2 /* filter_neon.S - NEON optimised filter functions
4 * Copyright (c) 2018 Cosmin Truta
5 * Copyright (c) 2014,2017 Glenn Randers-Pehrson
6 * Written by Mans Rullgard, 2011.
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 the
14 * definitions (or not) of PNG_ARM_NEON_OPT and PNG_ARM_NEON_IMPLEMENTATION.
16 #define PNG_VERSION_INFO_ONLY
17 #define PNG_ARM_NEON_IMPLEMENTATION 1 // apply patch for tizen
18 #include "../pngpriv.h"
20 #if (defined(__linux__) || defined(__FreeBSD__)) && defined(__ELF__)
21 .section .note.GNU-stack,"",%progbits /* mark stack as non-executable */
24 #ifdef PNG_READ_SUPPORTED
26 /* Assembler NEON support - only works for 32-bit ARM (i.e. it does not work for
27 * ARM64). The code in arm/filter_neon_intrinsics.c supports ARM64, however it
28 * only works if -mfpu=neon is specified on the GCC command line. See pngpriv.h
29 * for the logic which sets PNG_USE_ARM_NEON_ASM:
31 #if PNG_ARM_NEON_IMPLEMENTATION == 2 /* hand-coded assembler */
33 #if PNG_ARM_NEON_OPT > 0
44 .macro func name, export=0
46 ELF .size \name, . - \name
52 /* Explicitly specifying alignment here because some versions of
53 * GAS don't align code correctly. This is harmless in correctly
54 * written versions of GAS.
61 ELF .type \name, STT_FUNC
66 func png_read_filter_row_sub4_neon, export=1
67 ldr r3, [r0, #4] @ rowbytes
70 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
75 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
82 func png_read_filter_row_sub3_neon, export=1
83 ldr r3, [r0, #4] @ rowbytes
88 vld1.8 {q11}, [r0], r12
90 vext.8 d5, d22, d23, #3
92 vext.8 d6, d22, d23, #6
94 vext.8 d7, d23, d23, #1
95 vld1.8 {q11}, [r0], r12
96 vst1.32 {d0[0]}, [r1,:32], r2
98 vst1.32 {d1[0]}, [r1], r2
100 vst1.32 {d2[0]}, [r1], r2
101 vst1.32 {d3[0]}, [r1], r2
108 func png_read_filter_row_up_neon, export=1
109 ldr r3, [r0, #4] @ rowbytes
111 vld1.8 {q0}, [r1,:128]
112 vld1.8 {q1}, [r2,:128]!
114 vst1.8 {q0}, [r1,:128]!
121 func png_read_filter_row_avg4_neon, export=1
122 ldr r12, [r0, #4] @ rowbytes
125 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
126 vld4.32 {d16[],d17[],d18[],d19[]},[r2,:128]!
135 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
142 func png_read_filter_row_avg3_neon, export=1
144 ldr r12, [r0, #4] @ rowbytes
149 vld1.8 {q11}, [r0], lr
151 vld1.8 {q10}, [r2], lr
152 vext.8 d5, d22, d23, #3
154 vext.8 d17, d20, d21, #3
156 vext.8 d6, d22, d23, #6
158 vext.8 d18, d20, d21, #6
160 vext.8 d7, d23, d23, #1
161 vld1.8 {q11}, [r0], lr
162 vst1.32 {d0[0]}, [r1,:32], r4
164 vst1.32 {d1[0]}, [r1], r4
165 vext.8 d19, d21, d21, #1
168 vst1.32 {d2[0]}, [r1], r4
170 vst1.32 {d3[0]}, [r1], r4
177 .macro paeth rx, ra, rb, rc
178 vaddl.u8 q12, \ra, \rb @ a + b
179 vaddl.u8 q15, \rc, \rc @ 2*c
180 vabdl.u8 q13, \rb, \rc @ pa
181 vabdl.u8 q14, \ra, \rc @ pb
182 vabd.u16 q15, q12, q15 @ pc
183 vcle.u16 q12, q13, q14 @ pa <= pb
184 vcle.u16 q13, q13, q15 @ pa <= pc
185 vcle.u16 q14, q14, q15 @ pb <= pc
186 vand q12, q12, q13 @ pa <= pb && pa <= pc
193 func png_read_filter_row_paeth4_neon, export=1
194 ldr r12, [r0, #4] @ rowbytes
198 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
199 vld4.32 {d16[],d17[],d18[],d19[]},[r2,:128]!
200 paeth d0, d3, d16, d20
202 paeth d1, d0, d17, d16
204 paeth d2, d1, d18, d17
206 paeth d3, d2, d19, d18
209 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
216 func png_read_filter_row_paeth3_neon, export=1
218 ldr r12, [r0, #4] @ rowbytes
224 vld1.8 {q11}, [r0], lr
226 vld1.8 {q10}, [r2], lr
227 paeth d0, d3, d20, d4
228 vext.8 d5, d22, d23, #3
230 vext.8 d17, d20, d21, #3
231 paeth d1, d0, d17, d20
232 vst1.32 {d0[0]}, [r1,:32], r4
233 vext.8 d6, d22, d23, #6
235 vext.8 d18, d20, d21, #6
236 paeth d2, d1, d18, d17
237 vext.8 d7, d23, d23, #1
238 vld1.8 {q11}, [r0], lr
239 vst1.32 {d1[0]}, [r1], r4
241 vext.8 d19, d21, d21, #1
242 paeth d3, d2, d19, d18
243 vst1.32 {d2[0]}, [r1], r4
246 vst1.32 {d3[0]}, [r1], r4
252 #endif /* PNG_ARM_NEON_OPT > 0 */
253 #endif /* PNG_ARM_NEON_IMPLEMENTATION == 2 (assembler) */