adding implicit gemm v3

driver/device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp

@@ -38,7 +38,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
     constexpr index_t X = wei_kcyx_desc.GetLength(I3);
 
     // reorder weight
-    auto wei_cyxk_desc = make_packed_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
+    auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
     ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
 
     Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
@@ -51,7 +51,7 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
         std::thread::hardware_concurrency());
 
     // reorder input
-    auto in_chwn_desc = make_packed_ConstantTensorDescriptor(Sequence<C, Hi, Wi, N>{});
+    auto in_chwn_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Hi, Wi, N>{});
     ostream_ConstantTensorDescriptor(in_chwn_desc, std::cout << "in_chwn_desc: ");
 
     Tensor<T> in_chwn(make_TensorDescriptor(in_chwn_desc));
@@ -64,7 +64,8 @@ void device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(InDesc,
         std::thread::hardware_concurrency());
 
     // output
-    auto out_khwn_desc = make_packed_ConstantTensorDescriptor(Sequence<K, Ho, Wo, N>{});
+    auto out_khwn_desc =
+        make_ConstantTensorDescriptor_default_rank_packed(Sequence<K, Ho, Wo, N>{});
     ostream_ConstantTensorDescriptor(out_khwn_desc, std::cout << "out_khwn_desc: ");
 
     Tensor<T> out_khwn(make_TensorDescriptor(out_khwn_desc));

driver/device_convolution_implicit_gemm_v1_nchw_cyxk_khwn.hpp

@@ -37,7 +37,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
     constexpr index_t X = wei_kcyx_desc.GetLength(I3);
 
     // reorder weight
-    auto wei_cyxk_desc = make_packed_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
+    auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
     ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
 
     Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
@@ -50,7 +50,8 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_khwn(InDesc,
         std::thread::hardware_concurrency());
 
     // output
-    auto out_khwn_desc = make_packed_ConstantTensorDescriptor(Sequence<K, Ho, Wo, N>{});
+    auto out_khwn_desc =
+        make_ConstantTensorDescriptor_default_rank_packed(Sequence<K, Ho, Wo, N>{});
     ostream_ConstantTensorDescriptor(out_khwn_desc, std::cout << "out_khwn_desc: ");
 
     Tensor<T> out_khwn(make_TensorDescriptor(out_khwn_desc));

driver/device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw.hpp

@@ -36,7 +36,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
     constexpr index_t X = wei_kcyx_desc.GetLength(I3);
 
     // reorder weight
-    auto wei_cyxk_desc = make_packed_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
+    auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
     ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
 
     Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
@@ -57,7 +57,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
     wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
     out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
 
-#if 0
+#if 1
     // for 3x3, 34x34, v1r3, Pascal
     constexpr index_t BlockSize = 128;
 
@@ -92,7 +92,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
     constexpr index_t WeiBlockCopyDataPerRead_K = 4;
 
     constexpr index_t OutThreadCopyDataPerWrite_W = 2;
-#elif 0
+#elif 1
     // for 3x3, 34x34, v1r3, Vega 20, WoPerBlock = 32
     constexpr index_t BlockSize = 256;
 
@@ -162,7 +162,7 @@ void device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(InDesc,
     constexpr index_t WeiBlockCopyDataPerRead_K = 4;
 
     constexpr index_t OutThreadCopyDataPerWrite_W = 2;
-#elif 1
+#elif 0
     // for 3x3, 34x34, v1r3, Vega 20, WoPerBlock = 8
     constexpr index_t BlockSize = 256;
 

driver/device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp

@@ -35,7 +35,7 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
     constexpr index_t X = wei_kcyx_desc.GetLength(I3);
 
     // reorder weight
-    auto wei_cyxk_desc = make_ConstantTensorDescriptor(Sequence<C, Y, X, K>{});
+    auto wei_cyxk_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<C, Y, X, K>{});
     ostream_ConstantTensorDescriptor(wei_cyxk_desc, std::cout << "wei_cyxk_desc: ");
 
     Tensor<T> wei_cyxk(make_TensorDescriptor(wei_cyxk_desc));
@@ -56,37 +56,40 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
     wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
     out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
 
+    constexpr index_t N1 = 2;
+    constexpr index_t N2 = 4;
+
+    constexpr index_t B = (N * Ho * Wo) / (N1 * N2);
+
 #if 1
-    // for 3x3, 28x28, v3, Pascal
-    constexpr index_t BlockSize = 128;
+    // for 3x3, 28x28, v3
+    constexpr index_t BlockSize = 256;
 
     constexpr index_t BPerBlock = 16;
     constexpr index_t KPerBlock = 128;
     constexpr index_t CPerBlock = 8;
 
-    constexpr index_t BPerThread = 1;
-    constexpr index_t KPerThread = 8;
-
     constexpr index_t GemmMPerThreadSubC = 4;
     constexpr index_t GemmNPerThreadSubC = 4;
     constexpr index_t GemmMLevel0Cluster = 4;
-    constexpr index_t GemmNLevel0Cluster = 2;
+    constexpr index_t GemmNLevel0Cluster = 4;
     constexpr index_t GemmMLevel1Cluster = 4;
-    constexpr index_t GemmNLevel1Cluster = 2;
+    constexpr index_t GemmNLevel1Cluster = 4;
     constexpr index_t GemmKPerThreadLoop = 1;
     constexpr index_t GemmDataPerReadA   = 4;
     constexpr index_t GemmDataPerReadB   = 4;
 
-    using InBlockReorderSrcSubLengths_NCHW                    = Sequence<4, 1, 1, 1>;
-    using InBlockReorderSrcClusterLengths_NCHW                = Sequence<4, 8, 2, 2>;
-    using InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW = Sequence<1, 2, 0, 3>;
+    using InBlockCopySubLengths_N1_N2_C_B     = Sequence<1, 4, 1, 1>;
+    using InBlockCopyClusterLengths_N1_N2_C_B = Sequence<2, 1, 8, 16>;
+
+    constexpr index_t InBlockCopySrcDataPerRead_B   = 1;
+    constexpr index_t InBlockCopyDstDataPerWrite_N2 = 4;
 
