1 | |
2 | /* filter_sse2_intrinsics.c - SSE2 optimized filter functions |
3 | * |
4 | * Copyright (c) 2018 Cosmin Truta |
5 | * Copyright (c) 2016-2017 Glenn Randers-Pehrson |
6 | * Written by Mike Klein and Matt Sarett |
7 | * Derived from arm/filter_neon_intrinsics.c |
8 | * |
9 | * This code is released under the libpng license. |
10 | * For conditions of distribution and use, see the disclaimer |
11 | * and license in png.h |
12 | */ |
13 | |
14 | #include "../pngpriv.h" |
15 | |
16 | #ifdef PNG_READ_SUPPORTED |
17 | |
18 | #if PNG_INTEL_SSE_IMPLEMENTATION > 0 |
19 | |
20 | #include <immintrin.h> |
21 | |
22 | /* Functions in this file look at most 3 pixels (a,b,c) to predict the 4th (d). |
23 | * They're positioned like this: |
24 | * prev: c b |
25 | * row: a d |
26 | * The Sub filter predicts d=a, Avg d=(a+b)/2, and Paeth predicts d to be |
27 | * whichever of a, b, or c is closest to p=a+b-c. |
28 | */ |
29 | |
30 | static __m128i load4(const void* p) { |
31 | int tmp; |
32 | memcpy(&tmp, p, sizeof(tmp)); |
33 | return _mm_cvtsi32_si128(tmp); |
34 | } |
35 | |
36 | static void store4(void* p, __m128i v) { |
37 | int tmp = _mm_cvtsi128_si32(v); |
38 | memcpy(p, &tmp, sizeof(int)); |
39 | } |
40 | |
41 | static __m128i load3(const void* p) { |
42 | png_uint_32 tmp = 0; |
43 | memcpy(&tmp, p, 3); |
44 | return _mm_cvtsi32_si128(tmp); |
45 | } |
46 | |
47 | static void store3(void* p, __m128i v) { |
48 | int tmp = _mm_cvtsi128_si32(v); |
49 | memcpy(p, &tmp, 3); |
50 | } |
51 | |
52 | void png_read_filter_row_sub3_sse2(png_row_infop row_info, png_bytep row, |
53 | png_const_bytep prev) |
54 | { |
55 | /* The Sub filter predicts each pixel as the previous pixel, a. |
56 | * There is no pixel to the left of the first pixel. It's encoded directly. |
57 | * That works with our main loop if we just say that left pixel was zero. |
58 | */ |
59 | size_t rb; |
60 | |
61 | __m128i a, d = _mm_setzero_si128(); |
62 | |
63 | png_debug(1, "in png_read_filter_row_sub3_sse2" ); |
64 | |
65 | rb = row_info->rowbytes; |
66 | while (rb >= 4) { |
67 | a = d; d = load4(row); |
68 | d = _mm_add_epi8(d, a); |
69 | store3(row, d); |
70 | |
71 | row += 3; |
72 | rb -= 3; |
73 | } |
74 | if (rb > 0) { |
75 | a = d; d = load3(row); |
76 | d = _mm_add_epi8(d, a); |
77 | store3(row, d); |
78 | |
79 | row += 3; |
80 | rb -= 3; |
81 | } |
82 | PNG_UNUSED(prev) |
83 | } |
84 | |
85 | void png_read_filter_row_sub4_sse2(png_row_infop row_info, png_bytep row, |
86 | png_const_bytep prev) |
87 | { |
88 | /* The Sub filter predicts each pixel as the previous pixel, a. |
89 | * There is no pixel to the left of the first pixel. It's encoded directly. |
90 | * That works with our main loop if we just say that left pixel was zero. |
91 | */ |
92 | size_t rb; |
93 | |
94 | __m128i a, d = _mm_setzero_si128(); |
95 | |
96 | png_debug(1, "in png_read_filter_row_sub4_sse2" ); |
97 | |
98 | rb = row_info->rowbytes+4; |
99 | while (rb > 4) { |
100 | a = d; d = load4(row); |
101 | d = _mm_add_epi8(d, a); |
102 | store4(row, d); |
103 | |
104 | row += 4; |
105 | rb -= 4; |
106 | } |
107 | PNG_UNUSED(prev) |
108 | } |
109 | |
