benji/lib94
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

allow more control over init_round, rewrite tabulator and make it run every possible separation and turn order exactly once

Browse source
This commit is contained in:
Benji Dial 2023-11-25 18:15:30 -05:00
parent 5141b88157
commit eab56a7f00
11 changed files with 283 additions and 265 deletions

4
bench/bench_window.cpp
View file

@ -194,7 +194,7 @@ void bench_window::on_click_new_round() {
.bnumber = 0
});
if (!lib94::init_round(warriors.data(), warriors.size())) {
if (!lib94::init_round(warriors.data(), warriors.size(), 0, true)) {
Gtk::MessageDialog *md = new Gtk::MessageDialog("warriors do not fit in core; removing last warrior");
md->set_transient_for(*this);
md->set_modal();
@ -208,7 +208,7 @@ void bench_window::on_click_new_round() {
update_ui();
return;
}
assert(lib94::init_round(warriors.data(), warriors.size()));
assert(lib94::init_round(warriors.data(), warriors.size(), 0, true));
}
core.mut.lock();

25
include/lib94/lib94.hpp
View file

@ -91,15 +91,22 @@ namespace lib94 {
//does not effect the address sets.
void clear_core_random();
//clears the address sets, places the supplied warriors into the core, and starts one process for
//each supplied warrior at its org. the count parameter specifies the number of warriors in the
//array. each of the warriors in the array must not be deleted for the duration of the round, but
//the array itself may be deleted after this function returns. on success, this function returns
//true. on failure (i.e. when the warriors do not all fit into the core), this function returns
//false. note that this function does not clear the core before placing the warriors, so you may
//want to call clear_core or clear_core_random before calling this. note also that you probably
//want to call seed_prng, for example with the current time, before the first time you call this.
bool init_round(const warrior *const *warriors, size_t count);
//clears the address sets, places the supplied warriors into the core, and starts one
//process for each supplied warrior at its org. the count parameter specifies the number
//of warriors in the array. each of the warriors in the array must not be deleted for
//the duration of the round, but the array itself may be deleted after this function
//returns. if the offsets parameter is not null, it specifies where in the core the first
//instruction of each warrior is put. if it is null, the warriors are placed at random
//non-overlapping locations, respecting the order that they appear in the array. if
//shuffle is true, the warriors' turn order is decided randomly. if shuffle is false,
//the turn order is the order that they appear in the warriors array. on success, this
//function returns true. on failure (i.e. when offsets is null and the warriors do not
//all fit into the core), this function returns false. note that this function does not
//clear the core before placing the warriors, so you may want to call clear_core or
//clear_core_random before calling this. note also that if you are passing null for
//offsets or setting shuffle to true, you probably want to call seed_prng, for example
//with the current time, before the first time you call this.
bool init_round(const warrior *const *warriors, size_t count, const number_t *offsets, bool shuffle);
//returns the number of warriors who have at least one process
size_t alive_warrior_count();

56
lib94/core.cpp
View file

@ -55,51 +55,59 @@ namespace lib94 {
static std::vector<warrior_info> warrior_infos;
std::deque<warrior_info *> alive_warriors;
bool init_round(const warrior *const *warriors, size_t count) {
bool init_round(const warrior *const *warriors, size_t count, const number_t *offsets, bool shuffle) {
clear_address_sets();
warrior_infos.clear();
alive_warriors = std::deque<warrior_info *>();
std::vector<number_t> gap_sizes;
gap_sizes.resize(count);
std::vector<number_t> offsets_vector;
std::uniform_real_distribution phi_dist(0.0, 1.0);
number_t gap_remaining = LIB94_CORE_SIZE;
if (!offsets) {
offsets_vector.resize(count);
offsets_vector[0] = 0;
std::uniform_real_distribution phi_dist(0.0, 1.0);
number_t gap_remaining = LIB94_CORE_SIZE;
for (size_t i = 0; i < count; ++i) {
number_t wlength = warriors[i]->instructions.size();
if (wlength > gap_remaining)
return false;
gap_remaining -= wlength;
}
for (size_t i = 1; i < count; ++i) {
number_t gap_size = std::floor(gap_remaining * (1.0 - std::pow(phi_dist(prng), 1.0 / (count - i))));
offsets_vector[i] = offsets_vector[i - 1] + warriors[i - 1]->instructions.size() + gap_size;
gap_remaining -= gap_size;
}
offsets = offsets_vector.data();
for (size_t i = 0; i < count; ++i) {
number_t wlength = warriors[i]->instructions.size();
if (wlength > gap_remaining)
return false;
gap_remaining -= wlength;
}
for (size_t i = 0; i < count; ++i) {
number_t gap_size = std::floor(gap_remaining * (1.0 - std::pow(phi_dist(prng), 1.0 / (count - i))));
gap_sizes[i] = gap_size;
gap_remaining -= gap_size;
}
size_t place_at = 0;
for (size_t i = 0; i < count; ++i) {
const warrior *w = warriors[i];
const number_t offset = offsets[i];
for (number_t i = 0; i < (number_t)w->instructions.size(); ++i) {
assert(place_at + i < LIB94_CORE_SIZE);
core[place_at + i] = w->instructions[i];
add_written_instruction(core + place_at + i);
assert(offset + i < LIB94_CORE_SIZE);
core[offset + i] = w->instructions[i];
add_written_instruction(core + offset + i);
}
warrior_infos.push_back({});
warrior_info *wi = &warrior_infos.back();
wi->the_warrior = w;
assert(place_at + w->org < LIB94_CORE_SIZE);
wi->processes.push_back(place_at + w->org);
assert(offset + w->org < LIB94_CORE_SIZE);
wi->processes.push_back(offsets[i] + w->org);
place_at += w->instructions.size() + gap_sizes[i];
}
std::shuffle(warrior_infos.begin(), warrior_infos.end(), prng);
if (shuffle)
std::shuffle(warrior_infos.begin(), warrior_infos.end(), prng);
for (warrior_info &wi : warrior_infos)
alive_warriors.push_back(&wi);

