F# Tensor 0.4.9 Hostname: MINT2WIN CPU: Intel(R) Xeon(R) CPU E5-1650 v2 @ 3.50GHz GPU: nVidia Quadro K2200 OS: Microsoft Windows 10.0.17134 Runtime: .NET SDK 2.1.300-preview2-008533 Time: 2018-05-01 00:00:00Z		NumPy 1.14.0 Hostname: MINT2WIN CPU: Intel(R) Xeon(R) CPU E5-1650 v2 @ 3.50GHz GPU: nVidia Quadro K2200 OS: Microsoft Windows 10.0.17134 Runtime: Anaconda3 5.1.0 Time: 2018-05-01 00:00:00Z
Function	Shape	Type	Device	Mean	StdDev	Mean	StdDev
zeros	1000x1000	bool	Host	0.08 ms	±0.01 ms	—	—
		int32	Host	0.17 ms	±0.00 ms	0.27 ms	±0.01 ms
			Cuda	0.60 ms	±0.02 ms	—	—
		int64	Host	0.41 ms	±0.00 ms	0.56 ms	±0.01 ms
			Cuda	0.67 ms	±0.04 ms	—	—
		single	Host	0.16 ms	±0.01 ms	0.27 ms	±0.00 ms
			Cuda	0.61 ms	±0.02 ms	—	—
		double	Host	0.42 ms	±0.00 ms	0.58 ms	±0.01 ms
			Cuda	0.72 ms	±0.09 ms	—	—
	2000x2000	bool	Host	0.17 ms	±0.00 ms	—	—
		int32	Host	1.07 ms	±0.05 ms	1.59 ms	±0.02 ms
			Cuda	1.13 ms	±0.03 ms	—	—
		int64	Host	2.62 ms	±0.05 ms	3.56 ms	±0.04 ms
			Cuda	2.46 ms	±0.74 ms	—	—
		single	Host	1.13 ms	±0.03 ms	1.59 ms	±0.02 ms
			Cuda	1.14 ms	±0.02 ms	—	—
		double	Host	2.69 ms	±0.24 ms	3.56 ms	±0.04 ms
			Cuda	2.35 ms	±0.64 ms	—	—
	4000x4000	bool	Host	1.09 ms	±0.02 ms	—	—
		int32	Host	2.35 ms	±0.09 ms	0.01 ms	±0.00 ms
			Cuda	3.79 ms	±1.60 ms	—	—
		int64	Host	4.64 ms	±0.15 ms	0.01 ms	±0.00 ms
			Cuda	9.13 ms	±1.35 ms	—	—
		single	Host	2.46 ms	±0.06 ms	0.01 ms	±0.00 ms
			Cuda	5.06 ms	±0.93 ms	—	—
		double	Host	5.15 ms	±0.14 ms	0.01 ms	±0.00 ms
			Cuda	8.41 ms	±3.11 ms	—	—
ones	1000x1000	int32	Host	2.05 ms	±0.02 ms	0.28 ms	±0.00 ms
			Cuda	0.86 ms	±0.04 ms	—	—
		int64	Host	4.26 ms	±0.07 ms	0.66 ms	±0.01 ms
			Cuda	0.65 ms	±0.01 ms	—	—
		single	Host	2.14 ms	±0.05 ms	0.28 ms	±0.00 ms
			Cuda	0.64 ms	±0.04 ms	—	—
		double	Host	4.18 ms	±0.07 ms	0.66 ms	±0.01 ms
			Cuda	0.84 ms	±0.03 ms	—	—
	2000x2000	int32	Host	8.87 ms	±0.13 ms	1.60 ms	±0.02 ms
			Cuda	1.12 ms	±0.02 ms	—	—
		int64	Host	18.44 ms	±0.54 ms	3.54 ms	±0.01 ms
			Cuda	2.32 ms	±0.40 ms	—	—
		single	Host	8.81 ms	±0.04 ms	1.61 ms	±0.02 ms
			Cuda	1.11 ms	±0.02 ms	—	—
		double	Host	18.91 ms	±0.32 ms	3.54 ms	±0.01 ms
			Cuda	2.23 ms	±0.07 ms	—	—
	4000x4000	int32	Host	30.20 ms	±0.03 ms	28.85 ms	±0.03 ms
			Cuda	5.52 ms	±1.47 ms	—	—
		int64	Host	61.23 ms	±0.14 ms	61.94 ms	±0.06 ms
			Cuda	10.67 ms	±0.68 ms	—	—
		single	Host	31.15 ms	±0.07 ms	28.86 ms	±0.02 ms
			Cuda	3.86 ms	±1.24 ms	—	—
		double	Host	60.22 ms	±0.12 ms	61.97 ms	±0.06 ms
			Cuda	8.29 ms	±2.86 ms	—	—
arange	1000x1000	int32	Host	9.23 ms	±0.40 ms	—	—
			Cuda	0.64 ms	±0.04 ms	—	—
		int64	Host	11.70 ms	±0.32 ms	—	—
			Cuda	0.64 ms	±0.03 ms	—	—
		single	Host	9.71 ms	±0.06 ms	—	—
			Cuda	0.56 ms	±0.01 ms	—	—
		double	Host	11.40 ms	±0.21 ms	—	—
			Cuda	0.80 ms	±0.02 ms	—	—
	2000x2000	int32	Host	35.72 ms	±1.65 ms	—	—
			Cuda	1.21 ms	±0.05 ms	—	—
		int64	Host	43.38 ms	±1.18 ms	—	—
			Cuda	2.29 ms	±0.03 ms	—	—
		single	Host	37.54 ms	±0.38 ms	—	—
			Cuda	1.15 ms	±0.04 ms	—	—
		double	Host	43.32 ms	±0.02 ms	—	—
			Cuda	2.33 ms	±0.06 ms	—	—
	4000x4000	int32	Host	149.09 ms	±3.17 ms	—	—
			Cuda	3.65 ms	±0.98 ms	—	—
		int64	Host	169.28 ms	±0.17 ms	—	—
			Cuda	10.61 ms	±0.45 ms	—	—
		single	Host	148.52 ms	±0.11 ms	—	—
			Cuda	3.13 ms	±2.01 ms	—	—
		double	Host	165.32 ms	±0.27 ms	—	—
			Cuda	10.82 ms	±0.34 ms	—	—
copy	1000x1000	bool	Host	0.29 ms	±0.01 ms	—	—
			Cuda	0.74 ms	±0.05 ms	—	—
		int32	Host	2.38 ms	±0.04 ms	0.38 ms	±0.01 ms
			Cuda	0.87 ms	±0.03 ms	—	—
		int64	Host	5.19 ms	±0.11 ms	0.79 ms	±0.01 ms
			Cuda	1.02 ms	±0.04 ms	—	—
		single	Host	2.37 ms	±0.05 ms	0.38 ms	±0.01 ms
			Cuda	0.76 ms	±0.02 ms	—	—
		double	Host	5.26 ms	±0.28 ms	0.80 ms	±0.01 ms
			Cuda	0.96 ms	±0.05 ms	—	—
	2000x2000	bool	Host	2.46 ms	±0.07 ms	—	—
			Cuda	1.21 ms	±0.04 ms	—	—
		int32	Host	10.34 ms	±0.18 ms	1.77 ms	±0.01 ms
			Cuda	1.26 ms	±0.04 ms	—	—
		int64	Host	18.47 ms	±0.10 ms	3.60 ms	±0.01 ms
			Cuda	2.01 ms	±0.11 ms	—	—
		single	Host	10.58 ms	±0.30 ms	1.82 ms	±0.03 ms
			Cuda	1.33 ms	±0.05 ms	—	—
		double	Host	18.35 ms	±0.04 ms	3.65 ms	±0.01 ms
			Cuda	2.08 ms	±0.10 ms	—	—
	4000x4000	bool	Host	10.86 ms	±0.49 ms	—	—
			Cuda	4.01 ms	±1.23 ms	—	—
		int32	Host	37.76 ms	±0.02 ms	40.94 ms	±0.02 ms
			Cuda	5.53 ms	±1.02 ms	—	—
		int64	Host	75.04 ms	±0.22 ms	82.02 ms	±0.04 ms
			Cuda	10.66 ms	±0.65 ms	—	—
		single	Host	37.39 ms	±0.07 ms	40.95 ms	±0.04 ms
			Cuda	3.55 ms	±0.21 ms	—	—
		double	Host	75.84 ms	±0.21 ms	82.04 ms	±0.02 ms
			Cuda	10.85 ms	±0.35 ms	—	—
negate	1000x1000	bool	Cuda	0.87 ms	±0.04 ms	—	—
		int32	Host	2.36 ms	±0.05 ms	0.46 ms	±0.01 ms
			Cuda	0.97 ms	±0.05 ms	—	—
		int64	Host	5.09 ms	±0.19 ms	1.27 ms	±0.01 ms
			Cuda	1.04 ms	±0.04 ms	—	—
		single	Host	2.32 ms	±0.03 ms	0.45 ms	±0.01 ms
			Cuda	0.90 ms	±0.04 ms	—	—
		double	Host	4.96 ms	±0.31 ms	1.28 ms	±0.01 ms
			Cuda	0.95 ms	±0.05 ms	—	—
	2000x2000	bool	Cuda	1.51 ms	±0.06 ms	—	—
		int32	Host	10.39 ms	±0.54 ms	2.78 ms	±0.01 ms
			Cuda	1.52 ms	±0.03 ms	—	—
		int64	Host	17.69 ms	±0.16 ms	5.85 ms	±0.76 ms
			Cuda	2.11 ms	±0.16 ms	—	—
		single	Host	9.86 ms	±0.29 ms	2.88 ms	±0.02 ms
			Cuda	1.55 ms	±0.07 ms	—	—
		double	Host	17.61 ms	±0.07 ms	5.65 ms	±0.01 ms
			Cuda	2.48 ms	±0.16 ms	—	—
	4000x4000	bool	Cuda	4.21 ms	±0.14 ms	—	—
		int32	Host	35.61 ms	±0.10 ms	33.18 ms	±0.10 ms
			Cuda	4.30 ms	±0.12 ms	—	—
		int64	Host	71.59 ms	±0.95 ms	65.78 ms	±0.04 ms
			Cuda	10.94 ms	±0.36 ms	—	—
		single	Host	36.18 ms	±0.24 ms	33.74 ms	±0.07 ms
			Cuda	4.88 ms	±0.23 ms	—	—
		double	Host	71.45 ms	±1.32 ms	66.81 ms	±0.06 ms
