Public bug reported:
I think I found a bug/regression with the ubuntu server 23.04 riscv64
Nezha/Sipeed kernels.
Somehow, the CPU speed on my mangopi mq pro seems to have decreased by 2.7x
compared to the 22.10 image.
The allwinner D1 should run at 1Ghz, and execute a single add per cycle.
On 22.10 I was able to confirm that executing `1024*1024*1024` add instructions
took roughly a second, but on 23.04 it took 2.7 seconds (both fresh installs,
with the same code [1]).
There were no resource hungry background processes running, when I
checked top.
I also binary patched the opensbi image afterward, to disable mcountinhibit for
`rdcycle`, and confirmed that even the hardware cycle counter was slower than
1Ghz. [2]
One second had 410398820 cycles instead of something closer to the expected
1073741824.
On suggestion from the #ubuntu webchat, I also tested adding
`mitigations=off` to the linux cmd options, but that didn't change
anything.
[1]
// test.c
#include <stdio.h>
#include <time.h>
int main(void)
{
for (int i = 0; i < 10; ++i) {
clock_t beg = clock();
extern void foo(void);
foo();
printf("%f\n", (clock() - beg) * 1.0/CLOCKS_PER_SEC);
}
}
// test.S
.text
.global foo
foo:
li t0, (1024*1024*1024/(8*32))
1:
.rept 32
add a1, a1, a0
add a2, a2, a1
add a3, a3, a2
add a4, a4, a3
add a5, a5, a4
add a6, a6, a5
add a7, a7, a6
add a0, a0, a7
.endr
addi t0, t0, -1
bnez t0, 1b
ret
$ gcc -Ofast -march=rv64gc test.* && ./a.out
[2] https://github.com/camel-cdr/rvv-d1/#extra-enable-rdcycle
ProblemType: Bug
DistroRelease: Ubuntu 23.04
Package: linux-image-5.19.0-1009-allwinner 5.19.0-1009.9
ProcVersionSignature: User Name 5.19.0-1009.9-allwinner 5.19.17
Uname: Linux 5.19.0-1009-allwinner riscv64
ApportVersion: 2.26.1-0ubuntu2
Architecture: riscv64
CasperMD5CheckResult: unknown
CloudArchitecture: riscv64
CloudBuildName: server
CloudID: nocloud
CloudName: unknown
CloudPlatform: nocloud
CloudSerial: 20230415.2
CloudSubPlatform: seed-dir (/var/lib/cloud/seed/nocloud-net)
Date: Fri Jun 23 20:10:14 2023
ProcCpuinfoMinimal:
processor : 0
hart : 0
isa : rv64imafdc
mmu : sv39
uarch : thead,c906
ProcEnviron:
LANG=C.UTF-8
PATH=(custom, no user)
SHELL=/bin/bash
TERM=st-256color
XDG_RUNTIME_DIR=<set>
SourcePackage: linux-allwinner
UpgradeStatus: No upgrade log present (probably fresh install)
** Affects: linux-allwinner (Ubuntu)
Importance: Undecided
Status: New
** Tags: apport-bug cloud-image lunar riscv64
--
You received this bug notification because you are a member of Kernel
Packages, which is subscribed to linux-allwinner in Ubuntu.
https://bugs.launchpad.net/bugs/2024935
Title:
cpu 2.7x slower on ubuntu server 23.04 riscv64 Nezha/Sipeed images
compared to 22.10 on mangopi mq-pro
Status in linux-allwinner package in Ubuntu:
New
Bug description:
I think I found a bug/regression with the ubuntu server 23.04 riscv64
Nezha/Sipeed kernels.
Somehow, the CPU speed on my mangopi mq pro seems to have decreased by 2.7x
compared to the 22.10 image.
The allwinner D1 should run at 1Ghz, and execute a single add per cycle.
On 22.10 I was able to confirm that executing `1024*1024*1024` add
instructions took roughly a second, but on 23.04 it took 2.7 seconds (both
fresh installs, with the same code [1]).
There were no resource hungry background processes running, when I
checked top.
I also binary patched the opensbi image afterward, to disable mcountinhibit
for `rdcycle`, and confirmed that even the hardware cycle counter was slower
than 1Ghz. [2]
One second had 410398820 cycles instead of something closer to the expected
1073741824.
On suggestion from the #ubuntu webchat, I also tested adding
`mitigations=off` to the linux cmd options, but that didn't change
anything.
[1]
// test.c
#include <stdio.h>
#include <time.h>
int main(void)
{
for (int i = 0; i < 10; ++i) {
clock_t beg = clock();
extern void foo(void);
foo();
printf("%f\n", (clock() - beg) * 1.0/CLOCKS_PER_SEC);
}
}
// test.S
.text
.global foo
foo:
li t0, (1024*1024*1024/(8*32))
1:
.rept 32
add a1, a1, a0
add a2, a2, a1
add a3, a3, a2
add a4, a4, a3
add a5, a5, a4
add a6, a6, a5
add a7, a7, a6
add a0, a0, a7
.endr
addi t0, t0, -1
bnez t0, 1b
ret
$ gcc -Ofast -march=rv64gc test.* && ./a.out
[2] https://github.com/camel-cdr/rvv-d1/#extra-enable-rdcycle
ProblemType: Bug
DistroRelease: Ubuntu 23.04
Package: linux-image-5.19.0-1009-allwinner 5.19.0-1009.9
ProcVersionSignature: User Name 5.19.0-1009.9-allwinner 5.19.17
Uname: Linux 5.19.0-1009-allwinner riscv64
ApportVersion: 2.26.1-0ubuntu2
Architecture: riscv64
CasperMD5CheckResult: unknown
CloudArchitecture: riscv64
CloudBuildName: server
CloudID: nocloud
CloudName: unknown
CloudPlatform: nocloud
CloudSerial: 20230415.2
CloudSubPlatform: seed-dir (/var/lib/cloud/seed/nocloud-net)
Date: Fri Jun 23 20:10:14 2023
ProcCpuinfoMinimal:
processor : 0
hart : 0
isa : rv64imafdc
mmu : sv39
uarch : thead,c906
ProcEnviron:
LANG=C.UTF-8
PATH=(custom, no user)
SHELL=/bin/bash
TERM=st-256color
XDG_RUNTIME_DIR=<set>
SourcePackage: linux-allwinner
UpgradeStatus: No upgrade log present (probably fresh install)
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux-allwinner/+bug/2024935/+subscriptions
--
Mailing list: https://launchpad.net/~kernel-packages
Post to : kernel-packages@lists.launchpad.net
Unsubscribe : https://launchpad.net/~kernel-packages
More help : https://help.launchpad.net/ListHelp
