:keywords:ROCprofiler-SDK tool usage, rocprofv3 user manual, rocprofv3 usage, rocprofv3 user guide, using rocprofv3, ROCprofiler-SDK tool user guide, ROCprofiler-SDK tool user manual, using ROCprofiler-SDK tool, ROCprofiler-SDK command-line tool, ROCprofiler-SDK CLI, ROCprofiler-SDK command line tool
``rocprofv3`` is a CLI tool that helps you optimize applications and analyze the low-level kernel details without requiring any modification in the source code.
It's backward compatible with its predecessor, `rocprof <https://rocm.docs.amd.com/projects/rocprofiler/en/latest/index.html>`_, and provides enhanced features for application profiling with better accuracy.
- | Specifies the path to the input file. JSON and YAML formats support configuration of all command-line options for tracing and profiling whereas the text format supports only the specification of HW counters. |br| |br|
| Specifies output file name. If nothing is specified, the default path is ``%hostname%/%pid%``. |br| |br|
| Specifies the output path for saving the output files. If nothing is specified, the default path is ``%hostname%/%pid%``. |br| |br|
- | Collects tracing data for HIP runtime API, marker (ROCTx) API, RCCL API, memory operations (copies, scratch, and allocation), and kernel dispatches. Similar to ``--sys-trace`` but without HIP compiler API and the underlying HSA API tracing. |br| |br|
| Collects tracing data for HIP API, HSA API, marker (ROCTx) API, RCCL API, memory operations (copies, scratch, and allocations), and kernel dispatches.
- | Enables PC sampling and sets the ROCPROFILER_PC_SAMPLING_BETA_ENABLED environment variable. Note that PC sampling support is in beta version. |br| |br|
| Specifies the unit for PC sampling type or method. Note that only units of time are supported. |br| |br|
| Specifies the PC sampling type. Note that only host trap method is supported. |br| |br|
- | Combination of ``--hip-runtime-trace`` and ``--hip-compiler-trace``. This option only enables the HIP API tracing. Unlike previous iterations of ``rocprof``, this option doesn't enable kernel tracing, memory copy tracing, and so on. |br| |br|
| Collects marker (ROCTx) traces. Similar to ``--roctx-trace`` option in earlier ``rocprof`` versions, but with improved ``ROCTx`` library with more features. |br| |br|
| Collects kernel dispatch traces. |br| |br|
| Collects memory copy traces. This was a part of the HIP and HSA traces in previous ``rocprof`` versions. |br| |br|
| Collects memory allocation traces. Displays starting address, allocation size, and the agent where allocation occurs. |br| |br|
| Collects scratch memory operations traces. Helps in determining scratch allocations and manage them efficiently. |br| |br|
| Collects ``--hsa-core-trace``, ``--hsa-amd-trace``, ``--hsa-image-trace``, and ``--hsa-finalizer-trace``. This option only enables the HSA API tracing. Unlike previous iterations of ``rocprof``, this doesn't enable kernel tracing, memory copy tracing, and so on. |br| |br|
| Collects traces for RCCL (ROCm Communication Collectives Library), which is also pronounced as 'Rickle'. |br| |br|
| Enables builtin Kokkos tools support, which implies enabling ``--marker-trace`` collection and ``--kernel-rename``. |br| |br|
- | Collects HIP Runtime API traces. For example, public HIP API functions starting with ``hip`` such as ``hipSetDevice``. |br| |br|
| Collects HIP Compiler generated code traces. For example, HIP API functions starting with ``__hip`` such as ``__hipRegisterFatBinary``. |br| |br|
| Collects HSA API traces (core API). For example, HSA functions prefixed with only ``hsa_`` such as ``hsa_init``. |br| |br|
| Collects HSA API traces (AMD-extension API). For example, HSA functions prefixed with ``hsa_amd_`` such as ``hsa_amd_coherency_get_type``. |br| |br|
| Collects HSA API traces (image-extenson API). For example, HSA functions prefixed with only ``hsa_ext_image_`` such as ``hsa_ext_image_get_capability``. |br| |br|
| Collects HSA API traces (Finalizer-extension API). For example, HSA functions prefixed with only ``hsa_ext_program_`` such as ``hsa_ext_program_create``.
