Any luck building the autoscheduler tutorial with the latest Halide

[tekstop]

After a few iterations, I was able to build tiramisu with the latest Halide and get the tutorials working. Now, when I try to get the autoscheduler tutorial working, I get the following errors:

% ../generator
{ conv[n, fout, y, x, fin, k_y, k_x] -> src[n', fin', y_pad, x_pad] : -y + y_pad = 0 }
{ conv[n, fout, y, x, fin, k_y, k_x] -> src[n', fin', y_pad, x_pad] : -y - k_y + y_pad = 0 }
{ conv[n, fout, y, x, fin, k_y, k_x] -> src[n', fin', y_pad, x_pad] : -x + x_pad = 0 }
{ conv[n, fout, y, x, fin, k_y, k_x] -> src[n', fin', y_pad, x_pad] : -x - k_x + x_pad = 0 }

Generated Halide IR:
assert(reinterpret<uint64>((halide_buffer_t *)buf_weights.buffer) != (uint64)0, halide_error_buffer_argument_is_null("buf_weights"))
assert(reinterpret<uint64>((halide_buffer_t *)buf_src.buffer) != (uint64)0, halide_error_buffer_argument_is_null("buf_src"))
assert(reinterpret<uint64>((halide_buffer_t *)buf_output.buffer) != (uint64)0, halide_error_buffer_argument_is_null("buf_output"))
assert(reinterpret<uint64>((halide_buffer_t *)buf_bias.buffer) != (uint64)0, halide_error_buffer_argument_is_null("buf_bias"))
let buf_bias = (void *)_halide_buffer_get_host((halide_buffer_t *)buf_bias.buffer)
let buf_output = (void *)_halide_buffer_get_host((halide_buffer_t *)buf_output.buffer)
let buf_src = (void *)_halide_buffer_get_host((halide_buffer_t *)buf_src.buffer)
let buf_weights = (void *)_halide_buffer_get_host((halide_buffer_t *)buf_weights.buffer)
produce  {
 allocate _weights_b2[int32 * 1024 * 1024 * 2 * 8]
 allocate _src_b1[int32 * 1026 * 1026 * 3 * 8]
 allocate _conv_init_b3[int32 * 1024 * 1024 * 2 * 8]
 allocate _conv_b4[int32 * 3 * 3 * 3 * 1024 * 1024 * 2 * 8]
 allocate _bias_b0[int32 * 2]
 for (c1, 0, 8) {
  for (c3, 0, 2) {
   for (c5, 0, 1024) {
    for (c7, 0, 1024) {
     buf_output[(c1*2097152) + ((c3*1048576) + ((c5*1024) + c7))] = buf_bias[c3]
     for (c9, 0, 3) {
      for (c11, 0, 3) {
       for (c13, 0, 3) {
        buf_output[(c1*2097152) + ((c3*1048576) + ((c5*1024) + c7))] = buf_output[(c1*2097152) + ((c3*1048576) + ((c5*1024) + c7))] + (buf_src[(c1*3158028) + ((c9*1052676) + (((c11 + c5)*1026) + (c13 + c7)))]*buf_weights[(c3*27) + ((c9*9) + ((c11*3) + c13))])
       }
      }
     }
    }
   }
  }
 }
}
Traceback (most recent call last):
  File "/local/tiramisu/tiramisu/tutorials/tutorial_autoscheduler/model/main.py", line 39, in <module>
    tree_tensor = get_representation(prog_json, sched_json)
  File "/local/tiramisu/tiramisu/tutorials/tutorial_autoscheduler/model/json_to_tensor.py", line 54, in get_representation
    iterators_repr.append(+(iterator_name in comp_dict['real_dimensions']))
KeyError: 'real_dimensions'
Segmentation fault

Note, without the autoscheduler boolean set to true, I am able to generate as well as compile and execute with the wrapper.

Python 3.9.2 Torch version: 1.13.0+cpu

Any pointers would be helpful

[Baxi-codes]

I am facing the same issue. Even if we comment the real dimensions part, it would say mat1, mat2 shapes are not compatible to multiply (in forward method of the model). The tutorial is very old on this branch, but apparently it was being maintained on the dev branch, but couldn't build it, too many incompatibilities. If someone has a way to build a version of the autoscheduler, please enlighten me!