-    constexpr index_t WeiBlockCopyDataPerRead_K = 4;
+    constexpr index_t WeiBlockCopyDataPerAccess_K = 4;
 #endif
 
     constexpr index_t GridSize =
-        ((N + NPerBlock - 1) / NPerBlock) * ((K + KPerBlock - 1) / KPerBlock) *
-        ((Ho + HoPerBlock - 1) / HoPerBlock) * ((Wo + WoPerBlock - 1) / WoPerBlock);
+        ((B + BPerBlock - 1) / BPerBlock) * ((K + KPerBlock - 1) / KPerBlock);
 
     printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
 
@@ -102,15 +105,11 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
              decltype(in_nchw_desc),
              decltype(wei_cyxk_desc),
              decltype(out_nkhw_desc),
-             NPerBlock,
+             BPerBlock,
              KPerBlock,
              CPerBlock,
-             HoPerBlock,
-             WoPerBlock,
-             NPerThread,
-             KPerThread,
-             HoPerThread,
-             WoPerThread,
+             N1,
+             N2,
              GemmMPerThreadSubC,
              GemmNPerThreadSubC,
              GemmMLevel0Cluster,
@@ -120,14 +119,11 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
              GemmKPerThreadLoop,
              GemmDataPerReadA,
              GemmDataPerReadB,
-             InBlockReorderSrcSubLengths_NCHW,
-             InBlockReorderSrcClusterLengths_NCHW,
-             InBlockReorderMapThreadCluster2SrcCluster_CHNW2NCHW,
-             InBlockReorderDataPerRead_W,
-             InBlockReorderDataPerWrite_N,
-             WeiBlockCopyClusterLengths,
-             WeiBlockCopyDataPerRead_K,
-             OutThreadCopyDataPerWrite_W>{};
+             InBlockCopySubLengths_N1_N2_C_B,
+             InBlockCopyClusterLengths_N1_N2_C_B,
+             InBlockCopySrcDataPerRead_B,
+             InBlockCopyDstDataPerWrite_N2,
+             WeiBlockCopyDataPerAccess_K>{};
 
         float time = launch_kernel(run_gridwise_convolution<decltype(gridwise_conv), T>,
                                    dim3(GridSize),

driver/driver.hip.cpp

@@ -13,7 +13,7 @@
 #include "device_convolution_implicit_gemm_v1_nchw_cyxk_khwn.hpp"
 #include "device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw.hpp"
 #include "device_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hpp"
-//#include "device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
+#include "device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
 
 struct GeneratorTensor_1
 {
@@ -548,8 +548,8 @@ int main(int argc, char* argv[])
     auto lower_pads = Sequence<HPad, WPad>{};
     auto upper_pads = Sequence<HPad, WPad>{};
 
-    auto in_nchw_desc  = make_packed_ConstantTensorDescriptor(Sequence<N, C, HI, WI>{});
-    auto wei_kcyx_desc = make_packed_ConstantTensorDescriptor(Sequence<K, C, Y, X>{});
+    auto in_nchw_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<N, C, HI, WI>{});
+    auto wei_kcyx_desc = make_ConstantTensorDescriptor_default_rank_packed(Sequence<K, C, Y, X>{});
     auto out_nkhw_desc = get_convolution_with_padding_output_default_4d_tensor_descriptor(
         in_nchw_desc, wei_kcyx_desc, lower_pads, upper_pads);
 
@@ -612,11 +612,11 @@ int main(int argc, char* argv[])
     device_convolution_implicit_gemm_v1_chwn_cyxk_khwn
 #elif 0
     device_convolution_implicit_gemm_v1_nchw_cyxk_khwn
-#elif 1
+#elif 0
     device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw
 #elif 0
     device_convolution_implicit_gemm_v2_chwn_cyxk_khwn
-#elif 0
+#elif 1
     device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw
 #endif
     (in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);

src/include/Array.hip.hpp

@@ -12,7 +12,7 @@ struct Array
     index_t mData[nSize];
 
     template <class... Xs>
-    __host__ __device__ Array(Xs... xs) : mData{static_cast<TData>(xs)...}
+    __host__ __device__ constexpr Array(Xs... xs) : mData{static_cast<TData>(xs)...}
     {
     }
 
@@ -37,6 +37,25 @@ struct Array
     }
 };
 
+template <index_t... Is>
+__host__ __device__ constexpr auto sequence2array(Sequence<Is...>)
+{
+    return Array<index_t, sizeof...(Is)>{Is...};
+}
+
+template <class TData, index_t NSize>
+__host__ __device__ constexpr auto make_zero_array()
+{
+    Array<TData, NSize> a;
+
+    static_for<0, NSize, 1>{}([&](auto I) {
+        constexpr index_t i = I.Get();
+        a[i]                = static_cast<TData>(0);
+    });
+
+    return a;
+}
+
 template <class TData, index_t NSize, index_t... IRs>
 __host__ __device__ auto reorder_array_given_new2old(const Array<TData, NSize>& old_array,
                                                      Sequence<IRs...> new2old)
@@ -80,15 +99,14 @@ __host__ __device__ auto extract_array(const Array<TData, NSize>& old_array, Ext
 
     static_for<0, new_size, 1>{}([&](auto I) {
         constexpr index_t i = I.Get();
-        new_array[i]        = old_array[ExtractSeq{}.Get(I)];
+        new_array[i]        = old_array[ExtractSeq::Get(I)];
     });
 
     return new_array;
 }
 
 template <class TData, index_t NSize>
-__host__ __device__ constexpr auto operator+(const Array<TData, NSize>& a,
-                                             const Array<TData, NSize>& b)
+__host__ __device__ constexpr auto operator+(Array<TData, NSize> a, Array<TData, NSize> b)
 {
     Array<TData, NSize> result;
 
@@ -99,3 +117,20 @@ __host__ __device__ constexpr auto operator+(const Array<TData, NSize>& a,
 
     return result;
 }
+
+// Array = Array * Sequence
+template <class TData, index_t NSize, index_t... Is>
+__host__ __device__ constexpr auto operator*(Array<TData, NSize> a, Sequence<Is...> b)
+{
+    static_assert(sizeof...(Is) == NSize, "wrong! size not the same");
+
+    Array<TData, NSize> result;
+
+    static_for<0, NSize, 1>{}([&](auto I) {
+        constexpr index_t i = I.Get();
+
+        result[i] = a[i] + b.Get(I);
+    });
+
+    return result;
+}

src/include/ConstantMatrixDescriptor.hip.hpp

@@ -9,26 +9,26 @@ struct ConstantMatrixDescriptor
         static_assert(NCol_ <= RowStride_, "wrong! NCol > RowStride!");
     }
 
-    __host__ __device__ constexpr index_t NRow() const { return NRow_; }
+    __host__ __device__ static constexpr index_t NRow() { return NRow_; }
 
-    __host__ __device__ constexpr index_t NCol() const { return NCol_; }
+    __host__ __device__ static constexpr index_t NCol() { return NCol_; }
 
-    __host__ __device__ constexpr index_t RowStride() const { return RowStride_; }
+    __host__ __device__ static constexpr index_t RowStride() { return RowStride_; }
 
-    __host__ __device__ constexpr auto GetLengths() const { return Sequence<NRow_, NCol_>{}; }
+    __host__ __device__ static constexpr auto GetLengths() { return Sequence<NRow_, NCol_>{}; }
 
-    __host__ __device__ constexpr index_t GetElementSize() const { return NRow_ * NCol_; }
+    __host__ __device__ static constexpr index_t GetElementSize() { return NRow_ * NCol_; }
 
-    __host__ __device__ constexpr index_t GetElementSpace() const { return NRow_ * RowStride_; }
+    __host__ __device__ static constexpr index_t GetElementSpace() { return NRow_ * RowStride_; }
 