* - Counter collection
- | ``--pmc`` [PMC ...]
- | Specifies performance monitoring counters to be collected. Use comma or space to specify more than one counter. Also note that the job fails if the entire set of counters can't be collected in single pass.
- | Collects statistics of enabled tracing types. Must be combined with one or more tracing options. Doesn't include default kernel stats unlike previous ``rocprof`` versions. |br| |br|
| Displays single summary of tracing data for the enabled tracing type, after conclusion of the profiling session. Displays a summary of tracing data for the enabled tracing type, after conclusion of the profiling session. |br| |br|
| Displays a summary of each tracing domain for the enabled tracing type, after conclusion of the profiling session. |br| |br|
| Displays a summary for each set of domains matching the specified regular expression. For example, 'KERNEL_DISPATCH\|MEMORY_COPY' generates a summary of all the tracing data in the `KERNEL_DISPATCH` and `MEMORY_COPY` domains. Similarly '\*._API' generates a summary of all the tracing data in the ``HIP_API``, ``HSA_API``, and ``MARKER_API`` domains.
- | Overrides the default demangling of kernel names. |br| |br|
| Truncates the demangled kernel names for improved readability. In earlier ``rocprof`` versions, this was known as ``--basenames [on/off]``. |br| |br|
| Uses region names defined using ``roctxRangePush`` or ``roctxRangePop`` to rename the kernels. Was known as ``--roctx-rename`` in earlier ``rocprof`` versions.
- | Filters counter-collection and thread-trace data to include the kernels matching the specified regular expression. Non-matching kernels are excluded. |br| |br|
| Filters counter-collection and thread-trace data to exclude the kernels matching the specified regular expression. It is applied after ``--kernel-include-regex`` option. |br| |br|
| Specifies iteration range for each kernel matching the filter [start-stop]. |br| |br| |br|
| START_DELAY_TIME\: Time in seconds before the data collection begins. |br| COLLECTION_TIME\: Duration of data collection in seconds. |br| REPEAT\: Number of times the data collection cycle is repeated. |br| The default unit for time is seconds, which can be changed using the ``--collection-period-unit`` option. To repeat the cycle indefinitely, specify ``repeat`` as 0. You can specify multiple configurations, each defined by a triplet in the format ``start_delay_time:collection_time:repeat``. For example, the command ``-P 10:10:1 5:3:0`` specifies two configurations, the first one with a start delay time of 10 seconds, a collection time of 10 seconds, and a repeat of 1 (the cycle repeats once), and the second with a start delay time of 5 seconds, a collection time of 3 seconds, and a repeat of 0 (the cycle repeats indefinitely). |br| |br| |br|
| To change the unit of time used in ``--collection-period`` or ``-P``, specify the desired unit using the ``--collection-period-unit`` option. The available units are ``hour`` for hours, ``min`` for minutes, ``sec`` for seconds, ``msec`` for milliseconds, ``usec`` for microseconds, and ``nsec`` for nanoseconds.
- | Specifies backend for Perfetto data collection. When selecting 'system' mode, ensure to run the Perfetto ``traced`` daemon and then start a Perfetto session. |br| |br|
| Specifies buffer size for Perfetto output in KB. Default: 1 GB. |br| |br|
| Specifies policy for handling new records when Perfetto reaches the buffer limit. |br| |br|
| Specifies Perfetto shared memory size hint in KB. Default: 64 KB.
- | Lists the PC sampling configurations and metrics available in the counter_defs.yaml file for counter collection. In earlier ``rocprof`` versions, this was known as ``--list-basic``, ``--list-derived``, and ``--list-counters``. |br| |br|
| For displaying the HSA Queues that kernels and memory copy operations are submitted to rather than the default grouping of HIP Streams for perfetto.
- | Specifies libraries to prepend to ``LD_PRELOAD``. It is useful for sanitizer libraries. |br| |br|
| Output files are generated only if output data size is greater than minimum output data size. It can be used for controlling the generation of output files so that user don't recieve empty files. The input is in KB units. |br| |br|
| Disables the signal handlers in the rocprofv3 tool. It disables the prioritizing of rocprofv3 signal handler over application installed signal handler. When --disable-signal-handlers is set to true, and application has its signal handler on SIGSEGV or similar installed, then its signal handler will be used not the rocprofv3 signal handler. Note: glog still installs signal handlers which provide backtraces.
Application tracing provides the big picture of a program’s execution by collecting data on the execution times of API calls and GPU commands, such as kernel execution, async memory copy, and barrier packets. This information can be used as the first step in the profiling process to answer important questions, such as how much percentage of time was spent on memory copy and which kernel took the longest time to execute.
To use ``rocprofv3`` for application tracing, run:
HIP trace comprises execution traces for the entire application at the HIP level. This includes HIP API functions and their asynchronous activities at the runtime level. In general, HIP APIs directly interact with the user program. It is easier to analyze HIP traces as you can directly map them to the program.
Unlike previous iterations of `rocprof`, this does not enable kernel tracing, memory copy tracing, etc. If you want to enable kernel tracing, memory copy tracing, they need to be provided explicitly.
`rocprofv3` provides options to collect traces at more granular level. For HIP, user can collect traces for HIP compile time APIs and runtime APIs separately.
For the description of the fields in the output file, see :ref:`output-file-fields`.
HSA trace
+++++++++++++
The HIP runtime library is implemented with the low-level HSA runtime. HSA API tracing is more suited for advanced users who want to understand the application behavior at the lower level. In general, tracing at the HIP level is recommended for most users. You should use HSA trace only if you are familiar with HSA runtime.
HSA trace contains the start and end time of HSA runtime API calls and their asynchronous activities.
The preceding command generates a ``hsa_api_trace.csv`` file prefixed with process ID. Note that the contents of this file have been truncated for demonstration purposes.
`rocprofv3` provides options to collect traces at more granular level. HSA traces can be collected separately for four API domains: `HSA_AMD_EXT_API`, `HSA_CORE_API`, `HSA_IMAGE_EXT_API` and `HSA_FINALIZE_EXT_API`.
The preceding command generates a ``hsa_api_trace.csv`` file prefixed with process ID. Note that the contents of this file have been truncated for demonstration purposes.
To use ``rocprofv3`` for marker tracing, including and linking to old ``ROCTx`` works but it's recommended to switch to the new ``ROCTx`` to utilize new APIs.
To use the new ``ROCTx``, include header ``"rocprofiler-sdk-roctx/roctx.h"`` and link your application with ``librocprofiler-sdk-roctx.so``.
To see the complete list of ``ROCTx`` APIs, see public header file ``"rocprofiler-sdk-roctx/roctx.h"``.
`Kokkos <https://github.com/kokkos/kokkos>`_ is a C++ library for writing performance portable applications. Kokkos is widely used in scientific applications to write performance-portable code for CPUs, GPUs, and other accelerators.
``rocprofv3`` loads an inbuilt `Kokkos Tools library <https://github.com/kokkos/kokkos-tools>`_, which emits roctx ranges with the labels passed using Kokkos APIs. For example, ``Kokkos::parallel_for(“MyParallelForLabel”, …)`` calls ``roctxRangePush`` internally and enables the kernel renaming option to replace the highly templated kernel names with the Kokkos labels.
To enable the inbuilt marker support, use the ``kokkos-trace`` option. Internally, this option automatically enables ``marker-trace`` and ``kernel-rename``:
Memory copy traces track `hipMemcpy` and `hipMemcpyAsync` functions, which uses the `hsa_amd_memory_async_copy_on_engine` HSA functions internally. To trace memory moves across the application, use:
thus, tends to be an implementation detail irrelevant to most users. Similarly, the HIP compiler API is also excluded for being an implementation detail as these functions are automatically inserted during HIP compilation.
