[Bug math/27461] New: Unixbench/whetstone-double performance regression on glibc2.32

[Bug math/27461] New: Unixbench/whetstone-double performance regression on glibc2.32

[xuchunmei at linux dot alibaba.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

            Bug ID: 27461
           Summary: Unixbench/whetstone-double performance regression on
                    glibc2.32
           Product: glibc
           Version: 2.32
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: math
          Assignee: unassigned at sourceware dot org
          Reporter: xuchunmei at linux dot alibaba.com
  Target Milestone: ---

I am comparing unixbench performance on glibc2.28 and glibc2.32 on centos8,
found that whetstone-double will get performance regression.
I found the discuss
https://sourceware.org/legacy-ml/libc-alpha/2019-03/msg00395.html and related
commit
https://patchwork.ozlabs.org/project/glibc/patch/VI1PR0801MB212753501D9DA1AA00BC7BA583E20@VI1PR0801MB2127.eurprd08.prod.outlook.com/

Since glibc2.31, math-finite.h is removed and __log_finite is just alias. And
since glibc2.29, exp and log are optimized.
But the test result of unixbench/whetstone-double seems performance regression
compared with optimized before when -ffast-math is added.

Testcase is unixbench, and command is “./Run whetstone-double -c 8 -i 1”,
before test on different glibc, recompile whetstone-double.

Glibc2.28 test result:
8 CPUs in system; running 8 parallel copies of tests

Double-Precision Whetstone                    41535.6 MWIPS (9.6 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      41535.6   7551.9
                                                                   ========
System Benchmarks Index Score (Partial Only)                         7551.9

glibc2.32 test result:
8 CPUs in system; running 8 parallel copies of tests

Double-Precision Whetstone                    37152.0 MWIPS (10.0 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      37152.0   6754.9
                                                                   ========
System Benchmarks Index Score (Partial Only)                         6754.9

Perf record data:
Glibc2.32:
+   49.12%  whetstone-doubl  whetstone-double   [.] whetstones.constprop.1
+   10.57%  whetstone-doubl  libm-2.32.so       [.] __atan_fma
+   10.26%  whetstone-doubl  libm-2.32.so       [.] __ieee754_log_fma
+    8.76%  whetstone-doubl  libm-2.32.so       [.] __cos_fma
+    8.22%  whetstone-doubl  libm-2.32.so       [.] __ieee754_exp_fma
+    7.65%  whetstone-doubl  libm-2.32.so       [.] __sincos
+    3.52%  whetstone-doubl  libm-2.32.so       [.] log@@GLIBC_2.29
+    1.42%  whetstone-doubl  libm-2.32.so       [.] exp@@GLIBC_2.29
     0.14%  whetstone-doubl  whetstone-double   [.] log@plt
     0.11%  whetstone-doubl  whetstone-double   [.] cos@plt
     0.05%  whetstone-doubl  whetstone-double   [.] sincos@plt
     0.01%  whetstone-doubl  [kernel.kallsyms]  [k] _raw_spin_unlock_irqrestore

glibc2.28:
+   53.86%  whetstone-doubl  whetstone-double    [.] whetstones.constprop.1
+   11.51%  whetstone-doubl  libm-2.28.so        [.] __ieee754_log_fma
+   11.16%  whetstone-doubl  libm-2.28.so        [.] __atan_fma
+    9.19%  whetstone-doubl  libm-2.28.so        [.] __cos_fma
+    8.75%  whetstone-doubl  libm-2.28.so        [.] __sincos
+    5.00%  whetstone-doubl  libm-2.28.so        [.] __ieee754_exp_fma
     0.17%  whetstone-doubl  whetstone-double    [.] __log_finite@plt
     0.11%  whetstone-doubl  whetstone-double    [.] cos@plt
     0.06%  whetstone-doubl  whetstone-double    [.] sincos@plt
     0.01%  whetstone-doubl  [kernel.kallsyms]   [k]
_raw_spin_unlock_irqrestore


When compile whetstone-double, -O3 -ffast-math is added default, and I found
that on glibc2.28, fast-math can speed up whetstone performance.
Without fast-math, the data is slow as glibc2.32. But on glibc2.32, fast-math
has no effect.

Without -fast-math on glibc2.28: 
8 CPUs in system; running 8 parallel copies of tests

Double-Precision Whetstone                    37312.4 MWIPS (9.9 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      37312.4   6784.1
                                                                   ========
System Benchmarks Index Score (Partial Only)                         6784.1

Also I test bench-log and bench-exp on glibc2.28 and glibc2.32, bench-log and
bench-exp result show that glibc2.32 is faster than glibc2.28 on log and exp.

