Implement selective GPU dispatch for optimal CPU-GPU heterogeneous computing

[suzannejin]

Problem

Current implementation sends all functions to GPU when options(propr.use_gpu = TRUE) is set. Benchmarks show some functions are much slower on GPU due to transfer overhead and kernel launch costs.

Benchmark Results

GPU Winners (always beneficial):

• lrm, lrv: 500x-2500x speedup 🔥
• corRcpp, linRcpp, rhoRcpp: 100x-700x speedup
• clrRcpp, phiRcpp: 20x-370x speedup

GPU Losers (slower than CPU):

• coordToIndex: 0.1x-0.4x (3-9x SLOWER)
• count_* functions: 0.01x-0.03x (30-100x SLOWER)
• wtvRcpp, wtmRcpp: 0.6x-2x

Solution

Implement internal selective dispatch: automatically route functions to GPU only when beneficial, regardless of global options(propr.use_gpu = TRUE) setting.

bool should_use_gpu_internal(const char* func_name) {
    // Never use CPU for these
    if (func_name in ["coordToIndex", "count_*", "wtmRcpp", "wtvRcpp"]) {
        return false;
    }
    // Always use GPU for these (when available)
    if (func_name in ["lrm", "lrv", "corRcpp", "linRcpp", "rhoRcpp"]) {
        return true;
    }
}

User Experience

options(propr.use_gpu = TRUE)  # Enable GPU globally

# Package automatically decides:
# - lrm/lrv/corRcpp → GPU (fast!)
# - count_*/coordToIndex → CPU (avoid slowdown)

pr <- propr(counts, metric = "rho")  # Optimal performance by default

Benefits

• Optimal performance by default
• No performance regressions
• Fewer unnecessary CPU↔GPU transfers
• Users don't need to think about which functions benefit from GPU

Next Steps

1. Implement should_use_gpu_internal() dispatch logic
2. Keep global options(propr.use_gpu) as master switch
3. Test on different hardware configurations

---

Issue created by Luna (AI assistant for @suzannejin)