|
150
|
1 =========================
|
|
|
2 Compiling CUDA with clang
|
|
|
3 =========================
|
|
|
4
|
|
|
5 .. contents::
|
|
|
6 :local:
|
|
|
7
|
|
|
8 Introduction
|
|
|
9 ============
|
|
|
10
|
|
|
11 This document describes how to compile CUDA code with clang, and gives some
|
|
|
12 details about LLVM and clang's CUDA implementations.
|
|
|
13
|
|
|
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>`_.
|
|
|
18
|
|
|
19 Compiling CUDA Code
|
|
|
20 ===================
|
|
|
21
|
|
|
22 Prerequisites
|
|
|
23 -------------
|
|
|
24
|
|
|
25 CUDA is supported since llvm 3.9. Current release of clang (7.0.0) supports CUDA
|
|
|
26 7.0 through 9.2. If you need support for CUDA 10, you will need to use clang
|
|
|
27 built from r342924 or newer.
|
|
|
28
|
|
|
29 Before you build CUDA code, you'll need to have installed the appropriate driver
|
|
|
30 for your nvidia GPU and the CUDA SDK. See `NVIDIA's CUDA installation guide
|
|
|
31 <https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html>`_ for
|
|
|
32 details. Note that clang `does not support
|
|
|
33 <https://llvm.org/bugs/show_bug.cgi?id=26966>`_ the CUDA toolkit as installed by
|
|
|
34 many Linux package managers; you probably need to install CUDA in a single
|
|
|
35 directory from NVIDIA's package.
|
|
|
36
|
|
|
37 CUDA compilation is supported on Linux. Compilation on MacOS and Windows may or
|
|
|
38 may not work and currently have no maintainers. Compilation with CUDA-9.x is
|
|
|
39 `currently broken on Windows <https://bugs.llvm.org/show_bug.cgi?id=38811>`_.
|
|
|
40
|
|
|
41 Invoking clang
|
|
|
42 --------------
|
|
|
43
|
|
|
44 Invoking clang for CUDA compilation works similarly to compiling regular C++.
|
|
|
45 You just need to be aware of a few additional flags.
|
|
|
46
|
|
|
47 You can use `this <https://gist.github.com/855e277884eb6b388cd2f00d956c2fd4>`_
|
|
|
48 program as a toy example. Save it as ``axpy.cu``. (Clang detects that you're
|
|
|
49 compiling CUDA code by noticing that your filename ends with ``.cu``.
|
|
|
50 Alternatively, you can pass ``-x cuda``.)
|
|
|
51
|
|
|
52 To build and run, run the following commands, filling in the parts in angle
|
|
|
53 brackets as described below:
|
|
|
54
|
|
|
55 .. code-block:: console
|
|
|
56
|
|
|
57 $ clang++ axpy.cu -o axpy --cuda-gpu-arch=<GPU arch> \
|
|
|
58 -L<CUDA install path>/<lib64 or lib> \
|
|
|
59 -lcudart_static -ldl -lrt -pthread
|
|
|
60 $ ./axpy
|
|
|
61 y[0] = 2
|
|
|
62 y[1] = 4
|
|
|
63 y[2] = 6
|
|
|
64 y[3] = 8
|
|
|
65
|
|
|
66 On MacOS, replace `-lcudart_static` with `-lcudart`; otherwise, you may get
|
|
|
67 "CUDA driver version is insufficient for CUDA runtime version" errors when you
|
|
|
68 run your program.
|
|
|
69
|
|
|
70 * ``<CUDA install path>`` -- the directory where you installed CUDA SDK.
