Difference between revisions of "Power ISA/Vector Operations"

From RCS Wiki
Jump to navigation Jump to search
(Power ISA v3.1 adds MMA)
 
(17 intermediate revisions by 2 users not shown)
Line 1: Line 1:
The Power Architecture ISA includes a specification of vector or SIMD operations. Prior to the Power ISA, i.e. PowerPC, some of these operations were available, but defined in an external standard, called Altivec by Motorola (and then Freescale who bought them), VMX by IBM, and Velocity Engine by Apple.
+
The Power Architecture ISA includes a specification of vector or SIMD operations. Prior to the Power ISA, i.e. PowerPC, some of these operations were available, but defined in an external standard, called Altivec by Freescale (Motorola spin-off), Vector Multimedia Extension (VMX) by IBM, and Velocity Engine by Apple.
  
The Vector operations are classified as Vector Multimedia Extension (VMX) and Vector Scalar Extension (VSX) in current versions of the Power ISA.
+
The Vector operations are classified as Vector Facility and Vector Scalar Extension (VSX) in current versions of the Power ISA.
  
 
Power ISA v2.07 still refers to some instructions as VMX in its summary of changes since the previous version, but the rest of the document avoids mentioning VMX completely.
 
Power ISA v2.07 still refers to some instructions as VMX in its summary of changes since the previous version, but the rest of the document avoids mentioning VMX completely.
  
Power ISA v3 no longer mentions VMX at all.
+
Power ISA v3.0 no longer mentions VMX at all.
 +
 
 +
According to [[File:POWER9-Features-and-Specifications.pdf]] page 7, the Vector Scalar Unit (VSU)'s 128-bit hardware is dedicated per super-slice (2 threads).  This may indicate that trying to aggressively use 128-bit VSX instructions in two threads that use the same super-slice will be inefficient.  It is possible that clever usage of <code>taskset</code> may improve this situation.
 +
 
 +
Power ISA v3.1 adds an optional VSX extension, MMA, targeted at matrix math applications.
  
 
== External Links ==
 
== External Links ==
  
* Seebach, Peter. [https://www.ibm.com/developerworks/library/pa-unrollav1/ Unrolling AltiVec, Part 1, Introducing the PowerPC SIMD unit]. Published March 01, 2005
+
* Seebach, Peter. [https://web.archive.org/web/20080614023936/https://www.ibm.com/developerworks/library/pa-unrollav1/ Unrolling AltiVec, Part 1, Introducing the PowerPC SIMD unit]. Published March 01, 2005
* Seebach, Peter. [https://www.ibm.com/developerworks/library/pa-unrollav2/ Unrolling AltiVec, Part 2, Optimize code for SIMD processing]. Published March 16, 2005
+
* Seebach, Peter. [https://www.sucs.org/~grepwood/ps3/altivec/pa-unrollav2-pdf.pdf Unrolling AltiVec, Part 2, Optimize code for SIMD processing]. Published March 16, 2005
* Clarke, Paul. [https://www.ibm.com/developerworks/community/blogs/fe313521-2e95-46f2-817d-44a4f27eba32/entry/vectorizing_for_fun_and_performance?lang=en/ Vectorizing for fun and performance]. Published August 18, 2014
 
* Thomas, Francois. [https://www.ibm.com/developerworks/community/wikis/home?lang=en#!/wiki/W51a7ffcf4dfd_4b40_9d82_446ebc23c550/page/Intel%20SSE%20to%20PowerPC%20AltiVec%20migration/ Intel SSE to PowerPC AltiVec migration]. Published June 01, 2015
 
 
* Gschwind, Michael. [https://www.researchgate.net/publication/299472451_Workload_acceleration_with_the_IBM_POWER_vector-scalar_architecture Workload acceleration with the IBM POWER vector-scalar architecture]. IBM Journal of Research and Development. Published March, 2016
 
