/*
 * SPDX-License-Identifier: Apache-2.0
 */

#include "onnx/defs/math/utils.h"

#include <string>

namespace ONNX_NAMESPACE {
namespace defs {
namespace math {
namespace utils {

static constexpr const char* TopK_ver11_doc = R"DOC(
Retrieve the top-K largest or smallest elements along a specified axis. Given an input tensor of
shape [a_0, a_1, ..., a_{n-1}] and integer argument k, return two outputs:

* Value tensor of shape [a_0, a_1, ..., a_{axis-1}, k, a_{axis+1}, ... a_{n-1}]
  which contains the values of the top k elements along the specified axis
* Index tensor of shape [a_0, a_1, ..., a_{axis-1}, k, a_{axis+1}, ... a_{n-1}] which
  contains the indices of the top k elements (original indices from the input
  tensor).

* If "largest" is 1 (the default value) then the k largest elements are returned.
* If "sorted" is 1 (the default value) then the resulting k elements will be sorted.
* If "sorted" is 0, order of returned 'Values' and 'Indices' are undefined.

Given two equivalent values, this operator uses the indices along the axis as
a tiebreaker. That is, the element with the lower index will appear first.
)DOC";

std::function<void(OpSchema&)> TopKOpGenerator(const std::vector<std::string>& allowed_types) {
  return [=](OpSchema& schema) {
    schema.SetDoc(TopK_ver11_doc)
        .Input(
            0,
            "X",
            "Tensor of shape [a_0, a_1, ..., a_{n-1}]",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Input(
            1,
            "K",
            "A 1-D tensor containing a single positive value corresponding to the number of top elements to retrieve",
            "tensor(int64)",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "Values",
            "Tensor of shape [a_0, a_1, ..., a_{axis-1}, k, a_{axis+1}, ... a_{n-1}] "
            "containing top K values from the input tensor",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Output(
            1,
            "Indices",
            "Tensor of shape [a_0, a_1, ..., a_{axis-1}, k, a_{axis+1}, ... a_{n-1}] "
            "containing the corresponding input tensor indices for the top K "
            "values.",
            "I",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .TypeConstraint("T", allowed_types, "Constrain input and output types to numeric tensors.")
        .TypeConstraint("I", {"tensor(int64)"}, "Constrain index tensor to int64")
        .Attr(
            "axis",
            "Dimension on which to do the sort. Negative value means counting dimensions "
            "from the back. Accepted range is [-r, r-1] where r = rank(input).",
            AttributeProto::INT,
            static_cast<int64_t>(-1))
        .Attr(
            "largest",
            "Whether to return the top-K largest or smallest elements.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .Attr("sorted", "Whether to return the elements in sorted order.", AttributeProto::INT, static_cast<int64_t>(1))
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          // Type inference:
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          updateOutputElemType(ctx, 1, TensorProto::INT64);
          // Shape inference:
          if (!hasInputShape(ctx, 0))
            return;
          auto& input_shape = getInputShape(ctx, 0);
          int64_t rank = input_shape.dim_size();
          int64_t axis = getAttribute(ctx, "axis", -1);
          if (axis < 0)
            axis += rank;
          if (axis < 0 || axis >= rank) {
            fail_shape_inference("Invalid value for attribute axis");
          }

          const auto& axis_dim = input_shape.dim(static_cast<int>(axis));
          const auto* k = ctx.getInputData(1);

          // Infer output shape if:
          // (1) 'K' is available
          // (2) axis_dim has dim value
          // Otherwise cannot reliably compute output shape as axis dim value is
          // unknown and hence cannot determine if axis dim value >= k (which
          // should be enforced)
          if (nullptr != k && axis_dim.has_dim_value()) {
            int64_t k_value = 0;
            if (k->dims_size() != 1 || k->dims(0) != 1) {
              fail_shape_inference("K input must be a one-dimensional tensor of size 1.");
            }
            if (k->data_type() == TensorProto::INT64) {
              const auto data = ParseData<int64_t>(k);
              k_value = data[0];
            } else {
              fail_shape_inference("K input must be of type int64.");
            }
            if (axis_dim.dim_value() < k_value) {
              fail_shape_inference("Axis has less than the requested k elements.");
            }

            TensorShapeProto result_shape = input_shape;
            result_shape.mutable_dim(static_cast<int>(axis))->set_dim_value(k_value);

            updateOutputShape(ctx, 0, result_shape);
            updateOutputShape(ctx, 1, result_shape);

            return;
          }

          // Infer output shapes' rank in any case
          auto* output_shape_0 = getOutputShape(ctx, 0);
          auto* output_shape_1 = getOutputShape(ctx, 1);
          for (int i = 0; i < input_shape.dim_size(); ++i) {
            output_shape_0->add_dim();
            output_shape_1->add_dim();
          }

          return;
        });
  };
}

int MathOpTwoIntegers(const std::string& op_type, int a, int b) {
  if (op_type == "Add") {
    return a + b;
  } else if (op_type == "Sub") {
    return a - b;
  } else if (op_type == "Mul") {
    return a * b;
  }
  fail_shape_inference("Wrong op_type name for running propagation: ", op_type);
}

void MatMulShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, int input1Idx, int input2Idx) {
  if (!hasInputShape(ctx, input1Idx) || !hasInputShape(ctx, input2Idx)) {
    return;
  }

  const auto shape0 = ctx.getInputType(input1Idx)->tensor_type().shape();
  const auto shape1 = ctx.getInputType(input2Idx)->tensor_type().shape();

  if (shape0.dim_size() == 0 || shape1.dim_size() == 0) {
    fail_shape_inference("Input tensors of wrong rank (0).");
  }

