Skip to main content

Tensor Operations

Morph supports element-wise arithmetic and matrix multiplication on tensors.

Element-Wise Operations​

t3 is t1 + t2;    // addition
t4 is t1 - t2;    // subtraction
t5 is t1 * t2;    // element-wise multiplication
t6 is t1 / t2;    // element-wise division

Matrix Multiplication​

The @ operator performs matrix multiplication:

t7 is t1 @ t2;    // matrix multiply

Fused Operations (Kernel Fusion)​

Multiple tensor operations are automatically fused into a single GPU/CPU kernel to minimize memory bandwidth usage:

result is (a * b) + c;    // Single kernel: out[i] = (a[i] * b[i]) + c[i]

Optimization Techniques​

  • Cache Blocking: Tiling matrices to fit in L1/L2 cache.
  • SIMD Vectorization: Using AVX/SSE for CPU speedup.
  • Loop Tiling: Restructuring nested loops for spatial locality.

Real Example​

From TensorTest.mx:

module Tensor;

TensorDemo method(t1 as Tensor<float>, t2 as Tensor<float>) {
    t3 is t1 + t2;
    t4 is t1 - t2;
    t5 is t1 * t2;
    t6 is t1 / t2;
    t7 is t1 @ t2;
    Print(t3);
    Print(t7);
}

Next Steps​