			Cuda	11.16 ms	±0.62 ms	—	—
add	1000x1000	bool	Cuda	1.02 ms	±0.06 ms	—	—
		int32	Host	2.61 ms	±0.08 ms	0.78 ms	±0.01 ms
			Cuda	1.01 ms	±0.04 ms	—	—
		int64	Host	6.03 ms	±0.37 ms	1.97 ms	±0.11 ms
			Cuda	1.15 ms	±0.04 ms	—	—
		single	Host	2.61 ms	±0.07 ms	0.79 ms	±0.01 ms
			Cuda	1.02 ms	±0.04 ms	—	—
		double	Host	5.83 ms	±0.18 ms	1.93 ms	±0.01 ms
			Cuda	1.15 ms	±0.05 ms	—	—
	2000x2000	bool	Cuda	1.82 ms	±0.07 ms	—	—
		int32	Host	11.11 ms	±0.42 ms	3.89 ms	±0.01 ms
			Cuda	1.81 ms	±0.11 ms	—	—
		int64	Host	19.92 ms	±0.21 ms	8.27 ms	±0.81 ms
			Cuda	2.58 ms	±0.06 ms	—	—
		single	Host	11.06 ms	±0.33 ms	3.88 ms	±0.01 ms
			Cuda	1.80 ms	±0.05 ms	—	—
		double	Host	19.54 ms	±0.32 ms	7.81 ms	±0.01 ms
			Cuda	2.54 ms	±0.09 ms	—	—
	4000x4000	bool	Cuda	5.11 ms	±0.12 ms	—	—
		int32	Host	39.14 ms	±0.10 ms	36.95 ms	±0.08 ms
			Cuda	5.06 ms	±0.04 ms	—	—
		int64	Host	82.45 ms	±4.69 ms	73.74 ms	±0.02 ms
			Cuda	9.79 ms	±1.28 ms	—	—
		single	Host	39.32 ms	±0.23 ms	37.40 ms	±0.08 ms
			Cuda	4.95 ms	±0.07 ms	—	—
		double	Host	76.73 ms	±2.82 ms	74.86 ms	±0.02 ms
			Cuda	11.85 ms	±3.39 ms	—	—
subtract	1000x1000	bool	Cuda	1.01 ms	±0.05 ms	—	—
		int32	Host	2.81 ms	±0.16 ms	0.78 ms	±0.01 ms
			Cuda	0.98 ms	±0.03 ms	—	—
		int64	Host	5.74 ms	±0.22 ms	1.95 ms	±0.11 ms
			Cuda	1.12 ms	±0.04 ms	—	—
		single	Host	2.54 ms	±0.02 ms	0.79 ms	±0.01 ms
			Cuda	0.98 ms	±0.01 ms	—	—
		double	Host	5.91 ms	±0.23 ms	1.92 ms	±0.01 ms
			Cuda	1.14 ms	±0.05 ms	—	—
	2000x2000	bool	Cuda	1.78 ms	±0.04 ms	—	—
		int32	Host	10.88 ms	±0.32 ms	3.91 ms	±0.02 ms
			Cuda	1.78 ms	±0.09 ms	—	—
		int64	Host	19.54 ms	±0.30 ms	8.27 ms	±0.81 ms
			Cuda	2.56 ms	±0.03 ms	—	—
		single	Host	10.81 ms	±0.23 ms	3.90 ms	±0.01 ms
			Cuda	1.79 ms	±0.07 ms	—	—
		double	Host	19.31 ms	±0.04 ms	7.81 ms	±0.01 ms
			Cuda	2.57 ms	±0.06 ms	—	—
	4000x4000	bool	Cuda	5.08 ms	±0.07 ms	—	—
		int32	Host	40.09 ms	±0.37 ms	36.97 ms	±0.08 ms
			Cuda	4.97 ms	±0.05 ms	—	—
		int64	Host	76.80 ms	±0.22 ms	73.91 ms	±0.01 ms
			Cuda	8.69 ms	±0.30 ms	—	—
		single	Host	39.44 ms	±0.03 ms	37.43 ms	±0.08 ms
			Cuda	5.05 ms	±0.12 ms	—	—
		double	Host	78.04 ms	±1.11 ms	74.93 ms	±0.03 ms
			Cuda	14.08 ms	±0.14 ms	—	—
multiply	1000x1000	bool	Cuda	0.93 ms	±0.02 ms	—	—
		int32	Host	2.66 ms	±0.11 ms	0.78 ms	±0.01 ms
			Cuda	0.99 ms	±0.03 ms	—	—
		int64	Host	12.78 ms	±0.65 ms	1.91 ms	±0.12 ms
			Cuda	1.13 ms	±0.04 ms	—	—
		single	Host	2.78 ms	±0.08 ms	0.78 ms	±0.01 ms
			Cuda	1.00 ms	±0.03 ms	—	—
		double	Host	5.75 ms	±0.11 ms	1.93 ms	±0.01 ms
			Cuda	1.14 ms	±0.03 ms	—	—
	2000x2000	bool	Cuda	1.75 ms	±0.02 ms	—	—
		int32	Host	11.41 ms	±0.39 ms	3.91 ms	±0.02 ms
			Cuda	1.83 ms	±0.08 ms	—	—
		int64	Host	45.04 ms	±0.58 ms	7.90 ms	±0.88 ms
			Cuda	2.55 ms	±0.10 ms	—	—
		single	Host	11.29 ms	±0.18 ms	3.88 ms	±0.01 ms
			Cuda	1.79 ms	±0.05 ms	—	—
		double	Host	19.92 ms	±0.14 ms	7.83 ms	±0.03 ms
			Cuda	2.59 ms	±0.04 ms	—	—
	4000x4000	bool	Cuda	5.02 ms	±0.11 ms	—	—
		int32	Host	39.89 ms	±0.21 ms	37.25 ms	±0.08 ms
			Cuda	5.11 ms	±0.15 ms	—	—
		int64	Host	187.02 ms	±10.77 ms	77.64 ms	±0.01 ms
			Cuda	10.28 ms	±0.68 ms	—	—
		single	Host	39.47 ms	±0.01 ms	37.36 ms	±0.06 ms
			Cuda	5.02 ms	±0.05 ms	—	—
		double	Host	78.73 ms	±1.30 ms	74.78 ms	±0.02 ms
			Cuda	8.60 ms	±0.30 ms	—	—
divide	1000x1000	bool	Cuda	1.08 ms	±0.02 ms	—	—
		int32	Host	—	—	4.74 ms	±0.01 ms
			Cuda	1.06 ms	±0.05 ms	—	—
		int64	Host	—	—	5.58 ms	±0.01 ms
			Cuda	1.40 ms	±0.05 ms	—	—
		single	Host	2.58 ms	±0.03 ms	0.90 ms	±0.01 ms
			Cuda	1.11 ms	±0.03 ms	—	—
		double	Host	5.98 ms	±0.23 ms	3.54 ms	±0.01 ms
			Cuda	1.18 ms	±0.03 ms	—	—
	2000x2000	bool	Cuda	1.97 ms	±0.06 ms	—	—
		int32	Host	—	—	18.89 ms	±0.01 ms
			Cuda	1.98 ms	±0.05 ms	—	—
		int64	Host	—	—	22.76 ms	±0.10 ms
			Cuda	3.26 ms	±0.05 ms	—	—
		single	Host	11.04 ms	±0.21 ms	3.87 ms	±0.01 ms
			Cuda	1.94 ms	±0.11 ms	—	—
		double	Host	21.78 ms	±0.16 ms	14.22 ms	±0.02 ms
			Cuda	2.66 ms	±0.05 ms	—	—
	4000x4000	bool	Cuda	5.78 ms	±0.09 ms	—	—
		int32	Host	—	—	130.69 ms	±0.02 ms
			Cuda	5.90 ms	±0.08 ms	—	—
		int64	Host	—	—	142.73 ms	±0.04 ms
			Cuda	11.17 ms	±0.12 ms	—	—
		single	Host	40.92 ms	±0.93 ms	43.09 ms	±0.02 ms
			Cuda	5.55 ms	±0.11 ms	—	—
		double	Host	87.25 ms	±1.90 ms	112.29 ms	±0.03 ms
			Cuda	9.89 ms	±0.30 ms	—	—
power	1000x1000	single	Host	23.07 ms	±0.57 ms	83.52 ms	±0.03 ms
			Cuda	1.21 ms	±0.05 ms	—	—
		double	Host	23.60 ms	±0.59 ms	67.25 ms	±0.03 ms
			Cuda	4.98 ms	±0.13 ms	—	—
	2000x2000	single	Host	85.71 ms	±0.98 ms	333.76 ms	±0.08 ms
			Cuda	2.53 ms	±0.06 ms	—	—
		double	Host	89.66 ms	±3.49 ms	267.09 ms	±0.09 ms
			Cuda	16.55 ms	±0.06 ms	—	—
	4000x4000	single	Host	334.92 ms	±1.75 ms	1365.24 ms	±0.32 ms
			Cuda	7.72 ms	±0.09 ms	—	—
		double	Host	348.07 ms	±4.73 ms	1135.26 ms	±0.19 ms
			Cuda	63.69 ms	±0.57 ms	—	—
modulo	1000x1000	int32	Host	—	—	16.17 ms	±0.01 ms
			Cuda	1.03 ms	±0.05 ms	—	—
		int64	Host	—	—	27.20 ms	±0.01 ms
			Cuda	1.42 ms	±0.06 ms	—	—
		single	Host	8.16 ms	±0.23 ms	38.89 ms	±0.06 ms
			Cuda	1.30 ms	±0.02 ms	—	—
		double	Host	8.99 ms	±0.57 ms	38.64 ms	±0.03 ms
			Cuda	1.25 ms	±0.04 ms	—	—
	2000x2000	int32	Host	—	—	64.71 ms	±0.03 ms
			Cuda	2.02 ms	±0.06 ms	—	—
		int64	Host	—	—	108.82 ms	±0.03 ms
			Cuda	3.20 ms	±0.04 ms	—	—
		single	Host	28.89 ms	±0.80 ms	155.63 ms	±0.09 ms
			Cuda	2.80 ms	±0.03 ms	—	—
		double	Host	29.32 ms	±0.58 ms	154.72 ms	±0.05 ms
			Cuda	2.86 ms	±0.06 ms	—	—
	4000x4000	int32	Host	—	—	287.69 ms	±0.11 ms
			Cuda	5.83 ms	±0.03 ms	—	—
		int64	Host	—	—	492.06 ms	±0.25 ms
