[FEAT] Microbenchmark add visualization

benchmarks/microbenchmarks/README.md

@@ -75,3 +75,64 @@ python compare_results.py baseline.csv candidate.csv --bench-name GEMM
 
 The script auto-detects metric columns, computes speedups for overlapping rows,
 and reports rows that exist only in the baseline or only in the candidate.
+
+## Visualizing results
+
+`visualize.py` turns the produced CSVs into interactive
+[Plotly](https://plotly.com/python/) charts and writes a single, self-contained
+HTML file (no server needed — convenient over SSH or as a PR attachment).
+`dashboard.py` is a live [Panel](https://panel.holoviz.org/) companion for
+ad-hoc exploration. Install the extra dependencies first:
+
+```bash
+pip install -r requirements-viz.txt
+```
+
+Both consume the aggregate (`--csv`) and per-sample (`--csv-samples`) CSVs and
+auto-detect which format they were given. Parameter columns are detected
+generically, so every benchmark in this directory is supported (dense GEMM,
+FP8 GEMM, grouped GEMM, normalization, casting). When a benchmark has a
+dimension the default axes don't show — e.g. grouped GEMM's expert count `B` —
+the tool prints a `[note]` that medians are pooling across it; add `--color`,
+`--facet`, or `--pass` (or use a dashboard filter) to separate the series.
+
+### Static HTML report
+
+```bash
+# All applicable views for one CSV -> benchmark_gemm_samples.html
+python visualize.py benchmark_gemm_samples.csv
+
+# A single view, choosing the metric
+python visualize.py benchmark_gemm_samples.csv --kind scaling --value throughput
+
+# Baseline vs candidate speedup (visual complement to compare_results.py)
+python visualize.py bench_gemm_candidate.csv --baseline benchmark_gemm_samples.csv
+```
+
+Plot kinds (`--kind`, default `report` = all applicable):
+
+- `distribution`: box plus every raw sample point per group, with percentile
+  trimming (`--trim-upper` / `--trim-lower`). The honest distribution view for
+  the suite's small (~12) sample counts, where a violin/KDE would over-smooth.
+  Requires a samples CSV.
+- `scaling`: median throughput (or time) vs token count `M` per case, with a
+  shaded min–max band.
+- `bars`: grouped median-throughput bars per case with IQR error bars.
+- `comparison`: baseline-vs-candidate speedup bars, one per benchmark group
+  (needs `--baseline`).
+
+Axes default sensibly per kind and can be overridden with `--x`, `--color`,
+`--facet`, `--pass`, and `--value`. Pass `--cdn` to load plotly.js from a CDN
+instead of inlining it (much smaller file, needs internet to open).
+
+### Interactive dashboard
+
+```bash
+panel serve dashboard.py --show --args benchmark_gemm_samples.csv
+```
+
+The CSV path may also be set via the `BENCH_CSV` environment variable. The
+sidebar exposes the plot kind, metric, independent variable, hue, facet, a
+percentile-trim slider, and a per-attribute filter for every parameter column.
+Figure builders are shared with `visualize.py`, so the static and interactive
+views stay in sync.

