K1504296 commented on issue #7799:
URL: https://github.com/apache/tvm/issues/7799#issuecomment-822023089
It seems to be a problem related to LLVM. LLVM 6.0 returns a Segmentation
Fault; however, LLVM 10.0 gets stuck during execution. Running with gdb and
LLVM 10.0 the following is returned after interrupting the execution:
```
Thread 1 "python3" received signal SIGINT, Interrupt.
0x0000ffff90fd60e8 in ?? () from /usr/lib/llvm-10/lib/libLLVM-10.so.1
```
The segmentation fault occurs when trying to get the assembly code of the
TVM model via : lib.get_source("asm") with an oc value of 32.
Here is a minimal example which still causes the segmentation fault on
aarch64:
```
import numpy as np
import tvm
from tvm import relay
from tvm.relay import transform
def _compile(ic, oc, target, data_layout, kernel_layout, dtypes):
input_dtype, weight_dtype, output_dtype = dtypes
n, h, w, ch, cw = 1, 64, 64, 3, 3
data_shape = (n, ic, h, w)
x = relay.var("x", relay.TensorType(data_shape, input_dtype))
kernel_shape = (oc, ic, ch, cw)
weight = relay.var("weight", relay.TensorType(kernel_shape,
weight_dtype))
y = relay.nn.conv2d(
x,
weight,
kernel_size=(ch, cw),
channels=oc,
padding=(1, 1),
dilation=(1, 1),
data_layout=data_layout,
kernel_layout=kernel_layout,
out_dtype=output_dtype,
)
func = relay.Function([x, weight], y)
wdata = np.random.rand(*kernel_shape) * 10
parameters = {"weight": tvm.nd.array(wdata.astype(weight_dtype))}
with tvm.transform.PassContext(opt_level=3):
graph, lib, params = relay.build(func, target, params=parameters)
assembly = lib.get_source("asm")
return assembly
target = "llvm -mcpu=core-avx2"
fast_int8_dtypes = ("uint8", "int8", "int32")
asm = _compile(
ic=16,
oc=32,
target=target,
data_layout="NCHW",
kernel_layout="OIHW",
dtypes=fast_int8_dtypes,
)
```
--
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.
For queries about this service, please contact Infrastructure at:
[email protected]
