1// Copyright 2011 Google Inc. All Rights Reserved.
2//
3// Use of this source code is governed by a BSD-style license
4// that can be found in the COPYING file in the root of the source
5// tree. An additional intellectual property rights grant can be found
6// in the file PATENTS. All contributing project authors may
7// be found in the AUTHORS file in the root of the source tree.
8// -----------------------------------------------------------------------------
9//
10// SSE2 version of speed-critical encoding functions.
11//
12// Author: Christian Duvivier (cduvivier@google.com)
13
14#include "./dsp.h"
15
16#if defined(WEBP_USE_SSE2)
17#include <stdlib.h>  // for abs()
18#include <emmintrin.h>
19
20#include "../enc/cost.h"
21#include "../enc/vp8enci.h"
22
23//------------------------------------------------------------------------------
24// Quite useful macro for debugging. Left here for convenience.
25
26#if 0
27#include <stdio.h>
28static void PrintReg(const __m128i r, const char* const name, int size) {
29  int n;
30  union {
31    __m128i r;
32    uint8_t i8[16];
33    uint16_t i16[8];
34    uint32_t i32[4];
35    uint64_t i64[2];
36  } tmp;
37  tmp.r = r;
38  fprintf(stderr, "%s\t: ", name);
39  if (size == 8) {
40    for (n = 0; n < 16; ++n) fprintf(stderr, "%.2x ", tmp.i8[n]);
41  } else if (size == 16) {
42    for (n = 0; n < 8; ++n) fprintf(stderr, "%.4x ", tmp.i16[n]);
43  } else if (size == 32) {
44    for (n = 0; n < 4; ++n) fprintf(stderr, "%.8x ", tmp.i32[n]);
45  } else {
46    for (n = 0; n < 2; ++n) fprintf(stderr, "%.16lx ", tmp.i64[n]);
47  }
48  fprintf(stderr, "\n");
49}
50#endif
51
52//------------------------------------------------------------------------------
53// Transforms (Paragraph 14.4)
54
55// Does one or two inverse transforms.
56static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
57                       int do_two) {
58  // This implementation makes use of 16-bit fixed point versions of two
59  // multiply constants:
60  //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
61  //    K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16
62  //
63  // To be able to use signed 16-bit integers, we use the following trick to
64  // have constants within range:
65  // - Associated constants are obtained by subtracting the 16-bit fixed point
66  //   version of one:
67  //      k = K - (1 << 16)  =>  K = k + (1 << 16)
68  //      K1 = 85267  =>  k1 =  20091
69  //      K2 = 35468  =>  k2 = -30068
70  // - The multiplication of a variable by a constant become the sum of the
71  //   variable and the multiplication of that variable by the associated
72  //   constant:
73  //      (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
74  const __m128i k1 = _mm_set1_epi16(20091);
75  const __m128i k2 = _mm_set1_epi16(-30068);
76  __m128i T0, T1, T2, T3;
77
78  // Load and concatenate the transform coefficients (we'll do two inverse
79  // transforms in parallel). In the case of only one inverse transform, the
80  // second half of the vectors will just contain random value we'll never
81  // use nor store.
82  __m128i in0, in1, in2, in3;
83  {
84    in0 = _mm_loadl_epi64((const __m128i*)&in[0]);
85    in1 = _mm_loadl_epi64((const __m128i*)&in[4]);
86    in2 = _mm_loadl_epi64((const __m128i*)&in[8]);
87    in3 = _mm_loadl_epi64((const __m128i*)&in[12]);
88    // a00 a10 a20 a30   x x x x
89    // a01 a11 a21 a31   x x x x
90    // a02 a12 a22 a32   x x x x
91    // a03 a13 a23 a33   x x x x
92    if (do_two) {
93      const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]);
94      const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]);
95      const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]);
96      const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]);
97      in0 = _mm_unpacklo_epi64(in0, inB0);
98      in1 = _mm_unpacklo_epi64(in1, inB1);
99      in2 = _mm_unpacklo_epi64(in2, inB2);
100      in3 = _mm_unpacklo_epi64(in3, inB3);
101      // a00 a10 a20 a30   b00 b10 b20 b30
102      // a01 a11 a21 a31   b01 b11 b21 b31
103      // a02 a12 a22 a32   b02 b12 b22 b32
104      // a03 a13 a23 a33   b03 b13 b23 b33
105    }
106  }
107
108  // Vertical pass and subsequent transpose.
109  {
110    // First pass, c and d calculations are longer because of the "trick"
111    // multiplications.
112    const __m128i a = _mm_add_epi16(in0, in2);
113    const __m128i b = _mm_sub_epi16(in0, in2);
114    // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
115    const __m128i c1 = _mm_mulhi_epi16(in1, k2);
116    const __m128i c2 = _mm_mulhi_epi16(in3, k1);
117    const __m128i c3 = _mm_sub_epi16(in1, in3);
118    const __m128i c4 = _mm_sub_epi16(c1, c2);
119    const __m128i c = _mm_add_epi16(c3, c4);
120    // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
121    const __m128i d1 = _mm_mulhi_epi16(in1, k1);
122    const __m128i d2 = _mm_mulhi_epi16(in3, k2);
123    const __m128i d3 = _mm_add_epi16(in1, in3);
124    const __m128i d4 = _mm_add_epi16(d1, d2);
125    const __m128i d = _mm_add_epi16(d3, d4);
126
127    // Second pass.
128    const __m128i tmp0 = _mm_add_epi16(a, d);
129    const __m128i tmp1 = _mm_add_epi16(b, c);
130    const __m128i tmp2 = _mm_sub_epi16(b, c);
131    const __m128i tmp3 = _mm_sub_epi16(a, d);
132
133    // Transpose the two 4x4.
134    // a00 a01 a02 a03   b00 b01 b02 b03
135    // a10 a11 a12 a13   b10 b11 b12 b13
136    // a20 a21 a22 a23   b20 b21 b22 b23
137    // a30 a31 a32 a33   b30 b31 b32 b33
138    const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1);
139    const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3);
140    const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1);
141    const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3);
142    // a00 a10 a01 a11   a02 a12 a03 a13
143    // a20 a30 a21 a31   a22 a32 a23 a33
144    // b00 b10 b01 b11   b02 b12 b03 b13
145    // b20 b30 b21 b31   b22 b32 b23 b33
146    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
147    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
148    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
149    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
150    // a00 a10 a20 a30 a01 a11 a21 a31
151    // b00 b10 b20 b30 b01 b11 b21 b31
152    // a02 a12 a22 a32 a03 a13 a23 a33
153    // b02 b12 a22 b32 b03 b13 b23 b33
154    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
155    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
156    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
157    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
158    // a00 a10 a20 a30   b00 b10 b20 b30
159    // a01 a11 a21 a31   b01 b11 b21 b31
160    // a02 a12 a22 a32   b02 b12 b22 b32
161    // a03 a13 a23 a33   b03 b13 b23 b33
162  }
163
164  // Horizontal pass and subsequent transpose.
165  {
166    // First pass, c and d calculations are longer because of the "trick"
167    // multiplications.
168    const __m128i four = _mm_set1_epi16(4);
169    const __m128i dc = _mm_add_epi16(T0, four);
170    const __m128i a =  _mm_add_epi16(dc, T2);
171    const __m128i b =  _mm_sub_epi16(dc, T2);
172    // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
173    const __m128i c1 = _mm_mulhi_epi16(T1, k2);
174    const __m128i c2 = _mm_mulhi_epi16(T3, k1);
175    const __m128i c3 = _mm_sub_epi16(T1, T3);
176    const __m128i c4 = _mm_sub_epi16(c1, c2);
177    const __m128i c = _mm_add_epi16(c3, c4);
178    // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
179    const __m128i d1 = _mm_mulhi_epi16(T1, k1);
180    const __m128i d2 = _mm_mulhi_epi16(T3, k2);
181    const __m128i d3 = _mm_add_epi16(T1, T3);
182    const __m128i d4 = _mm_add_epi16(d1, d2);
183    const __m128i d = _mm_add_epi16(d3, d4);
184
185    // Second pass.
