feat: fix mulmat

[chraac]

Summary

In latest dev-refactoring, the test-backend-ops runs well on ADD operator, but failed on MUL_MAT.
After investigation, we found that the input tensor should be transposed (see: 63dc587).

Here's how we fix it:

1. Add tensor types INTERMEDIATE and PARAMETER. These types are graph-private and invisible to other graphs.
   1. The INTERMEDIATE tensor is used for connecting multiple operations within a single graph.
   2. The PARAMETER tensor is used for setting the parameters of certain operators, such as transpose.
2. Add ggml_qnn_matmul_op_config class, for creating necessary nodes to matmul graph
   1. At function ggml_qnn_matmul_op_config::create_tensors, we will add several nodes, and finally make a graph like this:

    graph TD;
         i1>input_tensor_a] --src0--> mat_mul0;
         i2>input_tensor_b] --src1--> transpose0;
         transpose0 --intermediate0--> mat_mul0;
         mat_mul0 --intermediate1--> transpose1;
         transpose1 --dst0--> o1>output_tensor_c];
   

   Loading

Log

Backend 4/4: CPU
  Skipping CPU backend
4/4 backends passed
�[1;32mOK�[0m

test-backend-ops_MUL_MAT.log

Changes

• Modify the ggml_qnn_op_config, to manage the rensors, and support multi-op
• Add transpose op before mulmat

Self-check

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[chraac]

Here we add nodes to the mat_mul graph, after this function ,the graph will be like:

 graph TD;
      i1>input_tensor_a] --src0--> mat_mul0;
      i2>input_tensor_b] --src1--> transpose0;
      transpose0 --intermediate0--> mat_mul0;
      mat_mul0 --intermediate1--> transpose1;
      transpose1 --dst0--> o1>output_tensor_c];

Loading