1 files changed, 28 insertions, 2 deletions

diff --git a/src/core/runtime.c b/src/core/runtime.c
index 0e5cf9d..950ec7c 100644
--- a/src/core/runtime.c
+++ b/src/core/runtime.c
@@ -4,6 +4,7 @@
 #include "core/gates.h"
 #include "core/kernel.h"
 #include "maths/complex.h"
+#include "maths/simd.h"
 
 struct PsiRuntimeConfig psi_new_runtime_config(void)
 {
@@ -121,6 +122,31 @@ static struct PsiVector new_zero_state(size_t num_qubits)
 	return state;
 }
 
+static void execute_kernels(struct PsiVector *state, const struct PsiKernel *kernels, size_t count, size_t num_qubits, struct PsiRuntimeConfig config)
+{
+	bool use_parallel = config.parallel && num_qubits >= config.parallel_threshold;
+
+	for (size_t i = 0; i < count; i++)
+	{
+		struct PsiKernel kernel = kernels[i];
+
+		if (config.simd && kernel.target_count == 1)
+		{
+			struct PsiComplex gate[2][2] = {
+				{ kernel.matrix.data[0], kernel.matrix.data[1] },
+				{ kernel.matrix.data[2], kernel.matrix.data[3] },
+			};
+
+			if (use_parallel)
+				psi_apply_single_qubit_gate_simd_parallel(state->data, gate, kernel.targets[0], num_qubits);
+			else
+				psi_apply_single_qubit_gate_simd(state->data, gate, kernel.targets[0], num_qubits);
+		}
+		else
+			psi_apply_kernel(state, kernel, num_qubits);
+	}
+}
+
 struct PsiVector psi_compute_runtime_config(struct PsiRuntimeConfig config, size_t num_qubits, const struct PsiGateOp *operations, size_t op_count)
 {
 	struct PsiVector state = new_zero_state(num_qubits);
@@ -136,7 +162,7 @@ struct PsiVector psi_compute_runtime_config(struct PsiRuntimeConfig config, size
 		}
 
 		psi_optimize_structure_aware_batch(&batch);
-		psi_execute_structure_aware_batch(batch, &state);
+		execute_kernels(&state, batch.kernels, batch.count, num_qubits, config);
 		psi_free_structure_aware_batch(&batch);
 
 		return state;
@@ -153,7 +179,7 @@ struct PsiVector psi_compute_runtime_config(struct PsiRuntimeConfig config, size
 	if (config.batched)
 		psi_optimize_kernel_batch(&batch);
 
-	psi_execute_kernel_batch(batch, &state);
+	execute_kernels(&state, batch.kernels, batch.count, num_qubits, config);
 	psi_free_kernel_batch(&batch);
 
 	return state;