Enable SymbolicSgdLogisticRegression (SymSgdNative) on arm64

Directory.Build.targets

@@ -30,10 +30,10 @@
 		</ItemGroup>
 
     <ItemGroup Condition="'$(TargetArchitecture)' == 'arm64' Or '$(TargetArchitecture)' == 'arm'">
-      <NativeAssemblyReference Remove="MklImports"/>
+      <!-- MklImports (arm shim) and SymSgdNative are built on arm/arm64, so keep
+           them here so they are copied next to the managed assemblies. -->
       <NativeAssemblyReference Remove="CpuMathNative"/>
       <NativeAssemblyReference Remove="FastTreeNative"/>
-      <NativeAssemblyReference Remove="SymSgdNative"/>
       <NativeAssemblyReference Remove="MklProxyNative"/>
       <NativeAssemblyReference Remove="libiomp5md"/>
     </ItemGroup>

src/Microsoft.ML.Mkl.Components/SymSgdClassificationTrainer.cs

@@ -834,8 +834,8 @@ private static unsafe class Native
 
             [DllImport(NativePath), SuppressUnmanagedCodeSecurity]
             private static extern void LearnAll(int totalNumInstances, int* instSizes, int** instIndices,
-                float** instValues, float* labels, bool tuneLR, ref float lr, float l2Const, float piw, float* weightVector, ref float bias,
-                int numFeatres, int numPasses, int numThreads, bool tuneNumLocIter, ref int numLocIter, float tolerance, bool needShuffle, bool shouldInitialize,
+                float** instValues, float* labels, [MarshalAs(UnmanagedType.I1)] bool tuneLR, ref float lr, float l2Const, float piw, float* weightVector, ref float bias,
+                int numFeatres, int numPasses, int numThreads, [MarshalAs(UnmanagedType.I1)] bool tuneNumLocIter, ref int numLocIter, float tolerance, [MarshalAs(UnmanagedType.I1)] bool needShuffle, [MarshalAs(UnmanagedType.I1)] bool shouldInitialize,
                 State* state, ChannelCallBack info);
 
             /// <summary>

src/Native/CMakeLists.txt

@@ -265,9 +265,11 @@ if(NOT ${ARCHITECTURE} MATCHES "arm.*")
     add_subdirectory(CpuMathNative)
     add_subdirectory(FastTreeNative)
     add_subdirectory(MklProxyNative)
-    # TODO: once we fix the 4 intel MKL methods, SymSgdNative will need to go back in.
     add_subdirectory(SymSgdNative)
-  endif()
+else()
+    add_subdirectory(MklImportsArm)
+    add_subdirectory(SymSgdNative)
+endif()
 
 if(${ARCHITECTURE} MATCHES "[xX].*64")
   add_subdirectory(OneDalNative)

src/Native/MklImportsArm/CMakeLists.txt

@@ -0,0 +1,23 @@
+project(MklImportsArm)
+
+# On ARM platforms, Intel MKL is not available. This target provides a small,
+# self-contained libMklImports covering exactly the symbols SymSGD needs
+# (dense/sparse level-1 CBLAS) plus DFTI stubs, with NO external BLAS
+# dependency. This is required because the CI cross-compilation sysroots do
+# not ship OpenBLAS or any system BLAS.
+
+set(SOURCES
+    MklImportsArm.c
+)
+
+if(NOT WIN32)
+    list(APPEND SOURCES ${VERSION_FILE_PATH})
+    SET(CMAKE_SKIP_BUILD_RPATH  FALSE)
+    SET(CMAKE_BUILD_WITH_INSTALL_RPATH FALSE)
+    SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
+    SET(CMAKE_INSTALL_RPATH "$ORIGIN/")
+endif()
+
+add_library(MklImports SHARED ${SOURCES} ${RESOURCES})
+
+install_library_and_symbols(MklImports)

src/Native/MklImportsArm/MklImportsArm.c

@@ -0,0 +1,118 @@
+// Licensed to the .NET Foundation under one or more agreements.
+// The .NET Foundation licenses this file to you under the MIT license.
+// See the LICENSE file in the project root for more information.
+
+// ARM replacement for Intel MKL (libMklImports.so).
+//
+// This provides a small, self-contained libMklImports for arm/arm64 that
+// covers exactly the symbols SymSGD needs, with no external BLAS dependency.
+// That is important because the cross-compilation sysroots used in CI do not
+// ship OpenBLAS (or any system BLAS), so linking against one is not an option.
+//
+// SymSGD uses only four CBLAS routines:
+//   * cblas_sdot / cblas_saxpy   - dense single-precision dot and AXPY,
+//   * cblas_sdoti / cblas_saxpyi - their sparse counterparts (MKL extensions).
+// All four are implemented below as plain C loops. With -O3 the compiler
+// autovectorizes the dense paths to NEON, matching hand-written BLAS closely.
+//
+// MKL DFTI (FFT) functions are stubbed — they are referenced by the managed
+// MKL Components initializer but not used by SymSGD. The stubs return error
+// codes so any actual FFT call fails cleanly rather than crashing.
+
+// The native build is compiled with -fvisibility=hidden, so every symbol that
+// must be visible to SymSgdNative (the CBLAS routines) or to the managed
+// P/Invoke layer (DftiErrorMessage) has to be exported explicitly.
+#if defined(_WIN32)
+#define MKLIMPORTS_EXPORT __declspec(dllexport)
+#else
+#define MKLIMPORTS_EXPORT __attribute__((visibility("default")))
+#endif
+
+// --- Dense BLAS (CBLAS, level 1) ---
+
+MKLIMPORTS_EXPORT float cblas_sdot(const int n, const float *x, const int incx,
+                 const float *y, const int incy)
+{
+    float result = 0.0f;
+    if (incx == 1 && incy == 1)
+    {
+        for (int i = 0; i < n; i++)
+            result += x[i] * y[i];
+    }
+    else
+    {
+        int ix = incx < 0 ? (1 - n) * incx : 0;
+        int iy = incy < 0 ? (1 - n) * incy : 0;
+        for (int i = 0; i < n; i++, ix += incx, iy += incy)
+            result += x[ix] * y[iy];
+    }
+    return result;
+}
+
+MKLIMPORTS_EXPORT void cblas_saxpy(const int n, const float a, const float *x, const int incx,
+                 float *y, const int incy)
+{
+    if (a == 0.0f)
+        return;
+    if (incx == 1 && incy == 1)
+    {
+        for (int i = 0; i < n; i++)
+            y[i] += a * x[i];
+    }
+    else
+    {
+        int ix = incx < 0 ? (1 - n) * incx : 0;
+        int iy = incy < 0 ? (1 - n) * incy : 0;
+        for (int i = 0; i < n; i++, ix += incx, iy += incy)
+            y[iy] += a * x[ix];
+    }
+}
+
+// --- Sparse BLAS (MKL extensions, not in standard BLAS) ---
+
+MKLIMPORTS_EXPORT void cblas_saxpyi(const int nz, const float a,
+                  const float *x, const int *indx, float *y)
+{
+    for (int i = 0; i < nz; i++)
+        y[indx[i]] += a * x[i];
+}
+
+MKLIMPORTS_EXPORT float cblas_sdoti(const int nz, const float *x,
+                  const int *indx, const float *y)
+{
+    float result = 0.0f;
+    for (int i = 0; i < nz; i++)
+        result += x[i] * y[indx[i]];
+    return result;
+}
+
+// --- DFTI (FFT) stubs ---
+
+MKLIMPORTS_EXPORT const char* DftiErrorMessage(long status)
+{
+    return "DFTI not available (arm64 MKL shim build)";
+}
+
+MKLIMPORTS_EXPORT long DftiCreateDescriptor(void **h, int precision, int domain, int dim, ...)
+{
+    *h = (void*)0;
+    return -1;
+}
+
+MKLIMPORTS_EXPORT long DftiSetValue(void *h, int param, ...)
+{
+    return -1;
+}
+
+MKLIMPORTS_EXPORT long DftiCommitDescriptor(void *h) { return -1; }
+MKLIMPORTS_EXPORT long DftiComputeForward(void *h, ...) { return -1; }
+MKLIMPORTS_EXPORT long DftiComputeBackward(void *h, ...) { return -1; }
+MKLIMPORTS_EXPORT long DftiFreeDescriptor(void **h)
+{
+    // Match MKL's contract: clear the caller's handle after freeing so callers
+    // that rely on the descriptor being nulled out (e.g. the managed
+    // FreeDescriptor(ref IntPtr) P/Invoke) behave correctly.
+    if (h != (void*)0)
+        *h = (void*)0;
+    return 0;
+}