Running the preceding command generates ``hip_api_trace.csv``, ``kernel_trace.csv``, ``memory_copy_trace.csv``, ``scratch_memory_trace.csv``, ``memory_allocation_trace.csv``, and ``marker_api_trace.csv`` (if ``ROCTx`` APIs are specified in the application) files prefixed with the process ID.
This is an all-inclusive option to collect HIP, HSA, kernel, memory copy, memory allocation, and marker trace (if ``ROCTx`` APIs are specified in the application).
Running the above command generates ``hip_api_trace.csv``, ``hsa_api_trace.csv``, ``kernel_trace.csv``, ``memory_copy_trace.csv``, ``scratch_memory_trace.csv``, ``memory_allocation_trace.csv``, and ``marker_api_trace.csv`` (if ``ROCTx`` APIs are specified in the application).
This option collects scratch memory operation traces. Scratch is an address space on AMD GPUs roughly equivalent to the `local memory` in NVIDIA CUDA. The `local memory` in CUDA is a thread-local global memory with interleaved addressing, which is used for register spills or stack space. This option helps to trace when the ``rocr`` runtime allocates, frees, and tries to reclaim scratch memory.
`RCCL <https://github.com/ROCm/rccl>`_ (pronounced "Rickle") is a stand-alone library of standard collective communication routines for GPUs. This option traces those communication routines.
The preceding command generates a ``hip_trace.csv`` and ``hsa_trace.csv`` file prefixed with the process ID along with displaying the summary of each domain.
This option displays a summary of multiple domains for the domain names specified on the command line. The summary groups can be separated using a pipe ( | ) symbol.
The preceding command generates a ``summary`` file displaying the summary of each domain. This also generates the files for the enabled tracing types under `-sys-trace` option.
The preceding sections describe how to collect traces by specifying the desired tracing type on the command line. You can also specify the desired tracing types in an input file in YAML (.yaml/.yml), or JSON (.json) format. You can supply any command-line option for tracing in the input file.
When using aggregate tracing options like ``--runtime-trace`` or ``--sys-trace``, you can disable specific tracing options by setting them to ``False``. This allows fine-grained control over which traces are collected.
This feature is particularly useful when you want to collect most traces but exclude specific ones that might be unnecessary for your analysis or that generate excessive data.
The application tracing functionality allows you to evaluate the duration of kernel execution but is of little help in providing insight into kernel execution details. The kernel counter collection functionality allows you to select kernels for profiling and choose the basic counters or derived metrics to be collected for each kernel execution, thus providing a greater insight into kernel execution.
AMDGPUs are equipped with hardware performance counters that can be used to measure specific values during kernel execution, which are then exported from the GPU and written into the output files at the end of the kernel execution. These performance counters vary according to the GPU. Therefore, it is recommended to examine the hardware counters that can be collected before running the profile.
There are two types of data available for profiling: hardware basic counters and derived metrics.
The derived metrics are the counters derived from the basic counters using mathematical expressions. Note that the basic counters and derived metrics are collectively referred as counters in this document.
To see the counters available on the GPU, use:
..code-block::shell
rocprofv3 --list-avail
You can also customize the counters according to the requirement. Such counters are named :ref:`extra-counters`.
For a comprehensive list of counters available on MI200, see `MI200 performance counters and metrics <https://rocm.docs.amd.com/en/latest/conceptual/gpu-arch/mi300-mi200-performance-counters.html>`_.
When using input file in text format, the line consisting of the counter names must begin with ``pmc``. The number of counters that can be collected in one profiling run are limited by the GPU hardware resources. If too many counters are selected, the kernels need to be executed multiple times(multi-pass execution) to collect all the counters. For multi-pass execution, include multiple ``pmc`` rows in the input file. Counters in each ``pmc`` row can be collected in each application run.
