onnx-ir 0.21.0

//! Phase 2: Node Conversion
//!
//! Converts ONNX nodes to IR, performs node fusion and remapping:
//!
//! **Coalescing**: Gemm → Linear, MatMul+Add → Linear
//! **Remapping**: Conv → Conv1d/2d/3d, MaxPool → MaxPool1d/2d, etc.

use std::{cell::RefCell, collections::HashMap, iter::Peekable, rc::Rc, slice::Iter};

use crate::{
    graph_state::GraphState,
    ir::{ArgType, AttributeValue, NodeType, RawNode, TensorData, TensorDataExt},
    pipeline::Error,
    processor::get_processor_registry,
    proto_conversion::convert_node_proto,
    protos::{GraphProto, NodeProto},
};

/// Convert all ONNX nodes from GraphProto to IR nodes (for subgraphs)
///
/// # Errors
///
/// Returns an error if graph attribute processing fails
pub(crate) fn convert_nodes_from_graph(
    graph: &GraphProto,
    state_rc: &Rc<RefCell<GraphState>>,
    opset_version: usize,
) -> Result<(), Error> {
    convert_nodes_impl(&graph.node, state_rc, opset_version)
}

/// Internal implementation for node conversion
///
/// # Errors
///
/// Returns an error if graph attribute processing fails
fn convert_nodes_impl(
    nodes: &[NodeProto],
    state_rc: &Rc<RefCell<GraphState>>,
    opset_version: usize,
) -> Result<(), Error> {
    let mut node_name_counter: HashMap<NodeType, usize> = HashMap::new();

    // Get the name registry (if available)
    let name_registry = {
        let state = state_rc.borrow();

        if let Some(registry) = state.name_registry() {
            // Registry is shared, already initialized with constant count from initializers
            Some(registry.clone())
        } else {
            // Fall back to local counter for backwards compatibility
            let constant_count = state
                .processed_nodes
                .iter()
                .filter(|n| n.node_type == NodeType::Constant)
                .count();
            if constant_count > 0 {
                node_name_counter.insert(NodeType::Constant, constant_count);
            }
            None
        }
    };

    let mut node_iter = nodes.iter().peekable();

    while let Some(node_proto) = node_iter.next() {
        let mut node = convert_node_proto(node_proto, &state_rc.borrow());

        // Handle graph attributes for control flow nodes (If, Loop, Scan)
        if matches!(
            node.node_type,
            NodeType::If | NodeType::Loop | NodeType::Scan
        ) {
            // Extract outer-scope references from subgraphs BEFORE building them
            // These references need their types resolved from the parent graph
            let outer_refs =
                crate::proto_conversion::extract_node_outer_scope_references(node_proto);

            // Add outer-scope references as additional inputs to the node
            // This ensures:
            // 1. Type inference processes dependencies before this node
            // 2. The types are available when building subgraphs
            if !outer_refs.is_empty() {
                log::debug!(
                    "Found {} outer-scope references for {} node {}: {:?}",
                    outer_refs.len(),
                    node.node_type,
                    node.name,
                    outer_refs
                );
                let state = state_rc.borrow();
                let mut scope_ref_names: Vec<String> = Vec::new();
                for ref_name in outer_refs {
                    // Skip references that are already in the node's inputs
                    if node.inputs.iter().any(|i| i.name == ref_name) {
                        log::debug!("Skipping '{}' - already in node inputs", ref_name);
                        continue;
                    }

                    // Create an argument for this outer-scope reference
                    // The type will be resolved from the parent graph
                    let arg = state.init_in(&ref_name);
                    log::debug!(
                        "Adding outer-scope reference '{}' -> arg.name='{}' (type: {:?}) as input to {} node",
                        ref_name,
                        arg.name,
                        arg.ty,
                        node.node_type
                    );
                    node.inputs.push(arg);
                    scope_ref_names.push(ref_name);
                }

                // Store the count of regular ONNX inputs (before scope refs)
                // This allows node processors to distinguish ONNX inputs from scope refs
                let onnx_input_count = node.inputs.len() - scope_ref_names.len();
                node.attrs.insert(
                    "__onnx_input_count".to_string(),
                    AttributeValue::Int64(onnx_input_count as i64),
                );

