imshp - image shape tuple of 2 or 3: 2 for a 2d image, 3 for a stack of 2d images.
kshp - kernel shape 2
nkern - # kernels
bsize - batch size
dx - patch stride rows
dy - patch stride cols
out_mode - 'valid', 'full'
unroll_patch - c code generation option(used when no shape gived)
unroll_batch - c code generation option
unroll_kern - c code generation option
verbose - passed to GpuConv
version - passed to GpuConv
If the imshp, kshp, nkern and bsize are provided, we can generate more optimal code. This make a significant difference for the full mode with unroll_patch version.
This Op implement the convolution of a kernel(tensor 4d,(nkern, stacksize, nb row, nb col)) on an image(tensor 4d, (batchsize, stacksize, nb row, nb col). The batch size is multiple image that we want to apply the same kernel over. The nkern is numtiple kernel that we want to apply to each image. The stack size is mostly used when their is multiple layer in the network. It is the sum of the convolution of multiple 2d image and kernel.
The reason that this op does the summation over convolutions within the 'stack' is that
it allows us to be memory-efficient about how gradients are calculated. If, for
...
...
@@ -63,17 +48,52 @@ class ConvOp(Op):
gradient on the filters.
unroll_patch. If True will use a version that is faster then without not unroll by unroll the patch loop.
unroll_batch. If >0 will use a version that will unroll the batch loop by the value of the option. By default don't use this version of the code.
unroll_nkern. idem as unroll_batch but unroll the kernel loop.
If the imshp, kshp, nkern and bsize are provided, we can generate more optimal code. This make a significant difference for the full mode with unroll_patch version.
The most frequent faster code currently available on 64_x86 computer is unroll_batch=4, unroll_kern=4, unroll_patch=False and this request that all the optional shape information are gived. Those number are empirically tested and backed up by the article: Anatomy of High-Performance Matrix Multiplication by Kazushige Goto and Robert A. Van De Geijn, ACM Transactions on Mathematical Software, vol 34, No. 3, article 12, May 2008. It is in figure 12, it give the value mr x nr, those value are the optimum to use for unroll_batch and unroll_kern. For x86_64 bits computer it is 4x4. Other architecture can have different value.(2x4 for x86, 8x8 for itanium,...)
:type out_mode: string
:param out_mode: 'valid'(give an output smaller then the image, 'full'(give an output bigger then the image)
The version is with unroll_batch=4 and unroll_nkern if possible(currenctly it don't support logical shape != physical shape) as this is what give the best performance in practice. This also tell that to have the best performance, you should have a batch size and a number of kernel multiple of 4. In the article:
Anatomy of High-Performance Matrix Multiplication by Kazushige Goto and Robert A. Van De Geijn, ACM Transactions on Mathematical Software, vol 34, No. 3, article 12, May 2008.
In figure 12, it give the value mr x nr, those value are the optimum to use for unroll_batch and unroll_kern. For x86_64 bits computer it is 4x4. Other architecture can have different value.(2x4 for x86, 8x8 for itanium,...)
optional parameter(if provided will be used to generate more optinal c code):
:type imshp: tuple of len 2 or 3: 2 for 2d image, 3 for a stack of 2d images.
:param imshp: (stacksize, nb image row, nb image col)
:type kshp: tuple of len 2
:param kshp: (nb kernel row, nb kernel col)
:type nkern: int
:param nkern: the number of kernel
:type bsize: int
:param bsize: the size of the minibatch
:type dx: int
:param dx: patch stride rows
:type dy: int
:param dx: patch stride cols
param to select the version of code used:
:type unroll_patch: bool
:param unroll_patch: use a version of c_code that unroll the patch loop that don't request all shape information to work, but if all shape information are present, will use it to hardcode the value in the code for faster code.
:type unroll_batch:int
:param unroll_batch: use a version of c_code that unroll the batch(by unroll_batch) and the nkern(by unroll_kern) loop. The size must by a multiple of bsize or nkern respectively.
:type unroll_kern:int
:param unroll_kern: use a version of c_code that unroll the batch(by unroll_batch) and the nkern(by unroll_kern) loop. The size must by a multiple of bsize or nkern respectively.
:type verbose: int
:param verbose: passed to GpuConv
:type version: int
:param version: passed to GpuConv
:param imshp_logical: used internally when we generate the gradient when dx!=1 or dy!=1
