"""Comparative figures for milestone-gqa-prefill-long-ctx-4cases.

Mirror of paper_plot_gqa_decode_long_ctx_4cases but for the prefill
variant. Reads sweep.json (emitted by ``kernbench run --bench
milestone-gqa-prefill-long-ctx-4cases``) and writes four PNGs into
``docs/report/1H-codesign-paper/figures/``:

  gqa_prefill_long_ctx_4cases_latency.png       end-to-end latency per case
  gqa_prefill_long_ctx_4cases_traffic.png       ipcq/dma op-count breakdown
  gqa_prefill_long_ctx_4cases_memory.png        per-PE KV bytes per case
  gqa_prefill_long_ctx_4cases_parallelism.png   active-PE × S_local work load

Run (after the bench):
    GQA_PREFILL_LONG_CTX_4CASES_RUN=1 python -m kernbench.cli.main run \\
        --bench milestone-gqa-prefill-long-ctx-4cases --topology topology.yaml
    python scripts/paper/paper_plot_gqa_prefill_long_ctx_4cases.py
"""
from __future__ import annotations

import json
from pathlib import Path

import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt  # noqa: E402

_REPO_ROOT = Path(__file__).resolve().parents[2]
# Sweep JSON + PNGs live together under the bench output dir.
_FIG_DIR = (
    _REPO_ROOT / "src" / "kernbench" / "benches"
    / "1H_milestone_output" / "gqa" / "long_ctx"
)
_SWEEP_JSON = _FIG_DIR / "sweep_prefill.json"

# Panel name → (short label, case ordinal for left-to-right plot order).
_CASE_INFO = {
    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_tp": (
        "Case 1\nCube-SP × PE-TP", 1),
    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_tp": (
        "Case 2\nCube-Repl × PE-TP", 2),
    "single_kv_group_prefill_long_ctx_gqa_cube_repl_pe_sp": (
        "Case 3\nCube-Repl × PE-SP", 3),
    "single_kv_group_prefill_long_ctx_gqa_cube_sp_pe_sp": (
        "Case 4 ★\nCube-SP × PE-SP", 4),
}


def _load() -> list[dict]:
    return json.loads(_SWEEP_JSON.read_text())["rows"]


def _sorted_by_case(rows: list[dict]) -> list[dict]:
    return sorted(rows, key=lambda r: _CASE_INFO[r["panel"]][1])


def _plot_latency(rows: list[dict]) -> Path:
    rows = _sorted_by_case(rows)
    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
    lat_us = [r["latency_ns"] / 1e3 for r in rows]
    colors = ["#888", "#888", "#888", "#3b6ea5"]  # Case 4 highlighted
    fig, ax = plt.subplots(figsize=(8.0, 4.5))
    bars = ax.bar(labels, lat_us, color=colors, width=0.6)
    ax.set_ylabel("end-to-end latency (µs)")
    ax.set_title(
        "Long-context prefill 4-cases — end-to-end latency per case\n"
        "LLaMA-3.1-70B single-KV-head group (8 cubes × 8 PEs)"
    )
    ax.bar_label(bars, fmt="%.1f", padding=3, fontsize=9)
    ax.grid(axis="y", ls=":", alpha=0.5)
    ax.set_ylim(0, max(lat_us) * 1.15)
    fig.tight_layout()
    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_latency.png"
    fig.savefig(out, dpi=150)
    plt.close(fig)
    return out


