* Srnagara/tool counters collect (#331)

* Adding counter collection capability to tools

* Adding counter collection feature to tools

* Adding counter collection capability to tools

* Fixing merge down issues

* Small tool fixes for build + prevent profile realloc

* Reproducing the counter name query issue in buffered callback

* Minor fix for init order + sample that directly uses sdk-tool for debug purposes

* Adding a temporary fix to print the counter names

* Fixing the output file name and reverting the changes of caching the profile config

* Fixing SGPR_Count value

* cleaning up debug prints

* Adding header to counter collection file

* Adding kernel filtering support

* Remove threading

* Cleaning up the code

* Removing redundant prints

* Revert "Remove threading"

This reverts commit 05c58fb9de826e92cf8d2e3d1c31d5578525dcb4.

* Revert "Cleaning up the code"

This reverts commit 1d964882bf2396dee8ad020cbb6c83b36e0674e9.

* Changing the tools code to align with init-order fix

* cmake formatting (cmake-format) (#335)

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

* source formatting (clang-format v11) (#336)

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

* Adding support for async memory copy

* source formatting (clang-format v11) (#391)

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

* Fixing header typo

* Fixing tool_fini

* Replaceing the direction and kind fields values with description

* Update lib/rocprofiler-sdk-tool/helper.cpp

- Remove use of VLA

* Update lib/rocprofiler-sdk-tool/tool.cpp

- Formatting

* Migrate common/config.* to rocprofiler-sdk-tool

* Update lib/rocprofiler-sdk-tool/tool.cpp

- fix clang-tidy issues

* source formatting (clang-format v11) (#392)

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

* Update lib/common/mpl.hpp

- is_string_type / is_string_type_impl for deducing if type is a string type

* Update include/rocprofiler-sdk/fwd.h

- ROCPROFILER_BUFFER_TRACING_MEMORY_COPY_NONE starts at zero

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

- functions for operation ids and names

* Update lib/rocprofiler-sdk/buffer_tracing.cpp

- support iterating and getting names for ROCPROFILER_BUFFER_TRACING_MEMORY_COPY

* Update lib/rocprofiler-sdk-tool/config.*

- env ROCPROFILER_ prefix -> ROCPROF_ prefix
- add support for memory copy tracing, counter collection, etc.

* Update lib/rocprofiler-sdk-tool/helper.*

- removed TracerFlushRecord
- removed cxa_demangle (use one in common library)
- removed GetCounterNames (handled in config)
- removed GetKernelNames (handled in config)

* Add lib/rocprofiler-sdk-tool/output_file.*

- separate out get_output_stream function and output_file struct from tool.cpp

* Add lib/rocprofiler-sdk-tool/csv.hpp

- write_csv_entry automatically quotes strings
- csv_encoder struct enforces correct number of columns

* Update lib/rocprofiler-sdk-tool/CMakeLists.txt

- add new files

* Update lib/rocprofiler-sdk-tool/tool.cpp

- update construction of output_file class
- add kernel_symbol_data for serializing kernel trace data
- use config instead of env lookups
- optimize counter collection profile config lookup/creation

* Update bin/rocprofv3

- rocprofv3 --help exits with 0 (as it should)
- command-line arg for memory copy tracing
- command-line arg for mangled kernels
- command-line arg for truncated kernels
- env ROCPROFILER_ prefix -> env ROCPROF_ prefix

* Update tests/async-copy-tracing/validate.py

- update test_async_copy_direction to new enum values

* Update tests/kernel-tracing/validate.py

- update test_async_copy_direction to new enum values

* Update tests/tools/json-tool.cpp

- add ROCPROFILER_BUFFER_TRACING_MEMORY_COPY to supported buffer_name_info

* Update samples/counter_collection/{CMakeLists.txt,main.cpp}

- remove counter-collection-sdk-tool

* Update .github/workflows/docs.yml

- fix paths triggering running the workflow

---------

Co-authored-by: Benjamin Welton <bewelton@amd.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: SrirakshaNag <SrirakshaNag@users.noreply.github.com>
Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
Co-authored-by: jrmadsen <jrmadsen@users.noreply.github.com>

