llvm/docs/CompileCudaWithLLVM.rst

=========================
Compiling CUDA with clang
=========================

.. contents::
   :local:

Introduction
============

This document describes how to compile CUDA code with clang, and gives some
details about LLVM and clang's CUDA implementations.

This document assumes a basic familiarity with CUDA. Information about CUDA
programming can be found in the
`CUDA programming guide
<http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html>`_.

Compiling CUDA Code
===================

Prerequisites
-------------

CUDA is supported since llvm 3.9. Clang currently supports CUDA 7.0 through
11.5. If clang detects a newer CUDA version, it will issue a warning and will
attempt to use detected CUDA SDK it as if it were CUDA 11.5.

Before you build CUDA code, you'll need to have installed the CUDA SDK.  See
`NVIDIA's CUDA installation guide
<https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html>`_ for
details.  Note that clang `maynot support
<https://bugs.llvm.org/show_bug.cgi?id=26966>`_ the CUDA toolkit as installed by
some Linux package managers. Clang does attempt to deal with specific details of
CUDA installation on a handful of common Linux distributions, but in general the
most reliable way to make it work is to install CUDA in a single directory from
NVIDIA's `.run` package and specify its location via `--cuda-path=...` argument.

CUDA compilation is supported on Linux. Compilation on MacOS and Windows may or
may not work and currently have no maintainers.

Invoking clang
--------------

Invoking clang for CUDA compilation works similarly to compiling regular C++.
You just need to be aware of a few additional flags.

You can use `this <https://gist.github.com/855e277884eb6b388cd2f00d956c2fd4>`_
program as a toy example.  Save it as ``axpy.cu``.  (Clang detects that you're
compiling CUDA code by noticing that your filename ends with ``.cu``.
Alternatively, you can pass ``-x cuda``.)

To build and run, run the following commands, filling in the parts in angle
brackets as described below:

.. code-block:: console

  $ clang++ axpy.cu -o axpy --cuda-gpu-arch=<GPU arch> \
      -L<CUDA install path>/<lib64 or lib>             \
      -lcudart_static -ldl -lrt -pthread
  $ ./axpy
  y[0] = 2
  y[1] = 4
  y[2] = 6
  y[3] = 8

On MacOS, replace `-lcudart_static` with `-lcudart`; otherwise, you may get
"CUDA driver version is insufficient for CUDA runtime version" errors when you
run your program.

* ``<CUDA install path>`` -- the directory where you installed CUDA SDK.
  Typically, ``/usr/local/cuda``.

  Pass e.g. ``-L/usr/local/cuda/lib64`` if compiling in 64-bit mode; otherwise,
  pass e.g. ``-L/usr/local/cuda/lib``.  (In CUDA, the device code and host code
  always have the same pointer widths, so if you're compiling 64-bit code for
  the host, you're also compiling 64-bit code for the device.) Note that as of
  v10.0 CUDA SDK `no longer supports compilation of 32-bit
  applications <https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#deprecated-features>`_.

* ``<GPU arch>`` -- the `compute capability
  <https://developer.nvidia.com/cuda-gpus>`_ of your GPU. For example, if you
  want to run your program on a GPU with compute capability of 3.5, specify
  ``--cuda-gpu-arch=sm_35``.

  Note: You cannot pass ``compute_XX`` as an argument to ``--cuda-gpu-arch``;
  only ``sm_XX`` is currently supported.  However, clang always includes PTX in
  its binaries, so e.g. a binary compiled with ``--cuda-gpu-arch=sm_30`` would be
  forwards-compatible with e.g. ``sm_35`` GPUs.

  You can pass ``--cuda-gpu-arch`` multiple times to compile for multiple archs.

The `-L` and `-l` flags only need to be passed when linking.  When compiling,
you may also need to pass ``--cuda-path=/path/to/cuda`` if you didn't install
the CUDA SDK into ``/usr/local/cuda`` or ``/usr/local/cuda-X.Y``.

Flags that control numerical code
---------------------------------

If you're using GPUs, you probably care about making numerical code run fast.
GPU hardware allows for more control over numerical operations than most CPUs,
but this results in more compiler options for you to juggle.

