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(width > 0); \
31 assert(height > 0); \
32 assert(stride >= width); \
33 assert(row >= 0 && num_rows > 0 && row + num_rows <= height); \
34 (void)height; /* Silence unused warning. */ \
35 } while (0)
36
PredictLineTop_SSE2(const uint8_t * src,const uint8_t * pred,uint8_t * dst,int length)37 static void PredictLineTop_SSE2(const uint8_t* src, const uint8_t* pred,
38 uint8_t* 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 * src,uint8_t * dst,int length)56 static void PredictLineLeft_SSE2(const uint8_t* src, uint8_t* dst, int length) {
57 int i;
58 const int max_pos = length & ~31;
59 assert(length >= 0);
60 for (i = 0; i < max_pos; i += 32) {
61 const __m128i A0 = _mm_loadu_si128((const __m128i*)(src + i + 0 ));
62 const __m128i B0 = _mm_loadu_si128((const __m128i*)(src + i + 0 - 1));
63 const __m128i A1 = _mm_loadu_si128((const __m128i*)(src + i + 16 ));
64 const __m128i B1 = _mm_loadu_si128((const __m128i*)(src + i + 16 - 1));
65 const __m128i C0 = _mm_sub_epi8(A0, B0);
66 const __m128i C1 = _mm_sub_epi8(A1, B1);
67 _mm_storeu_si128((__m128i*)(dst + i + 0), C0);
68 _mm_storeu_si128((__m128i*)(dst + i + 16), C1);
69 }
70 for (; i < length; ++i) dst[i] = src[i] - src[i - 1];
71 }
72
73 //------------------------------------------------------------------------------
74 // Horizontal filter.
75
DoHorizontalFilter_SSE2(const uint8_t * in,int width,int height,int stride,int row,int num_rows,uint8_t * out)76 static WEBP_INLINE void DoHorizontalFilter_SSE2(const uint8_t* in,
77 int width, int height,
78 int stride,
79 int row, int num_rows,
80 uint8_t* out) {
81 const size_t start_offset = row * stride;
82 const int last_row = row + num_rows;
83 DCHECK(in, out);
84 in += start_offset;
85 out += start_offset;
86
87 if (row == 0) {
88 // Leftmost pixel is the same as input for topmost scanline.
89 out[0] = in[0];
90 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
91 row = 1;
92 in += stride;
93 out += stride;
94 }
95
96 // Filter line-by-line.
97 while (row < last_row) {
98 // Leftmost pixel is predicted from above.
99 out[0] = in[0] - in[-stride];
100 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
101 ++row;
102 in += stride;
103 out += stride;
104 }
105 }
106
107 //------------------------------------------------------------------------------
108 // Vertical filter.
109
DoVerticalFilter_SSE2(const uint8_t * in,int width,int height,int stride,int row,int num_rows,uint8_t * out)110 static WEBP_INLINE void DoVerticalFilter_SSE2(const uint8_t* in,
111 int width, int height, int stride,
112 int row, int num_rows,
113 uint8_t* out) {
114 const size_t start_offset = row * stride;
115 const int last_row = row + num_rows;
116 DCHECK(in, out);
117 in += start_offset;
118 out += start_offset;
119
120 if (row == 0) {
121 // Very first top-left pixel is copied.
122 out[0] = in[0];
123 // Rest of top scan-line is left-predicted.
124 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
125 row = 1;
126 in += stride;
127 out += stride;
128 }
129
130 // Filter line-by-line.
131 while (row < last_row) {
132 PredictLineTop_SSE2(in, in - stride, out, width);
133 ++row;
134 in += stride;
135 out += stride;
136 }
137 }
138
139 //------------------------------------------------------------------------------
140 // Gradient filter.
