AI/rocm-systems
2
0
Fork 0
Código Incidencias Pull Requests Acciones Paquetes Proyectos Lanzamientos Wiki Actividad

PC Sampling Support (#715)

Explorar el Código 
* cmake formatting (cmake-format) (#188)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* source formatting (clang-format v11) (#189)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: design of the pc sampling data struct; guarding parts of code that uses ROCr marker packets

* source formatting (clang-format v11) (#191)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* cmake formatting (cmake-format) (#192)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: shadow variable fix

* pcs: fix for compiler errors reported by CI/CD

* source formatting (clang-format v11) (#193)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: docs fix; samples uses rocprofiler::rocprofiler library

* cmake formatting (cmake-format) (#195)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: client in samples folder fixed

* pcs: client requires rocprofiler package as dependency

* pcs: client uses single context

* source formatting (clang-format v11) (#196)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: client using single buffer; no buffer destroy in client

* pcs: client::setup explicitly called from the example

* pcs: rocprofiler_pc_sample_record_t updated

* pcs: fixed init of external correlation id

* source formatting (clang-format v11) (#198)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: remove outdated files; update CMakeLists

* cmake formatting (cmake-format) (#212)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: using rocprofiler_agent_id_t

* pcs: Removing trailing whitespaces

Co-authored-by: Jonathan R. Madsen <jrmadsen@users.noreply.github.com>

* source formatting (clang-format v11) (#214)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: mapping agent_id to the agent

* source formatting (clang-format v11) (#215)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: const while iterating over agents

* source formatting (clang-format v11) (#216)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: calling get_buffer instead of get_buffers

* pcs: workgroup typo

* pcs: documentation for the public PC sampling API

* pcs: queue_cb_t signature adaptation

* pcs: mocks removed

* pcs: updating HsaApiTable with HSA/ROCr PC sampling API

* pcs: querying available PC sampling configs through IOCTL

* pcs: create the PCS session in IOCTL

* pcs: first actual PC samples delivered to the rocprofiler's client :)

* pcs: works with marker packet too

* pcs: using HSA table to call pc sampling related functions

* pcs: using ioctl instead of kfd in naming

* pcs: configuration service test fixed

* pcs: sample processing test fixed

* pcs: marker packet macro wrapper removed

* pcs: marker packet is part of the rocprofiler_packet union

* pcs: one fixme added

* pcs: client that uses pc-sampling and code obj tracing

* pcs: client that supprts PC sampling and code obj tracing refactored

* pcs: show more info for each PC sample

* pcs: hex output for the samples that do not belong to the matmul kernel

* pcs: querying avail configuration happens immediately before configuring

* pcs: hsa_ven_amd_pcs_create_from_id renamed

* pcs: using hsa_stop; accessing a buffer by id from parser

* pcs: includes reworked, tests returned to life

* pcs: rocrofiler dir removed as outdated

* cmake formatting (cmake-format) (#271)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* source formatting (clang-format v11) (#272)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: some warnings fixed

* source formatting (clang-format v11) (#273)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* cmake formatting (cmake-format) (#274)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: show MI200 relevant information in the sample

* pcs: queue cb fixed; rocr.h include fixed

* source formatting (clang-format v11) (#296)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: getting hsa_agent and the doorbell_id from hsa_queue

* source formatting (clang-format v11) (#297)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: correlation ID logic fixed

* source formatting (clang-format v11) (#303)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: pure pc sampling example fixed

* source formatting (clang-format v11) (#307)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* cmake formatting (cmake-format) (#308)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: interval value if the PC sampling is already configured

* pcs: ROCPROFILER_STATUS_ERROR_PC_SAMPLING_ALREADY_CONFIGURED

New status code if another process configured PC sampling service with different configuration.
Samples are extended to consider this case and retry if it happens.

* pcs: hsa_amd_queue_get_info mocked in tests

* source formatting (clang-format v11) (#328)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs (tests): query configs after configuring service

* source formatting (clang-format v11) (#329)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: sample checks workgroup_id_* and wave_id

* source formatting (clang-format v11) (#330)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs samples: running samples on the device 0

* pcs: kfd_ioctl updated

* pcs: ioctl config struct changed fields names

* pcs: status when PC sampling is configured by another process is renamed

* pcs: HSA PC sampling API table fixed

* pcs: tmp hack to be able to use HSA pc sampling table

* source formatting (clang-format v11) (#443)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs service use CIDs generated by HIP API tracing service

* source formatting (clang-format v11) (#455)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* cmake formatting (cmake-format) (#456)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: CID manager

* pcs: explicit flush with no delivered data executes retirement logic

* source formatting (clang-format v11) (#464)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: rocprofiler_query_pc_sampling_agent_configurations docs update