                // Store the original ONNX names (sanitized) for the scope refs
                // These are needed when building subgraphs since subgraphs reference
                // the original names, not the renamed node output names
                node.attrs.insert(
                    "__scope_ref_names".to_string(),
                    AttributeValue::Strings(scope_ref_names),
                );
            }

            // Pass the current graph's NameRegistry to ensure unique names across nested subgraphs
            // If no registry exists, create one and initialize with current node counts
            let parent_registry = if let Some(registry) = state_rc.borrow().name_registry().cloned()
            {
                registry
            } else {
                // Create new registry and initialize with current node name counters
                let registry = crate::graph_state::NameRegistry::new();
                // Initialize counters from the current graph's already-named nodes
                // IMPORTANT: Increment by 1 to account for the current node which will be
                // renamed later using the local counter
                for (node_type, count) in &node_name_counter {
                    registry.set_initial_counter(node_type, count + 1);
                }
                // Also account for the current node's type
                registry.set_initial_counter(&node.node_type, 1);
                registry
            };

            // Convert graph attributes - now creates DeferredGraph instead of building immediately
            let base_path = state_rc.borrow().base_path().map(|p| p.to_path_buf());
            let graph_attrs = crate::proto_conversion::convert_graph_attributes(
                node_proto,
                opset_version,
                Some(parent_registry),
                base_path.as_deref(),
            );
            // Merge graph attributes with existing attributes
            for (key, value) in graph_attrs {
                node.attrs.insert(key, value);
            }
        }

        // Attach value_store to all arguments
        attach_value_stores(&mut node, state_rc);

        // Extract constant from attributes (if Constant node)
        if node.node_type == NodeType::Constant {
            extract_constant_from_attributes(&mut node, state_rc);
        }

        // Remap node types based on patterns
        remap_node_type(&mut node);

        // Rename node with counter
        rename_node(&mut node, &mut node_name_counter, name_registry.as_ref());

        // Track node type before coalesce (may change it)
        let node_type_before = node.node_type.clone();

        // Coalesce with following nodes (fusion)
        coalesce(&mut node, &mut node_iter, &mut state_rc.borrow_mut());

        // Re-attach value_stores after coalesce (may have added inputs)
        attach_value_stores(&mut node, state_rc);

        // Rename if coalesce changed node type
        if node.node_type != node_type_before {
            rename_node(&mut node, &mut node_name_counter, name_registry.as_ref());
        }

        // Lift constants and extract config
        let registry = get_processor_registry();
        let processor = registry.get(&node.node_type);

        processor
            .lift_constants(&mut node, opset_version)
            .unwrap_or_else(|e| {
                panic!(
                    "Failed to lift constants for node {} (type: {:?}): {:?}",
                    node.name, node.node_type, e
                )
            });

        // Config extraction is now done in infer_types instead of here
        // This allows processors to call extract_config during type inference

        // Add to graph state
        state_rc.borrow_mut().add_node(node);
    }

    Ok(())
}

/// Extract constant data from node attributes and move to tensor store
fn extract_constant_from_attributes(node: &mut RawNode, state_rc: &Rc<RefCell<GraphState>>) {
    let keys = [
        "value",
        "value_float",
        "value_floats",
        "value_int",
        "value_ints",
        "value_string",
        "value_strings",
    ];

    if let Some(attr_key) = keys.iter().find(|&key| node.attrs.contains_key(*key))
        && let Some(attr_value) = node.attrs.get(*attr_key)
    {
        let tensor_data_opt: Option<TensorData> = match attr_value {
            AttributeValue::Tensor(tensor) => Some(tensor.clone()),
            AttributeValue::Float32(val) => Some(TensorData::new(vec![*val], [0usize; 0])),
            AttributeValue::Float32s(vals) => Some(TensorData::new(vals.clone(), [vals.len()])),
            AttributeValue::Int64(val) => Some(TensorData::new(vec![*val], [0usize; 0])),
            AttributeValue::Int64s(vals) => Some(TensorData::new(vals.clone(), [vals.len()])),
            _ => None,
        };

        if let Some(tensor_data) = tensor_data_opt {
            // Store in central tensor store (convert to TensorDataRef)
            let data_id = {
                let mut state = state_rc.borrow_mut();
                state.store_tensor_data(tensor_data.clone().into())
            };

