[Util] allow packing dim reversal in npy2apintstream and v.v.

src/finn/data/cpp/npy2apintstream.hpp

@@ -13,7 +13,7 @@
 #endif
 
 template <typename PackedT, typename ElemT, int ElemBits, typename NpyT>
-void npy2apintstream(const char * npy_path, hls::stream<PackedT> & out_stream) {
+void npy2apintstream(const char * npy_path, hls::stream<PackedT> & out_stream, bool reverse_inner = true) {
   cnpy::NpyArray arr = cnpy::npy_load(npy_path);
   DEBUG_NPY2APINTSTREAM("word_size " << arr.word_size << " num_vals " << arr.num_vals)
   if(arr.word_size != sizeof(NpyT)) {
@@ -28,7 +28,8 @@ void npy2apintstream(const char * npy_path, hls::stream<PackedT> & out_stream) {
   DEBUG_NPY2APINTSTREAM("n_outer " << outer_dim_elems << " n_inner " << inner_dim_elems)
   for(size_t outer_elem = 0; outer_elem < outer_dim_elems; outer_elem++) {
     PackedT packed_elem = 0;
-    for(size_t i = 0; i < inner_dim_elems; i++) {
+    for(size_t ii = 0; ii < inner_dim_elems; ii++) {
+      size_t i = reverse_inner ? inner_dim_elems-ii-1 : ii;
       NpyT loaded_elem_npyt = *loaded_data;
       ElemT loaded_elem = (ElemT) loaded_elem_npyt;
       DEBUG_NPY2APINTSTREAM("NpyT " << loaded_elem_npyt << " elem " << loaded_elem)
@@ -41,7 +42,7 @@ void npy2apintstream(const char * npy_path, hls::stream<PackedT> & out_stream) {
 }
 
 template <typename PackedT, typename ElemT, int ElemBits, typename NpyT>
-void apintstream2npy(hls::stream<PackedT> & in_stream, const std::vector<size_t> & shape, const char * npy_path) {
+void apintstream2npy(hls::stream<PackedT> & in_stream, const std::vector<size_t> & shape, const char * npy_path, bool reverse_inner = true) {
   std::vector<NpyT> data_to_save;
   size_t outer_dim_elems = 1;
   for(size_t dim = 0; dim < shape.size()-1; dim++) {
@@ -53,7 +54,8 @@ void apintstream2npy(hls::stream<PackedT> & in_stream, const std::vector<size_t>
     PackedT packed_elem;
     in_stream >> packed_elem;
     DEBUG_APINTSTREAM2NPY("packed hls elem " << std::hex << packed_elem << std::dec)
-    for(size_t i = 0; i < inner_dim_elems; i++) {
+    for(size_t ii = 0; ii < inner_dim_elems; ii++) {
+      size_t i = reverse_inner ? inner_dim_elems-ii-1 : ii;
       ElemT elem = packed_elem((i+1)*ElemBits-1, i*ElemBits);
       NpyT npyt = (NpyT) elem;
       DEBUG_APINTSTREAM2NPY("elem " << elem << " NpyT " << npyt)