|
|
|
71 Typically, ``/usr/local/cuda``.
|
|
|
72
|
|
|
73 Pass e.g. ``-L/usr/local/cuda/lib64`` if compiling in 64-bit mode; otherwise,
|
|
|
74 pass e.g. ``-L/usr/local/cuda/lib``. (In CUDA, the device code and host code
|
|
|
75 always have the same pointer widths, so if you're compiling 64-bit code for
|
|
|
76 the host, you're also compiling 64-bit code for the device.) Note that as of
|
|
|
77 v10.0 CUDA SDK `no longer supports compilation of 32-bit
|
|
|
78 applications <https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#deprecated-features>`_.
|
|
|
79
|
|
|
80 * ``<GPU arch>`` -- the `compute capability
|
|
|
81 <https://developer.nvidia.com/cuda-gpus>`_ of your GPU. For example, if you
|
|
|
82 want to run your program on a GPU with compute capability of 3.5, specify
|
|
|
83 ``--cuda-gpu-arch=sm_35``.
|
|
|
84
|
|
|
85 Note: You cannot pass ``compute_XX`` as an argument to ``--cuda-gpu-arch``;
|
|
|
86 only ``sm_XX`` is currently supported. However, clang always includes PTX in
|
|
|
87 its binaries, so e.g. a binary compiled with ``--cuda-gpu-arch=sm_30`` would be
|
|
|
88 forwards-compatible with e.g. ``sm_35`` GPUs.
|
|
|
89
|
|
|
90 You can pass ``--cuda-gpu-arch`` multiple times to compile for multiple archs.
|
|
|
91
|
|
|
92 The `-L` and `-l` flags only need to be passed when linking. When compiling,
|
|
|
93 you may also need to pass ``--cuda-path=/path/to/cuda`` if you didn't install
|
|
|
94 the CUDA SDK into ``/usr/local/cuda`` or ``/usr/local/cuda-X.Y``.
|
|
|
95
|
|
|
96 Flags that control numerical code
|
|
|
97 ---------------------------------
|
|
|
98
|
|
|
99 If you're using GPUs, you probably care about making numerical code run fast.
|
|
|
100 GPU hardware allows for more control over numerical operations than most CPUs,
|
|
|
101 but this results in more compiler options for you to juggle.
|
|
|
102
|
|
|
103 Flags you may wish to tweak include:
|
|
|
104
|
|
|
105 * ``-ffp-contract={on,off,fast}`` (defaults to ``fast`` on host and device when
|
|
|
106 compiling CUDA) Controls whether the compiler emits fused multiply-add
|
|
|
107 operations.
|
|
|
108
|
|
|
109 * ``off``: never emit fma operations, and prevent ptxas from fusing multiply
|
|
|
110 and add instructions.
|
|
|
111 * ``on``: fuse multiplies and adds within a single statement, but never
|
|
|
112 across statements (C11 semantics). Prevent ptxas from fusing other
|
|
|
113 multiplies and adds.
|
|
|
114 * ``fast``: fuse multiplies and adds wherever profitable, even across
|
|
|
115 statements. Doesn't prevent ptxas from fusing additional multiplies and
|
|
|
116 adds.
|
|
|
117
|
|
|
118 Fused multiply-add instructions can be much faster than the unfused
|
|
|
119 equivalents, but because the intermediate result in an fma is not rounded,
|
|
|
120 this flag can affect numerical code.
|
|
|
121
|
|
|
122 * ``-fcuda-flush-denormals-to-zero`` (default: off) When this is enabled,
|
|
|
123 floating point operations may flush `denormal
|
|
|
124 <https://en.wikipedia.org/wiki/Denormal_number>`_ inputs and/or outputs to 0.
|
|
|
125 Operations on denormal numbers are often much slower than the same operations
|
|
|
126 on normal numbers.
|
|
|
127
|
|
|
128 * ``-fcuda-approx-transcendentals`` (default: off) When this is enabled, the
|
|
|
129 compiler may emit calls to faster, approximate versions of transcendental
|
|
|
130 functions, instead of using the slower, fully IEEE-compliant versions. For
|
|
|
131 example, this flag allows clang to emit the ptx ``sin.approx.f32``
|
|
|
132 instruction.
