def main(A: T.handle, tensor: T.handle) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
# buffer definition
tensor_2 = T.buffer_decl([1, 10, 12, 16], dtype="int8", elem_offset=0, align=128, offset_factor=1)
A_1 = T.match_buffer(A, [1, 12, 14, 16], dtype="int8", elem_offset=0, align=128, offset_factor=1)
tensor_1 = T.match_buffer(tensor, [1, 8, 8, 16], dtype="int8", elem_offset=0, align=128, offset_factor=1)
# body
T.realize(tensor_1[0:1, 0:8, 0:8, 0:16], "")
T.realize(tensor_2[0:1, 0:6, 0:12, 0:16], "")
for ax1_outer in T.serial(0, 2):
for ax1 in T.serial(0, 6):
for ax2 in T.serial(0, 12):
for ax3 in T.serial(0, 16):
if T.likely(((ax1_outer < 1) or (ax1 >= 2)), dtype='bool') :
tensor_2[0, T.floormod((ax1 + (ax1_outer*4)), 6), ax2, ax3] = T.int8(0)
for dh in T.serial(0, 3):
for dw in T.serial(0, 3):
if T.likely(((ax1_outer < 1) or (ax1 >= 2)), dtype='bool'):
tensor_2[0, T.floormod((ax1 + (ax1_outer*4)), 6), ax2, ax3] = T.max(tensor_2[0, T.floormod((ax1 + (ax1_outer*4)), 6), ax2, ax3], A_1[0, ((ax1 + (ax1_outer*4)) + dh), (ax2 + dw), ax3])
for ax1_inner in T.serial(0, 4):
for ax2_inner in T.serial(0, 8):
for ax3_inner in T.serial(0, 16):
tensor_1[0, (ax1_inner + (ax1_outer*4)), ax2_inner, ax3_inner] = T.int8(0)
for dh_1 in T.serial(0, 3):
for dw_1 in T.serial(0, 5):
tensor_1[0, (ax1_inner + (ax1_outer*4)), ax2_inner, ax3_inner] = T.max(tensor_1[0, (ax1_inner + (ax1_outer*4)), ax2_inner, ax3_inner], tensor_2[0, T.floormod(((ax1_inner + (ax1_outer*4)) + dh_1), 6), (ax2_inner + dw_1), ax3_inner])
def non_perfect_tiling_cache(a: T.handle, b: T.handle) -> None:
X = T.match_buffer(a, [224, 224], dtype="float32")
Y = T.match_buffer(b, [224, 224], dtype="float32")
cache = T.alloc_buffer([224, 224], dtype="float32")
for hh_0, ww_0 in T.grid(28, 28):
for ax0 in T.serial(0, 10):
for ax1 in T.serial(0, 10):
with T.block("cache"):
h = T.axis.spatial(224, hh_0 * 8 - 1 + ax0)
w = T.axis.spatial(224, ww_0 * 8 - 1 + ax1)
T.where(1 <= hh_0 * 8 + ax0 and hh_0 * 8 + ax0 < 225
and 1 <= ww_0 * 8 + ax1 and ww_0 * 8 + ax1 < 225)
cache[h, w] = X[h, w]
for hh_1, ww_1, khh, kww in T.grid(8, 8, 3, 3):
with T.block("compute"):
h = T.axis.spatial(224, hh_0 * 8 + hh_1)
w = T.axis.spatial(224, ww_0 * 8 + ww_1)
kh, kw = T.axis.remap("RR", [khh, kww])
with T.init():
Y[h, w] = 0.0
Y[h, w] = T.max(
Y[h, w],
T.if_then_else(
T.likely(1 <= h + kh, dtype="bool")
and T.likely(h + kh < 225, dtype="bool")
and T.likely(1 <= w + kw, dtype="bool")
and T.likely(w + kw < 225, dtype="bool"),
cache[h + kh - 1, w + kw - 1],
0.0,
dtype="float32",
),
)
def compacted_spatial_tiled_pad_and_pooling(
X: T.Buffer[(64, 112, 112), "int32"], Y: T.Buffer[(64, 56, 56), "int32"]
) -> None:
for h_o, w_o in T.grid(14, 14):
with T.block():
T.reads(X[0:64, h_o * 8 - 1 : h_o * 8 + 8, w_o * 8 - 1 : w_o * 8 + 8])
T.writes(Y[h_o * 4 : h_o * 4 + 4, w_o * 4 : w_o * 4 + 4, 0:64])
X_cache = T.alloc_buffer([9, 9, 64], dtype="int32")
for ax0, ax1, ax2 in T.grid(64, 9, 9):
with T.block("cache"):
T.where(1 <= h_o * 8 + ax1 and 1 <= w_o * 8 + ax2)
T.reads(X[ax0, h_o * 8 + ax1 - 1, w_o * 8 + ax2 - 1])
T.writes(
X_cache[
h_o * 8 + ax1 - T.max(0, h_o * 8 - 1) - 1,
w_o * 8 + ax2 - T.max(0, w_o * 8 - 1) - 1,
ax0,
]
)
X_cache[
