if (pass == 1) {
322 overflow = check_epi16_overflow_x8(&q0, &q1, &q2, &q3,
323 &q4, &q5, &q6, &q7);
324 if (overflow) {
325 vp9_highbd_fdct8x8_c(input, output, stride);
326 return;
327 }
328 }
329 #endif // DCT_HIGH_BIT_DEPTH
330 // Work on first four results
331 {
332 // Add/subtract
333 const __m128i r0 = ADD_EPI16(q0, q3);
334 const __m128i r1 = ADD_EPI16(q1, q2);
335 const __m128i r2 = SUB_EPI16(q1, q2);
336 const __m128i r3 = SUB_EPI16(q0, q3);
337 #if DCT_HIGH_BIT_DEPTH
338 overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
339 if (overflow) {
340 vp9_highbd_fdct8x8_c(input, output, stride);
341 return;
342 }
343 #endif // DCT_HIGH_BIT_DEPTH
344 // Interleave to do the multiply by constants which gets us into 32bits
345 {
346 const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
347 const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
348 const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
349 const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
350 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
351 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
352 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
353 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
354 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
355 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
356 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
357 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
358 // dct_const_round_shift
359 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
360 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
361 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
362 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
363 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
364 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
365 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
366 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
367 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
368 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
369 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
370 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
371 const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
372 const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
373 const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
374 const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
375 // Combine
376 res0 = _mm_packs_epi32(w0, w1);
377 res4 = _mm_packs_epi32(w2, w3);
378 res2 = _mm_packs_epi32(w4, w5);
379 res6 = _mm_packs_epi32(w6, w7);
380 #if DCT_HIGH_BIT_DEPTH
381 overflow = check_epi16_overflow_x4(&res0, &res4, &res2, &res6);
382 if (overflow) {
383 vp9_highbd_fdct8x8_c(input, output, stride);
384 return;
385 }
386 #endif // DCT_HIGH_BIT_DEPTH
387 }
388 }
389 // Work on next four results
390 {
391 // Interleave to do the multiply by constants which gets us into 32bits
392 const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
393 const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
394 const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
395 const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
396 const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
397 const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
398 // dct_const_round_shift
399 const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
400 const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
401 const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
402 const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
403 const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
404 const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
405 const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
406 const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
407 // Combine
408 const __m128i r0 = _mm_packs_epi32(s0, s1);
409 const __m128i r1 = _mm_packs_epi32(s2, s3);
410 #if DCT_HIGH_BIT_DEPTH
411 overflow = check_epi16_overflow_x2(&r0, &r1);
412 if (overflow) {
413 vp9_highbd_fdct8x8_c(input, output, stride);
414 return;
415 }
416 #endif // DCT_HIGH_BIT_DEPTH
417 {
418 // Add/subtract
419 const __m128i x0 = ADD_EPI16(q4, r0);
420 const __m128i x1 = SUB_EPI16(q4, r0);
421 const __m128i x2 = SUB_EPI16(q7, r1);
422 const __m128i x3 = ADD_EPI16(q7, r1);
423 #if DCT_HIGH_BIT_DEPTH
424 overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
425 if (overflow) {
426 vp9_highbd_fdct8x8_c(input, output, stride);
427 return;
428 }
429 #endif // DCT_HIGH_BIT_DEPTH
430 // Interleave to do the multiply by constants which gets us into 32bits
431 {
432 const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
433 const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
434 const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
435 const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
436 const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
437 const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
438 const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
439 const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
440 const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
441 const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
442 const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
443 const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
444 // dct_const_round_shift
445 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
446 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
447 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
448 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
449 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
450 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
451 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
452 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
453 const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
454 const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
455 const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
456 const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
457 const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
458 const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
459 const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
460 const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
461 // Combine
462 res1 = _mm_packs_epi32(w0, w1);
463 res7 = _mm_packs_epi32(w2, w3);
464 res5 = _mm_packs_epi32(w4, w5);
465 res3 = _mm_packs_epi32(w6, w7);
466 #if DCT_HIGH_BIT_DEPTH
467 overflow = check_epi16_overflow_x4(&res1, &res7, &res5, &res3);
468 if (overflow) {
469 vp9_highbd_fdct8x8_c(input, output, stride);
470 return;
471 }
472 #endif // DCT_HIGH_BIT_DEPTH
473 }
474 }
475 }
476 // Transpose the 8x8.
