Skip to content

Add CUDA version of 'radial' #979

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 59 commits into from
Jul 11, 2022
Merged

Add CUDA version of 'radial' #979

merged 59 commits into from
Jul 11, 2022

Conversation

drroe
Copy link
Contributor

@drroe drroe commented Jul 11, 2022

Version 6.11.0.

Adds a GPU version of the radial command using CUDA. Currently this is only for atom-atom RDFs (one or two masks). The speedup can be significant. Here's a comparison of the previous OpenMP code (run on an Intel sandybridge, 4 cores) to the new CUDA code (run on a GTX 780):

  CPPTRAJ BENCH (3): Radial solvent-solvent benchmark.
    TOP: ../FtuFabI.tip3p.parm7  CRD: ../FtuFabI.tip3p.nc  NTRAJIN: 1
    TRAJIN_ARGS: 1 10
	/home/droe/Cpptraj/rdfgpu.cpptraj/bin/cpptraj.OMP
	Iter   Time Context_Switches  Waits      FPS
	0     91.12             5982     36    0.110
	1     91.26             6146     36    0.110
	2     91.22             7455     36    0.110
	/home/droe/Cpptraj/rdfgpu.cpptraj/bin/cpptraj.cuda
	Iter   Time Context_Switches  Waits      FPS
	0      1.26                1     92   12.518
	1      1.26               21     89   12.527
	2      1.23               22     85   12.506
  Baseline  :  91.2000
  Benchmark :   1.2500
  Speedup   :  72.96
  %Change   :  98.63 %

The GPU RDF is calculated in single precision, so small numerical differences on the order of 0.0002-0.0004 can be expected in the resulting histograms.

This PR also includes a 2x speedup in the CPU code when only a single mask is specified (e.g. when calculating water oxygens to water oxygens RDF).

drroe added 30 commits July 6, 2022 19:32
@drroe
Start rdf gpu kernel
45ea34b
@drroe
Add rdf entry function to Makefile
a1030d6
@drroe
Add test for RDF
cd046c6
@drroe
Fix num blocks calc
d849cf5
@drroe
Change block dim to 32 to avoid creating too many threads per block. …
39a1b3a 
…Start the rdf bin kernel
@drroe
Add maximum2 and one_over_spacing
ff9fae4
@drroe
Add atomicAdd to the RDF histogram
f54d9cf
@drroe
Add kernel launch. Ensure device histogram memory is zeroed
a729aba
@drroe
Fix fxn definition and call
75c3913
@drroe
Ensure results are copied back. Make sure coords are uploaded to device.
ea3c658
@drroe
Add check that spacing is not greater than the max
e00293f
@drroe
Add separate test
6fcd283
@drroe
Merge remote-tracking branch 'origin/rdfgpu' into rdfgpu
7245e38
@drroe
Enable some debug info. Do kernel error checking
2775909
@drroe
Report max threads per block
d560184
@drroe
Hide some debug info. Make cuda errors non-fatal while debugging
50dd4eb
@drroe
Better labeling
84aa280
@drroe
Improve error checking for memory allocation and setting. Correct inv…
957af26 
…alid arg to cudaMemset
@drroe
Report max threads in each dim
99ec3c4
@drroe
Report max threads per multiprocessor
1c34100
@drroe
Report major compute capability.
e69eb39
@drroe
Add Gpu namespace to store Gpu related info
77ed890
@drroe
Try to set block size based on compute capability
87ec305
@drroe
Change recip to frac
04ef617
@drroe
Add rdf with ortho imaging
c5ec2e5
@drroe
Add version with no imaging
8edb99b
@drroe
Enable ortho and no image cases
0bca1c5
@drroe
Attempt to make functions usable with overlapping coords
a068892
@drroe
Add overlapping test
9a2b476
@drroe
Enable for overlapping coords.
d271e89
drroe added 26 commits July 8, 2022 19:30
@drroe
Hide some debug info
b550b43
@drroe
Add more realistic test for water-protein rdf
4265b9a
@drroe
Add more realistic water rdf test
7ba3b45
@drroe
Add orthogonal tests
107f6fe
@drroe
Add more realistic tests with no imaging
3d911d3
@drroe
Improve CPU radial function when single mask is being used
c4d42c8
@drroe
Always try to build libcpptraj.cuda
eb12ee7
@drroe
Start code to store indices for upper triangle matrix
f314034
@drroe
Doesnt seem like flattening works so well.
315119a
@drroe
Test out using device function template
6c96927
@drroe
Add template for non-orthogonal distance calc
6ca1e96
@drroe
Use the ortho dist template in the closest kernels
5424013
@drroe
Remove old code. Add note about slower ortho calc method
f11d1ce
@drroe
Start adding a gpu floating point type definition
8e1b734
@drroe
Use CpptrajGpu::FpType for GPU radial calcs.
163e1d4
@drroe
Change to floating point
14412a7
@drroe
Add tolerance for GPU radial calc, now done in single precision
c82d698
@drroe
Add kernel_rdf.cu to CMakeLists.txt
8a1bf0e
@drroe
Add a note about the rdf on GPU being done in single precision.
c203562
@drroe
Merge branch 'master' into rdfgpu
0f4fdaa
@drroe
6.11.0. Minor version bump for gpu version of RDF (single precision,
d772fa0 
  results will change from previous cpptraj.cuda).
@drroe
Make openmp single mask version loop dynamic
e30e559
@drroe
Move to separate test repo
0b5dcce
@drroe
Remove obsolete code
acbb714
@drroe
Hide thread info unless debug specified. Report total compute
86b9c06 
capability version.
@drroe
Fix test clean
989b9ab
@drroe drroe self-assigned this Jul 11, 2022
@drroe drroe merged commit ef8117f into Amber-MD:master Jul 11, 2022
@drroe drroe deleted the rdfgpu branch July 11, 2022 19:29
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees

@drroe drroe

Labels
enhancement
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@drroe