1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
|
From d2f06cd65d7ac39c6dd6761eef162abc946b155b Mon Sep 17 00:00:00 2001
From: Adenilson Cavalcanti <adenilson.cavalcanti@arm.com>
Date: Tue, 11 Apr 2017 17:13:02 -0700
Subject: [PATCH] NEON implementation for Adler32
The checksum is calculated in the uncompressed PNG data
and can be made much faster by using SIMD.
Tests in ARMv8 yielded an improvement of about 3x
(e.g. walltime was 350ms x 125ms for a 4096x4096 bytes
executed 30 times). That results in at least 18% improvement
in image decoding in Chromium.
Further details at:
https://bugs.chromium.org/p/chromium/issues/detail?id=688601
---
CMakeLists.txt | 29 +++++++---
adler32.c | 5 ++
contrib/README.contrib | 3 +
contrib/arm/neon_adler32.c | 137 +++++++++++++++++++++++++++++++++++++++++++++
4 files changed, 166 insertions(+), 8 deletions(-)
create mode 100644 contrib/arm/neon_adler32.c
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0fe939df..8e75f664 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -7,6 +7,7 @@ set(VERSION "1.2.11")
option(ASM686 "Enable building i686 assembly implementation")
option(AMD64 "Enable building amd64 assembly implementation")
+option(ARMv8 "Enable building ARM NEON intrinsics implementation")
set(INSTALL_BIN_DIR "${CMAKE_INSTALL_PREFIX}/bin" CACHE PATH "Installation directory for executables")
set(INSTALL_LIB_DIR "${CMAKE_INSTALL_PREFIX}/lib" CACHE PATH "Installation directory for libraries")
@@ -132,14 +133,26 @@ endif()
if(CMAKE_COMPILER_IS_GNUCC)
if(ASM686)
set(ZLIB_ASMS contrib/asm686/match.S)
- elseif (AMD64)
+ elseif(AMD64)
set(ZLIB_ASMS contrib/amd64/amd64-match.S)
- endif ()
+ elseif(ARMv8)
+ set(ZLIB_ARMv8 contrib/arm/neon_adler32.c)
+ endif()
- if(ZLIB_ASMS)
- add_definitions(-DASMV)
- set_source_files_properties(${ZLIB_ASMS} PROPERTIES LANGUAGE C COMPILE_FLAGS -DNO_UNDERLINE)
- endif()
+ if(ZLIB_ASMS)
+ add_definitions(-DASMV)
+ set_source_files_properties(${ZLIB_ASMS} PROPERTIES LANGUAGE C COMPILE_FLAGS -DNO_UNDERLINE)
+ elseif(ZLIB_ARMv8)
+ add_definitions(-DARMv8)
+ set(COMPILER ${CMAKE_C_COMPILER})
+ # NEON is mandatory in ARMv8.
+ if(${COMPILER} MATCHES "aarch64")
+ set_source_files_properties(${ZLIB_ARMv8} PROPERTIES LANGUAGE C COMPILE_FLAGS -march=armv8-a)
+ # But it was optional for ARMv7.
+ elseif(${COMPILER} MATCHES "arm")
+ set_source_files_properties(${ZLIB_ARMv8} PROPERTIES LANGUAGE C COMPILE_FLAGS -mfpu=neon)
+ endif()
+ endif()
endif()
if(MSVC)
@@ -183,8 +196,8 @@ if(MINGW)
set(ZLIB_DLL_SRCS ${CMAKE_CURRENT_BINARY_DIR}/zlib1rc.obj)
endif(MINGW)
-add_library(zlib SHARED ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_DLL_SRCS} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
-add_library(zlibstatic STATIC ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
+add_library(zlib SHARED ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_ARMv8} ${ZLIB_DLL_SRCS} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
+add_library(zlibstatic STATIC ${ZLIB_SRCS} ${ZLIB_ASMS} ${ZLIB_ARMv8} ${ZLIB_PUBLIC_HDRS} ${ZLIB_PRIVATE_HDRS})
set_target_properties(zlib PROPERTIES DEFINE_SYMBOL ZLIB_DLL)
set_target_properties(zlib PROPERTIES SOVERSION 1)
diff --git a/adler32.c b/adler32.c
index d0be4380..45ebaa4b 100644
--- a/adler32.c
+++ b/adler32.c
@@ -136,7 +136,12 @@ uLong ZEXPORT adler32(adler, buf, len)
const Bytef *buf;
uInt len;
{
+#ifdef ARMv8
+# pragma message("Using NEON-ized Adler32.")
