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(__cplusplus) || defined(c_plusplus)
17extern "C" {
18#endif
19
20#if defined(WEBP_USE_SSE2)
21#include <stdlib.h>  // for abs()
22#include <emmintrin.h>
23
24#include "../enc/vp8enci.h"
25
26//------------------------------------------------------------------------------
27// Quite useful macro for debugging. Left here for convenience.
28
29#if 0
30#include <stdio.h>
31static void PrintReg(const __m128i r, const char* const name, int size) {
32  int n;
33  union {
34    __m128i r;
35    uint8_t i8[16];
36    uint16_t i16[8];
37    uint32_t i32[4];
38    uint64_t i64[2];
39  } tmp;
40  tmp.r = r;
41  printf("%s\t: ", name);
42  if (size == 8) {
43    for (n = 0; n < 16; ++n) printf("%.2x ", tmp.i8[n]);
44  } else if (size == 16) {
45    for (n = 0; n < 8; ++n) printf("%.4x ", tmp.i16[n]);
46  } else if (size == 32) {
47    for (n = 0; n < 4; ++n) printf("%.8x ", tmp.i32[n]);
48  } else {
49    for (n = 0; n < 2; ++n) printf("%.16lx ", tmp.i64[n]);
50  }
51  printf("\n");
52}
53#endif
54
55//------------------------------------------------------------------------------
56// Compute susceptibility based on DCT-coeff histograms:
57// the higher, the "easier" the macroblock is to compress.
58
59static void CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
60                                 int start_block, int end_block,
61                                 VP8Histogram* const histo) {
62  const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
63  int j;
64  for (j = start_block; j < end_block; ++j) {
65    int16_t out[16];
66    int k;
67
68    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
69
70    // Convert coefficients to bin (within out[]).
71    {
72      // Load.
73      const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
74      const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
75      // sign(out) = out >> 15  (0x0000 if positive, 0xffff if negative)
76      const __m128i sign0 = _mm_srai_epi16(out0, 15);
77      const __m128i sign1 = _mm_srai_epi16(out1, 15);
78      // abs(out) = (out ^ sign) - sign
79      const __m128i xor0 = _mm_xor_si128(out0, sign0);
80      const __m128i xor1 = _mm_xor_si128(out1, sign1);
81      const __m128i abs0 = _mm_sub_epi16(xor0, sign0);
82      const __m128i abs1 = _mm_sub_epi16(xor1, sign1);
83      // v = abs(out) >> 3
84      const __m128i v0 = _mm_srai_epi16(abs0, 3);
85      const __m128i v1 = _mm_srai_epi16(abs1, 3);
86      // bin = min(v, MAX_COEFF_THRESH)
87      const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
88      const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
89      // Store.
90      _mm_storeu_si128((__m128i*)&out[0], bin0);
91      _mm_storeu_si128((__m128i*)&out[8], bin1);
92    }
93
94    // Convert coefficients to bin.
95    for (k = 0; k < 16; ++k) {
96      histo->distribution[out[k]]++;
97    }
98  }
99}
100
101//------------------------------------------------------------------------------
102// Transforms (Paragraph 14.4)
103
104// Does one or two inverse transforms.
105static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
106                           int do_two) {
107  // This implementation makes use of 16-bit fixed point versions of two
108  // multiply constants:
109  //    K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
110  //    K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16
111  //
112  // To be able to use signed 16-bit integers, we use the following trick to
113  // have constants within range:
114  // - Associated constants are obtained by subtracting the 16-bit fixed point
115  //   version of one:
116  //      k = K - (1 << 16)  =>  K = k + (1 << 16)
117  //      K1 = 85267  =>  k1 =  20091
118  //      K2 = 35468  =>  k2 = -30068
119  // - The multiplication of a variable by a constant become the sum of the
120  //   variable and the multiplication of that variable by the associated
121  //   constant:
122  //      (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
123  const __m128i k1 = _mm_set1_epi16(20091);
124  const __m128i k2 = _mm_set1_epi16(-30068);
125  __m128i T0, T1, T2, T3;
126
127  // Load and concatenate the transform coefficients (we'll do two inverse
128  // transforms in parallel). In the case of only one inverse transform, the
129  // second half of the vectors will just contain random value we'll never
130  // use nor store.
131  __m128i in0, in1, in2, in3;
132  {
133    in0 = _mm_loadl_epi64((__m128i*)&in[0]);
134    in1 = _mm_loadl_epi64((__m128i*)&in[4]);
135    in2 = _mm_loadl_epi64((__m128i*)&in[8]);
136    in3 = _mm_loadl_epi64((__m128i*)&in[12]);
137    // a00 a10 a20 a30   x x x x
138    // a01 a11 a21 a31   x x x x
139    // a02 a12 a22 a32   x x x x
140    // a03 a13 a23 a33   x x x x
141    if (do_two) {
142      const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]);
143      const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]);
144      const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]);
145      const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]);
146      in0 = _mm_unpacklo_epi64(in0, inB0);
147      in1 = _mm_unpacklo_epi64(in1, inB1);
148      in2 = _mm_unpacklo_epi64(in2, inB2);
149      in3 = _mm_unpacklo_epi64(in3, inB3);
150      // a00 a10 a20 a30   b00 b10 b20 b30
151      // a01 a11 a21 a31   b01 b11 b21 b31
152      // a02 a12 a22 a32   b02 b12 b22 b32
153      // a03 a13 a23 a33   b03 b13 b23 b33
154    }
155  }
156
157  // Vertical pass and subsequent transpose.
158  {
159    // First pass, c and d calculations are longer because of the "trick"
160    // multiplications.
