• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2015 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 // SSE4 version of some encoding functions.
11 //
12 // Author: Skal (pascal.massimino@gmail.com)
13 
14 #include "src/dsp/dsp.h"
15 
16 #if defined(WEBP_USE_SSE41)
17 #include <smmintrin.h>
18 #include <stdlib.h>  // for abs()
19 
20 #include "src/dsp/common_sse2.h"
21 #include "src/enc/vp8i_enc.h"
22 
23 //------------------------------------------------------------------------------
24 // Compute susceptibility based on DCT-coeff histograms.
25 
CollectHistogram_SSE41(const uint8_t * ref,const uint8_t * pred,int start_block,int end_block,VP8Histogram * const histo)26 static void CollectHistogram_SSE41(const uint8_t* ref, const uint8_t* pred,
27                                    int start_block, int end_block,
28                                    VP8Histogram* const histo) {
29   const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
30   int j;
31   int distribution[MAX_COEFF_THRESH + 1] = { 0 };
32   for (j = start_block; j < end_block; ++j) {
33     int16_t out[16];
34     int k;
35 
36     VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
37 
38     // Convert coefficients to bin (within out[]).
39     {
40       // Load.
41       const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]);
42       const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]);
43       // v = abs(out) >> 3
44       const __m128i abs0 = _mm_abs_epi16(out0);
45       const __m128i abs1 = _mm_abs_epi16(out1);
46       const __m128i v0 = _mm_srai_epi16(abs0, 3);
47       const __m128i v1 = _mm_srai_epi16(abs1, 3);
48       // bin = min(v, MAX_COEFF_THRESH)
49       const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
50       const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
51       // Store.
52       _mm_storeu_si128((__m128i*)&out[0], bin0);
53       _mm_storeu_si128((__m128i*)&out[8], bin1);
54     }
55 
56     // Convert coefficients to bin.
57     for (k = 0; k < 16; ++k) {
58       ++distribution[out[k]];
59     }
60   }
61   VP8SetHistogramData(distribution, histo);
62 }
63 
64 //------------------------------------------------------------------------------
65 // Texture distortion
66 //
67 // We try to match the spectral content (weighted) between source and
68 // reconstructed samples.
69 
70 // Hadamard transform
71 // Returns the weighted sum of the absolute value of transformed coefficients.
72 // w[] contains a row-major 4 by 4 symmetric matrix.
TTransform_SSE41(const uint8_t * inA,const uint8_t * inB,const uint16_t * const w)73 static int TTransform_SSE41(const uint8_t* inA, const uint8_t* inB,
74                             const uint16_t* const w) {
75   int32_t sum[4];
76   __m128i tmp_0, tmp_1, tmp_2, tmp_3;
77 
78   // Load and combine inputs.
79   {
80     const __m128i inA_0 = _mm_loadu_si128((const __m128i*)&inA[BPS * 0]);
81     const __m128i inA_1 = _mm_loadu_si128((const __m128i*)&inA[BPS * 1]);
82     const __m128i inA_2 = _mm_loadu_si128((const __m128i*)&inA[BPS * 2]);
83     // In SSE4.1, with gcc 4.8 at least (maybe other versions),
84     // _mm_loadu_si128 is faster than _mm_loadl_epi64. But for the last lump
85     // of inA and inB, _mm_loadl_epi64 is still used not to have an out of
86     // bound read.
87     const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]);
88     const __m128i inB_0 = _mm_loadu_si128((const __m128i*)&inB[BPS * 0]);
89     const __m128i inB_1 = _mm_loadu_si128((const __m128i*)&inB[BPS * 1]);
90     const __m128i inB_2 = _mm_loadu_si128((const __m128i*)&inB[BPS * 2]);
91     const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]);
92 
93     // Combine inA and inB (we'll do two transforms in parallel).
