Parallella Community

hi! i just use two dimensional array in my code,but i dont kown whether the e_read ,e_write can be used like this:

int a[M][N];
for(i=0;i<platform.rows;i++)
for(j=0;j<platform.cols;j++)
e_write(&dev,i,j,0x2000,&a,M*N*sizeof(int));
....


can i pass the array to epiphany platform correctly? Or,how to pass a two-dimensional array between host and device platform

Thanks

Hi !

From what i see, this will only write the content of you two-dimensional array into each eCore on the chip. I guess it could work since you specifically address eCores but you have to keep in mind that each eCore only can have a maximum amout of 32kB of data in his local memory so you have to be careful about the size of your array.

Regards,

Thomas.

Hi,Thomas!

you are right,i just want to write arrays to each ecore,,since i want to use the 16 ecores separately run the code,i use the e_load_group function to load programs into them.But i'm not quite sure if there is any difference between one-dimensional array
and two-dimension array,especially the address ,and how to read/write the two-dimension arrays correctly in device/host side.there isn't any examples about two-dimension arrays tranferring

Regards

Anthony

In C, there are no N-dimensional arrays by default, only one-dimensional arrays. However, if (N-1) dimension sizes are known at compile-time (so that the memory layout is known), then C provides some syntactic sugar. Each element has its own address, and each address is compatible to a void* (and vice versa).

In your example, you copy the whole array into each eCore. But you could also do an e_write of &A[i][j] (one element of the array), with a sizeof(A[0]) for the size, which would only copy one (unique) element to each eCore.

I think in this case, &a and &a[0][0] are equivalent: The very first element of a C array has the same address as the array itself. You might want to read up on how (multi-dimensional) C arrays are laid out in memory, because that is the magic you need to know here.

If you do e_write(&dev, i, j, 0x2000, &a[0][0], sizeof(a[0])), then you only copy the first element of your array into each core.
If you do e_write(&dev, i, j, 0x2000, &a, M*N*sizeof(a[0][0])), then you copy the whole array into each core (*)
If you do e_write(&dev, i, j, 0x2000, &a[i][j], sizeof(a[0][0])), then you copy a[0][0] to core (0,0), a[0][1] to core (0,1) and so on.

(*) I think that sizeof(a) should work as well, but I am not sure. I tend to avoid these constructs.

No idea. I always called e_load() in a loop. I don't think all 16 cores will start simultaneously, so you shouldn't count on it.

I think that if you defined a workgroup containing the 16 cores, striclty regarding the given API you should be able to run your program on all the cores in the same time by using e_load_group. Or rather, if the API is correct, why wouldn't this work ?