Julia wrappers for generating PTX code with the NVIDIA NVVM library.

Let's start with some dummy IR that conforms to NVVM IR v2 (aka. LLVM 7.0):

target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
target triple = "nvptx64-nvidia-cuda"

define void @kernel() {
entry:
    ret void
}

!nvvm.annotations = !{!0}
!0 = !{void ()* @kernel, !"kernel", i32 1}

!nvvmir.version = !{!1}
!1 = !{i32 2, i32 0}

With NVVM.jl, you can compile this IR to PTX:

using NVVM

prog = Program()
add!(prog, ir)
verify(ir)
compile(prog)

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-32688072
// Cuda compilation tools, release 12.1, V12.1.105
// Based on NVVM 7.0.1
//

.version 8.1
.target sm_52
.address_size 64

	// .globl	kernel

.visible .entry kernel()
{
	ret;
}

To set compiler options, pass keyword arguments to compile or verify. You can use Julia keyword arguments to specify these options, using slightly user-friendlier syntax:

julia> compile(prog; arch=v"8.0", debug=true, opt=0, prec_sqrt=false)
// Generated by NVIDIA NVVM Compiler
.target sm_80
...

We always use the latest libNVVM, which means that you have to stick to the NVVM IR version it supports (currently v2.0), or use an older version of NVVM.jl. More problematically, that also means that the PTX version will be set to the latest ISA version (currently 8.1), which may not be supported by your driver. A feature request has been filed with NVIDIA to make this configurable.