Crop layer remove redundance

[xmyqsh]

remove the redundancy for loop in crop.cpp
remove the redundancy compute in crop.cu

[williford]

The change in the crop_layer.cpp file looks good to me. The code was in a for-loop when it does the same thing every loop (index is used and nothing is incremented).

I haven't figured out how the crop_layer.cu changes are equivalent yet. Can you give some explanation?

[xmyqsh]

@williford glad to see you interesting my PR
case
const int lines = top[0]->shape(cur_dim);
const int height = top[0]->shape(cur_dim);
so
lines is equal to height
from
__global__ void copy_kernel(const int n, const int height, const int width,
copy_kernel<<<CAFFE_GET_BLOCKS(lines), CAFFE_CUDA_NUM_THREADS>>>( lines, height, width,
we know n is equal to lines
from
CUDA_KERNEL_LOOP(index, n) {
we know index is less than n
so index is also less than height
so index / height is always zero
and index % height is always index

so
int src_start = index / height * src_outer_stride + index % height * src_inner_stride;
is equivalent to
int src_start = index * src_inner_stride;

these simplicity can remove N*C*H*2 times division, multiply, mod and add operations in one forward or backwark process

[shelhamer]

@BlGene could you take a look when you have the chance?

[BlGene]

It makes sense that we don't need to iterate over the size of the last spatial dimension because we are copying it in one go, as it is stored in memory continuously.

Sorry for overlooking this before.

@xmyqsh: Thanks for finding this, can you please squash the PR.
@shelhamer: This looks correct to me.

[shelhamer]

Thanks @xmyqsh for the simplification and removal of the redundant operations.