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 // SSE2 variant of alpha filters
11 //
12 // Author: Skal (pascal.massimino@gmail.com)
13
14 #include "src/dsp/dsp.h"
15
16 #if defined(WEBP_USE_SSE2)
17
18 #include <assert.h>
19 #include <emmintrin.h>
20 #include <stdlib.h>
21 #include <string.h>
22
23 //------------------------------------------------------------------------------
24 // Helpful macro.
25
26 #define DCHECK(in, out) \
27 do { \
28 assert((in) != NULL); \
29 assert((out) != NULL); \
30 assert((in) != (out)); \
31 assert(width > 0); \
32 assert(height > 0); \
33 assert(stride >= width); \
34 } while (0)
35
PredictLineTop_SSE2(const uint8_t * WEBP_RESTRICT src,const uint8_t * WEBP_RESTRICT pred,uint8_t * WEBP_RESTRICT dst,int length)36 static void PredictLineTop_SSE2(const uint8_t* WEBP_RESTRICT src,
37 const uint8_t* WEBP_RESTRICT pred,
38 uint8_t* WEBP_RESTRICT dst, int length) {
39 int i;
40 const int max_pos = length & ~31;
41 assert(length >= 0);
42 for (i = 0; i < max_pos; i += 32) {
43 const __m128i A0 = _mm_loadu_si128((const __m128i*)&src[i + 0]);
44 const __m128i A1 = _mm_loadu_si128((const __m128i*)&src[i + 16]);
45 const __m128i B0 = _mm_loadu_si128((const __m128i*)&pred[i + 0]);
46 const __m128i B1 = _mm_loadu_si128((const __m128i*)&pred[i + 16]);
47 const __m128i C0 = _mm_sub_epi8(A0, B0);
48 const __m128i C1 = _mm_sub_epi8(A1, B1);
49 _mm_storeu_si128((__m128i*)&dst[i + 0], C0);
50 _mm_storeu_si128((__m128i*)&dst[i + 16], C1);
51 }
52 for (; i < length; ++i) dst[i] = src[i] - pred[i];
53 }
54
55 // Special case for left-based prediction (when preds==dst-1 or preds==src-1).
PredictLineLeft_SSE2(const uint8_t * WEBP_RESTRICT src,uint8_t * WEBP_RESTRICT dst,int length)56 static void PredictLineLeft_SSE2(const uint8_t* WEBP_RESTRICT src,
57 uint8_t* WEBP_RESTRICT dst, int length) {
58 int i;
59 const int max_pos = length & ~31;
60 assert(length >= 0);
61 for (i = 0; i < max_pos; i += 32) {
62 const __m128i A0 = _mm_loadu_si128((const __m128i*)(src + i + 0 ));
63 const __m128i B0 = _mm_loadu_si128((const __m128i*)(src + i + 0 - 1));
64 const __m128i A1 = _mm_loadu_si128((const __m128i*)(src + i + 16 ));
65 const __m128i B1 = _mm_loadu_si128((const __m128i*)(src + i + 16 - 1));
66 const __m128i C0 = _mm_sub_epi8(A0, B0);
67 const __m128i C1 = _mm_sub_epi8(A1, B1);
68 _mm_storeu_si128((__m128i*)(dst + i + 0), C0);
69 _mm_storeu_si128((__m128i*)(dst + i + 16), C1);
70 }
71 for (; i < length; ++i) dst[i] = src[i] - src[i - 1];
72 }
73
74 //------------------------------------------------------------------------------
75 // Horizontal filter.
76
DoHorizontalFilter_SSE2(const uint8_t * WEBP_RESTRICT in,int width,int height,int stride,uint8_t * WEBP_RESTRICT out)77 static WEBP_INLINE void DoHorizontalFilter_SSE2(
78 const uint8_t* WEBP_RESTRICT in, int width, int height, int stride,
79 uint8_t* WEBP_RESTRICT out) {
80 int row;
81 DCHECK(in, out);
82
83 // Leftmost pixel is the same as input for topmost scanline.