			Cuda	11.10 ms	±0.13 ms	—	—
		single	Host	108.78 ms	±1.37 ms	651.76 ms	±0.22 ms
			Cuda	9.16 ms	±0.08 ms	—	—
		double	Host	110.98 ms	±1.32 ms	678.19 ms	±0.15 ms
			Cuda	9.30 ms	±0.09 ms	—	—
dot	1000x1000	int32	Host	—	—	473.22 ms	±0.07 ms
		int64	Host	—	—	1353.18 ms	±0.17 ms
		single	Host	21.14 ms	±0.17 ms	10.13 ms	±0.01 ms
			Cuda	3.32 ms	±0.10 ms	—	—
		double	Host	38.60 ms	±1.16 ms	20.17 ms	±0.44 ms
			Cuda	52.04 ms	±0.27 ms	—	—
	2000x2000	int32	Host	—	—	4041.68 ms	±0.42 ms
		int64	Host	—	—	11000.01 ms	±22.24 ms
		single	Host	104.04 ms	±21.00 ms	75.56 ms	±1.29 ms
			Cuda	15.93 ms	±0.07 ms	—	—
		double	Host	202.81 ms	±44.84 ms	147.94 ms	±1.54 ms
			Cuda	403.79 ms	±1.22 ms	—	—
	4000x4000	int32	Host	—	—	33160.89 ms	±2.34 ms
		int64	Host	—	—	86242.90 ms	±2.01 ms
		single	Host	605.88 ms	±64.61 ms	590.84 ms	±2.47 ms
			Cuda	117.63 ms	±3.59 ms	—	—
		double	Host	—	—	1177.82 ms	±2.07 ms
sgn	1000x1000	bool	Cuda	0.92 ms	±0.03 ms	—	—
		int32	Host	5.74 ms	±0.34 ms	0.56 ms	±0.01 ms
			Cuda	0.98 ms	±0.04 ms	—	—
		int64	Host	6.76 ms	±0.24 ms	6.89 ms	±0.11 ms
			Cuda	0.92 ms	±0.04 ms	—	—
		single	Host	6.43 ms	±0.11 ms	7.98 ms	±0.12 ms
			Cuda	0.92 ms	±0.02 ms	—	—
		double	Host	6.94 ms	±0.34 ms	8.25 ms	±0.11 ms
			Cuda	1.03 ms	±0.05 ms	—	—
	2000x2000	bool	Cuda	1.55 ms	±0.07 ms	—	—
		int32	Host	18.84 ms	±0.54 ms	2.91 ms	±0.08 ms
			Cuda	1.56 ms	±0.02 ms	—	—
		int64	Host	21.68 ms	±0.37 ms	27.67 ms	±0.28 ms
			Cuda	2.50 ms	±0.09 ms	—	—
		single	Host	20.16 ms	±0.33 ms	32.08 ms	±0.31 ms
			Cuda	1.62 ms	±0.08 ms	—	—
		double	Host	22.54 ms	±0.61 ms	33.00 ms	±0.04 ms
			Cuda	2.02 ms	±0.08 ms	—	—
	4000x4000	bool	Cuda	5.00 ms	±0.08 ms	—	—
		int32	Host	70.85 ms	±1.88 ms	37.13 ms	±0.43 ms
			Cuda	5.14 ms	±0.26 ms	—	—
		int64	Host	79.25 ms	±1.05 ms	165.27 ms	±0.71 ms
			Cuda	8.92 ms	±1.79 ms	—	—
		single	Host	77.25 ms	±0.66 ms	156.19 ms	±0.81 ms
			Cuda	4.46 ms	±0.06 ms	—	—
		double	Host	84.58 ms	±0.57 ms	187.36 ms	±0.81 ms
			Cuda	10.73 ms	±0.25 ms	—	—
log	1000x1000	int32	Host	—	—	32.64 ms	±0.47 ms
		int64	Host	—	—	33.00 ms	±0.49 ms
		single	Host	12.74 ms	±0.43 ms	23.53 ms	±4.29 ms
			Cuda	0.98 ms	±0.05 ms	—	—
		double	Host	13.13 ms	±0.25 ms	31.81 ms	±0.48 ms
			Cuda	2.16 ms	±0.06 ms	—	—
	2000x2000	int32	Host	—	—	130.48 ms	±0.91 ms
		int64	Host	—	—	132.12 ms	±0.91 ms
		single	Host	46.65 ms	±1.03 ms	82.64 ms	±0.92 ms
			Cuda	1.74 ms	±0.05 ms	—	—
		double	Host	47.91 ms	±0.79 ms	127.23 ms	±0.98 ms
			Cuda	6.27 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	580.22 ms	±1.26 ms
		int64	Host	—	—	655.68 ms	±47.15 ms
		single	Host	179.25 ms	±1.62 ms	361.09 ms	±1.52 ms
			Cuda	4.91 ms	±0.12 ms	—	—
		double	Host	183.58 ms	±0.84 ms	566.88 ms	±1.28 ms
			Cuda	23.17 ms	±0.04 ms	—	—
log10	1000x1000	int32	Host	—	—	34.15 ms	±0.53 ms
		int64	Host	—	—	34.59 ms	±0.55 ms
		single	Host	12.79 ms	±0.23 ms	25.56 ms	±0.40 ms
			Cuda	0.95 ms	±0.02 ms	—	—
		double	Host	13.27 ms	±0.39 ms	34.88 ms	±0.57 ms
			Cuda	2.24 ms	±0.02 ms	—	—
	2000x2000	int32	Host	—	—	136.53 ms	±1.12 ms
		int64	Host	—	—	138.04 ms	±1.03 ms
		single	Host	49.18 ms	±1.68 ms	102.44 ms	±1.11 ms
			Cuda	1.76 ms	±0.08 ms	—	—
		double	Host	47.70 ms	±1.15 ms	139.56 ms	±1.09 ms
			Cuda	6.79 ms	±0.05 ms	—	—
	4000x4000	int32	Host	—	—	604.82 ms	±1.64 ms
		int64	Host	—	—	610.75 ms	±1.76 ms
		single	Host	179.04 ms	±2.06 ms	440.81 ms	±1.90 ms
			Cuda	4.80 ms	±0.08 ms	—	—
		double	Host	182.97 ms	±3.08 ms	616.02 ms	±0.36 ms
			Cuda	25.00 ms	±0.18 ms	—	—
exp	1000x1000	int32	Host	—	—	27.33 ms	±0.00 ms
		int64	Host	—	—	29.07 ms	±3.99 ms
		single	Host	7.45 ms	±0.11 ms	16.57 ms	±0.02 ms
			Cuda	0.98 ms	±0.05 ms	—	—
		double	Host	8.13 ms	±0.22 ms	26.94 ms	±0.01 ms
			Cuda	1.51 ms	±0.04 ms	—	—
	2000x2000	int32	Host	—	—	109.37 ms	±0.01 ms
		int64	Host	—	—	111.06 ms	±0.01 ms
		single	Host	24.26 ms	±0.76 ms	66.35 ms	±0.01 ms
			Cuda	1.67 ms	±0.03 ms	—	—
		double	Host	25.63 ms	±0.52 ms	107.68 ms	±0.01 ms
			Cuda	3.90 ms	±0.11 ms	—	—
	4000x4000	int32	Host	—	—	575.90 ms	±103.34 ms
		int64	Host	—	—	576.50 ms	±110.41 ms
		single	Host	94.02 ms	±1.68 ms	293.30 ms	±0.07 ms
			Cuda	4.65 ms	±0.07 ms	—	—
		double	Host	97.30 ms	±1.56 ms	487.84 ms	±0.37 ms
			Cuda	13.43 ms	±0.03 ms	—	—
sin	1000x1000	int32	Host	—	—	54.91 ms	±0.09 ms
		int64	Host	—	—	55.34 ms	±0.09 ms
		single	Host	7.86 ms	±0.53 ms	21.02 ms	±0.03 ms
			Cuda	0.96 ms	±0.06 ms	—	—
		double	Host	9.21 ms	±0.68 ms	60.76 ms	±0.02 ms
			Cuda	1.66 ms	±0.03 ms	—	—
	2000x2000	int32	Host	—	—	219.60 ms	±0.09 ms
		int64	Host	—	—	221.39 ms	±0.10 ms
		single	Host	27.79 ms	±0.94 ms	84.46 ms	±0.02 ms
			Cuda	1.74 ms	±0.09 ms	—	—
		double	Host	28.69 ms	±0.37 ms	242.90 ms	±0.05 ms
			Cuda	4.46 ms	±0.06 ms	—	—
	4000x4000	int32	Host	—	—	939.10 ms	±0.35 ms
		int64	Host	—	—	944.28 ms	±0.40 ms
		single	Host	109.36 ms	±2.68 ms	369.04 ms	±0.05 ms
			Cuda	4.97 ms	±0.08 ms	—	—
		double	Host	109.69 ms	±2.07 ms	1053.12 ms	±37.89 ms
			Cuda	15.28 ms	±0.05 ms	—	—
cos	1000x1000	int32	Host	—	—	86.90 ms	±0.04 ms
		int64	Host	—	—	87.34 ms	±0.05 ms
		single	Host	8.57 ms	±0.57 ms	41.42 ms	±0.02 ms
			Cuda	0.94 ms	±0.04 ms	—	—
		double	Host	8.80 ms	±0.41 ms	82.03 ms	±0.10 ms
			Cuda	1.72 ms	±0.10 ms	—	—
	2000x2000	int32	Host	—	—	348.03 ms	±0.06 ms
		int64	Host	—	—	349.48 ms	±0.11 ms
		single	Host	27.14 ms	±0.53 ms	166.05 ms	±0.04 ms
			Cuda	1.85 ms	±0.02 ms	—	—
		double	Host	28.08 ms	±0.64 ms	328.65 ms	±0.40 ms
			Cuda	4.41 ms	±0.10 ms	—	—
	4000x4000	int32	Host	—	—	1461.91 ms	±29.70 ms
		int64	Host	—	—	1455.08 ms	±0.25 ms
		single	Host	103.46 ms	±1.39 ms	692.74 ms	±0.12 ms
			Cuda	4.93 ms	±0.06 ms	—	—
		double	Host	105.98 ms	±1.33 ms	1373.27 ms	±1.03 ms
			Cuda	15.37 ms	±0.20 ms	—	—
tan	1000x1000	single	Host	9.26 ms	±0.38 ms	—	—
			Cuda	0.83 ms	±0.01 ms	—	—
		double	Host	9.68 ms	±0.12 ms	—	—
			Cuda	2.36 ms	±0.03 ms	—	—
	2000x2000	single	Host	32.65 ms	±0.48 ms	—	—
			Cuda	2.11 ms	±0.06 ms	—	—
		double	Host	33.76 ms	±0.34 ms	—	—
			Cuda	6.76 ms	±0.17 ms	—	—
	4000x4000	single	Host	127.05 ms	±1.63 ms	—	—
			Cuda	6.06 ms	±0.10 ms	—	—
		double	Host	129.43 ms	±1.10 ms	—	—
			Cuda	24.58 ms	±0.02 ms	—	—
asin	1000x1000	int32	Host	—	—	11.11 ms	±0.01 ms
		int64	Host	—	—	11.50 ms	±0.01 ms
		single	Host	16.15 ms	±0.43 ms	11.96 ms	±0.00 ms
			Cuda	0.81 ms	±0.04 ms	—	—
		double	Host	17.38 ms	±0.44 ms	10.38 ms	±0.01 ms
			Cuda	2.26 ms	±0.07 ms	—	—
	2000x2000	int32	Host	—	—	44.41 ms	±0.01 ms
		int64	Host	—	—	47.40 ms	±4.59 ms
		single	Host	61.59 ms	±0.59 ms	47.89 ms	±0.02 ms
			Cuda	1.88 ms	±0.07 ms	—	—
		double	Host	64.00 ms	±1.46 ms	41.60 ms	±0.02 ms
			Cuda	6.85 ms	±0.04 ms	—	—
	4000x4000	int32	Host	—	—	239.15 ms	±9.48 ms
		int64	Host	—	—	237.64 ms	±0.04 ms
		single	Host	243.50 ms	±3.46 ms	219.26 ms	±0.04 ms
			Cuda	5.38 ms	±0.05 ms	—	—
		double	Host	248.15 ms	±2.13 ms	220.14 ms	±0.03 ms
			Cuda	25.13 ms	±0.18 ms	—	—
acos	1000x1000	int32	Host	—	—	12.10 ms	±0.01 ms
		int64	Host	—	—	12.48 ms	±0.01 ms
		single	Host	16.14 ms	±0.30 ms	11.54 ms	±0.01 ms
			Cuda	1.00 ms	±0.04 ms	—	—
		double	Host	17.13 ms	±0.20 ms	11.37 ms	±0.01 ms
			Cuda	1.87 ms	±0.08 ms	—	—
	2000x2000	int32	Host	—	—	48.36 ms	±0.03 ms
		int64	Host	—	—	49.96 ms	±0.01 ms
		single	Host	61.06 ms	±0.32 ms	46.28 ms	±0.02 ms
			Cuda	1.86 ms	±0.05 ms	—	—
		double	Host	62.61 ms	±0.74 ms	45.48 ms	±0.02 ms
			Cuda	5.30 ms	±0.03 ms	—	—
	4000x4000	int32	Host	—	—	304.26 ms	±32.94 ms
		int64	Host	—	—	253.44 ms	±0.06 ms
		single	Host	243.49 ms	±5.10 ms	212.83 ms	±0.03 ms
			Cuda	5.46 ms	±0.07 ms	—	—
		double	Host	244.80 ms	±1.37 ms	235.34 ms	±0.03 ms
			Cuda	18.92 ms	±0.02 ms	—	—
atan	1000x1000	int32	Host	—	—	38.09 ms	±0.02 ms
		int64	Host	—	—	38.50 ms	±0.03 ms
		single	Host	8.51 ms	±0.13 ms	18.98 ms	±0.02 ms
			Cuda	1.06 ms	±0.05 ms	—	—
		double	Host	9.30 ms	±0.22 ms	39.43 ms	±0.02 ms
			Cuda	2.09 ms	±0.09 ms	—	—
	2000x2000	int32	Host	—	—	152.32 ms	±0.06 ms
		int64	Host	—	—	169.61 ms	±24.22 ms
		single	Host	29.27 ms	±1.00 ms	76.03 ms	±0.02 ms
			Cuda	1.86 ms	±0.04 ms	—	—
		double	Host	29.66 ms	±0.41 ms	157.75 ms	±0.03 ms
			Cuda	6.50 ms	±0.06 ms	—	—
	4000x4000	int32	Host	—	—	668.09 ms	±0.16 ms
		int64	Host	—	—	673.69 ms	±0.18 ms
		single	Host	112.47 ms	±0.08 ms	334.32 ms	±0.11 ms
			Cuda	5.57 ms	±0.10 ms	—	—
		double	Host	114.61 ms	±0.85 ms	757.55 ms	±103.61 ms
			Cuda	23.62 ms	±0.01 ms	—	—
sinh	1000x1000	int32	Host	—	—	63.06 ms	±7.39 ms
		int64	Host	—	—	49.02 ms	±0.02 ms
		single	Host	8.72 ms	±0.21 ms	38.09 ms	±0.01 ms
			Cuda	0.96 ms	±0.02 ms	—	—
		double	Host	9.72 ms	±0.31 ms	48.73 ms	±0.01 ms
			Cuda	2.47 ms	±0.08 ms	—	—
	2000x2000	int32	Host	—	—	241.74 ms	±36.77 ms
		int64	Host	—	—	196.19 ms	±0.03 ms
		single	Host	31.37 ms	±0.77 ms	152.72 ms	±0.03 ms
			Cuda	1.82 ms	±0.08 ms	—	—
		double	Host	32.05 ms	±0.29 ms	194.96 ms	±0.03 ms
			Cuda	6.98 ms	±0.12 ms	—	—
	4000x4000	int32	Host	—	—	905.57 ms	±122.32 ms
		int64	Host	—	—	844.33 ms	±0.24 ms
		single	Host	118.98 ms	±0.55 ms	639.87 ms	±0.09 ms
			Cuda	5.23 ms	±0.11 ms	—	—
		double	Host	123.36 ms	±0.82 ms	839.24 ms	±0.20 ms
			Cuda	25.66 ms	±0.17 ms	—	—
cosh	1000x1000	int32	Host	—	—	35.42 ms	±0.03 ms
		int64	Host	—	—	35.87 ms	±0.01 ms
		single	Host	9.09 ms	±0.11 ms	22.92 ms	±0.03 ms
			Cuda	0.93 ms	±0.03 ms	—	—
		double	Host	9.82 ms	±0.01 ms	35.28 ms	±0.01 ms
			Cuda	1.75 ms	±0.12 ms	—	—
	2000x2000	int32	Host	—	—	141.86 ms	±0.14 ms
		int64	Host	—	—	159.11 ms	±26.40 ms
		single	Host	32.15 ms	±1.38 ms	91.70 ms	±0.09 ms
			Cuda	1.72 ms	±0.04 ms	—	—
		double	Host	32.51 ms	±0.50 ms	141.36 ms	±0.04 ms
			Cuda	4.70 ms	±0.08 ms	—	—
	4000x4000	int32	Host	—	—	626.83 ms	±0.35 ms
		int64	Host	—	—	632.70 ms	±0.39 ms
		single	Host	123.85 ms	±3.35 ms	431.62 ms	±72.79 ms
			Cuda	4.69 ms	±0.06 ms	—	—
		double	Host	125.60 ms	±0.39 ms	625.80 ms	±0.13 ms
			Cuda	16.64 ms	±0.17 ms	—	—
tanh	1000x1000	int32	Host	—	—	28.31 ms	±0.01 ms
		int64	Host	—	—	28.69 ms	±0.02 ms
		single	Host	8.36 ms	±0.40 ms	32.42 ms	±5.54 ms
			Cuda	0.92 ms	±0.04 ms	—	—
		double	Host	8.84 ms	±0.14 ms	28.56 ms	±0.02 ms
			Cuda	2.14 ms	±0.04 ms	—	—
	2000x2000	int32	Host	—	—	115.52 ms	±6.88 ms
		int64	Host	—	—	114.78 ms	±0.03 ms
		single	Host	27.77 ms	±0.56 ms	111.03 ms	±0.09 ms
			Cuda	1.84 ms	±0.07 ms	—	—
		double	Host	29.40 ms	±0.60 ms	114.38 ms	±0.03 ms
			Cuda	6.23 ms	±0.10 ms	—	—
	4000x4000	int32	Host	—	—	511.16 ms	±0.12 ms
		int64	Host	—	—	516.73 ms	±0.17 ms
		single	Host	110.60 ms	±2.60 ms	552.64 ms	±77.32 ms
			Cuda	5.04 ms	±0.05 ms	—	—
		double	Host	113.75 ms	±1.79 ms	621.19 ms	±77.36 ms
			Cuda	22.50 ms	±0.02 ms	—	—
sqrt	1000x1000	int32	Host	6.63 ms	±0.50 ms	3.73 ms	±0.01 ms
		int64	Host	11.66 ms	±0.81 ms	4.09 ms	±0.01 ms
		single	Host	2.44 ms	±0.04 ms	0.92 ms	±0.00 ms
			Cuda	1.01 ms	±0.06 ms	—	—
		double	Host	6.51 ms	±0.52 ms	3.21 ms	±0.01 ms
			Cuda	1.25 ms	±0.03 ms	—	—
	2000x2000	int32	Host	23.78 ms	±0.04 ms	14.92 ms	±0.01 ms
		int64	Host	41.70 ms	±0.46 ms	16.47 ms	±0.01 ms
		single	Host	10.54 ms	±0.21 ms	3.72 ms	±0.02 ms
			Cuda	1.85 ms	±0.04 ms	—	—
		double	Host	20.13 ms	±0.08 ms	12.82 ms	±0.01 ms
