# tensor: a scientific computing library in pure Go.
#
# A library: nothing to install, nothing to run. Below the variable
# block is the standard recipe contract; docs-check and fuzz-all are
# the project extensions. The portable build is the only build.

# What the test, race, unit and bench recipes sweep: every logic
# package, the examples aside. The examples are main programs with no
# tests; the build compiles them, and the coverage floor is a property
# of the library packages alone.
packages := ". ./internal/... ./grad/... ./integrate/... ./io/... ./linalg/... ./optim/... ./plot/... ./signal/... ./spmd/... ./stats/..."

# Where the fuzz targets live; the fuzz-all sweep discovers them by name.
fuzz_packages := "./io ./grad"

# The memory fence for the test recipes: a cgroup ceiling with swap off, so a
# runaway run dies as a failed run and never eats the machine. 4G is the
# default; raise it only with a reason recorded here.
memlimit := "4G"

# The representative set the bench-report measures, one benchmark per
# kernel family across the packages. The recipe measures the names this
# release tag and the head tree share, so an old tag never blocks the
# report.
bench_set := "Add1M Exp100k Dot SumAxis2D SumAxis3DDim1 MeanAxis2D Norm2D Prod2D CumSum2D ArgSort1M MatMul1024 MatMulSmall64 MatMulOdd130 MatMulTall512x64x2048 MatVec KernelTransposeTiled256 EinsumBatchedMatMul Cholesky256 Solve256 SVD128 FFT4096 FFT2_512x512 Conv1D CWTMorlet SavitzkyGolay WelchPSD IntegrateRK4 IntegrateBackwardEuler IntegrateHeat2D MinimiseLBFGSBounded MinimiseDifferentialEvolution BackwardTwoLayer LinearRegression LogisticRegression KernelDensity Median CovarianceMatrix"

default:
    @just --list

# Compile. Zero errors, zero warnings.
build:
    go build ./...

# The test gate: the suite, no cache, the coverage floor, under the memory fence.
test:
    #!/usr/bin/env perl
    my @fence = (q{systemd-run}, q{--user}, q{--scope},
                 q{-p}, q{MemoryMax={{memlimit}}}, q{-p}, q{MemorySwapMax=0});
    system(@fence, q{go}, q{test}, q{-count=1}, q{-timeout}, q{30m},
      q{-coverprofile}, q{coverage.out}, qw({{packages}})) == 0
      or die qq{the test suite failed\n};
    open(my $c, q{-|}, q{go}, q{tool}, q{cover}, q{-func=coverage.out}) or die qq{cover: $!};
    my $total;
    while (my $l = <$c>) { $total = $1 if $l =~ m{^total:\s+\S+\s+([0-9.]+)%} }
    close($c);
    die qq{no total line in coverage.out\n} unless defined $total;
    printf qq{Total coverage: %s%%\n}, $total;
    exit($total < 80 ? 1 : 0);

# The same suite under the race detector. The expensive one, still fenced.
race:
    systemd-run --user --scope -p MemoryMax={{memlimit}} -p MemorySwapMax=0 go test -race -count=1 -timeout 30m {{packages}}

# Fast scoped run for iterating. This is the one that runs after every edit.
unit pkgs=packages run=".*":
    systemd-run --user --scope -p MemoryMax={{memlimit}} -p MemorySwapMax=0 go test {{pkgs}} -run '{{run}}'

# Time-boxed fuzz of one target in one package. The package is required; never a gate.
fuzz target pkg fuzztime="60s":
    systemd-run --user --scope -p MemoryMax={{memlimit}} -p MemorySwapMax=0 go test -run '^$' -fuzz '{{target}}' -fuzztime={{fuzztime}} {{pkg}}

# Benchmarks. On an idle machine only, deliberately unfenced: a ceiling would distort the measurement.
bench pkgs=packages:
    go test -run '^$' -bench=. -benchmem -count=5 {{pkgs}}

