2 /* filter_neon_intrinsics.c - NEON optimised filter functions
4 * Copyright (c) 2013 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 1.6.8 [December 19, 2013]
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 /* This code requires -mfpu=neon on the command line: */
18 #if PNG_ARM_NEON_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */
22 /* libpng row pointers are not necessarily aligned to any particular boundary,
23 * however this code will only work with appropriate alignment. arm/arm_init.c
24 * checks for this (and will not compile unless it is done). This code uses
25 * variants of png_aligncast to avoid compiler warnings.
27 #define png_ptr(type,pointer) png_aligncast(type *,pointer)
28 #define png_ptrc(type,pointer) png_aligncastconst(const type *,pointer)
30 /* The following relies on a variable 'temp_pointer' being declared with type
31 * 'type'. This is written this way just to hide the GCC strict aliasing
32 * warning; note that the code is safe because there never is an alias between
33 * the input and output pointers.
35 #define png_ldr(type,pointer)\
36 (temp_pointer = png_ptr(type,pointer), *temp_pointer)
38 #ifdef PNG_READ_SUPPORTED
39 #if PNG_ARM_NEON_OPT > 0
42 png_read_filter_row_up_neon(png_row_infop row_info, png_bytep row,
43 png_const_bytep prev_row)
46 png_bytep rp_stop = row + row_info->rowbytes;
47 png_const_bytep pp = prev_row;
49 for (; rp < rp_stop; rp += 16, pp += 16)
55 qrp = vaddq_u8(qrp, qpp);
61 png_read_filter_row_sub3_neon(png_row_infop row_info, png_bytep row,
62 png_const_bytep prev_row)
65 png_bytep rp_stop = row + row_info->rowbytes;
67 uint8x16_t vtmp = vld1q_u8(rp);
68 uint8x8x2_t *vrpt = png_ptr(uint8x8x2_t, &vtmp);
69 uint8x8x2_t vrp = *vrpt;
72 vdest.val[3] = vdup_n_u8(0);
76 uint8x8_t vtmp1, vtmp2;
77 uint32x2_t *temp_pointer;
79 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
80 vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
81 vtmp2 = vext_u8(vrp.val[0], vrp.val[1], 6);
82 vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
84 vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
85 vdest.val[2] = vadd_u8(vdest.val[1], vtmp2);
86 vdest.val[3] = vadd_u8(vdest.val[2], vtmp1);
88 vtmp = vld1q_u8(rp + 12);
89 vrpt = png_ptr(uint8x8x2_t, &vtmp);
92 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
94 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
96 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
98 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
106 png_read_filter_row_sub4_neon(png_row_infop row_info, png_bytep row,
107 png_const_bytep prev_row)
110 png_bytep rp_stop = row + row_info->rowbytes;
113 vdest.val[3] = vdup_n_u8(0);
115 for (; rp < rp_stop; rp += 16)
117 uint32x2x4_t vtmp = vld4_u32(png_ptr(uint32_t,rp));
118 uint8x8x4_t *vrpt = png_ptr(uint8x8x4_t,&vtmp);
119 uint8x8x4_t vrp = *vrpt;
120 uint32x2x4_t *temp_pointer;
122 vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
123 vdest.val[1] = vadd_u8(vdest.val[0], vrp.val[1]);
124 vdest.val[2] = vadd_u8(vdest.val[1], vrp.val[2]);
125 vdest.val[3] = vadd_u8(vdest.val[2], vrp.val[3]);
126 vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
133 png_read_filter_row_avg3_neon(png_row_infop row_info, png_bytep row,
134 png_const_bytep prev_row)
137 png_const_bytep pp = prev_row;
138 png_bytep rp_stop = row + row_info->rowbytes;
144 vdest.val[3] = vdup_n_u8(0);
147 vrpt = png_ptr(uint8x8x2_t,&vtmp);
150 for (; rp < rp_stop; pp += 12)
152 uint8x8_t vtmp1, vtmp2, vtmp3;
157 uint32x2_t *temp_pointer;
160 vppt = png_ptr(uint8x8x2_t,&vtmp);
163 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
164 vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
165 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
167 vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
168 vtmp3 = vext_u8(vrp.val[0], vrp.val[1], 6);
169 vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
170 vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
172 vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 6);
173 vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
175 vtmp = vld1q_u8(rp + 12);
176 vrpt = png_ptr(uint8x8x2_t,&vtmp);
179 vdest.val[2] = vhadd_u8(vdest.val[1], vtmp2);
180 vdest.val[2] = vadd_u8(vdest.val[2], vtmp3);
182 vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
184 vdest.val[3] = vhadd_u8(vdest.val[2], vtmp2);
185 vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
187 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
189 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
191 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
193 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
199 png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row,
200 png_const_bytep prev_row)
203 png_bytep rp_stop = row + row_info->rowbytes;
204 png_const_bytep pp = prev_row;
207 vdest.val[3] = vdup_n_u8(0);
209 for (; rp < rp_stop; rp += 16, pp += 16)
212 uint8x8x4_t *vrpt, *vppt;
213 uint8x8x4_t vrp, vpp;
214 uint32x2x4_t *temp_pointer;
216 vtmp = vld4_u32(png_ptr(uint32_t,rp));
217 vrpt = png_ptr(uint8x8x4_t,&vtmp);
219 vtmp = vld4_u32(png_ptrc(uint32_t,pp));
220 vppt = png_ptr(uint8x8x4_t,&vtmp);
223 vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
224 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
225 vdest.val[1] = vhadd_u8(vdest.val[0], vpp.val[1]);
226 vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
227 vdest.val[2] = vhadd_u8(vdest.val[1], vpp.val[2]);
228 vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
229 vdest.val[3] = vhadd_u8(vdest.val[2], vpp.val[3]);
230 vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
232 vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
237 paeth(uint8x8_t a, uint8x8_t b, uint8x8_t c)
240 uint16x8_t p1, pa, pb, pc;
242 p1 = vaddl_u8(a, b); /* a + b */
243 pc = vaddl_u8(c, c); /* c * 2 */
244 pa = vabdl_u8(b, c); /* pa */
245 pb = vabdl_u8(a, c); /* pb */
246 pc = vabdq_u16(p1, pc); /* pc */
248 p1 = vcleq_u16(pa, pb); /* pa <= pb */
249 pa = vcleq_u16(pa, pc); /* pa <= pc */
250 pb = vcleq_u16(pb, pc); /* pb <= pc */
252 p1 = vandq_u16(p1, pa); /* pa <= pb && pa <= pc */
257 d = vbsl_u8(d, b, c);
258 e = vbsl_u8(e, a, d);
264 png_read_filter_row_paeth3_neon(png_row_infop row_info, png_bytep row,
265 png_const_bytep prev_row)
268 png_const_bytep pp = prev_row;
269 png_bytep rp_stop = row + row_info->rowbytes;
274 uint8x8_t vlast = vdup_n_u8(0);
276 vdest.val[3] = vdup_n_u8(0);
279 vrpt = png_ptr(uint8x8x2_t,&vtmp);
282 for (; rp < rp_stop; pp += 12)
286 uint8x8_t vtmp1, vtmp2, vtmp3;
287 uint32x2_t *temp_pointer;
290 vppt = png_ptr(uint8x8x2_t,&vtmp);
293 vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
294 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
296 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
297 vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
298 vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
299 vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
301 vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 6);
302 vtmp3 = vext_u8(vpp.val[0], vpp.val[1], 6);
303 vdest.val[2] = paeth(vdest.val[1], vtmp3, vtmp2);
304 vdest.val[2] = vadd_u8(vdest.val[2], vtmp1);
306 vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
307 vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
309 vtmp = vld1q_u8(rp + 12);
310 vrpt = png_ptr(uint8x8x2_t,&vtmp);
313 vdest.val[3] = paeth(vdest.val[2], vtmp2, vtmp3);
314 vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
318 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
320 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
322 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
324 vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
330 png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
331 png_const_bytep prev_row)
334 png_bytep rp_stop = row + row_info->rowbytes;
335 png_const_bytep pp = prev_row;
337 uint8x8_t vlast = vdup_n_u8(0);
339 vdest.val[3] = vdup_n_u8(0);
341 for (; rp < rp_stop; rp += 16, pp += 16)
344 uint8x8x4_t *vrpt, *vppt;
345 uint8x8x4_t vrp, vpp;
346 uint32x2x4_t *temp_pointer;
348 vtmp = vld4_u32(png_ptr(uint32_t,rp));
349 vrpt = png_ptr(uint8x8x4_t,&vtmp);
351 vtmp = vld4_u32(png_ptrc(uint32_t,pp));
352 vppt = png_ptr(uint8x8x4_t,&vtmp);
355 vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
356 vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
357 vdest.val[1] = paeth(vdest.val[0], vpp.val[1], vpp.val[0]);
358 vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
359 vdest.val[2] = paeth(vdest.val[1], vpp.val[2], vpp.val[1]);
360 vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
361 vdest.val[3] = paeth(vdest.val[2], vpp.val[3], vpp.val[2]);
362 vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
366 vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
370 #endif /* PNG_ARM_NEON_OPT > 0 */
371 #endif /* PNG_READ_SUPPORTED */
372 #endif /* PNG_ARM_NEON_IMPLEMENTATION == 1 (intrinsics) */