seu-3dgs / code /distributed.py
Lightcap's picture
Upload folder using huggingface_hub
f138992 verified
Raw
History Blame Contribute Delete
10.5 kB
"""E9: distributed (sort-first) tile-rendering experiment.
A sort-first parallel rasterizer partitions the screen into node regions; each
node renders the primitives whose projected footprint overlaps its region. Two
quantities follow directly from the real projected geometry that gsplat exposes
(per-primitive projected centre and radius): the communication amplification (how
many nodes each primitive must be sent to) and, under a single-bit upset, the
contamination (how many node regions a corrupted primitive reaches). We measure
both as a function of the node count and with/without the support guard, and we
validate the predicted contaminated regions against an actual per-region render.
"""
import argparse
import json
import math
import os
import time
import numpy as np
import torch
import faultlib as F
import gsmodel
def grid_for(T):
"""near-square (rows, cols) with rows*cols == T."""
r = int(round(math.sqrt(T)))
while T % r:
r -= 1
return r, T // r
def nodes_overlapped(cx, cy, rx, ry, W, H, rows, cols):
"""number of (rows x cols) screen-node regions a centre+radius box overlaps."""
x0 = np.clip(np.floor((cx - rx) / W * cols), 0, cols - 1)
x1 = np.clip(np.floor((cx + rx) / W * cols), 0, cols - 1)
y0 = np.clip(np.floor((cy - ry) / H * rows), 0, rows - 1)
y1 = np.clip(np.floor((cy + ry) / H * rows), 0, rows - 1)
return (x1 - x0 + 1) * (y1 - y0 + 1)
def visible_geometry(params, vm, K, W, H, sh):
_, _, info = gsmodel.render(params, vm, K, W, H, sh)
gid = info["gaussian_ids"].detach().cpu().numpy()
radii = info["radii"].detach().float().cpu().numpy() # [nnz,2]
m2d = info["means2d"].detach().cpu().numpy() # [nnz,2]
return gid, radii, m2d
def run(model_path, out, Ts, S, seed, log):
ck = torch.load(model_path, map_location="cuda", weights_only=False)
params = {k: v.cuda().float() for k, v in ck["params"].items()}
sh, W, H = ck["sh_degree"], ck["W"], ck["H"]
scene = ck["scene"]
bounds = F.compute_bounds(params)
vm = ck["test_viewmats"][:1].cuda()
K = ck["test_Ks"][:1].cuda()
rng = np.random.default_rng(seed)
def lg(*a):
m = " ".join(str(x) for x in a); print(m, flush=True)
open(log, "a").write(m + "\n")
gid, radii, m2d = visible_geometry(params, vm, K, W, H, sh)
rx = radii[:, 0]; ry = radii[:, 1]
cx = m2d[:, 0]; cy = m2d[:, 1]
vis = gid # global ids of visible primitives
lg(f"[{scene}] visible primitives={len(vis)} WxH={W}x{H}")
# sample scale-sign upsets among visible primitives
sample = rng.choice(len(vis), size=min(S, len(vis)), replace=False)
stored, work = F.quantize_params(params, "fp32")
results = {}
for T in Ts:
rows, cols = grid_for(T)
# communication amplification on the clean model
clean_nodes = nodes_overlapped(cx, cy, rx, ry, W, H, rows, cols)
comm_clean = float(clean_nodes.mean())
cont_ng, cont_g, comm_amp = [], [], []
for s in sample:
g = int(vis[s]); comp = 0
