Simple CUDA kernel

Author: pwilkin

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2144,6 +2144,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ADD_ID:
         case GGML_OP_ADD1:
         case GGML_OP_ACC:
+        case GGML_OP_XIELU:
             {
                 n_tasks = n_threads;
             } break;
@@ -2167,7 +2168,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_REPEAT:
         case GGML_OP_REPEAT_BACK:
         case GGML_OP_LEAKY_RELU:
-        case GGML_OP_XIELU:
             {
                 n_tasks = 1;
             } break;

ggml/src/ggml-cpu/unary-ops.cpp

@@ -207,9 +207,6 @@ void ggml_compute_forward_xielu(const ggml_compute_params * params, ggml_tensor
     const float beta = op_params[2];
     const float eps = op_params[3];
 
-//    alpha_p = softplus(alpha_p);
-//    alpha_n = beta + softplus(alpha_n);
-
     const auto xielu_op_params = [alpha_n, alpha_p, beta, eps](float f) {
         return op_xielu(f, alpha_n, alpha_p, beta, eps);
     };

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2517,6 +2517,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_SGD:
             ggml_cuda_opt_step_sgd(ctx, dst);
             break;
+        case GGML_OP_XIELU:
+            ggml_cuda_op_xielu(ctx, dst);
+            break;
         default:
             return false;
     }