cmakelist changes to exclude navi cards for gemv splitk & merge changes from dev

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp

@@ -471,7 +471,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
             auto launch_kernel = [&](auto has_main_k_block_loop) {
                 constexpr bool has_main_loop = has_main_k_block_loop.value;
 
-                const auto kernel = kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle<
+                const auto kernel = kernel_grouped_conv_multiple_d_wmma_cshuffle<
                     GridwiseOp,
                     ADataType,
                     BDataType,

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp

@@ -43,7 +43,13 @@ struct ComputePtrOffsetOfStridedBatch
         return ds_offset;
     }
 
-    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    [[maybe_unused]] __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    // alias for kernels without multiple D
+    [[maybe_unused]] __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
     {
         return g_idx * static_cast<long_index_t>(BatchStrideE_);
     }
@@ -52,6 +58,7 @@ struct ComputePtrOffsetOfStridedBatch
     index_t BatchStrideB_;
     Array<ck::index_t, NumDTensor> BatchStrideDs_;
     index_t BatchStrideE_;
+    index_t& BatchStrideC_ = BatchStrideE_; // alias for kernels without multiple D
 };
 
 } // namespace device

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -113,7 +113,6 @@ struct PassThrough
     }
 #endif
 
-#if defined CK_ENABLE_FP8
     template <>
     __host__ __device__ void operator()<f8_t, f8_t>(f8_t& y, const f8_t& x) const
     {
@@ -143,9 +142,7 @@ struct PassThrough
     {
         y = type_convert<f8_t>(x);
     }
-#endif
 
-#if defined CK_ENABLE_BF8
     template <>
     __host__ __device__ void operator()<bf8_t, bf8_t>(bf8_t& y, const bf8_t& x) const
     {
@@ -175,7 +172,6 @@ struct PassThrough
     {
         y = ck::type_convert<bf8_t>(x);
     }
-#endif
 };
 
 struct UnaryConvert
@@ -204,7 +200,6 @@ struct ConvertBF16RTN
     }
 };
 
-#if defined CK_ENABLE_FP8
 struct ConvertF8SR
 {
     // convert to fp8 using stochastic rounding (SR)
@@ -221,7 +216,6 @@ struct ConvertF8SR
         y = f8_convert_sr<Y>(x);
     }
 };
-#endif
 
 struct Scale
 {

include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp

@@ -428,7 +428,7 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
             [&](auto i) {
                 using ALayout = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
 
-                return MakeAGridDescriptor_M_N<ALayout, GemmSpec>(MRaws[i], KRaws[i], AsStride[i]);
+                return MakeAGridDescriptor_M_K<ALayout, GemmSpec>(MRaws[i], KRaws[i], AsStride[i]);
             },
             Number<NumATensor>{});
     }
@@ -656,7 +656,8 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ComputeDataType,
+            ComputeDataType, // ComputeDataType for A
+            ComputeDataType, // ComputeDataType for B
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp

@@ -36,7 +36,7 @@ __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle(
+        kernel_grouped_conv_multiple_d_wmma_cshuffle(
             const ADataType* __restrict__ p_a_grid,
             const BDataType* __restrict__ p_b_grid,
             DsPointer p_ds_grid,
@@ -452,11 +452,11 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
     }
 
     // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    // CheckValidity for kernels without multi D
     template <typename Block2CTileMap>
     __host__ __device__ static constexpr bool
     CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
                   const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
-                  const DsGridDesc_M_N& ds_grid_desc_m_n,
                   const EGridDesc_M_N& e_grid_desc_m_n,
                   const Block2CTileMap& block_2_ctile_map)
     {
@@ -471,18 +471,6 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
         const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
         const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
 
-        bool valid = true;
-
-        static_for<0, NumDTensor, 1>{}([&](auto i) {
-            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
-                              N == ds_grid_desc_m_n[i].GetLength(I1));
-        });
-
-        if(!valid)
-        {
-            return false;
-        }
-
         if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
              K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
              K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
@@ -517,6 +505,31 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
         return true;
     }
 
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const DsGridDesc_M_N& ds_grid_desc_m_n,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto M = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N = b_grid_desc_k0_n_k1.GetLength(I1);
+        bool valid   = true;
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        return CheckValidity(
+            a_grid_desc_k0_m_k1, b_grid_desc_k0_n_k1, e_grid_desc_m_n, block_2_ctile_map);
+    }
+
     __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
     {
         const index_t num_loop = K / (K0PerBlock * K1);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp

