utils.py

import sys

import gradio as gr
import numpy as np
import pandas as pd
import plotly.express as px
import plotly.graph_objects as go
import re

from config.constants import COLUMN_MAPPINGS, COLUMN_ORDER, TYPE_EMOJI, DISCARDED_MODELS


def model_hyperlink(link, model_name, release, thinking=False):
    ret = f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
    new_badge = f' <span class="badge new-badge">new</span>'
    reasoning_badge = f' <span class="badge reasoning-badge">reasoning</span>'

    if release == "V4":
        # show new badge only to the latest releases
        return ret + reasoning_badge + new_badge if thinking == "Reasoning" else ret + new_badge
    else:
        return ret + reasoning_badge if thinking == "Reasoning" else ret


def extract_name_from_link(html: str) -> str:
    """
    Extracts the model name from the HTML generated by model_hyperlink()
    """
    if not isinstance(html, str):
        return html

    match = re.search(r'<a[^>]*>(.*?)</a>', html)
    if match:
        return match.group(1).strip()

    return re.sub(r'<[^>]+>', '', html).strip()


def handle_special_cases(benchmark, metric):
    if metric == "Exact Matching (EM)":
        benchmark = "RTL-Repo"
    elif benchmark == "RTL-Repo":
        metric = "Exact Matching (EM)"
    return benchmark, metric


def filter_RTLRepo(subset: pd.DataFrame, name=str) -> pd.DataFrame:
    if subset.empty:
        return pd.DataFrame(columns=["Type", "Model", "Params", "Exact Matching (EM)"])

    subset = subset.drop(subset[subset.Score < 0.0].index)

    # Check again if empty after filtering
    if subset.empty:
        return pd.DataFrame(columns=["Type", "Model", "Params", "Exact Matching (EM)"])

    details = subset[["Model", "Model URL", "Model Type", "Params", "Release", "Thinking"]].drop_duplicates(
        "Model"
    )
    filtered_df = subset[["Model", "Score"]].rename(columns={"Score": "Exact Matching (EM)"})

    filtered_df = pd.merge(filtered_df, details, on="Model", how="left")
    filtered_df["Model"] = filtered_df.apply(
        lambda row: model_hyperlink(
            row["Model URL"],
            row["Model"],
            row["Release"],
        ),
        axis=1,
    )
    filtered_df["Type"] = filtered_df["Model Type"].map(lambda x: TYPE_EMOJI.get(x, ""))
    filtered_df = filtered_df[["Type", "Model", "Params", "Exact Matching (EM)"]]
    filtered_df = filtered_df.sort_values(by="Exact Matching (EM)", ascending=False).reset_index(drop=True)

    if name == "Other Models":
        filtered_df["Date Discarded"] = filtered_df["Model"].apply(lambda x: DISCARDED_MODELS.get(extract_name_from_link(x), "N/A"))

        # reorder to put Date Discarded between Params and Exact Matching (EM)
        cols = ["Type", "Model", "Parameters (B)", "Date Discarded", "Exact Matching (EM)"]
        filtered_df = filtered_df[[c for c in cols if c in filtered_df.columns]]
    return filtered_df


def filter_bench(subset: pd.DataFrame, df_agg=None, agg_column=None, name=str) -> pd.DataFrame:
    if subset.empty:
        return pd.DataFrame(columns=COLUMN_ORDER)

    details = subset[["Model", "Model URL", "Model Type", "Params", "Release", "Thinking"]].drop_duplicates(
        "Model"
    )
    if "RTLLM" in subset["Benchmark"].unique():
        pivot_df = (
            subset.pivot_table(index="Model", columns="Metric", values="Score", aggfunc=custom_agg_s2r)
            .reset_index()
            .round(2)
        )
    else:
        pivot_df = (
            subset.pivot_table(index="Model", columns="Metric", values="Score", aggfunc=custom_agg_cc)
            .reset_index()
            .round(2)
        )

    # if df_agg is not None and agg_column is not None and agg_column in df_agg.columns:
    #     agg_data = df_agg[["Model", agg_column]].rename(
    #         columns={agg_column: "Aggregated ⬆️"}
    #     )
    #     pivot_df = pd.merge(pivot_df, agg_data, on="Model", how="left")
    # else:  # fallback
    #     pivot_df["Aggregated ⬆️"] = pivot_df.mean(axis=1, numeric_only=True).round(2)

