[CK_TILE] wa prec, remove sgpr offset for inline asm (#1356)

* wa prec, remove sgpr offset for inline asm

* macro for set tile

* ignore unused param if no kernel instances in host API

* fix more prec issue

* cache buffer resource

* fix

* support pre-nop

* clear tile by vector type members

* add workaround to reduce scratch memory

* conditionally enable workaround code

* enable workaround start from certain build version

* fallback set_tile() implementation from certain build version

* undo template argument changes

* put dummy asm in load_raw()

* fix comments, refactor s_nop inside buffer_load

---------
Co-authored-by: PoYen, Chen <PoYen.Chen@amd.com>

example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py

@@ -271,7 +271,9 @@ class FmhaBwdApiPool:
                 per_hdim_case = per_hdim_case + FMHA_BWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
             if_i = 'if' if i == 0 else 'else if'
             per_dtypes = per_dtypes + FMHA_BWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
-
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
         return FMHA_BWD_KERNEL_HEADER + FMHA_BWD_API.format(F_dispatch = per_dtypes)
 
 # GEMM0: Q@K=S^T

example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py

@@ -278,6 +278,9 @@ class FmhaFwdApiPool:
                 per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
             if_i = 'if' if i == 0 else 'else if'
             per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
         return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_API.format(F_dispatch = per_dtypes)
 
 @dataclass

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py

@@ -331,6 +331,9 @@ class FmhaFwdSplitKVApiPool:
                 per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
             if_i = 'if' if i == 0 else 'else if'
             per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
         return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_SPLITKV_API.format(F_dispatch = per_dtypes)
 
 @dataclass

include/ck_tile/core/arch/arch.hpp

@@ -82,14 +82,12 @@ CK_TILE_DEVICE void block_sync_lds_direct_load()
     " ::);
 }
 
-CK_TILE_DEVICE void s_nop()
+CK_TILE_DEVICE void s_nop(index_t cnt = 0)
 {
 #if 1
-    asm volatile("\
-    s_nop 0 \n \
-    " ::);
+    asm volatile("s_nop %0" : : "n"(cnt) :);
 #else
-    __builtin_amdgcn_sched_barrier(0);
+    __builtin_amdgcn_sched_barrier(cnt);
 #endif
 }
 

include/ck_tile/core/config.hpp

@@ -21,6 +21,7 @@
 #define __gfx12__
 #endif
 
+#include "hip/hip_version.h"
 #ifndef CK_TILE_DONT_USE_HIP_RUNTIME_HEADERS
 #include "hip/hip_runtime.h"
 #include "hip/hip_fp16.h"
@@ -147,6 +148,14 @@
 #define CK_TILE_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE 1
 #endif
 
+#ifndef CK_TILE_WORKAROUND_ROCM_6_1_SCRATCH_MEMORY_ISSUE
+#if HIP_VERSION_MAJOR == 6 && HIP_VERSION_MINOR == 1 && HIP_VERSION_PATCH >= 40091
+#define CK_TILE_WORKAROUND_ROCM_6_1_SCRATCH_MEMORY_ISSUE 1
+#else
+#define CK_TILE_WORKAROUND_ROCM_6_1_SCRATCH_MEMORY_ISSUE 0
+#endif
+#endif
+
 #ifndef CK_TILE_DEBUG_LOG
 #define CK_TILE_DEBUG_LOG 0
 #endif

include/ck_tile/core/tensor/buffer_view.hpp

@@ -69,6 +69,8 @@ struct buffer_view<address_space_enum::generic,
     {
     }
 
+    CK_TILE_HOST_DEVICE void init_raw() {}
+
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::generic;
@@ -224,25 +226,36 @@ struct buffer_view<address_space_enum::global,
 
     T* p_data_ = nullptr;
     BufferSizeType buffer_size_;
+    int32x4_t cached_buf_res_;
     remove_cvref_t<T> invalid_element_value_ = T{0};
 
