• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2018 Gael Guennebaud <gael.guennebaud@inria.fr>
5 // Copyright (C) 2020, Arm Limited and Contributors
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 #ifndef EIGEN_CONFIGURE_VECTORIZATION_H
12 #define EIGEN_CONFIGURE_VECTORIZATION_H
13 
14 //------------------------------------------------------------------------------------------
15 // Static and dynamic alignment control
16 //
17 // The main purpose of this section is to define EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES
18 // as the maximal boundary in bytes on which dynamically and statically allocated data may be alignment respectively.
19 // The values of EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES can be specified by the user. If not,
20 // a default value is automatically computed based on architecture, compiler, and OS.
21 //
22 // This section also defines macros EIGEN_ALIGN_TO_BOUNDARY(N) and the shortcuts EIGEN_ALIGN{8,16,32,_MAX}
23 // to be used to declare statically aligned buffers.
24 //------------------------------------------------------------------------------------------
25 
26 
27 /* EIGEN_ALIGN_TO_BOUNDARY(n) forces data to be n-byte aligned. This is used to satisfy SIMD requirements.
28  * However, we do that EVEN if vectorization (EIGEN_VECTORIZE) is disabled,
29  * so that vectorization doesn't affect binary compatibility.
30  *
31  * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link
32  * vectorized and non-vectorized code.
33  *
34  * FIXME: this code can be cleaned up once we switch to proper C++11 only.
35  */
36 #if (defined EIGEN_CUDACC)
37   #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n)
38   #define EIGEN_ALIGNOF(x) __alignof(x)
39 #elif EIGEN_HAS_ALIGNAS
40   #define EIGEN_ALIGN_TO_BOUNDARY(n) alignas(n)
41   #define EIGEN_ALIGNOF(x) alignof(x)
42 #elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM
43   #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
44   #define EIGEN_ALIGNOF(x) __alignof(x)
45 #elif EIGEN_COMP_MSVC
46   #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n))
47   #define EIGEN_ALIGNOF(x) __alignof(x)
48 #elif EIGEN_COMP_SUNCC
49   // FIXME not sure about this one:
50   #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
51   #define EIGEN_ALIGNOF(x) __alignof(x)
52 #else
53   #error Please tell me what is the equivalent of alignas(n) and alignof(x) for your compiler
54 #endif
55 
56 // If the user explicitly disable vectorization, then we also disable alignment
57 #if defined(EIGEN_DONT_VECTORIZE)
58   #if defined(EIGEN_GPUCC)
59     // GPU code is always vectorized and requires memory alignment for
60     // statically allocated buffers.
61     #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16
62   #else
63     #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0
64   #endif
65 #elif defined(__AVX512F__)
66   // 64 bytes static alignment is preferred only if really required
67   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64
68 #elif defined(__AVX__)
69   // 32 bytes static alignment is preferred only if really required
70   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 32
71 #else
72   #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16
73 #endif
74 
75 
76 // EIGEN_MIN_ALIGN_BYTES defines the minimal value for which the notion of explicit alignment makes sense
77 #define EIGEN_MIN_ALIGN_BYTES 16
78 
79 // Defined the boundary (in bytes) on which the data needs to be aligned. Note
80 // that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be
81 // aligned at all regardless of the value of this #define.
82 
83 #if (defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN))  && defined(EIGEN_MAX_STATIC_ALIGN_BYTES) && EIGEN_MAX_STATIC_ALIGN_BYTES>0
84 #error EIGEN_MAX_STATIC_ALIGN_BYTES and EIGEN_DONT_ALIGN[_STATICALLY] are both defined with EIGEN_MAX_STATIC_ALIGN_BYTES!=0. Use EIGEN_MAX_STATIC_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN_STATICALLY.
