[rocpd] Schema Updates

.gitmodules

@@ -39,4 +39,4 @@
 	url = https://github.com/pybind/pybind11.git
 [submodule "external/gotcha"]
 	path = external/gotcha
-	url = https://jrmadsen@github.com/jrmadsen/GOTCHA
+	url = https://github.com/ROCm/GOTCHA

benchmark/source/bin/mandelbrot/CMakeLists.txt

@@ -29,6 +29,8 @@ set(CMAKE_HIP_STANDARD 17)
 set(CMAKE_HIP_EXTENSIONS OFF)
 set(CMAKE_HIP_STANDARD_REQUIRED ON)
 
+find_package(rocprofiler-sdk-roctx REQUIRED)
+
 set_source_files_properties(mandelbrot.cpp PROPERTIES LANGUAGE HIP)
 set_source_files_properties(utils.cpp PROPERTIES LANGUAGE HIP)
 
@@ -37,6 +39,7 @@ target_sources(mandelbrot PRIVATE mandelbrot.cpp utils.cpp)
 target_compile_options(mandelbrot PRIVATE -W -Wall -Wextra -Wpedantic -Werror
                                           -ffp-contract=fast)
 target_include_directories(mandelbrot PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+target_link_libraries(mandelbrot PRIVATE rocprofiler-sdk-roctx::rocprofiler-sdk-roctx)
 
 install(
     TARGETS mandelbrot

benchmark/source/bin/mandelbrot/mandelbrot.cpp

@@ -25,6 +25,8 @@
 
 #include "utils.hpp"
 
+#include <rocprofiler-sdk-roctx/roctx.h>
+
 #include <hip/hip_vector_types.h>
 #include <hip/math_functions.h>
 #include <omp.h>
@@ -473,6 +475,19 @@ hipPerfMandelBrot::printResults()
     std::cout << std::endl;
 }
 
+struct roctx_range
+{
+    template <typename... Args>
+    roctx_range(Args&&... args)
+    {
+        auto _ss = std::stringstream{};
+        ((_ss << args), ...);
+        roctxRangePush(_ss.str().c_str());
+    }
+
+    ~roctx_range() { roctxRangePop(); }
+};
+
 // Wrappers for the kernel launches
 void
 hipPerfMandelBrot::float_mad(uint* out,
@@ -487,6 +502,9 @@ hipPerfMandelBrot::float_mad(uint* out,
                              int          threads_per_block,
                              int          kernelCnt)
 {
+    auto _range =
+        roctx_range{__FUNCTION__, "(streams=", getNumStreams(), ", kernels=", kernelCnt, ")"};
+
     int streamCnt = getNumStreams();
     hipLaunchKernelGGL(float_mad_kernel<float>,
                        dim3(blocks),
@@ -515,6 +533,9 @@ hipPerfMandelBrot::float_mandel_unroll(uint* out,
                                        int          threads_per_block,
                                        int          kernelCnt)
 {
+    auto _range =
+        roctx_range{__FUNCTION__, "(streams=", getNumStreams(), ", kernels=", kernelCnt, ")"};
+
     int streamCnt = getNumStreams();
     hipLaunchKernelGGL(float_mandel_unroll_kernel<float>,
                        dim3(blocks),
@@ -543,6 +564,9 @@ hipPerfMandelBrot::double_mad(uint* out,
                               int          threads_per_block,
                               int          kernelCnt)
 {
+    auto _range =
+        roctx_range{__FUNCTION__, "(streams=", getNumStreams(), ", kernels=", kernelCnt, ")"};
+
     int streamCnt = getNumStreams();
     hipLaunchKernelGGL(double_mad_kernel<double>,
                        dim3(blocks),
@@ -571,6 +595,9 @@ hipPerfMandelBrot::double_mandel_unroll(uint* out,
                                         int          threads_per_block,
                                         int          kernelCnt)
 {
+    auto _range =
+        roctx_range{__FUNCTION__, "(streams=", getNumStreams(), ", kernels=", kernelCnt, ")"};
+
     int streamCnt = getNumStreams();
     hipLaunchKernelGGL(float_mandel_unroll_kernel<double>,
                        dim3(blocks),
@@ -589,6 +616,8 @@ hipPerfMandelBrot::double_mandel_unroll(uint* out,
 void
 hipPerfMandelBrot::run(unsigned int testCase, unsigned int /* deviceId */)
 {
+    auto _run_range = roctx_range{__FUNCTION__, "(testCase=", testCase, ")"};
+
     unsigned int numStreams = getNumStreams();
     coordIdx                = testCase % numCoords;
 
