Generalize the Conv() op to support input product scale > output scale, i.e.

output_multiplier > 1.

#20451
#19607

PiperOrigin-RevId: 206983654

4 files changed, 50 insertions, 19 deletions

diff --git a/tensorflow/contrib/lite/kernels/conv.cc b/tensorflow/contrib/lite/kernels/conv.cc
index 6f174763df..04c0263b78 100644
--- a/tensorflow/contrib/lite/kernels/conv.cc
+++ b/tensorflow/contrib/lite/kernels/conv.cc
@@ -256,10 +256,10 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
     double real_multiplier = 0.0;
     TF_LITE_ENSURE_STATUS(GetQuantizedConvolutionMultipler(
         context, input, filter, bias, output, &real_multiplier));
-    TF_LITE_ENSURE(context, real_multiplier < 1.0);
-    QuantizeMultiplierSmallerThanOneExp(
-        real_multiplier, &data->output_multiplier, &data->output_shift);
-    data->output_shift *= -1;
+
+    int exponent;
+    QuantizeMultiplier(real_multiplier, &data->output_multiplier, &exponent);
+    data->output_shift = -exponent;
     CalculateActivationRangeUint8(params->activation, output,
                                   &data->output_activation_min,
                                   &data->output_activation_max);
diff --git a/tensorflow/contrib/lite/kernels/conv_test.cc b/tensorflow/contrib/lite/kernels/conv_test.cc
index 0dcfc826fd..24633c2fd7 100644
--- a/tensorflow/contrib/lite/kernels/conv_test.cc
+++ b/tensorflow/contrib/lite/kernels/conv_test.cc
@@ -64,12 +64,6 @@ class BaseConvolutionOpModel : public SingleOpModel {
     }
 
     output_ = AddOutput(output);
-    if (input.type != TensorType_FLOAT32) {
-      // The following is required by quantized inference. It is the unittest's
-      // responsibility to make sure the output scale falls into the correct
-      // range.
-      CHECK_LT(GetScale(input_) * GetScale(filter_), GetScale(output_));
-    }
 
     SetBuiltinOp(BuiltinOperator_CONV_2D, BuiltinOptions_Conv2DOptions,
                  CreateConv2DOptions(
@@ -441,6 +435,44 @@ TEST_P(ConvolutionOpTest, SimpleTestQuantized) {
                              }));
 }
 
+TEST_P(ConvolutionOpTest, SimpleTestQuantizedOutputMultiplierGreaterThan1) {
+  // output_multiplier = 1.0118
+  QuantizedConvolutionOpModel quant_op(
+      GetRegistration(), {TensorType_UINT8, {2, 2, 4, 1}, -128.5, 128},
+      {TensorType_UINT8, {3, 2, 2, 1}, -128.5, 128},
+      {TensorType_UINT8, {}, -127, 128});
+  ConvolutionOpModel float_op(
+      GetRegistration(), {TensorType_FLOAT32, {2, 2, 4, 1}},
+      {TensorType_FLOAT32, {3, 2, 2, 1}}, {TensorType_FLOAT32, {}});
+  std::initializer_list<float> input = {
+      // First batch
+      1, 1, 1, 1,  // row = 1
+      2, 2, 2, 2,  // row = 2
+      // Second batch
+      1, 2, 3, 4,  // row = 1
+      1, 2, 3, 4,  // row = 2
+  };
+  std::initializer_list<float> filter = {
+      1,  2,  3,  4,  // first 2x2 filter
+      -1, 1,  -1, 1,  // second 2x2 filter
+      -1, -1, 1,  1,  // third 2x2 filter
+  };
+  std::initializer_list<float> bias = {1, 2, 3};
+
+  quant_op.SetInput(input);
+  quant_op.SetFilter(filter);
+  quant_op.SetBias(bias);
+  quant_op.Invoke();
+
+  float_op.SetInput(input);
+  float_op.SetFilter(filter);
+  float_op.SetBias(bias);
+  float_op.Invoke();
+
+  EXPECT_THAT(quant_op.GetDequantizedOutput(),
+              ElementsAreArray(ArrayFloatNear(float_op.GetOutput(), 1)));
+}
+
 TEST_P(ConvolutionOpTest, SimpleTestQuantizedWithAnisotropicStrides) {
   QuantizedConvolutionOpModel m(GetRegistration(),
                                 {TensorType_UINT8, {1, 3, 6, 1}, -63.5, 64},
diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h b/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h
index 78567d52ea..bb1dcdda6e 100644
--- a/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h
+++ b/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h
@@ -1018,10 +1018,10 @@ inline void FullyConnectedAsGEMV(
 struct GemmlowpOutputPipeline {
   typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
       ColVectorMap;
-  typedef std::tuple<
-      gemmlowp::OutputStageBiasAddition<ColVectorMap>,
-      gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint,
-      gemmlowp::OutputStageClamp, gemmlowp::OutputStageSaturatingCastToUint8>
+  typedef std::tuple<gemmlowp::OutputStageBiasAddition<ColVectorMap>,
+                     gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent,
+                     gemmlowp::OutputStageClamp,
+                     gemmlowp::OutputStageSaturatingCastToUint8>
       Pipeline;
   static Pipeline MakeExp(const int32* bias_data, int output_rows,
                           int32 output_offset, int32 output_multiplier,
@@ -1030,11 +1030,10 @@ struct GemmlowpOutputPipeline {
     ColVectorMap bias_vector(bias_data, output_rows);
     gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
     bias_addition_stage.bias_vector = bias_vector;
-    gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint
-        quantize_down_stage;
+    gemmlowp::OutputStageScaleInt32ByFixedPointAndExponent quantize_down_stage;
     quantize_down_stage.result_offset_after_shift = output_offset;
     quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
-    quantize_down_stage.result_shift = -output_left_shift;
+    quantize_down_stage.result_exponent = output_left_shift;
     gemmlowp::OutputStageClamp clamp_stage;
     clamp_stage.min = output_activation_min;
     clamp_stage.max = output_activation_max;
diff --git a/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h b/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h
index 714613b96e..3afa336413 100644
--- a/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h
+++ b/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h
@@ -322,8 +322,8 @@ inline void Conv(const uint8* input_data, const Dims<4>& input_dims,
           if (bias_data) {
             acc += bias_data[Offset(bias_dims, out_channel, 0, 0, 0)];
           }
-          acc = MultiplyByQuantizedMultiplierSmallerThanOneExp(
-              acc, output_multiplier, kReverseShift * output_shift);
+          acc = MultiplyByQuantizedMultiplier(acc, output_multiplier,
+                                              kReverseShift * output_shift);
           acc += output_offset;
           acc = std::max(acc, output_activation_min);
           acc = std::min(acc, output_activation_max);