While the input file in text format can only be used for counter collection, JSON and YAML formats support all the command-line options for profiling. The input file in YAML or JSON format has an array of profiling configurations called jobs. Each job is used to configure profiling for an application execution.
Here is the input schema (properties) of JSON or YAML input files:
While the basic counters and derived metrics are available for collection by default, you can also define counters as per requirement. These user-defined counters with custom definitions are named extra counters.
Using ``rocprofv3`` for counter collection using input file or command line generates a ``./pmc_n/counter_collection.csv`` file prefixed with the process ID. For each ``pmc`` row, a directory ``pmc_n`` containing a ``counter_collection.csv`` file is generated, where n = 1 for the first row and so on.
When using input file in JSON or YAML format, for each job, a directory ``pass_n`` containing a ``counter_collection.csv`` file is generated, where n = 1 for the first job and so on.
Counter multiplexing allows a single run of the program to collect groups of counters. This is useful when the counters you want to collect exceed the hardware limits and you cannot run the program multiple times for collection.
This feature is available when using YAML (.yaml/.yml) or JSON (.json) input formats. Two new fields are introduced, ``pmc_groups`` and ``pmc_group_interval``. The ``pmc_groups`` field is used to specify the groups of counters to be collected in each run. The ``pmc_group_interval`` field is used to specify the interval between each group of counters. Interval is per-device and increments per dispatch on the device (i.e. dispatch_id). When the interval is reached the next group is selected.
Here is a sample input.yaml file for specifying counter multiplexing:
This sample input will collect the first group of counters (``SQ_WAVES``, ``GRBM_COUNT``) for the first 4 kernel executions on the device, then the second group of counters (``GRBM_GUI_ACTIVE``) for the next 4 kernel executions on the device, and so on.
An example of the interval period for this input is given below:
Users can generate Perfetto trace files using the ``--output-format pftrace`` option. This allows users to visualize the traces in the Perfetto viewer.
Perfetto is an open-source tracing tool that provides a detailed view of system performance. It allows you to visualize the collected traces in a user-friendly interface, making it easier to analyze and understand the performance characteristics of your application.
To generate a Perfetto trace file, use the ``--output-format pftrace`` option along with the desired tracing options. For example, to collect system traces and generate a Perfetto trace file, use:
When collecting performance counter data, you can visualize the counter tracks per agent in the Perfetto viewer by using the PFTrace output format. This allows you to see how counter values change over time during kernel execution.
To generate a Perfetto trace file with counter data, use:
The generated Perfetto trace file can be opened in the `Perfetto UI <https://ui.perfetto.dev/>`_. In the viewer, performance counters will appear as counter tracks organized by agent, allowing you to visualize counter values changing over time alongside kernel executions and other traced activities.
You can also combine this with the system trace option to get a more comprehensive view of the system's performance. For example, you can use the following command to collect both system trace and performance counter data:
The agent index is a unique identifier for each agent in the system. It is used to identify the agent in the output files. Since, each runtime/tool has their own way to represent the agent's indices, `rocprofv3` provides a way to configure the agent index in the output files.
-**absolute** == *node_id* - absolute index of the agent regardless of cgroups masking. This is a monotonically increasing number that is incremented for every folder in `/sys/class/kfd/kfd/topology/nodes`. e.g. Agent-0, Agent-2, Agent-4.
-**relative** == *logical_node_id* - relative index of the agent accounting for cgroups masking. This is a monotonically increasing number which is incremented for every folder in `/sys/class/kfd/kfd/topology/nodes/` whose properties file was non-empty.e.g. Agent-0, Agent-1, Agent-2
-**type-relative** == *logical_node_type_id* - relative index of the agent accounting for cgroups masking where indexing starts at zero for each agent type. e.g. CPU-0, GPU-0, GPU-1
By default, `rocprofv3` shows which HIP streams kernel and memory copy operations were submitted to when outputting a perfetto trace. The ``--group-by-queue`` option allows users to display the HSA queues that these kernel and memory operations were submitted to instead.
