//go:build !cuda

// Package compute provides device-agnostic computation dispatching.
package compute

import (
	"fmt"

	"makarna/pkg/backend/cpu"
	"makarna/pkg/backend/cpu/matmul"
	"makarna/pkg/tensor"
)

// Linear performs a linear layer: output = input @ weight.T
// CPU-only build: always uses CPU path.
func Linear(ctx *Context, input, weight, output tensor.Tensor) error {
	return linearCPU(input, weight, output)
}

// linearCPU executes matmul on CPU
func linearCPU(input, weight, output tensor.Tensor) error {
	inCPU, ok := input.(*cpu.Tensor)
	if !ok {
		var err error
		inCPU, err = ToCPU(input)
		if err != nil {
			return fmt.Errorf("linear: failed to get CPU input: %w", err)
		}
	}

	wCPU, ok := weight.(*cpu.Tensor)
	if !ok {
		return fmt.Errorf("linear: weight must be CPU tensor for CPU path")
	}

	outCPU, ok := output.(*cpu.Tensor)
	if !ok {
		return fmt.Errorf("linear: output must be CPU tensor for CPU path")
	}

	return matmul.Linear(inCPU, wCPU, outCPU)
}