Skip to content

Latest commit

 

History

History
387 lines (338 loc) · 21 KB

PerformanceSmall.md

File metadata and controls

387 lines (338 loc) · 21 KB

Contents

Introduction

This document showcases performance results for the level-3 gemm operation on small matrices with BLIS and BLAS for select hardware architectures.

General information

Generally speaking, for level-3 operations on small matrices, we publish two "panels" for each type of hardware, one that reflects performance on row-stored matrices and another for column-stored matrices. Each panel will consist of a 4x7 grid of graphs, with each row representing a different transposition case (nn, nt, tn, tt) complex) and each column representing a different shape scenario, usually with one or two matrix dimensions bound to a fixed size for all problem sizes tested. Each of the 28 graphs within a panel will contain an x-axis that reports problem size, with one, two, or all three matrix dimensions equal to the problem size (e.g. m = 6; n = k, also encoded as m6npkp). The y-axis will report in units GFLOPS (or billions of floating-point operations per second) per core.

It's also worth pointing out that the top of some graphs (e.g. the maximum y-axis value depicted) correspond to the theoretical peak performance under the conditions associated with that graph, while in other graphs the y-axis has been adjusted to better show the difference between the various curves. (We strongly prefer to always use peak performance as the top of the graph; however, this is one of the few exceptions where we feel some scaling is warranted.) Theoretical peak performance on a single core, in units of GFLOPS, is calculated as the product of:

  1. the maximum sustainable clock rate in GHz; and
  2. the maximum number of floating-point operations (flops) that can be executed per cycle.

Note that the maximum sustainable clock rate may change depending on the conditions. For example, on some systems the maximum clock rate is higher when only one core is active (e.g. single-threaded performance) versus when all cores are active (e.g. multithreaded performance). The maximum number of flops executable per cycle (per core) is generally computed as the product of:

  1. the maximum number of fused multiply-add (FMA) vector instructions that can be issued per cycle (per core);
  2. the maximum number of elements that can be stored within a single vector register (for the datatype in question); and
  3. 2.0, since an FMA instruction fuses two operations (a multiply and an add).

Typically, organizations and individuals publish performance with square matrices, which can miss the problem sizes of interest to many applications. Here, in addition to square matrices (shown in the seventh column), we also show six other scenarios where one or two gemm dimensions (of m, n, and k) is small. In these six columns, the constant small matrix dimensions were chosen to be very small--in the neighborhood of 8--intentionally to showcase what happens when at least one of the matrices is abnormally "skinny."

The problem size range, represented on the x-axis, is sampled in increments that vary. These increments (and the overall range) are generally large for the cases where two dimensions are small (and constant), medium for cases where one dimension is small (and constant), and small for cases where all dimensions (e.g. m, n, and k) are variable and bound to the problem size (i.e., square matrices).

The legend in each graph contains two entries for BLIS, corresponding to the two black lines, one solid and one dotted. The dotted line, "BLIS conv", represents the conventional implementation that targets large matrices. This was the only implementation available in BLIS prior to the addition to the small/skinny matrix support. The solid line, "BLIS sup", makes use of the new small/skinny matrix implementation. Sometimes, the performance of "BLIS sup" drops below that of "BLIS conv" for somewhat larger problems. However, in practice, we use a threshold to determine when to switch from the former to the latter, and therefore the goal is for the performance of "BLIS conv" to serve as an approximate floor below which BLIS performance never drops.

Finally, each point along each curve represents the best of three trials.

Interpretation

In general, the the curves associated with higher-performing implementations will appear higher in the graphs than lower-performing implementations. Ideally, an implementation will climb in performance (as a function of problem size) as quickly as possible and asymptotically approach some high fraction of peak performance.

When corresponding with us, via email or when opening an issue on github, we kindly ask that you specify as closely as possible (though a range is fine) your problem size of interest so that we can better assist you.

Reproduction

In general, we do not offer any step-by-step guide for how to reproduce the performance graphs shown below.

That said, if you are keenly interested in running your own performance benchmarks, either in an attempt to reproduce the results shown here or to measure performance of different hardware, of different implementations (or versions), and/or for different problem sizes, you should begin by studying the source code, Makefile, and scripts in the test/sup directory of the BLIS source distribution. Then, you'll need to take time to build and/or install some (or all) of the implementations shown (e.g. OpenBLAS, MKL, Eigen, BLASFEO, and libxsmm), including BLIS. Be sure to consult the detailed notes provided below; they should be very helpful in successfully building the libraries. The runme.sh script in test/sup (or test/supmt) will help you run some (or all) of the test drivers produced by the Makefile, and the Matlab/Octave function plot_panel_trxsh() defined in the octave directory will help you turn the output of those test drivers into a PDF file of graphs. The runthese.m file will contain example invocations of the function.