            // Create type from tensor data
            let ty = if tensor_data.shape.is_empty() {
                crate::ir::ArgType::ScalarNative(tensor_data.elem_type())
            } else {
                crate::ir::ArgType::Tensor(crate::ir::TensorType {
                    dtype: tensor_data.elem_type(),
                    rank: tensor_data.shape.len(),
                    static_shape: Some(tensor_data.shape.iter().map(|&d| Some(d)).collect()),
                })
            };

            // Create input with Static value
            let mut input_arg = crate::ir::Argument {
                name: String::new(),
                ty: ty.clone(),
                value_source: crate::ir::ValueSource::Static(data_id),
                value_store: None,
            };
            let value_store = state_rc.borrow().build_value_store();
            input_arg.set_value_store(value_store);
            node.inputs.push(input_arg);

            // Set output type and value_source
            if !node.outputs.is_empty() {
                node.outputs[0].value_source = crate::ir::ValueSource::Constant;
                node.outputs[0].ty = ty;
            }

            // Remove from attributes
            node.attrs.remove(*attr_key);
        }
    }
}

/// Attach value_store references to node arguments
///
/// For most arguments, we set the current graph's value_store.
/// For outer-scope constant/static arguments (from parent graph), we preserve their
/// existing value_store which contains the tensor data they reference.
///
/// We identify outer-scope constants by checking if:
/// 1. The argument is Constant (value_source == Constant) and already has a value_store
///    with the constant in its map, OR
/// 2. The argument is Static (value_source == Static(data_id)) and already has a value_store
///    with that data_id in its tensor store
fn attach_value_stores(node: &mut RawNode, state_rc: &Rc<RefCell<GraphState>>) {
    let value_store = state_rc.borrow().build_value_store();
    for arg in &mut node.inputs {
        // Check if this is an outer-scope constant/static that should preserve its parent's store
        let should_preserve = match arg.value_source {
            crate::ir::ValueSource::Constant => {
                // For Constant: check if the name is in the argument's existing store
                arg.value_store
                    .as_ref()
                    .map(|store| store.get_constant_data_id(&arg.name).is_some())
                    .unwrap_or(false)
            }
            crate::ir::ValueSource::Static(data_id) => {
                // For Static: check if the data_id is in the argument's existing store
                arg.value_store
                    .as_ref()
                    .map(|store| store.get_tensor_data(data_id).is_some())
                    .unwrap_or(false)
            }
            _ => false,
        };

        if should_preserve {
            log::debug!(
                "Preserving outer-scope value_store for '{}' ({:?})",
                arg.name,
                arg.value_source
            );
        } else {
            arg.set_value_store(value_store.clone());
        }
    }
    for arg in &mut node.outputs {
        arg.set_value_store(value_store.clone());
    }
}

/// Rename node with type-based counter
fn rename_node(
    node: &mut RawNode,
    counters: &mut HashMap<NodeType, usize>,
    name_registry: Option<&crate::graph_state::NameRegistry>,
) {
    // If registry is available, use it to generate unique names across subgraphs
    if let Some(registry) = name_registry {
        let old_name = node.name.clone();
        node.name = registry.generate_node_name(&node.node_type);
        log::debug!(
            "Renamed node: '{}' -> '{}' (type: {:?})",
            old_name,
            node.name,
            node.node_type
        );
    } else {
        // Fall back to local counter for backwards compatibility
        counters
            .entry(node.node_type.clone())
            .and_modify(|e| *e += 1)
            .or_insert(1);

        let new_name = format!("{}{}", node.node_type, counters[&node.node_type]).to_lowercase();
        node.name = new_name;
    }
}

/// Remap node type using kernel shape
fn remap_node_with_kernel_shape<F>(node: &mut RawNode, new_node_type: F)
where
    F: FnOnce(usize) -> NodeType,
{
    let spatial_dims = match node.attrs.get("kernel_shape") {
        Some(AttributeValue::Int64s(ints)) => ints.len(),
        None if [NodeType::Conv, NodeType::ConvTranspose].contains(&node.node_type) => {
            // "kernel_shape" attribute is optional and should be inferred from weights
            // https://onnx.ai/onnx/operators/onnx__Conv.html
            if let ArgType::Tensor(weight) = &node.inputs[1].ty {
                // Skip leading channels in/out
                weight.rank - 2
            } else {
                panic!("Cannot infer kernel spatial dims");
            }
        }
        _ => panic!("Cannot infer kernel shape"),
    };
    node.node_type = new_node_type(spatial_dims);
}