477 {
478 // 00 01 02 03 04 05 06 07
479 // 10 11 12 13 14 15 16 17
480 // 20 21 22 23 24 25 26 27
481 // 30 31 32 33 34 35 36 37
482 // 40 41 42 43 44 45 46 47
483 // 50 51 52 53 54 55 56 57
484 // 60 61 62 63 64 65 66 67
485 // 70 71 72 73 74 75 76 77
486 const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
487 const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
488 const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
489 const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
490 const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
491 const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
492 const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
493 const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
494 // 00 10 01 11 02 12 03 13
495 // 20 30 21 31 22 32 23 33
496 // 04 14 05 15 06 16 07 17
497 // 24 34 25 35 26 36 27 37
498 // 40 50 41 51 42 52 43 53
499 // 60 70 61 71 62 72 63 73
500 // 54 54 55 55 56 56 57 57
501 // 64 74 65 75 66 76 67 77
502 const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
503 const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
504 const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
505 const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
506 const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
507 const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
508 const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
509 const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
510 // 00 10 20 30 01 11 21 31
511 // 40 50 60 70 41 51 61 71
512 // 02 12 22 32 03 13 23 33
513 // 42 52 62 72 43 53 63 73
514 // 04 14 24 34 05 15 21 36
515 // 44 54 64 74 45 55 61 76
516 // 06 16 26 36 07 17 27 37
517 // 46 56 66 76 47 57 67 77
518 in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
519 in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
520 in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
521 in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
522 in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
523 in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
524 in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
525 in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
526 // 00 10 20 30 40 50 60 70
527 // 01 11 21 31 41 51 61 71
528 // 02 12 22 32 42 52 62 72
529 // 03 13 23 33 43 53 63 73
530 // 04 14 24 34 44 54 64 74
531 // 05 15 25 35 45 55 65 75
532 // 06 16 26 36 46 56 66 76
533 // 07 17 27 37 47 57 67 77
534 }
535 }
536 // Post-condition output and store it
537 {
538 // Post-condition (division by two)
539 // division of two 16 bits signed numbers using shifts
540 // n / 2 = (n - (n >> 15)) >> 1
541 const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
542 const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
543 const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
544 const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
545 const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
546 const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
547 const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
548 const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
549 in0 = _mm_sub_epi16(in0, sign_in0);
550 in1 = _mm_sub_epi16(in1, sign_in1);
551 in2 = _mm_sub_epi16(in2, sign_in2);
552 in3 = _mm_sub_epi16(in3, sign_in3);
553 in4 = _mm_sub_epi16(in4, sign_in4);
554 in5 = _mm_sub_epi16(in5, sign_in5);
555 in6 = _mm_sub_epi16(in6, sign_in6);
556 in7 = _mm_sub_epi16(in7, sign_in7);
557 in0 = _mm_srai_epi16(in0, 1);
558 in1 = _mm_srai_epi16(in1, 1);
559 in2 = _mm_srai_epi16(in2, 1);
560 in3 = _mm_srai_epi16(in3, 1);
561 in4 = _mm_srai_epi16(in4, 1);
562 in5 = _mm_srai_epi16(in5, 1);
563 in6 = _mm_srai_epi16(in6, 1);
564 in7 = _mm_srai_epi16(in7, 1);
565 // store results
566 store_output(&in0, (output + 0 * 8));
567 store_output(&in1, (output + 1 * 8));
568 store_output(&in2, (output + 2 * 8));
569 store_output(&in3, (output + 3 * 8));
570 store_output(&in4, (output + 4 * 8));
571 store_output(&in5, (output + 5 * 8));
572 store_output(&in6, (output + 6 * 8));
573 store_output(&in7, (output + 7 * 8));
574 }
575 }
576
577 void FDCT16x16_2D(const int16_t *input, tran_low_t *output, int stride) {
578 // The 2D transform is done with two passes which are actually pretty
579 // similar. In the first one, we transform the columns and transpose
580 // the results. In the second one, we transform the rows. To achieve that,
581 // as the first pass results are transposed, we transpose the columns (that
582 // is the transposed rows) and transpose the results (so that it goes back
583 // in normal/row positions).