@@ -22,13 +22,19 @@ namespace ck {
 template <typename GridwiseGemm,
           bool HasMainKBlockLoop,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
-          typename Block2CTileMap>
+          typename Block2CTileMap,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
 __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
         kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
-                                             const Block2CTileMap& b2c_map)
+                                             const Block2CTileMap& b2c_map,
+                                             const AElementwiseOperation a_element_op,
+                                             const BElementwiseOperation b_element_op,
+                                             const CElementwiseOperation c_element_op)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
     defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
@@ -37,10 +43,13 @@ __global__ void
     __shared__ uint8_t p_shared[shared_size];
 
     GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
-        karg, static_cast<void*>(p_shared), b2c_map);
+        karg, static_cast<void*>(p_shared), b2c_map, a_element_op, b_element_op, c_element_op);
 #else
     ignore = karg;
     ignore = b2c_map;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
 #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
@@ -577,7 +586,10 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
               typename Block2CTileMap>
     __device__ static void Run(const Argument& karg,
                                void* __restrict__ p_shared_block,
-                               const Block2CTileMap& block_2_ctile_map)
+                               const Block2CTileMap& block_2_ctile_map,
+                               const AElementwiseOperation a_element_op = AElementwiseOperation{},
+                               const BElementwiseOperation b_element_op = BElementwiseOperation{},
+                               const CElementwiseOperation c_element_op = CElementwiseOperation{})
     {
         const FloatA* p_a_grid           = karg.p_a_grid;
         const FloatB* p_b_grid           = karg.p_b_grid;
@@ -590,9 +602,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
 
         const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
             MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
-        const AElementwiseOperation a_element_op = AElementwiseOperation{};
-        const BElementwiseOperation b_element_op = BElementwiseOperation{};
-        const CElementwiseOperation c_element_op = CElementwiseOperation{};
 
         const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
@@ -761,8 +770,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ComputeType,
-            ComputeType,
+            ComputeType, // ComputeType A
+            ComputeType, // ComputeType B
             FloatAcc,
             decltype(a_k0_m_k1_block_desc),
             decltype(b_k0_n_k1_block_desc),

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp

@@ -9,6 +9,7 @@
 #include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor/static_tensor.hpp"
+#include "ck/utility/is_detected.hpp"
 
 namespace ck {
 
@@ -211,10 +212,44 @@ struct ThreadwiseTensorSliceTransfer_v3r1
             auto src_vector_container = src_vector_type{
                 src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
 
+            using dst_vector_type = vector_type_maker_t<DstData, SrcScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+            dst_vector_type op_r_v;
+
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack4_invocable_t, decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack2_invocable_t, decltype(src_element_op_)>::value)
+                {
+                    if constexpr(decltype(src_element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                return 1;
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            using src_elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+            using dst_elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto idx) {
+                // apply the src elementwise op and convert to DstData under the hood if needed
+                src_element_op_(op_r_v.template AsType<dst_elem_op_vec_t>()(idx),
+                                src_vector_container.template AsType<src_elem_op_vec_t>()[idx]);
+            });
+
             // copy data from src_vector_container into src_thread_scratch_
             src_thread_scratch_tuple_(thread_scratch_id)
-                .template SetAsType<src_vector_t>(
-                    src_data_idx_seq, src_vector_container.template AsType<src_vector_t>()[I0]);
+                .template SetAsType<dst_vector_t>(src_data_idx_seq,
+                                                  op_r_v.template AsType<dst_vector_t>()[I0]);
 
             constexpr auto move_on_dim = [&]() constexpr
             {
@@ -267,19 +302,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1
     {
 #if !CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
         static_ford<SliceLengths>{}([&](auto idx) {
-            // convert from SrcData to DstData here
-            dst_thread_scratch_(idx) =
-                type_convert<DstData>(src_thread_scratch_tuple_[thread_scratch_id][idx]);
+            dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
         });
 #else
         // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
         // TODO make this logic more generic for more sub-dword datatype
         if constexpr(SrcVectorDim != DstVectorDim &&
-                     ((is_same<half_t, remove_cvref_t<SrcData>>::value &&
-                       is_same<half_t, remove_cvref_t<DstData>>::value &&
+                     ((is_same<half_t, remove_cvref_t<DstData>>::value &&
                        SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
-                      (is_same<int8_t, remove_cvref_t<SrcData>>::value &&
-                       is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                      (is_same<int8_t, remove_cvref_t<DstData>>::value &&
                        SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
         {
             // each transpose does
@@ -313,7 +344,7 @@ struct ThreadwiseTensorSliceTransfer_v3r1
                 constexpr auto data_idx_seq = generate_sequence_v2(
                     [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
 