161    const __m128i a = _mm_add_epi16(in0, in2);
162    const __m128i b = _mm_sub_epi16(in0, in2);
163    // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
164    const __m128i c1 = _mm_mulhi_epi16(in1, k2);
165    const __m128i c2 = _mm_mulhi_epi16(in3, k1);
166    const __m128i c3 = _mm_sub_epi16(in1, in3);
167    const __m128i c4 = _mm_sub_epi16(c1, c2);
168    const __m128i c = _mm_add_epi16(c3, c4);
169    // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
170    const __m128i d1 = _mm_mulhi_epi16(in1, k1);
171    const __m128i d2 = _mm_mulhi_epi16(in3, k2);
172    const __m128i d3 = _mm_add_epi16(in1, in3);
173    const __m128i d4 = _mm_add_epi16(d1, d2);
174    const __m128i d = _mm_add_epi16(d3, d4);
175
176    // Second pass.
177    const __m128i tmp0 = _mm_add_epi16(a, d);
178    const __m128i tmp1 = _mm_add_epi16(b, c);
179    const __m128i tmp2 = _mm_sub_epi16(b, c);
180    const __m128i tmp3 = _mm_sub_epi16(a, d);
181
182    // Transpose the two 4x4.
183    // a00 a01 a02 a03   b00 b01 b02 b03
184    // a10 a11 a12 a13   b10 b11 b12 b13
185    // a20 a21 a22 a23   b20 b21 b22 b23
186    // a30 a31 a32 a33   b30 b31 b32 b33
187    const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1);
188    const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3);
189    const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1);
190    const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3);
191    // a00 a10 a01 a11   a02 a12 a03 a13
192    // a20 a30 a21 a31   a22 a32 a23 a33
193    // b00 b10 b01 b11   b02 b12 b03 b13
194    // b20 b30 b21 b31   b22 b32 b23 b33
195    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
196    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
197    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
198    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
199    // a00 a10 a20 a30 a01 a11 a21 a31
200    // b00 b10 b20 b30 b01 b11 b21 b31
201    // a02 a12 a22 a32 a03 a13 a23 a33
202    // b02 b12 a22 b32 b03 b13 b23 b33
203    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
204    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
205    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
206    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
207    // a00 a10 a20 a30   b00 b10 b20 b30
208    // a01 a11 a21 a31   b01 b11 b21 b31
209    // a02 a12 a22 a32   b02 b12 b22 b32
210    // a03 a13 a23 a33   b03 b13 b23 b33
211  }
212
213  // Horizontal pass and subsequent transpose.
214  {
215    // First pass, c and d calculations are longer because of the "trick"
216    // multiplications.
217    const __m128i four = _mm_set1_epi16(4);
218    const __m128i dc = _mm_add_epi16(T0, four);
219    const __m128i a =  _mm_add_epi16(dc, T2);
220    const __m128i b =  _mm_sub_epi16(dc, T2);
221    // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
222    const __m128i c1 = _mm_mulhi_epi16(T1, k2);
223    const __m128i c2 = _mm_mulhi_epi16(T3, k1);
224    const __m128i c3 = _mm_sub_epi16(T1, T3);
225    const __m128i c4 = _mm_sub_epi16(c1, c2);
226    const __m128i c = _mm_add_epi16(c3, c4);
227    // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
228    const __m128i d1 = _mm_mulhi_epi16(T1, k1);
229    const __m128i d2 = _mm_mulhi_epi16(T3, k2);
230    const __m128i d3 = _mm_add_epi16(T1, T3);
231    const __m128i d4 = _mm_add_epi16(d1, d2);
232    const __m128i d = _mm_add_epi16(d3, d4);
233
234    // Second pass.
235    const __m128i tmp0 = _mm_add_epi16(a, d);
236    const __m128i tmp1 = _mm_add_epi16(b, c);
237    const __m128i tmp2 = _mm_sub_epi16(b, c);
238    const __m128i tmp3 = _mm_sub_epi16(a, d);
239    const __m128i shifted0 = _mm_srai_epi16(tmp0, 3);
240    const __m128i shifted1 = _mm_srai_epi16(tmp1, 3);
241    const __m128i shifted2 = _mm_srai_epi16(tmp2, 3);
242    const __m128i shifted3 = _mm_srai_epi16(tmp3, 3);
243
244    // Transpose the two 4x4.
245    // a00 a01 a02 a03   b00 b01 b02 b03
246    // a10 a11 a12 a13   b10 b11 b12 b13
247    // a20 a21 a22 a23   b20 b21 b22 b23
248    // a30 a31 a32 a33   b30 b31 b32 b33
249    const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1);
250    const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3);
251    const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1);
252    const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3);
253    // a00 a10 a01 a11   a02 a12 a03 a13
254    // a20 a30 a21 a31   a22 a32 a23 a33
255    // b00 b10 b01 b11   b02 b12 b03 b13
256    // b20 b30 b21 b31   b22 b32 b23 b33
257    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
258    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
259    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
260    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
261    // a00 a10 a20 a30 a01 a11 a21 a31
262    // b00 b10 b20 b30 b01 b11 b21 b31
263    // a02 a12 a22 a32 a03 a13 a23 a33
264    // b02 b12 a22 b32 b03 b13 b23 b33
265    T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
266    T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
267    T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
268    T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
269    // a00 a10 a20 a30   b00 b10 b20 b30
270    // a01 a11 a21 a31   b01 b11 b21 b31
271    // a02 a12 a22 a32   b02 b12 b22 b32
272    // a03 a13 a23 a33   b03 b13 b23 b33
273  }
274
275  // Add inverse transform to 'ref' and store.
