Now, I want to use a tf model to test BYOC and I met some problems. I think there some problems in dnnl/codegen.cc file
void GenDNNLFunc(const Function& func) {
CHECK(func.defined()) << "Input error: expect a Relay function.";
const auto* call = func->body.as<CallNode>();
CHECK(call) << "DNNL expects a single convolution or dense op";
// Record the external symbol for runtime lookup.
auto sid = GetExtSymbol(func);
CodegenDNNL builder(sid);
auto out = builder.VisitExpr(func->body);
code_stream_ << builder.JIT(out);
}
In this function, I think that not all func’s body is callNode. In my test model , some func’s body is tupleNode. So I change this function like this
void GenDNNLFunc(const Function& func) {
CHECK(func.defined()) << "Input error: expect a Relay function.";
if(func->body->IsInstance<TupleNode>()) {
auto* tuple = func->body.as<TupleNode>();
for(Expr field : tuple->fields) {
LOG(INFO) << "wda debug : " << field;
const auto* call = field.as<CallNode>();
//if(call == NULL)
// return;
CHECK(call) << "SSDNN expects a single convolution or dense op";
}
}else if(func->body->IsInstance<CallNode>()) {
const auto* call = func->body.as<CallNode>();
CHECK(call) << "DNNL expects a single convolution or dense op";
}
// Record the external symbol for runtime lookup.
auto sid = GetExtSymbol(func);
CodegenDNNL builder(sid);
builder.VisitExpr(func->body);
code_stream_ << builder.JIT();
}
I think this is also some problems. because this can not return a tuple. It can’t handle the following cases
def @dnnl_6(%dnnl_6_i0: Tensor[(1, 7, 7, 2048), float32], Inline=1, Compiler="dnnl", global_symbol="dnnl_6", Primitive=1) -> (Tensor[(1, 7, 7, 2048), float32], Tensor[(1, 7, 7, 2048), float32]) {
%10 = nn.relu(%dnnl_6_i0) /* ty=Tensor[(1, 7, 7, 2048), float32] */;
(%10, %10)
}