			Cuda	2.88 ms	±0.06 ms	—	—
	4000x4000	int32	Host	95.80 ms	±0.73 ms	115.76 ms	±1.60 ms
		int64	Host	163.14 ms	±0.06 ms	119.99 ms	±0.04 ms
		single	Host	35.85 ms	±0.12 ms	42.34 ms	±0.02 ms
			Cuda	5.35 ms	±0.10 ms	—	—
		double	Host	82.56 ms	±0.92 ms	105.91 ms	±0.03 ms
			Cuda	9.46 ms	±0.08 ms	—	—
ceiling	1000x1000	single	Host	8.01 ms	±0.30 ms	—	—
			Cuda	0.80 ms	±0.04 ms	—	—
		double	Host	6.94 ms	±0.16 ms	—	—
			Cuda	0.88 ms	±0.02 ms	—	—
	2000x2000	single	Host	17.86 ms	±0.51 ms	—	—
			Cuda	1.59 ms	±0.08 ms	—	—
		double	Host	19.72 ms	±0.55 ms	—	—
			Cuda	2.51 ms	±0.08 ms	—	—
	4000x4000	single	Host	67.21 ms	±1.18 ms	—	—
			Cuda	4.86 ms	±0.12 ms	—	—
		double	Host	74.65 ms	±1.16 ms	—	—
			Cuda	11.03 ms	±0.30 ms	—	—
floor	1000x1000	single	Host	5.91 ms	±0.24 ms	—	—
			Cuda	0.91 ms	±0.03 ms	—	—
		double	Host	6.69 ms	±0.11 ms	—	—
			Cuda	0.84 ms	±0.03 ms	—	—
	2000x2000	single	Host	17.84 ms	±0.46 ms	—	—
			Cuda	1.59 ms	±0.08 ms	—	—
		double	Host	20.52 ms	±0.20 ms	—	—
			Cuda	2.48 ms	±0.13 ms	—	—
	4000x4000	single	Host	68.82 ms	±0.60 ms	—	—
			Cuda	4.96 ms	±0.04 ms	—	—
		double	Host	75.01 ms	±0.76 ms	—	—
			Cuda	10.90 ms	±0.31 ms	—	—
round	1000x1000	single	Host	6.18 ms	±0.30 ms	—	—
			Cuda	0.90 ms	±0.03 ms	—	—
		double	Host	6.55 ms	±0.14 ms	—	—
			Cuda	0.94 ms	±0.03 ms	—	—
	2000x2000	single	Host	17.47 ms	±0.41 ms	—	—
			Cuda	1.60 ms	±0.06 ms	—	—
		double	Host	19.92 ms	±0.15 ms	—	—
			Cuda	2.55 ms	±0.11 ms	—	—
	4000x4000	single	Host	68.77 ms	±0.60 ms	—	—
			Cuda	4.44 ms	±0.11 ms	—	—
		double	Host	74.54 ms	±0.52 ms	—	—
			Cuda	12.39 ms	±3.97 ms	—	—
truncate	1000x1000	single	Host	6.86 ms	±0.44 ms	—	—
			Cuda	0.88 ms	±0.03 ms	—	—
		double	Host	7.24 ms	±0.21 ms	—	—
			Cuda	0.80 ms	±0.02 ms	—	—
	2000x2000	single	Host	20.98 ms	±0.43 ms	—	—
			Cuda	1.55 ms	±0.05 ms	—	—
		double	Host	22.74 ms	±0.18 ms	—	—
			Cuda	2.49 ms	±0.15 ms	—	—
	4000x4000	single	Host	81.93 ms	±1.56 ms	—	—
			Cuda	4.25 ms	±0.09 ms	—	—
		double	Host	89.32 ms	±1.77 ms	—	—
			Cuda	11.16 ms	±0.31 ms	—	—
maxElemwise	1000x1000	bool	Host	5.75 ms	±0.14 ms	—	—
			Cuda	1.00 ms	±0.05 ms	—	—
		int32	Host	2.83 ms	±0.10 ms	—	—
			Cuda	0.99 ms	±0.01 ms	—	—
		int64	Host	6.82 ms	±0.43 ms	—	—
			Cuda	1.17 ms	±0.07 ms	—	—
		single	Host	2.81 ms	±0.12 ms	—	—
			Cuda	1.02 ms	±0.04 ms	—	—
		double	Host	6.11 ms	±0.27 ms	—	—
			Cuda	1.13 ms	±0.02 ms	—	—
	2000x2000	bool	Host	15.93 ms	±0.78 ms	—	—
			Cuda	1.76 ms	±0.08 ms	—	—
		int32	Host	11.14 ms	±0.30 ms	—	—
			Cuda	1.72 ms	±0.03 ms	—	—
		int64	Host	19.10 ms	±0.28 ms	—	—
			Cuda	2.60 ms	±0.12 ms	—	—
		single	Host	11.41 ms	±0.23 ms	—	—
			Cuda	1.80 ms	±0.04 ms	—	—
		double	Host	19.61 ms	±0.29 ms	—	—
			Cuda	2.56 ms	±0.10 ms	—	—
	4000x4000	bool	Host	55.53 ms	±1.01 ms	—	—
			Cuda	4.95 ms	±0.07 ms	—	—
		int32	Host	39.05 ms	±0.46 ms	—	—
			Cuda	4.98 ms	±0.10 ms	—	—
		int64	Host	79.92 ms	±2.27 ms	—	—
			Cuda	11.67 ms	±2.89 ms	—	—
		single	Host	40.56 ms	±1.10 ms	—	—
			Cuda	4.98 ms	±0.08 ms	—	—
		double	Host	79.27 ms	±3.37 ms	—	—
			Cuda	12.12 ms	±1.67 ms	—	—
minElemwise	1000x1000	bool	Host	5.69 ms	±0.09 ms	—	—
			Cuda	1.01 ms	±0.03 ms	—	—
		int32	Host	2.70 ms	±0.06 ms	—	—
			Cuda	0.99 ms	±0.02 ms	—	—
		int64	Host	6.32 ms	±0.52 ms	—	—
			Cuda	1.14 ms	±0.04 ms	—	—
		single	Host	2.84 ms	±0.16 ms	—	—
			Cuda	1.04 ms	±0.05 ms	—	—
		double	Host	5.88 ms	±0.21 ms	—	—
			Cuda	1.17 ms	±0.03 ms	—	—
	2000x2000	bool	Host	15.18 ms	±0.20 ms	—	—
			Cuda	1.78 ms	±0.05 ms	—	—
		int32	Host	11.60 ms	±0.16 ms	—	—
			Cuda	1.87 ms	±0.10 ms	—	—
		int64	Host	19.90 ms	±0.79 ms	—	—
			Cuda	2.58 ms	±0.10 ms	—	—
		single	Host	11.06 ms	±0.10 ms	—	—
			Cuda	1.78 ms	±0.03 ms	—	—
		double	Host	19.81 ms	±0.27 ms	—	—
			Cuda	2.56 ms	±0.06 ms	—	—
	4000x4000	bool	Host	55.36 ms	±1.97 ms	—	—
			Cuda	5.01 ms	±0.06 ms	—	—
		int32	Host	41.39 ms	±1.70 ms	—	—
			Cuda	5.06 ms	±0.04 ms	—	—
		int64	Host	81.67 ms	±0.97 ms	—	—
			Cuda	11.74 ms	±1.70 ms	—	—
		single	Host	40.64 ms	±1.21 ms	—	—
			Cuda	5.02 ms	±0.13 ms	—	—
		double	Host	80.58 ms	±5.30 ms	—	—
			Cuda	8.66 ms	±0.30 ms	—	—
sumAxis	1000x1000	int32	Host	3.06 ms	±0.10 ms	1.40 ms	±0.01 ms
			Cuda	1.35 ms	±0.08 ms	—	—
		int64	Host	3.03 ms	±0.13 ms	0.94 ms	±0.01 ms
			Cuda	1.60 ms	±0.05 ms	—	—
		single	Host	3.28 ms	±0.22 ms	0.62 ms	±0.00 ms
			Cuda	1.42 ms	±0.01 ms	—	—
		double	Host	3.05 ms	±0.15 ms	0.63 ms	±0.01 ms
			Cuda	1.65 ms	±0.07 ms	—	—
	2000x2000	int32	Host	12.62 ms	±0.72 ms	5.67 ms	±0.01 ms
			Cuda	2.50 ms	±0.06 ms	—	—
		int64	Host	12.69 ms	±0.30 ms	3.75 ms	±0.01 ms
			Cuda	2.98 ms	±0.08 ms	—	—
		single	Host	11.83 ms	±0.33 ms	2.43 ms	±0.01 ms
			Cuda	2.49 ms	±0.03 ms	—	—
		double	Host	13.11 ms	±0.59 ms	2.69 ms	±0.01 ms
			Cuda	3.01 ms	±0.05 ms	—	—
	4000x4000	int32	Host	44.26 ms	±1.19 ms	22.98 ms	±0.01 ms
			Cuda	6.80 ms	±0.05 ms	—	—
		int64	Host	45.82 ms	±6.60 ms	14.94 ms	±0.08 ms
			Cuda	12.57 ms	±0.06 ms	—	—
		single	Host	49.03 ms	±5.63 ms	9.58 ms	±0.01 ms
			Cuda	6.76 ms	±0.13 ms	—	—
		double	Host	45.63 ms	±1.17 ms	10.71 ms	±0.02 ms
			Cuda	12.51 ms	±0.05 ms	—	—
productAxis	1000x1000	int32	Host	3.03 ms	±0.15 ms	—	—
			Cuda	1.47 ms	±0.04 ms	—	—
		int64	Host	3.14 ms	±0.07 ms	—	—
			Cuda	1.61 ms	±0.03 ms	—	—
		single	Host	3.01 ms	±0.11 ms	—	—
			Cuda	1.45 ms	±0.03 ms	—	—
		double	Host	3.30 ms	±0.40 ms	—	—
			Cuda	1.60 ms	±0.03 ms	—	—
	2000x2000	int32	Host	12.93 ms	±0.77 ms	—	—
			Cuda	2.54 ms	±0.09 ms	—	—
		int64	Host	12.50 ms	±0.49 ms	—	—
			Cuda	3.08 ms	±0.09 ms	—	—
		single	Host	12.51 ms	±0.57 ms	—	—
			Cuda	2.50 ms	±0.03 ms	—	—
