Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 1 | ========================= |
Justin Lebar | 7029cb5 | 2016-09-07 20:09:53 +0000 | [diff] [blame] | 2 | Compiling CUDA with clang |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 3 | ========================= |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 4 | |
| 5 | .. contents:: |
| 6 | :local: |
| 7 | |
| 8 | Introduction |
| 9 | ============ |
| 10 | |
Justin Lebar | 7029cb5 | 2016-09-07 20:09:53 +0000 | [diff] [blame] | 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 |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 16 | `CUDA programming guide |
| 17 | <http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html>`_. |
| 18 | |
Justin Lebar | 7029cb5 | 2016-09-07 20:09:53 +0000 | [diff] [blame] | 19 | Compiling CUDA Code |
| 20 | =================== |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 21 | |
Justin Lebar | 7029cb5 | 2016-09-07 20:09:53 +0000 | [diff] [blame] | 22 | Prerequisites |
| 23 | ------------- |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 24 | |
Justin Lebar | 7029cb5 | 2016-09-07 20:09:53 +0000 | [diff] [blame] | 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>`_. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 28 | |
Justin Lebar | 7029cb5 | 2016-09-07 20:09:53 +0000 | [diff] [blame] | 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. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 35 | |
Justin Lebar | 7880141 | 2016-11-18 00:41:40 +0000 | [diff] [blame] | 36 | You will need CUDA 7.0, 7.5, or 8.0 to compile with clang. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 37 | |
Justin Lebar | e2cd288 | 2017-01-05 16:54:28 +0000 | [diff] [blame] | 38 | CUDA compilation is supported on Linux, on MacOS as of 2016-11-18, and on |
| 39 | Windows as of 2017-01-05. |
Justin Lebar | 2d22920 | 2016-11-18 00:42:00 +0000 | [diff] [blame] | 40 | |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 41 | Invoking clang |
| 42 | -------------- |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 43 | |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 44 | Invoking clang for CUDA compilation works similarly to compiling regular C++. |
| 45 | You just need to be aware of a few additional flags. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 46 | |
Justin Lebar | 62d5b01 | 2016-09-07 20:42:24 +0000 | [diff] [blame] | 47 | You can use `this <https://gist.github.com/855e277884eb6b388cd2f00d956c2fd4>`_ |
Justin Lebar | 1c10257 | 2016-09-07 21:46:21 +0000 | [diff] [blame] | 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: |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 54 | |
| 55 | .. code-block:: console |
| 56 | |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 57 | $ clang++ axpy.cu -o axpy --cuda-gpu-arch=<GPU arch> \ |
| 58 | -L<CUDA install path>/<lib64 or lib> \ |
Jingyue Wu | 313496b | 2016-01-30 23:48:47 +0000 | [diff] [blame] | 59 | -lcudart_static -ldl -lrt -pthread |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 60 | $ ./axpy |
| 61 | y[0] = 2 |
| 62 | y[1] = 4 |
| 63 | y[2] = 6 |
| 64 | y[3] = 8 |
| 65 | |
Justin Lebar | e8c45e9 | 2016-11-22 23:13:29 +0000 | [diff] [blame] | 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 | |
Justin Lebar | 1c10257 | 2016-09-07 21:46:21 +0000 | [diff] [blame] | 70 | * ``<CUDA install path>`` -- the directory where you installed CUDA SDK. |
| 71 | Typically, ``/usr/local/cuda``. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 72 | |
Justin Lebar | 1c10257 | 2016-09-07 21:46:21 +0000 | [diff] [blame] | 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.) |
Justin Lebar | 84473cd | 2016-09-07 20:09:46 +0000 | [diff] [blame] | 77 | |
Justin Lebar | 1c10257 | 2016-09-07 21:46:21 +0000 | [diff] [blame] | 78 | * ``<GPU arch>`` -- the `compute capability |
| 79 | <https://developer.nvidia.com/cuda-gpus>`_ of your GPU. For example, if you |
| 80 | want to run your program on a GPU with compute capability of 3.5, specify |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 81 | ``--cuda-gpu-arch=sm_35``. |
