MATLAB: GpuArray of variable size for codegen

codegegpuarray

Hi,

I'm trying to figure out the actual potential of the codegen capabilities on the GPU. In particular there are two points that I are not addressed explicitly in the online documentation (or I couldn't find the info).

*) First, it seems that the gpuArray type cannot be used as input. This is surprising, as it reduces drastically the applicability of mex functions generated with codegen… Is there a way to create mex files that accept gpuArray vectors as rhs?

*) Second, codegen throws a warning when input arguments of matrices that will be sent to the GPU (by a coder.gpu.kernelfun pragma inside the .m file) are declared to be of variable size using coder.typeof(). Again, a significant restriction on the usability of codegen…. Am I missing something here? Is there a way to trick codegen into accepting dynamically sized arrays as input for GPU processing?

thanks! Daniel

Best Answer

Support for gpuArray input comes with R2018b. I don't know about the other thing, sorry.

Related Solutions

MATLAB: Specifying gpuArray input for codegen

Okay, I was able to fix this problem by getting gpucoder to produce the code that the app used to compile my function.

Install the GPU coder interface for deep learning libraries.
use the GPU Coder app to compile your code.
run

gpucoder -script yourscript.m -tocode yourgpucoderproject.prj

replacing the .m and .prj files with the names of you files.

This will output a script called yourscript.m with the code the app used. Now you can turn this script into a function that takes in the sizes of the arrays it should expect.

The critical difference between what I was doing and what the GPU Coder app does is it uses

coder.typeof(single(0), [N 1], 'Gpu', true)

to signal a gpuArray input. NOWHERE in the documentation is this syntax shown or explained.

MATLAB: When using GPU Coder, how should I try to minimize how often data is transferred between CPU AND GPU

GPU Coder tries to minimize copies between CPU and GPU. CPU/GPU copies purely depends on data access patterns.

https://www.mathworks.com/help/releases/R2020b/gpucoder/ug/gpu-memory-allocation-and-minimization.html#mw_f543c3ea-ff72-4316-94dd-0157526846f4

If you generate code for kernelfun1 and kernelfun2 separately (i.e. you call 'codegen' twice) and then you try to call the generated mex functions like kernelfun1(b) .* kernelfun2(c) and kernelfun1 or kernelfun2 try to return a 'half' data type, then there will be a transfer to the CPU in order to perform the multiplication. This is a current limitation of MATLAB because 'gpuArary' does not support the half data type. However, if you do the multiplication in a wrapper function, e.g.:

function a = kernelfun3(b,c)
    coder.gpu.kernelfun;
    a = kernelfun1(b) .* kernelfun2(c);
end

and only call 'codegen' on func3, then GPU Coder will generate code such that the multiplication is performed on the GPU.

The limitation above does not apply if the returned data type of kernelfun1 and kernelfun2 is 'single' or some other datatype supported by gpuArray. In that case, the following multiplication in will be performed on the GPU.

kernelfun1(b) .* kernelfun2(c)

CPU Coder tries to fuse the kernels as much as possible, so with the example above, you may find that the generated code contains a single GPU kernel instead of three separate ones for kernelfun1, kernelfun2 and kernelfun3. The effectiveness of this optimization depends on program structure and dataflow. However, we also noticed this optimization not happening in some cases. We recommend trying out code generation on your design and looking at the generated code to see whether the coder performed this optimization. If not, you can try altering your design to get the desired results.

Best Answer

Related Solutions

MATLAB: Specifying gpuArray input for codegen

MATLAB: When using GPU Coder, how should I try to minimize how often data is transferred between CPU AND GPU

Related Question