-                using src_vector_t = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+                using src_vector_t = vector_type_maker_t<DstData, SrcScalarPerVector>;
                 using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
 
                 // get DstScalarPerVector # of read-only references to src vectors from
@@ -336,17 +367,16 @@ struct ThreadwiseTensorSliceTransfer_v3r1
                     Number<num_dst_vector>{});
 
                 // do data transpose
-                transpose_vectors<SrcData, DstScalarPerVector, SrcScalarPerVector>{}(
+                transpose_vectors<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
                     src_vector_refs, dst_vector_refs);
             });
         }
-
-        static_ford<SliceLengths>{}([&](auto idx) {
-            // apply the src elementwise op and convert to DstData under the hood if needed
-            DstData dst_v;
-            src_element_op_(dst_v, src_thread_scratch_tuple_[thread_scratch_id][idx]);
-            dst_thread_scratch_(idx) = dst_v;
-        });
+        else
+        {
+            static_ford<SliceLengths>{}([&](auto idx) {
+                dst_thread_scratch_(idx) = src_thread_scratch_tuple_[thread_scratch_id][idx];
+            });
+        }
 #endif
     }
 
@@ -761,11 +791,12 @@ struct ThreadwiseTensorSliceTransfer_v3r1
     static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
     static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
 
-    using SrcThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
-                                                             SrcData,
-                                                             SrcScalarPerVector,
-                                                             decltype(src_thread_scratch_desc_),
-                                                             true>;
+    using SrcThreadScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        DstData, // apply data_convert with SrcThreadScratch
+                                        SrcScalarPerVector,
+                                        decltype(src_thread_scratch_desc_),
+                                        true>;
 
     using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
                                                              DstData,

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp

@@ -132,9 +132,6 @@ struct ThreadwiseTensorSliceTransfer_v7r2
             Number<num>{});
     }
 
-    template <typename T>
-    using has_vec_len = decltype(std::declval<T&>().vec_len);
-
     // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
     // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
     template <typename SrcBuffers,
@@ -159,94 +156,63 @@ struct ThreadwiseTensorSliceTransfer_v7r2
                                                            is_src_valid);
             });
 
-            if constexpr(is_detected<has_vec_len, decltype(element_op_)>::value)
-            {
-                constexpr auto elem_op_vec_len = decltype(element_op_)::vec_len;
-
-                static_assert(is_same<remove_cvref_t<decltype(elem_op_vec_len)>, index_t>::value,
-                              "vec_len in element_op_ type is not index_t");
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack4_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack2_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                return 1;
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            // apply pointwise function
+            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+
+                        return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+                        return dst_vectors(iDst).template AsType<elem_op_vec_t>()(i);
+                    },
+                    Number<nDst>{});
 