* source formatting (clang-format v11) (#465)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: rocprofiler_configure_pc_sampling_service docs update

* pcs: explicit sync introduced in PCSCIDManager

* pcs: new logic for retiring CIDs in PC sampling service documented

* pcs: queue interception cb signature updated

* source formatting (clang-format v11) (#471)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: if no agents supports PC sampling, fail gracefully

* elaborating when KFD returns EBUSY and EEXIST

* pcs: the second PC sampling examples fails gracefully

* code samples use only single kernel for now

* pcs: CID manager refactored

* source formatting (clang-format v11) (#481)

Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>

* pcs: ioctl update

* source formatting (clang-format v11) (#531)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs:code sample to test PC sampling applied on concurrent kernels

* source formatting (clang-format v11) (#533)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: pc sampling strest test included

* cmake formatting (cmake-format) (#539)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* source formatting (clang-format v11) (#540)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: standalone benchmark

* cmake formatting (cmake-format) (#555)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: glance in external correlation IDs

* source formatting (clang-format v11) (#557)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* another change in ioctl interface

* pcs: update queue interceptor callbacks and samples accroding to the agent 0 version

* source formatting (clang-format v11) (#611)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: avoid running problematic PC sampling test

* pcs: guarding tests not to fail on architectures not supporting PC sampling

* source formatting (clang-format v11) (#617)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: check IOCTL version prior to each KFD call

* pcs: ioctl refactoring

* pcs: PC sampling service increases the ref_count of the correlation ID of the kernel dispatch

* cmake formatting (cmake-format) (#631)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* source formatting (clang-format v11) (#632)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: PC sampling service provides external correlation IDs

* source formatting (clang-format v11) (#644)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: use rocprofiler_dim3_t for workgrou_ip

* source formatting (clang-format v11) (#645)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: minor fixes

* pcs: updating the documentation for the pc sampling API functions

* pcs: api table and queue controller fix

* pcs: don't generate marker packets for the agent if PC sampling is not configured on it

* pcs: multi-GPU and single-GPU clients

* source formatting (clang-format v11) (#700)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: warning and errors fixed

* source formatting (clang-format v11) (#702)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: clang compiler errors and warnings fixed

* source formatting (clang-format v11) (#716)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: const reference in cid manager

* source formatting (clang-format v11) (#717)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: const & func in manager explicit

* pcs: test to cover creating PC sampling service of agent that does not exist

* pcs: generate marker packets if service is active

* source formatting (clang-format v11) (#719)

Co-authored-by: vlaindic <139573562+vlaindic@users.noreply.github.com>

* pcs: refactoring hsa_adapter; use the correlation_id->thread_idx

* Update source/lib/rocprofiler-sdk/pc_sampling/cid_manager.cpp

* Update source/lib/rocprofiler-sdk/pc_sampling/cid_manager.cpp

* Update source/lib/rocprofiler-sdk/pc_sampling/hsa_adapter.cpp

* Update source/lib/rocprofiler-sdk/pc_sampling/hsa_adapter.cpp

* Update source/lib/rocprofiler-sdk/pc_sampling/hsa_adapter.cpp

* Update source/lib/rocprofiler-sdk/pc_sampling/hsa_adapter.cpp

* Update source/lib/rocprofiler-sdk/pc_sampling/utils.cpp

* Update utils.cpp

* moving pc-sampling tests and samples to pc-sampling label

* Format fix

* pcs: use configured instead of active service

* Update source/lib/rocprofiler-sdk/pc_sampling/service.cpp

* pcs: ensure configuring PC sampling on the HSA level is called only once

* pcs: minor fix

* Update CMakeLists.txt

* Update CMakeLists.txt

* Update CMakeLists.txt

* Update CMakeLists.txt

* pcs: refactoring IOCTL integration

* Update source/lib/rocprofiler-sdk/pc_sampling/tests/CMakeLists.txt

Co-authored-by: Ammar ELWazir <ammar.elwazir@amd.com>

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter.cpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter_types.hpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter.cpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter_types.hpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter.hpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* pcs: reverting back what bot doubled

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter_types.hpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* pcs: retesting the bot

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter_types.hpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* pcs: why bot fails on this IOCTL status

* pcs: why failing on <vector>

* Update source/lib/rocprofiler-sdk/pc_sampling/ioctl/ioctl_adapter.cpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* pcs: returning commits removed by bot

* pcs: formatting locally

* pcs: clients are flushing buffers inside the tool_fini

* pcs: sync function in public API

* pcs: sync prior to unloading the code object

* pcs: sync function requires context

* pcs: client uses CID retirement service

* pcs: test for flusing internal ROCr buffers

* pcs: source formatting

* Update source/lib/rocprofiler-sdk/pc_sampling/tests/CMakeLists.txt

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* pcs: code samples refactoring

* pcs: public API header refactored

* pcs: rocprofiler_buffer_flush drains internal PC sampling buffers too

* pcs: remove unnecessary functions

* pcs: do not call hsa's copytables

* pcs: include reordering

* pcs: using ROCP_ERROR inside PC sampling implementation

* pcs: pc_sampling sample uses ostream instean of printfs

* pcs: pc_sampling_codeobj tracing using ostream instead of prints

* pcs: registering once for interceptor callbacks

* pcs: do not generate internal CIDs if not in debug mode

* pcs: rebasing fixed; missing external correlation IDs

* pcs: code formatting

* enable kernel tracing service to receive external correlation IDs

* pcs: using ROCPROFILER_STATUS_ERROR_INCOMPATIBLE_KERNEL

* pcs: polishing parser

* formatting

* updating parser to use workgroup_id

* kfd_ioctl.h extracted in details folder

* refactoring

* pcs: preparing to generate code object information

* flush internal buffers prior to unloading code object

* pcs: generating marker records

* pcs: wrap code_object's shutdown function

* ROCR_VISIBLE_DEVICES and HIP_VISISBLE_DEVICES unsupported at the moment

* documenting the ignorance of ROCR/HIP_VISIBLE_DEVICES

* pcs: separate structs for code object loading/unloading markers

* pcs: inst_pkt_t changed the namespace

* pcs: removing wrapper around the shutdown function

* pcs: size in record field

* pcs: documentation refactoring + typdefs

* renaming PCSAgentConfig to PCSAgentSession

* pcs: service does not keep a pointer to the context

* pcs: static assertions related to the versioning

* pcs: rocprofiler_pc_sampling_configuration_t size field

* pcs: report API unimplemented unleass explicitly enabled

* pcs: skip tests if KFD does not support PC sampling

* pcs: if ROCr hides some devices, no PC samples will be delivered for it

* pcs: hip error check after kernel launch

* formatting

* removing PCS info from agent.h

* fix based on review

* Update continuous integration workflow

- use mi200 runner for code coverage (supports PC sampling)
- split sanitizer jobs across navi3, vega20, and mi300

