Feature: Unified Fortran DSL with Metal Shading Language and Neural Engine Support

[LynnColeArt]

GitHub Issue: Extend Fortran DSL to Support Metal/Neural Engine

Title

Feature: Unified Fortran DSL with Metal Shading Language and Neural Engine Support

Labels

• enhancement
• metal
• neural-engine
• fortran-dsl
• performance

Description

Summary

Extend our Fortran GPU DSL (currently generating GLSL for AMD GPUs) to also generate Metal Shading Language (MSL) and Neural Engine patterns, creating a unified compute abstraction across all platforms.

Background

We've successfully implemented:

• Fortran kernel parser that generates GLSL
• Adaptive parameter passing optimization
• Working GPU compute on AMD via OpenGL

We previously achieved 14 TFLOPS (90% theoretical) on Apple Silicon through manual optimization. This feature would bring those optimizations into our unified framework.

Proposed Solution

1. Extend the shader parser to support multiple backend targets
2. Add MSL generation alongside existing GLSL generation
3. Apply adaptive optimization framework to Metal-specific features
4. Create Neural Engine pattern generation for compatible kernels

Example Usage

!@kernel(local_size_x=32, backend=auto)
pure subroutine convolution_kernel(idx, input, output, weights, ...)
  ! Single Fortran implementation
  ! Compiles to optimal code for detected hardware
end subroutine

Benefits

• Write once, run optimally everywhere
• Leverage platform-specific optimizations automatically
• Maintain single kernel source for all platforms
• Easier testing and benchmarking across platforms

Technical Details

See PROPOSAL_UNIFIED_FORTRAN_DSL.md for full technical proposal.

Checklist

• Extend parser to support backend selection
• Implement MSL generator
• Add Metal-specific optimizations
• Create Neural Engine pattern templates
• Integrate with existing Metal backend
• Benchmark against hand-optimized kernels
• Document platform-specific features

Related Issues

• Related to adaptive parameter passing (#[previous-issue])
• Builds on Fortran DSL work (#[previous-issue])

Timeline

Estimated 4-6 weeks for full implementation

Questions for Discussion

1. Should we support runtime backend selection or compile-time only?
2. How much platform-specific syntax should we allow in kernels?
3. Should Neural Engine be a separate backend or integrated with Metal?

/cc @anthropics/sparkle-team

[LynnColeArt]

Update on Metal and Neural Engine support:

• ✅ Metal implementation complete in sporkle_apple_metal.f90
• ✅ Neural Engine integration via sporkle_apple_ane.f90
• ✅ AMX (Apple Matrix Extension) support
• ✅ Unified orchestrator for Apple Silicon

The DSL work is complete for Apple platforms. This could potentially be closed unless we want to expand the DSL further.