/// Remap node type to a more specific one
fn remap_node_type(node: &mut RawNode) {
    match node.node_type {
        NodeType::Conv => remap_node_with_kernel_shape(node, |spatial_dims| match spatial_dims {
            1 => NodeType::Conv1d,
            2 => NodeType::Conv2d,
            3 => NodeType::Conv3d,
            _ => panic!("Only conv 1d, 2d and 3d are supported"),
        }),
        NodeType::ConvTranspose => {
            remap_node_with_kernel_shape(node, |spatial_dims| match spatial_dims {
                1 => NodeType::ConvTranspose1d,
                2 => NodeType::ConvTranspose2d,
                3 => NodeType::ConvTranspose3d,
                _ => panic!("Only conv_transpose 1d, 2d and 3d are supported"),
            })
        }
        NodeType::MaxPool => {
            remap_node_with_kernel_shape(node, |spatial_dims| match spatial_dims {
                1 => NodeType::MaxPool1d,
                2 => NodeType::MaxPool2d,
                3 => NodeType::MaxPool3d,
                _ => panic!("Only max_pool 1d, 2d, and 3d are supported"),
            })
        }
        NodeType::AveragePool => {
            remap_node_with_kernel_shape(node, |spatial_dims| match spatial_dims {
                1 => NodeType::AveragePool1d,
                2 => NodeType::AveragePool2d,
                3 => NodeType::AveragePool3d,
                _ => panic!("Only avg_pool 1d, 2d, and 3d are supported"),
            })
        }
        NodeType::LpPool => remap_node_with_kernel_shape(node, |spatial_dims| match spatial_dims {
            1 => NodeType::LpPool1d,
            2 => NodeType::LpPool2d,
            _ => panic!("Only lp_pool 1d and 2d are supported"),
        }),
        _ => (),
    }
}

/// Coalesce adjacent nodes into a single node (Gemm→Linear, MatMul+Add→Linear)
fn coalesce(
    node: &mut RawNode,
    nodes_iter: &mut Peekable<Iter<NodeProto>>,
    graph_data: &mut GraphState,
) {
    #[allow(clippy::single_match)]
    match node.node_type {
        NodeType::Gemm => convert_gemm_to_linear(node),
        NodeType::MatMul => {
            convert_matmul_to_linear(node, nodes_iter, graph_data);
        }
        _ => {}
    }
}

/// Convert Gemm to Linear (when alpha=1, beta=1, transA=0, transB=1)
///
/// Note: The weight tensor is kept in its original ONNX layout [out_features, in_features].
/// The burn-onnx layer is responsible for transposing the weights to Burn's expected
/// layout [in_features, out_features] during code generation.
fn convert_gemm_to_linear(node: &mut RawNode) {
    if node.outputs.len() != 1 {
        panic!("Gemm node must have 1 output");
    }

    // Check transA - must be 0 (no transpose) or absent (default is 0)
    let trans_a = node
        .attrs
        .get("transA")
        .map(|v| matches!(v, AttributeValue::Int64(1)))
        .unwrap_or(false);

    if trans_a {
        log::debug!(
            "Keeping Gemm node {} (transA=1 not supported for Linear conversion)",
            node.name
        );
        return;
    }

    let straight_linear = match (
        node.attrs.get("alpha"),
        node.attrs.get("beta"),
        node.attrs.get("transB"),
    ) {
        (
            Some(AttributeValue::Float32(alpha)),
            Some(AttributeValue::Float32(beta)),
            Some(AttributeValue::Int64(trans_b)),
        ) => *alpha == 1.0 && *beta == 1.0 && *trans_b == 1,
        _ => false,
    };

    if straight_linear {
        log::debug!("Fusing Gemm → Linear for node {}", node.name);
        node.node_type = NodeType::Linear;
        node.attrs.remove("alpha");
        node.attrs.remove("beta");
        node.attrs.remove("transA");
        node.attrs.remove("transB");
        // Mark that weights need transposition (Gemm layout is [out, in])
        node.attrs
            .insert("transpose_weight".to_string(), AttributeValue::Int64(1));
    } else {
        log::debug!(
            "Keeping Gemm node {} (alpha={:?}, beta={:?}, transB={:?} don't match Linear pattern)",
            node.name,
            node.attrs.get("alpha"),
            node.attrs.get("beta"),
            node.attrs.get("transB")
        );
    }
}