84 out[0] = in[0];
85 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
86 in += stride;
87 out += stride;
88
89 // Filter line-by-line.
90 for (row = 1; row < height; ++row) {
91 // Leftmost pixel is predicted from above.
92 out[0] = in[0] - in[-stride];
93 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
94 in += stride;
95 out += stride;
96 }
97 }
98
99 //------------------------------------------------------------------------------
100 // Vertical filter.
101
DoVerticalFilter_SSE2(const uint8_t * WEBP_RESTRICT in,int width,int height,int stride,uint8_t * WEBP_RESTRICT out)102 static WEBP_INLINE void DoVerticalFilter_SSE2(const uint8_t* WEBP_RESTRICT in,
103 int width, int height, int stride,
104 uint8_t* WEBP_RESTRICT out) {
105 int row;
106 DCHECK(in, out);
107
108 // Very first top-left pixel is copied.
109 out[0] = in[0];
110 // Rest of top scan-line is left-predicted.
111 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
112 in += stride;
113 out += stride;
114
115 // Filter line-by-line.
116 for (row = 1; row < height; ++row) {
117 PredictLineTop_SSE2(in, in - stride, out, width);
118 in += stride;
119 out += stride;
120 }
121 }
122
123 //------------------------------------------------------------------------------
124 // Gradient filter.
125
GradientPredictor_SSE2(uint8_t a,uint8_t b,uint8_t c)126 static WEBP_INLINE int GradientPredictor_SSE2(uint8_t a, uint8_t b, uint8_t c) {
127 const int g = a + b - c;
128 return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255; // clip to 8bit
129 }
130
GradientPredictDirect_SSE2(const uint8_t * const row,const uint8_t * const top,uint8_t * WEBP_RESTRICT const out,int length)131 static void GradientPredictDirect_SSE2(const uint8_t* const row,
132 const uint8_t* const top,
133 uint8_t* WEBP_RESTRICT const out,
134 int length) {
135 const int max_pos = length & ~7;
136 int i;
137 const __m128i zero = _mm_setzero_si128();
138 for (i = 0; i < max_pos; i += 8) {
139 const __m128i A0 = _mm_loadl_epi64((const __m128i*)&row[i - 1]);
140 const __m128i B0 = _mm_loadl_epi64((const __m128i*)&top[i]);
141 const __m128i C0 = _mm_loadl_epi64((const __m128i*)&top[i - 1]);
142 const __m128i D = _mm_loadl_epi64((const __m128i*)&row[i]);
143 const __m128i A1 = _mm_unpacklo_epi8(A0, zero);
144 const __m128i B1 = _mm_unpacklo_epi8(B0, zero);
145 const __m128i C1 = _mm_unpacklo_epi8(C0, zero);
146 const __m128i E = _mm_add_epi16(A1, B1);
147 const __m128i F = _mm_sub_epi16(E, C1);
148 const __m128i G = _mm_packus_epi16(F, zero);
149 const __m128i H = _mm_sub_epi8(D, G);
150 _mm_storel_epi64((__m128i*)(out + i), H);
151 }
152 for (; i < length; ++i) {
153 const int delta = GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]);
154 out[i] = (uint8_t)(row[i] - delta);
155 }
156 }
157
DoGradientFilter_SSE2(const uint8_t * WEBP_RESTRICT in,int width,int height,int stride,uint8_t * WEBP_RESTRICT out)158 static WEBP_INLINE void DoGradientFilter_SSE2(const uint8_t* WEBP_RESTRICT in,
159 int width, int height, int stride,
160 uint8_t* WEBP_RESTRICT out) {
161 int row;
162 DCHECK(in, out);
163
164 // left prediction for top scan-line
165 out[0] = in[0];
166 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
167 in += stride;
168 out += stride;
169
170 // Filter line-by-line.