276  {
277    const __m128i zero = _mm_setzero_si128();
278    // Load the reference(s).
279    __m128i ref0, ref1, ref2, ref3;
280    if (do_two) {
281      // Load eight bytes/pixels per line.
282      ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
283      ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
284      ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
285      ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
286    } else {
287      // Load four bytes/pixels per line.
288      ref0 = _mm_cvtsi32_si128(*(int*)&ref[0 * BPS]);
289      ref1 = _mm_cvtsi32_si128(*(int*)&ref[1 * BPS]);
290      ref2 = _mm_cvtsi32_si128(*(int*)&ref[2 * BPS]);
291      ref3 = _mm_cvtsi32_si128(*(int*)&ref[3 * BPS]);
292    }
293    // Convert to 16b.
294    ref0 = _mm_unpacklo_epi8(ref0, zero);
295    ref1 = _mm_unpacklo_epi8(ref1, zero);
296    ref2 = _mm_unpacklo_epi8(ref2, zero);
297    ref3 = _mm_unpacklo_epi8(ref3, zero);
298    // Add the inverse transform(s).
299    ref0 = _mm_add_epi16(ref0, T0);
300    ref1 = _mm_add_epi16(ref1, T1);
301    ref2 = _mm_add_epi16(ref2, T2);
302    ref3 = _mm_add_epi16(ref3, T3);
303    // Unsigned saturate to 8b.
304    ref0 = _mm_packus_epi16(ref0, ref0);
305    ref1 = _mm_packus_epi16(ref1, ref1);
306    ref2 = _mm_packus_epi16(ref2, ref2);
307    ref3 = _mm_packus_epi16(ref3, ref3);
308    // Store the results.
309    if (do_two) {
310      // Store eight bytes/pixels per line.
311      _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
312      _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
313      _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
314      _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
315    } else {
316      // Store four bytes/pixels per line.
317      *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(ref0);
318      *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(ref1);
319      *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(ref2);
320      *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(ref3);
321    }
322  }
323}
324
325static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
326                           int16_t* out) {
327  const __m128i zero = _mm_setzero_si128();
328  const __m128i seven = _mm_set1_epi16(7);
329  const __m128i k937 = _mm_set1_epi32(937);
330  const __m128i k1812 = _mm_set1_epi32(1812);
331  const __m128i k51000 = _mm_set1_epi32(51000);
332  const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
333  const __m128i k5352_2217 = _mm_set_epi16(5352,  2217, 5352,  2217,
334                                           5352,  2217, 5352,  2217);
335  const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352,
336                                           2217, -5352, 2217, -5352);
337  const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8);
338  const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8);
339  const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352,
340                                            2217, 5352, 2217, 5352);
341  const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217,
342                                            -5352, 2217, -5352, 2217);
343  __m128i v01, v32;
344
345
346  // Difference between src and ref and initial transpose.
347  {
348    // Load src and convert to 16b.
349    const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]);
350    const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]);
351    const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]);
352    const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]);
353    const __m128i src_0 = _mm_unpacklo_epi8(src0, zero);
354    const __m128i src_1 = _mm_unpacklo_epi8(src1, zero);
355    const __m128i src_2 = _mm_unpacklo_epi8(src2, zero);
356    const __m128i src_3 = _mm_unpacklo_epi8(src3, zero);
357    // Load ref and convert to 16b.
358    const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
359    const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
360    const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
361    const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
362    const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero);
363    const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
364    const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
365    const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
366    // Compute difference. -> 00 01 02 03 00 00 00 00
367    const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
368    const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
369    const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
370    const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
371
372
373    // Unpack and shuffle
374    // 00 01 02 03   0 0 0 0
375    // 10 11 12 13   0 0 0 0
376    // 20 21 22 23   0 0 0 0
377    // 30 31 32 33   0 0 0 0
378    const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1);
379    const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3);
380    // 00 01 10 11 02 03 12 13
381    // 20 21 30 31 22 23 32 33
382    const __m128i shuf01_p =
383        _mm_shufflehi_epi16(shuf01, _MM_SHUFFLE(2, 3, 0, 1));
384    const __m128i shuf23_p =
385        _mm_shufflehi_epi16(shuf23, _MM_SHUFFLE(2, 3, 0, 1));
386    // 00 01 10 11 03 02 13 12
387    // 20 21 30 31 23 22 33 32
388    const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p);
389    const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p);
390    // 00 01 10 11 20 21 30 31
391    // 03 02 13 12 23 22 33 32
392    const __m128i a01 = _mm_add_epi16(s01, s32);
393    const __m128i a32 = _mm_sub_epi16(s01, s32);
394    // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ]
395    // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ]
396
397    const __m128i tmp0 = _mm_madd_epi16(a01, k88p);  // [ (a0 + a1) << 3, ... ]
398    const __m128i tmp2 = _mm_madd_epi16(a01, k88m);  // [ (a0 - a1) << 3, ... ]
399    const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p);
400    const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m);
401    const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812);
402    const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937);
403    const __m128i tmp1   = _mm_srai_epi32(tmp1_2, 9);
404    const __m128i tmp3   = _mm_srai_epi32(tmp3_2, 9);
405    const __m128i s03 = _mm_packs_epi32(tmp0, tmp2);
406    const __m128i s12 = _mm_packs_epi32(tmp1, tmp3);
407    const __m128i s_lo = _mm_unpacklo_epi16(s03, s12);   // 0 1 0 1 0 1...
