1. Relay算子注冊 (C++層)

(a) 算子屬性注冊

路徑: src/relay/op/nn/nn.cc

RELAY_REGISTER_OP("hardswish").set_num_inputs(1).add_argument("data", "Tensor", "Input tensor.").set_support_level(3).add_type_rel("Identity", Identity);

(b) 調用節點構造

路徑: src/relay/op/nn/activation.cc

TVM_REGISTER_GLOBAL("relay.op._make.hardswish").set_body_typed([](Expr data) {static const Op& op = Op::Get("hardswish");return Call(op, {data}, Attrs(), {});});

---

2. TOPI計算實現 (C++層)

? TOPI注冊入口

路徑: src/topi/elemwise.cc

TVM_REGISTER_GLOBAL("topi.hardswish").set_body([](TVMArgs args, TVMRetValue* rv) {*rv = hardswish(args[0]);});

(d) 數學內核實現

路徑: include/tvm/topi/nn.h

inline Tensor hardswish(const Tensor& x, std::string name = "T_hardswish") {auto three = make_const(x->dtype, 3);auto six = make_const(x->dtype, 6);return compute(x->shape,[&](const Array<Var>& i) {return x(i) * max(min(x(i) + three, six), 0) / six;},name, kElementWise);
}

---

3. Python接口層

(e) Relay Python API

路徑: python/tvm/relay/op/nn/_nn.py

def hardswish(data):return _make.hardswish(data)

(f) TOPI通用接口

路徑: python/tvm/topi/nn.py

@tvm.target.generic_func
def hardswish(x):return cpp.hardswish(x)

---

4. 計算調度注冊

(g) Compute注冊

路徑: python/tvm/relay/op/strategy/generic.py

@register_compute("hardswish")
def hardswish_compute(attrs, inputs, out_type):return [topi.hardswish(inputs[0])]

(h) 調度策略

路徑: `python/tvm/relay/op/op.py**

register_broadcast_schedule("hardswish")
register_shape_func("hardswish", False, elemwise_shape_func)

---

5. 硬件專用實現

(i) NPU支持聲明

路徑: `src/relay/backend/contrib/npu/src/op_map.cc**

const std::vector<std::string> _NPU_OP = {...,"hardswish"  // 添加算子名
};

(j) NPU內核實現

路徑: `python/tvm/relay/backend/contrib/npu/ops.py**

def custom_hardswish(x):x1 = custom_add(x, te.extern_scalar_value(3.0))x2 = custom_relu(x1)return npu_hardwish(x2, ...)

(k) NPU策略注冊

路徑: `python/tvm/relay/op/strategy/npu.py**

@hardswish.register("npu")
def hardswish_npu(x):return npu_api.custom_hardswish(x)

---

6. 前端框架對接

(l) PyTorch轉換器

路徑: `python/tvm/relay/frontend/pytorch.py**

def _hardswish():def _impl(inputs, input_types):return _op.hardswish(inputs[0])return _impl

---

關鍵文件路徑總結

功能模塊	關鍵路徑
Relay核心注冊	src/relay/op/nn/{nn.cc, activation.cc}
TOPI計算	{include,src}/topi/{nn.h, elemwise.cc}
Python接口	python/tvm/{relay/op/nn/_nn.py, topi/nn.py}
策略注冊	python/tvm/relay/op/strategy/{generic.py, npu.py}
硬件后端	src/relay/backend/contrib/npu/
前端對接	python/tvm/relay/frontend/pytorch.py

---

開發流程示意圖

通過這種清晰的路徑劃分，TVM實現了：

1. 模塊化開發：各層級代碼物理隔離
2. 可擴展性：新增硬件只需在對應目錄添加實現
3. 維護性：相關功能的代碼集中存放