-    __host__ __device__ index_t GetOffsetFromMultiIndex(index_t irow, index_t icol) const
+    __host__ __device__ static index_t GetOffsetFromMultiIndex(index_t irow, index_t icol)
     {
         return irow * RowStride_ + icol;
     }
 
     template <index_t SubNRow, index_t SubNCol>
-    __host__ __device__ constexpr auto MakeSubMatrixDescriptor(Number<SubNRow>,
-                                                               Number<SubNCol>) const
+    __host__ __device__ static constexpr auto MakeSubMatrixDescriptor(Number<SubNRow>,
+                                                                      Number<SubNCol>)
     {
         return ConstantMatrixDescriptor<SubNRow, SubNCol, RowStride_>{};
     }

src/include/ConstantMergedTensorDescriptor.hip.hpp

@@ -11,8 +11,8 @@ struct ConstantMergedTensorDescriptor
 {
     static constexpr auto mOriginalDimMergeSeqs = std::tuple<OriginalDimMergeSeqs...>{};
 
-    static constexpr index_t nDim         = std::tuple_size<mOriginalDimMergeSeqs>::value;
-    static constexpr index_t nOriginalDim = OriginalDesc::GetNumOfDimension();
+    static constexpr index_t nDim         = sizeof...(OriginalDimMergeSeqs);
+    static constexpr index_t nOriginalDim = OriginalTensorDesc::GetNumOfDimension();
 
     __host__ __device__ constexpr ConstantMergedTensorDescriptor()
     {
@@ -21,25 +21,28 @@ struct ConstantMergedTensorDescriptor
         // TODO: check each of OriginalDimMergeSeqs contains at least 1, and at most
         // OriginalTensorDesc::nDim number of dimensions
 
-        // TODO: check there is no duplication in OriginalDimMergeSeqs
-
         // TODO: check OriginalDimMergeSeqs contains all original dimensions
+
+        // TODO: check there is no duplication in OriginalDimMergeSeqs
     }
 
     __host__ __device__ static constexpr index_t GetNumOfDimension() { return nDim; }
 
-    __host__ __device__ static constexpr index_t GetNumOfOriginalDimension() { return nOriginalDim }
+    __host__ __device__ static constexpr index_t GetNumOfOriginalDimension()
+    {
+        return nOriginalDim;
+    }
 
     template <index_t IDim>
     __host__ __device__ static constexpr bool ContainMultipleOriginalDimensions(Number<IDim>)
     {
-        return (std::Get<IDIM>(mOriginalDimMergeSeqs).GetSize() > 1);
+        return (std::get<IDim>(mOriginalDimMergeSeqs).GetSize() > 1);
     }
 
     template <index_t IDim>
     __host__ __device__ static constexpr index_t GetLength(Number<IDim>)
     {
-        constexpr auto original_dims_partial = std::Get<IDim>(mOriginalDimMergeSeqs);
+        constexpr auto original_dims_partial = std::get<IDim>(mOriginalDimMergeSeqs);
 
         return OriginalTensorDesc::Extract(original_dims_partial).GetElementSize();
     }
@@ -50,14 +53,14 @@ struct ConstantMergedTensorDescriptor
         static_assert(!ContainMultipleOriginalDimensions(Number<IDim>{}),
                       "wrong! stride of a merged dimension is undefined");
 
-        constexpr auto idim_original = std::Get<IDim>(mOriginalDimMergeSeqs).Front();
+        constexpr auto idim_original = std::get<IDim>(mOriginalDimMergeSeqs).Front();
 
         return OriginalTensorDesc::GetStride(Number<idim_original>{});
     }
 
     __host__ __device__ static constexpr auto GetLengths()
     {
-        return Sequence<OriginalTensorDesc::Extract(OriginalDimMergeSeqs).GetElementSize()...>{};
+        return Sequence<OriginalTensorDesc::Extract(OriginalDimMergeSeqs{}).GetElementSize()...>{};
     }
 
     __host__ __device__ static constexpr index_t GetElementSize()
@@ -75,17 +78,16 @@ struct ConstantMergedTensorDescriptor
             constexpr auto original_dims_partial = std::get<idim>(mOriginalDimMergeSeqs);
 
             // get partial original-multi-id corresponding to this merged dimension
-            constexpr auto original_multi_id_partial =
+            const auto original_multi_id_partial =
                 OriginalTensorDesc::Extract(original_dims_partial)
                     .GetMultiIndexFrom1dIndex(multi_id[idim]);
 
-            // make sure compiler unroll this loop and propagate all the constants
-            for(index_t i = 0; i < original_dims_partial.GetSize(); ++i)
-            {
-                index_t idim_original = original_dims_partial[i];
+            static_for<0, original_dims_partial.GetSize(), 1>{}([&](auto I_) {
+                constexpr auto I                = decltype(I_){};
+                constexpr index_t idim_original = original_dims_partial.Get(I);
 
-                original_multi_id[idim_original] = original_multi_id_partial[i]
-            }
+                original_multi_id[idim_original] = original_multi_id_partial[I.Get()];
+            });
         });
 
         return original_multi_id;
@@ -95,10 +97,10 @@ struct ConstantMergedTensorDescriptor
     {
         const auto original_multi_id = GetOriginalMultiIndexFromMultiIndex(multi_id);
 
-        return OriginalTensorDesc::GetOffsetFromMultiIndex(orginal_multi_id);
+        return OriginalTensorDesc::GetOffsetFromMultiIndex(original_multi_id);
     }
 
-    template <index_t... Is>
+    template <class... Is>
     __host__ __device__ static index_t GetOffsetFromMultiIndex(Is... is)
     {
         return GetOffsetFromMultiIndex(Array<index_t, nDim>{is...});
@@ -106,14 +108,15 @@ struct ConstantMergedTensorDescriptor
 
     __host__ __device__ static Array<index_t, nDim> GetMultiIndexFrom1dIndex(index_t id)
     {
-        constexpr auto dummy_desc = make_packed_ConstantTensorDescriptor(GetLengths());
+        constexpr auto dummy_desc = make_ConstantTensorDescriptor_default_rank_packed(GetLengths());
 
         return dummy_desc.GetMultiIndexFrom1dIndex(id);
     }
 };
 
 template <class OriginalTensorDesc, class... OriginalDimMergeSeqs>
-constexpr auto make_ConstantMergedTensorDescriptor(OriginalTensorDesc, OriginalDimMergeSeqs...)
+__host__ __device__ constexpr auto make_ConstantMergedTensorDescriptor(OriginalTensorDesc,
+                                                                       OriginalDimMergeSeqs...)
 {
     return ConstantMergedTensorDescriptor<OriginalTensorDesc, OriginalDimMergeSeqs...>{};
 }

src/include/ConstantTensorDescriptor.hip.hpp

@@ -2,20 +2,20 @@
 #include "common.hip.hpp"
 
 template <class Lengths>
-__host__ __device__ constexpr auto calculate_packed_tensor_strides(Lengths)
+__host__ __device__ constexpr auto calculate_tensor_strides_default_rank_packed(Lengths)
 {
     return reverse_inclusive_scan_sequence(Lengths{}.PopFront(), std::multiplies<index_t>{})
         .PushBack(Number<1>{});
 }
 
 template <class Lengths, index_t Align>
-__host__ __device__ constexpr auto
-    calculate_rank_tensor_default_strides_with_alignment(Lengths, Number<Align>)
+__host__ __device__ constexpr auto calculate_tensor_strides_default_rank_aligned(Lengths,
+                                                                                 Number<Align>)
 {
     constexpr index_t L_back_align =
         Align * mod_conv::integer_divide_ceiler<index_t>{}(Lengths{}.Back(), Align);
 
