@cannbot-plugin/cannbot-harness 0.0.0-stage → 0.1.5

package/assets/ops-direct-invoke/skills/ops-profiling/msopprof-visualization/scripts/visualize_parsers.py

@@ -0,0 +1,1555 @@
+# ----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# ----------------------------------------------------------------------------------------------------------
+
+from __future__ import annotations
+
+import json
+import logging
+import math
+import re
+import statistics
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, Iterator, List, Mapping, NamedTuple, Optional, Sequence, Tuple
+from visualize_common import (
+    ArtifactSet,
+    flatten_numeric,
+    humanize_key,
+    iter_dicts,
+    mean,
+    metric_is_percent,
+    percentile,
+    to_float,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def _parse_embedded_blob(data: bytes, start: int, end: int) -> Tuple[int, Optional[Dict[str, Any]]]:
+    """Parse the balanced blob ``data[start:end + 1]``.
+
+    Returns ``(next_index, record)``. An unparseable blob yields
+    ``(start + 1, None)`` so the caller resumes its garbage-skipping rescan one
+    byte later; valid JSON that is not a dict yields ``(end + 1, None)``.
+    """
+    blob = data[start:end + 1]
+    try:
+        obj = json.loads(blob.decode("utf-8"))
+    except Exception:
+        return start + 1, None
+    record = obj if isinstance(obj, dict) else None
+    return end + 1, record
+
+
+def _string_scan_state(c: int, in_string: bool, escaped: bool) -> Tuple[bool, bool]:
+    """Advance the string/escape state machine for one byte inside a string."""
+    if escaped:
+        return in_string, False
+    if c == 92:
+        return in_string, True
+    if c == 34:
+        return False, False
+    return in_string, escaped
+
+
+def _scan_embedded_candidate(data: bytes, start: int, n: int) -> Tuple[int, Optional[Dict[str, Any]]]:
+    """Scan one candidate JSON object beginning at ``start``.
+
+    Returns ``(next_index, record)``; see ``_parse_embedded_blob`` for the
+    rescan semantics of unparseable candidates.
+    """
+    depth = 0
+    in_string = False
+    escaped = False
+    for j in range(start, n):
+        c = data[j]
+        if in_string:
+            in_string, escaped = _string_scan_state(c, in_string, escaped)
+            continue
+        if c == 34:
+            in_string = True
+            continue
+        if c == 123:
+            depth += 1
+            continue
+        if c != 125:
+            continue
+        depth -= 1
+        if depth == 0:
+            return _parse_embedded_blob(data, start, j)
+    return start + 1, None
+
+
+def extract_embedded_json_records(data: bytes) -> List[Dict[str, Any]]:
+    records: List[Dict[str, Any]] = []
+    i, n = 0, len(data)
+    while i < n:
+        start = data.find(b"{", i)
+        if start < 0:
+            break
+        i, record = _scan_embedded_candidate(data, start, n)
+        if record is not None:
+            records.append(record)
+    return records
+
+
+def load_visualize_records(path: Optional[Path]) -> List[Dict[str, Any]]:
+    if not path or not path.is_file() or path.stat().st_size == 0:
+        return []
+    try:
+        return extract_embedded_json_records(path.read_bytes())
+    except Exception as exc:
+        logger.debug("Failed to load visualize records from %s: %s", path, exc)
+        return []
+
+
+def collect_advice(records: Sequence[Mapping[str, Any]]) -> List[str]:
+    found: List[str] = []
+
+    def add(value: Any) -> None:
+        if value is None:
+            return
+        if isinstance(value, str):
+            for line in value.splitlines():
+                cleaned = re.sub(r"^[\s\t\-\d\)\.]+", "", line).strip()
+                if cleaned and cleaned not in found:
+                    found.append(cleaned)
+        elif isinstance(value, list):
+            for item in value:
+                add(item)
+        elif isinstance(value, Mapping):
+            for item in value.values():
+                add(item)
+
+    for record in records:
+        if "advice" in record:
+            add(record.get("advice"))
+        for obj in iter_dicts(record):
+            if obj is record:
+                continue
+            if "advice" in obj:
+                add(obj.get("advice"))
+    return found
+
+
+def parse_block_detail(records: Sequence[Mapping[str, Any]]) -> List[Dict[str, Any]]:
+    for record in records:
+        detail = record.get("block_detail")
+        if not isinstance(detail, Mapping):
+            continue
+        headers = [str(h) for h in (detail.get("head_name") or [])]
+        rows: List[Dict[str, Any]] = []
+        for item in detail.get("row") or []:
+            values = list(item.get("value") or []) if isinstance(item, Mapping) else []
+            row: Dict[str, Any] = {}
+            for h, v in zip(headers, values):
+                row[h] = v
+            rows.append(row)
+        return rows
+    return []
+
+
+def parse_basic(records: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
+    for record in records:
+        detail = record.get("block_detail")
+        if not isinstance(detail, Mapping):
+            continue
+        blocks: List[Dict[str, Any]] = []
+        for row in parse_block_detail([record]):
+            block_id = next((v for k, v in row.items() if "block" in k.lower() and "id" in k.lower()), None)
+            core_type = next((v for k, v in row.items() if "core" in k.lower() and "type" in k.lower()), "")
+            duration = next((v for k, v in row.items() if "duration" in k.lower()), None)
+            duration_us = to_float(duration)
+            if block_id is None or duration_us is None:
+                continue
+            blocks.append({"block_id": str(block_id), "core_type": str(core_type or ""), "duration_us": duration_us})
+        current_freq = to_float(record.get("cur_freq"))
+        rated_freq = to_float(record.get("rated_freq"))
+        frequency_ratio = None
+        if current_freq is not None and rated_freq not in (None, 0):
+            frequency_ratio = current_freq / rated_freq * 100.0
+        return {
+            "available": True,
+            "name": str(record.get("name", "")),
+            "duration_us": to_float(record.get("duration")),
+            "op_type": str(record.get("op_type", "")),
+            "block_dim": str(record.get("block_dim", "")),
+            "mix_block_dim": str(record.get("mix_block_dim", "")),
+            "soc": str(record.get("soc", "")),
+            "device_id": str(record.get("device_id", "")),
+            "current_frequency_mhz": current_freq,
+            "rated_frequency_mhz": rated_freq,
+            "frequency_ratio_percent": frequency_ratio,
+            "blocks": blocks,
+        }
+    return {"available": False, "blocks": []}
+
+
+OCCUPANCY_PACKED_LAYOUTS: Dict[str, Dict[str, Any]] = {
+    # Exact SoC string emitted by the supplied Ascend 950 profiler payload.
+    "ASCEND950PR9579": {
+        "lanes_per_core": 2,
+        # Source-slot order verified against the authoritative Ascend 950 occupancy view.
+        # Unlisted groups keep source order; raw IDs are always retained.
+        "slot_order": {
+            0: [1, 2, 3, 0],
+            1: [0, 1, 2, 3],
+            2: [0, 1, 2, 3],
+            3: [3, 2, 1, 0],
+            4: [0, 1, 2, 3],
+            5: [0, 1, 2, 3],
+            6: [0, 1, 2, 3],
+            128: [0, 3, 1, 2],
+            129: [0, 1, 2, 3],
+            130: [0, 1, 2, 3],
+            131: [3, 2, 1, 0],
+            132: [0, 1, 2, 3],
+            133: [0, 1, 2, 3],
+            134: [0, 1, 2, 3],
+        },
+    },
+}
+
+
+OCCUPANCY_PACKED_LAYOUT_ALIASES: Dict[str, str] = {
+    "910A5": "ASCEND950PR9579",
+}
+
+
+def normalize_soc_name(value: Any) -> str:
+    return re.sub(r"[^A-Za-z0-9]", "", str(value or "")).upper()
+
+
+def occupancy_packed_schema(raw_items: Sequence[Mapping[str, Any]], soc: str) -> Optional[Dict[str, Any]]:
+    if not raw_items:
+        return None
+    grouped: Dict[str, List[Mapping[str, Any]]] = defaultdict(list)
+    for item in raw_items:
+        grouped[str(item.get("raw_core_id", ""))].append(item)
+    duplicate_groups = [rows for rows in grouped.values() if len(rows) > 1]
+    if not duplicate_groups:
+        return None
+    # Packed schema signature: repeated raw core IDs, identical nested subcore IDs,
+    # and one engine family. This avoids rewriting ordinary multi-subcore payloads.