* adding counter collection support

* Adding counter collection test

* changing directory structure of counter collection tests

* Fixing test path for rocprofv3

* Adding hsa-tracing basic test

* cmake formatting (cmake-format) (#362)

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

* counter collection tests drop2

* fixing hsa-trace test for rocprofv3 path

* python formatting (black) (#371)

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

* both counter colleciton and tracing should work together

* Fixing rocprofv3 path

* Attempt to fix Segfault with AddressSanitizer

* fixing sanitizer segfault

* Update rocprofv3

* Update lib/rocprofiler-sdk-tool/README.md

- update env variables

* Update lib/rocprofiler-sdk/buffer_tracing.cpp

- return ROCPROFILER_STATUS_BUFFER_NOT_FOUND if buffer tracing service is configured with invalid buffer

* Update lib/rocprofiler-sdk-tool/tool.cpp

- designated hsa API trace buffer

* Update tests/hsa-tracing/CMakeLists.txt

- Fix environment

* Update rocprofv3

- do not override HSA_TOOLS_LIB
- support ROCPROF_PRELOAD
- LD_PRELOAD librocprofiler-sdk.so

* Restructure tests directory

- move all rocprofv3 integration tests into subfolder

* Update cmake/Templates/rocprofiler-sdk/config.cmake.in

- create rocprofiler-sdk::rocprofv3 cmake target

* Update tests/rocprofv3/hsa-tracing

- improve validate.py
- convert input to dict via csv.DictReader

* Update tests/apps/CMakeLists.txt

- fix build rpath for simple-transpose

* Update  cmake/rocprofiler_memcheck.cmake

- prefer libtsan.so.0

* Update tests/rocprofv3/hsa-tracing

- move to tests/rocprofv3/tracing
- include kernel tracing and memory copy tracing

* Update lib/rocprofiler-sdk-tool/tool.cpp

- normalize "_ID" vs. "_Id" in CSV column names (use "_Id")

* Update lib/rocprofiler-sdk/buffer.{hpp,cpp}

- change signature of buffer::get_buffers()
- buffer::get_buffers() uses static_object

* Update lib/rocprofiler-sdk/context/context.cpp

- update usage of buffer::get_buffers()
  - now returns pointer

* Update lib/rocprofiler-sdk/tests/buffer.cpp

- update to change for signature of buffer::get_buffers()

