Constants duplicated in generated C code

Troubleshooting
1

Hi, I am trying to generate C code from an imported onnx model. The weights and biases of my model (140 in total) are duplicated several times in the global_const_workspace:

static struct global_const_workspace {
      float fused_constant_9_let[140] __attribute__((aligned(16))); // 560 bytes, aligned offset: 0
      float fused_constant_8_let[140] __attribute__((packed, aligned(16))); // 560 bytes, aligned offset: 560
      float fused_constant_7_let[140] __attribute__((packed, aligned(16))); // 560 bytes, aligned offset: 1120
      float fused_constant_6_let[140] __attribute__((packed, aligned(16))); // 560 bytes, aligned offset: 1680

How can I ensure that constants are not copied?

I called tvmc with the following parameters:

python -m tvm.driver.tvmc compile --target="c" --runtime=crt --executor=aot --executor-aot-interface-api=c --executor-aot-unpacked-api=1 --runtime-crt-system-lib 1 --pass-config tir.usmp.enable=1 --pass-config tir.disable_storage_rewrite=1 --pass-config tir.disable_vectorize=1 --input-shapes="seq:[25,2]" --model-format=onnx model.onnx --output-format=mlf

2

Maybe a bit more context will help: my model is a LSTM with an input sequence length of 25. Looking at the relay function, it seems that the model weights are already duplicated there, after unrolling the LSTM loop. There are exactly 25 references to a (20, 7) tensor which contains the weights (I assume), each one with a different index in the relay.Constant list. Each one of these seems to be allocated separately, so that the generated code ends up using much more storage than necessary.

Has anyone seen something similar when importing LSTM models? Is this to be expected? Any hints would be appreciated.

Here is a part of the relay function, where you can see the same constant being used in %6, %27 and %48:

def @main(%seq: Tensor[(25, 2), float32] /* ty=Tensor[(25, 2), float32] span=/Reshape.seq:0:0 */) -> 
   Tensor[(1), float32] {
%0 = reshape(%seq, newshape=[25, 1, -1]) /* ty=Tensor[(25, 1, 2), float32] span=/Reshape:0:0 */;
%1 = split(%0, indices_or_sections=25) /* ty=(Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], 
   Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), 
   float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 
   2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 
  1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], 
  Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), 
  float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 2), float32], Tensor[(1, 1, 
  2), float32]) span=/lstm/LSTM:0:0 */;
%2 = %1.0 /* ty=Tensor[(1, 1, 2), float32] span=/lstm/LSTM:0:0 */;
%3 = squeeze(%2, axis=[0]) /* ty=Tensor[(1, 2), float32] span=/lstm/LSTM:0:0 */;
%4 = (%3, meta[relay.Constant][0] /* ty=Tensor[(1, 5), float32] span=/lstm/LSTM:0:0 */) /* ty= 
  (Tensor[(1, 2), float32], Tensor[(1, 5), float32]) span=/lstm/LSTM:0:0 */;
%5 = concatenate(%4, axis=1) /* ty=Tensor[(1, 7), float32] span=/lstm/LSTM:0:0 */;
%6 = nn.dense(%5, meta[relay.Constant][1] /* ty=Tensor[(20, 7), float32] span=/lstm/LSTM:0:0 */, units=None) /* ty=Tensor[(1, 20), float32] span=/lstm/LSTM:0:0 */;
...
%27 = nn.dense(%26, meta[relay.Constant][5] /* ty=Tensor[(20, 7), float32] span=/lstm/LSTM:0:0 */, 
  units=None) /* ty=Tensor[(1, 20), float32] span=/lstm/LSTM:0:0 */;
...
%48 = nn.dense(%47, meta[relay.Constant][6] /* ty=Tensor[(20, 7), float32] span=/lstm/LSTM:0:0 */, 
  units=None) /* ty=Tensor[(1, 20), float32] span=/lstm/LSTM:0:0 */;
...