F# Tensor 0.4.9 Hostname: mint3 CPU: Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz GPU: nVidia Tesla K40c OS: Linux 4.13.0-38-generic #43~16.04.1-Ubuntu SMP Wed Mar 14 17:48:43 UTC 2018 Runtime: .NET SDK 2.1.300-preview2-008533 Time: 2018-05-01 00:00:00Z		NumPy 1.14.0 Hostname: mint3 CPU: Intel(R) Core(TM) i7-4770 CPU @ 3.40GHz GPU: nVidia Tesla K40c OS: Linux 4.13.0-38-generic #43~16.04.1-Ubuntu SMP Wed Mar 14 17:48:43 UTC 2018 Runtime: Anaconda3 5.1.0 Time: 2018-05-01 00:00:00Z
Function	Shape	Type	Device	Mean	StdDev	Mean	StdDev
zeros	1000x1000	bool	Host	0.07 ms	±0.02 ms	—	—
		int32	Host	0.08 ms	±0.00 ms	0.27 ms	±0.01 ms
			Cuda	0.62 ms	±0.01 ms	—	—
		int64	Host	0.15 ms	±0.00 ms	0.56 ms	±0.01 ms
			Cuda	0.63 ms	±0.00 ms	—	—
		single	Host	0.08 ms	±0.00 ms	0.27 ms	±0.00 ms
			Cuda	0.62 ms	±0.02 ms	—	—
		double	Host	0.15 ms	±0.00 ms	0.58 ms	±0.01 ms
			Cuda	0.60 ms	±0.01 ms	—	—
	2000x2000	bool	Host	0.11 ms	±0.00 ms	—	—
		int32	Host	0.25 ms	±0.00 ms	1.59 ms	±0.02 ms
			Cuda	0.78 ms	±0.00 ms	—	—
		int64	Host	0.54 ms	±0.02 ms	3.56 ms	±0.04 ms
			Cuda	0.74 ms	±0.00 ms	—	—
		single	Host	0.25 ms	±0.00 ms	1.59 ms	±0.02 ms
			Cuda	0.77 ms	±0.00 ms	—	—
		double	Host	0.53 ms	±0.02 ms	3.56 ms	±0.04 ms
			Cuda	0.78 ms	±0.00 ms	—	—
	4000x4000	bool	Host	0.33 ms	±0.00 ms	—	—
		int32	Host	1.21 ms	±0.09 ms	0.01 ms	±0.00 ms
			Cuda	1.41 ms	±0.01 ms	—	—
		int64	Host	0.70 ms	±0.00 ms	0.01 ms	±0.00 ms
			Cuda	11.25 ms	±0.05 ms	—	—
		single	Host	1.05 ms	±0.07 ms	0.01 ms	±0.00 ms
			Cuda	1.41 ms	±0.01 ms	—	—
		double	Host	0.69 ms	±0.00 ms	0.01 ms	±0.00 ms
			Cuda	11.08 ms	±0.04 ms	—	—
ones	1000x1000	int32	Host	1.96 ms	±0.01 ms	0.28 ms	±0.00 ms
			Cuda	0.64 ms	±0.01 ms	—	—
		int64	Host	3.83 ms	±0.02 ms	0.66 ms	±0.01 ms
			Cuda	0.62 ms	±0.01 ms	—	—
		single	Host	1.94 ms	±0.01 ms	0.28 ms	±0.00 ms
			Cuda	0.61 ms	±0.00 ms	—	—
		double	Host	3.74 ms	±0.01 ms	0.66 ms	±0.01 ms
			Cuda	0.64 ms	±0.00 ms	—	—
	2000x2000	int32	Host	7.64 ms	±0.04 ms	1.60 ms	±0.02 ms
			Cuda	0.78 ms	±0.01 ms	—	—
		int64	Host	15.86 ms	±0.07 ms	3.54 ms	±0.01 ms
			Cuda	0.76 ms	±0.01 ms	—	—
		single	Host	7.87 ms	±0.02 ms	1.61 ms	±0.02 ms
			Cuda	0.78 ms	±0.01 ms	—	—
		double	Host	15.98 ms	±0.04 ms	3.54 ms	±0.01 ms
			Cuda	0.76 ms	±0.01 ms	—	—
	4000x4000	int32	Host	28.07 ms	±0.03 ms	28.85 ms	±0.03 ms
			Cuda	1.40 ms	±0.01 ms	—	—
		int64	Host	55.67 ms	±0.04 ms	61.94 ms	±0.06 ms
			Cuda	11.25 ms	±0.05 ms	—	—
		single	Host	28.03 ms	±0.02 ms	28.86 ms	±0.02 ms
			Cuda	1.39 ms	±0.03 ms	—	—
		double	Host	55.70 ms	±0.01 ms	61.97 ms	±0.06 ms
			Cuda	11.24 ms	±0.04 ms	—	—
arange	1000x1000	int32	Host	15.04 ms	±0.01 ms	—	—
			Cuda	0.60 ms	±0.01 ms	—	—
		int64	Host	17.01 ms	±0.12 ms	—	—
			Cuda	0.59 ms	±0.00 ms	—	—
		single	Host	17.92 ms	±5.11 ms	—	—
			Cuda	0.62 ms	±0.00 ms	—	—
		double	Host	17.56 ms	±0.10 ms	—	—
			Cuda	0.62 ms	±0.01 ms	—	—
	2000x2000	int32	Host	59.41 ms	±0.04 ms	—	—
			Cuda	0.67 ms	±0.00 ms	—	—
		int64	Host	65.63 ms	±0.15 ms	—	—
			Cuda	0.72 ms	±0.00 ms	—	—
		single	Host	60.09 ms	±0.07 ms	—	—
			Cuda	0.70 ms	±0.01 ms	—	—
		double	Host	67.80 ms	±0.09 ms	—	—
			Cuda	0.70 ms	±0.00 ms	—	—
	4000x4000	int32	Host	235.30 ms	±0.62 ms	—	—
			Cuda	7.32 ms	±0.01 ms	—	—
		int64	Host	282.10 ms	±42.06 ms	—	—
			Cuda	10.15 ms	±0.04 ms	—	—
		single	Host	240.71 ms	±0.64 ms	—	—
			Cuda	6.49 ms	±0.01 ms	—	—
		double	Host	269.27 ms	±0.51 ms	—	—
			Cuda	9.52 ms	±0.04 ms	—	—
copy	1000x1000	bool	Host	0.30 ms	±0.00 ms	—	—
			Cuda	0.58 ms	±0.02 ms	—	—
		int32	Host	2.24 ms	±0.01 ms	0.38 ms	±0.01 ms
			Cuda	0.62 ms	±0.01 ms	—	—
		int64	Host	4.95 ms	±0.03 ms	0.79 ms	±0.01 ms
			Cuda	0.65 ms	±0.01 ms	—	—
		single	Host	2.21 ms	±0.01 ms	0.38 ms	±0.01 ms
			Cuda	0.61 ms	±0.01 ms	—	—
		double	Host	4.97 ms	±0.02 ms	0.80 ms	±0.01 ms
			Cuda	0.65 ms	±0.01 ms	—	—
	2000x2000	bool	Host	2.28 ms	±0.01 ms	—	—
			Cuda	0.61 ms	±0.00 ms	—	—
		int32	Host	10.10 ms	±0.04 ms	1.77 ms	±0.01 ms
			Cuda	0.70 ms	±0.00 ms	—	—
		int64	Host	19.30 ms	±0.01 ms	3.60 ms	±0.01 ms
			Cuda	0.86 ms	±0.01 ms	—	—
		single	Host	10.27 ms	±0.05 ms	1.82 ms	±0.03 ms
			Cuda	0.72 ms	±0.00 ms	—	—
		double	Host	19.36 ms	±0.00 ms	3.65 ms	±0.01 ms
			Cuda	0.87 ms	±0.00 ms	—	—
	4000x4000	bool	Host	10.19 ms	±0.07 ms	—	—
			Cuda	5.08 ms	±0.01 ms	—	—
		int32	Host	38.32 ms	±0.01 ms	40.94 ms	±0.02 ms
			Cuda	7.04 ms	±0.02 ms	—	—
		int64	Host	76.16 ms	±0.24 ms	82.02 ms	±0.04 ms
			Cuda	14.06 ms	±0.02 ms	—	—
		single	Host	38.29 ms	±0.25 ms	40.95 ms	±0.04 ms
			Cuda	7.02 ms	±0.01 ms	—	—
		double	Host	75.81 ms	±0.01 ms	82.04 ms	±0.02 ms
			Cuda	14.16 ms	±0.06 ms	—	—
negate	1000x1000	bool	Cuda	0.66 ms	±0.00 ms	—	—
		int32	Host	2.25 ms	±0.01 ms	0.46 ms	±0.01 ms
			Cuda	0.63 ms	±0.03 ms	—	—
		int64	Host	4.68 ms	±0.02 ms	1.27 ms	±0.01 ms
			Cuda	0.66 ms	±0.02 ms	—	—
		single	Host	2.22 ms	±0.01 ms	0.45 ms	±0.01 ms
			Cuda	0.68 ms	±0.01 ms	—	—
		double	Host	4.67 ms	±0.03 ms	1.28 ms	±0.01 ms
			Cuda	0.67 ms	±0.01 ms	—	—
	2000x2000	bool	Cuda	0.86 ms	±0.01 ms	—	—
		int32	Host	9.17 ms	±0.04 ms	2.78 ms	±0.01 ms
			Cuda	0.89 ms	±0.01 ms	—	—
		int64	Host	17.20 ms	±0.00 ms	5.85 ms	±0.76 ms
			Cuda	0.90 ms	±0.01 ms	—	—
		single	Host	9.15 ms	±0.03 ms	2.88 ms	±0.02 ms
			Cuda	0.89 ms	±0.01 ms	—	—
		double	Host	17.19 ms	±0.00 ms	5.65 ms	±0.01 ms
			Cuda	0.89 ms	±0.00 ms	—	—
	4000x4000	bool	Cuda	1.77 ms	±0.02 ms	—	—
		int32	Host	34.49 ms	±0.00 ms	33.18 ms	±0.10 ms
			Cuda	1.85 ms	±0.02 ms	—	—
		int64	Host	67.27 ms	±0.02 ms	65.78 ms	±0.04 ms
			Cuda	12.30 ms	±0.05 ms	—	—