2
makefile
View file

@ -16,7 +16,7 @@ bin/bench: obj/bench/main.o obj/bench/bench_window.o obj/bench/core_widget.o obj
@mkdir -p $(dir $@)
g++ ${CPP_ARGS} $^ ${GTKMM_LD_ARGS} -o $@
bin/tabulator-mpi: obj/tabulator-mpi/main.o obj/tabulator-mpi/head.o obj/tabulator-mpi/worker.o obj/lib94.o
bin/tabulator-mpi: obj/tabulator-mpi/source.o obj/lib94.o
@mkdir -p $(dir $@)
g++ ${CPP_ARGS} $^ ${MPI_LD_ARGS} -o $@

6
readme.txt
View file

@ -28,9 +28,9 @@ To open bench, just run bin/bench after building as above.
=== tabulator ===
The "tabulator" program runs every possible pairing of warriors from a selection against each other a number of times, and then shows
the number of wins of each warrior against each other warrior in a table format. This program uses MPI to run batches of these rounds
in different processes, and communicate the results back to a head process.
The "tabulator" program runs every possible pairing of warriors from a selection against each other with ever possible separation,
and then shows the number of wins of each warrior against each other warrior in a table format. This program uses MPI to run batches
of these rounds in different processes, and communicate the results back to a head process.
To run all of the included warriors against each other, run
mpirun bin/tabulator-mpi warriors/*.red

3
tabulator-mpi/constants.hpp
View file

@ -1,3 +0,0 @@
#define STEPS_TO_TIE 1000000
#define ROUNDS_PER_CHUNK 200
#define CHUNKS_PER_PAIR 5

112
tabulator-mpi/head.cpp
View file

@ -1,112 +0,0 @@
#include <lib94/lib94.hpp>
#include <cstdio>
#include <mpi.h>
#include "constants.hpp"
static int **wins;
static int get_result() {
int result[4];
MPI_Status status;
MPI_Recv(result, 4, MPI_INT, MPI_ANY_SOURCE, 1, MPI_COMM_WORLD, &status);
if (result[0] != -1) {
wins[result[0]][result[1]] += result[2];
wins[result[1]][result[0]] += result[3];
}
return status.MPI_SOURCE;
}
void head_main(int comm_size, int warrior_count, const lib94::warrior *const *warriors) {
wins = new int *[warrior_count];
for (int i = 0; i < warrior_count; ++i) {
wins[i] = new int[warrior_count];
for (int j = 0; j < warrior_count; ++j)
wins[i][j] = 0;
}
int chunks = warrior_count * (warrior_count - 1) / 2 * CHUNKS_PER_PAIR;
int on_chunk = 0;
int right_name_width = 0;
for (int i = 0; i < warrior_count; ++i)
if ((int)warriors[i]->name.size() > right_name_width)
right_name_width = warriors[i]->name.size();
int rank_width = std::max(std::string("rank").size(), std::to_string(comm_size - 1).size());
int right_round_width = std::to_string(CHUNKS_PER_PAIR * ROUNDS_PER_CHUNK).size();
int right_chunk_width = std::to_string(chunks).size();
int left_name_width = std::max(right_name_width, (int)std::string("match").size() - right_name_width - 4);
int left_round_width = std::max((int)std::to_string((CHUNKS_PER_PAIR - 1) * ROUNDS_PER_CHUNK + 1).size(), (int)std::string("rounds").size() - right_round_width - 3);
int left_chunk_width = std::max(right_chunk_width, (int)std::string("chunk").size() - right_chunk_width - 3);
fprintf(stderr, "\x1b""7\x1b[?47h\x1b[?25l\x1b[2J\x1b[0H");
fprintf(stderr, "%*s | %*s | %*s | %*s", rank_width, "rank",
left_name_width + 4 + right_name_width, "match",
left_round_width + 3 + right_round_width, "rounds",
left_chunk_width + 3 + right_chunk_width, "chunk");
for (int i = 0; i < warrior_count; ++i)
for (int j = i + 1; j < warrior_count; ++j)
for (int x = 0; x < CHUNKS_PER_PAIR; ++x) {
++on_chunk;
int rank = get_result();
int message[4] = {i, j, x * ROUNDS_PER_CHUNK};
MPI_Send(message, 4, MPI_INT, rank, 0, MPI_COMM_WORLD);
fprintf(stderr, "\x1b[%d;0H%*d | %*s vs %*s | %*d - %*d | %*d / %*d\x1b[0K", rank + 1, rank_width, rank,
left_name_width, warriors[i]->name.c_str(), right_name_width, warriors[j]->name.c_str(),
left_round_width, x * ROUNDS_PER_CHUNK + 1, right_round_width, x * ROUNDS_PER_CHUNK + ROUNDS_PER_CHUNK,
left_chunk_width, on_chunk, right_chunk_width, chunks);
}
for (int i = 0; i < comm_size - 1; ++i) {
int rank = get_result();
int message[4] = {-1};
MPI_Send(message, 4, MPI_INT, rank, 0, MPI_COMM_WORLD);
fprintf(stderr, "\x1b[%d;0H%*d | %*s | %*s | %*s\x1b[0K",
rank + 1, rank_width, rank,
left_name_width + 4 + right_name_width, "",
left_round_width + 3 + right_round_width, "",
left_chunk_width + 3 + right_chunk_width, "done");
}
fprintf(stderr, "\x1b[?25h\x1b[?47l\x1b""8");
int *tab_widths = new int[warrior_count + 1];
tab_widths[0] = 0;
for (int i = 0; i < warrior_count; ++i) {
int len = warriors[i]->name.size();
if (len > tab_widths[0])
tab_widths[0] = len;
tab_widths[i + 1] = len > 5 ? len : 5;
}
printf(" %*s", tab_widths[0], "");
for (int j = 0; j < warrior_count; ++j)
printf(" | %*s", tab_widths[j + 1], warriors[j]->name.c_str());
putchar('\n');
putchar('-');
for (int x = 0; x < tab_widths[0]; ++x)
putchar('-');
for (int j = 0; j < warrior_count; ++j) {
printf("-+-");
for (int x = 0; x < tab_widths[j + 1]; ++x)
putchar('-');
}
printf("-\n");
for (int i = 0; i < warrior_count; ++i) {
printf(" %*s", tab_widths[0], warriors[i]->name.c_str());
for (int j = 0; j < warrior_count; ++j)
if (i == j)
printf(" | %*s", tab_widths[j + 1], "x");
else
printf(" | %*d", tab_widths[j + 1], wins[i][j]);
putchar('\n');
}
}

60
tabulator-mpi/main.cpp
View file

@ -1,60 +0,0 @@
#include <lib94/lib94.hpp>
#include <fstream>
#include <cstdio>
#include <mpi.h>
void head_main(int comm_size, int warrior_count, const lib94::warrior *const *warriors);
void worker_main(const lib94::warrior *const *warriors);
const lib94::warrior *load_warrior(const char *file) {
std::ifstream stream(file);
if (!stream) {
fprintf(stderr, "could not open %s\n", file);
exit(1);
}
std::string source(std::istreambuf_iterator<char>(stream), {});
try {
return lib94::compile_warrior(source);
}
catch (const lib94::compiler_exception &ex) {
fprintf(stderr, "error in %s on line %u: %s\n", file, ex.source_line_number, ex.message.c_str());
exit(1);
}
}
int main(int argc, char **argv) {
MPI_Init(&argc, &argv);
int comm_rank, comm_size;
MPI_Comm_rank(MPI_COMM_WORLD, &comm_rank);
MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
if (comm_size < 2) {
fprintf(stderr, "at least two processes are required\n");
return 1;
}
const lib94::warrior **warriors = new const lib94::warrior *[argc - 1];
for (int i = 0; i < argc - 1; ++i)
warriors[i] = load_warrior(argv[i + 1]);
for (int i = 0; i < argc - 1; ++i)
for (int j = i + 1; j < argc - 1; ++j) {
const lib94::warrior *wbuf[2] = {warriors[i], warriors[j]};
if (!lib94::init_round(wbuf, 2)) {
fprintf(stderr, "warriors do not fit in core\n");
return 1;
}
}
if (comm_rank == 0)
head_main(comm_size, argc - 1, warriors);
else
worker_main(warriors);
MPI_Finalize();
}

