Merge branch 'master' into implicit_gemm_v4_backward

109f1e90 · Chao Liu · 0b10c0bb · a68b16a5 · 109f1e90 · 109f1e90
Commit 109f1e90 authored Jun 07, 2019 by Chao Liu
20 changed files
--- a/src/include/Array.hip.hpp
+++ b/src/include/Array.hip.hpp
@@ -18,18 +18,22 @@ struct Array
    __host__ __device__ constexpr index_t GetSize() const { return NSize; }
-    __host__ __device__ constexpr TData operator[](index_t i) const { return mData[i]; }
+    template <index_t I>
+    __host__ __device__ constexpr TData operator[](Number<I>) const
+    {
+        return mData[I];
+    }
-    __host__ __device__ TData& operator[](index_t i) { return mData[i]; }
+    __host__ __device__ constexpr TData operator[](index_t i) const { return mData[i]; }
    template <index_t I>
-    __host__ __device__ constexpr TData Get(Number<I>) const
+    __host__ __device__ TData& operator()(Number<I>)
    {
-        static_assert(I < NSize, "wrong!");
        return mData[I];
    }
+    __host__ __device__ TData& operator()(index_t i) { return mData[i]; }
    template <index_t I>
    __host__ __device__ constexpr void Set(Number<I>, TData x)
    {
@@ -38,16 +42,33 @@ struct Array
        mData[I] = x;
    }
+    __host__ __device__ constexpr void Set(index_t I, TData x) { mData[I] = x; }
+    struct lambda_PushBack // emulate constexpr lambda
+    {
+        const Array<TData, NSize>& old_array;
+        Array<TData, NSize + 1>& new_array;
+        __host__ __device__ constexpr lambda_PushBack(const Array<TData, NSize>& old_array_,
+                                                      Array<TData, NSize + 1>& new_array_)
+            : old_array(old_array_), new_array(new_array_)
+        {
+        }
+        template <index_t I>
+        __host__ __device__ constexpr void operator()(Number<I>) const
+        {
+            new_array.Set(Number<I>{}, old_array[I]);
+        }
+    };
    __host__ __device__ constexpr auto PushBack(TData x) const
    {
        Array<TData, NSize + 1> new_array;
-        static_for<0, NSize, 1>{}([&](auto I) {
+        static_for<0, NSize, 1>{}(lambda_PushBack(*this, new_array));
-            constexpr index_t i = I.Get();
-            new_array[i]        = mData[i];
-        });
-        new_array[NSize] = x;
+        new_array.Set(Number<NSize>{}, x);
        return new_array;
    }
@@ -62,93 +83,60 @@ __host__ __device__ constexpr auto sequence2array(Sequence<Is...>)
 template <class TData, index_t NSize>
 __host__ __device__ constexpr auto make_zero_array()
 {
-#if 0
-    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;
-#else
    constexpr auto zero_sequence = typename uniform_sequence_gen<NSize, 0>::SeqType{};
    constexpr auto zero_array    = sequence2array(zero_sequence);
    return zero_array;
-#endif
 }
 template <class TData, index_t NSize, index_t... IRs>
 __host__ __device__ constexpr auto reorder_array_given_new2old(const Array<TData, NSize>& old_array,
-                                                               Sequence<IRs...> new2old)
+                                                               Sequence<IRs...> /*new2old*/)
 {
-    Array<TData, NSize> new_array;
    static_assert(NSize == sizeof...(IRs), "NSize not consistent");
-    static_for<0, NSize, 1>{}([&](auto IDim) {
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>::value, "wrong! invalid reorder map");
-        constexpr index_t idim = IDim.Get();
-        new_array[idim]        = old_array[new2old.Get(IDim)];
-    });
-    return new_array;
+    return Array<TData, NSize>{old_array.mSize[IRs]...};
 }
-#if 0
-template <class TData, index_t NSize, index_t... IRs>
-__host__ __device__ constexpr auto reorder_array_given_old2new(const Array<TData, NSize>& old_array,
-                                                               Sequence<IRs...> old2new)
-{
-    Array<TData, NSize> new_array;
-    static_assert(NSize == sizeof...(IRs), "NSize not consistent");
-    static_for<0, NSize, 1>{}([&](auto IDim) {
-        constexpr index_t idim       = IDim.Get();
-        new_array[old2new.Get(IDim)] = old_array[idim];
-    });
-    return new_array;
-}
-#else
 template <class TData, index_t NSize, class MapOld2New>
-struct reorder_array_given_old2new_impl
+struct lambda_reorder_array_given_old2new
 {
-    const Array<TData, NSize>& old_array_ref;
+    const Array<TData, NSize>& old_array;
-    Array<TData, NSize>& new_array_ref;
+    Array<TData, NSize>& new_array;
-    __host__
+    __host__ __device__ constexpr lambda_reorder_array_given_old2new(
-        __device__ constexpr reorder_array_given_old2new_impl(const Array<TData, NSize>& old_array,
+        const Array<TData, NSize>& old_array_, Array<TData, NSize>& new_array_)
-                                                              Array<TData, NSize>& new_array)
+        : old_array(old_array_), new_array(new_array_)
-        : old_array_ref(old_array), new_array_ref(new_array)
    {
    }
    template <index_t IOldDim>
    __host__ __device__ constexpr void operator()(Number<IOldDim>) const
    {
-        TData old_data = old_array_ref.Get(Number<IOldDim>{});
+        TData old_data = old_array[IOldDim];
        constexpr index_t INewDim = MapOld2New::Get(Number<IOldDim>{});
-        new_array_ref.Set(Number<INewDim>{}, old_data);
+        new_array.Set(Number<INewDim>{}, old_data);
    }
 };
 template <class TData, index_t NSize, index_t... IRs>
 __host__ __device__ constexpr auto reorder_array_given_old2new(const Array<TData, NSize>& old_array,
-                                                               Sequence<IRs...> old2new)
+                                                               Sequence<IRs...> /*old2new*/)
 {
    Array<TData, NSize> new_array;
    static_assert(NSize == sizeof...(IRs), "NSize not consistent");
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>::value, "wrong! invalid reorder map");
    static_for<0, NSize, 1>{}(
-        reorder_array_given_old2new_impl<TData, NSize, Sequence<IRs...>>(old_array, new_array));
+        lambda_reorder_array_given_old2new<TData, NSize, Sequence<IRs...>>(old_array, new_array));
    return new_array;
 }
-#endif
 template <class TData, index_t NSize, class ExtractSeq>
 __host__ __device__ constexpr auto extract_array(const Array<TData, NSize>& old_array, ExtractSeq)
@@ -159,25 +147,44 @@ __host__ __device__ constexpr auto extract_array(const Array<TData, NSize>& old_
    static_assert(new_size <= NSize, "wrong! too many extract");
-    static_for<0, new_size, 1>{}([&](auto I) {
+    static_for<0, new_size, 1>{}([&](auto I) { new_array(I) = old_array[ExtractSeq::Get(I)]; });
-        constexpr index_t i = I.Get();
-        new_array[i]        = old_array[ExtractSeq::Get(I)];
-    });
    return new_array;
 }
+template <class F, class X, class Y, class Z> // emulate constepxr lambda for array math
+struct lambda_array_math
+{
+    const F& f;
+    const X& x;
+    const Y& y;
+    Z& z;
+    __host__ __device__ constexpr lambda_array_math(const F& f_, const X& x_, const Y& y_, Z& z_)
+        : f(f_), x(x_), y(y_), z(z_)
+    {
+    }
+    template <index_t IDim_>
+    __host__ __device__ constexpr void operator()(Number<IDim_>) const
+    {
+        constexpr auto IDim = Number<IDim_>{};
+        z.Set(IDim, f(x[IDim], y[IDim]));
+    }
+};
 // Array = Array + Array
 template <class TData, index_t NSize>
 __host__ __device__ constexpr auto operator+(Array<TData, NSize> a, Array<TData, NSize> b)
 {
    Array<TData, NSize> result;
-    static_for<0, NSize, 1>{}([&](auto I) {
+    auto f = mod_conv::plus<index_t>{};
-        constexpr index_t i = I.Get();
-        result[i] = a[i] + b[i];
+    static_for<0, NSize, 1>{}(
-    });
+        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
+            f, a, b, result));
    return result;
 }
@@ -188,11 +195,11 @@ __host__ __device__ constexpr auto operator-(Array<TData, NSize> a, Array<TData,
 {
    Array<TData, NSize> result;
-    static_for<0, NSize, 1>{}([&](auto I) {
+    auto f = mod_conv::minus<index_t>{};
-        constexpr index_t i = I.Get();
-        result[i] = a[i] - b[i];
+    static_for<0, NSize, 1>{}(
-    });
+        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
+            f, a, b, result));
    return result;
 }
@@ -205,11 +212,11 @@ __host__ __device__ constexpr auto operator+(Array<TData, NSize> a, Sequence<Is.
    Array<TData, NSize> result;
-    static_for<0, NSize, 1>{}([&](auto I) {
+    auto f = mod_conv::plus<index_t>{};
-        constexpr index_t i = I.Get();
-        result[i] = a[i] + b.Get(I);
+    static_for<0, NSize, 1>{}(
-    });
+        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
+            f, a, b, result));
    return result;
 }
@@ -222,11 +229,11 @@ __host__ __device__ constexpr auto operator-(Array<TData, NSize> a, Sequence<Is.
    Array<TData, NSize> result;
-    static_for<0, NSize, 1>{}([&](auto I) {
+    auto f = mod_conv::minus<index_t>{};
-        constexpr index_t i = I.Get();
-        result[i] = a[i] - b.Get(I);
+    static_for<0, NSize, 1>{}(
-    });
+        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
+            f, a, b, result));
    return result;
 }
@@ -239,11 +246,11 @@ __host__ __device__ constexpr auto operator*(Array<TData, NSize> a, Sequence<Is.
    Array<TData, NSize> result;
-    static_for<0, NSize, 1>{}([&](auto I) {
+    auto f = mod_conv::multiplies<index_t>{};
-        constexpr index_t i = I.Get();
-        result[i] = a[i] * b.Get(I);
+    static_for<0, NSize, 1>{}(
-    });
+        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
+            f, a, b, result));
    return result;
 }
@@ -256,11 +263,11 @@ __host__ __device__ constexpr auto operator-(Sequence<Is...> a, Array<TData, NSi
    Array<TData, NSize> result;
-    static_for<0, NSize, 1>{}([&](auto I) {
+    auto f = mod_conv::minus<index_t>{};
-        constexpr index_t i = I.Get();
-        result[i] = a.Get(I) - b[i];
+    static_for<0, NSize, 1>{}(
-    });
+        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
+            f, a, b, result));
    return result;
 }
@@ -273,10 +280,7 @@ accumulate_on_array(const Array<TData, NSize>& a, Reduce f, TData init)
    static_assert(NSize > 0, "wrong");
-    static_for<0, NSize, 1>{}([&](auto I) {
+    static_for<0, NSize, 1>{}([&](auto I) { result = f(result, a[I]); });
-        constexpr index_t i = I.Get();
-        result              = f(result, a[i]);
-    });
    return result;
 }

--- a/src/include/ConstantMergedTensorDescriptor.hip.hpp
+++ b/src/include/ConstantMergedTensorDescriptor.hip.hpp
@@ -9,6 +9,8 @@
 template <class OriginalTensorDesc, class... OriginalDimMergeSeqs>
 struct ConstantMergedTensorDescriptor
 {
+    using Type = ConstantMergedTensorDescriptor;
    static constexpr auto mOriginalDimMergeSeqs = std::tuple<OriginalDimMergeSeqs...>{};
    static constexpr index_t nDim         = sizeof...(OriginalDimMergeSeqs);
@@ -74,43 +76,17 @@ struct ConstantMergedTensorDescriptor
        return OriginalTensorDesc::GetElementSize();
    }
-#if 0
-    __host__ __device__ static constexpr auto
-    GetOriginalMultiIndexFromMultiIndex(Array<index_t, nDim> multi_id)
-    {
-        Array<index_t, nOriginalDim> original_multi_id;
-        static_for<0, nDim, 1>{}([&](auto IDim) {
-            constexpr index_t idim               = IDim.Get();
-            constexpr auto original_dims_partial = std::get<idim>(mOriginalDimMergeSeqs);
-            // get partial original-multi-id corresponding to this merged dimension
-            const auto original_multi_id_partial =
-                OriginalTensorDesc::Extract(original_dims_partial)
-                    .GetMultiIndexFrom1dIndex(multi_id[idim]);
-            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.Get()];
-            });
-        });
-        return original_multi_id;
-    }
-#else
    template <class OriginalDimsPartial>
-    struct GetOriginalMultiIndexFromMultiIndex_impl1
+    struct lambda_1_GetOriginalMultiIndexFromMultiIndex
    {
-        const Array<index_t, OriginalDimsPartial::GetSize()>& original_multi_id_partial_ref;
+        const Array<index_t, OriginalDimsPartial::GetSize()>& original_multi_id_partial;
-        Array<index_t, nOriginalDim>& original_multi_id_ref;
+        Array<index_t, nOriginalDim>& original_multi_id;
-        __host__ __device__ constexpr GetOriginalMultiIndexFromMultiIndex_impl1(
+        __host__ __device__ constexpr lambda_1_GetOriginalMultiIndexFromMultiIndex(
-            const Array<index_t, OriginalDimsPartial::GetSize()>& original_multi_id_partial,
+            const Array<index_t, OriginalDimsPartial::GetSize()>& original_multi_id_partial_,
-            Array<index_t, nOriginalDim>& original_multi_id)
+            Array<index_t, nOriginalDim>& original_multi_id_)
-            : original_multi_id_partial_ref(original_multi_id_partial),
+            : original_multi_id_partial(original_multi_id_partial_),
-              original_multi_id_ref(original_multi_id)
+              original_multi_id(original_multi_id_)
        {
        }
@@ -119,37 +95,36 @@ struct ConstantMergedTensorDescriptor
        {
            constexpr index_t idim_original = OriginalDimsPartial::Get(Number<I>{});
-            index_t itmp = original_multi_id_partial_ref.Get(Number<I>{});
+            index_t itmp = original_multi_id_partial[I];
-            original_multi_id_ref.Set(Number<idim_original>{}, itmp);
+            original_multi_id.Set(Number<idim_original>{}, itmp);
        }
    };
-    struct GetOriginalMultiIndexFromMultiIndex_impl0
+    struct lambda_0_GetOriginalMultiIndexFromMultiIndex
    {
-        const Array<index_t, nDim>& multi_id_ref;
+        const Array<index_t, nDim>& multi_id;
-        Array<index_t, nOriginalDim>& original_multi_id_ref;
+        Array<index_t, nOriginalDim>& original_multi_id;
-        __host__ __device__ constexpr GetOriginalMultiIndexFromMultiIndex_impl0(
+        __host__ __device__ constexpr lambda_0_GetOriginalMultiIndexFromMultiIndex(
-            const Array<index_t, nDim>& multi_id, Array<index_t, nOriginalDim>& original_multi_id)
+            const Array<index_t, nDim>& multi_id_, Array<index_t, nOriginalDim>& original_multi_id_)
-            : multi_id_ref(multi_id), original_multi_id_ref(original_multi_id)
+            : multi_id(multi_id_), original_multi_id(original_multi_id_)
        {
        }
        template <index_t IDim>
        __host__ __device__ constexpr void operator()(Number<IDim>) const
        {
-            constexpr auto original_dims_partial =
+            constexpr auto original_dims_partial = std::get<IDim>(Type::mOriginalDimMergeSeqs);
-                std::get<IDim>(std::tuple<OriginalDimMergeSeqs...>{});
            // get partial original-multi-id corresponding to this merged dimension
            const auto original_multi_id_partial =
                OriginalTensorDesc::Extract(original_dims_partial)
-                    .GetMultiIndexFrom1dIndex(multi_id_ref[IDim]);
+                    .GetMultiIndexFrom1dIndex(multi_id[IDim]);
            static_for<0, original_dims_partial.GetSize(), 1>{}(
-                GetOriginalMultiIndexFromMultiIndex_impl1<decltype(original_dims_partial)>(
+                lambda_1_GetOriginalMultiIndexFromMultiIndex<decltype(original_dims_partial)>(
-                    original_multi_id_partial, original_multi_id_ref));
+                    original_multi_id_partial, original_multi_id));
        }
    };
@@ -160,7 +135,7 @@ struct ConstantMergedTensorDescriptor
        Array<index_t, nOriginalDim> original_multi_id;
        static_for<0, nDim, 1>{}(
-            GetOriginalMultiIndexFromMultiIndex_impl0(multi_id, original_multi_id));
+            lambda_0_GetOriginalMultiIndexFromMultiIndex(multi_id, original_multi_id));
        return original_multi_id;
    }
@@ -174,7 +149,6 @@ struct ConstantMergedTensorDescriptor
        return OriginalTensorDesc::GetOffsetFromMultiIndex(original_multi_id);
    }
-#endif
    __host__ __device__ static constexpr index_t
    GetOffsetFromMultiIndex(Array<index_t, nDim> multi_id)
@@ -192,9 +166,9 @@ struct ConstantMergedTensorDescriptor
    __host__ __device__ static constexpr Array<index_t, nDim> GetMultiIndexFrom1dIndex(index_t id)
    {
-        constexpr auto dummy_desc = make_ConstantTensorDescriptor_packed(GetLengths());
+        constexpr auto packed_desc = make_ConstantTensorDescriptor_packed(GetLengths());
-        return dummy_desc.GetMultiIndexFrom1dIndex(id);
+        return packed_desc.GetMultiIndexFrom1dIndex(id);
    }
 };

--- a/src/include/ConstantTensorDescriptor.hip.hpp
+++ b/src/include/ConstantTensorDescriptor.hip.hpp
@@ -48,26 +48,48 @@ struct ConstantTensorDescriptor
    template <index_t I>
    __host__ __device__ static constexpr index_t GetLength(Number<I>)
    {
-        return Lengths{}.Get(Number<I>{});
+        return Lengths::Get(Number<I>{});
    }
    template <index_t I>
    __host__ __device__ static constexpr index_t GetStride(Number<I>)
    {
-        return Strides{}.Get(Number<I>{});
+        return Strides::Get(Number<I>{});
    }
-    __host__ __device__ static constexpr bool AreStridesNonAscending()
+    struct lambda_AreDimensionsContinuous
    {
-        bool flag = true;
+        bool& is_continuous;
-        static_for<0, nDim - 1, 1>{}([&](auto IDim) {
+        __host__ __device__ constexpr lambda_AreDimensionsContinuous(bool& is_continuous_)
-            constexpr auto IDim_p1 = Number<IDim.Get() + 1>{};
+            : is_continuous(is_continuous_)
+        {
+        }
-            flag = flag && (GetLength(IDim) >= GetLength(IDim_p1));
+        template <index_t IDim_>
-        });
+        __host__ __device__ constexpr void operator()(Number<IDim_>) const
+        {
+            constexpr auto IDim    = Number<IDim_>{};
+            constexpr auto IDim_p1 = Number<IDim_ + 1>{};
+            is_continuous =
+                is_continuous && (GetStride(IDim) >= GetStride(IDim_p1) &&
+                                  GetStride(IDim) == GetStride(IDim_p1) * GetLength(IDim_p1));
+        }
+    };
-        return flag;
+    __host__ __device__ static constexpr bool AreDimensionsContinuous()
+    {
+        bool is_continuous = true;
+        static_for<0, nDim - 1, 1>{}(lambda_AreDimensionsContinuous(is_continuous));
+        return is_continuous;
+    }
+    __host__ __device__ static constexpr bool IsPackedTensor()
+    {
+        return AreDimensionsContinuous() && GetStride(Number<nDim - 1>{}) == 1;
    }
    template <class T>
@@ -92,40 +114,24 @@ struct ConstantTensorDescriptor
        return align.Get() * ((element_space_unaligned + align.Get() - 1) / align.Get());
    }
-#if 0
+    // emulate constexpr lambda
-    template <index_t NSize>
-    __host__ __device__ static constexpr index_t
-    GetOffsetFromMultiIndex(Array<index_t, NSize> multi_id)
-    {
-        static_assert(NSize == nDim, "wrong! Dimension not consistent");
-        index_t offset = 0;
-        static_for<0, nDim, 1>{}([&](auto IDim) {
-            constexpr index_t idim = IDim.Get();
-            offset += multi_id[idim] * GetStride(IDim);
-        });
-        return offset;
-    }
-#else
    template <index_t NSize>
-    struct GetOffsetFromMultiIndex_impl
+    struct lambda_GetOffsetFromMultiIndex
    {
-        Array<index_t, NSize>& multi_id_ref;
+        Array<index_t, NSize>& multi_id;
-        index_t& offset_ref;
+        index_t& offset;
-        __host__ __device__ constexpr GetOffsetFromMultiIndex_impl(Array<index_t, NSize>& multi_id,
+        __host__
-                                                                   index_t& offset)
+            __device__ constexpr lambda_GetOffsetFromMultiIndex(Array<index_t, NSize>& multi_id_,
-            : multi_id_ref(multi_id), offset_ref(offset)
+                                                                index_t& offset_)
+            : multi_id(multi_id_), offset(offset_)
        {
        }
-        template <index_t IDim>
+        template <class X>
-        __host__ __device__ constexpr bool operator()(Number<IDim>) const
+        __host__ __device__ constexpr void operator()(X IDim) const
        {
-            offset_ref += multi_id_ref.Get(Number<IDim>{}) * Type::GetStride(Number<IDim>{});
+            offset += multi_id[IDim] * Type::GetStride(IDim);
-            return true;
        }
    };
@@ -137,11 +143,10 @@ struct ConstantTensorDescriptor
        index_t offset = 0;
-        static_for<0, nDim, 1>{}(GetOffsetFromMultiIndex_impl<NSize>(multi_id, offset));
+        static_for<0, nDim, 1>{}(lambda_GetOffsetFromMultiIndex<NSize>(multi_id, offset));
        return offset;
    }
-#endif
    template <class... Is>
    __host__ __device__ static constexpr index_t GetOffsetFromMultiIndex(Is... is)
@@ -160,47 +165,27 @@ struct ConstantTensorDescriptor
            multi_id * GetStrides(), mod_conv::plus<index_t>{}, Number<0>{});
    }
-#if 0
+    // emulate constexpr lambda
-    __host__ __device__ static constexpr Array<index_t, nDim> GetMultiIndexFrom1dIndex(index_t id)
+    template <class PackedStrides>
-    {
+    struct lambda_GetMultiIndexFrom1dIndex
-        Array<index_t, nDim> multi_id;
-        constexpr auto dummy_strides = calculate_tensor_strides_packed(GetLengths());
-        // calculate index in each of the dimensions in the order of their dimension
-        static_for<0, nDim - 1, 1>{}([&](auto IDim) {
-            constexpr index_t idim   = IDim.Get();
-            constexpr index_t stride = dummy_strides.Get(Number<idim>{});
-            multi_id[idim]           = id / stride;
-            id -= multi_id[idim] * stride;
-        });
-        multi_id[nDim - 1] = id / dummy_strides.Get(Number<nDim - 1>{});
-        return multi_id;
-    }
-#else
-    struct GetMultiIndexFrom1dIndex_impl
    {
-        using DummyStrides = decltype(calculate_tensor_strides_packed(GetLengths()));
+        index_t& id;
+        Array<index_t, nDim>& multi_id;
-        index_t& id_ref;
-        Array<index_t, nDim>& multi_id_ref;
-        __host__ __device__ constexpr GetMultiIndexFrom1dIndex_impl(index_t& id,
+        __host__
-                                                                    Array<index_t, nDim>& multi_id)
+            __device__ constexpr lambda_GetMultiIndexFrom1dIndex(index_t& id_,
-            : id_ref(id), multi_id_ref(multi_id)
+                                                                 Array<index_t, nDim>& multi_id_)
+            : id(id_), multi_id(multi_id_)
        {
        }
-        template <index_t IDim>
+        template <class IDim_>
-        __host__ __device__ constexpr bool operator()(Number<IDim>) const
+        __host__ __device__ constexpr void operator()(IDim_) const
        {
-            constexpr index_t stride = DummyStrides::Get(Number<IDim>{});
+            constexpr auto IDim      = IDim_{};
-            multi_id_ref.Set(Number<IDim>{}, id_ref / stride);