94     const __m128i inAB_0 = _mm_unpacklo_epi32(inA_0, inB_0);
95     const __m128i inAB_1 = _mm_unpacklo_epi32(inA_1, inB_1);
96     const __m128i inAB_2 = _mm_unpacklo_epi32(inA_2, inB_2);
97     const __m128i inAB_3 = _mm_unpacklo_epi32(inA_3, inB_3);
98     tmp_0 = _mm_cvtepu8_epi16(inAB_0);
99     tmp_1 = _mm_cvtepu8_epi16(inAB_1);
100     tmp_2 = _mm_cvtepu8_epi16(inAB_2);
101     tmp_3 = _mm_cvtepu8_epi16(inAB_3);
102     // a00 a01 a02 a03   b00 b01 b02 b03
103     // a10 a11 a12 a13   b10 b11 b12 b13
104     // a20 a21 a22 a23   b20 b21 b22 b23
105     // a30 a31 a32 a33   b30 b31 b32 b33
106   }
107 
108   // Vertical pass first to avoid a transpose (vertical and horizontal passes
109   // are commutative because w/kWeightY is symmetric) and subsequent transpose.
110   {
111     // Calculate a and b (two 4x4 at once).
112     const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
113     const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
114     const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
115     const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
116     const __m128i b0 = _mm_add_epi16(a0, a1);
117     const __m128i b1 = _mm_add_epi16(a3, a2);
118     const __m128i b2 = _mm_sub_epi16(a3, a2);
119     const __m128i b3 = _mm_sub_epi16(a0, a1);
120     // a00 a01 a02 a03   b00 b01 b02 b03
121     // a10 a11 a12 a13   b10 b11 b12 b13
122     // a20 a21 a22 a23   b20 b21 b22 b23
123     // a30 a31 a32 a33   b30 b31 b32 b33
124 
125     // Transpose the two 4x4.
126     VP8Transpose_2_4x4_16b(&b0, &b1, &b2, &b3, &tmp_0, &tmp_1, &tmp_2, &tmp_3);
127   }
128 
129   // Horizontal pass and difference of weighted sums.
130   {
131     // Load all inputs.
132     const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]);
133     const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]);
134 
135     // Calculate a and b (two 4x4 at once).
136     const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
137     const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
138     const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
139     const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
140     const __m128i b0 = _mm_add_epi16(a0, a1);
141     const __m128i b1 = _mm_add_epi16(a3, a2);
142     const __m128i b2 = _mm_sub_epi16(a3, a2);
143     const __m128i b3 = _mm_sub_epi16(a0, a1);
144 
145     // Separate the transforms of inA and inB.
146     __m128i A_b0 = _mm_unpacklo_epi64(b0, b1);
147     __m128i A_b2 = _mm_unpacklo_epi64(b2, b3);
148     __m128i B_b0 = _mm_unpackhi_epi64(b0, b1);
149     __m128i B_b2 = _mm_unpackhi_epi64(b2, b3);
150 
151     A_b0 = _mm_abs_epi16(A_b0);
152     A_b2 = _mm_abs_epi16(A_b2);
153     B_b0 = _mm_abs_epi16(B_b0);
154     B_b2 = _mm_abs_epi16(B_b2);
155 
156     // weighted sums
157     A_b0 = _mm_madd_epi16(A_b0, w_0);
158     A_b2 = _mm_madd_epi16(A_b2, w_8);
159     B_b0 = _mm_madd_epi16(B_b0, w_0);
160     B_b2 = _mm_madd_epi16(B_b2, w_8);
161     A_b0 = _mm_add_epi32(A_b0, A_b2);
162     B_b0 = _mm_add_epi32(B_b0, B_b2);
163 
164     // difference of weighted sums
165     A_b2 = _mm_sub_epi32(A_b0, B_b0);
166     _mm_storeu_si128((__m128i*)&sum[0], A_b2);
167   }
168   return sum[0] + sum[1] + sum[2] + sum[3];
169 }
170 
Disto4x4_SSE41(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)171 static int Disto4x4_SSE41(const uint8_t* const a, const uint8_t* const b,
172                           const uint16_t* const w) {
173   const int diff_sum = TTransform_SSE41(a, b, w);
174   return abs(diff_sum) >> 5;
175 }
176 
Disto16x16_SSE41(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)177 static int Disto16x16_SSE41(const uint8_t* const a, const uint8_t* const b,
178                             const uint16_t* const w) {
179   int D = 0;
180   int x, y;
181   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
182     for (x = 0; x < 16; x += 4) {
183       D += Disto4x4_SSE41(a + x + y, b + x + y, w);
184     }
185   }
186   return D;
187 }
188 
189 //------------------------------------------------------------------------------
190 // Quantization
191 //
192 
193 // Generates a pshufb constant for shuffling 16b words.
