The Spatial Use Case

This page explains how to run HyperMapper with the Spatial language (https://github.com/stanford-ppl/spatial-lang). Here Spatial is the program that calls HyperMapper. HyperMapper replies with the points that have to be run by Spatial. Both software are writing on stdio and stdout and following a precise protocol of communication, like explained here. This is what in HyperMapper terminology is called "interactive mode". The developer responsible for Spatial is David Koeplinger ([email protected]).

This page supposes that we are working on the Stanford machine "london" and using the login "lnardi".

ssh [email protected]

Install Spatial and checkout the dse branch:

git clone https://github.com/stanford-ppl/spatial-lang.git
cd spatial-lang
git checkout dse
git submodule update --init
make apps

This runs Spatial on the benchmark BlackScholes, DotProduct and OuterProduct, using a heuristic based on pruning and random sampling design space exploration (DSE) which is the technique used to do DSE before HyperMapper. So this is the so called "previous approach". Previous approach runs 100000 random samples. The commands will generate 10 trials of data with timestamp, the pruning is counted in the timestamp calculation so the clock starts at the time Spatial starts the pruning.

./bin/spatial BlackScholes --experiment
./bin/spatial DotProduct --experiment
./bin/spatial OuterProduct --experiment

The data generated is in:

./results

Set HyperMapper environment variable and then run Spatial with HyperMapper:

export HYPERMAPPER_HOME=/home/lnardi/hypermapper
./bin/spatial DotProduct --hypermapper

This launches HyperMapper via the script hypermapper.py from $HYPERMAPPER/hypermapper.py.

Using exhaustive search to compare HyperMapper against the actual Pareto

For small Spatial benchmarks it is possible to generate the ground-truth exhaustive search.

./bin/spatial BlackScholes --bruteforce
./bin/spatial OuterProduct --bruteforce
./bin/spatial DotProduct --bruteforce

After running this commands a csv file for each benchmark is generated that will contain the exhaustive search. Now you can modify the json file to use the exhaustive mode like explained here.

Benchmarks available in Spatial

App	Total Space	Pruned Space	Description
BlackScholes	76,800	672	Financial option pricing
DotProduct			Dot product
OuterProduct		31068	Outer product
GDA	30,049,566,720	4,630,528	Gaussian Discriminant Analysis
MatMult_outer	262,144,000	143,724	GEMM
Kmeans	2,080,800	18,720	k-means
SW	2,097,152	32,768	Smith-Waterman DNA sequence alignment
TPCHQ6	3,538,944,000	2,265,600	Filter reduce database query

Additional comments

Experiments on the Stanford machines.

Machine	Max ALMs	Type of machine
Tucson	524800	TBD
London	262400	TBD

Note on the name of the parameters and their meaning for the DotProduct benchmark:

Parameter name	What it is	In/Out	More explanation
B	tile-sizeX_DP	input	TBD
x439	meta-pipe_DP	input	TBD
P1	outer-par_DP	input	TBD
P2	inner-par_DP	input	TBD
P3	burst-par_DP	input	TBD
DSPs	DSPs	output	TBD
Valid	Valid	output	TBD
ALMs	ALMs	output	TBD
BRAM	BRAMs	output	TBD
Cycles	Cycles	output	TBD
Channels	Channels	output	TBD
Reg	Regs	output	TBD
Timestamp	Timestamp	output	TBD

And for the BlackScholes benchmark:

Parameter name	What it is	In/Out	More explanation
B	tile-sizeX_BS	input	TBD
OP	outer-par_BS	input	TBD
IP	inner-par_BS	input	TBD
x1873	meta-pipe_BS	input	TBD
DSPs	DSPs	output	TBD
Valid	Valid	output	TBD
ALMs	ALMs	output	TBD
BRAM	BRAMs	output	TBD
Cycles	Cycles	output	TBD
Timestamp	Timestamp	output	TBD