================
@@ -1,119 +1,542 @@
-=============================================
-SYCL Compiler and Runtime architecture design
-=============================================
+============
+SYCL Support
+============
.. contents::
:local:
Introduction
============
+The `SYCL 2020 specification <SYCL-2020_>`_ defines a single-source programming
+model and C++ run-time library interface for writing portable programs that
+support heterogeneous devices including GPUs, CPUs, and accelerators.
+The specification is intended to allow for a wide range of implementation
+possibilities, examples of which include:
+
+- A SYCL run-time library written in standard C++ that executes kernels on a
+ homogeneous set of host and device processors, each of which can execute
+ common compiled code from shared memory.
+- A SYCL run-time library that executes kernels on a heterogeneous set of
+ device processors for which each kernel is pre-compiled for each supported
+ device processor (Ahead-Of-Time (AOT) compilation) or for a family of device
+ processors (Just-In-Time (JIT) compilation).
+
+Since Clang is a conforming implementation of the C++ standard, no additional
+features are required for support of the first implementation strategy.
+This document details the core language features Clang provides for use by
+SYCL run-time libraries that use the second implementation strategy.
+
+.. _SYCL-2020:
+ https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html
+
+
+Example Usage
+=============
+SYCL is designed as an extension of C++ rather than as a distinct programming
+language.
+SYCL support is enabled with the `-fsycl <opt-fsycl_>`_ option.
+
+.. code-block:: sh
+
+ clang++ -c -fsycl source-file.cpp
+
+The choice of which target devices will be supported is made at compile time.
----------------
tahonermann wrote:
I was unsure whether it was worthwhile to distinguish between concrete and
virtual device architectures and intermediate representations other than via
some vague hand waving about AOT and JIT. SPIR-V targets aren't completely
device agnostic since the LLVM IR is still constrained by ABI decisions (e.g.,
32-bit vs 64-bit).
What I was trying to convey is that the set of target devices that are eligible
for use at run-time is constrained by choices made at compile-time. I'll try to
clarify this. I'll look at adding some commentary about concrete vs virtual
architectures in the "Supported Targets" section too.
https://github.com/llvm/llvm-project/pull/170602
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
