//if we remove the restriction img_size_byte+kern_size_byte>8*1024, we can enter in condition where we will lower the occupency due to shared memory and/or registers.
if ((version == -1) && (out_size<64 || img_size_byte+kern_size_byte>8*1024) && out_size<=256){
//if we remove the restriction
//img_size_byte+kern_size_byte>8*1024, we can enter in condition where
//we will lower the occupency due to shared memory and/or registers.
if (verbose) fprintf(stderr, "INFO: used 'conv_rows_stack2' version %s with %d row(s).\n",(version==9?"'load full kernel'":"'load 1 kern row at a time'"),nb_row);
CudaNdarray_conv_full(const CudaNdarray *img, const CudaNdarray * kern, CudaNdarray * out, int subsample_rows, int subsample_cols, int version = -1, int verbose=0)
bool img_batch_stack_contiguous = (img_stride_stack==img_stride_row*img_len) && (img_stride_batch==img_stride_stack*nstack);//don't support stride for nbatch and nstack
//if the lower 2 dims are c_contiguous but flipped, unflipping the stride and not flipping the kernel in shared memroy
//if the lower 2 dims are c_contiguous but flipped, unflipping the
//stride and not flipping the kernel in shared memroy
//allow to use a version that use less registers(so is faster)
//the unflipped version of variable have the original value when we don't need to unflip it, but have the new value when we unflip it.
//the unflipped version of variable have the original value when
//we don't need to unflip it, but have the new value when we unflip it.
//we pass by ceil_intdiv in case the out_len is not a multiple of nb_split, we want nb_split the number of iteration.
//we pass by ceil_intdiv in case the out_len is not a multiple
//of nb_split, we want nb_split the number of iteration.
//Max of 16k of shared memory
if(version==5)
while ((((kern_len+ceil_intdiv(out_len,nb_split)-1)+2*kern_len-2)*img_wid_padded*sizeof(float) + kern_size_byte)>shared_avail) nb_split++;
//327 as we use 25 register
//version 5 will have only 1 block running at a time, so we can use 32 registers per threads, but their is some other stuff that for the limit to bu lower then 512.
//version 5 will have only 1 block running at a time, so we
//can use 32 registers per threads, but their is some other stuff that
//for the limit to bu lower then 512.
int max_thread = (version!=5?327:450);
while (ceil_intdiv(out_len,nb_split)*out_wid>max_thread) nb_split++;
