Hi all:
Recently I met infer bound error on split op:
TVMError: Check failed: match: iter_var(blockIdx.x, , blockIdx.x) domain already inferred, cannot prove their extents are the same floordiv(((({any_dim|any_dim>=0}*{any_dim|any_dim>=0})*({any_dim|any_dim>=0} - (floordiv({any_dim|any_dim>=0}, 4)*3))) + 511), 512) vs floordiv(((({any_dim|any_dim>=0}*{any_dim|any_dim>=0})*floordiv({any_dim|any_dim>=0}, 4)) + 511), 512)
Error during compile function
And this error can be reproduced by this code snippet:
import tvm
from tvm import relay
import numpy as np
def test_split():
# input var
data = relay.var("data", tvm.ir.TensorType(shape = (relay.Any(), relay.Any(), relay.Any()), dtype = "float32"))
data_1 = relay.var("data_1", tvm.ir.TensorType(shape = (relay.Any(), relay.Any(), relay.Any()), dtype = "float32"))
split_out = relay.op.split(data, 4, 2)
out = tvm.relay.op.multiply(split_out[0], data_1)
out_1 = tvm.relay.op.multiply(split_out[1], data_1)
mod = tvm.IRModule()
mod["main"] = relay.Function([data, data_1], tvm.relay.expr.Tuple([out, out_1]))
# tvm setting
target = tvm.target.Target("cuda -libs=cublas,cudnn")
ctx = tvm.gpu(0)
ex = relay.create_executor("vm", mod=mod, ctx=ctx, target=target)
# inputs data
data_np = np.random.randint(256, size=(1,1,256)).astype(np.float32)
data_1_np = np.random.randint(64, size=(1,1,64)).astype(np.float32)
result_tvm = ex.evaluate()(data=data_np, data_1=data_1_np)
print("=======tvm result=========")
print(result_tvm)
if __name__ == "__main__":
test_split()
It seems that the itervar was inferred twice, but the range->extent can not be proved the same.
Could anyone give me some advices on this issue, why tvm cannot correctly infer bound on this case?
Best regards!
