Release notes

Latest

Version 2025.1.0

  • Additions and improvements:

    • Implement udev cache (improves DPU allocation performance)

    • LLDB: reduce amount of memory transfer when transferring unaligned data during a debug session

    • dpu-diag: Add option to also disable DPUs which have IRAM or WRAM defects

  • Bug fixes:

    • Fix some conditions in which a subset of valid DPUs would become unusable (avoid resetting faulty DPUs, do not allocate DPUs related to a non-functional Control Interface)

    • Fix memory leaks in corner cases

    • upmem_utilities: Add pyelftools as a dependency since it is needed for dpu_sections and dpu_statistics

    • Fix DPU toolchain file for CMake

  • Documentation:

    • Improve documentation style

    • Clarify various comments in code

    • Clarify installation instructions

Operating system / DPU support matrix:

Operating system

V1A DPUs

V1B DPUs

Debian 10 (1)

Supported

Supported

Ubuntu 20.04 LTS (1)

Supported

Not Supported

Ubuntu 22.04 LTS (2)

Supported

Supported

Rocky 8 (1)

Supported

Supported

  • (1) In future releases, all systems with a Linux kernel older than version 6.8 will be deprecated.

  • (2) In future releases, Ubuntu 22.04 LTS will only be supported with the Hardware enablement stack (HWE) and the kernel 6.8.

Previous

Version 2024.2.0

  • Additions and improvements:

    • Implement WRAM access in parallel of the DPU execution

    • Add DPU v1B performance counters

    • Add chip versioning tags to DPU binaries

    • Fix lowlevel header files includes

    • IWYU users: dpu.h now exports most of the DPU API

    • Rely on dpu_types.h to get C standard integer types

    • Utilities / dpu-diag:

      • Added --no-update-vpd option

      • Improved frequency measurement stability

    • Utilities / upmem-dimm-configure:

      • Update MCU DB fields

      • General improvements

    • Functional simulator: remove broken nrDpusPerCi option

  • Bug fixes:

    • Fix rare Control Interface issue on V1B

    • LLVM/Clang: minor bug fixes

    • Run-time libraries:

      • sequential reader: fix correctness issue

      • Improvements to div32 and udiv32

      • mem-alloc: early clobber output variable

    • MCU firmware: fix utils alignment issue

  • Documentation:
    • Run-time library documentation now in doxygen format

      • Remove redundant dpurtldoc CLI tool

    • Use modern theme for doxygen sites

    • Support dark mode for documentation

    • Document Upmem PPA

    • Document UPMEM_PROFILE

Operating system / DPU support matrix:

Operating system

V1A DPUs

V1B DPUs

Debian 10

Supported

Supported

Ubuntu 20.04 LTS

Supported

Not Supported

Ubuntu 22.04 LTS

Supported

Supported

Rocky 8

Supported

Supported

Version 2024.1.0

  • Add support for V1B UPMEM DIMMs on Ubuntu 22.04 and Debian 10

  • Fix performance degradation caused by Gather Data Sampling microcode mitigation Courtesy of Manuel Penschuck from Goethe University Frankfurt. For further performance improvements, check out his awesome work on his fork of the UPMEM backend (benchmarks)

  • Add dpu_attach_first command to the debugger

  • Drop support for Ubuntu 18.04

  • Drop support for CentOS 7

  • Add documentation for:

    • DPU chips characteristics

    • Choosing the UPMEM chip version when compiling and running on simulator (cf. DPU Version Selection)

    • Configuring the UPMEM debugger port

  • Bugfixes:

    • Memory leak in scatter-gather transfers

    • Memory leaks and sanitization in LLVM tools

    • Pyelftools version conflict in dpu_profiling

    • Wrong python libraries path on rocky8

Operating system / DPU support matrix:

Operating system

V1A DPUs

V1B DPUs

Debian 10

Supported

Supported

Ubuntu 20.04 LTS

Supported

Not Supported

Ubuntu 22.04 LTS

Supported

Supported

Rocky 8

Supported

Supported

Version 2023.2.0

  • Scatter-gather transfers update:

    • The get_block_t struct now needs an additional args_size parameter to store the size of the function context

    • The SDK now copies and manages the lifetime of the args context used by the get_block_func_t function

    • Add C++ API for scatter-gather transfers

    • C++ API for scatter-gather transfers also accepts capturing lambdas

  • Add support for Ubuntu 22.04

  • Add support for Rocky 8

  • Add documentation for:

    • Setting up permissions for users access UPMEM hardware and for profiling

    • Valid values for nrDpusPerCi with the functional simulator

    • Using the clangd language server with the SDK

  • Fix for scatter-gather transfers in the functional simulator

  • Fix a bug in UFI config when counting faulty bits

  • Fix an edge-case compiler crash when using -O1 or higher with mram_read or mram_write.