15
tabulator-mpi/mpi.txt Normal file
View file

@ -0,0 +1,15 @@
message tag 0:
head -> worker
four ints:
warrior 1 index
warrior 2 index
start round
end round (exclusive, 0 for exit)
message tag 1:
worker -> head
four ints:
warrior 1 index
warrior 2 index
w1 wins against w2
w2 wins against w1

214
tabulator-mpi/source.cpp Normal file
View file

@ -0,0 +1,214 @@
#include <lib94/lib94.hpp>
#include <iostream>
#include <fstream>
#include <sstream>
#include <cstdio>
#include <vector>
#include <mpi.h>
const int default_rounds_per_chunk = 250;
const int steps_to_tie = 1000000;
int error(std::string msg, int rank) {
if (rank == 0) {
std::cerr << msg << std::endl;
return 1;
}
return 0;
}
int main(int argc, char **argv) {
MPI_Init(&argc, &argv);
int rank;
int size;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (size < 2)
error("this must be run under mpi with at least two processes.", rank);
std::vector<std::string> filenames = {};
int rounds_per_chunk = default_rounds_per_chunk;
for (int i = 1; i < argc; ++i) {
if (atoi(argv[i]) > 0)
rounds_per_chunk = atoi(argv[i]);
else
filenames.push_back(argv[i]);
}
if (filenames.size() == 0)
return error("no files specified.", rank);
if (filenames.size() == 1)
return error("only one file specified.", rank);
int count = filenames.size();
lib94::warrior **warriors = new lib94::warrior *[count];
for (int i = 0; i < count; ++i) {
std::ifstream file(filenames[i]);
if (!file)
return error("could not open " + filenames[i] + ".", rank);
std::stringstream stream;
stream << file.rdbuf();
try {
warriors[i] = lib94::compile_warrior(stream.str());
}
catch (const lib94::compiler_exception &ex) {
return error("could not compile " + filenames[i] + ": " + ex.message + " on line " + std::to_string(ex.source_line_number) + ".", rank);
}
}
//w1 * count + w2
int *placements = new int[count * count];
for (int i = 0; i < count; ++i)
for (int j = 0; j < count; ++j) {
int p = LIB94_CORE_SIZE - warriors[i]->instructions.size() - warriors[j]->instructions.size() + 1;
if (p <= 0)
return error(filenames[i] + " and " + filenames[j] + " do not fit in core together.", rank);
placements[i * count + j] = p;
}
if (rank == 0) {
std::cerr << "\x1b""7\x1b[?47h\x1b[2J" << std::flush;
//w1 * count + w2
int *wins_array = new int[count * count];
for (int i = 0; i < count * count; ++i)
wins_array[i] = 0;
for (int i = 0; i < count; ++i)
for (int j = 0; j < count; ++j) {
if (i == j)
continue;
int rounds = placements[i * count + j];
for (int r = 0; r < rounds; r += rounds_per_chunk) {
int e = std::min(r + rounds_per_chunk, rounds);
MPI_Status status;
int buffer[4];
MPI_Recv(buffer, 4, MPI_INT, MPI_ANY_SOURCE, 1, MPI_COMM_WORLD, &status);
int source = status.MPI_SOURCE;
wins_array[buffer[0] * count + buffer[1]] += buffer[2];
wins_array[buffer[1] * count + buffer[0]] += buffer[3];
buffer[0] = i;
buffer[1] = j;
buffer[2] = r;
buffer[3] = e;
MPI_Send(buffer, 4, MPI_INT, source, 0, MPI_COMM_WORLD);
std::cerr << "\x1b[" << source << ";1Hworker " << source << ": " << warriors[i]->name << " vs " << warriors[j]->name << " rounds " << r + 1 << " to " << e << " of " << rounds << ".\x1b[0K" << std::flush;
}
}
for (int i = 1; i < size; ++i) {
MPI_Status status;
int buffer[4];
MPI_Recv(buffer, 4, MPI_INT, MPI_ANY_SOURCE, 1, MPI_COMM_WORLD, &status);
int source = status.MPI_SOURCE;
wins_array[buffer[0] * count + buffer[1]] += buffer[2];
wins_array[buffer[1] * count + buffer[0]] += buffer[3];
buffer[3] = 0;
MPI_Send(buffer, 4, MPI_INT, source, 0, MPI_COMM_WORLD);
std::cerr << "\x1b[" << source << ";1Hworker " << source << ": complete.\x1b[0K" << std::flush;
}
std::cerr << "\x1b[?47l\x1b""8";
int col_width = 13;
for (int i = 0; i < count; ++i)
if ((int)warriors[i]->name.size() > col_width)
col_width = (int)warriors[i]->name.size();
printf(" %*s", col_width, "");
for (int i = 0; i < count; ++i)
printf(" | %*s", col_width, warriors[i]->name.c_str());
putchar('\n');
for (int k = 0; k < col_width + 2; ++k)
putchar('-');
for (int i = 0; i < count; ++i) {
putchar('+');
for (int k = 0; k < col_width + 2; ++k)
putchar('-');
}
putchar('\n');
for (int j = 0; j < count; ++j) {
printf(" %*s", col_width, warriors[j]->name.c_str());
for (int i = 0; i < count; ++i)
if (i == j)
printf(" | %*s", col_width, "");
else
printf(" | %5d / %5d%*s", wins_array[i * count + j], placements[i * count + j] * 2, col_width - 13, "");
putchar('\n');
}
}
else {
lib94::instruction core_background = {
.op = lib94::DAT,
.mod = lib94::F,
.amode = lib94::DIRECT,
.bmode = lib94::DIRECT,
.anumber = 0,
.bnumber = 0
};
int buffer[4] = { 0, 0, 0, 0 };
MPI_Send(buffer, 4, MPI_INT, 0, 1, MPI_COMM_WORLD);
while (true) {
MPI_Recv(buffer, 4, MPI_INT, 0, 0, MPI_COMM_WORLD, 0);
if (buffer[3] == 0)
break;
auto w1 = warriors[buffer[0]];
auto w2 = warriors[buffer[1]];
int start_offset = w1->instructions.size();
int w1_wins = 0;
int w2_wins = 0;
const lib94::warrior *const wlist[2] = { w1, w2 };
int offsets[2] = {};
for (int round = buffer[2]; round < buffer[3]; ++round) {
offsets[1] = start_offset + round;
lib94::clear_core(core_background);
lib94::init_round(wlist, 2, offsets, false);
for (int step = 0; step < steps_to_tie; ++step) {
auto lost = lib94::single_step<false>();
if (lost == w1) {
++w2_wins;
break;
}
if (lost == w2) {
++w1_wins;
break;
}
}
}
buffer[2] = w1_wins;
buffer[3] = w2_wins;
MPI_Send(buffer, 4, MPI_INT, 0, 1, MPI_COMM_WORLD);
}
}
MPI_Finalize();
return 0;
}

