2013 INDIN v6.7-jb

8/14/2019 2013 INDIN v6.7-jb

http://slidepdf.com/reader/full/2013-indin-v67-jb 1/27

The MATISSE MATLAB Compiler 1

The MATISSE MATLAB CompilerA MATrix(MATLAB)-aware compiler InfraStructure for embedded computing SystEms

INDIN 2013Bochum, Germany – July 29-31

João Bispo, Pedro Pinto, Ricardo Nobre, Tiago Carvalho, João M. P. Cardoso, Pedro C. Diniz

{jbispo, pmsp, ricardo.nobre, tiago.diogo.carvalho, jmpc}@fe.up.pt, [email protected]

8/14/2019 2013 INDIN v6.7-jb



Outline

Motivation

MATISSE Architecture

LARA Language

Examples

Results

Conclusions and Future Work

8/14/2019 2013 INDIN v6.7-jb



Motivation – MATLAB Language

High-level matrix-oriented language, adequate for

prototyping, modeling and simulation

Fast learning curve

Rich set of libraries (toolboxes)

(De facto) Standard Prototype Development

Language for System Engineering

8/14/2019 2013 INDIN v6.7-jb



Motivation – MATLAB in

Embedded Systems

MATLAB is translated to C

MATLAB characteristics make compilation complexand more difficult to achieve efficient solutions

On-the-fly interpretation of code

Dynamic types and shapes

8/14/2019 2013 INDIN v6.7-jb



Motivation – Related Work FALCON [DeRose, Padua]

MATLAB to Fortran90

[Joisha, Banerjee]

Focus on type and shape inference

[Navak , Banerjee]

Annotations

C targeting HW MathWorks

Real-Time Workshop

MATLAB Coder

8/14/2019 2013 INDIN v6.7-jb



Motivation – MATISSE Approach

Objectives:

Generate optimized C code

Code transformation, e.g.: instrumentation and monitoring

General several versions from the same MATLAB code

Guide compilation using user/domain information

Provide fine-grained control through LARA aspects

Define type of any variable

Behavior of induction variable (e.g., monotonicallyincreasing)

Indicate if matrix shape does not change

Static vs Dynamic arrays

Specialize functions for certain input values (e.g., image size)

8/14/2019 2013 INDIN v6.7-jb

http://slidepdf.com/reader/full/2013-indin-v67-jb 7/27The MATISSE MATLAB Compiler 7

MATISSE Architecture

MATISSE:

MATLAB Parser

MATLAB Weaver

MATLAB to MATLAB Code Instrumentation

Type specialization

Transformationspropagate to C

MATLAB to C

MATLAB Parser

MATLAB Weaver

MIR

MATLAB Code

Generator

MATLAB Code LARA Aspects

MATLAB Code

MATISSE

MATLAB to C

C Code

CIR

Aspect Data

MIR

C Code Generator

8/14/2019 2013 INDIN v6.7-jb


LARA Language

Select-Apply-Condition block

Modularity (aspects can call other aspects)

Javascript for arbitrary computation

aspectdef <aspect_name>

select <point_of_interest> end

apply

<actions>

end

[condition <boolean_expression> end ]

end

8/14/2019 2013 INDIN v6.7-jb


Example – FIR Dynamic

function output=fir(input, coef)

NTAPS = length(coef);

N = length(input);

output = zeros(1, N);

for i = NTAPS:1:N

sum = 0.0;

for j = 0:1:NTAPS-1

sum = sum + input(i-j) *

coef (j+1);

end

output(i) = sum;

end

end

aspectdef firSingle

// Define array implementation select file end apply

def array_implementation = "dynamic"; end

var typeDef = {input: "single[]",coef: "single[]",sum: "single", output: "single" };

// Define types call defineTypes("fir", typeDef);

end

MATLAB Code LARA Code (Aspect)

8/14/2019 2013 INDIN v6.7-jb





N = length(input);


for i = NTAPS:1:N

sum = 0.0;

for j = 0:1:NTAPS-1


coef (j+1);

end

output(i) = sum;

end

end

aspectdef defineTypes

input functionName, typeDef end

select function.var end

apply

def type = typeDef[$var.name]; end

condition $var.name in typeDef &&

$function.name == functionName

end

end


8/14/2019 2013 INDIN v6.7-jb


Example – FIR Dynamictensor_f** fir (tensor_f* input, tensor_f* coef, tensor_f** output) {

int NTAPS; int N; int i; float sum; int j;

NTAPS = length_alloc_f(coef);

N = length_alloc_f(input);

zeros_f2(1, N, output);

for(i = NTAPS; i <=N; i = i+1) { sum = 0.0f;

for(j = 0; j <=(NTAPS-1); j = j+1) {

sum = sum+(input-> data[(i-j)-1]*coef-> data[(j+1)-1]);

}

(*output)-> data[i-1] = sum;

}

return output;

}

C Code

8/14/2019 2013 INDIN v6.7-jb





N = length(input);


for i = NTAPS:1:N

sum = 0.0;

for j = 0:1:NTAPS-1


coef (j+1);

end

output(i) = sum;

end

end

aspectdef firSingle

// Define array implementation select file end apply

def array_implementation = "static"; end

var typeDef = {input: "single[1][1024]",coef: "single[1][32]",sum: "single", output: "single" };

// Define types call defineTypes("fir", typeDef);

end


8/14/2019 2013 INDIN v6.7-jb


Example – FIR Staticfloat* fir(float* input, float* coef, float* output) {

int NTAPS; int N; int i; float sum; int j;

NTAPS = 32;