Justin Lebar | 32835c8 | 2016-03-21 23:05:15 +0000 | [diff] [blame] | 82 | |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 83 | Note: You cannot pass ``compute_XX`` as an argument to ``--cuda-gpu-arch``; |
| 84 | only ``sm_XX`` is currently supported. However, clang always includes PTX in |
| 85 | its binaries, so e.g. a binary compiled with ``--cuda-gpu-arch=sm_30`` would be |
| 86 | forwards-compatible with e.g. ``sm_35`` GPUs. |
Justin Lebar | 32835c8 | 2016-03-21 23:05:15 +0000 | [diff] [blame] | 87 | |
Justin Lebar | 1c10257 | 2016-09-07 21:46:21 +0000 | [diff] [blame] | 88 | You can pass ``--cuda-gpu-arch`` multiple times to compile for multiple archs. |
Justin Lebar | 32835c8 | 2016-03-21 23:05:15 +0000 | [diff] [blame] | 89 | |
Justin Lebar | b5cb9df | 2016-09-07 21:46:49 +0000 | [diff] [blame] | 90 | The `-L` and `-l` flags only need to be passed when linking. When compiling, |
| 91 | you may also need to pass ``--cuda-path=/path/to/cuda`` if you didn't install |
| 92 | the CUDA SDK into ``/usr/local/cuda``, ``/usr/local/cuda-7.0``, or |
| 93 | ``/usr/local/cuda-7.5``. |
| 94 | |
Justin Lebar | b649e75 | 2016-05-25 23:11:31 +0000 | [diff] [blame] | 95 | Flags that control numerical code |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 96 | --------------------------------- |
Justin Lebar | b649e75 | 2016-05-25 23:11:31 +0000 | [diff] [blame] | 97 | |
| 98 | If you're using GPUs, you probably care about making numerical code run fast. |
| 99 | GPU hardware allows for more control over numerical operations than most CPUs, |
| 100 | but this results in more compiler options for you to juggle. |
| 101 | |
| 102 | Flags you may wish to tweak include: |
| 103 | |
| 104 | * ``-ffp-contract={on,off,fast}`` (defaults to ``fast`` on host and device when |
| 105 | compiling CUDA) Controls whether the compiler emits fused multiply-add |
| 106 | operations. |
| 107 | |
| 108 | * ``off``: never emit fma operations, and prevent ptxas from fusing multiply |
| 109 | and add instructions. |
| 110 | * ``on``: fuse multiplies and adds within a single statement, but never |
| 111 | across statements (C11 semantics). Prevent ptxas from fusing other |
| 112 | multiplies and adds. |
| 113 | * ``fast``: fuse multiplies and adds wherever profitable, even across |
| 114 | statements. Doesn't prevent ptxas from fusing additional multiplies and |
| 115 | adds. |
| 116 | |
| 117 | Fused multiply-add instructions can be much faster than the unfused |
| 118 | equivalents, but because the intermediate result in an fma is not rounded, |
| 119 | this flag can affect numerical code. |
| 120 | |
| 121 | * ``-fcuda-flush-denormals-to-zero`` (default: off) When this is enabled, |
| 122 | floating point operations may flush `denormal |
| 123 | <https://en.wikipedia.org/wiki/Denormal_number>`_ inputs and/or outputs to 0. |
| 124 | Operations on denormal numbers are often much slower than the same operations |
| 125 | on normal numbers. |
| 126 | |
| 127 | * ``-fcuda-approx-transcendentals`` (default: off) When this is enabled, the |
| 128 | compiler may emit calls to faster, approximate versions of transcendental |
| 129 | functions, instead of using the slower, fully IEEE-compliant versions. For |
| 130 | example, this flag allows clang to emit the ptx ``sin.approx.f32`` |
| 131 | instruction. |
| 132 | |
| 133 | This is implied by ``-ffast-math``. |
| 134 | |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 135 | Standard library support |
| 136 | ======================== |
| 137 | |
| 138 | In clang and nvcc, most of the C++ standard library is not supported on the |
| 139 | device side. |
| 140 | |
Justin Lebar | 4856e230 | 2016-09-16 04:14:02 +0000 | [diff] [blame] | 141 | ``<math.h>`` and ``<cmath>`` |
| 142 | ---------------------------- |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 143 | |
| 144 | In clang, ``math.h`` and ``cmath`` are available and `pass |
| 145 | <https://github.com/llvm-mirror/test-suite/blob/master/External/CUDA/math_h.cu>`_ |
| 146 | `tests |
| 147 | <https://github.com/llvm-mirror/test-suite/blob/master/External/CUDA/cmath.cu>`_ |
| 148 | adapted from libc++'s test suite. |
| 149 | |