     CK_TILE_HOST_DEVICE constexpr buffer_view()
-        : p_data_{}, buffer_size_{}, invalid_element_value_{}
+        : p_data_{}, buffer_size_{}, cached_buf_res_{0}, invalid_element_value_{}
     {
     }
 
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data, BufferSizeType buffer_size)
-        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
+        : p_data_{p_data}, buffer_size_{buffer_size}, cached_buf_res_{0}, invalid_element_value_{0}
     {
     }
 
     CK_TILE_HOST_DEVICE constexpr buffer_view(T* p_data,
                                               BufferSizeType buffer_size,
                                               T invalid_element_value)
-        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
+        : p_data_{p_data},
+          buffer_size_{buffer_size},
+          cached_buf_res_{0},
+          invalid_element_value_{invalid_element_value}
     {
     }
 
+    // this is non constexpr intentially (will call some intrinsic internally)
+    // Must call for buffers that need *_raw load/store
+    CK_TILE_HOST_DEVICE void init_raw()
+    {
+        cached_buf_res_ = make_wave_buffer_resource(p_data_, buffer_size_ * sizeof(type));
+    }
+
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::global;
@@ -333,12 +346,15 @@ struct buffer_view<address_space_enum::global,
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
+              bool pre_nop               = false,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
-    CK_TILE_DEVICE constexpr auto
-    get_raw(remove_cvref_t<X>& dst, index_t i, bool is_valid_element) const
+    CK_TILE_DEVICE constexpr auto get_raw(remove_cvref_t<X>& dst,
+                                          index_t i,
+                                          bool is_valid_element,
+                                          bool_constant<pre_nop> = {}) const
     {
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
 
@@ -349,18 +365,21 @@ struct buffer_view<address_space_enum::global,
 
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
 
-        amd_buffer_load_raw<remove_cvref_t<T>, t_per_x, Coherence, oob_conditional_check>(
-            dst, p_data_, i, buffer_size_, is_valid_element);
+        amd_buffer_load_raw<remove_cvref_t<T>, t_per_x, Coherence, oob_conditional_check, pre_nop>(
+            dst, cached_buf_res_, i, is_valid_element, bool_constant<pre_nop>{});
     }
 
     // i is offset of T, not X. i should be aligned to X
     template <typename X,
+              bool pre_nop = false,
               typename std::enable_if<
                   std::is_same<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                typename vector_traits<remove_cvref_t<T>>::scalar_type>::value,
                   bool>::type = false>
-    CK_TILE_DEVICE constexpr auto
-    async_get(remove_cvref_t<T>* smem, index_t i, bool /*is_valid_element*/) const
+    CK_TILE_DEVICE constexpr auto async_get_raw(remove_cvref_t<T>* smem,
+                                                index_t i,
+                                                bool /*is_valid_element*/,
+                                                bool_constant<pre_nop> = {}) const
     {
         // X is vector of T
         constexpr index_t scalar_per_t_vector = vector_traits<remove_cvref_t<T>>::vector_size;
@@ -371,8 +390,8 @@ struct buffer_view<address_space_enum::global,
 
         constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
 
-        amd_async_buffer_load_with_oob<remove_cvref_t<T>, t_per_x, Coherence>(
-            smem, p_data_, i, buffer_size_);
+        amd_async_buffer_load_with_oob_raw<remove_cvref_t<T>, t_per_x, Coherence>(
+            smem, cached_buf_res_, i, bool_constant<pre_nop>{});
     }
 
     // i is offset of T, not X. i should be aligned to X
@@ -627,6 +646,8 @@ struct buffer_view<address_space_enum::lds,
     {
     }
 
+    CK_TILE_HOST_DEVICE void init_raw() {}
+
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::lds;
@@ -909,6 +930,8 @@ struct buffer_view<address_space_enum::vgpr,
     {
     }
 
