WIP

Author: solaslin

projects/hipblaslt/tensilelite/client/src/DataInitialization.cpp

@@ -2070,6 +2070,12 @@ namespace TensileLite
                     auto swizzleKey
                         = std::make_tuple(toBitWidth(desc.dataType()), unrolledSize, tiledSize);
 
+                    // ::Tensor::Manipulation::TensorDesc paddedShapDesc(paddedShape);
+                    // std::cout << "original un-padded elems:" << desc.totalAllocatedElements() << std::endl;
+                    // std::cout << "auto padded elems for swizzle:" << paddedShapDesc.flattenSize() << std::endl;
+                    // std::cout << "default copy elems:" << p.maxElements << std::endl;
+
+                    // Cache-hit
                     if(g_swizzleCache.count(swizzleKey))
                     {
                         if(swizzleKey != g_swizzleCache.back())
@@ -2086,8 +2092,11 @@ namespace TensileLite
                             ptr = p.gpuInput.valid.get();
                         }
                     }
+                    // No Cache-hit, do pre-shuffle...
                     else
                     {
+                        // TODO- no need to do any of memory operation (cpy, shuffle, etc...)
+                        // but we still need MAKE SURE WE ALLOC/PASS THE INPUT SIZES (auto-padded = paddedShape)
                         auto tmpTensor = Tensor({tiledSize, unrolledSize}, desc.elementBytes());
 
                         memcpy(