add int32 support for mul

PiperOrigin-RevId: 204230461

tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h

@@ -3054,6 +3054,20 @@ void Mul(const float* input1_data, const Dims<4>& input1_dims,
       output_activation_max, output_data, output_dims);
 }
 
+inline void Mul(const int32* input1_data, const Dims<4>& input1_dims,
+                const int32* input2_data, const Dims<4>& input2_dims,
+                int32 output_activation_min, int32 output_activation_max,
+                int32* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Mul/int32");
+
+  const int flat_size = MatchingFlatSize(input1_dims, input2_dims, output_dims);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] * input2_data[i], output_activation_min,
+        output_activation_max);
+  }
+}
+
 template <FusedActivationFunctionType Ac>
 void Mul(const int32* input1_data, const Dims<4>& input1_dims,
          const int32* input2_data, const Dims<4>& input2_dims,

tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h

@@ -1429,10 +1429,11 @@ inline void BroadcastAddFivefold(
                        output_activation_max, output_data, output_dims);
 }
 
-inline void Mul(const float* input1_data, const Dims<4>& input1_dims,
+template <typename T>
-                const float* input2_data, const Dims<4>& input2_dims,
+inline void Mul(const T* input1_data, const Dims<4>& input1_dims,
-                float output_activation_min, float output_activation_max,
+                const T* input2_data, const Dims<4>& input2_dims,
-                float* output_data, const Dims<4>& output_dims) {
+                T output_activation_min, T output_activation_max,
+                T* output_data, const Dims<4>& output_dims) {
   const int flat_size = MatchingFlatSize(input1_dims, input2_dims, output_dims);
   for (int i = 0; i < flat_size; ++i) {
     output_data[i] = ActivationFunctionWithMinMax(

tensorflow/contrib/lite/kernels/mul.cc

@@ -100,29 +100,44 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
 }
 
 template <KernelType kernel_type>
-void EvalFloat(TfLiteContext* context, TfLiteNode* node,
+void EvalMul(TfLiteContext* context, TfLiteNode* node, TfLiteMulParams* params,
-               TfLiteMulParams* params, const OpData* data,
+             const OpData* data, const TfLiteTensor* input1,
-               const TfLiteTensor* input1, const TfLiteTensor* input2,
+             const TfLiteTensor* input2, TfLiteTensor* output) {
-               TfLiteTensor* output) {
+#define TF_LITE_MUL(type, opname, data_type)                            \
-  float output_activation_min, output_activation_max;
+  data_type output_activation_min, output_activation_max;               \
-  CalculateActivationRange(params->activation, &output_activation_min,
+  CalculateActivationRange(params->activation, &output_activation_min,  \
-                           &output_activation_max);
+                           &output_activation_max);                     \
-#define TF_LITE_MUL(type, opname)                                   \
+  type::opname(GetTensorData<data_type>(input1), GetTensorDims(input1), \
-  type::opname(GetTensorData<float>(input1), GetTensorDims(input1), \
+               GetTensorData<data_type>(input2), GetTensorDims(input2), \
-               GetTensorData<float>(input2), GetTensorDims(input2), \
+               output_activation_min, output_activation_max,            \
-               output_activation_min, output_activation_max,        \
+               GetTensorData<data_type>(output), GetTensorDims(output))
-               GetTensorData<float>(output), GetTensorDims(output))
+  if (output->type == kTfLiteInt32) {
-  if (kernel_type == kReference) {
+    if (kernel_type == kReference) {
-    if (data->requires_broadcast) {
+      if (data->requires_broadcast) {
-      TF_LITE_MUL(reference_ops, BroadcastMul);
+        TF_LITE_MUL(reference_ops, BroadcastMul, int32_t);
+      } else {
+        TF_LITE_MUL(reference_ops, Mul, int32_t);
+      }
     } else {
-      TF_LITE_MUL(reference_ops, Mul);
+      if (data->requires_broadcast) {
+        TF_LITE_MUL(optimized_ops, BroadcastMul, int32_t);
+      } else {
+        TF_LITE_MUL(optimized_ops, Mul, int32_t);
+      }
     }
-  } else {
+  } else if (output->type == kTfLiteFloat32) {
-    if (data->requires_broadcast) {
+    if (kernel_type == kReference) {
-      TF_LITE_MUL(optimized_ops, BroadcastMul);
+      if (data->requires_broadcast) {
+        TF_LITE_MUL(reference_ops, BroadcastMul, float);
+      } else {
+        TF_LITE_MUL(reference_ops, Mul, float);
+      }
     } else {
-      TF_LITE_MUL(optimized_ops, Mul);
+      if (data->requires_broadcast) {
+        TF_LITE_MUL(optimized_ops, BroadcastMul, float);
+      } else {
+        TF_LITE_MUL(optimized_ops, Mul, float);
+      }
     }
   }
 #undef TF_LITE_MUL
@@ -194,17 +209,17 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
   TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
 
