| Commit message (Collapse) | Author | Age |
|
|
|
| |
TensorInflation.h.
|
| |
|
| |
|
|
|
|
| |
verification operation for cxx11_tensorChipping.cpp test
|
|\ |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
| |
| |
| | |
unused parameter warning for eval_op_indices in TensorContraction.h
|
|\| |
|
| | |
|
| | |
|
| |
| |
| |
| |
| |
| | |
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.
|
| |\
| | |
| | |
| | | |
generalized operator() for indexed access and slicing
|
| | | |
|
| |/
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
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%
|
| | |
|
| |
| |
| |
| | |
buffer is not enough for tensorflow.
|
| | |
|
|\ \ |
|
| | |
| | |
| | |
| | | |
TensorConvolutionOp for ComputeCpp; fixing typos. modifying TensorDeviceSycl to use the LegacyPointer class.
|
|/ /
| |
| |
| | |
there is no need to externally call synchronise() for device memcopy.
|
| |
| |
| |
| | |
Tensor Contractsycl to be located in any place in the expression tree.
|
| | |
|
| | |
|
| |\ |
|
| | | |
|
| |/|
| |/
|/|
| | |
Fix for auto appearing in functor template argument.
|
| | |
|
| | |
|
|/ |
|
|\ |
|
| | |
|
| | |
|
| | |
|
|/ |
|
|\ |
|
| | |
|
| | |
|
|/
|
|
| |
duplication name error; adding tensorConcatinationOp backend for sycl.
|
| |
|
| |
|
| |
|
|\ |
|
| | |
|
| |
| |
| |
| | |
contractionOp sycl backend; adding temporary solution to prevent memory leak in buffer; cleaning up cxx11_tensor_buildins_sycl.h
|
| | |
|
| | |
|