AutoFDO and ARM Trace

Using CoreSight trace and perf with OpenCSD for AutoFDO.

Introduction

Feedback directed optimization (FDO, also know as profile guided optimization - PGO) uses a profile of a program's execution to guide the optmizations performed by the compiler. Traditionally, this involves building an instrumented version of the program, which records a profile of execution as it runs. The instrumentation adds significant runtime overhead, possibly changing the behaviour of the program and it may not be possible to run the instrumented program in a production environment (e.g. where performance criteria must be met).

AutoFDO uses facilities in the hardware to sample the behaviour of the program in the production environment and generate the execution profile. An improved profile can be obtained by including the branch history (i.e. a record of the last branches taken) when generating an instruction samples. On Arm systems, the ETM can be used to generate such records.

The process can be broken down into the following steps:

• Record execution trace of the program
• Convert the execution trace to instruction samples with branch histories
• Convert the instruction samples to source level profiles
• Use the source level profile with the compiler

This article describes how to enable ETM trace on Arm targets running Linux and use the ETM trace to generate AutoFDO profiles and compile an optimized program.

Execution trace on Arm targets

Debug and trace of Arm targets is provided by CoreSight. This consists of a set of components that allow access to debug logic, record (trace) the execution of a processor and route this data through the system, collecting it into a store.

To record the execution of a processor, we require the following components:

• A trace source. The core contains a trace unit, called an ETM that emits data describing the instructions executed by the core.
• Trace links. The trace data generated by the ETM must be moved through the system to the component that collects the data (sink). Links include:
  • Funnels: merge multiple streams of data
  • FIFOs: buffer data to smooth out bursts
  • Replicators: send a stream of data to multiple components
• Sinks. These receive the trace data and store it or send it to an external device:
  • ETB: A small circular buffer (64-128 kilobytes) that stores the most recent data
  • ETR: A larger (several megabytes) buffer that uses system RAM to store data
  • TPIU: Sends data to an off-chip capture device (e.g. Arm DSTREAM)

Each Arm SoC design may have a different layout (topology) of components. This topology is described to the OS drivers by the platform's devicetree or (in future) ACPI firmware.

For application profiling, we need to store several megabytes of data within the system, so will use ETR with the capture tool (perf) periodically draining the buffer to a file.

Even though we have a large capture buffer, the ETM can still generate a lot of data very quickly - typically an ETM will generate ~1 bit of data per instruction (depending on the workload), which results in 256Mbytes per second for a core running at 2GHz. This leads to problems storing and decoding such large volumes of data. AutoFDO uses samples of program execution, so we can avoid this problem by using the ETM's features to only record small slices of execution - e.g. collect ~5000 cycles of data every 50M cycles. This reduces the data rate to a manageable level - a few megabytes per minute. This technique is known as 'strobing'.

Enabling trace

Driver support

To collect ETM trace, the CoreSight drivers must be included in the kernel. Some of the driver support is not yet included in the mainline kernel and many targets are using older kernels. To enable CoreSight trace on these targets, Arm have provided backports of the latest CoreSight drivers and ETM strobing patch at:

[TODO: link to git repos for CoreSight backports]

You can include these backports in your kernel by either merging the appropriate branch using git or generating patches (using git format-patch).

For 4.9 based kernels, use the coresight-4.9-etr-etm_strobe branch:

git merge coresight-4.9-etr-etm_strobe

or

git format-patch --output-directory /output/dir v4.9..coresight-4.9-etr-etm_strobe
cd my_kernel
git am /output/dir/*.patch # or patch -p1 /output/dir/*.patch if not using git

For 4.14 based kernels, use the coresight-4.14-etm_strobe branch:

git merge coresight-4.14-etm_strobe

or

git format-patch --output-directory /output/dir v4.14..coresight-4.14-etm_strobe
cd my_kernel
git am /output/dir/*.patch # or patch -p1 /output/dir/*.patch if not using git

The CoreSight trace drivers must also be enabled in the kernel configuration. This can be done using the configuration menu (make menuconfig), selecting Kernel hacking / CoreSight Tracing Support and enabling all options, or by setting the following in the configuration file:

CONFIG_CORESIGHT=y
CONFIG_CORESIGHT_LINK_AND_SINK_TMC=y
CONFIG_CORESIGHT_SINK_TPIU=y
CONFIG_CORESIGHT_SOURCE_ETM4X=y
CONFIG_CORESIGHT_DYNAMIC_REPLICATOR=y
CONFIG_CORESIGHT_STM=y
CONFIG_CORESIGHT_CATU=y

Compile the kernel for your target in the usual way, e.g.

make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- 

Each target may have a different layout of CoreSight components. To collect trace into a sink, the kernel drivers need to know which other devices need to be configured to route data from the source to the sink. This is described in the devicetree (and in future, the ACPI tables). The device tree will define which CoreSight devices are present in the system, where they are located and how they are connected together. The devicetree for some platforms includes a description of the platform's CoreSight components, but in other cases you may have to ask the platform/SoC vendor to supply it or create it yourself (see Appendix: Describing CoreSight in Devicetree).

