perf: spread bitmap uint8 + pre-NMS prima refine (3.5x globale, 49x worst case)

Due ottimizzazioni chiave:

1. Spread bitmap uint8 invece di response map (N_BINS, H, W) float32
   - 32x meno memoria, cache-friendly
   - Nuovi kernel Numba: _jit_score_bitmap, _jit_popcount_density
   - Formato: spread[y,x] bit b = bin b attivo nel raggio di spread
   - _refine_angle usa slicing su bitmap con mask & bit

2. Pre-NMS prima di refine_angle/verify_ncc
   - Problema: loop 'for raw in candidati' applicava refine+verify A OGNI
     candidato prima del check NMS → 2000+ refine chiamati per ~25 match
   - Fix: pre-NMS su (cx, cy) subpixel, limita a max_matches*3 candidati,
     poi refine + verify solo su quelli
   - Esempio worst case: lama_full_fast 55.9s → 1.13s (49x)

Benchmark suite 16 scenari (4 immagini x full/part x fast/preciso):
  prima: totale find 94.6s
  dopo:  totale find 27.3s (3.5x globale)
  casi peggiori <5s (prima erano >50s)

ROI parziali (solo metà oggetto) funzionano in tutti i casi.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-04-24 02:11:33 +02:00
parent d27676cfe6
commit ba54b42fdc
3 changed files with 262 additions and 38 deletions
+96 -4
View File
@@ -45,13 +45,12 @@ if HAS_NUMBA:
resp: np.ndarray, # float32 (N_BINS, H, W)
dx: np.ndarray, # int32 (N,)
dy: np.ndarray, # int32 (N,)
bins: np.ndarray, # int8 or int32 (N,)
bins: np.ndarray, # int8 (N,)
bin_active: np.ndarray, # bool_ (N_BINS,)
) -> np.ndarray:
n_bins, H, W = resp.shape
_, H, W = resp.shape
N = dx.shape[0]
acc = np.zeros((H, W), dtype=np.float32)
# Parallelizza per riga: niente race (ogni y scrive solo acc[y, :])
for y in nb.prange(H):
for i in range(N):
b = bins[i]
@@ -73,7 +72,59 @@ if HAS_NUMBA:
acc[y, x] *= inv
return acc
# Warmup: precompila con dummy data
@nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
def _jit_score_bitmap(
spread: np.ndarray, # uint8 (H, W), bit b = bin b attivo
dx: np.ndarray, # int32 (N,)
dy: np.ndarray, # int32 (N,)
bins: np.ndarray, # int8 (N,) bin per ogni feature
bit_active: np.uint8, # bitmask bin attivi in scena
) -> np.ndarray:
"""score[y,x] = (Σ_i [bit bins[i] acceso in spread[y+dy_i, x+dx_i]]) / N.
32× meno memoria di response map float32 → cache-friendly.
"""
H, W = spread.shape
N = dx.shape[0]
acc = np.zeros((H, W), dtype=np.float32)
for y in nb.prange(H):
for i in range(N):
b = bins[i]
mask = np.uint8(1) << b
if (bit_active & mask) == 0:
continue
ddy = dy[i]
yy = y + ddy
if yy < 0 or yy >= H:
continue
ddx = dx[i]
x_lo = 0 if ddx >= 0 else -ddx
x_hi = W if ddx <= 0 else W - ddx
for x in range(x_lo, x_hi):
if spread[yy, x + ddx] & mask:
acc[y, x] += 1.0
if N > 0:
inv = 1.0 / N
for y in nb.prange(H):
for x in range(W):
acc[y, x] *= inv
return acc
@nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
def _jit_popcount_density(spread: np.ndarray) -> np.ndarray:
"""Conta bit set per pixel: ritorna (H, W) float32 in [0..8]."""
H, W = spread.shape
out = np.zeros((H, W), dtype=np.float32)
for y in nb.prange(H):
for x in range(W):
v = spread[y, x]
# popcount manuale
v = (v & 0x55) + ((v >> 1) & 0x55)
v = (v & 0x33) + ((v >> 2) & 0x33)
v = (v & 0x0F) + ((v >> 4) & 0x0F)
out[y, x] = float(v)
return out
def _warmup():
resp = np.zeros((8, 32, 32), dtype=np.float32)
dx = np.zeros(1, dtype=np.int32)
@@ -81,16 +132,57 @@ if HAS_NUMBA:
b = np.zeros(1, dtype=np.int8)
ba = np.ones(8, dtype=np.bool_)
_jit_score_by_shift(resp, dx, dy, b, ba)
spread = np.zeros((32, 32), dtype=np.uint8)
_jit_score_bitmap(spread, dx, dy, b, np.uint8(0xFF))
_jit_popcount_density(spread)
else: # pragma: no cover
def _jit_score_by_shift(resp, dx, dy, bins, bin_active):
raise RuntimeError("numba non disponibile")
def _jit_score_bitmap(spread, dx, dy, bins, bit_active):
raise RuntimeError("numba non disponibile")
def _jit_popcount_density(spread):
raise RuntimeError("numba non disponibile")
def _warmup():
pass
def score_bitmap(
spread: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
bit_active: int,
) -> np.ndarray:
"""Dispatch bitmap: JIT se numba, fallback numpy."""
if HAS_NUMBA and len(dx) > 0:
return _jit_score_bitmap(
np.ascontiguousarray(spread, dtype=np.uint8),
np.ascontiguousarray(dx, dtype=np.int32),
np.ascontiguousarray(dy, dtype=np.int32),
np.ascontiguousarray(bins, dtype=np.int8),
np.uint8(bit_active),
)
# Fallback numpy (lento): converte bitmap a response 3D
H, W = spread.shape
resp = np.zeros((8, H, W), dtype=np.float32)
for b in range(8):
resp[b] = ((spread >> b) & 1).astype(np.float32)
return _numpy_score_by_shift(resp, dx, dy, bins, None)
def popcount_density(spread: np.ndarray) -> np.ndarray:
if HAS_NUMBA:
return _jit_popcount_density(np.ascontiguousarray(spread, dtype=np.uint8))
# Fallback
H, W = spread.shape
out = np.zeros((H, W), dtype=np.float32)
for b in range(8):
out += ((spread >> b) & 1).astype(np.float32)
return out
def score_by_shift(
resp: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
bin_has_data: np.ndarray | None = None,