The caffe's timer has some overhead, and when our tunning kernel is
very tiny, the overhead may cause very unstable timing result, so
I increase the iteration count to reduce this type of overhead.

Signed-off-by: Zhigang Gong <[email protected]>

If the spatial dimension is relatively large, we should use the default code
path to achieve better parallelism.

Signed-off-by: Zhigang Gong <[email protected]>

Sometimes, the sub buffer creation may fail, we need to take care of it.

Signed-off-by: Zhigang Gong <[email protected]>

Some features e.g. opencl_unroll_hint are not allowed for beignet
compiler, use __BEIGNET__ macro to choose whether to build with these
features.

Also add an helper func to faciliate judging beignet driver.

Signed-off-by: Zhiwen Wu <[email protected]>

If the input image size changed during runtime, and the kernel type
change to 2 or 5, we need to swizzle the weights again.

Signed-off-by: Zhigang Gong <[email protected]>

Added a new basic convolution kernel that supports input image with
no padding, so that no image padding in host code need anymore.

Signed-off-by: Zhiwen Wu<[email protected]>

Signed-off-by: Zhigang Gong <[email protected]>

Change-Id: I392c4e73319fcfc18e628f9476b9bfdcba3cc206

Signed-off-by: Zhigang Gong <[email protected]>

If we simply use the cpu code path to copy the data, we will introduce
one race condition between the GPU queue and the CPU. The scenario is:
when we call it in an iteration loop. The data blob is in a zero-copy
blob, and the first pass may be still blocking on the GPU side. The
second pass will modify the data blob on CPU side before the data is
accessed at the first pass on GPU side.

We can simply add a synchronization point between the two iterations, but that
is not a good fix as we force the GPU queue to flush and wait it to finish.
The best way is to do the copy on the GPU side and in the same queue. Thus we
don't need to worry about this race condition any more and without any interfere
the GPU queue.

Signed-off-by: Zhigang Gong <[email protected]>

This will cause relu gradient fail.

Prepare to support varying sizes.

Signed-off-by: Zhigang Gong <[email protected]>

No need to tune different kernel for each different input size.

Signed-off-by: Zhigang Gong <[email protected]>

Add the platform and driver information to change to use
system cache directory if possible. After this change, we
can reuse a offline tuned configurations.

Signed-off-by: Zhigang Gong <[email protected]>

Signed-off-by: Zhigang Gong <[email protected]>

Signed-off-by: Zhigang Gong <[email protected]>

Signed-off-by: Zhigang Gong <[email protected]>

Signed-off-by: Zhigang Gong <[email protected]>