+    CK_TILE_HOST_DEVICE void init_raw() {}
+
     CK_TILE_DEVICE static constexpr address_space_enum get_address_space()
     {
         return address_space_enum::vgpr;

include/ck_tile/core/tensor/load_tile.hpp

@@ -36,30 +36,37 @@ template <typename T,
           typename WindowLengths_,
           typename TileDistribution_,
           index_t NumCoord,
-          bool oob_conditional_check = true>
+          bool oob_conditional_check = true,
+          bool pre_nop               = false>
 CK_TILE_DEVICE auto load_tile_raw(T& tile,
                                   const tile_window_with_static_distribution<BottomTensorView_,
                                                                              WindowLengths_,
                                                                              TileDistribution_,
                                                                              NumCoord>& tile_window,
-                                  bool_constant<oob_conditional_check> = {})
+                                  bool_constant<oob_conditional_check> = {},
+                                  bool_constant<pre_nop>               = {})
 {
-    tile_window.load_raw(tile, bool_constant<oob_conditional_check>{});
+    tile_window.load_raw(tile, bool_constant<oob_conditional_check>{}, bool_constant<pre_nop>{});
 }
 
 template <typename LdsTileWindow_,
           typename BottomTensorView_,
           typename WindowLengths_,
           typename TileDistribution_,
-          index_t NumCoord>
+          index_t NumCoord,
+          bool oob_conditional_check = true,
+          bool pre_nop               = false>
 CK_TILE_DEVICE auto
 async_load_tile_raw(LdsTileWindow_&& lds_tile,
                     const tile_window_with_static_distribution<BottomTensorView_,
                                                                WindowLengths_,
                                                                TileDistribution_,
-                                                               NumCoord>& tile_window)
+                                                               NumCoord>& tile_window,
+                    bool_constant<oob_conditional_check> = {},
+                    bool_constant<pre_nop>               = {})
 {
-    return tile_window.async_load(lds_tile);
+    return tile_window.async_load_raw(
+        lds_tile, bool_constant<oob_conditional_check>{}, bool_constant<pre_nop>{});
 }
 
 CK_TILE_DEVICE auto async_load_fence(index_t cnt = 0)

include/ck_tile/core/tensor/null_tile_window.hpp

@@ -35,6 +35,8 @@ struct null_tile_window
 
     CK_TILE_DEVICE constexpr auto get_window_origin() const { return BottomTensorIndex{}; }
 
+    CK_TILE_DEVICE void init_raw() {}
+
     WindowLengths window_lengths_;
 };
 

include/ck_tile/core/tensor/tensor_view.hpp

@@ -36,6 +36,8 @@ struct tensor_view
     {
     }
 
+    CK_TILE_HOST_DEVICE void init_raw() { buf_.init_raw(); }
+
     CK_TILE_HOST_DEVICE constexpr auto& get_tensor_descriptor() const { return desc_; }
 
     CK_TILE_HOST_DEVICE static constexpr index_t get_num_of_dimension()
@@ -85,30 +87,34 @@ struct tensor_view
     // "coord" is coordinate of DataType, not X. "coord" should be aligned to X
     template <typename X,
               bool oob_conditional_check = true,
+              bool pre_nop               = false,
               typename std::enable_if<
                   std::is_same_v<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                  typename vector_traits<remove_cvref_t<DataType>>::scalar_type>,
                   bool>::type = false>
-    CK_TILE_HOST_DEVICE void
-    get_vectorized_elements_raw(remove_cvref_t<X>& dst,
-                                const TensorCoord& coord,
-                                bool_constant<oob_conditional_check> = {}) const
+    CK_TILE_HOST_DEVICE void get_vectorized_elements_raw(remove_cvref_t<X>& dst,
+                                                         const TensorCoord& coord,
+                                                         bool_constant<oob_conditional_check> = {},
+                                                         bool_constant<pre_nop> = {}) const
     {
-        return buf_.template get_raw<X, oob_conditional_check>(
+        return buf_.template get_raw<X, oob_conditional_check, pre_nop>(
             dst,
             coord.get_offset(),
-            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord));
+            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord),
+            bool_constant<pre_nop>{});
     }
 
