Direct OpenCL clBuildProgram() support?

[goofyseeker311]

Does Intel implementation support full OpenCL 1.2+ specification including in-OpenCL "online" compilation of OpenCL C kernel code?
Or does Intel OpenCL always require external kernel compilation outside of OpenCL host application (for example LWJGL OpenCL API)?

Assuming there is the document section about running Intel OpenCL without the Intel OpenCL SDK suggest that Intel FPGA Runtime for OpenCL supports full online compilation of text OpenCL C code to FPGA kernels compilation. ie same full "online" kernel deploy method works for all of Intel CPUs, GPUs and FPGAs, without modification or external separate compile of kernels.

Can you verify this?

So far most of any documentation has pointed towards full off-line compilation of kernels only.

Edit: Can you provide a full working full online compile deploy/compile runtime environment install image for DE10-Nano (Cyclone V and other FPGA SoC families), for example sake. OpenCL API only without any external program compile of kernels. so that any kernel compile happens internally in the OpenCL device drivers.

[sophimao]

Hello, Intel FPGA runtime for OpenCL does not officially support online compilation of OpenCL C kernel code, if you query for device information CL_DEVICE_COMPILER_AVAILABLE, the OpenCL FPGA runtime will return false. Note that this is only the OpenCL runtime for Intel FPGA, it is possible that Intel GPU and CPU supports online kernel compilation.

[goofyseeker311]

If so, then its obvious why this much faster development route is not advertised or talked anywhere.

Terasic support said tho that the icd path including the buildkernel() and enqueuendrangekernel() would work with the compiled icd linux kernel drivers, for example for de10 nano. I did not recompile the kernel icd driver support so cant say anything about that. any test image did not include the icd drivers (similar to cpu/gpu os linux drivers). kernel recompile to get the icd drivers fully working is kinda tedious.

Terasic said that they have not tested the enqueuendrangekernel() but they have run tested the buildkernel().

Typically an opencl driver also provides the compilation and workload run support. even if the driver would use an external program (intel fpga compiler) to get its final executable form but run through the icd driver fully, including the running of the compiled programs.

edit: if intel fgpa opencl icd does not current actually support the full opencl compile-run path-way, then can you add this as implemented and default supported in any test operating system images, so that you can say that your device fully supports opencl 3.0. I think both sockit and de10-nano advertise full opencl 3.0 compatibility, which lacking online compile and run would not actually be opencl 3.0. just the offline compile mode of opencl 3.0.

this means the driver (open source) or the off-line compiler acting as the driver compiler would also fully compile opencl programs into fpga executable. supporting opencl 3.0 typically means the cpu/gpu way of opencl direct run capability, not only the offline code compile to fpga executable.

[goofyseeker311]

I have a java lwjgl opencl benchmarking program that needs to run directly on the fpga without any change (similar as it already runs on intel/any cpu/gpu). ie drivers must be the only things installed in the system, no other modifications to add an online opencl device.

edit: why do a runtime for opencl if its not used. ie used to compile and run opencl code to fpga executable.

[sophimao]

I'm not entirely sure if I fully understood your comments, but I'll try my best to respond.

if intel fgpa opencl icd does not current actually support the full opencl compile-run path-way, then can you add this as implemented and default supported in any test operating system images, so that you can say that your device fully supports opencl 3.0. I think both sockit and de10-nano advertise full opencl 3.0 compatibility, which lacking online compile and run would not actually be opencl 3.0. just the offline compile mode of opencl 3.0.

I believe the FPGA runtime for OpenCL currently implements OpenCL 1.2, we would like to support OpenCL 3.0 but due to other higher priority tasks this won't be implemented in the near future. It is possible that a device is conformant to OpenCL 3.0 while the runtime is not.

why do a runtime for opencl if its not used. ie used to compile and run opencl code to fpga executable.

Assuming the device is programmed with the device image (i.e., the output from the offline compiler), the runtime is used for running the OpenCL host code which interacts with the device.