- Install JQuante locally
git clone https://github.com/mjw99/jquante.git ; cd jquante ; mvn clean install - Compile and run the benchmark:
git clone https://github.com/mjw99/jquante-jmh.git
cd jquante-jmh
# Run the CPU at the maximum frequency.
sudo cpupower frequency-set -g performance
mvn clean package ; java -jar ./target/benchmarks.jarThis will produce output similar to the following:
Benchmark Mode Cnt Score Error Units
HGPTwoElectronTermBenchmark.sd avgt 10 13245.889 ± 23.475 ns/op
HGPTwoElectronTermBenchmark.sp avgt 10 1315.202 ± 3.614 ns/op
HGPTwoElectronTermBenchmark.ss avgt 10 182.214 ± 0.255 ns/op
HuzinagaTwoElectronTermBenchmark.sd avgt 10 2708.537 ± 9.896 ns/op
HuzinagaTwoElectronTermBenchmark.sp avgt 10 1131.101 ± 6.763 ns/op
HuzinagaTwoElectronTermBenchmark.ss avgt 10 506.775 ± 0.991 ns/op
RysTwoElectronTermBenchmark.sd avgt 10 629.575 ± 3.787 ns/op
RysTwoElectronTermBenchmark.sp avgt 10 337.277 ± 1.054 ns/op
RysTwoElectronTermBenchmark.ss avgt 10 185.916 ± 0.661 ns/op
SinglePointHFWater631gdp.doBenchmark avgt 10 0.041 ± 0.001 s/op
SinglePointHFWaterSTO3G.doBenchmark avgt 10 0.003 ± 0.001 s/op
SinglePointHFWaterccpvtz.doBenchmark avgt 10 1.132 ± 0.044 s/opBasic stack based profile:
java -jar ./target/benchmarks.jar -prof stack SinglePointHFWaterccpvtzThis will produce output similar to the following:
Secondary result "name.mjw.jquante.benchmarks.SinglePointHFWaterccpvtz.doBenchmark:·stack":
Stack profiler:
....[Thread state distributions]....................................................................
72.1% RUNNABLE
21.9% WAITING
6.1% TIMED_WAITING
....[Thread state: RUNNABLE]........................................................................
23.8% 33.0% name.mjw.jquante.math.qm.integral.RysTwoElectronTerm.int1d
16.0% 22.1% name.mjw.jquante.math.qm.GMatrix.lambda$makeGMatrix$0
5.6% 7.7% name.mjw.jquante.math.qm.integral.RysTwoElectronTerm.processGn
5.0% 7.0% name.mjw.jquante.math.qm.integral.RysTwoElectronTerm.selectRoots
5.0% 6.9% <stack is empty, everything is filtered?>
4.3% 6.0% name.mjw.jquante.math.qm.integral.RysTwoElectronTerm.finaliseG
4.3% 6.0% name.mjw.jquante.math.qm.integral.RysTwoElectronTerm.processGm
1.3% 1.8% name.mjw.jquante.math.qm.integral.RysTwoElectronTerm.rNode
1.0% 1.4% java.util.stream.ForEachOps$ForEachTask.compute
0.9% 1.3% org.hipparchus.linear.ArrayRealVector.dotProduct
4.9% 6.8% <other>
....[Thread state: WAITING].........................................................................
21.9% 100.0% jdk.internal.misc.Unsafe.park
....[Thread state: TIMED_WAITING]...................................................................
5.0% 82.5% java.lang.Object.wait
1.1% 17.5% jdk.internal.misc.Unsafe.parkThis profiles using kernel based counters.
- Install userspace perf tools and kernel modules
sudo apt-get install linux-tools-common linux-tools-generic- Note "-XX:CompileCommand=print" will not print assembly without the hsdis-amd64.so library. To solve this under Ubuntu 20.04:
sudo apt-get install libhsdis0-fcmlOr, one can obtain a more recent version from here or here.
- Perform perfasm profile:
sudo sh -c 'echo -1 >/proc/sys/kernel/perf_event_paranoid'
java -jar ./target/benchmarks.jar -prof 'perfasm:intelSyntax=true;tooBigThreshold=1500;top=3' SinglePointHFWaterccpvtzThis will generate much more detail, including assembly code of the hot regions:
# Processing profiler results: LinuxPerfAsmProfiler
Result "name.mjw.jquante.benchmarks.SinglePointHFWaterccpvtz.doBenchmark":
1.136 ±(99.9%) 0.033 s/op [Average]
(min, avg, max) = (1.093, 1.136, 1.167), stdev = 0.022
CI (99.9%): [1.103, 1.168] (assumes normal distribution)
Secondary result "name.mjw.jquante.benchmarks.SinglePointHFWaterccpvtz.doBenchmark:·asm":
PrintAssembly processed: 1103591 total address lines.
Perf output processed (skipped 63.433 seconds):
Column 1: cycles (1333215 events)
Hottest code regions (>10.00% "cycles" events):
Event counts are percents of total event count.
