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], "")
for ax1_outer in T.serial(0, 2):
T.realize(tensor_2[0:1, (ax1_outer*4):((ax1_outer*4) + 6), 0:12, 0:16], "")
T.attr(tensor_2, "rolling_buffer_scope", True)
for ax1 in T.serial(0, 6):
for ax2 in T.serial(0, 12):
for ax3 in T.serial(0, 16):
tensor_2[0, (ax1 + (ax1_outer*4)), ax2, ax3] = T.int8(0)
for dh in T.serial(0, 3):
for dw in T.serial(0, 3):
tensor_2[0, (ax1 + (ax1_outer*4)), ax2, ax3] = T.max(tensor_2[0, (ax1 + (ax1_outer*4)), 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, ((ax1_inner + (ax1_outer*4)) + dh_1), (ax2_inner + dw_1), ax3_inner])
def scale_by_two(A: T.Buffer[(8192, ), "int8"], C: T.Buffer[(8192, ), "int8"]):
for i in T.serial(
0,
8192,
):
with T.block("C"):
C[i] = A[i] * T.int8(2)
def scale_by_two(buffer_a: T.Buffer[(8192, ), "int8"],
buffer_c: T.Buffer[(8192, ), "int8"]):
for i in T.serial(
0,
8192,
):
with T.block("C"):
buffer_c[i] = buffer_a[i] * T.int8(2)
def pooling_decompose_3(
x: T.Buffer[(1, 16, 225, 225), "int8"],
tensor: T.Buffer[(1, 16, 225, 225), "int8"]) -> None:
pad_temp = T.alloc_buffer([1, 16, 231, 231], dtype="int8")
for i0, i2_0, i3_0 in T.grid(1, 3, 3):
for ax0, ax1, ax2 in T.grid(16, 86, 86):
with T.block("pad_temp_pad_const"):
ax0_1 = T.axis.spatial(1, 0)
ax1_1 = T.axis.spatial(16, ax0)
ax2_1 = T.axis.spatial(231, i2_0 * 80 + ax1)
ax3 = T.axis.spatial(231, i3_0 * 80 + ax2)
T.where(i2_0 * 80 + ax1 < 231 and i3_0 * 80 + ax2 < 231)
T.reads()
T.writes(pad_temp[ax0_1, ax1_1, ax2_1, ax3])
pad_temp[ax0_1, ax1_1, ax2_1, ax3] = T.int8(0)
for ax0, ax1, ax2 in T.grid(16, 86, 86):
with T.block("pad_temp"):
ax0_2 = T.axis.spatial(1, 0)
ax1_2 = T.axis.spatial(16, ax0)
ax2_2 = T.axis.spatial(225, i2_0 * 80 + ax1 - 3)
ax3 = T.axis.spatial(225, i3_0 * 80 + ax2 - 3)
T.where(3 <= i2_0 * 80 + ax1 and i2_0 * 80 + ax1 < 228
and 3 <= i3_0 * 80 + ax2 and i3_0 * 80 + ax2 < 228
and i2_0 * 80 + ax1 < 231
and i3_0 * 80 + ax2 < 231)
T.reads(x[ax0_2, ax1_2, ax2_2, ax3])
T.writes(pad_temp[ax0_2, ax1_2, ax2_2 + 3, ax3 + 3])
pad_temp[ax0_2, ax1_2, ax2_2 + 3,
ax3 + 3] = x[ax0_2, ax1_2, ax2_2, ax3]
for i1, i2_1, i3_1, i4, i5 in T.grid(16, 80, 80, 7, 7):
with T.block("tensor"):
ax0_3, ax1_3 = T.axis.remap("SS", [i0, i1])
ax2_3 = T.axis.spatial(225, i2_0 * 80 + i2_1)
ax3 = T.axis.spatial(225, i3_0 * 80 + i3_1)
rv0, rv1 = T.axis.remap("RR", [i4, i5])
T.where(i2_0 * 80 + i2_1 < 225 and i3_0 * 80 + i3_1 < 225)
T.reads(pad_temp[ax0_3, ax1_3, ax2_3 + rv0, ax3 + rv1])
T.writes(tensor[ax0_3, ax1_3, ax2_3, ax3])
with T.init():
tensor[ax0_3, ax1_3, ax2_3, ax3] = T.int8(0)
tensor[ax0_3, ax1_3, ax2_3, ax3] = (
tensor[ax0_3, ax1_3, ax2_3, ax3] +
pad_temp[ax0_3, ax1_3, ax2_3 + rv0, ax3 + rv1])
def pooling_decompose_0(x: T.Buffer[(1, 16, 225, 225), "int8"],
tensor: T.Buffer[(1, 16, 225, 225), "int8"]):
pad_temp = T.alloc_buffer([1, 16, 231, 231], dtype="int8")
