Optimize the performance of sum by using universal intrinsics

[Qiyu8]

Introduction

Inspired by #2887, we can take advantage of universal intrinsics, Compare with develop branch, the new implementation has about 13x faster in AVX2, about 4x faster in SSE2, about 6x faster in NEON.

Benchmark

Each test has run for 10 loops and then take the average result, The radio is the MFlops division of Baseline and optimized branch(The bigger the radio is, the better performance is achieved).

• X86 BaseLine

data size(10^3)	develop(float)
4000	568.87 MFlops 0.014063 sec
8000	564.95 MFlops 0.028321 sec

• X86 loop unrolling

data size(10^3)	develop(float)	ratio
4000	1820.37 MFlops 0.004395 sec	3.19
8000	1883.19 MFlops 0.008496 sec	3.33

• X86-SSE2 enabled

  data size(10^3)	usimd-sum(float)	ratio
  4000	2642.44 MFlops 0.003027 sec	4.64
  8000	2642.40 MFlops 0.006055 sec	4.67

• X86-AVX2 enabled

  data size(10^3)	usimd-sum(float)	ratio
  4000	7447.40 MFlops 0.001074 sec	13.09
  8000	5649.32 MFlops 0.002832 sec	9.99

• ARM BaseLine

  data size(10^3)	develop(float)
  4000	338.97 MFlops 0.023601 sec
  8000	338.20 MFlops 0.047309 sec

• ARM loop unrolling

data size	usimd-sum(float)	ratio
4000	1007.34 MFlops 0.007942 sec	2.97
8000	1009.55 MFlops 0.015849 sec	2.98

• ARM-Neon enabled

  data size	usimd-sum(float)	ratio
  4000	2151.17 MFlops 0.003719 sec	6.34
  8000	2154.59 MFlops 0.007426 sec	6.37

System Info

	Arm	x86
Hardware	KunPeng
Processor	ARMv8 2.6GMHZ 8 processors	Intel(R) Xeon(R) Gold 6161 CPU @ 2.20GHz
OS	Linux ecs-9d50 4.19.36-vhulk1905.1.0.h276.eulerosv2r8.aarch64	Windows Server 2008 R2 Enterprise
Compiler	gcc (GCC) 7.3.0	MSVC14.06

[martin-frbg]

Looks good to me, thanks

[martin-frbg]

Oops - that code is single-precision only, right ? arm/sum.c gets built for both SSUM and DSUM (where the -DDOUBLE automagically changes the FLOAT type to double instead of float)