onnx 1.17.0__cp311-cp311-win32.whl → 1.19.0__cp311-cp311-win32.whl

onnx/defs/math/defs.cc

@@ -3,10 +3,9 @@
  */
 
 #include <algorithm>
-#include <functional>
+#include <map>
 
 #include "onnx/common/assertions.h"
-#include "onnx/defs/data_type_utils.h"
 #include "onnx/defs/function.h"
 #include "onnx/defs/math/utils.h"
 #include "onnx/defs/schema.h"
@@ -14,7 +13,7 @@
 
 namespace ONNX_NAMESPACE {
 
-inline void MathOpDataPropagator(DataPropagationContext& ctx, std::string op_type) {
+static void MathOpDataPropagator(DataPropagationContext& ctx, const std::string& op_type) {
   const auto input_0 = ctx.getInputData(0);
   const auto input_1 = ctx.getInputData(1);
   if (input_0 == nullptr || input_1 == nullptr) {
@@ -27,7 +26,8 @@ inline void MathOpDataPropagator(DataPropagationContext& ctx, std::string op_typ
     fail_shape_inference("Invalid rank for ", op_type, " broadcasting: (", size_0, ") vs (", size_1, ").");
   }
   TensorShapeProto tsp;
-  for (int i = 0; i < std::max(size_0, size_1); ++i) {
+  int size_out = size_0 == 1 ? size_1 : size_0;
+  for (int i = 0; i < size_out; ++i) {
     auto& input_dim_0 = input_0->dim(size_0 == 1 ? 0 : i);
     auto& input_dim_1 = input_1->dim(size_1 == 1 ? 0 : i);
     if (input_dim_0.has_dim_value() && input_dim_1.has_dim_value()) {
@@ -41,18 +41,19 @@ inline void MathOpDataPropagator(DataPropagationContext& ctx, std::string op_typ
   ctx.addOutputData(0, std::move(tsp));
 }
 
-std::function<void(OpSchema&)> MathDocGenerator(const char* name) {
+static std::function<void(OpSchema&)> MathDocGenerator(const char* name) {
   return [=](OpSchema& schema) {
     std::string doc;
-    POPULATE_OP_DOC_STR(doc = R"DOC(
+    POPULATE_OP_DOC_STR(
+        doc = R"DOC(
 Performs element-wise binary {name} (with Numpy-style broadcasting support).
 
 {broadcast_doc}
 
 (Opset 14 change): Extend supported types to include uint8, int8, uint16, and int16.
 )DOC";
-                        ReplaceAll(doc, "{name}", name);
-                        ReplaceAll(doc, "{broadcast_doc}", GenerateBroadcastingDocMul().c_str()););
+        ReplaceAll(doc, "{name}", name);
+        ReplaceAll(doc, "{broadcast_doc}", GenerateBroadcastingDocMul().c_str()););
     schema.SetDoc(doc);
     schema.Input(0, "A", "First operand.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable);
     schema.Input(1, "B", "Second operand.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable);
@@ -93,19 +94,23 @@ ONNX_OPERATOR_SET_SCHEMA(
         .PartialDataPropagationFunction([](DataPropagationContext& ctx) { MathOpDataPropagator(ctx, "Sub"); }));
 
 static const char* Mod_doc = R"DOC(
-  Performs element-wise binary modulus (with Numpy-style broadcasting support).
-  The sign of the remainder is the same as that of the Divisor.
-
-  Mod operator can also behave like C fmod() or numpy.fmod. In this case, the sign of the remainder however, will be the same as the Dividend
-  (in contrast to integer mod). To force a behavior like numpy.fmod() an 'fmod' Attribute is provided.
-  This attribute is set to 0 by default causing the behavior to be like integer mod.
-  Setting this attribute to 1 causes the remainder to be calculated similar to that of numpy.fmod().
-
-  If the input type is floating point, then `fmod` attribute must be set to 1.
-
-  In case of dividend being zero, the results will be platform dependent.
-
-  This operator supports **multidirectional (i.e., Numpy-style) broadcasting**; for more details please check [the doc](Broadcasting.md).
+Performs an element-wise binary modulo operation.
+The semantics and supported data types depend on the value of the `fmod` attribute which must be `0` (default), or `1`.
+
+If the `fmod` attribute is set to `0`, `T` is constrained to integer data types and the semantics follow that of the Python `%`-operator.
+The sign of the result is that of the divisor.
+
+If `fmod` is set to `1`, the behavior of this operator follows that of the `fmod` function in C and `T` is constrained to floating point data types.
+The result of this operator is the remainder of the division operation `x / y` where `x` and `y` are respective elements of `A` and `B`. The result is exactly the value `x - n * y`, where `n` is `x / y` with its fractional part truncated.
+The returned value has the same sign as `x` (except if `x` is `-0`) and is less or equal to `|y|` in magnitude.
+The following special cases apply when `fmod` is set to `1`:
+- If `x` is `-0` and `y` is greater than zero, either `+0` or `-0` may be returned.
+- If `x` is `±∞` and `y` is not `NaN`, `NaN` is returned.
+- If `y` is `±0` and `x` is not `NaN`, `NaN` should be returned.
+- If `y` is `±∞` and `x` is finite, `x` is returned.
+- If either argument is `NaN`, `NaN` is returned.
+
+This operator supports **multidirectional (i.e., NumPy-style) broadcasting**; for more details please check [the doc](Broadcasting.md).
 )DOC";
 
