using AyCode.Core.Benchmarks.Reporting;
using AyCode.Core.Benchmarks.Workloads.Scenarios;
using AyCode.Core.Serializers;
using AyCode.Core.Serializers.Binaries;
using AyCode.Core.Tests.TestModels;
using BenchmarkDotNet.Reports;
using System.Text;

namespace AyCode.Core.Benchmarks;

/// <summary>
/// Translates <see cref="BenchmarkDotNet.Reports.Summary"/> (BDN's post-run aggregate) into the unified
/// <see cref="BenchmarkResult"/> rows consumed by <see cref="BenchmarkReportWriter"/>, then emits the
/// <c>Bdn.FullBenchmark_*.{log,LLM,output}</c> triplet alongside Console's counterparts in
/// <see cref="ReportingContext.ResolveResultsDirectory"/>'s <c>Test_Benchmark_Results/Benchmark/</c>.
///
/// <para><b>Why a separate adapter</b>: BDN's Summary is per-method (Serialize / Deserialize as separate
/// <see cref="BenchmarkReport"/>s, parameterised by TestData + Engine). The unified format collapses these
/// into per-cell rows (one row per <c>TestData × Engine</c> with both Ser and Des stats inline). The adapter
/// groups, transposes, and converts ns → ms before handing off to the shared writer.</para>
///
/// <para><b>Mean vs Median</b>: maps BDN's <see cref="BenchmarkDotNet.Mathematics.Statistics.Median"/> into
/// the BenchmarkResult's time columns — same convention as Console (which captures sample-median).
/// Min/Max/StdDev populate the inter-sample range surfaced in <see cref="BenchmarkReportWriter.FormatMicrosWithRange"/>
/// (incl. CV-warning ⚠️ marker when stddev/median exceeds <see cref="ReportingContext.UnstableCVThreshold"/>).</para>
///
/// <para><b>Iteration count = 1</b>: BDN reports <i>per-operation</i> time (ns) — already amortized across N
/// invocations. The unified BenchmarkResult expects total-batch time + iteration count (so <c>µs/op =
/// timeMs / iterations * 1000</c>). Storing Mean-in-ms with iterations = 1 makes the same formula yield
/// Mean-in-µs directly. The actual BDN N count is recorded in the BDN-native artifacts (<c>.../BDN/...</c>)
/// for anyone who wants the raw invocation count.</para>
/// </summary>
public static class BdnSummaryAdapter
{
    /// <summary>
    /// Post-run entry point — call once after <c>BenchmarkRunner.Run&lt;AcBinaryVsMemPackBenchmark&gt;(...)</c>
    /// returns. Produces the BDN-side <c>Bdn.*</c> file triplet AND prints the grouped-results console table
    /// (same view Console produces post-run) so the user sees the cell-level deltas immediately, without
    /// having to open the .LLM file.
    /// </summary>
    public static void WriteResults(Summary summary)
    {
        // Parent-process counterpart of AcBinaryVsMemPackBenchmark.Setup's charset pin: the BDN child
        // processes ran Latin1Short, but this adapter runs in the parent process where LongStringSuffix
        // would still be the compile-time default (Latin1Long). Set it so GetCharsetName() labels the
        // .LLM correctly AND the CreateTestDataSets()/CreateWorkload calls below compute matching Size(B).
        BenchmarkTestDataProvider.LongStringSuffix = CharsetSuffixes.Latin1Short;

        var allTestData = BenchmarkTestDataProvider_All_False.CreateTestDataSets();
        var results = Translate(summary, allTestData);
        var ctx = CreateContext();

        BenchmarkReportWriter.PrintGroupedResults(results, allTestData);
        BenchmarkReportWriter.SaveAll(ctx, results, allTestData);
    }

    private static ReportingContext CreateContext()
    {
#if DEBUG
        const string buildConfig = "Debug";
#elif SGEN_ONLY
        const string buildConfig = "SGenOnly";
#else
        const string buildConfig = "Release";
#endif
        return new ReportingContext(
            SourceTag: "Bdn",
            ResultsDirectory: ReportingContext.ResolveResultsDirectory(),
            BuildConfiguration: buildConfig,
            Utf8NoBom: new UTF8Encoding(encoderShouldEmitUTF8Identifier: false),
            CharsetName: GetCharsetName(),
            // Warmup / Samples / TargetSampleMs are BDN-managed (not Console's adaptive engine). Zeros here
            // signal "BDN handled internally" in the header; the BDN-native artifacts under .../BDN/ have
            // the exact BDN config (warmup count, iteration count, run strategy) for anyone who needs it.
            WarmupIterations: 0,
            BenchmarkSamples: 0,
            TargetSampleMs: 0,
            UnstableCVThreshold: 0.03);
    }