+ return NEON_adler32(adler, buf, len);
+#else
return adler32_z(adler, buf, len);
+#endif
}
/* ========================================================================= */
diff --git a/contrib/README.contrib b/contrib/README.contrib
index a411d5c3..3fd1d202 100644
--- a/contrib/README.contrib
+++ b/contrib/README.contrib
@@ -12,6 +12,9 @@ amd64/ by Mikhail Teterin <mi@ALDAN.algebra.com>
asm code for AMD64
See patch at http://www.freebsd.org/cgi/query-pr.cgi?pr=bin/96393
+arm/ by Adenilson Cavalcanti <cavalcantii@chromium.org>
+ ARM optimizations (NEON and ARMv8 code).
+
asm686/ by Brian Raiter <breadbox@muppetlabs.com>
asm code for Pentium and PPro/PII, using the AT&T (GNU as) syntax
See http://www.muppetlabs.com/~breadbox/software/assembly.html
diff --git a/contrib/arm/neon_adler32.c b/contrib/arm/neon_adler32.c
new file mode 100644
index 00000000..f173a74f
--- /dev/null
+++ b/contrib/arm/neon_adler32.c
@@ -0,0 +1,137 @@
+/* Copyright (C) 1995-2011, 2016 Mark Adler
+ * Copyright (C) 2017 ARM Holdings Inc.
+ * Authors: Adenilson Cavalcanti <adenilson.cavalcanti@arm.com>
+ * Simon Hosie <simon.hosie@arm.com>
+ * This software is provided 'as-is', without any express or implied
+ * warranty. In no event will the authors be held liable for any damages
+ * arising from the use of this software.
+ * Permission is granted to anyone to use this software for any purpose,
+ * including commercial applications, and to alter it and redistribute it
+ * freely, subject to the following restrictions:
+ * 1. The origin of this software must not be misrepresented; you must not
+ * claim that you wrote the original software. If you use this software
+ * in a product, an acknowledgment in the product documentation would be
+ * appreciated but is not required.
+ * 2. Altered source versions must be plainly marked as such, and must not be
+ * misrepresented as being the original software.
+ * 3. This notice may not be removed or altered from any source distribution.
