[FEA] Replace numpy mod op in `Series.hash_encode`

[cmgreen210]

The current implementation of hash_encode is bottlenecked at the numpy call https://github.com/rapidsai/cudf/blob/branch-0.6/python/cudf/dataframe/series.py#L1078. This computation should be moved to the gpu.

[cmgreen210]

I will work on a fix for this.

[cmgreen210]

This will be easier than expected. We'll change the line to hashed_values = np.mod(hashed_values.data.to_array(), stop)

In [44]: %time np.mod(hv, 100)
CPU times: user 24.8 s, sys: 404 ms, total: 25.2 s
Wall time: 25.2 s
Out[44]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

In [45]: %time np.mod(hv.data.to_array(), 100)
CPU times: user 4 ms, sys: 0 ns, total: 4 ms
Wall time: 1.58 ms
Out[45]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

[kkraus14]

This will be easier than expected. We'll change the line to hashed_values = np.mod(hashed_values.data.to_array(), stop)

In [44]: %time np.mod(hv, 100)
CPU times: user 24.8 s, sys: 404 ms, total: 25.2 s
Wall time: 25.2 s
Out[44]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

In [45]: %time np.mod(hv.data.to_array(), 100)
CPU times: user 4 ms, sys: 0 ns, total: 4 ms
Wall time: 1.58 ms
Out[45]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

That will still run it on the CPU and trigger a device to host copy on the .to_array() call.

[kkraus14]

I don't recall if we have a modulo binary op in libcudf that you'd probably want here, @jrhemstad?

[cmgreen210]

This will be easier than expected. We'll change the line to hashed_values = np.mod(hashed_values.data.to_array(), stop)

In [44]: %time np.mod(hv, 100)
CPU times: user 24.8 s, sys: 404 ms, total: 25.2 s
Wall time: 25.2 s
Out[44]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

In [45]: %time np.mod(hv.data.to_array(), 100)
CPU times: user 4 ms, sys: 0 ns, total: 4 ms
Wall time: 1.58 ms
Out[45]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

That will still run it on the CPU and trigger a device to host copy on the .to_array() call.

Good point. But the speed up is probably good enough for now. wdyt?

[kkraus14]

This will be easier than expected. We'll change the line to hashed_values = np.mod(hashed_values.data.to_array(), stop)

In [44]: %time np.mod(hv, 100)
CPU times: user 24.8 s, sys: 404 ms, total: 25.2 s
Wall time: 25.2 s
Out[44]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

In [45]: %time np.mod(hv.data.to_array(), 100)
CPU times: user 4 ms, sys: 0 ns, total: 4 ms
Wall time: 1.58 ms
Out[45]: array([ 9, 69, 85, ..., 87,  1, 22], dtype=int32)

That will still run it on the CPU and trigger a device to host copy on the .to_array() call.

Good point. But the speed up is probably good enough for now. wdyt?

Implementing a binary modulo op on the GPU should be trivial. If the arrays are larger or we call this on a bunch of columns the allocations / memory copies would become expensive.

[jrhemstad]

I don't recall if we have a modulo binary op in libcudf that you'd probably want here, @jrhemstad?

So you want a binary op of doing a modulo of a column by a scalar? This will be added in #892

[cmgreen210]

@jrhemstad yes. Will this be added by the 0.6 release? If not might want to do an intermediate fix.

[kkraus14]

@jrhemstead yes. Will this be added by the 0.6 release? If not might want to do an intermediate fix.

@cmgreen210 as a stop gap it should be pretty straightforward to write a numba kernel that does a modulo operator until we move it into libcudf.

[cmgreen210]

@jrhemstead yes. Will this be added by the 0.6 release? If not might want to do an intermediate fix.

@cmgreen210 as a stop gap it should be pretty straightforward to write a numba kernel that does a modulo operator until we move it into libcudf.

@kkraus14 Sounds good. I'll have that up asap.