[TODO] Make CR equal to 1 instead of 0.x (such as 0.9 or 0.8)

[jxybb]

Thanks for open source this awesome compressor!
I tried the Python compression API, but it seems the compressor didn't work for 1-D FP32 numpy array.
I tested eb_mode=1 on an array with a lot of zeros, but the compressor seems just compressed it to 120 bytes no matter what eb_rel value I use.
And Interestly, the compressor works after I covert the array to FP64.
Could someone help to look into this? Thanks!

[ayzk]

@jz0909 1D FP32 array works on my side. Could you share your data with me so I could dig into it? Thanks.

[jxybb]

Thanks @ayzk , looks the issue is not limited to Python. I also see it with C++.
For an 1D float array with 1000 elements on N(0, 1) distribution, when I set eb_rel to 1e-3, the compressed size is larger than uncompressed size. I tried to increase the array size to 5000, and I saw the same issue when eb_rel is 1e-4.

[ayzk]

@jz0909 Thanks for pointing out the problem. The size of 1000 is extremely small and SZ3 is not optimized for such cases yet. I verified the error control is correct. We will make the compressed size equal to 1 in an updated version.

A larger data is more suitable for SZ3. The following code will output a compression ratio around 1.9

data = np.random.normal(0, 1, 5000).astype(np.float32)
data_cmpr, cmpr_ratio = sz.compress(data, 1, 0, 1e-3, 0)