@@ -667,6 +696,8 @@ hipPerfMandelBrot::run(unsigned int testCase, unsigned int /* deviceId */)
 
     for(unsigned int k = 0; k < numLoops; k++)
     {
+        auto _loop_range = roctx_range{__FUNCTION__, "(testCase=", testCase, ") :: loop #", k};
+
         if((testCase == 0 || testCase == 1 || testCase == 2 || testCase == 5 || testCase == 6 ||
             testCase == 7 || testCase == 10 || testCase == 11 || testCase == 12))
         {
@@ -805,6 +836,8 @@ hipPerfMandelBrot::checkData(uint* ptr)
 int
 main(int argc, char* argv[])
 {
+    auto _range = roctx_range{argv[0]};
+
     // Default values for kernels and streams
     unsigned int numStreamsWarmup = 1, numKernelsWarmup = 1;
     unsigned int numStreamsSync = 1, numKernelsSync = 1;

cmake/rocprofiler_build_settings.cmake

@@ -93,10 +93,12 @@ rocprofiler_target_compile_options(
     rocprofiler-sdk-debug-flags INTERFACE "-g3" "-fno-omit-frame-pointer"
                                           "-fno-optimize-sibling-calls")
 
-target_compile_options(
-    rocprofiler-sdk-debug-flags
-    INTERFACE $<$<COMPILE_LANGUAGE:C>:$<$<C_COMPILER_ID:GNU>:-rdynamic>>
-              $<$<COMPILE_LANGUAGE:CXX>:$<$<CXX_COMPILER_ID:GNU>:-rdynamic>>)
+if(NOT ROCPROFILER_ENABLE_CLANG_TIDY)
+    target_compile_options(
+        rocprofiler-sdk-debug-flags
+        INTERFACE $<$<COMPILE_LANGUAGE:C>:$<$<C_COMPILER_ID:GNU>:-rdynamic>>
+                  $<$<COMPILE_LANGUAGE:CXX>:$<$<CXX_COMPILER_ID:GNU>:-rdynamic>>)
+endif()
 
 if(NOT APPLE AND NOT ROCPROFILER_ENABLE_CLANG_TIDY)
     target_link_options(rocprofiler-sdk-debug-flags INTERFACE

cmake/rocprofiler_options.cmake

@@ -51,6 +51,11 @@ rocprofiler_add_option(
     "Use (internal) <rocprofiler-sdk/rccl/details/api_trace.h> instead of RCCL-provided <rccl/amd_detail/api_trace.h>. Note: this should never be used in production"
     OFF
     ADVANCED)
+rocprofiler_add_option(
+    ROCPROFILER_BUILD_PYTHON
+    "Enable building the Python bindings for roctx and rocpd. Note: this should not be disabled unless absolutely necessary"
+    ON
+    ADVANCED)
 
 rocprofiler_add_option(
     ROCPROFILER_BUILD_GHC_FS
@@ -162,3 +167,30 @@ set(ROCPROFILER_DEFAULT_FAIL_REGEX
 # this should be defaulted to OFF by ROCm 7.0.1 or 7.1 this should only used to disable
 # sample tests in extreme circumstances
 option(ROCPROFILER_DISABLE_UNSTABLE_CTESTS "Disable unstable tests" ON)
+
+if(ROCPROFILER_BUILD_PYTHON)
+    # make sure we have all python version candidates
+    set(ROCPROFILER_PYTHON_VERSION_CANDIDATES
+        "3.20;3.19;3.18;3.17;3.16;3.15;3.14;3.13;3.12;3.11;3.10;3.9;3.8;3.7;3.6"
+        CACHE STRING "Python versions to search for, newest first")
+
+    if(NOT ROCPROFILER_PYTHON_VERSIONS)
+        unset(ROCPROFILER_PYTHON_VERSIONS CACHE)
+        rocprofiler_get_default_python_versions(DEFAULT_PYTHON_VERSIONS)
+        set(ROCPROFILER_PYTHON_VERSIONS
+            "${DEFAULT_PYTHON_VERSIONS}"
+            CACHE STRING "")
+    endif()
+
+    if(NOT ROCPROFILER_PYTHON_VERSIONS)
+        message(
+            FATAL_ERROR
+                "No python3 versions found for building rocprofiler-sdk Python bindings. Either install Python3 development package(s) (i.e. Python.h + python library) or set ROCPROFILER_BUILD_PYTHON=OFF"
+            )
+    endif()
+else()
+    set(ROCPROFILER_PYTHON_VERSIONS "")
+endif()
+
+rocprofiler_add_feature(ROCPROFILER_PYTHON_VERSIONS
+                        "ROCTx and ROCpd Python bindings build versions")

cmake/rocprofiler_utilities.cmake

@@ -1026,4 +1026,79 @@ function(rocprofiler_install_env_setup_files)
         COMPONENT ${RIEF_COMPONENT})
 endfunction()
 
