Cannot modify a value in Page locked memory

Accelerated Computing CUDA CUDA Programming and Performance
1

[Moving the thread to the right section: CUDA Programming and Development ]

link

I have a problem.

h_value remains 0 … why?

Could anyone help me please?

Code:

[codebox]

void my_function(void){

int *h_value = NULL;

int *d_value = NULL;

dim3 block(1,1);

dim3 grid(1,1);

cudaSetDeviceFlags(cudaDeviceMapHost);

cudaHostAlloc((void**) &h_value, sizeof( int ) , cudaHostAllocMapped | cudaHostAllocPortable);

cudaHostGetDevicePointer( &d_value , h_value , 0 );

    //init value on host

*h_value=0;

my_kernel<<<grid,block>>>( d_value );

printf("h_value = %d \n",*h_value);

}

[/codebox]

[codebox]global void ky_kernel( int *value )

{

*value=1;

}

[/codebox]

Tnk’s.

2

  1. Run the deviceQuery example to see whether your gpu supports host page-locked memory mapping.

  2. See section 3.2.6.3 of the programming guide:

N.

3

Many tnk’s

added a cudaThreadSynchronize(); after the kernel and go!

4

As Nico said was a problem of synchronization and inserting a cudaSynchronizeThread(); after the kernel execution permits to the host to read the correct value.

But now I’ve a new problem:

[codebox]

void my_function(void){

int *h_value = NULL;

int *d_value = NULL;

    int size=1024;

dim3 block(1,1);

dim3 grid(1,1);

cudaMalloc((void**)&d_input_data, size); //< This line makes kernel increment fail!

cudaSetDeviceFlags(cudaDeviceMapHost);

cudaHostAlloc((void**) &h_value, sizeof( int ) , cudaHostAllocMapped | cudaHostAllocPortable);

cudaHostGetDevicePointer( &d_value , h_value , 0 );

    //init value on host

*h_value=0;

my_kernel<<<grid,block>>>( d_input_data , d_value );

printf("h_value = %d \n",*h_value);

}

[/codebox]

[codebox]global void ky_kernel( char* device_array, int *value )

{

*value=1;

}

[/codebox]

Simply adding a cudaMalloc() and a parameter to kernel the mapped variable cannot be modified (!!!) and h_values remains 0.

Some hint?