-                static_assert(elem_op_vec_len == 1 || elem_op_vec_len == 2 ||
-                                  elem_op_vec_len == 4 || elem_op_vec_len == 8,
-                              "vec_len in element_op_ must be 1, 2, 4, 8");
-
-                static_assert(SrcScalarPerVector % elem_op_vec_len == 0,
-                              "vec_len in element_op_ cannot be divided by SrcScalarPerVector!");
-
-                // apply pointwise function
-                static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
-                    // get reference to src data
-                    const auto src_data_refs = generate_tie(
-                        // return type should be lvalue
-                        [&](auto iSrc) -> const auto& {
-                            using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
-
-                            using elem_op_vec_t =
-                                typename vector_type<SrcData, elem_op_vec_len>::type;
-
-                            return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
-                        },
-                        Number<nSrc>{});
-
-                    // get reference to dst data
-                    auto dst_data_refs = generate_tie(
-                        // return type should be lvalue
-                        [&](auto iDst) -> auto& {
-                            using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
-
-                            using elem_op_vec_t =
-                                typename vector_type<DstData, elem_op_vec_len>::type;
-
-                            return dst_vectors(iDst).template AsType<elem_op_vec_t>()(i);
-                        },
-                        Number<nDst>{});
-
-                    // apply pointwise function
-                    // pointwise function signature:
-                    // element_op_(dst_data_refs[I0],
-                    //             dst_data_refs[I1],
-                    //             ...,
-                    //             src_data_refs[I0],
-                    //             src_data_refs[I1],
-                    //             ...)
-                    unpack2(element_op_, dst_data_refs, src_data_refs);
-                });
-            }
-            else
-            {
                 // apply pointwise function
-                static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
-                    // get reference to src data
-                    const auto src_data_refs = generate_tie(
-                        // return type should be lvalue
-                        [&](auto iSrc) -> const auto& {
-                            using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
-
-                            return src_vectors[iSrc].template AsType<SrcData>()[i];
-                        },
-                        Number<nSrc>{});
-
-                    // get reference to dst data
-                    auto dst_data_refs = generate_tie(
-                        // return type should be lvalue
-                        [&](auto iDst) -> auto& {
-                            using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
-
-                            return dst_vectors(iDst).template AsType<DstData>()(i);
-                        },
-                        Number<nDst>{});
-
-                    // apply pointwise function
-                    // pointwise function signature:
-                    // element_op_(dst_data_refs[I0],
-                    //             dst_data_refs[I1],
-                    //             ...,
-                    //             src_data_refs[I0],
-                    //             src_data_refs[I1],
-                    //             ...)
-                    unpack2(element_op_, dst_data_refs, src_data_refs);
-                });
-            }
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
 
             dst_vectors_tuple_(iAccess) = dst_vectors;
 

include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp

@@ -462,7 +462,6 @@ struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
     }
 };
 
-#if defined CK_ENABLE_FP8
 template <>
 struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8f8>
 {
@@ -506,9 +505,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8f8>
         intrin_mfma_f32_16x16x32f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
     }
 };
-#endif
 
-#if defined CK_ENABLE_BF8
 template <>
 struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8bf8>
 {
@@ -552,9 +549,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8bf8>
         intrin_mfma_f32_16x16x32bf8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
     }
 };
-#endif
 
-#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
 template <>
 struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8bf8>
 {
@@ -598,9 +593,7 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8bf8>
         intrin_mfma_f32_16x16x32f8bf8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
     }
 };
-#endif
 
-#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
 template <>
 struct mfma_type<MfmaInstr::mfma_f32_32x32x16bf8f8>
 {
@@ -644,7 +637,6 @@ struct mfma_type<MfmaInstr::mfma_f32_16x16x32bf8f8>
         intrin_mfma_f32_16x16x32bf8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
     }
 };
-#endif
 
 template <typename base_type,
           index_t MPerXdlops,
@@ -792,7 +784,6 @@ struct MfmaSelector
     }
 #endif
 
-#if defined CK_ENABLE_FP8
     template <>
     static constexpr auto GetMfma<f8_t, 32, 32>()
     {
@@ -804,9 +795,7 @@ struct MfmaSelector
     {
         return MfmaInstr::mfma_f32_16x16x32f8f8;
     }
-#endif
 
-#if defined CK_ENABLE_BF8
     template <>
     static constexpr auto GetMfma<bf8_t, 32, 32>()
     {
@@ -818,9 +807,7 @@ struct MfmaSelector
     {
         return MfmaInstr::mfma_f32_16x16x32bf8bf8;
     }
-#endif
 
-#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
     template <>
     static constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
     {
@@ -832,9 +819,7 @@ struct MfmaSelector
     {
         return MfmaInstr::mfma_f32_16x16x32f8bf8;
     }
-#endif
 
