+++ /dev/null
-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
-
---- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -7,6 +7,7 @@ set(VERSION "1.2.12")
-
- 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)
-
---- a/adler32.c
-+++ b/adler32.c
-@@ -136,7 +136,14 @@ uLong ZEXPORT adler32(adler, buf, len)
- const Bytef *buf;
- uInt len;
- {
-+#ifdef ARMv8
-+# pragma message("Using NEON-ized Adler32.")
-+unsigned long NEON_adler32(unsigned long adler, const unsigned char *buf,
-+ const unsigned int len);
-+ return NEON_adler32(adler, buf, len);
-+#else
- return adler32_z(adler, buf, len);
-+#endif
- }
-
- /* ========================================================================= */
---- a/contrib/README.contrib
-+++ b/contrib/README.contrib
-@@ -8,6 +8,9 @@ ada/ by Dmitriy Anisimkov <anisim
- Support for Ada
- See http://zlib-ada.sourceforge.net/
-
-+arm/ by Adenilson Cavalcanti <cavalcantii@chromium.org>
-+ ARM optimizations (NEON and ARMv8 code).
-+
- blast/ by Mark Adler <madler@alumni.caltech.edu>
- Decompressor for output of PKWare Data Compression Library (DCL)
-
---- /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
3 files changed, 335 insertions(+), 62 deletions(-)
create mode 100644 contrib/arm/chunkcopy.h
-diff --git a/contrib/arm/chunkcopy.h b/contrib/arm/chunkcopy.h
-new file mode 100644
-index 00000000..2d6fd6f9
--- /dev/null
+++ b/contrib/arm/chunkcopy.h
@@ -0,0 +1,279 @@
+#undef Z_RESTRICT
+
+#endif /* CHUNKCOPY_H */
-diff --git a/contrib/arm/inffast.c b/contrib/arm/inffast.c
-index 0dbd1dbc..f7f50071 100644
--- a/contrib/arm/inffast.c
+++ b/contrib/arm/inffast.c
@@ -7,6 +7,7 @@
#ifdef ASMINF
# pragma message("Assembler code may have bugs -- use at your own risk")
-@@ -57,6 +58,7 @@ unsigned start; /* inflate()'s starting value for strm->avail_out */
+@@ -57,6 +58,7 @@ unsigned start; /* inflate()'s s
unsigned char FAR *out; /* local strm->next_out */
unsigned char FAR *beg; /* inflate()'s initial strm->next_out */
unsigned char FAR *end; /* while out < end, enough space available */
#ifdef INFLATE_STRICT
unsigned dmax; /* maximum distance from zlib header */
#endif
-@@ -84,12 +86,13 @@ unsigned start; /* inflate()'s starting value for strm->avail_out */
+@@ -84,12 +86,13 @@ unsigned start; /* inflate()'s s
out = strm->next_out;
beg = out - (start - strm->avail_out);
end = out + (strm->avail_out - 257);
window = state->window;
hold = state->hold;
bits = state->bits;
-@@ -197,70 +200,51 @@ unsigned start; /* inflate()'s starting value for strm->avail_out */
+@@ -197,70 +200,51 @@ unsigned start; /* inflate()'s s
#endif
}
from = window;
- } while (--op);
- from = out - dist; /* rest from output */
- }
-+ out = chunkcopy_safe(out, from, op, limit);
-+ from = window; /* more from start of window */
-+ op = wnext;
-+ /* This (rare) case can create a situation where
-+ the first chunkcopy below must be checked.
-+ */
- }
- }
+- }
+- }
- else { /* contiguous in window */
- from += wnext - op;
- if (op < len) { /* some from window */
- *out++ = *from++;
- if (len > 1)
- *out++ = *from++;
++ out = chunkcopy_safe(out, from, op, limit);
++ from = window; /* more from start of window */
++ op = wnext;
++ /* This (rare) case can create a situation where
++ the first chunkcopy below must be checked.
++ */
++ }
++ }
+ if (op < len) { /* still need some from output */
+ out = chunkcopy_safe(out, from, op, limit);
+ len -= op;
}
}
else if ((op & 64) == 0) { /* 2nd level distance code */
-diff --git a/contrib/arm/inflate.c b/contrib/arm/inflate.c
-index ac333e8c..e40322c3 100644
--- a/contrib/arm/inflate.c
+++ b/contrib/arm/inflate.c
@@ -84,6 +84,7 @@
projects / openwrt / openwrt.git / commitdiff