Re: [PATCH v5 2/3] perf vendor events amd: add Zen2 events
From: Arnaldo Carvalho de Melo
Date: Wed Mar 18 2020 - 10:21:06 EST
Em Mon, Mar 16, 2020 at 06:52:37PM -0400, Vijay Thakkar escreveu:
> This patch adds PMU events for AMD Zen2 core based processors, namely,
> Matisse (model 71h), Castle Peak (model 31h) and Rome (model 2xh), as
> documented in the AMD Processor Programming Reference for Matisse [1].
> The model number regex has been set to detect all the models under
> family 17 that do not match those of Zen1, as the range is larger for
> zen2.
>
> Zen2 adds some additional counters that are not present in Zen1 and
> events for them have been added in this patch. Some counters have also
> been removed for Zen2 thatwere previously present in Zen1 and have been
> confirmed to always sample zero on zen2. These added/removed counters
> have been omitted for brevity but can be found here:
> https://gist.github.com/thakkarV/5b12ca5fd7488eb2c42e451e40bdd5f3
>
> Note that PPR for Zen2 [1] does not include some counters that were
> documented in the PPR for Zen1 based processors [2]. After having tested
> these counters, some of them that still work for zen2 systems have been
> preserved in the events for zen2. The counters that are omitted in [1]
> but are still measurable and non-zero on zen2 (tested on a Ryzen 3900X
> system) are the following:
>
> PMC 0x000 fpu_pipe_assignment.{total|total0|total1|total2|total3}
So trying to test this a bit, can you take a look at the examples
below? I.e. it would be really nice to have at least some of these
examples tested programatically.
So, picking this one:
"EventName": "fpu_pipe_assignment.total",
"EventCode": "0x00",
"BriefDescription": "Total number of fp uOps.",
"PublicDescription":
"Total number of fp uOps. The number of operations (uOps) dispatched
to each of the 4 FPU execution pipelines. This event reflects how
busy the FPU pipelines are and may be used for workload
characterization. This includes all operations performed by x87,
MMX, and SSE instructions, including moves. Each increment
represents a one- cycle dispatch event. This event is a speculative
event. Since this event includes non-numeric operations it is not
suitable for measuring MFLOPS.",
Committer testing:
On a AMD Ryzen 5 3600X 6-Core Processor, model 113 (0x71):
Before:
[root@five ~]# perf list *fpu_pipe_assignment*
List of pre-defined events (to be used in -e):
[root@five ~]#
After:
[root@five ~]# perf list *fpu_pipe_assignment*
List of pre-defined events (to be used in -e):
floating point:
fpu_pipe_assignment.total
[Total number of fp uOps]
Metric Groups:
[root@five ~]#
Using it:
[root@five ~]# perf stat -I1000 -e fpu_pipe_assignment.total
# time counts unit events
1.000781022 71,830,536 fpu_pipe_assignment.total
2.001835955 69,017,141 fpu_pipe_assignment.total
3.002498539 79,561,420 fpu_pipe_assignment.total
4.002922910 79,834,539 fpu_pipe_assignment.total
5.003545481 76,197,236 fpu_pipe_assignment.total
^C 5.288876179 36,038,036 fpu_pipe_assignment.total
[root@five ~]#
[root@five ~]# perf stat -e fpu_pipe_assignment.total sleep 1
Performance counter stats for 'sleep 1':
396,722 fpu_pipe_assignment.total
1.000777226 seconds time elapsed
0.000712000 seconds user
0.000000000 seconds sys
[root@five ~]#
[root@five ~]# perf record -e fpu_pipe_assignment.total sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.020 MB perf.data (7 samples) ]
[root@five ~]# perf report --stdio --no-header
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 7 of event 'fpu_pipe_assignment.total'