-#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
     template <>
     static constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
     {
@@ -846,7 +831,6 @@ struct MfmaSelector
     {
         return MfmaInstr::mfma_f32_16x16x32bf8f8;
     }
-#endif
 
     static constexpr auto selected_mfma =
         mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops, additional_type>()>{};
@@ -1051,18 +1035,10 @@ struct XdlopsGemm
         static_assert(
             is_same<base_type, double>::value || is_same<base_type, float>::value ||
                 is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
-                is_same<base_type, int8_t>::value
-#if defined CK_ENABLE_FP8
-                || is_same<base_type, f8_t>::value
-#endif
-#if defined CK_ENABLE_BF8
-                || is_same<base_type, bf8_t>::value
-#endif
-#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
-                || (is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value) ||
-                (is_same<base_type, bf8_t>::value && is_same<additional_type, f8_t>::value)
-#endif
-                ,
+                is_same<base_type, int8_t>::value || is_same<base_type, f8_t>::value ||
+                is_same<base_type, bf8_t>::value ||
+                (is_same<base_type, f8_t>::value && is_same<additional_type, bf8_t>::value) ||
+                (is_same<base_type, bf8_t>::value && is_same<additional_type, f8_t>::value),
             "base base_type must be double, float, half, bfloat16, int8_t, f8_t or bf8_t!");
 
         static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {

include/ck/utility/amd_xdlops.hpp

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
-#ifndef CK_AMD_XDLOPS_HPP
-#define CK_AMD_XDLOPS_HPP
-
-#include "data_type.hpp"
+#pragma once
 
 namespace ck {
 
@@ -355,7 +352,6 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
     }
 };
 
-#if defined CK_ENABLE_FP8
 template <index_t MPerWave, index_t NPerWave>
 struct intrin_mfma_f32_32x32x16f8f8;
 
@@ -418,9 +414,7 @@ struct intrin_mfma_f32_16x16x32f8f8<16, 16>
 #endif
     }
 };
-#endif
 
-#if defined CK_ENABLE_BF8
 template <index_t MPerWave, index_t NPerWave>
 struct intrin_mfma_f32_32x32x16bf8bf8;
 
@@ -483,9 +477,7 @@ struct intrin_mfma_f32_16x16x32bf8bf8<16, 16>
 #endif
     }
 };
-#endif
 
-#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
 template <index_t MPerWave, index_t NPerWave>
 struct intrin_mfma_f32_32x32x16f8bf8;
 
@@ -548,9 +540,7 @@ struct intrin_mfma_f32_16x16x32f8bf8<16, 16>
 #endif
     }
 };
-#endif
 
-#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
 template <index_t MPerWave, index_t NPerWave>
 struct intrin_mfma_f32_32x32x16bf8f8;
 
@@ -613,6 +603,5 @@ struct intrin_mfma_f32_16x16x32bf8f8<16, 16>
 #endif
     }
 };
-#endif
+
 } // namespace ck
-#endif

include/ck/utility/data_type.hpp

@@ -9,15 +9,9 @@ namespace ck {
 
 using bhalf_t = ushort;
 using half_t  = _Float16;
-#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
-using int4_t = _BitInt(4);
-#endif
-#if defined CK_ENABLE_FP8
-using f8_t = _BitInt(8);
-#endif
-#if defined CK_ENABLE_BF8
-using bf8_t = unsigned _BitInt(8);
-#endif
+using int4_t  = _BitInt(4);
+using f8_t    = _BitInt(8);
+using bf8_t   = unsigned _BitInt(8);
 
 // vector_type
 template <typename T, index_t N>
@@ -148,23 +142,19 @@ struct scalar_type<int4_t>
 };
 #endif
 
-#if defined CK_ENABLE_FP8
 template <>
 struct scalar_type<f8_t>
 {
     using type                           = f8_t;
     static constexpr index_t vector_size = 1;
 };
-#endif
 
-#if defined CK_ENABLE_BF8
 template <>
 struct scalar_type<bf8_t>
 {
     using type                           = bf8_t;
     static constexpr index_t vector_size = 1;
 };
-#endif
 
 template <typename T>
 struct vector_type<T, 1>
@@ -968,24 +958,20 @@ using int8x32_t = typename vector_type<int8_t, 32>::type;
 using int8x64_t = typename vector_type<int8_t, 64>::type;
 
 // f8
-#if defined CK_ENABLE_FP8
 using f8x2_t  = typename vector_type<f8_t, 2>::type;
 using f8x4_t  = typename vector_type<f8_t, 4>::type;
 using f8x8_t  = typename vector_type<f8_t, 8>::type;
 using f8x16_t = typename vector_type<f8_t, 16>::type;
 using f8x32_t = typename vector_type<f8_t, 32>::type;
 using f8x64_t = typename vector_type<f8_t, 64>::type;
-#endif
 