584 int pass;
585 // We need an intermediate buffer between passes.
586 DECLARE_ALIGNED(16, int16_t, intermediate[256]);
587 const int16_t *in = input;
588 int16_t *out0 = intermediate;
589 tran_low_t *out1 = output;
590 // Constants
591 // When we use them, in one case, they are all the same. In all others
592 // it's a pair of them that we need to repeat four times. This is done
593 // by constructing the 32 bit constant corresponding to that pair.
594 const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
595 const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
596 const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
597 const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
598 const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
599 const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
600 const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
601 const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
602 const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
603 const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
604 const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
605 const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
606 const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
607 const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
608 const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
609 const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
610 const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
611 const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
612 const __m128i kOne = _mm_set1_epi16(1);
613 // Do the two transform/transpose passes
614 for (pass = 0; pass < 2; ++pass) {
615 // We process eight columns (transposed rows in second pass) at a time.
616 int column_start;
617 #if DCT_HIGH_BIT_DEPTH
618 int overflow;
619 #endif
620 for (column_start = 0; column_start < 16; column_start += 8) {
621 __m128i in00, in01, in02, in03, in04, in05, in06, in07;
622 __m128i in08, in09, in10, in11, in12, in13, in14, in15;
623 __m128i input0, input1, input2, input3, input4, input5, input6, input7;
624 __m128i step1_0, step1_1, step1_2, step1_3;
625 __m128i step1_4, step1_5, step1_6, step1_7;
626 __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
627 __m128i step3_0, step3_1, step3_2, step3_3;
628 __m128i step3_4, step3_5, step3_6, step3_7;
629 __m128i res00, res01, res02, res03, res04, res05, res06, res07;
630 __m128i res08, res09, res10, res11, res12, res13, res14, res15;
631 // Load and pre-condition input.
632 if (0 == pass) {
633 in00 = _mm_load_si128((const __m128i *)(in + 0 * stride));
634 in01 = _mm_load_si128((const __m128i *)(in + 1 * stride));
635 in02 = _mm_load_si128((const __m128i *)(in + 2 * stride));
636 in03 = _mm_load_si128((const __m128i *)(in + 3 * stride));
637 in04 = _mm_load_si128((const __m128i *)(in + 4 * stride));
638 in05 = _mm_load_si128((const __m128i *)(in + 5 * stride));
639 in06 = _mm_load_si128((const __m128i *)(in + 6 * stride));
640 in07 = _mm_load_si128((const __m128i *)(in + 7 * stride));
641 in08 = _mm_load_si128((const __m128i *)(in + 8 * stride));
642 in09 = _mm_load_si128((const __m128i *)(in + 9 * stride));
