1 AutoFDO and ARM Trace {#AutoFDO}
4 @brief Using CoreSight trace and perf with OpenCSD for AutoFDO.
8 Feedback directed optimization (FDO, also know as profile guided
9 optimization - PGO) uses a profile of a program's execution to guide the
10 optmizations performed by the compiler. Traditionally, this involves
11 building an instrumented version of the program, which records a profile of
12 execution as it runs. The instrumentation adds significant runtime
13 overhead, possibly changing the behaviour of the program and it may not be
14 possible to run the instrumented program in a production environment
15 (e.g. where performance criteria must be met).
17 AutoFDO uses facilities in the hardware to sample the behaviour of the
18 program in the production environment and generate the execution profile.
19 An improved profile can be obtained by including the branch history
20 (i.e. a record of the last branches taken) when generating an instruction
21 samples. On Arm systems, the ETM can be used to generate such records.
23 The process can be broken down into the following steps:
25 * Record execution trace of the program
26 * Convert the execution trace to instruction samples with branch histories
27 * Convert the instruction samples to source level profiles
28 * Use the source level profile with the compiler
30 This article describes how to enable ETM trace on Arm targets running Linux
31 and use the ETM trace to generate AutoFDO profiles and compile an optimized
35 ## Execution trace on Arm targets
37 Debug and trace of Arm targets is provided by CoreSight. This consists of
38 a set of components that allow access to debug logic, record (trace) the
39 execution of a processor and route this data through the system, collecting
42 To record the execution of a processor, we require the following
45 * A trace source. The core contains a trace unit, called an ETM that emits
46 data describing the instructions executed by the core.
47 * Trace links. The trace data generated by the ETM must be moved through
48 the system to the component that collects the data (sink). Links
50 * Funnels: merge multiple streams of data
51 * FIFOs: buffer data to smooth out bursts
52 * Replicators: send a stream of data to multiple components
53 * Sinks. These receive the trace data and store it or send it to an
55 * ETB: A small circular buffer (64-128 kilobytes) that stores the most
57 * ETR: A larger (several megabytes) buffer that uses system RAM to
59 * TPIU: Sends data to an off-chip capture device (e.g. Arm DSTREAM)
61 Each Arm SoC design may have a different layout (topology) of components.
62 This topology is described to the OS drivers by the platform's devicetree
63 or (in future) ACPI firmware.
65 For application profiling, we need to store several megabytes of data
66 within the system, so will use ETR with the capture tool (perf)
67 periodically draining the buffer to a file.
69 Even though we have a large capture buffer, the ETM can still generate a
70 lot of data very quickly - typically an ETM will generate ~1 bit of data
71 per instruction (depending on the workload), which results in 256Mbytes per
72 second for a core running at 2GHz. This leads to problems storing and
73 decoding such large volumes of data. AutoFDO uses samples of program
74 execution, so we can avoid this problem by using the ETM's features to
75 only record small slices of execution - e.g. collect ~5000 cycles of data
76 every 50M cycles. This reduces the data rate to a manageable level - a few
77 megabytes per minute. This technique is known as 'strobing'.
84 To collect ETM trace, the CoreSight drivers must be included in the
85 kernel. Some of the driver support is not yet included in the mainline
86 kernel and many targets are using older kernels. To enable CoreSight trace
87 on these targets, Arm have provided backports of the latest CoreSight
88 drivers and ETM strobing patch at:
90 [TODO: link to git repos for CoreSight backports]
92 You can include these backports in your kernel by either merging the
93 appropriate branch using git or generating patches (using `git
96 For 4.9 based kernels, use the `coresight-4.9-etr-etm_strobe` branch:
99 git merge coresight-4.9-etr-etm_strobe
105 git format-patch --output-directory /output/dir v4.9..coresight-4.9-etr-etm_strobe
107 git am /output/dir/*.patch # or patch -p1 /output/dir/*.patch if not using git
110 For 4.14 based kernels, use the `coresight-4.14-etm_strobe` branch:
113 git merge coresight-4.14-etm_strobe
119 git format-patch --output-directory /output/dir v4.14..coresight-4.14-etm_strobe
121 git am /output/dir/*.patch # or patch -p1 /output/dir/*.patch if not using git
124 The CoreSight trace drivers must also be enabled in the kernel
125 configuration. This can be done using the configuration menu (`make
126 menuconfig`), selecting `Kernel hacking` / `CoreSight Tracing Support` and
127 enabling all options, or by setting the following in the configuration
132 CONFIG_CORESIGHT_LINK_AND_SINK_TMC=y
133 CONFIG_CORESIGHT_SINK_TPIU=y
134 CONFIG_CORESIGHT_SOURCE_ETM4X=y
135 CONFIG_CORESIGHT_DYNAMIC_REPLICATOR=y
136 CONFIG_CORESIGHT_STM=y
137 CONFIG_CORESIGHT_CATU=y
140 Compile the kernel for your target in the usual way, e.g.
143 make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu-
146 Each target may have a different layout of CoreSight components. To
147 collect trace into a sink, the kernel drivers need to know which other
148 devices need to be configured to route data from the source to the sink.
