How to malloc a buffer on co-processor in offload mode?

[songlinhai]

Hi,

   How could I malloc a buffer on co-processor in offload mode? I only know that I can malloc a buffer on host, and use "offload in" to malloc the same buffer on co-processor. Is there some ways that I can malloc the buffer on co-processor directly?

   Thanks a lot!

[Bachar_Z_]

Hi,

You can use malloc directly inside the offload region!

if you want to use the buffer later on, you can use an output variable in the (size_t) and use it later as in variable for aother offload region.

you should do the cast!

[Kevin_D_Intel]

Perhaps Explicitly managed Heap-allocated Data discussed here provides the functionality you’re looking for.

[Bachar_Z_]

Kevin Davis (Intel) wrote:

Perhaps Explicitly managed Heap-allocated Data discussed here provides the functionality you’re looking for.

if we want to use the pointer p in more than one function, this way does not work!

[Kevin_D_Intel]

With the local pointer that is correct; however, with a global pointer as demonstrated below the method is usable with multiple functions. I also confirmed this with our Developers.

$ icc -V

Intel(R) C Intel(R) 64 Compiler XE for applications running on Intel(R) 64, Version 13.1.2.183 Build 20130514

$ icc example.c -o example && example

q = 77

[cpp]#include <stdio.h>

#include <offload.h>

__attribute__ ((target (mic))) int *p;

extern void f1();

extern void f2();

void f1()

{

  // The nocopy clause ensures CPU values pointed to by p

  // are not transferred to coprocessor

  #pragma offload target(mic:0) nocopy(p)

  {

    // Allocate dynamic memory for p on coprocessor

    p = (int *)malloc(sizeof (int) * 100);

    p[99] = 77;

  }

}

void f2()

{

int q = 0;

  // The nocopy clause ensures p is not altered by the offload process

  #pragma offload target(mic:0) nocopy(p) out(q)

  {

    // Reuse dynamic memory pointed to by p

    q = p[99]; // Will be 77

  }

  printf("q = %d\n",q);

}

int main()

{

f1();

f2();

}[/cpp]

[Bachar_Z_]

My case was the following:

I have an object with a pointer sould be in Phi mem, and I use that object in may functionality inside the system.

in my case, the global array( pointer) does not work because I have many of this object. " static does have the same problem"

And then I used the size_t variable as output for the alloc part ofthe object and I can use that value whenever I need to apply some functionality to any obect of that type!

[Corey_P_]

With reference to Kevin's code example above using nopy to reference a buffer that is created on the Mics. This method currently has a problem when using multiple Mics, though works fine for a single device. Read here for further discussion on that specific method:

http://software.intel.com/en-us/forums/topic/393649

Corey