1. Relay算子注册 (C++层)

(a) 算子属性注册

路径: 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) 调用节点构造

路径: 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计算实现 (C++层)

© TOPI注册入口

路径: src/topi/elemwise.cc

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

(d) 数学内核实现

路径: 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接口层

(e) Relay Python API

路径: python/tvm/relay/op/nn/_nn.py

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

(f) TOPI通用接口

路径: python/tvm/topi/nn.py

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

---

4. 计算调度注册

(g) Compute注册

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

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

(h) 调度策略

路径: `python/tvm/relay/op/op.py**

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

---

5. 硬件专用实现

(i) NPU支持声明

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

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

(j) NPU内核实现

路径: `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策略注册

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

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

---

6. 前端框架对接

(l) PyTorch转换器

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

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

---

关键文件路径总结

功能模块	关键路径
Relay核心注册	src/relay/op/nn/{nn.cc, activation.cc}
TOPI计算	{include,src}/topi/{nn.h, elemwise.cc}
Python接口	python/tvm/{relay/op/nn/_nn.py, topi/nn.py}
策略注册	python/tvm/relay/op/strategy/{generic.py, npu.py}
硬件后端	src/relay/backend/contrib/npu/
前端对接	python/tvm/relay/frontend/pytorch.py

---

开发流程示意图

通过这种清晰的路径划分，TVM实现了：

1. 模块化开发：各层级代码物理隔离
2. 可扩展性：新增硬件只需在对应目录添加实现
3. 维护性：相关功能的代码集中存放