* Gschwind, Michael. [https://www.researchgate.net/publication/299472451_Workload_acceleration_with_the_IBM_POWER_vector-scalar_architecture Workload acceleration with the IBM POWER vector-scalar architecture]. IBM Journal of Research and Development. Published March, 2016
* Clarke, Paul. [https://developer.ibm.com/linuxonpower/2018/01/24/porting-x86-vector-intrinsics-code-linux-power-hurry/ Porting x86 vector intrinsics code to Linux on Power in a hurry]. Published January 24, 2018
+
* Clarke, Paul. [https://www.ibm.com/support/pages/vectorizing-fun-and-performance Vectorizing for fun and performance]. Published January 24, 2018
* OpenPOWER. [https://openpowerfoundation.org/?resource_lib=linux-power-porting-guide-vector-intrinsics/ Linux on Power Porting Guide: Vector Intrinsics]. Published March 13, 2018
+
* OpenPOWER. [https://openpowerfoundation.org/specifications/vectorintrinsicportingguide/ Vector Intrinsics Porting Guide]. Published March 13, 2018
 +
* Talospace.  [https://www.talospace.com/2019/07/easier-power-vectorizing-for-fun-and.html Easier Power ISA vectorizing for fun and profit with GCC x86 intrinsics].  Published July 26, 2019
 +
* OpenPOWER. [https://openpowerfoundation.org/specifications/vectorintrinsicprogrammingreference/ Vector Intrinsics Programming Reference Specification]. Published August 11, 2020
 +
* Ruzhu Chen. [https://developer.ibm.com/learningpaths/port-your-app-to-lop/migrate-app-on-lop/ Linux on IBM Power Systems application porting and tuning guide]. Published September 27, 2020
 +
* OpenPOWER. [https://openpowerfoundation.org/compliance/vectorintrinsicprogrammingreference/ Vector Intrinsic Programming Reference Compliance Specification]. Published March 28, 2021
 +
* José E. Moreira. [https://arxiv.org/abs/2104.03142 A matrix math facility for Power ISA(TM) processors]. Published April 7, 2021
 +
* José E. Moreira. [https://www.redbooks.ibm.com/Redbooks.nsf/RedpieceAbstracts/redp5612.html Matrix-Multiply Assist Best Practices Guide]. Published April 15, 2021
 +
* Sridhar Venkat. [https://community.ibm.com/community/user/powerdeveloper/blogs/sridhar-venkat1/2022/08/29/power10-mma MMA in IBM Power10 processor]. Published August 29, 2022
 +
 
 +
== Github / Gitlab pages ==
 +
 
 +
* Eigen. [https://gitlab.com/libeigen/eigen A C++ template library for linear algebra: matrices, vectors, numerical solvers and related algorithms]
 +
* Simd Library. [https://github.com/ermig1979/Simd C++ image processing and machine learning library with using of SIMD]
 +
* SIMD Everywhere. [https://github.com/simd-everywhere/simde Implementations of SIMD instruction sets for systems which don't natively support them]
 +
* EVE - the Expressive Vector Engine. [https://github.com/jfalcou/eve SIMD in C++]
 +
* UniSIMD Assembler. [https://github.com/VectorChief/UniSIMD-assembler SIMD macro assembler unified for ARM, MIPS, PPC and x86]
 +
* Turbo Base64. [https://github.com/powturbo/Turbo-Base64 Fastest Base64 SIMD:SSE/AVX2/AVX512/Neon/Altivec]
 +
* Inastemp. [https://gitlab.inria.fr/bramas/inastemp Intrinsics as template - is a basic library to use vectorization easily in C++]
 +
* libsimdpp. [https://github.com/p12tic/libsimdpp Portable header-only C++ low level SIMD library]
 +
* pveclib. [https://github.com/open-power-sdk/pveclib Power Vector Library]
 +
* SLEEF. [https://github.com/shibatch/sleef SIMD Library for Evaluating Elementary Functions, vectorized libm and DFT]
 +
* libjpeg-turbo. [https://github.com/libjpeg-turbo/libjpeg-turbo A JPEG image codec that uses SIMD instructions to accelerate baseline JPEG compression and decompression]
 +
* libfreevec. [https://github.com/VectorCamp/libfreevec SIMD optimized C library]

Latest revision as of 06:54, 8 November 2023

The Power Architecture ISA includes a specification of vector or SIMD operations. Prior to the Power ISA, i.e. PowerPC, some of these operations were available, but defined in an external standard, called Altivec by Freescale (Motorola spin-off), Vector Multimedia Extension (VMX) by IBM, and Velocity Engine by Apple.

The Vector operations are classified as Vector Facility and Vector Scalar Extension (VSX) in current versions of the Power ISA.

Power ISA v2.07 still refers to some instructions as VMX in its summary of changes since the previous version, but the rest of the document avoids mentioning VMX completely.

Power ISA v3.0 no longer mentions VMX at all.

According to File:POWER9-Features-and-Specifications.pdf page 7, the Vector Scalar Unit (VSU)'s 128-bit hardware is dedicated per super-slice (2 threads). This may indicate that trying to aggressively use 128-bit VSX instructions in two threads that use the same super-slice will be inefficient. It is possible that clever usage of taskset may improve this situation.

Power ISA v3.1 adds an optional VSX extension, MMA, targeted at matrix math applications.

External Links

Github / Gitlab pages