 ONNX_OPERATOR_SET_SCHEMA(
@@ -420,8 +425,8 @@ ONNX_OPERATOR_SET_SCHEMA(
     OpSchema()
         .Attr("alpha", "Coefficient of ELU.", AttributeProto::FLOAT, 1.0f)
         .SetDoc(Elu_ver22_doc)
-        .Input(0, "X", "1D input tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
-        .Output(0, "Y", "1D output tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
+        .Input(0, "X", "Input tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
+        .Output(0, "Y", "Output tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
         .TypeConstraint("T", OpSchema::all_float_types_ir4(), "Constrain input and output types to float tensors.")
         .TypeAndShapeInferenceFunction(propagateShapeAndTypeFromFirstInput)
         .FunctionBody(
@@ -481,31 +486,7 @@ max(0,x) + min(0,alpha*(exp(x/alpha)-1))
 
 static float celu_default_alpha = 1.0;
 
-TensorProto ToDimensionOneFloatTensor(float value) {
-  auto t = ToTensor(std::vector<float>({value}));
-  t.add_dims(1);
-  return t;
-}
-
-TensorProto ToDimensionOneTensor(int32_t value) {
-  auto t = ToTensor(std::vector<int32_t>({value}));
-  t.add_dims(1);
-  return t;
-}
-
-TensorProto ToDimensionOneInt64Tensor(int64_t value) {
-  auto t = ToTensor(std::vector<int64_t>({value}));
-  t.add_dims(1);
-  return t;
-}
-
-TensorProto ToDimensionOneInt64Tensor(std::vector<int64_t> value) {
-  auto t = ToTensor(value);
-  t.add_dims(value.size());
-  return t;
-}
-
-bool BuildContextDependentFunctionBodyCelu(
+static bool BuildContextDependentFunctionBodyCelu(
     const FunctionBodyBuildContext& ctx,
     const OpSchema& schema,
     FunctionProto& functionProto) {
@@ -549,7 +530,7 @@ to the tensor elementwise.
 
 static std::string gelu_default_approx = "none";
 
-bool BuildContextDependentFunctionBodyGelu(
+static bool BuildContextDependentFunctionBodyGelu(
     const FunctionBodyBuildContext& ctx,
     const OpSchema& schema,
     FunctionProto& functionProto) {
@@ -621,6 +602,35 @@ ONNX_OPERATOR_SET_SCHEMA(
         .SetContextDependentFunctionBodyBuilder(BuildContextDependentFunctionBodyGelu)
         .TypeAndShapeInferenceFunction(propagateShapeAndTypeFromFirstInput));
 
+static const char* Swish_ver24_doc = R"DOC(
+Swish function takes one input data (Tensor<T>) and produces one output data (Tensor<T>) of the same shape,
+where $Swish(x) = x * sigmoid(alpha * x)$.
+)DOC";
+
+ONNX_OPERATOR_SET_SCHEMA(
+    Swish,
+    24,
+    OpSchema()
+        .Attr("alpha", "Coefficient to multiply with input before sigmoid.", AttributeProto::FLOAT, 1.0f)
+        .SetDoc(Swish_ver24_doc)
+        .Input(0, "X", "Input tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
+        .Output(0, "Y", "Output tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
+        .TypeConstraint(
+            "T",
+            {"tensor(float16)", "tensor(float)", "tensor(bfloat16)", "tensor(double)"},
+            "Constrain input and output types to float tensors.")
+        .TypeAndShapeInferenceFunction(propagateShapeAndTypeFromFirstInput)
+        .FunctionBody(
+            R"ONNX(
+            {
+                Alpha = Constant <value_float: float = @alpha>()
+                AlphaCast = CastLike (Alpha, X)
+                AlphaMulX = Mul (AlphaCast, X)
+                SigmoidAlphaMulX = Sigmoid(AlphaMulX)
+                Y = Mul (X, SigmoidAlphaMulX)
+            }
+            )ONNX"));
+
 static const char* Exp_ver13_doc = R"DOC(
 Calculates the exponential of the given input tensor, element-wise.
 )DOC";
@@ -873,16 +883,17 @@ ONNX_OPERATOR_SET_SCHEMA(
 
