9e162146a9
Only initialize the part of the jobs array that will get used
...
The jobs array is getting initialized in O(compiled cpus^2) complexity.
Distros and people with bigger systems will use pretty high values
(128 or 256 or more) for this value, leading to interesting bubbles
in performance.
Baseline (single threaded performance) gets roughly 13 - 15 multiplications per cycle
in the interesting range (threading kicks in at 65x65 mult by 65x65).
The hardware is capable of 32 multiplications per cycle theoretically.
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10703.9 10.6 0.0% 17990.6 6.3 0.0%
64 x 64 20778.4 12.8 0.0% 40629.2 6.5 0.0%
65 x 65 26869.9 10.3 0.0% 52545.7 5.3 0.0%
80 x 80 38104.5 13.5 0.0% 72492.7 7.1 0.0%
96 x 96 61626.4 14.4 0.0% 113983.8 7.8 0.0%
112 x 112 91803.8 15.3 0.0% 180987.3 7.8 0.0%
128 x 128 133161.4 15.8 0.0% 258374.3 8.1 0.0%
When threading is turned on
TARGET=SKYLAKEX F_COMPILER=GFORTRAN SHARED=1 DYNAMIC_THREADS=1 USE_OPENMP=0 NUM_THREADS=128
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10725.9 10.5 -0.2% 18134.9 6.2 -0.8%
64 x 64 20500.6 12.9 1.3% 40929.1 6.5 -0.7%
65 x 65 2040832.1 0.1 -7495.2% 2097633.6 0.1 -3892.0%
80 x 80 2063129.1 0.2 -5314.4% 2119925.2 0.2 -2824.3%
96 x 96 2070374.5 0.4 -3259.6% 2173604.4 0.4 -1806.9%
112 x 112 2111721.5 0.7 -2169.6% 2263330.8 0.6 -1170.0%
128 x 128 2276181.5 0.9 -1609.3% 2377228.9 0.9 -820.1%
There is a deep deep cliff once you hit 65x65
With this patch
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10630.0 10.6 0.7% 18112.8 6.2 -0.7%
64 x 64 20374.8 13.0 1.9% 40487.0 6.5 0.4%
65 x 65 141955.2 1.9 -428.3% 146708.8 1.9 -179.2%
80 x 80 178921.1 2.9 -369.6% 186032.7 2.8 -156.6%
96 x 96 205436.2 4.3 -233.4% 224513.1 3.9 -97.0%
112 x 112 244408.2 5.8 -162.7% 262158.7 5.4 -47.1%
128 x 128 321334.5 6.5 -141.3% 333829.0 6.3 -29.2%
The cliff is very significantly reduced.
(more to follow)
2018-06-17 15:32:03 +00:00
47bf0dba8f
Add build-time option for OMP scheduler; document MULTITHREAD_THRESHOLD range ( #1620 )
...
* Allow choosing the OpenMP scheduler and add range hint for GEMM_MULTITHREAD_THRESHOLD
* Amended description of GEMM_MULTITHREAD_THRESHOLD
to reflect #742 making it track floating point operations rather than matrix size
2018-06-15 11:25:05 +02:00
12603b7dbb
Merge pull request #1618 from oon3m0oo/less_locking
...
Remove the need for most locking in memory.c.
2018-06-15 00:10:29 +02:00
bf40f806ef
Remove the need for most locking in memory.c.
...
Using thread local storage for tracking memory allocations means that threads
no longer have to lock at all when doing memory allocations / frees. This
particularly helps the gemm driver since it does an allocation per invocation.
