cxx122 opened a new issue, #15205:
URL: https://github.com/apache/tvm/issues/15205
When I sum a constant value but with out of bound acess, the result of
noschedule and scheduled programs will be different. It may be related to the
optimization of 'tir.const'.
### Expected behavior
Same results.
### Actual behavior
Different results.
```
Not equal to tolerance rtol=1e-05, atol=1e-07
x and y nan location mismatch:
x: array(nan, dtype=float32)
y: array(-2177.8052, dtype=float32)
@main = primfn(B_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol": "main",
"tir.noalias": True}
buffers = {B: Buffer(B_2: Pointer(float32), float32, [1], [])}
buffer_map = {B_1: B}
preflattened_buffer_map = {B_1: B_3: Buffer(B_2, float32, [], [])} {
B[0] = 0f32
B[0] = (B[0] + -34.0282f32)
B[0] = (B[0] + -34.0282f32)
B[0] = (B[0] + -34.0282f32)
...
B[0] = (B[0] + -34.0282f32)
B[0] = (B[0] + -34.0282f32)
B[0] = (B[0] + -34.0282f32)
B[0] = (B[0] + -34.0282f32)
}
```
### Environment
Operating System: Ubuntu 18.04
TVM version: v0.10.dev0
### Steps to reproduce
```
import tvm
import json
import random
import numpy as np
from tvm import te
from tvm import tir
from tvm import testing
from tvm import auto_scheduler
from tvm.auto_scheduler.workload_registry import register_workload_tensors
POLICY_PARAMS = {
"eps_greedy": 0.05,
"retry_search_one_round_on_empty": 1,
"sample_init_min_population": 3,
"sample_init_use_measured_ratio": 0.2,
"evolutionary_search_population": 5,
"evolutionary_search_num_iters": 4,
"evolutionary_search_mutation_prob": 0.85,
"cpu_multi_level_tiling_structure": "SSRSRS",
"gpu_multi_level_tiling_structure": "SSSRRSRS",
# Notice: the default thread bind policy of GPU assumes the tiling
structure to have at
# least 3 spatial tiling levels in outermost
"max_innermost_split_factor": 64,
"max_vectorize_size": 16,
"disable_change_compute_location": 0,
}
def te_test():
inline_tensor = te.compute([3], lambda ax0 : tir.const(-3.40282e+1,
dtype="float32"), name='inline_tensor')
k = te.reduce_axis([0, 64], name='k')
B = te.compute([], lambda : te.sum(inline_tensor[k], axis=[k]),
name='B')
return [B]
# Get dag and print it.
tensors = te_test()
dag = auto_scheduler.ComputeDAG(tensors)
dict = json.loads(tvm.ir.save_json(tensors))
with open("./saved_json.txt", "w") as file:
file.write(tvm.ir.save_json(tensors))
print(dag)
# Get inputs.
inputs = []
for tensor in dag.tensors:
shape = [x.value if 'value' in dir(x) and isinstance(x.value, int) else
1 for x in tensor.shape]
inputs.append((2 * np.random.uniform(size=shape)+1).astype(tensor.dtype))
# Get program with no schedule.
results = []
mod_list = []
pre_schedule_list = dag.apply_steps_from_state(dag.get_init_state())
pre_mod = tvm.lower(pre_schedule_list[0], pre_schedule_list[1],
simple_mode=True)
mod_list.append(pre_mod)
with tvm.transform.PassContext(opt_level=0):
mod_exec = tvm.build(pre_mod)
print(pre_mod)
new_inputs = [tvm.nd.array(x, tvm.cpu()) for x in inputs.copy()]
mod_exec(*new_inputs)
result = []
for x in new_inputs:
try:
result.append(x.numpy() if isinstance(
x, tvm.runtime.NDArray) else x)
except (ValueError, tvm.TVMError):
result.append(None)
results.append(result)
# Get program with schedule.
register_workload_tensors(dag.workload_key(), tensors)
task = auto_scheduler.SearchTask(workload_key=dag.workload_key(),
target=tvm.target.Target("llvm"))
policy = auto_scheduler.SketchPolicy(task, verbose=0, params=POLICY_PARAMS)
states = policy.sample_initial_population()
for state in states:
schedule_list = dag.apply_steps_from_state(state)
print("Before Schedule")
mod = tvm.lower(schedule_list[0], schedule_list[1], simple_mode=True)
mod_list.append(mod)
print("After Schedule")
with tvm.transform.PassContext(opt_level=0):
mod_exec = tvm.build(mod)
new_inputs = [tvm.nd.array(x, tvm.cpu()) for x in inputs.copy()]
mod_exec(*new_inputs)
print("Got results")
result = []
for x in new_inputs:
try:
result.append(x.numpy() if isinstance(
x, tvm.runtime.NDArray) else x)
except (ValueError, tvm.TVMError):
result.append(None)
results.append(result)
for i in range(1, len(results)):
result = results[i]
for compare_idex in [-1]:
try:
tvm.testing.assert_allclose(results[0][compare_idex],
result[compare_idex], rtol=1e-5)
except AssertionError as e:
print(e)
print(mod_list[i])
break
```
--
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]