		double	Host	12.70 ms	±0.86 ms	—	—
			Cuda	3.02 ms	±0.06 ms	—	—
	4000x4000	int32	Host	44.99 ms	±1.78 ms	—	—
			Cuda	6.69 ms	±0.08 ms	—	—
		int64	Host	46.90 ms	±6.64 ms	—	—
			Cuda	12.48 ms	±0.01 ms	—	—
		single	Host	44.97 ms	±4.74 ms	—	—
			Cuda	6.76 ms	±0.06 ms	—	—
		double	Host	43.42 ms	±2.01 ms	—	—
			Cuda	12.52 ms	±0.04 ms	—	—
maxAxis	1000x1000	int32	Host	3.30 ms	±0.10 ms	—	—
			Cuda	1.46 ms	±0.02 ms	—	—
		int64	Host	3.46 ms	±0.06 ms	—	—
			Cuda	1.59 ms	±0.03 ms	—	—
		single	Host	3.47 ms	±0.29 ms	—	—
			Cuda	1.46 ms	±0.02 ms	—	—
		double	Host	3.35 ms	±0.06 ms	—	—
			Cuda	1.65 ms	±0.05 ms	—	—
	2000x2000	int32	Host	12.86 ms	±0.68 ms	—	—
			Cuda	2.47 ms	±0.03 ms	—	—
		int64	Host	13.05 ms	±0.73 ms	—	—
			Cuda	3.08 ms	±0.11 ms	—	—
		single	Host	12.88 ms	±0.35 ms	—	—
			Cuda	2.46 ms	±0.04 ms	—	—
		double	Host	12.74 ms	±0.68 ms	—	—
			Cuda	3.01 ms	±0.10 ms	—	—
	4000x4000	int32	Host	46.89 ms	±1.13 ms	—	—
			Cuda	6.69 ms	±0.04 ms	—	—
		int64	Host	48.34 ms	±1.49 ms	—	—
			Cuda	12.54 ms	±0.06 ms	—	—
		single	Host	48.35 ms	±1.66 ms	—	—
			Cuda	6.72 ms	±0.06 ms	—	—
		double	Host	48.30 ms	±1.67 ms	—	—
			Cuda	12.54 ms	±0.05 ms	—	—
minAxis	1000x1000	int32	Host	3.23 ms	±0.16 ms	—	—
			Cuda	1.40 ms	±0.01 ms	—	—
		int64	Host	3.18 ms	±0.15 ms	—	—
			Cuda	1.58 ms	±0.04 ms	—	—
		single	Host	3.27 ms	±0.09 ms	—	—
			Cuda	1.45 ms	±0.04 ms	—	—
		double	Host	3.29 ms	±0.10 ms	—	—
			Cuda	1.63 ms	±0.05 ms	—	—
	2000x2000	int32	Host	13.47 ms	±0.81 ms	—	—
			Cuda	2.53 ms	±0.03 ms	—	—
		int64	Host	13.15 ms	±0.75 ms	—	—
			Cuda	3.00 ms	±0.06 ms	—	—
		single	Host	13.56 ms	±1.48 ms	—	—
			Cuda	2.49 ms	±0.05 ms	—	—
		double	Host	13.80 ms	±1.07 ms	—	—
			Cuda	2.99 ms	±0.03 ms	—	—
	4000x4000	int32	Host	47.39 ms	±2.33 ms	—	—
			Cuda	6.80 ms	±0.04 ms	—	—
		int64	Host	53.98 ms	±0.61 ms	—	—
			Cuda	12.56 ms	±0.04 ms	—	—
		single	Host	51.34 ms	±1.03 ms	—	—
			Cuda	6.75 ms	±0.06 ms	—	—
		double	Host	48.59 ms	±1.59 ms	—	—
			Cuda	12.66 ms	±0.10 ms	—	—
argMaxAxis	1000x1000	int32	Host	4.32 ms	±0.15 ms	—	—
			Cuda	1.42 ms	±0.03 ms	—	—
		int64	Host	4.62 ms	±0.25 ms	—	—
			Cuda	1.54 ms	±0.01 ms	—	—
		single	Host	4.75 ms	±0.07 ms	—	—
			Cuda	1.50 ms	±0.05 ms	—	—
		double	Host	4.71 ms	±0.31 ms	—	—
			Cuda	1.61 ms	±0.04 ms	—	—
	2000x2000	int32	Host	16.13 ms	±0.61 ms	—	—
			Cuda	2.64 ms	±0.04 ms	—	—
		int64	Host	16.19 ms	±0.60 ms	—	—
			Cuda	2.90 ms	±0.04 ms	—	—
		single	Host	16.68 ms	±0.53 ms	—	—
			Cuda	2.52 ms	±0.05 ms	—	—
		double	Host	16.82 ms	±0.43 ms	—	—
			Cuda	2.99 ms	±0.10 ms	—	—
	4000x4000	int32	Host	61.87 ms	±1.80 ms	—	—
			Cuda	6.80 ms	±0.08 ms	—	—
		int64	Host	62.13 ms	±1.04 ms	—	—
			Cuda	12.56 ms	±0.03 ms	—	—
		single	Host	74.78 ms	±1.93 ms	—	—
			Cuda	6.79 ms	±0.06 ms	—	—
		double	Host	64.73 ms	±2.11 ms	—	—
			Cuda	12.56 ms	±0.06 ms	—	—
argMinAxis	1000x1000	int32	Host	4.51 ms	±0.30 ms	—	—
			Cuda	1.36 ms	±0.04 ms	—	—
		int64	Host	4.22 ms	±0.14 ms	—	—
			Cuda	1.60 ms	±0.05 ms	—	—
		single	Host	4.21 ms	±0.16 ms	—	—
			Cuda	1.49 ms	±0.05 ms	—	—
		double	Host	4.40 ms	±0.14 ms	—	—
			Cuda	1.57 ms	±0.04 ms	—	—
	2000x2000	int32	Host	16.16 ms	±0.47 ms	—	—
			Cuda	2.52 ms	±0.03 ms	—	—
		int64	Host	16.15 ms	±0.71 ms	—	—
			Cuda	2.90 ms	±0.04 ms	—	—
		single	Host	16.30 ms	±0.54 ms	—	—
			Cuda	2.47 ms	±0.02 ms	—	—
		double	Host	16.47 ms	±0.19 ms	—	—
			Cuda	2.95 ms	±0.05 ms	—	—
	4000x4000	int32	Host	63.24 ms	±2.04 ms	—	—
			Cuda	6.76 ms	±0.10 ms	—	—
		int64	Host	64.42 ms	±4.46 ms	—	—
			Cuda	12.53 ms	±0.05 ms	—	—
		single	Host	61.57 ms	±1.63 ms	—	—
			Cuda	6.77 ms	±0.10 ms	—	—
		double	Host	63.76 ms	±1.90 ms	—	—
			Cuda	12.60 ms	±0.05 ms	—	—
findAxis	1000x1000	int32	Host	1.10 ms	±0.01 ms	—	—
			Cuda	0.57 ms	±0.00 ms	—	—
		int64	Host	1.12 ms	±0.03 ms	—	—
			Cuda	0.74 ms	±0.02 ms	—	—
		single	Host	1.39 ms	±0.02 ms	—	—
			Cuda	1.48 ms	±0.04 ms	—	—
		double	Host	1.42 ms	±0.02 ms	—	—
			Cuda	1.50 ms	±0.04 ms	—	—
	2000x2000	int32	Host	3.64 ms	±0.04 ms	—	—
			Cuda	0.85 ms	±0.03 ms	—	—
		int64	Host	3.66 ms	±0.04 ms	—	—
			Cuda	0.88 ms	±0.02 ms	—	—
		single	Host	4.85 ms	±0.02 ms	—	—
			Cuda	2.44 ms	±0.07 ms	—	—
		double	Host	4.92 ms	±0.08 ms	—	—
			Cuda	2.69 ms	±0.09 ms	—	—
	4000x4000	int32	Host	13.17 ms	±0.24 ms	—	—
			Cuda	1.34 ms	±0.04 ms	—	—
		int64	Host	13.28 ms	±0.15 ms	—	—
			Cuda	1.42 ms	±0.06 ms	—	—
		single	Host	18.01 ms	±0.16 ms	—	—
			Cuda	6.82 ms	±0.10 ms	—	—
		double	Host	18.12 ms	±0.29 ms	—	—
			Cuda	12.53 ms	±0.03 ms	—	—
maskedGet	1000x1000	bool	Host	61.89 ms	±4.15 ms	—	—
		int32	Host	61.87 ms	±2.69 ms	—	—
		int64	Host	69.05 ms	±4.59 ms	—	—
		single	Host	61.39 ms	±2.30 ms	—	—
		double	Host	63.59 ms	±6.51 ms	—	—
	2000x2000	bool	Host	275.66 ms	±62.64 ms	—	—
		int32	Host	294.05 ms	±59.29 ms	—	—
		int64	Host	275.87 ms	±37.48 ms	—	—
		single	Host	269.70 ms	±49.99 ms	—	—
		double	Host	269.58 ms	±18.39 ms	—	—
	4000x4000	int32	Host	988.48 ms	±53.13 ms	—	—
		single	Host	960.94 ms	±18.94 ms	—	—
		double	Host	996.74 ms	±12.15 ms	—	—
maskedSet	1000x1000	bool	Host	62.88 ms	±2.11 ms	—	—
		int32	Host	66.68 ms	±6.03 ms	—	—
		int64	Host	65.94 ms	±5.41 ms	—	—
		single	Host	65.01 ms	±2.46 ms	—	—
		double	Host	67.73 ms	±3.74 ms	—	—
	2000x2000	bool	Host	254.70 ms	±21.87 ms	—	—
		int32	Host	249.84 ms	±8.72 ms	—	—
		int64	Host	267.47 ms	±16.21 ms	—	—
		single	Host	246.02 ms	±20.99 ms	—	—
		double	Host	257.04 ms	±25.52 ms	—	—
	4000x4000	bool	Host	979.76 ms	±22.66 ms	—	—
		int64	Host	979.39 ms	±10.59 ms	—	—
		double	Host	972.80 ms	±20.03 ms	—	—
invert	1000x1000	single	Host	22.04 ms	±2.59 ms	—	—
			Cuda	605.05 ms	±2.30 ms	—	—
		double	Host	44.59 ms	±6.42 ms	—	—
	2000x2000	single	Host	139.06 ms	±14.92 ms	—	—
		double	Host	229.70 ms	±32.49 ms	—	—
	4000x4000	single	Host	728.64 ms	±63.85 ms	—	—
pseudoInvert	1000x1000	single	Host	304.75 ms	±24.19 ms	—	—
		double	Host	430.58 ms	±24.67 ms	—	—
sVD	1000x1000	single	Host	272.86 ms	±4.36 ms	—	—
		double	Host	408.41 ms	±38.34 ms	—	—
symEigDec	1000x1000	single	Host	75.49 ms	±6.50 ms	—	—
		double	Host	127.58 ms	±18.01 ms	—	—
	2000x2000	single	Host	346.69 ms	±11.95 ms	—	—
		double	Host	593.21 ms	±21.09 ms	—	—
equal	1000x1000	bool	Host	5.80 ms	±0.17 ms	—	—
			Cuda	1.00 ms	±0.01 ms	—	—
		int32	Host	4.27 ms	±0.28 ms	0.44 ms	±0.01 ms
			Cuda	0.99 ms	±0.03 ms	—	—
		int64	Host	3.79 ms	±0.09 ms	1.16 ms	±0.01 ms
			Cuda	1.01 ms	±0.02 ms	—	—
		single	Host	4.43 ms	±0.12 ms	0.45 ms	±0.01 ms
			Cuda	1.02 ms	±0.02 ms	—	—
		double	Host	4.08 ms	±0.22 ms	1.15 ms	±0.02 ms
			Cuda	1.03 ms	±0.05 ms	—	—
	2000x2000	bool	Host	15.78 ms	±1.22 ms	—	—
			Cuda	1.76 ms	±0.08 ms	—	—
		int32	Host	14.85 ms	±0.81 ms	2.62 ms	±0.01 ms
			Cuda	1.71 ms	±0.07 ms	—	—
		int64	Host	14.65 ms	±0.63 ms	4.89 ms	±0.00 ms
			Cuda	1.86 ms	±0.05 ms	—	—
		single	Host	14.48 ms	±0.43 ms	2.62 ms	±0.02 ms
			Cuda	1.76 ms	±0.06 ms	—	—
		double	Host	14.59 ms	±0.70 ms	4.88 ms	±0.01 ms
			Cuda	1.91 ms	±0.09 ms	—	—
	4000x4000	bool	Host	51.58 ms	±2.01 ms	—	—
			Cuda	5.05 ms	±0.05 ms	—	—
		int32	Host	54.05 ms	±3.07 ms	10.69 ms	±0.01 ms
			Cuda	5.02 ms	±0.11 ms	—	—
		int64	Host	53.66 ms	±2.79 ms	19.84 ms	±0.02 ms
			Cuda	5.77 ms	±0.12 ms	—	—
		single	Host	52.47 ms	±1.64 ms	10.65 ms	±0.03 ms
			Cuda	4.97 ms	±0.13 ms	—	—
		double	Host	60.88 ms	±11.53 ms	19.95 ms	±0.06 ms
			Cuda	5.76 ms	±0.13 ms	—	—
notEqual	1000x1000	bool	Host	5.64 ms	±0.21 ms	—	—
			Cuda	1.09 ms	±0.02 ms	—	—
		int32	Host	4.33 ms	±0.19 ms	0.44 ms	±0.01 ms
			Cuda	0.93 ms	±0.03 ms	—	—
		int64	Host	3.82 ms	±0.16 ms	1.16 ms	±0.01 ms
			Cuda	0.98 ms	±0.04 ms	—	—
		single	Host	4.02 ms	±0.25 ms	0.45 ms	±0.01 ms
			Cuda	1.08 ms	±0.04 ms	—	—
		double	Host	3.83 ms	±0.08 ms	1.15 ms	±0.02 ms
			Cuda	0.98 ms	±0.04 ms	—	—
	2000x2000	bool	Host	14.22 ms	±0.80 ms	—	—
			Cuda	1.76 ms	±0.02 ms	—	—
		int32	Host	15.77 ms	±1.49 ms	2.62 ms	±0.01 ms
			Cuda	1.88 ms	±0.09 ms	—	—
		int64	Host	15.32 ms	±0.74 ms	4.82 ms	±0.01 ms
			Cuda	1.88 ms	±0.04 ms	—	—
		single	Host	14.82 ms	±0.47 ms	2.62 ms	±0.02 ms
			Cuda	1.82 ms	±0.04 ms	—	—
		double	Host	15.29 ms	±0.38 ms	4.81 ms	±0.02 ms
			Cuda	1.99 ms	±0.11 ms	—	—
	4000x4000	bool	Host	51.50 ms	±2.27 ms	—	—
			Cuda	4.94 ms	±0.05 ms	—	—
		int32	Host	53.56 ms	±0.97 ms	10.68 ms	±0.02 ms
			Cuda	5.02 ms	±0.06 ms	—	—
		int64	Host	63.66 ms	±15.63 ms	19.84 ms	±0.02 ms
			Cuda	5.79 ms	±0.08 ms	—	—
		single	Host	53.23 ms	±1.42 ms	10.76 ms	±0.03 ms
			Cuda	5.05 ms	±0.09 ms	—	—
		double	Host	55.27 ms	±1.34 ms	19.99 ms	±0.02 ms
			Cuda	5.76 ms	±0.18 ms	—	—
less	1000x1000	bool	Host	5.92 ms	±0.20 ms	—	—
			Cuda	1.08 ms	±0.02 ms	—	—
		int32	Host	4.18 ms	±0.32 ms	0.44 ms	±0.01 ms
			Cuda	0.97 ms	±0.04 ms	—	—
		int64	Host	3.94 ms	±0.12 ms	1.15 ms	±0.01 ms
			Cuda	0.95 ms	±0.03 ms	—	—
		single	Host	4.79 ms	±0.99 ms	0.45 ms	±0.01 ms
			Cuda	1.08 ms	±0.06 ms	—	—
		double	Host	3.96 ms	±0.27 ms	1.15 ms	±0.02 ms
			Cuda	0.98 ms	±0.05 ms	—	—
	2000x2000	bool	Host	14.27 ms	±0.73 ms	—	—
			Cuda	1.77 ms	±0.06 ms	—	—
		int32	Host	14.92 ms	±1.04 ms	2.60 ms	±0.01 ms
			Cuda	1.88 ms	±0.05 ms	—	—
		int64	Host	15.32 ms	±0.62 ms	4.81 ms	±0.01 ms
			Cuda	1.90 ms	±0.03 ms	—	—
		single	Host	15.71 ms	±0.60 ms	2.62 ms	±0.02 ms
			Cuda	1.76 ms	±0.02 ms	—	—
		double	Host	15.82 ms	±2.04 ms	4.83 ms	±0.02 ms
			Cuda	2.03 ms	±0.08 ms	—	—
	4000x4000	bool	Host	60.34 ms	±15.67 ms	—	—
			Cuda	5.03 ms	±0.17 ms	—	—
		int32	Host	55.27 ms	±1.66 ms	10.63 ms	±0.01 ms
			Cuda	5.07 ms	±0.07 ms	—	—
		int64	Host	55.28 ms	±2.07 ms	19.93 ms	±0.01 ms
			Cuda	5.81 ms	±0.20 ms	—	—
		single	Host	51.89 ms	±1.47 ms	10.76 ms	±0.03 ms
			Cuda	5.07 ms	±0.09 ms	—	—
		double	Host	54.03 ms	±2.59 ms	19.99 ms	±0.07 ms
			Cuda	5.69 ms	±0.17 ms	—	—
lessOrEqual	1000x1000	bool	Host	5.69 ms	±0.37 ms	—	—
			Cuda	1.10 ms	±0.04 ms	—	—
		int32	Host	4.19 ms	±0.34 ms	0.44 ms	±0.01 ms
			Cuda	0.97 ms	±0.02 ms	—	—
		int64	Host	4.04 ms	±0.09 ms	1.16 ms	±0.01 ms
			Cuda	1.01 ms	±0.02 ms	—	—
		single	Host	4.40 ms	±0.47 ms	0.45 ms	±0.01 ms
			Cuda	1.10 ms	±0.03 ms	—	—
		double	Host	3.84 ms	±0.23 ms	1.15 ms	±0.02 ms
			Cuda	1.02 ms	±0.02 ms	—	—
	2000x2000	bool	Host	14.76 ms	±0.76 ms	—	—
			Cuda	1.76 ms	±0.03 ms	—	—
		int32	Host	14.51 ms	±0.57 ms	2.61 ms	±0.01 ms
			Cuda	1.85 ms	±0.04 ms	—	—
		int64	Host	14.33 ms	±0.39 ms	4.82 ms	±0.01 ms
			Cuda	1.91 ms	±0.09 ms	—	—
		single	Host	14.26 ms	±0.54 ms	2.62 ms	±0.02 ms
			Cuda	1.76 ms	±0.04 ms	—	—
		double	Host	15.46 ms	±1.43 ms	4.83 ms	±0.02 ms
			Cuda	1.99 ms	±0.06 ms	—	—
	4000x4000	bool	Host	53.10 ms	±3.88 ms	—	—
			Cuda	4.98 ms	±0.06 ms	—	—
		int32	Host	53.08 ms	±0.79 ms	10.68 ms	±0.01 ms
			Cuda	5.05 ms	±0.15 ms	—	—
		int64	Host	54.34 ms	±1.42 ms	19.95 ms	±0.01 ms
			Cuda	5.86 ms	±0.12 ms	—	—
		single	Host	54.65 ms	±1.26 ms	10.81 ms	±0.17 ms
			Cuda	4.97 ms	±0.08 ms	—	—
		double	Host	54.10 ms	±2.05 ms	19.99 ms	±0.07 ms
			Cuda	5.88 ms	±0.18 ms	—	—
greater	1000x1000	bool	Host	5.86 ms	±0.38 ms	—	—
			Cuda	0.83 ms	±0.06 ms	—	—
		int32	Host	4.21 ms	±0.23 ms	0.44 ms	±0.01 ms
			Cuda	0.99 ms	±0.03 ms	—	—
		int64	Host	3.73 ms	±0.17 ms	1.16 ms	±0.01 ms
			Cuda	0.94 ms	±0.01 ms	—	—
		single	Host	3.91 ms	±0.11 ms	0.45 ms	±0.01 ms