  ONNX_NAMESPACE::TensorShapeProto shapeL, shapeR;

  // First promote each shape to at least rank-2. This logic is
  // specific to matmul, not generic broadcasting.
  {
    if (shape0.dim_size() == 1) {
      shapeL.add_dim()->set_dim_value(1);
      *shapeL.add_dim() = shape0.dim(0);
    } else {
      *shapeL.mutable_dim() = shape0.dim();
    }
    if (shape1.dim_size() == 1) {
      *shapeR.add_dim() = shape1.dim(0);
      shapeR.add_dim()->set_dim_value(1);
    } else {
      *shapeR.mutable_dim() = shape1.dim();
    }
  }

  // Check for compatible matrix multiply dimensions
  {
    const auto& dimL = shapeL.dim(shapeL.dim_size() - 1);
    const auto& dimR = shapeR.dim(shapeR.dim_size() - 2);
    if (dimL.has_dim_value() && dimR.has_dim_value() && dimL.dim_value() != dimR.dim_value()) {
      fail_shape_inference("Incompatible dimensions for matrix multiplication");
    }
  }

  ONNX_NAMESPACE::TensorShapeProto resultShape;

  // Now call out to generic multidimensional broadcasting for
  // the broadcastable prefixes.
  {
    ONNX_NAMESPACE::TensorShapeProto prefixShapeL, prefixShapeR;
    for (int i = 0; i < shapeL.dim_size() - 2; ++i) {
      *prefixShapeL.add_dim() = shapeL.dim(i);
    }
    for (int i = 0; i < shapeR.dim_size() - 2; ++i) {
      *prefixShapeR.add_dim() = shapeR.dim(i);
    }
    bidirectionalBroadcastShapeInference(prefixShapeL, prefixShapeR, resultShape);
  }

  // Back to matmul-specific. Add the trailing dimensions back in.
  {
    if (shape0.dim_size() != 1) {
      *resultShape.add_dim() = shapeL.dim(shapeL.dim_size() - 2);
    }
    if (shape1.dim_size() != 1) {
      *resultShape.add_dim() = shapeR.dim(shapeR.dim_size() - 1);
    }
  }

  *ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape() = resultShape;
}

void QLinearMatMulShapeInference(ONNX_NAMESPACE::InferenceContext& ctx) {
  auto a_type = ctx.getInputType(0);
  auto b_type = ctx.getInputType(3);
  if (nullptr == a_type || nullptr == b_type || a_type->value_case() != ONNX_NAMESPACE::TypeProto::kTensorType ||
      b_type->value_case() != ONNX_NAMESPACE::TypeProto::kTensorType) {
    fail_type_inference("inputs are expected to have tensor type.");
  }

  auto a_zero_point_type = ctx.getInputType(2);
  if (nullptr == a_zero_point_type ||
      a_zero_point_type->tensor_type().elem_type() != a_type->tensor_type().elem_type()) {
    fail_type_inference("input and zero_point pair is expected to have be same type.");
  }

  auto b_zero_point_type = ctx.getInputType(5);
  if (nullptr == b_zero_point_type ||
      b_zero_point_type->tensor_type().elem_type() != b_type->tensor_type().elem_type()) {
    fail_type_inference("input and zero_point pair is expected to have same type.");
  }

  propagateElemTypeFromInputToOutput(ctx, 7, 0);

  MatMulShapeInference(ctx, 0, 3);
}

const char* QLinearMatMulDoc() {
  static constexpr const char* QLinearMatMul_doc = R"DOC(
Matrix product that behaves like [numpy.matmul](https://numpy.org/doc/stable/reference/generated/numpy.matmul.html).
It consumes two quantized input tensors, their scales and zero points, scale and zero point of output,
and computes the quantized output. The quantization formula is y = saturate((x / y_scale) + y_zero_point).
For (x / y_scale), it is rounding to nearest ties to even. Refer to https://en.wikipedia.org/wiki/Rounding for details.
Scale and zero point must have same shape. They must be either scalar (per tensor) or N-D tensor
(per row for 'a' and per column for 'b'). Scalar refers to per tensor quantization whereas N-D refers to per row
or per column quantization. If the input is 2D of shape [M, K] then zero point and scale tensor may be
an M element vector [v_1, v_2, ..., v_M] for per row quantization and K element vector of shape [v_1, v_2, ..., v_K]
for per column quantization. If the input is N-D tensor with shape [D1, D2, M, K] then zero point and scale tensor may
have shape [D1, D2, M, 1] for per row quantization and shape [D1, D2, 1, K] for per column quantization.
Production must never overflow, and accumulation may overflow if and only if in 32 bits.
)DOC";
  return QLinearMatMul_doc;
}

} // namespace utils
} // namespace math
} // namespace defs
} // namespace ONNX_NAMESPACE