Level-3 performance

Kaby Lake

Kaby Lake experiment details

  • Location: undisclosed
  • Processor model: Intel Core i5-7500 (Kaby Lake)
  • Core topology: one socket, 4 cores total
  • SMT status: unavailable
  • Max clock rate: 3.8GHz (single-core)
  • Max vector register length: 256 bits (AVX2)
  • Max FMA vector IPC: 2
  • Peak performance:
    • single-core: 57.6 GFLOPS (double-precision), 115.2 GFLOPS (single-precision)
    • multicore: 57.6 GFLOPS/core (double-precision), 115.2 GFLOPS/core (single-precision)
  • Operating system: Gentoo Linux (Linux kernel 5.2.4)
  • Page size: 4096 bytes
  • Compiler: gcc 8.3.0
  • Results gathered: 3 March 2020
  • Implementations tested:
    • BLIS 90db88e (0.6.1-8)
      • configured with ./configure --enable-cblas auto (single-threaded)
      • configured with ./configure --enable-cblas -t openmp auto (multithreaded)
      • sub-configuration exercised: haswell
      • Multithreaded (4 cores) execution requested via export BLIS_NUM_THREADS=4
    • OpenBLAS 0.3.8
      • configured Makefile.rule with BINARY=64 NO_LAPACK=1 NO_LAPACKE=1 USE_THREAD=0 USE_LOCKING=1 (single-threaded)
      • configured Makefile.rule with BINARY=64 NO_LAPACK=1 NO_LAPACKE=1 USE_THREAD=1 NUM_THREADS=4 (multithreaded)
      • Multithreaded (4 cores) execution requested via export OPENBLAS_NUM_THREADS=4
    • BLASFEO f9b78c6
      • configured Makefile.rule with: BLAS_API=1 FORTRAN_BLAS_API=1 CBLAS_API=1.
    • Eigen 3.3.90
      • Obtained via the Eigen git mirror (36b9596)
      • Prior to compilation, modified top-level CMakeLists.txt to ensure that -march=native was added to CXX_FLAGS variable (h/t Sameer Agarwal):
        # These lines added after line 67.
        check_cxx_compiler_flag("-march=native" COMPILER_SUPPORTS_MARCH_NATIVE)
        if(COMPILER_SUPPORTS_MARCH_NATIVE)
          set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
        endif()
        
      • configured and built BLAS library via mkdir build; cd build; CC=gcc cmake ..; make blas
      • installed headers via cmake . -DCMAKE_INSTALL_PREFIX=$HOME/flame/eigen; make install
      • The gemm implementation was pulled in at compile-time via Eigen headers; other operations were linked to Eigen's BLAS library.
      • Single-threaded (1 core) execution requested via export OMP_NUM_THREADS=1
      • Multithreaded (4 cores) execution requested via export OMP_NUM_THREADS=4
    • MKL 2020 initial release
      • Single-threaded (1 core) execution requested via export MKL_NUM_THREADS=1
      • Multithreaded (4 cores) execution requested via export MKL_NUM_THREADS=4
    • libxsmm a40a833 (post-1.14)
      • compiled with make AVX=2; linked with netlib BLAS 3.6.0 as the fallback library to better show where libxsmm stops handling the computation internally.
  • Affinity:
    • Thread affinity for BLIS was specified manually via GOMP_CPU_AFFINITY="0-3". However, multithreaded OpenBLAS appears to revert to single-threaded execution if GOMP_CPU_AFFINITY is set. Therefore, when measuring OpenBLAS performance, the GOMP_CPU_AFFINITY environment variable was unset.
  • Frequency throttling (via cpupower):
    • Driver: intel_pstate
    • Governor: performance
    • Hardware limits: 800MHz - 3.8GHz
    • Adjusted minimum: 3.8GHz
  • Comments:
    • libxsmm is highly competitive for very small problems, but quickly gives up once the "large" dimension exceeds about 180-240 (or 64 in the case where all operands are square). Also, libxsmm's gemm cannot handle a transposition on matrix A and similarly dispatches the fallback implementation for those cases. libxsmm also does not export CBLAS interfaces, and therefore only appears on the graphs for column-stored matrices.

