Feature request: ability to defined fixed spins about 3

[barnex]

Disabling torque in one or more regions does seem feasible. I might write it as an extension. Does
region := 2; ext_disableTorque(region); other_region_as_well := 3; ext_disableTorque(other_region_as_well)
look reasonable to you?

[godsic]

Would it be more reasonable to make 'gamma' region-able (if it is not
yet), so that setting gamma to 0 in some regions freezes corresponding
spins?

-Mykola

On Fri, 29 Aug, 2014 at 5:09 PM, Arne Vansteenkiste
[email protected] wrote:

Disabling torque in one or more regions does seem feasible. I might
write it as an extension. Does
region := 2
ext_disableTorque(region)
other_region_as_well := 3
ext_disableTorque(other_region_as_well)

look reasonable to you?

—
Reply to this email directly or view it on GitHub.

[xfong]

Making 'gamma' region-able should work too. But mumax3 should also recognize that if 'gamma' is zero for a particular cell, it should just return 'torque' as zero so that small numerical errors are not introduced.

[barnex]

0 * something is zero, regardless of numerical noise.
Anyway, gamma is used in quite a few places so making it
region-settable will require some coding and extensive testing of, at
least, Landau-Lifshitz, STT and temperature.

On Sat, Aug 30, 2014 at 3:11 PM, xfong [email protected] wrote:

Making 'gamma' region-able should work too. But mumax3 should also
recognize that if 'gamma' is zero for a particular cell, it should
just return 'torque' as zero so that small numerical errors are not
introduced.

—
Reply to this email directly or view it on GitHub.

[xfong]

As a first step, making gamma region-settable will be good. Later optimizations can be done to speed up the code so that computations are skipped for cells with gamma set as zero.

[barnex]

There's no such thing as speeding up by skipping cells on a GPU.
On Sep 6, 2014 7:48 PM, "xfong" [email protected] wrote:

As a first step, making gamma region-settable will be good. Later
optimizations can be done to speed up the code so that computations are
skipped for cells with gamma set as zero.

—
Reply to this email directly or view it on GitHub
#27 (comment).

[xfong]

Oh, I see. But does't that depend on the code? If the warp contains calculations in which "gamma" is zero, it should terminate early (which is what I meant by "skipping cells"). Since the fixed spins are likely to be local in memory, the threads that execute on the data in those memory locations are also likely to be i the same warp, no? So if the code is such that the function returns if gamma is detected to be zero, wouldn't those warps terminate sooner and result in faster execution overall?

[barnex]

In the special case that the fixed spins perfectly coincide with thread
blocks, the GPU would be doing few cycles less, yes. In the general
case it would do more due to divergent branching. In any case we are
bandwidth limited so that merely reading gamma would slow things down,
regardless of computation.

On Sat, Sep 6, 2014 at 10:03 PM, xfong [email protected] wrote:

Oh, I see. But does't that depend on the code? If the warp contains
calculations in which "gamma" is zero, it should terminate early
(which is what I meant by "skipping cells"). Since the fixed spins
are likely to be local in memory, the threads that execute on the
data in those memory locations are also likely to be i the same warp,
no? So if the code is such that the function returns if gamma is
detected to be zero, wouldn't those warps terminate sooner and result
in faster execution overall?

—

Reply to this email directly or view it on GitHub.

[godsic]

Kelvin, please check out 8946d26.

[barnex]

Available in 3.6.1