85 #endif
86 
87 // EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprecated
88 // They imply EIGEN_MAX_STATIC_ALIGN_BYTES=0
89 #if defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN)
90   #ifdef EIGEN_MAX_STATIC_ALIGN_BYTES
91     #undef EIGEN_MAX_STATIC_ALIGN_BYTES
92   #endif
93   #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
94 #endif
95 
96 #ifndef EIGEN_MAX_STATIC_ALIGN_BYTES
97 
98   // Try to automatically guess what is the best default value for EIGEN_MAX_STATIC_ALIGN_BYTES
99 
100   // 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
101   // 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
102   // enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
103   // certain common platform (compiler+architecture combinations) to avoid these problems.
104   // Only static alignment is really problematic (relies on nonstandard compiler extensions),
105   // try to keep heap alignment even when we have to disable static alignment.
106   #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64 || EIGEN_ARCH_MIPS)
107   #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
108   #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6)
109   // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support.
110   // Not sure which version fixed it, hopefully it doesn't affect 4.7, which is still somewhat in use.
111   // 4.8 and newer seem definitely unaffected.
112   #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
113   #else
114   #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
115   #endif
116 
117   // static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX
118   #if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \
119   && !EIGEN_GCC3_OR_OLDER \
120   && !EIGEN_COMP_SUNCC \
121   && !EIGEN_OS_QNX
122     #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1
123   #else
124     #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0
125   #endif
126 
127   #if EIGEN_ARCH_WANTS_STACK_ALIGNMENT
128     #define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
129   #else
130     #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
131   #endif
132 
133 #endif
134 
135 // If EIGEN_MAX_ALIGN_BYTES is defined, then it is considered as an upper bound for EIGEN_MAX_STATIC_ALIGN_BYTES
136 #if defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES<EIGEN_MAX_STATIC_ALIGN_BYTES
137 #undef EIGEN_MAX_STATIC_ALIGN_BYTES
138 #define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
139 #endif
140 
141 #if EIGEN_MAX_STATIC_ALIGN_BYTES==0 && !defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT)
142   #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
143 #endif
144 
145 // At this stage, EIGEN_MAX_STATIC_ALIGN_BYTES>0 is the true test whether we want to align arrays on the stack or not.
146 // It takes into account both the user choice to explicitly enable/disable alignment (by setting EIGEN_MAX_STATIC_ALIGN_BYTES)
147 // and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT).
148 // Henceforth, only EIGEN_MAX_STATIC_ALIGN_BYTES should be used.
149 
150 
151 // Shortcuts to EIGEN_ALIGN_TO_BOUNDARY
152 #define EIGEN_ALIGN8  EIGEN_ALIGN_TO_BOUNDARY(8)
153 #define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16)
154 #define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32)
155 #define EIGEN_ALIGN64 EIGEN_ALIGN_TO_BOUNDARY(64)
156 #if EIGEN_MAX_STATIC_ALIGN_BYTES>0
157 #define EIGEN_ALIGN_MAX EIGEN_ALIGN_TO_BOUNDARY(EIGEN_MAX_STATIC_ALIGN_BYTES)
158 #else
159 #define EIGEN_ALIGN_MAX
160 #endif
161 
162 
163 // Dynamic alignment control
164 
165 #if defined(EIGEN_DONT_ALIGN) && defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES>0
166 #error EIGEN_MAX_ALIGN_BYTES and EIGEN_DONT_ALIGN are both defined with EIGEN_MAX_ALIGN_BYTES!=0. Use EIGEN_MAX_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN.