Version 2023.1.0

  • New APIs for scatter-gather memory transfers between host CPU and DPUs

  • Mutual exclusions extension : virtual mutexes and mutex pools

  • New APIs for partial support of unaligned MRAM accesses

  • Remove the dependency of dpu_profiling functions on UPMEM’s modified Linux perf package

  • Add compatibility of dpu_profiling functions with Perfetto UI

  • Add static assert to prevent usage of perfcounter_config with -pg option

  • Update the UPMEM driver to support Linux kernel 5.15

  • Firmware support for PCB P21F

Version 2021.4.0

  • Fix bug in calling convention when a DPU function has a large number of parameters

  • Fix bug with mux switch for sparse matrix MRAM transfer

  • Fix for putting DIMMs in power-saving mode

  • Static assertion when perfcounter_config is used with -pg option

  • Firmware support for PCB P21E

Version 2021.3.0

  • Upgrade compiler to LLVM-12.0.0

  • Add support for new P21D PCB

  • Update BIOS flashing procedure based on new flasher kit

  • Document functional simulator

  • Fix a corner case bug with Host <-> DPU MRAM transfer

Version 2021.2.0

  • New DPU profiling framework based on perfcounters

  • Fix Functional Simulator MRAM management when using multiple DPUs

  • Enhance Function Simulator instructions tracing

  • Enhance DPU 8-bit multiplication

  • Fix DPU 64-bit rotation

  • Fix MCU firmware bug

  • Update server installation documentation

Version 2021.1.0

  • Small improvements in Java host API (cf Java Library)

  • Upgrade compiler to LLVM-11.0.0

  • Fix various bugs in LLDB

  • Add CLOCKS_PER_SEC constant in DPU library

  • Fix bugs in host API

  • Improve DPU program loading performance for multi-rank call

  • Fix bug in dpugrind

  • Add documentation about C++/Java/Python APIs

Version 2020.5.0

  • Adding support for dual-socket servers

  • Adding an host Java API (cf Java Library)

  • Adding an host C++ API (cf C++ Host API)

  • Improving the performance of the Functional Simulator

Version 2020.4.0

  • Remove dependency on the JDK

  • DPU Runtime Library: memcpy/memmove/memset can now be used with MRAM pointers as arguments

  • Adding asynchronism in the Host API:

    • dpu_push_xfer now has a DPU_XFER_ASYNC flag to copy asynchronously

    • dpu_broadcast_to, a new function with the same behavior as dpu_copy_to, but contrary to its predecessor, can be used asynchronously

    • dpu_callback, a new function to insert callback that will be called asynchronously during the execution flow of the DPUs.

Version 2020.3.0

  • Upgrade compiler to LLVM-10.0.0

  • Remove dependency on Python 2: only Python 3 is needed

  • Add Python host API (cf Python Library)

  • Add support for ubuntu 20.4

  • Remove support for debian 9

  • Update driver ABI (and version):

    • Gather sysfs attributes in a single directory

    • Remove region index from device nodes’ name

  • Host API improvement (parallelize transfer when using dpu_set_t of multiple ranks)

  • Add tool dpu-profiling to unify all profiling features

Version 2020.2.0

  • New transparent management of the DPU SRAM repairs

  • Improved implementation of the MRAM sequential reader

  • Core dump binaries now contain all the debug sections

  • Both -flto=full and -flto=thin are supported for the DPU Runtime Library

Version 2020.1.0

  • Improve significantly communications performance including copy to/from MRAM

  • Add support for the centos8 Linux distribution

  • Add the dpu-diag self-diagnostic tool to dump information & check the SDK installation

  • Update the functions to access MRAM from the DPU (only use mram_read and mram_write)

  • Macros to get barriers, mutex and semaphore (<OBJECT>_GET) are no longer necessary

  • Interrogate the DIMM hardware to get the DPU clock frequency rather than declaring it through UPMEM_PROFILE

