Merge pull request #223 from mransan/wip-encode-backward-with-c-stubs

encode backward and use C stubs

.github/workflows/main.yml

@@ -4,8 +4,6 @@ on:
     branches:
       - master
   pull_request:
-    branches:
-      - master
 jobs:
   run:
     name: Build

.gitignore

@@ -14,3 +14,4 @@ ocaml_protoc.native
 *.orig
 _opam
 *.data
+*.exe

b_run.sh

@@ -0,0 +1,2 @@
+#!/bin/sh
+exec dune exec --profile=release -- benchs/bin/run.exe $@

benchs/benchs.ml

@@ -182,7 +182,7 @@ module Dec = struct
   (* make a buffer with the integers from [0] to [n] inside *)
   let mk_buf_n n : string =
     let enc = Pbrt.Encoder.create () in
-    for i = 0 to n do
+    for i = n downto 0 do
       Pbrt.Encoder.int_as_varint i enc
     done;
     Pbrt.Encoder.to_string enc
@@ -270,7 +270,7 @@ module Dec_bits64 = struct
   (* put the int64 integers from 0 to n in a dec *)
   let mk_buf_n n : string =
     let enc = Pbrt.Encoder.create () in
-    for i = 0 to n do
+    for i = n downto 0 do
       Pbrt.Encoder.int_as_bits64 i enc
     done;
     Pbrt.Encoder.to_string enc
@@ -493,7 +493,7 @@ module Varint_size = struct
 
   module C_while = struct
     external varint_size : (int64[@unboxed]) -> int
-      = "caml_pbrt_varint_size_byte" "caml_pbrt_varint_size"
+      = "b_caml_pbrt_varint_size_byte" "b_caml_pbrt_varint_size"
       [@@noalloc]
 
     let loop ~n =
@@ -596,10 +596,13 @@ module Nested = struct
     val nested : (t -> unit) -> t -> unit
   end
 
+  let[@inline] zigzag i : int64 =
+    Int64.(logxor (shift_left i 1) (shift_right i 63))
+
   module Make_enc (E : ENC) = struct
     let enc_person (p : person) (e : E.t) : unit =
       E.key 1 Bytes (E.string p.name) e;
-      E.key 2 Varint (E.int64_as_varint p.age) e;
+      E.key 2 Varint (E.int64_as_varint @@ zigzag p.age) e;
       ()
 
     let enc_store (st : store) (e : E.t) : unit =
@@ -649,26 +652,49 @@ module Nested = struct
             });
     }
 
-  module Make_bench (E : ENC) = struct
-    include Make_enc (E)
+  module type MK_COMPANY = sig
+    type t
+
+    val name_of_enc : string
+    val create : unit -> t
+    val clear : t -> unit
+    val enc_company : company -> t -> unit
+    val to_string : t -> string
+  end
 
+  module Make_bench_of_mk_company (E : MK_COMPANY) = struct
     let bench company =
-      mk_t (spf "nenc-%s" E.name_of_enc) @@ fun () ->
-      for _i = 1 to 10 do
-        let enc = E.create () in
-        for _j = 1 to 10 do
-          Sys.opaque_identity
-            (E.clear enc;
-             enc_company company enc)
-        done
-      done
+      (* create an encoder that we'll reuse every time, so we can measure
+         the allocations caused purely by the encoding itself *)
+      let enc = E.create () in
+      mk_t E.name_of_enc @@ fun () ->
+      E.clear enc;
+      Sys.opaque_identity (E.enc_company company enc)
 
     let string_of_company c =
       let e = E.create () in
-      enc_company c e;
+      E.enc_company c e;
       E.to_string e
   end
 
+  module Make_bench (E : ENC) = struct
+    module Arg = struct
+      include E
+      include Make_enc (E)
+    end
+
+    include Make_bench_of_mk_company (Arg)
+  end
+
+  module Cur = Make_bench_of_mk_company (struct
+    let name_of_enc = "current"
+
+    include Pbrt.Encoder
+
+    let create () = create ()
+    let enc_company = Foo.encode_pb_company
+  end)
+
   module Basic = Make_bench (struct
     let name_of_enc = "basic-buffer"
 
@@ -1062,7 +1088,7 @@ module Nested = struct
 
     (*
     external varint_size : (int64[@unboxed]) -> int
-      = "caml_pbrt_varint_size_byte" "caml_pbrt_varint_size"
+      = "b_caml_pbrt_varint_size_byte" "b_caml_pbrt_varint_size"
       [@@noalloc]
       *)
 