+            constexpr index_t stride = PackedStrides::Get(IDim);
-            id_ref -= multi_id_ref.Get(Number<IDim>{}) * stride;
+            multi_id.Set(IDim, id / stride);
+            id -= multi_id[IDim] * stride;
-            return true;
        }
    };
@@ -208,27 +193,15 @@ struct ConstantTensorDescriptor
    {
        Array<index_t, nDim> multi_id;
-        constexpr auto dummy_strides = calculate_tensor_strides_packed(GetLengths());
+        using PackedStrides = decltype(calculate_tensor_strides_packed(GetLengths()));
        // calculate index in each of the dimensions in the order of their dimension
-        static_for<0, nDim - 1, 1>{}(GetMultiIndexFrom1dIndex_impl(id, multi_id));
+        static_for<0, nDim - 1, 1>{}(lambda_GetMultiIndexFrom1dIndex<PackedStrides>(id, multi_id));
-        index_t itmp = id / dummy_strides.Get(Number<nDim - 1>{});
-        multi_id.Set(Number<nDim - 1>{}, itmp);
+        multi_id.Set(Number<nDim - 1>{}, id / PackedStrides::Get(Number<nDim - 1>{}));
        return multi_id;
    }
-#endif
-#if 0
-    // return type is Sequence<...>
-    template<index_t Id>
-    __host__ __device__ static constexpr auto GetMultiIndexFrom1dIndex(Number<Id>)
-    {
-        return inclusive_scan_sequence(f_impl, GetStrides(), Number<Id>{});
-    }
-#endif
    __host__ __device__ static constexpr auto
    GetOriginalMultiIndexFromMultiIndex(Array<index_t, nDim> multi_id)
@@ -236,9 +209,10 @@ struct ConstantTensorDescriptor
        return multi_id;
    }
-    // This function doesn't do carry check on the highest dimension, for performance reason.
+    // This function doesn't do carry check on the highest dimension for positive stepping (or
-    // It is the user's responsibility to make sure the result "new_mutli_id" is not out-of-bound
+    // borrow check on the lowest dimension for negative stepping) , for performance reason. It is
-    // on the highest dimension
+    // the user's responsibility to make sure the result "new_mutli_id" is not out-of-bound on the
+    // highest dimension for positive stepping (or on the lowest dimension for negative stepping)
    template <bool PositiveDirection>
    __host__ __device__ static Array<index_t, nDim>
    UpdateMultiIndexGivenStepSizeOf1dIndex(Array<index_t, nDim> old_multi_id,
@@ -262,14 +236,14 @@ struct ConstantTensorDescriptor
                if(carry)
                {
-                    ++new_multi_id[idim];
+                    ++new_multi_id(idim);
                }
                carry = false;
                if(new_multi_id[idim] >= GetLength(IDim))
                {
-                    new_multi_id[idim] -= GetLength(IDim);
+                    new_multi_id(idim) -= GetLength(IDim);
                    carry = true;
                }
            });
@@ -288,14 +262,14 @@ struct ConstantTensorDescriptor
                if(borrow)
                {
-                    --new_multi_id[idim];
+                    --new_multi_id(idim);
                }
                borrow = false;
                if(new_multi_id[idim] < GetLength(IDim))
                {
-                    new_multi_id[idim] += GetLength(IDim);
+                    new_multi_id(idim) += GetLength(IDim);
                    borrow = true;
                }
            });
@@ -382,15 +356,7 @@ struct ConstantTensorDescriptor
        return ConstantTensorDescriptor<decltype(new_lengths), decltype(new_strides)>{};
    }
-    template <index_t Threashold, index_t Delta>
+    // this function unfold dimension [FirstUnfoldDim, ..., LastUnfoldDim] into 1 dimension
-    struct f_unfold_impl
-    {
-        __host__ __device__ constexpr index_t operator()(index_t x) const
-        {
-            return x > Threashold ? x - Delta : x;
-        }
-    };
    template <index_t FirstUnfoldDim, index_t LastUnfoldDim>
    __host__ __device__ static constexpr auto Unfold(Number<FirstUnfoldDim>, Number<LastUnfoldDim>)
    {
@@ -398,24 +364,6 @@ 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_) {
-            constexpr auto IDim    = decltype(IDim_){};
-            constexpr auto IDim_p1 = IDim + Number<1>{};
-            // check stride
-            static_assert(
-                GetStride(IDim) >= GetStride(IDim_p1),
-                "wrong! dimensions to be unfolded need to be in descending order w.r.t strides");
-            // check if packed
-            static_assert(GetStride(IDim_p1) * GetLength(IDim_p1) == GetStride(IDim),
-                          "wrong! dimensions to be unfolded need to be packed");
-        });
-#endif
        // left and right
        constexpr auto left = typename arithmetic_sequence_gen<0, FirstUnfoldDim, 1>::SeqType{};
        constexpr auto middle =
@@ -423,6 +371,9 @@ struct ConstantTensorDescriptor
        constexpr auto right =
            typename arithmetic_sequence_gen<LastUnfoldDim + 1, GetNumOfDimension(), 1>::SeqType{};
+        // dimensions to be unfolded need to be continuous
+        static_assert(Type::Extract(middle).AreDimensionsContinuous(), "wrong! not unfoldable");
        // unfolded length, stride
        constexpr index_t unfold_length = accumulate_on_sequence(
            GetLengths().Extract(middle), mod_conv::multiplies<index_t>{}, Number<1>{});
@@ -446,16 +397,16 @@ struct ConstantTensorDescriptor
    template <class MapNew2Old>
    __host__ __device__ static constexpr auto ReorderGivenNew2Old(MapNew2Old)
    {
-        return ConstantTensorDescriptor<decltype(Lengths{}.ReorderGivenNew2Old(MapNew2Old{})),
+        return ConstantTensorDescriptor<decltype(Lengths::ReorderGivenNew2Old(MapNew2Old{})),
-                                        decltype(Strides{}.ReorderGivenNew2Old(MapNew2Old{}))>{};
+                                        decltype(Strides::ReorderGivenNew2Old(MapNew2Old{}))>{};
    }
 #if 0 // require sequence_sort, which is not implemented yet
    template <class MapOld2New>
    __host__ __device__ static constexpr auto ReorderGivenOld2New(MapOld2New)
    {
-        return ConstantTensorDescriptor<decltype(Lengths{}.ReorderGivenOld2New(MapOld2New{})),
+        return ConstantTensorDescriptor<decltype(Lengths::ReorderGivenOld2New(MapOld2New{})),
-                                        decltype(Strides{}.ReorderGivenOld2New(MapOld2New{}))>{}
+                                        decltype(Strides::ReorderGivenOld2New(MapOld2New{}))>{}
    }
 #endif
 };