* Updating pc sampling test labels

* ROCP_PC_SAMPLING_ENABLED env in CI

* ROCP_PC_SAMPLING_ENABLED for all CI mi200 jobs

* Rearrange sanitizer assignments

* fixes according to review

* removed unused functions

* pcs: rocprofiler_agent_id_t instead of handle as a key in map

* Update source/lib/rocprofiler-sdk/context/context.hpp

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

* removing drm_fd from the agent.h

* pcs: removing one sample due to complexity

* pcs: refactoring sample

* simplifying sample

* new lines

* Improve queue_control enable intercepter logic

* Update lib/rocprofiler-sdk/hsa/types.hpp

- handle amd_ext size for HSA 1.12.0

* ROCP_PC_SAMPLING_ENABLED -> ROCPROFILER_PC_SAMPLING_BETA_ENABLED

* Update hsa_adapter.cpp

- anonymous namespace + remove debug

* parser update

* Apply suggestions from code review

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: vlaindic <vlaindic@users.noreply.github.com>
Co-authored-by: vlaindic <vladimir.indic@amd.com>
Co-authored-by: vlaindic <vlaindic@amd.com>
Co-authored-by: Vladimir Indic <139573562+vlaindic@users.noreply.github.com>
Co-authored-by: Jonathan R. Madsen <jrmadsen@users.noreply.github.com>
Co-authored-by: gobhardw <gopesh.bhardwaj@amd.com>
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
Este commit está contenido en:
Ammar ELWazir
2024-05-24 09:49:44 -05:00
cometido por GitHub
padre f14f5a994e
commit 987ae3cc47
Se han modificado 64 ficheros con 6831 adiciones y 2012 borrados
Descargar archivo de parche Descargar archivo de diferencias Expandir todos los archivos Contraer todos los archivos
source/include/rocprofiler-sdk/agent.h
+4 -11
Ver fichero
@@ -187,17 +187,10 @@ typedef struct rocprofiler_agent_v0_t
const char* vendor_name; ///< Vendor of agent (will be AMD)
const char* product_name; ///< Marketing name
const char* model_name; ///< GPU only. Will be something like vega20, mi200, etc.
uint64_t num_pc_sampling_configs; ///< GPU only. Number of PC sampling modes available for this
///< device type. Note: if another process is currently using
///< PC sampling on this agent, this value will be zero so
///< do not assume the number of PC sampling configurations
///< based on the device type.
const rocprofiler_pc_sampling_configuration_t*
pc_sampling_configs; ///< GPU only. Array of PC sampling configuration types.
uint32_t node_id; ///< Node sequence number. This will be equivalent to the HSA-runtime
///< HSA_AMD_AGENT_INFO_DRIVER_NODE_ID property
int32_t logical_node_id; ///< Logical sequence number. This will always be [0..N) where N is
///< the total number of agents
uint32_t node_id; ///< Node sequence number. This will be equivalent to the HSA-runtime
///< HSA_AMD_AGENT_INFO_DRIVER_NODE_ID property
int32_t logical_node_id; ///< Logical sequence number. This will always be [0..N) where N is
///< the total number of agents
} rocprofiler_agent_v0_t;
typedef rocprofiler_agent_v0_t rocprofiler_agent_t;
source/include/rocprofiler-sdk/cxx/serialization.hpp
-1
Ver fichero
@@ -720,7 +720,6 @@ save(ArchiveT& ar, const rocprofiler_agent_v0_t& data)
ROCP_SDK_SAVE_DATA_CSTR(vendor_name);
ROCP_SDK_SAVE_DATA_CSTR(product_name);
ROCP_SDK_SAVE_DATA_CSTR(model_name);
ROCP_SDK_SAVE_DATA_FIELD(num_pc_sampling_configs);
ROCP_SDK_SAVE_DATA_FIELD(node_id);
ROCP_SDK_SAVE_DATA_FIELD(logical_node_id);
source/include/rocprofiler-sdk/fwd.h
+16 -4
Ver fichero
@@ -102,6 +102,9 @@ typedef enum // NOLINT(performance-enum-size)
ROCPROFILER_STATUS_ERROR_NO_PROFILE_QUEUE, ///< Profile queue creation failed
ROCPROFILER_STATUS_ERROR_NO_HARDWARE_COUNTERS, ///< No hardware counters were specified
ROCPROFILER_STATUS_ERROR_AGENT_MISMATCH, ///< Agent mismatch between profile and context.
ROCPROFILER_STATUS_ERROR_NOT_AVAILABLE, ///< The service is not available.
///< Please refer to API functions that return this
///< status code for more information.
ROCPROFILER_STATUS_LAST,
} rocprofiler_status_t;
@@ -400,6 +403,19 @@ typedef enum
ROCPROFILER_COUNTER_FLAG_LAST,
} rocprofiler_counter_flag_t;
/**
* @brief Enumeration for distinguishing different buffer record kinds within the
* ::ROCPROFILER_BUFFER_CATEGORY_PC_SAMPLING category
*/
typedef enum
{
ROCPROFILER_PC_SAMPLING_RECORD_NONE = 0,
ROCPROFILER_PC_SAMPLING_RECORD_SAMPLE, ///< ::rocprofiler_pc_sampling_record_t
ROCPROFILER_PC_SAMPLING_RECORD_CODE_OBJECT_LOAD_MARKER, ///< ::rocprofiler_pc_sampling_code_object_load_marker_t
ROCPROFILER_PC_SAMPLING_RECORD_CODE_OBJECT_UNLOAD_MARKER, ///< ::rocprofiler_pc_sampling_code_object_unload_marker_t
ROCPROFILER_PC_SAMPLING_RECORD_LAST,
} rocprofiler_pc_sampling_record_kind_t;
//--------------------------------------------------------------------------------------//
//
// ALIASES
@@ -442,10 +458,6 @@ typedef uint64_t rocprofiler_kernel_id_t;
// */
typedef uint64_t rocprofiler_dispatch_id_t;
// forward declaration of struct