for i0, i1, i2, i3 in T.grid(1, 16, 231, 231):
with T.block("pad_temp_pad_const"):
ax0, ax1, ax2, ax3 = T.axis.remap("SSSS", [i0, i1, i2, i3])
pad_temp[ax0, ax1, ax2, ax3] = T.int8(0)
for i0, i1, i2, i3 in T.grid(1, 16, 225, 225):
with T.block("pad_temp"):
ax0, ax1, ax2, ax3 = T.axis.remap("SSSS", [i0, i1, i2, i3])
pad_temp[ax0, ax1, ax2 + 3, ax3 + 3] = x[ax0, ax1, ax2, ax3]
for i0, i1, i2, i3, i4, i5 in T.grid(1, 16, 225, 225, 7, 7):
with T.block("tensor"):
ax0, ax1, ax2, ax3, rv0, rv1 = T.axis.remap(
"SSSSRR", [i0, i1, i2, i3, i4, i5])
with T.init():
tensor[ax0, ax1, ax2, ax3] = T.int8(0)
tensor[ax0, ax1, ax2,
ax3] = (tensor[ax0, ax1, ax2, ax3] +
pad_temp[ax0, ax1, ax2 + rv0, ax3 + rv1])
def sum_pool_2d(x: T.Buffer[(1, 16, 225, 225), "int8"],
tensor: T.Buffer[(1, 16, 225, 225), "int8"]):
pad_temp = T.alloc_buffer([1, 16, 231, 231], dtype="int8")
for i0, i1, i2, i3 in T.grid(1, 16, 231, 231):
with T.block("pad_temp"):
ax0, ax1, ax2, ax3 = T.axis.remap("SSSS", [i0, i1, i2, i3])
pad_temp[ax0, ax1, ax2, ax3] = T.if_then_else(
3 <= ax2 and ax2 < 228 and 3 <= ax3 and ax3 < 228,
x[ax0, ax1, ax2 - 3, ax3 - 3],
T.int8(0),
dtype="int8",
)
for i0, i1, i2, i3, i4, i5 in T.grid(1, 16, 225, 225, 7, 7):
with T.block("tensor"):
ax0, ax1, ax2, ax3, rv0, rv1 = T.axis.remap(
"SSSSRR", [i0, i1, i2, i3, i4, i5])
with T.init():
tensor[ax0, ax1, ax2, ax3] = T.int8(0)
tensor[ax0, ax1, ax2,
ax3] = (tensor[ax0, ax1, ax2, ax3] +
pad_temp[ax0, ax1, ax2 + rv0, ax3 + rv1])
def main(placeholder: T.Buffer[(301056, ), "int8"],
ethosu_write: T.Buffer[(75264, ), "int8"]) -> None:
T.func_attr({
"from_legacy_te_schedule": True,
"global_symbol": "main",
"tir.noalias": True
})
T.preflattened_buffer(placeholder, [1, 56, 56, 96],
dtype='int8',
data=placeholder.data)
T.preflattened_buffer(ethosu_write, [1, 56, 56, 24],
dtype='int8',
data=ethosu_write.data)
buffer1 = T.buffer_decl([2608], "uint8")
buffer2 = T.buffer_decl([240], "uint8")
buffer3 = T.buffer_decl([736], "uint8")
buffer4 = T.buffer_decl([240], "uint8")
p1 = T.allocate([2608],
"uint8",
"global",
annotations={"disable_lower_builtin": True})
p2 = T.allocate([240],
"uint8",
"global",
annotations={"disable_lower_builtin": True})
p3 = T.allocate([736],
"uint8",
"global",
annotations={"disable_lower_builtin": True})
p4 = T.allocate([240],
"uint8",
"global",
annotations={"disable_lower_builtin": True})
p5 = T.allocate([75264],
"int8",
"global",
annotations={"disable_lower_builtin": True})
p6 = T.allocate([75264],
"int8",
"global",
annotations={"disable_lower_builtin": True})
T.evaluate(
T.call_extern("ethosu_copy",
buffer1[0],
2608,
p1[0],
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_copy",
buffer2[0],
240,
p2[0],
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_copy",
buffer3[0],
736,
p3[0],
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_copy",
buffer4[0],
240,
p4[0],
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_conv2d",
