#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
"""Summary of a device profile of ``profile_ops.py`` (CPU only, standard library): the OPT_BASELINE.md tables.

    python code/scripts/profile_summary.py <ops_perf_results_*.csv> <out.json> [--md out.md]

Reads the Tracy ops CSV (``python -m tracy -r ...``) and its sections ``eager`` .. ``eager_end`` and ``trace`` ..
``trace_end``. The eager section carries per-module (``m:<module>``: the 21 sparse layers ``m:sparse.<layer>``,
``m:sparse.to_dense``, the dense modules, the query head's parts as ``m:query.<part>``) and per-conv (``c:<layer>``)
signposts; the traced frame (``sparse_<bucket>`` replay + ``dense`` replay: two trace ids) has the same op sequence
(checked op code by op code), so every replayed op inherits the module / layer of the eager op at the same position.
Per section: program count, device kernel sum, FW sum, op-to-op gap sum and median, span (first FW start -> last FW
end at the CSV's chip frequency), cores per op, math fidelities; tables per category, module, op code and layer; the
top single ops; the largest gaps; per trace segment (trace id) the same totals. If the trace section holds several
frames (``profile_ops.py --replays N``), the last replay session of each trace id is summarised. Input shapes of the CSV are those of
the first op that built a cached program (``--no-op-info-cache`` in tracy shows every op's own), so read shapes of
later ops of the same program with care; durations are per op.
"""
from __future__ import annotations

import argparse
import csv
import json
import re
import statistics
from collections import OrderedDict, defaultdict
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple

KERNEL = "DEVICE KERNEL DURATION [ns]"
FW = "DEVICE FW DURATION [ns]"
GAP = "OP TO OP LATENCY [ns]"
START = "DEVICE FW START CYCLE"
END = "DEVICE FW END CYCLE"

CATEGORIES = [  # (category, regex on the op code), first match wins
    ("conv2d", r"Conv2d|conv2d|Conv2D"),
    ("conv_transpose2d", r"ConvTranspose|conv_transpose"),
    ("halo", r"Halo|halo"),
    ("attention (SDPA)", r"SDPA|Sdpa|ScaledDotProduct|sdpa"),
    ("top-k", r"TopK|Topk|topk"),
    ("layer norm / softmax", r"LayerNorm|layernorm|Softmax|softmax"),
    ("pool (local max)", r"Pool|pool"),
    ("matmul (linear, 1x1 conv)", r"Matmul|matmul|Linear"),
    ("slice / DRAM-slicing copies", r"Slice|slice"),
    ("binary (two-term sums, bias adds, K-split adds, masks)", r"Binary|binary"),
    ("reduction (slot max)", r"Reduce|reduce|Max\b|MaxDevice"),
    ("embedding (sparse im2col gathers, to-dense gathers)", r"Embedding|embedding"),
    ("tilize / untilize / layout", r"Tilize|tilize|Untilize|untilize|ToLayout|Layout"),
    ("reshape / transpose / permute", r"Reshape|reshape|Transpose|transpose|Permute|permute|View"),
    ("move / copy / reshard", r"Move|move|Copy|copy|Shard|shard|Interleaved|interleaved|Reallocate"),
    ("pad / fill", r"Pad|pad|Fill|fill"),
    ("unary / typecast", r"Unary|unary|Typecast|typecast"),
]
COMPUTE = ("Conv2dDeviceOperation", "MatmulDeviceOperation", "ConvTranspose2dDeviceOperation")


def category(op: str) -> str:
    for name, pat in CATEGORIES:
        if re.search(pat, op):
            return name
    return "other"


def num(row: Dict[str, str], key: str) -> float:
    v = (row.get(key) or "").strip()
    try:
        return float(v) if v else 0.0
    except ValueError:
        return 0.0


def logical(v: str) -> str:
    """'32[16]' (padded[logical]) -> '16'; '7' -> '7'."""
    m = re.match(r"\s*([0-9]+)(?:\[([0-9]+)\])?", v or "")
    return (m.group(2) or m.group(1)) if m else "?"