110 | void png_read_filter_row_avg3_sse2(png_row_infop row_info, png_bytep row, |
111 | png_const_bytep prev) |
112 | { |
113 | /* The Avg filter predicts each pixel as the (truncated) average of a and b. |
114 | * There's no pixel to the left of the first pixel. Luckily, it's |
115 | * predicted to be half of the pixel above it. So again, this works |
116 | * perfectly with our loop if we make sure a starts at zero. |
117 | */ |
118 | |
119 | size_t rb; |
120 | |
121 | const __m128i zero = _mm_setzero_si128(); |
122 | |
123 | __m128i b; |
124 | __m128i a, d = zero; |
125 | |
126 | png_debug(1, "in png_read_filter_row_avg3_sse2" ); |
127 | rb = row_info->rowbytes; |
128 | while (rb >= 4) { |
129 | __m128i avg; |
130 | b = load4(prev); |
131 | a = d; d = load4(row ); |
132 | |
133 | /* PNG requires a truncating average, so we can't just use _mm_avg_epu8 */ |
134 | avg = _mm_avg_epu8(a,b); |
135 | /* ...but we can fix it up by subtracting off 1 if it rounded up. */ |
136 | avg = _mm_sub_epi8(avg, _mm_and_si128(_mm_xor_si128(a,b), |
137 | _mm_set1_epi8(1))); |
138 | d = _mm_add_epi8(d, avg); |
139 | store3(row, d); |
140 | |
141 | prev += 3; |
142 | row += 3; |
143 | rb -= 3; |
144 | } |
145 | if (rb > 0) { |
146 | __m128i avg; |
147 | b = load3(prev); |
148 | a = d; d = load3(row ); |
149 | |
150 | /* PNG requires a truncating average, so we can't just use _mm_avg_epu8 */ |
151 | avg = _mm_avg_epu8(a,b); |
152 | /* ...but we can fix it up by subtracting off 1 if it rounded up. */ |
153 | avg = _mm_sub_epi8(avg, _mm_and_si128(_mm_xor_si128(a,b), |
154 | _mm_set1_epi8(1))); |
155 | |
156 | d = _mm_add_epi8(d, avg); |
157 | store3(row, d); |
158 | |
159 | prev += 3; |
160 | row += 3; |
161 | rb -= 3; |
162 | } |
163 | } |
164 | |
165 | void png_read_filter_row_avg4_sse2(png_row_infop row_info, png_bytep row, |
166 | png_const_bytep prev) |
167 | { |
168 | /* The Avg filter predicts each pixel as the (truncated) average of a and b. |
169 | * There's no pixel to the left of the first pixel. Luckily, it's |
170 | * predicted to be half of the pixel above it. So again, this works |
171 | * perfectly with our loop if we make sure a starts at zero. |
172 | */ |
173 | size_t rb; |
174 | const __m128i zero = _mm_setzero_si128(); |
175 | __m128i b; |
176 | __m128i a, d = zero; |
177 | |
178 | png_debug(1, "in png_read_filter_row_avg4_sse2" ); |
179 | |
180 | rb = row_info->rowbytes+4; |
181 | while (rb > 4) { |
182 | __m128i avg; |
183 | b = load4(prev); |
184 | a = d; d = load4(row ); |
185 | |
186 | /* PNG requires a truncating average, so we can't just use _mm_avg_epu8 */ |
187 | avg = _mm_avg_epu8(a,b); |
188 | /* ...but we can fix it up by subtracting off 1 if it rounded up. */ |
189 | avg = _mm_sub_epi8(avg, _mm_and_si128(_mm_xor_si128(a,b), |
190 | _mm_set1_epi8(1))); |
191 | |
192 | d = _mm_add_epi8(d, avg); |
193 | store4(row, d); |
194 | |
195 | prev += 4; |
196 | row += 4; |
197 | rb -= 4; |
198 | } |
199 | } |
200 | |
201 | /* Returns |x| for 16-bit lanes. */ |
202 | static __m128i abs_i16(__m128i x) { |
203 | #if PNG_INTEL_SSE_IMPLEMENTATION >= 2 |
204 | return _mm_abs_epi16(x); |
205 | #else |
206 | /* Read this all as, return x<0 ? -x : x. |
