From 80f1188eb1a20f2dc43ade9ce3ec47fe4326290a Mon Sep 17 00:00:00 2001
From: Max Ren <maxren@meta.com>
Date: Wed, 23 Aug 2023 15:12:43 -0700
Subject: [PATCH] add preprocessor flag for enabling dq linear in runtime (#66)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/66

DQ Linear runtime logic for XNNPACK is FB internal only, so we are adding a flag in the dq linear logic to remove dq linear logic from xnnpack delegate in OSS

Reviewed By: digantdesai

Differential Revision: D48335277

fbshipit-source-id: b137a667ab204bf94c8332cb31eeedf26ed7672f
---
 backends/xnnpack/runtime/XNNCompiler.cpp | 5 +++++
 backends/xnnpack/runtime/XNNExecutor.cpp | 7 +++++++
 backends/xnnpack/targets.bzl             | 1 +
 3 files changed, 13 insertions(+)

diff --git a/backends/xnnpack/runtime/XNNCompiler.cpp b/backends/xnnpack/runtime/XNNCompiler.cpp
index 0c1c9e6d42c..8498bd84c5f 100644
--- a/backends/xnnpack/runtime/XNNCompiler.cpp
+++ b/backends/xnnpack/runtime/XNNCompiler.cpp
@@ -1517,6 +1517,7 @@ __ET_NODISCARD Error XNNCompiler::compileModel(
   if (!executor->qinputs_.empty() && flatbuffer_graph->xnodes()->size() > 0 &&
       flatbuffer_graph->xnodes()->Get(0)->xnode_union_type() ==
           fb_xnnpack::XNodeUnion::XNNFullyConnected) {
+#ifdef ENABLE_DYNAMIC_QUANTIZATION
     // This delegate is for DQLinear which supports dynamic input shapes
     if (executor->getNumInputs() < 1 || executor->getNumOutputs() != 1) {
       ET_LOG(
@@ -1525,6 +1526,10 @@ __ET_NODISCARD Error XNNCompiler::compileModel(
       return Error::NotSupported;
     }
     executor->setNeedsResizeOutput();
+#else
+    ET_LOG(Error, "DQ Linear is not supported");
+    return Error::NotSupported;
+#endif
   }
 
   return err;
diff --git a/backends/xnnpack/runtime/XNNExecutor.cpp b/backends/xnnpack/runtime/XNNExecutor.cpp
index 5e39c86c1ba..30b60ee329d 100644
--- a/backends/xnnpack/runtime/XNNExecutor.cpp
+++ b/backends/xnnpack/runtime/XNNExecutor.cpp
@@ -7,7 +7,9 @@
  */
 
 #include <executorch/backends/xnnpack/runtime/XNNExecutor.h>
+#ifdef ENABLE_DYNAMIC_QUANTIZATION
 #include <executorch/backends/xnnpack/runtime/utils/utils.h>
+#endif
 
 namespace torch {
 namespace executor {
@@ -17,6 +19,7 @@ namespace delegate {
 Error XNNExecutor::set_external_input(uint32_t id, Tensor* input) {
   auto qinput_pair = qinputs_.find(id);
   if (qinput_pair != qinputs_.end()) {
+#ifdef ENABLE_DYNAMIC_QUANTIZATION
     auto qinput = qinput_pair->second;
     // dq the input and copy it in to qinput
     float input_min, input_max;
@@ -60,6 +63,10 @@ Error XNNExecutor::set_external_input(uint32_t id, Tensor* input) {
         {static_cast<float>(input_qparam.scale),
          static_cast<int8_t>(input_qparam.zero_point)},
         batch_size});
+#else
+    ET_LOG(Error, "Dynamic Quantization is not supported");
+    return Error::NotSupported;
+#endif
   } else {
     externals_.emplace_back(xnn_external_value{id, input->mutable_data_ptr()});
   }
diff --git a/backends/xnnpack/targets.bzl b/backends/xnnpack/targets.bzl
index 6df70f654aa..0f80af77cd0 100644
--- a/backends/xnnpack/targets.bzl
+++ b/backends/xnnpack/targets.bzl
@@ -65,6 +65,7 @@ def define_common_targets():
             "//executorch/extension/pybindings/...",
             "@EXECUTORCH_CLIENTS",
         ],
+        preprocessor_flags = [] if runtime.is_oss else ["-DENABLE_DYNAMIC_QUANTIZATION"],
         deps = [
             third_party_dep("XNNPACK"),
             ":xnnpack_schema",