在OpenCl中模拟动态内存分配

Question

I ran into a problem which is making me crazy. I need to simulate dynamic memory allocation in OpenCl kernel. In this regard, I have the following malloc function defined in a *.cl file:

 __global void* malloc(size_t size, __global byte *heap, __global uint *next)
{
  uint index = atomic_add(next, size);
  return heap+index;
}

In the host program, I dynamically dedicate a large array of type cl_uchar for this virtual heap as follows:

int MAX_NUM_OF_HEADERS_PROCESSED_IN_PARALLEL = 1000;
cl_uchar* heap = new cl_byte[1000000];
cl_uint  *next  =  new cl_uint;
*next = 0;
cl_uint * test_result =
        new cl_uint[MAX_NUM_OF_HEADERS_PROCESSED_IN_PARALLEL];
cl_mem memory[3]= { 0, 0, 0};
cl_int error;

memory[0] = clCreateBuffer(GPU_context,
CL_MEM_READ_WRITE, sizeof(cl_uchar) * MAX_HEAP_SIZE, NULL,
NULL);

memory[1] = clCreateBuffer(GPU_context, CL_MEM_READ_WRITE, sizeof(cl_uint), NULL,
        &error);

memory[2] = clCreateBuffer(GPU_context, CL_MEM_READ_WRITE,
            sizeof(cl_uint) * MAX_NUM_OF_HEADERS_PROCESSED_IN_PARALLEL, NULL,
            &error);
clEnqueueWriteBuffer(command_queue, memory[0], CL_TRUE, 0,
        sizeof(cl_uchar) * MAX_HEAP_SIZE, heap, 0, NULL, NULL);

clEnqueueWriteBuffer(command_queue, memory[1], CL_TRUE, 0, sizeof(cl_uint),
        next, 0, NULL, NULL);
error = 0;
error |= clSetKernelArg(kernel, 0, sizeof(cl_mem), &memory[0]);
error |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &memory[1]);
error |= clSetKernelArg(kernel, 2, sizeof(cl_mem), &memory[2]);

size_t globalWorkSize[1] = { MAX_NUM_OF_HEADERS_PROCESSED_IN_PARALLEL };
size_t localWorkSize[1] = { 1 };


error = 0;
error = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL,
        globalWorkSize, localWorkSize, 0, NULL, NULL);

I also have the following kernel:

__kernel void packet_routing2(__global byte* heap_, __global uint* next, __global uint* test_result){

    int gid = get_global_id(0);

    __global uint*xx[100];

    for ( int i = 0 ; i < 100; i ++)
    {
        xx[i] = (__global uint*) malloc(sizeof(uint),heap_,next);
        *xx[i] = i*gid;

    result[gid] = *(xx[0]);
}   

I encounterd the following error when I run the program:

" %27 = load i32 addrspace(1)* %26, align 4, !tbaa !17
Illegal pointer which is not from a valid memory space.
Aborting..."

Could you please help me fix this issue. I also found out that if xx has only 10 elements, instead of 100, the code works well !!!!

Answer 1

Edit: Simplest solution: add a padding value to 'size' before malloc so all struct types (that are lesser in size than max-padding) receive necessary alignment conditions.

0=struct footprint in memory

*=heap

_=padding

***000_____*****0000____****0_______****00000___*****0000000_*******00______***
      |
      v
 save this unused padded memory space in its thread to use later.

it is important that first/starting address value needs to satisfy maximum alignment requirements. If there is a struct 256-byte long, it should have multiple of 256 to start.

struct size      malloc size    minimum 'next' value (address, not offset)
  1-4                 4            multiple of 4
  5-8                 8            multiple of 8
  9-16                16           multiple of 16
  17-32               32            32*k
  33-64               64            64*k

if there is 64-byte struct, even an int needs 64-byte malloc size now. Maybe you can save that values locally per thread to use that remaining unused areas.

So it doesnt give alignment errors and probably works faster for those don't.

Also float3 needs 16 byte natively.