1 =========================
2 Compiling CUDA with clang
3 =========================
11 This document describes how to compile CUDA code with clang, and gives some
12 details about LLVM and clang's CUDA implementations.
14 This document assumes a basic familiarity with CUDA. Information about CUDA
15 programming can be found in the
16 `CUDA programming guide
17 <http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html>`_.
25 CUDA is supported in llvm 3.9, but it's still in active development, so we
26 recommend you `compile clang/LLVM from HEAD
27 <http://llvm.org/docs/GettingStarted.html>`_.
29 Before you build CUDA code, you'll need to have installed the appropriate
30 driver for your nvidia GPU and the CUDA SDK. See `NVIDIA's CUDA installation
31 guide <https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html>`_
32 for details. Note that clang `does not support
33 <https://llvm.org/bugs/show_bug.cgi?id=26966>`_ the CUDA toolkit as installed
34 by many Linux package managers; you probably need to install nvidia's package.
36 You will need CUDA 7.0 or 7.5 to compile with clang. CUDA 8 support is in the
42 Invoking clang for CUDA compilation works similarly to compiling regular C++.
43 You just need to be aware of a few additional flags.
45 You can use `this <https://gist.github.com/855e277884eb6b388cd2f00d956c2fd4>_`
46 program as a toy example. Save it as ``axpy.cu``. To build and run, run the
49 .. code-block:: console
51 $ clang++ axpy.cu -o axpy --cuda-gpu-arch=<GPU arch> \
52 -L<CUDA install path>/<lib64 or lib> \
53 -lcudart_static -ldl -lrt -pthread
60 * clang detects that you're compiling CUDA by noticing that your source file ends
61 with ``.cu``. (Alternatively, you can pass ``-x cuda``.)
63 * ``<CUDA install path>`` is the root directory where you installed CUDA SDK,
64 typically ``/usr/local/cuda``.
66 Pass e.g. ``/usr/local/cuda/lib64`` if compiling in 64-bit mode; otherwise,
67 pass ``/usr/local/cuda/lib``. (In CUDA, the device code and host code always
68 have the same pointer widths, so if you're compiling 64-bit code for the
69 host, you're also compiling 64-bit code for the device.)
71 * ``<GPU arch>`` is `the compute capability of your GPU
72 <https://developer.nvidia.com/cuda-gpus>`_. For example, if you want to run
73 your program on a GPU with compute capability of 3.5, you should specify
74 ``--cuda-gpu-arch=sm_35``.
76 Note: You cannot pass ``compute_XX`` as an argument to ``--cuda-gpu-arch``;
77 only ``sm_XX`` is currently supported. However, clang always includes PTX in
78 its binaries, so e.g. a binary compiled with ``--cuda-gpu-arch=sm_30`` would be
79 forwards-compatible with e.g. ``sm_35`` GPUs.
81 You can pass ``--cuda-gpu-arch`` multiple times to compile for multiple
84 Flags that control numerical code
85 ---------------------------------
87 If you're using GPUs, you probably care about making numerical code run fast.
88 GPU hardware allows for more control over numerical operations than most CPUs,
89 but this results in more compiler options for you to juggle.
91 Flags you may wish to tweak include:
93 * ``-ffp-contract={on,off,fast}`` (defaults to ``fast`` on host and device when
94 compiling CUDA) Controls whether the compiler emits fused multiply-add
97 * ``off``: never emit fma operations, and prevent ptxas from fusing multiply
99 * ``on``: fuse multiplies and adds within a single statement, but never
100 across statements (C11 semantics). Prevent ptxas from fusing other
102 * ``fast``: fuse multiplies and adds wherever profitable, even across
103 statements. Doesn't prevent ptxas from fusing additional multiplies and
106 Fused multiply-add instructions can be much faster than the unfused
107 equivalents, but because the intermediate result in an fma is not rounded,
108 this flag can affect numerical code.
110 * ``-fcuda-flush-denormals-to-zero`` (default: off) When this is enabled,
111 floating point operations may flush `denormal
112 <https://en.wikipedia.org/wiki/Denormal_number>`_ inputs and/or outputs to 0.
113 Operations on denormal numbers are often much slower than the same operations
116 * ``-fcuda-approx-transcendentals`` (default: off) When this is enabled, the
117 compiler may emit calls to faster, approximate versions of transcendental
118 functions, instead of using the slower, fully IEEE-compliant versions. For
119 example, this flag allows clang to emit the ptx ``sin.approx.f32``
122 This is implied by ``-ffast-math``.
124 Detecting clang vs NVCC from code
125 =================================
127 Although clang's CUDA implementation is largely compatible with NVCC's, you may
128 still want to detect when you're compiling CUDA code specifically with clang.