Flags you may wish to tweak include:

* ``-ffp-contract={on,off,fast}`` (defaults to ``fast`` on host and device when
  compiling CUDA) Controls whether the compiler emits fused multiply-add
  operations.

  * ``off``: never emit fma operations, and prevent ptxas from fusing multiply
    and add instructions.
  * ``on``: fuse multiplies and adds within a single statement, but never
    across statements (C11 semantics).  Prevent ptxas from fusing other
    multiplies and adds.
  * ``fast``: fuse multiplies and adds wherever profitable, even across
    statements.  Doesn't prevent ptxas from fusing additional multiplies and
    adds.

  Fused multiply-add instructions can be much faster than the unfused
  equivalents, but because the intermediate result in an fma is not rounded,
  this flag can affect numerical code.

* ``-fcuda-flush-denormals-to-zero`` (default: off) When this is enabled,
  floating point operations may flush `denormal
  <https://en.wikipedia.org/wiki/Denormal_number>`_ inputs and/or outputs to 0.
  Operations on denormal numbers are often much slower than the same operations
  on normal numbers.

* ``-fcuda-approx-transcendentals`` (default: off) When this is enabled, the
  compiler may emit calls to faster, approximate versions of transcendental
  functions, instead of using the slower, fully IEEE-compliant versions.  For
  example, this flag allows clang to emit the ptx ``sin.approx.f32``
  instruction.

  This is implied by ``-ffast-math``.

Standard library support
========================

In clang and nvcc, most of the C++ standard library is not supported on the
device side.

``<math.h>`` and ``<cmath>``
----------------------------

In clang, ``math.h`` and ``cmath`` are available and `pass
<https://github.com/llvm/llvm-test-suite/blob/main/External/CUDA/math_h.cu>`_
`tests
<https://github.com/llvm/llvm-test-suite/blob/main/External/CUDA/cmath.cu>`_
adapted from libc++'s test suite.

In nvcc ``math.h`` and ``cmath`` are mostly available.  Versions of ``::foof``
in namespace std (e.g. ``std::sinf``) are not available, and where the standard
calls for overloads that take integral arguments, these are usually not
available.

.. code-block:: c++

  #include <math.h>
  #include <cmath.h>

  // clang is OK with everything in this function.
  __device__ void test() {
    std::sin(0.); // nvcc - ok
    std::sin(0);  // nvcc - error, because no std::sin(int) override is available.
    sin(0);       // nvcc - same as above.

    sinf(0.);       // nvcc - ok
    std::sinf(0.);  // nvcc - no such function
  }

``<std::complex>``
------------------

nvcc does not officially support ``std::complex``.  It's an error to use
``std::complex`` in ``__device__`` code, but it often works in ``__host__
__device__`` code due to nvcc's interpretation of the "wrong-side rule" (see
below).  However, we have heard from implementers that it's possible to get
into situations where nvcc will omit a call to an ``std::complex`` function,
especially when compiling without optimizations.

As of 2016-11-16, clang supports ``std::complex`` without these caveats.  It is
tested with libstdc++ 4.8.5 and newer, but is known to work only with libc++
newer than 2016-11-16.

``<algorithm>``
---------------

In C++14, many useful functions from ``<algorithm>`` (notably, ``std::min`` and
``std::max``) become constexpr.  You can therefore use these in device code,
when compiling with clang.

Detecting clang vs NVCC from code
=================================

Although clang's CUDA implementation is largely compatible with NVCC's, you may
still want to detect when you're compiling CUDA code specifically with clang.

This is tricky, because NVCC may invoke clang as part of its own compilation
process!  For example, NVCC uses the host compiler's preprocessor when
compiling for device code, and that host compiler may in fact be clang.

When clang is actually compiling CUDA code -- rather than being used as a
subtool of NVCC's -- it defines the ``__CUDA__`` macro.  ``__CUDA_ARCH__`` is
defined only in device mode (but will be defined if NVCC is using clang as a
preprocessor).  So you can use the following incantations to detect clang CUDA
compilation, in host and device modes:

.. code-block:: c++

  #if defined(__clang__) && defined(__CUDA__) && !defined(__CUDA_ARCH__)
  // clang compiling CUDA code, host mode.
  #endif

  #if defined(__clang__) && defined(__CUDA__) && defined(__CUDA_ARCH__)
  // clang compiling CUDA code, device mode.
  #endif

Both clang and nvcc define ``__CUDACC__`` during CUDA compilation.  You can
detect NVCC specifically by looking for ``__NVCC__``.

