Remove program server (#10)

* removing program server

* specify launch bound per kernel instance

include/ck/tile_program/grid/grid_gemm.hpp

@@ -24,8 +24,7 @@ struct GridGemm
     using BlockGemmPipeline = typename Policy::template BlockGemmPipeline<Problem>;
 
     template <typename AGridTensorView, typename BGridTensorView, typename CGridTensorView>
-    __host__ __device__ void operator()(ProgramServer& ps,
-                                        const AGridTensorView& a_grid,
+    __device__ void operator()(const AGridTensorView& a_grid,
                                const BGridTensorView& b_grid,
                                CGridTensorView& c_grid,
                                const AElementFunction& a_element_func,
@@ -41,17 +40,17 @@ struct GridGemm
         const auto K = a_grid.desc_.GetLength(Number<1>{});
 
         // divide problem
-        const auto id_block = ps.get_block_id();
+        const auto id_block = get_block_id();
 
         const auto num_tile_m = M / kMPerBlock;
         const auto num_tile_n = N / kNPerBlock;
 
-        const auto block2tile = ps(Policy::MakeBlock2TileMap(num_tile_m, num_tile_n));
+        const auto block2tile = Policy::MakeBlock2TileMap(num_tile_m, num_tile_n);
 
         const auto id_tile = block2tile(id_block);
 
-        const auto iM = ps.read_first_lane(id_tile.template At<0>() * kMPerBlock);
-        const auto iN = ps.read_first_lane(id_tile.template At<1>() * kNPerBlock);
+        const auto iM = __builtin_amdgcn_readfirstlane(id_tile.template At<0>() * kMPerBlock);
+        const auto iN = __builtin_amdgcn_readfirstlane(id_tile.template At<1>() * kNPerBlock);
 
         // A block window
         auto a_block_window = make_tile_window(

include/ck/tile_program/tile/distributed_tile_sweep.hpp

@@ -14,7 +14,7 @@ namespace tile_program {
 template <typename TileDistributedSpan_, // TileDistributedSpan<...>
           typename F                     // signature: F(TileDistributedIndex<...>)
           >
-__host__ __device__ void sweep_tile_span(TileDistributedSpan_, const F& f)
+__device__ void sweep_tile_span(TileDistributedSpan_, const F& f)
 {
     using DstrSpan = remove_cvref_t<TileDistributedSpan_>;
 

include/ck/tile_program/tile/load_tile.hpp

@@ -160,25 +160,6 @@ __device__ auto load_sliced_thread_data_from_tile_window(
 
 } // namespace detail
 
-// FIXME: host dummy function for tile program
-template <typename BottomTensorView_, typename WindowLengths_, typename TileDistribution_>
-__host__ auto load_tile(
-    const TileWindowWithStaticDistribution<BottomTensorView_, WindowLengths_, TileDistribution_>&
-        tile_window)
-{
-    using DataType         = remove_cvref_t<typename BottomTensorView_::DataType>;
-    using BottomTensorView = remove_cvref_t<BottomTensorView_>;
-    using WindowLengths    = remove_cvref_t<WindowLengths_>;
-    using TileDstr         = remove_cvref_t<TileDistribution_>;
-    using TileWindow = TileWindowWithStaticDistribution<BottomTensorView, WindowLengths, TileDstr>;
-
-    static_assert(is_known_at_compile_time<WindowLengths>::value,
-                  "wrong! lengths should be static");
-    static_assert(TileWindow::HasStaticTileDistribution(), "wrong!");
-
-    return make_static_distributed_tensor<DataType>(tile_window.GetTileDistribution());
-}
-
 template <typename BottomTensorView_, typename WindowLengths_, typename TileDistribution_>
 __device__ auto
 load_tile(TileWindowWithStaticDistribution<BottomTensorView_, WindowLengths_, TileDistribution_>&

include/ck/tile_program/tile/store_tile_impl_static_distribution.hpp

@@ -14,17 +14,6 @@
 namespace ck {
 namespace tile_program {
 
-// FIXME: host dummy function for tile program
-template <typename BottomTensorView_,
-          typename WindowLengths_,
-          typename TileDistribution_,
-          typename DataType_>
-__host__ void
-store_tile(TileWindowWithStaticDistribution<BottomTensorView_, WindowLengths_, TileDistribution_>&,
-           const StaticDistributedTensor<DataType_, TileDistribution_>&)
-{
-}
-
 template <typename BottomTensorView_,
           typename WindowLengths_,
           typename TileDistribution_,

include/ck/tile_program/tile/store_tile_impl_static_lengths.hpp

@@ -15,16 +15,6 @@
 namespace ck {
 namespace tile_program {
 
-// FIXME: host dummy function for tile program
-template <typename BottomTensorView_,
-          typename WindowLengths_,
-          typename TileDistribution_,
-          typename DataType_>
-__host__ void store_tile(TileWindowWithStaticLengths<BottomTensorView_, WindowLengths_>&,
-                         const StaticDistributedTensor<DataType_, TileDistribution_>&)
-{
-}
-
 template <typename BottomTensorView_,
           typename WindowLengths_,
           typename TileDistribution_,

include/ck/tile_program/tile/tile_elementwise.hpp

@@ -16,7 +16,7 @@ namespace tile_program {
 
 // TODO: support tensors with different distribution
 template <typename InOutElementFunc, typename... InOutDstrTensors>
-__host__ __device__ void tile_elementwise_inout(const InOutElementFunc& inout_element_func,
+__device__ void tile_elementwise_inout(const InOutElementFunc& inout_element_func,
                                        InOutDstrTensors&... inout_dstr_tensors)
 {
     // TODO: make sure all distributed tensors have same lengths and distribution
@@ -30,7 +30,7 @@ __host__ __device__ void tile_elementwise_inout(const InOutElementFunc& inout_el
 }
 
 template <typename InElementFunc, typename... InDstrTensors>
-__host__ __device__ auto tile_elementwise_in(const InElementFunc& in_element_func,
+__device__ auto tile_elementwise_in(const InElementFunc& in_element_func,
                                     const InDstrTensors&... in_dstr_tensors)
 {
     using OutDataType = decltype(in_element_func(typename InDstrTensors::DataType{}...));

include/ck/tile_program/tile/tile_window_impl_static_distribution.hpp

@@ -45,20 +45,7 @@ struct TileWindowWithStaticDistribution
     using BottomTensorCoord =
         decltype(make_tensor_coordinate(BottomTensorDesc{}, BottomTensorIndex{}));
 
-    __host__ __device__ constexpr TileWindowWithStaticDistribution() = default;
-
-    // FIXME: host dummy constructor for tile program
-    __host__ constexpr TileWindowWithStaticDistribution(const BottomTensorView& bottom_tensor_view,
-                                                        const WindowLengths&,
-                                                        const BottomTensorIndex&,
-                                                        const TileDstr&)
-        : bottom_tensor_view_{bottom_tensor_view},
-          window_lengths_{},
-          bottom_tensor_thread_coord_{},
-          tile_dstr_{},
-          window_adaptor_thread_coord_{}
-    {
-    }
+    __device__ constexpr TileWindowWithStaticDistribution() = default;
 
     __device__ constexpr TileWindowWithStaticDistribution(
         const BottomTensorView& bottom_tensor_view,
@@ -86,22 +73,19 @@ struct TileWindowWithStaticDistribution
             bottom_tensor_view_.GetTensorDescriptor(), bottom_tensor_thread_origin_idx);
     }
 
-    __host__ __device__ static constexpr index_t GetNumOfDimension() { return NDimBottomTensor; }
+    __device__ static constexpr index_t GetNumOfDimension() { return NDimBottomTensor; }
 
-    __host__ __device__ static constexpr bool HasStaticTileDistribution()
-    {
-        return TileDstr::IsStatic();
-    }
+    __device__ static constexpr bool HasStaticTileDistribution() { return TileDstr::IsStatic(); }
 
-    __host__ __device__ constexpr auto GetWindowLengths() const { return window_lengths_; }
+    __device__ constexpr auto GetWindowLengths() const { return window_lengths_; }
 
-    __host__ __device__ constexpr auto GetTileDistribution() const { return tile_dstr_; }
+    __device__ constexpr auto GetTileDistribution() const { return tile_dstr_; }
 
-    __host__ __device__ constexpr auto GetBottomTensorView() const { return bottom_tensor_view_; }
+    __device__ constexpr auto GetBottomTensorView() const { return bottom_tensor_view_; }
 
-    __host__ __device__ constexpr auto GetWindowOrigin() const { return window_origin_; }
+    __device__ constexpr auto GetWindowOrigin() const { return window_origin_; }
 
-    __host__ __device__ constexpr auto GetBottomTensorThreadCoordinate() const
+    __device__ constexpr auto GetBottomTensorThreadCoordinate() const
     {
         return bottom_tensor_thread_coord_;
     }
@@ -141,7 +125,7 @@ struct TileWindowWithStaticDistribution
     }
 
