blob: f57839cec961580705806dfe698649ea5f53221c [file] [log] [blame]
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +00001===================================
2Compiling CUDA C/C++ with LLVM
3===================================
4
5.. contents::
6 :local:
7
8Introduction
9============
10
11This document contains the user guides and the internals of compiling CUDA
12C/C++ with LLVM. It is aimed at both users who want to compile CUDA with LLVM
13and developers who want to improve LLVM for GPUs. This document assumes a basic
14familiarity with CUDA. Information about CUDA programming can be found in the
15`CUDA programming guide
16<http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html>`_.
17
18How to Build LLVM with CUDA Support
19===================================
20
Jingyue Wu313496b2016-01-30 23:48:47 +000021CUDA support is still in development and works the best in the trunk version
22of LLVM. Below is a quick summary of downloading and building the trunk
23version. Consult the `Getting Started
24<http://llvm.org/docs/GettingStarted.html>`_ page for more details on setting
25up LLVM.
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +000026
27#. Checkout LLVM
28
29 .. code-block:: console
30
31 $ cd where-you-want-llvm-to-live
32 $ svn co http://llvm.org/svn/llvm-project/llvm/trunk llvm
33
34#. Checkout Clang
35
36 .. code-block:: console
37
38 $ cd where-you-want-llvm-to-live
39 $ cd llvm/tools
40 $ svn co http://llvm.org/svn/llvm-project/cfe/trunk clang
41
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +000042#. Configure and build LLVM and Clang
43
44 .. code-block:: console
45
46 $ cd where-you-want-llvm-to-live
47 $ mkdir build
48 $ cd build
49 $ cmake [options] ..
50 $ make
51
52How to Compile CUDA C/C++ with LLVM
53===================================
54
55We assume you have installed the CUDA driver and runtime. Consult the `NVIDIA
Jingyue Wuf190ed42016-03-30 05:05:40 +000056CUDA installation guide
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +000057<https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html>`_ if
58you have not.
59
60Suppose you want to compile and run the following CUDA program (``axpy.cu``)
61which multiplies a ``float`` array by a ``float`` scalar (AXPY).
62
63.. code-block:: c++
64
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +000065 #include <iostream>
66
67 __global__ void axpy(float a, float* x, float* y) {
68 y[threadIdx.x] = a * x[threadIdx.x];
69 }
70
71 int main(int argc, char* argv[]) {
72 const int kDataLen = 4;
73
74 float a = 2.0f;
75 float host_x[kDataLen] = {1.0f, 2.0f, 3.0f, 4.0f};
76 float host_y[kDataLen];
77
78 // Copy input data to device.
79 float* device_x;
80 float* device_y;
Jingyue Wu313496b2016-01-30 23:48:47 +000081 cudaMalloc(&device_x, kDataLen * sizeof(float));
82 cudaMalloc(&device_y, kDataLen * sizeof(float));
83 cudaMemcpy(device_x, host_x, kDataLen * sizeof(float),
84 cudaMemcpyHostToDevice);
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +000085
86 // Launch the kernel.
87 axpy<<<1, kDataLen>>>(a, device_x, device_y);
88
89 // Copy output data to host.
Jingyue Wu313496b2016-01-30 23:48:47 +000090 cudaDeviceSynchronize();
91 cudaMemcpy(host_y, device_y, kDataLen * sizeof(float),
92 cudaMemcpyDeviceToHost);
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +000093
94 // Print the results.
95 for (int i = 0; i < kDataLen; ++i) {
96 std::cout << "y[" << i << "] = " << host_y[i] << "\n";
97 }
98
Jingyue Wu313496b2016-01-30 23:48:47 +000099 cudaDeviceReset();
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +0000100 return 0;
101 }
102
103The command line for compilation is similar to what you would use for C++.