167 #endif
168 
169 #ifdef EIGEN_DONT_ALIGN
170   #ifdef EIGEN_MAX_ALIGN_BYTES
171     #undef EIGEN_MAX_ALIGN_BYTES
172   #endif
173   #define EIGEN_MAX_ALIGN_BYTES 0
174 #elif !defined(EIGEN_MAX_ALIGN_BYTES)
175   #define EIGEN_MAX_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
176 #endif
177 
178 #if EIGEN_IDEAL_MAX_ALIGN_BYTES > EIGEN_MAX_ALIGN_BYTES
179 #define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
180 #else
181 #define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
182 #endif
183 
184 
185 #ifndef EIGEN_UNALIGNED_VECTORIZE
186 #define EIGEN_UNALIGNED_VECTORIZE 1
187 #endif
188 
189 //----------------------------------------------------------------------
190 
191 // if alignment is disabled, then disable vectorization. Note: EIGEN_MAX_ALIGN_BYTES is the proper check, it takes into
192 // account both the user's will (EIGEN_MAX_ALIGN_BYTES,EIGEN_DONT_ALIGN) and our own platform checks
193 #if EIGEN_MAX_ALIGN_BYTES==0
194   #ifndef EIGEN_DONT_VECTORIZE
195     #define EIGEN_DONT_VECTORIZE
196   #endif
197 #endif
198 
199 
200 // The following (except #include <malloc.h> and _M_IX86_FP ??) can likely be
201 // removed as gcc 4.1 and msvc 2008 are not supported anyways.
202 #if EIGEN_COMP_MSVC
203   #include <malloc.h> // for _aligned_malloc -- need it regardless of whether vectorization is enabled
204   #if (EIGEN_COMP_MSVC >= 1500) // 2008 or later
205     // a user reported that in 64-bit mode, MSVC doesn't care to define _M_IX86_FP.
206     #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || EIGEN_ARCH_x86_64
207       #define EIGEN_SSE2_ON_MSVC_2008_OR_LATER
208     #endif
209   #endif
210 #else
211   #if (defined __SSE2__) && ( (!EIGEN_COMP_GNUC) || EIGEN_COMP_ICC || EIGEN_GNUC_AT_LEAST(4,2) )
212     #define EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC
213   #endif
214 #endif
215 
216 #if !(defined(EIGEN_DONT_VECTORIZE) || defined(EIGEN_GPUCC))
217 
218   #if defined (EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC) || defined(EIGEN_SSE2_ON_MSVC_2008_OR_LATER)
219 
220     // Defines symbols for compile-time detection of which instructions are
221     // used.
222     // EIGEN_VECTORIZE_YY is defined if and only if the instruction set YY is used
223     #define EIGEN_VECTORIZE
224     #define EIGEN_VECTORIZE_SSE
225     #define EIGEN_VECTORIZE_SSE2
226 
227     // Detect sse3/ssse3/sse4:
228     // gcc and icc defines __SSE3__, ...
229     // there is no way to know about this on msvc. You can define EIGEN_VECTORIZE_SSE* if you
230     // want to force the use of those instructions with msvc.
231     #ifdef __SSE3__
232       #define EIGEN_VECTORIZE_SSE3
233     #endif
234     #ifdef __SSSE3__
235       #define EIGEN_VECTORIZE_SSSE3
236     #endif
237     #ifdef __SSE4_1__
238       #define EIGEN_VECTORIZE_SSE4_1
239     #endif
240     #ifdef __SSE4_2__
241       #define EIGEN_VECTORIZE_SSE4_2
242     #endif
243     #ifdef __AVX__
244       #ifndef EIGEN_USE_SYCL
245         #define EIGEN_VECTORIZE_AVX
246       #endif
247       #define EIGEN_VECTORIZE_SSE3
248       #define EIGEN_VECTORIZE_SSSE3
249       #define EIGEN_VECTORIZE_SSE4_1
250       #define EIGEN_VECTORIZE_SSE4_2
251     #endif
252     #ifdef __AVX2__
253       #ifndef EIGEN_USE_SYCL
254         #define EIGEN_VECTORIZE_AVX2
255         #define EIGEN_VECTORIZE_AVX
256       #endif
257       #define EIGEN_VECTORIZE_SSE3
258       #define EIGEN_VECTORIZE_SSSE3
259       #define EIGEN_VECTORIZE_SSE4_1
260       #define EIGEN_VECTORIZE_SSE4_2
261     #endif
262     #if defined(__FMA__) || (EIGEN_COMP_MSVC && defined(__AVX2__))
263       // MSVC does not expose a switch dedicated for FMA
264       // For MSVC, AVX2 => FMA
265       #define EIGEN_VECTORIZE_FMA
266     #endif
267     #if defined(__AVX512F__)
268       #ifndef EIGEN_VECTORIZE_FMA
269       #if EIGEN_COMP_GNUC
270       #error Please add -mfma to your compiler flags: compiling with -mavx512f alone without SSE/AVX FMA is not supported (bug 1638).
