Optimized <cstring> functions. More...

Classes
class	vsmc::CStringNonTemporalThreshold
	The threshold of buffer size above which `memcpy` use non-temporal instructions. More...

Functions
void *	vsmc::memcpy (void dst, const void src, std::size_t n)
	SIMD optimized `memcpy` with non-temporal store for large buffers. More...

void *	vsmc::memcpy_avx (void dst, const void src, std::size_t n)
	AVX optimized `memcpy` with non-temporal store for large buffers. More...

void *	vsmc::memcpy_avx_nt (void dst, const void src, std::size_t n)
	AVX optimized `memcpy` with non-temporal store regardless of size. More...

void *	vsmc::memcpy_nt (void dst, const void src, std::size_t n)
	SIMD optimized `memcpy` with non-temporal store regardless of size. More...

void *	vsmc::memcpy_sse2 (void dst, const void src, std::size_t n)
	SSE2 optimized `memcpy` with non-temporal store for large buffers. More...

void *	vsmc::memcpy_sse2_nt (void dst, const void src, std::size_t n)
	SSE2 optimized `memcpy` with non-temporal store regardless of size. More...

void *	vsmc::memcpy_std (void dst, const void src, std::size_t n)
	Direct call to `std::memcpy` More...

void *	vsmc::memset (void *dst, int ch, std::size_t n)
	SIMD optimized `memset` with non-temporal store for large buffers. More...

void *	vsmc::memset_avx (void *dst, int ch, std::size_t n)
	AVX optimized `memset` with non-temporal store for large buffers. More...

void *	vsmc::memset_avx_nt (void *dst, int ch, std::size_t n)
	AVX optimized `memset` with non-temporal store regardless of size. More...

void *	vsmc::memset_nt (void *dst, int ch, std::size_t n)
	SIMD optimized `memset` with non-temporal store regardless of size. More...

void *	vsmc::memset_sse2 (void *dst, int ch, std::size_t n)
	SSE2 optimized `memset` with non-temporal store for large buffers. More...

void *	vsmc::memset_sse2_nt (void *dst, int ch, std::size_t n)
	SSE2 optimized `memset` with non-temporal store regardless of size. More...

void *	vsmc::memset_std (void *dst, int ch, std::size_t n)
	Direct call to `std::memset` More...

Detailed Description

Optimized <cstring> functions.

This module implement the memcpy, etc., functions in the vsmc namespace. The implementaions are optimzied with SIMD instructions. Three groups of functions are provided.

memcpy_std etc., they simply call std::memcpy etc.
memcpy_sse2 etc., they are avialable if at least SSE2 is supported and are optimized with SSE2 instructions
memcpy_avx etc., they are avialable if at least AVX is supported and are optimized with AVX instructions

There are also generic vsmc::memcpy etc. They dispatch the call based on the following rules.

If AVX is available, then call memcpy_avx etc.
Else if SSE2 is available, then call memcpy_sse2 etc.
Else call memcpy_std.

This dispatch can be done at compile time if the configuration macro VSMC_CSTRING_RUNTIME_DISPATCH is zero. If the macro is non-zero, then it will be done at runtime using CPUID information.

Before using any of these vSMC provided functions. A few factors shall be considered.

For large buffers, vSMC functions use non-temporal store instructions. In this case, the performance is likely to be better than the system library or compiler builtins. On Haswell and Nehalem, about 30% performance gain were observed, though 70% were observed in some situations. A more important benefits is that cache pollution may be avoided in this case. The threshold is set to be half the size of the LLC (last level cache). Note that, only up to level 3 cache is considered. Some newer Intel CPUs have level 4 cache, which is shared with the integrated GPU. This threshold can be changed at compile time by define VSMC_CSTRING_NON_TEMPORAL_THRESHOLD or at runtime via CStringNonTemporalThreshold singleton.
For moderate size buffers (from 1KB upto the non-temporal threshold), the performance is most likely to be only comparable to the system. At worst, 20% perforamnce degrade was observed, though in most cases, it is almost as fast as or slightly faster than system library. To change the threshold,
vsmc::CStringNonTemporalThreshold::instance().set(new_threshold);

To change the maximum level of cache to be considered and reset the threshold accordingly.
vsmc::CStringNonTemporalThreshold::instance().max_level(new_level);
The performance is only tested on limited models of CPUs. As a single person I don't have much resources. In particle, AMD CPUs were not tested at all.

In any case, most systems's standard C library is likely to be optimized enough to suffice most usage situations. And even when there is a noticeable perforamnce difference, unless all the program do is copy and moving memories, the difference is not likely to be big enough to make a difference. And taking caching into consideration, those difference seem in memory dedicated benchmarks might well not exist at all in real programs.

In summary, do some benchmark of real programs before deciding if using these functions are beneficial.

Function Documentation

void* vsmc::memcpy	(	void *	dst,
		const void *	src,
		std::size_t	n
	)

inline

SIMD optimized memcpy with non-temporal store for large buffers.

Definition at line 923 of file cstring.hpp.

void* vsmc::memcpy_avx	(	void *	dst,
		const void *	src,
		std::size_t	n
	)

inline

AVX optimized memcpy with non-temporal store for large buffers.

Definition at line 820 of file cstring.hpp.

void* vsmc::memcpy_avx_nt	(	void *	dst,
		const void *	src,
		std::size_t	n
	)

inline

AVX optimized memcpy with non-temporal store regardless of size.

Note: The destination and source pointers must be aligned to 32 bytes

Definition at line 834 of file cstring.hpp.

void* vsmc::memcpy_nt	(	void *	dst,
		const void *	src,
		std::size_t	n
	)

inline

SIMD optimized memcpy with non-temporal store regardless of size.

Note: The destination and source pointers must be aligned to 32 bytes or 16 bytes without AVX support

Definition at line 964 of file cstring.hpp.

void* vsmc::memcpy_sse2	(	void *	dst,
		const void *	src,
		std::size_t	n
	)

inline

SSE2 optimized memcpy with non-temporal store for large buffers.

Definition at line 792 of file cstring.hpp.

void* vsmc::memcpy_sse2_nt	(	void *	dst,
		const void *	src,
		std::size_t	n
	)

inline

SSE2 optimized memcpy with non-temporal store regardless of size.

Note: The destination and source pointers must be aligned to 16 bytes

Definition at line 806 of file cstring.hpp.

void* vsmc::memcpy_std	(	void *	dst,
		const void *	src,
		std::size_t	n
	)

inline

Direct call to std::memcpy

Definition at line 780 of file cstring.hpp.

void* vsmc::memset	(	void *	dst,
		int	ch,
		std::size_t	n
	)

inline

SIMD optimized memset with non-temporal store for large buffers.

Definition at line 906 of file cstring.hpp.

void* vsmc::memset_avx	(	void *	dst,
		int	ch,
		std::size_t	n
	)

inline

AVX optimized memset with non-temporal store for large buffers.

Definition at line 815 of file cstring.hpp.

void* vsmc::memset_avx_nt	(	void *	dst,
		int	ch,
		std::size_t	n
	)

inline

AVX optimized memset with non-temporal store regardless of size.

Note: The destination pointer must be aligned to 32 bytes

Definition at line 827 of file cstring.hpp.

void* vsmc::memset_nt	(	void *	dst,
		int	ch,
		std::size_t	n
	)

inline

SIMD optimized memset with non-temporal store regardless of size.

Note: The destination pointer must be aligned to 32 bytes or 16 bytes without AVX support

Definition at line 944 of file cstring.hpp.

void* vsmc::memset_sse2	(	void *	dst,
		int	ch,
		std::size_t	n
	)

inline

SSE2 optimized memset with non-temporal store for large buffers.

Definition at line 787 of file cstring.hpp.

void* vsmc::memset_sse2_nt	(	void *	dst,
		int	ch,
		std::size_t	n
	)

inline

SSE2 optimized memset with non-temporal store regardless of size.

Note: The destination pointer must be aligned to 16 bytes

Definition at line 799 of file cstring.hpp.

void* vsmc::memset_std	(	void *	dst,
		int	ch,
		std::size_t	n
	)

inline

Direct call to std::memset

Definition at line 775 of file cstring.hpp.

Classes

Functions

Detailed Description

Function Documentation