Skip to main content

Neural Operations

Specialized tensor operations for neural network workloads.

Fused Multiply-Add (FMA)​

The compiler automatically fuses a * b + c into a single FMA operation:

result is (a * b) + c;    // fused to FMA for performance

BLAS Operations​

Matrix multiplication uses optimized BLAS routines when available:

output is weights @ input;    // calls BLAS gemm

Activation Functions​

Apply activation functions from the NN module:

module NN;

activated is NN.ReLU(preactivation);
softmaxed is NN.Softmax(logits);

GPU Acceleration​

Wrap tensor operations in gpu { } for GPU dispatch:

gpu {
    result is A @ B;        // GPU matrix multiply
    activated is NN.ReLU(result);
}

Next Steps​