+    repeated_subcore_ids = all(len({str(x.get("raw_subcore_id", "")) for x in rows}) == 1 for rows in duplicate_groups)
+    engine_types = {str(x.get("subcore_type", "")).lower() for x in raw_items if str(x.get("subcore_type", "")).strip()}
+    if not repeated_subcore_ids or len(engine_types) != 1:
+        return None
+    normalized_soc = normalize_soc_name(soc)
+    layout_key = OCCUPANCY_PACKED_LAYOUT_ALIASES.get(normalized_soc, normalized_soc)
+    layout = OCCUPANCY_PACKED_LAYOUTS.get(layout_key)
+    if layout:
+        result = dict(layout)
+        result["layout_key"] = layout_key
+        return result
+    # Conservative generic fallback for vector payloads: only expand when every
+    # repeated group can form two-lane physical cores exactly.
+    if engine_types == {"vector"} and all(len(rows) % 2 == 0 for rows in duplicate_groups):
+        return {"lanes_per_core": 2, "slot_order": {}}
+    return None
+
+
+def _raw_core_sort_key(value: str) -> Tuple[int, Any]:
+    try:
+        return (0, int(value))
+    except (TypeError, ValueError):
+        return (1, value)
+
+
+def _assemble_packed_logical(
+    grouped: Mapping[str, List[Dict[str, Any]]],
+    slot_orders: Mapping[Any, Any],
+    lanes_per_core: int,
+) -> Tuple[List[Dict[str, Any]], int]:
+    logical: List[Dict[str, Any]] = []
+    logical_core_id = 0
+    for raw_id in sorted(grouped, key=_raw_core_sort_key):
+        rows = grouped[raw_id]
+        try:
+            numeric_raw = int(raw_id)
+        except (TypeError, ValueError):
+            numeric_raw = None
+        order = slot_orders.get(numeric_raw)
+        if isinstance(order, list) and len(order) == len(rows) and sorted(order) == list(range(len(rows))):
+            rows = [rows[i] for i in order]
+        for offset in range(0, len(rows), lanes_per_core):
+            # Do not invent a missing lane. Keep the residual entry visible.
+            chunk = rows[offset:offset + lanes_per_core]
+            for lane_id, item in enumerate(chunk):
+                item["core_id"] = logical_core_id
+                item["subcore_id"] = lane_id
+                item["logical_core_id"] = logical_core_id
+                item["logical_subcore_id"] = lane_id
+                logical.append(item)
+            logical_core_id += 1
+    return logical, logical_core_id
+
+
+def normalize_packed_occupancy(raw_items: Sequence[Mapping[str,
+    Any]], soc: str) -> Tuple[List[Dict[str, Any]], Dict[str, Any]]:
+    schema = occupancy_packed_schema(raw_items, soc)
+    if not schema:
+        return [dict(x) for x in raw_items], {
+            "adapter": "native",
+            "soc": soc,
+            "logical_core_count": len({str(x.get('core_id')) for x in raw_items}),
+            "raw_core_count": len({str(x.get('raw_core_id')) for x in raw_items}),
+        }
+
+    lanes_per_core = max(1, int(schema.get("lanes_per_core", 2)))
+    slot_orders = schema.get("slot_order") or {}
+    grouped: Dict[str, List[Dict[str, Any]]] = defaultdict(list)
+    for item in raw_items:
+        grouped[str(item.get("raw_core_id", ""))].append(dict(item))
+    logical, logical_core_id = _assemble_packed_logical(grouped, slot_orders, lanes_per_core)
+
+    return logical, {
+        "adapter": "packed-core-group",
+        "soc": soc,
+        "lanes_per_core": lanes_per_core,
+        "logical_core_count": logical_core_id,
+        "raw_core_count": len(grouped),
+        "raw_core_ids": sorted(grouped, key=_raw_core_sort_key),
+        "layout_key": schema.get("layout_key"),
+    }
+
+
+def _occupancy_op_entries(records: Sequence[Mapping[str, Any]]) -> Tuple[List[Mapping[str, Any]], str]:
+    op_entries: List[Mapping[str, Any]] = []
+    soc = ""
+    for record in records:
+        if isinstance(record.get("op_detail"), list):
+            op_entries.extend(x for x in record["op_detail"] if isinstance(x, Mapping))
+            if record.get("soc") and not soc:
+                soc = str(record.get("soc"))
+    return op_entries, soc
+
+
+def _occupancy_metrics_of(sub: Mapping[str, Any]) -> Dict[str, Any]:
+    metrics: Dict[str, Any] = {}
+    for key, value in flatten_numeric(sub).items():
+        if key in {"subcore_id", "core_id"}:
+            continue
+        if key.lower().endswith("_id") or key.lower().endswith(".id"):
+            continue
+        metrics[key] = value
+    return metrics
+
+
+def _occupancy_raw_item(
+    core: Mapping[str, Any],
+    entry_index: int,
+    detail_index: int,
+    sub: Mapping[str, Any],
+) -> Optional[Dict[str, Any]]:
+    metrics = _occupancy_metrics_of(sub)
+    if not metrics:
+        return None
+    core_id = core.get("core_id")
+    return {
+        "core_id": core_id,
+        "subcore_id": sub.get("subcore_id"),
+        "subcore_type": sub.get("subcore_type", ""),
+        "raw_core_id": core_id,
+        "raw_subcore_id": sub.get("subcore_id"),
+        "source_entry_index": entry_index,
+        "source_detail_index": detail_index,
+        "metrics": metrics,
+    }
+
+
+def _occupancy_raw_items(op_entries: Sequence[Mapping[str, Any]]) -> List[Dict[str, Any]]:
+    raw_items: List[Dict[str, Any]] = []
+    for entry_index, core in enumerate(op_entries):
+        details = core.get("core_detail") or []
+        for detail_index, sub in enumerate(details):
+            if not isinstance(sub, Mapping):
+                continue
+            item = _occupancy_raw_item(core, entry_index, detail_index, sub)
+            if item is not None:
+                raw_items.append(item)
+    return raw_items
+
+
+def _occupancy_metrics(cores: Sequence[Mapping[str, Any]]) -> List[Dict[str, Any]]:
+    metric_values: Dict[str, List[Optional[float]]] = defaultdict(list)
+    for item in cores:
+        for key, value in (item.get("metrics") or {}).items():
+            metric_values[key].append(value)
+
+    metrics: List[Dict[str, Any]] = []
+    preferred_tokens = ["cycles", "throughput", "hit", "instructions", "instruction_per_cycle"]
+    for key, values in metric_values.items():
+        finite = [v for v in values if v is not None]
+        if not finite:
+            continue
+        priority = min((preferred_tokens.index(t) for t in preferred_tokens if t in key.lower()), default=99)
+        signal = max(abs(v) for v in finite) + (max(finite) - min(finite))
+        med = statistics.median(finite)
+        metrics.append({
+            "key": key,
+            "label": humanize_key(key),
+            "format": "percent" if metric_is_percent(key, values) else "number",
+            "min": min(finite),
+            "max": max(finite),
+            "median": med,
+            "p10": percentile(finite, 0.10),
+            "p90": percentile(finite, 0.90),
+            "spread_pct": ((max(finite) - min(finite)) / med * 100.0) if med not in {0, None} else None,
+            "priority": priority,
+            "signal": signal,
+        })
+    # Do not default to an all-zero metric merely because its name is normally preferred.
