Move softmax partial kernels to matmul

This commit is contained in:
2024-04-11 22:01:47 +02:00
parent bf7c961b9e
commit 710a33bdde
6 changed files with 274 additions and 212 deletions

View File

@@ -29,45 +29,6 @@ __global__ void relu(
const unsigned int len
);
/**
* @brief Softmax activation exponentiation kernel
*
* @param src Pointer to the source array
* @param dst Pointer to the destination array
* @param len Length of the arrays
*/
__global__ void softmax_exp(
const float* __restrict__ src,
float* __restrict__ dst,
const unsigned int len
);
/**
* @brief
*
* @param d_vector Device pointer to vector
* @param d_output Device pointer to output vector
* @param w Length of the vector
*/
__global__ void softmax_sum(
const float* __restrict__ d_vector,
float* __restrict__ d_output
);
/**
* @brief Softmax activation function kernel
*
* @param src Pointer to the source array
* @param dst Pointer to the destination array
* @param len Length of the arrays
*/
__global__ void softmax_div(
const float* __restrict__ src,
float* __restrict__ dst,
const float* __restrict__ sum,
const unsigned int len
);
} // namespace CUDANet::Kernels
#endif // CUDANET_ACTIVATION_FUNCTIONS_H