		single	Host	34.36 ms	±0.03 ms	33.74 ms	±0.07 ms
			Cuda	1.84 ms	±0.02 ms	—	—
		double	Host	68.45 ms	±0.38 ms	66.81 ms	±0.06 ms
			Cuda	12.24 ms	±0.09 ms	—	—
add	1000x1000	bool	Cuda	0.75 ms	±0.00 ms	—	—
		int32	Host	2.64 ms	±0.00 ms	0.78 ms	±0.01 ms
			Cuda	1.11 ms	±0.01 ms	—	—
		int64	Host	5.53 ms	±0.07 ms	1.97 ms	±0.11 ms
			Cuda	1.56 ms	±0.02 ms	—	—
		single	Host	2.65 ms	±0.01 ms	0.79 ms	±0.01 ms
			Cuda	0.95 ms	±0.23 ms	—	—
		double	Host	5.39 ms	±0.03 ms	1.93 ms	±0.01 ms
			Cuda	1.55 ms	±0.01 ms	—	—
	2000x2000	bool	Cuda	0.97 ms	±0.00 ms	—	—
		int32	Host	10.55 ms	±0.04 ms	3.89 ms	±0.01 ms
			Cuda	0.95 ms	±0.01 ms	—	—
		int64	Host	19.43 ms	±0.09 ms	8.27 ms	±0.81 ms
			Cuda	1.07 ms	±0.00 ms	—	—
		single	Host	10.62 ms	±0.04 ms	3.88 ms	±0.01 ms
			Cuda	0.99 ms	±0.01 ms	—	—
		double	Host	19.80 ms	±0.00 ms	7.81 ms	±0.01 ms
			Cuda	1.09 ms	±0.02 ms	—	—
	4000x4000	bool	Cuda	2.22 ms	±0.02 ms	—	—
		int32	Host	39.57 ms	±0.02 ms	36.95 ms	±0.08 ms
			Cuda	2.24 ms	±0.02 ms	—	—
		int64	Host	76.88 ms	±0.01 ms	73.74 ms	±0.02 ms
			Cuda	2.70 ms	±0.06 ms	—	—
		single	Host	39.70 ms	±0.21 ms	37.40 ms	±0.08 ms
			Cuda	2.26 ms	±0.06 ms	—	—
		double	Host	77.70 ms	±0.02 ms	74.86 ms	±0.02 ms
			Cuda	2.67 ms	±0.06 ms	—	—
subtract	1000x1000	bool	Cuda	1.12 ms	±0.00 ms	—	—
		int32	Host	2.61 ms	±0.05 ms	0.78 ms	±0.01 ms
			Cuda	1.14 ms	±0.00 ms	—	—
		int64	Host	5.34 ms	±0.03 ms	1.95 ms	±0.11 ms
			Cuda	1.55 ms	±0.00 ms	—	—
		single	Host	2.64 ms	±0.01 ms	0.79 ms	±0.01 ms
			Cuda	0.76 ms	±0.00 ms	—	—
		double	Host	5.34 ms	±0.06 ms	1.92 ms	±0.01 ms
			Cuda	0.72 ms	±0.00 ms	—	—
	2000x2000	bool	Cuda	0.98 ms	±0.00 ms	—	—
		int32	Host	10.50 ms	±0.04 ms	3.91 ms	±0.02 ms
			Cuda	0.98 ms	±0.01 ms	—	—
		int64	Host	19.86 ms	±0.10 ms	8.27 ms	±0.81 ms
			Cuda	1.08 ms	±0.01 ms	—	—
		single	Host	10.57 ms	±0.05 ms	3.90 ms	±0.01 ms
			Cuda	0.97 ms	±0.01 ms	—	—
		double	Host	19.72 ms	±0.01 ms	7.81 ms	±0.01 ms
			Cuda	1.06 ms	±0.01 ms	—	—
	4000x4000	bool	Cuda	2.23 ms	±0.02 ms	—	—
		int32	Host	39.28 ms	±0.03 ms	36.97 ms	±0.08 ms
			Cuda	2.25 ms	±0.03 ms	—	—
		int64	Host	77.18 ms	±0.31 ms	73.91 ms	±0.01 ms
			Cuda	2.69 ms	±0.06 ms	—	—
		single	Host	39.50 ms	±0.01 ms	37.43 ms	±0.08 ms
			Cuda	2.22 ms	±0.01 ms	—	—
		double	Host	78.55 ms	±0.84 ms	74.93 ms	±0.03 ms
			Cuda	2.67 ms	±0.06 ms	—	—
multiply	1000x1000	bool	Cuda	1.09 ms	±0.02 ms	—	—
		int32	Host	2.69 ms	±0.01 ms	0.78 ms	±0.01 ms
			Cuda	0.73 ms	±0.00 ms	—	—
		int64	Host	13.25 ms	±0.11 ms	1.91 ms	±0.12 ms
			Cuda	0.72 ms	±0.01 ms	—	—
		single	Host	2.66 ms	±0.02 ms	0.78 ms	±0.01 ms
			Cuda	1.13 ms	±0.01 ms	—	—
		double	Host	5.36 ms	±0.05 ms	1.93 ms	±0.01 ms
			Cuda	0.70 ms	±0.00 ms	—	—
	2000x2000	bool	Cuda	0.99 ms	±0.00 ms	—	—
		int32	Host	10.93 ms	±0.03 ms	3.91 ms	±0.02 ms
			Cuda	1.00 ms	±0.00 ms	—	—
		int64	Host	50.69 ms	±0.03 ms	7.90 ms	±0.88 ms
			Cuda	1.08 ms	±0.00 ms	—	—
		single	Host	10.82 ms	±0.11 ms	3.88 ms	±0.01 ms
			Cuda	0.97 ms	±0.00 ms	—	—
		double	Host	19.64 ms	±0.01 ms	7.83 ms	±0.03 ms
			Cuda	1.06 ms	±0.00 ms	—	—
	4000x4000	bool	Cuda	2.21 ms	±0.02 ms	—	—
		int32	Host	39.95 ms	±0.02 ms	37.25 ms	±0.08 ms
			Cuda	2.27 ms	±0.02 ms	—	—
		int64	Host	202.93 ms	±0.11 ms	77.64 ms	±0.01 ms
			Cuda	2.70 ms	±0.06 ms	—	—
		single	Host	39.84 ms	±0.02 ms	37.36 ms	±0.06 ms
			Cuda	2.25 ms	±0.04 ms	—	—
		double	Host	78.21 ms	±0.90 ms	74.78 ms	±0.02 ms
			Cuda	2.66 ms	±0.06 ms	—	—
divide	1000x1000	bool	Cuda	0.72 ms	±0.00 ms	—	—
		int32	Host	—	—	4.74 ms	±0.01 ms
			Cuda	0.71 ms	±0.01 ms	—	—
		int64	Host	—	—	5.58 ms	±0.01 ms
			Cuda	0.83 ms	±0.01 ms	—	—
		single	Host	2.86 ms	±0.01 ms	0.90 ms	±0.01 ms
			Cuda	0.72 ms	±0.00 ms	—	—
		double	Host	7.26 ms	±0.04 ms	3.54 ms	±0.01 ms
			Cuda	1.53 ms	±0.01 ms	—	—
	2000x2000	bool	Cuda	1.04 ms	±0.00 ms	—	—
		int32	Host	—	—	18.89 ms	±0.01 ms
			Cuda	1.06 ms	±0.00 ms	—	—
		int64	Host	—	—	22.76 ms	±0.10 ms
			Cuda	1.59 ms	±0.01 ms	—	—
		single	Host	11.58 ms	±0.11 ms	3.87 ms	±0.01 ms
			Cuda	1.03 ms	±0.00 ms	—	—
		double	Host	27.79 ms	±0.01 ms	14.22 ms	±0.02 ms
			Cuda	1.13 ms	±0.04 ms	—	—
	4000x4000	bool	Cuda	2.43 ms	±0.02 ms	—	—
		int32	Host	—	—	130.69 ms	±0.02 ms
			Cuda	2.53 ms	±0.02 ms	—	—
		int64	Host	—	—	142.73 ms	±0.04 ms
			Cuda	4.80 ms	±0.10 ms	—	—
		single	Host	42.34 ms	±0.01 ms	43.09 ms	±0.02 ms
			Cuda	2.42 ms	±0.02 ms	—	—
		double	Host	110.59 ms	±0.04 ms	112.29 ms	±0.03 ms
			Cuda	2.81 ms	±0.06 ms	—	—
power	1000x1000	single	Host	43.41 ms	±0.94 ms	83.52 ms	±0.03 ms
			Cuda	0.76 ms	±0.01 ms	—	—
		double	Host	42.22 ms	±0.65 ms	67.25 ms	±0.03 ms
			Cuda	0.92 ms	±0.00 ms	—	—
	2000x2000	single	Host	165.51 ms	±0.66 ms	333.76 ms	±0.08 ms
			Cuda	1.22 ms	±0.00 ms	—	—
		double	Host	164.29 ms	±2.15 ms	267.09 ms	±0.09 ms
			Cuda	1.76 ms	±0.01 ms	—	—
	4000x4000	single	Host	657.84 ms	±3.20 ms	1365.24 ms	±0.32 ms
			Cuda	3.17 ms	±0.03 ms	—	—
		double	Host	642.87 ms	±7.76 ms	1135.26 ms	±0.19 ms
			Cuda	5.54 ms	±0.12 ms	—	—
modulo	1000x1000	int32	Host	—	—	16.17 ms	±0.01 ms
			Cuda	1.16 ms	±0.01 ms	—	—
		int64	Host	—	—	27.20 ms	±0.01 ms
			Cuda	0.84 ms	±0.00 ms	—	—
		single	Host	15.77 ms	±0.29 ms	38.89 ms	±0.06 ms
			Cuda	0.78 ms	±0.01 ms	—	—
		double	Host	16.59 ms	±0.50 ms	38.64 ms	±0.03 ms
			Cuda	0.80 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	64.71 ms	±0.03 ms
			Cuda	1.04 ms	±0.01 ms	—	—
		int64	Host	—	—	108.82 ms	±0.03 ms
			Cuda	1.59 ms	±0.01 ms	—	—
		single	Host	59.83 ms	±1.17 ms	155.63 ms	±0.09 ms
			Cuda	1.29 ms	±0.00 ms	—	—