+ */
+
+#if (defined(__ARM_NEON__) || defined(__ARM_NEON))
+#include <arm_neon.h>
+
+static void NEON_accum32(uint32_t *s, const unsigned char *buf,
+ unsigned int len)
+{
+ static const uint8_t taps[32] = {
+ 32, 31, 30, 29, 28, 27, 26, 25,
+ 24, 23, 22, 21, 20, 19, 18, 17,
+ 16, 15, 14, 13, 12, 11, 10, 9,
+ 8, 7, 6, 5, 4, 3, 2, 1 };
+
+ uint32x2_t adacc2, s2acc2, as;
+ uint8x16_t t0 = vld1q_u8(taps), t1 = vld1q_u8(taps + 16);
+
+ uint32x4_t adacc = vdupq_n_u32(0), s2acc = vdupq_n_u32(0);
+ adacc = vsetq_lane_u32(s[0], adacc, 0);
+ s2acc = vsetq_lane_u32(s[1], s2acc, 0);
+
+ while (len >= 2) {
+ uint8x16_t d0 = vld1q_u8(buf), d1 = vld1q_u8(buf + 16);
+ uint16x8_t adler, sum2;
+ s2acc = vaddq_u32(s2acc, vshlq_n_u32(adacc, 5));
+ adler = vpaddlq_u8( d0);
+ adler = vpadalq_u8(adler, d1);
+ sum2 = vmull_u8( vget_low_u8(t0), vget_low_u8(d0));
+ sum2 = vmlal_u8(sum2, vget_high_u8(t0), vget_high_u8(d0));
+ sum2 = vmlal_u8(sum2, vget_low_u8(t1), vget_low_u8(d1));
+ sum2 = vmlal_u8(sum2, vget_high_u8(t1), vget_high_u8(d1));
+ adacc = vpadalq_u16(adacc, adler);
+ s2acc = vpadalq_u16(s2acc, sum2);
+ len -= 2;
+ buf += 32;
+ }
+
+ while (len > 0) {
+ uint8x16_t d0 = vld1q_u8(buf);
+ uint16x8_t adler, sum2;
+ s2acc = vaddq_u32(s2acc, vshlq_n_u32(adacc, 4));
+ adler = vpaddlq_u8(d0);
+ sum2 = vmull_u8( vget_low_u8(t1), vget_low_u8(d0));
+ sum2 = vmlal_u8(sum2, vget_high_u8(t1), vget_high_u8(d0));
+ adacc = vpadalq_u16(adacc, adler);
+ s2acc = vpadalq_u16(s2acc, sum2);
+ buf += 16;
+ len--;
+ }
+
+ adacc2 = vpadd_u32(vget_low_u32(adacc), vget_high_u32(adacc));
+ s2acc2 = vpadd_u32(vget_low_u32(s2acc), vget_high_u32(s2acc));
+ as = vpadd_u32(adacc2, s2acc2);
+ s[0] = vget_lane_u32(as, 0);
+ s[1] = vget_lane_u32(as, 1);
+}
+
+static void NEON_handle_tail(uint32_t *pair, const unsigned char *buf,
+ unsigned int len)
+{
+ /* Oldie K&R code integration. */
+ unsigned int i;
+ for (i = 0; i < len; ++i) {
+ pair[0] += buf[i];
+ pair[1] += pair[0];
+ }
+}
+
+extern unsigned long NEON_adler32(unsigned long adler, const unsigned char *buf,
+ const unsigned int len)
+{
+ /* initial Adler-32 value (deferred check for len == 1 speed) */
+ if (!buf)
+ return 1L;
+
+ /* The largest prime smaller than 65536. */
+ const uint32_t M_BASE = 65521;
+ /* This is the threshold where doing accumulation may overflow. */
+ const int M_NMAX = 5552;
+
+ unsigned long sum2;
+ uint32_t pair[2];
+ int n = M_NMAX;
+ unsigned int done = 0;
+ /* Oldie K&R code integration. */
+ unsigned int i;
+
+ /* Split Adler-32 into component sums, it can be supplied by
+ * the caller sites (e.g. in a PNG file).
+ */
+ sum2 = (adler >> 16) & 0xffff;
+ adler &= 0xffff;
+ pair[0] = adler;
+ pair[1] = sum2;
+
+ for (i = 0; i < len; i += n) {
+ if ((i + n) > len)
+ n = len - i;
+
+ if (n < 16)
+ break;
+
+ NEON_accum32(pair, buf + i, n / 16);
+ pair[0] %= M_BASE;
+ pair[1] %= M_BASE;
+
+ done += (n / 16) * 16;
+ }
+
+ /* Handle the tail elements. */
+ if (done < len) {
+ NEON_handle_tail(pair, (buf + done), len - done);
+ pair[0] %= M_BASE;
+ pair[1] %= M_BASE;
+ }
+
+ /* D = B * 65536 + A, see: https://en.wikipedia.org/wiki/Adler-32. */
+ return (pair[1] << 16) | pair[0];
+}
+#endif
|