Even without threading at all, this helps, since even calling a lock with
no contention has a cost:
Before this change, no threading:
```
----------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------
BM_SGEMM/4 102 ns 102 ns 13504412
BM_SGEMM/6 175 ns 175 ns 7997580
BM_SGEMM/8 205 ns 205 ns 6842073
BM_SGEMM/10 266 ns 266 ns 5294919
BM_SGEMM/16 478 ns 478 ns 2963441
BM_SGEMM/20 690 ns 690 ns 2144755
BM_SGEMM/32 1906 ns 1906 ns 716981
BM_SGEMM/40 2983 ns 2983 ns 473218
BM_SGEMM/64 9421 ns 9422 ns 148450
BM_SGEMM/72 12630 ns 12631 ns 112105
BM_SGEMM/80 15845 ns 15846 ns 89118
BM_SGEMM/90 25675 ns 25676 ns 54332
BM_SGEMM/100 29864 ns 29865 ns 47120
BM_SGEMM/112 37841 ns 37842 ns 36717
BM_SGEMM/128 56531 ns 56532 ns 25361
BM_SGEMM/140 75886 ns 75888 ns 18143
BM_SGEMM/150 98493 ns 98496 ns 14299
BM_SGEMM/160 102620 ns 102622 ns 13381
BM_SGEMM/170 135169 ns 135173 ns 10231
BM_SGEMM/180 146170 ns 146172 ns 9535
BM_SGEMM/189 190226 ns 190231 ns 7397
BM_SGEMM/200 194513 ns 194519 ns 7210
BM_SGEMM/256 396561 ns 396573 ns 3531
```
with this change:
```
----------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------
BM_SGEMM/4 95 ns 95 ns 14500387
BM_SGEMM/6 166 ns 166 ns 8381763
BM_SGEMM/8 196 ns 196 ns 7277044
BM_SGEMM/10 256 ns 256 ns 5515721
BM_SGEMM/16 463 ns 463 ns 3025197
BM_SGEMM/20 636 ns 636 ns 2070213
BM_SGEMM/32 1885 ns 1885 ns 739444
BM_SGEMM/40 2969 ns 2969 ns 472152
BM_SGEMM/64 9371 ns 9372 ns 148932
BM_SGEMM/72 12431 ns 12431 ns 112919
BM_SGEMM/80 15615 ns 15616 ns 89978
BM_SGEMM/90 25397 ns 25398 ns 55041
BM_SGEMM/100 29445 ns 29446 ns 47540
BM_SGEMM/112 37530 ns 37531 ns 37286
BM_SGEMM/128 55373 ns 55375 ns 25277
BM_SGEMM/140 76241 ns 76241 ns 18259
BM_SGEMM/150 102196 ns 102200 ns 13736
BM_SGEMM/160 101521 ns 101525 ns 13556
BM_SGEMM/170 136182 ns 136184 ns 10567
BM_SGEMM/180 146861 ns 146864 ns 9035
BM_SGEMM/189 192632 ns 192632 ns 7231
BM_SGEMM/200 198547 ns 198555 ns 6995
BM_SGEMM/256 392316 ns 392330 ns 3539
```
Before, when built with USE_THREAD=1, GEMM_MULTITHREAD_THRESHOLD = 4, the cost
of small matrix operations was overshadowed by thread locking (look smaller than
32) even when not explicitly spawning threads:
```
----------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------
BM_SGEMM/4 328 ns 328 ns 4170562
BM_SGEMM/6 396 ns 396 ns 3536400
BM_SGEMM/8 418 ns 418 ns 3330102
BM_SGEMM/10 491 ns 491 ns 2863047
BM_SGEMM/16 710 ns 710 ns 2028314
BM_SGEMM/20 871 ns 871 ns 1581546
BM_SGEMM/32 2132 ns 2132 ns 657089
BM_SGEMM/40 3197 ns 3196 ns 437969
BM_SGEMM/64 9645 ns 9645 ns 144987
BM_SGEMM/72 35064 ns 32881 ns 50264
BM_SGEMM/80 37661 ns 35787 ns 42080
BM_SGEMM/90 36507 ns 36077 ns 40091
BM_SGEMM/100 32513 ns 31850 ns 48607
BM_SGEMM/112 41742 ns 41207 ns 37273
BM_SGEMM/128 67211 ns 65095 ns 21933
BM_SGEMM/140 68263 ns 67943 ns 19245
BM_SGEMM/150 121854 ns 115439 ns 10660
BM_SGEMM/160 116826 ns 115539 ns 10000
BM_SGEMM/170 126566 ns 122798 ns 11960
BM_SGEMM/180 130088 ns 127292 ns 11503
BM_SGEMM/189 120309 ns 116634 ns 13162
BM_SGEMM/200 114559 ns 110993 ns 10000
BM_SGEMM/256 217063 ns 207806 ns 6417
```
and after, it's gone (note this includes my other change which reduces calls
to num_cpu_avail):
```
----------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------
BM_SGEMM/4 95 ns 95 ns 12347650
BM_SGEMM/6 166 ns 166 ns 8259683
