gvsafronov/rapidjson
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Added ARM-Neon support for SIMD.SkipWhitespace*

Browse Source 
Change-Id: Iaf210d029758723a7eeb7f28fc10cab7467889a9
Signed-off-by: Jun He <jun.he@arm.com>
This commit is contained in:
Alejandro Martinez 2017-04-11 02:02:15 +00:00 committed by Jun He
parent e6d7247ed9
commit 2291258bb8
10 changed files with 365 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
doc/faq.md
View File

@ -256,7 +256,7 @@ Alternatively, if we don't want to explicitly refer to the root value of `addres
3. What is SIMD? How it is applied in RapidJSON?
[SIMD](http://en.wikipedia.org/wiki/SIMD) instructions can perform parallel computation in modern CPUs. RapidJSON support Intel's SSE2/SSE4.2 to accelerate whitespace skipping. This improves performance of parsing indent formatted JSON. Define `RAPIDJSON_SSE2` or `RAPIDJSON_SSE42` macro to enable this feature. However, running the executable on a machine without such instruction set support will make it crash.
[SIMD](http://en.wikipedia.org/wiki/SIMD) instructions can perform parallel computation in modern CPUs. RapidJSON support Intel's SSE2/SSE4.2 and ARM's Neon to accelerate whitespace/tabspace/carriage-return/line-feed skipping. This improves performance of parsing indent formatted JSON. Define `RAPIDJSON_SSE2`, `RAPIDJSON_SSE42` or `RAPIDJSON_NEON` macro to enable this feature. However, running the executable on a machine without such instruction set support will make it crash.
4. Does it consume a lot of memory?

2
doc/faq.zh-cn.md
View File

@ -257,7 +257,7 @@
3. 什是是 SIMD？它如何用于 RapidJSON？
[SIMD](http://en.wikipedia.org/wiki/SIMD) 指令可以在现代 CPU 中执行并行运算。RapidJSON 支持了 Intel 的 SSE2/SSE4.2 去加速跳过空白字符。在解析含缩进的 JSON 时，这能提升性能。只要定义名为 `RAPIDJSON_SSE2` 或 `RAPIDJSON_SSE42` 的宏，就能启动这个功能。然而，若在不支持这些指令集的机器上执行这些可执行文件，会导致崩溃。
[SIMD](http://en.wikipedia.org/wiki/SIMD) 指令可以在现代 CPU 中执行并行运算。RapidJSON 支持使用 Intel 的 SSE2/SSE4.2 和 ARM 的 Neon 来加速对空白符、制表符、回车符和换行符的过滤处理。在解析含缩进的 JSON 时，这能提升性能。只要定义名为 `RAPIDJSON_SSE2` ，`RAPIDJSON_SSE42` 或 `RAPIDJSON_NEON` 的宏，就能启动这个功能。然而，若在不支持这些指令集的机器上执行这些可执行文件，会导致崩溃。
4. 它会消耗许多内存么？

7
doc/internals.md
View File

@ -183,17 +183,20 @@ void SkipWhitespace(InputStream& s) {
However, this requires 4 comparisons and a few branching for each character. This was found to be a hot spot.
To accelerate this process, SIMD was applied to compare 16 characters with 4 white spaces for each iteration. Currently RapidJSON only supports SSE2 and SSE4.2 instructions for this. And it is only activated for UTF-8 memory streams, including string stream or *in situ* parsing.
To accelerate this process, SIMD was applied to compare 16 characters with 4 white spaces for each iteration. Currently RapidJSON supports SSE2, SSE4.2 and ARM Neon instructions for this. And it is only activated for UTF-8 memory streams, including string stream or *in situ* parsing.
To enable this optimization, need to define `RAPIDJSON_SSE2` or `RAPIDJSON_SSE42` before including `rapidjson.h`. Some compilers can detect the setting, as in `perftest.h`:
To enable this optimization, need to define `RAPIDJSON_SSE2`, `RAPIDJSON_SSE42` or `RAPIDJSON_NEON` before including `rapidjson.h`. Some compilers can detect the setting, as in `perftest.h`:
~~~cpp
// __SSE2__ and __SSE4_2__ are recognized by gcc, clang, and the Intel compiler.
// We use -march=native with gmake to enable -msse2 and -msse4.2, if supported.
// Likewise, __ARM_NEON is used to detect Neon.
#if defined(__SSE4_2__)
# define RAPIDJSON_SSE42
#elif defined(__SSE2__)
# define RAPIDJSON_SSE2
#elif defined(__ARM_NEON)
# define RAPIDJSON_NEON
#endif
~~~

7
doc/internals.zh-cn.md
View File

@ -183,17 +183,20 @@ void SkipWhitespace(InputStream& s) {
但是，这需要对每个字符进行4次比较以及一些分支。这被发现是一个热点。
为了加速这一处理，RapidJSON 使用 SIMD 来在一次迭代中比较16个字符和4个空格。目前 RapidJSON 只支持 SSE2 和 SSE4.2 指令。同时它也只会对 UTF-8 内存流启用，包括字符串流或 *原位* 解析。
为了加速这一处理，RapidJSON 使用 SIMD 来在一次迭代中比较16个字符和4个空格。目前 RapidJSON 支持 SSE2 ， SSE4.2 和 ARM Neon 指令。同时它也只会对 UTF-8 内存流启用，包括字符串流或 *原位* 解析。
你可以通过在包含 `rapidjson.h` 之前定义 `RAPIDJSON_SSE2` 或 `RAPIDJSON_SSE42` 来启用这个优化。一些编译器可以检测这个设置，如 `perftest.h`：
你可以通过在包含 `rapidjson.h` 之前定义 `RAPIDJSON_SSE2` ， `RAPIDJSON_SSE42` 或 `RAPIDJSON_NEON` 来启用这个优化。一些编译器可以检测这个设置，如 `perftest.h`：
~~~cpp
// __SSE2__ 和 __SSE4_2__ 可被 gcc、clang 和 Intel 编译器识别：
// 如果支持的话，我们在 gmake 中使用了 -march=native 来启用 -msse2 和 -msse4.2
// 同样的， __ARM_NEON 被用于识别Neon
#if defined(__SSE4_2__)
# define RAPIDJSON_SSE42
#elif defined(__SSE2__)
# define RAPIDJSON_SSE2
#elif defined(__ARM_NEON)
# define RAPIDJSON_NEON
#endif
~~~

18
include/rapidjson/rapidjson.h
View File

@ -325,17 +325,17 @@
#endif
///////////////////////////////////////////////////////////////////////////////
// RAPIDJSON_SSE2/RAPIDJSON_SSE42/RAPIDJSON_SIMD
// RAPIDJSON_SSE2/RAPIDJSON_SSE42/RAPIDJSON_NEON/RAPIDJSON_SIMD
/*! \def RAPIDJSON_SIMD
\ingroup RAPIDJSON_CONFIG
\brief Enable SSE2/SSE4.2 optimization.
\brief Enable SSE2/SSE4.2/Neon optimization.
RapidJSON supports optimized implementations for some parsing operations
based on the SSE2 or SSE4.2 SIMD extensions on modern Intel-compatible
processors.
based on the SSE2, SSE4.2 or NEon SIMD extensions on modern Intel
or ARM compatible processors.
To enable these optimizations, two different symbols can be defined;
To enable these optimizations, three different symbols can be defined;
\code
// Enable SSE2 optimization.
#define RAPIDJSON_SSE2
@ -344,13 +344,17 @@
#define RAPIDJSON_SSE42
\endcode
\c RAPIDJSON_SSE42 takes precedence, if both are defined.
// Enable ARM Neon optimization.
#define RAPIDJSON_NEON
\endcode
\c RAPIDJSON_SSE42 takes precedence over SSE2, if both are defined.
If any of these symbols is defined, RapidJSON defines the macro
\c RAPIDJSON_SIMD to indicate the availability of the optimized code.
*/
#if defined(RAPIDJSON_SSE2) || defined(RAPIDJSON_SSE42) \
|| defined(RAPIDJSON_DOXYGEN_RUNNING)
|| defined(RAPIDJSON_NEON) || defined(RAPIDJSON_DOXYGEN_RUNNING)
#define RAPIDJSON_SIMD
#endif

264
include/rapidjson/reader.h
View File

@ -33,6 +33,8 @@
#include <nmmintrin.h>
#elif defined(RAPIDJSON_SSE2)
#include <emmintrin.h>
#elif defined(RAPIDJSON_NEON)
#include <arm_neon.h>
#endif
#ifdef _MSC_VER
@ -411,7 +413,92 @@ inline const char *SkipWhitespace_SIMD(const char* p, const char* end) {
return SkipWhitespace(p, end);
}
#endif // RAPIDJSON_SSE2
#elif defined(RAPIDJSON_NEON)
//! Skip whitespace with ARM Neon instructions, testing 16 8-byte characters at once.
inline const char *SkipWhitespace_SIMD(const char* p) {
// Fast return for single non-whitespace
if (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
++p;
else
return p;
// 16-byte align to the next boundary
const char* nextAligned = reinterpret_cast<const char*>((reinterpret_cast<size_t>(p) + 15) & static_cast<size_t>(~15));
while (p != nextAligned)
if (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t')
++p;
else
return p;
const uint8x16_t w0 = vmovq_n_u8(' ');
const uint8x16_t w1 = vmovq_n_u8('\n');
const uint8x16_t w2 = vmovq_n_u8('\r');
const uint8x16_t w3 = vmovq_n_u8('\t');
for (;; p += 16) {
const uint8x16_t s = vld1q_u8(reinterpret_cast<const uint8_t *>(p));
uint8x16_t x = vceqq_u8(s, w0);
x = vorrq_u8(x, vceqq_u8(s, w1));
x = vorrq_u8(x, vceqq_u8(s, w2));
x = vorrq_u8(x, vceqq_u8(s, w3));
x = vmvnq_u8(x); // Negate
x = vrev64q_u8(x); // Rev in 64
uint64_t low = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 0); // extract
uint64_t high = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 1); // extract
if (low == 0) {
if (high != 0) {
int lz =__builtin_clzll(high);;
return p + 8 + (lz >> 3);
}
} else {
int lz = __builtin_clzll(low);;
return p + (lz >> 3);
}
}
}
inline const char *SkipWhitespace_SIMD(const char* p, const char* end) {
// Fast return for single non-whitespace
if (p != end && (*p == ' ' || *p == '\n' || *p == '\r' || *p == '\t'))
++p;
else
return p;
const uint8x16_t w0 = vmovq_n_u8(' ');
const uint8x16_t w1 = vmovq_n_u8('\n');
const uint8x16_t w2 = vmovq_n_u8('\r');
const uint8x16_t w3 = vmovq_n_u8('\t');
for (; p <= end - 16; p += 16) {
const uint8x16_t s = vld1q_u8(reinterpret_cast<const uint8_t *>(p));
uint8x16_t x = vceqq_u8(s, w0);
x = vorrq_u8(x, vceqq_u8(s, w1));
x = vorrq_u8(x, vceqq_u8(s, w2));
x = vorrq_u8(x, vceqq_u8(s, w3));
x = vmvnq_u8(x); // Negate
x = vrev64q_u8(x); // Rev in 64
uint64_t low = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 0); // extract
uint64_t high = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 1); // extract
if (low == 0) {
if (high != 0) {
int lz = __builtin_clzll(high);
return p + 8 + (lz >> 3);
}
} else {
int lz = __builtin_clzll(low);
return p + (lz >> 3);
}
}
return SkipWhitespace(p, end);
}
#endif // RAPIDJSON_NEON
#ifdef RAPIDJSON_SIMD
//! Template function specialization for InsituStringStream
@ -1129,7 +1216,180 @@ private:
is.src_ = is.dst_ = p;
}
#endif
#elif defined(RAPIDJSON_NEON)
// StringStream -> StackStream<char>
static RAPIDJSON_FORCEINLINE void ScanCopyUnescapedString(StringStream& is, StackStream<char>& os) {
const char* p = is.src_;
// Scan one by one until alignment (unaligned load may cross page boundary and cause crash)
const char* nextAligned = reinterpret_cast<const char*>((reinterpret_cast<size_t>(p) + 15) & static_cast<size_t>(~15));
while (p != nextAligned)
if (RAPIDJSON_UNLIKELY(*p == '\"') || RAPIDJSON_UNLIKELY(*p == '\\') || RAPIDJSON_UNLIKELY(static_cast<unsigned>(*p) < 0x20)) {
is.src_ = p;
return;
}
else
os.Put(*p++);
// The rest of string using SIMD
const uint8x16_t s0 = vmovq_n_u8('"');
const uint8x16_t s1 = vmovq_n_u8('\\');
const uint8x16_t s2 = vmovq_n_u8('\b');
const uint8x16_t s3 = vmovq_n_u8(32);
for (;; p += 16) {
const uint8x16_t s = vld1q_u8(reinterpret_cast<const uint8_t *>(p));
uint8x16_t x = vceqq_u8(s, s0);
x = vorrq_u8(x, vceqq_u8(s, s1));
x = vorrq_u8(x, vceqq_u8(s, s2));
x = vorrq_u8(x, vcltq_u8(s, s3));
x = vrev64q_u8(x); // Rev in 64
uint64_t low = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 0); // extract
uint64_t high = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 1); // extract
SizeType length = 0;
bool escaped = false;
if (low == 0) {
if (high != 0) {
unsigned lz = (unsigned)__builtin_clzll(high);;
length = 8 + (lz >> 3);
escaped = true;
}
} else {
unsigned lz = (unsigned)__builtin_clzll(low);;
length = lz >> 3;
escaped = true;
}
if (RAPIDJSON_UNLIKELY(escaped)) { // some of characters is escaped
if (length != 0) {
char* q = reinterpret_cast<char*>(os.Push(length));
for (size_t i = 0; i < length; i++)
q[i] = p[i];
p += length;
}
break;
}
vst1q_u8(reinterpret_cast<uint8_t *>(os.Push(16)), s);
}
is.src_ = p;
}
// InsituStringStream -> InsituStringStream
static RAPIDJSON_FORCEINLINE void ScanCopyUnescapedString(InsituStringStream& is, InsituStringStream& os) {
RAPIDJSON_ASSERT(&is == &os);
(void)os;
if (is.src_ == is.dst_) {
SkipUnescapedString(is);
return;
}
char* p = is.src_;
char *q = is.dst_;
// Scan one by one until alignment (unaligned load may cross page boundary and cause crash)
const char* nextAligned = reinterpret_cast<const char*>((reinterpret_cast<size_t>(p) + 15) & static_cast<size_t>(~15));
while (p != nextAligned)
if (RAPIDJSON_UNLIKELY(*p == '\"') || RAPIDJSON_UNLIKELY(*p == '\\') || RAPIDJSON_UNLIKELY(static_cast<unsigned>(*p) < 0x20)) {
is.src_ = p;