 // bf8
-#if defined CK_ENABLE_BF8
 using bf8x2_t  = typename vector_type<bf8_t, 2>::type;
 using bf8x4_t  = typename vector_type<bf8_t, 4>::type;
 using bf8x8_t  = typename vector_type<bf8_t, 8>::type;
 using bf8x16_t = typename vector_type<bf8_t, 16>::type;
 using bf8x32_t = typename vector_type<bf8_t, 32>::type;
 using bf8x64_t = typename vector_type<bf8_t, 64>::type;
-#endif
 
 template <typename T>
 struct NumericLimits
@@ -1033,7 +1019,6 @@ struct NumericLimits<int4_t>
 };
 #endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
 
-#if defined CK_ENABLE_FP8
 template <>
 struct NumericLimits<f8_t>
 {
@@ -1056,9 +1041,7 @@ struct NumericLimits<f8_t>
 
     __host__ __device__ static constexpr f8_t QuietNaN() { return f8_t(binary_qnan); }
 };
-#endif
 
-#if defined CK_ENABLE_BF8
 template <>
 struct NumericLimits<bf8_t>
 {
@@ -1081,7 +1064,6 @@ struct NumericLimits<bf8_t>
 
     __host__ __device__ static constexpr bf8_t QuietNaN() { return bf8_t(binary_qnan); }
 };
-#endif
 
 template <typename T>
 struct NumericUtils
@@ -1120,22 +1102,18 @@ struct NumericUtils<half_t>
     using bitwise_type                  = uint16_t;
 };
 
-#if defined CK_ENABLE_FP8
 template <>
 struct NumericUtils<f8_t>
 {
     static constexpr int exp  = 4;
     static constexpr int mant = 3;
 };
-#endif
 
-#if defined CK_ENABLE_BF8
 template <>
 struct NumericUtils<bf8_t>
 {
     static constexpr int exp  = 5;
     static constexpr int mant = 2;
 };
-#endif
-
+//
 } // namespace ck

include/ck/utility/f8_utils.hpp

@@ -6,8 +6,6 @@
 #include "ck/utility/data_type.hpp"
 
 // these conversions are disabled if native conversions available
-#if !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__)
-#if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8
 namespace ck {
 
 // fp8 rounding modes
@@ -244,5 +242,3 @@ __host__ __device__ Y cast_from_f8(X x)
 }
 
 } // namespace ck::utils
-#endif // #if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8
-#endif // #if !defined(__gfx940__) && !defined(__gfx941__) && !defined(__gfx942__)

include/ck/utility/is_detected.hpp

@@ -31,4 +31,13 @@ struct nonesuch
 template <template <class...> class Op, class... Args>
 using is_detected = typename detail::detector<nonesuch, void, Op, Args...>::value_t;
 
+template <typename T>
+using is_pack2_invocable_t = decltype(std::declval<T&>().is_pack2_invocable);
+
+template <typename T>
+using is_pack4_invocable_t = decltype(std::declval<T&>().is_pack4_invocable);
+
+template <typename T>
+using is_pack8_invocable_t = decltype(std::declval<T&>().is_pack8_invocable);
+
 } // namespace ck

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -20,12 +20,8 @@ using F16  = ck::half_t;
 using BF16 = ck::bhalf_t;
 using I8   = int8_t;
 using I32  = int32_t;
-#if defined CK_ENABLE_FP8
-using F8 = ck::f8_t;
-#endif
-#if defined CK_ENABLE_BF8
-using BF8 = ck::bf8_t;
-#endif
+using F8   = ck::f8_t;
+using BF8  = ck::bf8_t;
 
 using Empty_Tuple = ck::Tuple<>;
 