| 150 | In nvcc ``math.h`` and ``cmath`` are mostly available. Versions of ``::foof`` |
| 151 | in namespace std (e.g. ``std::sinf``) are not available, and where the standard |
| 152 | calls for overloads that take integral arguments, these are usually not |
| 153 | available. |
| 154 | |
| 155 | .. code-block:: c++ |
| 156 | |
| 157 | #include <math.h> |
| 158 | #include <cmath.h> |
| 159 | |
| 160 | // clang is OK with everything in this function. |
| 161 | __device__ void test() { |
| 162 | std::sin(0.); // nvcc - ok |
| 163 | std::sin(0); // nvcc - error, because no std::sin(int) override is available. |
| 164 | sin(0); // nvcc - same as above. |
| 165 | |
| 166 | sinf(0.); // nvcc - ok |
| 167 | std::sinf(0.); // nvcc - no such function |
| 168 | } |
| 169 | |
Justin Lebar | 4856e230 | 2016-09-16 04:14:02 +0000 | [diff] [blame] | 170 | ``<std::complex>`` |
| 171 | ------------------ |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 172 | |
| 173 | nvcc does not officially support ``std::complex``. It's an error to use |
| 174 | ``std::complex`` in ``__device__`` code, but it often works in ``__host__ |
| 175 | __device__`` code due to nvcc's interpretation of the "wrong-side rule" (see |
| 176 | below). However, we have heard from implementers that it's possible to get |
| 177 | into situations where nvcc will omit a call to an ``std::complex`` function, |
| 178 | especially when compiling without optimizations. |
| 179 | |
Justin Lebar | be0cfcc | 2016-11-17 01:03:42 +0000 | [diff] [blame] | 180 | As of 2016-11-16, clang supports ``std::complex`` without these caveats. It is |
| 181 | tested with libstdc++ 4.8.5 and newer, but is known to work only with libc++ |
| 182 | newer than 2016-11-16. |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 183 | |
Justin Lebar | 4856e230 | 2016-09-16 04:14:02 +0000 | [diff] [blame] | 184 | ``<algorithm>`` |
| 185 | --------------- |
| 186 | |
| 187 | In C++14, many useful functions from ``<algorithm>`` (notably, ``std::min`` and |
| 188 | ``std::max``) become constexpr. You can therefore use these in device code, |
| 189 | when compiling with clang. |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 190 | |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 191 | Detecting clang vs NVCC from code |
| 192 | ================================= |
| 193 | |
| 194 | Although clang's CUDA implementation is largely compatible with NVCC's, you may |
| 195 | still want to detect when you're compiling CUDA code specifically with clang. |
| 196 | |
| 197 | This is tricky, because NVCC may invoke clang as part of its own compilation |
| 198 | process! For example, NVCC uses the host compiler's preprocessor when |
| 199 | compiling for device code, and that host compiler may in fact be clang. |
| 200 | |
| 201 | When clang is actually compiling CUDA code -- rather than being used as a |
| 202 | subtool of NVCC's -- it defines the ``__CUDA__`` macro. ``__CUDA_ARCH__`` is |
| 203 | defined only in device mode (but will be defined if NVCC is using clang as a |
| 204 | preprocessor). So you can use the following incantations to detect clang CUDA |
| 205 | compilation, in host and device modes: |
| 206 | |
| 207 | .. code-block:: c++ |
| 208 | |
| 209 | #if defined(__clang__) && defined(__CUDA__) && !defined(__CUDA_ARCH__) |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 210 | // clang compiling CUDA code, host mode. |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 211 | #endif |
| 212 | |
| 213 | #if defined(__clang__) && defined(__CUDA__) && defined(__CUDA_ARCH__) |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 214 | // clang compiling CUDA code, device mode. |
Justin Lebar | 6f04ed9 | 2016-09-07 20:37:41 +0000 | [diff] [blame] | 215 | #endif |
| 216 | |
| 217 | Both clang and nvcc define ``__CUDACC__`` during CUDA compilation. You can |
| 218 | detect NVCC specifically by looking for ``__NVCC__``. |
| 219 | |
Justin Lebar | a4fa359 | 2016-09-15 02:04:32 +0000 | [diff] [blame] | 220 | Dialect Differences Between clang and nvcc |
| 221 | ========================================== |
| 222 | |
| 223 | There is no formal CUDA spec, and clang and nvcc speak slightly different |
| 224 | dialects of the language. Below, we describe some of the differences. |
| 225 | |
| 226 | This section is painful; hopefully you can skip this section and live your life |
| 227 | blissfully unaware. |
| 228 | |
| 229 | Compilation Models |
| 230 | ------------------ |
| 231 | |
| 232 | Most of the differences between clang and nvcc stem from the different |
| 233 | compilation models used by clang and nvcc. nvcc uses *split compilation*, |
| 234 | which works roughly as follows: |
| 235 | |
| 236 | * Run a preprocessor over the input ``.cu`` file to split it into two source |
| 237 | files: ``H``, containing source code for the host, and ``D``, containing |
| 238 | source code for the device. |
| 239 | |
| 240 | * For each GPU architecture ``arch`` that we're compiling for, do: |
| 241 | |
| 242 | * Compile ``D`` using nvcc proper. The result of this is a ``ptx`` file for |
| 243 | ``P_arch``. |
| 244 | |
| 245 | * Optionally, invoke ``ptxas``, the PTX assembler, to generate a file, |
| 246 | ``S_arch``, containing GPU machine code (SASS) for ``arch``. |
| 247 | |
| 248 | * Invoke ``fatbin`` to combine all ``P_arch`` and ``S_arch`` files into a |
| 249 | single "fat binary" file, ``F``. |
| 250 | |
| 251 | * Compile ``H`` using an external host compiler (gcc, clang, or whatever you |
| 252 | like). ``F`` is packaged up into a header file which is force-included into |
| 253 | ``H``; nvcc generates code that calls into this header to e.g. launch |
| 254 | kernels. |
| 255 | |
| 256 | clang uses *merged parsing*. This is similar to split compilation, except all |
| 257 | of the host and device code is present and must be semantically-correct in both |
| 258 | compilation steps. |
| 259 | |
| 260 | * For each GPU architecture ``arch`` that we're compiling for, do: |
| 261 | |
| 262 | * Compile the input ``.cu`` file for device, using clang. ``__host__`` code |
| 263 | is parsed and must be semantically correct, even though we're not |
| 264 | generating code for the host at this time. |
| 265 | |
| 266 | The output of this step is a ``ptx`` file ``P_arch``. |
| 267 | |
| 268 | * Invoke ``ptxas`` to generate a SASS file, ``S_arch``. Note that, unlike |
| 269 | nvcc, clang always generates SASS code. |
| 270 | |
| 271 | * Invoke ``fatbin`` to combine all ``P_arch`` and ``S_arch`` files into a |
| 272 | single fat binary file, ``F``. |
| 273 | |
| 274 | * Compile ``H`` using clang. ``__device__`` code is parsed and must be |
| 275 | semantically correct, even though we're not generating code for the device |
| 276 | at this time. |
| 277 | |
| 278 | ``F`` is passed to this compilation, and clang includes it in a special ELF |
| 279 | section, where it can be found by tools like ``cuobjdump``. |
| 280 | |
| 281 | (You may ask at this point, why does clang need to parse the input file |
| 282 | multiple times? Why not parse it just once, and then use the AST to generate |
| 283 | code for the host and each device architecture? |
| 284 | |
| 285 | Unfortunately this can't work because we have to define different macros during |
| 286 | host compilation and during device compilation for each GPU architecture.) |
| 287 | |
| 288 | clang's approach allows it to be highly robust to C++ edge cases, as it doesn't |
| 289 | need to decide at an early stage which declarations to keep and which to throw |
| 290 | away. But it has some consequences you should be aware of. |
| 291 | |
| 292 | Overloading Based on ``__host__`` and ``__device__`` Attributes |
| 293 | --------------------------------------------------------------- |
| 294 | |
| 295 | Let "H", "D", and "HD" stand for "``__host__`` functions", "``__device__`` |
| 296 | functions", and "``__host__ __device__`` functions", respectively. Functions |
| 297 | with no attributes behave the same as H. |
| 298 | |
| 299 | nvcc does not allow you to create H and D functions with the same signature: |
| 300 | |
| 301 | .. code-block:: c++ |
| 302 | |
| 303 | // nvcc: error - function "foo" has already been defined |
| 304 | __host__ void foo() {} |
| 305 | __device__ void foo() {} |
| 306 | |
