Repository navigation
Higher than 1-D arrays in dot #468
Description
Trying to use dot from pygpu with OpenCL. However it seems to throw an error due to the lack of collective operations. Am able to work around this using BLAS functions (e.g. gemm). After looking around in the code was able to find the error and the C function involved, but not exactly how we got to it.
Have a few questions. How are collective operations involved here (higher dimensions?)? Would it make sense to special case some ranks of arrays to avoid collective operations and go straight to BLAS operations? If I'm misunderstanding things here, please feel free to correct me.
Details
In [1]: import pygpu
In [2]: import pygpu.blas
In [3]: import numpy
In [4]: import numpy as np
In [5]: c = pygpu.init("opencl0:1")
In [6]: a = pygpu.gpuarray.zeros((10, 11), dtype=np.float32, context=c, cls=pygpu.ndgpuarray)
In [7]: pygpu.blas.gemm(1, a, a.T, 0)
OpenCL Build Warning : Compiler build log:
<program source>:101:54: warning: double precision constant requires cl_khr_fp64, casting to single precision
lA[ 0*localAStride ] = ( globalARow(0) >= M) ? 0.0 : A[ GET_GLOBAL_INDEX_A( globalARow(0), globalACol(0) ) ];
^
<program source>:104:53: warning: double precision constant requires cl_khr_fp64, casting to single precision
lB[ 0*localBStride ] = (globalBCol(0) >= N) ? 0.0 : B[ GET_GLOBAL_INDEX_B( globalBRow(0), globalBCol(0) ) ];
^
Out[7]:
gpuarray.array([[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]], dtype=float32)
In [8]: pygpu.blas.dot(a, a.T)
---------------------------------------------------------------------------
GpuArrayException Traceback (most recent call last)
<ipython-input-8-ee1cc1b223fb> in <module>()
----> 1 pygpu.blas.dot(a, a.T)
pygpu/blas.pyx in pygpu.blas.dot (pygpu/blas.c:2400)()
pygpu/blas.pyx in pygpu.blas.pygpu_blas_rdot (pygpu/blas.c:1713)()
GpuArrayException: ('Collectives operations not supported on OpenCL', 2)
Activity
[-]Dispatch to BLAS if dot can[/-][+]Collective operations and dot[/+]abergeron commentedon Jun 27, 2017
The message is most certainly an mistake somewhere. Possibly a reuse of the message from another failure (which is an issue still present on 0.6.5 that should be fixed with 0.6.6).
Can you try with the current master and see if you get the same problem?
jakirkham commentedon Jun 27, 2017
Sure, retried with 0.6.8, which I built locally. Am getting an error related to dimensionality instead. Looks like 2-D arrays are not supported by
dot. Is this expected? If so, what are your thoughts on expandingdotto N-D arrays of rank greater than 1?Details
nouiz commentedon Jun 27, 2017
[-]Collective operations and dot[/-][+]Higher than 1-D arrays in dot[/+]abergeron commentedon Jun 27, 2017
There isn't a pygpu.dot() akin to numpy yet. I welcome patches.
Otherwise, you're right this it the blas dot, not the numpy dot.