nvvm- FFI bindings to NVVM

Copyright[2016] Trevor L. McDonell
Safe HaskellNone



This module defines an interface to the libNVVM library provided by NVIDIA as part of the CUDA toolkit. It compiles NVVM IR, a compiler intermediate representation based on LLVM IR, into PTX code suitable for execution on NVIDIA GPUs. NVVM IR is a subset of LLVM IR, with a set of rules, restrictions, conventions, and intrinsic functions.

NVIDIA's own nvcc compiler uses NVVM IR and libNVVM internally as part of the CUDA C compilation process. In contrast to the (open-source) NVPTX target included with the standard LLVM toolchain (which also compiles NVVM IR into PTX code), libNVVM includes a set of proprietary optimisation passes, which may result in faster GPU code. More information on NVVM IR can be found here:


The following is a short tutorial on using this library. The steps can be copied into a file, or run directly in ghci, in which case ghci should be launched with the option -fno-ghci-sandbox. This is because CUDA maintains CPU-local state, so operations must be run from a bound thread.

Note that the focus of this library is the generation of executable PTX code from NVVM IR, so we will additionally need to use the cuda package to control and execute the compiled program. In this tutorial we will go over those steps quickly, but see the cuda package for more information and a similar tutorial:


Initialise the CUDA environment

Before any operation can be performed, we must initialise the CUDA Driver API.

>>> import Foreign.CUDA.Driver as CUDA
>>> CUDA.initialise []

Select a GPU and create an execution context for that device. Each available device is given a unique numeric identifier (beginning at zero). For this example we just select the first device (the default).

>>> dev0 <- CUDA.device 0
>>> prp0 <- CUDA.props dev0
>>> ctx0 <- CUDA.create dev0 []

Remember that once the context is no longer needed, it should be destroyed in order to free up any resources that were allocated into it.

Compiling kernels with NVVM

For this example we will step through executing the equivalent of the following Haskell function, which element-wise adds the elements of two arrays:

>>> vecAdd xs ys = zipWith (+) xs ys

The following NVVM IR implements this for the GPU. Note that this example is written using NVVM IR version 1.2 syntax (corresponding to CUDA toolkit 7.5), which is based on LLVM IR version 3.4. The human readable representation of LLVM IR (and by extension NVVM IR) is notorious for changing between releases, whereas the bitcode representation is somewhat more stable. You may wish to keep this in mind for your own programs.

target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-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 @vecAdd(float* %A, float* %B, float* %C) {
  ; What is my ID?
  %id = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind

  ; Compute pointers into A, B, and C
  %ptrA = getelementptr float* %A, i32 %id
  %ptrB = getelementptr float* %B, i32 %id
  %ptrC = getelementptr float* %C, i32 %id

  ; Read A, B
  %valA = load float* %ptrA, align 4
  %valB = load float* %ptrB, align 4

  ; Compute C = A + B
  %valC = fadd float %valA, %valB

  ; Store back to C
  store float %valC, float* %ptrC, align 4

  ret void

; Intrinsic to read threadIdx.x
declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind

!nvvm.annotations = !{!0}
!0 = metadata !{void (float*, float*, float*)* @vecAdd, metadata !"kernel", i64 1}

For reference, in CUDA this kernel would have been written as:

extern "C" __global__ void vecAdd(float *xs, float* ys, float *zs)
    int ix = threadIdx.x;

    zs[ix] = xs[ix] + ys[ix];

The NVVM IR can be stored directly in the program as a [Byte]String, but here I will assume that it is saved to a file vector_add.ll:

>>> import Data.ByteString as B
>>> ll <- B.readFile "vector_add.ll"

Now we can use NVVM to compile this into PTX code:

>>> import Foreign.NVVM as NVVM
>>> ptx <- NVVM.compileModule "vecAdd" ll [ NVVM.Target (CUDA.computeCapability prp0) ]

Notice that we asked NVVM to specialise the generated PTX code for our current device. By default the code will be compiled for compute capability 2.0 (the earliest supported target).

The generated PTX code can then be loaded into the current CUDA execution context, from which we can extract a reference to the GPU kernel that we will later execute.

>>> mdl    <- CUDA.loadData (NVVM.compileResult ptx)
>>> vecAdd <- CUDA.getFun mdl "vecAdd"

After we are finished with the module, it is a good idea to unload it in order to free any resources it used.

Executing the kernel

Executing the vecAdd kernel now proceeds exactly like executing any other kernel function using the CUDA Driver API. The following is a quick overview; see the tutorial in the cuda package for more information.

First, generate some data and copy it to the device. We also allocate an (uninitialised) array on the device to store the results.

>>> let xs = [1..256]   :: [Float]
>>> let ys = [2,4..512] :: [Float]
>>> xs_dev <- CUDA.newListArray xs
>>> ys_dev <- CUDA.newListArray ys
>>> zs_dev <- CUDA.mallocArray 256 :: IO (CUDA.DevicePtr Float)

For this simple kernel we execute it using a single (one dimensional) thread block, with one thread computing each element of the output.

>>> CUDA.launchKernel vecAdd (1,1,1) (256,1,1) 0 Nothing [CUDA.VArg xs_dev, CUDA.VArg ys_dev, CUDA.VArg zs_dev]

Finally, we can copy the results back to the host, and deallocate the arrays from the GPU.

>>> zs <- CUDA.peekListArray 256 zs_dev
>>> CUDA.free xs_dev
>>> CUDA.free ys_dev
>>> CUDA.free zs_dev
Next steps

The library also provides functions for compiling several NVVM IR sources into a single module. In particular this is useful when linking against libdevice, a standard library of functions in NVVM IR which implement, for example, math primitives and bitwise operations. More information on libdevice can be found here: