How to get pixeldata without using LockRect? (Performance issue) - c++

Question

I'm using these function someone post on StackOverflow, but the performance is toooooooo bad..
With out using this function( not capture pixel ) FPS show over 6k, but using this function, FPS show only 60!
So is there any fast method that can copy between video memory and system memory?
Anyone please?
HRESULT capture(IDirect3DDevice9* m_d3ddev, void* buffer, int& width, int& height, D3DFORMAT format)
{
IDirect3DSurface9 *renderTarget = NULL;
IDirect3DSurface9 *destTarget = NULL;
HRESULT hr = m_d3ddev->GetRenderTarget(0, &renderTarget);
hr = m_d3ddev->CreateOffscreenPlainSurface(width, height, format, D3DPOOL_SYSTEMMEM, &destTarget, NULL);
if (FAILED(hr))
{
printf("Failed CreateOffscreenPlainSurface!");
}
hr = m_d3ddev->GetRenderTargetData(renderTarget, destTarget);
if (FAILED(hr))
{
printf("Failed GetRenderTargetData!");
}
D3DLOCKED_RECT lr;
ZeroMemory(&lr, sizeof(D3DLOCKED_RECT));
hr = destTarget->LockRect(&lr, 0, D3DLOCK_READONLY);
if (FAILED(hr))
{
printf("Cannot lock rect!");
}
if (lr.pBits)
{
memcpy(buffer, lr.pBits, width * height * 4);
}
hr = destTarget->UnlockRect();
if (FAILED(hr))
{
printf("Cannot unlock rect!");
}
renderTarget->Release();
destTarget->Release();
return hr;
}