changed gpu interface for mpi

GPy/kern/_src/psi_comp/rbf_psi_comp.py

@@ -6,12 +6,6 @@ import numpy as np
 from GPy.util.caching import Cacher
 
 def psicomputations(variance, lengthscale, Z, variational_posterior, return_psi2_n=False):
-    """
-    Z - MxQ
-    mu - NxQ
-    S - NxQ
-    gamma - NxQ
-    """
     # here are the "statistics" for psi0, psi1 and psi2
     # Produced intermediate results:
     # _psi1                NxM
@@ -26,12 +20,6 @@ def psicomputations(variance, lengthscale, Z, variational_posterior, return_psi2
     return psi0, psi1, psi2
 
 def __psi1computations(variance, lengthscale, Z, mu, S):
-    """
-    Z - MxQ
-    mu - NxQ
-    S - NxQ
-    gamma - NxQ
-    """
     # here are the "statistics" for psi1
     # Produced intermediate results:
     # _psi1                NxM
@@ -46,12 +34,6 @@ def __psi1computations(variance, lengthscale, Z, mu, S):
     return _psi1
 
 def __psi2computations(variance, lengthscale, Z, mu, S):
-    """
-    Z - MxQ
-    mu - NxQ
-    S - NxQ
-    gamma - NxQ
-    """
     # here are the "statistics" for psi2
     # Produced intermediate results:
     # _psi2                MxM
@@ -86,13 +68,6 @@ def psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscal
     return dL_dvar, dL_dlengscale, dL_dZ, dL_dmu, dL_dS
 
 def _psi1compDer(dL_dpsi1, variance, lengthscale, Z, mu, S):
-    """
-    dL_dpsi1 - NxM
-    Z - MxQ
-    mu - NxQ
-    S - NxQ
-    gamma - NxQ
-    """
     # here are the "statistics" for psi1
     # Produced intermediate results: dL_dparams w.r.t. psi1
     # _dL_dvariance     1
@@ -118,13 +93,6 @@ def _psi1compDer(dL_dpsi1, variance, lengthscale, Z, mu, S):
     return _dL_dvar, _dL_dl, _dL_dZ, _dL_dmu, _dL_dS
 
 def _psi2compDer(dL_dpsi2, variance, lengthscale, Z, mu, S):
-    """
-    Z - MxQ
-    mu - NxQ
-    S - NxQ
-    gamma - NxQ
-    dL_dpsi2 - MxM
-    """
     # here are the "statistics" for psi2
     # Produced the derivatives w.r.t. psi2:
     # _dL_dvariance      1

GPy/kern/_src/psi_comp/rbf_psi_gpucomp.py

@@ -7,13 +7,6 @@ from ....util.caching import Cache_this
 from . import PSICOMP_RBF
 from ....util import gpu_init
 
-try:
-    import pycuda.gpuarray as gpuarray
-    from pycuda.compiler import SourceModule
-    from ....util.linalg_gpu import sum_axis
-except:
-    pass    
-
 gpu_code = """
     // define THREADNUM
 
@@ -242,6 +235,10 @@ gpu_code = """
 class PSICOMP_RBF_GPU(PSICOMP_RBF):
 
     def __init__(self, threadnum=256, blocknum=30, GPU_direct=False):
+        from pycuda.compiler import SourceModule
+        from ....util.gpu_init import initGPU
+        initGPU()
+        
         self.GPU_direct = GPU_direct
         self.gpuCache = None
         
@@ -265,6 +262,7 @@ class PSICOMP_RBF_GPU(PSICOMP_RBF):
         return s
     
     def _initGPUCache(self, N, M, Q):
+        import pycuda.gpuarray as gpuarray
         if self.gpuCache == None:
             self.gpuCache = {
                              'l_gpu'                :gpuarray.empty((Q,),np.float64,order='F'),
@@ -320,13 +318,14 @@ class PSICOMP_RBF_GPU(PSICOMP_RBF):
     def get_dimensions(self, Z, variational_posterior):
         return variational_posterior.mean.shape[0], Z.shape[0], Z.shape[1]
 
-    @Cache_this(limit=1, ignore_args=(0,))
-    def psicomputations(self, variance, lengthscale, Z, variational_posterior):
+    @Cache_this(limit=5, ignore_args=(0,))
+    def psicomputations(self, kern, Z, variational_posterior, return_psi2_n=False):
         """
         Z - MxQ
         mu - NxQ
         S - NxQ
         """
+        variance, lengthscale = kern.variance, kern.lengthscale
         N,M,Q = self.get_dimensions(Z, variational_posterior)
         self._initGPUCache(N,M,Q)
         self.sync_params(lengthscale, Z, variational_posterior.mean, variational_posterior.variance)
@@ -355,8 +354,10 @@ class PSICOMP_RBF_GPU(PSICOMP_RBF):
         else:
             return psi0, psi1_gpu.get(), psi2_gpu.get()
 