BM_SGEMM/8 193 ns 193 ns 7162210
BM_SGEMM/10 258 ns 258 ns 5415657
BM_SGEMM/16 471 ns 471 ns 2981009
BM_SGEMM/20 666 ns 666 ns 2148002
BM_SGEMM/32 1903 ns 1903 ns 738245
BM_SGEMM/40 2969 ns 2969 ns 473239
BM_SGEMM/64 9440 ns 9440 ns 148442
BM_SGEMM/72 37239 ns 33330 ns 46813
BM_SGEMM/80 57350 ns 55949 ns 32251
BM_SGEMM/90 36275 ns 36249 ns 42259
BM_SGEMM/100 31111 ns 31008 ns 45270
BM_SGEMM/112 43782 ns 40912 ns 34749
BM_SGEMM/128 67375 ns 64406 ns 22443
BM_SGEMM/140 76389 ns 67003 ns 21430
BM_SGEMM/150 72952 ns 71830 ns 19793
BM_SGEMM/160 97039 ns 96858 ns 11498
BM_SGEMM/170 123272 ns 122007 ns 11855
BM_SGEMM/180 126828 ns 126505 ns 11567
BM_SGEMM/189 115179 ns 114665 ns 11044
BM_SGEMM/200 89289 ns 87259 ns 16147
BM_SGEMM/256 226252 ns 222677 ns 7375
```
I've also tested this with ThreadSanitizer and found no data races during
execution. I'm not sure why 200 is always faster than it's neighbors, we must
be hitting some optimal cache size or something.
2018-06-14 16:54:58 +01:00
ed682a4a0c
Merge pull request #1619 from martin-frbg/issue1580
...
Update OSX deployment target to 10.8
2018-06-14 17:48:51 +02:00
fcb77ab129
Update OSX deployment target to 10.8
...
fixes #1580
2018-06-14 16:57:58 +02:00
26e1cfb653
Merge pull request #1607 from martin-frbg/dynarch
...
Move some x86_64 DYNAMIC_ARCH targets to new DYNAMIC_OLDER option
2018-06-14 16:52:55 +02:00
c628c6fa59
Merge pull request #1612 from oon3m0oo/cpus
...
Fixed a few more unnecessary calls to num_cpu_avail.
2018-06-14 16:51:31 +02:00
67d81ab49d
Merge pull request #1609 from martin-frbg/issue1529
...
Create OpenBLASConfig.cmake in cmake builds as well
2018-06-12 23:00:24 +02:00
2f957947a6
Merge pull request #1613 from xianyi/revert-1600-noyield
...
Revert "Use usleep instead of sched_yield by default"
2018-06-11 17:14:49 +02:00
de8fff671d
Revert "Use usleep instead of sched_yield by default"
2018-06-11 17:05:27 +02:00
6f71c0fce4
Return a somewhat sane default value for L2 cache size if cpuid retur… ( #1611 )
...
* Return a somewhat sane default value for L2 cache size if cpuid returned something unexpected
Fixes #1610 , the KVM hypervisor on Google Chromebooks returning zero for CPUID 0x80000006, causing DYNAMIC_ARCH
builds of OpenBLAS to hang
2018-06-11 13:26:19 +02:00
c2545b0fd6
Fixed a few more unnecessary calls to num_cpu_avail.
...
I don't have as many benchmarks for these as for gemm, but it should still
make a difference for small matrices.
2018-06-11 10:17:16 +01:00
e65f451409
include CMakePackageConfigHelpers
2018-06-10 15:09:43 +02:00
02634b549b
Add template for OpenBLASConfig.cmake
2018-06-10 09:25:46 +02:00
0bea6bb9e7
Create OpenBLASConfig.cmake from cmake as well
2018-06-10 09:24:37 +02:00
3313e4b946
Merge pull request #1608 from martin-frbg/issue874
...
Enable parallel make on MS Windows by default
2018-06-09 19:57:33 +02:00
e9cd11768c
Enable parallel make on MS Windows by default
...
fixes #874
2018-06-09 17:54:36 +02:00
63f7395fb4
Move some DYNAMIC_ARCH targets to new DYNAMIC_OLDER option
2018-06-09 16:31:38 +02:00
1cbd8f3ae4
Move some DYNAMIC_ARCH targets to new DYNAMIC_OLDER option
2018-06-09 16:30:46 +02:00
6c2d90ba77
Move some DYNAMIC_ARCH targets to new DYNAMIC_OLDER option
2018-06-09 16:29:17 +02:00
0297b3211a
Merge pull request #1605 from oon3m0oo/develop
...
