Expand `ktype` support in codes_get_array?

[TheoMathurin]

Is your feature request related to a problem? Please describe.

I installed the 2.43.0 version of the python bindings via pip.

The grib_get_array function allows the user to specify a ktype. Among numerical dtypes, int, float and float32 seem supported:

    if ktype is int:
        result = grib_get_long_array(msgid, key)
    elif ktype is float or ktype is np.float64:
        result = grib_get_double_array(msgid, key)
    elif ktype is np.float32:
        result = grib_get_float_array(msgid, key)
    elif ktype is str:
        result = grib_get_string_array(msgid, key)
    elif ktype is bytes:
        result = grib_get_string_array(msgid, key)

However, I've tried calling it with float32 but received the following error: gribapi.errors.FunctionNotImplementedError: Function not yet implemented

Besides, it would be great to expand to range of integer precisions supported, for instance I have files with int16 and would like to retrieve the values in their native precision.

NB: if ktype is not provided, the grib_get_native_type gets called and its doctrings says possible returned numerical dtypes only include int or float.

Describe the solution you'd like

I'd like the function to accept more dtypes, clearly separating various integer dtypes: at least int/int64, int32 and int16. Also, if confirmed, the bug with float32 should be addressed.

Describe alternatives you've considered

Casting afterwards is possible but inelegant and probably sub-optimal.

Additional context

No response

Organisation

No response

[shahramn]

Re: the error message
What kind of data are you using? Can you send us one message from your input (is it GRIB or BUFR?)
Also a small code snippet so we can reproduce the problem.
Thanks

[TheoMathurin]

Thank you for your response.

Below is a message extracted from a IASI L1 BUFR file that I use. In this file, the most important data correspond to the “scaledIasiRadiance” key, and they are also by far the largest. Since they are encoded on 16 bits, I’d like to read them at their native precision rather than retrieving them as 64‑bit values and then immediately casting them to reduce memory usage.

W_XX-EUMETSAT-Darmstadt,SOUNDING+SATELLITE,METOPC+IASI_C_EUMR_20240505124454_28506_eps_o_l1_001.zip

Here's a small script to read it and confirm that you get int64:

file = "W_XX-EUMETSAT-Darmstadt,SOUNDING+SATELLITE,METOPC+IASI_C_EUMR_20240505124454_28506_eps_o_l1.bin_001"
with open(file, "rb") as f:
    while 1:
        bufr = codes_bufr_new_from_file(f)
        if bufr is None:
            break
        codes_set(bufr, "unpack", 1)
        radiance = codes_get_array(bufr, "scaledIasiRadiance")

print(radiance.dtype)