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 /* These are required because Mozilla's moz.build system doesn't pass
14 * -DDefined macros to the assembler.
16 #define PNG_READ_SUPPORTED
17 #define MOZ_PNG_HAVE_ARM_NEON
19 /* This is required to get the symbol renames, which are #defines, and also
20 * includes the definition (or not) of PNG_ARM_NEON_OPT and
21 * PNG_ARM_NEON_IMPLEMENTATION.
23 #define PNG_VERSION_INFO_ONLY
24 #include "../pngpriv.h"
26 #if defined(__linux__) && defined(__ELF__)
27 .section .note.GNU-stack,"",%progbits /* mark stack as non-executable */
30 /* Assembler NEON support - only works for 32-bit ARM (i.e. it does not work for
31 * ARM64). The code in arm/filter_neon_intrinsics.c supports ARM64, however it
32 * only works if -mfpu=neon is specified on the GCC command line. See pngpriv.h
33 * for the logic which sets PNG_USE_ARM_NEON_ASM:
35 #if PNG_ARM_NEON_IMPLEMENTATION == 2 /* hand-coded assembler */
37 #ifdef PNG_READ_SUPPORTED
38 #if PNG_ARM_NEON_OPT > 0
49 .macro func name, export=0
51 ELF .size \name, . - \name
57 /* Explicitly specifying alignment here because some versions of
58 gas don't align code correctly. See
59 http://lists.gnu.org/archive/html/bug-binutils/2011-06/msg00199.html
60 and https://bugzilla.mozilla.org/show_bug.cgi?id=920992
67 ELF .type \name, STT_FUNC
72 func png_read_filter_row_sub4_neon, export=1
73 ldr r3, [r0, #4] @ rowbytes
76 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
81 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
88 func png_read_filter_row_sub3_neon, export=1
89 ldr r3, [r0, #4] @ rowbytes
94 vld1.8 {q11}, [r0], r12
96 vext.8 d5, d22, d23, #3
98 vext.8 d6, d22, d23, #6
100 vext.8 d7, d23, d23, #1
101 vld1.8 {q11}, [r0], r12
102 vst1.32 {d0[0]}, [r1,:32], r2
104 vst1.32 {d1[0]}, [r1], r2
106 vst1.32 {d2[0]}, [r1], r2
107 vst1.32 {d3[0]}, [r1], r2
114 func png_read_filter_row_up_neon, export=1
115 ldr r3, [r0, #4] @ rowbytes
117 vld1.8 {q0}, [r1,:128]
118 vld1.8 {q1}, [r2,:128]!
120 vst1.8 {q0}, [r1,:128]!
127 func png_read_filter_row_avg4_neon, export=1
128 ldr r12, [r0, #4] @ rowbytes
131 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
132 vld4.32 {d16[],d17[],d18[],d19[]},[r2,:128]!
141 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
148 func png_read_filter_row_avg3_neon, export=1
150 ldr r12, [r0, #4] @ rowbytes
155 vld1.8 {q11}, [r0], lr
157 vld1.8 {q10}, [r2], lr
158 vext.8 d5, d22, d23, #3
160 vext.8 d17, d20, d21, #3
162 vext.8 d6, d22, d23, #6
164 vext.8 d18, d20, d21, #6
166 vext.8 d7, d23, d23, #1
167 vld1.8 {q11}, [r0], lr
168 vst1.32 {d0[0]}, [r1,:32], r4
170 vst1.32 {d1[0]}, [r1], r4
171 vext.8 d19, d21, d21, #1
174 vst1.32 {d2[0]}, [r1], r4
176 vst1.32 {d3[0]}, [r1], r4
183 .macro paeth rx, ra, rb, rc
184 vaddl.u8 q12, \ra, \rb @ a + b
185 vaddl.u8 q15, \rc, \rc @ 2*c
186 vabdl.u8 q13, \rb, \rc @ pa
187 vabdl.u8 q14, \ra, \rc @ pb
188 vabd.u16 q15, q12, q15 @ pc
189 vcle.u16 q12, q13, q14 @ pa <= pb
190 vcle.u16 q13, q13, q15 @ pa <= pc
191 vcle.u16 q14, q14, q15 @ pb <= pc
192 vand q12, q12, q13 @ pa <= pb && pa <= pc
199 func png_read_filter_row_paeth4_neon, export=1
200 ldr r12, [r0, #4] @ rowbytes
204 vld4.32 {d4[],d5[],d6[],d7[]}, [r1,:128]
205 vld4.32 {d16[],d17[],d18[],d19[]},[r2,:128]!
206 paeth d0, d3, d16, d20
208 paeth d1, d0, d17, d16
210 paeth d2, d1, d18, d17
212 paeth d3, d2, d19, d18
215 vst4.32 {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
222 func png_read_filter_row_paeth3_neon, export=1
224 ldr r12, [r0, #4] @ rowbytes
230 vld1.8 {q11}, [r0], lr
232 vld1.8 {q10}, [r2], lr
233 paeth d0, d3, d20, d4
234 vext.8 d5, d22, d23, #3
236 vext.8 d17, d20, d21, #3
237 paeth d1, d0, d17, d20
238 vst1.32 {d0[0]}, [r1,:32], r4
239 vext.8 d6, d22, d23, #6
241 vext.8 d18, d20, d21, #6
242 paeth d2, d1, d18, d17
243 vext.8 d7, d23, d23, #1
244 vld1.8 {q11}, [r0], lr
245 vst1.32 {d1[0]}, [r1], r4
247 vext.8 d19, d21, d21, #1
248 paeth d3, d2, d19, d18
249 vst1.32 {d2[0]}, [r1], r4
252 vst1.32 {d3[0]}, [r1], r4
258 #endif /* PNG_ARM_NEON_OPT > 0 */
259 #endif /* PNG_READ_SUPPORTED */
260 #endif /* PNG_ARM_NEON_IMPLEMENTATION == 2 (assembler) */