-    return calculate_packed_tensor_strides(
+    return calculate_tensor_strides_default_rank_packed(
         Lengths{}.Modify(Number<Lengths{}.GetSize() - 1>{}, Number<L_back_align>{}));
 }
 
@@ -66,6 +66,12 @@ struct ConstantTensorDescriptor
         return MemoryRanks{}.Get(Number<I>{});
     }
 
+    template <class T>
+    __host__ __device__ static constexpr bool ContainMultipleOriginalDimensions(T)
+    {
+        return false;
+    }
+
     __host__ __device__ static constexpr index_t GetElementSize()
     {
         return accumulate_on_sequence(Lengths{}, std::multiplies<index_t>{}, Number<1>{});
@@ -146,7 +152,7 @@ struct ConstantTensorDescriptor
     {
         Array<index_t, nDim> multi_id;
 
-        constexpr auto dummy_strides = calculate_packed_tensor_strides(GetLengths());
+        constexpr auto dummy_strides = calculate_tensor_strides_default_rank_packed(GetLengths());
 
         // calculate index in each of the dimensions in the order of their dimension (not rank)
         static_for<0, nDim - 1, 1>{}([&](auto IDim) {
@@ -181,6 +187,12 @@ struct ConstantTensorDescriptor
         return ConstantTensorDescriptor<extract_lengths, extract_strides, new_ranks>{};
     }
 
+    template <index_t... IDims>
+    __host__ __device__ static constexpr auto Extract(Sequence<IDims...>)
+    {
+        return Extract(Number<IDims>{}...);
+    }
+
     template <index_t IDim, index_t SliceLen>
     __host__ __device__ static constexpr auto Slice(Number<IDim>, Number<SliceLen>)
     {
@@ -271,9 +283,11 @@ struct ConstantTensorDescriptor
                           FirstUnfoldDim <= LastUnfoldDim,
                       "wrong! should have FirstUnfoldDim <= LastUnfoldDim!");
 
+#if 0 // cannot compile: compiler complain about constexpr
         // dimensions to be unfold need to be in descending order (w.r.t. strides), and need to be
         // packed in memory, otherwise, unfolding is invalid
-        static_for<FirstUnfoldDim, LastUnfoldDim, 1>{}([&](auto IDim) {
+        static_for<FirstUnfoldDim, LastUnfoldDim, 1>{}([&](auto IDim_) {
+            constexpr auto IDim    = decltype(IDim_){};
             constexpr auto IDim_p1 = IDim + Number<1>{};
 
             // check stride
@@ -285,11 +299,12 @@ struct ConstantTensorDescriptor
             static_assert(GetStride(IDim_p1) * GetLength(IDim_p1) == GetStride(IDim),
                           "wrong! dimensions to be unfolded need to be packed");
 
-            // checkt ranks
+            // check ranks
             static_assert(GetMemoryRank(IDim_p1) == GetMemoryRank(IDim) + 1,
                           "wrong! ranks of dimensions to be unfolded need to be in increasing and "
                           "continuous ranks");
         });
+#endif
 
         // left and right
         constexpr auto left = typename arithmetic_sequence_gen<0, FirstUnfoldDim, 1>::SeqType{};
@@ -308,9 +323,9 @@ struct ConstantTensorDescriptor
 
         // decrease the ranks that are larger than the rank of LastUnfoldDim
         constexpr auto tmp_ranks =
-            transform_sequences(GetMemoryRanks(),
-                                f_unfold_impl<GetMemoryRank(Number<LastUnfoldDim>{}),
-                                              LastUnfoldDim - FirstUnfoldDim + 1>{});
+            transform_sequences(f_unfold_impl<GetMemoryRank(Number<LastUnfoldDim>{}),
+                                              LastUnfoldDim - FirstUnfoldDim + 1>{},
+                                GetMemoryRanks());
 
         // new lengths, strides and ranks
         constexpr auto new_lengths = GetLengths()
@@ -354,26 +369,26 @@ struct ConstantTensorDescriptor
 };
 
 template <class Lengths>
-__host__ __device__ constexpr auto make_packed_ConstantTensorDescriptor(Lengths)
+__host__ __device__ constexpr auto make_ConstantTensorDescriptor_default_rank_packed(Lengths)
 {
-    using Strides     = decltype(calculate_packed_tensor_strides(Lengths{}));
+    using Strides     = decltype(calculate_tensor_strides_default_rank_packed(Lengths{}));
     using MemoryRanks = typename arithmetic_sequence_gen<0, Lengths::GetSize(), 1>::SeqType;
     return ConstantTensorDescriptor<Lengths, Strides, MemoryRanks>{};
 }
 
 template <class Lengths, class Strides>
-__host__ __device__ constexpr auto make_ranked_ConstantTensorDescriptor(Lengths, Strides)
+__host__ __device__ constexpr auto make_ConstantTensorDescriptor_default_rank(Lengths, Strides)
 {
     using MemoryRanks = typename arithmetic_sequence_gen<0, Lengths::GetSize(), 1>::SeqType;
     return ConstantTensorDescriptor<Lengths, Strides, MemoryRanks>{};
 }
 
 template <class Lengths, index_t Align>
-__host__ __device__ constexpr auto
-    make_ranked_ConstantTensorDescriptor_with_alignment(Lengths, Number<Align>)
+__host__ __device__ constexpr auto make_ConstantTensorDescriptor_default_rank_aligned(Lengths,
+                                                                                      Number<Align>)
 {
     using Strides =
-        decltype(calculate_rank_tensor_default_strides_with_alignment(Lengths{}, Number<Align>{}));
+        decltype(calculate_tensor_strides_default_rank_aligned(Lengths{}, Number<Align>{}));
     using MemoryRanks = typename arithmetic_sequence_gen<0, Lengths::GetSize(), 1>::SeqType;
     return ConstantTensorDescriptor<Lengths, Strides, MemoryRanks>{};
 }

src/include/Sequence.hip.hpp

 #pragma once
-#include "constant_integral.hip.hpp"
+#include "integral_constant.hip.hpp"
 #include "functional.hip.hpp"
 
 template <index_t... Is>
@@ -21,12 +21,6 @@ struct Sequence
         return mData[I];
     }
 
-    __host__ __device__ index_t operator[](index_t i) const
-    {
-        const index_t mData[mSize + 1] = {Is..., 0};
-        return mData[i];
-    }
-
     template <index_t... IRs>
     __host__ __device__ static constexpr auto ReorderGivenNew2Old(Sequence<IRs...> /*new2old*/)
     {
@@ -164,6 +158,12 @@ struct sequence_reverse_inclusive_scan<Sequence<I>, Reduce>
     using SeqType = Sequence<I>;
 };
 
+template <class Reduce>
+struct sequence_reverse_inclusive_scan<Sequence<>, Reduce>
+{
+    using SeqType = Sequence<>;
+};
+
 template <class, class>
 struct sequence_extract;
 

src/include/blockwise_2d_tensor_op.hip.hpp

@@ -457,7 +457,8 @@ struct Blockwise2dTensorCopy3
     index_t mSrcMyThreadOffset;
     index_t mDstMyThreadOffset;
 
-    __device__ Blockwise2dTensorCopy3()
+    __device__ Blockwise2dTensorCopy3(Array<index_t, 2> src_block_data_multi_id_begin,
+                                      Array<index_t, 2> dst_block_data_multi_id_begin)
     {
         constexpr auto I0 = Number<0>{};
         constexpr auto I1 = Number<1>{};
@@ -499,10 +500,13 @@ struct Blockwise2dTensorCopy3
         const index_t thread_id_d0 = get_thread_local_1d_id() / thread_per_d1;
         const index_t thread_id_d1 = get_thread_local_1d_id() - thread_id_d0 * thread_per_d1;
 
-        mSrcMyThreadOffset =
-            SrcDesc{}.GetOffsetFromMultiIndex(thread_id_d0, thread_id_d1 * DataPerRead);
-        mDstMyThreadOffset =
-            DstDesc{}.GetOffsetFromMultiIndex(thread_id_d0, thread_id_d1 * DataPerRead);
+        mSrcMyThreadOffset = SrcDesc{}.GetOffsetFromMultiIndex(
+            src_block_data_multi_id_begin +
+            Array<index_t, 2>{thread_id_d0, thread_id_d1 * DataPerRead});
+
+        mDstMyThreadOffset = DstDesc{}.GetOffsetFromMultiIndex(
+            dst_block_data_multi_id_begin +
+            Array<index_t, 2>{thread_id_d0, thread_id_d1 * DataPerRead});
     }
 
     __device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const

src/include/blockwise_4d_tensor_op.hip.hpp

@@ -13,7 +13,7 @@ blockwise_4d_tensor_pointwise_operation_unary(DstDesc, Float* __restrict__ p_dst
 
     constexpr auto dst_desc = DstDesc{};
 
-    constexpr auto desc = make_packed_ConstantTensorDescriptor(dst_desc.GetLengths());
+    constexpr auto desc = make_ConstantTensorDescriptor_default_rank_packed(dst_desc.GetLengths());
 
 #if 0
     if(get_thread_local_1d_id() == 0)
@@ -108,7 +108,7 @@ __device__ void blockwise_4d_tensor_pointwise_operation_binary_reorder_by_get_ds
 
     constexpr auto src_desc = SrcDesc{};
     constexpr auto dst_desc = DstDesc{};
-    constexpr auto ref_desc = make_packed_ConstantTensorDescriptor(SrcOpLengths{});
+    constexpr auto ref_desc = make_ConstantTensorDescriptor_default_rank_packed(SrcOpLengths{});
 
     constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
 
@@ -259,7 +259,7 @@ struct Blockwise4dTensorCopy1
         constexpr index_t read_per_d3 = mod_conv::integer_divide_ceil(L3, DataPerRead);
 
         constexpr auto ref_desc =
-            make_packed_ConstantTensorDescriptor(Sequence<L0, L1, L2, read_per_d3>{});
+            make_ConstantTensorDescriptor_default_rank_packed(Sequence<L0, L1, L2, read_per_d3>{});
 
         constexpr index_t NLoop = ref_desc.GetElementSize() / BlockSize;
 
@@ -336,7 +336,7 @@ struct BlockwiseChwnTensorCopyPadded
 
         constexpr auto src_desc = SrcDesc{};
         constexpr auto dst_desc = DstDesc{};
-        constexpr auto ref_desc = make_packed_ConstantTensorDescriptor(DstOpLengths{});
+        constexpr auto ref_desc = make_ConstantTensorDescriptor_default_rank_packed(DstOpLengths{});
 
         constexpr auto h_global_pad_low = GlobalLowerPads{}.Get(I0);
         constexpr auto w_global_pad_low = GlobalLowerPads{}.Get(I1);
@@ -510,7 +510,8 @@ struct Blockwise4dTensorCopy3
             }
         }
 
-        constexpr auto thread_cluster_desc = make_packed_ConstantTensorDescriptor(ThreadPerDims{});
+        constexpr auto thread_cluster_desc =
+            make_ConstantTensorDescriptor_default_rank_packed(ThreadPerDims{});
         const auto thread_multi_id =
             thread_cluster_desc.GetMultiIndexFrom1dIndex(get_thread_local_1d_id());
 
@@ -652,7 +653,7 @@ struct Blockwise4dTensorCopy3
         constexpr index_t nloop_d2 = L2 / thread_per_d2;
         constexpr index_t nloop_d3 = mod_conv::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
 
-        constexpr auto clipboard_desc = make_packed_ConstantTensorDescriptor(
+        constexpr auto clipboard_desc = make_ConstantTensorDescriptor_default_rank_packed(
             Sequence<nloop_d0, nloop_d1, nloop_d2, nloop_d3 * DataPerRead>{});
 
 #pragma unroll
@@ -719,7 +720,7 @@ struct Blockwise4dTensorCopy3
         constexpr index_t nloop_d2 = L2 / thread_per_d2;
         constexpr index_t nloop_d3 = mod_conv::integer_divide_ceil(L3, thread_per_d3 * DataPerRead);
 
-        constexpr auto clipboard_desc = make_packed_ConstantTensorDescriptor(
+        constexpr auto clipboard_desc = make_ConstantTensorDescriptor_default_rank_packed(
             Sequence<nloop_d0, nloop_d1, nloop_d2, nloop_d3 * DataPerRead>{});
 
 #pragma unroll

src/include/blockwise_gemm.hip.hpp

@@ -46,7 +46,7 @@ struct BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2
                           N % (NPerThreadSubC * NLevel0Cluster * NLevel1Cluster) == 0,
                       "wrong! Cannot evenly divide work among\n");
 
-        static_assert(ThreadMatrixC::GetLengths() == GetThreadMatrixCLengths(),
+        static_assert(is_same_type(ThreadMatrixC::GetLengths(), GetThreadMatrixCLengths()),
                       "wrong! ThreadMatrixC lengths is wrong");
 
         auto c_thread_mtx_index = GetBeginOfThreadMatrixC(get_thread_local_1d_id());
@@ -55,7 +55,7 @@ struct BlockwiseGemmBlockABlockBThreadCTransANormalBNormalC_v2
         mMyThreadOffsetB = BlockMatrixB::GetOffsetFromMultiIndex(0, c_thread_mtx_index.col);
     }
 
