viva_tensor

-module(viva_tensor@bench@gflops).
-compile([no_auto_import, nowarn_unused_vars, nowarn_unused_function, nowarn_nomatch, inline]).
-define(FILEPATH, "src/viva_tensor/bench/gflops.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/gflops.gleam", 369).
?DOC(false).
-spec repeat_string(binary(), integer()) -> binary().
repeat_string(S, N) ->
    case N =< 0 of
        true ->
            <<""/utf8>>;

        false ->
            <<S/binary, (repeat_string(S, N - 1))/binary>>
    end.

-file("src/viva_tensor/bench/gflops.gleam", 360).
?DOC(false).
-spec pad_right(binary(), integer()) -> binary().
pad_right(S, Width) ->
    Len = string:length(S),
    Padding = Width - Len,
    case Padding > 0 of
        true ->
            <<S/binary, (repeat_string(<<" "/utf8>>, Padding))/binary>>;

        false ->
            S
    end.

-file("src/viva_tensor/bench/gflops.gleam", 354).
?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/gflops.gleam", 293).
?DOC(false).
-spec print_gflops_row(binary(), float(), float()) -> nil.
print_gflops_row(Name, Time_ms, Gflops) ->
    Time_str = case Time_ms > +0.0 of
        true ->
            float_to_str(Time_ms);

        false ->
            <<"N/A"/utf8>>
    end,
    Gflops_str = case Gflops > +0.0 of
        true ->
            float_to_str(Gflops);

        false ->
            <<"N/A"/utf8>>
    end,
    gleam_stdlib:println(
        <<<<<<<<<<<<"  │ "/utf8, (pad_right(Name, 18))/binary>>/binary,
                            " │ "/utf8>>/binary,
                        (pad_right(Time_str, 12))/binary>>/binary,
                    " │ "/utf8>>/binary,
                (pad_right(Gflops_str, 12))/binary>>/binary,
            " │"/utf8>>
    ).

-file("src/viva_tensor/bench/gflops.gleam", 286).
?DOC(false).
-spec compute_gflops(integer(), float()) -> float().
compute_gflops(Flops, Time_ms) ->
    case Time_ms > +0.0 of
        true ->
            case (Time_ms * 1000000.0) of
                +0.0 -> +0.0;
                -0.0 -> -0.0;
                Gleam@denominator -> erlang:float(Flops) / Gleam@denominator
            end;

        false ->
            +0.0
    end.

-file("src/viva_tensor/bench/gflops.gleam", 317).
?DOC(false).
-spec time_it(fun(() -> LAD)) -> {float(), LAD}.
time_it(F) ->
    Start = viva_tensor@core@ffi:now_microseconds(),
    Result = F(),
    End = viva_tensor@core@ffi:now_microseconds(),
    Time_ms = erlang:float(End - Start) / 1000.0,
    {Time_ms, Result}.

-file("src/viva_tensor/bench/gflops.gleam", 325).
?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/gflops.gleam", 231).
?DOC(false).
-spec benchmark_gflops(integer(), integer(), integer()) -> nil.
benchmark_gflops(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>>,
    Flops = ((2 * M) * N) * K,
    gleam_stdlib:println(
        <<<<<<<<"  "/utf8, Size_str/binary>>/binary, " ("/utf8>>/binary,
                (erlang:integer_to_binary(Flops))/binary>>/binary,
            " FLOPs):"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  ┌────────────────────┬──────────────┬──────────────┐"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  │ Backend            │ Time (ms)    │ GFLOPS       │"/utf8>>
    ),
    gleam_stdlib:println(
        <<"  ├────────────────────┼──────────────┼──────────────┤"/utf8>>
    ),
    A_data = random_floats(M * K),
    B_data = random_floats(K * N),
    A_arr = viva_tensor@core@ffi:list_to_array(A_data),
    B_arr = viva_tensor@core@ffi:list_to_array(B_data),
    {Pure_time, _} = time_it(
        fun() -> viva_tensor@core@ffi:array_matmul(A_arr, B_arr, M, N, K) end
    ),
    Pure_gflops = compute_gflops(Flops, Pure_time),
    print_gflops_row(<<"Pure Erlang"/utf8>>, Pure_time, Pure_gflops),
    case viva_tensor@core@ffi:is_nif_loaded() of
        true ->
            {Accel_time, _} = time_it(
                fun() ->
                    viva_tensor@core@ffi:nif_matmul(A_data, B_data, M, N, K)
                end
            ),
            Accel_gflops = compute_gflops(Flops, Accel_time),
            print_gflops_row(
                <<"Apple Accelerate"/utf8>>,
                Accel_time,
                Accel_gflops
            );