"int8",
56,
56,
96,
56,
0,
56,
placeholder[0],
0,
0,
0,
T.float32(0.5),
10,
"NHWC",
5376,
96,
1,
"int8",
56,
56,
24,
56,
0,
56,
p5[0],
0,
0,
0,
T.float32(0.25),
14,
"NHWC",
1344,
24,
1,
1,
1,
1,
1,
1,
1,
p1[0],
2608,
T.int8(-1),
T.int8(-1),
12,
p2[0],
240,
T.int8(-1),
T.int8(-1),
0,
0,
0,
0,
"NONE",
0,
0,
"TFL",
"NONE",
0,
0,
0,
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_conv2d",
"int8",
56,
56,
24,
56,
0,
56,
p5[0],
0,
0,
0,
T.float32(0.5),
10,
"NHWC",
1344,
24,
1,
"int8",
56,
56,
24,
56,
0,
56,
p6[0],
0,
0,
0,
T.float32(0.25),
14,
"NHWC",
1344,
24,
1,
1,
1,
1,
1,
1,
1,
p3[0],
736,
T.int8(-1),
T.int8(-1),
12,
p4[0],
240,
T.int8(-1),
T.int8(-1),
0,
0,
0,
0,
"NONE",
0,
0,
"TFL",
"NONE",
0,
0,
0,
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_binary_elementwise",
"int8",
56,
56,
24,
56,
0,
56,
p5[0],
0,
0,
0,
T.float32(1),
0,
"NHWC",
1344,
24,
1,
"int8",
56,
56,
24,
56,
0,
56,
p6[0],
0,
0,
0,
T.float32(1),
0,
"NHWC",
1344,
24,
1,
"int8",
56,
56,
24,
56,
0,
56,
ethosu_write[0],
0,
0,
0,
T.float32(1),
0,
"NHWC",
1344,
24,
1,
"ADD",
0,
"NONE",
0,
0,
"TFL",
0,
0,
0,
dtype="handle"))
def main(placeholder: T.Buffer[(1536, ), "int8"],
placeholder_1: T.Buffer[(1280, ), "int8"],
T_concat: T.Buffer[(4096, ), "int8"]) -> None:
# function attr dict
T.func_attr({
"from_legacy_te_schedule": True,
"global_symbol": "main",
"tir.noalias": True
})
buffer = T.buffer_decl([2992], "uint8")
buffer_1 = T.buffer_decl([160], "uint8")
buffer_2 = T.buffer_decl([2992], "uint8")
buffer_3 = T.buffer_decl([160], "uint8")
buffer_4 = T.buffer_decl([2992], "uint8")
buffer_5 = T.buffer_decl([160], "uint8")
buffer_6 = T.buffer_decl([2992], "uint8")
buffer_7 = T.buffer_decl([160], "uint8")
# body
T_concat_1 = T.allocate([2816],
"int8",
"global",
annotations={"disable_lower_builtin": True})
T.evaluate(
T.call_extern("ethosu_conv2d",
"int8",
8,
10,
16,
8,
0,
10,
placeholder_1[0],
0,
0,
0,
T.float32(0.5),
10,
"NHWC",
160,
16,
1,
"int8",
8,
10,
16,
8,
0,
10,
T_concat_1[192],
0,
0,
0,
T.float32(0.25),
14,
"NHWC",
352,
16,
1,
3,
3,
1,
1,
1,
1,
buffer[0],
2992,
T.int8(-1),
T.int8(-1),
12,
buffer_1[0],
160,
T.int8(-1),
T.int8(-1),
1,
1,
1,
1,
"NONE",
0,
0,
"TFL",
"NONE",
0,
0,
0,
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_conv2d",
"int8",
8,
10,
16,
8,
0,
10,
T_concat_1[192],
0,
0,
0,
T.float32(0.5),
10,
"NHWC",
352,
16,
1,
"int8",
8,
10,
16,
8,
0,
10,
T_concat[352],
0,
0,
0,
T.float32(0.25),
14,
"NHWC",
512,
16,
1,
3,
3,
1,
1,
1,
1,
buffer_2[0],
2992,
T.int8(-1),
T.int8(-1),
12,
buffer_3[0],
160,
T.int8(-1),
T.int8(-1),
1,
1,
1,
1,
"NONE",
0,
0,
"TFL",
"NONE",
0,
0,
0,
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_conv2d",
"int8",
8,
12,
16,
8,
0,
12,
placeholder[0],
0,
0,
0,
T.float32(0.5),
10,
"NHWC",
192,
16,
1,
"int8",
8,
12,
16,
8,
0,
12,
T_concat_1[0],
0,
0,
0,
T.float32(0.25),
14,
"NHWC",
352,
16,
1,
3,
3,