-  if (output->type == kTfLiteFloat32) {
+  if (output->type == kTfLiteFloat32 || output->type == kTfLiteInt32) {
-    EvalFloat<kernel_type>(context, node, params, data, input1, input2, output);
+    EvalMul<kernel_type>(context, node, params, data, input1, input2, output);
   } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt16) {
     TF_LITE_ENSURE_OK(
         context, EvalQuantized<kernel_type>(context, node, params, data, input1,
                                             input2, output));
   } else {
-    context->ReportError(
+    context->ReportError(context,
-        context,
+                         "Mul only supports FLOAT32, INT32 and quantized UINT8 "
-        "Mul only supports FLOAT32 and quantized UINT8 and INT16 now, got %d.",
+                         "and INT16 now, got %d.",
-        output->type);
+                         output->type);
     return kTfLiteError;
   }
 

tensorflow/contrib/lite/kernels/mul_test.cc

@@ -52,6 +52,13 @@ class FloatMulOpModel : public BaseMulOpModel {
   std::vector<float> GetOutput() { return ExtractVector<float>(output_); }
 };
 
+class IntegerMulOpModel : public BaseMulOpModel {
+ public:
+  using BaseMulOpModel::BaseMulOpModel;
+
+  std::vector<int32_t> GetOutput() { return ExtractVector<int32_t>(output_); }
+};
+
 // For quantized Mul, the error shouldn't exceed (2*step + step^2).
 // The param min=-1.0 & max=1.0 is used in the following tests.
 // The tolerance value is ~0.0157.
@@ -133,6 +140,57 @@ TEST(FloatMulOpTest, WithBroadcast) {
   }
 }
 
+TEST(IntegerMulOpTest, NoActivation) {
+  IntegerMulOpModel m({TensorType_INT32, {1, 2, 2, 1}},
+                      {TensorType_INT32, {1, 2, 2, 1}}, {TensorType_INT32, {}},
+                      ActivationFunctionType_NONE);
+  m.PopulateTensor<int32_t>(m.input1(), {-20, 2, 7, 8});
+  m.PopulateTensor<int32_t>(m.input2(), {1, 2, 3, 5});
+  m.Invoke();
+  EXPECT_THAT(m.GetOutput(), ElementsAreArray({-20, 4, 21, 40}));
+}
+
+TEST(IntegerMulOpTest, ActivationRELU_N1_TO_1) {
+  IntegerMulOpModel m({TensorType_INT32, {1, 2, 2, 1}},
+                      {TensorType_INT32, {1, 2, 2, 1}}, {TensorType_INT32, {}},
+                      ActivationFunctionType_RELU_N1_TO_1);
+  m.PopulateTensor<int32_t>(m.input1(), {-20, 2, 7, 8});
+  m.PopulateTensor<int32_t>(m.input2(), {1, 2, 3, 5});
+  m.Invoke();
+  EXPECT_THAT(m.GetOutput(), ElementsAreArray({-1, 1, 1, 1}));
+}
+
+TEST(IntegerMulOpTest, VariousInputShapes) {
+  std::vector<std::initializer_list<int>> test_shapes = {
+      {6}, {2, 3}, {2, 1, 3}, {1, 3, 1, 2}};
+  for (int i = 0; i < test_shapes.size(); ++i) {
+    IntegerMulOpModel m({TensorType_INT32, test_shapes[i]},
+                        {TensorType_INT32, test_shapes[i]},
+                        {TensorType_INT32, {}}, ActivationFunctionType_NONE);
+    m.PopulateTensor<int32_t>(m.input1(), {-20, 2, 7, 8, 11, 20});
+    m.PopulateTensor<int32_t>(m.input2(), {1, 2, 3, 5, 11, 1});
+    m.Invoke();
+    EXPECT_THAT(m.GetOutput(), ElementsAreArray({-20, 4, 21, 40, 121, 20}))
+        << "With shape number " << i;
+  }
+}
+
+TEST(IntegerMulOpTest, WithBroadcast) {
+  std::vector<std::initializer_list<int>> test_shapes = {
+      {6}, {2, 3}, {2, 1, 3}, {1, 3, 1, 2}};
+  for (int i = 0; i < test_shapes.size(); ++i) {
+    IntegerMulOpModel m({TensorType_INT32, test_shapes[i]},
+                        {TensorType_INT32, {}},  // always a scalar
+                        {TensorType_INT32, {}}, ActivationFunctionType_NONE);
+    m.PopulateTensor<int32_t>(m.input1(), {-20, 2, 7, 8, 11, 20});
+    m.PopulateTensor<int32_t>(m.input2(), {1});
+    m.Invoke();
+    EXPECT_THAT(m.GetOutput(),
+                ElementsAreArray(ArrayFloatNear({-20, 2, 7, 8, 11, 20})))
+        << "With shape number " << i;
+  }
+}
+
 TEST(QuantizedMulOpTest, NoActivation) {
   QuantizedMulOpModel m({TensorType_UINT8, {1, 2, 2, 1}, -1.0, 1.0},
                         {TensorType_UINT8, {1, 2, 2, 1}, -1.0, 1.0},

tensorflow/contrib/lite/testing/generated_examples_zip_test.cc

@@ -53,7 +53,6 @@ tensorflow::Env* env = tensorflow::Env::Default();
 // Key is a substring of the test name and value is a bug number.
 // TODO(ahentz): make sure we clean this list up frequently.
 std::map<string, string> kBrokenTests = {
-    {R"(^\/mul.*int32)", "68808744"},
     {R"(^\/div.*int32)", "68808744"},
     {R"(^\/sub.*int32)", "68808744"},