194 #define PSHUFB_CST(A,B,C,D,E,F,G,H) \
195   _mm_set_epi8(2 * (H) + 1, 2 * (H) + 0, 2 * (G) + 1, 2 * (G) + 0, \
196                2 * (F) + 1, 2 * (F) + 0, 2 * (E) + 1, 2 * (E) + 0, \
197                2 * (D) + 1, 2 * (D) + 0, 2 * (C) + 1, 2 * (C) + 0, \
198                2 * (B) + 1, 2 * (B) + 0, 2 * (A) + 1, 2 * (A) + 0)
199 
DoQuantizeBlock_SSE41(int16_t in[16],int16_t out[16],const uint16_t * const sharpen,const VP8Matrix * const mtx)200 static WEBP_INLINE int DoQuantizeBlock_SSE41(int16_t in[16], int16_t out[16],
201                                              const uint16_t* const sharpen,
202                                              const VP8Matrix* const mtx) {
203   const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
204   const __m128i zero = _mm_setzero_si128();
205   __m128i out0, out8;
206   __m128i packed_out;
207 
208   // Load all inputs.
209   __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
210   __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
211   const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
212   const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
213   const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
214   const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
215 
216   // coeff = abs(in)
217   __m128i coeff0 = _mm_abs_epi16(in0);
218   __m128i coeff8 = _mm_abs_epi16(in8);
219 
220   // coeff = abs(in) + sharpen
221   if (sharpen != NULL) {
222     const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
223     const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
224     coeff0 = _mm_add_epi16(coeff0, sharpen0);
225     coeff8 = _mm_add_epi16(coeff8, sharpen8);
226   }
227 
228   // out = (coeff * iQ + B) >> QFIX
229   {
230     // doing calculations with 32b precision (QFIX=17)
231     // out = (coeff * iQ)
232     const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
233     const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
234     const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
235     const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
236     __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
237     __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
238     __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
239     __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
240     // out = (coeff * iQ + B)
241     const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
242     const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
243     const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
244     const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
245     out_00 = _mm_add_epi32(out_00, bias_00);
246     out_04 = _mm_add_epi32(out_04, bias_04);
247     out_08 = _mm_add_epi32(out_08, bias_08);
248     out_12 = _mm_add_epi32(out_12, bias_12);
249     // out = QUANTDIV(coeff, iQ, B, QFIX)
250     out_00 = _mm_srai_epi32(out_00, QFIX);
251     out_04 = _mm_srai_epi32(out_04, QFIX);
252     out_08 = _mm_srai_epi32(out_08, QFIX);
253     out_12 = _mm_srai_epi32(out_12, QFIX);
254 
255     // pack result as 16b
256     out0 = _mm_packs_epi32(out_00, out_04);
257     out8 = _mm_packs_epi32(out_08, out_12);
258 
259     // if (coeff > 2047) coeff = 2047
260     out0 = _mm_min_epi16(out0, max_coeff_2047);
261     out8 = _mm_min_epi16(out8, max_coeff_2047);
262   }
263 
264   // put sign back
265   out0 = _mm_sign_epi16(out0, in0);
266   out8 = _mm_sign_epi16(out8, in8);
267 
268   // in = out * Q
269   in0 = _mm_mullo_epi16(out0, q0);
270   in8 = _mm_mullo_epi16(out8, q8);
271 
272   _mm_storeu_si128((__m128i*)&in[0], in0);
273   _mm_storeu_si128((__m128i*)&in[8], in8);
274 
275   // zigzag the output before storing it. The re-ordering is:
276   //    0 1 2 3 4 5 6 7 | 8  9 10 11 12 13 14 15
277   // -> 0 1 4[8]5 2 3 6 | 9 12 13 10 [7]11 14 15
278   // There's only two misplaced entries ([8] and [7]) that are crossing the
279   // reg's boundaries.
