<s>
Heterogeneous	B-Architecture
System	I-Architecture
Architecture	I-Architecture
(	O
HSA	O
)	O
is	O
a	O
cross-vendor	O
set	O
of	O
specifications	O
that	O
allow	O
for	O
the	O
integration	O
of	O
central	B-General_Concept
processing	I-General_Concept
units	I-General_Concept
and	O
graphics	B-Architecture
processors	I-Architecture
on	O
the	O
same	O
bus	O
,	O
with	O
shared	B-Operating_System
memory	I-Operating_System
and	O
tasks	B-General_Concept
.	O
</s>
<s>
The	O
HSA	O
is	O
being	O
developed	O
by	O
the	O
HSA	B-Application
Foundation	I-Application
,	O
which	O
includes	O
(	O
among	O
many	O
others	O
)	O
AMD	O
and	O
ARM	O
.	O
</s>
<s>
The	O
platform	O
's	O
stated	O
aim	O
is	O
to	O
reduce	O
communication	B-Protocol
latency	I-Protocol
between	O
CPUs	O
,	O
GPUs	B-Architecture
and	O
other	O
compute	O
devices	O
,	O
and	O
make	O
these	O
various	O
devices	O
more	O
compatible	O
from	O
a	O
programmer	O
's	O
perspective	O
,	O
relieving	O
the	O
programmer	O
of	O
the	O
task	O
of	O
planning	O
the	O
moving	O
of	O
data	O
between	O
devices	O
 '	O
disjoint	O
memories	O
(	O
as	O
must	O
currently	O
be	O
done	O
with	O
OpenCL	B-Application
or	O
CUDA	B-Architecture
)	O
.	O
</s>
<s>
CUDA	B-Architecture
and	O
OpenCL	B-Application
as	O
well	O
as	O
most	O
other	O
fairly	O
advanced	O
programming	O
languages	O
can	O
use	O
HSA	O
to	O
increase	O
their	O
execution	O
performance	O
.	O
</s>
<s>
Heterogeneous	O
computing	O
is	O
widely	O
used	O
in	O
system-on-chip	B-General_Concept
devices	O
such	O
as	O
tablets	B-Device
,	O
smartphones	B-Application
,	O
other	O
mobile	O
devices	O
,	O
and	O
video	B-Device
game	I-Device
consoles	I-Device
.	O
</s>
<s>
HSA	O
allows	O
programs	O
to	O
use	O
the	O
graphics	B-Architecture
processor	I-Architecture
for	O
floating	B-Algorithm
point	I-Algorithm
calculations	O
without	O
separate	O
memory	O
or	O
scheduling	O
.	O
</s>
<s>
The	O
rationale	O
behind	O
HSA	O
is	O
to	O
ease	O
the	O
burden	O
on	O
programmers	O
when	O
offloading	O
calculations	O
to	O
the	O
GPU	B-Device
.	O
</s>
<s>
Originally	O
driven	O
solely	O
by	O
AMD	O
and	O
called	O
the	O
FSA	O
,	O
the	O
idea	O
was	O
extended	O
to	O
encompass	O
processing	O
units	O
other	O
than	O
GPUs	B-Architecture
,	O
such	O
as	O
other	O
manufacturers	O
 '	O
DSPs	B-Architecture
,	O
as	O
well	O
.	O
</s>
<s>
Modern	O
GPUs	B-Architecture
are	O
very	O
well	O
suited	O
to	O
perform	O
single	B-Device
instruction	I-Device
,	I-Device
multiple	I-Device
data	I-Device
(	O
SIMD	B-Device
)	O
and	O
single	B-General_Concept
instruction	I-General_Concept
,	I-General_Concept
multiple	I-General_Concept
threads	I-General_Concept
(	O
SIMT	O
)	O
,	O
while	O
modern	O
CPUs	O
are	O
still	O
being	O
optimized	O
for	O
branching	O
.	O
</s>
<s>
