ptxas error : Entry function '_Z15explore_subsetsP3curS0_PbPiS2_PfP17curandStateXORWOW' uses too m

Accelerated Computing CUDA CUDA Programming and Performance
1

I do not know why this happens.

compile : nvcc -o scatter.cu -w
output : ptxas error : Entry function ‘_Z15explore_subsetsP3curS0_PbPiS2_PfP17curandStateXORWOW’ uses too much local data (0x6db8 bytes, 0x4000 max)

code :

#define PROBLEM_SIZE 1000;

typedef struct cur{
float vector[PROBLEM_SIZE];
float cost;
float dist;
bool ref_new;
}ref_set_str;

i use the memcpy.

memcpy(best, &candidate, sizeof(ref_set_str));

If you set the size of the problem_size on the 600 error.

ptxas error : Entry function ‘_Z15explore_subsetsP3curS0_PbPiS2_PfP17curandStateXORWOW’ uses too much local data (0x6db8 bytes, 0x4000 max)

I do not know why that error.

please help me.

2

Probably you haven’t shown enough code to work things out. GPUs are limited in the amount of local memory per thread they can allocate.

Apparently you have a kernel like this:

explore_subsets(cur*, cur*, bool*, int*, int*, float*, curandStateXORWOW*)

That kernel is using too much local data. The limit on local data usage varies by compute capability:

[url]Programming Guide :: CUDA Toolkit Documentation

Your limit is 0x4000 i.e. 16KB per thread, so I guess you are compiling for a cc1.x device.

If you are actually using a cc2.x or newer device, you may be able to work around this error just by compiling for the proper compute capability. If you are actually using a cc1.x device, then you will need to reduce the local memory usage per thread.

An example of local memory usage would be:

float vector[PROBLEM_SIZE];

if that declaration occurred in ordinary thread code (i.e. not in global memory or shared memory). So each ref_set_str that you declare in thread code will use at least 4Kbytes (and the limit is 16KB for cc1.x).