| 307 | However, nvcc allows you to "overload" H and D functions with different |
| 308 | signatures: |
| 309 | |
| 310 | .. code-block:: c++ |
| 311 | |
| 312 | // nvcc: no error |
| 313 | __host__ void foo(int) {} |
| 314 | __device__ void foo() {} |
| 315 | |
| 316 | In clang, the ``__host__`` and ``__device__`` attributes are part of a |
| 317 | function's signature, and so it's legal to have H and D functions with |
| 318 | (otherwise) the same signature: |
| 319 | |
| 320 | .. code-block:: c++ |
| 321 | |
| 322 | // clang: no error |
| 323 | __host__ void foo() {} |
| 324 | __device__ void foo() {} |
| 325 | |
| 326 | HD functions cannot be overloaded by H or D functions with the same signature: |
| 327 | |
| 328 | .. code-block:: c++ |
| 329 | |
| 330 | // nvcc: error - function "foo" has already been defined |
| 331 | // clang: error - redefinition of 'foo' |
| 332 | __host__ __device__ void foo() {} |
| 333 | __device__ void foo() {} |
| 334 | |
| 335 | // nvcc: no error |
| 336 | // clang: no error |
| 337 | __host__ __device__ void bar(int) {} |
| 338 | __device__ void bar() {} |
| 339 | |
| 340 | When resolving an overloaded function, clang considers the host/device |
| 341 | attributes of the caller and callee. These are used as a tiebreaker during |
| 342 | overload resolution. See `IdentifyCUDAPreference |
| 343 | <http://clang.llvm.org/doxygen/SemaCUDA_8cpp.html>`_ for the full set of rules, |
| 344 | but at a high level they are: |
| 345 | |
| 346 | * D functions prefer to call other Ds. HDs are given lower priority. |
| 347 | |
| 348 | * Similarly, H functions prefer to call other Hs, or ``__global__`` functions |
| 349 | (with equal priority). HDs are given lower priority. |
| 350 | |
| 351 | * HD functions prefer to call other HDs. |
| 352 | |
| 353 | When compiling for device, HDs will call Ds with lower priority than HD, and |
| 354 | will call Hs with still lower priority. If it's forced to call an H, the |
| 355 | program is malformed if we emit code for this HD function. We call this the |
| 356 | "wrong-side rule", see example below. |
| 357 | |
| 358 | The rules are symmetrical when compiling for host. |
| 359 | |
| 360 | Some examples: |
| 361 | |
| 362 | .. code-block:: c++ |
| 363 | |
| 364 | __host__ void foo(); |
| 365 | __device__ void foo(); |
| 366 | |
| 367 | __host__ void bar(); |
| 368 | __host__ __device__ void bar(); |
| 369 | |
| 370 | __host__ void test_host() { |
| 371 | foo(); // calls H overload |
| 372 | bar(); // calls H overload |
| 373 | } |
| 374 | |
| 375 | __device__ void test_device() { |
| 376 | foo(); // calls D overload |
| 377 | bar(); // calls HD overload |
| 378 | } |
| 379 | |
| 380 | __host__ __device__ void test_hd() { |
| 381 | foo(); // calls H overload when compiling for host, otherwise D overload |
| 382 | bar(); // always calls HD overload |
| 383 | } |
| 384 | |
| 385 | Wrong-side rule example: |
| 386 | |
| 387 | .. code-block:: c++ |
| 388 | |
| 389 | __host__ void host_only(); |
| 390 | |
| 391 | // We don't codegen inline functions unless they're referenced by a |
| 392 | // non-inline function. inline_hd1() is called only from the host side, so |
| 393 | // does not generate an error. inline_hd2() is called from the device side, |
| 394 | // so it generates an error. |
| 395 | inline __host__ __device__ void inline_hd1() { host_only(); } // no error |
| 396 | inline __host__ __device__ void inline_hd2() { host_only(); } // error |
| 397 | |
| 398 | __host__ void host_fn() { inline_hd1(); } |
| 399 | __device__ void device_fn() { inline_hd2(); } |
| 400 | |
| 401 | // This function is not inline, so it's always codegen'ed on both the host |
| 402 | // and the device. Therefore, it generates an error. |
| 403 | __host__ __device__ void not_inline_hd() { host_only(); } |
| 404 | |
| 405 | For the purposes of the wrong-side rule, templated functions also behave like |
| 406 | ``inline`` functions: They aren't codegen'ed unless they're instantiated |
| 407 | (usually as part of the process of invoking them). |
| 408 | |
| 409 | clang's behavior with respect to the wrong-side rule matches nvcc's, except |