library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp

@@ -240,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
         if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, NWC> && is_same_v<WeiLayout, KXC> &&
                      is_same_v<OutLayout, NWK>)
         {
+#ifdef CK_ENABLE_FP32
             if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
                          is_same_v<OutDataType, float>)
             {
                 add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
             }
+#endif
 #ifdef CK_ENABLE_FP16
             if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
                          is_same_v<OutDataType, half_t>)
@@ -267,17 +269,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             }
 #endif
         }
-        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
-                          is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
+                     is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
         {
+#ifdef CK_ENABLE_FP32
             if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
                          is_same_v<OutDataType, float>)
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
-#ifdef DL_KERNELS
-                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+            }
 #endif
+#if defined(DL_KERNELS) && defined(CK_ENABLE_FP32)
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
             }
+#endif
 #ifdef CK_ENABLE_FP16
             if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
                          is_same_v<OutDataType, half_t>)
@@ -306,14 +314,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             }
 #endif
         }
-        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
-                          is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
+        if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
+                     is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
         {
+#ifdef CK_ENABLE_FP32
             if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
                          is_same_v<OutDataType, float>)
             {
                 add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
             }
+#endif
 #ifdef CK_ENABLE_FP16
             if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
                          is_same_v<OutDataType, half_t>)

library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp

@@ -98,30 +98,31 @@ struct DeviceOperationInstanceFactory<
         if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
                      is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
         {
+#ifdef CK_ENABLE_FP32
             if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
                          is_same_v<OutDataType, float>)
             {
                 add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
             }
+#endif
 #ifdef CK_ENABLE_FP16
-            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
-                              is_same_v<OutDataType, half_t>)
+            if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                         is_same_v<OutDataType, half_t>)
             {
                 add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
                 add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
             }
 #endif
 #ifdef CK_ENABLE_BF16
-            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
-                              is_same_v<WeiDataType, ck::bhalf_t> &&
-                              is_same_v<OutDataType, ck::bhalf_t>)
+            if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                         is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
             {
                 add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
             }
 #endif
 #ifdef CK_ENABLE_INT8
-            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
-                              is_same_v<OutDataType, int8_t>)
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
             }

library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp

@@ -98,6 +98,26 @@ void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
+        instances);
 #endif
 
 template <typename ADataType,
@@ -105,7 +125,8 @@ template <typename ADataType,
           typename CDataType,
           typename ALayout,
           typename BLayout,
-          typename CLayout>
+          typename CLayout,
+          typename ComputeType>
 struct DeviceOperationInstanceFactory<
     ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
                                                    BLayout,
@@ -115,7 +136,8 @@ struct DeviceOperationInstanceFactory<
                                                    CDataType,
                                                    ck::tensor_operation::element_wise::PassThrough,
                                                    ck::tensor_operation::element_wise::PassThrough,
-                                                   ck::tensor_operation::element_wise::PassThrough>>
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ComputeType>>
 {
     using DeviceOp = DeviceGemmSplitK<ALayout,
                                       BLayout,
@@ -125,14 +147,15 @@ struct DeviceOperationInstanceFactory<
                                       CDataType,
                                       ck::tensor_operation::element_wise::PassThrough,
                                       ck::tensor_operation::element_wise::PassThrough,
-                                      ck::tensor_operation::element_wise::PassThrough>;
+                                      ck::tensor_operation::element_wise::PassThrough,
+                                      ComputeType>;
 
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
 #ifdef CK_ENABLE_FP32
         if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
-                     is_same_v<CDataType, float>)
+                     is_same_v<CDataType, float> && is_same_v<ComputeType, float>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
@@ -157,8 +180,8 @@ struct DeviceOperationInstanceFactory<
         }
 #endif
 #ifdef CK_ENABLE_FP16
-        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
-                          is_same_v<CDataType, half_t>)
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
@@ -183,8 +206,8 @@ struct DeviceOperationInstanceFactory<
         }
 #endif
 #if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
-        else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
-                          is_same_v<CDataType, half_t>)
+        if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
@@ -207,8 +230,8 @@ struct DeviceOperationInstanceFactory<
                 add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(op_ptrs);
             }
         }
-        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
-                          is_same_v<CDataType, half_t>)
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
+                     is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
         {
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
@@ -231,6 +254,31 @@ struct DeviceOperationInstanceFactory<
                 add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs);
             }
         }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t> && is_same_v<ComputeType, f8_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(op_ptrs);
+            }
+        }
+
 #endif
         return op_ptrs;
     }

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp

@@ -6,8 +6,6 @@
 #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
-#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-
 namespace ck {
 namespace tensor_operation {
 namespace device {