        false ->
            print_gflops_row(<<"Apple Accelerate"/utf8>>, +0.0, +0.0)
    end,
    case viva_tensor@core@ffi:zig_is_loaded() of
        true ->
            {Zig_time, _} = time_it(
                fun() ->
                    viva_tensor@core@ffi:zig_matmul(A_data, B_data, M, N, K)
                end
            ),
            Zig_gflops = compute_gflops(Flops, Zig_time),
            print_gflops_row(<<"Zig SIMD"/utf8>>, Zig_time, Zig_gflops);

        false ->
            print_gflops_row(<<"Zig SIMD"/utf8>>, +0.0, +0.0)
    end,
    gleam_stdlib:println(
        <<"  └────────────────────┴──────────────┴──────────────┘"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>).

-file("src/viva_tensor/bench/gflops.gleam", 347).
?DOC(false).
-spec result_to_string(boolean()) -> binary().
result_to_string(Ok) ->
    case Ok of
        true ->
            <<"✓ PASS"/utf8>>;

        false ->
            <<"✗ FAIL"/utf8>>
    end.

-file("src/viva_tensor/bench/gflops.gleam", 339).
?DOC(false).
-spec float_close(float(), float(), float()) -> boolean().
float_close(A, B, Tol) ->
    Diff = case (A - B) < +0.0 of
        true ->
            B - A;

        false ->
            A - B
    end,
    Diff < Tol.

-file("src/viva_tensor/bench/gflops.gleam", 185).
?DOC(false).
-spec test_sum_correctness() -> nil.
test_sum_correctness() ->
    gleam_stdlib:println(<<"  Sum Reduction Correctness:"/utf8>>),
    Data = [1.0, 2.0, 3.0, 4.0, 5.0],
    Expected = 15.0,
    Arr = viva_tensor@core@ffi:list_to_array(Data),
    Pure_result = viva_tensor@core@ffi:array_sum(Arr),
    Pure_ok = float_close(Pure_result, Expected, 1.0e-10),
    gleam_stdlib:println(
        <<"    Pure Erlang:      "/utf8, (result_to_string(Pure_ok))/binary>>
    ),
    case viva_tensor@core@ffi:is_nif_loaded() of
        true ->
            case viva_tensor@core@ffi:nif_sum(Data) of
                {ok, Result} ->
                    Ok = float_close(Result, Expected, 1.0e-10),
                    gleam_stdlib:println(
                        <<"    Apple Accelerate: "/utf8,
                            (result_to_string(Ok))/binary>>
                    );

                {error, _} ->
                    gleam_stdlib:println(
                        <<"    Apple Accelerate: ✗ Error"/utf8>>
                    )
            end;

        false ->
            gleam_stdlib:println(
                <<"    Apple Accelerate: - Not available"/utf8>>
            )
    end,
    case viva_tensor@core@ffi:zig_is_loaded() of
        true ->
            case viva_tensor@core@ffi:zig_sum(Data) of
                {ok, Result@1} ->
                    Ok@1 = float_close(Result@1, Expected, 1.0e-10),
                    gleam_stdlib:println(
                        <<"    Zig SIMD:         "/utf8,
                            (result_to_string(Ok@1))/binary>>
                    );

                {error, _} ->
                    gleam_stdlib:println(
                        <<"    Zig SIMD:         ✗ Error"/utf8>>
                    )
            end;

        false ->
            gleam_stdlib:println(
                <<"    Zig SIMD:         - Not available"/utf8>>
            )
    end.

