tribev2 0.0.4

TRIBE v2 — multimodal fMRI brain encoding model inference in Rust
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295

#!/usr/bin/env python3
"""Generate benchmark charts from JSON results.

Charts produced
---------------
bench_latency.png       — all backends, log-scale bar chart
bench_speedup.png       — speedup vs naive Rust CPU
bench_gpu_detail.png    — GPU & accelerated backends, linear scale
bench_optimization.png  — step-by-step wgpu improvement story
bench_table.md          — markdown summary table
"""

import json, os, glob
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
import matplotlib.ticker as ticker
import numpy as np

FIGURES_DIR = os.path.join(os.path.dirname(__file__), "..", "figures")
BENCH_DIR   = os.path.dirname(__file__)


def load_results():
    results = {}
    for path in sorted(glob.glob(os.path.join(BENCH_DIR, "results_*.json"))):
        with open(path) as f:
            results.update(json.load(f))
    return results


# ── Colour palette ──────────────────────────────────────────────────────────
C = {
    "rust_cpu":                     "#BBBBBB",
    "rust_cpu_accelerate":          "#8172B3",
    "rust_burn_ndarray":            "#DD8452",
    "rust_burn_ndarray_accelerate": "#64B5CD",
    "python_cpu_1thread":           "#4C72B0",
    "python_cpu_multithread":       "#6EA6D0",
    "python_mps":                   "#C44E52",
    "rust_burn_wgpu_metal":         "#55A868",
    "rust_burn_wgpu_metal_f16":     "#2CA02C",
    "rust_burn_wgpu_metal_kernels": "#1B7837",
}

# Master display order (key, short label, colour key)
ALL = [
    ("rust_cpu",                     "Rust CPU\n(naive)",          "rust_cpu"),
    ("rust_burn_ndarray",            "Burn\nNdArray",              "rust_burn_ndarray"),
    ("rust_burn_ndarray_accelerate", "Burn NdArray\n+Accelerate",  "rust_burn_ndarray_accelerate"),
    ("rust_cpu_accelerate",          "Rust CPU\nAccelerate",       "rust_cpu_accelerate"),
    ("python_cpu_1thread",           "Python\nCPU",                "python_cpu_1thread"),
    ("rust_burn_wgpu_metal",         "Burn wgpu\nf32",             "rust_burn_wgpu_metal"),
    ("rust_burn_wgpu_metal_f16",     "Burn wgpu\nf16",             "rust_burn_wgpu_metal_f16"),
    ("rust_burn_wgpu_metal_kernels", "Burn wgpu\nf32+kernels",     "rust_burn_wgpu_metal_kernels"),
    ("python_mps",                   "Python\nMPS GPU",            "python_mps"),
]


def _bar(ax, indices, means, stds, colors, labels,
         value_fmt="{:.0f}ms", value_offset_frac=0.04):
    bars = ax.bar(indices, means, yerr=stds if any(s > 0 for s in stds) else None,
                  capsize=4, color=colors, edgecolor="black", linewidth=0.6,
                  error_kw=dict(lw=1, capthick=1))
    top = max(means)
    for bar, m, s in zip(bars, means, stds):
        lbl = value_fmt.format(m)
        ax.text(bar.get_x() + bar.get_width() / 2,
                bar.get_height() + s + top * value_offset_frac,
                lbl, ha="center", va="bottom", fontsize=9, fontweight="bold")
    ax.set_xticks(indices)
    ax.set_xticklabels(labels, fontsize=9)
    return bars


# ── 1. Full latency (log scale) ─────────────────────────────────────────────
def make_latency_chart(r):
    keys = [(k, l, C[ck]) for k, l, ck in ALL if k in r]
    if not keys: return
    idx    = list(range(len(keys)))
    means  = [r[k]["mean_ms"] for k, _, _ in keys]
    stds   = [r[k].get("std_ms", 0) for k, _, _ in keys]
    colors = [c for _, _, c in keys]
    labels = [l for _, l, _ in keys]

    fig, ax = plt.subplots(figsize=(max(9, len(keys) * 1.7), 5.5))
    _bar(ax, idx, means, stds, colors, labels,
         value_fmt="{:.0f}ms", value_offset_frac=0.0)
    ax.set_yscale("log")
    ax.yaxis.set_major_formatter(ticker.FuncFormatter(
        lambda x, _: f"{int(x):,}" if x < 1000 else f"{x/1000:.1f}k"))
    ax.set_ylabel("Latency  (ms, log scale)", fontsize=11)
    ax.set_title(
        "TRIBE v2  Forward Pass Latency — All Backends\n"
        "batch=1 · T=100 · 3 modalities · 20 484 cortical vertices",
        fontsize=12, fontweight="bold")
    ax.grid(axis="y", alpha=0.25)
    ax.set_axisbelow(True)
    fig.tight_layout()
    _save(fig, "bench_latency.png")