is.dst_ = q;
return;
}
else
*q++ = *p++;
// The rest of string using SIMD
const uint8x16_t s0 = vmovq_n_u8('"');
const uint8x16_t s1 = vmovq_n_u8('\\');
const uint8x16_t s2 = vmovq_n_u8('\b');
const uint8x16_t s3 = vmovq_n_u8(32);
for (;; p += 16, q += 16) {
const uint8x16_t s = vld1q_u8(reinterpret_cast<uint8_t *>(p));
uint8x16_t x = vceqq_u8(s, s0);
x = vorrq_u8(x, vceqq_u8(s, s1));
x = vorrq_u8(x, vceqq_u8(s, s2));
x = vorrq_u8(x, vcltq_u8(s, s3));
x = vrev64q_u8(x); // Rev in 64
uint64_t low = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 0); // extract
uint64_t high = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 1); // extract
SizeType length = 0;
bool escaped = false;
if (low == 0) {
if (high != 0) {
unsigned lz = (unsigned)__builtin_clzll(high);
length = 8 + (lz >> 3);
escaped = true;
}
} else {
unsigned lz = (unsigned)__builtin_clzll(low);
length = lz >> 3;
escaped = true;
}
if (RAPIDJSON_UNLIKELY(escaped)) { // some of characters is escaped
for (const char* pend = p + length; p != pend; ) {
*q++ = *p++;
}
break;
}
vst1q_u8(reinterpret_cast<uint8_t *>(q), s);
}
is.src_ = p;
is.dst_ = q;
}
// When read/write pointers are the same for insitu stream, just skip unescaped characters
static RAPIDJSON_FORCEINLINE void SkipUnescapedString(InsituStringStream& is) {
RAPIDJSON_ASSERT(is.src_ == is.dst_);
char* p = is.src_;
// Scan one by one until alignment (unaligned load may cross page boundary and cause crash)
const char* nextAligned = reinterpret_cast<const char*>((reinterpret_cast<size_t>(p) + 15) & static_cast<size_t>(~15));
for (; p != nextAligned; p++)
if (RAPIDJSON_UNLIKELY(*p == '\"') || RAPIDJSON_UNLIKELY(*p == '\\') || RAPIDJSON_UNLIKELY(static_cast<unsigned>(*p) < 0x20)) {
is.src_ = is.dst_ = p;
return;
}
// The rest of string using SIMD
const uint8x16_t s0 = vmovq_n_u8('"');
const uint8x16_t s1 = vmovq_n_u8('\\');
const uint8x16_t s2 = vmovq_n_u8('\b');
const uint8x16_t s3 = vmovq_n_u8(32);
for (;; p += 16) {
const uint8x16_t s = vld1q_u8(reinterpret_cast<uint8_t *>(p));
uint8x16_t x = vceqq_u8(s, s0);
x = vorrq_u8(x, vceqq_u8(s, s1));
x = vorrq_u8(x, vceqq_u8(s, s2));
x = vorrq_u8(x, vcltq_u8(s, s3));
x = vrev64q_u8(x); // Rev in 64
uint64_t low = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 0); // extract
uint64_t high = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 1); // extract
if (low == 0) {
if (high != 0) {
int lz = __builtin_clzll(high);
p += 8 + (lz >> 3);
break;
}
} else {
int lz = __builtin_clzll(low);
p += lz >> 3;
break;
}
}
is.src_ = is.dst_ = p;
}
#endif // RAPIDJSON_NEON
template<typename InputStream, bool backup, bool pushOnTake>
class NumberStream;