+    metrics.sort(key=lambda x: (0 if x["signal"] > 1e-12 else 1, x["priority"], x["label"]))
+    for metric in metrics:
+        metric.pop("priority", None)
+        metric.pop("signal", None)
+    return metrics
+
+
+def parse_occupancy(records: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
+    op_entries, soc = _occupancy_op_entries(records)
+    raw_items = _occupancy_raw_items(op_entries)
+    cores, layout = normalize_packed_occupancy(raw_items, soc)
+    metrics = _occupancy_metrics(cores)
+    return {"available": bool(cores and metrics), "cores": cores, "metrics": metrics, "layout": layout}
+
+
+def parse_compute(records: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
+    raw_items: List[Mapping[str, Any]] = []
+    for record in records:
+        if isinstance(record.get("subblock_detail"), list):
+            raw_items.extend(x for x in record["subblock_detail"] if isinstance(x, Mapping))
+
+    items: List[Dict[str, Any]] = []
+    for item in raw_items:
+        value = to_float(item.get("value"))
+        if value is None:
+            continue
+        items.append({
+            "block_id": str(item.get("block_id", "")),
+            "block_type": str(item.get("block_type", "") or "Compute"),
+            "name": str(item.get("name", "Metric") or "Metric"),
+            "unit": item.get("unit"),
+            "value": value,
+            "origin_value": to_float(item.get("origin_value")),
+        })
+
+    items.sort(key=lambda x: (
+        int(x["block_id"]) if str(x["block_id"]).isdigit() else 10**9,
+        str(x["block_type"]),
+        str(x["name"]),
+        str(x.get("unit")),
+    ))
+    # Mixed numeric/non-numeric Block IDs cannot be compared directly; order
+    # numeric IDs first (by value), then the rest lexicographically.
+    block_ids = sorted({x["block_id"] for x in items}, key=lambda x: (0, int(x)) if x.isdigit() else (1, x))
+    block_types = sorted({x["block_type"] for x in items})
+
+    # Keep a compact aggregate view for compatibility with earlier consumers and tests.
+    grouped: Dict[Tuple[str, str, str], List[Dict[str, Any]]] = defaultdict(list)
+    for item in items:
+        grouped[(item["name"], item["block_type"], str(item.get("unit")))].append(item)
+    metrics: List[Dict[str, Any]] = []
+    for (name, block_type, unit_key), group in grouped.items():
+        vals = [x["value"] for x in group]
+        is_percent = str(group[0].get("unit")) == "3" or metric_is_percent(name, vals)
+        metrics.append({
+            "key": f"{block_type}::{name}::unit={unit_key}",
+            "name": name,
+            "display_name": name + (" (%)" if is_percent else " (raw)"),
+            "block_type": block_type,
+            "format": "percent" if is_percent else "number",
+            "mean": sum(vals) / len(vals),
+            "min": min(vals),
+            "max": max(vals),
+            "blocks": [{"block_id": x["block_id"], "value": x["value"]} for x in group],
+        })
+    return {"available": bool(items), "items": items, "block_ids": block_ids,
+        "block_types": block_types, "metrics": metrics}
+
+
+def _tpb_row_item(row: Mapping[str, Any], block_id: str, headers: Sequence[str]) -> Dict[str, Any]:
+    values = list(row.get("value") or [])
+    metric_headers = headers[1:] if len(headers) == len(values) + 1 else headers[:len(values)]
+    item: Dict[str, Any] = {"block_id": block_id, "name": str(row.get("name", ""))}
+    for h, v in zip(metric_headers, values):
+        item[str(h)] = v
+    return item
+
+
+def _tpb_table_item(
+    table: Mapping[str, Any],
+    block_id: str,
+    grouped: Dict[str, List[Dict[str, Any]]],
+) -> Dict[str, Any]:
+    table_name = str(table.get("table_name", "Unknown"))
+    headers = [str(h) for h in (table.get("header_name") or [])]
+    rows: List[Dict[str, Any]] = []
+    for row in table.get("row") or []:
+        if not isinstance(row, Mapping):
+            continue
+        item = _tpb_row_item(row, block_id, headers)
+        rows.append(item)
+        grouped[table_name].append(item)
+    return {"name": table_name, "headers": headers, "rows": rows}
+
+
+def _tpb_block_item(
+    block: Mapping[str, Any],
+    grouped: Dict[str, List[Dict[str, Any]]],
+) -> Dict[str, Any]:
+    block_id = str(block.get("block_id", ""))
+    normalized_tables: List[Dict[str, Any]] = []
+    for table in block.get("table_detail") or []:
+        if not isinstance(table, Mapping):
+            continue
+        normalized_tables.append(_tpb_table_item(table, block_id, grouped))
+    return {"block_id": block_id, "tables": normalized_tables}
+
+
+def normalize_table_per_block(records: Sequence[Mapping[str,
+    Any]]) -> Tuple[List[Dict[str, Any]], Dict[str, List[Dict[str, Any]]]]:
+    blocks: List[Dict[str, Any]] = []
+    grouped: Dict[str, List[Dict[str, Any]]] = defaultdict(list)
+    for record in records:
+        source = record.get("table_per_block")
+        if not isinstance(source, list):
+            continue
+        for block in source:
+            if not isinstance(block, Mapping):
+                continue
+            blocks.append(_tpb_block_item(block, grouped))
+    return blocks, dict(grouped)
+
+
+def _memory_table_item(table: Mapping[str, Any]) -> Optional[Dict[str, Any]]:
+    headers = [str(h) for h in (table.get("headers") or [])]
+    value_headers = headers[1:] if headers else []
+    rows: List[Dict[str, Any]] = []
+    for row in table.get("rows") or []:
+        values: Dict[str, Any] = {}
+        for key, value in row.items():
+            if key in {"block_id", "name"}:
+                continue
+            values[str(key)] = value
+        rows.append({"name": str(row.get("name", "")), "values": values})
+    if not rows:
+        return None
+    return {"name": str(table.get("name", "Unknown")),
+        "headers": headers or (["Name"] + value_headers), ("r"
+        "ows"): rows}
+
+
+def _memory_block_tables(raw_blocks: Sequence[Mapping[str, Any]]) -> List[Dict[str, Any]]:
+    blocks: List[Dict[str, Any]] = []
+    for block in raw_blocks:
+        tables: List[Dict[str, Any]] = []
+        for table in block.get("tables") or []:
+            item = _memory_table_item(table)
+            if item is not None:
+                tables.append(item)
+        if tables:
+            blocks.append({"block_id": str(block.get("block_id", "")), "tables": tables})
+    return blocks
+
+
+def _memory_core_map_entry(core: Mapping[str, Any]) -> Tuple[str, List[Dict[str, Any]], Dict[str, Any]]:
+    core_id = str(core.get("core_no", ""))
+    paths: List[Dict[str, Any]] = []
+    for unit in core.get("memory_unit") or []:
+        if not isinstance(unit, Mapping):
+            continue
+        paths.append({
+            "memory_path": unit.get("memory_path"),
+            "request": to_float(unit.get("request")),
+            "bandwidth": to_float(unit.get("bandwidth")),
+            "peak_ratio": to_float(unit.get("peak_ratio")),
+            "display": unit.get("display"),
+        })
+    summary: Dict[str, Any] = {}
+    for key, value in core.items():
+        if key in {"core_no", "memory_unit", "advice", "soc", "op_type"}:
+            continue
+        if not isinstance(value, (Mapping, list)):
+            continue
+        summary[key] = value
+    return core_id, paths, summary
+
+
+def _memory_core_maps(
+    records: Sequence[Mapping[str, Any]],
+) -> Tuple[Dict[str, List[Dict[str, Any]]], Dict[str, Dict[str, Any]]]:
+    core_paths: Dict[str, List[Dict[str, Any]]] = {}
+    core_summaries: Dict[str, Dict[str, Any]] = {}
+    for record in records:
+        source = record.get("core_memory_map")
+        if not isinstance(source, list):
+            continue
+        for core in source:
+            if not isinstance(core, Mapping):
+                continue
+            core_id, paths, summary = _memory_core_map_entry(core)
+            core_paths[core_id] = paths
+            core_summaries[core_id] = summary
+    return core_paths, core_summaries
+
+
+def _memory_row_keys(row_items: Sequence[Mapping[str, Any]]) -> List[str]:
+    keys: List[str] = []
+    for row_item in row_items:
+        for key in (row_item.get("values") or {}):
+            if key in keys:
+                continue
+            keys.append(key)