		double	Host	60.49 ms	±0.56 ms	154.72 ms	±0.05 ms
			Cuda	1.39 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	287.69 ms	±0.11 ms
			Cuda	2.52 ms	±0.03 ms	—	—
		int64	Host	—	—	492.06 ms	±0.25 ms
			Cuda	4.78 ms	±0.11 ms	—	—
		single	Host	232.45 ms	±2.63 ms	651.76 ms	±0.22 ms
			Cuda	3.42 ms	±0.03 ms	—	—
		double	Host	235.98 ms	±1.44 ms	678.19 ms	±0.15 ms
			Cuda	3.96 ms	±0.09 ms	—	—
dot	1000x1000	int32	Host	—	—	473.22 ms	±0.07 ms
		int64	Host	—	—	1353.18 ms	±0.17 ms
		single	Host	28.20 ms	±1.87 ms	10.13 ms	±0.01 ms
			Cuda	2.39 ms	±0.07 ms	—	—
		double	Host	42.59 ms	±0.32 ms	20.17 ms	±0.44 ms
			Cuda	3.24 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	4041.68 ms	±0.42 ms
		int64	Host	—	—	11000.01 ms	±22.24 ms
		single	Host	129.17 ms	±6.90 ms	75.56 ms	±1.29 ms
			Cuda	8.83 ms	±0.05 ms	—	—
		double	Host	206.18 ms	±8.52 ms	147.94 ms	±1.54 ms
			Cuda	16.70 ms	±0.09 ms	—	—
	4000x4000	int32	Host	—	—	33160.89 ms	±2.34 ms
		int64	Host	—	—	86242.90 ms	±2.01 ms
		single	Host	734.55 ms	±4.42 ms	590.84 ms	±2.47 ms
			Cuda	56.39 ms	±0.62 ms	—	—
		double	Host	—	—	1177.82 ms	±2.07 ms
			Cuda	119.68 ms	±2.68 ms	—	—
sgn	1000x1000	bool	Cuda	0.66 ms	±0.00 ms	—	—
		int32	Host	10.34 ms	±0.02 ms	0.56 ms	±0.01 ms
			Cuda	0.65 ms	±0.02 ms	—	—
		int64	Host	11.83 ms	±0.51 ms	6.89 ms	±0.11 ms
			Cuda	0.65 ms	±0.01 ms	—	—
		single	Host	11.34 ms	±0.34 ms	7.98 ms	±0.12 ms
			Cuda	0.65 ms	±0.00 ms	—	—
		double	Host	11.97 ms	±0.18 ms	8.25 ms	±0.11 ms
			Cuda	0.65 ms	±0.00 ms	—	—
	2000x2000	bool	Cuda	0.87 ms	±0.01 ms	—	—
		int32	Host	39.56 ms	±0.72 ms	2.91 ms	±0.08 ms
			Cuda	0.88 ms	±0.00 ms	—	—
		int64	Host	42.52 ms	±0.77 ms	27.67 ms	±0.28 ms
			Cuda	0.92 ms	±0.01 ms	—	—
		single	Host	40.94 ms	±0.65 ms	32.08 ms	±0.31 ms
			Cuda	0.89 ms	±0.00 ms	—	—
		double	Host	42.48 ms	±0.85 ms	33.00 ms	±0.04 ms
			Cuda	0.93 ms	±0.00 ms	—	—
	4000x4000	bool	Cuda	1.78 ms	±0.02 ms	—	—
		int32	Host	151.74 ms	±0.42 ms	37.13 ms	±0.43 ms
			Cuda	1.85 ms	±0.02 ms	—	—
		int64	Host	160.62 ms	±0.19 ms	165.27 ms	±0.71 ms
			Cuda	12.41 ms	±0.05 ms	—	—
		single	Host	158.98 ms	±1.68 ms	156.19 ms	±0.81 ms
			Cuda	1.87 ms	±0.02 ms	—	—
		double	Host	168.05 ms	±1.64 ms	187.36 ms	±0.81 ms
			Cuda	12.53 ms	±0.05 ms	—	—
log	1000x1000	int32	Host	—	—	32.64 ms	±0.47 ms
		int64	Host	—	—	33.00 ms	±0.49 ms
		single	Host	16.94 ms	±0.57 ms	23.53 ms	±4.29 ms
			Cuda	0.66 ms	±0.00 ms	—	—
		double	Host	17.26 ms	±0.27 ms	31.81 ms	±0.48 ms
			Cuda	0.70 ms	±0.01 ms	—	—
	2000x2000	int32	Host	—	—	130.48 ms	±0.91 ms
		int64	Host	—	—	132.12 ms	±0.91 ms
		single	Host	63.88 ms	±0.32 ms	82.64 ms	±0.92 ms
			Cuda	0.93 ms	±0.00 ms	—	—
		double	Host	63.36 ms	±0.61 ms	127.23 ms	±0.98 ms
			Cuda	1.15 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	580.22 ms	±1.26 ms
		int64	Host	—	—	655.68 ms	±47.15 ms
		single	Host	247.31 ms	±1.26 ms	361.09 ms	±1.52 ms
			Cuda	2.03 ms	±0.02 ms	—	—
		double	Host	248.18 ms	±2.65 ms	566.88 ms	±1.28 ms
			Cuda	3.07 ms	±0.07 ms	—	—
log10	1000x1000	int32	Host	—	—	34.15 ms	±0.53 ms
		int64	Host	—	—	34.59 ms	±0.55 ms
		single	Host	20.34 ms	±0.23 ms	25.56 ms	±0.40 ms
			Cuda	0.66 ms	±0.00 ms	—	—
		double	Host	18.70 ms	±0.38 ms	34.88 ms	±0.57 ms
			Cuda	0.73 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	136.53 ms	±1.12 ms
		int64	Host	—	—	138.04 ms	±1.03 ms
		single	Host	68.15 ms	±0.36 ms	102.44 ms	±1.11 ms
			Cuda	0.96 ms	±0.00 ms	—	—
		double	Host	68.95 ms	±0.58 ms	139.56 ms	±1.09 ms
			Cuda	1.18 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	604.82 ms	±1.64 ms
		int64	Host	—	—	610.75 ms	±1.76 ms
		single	Host	265.11 ms	±1.87 ms	440.81 ms	±1.90 ms
			Cuda	2.10 ms	±0.02 ms	—	—
		double	Host	267.87 ms	±2.90 ms	616.02 ms	±0.36 ms
			Cuda	3.12 ms	±0.06 ms	—	—
exp	1000x1000	int32	Host	—	—	27.33 ms	±0.00 ms
		int64	Host	—	—	29.07 ms	±3.99 ms
		single	Host	16.77 ms	±0.09 ms	16.57 ms	±0.02 ms
			Cuda	0.69 ms	±0.00 ms	—	—
		double	Host	17.62 ms	±0.24 ms	26.94 ms	±0.01 ms
			Cuda	0.68 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	109.37 ms	±0.01 ms
		int64	Host	—	—	111.06 ms	±0.01 ms
		single	Host	65.02 ms	±0.38 ms	66.35 ms	±0.01 ms
			Cuda	0.92 ms	±0.00 ms	—	—
		double	Host	66.10 ms	±0.61 ms	107.68 ms	±0.01 ms
			Cuda	1.04 ms	±0.00 ms	—	—
	4000x4000	int32	Host	—	—	575.90 ms	±103.34 ms
		int64	Host	—	—	576.50 ms	±110.41 ms
		single	Host	250.42 ms	±1.48 ms	293.30 ms	±0.07 ms
			Cuda	2.02 ms	±0.02 ms	—	—
		double	Host	255.47 ms	±3.57 ms	487.84 ms	±0.37 ms
			Cuda	14.86 ms	±0.06 ms	—	—
sin	1000x1000	int32	Host	—	—	54.91 ms	±0.09 ms
		int64	Host	—	—	55.34 ms	±0.09 ms
		single	Host	21.54 ms	±0.07 ms	21.02 ms	±0.03 ms
			Cuda	0.68 ms	±0.01 ms	—	—
		double	Host	22.07 ms	±0.67 ms	60.76 ms	±0.02 ms
			Cuda	0.70 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	219.60 ms	±0.09 ms
		int64	Host	—	—	221.39 ms	±0.10 ms
		single	Host	84.74 ms	±0.94 ms	84.46 ms	±0.02 ms
			Cuda	0.96 ms	±0.00 ms	—	—
		double	Host	85.50 ms	±0.96 ms	242.90 ms	±0.05 ms
			Cuda	1.08 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	939.10 ms	±0.35 ms
		int64	Host	—	—	944.28 ms	±0.40 ms
		single	Host	326.39 ms	±3.30 ms	369.04 ms	±0.05 ms
			Cuda	2.14 ms	±0.02 ms	—	—
		double	Host	328.68 ms	±2.74 ms	1053.12 ms	±37.89 ms
			Cuda	2.76 ms	±0.04 ms	—	—
cos	1000x1000	int32	Host	—	—	86.90 ms	±0.04 ms
		int64	Host	—	—	87.34 ms	±0.05 ms
		single	Host	21.39 ms	±0.90 ms	41.42 ms	±0.02 ms
			Cuda	0.68 ms	±0.01 ms	—	—
		double	Host	21.39 ms	±0.29 ms	82.03 ms	±0.10 ms
			Cuda	0.72 ms	±0.02 ms	—	—
