This is the list of SYCL specific options supported by compiler and some examples.
Options marked as [DEPRECATED] are going to be removed in some future updates. Options marked as [EXPERIMENTAL] are expected to be used only in limitted cases and not recommended to use in production environment.
General enabling option for SYCL compilation and linking mode. List of targets can be specified with `-fsycl-targets`. This is fundamental option for any SYCL compilation. All other SYCL specific options require it.
Enables ahead of time (AOT) compilation for specified device targets. T is a compiler target triple string, representing a target device architecture. You can specify more than one target, comma separated. Default just in time (JIT) compilation target can be added to the list to produce a combination of AOT and JIT code in the resulting fat binary. The following triples are supported by default: * spir64 - this is the default generic SPIR-V target; * spir64_x86_64 - generate code ahead of time for x86_64 CPUs; * spir64_fpga - generate code ahead of time for Intel FPGA; * spir64_gen - generate code ahead of time for Intel Processor Graphics; Full target triples can also be used: * spir64-unknown-unknown, spir64_x86_64-unknown-unknown, spir64_fpga-unknown-unknown, spir64_gen-unknown-unknown Available in special build configuration: * nvptx64-nvidia-cuda - generate code ahead of time for CUDA target;
SYCL language standard to compile for. Possible values: * 121 - SYCL 1.2.1 * 2020 - SYCL 2020 It doesn't guarantee specific standard compliance, but some selected compiler features change behavior. It is under development and not recommended to use in production environment. Default value is 2020.
Enables/Disables unnamed SYCL lambda kernels support. Disabled by default.
The option was used to enable/disable SYCL explicit SIMD extension. Not used anymore.
Enables (or disables) intermediate representation optimization pipeline before translation to SPIR-V. Have effect only if optimizations are turned on by standard compiler options (-O1 or higher). Enabled by default.
Enables (or disables) LLVM IR dead argument elimination pass to remove unused arguments for the kernel functions before translation to SPIR-V. Currently has effect only on spir64\* targets. Disabled by default.
Assume/Do not assume that SYCL ID queries fit within MAX_INT. It assumes that these values fit within MAX_INT: * id class get() member function and operator * item class get_id() member function and operator * nd_item class get_global_id()/get_global_linear_id() member functions Enabled by default.
Sets the inline threshold for device compilation to <n>. Note that this option only affects the behaviour of the DPC++ compiler, not target- specific compilers (e.g. OpenCL/Level Zero/Nvidia/AMD target compilers) which may or may not perform additional inlining. Default value is 225.
Target toolchain options¶
Pass "options" to the backend of target device compiler, specified by triple T. The backend of device compiler generates target machine code from intermediate representation. This option can be used to tune code generation for a specific target. The "options" are used during AOT compilation. For JIT compilation "options" are saved in a fat binary and used when code is JITed during runtime. -Xs is a shortcut to pass "options" to all backends specified via the '-fsycl-targets' option (or default one).
Pass "options" to the frontend of target device compiler, specified by triple T. This option can be used to control of intermediate representation generation during offline or online compilation.
Pass "options" to the device code linker, when linking multiple device object modules. T is specific target device triple.
Intel FPGA specific options¶
Perform ahead of time compilation for Intel FPGA. It sets the target to FPGA and turns on the debug options that are needed to generate FPGA reports. It is functionally equivalent shortcut to `-fsycl-targets=spir64_fpga -g -MMD` on Linux and `-fsycl-targets=spir64_fpga -Zi -MMD` on Windows.
Controls FPGA target binary output format. Same as -fsycl-link, but optional output can be one of the following: * early - generate html reports and an intermediate object file that avoids a full Quartus compile. Usually takes minutes to generate. Link can later be resumed from this point using -fsycl-link=image. * image - generate a bitstream which is ready to be linked and used on a FPGA board. Usually takes hours to generate.
Speed up FPGA backend compilation if the device code in <binary> is unchanged. If it's safe to do so the compiler will re-use the device binary embedded within it. This can be used to minimize or avoid long Quartus compile times for FPGA targets when the device code is unchanged.
Compile only device part of the code and ignore host part.
Emit SYCL device code in LLVM-IR bitcode format. When disabled, SPIR-V is emitted. Enabled by default.
Emit help information from device compiler backend. Backend can be one of the following: "x86_64", "fpga", "gen", or "all". Specifying "all" is the same as specifying -fsycl-help with no argument and emits help for all backends.
Informs the compiler driver that the host compilation step that is performed as part of the greater compilation flow will be performed by the compiler <arg>. It is expected that <arg> is the compiler to be called, either by name (in which the PATH will be used to discover it) or a fully qualified directory with compiler to invoke. This option is only useful when -fsycl is provided on the command line.
Passes along the space separated quoted "opts" string as option arguments to the compiler specified with the -fsycl-host-compiler=<arg> option. It is expected that the options used here are compatible with the compiler specified via -fsycl-host-compiler=<arg>. NOTE: Using -fsycl-host-compiler-options to pass any kind of phase limiting options (e.g. -c, -E, -S) may interfere with the expected output set during the host compilation. Doing so is considered undefined behavior.
Example: SYCL device code compilation¶
To invoke SYCL device compiler set
$ clang++ -fsycl-device-only sycl-app.cpp -o sycl-app.bc
By default the output format for SYCL device is LLVM bytecode.
-fno-sycl-use-bitcode can be used to emit device code in SPIR-V format.
$ clang++ -fsycl-device-only -fno-sycl-use-bitcode sycl-app.cpp -o sycl-app.spv