     template <typename X,
+              bool pre_nop = false,
               typename std::enable_if<
                   std::is_same_v<typename vector_traits<remove_cvref_t<X>>::scalar_type,
                                  typename vector_traits<remove_cvref_t<DataType>>::scalar_type>,
                   bool>::type = false>
-    CK_TILE_HOST_DEVICE constexpr void async_get_vectorized_elements(remove_cvref_t<DataType>* smem,
-                                                                     const TensorCoord& coord) const
+    CK_TILE_HOST_DEVICE constexpr void async_get_vectorized_elements_raw(
+        remove_cvref_t<DataType>* smem, const TensorCoord& coord, bool_constant<pre_nop> = {}) const
     {
-        return buf_.template async_get<X>(smem, coord.get_offset(), true /*not used*/);
+        return buf_.template async_get_raw<X>(
+            smem, coord.get_offset(), true /*not used*/, bool_constant<pre_nop>{});
     }
 
     // X is vector of DataType.

include/ck_tile/core/tensor/tile_elementwise.hpp

@@ -76,23 +76,63 @@ CK_TILE_DEVICE void set_tile(null_tensor&, const T&)
 
 // TODO: prefer to use per-dword value to set a tensor, in case compiler not doing well with
 // sub-dword tensor...
-template <typename DstrTensors, index_t v>
-CK_TILE_DEVICE void set_tile(DstrTensors& dstr_tensor, number<v>)
+template <typename DstrTensors, index_t v, bool skip_subdword_opt = false>
+CK_TILE_DEVICE void
+set_tile(DstrTensors& dstr_tensor, number<v>, bool_constant<skip_subdword_opt> = {})
 {
-    constexpr index_t tensor_bytes =
-        DstrTensors::get_thread_buffer_size() * sizeof(typename DstrTensors::DataType);
-    if constexpr(v == 0 && tensor_bytes % 4 == 0)
+    using elem_type             = typename DstrTensors::DataType;
+    constexpr index_t elem_size = sizeof(elem_type);
+
+    constexpr index_t tensor_bytes = DstrTensors::get_thread_buffer_size() * elem_size;
+
+    // # bytes per write = 4
+    if constexpr(v == 0 && tensor_bytes % 4 == 0 && !skip_subdword_opt)
     {
+#if CK_TILE_WORKAROUND_ROCM_6_1_SCRATCH_MEMORY_ISSUE
+        auto& buffer = dstr_tensor.get_thread_buffer();
+
+        static_for<0, tensor_bytes / 4, 1>{}([&](auto i_write) {
+            if constexpr(elem_size == 1)
+            {
+                // # elements per write = 4
+                constexpr auto values = ext_vector_t<elem_type, 4>{0, 0, 0, 0};
+
+                buffer[i_write * 4 + 0] = values.x;
+                buffer[i_write * 4 + 1] = values.y;
+                buffer[i_write * 4 + 2] = values.z;
+                buffer[i_write * 4 + 3] = values.w;
+            }
+            else if constexpr(elem_size == 2)
+            {
+                // # elements per write = 2
+                constexpr auto values = ext_vector_t<elem_type, 2>{0, 0};
+
+                buffer[i_write * 2 + 0] = values.x;
+                buffer[i_write * 2 + 1] = values.y;
+            }
+            else if constexpr(elem_size == 4)
+            {
+                // # elements per write = 1
+                constexpr elem_type value = 0;
+
+                buffer[i_write] = value;
+            }
+            else
+            {
+                static_assert(false, "type not supported");
+            }
+        });
+#else
         using dvec_t = array<index_t, tensor_bytes / 4>;
         auto& tensor = reinterpret_cast<dvec_t&>(dstr_tensor.get_thread_buffer());
         for(auto i = 0; i < tensor.size(); i++)
             tensor.get(i) = v;
+#endif
     }
     else
     {
-        tile_elementwise_inout(
-            [](auto& x) { x = type_convert<typename DstrTensors::DataType, index_t>(v); },
-            dstr_tensor);
+        tile_elementwise_inout([](auto& x) { x = type_convert<elem_type, index_t>(v); },
+                               dstr_tensor);
     }
 }
 

include/ck_tile/core/tensor/tile_window.hpp

@@ -344,9 +344,10 @@ struct tile_window_with_static_distribution
         return dst_tensor;
     }
 