408    const __m128i s_hi = _mm_unpackhi_epi16(s03, s12);   // 2 3 2 3 2 3
409    const __m128i v23 = _mm_unpackhi_epi32(s_lo, s_hi);
410    v01 = _mm_unpacklo_epi32(s_lo, s_hi);
411    v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));  // 3 2 3 2 3 2..
412  }
413
414  // Second pass
415  {
416    // Same operations are done on the (0,3) and (1,2) pairs.
417    // a0 = v0 + v3
418    // a1 = v1 + v2
419    // a3 = v0 - v3
420    // a2 = v1 - v2
421    const __m128i a01 = _mm_add_epi16(v01, v32);
422    const __m128i a32 = _mm_sub_epi16(v01, v32);
423    const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
424    const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
425    const __m128i a01_plus_7 = _mm_add_epi16(a01, seven);
426
427    // d0 = (a0 + a1 + 7) >> 4;
428    // d2 = (a0 - a1 + 7) >> 4;
429    const __m128i c0 = _mm_add_epi16(a01_plus_7, a11);
430    const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11);
431    const __m128i d0 = _mm_srai_epi16(c0, 4);
432    const __m128i d2 = _mm_srai_epi16(c2, 4);
433
434    // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16)
435    // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16)
436    const __m128i b23 = _mm_unpacklo_epi16(a22, a32);
437    const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
438    const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
439    const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one);
440    const __m128i d3 = _mm_add_epi32(c3, k51000);
441    const __m128i e1 = _mm_srai_epi32(d1, 16);
442    const __m128i e3 = _mm_srai_epi32(d3, 16);
443    const __m128i f1 = _mm_packs_epi32(e1, e1);
444    const __m128i f3 = _mm_packs_epi32(e3, e3);
445    // f1 = f1 + (a3 != 0);
446    // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
447    // desired (0, 1), we add one earlier through k12000_plus_one.
448    // -> f1 = f1 + 1 - (a3 == 0)
449    const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
450
451    _mm_storel_epi64((__m128i*)&out[ 0], d0);
452    _mm_storel_epi64((__m128i*)&out[ 4], g1);
453    _mm_storel_epi64((__m128i*)&out[ 8], d2);
454    _mm_storel_epi64((__m128i*)&out[12], f3);
455  }
456}
457
458static void FTransformWHTSSE2(const int16_t* in, int16_t* out) {
459  int16_t tmp[16];
460  int i;
461  for (i = 0; i < 4; ++i, in += 64) {
462    const int a0 = (in[0 * 16] + in[2 * 16]);
463    const int a1 = (in[1 * 16] + in[3 * 16]);
464    const int a2 = (in[1 * 16] - in[3 * 16]);
465    const int a3 = (in[0 * 16] - in[2 * 16]);
466    tmp[0 + i * 4] = a0 + a1;
467    tmp[1 + i * 4] = a3 + a2;
468    tmp[2 + i * 4] = a3 - a2;
469    tmp[3 + i * 4] = a0 - a1;
470  }
471  {
472    const __m128i src0 = _mm_loadl_epi64((__m128i*)&tmp[0]);
473    const __m128i src1 = _mm_loadl_epi64((__m128i*)&tmp[4]);
474    const __m128i src2 = _mm_loadl_epi64((__m128i*)&tmp[8]);
475    const __m128i src3 = _mm_loadl_epi64((__m128i*)&tmp[12]);
476    const __m128i a0 = _mm_add_epi16(src0, src2);
477    const __m128i a1 = _mm_add_epi16(src1, src3);
478    const __m128i a2 = _mm_sub_epi16(src1, src3);
479    const __m128i a3 = _mm_sub_epi16(src0, src2);
480    const __m128i b0 = _mm_srai_epi16(_mm_adds_epi16(a0, a1), 1);
481    const __m128i b1 = _mm_srai_epi16(_mm_adds_epi16(a3, a2), 1);
482    const __m128i b2 = _mm_srai_epi16(_mm_subs_epi16(a3, a2), 1);
483    const __m128i b3 = _mm_srai_epi16(_mm_subs_epi16(a0, a1), 1);
484    _mm_storel_epi64((__m128i*)&out[ 0], b0);
485    _mm_storel_epi64((__m128i*)&out[ 4], b1);
486    _mm_storel_epi64((__m128i*)&out[ 8], b2);
487    _mm_storel_epi64((__m128i*)&out[12], b3);
488  }
489}
490
491//------------------------------------------------------------------------------
492// Metric
493
494static int SSE_Nx4SSE2(const uint8_t* a, const uint8_t* b,
495                       int num_quads, int do_16) {
496  const __m128i zero = _mm_setzero_si128();
497  __m128i sum1 = zero;
498  __m128i sum2 = zero;
499
500  while (num_quads-- > 0) {
501    // Note: for the !do_16 case, we read 16 pixels instead of 8 but that's ok,
502    // thanks to buffer over-allocation to that effect.