# Format in place.
fmt:
    gofmt -w .

# Zero diff. Prints nothing when everything is formatted.
fmt-check:
    #!/usr/bin/env perl
    open(my $g, q{-|}, q{gofmt}, q{-l}, q{.}) or die qq{gofmt: $!};
    my @bad = <$g>;
    close($g);
    print @bad;
    exit(@bad ? 1 : 0);

# Both static gates: go vet and go fix -diff.
vet:
    go vet ./...
    go fix -diff ./...

# The definition of done, in one command. Once per task, never per edit.
gates: build fmt-check vet test race

# Remove the build artefacts.
clean:
    rm -rf bin/ coverage.out

# Run every Go program in README.md. An extension: no standard gate covers prose.
docs-check:
    #!/usr/bin/env perl
    my $root = `pwd`; chomp $root;
    my $tmp = $ENV{TMPDIR} || q{/tmp};
    my $base = qq{$tmp/tensor-docs-check};
    system(q{rm}, q{-rf}, $base) == 0 or die qq{rm $base failed\n};
    mkdir $base or die qq{mkdir $base: $!};
    open(my $md, q{<}, q{README.md}) or die qq{README.md: $!};
    my (@blocks, $cur);
    while (my $l = <$md>) {
        if ($l =~ m{^```go\s*$}) { $cur = q{}; next }
        if (defined $cur && $l =~ m{^```\s*$}) { push @blocks, $cur; undef $cur; next }
        $cur .= $l if defined $cur;
    }
    close($md) or die qq{README.md close failed\n};
    my ($i, $failed) = (0, 0);
    for my $b (@blocks) {
        $i++;
        my $dir = sprintf qq{%s/%02d}, $base, $i;
        mkdir $dir or die qq{mkdir $dir: $!};
        open(my $g, q{>}, qq{$dir/go.mod}) or die qq{$dir/go.mod: $!};
        print $g qq{module readme$i\n\ngo 1.27.1\n\n}
          . qq{require sourcedock.dev/petrbalvin/tensor v0.0.0\n\n}
          . qq{replace sourcedock.dev/petrbalvin/tensor => $root\n};
        close($g) or die qq{go.mod close failed\n};
        open(my $m, q{>}, qq{$dir/main.go}) or die qq{$dir/main.go: $!};
        print $m $b;
        close($m) or die qq{main.go close failed\n};
        chdir $dir or die qq{chdir $dir: $!};
        local $ENV{GOFLAGS} = q{-mod=mod};
        open(my $r, q{-|}, q{go}, q{run}, q{.}) or die qq{go run $dir: $!};
        my @out = <$r>;
        # close reports true on a clean exit; $? carries the status of one that failed.
        my $ok = close($r);
        my $status = $? >> 8;
        unless ($ok) {
            $failed++;
            print qq{FAILED: README program $i in $dir (exit $status)\n}, @out;
        }
        chdir $root or die qq{chdir $root: $!};
    }
    print qq{README programs: $i, failures: $failed\n};
    exit($failed ? 1 : 0);

# Fuzz every target for FUZZTIME each, under the same fence. Exploration, never a gate.
fuzz-all fuzztime="5s":
    #!/usr/bin/env perl
    my @fence = (q{systemd-run}, q{--user}, q{--scope},
                 q{-p}, q{MemoryMax={{memlimit}}}, q{-p}, q{MemorySwapMax=0});
    my $bad = 0;
    for my $pkg (split q{ }, q({{fuzz_packages}})) {
        open(my $l, q{-|}, q{go}, q{test}, q{-list}, q{^Fuzz}, $pkg)
          or die qq{go test -list $pkg: $!};
        my @targets = map { chomp; $_ } grep { m{^Fuzz} } <$l>;
        close($l) or die qq{go test -list $pkg failed\n};
        for my $t (@targets) {
            print qq{=== fuzz $t in $pkg for {{fuzztime}} ===\n};
            system(@fence, q{go}, q{test}, q{-run}, q{^$}, q{-fuzz}, qq{^$t\$},
              q{-fuzztime}, q({{fuzztime}}), $pkg) == 0
              or do { $bad = 1; print qq{FAILED: $t\n} };
        }
    }
    exit($bad ? 1 : 0);

