feat: kernel JIT batch top-max-per-variant (opt-in)

Nuovo kernel _jit_top_max_per_variant: prange esterno sulle varianti
invece di n_vars chiamate JIT separate via ThreadPoolExecutor.
Wrapper Python top_max_per_variant prepara buffer flat (offsets +
dx_flat/dy_flat/bins_flat) e bg per scala.

Default batch_top=False perche su benchmark realistici (Linux 13 core,
72-180 varianti) ThreadPoolExecutor + kernel singolo che rilascia GIL
e gia ottimale. Path batch_top=True utile come opzione per scenari
con n_vars >>> n_threads o overhead chiamate JIT dominante.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-04 15:35:51 +02:00
parent 89b59b3ea3
commit 6704d66cd5
2 changed files with 135 additions and 1 deletions
+114
View File
@@ -159,6 +159,63 @@ if HAS_NUMBA:
acc[y, x] = 0.0
return acc
@nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
def _jit_top_max_per_variant(
spread: np.ndarray, # uint8 (H, W)
dx_flat: np.ndarray, # int32 (sum_N,)
dy_flat: np.ndarray, # int32 (sum_N,)
bins_flat: np.ndarray, # int8 (sum_N,)
offsets: np.ndarray, # int32 (n_vars+1,) prefix sum
bit_active: np.uint8,
bg_per_variant: np.ndarray, # float32 (n_vars, H, W) - 1 per scala
scale_idx: np.ndarray, # int32 (n_vars,) idx in bg_per_variant
) -> np.ndarray:
"""Batch: per ogni variante calcola max score (rescored bg), ritorna
array float32 (n_vars,). Parallelismo prange ESTERNO sulle varianti
elimina overhead di n_vars chiamate JIT separate (avg ~20us per
chiamata su template piccoli) + pool thread Python.
Pensato per fase TOP del pruning quando n_vars >> n_threads.
"""
n_vars = offsets.shape[0] - 1
H, W = spread.shape
out = np.zeros(n_vars, dtype=np.float32)
for vi in nb.prange(n_vars):
i0 = offsets[vi]; i1 = offsets[vi + 1]
N = i1 - i0
if N == 0:
out[vi] = -1.0
continue
si = scale_idx[vi]
inv = nb.float32(1.0 / N)
best = nb.float32(-1.0)
for y in range(H):
for x in range(W):
s = nb.float32(0.0)
for k in range(N):
b = bins_flat[i0 + k]
mask = np.uint8(1) << b
if (bit_active & mask) == 0:
continue
ddy = dy_flat[i0 + k]
yy = y + ddy
if yy < 0 or yy >= H:
continue
ddx = dx_flat[i0 + k]
xx = x + ddx
if xx < 0 or xx >= W:
continue
if spread[yy, xx] & mask:
s += nb.float32(1.0)
s *= inv
bgv = bg_per_variant[si, y, x]
if bgv < 1.0:
r = (s - bgv) / (1.0 - bgv + 1e-6)
if r > best:
best = r
out[vi] = best if best > 0.0 else 0.0
return out
@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]."""
@@ -185,6 +242,12 @@ if HAS_NUMBA:
_jit_score_bitmap(spread, dx, dy, b, np.uint8(0xFF))
bg = np.zeros((32, 32), dtype=np.float32)
_jit_score_bitmap_rescored(spread, dx, dy, b, np.uint8(0xFF), bg)
offsets = np.array([0, 1], dtype=np.int32)
scale_idx = np.zeros(1, dtype=np.int32)
bg_pv = np.zeros((1, 32, 32), dtype=np.float32)
_jit_top_max_per_variant(
spread, dx, dy, b, offsets, np.uint8(0xFF), bg_pv, scale_idx,
)
_jit_popcount_density(spread)
else: # pragma: no cover
@@ -198,6 +261,12 @@ else: # pragma: no cover
def _jit_score_bitmap_rescored(spread, dx, dy, bins, bit_active, bg):
raise RuntimeError("numba non disponibile")
def _jit_top_max_per_variant(
spread, dx_flat, dy_flat, bins_flat, offsets, bit_active,
bg_per_variant, scale_idx,
):
raise RuntimeError("numba non disponibile")
def _jit_popcount_density(spread):
raise RuntimeError("numba non disponibile")
@@ -246,6 +315,51 @@ def score_bitmap_rescored(
return np.maximum(0.0, out).astype(np.float32)
def top_max_per_variant(
spread: np.ndarray,
dx_list: list, dy_list: list, bin_list: list,
bg_per_scale: dict,
variant_scales: list,
bit_active: int,
) -> np.ndarray:
"""Wrapper: prepara buffer flat e chiama kernel batch su tutte le varianti.
Parallelismo Numba prange-esterno sulle varianti (n_vars >> n_threads
tipicamente per top-pruning) → meglio del thread-pool Python che paga
overhead di n_vars chiamate JIT separate.
"""
if not HAS_NUMBA or len(dx_list) == 0:
return np.array([], dtype=np.float32)
n_vars = len(dx_list)
sizes = [len(d) for d in dx_list]
offsets = np.zeros(n_vars + 1, dtype=np.int32)
offsets[1:] = np.cumsum(sizes)
total = int(offsets[-1])
dx_flat = np.empty(total, dtype=np.int32)
dy_flat = np.empty(total, dtype=np.int32)
bins_flat = np.empty(total, dtype=np.int8)
for vi, (dx, dy, bn) in enumerate(zip(dx_list, dy_list, bin_list)):
i0 = int(offsets[vi]); i1 = int(offsets[vi + 1])
dx_flat[i0:i1] = dx
dy_flat[i0:i1] = dy
bins_flat[i0:i1] = bn
# bg per variante: indicizzato per scala
scales_unique = sorted(bg_per_scale.keys())
scale_to_idx = {s: i for i, s in enumerate(scales_unique)}
H, W = spread.shape
bg_pv = np.empty((len(scales_unique), H, W), dtype=np.float32)
for s, idx in scale_to_idx.items():
bg_pv[idx] = bg_per_scale[s]
scale_idx = np.array(
[scale_to_idx[s] for s in variant_scales], dtype=np.int32,
)
return _jit_top_max_per_variant(
np.ascontiguousarray(spread, dtype=np.uint8),
dx_flat, dy_flat, bins_flat, offsets, np.uint8(bit_active),
bg_pv, scale_idx,
)
def popcount_density(spread: np.ndarray) -> np.ndarray:
if HAS_NUMBA:
return _jit_popcount_density(np.ascontiguousarray(spread, dtype=np.uint8))