+macro(rocprofiler_reset_python3_cache)
+    foreach(
+        _VAR
+        _Python3_Compiler_REASON_FAILURE
+        _Python3_Development_REASON_FAILURE
+        _Python3_EXECUTABLE
+        _Python3_INCLUDE_DIR
+        _Python3_INTERPRETER_PROPERTIES
+        _Python3_INTERPRETER_SIGNATURE
+        _Python3_LIBRARY_RELEASE
+        _Python3_NumPy_REASON_FAILURE
+        Python3_EXECUTABLE
+        Python3_INCLUDE_DIR
+        Python3_INTERPRETER_ID
+        Python3_STDLIB
+        Python3_STDARCH
+        Python3_SITELIB
+        Python3_SOABI
+        ${ARGN})
+        unset(${_VAR} CACHE)
+        unset(${_VAR})
+    endforeach()
+endmacro()
+
+macro(rocprofiler_find_python3 _VERSION)
+    rocprofiler_reset_python3_cache()
+
+    if("${_VERSION}" MATCHES "^([0-9]+)\\.([0-9]+)\\.([0-9]+)$")
+        find_package(Python3 ${_VERSION} EXACT ${ARGN} REQUIRED MODULE
+                     COMPONENTS Interpreter Development)
+    elseif("${_VERSION}" MATCHES "^([0-9]+)\\.([0-9]+)$")
+        find_package(Python3 ${_VERSION}.0...${_VERSION}.999 ${ARGN} REQUIRED MODULE
+                     COMPONENTS Interpreter Development)
+    else()
+        message(
+            FATAL_ERROR
+                "Invalid Python3 version (${_VERSION}). Specify <MAJOR>.<MINOR> or <MAJOR>.<MINOR>.<PATCH>"
+            )
+    endif()
+endmacro()
+
+function(rocprofiler_get_default_python_versions _VAR)
+    set(_PYTHON_FOUND_VERSIONS)
+
+    foreach(_VER IN LISTS ROCPROFILER_PYTHON_VERSION_CANDIDATES)
+        rocprofiler_reset_python3_cache()
+        find_package(Python3 ${_VER} EXACT QUIET COMPONENTS Interpreter Development)
+
+        if(Python3_FOUND)
+            list(APPEND _PYTHON_FOUND_VERSIONS
+                 "${Python3_VERSION_MAJOR}.${Python3_VERSION_MINOR}")
+        endif()
+    endforeach()
+
+    # If none found, do one last check for 3.6 (no EXACT)
+    if(NOT _PYTHON_FOUND_VERSIONS)
+        rocprofiler_reset_python3_cache()
+        find_package(Python3 3.6 COMPONENTS Interpreter Development)
+
+        if(Python3_FOUND)
+            list(APPEND _PYTHON_FOUND_VERSIONS
+                 "${Python3_VERSION_MAJOR}.${Python3_VERSION_MINOR}")
+        endif()
+    endif()
+
+    # Set the output variable to the first found version, if any
+    if(_PYTHON_FOUND_VERSIONS)
+        set(${_VAR}
+            "${_PYTHON_FOUND_VERSIONS}"
+            PARENT_SCOPE)
+    endif()
+
+    rocprofiler_reset_python3_cache()
+endfunction()
+
 cmake_policy(POP)

external/CMakeLists.txt

@@ -185,6 +185,8 @@ target_sources(
     rocprofiler-sdk-perfetto-static-library
     PRIVATE ${PROJECT_SOURCE_DIR}/external/perfetto/sdk/perfetto.h
             ${PROJECT_SOURCE_DIR}/external/perfetto/sdk/perfetto.cc)
+target_compile_definitions(rocprofiler-sdk-perfetto-static-library
+                           PRIVATE $<BUILD_INTERFACE:PERFETTO_DISABLE_LOG=1>)
 target_include_directories(
     rocprofiler-sdk-perfetto-static-library SYSTEM
     INTERFACE $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/external/perfetto/sdk>)