Once the target has been booted with the devicetree describing the CoreSight devices, you should find the devices in sysfs:

# ls /sys/bus/coresight/devices/
28440000.etm  28540000.etm  28640000.etm  28740000.etm
28c03000.funnel  28c04000.etf  28c05000.replicator  28c06000.etr
28c07000.tpiu

Perf tools

The perf tool is used to capture execution trace, configuring the trace sources to generate trace, routing the data to the sink and collecting the data from the sink.

Arm recommends to use the perf version corresponding to the kernel running on the target. This can be built from the same kernel sources with

make -C tools/perf ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- 

If the post-processing (perf inject) of the captured data is not being done on the target, then the OpenCSD library is not required for this build of perf.

Trace is captured by collecting the cs_etm event from perf. The sink to collect data into is specified as a parameter of this event. Trace can also be restricted to user space or kernel space with 'u' or 'k' parameters. For example:

perf record -e cs_etm/@28c06000.etr/u --per-thread -- /bin/ls

Will record the userspace execution of '/bin/ls' into the ETR located at 0x28c06000. Note the --per-thread option is required - perf currently only supports trace of a single thread of execution. CPU wide trace is a work in progresss.

Processing trace and profiles

perf is also used to convert the execution trace an instruction profile. This requires a different build of perf, using the version of perf from Linux v4.17 or later, as the trace processing code isn't included in the driver backports. Trace decode is provided by the OpenCSD library (https://github.com/Linaro/OpenCSD), v0.9.1 or later. This is packaged for debian testing (install the libopencsd0, libopencsd-dev packages) or can be compiled from source and installed.

The autoFDO tool https://github.com/google/autofdo is used to convert the instruction profiles to source profiles for the GCC and clang/llvm compilers.

Recording and profiling

Once trace collection using perf is working, we can now use it to profile an application.

The application must be compiled to include sufficient debug information to map instructions back to source lines. For GCC, use the -g1 or -gmlt options. For clang/llvm, also add the -fdebug-info-for-profiling option.

perf identifies the active program or library using the build identifier stored in the elf file. This should be added at link time with the compiler flag -Wl,--build-id=sha1.

The next step is to record the execution trace of the application using the perf tool. The ETM strobing should be configured before running the perf tool. There are two parameters:

• window size: A number of CPU cycles (W)
• period: Trace is enabled for W cycle every period * W cycles.

For example, a typical configuration is to use a window size of 5000 cycles and a period of 10000 - this will collect 5000 cycles of trace every 50M cycles. With these proof-of-concept patches, the strobe parameters are configured via sysfs - each ETM will have strobe_window and strobe_period parameters in /sys/bus/coresight/devices/NNNNNNNN.etm and these values will have to be written to each (In a future version, this will be integrated into the drivers and perf tool). The attached record.sh (TODO: attach!) script automates this process.

To collect trace from an application using ETM strobing, run:

taskset -c 0 ./record.sh --strobe 5000 10000 28c06000.etr ./my_application arg1 arg2

The taskset command is used to ensure the process stays on the same CPU during execution.

The raw trace can be examined using the perf report command:

perf report -D -i perf.data --stdio

For example:

0x1d370 [0x30]: PERF_RECORD_AUXTRACE size: 0x2003c0  offset: 0  ref: 0x39ba881d145f8639  idx: 0  tid: 4551  cpu: -1