1,
1,
1,
1,
buffer_4[0],
2992,
T.int8(-1),
T.int8(-1),
12,
buffer_5[0],
160,
T.int8(-1),
T.int8(-1),
1,
1,
1,
1,
"NONE",
0,
0,
"TFL",
"NONE",
0,
0,
0,
dtype="handle"))
T.evaluate(
T.call_extern("ethosu_conv2d",
"int8",
8,
22,
16,
8,
0,
22,
T_concat_1[0],
0,
0,
0,
T.float32(0.5),
10,
"NHWC",
352,
16,
1,
"int8",
8,
22,
16,
8,
0,
22,
T_concat[0],
0,
0,
0,
T.float32(0.25),
14,
"NHWC",
512,
16,
1,
3,
3,
1,
1,
1,
1,
buffer_6[0],
2992,
T.int8(-1),
T.int8(-1),
12,
buffer_7[0],
160,
T.int8(-1),
T.int8(-1),
1,
1,
1,
1,
"NONE",
0,
0,
"TFL",
"NONE",
0,
0,
0,
dtype="handle"))
def main(placeholder_5: T.Buffer[(8192,), "int8"], ethosu_write_1: T.Buffer[(4096,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([416], "uint8")
buffer_1 = T.buffer_decl([112], "uint8")
buffer_2 = T.buffer_decl([272], "uint8")
buffer_3 = T.buffer_decl([64], "uint8")
T.preflattened_buffer(placeholder_5, [1, 16, 16, 32], dtype="int8", data=placeholder_5.data)
T.preflattened_buffer(ethosu_write_1, [1, 16, 16, 16], dtype="int8", data=ethosu_write_1.data)
# body
placeholder_global_unrolled_iter_0 = T.allocate([416], "uint8", "global", annotations={"disable_lower_builtin": True})
placeholder_d_global_unrolled_iter_0 = T.allocate([112], "uint8", "global", annotations={"disable_lower_builtin": True})
placeholder_global_unrolled_iter_1 = T.allocate([272], "uint8", "global", annotations={"disable_lower_builtin": True})
placeholder_d_global_unrolled_iter_1 = T.allocate([64], "uint8", "global", annotations={"disable_lower_builtin": True})
T.evaluate(T.call_extern("ethosu_copy", buffer[0], 416, placeholder_global_unrolled_iter_0[0], dtype="handle"))
T.evaluate(T.call_extern("ethosu_copy", buffer_1[0], 112, placeholder_d_global_unrolled_iter_0[0], dtype="handle"))
T.evaluate(T.call_extern("ethosu_copy", buffer_2[0], 272, placeholder_global_unrolled_iter_1[0], dtype="handle"))
T.evaluate(T.call_extern("ethosu_copy", buffer_3[0], 64, placeholder_d_global_unrolled_iter_1[0], dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 16, 16, 32, 16, 0, 16, placeholder_5[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 512, 32, 1, "int8", 16, 16, 10, 16, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 256, 16, 1, 1, 1, 1, 1, 1, 1, placeholder_global_unrolled_iter_0[0], 416, T.int8(-1), T.int8(-1), 12, placeholder_d_global_unrolled_iter_0[0], 112, T.int8(-1), T.int8(-1), 0, 0, 0, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 16, 16, 32, 16, 0, 16, placeholder_5[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 512, 32, 1, "int8", 16, 16, 6, 16, 0, 16, ethosu_write_1[10], 0, 0, 0, T.float32(0.25), 14, "NHWC", 256, 16, 1, 1, 1, 1, 1, 1, 1, placeholder_global_unrolled_iter_1[0], 272, T.int8(-1), T.int8(-1), 12, placeholder_d_global_unrolled_iter_1[0], 64, T.int8(-1), T.int8(-1), 0, 0, 0, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
def main(placeholder_3: T.Buffer[(8192,), "int8"], ethosu_write_1: T.Buffer[(2048,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([80], "uint8")
buffer_1 = T.buffer_decl([304], "uint8")
T.preflattened_buffer(placeholder_3, [1, 16, 16, 32], dtype="int8", data=placeholder_3.data)
T.preflattened_buffer(ethosu_write_1, [1, 16, 16, 8], dtype="int8", data=ethosu_write_1.data)
# body
placeholder_global = T.allocate([304], "uint8", "global", annotations={"disable_lower_builtin": True})
placeholder_d_global = T.allocate([80], "uint8", "global", annotations={"disable_lower_builtin": True})
T.evaluate(T.call_extern("ethosu_copy", buffer_1[0], 304, placeholder_global[0], dtype="handle"))
T.evaluate(T.call_extern("ethosu_copy", buffer[0], 80, placeholder_d_global[0], dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 16, 16, 32, 16, 0, 16, placeholder_3[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 512, 32, 1, "int8", 16, 16, 8, 16, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 128, 8, 1, 1, 1, 1, 1, 1, 1, placeholder_global[0], 304, T.int8(-1), T.int8(-1), 12, placeholder_d_global[0], 80, T.int8(-1), T.int8(-1), 0, 0, 0, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
def main(placeholder_3: T.Buffer[(192,), "int8"], ethosu_write_1: T.Buffer[(768,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([160], "uint8")
buffer_1 = T.buffer_decl([848], "uint8")
T.preflattened_buffer(placeholder_3, [192], 'int8', data=placeholder_3.data)
T.preflattened_buffer(ethosu_write_1, [1, 8, 6, 16], 'int8', data=ethosu_write_1.data)
# body
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 5, 6, 4, 5, 0, 6, placeholder_3[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 24, 4, 1, "int8", 4, 6, 16, 4, 0, 6, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 96, 16, 1, 3, 3, 1, 1, 1, 1, buffer_1[0], 848, T.int8(-1), T.int8(-1), 12, buffer[0], 160, T.int8(-1), T.int8(-1), 1, 1, 0, 1, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 5, 6, 4, 5, 0, 6, placeholder_3[72], 0, 0, 0, T.float32(0.5), 10, "NHWC", 24, 4, 1, "int8", 4, 6, 16, 4, 0, 6, ethosu_write_1[384], 0, 0, 0, T.float32(0.25), 14, "NHWC", 96, 16, 1, 3, 3, 1, 1, 1, 1, buffer_1[0], 848, T.int8(-1), T.int8(-1), 12, buffer[0], 160, T.int8(-1), T.int8(-1), 0, 1, 1, 1, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
def main(placeholder_3: T.Buffer[(315,), "int8"], ethosu_write_1: T.Buffer[(240,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([160], "uint8")
buffer_1 = T.buffer_decl([656], "uint8")
T.preflattened_buffer(placeholder_3, [1, 7, 9, 5], 'int8', data=placeholder_3.data)
T.preflattened_buffer(ethosu_write_1, [1, 3, 5, 16], 'int8', data=ethosu_write_1.data)
# body
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 3, 5, 3, 3, 0, 5, placeholder_3[146], 0, 0, 0, T.float32(0.5), 10, "NHWC", 45, 5, 1, "int8", 3, 5, 16, 3, 0, 5, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 80, 16, 1, 3, 3, 1, 1, 1, 1, buffer_1[0], 656, T.int8(-1), T.int8(-1), 12, buffer[0], 160, T.int8(-1), T.int8(-1), 1, 1, 1, 1, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
def main(placeholder: T.Buffer[(1024,), "int8"], ethosu_write: T.Buffer[(32768,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([1456], "uint8")
buffer_1 = T.buffer_decl([352], "uint8")
buffer_2 = T.buffer_decl([11040], "uint8")
buffer_3 = T.buffer_decl([272], "uint8")
T.preflattened_buffer(placeholder, [1, 8, 1, 8, 16], 'int8', data=placeholder.data)
T.preflattened_buffer(ethosu_write, [1, 32, 2, 32, 16], 'int8', data=ethosu_write.data)
# body
ethosu_write_1 = T.allocate([12288], "int8", "global", annotations={"disable_lower_builtin":True})
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 8, 8, 3, 8, 0, 8, placeholder[0], 0, 0, 0, T.float32(0.5), 10, "NHCWB16", 128, 16, 1, "int8", 16, 16, 35, 16, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHCWB16", 768, 16, 256, 3, 3, 1, 1, 1, 1, buffer[0], 1456, T.int8(-1), T.int8(-1), 12, buffer_1[0], 352, T.int8(-1), T.int8(-1), 1, 1, 1, 1, "NONE", 0, 0, "TFL", "NEAREST", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 16, 16, 35, 16, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.5), 10, "NHCWB16", 768, 16, 256, "int8", 32, 32, 26, 32, 0, 32, ethosu_write[0], 0, 0, 0, T.float32(0.25), 14, "NHCWB16", 1024, 16, 512, 3, 3, 1, 1, 1, 1, buffer_2[0], 11040, T.int8(-1), T.int8(-1), 12, buffer_3[0], 272, T.int8(-1), T.int8(-1), 1, 1, 1, 1, "NONE", 0, 0, "TFL", "NEAREST", 0, 0, 0, dtype="handle"))
def main(placeholder: T.Buffer[(192,), "int8"], ethosu_write: T.Buffer[(8192,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([160], "uint8")
buffer_1 = T.buffer_decl([320], "uint8")
buffer_2 = T.buffer_decl([304], "uint8")
buffer_3 = T.buffer_decl([80], "uint8")
T.preflattened_buffer(placeholder, [1, 8, 8, 3], 'int8', data=placeholder.data)
T.preflattened_buffer(ethosu_write, [1, 32, 32, 8], 'int8', data=ethosu_write.data)
# body
ethosu_write_1 = T.allocate([4096], "int8", "global", annotations={"disable_lower_builtin":True})
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 4, 8, 3, 4, 0, 8, placeholder[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 24, 3, 1, "int8", 8, 16, 32, 8, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 512, 32, 1, 1, 1, 1, 1, 1, 1, buffer[0], 160, T.int8(-1), T.int8(-1), 12, buffer_1[0], 320, T.int8(-1), T.int8(-1), 0, 0, 0, 0, "NONE", 0, 0, "TFL", "ZEROS", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 8, 16, 32, 8, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 512, 32, 1, "int8", 16, 32, 8, 16, 0, 32, ethosu_write[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 256, 8, 1, 1, 1, 1, 1, 1, 1, buffer_2[0], 304, T.int8(-1), T.int8(-1), 12, buffer_3[0], 80, T.int8(-1), T.int8(-1), 0, 0, 0, 0, "NONE", 0, 