-    @Cache_this(limit=1, ignore_args=(0,1,2,3))
-    def psiDerivativecomputations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior):
+    @Cache_this(limit=5, ignore_args=(0,2,3,4))
+    def psiDerivativecomputations(self, kern, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
+        variance, lengthscale = kern.variance, kern.lengthscale
+        from ....util.linalg_gpu import sum_axis
         ARD = (len(lengthscale)!=1)
         
         N,M,Q = self.get_dimensions(Z, variational_posterior)

GPy/kern/_src/psi_comp/sslinear_psi_comp.py

@@ -9,7 +9,7 @@ from ....util.linalg import tdot
 
 import numpy as np
 
-def psicomputations(variance, Z, variational_posterior):
+def psicomputations(variance, Z, variational_posterior, return_psi2_n=False):
     """
     Compute psi-statistics for ss-linear kernel
     """

GPy/kern/_src/psi_comp/ssrbf_psi_gpucomp.py

@@ -6,14 +6,7 @@ The module for psi-statistics for RBF kernel for Spike-and-Slab GPLVM
 import numpy as np
 from ....util.caching import Cache_this
 from . import PSICOMP_RBF
-from ....util import gpu_init
 
-try:
-    import pycuda.gpuarray as gpuarray
-    from pycuda.compiler import SourceModule
-    from ....util.linalg_gpu import sum_axis
-except:
-    pass    
 
 gpu_code = """
     // define THREADNUM
@@ -292,6 +285,11 @@ gpu_code = """
 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
         
@@ -315,6 +313,7 @@ class PSICOMP_SSRBF_GPU(PSICOMP_RBF):
         return s
 
     def _initGPUCache(self, N, M, Q):
+        import pycuda.gpuarray as gpuarray
         if self.gpuCache == None:
             self.gpuCache = {
                              'l_gpu'                :gpuarray.empty((Q,),np.float64,order='F'),
@@ -377,12 +376,13 @@ class PSICOMP_SSRBF_GPU(PSICOMP_RBF):
         return variational_posterior.mean.shape[0], Z.shape[0], Z.shape[1]
 
     @Cache_this(limit=1, ignore_args=(0,))
-    def psicomputations(self, variance, lengthscale, Z, variational_posterior):
+    def psicomputations(self, kern, Z, variational_posterior, return_psi2_n=False):
         """
         Z - MxQ
         mu - NxQ
         S - NxQ
         """
+        variance, lengthscale = kern.variance, kern.lengthscale
         N,M,Q = self.get_dimensions(Z, variational_posterior)
         self._initGPUCache(N,M,Q)
         self.sync_params(lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
@@ -409,8 +409,10 @@ class PSICOMP_SSRBF_GPU(PSICOMP_RBF):
         else:
             return psi0, psi1_gpu.get(), psi2_gpu.get()
 
-    @Cache_this(limit=1, ignore_args=(0,1,2,3))
-    def psiDerivativecomputations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior):
+    @Cache_this(limit=1, ignore_args=(0,2,3,4))
+    def psiDerivativecomputations(self, kern, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
+        variance, lengthscale = kern.variance, kern.lengthscale
+        from ....util.linalg_gpu import sum_axis
         ARD = (len(lengthscale)!=1)
         
         N,M,Q = self.get_dimensions(Z, variational_posterior)

GPy/util/__init__.py

@@ -15,6 +15,5 @@ from . import caching
 from . import diag
 from . import initialization
 from . import multioutput
-from . import linalg_gpu
 from . import parallel
 

GPy/util/gpu_init.py

@@ -16,7 +16,10 @@ try:
 except:
     pass
 
-try:
+
+
+def initGPU():
+    try:
         if MPI_enabled and MPI.COMM_WORLD.size>1:
             from .parallel import get_id_within_node
             gpuid = get_id_within_node()
@@ -24,7 +27,6 @@ try:
             pycuda.driver.init()
             if gpuid>=pycuda.driver.Device.count():
                 print('['+MPI.Get_processor_name()+'] more processes than the GPU numbers!')
-            #MPI.COMM_WORLD.Abort()
                 raise
             gpu_device = pycuda.driver.Device(gpuid)
             gpu_context = gpu_device.make_context()
@@ -32,15 +34,7 @@ try:
         else:
             import pycuda.autoinit
             gpu_initialized = True
-except:
-    pass
-
-try:
-    from scikits.cuda import cublas
-    import scikits.cuda.linalg as culinalg
-    culinalg.init()
-    cublas_handle = cublas.cublasCreate()
-except:
+    except:
         pass
     
 def closeGPU():

GPy/util/linalg_gpu.py

@@ -61,12 +61,5 @@ try:
 except:
     pass
 
-try:
-    import scikits.cuda.linalg as culinalg
-    from scikits.cuda import cublas
-    from scikits.cuda.cula import culaExceptions
-except:
-    pass
-