-    __device__ static auto GetThreadMatrixCLengths()
+    __device__ static constexpr auto GetThreadMatrixCLengths()
     {
         constexpr index_t M = BlockMatrixA::NCol(); // A is transposed
         constexpr index_t N = BlockMatrixB::NCol();

src/include/blockwise_merged_tensor_slice_op.hip.hpp

 #pragma once
 #include "threadwise_tensor_slice_op.hip.hpp"
 
-// slice a merged tensor, reorder and copy it into a normal tensor
-// src: a merged tensor,
-// dst: a normal tensor
+// slice a (normal or merged) tensor, reorder and copy it into another (normal or merged) tensor
 template <index_t BlockSize,
           class Float,
           class SrcDesc,
           class DstDesc,
           class SliceLengths,
           class SubLengths,
-          class ClusterLengths,
+          class DataClusterLengths,
           class ThreadClusterArrangeOrder,
           class SrcAccessOrder,
-          class DstAccessOrder>
+          class DstAccessOrder,
+          index_t SrcDataPerRead,
+          index_t DstDataPerRead>
 struct BlockwiseTensorSliceCopy_generic_v1
 {
     static constexpr index_t nDim = SrcDesc::GetNumOfDimension();
@@ -21,39 +21,44 @@ struct BlockwiseTensorSliceCopy_generic_v1
     index_t mSrcMyThreadOffset;
     index_t mDstMyThreadOffset;
 
-    __device__ BlockwiseTensorSliceCopy_generic_v1(Array<index_t, nDim> src_block_multi_offset,
-                                                   Array<index_t, nDim> dst_block_multi_offset)
+    __device__
+    BlockwiseTensorSliceCopy_generic_v1(Array<index_t, nDim> src_block_data_multi_id_begin,
+                                        Array<index_t, nDim> dst_block_data_multi_id_begin)
     {
         // check NDim consistent
         static_assert(SrcDesc::GetNumOfDimension() == DstDesc::GetNumOfDimension(), "wrong");
 
-        constexpr auto thread_cluster_desc = make_packed_ConstantTensorDescriptor(
-            ClusterLengths{}.ReorderGivenNew2Old(ThreadClusterArrangeOrder{}));
+        // thread cluster
+        constexpr auto thread_cluster_desc = make_ConstantTensorDescriptor_default_rank_packed(
+            DataClusterLengths{}.ReorderGivenNew2Old(ThreadClusterArrangeOrder{}));
 
         // BlockSize
         static_assert(BlockSize == thread_cluster_desc.GetElementSize(), "wrong! BlockSize");
 
         // divide work
-        static_for<0, nDim, 1>{}([&](auto IDim) {
-            static_assert(SliceLengths{}.Get(IDim) % SubLenghs{}.Get(IDim) == 0,
+        constexpr auto data_per_cluster_per_dims = SubLengths{} * DataClusterLengths{};
+
+        static_for<0, nDim, 1>{}([&](auto IDim_) {
+            constexpr auto IDim = decltype(IDim_){};
+
+            static_assert(SliceLengths::Get(IDim) % SubLengths::Get(IDim) == 0,
                           "wrong! cannot evenly divide sliced tensor into sub-tensor");
+
+            static_assert(SliceLengths::Get(IDim) % data_per_cluster_per_dims.Get(IDim) == 0,
+                          "wrong! cannot evenly divide sliced tensor into cluster");
         });
 
-        constexpr auto thread_work_desc =
-            make_packed_ConstantTensorDescriptor(SliceLengths{} / SliceSubLengths{});
+        constexpr auto repeat_lengths = SliceLengths{} / data_per_cluster_per_dims;
 
-        static_for<0, nDim, 1>{}([&](auto IDim) {
-            static_assert(thread_work_desc.GetLength(IDim) % thread_cluster_desc.Get(IDim) == 0,
-                          "wrong! cannot evenly divide work to cluster");
-        });
+        // for now, only support SubLengths.Get() == 1 on a merged dimension that is merge from
+        // multiple dimensions
+        static_for<0, nDim, 1>{}([&](auto IDim_) {
+            constexpr auto IDim = decltype(IDim_){};
 
-        // only support SubLengths.Get() == 1 on merged dimension, for now
-        static_for<0, nDim, 1>{}([&](auto IDim) {
-            static_if<(SrcDesc::ContainMultipleOriginalDimensions(IDim) ||
-                       DstDesc::ContainMultipleOriginalDimensions(IDim))>{}([&](auto fwd) {
-                static_assert(fwd(SubLengths{}).Get(IDim) == 1,
-                              "wrong! Sub-Lengths on merged dimension should be 1");
-            });
+            static_assert(SubLengths::Get(IDim) == 1 ||
+                              (!SrcDesc::ContainMultipleOriginalDimensions(IDim) &&
+                               !DstDesc::ContainMultipleOriginalDimensions(IDim)),
+                          "wrong! only surpport Sub-Length == 1 on a merged dimension");
         });
 
         // calculate mSrcMyThreadOffset, mDstMyThreadOffset
@@ -63,22 +68,23 @@ struct BlockwiseTensorSliceCopy_generic_v1
         const auto data_cluster_multi_id =
             reorder_array_given_old2new(thread_cluster_multi_id, ThreadClusterArrangeOrder{});
 
-        const auto thread_data_multi_offset = data_cluster_multi_id * SubLengths{};
+        const auto thread_data_multi_id_begin = data_cluster_multi_id * SubLengths{};
+
+        mSrcMyThreadOffset = SrcDesc::GetOffsetFromMultiIndex(src_block_data_multi_id_begin +
+                                                              thread_data_multi_id_begin);
 
-        mSrcMythreadOffset =
-            SrcDesc::GetOffsetFromMultiIndex(src_block_multi_offset + thread_data_multi_offset);
-        mSrcMythreadOffset =
-            DstDesc::GetOffsetFromMultiIndex(dst_block_multi_offset + thread_data_multi_offset);
+        mSrcMyThreadOffset = DstDesc::GetOffsetFromMultiIndex(dst_block_data_multi_id_begin +
+                                                              thread_data_multi_id_begin);
     }
 
     __device__ static constexpr index_t GetRegisterClipboardSize()
     {
-        constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * ClusterLengths{});
+        constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * DataClusterLengths{});
 
         constexpr auto thread_tensor_desc =
-            make_packed_ConstantTensorDescriptor(SubLengths{} * repeat_lengths);
+            make_ConstantTensorDescriptor_default_rank_packed(SubLengths{} * repeat_lengths);
 
-        return thread_tensor_desc.GetElementSpaceSize();
+        return thread_tensor_desc.GetElementSpace();
     }
 
     __device__ void RunLoadRegisterClipboard(const Float* __restrict__ p_src,
@@ -86,32 +92,34 @@ struct BlockwiseTensorSliceCopy_generic_v1
     {
         constexpr auto thread_sub_tensor_lengths = SubLengths{};
 
-        constexpr auto data_per_cluster_per_dims = thread_sub_tensor_lengths * ClusterLengths{};
+        constexpr auto data_per_cluster_per_dims = thread_sub_tensor_lengths * DataClusterLengths{};
 
-        constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * ClusterLengths{});
+        constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * DataClusterLengths{});
 
-        constexpr auto thread_tensor_desc =
-            make_packed_ConstantTensorDescriptor(thread_sub_tensor_lengths * repeat_lengths);
+        constexpr auto thread_tensor_desc = make_ConstantTensorDescriptor_default_rank_packed(
+            thread_sub_tensor_lengths * repeat_lengths);
 
         static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
-            constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
+            constexpr auto repeat_multi_id = sequence2array(decltype(repeat_multi_id_){});
+
+            const auto src_thread_data_multi_id_begin =
+                repeat_multi_id * data_per_cluster_per_dims; // cannot not constexpr, why?
 