503    const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]);
504    const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]);
505    const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]);
506    const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]);
507    const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]);
508    const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]);
509    const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]);
510    const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]);
511
512    // compute clip0(a-b) and clip0(b-a)
513    const __m128i a0p = _mm_subs_epu8(a0, b0);
514    const __m128i a0m = _mm_subs_epu8(b0, a0);
515    const __m128i a1p = _mm_subs_epu8(a1, b1);
516    const __m128i a1m = _mm_subs_epu8(b1, a1);
517    const __m128i a2p = _mm_subs_epu8(a2, b2);
518    const __m128i a2m = _mm_subs_epu8(b2, a2);
519    const __m128i a3p = _mm_subs_epu8(a3, b3);
520    const __m128i a3m = _mm_subs_epu8(b3, a3);
521
522    // compute |a-b| with 8b arithmetic as clip0(a-b) | clip0(b-a)
523    const __m128i diff0 = _mm_or_si128(a0p, a0m);
524    const __m128i diff1 = _mm_or_si128(a1p, a1m);
525    const __m128i diff2 = _mm_or_si128(a2p, a2m);
526    const __m128i diff3 = _mm_or_si128(a3p, a3m);
527
528    // unpack (only four operations, instead of eight)
529    const __m128i low0 = _mm_unpacklo_epi8(diff0, zero);
530    const __m128i low1 = _mm_unpacklo_epi8(diff1, zero);
531    const __m128i low2 = _mm_unpacklo_epi8(diff2, zero);
532    const __m128i low3 = _mm_unpacklo_epi8(diff3, zero);
533
534    // multiply with self
535    const __m128i low_madd0 = _mm_madd_epi16(low0, low0);
536    const __m128i low_madd1 = _mm_madd_epi16(low1, low1);
537    const __m128i low_madd2 = _mm_madd_epi16(low2, low2);
538    const __m128i low_madd3 = _mm_madd_epi16(low3, low3);
539
540    // collect in a cascading way
541    const __m128i low_sum0 = _mm_add_epi32(low_madd0, low_madd1);
542    const __m128i low_sum1 = _mm_add_epi32(low_madd2, low_madd3);
543    sum1 = _mm_add_epi32(sum1, low_sum0);
544    sum2 = _mm_add_epi32(sum2, low_sum1);
545
546    if (do_16) {  // if necessary, process the higher 8 bytes similarly
547      const __m128i hi0 = _mm_unpackhi_epi8(diff0, zero);
548      const __m128i hi1 = _mm_unpackhi_epi8(diff1, zero);
549      const __m128i hi2 = _mm_unpackhi_epi8(diff2, zero);
550      const __m128i hi3 = _mm_unpackhi_epi8(diff3, zero);
551
552      const __m128i hi_madd0 = _mm_madd_epi16(hi0, hi0);
553      const __m128i hi_madd1 = _mm_madd_epi16(hi1, hi1);
554      const __m128i hi_madd2 = _mm_madd_epi16(hi2, hi2);
555      const __m128i hi_madd3 = _mm_madd_epi16(hi3, hi3);
556      const __m128i hi_sum0 = _mm_add_epi32(hi_madd0, hi_madd1);
557      const __m128i hi_sum1 = _mm_add_epi32(hi_madd2, hi_madd3);
558      sum1 = _mm_add_epi32(sum1, hi_sum0);
559      sum2 = _mm_add_epi32(sum2, hi_sum1);
560    }
561    a += 4 * BPS;
562    b += 4 * BPS;
563  }
564  {
565    int32_t tmp[4];
566    const __m128i sum = _mm_add_epi32(sum1, sum2);
567    _mm_storeu_si128((__m128i*)tmp, sum);
568    return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
569  }
570}
571
572static int SSE16x16SSE2(const uint8_t* a, const uint8_t* b) {
573  return SSE_Nx4SSE2(a, b, 4, 1);
574}
575
576static int SSE16x8SSE2(const uint8_t* a, const uint8_t* b) {
577  return SSE_Nx4SSE2(a, b, 2, 1);
578}
579
580static int SSE8x8SSE2(const uint8_t* a, const uint8_t* b) {
581  return SSE_Nx4SSE2(a, b, 2, 0);
582}
583
584static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
585  const __m128i zero = _mm_setzero_si128();
586
587  // Load values. Note that we read 8 pixels instead of 4,
588  // but the a/b buffers are over-allocated to that effect.
589  const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]);
590  const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]);
591  const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]);
592  const __m128i a3 = _mm_loadl_epi64((__m128i*)&a[BPS * 3]);
593  const __m128i b0 = _mm_loadl_epi64((__m128i*)&b[BPS * 0]);
594  const __m128i b1 = _mm_loadl_epi64((__m128i*)&b[BPS * 1]);
595  const __m128i b2 = _mm_loadl_epi64((__m128i*)&b[BPS * 2]);
596  const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]);
597
598  // Combine pair of lines and convert to 16b.
599  const __m128i a01 = _mm_unpacklo_epi32(a0, a1);
600  const __m128i a23 = _mm_unpacklo_epi32(a2, a3);
601  const __m128i b01 = _mm_unpacklo_epi32(b0, b1);
602  const __m128i b23 = _mm_unpacklo_epi32(b2, b3);
603  const __m128i a01s = _mm_unpacklo_epi8(a01, zero);
604  const __m128i a23s = _mm_unpacklo_epi8(a23, zero);
605  const __m128i b01s = _mm_unpacklo_epi8(b01, zero);
606  const __m128i b23s = _mm_unpacklo_epi8(b23, zero);
607
608  // Compute differences; (a-b)^2 = (abs(a-b))^2 = (sat8(a-b) + sat8(b-a))^2
609  // TODO(cduvivier): Dissassemble and figure out why this is fastest. We don't
610  //                  need absolute values, there is no need to do calculation
611  //                  in 8bit as we are already in 16bit, ... Yet this is what
612  //                  benchmarks the fastest!
613  const __m128i d0 = _mm_subs_epu8(a01s, b01s);
614  const __m128i d1 = _mm_subs_epu8(b01s, a01s);
615  const __m128i d2 = _mm_subs_epu8(a23s, b23s);
616  const __m128i d3 = _mm_subs_epu8(b23s, a23s);
617
618  // Square and add them all together.