+    return keys
+
+
+def _memory_group_rows(rows: Sequence[Mapping[str, Any]]) -> List[Dict[str, Any]]:
+    by_name: Dict[str, List[Dict[str, Any]]] = defaultdict(list)
+    for row in rows:
+        by_name[str(row.get("name", ""))].append(row)
+    normalized_rows: List[Dict[str, Any]] = []
+    for row_name, row_items in by_name.items():
+        metrics: Dict[str, Optional[float]] = {}
+        for key in _memory_row_keys(row_items):
+            metrics[key] = mean((ri.get("values") or {}).get(key) for ri in row_items)
+        normalized_rows.append({
+            "name": row_name,
+            "metrics": metrics,
+        })
+    return normalized_rows
+
+
+def _memory_groups(blocks: Sequence[Mapping[str, Any]]) -> List[Dict[str, Any]]:
+    groups: List[Dict[str, Any]] = []
+    by_table: Dict[str, List[Dict[str, Any]]] = defaultdict(list)
+    for block in blocks:
+        for table in block.get("tables") or []:
+            for row in table.get("rows") or []:
+                by_table[table["name"]].append(row)
+    for table_name, rows in by_table.items():
+        field_names = set()
+        for row in rows:
+            for key in (row.get("values") or {}).keys():
+                field_names.add(str(key).lower())
+        cache_summary = (
+            "cache" in table_name.lower()
+            and any("hit" in f or "miss" in f for f in field_names)
+            and not any("throughput" in f or "bandwidth" in f or "request" in f for f in field_names)
+        )
+        if cache_summary:
+            continue
+        normalized_rows = _memory_group_rows(rows)
+        if normalized_rows:
+            groups.append({"name": table_name, "rows": normalized_rows})
+    return groups
+
+
+def parse_memory(records: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
+    raw_blocks, _ = normalize_table_per_block(records)
+    blocks = _memory_block_tables(raw_blocks)
+    core_paths, core_summaries = _memory_core_maps(records)
+    groups = _memory_groups(blocks)
+    advice = collect_advice(records)
+    return {
+        "available": bool(blocks or core_paths),
+        "blocks": blocks,
+        "core_paths": core_paths,
+        "core_summaries": core_summaries,
+        "groups": groups,
+        "advice": advice,
+    }
+
+
+def _canonical_cache_field(name: Any) -> Optional[str]:
+    low = str(name).strip().lower().replace("_", " ")
+    if "hit rate" in low or ("hit" in low and "rate" in low):
+        return "hit_rate"
+    if low == "hit" or low.endswith(" hit"):
+        return "hit"
+    if low == "miss" or low.endswith(" miss"):
+        return "miss"
+    if low == "total" or low.endswith(" total"):
+        return "total"
+    return None
+
+
+class _CacheAccumulators(NamedTuple):
+    entries: Dict[str, float]
+    metric_keys: List[str]
+    aggregate: Dict[str, List[float]]
+
+
+def _cache_row_fields(
+    row: Mapping[str, Any],
+    row_name: str,
+    family: Dict[str, Optional[float]],
+    accumulators: _CacheAccumulators,
+) -> None:
+    for field, raw in row.items():
+        if field in {"block_id", "name"}:
+            continue
+        value = to_float(raw)
+        if value is None:
+            continue
+        key = f"{row_name} · {field}"
+        accumulators.entries[key] = value
+        accumulators.aggregate[key].append(value)
+        if key not in accumulators.metric_keys:
+            accumulators.metric_keys.append(key)
+        canonical = _canonical_cache_field(field)
+        if canonical:
+            family[canonical] = value
+
+
+def _cache_table_field_names(table: Mapping[str, Any]) -> set:
+    fields = set()
+    for row in table["rows"]:
+        for field_name in row.keys():
+            if field_name in {"block_id", "name"}:
+                continue
+            fields.add(str(field_name).lower())
+    return fields
+
+
+def _cache_block_entries(
+    block: Mapping[str, Any],
+    family_by_block: Dict[str, Dict[str, Dict[str, Optional[float]]]],
+    metric_keys: List[str],
+    series_keys: List[str],
+    aggregate: Dict[str, List[float]],
+) -> Dict[str, float]:
+    entries: Dict[str, float] = {}
+    accumulators = _CacheAccumulators(entries, metric_keys, aggregate)
+    block_id = str(block["block_id"])
+    for table in block["tables"]:
+        fields = _cache_table_field_names(table)
+        is_cache_summary = "cache" in table["name"].lower() and any("hit" in f or "miss" in f for f in fields)
+        if not is_cache_summary:
+            continue
+        for row in table["rows"]:
+            row_name = str(row.get("name", "")).strip() or str(table["name"])
+            family = family_by_block[block_id].setdefault(row_name, {"hit": None, "miss": None, "total": None, ("h"
+                "it_rate"): None})
+            _cache_row_fields(row, row_name, family, accumulators)
+            if row_name not in series_keys:
+                series_keys.append(row_name)
+    return entries
+
+
+def _cache_summary(aggregate: Mapping[str, List[float]]) -> List[Dict[str, Any]]:
+    summary = []
+    for key, vals in aggregate.items():
+        summary.append({
+            "key": key,
+            "label": key,
+            "mean": sum(vals) / len(vals),
+            "min": min(vals),
+            "max": max(vals),
+            "format": "percent" if metric_is_percent(key, vals) else "number",
+        })
+    summary.sort(key=lambda x: (0 if "hit rate" in x["key"].lower() else 1, x["key"]))
+    return summary
+
+
+def _cache_family_cells(
+    family_name: str,
+    cache_blocks: Sequence[Mapping[str, Any]],
+    family_by_block: Mapping[str, Any],
+) -> Tuple[List[Dict[str, Any]], float, float]:
+    raw_cells: List[Dict[str, Any]] = []
+    hit_sum = miss_sum = 0.0
+    for block in cache_blocks:
+        values = family_by_block.get(str(block["block_id"]), {}).get(family_name)
+        if not values:
+            continue
+        hit = to_float(values.get("hit")) or 0.0
+        miss = to_float(values.get("miss")) or 0.0
+        total = to_float(values.get("total"))
+        if total is None:
+            total = hit + miss
+        hit_rate = to_float(values.get("hit_rate"))
+        if hit_rate is None and total > 0:
+            hit_rate = hit / total * 100.0
+        hit_sum += hit
+        miss_sum += miss
+        raw_cells.append({
+            "index": str(block["block_id"]),
+            "index_kind": "block_id",
+            "block_id": str(block["block_id"]),
+            "hit": hit,
+            "miss": miss,
+            "total": total,
+            "hit_rate": hit_rate,
+            "source_links": [],
+        })
+    for cell in raw_cells:
+        cell["hit_share_percent"] = (cell["hit"] / hit_sum * 100.0) if hit_sum > 0 else 0.0
+        cell["miss_share_percent"] = (cell["miss"] / miss_sum * 100.0) if miss_sum > 0 else 0.0
+    return raw_cells, hit_sum, miss_sum
+
+
+def _cache_family_models(
+    series_keys: Sequence[str],
+    cache_blocks: Sequence[Mapping[str, Any]],
+    family_by_block: Mapping[str, Any],
+) -> List[Dict[str, Any]]:
+    family_models: List[Dict[str, Any]] = []
+    for family_name in series_keys:
+        raw_cells, hit_sum, miss_sum = _cache_family_cells(family_name, cache_blocks, family_by_block)
+        if raw_cells:
+            cols = max(1, math.ceil(math.sqrt(len(raw_cells))))
+            rows_count = math.ceil(len(raw_cells) / cols)
+            family_models.append({
+                "name": family_name,
+                "index_kind": "block_id",
+                "index_label": "Profiler Block ID",
+                "granularity_note": ("This capture exposes per-block cache aggregates; native cacheline "
+                    "indices and instruction linkage are not present in the payload."),
+                "columns": cols,
+                "rows": rows_count,
+                "cells": raw_cells,
+                "totals": {"hit": hit_sum, "miss": miss_sum, "events": hit_sum + miss_sum},
+            })
+    family_models.sort(key=lambda x: (0 if x["name"].lower() == "l2 cache total" else 1,
+        0 if "l2 cache" in x["name"].lower() else 1, x["name"].lower()))
+    return family_models
+
+
+def parse_cache(records: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
+    """Normalize cache summaries into paired Hit/Miss event maps.