|
|
|
133
|
|
|
134 This is implied by ``-ffast-math``.
|
|
|
135
|
|
|
136 Standard library support
|
|
|
137 ========================
|
|
|
138
|
|
|
139 In clang and nvcc, most of the C++ standard library is not supported on the
|
|
|
140 device side.
|
|
|
141
|
|
|
142 ``<math.h>`` and ``<cmath>``
|
|
|
143 ----------------------------
|
|
|
144
|
|
|
145 In clang, ``math.h`` and ``cmath`` are available and `pass
|
|
|
146 <https://github.com/llvm/llvm-test-suite/blob/master/External/CUDA/math_h.cu>`_
|
|
|
147 `tests
|
|
|
148 <https://github.com/llvm/llvm-test-suite/blob/master/External/CUDA/cmath.cu>`_
|
|
|
149 adapted from libc++'s test suite.
|
|
|
150
|
|
|
151 In nvcc ``math.h`` and ``cmath`` are mostly available. Versions of ``::foof``
|
|
|
152 in namespace std (e.g. ``std::sinf``) are not available, and where the standard
|
|
|
153 calls for overloads that take integral arguments, these are usually not
|
|
|
154 available.
|
|
|
155
|
|
|
156 .. code-block:: c++
|
|
|
157
|
|
|
158 #include <math.h>
|
|
|
159 #include <cmath.h>
|
|
|
160
|
|
|
161 // clang is OK with everything in this function.
|
|
|
162 __device__ void test() {
|
|
|
163 std::sin(0.); // nvcc - ok
|
|
|
164 std::sin(0); // nvcc - error, because no std::sin(int) override is available.
|
|
|
165 sin(0); // nvcc - same as above.
|
|
|
166
|
|
|
167 sinf(0.); // nvcc - ok
|
|
|
168 std::sinf(0.); // nvcc - no such function
|
|
|
169 }
|
|
|
170
|
|
|
171 ``<std::complex>``
|
|
|
172 ------------------
|
|
|
173
|
|
|
174 nvcc does not officially support ``std::complex``. It's an error to use
|
|
|
175 ``std::complex`` in ``__device__`` code, but it often works in ``__host__
|
|
|
176 __device__`` code due to nvcc's interpretation of the "wrong-side rule" (see
|
|
|
177 below). However, we have heard from implementers that it's possible to get
|
|
|
178 into situations where nvcc will omit a call to an ``std::complex`` function,
|
|
|
179 especially when compiling without optimizations.
|
|
|
180
|
|
|
181 As of 2016-11-16, clang supports ``std::complex`` without these caveats. It is
|
|
|
182 tested with libstdc++ 4.8.5 and newer, but is known to work only with libc++
|
|
|
183 newer than 2016-11-16.
|
|
|
184
|
|
|
185 ``<algorithm>``
|
|
|
186 ---------------
|
|
|
187
|
|
|
188 In C++14, many useful functions from ``<algorithm>`` (notably, ``std::min`` and
|
|
|
189 ``std::max``) become constexpr. You can therefore use these in device code,
|
|
|
190 when compiling with clang.
|
|
|
191
|
|
|
192 Detecting clang vs NVCC from code
|
|
|
193 =================================
|
|
|
194
|
|
|
195 Although clang's CUDA implementation is largely compatible with NVCC's, you may
|
|
|
196 still want to detect when you're compiling CUDA code specifically with clang.
|
|
|
197
|
|
|
198 This is tricky, because NVCC may invoke clang as part of its own compilation
|
|
|
199 process! For example, NVCC uses the host compiler's preprocessor when
|
|
|
200 compiling for device code, and that host compiler may in fact be clang.