619  const __m128i madd0 = _mm_madd_epi16(d0, d0);
620  const __m128i madd1 = _mm_madd_epi16(d1, d1);
621  const __m128i madd2 = _mm_madd_epi16(d2, d2);
622  const __m128i madd3 = _mm_madd_epi16(d3, d3);
623  const __m128i sum0 = _mm_add_epi32(madd0, madd1);
624  const __m128i sum1 = _mm_add_epi32(madd2, madd3);
625  const __m128i sum2 = _mm_add_epi32(sum0, sum1);
626
627  int32_t tmp[4];
628  _mm_storeu_si128((__m128i*)tmp, sum2);
629  return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
630}
631
632//------------------------------------------------------------------------------
633// Texture distortion
634//
635// We try to match the spectral content (weighted) between source and
636// reconstructed samples.
637
638// Hadamard transform
639// Returns the difference between the weighted sum of the absolute value of
640// transformed coefficients.
641static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
642                          const uint16_t* const w) {
643  int32_t sum[4];
644  __m128i tmp_0, tmp_1, tmp_2, tmp_3;
645  const __m128i zero = _mm_setzero_si128();
646
647  // Load, combine and tranpose inputs.
648  {
649    const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]);
650    const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]);
651    const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]);
652    const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]);
653    const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]);
654    const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]);
655    const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]);
656    const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]);
657
658    // Combine inA and inB (we'll do two transforms in parallel).
659    const __m128i inAB_0 = _mm_unpacklo_epi8(inA_0, inB_0);
660    const __m128i inAB_1 = _mm_unpacklo_epi8(inA_1, inB_1);
661    const __m128i inAB_2 = _mm_unpacklo_epi8(inA_2, inB_2);
662    const __m128i inAB_3 = _mm_unpacklo_epi8(inA_3, inB_3);
663    // a00 b00 a01 b01 a02 b03 a03 b03   0 0 0 0 0 0 0 0
664    // a10 b10 a11 b11 a12 b12 a13 b13   0 0 0 0 0 0 0 0
665    // a20 b20 a21 b21 a22 b22 a23 b23   0 0 0 0 0 0 0 0
666    // a30 b30 a31 b31 a32 b32 a33 b33   0 0 0 0 0 0 0 0
667
668    // Transpose the two 4x4, discarding the filling zeroes.
669    const __m128i transpose0_0 = _mm_unpacklo_epi8(inAB_0, inAB_2);
670    const __m128i transpose0_1 = _mm_unpacklo_epi8(inAB_1, inAB_3);
671    // a00 a20  b00 b20  a01 a21  b01 b21  a02 a22  b02 b22  a03 a23  b03 b23
672    // a10 a30  b10 b30  a11 a31  b11 b31  a12 a32  b12 b32  a13 a33  b13 b33
673    const __m128i transpose1_0 = _mm_unpacklo_epi8(transpose0_0, transpose0_1);
674    const __m128i transpose1_1 = _mm_unpackhi_epi8(transpose0_0, transpose0_1);
675    // a00 a10 a20 a30  b00 b10 b20 b30  a01 a11 a21 a31  b01 b11 b21 b31
676    // a02 a12 a22 a32  b02 b12 b22 b32  a03 a13 a23 a33  b03 b13 b23 b33
677
678    // Convert to 16b.
679    tmp_0 = _mm_unpacklo_epi8(transpose1_0, zero);
680    tmp_1 = _mm_unpackhi_epi8(transpose1_0, zero);
681    tmp_2 = _mm_unpacklo_epi8(transpose1_1, zero);
682    tmp_3 = _mm_unpackhi_epi8(transpose1_1, zero);
683    // a00 a10 a20 a30   b00 b10 b20 b30
684    // a01 a11 a21 a31   b01 b11 b21 b31
685    // a02 a12 a22 a32   b02 b12 b22 b32
686    // a03 a13 a23 a33   b03 b13 b23 b33
687  }
688
689  // Horizontal pass and subsequent transpose.
690  {
691    // Calculate a and b (two 4x4 at once).
692    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
693    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
694    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
695    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
696    const __m128i b0 = _mm_add_epi16(a0, a1);
697    const __m128i b1 = _mm_add_epi16(a3, a2);
698    const __m128i b2 = _mm_sub_epi16(a3, a2);
699    const __m128i b3 = _mm_sub_epi16(a0, a1);
700    // a00 a01 a02 a03   b00 b01 b02 b03
701    // a10 a11 a12 a13   b10 b11 b12 b13
702    // a20 a21 a22 a23   b20 b21 b22 b23
703    // a30 a31 a32 a33   b30 b31 b32 b33
704
705    // Transpose the two 4x4.
706    const __m128i transpose0_0 = _mm_unpacklo_epi16(b0, b1);
707    const __m128i transpose0_1 = _mm_unpacklo_epi16(b2, b3);
708    const __m128i transpose0_2 = _mm_unpackhi_epi16(b0, b1);
709    const __m128i transpose0_3 = _mm_unpackhi_epi16(b2, b3);
710    // a00 a10 a01 a11   a02 a12 a03 a13
711    // a20 a30 a21 a31   a22 a32 a23 a33
712    // b00 b10 b01 b11   b02 b12 b03 b13
713    // b20 b30 b21 b31   b22 b32 b23 b33
714    const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
715    const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
716    const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
717    const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
718    // a00 a10 a20 a30 a01 a11 a21 a31
719    // b00 b10 b20 b30 b01 b11 b21 b31
720    // a02 a12 a22 a32 a03 a13 a23 a33
721    // b02 b12 a22 b32 b03 b13 b23 b33
722    tmp_0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
723    tmp_1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
724    tmp_2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
725    tmp_3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
726    // a00 a10 a20 a30   b00 b10 b20 b30
727    // a01 a11 a21 a31   b01 b11 b21 b31
728    // a02 a12 a22 a32   b02 b12 b22 b32
729    // a03 a13 a23 a33   b03 b13 b23 b33
730  }
731
732  // Vertical pass and difference of weighted sums.