186    const __m128i tmp0 = _mm_add_epi16(a, d);
187    const __m128i tmp1 = _mm_add_epi16(b, c);
188    const __m128i tmp2 = _mm_sub_epi16(b, c);
189    const __m128i tmp3 = _mm_sub_epi16(a, d);
190    const __m128i shifted0 = _mm_srai_epi16(tmp0, 3);
191    const __m128i shifted1 = _mm_srai_epi16(tmp1, 3);
192    const __m128i shifted2 = _mm_srai_epi16(tmp2, 3);
193    const __m128i shifted3 = _mm_srai_epi16(tmp3, 3);
194
195    // Transpose the two 4x4.
196    // a00 a01 a02 a03   b00 b01 b02 b03
197    // a10 a11 a12 a13   b10 b11 b12 b13
198    // a20 a21 a22 a23   b20 b21 b22 b23
199    // a30 a31 a32 a33   b30 b31 b32 b33
200    const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1);
201    const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3);
202    const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1);
203    const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3);
204    // a00 a10 a01 a11   a02 a12 a03 a13
205    // a20 a30 a21 a31   a22 a32 a23 a33
206    // b00 b10 b01 b11   b02 b12 b03 b13
207    // b20 b30 b21 b31   b22 b32 b23 b33
208    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
209    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
210    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
211    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
212    // a00 a10 a20 a30 a01 a11 a21 a31
213    // b00 b10 b20 b30 b01 b11 b21 b31
214    // a02 a12 a22 a32 a03 a13 a23 a33
215    // b02 b12 a22 b32 b03 b13 b23 b33
216    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
217    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
218    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
219    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
220    // a00 a10 a20 a30   b00 b10 b20 b30
221    // a01 a11 a21 a31   b01 b11 b21 b31
222    // a02 a12 a22 a32   b02 b12 b22 b32
223    // a03 a13 a23 a33   b03 b13 b23 b33
224  }
225
226  // Add inverse transform to 'ref' and store.
227  {
228    const __m128i zero = _mm_setzero_si128();
229    // Load the reference(s).
230    __m128i ref0, ref1, ref2, ref3;
231    if (do_two) {
232      // Load eight bytes/pixels per line.
233      ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
234      ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
235      ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
236      ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
237    } else {
238      // Load four bytes/pixels per line.
239      ref0 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[0 * BPS]));
240      ref1 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[1 * BPS]));
241      ref2 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[2 * BPS]));
242      ref3 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[3 * BPS]));
243    }
244    // Convert to 16b.
245    ref0 = _mm_unpacklo_epi8(ref0, zero);
246    ref1 = _mm_unpacklo_epi8(ref1, zero);
247    ref2 = _mm_unpacklo_epi8(ref2, zero);
248    ref3 = _mm_unpacklo_epi8(ref3, zero);
249    // Add the inverse transform(s).
250    ref0 = _mm_add_epi16(ref0, T0);
251    ref1 = _mm_add_epi16(ref1, T1);
252    ref2 = _mm_add_epi16(ref2, T2);
253    ref3 = _mm_add_epi16(ref3, T3);
254    // Unsigned saturate to 8b.
255    ref0 = _mm_packus_epi16(ref0, ref0);
256    ref1 = _mm_packus_epi16(ref1, ref1);
257    ref2 = _mm_packus_epi16(ref2, ref2);
258    ref3 = _mm_packus_epi16(ref3, ref3);
259    // Store the results.
260    if (do_two) {
261      // Store eight bytes/pixels per line.
262      _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
263      _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
264      _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
265      _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
266    } else {
267      // Store four bytes/pixels per line.
268      WebPUint32ToMem(&dst[0 * BPS], _mm_cvtsi128_si32(ref0));
269      WebPUint32ToMem(&dst[1 * BPS], _mm_cvtsi128_si32(ref1));
270      WebPUint32ToMem(&dst[2 * BPS], _mm_cvtsi128_si32(ref2));
271      WebPUint32ToMem(&dst[3 * BPS], _mm_cvtsi128_si32(ref3));
272    }
273  }
274}
275
276static void FTransformPass1(const __m128i* const in01,
277                            const __m128i* const in23,
278                            __m128i* const out01,
279                            __m128i* const out32) {
280  const __m128i k937 = _mm_set1_epi32(937);
281  const __m128i k1812 = _mm_set1_epi32(1812);
282
283  const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8);
284  const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8);
285  const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352,
286                                            2217, 5352, 2217, 5352);
287  const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217,
288                                            -5352, 2217, -5352, 2217);
289
290  // *in01 = 00 01 10 11 02 03 12 13
291  // *in23 = 20 21 30 31 22 23 32 33
292  const __m128i shuf01_p = _mm_shufflehi_epi16(*in01, _MM_SHUFFLE(2, 3, 0, 1));
293  const __m128i shuf23_p = _mm_shufflehi_epi16(*in23, _MM_SHUFFLE(2, 3, 0, 1));
294  // 00 01 10 11 03 02 13 12
295  // 20 21 30 31 23 22 33 32
296  const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p);
297  const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p);
298  // 00 01 10 11 20 21 30 31
299  // 03 02 13 12 23 22 33 32
300  const __m128i a01 = _mm_add_epi16(s01, s32);
301  const __m128i a32 = _mm_sub_epi16(s01, s32);
302  // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ]
303  // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ]
304
305  const __m128i tmp0   = _mm_madd_epi16(a01, k88p);  // [ (a0 + a1) << 3, ... ]
306  const __m128i tmp2   = _mm_madd_epi16(a01, k88m);  // [ (a0 - a1) << 3, ... ]
307  const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p);
308  const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m);
309  const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812);
310  const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937);
311  const __m128i tmp1   = _mm_srai_epi32(tmp1_2, 9);
312  const __m128i tmp3   = _mm_srai_epi32(tmp3_2, 9);
313  const __m128i s03    = _mm_packs_epi32(tmp0, tmp2);
314  const __m128i s12    = _mm_packs_epi32(tmp1, tmp3);
315  const __m128i s_lo   = _mm_unpacklo_epi16(s03, s12);   // 0 1 0 1 0 1...
316  const __m128i s_hi   = _mm_unpackhi_epi16(s03, s12);   // 2 3 2 3 2 3
317  const __m128i v23    = _mm_unpackhi_epi32(s_lo, s_hi);
318  *out01 = _mm_unpacklo_epi32(s_lo, s_hi);
319  *out32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));  // 3 2 3 2 3 2..
320}
321
322static void FTransformPass2(const __m128i* const v01, const __m128i* const v32,
323                            int16_t* out) {
324  const __m128i zero = _mm_setzero_si128();
325  const __m128i seven = _mm_set1_epi16(7);
326  const __m128i k5352_2217 = _mm_set_epi16(5352,  2217, 5352,  2217,
327                                           5352,  2217, 5352,  2217);
328  const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352,
329                                           2217, -5352, 2217, -5352);
330  const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
331  const __m128i k51000 = _mm_set1_epi32(51000);
332
333  // Same operations are done on the (0,3) and (1,2) pairs.