643 in10 = _mm_load_si128((const __m128i *)(in + 10 * stride));
644 in11 = _mm_load_si128((const __m128i *)(in + 11 * stride));
645 in12 = _mm_load_si128((const __m128i *)(in + 12 * stride));
646 in13 = _mm_load_si128((const __m128i *)(in + 13 * stride));
647 in14 = _mm_load_si128((const __m128i *)(in + 14 * stride));
648 in15 = _mm_load_si128((const __m128i *)(in + 15 * stride));
649 // x = x << 2
650 in00 = _mm_slli_epi16(in00, 2);
651 in01 = _mm_slli_epi16(in01, 2);
652 in02 = _mm_slli_epi16(in02, 2);
653 in03 = _mm_slli_epi16(in03, 2);
654 in04 = _mm_slli_epi16(in04, 2);
655 in05 = _mm_slli_epi16(in05, 2);
656 in06 = _mm_slli_epi16(in06, 2);
657 in07 = _mm_slli_epi16(in07, 2);
658 in08 = _mm_slli_epi16(in08, 2);
659 in09 = _mm_slli_epi16(in09, 2);
660 in10 = _mm_slli_epi16(in10, 2);
661 in11 = _mm_slli_epi16(in11, 2);
662 in12 = _mm_slli_epi16(in12, 2);
663 in13 = _mm_slli_epi16(in13, 2);
664 in14 = _mm_slli_epi16(in14, 2);
665 in15 = _mm_slli_epi16(in15, 2);
666 } else {
667 in00 = _mm_load_si128((const __m128i *)(in + 0 * 16));
668 in01 = _mm_load_si128((const __m128i *)(in + 1 * 16));
669 in02 = _mm_load_si128((const __m128i *)(in + 2 * 16));
670 in03 = _mm_load_si128((const __m128i *)(in + 3 * 16));
671 in04 = _mm_load_si128((const __m128i *)(in + 4 * 16));
672 in05 = _mm_load_si128((const __m128i *)(in + 5 * 16));
673 in06 = _mm_load_si128((const __m128i *)(in + 6 * 16));
674 in07 = _mm_load_si128((const __m128i *)(in + 7 * 16));
675 in08 = _mm_load_si128((const __m128i *)(in + 8 * 16));
676 in09 = _mm_load_si128((const __m128i *)(in + 9 * 16));
677 in10 = _mm_load_si128((const __m128i *)(in + 10 * 16));
678 in11 = _mm_load_si128((const __m128i *)(in + 11 * 16));
679 in12 = _mm_load_si128((const __m128i *)(in + 12 * 16));
680 in13 = _mm_load_si128((const __m128i *)(in + 13 * 16));
681 in14 = _mm_load_si128((const __m128i *)(in + 14 * 16));
682 in15 = _mm_load_si128((const __m128i *)(in + 15 * 16));
683 // x = (x + 1) >> 2
684 in00 = _mm_add_epi16(in00, kOne);
685 in01 = _mm_add_epi16(in01, kOne);
686 in02 = _mm_add_epi16(in02, kOne);
687 in03 = _mm_add_epi16(in03, kOne);
688 in04 = _mm_add_epi16(in04, kOne);
689 in05 = _mm_add_epi16(in05, kOne);
690 in06 = _mm_add_epi16(in06, kOne);
691 in07 = _mm_add_epi16(in07, kOne);
692 in08 = _mm_add_epi16(in08, kOne);
693 in09 = _mm_add_epi16(in09, kOne);
694 in10 = _mm_add_epi16(in10, kOne);
695 in11 = _mm_add_epi16(in11, kOne);
696 in12 = _mm_add_epi16(in12, kOne);
697 in13 = _mm_add_epi16(in13, kOne);
698 in14 = _mm_add_epi16(in14, kOne);
699 in15 = _mm_add_epi16(in15, kOne);
700 in00 = _mm_srai_epi16(in00, 2);
701 in01 = _mm_srai_epi16(in01, 2);
702 in02 = _mm_srai_epi16(in02, 2);
703 in03 = _mm_srai_epi16(in03, 2);
704 in04 = _mm_srai_epi16(in04, 2);
705 in05 = _mm_srai_epi16(in05, 2);
706 in06 = _mm_srai_epi16(in06, 2);
707 in07 = _mm_srai_epi16(in07, 2);
708 in08 = _mm_srai_epi16(in08, 2);
709 in09 = _mm_srai_epi16(in09, 2);
710 in10 = _mm_srai_epi16(in10, 2);
711 in11 = _mm_srai_epi16(in11, 2);
712 in12 = _mm_srai_epi16(in12, 2);
713 in13 = _mm_srai_epi16(in13, 2);
714 in14 = _mm_srai_epi16(in14, 2);
715 in15 = _mm_srai_epi16(in15, 2);
716 }
717 in += 8;
718 // Calculate input for the first 8 results.
