vector-on-power-reference

This repository contains IBM Power based optimizations for vector search based on the vector search functions from FAISS, knowhere and PGVector libraries.

Structure of the repo

• src - Folder contains helper functions, test framework and distance computation kernels. Inside distances folder you will be able to see
  • base - Includes reference code sourced from PGVector database, FAISS and knowhere libraries
  • optimized - Power Optimized version of distance computation kernels using code change leveraging vector data types
  • intrinsic - Intrinsic based optimization - by using IBM specific vector built-in functions and the AltiVec built-in functions with IBM extensions for increased performance
• MakefileAIX - Makefile for building the code in AIX with the IBM Open XL C/C++ compiler
• Makefile - Makefile for building the code in Linux on RHEL with the GCC compiler
• Other license files and README.md

Building the repo in an RHEL environment

1. Use the Makefile for Linux and Run make command in the same folder as the Makefile - This will create an executable test inside bin directory
2. Run ./bin/test -h to see the options to run the test utility
3. Example Run command : ./bin/test -s 32 for testing dimension of 32 - Once you run this, the results will be inside "Results" directory with file names test_result and test_time suffixed with timestamp.

To clean all object files run make clean

Reference for optimizations are provided in two formats:

1. Compiler intrinsic based optimization

   Path: src/distances/intrinsic/
   To trigger intrinsic based optimization code, specify --run_intrinsic_code command line option. For example,

    ./bin/test -s 32 --run_intrinsic_code 
   

2. Code based optimization using vector data types

   Path: src/distances/optimized/
   To trigger source code optimized version, specify --run_optimized_code command line option
   

    ./bin/test -s 32 --run_optimized_code 
   

Building the repo in an AIX environment

Prerequisites : Install make and IBM Clang from AIX toolchain and export their installation path to PATH variable Note: Currently we only support IBM Clang 17.1.2

1. Use the Makefile for AIX and Run gmake -f MakefileAIX command in the same folder as the Makefile - This will create an executable test inside bin directory
2. Follow steps 2 and 3 from RHEL environment commands

Run the executable to test functionality

Running ./bin/test will generate two text files inside Results folder

1. test_results.txt
2. test_time.txt

By default, the test_results.txt file shows the functional behavior of both base and optimized function results and it is expected to have both results same. If you run with --run_optimized_code the functional behavior of both base and intrinsic based implementations are compared.

Sample output - test_results.txt

    Original code execution results.
    Function name   array size
                        32
    Euclidean:
    fvec_L2sqr_ref  6.11873e+06

    PowerPC code execution results.
    Function name   array size
                        32
    Euclidean:
    fvec_L2sqr_ref  6.11873e+06

This indicates that we have run Euclidean fvec_L2sqr_ref function with dimension 32. We have results of original code execution v/s PowerPC(PPC) code execution listed.

The test_time.txt shows the execution time differences between base and optimized functions or between base and intrinsic based on the flags used.

Sample output - test_time.txt

    Original code execution time in ns.
    Function name 	 array size
                         32	
    Euclidean:
    fvec_L2sqr_ref        219456057

    Optimized PowerPC code execution time in ns.
    Function name 	 array size
                         32	       
    Euclidean:
    fvec_L2sqr_ref_ppc    86881094

    Percentage of optimized PPC execution time versus original time.

    Function name 	 array size
                         32	
    Euclidean:
    fvec_L2sqr_ref_ppc    39.59%  (2.53 times)

This files shows the time taken of original code execution v/s PowerPC code execution compared. The results shows that PPC based optimized code takes only 39.59% or 2.53X of the original code execution time.

Notes

1. The intrinsic and optimized versions of the Hamming distance test make use of the overloaded PowerPC builtin vec_popcnt(). The vec_popcnt() is only supported on Power 8 and newer CPUS. If you wish to compile and/or run on Power 7, you must set the #define VEC_POPCNT_SUPPORTED in main-supported.h to 0 to get the code to compile. Setting VEC_POPCNT_SUPPORTED to 0 will revert the optimized Hamming distance code to the base version. Additionally the -mcpu=pwr7 flag will be needed in the makefile to compile for Power7.
2. You can add all the compilation flags required in Makefile including -mcpu and -mvsx depending on the use case.