2 /* filter_neon.S - NEON optimised filter functions
4 * Copyright (c) 2013 Glenn Randers-Pehrson
5 * Written by Mans Rullgard, 2011.
6 * Last changed in libpng 1.6.8 [December 19, 2013]
8 * This code is released under the libpng license.
9 * For conditions of distribution and use, see the disclaimer
10 * and license in png.h
13 /* This is required to get the symbol renames, which are #defines, and also
14 * includes the definition (or not) of PNG_ARM_NEON_OPT and
15 * PNG_ARM_NEON_IMPLEMENTATION.
17 #define PNG_VERSION_INFO_ONLY
18 #include "../pngpriv.h"
20 #if defined(__linux__) && defined(__ELF__)
21 .section .note.GNU-stack,"",%progbits /* mark stack as non-executable */
24 /* Assembler NEON support - only works for 32-bit ARM (i.e. it does not work for
25 * ARM64). The code in arm/filter_neon_intrinsics.c supports ARM64, however it
26 * only works if -mfpu=neon is specified on the GCC command line. See pngpriv.h
27 * for the logic which sets PNG_USE_ARM_NEON_ASM:
29 #if PNG_ARM_NEON_IMPLEMENTATION == 2 /* hand-coded assembler */
31 #ifdef PNG_READ_SUPPORTED
32 #if PNG_ARM_NEON_OPT > 0
43 .macro func name, export=0
45 ELF .size \name, . - \name
53 ELF .type \name, STT_FUNC
58 func png_read_filter_row_sub4_neon, export=1
59 ldr r3, [r0, #4] @ rowbytes
62 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
67 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
74 func png_read_filter_row_sub3_neon, export=1
75 ldr r3, [r0, #4] @ rowbytes
80 vld1.8 {q11}, [r0], r12
82 vext.8 d5, d22, d23, #3
84 vext.8 d6, d22, d23, #6
86 vext.8 d7, d23, d23, #1
87 vld1.8 {q11}, [r0], r12
88 vst1.32 {d0[0]}, [r1,:32], r2
90 vst1.32 {d1[0]}, [r1], r2
92 vst1.32 {d2[0]}, [r1], r2
93 vst1.32 {d3[0]}, [r1], r2
100 func png_read_filter_row_up_neon, export=1
101 ldr r3, [r0, #4] @ rowbytes
103 vld1.8 {q0}, [r1,:128]
104 vld1.8 {q1}, [r2,:128]!
106 vst1.8 {q0}, [r1,:128]!
113 func png_read_filter_row_avg4_neon, export=1
114 ldr r12, [r0, #4] @ rowbytes
117 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
118 vld4.32 {d16[],d17[],d18[],d19[]},[r2,:128]!
127 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
134 func png_read_filter_row_avg3_neon, export=1
136 ldr r12, [r0, #4] @ rowbytes
141 vld1.8 {q11}, [r0], lr
143 vld1.8 {q10}, [r2], lr
144 vext.8 d5, d22, d23, #3
146 vext.8 d17, d20, d21, #3
148 vext.8 d6, d22, d23, #6
150 vext.8 d18, d20, d21, #6
152 vext.8 d7, d23, d23, #1
153 vld1.8 {q11}, [r0], lr
154 vst1.32 {d0[0]}, [r1,:32], r4
156 vst1.32 {d1[0]}, [r1], r4
157 vext.8 d19, d21, d21, #1
160 vst1.32 {d2[0]}, [r1], r4
162 vst1.32 {d3[0]}, [r1], r4
169 .macro paeth rx, ra, rb, rc
170 vaddl.u8 q12, \ra, \rb @ a + b
171 vaddl.u8 q15, \rc, \rc @ 2*c
172 vabdl.u8 q13, \rb, \rc @ pa
173 vabdl.u8 q14, \ra, \rc @ pb
174 vabd.u16 q15, q12, q15 @ pc
175 vcle.u16 q12, q13, q14 @ pa <= pb
176 vcle.u16 q13, q13, q15 @ pa <= pc
177 vcle.u16 q14, q14, q15 @ pb <= pc
178 vand q12, q12, q13 @ pa <= pb && pa <= pc
185 func png_read_filter_row_paeth4_neon, export=1
186 ldr r12, [r0, #4] @ rowbytes
190 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
191 vld4.32 {d16[],d17[],d18[],d19[]},[r2,:128]!
192 paeth d0, d3, d16, d20
194 paeth d1, d0, d17, d16
196 paeth d2, d1, d18, d17
198 paeth d3, d2, d19, d18
201 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
208 func png_read_filter_row_paeth3_neon, export=1
210 ldr r12, [r0, #4] @ rowbytes
216 vld1.8 {q11}, [r0], lr
218 vld1.8 {q10}, [r2], lr
219 paeth d0, d3, d20, d4
220 vext.8 d5, d22, d23, #3
222 vext.8 d17, d20, d21, #3
223 paeth d1, d0, d17, d20
224 vst1.32 {d0[0]}, [r1,:32], r4
225 vext.8 d6, d22, d23, #6
227 vext.8 d18, d20, d21, #6
228 paeth d2, d1, d18, d17
229 vext.8 d7, d23, d23, #1
230 vld1.8 {q11}, [r0], lr
231 vst1.32 {d1[0]}, [r1], r4
233 vext.8 d19, d21, d21, #1
234 paeth d3, d2, d19, d18
235 vst1.32 {d2[0]}, [r1], r4
238 vst1.32 {d3[0]}, [r1], r4
244 #endif /* PNG_ARM_NEON_OPT > 0 */
245 #endif /* PNG_READ_SUPPORTED */
246 #endif /* PNG_ARM_NEON_IMPLEMENTATION == 2 (assembler) */