	2000x2000	int32	Host	—	—	348.03 ms	±0.06 ms
		int64	Host	—	—	349.48 ms	±0.11 ms
		single	Host	82.31 ms	±0.89 ms	166.05 ms	±0.04 ms
			Cuda	0.97 ms	±0.00 ms	—	—
		double	Host	81.96 ms	±0.89 ms	328.65 ms	±0.40 ms
			Cuda	1.09 ms	±0.00 ms	—	—
	4000x4000	int32	Host	—	—	1461.91 ms	±29.70 ms
		int64	Host	—	—	1455.08 ms	±0.25 ms
		single	Host	314.67 ms	±3.43 ms	692.74 ms	±0.12 ms
			Cuda	2.18 ms	±0.03 ms	—	—
		double	Host	317.27 ms	±2.71 ms	1373.27 ms	±1.03 ms
			Cuda	2.75 ms	±0.06 ms	—	—
tan	1000x1000	single	Host	25.95 ms	±0.16 ms	—	—
			Cuda	0.67 ms	±0.01 ms	—	—
		double	Host	26.20 ms	±0.52 ms	—	—
			Cuda	0.73 ms	±0.00 ms	—	—
	2000x2000	single	Host	99.90 ms	±0.21 ms	—	—
			Cuda	1.03 ms	±0.00 ms	—	—
		double	Host	101.22 ms	±0.45 ms	—	—
			Cuda	1.13 ms	±0.01 ms	—	—
	4000x4000	single	Host	392.67 ms	±1.59 ms	—	—
			Cuda	2.44 ms	±0.02 ms	—	—
		double	Host	398.98 ms	±1.08 ms	—	—
			Cuda	3.02 ms	±0.07 ms	—	—
asin	1000x1000	int32	Host	—	—	11.11 ms	±0.01 ms
		int64	Host	—	—	11.50 ms	±0.01 ms
		single	Host	12.61 ms	±0.51 ms	11.96 ms	±0.00 ms
			Cuda	0.67 ms	±0.00 ms	—	—
		double	Host	13.42 ms	±0.50 ms	10.38 ms	±0.01 ms
			Cuda	0.72 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	44.41 ms	±0.01 ms
		int64	Host	—	—	47.40 ms	±4.59 ms
		single	Host	47.05 ms	±0.87 ms	47.89 ms	±0.02 ms
			Cuda	0.99 ms	±0.02 ms	—	—
		double	Host	48.76 ms	±0.38 ms	41.60 ms	±0.02 ms
			Cuda	1.14 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	239.15 ms	±9.48 ms
		int64	Host	—	—	237.64 ms	±0.04 ms
		single	Host	183.29 ms	±0.92 ms	219.26 ms	±0.04 ms
			Cuda	2.30 ms	±0.02 ms	—	—
		double	Host	183.01 ms	±0.26 ms	220.14 ms	±0.03 ms
			Cuda	3.03 ms	±0.07 ms	—	—
acos	1000x1000	int32	Host	—	—	12.10 ms	±0.01 ms
		int64	Host	—	—	12.48 ms	±0.01 ms
		single	Host	12.96 ms	±0.25 ms	11.54 ms	±0.01 ms
			Cuda	0.70 ms	±0.00 ms	—	—
		double	Host	12.95 ms	±0.73 ms	11.37 ms	±0.01 ms
			Cuda	0.71 ms	±0.01 ms	—	—
	2000x2000	int32	Host	—	—	48.36 ms	±0.03 ms
		int64	Host	—	—	49.96 ms	±0.01 ms
		single	Host	46.31 ms	±0.11 ms	46.28 ms	±0.02 ms
			Cuda	1.00 ms	±0.00 ms	—	—
		double	Host	48.13 ms	±0.48 ms	45.48 ms	±0.02 ms
			Cuda	1.12 ms	±0.03 ms	—	—
	4000x4000	int32	Host	—	—	304.26 ms	±32.94 ms
		int64	Host	—	—	253.44 ms	±0.06 ms
		single	Host	181.65 ms	±0.41 ms	212.83 ms	±0.03 ms
			Cuda	2.28 ms	±0.02 ms	—	—
		double	Host	185.92 ms	±4.50 ms	235.34 ms	±0.03 ms
			Cuda	2.91 ms	±0.07 ms	—	—
atan	1000x1000	int32	Host	—	—	38.09 ms	±0.02 ms
		int64	Host	—	—	38.50 ms	±0.03 ms
		single	Host	18.85 ms	±0.24 ms	18.98 ms	±0.02 ms
			Cuda	0.69 ms	±0.00 ms	—	—
		double	Host	19.43 ms	±0.71 ms	39.43 ms	±0.02 ms
			Cuda	0.71 ms	±0.02 ms	—	—
	2000x2000	int32	Host	—	—	152.32 ms	±0.06 ms
		int64	Host	—	—	169.61 ms	±24.22 ms
		single	Host	72.07 ms	±0.37 ms	76.03 ms	±0.02 ms
			Cuda	0.97 ms	±0.02 ms	—	—
		double	Host	72.49 ms	±0.81 ms	157.75 ms	±0.03 ms
			Cuda	1.10 ms	±0.00 ms	—	—
	4000x4000	int32	Host	—	—	668.09 ms	±0.16 ms
		int64	Host	—	—	673.69 ms	±0.18 ms
		single	Host	280.84 ms	±1.60 ms	334.32 ms	±0.11 ms
			Cuda	2.23 ms	±0.02 ms	—	—
		double	Host	280.42 ms	±2.28 ms	757.55 ms	±103.61 ms
			Cuda	2.84 ms	±0.06 ms	—	—
sinh	1000x1000	int32	Host	—	—	63.06 ms	±7.39 ms
		int64	Host	—	—	49.02 ms	±0.02 ms
		single	Host	20.80 ms	±0.39 ms	38.09 ms	±0.01 ms
			Cuda	0.70 ms	±0.01 ms	—	—
		double	Host	20.75 ms	±0.23 ms	48.73 ms	±0.01 ms
			Cuda	0.79 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	241.74 ms	±36.77 ms
		int64	Host	—	—	196.19 ms	±0.03 ms
		single	Host	79.67 ms	±0.69 ms	152.72 ms	±0.03 ms
			Cuda	0.99 ms	±0.01 ms	—	—
		double	Host	79.47 ms	±0.69 ms	194.96 ms	±0.03 ms
			Cuda	1.29 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	905.57 ms	±122.32 ms
		int64	Host	—	—	844.33 ms	±0.24 ms
		single	Host	312.08 ms	±1.74 ms	639.87 ms	±0.09 ms
			Cuda	2.27 ms	±0.02 ms	—	—
		double	Host	318.69 ms	±5.96 ms	839.24 ms	±0.20 ms
			Cuda	3.63 ms	±0.08 ms	—	—
cosh	1000x1000	int32	Host	—	—	35.42 ms	±0.03 ms
		int64	Host	—	—	35.87 ms	±0.01 ms
		single	Host	18.30 ms	±0.08 ms	22.92 ms	±0.03 ms
			Cuda	0.67 ms	±0.04 ms	—	—
		double	Host	18.86 ms	±0.55 ms	35.28 ms	±0.01 ms
			Cuda	0.72 ms	±0.01 ms	—	—
	2000x2000	int32	Host	—	—	141.86 ms	±0.14 ms
		int64	Host	—	—	159.11 ms	±26.40 ms
		single	Host	70.65 ms	±0.32 ms	91.70 ms	±0.09 ms
			Cuda	0.95 ms	±0.01 ms	—	—
		double	Host	70.22 ms	±0.55 ms	141.36 ms	±0.04 ms
			Cuda	1.07 ms	±0.01 ms	—	—
	4000x4000	int32	Host	—	—	626.83 ms	±0.35 ms
		int64	Host	—	—	632.70 ms	±0.39 ms
		single	Host	279.45 ms	±4.32 ms	431.62 ms	±72.79 ms
			Cuda	2.12 ms	±0.02 ms	—	—
		double	Host	276.39 ms	±2.95 ms	625.80 ms	±0.13 ms
			Cuda	2.73 ms	±0.06 ms	—	—
tanh	1000x1000	int32	Host	—	—	28.31 ms	±0.01 ms
		int64	Host	—	—	28.69 ms	±0.02 ms
		single	Host	16.35 ms	±0.17 ms	32.42 ms	±5.54 ms
			Cuda	0.70 ms	±0.00 ms	—	—
		double	Host	16.86 ms	±0.47 ms	28.56 ms	±0.02 ms
			Cuda	0.69 ms	±0.00 ms	—	—
	2000x2000	int32	Host	—	—	115.52 ms	±6.88 ms
		int64	Host	—	—	114.78 ms	±0.03 ms
		single	Host	62.02 ms	±0.51 ms	111.03 ms	±0.09 ms
			Cuda	0.96 ms	±0.01 ms	—	—
		double	Host	63.09 ms	±1.71 ms	114.38 ms	±0.03 ms
			Cuda	1.13 ms	±0.00 ms	—	—
	4000x4000	int32	Host	—	—	511.16 ms	±0.12 ms
		int64	Host	—	—	516.73 ms	±0.17 ms
		single	Host	245.96 ms	±3.37 ms	552.64 ms	±77.32 ms
			Cuda	2.24 ms	±0.02 ms	—	—
		double	Host	240.44 ms	±1.69 ms	621.19 ms	±77.36 ms
			Cuda	2.92 ms	±0.07 ms	—	—
sqrt	1000x1000	int32	Host	8.66 ms	±0.17 ms	3.73 ms	±0.01 ms
		int64	Host	11.41 ms	±0.03 ms	4.09 ms	±0.01 ms
