ARM CPU core 毎の浮動小数点演算速度の比較 (VFP/NEON)

計測アプリを公開しました。

命令(AArch32)	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)	(12)	(13)	(14)	命令(AArch64)
	iPhone3GS	iPod touch4	Ziio7	LTNote	iPad3	Nexus7	EVO 3D	iPhone5	iPad4	HTL21	Nexus10	iPhone5s	iPhone5s	Kin HDX7
	C-A8	C-A8	C-A8	C-A9	C-A9	C-A9	Scorpion	Swift	Swift	Krait	C-A15	Cyclone	Cyclone	Krait4
	S5PC100	A4	ZMS-08	Tegra2	A5X	Tegra3	MSM8660	A6	A6X	APQ8064	Exynos5D	A7	A7	MSM8974
	ARMv7	ARMv7	ARMv7	ARMv7	ARMv7	ARMv7	ARMv7	ARMv7	ARMv7	ARMv7	ARMv7	ARMv8 A32	ARMv8 A64	ARMv7
	VFPv3	VFPv3	VFPv3	VFPv3	VFPv3	VFPv3	VFPv3	VFPv4	VFPv4	VFPv4	VFPv4	VFPv4	AdvNEON	VFPv4
	0.6GHz	0.8GHz	1.0GHz	1.0GHz	1.0GHz	1.2GHz	1.2GHz	1.3GHz	1.4GHz	1.5GHz	1.7GHz	1.3GHz?	1.3GHz?	2.2GHz
a:m44 vmla_A Q	9.907	7.725	7.387	—–	4.784	3.959	2.859	1.293	1.204	1.337	0.619	0.700	—–	0.661
b:m44 vmla_B Q	5.874	4.200	4.542	—–	2.408	2.002	1.136	1.359	1.266	0.931	0.569	0.670	—–	0.542
c:m44 vmla_A D	9.816	7.052	7.331	—–	4.781	3.980	3.053	1.669	1.554	1.889	0.557	0.649	—–	0.888
d:m44 vmla_B D	5.894	4.162	4.528	—–	2.406	2.003	1.434	1.329	1.238	1.532	0.568	0.745	—–	0.768
A:m44 vfma_A Q	—–	—–	—–	—–	—–	—–	—–	1.632	1.519	1.882	0.746	0.707	0.692	1.178	fmla v AQ
B:m44 vfma_B Q	—–	—–	—–	—–	—–	—–	—–	1.594	1.484	0.695	0.840	0.699	0.696	0.463	fmla v BQ
e:fadds A	50.964	36.793	49.616	4.611	4.010	3.343	3.383	3.090	2.878	2.774	2.383	3.551	1.043	1.864	fadd s A
f:fmuls A	49.512	36.148	55.088	4.513	4.010	3.337	3.383	3.167	2.953	2.747	2.369	3.475	1.548	1.867	fmul s A
g:fmacs A	78.228	56.711	99.153	4.310	4.012	3.337	3.379	6.180	5.757	5.574	2.956	3.480	—–	2.052
h:vfma.f32 A	—–	—–	—–	—–	—–	—–	—–	6.180	5.756	2.747	2.957	3.480	3.185	1.864	fmadd s A
i:vadd.f32 D A	7.331	5.136	5.509	—–	4.111	3.426	3.377	3.091	2.877	2.762	1.183	1.031	1.031	1.866	fadd.2s D A
j:vmul.f32 D A	7.044	5.134	5.511	—–	4.110	3.421	3.383	3.168	2.950	2.746	1.478	1.545	1.545	1.864	fmul.2s D A
k:vmla.f32 D A	7.831	5.891	6.195	—–	4.512	3.792	3.380	3.166	2.951	5.604	1.480	1.567	—–	2.051
o:vfma.f32 D A	—–	—–	—–	—–	—–	—–	—–	3.167	2.494	2.833	1.479	1.574	1.753	1.871	fmla.2s D A
l:vadd.f32 Q A	14.065	10.339	11.016	—–	8.023	6.688	3.377	3.090	2.878	2.801	2.365	1.031	1.039	1.872	fadd.4s Q A
m:vmul.f32 Q A	14.413	10.302	11.016	—–	8.022	6.681	3.384	3.166	2.952	2.761	2.364	1.548	1.548	1.879	fmul.4s Q A
n:vmla.f32 Q A	14.381	10.287	11.035	—–	8.025	6.681	3.380	3.167	2.950	5.606	2.367	1.574	—–	2.059
N:vfma.f32 Q A	—–	—–	—–	—–	—–	—–	—–	—–	—–	—–	—–	—–	1.696	—–	fmla.4s Q A
p:fadds B	49.628	36.236	50.921	4.611	4.014	3.347	5.917	6.181	5.756	3.467	2.956	6.953	3.663	—–	fadd s B
q:fmuls B	49.711	36.011	89.488	5.315	5.013	4.195	5.917	6.180	5.756	3.556	3.558	6.652	3.296	—–	fmul s B
r:fmacs B	78.259	56.706	172.132	8.033	8.023	6.688	8.451	12.361	11.514	6.298	5.912	9.867	—–	—–
s:vfma.f32 B	—–	—–	—–	—–	—–	—–	—–	12.363	11.513	3.430	5.910	9.859	3.292	—–	fmadd s B
t:vadd.f32 D B	7.280	5.139	5.507	—–	4.113	3.421	5.916	3.090	2.881	3.529	2.958	3.663	3.643	1.865	fadd.2s D B
u:vmul.f32 D B	7.044	5.140	5.511	—–	4.118	3.422	5.073	3.169	2.949	3.447	2.364	3.114	3.289	2.339	fmul.2s D B
v:vmla.f32 D B	16.371	11.661	12.392	—–	9.027	7.561	8.451	6.180	5.755	6.293	4.728	6.185	—–	3.773
z:vfma.f32 D B	—–	—–	—–	—–	—–	—–	—–	6.181	5.755	3.437	4.730	6.188	6.237	2.340	fmla.2s D B
w:vadd.f32 Q B	14.209	10.354	11.020	—–	8.021	6.705	5.916	3.090	2.879	3.457	2.961	3.659	3.641	1.875	fadd.4s Q B
x:vmul.f32 Q B	13.879	10.365	11.016	—–	8.029	6.683	5.074	3.167	2.950	3.428	2.363	3.101	3.276	2.340	fmul.4s Q B
y:vmla.f32 Q B	16.026	11.578	12.394	—–	9.026	7.532	8.457	6.179	5.759	6.372	4.729	6.199	—–	3.746
Y:vfma.f32 Q B	—–	—–	—–	—–	—–	—–	—–	—–	—–	—–	—–	—–	6.226	—–	fmla.4s Q B

