|
| 1 | +module LibTBLIS |
| 2 | + |
| 3 | +using tblis_jll |
| 4 | +using LinearAlgebra: BlasFloat |
| 5 | + |
| 6 | +export tblis_scalar, tblis_tensor |
| 7 | +export tblis_tensor_add, tblis_tensor_mult, tblis_tensor_dot |
| 8 | +export tblis_set_num_threads, tblis_get_num_threads |
| 9 | + |
| 10 | +const ptrdiff_t = Cptrdiff_t |
| 11 | + |
| 12 | +const scomplex = ComplexF32 |
| 13 | +const dcomplex = ComplexF64 |
| 14 | + |
| 15 | +const IS_LIBC_MUSL = occursin("musl", Base.BUILD_TRIPLET) |
| 16 | +if Sys.isapple() && Sys.ARCH === :aarch64 |
| 17 | + include("lib/aarch64-apple-darwin20.jl") |
| 18 | +elseif Sys.islinux() && Sys.ARCH === :aarch64 && !IS_LIBC_MUSL |
| 19 | + include("lib/aarch64-linux-gnu.jl") |
| 20 | +elseif Sys.islinux() && Sys.ARCH === :aarch64 && IS_LIBC_MUSL |
| 21 | + include("lib/aarch64-linux-musl.jl") |
| 22 | +elseif Sys.islinux() && startswith(string(Sys.ARCH), "arm") && !IS_LIBC_MUSL |
| 23 | + include("lib/armv7l-linux-gnueabihf.jl") |
| 24 | +elseif Sys.islinux() && startswith(string(Sys.ARCH), "arm") && IS_LIBC_MUSL |
| 25 | + include("lib/armv7l-linux-musleabihf.jl") |
| 26 | +elseif Sys.islinux() && Sys.ARCH === :i686 && !IS_LIBC_MUSL |
| 27 | + include("lib/i686-linux-gnu.jl") |
| 28 | +elseif Sys.islinux() && Sys.ARCH === :i686 && IS_LIBC_MUSL |
| 29 | + include("lib/i686-linux-musl.jl") |
| 30 | +elseif Sys.iswindows() && Sys.ARCH === :i686 |
| 31 | + include("lib/i686-w64-mingw32.jl") |
| 32 | +elseif Sys.islinux() && Sys.ARCH === :powerpc64le |
| 33 | + include("lib/powerpc64le-linux-gnu.jl") |
| 34 | +elseif Sys.isapple() && Sys.ARCH === :x86_64 |
| 35 | + include("lib/x86_64-apple-darwin14.jl") |
| 36 | +elseif Sys.islinux() && Sys.ARCH === :x86_64 && !IS_LIBC_MUSL |
| 37 | + include("lib/x86_64-linux-gnu.jl") |
| 38 | +elseif Sys.islinux() && Sys.ARCH === :x86_64 && IS_LIBC_MUSL |
| 39 | + include("lib/x86_64-linux-musl.jl") |
| 40 | +elseif Sys.isbsd() && !Sys.isapple() |
| 41 | + include("lib/x86_64-unknown-freebsd.jl") |
| 42 | +elseif Sys.iswindows() && Sys.ARCH === :x86_64 |
| 43 | + include("lib/x86_64-w64-mingw32.jl") |
| 44 | +else |
| 45 | + error("Unknown platform: $(Base.BUILD_TRIPLET)") |
| 46 | +end |
| 47 | + |
| 48 | +# tblis_scalar and tblis_tensor |
| 49 | +# ----------------------------- |
| 50 | +""" |
| 51 | + tblis_scalar(s::Number) |
| 52 | +
|
| 53 | +Initializes a tblis scalar from a number. |
| 54 | +""" |
| 55 | +function tblis_scalar end |
| 56 | + |
| 57 | +""" |
| 58 | + tblis_tensor(A::AbstractArray{T<:BlasFloat}, [szA::Vector{Int}, strA::Vector{Int}, scalar::Number]) |
| 59 | +
|
| 60 | +Initializes a tblis tensor from an array that should be strided and admit a pointer to its |
| 61 | +data. This operation is deemed unsafe, in the sense that the user is responsible for ensuring |
| 62 | +that the reference to the array and the sizes and strides stays alive during the lifetime of |
| 63 | +the tensor. |
| 64 | +""" |
| 65 | +function tblis_tensor end |
| 66 | + |
| 67 | +for (T, tblis_init_scalar, tblis_init_tensor, tblis_init_tensor_scaled) in |
| 68 | + ((:Float32, :tblis_init_scalar_s, :tblis_init_tensor_s, :tblis_init_tensor_scaled_s), |
| 69 | + (:Float64, :tblis_init_scalar_d, :tblis_init_tensor_d, :tblis_init_tensor_scaled_d), |
| 70 | + (:ComplexF32, :tblis_init_scalar_c, :tblis_init_tensor_c, :tblis_init_tensor_scaled_c), |
| 71 | + (:ComplexF64, :tblis_init_scalar_z, :tblis_init_tensor_z, :tblis_init_tensor_scaled_z)) |
| 72 | + @eval begin |
| 73 | + function tblis_scalar(s::$T) |
| 74 | + t = Ref{tblis_scalar}() |
| 75 | + $tblis_init_scalar(t, s) |
| 76 | + return t[] |
| 77 | + end |
| 78 | + function tblis_tensor(A::AbstractArray{$T,N}, |
| 79 | + s::Number=one(T), |
| 80 | + szA::Vector{len_type}=collect(len_type, size(A)), |
| 81 | + strA::Vector{stride_type}=collect(stride_type, strides(A))) where {N} |
| 82 | + t = Ref{tblis_tensor}() |
| 83 | + if isone(s) |
| 84 | + $tblis_init_tensor(t, N, pointer(szA), pointer(A), pointer(strA)) |
| 85 | + else |
| 86 | + $tblis_init_tensor_scaled(t, $T(s), N, pointer(szA), pointer(A), |
| 87 | + pointer(strA)) |
| 88 | + end |
| 89 | + return t[] |
| 90 | + end |
| 91 | + end |
| 92 | +end |
| 93 | + |
| 94 | +# tensor operations |
| 95 | +# ------------------ |
| 96 | +function tblis_tensor_add(A::tblis_tensor, idxA, B::tblis_tensor, idxB) |
| 97 | + return tblis_tensor_add(C_NULL, C_NULL, Ref(A), idxA, Ref(B), idxB) |
| 98 | +end |
| 99 | + |
| 100 | +function tblis_tensor_mult(A::tblis_tensor, idxA, B::tblis_tensor, idxB, C::tblis_tensor, |
| 101 | + idxC) |
| 102 | + return tblis_tensor_mult(C_NULL, C_NULL, Ref(A), idxA, Ref(B), idxB, Ref(C), idxC) |
| 103 | +end |
| 104 | + |
| 105 | +function tblis_tensor_dot(A::tblis_tensor, idxA, B::tblis_tensor, idxB, C::tblis_scalar) |
| 106 | + return tblis_tensor_dot(C_NULL, C_NULL, Ref(A), idxA, Ref(B), idxB, Ref(C)) |
| 107 | +end |
| 108 | + |
| 109 | +end |
0 commit comments