/// Convert MatMul to Linear, fusing the following Add node as bias if present
fn convert_matmul_to_linear(
    node: &mut RawNode,
    iter_mut: &mut Peekable<Iter<NodeProto>>,
    graph_data: &mut GraphState,
) {
    if node.inputs.len() != 2 {
        panic!("MatMul node must have 2 inputs");
    }

    // if the second input does not have a value, it is not a weight, then proceed to the next node
    if !graph_data.has_value(&node.inputs[1].name) {
        log::debug!(
            "Keeping MatMul node {} (second input is not a constant weight)",
            node.name
        );
        return;
    }

    // Check if the second input is a 2D tensor; if not, skip conversion
    match &node.inputs[1].ty {
        ArgType::Tensor(tensor_type) if tensor_type.rank == 2 => {
            // Weight is 2D, proceed with conversion
        }
        ArgType::Tensor(tensor_type) => {
            log::debug!(
                "Keeping MatMul node {} (weight is {}D, not 2D)",
                node.name,
                tensor_type.rank
            );
            return;
        }
        _ => {
            log::debug!("Keeping MatMul node {} (weight is not a tensor)", node.name);
            return;
        }
    }

    // Convert the node to Linear
    node.node_type = NodeType::Linear;
    log::debug!("Converting MatMul → Linear for node {}", node.name);

    // Check the next node for potential conversion
    if let Some(peek_node) = iter_mut.peek() {
        let peek_node = convert_node_proto(peek_node, graph_data);
        if is_add_node_with_bias(&peek_node, node, graph_data) {
            convert_and_remove_add_node(&peek_node, node);
            log::debug!("Fused Add bias into Linear node {}", node.name);

            // You don't have to remove it if it's never stored in the first place
            let _ = iter_mut.next();
        }
    }
}

/// Check if the peeked node is an Add with bias for the current MatMul
fn is_add_node_with_bias(
    peek_node: &RawNode,
    current_node: &RawNode,
    graph_data: &GraphState,
) -> bool {
    // Check structural requirements first
    if peek_node.node_type != NodeType::Add || peek_node.inputs.len() != 2 {
        return false;
    }

    // Identify which Add input is the bias (constant) and which is the matmul output
    let bias_input = if peek_node.inputs[0].name == current_node.outputs[0].name
        && graph_data.has_value(&peek_node.inputs[1].name)
    {
        &peek_node.inputs[1]
    } else if peek_node.inputs[1].name == current_node.outputs[0].name
        && graph_data.has_value(&peek_node.inputs[0].name)
    {
        &peek_node.inputs[0]
    } else {
        return false;
    };

    // The bias must be a 1D tensor whose size matches out_features (weight dim[1]).
    // A scalar bias (e.g. shape [1]) would cause a shape mismatch in the Linear layer.
    let weight_input = &current_node.inputs[1];
    match (&bias_input.ty, &weight_input.ty) {
        (ArgType::Tensor(bias_ty), ArgType::Tensor(weight_ty)) => {
            if bias_ty.rank != 1 {
                log::debug!(
                    "Skipping Add fusion: bias rank is {}, expected 1",
                    bias_ty.rank
                );
                return false;
            }
            // If both static shapes are known, verify the bias dimension matches out_features
            if let (Some(bias_shape), Some(weight_shape)) =
                (&bias_ty.static_shape, &weight_ty.static_shape)
            {
                let out_features = weight_shape.get(1).copied().flatten();
                let bias_dim = bias_shape.first().copied().flatten();
                if let (Some(out_f), Some(bias_d)) = (out_features, bias_dim)
                    && out_f != bias_d
                {
                    log::debug!(
                        "Skipping Add fusion: bias size {} != out_features {}",
                        bias_d,
                        out_f
                    );
                    return false;
                }
            } else {
                log::debug!("Add fusion: static shapes not fully known, proceeding optimistically");
            }
            true
        }
        _ => {
            log::debug!(
                "Skipping Add fusion: expected (Tensor, Tensor), got ({}, {})",
                bias_input.ty,
                weight_input.ty
            );
            false
        }
    }
}

