viva_tensor

-module(viva_tensor@blas).
-compile([no_auto_import, nowarn_unused_vars, nowarn_unused_function, nowarn_nomatch, inline]).
-define(FILEPATH, "src/viva_tensor/blas.gleam").
-export([detect_backend/0, get_cpu_topology/0, configure_threads/1, set_affinity/1, auto_configure/0, backend_name/1, expected_gflops/2]).
-export_type([blas_backend/0, cpu_topology/0]).

-if(?OTP_RELEASE >= 27).
-define(MODULEDOC(Str), -moduledoc(Str)).
-define(DOC(Str), -doc(Str)).
-else.
-define(MODULEDOC(Str), -compile([])).
-define(DOC(Str), -compile([])).
-endif.

-type blas_backend() :: intel_m_k_l | open_b_l_a_s | zig_s_i_m_d | unknown.

-type cpu_topology() :: {cpu_topology,
        integer(),
        integer(),
        integer(),
        integer(),
        boolean(),
        boolean()}.

-file("src/viva_tensor/blas.gleam", 34).
?DOC(" Detect the best available BLAS backend\n").
-spec detect_backend() -> blas_backend().
detect_backend() ->
    viva_tensor_blas:detect_backend().

-file("src/viva_tensor/blas.gleam", 38).
?DOC(" Get CPU topology for tuning\n").
-spec get_cpu_topology() -> cpu_topology().
get_cpu_topology() ->
    viva_tensor_blas:get_cpu_topology().

-file("src/viva_tensor/blas.gleam", 42).
?DOC(" Configure MKL/OpenBLAS threads for optimal performance\n").
-spec configure_threads(integer()) -> {ok, nil} | {error, binary()}.
configure_threads(Num_threads) ->
    viva_tensor_blas:configure_threads(Num_threads).

-file("src/viva_tensor/blas.gleam", 46).
?DOC(" Set thread affinity (scatter for hybrid CPUs)\n").
-spec set_affinity(binary()) -> {ok, nil} | {error, binary()}.
set_affinity(Mode) ->
    viva_tensor_blas:set_affinity(Mode).

-file("src/viva_tensor/blas.gleam", 49).
?DOC(" Auto-configure for maximum performance\n").
-spec auto_configure() -> {ok, blas_backend()} | {error, binary()}.
auto_configure() ->
    Backend = viva_tensor_blas:detect_backend(),
    Topo = viva_tensor_blas:get_cpu_topology(),
    Optimal_threads = erlang:element(3, Topo),
    case viva_tensor_blas:configure_threads(Optimal_threads) of
        {ok, _} ->
            case viva_tensor_blas:set_affinity(<<"scatter"/utf8>>) of
                {ok, _} ->
                    {ok, Backend};

                {error, E} ->
                    {error, E}
            end;

        {error, E@1} ->
            {error, E@1}
    end.

-file("src/viva_tensor/blas.gleam", 69).
?DOC(" Get backend name as string\n").
-spec backend_name(blas_backend()) -> binary().
backend_name(Backend) ->
    case Backend of
        intel_m_k_l ->
            <<"Intel MKL (800+ GFLOPS)"/utf8>>;

        open_b_l_a_s ->
            <<"OpenBLAS DYNAMIC_ARCH (500+ GFLOPS)"/utf8>>;

        zig_s_i_m_d ->
            <<"Zig SIMD GEMM (400+ GFLOPS)"/utf8>>;

        unknown ->
            <<"Unknown"/utf8>>
    end.

-file("src/viva_tensor/blas.gleam", 79).
?DOC(" Performance tier for the backend\n").
-spec expected_gflops(blas_backend(), integer()) -> integer().
expected_gflops(Backend, Matrix_size) ->
    case Backend of
        intel_m_k_l ->
            case Matrix_size of
                S when S >= 4000 ->
                    800;

                S@1 when S@1 >= 2000 ->
                    600;

                S@2 when S@2 >= 1000 ->
                    450;

                _ ->
                    300
            end;

        open_b_l_a_s ->
            case Matrix_size of
                S@3 when S@3 >= 4000 ->
                    500;

                S@4 when S@4 >= 2000 ->
                    400;

                S@5 when S@5 >= 1000 ->
                    300;

                _ ->
                    200
            end;

        zig_s_i_m_d ->
            case Matrix_size of
                S@6 when S@6 >= 4000 ->
                    450;

                S@7 when S@7 >= 2000 ->
                    350;

                S@8 when S@8 >= 1000 ->
                    250;

                _ ->
                    150
            end;

        unknown ->
            100
    end.