ii
CUDA C Programming Guide Version 4.1
Changes from Version 4.0
Documented the
cudaAddressModeBorder
and
cudaAddressModeMirror
Updated Section3.2.10.2.1to mention the change in how surface references aredeclared.
Added Sections3.2.10.2.3, 3.2.10.2.4,B.9.5,B.9.6, B.9.7, B.9.8, B.9.9,B.9.10,
B.9.11, B.9.12, B.9.13,andB.9.14on three-dimensional, cubemap, layered, and
cubemap layered surfaces.
Mentioned the new function
cudaGLGetDevices()
Mentioned in SectionB.8.1that
tex1Dfetch()
only works with non-normalized coordinates, so only the border and clamp addressing modes aresupported.
Shortened the description of the driver API and moved it toAppendix G.
Specified in SectionB.12that the warp vote functions operate on the activethreads of the warp.
Added new SectionB.14on assertions.
Deleted fromAppendix Cthe descriptions of a few functions that have beenmoved to the reference manual.
Added mathematical functions
erfcxf()
and
erfcx()
erfcx()
is the scaled complementary error function:
erfcx(x) = exp(x*x) * erfc(x)
.
Added the Bessel functions
j0()
,
j1()
,
jn()
,
y0()
,
y1()
, and
yn()
j0f()
,
j1f()
,
jnf()
,
y0f()
,
y1f()
, and
ynf()
Removed Section D.2.4.5 since
__global__
functions can now be defined within an anonymous namespace.
Added toTable F-2the maximum sizes for cubemap textures, cubemap layeredtextures, three-dimensional surfaces, cubemap surfaces, and cubemap layeredsurfaces.
Updated the maximum width and height of a texture bound to linear memory in Table F-2.
Updated the maximum number of instructions per kernel for devices of compute capability 2.x from 2 million to 512 million inTable F-2.