....[Hottest Region 1]..............................................................................
c2, level 4, name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d, version 2, compile id 2291
# parm11: xmm7:xmm7 = double
# parm12: [sp+0x110] = double (sp of caller)
0x00007f92b4e6e2c0: mov r10d,DWORD PTR [rsi+0x8] ; {no_reloc}
0x00007f92b4e6e2c4: movabs r11,0x800000000
0x00007f92b4e6e2ce: add r10,r11
0x00007f92b4e6e2d1: cmp rax,r10
0x00007f92b4e6e2d4: jne 0x00007f92b45edd80 ; {runtime_call ic_miss_stub}
0x00007f92b4e6e2da: xchg ax,ax
0x00007f92b4e6e2dc: nop DWORD PTR [rax+0x0]
[Verified Entry Point]
0.03% 0x00007f92b4e6e2e0: mov DWORD PTR [rsp-0x14000],eax
0.09% 0x00007f92b4e6e2e7: push rbp
0.04% 0x00007f92b4e6e2e8: sub rsp,0x100 ;*synchronization entry
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d@-1 (line 1688)
0.09% 0x00007f92b4e6e2ef: vmovsd QWORD PTR [rsp+0x68],xmm6
0.10% 0x00007f92b4e6e2f5: vmovsd QWORD PTR [rsp+0x30],xmm0
0.05% 0x00007f92b4e6e2fb: vmovsd QWORD PTR [rsp+0x28],xmm4
0.04% 0x00007f92b4e6e301: vmovsd QWORD PTR [rsp+0x20],xmm3
0.04% 0x00007f92b4e6e307: vmovsd QWORD PTR [rsp+0x18],xmm2
0.12% 0x00007f92b4e6e30d: vmovsd QWORD PTR [rsp+0x10],xmm1
0.07% 0x00007f92b4e6e313: mov DWORD PTR [rsp+0xc],r9d
0.09% 0x00007f92b4e6e318: mov DWORD PTR [rsp+0x8],r8d
0.08% 0x00007f92b4e6e31d: mov DWORD PTR [rsp+0x4],ecx
0.17% 0x00007f92b4e6e321: mov DWORD PTR [rsp],edx
0.06% 0x00007f92b4e6e324: mov r8d,edx
0.00% 0x00007f92b4e6e327: add r8d,ecx ;*iadd {reexecute=0 rethrow=0 return_oop=0}
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d@3 (line 1688)
0.00% 0x00007f92b4e6e32a: mov r10d,DWORD PTR [rsp+0x8]
0.70% 0x00007f92b4e6e32f: add r10d,r9d ;*iadd {reexecute=0 rethrow=0 return_oop=0}
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d@10 (line 1689)
0.00% 0x00007f92b4e6e332: vaddsd xmm6,xmm5,xmm6 ;*dadd {reexecute=0 rethrow=0 return_oop=0}
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::recur@4 (line 1703)
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d@34 (line 1691)
0.01% 0x00007f92b4e6e336: vmovapd xmm0,xmm7
0.01% 0x00007f92b4e6e33a: vaddsd xmm7,xmm7,QWORD PTR [rsp+0x110]
0.25% 0x00007f92b4e6e343: mov r11,QWORD PTR [r15+0x280]
0.03% 0x00007f92b4e6e34a: mov rsi,QWORD PTR [r11] ;*invokestatic currentThread {reexecute=0 rethrow=0 return_oop=0}
; - java.lang.ThreadLocal::get@0 (line 162)
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::initialiseG@3 (line 1737)
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::recur@34 (line 1706)
; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d@34 (line 1691)
<----SNIP---->
0.03% 0x00007f92b4e6e9fa: cmp r9d,0x1
0.01% ╭ 0x00007f92b4e6e9fe: jbe 0x00007f92b4e6ea1a
0.03% │ 0x00007f92b4e6ea00: movsxd rcx,r9d
0.00% │ 0x00007f92b4e6ea03: cmp rbp,rcx
0.02% │╭ 0x00007f92b4e6ea06: jae 0x00007f92b4e6ea1a
││ 0x00007f92b4e6ea08: test r9d,r9d
0.01% ││╭ 0x00007f92b4e6ea0b: jbe 0x00007f92b4e6ea1a
0.00% │││ 0x00007f92b4e6ea0d: add rdx,0xfffffffffffffffe
0.00% │││ 0x00007f92b4e6ea11: cmp rdx,rcx
0.01% │││ 0x00007f92b4e6ea14: jb 0x00007f92b4e6eae9 ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0}
│││ ; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::finaliseG@20 (line 1776)
│││ ; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::recur@229 (line 1729)
│││ ; - name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d@34 (line 1691)
↘↘↘ 0x00007f92b4e6ea1a: mov esi,0xffffff76
0x00007f92b4e6ea1f: mov ebp,DWORD PTR [rsp]
0x00007f92b4e6ea22: mov r11d,DWORD PTR [rsp+0x4]
0x00007f92b4e6ea27: mov DWORD PTR [rsp+0x110],r11d
0x00007f92b4e6ea2f: mov r9d,DWORD PTR [rsp+0x8]
0x00007f92b4e6ea34: mov DWORD PTR [rsp+0x114],r9d
0x00007f92b4e6ea3c: mov r11d,DWORD PTR [rsp+0xc]
....................................................................................................
30.59% <total for region 1>
<----SNIP---->
....[Hottest Methods (after inlining)]..............................................................
33.01% c2, level 4 name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::int1d, version 2, compile id 2291
27.41% c2, level 4 name.mjw.jquante.math.qm.GMatrix::lambda$makeGMatrix$0, version 3, compile id 2654
19.00% c2, level 4 name.mjw.jquante.math.qm.integral.RysTwoElectronTerm::shift, version 2, compile id 2232
20.58% <...other 228 warm methods...>
....................................................................................................
100.00% <totals>