clean up

composable_kernel/include/kernel_algorithm/dummy_dynamic_transform.hpp

@@ -696,7 +696,7 @@ struct DummyDynamicTransform
         auto in_gemmk_gemmn_coord =
             make_dynamic_tensor_coordinate_v2(in_gemmk_gemmn_global_desc, idx);
 
-        constexpr auto in_gemmk_gemmn_coord_step = make_dynamic_tensor_coordinate_step_v2(
+        const auto in_gemmk_gemmn_coord_step = make_dynamic_tensor_coordinate_step_v2(
             in_gemmk_gemmn_global_desc, MultiIndex<2>{{1, 0}});
 
         for(index_t iter = 0; iter < niter; ++iter)

composable_kernel/include/utility/array.hpp

@@ -16,23 +16,6 @@ struct Array
     // TODO: implement empty Array
     TData mData[NSize + 1] = {0};
 
-#if 0
-    __host__ __device__ explicit constexpr Array(TData x)
-        : mData{x}
-    {}
-
-    __host__ __device__ explicit constexpr Array()
-        : Array(TData{0})
-    {}
-
-    template <typename... Xs>
-    __host__ __device__ constexpr Array(Xs... xs)
-        : mData{static_cast<TData>(xs)...}
-    {
-        static_assert(sizeof...(Xs) == NSize, "wrong! size");
-    }
-#endif
-
     __host__ __device__ static constexpr index_t Size() { return NSize; }
 
     // TODO: remove

composable_kernel/include/utility/array_helper.hpp

@@ -115,104 +115,6 @@ struct lambda_array_math
     }
 };
 
-#if 0
-// Array = Array + Array
-template <typename TData, index_t NSize>
-__host__ __device__ constexpr auto operator+(Array<TData, NSize> a, Array<TData, NSize> b)
-{
-    Array<TData, NSize> result;
-
-    auto f = math::plus<index_t>{};
-
-    static_for<0, NSize, 1>{}(
-        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
-            f, a, b, result));
-
-    return result;
-}
-
-// Array = Array - Array
-template <typename TData, index_t NSize>
-__host__ __device__ constexpr auto operator-(Array<TData, NSize> a, Array<TData, NSize> b)
-{
-    Array<TData, NSize> result;
-
-    auto f = math::minus<index_t>{};
-
-    static_for<0, NSize, 1>{}(
-        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
-            f, a, b, result));
-
-    return result;
-}
-
-// Array += Array
-template <typename TData, index_t NSize>
-__host__ __device__ constexpr auto operator+=(Array<TData, NSize>& a, Array<TData, NSize> b)
-{
-    a = a + b;
-    return a;
-}
-
-// Array -= Array
-template <typename TData, index_t NSize>
-__host__ __device__ constexpr auto operator-=(Array<TData, NSize>& a, Array<TData, NSize> b)
-{
-    a = a - b;
-    return a;
-}
-// Array = Array + Sequence
-template <typename 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;
-
-    auto f = math::plus<index_t>{};
-
-    static_for<0, NSize, 1>{}(
-        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
-            f, a, b, result));
-
-    return result;
-}
-
-// Array = Array - Sequence
-template <typename 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;
-
-    auto f = math::minus<index_t>{};
-
-    static_for<0, NSize, 1>{}(
-        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
-            f, a, b, result));
-
-    return result;
-}
-
-// Array = Array * Sequence
-template <typename 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;
-
-    auto f = math::multiplies<index_t>{};
-
-    static_for<0, NSize, 1>{}(
-        lambda_array_math<decltype(f), decltype(a), decltype(b), decltype(result)>(
-            f, a, b, result));
-
-    return result;
-}
-#endif
-
 // Array = Sequence - Array
 template <typename TData, index_t NSize, index_t... Is>
 __host__ __device__ constexpr auto operator-(Sequence<Is...> a, Array<TData, NSize> b)

driver/src/conv_driver.cpp

@@ -560,7 +560,7 @@ int main(int argc, char* argv[])
                                                          LeftPads{},
                                                          RightPads{},
                                                          nrepeat);
-#elif 1
+#elif 0
     device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc,
                                                          in_nchw,
                                                          wei_kcyx_desc,