104
105.. code-block:: console
106
Jingyue Wu313496b2016-01-30 23:48:47 +0000107 $ clang++ axpy.cu -o axpy --cuda-gpu-arch=<GPU arch> \
108 -L<CUDA install path>/<lib64 or lib> \
109 -lcudart_static -ldl -lrt -pthread
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +0000110 $ ./axpy
111 y[0] = 2
112 y[1] = 4
113 y[2] = 6
114 y[3] = 8
115
Jingyue Wu313496b2016-01-30 23:48:47 +0000116``<CUDA install path>`` is the root directory where you installed CUDA SDK,
117typically ``/usr/local/cuda``. ``<GPU arch>`` is `the compute capability of
118your GPU <https://developer.nvidia.com/cuda-gpus>`_. For example, if you want
119to run your program on a GPU with compute capability of 3.5, you should specify
120``--cuda-gpu-arch=sm_35``.
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +0000121
Justin Lebar32835c82016-03-21 23:05:15 +0000122Detecting clang vs NVCC
123=======================
124
125Although clang's CUDA implementation is largely compatible with NVCC's, you may
126still want to detect when you're compiling CUDA code specifically with clang.
127
Justin Lebar068a7942016-03-23 22:43:10 +0000128This is tricky, because NVCC may invoke clang as part of its own compilation
129process! For example, NVCC uses the host compiler's preprocessor when
130compiling for device code, and that host compiler may in fact be clang.
Justin Lebar32835c82016-03-21 23:05:15 +0000131
132When clang is actually compiling CUDA code -- rather than being used as a
133subtool of NVCC's -- it defines the ``__CUDA__`` macro. ``__CUDA_ARCH__`` is
134defined only in device mode (but will be defined if NVCC is using clang as a
135preprocessor). So you can use the following incantations to detect clang CUDA
136compilation, in host and device modes:
137
138.. code-block:: c++
139
140 #if defined(__clang__) && defined(__CUDA__) && !defined(__CUDA_ARCH__)
141 // clang compiling CUDA code, host mode.
142 #endif
143
144 #if defined(__clang__) && defined(__CUDA__) && defined(__CUDA_ARCH__)
145 // clang compiling CUDA code, device mode.
146 #endif
147
Justin Lebar068a7942016-03-23 22:43:10 +0000148Both clang and nvcc define ``__CUDACC__`` during CUDA compilation. You can
149detect NVCC specifically by looking for ``__NVCC__``.
Justin Lebar32835c82016-03-21 23:05:15 +0000150
Justin Lebarb649e752016-05-25 23:11:31 +0000151Flags that control numerical code
152=================================
153
154If you're using GPUs, you probably care about making numerical code run fast.
155GPU hardware allows for more control over numerical operations than most CPUs,
156but this results in more compiler options for you to juggle.
157
158Flags you may wish to tweak include:
159
160* ``-ffp-contract={on,off,fast}`` (defaults to ``fast`` on host and device when
161 compiling CUDA) Controls whether the compiler emits fused multiply-add
162 operations.
163
164 * ``off``: never emit fma operations, and prevent ptxas from fusing multiply
165 and add instructions.
166 * ``on``: fuse multiplies and adds within a single statement, but never
167 across statements (C11 semantics). Prevent ptxas from fusing other
168 multiplies and adds.
169 * ``fast``: fuse multiplies and adds wherever profitable, even across
170 statements. Doesn't prevent ptxas from fusing additional multiplies and
171 adds.
172
173 Fused multiply-add instructions can be much faster than the unfused
174 equivalents, but because the intermediate result in an fma is not rounded,
175 this flag can affect numerical code.
176
177* ``-fcuda-flush-denormals-to-zero`` (default: off) When this is enabled,
178 floating point operations may flush `denormal
179 <https://en.wikipedia.org/wiki/Denormal_number>`_ inputs and/or outputs to 0.
180 Operations on denormal numbers are often much slower than the same operations
181 on normal numbers.
182
183* ``-fcuda-approx-transcendentals`` (default: off) When this is enabled, the
184 compiler may emit calls to faster, approximate versions of transcendental
185 functions, instead of using the slower, fully IEEE-compliant versions. For
186 example, this flag allows clang to emit the ptx ``sin.approx.f32``
187 instruction.
188
189 This is implied by ``-ffast-math``.
190
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +0000191Optimizations
192=============
193
194CPU and GPU have different design philosophies and architectures. For example, a
195typical CPU has branch prediction, out-of-order execution, and is superscalar,
196whereas a typical GPU has none of these. Due to such differences, an
197optimization pipeline well-tuned for CPUs may be not suitable for GPUs.