@@ -1083,7 +1109,7 @@ module Nested = struct
 
     external varint_slice :
       bytes -> (int[@untagged]) -> (int64[@unboxed]) -> unit
-      = "caml_pbrt_varint_byte" "caml_pbrt_varint"
+      = "b_caml_pbrt_varint_byte" "b_caml_pbrt_varint"
       [@@noalloc]
 
     let[@inline] varint (i : int64) (e : t) : unit =
@@ -1132,6 +1158,7 @@ module Nested = struct
       int_as_varint size e
   end)
 
+  let bench_cur = Cur.bench
   let bench_basic = Basic.bench
   let bench_buffers_nested = Buffers_nested.bench
   let bench_from_back = From_back.bench
@@ -1141,19 +1168,23 @@ module Nested = struct
   let pp_size ~n ~depth =
     Printf.printf "bench nested enc: length for n=%d, depth=%d is %d B\n" n
       depth
-      (String.length (Basic.string_of_company @@ mk_company ~n ~depth))
+      (String.length (Cur.string_of_company @@ mk_company ~n ~depth))
 
   (* sanity check *)
   let check ~n ~depth () =
-    let s_basic = Basic.string_of_company (mk_company ~n ~depth) in
-    let s_buffers_nested =
-      Buffers_nested.string_of_company (mk_company ~n ~depth)
-    in
-    let s_from_back = From_back.string_of_company (mk_company ~n ~depth) in
-    let s_from_back2 =
-      From_back_noinline.string_of_company (mk_company ~n ~depth)
-    in
-    let s_from_backc = From_back_c.string_of_company (mk_company ~n ~depth) in
+    let comp = mk_company ~n ~depth in
+    let s_cur = Cur.string_of_company comp in
+    let s_basic = Basic.string_of_company comp in
+
+    (*
+    Printf.printf "###### n=%d, depth=%d\n" n depth;
+    Printf.printf "s_cur[%d]=%S\n" (String.length s_cur) s_cur;
+    Printf.printf "s_basic[%d]=%S\n" (String.length s_basic) s_basic;
+    *)
+    let s_buffers_nested = Buffers_nested.string_of_company comp in
+    let s_from_back = From_back.string_of_company comp in
+    let s_from_back2 = From_back_noinline.string_of_company comp in
+    let s_from_backc = From_back_c.string_of_company comp in
     (*
        Printf.printf "basic:\n(len=%d) %S\n" (String.length s_basic) s_basic;
        Printf.printf "from_back:\n(len=%d) %S\n"
@@ -1169,14 +1200,18 @@ module Nested = struct
         Foo.decode_pb_company dec)
     in
     let c_basic = dec_s s_basic in
+    let c_cur = dec_s s_cur in
     let c_buffers_nested = dec_s s_buffers_nested in
     let c_from_back = dec_s s_from_back in
     let c_from_back2 = dec_s s_from_back2 in
     let c_from_backc = dec_s s_from_backc in
     (*
-    Format.printf "c_basic=%a@." Foo_pp.pp_company c_basic;
-    Format.printf "c_from_back=%a@." Foo_pp.pp_company c_from_back;
+    Format.printf "comp=%a@." Foo.pp_company comp;
+    Format.printf "c_basic=%a@." Foo.pp_company c_basic;
+    Format.printf "c_cur=%a@." Foo.pp_company c_cur;
        *)
+    assert (c_basic = comp);
+    assert (c_basic = c_cur);
     assert (c_basic = c_buffers_nested);
     assert (c_basic = c_from_back);
     assert (c_basic = c_from_back2);
@@ -1186,7 +1221,7 @@ module Nested = struct
   let () =
     List.iter
       (fun (n, depth) -> check ~n ~depth ())
-      [ 1, 3; 2, 4; 10, 1; 20, 2 ]
+      [ 1, 3; 2, 4; 10, 1; 20, 2; 1, 10 ]
 end
 
 let test_nested_enc ~n ~depth =
@@ -1198,6 +1233,7 @@ let test_nested_enc ~n ~depth =
         B.throughputN ~repeat:4 3
           [
             Nested.bench_basic company;
+            Nested.bench_cur company;
             Nested.bench_buffers_nested company;
             Nested.bench_from_back company;
             Nested.bench_from_back_noinline company;

benchs/bin/dune

@@ -0,0 +1,11 @@
+
+(executable
+ (name run)
+ (ocamlopt_flags :standard -inline 100)
+ (libraries pbrt))
+
+(rule
+ (targets foo.ml foo.mli)
+ (deps foo.proto)
+ (action
+  (run ocaml-protoc %{deps} --binary --pp --ml_out .)))