 // Generate opschema for element-wise ops. Leaves type constraint "T"
 // unspecified.
-std::function<void(OpSchema&)> ElementwiseMultiOpDocGenerator(const char* name) {
+static std::function<void(OpSchema&)> ElementwiseMultiOpDocGenerator(const char* name) {
   return [=](OpSchema& schema) {
     std::string doc;
-    POPULATE_OP_DOC_STR(doc = R"DOC(
+    POPULATE_OP_DOC_STR(
+        doc = R"DOC(
 Element-wise {name} of each of the input tensors (with Numpy-style broadcasting support).
 All inputs and outputs must have the same data type.
 {broadcast_doc}
 )DOC";
-                        ReplaceAll(doc, "{name}", name);
-                        ReplaceAll(doc, "{broadcast_doc}", GenerateBroadcastingDocMul().c_str()););
+        ReplaceAll(doc, "{name}", name);
+        ReplaceAll(doc, "{broadcast_doc}", GenerateBroadcastingDocMul().c_str()););
     schema.SetDoc(doc);
     schema.Input(
         0,
@@ -956,9 +967,11 @@ static const char* Clip_ver13_doc = R"DOC(
 Clip operator limits the given input within an interval. The interval is
 specified by the inputs 'min' and 'max'. They default to
 numeric_limits::lowest() and numeric_limits::max(), respectively.
+When 'min' is greater than 'max', the clip operator sets all the 'input' values to
+the value of 'max'. Thus, this is equivalent to 'Min(max, Max(input, min))'.
 )DOC";
 
-bool BuildContextDependentFunctionBodyClip(
+static bool BuildContextDependentFunctionBodyClip(
     const FunctionBodyBuildContext& ctx,
     const OpSchema& schema,
     FunctionProto& functionProto) {
@@ -1035,11 +1048,12 @@ ONNX_OPERATOR_SET_SCHEMA(
         .SetContextDependentFunctionBodyBuilder(BuildContextDependentFunctionBodyClip)
         .TypeAndShapeInferenceFunction(propagateShapeAndTypeFromFirstInput));
 
-std::function<void(OpSchema&)>
+static std::function<void(OpSchema&)>
 SoftmaxFamilyDocGenerator(const char* name, const char* description, const char* equation) {
   return [=](OpSchema& schema) {
     std::string doc;
-    POPULATE_OP_DOC_STR(doc = R"DOC(
+    POPULATE_OP_DOC_STR(
+        doc = R"DOC(
 The operator computes the {description} values for the given input:
 
  {equation}
@@ -1048,16 +1062,17 @@ The "axis" attribute indicates the dimension along which {name}
 will be performed. The output tensor has the same shape
 and contains the {name} values of the corresponding input.
 )DOC";
-                        ReplaceAll(doc, "{name}", name);
-                        ReplaceAll(doc, "{description}", description);
-                        ReplaceAll(doc, "{equation}", equation););
+        ReplaceAll(doc, "{name}", name);
+        ReplaceAll(doc, "{description}", description);
+        ReplaceAll(doc, "{equation}", equation););
     std::string axis_attr;
-    POPULATE_OP_DOC_STR(axis_attr = R"DOC(
+    POPULATE_OP_DOC_STR(
+        axis_attr = R"DOC(
 Describes the dimension {name} will be performed on.
 Negative value means counting dimensions
 from the back. Accepted range is [-r, r-1] where r = rank(input).
 )DOC";
-                        ReplaceAll(axis_attr, "{name}", name););
+        ReplaceAll(axis_attr, "{name}", name););
     schema.SetDoc(doc);
     schema.Attr("axis", axis_attr, AttributeProto::INT, static_cast<int64_t>(-1));
     schema.Input(
@@ -1241,8 +1256,8 @@ ONNX_OPERATOR_SET_SCHEMA(
     22,
     OpSchema()
         .SetDoc(Softplus_ver22_doc)
-        .Input(0, "X", "1D input tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
-        .Output(0, "Y", "1D input tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
+        .Input(0, "X", "Input tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
+        .Output(0, "Y", "Output tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
         .TypeConstraint("T", OpSchema::all_float_types_ir4(), "Constrain input and output types to float tensors.")
         .TypeAndShapeInferenceFunction(propagateShapeAndTypeFromFirstInput)
         .FunctionBody(
@@ -1372,141 +1387,10 @@ ONNX_OPERATOR_SET_SCHEMA(
           defs::math::utils::MatMulShapeInference(ctx, 0, 1);
         }));
 