↑数値は実行時間(秒) 数値が小さい方が速い
すべて単精度 32bit float の演算です。

id	device	SoC	CPU core	core	clock	arch	VFP/SIMD	OS
(1)	Apple iPhone 3GS	S5PC100	Cortex-A8	1	0.6GHz	ARMv7A	VFPv3-D32 + NEON	i6.1
(2)	Apple iPod touch 4	A4	Cortex-A8	1	0.8GHz	ARMv7A	VFPv3-D32 + NEON	i6.1
(3)	Creative Ziio7	ZMS-08	Cortex-A8	1	1.0GHz	ARMv7A	VFPv3-D32 + NEON	A2.2
(4)	Life Touch Note NA70W	Tegra250	Cortex-A9	2	1.0GHz	ARMv7A	VFPv3-D16	A2.2
(5)	Apple iPad 3	A5X	Cortex-A9	2	1.0GHz	ARMv7A	VFPv3-D32 + NEON	i6.1
(6)	ASUS Nexus 7	Tegra 3	Cortex-A9	4	1.2GHz	ARMv7A	VFPv3-D32 + NEON	A4.2
(7)	HTC EVO 3D ISW12HT	MSM8660	Scorpion	2	1.2GHz	ARMv7A	VFPv3-D32 + NEON	A4.0
(8)	Apple iPhone 5	A6	Swift	2	1.3GHz	ARMv7A	VFPv4-D32 + NEON	i6.1
(9)	Apple iPad 4	A6X	Swift	2	1.4GHz	ARMv7A	VFPv4-D32 + NEON	i6.1
(10)	HTC J butterfly HTL21	APQ8064	Krait	4	1.5GHz	ARMv7A	VFPv4-D32 + NEON	A4.1
(11)	Samsung Nexus 10	Exynos5D	Cortex-A15	2	1.7GHz	ARMv7A	VFPv4-D32 + NEON	A4.2
(12)	Apple iPhone 5s	A7	Cyclone	2	1.3GHz	ARMv8A + AArch32	VFPv4-D32 + NEON	i7.0
(13)	Apple iPhone 5s	A7	Cyclone	2	1.3GHz	ARMv8A + AArch64	Advanced NEON	i7.0
(14)	Amazon Kindle HDX 7	MSM8974	Krait 400	4	2.2GHz	ARMv7A	VFPv4-D32 + NEON	A4.2

テスト項目の詳細は下記の解説記事を参照 (a:～z:)

ARM CPU core 毎の浮動小数点演算速度の比較 (VFP/NEON)

解説