207 | * To negate two's complement, you flip all the bits then add 1. |
208 | */ |
209 | __m128i is_negative = _mm_cmplt_epi16(x, _mm_setzero_si128()); |
210 | |
211 | /* Flip negative lanes. */ |
212 | x = _mm_xor_si128(x, is_negative); |
213 | |
214 | /* +1 to negative lanes, else +0. */ |
215 | x = _mm_sub_epi16(x, is_negative); |
216 | return x; |
217 | #endif |
218 | } |
219 | |
220 | /* Bytewise c ? t : e. */ |
221 | static __m128i if_then_else(__m128i c, __m128i t, __m128i e) { |
222 | #if PNG_INTEL_SSE_IMPLEMENTATION >= 3 |
223 | return _mm_blendv_epi8(e,t,c); |
224 | #else |
225 | return _mm_or_si128(_mm_and_si128(c, t), _mm_andnot_si128(c, e)); |
226 | #endif |
227 | } |
228 | |
229 | void png_read_filter_row_paeth3_sse2(png_row_infop row_info, png_bytep row, |
230 | png_const_bytep prev) |
231 | { |
232 | /* Paeth tries to predict pixel d using the pixel to the left of it, a, |
233 | * and two pixels from the previous row, b and c: |
234 | * prev: c b |
235 | * row: a d |
236 | * The Paeth function predicts d to be whichever of a, b, or c is nearest to |
237 | * p=a+b-c. |
238 | * |
239 | * The first pixel has no left context, and so uses an Up filter, p = b. |
240 | * This works naturally with our main loop's p = a+b-c if we force a and c |
241 | * to zero. |
242 | * Here we zero b and d, which become c and a respectively at the start of |
243 | * the loop. |
244 | */ |
245 | size_t rb; |
246 | const __m128i zero = _mm_setzero_si128(); |
247 | __m128i c, b = zero, |
248 | a, d = zero; |
249 | |
250 | png_debug(1, "in png_read_filter_row_paeth3_sse2" ); |
251 | |
252 | rb = row_info->rowbytes; |
253 | while (rb >= 4) { |
254 | /* It's easiest to do this math (particularly, deal with pc) with 16-bit |
255 | * intermediates. |
256 | */ |
257 | __m128i pa,pb,pc,smallest,nearest; |
258 | c = b; b = _mm_unpacklo_epi8(load4(prev), zero); |
259 | a = d; d = _mm_unpacklo_epi8(load4(row ), zero); |
260 | |
261 | /* (p-a) == (a+b-c - a) == (b-c) */ |
262 | |
263 | pa = _mm_sub_epi16(b,c); |
264 | |
265 | /* (p-b) == (a+b-c - b) == (a-c) */ |
266 | pb = _mm_sub_epi16(a,c); |
267 | |
268 | /* (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c) */ |
269 | pc = _mm_add_epi16(pa,pb); |
270 | |
271 | pa = abs_i16(pa); /* |p-a| */ |
272 | pb = abs_i16(pb); /* |p-b| */ |
273 | pc = abs_i16(pc); /* |p-c| */ |
274 | |
275 | smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb)); |
276 | |
277 | /* Paeth breaks ties favoring a over b over c. */ |
278 | nearest = if_then_else(_mm_cmpeq_epi16(smallest, pa), a, |
279 | if_then_else(_mm_cmpeq_epi16(smallest, pb), b, |
280 | c)); |
281 | |
282 | /* Note `_epi8`: we need addition to wrap modulo 255. */ |
283 | d = _mm_add_epi8(d, nearest); |
284 | store3(row, _mm_packus_epi16(d,d)); |
285 | |
286 | prev += 3; |
287 | row += 3; |
288 | rb -= 3; |
289 | } |
290 | if (rb > 0) { |
291 | /* It's easiest to do this math (particularly, deal with pc) with 16-bit |
292 | * intermediates. |
293 | */ |
294 | __m128i pa,pb,pc,smallest,nearest; |
295 | c = b; b = _mm_unpacklo_epi8(load3(prev), zero); |
296 | a = d; d = _mm_unpacklo_epi8(load3(row ), zero); |
297 | |