Dialect Differences Between clang and nvcc
==========================================

There is no formal CUDA spec, and clang and nvcc speak slightly different
dialects of the language.  Below, we describe some of the differences.

This section is painful; hopefully you can skip this section and live your life
blissfully unaware.

Compilation Models
------------------

Most of the differences between clang and nvcc stem from the different
compilation models used by clang and nvcc.  nvcc uses *split compilation*,
which works roughly as follows:

 * Run a preprocessor over the input ``.cu`` file to split it into two source
   files: ``H``, containing source code for the host, and ``D``, containing
   source code for the device.

 * For each GPU architecture ``arch`` that we're compiling for, do:

   * Compile ``D`` using nvcc proper.  The result of this is a ``ptx`` file for
     ``P_arch``.

   * Optionally, invoke ``ptxas``, the PTX assembler, to generate a file,
     ``S_arch``, containing GPU machine code (SASS) for ``arch``.

 * Invoke ``fatbin`` to combine all ``P_arch`` and ``S_arch`` files into a
   single "fat binary" file, ``F``.

 * Compile ``H`` using an external host compiler (gcc, clang, or whatever you
   like).  ``F`` is packaged up into a header file which is force-included into
   ``H``; nvcc generates code that calls into this header to e.g. launch
   kernels.

clang uses *merged parsing*.  This is similar to split compilation, except all
of the host and device code is present and must be semantically-correct in both
compilation steps.

  * For each GPU architecture ``arch`` that we're compiling for, do:

    * Compile the input ``.cu`` file for device, using clang.  ``__host__`` code
      is parsed and must be semantically correct, even though we're not
      generating code for the host at this time.

      The output of this step is a ``ptx`` file ``P_arch``.

    * Invoke ``ptxas`` to generate a SASS file, ``S_arch``.  Note that, unlike
      nvcc, clang always generates SASS code.

  * Invoke ``fatbin`` to combine all ``P_arch`` and ``S_arch`` files into a
    single fat binary file, ``F``.

  * Compile ``H`` using clang.  ``__device__`` code is parsed and must be
    semantically correct, even though we're not generating code for the device
    at this time.

    ``F`` is passed to this compilation, and clang includes it in a special ELF
    section, where it can be found by tools like ``cuobjdump``.

(You may ask at this point, why does clang need to parse the input file
multiple times?  Why not parse it just once, and then use the AST to generate
code for the host and each device architecture?

Unfortunately this can't work because we have to define different macros during
host compilation and during device compilation for each GPU architecture.)

clang's approach allows it to be highly robust to C++ edge cases, as it doesn't
need to decide at an early stage which declarations to keep and which to throw
away.  But it has some consequences you should be aware of.

Overloading Based on ``__host__`` and ``__device__`` Attributes
---------------------------------------------------------------

Let "H", "D", and "HD" stand for "``__host__`` functions", "``__device__``
functions", and "``__host__ __device__`` functions", respectively.  Functions
with no attributes behave the same as H.

nvcc does not allow you to create H and D functions with the same signature:

.. code-block:: c++

  // nvcc: error - function "foo" has already been defined
  __host__ void foo() {}
  __device__ void foo() {}

However, nvcc allows you to "overload" H and D functions with different
signatures:

.. code-block:: c++

  // nvcc: no error
  __host__ void foo(int) {}
  __device__ void foo() {}

In clang, the ``__host__`` and ``__device__`` attributes are part of a
function's signature, and so it's legal to have H and D functions with
(otherwise) the same signature:

.. code-block:: c++

  // clang: no error
  __host__ void foo() {}
  __device__ void foo() {}

HD functions cannot be overloaded by H or D functions with the same signature:

.. code-block:: c++

  // nvcc: error - function "foo" has already been defined
  // clang: error - redefinition of 'foo'
  __host__ __device__ void foo() {}
  __device__ void foo() {}

  // nvcc: no error
  // clang: no error
  __host__ __device__ void bar(int) {}
  __device__ void bar() {}

When resolving an overloaded function, clang considers the host/device
attributes of the caller and callee.  These are used as a tiebreaker during
overload resolution.  See `IdentifyCUDAPreference
<https://clang.llvm.org/doxygen/SemaCUDA_8cpp.html>`_ for the full set of rules,
but at a high level they are:

 * D functions prefer to call other Ds.  HDs are given lower priority.

 * Similarly, H functions prefer to call other Hs, or ``__global__`` functions
   (with equal priority).  HDs are given lower priority.

 * HD functions prefer to call other HDs.

   When compiling for device, HDs will call Ds with lower priority than HD, and
   will call Hs with still lower priority.  If it's forced to call an H, the
   program is malformed if we emit code for this HD function.  We call this the
   "wrong-side rule", see example below.

   The rules are symmetrical when compiling for host.

Some examples:

.. code-block:: c++

   __host__ void foo();
   __device__ void foo();

   __host__ void bar();
   __host__ __device__ void bar();

   __host__ void test_host() {
     foo();  // calls H overload
     bar();  // calls H overload
   }

   __device__ void test_device() {
     foo();  // calls D overload
     bar();  // calls HD overload
   }

   __host__ __device__ void test_hd() {
     foo();  // calls H overload when compiling for host, otherwise D overload
     bar();  // always calls HD overload
   }

Wrong-side rule example:

.. code-block:: c++

  __host__ void host_only();

  // We don't codegen inline functions unless they're referenced by a
  // non-inline function.  inline_hd1() is called only from the host side, so
  // does not generate an error.  inline_hd2() is called from the device side,
  // so it generates an error.
  inline __host__ __device__ void inline_hd1() { host_only(); }  // no error
  inline __host__ __device__ void inline_hd2() { host_only(); }  // error

  __host__ void host_fn() { inline_hd1(); }
  __device__ void device_fn() { inline_hd2(); }

  // This function is not inline, so it's always codegen'ed on both the host
  // and the device.  Therefore, it generates an error.
  __host__ __device__ void not_inline_hd() { host_only(); }

For the purposes of the wrong-side rule, templated functions also behave like
``inline`` functions: They aren't codegen'ed unless they're instantiated
(usually as part of the process of invoking them).

clang's behavior with respect to the wrong-side rule matches nvcc's, except
nvcc only emits a warning for ``not_inline_hd``; device code is allowed to call
``not_inline_hd``.  In its generated code, nvcc may omit ``not_inline_hd``'s
call to ``host_only`` entirely, or it may try to generate code for
``host_only`` on the device.  What you get seems to depend on whether or not
the compiler chooses to inline ``host_only``.

Member functions, including constructors, may be overloaded using H and D
attributes.  However, destructors cannot be overloaded.

Using a Different Class on Host/Device
--------------------------------------

Occasionally you may want to have a class with different host/device versions.

If all of the class's members are the same on the host and device, you can just
provide overloads for the class's member functions.

However, if you want your class to have different members on host/device, you
won't be able to provide working H and D overloads in both classes. In this
case, clang is likely to be unhappy with you.

.. code-block:: c++

  #ifdef __CUDA_ARCH__
  struct S {
    __device__ void foo() { /* use device_only */ }
    int device_only;
  };
  #else
  struct S {
    __host__ void foo() { /* use host_only */ }
    double host_only;
  };

  __device__ void test() {
    S s;
    // clang generates an error here, because during host compilation, we
    // have ifdef'ed away the __device__ overload of S::foo().  The __device__
    // overload must be present *even during host compilation*.
    S.foo();
  }
  #endif

We posit that you don't really want to have classes with different members on H
and D.  For example, if you were to pass one of these as a parameter to a
kernel, it would have a different layout on H and D, so would not work
properly.

To make code like this compatible with clang, we recommend you separate it out
into two classes.  If you need to write code that works on both host and
device, consider writing an overloaded wrapper function that returns different
types on host and device.

