Switch TFLM Conv2D kernels to flat namespace.

This is incremental progress towards a flat namespace for TFLM. See
https://abseil.io/tips/130 for more context.

Best effort change to the arc_mli implementation. All the others (reference,
cmsis-nn and xtensa_hifimini build).

PiperOrigin-RevId: 335942662
Change-Id: I97271a681f46337a7c49f3fdf89ae73567b2ed7d

tensorflow/lite/micro/kernels/arc_mli/conv.cc

@@ -30,9 +30,7 @@ limitations under the License.
 #include "tensorflow/lite/micro/kernels/arc_mli/scratch_buffers.h"
 
 namespace tflite {
-namespace ops {
-namespace micro {
-namespace conv {
+namespace {
 
 constexpr int kInputTensor = 0;
 constexpr int kFilterTensor = 1;
@@ -498,19 +496,17 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   return kTfLiteOk;
 }
 
-}  // namespace conv
+}  // namespace
 
 TfLiteRegistration Register_CONV_2D() {
-  return {/*init=*/conv::Init,
+  return {/*init=*/Init,
           /*free=*/nullptr,
-          /*prepare=*/conv::Prepare,
-          /*invoke=*/conv::Eval,
+          /*prepare=*/Prepare,
+          /*invoke=*/Eval,
           /*profiling_string=*/nullptr,
           /*builtin_code=*/0,
           /*custom_name=*/nullptr,
           /*version=*/0};
 }
 
-}  // namespace micro
-}  // namespace ops
 }  // namespace tflite

tensorflow/lite/micro/kernels/cmsis-nn/conv.cc

@@ -28,9 +28,7 @@ limitations under the License.
 #include "tensorflow/lite/micro/kernels/kernel_util.h"
 
 namespace tflite {
-namespace ops {
-namespace micro {
-namespace conv {
+namespace {
 
 constexpr int kInputTensor = 0;
 constexpr int kFilterTensor = 1;
@@ -442,19 +440,17 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   return kTfLiteOk;
 }
 
-}  // namespace conv
+}  // namespace
 
 TfLiteRegistration Register_CONV_2D() {
-  return {/*init=*/conv::Init,
+  return {/*init=*/Init,
           /*free=*/nullptr,
-          /*prepare=*/conv::Prepare,
-          /*invoke=*/conv::Eval,
+          /*prepare=*/Prepare,
+          /*invoke=*/Eval,
           /*profiling_string=*/nullptr,
           /*builtin_code=*/0,
           /*custom_name=*/nullptr,
           /*version=*/0};
 }
 
-}  // namespace micro
-}  // namespace ops
 }  // namespace tflite

tensorflow/lite/micro/kernels/conv.cc

@@ -26,9 +26,7 @@ limitations under the License.
 #include "tensorflow/lite/micro/kernels/kernel_util.h"
 
 namespace tflite {
-namespace ops {
-namespace micro {
-namespace conv {
+namespace {
 
 constexpr int kInputTensor = 0;
 constexpr int kFilterTensor = 1;
@@ -322,19 +320,17 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   return kTfLiteOk;
 }
 
-}  // namespace conv
+}  // namespace
 
 TfLiteRegistration Register_CONV_2D() {
-  return {/*init=*/conv::Init,
+  return {/*init=*/Init,
           /*free=*/nullptr,
-          /*prepare=*/conv::Prepare,
-          /*invoke=*/conv::Eval,
+          /*prepare=*/Prepare,
+          /*invoke=*/Eval,
           /*profiling_string=*/nullptr,
           /*builtin_code=*/0,
           /*custom_name=*/nullptr,
           /*version=*/0};
 }
 
-}  // namespace micro
-}  // namespace ops
 }  // namespace tflite

tensorflow/lite/micro/kernels/conv_test.cc

@@ -59,8 +59,7 @@ TfLiteStatus ValidateConvGoldens(TfLiteTensor* tensors, int tensors_size,
   int outputs_array_data[] = {1, 3};
   TfLiteIntArray* outputs_array = IntArrayFromInts(outputs_array_data);
 
-  const TfLiteRegistration registration =
-      tflite::ops::micro::Register_CONV_2D();
+  const TfLiteRegistration registration = Register_CONV_2D();
   micro::KernelRunner runner(
       registration, tensors, tensors_size, inputs_array, outputs_array,
       reinterpret_cast<void*>(conv_params), micro_test::reporter);

tensorflow/lite/micro/kernels/micro_ops.h

@@ -31,6 +31,7 @@ namespace tflite {
 // (https://abseil.io/tips/130). Any new ops (or cleanup of existing ops should
 // have their Register function declarations in the tflite namespace.
 
+TfLiteRegistration Register_CONV_2D();
 TfLiteRegistration Register_SHAPE();
 
 namespace ops {
@@ -44,7 +45,6 @@ TfLiteRegistration Register_AVERAGE_POOL_2D();
 TfLiteRegistration Register_CEIL();
 // TODO(b/160234179): Change custom OPs to also return by value.
 TfLiteRegistration* Register_CIRCULAR_BUFFER();
-TfLiteRegistration Register_CONV_2D();
 TfLiteRegistration Register_CONCATENATION();
 TfLiteRegistration Register_COS();
 TfLiteRegistration Register_DEPTHWISE_CONV_2D();

tensorflow/lite/micro/kernels/xtensa_hifimini/conv.cc

@@ -28,11 +28,37 @@ limitations under the License.
 #include "tensorflow/lite/micro/kernels/xtensa_hifimini/fixedpoint_utils.h"
 
 namespace tflite {
-namespace ops {
-namespace micro {
-namespace conv {
-namespace xtensa {
-namespace hifimini {
+namespace {
+
+constexpr int kInputTensor = 0;
+constexpr int kFilterTensor = 1;
+constexpr int kBiasTensor = 2;
+constexpr int kOutputTensor = 0;
+
+// Conv is quantized along dimension 0:
+// https://www.tensorflow.org/lite/performance/quantization_spec
+constexpr int kConvQuantizedDimension = 0;
+
+struct OpData {
+  TfLitePaddingValues padding;
+  // The scaling factor from input to output (aka the 'real multiplier') can
+  // be represented as a fixed point multiplier plus a left shift.
+  int32_t output_multiplier;
+  int output_shift;
+
+  // Cached tensor zero point values for quantized operations.
+  int32_t input_zero_point;
+  int32_t output_zero_point;
+
+  // Per channel output multiplier and shift.
+  int32_t* per_channel_output_multiplier;
+  int32_t* per_channel_output_shift;
+
+  // The range of the fused activation layer. For example for kNone and
+  // uint8_t these would be 0 and 255.
+  int32_t output_activation_min;
+  int32_t output_activation_max;
+};
 
 void ConvPerChannel(const ConvParams& params, const int32_t* output_multiplier,
                     const int32_t* output_shift,
@@ -146,7 +172,7 @@ void ConvPerChannel(const ConvParams& params, const int32_t* output_multiplier,
 
           // Apply quantized multiplier and accumulate result at 48bit
           // alignment:
-          acc_56 = micro::xtensa::hifimini::MultiplyByQuantizedMultiplier(
+          acc_56 = ops::micro::xtensa::hifimini::MultiplyByQuantizedMultiplier(
               acc_24x2, output_multiplier[out_channel],
               output_shift[out_channel]);
 
@@ -223,7 +249,7 @@ inline void Conv1x32Input32x32Filter(
 
     // Apply quantized multiplier and accumulate result at 48bit alignment.
     // Convert the (unsigned) 32-bit multiplier down to a 24-bit multiplier.
-    acc_56 = micro::xtensa::hifimini::MultiplyByQuantizedMultiplier(
+    acc_56 = ops::micro::xtensa::hifimini::MultiplyByQuantizedMultiplier(
         acc_24x2, output_multiplier[ch] >> 8, output_shift[ch]);
 
     // Add output offset, cap activation, and assign to the output:
@@ -235,39 +261,6 @@ inline void Conv1x32Input32x32Filter(
   }
 }
 