280   // We use pshufb instead of pshuflo/pshufhi.
281   {
282     const __m128i kCst_lo = PSHUFB_CST(0, 1, 4, -1, 5, 2, 3, 6);
283     const __m128i kCst_7 = PSHUFB_CST(-1, -1, -1, -1, 7, -1, -1, -1);
284     const __m128i tmp_lo = _mm_shuffle_epi8(out0, kCst_lo);
285     const __m128i tmp_7 = _mm_shuffle_epi8(out0, kCst_7);  // extract #7
286     const __m128i kCst_hi = PSHUFB_CST(1, 4, 5, 2, -1, 3, 6, 7);
287     const __m128i kCst_8 = PSHUFB_CST(-1, -1, -1, 0, -1, -1, -1, -1);
288     const __m128i tmp_hi = _mm_shuffle_epi8(out8, kCst_hi);
289     const __m128i tmp_8 = _mm_shuffle_epi8(out8, kCst_8);  // extract #8
290     const __m128i out_z0 = _mm_or_si128(tmp_lo, tmp_8);
291     const __m128i out_z8 = _mm_or_si128(tmp_hi, tmp_7);
292     _mm_storeu_si128((__m128i*)&out[0], out_z0);
293     _mm_storeu_si128((__m128i*)&out[8], out_z8);
294     packed_out = _mm_packs_epi16(out_z0, out_z8);
295   }
296 
297   // detect if all 'out' values are zeroes or not
298   return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
299 }
300 
301 #undef PSHUFB_CST
302 
QuantizeBlock_SSE41(int16_t in[16],int16_t out[16],const VP8Matrix * const mtx)303 static int QuantizeBlock_SSE41(int16_t in[16], int16_t out[16],
304                                const VP8Matrix* const mtx) {
305   return DoQuantizeBlock_SSE41(in, out, &mtx->sharpen_[0], mtx);
306 }
307 
QuantizeBlockWHT_SSE41(int16_t in[16],int16_t out[16],const VP8Matrix * const mtx)308 static int QuantizeBlockWHT_SSE41(int16_t in[16], int16_t out[16],
309                                   const VP8Matrix* const mtx) {
310   return DoQuantizeBlock_SSE41(in, out, NULL, mtx);
311 }
312 
Quantize2Blocks_SSE41(int16_t in[32],int16_t out[32],const VP8Matrix * const mtx)313 static int Quantize2Blocks_SSE41(int16_t in[32], int16_t out[32],
314                                  const VP8Matrix* const mtx) {
315   int nz;
316   const uint16_t* const sharpen = &mtx->sharpen_[0];
317   nz  = DoQuantizeBlock_SSE41(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0;
318   nz |= DoQuantizeBlock_SSE41(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1;
319   return nz;
320 }
321 
322 //------------------------------------------------------------------------------
323 // Entry point
324 
325 extern void VP8EncDspInitSSE41(void);
VP8EncDspInitSSE41(void)326 WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE41(void) {
327   VP8CollectHistogram = CollectHistogram_SSE41;
328   VP8EncQuantizeBlock = QuantizeBlock_SSE41;
329   VP8EncQuantize2Blocks = Quantize2Blocks_SSE41;
330   VP8EncQuantizeBlockWHT = QuantizeBlockWHT_SSE41;
331   VP8TDisto4x4 = Disto4x4_SSE41;
332   VP8TDisto16x16 = Disto16x16_SSE41;
333 }
334 
335 #else  // !WEBP_USE_SSE41
336 
337 WEBP_DSP_INIT_STUB(VP8EncDspInitSSE41)
338 
339 #endif  // WEBP_USE_SSE41
340