-    template <typename DstTile, bool oob_conditional_check = true>
+    template <typename DstTile, bool oob_conditional_check = true, bool pre_nop = false>
     CK_TILE_DEVICE void load_raw(DstTile& dst_tensor,
-                                 bool_constant<oob_conditional_check> = {}) const
+                                 bool_constant<oob_conditional_check> = {},
+                                 bool_constant<pre_nop>               = {}) const
     {
         using Traits = load_store_traits;
 
@@ -373,7 +374,13 @@ struct tile_window_with_static_distribution
             auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
 
             static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
-                constexpr auto iAccess = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
+                constexpr auto iAccess  = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
+                constexpr auto pre_nop_ = [&]() {
+                    if constexpr(pre_nop && iCoord == 0 && iCoordAccess == 0)
+                        return bool_constant<true>{};
+                    else
+                        return bool_constant<false>{};
+                }();
 
                 // data index [y0, y1, ...]
                 constexpr auto idx_ys_start = SFC_Ys::get_index(iAccess);
@@ -384,7 +391,8 @@ struct tile_window_with_static_distribution
                 get_bottom_tensor_view().template get_vectorized_elements_raw<vector_t>(
                     dst_vec_tbuf.template at<d / Traits::ScalarPerVector>(),
                     bottom_tensor_thread_coord,
-                    bool_constant<oob_conditional_check>{});
+                    bool_constant<oob_conditional_check>{},
+                    pre_nop_);
 
                 // move thread coordinate
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
@@ -399,12 +407,17 @@ struct tile_window_with_static_distribution
                 }
             });
         });
+#if CK_TILE_WORKAROUND_ROCM_6_1_SCRATCH_MEMORY_ISSUE
+        asm volatile("; this inline asm is workaround to prevent compiler from using too much "
+                     "scratch memory" ::);
+#endif
     }
 
     // TODO: currently async load only implemented in inline asm
-    template <typename LdsTileWindow_, bool oob_conditional_check = true>
-    CK_TILE_DEVICE auto async_load(LdsTileWindow_&& lds_tile,
-                                   bool_constant<oob_conditional_check> = {}) const
+    template <typename LdsTileWindow_, bool oob_conditional_check = true, bool pre_nop = false>
+    CK_TILE_DEVICE auto async_load_raw(LdsTileWindow_&& lds_tile,
+                                       bool_constant<oob_conditional_check> = {},
+                                       bool_constant<pre_nop>               = {}) const
     {
         using LdsTileWindow = remove_cvref_t<LdsTileWindow_>;
         // using LdsTensorView = typename LdsTileWindow::BottomTensorView;
@@ -449,11 +462,17 @@ struct tile_window_with_static_distribution
             auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
 
             static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
-                constexpr auto iAccess = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
+                constexpr auto iAccess  = number<iCoord * NumAccessPerCoord + iCoordAccess>{};
+                constexpr auto pre_nop_ = [&]() {
+                    if constexpr(pre_nop && iCoord == 0 && iCoordAccess == 0)
+                        return bool_constant<true>{};
+                    else
+                        return bool_constant<false>{};
+                }();
 
                 // read from bottom tensor
-                get_bottom_tensor_view().template async_get_vectorized_elements<vector_t>(
-                    smem, bottom_tensor_thread_coord);
+                get_bottom_tensor_view().template async_get_vectorized_elements_raw<vector_t>(
+                    smem, bottom_tensor_thread_coord, pre_nop_);
 