/// Merge the Add node's bias into the MatMul node
fn convert_and_remove_add_node(bias_node: &RawNode, current_node: &mut RawNode) {
    // The bias is whichever input is NOT the matmul output
    let bias_input = if bias_node.inputs[0].name == current_node.outputs[0].name {
        bias_node.inputs[1].clone()
    } else {
        bias_node.inputs[0].clone()
    };

    // Push the bias input and update the output name
    current_node.inputs.push(bias_input);
    current_node.outputs[0]
        .name
        .clone_from(&bias_node.outputs[0].name);
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::node::test_utils::TestNodeBuilder;

    #[test]
    fn should_infer_conv2d_node_from_weights_rank() {
        // Weight tensor data - not important for the test
        let weight_data = vec![0.0; 16];
        // [out_channels, in_channels, k_h, k_w] = [2, 2, 2, 2] = 16 elements
        let weight_shape = vec![2, 2, 2, 2];

        let mut node = TestNodeBuilder::new(NodeType::Conv, "test_conv2d")
            .input_tensor_f32("data", 4, None)
            .input_tensor_f32_data("weight", weight_data.clone(), weight_shape)
            .output_tensor_f32("output", 4, None)
            // .attr_ints("kernel_shape", kernel_shape)
            .attr_ints("strides", vec![1, 1])
            .attr_ints("pads", vec![0, 0, 0, 0])
            .attr_ints("dilations", vec![1, 1])
            .attr_int("group", 1)
            .build();

        assert_eq!(node.node_type, NodeType::Conv);
        remap_node_type(&mut node);
        assert_eq!(node.node_type, NodeType::Conv2d);
    }

    #[test]
    fn should_infer_conv_transpose1d_node_from_weights_rank() {
        // Weight tensor data - not important for the test
        let weight_data = vec![0.0; 16];
        // [.., kernel_size]
        let weight_shape = vec![2, 2, 4];

        let mut node = TestNodeBuilder::new(NodeType::ConvTranspose, "test_conv2d")
            .input_tensor_f32("data", 3, None)
            .input_tensor_f32_data("weight", weight_data, weight_shape)
            .output_tensor_f32("output", 3, None)
            .build();

        assert_eq!(node.node_type, NodeType::ConvTranspose);
        remap_node_type(&mut node);
        assert_eq!(node.node_type, NodeType::ConvTranspose1d);
    }

    #[test]
    fn should_skip_matmul_to_linear_for_1d_weight() {
        use crate::ir::TensorData;

        // Create MatMul node with 1D weight (should NOT be converted to Linear)
        let weight_data = vec![1.0, 2.0, 3.0, 4.0];
        let weight_shape = vec![4]; // 1D tensor

        let node = TestNodeBuilder::new(NodeType::MatMul, "test_matmul")
            .input_tensor_f32("input", 2, None)
            .input_tensor_f32_data("weight", weight_data.clone(), weight_shape.clone())
            .output_tensor_f32("output", 1, None)
            .build();

        // Create GraphState and register the weight as a constant
        let mut graph_state = crate::graph_state::GraphState::new(&[], &[], &[], &[]);
        graph_state.register_test_constant(
            "weight".to_string(),
            TensorData::new(weight_data, weight_shape),
        );

        let mut node = node;
        let empty_nodes: Vec<NodeProto> = vec![];
        let mut iter = empty_nodes.iter().peekable();

        // This should NOT panic - it should gracefully skip conversion
        convert_matmul_to_linear(&mut node, &mut iter, &mut graph_state);

        // Node type should remain MatMul (not converted to Linear)
        assert_eq!(node.node_type, NodeType::MatMul);
    }

