Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 1 | ========================== |
| 2 | Auto-Vectorization in LLVM |
| 3 | ========================== |
| 4 | |
Sean Silva | 12ae515 | 2012-12-20 22:42:20 +0000 | [diff] [blame] | 5 | .. contents:: |
| 6 | :local: |
| 7 | |
Nadav Rotem | 3fe91a4 | 2013-04-15 22:21:25 +0000 | [diff] [blame] | 8 | LLVM has two vectorizers: The :ref:`Loop Vectorizer <loop-vectorizer>`, |
Nadav Rotem | 96e0b96 | 2013-04-15 22:11:07 +0000 | [diff] [blame] | 9 | which operates on Loops, and the :ref:`SLP Vectorizer |
Nadav Rotem | b5a8a90 | 2013-06-26 17:59:35 +0000 | [diff] [blame] | 10 | <slp-vectorizer>`. These vectorizers |
Sean Silva | 12ae515 | 2012-12-20 22:42:20 +0000 | [diff] [blame] | 11 | focus on different optimization opportunities and use different techniques. |
Nadav Rotem | 3fe91a4 | 2013-04-15 22:21:25 +0000 | [diff] [blame] | 12 | The SLP vectorizer merges multiple scalars that are found in the code into |
Nadav Rotem | b5a8a90 | 2013-06-26 17:59:35 +0000 | [diff] [blame] | 13 | vectors while the Loop Vectorizer widens instructions in loops |
| 14 | to operate on multiple consecutive iterations. |
Sean Silva | 12ae515 | 2012-12-20 22:42:20 +0000 | [diff] [blame] | 15 | |
Nadav Rotem | 055028f | 2013-08-05 04:27:34 +0000 | [diff] [blame] | 16 | Both the Loop Vectorizer and the SLP Vectorizer are enabled by default. |
| 17 | |
Sean Silva | 12ae515 | 2012-12-20 22:42:20 +0000 | [diff] [blame] | 18 | .. _loop-vectorizer: |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 19 | |
| 20 | The Loop Vectorizer |
| 21 | =================== |
| 22 | |
Nadav Rotem | 649a33e | 2012-12-19 18:04:44 +0000 | [diff] [blame] | 23 | Usage |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 24 | ----- |
Nadav Rotem | 649a33e | 2012-12-19 18:04:44 +0000 | [diff] [blame] | 25 | |
Nadav Rotem | 055028f | 2013-08-05 04:27:34 +0000 | [diff] [blame] | 26 | The Loop Vectorizer is enabled by default, but it can be disabled |
| 27 | through clang using the command line flag: |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 28 | |
| 29 | .. code-block:: console |
| 30 | |
Nadav Rotem | df4381b | 2013-04-08 21:34:49 +0000 | [diff] [blame] | 31 | $ clang ... -fno-vectorize file.c |
Nadav Rotem | 3e6da7e | 2012-12-19 18:02:36 +0000 | [diff] [blame] | 32 | |
Nadav Rotem | 4aa55bb | 2013-01-04 17:49:45 +0000 | [diff] [blame] | 33 | Command line flags |
| 34 | ^^^^^^^^^^^^^^^^^^ |
| 35 | |
| 36 | The loop vectorizer uses a cost model to decide on the optimal vectorization factor |
| 37 | and unroll factor. However, users of the vectorizer can force the vectorizer to use |
| 38 | specific values. Both 'clang' and 'opt' support the flags below. |
| 39 | |
| 40 | Users can control the vectorization SIMD width using the command line flag "-force-vector-width". |
| 41 | |
| 42 | .. code-block:: console |
| 43 | |
| 44 | $ clang -mllvm -force-vector-width=8 ... |
| 45 | $ opt -loop-vectorize -force-vector-width=8 ... |
| 46 | |
Eli Friedman | 8a5dfea | 2017-05-31 23:02:55 +0000 | [diff] [blame] | 47 | Users can control the unroll factor using the command line flag "-force-vector-interleave" |
Nadav Rotem | 4aa55bb | 2013-01-04 17:49:45 +0000 | [diff] [blame] | 48 | |
| 49 | .. code-block:: console |
| 50 | |
Eli Friedman | 8a5dfea | 2017-05-31 23:02:55 +0000 | [diff] [blame] | 51 | $ clang -mllvm -force-vector-interleave=2 ... |
| 52 | $ opt -loop-vectorize -force-vector-interleave=2 ... |
Nadav Rotem | 4aa55bb | 2013-01-04 17:49:45 +0000 | [diff] [blame] | 53 | |
Tyler Nowicki | 9487d2a | 2014-06-27 18:30:08 +0000 | [diff] [blame] | 54 | Pragma loop hint directives |
| 55 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 56 | |
| 57 | The ``#pragma clang loop`` directive allows loop vectorization hints to be |
| 58 | specified for the subsequent for, while, do-while, or c++11 range-based for |
| 59 | loop. The directive allows vectorization and interleaving to be enabled or |
| 60 | disabled. Vector width as well as interleave count can also be manually |
| 61 | specified. The following example explicitly enables vectorization and |
| 62 | interleaving: |
| 63 | |
| 64 | .. code-block:: c++ |
| 65 | |
| 66 | #pragma clang loop vectorize(enable) interleave(enable) |
| 67 | while(...) { |
| 68 | ... |
| 69 | } |
| 70 | |
| 71 | The following example implicitly enables vectorization and interleaving by |
| 72 | specifying a vector width and interleaving count: |
| 73 | |
| 74 | .. code-block:: c++ |
| 75 | |
| 76 | #pragma clang loop vectorize_width(2) interleave_count(2) |
| 77 | for(...) { |
| 78 | ... |
| 79 | } |
| 80 | |
| 81 | See the Clang |
| 82 | `language extensions |
| 83 | <http://clang.llvm.org/docs/LanguageExtensions.html#extensions-for-loop-hint-optimizations>`_ |
| 84 | for details. |
| 85 | |
| 86 | Diagnostics |
| 87 | ----------- |
| 88 | |
| 89 | Many loops cannot be vectorized including loops with complicated control flow, |
| 90 | unvectorizable types, and unvectorizable calls. The loop vectorizer generates |
| 91 | optimization remarks which can be queried using command line options to identify |
| 92 | and diagnose loops that are skipped by the loop-vectorizer. |
| 93 | |
| 94 | Optimization remarks are enabled using: |
| 95 | |
| 96 | ``-Rpass=loop-vectorize`` identifies loops that were successfully vectorized. |
| 97 | |
| 98 | ``-Rpass-missed=loop-vectorize`` identifies loops that failed vectorization and |
| 99 | indicates if vectorization was specified. |
| 100 | |
| 101 | ``-Rpass-analysis=loop-vectorize`` identifies the statements that caused |
Ayal Zaks | f93293e | 2017-05-23 07:08:02 +0000 | [diff] [blame] | 102 | vectorization to fail. If in addition ``-fsave-optimization-record`` is |
| 103 | provided, multiple causes of vectorization failure may be listed (this behavior |
| 104 | might change in the future). |
Tyler Nowicki | 9487d2a | 2014-06-27 18:30:08 +0000 | [diff] [blame] | 105 | |
| 106 | Consider the following loop: |
| 107 | |
| 108 | .. code-block:: c++ |
| 109 | |
| 110 | #pragma clang loop vectorize(enable) |
| 111 | for (int i = 0; i < Length; i++) { |
| 112 | switch(A[i]) { |
| 113 | case 0: A[i] = i*2; break; |
| 114 | case 1: A[i] = i; break; |
| 115 | default: A[i] = 0; |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | The command line ``-Rpass-missed=loop-vectorized`` prints the remark: |
| 120 | |
| 121 | .. code-block:: console |
| 122 | |
| 123 | no_switch.cpp:4:5: remark: loop not vectorized: vectorization is explicitly enabled [-Rpass-missed=loop-vectorize] |
| 124 | |
| 125 | And the command line ``-Rpass-analysis=loop-vectorize`` indicates that the |
| 126 | switch statement cannot be vectorized. |
| 127 | |
| 128 | .. code-block:: console |
| 129 | |
| 130 | no_switch.cpp:4:5: remark: loop not vectorized: loop contains a switch statement [-Rpass-analysis=loop-vectorize] |
| 131 | switch(A[i]) { |
| 132 | ^ |
| 133 | |
| 134 | To ensure line and column numbers are produced include the command line options |
| 135 | ``-gline-tables-only`` and ``-gcolumn-info``. See the Clang `user manual |
| 136 | <http://clang.llvm.org/docs/UsersManual.html#options-to-emit-optimization-reports>`_ |
| 137 | for details |
| 138 | |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 139 | Features |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 140 | -------- |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 141 | |
| 142 | The LLVM Loop Vectorizer has a number of features that allow it to vectorize |
| 143 | complex loops. |
| 144 | |
| 145 | Loops with unknown trip count |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 146 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 147 | |
| 148 | The Loop Vectorizer supports loops with an unknown trip count. |
| 149 | In the loop below, the iteration ``start`` and ``finish`` points are unknown, |
| 150 | and the Loop Vectorizer has a mechanism to vectorize loops that do not start |
Sean Silva | 68d5b27 | 2012-12-20 02:23:25 +0000 | [diff] [blame] | 151 | at zero. In this example, 'n' may not be a multiple of the vector width, and |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 152 | the vectorizer has to execute the last few iterations as scalar code. Keeping |
| 153 | a scalar copy of the loop increases the code size. |
| 154 | |
| 155 | .. code-block:: c++ |
| 156 | |
| 157 | void bar(float *A, float* B, float K, int start, int end) { |
Sean Silva | 9baa6e4 | 2012-12-20 22:47:41 +0000 | [diff] [blame] | 158 | for (int i = start; i < end; ++i) |
| 159 | A[i] *= B[i] + K; |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 160 | } |
| 161 | |
| 162 | Runtime Checks of Pointers |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 163 | ^^^^^^^^^^^^^^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 164 | |
| 165 | In the example below, if the pointers A and B point to consecutive addresses, |
| 166 | then it is illegal to vectorize the code because some elements of A will be |
| 167 | written before they are read from array B. |
| 168 | |
| 169 | Some programmers use the 'restrict' keyword to notify the compiler that the |
| 170 | pointers are disjointed, but in our example, the Loop Vectorizer has no way of |
| 171 | knowing that the pointers A and B are unique. The Loop Vectorizer handles this |
| 172 | loop by placing code that checks, at runtime, if the arrays A and B point to |
| 173 | disjointed memory locations. If arrays A and B overlap, then the scalar version |
Sean Silva | 689858b | 2012-12-20 22:59:36 +0000 | [diff] [blame] | 174 | of the loop is executed. |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 175 | |
| 176 | .. code-block:: c++ |
| 177 | |
| 178 | void bar(float *A, float* B, float K, int n) { |
Sean Silva | 9baa6e4 | 2012-12-20 22:47:41 +0000 | [diff] [blame] | 179 | for (int i = 0; i < n; ++i) |
| 180 | A[i] *= B[i] + K; |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 181 | } |
| 182 | |
| 183 | |
| 184 | Reductions |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 185 | ^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 186 | |
Sean Silva | 689858b | 2012-12-20 22:59:36 +0000 | [diff] [blame] | 187 | In this example the ``sum`` variable is used by consecutive iterations of |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 188 | the loop. Normally, this would prevent vectorization, but the vectorizer can |
Sean Silva | 68d5b27 | 2012-12-20 02:23:25 +0000 | [diff] [blame] | 189 | detect that 'sum' is a reduction variable. The variable 'sum' becomes a vector |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 190 | of integers, and at the end of the loop the elements of the array are added |
Sean Silva | 689858b | 2012-12-20 22:59:36 +0000 | [diff] [blame] | 191 | together to create the correct result. We support a number of different |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 192 | reduction operations, such as addition, multiplication, XOR, AND and OR. |
| 193 | |
| 194 | .. code-block:: c++ |
| 195 | |
| 196 | int foo(int *A, int *B, int n) { |
| 197 | unsigned sum = 0; |
| 198 | for (int i = 0; i < n; ++i) |
Sean Silva | 689858b | 2012-12-20 22:59:36 +0000 | [diff] [blame] | 199 | sum += A[i] + 5; |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 200 | return sum; |
| 201 | } |
| 202 | |
Nadav Rotem | 2a92c10 | 2013-01-08 17:46:30 +0000 | [diff] [blame] | 203 | We support floating point reduction operations when `-ffast-math` is used. |
| 204 | |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 205 | Inductions |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 206 | ^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 207 | |
| 208 | In this example the value of the induction variable ``i`` is saved into an |
| 209 | array. The Loop Vectorizer knows to vectorize induction variables. |
| 210 | |
| 211 | .. code-block:: c++ |
| 212 | |
| 213 | void bar(float *A, float* B, float K, int n) { |
Sean Silva | 9baa6e4 | 2012-12-20 22:47:41 +0000 | [diff] [blame] | 214 | for (int i = 0; i < n; ++i) |
| 215 | A[i] = i; |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 216 | } |
| 217 | |
| 218 | If Conversion |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 219 | ^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 220 | |
| 221 | The Loop Vectorizer is able to "flatten" the IF statement in the code and |
| 222 | generate a single stream of instructions. The Loop Vectorizer supports any |
| 223 | control flow in the innermost loop. The innermost loop may contain complex |
| 224 | nesting of IFs, ELSEs and even GOTOs. |
| 225 | |
| 226 | .. code-block:: c++ |
| 227 | |
| 228 | int foo(int *A, int *B, int n) { |
| 229 | unsigned sum = 0; |
| 230 | for (int i = 0; i < n; ++i) |
| 231 | if (A[i] > B[i]) |
| 232 | sum += A[i] + 5; |
| 233 | return sum; |
| 234 | } |
| 235 | |
| 236 | Pointer Induction Variables |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 237 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 238 | |
| 239 | This example uses the "accumulate" function of the standard c++ library. This |
| 240 | loop uses C++ iterators, which are pointers, and not integer indices. |
| 241 | The Loop Vectorizer detects pointer induction variables and can vectorize |
| 242 | this loop. This feature is important because many C++ programs use iterators. |
| 243 | |
| 244 | .. code-block:: c++ |
| 245 | |
| 246 | int baz(int *A, int n) { |
| 247 | return std::accumulate(A, A + n, 0); |
| 248 | } |
| 249 | |
| 250 | Reverse Iterators |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 251 | ^^^^^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 252 | |
| 253 | The Loop Vectorizer can vectorize loops that count backwards. |
| 254 | |
| 255 | .. code-block:: c++ |
| 256 | |
| 257 | int foo(int *A, int *B, int n) { |
| 258 | for (int i = n; i > 0; --i) |
| 259 | A[i] +=1; |
| 260 | } |
| 261 | |
| 262 | Scatter / Gather |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 263 | ^^^^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 264 | |
Nadav Rotem | f574b88 | 2013-01-03 01:47:02 +0000 | [diff] [blame] | 265 | The Loop Vectorizer can vectorize code that becomes a sequence of scalar instructions |
| 266 | that scatter/gathers memory. |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 267 | |
| 268 | .. code-block:: c++ |
| 269 | |
Arnold Schwaighofer | 6a7d263 | 2014-03-12 23:23:44 +0000 | [diff] [blame] | 270 | int foo(int * A, int * B, int n) { |
| 271 | for (intptr_t i = 0; i < n; ++i) |
Arnold Schwaighofer | 8d46932 | 2014-03-12 23:58:07 +0000 | [diff] [blame] | 272 | A[i] += B[i * 4]; |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 273 | } |
| 274 | |
Arnold Schwaighofer | 6a7d263 | 2014-03-12 23:23:44 +0000 | [diff] [blame] | 275 | In many situations the cost model will inform LLVM that this is not beneficial |
| 276 | and LLVM will only vectorize such code if forced with "-mllvm -force-vector-width=#". |
| 277 | |
Nadav Rotem | af08627 | 2012-12-19 07:36:35 +0000 | [diff] [blame] | 278 | Vectorization of Mixed Types |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 279 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 280 | |
| 281 | The Loop Vectorizer can vectorize programs with mixed types. The Vectorizer |
| 282 | cost model can estimate the cost of the type conversion and decide if |
| 283 | vectorization is profitable. |
| 284 | |
| 285 | .. code-block:: c++ |
| 286 | |
| 287 | int foo(int *A, char *B, int n, int k) { |
Sean Silva | 9baa6e4 | 2012-12-20 22:47:41 +0000 | [diff] [blame] | 288 | for (int i = 0; i < n; ++i) |
Sean Silva | 5e81633 | 2012-12-20 22:49:13 +0000 | [diff] [blame] | 289 | A[i] += 4 * B[i]; |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 290 | } |
| 291 | |
Renato Golin | abafaba | 2013-02-23 13:25:41 +0000 | [diff] [blame] | 292 | Global Structures Alias Analysis |
| 293 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 294 | |
| 295 | Access to global structures can also be vectorized, with alias analysis being |
| 296 | used to make sure accesses don't alias. Run-time checks can also be added on |
| 297 | pointer access to structure members. |
| 298 | |
| 299 | Many variations are supported, but some that rely on undefined behaviour being |
| 300 | ignored (as other compilers do) are still being left un-vectorized. |
| 301 | |
| 302 | .. code-block:: c++ |
| 303 | |
| 304 | struct { int A[100], K, B[100]; } Foo; |
| 305 | |
| 306 | int foo() { |
| 307 | for (int i = 0; i < 100; ++i) |
| 308 | Foo.A[i] = Foo.B[i] + 100; |
| 309 | } |
| 310 | |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 311 | Vectorization of function calls |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 312 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 313 | |
Sanjay Patel | 337d711 | 2019-01-10 17:02:55 +0000 | [diff] [blame] | 314 | The Loop Vectorizer can vectorize intrinsic math functions. |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 315 | See the table below for a list of these functions. |
| 316 | |
| 317 | +-----+-----+---------+ |
| 318 | | pow | exp | exp2 | |
| 319 | +-----+-----+---------+ |
| 320 | | sin | cos | sqrt | |
| 321 | +-----+-----+---------+ |
| 322 | | log |log2 | log10 | |
| 323 | +-----+-----+---------+ |
| 324 | |fabs |floor| ceil | |
| 325 | +-----+-----+---------+ |
| 326 | |fma |trunc|nearbyint| |
| 327 | +-----+-----+---------+ |
Nadav Rotem | f7769e3 | 2012-12-26 06:03:35 +0000 | [diff] [blame] | 328 | | | | fmuladd | |
| 329 | +-----+-----+---------+ |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 330 | |
Sanjay Patel | 337d711 | 2019-01-10 17:02:55 +0000 | [diff] [blame] | 331 | Note that the optimizer may not be able to vectorize math library functions |
| 332 | that correspond to these intrinsics if the library calls access external state |
| 333 | such as "errno". To allow better optimization of C/C++ math library functions, |
| 334 | use "-fno-math-errno". |
| 335 | |
Benjamin Kramer | 19949d8 | 2013-02-28 19:33:46 +0000 | [diff] [blame] | 336 | The loop vectorizer knows about special instructions on the target and will |
| 337 | vectorize a loop containing a function call that maps to the instructions. For |
| 338 | example, the loop below will be vectorized on Intel x86 if the SSE4.1 roundps |
| 339 | instruction is available. |
| 340 | |
| 341 | .. code-block:: c++ |
| 342 | |
| 343 | void foo(float *f) { |
| 344 | for (int i = 0; i != 1024; ++i) |
| 345 | f[i] = floorf(f[i]); |
| 346 | } |
Nadav Rotem | f574b88 | 2013-01-03 01:47:02 +0000 | [diff] [blame] | 347 | |
| 348 | Partial unrolling during vectorization |
| 349 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 350 | |
| 351 | Modern processors feature multiple execution units, and only programs that contain a |
Nadav Rotem | 43f3928 | 2013-01-03 01:56:33 +0000 | [diff] [blame] | 352 | high degree of parallelism can fully utilize the entire width of the machine. |
Nadav Rotem | f574b88 | 2013-01-03 01:47:02 +0000 | [diff] [blame] | 353 | The Loop Vectorizer increases the instruction level parallelism (ILP) by |
| 354 | performing partial-unrolling of loops. |
| 355 | |
| 356 | In the example below the entire array is accumulated into the variable 'sum'. |
Nadav Rotem | 43f3928 | 2013-01-03 01:56:33 +0000 | [diff] [blame] | 357 | This is inefficient because only a single execution port can be used by the processor. |
Nadav Rotem | f574b88 | 2013-01-03 01:47:02 +0000 | [diff] [blame] | 358 | By unrolling the code the Loop Vectorizer allows two or more execution ports |
Nadav Rotem | 43f3928 | 2013-01-03 01:56:33 +0000 | [diff] [blame] | 359 | to be used simultaneously. |
Nadav Rotem | f574b88 | 2013-01-03 01:47:02 +0000 | [diff] [blame] | 360 | |
| 361 | .. code-block:: c++ |
| 362 | |
| 363 | int foo(int *A, int *B, int n) { |
| 364 | unsigned sum = 0; |
| 365 | for (int i = 0; i < n; ++i) |
| 366 | sum += A[i]; |
| 367 | return sum; |
| 368 | } |
| 369 | |
Nadav Rotem | 4aa55bb | 2013-01-04 17:49:45 +0000 | [diff] [blame] | 370 | The Loop Vectorizer uses a cost model to decide when it is profitable to unroll loops. |
| 371 | The decision to unroll the loop depends on the register pressure and the generated code size. |
Nadav Rotem | f574b88 | 2013-01-03 01:47:02 +0000 | [diff] [blame] | 372 | |
Nadav Rotem | 67a6ec8 | 2012-12-19 08:28:24 +0000 | [diff] [blame] | 373 | Performance |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 374 | ----------- |
Nadav Rotem | 67a6ec8 | 2012-12-19 08:28:24 +0000 | [diff] [blame] | 375 | |
Ed Maste | ffc045a | 2015-04-14 20:52:58 +0000 | [diff] [blame] | 376 | This section shows the execution time of Clang on a simple benchmark: |
James Y Knight | 8986b31 | 2019-01-14 22:27:32 +0000 | [diff] [blame] | 377 | `gcc-loops <https://github.com/llvm/llvm-test-suite/tree/master/SingleSource/UnitTests/Vectorizer>`_. |
Sean Silva | 689858b | 2012-12-20 22:59:36 +0000 | [diff] [blame] | 378 | This benchmarks is a collection of loops from the GCC autovectorization |
Nadav Rotem | 3e6da7e | 2012-12-19 18:02:36 +0000 | [diff] [blame] | 379 | `page <http://gcc.gnu.org/projects/tree-ssa/vectorization.html>`_ by Dorit Nuzman. |
Nadav Rotem | 67a6ec8 | 2012-12-19 08:28:24 +0000 | [diff] [blame] | 380 | |
Nadav Rotem | 6d1fc53 | 2012-12-20 00:03:36 +0000 | [diff] [blame] | 381 | The chart below compares GCC-4.7, ICC-13, and Clang-SVN with and without loop vectorization at -O3, tuned for "corei7-avx", running on a Sandybridge iMac. |
Sean Silva | 689858b | 2012-12-20 22:59:36 +0000 | [diff] [blame] | 382 | The Y-axis shows the time in msec. Lower is better. The last column shows the geomean of all the kernels. |
Nadav Rotem | 67a6ec8 | 2012-12-19 08:28:24 +0000 | [diff] [blame] | 383 | |
| 384 | .. image:: gcc-loops.png |
| 385 | |
Nadav Rotem | 13410a1 | 2013-01-04 19:00:42 +0000 | [diff] [blame] | 386 | And Linpack-pc with the same configuration. Result is Mflops, higher is better. |
| 387 | |
| 388 | .. image:: linpack-pc.png |
| 389 | |
Ayal Zaks | 98be03e | 2017-05-29 15:36:23 +0000 | [diff] [blame] | 390 | Ongoing Development Directions |
| 391 | ------------------------------ |
| 392 | |
| 393 | .. toctree:: |
| 394 | :hidden: |
| 395 | |
| 396 | Proposals/VectorizationPlan |
| 397 | |
| 398 | :doc:`Proposals/VectorizationPlan` |
| 399 | Modeling the process and upgrading the infrastructure of LLVM's Loop Vectorizer. |
| 400 | |
Nadav Rotem | fc175d9 | 2013-04-15 05:53:23 +0000 | [diff] [blame] | 401 | .. _slp-vectorizer: |
Sean Silva | 12ae515 | 2012-12-20 22:42:20 +0000 | [diff] [blame] | 402 | |
Nadav Rotem | fc175d9 | 2013-04-15 05:53:23 +0000 | [diff] [blame] | 403 | The SLP Vectorizer |
| 404 | ================== |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 405 | |
Nadav Rotem | 649a33e | 2012-12-19 18:04:44 +0000 | [diff] [blame] | 406 | Details |
Sean Silva | 6241703 | 2012-12-20 02:40:45 +0000 | [diff] [blame] | 407 | ------- |
Nadav Rotem | 649a33e | 2012-12-19 18:04:44 +0000 | [diff] [blame] | 408 | |
Nadav Rotem | fc175d9 | 2013-04-15 05:53:23 +0000 | [diff] [blame] | 409 | The goal of SLP vectorization (a.k.a. superword-level parallelism) is |
Nadav Rotem | b5a8a90 | 2013-06-26 17:59:35 +0000 | [diff] [blame] | 410 | to combine similar independent instructions |
| 411 | into vector instructions. Memory accesses, arithmetic operations, comparison |
| 412 | operations, PHI-nodes, can all be vectorized using this technique. |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 413 | |
| 414 | For example, the following function performs very similar operations on its |
| 415 | inputs (a1, b1) and (a2, b2). The basic-block vectorizer may combine these |
| 416 | into vector operations. |
| 417 | |
| 418 | .. code-block:: c++ |
| 419 | |
Nadav Rotem | fc175d9 | 2013-04-15 05:53:23 +0000 | [diff] [blame] | 420 | void foo(int a1, int a2, int b1, int b2, int *A) { |
| 421 | A[0] = a1*(a1 + b1)/b1 + 50*b1/a1; |
| 422 | A[1] = a2*(a2 + b2)/b2 + 50*b2/a2; |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 423 | } |
| 424 | |
Nadav Rotem | b5a8a90 | 2013-06-26 17:59:35 +0000 | [diff] [blame] | 425 | The SLP-vectorizer processes the code bottom-up, across basic blocks, in search of scalars to combine. |
Nadav Rotem | 59f2af9 | 2012-12-19 07:22:24 +0000 | [diff] [blame] | 426 | |
Nadav Rotem | fc175d9 | 2013-04-15 05:53:23 +0000 | [diff] [blame] | 427 | Usage |
| 428 | ------ |
Nadav Rotem | a15dedb | 2013-04-14 07:42:25 +0000 | [diff] [blame] | 429 | |
Nadav Rotem | 055028f | 2013-08-05 04:27:34 +0000 | [diff] [blame] | 430 | The SLP Vectorizer is enabled by default, but it can be disabled |
Nadav Rotem | fc175d9 | 2013-04-15 05:53:23 +0000 | [diff] [blame] | 431 | through clang using the command line flag: |
Nadav Rotem | a15dedb | 2013-04-14 07:42:25 +0000 | [diff] [blame] | 432 | |
Nadav Rotem | fc175d9 | 2013-04-15 05:53:23 +0000 | [diff] [blame] | 433 | .. code-block:: console |
| 434 | |
Nadav Rotem | 055028f | 2013-08-05 04:27:34 +0000 | [diff] [blame] | 435 | $ clang -fno-slp-vectorize file.c |