0, "TFL", "ZEROS", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 4, 8, 3, 4, 0, 8, placeholder[96], 0, 0, 0, T.float32(0.5), 10, "NHWC", 24, 3, 1, "int8", 8, 16, 32, 8, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 512, 32, 1, 1, 1, 1, 1, 1, 1, buffer[0], 160, T.int8(-1), T.int8(-1), 12, buffer_1[0], 320, T.int8(-1), T.int8(-1), 0, 0, 0, 0, "NONE", 0, 0, "TFL", "ZEROS", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 8, 16, 32, 8, 0, 16, ethosu_write_1[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 512, 32, 1, "int8", 16, 32, 8, 16, 0, 32, ethosu_write[4096], 0, 0, 0, T.float32(0.25), 14, "NHWC", 256, 8, 1, 1, 1, 1, 1, 1, 1, buffer_2[0], 304, T.int8(-1), T.int8(-1), 12, buffer_3[0], 80, T.int8(-1), T.int8(-1), 0, 0, 0, 0, "NONE", 0, 0, "TFL", "ZEROS", 0, 0, 0, dtype="handle"))
def main(placeholder_5: T.Buffer[(1024,), "int8"], ethosu_write_1: T.Buffer[(2048,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([1456], "uint8")
buffer_1 = T.buffer_decl([352], "uint8")
buffer_2 = T.buffer_decl([272], "uint8")
buffer_3 = T.buffer_decl([11040], "uint8")
T.preflattened_buffer(placeholder_5, [1, 8, 1, 8, 16], 'int8', data=placeholder_5.data)
T.preflattened_buffer(ethosu_write_1, [1, 8, 2, 8, 16], 'int8', data=ethosu_write_1.data)
# body
ethosu_write_2 = T.allocate([2304], "int8", "global", annotations={"disable_lower_builtin": True})
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 6, 8, 3, 6, 0, 8, placeholder_5[0], 0, 0, 0, T.float32(0.5), 10, "NHCWB16", 128, 16, 1, "int8", 5, 8, 35, 5, 0, 8, ethosu_write_2[384], 0, 0, 0, T.float32(0.25), 14, "NHCWB16", 384, 16, 128, 3, 3, 1, 1, 1, 1, buffer[0], 1456, T.int8(-1), T.int8(-1), 12, buffer_1[0], 352, T.int8(-1), T.int8(-1), 1, 1, 0, 1, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 5, 8, 35, 5, 0, 8, ethosu_write_2[384], 0, 0, 0, T.float32(0.5), 10, "NHCWB16", 384, 16, 128, "int8", 4, 8, 26, 4, 0, 8, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHCWB16", 256, 16, 128, 3, 3, 1, 1, 1, 1, buffer_3[0], 11040, T.int8(-1), T.int8(-1), 12, buffer_2[0], 272, T.int8(-1), T.int8(-1), 1, 1, 0, 1, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 6, 8, 3, 6, 0, 8, placeholder_5[256], 0, 0, 0, T.float32(0.5), 10, "NHCWB16", 128, 16, 1, "int8", 5, 8, 35, 5, 0, 8, ethosu_write_2[0], 0, 0, 0, T.float32(0.25), 14, "NHCWB16", 384, 16, 128, 3, 3, 1, 1, 1, 1, buffer[0], 1456, T.int8(-1), T.int8(-1), 12, buffer_1[0], 352, T.int8(-1), T.int8(-1), 0, 1, 1, 1, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 5, 8, 35, 5, 0, 8, ethosu_write_2[0], 0, 0, 0, T.float32(0.5), 10, "NHCWB16", 384, 16, 128, "int8", 4, 8, 26, 4, 0, 8, ethosu_write_1[1024], 0, 0, 0, T.float32(0.25), 14, "NHCWB16", 256, 16, 128, 3, 3, 1, 1, 1, 1, buffer_3[0], 11040, T.int8(-1), T.int8(-1), 12, buffer_2[0], 272, T.int8(-1), T.int8(-1), 0, 1, 1, 1, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
