perf(dnn/relayout): disable copy_last_contiguous when contiguous_size is

small

GitOrigin-RevId: 7bd5e793a4f252510d739ae8b39046cea969cb58

dnn/src/cuda/relayout/kern_contiguous.cuh

@@ -232,7 +232,7 @@ DEFINE_CONTIG_RECEIVER(7, cb_header, cb_dispatch, m_param[1].layout)
             typedef OpCallerBinaryContiguous<PVis0, PVis1> Caller;
             size_t size = m_param.size;
             int grid_size, block_size;
-            if (m_contiguous_size > 32) {
+            if (m_contiguous_size >= 32) {
                 void (*fptr)(Caller, uint32_t, uint32_t, uint32_t, uint32_t);
                 fptr = cuda_last_contiguous_large_kern<Caller>;
                 safe_size_in_kern(size);

dnn/src/cuda/relayout/opr_impl.cpp

@@ -176,7 +176,8 @@ bool RelayoutForwardImpl::Param::try_copy_last_contig() {
             !has_negative_stride(lsrc) && !has_negative_stride(ldst)) {
         size_t contiguous_size =
             gcd(lsrc.shape[lsrc.ndim - 1], ldst.shape[ldst.ndim - 1]);
-        if (contiguous_size > 1) {
+        // FIXME: disable copy_last_contiguous when contiguous_size < 32 due to performance issue
+        if (contiguous_size >= 32) {
             copy_last_contiguous(m_dst, m_src, contiguous_size,
                                  m_opr->stream());
             return true;