triSYCL | Generic system-wide modern C for heterogeneous platforms | GPU library

The problem is that triSYCL is a research project looking deeper at some aspects of SYCL while not providing a global generic SYCL support for an end-user. I have just clarified this on the README of the project. :-( Probably the problem here is that the OpenCL SPIR kernel has not been generated. So you need to first compile the specific (old) Clang & LLVM from triSYCL https://github.com/triSYCL/triSYCL/blob/master/doc/architecture.rst#trisycl-architecture-for-accelerator. But unfortunately there is no simple Clang driver to use all the specific Clang & LLVM to generate the kernels from the SYCL source. Right know it is done with some ad-hoc awful Makefiles (look around https://github.com/triSYCL/triSYCL/blob/master/tests/Makefile#L360) and, even if you can survive to this, you might encounter some bugs...

The good news is now there are several other implementations of SYCL which are quite easier to use, quite more complete and quite less buggy! :-) Look at ComputeCpp, DPC++ and hipSYCL for example.

I'm Idiot! I don't even have to run CMake on triSYCL because it's a headers only library. I found that on this website in the top of the documentation section

So I can use this CMakeLists.txt file and it works perfectly:

Recently there were interface changes in the SYCL specification, and therefore you need to make sure your TensorFlow source code matches up to the ComputeCpp version.

Follow the guide here and ensure you have downloaded the latest version of ComputeCpp and have an up to date version of the source code from GitHub. Note that you need to use the Codeplay branch here because changes take some time to be propagated into the main TensorFlow repository, this is outlined in the guide.

The main reason for SYCL 1.2 to require OpenCL 1.2 is because the Khronos intermediate representation SPIR 1.2 requires it. Without SPIR, or any other intermediate representation, a SYCL implementation cannot compile C++ code into device binaries, and would need to convert C++ to OpenCL C, which is quite problematic.

First, a disclaimer about the fact that cl::sycl::pipe are experimental things from the provisional SYCL 2.2 specification and that runs only on CPU (no accelerator, no FPGA...), and only in a very inefficient way.

But of course it is useful to experiment about how a real design would work and how SYCL can work.

A pipe is similar to some hardware FIFO.

You wrote

And the answer is: This is not how it's done, and I still don't think it's possible.

Even my first assumption was wrong. If all you have is an ordinary C++ compiler, then any SYCL kernel can only be executed "in software", by the host device (CPU) running the "controller" code.

To translate the kernels to OpenCL (or SPIR-V) for execution on any other device, either an "augmented" compiler is necessary; or two compilers, one for the host, and one for the compute device.

A good explanation can be found here: https://www.codeplay.com/portal/introduction-to-sycl

The most related section is "What Would A SYCL Work Flow Look Like?"