std::max(int(img_size_byte+2*kern_wid*sizeof(float)), out_size_byte*2)<shared_avail && //their is only 16k of shared memory and if we can't have the output at least twice in shared mem, we won't have any reduce!
!work_complete)
version = 7; //conv_patch_stack_reduce, switch to version 8/13 automatically if needed.
out_contiguous &&
out_size<512 &&//Maximum of 512 theads by block
std::max(int(img_size_byte+2*kern_wid*sizeof(float)), out_size_byte*2)<shared_avail && //their is only 16k of shared memory and if we can't have the output at least twice in shared mem, we won't have any reduce!
!work_complete)
version = 7; //conv_patch_stack_reduce, switch to version 8/13 automatically if needed.
}
if (!subsample && c_contiguous &&
(version==0||version==2||version==-1) &&
out_wid<512 &&//Maximum of 512 theads by block
nstack == 1 &&// don't implement the stack in the kernel.
img_size_byte+kern_size_byte<shared_avail && //their is only 16k of shared memory
!work_complete) //conv_patch
(version==0||version==2||version==-1) &&
out_wid<512 &&//Maximum of 512 theads by block
nstack == 1 &&// don't implement the stack in the kernel.
img_size_byte+kern_size_byte<shared_avail && //their is only 16k of shared memory
!work_complete) //conv_patch
{
int nb_split=1;//The number of split (i.e. the number of output pixel each thread compute.)
if(version==2 && out_len>1)nb_split++;//to force the use of split=true when testing.
//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.
while (ceil_intdiv(out_len,nb_split)*out_wid>512) nb_split++;
if(version==2 && out_len>1)nb_split++;//to force the use of split=true when testing.
//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.
while (ceil_intdiv(out_len,nb_split)*out_wid>512) nb_split++;
if (false && !subsample && //disabled as test fail for this kernel
(version==1||version==-1) &&
out_size<512 &&//Maximum of 512 theads by block
(version==1||version==-1) &&
out_size<512 &&//Maximum of 512 theads by block
(nbatch > 20 || version==1) && // we only launch nbatch blocks, so make sure there is enough to be worth it, but if we specify the version, this check should not be done to allow testing.
nstack*img_size_byte+nstack*kern_size_byte<shared_avail && //there is only 16k of shared memory
!work_complete) //conv_full_load_everything
nstack*img_size_byte+nstack*kern_size_byte<shared_avail && //there is only 16k of shared memory
!work_complete) //conv_full_load_everything
{
dim3 threads(out_wid, out_len);
dim3 grid(nbatch);
int shared_size=(img_size + kern_size)*nstack*sizeof(float);
//TODO assert c_continious for img, kern and out in the 2 inner dimensions.
//TODO assert c_continious for img, kern and out in the 2 inner dimensions.