719 {
720 input0 = ADD_EPI16(in00, in15);
721 input1 = ADD_EPI16(in01, in14);
722 input2 = ADD_EPI16(in02, in13);
723 input3 = ADD_EPI16(in03, in12);
724 input4 = ADD_EPI16(in04, in11);
725 input5 = ADD_EPI16(in05, in10);
726 input6 = ADD_EPI16(in06, in09);
727 input7 = ADD_EPI16(in07, in08);
728 #if DCT_HIGH_BIT_DEPTH
729 overflow = check_epi16_overflow_x8(&input0, &input1, &input2, &input3,
730 &input4, &input5, &input6, &input7);
731 if (overflow) {
732 vp9_highbd_fdct16x16_c(input, output, stride);
733 return;
734 }
735 #endif // DCT_HIGH_BIT_DEPTH
736 }
737 // Calculate input for the next 8 results.
738 {
739 step1_0 = SUB_EPI16(in07, in08);
740 step1_1 = SUB_EPI16(in06, in09);
741 step1_2 = SUB_EPI16(in05, in10);
742 step1_3 = SUB_EPI16(in04, in11);
743 step1_4 = SUB_EPI16(in03, in12);
744 step1_5 = SUB_EPI16(in02, in13);
745 step1_6 = SUB_EPI16(in01, in14);
746 step1_7 = SUB_EPI16(in00, in15);
747 #if DCT_HIGH_BIT_DEPTH
748 overflow = check_epi16_overflow_x8(&step1_0, &step1_1,
749 &step1_2, &step1_3,
750 &step1_4, &step1_5,
751 &step1_6, &step1_7);
752 if (overflow) {
753 vp9_highbd_fdct16x16_c(input, output, stride);
754 return;
755 }
756 #endif // DCT_HIGH_BIT_DEPTH
757 }
758 // Work on the first eight values; fdct8(input, even_results);
759 {
760 // Add/subtract
761 const __m128i q0 = ADD_EPI16(input0, input7);
762 const __m128i q1 = ADD_EPI16(input1, input6);
763 const __m128i q2 = ADD_EPI16(input2, input5);
764 const __m128i q3 = ADD_EPI16(input3, input4);
765 const __m128i q4 = SUB_EPI16(input3, input4);
766 const __m128i q5 = SUB_EPI16(input2, input5);
767 const __m128i q6 = SUB_EPI16(input1, input6);
768 const __m128i q7 = SUB_EPI16(input0, input7);
769 #if DCT_HIGH_BIT_DEPTH
770 overflow = check_epi16_overflow_x8(&q0, &q1, &q2, &q3,
771 &q4, &q5, &q6, &q7);
772 if (overflow) {
773 vp9_highbd_fdct16x16_c(input, output, stride);
774 return;
775 }
776 #endif // DCT_HIGH_BIT_DEPTH
777 // Work on first four results
778 {
779 // Add/subtract
780 const __m128i r0 = ADD_EPI16(q0, q3);
781 const __m128i r1 = ADD_EPI16(q1, q2);
782 const __m128i r2 = SUB_EPI16(q1, q2);
783 const __m128i r3 = SUB_EPI16(q0, q3);
784 #if DCT_HIGH_BIT_DEPTH
785 overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
786 if (overflow) {
787 vp9_highbd_fdct16x16_c(input, output, stride);
788 return;
789 }
790 #endif // DCT_HIGH_BIT_DEPTH
791 // Interleave to do the multiply by constants which gets us
792 // into 32 bits.
793 {
794 const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
795 const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
796 const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
797 const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
798 res00 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
799 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
800 res08 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
801 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
802 res04 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
803 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
804 res12 = mult_round_shift(&t2, &t3, &k__cospi_m08_p24,
805 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
806 #if DCT_HIGH_BIT_DEPTH
807 overflow = check_epi16_overflow_x4(&res00, &res08, &res04, &res12);
808 if (overflow) {
809 vp9_highbd_fdct16x16_c(input, output, stride);
810 return;
811 }
812 #endif // DCT_HIGH_BIT_DEPTH
813 }
814 }
815 // Work on next four results
816 {
817 // Interleave to do the multiply by constants which gets us
818 // into 32 bits.