N = 1024;

zeros_f1x1024(output);

for(i = NTAPS; i <=N; i = i+1) { sum = 0.0f;

for(j = 0; j <=(NTAPS-1); j = j+1) {

sum = sum+(input[(i-j)-1]*coef[(j+1)-1]);

}

output[i-1] = sum;

}

return output;

}

C Code

8/14/2019 2013 INDIN v6.7-jb


Example – Grid Iteratefunction v_old = grid_iterate(obstacle, v, iter_max, nx, ny, nz)

v_0 = 0; v_end = 1; c = 1/6;

for iter = 1:iter_max

for i=2:nx-1

for j=2:ny-1 for k=2:nz-1

if(obstacle(i,j,k)==1)

v(i,j,k) = v_0;

elseif(obstacle(i,j,k)==-1) v(i,j,k) = v_end;

else

temp = v(i-1,j,k) + v(i+1,j,k) + v(i,j-1,k) +v(i,j+1,k) + v(i,j,k-1) + v(i,j,k+1);

v(i,j,k) = temp * c; end

endend end

end

v_old = v;

end

MATLAB Code

8/14/2019 2013 INDIN v6.7-jb



v_0 = 0; v_end = 1; c = 1/6;


for i=2:nx-1



v(i,j,k) = v_0;


else



endend end

end

v_old = v;

end

MATLAB Code

8/14/2019 2013 INDIN v6.7-jb


Example – Grid Iterate

temp = v-> data[((i-1)-1)+(j-1)*1*v-> shape[0]...

int v_shape0 = v-> shape[0];

...

temp = v-> data[((i-1)-1)+(j-1)*1*v_shape0...

Aspect: “shape of „v‟ does not change”

Performance increase: 8%

8/14/2019 2013 INDIN v6.7-jb


Experimental Setup

Set of kernels from embedded computing

GridIterate, FIR, FilterSubband and FFT2D

C Hand-coded and MATLAB versions

Two targets:

(PPC-O2) Cycle-acc. simulation, PowerPC 604,

gcc 4.3.3 (MB-O2) Cycle-acc. simulation, 3-stages MicroBlaze,

gcc 4.1.2

8/14/2019 2013 INDIN v6.7-jb


Results – C code performance

Performance of MATISSE competitive to hand-made code

Different versions of the code specified through aspects

0.95 0.99 0.98 0.97

1.16

1.111.07

1.32

0.50

0.75

1.00

1.25

gridIt fir fsubband fft2D

P e r f o r m a n c e ( m a

n u a l / M A T I S S E )

PPC-O2 MB-O2

8/14/2019 2013 INDIN v6.7-jb


Results – fsubband

MATISSE-generated code suited for HW implementation

1.36

1.02

0.991.00

1.04

1.001.001.02

1.06

1.09

1.30

1.33

0.95

1.00

1.05

1.10

1.15

1.20

1.25

1.30

1.35

1.40

single double

Manual / MATISSE

SW 0

SW 1

SW 2

HW 0

HW 1

HW 2

8/14/2019 2013 INDIN v6.7-jb


Conclusions

Aspect-oriented approach

Separation of code and additional information

Fine-grained control over generated code

MATLAB-to-MATLAB transformations

E.g., code monitoring and instrumentation

Generation of efficient C code from MATLAB

Performance competitive to manual code

Code prepared for hardware generation

8/14/2019 2013 INDIN v6.7-jb


Future Work

Support for fixed-point data types and morefunctions

Continue evaluation and experiments

Improve code optimization and specialization with

other kinds of information

8/14/2019 2013 INDIN v6.7-jb



THANK YOU!

Questions?

http://specs.fe.up.pt/tools/matisse/

Check MATISSE at the DEMO NIGHT!



8/14/2019 2013 INDIN v6.7-jb




v_0 = 0; v_end = 1; c = 1/6;


for i=2:nx-1



v(i,j,k) = v_0;


else



endend end

end

v_old = v;

end

MATLAB Code

8/14/2019 2013 INDIN v6.7-jb



Example – Grid Iterate

Run C code on a MicroBlaze

Single precision implementation

No HW support for double precision

C = 1/6 -> Inferred as a double

Statement alone is not enough to infer single

LARA aspect -> c: "single“

Performance increase: 6.3x

8/14/2019 2013 INDIN v6.7-jb



Example – FFT 2D

theta = isign*(2.0*(scalar_division_f(PI, temp_div))); ...

wp_x = wp_x*(((-2.0))*wp_x); wp_y = sinf(theta);

theta = isign*(2.0f*(scalar_division_f(PI, temp_div))); ...

wp_x = wp_x*(((-2.0f))*wp_x); wp_y = sinf(theta);

Automatically performed by MATISSE

Performance increase: 13%

8/14/2019 2013 INDIN v6.7-jb



Motivation – LARA Toolchain

Generation of C code

according to different

non-functional

requirements

LARA-based toolchain:

C Optimized for

Embedded Systems

HW/SW Partitioning

DSE

Profiling H a r d w a r e

/ S o f t w a r e F l o w

Application

(C)

C Front-End

Aspects and Design

Patterns (LARA)

VHDL-RTL

Back-End (code

generators)

Optimizer

(Software/Hardware)

Harmonic

Application

(C)

CDFG-IR

Kernels for Swand HwComponents (C) +

Annotations

Aspect-IR

Design-Space

Exploration

(DSE)

LARA Front-End

Best Practices

CDFG-IR

Hardware/Softwar

e Templates

CoSy

Assembly

weaving

8/14/2019 2013 INDIN v6.7-jb


Motivation However, it is cumbersome to apply certain

transformations directly in C code

Too many implementations decisions already in the

code

Higher level language for description of the

problem

Documents

2013 INDIN v6.7-jb