def shutdown():
global _handles, _pid
pid = os.getpid()
if _pid != pid: # not initialized
return
for handle in _handles.itervalues():
libcudnn.cudnnDestroy(handle)
_handles = {}
_pid = None # mark as uninitialized
def shutdown():
global _handles, _pid
pid = os.getpid()
if _pid != pid: # not initialized
return
for handle in _handles.itervalues():
libcudnn.cudnnDestroy(handle)
_handles = {}
_pid = None # mark as uninitialized
libcudnn.cudnnConvolutionForward(cudnn_context, alpha, X_desc, X_data,
filters_desc, filters_data, conv_desc, algo, ws_data, ws_size.value, beta,
Y_desc, Y_data)
start_bench()
for i in range(10):
libcudnn.cudnnConvolutionForward(cudnn_context, alpha, X_desc, X_data,
filters_desc, filters_data, conv_desc, algo, ws_data, ws_size.value, beta,
Y_desc, Y_data)
ms = end_bench()
ws_ptr = None
libcudnn.cudnnDestroyTensorDescriptor(X_desc)
libcudnn.cudnnDestroyTensorDescriptor(Y_desc)
libcudnn.cudnnDestroyFilterDescriptor(filters_desc)
libcudnn.cudnnDestroyConvolutionDescriptor(conv_desc)
return ms / 10
# for kw in range(1, 11):
# for kh in range(1, 11):
# ms = benchmark_conv(kw, kh)
# print("%dx%d : %fms" % (kw, kh, ms))
for bsz in range(1, 32):
ms = benchmark_conv(11, 11, bsz)
print("%d : %.2fms => %f img/sec" % (bsz, ms, bsz/ms))
# Clean up
libcudnn.cudnnDestroy(cudnn_context)
maxU = parU[0:1,0:1]
maxo = ng.max(abs(cuO - nlO.T), partial=parO, out=maxO).get()[0,0]
maxb = ng.max(abs(cuB - nlB.T), partial=parB, out=maxB).get()[0,0]
maxu = ng.max(abs(cuU - nlU.T), partial=parU, out=maxU).get()[0,0]
meano = ng.mean(abs(cuO), partial=parO, out=maxO).get()[0,0]
meanb = ng.mean(abs(cuB), partial=parB, out=maxB).get()[0,0]
meanu = ng.mean(abs(cuU), partial=parU, out=maxU).get()[0,0]
print " maxerr mean pct"
print "fprop: %7.5f %6.2f %5.3f" % (maxo, meano, 100*maxo/meano)
print "bprop: %7.5f %6.2f %5.3f" % (maxb, meanb, 100*maxb/meanb)
print "updat: %7.5f %6.2f %5.3f" % (maxu, meanu, 100*maxu/meanu)
# free up memory from this layer before proceeding
cuB = cuU = cuO = None
nlB = nlU = nlO = None
parO = parB = parU = maxO = maxB = maxU = None
libcudnn.cudnnDestroyTensorDescriptor(I_desc)
libcudnn.cudnnDestroyTensorDescriptor(O_desc)
libcudnn.cudnnDestroyFilterDescriptor(F_desc)
libcudnn.cudnnDestroyTensorDescriptor(E_desc)
libcudnn.cudnnDestroyTensorDescriptor(B_desc)
libcudnn.cudnnDestroyFilterDescriptor(U_desc)
libcudnn.cudnnDestroyConvolutionDescriptor(C_desc)
libcudnn.cudnnDestroy(cudnn)
X_data = ctypes.c_void_p(int(X.gpudata))
filters_data = ctypes.c_void_p(int(filters.gpudata))
Y_data = ctypes.c_void_p(int(Y.gpudata))
# Perform convolution
algo = libcudnn.cudnnGetConvolutionForwardAlgorithm(cudnn_context, X_desc,
filters_desc, conv_desc, Y_desc, convolution_fwd_pref, 0)
print("Cudnn algorithm = %d" % algo.value)
ws_size = libcudnn.cudnnGetConvolutionForwardWorkspaceSize(cudnn_context, X_desc, filters_desc, conv_desc, Y_desc, algo)
ws_ptr = drv.mem_alloc(ws_size.value) if ws_size.value > 0 else 0
ws_data = ctypes.c_void_p(int(ws_ptr))
start_bench()
libcudnn.cudnnConvolutionForward(cudnn_context, alpha, X_desc, X_data,
filters_desc, filters_data, conv_desc, algo, ws_data, ws_size.value, beta,
Y_desc, Y_data)
end_bench("fprop")
ws_ptr = None
# Clean up
libcudnn.cudnnDestroyTensorDescriptor(X_desc)
libcudnn.cudnnDestroyTensorDescriptor(Y_desc)
libcudnn.cudnnDestroyFilterDescriptor(filters_desc)
libcudnn.cudnnDestroyConvolutionDescriptor(conv_desc)
libcudnn.cudnnDestroy(cudnn_context)
maxB = parB[0:1, 0:1]
maxU = parU[0:1, 0:1]
maxo = ng.max(abs(cuO - nlO.T), partial=parO, out=maxO).get()[0, 0]
maxb = ng.max(abs(cuB - nlB.T), partial=parB, out=maxB).get()[0, 0]
maxu = ng.max(abs(cuU - nlU.T), partial=parU, out=maxU).get()[0, 0]
meano = ng.mean(abs(cuO), partial=parO, out=maxO).get()[0, 0]
meanb = ng.mean(abs(cuB), partial=parB, out=maxB).get()[0, 0]
meanu = ng.mean(abs(cuU), partial=parU, out=maxU).get()[0, 0]
print " maxerr mean pct"
print "fprop: %7.5f %6.2f %5.3f" % (maxo, meano, 100 * maxo / meano)
print "bprop: %7.5f %6.2f %5.3f" % (maxb, meanb, 100 * maxb / meanb)
print "updat: %7.5f %6.2f %5.3f" % (maxu, meanu, 100 * maxu / meanu)
# free up memory from this layer before proceeding
cuB = cuU = cuO = None
nlB = nlU = nlO = None
parO = parB = parU = maxO = maxB = maxU = None
libcudnn.cudnnDestroyTensorDescriptor(I_desc)
libcudnn.cudnnDestroyTensorDescriptor(O_desc)
libcudnn.cudnnDestroyFilterDescriptor(F_desc)
libcudnn.cudnnDestroyTensorDescriptor(E_desc)
libcudnn.cudnnDestroyTensorDescriptor(B_desc)
libcudnn.cudnnDestroyFilterDescriptor(U_desc)
libcudnn.cudnnDestroyConvolutionDescriptor(C_desc)
libcudnn.cudnnDestroy(cudnn)