		single	Host	2.54 ms	±0.00 ms	0.92 ms	±0.00 ms
			Cuda	0.68 ms	±0.01 ms	—	—
		double	Host	6.33 ms	±0.00 ms	3.21 ms	±0.01 ms
			Cuda	0.70 ms	±0.00 ms	—	—
	2000x2000	int32	Host	33.44 ms	±0.01 ms	14.92 ms	±0.01 ms
		int64	Host	45.02 ms	±0.01 ms	16.47 ms	±0.01 ms
		single	Host	10.14 ms	±0.02 ms	3.72 ms	±0.02 ms
			Cuda	0.96 ms	±0.00 ms	—	—
		double	Host	25.17 ms	±0.05 ms	12.82 ms	±0.01 ms
			Cuda	1.02 ms	±0.01 ms	—	—
	4000x4000	int32	Host	135.70 ms	±4.13 ms	115.76 ms	±1.60 ms
		int64	Host	177.08 ms	±0.06 ms	119.99 ms	±0.04 ms
		single	Host	40.38 ms	±0.14 ms	42.34 ms	±0.02 ms
			Cuda	2.18 ms	±0.02 ms	—	—
		double	Host	100.98 ms	±0.02 ms	105.91 ms	±0.03 ms
			Cuda	14.71 ms	±0.06 ms	—	—
ceiling	1000x1000	single	Host	10.42 ms	±0.40 ms	—	—
			Cuda	0.64 ms	±0.00 ms	—	—
		double	Host	10.89 ms	±0.56 ms	—	—
			Cuda	0.67 ms	±0.00 ms	—	—
	2000x2000	single	Host	37.74 ms	±0.57 ms	—	—
			Cuda	0.90 ms	±0.00 ms	—	—
		double	Host	39.50 ms	±0.46 ms	—	—
			Cuda	0.91 ms	±0.00 ms	—	—
	4000x4000	single	Host	147.24 ms	±0.53 ms	—	—
			Cuda	1.83 ms	±0.02 ms	—	—
		double	Host	151.38 ms	±1.55 ms	—	—
			Cuda	14.42 ms	±0.06 ms	—	—
floor	1000x1000	single	Host	11.09 ms	±0.55 ms	—	—
			Cuda	0.68 ms	±0.02 ms	—	—
		double	Host	11.47 ms	±0.44 ms	—	—
			Cuda	0.67 ms	±0.01 ms	—	—
	2000x2000	single	Host	40.82 ms	±0.42 ms	—	—
			Cuda	0.88 ms	±0.00 ms	—	—
		double	Host	40.30 ms	±0.50 ms	—	—
			Cuda	0.91 ms	±0.01 ms	—	—
	4000x4000	single	Host	157.55 ms	±1.89 ms	—	—
			Cuda	1.80 ms	±0.02 ms	—	—
		double	Host	157.45 ms	±0.48 ms	—	—
			Cuda	14.41 ms	±0.06 ms	—	—
round	1000x1000	single	Host	11.14 ms	±0.45 ms	—	—
			Cuda	0.65 ms	±0.01 ms	—	—
		double	Host	10.93 ms	±0.46 ms	—	—
			Cuda	0.68 ms	±0.00 ms	—	—
	2000x2000	single	Host	37.14 ms	±0.45 ms	—	—
			Cuda	0.90 ms	±0.00 ms	—	—
		double	Host	38.48 ms	±0.32 ms	—	—
			Cuda	0.96 ms	±0.00 ms	—	—
	4000x4000	single	Host	146.54 ms	±1.87 ms	—	—
			Cuda	1.93 ms	±0.02 ms	—	—
		double	Host	149.05 ms	±0.66 ms	—	—
			Cuda	13.32 ms	±0.05 ms	—	—
truncate	1000x1000	single	Host	11.42 ms	±0.51 ms	—	—
			Cuda	0.66 ms	±0.00 ms	—	—
		double	Host	12.00 ms	±0.39 ms	—	—
			Cuda	0.67 ms	±0.02 ms	—	—
	2000x2000	single	Host	42.63 ms	±0.64 ms	—	—
			Cuda	0.88 ms	±0.01 ms	—	—
		double	Host	42.75 ms	±0.46 ms	—	—
			Cuda	0.91 ms	±0.00 ms	—	—
	4000x4000	single	Host	229.59 ms	±1.80 ms	—	—
			Cuda	1.83 ms	±0.02 ms	—	—
		double	Host	166.28 ms	±0.79 ms	—	—
			Cuda	12.20 ms	±0.05 ms	—	—
maxElemwise	1000x1000	bool	Host	10.89 ms	±0.70 ms	—	—
			Cuda	1.13 ms	±0.01 ms	—	—
		int32	Host	2.44 ms	±0.01 ms	—	—
			Cuda	1.12 ms	±0.01 ms	—	—
		int64	Host	4.95 ms	±0.02 ms	—	—
			Cuda	1.36 ms	±0.41 ms	—	—
		single	Host	2.46 ms	±0.00 ms	—	—
			Cuda	1.13 ms	±0.01 ms	—	—
		double	Host	4.90 ms	±0.05 ms	—	—
			Cuda	0.72 ms	±0.01 ms	—	—
	2000x2000	bool	Host	35.96 ms	±0.93 ms	—	—
			Cuda	0.98 ms	±0.00 ms	—	—
		int32	Host	9.79 ms	±0.12 ms	—	—
			Cuda	0.99 ms	±0.00 ms	—	—
		int64	Host	19.07 ms	±0.01 ms	—	—
			Cuda	1.06 ms	±0.01 ms	—	—
		single	Host	9.90 ms	±0.10 ms	—	—
			Cuda	0.98 ms	±0.01 ms	—	—
		double	Host	18.89 ms	±0.00 ms	—	—
			Cuda	1.06 ms	±0.00 ms	—	—
	4000x4000	bool	Host	140.14 ms	±1.67 ms	—	—
			Cuda	2.18 ms	±0.02 ms	—	—
		int32	Host	38.28 ms	±0.16 ms	—	—
			Cuda	2.25 ms	±0.02 ms	—	—
		int64	Host	76.69 ms	±0.05 ms	—	—
			Cuda	2.69 ms	±0.06 ms	—	—
		single	Host	38.91 ms	±0.01 ms	—	—
			Cuda	2.28 ms	±0.06 ms	—	—
		double	Host	76.31 ms	±0.05 ms	—	—
			Cuda	2.67 ms	±0.06 ms	—	—
minElemwise	1000x1000	bool	Host	10.52 ms	±1.35 ms	—	—
			Cuda	1.10 ms	±0.00 ms	—	—
		int32	Host	2.40 ms	±0.01 ms	—	—
			Cuda	1.10 ms	±0.01 ms	—	—
		int64	Host	4.93 ms	±0.06 ms	—	—
			Cuda	0.69 ms	±0.00 ms	—	—
		single	Host	2.40 ms	±0.03 ms	—	—
			Cuda	1.12 ms	±0.02 ms	—	—
		double	Host	4.91 ms	±0.01 ms	—	—
			Cuda	0.70 ms	±0.02 ms	—	—
	2000x2000	bool	Host	35.44 ms	±0.83 ms	—	—
			Cuda	0.96 ms	±0.00 ms	—	—
		int32	Host	9.72 ms	±0.04 ms	—	—
			Cuda	0.99 ms	±0.00 ms	—	—
		int64	Host	19.20 ms	±0.11 ms	—	—
			Cuda	1.06 ms	±0.01 ms	—	—
		single	Host	9.76 ms	±0.00 ms	—	—
			Cuda	0.99 ms	±0.00 ms	—	—
		double	Host	19.00 ms	±0.00 ms	—	—
			Cuda	1.05 ms	±0.00 ms	—	—
	4000x4000	bool	Host	141.37 ms	±1.41 ms	—	—
			Cuda	2.21 ms	±0.06 ms	—	—
		int32	Host	39.09 ms	±0.01 ms	—	—
			Cuda	2.26 ms	±0.03 ms	—	—
		int64	Host	76.98 ms	±0.01 ms	—	—
			Cuda	2.70 ms	±0.06 ms	—	—
		single	Host	38.28 ms	±0.01 ms	—	—
			Cuda	2.22 ms	±0.02 ms	—	—
		double	Host	75.90 ms	±0.02 ms	—	—
			Cuda	2.67 ms	±0.06 ms	—	—
sumAxis	1000x1000	int32	Host	6.70 ms	±0.16 ms	1.40 ms	±0.01 ms
			Cuda	3.23 ms	±0.00 ms	—	—
		int64	Host	6.71 ms	±0.24 ms	0.94 ms	±0.01 ms
			Cuda	3.37 ms	±0.00 ms	—	—
		single	Host	6.52 ms	±0.28 ms	0.62 ms	±0.00 ms
			Cuda	3.25 ms	±0.00 ms	—	—
		double	Host	6.85 ms	±0.61 ms	0.63 ms	±0.01 ms
			Cuda	3.39 ms	±0.01 ms	—	—
	2000x2000	int32	Host	23.86 ms	±0.33 ms	5.67 ms	±0.01 ms
			Cuda	6.06 ms	±0.00 ms	—	—
		int64	Host	24.38 ms	±0.39 ms	3.75 ms	±0.01 ms
			Cuda	6.31 ms	±0.02 ms	—	—
		single	Host	24.74 ms	±0.34 ms	2.43 ms	±0.01 ms
			Cuda	6.06 ms	±0.01 ms	—	—
		double	Host	24.56 ms	±0.39 ms	2.69 ms	±0.01 ms
			Cuda	6.33 ms	±0.00 ms	—	—
	4000x4000	int32	Host	90.28 ms	±1.04 ms	22.98 ms	±0.01 ms
			Cuda	12.06 ms	±0.00 ms	—	—
		int64	Host	93.62 ms	±1.49 ms	14.94 ms	±0.08 ms
			Cuda	12.63 ms	±0.00 ms	—	—
		single	Host	91.95 ms	±1.28 ms	9.58 ms	±0.01 ms
			Cuda	12.08 ms	±0.02 ms	—	—
		double	Host	93.12 ms	±1.27 ms	10.71 ms	±0.02 ms
			Cuda	12.67 ms	±0.02 ms	—	—
productAxis	1000x1000	int32	Host	6.34 ms	±0.30 ms	—	—