typedef struct rocprofiler_pc_sampling_configuration_s rocprofiler_pc_sampling_configuration_t;
typedef struct rocprofiler_pc_sampling_record_s rocprofiler_pc_sampling_record_t;
/**
* @brief Unique record id encoding both the counter
* and dimensional values (positions) for the record.
source/include/rocprofiler-sdk/pc_sampling.h
+164 -135
Ver fichero
@@ -38,53 +38,57 @@ ROCPROFILER_EXTERN_C_INIT
* @brief Function used to configure the PC sampling service on the GPU agent with @p agent_id.
*
* Prerequisites are the following:
* - The user must create a context and supply its @p context_id. By using this context,
* - The user must create a context and supply its @p context_id. By using this context,
* the user can start/stop PC sampling on the agent. For more information,
* please @see `rocprofiler_start_context`/`rocprofiler_stop_context`.
* - The user must create a buffer and supply its @p buffer_id. Rocprofiler uses the buffer
* - The user must create a buffer and supply its @p buffer_id. Rocprofiler uses the buffer
* to deliver the PC samples to the user. For more information about the data delivery,
* please @see `rocprofiler_create_buffer` and `rocprofiler_buffer_tracing_cb_t`.
* - The client must create a context and supply its @p context_id. By using this context,
* the client can start/stop PC sampling on the agent. For more information,
* please @see rocprofiler_start_context/rocprofiler_stop_context.
* - The user must create a buffer and supply its @p buffer_id. Rocprofiler-SDK uses the buffer
* to deliver the PC samples to the client. For more information about the data delivery,
* please @see rocprofiler_create_buffer and @see rocprofiler_buffer_tracing_cb_t.
*
* Before calling this function, we recommend querying PC sampling configurations
* supported by the GPU agent via the `rocprofiler_query_pc_sampling_agent_configurations`.
* The user then chooses the @p method, @p unit, and @p interval to match one of the
* available configurations. Note that the @p interval must belong to the range of values
* The user then chooses the @p method, @p unit, and @p interval to match one of the
* supported by the GPU agent via the @see rocprofiler_query_pc_sampling_agent_configurations.
* The client chooses the @p method, @p unit, and @p interval to match one of the
* available configurations. Note that the @p interval must belong to the range of values
* [available_config.min_interval, available_config.max_interval],
* where available_config is the instance of the `rocprofiler_pc_sampling_configuration_s`
* supported at the moment.
* where available_config is the instance of the @see rocprofiler_pc_sampling_configuration_s
* supported/available at the moment.
*
* Rocprofiler checks whether the requsted configuration is actually supported
* Rocprofiler-SDK checks whether the requsted configuration is actually supported
* at the moment of calling this function. If the answer is yes, it returns
* the ROCPROFILER_STATUS_SUCCESS. Otherwise, notifies the caller about the
* the @see ROCPROFILER_STATUS_SUCCESS. Otherwise, it notifies the client about the
* rejection reason via the returned status code. For more information
* about the status codes, please @see rocprofiler_status_t.
*
* There are a few constraints a client's code needs to be aware of.
*
* Constraint1: A GPU agent can be configured to support at most one running PC sampling
* configuration at any time, which implies some of the consequences described below.
* After the tool configures the PC sampling with one of the available configurations,
* rocprofiler guarantees that this configuration will be valid for the tool's
* rocprofiler-SDK guarantees that this configuration will be valid for the tool's
* lifetime. The tool can start and stop the configured PC sampling service whenever convenient.
*
* Constraint2: Since the same GPU agent can be used by multiple processes concurrently,
* Rocprofiler cannot guarantee the exclusive access to the PC sampling capability.
* Rocprofiler-SDK cannot guarantee the exclusive access to the PC sampling capability.
* The consequence is the following scenario. The tool TA that belongs to the process PA,
* calls the `rocprofiler_query_pc_sampling_agent_configurations` that returns the
* two supported configurations CA and CB by the agent. Then the toolb TB of the process PB,
* calls the @see rocprofiler_query_pc_sampling_agent_configurations that returns the
* two supported configurations CA and CB by the agent. Then the tool TB of the process PB,
* configures the PC sampling on the same agent by using the configuration CB.
* Subsequently, the TA tries configuring the CA on the agent, and it fails.
* To point out that this case happened, we introduce a special status code (TODO: ARE WE)?
* When this status code is observed by the tool TA, it queties all available configurations again