def _plot_traffic(rows: list[dict]) -> Path:
    rows = _sorted_by_case(rows)
    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
    x = list(range(len(rows)))
    keys = ["ipcq_copy_count", "dma_read_count", "dma_write_count"]
    disp = ["IPCQ copy", "DMA read", "DMA write"]
    colors = ["#c0504d", "#9bbb59", "#8064a2"]
    w = 0.25
    fig, ax = plt.subplots(figsize=(9.0, 4.5))
    for i, (k, d, c) in enumerate(zip(keys, disp, colors)):
        vals = [r["op_log_summary"][k] for r in rows]
        ax.bar([xi + (i - 1) * w for xi in x], vals, width=w, label=d, color=c)
    ax.set_xticks(list(x))
    ax.set_xticklabels(labels, fontsize=9)
    ax.set_ylabel("op count")
    ax.set_title("Long-context prefill 4-cases — op-count breakdown per case")
    ax.legend(fontsize=9)
    ax.grid(axis="y", ls=":", alpha=0.5)
    fig.tight_layout()
    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_traffic.png"
    fig.savefig(out, dpi=150)
    plt.close(fig)
    return out


def _s_local_per_pe(panel: str, *, S_kv: int, C: int, P: int) -> int:
    """S_local (token count) each PE attends over locally.

    Encodes the cube/pe sharding axes from the panel name:
      cube_sp_pe_tp (Case 1): S_kv / C       (pe=replicate within cube)
      cube_repl_pe_tp (Case 2): S_kv         (full S_kv per active PE)
      cube_repl_pe_sp (Case 3): S_kv / P     (pe=row_wise within cube)
      cube_sp_pe_sp (Case 4): S_kv / (C·P)   (★ 64-way split)
    """
    S_per_cube = S_kv if "cube_repl" in panel else S_kv // C
    return S_per_cube // P if "pe_sp" in panel else S_per_cube


def _active_pe_count(panel: str, *, C: int, P: int) -> int:
    """Number of PEs doing non-idle attention work.

    Prefill T_q≫1 means PE-TP is *useful* (not wasted as in decode):
      cube_sp_pe_tp (Case 1): C·P   (T_q sharded across all 64 ranks)
      cube_repl_pe_tp (Case 2): P   (CUBE 0 only; T_q sharded across its P PEs)
      cube_repl_pe_sp (Case 3): C·P (all PEs busy but cubes redundant)
      cube_sp_pe_sp (Case 4): C·P   (all 64 PEs doing unique work)
    """
    if "cube_repl" in panel and "pe_tp" in panel:
        return P
    return C * P


def _kv_bytes_per_pe(panel: str, *, S_kv: int, h_kv: int,
                     d_head: int, C: int, P: int) -> int:
    """KV bytes a single PE references (K + V, f16, 2 B/elem)."""
    s_local = _s_local_per_pe(panel, S_kv=S_kv, C=C, P=P)
    return 2 * s_local * h_kv * d_head * 2


def _plot_memory(rows: list[dict]) -> Path:
    """Per-PE KV bytes — Case 4 wins (64-way split)."""
    rows = _sorted_by_case(rows)
    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
    mib_per_pe = [
        _kv_bytes_per_pe(
            r["panel"], S_kv=r["S_kv"], h_kv=r["h_kv"],
            d_head=r["d_head"], C=r["C"], P=r["P"],
        ) / (1024 * 1024)
        for r in rows
    ]
    colors = ["#888", "#c0504d", "#888", "#3b6ea5"]
    fig, ax = plt.subplots(figsize=(8.0, 4.5))
    bars = ax.bar(labels, mib_per_pe, color=colors, width=0.6)
    ax.set_ylabel("KV bytes per PE (MiB, K + V, f16)")
    ax.set_title(
        "Long-context prefill 4-cases — KV memory per PE\n"
        "(one KV-head group; per-layer, full S_kv state)"
    )
    ax.bar_label(bars, fmt="%.3f", padding=3, fontsize=9)
    ax.grid(axis="y", ls=":", alpha=0.5)
    ax.set_ylim(0, max(mib_per_pe) * 1.15)
    fig.tight_layout()
    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_memory.png"
    fig.savefig(out, dpi=150)
    plt.close(fig)
    return out


def _t_q_per_pe(panel: str, *, T_q: int, C: int, P: int) -> int:
    """T_q row count each active PE computes attention for.

