k-Wave

1. huangchao
   

   Member
   Joined: Nov '10
   Posts: 57

   Hi, Dr. Cox and Dr. Treeby,

   I have some questions about the following codes in the 'kspaceFirstOrder2D' function:

   ux_sgx = bsxfun(@times, pml_x_sgx, ...
   bsxfun(@times, pml_x_sgx, ux_sgx) ...
   - dt./ rho0_sgx .* real(ifft2( bsxfun(@times, ddx_k_shift_pos, kappa .* p_k) )) ...
   );
   uy_sgy = bsxfun(@times, pml_y_sgy, ...
   bsxfun(@times, pml_y_sgy, uy_sgy) ...
   - dt./ rho0_sgy .* real(ifft2( bsxfun(@times, ddy_k_shift_pos, kappa .* p_k) )) ...
   );

   % force bsxfun compatability with Accelereyes GPU toolbox
   if strncmp(data_cast, 'g', 1);
   geval(ux_sgx, uy_sgy);
   end

   I am confused about the use of the 'geval' function. Since 'ux_sgx' and 'uy_sgy' have been computed in the above lines, what's the purpose to use 'geval'. I don't understand the comments 'force bsxfun compatability with Accelereyes GPU toolbox'.

   Also, do 'ux_sgx' and 'uy_sgy' comupted in the first 2 statements reside in the CPU memory, while 'ux_sgx' and 'uy_sgy' comupted by 'geval' reside in GPU memroy?

   Thanks,
   Chao

   Posted 12 years ago #

2. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Hi Chao,

   Accelereyes Jacket uses a concept known as lazy evaluation in which the lines of code are not necessarily executed immediately in the order they are written. Instead, it tries to calculate when the most efficient time to calculate these values might be. If you use the variable again in your code and the previous line of code hasn't yet executed, it is calculated then to make sure your computations are correct.

   At least in earlier versions of Jacket, this wasn't working correctly when multiple calls to bsxfun are used. The geval command forces Jacket to make sure the values for ux_sgx (etc) are up to date, rather than waiting for the lazy evaluation to catch up.

   When 'DataCast' is set to 'gsingle' or 'gdouble', the values of all the loop variables reside in GPU memory.

   I hope that helps,

   Brad.

   Posted 12 years ago #

3. pavanky
   

   Member
   Joined: Mar '12
   Posts: 1

   Hi Bradley,

   Just to clarify, the problem was with accuracy. for the following snippet of code

   C = A*B;
   E = C + D;

   The compiler merges the multiply and add into a single operation E = madd(A, B, D); which is less accurate (but much faster) than one seperate multiply and add.

   If you do geval(C); before doing E = C+D, the multiply and add are performed separately and you get the most accurate results.

   So if you are not using multiply / add combination the geval is not necessary.

   Posted 12 years ago #

4. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Hi Pavanky,

   Thanks for your comment. Actually, the problem isn't with accuracy in the sense you mention. Rather, when we switched to using bsxfun in the B.0.4 release of k-Wave, the simulations run on the GPU using Jacket were completely wrong. I tracked this down to a bug in Jacket when using multiple calls to bsxfun (see here for details). Adding the calls to geval fixed the problem. This may have been fixed in later releases of Jacket (I haven't tested this). In any case, we'll leave the geval calls in for the time being in case anyone is using older versions of the Jacket toolbox.

   Kind regards,

   Brad.

   Posted 12 years ago #

5. huangchao
   

   Member
   Joined: Nov '10
   Posts: 57

   Hi,

   I just tested the codes without using gevals calls on my machine, and it gave the same results as the one using geval, so I guess the bsxfun bugs had been fixed in the Jacket 2.1

   Posted 12 years ago #

6. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Thanks Chao, that's useful to know!

   Posted 12 years ago #

7. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Hi Chao,

   Interestingly enough, although the code runs using Jacket 2.1 without the geval calls, taking them out seems to make the code more than 30% slower. Following your post on the Accelereyes forums, I guess it's related to Jacket getting confused as to when it should evaluate the expressions. In the case of k-Wave, the blocks of code between the geval calls contain lots of operations (FFTs, matrix multiplies, etc), so forcing execution still gives the GPU cores plenty of work to do.

   Brad.

   Posted 12 years ago #

RSS feed for this topic

Reply

You must log in to post.