methods and implementation @ hams
TRANSCRIPT
Politecnico di MilanoDipartimento di Elettronica, Informazione e Bioingegneria (DEIB)
HAMSproject
Chiara [email protected]
Guido [email protected]
May 26th, 2016
NECST Lab, Politecnico di Milano
2
Our goal
Study an oxygenator for ECC through a lumped
parameters model in MATLAB and accelerate the
simulation by means of an FPGA-based system
Lumped
parameters
model
3
Methods
MATLAB
functionsHigh level
Low level
Vivado HLS Vivado SDK
FPGAs
Our framework
4
Implementation
(SW-HW)
5
Implementation
Jacobi method: accuracy and parallelism
for (n_times)
split matrix into columns
compute c,s for each couple
update columns
end
// until convergence
// multiply&accumulate
// streaming computation
Ãi Ãj
Coefficient
computation
Column
updateAi Aj
c,s
6
Coefficient
computation
multiplyx x x x x x x x
+ + + +
+ +
+
Ai
Aj
set_directive_dataflow coefficientComputation
set_directive_pipeline coefficientComputation
set_directive_pipeline outerLoops
set_directive_unroll innerLoops
set_directive_interface –mode axis
accumulate
c,s
7
Coefficient
computation
multiplyx x x x x x x x
+ + + +
+ +
+
Ai
Aj
set_directive_dataflow coefficientComputation
set_directive_pipeline coefficientComputation
set_directive_pipeline outerLoops
set_directive_unroll innerLoops
set_directive_interface –mode axis
accumulate
performances
vs area
c,s
8
set_directive_dataflow columnUpdate
set_directive_dataflow outerLoop
set_directive_pipeline innerLoop
set_directive_interface -mode axis
Column
update
Ai
Aj
Ãi
Ãj
- - t e m p - -
x x x x x x x xc,s
9
Sparse matrix
The matrix is generated through the device meshing
(with known meshing pattern)
10
Sparse matrix
The matrix is generated through the device meshing
(with known meshing pattern)
The modeled matrix is sparse
(with known pattern)
We identify non-null elements
through MATLAB script
11
QUESTIONS?
Contact us!
HAMSproject
www.facebook.com/hams.project
https://twitter.com/HAMS_project
http://www.slideshare.net/HAMSproject
https://www.youtube.com/channel/UCaovqRpUc7D_Uf2WJHL0rvA