733  {
734    // Load all inputs.
735    // TODO(cduvivier): Make variable declarations and allocations aligned so
736    //                  we can use _mm_load_si128 instead of _mm_loadu_si128.
737    const __m128i w_0 = _mm_loadu_si128((__m128i*)&w[0]);
738    const __m128i w_8 = _mm_loadu_si128((__m128i*)&w[8]);
739
740    // Calculate a and b (two 4x4 at once).
741    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
742    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
743    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
744    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
745    const __m128i b0 = _mm_add_epi16(a0, a1);
746    const __m128i b1 = _mm_add_epi16(a3, a2);
747    const __m128i b2 = _mm_sub_epi16(a3, a2);
748    const __m128i b3 = _mm_sub_epi16(a0, a1);
749
750    // Separate the transforms of inA and inB.
751    __m128i A_b0 = _mm_unpacklo_epi64(b0, b1);
752    __m128i A_b2 = _mm_unpacklo_epi64(b2, b3);
753    __m128i B_b0 = _mm_unpackhi_epi64(b0, b1);
754    __m128i B_b2 = _mm_unpackhi_epi64(b2, b3);
755
756    {
757      // sign(b) = b >> 15  (0x0000 if positive, 0xffff if negative)
758      const __m128i sign_A_b0 = _mm_srai_epi16(A_b0, 15);
759      const __m128i sign_A_b2 = _mm_srai_epi16(A_b2, 15);
760      const __m128i sign_B_b0 = _mm_srai_epi16(B_b0, 15);
761      const __m128i sign_B_b2 = _mm_srai_epi16(B_b2, 15);
762
763      // b = abs(b) = (b ^ sign) - sign
764      A_b0 = _mm_xor_si128(A_b0, sign_A_b0);
765      A_b2 = _mm_xor_si128(A_b2, sign_A_b2);
766      B_b0 = _mm_xor_si128(B_b0, sign_B_b0);
767      B_b2 = _mm_xor_si128(B_b2, sign_B_b2);
768      A_b0 = _mm_sub_epi16(A_b0, sign_A_b0);
769      A_b2 = _mm_sub_epi16(A_b2, sign_A_b2);
770      B_b0 = _mm_sub_epi16(B_b0, sign_B_b0);
771      B_b2 = _mm_sub_epi16(B_b2, sign_B_b2);
772    }
773
774    // weighted sums
775    A_b0 = _mm_madd_epi16(A_b0, w_0);
776    A_b2 = _mm_madd_epi16(A_b2, w_8);
777    B_b0 = _mm_madd_epi16(B_b0, w_0);
778    B_b2 = _mm_madd_epi16(B_b2, w_8);
779    A_b0 = _mm_add_epi32(A_b0, A_b2);
780    B_b0 = _mm_add_epi32(B_b0, B_b2);
781
782    // difference of weighted sums
783    A_b0 = _mm_sub_epi32(A_b0, B_b0);
784    _mm_storeu_si128((__m128i*)&sum[0], A_b0);
785  }
786  return sum[0] + sum[1] + sum[2] + sum[3];
787}
788
789static int Disto4x4SSE2(const uint8_t* const a, const uint8_t* const b,
790                        const uint16_t* const w) {
791  const int diff_sum = TTransformSSE2(a, b, w);
792  return abs(diff_sum) >> 5;
793}
794
795static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b,
796                          const uint16_t* const w) {
797  int D = 0;
798  int x, y;
799  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
800    for (x = 0; x < 16; x += 4) {
801      D += Disto4x4SSE2(a + x + y, b + x + y, w);
802    }
803  }
804  return D;
805}
806
807//------------------------------------------------------------------------------
808// Quantization
809//
810
811// Simple quantization
812static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
813                             int n, const VP8Matrix* const mtx) {
814  const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
815  const __m128i zero = _mm_setzero_si128();
816  __m128i coeff0, coeff8;
817  __m128i out0, out8;
818  __m128i packed_out;
819
820  // Load all inputs.
821  // TODO(cduvivier): Make variable declarations and allocations aligned so that
822  //                  we can use _mm_load_si128 instead of _mm_loadu_si128.