Kaby Lake results

pdf

png (inline)

  • Kaby Lake single-threaded row-stored single-threaded row-stored
  • Kaby Lake single-threaded column-stored single-threaded column-stored
  • Kaby Lake multithreaded (4 cores) row-stored multithreaded row-stored
  • Kaby Lake multithreaded (4 cores) column-stored multithreaded column-stored

Haswell

Haswell experiment details

  • Location: TACC (Lonestar5)
  • Processor model: Intel Xeon E5-2690 v3 (Haswell)
  • Core topology: two sockets, 12 cores per socket, 24 cores total
  • SMT status: enabled, but not utilized
  • Max clock rate: 3.5GHz (single-core), 3.1GHz (multicore)
  • Max vector register length: 256 bits (AVX2)
  • Max FMA vector IPC: 2
  • Peak performance:
    • single-core: 56 GFLOPS (double-precision), 112 GFLOPS (single-precision)
    • multicore: 49.6 GFLOPS/core (double-precision), 99.2 GFLOPS/core (single-precision)
  • Operating system: Cray Linux Environment 6 (Linux kernel 4.4.103)
  • Page size: 4096 bytes
  • Compiler: gcc 7.3.0
  • Results gathered: 3 March 2020
  • Implementations tested:
    • BLIS 90db88e (0.6.1-8)
      • configured with ./configure --enable-cblas auto (single-threaded)
      • configured with ./configure --enable-cblas -t openmp auto (multithreaded)
      • sub-configuration exercised: haswell
      • Multithreaded (12 cores) execution requested via export BLIS_NUM_THREADS=12
    • OpenBLAS 0.3.8
      • configured Makefile.rule with BINARY=64 NO_LAPACK=1 NO_LAPACKE=1 USE_THREAD=0 USE_LOCKING=1 (single-threaded)
      • configured Makefile.rule with BINARY=64 NO_LAPACK=1 NO_LAPACKE=1 USE_THREAD=1 NUM_THREADS=12 (multithreaded)
      • Multithreaded (12 cores) execution requested via export OPENBLAS_NUM_THREADS=12
    • BLASFEO f9b78c6
      • configured Makefile.rule with: BLAS_API=1 FORTRAN_BLAS_API=1 CBLAS_API=1.
    • Eigen 3.3.90
      • Obtained via the Eigen git mirror (36b9596)
      • Prior to compilation, modified top-level CMakeLists.txt to ensure that -march=native was added to CXX_FLAGS variable (h/t Sameer Agarwal):
        # These lines added after line 67.
        check_cxx_compiler_flag("-march=native" COMPILER_SUPPORTS_MARCH_NATIVE)
        if(COMPILER_SUPPORTS_MARCH_NATIVE)
          set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
        endif()
        
      • configured and built BLAS library via mkdir build; cd build; CC=gcc cmake ..; make blas
      • installed headers via cmake . -DCMAKE_INSTALL_PREFIX=$HOME/flame/eigen; make install
      • The gemm implementation was pulled in at compile-time via Eigen headers; other operations were linked to Eigen's BLAS library.
      • Single-threaded (1 core) execution requested via export OMP_NUM_THREADS=1
      • Multithreaded (12 cores) execution requested via export OMP_NUM_THREADS=12
    • MKL 2020 initial release
      • Single-threaded (1 core) execution requested via export MKL_NUM_THREADS=1
      • Multithreaded (12 cores) execution requested via export MKL_NUM_THREADS=12
    • libxsmm a40a833 (post-1.14)
      • compiled with make AVX=2; linked with netlib BLAS 3.6.0 as the fallback library to better show where libxsmm stops handling the computation internally.
  • Affinity:
    • Thread affinity for BLIS was specified manually via GOMP_CPU_AFFINITY="0-11". However, multithreaded OpenBLAS appears to revert to single-threaded execution if GOMP_CPU_AFFINITY is set. Therefore, when measuring OpenBLAS performance, the GOMP_CPU_AFFINITY environment variable was unset.
  • Frequency throttling (via cpupower):
    • No changes made.
  • Comments:
    • libxsmm is highly competitive for very small problems, but quickly gives up once the "large" dimension exceeds about 180-240 (or 64 in the case where all operands are square). Also, libxsmm's gemm cannot handle a transposition on matrix A and similarly dispatches the fallback implementation for those cases. libxsmm also does not export CBLAS interfaces, and therefore only appears on the graphs for column-stored matrices.

