[feat] Kernel-level fusion

[HobbitQia]

PR 251

Implements --fuse-task, a new pass at the Taskflow IR level that merges adjacent loop kernels before lowering to Neura. Two strategies are supported:

• Producer-consumer: fuses two tasks where the first writes an intermediate buffer consumed by the second; the intermediate store/load is eliminated.
• Sibling: fuses two tasks that share input arrays but have no data dependency.

Both strategies are gated by an MII-based profitability check: the pass speculatively lowers candidate kernels through the full Taskflow → Neura pipeline on a cloned module, measures rec_mii and res_mii, and only fuses when the fused MII does not exceed the cost of running the two tasks independently.

Example: Producer-Consumer
Before fusion:

// Task 0: C[i] = A[i] + B[i]
%t0 = taskflow.task @Task_0 read_memrefs(%A, %B) write_memrefs(%C) : ... {
  ^bb0(%i: index):
    %a = memref.load %A[%i] : memref<64xf32>
    %b = memref.load %B[%i] : memref<64xf32>
    %s = arith.addf %a, %b : f32
    memref.store %s, %C[%i] : memref<64xf32>
}
// Task 1: D[i] = C[i] * 2.0  (consumes C from Task 0)
%t1 = taskflow.task @Task_1 read_memrefs(%t0) write_memrefs(%D) value_inputs(%cst) : ... {
  ^bb0(%i: index):
    %v = memref.load %C[%i] : memref<64xf32>
    %r = arith.mulf %v, %cst : f32
    memref.store %r, %D[%i] : memref<64xf32>
}

After fusion:

// Fused: D[i] = (A[i] + B[i]) * 2.0
%t = taskflow.task @fused_pc read_memrefs(%A, %B) write_memrefs(%C, %D) value_inputs(%cst) : ... {
  ^bb0(%i: index):
    %a = memref.load %A[%i] : memref<64xf32>
    %b = memref.load %B[%i] : memref<64xf32>
    %s = arith.addf %a, %b : f32
    %r = arith.mulf %s, %cst : f32
    memref.store %r, %D[%i] : memref<64xf32>
}

[HobbitQia]

@ShangkunLi You can check the function computeRealMetrics where I created a cloned module and used pass manager to apply the all pass pipelines on the current codes to obtain statistics.