--- a/src/include/Sequence.hip.hpp
+++ b/src/include/Sequence.hip.hpp
@@ -2,6 +2,9 @@
 #include "integral_constant.hip.hpp"
 #include "functional.hip.hpp"
+template <class Seq>
+struct is_valid_sequence_map;
 template <index_t... Is>
 struct Sequence
 {
@@ -16,7 +19,23 @@ struct Sequence
    {
        static_assert(I < mSize, "wrong! I too large");
-        // the last dummy element is to prevent compiler complain about empty Sequence
+        // the last dummy element is to prevent compiler complain about empty array, when mSize = 0
+        const index_t mData[mSize + 1] = {Is..., 0};
+        return mData[I];
+    }
+    template <index_t I>
+    __host__ __device__ constexpr index_t operator[](Number<I>) const
+    {
+        static_assert(I < mSize, "wrong! I too large");
+        const index_t mData[mSize + 1] = {Is..., 0};
+        return mData[I];
+    }
+    // make sure I is constepxr
+    __host__ __device__ constexpr index_t operator[](index_t I) const
+    {
        const index_t mData[mSize + 1] = {Is..., 0};
        return mData[I];
    }
@@ -24,24 +43,24 @@ struct Sequence
    template <index_t... IRs>
    __host__ __device__ static constexpr auto ReorderGivenNew2Old(Sequence<IRs...> /*new2old*/)
    {
-#if 0 // require sequence_sort, which is not implemented yet
+        static_assert(sizeof...(Is) == sizeof...(IRs),
-        static_assert(is_same<sequence_sort<Sequence<IRs...>>::SortedSeqType,
+                      "wrong! reorder map should have the same size as Sequence to be rerodered");
-                              arithmetic_sequence_gen<0, mSize, 1>::SeqType>::value,
-                      "wrong! invalid new2old map");
+        static_assert(is_valid_sequence_map<Sequence<IRs...>>::value, "wrong! invalid reorder map");
-#endif
-        return Sequence<Type{}.Get(Number<IRs>{})...>{};
+        return Sequence<Type::Get(Number<IRs>{})...>{};
    }
 #if 0 // require sequence_sort, which is not implemented yet
    template <class MapOld2New>
    __host__ __device__ static constexpr auto ReorderGivenOld2New(MapOld2New /*old2new*/)
    {
-#if 0
+        static_assert(sizeof...(Is) == MapOld2New::GetSize(),
-        static_assert(is_same<sequence_sort<MapOld2New>::SortedSeqType,
+                      "wrong! reorder map should have the same size as Sequence to be rerodered");
-                              arithmetic_sequence_gen<0, mSize, 1>::SeqType>::value,
-                      "wrong! invalid old2new map");
+        static_assert(is_valid_sequence_map<MapOld2New>::value, 
-#endif
+                      "wrong! invalid reorder map");
        constexpr auto map_new2old = typename sequence_map_inverse<MapOld2New>::SeqMapType{};
        return ReorderGivenNew2Old(map_new2old);
@@ -87,13 +106,13 @@ struct Sequence
    template <index_t... Ns>
    __host__ __device__ static constexpr auto Extract(Number<Ns>...)
    {
-        return Sequence<Type{}.Get(Number<Ns>{})...>{};
+        return Sequence<Type::Get(Number<Ns>{})...>{};
    }
    template <index_t... Ns>
    __host__ __device__ static constexpr auto Extract(Sequence<Ns...>)
    {
-        return Sequence<Type{}.Get(Number<Ns>{})...>{};
+        return Sequence<Type::Get(Number<Ns>{})...>{};
    }
    template <index_t I, index_t X>
@@ -297,6 +316,7 @@ struct sequence_map_inverse<Sequence<Is...>>
 };
 #endif
 template <class Seq>
 struct is_valid_sequence_map
 {
@@ -322,11 +342,6 @@ __host__ __device__ constexpr auto operator-(Sequence<Xs...> seq_x, Sequence<Ys.
 {
    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
-#if 0
-    static_for<0, seq_x.GetSize(), 1>{}(
-        [&](auto I) { static_assert(seq_x.Get(I) >= seq_y.Get(I), "wrong! going to undeflow"); });
-#endif
    return Sequence<(Xs - Ys)...>{};
 }
@@ -363,15 +378,6 @@ __host__ __device__ constexpr auto operator+(Sequence<Xs...>, Number<Y>)
 template <index_t... Xs, index_t Y>
 __host__ __device__ constexpr auto operator-(Sequence<Xs...>, Number<Y>)
 {
-#if 0 // TODO: turn it on. Doesn't compile
-    constexpr auto seq_x = Sequence<Xs...>{};
-    static_for<0, sizeof...(Xs), 1>{}([&](auto Iter) {
-        constexpr auto I = decltype(Iter){};
-        static_assert(seq_x.Get(I) >= Y, "wrong! going to underflow");
-    });
-#endif
    return Sequence<(Xs - Y)...>{};
 }
@@ -404,13 +410,6 @@ __host__ __device__ constexpr auto operator-(Number<Y>, Sequence<Xs...>)
 {
    constexpr auto seq_x = Sequence<Xs...>{};
-#if 0
-    static_for<0, sizeof...(Xs), 1>{}([&](auto Iter) {
-        constexpr auto I = decltype(Iter){};
-        static_assert(seq_x.Get(I) <= Y, "wrong! going to underflow");
-    });
-#endif
    return Sequence<(Y - Xs)...>{};
 }
@@ -482,25 +481,6 @@ __host__ __device__ constexpr auto inclusive_scan_sequence(Seq, Reduce, Number<I
    return reverse_inclusive_scan_sequence(Seq{}.Reverse(), Reduce{}, Number<Init>{}).Reverse();
 }
-template <class Seq>
-struct accumulate_on_sequence_impl
-{
-    template <class IDim>
-    __host__ __device__ constexpr index_t operator()(IDim) const
-    {
-        return Seq{}.Get(IDim{});
-    }
-};
-template <class Seq, class Reduce, index_t I>
-__host__ __device__ constexpr index_t
-    accumulate_on_sequence(Seq, Reduce, Number<I> /*initial_value*/)
-{
-    constexpr index_t a =
-        static_const_reduce_n<Seq::mSize>{}(accumulate_on_sequence_impl<Seq>{}, Reduce{});
-    return Reduce{}(a, I);
-}
 template <index_t... Is>
 __host__ __device__ constexpr auto Sequence<Is...>::PopFront()
 {