			Cuda	3.22 ms	±0.00 ms	—	—
		int64	Host	6.98 ms	±0.13 ms	—	—
			Cuda	3.35 ms	±0.00 ms	—	—
		single	Host	6.99 ms	±0.36 ms	—	—
			Cuda	3.23 ms	±0.00 ms	—	—
		double	Host	6.94 ms	±0.72 ms	—	—
			Cuda	3.34 ms	±0.00 ms	—	—
	2000x2000	int32	Host	24.05 ms	±0.35 ms	—	—
			Cuda	6.06 ms	±0.00 ms	—	—
		int64	Host	24.03 ms	±0.18 ms	—	—
			Cuda	6.35 ms	±0.01 ms	—	—
		single	Host	24.33 ms	±0.42 ms	—	—
			Cuda	6.04 ms	±0.02 ms	—	—
		double	Host	24.31 ms	±0.24 ms	—	—
			Cuda	6.35 ms	±0.00 ms	—	—
	4000x4000	int32	Host	93.36 ms	±1.45 ms	—	—
			Cuda	12.06 ms	±0.02 ms	—	—
		int64	Host	90.85 ms	±0.35 ms	—	—
			Cuda	12.65 ms	±0.00 ms	—	—
		single	Host	93.97 ms	±0.43 ms	—	—
			Cuda	12.05 ms	±0.00 ms	—	—
		double	Host	93.06 ms	±0.35 ms	—	—
			Cuda	12.64 ms	±0.03 ms	—	—
maxAxis	1000x1000	int32	Host	7.43 ms	±0.37 ms	—	—
			Cuda	3.22 ms	±0.00 ms	—	—
		int64	Host	7.67 ms	±0.21 ms	—	—
			Cuda	3.35 ms	±0.01 ms	—	—
		single	Host	7.72 ms	±0.26 ms	—	—
			Cuda	3.22 ms	±0.00 ms	—	—
		double	Host	7.60 ms	±0.13 ms	—	—
			Cuda	3.34 ms	±0.00 ms	—	—
	2000x2000	int32	Host	28.20 ms	±0.81 ms	—	—
			Cuda	6.06 ms	±0.00 ms	—	—
		int64	Host	28.06 ms	±0.39 ms	—	—
			Cuda	6.34 ms	±0.00 ms	—	—
		single	Host	27.45 ms	±0.54 ms	—	—
			Cuda	6.06 ms	±0.00 ms	—	—
		double	Host	27.99 ms	±0.67 ms	—	—
			Cuda	6.34 ms	±0.01 ms	—	—
	4000x4000	int32	Host	109.06 ms	±1.51 ms	—	—
			Cuda	12.05 ms	±0.00 ms	—	—
		int64	Host	108.19 ms	±1.91 ms	—	—
			Cuda	12.60 ms	±0.00 ms	—	—
		single	Host	106.44 ms	±2.98 ms	—	—
			Cuda	12.05 ms	±0.00 ms	—	—
		double	Host	107.22 ms	±2.46 ms	—	—
			Cuda	12.62 ms	±0.02 ms	—	—
minAxis	1000x1000	int32	Host	7.64 ms	±0.69 ms	—	—
			Cuda	3.21 ms	±0.03 ms	—	—
		int64	Host	7.86 ms	±0.47 ms	—	—
			Cuda	3.33 ms	±0.00 ms	—	—
		single	Host	7.82 ms	±0.58 ms	—	—
			Cuda	3.22 ms	±0.00 ms	—	—
		double	Host	7.49 ms	±0.56 ms	—	—
			Cuda	3.34 ms	±0.01 ms	—	—
	2000x2000	int32	Host	28.79 ms	±0.41 ms	—	—
			Cuda	6.06 ms	±0.00 ms	—	—
		int64	Host	27.82 ms	±0.16 ms	—	—
			Cuda	6.33 ms	±0.01 ms	—	—
		single	Host	27.35 ms	±0.50 ms	—	—
			Cuda	6.05 ms	±0.02 ms	—	—
		double	Host	27.83 ms	±0.57 ms	—	—
			Cuda	6.34 ms	±0.00 ms	—	—
	4000x4000	int32	Host	107.32 ms	±1.38 ms	—	—
			Cuda	12.07 ms	±0.00 ms	—	—
		int64	Host	107.43 ms	±1.76 ms	—	—
			Cuda	12.65 ms	±0.01 ms	—	—
		single	Host	105.30 ms	±1.24 ms	—	—
			Cuda	12.05 ms	±0.00 ms	—	—
		double	Host	108.85 ms	±1.19 ms	—	—
			Cuda	12.63 ms	±0.01 ms	—	—
argMaxAxis	1000x1000	int32	Host	10.14 ms	±0.65 ms	—	—
			Cuda	3.22 ms	±0.00 ms	—	—
		int64	Host	9.66 ms	±0.40 ms	—	—
			Cuda	3.36 ms	±0.00 ms	—	—
		single	Host	9.95 ms	±0.59 ms	—	—
			Cuda	3.20 ms	±0.02 ms	—	—
		double	Host	9.52 ms	±0.20 ms	—	—
			Cuda	3.33 ms	±0.00 ms	—	—
	2000x2000	int32	Host	35.43 ms	±0.57 ms	—	—
			Cuda	6.06 ms	±0.02 ms	—	—
		int64	Host	35.24 ms	±0.47 ms	—	—
			Cuda	6.35 ms	±0.00 ms	—	—
		single	Host	37.19 ms	±0.62 ms	—	—
			Cuda	6.07 ms	±0.01 ms	—	—
		double	Host	35.63 ms	±0.64 ms	—	—
			Cuda	6.37 ms	±0.01 ms	—	—
	4000x4000	int32	Host	140.16 ms	±2.23 ms	—	—
			Cuda	12.08 ms	±0.00 ms	—	—
		int64	Host	138.56 ms	±1.64 ms	—	—
			Cuda	12.69 ms	±0.00 ms	—	—
		single	Host	140.24 ms	±2.14 ms	—	—
			Cuda	12.08 ms	±0.00 ms	—	—
		double	Host	140.01 ms	±1.29 ms	—	—
			Cuda	12.64 ms	±0.03 ms	—	—
argMinAxis	1000x1000	int32	Host	9.85 ms	±0.24 ms	—	—
			Cuda	3.22 ms	±0.00 ms	—	—
		int64	Host	9.57 ms	±0.13 ms	—	—
			Cuda	3.36 ms	±0.00 ms	—	—
		single	Host	9.52 ms	±0.09 ms	—	—
			Cuda	3.23 ms	±0.00 ms	—	—
		double	Host	9.85 ms	±0.03 ms	—	—
			Cuda	3.35 ms	±0.00 ms	—	—
	2000x2000	int32	Host	35.88 ms	±0.94 ms	—	—
			Cuda	6.08 ms	±0.00 ms	—	—
		int64	Host	35.41 ms	±0.65 ms	—	—
			Cuda	6.36 ms	±0.00 ms	—	—
		single	Host	36.36 ms	±0.77 ms	—	—
			Cuda	6.07 ms	±0.00 ms	—	—
		double	Host	35.49 ms	±0.87 ms	—	—
			Cuda	6.36 ms	±0.00 ms	—	—
	4000x4000	int32	Host	138.45 ms	±1.80 ms	—	—
			Cuda	12.07 ms	±0.00 ms	—	—
		int64	Host	137.86 ms	±1.56 ms	—	—
			Cuda	12.66 ms	±0.00 ms	—	—
		single	Host	141.17 ms	±1.25 ms	—	—
			Cuda	12.07 ms	±0.01 ms	—	—
		double	Host	137.50 ms	±1.48 ms	—	—
			Cuda	12.62 ms	±0.00 ms	—	—
findAxis	1000x1000	int32	Host	2.52 ms	±0.05 ms	—	—
			Cuda	0.85 ms	±0.01 ms	—	—
		int64	Host	2.73 ms	±0.04 ms	—	—
			Cuda	0.86 ms	±0.00 ms	—	—
		single	Host	3.47 ms	±0.19 ms	—	—
			Cuda	3.22 ms	±0.00 ms	—	—
		double	Host	3.52 ms	±0.16 ms	—	—
			Cuda	3.36 ms	±0.00 ms	—	—
	2000x2000	int32	Host	8.57 ms	±0.11 ms	—	—
			Cuda	1.03 ms	±0.00 ms	—	—
		int64	Host	9.18 ms	±0.35 ms	—	—
			Cuda	1.12 ms	±0.00 ms	—	—
		single	Host	12.28 ms	±0.13 ms	—	—
			Cuda	6.07 ms	±0.00 ms	—	—
		double	Host	12.38 ms	±0.25 ms	—	—
			Cuda	6.37 ms	±0.00 ms	—	—
	4000x4000	int32	Host	31.51 ms	±0.50 ms	—	—
			Cuda	1.39 ms	±0.00 ms	—	—
		int64	Host	33.43 ms	±0.54 ms	—	—
			Cuda	1.52 ms	±0.00 ms	—	—
		single	Host	46.50 ms	±0.10 ms	—	—
			Cuda	12.10 ms	±0.00 ms	—	—
		double	Host	47.01 ms	±0.32 ms	—	—
			Cuda	12.65 ms	±0.02 ms	—	—
maskedGet	1000x1000	bool	Host	113.69 ms	±0.07 ms	—	—
		int32	Host	118.98 ms	±0.18 ms	—	—
		int64	Host	115.16 ms	±0.07 ms	—	—
		single	Host	115.98 ms	±0.09 ms	—	—
		double	Host	116.40 ms	±0.03 ms	—	—
	2000x2000	bool	Host	449.91 ms	±0.14 ms	—	—
		int32	Host	454.92 ms	±0.39 ms	—	—
		int64	Host	458.18 ms	±0.21 ms	—	—
		single	Host	507.79 ms	±18.17 ms	—	—
		double	Host	464.75 ms	±0.22 ms	—	—
maskedSet	1000x1000	bool	Host	113.10 ms	±0.08 ms	—	—
		int32	Host	116.41 ms	±0.09 ms	—	—
		int64	Host	111.72 ms	±0.04 ms	—	—
		single	Host	112.88 ms	±0.09 ms	—	—
		double	Host	117.30 ms	±0.11 ms	—	—