# Benchmark the working tree against the latest release tag, rewriting docs/benchmarks/release-vs-head.md.
bench-report benchtime="0.5s":
    #!/usr/bin/env perl
    use v5.40;
    my $repo  = `git rev-parse --show-toplevel`; chomp $repo;
    my @tags  = grep { chomp; $_ } `git tag --list v* --sort=-v:refname`;
    @tags or die qq{no release tag in the repository\n};
    my ($tag) = @tags;
    my ($head, $dirty) = (`git rev-parse --short HEAD` =~ s/\s+\z//r);
    $dirty = `git status --porcelain` eq q{} ? q{} : q{, dirty working tree};
    chomp $head;
    my $rel_dir = $repo =~ s{/[^/]+\z}{}r . qq{/tensor-bench-release};
    END { system(q{git}, q{worktree}, q{remove}, q{--force}, $rel_dir) }
    if (-d $rel_dir) { system(q{git}, q{worktree}, q{remove}, q{--force}, $rel_dir) }
    system(q{git}, q{worktree}, q{add}, q{--detach}, $rel_dir, $tag) == 0
      or die qq{worktree for $tag failed\n};
    print qq{=== $tag ($rel_dir) against head $head$dirty ===\n};
    my %wanted = map { $_ => 1 } split q{ }, q({{bench_set}});
    my %map;    # side -> name -> package
    for my $side ([release => $rel_dir], [head => $repo]) {
        my @pkgs = split /\n/, `go -C $side->[1] list ./...`;
        for my $p (@pkgs) {
        open(my $l, q{-|}, q{go}, q{-C}, $side->[1], q{test}, $p, q{-list}, q{Benchmark})
          or die qq{go -C $side->[1] test -list $p: $!};
        while (my $n = <$l>) {
            next unless $n =~ m{^Benchmark(\w+)\s*\z};
            chomp $n;
            $map{$side->[0]}{$1} = $p if $wanted{$1};
        }
        close($l);
        }
    }
    my %pkgs;   # side -> package -> [names both sides know]
    for my $n (sort keys %{$map{release}}) {
        next unless $map{head}{$n};
        push @{$pkgs{release}{$map{release}{$n}}}, $n;
        push @{$pkgs{head}{$map{head}{$n}}}, $n;
    }
    my (%ns, %bytes, %allocs);
    for my $round (1 .. 4) {
        my @sides = $round % 2 ? ([release => $rel_dir], [head => $repo])
                               : ([head => $repo], [release => $rel_dir]);
        for my $s (@sides) {
            my ($side, $dir) = @$s;
            for my $p (sort keys %{$pkgs{$side}}) {
            my @names = @{$pkgs{$side}{$p}};
            my $sel = q{^(} . join(q{|}, map { qq{Benchmark$_} } @names) . q{)$};
            open(my $r, q{-|}, q{go}, q{-C}, $dir, q{test}, $p, q{-run}, q{^$},
              q{-bench}, $sel, q{-benchmem}, qq{-benchtime={{benchtime}}}, q{-count=2})
              or die qq{go -C $dir bench $p: $!};
            while (my $l = <$r>) {
                next unless $l =~ m{^Benchmark(\w+)\S*\s+\d+\s+([\d.]+) ns/op(?:\s+(\d+) B/op(?:\s+(\d+) allocs/op)?)?};
                my $n = $1;
                next unless $wanted{$n} && $map{release}{$n} && $map{head}{$n};
                push @{$ns{$n}{$side}}, $2;
                push @{$bytes{$n}{$side}}, $3 if defined $3;
                push @{$allocs{$n}{$side}}, $4 if defined $4;
            }
            close($r) or die qq{the bench run failed: $p ($side)\n};
        }
        }
        print qq{=== round $round done ===\n};
    }
    sub median { my @v = sort { $a <=> $b } @_; my $m = int(@v / 2); @v % 2 ? $v[$m] : ($v[$m - 1] + $v[$m]) / 2 }
    sub unit {
        my ($ns) = @_;
        return sprintf qq{%.2f ms}, $ns / 1e6 if $ns >= 1e6;
        return sprintf qq{%.2f µs}, $ns / 1e3 if $ns >= 1e3;
        return sprintf qq{%.0f ns}, $ns;
    }
    my (@rows, $faster, $slower, $same);
    ($faster, $slower, $same) = (0, 0, 0);
    for my $n (sort keys %ns) {
        next unless @{$ns{$n}{release}} && @{$ns{$n}{head}};
        my ($mr, $mh) = (median(@{$ns{$n}{release}}), median(@{$ns{$n}{head}}));
        my $f = $mr / $mh;
        $f >= 1.05 ? $faster++ : $f <= 0.95 ? $slower++ : $same++;
        my $bs = defined $bytes{$n}{release} && defined $bytes{$n}{head}
          ? sprintf qq{%d → %d}, median(@{$bytes{$n}{release}}), median(@{$bytes{$n}{head}}) : q{—};
        my $as = defined $allocs{$n}{release} && defined $allocs{$n}{head}
          ? sprintf qq{%d → %d}, median(@{$allocs{$n}{release}}), median(@{$allocs{$n}{head}}) : q{—};
        push @rows, [$n, unit($mr), unit($mh), sprintf(qq{%.2f}, $f), $bs, $as];
    }
    @rows or die qq{no benchmark measured on both sides\n};
    @rows = sort { $b->[3] <=> $a->[3] } @rows;
    open(my $ci, q{<}, q{/proc/cpuinfo}) or die qq{/proc/cpuinfo: $!};
    my ($cpu_name) = grep { m{^model name} } <$ci>;
    close($ci);
    ($cpu_name = $cpu_name // q{unknown CPU}) =~ s{^model name\s*:\s*}{};
    chomp $cpu_name;
    my $cores = do { open(my $c3, q{<}, q{/proc/cpuinfo}); grep { m{^processor} } <$c3> };
    my $mem = do { open(my $m, q{<}, q{/proc/meminfo}); (grep { m{^MemTotal} } <$m>)[0] };
    $mem =~ s{^MemTotal:\s+([0-9]+) kB.*}{sprintf qq{%.0f GiB}, $1 / 1048576}e;
    chomp(my $go_v = `go version`); chomp(my $os = `uname -sr`); chomp($mem);
    my $report = qq{$repo/docs/benchmarks/release-vs-head.md};
    mkdir qq{$repo/docs/benchmarks};
    open(my $o, q{>}, $report) or die qq{$report: $!};
    print $o <<"EOF";
    # Benchmark report: $tag against head $head

