CUDA animation from Nsight Data

cuda_anim_nsight.py

# this needs a sqlite dump from nsight nsys utility
# for example on windows i ran this:
# "C:\Program Files\NVIDIA Corporation\Nsight Systems 2025.1.3\target-windows-x64\nsys.exe" profile --trace=cuda,nvtx --export sqlite --output profile_matmul you_cuda_program.exe
# once that dump is there run the follow python code on it basically it takes few key tables and animate a timeline

import sqlite3
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.patches import Patch
from mpl_toolkits.mplot3d import Axes3D
from matplotlib.animation import FuncAnimation

# --- CONFIG ---
db_path = "profile_matmul.sqlite"  # Adjust if needed
bin_width = 0.5
time_start = 480
time_end = 580
time_bins = np.arange(time_start, time_end, bin_width)

# 💡 Neon palette (adjusted for dark background)
yticks = {
    "kernel": 0,
    "memcpy": 1,
    "sync": 2,
    "runtime": 3,
    "driver": 4,
    "overhead": 5,
}
colors = {
    "kernel": "#00ffff",    # neon cyan
    "memcpy": "#39ff14",    # neon green
    "sync":   "#ff007f",    # hot pink
    "runtime": "#fffb00",   # neon yellow
    "driver": "#9400ff",    # neon purple
    "overhead": "#ff4c00",  # neon orange
}

conn = sqlite3.connect(db_path)
cursor = conn.cursor()

# --- Collect Events ---
events = []

def collect_tiles(label, sql, scale_fn):
    try:
        rows = cursor.execute(sql).fetchall()
    except sqlite3.OperationalError:
        return
    for row in rows:
        start, end = row[:2]
        z = scale_fn(row)
        for t in time_bins:
            if start / 1e6 <= t <= end / 1e6:
                events.append((t, yticks[label], z, colors[label]))

collect_tiles("kernel",   "SELECT start, end FROM CUPTI_ACTIVITY_KIND_KERNEL", lambda r: 1.0)
collect_tiles("memcpy",   "SELECT start, end, bytes FROM CUPTI_ACTIVITY_KIND_MEMCPY", lambda r: max(0.3, np.log2(r[2]) / 10))
collect_tiles("sync",     "SELECT start, end FROM CUPTI_ACTIVITY_KIND_SYNCHRONIZATION", lambda r: 0.6)
collect_tiles("runtime",  "SELECT start, end FROM CUPTI_ACTIVITY_KIND_RUNTIME", lambda r: 0.6)
collect_tiles("driver",   "SELECT start, end FROM CUPTI_ACTIVITY_KIND_DRIVER", lambda r: 0.6)
collect_tiles("overhead", "SELECT start, end FROM PROFILER_OVERHEAD", lambda r: 0.5)

events.sort(key=lambda e: e[0])  # Sort by time

# --- Setup Matplotlib 3D ---
fig = plt.figure(figsize=(14, 7))
ax = fig.add_subplot(111, projection='3d')
fig.patch.set_facecolor("black")
ax.set_facecolor("black")

# Clean background
ax.xaxis.pane.set_visible(False)
ax.yaxis.pane.set_visible(False)
ax.zaxis.pane.set_visible(False)

for axis in [ax.xaxis, ax.yaxis, ax.zaxis]:
    axis._axinfo["grid"]["linewidth"] = 0
    axis._axinfo["tick"]["in"] = False
    axis._axinfo["tick"]["out"] = False
    axis._axinfo["tick"]["color"] = (1, 1, 1, 0)
    axis._axinfo["axisline"]["color"] = (1, 1, 1, 0)

ax.set_xlim(time_start, time_end)
ax.set_ylim(-1, len(yticks))
ax.set_zlim(0, 5)
ax.set_xticks([])
ax.set_yticks([])
ax.set_zticks([])
ax.grid(False)

# Title (neon style)
ax.set_title("🔮 GPU Timeline — Neon Voxel Animation", fontsize=14, color="white")

# --- Legend ---
legend_elements = [
    Patch(facecolor=colors[name], edgecolor='white', label=name)
    for name in yticks.keys()
]
ax.legend(
    handles=legend_elements,
    loc='upper left',
    bbox_to_anchor=(1.05, 1),
    fontsize=9,
    title="Activity Type",
    title_fontsize=10,
    frameon=False
)

# --- Animation Logic ---
bars = []

def animate(frame):
    if frame >= len(events):
        return
    x, y, z, color = events[frame]
    bar = ax.bar3d(
        x, y, 0,
        bin_width * 0.95, 0.6, z,
        color=color,
        edgecolor="white",
        linewidth=0.3,
        alpha=0.9
    )
    bars.append(bar)

# --- Animate ---
anim = FuncAnimation(fig, animate, frames=len(events), interval=10, repeat=False)

plt.tight_layout()
plt.show()