823  __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
824  __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
825  const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]);
826  const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]);
827  const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]);
828  const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]);
829  const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]);
830  const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]);
831  const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]);
832  const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]);
833  const __m128i zthresh0 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[0]);
834  const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]);
835
836  // sign(in) = in >> 15  (0x0000 if positive, 0xffff if negative)
837  const __m128i sign0 = _mm_srai_epi16(in0, 15);
838  const __m128i sign8 = _mm_srai_epi16(in8, 15);
839
840  // coeff = abs(in) = (in ^ sign) - sign
841  coeff0 = _mm_xor_si128(in0, sign0);
842  coeff8 = _mm_xor_si128(in8, sign8);
843  coeff0 = _mm_sub_epi16(coeff0, sign0);
844  coeff8 = _mm_sub_epi16(coeff8, sign8);
845
846  // coeff = abs(in) + sharpen
847  coeff0 = _mm_add_epi16(coeff0, sharpen0);
848  coeff8 = _mm_add_epi16(coeff8, sharpen8);
849
850  // out = (coeff * iQ + B) >> QFIX;
851  {
852    // doing calculations with 32b precision (QFIX=17)
853    // out = (coeff * iQ)
854    __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
855    __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
856    __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
857    __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
858    __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
859    __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
860    __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
861    __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
862    // expand bias from 16b to 32b
863    __m128i bias_00 = _mm_unpacklo_epi16(bias0, zero);
864    __m128i bias_04 = _mm_unpackhi_epi16(bias0, zero);
865    __m128i bias_08 = _mm_unpacklo_epi16(bias8, zero);
866    __m128i bias_12 = _mm_unpackhi_epi16(bias8, zero);
867    // out = (coeff * iQ + B)
868    out_00 = _mm_add_epi32(out_00, bias_00);
869    out_04 = _mm_add_epi32(out_04, bias_04);
870    out_08 = _mm_add_epi32(out_08, bias_08);
871    out_12 = _mm_add_epi32(out_12, bias_12);
872    // out = (coeff * iQ + B) >> QFIX;
873    out_00 = _mm_srai_epi32(out_00, QFIX);
874    out_04 = _mm_srai_epi32(out_04, QFIX);
875    out_08 = _mm_srai_epi32(out_08, QFIX);
876    out_12 = _mm_srai_epi32(out_12, QFIX);
877
878    // pack result as 16b
879    out0 = _mm_packs_epi32(out_00, out_04);
880    out8 = _mm_packs_epi32(out_08, out_12);
881
882    // if (coeff > 2047) coeff = 2047
883    out0 = _mm_min_epi16(out0, max_coeff_2047);
884    out8 = _mm_min_epi16(out8, max_coeff_2047);
885  }
886
887  // get sign back (if (sign[j]) out_n = -out_n)
888  out0 = _mm_xor_si128(out0, sign0);
889  out8 = _mm_xor_si128(out8, sign8);
890  out0 = _mm_sub_epi16(out0, sign0);
891  out8 = _mm_sub_epi16(out8, sign8);
892
893  // in = out * Q
894  in0 = _mm_mullo_epi16(out0, q0);
895  in8 = _mm_mullo_epi16(out8, q8);
896
897  // if (coeff <= mtx->zthresh_) {in=0; out=0;}
898  {
899    __m128i cmp0 = _mm_cmpgt_epi16(coeff0, zthresh0);
900    __m128i cmp8 = _mm_cmpgt_epi16(coeff8, zthresh8);
901    in0 = _mm_and_si128(in0, cmp0);
902    in8 = _mm_and_si128(in8, cmp8);
903    _mm_storeu_si128((__m128i*)&in[0], in0);
904    _mm_storeu_si128((__m128i*)&in[8], in8);
905    out0 = _mm_and_si128(out0, cmp0);
906    out8 = _mm_and_si128(out8, cmp8);
907  }
908
909  // zigzag the output before storing it.
910  //
911  // The zigzag pattern can almost be reproduced with a small sequence of
912  // shuffles. After it, we only need to swap the 7th (ending up in third
913  // position instead of twelfth) and 8th values.
914  {
915    __m128i outZ0, outZ8;
916    outZ0 = _mm_shufflehi_epi16(out0,  _MM_SHUFFLE(2, 1, 3, 0));
917    outZ0 = _mm_shuffle_epi32  (outZ0, _MM_SHUFFLE(3, 1, 2, 0));
918    outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2));
919    outZ8 = _mm_shufflelo_epi16(out8,  _MM_SHUFFLE(3, 0, 2, 1));
920    outZ8 = _mm_shuffle_epi32  (outZ8, _MM_SHUFFLE(3, 1, 2, 0));
921    outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0));
922    _mm_storeu_si128((__m128i*)&out[0], outZ0);
923    _mm_storeu_si128((__m128i*)&out[8], outZ8);
924    packed_out = _mm_packs_epi16(outZ0, outZ8);
925  }
926  {
927    const int16_t outZ_12 = out[12];
928    const int16_t outZ_3 = out[3];
929    out[3] = outZ_12;
930    out[12] = outZ_3;
931  }
932
933  // detect if all 'out' values are zeroes or not
934  {
935    int32_t tmp[4];
936    _mm_storeu_si128((__m128i*)tmp, packed_out);
937    if (n) {
938      tmp[0] &= ~0xff;
939    }
940    return (tmp[3] || tmp[2] || tmp[1] || tmp[0]);
941  }
942}
943
944#endif   // WEBP_USE_SSE2
945
946//------------------------------------------------------------------------------
947// Entry point
948
949extern void VP8EncDspInitSSE2(void);
950
951void VP8EncDspInitSSE2(void) {
952#if defined(WEBP_USE_SSE2)
953  VP8CollectHistogram = CollectHistogramSSE2;
954  VP8EncQuantizeBlock = QuantizeBlockSSE2;
955  VP8ITransform = ITransformSSE2;
956  VP8FTransform = FTransformSSE2;
957  VP8FTransformWHT = FTransformWHTSSE2;
958  VP8SSE16x16 = SSE16x16SSE2;
959  VP8SSE16x8 = SSE16x8SSE2;
960  VP8SSE8x8 = SSE8x8SSE2;
961  VP8SSE4x4 = SSE4x4SSE2;
962  VP8TDisto4x4 = Disto4x4SSE2;
963  VP8TDisto16x16 = Disto16x16SSE2;
964#endif   // WEBP_USE_SSE2
965}
966
967#if defined(__cplusplus) || defined(c_plusplus)
968}    // extern "C"
969#endif
970