migrate aqlprofile docs 7.0.1 from standalone repo (#1379)

This PR migrates the aqlprofile/docs folder from standalone repo to monorepo
Link to the docs branch:
https://github.com/ROCm/aqlprofile/commits/docs/7.0.1

---------

Co-authored-by: Matt Williams <matt.williams@amd.com>
Co-authored-by: pbhandar-amd <138039281+pbhandar-amd@users.noreply.github.com>
This commit is contained in:
Saurabh Verma
2025-10-15 19:31:36 +05:30
committed by GitHub
parent fc5551a724
commit b6a187aed1
15 changed files with 1180 additions and 0 deletions
+109
View File
@@ -0,0 +1,109 @@
.. meta::
:description: A typical workflow for collecting PMC data
:keywords: AQLprofile, ROCm, API, how-to, PMC
**********************************************************
Performance Monitor Control (PMC) workflow with AQLprofile
**********************************************************
This page describes a typical workflow for collecting PMC data using AQLprofile (as integrated in `ROCprofiler-SDK <https://github.com/ROCm/rocprofiler-sdk>`__).
This workflow relies on creating a profile object, generating command packets, and iterating over output buffers:
1. **Intercept kernel dispatch**: The SDK intercepts kernel dispatch packets submitted to the GPU queue.
2. **Create a profile object**: A profile/session object is created, specifying the agent (GPU), events (counters), and output buffers.
3. **Generate command packets**: Start, stop, and read command packets are generated and injected into the queue around the kernel dispatch.
4. **Submit packets and run the kernel**: The kernel and profiling packets are submitted to the GPU queue for execution.
5. **Collect the output buffer**: After execution, the output buffer is read back from the GPU.
6. **Iterate and extract the results**: The SDK iterates over the output buffer to extract and report counter results.
The SDK abstracts queue interception and packet management so tool developers can focus on results.
Key API code snippets
=====================
These API snippets use the legacy interfaces from ``hsa_ven_amd_aqlprofile.h``. These are provided for understanding purposes only.
For new development, refer to the updated APIs in ``aql_profile_v2.h``.
.. note::
The ROCprofiler-SDK is migrating to these newer interfaces in ``aql_profile_v2.h``. You should use the APIs in ``aql_profile_v2.h`` to stay up-to-date.
Define the events and profile
-----------------------------
.. code:: cpp
// Select events (counters) to collect
hsa_ven_amd_aqlprofile_event_t events[] = {
{ HSA_VEN_AMD_AQLPROFILE_BLOCK_NAME_SQ, 0, 2 }, // Example: SQ block, instance 0, counter 2
{ HSA_VEN_AMD_AQLPROFILE_BLOCK_NAME_SQ, 0, 3 }
};
// Create profile object
hsa_ven_amd_aqlprofile_profile_t profile = {
.agent = agent, // hsa_agent_t
.type = HSA_VEN_AMD_AQLPROFILE_EVENT_TYPE_PMC,
.events = events,
.event_count = sizeof(events)/sizeof(events[0]),
.parameters = nullptr,
.parameter_count = 0,
.output_buffer = {output_ptr, output_size},
.command_buffer = {cmd_ptr, cmd_size}
};
Validate events
---------------
.. code:: cpp
bool valid = false;
hsa_ven_amd_aqlprofile_validate_event(agent, &events[0], &valid);
if (!valid) {
// Handle invalid event
}
Generate command packets
-------------------------
.. code:: cpp
hsa_ext_amd_aql_pm4_packet_t start_pkt, stop_pkt, read_pkt;
hsa_ven_amd_aqlprofile_start(&profile, &start_pkt);
hsa_ven_amd_aqlprofile_stop(&profile, &stop_pkt);
hsa_ven_amd_aqlprofile_read(&profile, &read_pkt);
Submit packets and run the kernel
---------------------------------
.. code:: cpp
// Pseudocode: inject packets into HSA queue
queue->Submit(&start_pkt);
queue->Submit(&kernel_pkt);
queue->Submit(&stop_pkt);
queue->Submit(&read_pkt);
Iterate and extract results
----------------------------
.. code:: cpp