+
+    The profiler payload used by many on-board captures is block aggregated, not
+    cacheline resolved.  The model therefore records the index semantics instead
+    of labelling block IDs as cacheline indices.  A future native cacheline
+    provider can populate the same family/cell contract with
+    ``index_kind=cacheline`` without changing the renderer.
+    """
+    blocks, _ = normalize_table_per_block(records)
+    cache_blocks: List[Dict[str, Any]] = []
+    metric_keys: List[str] = []
+    series_keys: List[str] = []
+    aggregate: Dict[str, List[float]] = defaultdict(list)
+    family_by_block: Dict[str, Dict[str, Dict[str, Optional[float]]]] = defaultdict(dict)
+
+    for block in blocks:
+        entries = _cache_block_entries(block, family_by_block, metric_keys, series_keys, aggregate)
+        if entries:
+            cache_blocks.append({"block_id": str(block["block_id"]), "values": entries})
+
+    summary = _cache_summary(aggregate)
+    metric_keys.sort(key=lambda x: (0 if "hit rate" in x.lower() else 1, x))
+    family_models = _cache_family_models(series_keys, cache_blocks, family_by_block)
+    supports_source_linkage = False
+    for family in family_models:
+        for cell in family.get("cells") or []:
+            if cell.get("source_links"):
+                supports_source_linkage = True
+    return {
+        "available": bool(cache_blocks),
+        "schema": "msopprof-cache-heatmap/v2",
+        "granularity": "block",
+        "index_kind": "block_id",
+        "blocks": cache_blocks,
+        "metrics": metric_keys,
+        "summary": summary,
+        "families": family_models,
+        "default_family": family_models[0]["name"] if family_models else None,
+        "supports_cacheline_detail": False,
+        "supports_source_linkage": supports_source_linkage,
+    }
+
+
+def _walk_json_nodes(value: Any) -> Iterator[Mapping[str, Any]]:
+    if isinstance(value, Mapping):
+        yield value
+        for child in value.values():
+            yield from _walk_json_nodes(child)
+    elif isinstance(value, list):
+        for child in value:
+            yield from _walk_json_nodes(child)
+
+
+def _first_key(mapping: Mapping[str, Any], names: Sequence[str]) -> Any:
+    normalized = {str(k).strip().lower().replace("-", "_"): v for k, v in mapping.items()}
+    for name in names:
+        key = name.lower().replace("-", "_")
+        if key in normalized:
+            return normalized[key]
+    return None
+
+
+def _numeric_pair(value: Any) -> Tuple[Optional[float], Optional[float]]:
+    if isinstance(value, (list, tuple)) and len(value) >= 2:
+        return to_float(value[0]), to_float(value[1])
+    if isinstance(value, Mapping):
+        return to_float(_first_key(value, ["start", "begin", "min"])), to_float(_first_key(value, ["end", "finish", ("m"
+            "ax")]))
+    return None, None
+
+
+def _onchip_missing_diagnostic() -> Dict[str, Any]:
+    return {
+        "available": False,
+        "diagnostic": {
+            "code": "memory_info_missing",
+            "title": "On-Chip Memory unavailable",
+            "explanation": ("The collection does not contain a manifest-declared memory_info.json file. "
+                "MemoryDetail counters cannot be substituted for allocation lifetime/address data."),
+            "evidence": ["Required artifact: memory_info.json", ("Required views: memory block map, "
+                "selected allocation detail, bank/group distribution")],
+            "requirements": ["Generate memory_info.json for the selected kernel/build.", ("Declare or "
+                "pass the exact file to the visualization pipeline.")],
+        },
+    }
+
+
+def _onchip_errors(raw: Any) -> List[str]:
+    errors: List[str] = []
+    if isinstance(raw, Mapping):
+        for key in ["error", "errors", "error_message", "compile_error", "message"]:
+            value = _first_key(raw, [key])
+            if value and any(token in str(value).lower() for token in ["fail", "error", "失败"]):
+                items = value if isinstance(value, list) else [value]
+                errors.extend(str(x) for x in items)
+    return errors
+
+
+def _onchip_int_list(value: Any) -> List[int]:
+    if isinstance(value, (int, float, str)):
+        return [int(x) for x in re.findall(r"\d+", str(value))]
+    if isinstance(value, list):
+        numbers: List[int] = []
+        for item in value:
+            for token in re.findall(r"\d+", str(item)):
+                numbers.append(int(token))
+        return numbers
+    return []
+
+
+def _onchip_candidate(node: Mapping[str, Any], ordinal: int) -> Optional[Dict[str, Any]]:
+    start = to_float(_first_key(node, ["start", "start_time",
+        "life_start", "begin", "alloc_order", "start_line", ("a"
+        "lloc_line"), "birth"]))
+    end = to_float(_first_key(node, ["end", "end_time", "life_end", "finish", "free_order", "end_line", ("free_l"
+        "ine"), "death"]))
+    life_start, life_end = _numeric_pair(_first_key(node, ["lifetime", "life_time", "life", "interval"]))
+    start = start if start is not None else life_start
+    end = end if end is not None else life_end
+    size = to_float(_first_key(node, ["size", "size_bytes", "buffer_size", "memory_size", "bytes", "tensor_size"]))
+    addr_start = to_float(_first_key(node, ["address", "offset", "start_address", "address_start", "addr", ("ub_"
+        "offset")]))
+    addr_end = to_float(_first_key(node, ["end_address", "address_end", "addr_end"]))
+    if addr_end is None and addr_start is not None and size is not None:
+        addr_end = addr_start + size
+    signals = sum(x is not None for x in [start, end, size, addr_start, addr_end])
+    if signals < 3 or start is None or end is None:
+        return None
+    name = _first_key(node, ["name", "tensor_name", "buffer_name", "id", "block_id", "allocation", "var_name"])
+    block_id = str(name if name not in {None, ""} else f"allocation-{ordinal}")
+    banks_raw = _first_key(node, ["banks", "bank_ids", "bank_id", "bank", "physical_banks"])
+    groups_raw = _first_key(node, ["groups", "group_ids", "group_id", "bank_group"])
+    temporary_raw = _first_key(node, ["temporary", "is_temporary", "is_temp", "temp"])
+    temporary = None if temporary_raw is None else str(temporary_raw).lower() in {"1", "true", "yes", "y", ("tem"
+        "porary"), "temp"}
+    return {
+        "id": f"mem-{ordinal}",
+        "name": block_id,
+        "start": start,
+        "end": end,
+        "lifetime": max(0.0, end - start),
+        "size_bytes": size,
+        "address_start": addr_start,
+        "address_end": addr_end,
+        "allocation_location": str(_first_key(node, ["allocation_location", "apply_position", "alloc_position", ("s"
+            "ource"), "source_location", "ir_line"]) or ""),
+        "scope": str(_first_key(node, ["scope", "scope_type", "memory_scope"]) or ""),
+        "temporary": temporary,
+        "bank_ids": sorted(set(_onchip_int_list(banks_raw))),
+        "group_ids": sorted(set(_onchip_int_list(groups_raw))),
+        "_signature": (block_id, start, end, addr_start, size),
+    }
+
+
+def _onchip_candidates(raw: Any) -> List[Dict[str, Any]]:
+    candidates: List[Dict[str, Any]] = []
+    seen = set()
+    for node in _walk_json_nodes(raw):
+        candidate = _onchip_candidate(node, len(candidates))
+        if candidate is None:
+            continue
+        signature = candidate.pop("_signature")
+        if signature in seen:
+            continue
+        seen.add(signature)
+        candidates.append(candidate)
+    return candidates
+
+
+def _onchip_bank_layout(raw: Any, candidates: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
+    bank_count = group_count = banks_per_group = None
+    for node in _walk_json_nodes(raw):
+        bank_count = bank_count or to_float(_first_key(node, ["bank_count", "num_banks", "banks_count"]))
+        group_count = group_count or to_float(_first_key(node, ["group_count", "num_groups", "bank_group_count"]))
+        banks_per_group = banks_per_group or to_float(_first_key(node, ["banks_per_group", "bank_num_per_group"]))
+    max_bank = max((max(x["bank_ids"]) for x in candidates if x["bank_ids"]), default=-1)
+    if bank_count is None and max_bank >= 0:
+        bank_count = max_bank + 1
+    if group_count is None and candidates:
+        max_group = max((max(x["group_ids"]) for x in candidates if x["group_ids"]), default=-1)
+        if max_group >= 0:
+            group_count = max_group + 1
+    if banks_per_group is None and bank_count and group_count:
+        banks_per_group = math.ceil(bank_count / group_count)
+
+    bank_cells: List[Dict[str, Any]] = []
+    if bank_count:
+        for bank in range(int(bank_count)):
+            group = int(bank // banks_per_group) if banks_per_group else None
+            owners = [x["id"] for x in candidates if bank in x["bank_ids"]]
+            bank_cells.append({"bank_id": bank, "group_id": group, "owner_ids": owners})
+    return {"bank_count": int(bank_count) if bank_count else 0,
+        "group_count": int(group_count) if group_count else 0, ("b"
+        "anks_per_group"): int(banks_per_group) if banks_per_group else 0, "cells": bank_cells}
+
+
+def parse_onchip_memory(path: Optional[Path]) -> Dict[str, Any]:
+    """Parse a memory_info.json sidecar into the three-view On-Chip Memory contract."""
