/* * Copyright (C) 2018 The Android Open Source Project * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #include "CpuOperationUtils.h" #include "OperationResolver.h" #include "OperationsUtils.h" #include #include #include "Tracing.h" #include "tensorflow/lite/kernels/internal/common.h" namespace android { namespace nn { namespace roi_align { constexpr char kOperationName[] = "ROI_ALIGN"; constexpr uint32_t kNumInputs = 10; constexpr uint32_t kInputTensor = 0; constexpr uint32_t kRoiTensor = 1; constexpr uint32_t kBatchSplitTensor = 2; constexpr uint32_t kOutputHeightScalar = 3; constexpr uint32_t kOutputWidthScalar = 4; constexpr uint32_t kHeightStrideSalar = 5; constexpr uint32_t kWidthStrideScalar = 6; constexpr uint32_t kHeightSamplingRatioScalar = 7; constexpr uint32_t kWidthSamplingRatioScalar = 8; constexpr uint32_t kLayoutScalar = 9; constexpr uint32_t kNumOutputs = 1; constexpr uint32_t kOutputTensor = 0; namespace { template inline bool roiAlignNhwc(const T_Input* inputData, const Shape& inputShape, const T_Roi* roiData, const Shape& roiShape, const int32_t* batchSplitData, const Shape& batchSplitShape, float heightStride, float widthStride, int32_t heightSamplingRatio, int32_t widthSamplingRatio, T_Input* outputData, const Shape& outputShape) { NNTRACE_TRANS("RoiAlign"); const uint32_t kRoiDim = 4; const T_Roi heightScale = 1.0f / heightStride; const T_Roi widthScale = 1.0f / widthStride; uint32_t numBatches = getSizeOfDimension(inputShape, 0); uint32_t inHeight = getSizeOfDimension(inputShape, 1); uint32_t inWidth = getSizeOfDimension(inputShape, 2); uint32_t inDepth = getSizeOfDimension(inputShape, 3); uint32_t outHeight = getSizeOfDimension(outputShape, 1); uint32_t outWidth = getSizeOfDimension(outputShape, 2); uint32_t numRois = getSizeOfDimension(roiShape, 0); uint32_t roiInfoLength = getSizeOfDimension(roiShape, 1); T_Input* outPtr = outputData; const T_Roi* roiDataEnd = roiData + numRois * roiInfoLength; uint32_t roiIndex = 0; for (const T_Roi* roiInfo = roiData; roiInfo < roiDataEnd; roiInfo += kRoiDim, roiIndex++) { uint32_t batchId = static_cast(batchSplitData[roiIndex]); // Check for malformed data // 1. invalid batch id // 2. Region out of bound: x1|x2|y1|y2 < 0 || x1|x2 > inWidth || y1|y2 > inHeight // 3. Invalid region: x2 < x1 || y2 < y1 NN_RET_CHECK_GE(batchId, 0); NN_RET_CHECK_LT(batchId, numBatches); NN_RET_CHECK(roiInfo[0] >= 0); NN_RET_CHECK(roiInfo[1] >= 0); NN_RET_CHECK(roiInfo[2] >= 0); NN_RET_CHECK(roiInfo[3] >= 0); NN_RET_CHECK(roiInfo[0] * widthScale <= inWidth); NN_RET_CHECK(roiInfo[1] * heightScale <= inHeight); NN_RET_CHECK(roiInfo[2] * widthScale <= inWidth); NN_RET_CHECK(roiInfo[3] * heightScale <= inHeight); NN_RET_CHECK(roiInfo[0] <= roiInfo[2]); NN_RET_CHECK(roiInfo[1] <= roiInfo[3]); T_Roi wRoiStart = roiInfo[0] * widthScale; T_Roi hRoiStart = roiInfo[1] * heightScale; T_Roi wRoiEnd = roiInfo[2] * widthScale; T_Roi hRoiEnd = roiInfo[3] * heightScale; T_Roi roiWidth = std::max(static_cast(wRoiEnd - wRoiStart), 1.0f); T_Roi roiHeight = std::max(static_cast(hRoiEnd - hRoiStart), 1.0f); T_Roi wStepSize = roiWidth / static_cast(outWidth); T_Roi hStepSize = roiHeight / static_cast(outHeight); // if samplingRatio = 0, use adaptive value of ceil(roiWidth/outWidth), same for height uint32_t wSamplingRatio = widthSamplingRatio > 0 ? widthSamplingRatio : std::ceil(static_cast(wStepSize)); uint32_t hSamplingRatio = heightSamplingRatio > 0 ? heightSamplingRatio : std::ceil(static_cast(hStepSize)); int32_t numSamplingPoints = wSamplingRatio * hSamplingRatio; T_Roi wBinSize = wStepSize / static_cast(wSamplingRatio); T_Roi hBinSize = hStepSize / static_cast(hSamplingRatio); const T_Input* batchBase = inputData + batchId * inHeight * inWidth * inDepth; for (uint32_t i = 0; i < outHeight; i++) { for (uint32_t j = 0; j < outWidth; j++) { T_Roi wStart = wStepSize * j + wRoiStart; T_Roi wEnd = wStepSize * (j + 1) + wRoiStart; T_Roi hStart = hStepSize * i + hRoiStart; T_Roi hEnd = hStepSize * (i + 1) + hRoiStart; // initialize output to zero for (uint32_t k = 0; k < inDepth; k++) outPtr[k] = 0; // calculate the sum of the sampling points for (uint32_t yInd = 0; yInd < hSamplingRatio; yInd++) { for (uint32_t xInd = 0; xInd < wSamplingRatio; xInd++) { T_Roi y = hStart + hBinSize / 2 + hBinSize * yInd; T_Roi x = wStart + wBinSize / 2 + wBinSize * xInd; // bilinear interpolation of point (x,y) // w.r.t box [(x1,y1), (x1,y2), (x2,y1), (x2,y2)] uint32_t x1 = std::floor(static_cast(x)); uint32_t y1 = std::floor(static_cast(y)); uint32_t x2 = x1 + 1, y2 = y1 + 1; T_Roi dx1 = x - static_cast(x1); T_Roi dy1 = y - static_cast(y1); // dealing with out of bound samples if (x1 >= inWidth - 1) { x1 = x2 = inWidth - 1; dx1 = 0; } if (y1 >= inHeight - 1) { y1 = y2 = inHeight - 1; dy1 = 0; } T_Roi dx2 = 1.0f - dx1, dy2 = 1.0f - dy1; T_Roi ws[] = {dx2 * dy2, dx1 * dy2, dx2 * dy1, dx1 * dy1}; uint32_t offsets[] = {y1 * inWidth * inDepth + x1 * inDepth, y1 * inWidth * inDepth + x2 * inDepth, y2 * inWidth * inDepth + x1 * inDepth, y2 * inWidth * inDepth + x2 * inDepth}; for (uint32_t k = 0; k < inDepth; k++) { T_Input interpolation = 0; for (uint32_t c = 0; c < 4; c++) { interpolation += ws[c] * batchBase[offsets[c] + k]; } outPtr[k] += interpolation; } } } // take average for (uint32_t k = 0; k < inDepth; k++) outPtr[k] /= static_cast(numSamplingPoints); outPtr += inDepth; } } } return true; } template <> inline bool roiAlignNhwc(const uint8_t* inputData, const Shape& inputShape, const uint16_t* roiData, const Shape& roiShape, const int32_t* batchSplitData, const Shape& batchSplitShape, float heightStride, float widthStride, int32_t heightSamplingRatio, int32_t widthSamplingRatio, uint8_t* outputData, const Shape& outputShape) { NNTRACE_TRANS("RoiAlignQuant8"); constexpr float wScale = 1.0f / 255.0f; constexpr uint32_t kRoiDim = 4; const float heightScale = 1.0f / heightStride; const float widthScale = 1.0f / widthStride; uint32_t numBatches = getSizeOfDimension(inputShape, 0); uint32_t inHeight = getSizeOfDimension(inputShape, 1); uint32_t inWidth = getSizeOfDimension(inputShape, 2); uint32_t inDepth = getSizeOfDimension(inputShape, 3); uint32_t outHeight = getSizeOfDimension(outputShape, 1); uint32_t outWidth = getSizeOfDimension(outputShape, 2); uint32_t numRois = getSizeOfDimension(roiShape, 0); uint32_t roiInfoLength = getSizeOfDimension(roiShape, 1); uint8_t* outPtr = outputData; const uint16_t* roiDataEnd = roiData + numRois * roiInfoLength; uint32_t roiIndex = 0; for (const uint16_t* roiInfo = roiData; roiInfo < roiDataEnd; roiInfo += kRoiDim, roiIndex++) { uint32_t batchId = static_cast(batchSplitData[roiIndex]); float wRoiStart = static_cast(roiInfo[0]) * widthScale * 0.125f; float hRoiStart = static_cast(roiInfo[1]) * heightScale * 0.125f; float wRoiEnd = static_cast(roiInfo[2]) * widthScale * 0.125f; float hRoiEnd = static_cast(roiInfo[3]) * heightScale * 0.125f; // Check for malformed data // 1. invalid batch id // 2. Region out of bound: x1|x2|y1|y2 < 0 || x1|x2 > inWidth || y1|y2 > inHeight // 3. Invalid region: x2 < x1 || y2 < y1 NN_RET_CHECK_GE(batchId, 0); NN_RET_CHECK_LT(batchId, numBatches); NN_RET_CHECK(wRoiStart <= inWidth); NN_RET_CHECK(hRoiStart <= inHeight); NN_RET_CHECK(wRoiEnd <= inWidth); NN_RET_CHECK(hRoiEnd <= inHeight); NN_RET_CHECK_LE(wRoiStart, wRoiEnd); NN_RET_CHECK_LE(hRoiStart, hRoiEnd); float roiWidth = std::max(wRoiEnd - wRoiStart, 1.0f); float roiHeight = std::max(hRoiEnd - hRoiStart, 1.0f); float wStepSize = roiWidth / static_cast(outWidth); float hStepSize = roiHeight / static_cast(outHeight); // if samplingRatio = 0, use adaptive value of ceil(roiWidth/outWidth), same for height uint32_t wSamplingRatio = widthSamplingRatio > 0 ? widthSamplingRatio : std::ceil(wStepSize); uint32_t hSamplingRatio = heightSamplingRatio > 0 ? heightSamplingRatio : std::ceil(hStepSize); int32_t numSamplingPoints = wSamplingRatio * hSamplingRatio; float wBinSize = wStepSize / static_cast(wSamplingRatio); float hBinSize = hStepSize / static_cast(hSamplingRatio); float realMultiplier = inputShape.scale * wScale / outputShape.scale / numSamplingPoints; int32_t outputMultiplier = 0; int32_t outputShift = 0; if (!QuantizeMultiplierSmallerThanOne(realMultiplier, &outputMultiplier, &outputShift)) { return false; } const uint8_t* batchBase = inputData + batchId * inHeight * inWidth * inDepth; for (uint32_t i = 0; i < outHeight; i++) { for (uint32_t j = 0; j < outWidth; j++) { float wStart = wStepSize * j + wRoiStart; float wEnd = wStepSize * (j + 1) + wRoiStart; float hStart = hStepSize * i + hRoiStart; float hEnd = hStepSize * (i + 1) + hRoiStart; std::vector outTemp(inDepth, 0); // calculate the sum of the sampling points for (uint32_t yInd = 0; yInd < hSamplingRatio; yInd++) { for (uint32_t xInd = 0; xInd < wSamplingRatio; xInd++) { float y = hStart + hBinSize / 2 + hBinSize * yInd; float x = wStart + wBinSize / 2 + wBinSize * xInd; // bilinear interpolation of point (x,y) // w.r.t box [(x1,y1), (x1,y2), (x2,y1), (x2,y2)] uint32_t x1 = std::floor(x), y1 = std::floor(y); uint32_t x2 = x1 + 1, y2 = y1 + 1; float dx1 = x - static_cast(x1); float dy1 = y - static_cast(y1); // dealing with out of bound samples if (x1 >= inWidth - 1) { x1 = x2 = inWidth - 1; dx1 = 0; } if (y1 >= inHeight - 1) { y1 = y2 = inHeight - 1; dy1 = 0; } float dx2 = 1.0f - dx1, dy2 = 1.0f - dy1; float ws[] = {dx2 * dy2, dx1 * dy2, dx2 * dy1, dx1 * dy1}; uint32_t offsets[] = {y1 * inWidth * inDepth + x1 * inDepth, y1 * inWidth * inDepth + x2 * inDepth, y2 * inWidth * inDepth + x1 * inDepth, y2 * inWidth * inDepth + x2 * inDepth}; for (uint32_t k = 0; k < inDepth; k++) { int32_t interpolation = 0; for (uint32_t c = 0; c < 4; c++) { int32_t wQuant = static_cast(std::round(ws[c] / wScale)); interpolation += wQuant * (static_cast(batchBase[offsets[c] + k]) - inputShape.offset); } outTemp[k] += interpolation; } } } // take average and cast to output quantization for (uint32_t k = 0; k < inDepth; k++) { int32_t raw_out = tflite::MultiplyByQuantizedMultiplier( outTemp[k], outputMultiplier, -outputShift) + outputShape.offset; int32_t clamped_out = std::min(255, std::max(0, raw_out)); outPtr[k] = static_cast(clamped_out); } outPtr += inDepth; } } } return true; } template inline bool roiAlign(const T_Input* inputData, const Shape& inputShape, const T_Roi* roiData, const