Retire legacy-Accelerate accommodations: macOS runs the full library

.github/workflows/core-macos.yaml

@@ -26,14 +26,22 @@ jobs:
           set -x
           brew install googletest
 
-      - name: install BLAS++
+      # macOS builds use Apple's NEW Accelerate interface (LAPACK 3.12 on
+      # current SDKs) -- the legacy default is LAPACK 3.2.1 from 2009, with a
+      # broken divide-and-conquer gesdd and none of the routines BQRRP/HQRRP
+      # need. The defines.h check makes a silent fallback to the legacy
+      # interface a hard failure. Requires blaspp/lapackpp with new-Accelerate
+      # support (icl-utk-edu masters after the BallisticLA PRs merge).
+      - name: install BLAS++ (new Accelerate)
         run: |
           cd ..
           git clone https://github.com/icl-utk-edu/blaspp.git
           mkdir blaspp-build
           cd blaspp-build
-          cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=`pwd`/../blaspp-install -Dbuild_tests=OFF ../blaspp
+          cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=`pwd`/../blaspp-install -Dblas=accelerate -Dgpu_backend=none -Dbuild_tests=OFF ../blaspp
           make -j2 install
+          grep -q ACCELERATE_NEW_LAPACK ../blaspp-install/include/blas/defines.h \
+            || { echo "::error::BLAS++ did not select Apple's new Accelerate interface"; exit 1; }
 
       - name: install Random123
         run: |

INSTALL.md

@@ -43,6 +43,18 @@ For GPU/CUDA support (enabled with `-DRequireCUDA=ON`), you need:
 **Verified Working Configuration** (as of 2025-11-26):
 - CUDA 12.9.0 + GCC 13.3.0 + CMake 3.31.9 + Driver v581.80 ✅
 
+### macOS: Apple Accelerate requirements
+Apple ships two LAPACK implementations inside Accelerate. The legacy default
+is LAPACK 3.2.1 (2009): its divide-and-conquer SVD (`gesdd`) is broken on
+Apple Silicon, and it lacks routines RandLAPACK requires (`orhr_col`,
+`geqrt`, ...). **RandLAPACK supports only the new interface** (macOS 13.3+,
+LAPACK 3.12 on current SDKs): build BLAS++ with `-Dblas=accelerate` using a
+BLAS++/LAPACK++ version that carries new-Accelerate support, and the correct
+interface propagates to LAPACK++ and RandLAPACK automatically. Verify by
+checking that the installed `include/blas/defines.h` contains
+`ACCELERATE_NEW_LAPACK`. Alternatively, use Homebrew OpenBLAS (what
+`install.sh` does on macOS).
+
 ### Note on Directory Names
 On some systems, library directories are called `lib` while on others they're called `lib64`. Adjust paths accordingly in the CMake configuration commands below.
 

RandLAPACK/drivers/rl_bqrrp.hh

@@ -162,10 +162,6 @@ int BQRRP<T, RNG>::call(
     int64_t* J,
     RandBLAS::RNGState<RNG> &state
 ){
-    #ifdef __APPLE__
-    UNUSED(m); UNUSED(n); UNUSED(A); UNUSED(lda); UNUSED(d_factor); UNUSED(tau); UNUSED(J); UNUSED(state);
-    throw std::runtime_error("BQRRP is not supported when BLAS is linked against Apple Accelerate.");
-    #else
     // Input parameter validation. Bad inputs would otherwise lead to a
     // downstream BLAS/LAPACK failure or, worse, a segfault -- both fatal
     // when BQRRP is called through a binding layer (e.g. MEX/MATLAB).
@@ -660,7 +656,6 @@ int BQRRP<T, RNG>::call(
         cols -= b_sz;
     }
     delete[] J_buffer;
-    #endif
     return 0;
 }
 

RandLAPACK/drivers/rl_cqrrpt.hh

@@ -231,7 +231,6 @@ int CQRRPT<T, RNG>::call(
         hqrrp(d, n, A_hat, d, J, tau, this->nb_alg, this->oversampling, this->panel_pivoting, this->use_cholqr, state, (T**) nullptr);
     } else if(this -> qrcp == Subroutines::QRCP::bqrrp) {
 
-        #if !defined(__APPLE__)
         if (n <= 2000) {
             this->bqrrp_block_ratio = 1.0;
         } else if (n <= 8000) {
@@ -242,7 +241,6 @@ int CQRRPT<T, RNG>::call(
 
         RandLAPACK::BQRRP<T, RNG> BQRRP(false, n * this->bqrrp_block_ratio);
         BQRRP.call(d, n, A_hat, d, 1.0, tau, J, state);
-        #endif
     } else {
         lapack::geqp3(d, n, A_hat, d, J, tau);
     }

RandLAPACK/drivers/rl_hqrrp.hh

@@ -527,12 +527,6 @@ static int64_t CHOLQR_mod_WY(
     //
     // Simplification of NoFLA_QRPmod_WY_unb_var4 for the case when pivoting=0.
     //
-    #if defined(__APPLE__)
-    UNUSED(num_stages); UNUSED(m_A); UNUSED(n_A); UNUSED(buff_A); UNUSED(ldim_A);
-    UNUSED(buff_t); UNUSED(buff_T); UNUSED(ldim_T); UNUSED(buff_R); UNUSED(ldim_R);
-    UNUSED(buff_D);
-    throw std::runtime_error("Unsupported on macOS.");
-    #else
 
     // Some initializations.
     if( num_stages < 0 )
@@ -563,7 +557,6 @@ static int64_t CHOLQR_mod_WY(
     // Entries of tau will be placed on the main diagonal of matrix T from orhr_col().
     for(i = 0; i < n_A; ++i)
         buff_t[i] = buff_T[(ldim_T + 1) * i];
-    #endif
     return 0;
 }
 

benchmark/bench_BQRRP/BQRRP_error_analysis.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 Performs computations in order to assess the pivot quality of BQRRP.
 The setup is described in detail in Section 4 of The arXiv version 2 CQRRPT (https://arxiv.org/pdf/2311.08316.pdf) paper.
@@ -260,4 +253,3 @@ int main(int argc, char *argv[]) {
         j = 0;
     }
 }
-#endif

benchmark/bench_BQRRP/BQRRP_pivot_quality.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 Performs computations in order to assess the pivot quality of BQRRP.
 The setup is described in detail in Section 4 of The arXiv version 2 CQRRPT (https://arxiv.org/pdf/2311.08316.pdf) paper.
@@ -332,4 +325,3 @@ int main(int argc, char *argv[]) {
     sv_ratio(m_info, b_sz, all_data, state_constant2, path);
     std::cout << "Pivot quality metric 2 done\n\n";
 }
-#endif

benchmark/bench_BQRRP/BQRRP_runtime_breakdown.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 IBQRRP runtime breakdown benchmark - assesses the time taken by each subcomponent of IBQRRP.
 There are 10 things that we time:
@@ -182,4 +175,3 @@ int main(int argc, char *argv[]) {
     file << "Total benchmark execution time:" +  std::to_string(dur_time_all) + "\n";
     file.flush();   
 }
-#endif

benchmark/bench_BQRRP/BQRRP_speed_comparisons_block_size.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 BQRRP speed comparison benchmark - runs:
     1. BQRRP_CQR and BQRRP_HQR
@@ -267,4 +260,3 @@ int main(int argc, char *argv[]) {
     file << "Total benchmark execution time:" +  std::to_string(dur_time_all) + "\n";
     file.flush();   
 }
-#endif

benchmark/bench_BQRRP/BQRRP_speed_comparisons_mat_size.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 BQRRP speed comparison benchmark - runs:
     1. BQRRP_CQR and BQRRP_HQR
@@ -279,4 +272,3 @@ int main(int argc, char *argv[]) {
     file << "Total benchmark execution time:" +  std::to_string(dur_time_all) + "\n";
     file.flush();   
 }
-#endif

benchmark/bench_BQRRP/BQRRP_subroutines_speed.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 QR speed comparison benchmark - runs:
     1. GEQRF
@@ -436,4 +429,3 @@ int main(int argc, char *argv[]) {
     file << "Total benchmark execution time:" +  std::to_string(dur_time_all) + "\n";
     file.flush();   
 }
-#endif

benchmark/bench_BQRRP/HQRRP_runtime_breakdown.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 HQRRP runtime breakdown benchmark - assesses the time taken by each subcomponent of HQRRP.
 There are 7 things that we time:
@@ -167,4 +160,3 @@ int main(int argc, char *argv[]) {
     file << "Total benchmark execution time:" +  std::to_string(dur_time_all) + "\n";
     file.flush();   
 }
-#endif

benchmark/bench_BQRRP/HQRRP_sanity_check.cc

@@ -1,12 +1,5 @@
 
 
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 
 // Making sure that HQRRP's performance instability is specific to HQRRP and not related to the flaws in behcnmarking logic
 // by comparing HQRRP and GEMM side-by-side.
@@ -183,4 +176,3 @@ int main(int argc, char *argv[]) {
     file << "Total benchmark execution time:" +  std::to_string(dur_time_all) + "\n";
     file.flush();   
 }
-#endif

benchmark/bench_CQRRPT/CQRRPT_error_analysis.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 Performs computations in order to assess the pivot quality of BQRRP.
 The setup is described in detail in Section 4 of The arXiv version 2 CQRRPT (https://arxiv.org/pdf/2311.08316.pdf) paper.
@@ -255,4 +248,3 @@ int main(int argc, char *argv[]) {
         j = 0;
     }
 }
-#endif

benchmark/bench_CQRRPT/CQRRPT_speed_comparisons.cc

@@ -1,10 +1,3 @@
-#if defined(__APPLE__)
-#include <iostream>
-int main() {
-    std::cout << "This benchmark cannot run on Apple machines." << std::endl;
-    return 1;
-}
-#else
 /*
 CQRRPT speed comparison benchmark - runs:
     1. CQRRPT
@@ -274,4 +267,3 @@ int main(int argc, char *argv[]) {
     file << "Total benchmark execution time:" +  std::to_string(dur_time_all) + "\n";
     file.flush();   
 }
-#endif

build/RandBLAS/RandBLAS/config.h

@@ -0,0 +1,69 @@
+#pragma once
+
+#define RandBLAS_FULL_VERSION "1.1.0-32-g8417f4b"
+#define RandBLAS_VERSION_MAJOR 1
+#define RandBLAS_VERSION_MINOR 1
+#define RandBLAS_VERSION_PATCH 0
+#define RandBLAS_COMMITS_SINCE_RELEASE 32 
+#define RandBLAS_COMMIT_HASH "8417f4b"
+// ^ CMake sets those based on the output of a shell command executed in the RandBLAS git repo.
+//
+//   The specific command is
+//       git describe --tags --match "[0-9]*.[0-9]*.[0-9]*"
+//   That should return a string of the form
+//       [X].[Y].[Z]-[C]-g[H]
+//   The contents of this string are interpreted as follows.
+//       [X] is a nonnegative integer equal to the major version number
+//       [Y] is a nonnegative integer equal to the minor version number
+//       [Z] is a nonnegative integer equal to the patch number
+//       [C] is a nonnegative integer equal to the number of commits since the version number was assigned
+//       [H] is a short-form commit hash, indicating the state of repository's source code.
+//   RandBLAS_FULL_VERSION is the entire string returned by the git command.
+//
+//   If you want to use RandBLAS without using CMake, you need to create config.h manually.
+//   We explain how to do that below. But first
+//
+//      DISCLAIMER: if you're looking at a config.h file actually generated by CMake parsing
+//      this config.h.in file, then the instructions below will mostly look like nonsense.
+//      This is because CMake will have performed string substitution on the original
+//      comments as given in config.h.in.
+//
+//   Start by making a copy of this config.h.in file and giving it the name config.h.
+//   From there, substitute every expression the form  with a value based on the
+//   ``git describe`` command. For example, if ``git describe``returned 0.1.0-456-gcb84abc,
+//   then your config.h file would make the substitutions
+//
+//      1.1.0-32-g8417f4b           -->    0.1.0-456-gcb84abc
+//      1          -->    0
+//      1          -->    1
+//      0          -->    0
+//      32  -->    456
+//      8417f4b            -->    cb84abc
+//
+
+#define RandBLAS_HAS_OpenMP
+// ^ CMake determines whether or not to #define RandBLAS_HAS_OpenMP
+//
+//   If you don't want to use CMake, then your config.h file should
+//   either delete that line (if you aren't linking to OpenMP) or
+//   it should replace it with
+//
+//       #define RandBLAS_HAS_OpenMP
+//
+//   if you are linking to OpenMP.
+//
+
+#define RandBLAS_HAS_MKL
+// ^ CMake determines whether or not to #define RandBLAS_HAS_MKL
+//
+//   This is set when BLAS++ was built with Intel MKL, the MKL sparse
+//   BLAS header (mkl_spblas.h) is found, AND the user has not disabled
+//   MKL sparse features via -DRandBLAS_USE_MKL_SPARSE=OFF.
+//
+//   When defined, RandBLAS uses MKL's Inspector-Executor sparse BLAS
+//   for accelerated sparse matrix operations (sparse x dense and
+//   sparse x sparse multiplication).
+//
+//   If you don't want to use CMake, define this only if you are linking
+//   to Intel MKL and have mkl_spblas.h available.
+//

build/RandLAPACK/rl_config.hh

@@ -0,0 +1,9 @@
+#pragma once
+#include "../RandBLAS/RandBLAS/config.h.in"
+
+#define RandLAPACK_VERSION "0.1.0-162-g54cf3a7"
+#define RandLAPACK_VERSION_MAJOR 0
+#define RandLAPACK_VERSION_MINOR 1
+#define RandLAPACK_VERSION_PATCH 0
+
+#endif