334  // a0 = v0 + v3
335  // a1 = v1 + v2
336  // a3 = v0 - v3
337  // a2 = v1 - v2
338  const __m128i a01 = _mm_add_epi16(*v01, *v32);
339  const __m128i a32 = _mm_sub_epi16(*v01, *v32);
340  const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
341  const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
342  const __m128i a01_plus_7 = _mm_add_epi16(a01, seven);
343
344  // d0 = (a0 + a1 + 7) >> 4;
345  // d2 = (a0 - a1 + 7) >> 4;
346  const __m128i c0 = _mm_add_epi16(a01_plus_7, a11);
347  const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11);
348  const __m128i d0 = _mm_srai_epi16(c0, 4);
349  const __m128i d2 = _mm_srai_epi16(c2, 4);
350
351  // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
352  // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
353  const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
354  const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
355  const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
356  const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one);
357  const __m128i d3 = _mm_add_epi32(c3, k51000);
358  const __m128i e1 = _mm_srai_epi32(d1, 16);
359  const __m128i e3 = _mm_srai_epi32(d3, 16);
360  const __m128i f1 = _mm_packs_epi32(e1, e1);
361  const __m128i f3 = _mm_packs_epi32(e3, e3);
362  // f1 = f1 + (a3 != 0);
363  // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
364  // desired (0, 1), we add one earlier through k12000_plus_one.
365  // -> f1 = f1 + 1 - (a3 == 0)
366  const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
367
368  const __m128i d0_g1 = _mm_unpacklo_epi64(d0, g1);
369  const __m128i d2_f3 = _mm_unpacklo_epi64(d2, f3);
370  _mm_storeu_si128((__m128i*)&out[0], d0_g1);
371  _mm_storeu_si128((__m128i*)&out[8], d2_f3);
372}
373
374static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
375  const __m128i zero = _mm_setzero_si128();
376
377  // Load src and convert to 16b.
378  const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]);
379  const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]);
380  const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]);
381  const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]);
382  const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
383  const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
384  const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
385  const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
386  // Load ref and convert to 16b.
387  const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
388  const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
389  const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
390  const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
391  const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
392  const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
393  const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
394  const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
395  // Compute difference. -> 00 01 02 03 00 00 00 00
396  const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
397  const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
398  const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
399  const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
400
401  // Unpack and shuffle
402  // 00 01 02 03   0 0 0 0
403  // 10 11 12 13   0 0 0 0
404  // 20 21 22 23   0 0 0 0
405  // 30 31 32 33   0 0 0 0
406  const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1);
407  const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3);
408  __m128i v01, v32;
409
410  // First pass
411  FTransformPass1(&shuf01, &shuf23, &v01, &v32);
412
413  // Second pass
414  FTransformPass2(&v01, &v32, out);
415}
416
417static void FTransform2(const uint8_t* src, const uint8_t* ref, int16_t* out) {
418  const __m128i zero = _mm_setzero_si128();
419
420  // Load src and convert to 16b.
421  const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]);
422  const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]);
423  const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]);
424  const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]);
425  const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
426  const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
427  const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
428  const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
429  // Load ref and convert to 16b.
430  const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]);
431  const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]);
432  const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]);
433  const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]);
434  const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
435  const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
436  const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
437  const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
438  // Compute difference. -> 00 01 02 03  00' 01' 02' 03'
439  const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
440  const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
441  const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
442  const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
443
444  // Unpack and shuffle
445  // 00 01 02 03   0 0 0 0
446  // 10 11 12 13   0 0 0 0
447  // 20 21 22 23   0 0 0 0
448  // 30 31 32 33   0 0 0 0
449  const __m128i shuf01l = _mm_unpacklo_epi32(diff0, diff1);
450  const __m128i shuf23l = _mm_unpacklo_epi32(diff2, diff3);
451  const __m128i shuf01h = _mm_unpackhi_epi32(diff0, diff1);
452  const __m128i shuf23h = _mm_unpackhi_epi32(diff2, diff3);
453  __m128i v01l, v32l;
454  __m128i v01h, v32h;
455
456  // First pass
457  FTransformPass1(&shuf01l, &shuf23l, &v01l, &v32l);
458  FTransformPass1(&shuf01h, &shuf23h, &v01h, &v32h);
459
460  // Second pass
461  FTransformPass2(&v01l, &v32l, out + 0);
462  FTransformPass2(&v01h, &v32h, out + 16);
463}
464
465static void FTransformWHTRow(const int16_t* const in, __m128i* const out) {
466  const __m128i kMult1 = _mm_set_epi16(0, 0, 0, 0, 1, 1, 1, 1);
467  const __m128i kMult2 = _mm_set_epi16(0, 0, 0, 0, -1, 1, -1, 1);
468  const __m128i src0 = _mm_loadl_epi64((__m128i*)&in[0 * 16]);
469  const __m128i src1 = _mm_loadl_epi64((__m128i*)&in[1 * 16]);
470  const __m128i src2 = _mm_loadl_epi64((__m128i*)&in[2 * 16]);
471  const __m128i src3 = _mm_loadl_epi64((__m128i*)&in[3 * 16]);
472  const __m128i A01 = _mm_unpacklo_epi16(src0, src1);  // A0 A1 | ...
473  const __m128i A23 = _mm_unpacklo_epi16(src2, src3);  // A2 A3 | ...
474  const __m128i B0 = _mm_adds_epi16(A01, A23);    // a0 | a1 | ...
475  const __m128i B1 = _mm_subs_epi16(A01, A23);    // a3 | a2 | ...
476  const __m128i C0 = _mm_unpacklo_epi32(B0, B1);  // a0 | a1 | a3 | a2
477  const __m128i C1 = _mm_unpacklo_epi32(B1, B0);  // a3 | a2 | a0 | a1
478  const __m128i D0 = _mm_madd_epi16(C0, kMult1);  // out0, out1
479  const __m128i D1 = _mm_madd_epi16(C1, kMult2);  // out2, out3
480  *out = _mm_unpacklo_epi64(D0, D1);
481}
482
483static void FTransformWHT(const int16_t* in, int16_t* out) {
484  __m128i row0, row1, row2, row3;
485  FTransformWHTRow(in + 0 * 64, &row0);
486  FTransformWHTRow(in + 1 * 64, &row1);
487  FTransformWHTRow(in + 2 * 64, &row2);
488  FTransformWHTRow(in + 3 * 64, &row3);
489
490  {
491    const __m128i a0 = _mm_add_epi32(row0, row2);
492    const __m128i a1 = _mm_add_epi32(row1, row3);
493    const __m128i a2 = _mm_sub_epi32(row1, row3);
494    const __m128i a3 = _mm_sub_epi32(row0, row2);
495    const __m128i b0 = _mm_srai_epi32(_mm_add_epi32(a0, a1), 1);
496    const __m128i b1 = _mm_srai_epi32(_mm_add_epi32(a3, a2), 1);
497    const __m128i b2 = _mm_srai_epi32(_mm_sub_epi32(a3, a2), 1);
498    const __m128i b3 = _mm_srai_epi32(_mm_sub_epi32(a0, a1), 1);
499    const __m128i out0 = _mm_packs_epi32(b0, b1);
500    const __m128i out1 = _mm_packs_epi32(b2, b3);
501    _mm_storeu_si128((__m128i*)&out[0], out0);
502    _mm_storeu_si128((__m128i*)&out[8], out1);
503  }
504}
505
506//------------------------------------------------------------------------------
507// Compute susceptibility based on DCT-coeff histograms:
508// the higher, the "easier" the macroblock is to compress.
509
510static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
511                             int start_block, int end_block,
512                             VP8Histogram* const histo) {
513  const __m128i zero = _mm_setzero_si128();
514  const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
515  int j;
516  int distribution[MAX_COEFF_THRESH + 1] = { 0 };
517  for (j = start_block; j < end_block; ++j) {
518    int16_t out[16];
519    int k;
520
521    FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
522
523    // Convert coefficients to bin (within out[]).
524    {
525      // Load.