    /// <summary>
    /// Looks up the human-readable name for the currently-active <see cref="BenchmarkTestDataProvider.LongStringSuffix"/>
    /// charset. Mirrors Console's <c>Configuration.GetCurrentCharsetName</c>. The BDN serializer benchmark
    /// pins the charset to <c>Latin1Short</c> — set in <see cref="WriteResults"/> (parent process) and in
    /// <c>AcBinaryVsMemPackBenchmark.Setup</c> (child process); see those sites for the process-isolation
    /// rationale (BDN's per-benchmark child processes don't inherit a parent static-field mutation).
    /// </summary>
    private static string GetCharsetName()
    {
        var s = BenchmarkTestDataProvider.LongStringSuffix;
        return s switch
        {
            CharsetSuffixes.AsciiFix => nameof(CharsetSuffixes.AsciiFix),
            CharsetSuffixes.AsciiShort => nameof(CharsetSuffixes.AsciiShort),
            CharsetSuffixes.AsciiLong => nameof(CharsetSuffixes.AsciiLong),
            CharsetSuffixes.Latin1Fix => nameof(CharsetSuffixes.Latin1Fix),
            CharsetSuffixes.Latin1Short => nameof(CharsetSuffixes.Latin1Short),
            CharsetSuffixes.Latin1Long => nameof(CharsetSuffixes.Latin1Long),
            CharsetSuffixes.CjkBmpShort => nameof(CharsetSuffixes.CjkBmpShort),
            CharsetSuffixes.CjkBmpLong => nameof(CharsetSuffixes.CjkBmpLong),
            CharsetSuffixes.CyrillicShort => nameof(CharsetSuffixes.CyrillicShort),
            CharsetSuffixes.CyrillicLong => nameof(CharsetSuffixes.CyrillicLong),
            CharsetSuffixes.MixedShort => nameof(CharsetSuffixes.MixedShort),
            CharsetSuffixes.MixedLong => nameof(CharsetSuffixes.MixedLong),
            _ => "Custom"
        };
    }

    private static List<BenchmarkResult> Translate(Summary summary, List<TestDataSet> allTestData)
    {
        var grouped = summary.Reports
            .Where(r => r.Success && r.ResultStatistics != null)
            .GroupBy(r => (
                TestData: GetParam(r, "TestData"),
                Engine: GetParam(r, "Engine")
            ))
            .Where(g => !string.IsNullOrEmpty(g.Key.TestData) && !string.IsNullOrEmpty(g.Key.Engine))
            .ToList();

        var results = new List<BenchmarkResult>(grouped.Count);

        foreach (var group in grouped)
        {
            var testDataSet = (TestDataSet<TestOrder_All_False>)allTestData.First(t => t.Name.StartsWith(group.Key.TestData));
            var engineEnum = group.Key.Engine switch
            {
                "AcBinary" => BenchmarkEngine.AcBinary,
                "MemoryPack" => BenchmarkEngine.MemoryPack,
                _ => throw new InvalidOperationException($"Unknown engine in BDN params: {group.Key.Engine}")
            };

            // Construct the same workload instance AcBinaryVsMemPackBenchmark.Setup would build — same options,
            // same wire mode. Reading SerializedSize + OptionsDescription from it keeps the BDN-side metadata
            // in lockstep with what the workload actually serialised (no drift between hardcoded BDN strings
            // and the workload's own OptionsDescription / SerializedSize).
            var workload = CreateWorkload(testDataSet, group.Key.Engine);

            var result = new BenchmarkResult
            {
                TestDataName = testDataSet.DisplayName,
                Engine = engineEnum,
                IoMode = BenchmarkIoMode.ByteArray,
                DispatchMode = BenchmarkDispatchMode.SGen,
                OptionsPreset = group.Key.Engine == "AcBinary" ? "FastMode" : "Default",
                OrderTypeName = nameof(TestOrder_All_False),
                SerializedSize = workload.SerializedSize,
                OptionsDescription = workload.OptionsDescription,
            };