    

    pivot_df = pd.merge(pivot_df, details, on="Model", how="left")
    pivot_df["Model"] = pivot_df.apply(
        lambda row: model_hyperlink(row["Model URL"], row["Model"], row["Release"], row["Thinking"]),
        axis=1,
    )
    pivot_df["Type"] = pivot_df["Model Type"].map(lambda x: TYPE_EMOJI.get(x, ""))

    if all(col in pivot_df.columns for col in ["Power", "Performance", "Area"]):
        pivot_df["Post-Synthesis (PSQ)"] = pivot_df[["Power", "Performance", "Area"]].mean(axis=1).round(2)

    
    pivot_df.rename(columns=COLUMN_MAPPINGS, inplace=True)
    pivot_df = pivot_df[[col for col in COLUMN_ORDER if col in pivot_df.columns]]

    if "Functionality" in pivot_df.columns:
        pivot_df = pivot_df.sort_values(by="Functionality", ascending=False).reset_index(drop=True)

    if name == "Other Models":
        pivot_df["Date Discarded"] = pivot_df["Model"].apply(lambda x: DISCARDED_MODELS.get(extract_name_from_link(x), "N/A"))
        # reorder to put Date Discarded between Params Syntax
        cols = ["Type", "Model", "Parameters (B)", "Date Discarded", "Syntax", "Functionality", "Synthesis", "Post-Synthesis"]
        pivot_df = pivot_df[[c for c in cols if c in pivot_df.columns]]
    return pivot_df


def custom_agg_s2r(vals):
    if len(vals) == 2:
        s2r_val = vals.iloc[0]
        rtllm_val = vals.iloc[1]
        w1 = 155
        w2 = 47
        result = (w1 * s2r_val + w2 * rtllm_val) / (w1 + w2)
    else:
        result = vals.iloc[0]
    return round(result, 2)


def custom_agg_cc(vals):
    if len(vals) == 2:
        veval_val = vals.iloc[0]
        vgen_val = vals.iloc[1]
        w1 = 155
        w2 = 17
        result = (w1 * veval_val + w2 * vgen_val) / (w1 + w2)
    else:
        result = vals.iloc[0]
    return round(result, 2)


def filter_bench_all(subset: pd.DataFrame, df_agg=None, agg_column=None, name=str) -> pd.DataFrame:
    if subset.empty:
        return pd.DataFrame(columns=COLUMN_ORDER)

    details = subset[["Model", "Model URL", "Model Type", "Params", "Release", "Thinking"]].drop_duplicates(
        "Model"
    )
    if "RTLLM" in subset["Benchmark"].unique():
        pivot_df = (
            subset.pivot_table(index="Model", columns="Metric", values="Score", aggfunc=custom_agg_s2r)
            .reset_index()
            .round(2)
        )
    else:
        pivot_df = (
            subset.pivot_table(index="Model", columns="Metric", values="Score", aggfunc=custom_agg_cc)
            .reset_index()
            .round(2)
        )

    pivot_df = pd.merge(pivot_df, details, on="Model", how="left")
    pivot_df["Model"] = pivot_df.apply(
        lambda row: model_hyperlink(row["Model URL"], row["Model"], row["Release"], row["Thinking"]),
        axis=1,
    )
    pivot_df["Type"] = pivot_df["Model Type"].map(lambda x: TYPE_EMOJI.get(x, ""))

    if all(col in pivot_df.columns for col in ["Power", "Performance", "Area"]):
        pivot_df["Post-Synthesis (PSQ)"] = pivot_df[["Power", "Performance", "Area"]].mean(axis=1).round(2)

    pivot_df.rename(columns=COLUMN_MAPPINGS, inplace=True)
    pivot_df = pivot_df[[col for col in COLUMN_ORDER if col in pivot_df.columns]]

    if "Functionality" in pivot_df.columns:
        pivot_df = pivot_df.sort_values(by="Functionality", ascending=False).reset_index(drop=True)

    if name == "Other Models":
        pivot_df["Date Discarded"] = pivot_df["Model"].apply(lambda x: DISCARDED_MODELS.get(extract_name_from_link(x), "N/A"))
        # reorder to put Date Discarded between Params Syntax
        cols = ["Type", "Model", "Parameters (B)", "Date Discarded", "Syntax", "Functionality", "Synthesis", "Post-Synthesis"]
        pivot_df = pivot_df[[c for c in cols if c in pivot_df.columns]]

    return pivot_df