      cube_sp_pe_tp (Case 1): T_q / (C·P)  (T_q sharded across all 64 ranks)
      cube_repl_pe_tp (Case 2): T_q / P    (T_q sharded across CUBE 0's P PEs)
      cube_repl_pe_sp (Case 3): T_q        (Q replicated on every PE)
      cube_sp_pe_sp (Case 4): T_q / C      (Q sharded by cube, replicated within)
    """
    if "cube_sp" in panel and "pe_tp" in panel:
        return T_q // (C * P)
    if "cube_repl" in panel and "pe_tp" in panel:
        return T_q // P
    if "cube_repl" in panel and "pe_sp" in panel:
        return T_q
    return T_q // C  # cube_sp_pe_sp


def _s_kv_processed_per_pe(panel: str, *, S_kv: int, C: int, P: int) -> int:
    """S_kv tokens each PE actually processes attention over.

    Differs from ``_s_local_per_pe`` (OWNED KV bytes): for cases with
    a Ring (Case 1, 4) each PE sees C ring steps so processes more
    tokens than it locally owns.

      cube_sp_pe_tp (Case 1): S_kv          (Ring + pe=replicate within cube)
      cube_repl_pe_tp (Case 2): S_kv        (full KV per PE)
      cube_repl_pe_sp (Case 3): S_kv / P    (pe=row_wise; no Ring)
      cube_sp_pe_sp (Case 4): S_kv / P      (Ring restores full S_kv/P per PE)
    """
    if "cube_sp" in panel and "pe_tp" in panel:
        return S_kv
    if "cube_repl" in panel and "pe_tp" in panel:
        return S_kv
    return S_kv // P  # both pe_sp cases


def _plot_parallelism(rows: list[dict]) -> Path:
    """Total compute work (PE × T_q × S_kv token-pairs) — exposes Case 3's
    redundancy. Cases 1, 2, 4 all do the same total work (correct
    attention over T_q × S_kv); Case 3 does C× more (cubes redundantly
    repeat the same compute because K/V is cube-replicated).
    """
    rows = _sorted_by_case(rows)
    labels = [_CASE_INFO[r["panel"]][0] for r in rows]
    total_work = [
        _active_pe_count(r["panel"], C=r["C"], P=r["P"])
        * _t_q_per_pe(r["panel"], T_q=r["T_q"], C=r["C"], P=r["P"])
        * _s_kv_processed_per_pe(
            r["panel"], S_kv=r["S_kv"], C=r["C"], P=r["P"],
        )
        for r in rows
    ]
    colors = ["#888", "#888", "#c0504d", "#3b6ea5"]  # Case 3 red, Case 4 highlighted
    fig, ax = plt.subplots(figsize=(8.0, 4.5))
    bars = ax.bar(labels, total_work, color=colors, width=0.6)
    ax.set_ylabel(
        "total compute (PE × T_q × S_kv token-pairs; lower ⇒ less wasted work)"
    )
    ax.set_title(
        "Long-context prefill 4-cases — total compute work across active PEs\n"
        "(Case 3 replicates K/V across 8 cubes ⇒ 8× redundant compute)"
    )
    ax.bar_label(bars, fmt="%d", padding=3, fontsize=9)
    ax.grid(axis="y", ls=":", alpha=0.5)
    ax.set_ylim(0, max(total_work) * 1.15)
    fig.tight_layout()
    out = _FIG_DIR / "gqa_prefill_long_ctx_4cases_parallelism.png"
    fig.savefig(out, dpi=150)
    plt.close(fig)
    return out


def main() -> None:
    rows = _load()
    _FIG_DIR.mkdir(parents=True, exist_ok=True)
    p1 = _plot_latency(rows)
    p2 = _plot_traffic(rows)
    p3 = _plot_memory(rows)
    p4 = _plot_parallelism(rows)
    print(f"wrote {p1}")
    print(f"wrote {p2}")
    print(f"wrote {p3}")
    print(f"wrote {p4}")


if __name__ == "__main__":
    main()