. ... CoreSight ETM Trace data: size 2098112 bytes
        Idx:0; ID:12;   I_ASYNC : Alignment Synchronisation.
        Idx:12; ID:12;  I_TRACE_INFO : Trace Info.; INFO=0x0
        Idx:17; ID:12;  I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0xFFFF000008A4991C; 
        Idx:48; ID:14;  I_ASYNC : Alignment Synchronisation.
        Idx:60; ID:14;  I_TRACE_INFO : Trace Info.; INFO=0x0
        Idx:65; ID:14;  I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0xFFFF000008A4991C; 
        Idx:96; ID:14;  I_ASYNC : Alignment Synchronisation.
        Idx:108; ID:14; I_TRACE_INFO : Trace Info.; INFO=0x0
        Idx:113; ID:14; I_ADDR_L_64IS0 : Address, Long, 64 bit, IS0.; Addr=0xFFFF000008A4991C; 
        Idx:122; ID:14; I_TRACE_ON : Trace On.
        Idx:123; ID:14; I_ADDR_CTXT_L_64IS0 : Address & Context, Long, 64 bit, IS0.; Addr=0x0000000000407B00; Ctxt: AArch64,EL0, NS; 
        Idx:134; ID:14; I_ATOM_F3 : Atom format 3.; ENN
        Idx:135; ID:14; I_ATOM_F5 : Atom format 5.; NENEN
        Idx:136; ID:14; I_ATOM_F5 : Atom format 5.; ENENE
        Idx:137; ID:14; I_ATOM_F5 : Atom format 5.; NENEN
        Idx:138; ID:14; I_ATOM_F3 : Atom format 3.; ENN
        Idx:139; ID:14; I_ATOM_F3 : Atom format 3.; NNE
        Idx:140; ID:14; I_ATOM_F1 : Atom format 1.; E
.....

The execution trace is then converted to an instruction profile using the perf build with trace decode support. This may be done on a different machine than that which collected the trace (e.g. when cross compiling for an embedded target). The perf inject command decodes the execution trace and generates periodic instruction samples, with branch histories:

perf inject -i perf.data -o inj.data --itrace=i100000il

The --itrace option configures the instruction sample behaviour:

• i100000i generates an instruction sample every 100000 instructions (only instruction count periods are currently supported, future versions may support time or cycle count periods)
• l includes the branch histories on each sample
• b generates a sample on each branch (not used here)

Perf requires the original program binaries to decode the execution trace. If running the inject command on a different system than the trace was captured on, then the binary and any shared libraries must be added to perf's cache with:

perf buildid-cache -a /path/to/binary_or_library

perf report can also be used to show the instruction samples:

perf report -D -i inj.data --stdio
.......
0x1528 [0x630]: PERF_RECORD_SAMPLE(IP, 0x2): 4551/4551: 0x434b98 period: 3093 addr: 0
... branch stack: nr:64
.....  0: 0000000000434b58 -> 0000000000434b68 0 cycles  P   0
.....  1: 0000000000436a88 -> 0000000000434b4c 0 cycles  P   0
.....  2: 0000000000436a64 -> 0000000000436a78 0 cycles  P   0
.....  3: 00000000004369d0 -> 0000000000436a60 0 cycles  P   0
.....  4: 000000000043693c -> 00000000004369cc 0 cycles  P   0
.....  5: 00000000004368a8 -> 0000000000436928 0 cycles  P   0
.....  6: 000000000042d070 -> 00000000004368a8 0 cycles  P   0
.....  7: 000000000042d108 -> 000000000042d070 0 cycles  P   0
.......
..... 57: 0000000000448ee0 -> 0000000000448f24 0 cycles  P   0
..... 58: 0000000000448ea4 -> 0000000000448ebc 0 cycles  P   0
..... 59: 0000000000448e20 -> 0000000000448e94 0 cycles  P   0
..... 60: 0000000000448da8 -> 0000000000448ddc 0 cycles  P   0
..... 61: 00000000004486f4 -> 0000000000448da8 0 cycles  P   0
..... 62: 00000000004480fc -> 00000000004486d4 0 cycles  P   0
..... 63: 0000000000448658 -> 00000000004480ec 0 cycles  P   0
 ... thread: program1:4551
 ...... dso: /home/root/program1
.......

The instruction samples produced by perf inject is then passed to the autofdo tool to generate source level profiles for the compiler. For clang/LLVM:

create_llvm_prof -binary=/path/to/binary -profile=inj.data -out=program.llvmprof

And for GCC:

create_gcov -binary=/path/to/binary -profile=inj.data -gcov_version=1 -gcov=program.gcov

The profiles can be viewed with:

llvm-profdata show -sample program.llvmprof

Or, for GCC:

dump_gcov -gcov_version=1 program.gcov

Using profile in the compiler

The profile produced by the above steps can then be passed to the compiler to optimize the next build of the program.