| 410 | nvcc only emits a warning for ``not_inline_hd``; device code is allowed to call |
| 411 | ``not_inline_hd``. In its generated code, nvcc may omit ``not_inline_hd``'s |
| 412 | call to ``host_only`` entirely, or it may try to generate code for |
| 413 | ``host_only`` on the device. What you get seems to depend on whether or not |
| 414 | the compiler chooses to inline ``host_only``. |
| 415 | |
| 416 | Member functions, including constructors, may be overloaded using H and D |
| 417 | attributes. However, destructors cannot be overloaded. |
| 418 | |
| 419 | Using a Different Class on Host/Device |
| 420 | -------------------------------------- |
| 421 | |
| 422 | Occasionally you may want to have a class with different host/device versions. |
| 423 | |
| 424 | If all of the class's members are the same on the host and device, you can just |
| 425 | provide overloads for the class's member functions. |
| 426 | |
| 427 | However, if you want your class to have different members on host/device, you |
| 428 | won't be able to provide working H and D overloads in both classes. In this |
| 429 | case, clang is likely to be unhappy with you. |
| 430 | |
| 431 | .. code-block:: c++ |
| 432 | |
| 433 | #ifdef __CUDA_ARCH__ |
| 434 | struct S { |
| 435 | __device__ void foo() { /* use device_only */ } |
| 436 | int device_only; |
| 437 | }; |
| 438 | #else |
| 439 | struct S { |
| 440 | __host__ void foo() { /* use host_only */ } |
| 441 | double host_only; |
| 442 | }; |
| 443 | |
| 444 | __device__ void test() { |
| 445 | S s; |
| 446 | // clang generates an error here, because during host compilation, we |
| 447 | // have ifdef'ed away the __device__ overload of S::foo(). The __device__ |
| 448 | // overload must be present *even during host compilation*. |
| 449 | S.foo(); |
| 450 | } |
| 451 | #endif |
| 452 | |
| 453 | We posit that you don't really want to have classes with different members on H |
| 454 | and D. For example, if you were to pass one of these as a parameter to a |
| 455 | kernel, it would have a different layout on H and D, so would not work |
| 456 | properly. |
| 457 | |
| 458 | To make code like this compatible with clang, we recommend you separate it out |
| 459 | into two classes. If you need to write code that works on both host and |
| 460 | device, consider writing an overloaded wrapper function that returns different |
| 461 | types on host and device. |
| 462 | |
| 463 | .. code-block:: c++ |
| 464 | |
| 465 | struct HostS { ... }; |
| 466 | struct DeviceS { ... }; |
| 467 | |
| 468 | __host__ HostS MakeStruct() { return HostS(); } |
| 469 | __device__ DeviceS MakeStruct() { return DeviceS(); } |
| 470 | |
| 471 | // Now host and device code can call MakeStruct(). |
| 472 | |
| 473 | Unfortunately, this idiom isn't compatible with nvcc, because it doesn't allow |
| 474 | you to overload based on the H/D attributes. Here's an idiom that works with |
| 475 | both clang and nvcc: |
| 476 | |
| 477 | .. code-block:: c++ |
| 478 | |
| 479 | struct HostS { ... }; |
| 480 | struct DeviceS { ... }; |
| 481 | |
| 482 | #ifdef __NVCC__ |
| 483 | #ifndef __CUDA_ARCH__ |
| 484 | __host__ HostS MakeStruct() { return HostS(); } |
| 485 | #else |
| 486 | __device__ DeviceS MakeStruct() { return DeviceS(); } |
| 487 | #endif |
| 488 | #else |
| 489 | __host__ HostS MakeStruct() { return HostS(); } |
| 490 | __device__ DeviceS MakeStruct() { return DeviceS(); } |
| 491 | #endif |
| 492 | |
| 493 | // Now host and device code can call MakeStruct(). |
| 494 | |
| 495 | Hopefully you don't have to do this sort of thing often. |
| 496 | |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 497 | Optimizations |
| 498 | ============= |
| 499 | |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 500 | Modern CPUs and GPUs are architecturally quite different, so code that's fast |
| 501 | on a CPU isn't necessarily fast on a GPU. We've made a number of changes to |
| 502 | LLVM to make it generate good GPU code. Among these changes are: |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 503 | |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 504 | * `Straight-line scalar optimizations <https://goo.gl/4Rb9As>`_ -- These |
