how to generate C for host and OpenCL for device?

[ronghongbo]

Hello,

I would like to generate OpenCL code for a device (from a parallel loop in a loop nest), and generate C code for the host (from the rest of the loop nest). By using "compile_jit", I can see the OpenCL code generated, and the host code is automatically executed. How can I dump out the host code as C so that I may play with the C and OpenCL code manually (e.g. modify and compile them in a command shell)?

thanks!
hongbo

[zvookin]

You will need to use the "compile_to_c" method to generate C code. (I believe this works with GPU backends now, but it may not be very solid. This is a little used path.)

…

-Z-

On Tue, Mar 6, 2018 at 5:04 PM Hongbo Rong ***@***.***> wrote: Hello, I would like to generate OpenCL code for a device (from a parallel loop in a loop nest), and generate C code for the host (from the rest of the loop nest). By using "compile_jit", I can see the OpenCL code generated, and the host code is automatically executed. How can I dump out the host code as C so that I may play with the C and OpenCL code manually (e.g. modify and compile them in a command shell)? thanks! hongbo — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2797>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABbqFDRY-O7YmzzIe_jUZNibFqPRNFGTks5tbzItgaJpZM4Sfsf1> .

[ronghongbo]

Thanks! I tried that, it generates C code, but the host and device code are not separated. The code is like this:
halide_copy_to_device(...)
#pragma omp parallel for
for (....) {
// the code that should be a device kernel
}
halide_device_free(...)
I hope the above bold lines can be separated out as a device kernel, and the activation of the kernel is explicitly shown.

[zvookin]

Halide is not designed to generate separate kernels for GPUs that run independently of the Halide generated host code (in whatever system the host code is running).

…

-Z-

On Wed, Mar 7, 2018 at 9:19 AM, Hongbo Rong ***@***.***> wrote: Thanks! I tried that, it generates C code, but the host and device code are not separated. The code is like this: halide_copy_to_device(...) *#pragma omp parallel for for (....) { // the code that should be a device kernel }* halide_device_free(...) I hope the above bold lines can be separated out as a device kernel, and the activation of the kernel is explicitly shown. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#2797 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABbqFGgDDlEf3rKpEHyvYFju1QeqEeutks5tcBaxgaJpZM4Sfsf1> .