* by calling `rocprofiler_query_pc_sampling_agent_configurations`,
* To point out that this case happened, we introduce a special status code
* @see ROCPROFILER_STATUS_ERROR_NOT_AVAILABLE.
* When this status code is observed by the tool TA, it queries all available configurations again
* by calling @see rocprofiler_query_pc_sampling_agent_configurations,
* that returns only CB this time. The tool TA can choose CB, so that both
* TA and TB use the PC sampling capability in the separate processes.
* Both TA and TB receives samples generated by the kernels launched by the
* corresponding processes PA and PB, respectively.
*
* Constraints3: We allow only one context to contain the configured PC sampling service
* within the process, that implies that at most one of the loaded tools can use PC sampling.
* One context can contains multiple PC sampling services configured for different GPU agents.
* Constraint3: Rocprofiler-SDK allows only one context to contain the configured PC sampling
* service within the process, that implies that at most one of the loaded tools can use PC
* sampling. One context can contains multiple PC sampling services configured for different GPU
* agents.
*
* Constraint4: PC sampling feature is not available within the ROCgdb.
*
* @param [in] context_id - id of the context used for starting/stopping PC sampling service
* @param [in] agent_id - id of the agent on which caller tries using PC sampling capability
@@ -93,6 +97,14 @@ ROCPROFILER_EXTERN_C_INIT
* @param [in] interval - frequency at which PC samples are generated
* @param [in] buffer_id - id of the buffer used for delivering PC samples
* @return ::rocprofiler_status_t
* @retval ::ROCPROFILER_STATUS_SUCCESS PC sampling service configured successfully
* @retval ::ROCPROFILER_STATUS_ERROR_NOT_AVAILABLE One of the scenarios is present:
* 1. PC sampling is already configured with configuration different than requested,
* 2. PC sampling is requested from a process that runs within the ROCgdb.
* 3. HSA runtime does not support PC sampling.
* @retval ::ROCPROFILER_STATUS_ERROR_INCOMPATIBLE_KERNEL the amdgpu driver installed on the system
* does not support the PC sampling feature
* @retval ::ROCPROFILER_STATUS_ERROR a general error caused by the amdgpu driver
*
*/
rocprofiler_status_t ROCPROFILER_API
@@ -105,45 +117,45 @@ rocprofiler_configure_pc_sampling_service(rocprofiler_context_id_t conte
/**
* @brief PC sampling configuration supported by a GPU agent.
* @var rocprofiler_pc_sampling_configuration_s::method
* Sampling method supported by the GPU
* agent. Currenlty, it can take one of the following two values:
* - ROCPROFILER_PC_SAMPLING_METHOD_HOST_TRAP: a background host thread
* periodically interrupts waves execution on the GPU to generate PC samples
* - ROCPROFILER_PC_SAMPLING_METHOD_STOCHASTIC: performance monitoring hardware
* on the GPU periodically interrupts waves to generate PC samples.
* @var rocprofiler_pc_sampling_configuration_s::unit
* A unit used to specify the period of the
* @ref method for samples generation.
* @var rocprofiler_pc_sampling_configuration_s::min_interval
* the highest possible frequencey for
* generating samples using @ref method.
* @var rocprofiler_pc_sampling_configuration_s::max_interval
* the lowest possible frequency for
* generating samples using @ref method
* @var rocprofiler_pc_sampling_configuration_s::flags
* TODO: ???
*/
struct rocprofiler_pc_sampling_configuration_s
typedef struct
{
uint64_t size; ///< Size of this struct
rocprofiler_pc_sampling_method_t method;
rocprofiler_pc_sampling_unit_t unit;
size_t min_interval;
size_t max_interval;
uint64_t flags;
};
uint64_t flags; /// for future use
/// @var method
/// @brief Sampling method supported by the GPU agent.
/// Currently, it can take one of the following two values:
/// - ::ROCPROFILER_PC_SAMPLING_METHOD_HOST_TRAP: a background host kernel thread
/// periodically interrupts waves execution on the GPU to generate PC samples
/// - ::ROCPROFILER_PC_SAMPLING_METHOD_STOCHASTIC: performance monitoring hardware
/// on the GPU periodically interrupts waves to generate PC samples.
/// @var unit
/// @brief A unit used to specify the interval of the @ref method for samples generation.
/// @var min_interval
/// @brief the highest possible frequencey for generating samples using @ref method.
/// @var max_interva
/// @brief the lowest possible frequency for generating samples using @ref method
} rocprofiler_pc_sampling_configuration_t;
/**
* @brief The rocprofiler calls the tool's callback to deliver the list
* of available configurations upon the calls to the @ref
* rocprofiler_query_pc_sampling_agent_configurations.
* @brief Rocprofiler SDK's callback function to deliver the list of available PC
* sampling configurations upon the call to the
* @ref rocprofiler_query_pc_sampling_agent_configurations.
*
