cudaMemcpyAsync return invalid paramater

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hello,

I’m having a problem now.
I use host mapping and stream API.
It is, in order to exceed the wall of the device memory size.
And faster?

init
{
cudaSetDevice(*DeviceID)
cudaSetDeviceFlags(cudaDeviceMapHost);
・
・
}

processing device memory case
{
cudaMemcpyAsync(dst, src, memsize, cudaMemcpyHostToDevice, stream);
kernel1<<< , 0,stream>>>();
kernel2<<< , 0,stream>>>();
cudaMemcpyAsync(dst, src, memsize, cudaMemcpyDeviceToHost, stream);
cudaStreamSynchronize(stream);
}

processing host mapped memory case
{
cudaHostRegister();
cudaHostGetDevicePointer();
kernel1<<< , 0,stream>>>();
kernel2<<< , 0,stream>>>();
cudaStreamSynchronize(stream);
cudaHostUnregister();
}

It may not work cudaMemcpyAsync()
memsize=10467126 NG
memsize=11520000 OK
There a limitation of size, aligned and MapHost?

my environment
CUDA5.0 + Capability 3.0 GPU + VS2010
Driver Vrsion 327.23

#2

Because have a difficult how to use?
That I am using to split to get the large size seems to cause

cudaMalloc(&ptr, size);
・
・
CalcBuffer(ptr, x, y, z)
{
p[0] = aligin64( x * y ); <- aligin 64byte
p[1] = p[0] + aligin64( x * y );
p[2] =
・
・
}