Rasmus Munk Larsen
3815aeed7a
Parallelize tensor contraction over the inner dimension in cases where where one or both of the outer dimensions (m and n) are small but k is large. This speeds up individual matmul microbenchmarks by up to 85%.
...
Naming below is BM_Matmul_M_K_N_THREADS, measured on a 2-socket Intel Broadwell-based server.
Benchmark Base (ns) New (ns) Improvement
------------------------------------------------------------------
BM_Matmul_1_80_13522_1 387457 396013 -2.2%
BM_Matmul_1_80_13522_2 406487 230789 +43.2%
BM_Matmul_1_80_13522_4 395821 123211 +68.9%
BM_Matmul_1_80_13522_6 391625 97002 +75.2%
BM_Matmul_1_80_13522_8 408986 113828 +72.2%
BM_Matmul_1_80_13522_16 399988 67600 +83.1%
BM_Matmul_1_80_13522_22 411546 60044 +85.4%
BM_Matmul_1_80_13522_32 393528 57312 +85.4%
BM_Matmul_1_80_13522_44 390047 63525 +83.7%
BM_Matmul_1_80_13522_88 387876 63592 +83.6%
BM_Matmul_1_1500_500_1 245359 248119 -1.1%
BM_Matmul_1_1500_500_2 401833 143271 +64.3%
BM_Matmul_1_1500_500_4 210519 100231 +52.4%
BM_Matmul_1_1500_500_6 251582 86575 +65.6%
BM_Matmul_1_1500_500_8 211499 80444 +62.0%
BM_Matmul_3_250_512_1 70297 68551 +2.5%
BM_Matmul_3_250_512_2 70141 52450 +25.2%
BM_Matmul_3_250_512_4 67872 58204 +14.2%
BM_Matmul_3_250_512_6 71378 63340 +11.3%
BM_Matmul_3_250_512_8 69595 41652 +40.2%
BM_Matmul_3_250_512_16 72055 42549 +40.9%
BM_Matmul_3_250_512_22 70158 54023 +23.0%
BM_Matmul_3_250_512_32 71541 56042 +21.7%
BM_Matmul_3_250_512_44 71843 57019 +20.6%
BM_Matmul_3_250_512_88 69951 54045 +22.7%
BM_Matmul_3_1500_512_1 369328 374284 -1.4%
BM_Matmul_3_1500_512_2 428656 223603 +47.8%
BM_Matmul_3_1500_512_4 205599 139508 +32.1%
BM_Matmul_3_1500_512_6 214278 139071 +35.1%
BM_Matmul_3_1500_512_8 184149 142338 +22.7%
BM_Matmul_3_1500_512_16 156462 156983 -0.3%
BM_Matmul_3_1500_512_22 163905 158259 +3.4%
BM_Matmul_3_1500_512_32 155314 157662 -1.5%
BM_Matmul_3_1500_512_44 235434 158657 +32.6%
BM_Matmul_3_1500_512_88 156779 160275 -2.2%
BM_Matmul_1500_4_512_1 363358 349528 +3.8%
BM_Matmul_1500_4_512_2 303134 263319 +13.1%
BM_Matmul_1500_4_512_4 176208 130086 +26.2%
BM_Matmul_1500_4_512_6 148026 115449 +22.0%
BM_Matmul_1500_4_512_8 131656 98421 +25.2%
BM_Matmul_1500_4_512_16 134011 82861 +38.2%
BM_Matmul_1500_4_512_22 134950 85685 +36.5%
BM_Matmul_1500_4_512_32 133165 90081 +32.4%
BM_Matmul_1500_4_512_44 133203 90644 +32.0%
BM_Matmul_1500_4_512_88 134106 100566 +25.0%
BM_Matmul_4_1500_512_1 439243 435058 +1.0%
BM_Matmul_4_1500_512_2 451830 257032 +43.1%
BM_Matmul_4_1500_512_4 276434 164513 +40.5%
BM_Matmul_4_1500_512_6 182542 144827 +20.7%
BM_Matmul_4_1500_512_8 179411 166256 +7.3%
BM_Matmul_4_1500_512_16 158101 155560 +1.6%
BM_Matmul_4_1500_512_22 152435 155448 -1.9%
BM_Matmul_4_1500_512_32 155150 149538 +3.6%
BM_Matmul_4_1500_512_44 193842 149777 +22.7%
BM_Matmul_4_1500_512_88 149544 154468 -3.3%
2018-09-26 16:47:13 -07:00
Gael Guennebaud
b3fd93207b
Fix typos found using codespell
2018-06-07 14:43:02 +02:00
Rasmus Munk Larsen
1b7294f6fc
Fix cut-and-paste error.
2017-09-08 16:35:58 -07:00
Rasmus Munk Larsen
94e2213b38
Avoid undefined behavior in Eigen::TensorCostModel::numThreads.
...
If the cost is large enough then the thread count can be larger than the maximum
representable int, so just casting it to an int is undefined behavior.
Contributed by phurst@google.com .
2017-09-08 15:49:55 -07:00
Benoit Steiner
488ad7dd1b
Added missing EIGEN_DEVICE_FUNC qualifiers
2016-09-14 13:35:00 -07:00
Benoit Steiner
7944d4431f
Made the cost model cwiseMax and cwiseMin methods consts to help the PowerPC cuda compiler compile this code.
2016-08-18 13:46:36 -07:00
Rasmus Munk Larsen
7df811cfe5
Minor cleanups: 1. Get rid of unused variables. 2. Get rid of last uses of EIGEN_USE_COST_MODEL.
2016-05-18 15:09:48 -07:00
Benoit Steiner
83ef39e055
Turn on the cost model by default. This results in some significant speedups for smaller tensors. For example, below are the results for the various tensor reductions.
...
Before:
BM_colReduction_12T/10 1000000 1949 51.29 MFlops/s
BM_colReduction_12T/80 100000 15636 409.29 MFlops/s
BM_colReduction_12T/640 20000 95100 4307.01 MFlops/s
BM_colReduction_12T/4K 500 4573423 5466.36 MFlops/s
BM_colReduction_4T/10 1000000 1867 53.56 MFlops/s
BM_colReduction_4T/80 500000 5288 1210.11 MFlops/s
BM_colReduction_4T/640 10000 106924 3830.75 MFlops/s
BM_colReduction_4T/4K 500 9946374 2513.48 MFlops/s
BM_colReduction_8T/10 1000000 1912 52.30 MFlops/s
BM_colReduction_8T/80 200000 8354 766.09 MFlops/s
BM_colReduction_8T/640 20000 85063 4815.22 MFlops/s
BM_colReduction_8T/4K 500 5445216 4591.19 MFlops/s
BM_rowReduction_12T/10 1000000 2041 48.99 MFlops/s
BM_rowReduction_12T/80 100000 15426 414.87 MFlops/s
BM_rowReduction_12T/640 50000 39117 10470.98 MFlops/s
BM_rowReduction_12T/4K 500 3034298 8239.14 MFlops/s
BM_rowReduction_4T/10 1000000 1834 54.51 MFlops/s
BM_rowReduction_4T/80 500000 5406 1183.81 MFlops/s
BM_rowReduction_4T/640 50000 35017 11697.16 MFlops/s
BM_rowReduction_4T/4K 500 3428527 7291.76 MFlops/s
BM_rowReduction_8T/10 1000000 1925 51.95 MFlops/s
BM_rowReduction_8T/80 200000 8519 751.23 MFlops/s
BM_rowReduction_8T/640 50000 33441 12248.42 MFlops/s
BM_rowReduction_8T/4K 1000 2852841 8763.19 MFlops/s
After:
BM_colReduction_12T/10 50000000 59 1678.30 MFlops/s
BM_colReduction_12T/80 5000000 725 8822.71 MFlops/s
BM_colReduction_12T/640 20000 90882 4506.93 MFlops/s
BM_colReduction_12T/4K 500 4668855 5354.63 MFlops/s
BM_colReduction_4T/10 50000000 59 1687.37 MFlops/s
BM_colReduction_4T/80 5000000 737 8681.24 MFlops/s
BM_colReduction_4T/640 50000 108637 3770.34 MFlops/s
BM_colReduction_4T/4K 500 7912954 3159.38 MFlops/s
BM_colReduction_8T/10 50000000 60 1657.21 MFlops/s
BM_colReduction_8T/80 5000000 726 8812.48 MFlops/s
BM_colReduction_8T/640 20000 91451 4478.90 MFlops/s
BM_colReduction_8T/4K 500 5441692 4594.16 MFlops/s
BM_rowReduction_12T/10 20000000 93 1065.28 MFlops/s
BM_rowReduction_12T/80 2000000 950 6730.96 MFlops/s
BM_rowReduction_12T/640 50000 38196 10723.48 MFlops/s
BM_rowReduction_12T/4K 500 3019217 8280.29 MFlops/s
BM_rowReduction_4T/10 20000000 93 1064.30 MFlops/s
BM_rowReduction_4T/80 2000000 959 6667.71 MFlops/s
BM_rowReduction_4T/640 50000 37433 10941.96 MFlops/s
BM_rowReduction_4T/4K 500 3036476 8233.23 MFlops/s
BM_rowReduction_8T/10 20000000 93 1072.47 MFlops/s
BM_rowReduction_8T/80 2000000 959 6670.04 MFlops/s
BM_rowReduction_8T/640 50000 38069 10759.37 MFlops/s
BM_rowReduction_8T/4K 1000 2758988 9061.29 MFlops/s
2016-05-16 08:55:21 -07:00
Benoit Steiner
09653e1f82
Improved the portability of the tensor code
2016-05-11 23:29:09 -07:00
Benoit Steiner
968ec1c2ae
Use numext::isfinite instead of std::isfinite
2016-05-03 19:56:40 -07:00
Benoit Steiner
c07404f6a1
Restore Tensor support for non c++11 compilers
2016-04-29 15:19:19 -07:00
Rasmus Munk Larsen
07ac4f7e02
Eigen Tensor cost model part 2: Thread scheduling for standard evaluators and reductions. The cost model is turned off by default.
2016-04-14 18:28:23 -07:00
Rasmus Munk Larsen
aeb5494a0b
Improvements to cost model.
2016-04-14 15:52:58 -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