298 | /* (p-a) == (a+b-c - a) == (b-c) */ |
299 | pa = _mm_sub_epi16(b,c); |
300 | |
301 | /* (p-b) == (a+b-c - b) == (a-c) */ |
302 | pb = _mm_sub_epi16(a,c); |
303 | |
304 | /* (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c) */ |
305 | pc = _mm_add_epi16(pa,pb); |
306 | |
307 | pa = abs_i16(pa); /* |p-a| */ |
308 | pb = abs_i16(pb); /* |p-b| */ |
309 | pc = abs_i16(pc); /* |p-c| */ |
310 | |
311 | smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb)); |
312 | |
313 | /* Paeth breaks ties favoring a over b over c. */ |
314 | nearest = if_then_else(_mm_cmpeq_epi16(smallest, pa), a, |
315 | if_then_else(_mm_cmpeq_epi16(smallest, pb), b, |
316 | c)); |
317 | |
318 | /* Note `_epi8`: we need addition to wrap modulo 255. */ |
319 | d = _mm_add_epi8(d, nearest); |
320 | store3(row, _mm_packus_epi16(d,d)); |
321 | |
322 | prev += 3; |
323 | row += 3; |
324 | rb -= 3; |
325 | } |
326 | } |
327 | |
328 | void png_read_filter_row_paeth4_sse2(png_row_infop row_info, png_bytep row, |
329 | png_const_bytep prev) |
330 | { |
331 | /* Paeth tries to predict pixel d using the pixel to the left of it, a, |
332 | * and two pixels from the previous row, b and c: |
333 | * prev: c b |
334 | * row: a d |
335 | * The Paeth function predicts d to be whichever of a, b, or c is nearest to |
336 | * p=a+b-c. |
337 | * |
338 | * The first pixel has no left context, and so uses an Up filter, p = b. |
339 | * This works naturally with our main loop's p = a+b-c if we force a and c |
340 | * to zero. |
341 | * Here we zero b and d, which become c and a respectively at the start of |
342 | * the loop. |
343 | */ |
344 | size_t rb; |
345 | const __m128i zero = _mm_setzero_si128(); |
346 | __m128i pa,pb,pc,smallest,nearest; |
347 | __m128i c, b = zero, |
348 | a, d = zero; |
349 | |
350 | png_debug(1, "in png_read_filter_row_paeth4_sse2" ); |
351 | |
352 | rb = row_info->rowbytes+4; |
353 | while (rb > 4) { |
354 | /* It's easiest to do this math (particularly, deal with pc) with 16-bit |
355 | * intermediates. |
356 | */ |
357 | c = b; b = _mm_unpacklo_epi8(load4(prev), zero); |
358 | a = d; d = _mm_unpacklo_epi8(load4(row ), zero); |
359 | |
360 | /* (p-a) == (a+b-c - a) == (b-c) */ |
361 | pa = _mm_sub_epi16(b,c); |
362 | |
363 | /* (p-b) == (a+b-c - b) == (a-c) */ |
364 | pb = _mm_sub_epi16(a,c); |
365 | |
366 | /* (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c) */ |
367 | pc = _mm_add_epi16(pa,pb); |
368 | |
369 | pa = abs_i16(pa); /* |p-a| */ |
370 | pb = abs_i16(pb); /* |p-b| */ |
371 | pc = abs_i16(pc); /* |p-c| */ |
372 | |
373 | smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb)); |
374 | |
375 | /* Paeth breaks ties favoring a over b over c. */ |
376 | nearest = if_then_else(_mm_cmpeq_epi16(smallest, pa), a, |
377 | if_then_else(_mm_cmpeq_epi16(smallest, pb), b, |
378 | c)); |
379 | |
380 | /* Note `_epi8`: we need addition to wrap modulo 255. */ |
381 | d = _mm_add_epi8(d, nearest); |
382 | store4(row, _mm_packus_epi16(d,d)); |
383 | |
384 | prev += 4; |
385 | row += 4; |
386 | rb -= 4; |
387 | } |
388 | } |
389 | |
390 | #endif /* PNG_INTEL_SSE_IMPLEMENTATION > 0 */ |
391 | #endif /* READ */ |
392 | |