flat = g * 3 + comp
cval, _ = F.flip_one(stored["scales"], work["scales"], flat, 31, "fp32")
# corrupted geometry: re-render, read this primitive's projected radius
gid2, radii2, m2d2 = visible_geometry(work, vm, K, W, H, sh)
loc = np.where(gid2 == g)[0]
if len(loc):
i = loc[0]
n_ng = nodes_overlapped(m2d2[i, 0], m2d2[i, 1], radii2[i, 0], radii2[i, 1], W, H, rows, cols)
else:
n_ng = T # off-screen-huge: treat as global
# guarded
gw = F.apply_guard(work, bounds)
gid3, radii3, m2d3 = visible_geometry(gw, vm, K, W, H, sh)
loc3 = np.where(gid3 == g)[0]
if len(loc3):
j = loc3[0]
n_g = nodes_overlapped(m2d3[j, 0], m2d3[j, 1], radii3[j, 0], radii3[j, 1], W, H, rows, cols)
else:
n_g = 1
F.restore_one(work["scales"], flat, cval)
cont_ng.append(float(n_ng)); cont_g.append(float(n_g))
comm_amp.append(float(n_ng)) # extra transmissions for the corrupted primitive
cont_ng = np.array(cont_ng); cont_g = np.array(cont_g)
results[T] = {
"rows": rows, "cols": cols, "comm_clean": comm_clean,
"contam_mean_noguard": float(cont_ng.mean()),
"contam_p99_noguard": float(np.percentile(cont_ng, 99)),
"contam_frac_noguard": float((cont_ng / T).mean()),
"contam_mean_guard": float(cont_g.mean()),
"contam_p99_guard": float(np.percentile(cont_g, 99)),
"contam_frac_guard": float((cont_g / T).mean()),
}
lg(f" T={T:3d} ({rows}x{cols}) comm_clean={comm_clean:.2f} "
f"contam(no guard) mean={cont_ng.mean():.1f}/{T} p99={np.percentile(cont_ng,99):.0f} | "
f"contam(guard) mean={cont_g.mean():.2f}")
# rendering-based validation: predicted contaminated regions vs actually changed regions
val = validate_render(work, stored, bounds, vm, K, W, H, sh, vis, sample[:20], rng)
rt = rank_timing(work, stored, bounds, vm, K, W, H, sh, vis, sample, T=16)
lg(f" rank timing T=16: barrier(max rank) clean={rt['clean']['max_ms']:.2f}ms "
f"corrupt={rt['corrupt']['max_ms']:.2f}ms guard={rt['guard']['max_ms']:.2f}ms; "
f"imbalance corrupt={rt['corrupt']['imbalance']:.2f} guard={rt['guard']['imbalance']:.2f}")
out_obj = {"scene": scene, "Ts": {str(k): v for k, v in results.items()},
"validation": val, "rank_timing": rt}
json.dump(out_obj, open(os.path.join(out, f"distributed_{scene}.json"), "w"), indent=2)
lg(f" validation pred-vs-render IoU={val['mean_iou']:.3f} over {val['n']} cases")
def regions_of_bbox(cx, cy, rx, ry, W, H, rows, cols):
x0 = int(np.clip(np.floor((cx - rx) / W * cols), 0, cols - 1))
x1 = int(np.clip(np.floor((cx + rx) / W * cols), 0, cols - 1))
y0 = int(np.clip(np.floor((cy - ry) / H * rows), 0, rows - 1))
y1 = int(np.clip(np.floor((cy + ry) / H * rows), 0, rows - 1))
return {(r, c) for r in range(y0, y1 + 1) for c in range(x0, x1 + 1)}
def validate_render(work, stored, bounds, vm, K, W, H, sh, vis, sample, rng, T=16):
"""Compare the node regions predicted contaminated (from the corrupted
primitive's projected radius) against the regions whose pixels actually
change when the scene is rendered. Reports mean intersection-over-union."""