hsa_ven_amd_aqlprofile_iterate_data(
&profile,
[](hsa_ven_amd_aqlprofile_info_type_t info_type,
hsa_ven_amd_aqlprofile_info_data_t* info_data,
void* user_data) -> hsa_status_t {
if (info_type == HSA_VEN_AMD_AQLPROFILE_INFO_PMC_DATA) {
printf("Event: block %d, id %d, value: %llu\n",
info_data->pmc_data.event.block_name,
info_data->pmc_data.event.counter_id,
info_data->pmc_data.result);
}
return HSA_STATUS_SUCCESS;
},
nullptr
);
+93
View File
@@ -0,0 +1,93 @@
.. meta::
:description: A typical workflow for collecting detailed instruction-level traces
:keywords: AQLprofile, ROCm, API, how-to, SQTT
***********************************************
SQ Thread Trace (SQTT) workflow with AQLprofile
***********************************************
The SQ Thread Trace workflow focuses on collecting detailed instruction-level traces.
This workflow relies on creating a profile object, generating command packets, and iterating over output buffers:
1. **Intercept the kernel dispatch**: The SDK intercepts the kernel dispatch.
2. **Create a SQTT profile object**: A profile object is created for SQTT, specifying trace parameters and output buffers.
3. **Generate SQTT command packets**: Start, stop, and read packets for SQTT are generated and injected into the queue.
4. **Submit packets and run the kernel**: The kernel and SQTT packets are submitted for execution.
5. **Collect the trace buffer**: The trace output buffer is collected after execution.
6. **Iterate and decode trace data**: The SDK iterates over the trace buffer and decodes the SQTT data for analysis.
The SDK abstracts queue interception and packet management so tool developers can focus on results.
Key API code snippets
=====================
These API snippets use the legacy interfaces from ``hsa_ven_amd_aqlprofile.h``. These are provided for understanding purposes only.
For new development, refer to the updated APIs in ``aql_profile_v2.h``.
In the `ROCprofiler-SDK <https://github.com/ROCm/rocprofiler-sdk>`__ codebase, these APIs are wrapped and orchestrated in the ``aql``, ``hsa``, and ``thread_trace`` folders for queue interception, packet construction, and result iteration.
.. note::
The`ROCprofiler-SDK is migrating to these newer interfaces in ``aql_profile_v2.h``. You should use the APIs in ``aql_profile_v2.h`` to stay up-to-date.
Define parameters and profile
------------------------------
.. code:: cpp
hsa_ven_amd_aqlprofile_parameter_t params[] = {
{ HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_ATT_BUFFER_SIZE, 0x1000000} // 16 MB buffer
};
hsa_ven_amd_aqlprofile_profile_t profile = {
.agent = agent,
.type = HSA_VEN_AMD_AQLPROFILE_EVENT_TYPE_TRACE,
.events = nullptr,
.event_count = 0,
.parameters = params,
.parameter_count = sizeof(params)/sizeof(params[0]),
.output_buffer = {trace_ptr, trace_size},
.command_buffer = {cmd_ptr, cmd_size}
};
Generate SQTT start/stop packets
---------------------------------
.. code:: cpp
hsa_ext_amd_aql_pm4_packet_t sqtt_start_pkt, sqtt_stop_pkt;
hsa_ven_amd_aqlprofile_start(&profile, &sqtt_start_pkt);
hsa_ven_amd_aqlprofile_stop(&profile, &sqtt_stop_pkt);
Submit packets and run the kernel
---------------------------------
.. code:: cpp
queue->Submit(&sqtt_start_pkt);
queue->Submit(&kernel_pkt);
queue->Submit(&sqtt_stop_pkt);
Iterate and decode trace data
-----------------------------
.. code:: cpp
hsa_ven_amd_aqlprofile_iterate_data(
&profile,
[](hsa_ven_amd_aqlprofile_info_type_t info_type,
hsa_ven_amd_aqlprofile_info_data_t* info_data,
void* user_data) -> hsa_status_t {
if (info_type == HSA_VEN_AMD_AQLPROFILE_INFO_TRACE_DATA) {
// info_data->trace_data.ptr, info_data->trace_data.size
decode_trace(info_data->trace_data.ptr, info_data->trace_data.size);
}
return HSA_STATUS_SUCCESS;
},
nullptr
);