def main(placeholder_5: T.Buffer[(768,), "int8"], ethosu_write_1: T.Buffer[(640,), "int8"]) -> None:
# function attr dict
T.func_attr({"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True})
buffer = T.buffer_decl([1744], "uint8")
buffer_1 = T.buffer_decl([80], "uint8")
buffer_2 = T.buffer_decl([320], "uint8")
buffer_3 = T.buffer_decl([880], "uint8")
T.preflattened_buffer(placeholder_5, [1, 16, 16, 3], 'int8', data=placeholder_5.data)
T.preflattened_buffer(ethosu_write_1, [1, 20, 4, 8], 'int8', data=ethosu_write_1.data)
# body
ethosu_write_2 = T.allocate([2560], "int8", "global", annotations={"disable_lower_builtin": True})
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 8, 16, 3, 8, 0, 16, placeholder_5[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 48, 3, 1, "int8", 8, 8, 32, 8, 0, 8, ethosu_write_2[512], 0, 0, 0, T.float32(0.25), 14, "NHWC", 256, 32, 1, 2, 3, 2, 1, 2, 1, buffer_3[0], 880, T.int8(-1), T.int8(-1), 12, buffer_2[0], 320, T.int8(-1), T.int8(-1), 2, 1, 0, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 8, 8, 32, 8, 0, 8, ethosu_write_2[512], 0, 0, 0, T.float32(0.5), 10, "NHWC", 256, 32, 1, "int8", 8, 4, 8, 8, 0, 4, ethosu_write_1[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 32, 8, 1, 2, 3, 2, 1, 2, 1, buffer[0], 1744, T.int8(-1), T.int8(-1), 12, buffer_1[0], 80, T.int8(-1), T.int8(-1), 2, 1, 0, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 12, 16, 3, 12, 0, 16, placeholder_5[192], 0, 0, 0, T.float32(0.5), 10, "NHWC", 48, 3, 1, "int8", 10, 8, 32, 10, 0, 8, ethosu_write_2[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 256, 32, 1, 2, 3, 2, 1, 2, 1, buffer_3[0], 880, T.int8(-1), T.int8(-1), 12, buffer_2[0], 320, T.int8(-1), T.int8(-1), 0, 1, 0, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 10, 8, 32, 10, 0, 8, ethosu_write_2[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 256, 32, 1, "int8", 8, 4, 8, 8, 0, 4, ethosu_write_1[256], 0, 0, 0, T.float32(0.25), 14, "NHWC", 32, 8, 1, 2, 3, 2, 1, 2, 1, buffer[0], 1744, T.int8(-1), T.int8(-1), 12, buffer_1[0], 80, T.int8(-1), T.int8(-1), 0, 1, 0, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 4, 16, 3, 4, 0, 16, placeholder_5[576], 0, 0, 0, T.float32(0.5), 10, "NHWC", 48, 3, 1, "int8", 4, 8, 32, 4, 0, 8, ethosu_write_2[0], 0, 0, 0, T.float32(0.25), 14, "NHWC", 256, 32, 1, 2, 3, 2, 1, 2, 1, buffer_3[0], 880, T.int8(-1), T.int8(-1), 12, buffer_2[0], 320, T.int8(-1), T.int8(-1), 0, 1, 2, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
T.evaluate(T.call_extern("ethosu_conv2d", "int8", 4, 8, 32, 4, 0, 8, ethosu_write_2[0], 0, 0, 0, T.float32(0.5), 10, "NHWC", 256, 32, 1, "int8", 4, 4, 8, 4, 0, 4, ethosu_write_1[512], 0, 0, 0, T.float32(0.25), 14, "NHWC", 32, 8, 1, 2, 3, 2, 1, 2, 1, buffer[0], 1744, T.int8(-1), T.int8(-1), 12, buffer_1[0], 80, T.int8(-1), T.int8(-1), 0, 1, 2, 0, "NONE", 0, 0, "TFL", "NONE", 0, 0, 0, dtype="handle"))