# ── 2. Speedup vs naive CPU ─────────────────────────────────────────────────
def make_speedup_chart(r):
    base_key = "rust_cpu"
    if base_key not in r: return
    base   = r[base_key]["mean_ms"]
    keys   = [(k, l, C[ck]) for k, l, ck in ALL if k in r]
    idx    = list(range(len(keys)))
    sp     = [base / r[k]["mean_ms"] for k, _, _ in keys]
    colors = [c for _, _, c in keys]
    labels = [l for _, l, _ in keys]

    fig, ax = plt.subplots(figsize=(max(9, len(keys) * 1.7), 5.5))
    bars = ax.bar(idx, sp, color=colors, edgecolor="black", linewidth=0.6)
    for bar, s in zip(bars, sp):
        ax.text(bar.get_x() + bar.get_width() / 2,
                bar.get_height() + max(sp) * 0.02,
                f"{s:.0f}×", ha="center", va="bottom", fontsize=9, fontweight="bold")
    ax.set_xticks(idx)
    ax.set_xticklabels(labels, fontsize=9)
    ax.set_ylabel("Speedup vs Rust CPU naive", fontsize=11)
    ax.set_title(
        "TRIBE v2  Speedup vs Rust CPU (naive loops)\n"
        "20 484 vertices · 100 timesteps",
        fontsize=12, fontweight="bold")
    ax.grid(axis="y", alpha=0.25)
    ax.set_axisbelow(True)
    fig.tight_layout()
    _save(fig, "bench_speedup.png")


# ── 3. GPU detail (linear scale) ────────────────────────────────────────────
def make_gpu_detail_chart(r):
    order = [
        ("python_cpu_1thread",           "Python\nCPU",          "python_cpu_1thread"),
        ("rust_cpu_accelerate",          "Rust CPU\nAccelerate",  "rust_cpu_accelerate"),
        ("rust_burn_wgpu_metal",         "Burn wgpu\nf32",        "rust_burn_wgpu_metal"),
        ("rust_burn_wgpu_metal_f16",     "Burn wgpu\nf16",        "rust_burn_wgpu_metal_f16"),
        ("rust_burn_wgpu_metal_kernels", "Burn wgpu\nf32+kernels","rust_burn_wgpu_metal_kernels"),
        ("python_mps",                   "Python\nMPS GPU",       "python_mps"),
    ]
    keys = [(k, l, C[ck]) for k, l, ck in order if k in r]
    if len(keys) < 2: return
    idx    = list(range(len(keys)))
    means  = [r[k]["mean_ms"] for k, _, _ in keys]
    stds   = [r[k].get("std_ms", 0) for k, _, _ in keys]
    colors = [c for _, _, c in keys]
    labels = [l for _, l, _ in keys]

    fig, ax = plt.subplots(figsize=(max(7, len(keys) * 1.9), 5.5))
    _bar(ax, idx, means, stds, colors, labels, value_fmt="{:.1f}ms")

    # reference line: Python MPS
    if "python_mps" in r:
        mps = r["python_mps"]["mean_ms"]
        ax.axhline(mps, color=C["python_mps"], linestyle="--", linewidth=1.3, alpha=0.7,
                   label=f"Python MPS  {mps:.1f} ms")
        ax.legend(fontsize=9, loc="upper right")

    ax.set_ylabel("Latency  (ms)", fontsize=11)
    ax.set_title(
        "TRIBE v2  GPU & Accelerated Backends\n"
        "20 484 vertices · 100 timesteps · batch=1",
        fontsize=12, fontweight="bold")
    ax.grid(axis="y", alpha=0.25)
    ax.set_axisbelow(True)
    fig.tight_layout()
    _save(fig, "bench_gpu_detail.png")


# ── 4. Optimisation story (wgpu step-by-step) ───────────────────────────────
def make_optimization_chart(r):
    """Waterfall showing each optimisation applied to the wgpu Metal backend."""
    steps = [
        # (result_key,             label,                             color)
        ("rust_burn_wgpu_metal_baseline", "Original\n(pre-opt)",      "#BBBBBB"),
        # synthesised from history if not present:
        ("rust_burn_wgpu_metal",          "Separate QKV\n+manual attn\n+RoPE cache\n+w_avg_t",
                                                                       C["rust_burn_wgpu_metal"]),
        ("rust_burn_wgpu_metal_f16",      "  + f16\n  dtype",         C["rust_burn_wgpu_metal_f16"]),
        ("rust_burn_wgpu_metal_kernels",  "+ Fused CubeCL\n  kernels", C["rust_burn_wgpu_metal_kernels"]),
        ("python_mps",                    "Python\nMPS",               C["python_mps"]),
    ]