149 This is described in the devicetree (and in future, the ACPI tables). The
150 device tree will define which CoreSight devices are present in the system,
151 where they are located and how they are connected together. The devicetree
152 for some platforms includes a description of the platform's CoreSight
153 components, but in other cases you may have to ask the platform/SoC vendor
154 to supply it or create it yourself (see Appendix: Describing CoreSight in
157 Once the target has been booted with the devicetree describing the
158 CoreSight devices, you should find the devices in sysfs:
161 # ls /sys/bus/coresight/devices/
162 28440000.etm 28540000.etm 28640000.etm 28740000.etm
163 28c03000.funnel 28c04000.etf 28c05000.replicator 28c06000.etr
169 The perf tool is used to capture execution trace, configuring the trace
170 sources to generate trace, routing the data to the sink and collecting the
173 Arm recommends to use the perf version corresponding to the kernel running
174 on the target. This can be built from the same kernel sources with
177 make -C tools/perf ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu-
180 If the post-processing (`perf inject`) of the captured data is not being
181 done on the target, then the OpenCSD library is not required for this build
184 Trace is captured by collecting the `cs_etm` event from perf. The sink
185 to collect data into is specified as a parameter of this event. Trace can
186 also be restricted to user space or kernel space with 'u' or 'k'
187 parameters. For example:
190 perf record -e cs_etm/@28c06000.etr/u --per-thread -- /bin/ls
193 Will record the userspace execution of '/bin/ls' into the ETR located at
194 0x28c06000. Note the `--per-thread` option is required - perf currently
195 only supports trace of a single thread of execution. CPU wide trace is a
199 ## Processing trace and profiles
201 perf is also used to convert the execution trace an instruction profile.
202 This requires a different build of perf, using the version of perf from
203 Linux v4.17 or later, as the trace processing code isn't included in the
204 driver backports. Trace decode is provided by the OpenCSD library
205 (<https://github.com/Linaro/OpenCSD>), v0.9.1 or later. This is packaged
206 for debian testing (install the libopencsd0, libopencsd-dev packages) or
207 can be compiled from source and installed.
209 The autoFDO tool <https://github.com/google/autofdo> is used to convert the
210 instruction profiles to source profiles for the GCC and clang/llvm
214 ## Recording and profiling
216 Once trace collection using perf is working, we can now use it to profile
219 The application must be compiled to include sufficient debug information to
220 map instructions back to source lines. For GCC, use the `-g1` or `-gmlt`
221 options. For clang/llvm, also add the `-fdebug-info-for-profiling` option.
223 perf identifies the active program or library using the build identifier
224 stored in the elf file. This should be added at link time with the compiler
225 flag `-Wl,--build-id=sha1`.
227 The next step is to record the execution trace of the application using the
228 perf tool. The ETM strobing should be configured before running the perf
229 tool. There are two parameters:
231 * window size: A number of CPU cycles (W)
232 * period: Trace is enabled for W cycle every _period_ * W cycles.
234 For example, a typical configuration is to use a window size of 5000 cycles
235 and a period of 10000 - this will collect 5000 cycles of trace every 50M
236 cycles. With these proof-of-concept patches, the strobe parameters are
237 configured via sysfs - each ETM will have `strobe_window` and
238 `strobe_period` parameters in `/sys/bus/coresight/devices/NNNNNNNN.etm` and
239 these values will have to be written to each (In a future version, this
240 will be integrated into the drivers and perf tool). The attached `record.sh`
241 (TODO: attach!) script automates this process.
243 To collect trace from an application using ETM strobing, run:
246 taskset -c 0 ./record.sh --strobe 5000 10000 28c06000.etr ./my_application arg1 arg2
249 The taskset command is used to ensure the process stays on the same CPU
252 The raw trace can be examined using the `perf report` command:
255 perf report -D -i perf.data --stdio
261 0x1d370 [0x30]: PERF_RECORD_AUXTRACE size: 0x2003c0 offset: 0 ref: 0x39ba881d145f8639 idx: 0 tid: 4551 cpu: -1