	2000x2000	bool	Host	449.53 ms	±0.08 ms	—	—
		int32	Host	457.24 ms	±0.44 ms	—	—
		int64	Host	448.58 ms	±0.41 ms	—	—
		single	Host	457.84 ms	±0.56 ms	—	—
		double	Host	449.36 ms	±0.26 ms	—	—
invert	1000x1000	single	Host	18.90 ms	±0.02 ms	—	—
		double	Host	35.61 ms	±0.03 ms	—	—
	2000x2000	single	Host	112.00 ms	±2.17 ms	—	—
		double	Host	223.55 ms	±2.12 ms	—	—
	4000x4000	single	Host	782.36 ms	±2.50 ms	—	—
pseudoInvert	1000x1000	single	Host	250.96 ms	±12.57 ms	—	—
		double	Host	419.21 ms	±5.06 ms	—	—
sVD	1000x1000	single	Host	217.31 ms	±9.02 ms	—	—
		double	Host	388.39 ms	±14.50 ms	—	—
symEigDec	1000x1000	single	Host	85.23 ms	±2.30 ms	—	—
		double	Host	130.96 ms	±3.12 ms	—	—
	2000x2000	single	Host	444.07 ms	±10.01 ms	—	—
		double	Host	910.45 ms	±11.48 ms	—	—
equal	1000x1000	bool	Host	9.45 ms	±0.60 ms	—	—
			Cuda	0.69 ms	±0.01 ms	—	—
		int32	Host	8.27 ms	±0.41 ms	0.44 ms	±0.01 ms
			Cuda	1.13 ms	±0.00 ms	—	—
		int64	Host	8.62 ms	±0.66 ms	1.16 ms	±0.01 ms
			Cuda	0.72 ms	±0.00 ms	—	—
		single	Host	8.49 ms	±0.38 ms	0.45 ms	±0.01 ms
			Cuda	1.16 ms	±0.01 ms	—	—
		double	Host	8.91 ms	±0.37 ms	1.15 ms	±0.02 ms
			Cuda	1.15 ms	±0.00 ms	—	—
	2000x2000	bool	Host	31.19 ms	±0.28 ms	—	—
			Cuda	0.98 ms	±0.00 ms	—	—
		int32	Host	31.54 ms	±0.65 ms	2.62 ms	±0.01 ms
			Cuda	0.97 ms	±0.00 ms	—	—
		int64	Host	31.58 ms	±0.78 ms	4.89 ms	±0.00 ms
			Cuda	0.99 ms	±0.01 ms	—	—
		single	Host	31.95 ms	±0.88 ms	2.62 ms	±0.02 ms
			Cuda	0.99 ms	±0.00 ms	—	—
		double	Host	33.75 ms	±0.60 ms	4.88 ms	±0.01 ms
			Cuda	0.98 ms	±0.00 ms	—	—
	4000x4000	bool	Host	118.73 ms	±0.43 ms	—	—
			Cuda	2.23 ms	±0.02 ms	—	—
		int32	Host	118.28 ms	±0.83 ms	10.69 ms	±0.01 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		int64	Host	121.11 ms	±3.16 ms	19.84 ms	±0.02 ms
			Cuda	2.32 ms	±0.03 ms	—	—
		single	Host	120.43 ms	±0.25 ms	10.65 ms	±0.03 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		double	Host	124.86 ms	±2.14 ms	19.95 ms	±0.06 ms
			Cuda	2.29 ms	±0.02 ms	—	—
notEqual	1000x1000	bool	Host	10.13 ms	±1.39 ms	—	—
			Cuda	1.16 ms	±0.00 ms	—	—
		int32	Host	8.27 ms	±0.35 ms	0.44 ms	±0.01 ms
			Cuda	0.83 ms	±0.04 ms	—	—
		int64	Host	8.22 ms	±0.26 ms	1.16 ms	±0.01 ms
			Cuda	1.16 ms	±0.01 ms	—	—
		single	Host	8.92 ms	±0.78 ms	0.45 ms	±0.01 ms
			Cuda	1.12 ms	±0.01 ms	—	—
		double	Host	8.51 ms	±0.23 ms	1.15 ms	±0.02 ms
			Cuda	1.14 ms	±0.01 ms	—	—
	2000x2000	bool	Host	31.55 ms	±0.78 ms	—	—
			Cuda	0.99 ms	±0.00 ms	—	—
		int32	Host	32.64 ms	±1.07 ms	2.62 ms	±0.01 ms
			Cuda	0.98 ms	±0.01 ms	—	—
		int64	Host	32.10 ms	±0.66 ms	4.82 ms	±0.01 ms
			Cuda	0.99 ms	±0.00 ms	—	—
		single	Host	31.60 ms	±0.11 ms	2.62 ms	±0.02 ms
			Cuda	0.98 ms	±0.01 ms	—	—
		double	Host	31.78 ms	±0.55 ms	4.81 ms	±0.02 ms
			Cuda	1.00 ms	±0.00 ms	—	—
	4000x4000	bool	Host	122.58 ms	±0.64 ms	—	—
			Cuda	2.22 ms	±0.02 ms	—	—
		int32	Host	118.26 ms	±0.60 ms	10.68 ms	±0.02 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		int64	Host	120.14 ms	±1.01 ms	19.84 ms	±0.02 ms
			Cuda	2.31 ms	±0.02 ms	—	—
		single	Host	120.93 ms	±0.73 ms	10.76 ms	±0.03 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		double	Host	121.79 ms	±1.38 ms	19.99 ms	±0.02 ms
			Cuda	2.28 ms	±0.02 ms	—	—
less	1000x1000	bool	Host	10.27 ms	±1.49 ms	—	—
			Cuda	1.15 ms	±0.00 ms	—	—
		int32	Host	8.54 ms	±0.23 ms	0.44 ms	±0.01 ms
			Cuda	0.80 ms	±0.01 ms	—	—
		int64	Host	8.81 ms	±0.31 ms	1.15 ms	±0.01 ms
			Cuda	1.16 ms	±0.01 ms	—	—
		single	Host	8.10 ms	±0.17 ms	0.45 ms	±0.01 ms
			Cuda	1.15 ms	±0.00 ms	—	—
		double	Host	8.48 ms	±0.12 ms	1.15 ms	±0.02 ms
			Cuda	1.13 ms	±0.01 ms	—	—
	2000x2000	bool	Host	33.69 ms	±0.92 ms	—	—
			Cuda	0.97 ms	±0.01 ms	—	—
		int32	Host	30.89 ms	±0.50 ms	2.60 ms	±0.01 ms
			Cuda	0.98 ms	±0.01 ms	—	—
		int64	Host	31.01 ms	±0.37 ms	4.81 ms	±0.01 ms
			Cuda	0.99 ms	±0.00 ms	—	—
		single	Host	31.41 ms	±0.22 ms	2.62 ms	±0.02 ms
			Cuda	0.99 ms	±0.00 ms	—	—
		double	Host	31.63 ms	±0.05 ms	4.83 ms	±0.02 ms
			Cuda	1.00 ms	±0.00 ms	—	—
	4000x4000	bool	Host	125.67 ms	±0.61 ms	—	—
			Cuda	2.21 ms	±0.02 ms	—	—
		int32	Host	119.32 ms	±1.91 ms	10.63 ms	±0.01 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		int64	Host	119.51 ms	±2.08 ms	19.93 ms	±0.01 ms
			Cuda	2.31 ms	±0.02 ms	—	—
		single	Host	119.36 ms	±0.99 ms	10.76 ms	±0.03 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		double	Host	124.35 ms	±1.77 ms	19.99 ms	±0.07 ms
			Cuda	2.27 ms	±0.02 ms	—	—
lessOrEqual	1000x1000	bool	Host	9.41 ms	±0.22 ms	—	—
			Cuda	1.16 ms	±0.00 ms	—	—
		int32	Host	8.56 ms	±0.46 ms	0.44 ms	±0.01 ms
			Cuda	1.17 ms	±0.00 ms	—	—
		int64	Host	8.40 ms	±0.27 ms	1.16 ms	±0.01 ms
			Cuda	1.16 ms	±0.02 ms	—	—
		single	Host	8.75 ms	±0.59 ms	0.45 ms	±0.01 ms
			Cuda	1.15 ms	±0.01 ms	—	—
		double	Host	8.12 ms	±0.15 ms	1.15 ms	±0.02 ms
			Cuda	0.71 ms	±0.01 ms	—	—
	2000x2000	bool	Host	32.03 ms	±0.29 ms	—	—
			Cuda	0.98 ms	±0.00 ms	—	—
		int32	Host	31.52 ms	±0.49 ms	2.61 ms	±0.01 ms
			Cuda	0.96 ms	±0.01 ms	—	—
		int64	Host	31.51 ms	±0.47 ms	4.82 ms	±0.01 ms
			Cuda	0.99 ms	±0.00 ms	—	—
		single	Host	31.95 ms	±0.35 ms	2.62 ms	±0.02 ms
			Cuda	0.99 ms	±0.02 ms	—	—
		double	Host	32.41 ms	±0.18 ms	4.83 ms	±0.02 ms
			Cuda	1.00 ms	±0.00 ms	—	—
	4000x4000	bool	Host	124.45 ms	±0.99 ms	—	—
			Cuda	2.22 ms	±0.03 ms	—	—
		int32	Host	116.87 ms	±1.18 ms	10.68 ms	±0.01 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		int64	Host	123.70 ms	±2.69 ms	19.95 ms	±0.01 ms
			Cuda	2.32 ms	±0.02 ms	—	—
		single	Host	119.34 ms	±1.21 ms	10.81 ms	±0.17 ms