Originally	O
introduced	O
by	O
embedded	B-Architecture
systems	I-Architecture
such	O
as	O
the	O
Cell	O
Broadband	O
Engine	O
,	O
sharing	O
system	O
memory	O
directly	O
between	O
multiple	O
system	O
actors	O
makes	O
heterogeneous	O
computing	O
more	O
mainstream	O
.	O
</s>
<s>
Heterogeneous	O
computing	O
itself	O
refers	O
to	O
systems	O
that	O
contain	O
multiple	O
processing	O
units	O
central	B-General_Concept
processing	I-General_Concept
units	I-General_Concept
(	O
CPUs	O
)	O
,	O
graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
(	O
GPUs	B-Architecture
)	O
,	O
digital	B-Architecture
signal	I-Architecture
processors	I-Architecture
(	O
DSPs	B-Architecture
)	O
,	O
or	O
any	O
type	O
of	O
application-specific	O
integrated	O
circuits	O
(	O
ASICs	O
)	O
.	O
</s>
<s>
The	O
system	O
architecture	O
allows	O
any	O
accelerator	O
,	O
for	O
instance	O
a	O
graphics	B-Architecture
processor	I-Architecture
,	O
to	O
operate	O
at	O
the	O
same	O
processing	O
level	O
as	O
the	O
system	O
's	O
CPU	O
.	O
</s>
<s>
Among	O
its	O
main	O
features	O
,	O
HSA	O
defines	O
a	O
unified	O
virtual	B-General_Concept
address	I-General_Concept
space	I-General_Concept
for	O
compute	O
devices	O
:	O
where	O
GPUs	B-Architecture
traditionally	O
have	O
their	O
own	O
memory	O
,	O
separate	O
from	O
the	O
main	O
(	O
CPU	O
)	O
memory	O
,	O
HSA	O
requires	O
these	O
devices	O
to	O
share	O
page	B-General_Concept
tables	I-General_Concept
so	O
that	O
devices	O
can	O
exchange	O
data	O
by	O
sharing	O
pointers	O
.	O
</s>
<s>
This	O
is	O
to	O
be	O
supported	O
by	O
custom	O
memory	B-General_Concept
management	I-General_Concept
units	I-General_Concept
.	O
</s>
<s>
To	O
render	O
interoperability	O
possible	O
and	O
also	O
to	O
ease	O
various	O
aspects	O
of	O
programming	O
,	O
HSA	O
is	O
intended	O
to	O
be	O
ISA-agnostic	O
for	O
both	O
CPUs	O
and	O
accelerators	O
,	O
and	O
to	O
support	O
high-level	O
programming	O
languages	O
.	O
</s>
<s>
Java	B-Language
)	O
and	O
unmanaged	O
languages	O
(	O
e.g.	O
</s>
<s>
will	O
make	O
it	O
much	O
easier	O
to	O
develop	O
3rd-party	O
compilers	O
for	O
a	O
wide	O
range	O
of	O
heterogeneous	O
products	O
programmed	O
in	O
Fortran	B-Application
,	O
C++	O
,	O
C++	B-Language
AMP	I-Language
,	O
Java	B-Language
,	O
et	O
al	O
.	O
</s>
<s>
The	O
illustrations	O
below	O
compare	O
CPU-GPU	O
coordination	O
under	O
HSA	O
versus	O
under	O
traditional	O
architectures	O
.	O
</s>
<s>
Some	O
of	O
the	O
HSA-specific	O
features	O
implemented	O
in	O
the	O
hardware	O
need	O
to	O
be	O
supported	O
by	O
the	O
operating	B-Operating_System
system	I-Operating_System
kernel	I-Operating_System
and	O
specific	O
device	O
drivers	O
.	O
</s>
<s>
For	O
example	O
,	O
support	O
for	O
AMD	B-Device
Radeon	I-Device