* @param[out] configs - The list of PC sampling configurations supported by the agent of the
* moment of invoking @ref rocprofiler_query_pc_sampling_agent_configurations.
* @param[out] num_config - The number of configuration contained in the underlying
* @param[out] configs - The array of PC sampling configurations supported by the agent
* at the moment of invoking @ref rocprofiler_query_pc_sampling_agent_configurations.
* @param[out] num_config - The number of configurations contained in the underlying array
* @p configs.
* In case the GPU agent does not support PC sampling, the value is 0.
* @param[in] user_data - A pointer passed as the last argument of the
* @param[in] user_data - client's private data passed via
* @ref rocprofiler_query_pc_sampling_agent_configurations
* @return ::rocprofiler_status_t
*/
typedef rocprofiler_status_t (*rocprofiler_available_pc_sampling_configurations_cb_t)(
const rocprofiler_pc_sampling_configuration_t* configs,
@@ -153,10 +165,24 @@ typedef rocprofiler_status_t (*rocprofiler_available_pc_sampling_configurations_
/**
* @brief Query PC Sampling Configuration.
*
* @param [in] agent_id - id of the agent for which available configuration will be listed
* Lists PC sampling configurations a GPU agent with @p agent_id supports at the moment
* of invoking the function. Delivers configurations via @p cb.
* In case the PC sampling is configured on the GPU agent, the @p cb delivers information
* about the active PC sampling configuration.
* In case the GPU agent does not support PC sampling capability,
* the @p cb delivers none PC sampling configurations.
*
* @param [in] agent_id - id of the agent for which available configurations will be listed
* @param [in] cb - User callback that delivers the available PC sampling configurations
* @param [in] user_data - passed to the @p cb
* @return ::rocprofiler_status_t
* @retval ::ROCPROFILER_STATUS_ERROR_NOT_AVAILABLE One of the scenarios is present:
* 1. PC sampling is requested from a process that runs within the ROCgdb.
* 2. HSA runtime does not support PC sampling.
* @retval ::ROCPROFILER_STATUS_ERROR_INCOMPATIBLE_KERNEL the amdgpu driver installed on the system
* does not support the PC sampling feature.
* @retval ::ROCPROFILER_STATUS_ERROR a general error caused by the amdgpu driver
* @retval ::ROCPROFILER_STATUS_SUCCESS @p cb successfully finished
*/
rocprofiler_status_t ROCPROFILER_API
rocprofiler_query_pc_sampling_agent_configurations(
@@ -165,36 +191,31 @@ rocprofiler_query_pc_sampling_agent_configurations(
void* user_data) ROCPROFILER_NONNULL(2, 3);
/**
* @brief The header of the @ref rocprofiler_pc_sampling_record_s, indicating
* what fields of the @ref rocprofiler_pc_sampling_record_s instance are meaningful
* @brief The header of the @ref rocprofiler_pc_sampling_record_s, indicating
* what fields of the @ref rocprofiler_pc_sampling_record_s instance are meaningful
* @brief The header of the @ref rocprofiler_pc_sampling_record_t, indicating
* what fields of the @ref rocprofiler_pc_sampling_record_t instance are meaningful
* for the sample.
* @var rocprofiler_pc_sampling_header_v1_t::valid
* the sample is valid
* @var rocprofiler_pc_sampling_header_v1_t::type
* The following values are possible:
* - 0 - reserved
* - 1 - host trap pc sample
* - 2 - stochastic pc sample
* - 3 - perfcounter (unsupported at the moment)
* - other values does not mean anything at the moment
* @var rocprofiler_pc_sampling_header_v1_t::has_stall_reason
* whether the sample contains
* information about the stall reason. If so, please @see rocprofiler_pc_sampling_snapshot_v1_t.
* @var rocprofiler_pc_sampling_header_v1_t::has_wave_cnt
* whether the @ref rocprofiler_pc_sampling_record_s::wave_count contains
* meaningful value
* @var rocprofiler_pc_sampling_header_v1_t::reserved
* for future use
*/
typedef struct
{
uint8_t valid : 1;
uint8_t type : 4; // 0=reserved, 1=hosttrap, 2=stochastic
uint8_t valid : 1; /// sample is valid
uint8_t type : 4;
uint8_t has_stall_reason : 1;
uint8_t has_wave_cnt : 1;
uint8_t reserved : 1;
uint8_t reserved : 1; /// for future use
/// @var type
/// @brief The following values are possible:
/// - 0 - reserved
/// - 1 - host trap pc sample
/// - 2 - stochastic pc sample
/// - 3 - perfcounter (unsupported at the moment)
/// - other values does not mean anything at the moment
/// @var has_stall_reason
/// @brief whether the sample contains information about the stall reason.
/// If so, please @see rocprofiler_pc_sampling_snapshot_v1_t.
/// @var has_wave_cnt
/// @brief whether the @ref rocprofiler_pc_sampling_record_t::wave_count
/// contains meaningful value
} rocprofiler_pc_sampling_header_v1_t;
/**
@@ -213,65 +234,73 @@ typedef struct
// to reduce the space needed to represent a single sample.
/**
* @brief ROCProfiler PC Sampling Record corresponding to the interrupted wave.