benchmarks/microbenchmarks/dashboard.py

@@ -0,0 +1,144 @@
+#!/usr/bin/env python
+###############################################################################
+# Copyright (c) 2026, Advanced Micro Devices, Inc. All rights reserved.
+#
+# See LICENSE for license information.
+###############################################################################
+"""Interactive Panel dashboard for microbenchmark CSVs.
+
+A live-exploration companion to ``visualize.py`` (which emits static,
+self-contained HTML). It mirrors the JAX fused-attention dashboard
+(``benchmarks/attention/panel_app.py``): pick the independent variable, the
+hue/series, the metric, per-attribute filters, and a percentile trim, and the
+Plotly figure re-renders. Figure builders are shared with ``visualize.py`` so
+the static and interactive views never diverge.
+
+Usage
+-----
+    panel serve dashboard.py --show --args benchmark_gemm_samples.csv
+
+The CSV path may also be given via the ``BENCH_CSV`` environment variable. If
+neither is provided, the first ``*_samples.csv`` in the working directory is
+used.
+"""
+
+import glob
+import os
+import sys
+
+import panel as pn
+
+import viz_data as vd
+import visualize as viz
+
+pn.extension("plotly", design="material", sizing_mode="stretch_width")
+
+
+def _resolve_csv_path():
+    """Find the CSV to load from --args, $BENCH_CSV, or the cwd."""
+    for arg in sys.argv[1:]:
+        if arg.endswith(".csv"):
+            return arg
+    if os.environ.get("BENCH_CSV"):
+        return os.environ["BENCH_CSV"]
+    candidates = sorted(glob.glob("*_samples.csv")) or sorted(glob.glob("*.csv"))
+    if not candidates:
+        raise SystemExit(
+            "No CSV given. Pass one via '--args FILE' or set BENCH_CSV."
+        )
+    return candidates[0]
+
+
+CSV_PATH = _resolve_csv_path()
+DF = vd.load_any(CSV_PATH)
+IS_SAMPLES = DF.attrs["source"] == "samples"
+PARAMS = list(DF.attrs["params"])
+AXIS_COLS = PARAMS + ["pass"]
+
+KINDS = (["distribution", "scaling", "bars"] if IS_SAMPLES else ["scaling", "bars"])
+
+# ---------------------------------------------------------------------------
+# Controls
+# ---------------------------------------------------------------------------
+kind_w = pn.widgets.Select(name="Plot", options=KINDS, value=KINDS[0])
+value_w = pn.widgets.Select(
+    name="Metric", options=["auto", "time_ms", "throughput"], value="auto",
+)
+x_w = pn.widgets.Select(
+    name="Independent variable (x)", options=AXIS_COLS,
+    value="M" if "M" in AXIS_COLS else AXIS_COLS[0],
+)
+color_w = pn.widgets.Select(
+    name="Hue / series", options=AXIS_COLS,
+    value="Case" if "Case" in AXIS_COLS else AXIS_COLS[-1],
+)
+facet_w = pn.widgets.Select(
+    name="Facet", options=["none"] + AXIS_COLS,
+    value="pass" if "pass" in AXIS_COLS else "none",
+)
+trim_w = pn.widgets.FloatSlider(
+    name="Percentile trim (upper)", start=0.5, end=1.0, step=0.01, value=0.95,
+    disabled=not IS_SAMPLES,
+)
+
+# One filter per parameter column (+ pass). Empty selection means "all".
+filter_ws = {
+    col: pn.widgets.MultiChoice(
+        name=col, options=sorted(DF[col].astype(str).unique(), key=str), value=[],
+    )
+    for col in AXIS_COLS
+}
+
+
+def _resolve_value(kind, value):
+    if value != "auto":
+        return value
+    if kind == "distribution":
+        return "time_ms"
+    return vd.default_value_column(DF, "Forward")
+
+
+def _apply_filters(df):
+    for col, widget in filter_ws.items():
+        if widget.value:
+            df = df[df[col].astype(str).isin(widget.value)]
+    return df
+
+
+@pn.depends(
+    kind=kind_w, value=value_w, x=x_w, color=color_w, facet=facet_w, trim=trim_w,
+    **{f"f_{c}": w for c, w in filter_ws.items()},
+)
+def make_plot(kind, value, x, color, facet, trim, **_filters):
+    df = _apply_filters(DF)
+    if df.empty:
+        return pn.pane.Alert("No rows match the current filters.", alert_type="warning")
+
+    value = _resolve_value(kind, value)
+    facet_arg = None if facet == "none" else facet
+    try:
+        if kind == "distribution":
+            fig = viz.fig_distribution(df, x=x, value=value, color=color,
+                                       facet=facet_arg, trim_upper=trim)
+        elif kind == "scaling":
+            fig = viz.fig_scaling(vd.trim_percentile(df, value=value, upper=trim),
+                                  x=x, value=value, color=color, facet=facet_arg)
+        else:
+            fig = viz.fig_throughput_bars(vd.trim_percentile(df, value=value, upper=trim),
+                                          x=x, value=value, color=color, facet=facet_arg)
+    except Exception as exc:  # surface builder errors in the UI instead of 500s
+        return pn.pane.Alert(f"Could not build plot: {exc}", alert_type="danger")
+    return pn.pane.Plotly(fig, sizing_mode="stretch_width", height=640)
+
+
+template = pn.template.BootstrapTemplate(
+    title=f"Microbenchmark Explorer — {os.path.basename(CSV_PATH)}",
+    sidebar=[
+        pn.pane.Markdown(f"**Source:** `{CSV_PATH}`  \n**Format:** {DF.attrs['source']}"),
+        kind_w, value_w, x_w, color_w, facet_w, trim_w,
+        pn.pane.Markdown("### Filters (empty = all)"),
+        *filter_ws.values(),
+    ],
+)
+template.main.append(pn.Column(make_plot))
+template.servable()

benchmarks/microbenchmarks/requirements-viz.txt

@@ -0,0 +1,5 @@
+# Dependencies for the visualization tooling (visualize.py, dashboard.py).
+# Not required to run the benchmarks themselves.
+pandas
+plotly
+panel