From 9ad5661482ef8faf1f2ad1a320a24d40f112d802 Mon Sep 17 00:00:00 2001 From: Rasmus Munk Larsen Date: Tue, 15 Mar 2022 20:51:03 +0000 Subject: [PATCH] Revert "Fix up PowerPC MMA flags so it builds by default." --- Eigen/src/Core/arch/AltiVec/MatrixProduct.h | 211 +++++++++--------- .../src/Core/arch/AltiVec/MatrixProductMMA.h | 11 +- 2 files changed, 108 insertions(+), 114 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/MatrixProduct.h b/Eigen/src/Core/arch/AltiVec/MatrixProduct.h index 00411fcec..2919dda8b 100644 --- a/Eigen/src/Core/arch/AltiVec/MatrixProduct.h +++ b/Eigen/src/Core/arch/AltiVec/MatrixProduct.h @@ -17,27 +17,24 @@ #include "MatrixProductCommon.h" -// Check for MMA builtin support. -#if !defined(EIGEN_ALTIVEC_DISABLE_MMA) && defined(__has_builtin) +// Since LLVM doesn't support dynamic dispatching, force either always MMA or VSX +#if EIGEN_COMP_LLVM +#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 EIGEN_ALTIVEC_MMA_SUPPORT + #define ALTIVEC_MMA_SUPPORT #endif #endif -// Check if and how we should actually use MMA if supported. -#if defined(EIGEN_ALTIVEC_MMA_SUPPORT) - -// Use MMA by default if available. -#if defined(__MMA__) -#define EIGEN_ALTIVEC_MMA_ONLY 1 -// Otherwise, check if we want to enable dynamic dispatch. Not supported by LLVM. -#elif defined(EIGEN_ALTIVEC_ENABLE_MMA_DYNAMIC_DISPATCH) && !defined(EIGEN_COMP_LLVM) -#define EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH 1 -#endif - -#endif // EIGEN_ALTIVEC_MMA_SUPPORT - -#if defined(EIGEN_ALTIVEC_MMA_ONLY) || defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH) +#if defined(ALTIVEC_MMA_SUPPORT) && !defined(EIGEN_ALTIVEC_DISABLE_MMA) #include "MatrixProductMMA.h" #endif @@ -2484,10 +2481,10 @@ void gebp_kernel::size; void (*gemm_function)(const DataMapper&, const float*, const float*, Index, Index, Index, float, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) + #ifdef EIGEN_ALTIVEC_MMA_ONLY //generate with MMA only gemm_function = &Eigen::internal::gemmMMA; - #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH) + #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; } @@ -2497,7 +2494,7 @@ void gebp_kernel; #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } template @@ -2523,20 +2520,20 @@ void gebp_kernel, std::complex, Index, DataMapper, mr void (*gemm_function)(const DataMapper&, const std::complex*, const std::complex*, Index, Index, Index, std::complex, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) - //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(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH) - 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>; - } - #else - gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; - #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + #ifdef EIGEN_ALTIVEC_MMA_ONLY + //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>; + } + #else + gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + #endif + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } template @@ -2561,20 +2558,20 @@ void gebp_kernel, Index, DataMapper, mr, nr, Conjugat const Index accCols = quad_traits::size; void (*gemm_function)(const DataMapper&, const float*, const std::complex*, Index, Index, Index, std::complex, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) - //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(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, 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 - gemm_function = &Eigen::internal::gemm_complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; - #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + #ifdef EIGEN_ALTIVEC_MMA_ONLY + //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>; + } + #else + gemm_function = &Eigen::internal::gemm_complex, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + #endif + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } template @@ -2599,20 +2596,20 @@ void gebp_kernel, float, Index, DataMapper, mr, nr, Conjugat const Index accCols = quad_traits::size; void (*gemm_function)(const DataMapper&, const std::complex*, const float*, Index, Index, Index, std::complex, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) - //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(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, 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 - gemm_function = &Eigen::internal::gemm_complex, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; - #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + #ifdef EIGEN_ALTIVEC_MMA_ONLY + //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>; + } + #else + gemm_function = &Eigen::internal::gemm_complex, float, std::complex, float, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + #endif + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } template @@ -2636,10 +2633,10 @@ void gebp_kernel::size; void (*gemm_function)(const DataMapper&, const double*, const double*, Index, Index, Index, double, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) + #ifdef EIGEN_ALTIVEC_MMA_ONLY //generate with MMA only gemm_function = &Eigen::internal::gemmMMA; - #elif defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH) + #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; } @@ -2649,7 +2646,7 @@ void gebp_kernel; #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } template @@ -2674,20 +2671,20 @@ void gebp_kernel, std::complex, Index, DataMapper, const Index accCols = quad_traits::size; void (*gemm_function)(const DataMapper&, const std::complex*, const std::complex*, Index, Index, Index, std::complex, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) - //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(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH) - 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>; - } - #else - gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; - #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + #ifdef EIGEN_ALTIVEC_MMA_ONLY + //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>; + } + #else + gemm_function = &Eigen::internal::gemm_complex, std::complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, false>; + #endif + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } template @@ -2712,20 +2709,20 @@ void gebp_kernel, double, Index, DataMapper, mr, nr, Conjug const Index accCols = quad_traits::size; void (*gemm_function)(const DataMapper&, const std::complex*, const double*, Index, Index, Index, std::complex, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) - //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(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, 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 - gemm_function = &Eigen::internal::gemm_complex, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; - #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + #ifdef EIGEN_ALTIVEC_MMA_ONLY + //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>; + } + #else + gemm_function = &Eigen::internal::gemm_complex, double, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, false, true>; + #endif + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } template @@ -2750,20 +2747,20 @@ void gebp_kernel, Index, DataMapper, mr, nr, Conjug const Index accCols = quad_traits::size; void (*gemm_function)(const DataMapper&, const double*, const std::complex*, Index, Index, Index, std::complex, Index, Index, Index, Index); - #if defined(EIGEN_ALTIVEC_MMA_ONLY) - //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(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, 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 - gemm_function = &Eigen::internal::gemm_complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; - #endif - gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); + #ifdef EIGEN_ALTIVEC_MMA_ONLY + //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>; + } + #else + gemm_function = &Eigen::internal::gemm_complex, std::complex, double, Index, Packet, Packetc, RhsPacket, DataMapper, accRows, accCols, ConjugateLhs, ConjugateRhs, true, false>; + #endif + gemm_function(res, blockA, blockB, rows, depth, cols, alpha, strideA, strideB, offsetA, offsetB); } } // end namespace internal diff --git a/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h b/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h index 5a0834758..8104697a1 100644 --- a/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h +++ b/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h @@ -11,9 +11,7 @@ #ifndef EIGEN_MATRIX_PRODUCT_MMA_ALTIVEC_H #define EIGEN_MATRIX_PRODUCT_MMA_ALTIVEC_H -// If using dynamic dispatch, set the CPU target. -#if defined(EIGEN_ALTIVEC_MMA_DYNAMIC_DISPATCH) -#pragma GCC push_options +#if !EIGEN_COMP_LLVM #pragma GCC target("cpu=power10,htm") #endif @@ -617,13 +615,12 @@ 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