+    if path is None or not path.is_file():
+        return _onchip_missing_diagnostic()
+    try:
+        raw = json.loads(path.read_text(encoding="utf-8"))
+    except Exception as exc:
+        return {"available": False, "diagnostic": {"code": "memory_info_invalid_json", "title": ("On-Chip "
+            "Memory payload is invalid"), "explanation": str(exc), "evidence": [path.name]}}
+
+    errors = _onchip_errors(raw)
+    candidates = _onchip_candidates(raw)
+    if not candidates:
+        return {
+            "available": False,
+            "diagnostic": {
+                "code": "memory_info_no_usable_allocations",
+                "title": "On-Chip Memory payload has no usable allocation records",
+                "explanation": ("memory_info.json was found, but no records contained a lifetime plus "
+                    "size/address information required by the memory block map."),
+                "evidence": errors or [path.name],
+            },
+        }
+    start_min = min(x["start"] for x in candidates)
+    end_max = max(x["end"] for x in candidates)
+    known_addr: List[float] = []
+    for candidate in candidates:
+        for addr_value in (candidate.get("address_start"), candidate.get("address_end")):
+            if addr_value is not None:
+                known_addr.append(addr_value)
+    return {
+        "available": True,
+        "schema": "msopprof-onchip-memory/v1",
+        "source_name": path.name,
+        "status": "partial" if errors else "ok",
+        "errors": errors,
+        "blocks": candidates,
+        "execution_range": [start_min, end_max],
+        "address_range": [min(known_addr), max(known_addr)] if known_addr else None,
+        "bank_layout": _onchip_bank_layout(raw, candidates),
+        "summary": {"allocation_count": len(candidates),
+            "peak_known_bytes": max((x.get("size_bytes") or 0 for x in candidates), default=0), ("t"
+            "emporary_count"): sum(x.get("temporary") is True for x in candidates)},
+    }
+
+
+def recursive_find_key_values(obj: Any, key_name: str) -> List[Any]:
+    out: List[Any] = []
+    if isinstance(obj, Mapping):
+        for key, value in obj.items():
+            if str(key) == key_name:
+                out.append(value)
+            out.extend(recursive_find_key_values(value, key_name))
+    elif isinstance(obj, list):
+        for value in obj:
+            out.extend(recursive_find_key_values(value, key_name))
+    return out
+
+
+X_KEY_PATTERNS = ["arithmetic_intensity", "arithmetic intensity", "intensity", "ai", "x"]
+
+
+Y_KEY_PATTERNS = ["performance", "perf", "throughput", "flops", "gflops", "tflops", "y"]
+
+
+def first_matching_numeric(mapping: Mapping[str, Any],
+    patterns: Sequence[str]) -> Tuple[Optional[str], Optional[float]]:
+    for key, value in mapping.items():
+        low = str(key).lower()
+        if any(low == p or p in low for p in patterns):
+            x = to_float(value)
+            if x is not None:
+                return str(key), x
+    return None, None
+
+
+def _point_label(obj: Mapping[str, Any], path: str) -> str:
+    label = path or ("P"
+        "oint")
+    for key in obj:
+        if str(key).lower() in {"name", "label", "series", "type"}:
+            label = str(obj[key])
+            break
+    return label
+
+
+def collect_point_candidates(obj: Any, path: str = "") -> List[Dict[str, Any]]:
+    points: List[Dict[str, Any]] = []
+    if isinstance(obj, Mapping):
+        xk, x = first_matching_numeric(obj, X_KEY_PATTERNS)
+        yk, y = first_matching_numeric(obj, Y_KEY_PATTERNS)
+        if x is not None and y is not None and xk != yk:
+            points.append({"x": x, "y": y, "label": _point_label(obj, path)})
+        for key, value in obj.items():
+            points.extend(collect_point_candidates(value, f"{path}.{key}" if path else str(key)))
+    elif isinstance(obj, list):
+        for i, value in enumerate(obj):
+            points.extend(collect_point_candidates(value, f"{path}[{i}]"))
+    return points
+
+
+def _advice_item_lines(item: Any) -> List[str]:
+    cleaned_lines: List[str] = []
+    for line in str(item).splitlines():
+        cleaned = re.sub(r"^[\s\t\-\d\)\.]+", "", line).strip()
+        if cleaned:
+            cleaned_lines.append(cleaned)
+    return cleaned_lines
+
+
+def _collect_advice_item(found: List[str], item: Any) -> None:
+    for cleaned in _advice_item_lines(item):
+        if cleaned not in found:
+            found.append(cleaned)
+
+
+def collect_top_level_advice(records: Sequence[Mapping[str, Any]]) -> List[str]:
+    found: List[str] = []
+    for record in records:
+        value = record.get("advice")
+        values = value if isinstance(value, list) else [value]
+        for item in values:
+            if item is None:
+                continue
+            _collect_advice_item(found, item)
+    return found
+
+
+def _normalized_ratio_fraction(value: Any, measured: Optional[float], limit: Optional[float]) -> Optional[float]:
+    """Return a ratio as a fraction while preserving profiler variability.
+
+    Native profiler versions may emit either a fraction (0.14) or percentage
+    points (14.0). When the native value is absent, derive the ratio only from
+    the canonical measured point and its active roofline limit.
