Add NVML Provider and GPU Energy Monitoring Tools

This commit adds comprehensive NVML-based GPU energy monitoring:

NVML Provider:
- NVMLProvider class for GPU power monitoring (provider_nvml.cpp/hpp)
- Device discovery and power usage APIs
- Comprehensive error handling and status checking
- Support for multiple GPUs with individual device queries

Kokkos Tools:
- kp_nvml_power: Background power monitoring with daemon
- kp_nvml_direct_power: Direct power measurement tool
- kp_nvml_energy_consumption: Energy consumption tracking

Testing:
- Unit tests for NVML provider functionality
- Integration tests with daemon system
- Fast sampling tests (20ms intervals)

Features:
- Configurable sampling intervals
- CSV export of power data with timestamps
- Thread-safe data collection
- Comprehensive power statistics (min/max/avg/total energy)

Depends on: feature/energy-profiler-daemon

Author: DaKerboul

profiling/energy-profiler/CMakeLists.txt

@@ -1,10 +1,12 @@
-# Energy Profiler Infrastructure with Daemon Support
-# This provides the foundational infrastructure for energy profiling tools:
+# Energy Profiler Infrastructure with NVML Provider
+# Complete infrastructure for NVML-based GPU energy profiling:
 # - Timer system for kernel and region tracking
 # - Common utilities (filename generation, error handling)
 # - Tool interface definitions
 # - Basic kernel timer tool
 # - Daemon system for background monitoring tasks
+# - NVML provider for NVIDIA GPU power monitoring
+# - Multiple NVML-based profiling tools
 
 option(KOKKOS_TOOLS_ENABLE_ENERGY_PROFILER_VERBOSE "Enable verbose output for energy profiler tools" OFF)
 

profiling/energy-profiler/kokkos-tools/CMakeLists.txt

@@ -1,5 +1,8 @@
-# Energy Profiler Infrastructure - Kokkos Tools
-# Basic kernel timer tool using the timing infrastructure
+# Energy Profiler Infrastructure - Kokkos Tools with NVML Support
+# Basic kernel timer tool and NVML-based energy monitoring tools
+
+# Find Threads package for pthread support (needed by std::thread in daemon.cpp)
+find_package(Threads REQUIRED)
 
 # Basic kernel timer tool (no energy monitoring dependencies)
 kp_add_library(kp_energy_kernel_timer kp_energy_kernel_timer.cpp
@@ -12,6 +15,59 @@ target_include_directories(kp_energy_kernel_timer PRIVATE
 )
 
 # Add verbose output option
-if(KOKKOS_TOOLS_ENABLE_ENERGY_PROFILER_VERBOSE)
-    target_compile_definitions(kp_energy_kernel_timer PRIVATE KOKKOS_TOOLS_ENABLE_VERBOSE_OUTPUT)
+if(ENABLE_ENERGY_PROFILER_VERBOSE)
+    target_compile_definitions(kp_energy_kernel_timer PRIVATE ENABLE_VERBOSE_OUTPUT)
+endif()
+
+# NVML-based tools
+find_package(CUDAToolkit QUIET)
+
+if (CUDAToolkit_FOUND)
+    find_package(CUDA::nvml QUIET)
+    if(TARGET CUDA::nvml)
+        message(STATUS "Found CUDA NVML, making NVML power profiler available.")
+        
+        # NVML Power Tool (with daemon)
+        kp_add_library(kp_nvml_power kp_nvml_power.cpp
+            ../common/daemon.cpp
+            ../common/filename_prefix.cpp
+            ../common/timer_system.cpp
+            ../provider/provider_nvml.cpp
+        )
+        target_link_libraries(kp_nvml_power PRIVATE CUDA::nvml Threads::Threads)
+
+        target_include_directories(kp_nvml_power PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}
+        )
+
+        # NVML Direct Power Tool (with daemon)
+        kp_add_library(kp_nvml_direct_power kp_nvml_direct_power.cpp
+            ../common/daemon.cpp
+            ../common/filename_prefix.cpp
+            ../common/timer_system.cpp
+            ../provider/provider_nvml.cpp
+        )
+        target_link_libraries(kp_nvml_direct_power PRIVATE CUDA::nvml Threads::Threads)
+
+        target_include_directories(kp_nvml_direct_power PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}
+        )
+
+        # NVML Energy Consumption Tool (no daemon needed)
+        kp_add_library(kp_nvml_energy_consumption kp_nvml_energy_consumption.cpp
+            ../common/filename_prefix.cpp
+            ../common/timer_system.cpp
+            ../provider/provider_nvml.cpp
+        )
+        target_link_libraries(kp_nvml_energy_consumption PRIVATE CUDA::nvml)
+
+        target_include_directories(kp_nvml_energy_consumption PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}
+        )
+    else()
+        message(STATUS "CUDA::nvml target not found, skipping NVML power profiler.")
+    endif()
+else()
+    message(STATUS "CUDAToolkit not found, skipping NVML power profiler.")
+endif()
 endif()