third_party/blink/renderer/modules/ml/ml_context.idl

// Copyright 2022 The Chromium Authors
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

// This will be a shared interface by two APIs:
// - The Model Loader API,
//     https://github.com/webmachinelearning/model-loader/blob/main/explainer.md
// - The WebNN API,
//     https://github.com/webmachinelearning/webnn/blob/main/explainer.md
// We follow WebNN spec's definition,
//   - https://webmachinelearning.github.io/webnn/#api-mlcontext
// In the future, this interface may be extended when WebNN is implemented so
// a dictionary may be insufficient.

dictionary MLContextLostInfo {
  DOMString message;
};

dictionary MLRankRange {
  unsigned long min;
  unsigned long max;
};

dictionary MLTensorLimits {
  sequence<DOMString> dataTypes;
  MLRankRange rankRange;
};

dictionary MLBinarySupportLimits {
  MLTensorLimits a;
  MLTensorLimits b;
  MLTensorLimits output;
};

dictionary MLBatchNormalizationSupportLimits {
  MLTensorLimits input;
  MLTensorLimits mean;
  MLTensorLimits variance;
  MLTensorLimits scale;
  MLTensorLimits bias;
  MLTensorLimits output;
};

dictionary MLConcatSupportLimits {
  MLTensorLimits inputs;
  MLTensorLimits output;
};

dictionary MLConv2dSupportLimits {
  MLTensorLimits input;
  MLTensorLimits filter;
  MLTensorLimits bias;
  MLTensorLimits output;
};

dictionary MLGatherSupportLimits {
  MLTensorLimits input;
  MLTensorLimits indices;
  MLTensorLimits output;
};

dictionary MLGemmSupportLimits {
  MLTensorLimits a;
  MLTensorLimits b;
  MLTensorLimits c;
  MLTensorLimits output;
};

dictionary MLGruSupportLimits {
  MLTensorLimits input;
  MLTensorLimits weight;
  MLTensorLimits recurrentWeight;
  MLTensorLimits bias;
  MLTensorLimits recurrentBias;
  MLTensorLimits initialHiddenState;
  MLTensorLimits output0;
  MLTensorLimits output1;
};

dictionary MLGruCellSupportLimits {
  MLTensorLimits input;
  MLTensorLimits weight;
  MLTensorLimits recurrentWeight;
  MLTensorLimits hiddenState;
  MLTensorLimits bias;
  MLTensorLimits recurrentBias;
  MLTensorLimits output;
};

dictionary MLLstmSupportLimits {
  MLTensorLimits input;
  MLTensorLimits weight;
  MLTensorLimits recurrentWeight;
  MLTensorLimits bias;
  MLTensorLimits recurrentBias;
  MLTensorLimits peepholeWeight;
  MLTensorLimits initialHiddenState;
  MLTensorLimits initialCellState;
  MLTensorLimits output0;
  MLTensorLimits output1;
  MLTensorLimits output2;
};

dictionary MLLstmCellSupportLimits {
  MLTensorLimits input;
  MLTensorLimits weight;
  MLTensorLimits recurrentWeight;
  MLTensorLimits hiddenState;
  MLTensorLimits cellState;
  MLTensorLimits bias;
  MLTensorLimits recurrentBias;
  MLTensorLimits peepholeWeight;
  MLTensorLimits output0;
  MLTensorLimits output1;
};

dictionary MLNormalizationSupportLimits {
  MLTensorLimits input;
  MLTensorLimits scale;
  MLTensorLimits bias;
  MLTensorLimits output;
};

dictionary MLPreluSupportLimits {
  MLTensorLimits input;
  MLTensorLimits slope;
  MLTensorLimits output;
};

dictionary MLQuantizeDequantizeLinearSupportLimits {
  MLTensorLimits input;
  MLTensorLimits scale;
  MLTensorLimits zeroPoint;
  MLTensorLimits output;
};

dictionary MLLogicalNotSupportLimits {
  MLTensorLimits a;
  MLTensorLimits output;
};

dictionary MLSingleInputSupportLimits {
  MLTensorLimits input;
  MLTensorLimits output;
};

dictionary MLScatterSupportLimits {
  MLTensorLimits input;
  MLTensorLimits indices;
  MLTensorLimits updates;
  MLTensorLimits output;
};

dictionary MLSplitSupportLimits {
  MLTensorLimits input;
  MLTensorLimits outputs;
};

dictionary MLWhereSupportLimits {
  MLTensorLimits condition;
  MLTensorLimits trueValue;
  MLTensorLimits falseValue;
  MLTensorLimits output;
};

dictionary MLOpSupportLimits {
  MLInputOperandLayout preferredInputLayout;
  [EnforceRange] unsigned long long maxTensorByteLength;
  MLTensorLimits input;
  MLTensorLimits constant;
  MLTensorLimits output;

  MLSingleInputSupportLimits argMin;
  MLSingleInputSupportLimits argMax;
  MLBatchNormalizationSupportLimits batchNormalization;
  MLSingleInputSupportLimits cast;
  MLSingleInputSupportLimits clamp;
  MLConcatSupportLimits concat;
  MLConv2dSupportLimits conv2d;
  MLConv2dSupportLimits convTranspose2d;
  MLSingleInputSupportLimits cumulativeSum;
  MLQuantizeDequantizeLinearSupportLimits dequantizeLinear;

