devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove unused LhsPacketType typedef in gebp_peeled_loop

Browse Source 
libeigen/eigen!2162

Closes #3029

Co-authored-by: Rasmus Munk Larsen <rmlarsen@gmail.com>
This commit is contained in:
Rasmus Munk Larsen
2026-02-19 20:15:15 -08:00
parent 4fdc82d695
commit 2c898e8b95
1 changed files with 32 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@@ -1154,12 +1154,14 @@ struct gebp_micro_step {
}; };
// Compiler workaround macros used inside gebp_peeled_loop. // Compiler workaround macros used inside gebp_peeled_loop.
#if EIGEN_ARCH_ARM64 && defined(EIGEN_VECTORIZE_NEON) && EIGEN_GNUC_STRICT_LESS_THAN(9, 0, 0) #if EIGEN_ARCH_ARM64 && defined(EIGEN_VECTORIZE_NEON) && EIGEN_GNUC_STRICT_LESS_THAN(9, 0, 0)
#define EIGEN_GEBP_ARM64_3P_WORKAROUND(MrPackets, A, LhsPacketType, FullLhsPacket) \ #define EIGEN_GEBP_ARM64_3P_WORKAROUND(MrPackets, A, LhsArray, FullLhsPacket) \
EIGEN_IF_CONSTEXPR((MrPackets == 3 && std::is_same<LhsPacketType, FullLhsPacket>::value)) { \ EIGEN_IF_CONSTEXPR( \
__asm__("" : "+w,m"(A[0]), "+w,m"(A[1]), "+w,m"(A[2])); \ (MrPackets == 3 && \
std::is_same<std::remove_all_extents_t<std::remove_reference_t<LhsArray>>, FullLhsPacket>::value)) { \
__asm__("" : "+w,m"(A[0]), "+w,m"(A[1]), "+w,m"(A[2])); \
} }
#else #else
#define EIGEN_GEBP_ARM64_3P_WORKAROUND(MrPackets, A, LhsPacketType, FullLhsPacket) #define EIGEN_GEBP_ARM64_3P_WORKAROUND(MrPackets, A, LhsArray, FullLhsPacket)
#endif #endif
// GCC's register allocator can fail to keep array-based accumulators in XMM // GCC's register allocator can fail to keep array-based accumulators in XMM
@@ -1172,9 +1174,11 @@ struct gebp_micro_step {
#ifdef EIGEN_HAS_CXX17_IFCONSTEXPR #ifdef EIGEN_HAS_CXX17_IFCONSTEXPR
// C++17: pin accumulators when they're plain SSE vectors (sizeof matches FullLhsPacket). // C++17: pin accumulators when they're plain SSE vectors (sizeof matches FullLhsPacket).
// For complex types, AccPacket is a struct (DoublePacket) and the asm is safely discarded. // For complex types, AccPacket is a struct (DoublePacket) and the asm is safely discarded.
#define EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsPacketType, FullLhsPacket) \ #define EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsArray, FullLhsPacket) \
EIGEN_IF_CONSTEXPR((MrPackets <= 2 && NrCols >= 4 && std::is_same<LhsPacketType, FullLhsPacket>::value && \ EIGEN_IF_CONSTEXPR( \
sizeof(ACC[0]) == sizeof(FullLhsPacket))) { \ (MrPackets <= 2 && NrCols >= 4 && \
std::is_same<std::remove_all_extents_t<std::remove_reference_t<LhsArray>>, FullLhsPacket>::value && \
sizeof(ACC[0]) == sizeof(FullLhsPacket))) { \
EIGEN_IF_CONSTEXPR(MrPackets == 2 && NrCols == 4) { \ EIGEN_IF_CONSTEXPR(MrPackets == 2 && NrCols == 4) { \
__asm__("" \ __asm__("" \
: "+x"(ACC[0]), "+x"(ACC[1]), "+x"(ACC[2]), "+x"(ACC[3]), "+x"(ACC[4]), "+x"(ACC[5]), "+x"(ACC[6]), \ : "+x"(ACC[0]), "+x"(ACC[1]), "+x"(ACC[2]), "+x"(ACC[3]), "+x"(ACC[4]), "+x"(ACC[5]), "+x"(ACC[6]), \
@@ -1184,10 +1188,12 @@ struct gebp_micro_step {
} }
#else #else
// C++14: only pin LHS packets (A), not accumulators, to avoid asm errors with complex types. // C++14: only pin LHS packets (A), not accumulators, to avoid asm errors with complex types.
#define EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsPacketType, FullLhsPacket) \ #define EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsArray, FullLhsPacket) \
EIGEN_IF_CONSTEXPR((MrPackets <= 2 && NrCols >= 4 && std::is_same<LhsPacketType, FullLhsPacket>::value)) { \ EIGEN_IF_CONSTEXPR( \
EIGEN_IF_CONSTEXPR(MrPackets == 2) { __asm__("" : "+x,m"(A[0]), "+x,m"(A[1])); } \ (MrPackets <= 2 && NrCols >= 4 && \
EIGEN_GEBP_SSE_1P_WORKAROUND(MrPackets, NrCols, A, ACC) \ std::is_same<std::remove_all_extents_t<std::remove_reference_t<LhsArray>>, FullLhsPacket>::value)) { \
EIGEN_IF_CONSTEXPR(MrPackets == 2) { __asm__("" : "+x,m"(A[0]), "+x,m"(A[1])); } \
EIGEN_GEBP_SSE_1P_WORKAROUND(MrPackets, NrCols, A, ACC) \
} }
#endif #endif
#if !(EIGEN_COMP_LCC) #if !(EIGEN_COMP_LCC)
@@ -1197,7 +1203,7 @@ struct gebp_micro_step {
#define EIGEN_GEBP_SSE_1P_WORKAROUND(MrPackets, NrCols, A, ACC) #define EIGEN_GEBP_SSE_1P_WORKAROUND(MrPackets, NrCols, A, ACC)
#endif #endif
#else #else
#define EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsPacketType, FullLhsPacket) #define EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsArray, FullLhsPacket)
#endif #endif
// Unrolled peeled loop body: calls gebp_micro_step for K=0..7, handling // Unrolled peeled loop body: calls gebp_micro_step for K=0..7, handling
@@ -1208,27 +1214,26 @@ struct gebp_peeled_loop {
typename RhsPacketType, typename AccArray, typename AccArrayD, typename FullLhsPacket> typename RhsPacketType, typename AccArray, typename AccArrayD, typename FullLhsPacket>
static EIGEN_ALWAYS_INLINE void run(GEBPTraits& traits, const LhsScalar_* blA, const RhsScalar_* blB, LhsArray& A, static EIGEN_ALWAYS_INLINE void run(GEBPTraits& traits, const LhsScalar_* blA, const RhsScalar_* blB, LhsArray& A,
RhsPanelType& rhs_panel, RhsPacketType& T0, AccArray& C, AccArrayD& D) { RhsPanelType& rhs_panel, RhsPacketType& T0, AccArray& C, AccArrayD& D) {
using LhsPacketType = std::remove_all_extents_t<std::remove_reference_t<LhsArray>>;
constexpr bool use_double_accum = (MrPackets == 1 && NrCols == 4); constexpr bool use_double_accum = (MrPackets == 1 && NrCols == 4);
// Prefetch for 4-col paths // Prefetch for 4-col paths
EIGEN_IF_CONSTEXPR(NrCols == 4) { internal::prefetch(blB + (48 + 0)); } EIGEN_IF_CONSTEXPR(NrCols == 4) { internal::prefetch(blB + (48 + 0)); }
// Helper to do one step with workarounds // Helper to do one step with workarounds
#define EIGEN_GEBP_DO_STEP(KVAL, ACC) \ #define EIGEN_GEBP_DO_STEP(KVAL, ACC) \
do { \ do { \
gebp_micro_step<KVAL, MrPackets, NrCols>::run(traits, blA, blB, A, rhs_panel, T0, ACC); \ gebp_micro_step<KVAL, MrPackets, NrCols>::run(traits, blA, blB, A, rhs_panel, T0, ACC); \
/* ARM64 NEON register alloc workaround for 3-packet paths */ \ /* ARM64 NEON register alloc workaround for 3-packet paths */ \
EIGEN_GEBP_ARM64_3P_WORKAROUND(MrPackets, A, LhsPacketType, FullLhsPacket) \ EIGEN_GEBP_ARM64_3P_WORKAROUND(MrPackets, A, LhsArray, FullLhsPacket) \
/* GCC SSE spilling workaround: pin LHS packets and accumulators in registers */ \ /* GCC SSE spilling workaround: pin LHS packets and accumulators in registers */ \
EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsPacketType, FullLhsPacket) \ EIGEN_GEBP_SSE_SPILLING_WORKAROUND(MrPackets, NrCols, A, ACC, LhsArray, FullLhsPacket) \
/* LHS prefetch for 2pX4 and 3pX4 */ \ /* LHS prefetch for 2pX4 and 3pX4 */ \
EIGEN_IF_CONSTEXPR((MrPackets == 2 || MrPackets == 3) && NrCols == 4) { \ EIGEN_IF_CONSTEXPR((MrPackets == 2 || MrPackets == 3) && NrCols == 4) { \
internal::prefetch(blA + (MrPackets * KVAL + 16) * GEBPTraits::LhsProgress); \ internal::prefetch(blA + (MrPackets * KVAL + 16) * GEBPTraits::LhsProgress); \
if (EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) { \ if (EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) { \
internal::prefetch(blB + (NrCols * KVAL + 16) * GEBPTraits::RhsProgress); \ internal::prefetch(blB + (NrCols * KVAL + 16) * GEBPTraits::RhsProgress); \
} \ } \
} \ } \
} while (false) } while (false)
EIGEN_IF_CONSTEXPR(use_double_accum) { EIGEN_IF_CONSTEXPR(use_double_accum) {