Rasmus Munk Larsen
edaa0fc5d1
Revert PR-292. After further investigation, the memcpy->memmove change was only good for Haswell on older versions of glibc. Adding a switch for small sizes is perhaps useful for string copies, but also has an overhead for larger sizes, making it a poor trade-off for general memcpy.
...
This PR also removes a couple of unnecessary semi-colons in Eigen/src/Core/AssignEvaluator.h that caused compiler warning everywhere.
2017-01-26 12:46:06 -08:00
Rasmus Munk Larsen
e6b1020221
Adds a fast memcpy function to Eigen. This takes advantage of the following:
...
1. For small fixed sizes, the compiler generates inline code for memcpy, which is much faster.
2. My colleague eriche at googl dot com discovered that for large sizes, memmove is significantly faster than memcpy (at least on Linux with GCC or Clang). See benchmark numbers measured on a Haswell (HP Z440) workstation here: https://docs.google.com/a/google.com/spreadsheets/d/1jLs5bKzXwhpTySw65MhG1pZpsIwkszZqQTjwrd_n0ic/pubhtml This is of course surprising since memcpy is a less constrained version of memmove. This stackoverflow thread contains some speculation as to the causes: http://stackoverflow.com/questions/22793669/poor-memcpy-performance-on-linux
Below are numbers for copying and slicing tensors using the multithreaded TensorDevice. The numbers show significant improvements for memcpy of very small blocks and for memcpy of large blocks single threaded (we were already able to saturate memory bandwidth for >1 threads before on large blocks). The "slicingSmallPieces" benchmark also shows small consistent improvements, since memcpy cost is a fair portion of that particular computation.
The benchmarks operate on NxN matrices, and the names are of the form BM_$OP_${NUMTHREADS}T/${N}.
Measured improvements in wall clock time:
Run on rmlarsen3.mtv (12 X 3501 MHz CPUs); 2017-01-20T11:26:31.493023454-08:00
CPU: Intel Haswell with HyperThreading (6 cores) dL1:32KB dL2:256KB dL3:15MB
Benchmark Base (ns) New (ns) Improvement
------------------------------------------------------------------
BM_memcpy_1T/2 3.48 2.39 +31.3%
BM_memcpy_1T/8 12.3 6.51 +47.0%
BM_memcpy_1T/64 371 383 -3.2%
BM_memcpy_1T/512 66922 66720 +0.3%
BM_memcpy_1T/4k 9892867 6849682 +30.8%
BM_memcpy_1T/5k 14951099 10332856 +30.9%
BM_memcpy_2T/2 3.50 2.46 +29.7%
BM_memcpy_2T/8 12.3 7.66 +37.7%
BM_memcpy_2T/64 371 376 -1.3%
BM_memcpy_2T/512 66652 66788 -0.2%
BM_memcpy_2T/4k 6145012 6117776 +0.4%
BM_memcpy_2T/5k 9181478 9010942 +1.9%
BM_memcpy_4T/2 3.47 2.47 +31.0%
BM_memcpy_4T/8 12.3 6.67 +45.8
BM_memcpy_4T/64 374 376 -0.5%
BM_memcpy_4T/512 67833 68019 -0.3%
BM_memcpy_4T/4k 5057425 5188253 -2.6%
BM_memcpy_4T/5k 7555638 7779468 -3.0%
BM_memcpy_6T/2 3.51 2.50 +28.8%
BM_memcpy_6T/8 12.3 7.61 +38.1%
BM_memcpy_6T/64 373 378 -1.3%
BM_memcpy_6T/512 66871 66774 +0.1%
BM_memcpy_6T/4k 5112975 5233502 -2.4%
BM_memcpy_6T/5k 7614180 7772246 -2.1%
BM_memcpy_8T/2 3.47 2.41 +30.5%
BM_memcpy_8T/8 12.4 10.5 +15.3%
BM_memcpy_8T/64 372 388 -4.3%
BM_memcpy_8T/512 67373 66588 +1.2%
BM_memcpy_8T/4k 5148462 5254897 -2.1%
BM_memcpy_8T/5k 7660989 7799058 -1.8%
BM_memcpy_12T/2 3.50 2.40 +31.4%
BM_memcpy_12T/8 12.4 7.55 +39.1
BM_memcpy_12T/64 374 378 -1.1%
BM_memcpy_12T/512 67132 66683 +0.7%
BM_memcpy_12T/4k 5185125 5292920 -2.1%
BM_memcpy_12T/5k 7717284 7942684 -2.9%
BM_slicingSmallPieces_1T/2 47.3 47.5 +0.4%
BM_slicingSmallPieces_1T/8 53.6 52.3 +2.4%
BM_slicingSmallPieces_1T/64 491 476 +3.1%
BM_slicingSmallPieces_1T/512 21734 18814 +13.4%
BM_slicingSmallPieces_1T/4k 394660 396760 -0.5%
BM_slicingSmallPieces_1T/5k 218722 209244 +4.3%
BM_slicingSmallPieces_2T/2 80.7 79.9 +1.0%
