Add threading support for Apple Accelerate

src/threading.c

@@ -31,6 +31,15 @@
  } MKLVersion;
 
 
+/* Apple Accelerate doesn't allow setting the number of threads directly, it only has an
+ * option to do single-threaded or multi-threaded. That is controlled via the BLASSetThreading
+ * API introduced in macOS 15.
+ *
+ * These constants are from the vecLib.framework/Headers/thread_api.h file
+ */
+#define ACCELERATE_BLAS_THREADING_MULTI_THREADED    0
+#define ACCELERATE_BLAS_THREADING_SINGLE_THREADED   1
+
 /*
  * We provide a flexible thread getter/setter interface here; by calling `lbt_set_num_threads()`
  * libblastrampoline will propagate the call through to its loaded libraries as long as the
@@ -50,6 +59,7 @@ static char * getter_names[MAX_THREADING_NAMES] = {
     "nvpl_lapack_get_max_threads",
     // We special-case MKL in the lookup loop below
     //"MKL_Domain_Get_Max_Threads",
+    // We special-case Apple Accelerate below
     NULL
 };
 
@@ -60,6 +70,7 @@ static char * setter_names[MAX_THREADING_NAMES] = {
     "nvpl_lapack_set_num_threads",
     // We special-case MKL in the lookup loop below
     //"MKL_Domain_Set_Num_Threads",
+    // We special-case Apple Accelerate below
     NULL
 };
 
@@ -129,6 +140,37 @@ LBT_DLLEXPORT int32_t lbt_get_num_threads() {
                 }
             }
         }
+
+        // Special case Apple Accelerate because we have to determine if we are single-threaded or multi-threaded
+        // This API only exists on macOS 15+.
+        int (*fptr_acc)(void) = lookup_symbol(lib->handle, "BLASGetThreading");
+        if (fptr_acc != NULL) {
+            int nthreads = fptr_acc();
+
+            if(nthreads == ACCELERATE_BLAS_THREADING_MULTI_THREADED) {
+                int (*fptr_acc_nthreads)(void) = lookup_symbol(lib->handle, "APPLE_NTHREADS");
+                if (fptr_acc != NULL) {
+                    // In Accelerate, there is a symbol called APPLE_NTHREADS, which appears to be a function we
+                    // can call to get an integer saying the number of CPU threads. There is no documentation for this
+                    // anywhere accessible online, but testing two different CPUs seem to suggest it is CPU cores.
+                    //
+                    // Doing this:
+                    //     julia> @ccall AppleAccelerate.libacc.APPLE_NTHREADS()::Int
+                    //
+                    // The M2 Max returned 12, M4 Max returned 16, which is the total number of cores (both big and little)
+                    // in each processor.
+                    int nthreads = fptr_acc_nthreads();
+                    max_threads = max(max_threads, nthreads);
+                } else {
+                    // This number is arbitrary because we have no idea how many threads are actually in use,
+                    // but greater than 1 to mean multi-threaded.
+                    max_threads = max(max_threads, 2);
+                }
+            } else {
+                // Single-threaded
+                max_threads = max(max_threads, 1);
+            }
+        }
     }
     return max_threads;
 }
@@ -157,5 +199,16 @@ LBT_DLLEXPORT void lbt_set_num_threads(int32_t nthreads) {
             fptr(nthreads, MKL_DOMAIN_BLAS);
             fptr(nthreads, MKL_DOMAIN_LAPACK);
         }
+
+        // Special case Apple Accelerate because we have to determine if we must set multi-threaded or single-threaded
+        // This API only exists on macOS 15+.
+        int (*fptr_acc)(int) = lookup_symbol(lib->handle, "BLASSetThreading");
+        if (fptr_acc != NULL) {
+            if(nthreads > 1) {
+                fptr_acc(ACCELERATE_BLAS_THREADING_MULTI_THREADED);
+            } else {
+                fptr_acc(ACCELERATE_BLAS_THREADING_SINGLE_THREADED);
+            }
+        }
     }
 }

test/accelerate.jl

@@ -0,0 +1,93 @@
+using Libdl, Test
+
+# Taken from AppleAccelerate.jl to avoid a dependency on it
+const libacc = "/System/Library/Frameworks/Accelerate.framework/Accelerate"
+const libacc_info_plist = "/System/Library/Frameworks/Accelerate.framework/Versions/Current/Resources/Info.plist"
+
+function get_macos_version(normalize=true)
+    @static if !Sys.isapple()
+        return nothing
+    end
+
+    plist_lines = split(String(read("/System/Library/CoreServices/SystemVersion.plist")), "\n")
+    vers_idx = findfirst(l -> occursin("ProductVersion", l), plist_lines)
+    if vers_idx === nothing
+        return nothing
+    end
+
+    m = match(r">([\d\.]+)<", plist_lines[vers_idx+1])
+    if m === nothing
+        return nothing
+    end
+
+    ver = VersionNumber(only(m.captures))
+    if normalize && ver.major == 16
+        return VersionNumber(26, ver.minor, ver.patch)
+    end
+    return ver
+end
+
+
+# Load the Accelerate library
+libacc_handle = dlopen(libacc)
+@testset "Accelerate ILP64 loading" begin
+    # ILP64 requires macOS 13.3+
+    if get_macos_version() >= v"13.3"
+        # Load the ILP64 interface
+        lbt_forward(lbt_handle, libacc; clear=true, suffix_hint="\x1a\$NEWLAPACK\$ILP64")
+
+        # Test that we have only one library loaded
+        config = lbt_get_config(lbt_handle)
+        libs = unpack_loaded_libraries(config)
+        @test length(libs) == 1
+
+        # Test that it's Accelerate and it's correctly identified
+        @test libs[1].libname == libacc
+        @test libs[1].interface == LBT_INTERFACE_ILP64
+
+        # Test that `dgemm` forwards to `dgemm_` within the Accelerate library
+        acc_dgemm = dlsym(libacc_handle, "dgemm\$NEWLAPACK\$ILP64")
+        @test lbt_get_forward(lbt_handle, "dgemm_", LBT_INTERFACE_ILP64) == acc_dgemm
+    end
+end
+
+@testset "Accelerate LP64 loading" begin
+    # New LAPACK interface requires macOS 13.3+
+    if get_macos_version() >= v"13.3"
+        # Load the LP64 interface
+        lbt_forward(lbt_handle, libacc; clear=true, suffix_hint="\x1a\$NEWLAPACK")
+
+        # Test that we have only one library loaded
+        config = lbt_get_config(lbt_handle)
+        libs = unpack_loaded_libraries(config)
+        @test length(libs) == 1
+
+        # Test that it's Accelerate and it's correctly identified
+        @test libs[1].libname == libacc
+        @test libs[1].interface == LBT_INTERFACE_LP64
+
+        # Test that `dgemm` forwards to `dgemm_` within the Accelerate library
+        acc_dgemm = dlsym(libacc_handle, "dgemm\$NEWLAPACK")
+        @test lbt_get_forward(lbt_handle, "dgemm_", LBT_INTERFACE_LP64) == acc_dgemm
+    end
+end
+
+@testset "Accelerate threading" begin
+    # This threading API will only work on v15 and above
+    if get_macos_version() >= v"15"
+        lbt_forward(lbt_handle, libacc; clear=true)
+
+        # Set to single-threaded
+        lbt_set_num_threads(lbt_handle, 1)
+        @test lbt_get_num_threads(lbt_handle) == 1
+
+        # Set to multi-threaded
+        # Accelerate doesn't actually let us say how many threads, so we must test for greater than
+        lbt_set_num_threads(lbt_handle, 2)
+        @test lbt_get_num_threads(lbt_handle) > 1
+
+        # Set back to single-threaded
+        lbt_set_num_threads(lbt_handle, 1)
+        @test lbt_get_num_threads(lbt_handle) == 1
+    end
+end

test/runtests.jl

@@ -255,3 +255,8 @@ end
 
 # Run our "direct" tests within Julia
 include("direct.jl")
+
+# Run some Apple Accelerate tests, but only on Apple
+@static if Sys.isapple()
+    include("accelerate.jl")
+end