def chip_freq_mhz(csv_path: Path) -> Optional[float]:
    for cand in [csv_path.parent / "profile_log_device.csv", csv_path.parent.parent / "profile_log_device.csv",
                 csv_path.parent / ".logs" / "profile_log_device.csv",
                 csv_path.parent.parent / ".logs" / "profile_log_device.csv"]:
        try:
            head = cand.read_text(errors="ignore").splitlines()[0]
        except (OSError, IndexError):
            continue
        m = re.search(r"CHIP_FREQ\[MHz\]:\s*([0-9.]+)", head)
        if m:
            return float(m.group(1))
    return None


def sections(rows: List[Dict[str, str]], name_key: str) -> Dict[str, List[Tuple[Dict[str, str], str, str]]]:
    """{"eager": [(row, module, layer)], "trace": [...]} from the signposts."""
    out: Dict[str, List[Tuple[Dict[str, str], str, str]]] = {"eager": [], "trace": []}
    cur: Optional[str] = None
    module, layer = "", ""
    for r in rows:
        if (r.get("OP TYPE") or "") == "signpost":
            code = r.get(name_key, "")
            if code in ("eager", "trace"):
                cur, module, layer = code, "", ""
            elif code in ("eager_end", "trace_end"):
                cur = None
            elif code.startswith("m:"):
                module, layer = code[2:], ""
            elif code.startswith("c:"):
                layer = code[2:]
            continue
        if cur is not None:
            out[cur].append((r, module, layer))
    return out


def last_replay(sec: List[Tuple[Dict[str, str], str, str]]) -> List[Tuple[Dict[str, str], str, str]]:
    """The ops of the last replay session of every trace id, in CSV order (all ops when the CSV has no ids)."""
    tids = [(r.get("METAL TRACE ID") or "").strip() for r, _, _ in sec]
    sids = [(r.get("METAL TRACE REPLAY SESSION ID") or "").strip() for r, _, _ in sec]
    if not any(sids):
        return sec
    last: Dict[str, str] = {}
    for t, s in zip(tids, sids):
        if s:
            last[t] = s
    return [x for x, t, s in zip(sec, tids, sids) if s and last.get(t) == s]


def summarize(sec: List[Tuple[Dict[str, str], str, str]], name_key: str, freq: float, top: int = 20) -> Dict[str, Any]:
    rows = [r for r, _, _ in sec]
    kern = [num(r, KERNEL) / 1e3 for r in rows]
    fw = [num(r, FW) / 1e3 for r in rows]
    gaps = [num(r, GAP) / 1e3 for r in rows]
    if gaps:
        gaps[0] = 0.0            # the first op's gap reaches back across the section boundary
    starts = [num(r, START) for r in rows if num(r, START) > 0]
    ends = [num(r, END) for r in rows if num(r, END) > 0]
    span = (max(ends) - min(starts)) / freq if starts and ends else None
    total = sum(kern) or 1.0

    def group(keyfn) -> List[Dict[str, Any]]:
        agg: Dict[str, List[float]] = OrderedDict()
        for (r, m, l), k, g in zip(sec, kern, gaps):
            key = keyfn(r, m, l)
            a = agg.setdefault(key, [0, 0.0, 0.0, 0.0])
            a[0] += 1
            a[1] += k
            a[2] += g
            if r.get(name_key, "") in COMPUTE:
                a[3] += k
        return [{"key": key, "ops": int(c), "kernel_us": round(t, 1), "share": round(t / total, 4),
                 "gap_us": round(gs, 1), "compute_us": round(cu, 1)}
                for key, (c, t, gs, cu) in sorted(agg.items(), key=lambda kv: -kv[1][1])]