            // ns → ms (BenchmarkResult expects ms per op with iter=1, so µs/op = ms * 1000 / 1 = ms*1000).
            const double nsToMs = 1.0 / 1_000_000.0;

            foreach (var report in group)
            {
                var methodName = report.BenchmarkCase.Descriptor.WorkloadMethod.Name;
                var stats = report.ResultStatistics!;
                var allocBytes = report.GcStats.GetBytesAllocatedPerOperation(report.BenchmarkCase) ?? 0;

                if (methodName == "Serialize")
                {
                    result.SerializeTimeMs = stats.Median * nsToMs;
                    result.SerializeTimeMinMs = stats.Min * nsToMs;
                    result.SerializeTimeMaxMs = stats.Max * nsToMs;
                    result.SerializeTimeStdDevMs = stats.StandardDeviation * nsToMs;
                    result.SerializeIterations = 1; // see class-doc "Iteration count = 1" note
                    result.SerializeAllocBytesPerOp = allocBytes;
                }
                else if (methodName == "Deserialize")
                {
                    result.DeserializeTimeMs = stats.Median * nsToMs;
                    result.DeserializeTimeMinMs = stats.Min * nsToMs;
                    result.DeserializeTimeMaxMs = stats.Max * nsToMs;
                    result.DeserializeTimeStdDevMs = stats.StandardDeviation * nsToMs;
                    result.DeserializeIterations = 1;
                    result.DeserializeAllocBytesPerOp = allocBytes;
                }
            }

            // Compose RT from Ser + Des per-op µs (same logic as Console BenchmarkLoop's in-memory
            // composition — since BDN measures Ser and Des independently, RT here is the analytic sum).
            var serPerOp = BenchmarkReportWriter.ToPerOpMicros(result.SerializeTimeMs, result.SerializeIterations);
            var desPerOp = BenchmarkReportWriter.ToPerOpMicros(result.DeserializeTimeMs, result.DeserializeIterations);
            var rtPerOp = serPerOp + desPerOp;
            result.RoundTripIterations = Math.Max(result.SerializeIterations, result.DeserializeIterations);
            result.RoundTripTimeMs = rtPerOp / 1000.0 * result.RoundTripIterations;
            result.RoundTripAllocBytesPerOp = result.SerializeAllocBytesPerOp + result.DeserializeAllocBytesPerOp;

            results.Add(result);
        }

        return results;
    }

    private static string GetParam(BenchmarkReport report, string name) =>
        report.BenchmarkCase.Parameters.Items.FirstOrDefault(p => p.Name == name)?.Value?.ToString() ?? "";

    /// <summary>
    /// Constructs the same workload instance <see cref="AcBinaryVsMemPackBenchmark.Setup"/> would build —
    /// same options, same wire mode. The adapter reads <see cref="ISerializerBenchmark.SerializedSize"/> and
    /// <see cref="ISerializerBenchmark.OptionsDescription"/> from this instance so the BDN-side BenchmarkResult
    /// rows carry the same workload-side metadata the Console rows have (no risk of drift between hardcoded
    /// adapter strings and what the workload actually used).
    ///
    /// Cost: one Serialize call inside the ctor per (TestData × Engine) cell — runs once during summary
    /// translation, NOT in BDN's measured hot path. Negligible vs BDN's per-run cost.
    /// </summary>
    private static ISerializerBenchmark CreateWorkload(TestDataSet<TestOrder_All_False> testDataSet, string engine)
    {
        if (engine == "AcBinary")
        {
            var options = AcBinarySerializerOptions.FastMode;
            options.WireMode = WireMode.Compact;
            return new AcBinaryBenchmark<TestOrder_All_False>(testDataSet.Order, options, "FastMode");
        }
        return new MemoryPackBenchmark<TestOrder_All_False>(testDataSet.Order, WireMode.Compact, "Default");
    }
}