Yeppp!

About the project

Yeppp! library is a collection of low-level functions optimized for modern processors using SIMD instructions.

• Available on all major platforms (see full list of platforms):
• Compatible with Windows, Linux, and Android.
• Maintains binary compatibility across Linux distributions.
• Supports x86, x86-64, ARM, and MIPS processors.
• Code once, high performance everywhere!
• Each library function has several versions optimized for different microarchitectures.
• The optimal function implementation is chosen in run-time depending on processor microarchitecture and available instruction extensions.
• Highly optimized vector mathematical functions (see benchmarks)
• High-performance polynomial evaluation functions (see benchmarks and usage example).
• Cross-platform access to CPU information:
• Detection of processor microarchitecture and instruction sets.
• Portable access to CPU cycle counter and high-resolution system timer.
• Supports reading CPU cycle counter on ARM and MIPS architectures.
• Easy to use right out of the box:
• Precompiled binaries for Windows, Linux, and Android.
• Comprehensive documentation (browse online).
• C and C++-compatible header files.
• Official bindings for Java (JNI), FORTRAN, and .Net
• Usage examples for R.
• Free and open-source under permissive license.
• Licensed under 3-clause BSD license.
• Code available in Yeppp! Bitbucket repository

Support

You are welcome to post questions or suggestions about the Yeppp! library in yeppp mailing list

Releases

Next release: Beta 3 (library version 0.9.9.0).

To be released in mid-June.

Improvements:

• Even faster polynomial evaluation function for Haswell (see Yeppp! vs GCC 4.8 vs Clang 3.3).
• Improved performance of vector exp function on Bulldozer and Haswell (see benchmark).
• Haswell-optimized version of log function (see benchmark)
• New software-pipelined implementations in yepCore module (Haswell versions included).

New features:

• Mac OS X support (x86 and x86-64 architectures).
• Additional data-processing function: sum/max/min of all elements/absolute values of elements, operations with constants (e.g. add constant, subtract from constant, etc), in-place operations.

Current release: Beta 2 (library version 0.9.8.0).

Released on May 28, 2013; updated on May 29, 2013.

Improvements in vector functions:

• New function for evaluation of the specified polynomial on an array with optimizations for all significant x64 microarchitectures and great performance (see benchmarks and usage example).
• Improved performance of vector sine function on AMD Bulldozer & Piledriver.
• Additional version of vector sine function with optimization for Haswell.
• Vector cosine function with optimizations for Nehalem, Bulldozer, Sandy Bridge, and Haswell.

Improvements in language bindings:

• C# and FORTRAN bindings are now auto-generated and cover all functionality of yepCore and yepMath modules.
• Improved Makefile for FORTRAN bindings to recognize both Intel and GNU compilers, and choose options accordingly.

Other improvements:

• Windows builds now have version information built-in
• Android builds and most Linux builds have separate files with library binary and debug information.

Previous release: Beta 1 (library version 0.9.7.0).

Released on May 19, 2013; updated on May 20, 2013.

Improvements:

• Super fast exponential and logarithm functions (see benchmarks) optimized for all significant x64 architectures (Haswell (exp only), Sandy Bridge, Nehalem, Bulldozer, K10, Bobcat).
• Sin and Tan functions.
• Builds for ARM, MIPS, and Android/x86 are provided again.

Other changes:

• Project switched to 3-clause BSD license for compatibility with other open-source codes which might be used in future versions of Yeppp!

Previous release: Technology Preview 6 (library version 0.9.6.1).

Released on April 30, 2013.

Improvements:

• New versions of vector exponent and logarithm functions optimized for AMD Bulldozer/Piledriver and Intel Sandy Bridge/Ivy Bridge.
• Refactored initialization code.
• Debug information files.
• Simplified building the library with new build system.

New functionality:

• Functions to retrieve cleanly formatted human-readable processor name (e.g. "Intel Xeon X5550", "AMD FX-6300", or "Samsung Exynos 5250").
• New yepAtomic subsystem with functions for atomic operations.
• New yepRandom subsystem for random number generation.
• New functions for reading energy and power counters on Sandy Bridge and Ivy Bridge.
• Support for Xeon Phi architecture

Regressions:

• Binaries for bindings are not provided with these release
• Binaries for ARM, MIPS, and Android/x86 are not provided
• No VERSIONINFO resource in Windows binaries
• Bug in dot product kernel on Sandy Bridge

Improved language bindings:

• All functions (including yepLibrary module) in are made accessible from Java.
• Complete documentation for Java bindings (browse online)
• Initial release of CLR (.Net/Mono via P/Invoke) and FORTRAN 2003 bindings.
• Examples of using Java and CLR bindings:
• All examples are ported to Java
• Entropy example ported to C#

New functionality:

• Vector exponent function added (with optimized kernel for Nehalem)
• Java bindings now allow to process arbitrary slice of an array

This release includes sources for all library components:

• Yeppp! library (optimized kernels, CPU dispatcher, and CPU information retrieval)
• PEACH-Py code-generator (generates kernels and support code)
• Yeppp! build framework (builds the library; unusable in this release)
• Yeppp! runtime (support functions typically provided by compiler runtime; uses MIT license)
• Bindings for Java, D, and R
• Build scripts

New functionality:

• Two new ARM platforms are supported:
• arm-linux-softeabi-android corresponds to "armeabi" in Android NDK
• arm-linux-softeabi-v5t provides binaries compatible with soft-float ARM EABI on Linux (e.g. Debian armel port)
• Sum-of-squares function (with optimized kernels for Nehalem, Sandy Bridge, and Bulldozer)
• Multiply-add vector function
• Additional optimized kernels

Improved CPU detection:

• Implemented detection of microarchitecture for quad-core Qualcomm Krait cores.
• Implemented detection of microarchitecture for Intel Cloverview cores (new Atom for Tablets).
• Fixed erroneous detection of Piledriver cores as Bulldozer cores.
• Implemented detection of MIPS R2, MIPS Paired-Single and MIPS DSP R2 extensions.
• Improved detection of VFPv3-D32 on ARM to work properly on misconfigured Linux kernels.

Other improvements:

• Documentation lists optimized implementations for every function.
• Linux and Android binaries now provide version information in .version  ELF section

Note: this version is not binary compatible with previous releases

Supported functionality:

• Detection of processor vendor, architecture, microarchitecture, instruction set extensions, and other extended features.
• Cross-platform access to processor cycle counter and high-performance system timer.
• Five types of vector operations: addition, subtraction, multiplication, negation, and dot product.
• Addition, subtraction, dot product, and most multiplication functions have versions optimized for Nehalem on x64 architecture (both Linux and Windows).
• This optimized implementations are also used for other x64 microarchitectures when the required instruction sets are available.

Supported platforms: