TFLM: Fix double-promotion error.

Some of these double promotion is not obvious as va_args implicitly promotes float to double.

PiperOrigin-RevId: 290883516
Change-Id: I23fc883afaac7876f881c85f15d5f21dd7bf4f43

tensorflow/lite/experimental/ruy/check_macros.h

@@ -35,7 +35,7 @@ struct ToString {
 template <>
 struct ToString<float, void> {
   static void Run(float value, char* buf) {
-    snprintf(buf, kValueBufSize, "%.9g", static_cast<double>(value));
+    snprintf(buf, kValueBufSize, "%.9g", value);
   }
 };
 

tensorflow/lite/kernels/BUILD

@@ -1,5 +1,5 @@
 load("//tensorflow/lite:build_def.bzl", "tflite_copts")
-load("//tensorflow/lite/micro:build_def.bzl", "cc_library", "micro_copts")
+load("//tensorflow/lite/micro:build_def.bzl", "cc_library")
 load("//tensorflow/lite:special_rules.bzl", "tflite_portable_test_suite_combined")
 load("//tensorflow:tensorflow.bzl", "tf_opts_nortti_if_android")
 
@@ -373,7 +373,7 @@ cc_library(
     hdrs = [
         "kernel_util.h",
     ],
-    copts = tflite_copts() + micro_copts(),
+    copts = tflite_copts(),
     deps = [
         "//tensorflow/lite/c:common",
         "//tensorflow/lite/kernels/internal:quantization_util",

tensorflow/lite/kernels/internal/BUILD

@@ -1,6 +1,6 @@
 load("//tensorflow:tensorflow.bzl", "transitive_hdrs")
 load("//tensorflow/lite:build_def.bzl", "tflite_copts")
-load("//tensorflow/lite/micro:build_def.bzl", "cc_library", "micro_copts")
+load("//tensorflow/lite/micro:build_def.bzl", "cc_library")
 load("//tensorflow/lite:special_rules.bzl", "tflite_portable_test_suite_combined")
 
 package(
@@ -353,7 +353,7 @@ cc_library(
     name = "quantization_util",
     srcs = ["quantization_util.cc"],
     hdrs = ["quantization_util.h"],
-    copts = tflite_copts() + micro_copts(),
+    copts = tflite_copts(),
     deps = [
         ":compatibility",
         ":round",
@@ -645,7 +645,7 @@ cc_library(
     name = "kernel_utils",
     srcs = ["kernel_utils.cc"],
     hdrs = ["kernel_utils.h"],
-    copts = tflite_copts() + micro_copts(),
+    copts = tflite_copts(),
     deps = [
         ":tensor_utils",
         "//tensorflow/lite/c:common",

tensorflow/lite/kernels/internal/quantization_util.cc

@@ -183,11 +183,11 @@ double DoubleFromFractionAndShift(int64_t fraction, int shift) {
   // Detect NaNs and infinities.
   if (shift == std::numeric_limits<int>::max()) {
     if (fraction == 0) {
-      return std::numeric_limits<double>::quiet_NaN();
+      return NAN;
     } else if (fraction > 0) {
-      return std::numeric_limits<double>::infinity();
+      return INFINITY;
     } else {
-      return -std::numeric_limits<double>::infinity();
+      return -INFINITY;
     }
   }
 
@@ -229,7 +229,7 @@ double IntegerDoubleMultiply(double a, double b) {
   // Detect NaNs and infinities.
   if (a_shift == std::numeric_limits<int>::max() ||
       (b_shift == std::numeric_limits<int>::max())) {
-    return std::numeric_limits<double>::quiet_NaN();
+    return NAN;
   }
   const int result_shift = a_shift + b_shift + 1;
   const int64_t result_fraction = (a_fraction * b_fraction) >> 32;
@@ -379,7 +379,7 @@ bool CheckedLog2(const float x, int* log2_result) {
   const float x_log2_fracpart = x_log2 - x_log2_rounded;
 
   *log2_result = static_cast<int>(x_log2_rounded);
-  return std::abs(x_log2_fracpart) < 1e-3f;
+  return std::abs(x_log2_fracpart) < 1e-3;
 }
 
 void QuantizeMultiplierArray(const double* effective_scales, size_t size,

tensorflow/lite/kernels/internal/reference/quantize.h

@@ -36,9 +36,7 @@ inline void AffineQuantize(const tflite::QuantizationParams& op_params,
 
   for (int i = 0; i < flat_size; i++) {
     const float val = input_data[i];
-    int32 unclamped =
-        static_cast<int32>(TfLiteRound(val / static_cast<float>(scale))) +
-        zero_point;
+    int32 unclamped = static_cast<int32>(TfLiteRound(val / scale)) + zero_point;
     int32 clamped = std::min(std::max(unclamped, min_val), max_val);
     output_data[i] = clamped;
   }

tensorflow/lite/kernels/internal/reference/softmax.h

@@ -43,20 +43,16 @@ inline void Softmax(const SoftmaxParams& params,
       max = std::max(max, input_data[i * depth + c]);
     }
 
-    // TODO(b/148114827): Improve this code.
     // Compute sum.
     float sum = 0.f;
     for (int c = 0; c < depth; ++c) {
-      sum += std::exp(static_cast<double>(input_data[i * depth + c] - max) *
-                      params.beta);
+      sum += std::exp((input_data[i * depth + c] - max) * params.beta);
     }
 
     // Compute result.
     for (int c = 0; c < depth; ++c) {
       output_data[i * depth + c] =
-          std::exp(static_cast<double>(input_data[i * depth + c] - max) *
-                   params.beta) /
-          static_cast<double>(sum);
+          std::exp((input_data[i * depth + c] - max) * params.beta) / sum;
     }
   }
 }

tensorflow/lite/kernels/kernel_util.cc

@@ -118,12 +118,11 @@ TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context,
                                               const TfLiteTensor* bias,
                                               TfLiteTensor* output,
                                               double* multiplier) {
-  const double input_product_scale = static_cast<double>(input->params.scale) *
-                                     static_cast<double>(filter->params.scale);
+  const double input_product_scale = input->params.scale * filter->params.scale;
   // TODO(ahentz): The following conditions must be guaranteed by the training
   // pipeline.
   if (bias) {
-    const double bias_scale = static_cast<double>(bias->params.scale);
+    const double bias_scale = bias->params.scale;
     TF_LITE_ENSURE(context,
                    std::abs(input_product_scale - bias_scale) <=
                        1e-6 * std::min(input_product_scale, bias_scale));
@@ -137,10 +136,9 @@ TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context,
                                               const TfLiteTensor* filter,
                                               TfLiteTensor* output,
                                               double* multiplier) {
-  const double input_product_scale = static_cast<double>(input->params.scale) *
-                                     static_cast<double>(filter->params.scale);
+  const double input_product_scale = input->params.scale * filter->params.scale;
   TF_LITE_ENSURE(context, input_product_scale >= 0);
-  *multiplier = input_product_scale / static_cast<double>(output->params.scale);
+  *multiplier = input_product_scale / output->params.scale;
 
   return kTfLiteOk;
 }

tensorflow/lite/micro/build_def.bzl

@@ -10,9 +10,10 @@ load(
 def micro_copts():
     # TODO(b/139024129): include the followings as well:
     # -Wmissing-field-initializers
+    # -Wdouble-promotion
     # -Wunused-const-variable
     # -Wshadow
-    copts = ["-Werror", "-Wsign-compare", "-Wdouble-promotion"]
+    copts = ["-Werror", "-Wsign-compare"]
     return copts
 
 def cc_library(**kwargs):

tensorflow/lite/micro/examples/hello_world/output_handler.cc

@@ -18,7 +18,5 @@ limitations under the License.
 void HandleOutput(tflite::ErrorReporter* error_reporter, float x_value,
                   float y_value) {
   // Log the current X and Y values
-  error_reporter->Report("x_value: %f, y_value: %f\n",
-                         static_cast<double>(x_value),
-                         static_cast<double>(y_value));
+  error_reporter->Report("x_value: %f, y_value: %f\n", x_value, y_value);
 }

tensorflow/lite/micro/kernels/add.cc

@@ -77,15 +77,14 @@ TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteAddParams* params,
     data->output_offset = output->params.zero_point;
     data->left_shift = 20;
     const double twice_max_input_scale =
-        2 * static_cast<double>(
-                std::max(input1->params.scale, input2->params.scale));
+        2 * std::max(input1->params.scale, input2->params.scale);
     const double real_input1_multiplier =
-        static_cast<double>(input1->params.scale) / twice_max_input_scale;
+        input1->params.scale / twice_max_input_scale;
     const double real_input2_multiplier =
-        static_cast<double>(input2->params.scale) / twice_max_input_scale;
+        input2->params.scale / twice_max_input_scale;
     const double real_output_multiplier =
         twice_max_input_scale /
-        ((1 << data->left_shift) * static_cast<double>(output->params.scale));
+        ((1 << data->left_shift) * output->params.scale);
 
     QuantizeMultiplierSmallerThanOneExp(
         real_input1_multiplier, &data->input1_multiplier, &data->input1_shift);

tensorflow/lite/micro/kernels/comparisons.cc

@@ -43,14 +43,12 @@ constexpr int kOutputTensor = 0;
                                                                                \
       int32 input1_multiplier;                                                 \
       int input1_shift;                                                        \
-      QuantizeMultiplierSmallerThanOneExp(                                     \
-          static_cast<double>(input1->params.scale), &input1_multiplier,       \
-          &input1_shift);                                                      \
+      QuantizeMultiplierSmallerThanOneExp(input1->params.scale,                \
+                                          &input1_multiplier, &input1_shift);  \
       int32 input2_multiplier;                                                 \
       int input2_shift;                                                        \
-      QuantizeMultiplierSmallerThanOneExp(                                     \
-          static_cast<double>(input2->params.scale), &input2_multiplier,       \
-          &input2_shift);                                                      \
+      QuantizeMultiplierSmallerThanOneExp(input2->params.scale,                \
+                                          &input2_multiplier, &input2_shift);  \
                                                                                \
       ComparisonParams op_params;                                              \
       op_params.left_shift = left_shift;                                       \

tensorflow/lite/micro/kernels/dequantize.cc

@@ -46,7 +46,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
 
   tflite::DequantizationParams op_params;
   op_params.zero_point = input->params.zero_point;
-  op_params.scale = static_cast<double>(input->params.scale);
+  op_params.scale = input->params.scale;
   switch (input->type) {
     case kTfLiteUInt8:
       reference_ops::Dequantize(

tensorflow/lite/micro/kernels/mul.cc

@@ -55,9 +55,8 @@ TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node,
       &data->output_activation_max));
 
   if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
-    double real_multiplier = static_cast<double>(input1->params.scale) *
-                             static_cast<double>(input2->params.scale) /
-                             static_cast<double>(output->params.scale);
+    double real_multiplier =
+        input1->params.scale * input2->params.scale / output->params.scale;
     QuantizeMultiplier(real_multiplier, &data->output_multiplier,
                        &data->output_shift);
   }

tensorflow/lite/micro/kernels/pad.cc

@@ -152,9 +152,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
         // same quantized range as the input and output tensors.
         TF_LITE_ENSURE_EQ(context, op_context.output->params.zero_point,
                           op_context.constant_values->params.zero_point);
-        TF_LITE_ENSURE_EQ(
-            context, static_cast<double>(op_context.output->params.scale),
-            static_cast<double>(op_context.constant_values->params.scale));
+        TF_LITE_ENSURE_EQ(context, op_context.output->params.scale,
+                          op_context.constant_values->params.scale);
         pad_value = *GetTensorData<uint8_t>(op_context.constant_values);
       }
       if (op_context.resizing_category == ResizingCategory::kImageStyle) {

tensorflow/lite/micro/kernels/prelu.cc

@@ -53,7 +53,7 @@ inline void BroadcastPrelu4DSlowFloat(
           auto in2_idx = SubscriptToIndex(desc2, b, y, x, c);
           auto in1_val = input1_data[in1_idx];
           auto in2_val = input2_data[in2_idx];
-          output_data[out_idx] = in1_val >= 0.0f ? in1_val : in1_val * in2_val;
+          output_data[out_idx] = in1_val >= 0.0 ? in1_val : in1_val * in2_val;
         }
       }
     }
@@ -67,9 +67,8 @@ TfLiteStatus PreluEval(TfLiteContext* context, TfLiteNode* node) {
   int32_t output_multiplier = 0;
   int output_shift = 0;
   if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt16) {
-    double real_multiplier = static_cast<double>(input->params.scale) *
-                             static_cast<double>(alpha->params.scale) /
-                             static_cast<double>(output->params.scale);
+    double real_multiplier =
+        input->params.scale * alpha->params.scale / output->params.scale;
     QuantizeMultiplierSmallerThanOneExp(real_multiplier, &output_multiplier,
                                         &output_shift);
   }

