Relay VM executor performance regression for mlp model

Questions
1

I’m seeing 50X performance regression between the relay.build_module.build(mod, target, params=params) and relay.build_module.create_executor(‘vm’, mod, ctx, target).
Here is the script to reproduce in mlp instance:
import numpy as np

from tvm import relay
from tvm.relay import testing
import tvm

batch_size = 1
num_class = 10
image_shape=(1, 28, 28) # for mlp
data_shape = (batch_size,) + image_shape
out_shape = (batch_size, num_class)

mod, params = relay.testing.mlp.get_workload(
    batch_size=batch_size, image_shape=image_shape)

opt_level = 3
target = "llvm"
ctx = tvm.cpu()
with relay.build_config(opt_level=opt_level):
    executor = relay.build_module.create_executor('graph', mod, ctx, target)

dtype = 'float32'
data = tvm.nd.array(np.random.uniform(-1, 1, size=data_shape).astype("float32"))
tvm_out = executor.evaluate()(data, **params)
out =tvm_out.asnumpy()
print(out.flatten()[0:10])
test_times = 100
import time

start = time.time()
for i in range(test_times):
    executor.evaluate()(data, **params)
end = time.time()
print("Mean inference time", (end - start) / test_times * 1000)

##########################################################

 import numpy as np
from tvm import relay
from tvm.relay import testing
import tvm
batch_size = 1
num_class = 10
   image_shape=(1, 28, 28) # for mlp
   data_shape = (batch_size,) + image_shape
     out_shape = (batch_size, num_class)

    mod, params = relay.testing.mlp.get_workload(
        batch_size=batch_size, image_shape=image_shape)

    print(mod.astext(show_meta_data=False))

    opt_level = 3
    target = "llvm"
    ctx = tvm.cpu()
    with relay.build_config(opt_level=opt_level):
        executor = relay.build_module.create_executor('vm', mod, ctx, target)

    dtype = "float32"
    data = tvm.nd.array(np.random.uniform(-1, 1, size=data_shape).astype("float32"))
    tvm_out = executor.evaluate()(data, **params)
    out = tvm_out.asnumpy()

    test_times = 100
    import time

    start = time.time()
    for i in range(test_times):
        executor.evaluate()(data, **params)
    end = time.time()
    print("Mean inference time", (end - start) / test_times * 1000)

@ wweic