-file("src/viva_tensor/bench/gflops.gleam", 139).
?DOC(false).
-spec test_dot_correctness() -> nil.
test_dot_correctness() ->
    gleam_stdlib:println(<<"  Dot Product Correctness:"/utf8>>),
    A_data = [1.0, 2.0, 3.0, 4.0],
    B_data = [5.0, 6.0, 7.0, 8.0],
    Expected = 70.0,
    A_arr = viva_tensor@core@ffi:list_to_array(A_data),
    B_arr = viva_tensor@core@ffi:list_to_array(B_data),
    Pure_result = viva_tensor@core@ffi:array_dot(A_arr, B_arr),
    Pure_ok = float_close(Pure_result, Expected, 1.0e-10),
    gleam_stdlib:println(
        <<"    Pure Erlang:      "/utf8, (result_to_string(Pure_ok))/binary>>
    ),
    case viva_tensor@core@ffi:is_nif_loaded() of
        true ->
            case viva_tensor@core@ffi:nif_dot(A_data, B_data) of
                {ok, Result} ->
                    Ok = float_close(Result, Expected, 1.0e-10),
                    gleam_stdlib:println(
                        <<"    Apple Accelerate: "/utf8,
                            (result_to_string(Ok))/binary>>
                    );

                {error, _} ->
                    gleam_stdlib:println(
                        <<"    Apple Accelerate: ✗ Error"/utf8>>
                    )
            end;

        false ->
            gleam_stdlib:println(
                <<"    Apple Accelerate: - Not available"/utf8>>
            )
    end,
    case viva_tensor@core@ffi:zig_is_loaded() of
        true ->
            case viva_tensor@core@ffi:zig_dot(A_data, B_data) of
                {ok, Result@1} ->
                    Ok@1 = float_close(Result@1, Expected, 1.0e-10),
                    gleam_stdlib:println(
                        <<"    Zig SIMD:         "/utf8,
                            (result_to_string(Ok@1))/binary>>
                    );

                {error, _} ->
                    gleam_stdlib:println(
                        <<"    Zig SIMD:         ✗ Error"/utf8>>
                    )
            end;

        false ->
            gleam_stdlib:println(
                <<"    Zig SIMD:         - Not available"/utf8>>
            )
    end,
    gleam_stdlib:println(<<""/utf8>>).

-file("src/viva_tensor/bench/gflops.gleam", 330).
?DOC(false).
-spec check_close(list(float()), list(float()), float()) -> boolean().
check_close(Actual, Expected, Tol) ->
    case erlang:length(Actual) =:= erlang:length(Expected) of
        true ->
            _pipe = gleam@list:map2(
                Actual,
                Expected,
                fun(A, E) -> float_close(A, E, Tol) end
            ),
            gleam@list:all(_pipe, fun(X) -> X end);

        false ->
            false
    end.

-file("src/viva_tensor/bench/gflops.gleam", 87).
?DOC(false).
-spec test_matmul_correctness() -> nil.
test_matmul_correctness() ->
    gleam_stdlib:println(<<"  Matrix Multiplication Correctness:"/utf8>>),
    A_data = [1.0, 2.0, 3.0, 4.0],
    B_data = [5.0, 6.0, 7.0, 8.0],
    Expected = [19.0, 22.0, 43.0, 50.0],
    A = gleam@result:unwrap(
        viva_tensor@core@tensor:new(A_data, [2, 2]),
        viva_tensor@core@tensor:zeros([2, 2])
    ),
    B = gleam@result:unwrap(
        viva_tensor@core@tensor:new(B_data, [2, 2]),
        viva_tensor@core@tensor:zeros([2, 2])
    ),
    Pure_result = viva_tensor@core@ops:matmul_fast(A, B),
    Pure_ok = case Pure_result of
        {ok, T} ->
            check_close(viva_tensor@core@tensor:to_list(T), Expected, 1.0e-10);

