avx2.hpp

Go to the documentation of this file.

#pragma once
 
#include <batmat/config.hpp>
#include <batmat/simd.hpp>
 
namespace batmat {
namespace linalg::micro_kernels::gemm {
 
/// Register block size of the matrix-matrix multiplication micro-kernels.
/// AVX2 has 16 vector registers, we use 9 registers for a 3×3 accumulator
/// block of matrix C (leaving some registers for loading A and B):
/// @note   A block size of 4×4 is slightly faster than 3×3 for large matrices, because the even
///         block size results in full cache lines being consumed. For small matrices, 3×3 is faster
///         because it does not spill any registers in the micro-kernels. 2×2 is slower than 3×3 for
///         both small and large matrices (tested using GCC 15.1 on an i7-10750H).
template <class T, class Abi>
inline constexpr index_t RowsReg = 3;
// Vectors greater than the physical vector length use more registers, so decrease the block size.
template <class T, class Abi>
    requires(datapar::simd_size<T, Abi>::value * sizeof(T) > 32)
inline constexpr index_t RowsReg<T, Abi> = 2;
 
} // namespace linalg::micro_kernels::gemm
namespace ops {
 
template <class T>
inline constexpr index_t RowsRegTranspose = 4;
template <class T>
inline constexpr index_t ColsRegTranspose = 4;
 
} // namespace ops
} // namespace batmat