198
199LLVM performs several general and CUDA-specific optimizations for GPUs. The
200list below shows some of the more important optimizations for GPUs. Most of
201them have been upstreamed to ``lib/Transforms/Scalar`` and
202``lib/Target/NVPTX``. A few of them have not been upstreamed due to lack of a
203customizable target-independent optimization pipeline.
204
205* **Straight-line scalar optimizations**. These optimizations reduce redundancy
206 in straight-line code. Details can be found in the `design document for
207 straight-line scalar optimizations <https://goo.gl/4Rb9As>`_.
208
209* **Inferring memory spaces**. `This optimization
Jingyue Wuf190ed42016-03-30 05:05:40 +0000210 <https://github.com/llvm-mirror/llvm/blob/master/lib/Target/NVPTX/NVPTXInferAddressSpaces.cpp>`_
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +0000211 infers the memory space of an address so that the backend can emit faster
Jingyue Wuf190ed42016-03-30 05:05:40 +0000212 special loads and stores from it.
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +0000213
214* **Aggressive loop unrooling and function inlining**. Loop unrolling and
215 function inlining need to be more aggressive for GPUs than for CPUs because
216 control flow transfer in GPU is more expensive. They also promote other
217 optimizations such as constant propagation and SROA which sometimes speed up
218 code by over 10x. An empirical inline threshold for GPUs is 1100. This
219 configuration has yet to be upstreamed with a target-specific optimization
220 pipeline. LLVM also provides `loop unrolling pragmas
221 <http://clang.llvm.org/docs/AttributeReference.html#pragma-unroll-pragma-nounroll>`_
222 and ``__attribute__((always_inline))`` for programmers to force unrolling and
223 inling.
224
225* **Aggressive speculative execution**. `This transformation
226 <http://llvm.org/docs/doxygen/html/SpeculativeExecution_8cpp_source.html>`_ is
227 mainly for promoting straight-line scalar optimizations which are most
228 effective on code along dominator paths.
229
230* **Memory-space alias analysis**. `This alias analysis
Jingyue Wu03d90e52015-11-18 22:01:44 +0000231 <http://reviews.llvm.org/D12414>`_ infers that two pointers in different
Jingyue Wu4f2a6cb2015-11-10 22:35:47 +0000232 special memory spaces do not alias. It has yet to be integrated to the new
233 alias analysis infrastructure; the new infrastructure does not run
234 target-specific alias analysis.
235
236* **Bypassing 64-bit divides**. `An existing optimization
237 <http://llvm.org/docs/doxygen/html/BypassSlowDivision_8cpp_source.html>`_
238 enabled in the NVPTX backend. 64-bit integer divides are much slower than
239 32-bit ones on NVIDIA GPUs due to lack of a divide unit. Many of the 64-bit
240 divides in our benchmarks have a divisor and dividend which fit in 32-bits at
241 runtime. This optimization provides a fast path for this common case.
Jingyue Wubec78182016-02-23 23:34:49 +0000242
Jingyue Wuf190ed42016-03-30 05:05:40 +0000243Publication
244===========
245
246| `gpucc: An Open-Source GPGPU Compiler <http://dl.acm.org/citation.cfm?id=2854041>`_
247| Jingyue Wu, Artem Belevich, Eli Bendersky, Mark Heffernan, Chris Leary, Jacques Pienaar, Bjarke Roune, Rob Springer, Xuetian Weng, Robert Hundt
248| *Proceedings of the 2016 International Symposium on Code Generation and Optimization (CGO 2016)*
249| `Slides for the CGO talk <http://wujingyue.com/docs/gpucc-talk.pdf>`_
250
251Tutorial
252========
253
254`CGO 2016 gpucc tutorial <http://wujingyue.com/docs/gpucc-tutorial.pdf>`_
255
Jingyue Wubec78182016-02-23 23:34:49 +0000256Obtaining Help
257==============
258
259To obtain help on LLVM in general and its CUDA support, see `the LLVM
260community <http://llvm.org/docs/#mailing-lists>`_.