			Cuda	0.96 ms	±0.02 ms	—	—
		double	Host	4.09 ms	±0.42 ms	1.15 ms	±0.02 ms
			Cuda	1.11 ms	±0.03 ms	—	—
	2000x2000	bool	Host	14.71 ms	±1.57 ms	—	—
			Cuda	1.82 ms	±0.04 ms	—	—
		int32	Host	15.23 ms	±0.59 ms	2.61 ms	±0.01 ms
			Cuda	1.77 ms	±0.04 ms	—	—
		int64	Host	14.67 ms	±0.73 ms	4.84 ms	±0.01 ms
			Cuda	1.93 ms	±0.04 ms	—	—
		single	Host	14.67 ms	±0.20 ms	2.61 ms	±0.02 ms
			Cuda	1.79 ms	±0.03 ms	—	—
		double	Host	15.08 ms	±0.72 ms	4.83 ms	±0.02 ms
			Cuda	1.90 ms	±0.05 ms	—	—
	4000x4000	bool	Host	62.81 ms	±15.01 ms	—	—
			Cuda	5.09 ms	±0.12 ms	—	—
		int32	Host	54.15 ms	±2.34 ms	10.70 ms	±0.13 ms
			Cuda	5.02 ms	±0.12 ms	—	—
		int64	Host	53.61 ms	±1.97 ms	19.95 ms	±0.03 ms
			Cuda	5.78 ms	±0.13 ms	—	—
		single	Host	54.14 ms	±3.81 ms	10.75 ms	±0.03 ms
			Cuda	4.98 ms	±0.06 ms	—	—
		double	Host	54.40 ms	±1.30 ms	19.95 ms	±0.05 ms
			Cuda	5.71 ms	±0.13 ms	—	—
greaterOrEqual	1000x1000	bool	Host	5.86 ms	±0.28 ms	—	—
			Cuda	1.01 ms	±0.03 ms	—	—
		int32	Host	3.99 ms	±0.30 ms	0.44 ms	±0.01 ms
			Cuda	0.95 ms	±0.04 ms	—	—
		int64	Host	3.90 ms	±0.19 ms	1.16 ms	±0.01 ms
			Cuda	0.98 ms	±0.02 ms	—	—
		single	Host	3.98 ms	±0.11 ms	0.45 ms	±0.01 ms
			Cuda	1.02 ms	±0.07 ms	—	—
		double	Host	4.21 ms	±0.24 ms	1.15 ms	±0.02 ms
			Cuda	0.96 ms	±0.02 ms	—	—
	2000x2000	bool	Host	14.82 ms	±0.81 ms	—	—
			Cuda	1.77 ms	±0.07 ms	—	—
		int32	Host	14.86 ms	±0.80 ms	2.59 ms	±0.01 ms
			Cuda	1.73 ms	±0.05 ms	—	—
		int64	Host	14.69 ms	±0.51 ms	4.82 ms	±0.01 ms
			Cuda	1.89 ms	±0.04 ms	—	—
		single	Host	15.06 ms	±1.13 ms	2.60 ms	±0.02 ms
			Cuda	1.75 ms	±0.05 ms	—	—
		double	Host	15.63 ms	±0.54 ms	4.83 ms	±0.02 ms
			Cuda	1.94 ms	±0.05 ms	—	—
	4000x4000	bool	Host	54.91 ms	±0.60 ms	—	—
			Cuda	4.94 ms	±0.03 ms	—	—
		int32	Host	54.11 ms	±1.68 ms	10.75 ms	±0.02 ms
			Cuda	5.03 ms	±0.16 ms	—	—
		int64	Host	54.11 ms	±1.69 ms	19.91 ms	±0.01 ms
			Cuda	5.77 ms	±0.20 ms	—	—
		single	Host	54.39 ms	±2.23 ms	10.72 ms	±0.03 ms
			Cuda	4.98 ms	±0.13 ms	—	—
		double	Host	53.81 ms	±2.73 ms	19.97 ms	±0.01 ms
			Cuda	5.83 ms	±0.24 ms	—	—
not	1000x1000	bool	Host	5.52 ms	±0.15 ms	0.11 ms	±0.00 ms
			Cuda	0.90 ms	±0.05 ms	—	—
	2000x2000		Host	14.59 ms	±1.19 ms	0.45 ms	±0.01 ms
			Cuda	1.54 ms	±0.06 ms	—	—
	4000x4000		Host	45.90 ms	±2.41 ms	2.82 ms	±0.02 ms
			Cuda	4.78 ms	±0.16 ms	—	—
and	1000x1000	bool	Host	5.80 ms	±0.17 ms	0.14 ms	±0.00 ms
			Cuda	1.00 ms	±0.05 ms	—	—
	2000x2000		Host	17.41 ms	±0.78 ms	0.79 ms	±0.01 ms
			Cuda	1.64 ms	±0.05 ms	—	—
	4000x4000		Host	57.37 ms	±8.14 ms	3.90 ms	±0.02 ms
			Cuda	5.00 ms	±0.05 ms	—	—
or	1000x1000	bool	Host	6.05 ms	±1.03 ms	0.14 ms	±0.00 ms
			Cuda	1.05 ms	±0.02 ms	—	—
	2000x2000		Host	16.03 ms	±1.36 ms	0.78 ms	±0.01 ms
			Cuda	1.75 ms	±0.05 ms	—	—
	4000x4000		Host	54.88 ms	±4.52 ms	3.87 ms	±0.01 ms
			Cuda	4.88 ms	±0.08 ms	—	—
xor	1000x1000	bool	Host	5.79 ms	±0.09 ms	1.53 ms	±0.00 ms
			Cuda	1.01 ms	±0.02 ms	—	—
	2000x2000		Host	15.03 ms	±0.24 ms	6.15 ms	±0.01 ms
			Cuda	1.78 ms	±0.06 ms	—	—
	4000x4000		Host	52.76 ms	±1.94 ms	24.66 ms	±0.01 ms
			Cuda	4.96 ms	±0.10 ms	—	—
countTrueAxis	1000x1000	bool	Host	3.42 ms	±0.11 ms	1.59 ms	±0.00 ms
	2000x2000		Host	12.76 ms	±0.70 ms	6.25 ms	±0.01 ms
	4000x4000		Host	44.79 ms	±1.56 ms	24.97 ms	±0.05 ms
allAxis	1000x1000	bool	Host	3.04 ms	±0.43 ms	—	—
			Cuda	1.14 ms	±0.02 ms	—	—
	2000x2000		Host	10.79 ms	±0.52 ms	—	—
			Cuda	2.12 ms	±0.03 ms	—	—
	4000x4000		Host	43.31 ms	±2.21 ms	—	—
			Cuda	4.11 ms	±0.01 ms	—	—
anyAxis	1000x1000	bool	Host	3.10 ms	±0.28 ms	—	—
			Cuda	1.15 ms	±0.04 ms	—	—
	2000x2000		Host	12.73 ms	±0.49 ms	—	—
			Cuda	2.16 ms	±0.05 ms	—	—
	4000x4000		Host	42.06 ms	±2.30 ms	—	—
			Cuda	4.12 ms	±0.02 ms	—	—
trueIndices	1000x1000	bool	Host	33.62 ms	±2.85 ms	—	—
	2000x2000		Host	114.49 ms	±4.33 ms	—	—
	4000x4000		Host	459.89 ms	±2.51 ms	—	—
ifThenElse	1000x1000	bool	Host	6.35 ms	±0.19 ms	—	—
			Cuda	1.12 ms	±0.03 ms	—	—
		int32	Host	5.80 ms	±0.25 ms	7.00 ms	±0.01 ms
			Cuda	0.98 ms	±0.05 ms	—	—
		int64	Host	6.49 ms	±0.20 ms	7.22 ms	±0.01 ms
			Cuda	1.17 ms	±0.05 ms	—	—
		single	Host	6.39 ms	±0.58 ms	7.00 ms	±0.01 ms
			Cuda	1.09 ms	±0.03 ms	—	—
		double	Host	6.72 ms	±0.37 ms	7.22 ms	±0.01 ms
			Cuda	1.13 ms	±0.04 ms	—	—
	2000x2000	bool	Host	17.06 ms	±0.77 ms	—	—
			Cuda	2.05 ms	±0.06 ms	—	—
		int32	Host	17.83 ms	±0.23 ms	28.11 ms	±0.01 ms
			Cuda	2.19 ms	±0.06 ms	—	—
		int64	Host	20.08 ms	±0.81 ms	28.98 ms	±0.01 ms
			Cuda	2.63 ms	±0.10 ms	—	—
		single	Host	17.64 ms	±0.34 ms	28.12 ms	±0.01 ms
			Cuda	2.13 ms	±0.09 ms	—	—
		double	Host	20.90 ms	±0.59 ms	29.01 ms	±0.02 ms
			Cuda	2.55 ms	±0.06 ms	—	—
	4000x4000	bool	Host	59.95 ms	±1.70 ms	—	—
			Cuda	6.08 ms	±0.05 ms	—	—
		int32	Host	68.17 ms	±2.13 ms	140.58 ms	±0.02 ms
			Cuda	6.33 ms	±0.11 ms	—	—
		int64	Host	74.97 ms	±2.84 ms	173.15 ms	±0.03 ms
			Cuda	9.94 ms	±0.42 ms	—	—
		single	Host	66.09 ms	±1.34 ms	140.63 ms	±0.03 ms
			Cuda	6.29 ms	±0.19 ms	—	—
		double	Host	74.37 ms	±0.80 ms	173.27 ms	±0.10 ms
			Cuda	10.17 ms	±1.07 ms	—	—
identity	1000x1000	int32	Host	—	—	0.31 ms	±0.00 ms
		int64	Host	—	—	0.61 ms	±0.01 ms
		single	Host	—	—	0.31 ms	±0.00 ms
		double	Host	—	—	0.61 ms	±0.01 ms
	2000x2000	int32	Host	—	—	1.65 ms	±0.02 ms
		int64	Host	—	—	3.64 ms	±0.01 ms
		single	Host	—	—	1.65 ms	±0.02 ms
		double	Host	—	—	3.64 ms	±0.01 ms
	4000x4000	int32	Host	—	—	6.83 ms	±0.02 ms
		int64	Host	—	—	6.87 ms	±0.02 ms
		single	Host	—	—	6.84 ms	±0.02 ms
		double	Host	—	—	6.87 ms	±0.02 ms

Report generated on Tue, 08 May 2018 12:47:12 GMT.