cuda : use CUBLAS_COMPTE_F32 insted of CUBLAS_COMPUTE_F16

[ggerganov]

On some video cards, this can be faster.

GTX 1660

./extra/bench-all.sh 1 1

• master

GPU	Config	Model	Th	Enc.	Dec.	Bch5	PP	Commit
GTX 1660	AVX2 BLAS CUDA	tiny	1	105.95	5.90	0.63	0.20	f52e74d
GTX 1660	AVX2 BLAS CUDA	tiny-q5_0	1	106.29	2.63	0.31	0.21	f52e74d
GTX 1660	AVX2 BLAS CUDA	tiny-q5_1	1	106.35	2.62	0.30	0.20	f52e74d
GTX 1660	AVX2 BLAS CUDA	base	1	236.20	7.56	1.01	0.33	f52e74d
GTX 1660	AVX2 BLAS CUDA	base-q5_0	1	237.01	4.18	0.44	0.33	f52e74d
GTX 1660	AVX2 BLAS CUDA	base-q5_1	1	237.01	4.17	0.43	0.33	f52e74d
GTX 1660	AVX2 BLAS CUDA	small	1	890.34	21.70	2.33	1.00	f52e74d
GTX 1660	AVX2 BLAS CUDA	small-q5_0	1	893.45	11.79	1.08	1.02	f52e74d
GTX 1660	AVX2 BLAS CUDA	small-q5_1	1	893.44	11.77	1.04	1.02	f52e74d
GTX 1660	AVX2 BLAS CUDA	medium	1	2767.58	54.52	6.80	2.40	f52e74d
GTX 1660	AVX2 BLAS CUDA	medium-q5_0	1	2779.65	26.92	2.57	2.46	f52e74d
GTX 1660	AVX2 BLAS CUDA	medium-q5_1	1	2779.30	26.96	2.42	2.46	f52e74d
GTX 1660	AVX2 BLAS CUDA	medium-dis	1	2766.26	7.75	1.27	0.40	f52e74d

• PR

GPU	Config	Model	Th	Enc.	Dec.	Bch5	PP	Commit
GTX 1660	AVX2 BLAS CUDA	tiny	1	105.09	1.19	0.72	0.20	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	tiny-q5_0	1	107.15	0.96	0.34	0.21	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	tiny-q5_1	1	107.12	0.96	0.33	0.21	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	base	1	232.68	1.73	1.18	0.33	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	base-q5_0	1	238.91	1.42	0.49	0.34	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	base-q5_1	1	238.63	1.44	0.48	0.34	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	small	1	921.25	6.14	4.09	1.02	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	small-q5_0	1	899.08	3.48	1.27	1.03	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	small-q5_1	1	899.34	3.44	1.22	1.03	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	medium	1	2891.58	18.21	9.18	2.53	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	medium-q5_0	1	2792.70	7.89	2.94	2.47	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	medium-q5_1	1	2792.82	7.94	2.80	2.48	c8b3bc6
GTX 1660	AVX2 BLAS CUDA	medium-dis	1	2890.40	5.22	1.46	0.42	c8b3bc6

[Alvarocda]

Is there any prediction on when this PR will be merged?
It is working correctly in the tests I did with my Quadro T1000

[RaitoBezarius]

I tested this PR on a Compute Capability 3.7 GPU, and it doesn't seem to be supported, FWIW. I assume this is because there's a mixing between FP16 and FP32, I guess.