        {error, _} ->
            false
    end,
    gleam_stdlib:println(
        <<"    Pure Erlang:      "/utf8, (result_to_string(Pure_ok))/binary>>
    ),
    case viva_tensor@core@ffi:is_nif_loaded() of
        true ->
            case viva_tensor@core@ffi:nif_matmul(A_data, B_data, 2, 2, 2) of
                {ok, Result} ->
                    Ok = check_close(Result, Expected, 1.0e-10),
                    gleam_stdlib:println(
                        <<"    Apple Accelerate: "/utf8,
                            (result_to_string(Ok))/binary>>
                    );

                {error, _} ->
                    gleam_stdlib:println(
                        <<"    Apple Accelerate: ✗ Error"/utf8>>
                    )
            end;

        false ->
            gleam_stdlib:println(
                <<"    Apple Accelerate: - Not available"/utf8>>
            )
    end,
    case viva_tensor@core@ffi:zig_is_loaded() of
        true ->
            case viva_tensor@core@ffi:zig_matmul(A_data, B_data, 2, 2, 2) of
                {ok, Result@1} ->
                    Ok@1 = check_close(Result@1, Expected, 1.0e-10),
                    gleam_stdlib:println(
                        <<"    Zig SIMD:         "/utf8,
                            (result_to_string(Ok@1))/binary>>
                    );

                {error, _} ->
                    gleam_stdlib:println(
                        <<"    Zig SIMD:         ✗ Error"/utf8>>
                    )
            end;

        false ->
            gleam_stdlib:println(
                <<"    Zig SIMD:         - Not available"/utf8>>
            )
    end,
    gleam_stdlib:println(<<""/utf8>>).

-file("src/viva_tensor/bench/gflops.gleam", 21).
?DOC(false).
-spec main() -> nil.
main() ->
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(
        <<"╔═══════════════════════════════════════════════════════════════════════════╗"/utf8>>
    ),
    gleam_stdlib:println(
        <<"║              viva_tensor - GFLOPS BENCHMARK                               ║"/utf8>>
    ),
    gleam_stdlib:println(
        <<"║          Performance measurement & numerical correctness                  ║"/utf8>>
    ),
    gleam_stdlib:println(
        <<"╚═══════════════════════════════════════════════════════════════════════════╝"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(<<"━━━ BACKEND STATUS ━━━"/utf8>>),
    gleam_stdlib:println(viva_tensor@core@ops:all_backends_info()),
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<"                     NUMERICAL CORRECTNESS"/utf8>>),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>),
    test_matmul_correctness(),
    test_dot_correctness(),
    test_sum_correctness(),
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<"                        GFLOPS MEASUREMENT"/utf8>>),
    gleam_stdlib:println(
        <<"━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"/utf8>>
    ),
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(<<"  Matrix multiplication: 2*M*N*K FLOPs"/utf8>>),
    gleam_stdlib:println(<<""/utf8>>),
    benchmark_gflops(128, 128, 128),
    benchmark_gflops(256, 256, 256),
    benchmark_gflops(512, 512, 512),
    gleam_stdlib:println(<<""/utf8>>),
    gleam_stdlib:println(
        <<"═══════════════════════════════════════════════════════════════════════════"/utf8>>
    ),
    gleam_stdlib:println(<<"                       BENCHMARK COMPLETE!"/utf8>>),
    gleam_stdlib:println(
        <<"═══════════════════════════════════════════════════════════════════════════"/utf8>>
    ).