Hi Troels, Again a great change for speed! I didn't see that these operations could occur in the target function initialisation - that should make things faster. Just running a test with a single iteration and the 'NS R1rho 2-site' model between two different revisions of your branch, I see:
""" $ ./disp_profile_all.py /data/relax/branches/disp_spin_speed2 /data/relax/branches/disp_spin_speed [snip] New relax version: relax repository checkout r24287 svn+ssh://[email protected]/svn/relax/branches/disp_spin_speed Old relax version: relax repository checkout r24274 svn+ssh://[email protected]/svn/relax/branches/disp_spin_speed Execution iteration 1 $ python profiling_ns_r1rho_2site.py /data/relax/branches/disp_spin_speed2 1000 0.018 0.000 5.730 0.006 relax_disp.py:1575(func_ns_r1rho_2site) 10 0.017 0.002 5.358 0.536 relax_disp.py:1575(func_ns_r1rho_2site) $ python profiling_ns_r1rho_2site.py /data/relax/branches/disp_spin_speed 1000 0.018 0.000 6.333 0.006 relax_disp.py:1552(func_ns_r1rho_2site) 10 0.017 0.002 5.834 0.583 relax_disp.py:1552(func_ns_r1rho_2site) 60 20 [snip] 100 single spins analysis: NS R1rho 2-site: 63.330+/-0.000 -> 57.300+/-0.000, 1.105x faster. Cluster of 100 spins analysis: NS R1rho 2-site: 58.340+/-0.000 -> 53.580+/-0.000, 1.089x faster. """ This is significant, but there is an obvious bottleneck preventing this useful change from making the model much faster. If this bottleneck is solved, I think this change will have a much higher and immediate impact as it does save a total of 5-6 seconds which, when you compare this to the total speed of the other models, is a lot. We can call this a latent speed up that will be revealed at a later date :) If the dot product for the magnetisation evolution is shifted out of the loops (using tensordot() or einsum(), which I guess you're looking at now anyway), then all the loops will disappear and the bottleneck might be removed. I also have some some small points about this change: - Maybe add blank lines before each comment. - The state A parameters all have A in capitals in the code so, for consistency and better readability, da_mat might be better as dA_mat. - Check the spacing around '=' signs. Cheers, Edward On 24 June 2014 16:35, <[email protected]> wrote: > Author: tlinnet > Date: Tue Jun 24 16:35:01 2014 > New Revision: 24286 > > URL: http://svn.gna.org/viewcvs/relax?rev=24286&view=rev > Log: > Speeded up ns r1rho 2site, by preforming the evolution matrices, and the M0 > matrix in the init part of the target 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_r1rho_2site.py > branches/disp_spin_speed/target_functions/relax_disp.py > > Modified: branches/disp_spin_speed/lib/dispersion/ns_r1rho_2site.py > URL: > http://svn.gna.org/viewcvs/relax/branches/disp_spin_speed/lib/dispersion/ns_r1rho_2site.py?rev=24286&r1=24285&r2=24286&view=diff > ============================================================================== > --- branches/disp_spin_speed/lib/dispersion/ns_r1rho_2site.py (original) > +++ branches/disp_spin_speed/lib/dispersion/ns_r1rho_2site.py Tue Jun 24 > 16:35:01 2014 > @@ -50,8 +50,7 @@ > """ > > # Python module imports. > -from math import atan2 > -from numpy import array, cos, dot, float64, log, multiply, sin, sum > +from numpy import array, cos, dot, einsum, float64, log, multiply, sin, sum > > # relax module imports. > from lib.float import isNaN > @@ -236,6 +235,9 @@ > # This matrix is a propagator that will evolve the magnetization with > the matrix R. > Rexpo_mat = matrix_exponential_rank_NE_NS_NM_NO_ND_x_x(R_mat) > > + # Magnetization evolution. > + Rexpo_M0_mat = einsum('...ij,...jk', Rexpo_mat, M0) > + > # Loop over spins. > for si in range(NS): > # Loop over the spectrometer frequencies. > @@ -245,23 +247,19 @@ > # Extract number of points. > num_points_i = num_points[0, si, mi, oi] > > - # Repetitive calculations (to speed up calculations). > - # Offset of spin-lock from A. > - dA = omega[0, si, mi, oi, 0] - offset[0, si, mi, oi, 0] > - > # Loop over the time points, back calculating the R2eff > values. > for j in range(num_points_i): > - # The following lines rotate the magnetization previous > to spin-lock into the weff frame. > - theta = atan2(spin_lock_fields[0, si, mi, oi, j], dA) > - M0[0] = sin(theta) # The A state initial X > magnetisation. > - M0[2] = cos(theta) # The A state initial Z > magnetisation. > + # Extract the preformed matrix that rotate the > magnetization previous to spin-lock into the weff frame. > + M0_i= M0[0, si, mi, oi, j, :, 0] > > # This matrix is a propagator that will evolve the > magnetization with the matrix R. > Rexpo_i = Rexpo_mat[0, si, mi, oi, j] > > + # Extract from the pre-formed Magnetization evolution > matrix. > + Rexpo_M0_mat_i = Rexpo_M0_mat[0, si, mi, oi, j, :, 0] > + > # Magnetization evolution. > - MA = dot(Rexpo_i, M0) > - MA = dot(M0, MA) > + MA = dot(M0_i, Rexpo_M0_mat_i) > > # The next lines calculate the R1rho using a two-point > approximation, i.e. assuming that the decay is mono-exponential. > if MA <= 0.0 or isNaN(MA): > > Modified: branches/disp_spin_speed/target_functions/relax_disp.py > URL: > http://svn.gna.org/viewcvs/relax/branches/disp_spin_speed/target_functions/relax_disp.py?rev=24286&r1=24285&r2=24286&view=diff > ============================================================================== > --- branches/disp_spin_speed/target_functions/relax_disp.py (original) > +++ branches/disp_spin_speed/target_functions/relax_disp.py Tue Jun 24 > 16:35:01 2014 > @@ -27,7 +27,7 @@ > # Python module imports. > from copy import deepcopy > from math import pi > -from numpy import add, array, asarray, complex64, dot, float64, int16, max, > multiply, ones, sqrt, sum, tile, zeros > +from numpy import add, array, arctan2, asarray, cos, complex64, dot, > float64, int16, max, multiply, ones, sin, sqrt, sum, tile, zeros > from numpy.ma import masked_equal > > # relax module imports. > @@ -404,7 +404,17 @@ > self.M0 = zeros(7, float64) > self.M0[0] = 0.5 > if model in [MODEL_NS_R1RHO_2SITE]: > - self.M0 = zeros(6, float64) > + # Offset of spin-lock from A. > + da_mat = self.chemical_shifts - self.offset > + # The following lines rotate the magnetization previous to > spin-lock into the weff frame. > + theta_mat = arctan2(self.spin_lock_omega1, da_mat) > + M0_0 = zeros([6, 1], float64) > + M0_0[0, 0] = 1 > + M0_sin = multiply.outer( sin(theta_mat), M0_0 ) > + M0_2 = zeros([6, 1], float64) > + M0_2[2, 0] = 1 > + M0_cos = multiply.outer( cos(theta_mat), M0_2 ) > + self.M0 = M0_sin + M0_cos > if model in [MODEL_NS_R1RHO_3SITE, MODEL_NS_R1RHO_3SITE_LINEAR]: > self.M0 = zeros(9, float64) > > > > _______________________________________________ > relax (http://www.nmr-relax.com) > > This is the relax-commits mailing list > [email protected] > > 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 _______________________________________________ relax (http://www.nmr-relax.com) This is the relax-devel mailing list [email protected] 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-devel