-static 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";
-
 ONNX_OPERATOR_SET_SCHEMA(
     TopK,
-    11,
-    OpSchema()
-        .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", OpSchema::all_numeric_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;
-        }));
+    24,
+    OpSchema().FillUsing(defs::math::utils::TopKOpGenerator(OpSchema::all_numeric_types_ir4())));
 
 static const char* Sin_ver22_doc = R"DOC(
 Calculates the sine of the given input tensor, element-wise.
@@ -2158,7 +2042,7 @@ ONNX_OPERATOR_SET_SCHEMA(
           }
         }));
 
-bool BuildContextDependentFunctionBody(
+static bool BuildContextDependentFunctionBody(
     const FunctionBodyBuildContext& ctx,
     const OpSchema& schema,
     FunctionProto& functionProto) {
@@ -2491,12 +2375,12 @@ ONNX_OPERATOR_SET_SCHEMA(
           }
         }));
 
-void einsumShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, std::string const& equation) {
+static void einsumShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, std::string const& equation) {
   // Only accept letters for indices
   auto is_letter = [](char c) { return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z'); };
 
   const size_t num_inputs = ctx.getNumInputs();
-  if (num_inputs < 1 || !hasNInputShapes(ctx, static_cast<int>(num_inputs))) {
+  if (num_inputs < 1 || !hasNInputShapes(ctx, num_inputs)) {
     return;
   }
   ONNX_NAMESPACE::TensorShapeProto output_shape;
@@ -2519,7 +2403,7 @@ void einsumShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, std::string con
   // Parse the left-hand side
   std::stringstream str(left_equation);
   std::map<char, size_t> label_maps;
-  std::set<char> repeated_labels;
+  std::unordered_set<char> repeated_labels;
   ONNX_NAMESPACE::TensorShapeProto dims_value, ellipsis_dims_value;
   size_t num_labels = 0;
   bool ellipsis_flag = true;
@@ -2537,12 +2421,33 @@ void einsumShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, std::string con
     size_t term_size = 0; // number of legal indices for the current term
     size_t num_illegal_char = 0; // number of illegal char before the current 'index' in the current term
 
-    for (size_t index = 0; index < term.size(); ++index) {
-      if (is_letter(term[index])) {
+    for (char index : term) {
+      if (is_letter(index)) {
         term_size += 1;
       }
     }
 
+    // Validate that term_size is compatible with rank before accessing dimensions
+    if (ellipsis_index != std::string::npos) {
+      // For ellipsis case, rank must be at least term_size
+      if (rank < term_size) {
+        fail_shape_inference(
+            "Ellipsis represents incompatible dimensions for input ",
+            num_operands,
+            ". Rank ",
+            rank,
+            " is less than term size ",
+            term_size,
+            ".");
+      }
+    } else {
+      // For non-ellipsis case, rank must equal term_size
+      if (rank != term_size) {
+        fail_shape_inference(
+            "Rank of input ", num_operands, " (", rank, ") does not match the equation indices (", term_size, ").");
+      }
+    }
+
     for (size_t index = 0; index < term.size(); ++index) {
       if (index == ellipsis_index) {
         // find ellipsis and record the dims represented by ellipsis
@@ -2571,7 +2476,7 @@ void einsumShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, std::string con
         continue;
       }
 