tensorflow/lite/micro/kernels/quantize.cc

@@ -60,7 +60,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
 
   tflite::QuantizationParams op_params;
   op_params.zero_point = output->params.zero_point;
-  op_params.scale = static_cast<double>(output->params.scale);
+  op_params.scale = output->params.scale;
   switch (output->type) {
     case kTfLiteInt8:
       reference_ops::AffineQuantize(

tensorflow/lite/micro/kernels/softmax.cc

@@ -53,8 +53,7 @@ TfLiteStatus CalculateSoftmaxOpData(TfLiteContext* context,
     static const int kScaledDiffIntegerBits = 5;
 
     tflite::PreprocessSoftmaxScaling(
-        static_cast<double>(params->beta),
-        static_cast<double>(input->params.scale), kScaledDiffIntegerBits,
+        params->beta, input->params.scale, kScaledDiffIntegerBits,
         &data->input_multiplier, &data->input_left_shift);
     data->diff_min = -1.0 * tflite::CalculateInputRadius(
                                 kScaledDiffIntegerBits, data->input_left_shift);
@@ -144,7 +143,7 @@ void Softmax2DQuantized(const TfLiteTensor* input, TfLiteTensor* output,
 void Softmax4DFloat(const TfLiteTensor* input, TfLiteTensor* output,
                     TfLiteSoftmaxParams* params) {
   SoftmaxParams op_params;
-  op_params.beta = static_cast<double>(params->beta);
+  op_params.beta = params->beta;
   tflite::reference_ops::Softmax(
       op_params, GetTensorShape(input), GetTensorData<float>(input),
       GetTensorShape(output), GetTensorData<float>(output));

tensorflow/lite/micro/kernels/svdf.cc

@@ -526,12 +526,12 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
         auto* output_params = reinterpret_cast<TfLiteAffineQuantization*>(
             output->quantization.params);
         const double effective_scale_1 =
-            static_cast<double>(input_params->scale->data[0] *
-                                weights_feature_params->scale->data[0] /
-                                state_params->scale->data[0]);
-        const double effective_scale_2 = static_cast<double>(
-            state_params->scale->data[0] * weight_time_params->scale->data[0] /
-            output_params->scale->data[0]);
+            input_params->scale->data[0] *
+            weights_feature_params->scale->data[0] /
+            state_params->scale->data[0];
+        const double effective_scale_2 = state_params->scale->data[0] *
+                                         weight_time_params->scale->data[0] /
+                                         output_params->scale->data[0];
         QuantizeMultiplier(effective_scale_1, &op_data.effective_scale_1_a,
                            &op_data.effective_scale_1_b);
         QuantizeMultiplier(effective_scale_2, &op_data.effective_scale_2_a,

tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc

@@ -54,7 +54,6 @@ $(MAKEFILE_DIR)/downloads/$(AM_SDK_DEST)/$(SF_BSPS_DEST): $(MAKEFILE_DIR)/downlo
     -Wall \
     -Wextra \
     -Wsign-compare \
-    -Wdouble-promotion \
     -Wno-unused-parameter \
     -Wno-missing-field-initializers \
     -Wno-write-strings \

tensorflow/lite/micro/tools/make/targets/bluepill_makefile.inc

@@ -28,7 +28,6 @@ ifeq ($(TARGET), bluepill)
     -Wall \
     -Wextra \
     -Wsign-compare \
-    -Wdouble-promotion \
     -Wno-unused-parameter \
     -Wno-missing-field-initializers \
     -Wno-write-strings \

tensorflow/lite/micro/tools/make/targets/ecm3531_makefile.inc

@@ -41,7 +41,6 @@ ifeq ($(TARGET), ecm3531)
     -Wall \
     -Wextra \
     -Wsign-compare \
-    -Wdouble-promotion \
     -Wno-unused-parameter \
     -Wno-missing-field-initializers \
     -Wno-write-strings \

tensorflow/lite/micro/tools/make/targets/mcu_riscv_makefile.inc

@@ -24,7 +24,6 @@ ifeq ($(TARGET), riscv32_mcu)
     -Wall \
     -Wextra \
     -Wsign-compare \
-    -Wdouble-promotion \
     -Wno-unused-parameter \
     -Wno-missing-field-initializers \
     -Wno-write-strings \