glibc2.28:
# ./bench-log
  "log": {
   "": {
    "duration": 2.51381e+09,
    "iterations": 3.212e+07,
    "max": 780.74,
    "min": 74.124,
    "mean": 78.2631
   }
  }
# ./bench-exp
  "exp": {
   "": {
    "duration": 2.50406e+09,
    "iterations": 2.57e+07,
    "max": 166.716,
    "min": 81.516,
    "mean": 97.4344
   },
   "144bits": {
    "duration": 2.50084e+09,
    "iterations": 2.7023e+07,
    "max": 2649.79,
    "min": 88.826,
    "mean": 92.5448
   },
   "768bits": {
    "duration": 2.49121e+09,
    "iterations": 2.714e+07,
    "max": 139.302,
    "min": 88.924,
    "mean": 91.7911
   }
  }

Glibc2.32:
# ./bench-log
  "log": {
   "": {
    "duration": 2.5072e+09,
    "iterations": 3.5332e+07,
    "max": 112.336,
    "min": 69.486,
    "mean": 70.9611
   }
  }
# ./bench-exp
  "exp": {
   "": {
    "duration": 2.50791e+09,
    "iterations": 2.9812e+07,
    "max": 2759.88,
    "min": 76.894,
    "mean": 84.1241
   },
   "144bits": {
    "duration": 2.4979e+09,
    "iterations": 2.8182e+07,
    "max": 479.622,
    "min": 77.824,
    "mean": 88.6346
   },
   "768bits": {
    "duration": 2.49041e+09,
    "iterations": 2.9995e+07,
    "max": 450.432,
    "min": 77.01,
    "mean": 83.0275
   }
  }


Cpuinfo:
# lscpu
Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              8
On-line CPU(s) list: 0-7
Thread(s) per core:  2
Core(s) per socket:  2
Socket(s):           2
NUMA node(s):        1
Vendor ID:           GenuineIntel
CPU family:          6
Model:               85
Model name:          Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz
Stepping:            4
CPU MHz:             2499.442
BogoMIPS:            4998.88
Hypervisor vendor:   KVM
Virtualization type: full
L1d cache:           32K
L1i cache:           32K
L2 cache:            1024K
L3 cache:            33792K
NUMA node0 CPU(s):   0-7
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca
cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm
constant_tsc rep_good nopl cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16
pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c
rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single pti ibrs ibpb stibp
fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm mpx avx512f
avx512dq rdseed adx smap avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[xuchunmei at linux dot alibaba.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

xuchunmei <xuchunmei at linux dot alibaba.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |wdijkstr at arm dot com,
                   |                            |xuchunmei at linux dot alibaba.com

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[xuchunmei at linux dot alibaba.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

xuchunmei <xuchunmei at linux dot alibaba.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |nszabolcs at gmail dot com

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[siddhesh at sourceware dot org]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

Siddhesh Poyarekar <siddhesh at sourceware dot org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |siddhesh at sourceware dot org

--- Comment #1 from Siddhesh Poyarekar <siddhesh at sourceware dot org> ---
ISTM that the core assumption when we merged _finite with regular functions was
that the overhead of error checking was negligible, but that's clearly not true
for the whetstone benchmark.

It would be interesting to see what's different between the whetstone workload
and the ones we have in the glibc microbenchmark and find a way to harmonize
them.

FWIW, we had noticed rounding mode changing overheads were most noticeable in a
UnixBench benchmark (can't remember which one but it led to [1]) too compared
to others, so it's likely that the workloads there are heavy on inputs that
need more error/mode processing.


[1] commit 2506109403de69bd454de27835d42e6eb6ec3abc
Author: Siddhesh Poyarekar <siddhesh@redhat.com>
Date:   Wed Jun 12 10:36:48 2013 +0530

    Set/restore rounding mode only when needed

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[nsz at gcc dot gnu.org]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

Szabolcs Nagy <nsz at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |nsz at gcc dot gnu.org

--- Comment #2 from Szabolcs Nagy <nsz at gcc dot gnu.org> ---
1) this benchmark is not representative: it only calls exp with values near
-0.575 and log with values near 0.75 so it's possible that those cases had a
fast path in the old implementation, but that's not relevant in practice.

2) exp@@GLIBC_2.29 is the same as __exp and __exp_finite and __ieee754_exp.
only on x86 there is a difference because the ifuncs are not sorted out: that's
an x86 only issue.

so somebody with the right hw should do a more direct comparision of exp and
log performance (e.g. glibc has its own benchmarks but you can just measure a
loop).
i don't have an x86 with fma the x86 improvement work should be done by
somebody who has.

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[wdijkstr at arm dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

--- Comment #3 from Wilco <wdijkstr at arm dot com> ---
(In reply to Siddhesh Poyarekar from comment #1)
> ISTM that the core assumption when we merged _finite with regular functions
> was that the overhead of error checking was negligible, but that's clearly
> not true for the whetstone benchmark.

It's the opposite, using a separate veneer for error checks is expensive as
this benchmark shows. So with -Ofast you might see a regression on functions
that were special cased in math-finite.h.

So what needs to happen is removal of these veneers so that math functions are
always directly called and do their own error checks.

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[xuchunmei at linux dot alibaba.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

xuchunmei <xuchunmei at linux dot alibaba.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hjl.tools at gmail dot com

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[xuchunmei at linux dot alibaba.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

xuchunmei <xuchunmei at linux dot alibaba.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |adhemerval.zanella at linaro dot o
                   |                            |rg

-- 
You are receiving this mail because:
You are on the CC list for the bug.

