gpu memory allocation and minimization -凯发k8网页登录

discrete and managed modes

gpu coder™ provides you access to two different memory allocation (malloc) modes available in the cuda^® programming model, cudamalloc and cudamallocmanaged. cudamalloc api is applicable to the traditionally separate cpu, and gpu global memories. cudamallocmanaged is applicable to unified memory.

from a programmer point of view, a traditional computer architecture requires that data be allocated and shared between the cpu and gpu memory spaces. the need for applications to manage data transfers between these two memory spaces adds to increased complexity. unified memory creates a pool of managed memory, shared between the cpu and the gpu. the managed memory is accessible to both the cpu and the gpu through a single pointer. unified memory attempts to optimize memory performance by migrating data to the device that needs it, at the same time hiding the migration details from the program. though unified memory simplifies the programming model, it requires device-sync calls when data written on the gpu is being accessed on the cpu. gpu coder inserts these synchronization calls. according to nvidia^®, unified memory can provide significant performance benefits when targeting embedded hardware like the nvidia tegra^®.

to change the memory allocation mode in the gpu coder app, use the malloc mode drop-down box under more settings->gpu coder. when using the command-line interface, use the mallocmode build configuration property and set it to either 'discrete' or 'unified'.

gpu memory manager

you can use the gpu memory manager for efficient memory allocation, management, and improving run-time performance. the gpu memory manager creates a collection of large gpu memory pools and manages allocation and deallocation of chunks of memory blocks within these pools. by creating large memory pools, the memory manager reduces the number of calls to the cuda memory apis, improving run-time performance. you can use the gpu memory manager for mex and standalone cuda code generation.

to enable the gpu memory manager, use one of these methods:

for cuda code that uses nvidia cuda libraries, such as cufft, cublas, and cusolver, you can enable the use of gpu memory manager for efficient memory allocation and management.

to use memory pools with cuda libraries, enable the memory manager using one the methods above and:

memory minimization

gpu coder analyzes the data dependency between cpu and gpu partitions and performs optimizations to minimize the number of cudamemcpy function calls in the generated code. the analysis also determines the minimum set of locations where data must be copied between cpu and gpu by using cudamemcpy.

for example, the function foo has sections of code that process data sequentially on the cpu and in parallel on the gpu.

function [out] = foo(input1,input2)
	   …
     % cpu work
			input1 = …
			input2 = …
			tmp1 = …
			tmp2 = …
   	…
     % gpu work
			kernel1(gpuinput1, gputmp1);
       kernel2(gpuinput2, gputmp1, gputmp2);
       kernel3(gputmp1, gputmp2, gpuout);
   	…
     % cpu work
       … = out
end

an unoptimized cuda implementation can potentially have multiple cudamemcpy function calls to transfer all inputs gpuinput1,gpuinput2, and the temporary results gputmp1,gputmp2 between kernel calls. because the intermediate results gputmp1,gputmp2 are not used outside the gpu, they can be stored within the gpu memory resulting in fewer cudamemcpy function calls. these optimizations improve overall performance of the generated code. the optimized implementation is:

gpuinput1 = input1;
gpuinput2 = input2;
kernel1<<< >>>(gpuinput1, gputmp1);
kernel2<<< >>>(gpuinput2, gputmp1, gputmp2);
kernel3<<< >>>(gputmp1, gputmp2, gpuout);
out = gpuout;

to eliminate redundant cudamemcpy calls, gpu coder analyzes all uses and definitions of a given variable and uses status flags to perform minimization. an example of the original code and what the generated code looks like is shown in this table.

a(:) = …
…
for i = 1:n
   gb = kernel1(ga);
   ga = kernel2(gb);
   if (somecondition)
      gc = kernel3(ga, gb);
   end
   …
end
…
… = c;

a(:) = …
a_isdirtyoncpu = true;
…
for i = 1:n
   if (a_isdirtyoncpu)
      ga = a;
      a_isdirtyoncpu = false;
   end
   gb = kernel1(ga);
   ga = kernel2(gb);
   if (somecondition)
      gc = kernel3(ga, gb);
      c_isdirtyongpu = true;
   end
   …
end
…
if (c_isdirtyongpu)
   c = gc;
   c_isdirtyongpu = false;
end
… = c;

the _isdirtyoncpu flag tells the gpu coder memory optimization about routines where the given variable is declared and used either on the cpu or on then gpu.