    /// Helper to create a MatMul node and Add node pair for bias fusion tests.
    /// Sets static_shape on weight and bias inputs (matching the real pipeline).
    fn make_matmul_add_pair(
        out_features: usize,
        bias_shape: Vec<usize>,
    ) -> (RawNode, RawNode, crate::graph_state::GraphState) {
        use crate::ir::{TensorData, TensorType};

        let in_features = 4;
        let weight_data = vec![0.0; in_features * out_features];
        let weight_shape = vec![in_features, out_features];

        let matmul_node = TestNodeBuilder::new(NodeType::MatMul, "matmul")
            .input_tensor_f32("input", 2, Some(vec![8, in_features]))
            .input_tensor_f32_data("weight", weight_data.clone(), weight_shape.clone())
            .output_tensor_f32("matmul_out", 2, Some(vec![8, out_features]))
            .build();

        // Patch the weight input to have static_shape (as it would in the real pipeline)
        let mut matmul_node = matmul_node;
        matmul_node.inputs[1].ty = ArgType::Tensor(TensorType {
            dtype: burn_tensor::DType::F32,
            rank: 2,
            static_shape: Some(weight_shape.iter().map(|&d| Some(d)).collect()),
        });

        let bias_data = vec![0.0; bias_shape.iter().product::<usize>()];
        let mut add_node = TestNodeBuilder::new(NodeType::Add, "add")
            .input_tensor_f32("matmul_out", 2, Some(vec![8, out_features]))
            .input_tensor_f32_data("bias", bias_data.clone(), bias_shape.clone())
            .output_tensor_f32("add_out", 2, Some(vec![8, out_features]))
            .build();

        // Patch the bias input to have static_shape (as it would in the real pipeline)
        add_node.inputs[1].ty = ArgType::Tensor(TensorType {
            dtype: burn_tensor::DType::F32,
            rank: 1,
            static_shape: Some(bias_shape.iter().map(|&d| Some(d)).collect()),
        });

        let mut graph_state = crate::graph_state::GraphState::new(&[], &[], &[], &[]);
        graph_state.register_test_constant(
            "weight".to_string(),
            TensorData::new(weight_data, weight_shape),
        );
        graph_state
            .register_test_constant("bias".to_string(), TensorData::new(bias_data, bias_shape));

        (matmul_node, add_node, graph_state)
    }

    #[test]
    fn should_fuse_add_bias_when_shape_matches() {
        let (matmul_node, add_node, graph_state) = make_matmul_add_pair(32, vec![32]);
        assert!(is_add_node_with_bias(&add_node, &matmul_node, &graph_state));
    }

    #[test]
    fn should_skip_add_fusion_when_bias_is_scalar() {
        // Bias shape [1] does not match out_features=32
        let (matmul_node, add_node, graph_state) = make_matmul_add_pair(32, vec![1]);
        assert!(!is_add_node_with_bias(
            &add_node,
            &matmul_node,
            &graph_state
        ));
    }

    #[test]
    fn should_skip_add_fusion_when_bias_rank_is_2d() {
        use crate::ir::{TensorData, TensorType};

        let out_features = 32;
        let in_features = 4;
        let weight_data = vec![0.0; in_features * out_features];
        let weight_shape = vec![in_features, out_features];

        let matmul_node = {
            let mut n = TestNodeBuilder::new(NodeType::MatMul, "matmul")
                .input_tensor_f32("input", 2, Some(vec![8, in_features]))
                .input_tensor_f32_data("weight", weight_data.clone(), weight_shape.clone())
                .output_tensor_f32("matmul_out", 2, Some(vec![8, out_features]))
                .build();
            n.inputs[1].ty = ArgType::Tensor(TensorType {
                dtype: burn_tensor::DType::F32,
                rank: 2,
                static_shape: Some(weight_shape.iter().map(|&d| Some(d)).collect()),
            });
            n
        };

        // Bias is rank 2 [1, 32], should be rejected
        let bias_data = vec![0.0; out_features];
        let bias_shape = vec![1, out_features];
        let mut add_node = TestNodeBuilder::new(NodeType::Add, "add")
            .input_tensor_f32("matmul_out", 2, Some(vec![8, out_features]))
            .input_tensor_f32_data("bias", bias_data.clone(), bias_shape.clone())
            .output_tensor_f32("add_out", 2, Some(vec![8, out_features]))
            .build();
        add_node.inputs[1].ty = ArgType::Tensor(TensorType {
            dtype: burn_tensor::DType::F32,
            rank: 2,
            static_shape: Some(bias_shape.iter().map(|&d| Some(d)).collect()),
        });

        let mut graph_state = crate::graph_state::GraphState::new(&[], &[], &[], &[]);
        graph_state.register_test_constant(
            "weight".to_string(),
            TensorData::new(weight_data, weight_shape),
        );
        graph_state
            .register_test_constant("bias".to_string(), TensorData::new(bias_data, bias_shape));

        assert!(!is_add_node_with_bias(
            &add_node,
            &matmul_node,
            &graph_state
        ));
    }

    #[test]
    fn should_fuse_add_bias_with_reversed_operand_order() {
        use crate::ir::{TensorData, TensorType};

        let out_features = 32;
        let in_features = 4;
        let weight_data = vec![0.0; in_features * out_features];
        let weight_shape = vec![in_features, out_features];

        let matmul_node = {
            let mut n = TestNodeBuilder::new(NodeType::MatMul, "matmul")
                .input_tensor_f32("input", 2, Some(vec![8, in_features]))
                .input_tensor_f32_data("weight", weight_data.clone(), weight_shape.clone())
                .output_tensor_f32("matmul_out", 2, Some(vec![8, out_features]))
                .build();
            n.inputs[1].ty = ArgType::Tensor(TensorType {
                dtype: burn_tensor::DType::F32,
                rank: 2,
                static_shape: Some(weight_shape.iter().map(|&d| Some(d)).collect()),
            });
            n
        };

        // Bias as inputs[0], matmul output as inputs[1] (reversed order)
        let bias_data = vec![0.0; out_features];
        let bias_shape = vec![out_features];
        let mut add_node = TestNodeBuilder::new(NodeType::Add, "add")
            .input_tensor_f32_data("bias", bias_data.clone(), bias_shape.clone())
            .input_tensor_f32("matmul_out", 2, Some(vec![8, out_features]))
            .output_tensor_f32("add_out", 2, Some(vec![8, out_features]))
            .build();
        add_node.inputs[0].ty = ArgType::Tensor(TensorType {
            dtype: burn_tensor::DType::F32,
            rank: 1,
            static_shape: Some(bias_shape.iter().map(|&d| Some(d)).collect()),
        });

        let mut graph_state = crate::graph_state::GraphState::new(&[], &[], &[], &[]);
        graph_state.register_test_constant(
            "weight".to_string(),
            TensorData::new(weight_data, weight_shape),
        );
        graph_state
            .register_test_constant("bias".to_string(), TensorData::new(bias_data, bias_shape));

        assert!(is_add_node_with_bias(&add_node, &matmul_node, &graph_state));
    }

    #[test]
    fn should_fuse_add_bias_when_static_shapes_unknown() {
        use crate::ir::TensorData;

        // When static shapes are not available, fusion proceeds optimistically
        let matmul_node = TestNodeBuilder::new(NodeType::MatMul, "matmul")
            .input_tensor_f32("input", 2, None)
            .input_tensor_f32_data("weight", vec![0.0; 128], vec![4, 32])
            .output_tensor_f32("matmul_out", 2, None)
            .build();
        // Note: input_tensor_f32_data does not set static_shape, simulating unknown shapes

        let add_node = TestNodeBuilder::new(NodeType::Add, "add")
            .input_tensor_f32("matmul_out", 2, None)
            .input_tensor_f32_data("bias", vec![0.0; 32], vec![32])
            .output_tensor_f32("add_out", 2, None)
            .build();

        let mut graph_state = crate::graph_state::GraphState::new(&[], &[], &[], &[]);
        graph_state.register_test_constant(
            "weight".to_string(),
            TensorData::new(vec![0.0; 128], vec![4, 32]),
        );
        graph_state
            .register_test_constant("bias".to_string(), TensorData::new(vec![0.0; 32], vec![32]));

        // Should return true (optimistic fallback)
        assert!(is_add_node_with_bias(&add_node, &matmul_node, &graph_state));
    }
}