51
tabulator-mpi/worker.cpp
View file

@ -1,51 +0,0 @@
#include <lib94/lib94.hpp>
#include <cassert>
#include <ctime>
#include <mpi.h>
#include "constants.hpp"
static void do_round(const lib94::warrior *w1, const lib94::warrior *w2, int &w1_wins, int& w2_wins) {
const lib94::warrior *ws[2] = {w1, w2};
lib94::instruction background = {
.op = lib94::DAT, .mod = lib94::F,
.amode = lib94::DIRECT, .bmode = lib94::DIRECT,
.anumber = 0, .bnumber = 0
};
lib94::clear_core(background);
assert(lib94::init_round(ws, 2));
for (int i = 0; i < STEPS_TO_TIE; ++i) {
const lib94::warrior *result = lib94::single_step<false>();
if (result == w1) {
++w2_wins;
return;
}
if (result == w2) {
++w1_wins;
return;
}
}
}
void worker_main(const lib94::warrior *const *warriors) {
lib94::seed_prng(time(0));
int buffer[4] = {-1};
while (1) {
MPI_Send(buffer, 4, MPI_INT, 0, 1, MPI_COMM_WORLD);
MPI_Recv(buffer, 4, MPI_INT, 0, 0, MPI_COMM_WORLD, 0);
if (buffer[0] == -1)
return;
buffer[2] = 0;
buffer[3] = 0;
for (int i = 0; i < ROUNDS_PER_CHUNK; ++i)
do_round(warriors[buffer[0]], warriors[buffer[1]], buffer[2], buffer[3]);
}
}