and	O
AMD	B-Application
FirePro	I-Application
graphics	B-Device
cards	I-Device
,	O
and	O
APUs	B-Architecture
based	O
on	O
Graphics	B-Architecture
Core	I-Architecture
Next	I-Architecture
(	O
GCN	O
)	O
,	O
was	O
merged	O
into	O
version	O
3.19	O
of	O
the	O
Linux	O
kernel	B-Operating_System
mainline	O
,	O
released	O
on	O
8	O
February	O
2015	O
.	O
</s>
<s>
This	O
very	O
first	O
implementation	O
,	O
known	O
as	O
,	O
focuses	O
on	O
"	O
Kaveri	O
"	O
or	O
"	O
Berlin	O
"	O
APUs	B-Architecture
and	O
works	O
alongside	O
the	O
existing	O
Radeon	B-Device
kernel	B-Operating_System
graphics	O
driver	O
.	O
</s>
<s>
Additionally	O
,	O
supports	O
heterogeneous	O
queuing	O
(	O
HQ	O
)	O
,	O
which	O
aims	O
to	O
simplify	O
the	O
distribution	O
of	O
computational	O
jobs	O
among	O
multiple	O
CPUs	O
and	O
GPUs	B-Architecture
from	O
the	O
programmer	O
's	O
perspective	O
.	O
</s>
<s>
Support	O
for	O
heterogeneous	O
memory	O
management	O
(	O
HMM	O
)	O
,	O
suited	O
only	O
for	O
graphics	O
hardware	O
featuring	O
version	O
2	O
of	O
the	O
AMD	O
's	O
IOMMU	B-General_Concept
,	O
was	O
accepted	O
into	O
the	O
Linux	O
kernel	B-Operating_System
mainline	O
version	O
4.14	O
.	O
</s>
<s>
Integrated	O
support	O
for	O
HSA	O
platforms	O
has	O
been	O
announced	O
for	O
the	O
"	O
Sumatra	O
"	O
release	O
of	O
OpenJDK	B-Language
,	O
due	O
in	O
2015	O
.	O
</s>
<s>
AMD	B-Application
APP	I-Application
SDK	I-Application
is	O
AMD	O
's	O
proprietary	O
software	O
development	O
kit	O
targeting	O
parallel	O
computing	O
,	O
available	O
for	O
Microsoft	O
Windows	O
and	O
Linux	O
.	O
</s>
<s>
Bolt	O
is	O
a	O
C++	O
template	O
library	O
optimized	O
for	O
heterogeneous	O
computing	O
.	O
</s>
<s>
GPUOpen	B-Application
comprehends	O
a	O
couple	O
of	O
other	O
software	O
tools	O
related	O
to	O
HSA	O
.	O
</s>
<s>
CodeXL	B-Language
version	O
2.0	O
includes	O
an	O
HSA	O
profiler	O
.	O
</s>
<s>
,	O
only	O
AMD	O
's	O
"	O
Kaveri	O
"	O
A-series	O
APUs	B-Architecture
(	O
cf	O
.	O
</s>
<s>
"	O
Kaveri	O
"	O
desktop	O
processors	O
and	O
"	O
Kaveri	O
"	O
mobile	O
processors	O
)	O
and	O
Sony	O
's	O
PlayStation	B-Operating_System
4	I-Operating_System
allowed	O
the	O
integrated	O
GPU	B-Device
to	O
access	O
memory	O
via	O
version	O
2	O
of	O
the	O
AMD	O
's	O
IOMMU	B-General_Concept
.	O
</s>
<s>
Earlier	O
APUs	B-Architecture
(	O
Trinity	O
and	O
Richland	O
)	O
included	O
the	O
version	O
2	O
IOMMU	B-General_Concept
functionality	O
,	O
but	O
only	O
for	O
use	O
by	O
an	O
external	O
GPU	B-Device
connected	O
via	O
PCI	O
Express	O
.	O
</s>
<s>
Post-2015	O
Carrizo	O
and	O
Bristol	O
Ridge	O
APUs	B-Architecture
also	O
include	O
the	O
version	O
2	O
IOMMU	B-General_Concept
functionality	O
for	O
the	O
integrated	O
GPU	B-Device
.	O
</s>
