Blender Git Loki

Cycles: Rework work stealing

Previous implementation of work stealing didn't do what it claimed to do
and made some odd assumptions. This new implementation is cleaner,
organizes work in a slightly more optimal way, and most importantly
decouples work from tile size, which will allow for greater speed ups.

Cycles: Silence warning about redefined macro

Cycles: Remove everything parallel samples from the split kernel

Parallel samples never actually worked, was producing incorrect results
or crashes, and wasn't any faster than work stealing, so removing it.

Merge branch 'master' into cycles_split_kernel

Cycles: Dont use array indexing for non array

Cycles: Seperate kernel loading time from render time

Cycles: More variable clean up

Cycles: Clean up variables in host side split kernel logic

Was a bit messy, lots of duplication, hard to tell what was what.

Cycles: Fix building with network device enabled

Cycles: Implement split kernel for CUDA

Helpful to have another device to test against and check that the split
kernel remains compatible with all devices.

Cycles: Tile updates for split kernel

Better feedback is given from tiles by updating while rendering, where
as before tiles would only be displayed when fully complete. Rendering
is a little slower from this, but there may be ways to reduce that
effect at some point.

Cycles: Add names to buffer allocations

This is to help debug and track memory usage for generic buffers. We
have similar for textures already since those require a name, but for
buffers the name is only for debugging proposes.

Merge branch 'master' into cycles_split_kernel

Cycles: Add a bit more error checking for split kernel

This helps catch some memory issues and avoids some crashes, but still
need to figure out a way to handle this better.

Merge branch 'master' into cycles_split_kernel

Cycles: Keep consistent memory management style

`unique_ptr` avoids some issues but we aren't using it anywhere else so
to keep things consistent just using plain `delete` for now.

Cycles: Make split kernel methods in Device private

These methods shouldn't really be called from anywhere except
`DeviceSplitKernel`.

Cycles: Remove overloaded Device::mem_alloc function

It was a bit confusing having two versions of this function, only one is
needed really. Added `device_memory::resize()` to take over the behavior
the overload provided.

Cycles: Ensure we get a compiler error if data_init parameters fall out of sync

Code isn't so pretty and the errors that we get if this does happen are kind
of cryptic, but it's better than nothing.

Cycles: Rename split_state and split_params

Now these are `kernel_split_state` and `kernel_split_params`, as well as
being references instead of pointers. This is to match `kernel_data`.