|
|
|
201
|
|
|
202 When clang is actually compiling CUDA code -- rather than being used as a
|
|
|
203 subtool of NVCC's -- it defines the ``__CUDA__`` macro. ``__CUDA_ARCH__`` is
|
|
|
204 defined only in device mode (but will be defined if NVCC is using clang as a
|
|
|
205 preprocessor). So you can use the following incantations to detect clang CUDA
|
|
|
206 compilation, in host and device modes:
|
|
|
207
|
|
|
208 .. code-block:: c++
|
|
|
209
|
|
|
210 #if defined(__clang__) && defined(__CUDA__) && !defined(__CUDA_ARCH__)
|
|
|
211 // clang compiling CUDA code, host mode.
|
|
|
212 #endif
|
|
|
213
|
|
|
214 #if defined(__clang__) && defined(__CUDA__) && defined(__CUDA_ARCH__)
|
|
|
215 // clang compiling CUDA code, device mode.
|
|
|
216 #endif
|
|
|
217
|
|
|
218 Both clang and nvcc define ``__CUDACC__`` during CUDA compilation. You can
|
|
|
219 detect NVCC specifically by looking for ``__NVCC__``.
|
|
|
220
|
|
|
221 Dialect Differences Between clang and nvcc
|
|
|
222 ==========================================
|
|
|
223
|
|
|
224 There is no formal CUDA spec, and clang and nvcc speak slightly different
|
|
|
225 dialects of the language. Below, we describe some of the differences.
|
|
|
226
|
|
|
227 This section is painful; hopefully you can skip this section and live your life
|
|
|
228 blissfully unaware.
|
|
|
229
|
|
|
230 Compilation Models
|
|
|
231 ------------------
|
|
|
232
|
|
|
233 Most of the differences between clang and nvcc stem from the different
|
|
|
234 compilation models used by clang and nvcc. nvcc uses *split compilation*,
|
|
|
235 which works roughly as follows:
|
|
|
236
|
|
|
237 * Run a preprocessor over the input ``.cu`` file to split it into two source
|
|
|
238 files: ``H``, containing source code for the host, and ``D``, containing
|
|
|
239 source code for the device.
|
|
|
240
|
|
|
241 * For each GPU architecture ``arch`` that we're compiling for, do:
|
|
|
242
|
|
|
243 * Compile ``D`` using nvcc proper. The result of this is a ``ptx`` file for
|
|
|
244 ``P_arch``.
|
|
|
245
|
|
|
246 * Optionally, invoke ``ptxas``, the PTX assembler, to generate a file,
|
|
|
247 ``S_arch``, containing GPU machine code (SASS) for ``arch``.
|
|
|
248
|
|
|
249 * Invoke ``fatbin`` to combine all ``P_arch`` and ``S_arch`` files into a
|
|
|
250 single "fat binary" file, ``F``.
|
|
|
251
|
|
|
252 * Compile ``H`` using an external host compiler (gcc, clang, or whatever you
|
|
|
253 like). ``F`` is packaged up into a header file which is force-included into
|
|
|
254 ``H``; nvcc generates code that calls into this header to e.g. launch
|
|
|
255 kernels.
|
|
|
256
|
|
|
257 clang uses *merged parsing*. This is similar to split compilation, except all
|
|
|
258 of the host and device code is present and must be semantically-correct in both
|
|
|
259 compilation steps.
|
|
|
260
|
|
|
261 * For each GPU architecture ``arch`` that we're compiling for, do:
|
|
|
262
|
|
|
263 * Compile the input ``.cu`` file for device, using clang. ``__host__`` code
|
|
|
264 is parsed and must be semantically correct, even though we're not
|
|
|
265 generating code for the host at this time.
|
|
|
266
|
|
|
267 The output of this step is a ``ptx`` file ``P_arch``.
|
|
|
268
|
|
|
269 * Invoke ``ptxas`` to generate a SASS file, ``S_arch``. Note that, unlike
|
|
|
270 nvcc, clang always generates SASS code.
|
|
|
271
|
|
|
272 * Invoke ``fatbin`` to combine all ``P_arch`` and ``S_arch`` files into a
|
|
|
273 single fat binary file, ``F``.