BM_slicingSmallPieces_2T/8 54.2 53.1 +2.0
BM_slicingSmallPieces_2T/64 497 477 +4.0%
BM_slicingSmallPieces_2T/512 21732 18822 +13.4%
BM_slicingSmallPieces_2T/4k 392885 390490 +0.6%
BM_slicingSmallPieces_2T/5k 221988 208678 +6.0%
BM_slicingSmallPieces_4T/2 80.8 80.1 +0.9%
BM_slicingSmallPieces_4T/8 54.1 53.2 +1.7%
BM_slicingSmallPieces_4T/64 493 476 +3.4%
BM_slicingSmallPieces_4T/512 21702 18758 +13.6%
BM_slicingSmallPieces_4T/4k 393962 404023 -2.6%
BM_slicingSmallPieces_4T/5k 249667 211732 +15.2%
BM_slicingSmallPieces_6T/2 80.5 80.1 +0.5%
BM_slicingSmallPieces_6T/8 54.4 53.4 +1.8%
BM_slicingSmallPieces_6T/64 488 478 +2.0%
BM_slicingSmallPieces_6T/512 21719 18841 +13.3%
BM_slicingSmallPieces_6T/4k 394950 397583 -0.7%
BM_slicingSmallPieces_6T/5k 223080 210148 +5.8%
BM_slicingSmallPieces_8T/2 81.2 80.4 +1.0%
BM_slicingSmallPieces_8T/8 58.1 53.5 +7.9%
BM_slicingSmallPieces_8T/64 489 480 +1.8%
BM_slicingSmallPieces_8T/512 21586 18798 +12.9%
BM_slicingSmallPieces_8T/4k 394592 400165 -1.4%
BM_slicingSmallPieces_8T/5k 219688 208301 +5.2%
BM_slicingSmallPieces_12T/2 80.2 79.8 +0.7%
BM_slicingSmallPieces_12T/8 54.4 53.4 +1.8
BM_slicingSmallPieces_12T/64 488 476 +2.5%
BM_slicingSmallPieces_12T/512 21931 18831 +14.1%
BM_slicingSmallPieces_12T/4k 393962 396541 -0.7%
BM_slicingSmallPieces_12T/5k 218803 207965 +5.0%
2017-01-24 13:55:18 -08:00
Benoit Steiner
4236aebe10
Simplified the contraction code`
2016-12-21 16:42:56 -08:00
Benoit Steiner
f9eff17e91
Leverage libxsmm kernels within signle threaded contractions
2016-12-21 12:32:06 -08:00
Mehdi Goli
2d4a091beb
Adding tensor contraction operation backend for Sycl; adding test for contractionOp sycl backend; adding temporary solution to prevent memory leak in buffer; cleaning up cxx11_tensor_buildins_sycl.h
2016-12-14 15:30:37 +00:00
Benoit Steiner
c53f783705
Updated the contraction code to support constant inputs.
2016-09-01 11:41:27 -07:00
Benoit Steiner
2b17f34574
Properly detect the type of the result of a contraction.
2016-08-16 16:00:30 -07:00
Benoit Steiner
a071629fec
Replace implicit cast with an explicit one
2016-05-12 10:40:07 -07:00
Benoit Steiner
f363e533aa
Added tests for full contractions using thread pools and gpu devices.
...
Fixed a couple of issues in the corresponding code.
2016-05-05 09:05:45 -07:00
Benoit Steiner
06d774bf58
Updated the contraction code to ensure that full contraction return a tensor of rank 0
2016-05-05 08:37:47 -07:00
Benoit Steiner
17fe7f354e
Deleted trailing commas
2016-04-29 18:39:01 -07:00
Benoit Steiner
e5f71aa6b2
Deleted useless trailing commas
2016-04-29 18:36:10 -07:00
Benoit Steiner
c8e8f93d6c
Move the evalGemm method into the TensorContractionEvaluatorBase class to make it accessible from both the single and multithreaded contraction evaluators.
2016-04-15 16:48:10 -07:00
Benoit Steiner
6c43c49e4a
Fixed a few compilation warnings
2016-04-15 15:34:34 -07:00
Rasmus Munk Larsen
235e83aba6
Eigen cost model part 1. This implements a basic recursive framework to estimate the cost of evaluating tensor expressions.
2016-04-14 13:57:35 -07:00
Benoit Steiner
f7329619da
Fix bug in tensor contraction. The code assumes that contraction axis indices for the LHS (after possibly swapping to ColMajor!) is increasing. Explicitly sort the contraction axis pairs to make it so.
2016-03-17 15:08:02 -07:00
Benoit Steiner
8768c063f5
Fixed the tensor chipping code.
2016-03-08 12:26:49 -08:00
Benoit Steiner
e09eb835db
Decoupled the packet type definition from the definition of the tensor ops. All the vectorization is now defined in the tensor evaluators. This will make it possible to relialably support devices with different packet types in the same compilation unit.
