zxybazh opened a new issue, #10900:
URL: https://github.com/apache/tvm/issues/10900
In the following script `compute_inline` accidentally accepts invalid pure
spatial block with init block inside, causing undefined reference to a tensor
created by `buffer_decl`.
```
from tvm.tir import Schedule
import tvm
from tvm.script import tir as T
# pylint:
disable=invalid-name,no-member,line-too-long,too-many-nested-blocks,no-self-argument
# fmt: off
# from tvm.script import tir as T
@tvm.script.ir_module
class Module:
@T.prim_func
def main(placeholder: T.Buffer[(1, 512, 7, 7), "float32"], tensor:
T.Buffer[(1, 512, 1, 1), "float32"]) -> None:
# function attr dict
T.func_attr({"global_symbol": "main", "tir.noalias": True})
# body
with T.block("root"):
T.reads()
T.writes()
T.block_attr({"meta_schedule.parallel":64,
"meta_schedule.unroll_explicit":0, "meta_schedule.vectorize":64})
tensor_1 = T.alloc_buffer([1, 512, 1, 1], dtype="float32")
tensor_1_rf = T.alloc_buffer([1, 512, 1, 1, 49], dtype="float32")
for i0, i1, i2, i3, i4_i5_fused_0_i4_i5_fused_1_fused_0,
i4_i5_fused_0_i4_i5_fused_1_fused_1 in T.grid(1, 512, 1, 1, 49, 1):
with T.block("tensor_rf"):
vi4_i5_fused_0_i4_i5_fused_1_fused_0 =
T.axis.spatial(49, i4_i5_fused_0_i4_i5_fused_1_fused_0)
ax0 = T.axis.spatial(1, 0)
ax1 = T.axis.spatial(512, i1)
ax2 = T.axis.spatial(1, 0)
ax3 = T.axis.spatial(1, 0)
T.reads(placeholder[ax0, ax1, ax2 * 7 +
vi4_i5_fused_0_i4_i5_fused_1_fused_0 // 7, ax3 * 7 +
vi4_i5_fused_0_i4_i5_fused_1_fused_0 % 7])
T.writes(tensor_1_rf[ax0, ax1, ax2, ax3,
vi4_i5_fused_0_i4_i5_fused_1_fused_0])
with T.init():
tensor_1_rf[ax0, ax1, ax2, ax3,
vi4_i5_fused_0_i4_i5_fused_1_fused_0] = T.float32(0)
tensor_1_rf[ax0, ax1, ax2, ax3,
vi4_i5_fused_0_i4_i5_fused_1_fused_0] = tensor_1_rf[ax0, ax1, ax2, ax3,
vi4_i5_fused_0_i4_i5_fused_1_fused_0] + placeholder[ax0, ax1, ax2 * 7 +
vi4_i5_fused_0_i4_i5_fused_1_fused_0 // 7, ax3 * 7 +
vi4_i5_fused_0_i4_i5_fused_1_fused_0 % 7]
for i0, i1 in T.grid(1, 512):
for ax0, ax1, ax2, ax3, ax4 in T.grid(49, 1, 1, 1, 1):
with T.block("tensor"):
vi4_i5_fused_0_i4_i5_fused_1_fused_0, ax0_1 =
T.axis.remap("RS", [ax0, ax1])
ax1_1 = T.axis.spatial(512, i1 + ax2)
ax2_1, ax3_1 = T.axis.remap("SS", [ax3, ax4])
T.reads(tensor_1_rf[ax0_1, ax1_1, ax2_1, ax3_1,
vi4_i5_fused_0_i4_i5_fused_1_fused_0])
T.writes(tensor_1[ax0_1, ax1_1, ax2_1, ax3_1])
with T.init():
tensor_1[ax0_1, ax1_1, ax2_1, ax3_1] =
T.float32(0)
tensor_1[ax0_1, ax1_1, ax2_1, ax3_1] =
tensor_1[ax0_1, ax1_1, ax2_1, ax3_1] + tensor_1_rf[ax0_1, ax1_1, ax2_1, ax3_1,
vi4_i5_fused_0_i4_i5_fused_1_fused_0]
for i2, i3 in T.grid(1, 1):
with T.block("tensor_1"):
ax0, ax1, ax2, ax3 = T.axis.remap("SSSS", [i0, i1,
i2, i3])
T.reads(tensor_1[ax0, ax1, ax2, ax3])
T.writes(tensor[ax0, ax1, ax2, ax3])
tensor[ax0, ax1, ax2, ax3] = tensor_1[ax0, ax1, ax2,
ax3] * T.float32(0.020408163265306121)
# fmt: on
# pylint:
enable=invalid-name,no-member,line-too-long,too-many-nested-blocks,no-self-argument
sch = Schedule(Module)
print(sch.mod.script())
b15 = sch.get_block(name="tensor_rf", func_name="main")
sch.compute_inline(block=b15)
print(sch.mod.script())
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