* Update tests/rocprofv3/tracing/CMakeLists.txt

- use %argt% with -d argument

* Update lib/rocprofiler-sdk-tool/tool.cpp

- use atexit for finalization

* Update tests/rocprofv3/tracing/CMakeLists.txt

- tweaked name of tests

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

- async_copy_fini + reference counting signals

* Update lib/rocprofiler-sdk/registration.cpp

- invoke hsa::async_copy_fini() to prevent data race on signals

---------

Co-authored-by: SrirakshaNag <104580803+SrirakshaNag@users.noreply.github.com>
Co-authored-by: Benjamin Welton <bewelton@amd.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: SrirakshaNag <SrirakshaNag@users.noreply.github.com>
Co-authored-by: gobhardw <gopesh.bhardwaj@amd.com>
Co-authored-by: bgopesh <bgopesh@users.noreply.github.com>
Этот коммит содержится в:
Jonathan R. Madsen
2024-01-22 19:06:25 -06:00
коммит произвёл GitHub
родитель bd80fa168e
Коммит 1f4cf1aa39
51 изменённых файлов: 2031 добавлений и 648 удалений
+18 -13
Просмотреть файл
@@ -26,6 +26,7 @@
#include <rocprofiler-sdk/rocprofiler.h>
#include "lib/common/container/stable_vector.hpp"
#include "lib/common/static_object.hpp"
#include "lib/common/utility.hpp"
#include "lib/rocprofiler-sdk/context/context.hpp"
#include "lib/rocprofiler-sdk/context/domain.hpp"
@@ -70,24 +71,26 @@ get_buffer_offset()
bool
is_valid_buffer_id(rocprofiler_buffer_id_t id)
{
auto nbuffers = get_buffers().size();
if(!get_buffers()) return false;
auto nbuffers = get_buffers()->size();
auto offset = get_buffer_offset();
return (id.handle >= offset && id.handle < (offset + nbuffers));
}
unique_buffer_vec_t&
unique_buffer_vec_t*
get_buffers()
{
static auto _v = unique_buffer_vec_t{reserve_size_t{unique_buffer_vec_t::chunk_size}};
static auto*& _v = common::static_object<unique_buffer_vec_t>::construct(
reserve_size_t{unique_buffer_vec_t::chunk_size});
return _v;
}
instance*
get_buffer(rocprofiler_buffer_id_t buffer_id)
{
if(is_valid_buffer_id(buffer_id))
if(is_valid_buffer_id(buffer_id) && get_buffers())
{
for(auto& itr : get_buffers())
for(auto& itr : *get_buffers())
{
if(itr && itr->buffer_id == buffer_id.handle)
{
@@ -101,6 +104,8 @@ get_buffer(rocprofiler_buffer_id_t buffer_id)
std::optional<rocprofiler_buffer_id_t>
allocate_buffer()
{
if(registration::get_fini_status() > 0) return std::nullopt;
// ensure buffer has thread to handle flushing it
static auto _init_threads_once = std::once_flag{};
std::call_once(_init_threads_once, []() { internal_threading::initialize(); });
@@ -109,13 +114,13 @@ allocate_buffer()
auto _lk = std::unique_lock<std::mutex>{get_buffers_mutex()};
// initial context identifier number
auto _idx = get_buffer_offset() + get_buffers().size();
auto _idx = get_buffer_offset() + CHECK_NOTNULL(get_buffers())->size();
// make space in registered
get_buffers().emplace_back(nullptr);
CHECK_NOTNULL(get_buffers())->emplace_back(nullptr);
// create an entry in the registered
auto& _cfg_v = get_buffers().back();
auto& _cfg_v = CHECK_NOTNULL(get_buffers())->back();
_cfg_v = allocator::make_unique_static<buffer::instance>();
auto* _cfg = _cfg_v.get();
@@ -157,7 +162,7 @@ flush(rocprofiler_buffer_id_t buffer_id, bool wait)
auto idx = buff->buffer_idx++;
auto _task = [buffer_id, idx, offset]() {
auto& buff_v = get_buffers().at(buffer_id.handle - offset);
auto& buff_v = CHECK_NOTNULL(get_buffers())->at(buffer_id.handle - offset);
auto& buff_internal_v = buff_v->get_internal_buffer(idx);
if(!buff_internal_v.is_empty())
@@ -232,8 +237,8 @@ rocprofiler_create_buffer(rocprofiler_context_id_t context,
if(!opt_buff_id) return ROCPROFILER_STATUS_ERROR_BUFFER_NOT_FOUND;
buffer_id->handle = opt_buff_id->handle;
auto& buff = rocprofiler::buffer::get_buffers().at(opt_buff_id->handle -
rocprofiler::buffer::get_buffer_offset());
auto& buff = CHECK_NOTNULL(rocprofiler::buffer::get_buffers())
->at(opt_buff_id->handle - rocprofiler::buffer::get_buffer_offset());
// allocate the buffers. if it is lossless, we allocate a second buffer to store data while
// other buffer is being flushed
@@ -264,8 +269,8 @@ rocprofiler_destroy_buffer(rocprofiler_buffer_id_t buffer_id)
return ROCPROFILER_STATUS_ERROR_BUFFER_NOT_FOUND;
auto offset = rocprofiler::buffer::get_buffer_offset();
auto& buffers = rocprofiler::buffer::get_buffers();
auto& buff = buffers.at(buffer_id.handle - offset);
auto* buffers = CHECK_NOTNULL(rocprofiler::buffer::get_buffers());
auto& buff = buffers->at(buffer_id.handle - offset);
// buffer is currently being flushed or destroyed
if(buff->syncer.test_and_set()) return ROCPROFILER_STATUS_ERROR_BUFFER_BUSY;