171 for (row = 1; row < height; ++row) {
172 out[0] = (uint8_t)(in[0] - in[-stride]);
173 GradientPredictDirect_SSE2(in + 1, in + 1 - stride, out + 1, width - 1);
174 in += stride;
175 out += stride;
176 }
177 }
178
179 #undef DCHECK
180
181 //------------------------------------------------------------------------------
182
HorizontalFilter_SSE2(const uint8_t * WEBP_RESTRICT data,int width,int height,int stride,uint8_t * WEBP_RESTRICT filtered_data)183 static void HorizontalFilter_SSE2(const uint8_t* WEBP_RESTRICT data,
184 int width, int height, int stride,
185 uint8_t* WEBP_RESTRICT filtered_data) {
186 DoHorizontalFilter_SSE2(data, width, height, stride, filtered_data);
187 }
188
VerticalFilter_SSE2(const uint8_t * WEBP_RESTRICT data,int width,int height,int stride,uint8_t * WEBP_RESTRICT filtered_data)189 static void VerticalFilter_SSE2(const uint8_t* WEBP_RESTRICT data,
190 int width, int height, int stride,
191 uint8_t* WEBP_RESTRICT filtered_data) {
192 DoVerticalFilter_SSE2(data, width, height, stride, filtered_data);
193 }
194
GradientFilter_SSE2(const uint8_t * WEBP_RESTRICT data,int width,int height,int stride,uint8_t * WEBP_RESTRICT filtered_data)195 static void GradientFilter_SSE2(const uint8_t* WEBP_RESTRICT data,
196 int width, int height, int stride,
197 uint8_t* WEBP_RESTRICT filtered_data) {
198 DoGradientFilter_SSE2(data, width, height, stride, filtered_data);
199 }
200
201 //------------------------------------------------------------------------------
202 // Inverse transforms
203
HorizontalUnfilter_SSE2(const uint8_t * prev,const uint8_t * in,uint8_t * out,int width)204 static void HorizontalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in,
205 uint8_t* out, int width) {
206 int i;
207 __m128i last;
208 out[0] = (uint8_t)(in[0] + (prev == NULL ? 0 : prev[0]));
209 if (width <= 1) return;
210 last = _mm_set_epi32(0, 0, 0, out[0]);
211 for (i = 1; i + 8 <= width; i += 8) {
212 const __m128i A0 = _mm_loadl_epi64((const __m128i*)(in + i));
213 const __m128i A1 = _mm_add_epi8(A0, last);
214 const __m128i A2 = _mm_slli_si128(A1, 1);
215 const __m128i A3 = _mm_add_epi8(A1, A2);
216 const __m128i A4 = _mm_slli_si128(A3, 2);
217 const __m128i A5 = _mm_add_epi8(A3, A4);
218 const __m128i A6 = _mm_slli_si128(A5, 4);
219 const __m128i A7 = _mm_add_epi8(A5, A6);
220 _mm_storel_epi64((__m128i*)(out + i), A7);
221 last = _mm_srli_epi64(A7, 56);
222 }
223 for (; i < width; ++i) out[i] = (uint8_t)(in[i] + out[i - 1]);
224 }
225
VerticalUnfilter_SSE2(const uint8_t * prev,const uint8_t * in,uint8_t * out,int width)226 static void VerticalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in,
227 uint8_t* out, int width) {
228 if (prev == NULL) {
229 HorizontalUnfilter_SSE2(NULL, in, out, width);
230 } else {
231 int i;
232 const int max_pos = width & ~31;
233 assert(width >= 0);
234 for (i = 0; i < max_pos; i += 32) {
235 const __m128i A0 = _mm_loadu_si128((const __m128i*)&in[i + 0]);
236 const __m128i A1 = _mm_loadu_si128((const __m128i*)&in[i + 16]);
237 const __m128i B0 = _mm_loadu_si128((const __m128i*)&prev[i + 0]);
238 const __m128i B1 = _mm_loadu_si128((const __m128i*)&prev[i + 16]);
239 const __m128i C0 = _mm_add_epi8(A0, B0);
240 const __m128i C1 = _mm_add_epi8(A1, B1);
241 _mm_storeu_si128((__m128i*)&out[i + 0], C0);