benchs/bin/foo.proto

@@ -0,0 +1,19 @@
+
+syntax = "proto3";
+
+message Person {
+  string name = 1;
+  sint64 age = 2;
+}
+
+message Store {
+  string address = 1;
+  repeated Person employees = 2;
+  repeated Person clients = 3;
+}
+
+message Company {
+  string name = 1;
+  repeated Store stores = 2;
+  repeated Company subsidiaries = 3;
+}

benchs/bin/run.ml

@@ -0,0 +1,64 @@
+open Foo
+
+let spf = Printf.sprintf
+
+(* company, with [n] stores and [2^depth] subsidiaries *)
+let rec mk_company ~n ~depth : company =
+  {
+    name = "bigcorp";
+    subsidiaries =
+      (if depth = 0 then
+        []
+      else (
+        let c = mk_company ~n ~depth:(depth - 1) in
+        [ c; c ]
+      ));
+    stores =
+      List.init n (fun i ->
+          {
+            address = spf "%d foobar street" i;
+            clients =
+              List.init 2 (fun j ->
+                  {
+                    name = spf "client_%d_%d" i j;
+                    age = Int64.of_int ((j mod 30) + 15);
+                  });
+            employees =
+              List.init 2 (fun j ->
+                  {
+                    name = spf "employee_%d_%d" i j;
+                    age = Int64.of_int ((j mod 30) + 18);
+                  });
+          });
+  }
+
+let comp = mk_company ~n:3 ~depth:2
+
+let () =
+  let n = ref 3 in
+  let depth = ref 2 in
+  let iters = ref 100 in
+
+  let opts =
+    [
+      "-n", Arg.Set_int n, " size for data";
+      "--depth", Arg.Set_int depth, " nesting depth for data";
+      "--iters", Arg.Set_int iters, " number of iterations";
+    ]
+    |> Arg.align
+  in
+  Arg.parse opts ignore "";
+
+  Printf.printf "n=%d, depth=%d, iters=%n\n%!" !n !depth !iters;
+  let comp = mk_company ~n:!n ~depth:!depth in
+
+  let enc = Pbrt.Encoder.create ~size:(64 * 1024) () in
+
+  Sys.opaque_identity (Foo.encode_pb_company comp enc);
+  let size = String.length @@ Pbrt.Encoder.to_string enc in
+  Printf.printf "size=%d B\n%!" size;
+
+  for _i = 1 to !iters do
+    Pbrt.Encoder.clear enc;
+    Sys.opaque_identity (Foo.encode_pb_company comp enc)
+  done

benchs/stubs.c

@@ -5,7 +5,7 @@
 #include <stdbool.h>
 #include <stdint.h>
 
-inline int pbrt_varint_size(int64_t i) {
+static inline int pbrt_varint_size(int64_t i) {
   int n = 0;
   while (1) {
     n++;
@@ -18,12 +18,12 @@ inline int pbrt_varint_size(int64_t i) {
 }
 
 // number of bytes for i
-CAMLprim value caml_pbrt_varint_size(int64_t i) {
+CAMLprim value b_caml_pbrt_varint_size(int64_t i) {
   int res = pbrt_varint_size(i);
   return Val_int(res);
 }
 
-CAMLprim value caml_pbrt_varint_size_byte(value v_i) {
+CAMLprim value b_caml_pbrt_varint_size_byte(value v_i) {
   CAMLparam1(v_i);
 
   int64_t i = Int64_val(v_i);
@@ -32,7 +32,7 @@ CAMLprim value caml_pbrt_varint_size_byte(value v_i) {
 }
 
 // write i at str[idx…]
-inline void pbrt_varint(unsigned char *str, int64_t i) {
+static inline void pbrt_varint(unsigned char *str, int64_t i) {
   while (true) {
     int64_t cur = i & 0x7f;
     if (cur == i) {
@@ -63,14 +63,14 @@ inline void pbrt_varint(unsigned char *str, int64_t i) {
 //      done
 
 // write `i` starting at `idx`
-CAMLprim value caml_pbrt_varint(value _str, intnat idx, int64_t i) {
+CAMLprim value b_caml_pbrt_varint(value _str, intnat idx, int64_t i) {
   CAMLparam1(_str);
   char *str = Bytes_val(_str);
   pbrt_varint(str + idx, i);
   CAMLreturn(Val_unit);
 }
 
-CAMLprim value caml_pbrt_varint_bytes(value _str, value _idx, value _i) {
+CAMLprim value b_caml_pbrt_varint_bytes(value _str, value _idx, value _i) {
   CAMLparam3(_str, _idx, _i);
   char *str = Bytes_val(_str);
   int idx = Int_val(_idx);