April 2019 - PyOpenCL - tiker.net Mailing Lists

Re: Is there any options to set the data type for device which don't support 64bit.

by Andreas Kloeckner

Wensheng, Wensheng Zou <wzou(a)panyi.ai> writes: > I know that the api has_double_support(dev) will return back if dev support double type, and with this api we can decide which data type can be used. > But I found that even if I don't use float64 or int64 for test data type, pyopencl will generate kernel code that contains double or long type as arguments. > Such as: > __kernel void plb_build_list_write(__global int *plb_mylist_list, __global int *plb_mylist_start_index, __global double *input_list, index_type n) > And so on. > All those double or long type in the kernel code may cause building errors on the devices don't support 64 bit. > So what should I do to avoid pyopencl generate kernel code contains double or long type, is there any options to deal with it? Thanks for your message! The behavior you describe definitely sounds like a bug. If there are no 64-bit types in the input, there shouldn't be any in the code. So could you open an issue on Github for this? Secondly, in order to act on your bug report, we'll need some way of reproducing your observation, i.e. some self-contained, instantly runnable code that shows the issue you observe. Andreas

9 months, 2 weeks

1
0
0 / 0

Is there any options to set the data type for device which don't support 64bit.

by Wensheng Zou

Hi I know that the api has_double_support(dev) will return back if dev support double type, and with this api we can decide which data type can be used. But I found that even if I don't use float64 or int64 for test data type, pyopencl will generate kernel code that contains double or long type as arguments. Such as: __kernel void plb_build_list_write(__global int *plb_mylist_list, __global int *plb_mylist_start_index, __global double *input_list, index_type n) And so on. All those double or long type in the kernel code may cause building errors on the devices don't support 64 bit. So what should I do to avoid pyopencl generate kernel code contains double or long type, is there any options to deal with it? Thanks! Wensheng

9 months, 2 weeks

1
0
0 / 0

Option to choose 64bit or 32bit supported on device.

by wzou＠panyi.ai

The API has_double_support(dev) decides if double and long type supported on the dev. But I found that even I didn't use float64 or int64, the kernel code generate by pyopencl contains double or long type args too, how to deal with it?

9 months, 2 weeks

1
0
0 / 0

PyOpenCL Installation

by DITTER, CHRISTOPHER J 1st Lt USAF HAF NASIC/ACNW

Hi, I am a new user trying to install PyOpenCL with the latest version of Anaconda (for python 3.7) to a standalone HPC. After installing Anaconda to my work directory, installing the correct PyOpenCL package for python 3.7, symbolically linking the nvidia.icd from /etc/OpenCL/vendors/nvidia.icd to /p/work1/myUsername/myAnaconda3/etc/OpenCL/vendors/nvidia.icd as the documentation states, launching an interactive GPU node, launching python, and running “import pyopencl”, I get the following error: “ImportError: /lib64/libOpenCL.so.1: version ‘OPENCL_2.1’ not found (required by /p/work1/ myUsername/myAnaconda3/lib/python3.7/pyopencl/_cl.cpython-37m-x86_64-linux-gnu.so”. The file /etc/OpenCL/vendors/nvidia.icd contains the text “libnvidia-opencl.so.1”. I originally submitted a ticket to the HPC admins asking them to update OpenCL to version 2.1 due to this error I was getting, but later learned that the GPU’s on this HPC only support up to OpenCL 1.2. After looking over some posts here, I’m tempted to think that there may be an issue with the OpenCL header files specifying that the OpenCL version is 2.1 even though the nvidia drivers only support OpenCL 1.2. Does this sound like a good path for me to pursue with the admins or do you have any other suggestions for this error? Thanks! Chris

9 months, 3 weeks

2
2
0 / 0

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

PyOpenCL April 2019