Gray node limit

Author: WrathfulSpatula

pyqrackising/kernels.cl

@@ -925,7 +925,7 @@ __kernel void gray(
     int i = get_global_id(0);
     const int max_i = get_global_size(0);
 
-    ulong theta_local[1024];
+    ulong theta_local[2048];
     for (int b = 0; b < blocks; ++b) {
         theta_local[b] = theta[b];
     }

pyqrackising/maxcut_tfim_util.py

@@ -8,6 +8,7 @@
 
 class OpenCLContext:
     def __init__(self, a, b, g, d, e, f, c, q, i, j, k, l, m, n, o, p, x, y, z, w, s, t):
+        self.GRAY_NODE_LIMIT = 131072
         self.MAX_GPU_PROC_ELEM = a
         self.IS_OPENCL_AVAILABLE = b
         self.work_group_size = g

pyqrackising/spin_glass_solver.py

@@ -15,6 +15,7 @@
 
 dtype = opencl_context.dtype
 wgs = opencl_context.work_group_size
+gnl = opencl_context.GRAY_NODE_LIMIT
 
 
 @njit(parallel=True)
@@ -396,9 +397,7 @@ def spin_glass_solver(
 
         local_work_group_size = min(wgs, n_qubits)
         global_work_group_size = n_qubits
-        gray_work_group_size = opencl_context.MAX_GPU_PROC_ELEM
         opencl_args = setup_opencl(local_work_group_size, global_work_group_size, np.array([n_qubits, is_spin_glass, segment_size]))
-        gray_args = setup_opencl(1, gray_work_group_size, np.array([n_qubits, is_spin_glass, (gray_iterations + gray_work_group_size - 1) // gray_work_group_size, segment_size]), True)
 
         if is_segmented:
             single_bit_flips_kernel = opencl_context.single_bit_flips_segmented_kernel
@@ -409,6 +408,11 @@ def spin_glass_solver(
             double_bit_flips_kernel = opencl_context.double_bit_flips_kernel
             gray_kernel = opencl_context.gray_kernel
 
+        if n_qubits <= gnl:
+            gray_work_group_size = opencl_context.MAX_GPU_PROC_ELEM
+            gray_args = setup_opencl(1, gray_work_group_size, np.array([n_qubits, is_spin_glass, (gray_iterations + gray_work_group_size - 1) // gray_work_group_size, segment_size]), True)
+            gray_kernel = opencl_context.gray_segmented_kernel if is_segmented else opencl_context.gray_kernel
+
     thread_count = os.cpu_count() ** 2
     improved = True
     while improved:
@@ -438,7 +442,7 @@ def spin_glass_solver(
             improved = True
             continue
 
-        if is_opencl:
+        if is_opencl and (n_qubits <= gnl):
             theta_buf_64 = make_best_theta_buf_64(best_theta)
             energy, state = run_gray_search_opencl(n_qubits, gray_kernel, max_energy, best_theta, theta_buf_64, G_m_buf, is_segmented, *gray_args)
         else: