viva_tensor

-module(viva_tensor@bench@nif).
-compile([no_auto_import, nowarn_unused_vars, nowarn_unused_function, nowarn_nomatch, inline]).
-define(FILEPATH, "src/viva_tensor/bench/nif.gleam").
-export([main/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.

?MODULEDOC(false).

-file("src/viva_tensor/bench/nif.gleam", 388).
?DOC(false).
-spec repeat_char(binary(), integer()) -> binary().
repeat_char(Char, N) ->
    case N =< 0 of
        true ->
            <<""/utf8>>;

        false ->
            <<Char/binary, (repeat_char(Char, N - 1))/binary>>
    end.

-file("src/viva_tensor/bench/nif.gleam", 381).
?DOC(false).
-spec string_length(binary()) -> integer().
string_length(S) ->
    string:length(S).

-file("src/viva_tensor/bench/nif.gleam", 368).
?DOC(false).
-spec float_to_str(float()) -> binary().
float_to_str(F) ->
    Rounded = begin
        _pipe = erlang:round(F * 100.0),
        erlang:float(_pipe)
    end,
    Result = Rounded / 100.0,
    gleam_stdlib:float_to_string(Result).

-file("src/viva_tensor/bench/nif.gleam", 349).
?DOC(false).
-spec pad_speedup(float()) -> binary().
pad_speedup(Speedup) ->
    Str = <<(float_to_str(Speedup))/binary, "x"/utf8>>,
    Len = string_length(Str),
    Padding = 12 - Len,
    <<Str/binary, (repeat_char(<<" "/utf8>>, Padding))/binary>>.

-file("src/viva_tensor/bench/nif.gleam", 356).
?DOC(false).
-spec format_time(float()) -> binary().
format_time(Ms) ->
    case Ms < 1.0 of
        true ->
            <<(float_to_str(Ms * 1000.0))/binary, " μs"/utf8>>;

        false ->
            case Ms < 1000.0 of
                true ->
                    <<(float_to_str(Ms))/binary, " ms"/utf8>>;

                false ->
                    <<(float_to_str(Ms / 1000.0))/binary, " s"/utf8>>
            end
    end.

-file("src/viva_tensor/bench/nif.gleam", 342).
?DOC(false).
-spec pad_time(float()) -> binary().
pad_time(Ms) ->
    Str = format_time(Ms),
    Len = string_length(Str),
    Padding = 12 - Len,
    <<Str/binary, (repeat_char(<<" "/utf8>>, Padding))/binary>>.

-file("src/viva_tensor/bench/nif.gleam", 395).
?DOC(false).
-spec now_microseconds() -> integer().
now_microseconds() ->
    viva_tensor@core@ffi:now_microseconds().

-file("src/viva_tensor/bench/nif.gleam", 337).
?DOC(false).
-spec random_floats(integer()) -> list(float()).
random_floats(N) ->
    _pipe = gleam@list:range(0, N - 1),
    gleam@list:map(_pipe, fun(_) -> viva_tensor@core@ffi:random_uniform() end).

-file("src/viva_tensor/bench/nif.gleam", 277).
?DOC(false).
-spec benchmark_sum(integer()) -> nil.
benchmark_sum(Size) ->
    gleam_stdlib:println(
        <<"  Vector size: "/utf8, (erlang:integer_to_binary(Size))/binary>>
    ),
    gleam_stdlib:println(
        <<"  ┌────────────────────┬──────────────┬──────────────┐"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  │ Backend            │ Time         │ Speedup      │"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  ├────────────────────┼──────────────┼──────────────┤"/utf8>>
    ),
    Data = random_floats(Size),
    Start_pure = now_microseconds(),
    Arr = viva_tensor@core@ffi:list_to_array(Data),
    _ = viva_tensor@core@ffi:array_sum(Arr),
    End_pure = now_microseconds(),
    Pure_time = erlang:float(End_pure - Start_pure) / 1000.0,
    gleam_stdlib:println(
        <<<<"  │ Pure Erlang        │ "/utf8, (pad_time(Pure_time))/binary>>/binary,
            " │ 1.00x        │"/utf8>>
    ),
    Start_c = now_microseconds(),
    _ = viva_tensor@core@ffi:nif_sum(Data),
    End_c = now_microseconds(),
    C_time = erlang:float(End_c - Start_c) / 1000.0,
    C_speedup = case C_time > +0.0 of
        true ->
            case C_time of
                +0.0 -> +0.0;
                -0.0 -> -0.0;
                Gleam@denominator -> Pure_time / Gleam@denominator
            end;

        false ->
            +0.0
    end,
    gleam_stdlib:println(
        <<<<<<<<"  │ Apple Accelerate   │ "/utf8, (pad_time(C_time))/binary>>/binary,
                    " │ "/utf8>>/binary,
                (pad_speedup(C_speedup))/binary>>/binary,
            " │"/utf8>>
    ),
    Start_zig = now_microseconds(),
    _ = viva_tensor@core@ffi:zig_sum(Data),
    End_zig = now_microseconds(),
    Zig_time = erlang:float(End_zig - Start_zig) / 1000.0,
    Zig_speedup = case Zig_time > +0.0 of
        true ->
            case Zig_time of
                +0.0 -> +0.0;
                -0.0 -> -0.0;
                Gleam@denominator@1 -> Pure_time / Gleam@denominator@1
            end;

        false ->
            +0.0
    end,
    gleam_stdlib:println(
        <<<<<<<<"  │ Zig SIMD           │ "/utf8, (pad_time(Zig_time))/binary>>/binary,
                    " │ "/utf8>>/binary,
                (pad_speedup(Zig_speedup))/binary>>/binary,
            " │"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  └────────────────────┴──────────────┴──────────────┘"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>).

-file("src/viva_tensor/bench/nif.gleam", 215).
?DOC(false).
-spec benchmark_dot(integer()) -> nil.
benchmark_dot(Size) ->
    gleam_stdlib:println(
        <<"  Vector size: "/utf8, (erlang:integer_to_binary(Size))/binary>>
    ),
    gleam_stdlib:println(
        <<"  ┌────────────────────┬──────────────┬──────────────┐"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  │ Backend            │ Time         │ Speedup      │"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  ├────────────────────┼──────────────┼──────────────┤"/utf8>>
    ),
    A_data = random_floats(Size),
    B_data = random_floats(Size),
    Start_pure = now_microseconds(),
    A_arr = viva_tensor@core@ffi:list_to_array(A_data),
    B_arr = viva_tensor@core@ffi:list_to_array(B_data),
    _ = viva_tensor@core@ffi:array_dot(A_arr, B_arr),
    End_pure = now_microseconds(),
    Pure_time = erlang:float(End_pure - Start_pure) / 1000.0,
    gleam_stdlib:println(
        <<<<"  │ Pure Erlang        │ "/utf8, (pad_time(Pure_time))/binary>>/binary,
            " │ 1.00x        │"/utf8>>
    ),
    Start_c = now_microseconds(),
    _ = viva_tensor@core@ffi:nif_dot(A_data, B_data),
    End_c = now_microseconds(),
    C_time = erlang:float(End_c - Start_c) / 1000.0,
    C_speedup = case C_time > +0.0 of
        true ->
            case C_time of
                +0.0 -> +0.0;
                -0.0 -> -0.0;
                Gleam@denominator -> Pure_time / Gleam@denominator
            end;

        false ->
            +0.0
    end,
    gleam_stdlib:println(
        <<<<<<<<"  │ Apple Accelerate   │ "/utf8, (pad_time(C_time))/binary>>/binary,
                    " │ "/utf8>>/binary,
                (pad_speedup(C_speedup))/binary>>/binary,
            " │"/utf8>>
    ),
    Start_zig = now_microseconds(),
    _ = viva_tensor@core@ffi:zig_dot(A_data, B_data),
    End_zig = now_microseconds(),
    Zig_time = erlang:float(End_zig - Start_zig) / 1000.0,
    Zig_speedup = case Zig_time > +0.0 of
        true ->
            case Zig_time of
                +0.0 -> +0.0;
                -0.0 -> -0.0;
                Gleam@denominator@1 -> Pure_time / Gleam@denominator@1
            end;

        false ->
            +0.0
    end,
    gleam_stdlib:println(
        <<<<<<<<"  │ Zig SIMD           │ "/utf8, (pad_time(Zig_time))/binary>>/binary,
                    " │ "/utf8>>/binary,
                (pad_speedup(Zig_speedup))/binary>>/binary,
            " │"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  └────────────────────┴──────────────┴──────────────┘"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>).

-file("src/viva_tensor/bench/nif.gleam", 195).
?DOC(false).
-spec time_zig_matmul(
    list(float()),
    list(float()),
    integer(),
    integer(),
    integer()
) -> {list(float()), float()}.
time_zig_matmul(A, B, M, N, K) ->
    Start = now_microseconds(),
    Result = viva_tensor@core@ffi:zig_matmul(A, B, M, N, K),
    End = now_microseconds(),
    case Result of
        {ok, Data} ->
            {Data, erlang:float(End - Start) / 1000.0};

        {error, _} ->
            {[], erlang:float(End - Start) / 1000.0}
    end.

-file("src/viva_tensor/bench/nif.gleam", 179).
?DOC(false).
-spec time_c_matmul(
    list(float()),
    list(float()),
    integer(),
    integer(),
    integer()
) -> {list(float()), float()}.
time_c_matmul(A, B, M, N, K) ->
    Start = now_microseconds(),
    Result = viva_tensor@core@ffi:nif_matmul(A, B, M, N, K),
    End = now_microseconds(),
    case Result of
        {ok, Data} ->
            {Data, erlang:float(End - Start) / 1000.0};

        {error, _} ->
            {[], erlang:float(End - Start) / 1000.0}
    end.

-file("src/viva_tensor/bench/nif.gleam", 163).
?DOC(false).
-spec time_pure_matmul(
    list(float()),
    list(float()),
    integer(),
    integer(),
    integer()
) -> {list(float()), float()}.
time_pure_matmul(A, B, M, N, K) ->
    Start = now_microseconds(),
    A_arr = viva_tensor@core@ffi:list_to_array(A),
    B_arr = viva_tensor@core@ffi:list_to_array(B),
    Result_arr = viva_tensor@core@ffi:array_matmul(A_arr, B_arr, M, N, K),
    Result = viva_tensor@core@ffi:array_to_list(Result_arr),
    End = now_microseconds(),
    {Result, erlang:float(End - Start) / 1000.0}.

-file("src/viva_tensor/bench/nif.gleam", 107).
?DOC(false).
-spec benchmark_matmul(integer(), integer(), integer()) -> nil.
benchmark_matmul(M, N, K) ->
    Size_str = <<<<<<<<<<<<(erlang:integer_to_binary(M))/binary, "x"/utf8>>/binary,
                        (erlang:integer_to_binary(K))/binary>>/binary,
                    " @ "/utf8>>/binary,
                (erlang:integer_to_binary(K))/binary>>/binary,
            "x"/utf8>>/binary,
        (erlang:integer_to_binary(N))/binary>>,
    gleam_stdlib:println(<<"  Matrix: "/utf8, Size_str/binary>>),
    gleam_stdlib:println(
        <<"  ┌────────────────────┬──────────────┬──────────────┐"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  │ Backend            │ Time         │ Speedup      │"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  ├────────────────────┼──────────────┼──────────────┤"/utf8>>
    ),
    A_data = random_floats(M * K),
    B_data = random_floats(K * N),
    {_, Pure_time} = time_pure_matmul(A_data, B_data, M, N, K),
    gleam_stdlib:println(
        <<<<"  │ Pure Erlang        │ "/utf8, (pad_time(Pure_time))/binary>>/binary,
            " │ 1.00x        │"/utf8>>
    ),
    {_, C_time} = time_c_matmul(A_data, B_data, M, N, K),
    C_speedup = case C_time > +0.0 of
        true ->
            case C_time of
                +0.0 -> +0.0;
                -0.0 -> -0.0;
                Gleam@denominator -> Pure_time / Gleam@denominator
            end;