--- a/src/include/base.hip.hpp
+++ b/src/include/base.hip.hpp
+#pragma once
+__device__ index_t get_thread_local_1d_id() { return threadIdx.x; }
+__device__ index_t get_block_1d_id() { return blockIdx.x; }
+template <class T1, class T2>
+struct is_same
+{
+    static constexpr bool value = false;
+};
+template <class T>
+struct is_same<T, T>
+{
+    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
+{
+    __host__ __device__ constexpr T operator()(T a) const { return s * a; }
+};
+template <class T>
+struct plus
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a + b; }
+};
+template <class T>
+struct minus
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a - b; }
+};
+template <class T>
+struct multiplies
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a * b; }
+};
+template <class T>
+struct integer_divide_ceiler
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const
+    {
+        static_assert(is_same<T, index_t>::value || is_same<T, int>::value, "wrong type");
+        return (a + b - 1) / b;
+    }
+};
+template <class T>
+__host__ __device__ constexpr T integer_divide_ceil(T a, T b)
+{
+    static_assert(is_same<T, index_t>::value || is_same<T, int>::value, "wrong type");
+    return (a + b - 1) / b;
+}
+template <class T>
+__host__ __device__ constexpr T max(T x, T y)
+{
+    return x > y ? x : y;
+}
+template <class T, class... Ts>
+__host__ __device__ constexpr T max(T x, Ts... xs)
+{
+    static_assert(sizeof...(xs) > 0, "not enough argument");
+    auto y = max(xs...);
+    static_assert(is_same<decltype(y), T>::value, "not the same type");
+    return x > y ? x : y;
+}
+template <class T>
+__host__ __device__ constexpr T min(T x, T y)
+{
+    return x < y ? x : y;
+}
+template <class T, class... Ts>
+__host__ __device__ constexpr T min(T x, Ts... xs)
+{
+    static_assert(sizeof...(xs) > 0, "not enough argument");
+    auto y = min(xs...);
+    static_assert(is_same<decltype(y), T>::value, "not the same type");
+    return x < y ? x : y;
+}
+// this is wrong
+// TODO: implement correct least common multiple, instead of calling max()
+template <class T, class... Ts>
+__host__ __device__ constexpr T lcm(T x, Ts... xs)
+{
+    return max(x, xs...);
+}
+} // namespace mod_conv
--- a/src/include/blockwise_generic_tensor_slice_op.hip.hpp
+++ b/src/include/blockwise_generic_tensor_slice_op.hip.hpp
@@ -122,7 +122,7 @@ struct BlockwiseGenericTensorSliceCopy_v1
            constexpr auto src_partial_original_desc =
                SrcDesc::GetOriginalTensorDescriptor().Extract(src_partial_original_dims);
-            mThreadSrcPartialOffsets[idim] = src_partial_original_desc.GetOffsetFromMultiIndex(
+            mThreadSrcPartialOffsets(idim) = src_partial_original_desc.GetOffsetFromMultiIndex(
                extract_array(mThreadSrcOriginalMultiId, src_partial_original_dims));
        });
@@ -136,7 +136,7 @@ struct BlockwiseGenericTensorSliceCopy_v1
            constexpr auto dst_partial_original_desc =
                DstDesc::GetOriginalTensorDescriptor().Extract(dst_partial_original_dims);
-            mThreadDstPartialOffsets[idim] = dst_partial_original_desc.GetOffsetFromMultiIndex(
+            mThreadDstPartialOffsets(idim) = dst_partial_original_desc.GetOffsetFromMultiIndex(
                extract_array(mThreadDstOriginalMultiId, dst_partial_original_dims));
        });
@@ -206,18 +206,16 @@ struct BlockwiseGenericTensorSliceCopy_v1
 #if 0
            constexpr auto repeat_multi_id = sequence2array(decltype(repeat_multi_id_){});
-            const auto src_thread_data_multi_id_begin =
+            const auto src_thread_data_multi_id_begin = repeat_multi_id * data_per_cluster_per_dims;
-                repeat_multi_id * data_per_cluster_per_dims; // cannot not constexpr, why?
-            const auto clipboard_data_multi_id_begin =
+            const auto clipboard_data_multi_id_begin = repeat_multi_id * thread_sub_tensor_lengths;
-                repeat_multi_id * thread_sub_tensor_lengths; // cannot not constexpr, why?
-            const index_t src_offset = SrcDesc{}.GetOffsetFromMultiIndex(
+            const index_t src_offset =
-                src_thread_data_multi_id_begin); // cannot not constexpr, why?
+                SrcDesc{}.GetOffsetFromMultiIndex(src_thread_data_multi_id_begin);
-            const index_t clipboard_offset = thread_tensor_desc.GetOffsetFromMultiIndex(
+            const index_t clipboard_offset =
-                clipboard_data_multi_id_begin); // cannot not constexpr, why?
+                thread_tensor_desc.GetOffsetFromMultiIndex(clipboard_data_multi_id_begin);
-#else
+#else // HIP compiler performs better with these codes
            constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
            constexpr auto src_thread_data_multi_id_begin =
@@ -261,18 +259,15 @@ struct BlockwiseGenericTensorSliceCopy_v1
 #if 0
            constexpr auto repeat_multi_id = sequence2array(decltype(repeat_multi_id_){});
-            const auto clipboard_data_multi_id_begin =
+            const auto clipboard_data_multi_id_begin = repeat_multi_id * thread_sub_tensor_lengths;
-                repeat_multi_id * thread_sub_tensor_lengths; // cannot not constexpr, why?
-            const auto dst_data_multi_id_begin =
+            const auto dst_data_multi_id_begin = repeat_multi_id * data_per_cluster_per_dims;
-                repeat_multi_id * data_per_cluster_per_dims; // cannot not constexpr, why?
-            const index_t clipboard_offset = thread_tensor_desc.GetOffsetFromMultiIndex(
+            const index_t clipboard_offset =
-                clipboard_data_multi_id_begin); // cannot not constexpr, why?
+                thread_tensor_desc.GetOffsetFromMultiIndex(clipboard_data_multi_id_begin);
-            const index_t dst_offset = DstDesc{}.GetOffsetFromMultiIndex(
+            const index_t dst_offset = DstDesc{}.GetOffsetFromMultiIndex(dst_data_multi_id_begin);
-                dst_data_multi_id_begin); // cannot not constexpr, why?
+#else // HIP compiler performs better with these codes
-#else
            constexpr auto repeat_multi_id = decltype(repeat_multi_id_){};
            constexpr auto clipboard_data_multi_id_begin =
@@ -343,33 +338,12 @@ struct BlockwiseGenericTensorSliceCopy_v1
                src_partial_original_desc.UpdateMultiIndexGivenStepSizeOf1dIndex(
                    old_src_partial_original_multi_id, StepSize, direction);
-#if 0
-            {
-                if(debug_flag && get_block_1d_id() == 0)
-                {
-                    printf("id %5u %5u: "
-                           "old_src_partial_original_multi_id %u %u %u, "
-                           "new_src_partial_original_multi_id %u %u %u, "
-                           "mThreadSrcOffset %u, mThreadDstOffset %u \n",
-                           get_block_1d_id(),
-                           get_thread_local_1d_id(),
-                           old_src_partial_original_multi_id[0],
-                           old_src_partial_original_multi_id[1],
-                           old_src_partial_original_multi_id[2],
-                           new_src_partial_original_multi_id[0],
-                           new_src_partial_original_multi_id[1],
-                           new_src_partial_original_multi_id[2]
-                           );
-                }
-            }
-#endif
            // update "mThreadSrcOriginalMultiId"
            static_for<0, decltype(src_partial_original_dims)::GetSize(), 1>{}([&](auto I_) {
                constexpr auto I                = decltype(I_){};
                constexpr index_t idim_original = src_partial_original_dims.Get(I);
-                mThreadSrcOriginalMultiId[idim_original] =
+                mThreadSrcOriginalMultiId(idim_original) =
                    new_src_partial_original_multi_id[I.Get()];
            });
@@ -381,7 +355,7 @@ struct BlockwiseGenericTensorSliceCopy_v1
                    new_src_partial_original_multi_id);
            // update "mThreadSrcPartialOffsets"
-            mThreadSrcPartialOffsets[idim] = new_src_partial_offset;
+            mThreadSrcPartialOffsets(idim) = new_src_partial_offset;
            // update "mThreadSrcOffset", do "+" before "-" to avoid underflow
            mThreadSrcOffset = (mThreadSrcOffset + new_src_partial_offset) - old_src_partial_offset;
@@ -401,15 +375,15 @@ struct BlockwiseGenericTensorSliceCopy_v1
            static_if<PositiveDirection>{}([&](auto fwd) {
                mThreadSrcOffset += StepSize * fwd(SrcDesc{}).GetStride(IDim);
-                mThreadSrcOriginalMultiId[idim_original] += StepSize;
+                mThreadSrcOriginalMultiId(idim_original) += StepSize;
-                mThreadSrcPartialOffsets[idim] += StepSize * fwd(SrcDesc{}).GetStride(IDim);
+                mThreadSrcPartialOffsets(idim) += StepSize * fwd(SrcDesc{}).GetStride(IDim);
            }).Else([&](auto fwd) {
                mThreadSrcOffset -= StepSize * fwd(SrcDesc{}).GetStride(IDim);
-                mThreadSrcOriginalMultiId[idim_original] -= StepSize;
+                mThreadSrcOriginalMultiId(idim_original) -= StepSize;
-                mThreadSrcPartialOffsets[idim] -= StepSize * fwd(SrcDesc{}).GetStride(IDim);
+                mThreadSrcPartialOffsets(idim) -= StepSize * fwd(SrcDesc{}).GetStride(IDim);
            });
        });
    }