  // Element-wise binary ops.
  MLBinarySupportLimits add;
  MLBinarySupportLimits sub;
  MLBinarySupportLimits mul;
  MLBinarySupportLimits div;
  MLBinarySupportLimits max;
  MLBinarySupportLimits min;
  MLBinarySupportLimits pow;

  // Element-wise logical ops.
  MLBinarySupportLimits equal;
  MLBinarySupportLimits greater;
  MLBinarySupportLimits greaterOrEqual;
  MLBinarySupportLimits lesser;
  MLBinarySupportLimits lesserOrEqual;
  MLBinarySupportLimits notEqual;
  MLBinarySupportLimits logicalAnd;
  MLBinarySupportLimits logicalOr;
  MLBinarySupportLimits logicalXor;
  MLLogicalNotSupportLimits logicalNot;
  MLLogicalNotSupportLimits isNaN;
  MLLogicalNotSupportLimits isInfinite;

  // Element-wise unary ops.
  MLSingleInputSupportLimits abs;
  MLSingleInputSupportLimits ceil;
  MLSingleInputSupportLimits cos;
  MLSingleInputSupportLimits erf;
  MLSingleInputSupportLimits exp;
  MLSingleInputSupportLimits floor;
  MLSingleInputSupportLimits identity;
  MLSingleInputSupportLimits log;
  MLSingleInputSupportLimits neg;
  MLSingleInputSupportLimits reciprocal;
  MLSingleInputSupportLimits roundEven;
  MLSingleInputSupportLimits sign;
  MLSingleInputSupportLimits sin;
  MLSingleInputSupportLimits sqrt;
  MLSingleInputSupportLimits tan;

  MLSingleInputSupportLimits elu;
  MLSingleInputSupportLimits expand;
  MLGatherSupportLimits gather;
  MLGatherSupportLimits gatherElements;
  MLGatherSupportLimits gatherND;
  MLSingleInputSupportLimits gelu;
  MLGemmSupportLimits gemm;
  MLGruSupportLimits gru;
  MLGruCellSupportLimits gruCell;
  MLSingleInputSupportLimits hardSigmoid;
  MLSingleInputSupportLimits hardSwish;
  MLNormalizationSupportLimits instanceNormalization;
  MLNormalizationSupportLimits layerNormalization;
  MLSingleInputSupportLimits leakyRelu;
  MLSingleInputSupportLimits linear;
  MLLstmSupportLimits lstm;
  MLLstmCellSupportLimits lstmCell;
  MLBinarySupportLimits matmul;
  MLSingleInputSupportLimits pad;
  MLPreluSupportLimits prelu;
  MLQuantizeDequantizeLinearSupportLimits quantizeLinear;

  // Pool2d.
  MLSingleInputSupportLimits averagePool2d;
  MLSingleInputSupportLimits l2Pool2d;
  MLSingleInputSupportLimits maxPool2d;

  // Reduction ops.
  MLSingleInputSupportLimits reduceL1;
  MLSingleInputSupportLimits reduceL2;
  MLSingleInputSupportLimits reduceLogSum;
  MLSingleInputSupportLimits reduceLogSumExp;
  MLSingleInputSupportLimits reduceMax;
  MLSingleInputSupportLimits reduceMean;
  MLSingleInputSupportLimits reduceMin;
  MLSingleInputSupportLimits reduceProduct;
  MLSingleInputSupportLimits reduceSum;
  MLSingleInputSupportLimits reduceSumSquare;

  MLSingleInputSupportLimits relu;
  MLSingleInputSupportLimits resample2d;
  MLSingleInputSupportLimits reshape;
  MLSingleInputSupportLimits reverse;
  MLScatterSupportLimits scatterElements;
  MLScatterSupportLimits scatterND;
  MLSingleInputSupportLimits sigmoid;
  MLSingleInputSupportLimits slice;
  MLSingleInputSupportLimits softmax;
  MLSingleInputSupportLimits softplus;
  MLSingleInputSupportLimits softsign;
  MLSplitSupportLimits split;
  MLSingleInputSupportLimits tanh;
  MLSingleInputSupportLimits tile;
  MLSingleInputSupportLimits transpose;
  MLSingleInputSupportLimits triangular;
  MLWhereSupportLimits where;
};

typedef record<USVString, MLTensor> MLNamedTensors;

[
  RuntimeEnabled=MachineLearningNeuralNetwork,
  SecureContext,
  Exposed=(Window, Worker)
] interface MLContext {
  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState
  ] readonly attribute Promise<MLContextLostInfo> lost;

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] void destroy();

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] Promise<MLTensor> createTensor(MLTensorDescriptor descriptor);

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] Promise<MLTensor> createConstantTensor(
        MLOperandDescriptor descriptor,
        AllowSharedBufferSource sourceData);

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] Promise<MLTensor> createExportableTensor(MLTensorDescriptor descriptor, GPUDevice device);

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] void writeTensor(MLTensor dstTensor, AllowSharedBufferSource srcData);

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] Promise<ArrayBuffer> readTensor(
        MLTensor sourceTensor);

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] Promise<undefined> readTensor(
        MLTensor sourceTensor,
        AllowSharedBufferSource destinationData);

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException,
    Measure
  ] void dispatch(
      MLGraph graph, MLNamedTensors inputs, MLNamedTensors outputs);
  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState
  ] MLOpSupportLimits opSupportLimits();

  [
    RuntimeEnabled=MachineLearningNeuralNetwork,
    CallWith=ScriptState,
    RaisesException
  ] Promise<GPUBuffer> exportToGPU(MLTensor tensor);
};