diff --git a/Eigen/src/Core/arch/AltiVec/MatrixProduct.h b/Eigen/src/Core/arch/AltiVec/MatrixProduct.h index 8feb88ea7..01b50b7c4 100644 --- a/Eigen/src/Core/arch/AltiVec/MatrixProduct.h +++ b/Eigen/src/Core/arch/AltiVec/MatrixProduct.h @@ -15,6 +15,21 @@ #define EIGEN_ALTIVEC_USE_CUSTOM_PACK 1 #endif +#ifdef _AIX +#define POWER_10 0x40000 +#define POWER_10_ANDUP (POWER_10) +#include +#define __power_10_andup() (_system_configuration.implementation & POWER_10_ANDUP) + +static bool check_mma_support() +{ + if(__power_10_andup() && __power_mma_version() == MMA_V31) + return true; + else + return false; +} +#endif + #include "MatrixProductCommon.h" // Since LLVM doesn't support dynamic dispatching, force either always MMA or VSX @@ -2481,12 +2496,21 @@ void gebp_kernel; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemmMMA; - } - else{ - gemm_function = &Eigen::internal::gemm; - } + #if defined(_AIX) + if(check_mma_support()){ + gemm_function = &Eigen::internal::gemmMMA; + } + else{ + gemm_function = &Eigen::internal::gemm; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemmMMA; + } + else{ + gemm_function = &Eigen::internal::gemm; + } + #endif #else gemm_function = &Eigen::internal::gemm; #endif @@ -2520,12 +2544,21 @@ void gebp_kernel, std::complex, Index, DataMapper, mr //generate with MMA only gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; - } - else{ - gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; - } + #if defined(_AIX) + if (check_mma_support()){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + #endif #else gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; #endif @@ -2558,12 +2591,21 @@ void gebp_kernel, Index, DataMapper, mr, nr, Conjugat //generate with MMA only gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; - } - else{ - gemm_function = &Eigen::internal::gemm_complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; - } + #if defined(_AIX) + if (check_mma_support()){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + #endif #else gemm_function = &Eigen::internal::gemm_complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; #endif @@ -2596,12 +2638,21 @@ void gebp_kernel, float, Index, DataMapper, mr, nr, Conjugat //generate with MMA only gemm_function = &Eigen::internal::gemm_complexMMA, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemm_complexMMA, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; - } - else{ - gemm_function = &Eigen::internal::gemm_complex, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; - } + #if defined(_AIX) + if (check_mma_support()){ + gemm_function = &Eigen::internal::gemm_complexMMA, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemm_complexMMA, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + #endif #else gemm_function = &Eigen::internal::gemm_complex, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; #endif @@ -2633,12 +2684,21 @@ void gebp_kernel; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemmMMA; - } - else{ - gemm_function = &Eigen::internal::gemm; - } + #if defined(_AIX) + if (check_mma_support()){ + gemm_function = &Eigen::internal::gemmMMA; + } + else{ + gemm_function = &Eigen::internal::gemm; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemmMMA; + } + else{ + gemm_function = &Eigen::internal::gemm; + } + #endif #else gemm_function = &Eigen::internal::gemm; #endif @@ -2671,12 +2731,21 @@ void gebp_kernel, std::complex, Index, DataMapper, //generate with MMA only gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; - } - else{ - gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; - } + #if defined(_AIX) + if (check_mma_support()){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + } + #endif #else gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; #endif @@ -2709,12 +2778,21 @@ void gebp_kernel, double, Index, DataMapper, mr, nr, Conjug //generate with MMA only gemm_function = &Eigen::internal::gemm_complexMMA, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemm_complexMMA, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; - } - else{ - gemm_function = &Eigen::internal::gemm_complex, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; - } + #if defined(_AIX) + if (check_mma_support()){ + gemm_function = &Eigen::internal::gemm_complexMMA, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemm_complexMMA, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + } + #endif #else gemm_function = &Eigen::internal::gemm_complex, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; #endif @@ -2747,12 +2825,21 @@ void gebp_kernel, Index, DataMapper, mr, nr, Conjug //generate with MMA only gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; #elif defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) - if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ - gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; - } - else{ - gemm_function = &Eigen::internal::gemm_complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; - } + #if defined(_AIX) + if (check_mma_support()){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + #else + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")){ + gemm_function = &Eigen::internal::gemm_complexMMA, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + else{ + gemm_function = &Eigen::internal::gemm_complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + } + #endif #else gemm_function = &Eigen::internal::gemm_complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; #endif