819 const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
820 const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
821 const __m128i r0 = mult_round_shift(&d0, &d1, &k__cospi_p16_m16,
822 &k__DCT_CONST_ROUNDING,
823 DCT_CONST_BITS);
824 const __m128i r1 = mult_round_shift(&d0, &d1, &k__cospi_p16_p16,
825 &k__DCT_CONST_ROUNDING,
826 DCT_CONST_BITS);
827 #if DCT_HIGH_BIT_DEPTH
828 overflow = check_epi16_overflow_x2(&r0, &r1);
829 if (overflow) {
830 vp9_highbd_fdct16x16_c(input, output, stride);
831 return;
832 }
833 #endif // DCT_HIGH_BIT_DEPTH
834 {
835 // Add/subtract
836 const __m128i x0 = ADD_EPI16(q4, r0);
837 const __m128i x1 = SUB_EPI16(q4, r0);
838 const __m128i x2 = SUB_EPI16(q7, r1);
839 const __m128i x3 = ADD_EPI16(q7, r1);
840 #if DCT_HIGH_BIT_DEPTH
841 overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
842 if (overflow) {
843 vp9_highbd_fdct16x16_c(input, output, stride);
844 return;
845 }
846 #endif // DCT_HIGH_BIT_DEPTH
847 // Interleave to do the multiply by constants which gets us
848 // into 32 bits.
849 {
850 const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
851 const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
852 const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
853 const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
854 res02 = mult_round_shift(&t0, &t1, &k__cospi_p28_p04,
855 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
856 res14 = mult_round_shift(&t0, &t1, &k__cospi_m04_p28,
857 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
858 res10 = mult_round_shift(&t2, &t3, &k__cospi_p12_p20,
859 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
860 res06 = mult_round_shift(&t2, &t3, &k__cospi_m20_p12,
861 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
862 #if DCT_HIGH_BIT_DEPTH
863 overflow = check_epi16_overflow_x4(&res02, &res14,
864 &res10, &res06);
865 if (overflow) {
866 vp9_highbd_fdct16x16_c(input, output, stride);
867 return;
868 }
869 #endif // DCT_HIGH_BIT_DEPTH
870 }
871 }
872 }
873 }
874 // Work on the next eight values; step1 -> odd_results
875 {
876 // step 2
877 {
878 const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
879 const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
880 const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
881 const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
882 step2_2 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
883 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
884 step2_3 = mult_round_shift(&t2, &t3, &k__cospi_p16_m16,
885 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
886 step2_5 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
887 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
888 step2_4 = mult_round_shift(&t2, &t3, &k__cospi_p16_p16,
889 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
890 #if DCT_HIGH_BIT_DEPTH
891 overflow = check_epi16_overflow_x4(&step2_2, &step2_3, &step2_5,
892 &step2_4);
893 if (overflow) {
894 vp9_highbd_fdct16x16_c(input, output, stride);
895 return;
896 }
897 #endif // DCT_HIGH_BIT_DEPTH
898 }
899 // step 3
900 {
901 step3_0 = ADD_EPI16(step1_0, step2_3);
902 step3_1 = ADD_EPI16(step1_1, step2_2);
903 step3_2 = SUB_EPI16(step1_1, step2_2);
904 step3_3 = SUB_EPI16(step1_0, step2_3);
905 step3_4 = SUB_EPI16(step1_7, step2_4);
906 step3_5 = SUB_EPI16(step1_6, step2_5);
907 step3_6 = ADD_EPI16(step1_6, step2_5);
908 step3_7 = ADD_EPI16(step1_7, step2_4);
909 #if DCT_HIGH_BIT_DEPTH
910 overflow = check_epi16_overflow_x8(&step3_0, &step3_1,
911 &step3_2, &step3_3,
912 &step3_4, &step3_5,
913 &step3_6, &step3_7);
914 if (overflow) {
915 vp9_highbd_fdct16x16_c(input, output, stride);
916 return;
917 }
918 #endif // DCT_HIGH_BIT_DEPTH
919 }
920 // step 4
921 {
922 const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
923 const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
924 const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
