WindowsPerf is a (Linux perf inspired) Windows on Arm performance profiling tool. Profiling is based on ARM64 PMU and its hardware counters. WindowsPerf supports the counting model for obtaining aggregate counts of occurrences of special events, and sampling model for determining the frequencies of event occurrences produced by program locations at the function, basic block, and/or instruction levels.
Learn more in this blog announcing the first release.
WindowsPerf is composed of two main components:
You can find all binary releases of WindowsPerf
here
.
For installation instructions see the install guide .
For a list of all available options, enter:
wperf -h
WindowsPerf uses Arm processor Performance Monitoring Unit (PMU) counters to generate its data. The available events that can be profiled will vary per target.
To generate a list of available events, use:
wperf list
The output should be similar to:
List of pre-defined events (to be used in -e)
Alias Name Raw Index Event Type
========== ========= ==========
sw_incr 0x00 [core PMU event]
l1i_cache_refill 0x01 [core PMU event]
...
List of supported metrics (to be used in -m)
Metric Events
====== ======
dcache {l1d_cache,l1d_cache_refill,l2d_cache,l2d_cache_refill,inst_retired}
dtlb {l1d_tlb,l1d_tlb_refill,l2d_tlb,l2d_tlb_refill,inst_retired}
icache {l1i_cache,l1i_cache_refill,l2i_cache,l2i_cache_refill,inst_retired}
...
You can extend wperf list command output with additional information like event and metrics description with -v command line option.
Command line option test prints on screen various wperf configuration settings:
wperf test
Test Name Result
========= ======
request.ioctl_events [EVT_CORE] False
request.ioctl_events [EVT_DSU] False
request.ioctl_events [EVT_DMC_CLK/EVT_DMC_CLKDIV2] False
pmu_device.vendor_name Arm Limited
pmu_device.product_name neoverse-n1
pmu_device.product_name(extended) Neoverse N1 (neoverse-n1), armv8.1, pmu_v3
pmu_device.product [] armv8-a,armv9-a,neoverse-n1,neoverse-n2,neoverse-n2-r0p0,neoverse-n2-r0p1,neoverse-n2-r0p3,neoverse-v1
pmu_device.m_product_alias (neoverse-n2-r0p0:neoverse-n2),(neoverse-n2-r0p1:neoverse-n2)
pmu_device.events_query(events) [EVT_CORE] 110
pmu_device.events_query(events) [EVT_DSU] 9
pmu_device.events_query(events) [EVT_DMC_CLK] 3
pmu_device.events_query(events) [EVT_DMC_CLKDIV2] 26
PMU_CTL_QUERY_HW_CFG [arch_id] 0x000f
PMU_CTL_QUERY_HW_CFG [core_num] 0x0050
PMU_CTL_QUERY_HW_CFG [fpc_num] 0x0001
PMU_CTL_QUERY_HW_CFG [gpc_num] 0x0006
PMU_CTL_QUERY_HW_CFG [total_gpc_num] 0x0006
PMU_CTL_QUERY_HW_CFG [part_id] 0x0d0c
PMU_CTL_QUERY_HW_CFG [pmu_ver] 0x0004
PMU_CTL_QUERY_HW_CFG [rev_id] 0x0001
PMU_CTL_QUERY_HW_CFG [variant_id] 0x0003
PMU_CTL_QUERY_HW_CFG [vendor_id] 0x0041
PMU_CTL_QUERY_HW_CFG [midr_value] 0x000000000000413fd0c1
...
You can output wperf test command in JSON format. Use --json command line option to enable JSON output.
Specify the event to profile with -e. Groups of events, known as metrics can be specified with -m.
For example, generate a report for CPU core 0 (-c 0) for two seconds (sleep 2) with:
wperf stat -e cpu_cycles -m icache -c 0 sleep 2
This will output a report similar to:
counting ... done
Performance counter stats for core 0, no multiplexing, kernel mode excluded, on Arm Limited core implementation:
note: 'e' - normal event, 'gN' - grouped event with group number N, metric name will be appended if 'e' or 'g' comes from it
counter value event name event idx event note
============= ========== ========= ==========
7,408,075 cycle fixed e
2,271,166 l1i_cache 0x14 g0,icache
126,875 l1i_cache_refill 0x01 g0,icache
0 l2i_cache 0x27 g0,icache
0 l2i_cache_refill 0x28 g0,icache
6,247,674 inst_retired 0x08 g0,icache
7,408,075 cpu_cycles 0x11 e
2.281 seconds time elapsed
You can output wperf stat command in JSON format. Use --json command line option to enable JSON output.
Example use cases are provided in the WindowsPerf documentation .