generate and share OpenCL local memory

[ronghongbo]

Hello,

Here is a skeleton of OpenCL code that I would like to generate from Halide. The requirement is to generate a local memory; all threads in the same work group fill the memory together, each filling one element. Finally, the work group size must be specified (by __attribute).

__kernel
__attribute((reqd_work_group_size(SIZE1, SIZE2,SIZE3)))
void f( __global float _A)
{
___local float A1[...];*
for (...)
{
copy 1 A element to A1;
barrier(CLK_LOCAL_MEM_FENCE); // wait for all elements of the block to be copied
do work on A1
copy back work to A
}
}

Any idea how to do so?
Thanks,
Hongbo

[zvookin]

Look at test/correctness/gpu_dynamic_shared.cpp for the part where
compute_at is used to get shared memory and the barrier. In order to copy
back to A, you will likely need to pass the same buffer as both an input
and an output, though it may be possible to do something with update stages
on the same buffer.

-Z-

On Mon, Jul 25, 2016 at 1:49 PM, ronghongbo [email protected]
wrote:

Hello,

Here is a skeleton of OpenCL code that I would like to generate from
Halide. The requirement is to generate a local memory; all threads in the
same work group fill the memory together, each filling one element.
Finally, the work group size must be specified (by __attribute).

__kernel
__attribute((reqd_work_group_size(SIZE1, SIZE2,SIZE3)))
void f( __global float

_A) { ___local float A1[...];*
for (...)
{
copy 1 A element to A1;
barrier(CLK_LOCAL_MEM_FENCE); // wait for all elements of the block to be
copied
do work on A1
copy back work to A
}
}

Any idea how to do so?
Thanks,
Hongbo

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#1406, or mute the thread
https://github.com/notifications/unsubscribe-auth/ABbqFDHK9A_sAmq9GBH07l2sgDiHECxzks5qZSFlgaJpZM4JUiDV
.

[ronghongbo]

Thanks, Zalman. That is useful info. Yes, the source and the code generated contain shared memory and barrier. It is a good starting point.

Hongbo

From: Zalman Stern [mailto:[email protected]]
Sent: Tuesday, July 26, 2016 10:50 AM
To: halide/Halide [email protected]
Cc: Rong, Hongbo [email protected]; Author [email protected]
Subject: Re: [halide/Halide] generate and share OpenCL local memory (#1406)

Look at test/correctness/gpu_dynamic_shared.cpp for the part where
compute_at is used to get shared memory and the barrier. In order to copy
back to A, you will likely need to pass the same buffer as both an input
and an output, though it may be possible to do something with update stages
on the same buffer.

-Z-

On Mon, Jul 25, 2016 at 1:49 PM, ronghongbo [email protected]
wrote:

Hello,

Here is a skeleton of OpenCL code that I would like to generate from
Halide. The requirement is to generate a local memory; all threads in the
same work group fill the memory together, each filling one element.
Finally, the work group size must be specified (by __attribute).

__kernel
__attribute((reqd_work_group_size(SIZE1, SIZE2,SIZE3)))
void f( __global float

_A) { ___local float A1[...];*
for (...)
{
copy 1 A element to A1;
barrier(CLK_LOCAL_MEM_FENCE); // wait for all elements of the block to be
copied
do work on A1
copy back work to A
}
}

Any idea how to do so?
Thanks,
Hongbo

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#1406, or mute the thread
https://github.com/notifications/unsubscribe-auth/ABbqFDHK9A_sAmq9GBH07l2sgDiHECxzks5qZSFlgaJpZM4JUiDV
.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHubhttps://github.com//issues/1406#issuecomment-235349012, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AJDLrUPLbfgPgdf2WQ7z5SJzxZ2M7r9gks5qZki-gaJpZM4JUiDV.