242 _mm_storeu_si128((__m128i*)&out[i + 16], C1);
243 }
244 for (; i < width; ++i) out[i] = (uint8_t)(in[i] + prev[i]);
245 }
246 }
247
GradientPredictInverse_SSE2(const uint8_t * const in,const uint8_t * const top,uint8_t * const row,int length)248 static void GradientPredictInverse_SSE2(const uint8_t* const in,
249 const uint8_t* const top,
250 uint8_t* const row, int length) {
251 if (length > 0) {
252 int i;
253 const int max_pos = length & ~7;
254 const __m128i zero = _mm_setzero_si128();
255 __m128i A = _mm_set_epi32(0, 0, 0, row[-1]); // left sample
256 for (i = 0; i < max_pos; i += 8) {
257 const __m128i tmp0 = _mm_loadl_epi64((const __m128i*)&top[i]);
258 const __m128i tmp1 = _mm_loadl_epi64((const __m128i*)&top[i - 1]);
259 const __m128i B = _mm_unpacklo_epi8(tmp0, zero);
260 const __m128i C = _mm_unpacklo_epi8(tmp1, zero);
261 const __m128i D = _mm_loadl_epi64((const __m128i*)&in[i]); // base input
262 const __m128i E = _mm_sub_epi16(B, C); // unclipped gradient basis B - C
263 __m128i out = zero; // accumulator for output
264 __m128i mask_hi = _mm_set_epi32(0, 0, 0, 0xff);
265 int k = 8;
266 while (1) {
267 const __m128i tmp3 = _mm_add_epi16(A, E); // delta = A + B - C
268 const __m128i tmp4 = _mm_packus_epi16(tmp3, zero); // saturate delta
269 const __m128i tmp5 = _mm_add_epi8(tmp4, D); // add to in[]
270 A = _mm_and_si128(tmp5, mask_hi); // 1-complement clip
271 out = _mm_or_si128(out, A); // accumulate output
272 if (--k == 0) break;
273 A = _mm_slli_si128(A, 1); // rotate left sample
274 mask_hi = _mm_slli_si128(mask_hi, 1); // rotate mask
275 A = _mm_unpacklo_epi8(A, zero); // convert 8b->16b
276 }
277 A = _mm_srli_si128(A, 7); // prepare left sample for next iteration
278 _mm_storel_epi64((__m128i*)&row[i], out);
279 }
280 for (; i < length; ++i) {
281 const int delta = GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]);
282 row[i] = (uint8_t)(in[i] + delta);
283 }
284 }
285 }
286
GradientUnfilter_SSE2(const uint8_t * prev,const uint8_t * in,uint8_t * out,int width)287 static void GradientUnfilter_SSE2(const uint8_t* prev, const uint8_t* in,
288 uint8_t* out, int width) {
289 if (prev == NULL) {
290 HorizontalUnfilter_SSE2(NULL, in, out, width);
291 } else {
292 out[0] = (uint8_t)(in[0] + prev[0]); // predict from above
293 GradientPredictInverse_SSE2(in + 1, prev + 1, out + 1, width - 1);
294 }
295 }
296
297 //------------------------------------------------------------------------------
298 // Entry point
299
300 extern void VP8FiltersInitSSE2(void);
301
VP8FiltersInitSSE2(void)302 WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitSSE2(void) {
303 WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_SSE2;
304 #if defined(CHROMIUM)
305 // TODO(crbug.com/654974)
306 (void)VerticalUnfilter_SSE2;
307 #else
308 WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_SSE2;
309 #endif
310 WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_SSE2;
311
312 WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_SSE2;
313 WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter_SSE2;
314 WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter_SSE2;
315 }
316
317 #else // !WEBP_USE_SSE2
318
319 WEBP_DSP_INIT_STUB(VP8FiltersInitSSE2)
320
321 #endif // WEBP_USE_SSE2
322