Introduce ConvDims/PoolDims parameter structs to reduce argument counts

[kolkov]

Background

Following feedback from @marcelloh in #14, our conv2d and maxpool2d functions have 10-14 arguments each, which exceeds the recommended maximum of 5.

Problem

Functions like conv2dInputBackwardFloat32 have 14 parameters:

func conv2dInputBackwardFloat32(
    inputGrad, grad, kernel *tensor.RawTensor,
    n, cIn, h, w, cOut, kH, kW, hOut, wOut, stride, padding int,
)

Proposed Solution

Introduce parameter structs following the Burn framework pattern:

// ConvDims groups convolution dimension parameters
type ConvDims struct {
    N, CIn, H, W       int  // Input dimensions
    COut, KH, KW       int  // Kernel dimensions  
    HOut, WOut         int  // Output dimensions
    Stride, Padding    int  // Convolution parameters
}

// PoolDims groups pooling dimension parameters
type PoolDims struct {
    N, C, H, W         int  // Input dimensions
    KH, KW             int  // Kernel dimensions
    HOut, WOut         int  // Output dimensions
    Stride, Padding    int  // Pooling parameters
}

This reduces function signatures from 14 to 4-5 parameters:

func conv2dInputBackwardFloat32(
    inputGrad, grad, kernel *tensor.RawTensor,
    dims *ConvDims,
)

Files to Update

• internal/backend/cpu/conv2d.go
• internal/backend/cpu/conv2d_backward.go
• internal/backend/cpu/maxpool2d.go
• internal/autodiff/ops/conv2d.go
• internal/autodiff/ops/maxpool2d.go

Acceptance Criteria

• ConvDims struct introduced
• PoolDims struct introduced
• All conv2d functions updated to use structs
• All maxpool2d functions updated to use structs
• No performance regression (benchmark comparison)
• All tests pass
• golangci-lint: 0 issues

Risk Assessment

Low risk - This is a refactoring that doesn't change algorithmic logic.

Related

• Issue Avoid too complex code #14 - Original complexity analysis

Labels

enhancement, refactoring, code-quality