Kernel name truncation is a feature that allows you to limit the length of kernel names in the output files. This is useful when dealing with long kernel names that can make the output files difficult to read.
To enable kernel name truncation, use the ``--truncate-kernels`` option.
Kernel filtering allows you to include or exclude the kernels for profiling by specifying a filter using a regex string. You can also specify an iteration range for profiling the included kernels. If the iteration range is not provided, then all iterations of the included kernels are profiled.
The ``roctxRangePush`` and ``roctxRangePop`` also let you rename the enclosed kernel with the supplied message. In the legacy ``rocprof``, this functionality was known as ``--roctx-rename``.
See how to use ``roctxRangePush`` and ``roctxRangePop`` for renaming the enclosed kernel:
Output prefix keys are useful in multiple use cases but are most helpful when dealing with multiple profiling runs or large MPI jobs. Here is the list of available keys:
The preceding command runs the application with ``rocprofv3`` and generates the trace file for each rank. The trace files are prefixed with hostname, process ID, and MPI rank.
Assuming the hostname as `ubuntu-latest` and the process IDs as 3000020 and 3000019, the output file names are:
The collection period is the time interval during which the profiling data is collected. You can specify the collection period using the ``--collection-period`` or ``-P`` option.
The above command collects the profiling data for 1 seconds, starting after 5 seconds of the application starts, and this cycle will be repeated 1 time.
The collection period can be specified in different units, such as seconds, milliseconds, microseconds, and nanoseconds. The default unit is `seconds`. You can change the unit using the ``--collection-period-unit`` option.
The above command collects the profiling data for 1 milliseconds, starting after 5 milliseconds of the application starts, and this cycle will be repeated 1 time.
Perfetto-specific options
++++++++++++++++++++++++++
The following options are specific to Perfetto tracing and are used to control the behavior of the Perfetto data collection:
-**--perfetto-buffer-fill-policy {discard,ring_buffer}**: Policy for handling new records when perfetto has reached the buffer limit.
-**RING_BUFFER (default)**: The buffer behaves like a ring buffer and writes when full will wrap over and replace the oldest trace data in the buffer.
-**DISCARD**: The buffer stops accepting data once full. Further write attempts are dropped.
-**--perfetto-buffer-size KB**: Size of buffer for perfetto output in KB. default: 1 GB. If set, stops the tracing session after N bytes have been written. Used to cap the size of the trace.
-**--perfetto-backend {inprocess,system}**: Perfetto data collection backend. 'system' mode requires starting traced and perfetto daemons.By default Perfetto keeps the full trace buffer(s) in memory.
-**--perfetto-shmem-size-hint KB**: Perfetto shared memory size hint in KB. default: 64 KB. This option gives you control over shared memory buffer sizing. Thisoption can be tweaked to avoid data loses when data is produced at a higher rate.
- The amount of memory required in bytes for the combined private, spill, and arg segments for a work item.
* - Group_Segment_Size
- The group segment memory required by a workgroup in bytes. This does not include any dynamically allocated group segment memory that may be added when the kernel is dispatched.
* - Workgroup_Size
- Size of the workgroup as declared by the compute shader.
* - Workgroup_Size_n
- Size of the workgroup in the nth dimension as declared by the compute shader, where n = X, Y, or Z.
* - Grid_Size
- Number of thread blocks required to launch the kernel.
* - Grid_Size_n
- Number of thread blocks in the nth dimension required to launch the kernel, where n = X, Y, or Z.
* - LDS_Block_Size
- Thread block size for the kernel's Local Data Share (LDS) memory.
* - Scratch_Size
- Kernel’s scratch memory size.
* - SGPR_Count
- Kernel's Scalar General Purpose Register (SGPR) count.
Format selection is case-insensitive and multiple output formats are supported. While ``--output-format json`` exclusively enables JSON output, ``--output-format csv json pftrace otf2`` enables all four output formats for the run.
