Hi,
If this works, that'd be a great speed up! For the t_mat matrix,
maybe you could name it as the evolution_matrix. That would make it
clear as to what it really is. Remember that long and descriptive
variable/class/function/method names are a feature of the relax source
code - with the aim of helping new developers and/or power users to be
able to read the code as text. See the "Long names" subsection at the
end of http://www.nmr-relax.com/manual/Variable_function_class_names.html.
Also, the following extensions of your current improvements would make
things a lot quicker:
1) Create the evolution_matrix data structure at the start before the
loop. Maybe with "evolution_matrix = R * 0.0" to create an empty
structure. This, together with the following points, will ensure that
only one data structure is created and deleted for the magnetisation
evolution for one target function call, rather than the 100+ created
and garbage collected at the moment.
2) Use the out argument:
dot(Rexpo, r180x, evolution_matrix)
dot(evolution_matrix, Rexpo, evolution_matrix)
3) The operation "t_mat =
Rexpo.dot(r180x).dot(Rexpo).dot(Rexpo).dot(r180x).dot(Rexpo).dot(Rexpo).dot(r180x).dot(Rexpo).dot(Rexpo).dot(r180x).dot(Rexpo)"
is 4x repetitive, so we can use that to minimise the operations.
Starting from the evolution matrix created in point 2), you just need
this 4 times, which you can do quickly with the following curious
looking two lines:
dot(evolution_matrix, evolution_matrix, evolution_matrix)
dot(evolution_matrix, evolution_matrix, evolution_matrix)
The first doubles the matrix block from Rexpo.r180x.Rexpo to
Rexpo.r180x.Rexpo.Rexpo.r180x.Rexpo, and the second line doubles that
again to
Rexpo.r180x.Rexpo.Rexpo.r180x.Rexpo.Rexpo.r180x.Rexpo.Rexpo.r180x.Rexpo.
By counting the dot products in the above steps, you can see that
there are only 4 compared to the current 11 in the disp_spin_speed
branch. And as the dot() call with out argument seems to be twice as
fast (https://gna.org/task/index.php?7807#comment199), this should
massively help your speed ambitions :)
Regards,
Edward
On 15 June 2014 08:53, <tlinnet@xxxxxxxxxxxxx> wrote:
Author: tlinnet
Date: Sun Jun 15 08:53:45 2014
New Revision: 23956
URL: http://svn.gna.org/viewcvs/relax?rev=23956&view=rev
Log:
Lowered the number of dot operations, by pre-preparing the evolution matrix
another round.
The power is in system tests always even.
The trick to removing this for loop, would be to make a general multi dot
function.
Task #7807 (https://gna.org/task/index.php?7807): Speed-up of dispersion
models for Clustered analysis.
Modified:
branches/disp_spin_speed/lib/dispersion/ns_cpmg_2site_3d.py
Modified: branches/disp_spin_speed/lib/dispersion/ns_cpmg_2site_3d.py
URL:
http://svn.gna.org/viewcvs/relax/branches/disp_spin_speed/lib/dispersion/ns_cpmg_2site_3d.py?rev=23956&r1=23955&r2=23956&view=diff
==============================================================================
--- branches/disp_spin_speed/lib/dispersion/ns_cpmg_2site_3d.py (original)
+++ branches/disp_spin_speed/lib/dispersion/ns_cpmg_2site_3d.py Sun Jun 15
08:53:45 2014
@@ -138,10 +138,10 @@
Rexpo = matrix_exponential(R*tcp[i])
# Temp matrix.
- t_mat =
Rexpo.dot(r180x).dot(Rexpo).dot(Rexpo).dot(r180x).dot(Rexpo)
+ t_mat =
Rexpo.dot(r180x).dot(Rexpo).dot(Rexpo).dot(r180x).dot(Rexpo).dot(Rexpo).dot(r180x).dot(Rexpo).dot(Rexpo).dot(r180x).dot(Rexpo)
# Loop over the CPMG elements, propagating the magnetisation.
- for j in range(power[i]):
+ for j in range(power[i]/2):
Mint = t_mat.dot(Mint)
# The next lines calculate the R2eff using a two-point
approximation, i.e. assuming that the decay is mono-exponential.
_______________________________________________
relax (http://www.nmr-relax.com)
This is the relax-commits mailing list
relax-commits@xxxxxxx
To unsubscribe from this list, get a password
reminder, or change your subscription options,
visit the list information page at
https://mail.gna.org/listinfo/relax-commits
Re: r23956 - /branches/disp_spin_speed/lib/dispersion/ns_cpmg_2site_3d.py