| Commit message (Collapse) | Author | Age |
| |
|
| |
|
| |
|
|
|
|
|
|
| |
553caeb6a3bb545aef895f8fc9f219be44679017
.
|
| |
|
| |
|
| |
|
| |
|
| |
|
|
|
|
| |
generate too many build warnings.
|
|
|
|
|
|
| |
dimension.
https://bitbucket.org/snippets/rmlarsen/MexxLo
|
| |
|
| |
|
| |
|
| |
|
|
|
|
| |
'thread-local' memory for packing
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
evalShardedByInnerDim ensures that the values it passes for start_k and
end_k to evalGemmPartialWithoutOutputKernel are multiples of 8 as the kernel
does not work correctly when the values of k are not multiples of the
packet_size. While this precaution works for AVX builds, it is insufficient
for AVX512 builds where the maximum packet size is 16. The result is slightly
incorrect float32 contractions on AVX512 builds.
This commit fixes the problem by ensuring that k is always a multiple of
the packet_size if the packet_size is > 8.
|
| |
|
| |
|
|\
| |
| |
| | |
Add tests for evalShardedByInnerDim contraction + fix bugs
|
| | |
|
|/ |
|
|\ |
|
| | |
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
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%
|
|/ |
|
|
|
|
| |
It brokes the complex-complex case on SSE.
|
|\
| |
| |
| | |
Use device's allocate function instead of internal::aligned_malloc.
|
| |
| |
| |
| | |
contraction to reduce binary size.
|
| | |
|
| |
| |
| |
| | |
would make it easier to track memory usage in device instances.
|
| | |
|
| | |
|
|/ |
|
| |
|
|
|
|
| |
to evaluate a tensor expression.
|
| |
|
|\
| |
| |
| | |
Minor cleanups: 1. Get rid of a few unused variables. 2. Get rid of last uses of EIGEN_USE_COST_MODEL.
|
| | |
|
|/
|
|
| |
EIGEN_USE_COST_MODEL.
|
|
|
|
| |
multiple cores. It is still possible to revert to the old thread pool by compiling with the EIGEN_USE_SIMPLE_THREAD_POOL define.
|
|
|
|
| |
the closure in the order in which they are enqueued. This is needed in order to switch to the new non blocking thread pool since this new thread pool can execute the closure in any order.
|
| |
|
|
|
|
| |
Fixed a couple of issues in the corresponding code.
|
|
|
|
| |
compiling with nvcc and avx enabled leads to many issues.
|
| |
|