wehub-resource-sync
95 lines
3.1 KiB
C++
95 lines
3.1 KiB
C++
//
|
|
// ShapeGatherV2.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/01/10.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "shape/SizeComputer.hpp"
|
|
#include "core/Macro.h"
|
|
|
|
namespace MNN {
|
|
|
|
class GatherV2Computer : public SizeComputer {
|
|
virtual bool onComputeSize(const MNN::Op* op, const std::vector<Tensor*>& inputs,
|
|
const std::vector<Tensor*>& outputs) const override {
|
|
Tensor* indices = nullptr;
|
|
int32_t paramShape[MNN_MAX_TENSOR_DIM];
|
|
int32_t paramDim = 0;
|
|
if (inputs.size() == 1) {
|
|
if (nullptr == op->main_as_Input()) {
|
|
MNN_ERROR("One input GatherV2 should has blob parameter\n");
|
|
return false;
|
|
}
|
|
indices = inputs[0];
|
|
auto blob = op->main_as_Input();
|
|
outputs[0]->setType(blob->dtype());
|
|
if (nullptr != blob->dims()) {
|
|
paramDim = blob->dims()->size();
|
|
::memcpy(paramShape, blob->dims()->data(), paramDim * sizeof(int));
|
|
}
|
|
TensorUtils::getDescribe(outputs[0])->dimensionFormat = blob->dformat();
|
|
} else {
|
|
auto params = inputs[0];
|
|
indices = inputs[1];
|
|
paramDim = params->dimensions();
|
|
for (int i = 0; i < params->dimensions(); ++i) {
|
|
paramShape[i] = params->length(i);
|
|
}
|
|
outputs[0]->buffer().type = params->buffer().type;
|
|
TensorUtils::getDescribe(outputs[0])->dimensionFormat = TensorUtils::getDescribe(inputs[0])->dimensionFormat;
|
|
}
|
|
if (indices->getType().code != halide_type_int) {
|
|
return false;
|
|
}
|
|
int axis = 0;
|
|
if (inputs.size() == 3) {
|
|
auto axis_tensor = inputs[2];
|
|
axis = axis_tensor->host<int32_t>()[0];
|
|
}
|
|
if (op->main_type() == OpParameter_Axis && op->main_as_Axis()) {
|
|
axis = op->main_as_Axis()->axis();
|
|
}
|
|
|
|
if( axis <= -paramDim || axis >= paramDim) {
|
|
return false;
|
|
}
|
|
|
|
if (axis < 0) {
|
|
axis = paramDim + axis;
|
|
}
|
|
|
|
const int gather_dim_size = paramShape[axis];
|
|
MNN_ASSERT(gather_dim_size <= std::numeric_limits<int32_t>::max());
|
|
|
|
const int numDimensions = paramDim + indices->buffer().dimensions - 1;
|
|
MNN_ASSERT(axis <= numDimensions);
|
|
|
|
std::vector<int> result_shape;
|
|
|
|
for (int i = 0; i < axis; i++) {
|
|
result_shape.push_back(paramShape[i]);
|
|
}
|
|
|
|
for (int i = 0; i < indices->buffer().dimensions; i++) {
|
|
result_shape.push_back(indices->buffer().dim[i].extent);
|
|
}
|
|
|
|
for (int i = axis + 1; i < paramDim; i++) {
|
|
result_shape.push_back(paramShape[i]);
|
|
}
|
|
|
|
outputs[0]->buffer().dimensions = (int)result_shape.size();
|
|
for (int i = 0; i < result_shape.size(); i++) {
|
|
outputs[0]->buffer().dim[i].extent = result_shape.at(i);
|
|
}
|
|
|
|
return true;
|
|
}
|
|
};
|
|
|
|
REGISTER_SHAPE_INPUTS(GatherV2Computer, OpType_GatherV2, (std::vector<int>{2}));
|
|
REGISTER_SHAPE(GatherV2Computer, OpType_Gather);
|
|
} // namespace MNN
|