Files
rocm-systems/source/lib/rocprofiler-sdk/hsa/profile_serializer.hpp
T
Jonathan R. Madsen 2f9b1767e9 Handle hsa_queue_destroy after finalization (#679)
* Handle hsa_queue_destroy after finalization

- fixes issue where hsa_queue_destroy(...) is invoked after rocprofiler-sdk has finalized
- hsa::get_queue_controller() returns pointer
- if queue controller is a null pointer, skip invoking QueueController::destroy_queue

* Update HIP/HSA/marker update_table logging

* Update rocprofv3 tests

- remove HSA_TOOLS_LIB env variable
- remove setting ROCPROFILER_LOG_LEVEL env variable
- add timeouts to tests which are missing them

* Disable thread sanitizer deadlock detection

* Update CI workflow

- rename vega20-ubuntu job to core-ci
- enable navi32 in core-ci and sanitizers

* Update run-ci.py

- set gcovr html medium and high threshold

* Update lib/rocprofiler-sdk/hsa/queue_controller.cpp

- remove this capture from enable/disable serialization

* Update lib/rocprofiler-sdk/hsa/{hsa_barrier,profile_serializer}.*

- hsa_barrier::set_barrier accepts const-ref to queue map
- profile_serializer::enable and profile_serializer::disable accept const-ref to queue map

* Logging for HIP/HSA/marker/profile_serializer

* Logging for HIP/HSA/marker/queue_controller

* Improve test_retired_correlation_ids asserts

* Fix tests/counter-collection/validate.py

- scale expected SQ_WAVES counter value based on warp size of GPU

* Tweak github comment for code coverage

* Remove gcovr html high/medium threshold args

* Fix tests/counter-collection/validate.py

- round before casting to int in test_counter_values

* operator bool for profile_serializer

- only wait on CV if profile_serializer is used

* Logging updates (profile_serializer + code_object)

* Update counter-collection validate.py

* QueueController does not wait on CV if finalizing/finalized

* Update CI workflow

- remove navi32 from core job

* Improve HIP/HSA/marker tracing get_functor/functor

- remove lambda wrapper around functor

* Update lib/rocprofiler-sdk/hsa/queue_controller.cpp

- do not acquire cvmutex lock during finalization

* Update lib/rocprofiler-sdk/hsa/hsa_barrier.*

- move ctor and dtor to implementation
- skip signal store screlease and destroy if already finalized

* Update CI workflow

- remove navi32 runners

* bwelton fixes for hangs

* CMake improvements + simplified demangle

- remove amd-comgr from common target (and thus removed from roctx DT_NEEDED)

---------

Co-authored-by: Benjamin Welton <bewelton@amd.com>
2024-03-21 17:52:15 -05:00

99 baris
3.5 KiB
C++

// MIT License
//
// Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
#pragma once
#include <rocprofiler-sdk/rocprofiler.h>
#include "lib/common/container/small_vector.hpp"
#include "lib/rocprofiler-sdk/hsa/hsa_barrier.hpp"
#include "lib/rocprofiler-sdk/hsa/queue.hpp"
#include <cstdint>
#include <functional>
#include <optional>
#include <unordered_map>
#include <vector>
namespace rocprofiler
{
namespace hsa
{
/*This is a profiler serializer. It should be instantiated
only once for the profiler. The following is the
description of each field.
1. _dispatch_queue - The queue to which the currently dispatched kernel
belongs to.
At any given time, in serialization only one kernel
can be executing.
2. _dispatch_ready- It is a software data structure which holds
the queues which have a kernel ready to be dispatched.
This stores the queues in FIFO order.
3. serializer_mutex - The mutex is used for thread synchronization
while accessing the singleton instance of this structure.
Currently, in case of profiling kernels are serialized by default.
*/
class profiler_serializer
{
public:
enum class Status
{
ENABLED,
DISABLED,
};
struct barrier_with_state
{
barrier_with_state(Status _state, std::unique_ptr<hsa_barrier> _barrier)
: state(_state)
, barrier(std::move(_barrier))
{}
Status state;
std::unique_ptr<hsa_barrier> barrier;
};
using queue_map_t = std::unordered_map<hsa_queue_t*, std::unique_ptr<Queue>>;
void kernel_completion_signal(const Queue&);
// Signal a kernel dispatch is taking place, generates packets needed to be
// inserted to support kernel dispatch
common::container::small_vector<hsa::rocprofiler_packet, 3> kernel_dispatch(const Queue&) const;
void queue_ready(hsa_queue_t* hsa_queue, const Queue& queue);
// Enable the serializer
void enable(const queue_map_t& queues);
// Disable the serializer
void disable(const queue_map_t& queues);
void destroy_queue(hsa_queue_t* id, const Queue& queue);
static void add_queue(hsa_queue_t** hsa_queues, const Queue& queue);
private:
const Queue* _dispatch_queue{nullptr};
std::deque<const Queue*> _dispatch_ready;
std::atomic<Status> _serializer_status{Status::DISABLED};
std::deque<barrier_with_state> _barrier;
};
} // namespace hsa
} // namespace rocprofiler