<s>
The	O
Sieve	B-Language
C++	I-Language
Parallel	I-Language
Programming	I-Language
System	I-Language
is	O
a	O
C++	B-Language
compiler	B-Language
and	O
parallel	O
runtime	O
designed	O
and	O
released	O
by	O
Codeplay	O
that	O
aims	O
to	O
simplify	O
the	O
parallelization	B-Operating_System
of	O
code	O
so	O
that	O
it	O
may	O
run	O
efficiently	O
on	O
multi-processor	O
or	O
multi-core	O
systems	O
.	O
</s>
<s>
It	O
is	O
an	O
alternative	O
to	O
other	O
well-known	O
parallelisation	B-Operating_System
methods	O
such	O
as	O
OpenMP	B-Application
,	O
the	O
RapidMind	O
Development	O
Platform	O
and	O
Threading	B-Application
Building	I-Application
Blocks	I-Application
(	O
TBB	O
)	O
.	O
</s>
<s>
Sieve	O
is	O
a	O
C++	B-Language
compiler	B-Language
that	O
will	O
take	O
a	O
section	O
of	O
serial	O
code	O
,	O
which	O
is	O
annotated	O
with	O
sieve	O
markers	O
,	O
and	O
parallelize	O
it	O
automatically	O
.	O
</s>
<s>
Delaying	O
side-effects	O
removes	O
many	O
small	O
dependencies	O
which	O
would	O
usually	O
impede	O
automatic	O
parallelization	B-Operating_System
.	O
</s>
<s>
Reads	O
and	O
writes	O
can	O
be	O
safely	O
reordered	O
by	O
the	O
compiler	B-Language
as	O
to	O
allow	O
better	O
use	O
of	O
various	O
data	O
movement	O
mechanisms	O
,	O
such	O
as	O
Direct	O
Memory	O
Access(DMA )	O
.	O
</s>
<s>
The	O
compiler	B-Language
can	O
then	O
split	O
up	O
code	O
within	O
the	O
sieve	O
block	O
much	O
easier	O
,	O
to	O
exploit	O
parallelism	B-Operating_System
.	O
</s>
<s>
Multi-core	O
CPUs	O
such	O
as	O
the	O
Cell	O
microprocessor	O
used	O
in	O
the	O
PlayStation	B-Operating_System
3	I-Operating_System
are	O
of	O
this	O
type	O
,	O
in	O
which	O
the	O
fast	O
cores	O
have	O
local	O
memories	O
that	O
must	O
be	O
utilized	O
to	O
exploit	O
performance	O
inherent	O
in	O
the	O
system	O
.	O
</s>
<s>
The	O
sieve	O
compiler	B-Language
can	O
split	O
code	O
within	O
a	O
sieve	O
block	O
into	O
chunks	O
either	O
implicitly	O
or	O
explicitly	O
though	O
a	O
'	O
splithere	O
 '	O
statement	O
.	O
</s>
<s>
The	O
compiler	B-Language
will	O
implicitly	O
add	O
a	O
splitpoint	O
above	O
the	O
for	O
loop	O
construct	O
body	O
,	O
as	O
an	O
entry	O
point	O
.	O
</s>
<s>
However	O
,	O
these	O
dependencies	O
must	O
not	O
cross	O
splitpoints	O
;	O
they	O
will	O
generate	O
compiler	B-Language
warnings	O
.	O
</s>
<s>
It	O
is	O
safe	O
for	O
parallelization	B-Operating_System
,	O
and	O
the	O
programmer	O
is	O
free	O
to	O
create	O
new	O
Iterator	O
classes	O
at	O
will	O
.	O
</s>
<s>
The	O
Sieve	O
Parallel	O
Runtime	O
employs	O
dynamic	O
speculative	B-General_Concept
execution	I-General_Concept
when	O
executing	O
on	O
a	O
target	O
platform	O
.	O
</s>
<s>
This	O
has	O
the	O
advantage	O
of	O
eliminating	O
race	B-Operating_System
conditions	I-Operating_System
,	O
one	O
of	O
the	O
most	O
common	O
bugs	O
in	O
concurrent	B-Architecture
programming	I-Architecture
.	O
</s>
<s>
The	O
removal	O
of	O
the	O
need	O
to	O
consider	O
concurrency	B-Operating_System
control	I-Operating_System
structures	O
within	O
a	O
sieve	O
block	O
can	O
speed	O
up	O
development	O
time	O
and	O
results	O
in	O
safer	O
code	O
.	O
</s>
<s>
The	O
system	O
is	O
designed	O
for	O
hierarchical	O
based	O
systems	O
with	O
homogeneous	O
or	O
heterogeneous	O
CPU	O
cores	O
which	O
have	O
local	O
memories	O
,	O
connected	O
via	O
DMA	B-General_Concept
engines	I-General_Concept
or	O
similar	O
memory	O
transfer	O
models	O
.	O
</s>
<s>
Sieve	O
has	O
been	O
shown	O
successfully	O
operating	O
on	O
multi-core	O
x86	O
systems	O
,	O
the	O
Ageia	B-Operating_System
PhysX	I-Operating_System
Physics	O
Processing	O
Unit	O
,	O
and	O
the	O
IBM	O
Cell	O
microprocessor	O
.	O
</s>
<s>
ANSI	O
C	O
is	O
generated	O
if	O
a	O
compiler	B-Language
code	B-Application
generator	I-Application
is	O
not	O
available	O
for	O
a	O
certain	O
target	O
platform	O
.	O
</s>
<s>
This	O
allows	O
for	O
autoparallelization	O
using	O
existing	O
C	O
compilation	B-Language
toolkits	O
.	O
</s>