--- a/src/include/common.hip.hpp
+++ b/src/include/common.hip.hpp
 #pragma once
+#include "base.hip.hpp"
 #include "vector_type.hip.hpp"
 #include "integral_constant.hip.hpp"
 #include "Sequence.hip.hpp"
@@ -10,109 +11,3 @@
 #if USE_AMD_INLINE_ASM
 #include "amd_inline_asm.hip.hpp"
 #endif
-__device__ index_t get_thread_local_1d_id() { return threadIdx.x; }
-__device__ index_t get_block_1d_id() { return blockIdx.x; }
-template <class T1, class T2>
-struct is_same
-{
-    static constexpr bool value = false;
-};
-template <class T>
-struct is_same<T, T>
-{
-    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
-{
-    __host__ __device__ constexpr T operator()(T a) const { return s * a; }
-};
-template <class T>
-struct plus
-{
-    __host__ __device__ constexpr T operator()(T a, T b) const { return a + b; }
-};
-template <class T>
-struct multiplies
-{
-    __host__ __device__ constexpr T operator()(T a, T b) const { return a * b; }
-};
-template <class T>
-struct integer_divide_ceiler
-{
-    __host__ __device__ constexpr T operator()(T a, T b) const
-    {
-        static_assert(is_same<T, index_t>::value || is_same<T, int>::value, "wrong type");
-        return (a + b - 1) / b;
-    }
-};
-template <class T>
-__host__ __device__ constexpr T integer_divide_ceil(T a, T b)
-{
-    static_assert(is_same<T, index_t>::value || is_same<T, int>::value, "wrong type");
-    return (a + b - 1) / b;
-}
-template <class T>
-__host__ __device__ constexpr T max(T x, T y)
-{
-    return x > y ? x : y;
-}
-template <class T, class... Ts>
-__host__ __device__ constexpr T max(T x, Ts... xs)
-{
-    static_assert(sizeof...(xs) > 0, "not enough argument");
-    auto y = max(xs...);
-    static_assert(is_same<decltype(y), T>::value, "not the same type");
-    return x > y ? x : y;
-}
-template <class T>
-__host__ __device__ constexpr T min(T x, T y)
-{
-    return x < y ? x : y;
-}
-template <class T, class... Ts>
-__host__ __device__ constexpr T min(T x, Ts... xs)
-{
-    static_assert(sizeof...(xs) > 0, "not enough argument");
-    auto y = min(xs...);
-    static_assert(is_same<decltype(y), T>::value, "not the same type");
-    return x < y ? x : y;
-}
-// this is wrong
-// TODO: implement correct least common multiple, instead of calling max()
-template <class T, class... Ts>
-__host__ __device__ constexpr T least_common_multiple(T x, Ts... xs)
-{
-    return max(x, xs...);
-}
-} // namespace mod_conv
--- a/src/include/functional.hip.hpp
+++ b/src/include/functional.hip.hpp
@@ -19,18 +19,7 @@ struct swallow
    }
 };
-#if 0
+// Emulate if constexpr
-template<class F>
-__host__ __device__ constexpr auto unpacker(F f)
-{
-    return [=](auto xs_array){ f(xs...); };
-}
-#endif
-// Emulate compile time if statement for C++14
-//   Get the idea from
-//   "https://baptiste-wicht.com/posts/2015/07/simulate-static_if-with-c11c14.html"
-// TODO: use if constexpr, when C++17 is supported
 template <bool Predicate>
 struct static_if
 {
@@ -81,28 +70,3 @@ struct static_if<false>
        return Type{};
    }
 };
-template <index_t NLoop>
-struct static_const_reduce_n
-{
-    // signature of F: F(Number<I>)
-    template <class F, class Reduce>
-    __host__ __device__ constexpr auto operator()(F f, Reduce r) const
-    {
-        static_assert(NLoop > 1, "out-of-range");
-        constexpr auto a = f(Number<NLoop - 1>{});
-        auto b = static_const_reduce_n<NLoop - 1>{}(f, r); // TODO: cannot use constexpr here, weird
-        return r(a, b);
-    }
-};
-template <>
-struct static_const_reduce_n<1>
-{
-    template <class F, class Reduce>
-    __host__ __device__ constexpr auto operator()(F f, Reduce) const
-    {
-        return f(Number<0>{});
-    }
-};
--- a/src/include/functional2.hip.hpp
+++ b/src/include/functional2.hip.hpp
@@ -2,29 +2,16 @@
 #include "functional.hip.hpp"
 #include "Sequence.hip.hpp"
-#if 0
+template <class>
-template <index_t Iter, index_t Remaining, index_t Increment>
+struct static_for_impl;
-struct static_for_impl
-{
-    template <class F>
-    constexpr __host__ __device__ void operator()(F f) const
-    {
-        static_assert(Remaining % Increment == 0, "wrong! Remaining % Increment != 0");
-        static_assert(Increment <= Remaining, "will go out-of-range");
-        f(Number<Iter>{});
-        static_for_impl<Iter + Increment, Remaining - Increment, Increment>{}(f);
-    }
-};
-template <index_t Iter, index_t Increment>
+template <index_t... Is>
-struct static_for_impl<Iter, 0, Increment>
+struct static_for_impl<Sequence<Is...>>
 {
    template <class F>
-    constexpr __host__ __device__ void operator()(F) const
+    __host__ __device__ constexpr void operator()(F f) const
    {
-        // no work left, just return
+        swallow{(f(Number<Is>{}), 0)...};
-        return;
    }
 };
@@ -33,48 +20,42 @@ template <index_t NBegin, index_t NEnd, index_t Increment>
 struct static_for
 {
    template <class F>
-    constexpr __host__ __device__ void operator()(F f) const
+    __host__ __device__ constexpr void operator()(F f) const
    {
        static_assert(NBegin <= NEnd, "wrongs! should have NBegin <= NEnd");
        static_assert((NEnd - NBegin) % Increment == 0,
                      "Wrong! should satisfy (NEnd - NBegin) % Increment == 0");
-#if 0
+        static_for_impl<typename arithmetic_sequence_gen<NBegin, NEnd, Increment>::SeqType>{}(f);
-        static_if<(NBegin < NEnd)>{}(
-            [&](auto fwd) { static_for_impl<NBegin, NEnd - NBegin, fwd(Increment)>{}(f); });
-#else
-        static_for_impl<NBegin, NEnd - NBegin, Increment>{}(f);
-#endif
    }
 };
-#else
-template <class>
-struct static_for_impl;
-template <index_t... Is>
+template <class Seq, class Reduce>
-struct static_for_impl<Sequence<Is...>>
+struct lambda_accumulate_on_sequence
 {
-    template <class F>
+    const Reduce& f;
-    __host__ __device__ constexpr void operator()(F f) const
+    index_t& result;
+    __host__ __device__ constexpr lambda_accumulate_on_sequence(const Reduce& f_, index_t& result_)
+        : f(f_), result(result_)
    {
-        swallow{(f(Number<Is>{}), 0)...};
+    }
+    template <class IDim>
+    __host__ __device__ constexpr index_t operator()(IDim) const
+    {
+        return result = f(result, Seq::Get(IDim{}));
    }
 };
-// F signature: F(Number<Iter>)
+template <class Seq, class Reduce, index_t Init>
-template <index_t NBegin, index_t NEnd, index_t Increment>
+__host__ __device__ constexpr index_t
-struct static_for
+accumulate_on_sequence(Seq, Reduce f, Number<Init> /*initial_value*/)
 {
-    template <class F>
+    index_t result = Init;
-    __host__ __device__ constexpr void operator()(F f) const
-    {
-        static_assert(NBegin <= NEnd, "wrongs! should have NBegin <= NEnd");
-        static_assert((NEnd - NBegin) % Increment == 0,
+    static_for<0, Seq::mSize, 1>{}(lambda_accumulate_on_sequence<Seq, Reduce>(f, result));
-                      "Wrong! should satisfy (NEnd - NBegin) % Increment == 0");
-        static_for_impl<typename arithmetic_sequence_gen<NBegin, NEnd, Increment>::SeqType>{}(f);
+    return result;
-    }
+}
-};
-#endif
--- a/src/include/functional3.hip.hpp
+++ b/src/include/functional3.hip.hpp
@@ -11,7 +11,7 @@ struct static_ford_impl
    // F signature: F(Sequence<...> multi_id)
    // CurrentMultiIndex: Sequence<...>
    template <class F, class CurrentMultiIndex>
-    __host__ __device__ void operator()(F f, CurrentMultiIndex) const
+    __host__ __device__ constexpr void operator()(F f, CurrentMultiIndex) const
    {
        static_assert(RemainLengths::GetSize() > 0, "wrong! should not get here");
@@ -28,7 +28,7 @@ struct static_ford_impl<Sequence<>>
    // F signature: F(Sequence<...> multi_id)
    // CurrentMultiIndex: Sequence<...>
    template <class F, class CurrentMultiIndex>
-    __host__ __device__ void operator()(F f, CurrentMultiIndex) const
+    __host__ __device__ constexpr void operator()(F f, CurrentMultiIndex) const
    {
        f(CurrentMultiIndex{});
    }
@@ -40,7 +40,7 @@ struct static_ford
 {
    // F signature: F(Sequence<...> multi_id)
    template <class F>
-    __host__ __device__ void operator()(F f) const
+    __host__ __device__ constexpr void operator()(F f) const
    {
        static_assert(Lengths::GetSize() > 0, "wrong! Lengths is empty");
@@ -55,7 +55,7 @@ struct ford_impl
    // CurrentMultiIndex: Array<...>
    // RemainLengths: Sequence<...>
    template <class F, class CurrentMultiIndex, class RemainLengths>
-    __host__ __device__ void
+    __host__ __device__ constexpr void
    operator()(F f, CurrentMultiIndex current_multi_id, RemainLengths) const
    {
        static_assert(RemainLengths::GetSize() == RemainDim, "wrong!");
@@ -77,7 +77,7 @@ struct ford_impl<1>
    // CurrentMultiIndex: Array<...>
    // RemainLengths: Sequence<...>
    template <class F, class CurrentMultiIndex, class RemainLengths>
-    __host__ __device__ void
+    __host__ __device__ constexpr void
    operator()(F f, CurrentMultiIndex current_multi_id, RemainLengths) const
    {
        static_assert(RemainLengths::GetSize() == 1, "wrong!");
@@ -97,7 +97,7 @@ struct ford
 {
    // F signature: F(Array<...> multi_id)
    template <class F>
-    __host__ __device__ void operator()(F f) const
+    __host__ __device__ constexpr void operator()(F f) const
    {
        constexpr index_t first_length = Lengths{}.Front();

