1d
comment Segmented Sort with CUDPP/Thrust
If the data type is of size 32, you could also sort the array of type u64 which will just be scanned_flag<<32 | A. That way you do not need to worry about the range of intermediate values.
Jan
26
comment About warp voting function
Cool. Found your email. I will write to you soon.
Jan
26
comment About warp voting function
Also, I think you should add a LICENSE.txt file to your repo.
Jan
26
comment About warp voting function
Actually it would be nice if you have some benchmarks against thrust library. Also I think that in line 78 of cuCompactor.cuh, it should be possible to have another global array called d_output_index which would contain the the value of idx corresponding to where the original data came from. Am I correct?
Jan
26
comment About warp voting function
This sounds super interesting. Is there any implementation I can look at?
Jan
6
comment Small sized binary searches on CUDA GPUs
@m.s. I have clearly mentioned in the edit that 'The values could range from 0 toINT64_MAX.'
Jan
5
comment Small sized binary searches on CUDA GPUs
@RobertCrovella I have added an example. The inputValues remain unchanged. Elements in removeValues are replaced with -1 if these elements exist in inputValues. And yes, all elements involved will be positive. If the signedness of the datatypes is a problem, the values can be marked with UINT64_MAX as well. My only requirement is to mark existing values with a conspicuous number so as to identify 'empty' entries sometime later in the code.
Jan
5
revised Small sized binary searches on CUDA GPUs
added 537 characters in body
Jan
5
comment Small sized binary searches on CUDA GPUs
@m.s. Which thrust function do you think suits my needs in this situation?
Jan
5
comment Small sized binary searches on CUDA GPUs
There are quite a few reasons for that, performance being one of them. This is a substep of a larger piece of code. The array inputValues is generated in a very nontrivial way. You could say that the sorted 32 element chunks of inputValues were not contiguous to begin. I do not wish to copy them to a contiguous chunk of memory but rather keep them in registers or shared memory at best. Gathering from separated global memory to contiguous memory does not seem like a good idea to me.
Jan
5
asked Small sized binary searches on CUDA GPUs
Aug
30
accepted CUDA Thrust reduce_by_key using less memory
Aug
26
asked CUDA Thrust reduce_by_key using less memory
Aug
20
awarded Supporter
Aug
20
awarded Scholar
Aug
20
accepted Tmux bindings using Ctrl+comma
Aug
19
answered How can I set up seamless navigation of vim and iterm2 splits?
Aug
4
asked OpenMP Single Producer Multiple Consumer
Jul
24
awarded Custodian
Jul
24
reviewed Approve suggested edit on Tmux bindings using Ctrl+comma
1 2 3 4 5