From a00d87510dd6619072eaf49f36982194a9d3b1d3 Mon Sep 17 00:00:00 2001
From: PaulPalomeroBernardo <paul.palomero-bernardo@uni-tuebingen.de>
Date: Tue, 31 May 2022 22:12:44 +0200
Subject: [PATCH] Move backend usage to guide-level explanation

---
 ...A_Unified_Modular_Accelerator_Interface.md | 46 +++++++++----------
 1 file changed, 23 insertions(+), 23 deletions(-)

diff --git a/rfcs/0060_UMA_Unified_Modular_Accelerator_Interface.md b/rfcs/0060_UMA_Unified_Modular_Accelerator_Interface.md
index a66d1d46..6b402a52 100644
--- a/rfcs/0060_UMA_Unified_Modular_Accelerator_Interface.md
+++ b/rfcs/0060_UMA_Unified_Modular_Accelerator_Interface.md
@@ -140,6 +140,29 @@ class UltraTrailBackend(UMABackend):
         return "ultra_trail"
 ```
 
+### Using a Custom Backend
+
+A new custom backend is created by implementing the `UMABackend` as shown in the [previous section](#adding-a-new-custom-accelerator). To use the backend it simply needs to be registered using `backend.register()`. Once a `UMABackend` is registered, it hooks into the usual `relay.build` process to create the code for the target accelerator.
+```
+# Load model
+mod, params = relay.frontend.from_pytorch(scripted_model, [("input_data", input_shape)])
+
+# Register a UMA backend
+ut_backend = UltraTrailBackend()
+ut_backend.register()
+mod = ut_backend.partition(mod)
+
+# Relay build (AOT C target)
+TARGET = tvm.target.Target("c")
+RUNTIME = tvm.relay.backend.Runtime("crt")
+EXECUTOR = tvm.relay.backend.Executor("aot", {"unpacked-api": True})
+
+with tvm.transform.PassContext(
+    opt_level=3, config={"tir.disable_vectorize": True}, disabled_pass=["AlterOpLayout"]
+):
+    module = relay.build(mod, target=TARGET, runtime=RUNTIME, executor=EXECUTOR, params=params)
+```
+
 ## Reference-level explanation 
 
 ### File and class structure and Snippets as example for integration
@@ -163,29 +186,6 @@ UMA provides a fully python-based API. The API is structured as shown below. The
     └── ...
 ```
 
-### UMABackend class
-
-A new custom backend is created by implementing the `UMABackend` as shown in the [Guide-level explanation](#adding-a-new-custom-accelerator). To use the backend it simply needs to be registered using `backend.register()`. Once a `UMABackend` is registered, it hooks into the usual `relay.build` process to create the code for the target accelerator.
-```
-# Load model
-mod, params = relay.frontend.from_pytorch(scripted_model, [("input_data", input_shape)])
-
-# Register a UMA backend
-ut_backend = UltraTrailBackend()
-ut_backend.register()
-mod = ut_backend.partition(mod)
-
-# Relay build (AOT C target)
-TARGET = tvm.target.Target("c")
-RUNTIME = tvm.relay.backend.Runtime("crt")
-EXECUTOR = tvm.relay.backend.Executor("aot", {"unpacked-api": True})
-
-with tvm.transform.PassContext(
-    opt_level=3, config={"tir.disable_vectorize": True}, disabled_pass=["AlterOpLayout"]
-):
-    module = relay.build(mod, target=TARGET, runtime=RUNTIME, executor=EXECUTOR, params=params)
-```
-
 ### Pass Phases
 
 UMA allows users to register passes in different stages of the compilation pipeline called pass phases. They are motivated by the existing [TIR pass phases](https://tvm.apache.org/docs/how_to/extend_tvm/low_level_custom_pass.html?highlight=phase#glue-to-lowering) and provide new users with some orientation as to where they should put their custom passes in the entire compilation process. To be more expressive, pass phases are named enums of type `PassPhase`, e.g., `PassPhase.PRE_PARTITIONING`, `PassPhase.POST_PARTITIONING`. This makes them easier to extend and more intuitive to use than an int-based representation.