Version 2019.4.0

  • Upgrade compiler to LLVM-9.0.1

  • dpu-lldb is now the default debugger

  • Add tool dputrace for low level execution tracing (cf Using dputrace tool)

  • Add tool dpu_stack_analyzer for static analysis of the DPU threads stack size (cf Stack Analyzer)

  • Remove tasklet mailbox concept

  • Refactor the Host API to be based on DPU sets

  • Experimental tool: dpugrind is now available (cf Verifying memory accesses with dpugrind)

Version 2019.3.0

  • Unification of the DPU Host libraries libdpu and libdpucni

  • Upgrade compiler to LLVM-8.0.1

  • Clang is now able to compile x86 code, the default target is x86

  • Add tool dpu-upmem-dpurte-clang which uses clang to target dpu backend by default

  • The DPU runtime library is now compiled with debug information

  • Experimental tool: LLDB is now available for dpu

Version 2019.2.0

  • Adding DPU logging module based on MRAM buffer (cf printf in stdio.h)

  • KTRACE and printf support print of 64-bit integers/floats.

  • The Assembler syntax changed:

    • % before a register is now optional and not print by llvm-objdump.

    • ? and ! before a condition is now optional and not print by llvm-objdump.

    • Condition in_buffX has been replaced by ncX.

  • Stabilization of MRAM variables

Version 2019.1.0

  • DPU program binaries are entirely built by clang, dpucc has been removed

  • the runtime library is pure C, there is no longer assembly generated by a configuration file and dpukconfig has been removed

  • the tasklets configuration is done by declaring TASKLETS_INITIALIZER and including rt.h header (cf Tasklet management and synchronization)

  • updated runtime API in barrier.h, mutex.h, sem.h

  • the default optimization level for clang is -O0, the developer must pass explicitly -O2

  • the compiler handles functions with variable arguments properly

  • Changes in the assembly syntax

  • Experimental feature: add variables stored in MRAM

  • Experimental feature: a profiler for DPU binaries (cf Profiling DPU binary)

Version 2018.2.0

  • Upgrade compiler to LLVM-7.0

  • Experimental feature: DPUGrind tool

Version 2018.1.0-EAP (January 2018)

Version 2017.1.5-EAP

  • Support for klog logging source in the host API

  • Finely tune the profiling parameters to select the instrumented data and the output file

Version 2017.1.4-EAP

  • DPU program binary format is now standard ELF

  • DPU programs built with an older toolchain version need to be rebuilt

  • Optimized procedures in the compiler for 8x8 and 16x16 multiplications

  • Optimized procedures in the compiler for variable initialization

  • Enriching and optimizing string.h (cf Standard library functions)

  • Adding dynamic allocation with the implementation of a buddy allocator (cf Memory management)

  • Experimental feature: adding a Link Time Optimizer in dpucc (enabled when building with -flto)

Version 2017.1.3-EAP

  • Bug fixes, including KI0284

  • ktrace correctly handles %u format specifier

Version 2017.1.2-EAP

  • 64-bits types and operators support in the toolchain

  • Fixed-Size Block Allocator (available in fsb_allocator.h or directly from alloc.h)

  • Make the DPU-side API of ktrace mimic printf

  • Build host APIs for ktrace, so that a x86 program can collect DPU traces programmatically

  • IRAM and WRAM sizes when using the simulator are now configurable

  • Software caches library functions optimized with 64-bits DMA instructions

  • dpucc accepts static libraries as inputs

Version 2017.1.1-EAP

  • DMA instructions granularity and alignment on 64 bits

  • -Os support for code optimization in size

  • ktrace service infrastructure to print execution string and variables values in dpushell without using processor resources (cf Logging)

Version 0.7.5-EAP

  • -O2 is now the default optimization of dpucc

  • Command-line profiling tool

  • Host API now accepts an “unlimited” number of simulators

  • dpukconfig has the ‘-s’ option to run scripts

  • dpugeni available to create native interfaces to DPUs

  • Assembler anonymous labels to easy asm code inlining

Version 0.7.4-EAP

  • Host API allows callback parameters to tune DPU allocation on the fly

  • Many optimizations and bug fixes of the C compiler

  • Actor programming model available (deprecated)

  • Software caches available to simplify prototyping

Version 0.7.3-EAP

Version 0.7.2-EAP

  • -O1 is now the default optimization of dpucc

  • Many bug fixes and improvements

Version 0.7.1-EAP