Fix up PowerPC MMA flags so it builds by default.

Eigen/src/Core/arch/AltiVec/MatrixProduct.h

@@ -17,24 +17,28 @@
 
 #include "MatrixProductCommon.h"
 
-// Since LLVM doesn't support dynamic dispatching, force either always MMA or VSX
+// Check for MMA builtin support. 
-#if EIGEN_COMP_LLVM
+#if !defined(EIGEN_ALTIVEC_DISABLE_MMA) && defined(__has_builtin)
-#if !defined(EIGEN_ALTIVEC_DISABLE_MMA) && !defined(EIGEN_ALTIVEC_MMA_ONLY)
-#ifdef __MMA__
-#define EIGEN_ALTIVEC_MMA_ONLY
-#else
-#define EIGEN_ALTIVEC_DISABLE_MMA
-#endif
-#endif
-#endif
-
-#ifdef __has_builtin
 #if __has_builtin(__builtin_mma_assemble_acc)
-  #define ALTIVEC_MMA_SUPPORT
+  #define EIGEN_ALTIVEC_MMA_SUPPORT
 #endif
 #endif
 
-#if defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+// Check if and how we should actually use MMA if supported.
+#if defined(EIGEN_ALTIVEC_MMA_SUPPORT)
+
+
+// Check if we want to enable dynamic dispatch. Not supported by LLVM.
+if defined(EIGEN_ALTIVEC_ENABLE_MMA_DYNAMIC_DISPATCH) && !EIGEN_COMP_LLVM
+#define EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH 1
+// Otherwise, use MMA by default if available.
+#elif defined(__MMA__)
+#define EIGEN_ALTIVEC_MMA_ONLY 1
+#endif
+
+#endif // EIGEN_ALTIVEC_MMA_SUPPORT
+
+#if defined(EIGEN_ALTIVEC_MMA_ONLY) || defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
   #include "MatrixProductMMA.h"
 #endif
 
@@ -2481,10 +2485,10 @@ void gebp_kernel<float, float, Index, DataMapper, mr, nr, ConjugateLhs, Conjugat
     const Index accCols = quad_traits<float>::size;
     void (*gemm_function)(const DataMapper&, const float*, const float*, Index, Index, Index, float, Index, Index, Index, Index);
 
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemmMMA<float, Index, Packet, RhsPacket, DataMapper, accRows, accCols>;
-    #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemmMMA<float, Index, Packet, RhsPacket, DataMapper, accRows, accCols>;
       }
@@ -2520,10 +2524,10 @@ void gebp_kernel<std::complex<float>, std::complex<float>, Index, DataMapper, mr
     void (*gemm_function)(const DataMapper&, const std::complex<float>*, const std::complex<float>*,
           Index, Index, Index, std::complex<float>, Index, Index, Index, Index);
 
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<float>, std::complex<float>, std::complex<float>, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>;
-     #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<float>, std::complex<float>, std::complex<float>, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>;
       }
@@ -2558,10 +2562,10 @@ void gebp_kernel<float, std::complex<float>, Index, DataMapper, mr, nr, Conjugat
     const Index accCols = quad_traits<float>::size;
     void (*gemm_function)(const DataMapper&, const float*, const std::complex<float>*,
           Index, Index, Index, std::complex<float>, Index, Index, Index, Index);
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemm_complexMMA<float, std::complex<float>, std::complex<float>, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>;
-     #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemm_complexMMA<float, std::complex<float>, std::complex<float>, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>;
       }
@@ -2596,10 +2600,10 @@ void gebp_kernel<std::complex<float>, float, Index, DataMapper, mr, nr, Conjugat
     const Index accCols = quad_traits<float>::size;
     void (*gemm_function)(const DataMapper&, const std::complex<float>*, const float*,
           Index, Index, Index, std::complex<float>, Index, Index, Index, Index);
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<float>, float, std::complex<float>, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>;
-     #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<float>, float, std::complex<float>, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>;
       }
@@ -2633,10 +2637,10 @@ void gebp_kernel<double, double, Index, DataMapper, mr, nr, ConjugateLhs, Conjug
     const Index accCols = quad_traits<double>::size;
     void (*gemm_function)(const DataMapper&, const double*, const double*, Index, Index, Index, double, Index, Index, Index, Index);
 
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemmMMA<double, Index, Packet, RhsPacket, DataMapper, accRows, accCols>;
-    #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemmMMA<double, Index, Packet, RhsPacket, DataMapper, accRows, accCols>;
       }
@@ -2671,10 +2675,10 @@ void gebp_kernel<std::complex<double>, std::complex<double>, Index, DataMapper,
     const Index accCols = quad_traits<double>::size;
     void (*gemm_function)(const DataMapper&, const std::complex<double>*, const std::complex<double>*,
           Index, Index, Index, std::complex<double>, Index, Index, Index, Index);
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<double>, std::complex<double>, std::complex<double>, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>;
-     #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<double>, std::complex<double>, std::complex<double>, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>;
       }
@@ -2709,10 +2713,10 @@ void gebp_kernel<std::complex<double>, double, Index, DataMapper, mr, nr, Conjug
     const Index accCols = quad_traits<double>::size;
     void (*gemm_function)(const DataMapper&, const std::complex<double>*, const double*,
           Index, Index, Index, std::complex<double>, Index, Index, Index, Index);
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<double>, double, std::complex<double>, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>;
-     #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemm_complexMMA<std::complex<double>, double, std::complex<double>, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>;
       }
@@ -2747,10 +2751,10 @@ void gebp_kernel<double, std::complex<double>, Index, DataMapper, mr, nr, Conjug
     const Index accCols = quad_traits<double>::size;
     void (*gemm_function)(const DataMapper&, const double*, const std::complex<double>*,
           Index, Index, Index, std::complex<double>, Index, Index, Index, Index);
-    #ifdef EIGEN_ALTIVEC_MMA_ONLY
+    #if defined(EIGEN_ALTIVEC_MMA_ONLY)
       //generate with MMA only
       gemm_function = &Eigen::internal::gemm_complexMMA<double, std::complex<double>, std::complex<double>, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>;
-     #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA)
+    #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
       if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){
         gemm_function = &Eigen::internal::gemm_complexMMA<double, std::complex<double>, std::complex<double>, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>;
       }

Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h

@@ -11,7 +11,9 @@
 #ifndef EIGEN_MATRIX_PRODUCT_MMA_ALTIVEC_H
 #define EIGEN_MATRIX_PRODUCT_MMA_ALTIVEC_H
 
-#if !EIGEN_COMP_LLVM
+// If using dynamic dispatch, set the CPU target.
+#if defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
+#pragma GCC push_options
 #pragma GCC target("cpu=power10,htm")
 #endif
 
@@ -615,12 +617,13 @@ void gemm_complexMMA(const DataMapper& res, const LhsScalar* blockAc, const RhsS
 #undef advanceRows
 #undef advanceCols
 
-#if !EIGEN_COMP_LLVM
-#pragma GCC reset_options
-#endif
 } // end namespace internal
 
 } // end namespace Eigen
 
+#if defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH)
+#pragma GCC pop_options
+#endif
+
 #endif // EIGEN_MATRIX_PRODUCT_MMA_ALTIVEC_H