From f0d6424da5f77d3e23491593027330b53517d591 Mon Sep 17 00:00:00 2001
From: Jian Li <jianlijianli@google.com>
Date: Wed, 17 Jul 2019 07:18:23 -0700
Subject: [PATCH] Add int16 support to Quant.

PiperOrigin-RevId: 258563058
---
 tensorflow/lite/kernels/quantize.cc                 | 13 ++++++++++---
 tensorflow/lite/kernels/quantize_test.cc            | 11 +++++++++++
 tensorflow/lite/kernels/register.cc                 |  4 +++-
 tensorflow/lite/tools/optimize/operator_property.cc |  2 +-
 .../lite/tools/optimize/quantize_model_test.cc      |  2 +-
 5 files changed, 26 insertions(+), 6 deletions(-)

diff --git a/tensorflow/lite/kernels/quantize.cc b/tensorflow/lite/kernels/quantize.cc
index 066bc72618b..35a0bb54f35 100644
--- a/tensorflow/lite/kernels/quantize.cc
+++ b/tensorflow/lite/kernels/quantize.cc
@@ -55,7 +55,8 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
   OpContext op_context(context, node);
 
   TF_LITE_ENSURE(context, op_context.output->type == kTfLiteUInt8 ||
-                              op_context.output->type == kTfLiteInt8);
+                              op_context.output->type == kTfLiteInt8 ||
+                              op_context.output->type == kTfLiteInt16);
 
   // TODO(b/128934713): Add support for fixed-point per-channel quantization.
   // Currently this only support affine per-layer quantization.
@@ -69,9 +70,11 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
 
   // For requantize use case.
   const bool is_requantize = (op_context.input->type == kTfLiteUInt8 ||
-                              op_context.input->type == kTfLiteInt8) &&
+                              op_context.input->type == kTfLiteInt8 ||
+                              op_context.input->type == kTfLiteInt16) &&
                              (op_context.output->type == kTfLiteUInt8 ||
-                              op_context.output->type == kTfLiteInt8);
+                              op_context.output->type == kTfLiteInt8 ||
+                              op_context.output->type == kTfLiteInt16);
   if (is_requantize) {
     const double effective_output_scale =
         static_cast<double>(op_context.input->params.scale) /
@@ -104,6 +107,10 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
         optimized_ops::AffineQuantize(
             op_params, GetTensorShape(input), GetTensorData<float>(input),
             GetTensorShape(output), GetTensorData<uint8_t>(output));
+      } else if (output->type == kTfLiteInt16) {
+        optimized_ops::AffineQuantize(
+            op_params, GetTensorShape(input), GetTensorData<float>(input),
+            GetTensorShape(output), GetTensorData<int16_t>(output));
       } else {
         context->ReportError(
             context,
diff --git a/tensorflow/lite/kernels/quantize_test.cc b/tensorflow/lite/kernels/quantize_test.cc
index b381c041859..e720f74728e 100644
--- a/tensorflow/lite/kernels/quantize_test.cc
+++ b/tensorflow/lite/kernels/quantize_test.cc
@@ -79,6 +79,17 @@ TEST(QuantizeOpTest, INT8) {
                   {-128, -127, -126, -125, -124, 123, 124, 125, 126, 127}));
 }
 
+TEST(QuantizeOpTest, INT16) {
+  QuantizeOpModel m({TensorType_FLOAT32, {2, 5}},
+                    {TensorType_INT16, {2, 5}, 0, 0, 0.005, 0});
+
+  m.SetInput({-63.5, -63, -3, -2, -1, 1, 2, 3, 63.5, 64});
+  m.Invoke();
+  EXPECT_THAT(m.GetOutput<int16_t>(),
+              ElementsAreArray({-12700, -12600, -600, -400, -200, 200, 400, 600,
+                                12700, 12800}));
+}
+
 // Input scale 0.500000, output scale 0.500000, input zeropoint -1, output
 // zeropoint -1
 TEST(QuantizeOpTest, Int8Int8SameScale) {
diff --git a/tensorflow/lite/kernels/register.cc b/tensorflow/lite/kernels/register.cc
index 700b321c7a9..f17c445a063 100644
--- a/tensorflow/lite/kernels/register.cc
+++ b/tensorflow/lite/kernels/register.cc
@@ -376,7 +376,9 @@ BuiltinOpResolver::BuiltinOpResolver() {
   AddBuiltin(BuiltinOperator_ELU, Register_ELU());
   AddBuiltin(BuiltinOperator_REVERSE_SEQUENCE, Register_REVERSE_SEQUENCE());
   AddBuiltin(BuiltinOperator_MATRIX_DIAG, Register_MATRIX_DIAG());
-  AddBuiltin(BuiltinOperator_QUANTIZE, Register_QUANTIZE());
+  AddBuiltin(BuiltinOperator_QUANTIZE, Register_QUANTIZE(),
+             /* min_version */ 1,
+             /* max_version */ 2);
   AddBuiltin(BuiltinOperator_MATRIX_SET_DIAG, Register_MATRIX_SET_DIAG());
 
   // TODO(andrewharp, ahentz): Move these somewhere more appropriate so that
diff --git a/tensorflow/lite/tools/optimize/operator_property.cc b/tensorflow/lite/tools/optimize/operator_property.cc
index f5b2736415e..ad974e7aca3 100644
--- a/tensorflow/lite/tools/optimize/operator_property.cc
+++ b/tensorflow/lite/tools/optimize/operator_property.cc
@@ -169,7 +169,7 @@ OperatorProperty GetOperatorProperty(const BuiltinOperator& op) {
     case BuiltinOperator_QUANTIZE:
       property.inputs = {{0, {}}};
       property.outputs = {{0, {}}};
-      property.version = 1;
+      property.version = 2;
       break;
     case BuiltinOperator_RESHAPE:
       property.inputs = {{0, {}}};
diff --git a/tensorflow/lite/tools/optimize/quantize_model_test.cc b/tensorflow/lite/tools/optimize/quantize_model_test.cc
index f41bf077cd3..7de8398c287 100644
--- a/tensorflow/lite/tools/optimize/quantize_model_test.cc
+++ b/tensorflow/lite/tools/optimize/quantize_model_test.cc
@@ -370,7 +370,7 @@ TEST_F(QuantizeConcatModelTest, AddRequantBeforeConcat) {
             BuiltinOperator_CONCATENATION);
   EXPECT_EQ(model_.operator_codes[0]->version, 2);
   EXPECT_EQ(model_.operator_codes[1]->builtin_code, BuiltinOperator_QUANTIZE);
-  EXPECT_EQ(model_.operator_codes[1]->version, 1);
+  EXPECT_EQ(model_.operator_codes[1]->version, 2);
 }
 
 class QuantizeConvModel1Test : public QuantizeModelTest {
-- 
GitLab