For GCC, use the -fauto-profile option:

gcc -O2 -fauto-profile=program.gcov -o program program.c

For Clang, use the -fprofile-sample-use option:

clang -O2 -fprofile-sample-use=program.llvmprof -o program program.c

Summary

The basic commands to run an application and create a compiler profile are:

taskset -c 0 ./record.sh --strobe 5000 10000 28c06000.etr ./my_application arg1 arg2
perf inject -i perf.data -o inj.data --itrace=i100000il
create_llvm_prof -binary=/path/to/binary -profile=inj.data -out=program.llvmprof

Use create_gcov for gcc.

References

• AutoFDO tool: https://github.com/google/autofdo
  • Build fix: https://github.com/google/autofdo/pull/73
• GCC's wiki on autofdo: https://gcc.gnu.org/wiki/AutoFDO, https://gcc.gnu.org/wiki/AutoFDO/Tutorial
• Google paper: https://ai.google/research/pubs/pub45290
• CoreSight kernel docs: Documentation/trace/coresight.txt

Troubleshooting

TODO:

• Record simple program (e.g. /bin/ls)
  • examine raw trace - look for overflows, corruption
• Check no errors reported
  • mmap error indicates no route from source to sink - bad device tree
  • try nearer sink
  • data loss warning - bandwidth problems
    • What if data loss is reported?
    • Don't worry - strobing

Appendix: Describing CoreSight in Devicetree

Each component has an entry in the device tree that describes its:

• type: The compatible field defines which driver to use
• location: A reg defines the component's address and size on the bus
• clocks: The clocks and clock-names fields state which clock provides the apb_pclk clock.
• connections to other components: port and ports field link the component to ports of other components

To create the device tree, some information about the platform is required:

• The memory address of the CoreSight components. This is the address in the CPU's address space where the CPU can access each CoreSight component.
• The connections between the components.

This information can be found in the SoC's reference manual or you may need to ask the platform/SoC vendor to supply it.

An ETMv4 source is declared with a section like this:

etm0: etm@22040000 {
    compatible = "arm,coresight-etm4x", "arm,primecell";
    reg = <0 0x22040000 0 0x1000>;

    cpu = <&A72_0>;
    clocks = <&soc_smc50mhz>;
    clock-names = "apb_pclk";
    port {
        cluster0_etm0_out_port: endpoint {
            remote-endpoint = <&cluster0_funnel_in_port0>;
        };
    };
};

This describes an ETMv4 attached to core A72_0, located at 0x22040000, with its output linked to port 0 of a funnel. The funnel is described with:

funnel@220c0000 { /* cluster0 funnel */
    compatible = "arm,coresight-funnel", "arm,primecell";
    reg = <0 0x220c0000 0 0x1000>;

    clocks = <&soc_smc50mhz>;
    clock-names = "apb_pclk";
    power-domains = <&scpi_devpd 0>;
    ports {
        #address-cells = <1>;
        #size-cells = <0>;

        port@0 {
            reg = <0>;
            cluster0_funnel_out_port: endpoint {
                remote-endpoint = <&main_funnel_in_port0>;
            };
        };

        port@1 {
            reg = <0>;
            cluster0_funnel_in_port0: endpoint {
                slave-mode;
                remote-endpoint = <&cluster0_etm0_out_port>;
            };
        };

        port@2 {
            reg = <1>;
            cluster0_funnel_in_port1: endpoint {
                slave-mode;
                remote-endpoint = <&cluster0_etm1_out_port>;
            };
        };
    };
};

This describes a funnel located at 0x220c0000, receiving data from 2 ETMs and sending the merged data to another funnel. We continue describing components with similar blocks until we reach the sink (an ETR):

etr@20070000 {
    compatible = "arm,coresight-tmc", "arm,primecell";
    reg = <0 0x20070000 0 0x1000>;
    iommus = <&smmu_etr 0>;

    clocks = <&soc_smc50mhz>;
    clock-names = "apb_pclk";
    power-domains = <&scpi_devpd 0>;
    port {
        etr_in_port: endpoint {
            slave-mode;
            remote-endpoint = <&replicator_out_port1>;
        };
    };
};

Full descriptions of the properties of each component can be found in the Linux source at Documentation/devicetree/bindings/arm/coresight.txt. The Arm Juno platform's devicetree (arch/arm64/boot/dts/arm) provides an example description of CoreSight description.

Many systems include a TPIU for off-chip trace. While this isn't required for self-hosted trace, it should still be included in the devicetree. This allows the drivers to access it to ensure it is put into a disabled state, otherwise it may limit the trace bandwidth causing data loss.