# Event count (approx.): 120733
#
# Overhead Command Shared Object Symbol
# ........ ....... ................ ..........................
#
96.44% sleep ld-2.30.so [.] _dl_map_object_deps
3.44% sleep [kernel.vmlinux] [k] __do_munmap
0.12% sleep [kernel.vmlinux] [k] kfree
0.00% sleep [kernel.vmlinux] [k] perf_event_mmap_output
0.00% sleep [kernel.vmlinux] [k] kmem_cache_alloc_trace
0.00% sleep [kernel.vmlinux] [k] prepend_name
0.00% sleep [kernel.vmlinux] [k] shift_arg_pages
#
# (Tip: If you have debuginfo enabled, try: perf report -s sym,srcline)
#
[root@five ~]#
[root@five ~]# perf annotate --stdio2 kfree | egrep '^ {0,2}[0-9]+' -B5 -A5
mov 0x8(%rbp),%rax
lea -0x1(%rax),%rdx
test $0x1,%al
cmovne %rdx,%rbp
mov 0x8(%rbp),%rdx
100.00 lea -0x1(%rdx),%rax
and $0x1,%edx
cmove %rbp,%rax
mov (%rax),%rax
test $0x2,%ah
â je 1bc
[root@five ~]#
[root@five ~]# perf annotate --stdio2 _dl_map_object_deps | egrep '^ {0,2}[0-9]+' -B5 -A5
cmp $0xfffffffffffffffd,%rax
â jbe 559
c7b: xor %ecx,%ecx
â jmpq 56f
c82: mov %r10,-0x4c0(%rbp)
100.00 mov -0x4b8(%rbp),%r14
movl $0x1,-0x4dc(%rbp)
â jmpq 13b
c9f: mov -0x4b8(%rbp),%rax
cmp %rax,_rtld_global
â jne 7a0
[root@five ~]#
These look reassuring:
[root@five ~]# perf record -e fpu_pipe_assignment.total -a sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.351 MB perf.data (2112 samples) ]
[root@five ~]# perf report --stdio --no-header | head -30
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 2K of event 'fpu_pipe_assignment.total'
# Event count (approx.): 1408125294
#
# Overhead Command Shared Object Symbol
# ........ ............... ............................ .............................................
#
10.24% Compositor libxul.so [.] 0x00000000041bf629
5.90% Compositor libxul.so [.] 0x00000000041bf63c
5.15% DecodingThread libxul.so [.] vp9_variance_and_sad_16x16_sse2
2.87% Compositor libxul.so [.] 0x00000000041c421b
2.56% Compositor libxul.so [.] 0x00000000041bf611
2.30% Compositor libc-2.30.so [.] __memmove_avx_unaligned_erms
1.39% Compositor libxul.so [.] 0x00000000041bf307
1.21% Compositor libxul.so [.] 0x0000000003b485de
1.09% Compositor libxul.so [.] 0x0000000003b4a9b0
1.02% Compositor libxul.so [.] 0x00000000041bf2f9
0.98% DecodingThread libxul.so [.] vp8_variance_and_sad_16x16_sse2
0.97% Compositor libxul.so [.] 0x00000000041c41cf
0.81% Compositor libxul.so [.] 0x00000000041c41bc
0.79% Compositor libxul.so [.] 0x00000000041bf2cd
0.73% Compositor libxul.so [.] 0x00000000041bf2de
0.72% Compositor libxul.so [.] 0x00000000041c4225
