What’s New

What’s New#

Warning

DRAFT DOCUMENTATION - This documentation is currently in draft status and subject to change.

Version 1.2.0#

Current release

Updates in this release:

  • Add the new ptiMetricsEnable and ptiMetricsDisable APIs, which enable metrics collection at runtime and remove the need to set the ZET_ENABLE_METRICS=1 environment variable before the process starts. Calls are reference counted per device and must be matched. ptiMetricsEnable returns the new PTI_ERROR_METRICS_RUNTIME_ENABLE_UNSUPPORTED result code when the installed driver does not support runtime enabling, in which case the environment variable is still required; other driver failures of both APIs are reported as PTI_ERROR_DRIVER. The new PTI_ERROR_METRICS_ENABLE_FAILED result code is also added, for metrics that could not be enabled on all devices.

  • Make ptiPcSamplingConfigure optional. If it is not called, ptiPcSamplingStartCollection applies the default configuration (the first device that supports PC sampling, sampling period of 100000 ns). Configuration failures are now reported with the new PTI_ERROR_PC_SAMPLING_CONFIGURATION_FAIL result code; PTI_ERROR_PC_SAMPLING_NOT_CONFIGURED is kept for backward compatibility.

  • Behavior change: libpti_view.so now registers itself as the ITT collector at load time, so oneCCL communication tracing works without extra environment setup. When INTEL_LIBITTNOTIFY64 is unset, a static initializer sets it to the absolute path of libpti_view.so before main(). As a result, ITT is loaded in every process that loads libpti_view.so, even if PTI_VIEW_COMMUNICATION is never enabled, and child processes (for example torchrun or mpirun workers) inherit the variable. To opt out, set the variable to another collector before the process starts; PTI then leaves it unchanged and disables its own oneCCL tracing. See Known Issues.

  • Significantly reduce tracing overhead for Level Zero native graph execution (about 40% lower in internal measurements) by gathering the timestamps of the whole graph in one batched submission and removing an unnecessary wait on in-order command lists. The first graph execution under device tracing is still slow; see Known Issues.

  • Behavior change: Add zeCommandQueueExecuteCommandLists to the PTI_API_CLASS_GPU_OPERATION_CORE API class, covering applications that submit work through non-immediate command lists. Applications that enable this class with ptiViewEnableDriverApiClass now receive additional PTI_VIEW_DRIVER_API records (and the associated overhead) for these calls.

  • Behavior change: ptiViewOverheadKindToString now returns "Instrumentation" for PTI_VIEW_OVERHEAD_KIND_TIME (previously "BUFFER_TIME"). Code that compares the returned string must be updated; compare the enum value where possible.

  • Fix a stale external correlation ID being reported alongside the live one when an ID is popped inside a ptiCallback subscriber callback.

  • Fix duplicate kernel records when tracing Level Zero graph creation with compute-runtime 26.31 and later.

  • Add build support for the open-source DPC++/DPCLang compiler (DPC++ 7.0), Rocky Linux 8, and building PTI as a CMake subproject, and fix compilation errors with the Intel oneAPI Compiler 2026.2.

Version 1.1.0#

Updates in this release:

  • Add support for on-demand profiling of Level Zero native graphs, validated on Intel(R) Arc(TM) B-Series (BMG) with a single GPU. Requires Intel oneAPI DPC++/C++ Compiler 2026.1.1 and compute-runtime 1.15.39122+12 or newer.

  • Update to Level Zero Loader version 1.32.0 and add the new Level Zero graph and SYCL Unified Runtime API IDs to pti_driver_levelzero_api_ids.h and pti_runtime_sycl_api_ids.h.

  • Intern kernel names to avoid unbounded memory growth in applications that launch many kernels.

  • Keep demangled kernel names in aggregated ptiPcSampling results.

  • Fix a data race on the buffer handoff between the thread producing records and the thread delivering buffers to the user.

  • Remove the obsolete bridge kernel GPU tracing path; the PTI_FORCE_BRIDGE_COMMANDS environment variable no longer has any effect.

  • Fix loading the PTI core library from the Python wheel: the dependencies it cannot find through its own RPATH are now pre-loaded by the PTI interface library. Fix builds with Intel oneAPI Compiler 2025.x and 2026.0.

Version 1.0.1#

Updates in this release:

  • Document that ptiPcSamplingEnable returns PTI_ERROR_PC_SAMPLING_UNSUPPORTED when PC sampling is not supported on the system.

  • Fix undefined symbols when running with LTS drivers on Intel(R) Data Center GPU Max Series: Level Zero symbols newer than loader version 1.14.0 are loaded dynamically.

  • Fix a crash on Windows when the PTI library is unloaded.

  • Fix builds with newer Level Zero Loader headers and MSVC Debug builds, and reconfigure the build when the VERSION file changes.

Version 1.0.0#

Updates in this release:

  • Declare the ptiView API stable for the 1.0 release, with fixed 4-byte enum sizes for a stable ABI.

  • Extend the pti_view_record_kernel, pti_view_record_memory_copy, pti_view_record_memory_copy_p2p, and pti_view_record_memory_fill records with new device handle and engine (ordinal and index) fields, defined as the new _v2 record structures pti_view_record_kernel_v2, pti_view_record_memory_copy_v2, pti_view_record_memory_copy_p2p_v2, and pti_view_record_memory_fill_v2. The previous records remain compatible with the new ones.

  • Add a proof-of-concept of GPU PC Sampling via the new ptiPcSampling API (Experimental) for statistical per-instruction stall sampling on GPU kernels. The API is complete, but the functionality is not: specifically, “address to source line mapping” is not yet implemented and is expected in the next version.

  • Add the new ptiCallback API (Experimental), providing synchronous notifications for GPU operation lifecycle events, with kernel handle context and driver/kernel callback domains.

  • Extend ptiMetricsScope with multi-device support, including multi-process auto-mode collection running in parallel.

  • Add an initial proof-of-concept for on-demand SYCL Graph profiling with reduced overhead for repeatedly executed graphs.

  • Reduce host-side runtime collection overhead significantly, in part by switching to GPU-time interpolation. Further reduction is in progress.

  • Fix device timestamp desynchronization that could produce incorrect kernel start/end times when the device submit time exceeded the device start time.

  • Improve tracing of Intel(R) oneCCL library operations.

  • Add compatibility with Intel oneAPI Compiler 2026.0 and 2026.1, the open-source DPC++ compiler, and newer Level Zero Loader headers, including the Level Zero driver extension headers.

Version 0.17.0#

Updates in this release:

  • Add support for tracing Intel(R) oneCCL library operations (via ptiViewEnable(PTI_VIEW_COMMUNICATION))

  • Add tracing of SYCL Graph operations with capture of repeatedly executed command lists. Improvements of this capability are in progress.

  • Add compatibility with Intel oneAPI Compiler 2026.

  • Update to Level Zero Loader version 1.28.0.