526      const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
527      const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
528      const __m128i d0 = _mm_sub_epi16(zero, out0);
529      const __m128i d1 = _mm_sub_epi16(zero, out1);
530      const __m128i abs0 = _mm_max_epi16(out0, d0);   // abs(v), 16b
531      const __m128i abs1 = _mm_max_epi16(out1, d1);
532      // v = abs(out) >> 3
533      const __m128i v0 = _mm_srai_epi16(abs0, 3);
534      const __m128i v1 = _mm_srai_epi16(abs1, 3);
535      // bin = min(v, MAX_COEFF_THRESH)
536      const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
537      const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
538      // Store.
539      _mm_storeu_si128((__m128i*)&out[0], bin0);
540      _mm_storeu_si128((__m128i*)&out[8], bin1);
541    }
542
543    // Convert coefficients to bin.
544    for (k = 0; k < 16; ++k) {
545      ++distribution[out[k]];
546    }
547  }
548  VP8SetHistogramData(distribution, histo);
549}
550
551//------------------------------------------------------------------------------
552// Intra predictions
553
554// helper for chroma-DC predictions
555static WEBP_INLINE void Put8x8uv(uint8_t v, uint8_t* dst) {
556  int j;
557  const __m128i values = _mm_set1_epi8(v);
558  for (j = 0; j < 8; ++j) {
559    _mm_storel_epi64((__m128i*)(dst + j * BPS), values);
560  }
561}
562
563static WEBP_INLINE void Put16(uint8_t v, uint8_t* dst) {
564  int j;
565  const __m128i values = _mm_set1_epi8(v);
566  for (j = 0; j < 16; ++j) {
567    _mm_store_si128((__m128i*)(dst + j * BPS), values);
568  }
569}
570
571static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
572  if (size == 4) {
573    int j;
574    for (j = 0; j < 4; ++j) {
575      memset(dst + j * BPS, value, 4);
576    }
577  } else if (size == 8) {
578    Put8x8uv(value, dst);
579  } else {
580    Put16(value, dst);
581  }
582}
583
584static WEBP_INLINE void VE8uv(uint8_t* dst, const uint8_t* top) {
585  int j;
586  const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
587  for (j = 0; j < 8; ++j) {
588    _mm_storel_epi64((__m128i*)(dst + j * BPS), top_values);
589  }
590}
591
592static WEBP_INLINE void VE16(uint8_t* dst, const uint8_t* top) {
593  const __m128i top_values = _mm_load_si128((const __m128i*)top);
594  int j;
595  for (j = 0; j < 16; ++j) {
596    _mm_store_si128((__m128i*)(dst + j * BPS), top_values);
597  }
598}
599
600static WEBP_INLINE void VerticalPred(uint8_t* dst,
601                                     const uint8_t* top, int size) {
602  if (top != NULL) {
603    if (size == 8) {
604      VE8uv(dst, top);
605    } else {
606      VE16(dst, top);
607    }
608  } else {
609    Fill(dst, 127, size);
610  }
611}
612
613static WEBP_INLINE void HE8uv(uint8_t* dst, const uint8_t* left) {
614  int j;
615  for (j = 0; j < 8; ++j) {
616    const __m128i values = _mm_set1_epi8(left[j]);
617    _mm_storel_epi64((__m128i*)dst, values);
618    dst += BPS;
619  }
620}
621
622static WEBP_INLINE void HE16(uint8_t* dst, const uint8_t* left) {
623  int j;
624  for (j = 0; j < 16; ++j) {
625    const __m128i values = _mm_set1_epi8(left[j]);
626    _mm_store_si128((__m128i*)dst, values);
627    dst += BPS;
628  }
629}
630
631static WEBP_INLINE void HorizontalPred(uint8_t* dst,
632                                       const uint8_t* left, int size) {
633  if (left != NULL) {
634    if (size == 8) {
635      HE8uv(dst, left);
636    } else {
637      HE16(dst, left);
638    }
639  } else {
640    Fill(dst, 129, size);
641  }
642}
643
644static WEBP_INLINE void TM(uint8_t* dst, const uint8_t* left,
645                           const uint8_t* top, int size) {
646  const __m128i zero = _mm_setzero_si128();
647  int y;
648  if (size == 8) {
649    const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
650    const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
651    for (y = 0; y < 8; ++y, dst += BPS) {
652      const int val = left[y] - left[-1];
653      const __m128i base = _mm_set1_epi16(val);
654      const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
655      _mm_storel_epi64((__m128i*)dst, out);
656    }
657  } else {
658    const __m128i top_values = _mm_load_si128((const __m128i*)top);
659    const __m128i top_base_0 = _mm_unpacklo_epi8(top_values, zero);
660    const __m128i top_base_1 = _mm_unpackhi_epi8(top_values, zero);
661    for (y = 0; y < 16; ++y, dst += BPS) {
662      const int val = left[y] - left[-1];
663      const __m128i base = _mm_set1_epi16(val);
664      const __m128i out_0 = _mm_add_epi16(base, top_base_0);
665      const __m128i out_1 = _mm_add_epi16(base, top_base_1);
666      const __m128i out = _mm_packus_epi16(out_0, out_1);
667      _mm_store_si128((__m128i*)dst, out);
668    }
669  }
670}
671
672static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
673                                   const uint8_t* top, int size) {
674  if (left != NULL) {
675    if (top != NULL) {
676      TM(dst, left, top, size);
677    } else {
678      HorizontalPred(dst, left, size);
679    }
680  } else {
681    // true motion without left samples (hence: with default 129 value)
682    // is equivalent to VE prediction where you just copy the top samples.
683    // Note that if top samples are not available, the default value is
684    // then 129, and not 127 as in the VerticalPred case.
685    if (top != NULL) {
686      VerticalPred(dst, top, size);
687    } else {
688      Fill(dst, 129, size);
689    }
690  }
691}
692
693static WEBP_INLINE void DC8uv(uint8_t* dst, const uint8_t* left,
694                              const uint8_t* top) {
695  const __m128i zero = _mm_setzero_si128();
696  const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
697  const __m128i left_values = _mm_loadl_epi64((const __m128i*)left);
698  const __m128i sum_top = _mm_sad_epu8(top_values, zero);
699  const __m128i sum_left = _mm_sad_epu8(left_values, zero);
700  const int DC = _mm_cvtsi128_si32(sum_top) + _mm_cvtsi128_si32(sum_left) + 8;
701  Put8x8uv(DC >> 4, dst);
702}
703
704static WEBP_INLINE void DC8uvNoLeft(uint8_t* dst, const uint8_t* top) {
705  const __m128i zero = _mm_setzero_si128();
706  const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
707  const __m128i sum = _mm_sad_epu8(top_values, zero);
708  const int DC = _mm_cvtsi128_si32(sum) + 4;
709  Put8x8uv(DC >> 3, dst);
710}
711
712static WEBP_INLINE void DC8uvNoTop(uint8_t* dst, const uint8_t* left) {
713  // 'left' is contiguous so we can reuse the top summation.
714  DC8uvNoLeft(dst, left);
715}
716
717static WEBP_INLINE void DC8uvNoTopLeft(uint8_t* dst) {
718  Put8x8uv(0x80, dst);
719}
720
721static WEBP_INLINE void DC8uvMode(uint8_t* dst, const uint8_t* left,
722                                  const uint8_t* top) {
723  if (top != NULL) {
724    if (left != NULL) {  // top and left present
725      DC8uv(dst, left, top);
726    } else {  // top, but no left
727      DC8uvNoLeft(dst, top);
728    }
729  } else if (left != NULL) {  // left but no top
730    DC8uvNoTop(dst, left);
731  } else {  // no top, no left, nothing.