    fid: Dict[str, int] = defaultdict(int)
    cores = []
    for r in rows:
        f = (r.get("MATH FIDELITY") or "").strip()
        if f:
            fid[f] += 1
        cores.append(num(r, "CORE COUNT"))
    order = sorted(range(len(rows)), key=lambda i: -kern[i])[:top]
    top_ops = []
    for i in order:
        r, m, l = sec[i]
        shape = "x".join(logical(r.get(f"INPUT_0_{d}_PAD[LOGICAL]") or r.get(f"INPUT_0_{d}") or "") for d in "WZYX")
        oshape = "x".join(logical(r.get(f"OUTPUT_0_{d}_PAD[LOGICAL]") or r.get(f"OUTPUT_0_{d}") or "") for d in "WZYX")
        top_ops.append({"index": i, "op": r.get(name_key), "module": m, "layer": l, "kernel_us": round(kern[i], 1),
                        "fw_us": round(fw[i], 1), "cores": int(num(r, "CORE COUNT")),
                        "fidelity": (r.get("MATH FIDELITY") or "").strip(), "input0": shape, "output0": oshape,
                        "in0_dtype": (r.get("INPUT_0_DATATYPE") or "").strip(),
                        "in0_mem": (r.get("INPUT_0_MEMORY") or "").strip(),
                        "out0_mem": (r.get("OUTPUT_0_MEMORY") or "").strip()})
    gap_order = sorted(range(1, len(rows)), key=lambda i: -gaps[i])[:top]
    top_gaps = [{"index": i, "op": rows[i].get(name_key), "after": rows[i - 1].get(name_key),
                 "gap_us": round(gaps[i], 1), "module": sec[i][1], "layer": sec[i][2]} for i in gap_order]
    nz_gaps = [g for g in gaps[1:] if g > 0]
    ops_list = [{"index": i, "op": r.get(name_key), "module": m, "layer": l, "kernel_us": round(k, 2),
                 "gap_us": round(g, 2), "cores": int(num(r, "CORE COUNT")),
                 "fidelity": (r.get("MATH FIDELITY") or "").strip(),
                 "input0": "x".join(logical(r.get(f"INPUT_0_{d}_PAD[LOGICAL]") or r.get(f"INPUT_0_{d}") or "")
                                    for d in "WZYX"),
                 "in0_mem": (r.get("INPUT_0_MEMORY") or "").strip(),
                 "out0_mem": (r.get("OUTPUT_0_MEMORY") or "").strip()}
                for i, ((r, m, l), k, g) in enumerate(zip(sec, kern, gaps))]
    return {
        "ops": len(rows), "kernel_sum_us": round(sum(kern), 1), "fw_sum_us": round(sum(fw), 1),
        "op2op_sum_us": round(sum(gaps), 1),
        "op2op_median_us": round(statistics.median(nz_gaps), 2) if nz_gaps else None,
        "span_us": round(span, 1) if span else None, "freq_mhz": freq,
        "cores_median": statistics.median(cores) if cores else None,
        "ops_below_120_cores": sum(1 for c in cores if 0 < c < 120),
        "compute_us": round(sum(k for r, k in zip(rows, kern) if r.get(name_key, "") in COMPUTE), 1),
        "fidelity": dict(fid),
        "by_op": group(lambda r, m, l: r.get(name_key, "?")),
        "by_category": group(lambda r, m, l: category(r.get(name_key, "?"))),
        "by_module": group(lambda r, m, l: m or "?"),
        "by_layer": group(lambda r, m, l: l or f"[{m or '?'}]"),
        "top_ops": top_ops, "top_gaps": top_gaps, "ops_list": ops_list,
    }


def md_table(rows: List[Dict[str, Any]], n: int, label: str) -> str:
    out = [f"| {label} | ops | kernel us | share | compute us | op-to-op us |", "|---|---:|---:|---:|---:|---:|"]
    for r in rows[:n]:
        out.append(f"| {r['key']} | {r['ops']} | {r['kernel_us']:.1f} | {100 * r['share']:.1f} % | "
                   f"{r['compute_us']:.1f} | {r['gap_us']:.1f} |")
    return "\n".join(out)