			Cuda	2.22 ms	±0.02 ms	—	—
		double	Host	128.21 ms	±1.84 ms	19.99 ms	±0.07 ms
			Cuda	2.29 ms	±0.02 ms	—	—
greater	1000x1000	bool	Host	10.55 ms	±1.31 ms	—	—
			Cuda	1.15 ms	±0.00 ms	—	—
		int32	Host	8.08 ms	±0.22 ms	0.44 ms	±0.01 ms
			Cuda	1.16 ms	±0.02 ms	—	—
		int64	Host	8.72 ms	±0.43 ms	1.16 ms	±0.01 ms
			Cuda	0.71 ms	±0.01 ms	—	—
		single	Host	8.52 ms	±0.38 ms	0.45 ms	±0.01 ms
			Cuda	1.13 ms	±0.01 ms	—	—
		double	Host	8.63 ms	±0.23 ms	1.15 ms	±0.02 ms
			Cuda	1.13 ms	±0.01 ms	—	—
	2000x2000	bool	Host	32.99 ms	±0.81 ms	—	—
			Cuda	0.98 ms	±0.01 ms	—	—
		int32	Host	31.27 ms	±0.38 ms	2.61 ms	±0.01 ms
			Cuda	0.98 ms	±0.00 ms	—	—
		int64	Host	32.70 ms	±0.84 ms	4.84 ms	±0.01 ms
			Cuda	1.01 ms	±0.01 ms	—	—
		single	Host	31.43 ms	±0.55 ms	2.61 ms	±0.02 ms
			Cuda	0.98 ms	±0.00 ms	—	—
		double	Host	32.14 ms	±0.56 ms	4.83 ms	±0.02 ms
			Cuda	1.00 ms	±0.00 ms	—	—
	4000x4000	bool	Host	124.18 ms	±0.52 ms	—	—
			Cuda	2.22 ms	±0.02 ms	—	—
		int32	Host	118.52 ms	±2.09 ms	10.70 ms	±0.13 ms
			Cuda	2.23 ms	±0.02 ms	—	—
		int64	Host	123.09 ms	±0.75 ms	19.95 ms	±0.03 ms
			Cuda	2.32 ms	±0.02 ms	—	—
		single	Host	123.16 ms	±2.51 ms	10.75 ms	±0.03 ms
			Cuda	2.23 ms	±0.02 ms	—	—
		double	Host	122.09 ms	±1.79 ms	19.95 ms	±0.05 ms
			Cuda	2.29 ms	±0.02 ms	—	—
greaterOrEqual	1000x1000	bool	Host	10.01 ms	±0.44 ms	—	—
			Cuda	1.13 ms	±0.01 ms	—	—
		int32	Host	8.58 ms	±0.34 ms	0.44 ms	±0.01 ms
			Cuda	1.13 ms	±0.01 ms	—	—
		int64	Host	8.45 ms	±0.31 ms	1.16 ms	±0.01 ms
			Cuda	0.70 ms	±0.02 ms	—	—
		single	Host	8.19 ms	±0.19 ms	0.45 ms	±0.01 ms
			Cuda	1.16 ms	±0.00 ms	—	—
		double	Host	8.76 ms	±0.32 ms	1.15 ms	±0.02 ms
			Cuda	0.69 ms	±0.00 ms	—	—
	2000x2000	bool	Host	33.67 ms	±0.55 ms	—	—
			Cuda	0.96 ms	±0.00 ms	—	—
		int32	Host	31.55 ms	±0.52 ms	2.59 ms	±0.01 ms
			Cuda	0.98 ms	±0.00 ms	—	—
		int64	Host	32.30 ms	±0.90 ms	4.82 ms	±0.01 ms
			Cuda	1.01 ms	±0.00 ms	—	—
		single	Host	31.42 ms	±0.60 ms	2.60 ms	±0.02 ms
			Cuda	0.99 ms	±0.00 ms	—	—
		double	Host	32.14 ms	±0.91 ms	4.83 ms	±0.02 ms
			Cuda	0.99 ms	±0.00 ms	—	—
	4000x4000	bool	Host	123.79 ms	±0.94 ms	—	—
			Cuda	2.22 ms	±0.02 ms	—	—
		int32	Host	117.71 ms	±1.87 ms	10.75 ms	±0.02 ms
			Cuda	2.22 ms	±0.04 ms	—	—
		int64	Host	121.48 ms	±1.93 ms	19.91 ms	±0.01 ms
			Cuda	2.31 ms	±0.02 ms	—	—
		single	Host	117.64 ms	±0.69 ms	10.72 ms	±0.03 ms
			Cuda	2.20 ms	±0.02 ms	—	—
		double	Host	122.14 ms	±1.16 ms	19.97 ms	±0.01 ms
			Cuda	2.28 ms	±0.02 ms	—	—
not	1000x1000	bool	Host	7.60 ms	±0.24 ms	0.11 ms	±0.00 ms
			Cuda	0.66 ms	±0.02 ms	—	—
	2000x2000		Host	26.01 ms	±0.41 ms	0.45 ms	±0.01 ms
			Cuda	0.86 ms	±0.00 ms	—	—
	4000x4000		Host	98.41 ms	±0.85 ms	2.82 ms	±0.02 ms
			Cuda	1.78 ms	±0.02 ms	—	—
and	1000x1000	bool	Host	8.85 ms	±0.12 ms	0.14 ms	±0.00 ms
			Cuda	1.16 ms	±0.01 ms	—	—
	2000x2000		Host	33.31 ms	±0.76 ms	0.79 ms	±0.01 ms
			Cuda	0.96 ms	±0.00 ms	—	—
	4000x4000		Host	127.27 ms	±1.62 ms	3.90 ms	±0.02 ms
			Cuda	2.20 ms	±0.03 ms	—	—
or	1000x1000	bool	Host	9.41 ms	±0.61 ms	0.14 ms	±0.00 ms
			Cuda	1.16 ms	±0.01 ms	—	—
	2000x2000		Host	33.02 ms	±0.63 ms	0.78 ms	±0.01 ms
			Cuda	0.98 ms	±0.00 ms	—	—
	4000x4000		Host	123.16 ms	±1.25 ms	3.87 ms	±0.01 ms
			Cuda	2.22 ms	±0.02 ms	—	—
xor	1000x1000	bool	Host	8.84 ms	±0.12 ms	1.53 ms	±0.00 ms
			Cuda	1.15 ms	±0.01 ms	—	—
	2000x2000		Host	30.07 ms	±0.54 ms	6.15 ms	±0.01 ms
			Cuda	0.98 ms	±0.01 ms	—	—
	4000x4000		Host	114.15 ms	±1.93 ms	24.66 ms	±0.01 ms
			Cuda	2.22 ms	±0.02 ms	—	—
countTrueAxis	1000x1000	bool	Host	8.47 ms	±0.98 ms	1.59 ms	±0.00 ms
	2000x2000		Host	27.86 ms	±0.73 ms	6.25 ms	±0.01 ms
	4000x4000		Host	104.74 ms	±0.99 ms	24.97 ms	±0.05 ms
allAxis	1000x1000	bool	Host	6.31 ms	±0.24 ms	—	—
			Cuda	3.11 ms	±0.01 ms	—	—
	2000x2000		Host	23.09 ms	±0.53 ms	—	—
			Cuda	5.87 ms	±0.00 ms	—	—
	4000x4000		Host	87.70 ms	±1.47 ms	—	—
			Cuda	11.66 ms	±0.00 ms	—	—
anyAxis	1000x1000	bool	Host	6.03 ms	±0.20 ms	—	—
			Cuda	3.13 ms	±0.00 ms	—	—
	2000x2000		Host	23.74 ms	±1.11 ms	—	—
			Cuda	5.90 ms	±0.00 ms	—	—
	4000x4000		Host	87.91 ms	±1.60 ms	—	—
			Cuda	11.70 ms	±0.00 ms	—	—
trueIndices	1000x1000	bool	Host	50.63 ms	±0.12 ms	—	—
	2000x2000		Host	199.15 ms	±0.26 ms	—	—
	4000x4000		Host	800.91 ms	±1.53 ms	—	—
ifThenElse	1000x1000	bool	Host	10.53 ms	±0.66 ms	—	—
			Cuda	1.04 ms	±0.01 ms	—	—
		int32	Host	10.87 ms	±0.63 ms	7.00 ms	±0.01 ms
			Cuda	1.09 ms	±0.00 ms	—	—
		int64	Host	11.38 ms	±0.43 ms	7.22 ms	±0.01 ms
			Cuda	1.45 ms	±0.01 ms	—	—
		single	Host	11.23 ms	±0.63 ms	7.00 ms	±0.01 ms
			Cuda	1.12 ms	±0.00 ms	—	—
		double	Host	12.05 ms	±1.20 ms	7.22 ms	±0.01 ms
			Cuda	1.45 ms	±0.00 ms	—	—
	2000x2000	bool	Host	37.32 ms	±0.42 ms	—	—
			Cuda	1.09 ms	±0.00 ms	—	—
		int32	Host	39.21 ms	±0.33 ms	28.11 ms	±0.01 ms
			Cuda	1.11 ms	±0.01 ms	—	—
		int64	Host	39.94 ms	±0.28 ms	28.98 ms	±0.01 ms
			Cuda	1.17 ms	±0.00 ms	—	—
		single	Host	39.09 ms	±0.50 ms	28.12 ms	±0.01 ms
			Cuda	1.12 ms	±0.00 ms	—	—
		double	Host	40.67 ms	±0.58 ms	29.01 ms	±0.02 ms
			Cuda	1.18 ms	±0.02 ms	—	—
	4000x4000	bool	Host	144.07 ms	±0.53 ms	—	—
			Cuda	2.61 ms	±0.02 ms	—	—
		int32	Host	145.52 ms	±0.38 ms	140.58 ms	±0.02 ms
			Cuda	2.75 ms	±0.02 ms	—	—
		int64	Host	152.34 ms	±1.49 ms	173.15 ms	±0.03 ms
			Cuda	3.05 ms	±0.07 ms	—	—
		single	Host	148.15 ms	±1.82 ms	140.63 ms	±0.03 ms
			Cuda	2.73 ms	±0.03 ms	—	—
		double	Host	153.73 ms	±1.83 ms	173.27 ms	±0.10 ms
			Cuda	3.06 ms	±0.06 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 12:47:09 GMT.