72
include/rapidjson/writer.h
View File

@ -32,6 +32,8 @@
#include <nmmintrin.h>
#elif defined(RAPIDJSON_SSE2)
#include <emmintrin.h>
#elif defined(RAPIDJSON_NEON)
#include <arm_neon.h>
#endif
#ifdef _MSC_VER
@ -619,7 +621,75 @@ inline bool Writer<StringBuffer>::ScanWriteUnescapedString(StringStream& is, siz
is.src_ = p;
return RAPIDJSON_LIKELY(is.Tell() < length);
}
#endif // defined(RAPIDJSON_SSE2) || defined(RAPIDJSON_SSE42)
#elif defined(RAPIDJSON_NEON)
template<>
inline bool Writer<StringBuffer>::ScanWriteUnescapedString(StringStream& is, size_t length) {
if (length < 16)
return RAPIDJSON_LIKELY(is.Tell() < length);
if (!RAPIDJSON_LIKELY(is.Tell() < length))
return false;
const char* p = is.src_;
const char* end = is.head_ + length;
const char* nextAligned = reinterpret_cast<const char*>((reinterpret_cast<size_t>(p) + 15) & static_cast<size_t>(~15));
const char* endAligned = reinterpret_cast<const char*>(reinterpret_cast<size_t>(end) & static_cast<size_t>(~15));
if (nextAligned > end)
return true;
while (p != nextAligned)
if (*p < 0x20 || *p == '\"' || *p == '\\') {
is.src_ = p;
return RAPIDJSON_LIKELY(is.Tell() < length);
}
else
os_->PutUnsafe(*p++);
// The rest of string using SIMD
const uint8x16_t s0 = vmovq_n_u8('"');
const uint8x16_t s1 = vmovq_n_u8('\\');
const uint8x16_t s2 = vmovq_n_u8('\b');
const uint8x16_t s3 = vmovq_n_u8(32);
for (; p != endAligned; p += 16) {
const uint8x16_t s = vld1q_u8(reinterpret_cast<const uint8_t *>(p));
uint8x16_t x = vceqq_u8(s, s0);
x = vorrq_u8(x, vceqq_u8(s, s1));
x = vorrq_u8(x, vceqq_u8(s, s2));
x = vorrq_u8(x, vcltq_u8(s, s3));
x = vrev64q_u8(x); // Rev in 64
uint64_t low = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 0); // extract
uint64_t high = vgetq_lane_u64(reinterpret_cast<uint64x2_t>(x), 1); // extract
SizeType len = 0;
bool escaped = false;
if (low == 0) {
if (high != 0) {
unsigned lz = (unsigned)__builtin_clzll(high);
len = 8 + (lz >> 3);
escaped = true;
}
} else {
unsigned lz = (unsigned)__builtin_clzll(low);
len = lz >> 3;
escaped = true;
}
if (RAPIDJSON_UNLIKELY(escaped)) { // some of characters is escaped
char* q = reinterpret_cast<char*>(os_->PushUnsafe(len));
for (size_t i = 0; i < len; i++)
q[i] = p[i];
p += len;
break;
}
vst1q_u8(reinterpret_cast<uint8_t *>(os_->PushUnsafe(16)), s);
}
is.src_ = p;
return RAPIDJSON_LIKELY(is.Tell() < length);
}
#endif // RAPIDJSON_NEON
RAPIDJSON_NAMESPACE_END

3
test/perftest/perftest.h
View File

@ -24,10 +24,13 @@
// __SSE2__ and __SSE4_2__ are recognized by gcc, clang, and the Intel compiler.
// We use -march=native with gmake to enable -msse2 and -msse4.2, if supported.
// Likewise, __ARM_NEON is used to detect Neon.
#if defined(__SSE4_2__)
# define RAPIDJSON_SSE42
#elif defined(__SSE2__)
# define RAPIDJSON_SSE2
#elif defined(__ARM_NEON)
# define RAPIDJSON_NEON
#endif
#define RAPIDJSON_HAS_STDSTRING 1

2
test/perftest/rapidjsontest.cpp
View File

@ -28,6 +28,8 @@
#define SIMD_SUFFIX(name) name##_SSE2
#elif defined(RAPIDJSON_SSE42)
#define SIMD_SUFFIX(name) name##_SSE42
#elif defined(RAPIDJSON_NEON)
#define SIMD_SUFFIX(name) name##_NEON
#else
#define SIMD_SUFFIX(name) name
#endif

4
test/unittest/simdtest.cpp
View File

@ -21,6 +21,8 @@
# define RAPIDJSON_SSE42
#elif defined(__SSE2__)
# define RAPIDJSON_SSE2
#elif defined(__ARM_NEON)
# define RAPIDJSON_NEON
#endif
#define RAPIDJSON_NAMESPACE rapidjson_simd
@ -41,6 +43,8 @@ using namespace rapidjson_simd;
#define SIMD_SUFFIX(name) name##_SSE2
#elif defined(RAPIDJSON_SSE42)
#define SIMD_SUFFIX(name) name##_SSE42
#elif defined(RAPIDJSON_NEON)
#define SIMD_SUFFIX(name) name##_NEON
#else
#define SIMD_SUFFIX(name) name
#endif