130 This is tricky, because NVCC may invoke clang as part of its own compilation
131 process! For example, NVCC uses the host compiler's preprocessor when
132 compiling for device code, and that host compiler may in fact be clang.
134 When clang is actually compiling CUDA code -- rather than being used as a
135 subtool of NVCC's -- it defines the ``__CUDA__`` macro. ``__CUDA_ARCH__`` is
136 defined only in device mode (but will be defined if NVCC is using clang as a
137 preprocessor). So you can use the following incantations to detect clang CUDA
138 compilation, in host and device modes:
142 #if defined(__clang__) && defined(__CUDA__) && !defined(__CUDA_ARCH__)
143 // clang compiling CUDA code, host mode.
146 #if defined(__clang__) && defined(__CUDA__) && defined(__CUDA_ARCH__)
147 // clang compiling CUDA code, device mode.
150 Both clang and nvcc define ``__CUDACC__`` during CUDA compilation. You can
151 detect NVCC specifically by looking for ``__NVCC__``.
156 CPU and GPU have different design philosophies and architectures. For example, a
157 typical CPU has branch prediction, out-of-order execution, and is superscalar,
158 whereas a typical GPU has none of these. Due to such differences, an
159 optimization pipeline well-tuned for CPUs may be not suitable for GPUs.
161 LLVM performs several general and CUDA-specific optimizations for GPUs. The
162 list below shows some of the more important optimizations for GPUs. Most of
163 them have been upstreamed to ``lib/Transforms/Scalar`` and
164 ``lib/Target/NVPTX``. A few of them have not been upstreamed due to lack of a
165 customizable target-independent optimization pipeline.
167 * **Straight-line scalar optimizations**. These optimizations reduce redundancy
168 in straight-line code. Details can be found in the `design document for
169 straight-line scalar optimizations <https://goo.gl/4Rb9As>`_.
171 * **Inferring memory spaces**. `This optimization
172 <https://github.com/llvm-mirror/llvm/blob/master/lib/Target/NVPTX/NVPTXInferAddressSpaces.cpp>`_
173 infers the memory space of an address so that the backend can emit faster
174 special loads and stores from it.
176 * **Aggressive loop unrooling and function inlining**. Loop unrolling and
177 function inlining need to be more aggressive for GPUs than for CPUs because
178 control flow transfer in GPU is more expensive. They also promote other
179 optimizations such as constant propagation and SROA which sometimes speed up
180 code by over 10x. An empirical inline threshold for GPUs is 1100. This
181 configuration has yet to be upstreamed with a target-specific optimization
182 pipeline. LLVM also provides `loop unrolling pragmas
183 <http://clang.llvm.org/docs/AttributeReference.html#pragma-unroll-pragma-nounroll>`_
184 and ``__attribute__((always_inline))`` for programmers to force unrolling and
187 * **Aggressive speculative execution**. `This transformation
188 <http://llvm.org/docs/doxygen/html/SpeculativeExecution_8cpp_source.html>`_ is
189 mainly for promoting straight-line scalar optimizations which are most
190 effective on code along dominator paths.
192 * **Memory-space alias analysis**. `This alias analysis
193 <http://reviews.llvm.org/D12414>`_ infers that two pointers in different
194 special memory spaces do not alias. It has yet to be integrated to the new
195 alias analysis infrastructure; the new infrastructure does not run
196 target-specific alias analysis.
198 * **Bypassing 64-bit divides**. `An existing optimization
199 <http://llvm.org/docs/doxygen/html/BypassSlowDivision_8cpp_source.html>`_
200 enabled in the NVPTX backend. 64-bit integer divides are much slower than
201 32-bit ones on NVIDIA GPUs due to lack of a divide unit. Many of the 64-bit
202 divides in our benchmarks have a divisor and dividend which fit in 32-bits at
203 runtime. This optimization provides a fast path for this common case.
208 | `gpucc: An Open-Source GPGPU Compiler <http://dl.acm.org/citation.cfm?id=2854041>`_
209 | Jingyue Wu, Artem Belevich, Eli Bendersky, Mark Heffernan, Chris Leary, Jacques Pienaar, Bjarke Roune, Rob Springer, Xuetian Weng, Robert Hundt
210 | *Proceedings of the 2016 International Symposium on Code Generation and Optimization (CGO 2016)*
211 | `Slides for the CGO talk <http://wujingyue.com/docs/gpucc-talk.pdf>`_
216 `CGO 2016 gpucc tutorial <http://wujingyue.com/docs/gpucc-tutorial.pdf>`_
221 To obtain help on LLVM in general and its CUDA support, see `the LLVM
222 community <http://llvm.org/docs/#mailing-lists>`_.