|
|
|
274
|
|
|
275 * Compile ``H`` using clang. ``__device__`` code is parsed and must be
|
|
|
276 semantically correct, even though we're not generating code for the device
|
|
|
277 at this time.
|
|
|
278
|
|
|
279 ``F`` is passed to this compilation, and clang includes it in a special ELF
|
|
|
280 section, where it can be found by tools like ``cuobjdump``.
|
|
|
281
|
|
|
282 (You may ask at this point, why does clang need to parse the input file
|
|
|
283 multiple times? Why not parse it just once, and then use the AST to generate
|
|
|
284 code for the host and each device architecture?
|
|
|
285
|
|
|
286 Unfortunately this can't work because we have to define different macros during
|
|
|
287 host compilation and during device compilation for each GPU architecture.)
|
|
|
288
|
|
|
289 clang's approach allows it to be highly robust to C++ edge cases, as it doesn't
|
|
|
290 need to decide at an early stage which declarations to keep and which to throw
|
|
|
291 away. But it has some consequences you should be aware of.
|
|
|
292
|
|
|
293 Overloading Based on ``__host__`` and ``__device__`` Attributes
|
|
|
294 ---------------------------------------------------------------
|
|
|
295
|
|
|
296 Let "H", "D", and "HD" stand for "``__host__`` functions", "``__device__``
|
|
|
297 functions", and "``__host__ __device__`` functions", respectively. Functions
|
|
|
298 with no attributes behave the same as H.
|
|
|
299
|
|
|
300 nvcc does not allow you to create H and D functions with the same signature:
|
|
|
301
|
|
|
302 .. code-block:: c++
|
|
|
303
|
|
|
304 // nvcc: error - function "foo" has already been defined
|
|
|
305 __host__ void foo() {}
|
|
|
306 __device__ void foo() {}
|
|
|
307
|
|
|
308 However, nvcc allows you to "overload" H and D functions with different
|
|
|
309 signatures:
|
|
|
310
|
|
|
311 .. code-block:: c++
|
|
|
312
|
|
|
313 // nvcc: no error
|
|
|
314 __host__ void foo(int) {}
|
|
|
315 __device__ void foo() {}
|
|
|
316
|
|
|
317 In clang, the ``__host__`` and ``__device__`` attributes are part of a
|
|
|
318 function's signature, and so it's legal to have H and D functions with
|
|
|
319 (otherwise) the same signature:
|
|
|
320
|
|
|
321 .. code-block:: c++
|
|
|
322
|
|
|
323 // clang: no error
|
|
|
324 __host__ void foo() {}
|
|
|
325 __device__ void foo() {}
|
|
|
326
|
|
|
327 HD functions cannot be overloaded by H or D functions with the same signature:
|
|
|
328
|
|
|
329 .. code-block:: c++
|
|
|
330
|
|
|
331 // nvcc: error - function "foo" has already been defined
|
|
|
332 // clang: error - redefinition of 'foo'
|
|
|
333 __host__ __device__ void foo() {}
|
|
|
334 __device__ void foo() {}
|
|
|
335
|
|
|
336 // nvcc: no error
|
|
|
337 // clang: no error
|
|
|
338 __host__ __device__ void bar(int) {}
|
|
|
339 __device__ void bar() {}
|
|
|
340
|
|
|
341 When resolving an overloaded function, clang considers the host/device
|
|
|
342 attributes of the caller and callee. These are used as a tiebreaker during
|
|
|
343 overload resolution. See `IdentifyCUDAPreference
|
|
|
344 <http://clang.llvm.org/doxygen/SemaCUDA_8cpp.html>`_ for the full set of rules,
|
|
|
345 but at a high level they are:
|
|
|
346
|
|
|
347 * D functions prefer to call other Ds. HDs are given lower priority.
