[Bug][Auto-scheduler] generated cuda code

[Haoran-Solar]

To be simple, the cuda code that generated by auto-scheduler from file

/tutorials/auto_scheduler/tune_conv2d_layer_cuda.py

I call it as following:

dim3 dimBlock(128);
dim3 dimGrid(16);

default_function_kernel0<<<dimGrid, dimBlock>>>(data, kernel, compute, bias);

cudaDeviceSynchronize();

Then bug occurred :

kernel ruined cuda api? or something
If I do not run the kernel, nothing wrong.

I cannot figure out what happened, can you help me?
P4 and T4, I used, with cuda-10.2

[comaniac]

How many trials did you try and how did you apply the schedule from the log file? It's possible for auto-scheduler to generate invalid kernels during the tuning, but they shouldn't be picked ultimately when compiling the model.

[Haoran-Solar]

Thanks for reply. I just started to try auto-scheduler, I change nothing in file tune_conv2d, just want to have a test.
Using prams from last layer in ResNet-50, just the tutorials did.
And I solve the problem. To run the generated code, I have to set the number of each dim less than 64 in a threadblock.
Which mean:

dim3 dimBlock(number);
*number<64*
*64, 32,16 ....*

Is that some special limit for auto-scheduler code? Cause some kernels that I wrote do not such limit.

[comaniac]

You are not supposed to run the generated kernel by your own, so this is not a limitation to me. Since this is not a bug, I'm going to close this issue. For future questions, please post them to https://discuss.tvm.apache.org/