[Bug math/27461] Unixbench/whetstone-double performance regression on glibc2.32

[xuchunmei at linux dot alibaba.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=27461

--- Comment #4 from xuchunmei <xuchunmei at linux dot alibaba.com> ---
I tested on another x86 cpu without fma, the hot function is __ieee754_exp_avx
and __ieee754_log_avx, the result shows that glibc2.32 is a little better than
glibc2.28.

on glibc2.32:
Benchmark Run: Tue Mar 02 2021 12:51:38 - 12:51:55
8 CPUs in system; running 8 parallel copies of tests

Double-Precision Whetstone                    28050.2 MWIPS (10.0 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      28050.2   5100.0
                                                                   ========
System Benchmarks Index Score (Partial Only)                         5100.0

on glibc2.28:
Benchmark Run: Tue Mar 02 2021 12:53:01 - 12:53:19
8 CPUs in system; running 8 parallel copies of tests

Double-Precision Whetstone                    27715.3 MWIPS (9.5 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      27715.3   5039.2
                                                                   ========
System Benchmarks Index Score (Partial Only)                         5039.2

perf data:
glibc2.32:
+   53.99%  whetstone-doubl  whetstone-double   [.] whetstones.constprop.1
+   13.42%  whetstone-doubl  libm-2.32.so       [.] __atan_avx
+    9.95%  whetstone-doubl  libm-2.32.so       [.] __cos_avx
+    7.03%  whetstone-doubl  libm-2.32.so       [.] __ieee754_log_avx
+    6.86%  whetstone-doubl  libm-2.32.so       [.] __ieee754_exp_avx
+    6.44%  whetstone-doubl  libm-2.32.so       [.] __sincos
+    1.13%  whetstone-doubl  libm-2.32.so       [.] exp@@GLIBC_2.29
     0.31%  whetstone-doubl  libm-2.32.so       [.] log@@GLIBC_2.29
     0.22%  whetstone-doubl  [kernel.kallsyms]  [k] __softirqentry_text_start
     0.15%  whetstone-doubl  whetstone-double   [.] log@plt
     0.11%  whetstone-doubl  libm-2.32.so       [.] 0x000000000000f3a4
     0.05%  whetstone-doubl  whetstone-double   [.] sincos@plt
     0.03%  whetstone-doubl  whetstone-double   [.] exp@plt
     0.02%  whetstone-doubl  [kernel.kallsyms]  [k] run_timer_softirq
     0.01%  whetstone-doubl  whetstone-double   [.] cos@plt
     0.01%  whetstone-doubl  whetstone-double   [.] atan@plt
     0.01%  whetstone-doubl  [kernel.kallsyms]  [k] finish_task_switch

glibc2.28:
+   52.14%  whetstone-doubl  whetstone-double    [.] whetstones.constprop.1
+   12.88%  whetstone-doubl  libm-2.28.so        [.] __atan_avx
+   12.64%  whetstone-doubl  libm-2.28.so        [.] __ieee754_log_avx
+    9.85%  whetstone-doubl  libm-2.28.so        [.] __cos_avx
+    6.43%  whetstone-doubl  libm-2.28.so        [.] __sincos
+    5.32%  whetstone-doubl  libm-2.28.so        [.] __ieee754_exp_avx
     0.14%  whetstone-doubl  whetstone-double    [.] __log_finite@plt
     0.14%  whetstone-doubl  whetstone-double    [.] __exp_finite@plt
     0.12%  whetstone-doubl  [kernel.kallsyms]   [k] __softirqentry_text_start
     0.05%  whetstone-doubl  whetstone-double    [.] sincos@plt
     0.01%  whetstone-doubl  [kernel.kallsyms]   [k] run_timer_softirq
     0.01%  whetstone-doubl  [kernel.kallsyms]   [k]
_raw_spin_unlock_irqrestore

cpuinfo:
# lscpu
Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              8
On-line CPU(s) list: 0-7
Thread(s) per core:  2
Core(s) per socket:  2
Socket(s):           2
NUMA node(s):        1
Vendor ID:           GenuineIntel
CPU family:          6
Model:               45
Model name:          Intel(R) Xeon(R) CPU E5-2430 0 @ 2.20GHz
Stepping:            7
CPU MHz:             2194.710
BogoMIPS:            4389.42
Virtualization:      VT-x
Hypervisor vendor:   KVM
Virtualization type: full
L1d cache:           32K
L1i cache:           32K
L2 cache:            256K
L3 cache:            15360K
NUMA node0 CPU(s):   0-7
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca
cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm
constant_tsc rep_good nopl cpuid tsc_known_freq pni pclmulqdq vmx ssse3 cx16
pcid sse4_1 sse4_2 x2apic popcnt tsc_deadline_timer aes xsave avx hypervisor
lahf_lm cpuid_fault ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid
tsc_adjust xsaveopt

-- 
You are receiving this mail because:
You are on the CC list for the bug.