--- a/src/include/gridwise_convolution_implicit_gemm_v1r1_chwn_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r1_chwn_cyxk_khwn.hip.hpp
@@ -103,7 +103,7 @@ struct GridwiseConvolutionImplicitGemm_v1r1_chwn_cyxk_khwn
        // tensor view of blockwise input and weight in LDS
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockCopyDataPerRead_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockCopyDataPerRead_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -119,11 +119,11 @@ struct GridwiseConvolutionImplicitGemm_v1r1_chwn_cyxk_khwn
        constexpr auto wei_cyx_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock * Y * X, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        constexpr auto wei_c_y_x_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, Y, X, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(

--- a/src/include/gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hip.hpp
@@ -104,7 +104,7 @@ struct GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockCopyDataPerRead_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockCopyDataPerRead_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -120,7 +120,7 @@ struct GridwiseConvolutionImplicitGemm_v1r2_chwn_cyxk_khwn
        constexpr auto wei_c_x_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, X, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(

--- a/src/include/gridwise_convolution_implicit_gemm_v1r2_nchw_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r2_nchw_cyxk_khwn.hip.hpp
@@ -108,7 +108,7 @@ struct GridwiseConvolutionImplicitGemm_v1r2_nchw_cyxk_khwn
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockReorderDataPerWrite_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockReorderDataPerWrite_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);

--- a/src/include/gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn.hip.hpp
@@ -99,7 +99,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockCopyDataPerRead_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockCopyDataPerRead_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -115,7 +115,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(

--- a/src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_chwn_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_chwn_cyxk_khwn.hip.hpp
@@ -104,7 +104,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockCopyDataPerRead_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockCopyDataPerRead_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -120,7 +120,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_chwn_cyxk_khwn
        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(

--- a/src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_khwn.hip.hpp
@@ -106,7 +106,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockReorderDataPerWrite_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockReorderDataPerWrite_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -122,7 +122,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_khwn
        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(

--- a/src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw.hip.hpp
@@ -105,7 +105,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockReorderDataPerWrite_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockReorderDataPerWrite_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -121,7 +121,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_lds_double_buffer_nchw_cyxk_nkhw
        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(

--- a/src/include/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_khwn.hip.hpp
@@ -104,7 +104,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_khwn
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockReorderDataPerWrite_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockReorderDataPerWrite_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -120,7 +120,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_khwn
        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor(

--- a/src/include/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw.hip.hpp
@@ -103,7 +103,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
        // LDS tensor view
        //   be careful of alignment
-        constexpr index_t max_align = mod_conv::max(InBlockReorderDataPerWrite_N,
+        constexpr index_t max_align = mod_conv::lcm(InBlockReorderDataPerWrite_N,
                                                    WeiBlockCopyDataPerRead_K,
                                                    GemmDataPerReadA,
                                                    GemmDataPerReadB);
@@ -119,7 +119,7 @@ struct GridwiseConvolutionImplicitGemm_v1r3_nchw_cyxk_nkhw
        constexpr auto wei_c_k_block_desc = make_ConstantTensorDescriptor_aligned(
            Sequence<CPerBlock, KPerBlock>{},
-            Number<mod_conv::max(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
+            Number<mod_conv::lcm(WeiBlockCopyDataPerRead_K, GemmDataPerReadA)>{});
        // tensor view of threadwise output in register
        constexpr auto out_k_h_w_n_thread_desc = make_ConstantTensorDescriptor_packed(

--- a/src/include/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hip.hpp
+++ b/src/include/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hip.hpp
@@ -181,7 +181,7 @@ struct GridwiseConvolutionImplicitGemm_v2_chwn_cyxk_khwn
        // LDS: be careful of alignment
        constexpr index_t max_align =
-            mod_conv::max(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead);
+            mod_conv::lcm(index_t(4), InBlockCopyDataPerRead, WeiBlockCopyDataPerRead);
        constexpr index_t in_block_space = in_cb_block_desc.GetElementSpace(Number<max_align>{});