Auto TVM's Pre-tuned parameters


#1

Hi, in the auto-tuning tutorial on the TVM website it mentions that the pre-tuned parameters have been released for some arm devices. Where can i find these parameters. I’m using a JetsonTX2 board and have run into the following bottleneck

WARNING:autotvm:Cannot find config for target=cuda -model=tx2, workload=('conv2d', (1, 1024, 11, 20, 'float32'), (256, 1024, 1, 1, 'float32'), (1, 1), (0, 0), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
WARNING:autotvm:Cannot find config for target=cuda -model=tx2, workload=('conv2d', (1, 512, 11, 20, 'float32'), (1024, 512, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
WARNING:autotvm:Cannot find config for target=cuda -model=tx2, workload=('conv2d', (1, 128, 22, 40, 'float32'), (256, 128, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
WARNING:autotvm:Cannot find config for target=cuda -model=tx2, workload=('conv2d', (1, 64, 44, 80, 'float32'), (128, 64, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
WARNING:autotvm:Cannot find config for target=cuda -model=tx2, workload=('conv2d', (1, 32, 88, 160, 'float32'), (64, 32, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
WARNING:autotvm:Cannot find config for target=cuda -model=tx2, workload=('conv2d', (1, 16, 176, 320, 'float32'), (32, 16, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.
WARNING:autotvm:Cannot find config for target=cuda -model=tx2, workload=('conv2d', (1, 3, 352, 640, 'float32'), (16, 3, 3, 3, 'float32'), (1, 1), (1, 1), (1, 1), 'NCHW', 'float32'). A fallback configuration is used, which may bring great performance regression.

#2

Have you checked reproducer, if this is working for you?


#3

Ive tried running the benchmark file but that just errors out. I have however got the tuning to work so ive tuned the model but the main runtime still cant find the log file it generated. I cant seem to find the documentation on how to load the log file for the tuned parameters