Package: Major fall update:
-> Major performance improvement of CONV for small batches -> Adopted NCHW image layout -> 3-D convolutions supported -> HFMA2 deprecated -> Multi-threaded auto-tuning -> Tensorflow Wrapper
This commit is contained in:
@@ -33,25 +33,28 @@ namespace isaac
|
||||
namespace driver
|
||||
{
|
||||
|
||||
|
||||
Buffer::Buffer(Context const & context, size_t size) : context_(context), size_(size)
|
||||
Buffer::Buffer(Context const & context, size_t size) : context_(context)
|
||||
{
|
||||
ContextSwitcher ctx_switch(context_);
|
||||
dispatch::cuMemAlloc(&*cu_, size);
|
||||
}
|
||||
|
||||
Buffer::Buffer(Context const & context, CUdeviceptr cu, bool take_ownership):
|
||||
context_(context), cu_(cu, take_ownership)
|
||||
{ }
|
||||
|
||||
void Buffer::set_zero(Stream const & queue, size_t size)
|
||||
{
|
||||
ContextSwitcher ctx_switch(context_);
|
||||
dispatch::cuMemsetD8Async(*cu_, 0, size, queue);
|
||||
}
|
||||
|
||||
void Buffer::set_zero(Stream const & queue)
|
||||
{ set_zero(queue, size_); }
|
||||
|
||||
Handle<CUdeviceptr> const & Buffer::cu() const
|
||||
{ return cu_; }
|
||||
|
||||
Handle<CUdeviceptr> & Buffer::cu()
|
||||
{ return cu_; }
|
||||
|
||||
}
|
||||
|
||||
}
|
||||
|
Reference in New Issue
Block a user