[P0] Enable compiler e2e test with arch spec

[tancheng]

We are able to compile/lower a C++ file via lit test. However, we do want to directly call neura-compiler:

dataflow/test/visualize/test2.mlir

Line 2 in 1c1360e

// RUN: neura-compiler --neura-conversion %s

rather than mlir-neura-opt:

dataflow/test/neura/for_loop/relu_test.mlir

Line 5 in 1c1360e

// RUN: mlir-neura-opt %t-relu.mlir\

For neura-compiler, we need to pass in architecture.yaml:

dataflow/test/neura/ctrl/branch_for.mlir

Line 59 in 1c1360e

// RUN: --architecture-spec=../../arch_spec/architecture.yaml \

Please check/track these code to make sure architecture.yaml can be recognized by the map-to-accelerator pass:

dataflow/lib/NeuraDialect/Transforms/MapToAcceleratorPass.cpp

Line 447 in 1c1360e

std::string architecture_spec_file = mlir::neura::getArchitectureSpecFile();

, and also be captured by the GenerateCodePass.

A proper design seems make the neura-compiler itself could build the architecture based on the spec, and pass in the const arch to both passes.

[tancheng]

This can be started after tancheng/CGRA-Flow#77 (comment)