925 const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
926 step2_1 = mult_round_shift(&t0, &t1, &k__cospi_m08_p24,
927 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
928 step2_2 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
929 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
930 step2_6 = mult_round_shift(&t0, &t1, &k__cospi_p24_p08,
931 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
932 step2_5 = mult_round_shift(&t2, &t3, &k__cospi_p08_m24,
933 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
934 #if DCT_HIGH_BIT_DEPTH
935 overflow = check_epi16_overflow_x4(&step2_1, &step2_2, &step2_6,
936 &step2_5);
937 if (overflow) {
938 vp9_highbd_fdct16x16_c(input, output, stride);
939 return;
940 }
941 #endif // DCT_HIGH_BIT_DEPTH
942 }
943 // step 5
944 {
945 step1_0 = ADD_EPI16(step3_0, step2_1);
946 step1_1 = SUB_EPI16(step3_0, step2_1);
947 step1_2 = ADD_EPI16(step3_3, step2_2);
948 step1_3 = SUB_EPI16(step3_3, step2_2);
949 step1_4 = SUB_EPI16(step3_4, step2_5);
950 step1_5 = ADD_EPI16(step3_4, step2_5);
951 step1_6 = SUB_EPI16(step3_7, step2_6);
952 step1_7 = ADD_EPI16(step3_7, step2_6);
953 #if DCT_HIGH_BIT_DEPTH
954 overflow = check_epi16_overflow_x8(&step1_0, &step1_1,
955 &step1_2, &step1_3,
956 &step1_4, &step1_5,
957 &step1_6, &step1_7);
958 if (overflow) {
959 vp9_highbd_fdct16x16_c(input, output, stride);
960 return;
961 }
962 #endif // DCT_HIGH_BIT_DEPTH
963 }
964 // step 6
965 {
966 const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
967 const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
968 const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
969 const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
970 res01 = mult_round_shift(&t0, &t1, &k__cospi_p30_p02,
971 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
972 res09 = mult_round_shift(&t2, &t3, &k__cospi_p14_p18,
973 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
974 res15 = mult_round_shift(&t0, &t1, &k__cospi_m02_p30,
975 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
976 res07 = mult_round_shift(&t2, &t3, &k__cospi_m18_p14,
977 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
978 #if DCT_HIGH_BIT_DEPTH
979 overflow = check_epi16_overflow_x4(&res01, &res09, &res15, &res07);
980 if (overflow) {
981 vp9_highbd_fdct16x16_c(input, output, stride);
982 return;
983 }
984 #endif // DCT_HIGH_BIT_DEPTH
985 }
986 {
987 const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
988 const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
989 const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
990 const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
991 res05 = mult_round_shift(&t0, &t1, &k__cospi_p22_p10,
992 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
993 res13 = mult_round_shift(&t2, &t3, &k__cospi_p06_p26,
994 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
995 res11 = mult_round_shift(&t0, &t1, &k__cospi_m10_p22,
996 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
997 res03 = mult_round_shift(&t2, &t3, &k__cospi_m26_p06,
998 &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
999 #if DCT_HIGH_BIT_DEPTH
1000 overflow = check_epi16_overflow_x4(&res05, &res13, &res11, &res03);
1001 if (overflow) {
1002 vp9_highbd_fdct16x16_c(input, output, stride);
1003 return;
1004 }
1005 #endif // DCT_HIGH_BIT_DEPTH
1006 }
1007 }
1008 // Transpose the results, do it as two 8x8 transposes.
1009 transpose_and_output8x8(&res00, &res01, &res02, &res03,
1010 &res04, &res05, &res06, &res07,
1011 pass, out0, out1);
1012 transpose_and_output8x8(&res08, &res09, &res10, &res11,
1013 &res12, &res13, &res14, &res15,
1014 pass, out0 + 8, out1 + 8);
1015 if (pass == 0) {
1016 out0 += 8*16;
1017 } else {
1018 out1 += 8*16;
1019 }
1020 }
1021 // Setup in/out for next pass.
1022 in = intermediate;
1023 }
1024 }
1025
1026 #undef ADD_EPI16
1027 #undef SUB_EPI16
AV1 Codec Library