From a77a6755495fe3871ca9aa53dd107554f26cc01b Mon Sep 17 00:00:00 2001
From: Zhenwen Dai <z.dai@sheffield.ac.uk>
Date: Tue, 9 Feb 2016 17:22:38 +0000
Subject: [PATCH] fix gpu initialziation

---
 GPy/kern/src/psi_comp/rbf_psi_gpucomp.py   |  3 --
 GPy/kern/src/psi_comp/ssrbf_psi_gpucomp.py |  2 -
 GPy/util/gpu_init.py                       | 52 ++++++++++++----------
 3 files changed, 29 insertions(+), 28 deletions(-)

diff --git a/GPy/kern/src/psi_comp/rbf_psi_gpucomp.py b/GPy/kern/src/psi_comp/rbf_psi_gpucomp.py
index baab83ec..e9268a3e 100644
--- a/GPy/kern/src/psi_comp/rbf_psi_gpucomp.py
+++ b/GPy/kern/src/psi_comp/rbf_psi_gpucomp.py
@@ -5,7 +5,6 @@ The module for psi-statistics for RBF kernel
 import numpy as np
 from paramz.caching import Cache_this
 from . import PSICOMP_RBF
-from ....util import gpu_init
 
 gpu_code = """
     // define THREADNUM
@@ -238,8 +237,6 @@ class PSICOMP_RBF_GPU(PSICOMP_RBF):
         self.fall_back = PSICOMP_RBF()
         
         from pycuda.compiler import SourceModule
-        from ....util.gpu_init import initGPU
-        initGPU()
         
         self.GPU_direct = GPU_direct
         self.gpuCache = None
diff --git a/GPy/kern/src/psi_comp/ssrbf_psi_gpucomp.py b/GPy/kern/src/psi_comp/ssrbf_psi_gpucomp.py
index 844f944e..62876785 100644
--- a/GPy/kern/src/psi_comp/ssrbf_psi_gpucomp.py
+++ b/GPy/kern/src/psi_comp/ssrbf_psi_gpucomp.py
@@ -287,8 +287,6 @@ class PSICOMP_SSRBF_GPU(PSICOMP_RBF):
     def __init__(self, threadnum=128, blocknum=15, GPU_direct=False):
         
         from pycuda.compiler import SourceModule
-        from ....util.gpu_init import initGPU
-        initGPU()
         
         self.GPU_direct = GPU_direct
         self.gpuCache = None
diff --git a/GPy/util/gpu_init.py b/GPy/util/gpu_init.py
index 94763d8b..19339b91 100644
--- a/GPy/util/gpu_init.py
+++ b/GPy/util/gpu_init.py
@@ -10,29 +10,35 @@ gpu_device = None
 gpu_context = None
 MPI_enabled = False
 
-def initGPU():
-    try:
-        from mpi4py import MPI
-        MPI_enabled = True
-    except:
-        pass
-    try:
-        if MPI_enabled and MPI.COMM_WORLD.size>1:
-            from .parallel import get_id_within_node
-            gpuid = get_id_within_node()
-            import pycuda.driver
-            pycuda.driver.init()
-            if gpuid>=pycuda.driver.Device.count():
-                print('['+MPI.Get_processor_name()+'] more processes than the GPU numbers!')
-                raise
-            gpu_device = pycuda.driver.Device(gpuid)
-            gpu_context = gpu_device.make_context()
-            gpu_initialized = True
-        else:
-            import pycuda.autoinit
-            gpu_initialized = True
-    except:
-        pass
+try:
+    import pycuda.autoinit
+    gpu_initialized = True
+except:
+    pass
+
+# def initGPU():
+#     try:
+#         from mpi4py import MPI
+#         MPI_enabled = True
+#     except:
+#         pass
+#     try:
+#         if MPI_enabled and MPI.COMM_WORLD.size>1:
+#             from .parallel import get_id_within_node
+#             gpuid = get_id_within_node()
+#             import pycuda.driver
+#             pycuda.driver.init()
+#             if gpuid>=pycuda.driver.Device.count():
+#                 print('['+MPI.Get_processor_name()+'] more processes than the GPU numbers!')
+#                 raise
+#             gpu_device = pycuda.driver.Device(gpuid)
+#             gpu_context = gpu_device.make_context()
+#             gpu_initialized = True
+#         else:
+#             import pycuda.autoinit
+#             gpu_initialized = True
+#     except:
+#         pass
     
 def closeGPU():
     if gpu_context is not None: