[SYSTEMML-1034] implemented gpu solve

[nakul02]

Implemented the GPU solve() function.
Ping @niketanpansare, @bertholdreinwald, @dusenberrymw

@iyounus - can you please try this out? (and also check for correctness?, I've checked on smaller data)

This will benefit us in some sense. I see it being used in these algorithms (based on a simple grep search):

• ALS-DS.dml
• CsplineDS.dml
• LinearRegDS.dml
• StepLinearRegDS

For me, I seem to get a 30x speedup in an example that I tried on my own machine (core i7 quad core, 32gb ram, GTX1070).

Program:

m = 12345
n = 4321

A = rand(rows=m, cols=n)
B = rand(rows=m, cols=1)

x = solve(A,B)
write(x, "xout")%                     

Output

➜  incubator-systemml git:(gpu_solve) ✗ bin/systemml solve.dml -gpu force -stats
================================================================================
Output dir: /home/njindal/git/incubator-systemml/temp
================================================================================
17/04/28 00:14:42 INFO api.DMLScript: BEGIN DML run 04/28/2017 00:14:42
17/04/28 00:14:42 WARN util.NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
17/04/28 00:14:42 INFO context.GPUContext: Initializing CUDA
17/04/28 00:14:43 INFO context.GPUContext:  GPU memory - Total: 8506.769408 MB, Available: 6889.2098559999995 MB on GPUContext{deviceNum=0}
17/04/28 00:14:43 INFO context.GPUContext: Total number of GPUs on the machine: 1
17/04/28 00:14:46 INFO api.DMLScript: SystemML Statistics:
Total elapsed time:		4.561 sec.
Total compilation time:		0.348 sec.
Total execution time:		4.213 sec.
Number of compiled MR Jobs:	0.
Number of executed MR Jobs:	0.
CUDA/CuLibraries init time:	0.660/0.514 sec.
Number of executed GPU inst:	1.
GPU mem tx time  (alloc/dealloc/set0/toDev/fromDev):	0.004/0.000/0.000/0.051/0.000 sec.
GPU mem tx count (alloc/dealloc/set0/toDev/fromDev/evict):	10/0/11/0/2/1/0.
GPU conversion time  (sparseConv/sp2dense/dense2sp):	0.000/0.000/0.000 sec.
GPU conversion count (sparseConv/sp2dense/dense2sp):	0/0/0.
Cache hits (Mem, WB, FS, HDFS):	2/0/0/0.
Cache writes (WB, FS, HDFS):	3/0/1.
Cache times (ACQr/m, RLS, EXP):	0.000/0.000/0.001/0.040 sec.
HOP DAGs recompiled (PRED, SB):	0/0.
HOP DAGs recompile time:	0.000 sec.
Total JIT compile time:		0.577 sec.
Total JVM GC count:		0.
Total JVM GC time:		0.0 sec.
Heavy hitter instructions (name, time, count):
-- 1) 	gpu_solve 	3.118 sec 	1	
-- 2) 	rand 	0.392 sec 	2	
-- 3) 	write 	0.040 sec 	1	
-- 4) 	createvar 	0.001 sec 	3	
-- 5) 	rmvar 	0.000 sec 	3	

17/04/28 00:14:46 INFO api.DMLScript: END DML run 04/28/2017 00:14:46
➜  incubator-systemml git:(gpu_solve) ✗ bin/systemml solve.dml  -stats       
================================================================================
Output dir: /home/njindal/git/incubator-systemml/temp
================================================================================
17/04/28 00:14:53 INFO api.DMLScript: BEGIN DML run 04/28/2017 00:14:53
17/04/28 00:14:53 WARN util.NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
17/04/28 00:17:06 INFO api.DMLScript: SystemML Statistics:
Total elapsed time:		132.744 sec.
Total compilation time:		0.345 sec.
Total execution time:		132.398 sec.
Number of compiled MR Jobs:	0.
Number of executed MR Jobs:	0.
Cache hits (Mem, WB, FS, HDFS):	2/0/0/0.
Cache writes (WB, FS, HDFS):	3/0/1.
Cache times (ACQr/m, RLS, EXP):	0.000/0.000/0.001/0.031 sec.
HOP DAGs recompiled (PRED, SB):	0/0.
HOP DAGs recompile time:	0.000 sec.
Total JIT compile time:		1.233 sec.
Total JVM GC count:		2.
Total JVM GC time:		0.448 sec.
Heavy hitter instructions (name, time, count):
-- 1) 	solve 	131.950 sec 	1	
-- 2) 	rand 	0.413 sec 	2	
-- 3) 	write 	0.031 sec 	1	
-- 4) 	createvar 	0.001 sec 	3	
-- 5) 	rmvar 	0.000 sec 	3	

17/04/28 00:17:06 INFO api.DMLScript: END DML run 04/28/2017 00:17:06

[akchinSTC]

Build failed, see build log for details

[akchinSTC]

Refer to this link for build results (access rights to CI server needed):
https://sparktc.ibmcloud.com/jenkins/job/SystemML-PullRequestBuilder/1411/

[mboehm7]

well, I could indeed imagine such as speedup as we're currently only calling out to commons-math because solve is by far not the bottleneck in ALS or LinregDS (only called for tiny matrices in the rank or number of features).

[nakul02]

@mboehm7 - understood, still this PR provides value. The more operations in a loop that are on the GPU, the lesser the ping pong of data between host and device memories.

[mboehm7]

sure - this is absolutely fine; I'm just setting the expectations straight: for example for LinregDS, it's called once and is even for 1k features in the sub-second range. However, down the road, once we have a distributed solve, there might be more algorithms that could benefit from it.

[nakul02]

distributing solve is a great idea. In fact, that is exactly what @iyounus is trying to do through DML using the single node builtin functions qr, cholesky and lu. I think #368 is some work towards that. There was an earlier version which used parfor to do the distribution.

[akchinSTC]

Refer to this link for build results (access rights to CI server needed):
https://sparktc.ibmcloud.com/jenkins/job/SystemML-PullRequestBuilder/1412/

[deroneriksson]

LGTM. If this is GPU, I feel @nakul02 and @niketanpansare are the owners of this area and need to merge and move forward for our 1.0.0 release.

[nakul02]

thanks @deroneriksson !

[iyounus]

I've checked the results from gpu solve and these are correct.

[niketanpansare]

LGTM, Thanks Nakul 👍

[nakul02]

Thanks, I shall merge.