1  // kernels.cuh
 2  // CUDA kernel declarations for Python bindings
 3  #pragma once
 4  
 5  // Scale vector elements: data[i] *= scale
 6  void launch_vector_scale(float *data, float scale, int n);
 7  
 8  // SAXPY: y = a*x + y
 9  void launch_saxpy(float *y, float a, const float *x, int n);
10  
11  // Dot product (result written to device memory)
12  void launch_dot_product(float *result, const float *a, const float *b, int n);