+    """
+    ratio = to_float(value)
+    if ratio is not None and ratio >= 0:
+        if ratio > 1.5 and ratio <= 100:
+            ratio /= 100.0
+        return ratio
+    if measured is not None and limit is not None and limit > 0:
+        return measured / limit
+    return None
+
+
+def _roofline_bound(arithmetic_intensity: float, knee_intensity: Optional[float]) -> str:
+    if knee_intensity is None or knee_intensity <= 0:
+        return "unknown"
+    tolerance = 0.02
+    if arithmetic_intensity < knee_intensity * (1 - tolerance):
+        return "bandwidth-bound region"
+    if arithmetic_intensity > knee_intensity * (1 + tolerance):
+        return "compute-bound region"
+    return "ridge region"
+
+
+def _roofline_chart_rank(title: str, source_index: int) -> Tuple[int, int]:
+    """Order architecture-level memory roofs before execution-unit roofs.
+
+    This is a semantic display rule, not an operator-specific title whitelist.
+    Unknown profiler titles retain their source order after known categories.
+    """
+    text = title.casefold()
+    hierarchy_tokens = ("gm", "l2", "dram", "global memory", "memory hierarchy")
+    unit_tokens = ("memory unit", "vector", "cube", "simt", "simd", "tensor")
+    if any(token in text for token in hierarchy_tokens):
+        return 0, source_index
+    if any(token in text for token in unit_tokens):
+        return 1, source_index
+    return 2, source_index
+
+
+def _roofline_curve_metrics(
+    raw: Mapping[str, Any],
+    x: float,
+    y: float,
+) -> Tuple[Optional[float], Optional[float], Optional[float], Optional[float], Optional[float]]:
+    bandwidth = to_float(raw.get("bw"))
+    compute_ceiling = to_float(raw.get("computility"))
+    has_curve = bandwidth is not None and bandwidth > 0 and compute_ceiling is not None and compute_ceiling > 0
+    knee_intensity = compute_ceiling / bandwidth if has_curve else None
+    roofline_limit = min(bandwidth * x, compute_ceiling) if has_curve else None
+    ratio_fraction = _normalized_ratio_fraction(raw.get("ratio"), y, roofline_limit)
+    return bandwidth, compute_ceiling, knee_intensity, roofline_limit, ratio_fraction
+
+
+def _roofline_row_item(
+    raw: Mapping[str, Any],
+    group_index: int,
+    row_index: int,
+    title: str,
+) -> Optional[Tuple[Dict[str, Any], Dict[str, Any]]]:
+    point = raw.get("point")
+    x = y = None
+    if isinstance(point, (list, tuple)) and len(point) >= 2:
+        x = to_float(point[0])
+        y = to_float(point[1])
+    if x is None or y is None:
+        _, x = first_matching_numeric(raw, X_KEY_PATTERNS)
+        _, y = first_matching_numeric(raw, Y_KEY_PATTERNS)
+        if x is None or y is None:
+            return None
+    if x <= 0 or y <= 0:
+        return None
+
+    bandwidth, compute_ceiling, knee_intensity, roofline_limit, ratio_fraction = _roofline_curve_metrics(raw, x, y)
+    bandwidth_name = str(raw.get("bw_name") or "Bandwidth")
+    compute_name = str(raw.get("computility_name") or "Compute")
+    series_id = f"{group_index}:{row_index}"
+    item = {
+        "series_id": series_id,
+        "curve_id": f"{bandwidth_name}|{compute_name}|{bandwidth}|{compute_ceiling}",
+        "bandwidth": bandwidth,
+        "bandwidth_name": bandwidth_name,
+        "compute_ceiling": compute_ceiling,
+        "compute_name": compute_name,
+        "point": {"x": x, "y": y},
+        "arithmetic_intensity": x,
+        "measured_performance": y,
+        "knee_intensity": knee_intensity,
+        "roofline_limit": roofline_limit,
+        "ratio": ratio_fraction,
+        "ratio_percent": ratio_fraction * 100 if ratio_fraction is not None else None,
+        "bound_region": _roofline_bound(x, knee_intensity),
+    }
+    flat_point = {
+        "series_id": series_id,
+        "chart_title": title,
+        "x": x,
+        "y": y,
+        "label": f"{title} · {bandwidth_name} · {compute_name}",
+    }
+    for shared_key in ("bandwidth", "compute_ceiling", "knee_intensity", "roofline_limit",
+                       "ratio", "ratio_percent", "bound_region"):
+        flat_point[shared_key] = item.get(shared_key)
+    return item, flat_point
+
+
+def _roofline_group_rows(
+    group: Mapping[str, Any],
+    group_index: int,
+    title: str,
+) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
+    rows: List[Dict[str, Any]] = []
+    flat_points: List[Dict[str, Any]] = []
+    for row_index, raw in enumerate(group.get("rooflines") or []):
+        if not isinstance(raw, Mapping):
+            continue
+        parsed = _roofline_row_item(raw, group_index, row_index, title)
+        if parsed is None:
+            continue
+        item, flat_point = parsed
+        rows.append(item)
+        flat_points.append(flat_point)
+    return rows, flat_points
+
+
+def _roofline_native_charts(
+    records: Sequence[Mapping[str, Any]],
+) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
+    charts: List[Dict[str, Any]] = []
+    flat_points: List[Dict[str, Any]] = []
+    seen_chart_signatures: set[Tuple[Any, ...]] = set()
+
+    for record in records:
+        source = record.get("multiple_rooflines")
+        if not isinstance(source, list):
+            continue
+        for group_index, group in enumerate(source):
+            if not isinstance(group, Mapping):
+                continue
+            title = str(group.get("title") or f"Roofline {group_index + 1}")
+            rows, group_flat_points = _roofline_group_rows(group, group_index, title)
+            flat_points.extend(group_flat_points)
+            if not rows:
+                continue
+            signature = (
+                title,
+                tuple((r["bandwidth_name"], r["compute_name"], r["point"]["x"], r["point"]["y"]) for r in rows),
+            )
+            if signature in seen_chart_signatures:
+                continue
+            seen_chart_signatures.add(signature)
+            charts.append({
+                "title": title,
+                "source_index": group_index,
+                "series": rows,
+                "bandwidth_names": list(dict.fromkeys(r["bandwidth_name"] for r in rows)),
+                "compute_names": list(dict.fromkeys(r["compute_name"] for r in rows)),
+                "axis": {
+                    "x_label": "Arithmetic Intensity",
+                    "x_unit": "Ops/Byte",
+                    "y_label": "Performance",
+                    "y_unit": "TOPS/s",
+                    "scale": "logarithmic",
+                },
+            })
+    return charts, flat_points
+
+
+def _roofline_fallback_charts(
+    records: Sequence[Mapping[str, Any]],
+) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
+    items: List[Any] = []
+    for record in records:
+        if isinstance(record.get("multiple_rooflines"), list):
+            items.extend(record["multiple_rooflines"])
+    points = collect_point_candidates(items)
+    seen = set()
+    unique_points = []
+    for point in points:
+        key = (round(point["x"], 12), round(point["y"], 12), point["label"])
+        if key not in seen and point["x"] > 0 and point["y"] > 0:
+            seen.add(key)
+            unique_points.append(point)
+    if not unique_points:
+        return [], unique_points
+    series: List[Dict[str, Any]] = []
+    for index, point in enumerate(unique_points):
+        series.append({
+            "series_id": f"fallback:{index}",
+            "curve_id": f"fallback:{index}",
+            "bandwidth": None,
+            "bandwidth_name": "Measured point",
+            "compute_ceiling": None,
+            "compute_name": str(point.get("label") or "Point"),
+            "point": {"x": point["x"], "y": point["y"]},
+            "arithmetic_intensity": point["x"],
+            "measured_performance": point["y"],
+            "knee_intensity": None,
+            "roofline_limit": None,
+            "ratio": None,
+            "ratio_percent": None,
+            "bound_region": "unknown",
+        })
+    charts = [{
+        "title": "Roofline",
+        "source_index": 0,
+        "series": series,
+        "bandwidth_names": ["Measured point"],
+        "compute_names": list(dict.fromkeys(str(point.get("label") or "Point") for point in unique_points)),
+        "axis": {
+            "x_label": "Arithmetic Intensity",
+            "x_unit": "Ops/Byte",
+            "y_label": "Performance",
+            "y_unit": "TOPS/s",
+            "scale": "logarithmic",
+        },
+    }]
+    return charts, unique_points
+
+
+def parse_roofline(records: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
+    """Normalize native and fallback Roofline payloads into an interactive model.
