<s>
In	O
computing	O
,	O
a	O
compute	B-Operating_System
kernel	I-Operating_System
is	O
a	O
routine	O
compiled	O
for	O
high	O
throughput	O
accelerators	B-General_Concept
(	O
such	O
as	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	O
)	O
or	O
field-programmable	B-Architecture
gate	I-Architecture
arrays	I-Architecture
(	O
FPGAs	B-Architecture
)	O
)	O
,	O
separate	O
from	O
but	O
used	O
by	O
a	O
main	O
program	O
(	O
typically	O
running	O
on	O
a	O
central	B-General_Concept
processing	I-General_Concept
unit	I-General_Concept
)	O
.	O
</s>
<s>
They	O
are	O
sometimes	O
called	O
compute	B-Operating_System
shaders	I-Operating_System
,	O
sharing	O
execution	B-General_Concept
units	I-General_Concept
with	O
vertex	O
shaders	O
and	O
pixel	O
shaders	O
on	O
GPUs	B-Architecture
,	O
but	O
are	O
not	O
limited	O
to	O
execution	O
on	O
one	O
class	O
of	O
device	O
,	O
or	O
graphics	B-Library
APIs	I-Library
.	O
</s>
<s>
Compute	B-Operating_System
kernels	I-Operating_System
roughly	O
correspond	O
to	O
inner	O
loops	O
when	O
implementing	O
algorithms	O
in	O
traditional	O
languages	O
(	O
except	O
there	O
is	O
no	O
implied	O
sequential	O
operation	O
)	O
,	O
or	O
to	O
code	O
passed	O
to	O
internal	O
iterators	O
.	O
</s>
<s>
They	O
may	O
be	O
specified	O
by	O
a	O
separate	O
programming	O
language	O
such	O
as	O
"	O
OpenCL	B-Application
C	O
"	O
(	O
managed	O
by	O
the	O
OpenCL	B-Application
API	O
)	O
,	O
as	O
"	O
compute	B-Operating_System
shaders	I-Operating_System
"	O
written	O
in	O
a	O
shading	B-Language
language	I-Language
(	O
managed	O
by	O
a	O
graphics	B-Library
API	I-Library
such	O
as	O
OpenGL	B-Application
)	O
,	O
or	O
embedded	O
directly	O
in	O
application	O
code	O
written	O
in	O
a	O
high	B-Language
level	I-Language
language	I-Language
,	O
as	O
in	O
the	O
case	O
of	O
C++AMP	B-Language
.	O
</s>
<s>
This	O
programming	O
paradigm	O
maps	O
well	O
to	O
vector	B-Operating_System
processors	I-Operating_System
:	O
there	O
is	O
an	O
assumption	O
that	O
each	O
invocation	O
of	O
a	O
kernel	B-Algorithm
within	O
a	O
batch	O
is	O
independent	O
,	O
allowing	O
for	O
data	B-Operating_System
parallel	I-Operating_System
execution	O
.	O
</s>
<s>
However	O
,	O
atomic	B-General_Concept
operations	I-General_Concept
may	O
sometimes	O
be	O
used	O
for	O
synchronization	O
between	O
elements	O
(	O
for	O
interdependent	O
work	O
)	O
,	O
in	O
some	O
scenarios	O
.	O
</s>
<s>
Individual	O
invocations	O
are	O
given	O
indices	O
(	O
in	O
1	O
or	O
more	O
dimensions	O
)	O
from	O
which	O
arbitrary	O
addressing	O
of	O
buffer	O
data	O
may	O
be	O
performed	O
(	O
including	O
scatter	B-General_Concept
gather	I-General_Concept
operations	O
)	O
,	O
so	O
long	O
as	O
the	O
non-overlapping	O
assumption	O
is	O
respected	O
.	O
</s>
<s>
The	O
Vulkan	B-Application
API	I-Application
provides	O
the	O
intermediate	O
SPIR-V	B-Application
representation	O
to	O
describe	O
both	O
Graphical	O
Shaders	O
,	O
and	O
Compute	B-Operating_System
Kernels	I-Operating_System
,	O
in	O
a	O
language	B-Application
independent	I-Application
and	O
machine	B-Operating_System
independent	I-Operating_System
manner	O
.	O
</s>
<s>
The	O
intention	O
is	O
to	O
facilitate	O
language	O
evolution	O
and	O
provide	O
a	O
more	O
natural	O
ability	O
to	O
leverage	O
GPU	B-Architecture
compute	O
capabilities	O
,	O
in	O
line	O
with	O
hardware	O
developments	O
such	O
as	O
Unified	O
Memory	O
Architecture	O
and	O
Heterogeneous	B-Architecture
System	I-Architecture
Architecture	I-Architecture
.	O
</s>
<s>
This	O
allows	O
closer	O
cooperation	O
between	O
a	O
CPU	O
and	O
GPU	B-Architecture
.	O
</s>
