Array copy in parallel_for_each context - c++

Question

I’m very newbie in AMP C++. Everything works fine if I use ‘memcpy’ inside the ‘parallel_for_each’ function, but I do know it is not the best practice. I tried to use ‘copy_to’, but it raises an exception. Below follows a simplified code, focusing the issue, that I am having troubles. Thanks in advance.
typedef std::vector<DWORD> CArrDwData;
class CdataMatrix
{
public:
CdataMatrix(int nChCount) : m_ChCount(nChCount)
{
}
void SetSize(UINT uSize)
{
// MUST be multiple of m_ChCount*DWORD
ASSERT(uSize%sizeof(DWORD) == 0);
m_PackedLength = uSize/sizeof(DWORD);
m_arrChannels.resize(m_ChCount*m_PackedLength);
}
UINT GetChannelPackedLen() const
{
return m_PackedLength;
}
const LPBYTE GetChannelBuffer(UINT uChannel) const
{
CArrDwData::const_pointer cPtr = m_arrChannels.data() + m_PackedLength*uChannel;
return (const LPBYTE)cPtr;
}
public:
CArrDwData m_arrChannels;
protected:
UINT m_ChCount;
UINT m_PackedLength;
};
void CtypDiskHeader::ParalelProcess()
{
const int nJobs = 6;
const int nChannelCount = 3;
UINT uAmount = 250000;
int vch;
CArrDwData arrCompData;
// Check buffers sizes
ASSERT((~uAmount & 0x00000003) == 3); // DWORD aligned
const UINT uInDWSize = uAmount/sizeof(DWORD); // in size give in DWORDs
CdataMatrix arrChData(nJobs);
arrCompData.resize(nJobs*uInDWSize);
vector<int> a(nJobs);
for(vch = 0; vch < nJobs; vch++)
a[vch] = vch;
arrChData.SetSize(uAmount+16); // note: 16 bytes or 4 DWORDs larger than uInDWSize
accelerator_view acc_view = accelerator().default_view;
Concurrency::extent<2> eIn(nJobs, uInDWSize);
Concurrency::extent<2> eOut(nJobs, arrChData.GetChannelPackedLen());
array_view<DWORD, 2> viewOut(eOut, arrChData.m_arrChannels);
array_view<DWORD, 2> viewIn(eIn, arrCompData);
concurrency::parallel_for_each(begin(a), end(a), [&](int vch)
{
vector<DWORD>::pointer ptr = (LPDWORD)viewIn(vch).data();
LPDWORD bufCompIn = (LPDWORD)ptr;
ptr = viewOut(vch).data();
LPDWORD bufExpandedIn = (LPDWORD)ptr;
if(ConditionNotOk())
{
// Copy raw data bufCompIn to bufExpandedIn
// Works fine, but not the best way, I suppose:
memcpy(bufExpandedIn, bufCompIn, uAmount);
// Raises exception:
//viewIn(vch).copy_to(viewOut(vch));
}
else
{
// Some data processing here
}
});
}

Answer 1

It was my fault. In the original code, the extent of viewOut(vch) is a little bit larger than viewIn(vch) extent. Using this way, it raises an exception 'runtime_exception'. When catching it, it supplies the following message xcp.what() = "Failed to copy because extents do not match".
I fixed the code replacing the original code by: viewIn(vch).copy_to(viewOut(vch).section(viewIn(vch).extent));
It copies only the source extent, that is what I need. But only compiles without restricted AMP.

Answer 2

The has nothing to do with the parallel_for_each it looks like it is a known bug with array_view::copy_to. See the following post:
Curiosity about concurrency::copy and array_view projection interactions
You can fix this using an explicit view_as() instead. I believe in your case your code should look something like this.
viewIn(vch).copy_to(viewOut(vch));
// Becomes...
viewIn[vch].view_as<1>(concurrency::extent<1>(uInDWSize)).copy_to(viewOut(vch));
I can't compile your example so was unable to verify this but I was able to get an exception from similar code and fix it using view_as().
If you want to copy data within a C++ AMP kernel then you need to do it as assignment operations on a series of threads. The following code copies the first 500 elements of source into the smaller dest array.
array<int, 1> source(1000);
array<int, 1> dest(500);
parallel_for_each(source.extent, [=, &source, &dest](index<1> idx)
{
if (dest.extent.contains(idx))
dest[idx] = source[idx];
});