h_o * 8 + ax1 - T.max(0, h_o * 8 - 1) - 1,
w_o * 8 + ax2 - T.max(0, w_o * 8 - 1) - 1,
ax0,
] = X[ax0, h_o * 8 + ax1 - 1, w_o * 8 + ax2 - 1]
for h_i, w_i, kh, kw, c in T.grid(4, 4, 3, 3, 64):
with T.block("compute"):
T.reads(
X_cache[
h_o * 8 + h_i * 2 + kh - T.max(0, h_o * 8 - 1) - 1,
w_o * 8 + w_i * 2 + kw - T.max(0, w_o * 8 - 1) - 1,
c,
]
)
T.writes(Y[h_o * 4 + h_i, w_o * 4 + w_i, c])
if kh == 0 and kw == 0:
Y[h_o * 4 + h_i, w_o * 4 + w_i, c] = 0
Y[h_o * 4 + h_i, w_o * 4 + w_i, c] = T.max(
Y[h_o * 4 + h_i, w_o * 4 + w_i, c],
T.if_then_else(
T.likely(1 <= h_o * 8 + h_i * 2 + kh, dtype="bool")
and T.likely(1 <= w_o * 8 + w_i * 2 + kw, dtype="bool"),
X_cache[
h_o * 8 + h_i * 2 + kh - T.max(0, h_o * 8 - 1) - 1,
w_o * 8 + w_i * 2 + kw - T.max(0, w_o * 8 - 1) - 1,
c,
],
0,
dtype="int32",
),
)
def compacted_padding_pattern_inlined(
X: T.Buffer[(224, 224), "float32"], Y: T.Buffer[(224, 224), "float32"]
) -> None:
cache = T.alloc_buffer([224, 224], dtype="float32")
for h, w in T.grid(224, 224):
with T.block("cache"):
cache[h, w] = X[h, w]
for h, w, kh, kw in T.grid(224, 224, 3, 3):
with T.block("compute"):
Y[h, w] = T.max(
Y[h, w],
T.if_then_else(
T.likely(1 <= h + kh, dtype="bool")
and T.likely(h + kh < 225, dtype="bool")
and T.likely(1 <= w + kw, dtype="bool")
and T.likely(w + kw < 225, dtype="bool"),
cache[h + kh - 1, w + kw - 1],
0.0,
dtype="float32",
),
)
def padding_pattern_inlined(a: T.handle, b: T.handle) -> None:
X = T.match_buffer(a, [224, 224], dtype="float32")
Y = T.match_buffer(b, [224, 224], dtype="float32")
cache = T.alloc_buffer([224, 224], dtype="float32")
for h, w in T.grid(224, 224):
with T.block("cache"):
cache[h, w] = X[h, w]
for h, w, kh, kw in T.grid(224, 224, 3, 3):
with T.block("compute"):
Y[h, w] = T.max(
Y[h, w],
T.if_then_else(
T.likely(1 <= h + kh, dtype="bool")
and T.likely(h + kh < 225, dtype="bool")
and T.likely(1 <= w + kw, dtype="bool")
and T.likely(w + kw < 225, dtype="bool"),
cache[h + kh - 1, w + kw - 1],
0.0,
dtype="float32",
),
)
def spatial_tiled_pad_and_pooling(
X: T.Buffer[(64, 112, 112), "int32"], Y: T.Buffer[(64, 56, 56), "int32"]
) -> None:
for h_o, w_o in T.grid(14, 14):
with T.block():
X_cache = T.alloc_buffer([112, 112, 64], dtype="int32")
for ax0, ax1, ax2 in T.grid(64, 9, 9):
with T.block("cache"):
T.where(1 <= h_o * 8 + ax1 and 1 <= w_o * 8 + ax2)
T.reads(X[ax0, h_o * 8 - 1 + ax1, w_o * 8 - 1 + ax2])
T.writes(X_cache[h_o * 8 - 1 + ax1, w_o * 8 - 1 + ax2, ax0])
X_cache[h_o * 8 - 1 + ax1, w_o * 8 - 1 + ax2, ax0] = X[
ax0, h_o * 8 - 1 + ax1, w_o * 8 - 1 + ax2
]
for h_i, w_i, kh, kw, c in T.grid(4, 4, 3, 3, 64):
with T.block("compute"):
T.reads(
X_cache[(h_o * 4 + h_i) * 2 + kh - 1, (w_o * 4 + w_i) * 2 + kw - 1, c]
)
T.writes(Y[h_o * 4 + h_i, w_o * 4 + w_i, c])
if kh == 0 and kw == 0:
Y[h_o * 4 + h_i, w_o * 4 + w_i, c] = 0
Y[h_o * 4 + h_i, w_o * 4 + w_i, c] = T.max(
Y[h_o * 4 + h_i, w_o * 4 + w_i, c],
T.if_then_else(
T.likely(1 <= (h_o * 4 + h_i) * 2 + kh, dtype="bool")
and T.likely((h_o * 4 + h_i) * 2 + kh < 113, dtype="bool")
and T.likely(1 <= (w_o * 4 + w_i) * 2 + kw, dtype="bool")
and T.likely((w_o * 4 + w_i) * 2 + kw < 113, dtype="bool"),
X_cache[
(h_o * 4 + h_i) * 2 + kh - 1,
(w_o * 4 + w_i) * 2 + kw - 1,
c,
],
0,
dtype="int32",
),
)