-      const auto inserted = label_maps.insert({term[index], num_labels}).second;
+      const auto inserted = label_maps.emplace(term[index], num_labels).second;
       if (inserted) {
         *dims_value.add_dim() = shape.dim(index + ellipsis_dims - num_illegal_char);
         ++num_labels;
@@ -2584,9 +2489,6 @@ void einsumShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, std::string con
       // If there is an ellipsis, the number of dimensions it represents
       // must be total dim - letter dimensions
       if (num_ellipsis == 0) {
-        if (rank < term_size) {
-          fail_shape_inference("Ellipsis represents incompatible dimensions.");
-        }
         num_ellipsis_indices = rank - term_size;
       } else { // ellipsis has been seen before. Check that if dimensions
                // are compatible
@@ -2595,10 +2497,6 @@ void einsumShapeInference(ONNX_NAMESPACE::InferenceContext& ctx, std::string con
         }
       }
       num_ellipsis++;
-    } else {
-      if (rank != term_size) {
-        fail_shape_inference("Rank of input ", num_operands, " does not match the equation indices.");
-      }
     }
     num_operands++;
   }
@@ -2688,7 +2586,7 @@ ONNX_OPERATOR_SET_SCHEMA(
           // Type inference
           propagateElemTypeFromInputToOutput(ctx, 0, 0);
           std::string equation = getAttribute(ctx, "equation", "");
-          if (equation.compare("") == 0) {
+          if (equation.empty()) {
             return;
           }
 
@@ -2697,7 +2595,7 @@ ONNX_OPERATOR_SET_SCHEMA(
           einsumShapeInference(ctx, equation);
         }));
 
-const char* reduction_doc_sce =
+static const char* reduction_doc_sce =
     "Type of reduction to apply to loss: none, sum, mean(default). "
     "'none': no reduction will be applied, "
     "'sum': the output will be summed. "
@@ -2747,7 +2645,7 @@ Finally, L is optionally reduced:
   where tensor W is of shape `(N, D1, D2, ..., Dk)` and `W[n][d1][d2]...[dk] = weights[labels[i][d1][d2]...[dk]]`.
 )DOC";
 
-bool BuildContextDependentFunctionBodySCE(
+static bool BuildContextDependentFunctionBodySCE(
     const FunctionBodyBuildContext& ctx,
     const OpSchema& schema,
     FunctionProto& functionProto) {
@@ -2858,7 +2756,7 @@ ONNX_OPERATOR_SET_SCHEMA(
         .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
           propagateElemTypeFromInputToOutput(ctx, 0, 0);
           std::string reduction = getAttribute(ctx, "reduction", "mean");
-          if (reduction.compare("none") == 0) {
+          if (reduction == "none") {
             if (hasInputShape(ctx, 1)) {
               propagateShapeFromInputToOutput(ctx, 1, 0);
             }
@@ -3020,15 +2918,16 @@ ONNX_OPERATOR_SET_SCHEMA(
             axis = -2;
           } else {
             const TensorProto* axis_tensor = ctx.getInputData(axis_arg_index);
-            ONNX_ASSERTM(axis_tensor != nullptr, "axis should not be nullptr at this point");
+            ONNX_ASSERTM(axis_tensor != nullptr, "axis should not be nullptr at this point")
             // TODO(justinchuby): Create invariance checking functions to ensure shapes and sizes
-            // to abstrct the following logic out.
+            // to abstract the following logic out.
             if (axis_tensor->dims_size() != 0) {
               fail_shape_inference("axis input must be a scalar.");
             }
             axis = defs::math::utils::GetScalarValueFromTensor<int64_t>(axis_tensor);
           }
 
+          // NOLINTNEXTLINE(readability-simplify-boolean-expr)
           if (!(-rank <= axis && axis != -1 && axis < rank - 1)) {
             fail_shape_inference(
                 "axis attribute value ",
@@ -3082,13 +2981,14 @@ ONNX_OPERATOR_SET_SCHEMA(
           updateOutputShape(ctx, output_index, result_shape_proto);
         }));
 
-std::function<void(OpSchema&)> CosineSumWindowOpDocGenerator(const char* name) {
+static std::function<void(OpSchema&)> CosineSumWindowOpDocGenerator(const char* name) {
   return [=](OpSchema& schema) {
     std::string doc;
-    POPULATE_OP_DOC_STR(doc = R"DOC(
+    POPULATE_OP_DOC_STR(
+        doc = R"DOC(
 Generates a {name} window as described in the paper https://ieeexplore.ieee.org/document/1455106.
 )DOC";
-                        ReplaceAll(doc, "{name}", name););
+        ReplaceAll(doc, "{name}", name););
 
     schema.SetDoc(doc);
     schema.Attr(