Shape& roiShape, const int32_t* batchSplitData, const Shape& batchSplitShape, float heightStride, float widthStride, int32_t heightSamplingRatio, int32_t widthSamplingRatio, bool useNchw, T_Input* outputData, const Shape& outputShape) { InputWithLayout input(useNchw); OutputWithLayout output(useNchw); NN_RET_CHECK(input.initialize(inputData, inputShape)); NN_RET_CHECK(output.initialize(outputData, outputShape)); NN_RET_CHECK(roiAlignNhwc(input.getNhwcBuffer(), input.getNhwcShape(), roiData, roiShape, batchSplitData, batchSplitShape, heightStride, widthStride, heightSamplingRatio, widthSamplingRatio, output.getNhwcBuffer(), output.getNhwcShape())); NN_RET_CHECK(output.commit()); return true; } } // namespace bool validate(const IOperationValidationContext* context) { NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs); NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs); std::vector inExpectedTypes; auto inputType = context->getInputType(kInputTensor); if (inputType == OperandType::TENSOR_FLOAT32) { inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32, OperandType::TENSOR_INT32, OperandType::INT32, OperandType::INT32, OperandType::FLOAT32, OperandType::FLOAT32, OperandType::INT32, OperandType::INT32, OperandType::BOOL}; } else if (inputType == OperandType::TENSOR_FLOAT16) { inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16, OperandType::TENSOR_INT32, OperandType::INT32, OperandType::INT32, OperandType::FLOAT16, OperandType::FLOAT16, OperandType::INT32, OperandType::INT32, OperandType::BOOL}; } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM) { inExpectedTypes = {OperandType::TENSOR_QUANT8_ASYMM, OperandType::TENSOR_QUANT16_ASYMM, OperandType::TENSOR_INT32, OperandType::INT32, OperandType::INT32, OperandType::FLOAT32, OperandType::FLOAT32, OperandType::INT32, OperandType::INT32, OperandType::BOOL}; } else { LOG(ERROR) << "Unsupported input tensor type for operation " << kOperationName; return false; } NN_RET_CHECK(validateInputTypes(context, inExpectedTypes)); NN_RET_CHECK(validateOutputTypes(context, {inputType})); return validateHalVersion(context, HalVersion::V1_2); } bool prepare(IOperationExecutionContext* context) { bool useNchw = context->getInputValue(kLayoutScalar); Shape input = context->getInputShape(kInputTensor); Shape roiShape = context->getInputShape(kRoiTensor); Shape batchSplitShape = context->getInputShape(kBatchSplitTensor); NN_RET_CHECK_EQ(getNumberOfDimensions(input), 4); NN_RET_CHECK_EQ(getNumberOfDimensions(roiShape), 2); uint32_t numBatches = getSizeOfDimension(input, 0); uint32_t inHeight = getSizeOfDimension(input, useNchw ? 2 : 1); uint32_t inWidth = getSizeOfDimension(input, useNchw ? 3 : 2); uint32_t inDepth = getSizeOfDimension(input, useNchw ? 1 : 3); uint32_t numRois = getSizeOfDimension(roiShape, 0); // Every dimension must be positive except for numRois. NN_RET_CHECK_GT(numBatches, 0); NN_RET_CHECK_GT(inHeight, 0); NN_RET_CHECK_GT(inWidth, 0); NN_RET_CHECK_GT(inDepth, 0); NN_RET_CHECK_EQ(getSizeOfDimension(roiShape, 1), 4); NN_RET_CHECK_EQ(getSizeOfDimension(batchSplitShape, 0), numRois); int32_t outputHeight = context->getInputValue(kOutputHeightScalar); int32_t outputWidth = context->getInputValue(kOutputWidthScalar); int32_t heightSamplingRatio = context->getInputValue(kHeightSamplingRatioScalar); int32_t widthSamplingRatio = context->getInputValue(kWidthSamplingRatioScalar); float heightScale, widthScale; if (context->getInputType(kInputTensor) == OperandType::TENSOR_FLOAT16) { heightScale = context->getInputValue<_Float16>(kHeightStrideSalar); widthScale = context->getInputValue<_Float16>(kWidthStrideScalar); } else { heightScale = context->getInputValue(kHeightStrideSalar); widthScale = context->getInputValue(kWidthStrideScalar); } NN_RET_CHECK_GT(outputHeight, 0); NN_RET_CHECK_GT(outputWidth, 0); NN_RET_CHECK_GT(heightScale, 0); NN_RET_CHECK_GT(widthScale, 0); // Sampling ratio can set to 0 for adaptive value. NN_RET_CHECK_GE(heightSamplingRatio, 0); NN_RET_CHECK_GE(widthSamplingRatio, 0); if (roiShape.type == OperandType::TENSOR_QUANT16_ASYMM) { NN_RET_CHECK_EQ(roiShape.scale, 0.125f); NN_RET_CHECK_EQ(roiShape.offset, 0); } Shape output = context->getOutputShape(kOutputTensor); output.type = input.type; if (useNchw) { output.dimensions = {numRois, inDepth, static_cast(outputHeight), static_cast(outputWidth)}; } else { output.dimensions = {numRois, static_cast(outputHeight), static_cast(outputWidth), inDepth}; } return context->setOutputShape(kOutputTensor, output); } bool execute(IOperationExecutionContext* context) { // Bypass execution in the case of zero-sized input. if (getNumberOfElements(context->getInputShape(kRoiTensor)) == 0) return true; switch (context->getInputType(kInputTensor)) { case OperandType::TENSOR_FLOAT16: return roiAlign(context->getInputBuffer<_Float16>(kInputTensor), context->getInputShape(kInputTensor), context->getInputBuffer<_Float16>(kRoiTensor), context->getInputShape(kRoiTensor), context->getInputBuffer(kBatchSplitTensor), context->getInputShape(kBatchSplitTensor), context->getInputValue<_Float16>(kHeightStrideSalar), context->getInputValue<_Float16>(kWidthStrideScalar), context->getInputValue(kHeightSamplingRatioScalar), context->getInputValue(kWidthSamplingRatioScalar), context->getInputValue(kLayoutScalar), context->getOutputBuffer<_Float16>(kOutputTensor), context->getOutputShape(kOutputTensor)); case OperandType::TENSOR_FLOAT32: return roiAlign(context->getInputBuffer(kInputTensor), context->getInputShape(kInputTensor), context->getInputBuffer(kRoiTensor), context->getInputShape(kRoiTensor), context->getInputBuffer(kBatchSplitTensor), context->getInputShape(kBatchSplitTensor), context->getInputValue(kHeightStrideSalar), context->getInputValue(kWidthStrideScalar), context->getInputValue(kHeightSamplingRatioScalar), context->getInputValue(kWidthSamplingRatioScalar), context->getInputValue(kLayoutScalar), context->getOutputBuffer(kOutputTensor), context->getOutputShape(kOutputTensor)); case OperandType::TENSOR_QUANT8_ASYMM: return roiAlign(context->getInputBuffer(kInputTensor), context->getInputShape(kInputTensor), context->getInputBuffer(kRoiTensor), context->getInputShape(kRoiTensor), context->getInputBuffer(kBatchSplitTensor), context->getInputShape(kBatchSplitTensor), context->getInputValue(kHeightStrideSalar), context->getInputValue(kWidthStrideScalar), context->getInputValue(kHeightSamplingRatioScalar), context->getInputValue(kWidthSamplingRatioScalar), context->getInputValue(kLayoutScalar), context->getOutputBuffer(kOutputTensor), context->getOutputShape(kOutputTensor)); default: NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName; } } } // namespace roi_align NN_REGISTER_OPERATION(ROI_ALIGN, roi_align::kOperationName, roi_align::validate, roi_align::prepare, roi_align::execute, .allowZeroSizedInput = true); } // namespace nn } // namespace android