Syoyo Fujita syoyo

💗

ray tracing

No ray tracing no life.

syoyo / gist:d80cb6f9936aa5f290da

Created April 18, 2015 16:45

exp(x) using trigonometric function in HPC-ACE

syoyo / gist:d9fb3092860614861efb

Created April 24, 2015 12:51

SIMD class

	PRE_ALIGN(128) struct __vec8_d {
	__vec8_d() { }

	FORCEINLINE __vec8_d(const double v0) {
	u.v0 = _mm_set_pd(v0, v0);
	u.v1 = _mm_set_pd(v0, v0);
	u.v2 = _mm_set_pd(v0, v0);
	u.v3 = _mm_set_pd(v0, v0);
	}

syoyo / gist:a8ba3b6fb1f2d51e4fff

Created May 4, 2015 16:15

1/x with -Kfast,nounroll,noswp

syoyo / gist:78f6ddb6f99deb49f8ae

Created May 6, 2015 15:41

fmath performance(2015/05/06)


	# K frontend
	$ FCCpx -Kfast,nounroll,noswp exp_bench.cpp fmath.cpp
	exp_bench.cpp:
	fmath.cpp:
	"fmath.cpp", line 30: warning: variable "fmath::local::LOG_TABLE_SIZE" was declared but never referenced
	const size_t LOG_TABLE_SIZE = 12;
	^

syoyo / gist:861c3d78ffc6ac5aabd0

Created May 7, 2015 10:03

disassemble of double fmath::exp(double) in SPARC/HPC-ACE

syoyo / gist:1516aa3e8e5489871fdd

Created May 8, 2015 06:59

doubl2 exp(double2) performance on SPARC/HPC-ACE

	#include <cmath>
	#include <emmintrin.h>

	#define _mm_set1_pd(x) _mm_set_pd((x), (x))

	#define FORCE_INLINE __attriabute__((force_inline))

	// Compute exp(x) using trigonometric function.
	inline __m128d exp_tri(__m128d v)
	{

syoyo / gist:9484d27be95e3303789b

Created June 8, 2015 14:15

expapprox() compiled for Parallella Epiphany

	// code From http://gallium.inria.fr/blog/fast-vectorizable-math-approx/

	$ e-gcc compile options: -O3 -mno-soft-cmpsf -mcmove -mfp-mode=truncate

	// 73 clocks
	00000f40 <_expapprox>:
	f40: 200b 0002 mov r1,0x0
	f44: 476b 0aa2 mov r2,0xaa3b
	f48: 470b 14b2 movt r2,0x4b38
	f4c: 2fcb 14e2 movt r1,0x4e7e

syoyo / gist:ef68a9c5b46b040e88db

Created June 12, 2015 04:24

exp() approximate function on Epiphany.

	The clock cycle count for "expf()" (reference) is 141645.

	The clock cycle count for "expapprox()" is 74.

	The clock cycle count for "expapprox4()" is 127 (/4 = 31).

	// GCC
	#define RESTRICT __restrict__

	// Disable range check makes faster evaluation of exp().

syoyo / gist:f98f95206c9418588645

Created June 25, 2015 12:14

new_exp on Epiphany.

syoyo / gist:956846067e49eb42c4b1

Created June 29, 2015 10:05

new_exp, new_exp4 sim