Home | History | Annotate | Download | only in kernels
      1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
      2 
      3 Licensed under the Apache License, Version 2.0 (the "License");
      4 you may not use this file except in compliance with the License.
      5 You may obtain a copy of the License at
      6 
      7     http://www.apache.org/licenses/LICENSE-2.0
      8 
      9 Unless required by applicable law or agreed to in writing, software
     10 distributed under the License is distributed on an "AS IS" BASIS,
     11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     12 See the License for the specific language governing permissions and
     13 limitations under the License.
     14 ==============================================================================*/
     15 #include <string.h>
     16 #include "tensorflow/contrib/lite/builtin_op_data.h"
     17 #include "tensorflow/contrib/lite/context.h"
     18 #include "tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h"
     19 #include "tensorflow/contrib/lite/kernels/internal/tensor.h"
     20 #include "tensorflow/contrib/lite/kernels/kernel_util.h"
     21 #include "tensorflow/contrib/lite/kernels/op_macros.h"
     22 #include "tensorflow/contrib/lite/string_util.h"
     23 
     24 namespace tflite {
     25 namespace ops {
     26 namespace builtin {
     27 namespace gather {
     28 constexpr int kInputTensor = 0;
     29 constexpr int kInputPositions = 1;
     30 constexpr int kOutputTensor = 0;
     31 
     32 TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
     33   TF_LITE_ENSURE_EQ(context, NumInputs(node), 2);
     34   TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);
     35 
     36   const auto* params =
     37       reinterpret_cast<const TfLiteGatherParams*>(node->builtin_data);
     38   TfLiteTensor* input = GetInput(context, node, kInputTensor);
     39   TfLiteTensor* positions = GetInput(context, node, kInputPositions);
     40   TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
     41   // Only INT32 positions are supported.
     42   TF_LITE_ENSURE_EQ(context, positions->type, kTfLiteInt32);
     43   // Check that input and output types match.
     44   TF_LITE_ENSURE_EQ(context, input->type, output->type);
     45   // TODO(mgubin): only 0D or 1D positions are currently supported.
     46   TF_LITE_ENSURE(context, NumDimensions(positions) <= 1);
     47   // TODO(mgubin): Only default axis == 0 is supported.
     48   TF_LITE_ENSURE_EQ(context, params->axis, 0);
     49   // Check conditions for different types.
     50   switch (input->type) {
     51     case kTfLiteFloat32:
     52     case kTfLiteUInt8:
     53     case kTfLiteInt32: {
     54       // Fully supported by reference_ops::Gather.
     55     } break;
     56 
     57     case kTfLiteString: {
     58       // Only 1D input is supported.
     59       TF_LITE_ENSURE_EQ(context, NumDimensions(input), 1);
     60     } break;
     61     default:
     62       context->ReportError(context,
     63                            "Only float32 and string types are supported");
     64       return kTfLiteError;
     65   }
     66   const int num_dimensions =
     67       NumDimensions(input) + NumDimensions(positions) - 1;
     68   TF_LITE_ENSURE(context, params->axis <= num_dimensions);
     69   TfLiteIntArray* output_shape = TfLiteIntArrayCreate(num_dimensions);
     70   int output_index = 0;
     71   for (int i = 0; i < params->axis; ++i) {
     72     output_shape->data[output_index++] = input->dims->data[i];
     73   }
     74   for (int i = 0; i < positions->dims->size; ++i) {
     75     output_shape->data[output_index++] = positions->dims->data[i];
     76   }
     77   for (int i = params->axis + 1; i < input->dims->size; ++i) {
     78     output_shape->data[output_index++] = input->dims->data[i];
     79   }
     80   return context->ResizeTensor(context, output, output_shape);
     81 }
     82 
     83 TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
     84   TfLiteTensor* input = GetInput(context, node, kInputTensor);
     85   TfLiteTensor* positions = GetInput(context, node, kInputPositions);
     86   TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
     87   const int input_rank = NumDimensions(input);
     88 #define TF_LITE_GATHER(data_type, index_type)                            \
     89   optimized_ops::Gather(                                                 \
     90       GetTensorData<data_type>(input), GetTensorDims(input), input_rank, \
     91       GetTensorData<index_type>(positions), GetTensorDims(positions),    \
     92       GetTensorData<data_type>(output), GetTensorDims(output));
     93   switch (input->type) {
     94     case kTfLiteFloat32:
     95       TF_LITE_GATHER(float, int32_t);
     96       break;
     97     case kTfLiteUInt8:
     98       TF_LITE_GATHER(uint8_t, int32_t);
     99       break;
    100     case kTfLiteInt32:
    101       TF_LITE_GATHER(int32_t, int32_t);
    102       break;
    103     case kTfLiteString: {
    104       DynamicBuffer buffer;
    105       const int32* indexes = positions->data.i32;
    106       const int num_strings = GetStringCount(input);
    107       for (int i = 0; i < positions->dims->data[0]; ++i) {
    108         const int pos = indexes[i];
    109         TF_LITE_ENSURE(context, pos < num_strings);
    110         const auto string_ref = GetString(input, pos);
    111         buffer.AddString(string_ref.str, string_ref.len);
    112       }
    113       buffer.WriteToTensor(output);
    114     } break;
    115     default:
    116       return kTfLiteError;
    117   }
    118 #undef TF_LITE_GATHER
    119   return kTfLiteOk;
    120 }
    121 }  // namespace gather
    122 
    123 TfLiteRegistration* Register_GATHER() {
    124   static TfLiteRegistration r = {nullptr, nullptr, gather::Prepare,
    125                                  gather::Eval};
    126   return &r;
    127 }
    128 
    129 }  // namespace builtin
    130 }  // namespace ops
    131 }  // namespace tflite
    132