[MetaSchedule] Logging Interface Unification

[zxybazh]

This PR introduces a new interface to unify meta schedule logging from both c++ and python side. It will be easier to control logging level and format content. From now on, only task scheduler level will be printed to screen by default, and task level information will be logged to a file in given folder. We still support any logging configuration from outside, and use tvm.meta_schedule as the main logger.

The interface uses logger on the python side while on the c++ side we created a new macro called TVM_PY_LOG to work int the following style where logging_func is the optional packed function for logging (thanks to #10032 ):

TVM_PY_LOG(DEBUG, logging_func)
TVM_PY_LOG(INFO, logging_func)
TVM_PY_LOG(WARNING, logging_func)
TVM_PY_LOG(ERROR, logging_func)

The logger would fall back to original c++ logger function LOG(Level) if given logging function is not defined.

Recomended setup of logger is as follows:

logging.getLogger("tvm.meta_schedule").setLevel(logging.INFO)
# Use logging.DEBUG if debugging

And further configuration of task level loggers can be set in TuneConfig for features like whether to stream to screen, whether to propagate to main logger, patterns, etc, by passing a Dict to TuneConfig as follows:

TuneConfig(
    ...,
    logger_config = {
        "loggers": ...,
        "handlers": ...,
        "formatters": ...
    }
    # None if use default setting.
    # The strings can be parameterized using any parameter given in create_loggers function
    # for example, `{logger_name}` and `{log_dir}` can be used for parameterization in tune_extracted_tasks function.
)

Default setting is to output to files in log_dir folder and the logging structure looks like the following directory:

tvm.meta_schedule.tune.task_00_fused_nn_conv2d_add.log
tvm.meta_schedule.tune.task_01_fused_nn_conv2d_add_1.log
tvm.meta_schedule.tune.task_02_fused_nn_conv2d_add_2.log
tvm.meta_schedule.tune.task_03_fused_layout_transform.log
tvm.meta_schedule.tune.task_04_fused_nn_conv2d_add_nn_relu.log
tvm.meta_schedule.tune.task_05_fused_nn_max_pool2d.log
tvm.meta_schedule.tune.task_06_fused_nn_conv2d_add_nn_relu_1.log
tvm.meta_schedule.tune.task_07_fused_nn_conv2d_add_add_nn_relu.log
tvm.meta_schedule.tune.task_08_fused_nn_conv2d_add_nn_relu_2.log
tvm.meta_schedule.tune.task_09_fused_nn_contrib_conv2d_winograd_without_weight_transform_add_nn_relu.log
tvm.meta_schedule.tune.task_10_fused_nn_contrib_conv2d_winograd_without_weight_transform_add_add_nn_relu.log
tvm.meta_schedule.tune.task_11_fused_nn_conv2d_add_nn_relu_3.log
tvm.meta_schedule.tune.task_12_fused_nn_contrib_conv2d_winograd_without_weight_transform_add_nn_relu_1.log
tvm.meta_schedule.tune.task_13_fused_nn_contrib_conv2d_winograd_without_weight_transform_add_add_nn_relu_1.log
tvm.meta_schedule.tune.task_14_fused_nn_conv2d_add_nn_relu_4.log
tvm.meta_schedule.tune.task_15_fused_nn_conv2d_add_nn_relu_5.log
tvm.meta_schedule.tune.task_16_fused_nn_conv2d_add_add_nn_relu_1.log
tvm.meta_schedule.tune.task_17_fused_nn_adaptive_avg_pool2d.log
tvm.meta_schedule.tune.task_18_fused_layout_transform_reshape_squeeze.log
tvm.meta_schedule.tune.task_19_fused_nn_dense_add.log
tvm.meta_schedule.tune.task_scheduler.log

Suggestions are welcome!

CC: @junrushao1994 @shingjan @comaniac

[junrushao]

Thanks for massively enhancing the current logging system!

However, on the implementation, we do need to follow the standard best practice in python, namely, using:

logger = logging.getLogger(__name__)

to get a logger for this file.

[zxybazh]

Thanks for the careful review, I've fixed all the logger usage and made sure it follows the current recommended practice. Ready for review.

[junrushao]

Pathlib is definitely a good library, but perhaps it's less plausible if we mix os.path and Pathlib. In our particular case, shall we just do:

os.makedirs(path, exist_ok=False)

[zxybazh]

I think it might be acceptable to use exist_ok = True in this case cause the logs folder could be reused.

[junrushao]

I was thinking if it's better to make it a LOG(FATAL) instead, but it's do that in a separate PR

[junrushao]

Any updates?

[junrushao]

LGTM! Please try it out for the last time and confirm it prints the dates properly, and then feel free to merge it yourself

[zxybazh]

Thanks @junrushao1994 @Hzfengsy for reviewing!