2016-03-08 12:07:33 -08:00
Benoit Steiner
963f2d2a8f
Marked several methods EIGEN_DEVICE_FUNC
2016-01-28 23:37:48 -08:00
Benoit Steiner
3aeeca32af
Leverage the new blocking code in the tensor contraction code.
2016-01-22 16:36:30 -08:00
Benoit Steiner
7ce932edd3
Small cleanup and small fix to the contraction of row major tensors
2016-01-20 18:12:08 -08:00
Benoit Steiner
47076bf00e
Reduce the register pressure exerted by the tensor mappers whenever possible. This improves the performance of the contraction of a matrix with a vector by about 35%.
2016-01-20 14:51:48 -08:00
Benoit Steiner
6d472d8375
Moved the contraction mapping code to its own file to make the code more manageable.
2016-01-19 17:22:05 -08:00
Benoit Steiner
79b69b7444
Trigger the optimized matrix vector path more conservatively.
2016-01-12 15:21:09 -08:00
Benoit Steiner
d920d57f38
Improved the performance of the contraction of a 2d tensor with a 1d tensor by a factor of 3 or more. This helps speedup LSTM neural networks.
2016-01-12 11:32:27 -08:00
Benoit Steiner
20e2ab1121
Fixed another compilation warning
2015-12-07 16:17:57 -08:00
Benoit Steiner
7dfe75f445
Fixed compilation warnings
2015-12-07 08:12:30 -08:00
Benoit Steiner
f4ca8ad917
Use signed integers instead of unsigned ones more consistently in the codebase.
2015-12-04 18:14:16 -08:00
Benoit Steiner
6a02c2a85d
Fixed a compilation warning
2015-10-29 20:21:29 -07:00
Benoit Steiner
a586fdaa91
Reworked the tensor contraction mapper code to make it compile on Android
2015-10-23 09:33:41 -07:00
Benoit Steiner
2ab603316a
Use numext::mini/numext::maxi instead of std::min/std::max in the tensor code
2015-08-28 08:14:15 -07:00
Gael Guennebaud
aec4814370
Many files were missing in previous changeset.
2015-07-29 11:11:23 +02:00
Godeffroy Valet
2195822df6
Allowed tensor contraction operation with an empty array of dimension pairs, which performs a tensor product.
2015-07-25 11:58:36 +02:00
Benoit Steiner
3f2101b03b
Use numext::swap instead of std::swap
2015-07-06 17:02:29 -07:00
Benoit Steiner
3625734bc8
Moved some utilities to TensorMeta.h to make it easier to reuse them accross several tensor operations.
...
Created the TensorDimensionList class to encode the list of all the dimensions of a tensor of rank n. This could be done using TensorIndexList, however TensorIndexList require cxx11 which isn't yet supported as widely as we'd like.
2015-06-29 10:49:55 -07:00
Benoit Steiner
3b429b71e6
Fixed compilation warning triggered by gcc 4.7
2015-04-18 13:41:06 -07:00
Benoit Steiner
bb483313f6
Fixed another batch of compilation warnings
2015-02-28 02:32:46 -08:00
Benoit Steiner
114e863f08
Silcenced a few compilation warnings
2015-02-10 12:20:24 -08:00
Benoit Steiner
410895a7e4
Silenced several compilation warnings
2015-02-10 12:13:19 -08:00
Benoit Steiner
e896c0ade7
Marked the contraction operation as read only, since its result can't be assigned.
2015-01-29 10:29:47 -08:00
Benoit Steiner
71676eaddd
Added support for RowMajor inputs to the contraction code.
2015-01-14 12:36:57 -08:00
Benoit Steiner
1d3c8306f8
Fixed compilation errors with clang.
...
H: Enter commit message. Lines beginning with 'HG:' are removed.
2014-11-13 19:13:17 -08:00
Benoit Steiner
b1789c112b
Improved handling of 1d tensors
2014-11-03 08:51:33 -08:00
Benoit Steiner
65af852b54
Silenced one last warning
2014-10-16 15:02:30 -07:00
Benoit Steiner
ae697b471c
Silenced a few compilation warnings
...
Generalized a TensorMap constructor
2014-10-16 14:52:50 -07:00
Benoit Steiner
bfdd9f3ac9
Made the blocking computation aware of the l3 cache
...
Also optimized the blocking parameters to take into account the number of threads used for a computation
2014-10-15 15:32:59 -07:00
Benoit Steiner
1269392822
Created the IndexPair type to store pair of tensor indices. CUDA doesn't support std::pair so we can't use them when targeting GPUs.
...
Improved the performance on tensor contractions
2014-10-03 10:16:59 -07:00
Benoit Steiner
efdff15749
Fixed a typo in the contraction code
2014-09-06 13:28:24 -07:00
Benoit Steiner
f1d8c13dbc
Fixed misc typos.
2014-08-13 08:40:26 -07:00
Benoit Steiner
9faad2932f
Added missing apis.
2014-08-13 08:36:33 -07:00
