Use faster lud_diagonal

[Munksgaard]

This version of lud_diagonal uses intra-group parallelism and is faster
with the right tuning parameters and incremental flattening:

Before:

$ cat lud-clean.fut.tuning
main.suff_intra_par_14=32
main.suff_intra_par_16=1024
main.suff_intra_par_18=2000000000
main.suff_intra_par_20=32
main.suff_outer_par_13=2000000000
main.suff_outer_par_15=2000000000
main.suff_outer_par_17=2000000000
main.suff_outer_par_19=1024

$ FUTHARK_INCREMENTAL_FLATTENING=1 futhark bench --backend=opencl lud-clean.fut   
Compiling lud-clean.fut...
Results for lud-clean.fut (using lud-clean.fut.tuning):
dataset data/16by16.in:     292.60μs (avg. of 10 runs; RSD: 0.12)
dataset data/64.in:         517.90μs (avg. of 10 runs; RSD: 0.11)
dataset data/256.in:       2167.60μs (avg. of 10 runs; RSD: 0.14)
dataset data/512.in:       4181.10μs (avg. of 10 runs; RSD: 0.12)
dataset data/2048.in:     20156.60μs (avg. of 10 runs; RSD: 0.12)

After

$ cat lud-clean.fut.tuning 
main.suff_intra_par_14=32
main.suff_intra_par_16=1024
main.suff_intra_par_18=32
main.suff_intra_par_20=32
main.suff_intra_par_24=32
main.suff_intra_par_6=32
main.suff_outer_par_13=2000000000
main.suff_outer_par_15=2000000000
main.suff_outer_par_17=1568
main.suff_outer_par_19=1024

$ FUTHARK_INCREMENTAL_FLATTENING=1 futhark bench --backend=opencl lud-clean.fut   
Compiling lud-clean.fut...
Results for lud-clean.fut (using lud-clean.fut.tuning):
dataset data/16by16.in:     142.40μs (avg. of 10 runs; RSD: 0.10)
dataset data/64.in:         215.80μs (avg. of 10 runs; RSD: 0.03)
dataset data/256.in:       1260.80μs (avg. of 10 runs; RSD: 0.03)
dataset data/512.in:       2692.20μs (avg. of 10 runs; RSD: 0.04)
dataset data/2048.in:     19300.50μs (avg. of 10 runs; RSD: 0.09)

on gpu04-diku-apl.

[Munksgaard]

In comparison, lud.fut performs about similar to the old lud-clean.fut:

$ cat lud.fut.tuning 
main.suff_intra_par_19=1024
main.suff_intra_par_21=1024
main.suff_intra_par_23=32
main.suff_intra_par_25=32
main.suff_outer_par_18=2000000000
main.suff_outer_par_20=2000000000
main.suff_outer_par_22=7200
main.suff_outer_par_24=50176

$ FUTHARK_INCREMENTAL_FLATTENING=1 futhark bench --backend=opencl lud.fut
Compiling lud.fut...
Results for lud.fut (using lud.fut.tuning):
dataset data/16by16.in:     303.60μs (avg. of 10 runs; RSD: 0.09)
dataset data/64.in:         548.50μs (avg. of 10 runs; RSD: 0.10)
dataset data/256.in:       2297.50μs (avg. of 10 runs; RSD: 0.12)
dataset data/512.in:       6138.60μs (avg. of 10 runs; RSD: 0.05)
dataset data/2048.in:     24013.00μs (avg. of 10 runs; RSD: 0.12)

whereas the Rodinia implementation is a lot faster than ours:

$ ./lud -i ../../../data/lud/64.dat                                               
WG size of kernel = 16 X 16
Reading matrix from file ../../../data/lud/64.dat
num_devices = 1
Time consumed(microseconds): 82.000000
Time consumed(microseconds): 61.000000
Time consumed(microseconds): 59.000000
Time consumed(microseconds): 68.000000
Time consumed(microseconds): 58.000000
Time consumed(microseconds): 57.000000
Time consumed(microseconds): 57.000000
Time consumed(microseconds): 58.000000
Time consumed(microseconds): 58.000000
Time consumed(microseconds): 58.000000
Time consumed(microseconds): 67.000000

$ ./lud -i ../../../data/lud/256.dat                                              
WG size of kernel = 16 X 16
Reading matrix from file ../../../data/lud/256.dat
num_devices = 1
Time consumed(microseconds): 287.000000
Time consumed(microseconds): 293.000000
Time consumed(microseconds): 260.000000
Time consumed(microseconds): 264.000000
Time consumed(microseconds): 322.000000
Time consumed(microseconds): 266.000000
Time consumed(microseconds): 278.000000
Time consumed(microseconds): 299.000000
Time consumed(microseconds): 258.000000
Time consumed(microseconds): 274.000000
Time consumed(microseconds): 268.000000

$ ./lud -i ../../../data/lud/512.dat                                              
WG size of kernel = 16 X 16
Reading matrix from file ../../../data/lud/512.dat
num_devices = 1
Time consumed(microseconds): 543.000000
Time consumed(microseconds): 564.000000
Time consumed(microseconds): 653.000000
Time consumed(microseconds): 564.000000
Time consumed(microseconds): 536.000000
Time consumed(microseconds): 558.000000
Time consumed(microseconds): 538.000000
Time consumed(microseconds): 526.000000
Time consumed(microseconds): 544.000000
Time consumed(microseconds): 638.000000
Time consumed(microseconds): 537.000000

$ ./lud -i ../../../data/lud/2048.dat                                             
WG size of kernel = 16 X 16
Reading matrix from file ../../../data/lud/2048.dat
num_devices = 1
Time consumed(microseconds): 2237.000000
Time consumed(microseconds): 2176.000000
Time consumed(microseconds): 2167.000000
Time consumed(microseconds): 2185.000000
Time consumed(microseconds): 2175.000000
Time consumed(microseconds): 2155.000000
Time consumed(microseconds): 2172.000000
Time consumed(microseconds): 2158.000000
Time consumed(microseconds): 2155.000000
Time consumed(microseconds): 2157.000000
Time consumed(microseconds): 2155.000000

[athas]

What is the difference between lud.fut and lud-clean.fut? And if we really want a version that uses this kind of hack, should it then go in the "clean" version?

[Munksgaard]

I'm not sure there's much point in keeping both lud.fut and lud-clean.fut, as both seem to perform about the same (before this PR). I believe the original intent was to have a nice implementation that was easy (or easier) to understand and modify and one that ran fast, but that doesn't seem necessary any more.

[athas]

I'll merge this if you create one single lud.fut that contains your grand vision and get rid of lud-clean.fut.

[Munksgaard]

I'll merge this if you create one single lud.fut that contains your grand vision and get rid of lud-clean.fut.

I've pushed new commits that merge lud-clean and lud, and apply the changes to lud_diagonal that I've been working on.