Improve performance of GEMM for small matrices when SMP is defined.
2018-06-09 12:42:34 +02:00
66316b9f4c
Improve performance of GEMM for small matrices when SMP is defined.
...
Always checking num_cpu_avail() regardless of whether threading will actually
be used adds noticeable overhead for small matrices. Most other uses of
num_cpu_avail() do so only if threading will be used, so do the same here.
2018-06-07 15:29:13 +01:00
6adc4b7b36
Merge pull request #1601 from martin-frbg/zaxpy
...
Use a single thread for small input size in zaxpy
2018-06-07 14:09:58 +02:00
2ade0ef085
Merge pull request #1600 from martin-frbg/noyield
...
Use usleep instead of sched_yield by default
2018-06-07 12:42:00 +02:00
e8880c1699
Use a single thread for small input size
...
copies daxpy improvement from #27 , see #1560
2018-06-07 10:26:55 +02:00
ed7c4a043b
Use usleep instead of sched_yield by default
...
sched_yield only burns cpu cycles, fixes #900 , see also #923 , #1560
2018-06-07 10:18:26 +02:00
cf234a0561
Merge pull request #1589 from fenrus75/skylakex
...
Initial support for SkylakeX / AVX512
2018-06-06 22:07:09 +02:00
ae2a33128b
Merge pull request #1599 from martin-frbg/c_check_avx512
...
Improved AVX512 test case for c_check
2018-06-06 18:42:42 +02:00
e4718b1fee
Better AVX512 test case
2018-06-06 16:51:30 +02:00
9b87b64262
Improve AVX512 testcase
...
clang 3.4 managed to accept the original test code, only to fail on the actual Skylake asm later
2018-06-06 16:49:00 +02:00
0218b884c1
Merge pull request #1598 from martin-frbg/issue1593-2
...
Restore _Atomic define before stdatomic.h for old gcc
2018-06-06 12:48:26 +02:00
83da278093
Update common.h
2018-06-06 09:27:49 +02:00
358d4df2bd
Merge branch 'develop' into issue1593-2
2018-06-06 09:21:41 +02:00
06d43760e4
Restore _Atomic define before stdatomic.h for old gcc
...
see #1593
2018-06-06 09:18:10 +02:00
a4af8861ff
Merge pull request #1597 from martin-frbg/cmake-avx512
...
Check build system support for AVX512 instructions
2018-06-06 07:22:20 +02:00
7fb62aed7e
Check build system support for AVX512 instructions
2018-06-05 23:29:33 +02:00
f6021c798d
Re-enable QUIET_MAKE
2018-06-05 19:09:38 +02:00
e8002536ec
disable quiet_make for the moment
2018-06-05 18:23:01 +02:00
ce6317f6c0
Merge pull request #1594 from martin-frbg/issue1593
...
Fix inverted condition in _Atomic declaration
2018-06-05 16:02:51 +02:00
15a78d6b66
export NO_AVX512 setting
2018-06-05 15:58:34 +02:00
354a976a59
Fix inverted condition in _Atomic declaration
...
fixes #1593
2018-06-05 10:31:34 +02:00
38ad05bd04
Extend loop range to find SkylakeX in force_coretype
2018-06-05 10:26:49 +02:00
b7feded85a
Propagate NO_AVX512 via CCOMMON_OPT
2018-06-05 10:24:05 +02:00
dc9fe05ab5
Update cpuid_x86.c
2018-06-04 17:10:19 +02:00
8be027e4c6
Update dynamic.c
2018-06-04 14:36:39 +02:00
ac7b6e3e9a
Fix misplaced endif
2018-06-04 08:23:40 +02:00
fc66a0ec0b
Merge pull request #1590 from martin-frbg/avx512_check
...
Disable AVX512 (Skylake X) support if the build system is too old
2018-06-04 08:18:38 +02:00
89372e0993
Use AVX512 also for DGEMM
...
this required switching to the generic gemm_beta code (which is faster anyway on SKX)
for both DGEMM and SGEMM
Performance for the not-retuned version is in the 30% range
2018-06-03 22:17:27 +00:00
ef626c6824
typo fix
2018-06-04 00:13:19 +02:00
Arjan van de Ven
Martin Kroeker
Craig Donner