Now，I woulde like to use te.extern to add an extern op level library ，like
cuDNN. For complex ops, like conv2d or dense , we can add new dispatch rules
in op strategy,like @softmax_strategy.register(["cuda", "gpu"]), and then
redefine the compute with te.extern. But for some injective or broadcast ops,
TVM
define the same compute function across all targets. As follows,
RELAY_REGISTER_OP("nn.relu")
.describe(R"code(Returns the relu input array, computed element-wise.
.. math::
max(x, 0)
)code" TVM_ADD_FILELINE)
.set_num_inputs(1)
.add_argument("data", "Tensor", "The input tensor.")
.set_support_level(1)
.add_type_rel("Identity", IdentityRel)
.set_attr<FInferCorrectLayout>("FInferCorrectLayout",
ElemwiseArbitraryLayout)
.set_attr<FTVMCompute>("FTVMCompute", [](const Attrs& attrs, const
Array<te::Tensor>& inputs,
const Type& out_type) {
return Array<te::Tensor>{topi::relu(inputs[0], 0.0f)};
});
Now, It seems that dispatch logic has been removed from TOPI. So how can I add
new compute for relu? @haichen
---
[Visit Topic](https://discuss.tvm.ai/t/relay-op-strategy/5247/20) to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click
here](https://discuss.tvm.ai/email/unsubscribe/22c3a7394ca4504768dd594002cc489cf8ee29474346fa2992b5c01ee96a8168).