* @var rocprofiler_pc_sampling_record_s::flags
* header that indicates what fields are meaningful
* for the PC sample. The values depend on what the underlying GPU agent architecture supports.
* @var rocprofiler_pc_sampling_record_s::chiplet
* chiplet index
* @var rocprofiler_pc_sampling_record_s::wave_id
* wave identifier within the workgroup
* @var rocprofiler_pc_sampling_record_s::wave_issued
* a flags indicated whether the wave is
* issueing the instruction' represented by the @ref pc at the moment of interruption.
* @var rocprofiler_pc_sampling_record_s::reserved
* FIXME: reserved 7 bits, must be zero.
* @var rocprofiler_pc_sampling_record_s::hw_id
* compute unit identifier
* @var rocprofiler_pc_sampling_record_s::pc
* The current program counter of the wave at the moment
* of interruption
* @var rocprofiler_pc_sampling_record_s::exec_mask
* shows how many SIMD lanes of the wave were
* executing the instruction represented by the @ref pc. Useful to understand thread-divergance
* within the wave
* @var rocprofiler_pc_sampling_record_s::workgroup_id_x
* the x coordinate of the wave within the workgroup
* @var rocprofiler_pc_sampling_record_s::workgroup_id_y
* the y coordinate of the wave within the workgroup
* @var rocprofiler_pc_sampling_record_s::workgroup_id_z
* the y coordinate of the wave within the workgroup
* @var rocprofiler_pc_sampling_record_s::wave_count
* FIXME: number of waves active at the CU at the moment of sample generation???
* @var rocprofiler_pc_sampling_record_s::timestamp
* represents the GPU timestamp when the sample is generated
* @var rocprofiler_pc_sampling_record_s::correlation_id
* correlation id of the API call that
* initiated kernel laucnh. The interrupted wave is executed as part of the kernel.
* @var rocprofiler_pc_sampling_record_s::snapshot
* TODO:
* @var rocprofiler_pc_sampling_record_s::reserved2
* for future use
*/
struct rocprofiler_pc_sampling_record_s
typedef struct
{
rocprofiler_pc_sampling_header_v1_t flags;
uint8_t chiplet;
uint8_t wave_id;
uint8_t wave_issued : 1;
uint8_t reserved : 7;
uint32_t hw_id;
uint64_t pc;
uint64_t exec_mask;
uint32_t workgroup_id_x;
uint32_t workgroup_id_y;
uint32_t workgroup_id_z;
uint32_t wave_count;
uint64_t timestamp;
rocprofiler_correlation_id_t correlation_id;
rocprofiler_pc_sampling_snapshot_v1_t snapshot;
uint32_t reserved2;
};
uint64_t size; ///< Size of this struct
rocprofiler_pc_sampling_header_v1_t flags;
uint8_t chiplet; /// chiplet index
uint8_t wave_id; /// wave identifier within the workgroup
uint8_t wave_issued : 1;
uint8_t reserved : 7; /// reserved 7 bits, must be zero
uint32_t hw_id; /// compute unit identifier
uint64_t pc; /// Program counter of the wave of the moment of interruption
uint64_t exec_mask;
rocprofiler_dim3_t workgroup_id; /// wave coordinates within the workgroup
uint32_t wave_count;
uint64_t timestamp; /// timestamp when sample is generated
rocprofiler_correlation_id_t correlation_id;
rocprofiler_pc_sampling_snapshot_v1_t
snapshot; /// @see ::rocprofiler_pc_sampling_snapshot_v1_t
uint32_t reserved2; /// for future use
/// @var flags
/// @brief indicates what fields of this struct are meaningful for the represented sample.
/// The values depend on what the underlying GPU agent architecture supports.
/// @var wave_issue
/// @brief indicates whether the wave is issueing the instruction represented by the @ref pc
/// @var exec_mask
/// @brief shows how many SIMD lanes of the wave were executing the instruction
/// represented by the @ref pc. Useful to understand thread-divergance within the wave
/// @var wave_count
/// @brief number of active waves on the CU at the moment of sample generation
/// @var correlation_id
/// @brief correlation id of the API call that initiated kernel launch.
/// The interrupted wave is executed as part of the kernel.
} rocprofiler_pc_sampling_record_t;
/**
* @brief Marker representing code object loading event.
*
* @see rocprofiler_callback_tracing_code_object_load_data_t
* for more information
*/
typedef struct
{
uint64_t size; ///< Size of this struct
uint64_t code_object_id; /// unique code object identifier
} rocprofiler_pc_sampling_code_object_load_marker_t;
/**
* @brief Marker representing code object unloading event.
*
* @see rocprofiler_callback_tracing_code_object_load_data_t
* for more information
*/
typedef struct
{
uint64_t size; ///< Size of this struct
uint64_t code_object_id; /// unique code object identifier
} rocprofiler_pc_sampling_code_object_unload_marker_t;
/** @} */
ROCPROFILER_EXTERN_C_FINI
ROCPROFILER_CXX_CODE(
static_assert(sizeof(rocprofiler_pc_sampling_record_t) == 80,
"Increasing the size of the pc sampling record is not permitted."));
ROCPROFILER_CXX_CODE(static_assert(offsetof(rocprofiler_pc_sampling_record_t, chiplet) == 9 &&
offsetof(rocprofiler_pc_sampling_record_t, reserved2) == 76,
"PC sampling record layout changed."));