src/Native/SymSgdNative/CMakeLists.txt

@@ -33,7 +33,12 @@ else()
     endif()
 endif()
 
-if(NOT ${ARCHITECTURE} MATCHES "arm.*")
+if(${ARCHITECTURE} MATCHES "arm.*")
+    # On ARM, MklImports is built from MklImportsArm, a self-contained CBLAS
+    # shim implemented as plain C loops (no external BLAS dependency).
+    # Link against the CMake target directly.
+    set(MKL_LIBRARY MklImports)
+else()
     find_library(MKL_LIBRARY MklImports HINTS ${MKL_LIB_PATH})
 endif()
 

src/Native/SymSgdNative/SparseBLAS.h

@@ -5,17 +5,23 @@
 #pragma once
 #include "../Stdafx.h"
 
-extern "C" float __cdecl cblas_sdot(const int vecSize, const float* denseVecX, const int incX, const float* denseVecY, const int incY);
-extern "C" float __cdecl cblas_sdoti(const int sparseVecSize, const float* sparseVecValues, const int* sparseVecIndices, float* denseVec);
-extern "C" void __cdecl cblas_saxpy(const int vecSize, const float coef, const float* denseVecX, const int incX, float* denseVecY, const int incY);
-extern "C" void __cdecl cblas_saxpyi(const int sparseVecSize, const float coef, const float* sparseVecValues, const int* sparseVecIndices, float* denseVec);
+#ifdef _WIN32
+#define CBLAS_CALLING_CONV __cdecl
+#else
+#define CBLAS_CALLING_CONV
+#endif
+
+extern "C" float CBLAS_CALLING_CONV cblas_sdot(const int vecSize, const float* denseVecX, const int incX, const float* denseVecY, const int incY);
+extern "C" float CBLAS_CALLING_CONV cblas_sdoti(const int sparseVecSize, const float* sparseVecValues, const int* sparseVecIndices, const float* denseVec);
+extern "C" void CBLAS_CALLING_CONV cblas_saxpy(const int vecSize, const float coef, const float* denseVecX, const int incX, float* denseVecY, const int incY);
+extern "C" void CBLAS_CALLING_CONV cblas_saxpyi(const int sparseVecSize, const float coef, const float* sparseVecValues, const int* sparseVecIndices, float* denseVec);
 
 float SDOT(const int vecSize, const float* denseVecX, const float* denseVecY) 
 {
     return cblas_sdot(vecSize, denseVecX, 1, denseVecY, 1);
 }
 
-float SDOTI(const int sparseVecSize, const int* sparseVecIndices, const float* sparseVecValues, float* denseVec) 
+float SDOTI(const int sparseVecSize, const int* sparseVecIndices, const float* sparseVecValues, const float* denseVec) 
 {
     return cblas_sdoti(sparseVecSize, sparseVecValues, sparseVecIndices, denseVec);
 }
@@ -28,4 +34,4 @@ void SAXPY(const int vecSize, const float* denseVecX, float* denseVecY, float co
 void SAXPYI(const int sparseVecSize, const int* sparseVecIndices, const float* sparseVecValues, float* denseVec, float coef) 
 {
     cblas_saxpyi(sparseVecSize, coef, sparseVecValues, sparseVecIndices, denseVec);
-}
+}