learnTemplates: error indexing into W using gpuArray/subsref on batch with Nfilt=0

[mspacek]

After commenting out the print statement that was triggered by a batch with blank raw data with no spikes that was causing the error in #231, I'm now hitting this error slightly later, which seems to be triggered by having Nfilt=0:

>> dbstop if error
>> rez = learnAndSolve8b(rez);
Time 2728s. Optimizing templates ...
Error using gpuArray/subsref
An unexpected error occurred trying to launch a kernel. The CUDA error was:
invalid configuration argument

Error in learnTemplates (line 203)
            W(:,Nfilt + [1:size(dWU0,3)],:) = W0(:,ones(1,size(dWU0,3)),:); % initialize temporal components of waveforms

Error in learnAndSolve8b (line 15)
    rez     = learnTemplates(rez, iorder0);
 
203             W(:,Nfilt + [1:size(dWU0,3)],:) = W0(:,ones(1,size(dWU0,3)),:); % initialize temporal components of waveforms
K>> Nfilt

Nfilt =

     0

K>> size(W)

ans =

    61     0     3

K>> W(:,Nfilt + [1:size(dWU0,3)],:)
Error using gpuArray/subsref
Index exceeds matrix dimension.

K>> dataRAW

dataRAW = <page truncated: showing [1:1455, 1:1] of 65600-by-31>

  65600×31 single gpuArray matrix

    0.0500
   -0.1450
    0.6850
    1.0800
    1.4750
    1.2950
    1.1850
    0.4400
    1.1900    
    ...

K>> nBatches

nBatches =

        9205

K>> ibatch

ibatch =

        3986            

If I print out all the contents of dat (same size as dataRAW), most of it is zeros for this batch, so this is almost certainly at least partially overlapping with the blank periods of data described in #231. Not sure how to proceed.

[JoseGuzman]

I'm trying to look into this (look in issues #227 and #218). I think it happens when units/batches are having zero spikes. In my case, it prompts an error here:
https://github.com/MouseLand/Kilosort2/blob/bdf9366bc73b4d9c101595c10193a532e33b54eb/mainLoop/learnTemplates.m#L149

>>> dbstop if error
>> master_kilosort
Looking for data inside ./ 
Time   0s. Computing whitening matrix.. 
Getting channel whitening matrix... 
Channel-whitening matrix computed. 
Time   1s. Loading raw data and applying filters... 
Time  34s. Finished preprocessing 801 batches. 
random seed for clusterSingleBatches: 1
Time  36s. Optimizing templates ...
One or more output arguments not assigned during call to "mexMPnu8".

Error in learnTemplates (line 149)
        mexMPnu8(Params, dataRAW, single(U), single(W), single(mu), iC-1, iW-1, UtU, iList-1, ...

Error in learnAndSolve8b (line 15)
    rez     = learnTemplates(rez, iorder0);

Error in master_kilosort (line 55)
rez = learnAndSolve8b(rez);
 
149         mexMPnu8(Params, dataRAW, single(U), single(W), single(mu), iC-1, iW-1, UtU, iList-1, ...

my Nfilt and W variables have the same values as yours, but the variable dWU returns a zero in the third dimension, which may be the origin of the problem (see #218):

K>> size(dWU)

ans =

    61    62     0

[mspacek]

I worked around this by no longer exporting raw data with longish periods of zero-padding between recordings. I simply directly concatenate recordings from the same session, then later work out the absolute spike times relative to the start of the first recording of the session.