Archives Discussions

[xmuben]

the testing  code is

#define  GROUP_SIZE 16
Attribute[GroupSize(GROUP_SIZE,1,1)]
kernel void kernelGPU(float input1 <>,float input2 <>, out float output<>
{
 int2 index = instance().xy;
 int a=index.x;
 int b=index.y;
 output= input1+input2;
}

How can I solve this problem?

What is the diffirence between instanceInGroup() and instance()?

If I set the groupsize to GROUP_SIZE and input Vector is set to <4*GROUP_SIZE>

Does Max(instanceInGroup().x) equal  Max(instance().x)?

[gaurav_garg]

Did you enable pre-processing using -pp flag?

When you define Attribute of a kernel with GROUP_SIZE, threads are divided into groups and these threads can share data among each other using shared memory.

InstanceInGroup() shows the thread ID within a group, instance() shows the global id.

[xmuben]

Originally posted by: gaurav.garg Did you enable pre-processing using -pp flag?

When you define Attribute of a kernel with GROUP_SIZE, threads are divided into groups and these threads can share data among each other using shared memory.

InstanceInGroup() shows the thread ID within a group, instance() shows the global id.

Thank you. Now Attribute works well.

Here I got another question about 2D array( width,height).  The input and output  arrays are width x height. I want to define width threads and each thread handles one column data of the arrays.

Just see the following code. 

 uint4 offset(0,0,0,0);
 uint4 index(width,1,1,1);
 kernelGPU.domainOffset(offset);
 kernelGPU.domainSize(index);

kernel void kernelGPU(float input1 [][],float input2 [][], out float output [][])
{
 int2 index = instance().xy;
 int a=index.x;
 int b=index.y;
 for (i=0;i {
     output=input1+input2;
     b=b+1;  
 }
}

the result is incorrect. could anyone tell me how to arrange the code?

thank you very much.

[gaurav_garg]

I think you want to do something like this-

output[ b ]=input1[ b ]+input2[ b ];

The first index shows the row number.