source/include/rocprofiler-sdk-rocpd/sql.h

@@ -57,7 +57,7 @@ typedef enum ROCPD_EXPERIMENTAL rocpd_sql_schema_kind_t
     ROCPD_SQL_SCHEMA_ROCPD_VIEWS,
     ROCPD_SQL_SCHEMA_ROCPD_DATA_VIEWS,
     ROCPD_SQL_SCHEMA_ROCPD_SUMMARY_VIEWS,
-    ROCPD_SQL_SCHEMA_ROCPD_MARKER_VIEWS,
+    ROCPD_SQL_SCHEMA_ROCPD_METADATA,
     ROCPD_SQL_SCHEMA_LAST,
 } rocpd_sql_schema_kind_t;
 

source/lib/common/logging.hpp

@@ -52,10 +52,10 @@
 
 #if defined(ROCPROFILER_CI)
 #    define ROCP_CI_LOG_IF(NON_CI_LEVEL, ...) ROCP_FATAL_IF(__VA_ARGS__)
-#    define ROCP_CI_LOG(NON_CI_LEVEL, ...)    ROCP_FATAL
+#    define ROCP_CI_LOG(NON_CI_LEVEL)         ROCP_FATAL
 #else
 #    define ROCP_CI_LOG_IF(NON_CI_LEVEL, ...) ROCP_##NON_CI_LEVEL##_IF(__VA_ARGS__)
-#    define ROCP_CI_LOG(NON_CI_LEVEL, ...)    ROCP_##NON_CI_LEVEL
+#    define ROCP_CI_LOG(NON_CI_LEVEL)         ROCP_##NON_CI_LEVEL
 #endif
 
 namespace rocprofiler

source/lib/common/simple_timer.cpp

@@ -22,6 +22,7 @@
 
 #include "lib/common/simple_timer.hpp"
 #include "lib/common/logging.hpp"
+#include "lib/common/synchronized.hpp"
 
 #include <fmt/format.h>
 
@@ -35,14 +36,16 @@ namespace rocprofiler
 {
 namespace common
 {
-simple_timer::simple_timer(std::string&& label)
+simple_timer::simple_timer(std::string&& label, int log_level)
 : m_label{std::move(label)}
+, m_log_level{log_level}
 {
     start();
 }
 
-simple_timer::simple_timer(std::string&& label, defer_start)
+simple_timer::simple_timer(std::string&& label, defer_start, int log_level)
 : m_label{std::move(label)}
+, m_log_level{log_level}
 {}
 
 simple_timer::~simple_timer()
@@ -52,19 +55,21 @@ simple_timer::~simple_timer()
     else if(m_end <= m_beg)
         stop();
 
-    ROCP_WARNING << fmt::format("{} :: {:12.6f} sec", m_label, get());
+    report();
 }
 
-void
+simple_timer&
 simple_timer::start()
 {
     m_beg = clock_type::now();
+    return *this;
 }
 
-void
+simple_timer&
 simple_timer::stop()
 {
     m_end = clock_type::now();
+    return *this;
 }
 
 double
@@ -81,6 +86,51 @@ simple_timer::get_nsec() const
     return std::chrono::duration_cast<std::chrono::nanoseconds>(m_end - m_beg).count();
 }
 
+simple_timer&
+simple_timer::report()
+{
+    static auto max_width = Synchronized<uint64_t>{0};
+    max_width.wlock(
+        [](auto& _max_width, auto _w) {
+            if(_w > _max_width) _max_width = _w;
+            if(_max_width > 120) _max_width = 120;
+        },
+        m_label.size());
+
+    auto _width = max_width.get();
+
+    switch(m_log_level)
+    {
+        case ROCP_LOG_LEVEL_WARNING:
+        {
+            ROCP_WARNING << fmt::format("{:<{}} :: {:12.6f} sec", m_label, _width, get());
+            break;
+        }
+        case ROCP_LOG_LEVEL_ERROR:
+        {
+            ROCP_ERROR << fmt::format("{:<{}} :: {:12.6f} sec", m_label, _width, get());
+            break;
+        }
+        case ROCP_LOG_LEVEL_INFO:
+        {
+            ROCP_INFO << fmt::format("{:<{}} :: {:12.6f} sec", m_label, _width, get());
+            break;
+        }
+        case ROCP_LOG_LEVEL_TRACE:
+        {
+            ROCP_TRACE << fmt::format("{:<{}} :: {:12.6f} sec", m_label, _width, get());
+            break;
+        }
+        case ROCP_LOG_LEVEL_NONE:
+        default:
+        {
+            break;
+        }
+    }
+
+    return *this;
+}
+
 std::ostream&
 operator<<(std::ostream& _os, const simple_timer& _val)
 {