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("csv")
    ap.add_argument("out_json")
    ap.add_argument("--md")
    ap.add_argument("--freq-mhz", type=float, default=None)
    a = ap.parse_args()
    path = Path(a.csv)
    with open(path, newline="") as f:
        rows = list(csv.DictReader(f))
    name_key = "OP CODE" if rows and "OP CODE" in rows[0] else "OP NAME"
    freq = a.freq_mhz or chip_freq_mhz(path) or 1350.0
    secs = sections(rows, name_key)
    eager, trace_all = secs["eager"], secs["trace"]
    trace = last_replay(trace_all)
    # the replay inherits the module / layer of the eager op at the same position (identical op sequence)
    same = len(eager) == len(trace) and all(e[0].get(name_key) == t[0].get(name_key) for e, t in zip(eager, trace))
    if same:
        trace = [(r, e[1], e[2]) for (r, _, _), e in zip(trace, eager)]
    segments = OrderedDict()
    for x in trace:
        segments.setdefault((x[0].get("METAL TRACE ID") or "").strip(), []).append(x)
    res = {"csv": str(path), "columns": list(rows[0].keys()) if rows else [], "freq_mhz": freq,
           "trace_matches_eager_op_sequence": same, "trace_section_ops": len(trace_all),
           "trace_ids": sorted({(r.get("METAL TRACE ID") or "").strip() for r, _, _ in trace_all}),
           "replay_sessions": sorted({(r.get("METAL TRACE REPLAY SESSION ID") or "").strip() for r, _, _ in trace_all}),
           "eager": summarize(eager, name_key, freq), "trace": summarize(trace, name_key, freq),
           "segments": {}}
    for tid, sec in segments.items():
        x = summarize(sec, name_key, freq)
        mods = [m for _, m, _ in sec if m]
        res["segments"][tid] = {"first_module": mods[0] if mods else None, "last_module": mods[-1] if mods else None,
                                **{k: x[k] for k in ("ops", "kernel_sum_us", "fw_sum_us", "op2op_sum_us",
                                                     "op2op_median_us", "span_us", "compute_us", "cores_median",
                                                     "ops_below_120_cores", "fidelity")}}
    Path(a.out_json).write_text(json.dumps(res, indent=1) + "\n")
    t = res["trace"]
    lines = [f"csv: {path}", f"freq {freq} MHz; trace op sequence == eager: {same}; trace ids {res['trace_ids']}; "
             f"replay sessions {res['replay_sessions']} ({len(trace_all)} ops in the trace section)"]
    for s in ("eager", "trace"):
        x = res[s]
        lines.append(f"{s}: ops {x['ops']}  kernel {x['kernel_sum_us']:.1f} us (conv/matmul {x['compute_us']:.1f})  "
                     f"fw {x['fw_sum_us']:.1f} us  op2op {x['op2op_sum_us']:.1f} us (median {x['op2op_median_us']})  "
                     f"span {x['span_us']} us  cores median {x['cores_median']}  <120 cores: "
                     f"{x['ops_below_120_cores']}  fidelity {x['fidelity']}")
    for tid, x in res["segments"].items():
        lines.append(f"segment trace id {tid} ({x['first_module']} .. {x['last_module']}): ops {x['ops']}  kernel "
                     f"{x['kernel_sum_us']:.1f} us (conv/matmul {x['compute_us']:.1f})  op2op {x['op2op_sum_us']:.1f} us "
                     f"span {x['span_us']} us  <120 cores {x['ops_below_120_cores']}")
    lines += ["", md_table(t["by_category"], 20, "category (trace)"), "", md_table(t["by_module"], 60, "module (trace)"),
              "", md_table(t["by_op"], 25, "op code (trace)"), "", md_table(t["by_layer"], 60, "layer (trace)"), "",
              "| # | op | module | layer | kernel us | cores | fidelity | input0 | output0 | in mem | out mem |",
              "|---:|---|---|---|---:|---:|---|---|---|---|---|"]
    for k, o in enumerate(t["top_ops"], 1):
        lines.append(f"| {k} | {o['op']} | {o['module']} | {o['layer']} | {o['kernel_us']:.1f} | {o['cores']} | "
                     f"{o['fidelity']} | {o['input0']} | {o['output0']} | {o['in0_mem']} | {o['out0_mem']} |")
    lines += ["", "largest op-to-op gaps (trace): " + json.dumps(t["top_gaps"])]
    text = "\n".join(lines)
    print(text)
    if a.md:
        Path(a.md).write_text(text + "\n")


if __name__ == "__main__":
    main()