+
+    The model deliberately describes semantics rather than any specific
+    operator: every native chart, bandwidth source, compute ceiling and point
+    becomes a self-contained series that can be rendered alone or composed
+    with other report modules.
+    """
+    charts, flat_points = _roofline_native_charts(records)
+
+    # Backward-compatible fallback for earlier x/y-shaped payloads. It keeps
+    # measured coordinates visible without inventing bandwidths or ceilings.
+    if not charts:
+        charts, flat_points = _roofline_fallback_charts(records)
+
+    charts.sort(key=lambda chart: _roofline_chart_rank(str(chart.get("title", "")), int(chart.get("source_index", 0))))
+    chart_order = {str(chart.get("title")): i for i, chart in enumerate(charts)}
+    flat_points.sort(key=lambda point: (chart_order.get(str(point.get("chart_title",
+        str(point.get("label", "")).split((" "
+        "· "), 1)[0])), 999), str(point.get("label", ""))))
+
+    advice = collect_top_level_advice(records)
+    return {
+        "available": bool(charts or advice),
+        "plot_available": bool(charts),
+        "charts": charts,
+        "points": flat_points[:400],
+        "advice": advice,
+        "coordinate_contract": "point=[arithmetic_intensity_ops_per_byte, measured_performance_tops_per_s]",
+        "interaction_contract": {
+            "continuous_crosshair": True,
+            "axis_coordinate_readout": True,
+            "point_tooltip": True,
+            "curve_tooltip": True,
+            "curve_hover_emphasis": True,
+        },
+    }
+
+
+class _TraceSink(NamedTuple):
+    lanes: Dict[Tuple[str, str], List[Dict[str, Any]]]
+    stacks: Dict[Tuple[str, str], List[Dict[str, Any]]]
+    starts: List[float]
+    ends: List[float]
+
+
+class _TraceEvent(NamedTuple):
+    name: str
+    category: str
+    start: float
+    duration: float
+    args: Any
+
+
+def _trace_add_event(sink: _TraceSink, lane: Tuple[str, str], event: _TraceEvent) -> None:
+    if event.duration < 0:
+        return
+    item = {
+        "name": event.name,
+        "category": event.category,
+        "start": event.start,
+        "duration": event.duration,
+        "args": event.args if isinstance(event.args, Mapping) else {},
+    }
+    sink.lanes[lane].append(item)
+    sink.starts.append(event.start)
+    sink.ends.append(event.start + event.duration)
+
+
+def _trace_consume(events: Sequence[Any], sink: _TraceSink) -> None:
+    for raw in events:
+        if not isinstance(raw, Mapping):
+            continue
+        ph = str(raw.get("ph", ""))
+        pid = str(raw.get("pid", "process"))
+        tid = str(raw.get("tid", "lane"))
+        ts = to_float(raw.get("ts"))
+        if ph == "X":
+            dur = to_float(raw.get("dur"))
+            if ts is None or dur is None:
+                continue
+            _trace_add_event(sink, (pid, tid), _TraceEvent(str(raw.get("name", "event")),
+                str(raw.get("cat", raw.get("cname", ""))), ts, dur, raw.get(("a"
+                "rgs"), {})))
+        elif ph == "B" and ts is not None:
+            sink.stacks[(pid, tid)].append({
+                "name": str(raw.get("name", "event")),
+                "category": str(raw.get("cat", raw.get("cname", ""))),
+                "start": ts,
+                "args": raw.get("args", {}),
+            })
+        elif ph == "E" and ts is not None and sink.stacks[(pid, tid)]:
+            begin = sink.stacks[(pid, tid)].pop()
+            _trace_add_event(sink, (pid, tid), _TraceEvent(begin["name"], begin["category"],
+                begin["start"], ts - begin["start"], begin["args"]))
+
+
+def _trace_out_lanes(
+    lanes: Dict[Tuple[str, str], List[Dict[str, Any]]],
+    max_events: int,
+) -> Tuple[List[Dict[str, Any]], int]:
+    out_lanes: List[Dict[str, Any]] = []
+    embedded_count = 0
+    for (pid, tid), items in sorted(lanes.items(), key=lambda kv: (kv[0][0], kv[0][1])):
+        items.sort(key=lambda x: (x["start"], x["duration"]))
+        remaining = max_events - embedded_count
+        if remaining <= 0:
+            break
+        kept = items[:remaining]
+        embedded_count += len(kept)
+        out_lanes.append({"pid": pid, "tid": tid, "label": f"{pid} / {tid}", "events": kept})
+    return out_lanes, embedded_count
+
+
+def _trace_timeline_kind(lanes: Mapping[Tuple[str, str], List[Dict[str, Any]]]) -> Tuple[str, str]:
+    lane_tokens = " ".join(f"{pid} {tid}" for pid, tid in lanes).upper()
+    event_tokens = " ".join(
+        str(item.get("name", ""))
+        for items in lanes.values()
+        for item in items[:20]
+    ).upper()
+    combined_tokens = f"{lane_tokens} {event_tokens}"
+    communication_markers = ("HCCL", "AI CPU", "AIC BLOCK", "AIV BLOCK", "ASCENDC API", "TURN")
+    pipe_markers = ("SCALAR", "VECTOR", "MTE1", "MTE2", "MTE3", "CUBE", "FIXPIPE")
+    if any(token in combined_tokens for token in communication_markers):
+        return "communication", "Communication-Compute Timeline"
+    if any(token in combined_tokens for token in pipe_markers):
+        return "pipe", "Pipe Timeline"
+    return "generic", "Timeline"
+
+
+def parse_trace_events(events: Sequence[Any], max_events: int = 10000) -> Dict[str, Any]:
+    sink = _TraceSink(defaultdict(list), defaultdict(list), [], [])
+    _trace_consume(events, sink)
+    lanes, starts, ends = sink.lanes, sink.starts, sink.ends
+
+    if not starts:
+        return {"available": False, "lanes": []}
+    t0, t1 = min(starts), max(ends)
+    out_lanes, embedded_count = _trace_out_lanes(lanes, max_events)
+    timeline_type, timeline_title = _trace_timeline_kind(lanes)
+    return {
+        "available": True,
+        "timeline_type": timeline_type,
+        "title": timeline_title,
+        "start": t0,
+        "end": t1,
+        "duration": t1 - t0,
+        "event_count": sum(len(v) for v in lanes.values()),
+        "embedded_event_count": embedded_count,
+        "lanes": out_lanes,
+    }
+
+
+def load_trace_from_path(path: Optional[Path], max_events: int) -> Optional[Dict[str, Any]]:
+    if not path or not path.is_file() or path.stat().st_size == 0:
+        return None
+    try:
+        # The trace can be tens of MB while only ~max_events are embedded, but
+        # event_count, B/E stack pairing and the global t0/t1 range all require
+        # the full event list, so a partial/streaming parse would not be exact.
+        raw = json.loads(path.read_text(encoding="utf-8"))
+    except Exception as exc:
+        logger.debug("Failed to load trace from %s: %s", path, exc)
+        return None
+    events = raw.get("traceEvents", raw if isinstance(raw, list) else []) if isinstance(raw, (dict, list)) else []
+    if not isinstance(events, list):
+        return None
+    return parse_trace_events(events, max_events=max_events)
+
+
+def parse_timeline(artifacts: ArtifactSet, max_events: int) -> Dict[str, Any]:
+    parsed = load_trace_from_path(artifacts.trace, max_events)
+    if parsed and parsed.get("available"):
+        return parsed
+    records = load_visualize_records(artifacts.visualize_data)
+    for record in records:
+        if isinstance(record.get("traceEvents"), list):
+            return parse_trace_events(record["traceEvents"], max_events=max_events)
+    return {"available": False, "lanes": []}