comaniac opened a new pull request #7317:
URL: https://github.com/apache/tvm/pull/7317
In this PR, we attempt to enable schedule sharing as a workaround before the
dynamic shape support is fully landed. The idea is that if we have a schedule
for batch size 1, then it is actually applicable to all other batch sizes
(regardless the performance). This is useful when we only tune the workload
with batch size 1 but wish to use it for all batch sizes to at least make the
flow working.
To do so, we introduce "workload distance factor", which indicates the
similarity of two workloads. Specifically, it is calculated by the following
rules:
- If two workloads are not for the same compute DAG or function, then `inf`.
- If two workloads are for the same compute DAG/function, and
- their non-zero integer arguments are dividable and their zero and
non-integer arguments are the same, then `factor=prod(a / b) for a, b in
zip(wkl1.args, wkl2.args)`,
- otherwise `inf`.
As a result, the distance factor ranges from 1 to `inf`. When the distance
factor is not `inf`, meaning that it is safe to apply the schedule of workload
2 to workload 1.
The above mechanism works well for registered TE computes but not the
ComputeDAG extracted from Relay programs. This is because currently when
extracting tasks from Relay, we use MD5 to hash the ComputeDAG serialized
string to be its key, which includes not only the DAG structure but the shapes,
so it's impossible to calculate the distance factor. To make it work, this PR
also improves the hashing mechanism of ComputeDAG by separating the
input/output tensor shapes so that they can be accessed. For example, the
workload key of a ComputeDAG was:
```
["8d5a93959138dc7b2ee1f1b3219dfa14"]
```
and it now becomes:
```
["ad6cecbf5d85cb1cda3c2bb7af170211", 1, 7, 7, 512, 4, 4, 512, 512, 1, 7, 7,
512, 1, 1, 1, 512, 1, 1, 1, 512, 1, 7, 7, 512]
```
Please note that since we change the workload key format of ComputeDAG, the
tuning logs won't match anymore. To make it work again, we can use the
following script to update the keys in existing log files. This is also the way
I used to update the CI logs:
```python
import json
import hashlib
import os
import sys
from tvm.te import ComputeOp, PlaceholderOp
from tvm.auto_scheduler import save_records
from tvm.auto_scheduler.measure import MeasureInput
from tvm.auto_scheduler.measure_record import load_records
from tvm.auto_scheduler.utils import get_const_tuple
tasks = [] # Extract tasks from a Relay program
log_file = "old-log-file"
new_log_file = "new-log-file"
def get_old_hash_key(dag):
"""Return the hash key of a compute DAG."""
str_key = ""
for op in dag.ops:
t = op.output(0)
if isinstance(op, PlaceholderOp):
str_key += "placeholder,"
str_key += str(get_const_tuple(t.shape)) + ","
str_key += t.dtype + ";"
elif isinstance(op, ComputeOp):
str_key += str(t.op.body) + ","
str_key += str(get_const_tuple(t.shape)) + ","
str_key += t.dtype + ";"
else:
raise ValueError("Invalid op: " + op)
str_key = str_key.encode(encoding="utf-8")
return hashlib.md5(str_key).hexdigest()
# Establish the key mapping
old_key_to_task = {}
hit_count = {}
for idx, task in enumerate(tasks):
old_key = json.dumps((get_old_hash_key(task.compute_dag),))
old_key_to_task[old_key] = task
hit_count[old_key] = 0
print("Task %d %s -> %s" % (idx, old_key, task.workload_key))
# Update the workload key in an existing log file
new_inputs = []
new_results = []
for inp, res in load_records(log_file):
if inp.task.workload_key not in old_key_to_task:
print(
"Ignore key %s in log file due to no corresponding task found" %
inp.task.workload_key
)
continue
hit_count[inp.task.workload_key] += 1
new_inputs.append(MeasureInput(old_key_to_task[inp.task.workload_key],
inp.state))
new_results.append(res)
for key, cnt in hit_count.items():
print("Old key %s hits %d times" % (key, cnt))
if os.path.exists(new_log_file):
os.remove(new_log_file)
save_records(new_log_file, new_inputs, new_results)
```
cc @merrymercy @jcf94
----------------------------------------------------------------
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]