263 . ... CoreSight ETM Trace data: size 2098112 bytes
264 Idx:0; ID:12; I_ASYNC : Alignment Synchronisation.
265 Idx:12; ID:12; I_TRACE_INFO : Trace Info.; INFO=0x0
266 Idx:17; ID:12; I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0xFFFF000008A4991C;
267 Idx:48; ID:14; I_ASYNC : Alignment Synchronisation.
268 Idx:60; ID:14; I_TRACE_INFO : Trace Info.; INFO=0x0
269 Idx:65; ID:14; I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0xFFFF000008A4991C;
270 Idx:96; ID:14; I_ASYNC : Alignment Synchronisation.
271 Idx:108; ID:14; I_TRACE_INFO : Trace Info.; INFO=0x0
272 Idx:113; ID:14; I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0xFFFF000008A4991C;
273 Idx:122; ID:14; I_TRACE_ON : Trace On.
274 Idx:123; ID:14; I_ADDR_CTXT_L_64IS0 : Address & Context, Long, 64 bit, IS0.; Addr=0x0000000000407B00; Ctxt: AArch64,EL0, NS;
275 Idx:134; ID:14; I_ATOM_F3 : Atom format 3.; ENN
276 Idx:135; ID:14; I_ATOM_F5 : Atom format 5.; NENEN
277 Idx:136; ID:14; I_ATOM_F5 : Atom format 5.; ENENE
278 Idx:137; ID:14; I_ATOM_F5 : Atom format 5.; NENEN
279 Idx:138; ID:14; I_ATOM_F3 : Atom format 3.; ENN
280 Idx:139; ID:14; I_ATOM_F3 : Atom format 3.; NNE
281 Idx:140; ID:14; I_ATOM_F1 : Atom format 1.; E
285 The execution trace is then converted to an instruction profile using
286 the perf build with trace decode support. This may be done on a different
287 machine than that which collected the trace (e.g. when cross compiling for
288 an embedded target). The `perf inject` command
289 decodes the execution trace and generates periodic instruction samples,
290 with branch histories:
293 perf inject -i perf.data -o inj.data --itrace=i100000il
296 The `--itrace` option configures the instruction sample behaviour:
298 * `i100000i` generates an instruction sample every 100000 instructions
299 (only instruction count periods are currently supported, future versions
300 may support time or cycle count periods)
301 * `l` includes the branch histories on each sample
302 * `b` generates a sample on each branch (not used here)
304 Perf requires the original program binaries to decode the execution trace.
305 If running the `inject` command on a different system than the trace was
306 captured on, then the binary and any shared libraries must be added to
310 perf buildid-cache -a /path/to/binary_or_library
313 `perf report` can also be used to show the instruction samples:
316 perf report -D -i inj.data --stdio
318 0x1528 [0x630]: PERF_RECORD_SAMPLE(IP, 0x2): 4551/4551: 0x434b98 period: 3093 addr: 0
319 ... branch stack: nr:64
320 ..... 0: 0000000000434b58 -> 0000000000434b68 0 cycles P 0
321 ..... 1: 0000000000436a88 -> 0000000000434b4c 0 cycles P 0
322 ..... 2: 0000000000436a64 -> 0000000000436a78 0 cycles P 0
323 ..... 3: 00000000004369d0 -> 0000000000436a60 0 cycles P 0
324 ..... 4: 000000000043693c -> 00000000004369cc 0 cycles P 0
325 ..... 5: 00000000004368a8 -> 0000000000436928 0 cycles P 0
326 ..... 6: 000000000042d070 -> 00000000004368a8 0 cycles P 0
327 ..... 7: 000000000042d108 -> 000000000042d070 0 cycles P 0
329 ..... 57: 0000000000448ee0 -> 0000000000448f24 0 cycles P 0
330 ..... 58: 0000000000448ea4 -> 0000000000448ebc 0 cycles P 0
331 ..... 59: 0000000000448e20 -> 0000000000448e94 0 cycles P 0
332 ..... 60: 0000000000448da8 -> 0000000000448ddc 0 cycles P 0
333 ..... 61: 00000000004486f4 -> 0000000000448da8 0 cycles P 0
334 ..... 62: 00000000004480fc -> 00000000004486d4 0 cycles P 0
335 ..... 63: 0000000000448658 -> 00000000004480ec 0 cycles P 0
336 ... thread: program1:4551
337 ...... dso: /home/root/program1
341 The instruction samples produced by `perf inject` is then passed to the
342 autofdo tool to generate source level profiles for the compiler. For
346 create_llvm_prof -binary=/path/to/binary -profile=inj.data -out=program.llvmprof
352 create_gcov -binary=/path/to/binary -profile=inj.data -gcov_version=1 -gcov=program.gcov
355 The profiles can be viewed with:
358 llvm-profdata show -sample program.llvmprof
364 dump_gcov -gcov_version=1 program.gcov
367 ## Using profile in the compiler
369 The profile produced by the above steps can then be passed to the compiler
370 to optimize the next build of the program.