141
GradientPredictor_SSE2(uint8_t a,uint8_t b,uint8_t c)142 static WEBP_INLINE int GradientPredictor_SSE2(uint8_t a, uint8_t b, uint8_t c) {
143 const int g = a + b - c;
144 return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255; // clip to 8bit
145 }
146
GradientPredictDirect_SSE2(const uint8_t * const row,const uint8_t * const top,uint8_t * const out,int length)147 static void GradientPredictDirect_SSE2(const uint8_t* const row,
148 const uint8_t* const top,
149 uint8_t* const out, int length) {
150 const int max_pos = length & ~7;
151 int i;
152 const __m128i zero = _mm_setzero_si128();
153 for (i = 0; i < max_pos; i += 8) {
154 const __m128i A0 = _mm_loadl_epi64((const __m128i*)&row[i - 1]);
155 const __m128i B0 = _mm_loadl_epi64((const __m128i*)&top[i]);
156 const __m128i C0 = _mm_loadl_epi64((const __m128i*)&top[i - 1]);
157 const __m128i D = _mm_loadl_epi64((const __m128i*)&row[i]);
158 const __m128i A1 = _mm_unpacklo_epi8(A0, zero);
159 const __m128i B1 = _mm_unpacklo_epi8(B0, zero);
160 const __m128i C1 = _mm_unpacklo_epi8(C0, zero);
161 const __m128i E = _mm_add_epi16(A1, B1);
162 const __m128i F = _mm_sub_epi16(E, C1);
163 const __m128i G = _mm_packus_epi16(F, zero);
164 const __m128i H = _mm_sub_epi8(D, G);
165 _mm_storel_epi64((__m128i*)(out + i), H);
166 }
167 for (; i < length; ++i) {
168 const int delta = GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]);
169 out[i] = (uint8_t)(row[i] - delta);
170 }
171 }
172
DoGradientFilter_SSE2(const uint8_t * in,int width,int height,int stride,int row,int num_rows,uint8_t * out)173 static WEBP_INLINE void DoGradientFilter_SSE2(const uint8_t* in,
174 int width, int height, int stride,
175 int row, int num_rows,
176 uint8_t* out) {
177 const size_t start_offset = row * stride;
178 const int last_row = row + num_rows;
179 DCHECK(in, out);
180 in += start_offset;
181 out += start_offset;
182
183 // left prediction for top scan-line
184 if (row == 0) {
185 out[0] = in[0];
186 PredictLineLeft_SSE2(in + 1, out + 1, width - 1);
187 row = 1;
188 in += stride;
189 out += stride;
190 }
191
192 // Filter line-by-line.
193 while (row < last_row) {
194 out[0] = (uint8_t)(in[0] - in[-stride]);
195 GradientPredictDirect_SSE2(in + 1, in + 1 - stride, out + 1, width - 1);
196 ++row;
197 in += stride;
198 out += stride;
199 }
200 }
201
202 #undef DCHECK
203
204 //------------------------------------------------------------------------------
205
HorizontalFilter_SSE2(const uint8_t * data,int width,int height,int stride,uint8_t * filtered_data)206 static void HorizontalFilter_SSE2(const uint8_t* data, int width, int height,
207 int stride, uint8_t* filtered_data) {
208 DoHorizontalFilter_SSE2(data, width, height, stride, 0, height,
209 filtered_data);
210 }
211
VerticalFilter_SSE2(const uint8_t * data,int width,int height,int stride,uint8_t * filtered_data)212 static void VerticalFilter_SSE2(const uint8_t* data, int width, int height,
213 int stride, uint8_t* filtered_data) {
214 DoVerticalFilter_SSE2(data, width, height, stride, 0, height, filtered_data);
215 }
216
GradientFilter_SSE2(const uint8_t * data,int width,int height,int stride,uint8_t * filtered_data)217 static void GradientFilter_SSE2(const uint8_t* data, int width, int height,
218 int stride, uint8_t* filtered_data) {
219 DoGradientFilter_SSE2(data, width, height, stride, 0, height, filtered_data);
220 }
221
222 //------------------------------------------------------------------------------
223 // Inverse transforms
224
HorizontalUnfilter_SSE2(const uint8_t * prev,const uint8_t * in,uint8_t * out,int width)225 static void HorizontalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in,
226 uint8_t* out, int width) {
227 int i;
228 __m128i last;
229 out[0] = (uint8_t)(in[0] + (prev == NULL ? 0 : prev[0]));
230 if (width <= 1) return;
231 last = _mm_set_epi32(0, 0, 0, out[0]);
232 for (i = 1; i + 8 <= width; i += 8) {
233 const __m128i A0 = _mm_loadl_epi64((const __m128i*)(in + i));
234 const __m128i A1 = _mm_add_epi8(A0, last);
235 const __m128i A2 = _mm_slli_si128(A1, 1);
236 const __m128i A3 = _mm_add_epi8(A1, A2);
237 const __m128i A4 = _mm_slli_si128(A3, 2);
238 const __m128i A5 = _mm_add_epi8(A3, A4);
239 const __m128i A6 = _mm_slli_si128(A5, 4);