|
|
|
348
|
|
|
349 * Similarly, H functions prefer to call other Hs, or ``__global__`` functions
|
|
|
350 (with equal priority). HDs are given lower priority.
|
|
|
351
|
|
|
352 * HD functions prefer to call other HDs.
|
|
|
353
|
|
|
354 When compiling for device, HDs will call Ds with lower priority than HD, and
|
|
|
355 will call Hs with still lower priority. If it's forced to call an H, the
|
|
|
356 program is malformed if we emit code for this HD function. We call this the
|
|
|
357 "wrong-side rule", see example below.
|
|
|
358
|
|
|
359 The rules are symmetrical when compiling for host.
|
|
|
360
|
|
|
361 Some examples:
|
|
|
362
|
|
|
363 .. code-block:: c++
|
|
|
364
|
|
|
365 __host__ void foo();
|
|
|
366 __device__ void foo();
|
|
|
367
|
|
|
368 __host__ void bar();
|
|
|
369 __host__ __device__ void bar();
|
|
|
370
|
|
|
371 __host__ void test_host() {
|
|
|
372 foo(); // calls H overload
|
|
|
373 bar(); // calls H overload
|
|
|
374 }
|
|
|
375
|
|
|
376 __device__ void test_device() {
|
|
|
377 foo(); // calls D overload
|
|
|
378 bar(); // calls HD overload
|
|
|
379 }
|
|
|
380
|
|
|
381 __host__ __device__ void test_hd() {
|
|
|
382 foo(); // calls H overload when compiling for host, otherwise D overload
|
|
|
383 bar(); // always calls HD overload
|
|
|
384 }
|
|
|
385
|
|
|
386 Wrong-side rule example:
|
|
|
387
|
|
|
388 .. code-block:: c++
|
|
|
389
|
|
|
390 __host__ void host_only();
|
|
|
391
|
|
|
392 // We don't codegen inline functions unless they're referenced by a
|
|
|
393 // non-inline function. inline_hd1() is called only from the host side, so
|
|
|
394 // does not generate an error. inline_hd2() is called from the device side,
|
|
|
395 // so it generates an error.
|
|
|
396 inline __host__ __device__ void inline_hd1() { host_only(); } // no error
|
|
|
397 inline __host__ __device__ void inline_hd2() { host_only(); } // error
|
|
|
398
|
|
|
399 __host__ void host_fn() { inline_hd1(); }
|
|
|
400 __device__ void device_fn() { inline_hd2(); }
|
|
|
401
|
|
|
402 // This function is not inline, so it's always codegen'ed on both the host
|
|
|
403 // and the device. Therefore, it generates an error.
|
|
|
404 __host__ __device__ void not_inline_hd() { host_only(); }
|
|
|
405
|
|
|
406 For the purposes of the wrong-side rule, templated functions also behave like
|
|
|
407 ``inline`` functions: They aren't codegen'ed unless they're instantiated
|
|
|
408 (usually as part of the process of invoking them).
|
|
|
409
|
|
|
410 clang's behavior with respect to the wrong-side rule matches nvcc's, except
|
|
|
411 nvcc only emits a warning for ``not_inline_hd``; device code is allowed to call
|
|
|
412 ``not_inline_hd``. In its generated code, nvcc may omit ``not_inline_hd``'s
|
|
|
413 call to ``host_only`` entirely, or it may try to generate code for
|
|
|
414 ``host_only`` on the device. What you get seems to depend on whether or not
|
|
|
415 the compiler chooses to inline ``host_only``.
|
|
|
416
|
|
|
417 Member functions, including constructors, may be overloaded using H and D
|
|
|
418 attributes. However, destructors cannot be overloaded.
|
|
|
419
|
|
|
420 Using a Different Class on Host/Device
|
|
|
421 --------------------------------------
|
|
|
422
|
|
|
423 Occasionally you may want to have a class with different host/device versions.
|
|
|
424
|
|
|
425 If all of the class's members are the same on the host and device, you can just
|
|
|
426 provide overloads for the class's member functions.