-            constexpr auto src_data_multi_offset = repeat_multi_id * data_per_cluster_per_dims;
+            const auto clipboard_data_multi_id_begin =
+                repeat_multi_id * thread_sub_tensor_lengths; // cannot not constexpr, why?
 
-            constexpr auto clipboard_data_multi_offset =
-                repeat_multi_id * thread_sub_tensor_lengths;
+            const index_t src_offset = SrcDesc{}.GetOffsetFromMultiIndex(
+                src_thread_data_multi_id_begin); // cannot not constexpr, why?
 
-            constexpr index_t src_offset = SrcDesc{}.GetOffsetFromMultiIndex(src_data_multi_id);
-            constexpr index_t clipboard_offset =
-                thread_tensor_desc.GetOffsetFromMultiIndex(clipboard_data_multi_id);
+            const index_t clipboard_offset = thread_tensor_desc.GetOffsetFromMultiIndex(
+                clipboard_data_multi_id_begin); // cannot not constexpr, why?
 
             threadwise_tensor_slice_copy_generic(SrcDesc{},
                                                  p_src + src_offset + mSrcMyThreadOffset,
-                                                 thread_tensor_desc,
-                                                 zero_array<index_t, nDim>{},
+                                                 make_zero_array<index_t, nDim>(),
                                                  thread_tensor_desc,
                                                  p_clipboard + clipboard_offset,
-                                                 zero_array<index_t, nDim>{},
+                                                 make_zero_array<index_t, nDim>(),
                                                  thread_sub_tensor_lengths,
                                                  SrcAccessOrder{});
         });
@@ -122,34 +130,37 @@ struct BlockwiseTensorSliceCopy_generic_v1
     {
         constexpr auto thread_sub_tensor_lengths = SubLengths{};
 
-        constexpr auto data_per_cluster_per_dims = thread_sub_tensor_lengths * ClusterLengths{};
+        constexpr auto data_per_cluster_per_dims = thread_sub_tensor_lengths * DataClusterLengths{};
 
-        constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * ClusterLengths{});
+        constexpr auto repeat_lengths = SliceLengths{} / (SubLengths{} * DataClusterLengths{});
 
-        constexpr auto thread_tensor_desc =
-            make_packed_ConstantTensorDescriptor(thread_sub_tensor_lengths * repeat_lengths);
+        constexpr auto thread_tensor_desc = make_ConstantTensorDescriptor_default_rank_packed(
+            thread_sub_tensor_lengths * repeat_lengths);
 
         static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
-            constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
+            constexpr auto repeat_multi_id = sequence2array(decltype(repeat_multi_id_){});
 
-            constexpr auto clipboard_data_multi_offset =
-                repeat_multi_id * thread_sub_tensor_lengths;
+            const auto clipboard_data_multi_id_begin =
+                repeat_multi_id * thread_sub_tensor_lengths; // cannot not constexpr, why?
 
-            constexpr auto dst_data_multi_offset = repeat_multi_id * data_per_cluster_per_dims;
+            const auto dst_data_multi_id_begin =
+                repeat_multi_id * data_per_cluster_per_dims; // cannot not constexpr, why?
 
-            constexpr index_t clipboard_offset =
-                thread_tensor_desc.GetOffsetFromMultiIndex(clipboard_data_multi_offset);
+            const index_t clipboard_offset = thread_tensor_desc.GetOffsetFromMultiIndex(
+                clipboard_data_multi_id_begin); // cannot not constexpr, why?
 
-            constexpr index_t dst_offset = DstDesc{}.GetOffsetFromMultiIndex(dst_data_multi_offset);
+            const index_t dst_offset = DstDesc{}.GetOffsetFromMultiIndex(
+                dst_data_multi_id_begin); // cannot not constexpr, why?
 
             threadwise_tensor_slice_copy_generic(thread_tensor_desc,
                                                  p_clipboard + clipboard_offset,
-                                                 zero_array<index_t, nDim>{},
+                                                 make_zero_array<index_t, nDim>(),
                                                  DstDesc{},
                                                  p_dst + dst_offset + mDstMyThreadOffset,
-                                                 zero_array<index_t, nDim>{},
+                                                 make_zero_array<index_t, nDim>(),
                                                  thread_sub_tensor_lengths,
                                                  DstAccessOrder{});
+        });
     }
 
     __device__ void Run(const Float* __restrict__ p_src, Float* __restrict__ p_dst) const
@@ -159,4 +170,4 @@ struct BlockwiseTensorSliceCopy_generic_v1
         RunLoadRegisterClipboard(p_src, p_clipboard);
         RunStoreRegisterClipboard(p_clipboard, p_dst);
     }
-    };
+};

src/include/blockwise_tensor_slice_op.hip.hpp

@@ -40,7 +40,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
             src_cluster_lengths.ReorderGivenNew2Old(map_thread_cluster_2_src_cluster);
 
         constexpr auto thread_cluster_desc =
-            make_packed_ConstantTensorDescriptor(thread_cluster_lengths);
+            make_ConstantTensorDescriptor_default_rank_packed(thread_cluster_lengths);
 