-}  // namespace hifimini
-}  // namespace xtensa
-
-constexpr int kInputTensor = 0;
-constexpr int kFilterTensor = 1;
-constexpr int kBiasTensor = 2;
-constexpr int kOutputTensor = 0;
-
-// Conv is quantized along dimension 0:
-// https://www.tensorflow.org/lite/performance/quantization_spec
-constexpr int kConvQuantizedDimension = 0;
-
-struct OpData {
-  TfLitePaddingValues padding;
-  // The scaling factor from input to output (aka the 'real multiplier') can
-  // be represented as a fixed point multiplier plus a left shift.
-  int32_t output_multiplier;
-  int output_shift;
-
-  // Cached tensor zero point values for quantized operations.
-  int32_t input_zero_point;
-  int32_t output_zero_point;
-
-  // Per channel output multiplier and shift.
-  int32_t* per_channel_output_multiplier;
-  int32_t* per_channel_output_shift;
-
-  // The range of the fused activation layer. For example for kNone and
-  // uint8_t these would be 0 and 255.
-  int32_t output_activation_min;
-  int32_t output_activation_max;
-};
-
 TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node,
                              TfLiteConvParams* params, int width, int height,
                              int filter_width, int filter_height, int out_width,
@@ -386,16 +379,16 @@ void EvalQuantizedPerChannel(TfLiteContext* context, TfLiteNode* node,
   op_params.quantized_activation_min = data->output_activation_min;
   op_params.quantized_activation_max = data->output_activation_max;
 
-  xtensa::hifimini::ConvPerChannel(
-      op_params, data->per_channel_output_multiplier,
-      data->per_channel_output_shift, tflite::micro::GetTensorShape(input),
-      tflite::micro::GetTensorData<int8_t>(input),
-      tflite::micro::GetTensorShape(filter),
-      tflite::micro::GetTensorData<int8_t>(filter),
-      tflite::micro::GetTensorShape(bias),
-      tflite::micro::GetTensorData<int32_t>(bias),
-      tflite::micro::GetTensorShape(output),
-      tflite::micro::GetTensorData<int8_t>(output));
+  ConvPerChannel(op_params, data->per_channel_output_multiplier,
+                 data->per_channel_output_shift,
+                 tflite::micro::GetTensorShape(input),
+                 tflite::micro::GetTensorData<int8_t>(input),
+                 tflite::micro::GetTensorShape(filter),
+                 tflite::micro::GetTensorData<int8_t>(filter),
+                 tflite::micro::GetTensorShape(bias),
+                 tflite::micro::GetTensorData<int32_t>(bias),
+                 tflite::micro::GetTensorShape(output),
+                 tflite::micro::GetTensorData<int8_t>(output));
 }
 
 TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
@@ -420,7 +413,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   if (input_dims[0] == 1 && input_dims[1] == 1 && input_dims[2] == 1 &&
       input_dims[3] == 32 && filter_dims[0] == 32 && filter_dims[1] == 1 &&
       filter_dims[2] == 1 && filter_dims[3] == 32) {
-    xtensa::hifimini::Conv1x32Input32x32Filter(
+    Conv1x32Input32x32Filter(
         -op_data->input_zero_point, op_data->output_zero_point,
         op_data->output_activation_min, op_data->output_activation_max,
         op_data->per_channel_output_multiplier,
@@ -447,20 +440,17 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   }
   return kTfLiteOk;
 }
-
-}  // namespace conv
+}  // namespace
 
 TfLiteRegistration Register_CONV_2D() {
-  return {/*init=*/conv::Init,
+  return {/*init=*/Init,
           /*free=*/nullptr,
-          /*prepare=*/conv::Prepare,
-          /*invoke=*/conv::Eval,
+          /*prepare=*/Prepare,
+          /*invoke=*/Eval,
           /*profiling_string=*/nullptr,
           /*builtin_code=*/0,
           /*custom_name=*/nullptr,
           /*version=*/0};
 }
 
-}  // namespace micro
-}  // namespace ops
 }  // namespace tflite

tensorflow/lite/micro/micro_mutable_op_resolver.h

@@ -150,8 +150,7 @@ class MicroMutableOpResolver : public MicroOpResolver {
   }
 
   TfLiteStatus AddConv2D() {
-    return AddBuiltin(BuiltinOperator_CONV_2D,
-                      tflite::ops::micro::Register_CONV_2D(), ParseConv2D);
+    return AddBuiltin(BuiltinOperator_CONV_2D, Register_CONV_2D(), ParseConv2D);
   }
 
   TfLiteStatus AddCos() {
@@ -465,7 +464,6 @@ class MicroMutableOpResolver : public MicroOpResolver {
   unsigned int num_buitin_ops_ = 0;
 
   ErrorReporter* error_reporter_;
-
 };
 
 };  // namespace tflite