[GitHub] [tvm] lhutton1 commented on a diff in pull request #12384: [ETHOSN] Add support for Requantize

[GitBox <gi...@apache.org>]

lhutton1 commented on code in PR #12384:
URL: https://github.com/apache/tvm/pull/12384#discussion_r944311020


##########
src/relay/backend/contrib/ethosn/ethosn_api.cc:
##########
@@ -676,6 +677,35 @@ EthosnError EthosnAPI::Relu(const Expr& expr, ReluParams* params) {
   return err;
 }
 
+EthosnError EthosnAPI::Requantize(const Expr& expr, RequantizeParams* params) {
+  Call call = Downcast<Call>(expr);
+  const auto* input_dtype = call->args[0]->checked_type().as<TensorTypeNode>();
+  sl::TensorShape input_tensor_shape = {1, 1, 1, 1};
+  sl::DataType input_data_type;
+  EthosnError err = Tvm2Npu(input_dtype->shape, &input_tensor_shape);
+  err += Tvm2Npu(input_dtype->dtype, &input_data_type);
+
+  float input_sc, output_sc;
+  int input_zp, output_zp;
+  err += AsConstant(call->args[1], &input_sc);
+  err += AsConstant(call->args[2], &input_zp);
+  err += AsConstant(call->args[3], &output_sc);
+  err += AsConstant(call->args[4], &output_zp);
+
+  params->requantize_info = sl::RequantizeInfo(sl::QuantizationInfo(output_zp, output_sc));

Review Comment:
   Nit: Looks like this has been missed in other operators as well, but the same quantization info object is created below as `output_q_info`. Could we just reuse it instead of creating two?



##########
src/relay/backend/contrib/ethosn/codegen.cc:
##########
@@ -918,6 +943,19 @@ TVM_REGISTER_GLOBAL("relay.ethos-n.support.relu")
       err += EthosnError(reason);
     });
 
+TVM_REGISTER_GLOBAL("relay.ethos-n.support.requantize")
+    .set_body([](tvm::TVMArgs args, tvm::TVMRetValue* rv) {
+      Call call = args[0];
+      RequantizeParams params;
+      auto err = EthosnAPI::Requantize(call, &params);
+      err += EthosnCompiler::SupportedSetup();
+      char reason[kReasonMaxLength];
+      reason[0] = '\0';
+      *rv = !err && EthosnCompiler::GetSupported()->IsRequantizeSupported(
+                        params.requantize_info, params.input_info, nullptr, reason, sizeof(reason));

Review Comment:
   Can we also supply the output info to the issupported check?



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org