    One live comparison, regenerated by `just bench-report` and never
    accumulated: the newest release tag against the working tree, taken in
    one interleaved session on an idle machine. Re-run it before a release
    and commit the file the run writes.

    ## Environment

    - CPU: $cpu_name, $cores logical cores
    - Memory: $mem
    - OS: $os
    - Go: $go_v
    - Build: the portable build, no pinned `GOAMD64` level, no `GOEXPERIMENT`

    ## Revisions

    - Release: `$tag`
    - Head: `$head`$dirty
    - The set: the benchmarks `bench_set` names that both revisions carry;
      a name either side lacks is left out, never counted as a result.

    ## Method

    Four interleaved rounds, the side order alternating between rounds, two
    counts per side per round, `-benchtime={{benchtime}}` with `-benchmem`,
    medians over the eight samples a side collects. A factor at or above
    1.05 is faster, at or below 0.95 slower, anything between noise. Time
    of the release divides time of the head, so above one means the head is
    faster.

    ## Results

    | Benchmark | $tag | head $head | factor | B/op release → head | allocs/op release → head |
    |---|---|---|---|---|---|
    EOF
    for my $r (@rows) { print $o qq{| `$r->[0]` | $r->[1] | $r->[2] | $r->[3]x | $r->[4] | $r->[5] |\n} }
    print $o qq{\n## Summary\n\n$faster of }, scalar @rows,
      qq{ benchmarks sit above the noise band (faster), $slower below it (slower), $same inside it.\n};
    close($o) or die qq{$report close failed\n};
    print qq{=== $report written: $faster faster, $slower slower, $same unchanged ===\n};