     // return vector dimension among [y0, y1, ...]
-    __host__ __device__ static constexpr auto GetWindowAdaptorYsSafeVectorLengthStrides()
+    __device__ static constexpr auto GetWindowAdaptorYsSafeVectorLengthStrides()
     {
         // bottom tensor top dimension vector lengths and strides
         const auto [bottom_tensor_top_dim_vector_lengths, bottom_tensor_top_dim_vector_strides] =
@@ -201,7 +185,7 @@ struct TileWindowWithStaticDistribution
 
 // TODO: use strategy
 template <typename TensorView_, typename WindowLengths_, typename StaticTileDistribution_>
-__host__ __device__ constexpr auto
+__device__ constexpr auto
 make_tile_window(const TensorView_& tensor_view,
                  const WindowLengths_& window_lengths,
                  const MultiIndex<TensorView_::GetNumOfDimension()>& origin,
@@ -213,16 +197,6 @@ make_tile_window(const TensorView_& tensor_view,
         tensor_view, window_lengths, origin, tile_distribution};
 }
 
-// FIXME: dummy host function for tile program
-template <typename TensorView_, typename WindowLengths_, typename StaticTileDistribution_>
-__host__ void move_tile_window(
-    TileWindowWithStaticDistribution<TensorView_, WindowLengths_, StaticTileDistribution_>&,
-    const MultiIndex<
-        TileWindowWithStaticDistribution<TensorView_, WindowLengths_, StaticTileDistribution_>::
-            GetNumOfDimension()>&)
-{
-}
-
 template <typename TensorView_, typename WindowLengths_, typename StaticTileDistribution_>
 __device__ void move_tile_window(
     TileWindowWithStaticDistribution<TensorView_, WindowLengths_, StaticTileDistribution_>& window,

include/ck/tile_program/tile/tile_window_impl_static_lengths.hpp

@@ -27,15 +27,7 @@ struct TileWindowWithStaticLengths
 
     using BottomTensorIndex = Array<index_t, NDimBottomTensor>;
 
-    __host__ __device__ constexpr TileWindowWithStaticLengths() = default;
-
-    // FIXME: host dummy constructor for tile program
-    __host__ constexpr TileWindowWithStaticLengths(const BottomTensorView& bottom_tensor_view,
-                                                   const WindowLengths&,
-                                                   const BottomTensorIndex&)
-        : bottom_tensor_view_{bottom_tensor_view}, window_lengths_{}, window_origin_{}
-    {
-    }
+    __device__ constexpr TileWindowWithStaticLengths() = default;
 
     __device__ constexpr TileWindowWithStaticLengths(const BottomTensorView& bottom_tensor_view,
                                                      const WindowLengths& window_lengths,
@@ -46,13 +38,13 @@ struct TileWindowWithStaticLengths
     {
     }
 
-    __host__ __device__ static constexpr index_t GetNumOfDimension() { return NDimBottomTensor; }
+    __device__ static constexpr index_t GetNumOfDimension() { return NDimBottomTensor; }
 
-    __host__ __device__ constexpr auto GetWindowLengths() const { return window_lengths_; }
+    __device__ constexpr auto GetWindowLengths() const { return window_lengths_; }
 
-    __host__ __device__ constexpr auto GetBottomTensorView() const { return bottom_tensor_view_; }
+    __device__ constexpr auto GetBottomTensorView() const { return bottom_tensor_view_; }
 
-    __host__ __device__ constexpr auto GetWindowOrigin() const { return window_origin_; }
+    __device__ constexpr auto GetWindowOrigin() const { return window_origin_; }
 
     // this is the bottom tensor view
     // [x0', x1', ...] ==> [offset]
@@ -66,7 +58,7 @@ struct TileWindowWithStaticLengths
 };
 
 template <typename TensorView_, typename WindowLengths_>
-__host__ __device__ constexpr auto
+__device__ constexpr auto
 make_tile_window(const TensorView_& tensor_view,
                  const WindowLengths_& window_lengths,
                  const MultiIndex<TensorView_::GetNumOfDimension()>& origin)
@@ -78,15 +70,6 @@ make_tile_window(const TensorView_& tensor_view,
         tensor_view, window_lengths, origin};
 }
 
-// FIXME: dummy host function for tile program
-template <typename TensorView_, typename WindowLengths_>
-__host__ void move_tile_window(
-    TileWindowWithStaticLengths<TensorView_, WindowLengths_>&,
-    const MultiIndex<
-        TileWindowWithStaticLengths<TensorView_, WindowLengths_>::GetNumOfDimension()>&)
-{
-}
-
 template <typename TensorView_, typename WindowLengths_>
 __device__ void move_tile_window(
     TileWindowWithStaticLengths<TensorView_, WindowLengths_>& window,