| 505 | reduce redundancy within straight-line code. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 506 | |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 507 | * `Aggressive speculative execution |
| 508 | <http://llvm.org/docs/doxygen/html/SpeculativeExecution_8cpp_source.html>`_ |
| 509 | -- This is mainly for promoting straight-line scalar optimizations, which are |
| 510 | most effective on code along dominator paths. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 511 | |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 512 | * `Memory space inference |
| 513 | <http://llvm.org/doxygen/NVPTXInferAddressSpaces_8cpp_source.html>`_ -- |
| 514 | In PTX, we can operate on pointers that are in a paricular "address space" |
| 515 | (global, shared, constant, or local), or we can operate on pointers in the |
| 516 | "generic" address space, which can point to anything. Operations in a |
| 517 | non-generic address space are faster, but pointers in CUDA are not explicitly |
| 518 | annotated with their address space, so it's up to LLVM to infer it where |
| 519 | possible. |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 520 | |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 521 | * `Bypassing 64-bit divides |
| 522 | <http://llvm.org/docs/doxygen/html/BypassSlowDivision_8cpp_source.html>`_ -- |
| 523 | This was an existing optimization that we enabled for the PTX backend. |
| 524 | |
| 525 | 64-bit integer divides are much slower than 32-bit ones on NVIDIA GPUs. |
| 526 | Many of the 64-bit divides in our benchmarks have a divisor and dividend |
| 527 | which fit in 32-bits at runtime. This optimization provides a fast path for |
| 528 | this common case. |
| 529 | |
| 530 | * Aggressive loop unrooling and function inlining -- Loop unrolling and |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 531 | function inlining need to be more aggressive for GPUs than for CPUs because |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 532 | control flow transfer in GPU is more expensive. More aggressive unrolling and |
| 533 | inlining also promote other optimizations, such as constant propagation and |
| 534 | SROA, which sometimes speed up code by over 10x. |
| 535 | |
| 536 | (Programmers can force unrolling and inline using clang's `loop unrolling pragmas |
Jingyue Wu | 4f2a6cb | 2015-11-10 22:35:47 +0000 | [diff] [blame] | 537 | <http://clang.llvm.org/docs/AttributeReference.html#pragma-unroll-pragma-nounroll>`_ |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 538 | and ``__attribute__((always_inline))``.) |
Jingyue Wu | bec7818 | 2016-02-23 23:34:49 +0000 | [diff] [blame] | 539 | |
Jingyue Wu | f190ed4 | 2016-03-30 05:05:40 +0000 | [diff] [blame] | 540 | Publication |
| 541 | =========== |
| 542 | |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 543 | The team at Google published a paper in CGO 2016 detailing the optimizations |
| 544 | they'd made to clang/LLVM. Note that "gpucc" is no longer a meaningful name: |
| 545 | The relevant tools are now just vanilla clang/LLVM. |
| 546 | |
Jingyue Wu | f190ed4 | 2016-03-30 05:05:40 +0000 | [diff] [blame] | 547 | | `gpucc: An Open-Source GPGPU Compiler <http://dl.acm.org/citation.cfm?id=2854041>`_ |
| 548 | | Jingyue Wu, Artem Belevich, Eli Bendersky, Mark Heffernan, Chris Leary, Jacques Pienaar, Bjarke Roune, Rob Springer, Xuetian Weng, Robert Hundt |
| 549 | | *Proceedings of the 2016 International Symposium on Code Generation and Optimization (CGO 2016)* |
Justin Lebar | 66feaf9 | 2016-09-07 21:46:53 +0000 | [diff] [blame] | 550 | | |
| 551 | | `Slides from the CGO talk <http://wujingyue.com/docs/gpucc-talk.pdf>`_ |
| 552 | | |
| 553 | | `Tutorial given at CGO <http://wujingyue.com/docs/gpucc-tutorial.pdf>`_ |
Jingyue Wu | f190ed4 | 2016-03-30 05:05:40 +0000 | [diff] [blame] | 554 | |
Jingyue Wu | bec7818 | 2016-02-23 23:34:49 +0000 | [diff] [blame] | 555 | Obtaining Help |
| 556 | ============== |
| 557 | |
| 558 | To obtain help on LLVM in general and its CUDA support, see `the LLVM |
| 559 | community <http://llvm.org/docs/#mailing-lists>`_. |