732    DC8uvNoTopLeft(dst);
733  }
734}
735
736static WEBP_INLINE void DC16(uint8_t* dst, const uint8_t* left,
737                             const uint8_t* top) {
738  const __m128i zero = _mm_setzero_si128();
739  const __m128i top_row = _mm_load_si128((const __m128i*)top);
740  const __m128i left_row = _mm_load_si128((const __m128i*)left);
741  const __m128i sad8x2 = _mm_sad_epu8(top_row, zero);
742  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
743  const __m128i sum_top = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
744  const __m128i sad8x2_left = _mm_sad_epu8(left_row, zero);
745  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
746  const __m128i sum_left =
747      _mm_add_epi16(sad8x2_left, _mm_shuffle_epi32(sad8x2_left, 2));
748  const int DC = _mm_cvtsi128_si32(sum_top) + _mm_cvtsi128_si32(sum_left) + 16;
749  Put16(DC >> 5, dst);
750}
751
752static WEBP_INLINE void DC16NoLeft(uint8_t* dst, const uint8_t* top) {
753  const __m128i zero = _mm_setzero_si128();
754  const __m128i top_row = _mm_load_si128((const __m128i*)top);
755  const __m128i sad8x2 = _mm_sad_epu8(top_row, zero);
756  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
757  const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
758  const int DC = _mm_cvtsi128_si32(sum) + 8;
759  Put16(DC >> 4, dst);
760}
761
762static WEBP_INLINE void DC16NoTop(uint8_t* dst, const uint8_t* left) {
763  // 'left' is contiguous so we can reuse the top summation.
764  DC16NoLeft(dst, left);
765}
766
767static WEBP_INLINE void DC16NoTopLeft(uint8_t* dst) {
768  Put16(0x80, dst);
769}
770
771static WEBP_INLINE void DC16Mode(uint8_t* dst, const uint8_t* left,
772                                 const uint8_t* top) {
773  if (top != NULL) {
774    if (left != NULL) {  // top and left present
775      DC16(dst, left, top);
776    } else {  // top, but no left
777      DC16NoLeft(dst, top);
778    }
779  } else if (left != NULL) {  // left but no top
780    DC16NoTop(dst, left);
781  } else {  // no top, no left, nothing.
782    DC16NoTopLeft(dst);
783  }
784}
785
786//------------------------------------------------------------------------------
787// 4x4 predictions
788
789#define DST(x, y) dst[(x) + (y) * BPS]
790#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
791#define AVG2(a, b) (((a) + (b) + 1) >> 1)
792
793// We use the following 8b-arithmetic tricks:
794//     (a + 2 * b + c + 2) >> 2 = (AC + b + 1) >> 1
795//   where: AC = (a + c) >> 1 = [(a + c + 1) >> 1] - [(a^c) & 1]
796// and:
797//     (a + 2 * b + c + 2) >> 2 = (AB + BC + 1) >> 1 - (ab|bc)&lsb
798//   where: AC = (a + b + 1) >> 1,   BC = (b + c + 1) >> 1
799//   and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1
800
801static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) {  // vertical
802  const __m128i one = _mm_set1_epi8(1);
803  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(top - 1));
804  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
805  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
806  const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00);
807  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one);
808  const __m128i b = _mm_subs_epu8(a, lsb);
809  const __m128i avg = _mm_avg_epu8(b, BCDEFGH0);
810  const uint32_t vals = _mm_cvtsi128_si32(avg);
811  int i;
812  for (i = 0; i < 4; ++i) {
813    WebPUint32ToMem(dst + i * BPS, vals);
814  }
815}
816
817static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) {  // horizontal
818  const int X = top[-1];
819  const int I = top[-2];
820  const int J = top[-3];
821  const int K = top[-4];
822  const int L = top[-5];
823  WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));
824  WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));
825  WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));
826  WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
827}
828
829static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) {
830  uint32_t dc = 4;
831  int i;
832  for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
833  Fill(dst, dc >> 3, 4);
834}
835
836static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) {  // Down-Left
837  const __m128i one = _mm_set1_epi8(1);
838  const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top);
839  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
840  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
841  const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, top[7], 3);
842  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0);
843  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);
844  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
845  const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0);
846  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
847  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
848  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
849  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
850}
851
852static WEBP_INLINE void VR4(uint8_t* dst,
853                            const uint8_t* top) {  // Vertical-Right
854  const __m128i one = _mm_set1_epi8(1);
855  const int I = top[-2];
856  const int J = top[-3];
857  const int K = top[-4];
858  const int X = top[-1];
859  const __m128i XABCD = _mm_loadl_epi64((const __m128i*)(top - 1));
860  const __m128i ABCD0 = _mm_srli_si128(XABCD, 1);
861  const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0);
862  const __m128i _XABCD = _mm_slli_si128(XABCD, 1);
863  const __m128i IXABCD = _mm_insert_epi16(_XABCD, I | (X << 8), 0);
864  const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0);
865  const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one);
866  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
867  const __m128i efgh = _mm_avg_epu8(avg2, XABCD);
868  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));
869  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));
870  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));
871  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));
872
873  // these two are hard to implement in SSE2, so we keep the C-version:
874  DST(0, 2) = AVG3(J, I, X);
875  DST(0, 3) = AVG3(K, J, I);
876}
877
878static WEBP_INLINE void VL4(uint8_t* dst,
879                            const uint8_t* top) {  // Vertical-Left
880  const __m128i one = _mm_set1_epi8(1);
881  const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top);
882  const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1);
883  const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2);
884  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_);
885  const __m128i avg2 = _mm_avg_epu8(CDEFGH__, BCDEFGH_);
886  const __m128i avg3 = _mm_avg_epu8(avg1, avg2);
887  const __m128i lsb1 = _mm_and_si128(_mm_xor_si128(avg1, avg2), one);
888  const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_);
889  const __m128i bc = _mm_xor_si128(CDEFGH__, BCDEFGH_);
890  const __m128i abbc = _mm_or_si128(ab, bc);
891  const __m128i lsb2 = _mm_and_si128(abbc, lsb1);
892  const __m128i avg4 = _mm_subs_epu8(avg3, lsb2);
893  const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
894  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));
895  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));
896  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));
897  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));
898
899  // these two are hard to get and irregular
900  DST(3, 2) = (extra_out >> 0) & 0xff;
901  DST(3, 3) = (extra_out >> 8) & 0xff;
902}
903
904static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) {  // Down-right
905  const __m128i one = _mm_set1_epi8(1);
906  const __m128i LKJIXABC = _mm_loadl_epi64((const __m128i*)(top - 5));
907  const __m128i LKJIXABCD = _mm_insert_epi16(LKJIXABC, top[3], 4);
908  const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1);
909  const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2);
910  const __m128i avg1 = _mm_avg_epu8(JIXABCD__, LKJIXABCD);
911  const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one);