.. code-block:: c++

  struct HostS { ... };
  struct DeviceS { ... };

  __host__ HostS MakeStruct() { return HostS(); }
  __device__ DeviceS MakeStruct() { return DeviceS(); }

  // Now host and device code can call MakeStruct().

Unfortunately, this idiom isn't compatible with nvcc, because it doesn't allow
you to overload based on the H/D attributes.  Here's an idiom that works with
both clang and nvcc:

.. code-block:: c++

  struct HostS { ... };
  struct DeviceS { ... };

  #ifdef __NVCC__
    #ifndef __CUDA_ARCH__
      __host__ HostS MakeStruct() { return HostS(); }
    #else
      __device__ DeviceS MakeStruct() { return DeviceS(); }
    #endif
  #else
    __host__ HostS MakeStruct() { return HostS(); }
    __device__ DeviceS MakeStruct() { return DeviceS(); }
  #endif

  // Now host and device code can call MakeStruct().

Hopefully you don't have to do this sort of thing often.

Optimizations
=============

Modern CPUs and GPUs are architecturally quite different, so code that's fast
on a CPU isn't necessarily fast on a GPU.  We've made a number of changes to
LLVM to make it generate good GPU code.  Among these changes are:

* `Straight-line scalar optimizations <https://goo.gl/4Rb9As>`_ -- These
  reduce redundancy within straight-line code.

* `Aggressive speculative execution
  <https://llvm.org/docs/doxygen/html/SpeculativeExecution_8cpp_source.html>`_
  -- This is mainly for promoting straight-line scalar optimizations, which are
  most effective on code along dominator paths.

* `Memory space inference
  <https://llvm.org/doxygen/NVPTXInferAddressSpaces_8cpp_source.html>`_ --
  In PTX, we can operate on pointers that are in a particular "address space"
  (global, shared, constant, or local), or we can operate on pointers in the
  "generic" address space, which can point to anything.  Operations in a
  non-generic address space are faster, but pointers in CUDA are not explicitly
  annotated with their address space, so it's up to LLVM to infer it where
  possible.

* `Bypassing 64-bit divides
  <https://llvm.org/docs/doxygen/html/BypassSlowDivision_8cpp_source.html>`_ --
  This was an existing optimization that we enabled for the PTX backend.

  64-bit integer divides are much slower than 32-bit ones on NVIDIA GPUs.
  Many of the 64-bit divides in our benchmarks have a divisor and dividend
  which fit in 32-bits at runtime. This optimization provides a fast path for
  this common case.

* Aggressive loop unrolling and function inlining -- Loop unrolling and
  function inlining need to be more aggressive for GPUs than for CPUs because
  control flow transfer in GPU is more expensive. More aggressive unrolling and
  inlining also promote other optimizations, such as constant propagation and
  SROA, which sometimes speed up code by over 10x.

  (Programmers can force unrolling and inline using clang's `loop unrolling pragmas
  <https://clang.llvm.org/docs/AttributeReference.html#pragma-unroll-pragma-nounroll>`_
  and ``__attribute__((always_inline))``.)

Publication
===========

The team at Google published a paper in CGO 2016 detailing the optimizations
they'd made to clang/LLVM.  Note that "gpucc" is no longer a meaningful name:
The relevant tools are now just vanilla clang/LLVM.

| `gpucc: An Open-Source GPGPU Compiler <http://dl.acm.org/citation.cfm?id=2854041>`_
| Jingyue Wu, Artem Belevich, Eli Bendersky, Mark Heffernan, Chris Leary, Jacques Pienaar, Bjarke Roune, Rob Springer, Xuetian Weng, Robert Hundt
| *Proceedings of the 2016 International Symposium on Code Generation and Optimization (CGO 2016)*
|
| `Slides from the CGO talk <http://wujingyue.github.io/docs/gpucc-talk.pdf>`_
|
| `Tutorial given at CGO <http://wujingyue.github.io/docs/gpucc-tutorial.pdf>`_

Obtaining Help
==============

To obtain help on LLVM in general and its CUDA support, see `the LLVM
community <https://llvm.org/docs/#mailing-lists>`_.