271       #else
272       #error Please enable FMA in your compiler flags (e.g. -mfma): compiling with AVX512 alone without SSE/AVX FMA is not supported (bug 1638).
273       #endif
274       #endif
275       #ifndef EIGEN_USE_SYCL
276         #define EIGEN_VECTORIZE_AVX512
277         #define EIGEN_VECTORIZE_AVX2
278         #define EIGEN_VECTORIZE_AVX
279       #endif
280       #define EIGEN_VECTORIZE_FMA
281       #define EIGEN_VECTORIZE_SSE3
282       #define EIGEN_VECTORIZE_SSSE3
283       #define EIGEN_VECTORIZE_SSE4_1
284       #define EIGEN_VECTORIZE_SSE4_2
285       #ifndef EIGEN_USE_SYCL
286         #ifdef __AVX512DQ__
287           #define EIGEN_VECTORIZE_AVX512DQ
288         #endif
289         #ifdef __AVX512ER__
290           #define EIGEN_VECTORIZE_AVX512ER
291         #endif
292         #ifdef __AVX512BF16__
293           #define EIGEN_VECTORIZE_AVX512BF16
294         #endif
295       #endif
296     #endif
297 
298     // Disable AVX support on broken xcode versions
299     #if defined(__apple_build_version__) && (__apple_build_version__ == 11000033 ) && ( __MAC_OS_X_VERSION_MIN_REQUIRED == 101500 )
300       // A nasty bug in the clang compiler shipped with xcode in a common compilation situation
301       // when XCode 11.0 and Mac deployment target macOS 10.15 is https://trac.macports.org/ticket/58776#no1
302       #ifdef EIGEN_VECTORIZE_AVX
303         #undef EIGEN_VECTORIZE_AVX
304         #warning "Disabling AVX support: clang compiler shipped with XCode 11.[012] generates broken assembly with -macosx-version-min=10.15 and AVX enabled. "
305         #ifdef EIGEN_VECTORIZE_AVX2
306           #undef EIGEN_VECTORIZE_AVX2
307         #endif
308         #ifdef EIGEN_VECTORIZE_FMA
309           #undef EIGEN_VECTORIZE_FMA
310         #endif
311         #ifdef EIGEN_VECTORIZE_AVX512
312           #undef EIGEN_VECTORIZE_AVX512
313         #endif
314         #ifdef EIGEN_VECTORIZE_AVX512DQ
315           #undef EIGEN_VECTORIZE_AVX512DQ
316         #endif
317         #ifdef EIGEN_VECTORIZE_AVX512ER
318           #undef EIGEN_VECTORIZE_AVX512ER
319         #endif
320       #endif
321       // NOTE: Confirmed test failures in XCode 11.0, and XCode 11.2 with  -macosx-version-min=10.15 and AVX
322       // NOTE using -macosx-version-min=10.15 with Xcode 11.0 results in runtime segmentation faults in many tests, 11.2 produce core dumps in 3 tests
323       // NOTE using -macosx-version-min=10.14 produces functioning and passing tests in all cases
324       // NOTE __clang_version__ "11.0.0 (clang-1100.0.33.8)"  XCode 11.0 <- Produces many segfault and core dumping tests
325       //                                                                    with  -macosx-version-min=10.15 and AVX
326       // NOTE __clang_version__ "11.0.0 (clang-1100.0.33.12)" XCode 11.2 <- Produces 3 core dumping tests with
327       //                                                                    -macosx-version-min=10.15 and AVX
328     #endif
329 
330     // include files
331 
332     // This extern "C" works around a MINGW-w64 compilation issue
333     // https://sourceforge.net/tracker/index.php?func=detail&aid=3018394&group_id=202880&atid=983354
334     // In essence, intrin.h is included by windows.h and also declares intrinsics (just as emmintrin.h etc. below do).
335     // However, intrin.h uses an extern "C" declaration, and g++ thus complains of duplicate declarations
336     // with conflicting linkage.  The linkage for intrinsics doesn't matter, but at that stage the compiler doesn't know;
337     // so, to avoid compile errors when windows.h is included after Eigen/Core, ensure intrinsics are extern "C" here too.
338     // notice that since these are C headers, the extern "C" is theoretically needed anyways.
339     extern "C" {
340       // In theory we should only include immintrin.h and not the other *mmintrin.h header files directly.
341       // Doing so triggers some issues with ICC. However old gcc versions seems to not have this file, thus:
342       #if EIGEN_COMP_ICC >= 1110
343         #include <immintrin.h>
344       #else
345         #include <mmintrin.h>
346         #include <emmintrin.h>
347         #include <xmmintrin.h>
348         #ifdef  EIGEN_VECTORIZE_SSE3
349         #include <pmmintrin.h>
350         #endif
351         #ifdef EIGEN_VECTORIZE_SSSE3
352         #include <tmmintrin.h>
353         #endif
354         #ifdef EIGEN_VECTORIZE_SSE4_1
355         #include <smmintrin.h>
356         #endif
357         #ifdef EIGEN_VECTORIZE_SSE4_2
358         #include <nmmintrin.h>
359         #endif
360         #if defined(EIGEN_VECTORIZE_AVX) || defined(EIGEN_VECTORIZE_AVX512)
361         #include <immintrin.h>
362         #endif
363       #endif
364     } // end extern "C"
365 
366   #elif defined __VSX__
367 
368     #define EIGEN_VECTORIZE
369     #define EIGEN_VECTORIZE_VSX
370     #include <altivec.h>
371     // We need to #undef all these ugly tokens defined in <altivec.h>
372     // => use __vector instead of vector
373     #undef bool
374     #undef vector
375     #undef pixel
376 
377   #elif defined __ALTIVEC__
378 
379     #define EIGEN_VECTORIZE
380     #define EIGEN_VECTORIZE_ALTIVEC
381     #include <altivec.h>
382     // We need to #undef all these ugly tokens defined in <altivec.h>
383     // => use __vector instead of vector
384     #undef bool
385     #undef vector
386     #undef pixel
387 
388   #elif ((defined  __ARM_NEON) || (defined __ARM_NEON__)) && !(defined EIGEN_ARM64_USE_SVE)
389 
390     #define EIGEN_VECTORIZE
391     #define EIGEN_VECTORIZE_NEON
392     #include <arm_neon.h>
393 
394   // We currently require SVE to be enabled explicitly via EIGEN_ARM64_USE_SVE and
395   // will not select the backend automatically
396   #elif (defined __ARM_FEATURE_SVE) && (defined EIGEN_ARM64_USE_SVE)
397 
398     #define EIGEN_VECTORIZE
399     #define EIGEN_VECTORIZE_SVE
400     #include <arm_sve.h>
401 
402     // Since we depend on knowing SVE vector lengths at compile-time, we need
403     // to ensure a fixed lengths is set
404     #if defined __ARM_FEATURE_SVE_BITS
405       #define EIGEN_ARM64_SVE_VL __ARM_FEATURE_SVE_BITS
406     #else
407 #error "Eigen requires a fixed SVE lector length but EIGEN_ARM64_SVE_VL is not set."
408 #endif
409 
410 #elif (defined __s390x__ && defined __VEC__)
411 
412 #define EIGEN_VECTORIZE
413 #define EIGEN_VECTORIZE_ZVECTOR
414 #include <vecintrin.h>
415 
416 #elif defined __mips_msa
417 
418 // Limit MSA optimizations to little-endian CPUs for now.
419 // TODO: Perhaps, eventually support MSA optimizations on big-endian CPUs?
420 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
421 #if defined(__LP64__)
422 #define EIGEN_MIPS_64
423 #else
424 #define EIGEN_MIPS_32
425 #endif
426 #define EIGEN_VECTORIZE
427 #define EIGEN_VECTORIZE_MSA
428 #include <msa.h>
429 #endif
430 
431 #endif
432 #endif
433 
434 // Following the Arm ACLE arm_neon.h should also include arm_fp16.h but not all
435 // compilers seem to follow this. We therefore include it explicitly.
436 // See also: https://bugs.llvm.org/show_bug.cgi?id=47955
437 #if defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
438   #include <arm_fp16.h>
439 #endif
440 
441 #if defined(__F16C__) && (!defined(EIGEN_GPUCC) && (!defined(EIGEN_COMP_CLANG) || EIGEN_COMP_CLANG>=380))
442   // We can use the optimized fp16 to float and float to fp16 conversion routines
443   #define EIGEN_HAS_FP16_C
444 
445   #if defined(EIGEN_COMP_CLANG)
446     // Workaround for clang: The FP16C intrinsics for clang are included by
447     // immintrin.h, as opposed to emmintrin.h as suggested by Intel:
448     // https://software.intel.com/sites/landingpage/IntrinsicsGuide/#othertechs=FP16C&expand=1711
449     #include <immintrin.h>
450   #endif
451 #endif
452 
453 #if defined EIGEN_CUDACC
454   #define EIGEN_VECTORIZE_GPU
455   #include <vector_types.h>
456   #if EIGEN_CUDA_SDK_VER >= 70500
457     #define EIGEN_HAS_CUDA_FP16
458   #endif
459 #endif
460 
461 #if defined(EIGEN_HAS_CUDA_FP16)
462   #include <cuda_runtime_api.h>
463   #include <cuda_fp16.h>
464 #endif
465 
466 #if defined(EIGEN_HIPCC)
467   #define EIGEN_VECTORIZE_GPU
468   #include <hip/hip_vector_types.h>
469   #define EIGEN_HAS_HIP_FP16
470   #include <hip/hip_fp16.h>
471 #endif
472 
473 
474 /** \brief Namespace containing all symbols from the %Eigen library. */
475 namespace Eigen {
476 
SimdInstructionSetsInUse(void)477 inline static const char *SimdInstructionSetsInUse(void) {
478 #if defined(EIGEN_VECTORIZE_AVX512)
479   return "AVX512, FMA, AVX2, AVX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
480 #elif defined(EIGEN_VECTORIZE_AVX)
481   return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
482 #elif defined(EIGEN_VECTORIZE_SSE4_2)
483   return "SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
484 #elif defined(EIGEN_VECTORIZE_SSE4_1)
485   return "SSE, SSE2, SSE3, SSSE3, SSE4.1";
486 #elif defined(EIGEN_VECTORIZE_SSSE3)
487   return "SSE, SSE2, SSE3, SSSE3";
488 #elif defined(EIGEN_VECTORIZE_SSE3)
489   return "SSE, SSE2, SSE3";
490 #elif defined(EIGEN_VECTORIZE_SSE2)
491   return "SSE, SSE2";
492 #elif defined(EIGEN_VECTORIZE_ALTIVEC)
493   return "AltiVec";
494 #elif defined(EIGEN_VECTORIZE_VSX)
495   return "VSX";
496 #elif defined(EIGEN_VECTORIZE_NEON)
497   return "ARM NEON";
498 #elif defined(EIGEN_VECTORIZE_SVE)
499   return "ARM SVE";
500 #elif defined(EIGEN_VECTORIZE_ZVECTOR)
501   return "S390X ZVECTOR";
502 #elif defined(EIGEN_VECTORIZE_MSA)
503   return "MIPS MSA";
504 #else
505   return "None";
506 #endif
507 }
508 
509 } // end namespace Eigen
510 
511 
512 #endif // EIGEN_CONFIGURE_VECTORIZATION_H
513