0.70% Compositor [kernel.vmlinux] [k] clear_page_rep
0.69% Compositor libxul.so [.] 0x00000000041bf2d2
0.68% Compositor libxul.so [.] 0x0000000003b48524
[root@five ~]#
[root@five ~]# perf annotate --stdio2 vp9_variance_and_sad_16x16_sse2 | egrep '^ {0,2}[0-9]+' -B2 -A2
movdqa %xmm1,%xmm5
pavgb %xmm3,%xmm5
2.15 psubusb %xmm1,%xmm4
psubusb %xmm0,%xmm1
psubusb %xmm3,%xmm6
--
movdqa (%rsp),%xmm2
pxor %xmm1,%xmm1
0.74 movdqa %xmm2,%xmm7
psubusb %xmm4,%xmm2
0.99 psubusb %xmm6,%xmm7
pcmpeqb %xmm1,%xmm2
2.29 pcmpeqb %xmm1,%xmm7
por %xmm2,%xmm7
14.96 neg %rax
movdqu (%rsi,%rax,2),%xmm1
movdqu (%rsi,%rax,1),%xmm3
--
pavgb %xmm3,%xmm1
psubusb %xmm2,%xmm6
2.70 psubusb %xmm0,%xmm2
psubusb %xmm3,%xmm4
psubusb %xmm0,%xmm3
--
movdqa (%rsp),%xmm2
pxor %xmm1,%xmm1
1.68 movdqa %xmm2,%xmm3
psubusb %xmm6,%xmm2
0.92 psubusb %xmm4,%xmm3
pcmpeqb %xmm1,%xmm2
pcmpeqb %xmm1,%xmm3
1.84 por %xmm2,%xmm7
5.95 por %xmm3,%xmm7
2.93 pavgb %xmm0,%xmm5
pand %xmm7,%xmm0
10.03 pandn %xmm5,%xmm7
0.96 paddusb %xmm7,%xmm0
movdqu %xmm0,(%rdi)
2.60 neg %rax
add $0x10,%rsi
add $0x10,%rdi
add $0x10,%rdx
cmp -0x28(%rbp),%edx
1.20 â jge b67
movdqu (%rbx),%xmm2
movdqu %xmm2,(%rsp)
add $0x10,%rbx
2.50 â jmpq a78
b67: sub %rdx,%rsi
sub %rdx,%rdi
--
punpcklbw %mm1,%mm1
punpcklwd %mm1,%mm1
1.14 punpckldq %mm1,%mm1
mov $0xfffffffffffffff8,%rdx
movq %mm1,(%rdi,%rdx,1)
movslq -0x28(%rbp),%rdx
movq -0x1(%rdi,%rdx,1),%mm1
2.16 punpcklbw %mm1,%mm1
punpcklwd %mm1,%mm1
punpckldq %mm1,%mm1
--
bb6: movdqu (%rdi,%rdx,1),%xmm0
movdqu -0x2(%rdi,%rdx,1),%xmm1
0.60 movdqu -0x1(%rdi,%rdx,1),%xmm3
movdqa %xmm0,%xmm4
movdqa %xmm0,%xmm6
--
movdqa %xmm2,%xmm7
psubusb %xmm4,%xmm2
1.19 psubusb %xmm6,%xmm7
2.37 pcmpeqb %xmm1,%xmm2
pcmpeqb %xmm1,%xmm7
por %xmm2,%xmm7
4.40 movdqu 0x1(%rdi,%rdx,1),%xmm1
movdqu 0x2(%rdi,%rdx,1),%xmm3
movdqa %xmm0,%xmm6
movdqa %xmm0,%xmm4
3.07 movdqa %xmm1,%xmm2
pavgb %xmm3,%xmm1
psubusb %xmm2,%xmm6
psubusb %xmm0,%xmm2
2.05 psubusb %xmm3,%xmm4
psubusb %xmm0,%xmm3
paddusb %xmm2,%xmm6
paddusb %xmm3,%xmm4
1.98 pavgb %xmm1,%xmm5
movdqa (%rsp),%xmm2
pxor %xmm1,%xmm1
movdqa %xmm2,%xmm3
1.94 psubusb %xmm6,%xmm2
psubusb %xmm4,%xmm3
pcmpeqb %xmm1,%xmm2
pcmpeqb %xmm1,%xmm3
2.42 por %xmm2,%xmm7
1.49 por %xmm3,%xmm7
1.78 pavgb %xmm0,%xmm5
pand %xmm7,%xmm0
2.89 pandn %xmm5,%xmm7
3.24 paddusb %xmm7,%xmm0
movq %mm0,-0x10(%rdi,%rdx,1)
movq %mm1,-0x8(%rdi,%rdx,1)
movdq2q %xmm0,%mm0
6.83 psrldq $0x8,%xmm0
1.97 movdq2q %xmm0,%mm1
4.03 add $0x10,%rdx
cmp -0x28(%rbp),%edx
â jge caf
[root@five ~]#