    # Use hard-coded original if not in results
    baseline_ms = 27.6
    known = {}
    for k, l, c in steps:
        if k in r:
            known[k] = r[k]["mean_ms"]
    known.setdefault("rust_burn_wgpu_metal_baseline", baseline_ms)

    present = [(k, l, c) for k, l, c in steps if k in known]
    if len(present) < 3: return

    idx    = list(range(len(present)))
    means  = [known[k] for k, _, _ in present]
    labels = [l for _, l, _ in present]
    colors = [c for _, _, c in present]
    stds   = [r[k].get("std_ms", 0) if k in r else 0 for k, _, _ in present]

    fig, ax = plt.subplots(figsize=(max(8, len(present) * 2.0), 5.5))
    _bar(ax, idx, means, stds, colors, labels, value_fmt="{:.1f}ms", value_offset_frac=0.03)

    # Annotation: delta arrows between bars
    for i in range(1, len(means)):
        delta = means[i] - means[i - 1]
        if abs(delta) > 0.3:
            x0 = idx[i - 1] + 0.4
            x1 = idx[i]     - 0.4
            y  = max(means[i - 1], means[i]) * 1.18
            ax.annotate("", xy=(x1, y), xytext=(x0, y),
                        arrowprops=dict(arrowstyle="->", color="grey", lw=1.2))
            ax.text((x0 + x1) / 2, y * 1.03,
                    f"{delta:+.1f}ms",
                    ha="center", va="bottom", fontsize=8, color="grey")

    ax.set_ylabel("Latency  (ms)", fontsize=11)
    ax.set_title(
        "Burn wgpu Metal — Optimisation Journey\n"
        "Each bar = cumulative improvements applied",
        fontsize=12, fontweight="bold")
    ax.grid(axis="y", alpha=0.25)
    ax.set_axisbelow(True)
    fig.tight_layout()
    _save(fig, "bench_optimization.png")


# ── 5. Markdown table ────────────────────────────────────────────────────────
def make_table(r):
    entries = [
        ("rust_cpu",                     "Rust CPU (naive loops)"),
        ("rust_burn_ndarray",            "Burn NdArray (Rayon)"),
        ("rust_burn_ndarray_accelerate", "Burn NdArray + Accelerate"),
        ("rust_cpu_accelerate",          "Rust CPU + Accelerate BLAS"),
        ("python_cpu_1thread",           "Python CPU (1 thread)"),
        ("rust_burn_wgpu_metal",         "Burn wgpu Metal f32"),
        ("rust_burn_wgpu_metal_f16",     "Burn wgpu Metal f16"),
        ("rust_burn_wgpu_metal_kernels", "Burn wgpu Metal f32 + fused kernels"),
        ("python_mps",                   "Python MPS GPU"),
    ]
    base = r.get("rust_cpu", {}).get("mean_ms", 1.0)
    lines = [
        "# TRIBE v2 Forward Pass Benchmark",
        "",
        "**Setup:** batch=1, T=100, 3 modalities (text/audio/video), 20 484 cortical vertices",
        "",
        "| Backend | Mean (ms) | Min (ms) | Std (ms) | vs CPU naive |",
        "|---------|----------:|---------:|---------:|-------------:|",
    ]
    for k, label in entries:
        if k not in r: continue
        d  = r[k]
        sp = f"{base / d['mean_ms']:.0f}×"
        lines.append(
            f"| {label} | {d['mean_ms']:.1f} | {d['min_ms']:.1f} |"
            f" {d.get('std_ms', 0):.1f} | {sp} |"
        )
    table = "\n".join(lines)
    path  = os.path.join(FIGURES_DIR, "bench_table.md")
    with open(path, "w") as f:
        f.write(table + "\n")
    print(f"Saved  {path}")
    print(table)


# ── helpers ───────────────────────────────────────────────────────────────────
def _save(fig, name):
    os.makedirs(FIGURES_DIR, exist_ok=True)
    path = os.path.join(FIGURES_DIR, name)
    fig.savefig(path, dpi=150, bbox_inches="tight")
    print(f"Saved  {path}")
    plt.close(fig)


def main():
    os.makedirs(FIGURES_DIR, exist_ok=True)
    r = load_results()
    if not r:
        print("No results/*.json found — run benchmarks first.")
        return
    print(f"Loaded {len(r)} result entries\n")

    make_latency_chart(r)
    make_speedup_chart(r)
    make_gpu_detail_chart(r)
    make_optimization_chart(r)
    make_table(r)
    print("\nAll done — figures in figures/")


if __name__ == "__main__":
    main()