Use Apple Accelerate as BLAS/LAPACK backend on macOS

[Ionizing]

Summary

Apple shipps Accelerate Framework with macOS, which provides BLAS/LAPACK interface to accelerate related routines. I tested it and saw at least 1x improvement compared to OpenBLAS on my MacBook Air 2025 with M4 10cores.

So I wonder if there can be some improvements on the CMake configuration files, so that it can utilize Apple Accelerate for better performance on nda on macOS.

Motivation

Use Apple Accelerate to gain higher performance.

CMake commands build libnda_c and libh5

Use OpenBLAS as backend

cmake .. -DCMAKE_INSTALL_PREFIX=$HOME/apps/nda_openblas \
    -DCMAKE_BUILD_TYPE="Release" \
    -DCMAKE_C_COMPILER="gcc-14" \
    -DCMAKE_CXX_COMPILER="g++-14"

make -j4 && make install

There should be some message reads "Can't build with Accelerate Framework on OSX, trying to find OpenBLAS instead" in the log:

-- -------- Lapack detection -------------
-- Can't build with Accelerate Framework on OSX, trying to find OpenBLAS instead...
-- Looking for sgemm_
-- Looking for sgemm_ - found
-- Found BLAS: /opt/homebrew/opt/openblas/lib/libopenblas.dylib
-- Looking for cheev_
-- Looking for cheev_ - found
-- Found LAPACK: /opt/homebrew/opt/openblas/lib/libopenblas.dylib;/opt/homebrew/opt/openblas/lib/libopenblas.dylib

Use Apple Accelerate as backend

cmake .. -DCMAKE_INSTALL_PREFIX=$HOME/apps/nda_accelerate \
    -DCMAKE_BUILD_TYPE="Release" \
    -DCMAKE_C_COMPILER="gcc-14" \
    -DCMAKE_CXX_COMPILER="g++-14" \
    -DBLAS_LIBRARIES="/Library/Developer/CommandLineTools/SDKs/MacOSX15.2.sdk/System/Library/Frameworks/Accelerate.framework" \
    -DLAPACK_LIBRARIES="/Library/Developer/CommandLineTools/SDKs/MacOSX15.2.sdk/System/Library/Frameworks/Accelerate.framework"

make -j4 && make install

Then cmake should be able to find BLAS and LAPACK inside Accelerate Framework, and the log reads

-- -------- Lapack detection -------------
-- Can't build with Accelerate Framework on OSX, trying to find OpenBLAS instead...
-- Found BLAS: /Library/Developer/CommandLineTools/SDKs/MacOSX15.2.sdk/System/Library/Frameworks/Accelerate.framework
-- Found LAPACK: /Library/Developer/CommandLineTools/SDKs/MacOSX15.2.sdk/System/Library/Frameworks/Accelerate.framework

Test code

Here is the test source code main.cpp

#include <iostream>
#include <chrono>
#include <nda/nda.hpp>
#include <nda/h5.hpp>

using namespace nda;
using namespace std::literals;

const int N = 100;

int main() {
    array<double, 2> A = rand(1000, 1000);
    array<double, 2> B = rand(1000, 1000);

    auto t0 = std::chrono::steady_clock::now();
    for (int i=0; i!=N; ++i) {
        array<double, 2> C = matmul(A, B);
    }
    auto t1 = std::chrono::steady_clock::now();
    auto dt = (t1 - t0) * 1.0 / N;

    std::cout << "dt = " << dt/1.0ms << "ms" << std::endl;

    return 0;
}

The CMakeLists.txt:

cmake_minimum_required(VERSION 3.20)
project(nda_test CXX)

# set required standard
set(CMAKE_BUILD_TYPE Release)
set(CMAKE_CXX_STANDARD 20)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)

# find nda
find_package(nda REQUIRED CONFIG)

# build the example
add_executable(ex main.cpp)
target_link_libraries(ex nda::nda_c)

Build and test commands:

#!/bin/bash

source ~/apps/nda_openblas/share/nda/ndavars.sh
# source ~/apps/nda_accelerate/share/nda/ndavars.sh

cmake .. \
    -DCMAKE_BUILD_TYPE="Release" \
    -DCMAKE_CXX_COMPILER="g++-14"

make && ./ex

Result

I only performed the matrix multiplication with two 1000x1000 random matrices, so the benchmark is not that sufficient.

BLAS Vendor	Time used (ms)
OpenBLAS	12.488
Apple Accelerate	5.16634

[HugoStrand]

Dear @Ionizing,

I am happily using Apple Accelerate with triqs/nda using the vecLibFort that fixes full GNU Fortran compatibility of Apple's vecLib, see https://github.com/mcg1969/vecLibFort.

Could you please have a go and try it out? (there are available brew and MacPorts packages)

Cheers, Hugo

[Ionizing]

Hi @HugoStrand ,

Thanks for your advice. I tried it out and it worked well on my machine. Would it be possible to add some hints on how to integrate "Apple Accelerate" or "vecLibFort" with nda to the "Installation" page ? User may have to spend considerable time on making it work.

Best regards, Ionizing

[the-hampel]

Hi,
vecLibFort looks really cool. I gave it a try and it seems a nice drop-in replacement for lapack/blas support via accelerate. At least compiling linking works nicely!

However, I played around with this a bit and I have somehow the same problem that I also encountered with accelerate itself: for certain operations it is not really accurate and leads to wrong results compared to openblas on MacOS. I am using the fortran interface, so maybe in C this is better, but I think I will stay away from it for now since some tests in which I use gradient like approaches, i.e. computing eigenvectors etc fail horribly. But I am not sure why that is. Maybe others can share their experience as well. I am just saying that one should test carefully if this gives correct results. Maybe for triqs this works well enough?

[HugoStrand]

Dear @the-hampel,

I have no problems passing the triqs test suite using vecLibFort. I guess if you are using (Vasp? and) Fortran, depending on very legacy behaviour, you might still get in trouble.

Cheers, H

[the-hampel]

That is good to hear @HugoStrand . I asked @Wentzell once and he said he also had some trouble? But if the triqs test and I guess also tprf pass I think that should be relatively safe to use! Indeed, VASP for example has big trouble somehow. I could not debug why exactly.

[Thoemi09]

Hi all,

I had problems as well in the past. I can build and pass the tests in nda with vecLibFort (without it, some tests throw an exception). However, I can't pass the tests in TRIQS.

@HugoStrand Can you tell me how you use vecLibFort? I might be doing something wrong.

[HugoStrand]

Dear @Thoemi09,

Sorry for the silence. I am not doing anything special just specifying the BLAS and LAPACK libraries explicitly using the flags

    -DBLAS_LIBRARIES="/opt/local/lib/libvecLibFort.dylib" \
    -DLAPACK_LIBRARIES="/opt/local/lib/libvecLibFort.dylib" \

in the call to cmake.

Cheers, H