Haswell results

pdf

png (inline)

  • Haswell single-threaded row-stored single-threaded row-stored
  • Haswell single-threaded column-stored single-threaded column-stored
  • Haswell multithreaded (12 cores) row-stored multithreaded row-stored
  • Haswell multithreaded (12 cores) column-stored multithreaded column-stored

Epyc

Epyc experiment details

  • Location: Oracle cloud
  • Processor model: AMD Epyc 7551 (Zen1)
  • Core topology: two sockets, 4 dies per socket, 2 core complexes (CCX) per die, 4 cores per CCX, 64 cores total
  • SMT status: enabled, but not utilized
  • Max clock rate: 3.0GHz (single-core), 2.55GHz (multicore)
  • Max vector register length: 256 bits (AVX2)
  • Max FMA vector IPC: 1
    • Alternatively, FMA vector IPC is 2 when vectors are limited to 128 bits each.
  • Peak performance:
    • single-core: 24 GFLOPS (double-precision), 48 GFLOPS (single-precision)
    • multicore: 20.4 GFLOPS/core (double-precision), 40.8 GFLOPS/core (single-precision)
  • Operating system: Ubuntu 18.04 (Linux kernel 4.15.0)
  • Page size: 4096 bytes
  • Compiler: gcc 7.4.0
  • Results gathered: 3 March 2020
  • Implementations tested:
    • BLIS 90db88e (0.6.1-8)
      • configured with ./configure --enable-cblas auto (single-threaded)
      • configured with ./configure --enable-cblas -t openmp auto (multithreaded)
      • sub-configuration exercised: haswell
      • Multithreaded (32 cores) execution requested via export BLIS_NUM_THREADS=32
    • OpenBLAS 0.3.8
      • configured Makefile.rule with BINARY=64 NO_LAPACK=1 NO_LAPACKE=1 USE_THREAD=0 USE_LOCKING=1 (single-threaded)
      • configured Makefile.rule with BINARY=64 NO_LAPACK=1 NO_LAPACKE=1 USE_THREAD=1 NUM_THREADS=32 (multithreaded)
      • Multithreaded (32 cores) execution requested via export OPENBLAS_NUM_THREADS=32
    • BLASFEO f9b78c6
      • configured Makefile.rule with: BLAS_API=1 FORTRAN_BLAS_API=1 CBLAS_API=1.
      • built BLAS library via make CC=gcc
    • Eigen 3.3.90
      • Obtained via the Eigen git mirror (36b9596)
      • Prior to compilation, modified top-level CMakeLists.txt to ensure that -march=native was added to CXX_FLAGS variable (h/t Sameer Agarwal):
        # These lines added after line 67.
        check_cxx_compiler_flag("-march=native" COMPILER_SUPPORTS_MARCH_NATIVE)
        if(COMPILER_SUPPORTS_MARCH_NATIVE)
          set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
        endif()
        
      • configured and built BLAS library via mkdir build; cd build; CC=gcc cmake ..; make blas
      • installed headers via cmake . -DCMAKE_INSTALL_PREFIX=$HOME/flame/eigen; make install
      • The gemm implementation was pulled in at compile-time via Eigen headers; other operations were linked to Eigen's BLAS library.
      • Single-threaded (1 core) execution requested via export OMP_NUM_THREADS=1
      • Multithreaded (32 cores) execution requested via export OMP_NUM_THREADS=32
    • MKL 2020 initial release
      • Single-threaded (1 core) execution requested via export MKL_NUM_THREADS=1
      • Multithreaded (32 cores) execution requested via export MKL_NUM_THREADS=32
    • libxsmm a40a833 (post-1.14)
      • compiled with make AVX=2; linked with netlib BLAS 3.6.0 as the fallback library to better show where libxsmm stops handling the computation internally.
  • Affinity:
    • Thread affinity for BLIS was specified manually via GOMP_CPU_AFFINITY="0-31". However, multithreaded OpenBLAS appears to revert to single-threaded execution if GOMP_CPU_AFFINITY is set. Therefore, when measuring OpenBLAS performance, the GOMP_CPU_AFFINITY environment variable was unset.
  • Frequency throttling (via cpupower):
    • Driver: acpi-cpufreq
    • Governor: performance
    • Hardware limits: 1.2GHz - 2.0GHz
    • Adjusted minimum: 2.0GHz
  • Comments:
    • libxsmm is highly competitive for very small problems, but quickly gives up once the "large" dimension exceeds about 180-240 (or 64 in the case where all operands are square). Also, libxsmm's gemm cannot handle a transposition on matrix A and similarly dispatches the fallback implementation for those cases. libxsmm also does not export CBLAS interfaces, and therefore only appears on the graphs for column-stored matrices.

Epyc results

pdf

png (inline)

  • Epyc single-threaded row-stored single-threaded row-stored
  • Epyc single-threaded column-stored single-threaded column-stored
  • Epyc multithreaded (32 cores) row-stored multithreaded row-stored
  • Epyc multithreaded (32 cores) column-stored multithreaded column-stored

Feedback

Please let us know what you think of these performance results! Similarly, if you have any questions or concerns, or are interested in reproducing these performance experiments on your own hardware, we invite you to open an issue and start a conversation with BLIS developers.

Thanks for your interest in BLIS!