2 /* filter_neon_intrinsics.c - NEON optimised filter functions
4 * Copyright (c) 2014 Glenn Randers-Pehrson
5 * Written by James Yu <james.yu at linaro.org>, October 2013.
6 * Based on filter_neon.S, written by Mans Rullgard, 2011.
8 * Last changed in libpng 1.6.16 [December 22, 2014]
10 * This code is released under the libpng license.
11 * For conditions of distribution and use, see the disclaimer
12 * and license in png.h
15 #include "../pngpriv.h"
17 #ifdef PNG_READ_SUPPORTED
19 /* This code requires -mfpu=neon on the command line: */
20 #if PNG_ARM_NEON_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */
24 /* libpng row pointers are not necessarily aligned to any particular boundary,
25 * however this code will only work with appropriate alignment. arm/arm_init.c
26 * checks for this (and will not compile unless it is done). This code uses
27 * variants of png_aligncast to avoid compiler warnings.
29 #define png_ptr(type,pointer) png_aligncast(type *,pointer)
30 #define png_ptrc(type,pointer) png_aligncastconst(const type *,pointer)
32 /* The following relies on a variable 'temp_pointer' being declared with type
33 * 'type'. This is written this way just to hide the GCC strict aliasing
34 * warning; note that the code is safe because there never is an alias between
35 * the input and output pointers.
37 #define png_ldr(type,pointer)\
38 (temp_pointer = png_ptr(type,pointer), *temp_pointer)
40 #if PNG_ARM_NEON_OPT > 0
43 png_read_filter_row_up_neon(png_row_infop row_info, png_bytep row,
44 png_const_bytep prev_row)
47 png_bytep rp_stop = row + row_info->rowbytes;
48 png_const_bytep pp = prev_row;
50 for (; rp < rp_stop; rp += 16, pp += 16)
56 qrp = vaddq_u8(qrp, qpp);
62 png_read_filter_row_sub3_neon(png_row_infop row_info, png_bytep row,
63 png_const_bytep prev_row)
66 png_bytep rp_stop = row + row_info->rowbytes;
68 uint8x16_t vtmp = vld1q_u8(rp);
69 uint8x8x2_t *vrpt = png_ptr(uint8x8x2_t, &vtmp);
70 uint8x8x2_t vrp = *vrpt;
73 vdest.val[3] = vdup_n_u8(0);
77 uint8x8_t vtmp1, vtmp2;
78 uint32x2_t *temp_pointer;
80 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
81 vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
82 vtmp2 = vext_u8(vrp.val[0], vrp.val[1], 6);
83 vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
85 vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
86 vdest.val[2] = vadd_u8(vdest.val[1], vtmp2);
87 vdest.val[3] = vadd_u8(vdest.val[2], vtmp1);
89 vtmp = vld1q_u8(rp + 12);
90 vrpt = png_ptr(uint8x8x2_t, &vtmp);
93 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
95 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
97 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
99 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
107 png_read_filter_row_sub4_neon(png_row_infop row_info, png_bytep row,
108 png_const_bytep prev_row)
111 png_bytep rp_stop = row + row_info->rowbytes;
114 vdest.val[3] = vdup_n_u8(0);
116 for (; rp < rp_stop; rp += 16)
118 uint32x2x4_t vtmp = vld4_u32(png_ptr(uint32_t,rp));
119 uint8x8x4_t *vrpt = png_ptr(uint8x8x4_t,&vtmp);
120 uint8x8x4_t vrp = *vrpt;
121 uint32x2x4_t *temp_pointer;
123 vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
124 vdest.val[1] = vadd_u8(vdest.val[0], vrp.val[1]);
125 vdest.val[2] = vadd_u8(vdest.val[1], vrp.val[2]);
126 vdest.val[3] = vadd_u8(vdest.val[2], vrp.val[3]);
127 vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
134 png_read_filter_row_avg3_neon(png_row_infop row_info, png_bytep row,
135 png_const_bytep prev_row)
138 png_const_bytep pp = prev_row;
139 png_bytep rp_stop = row + row_info->rowbytes;
145 vdest.val[3] = vdup_n_u8(0);
148 vrpt = png_ptr(uint8x8x2_t,&vtmp);
151 for (; rp < rp_stop; pp += 12)
153 uint8x8_t vtmp1, vtmp2, vtmp3;
158 uint32x2_t *temp_pointer;
161 vppt = png_ptr(uint8x8x2_t,&vtmp);
164 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
165 vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
166 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
168 vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
169 vtmp3 = vext_u8(vrp.val[0], vrp.val[1], 6);
170 vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
171 vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
173 vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 6);
174 vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
176 vtmp = vld1q_u8(rp + 12);
177 vrpt = png_ptr(uint8x8x2_t,&vtmp);
180 vdest.val[2] = vhadd_u8(vdest.val[1], vtmp2);
181 vdest.val[2] = vadd_u8(vdest.val[2], vtmp3);
183 vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
185 vdest.val[3] = vhadd_u8(vdest.val[2], vtmp2);
186 vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
188 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
190 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
192 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
194 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
200 png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row,
201 png_const_bytep prev_row)
204 png_bytep rp_stop = row + row_info->rowbytes;
205 png_const_bytep pp = prev_row;
208 vdest.val[3] = vdup_n_u8(0);
210 for (; rp < rp_stop; rp += 16, pp += 16)
213 uint8x8x4_t *vrpt, *vppt;
214 uint8x8x4_t vrp, vpp;
215 uint32x2x4_t *temp_pointer;
217 vtmp = vld4_u32(png_ptr(uint32_t,rp));
218 vrpt = png_ptr(uint8x8x4_t,&vtmp);
220 vtmp = vld4_u32(png_ptrc(uint32_t,pp));
221 vppt = png_ptr(uint8x8x4_t,&vtmp);
224 vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
225 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
226 vdest.val[1] = vhadd_u8(vdest.val[0], vpp.val[1]);
227 vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
228 vdest.val[2] = vhadd_u8(vdest.val[1], vpp.val[2]);
229 vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
230 vdest.val[3] = vhadd_u8(vdest.val[2], vpp.val[3]);
231 vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
233 vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
238 paeth(uint8x8_t a, uint8x8_t b, uint8x8_t c)
241 uint16x8_t p1, pa, pb, pc;
243 p1 = vaddl_u8(a, b); /* a + b */
244 pc = vaddl_u8(c, c); /* c * 2 */
245 pa = vabdl_u8(b, c); /* pa */
246 pb = vabdl_u8(a, c); /* pb */
247 pc = vabdq_u16(p1, pc); /* pc */
249 p1 = vcleq_u16(pa, pb); /* pa <= pb */
250 pa = vcleq_u16(pa, pc); /* pa <= pc */
251 pb = vcleq_u16(pb, pc); /* pb <= pc */
253 p1 = vandq_u16(p1, pa); /* pa <= pb && pa <= pc */
258 d = vbsl_u8(d, b, c);
259 e = vbsl_u8(e, a, d);
265 png_read_filter_row_paeth3_neon(png_row_infop row_info, png_bytep row,
266 png_const_bytep prev_row)
269 png_const_bytep pp = prev_row;
270 png_bytep rp_stop = row + row_info->rowbytes;
275 uint8x8_t vlast = vdup_n_u8(0);
277 vdest.val[3] = vdup_n_u8(0);
280 vrpt = png_ptr(uint8x8x2_t,&vtmp);
283 for (; rp < rp_stop; pp += 12)
287 uint8x8_t vtmp1, vtmp2, vtmp3;
288 uint32x2_t *temp_pointer;
291 vppt = png_ptr(uint8x8x2_t,&vtmp);
294 vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
295 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
297 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
298 vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
299 vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
300 vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
302 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 6);
303 vtmp3 = vext_u8(vpp.val[0], vpp.val[1], 6);
304 vdest.val[2] = paeth(vdest.val[1], vtmp3, vtmp2);
305 vdest.val[2] = vadd_u8(vdest.val[2], vtmp1);
307 vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
308 vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
310 vtmp = vld1q_u8(rp + 12);
311 vrpt = png_ptr(uint8x8x2_t,&vtmp);
314 vdest.val[3] = paeth(vdest.val[2], vtmp2, vtmp3);
315 vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
319 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
321 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
323 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
325 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
331 png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
332 png_const_bytep prev_row)
335 png_bytep rp_stop = row + row_info->rowbytes;
336 png_const_bytep pp = prev_row;
338 uint8x8_t vlast = vdup_n_u8(0);
340 vdest.val[3] = vdup_n_u8(0);
342 for (; rp < rp_stop; rp += 16, pp += 16)
345 uint8x8x4_t *vrpt, *vppt;
346 uint8x8x4_t vrp, vpp;
347 uint32x2x4_t *temp_pointer;
349 vtmp = vld4_u32(png_ptr(uint32_t,rp));
350 vrpt = png_ptr(uint8x8x4_t,&vtmp);
352 vtmp = vld4_u32(png_ptrc(uint32_t,pp));
353 vppt = png_ptr(uint8x8x4_t,&vtmp);
356 vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
357 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
358 vdest.val[1] = paeth(vdest.val[0], vpp.val[1], vpp.val[0]);
359 vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
360 vdest.val[2] = paeth(vdest.val[1], vpp.val[2], vpp.val[1]);
361 vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
362 vdest.val[3] = paeth(vdest.val[2], vpp.val[3], vpp.val[2]);
363 vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
367 vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
371 #endif /* PNG_ARM_NEON_OPT > 0 */
372 #endif /* PNG_ARM_NEON_IMPLEMENTATION == 1 (intrinsics) */