        false ->
            +0.0
    end,
    gleam_stdlib:println(
        <<<<<<<<"  │ Apple Accelerate   │ "/utf8, (pad_time(C_time))/binary>>/binary,
                    " │ "/utf8>>/binary,
                (pad_speedup(C_speedup))/binary>>/binary,
            " │"/utf8>>
    ),
    {_, Zig_time} = time_zig_matmul(A_data, B_data, M, N, K),
    Zig_speedup = case Zig_time > +0.0 of
        true ->
            case Zig_time of
                +0.0 -> +0.0;
                -0.0 -> -0.0;
                Gleam@denominator@1 -> Pure_time / Gleam@denominator@1
            end;

        false ->
            +0.0
    end,
    gleam_stdlib:println(
        <<<<<<<<"  │ Zig SIMD           │ "/utf8, (pad_time(Zig_time))/binary>>/binary,
                    " │ "/utf8>>/binary,
                (pad_speedup(Zig_speedup))/binary>>/binary,
            " │"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  └────────────────────┴──────────────┴──────────────┘"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>).

-file("src/viva_tensor/bench/nif.gleam", 374).
?DOC(false).
-spec bool_to_string(boolean()) -> binary().
bool_to_string(B) ->
    case B of
        true ->
            <<"Yes ✓"/utf8>>;

        false ->
            <<"No ✗"/utf8>>
    end.

-file("src/viva_tensor/bench/nif.gleam", 20).
?DOC(false).
-spec main() -> nil.
main() ->
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(
        <<"╔═══════════════════════════════════════════════════════════════════════════╗"/utf8>>
    ),
    gleam_stdlib:println(
        <<"║           viva_tensor - THREE-WAY NIF BENCHMARK                           ║"/utf8>>
    ),
    gleam_stdlib:println(
        <<"║       Pure Erlang vs Apple Accelerate vs Zig SIMD                         ║"/utf8>>
    ),
    gleam_stdlib:println(
        <<"╚═══════════════════════════════════════════════════════════════════════════╝"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>),
    C_nif_loaded = viva_tensor@core@ffi:is_nif_loaded(),
    Zig_nif_loaded = viva_tensor@core@ffi:zig_is_loaded(),
    C_backend = viva_tensor@core@ffi:nif_backend_info(),
    Zig_backend = viva_tensor@core@ffi:zig_backend_info(),
    gleam_stdlib:println(<<"━━━ BACKEND STATUS ━━━"/utf8>>),
    gleam_stdlib:println(
        <<"  Apple Accelerate NIF: "/utf8,
            (bool_to_string(C_nif_loaded))/binary>>
    ),
    gleam_stdlib:println(<<"    Backend: "/utf8, C_backend/binary>>),
    gleam_stdlib:println(
        <<"  Zig SIMD NIF: "/utf8, (bool_to_string(Zig_nif_loaded))/binary>>
    ),
    gleam_stdlib:println(<<"    Backend: "/utf8, Zig_backend/binary>>),
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(
        <<"                        MATRIX MULTIPLICATION"/utf8>>
    ),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>),
    benchmark_matmul(32, 32, 32),
    benchmark_matmul(64, 64, 64),
    benchmark_matmul(128, 128, 128),
    benchmark_matmul(256, 256, 256),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<"                          DOT PRODUCT"/utf8>>),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>),
    benchmark_dot(10000),
    benchmark_dot(100000),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<"                              SUM"/utf8>>),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>),
    benchmark_sum(100000),
    benchmark_sum(1000000),
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(
        <<"═══════════════════════════════════════════════════════════════════════════"/utf8>>
    ),
    gleam_stdlib:println(
        <<"                         BENCHMARK COMPLETE!"/utf8>>
    ),
    gleam_stdlib:println(
        <<"═══════════════════════════════════════════════════════════════════════════"/utf8>>
    ).