|
|
|
427
|
|
|
428 However, if you want your class to have different members on host/device, you
|
|
|
429 won't be able to provide working H and D overloads in both classes. In this
|
|
|
430 case, clang is likely to be unhappy with you.
|
|
|
431
|
|
|
432 .. code-block:: c++
|
|
|
433
|
|
|
434 #ifdef __CUDA_ARCH__
|
|
|
435 struct S {
|
|
|
436 __device__ void foo() { /* use device_only */ }
|
|
|
437 int device_only;
|
|
|
438 };
|
|
|
439 #else
|
|
|
440 struct S {
|
|
|
441 __host__ void foo() { /* use host_only */ }
|
|
|
442 double host_only;
|
|
|
443 };
|
|
|
444
|
|
|
445 __device__ void test() {
|
|
|
446 S s;
|
|
|
447 // clang generates an error here, because during host compilation, we
|
|
|
448 // have ifdef'ed away the __device__ overload of S::foo(). The __device__
|
|
|
449 // overload must be present *even during host compilation*.
|
|
|
450 S.foo();
|
|
|
451 }
|
|
|
452 #endif
|
|
|
453
|
|
|
454 We posit that you don't really want to have classes with different members on H
|
|
|
455 and D. For example, if you were to pass one of these as a parameter to a
|
|
|
456 kernel, it would have a different layout on H and D, so would not work
|
|
|
457 properly.
|
|
|
458
|
|
|
459 To make code like this compatible with clang, we recommend you separate it out
|
|
|
460 into two classes. If you need to write code that works on both host and
|
|
|
461 device, consider writing an overloaded wrapper function that returns different
|
|
|
462 types on host and device.
|
|
|
463
|
|
|
464 .. code-block:: c++
|
|
|
465
|
|
|
466 struct HostS { ... };
|
|
|
467 struct DeviceS { ... };
|
|
|
468
|
|
|
469 __host__ HostS MakeStruct() { return HostS(); }
|
|
|
470 __device__ DeviceS MakeStruct() { return DeviceS(); }
|
|
|
471
|
|
|
472 // Now host and device code can call MakeStruct().
|
|
|
473
|
|
|
474 Unfortunately, this idiom isn't compatible with nvcc, because it doesn't allow
|
|
|
475 you to overload based on the H/D attributes. Here's an idiom that works with
|
|
|
476 both clang and nvcc:
|
|
|
477
|
|
|
478 .. code-block:: c++
|
|
|
479
|
|
|
480 struct HostS { ... };
|
|
|
481 struct DeviceS { ... };
|
|
|
482
|
|
|
483 #ifdef __NVCC__
|
|
|
484 #ifndef __CUDA_ARCH__
|
|
|
485 __host__ HostS MakeStruct() { return HostS(); }
|
|
|
486 #else
|
|
|
487 __device__ DeviceS MakeStruct() { return DeviceS(); }
|
|
|
488 #endif
|
|
|
489 #else
|
|
|
490 __host__ HostS MakeStruct() { return HostS(); }
|
|
|
491 __device__ DeviceS MakeStruct() { return DeviceS(); }
|
|
|
492 #endif
|
|
|
493
|
|
|
494 // Now host and device code can call MakeStruct().
|
|
|
495
|
|
|
496 Hopefully you don't have to do this sort of thing often.
|
|
|
497
|
|
|
498 Optimizations
|
|
|
499 =============
|
|
|
500
|
|
|
501 Modern CPUs and GPUs are architecturally quite different, so code that's fast
|
|
|
502 on a CPU isn't necessarily fast on a GPU. We've made a number of changes to
|
|
|
503 LLVM to make it generate good GPU code. Among these changes are:
|
|
|
504
|
|
|
505 * `Straight-line scalar optimizations <https://goo.gl/4Rb9As>`_ -- These
|
|
|
506 reduce redundancy within straight-line code.