912  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
913  const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_);
914  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
915  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
916  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
917  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
918}
919
920static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) {
921  const int I = top[-2];
922  const int J = top[-3];
923  const int K = top[-4];
924  const int L = top[-5];
925  DST(0, 0) =             AVG2(I, J);
926  DST(2, 0) = DST(0, 1) = AVG2(J, K);
927  DST(2, 1) = DST(0, 2) = AVG2(K, L);
928  DST(1, 0) =             AVG3(I, J, K);
929  DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
930  DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
931  DST(3, 2) = DST(2, 2) =
932  DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
933}
934
935static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) {
936  const int X = top[-1];
937  const int I = top[-2];
938  const int J = top[-3];
939  const int K = top[-4];
940  const int L = top[-5];
941  const int A = top[0];
942  const int B = top[1];
943  const int C = top[2];
944
945  DST(0, 0) = DST(2, 1) = AVG2(I, X);
946  DST(0, 1) = DST(2, 2) = AVG2(J, I);
947  DST(0, 2) = DST(2, 3) = AVG2(K, J);
948  DST(0, 3)             = AVG2(L, K);
949
950  DST(3, 0)             = AVG3(A, B, C);
951  DST(2, 0)             = AVG3(X, A, B);
952  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
953  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
954  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
955  DST(1, 3)             = AVG3(L, K, J);
956}
957
958static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) {
959  const __m128i zero = _mm_setzero_si128();
960  const __m128i top_values = _mm_cvtsi32_si128(WebPMemToUint32(top));
961  const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
962  int y;
963  for (y = 0; y < 4; ++y, dst += BPS) {
964    const int val = top[-2 - y] - top[-1];
965    const __m128i base = _mm_set1_epi16(val);
966    const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
967    WebPUint32ToMem(dst, _mm_cvtsi128_si32(out));
968  }
969}
970
971#undef DST
972#undef AVG3
973#undef AVG2
974
975//------------------------------------------------------------------------------
976// luma 4x4 prediction
977
978// Left samples are top[-5 .. -2], top_left is top[-1], top are
979// located at top[0..3], and top right is top[4..7]
980static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
981  DC4(I4DC4 + dst, top);
982  TM4(I4TM4 + dst, top);
983  VE4(I4VE4 + dst, top);
984  HE4(I4HE4 + dst, top);
985  RD4(I4RD4 + dst, top);
986  VR4(I4VR4 + dst, top);
987  LD4(I4LD4 + dst, top);
988  VL4(I4VL4 + dst, top);
989  HD4(I4HD4 + dst, top);
990  HU4(I4HU4 + dst, top);
991}
992
993//------------------------------------------------------------------------------
994// Chroma 8x8 prediction (paragraph 12.2)
995
996static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
997                             const uint8_t* top) {
998  // U block
999  DC8uvMode(C8DC8 + dst, left, top);
1000  VerticalPred(C8VE8 + dst, top, 8);
1001  HorizontalPred(C8HE8 + dst, left, 8);
1002  TrueMotion(C8TM8 + dst, left, top, 8);
1003  // V block
1004  dst += 8;
1005  if (top != NULL) top += 8;
1006  if (left != NULL) left += 16;
1007  DC8uvMode(C8DC8 + dst, left, top);
1008  VerticalPred(C8VE8 + dst, top, 8);
1009  HorizontalPred(C8HE8 + dst, left, 8);
1010  TrueMotion(C8TM8 + dst, left, top, 8);
1011}
1012
1013//------------------------------------------------------------------------------
1014// luma 16x16 prediction (paragraph 12.3)
1015
1016static void Intra16Preds(uint8_t* dst,
1017                         const uint8_t* left, const uint8_t* top) {
1018  DC16Mode(I16DC16 + dst, left, top);
1019  VerticalPred(I16VE16 + dst, top, 16);
1020  HorizontalPred(I16HE16 + dst, left, 16);
1021  TrueMotion(I16TM16 + dst, left, top, 16);
1022}
1023
1024//------------------------------------------------------------------------------
1025// Metric
1026
1027static WEBP_INLINE void SubtractAndAccumulate(const __m128i a, const __m128i b,
1028                                              __m128i* const sum) {
1029  // take abs(a-b) in 8b
1030  const __m128i a_b = _mm_subs_epu8(a, b);
1031  const __m128i b_a = _mm_subs_epu8(b, a);
1032  const __m128i abs_a_b = _mm_or_si128(a_b, b_a);
1033  // zero-extend to 16b
1034  const __m128i zero = _mm_setzero_si128();
1035  const __m128i C0 = _mm_unpacklo_epi8(abs_a_b, zero);
1036  const __m128i C1 = _mm_unpackhi_epi8(abs_a_b, zero);
1037  // multiply with self
1038  const __m128i sum1 = _mm_madd_epi16(C0, C0);
1039  const __m128i sum2 = _mm_madd_epi16(C1, C1);
1040  *sum = _mm_add_epi32(sum1, sum2);
1041}
1042
1043static WEBP_INLINE int SSE_16xN(const uint8_t* a, const uint8_t* b,
1044                                int num_pairs) {
1045  __m128i sum = _mm_setzero_si128();
1046  int32_t tmp[4];
1047  int i;
1048
1049  for (i = 0; i < num_pairs; ++i) {
1050    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[BPS * 0]);
1051    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[BPS * 0]);
1052    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[BPS * 1]);
1053    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[BPS * 1]);
1054    __m128i sum1, sum2;
1055    SubtractAndAccumulate(a0, b0, &sum1);
1056    SubtractAndAccumulate(a1, b1, &sum2);
1057    sum = _mm_add_epi32(sum, _mm_add_epi32(sum1, sum2));
1058    a += 2 * BPS;
1059    b += 2 * BPS;
1060  }
1061  _mm_storeu_si128((__m128i*)tmp, sum);
1062  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
1063}
1064
1065static int SSE16x16(const uint8_t* a, const uint8_t* b) {
1066  return SSE_16xN(a, b, 8);
1067}
1068
1069static int SSE16x8(const uint8_t* a, const uint8_t* b) {
1070  return SSE_16xN(a, b, 4);
1071}
1072
1073#define LOAD_8x16b(ptr) \
1074  _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(ptr)), zero)
1075
1076static int SSE8x8(const uint8_t* a, const uint8_t* b) {
1077  const __m128i zero = _mm_setzero_si128();
1078  int num_pairs = 4;
1079  __m128i sum = zero;
1080  int32_t tmp[4];
1081  while (num_pairs-- > 0) {
1082    const __m128i a0 = LOAD_8x16b(&a[BPS * 0]);
1083    const __m128i a1 = LOAD_8x16b(&a[BPS * 1]);
1084    const __m128i b0 = LOAD_8x16b(&b[BPS * 0]);
1085    const __m128i b1 = LOAD_8x16b(&b[BPS * 1]);
1086    // subtract
1087    const __m128i c0 = _mm_subs_epi16(a0, b0);
1088    const __m128i c1 = _mm_subs_epi16(a1, b1);
1089    // multiply/accumulate with self
1090    const __m128i d0 = _mm_madd_epi16(c0, c0);
1091    const __m128i d1 = _mm_madd_epi16(c1, c1);
1092    // collect
1093    const __m128i sum01 = _mm_add_epi32(d0, d1);
1094    sum = _mm_add_epi32(sum, sum01);
1095    a += 2 * BPS;
1096    b += 2 * BPS;
1097  }
1098  _mm_storeu_si128((__m128i*)tmp, sum);
1099  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
1100}
1101#undef LOAD_8x16b
1102
1103static int SSE4x4(const uint8_t* a, const uint8_t* b) {
1104  const __m128i zero = _mm_setzero_si128();
1105
1106  // Load values. Note that we read 8 pixels instead of 4,
1107  // but the a/b buffers are over-allocated to that effect.
1108  const __m128i a0 = _mm_loadl_epi64((const __m128i*)&a[BPS * 0]);
1109  const __m128i a1 = _mm_loadl_epi64((const __m128i*)&a[BPS * 1]);
1110  const __m128i a2 = _mm_loadl_epi64((const __m128i*)&a[BPS * 2]);
1111  const __m128i a3 = _mm_loadl_epi64((const __m128i*)&a[BPS * 3]);
1112  const __m128i b0 = _mm_loadl_epi64((const __m128i*)&b[BPS * 0]);
1113  const __m128i b1 = _mm_loadl_epi64((const __m128i*)&b[BPS * 1]);
1114  const __m128i b2 = _mm_loadl_epi64((const __m128i*)&b[BPS * 2]);
1115  const __m128i b3 = _mm_loadl_epi64((const __m128i*)&b[BPS * 3]);
1116  // Combine pair of lines.
1117  const __m128i a01 = _mm_unpacklo_epi32(a0, a1);
1118  const __m128i a23 = _mm_unpacklo_epi32(a2, a3);
1119  const __m128i b01 = _mm_unpacklo_epi32(b0, b1);
1120  const __m128i b23 = _mm_unpacklo_epi32(b2, b3);
1121  // Convert to 16b.
1122  const __m128i a01s = _mm_unpacklo_epi8(a01, zero);
1123  const __m128i a23s = _mm_unpacklo_epi8(a23, zero);
1124  const __m128i b01s = _mm_unpacklo_epi8(b01, zero);
1125  const __m128i b23s = _mm_unpacklo_epi8(b23, zero);
1126  // subtract, square and accumulate
1127  const __m128i d0 = _mm_subs_epi16(a01s, b01s);
1128  const __m128i d1 = _mm_subs_epi16(a23s, b23s);
1129  const __m128i e0 = _mm_madd_epi16(d0, d0);
1130  const __m128i e1 = _mm_madd_epi16(d1, d1);
1131  const __m128i sum = _mm_add_epi32(e0, e1);
1132
1133  int32_t tmp[4];
1134  _mm_storeu_si128((__m128i*)tmp, sum);
1135  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
1136}
1137
1138//------------------------------------------------------------------------------
1139// Texture distortion
1140//
1141// We try to match the spectral content (weighted) between source and
1142// reconstructed samples.
1143
1144// Hadamard transform
1145// Returns the difference between the weighted sum of the absolute value of
1146// transformed coefficients.
1147static int TTransform(const uint8_t* inA, const uint8_t* inB,
1148                      const uint16_t* const w) {
1149  int32_t sum[4];
1150  __m128i tmp_0, tmp_1, tmp_2, tmp_3;
1151  const __m128i zero = _mm_setzero_si128();
1152
1153  // Load, combine and transpose inputs.
1154  {
1155    const __m128i inA_0 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 0]);
1156    const __m128i inA_1 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 1]);
1157    const __m128i inA_2 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 2]);
1158    const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]);
1159    const __m128i inB_0 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 0]);
1160    const __m128i inB_1 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 1]);
1161    const __m128i inB_2 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 2]);
1162    const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]);
1163
1164    // Combine inA and inB (we'll do two transforms in parallel).
1165    const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0);
1166    const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1);
1167    const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2);
1168    const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3);
1169    // a00 b00 a01 b01 a02 b03 a03 b03   0 0 0 0 0 0 0 0
1170    // a10 b10 a11 b11 a12 b12 a13 b13   0 0 0 0 0 0 0 0
1171    // a20 b20 a21 b21 a22 b22 a23 b23   0 0 0 0 0 0 0 0
1172    // a30 b30 a31 b31 a32 b32 a33 b33   0 0 0 0 0 0 0 0
1173
1174    // Transpose the two 4x4, discarding the filling zeroes.
1175    const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2);
1176    const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3);
1177    // a00 a20  b00 b20  a01 a21  b01 b21  a02 a22  b02 b22  a03 a23  b03 b23
1178    // a10 a30  b10 b30  a11 a31  b11 b31  a12 a32  b12 b32  a13 a33  b13 b33
1179    const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1);
1180    const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1);
1181    // a00 a10 a20 a30  b00 b10 b20 b30  a01 a11 a21 a31  b01 b11 b21 b31
1182    // a02 a12 a22 a32  b02 b12 b22 b32  a03 a13 a23 a33  b03 b13 b23 b33
1183
1184    // Convert to 16b.
1185    tmp_0 = _mm_unpacklo_epi8(transpose1_0, zero);
1186    tmp_1 = _mm_unpackhi_epi8(transpose1_0, zero);
1187    tmp_2 = _mm_unpacklo_epi8(transpose1_1, zero);
1188    tmp_3 = _mm_unpackhi_epi8(transpose1_1, zero);
1189    // a00 a10 a20 a30   b00 b10 b20 b30
1190    // a01 a11 a21 a31   b01 b11 b21 b31
1191    // a02 a12 a22 a32   b02 b12 b22 b32
1192    // a03 a13 a23 a33   b03 b13 b23 b33
1193  }
1194
1195  // Horizontal pass and subsequent transpose.
1196  {
1197    // Calculate a and b (two 4x4 at once).
1198    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
1199    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
1200    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
1201    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
1202    const __m128i b0 = _mm_add_epi16(a0, a1);
1203    const __m128i b1 = _mm_add_epi16(a3, a2);
1204    const __m128i b2 = _mm_sub_epi16(a3, a2);
1205    const __m128i b3 = _mm_sub_epi16(a0, a1);
1206    // a00 a01 a02 a03   b00 b01 b02 b03
1207    // a10 a11 a12 a13   b10 b11 b12 b13
1208    // a20 a21 a22 a23   b20 b21 b22 b23
1209    // a30 a31 a32 a33   b30 b31 b32 b33
1210
1211    // Transpose the two 4x4.
1212    const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1);
1213    const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3);
1214    const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1);
1215    const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3);
1216    // a00 a10 a01 a11   a02 a12 a03 a13
1217    // a20 a30 a21 a31   a22 a32 a23 a33
1218    // b00 b10 b01 b11   b02 b12 b03 b13
1219    // b20 b30 b21 b31   b22 b32 b23 b33
1220    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
1221    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
1222    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
1223    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
1224    // a00 a10 a20 a30 a01 a11 a21 a31
1225    // b00 b10 b20 b30 b01 b11 b21 b31
1226    // a02 a12 a22 a32 a03 a13 a23 a33
1227    // b02 b12 a22 b32 b03 b13 b23 b33
1228    tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
1229    tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
1230    tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
1231    tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
1232    // a00 a10 a20 a30   b00 b10 b20 b30
1233    // a01 a11 a21 a31   b01 b11 b21 b31
1234    // a02 a12 a22 a32   b02 b12 b22 b32
1235    // a03 a13 a23 a33   b03 b13 b23 b33
1236  }
1237
1238  // Vertical pass and difference of weighted sums.
1239  {
1240    // Load all inputs.
1241    const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]);
1242    const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]);
1243
1244    // Calculate a and b (two 4x4 at once).
1245    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
1246    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
1247    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
1248    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
1249    const __m128i b0 = _mm_add_epi16(a0, a1);
1250    const __m128i b1 = _mm_add_epi16(a3, a2);
1251    const __m128i b2 = _mm_sub_epi16(a3, a2);
1252    const __m128i b3 = _mm_sub_epi16(a0, a1);
1253
1254    // Separate the transforms of inA and inB.
1255    __m128i A_b0 = _mm_unpacklo_epi64(b0, b1);
1256    __m128i A_b2 = _mm_unpacklo_epi64(b2, b3);
1257    __m128i B_b0 = _mm_unpackhi_epi64(b0, b1);
1258    __m128i B_b2 = _mm_unpackhi_epi64(b2, b3);
1259
1260    {
1261      const __m128i d0 = _mm_sub_epi16(zero, A_b0);
1262      const __m128i d1 = _mm_sub_epi16(zero, A_b2);
1263      const __m128i d2 = _mm_sub_epi16(zero, B_b0);
1264      const __m128i d3 = _mm_sub_epi16(zero, B_b2);
1265      A_b0 = _mm_max_epi16(A_b0, d0);   // abs(v), 16b
1266      A_b2 = _mm_max_epi16(A_b2, d1);
1267      B_b0 = _mm_max_epi16(B_b0, d2);
1268      B_b2 = _mm_max_epi16(B_b2, d3);
1269    }
1270
1271    // weighted sums
1272    A_b0 = _mm_madd_epi16(A_b0, w_0);
1273    A_b2 = _mm_madd_epi16(A_b2, w_8);
1274    B_b0 = _mm_madd_epi16(B_b0, w_0);
1275    B_b2 = _mm_madd_epi16(B_b2, w_8);
1276    A_b0 = _mm_add_epi32(A_b0, A_b2);
1277    B_b0 = _mm_add_epi32(B_b0, B_b2);
1278
1279    // difference of weighted sums
1280    A_b0 = _mm_sub_epi32(A_b0, B_b0);
1281    _mm_storeu_si128((__m128i*)&sum[0], A_b0);
1282  }
1283  return sum[0] + sum[1] + sum[2] + sum[3];
1284}
1285
1286static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
1287                    const uint16_t* const w) {
1288  const int diff_sum = TTransform(a, b, w);
1289  return abs(diff_sum) >> 5;
1290}
1291
1292static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
1293                      const uint16_t* const w) {
1294  int D = 0;
1295  int x, y;
1296  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
1297    for (x = 0; x < 16; x += 4) {
1298      D += Disto4x4(a + x + y, b + x + y, w);
1299    }
1300  }
1301  return D;
1302}
1303
1304//------------------------------------------------------------------------------
1305// Quantization
1306//
1307
1308static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16],
1309                                       const uint16_t* const sharpen,
1310                                       const VP8Matrix* const mtx) {
1311  const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
1312  const __m128i zero = _mm_setzero_si128();
1313  __m128i coeff0, coeff8;
1314  __m128i out0, out8;
1315  __m128i packed_out;
1316
1317  // Load all inputs.