rows, cols = grid_for(T)
clean, _ = F.render_views(work, vm, K, W, H, sh)
clean = clean[0]
th, tw = H // rows, W // cols
ious, n = [], 0
for s in sample:
g = int(vis[s]); flat = g * 3
cval, _ = F.flip_one(stored["scales"], work["scales"], flat, 31, "fp32")
corr, _ = F.render_views(work, vm, K, W, H, sh)
gid2, radii2, m2d2 = visible_geometry(work, vm, K, W, H, sh)
F.restore_one(work["scales"], flat, cval)
diff = (corr[0] - clean).abs().amax(-1) > (1 / 255)
actual = set()
for ri in range(rows):
for ci in range(cols):
if bool(diff[ri * th:(ri + 1) * th, ci * tw:(ci + 1) * tw].any()):
actual.add((ri, ci))
loc = np.where(gid2 == g)[0]
if len(loc):
i = loc[0]
pred = regions_of_bbox(m2d2[i, 0], m2d2[i, 1], radii2[i, 0], radii2[i, 1], W, H, rows, cols)
else:
pred = {(r, c) for r in range(rows) for c in range(cols)}
if actual or pred:
inter = len(actual & pred); union = len(actual | pred)
ious.append(inter / max(union, 1))
n += 1
return {"mean_iou": float(np.mean(ious)) if ious else 0.0, "n": n, "T": T}
def rank_timing(work, stored, bounds, vm, K, W, H, sh, vis, sample, T=16):
"""Per-rank (per-tile) render time under sort-first partitioning, for a clean
scene, an unguarded scale-sign explosion, and the guarded version. The slowest
rank sets the barrier-synchronized frame time; the sum is total compute."""
import time
rows, cols = grid_for(T)
th, tw = H // rows, W // cols
def render_tile(params, ri, ci):
Kt = K.clone()
Kt[0, 0, 2] -= ci * tw
Kt[0, 1, 2] -= ri * th
torch.cuda.synchronize(); t = time.time()
for _ in range(3):
gsmodel.render(params, vm, Kt, tw, th, sh)
torch.cuda.synchronize()
return (time.time() - t) / 3 * 1e3 # ms
def all_ranks(params):
ts = [render_tile(params, ri, ci) for ri in range(rows) for ci in range(cols)]
ts = np.array(ts)
return {"max_ms": float(ts.max()), "mean_ms": float(ts.mean()),
"sum_ms": float(ts.sum()), "imbalance": float(ts.max() / ts.mean())}
# pick a large-footprint scale-sign primitive
best = (-1, int(vis[sample[0]]))
clean, _ = F.render_views(work, vm, K, W, H, sh)
for s in sample[:40]:
g = int(vis[s]); flat = g * 3
cv, _ = F.flip_one(stored["scales"], work["scales"], flat, 31, "fp32")
img, _ = F.render_views(work, vm, K, W, H, sh)
F.restore_one(work["scales"], flat, cv)
fp = ((img[0] - clean[0]).abs().amax(-1) > 1 / 255).float().mean().item()
if fp > best[0]:
best = (fp, g)
g = best[1]; flat = g * 3
clean_t = all_ranks(work)
cv, _ = F.flip_one(stored["scales"], work["scales"], flat, 31, "fp32")
corr_t = all_ranks(work)
guard_t = all_ranks(F.apply_guard(work, bounds))
F.restore_one(work["scales"], flat, cv)
return {"T": T, "clean": clean_t, "corrupt": corr_t, "guard": guard_t}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--results_dir", default="/root/seu/results")
ap.add_argument("--scenes", default="chair,lego,ficus,hotdog")
ap.add_argument("--out", default="/root/seu/results/distributed")
ap.add_argument("--Ts", default="4,8,16,32,64")
ap.add_argument("--S", type=int, default=300)
ap.add_argument("--seed", type=int, default=0)
args = ap.parse_args()
os.makedirs(args.out, exist_ok=True)
log = os.path.join(args.out, "distributed.log")
Ts = [int(x) for x in args.Ts.split(",")]
for sc in args.scenes.split(","):
mp = os.path.join(args.results_dir, sc, "model.pt")
if os.path.exists(mp):
run(mp, args.out, Ts, args.S, args.seed, log)
print("DISTRIBUTED_DONE", flush=True)
if __name__ == "__main__":
main()