ENH: Add select methods for IPC

[TomAugspurger]

Allows for select statements that returns memory handles, rather than a result
set. ATM this is implemented on the connection object, rather than the cursor.
I'm not sure which is more appropriate. For it to belong on the Cursor object,
I think we would need to bring the parsing code from pygdf into pymapd, so we
can parse the result into a result set, and return an Iterator over that.

[TomAugspurger]

@sklam could you take a look at this on a machine with a GPU when you get a chance? I'm still working on getting access to one.

This script should work assuming you have a mapd server running on 9091 with the default login info, and the flights_2008_10k dataset loaded.

import numpy as np
from numba import cuda
from numba.cuda.cudadrv import drvapi
from pygdf.gpuarrow import GpuArrowReader
import pymapd

table = "flights_2008_10k"
con = pymapd.connect(user="mapd", password="HyperInteractive",
                     host="localhost", dbname="mapd")
statement = "select depdelay from {}".format(table)

results = con.select_ipc_gpu(statement)


# Now for pygdf

ipc_handle = drvapi.cu_ipc_mem_handle(*results.df_handle)
ipch = cuda.driver.IpcHandle(None, ipc_handle, size=results.df_size)
ctx = cuda.current_context()
dptr = ipch.open(ctx)

dtype = np.dtype(np.byte)
darr = cuda.devicearray.DeviceNDArray(shape=dptr.size, strides=dtype.itemsize,
                                      dtype=dtype, gpu_data=dptr)
reader = GpuArrowReader(darr)
reader.to_dict()

---

Some questions you may be able to answer

• I've put these on the Connection object, instead of Cursor, which is more about building up results sets and then fetching them. For it to make sense on a Cursor, I think we would have to move the GpuArrowReader code to pymapd (or pymapd could import it and call it?)
• Do the method names sound OK to you?