1318  __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
1319  __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
1320  const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
1321  const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
1322  const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
1323  const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
1324
1325  // extract sign(in)  (0x0000 if positive, 0xffff if negative)
1326  const __m128i sign0 = _mm_cmpgt_epi16(zero, in0);
1327  const __m128i sign8 = _mm_cmpgt_epi16(zero, in8);
1328
1329  // coeff = abs(in) = (in ^ sign) - sign
1330  coeff0 = _mm_xor_si128(in0, sign0);
1331  coeff8 = _mm_xor_si128(in8, sign8);
1332  coeff0 = _mm_sub_epi16(coeff0, sign0);
1333  coeff8 = _mm_sub_epi16(coeff8, sign8);
1334
1335  // coeff = abs(in) + sharpen
1336  if (sharpen != NULL) {
1337    const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
1338    const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
1339    coeff0 = _mm_add_epi16(coeff0, sharpen0);
1340    coeff8 = _mm_add_epi16(coeff8, sharpen8);
1341  }
1342
1343  // out = (coeff * iQ + B) >> QFIX
1344  {
1345    // doing calculations with 32b precision (QFIX=17)
1346    // out = (coeff * iQ)
1347    const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
1348    const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
1349    const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
1350    const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
1351    __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
1352    __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
1353    __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
1354    __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
1355    // out = (coeff * iQ + B)
1356    const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
1357    const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
1358    const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
1359    const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
1360    out_00 = _mm_add_epi32(out_00, bias_00);
1361    out_04 = _mm_add_epi32(out_04, bias_04);
1362    out_08 = _mm_add_epi32(out_08, bias_08);
1363    out_12 = _mm_add_epi32(out_12, bias_12);
1364    // out = QUANTDIV(coeff, iQ, B, QFIX)
1365    out_00 = _mm_srai_epi32(out_00, QFIX);
1366    out_04 = _mm_srai_epi32(out_04, QFIX);
1367    out_08 = _mm_srai_epi32(out_08, QFIX);
1368    out_12 = _mm_srai_epi32(out_12, QFIX);
1369
1370    // pack result as 16b
1371    out0 = _mm_packs_epi32(out_00, out_04);
1372    out8 = _mm_packs_epi32(out_08, out_12);
1373
1374    // if (coeff > 2047) coeff = 2047
1375    out0 = _mm_min_epi16(out0, max_coeff_2047);
1376    out8 = _mm_min_epi16(out8, max_coeff_2047);
1377  }
1378
1379  // get sign back (if (sign[j]) out_n = -out_n)
1380  out0 = _mm_xor_si128(out0, sign0);
1381  out8 = _mm_xor_si128(out8, sign8);
1382  out0 = _mm_sub_epi16(out0, sign0);
1383  out8 = _mm_sub_epi16(out8, sign8);
1384
1385  // in = out * Q
1386  in0 = _mm_mullo_epi16(out0, q0);
1387  in8 = _mm_mullo_epi16(out8, q8);
1388
1389  _mm_storeu_si128((__m128i*)&in[0], in0);
1390  _mm_storeu_si128((__m128i*)&in[8], in8);
1391
1392  // zigzag the output before storing it.
1393  //
1394  // The zigzag pattern can almost be reproduced with a small sequence of
1395  // shuffles. After it, we only need to swap the 7th (ending up in third
1396  // position instead of twelfth) and 8th values.
1397  {
1398    __m128i outZ0, outZ8;
1399    outZ0 = _mm_shufflehi_epi16(out0,  _MM_SHUFFLE(2, 1, 3, 0));
1400    outZ0 = _mm_shuffle_epi32  (outZ0, _MM_SHUFFLE(3, 1, 2, 0));
1401    outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2));
1402    outZ8 = _mm_shufflelo_epi16(out8,  _MM_SHUFFLE(3, 0, 2, 1));
1403    outZ8 = _mm_shuffle_epi32  (outZ8, _MM_SHUFFLE(3, 1, 2, 0));
1404    outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0));
1405    _mm_storeu_si128((__m128i*)&out[0], outZ0);
1406    _mm_storeu_si128((__m128i*)&out[8], outZ8);
1407    packed_out = _mm_packs_epi16(outZ0, outZ8);
1408  }
1409  {
1410    const int16_t outZ_12 = out[12];
1411    const int16_t outZ_3 = out[3];
1412    out[3] = outZ_12;
1413    out[12] = outZ_3;
1414  }
1415
1416  // detect if all 'out' values are zeroes or not
1417  return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
1418}
1419
1420static int QuantizeBlock(int16_t in[16], int16_t out[16],
1421                         const VP8Matrix* const mtx) {
1422  return DoQuantizeBlock(in, out, &mtx->sharpen_[0], mtx);
1423}
1424
1425static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
1426                            const VP8Matrix* const mtx) {
1427  return DoQuantizeBlock(in, out, NULL, mtx);
1428}
1429
1430static int Quantize2Blocks(int16_t in[32], int16_t out[32],
1431                           const VP8Matrix* const mtx) {
1432  int nz;
1433  const uint16_t* const sharpen = &mtx->sharpen_[0];
1434  nz  = DoQuantizeBlock(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0;
1435  nz |= DoQuantizeBlock(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1;
1436  return nz;
1437}
1438
1439//------------------------------------------------------------------------------
1440// Entry point
1441
1442extern void VP8EncDspInitSSE2(void);
1443
1444WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE2(void) {
1445  VP8CollectHistogram = CollectHistogram;
1446  VP8EncPredLuma16 = Intra16Preds;
1447  VP8EncPredChroma8 = IntraChromaPreds;
1448  VP8EncPredLuma4 = Intra4Preds;
1449  VP8EncQuantizeBlock = QuantizeBlock;
1450  VP8EncQuantize2Blocks = Quantize2Blocks;
1451  VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
1452  VP8ITransform = ITransform;
1453  VP8FTransform = FTransform;
1454  VP8FTransform2 = FTransform2;
1455  VP8FTransformWHT = FTransformWHT;
1456  VP8SSE16x16 = SSE16x16;
1457  VP8SSE16x8 = SSE16x8;
1458  VP8SSE8x8 = SSE8x8;
1459  VP8SSE4x4 = SSE4x4;
1460  VP8TDisto4x4 = Disto4x4;
1461  VP8TDisto16x16 = Disto16x16;
1462}
1463
1464#else  // !WEBP_USE_SSE2
1465
1466WEBP_DSP_INIT_STUB(VP8EncDspInitSSE2)
1467
1468#endif  // WEBP_USE_SSE2
1469