| 743 | /// @see LargeKernel for an explanation of the algorithm. |
| 744 | template <typename KerTraits> |
| 745 | struct LargeKernelPlan { |
| 746 | using Traits = KerTraits; |
| 747 | using uint_type = typename Traits::uint_type; |
| 748 | static constexpr auto kShape = Traits::kShape; |
| 749 | static constexpr auto kPlanSize = LargeKernelPlanSize::Build(kShape); |
| 750 | |
| 751 | /// Array of byte offsets relative to the Kernel input memory address. |
| 752 | using ReadsPerKernel = std::array<int, kPlanSize.reads_per_kernel()>; |
| 753 | |
| 754 | /// Array of indices used for byte swizzle/shuffle in an SIMD register. |
| 755 | using Swizzle = std::array<uint8_t, kShape.simd_byte_size()>; |
| 756 | using SwizzlesPerKernel = std::array<Swizzle, kPlanSize.reads_per_kernel()>; |
| 757 | |
| 758 | /// Array of integer shifts for each unpacked integer in an SIMD register. |
| 759 | using Shifts = std::array<uint_type, kShape.unpacked_per_simd()>; |
| 760 | using ShitsPerKernel = std::array<Shifts, kPlanSize.reads_per_kernel()>; |
| 761 | |
| 762 | static constexpr LargeKernelPlan Build(); |
| 763 | |
| 764 | static constexpr int bytes_per_read() { |
| 765 | const auto bits_per_read = kPlanSize.unpacked_per_read() * kShape.packed_bit_size(); |
| 766 | return adjust_bytes_per_read(bits_per_read, kShape.simd_byte_size()); |
| 767 | } |
| 768 | |
| 769 | constexpr int total_bytes_read() const { return reads.back() + bytes_per_read(); } |
| 770 | |
| 771 | ReadsPerKernel reads; |
| 772 | SwizzlesPerKernel low_swizzles; |
| 773 | SwizzlesPerKernel high_swizzles; |
| 774 | ShitsPerKernel low_rshifts; |
| 775 | ShitsPerKernel high_lshifts; |
| 776 | uint_type mask; |
| 777 | }; |
| 778 | |
| 779 | template <typename KerTraits> |
| 780 | constexpr auto LargeKernelPlan<KerTraits>::Build() -> LargeKernelPlan<KerTraits> { |
nothing calls this directly
no test coverage detected