372 For GCC, use the `-fauto-profile` option:
375 gcc -O2 -fauto-profile=program.gcov -o program program.c
378 For Clang, use the `-fprofile-sample-use` option:
381 clang -O2 -fprofile-sample-use=program.llvmprof -o program program.c
387 The basic commands to run an application and create a compiler profile are:
390 taskset -c 0 ./record.sh --strobe 5000 10000 28c06000.etr ./my_application arg1 arg2
391 perf inject -i perf.data -o inj.data --itrace=i100000il
392 create_llvm_prof -binary=/path/to/binary -profile=inj.data -out=program.llvmprof
395 Use `create_gcov` for gcc.
400 * AutoFDO tool: <https://github.com/google/autofdo>
401 * Build fix: <https://github.com/google/autofdo/pull/73>
402 * GCC's wiki on autofdo: <https://gcc.gnu.org/wiki/AutoFDO>, <https://gcc.gnu.org/wiki/AutoFDO/Tutorial>
403 * Google paper: <https://ai.google/research/pubs/pub45290>
404 * CoreSight kernel docs: Documentation/trace/coresight.txt
410 * Record simple program (e.g. /bin/ls)
411 * examine raw trace - look for overflows, corruption
412 * Check no errors reported
413 * mmap error indicates no route from source to sink - bad device tree
415 * data loss warning - bandwidth problems
416 * What if data loss is reported?
417 * Don't worry - strobing
420 ## Appendix: Describing CoreSight in Devicetree
423 Each component has an entry in the device tree that describes its:
425 * type: The `compatible` field defines which driver to use
426 * location: A `reg` defines the component's address and size on the bus
427 * clocks: The `clocks` and `clock-names` fields state which clock provides
428 the `apb_pclk` clock.
429 * connections to other components: `port` and `ports` field link the
430 component to ports of other components
432 To create the device tree, some information about the platform is required:
434 * The memory address of the CoreSight components. This is the address in
435 the CPU's address space where the CPU can access each CoreSight
437 * The connections between the components.
439 This information can be found in the SoC's reference manual or you may need
440 to ask the platform/SoC vendor to supply it.
442 An ETMv4 source is declared with a section like this:
446 compatible = "arm,coresight-etm4x", "arm,primecell";
447 reg = <0 0x22040000 0 0x1000>;
450 clocks = <&soc_smc50mhz>;
451 clock-names = "apb_pclk";
453 cluster0_etm0_out_port: endpoint {
454 remote-endpoint = <&cluster0_funnel_in_port0>;
460 This describes an ETMv4 attached to core A72_0, located at 0x22040000, with
461 its output linked to port 0 of a funnel. The funnel is described with:
464 funnel@220c0000 { /* cluster0 funnel */
465 compatible = "arm,coresight-funnel", "arm,primecell";
466 reg = <0 0x220c0000 0 0x1000>;
468 clocks = <&soc_smc50mhz>;
469 clock-names = "apb_pclk";
470 power-domains = <&scpi_devpd 0>;
472 #address-cells = <1>;
477 cluster0_funnel_out_port: endpoint {
478 remote-endpoint = <&main_funnel_in_port0>;
484 cluster0_funnel_in_port0: endpoint {
486 remote-endpoint = <&cluster0_etm0_out_port>;
492 cluster0_funnel_in_port1: endpoint {
494 remote-endpoint = <&cluster0_etm1_out_port>;
501 This describes a funnel located at 0x220c0000, receiving data from 2 ETMs
502 and sending the merged data to another funnel. We continue describing
503 components with similar blocks until we reach the sink (an ETR):
507 compatible = "arm,coresight-tmc", "arm,primecell";
508 reg = <0 0x20070000 0 0x1000>;
509 iommus = <&smmu_etr 0>;
511 clocks = <&soc_smc50mhz>;
512 clock-names = "apb_pclk";
513 power-domains = <&scpi_devpd 0>;
515 etr_in_port: endpoint {
517 remote-endpoint = <&replicator_out_port1>;
523 Full descriptions of the properties of each component can be found in the
524 Linux source at Documentation/devicetree/bindings/arm/coresight.txt.
525 The Arm Juno platform's devicetree (arch/arm64/boot/dts/arm) provides an example
526 description of CoreSight description.
528 Many systems include a TPIU for off-chip trace. While this isn't required
529 for self-hosted trace, it should still be included in the devicetree. This
530 allows the drivers to access it to ensure it is put into a disabled state,
531 otherwise it may limit the trace bandwidth causing data loss.