The ll_nvptx Plugin

Table of Contents

• Dependencies
• Command-Line Arguments
• Phases
  • %ll_nvptx.emit

See also

mim::plug::ll_nvptx

An LLVM backend that targets NVIDIA GPUs. Loading this plugin (mim foo.mim -p ll_nvptx) appends an emission phase to the compilation pipeline that splits the fully lowered program into a host world and a device world. It writes the host LLVM IR to <world>.ll and the device LLVM IR (with PTX intrinsics) to <world>_dev.ll, and — on Linux by default — compiles and embeds the device binary into the host LLVM IR. The output paths and the embedding behavior can be overridden via -X ll_nvptx:<arg>; see Command-Line Arguments.

Dependencies

plugin gpu;
plugin nvptx;
plugin opt;
import compile;

Command-Line Arguments

This plugin reads the following plugin arguments (-X ll_nvptx:<arg>):

Argument	Effect
o=<file> output=<file>	Write the (host) LLVM IR to <file> instead of the default <world>.ll/a.ll.
o-dev=<file> output-dev=<file>	Write the device LLVM IR to <file> instead of the default <world>_dev.ll/a_dev.ll.
rt=embed rt=extern	How the C runtime wrappers (e.g. @mim_cu_check) reach the host module: embed (default) splices their LLVM IR in; extern only declares them.
no-embed	Don't embed the compiled device binary into the host LLVM IR (embedding is the default on Linux).
no-ptx-embed	If embedding the device binary, don't embed the PTX image into the fat binary (default: embed both PTX and CUBIN).
no-cubin-embed	If embedding the device binary, don't embed the CUBIN image into the fat binary (default: embed both PTX and CUBIN).
sm=<SM>	If embedding the device binary, compile it for compute capability sm_<SM>.
libdevice=<path>	If embedding the device binary and linking libdevice, link against the libdevice NVVM library at <path> instead of locating it via CUDA paths.
Xlink_llvm=<args>	If embedding the device binary and linking libdevice, invoke link_llvm with <args> (default: none).
Xopt=<args>	If embedding the device binary and linking libdevice, invoke opt with <args> (default: -passes="default<O2>,nvvm-reflect").
Xllc=<args>	If embedding the device binary, invoke llc with <args> (default: none).
Xptxas=<args>	If embedding the device binary, invoke ptxas with <args> (default: none). Also passed to fatbinary via --cmdline if embedding the PTX image.
Xfatbinary=<args>	If embedding the device binary, invoke fatbinary with <args> (default: none).

Phases

%ll_nvptx.emit

Emits the LLVM IR with PTX intrinsics for the host and device worlds split from the current world. This phase is spliced into the default pipeline (see opt) whenever the ll_nvptx plugin is loaded.

axm %ll_nvptx.emit: %compile.Phase;