         // sanity check: data type
         static_assert(is_same<Float, float>::value, "wrong! only support float for now!\n");
@@ -149,7 +149,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
         constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
 
         constexpr auto thread_tensor_desc =
-            make_packed_ConstantTensorDescriptor(thread_tensor_lengths);
+            make_ConstantTensorDescriptor_default_rank_packed(thread_tensor_lengths);
 
         return thread_tensor_desc.GetElementSpace();
     }
@@ -170,7 +170,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
         constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
 
         constexpr auto thread_tensor_desc =
-            make_packed_ConstantTensorDescriptor(thread_tensor_lengths);
+            make_ConstantTensorDescriptor_default_rank_packed(thread_tensor_lengths);
 
         static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
             constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
@@ -208,7 +208,7 @@ struct BlockwiseTensorSliceReorderCopy_v3
         constexpr auto thread_tensor_lengths = thread_sub_tensor_lengths * repeat_lengths;
 
         constexpr auto thread_tensor_desc =
-            make_packed_ConstantTensorDescriptor(thread_tensor_lengths);
+            make_ConstantTensorDescriptor_default_rank_packed(thread_tensor_lengths);
 
         static_ford<decltype(repeat_lengths)>{}([&](auto repeat_multi_id_) {
             constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};

src/include/common.hip.hpp

 #pragma once
 #include "vector_type.hip.hpp"
-#include "constant_integral.hip.hpp"
+#include "integral_constant.hip.hpp"
 #include "Sequence.hip.hpp"
 #include "Array.hip.hpp"
 #include "functional.hip.hpp"
@@ -17,15 +17,21 @@ __device__ index_t get_block_1d_id() { return blockIdx.x; }
 template <class T1, class T2>
 struct is_same
 {
-    static const bool value = false;
+    static constexpr bool value = false;
 };
 
 template <class T>
 struct is_same<T, T>
 {
-    static const bool value = true;
+    static constexpr bool value = true;
 };
 
+template <class X, class Y>
+__host__ __device__ constexpr bool is_same_type(X, Y)
+{
+    return is_same<X, Y>::value;
+}
+
 namespace mod_conv { // namespace mod_conv
 template <class T, T s>
 struct scales

src/include/conv_common.hip.hpp

@@ -30,7 +30,7 @@ __host__ __device__ constexpr auto get_convolution_output_default_4d_tensor_desc
     constexpr auto HO = HI + 1 - Y;
     constexpr auto WO = WI + 1 - X;
 
-    return make_packed_ConstantTensorDescriptor(Sequence<N, K, HO, WO>{});
+    return make_ConstantTensorDescriptor_default_rank_packed(Sequence<N, K, HO, WO>{});
 }
 
 template <class InDesc, class WeiDesc, class LowerPads, class UpperPads>
@@ -67,7 +67,7 @@ __host__ __device__ constexpr auto get_convolution_with_padding_output_default_4
     constexpr auto HO = HI + HPadLow + HPadUp + 1 - Y;
     constexpr auto WO = WI + WPadLow + WPadUp + 1 - X;
 
-    return make_packed_ConstantTensorDescriptor(Sequence<N, K, HO, WO>{});
+    return make_ConstantTensorDescriptor_default_rank_packed(Sequence<N, K, HO, WO>{});
 }
 
 template <class InDesc, class WeiDesc, class OutDesc>

src/include/functional.hip.hpp

 #pragma once
-#include "constant_integral.hip.hpp"
+#include "integral_constant.hip.hpp"
 
 struct forwarder
 {

src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_chwn_cyxk_khwn.hip.hpp

@@ -85,7 +85,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
         constexpr index_t WBlockWork = mod_conv::integer_divide_ceil(Wo, WoPerBlock);
         constexpr index_t NBlockWork = mod_conv::integer_divide_ceil(N, NPerBlock);
 
-        constexpr auto block_work_desc = make_packed_ConstantTensorDescriptor(
+        constexpr auto block_work_desc = make_ConstantTensorDescriptor_default_rank_packed(
             Sequence<KBlockWork, HBlockWork, WBlockWork, NBlockWork>{});
 
         const auto block_work_multi_id =
@@ -109,7 +109,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
                                                     GemmDataPerReadA,
                                                     GemmDataPerReadB);
 
-        constexpr auto in_c_h_w_n_block_desc = make_ranked_ConstantTensorDescriptor_with_alignment(
+        constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
             Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
             Number<InBlockCopyDataPerRead_N>{});
 
@@ -118,12 +118,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
         static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
                       "GemmDataPerReadB alignment requirement is not meet");
 
-        constexpr auto wei_c_k_block_desc = make_ranked_ConstantTensorDescriptor_with_alignment(
+        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
             Sequence<CPerBlock, KPerBlock>{},
             Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
 
         // tensor view of threadwise output in register
-        constexpr auto out_k_h_w_n_thread_desc = make_packed_ConstantTensorDescriptor(
+        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
             Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
 
 // blockwise copy

src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_khwn.hip.hpp

@@ -86,7 +86,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
         constexpr index_t HBlockWork = mod_conv::integer_divide_ceil(Ho, HoPerBlock);
         constexpr index_t WBlockWork = mod_conv::integer_divide_ceil(Wo, WoPerBlock);
 
-        constexpr auto block_work_desc = make_packed_ConstantTensorDescriptor(
+        constexpr auto block_work_desc = make_ConstantTensorDescriptor_default_rank_packed(
             Sequence<NBlockWork, KBlockWork, HBlockWork, WBlockWork>{});
 
         const auto block_work_multi_id =
@@ -102,7 +102,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
 
         // global tensor view
         constexpr auto wei_c_k_global_desc =
-            make_ranked_ConstantTensorDescriptor(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
+            make_ConstantTensorDescriptor_default_rank(Sequence<C, K>{}, Sequence<Y * X * K, 1>{});
 
         // LDS tensor view
         //   be careful of alignment
@@ -111,7 +111,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
                                                     GemmDataPerReadA,
                                                     GemmDataPerReadB);
 
-        constexpr auto in_c_h_w_n_block_desc = make_ranked_ConstantTensorDescriptor_with_alignment(
+        constexpr auto in_c_h_w_n_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
             Sequence<CPerBlock, HoPerBlock, WoPerBlock, NPerBlock>{},
             Number<InBlockReorderDataPerWrite_N>{});
 
@@ -120,12 +120,12 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
         static_assert(in_c_h_w_n_block_desc.GetStride(I1) % GemmDataPerReadB == 0,
                       "GemmDataPerReadB alignment requirement is not meet");
 
-        constexpr auto wei_c_k_block_desc = make_ranked_ConstantTensorDescriptor_with_alignment(
+        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_default_rank_aligned(
             Sequence<CPerBlock, KPerBlock>{},
             Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
 
         // tensor view of threadwise output in register
-        constexpr auto out_k_h_w_n_thread_desc = make_packed_ConstantTensorDescriptor(
+        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
             Sequence<KPerThread, HoPerThread, WoPerThread, NPerThread>{});
 
         // blockwise copy
@@ -448,10 +448,10 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
             constexpr index_t K1 = KPerBlock / KPerThread;
 
 #if 0
-            constexpr auto out_10d_global_desc = make_packed_ConstantTensorDescriptor(
+            constexpr auto out_10d_global_desc = make_ConstantTensorDescriptor_default_rank_packed(
                 Sequence<K / (K1 * K2), K1, K2, Ho, Wo / (W1 * W2 * W3), W1, W2, W3, N / N1, N1>{});
 
-            constexpr auto out_10d_thread_desc = make_packed_ConstantTensorDescriptor(
+            constexpr auto out_10d_thread_desc = make_ConstantTensorDescriptor_default_rank_packed(
                 Sequence<KPerThread / K2, 1, K2, HoPerThread, 1, W1, 1, W3, 1, N1>{});
 #else
             constexpr auto out_10d_global_desc =