|
|
|
507
|
|
|
508 * `Aggressive speculative execution
|
|
|
509 <http://llvm.org/docs/doxygen/html/SpeculativeExecution_8cpp_source.html>`_
|
|
|
510 -- This is mainly for promoting straight-line scalar optimizations, which are
|
|
|
511 most effective on code along dominator paths.
|
|
|
512
|
|
|
513 * `Memory space inference
|
|
|
514 <http://llvm.org/doxygen/NVPTXInferAddressSpaces_8cpp_source.html>`_ --
|
|
|
515 In PTX, we can operate on pointers that are in a particular "address space"
|
|
|
516 (global, shared, constant, or local), or we can operate on pointers in the
|
|
|
517 "generic" address space, which can point to anything. Operations in a
|
|
|
518 non-generic address space are faster, but pointers in CUDA are not explicitly
|
|
|
519 annotated with their address space, so it's up to LLVM to infer it where
|
|
|
520 possible.
|
|
|
521
|
|
|
522 * `Bypassing 64-bit divides
|
|
|
523 <http://llvm.org/docs/doxygen/html/BypassSlowDivision_8cpp_source.html>`_ --
|
|
|
524 This was an existing optimization that we enabled for the PTX backend.
|
|
|
525
|
|
|
526 64-bit integer divides are much slower than 32-bit ones on NVIDIA GPUs.
|
|
|
527 Many of the 64-bit divides in our benchmarks have a divisor and dividend
|
|
|
528 which fit in 32-bits at runtime. This optimization provides a fast path for
|
|
|
529 this common case.
|
|
|
530
|
|
|
531 * Aggressive loop unrolling and function inlining -- Loop unrolling and
|
|
|
532 function inlining need to be more aggressive for GPUs than for CPUs because
|
|
|
533 control flow transfer in GPU is more expensive. More aggressive unrolling and
|
|
|
534 inlining also promote other optimizations, such as constant propagation and
|
|
|
535 SROA, which sometimes speed up code by over 10x.
|
|
|
536
|
|
|
537 (Programmers can force unrolling and inline using clang's `loop unrolling pragmas
|
|
|
538 <http://clang.llvm.org/docs/AttributeReference.html#pragma-unroll-pragma-nounroll>`_
|
|
|
539 and ``__attribute__((always_inline))``.)
|
|
|
540
|
|
|
541 Publication
|
|
|
542 ===========
|
|
|
543
|
|
|
544 The team at Google published a paper in CGO 2016 detailing the optimizations
|
|
|
545 they'd made to clang/LLVM. Note that "gpucc" is no longer a meaningful name:
|
|
|
546 The relevant tools are now just vanilla clang/LLVM.
|
|
|
547
|
|
|
548 | `gpucc: An Open-Source GPGPU Compiler <http://dl.acm.org/citation.cfm?id=2854041>`_
|
|
|
549 | Jingyue Wu, Artem Belevich, Eli Bendersky, Mark Heffernan, Chris Leary, Jacques Pienaar, Bjarke Roune, Rob Springer, Xuetian Weng, Robert Hundt
|
|
|
550 | *Proceedings of the 2016 International Symposium on Code Generation and Optimization (CGO 2016)*
|
|
|
551 |
|
|
|
552 | `Slides from the CGO talk <http://wujingyue.github.io/docs/gpucc-talk.pdf>`_
|
|
|
553 |
|
|
|
554 | `Tutorial given at CGO <http://wujingyue.github.io/docs/gpucc-tutorial.pdf>`_
|
|
|
555
|
|
|
556 Obtaining Help
|
|
|
557 ==============
|
|
|
558
|
|
|
559 To obtain help on LLVM in general and its CUDA support, see `the LLVM
|
|
|
560 community <http://llvm.org/docs/#mailing-lists>`_.
|