240 const __m128i A7 = _mm_add_epi8(A5, A6);
241 _mm_storel_epi64((__m128i*)(out + i), A7);
242 last = _mm_srli_epi64(A7, 56);
243 }
244 for (; i < width; ++i) out[i] = (uint8_t)(in[i] + out[i - 1]);
245 }
246
VerticalUnfilter_SSE2(const uint8_t * prev,const uint8_t * in,uint8_t * out,int width)247 static void VerticalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in,
248 uint8_t* out, int width) {
249 if (prev == NULL) {
250 HorizontalUnfilter_SSE2(NULL, in, out, width);
251 } else {
252 int i;
253 const int max_pos = width & ~31;
254 assert(width >= 0);
255 for (i = 0; i < max_pos; i += 32) {
256 const __m128i A0 = _mm_loadu_si128((const __m128i*)&in[i + 0]);
257 const __m128i A1 = _mm_loadu_si128((const __m128i*)&in[i + 16]);
258 const __m128i B0 = _mm_loadu_si128((const __m128i*)&prev[i + 0]);
259 const __m128i B1 = _mm_loadu_si128((const __m128i*)&prev[i + 16]);
260 const __m128i C0 = _mm_add_epi8(A0, B0);
261 const __m128i C1 = _mm_add_epi8(A1, B1);
262 _mm_storeu_si128((__m128i*)&out[i + 0], C0);
263 _mm_storeu_si128((__m128i*)&out[i + 16], C1);
264 }
265 for (; i < width; ++i) out[i] = (uint8_t)(in[i] + prev[i]);
266 }
267 }
268
GradientPredictInverse_SSE2(const uint8_t * const in,const uint8_t * const top,uint8_t * const row,int length)269 static void GradientPredictInverse_SSE2(const uint8_t* const in,
270 const uint8_t* const top,
271 uint8_t* const row, int length) {
272 if (length > 0) {
273 int i;
274 const int max_pos = length & ~7;
275 const __m128i zero = _mm_setzero_si128();
276 __m128i A = _mm_set_epi32(0, 0, 0, row[-1]); // left sample
277 for (i = 0; i < max_pos; i += 8) {
278 const __m128i tmp0 = _mm_loadl_epi64((const __m128i*)&top[i]);
279 const __m128i tmp1 = _mm_loadl_epi64((const __m128i*)&top[i - 1]);
280 const __m128i B = _mm_unpacklo_epi8(tmp0, zero);
281 const __m128i C = _mm_unpacklo_epi8(tmp1, zero);
282 const __m128i D = _mm_loadl_epi64((const __m128i*)&in[i]); // base input
283 const __m128i E = _mm_sub_epi16(B, C); // unclipped gradient basis B - C
284 __m128i out = zero; // accumulator for output
285 __m128i mask_hi = _mm_set_epi32(0, 0, 0, 0xff);
286 int k = 8;
287 while (1) {
288 const __m128i tmp3 = _mm_add_epi16(A, E); // delta = A + B - C
289 const __m128i tmp4 = _mm_packus_epi16(tmp3, zero); // saturate delta
290 const __m128i tmp5 = _mm_add_epi8(tmp4, D); // add to in[]
291 A = _mm_and_si128(tmp5, mask_hi); // 1-complement clip
292 out = _mm_or_si128(out, A); // accumulate output
293 if (--k == 0) break;
294 A = _mm_slli_si128(A, 1); // rotate left sample
295 mask_hi = _mm_slli_si128(mask_hi, 1); // rotate mask
296 A = _mm_unpacklo_epi8(A, zero); // convert 8b->16b
297 }
298 A = _mm_srli_si128(A, 7); // prepare left sample for next iteration
299 _mm_storel_epi64((__m128i*)&row[i], out);
300 }
301 for (; i < length; ++i) {
302 const int delta = GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]);
303 row[i] = (uint8_t)(in[i] + delta);
304 }
305 }
306 }
307
GradientUnfilter_SSE2(const uint8_t * prev,const uint8_t * in,uint8_t * out,int width)308 static void GradientUnfilter_SSE2(const uint8_t* prev, const uint8_t* in,
309 uint8_t* out, int width) {
310 if (prev == NULL) {
311 HorizontalUnfilter_SSE2(NULL, in, out, width);
312 } else {
313 out[0] = (uint8_t)(in[0] + prev[0]); // predict from above
314 GradientPredictInverse_SSE2(in + 1, prev + 1, out + 1, width - 1);
315 }
316 }
317
318 //------------------------------------------------------------------------------
319 // Entry point
320
321 extern void VP8FiltersInitSSE2(void);
322
VP8FiltersInitSSE2(void)323 WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitSSE2(void) {
324 WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_SSE2;
325 #if defined(CHROMIUM)
326 // TODO(crbug.com/654974)
327 (void)VerticalUnfilter_SSE2;
328 #else
329 WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_SSE2;
330 #endif
331 WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_SSE2;
332
333 WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_SSE2;
334 WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter_SSE2;
335 WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter_SSE2;
336 }
337
338 #else // !WEBP_USE_SSE2
339
340 WEBP_DSP_INIT_STUB(VP8FiltersInitSSE2)
341
342 #endif // WEBP_USE_SSE2
343