                 // move thread coordinate
                 if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
@@ -668,6 +687,67 @@ struct tile_window_with_static_distribution
         });
     }
 
+    CK_TILE_DEVICE void set_window_origin(const BottomTensorIndex& new_window_origin)
+    {
+        window_origin_ = new_window_origin;
+
+#if 0 // debug
+      // TODO: this use more register for FA, but less register for GEMM
+      // need investigation
+      // only support warp-tile and block-tile
+        static_assert(NDimP == 1 or NDimP == 2, "wrong!");
+
+        WindowAdaptorCoord window_adaptor_thread_coord_tmp;
+
+        if constexpr(NDimP == 1)
+        {
+            window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
+                tile_dstr_.get_ps_ys_to_xs_adaptor(), AdaptorTopIndex{get_lane_id(), 0});
+        }
+        else if constexpr(NDimP == 2)
+        {
+            window_adaptor_thread_coord_tmp =
+                make_tensor_adaptor_coordinate(tile_dstr_.get_ps_ys_to_xs_adaptor(),
+                                               AdaptorTopIndex{get_warp_id(), get_lane_id(), 0});
+        }
+#else
+        // TODO: this use less register for FA, but more register for GEMM
+        // need investigation
+        const auto window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
+            tile_dstr_.get_ps_ys_to_xs_adaptor(),
+            container_concat(detail::get_partition_index(tile_dstr_), array<index_t, NDimY>{0}));
+#endif
+
+        BottomTensorIndex bottom_tensor_thread_origin_idx_tmp =
+            window_origin_ + window_adaptor_thread_coord_tmp.get_bottom_index();
+
+        const auto bottom_tensor_thread_coord_tmp = make_tensor_coordinate(
+            bottom_tensor_view_.get_tensor_descriptor(), bottom_tensor_thread_origin_idx_tmp);
+
+        // pre-compute NumCoord (WindowAdaptorCoord, BottomTensorCoord) bundles to speed up
+        // future load/store() calls (might allocate more registers)
+        using Traits = load_store_traits;
+        using SFC_Ys = typename Traits::SFC_Ys;
+
+        static_for<0, NumCoord, 1>{}([&](auto iCoord) {
+            auto window_adaptor_thread_coord = window_adaptor_thread_coord_tmp;
+            auto bottom_tensor_thread_coord  = bottom_tensor_thread_coord_tmp;
+
+            constexpr auto idx_diff_ys =
+                SFC_Ys::get_step_between(number<0>{}, number<iCoord * NumAccessPerCoord>{});
+
+            constexpr auto idx_diff_ps_ys = container_concat(array<index_t, NDimP>{0}, idx_diff_ys);
+
+            move_window_adaptor_and_bottom_tensor_thread_coordinate(
+                window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
+
+            pre_computed_coords_(iCoord) =
+                make_tuple(window_adaptor_thread_coord, bottom_tensor_thread_coord);
+        });
+    }
+
+    CK_TILE_HOST_DEVICE void init_raw() { bottom_tensor_view_.init_raw(); }
+
     // this is the bottom tensor view
     // [x0', x1', ...] ==> [offset]
     BottomTensorView bottom_tensor_view_;

include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp

@@ -81,6 +81,12 @@ struct BlockFmhaPipelineQRKSVSAsync
             return Problem::kBlockPerCu;
         else
         {
+            // minimize occupancy
+            if constexpr(BiasEnum != BlockAttentionBiasEnum::NO_BIAS && kHasDropout)
+            {
+                return 1;
+            }
+
             if constexpr(kK0BlockLength <= 32)
             {
                 if constexpr(kPadSeqLenK && BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS &&
@@ -220,6 +226,7 @@ struct BlockFmhaPipelineQRKSVSAsync
             q_dram_block_window_tmp.get_window_lengths(),
             q_dram_block_window_tmp.get_window_origin(),
             Policy::template MakeQDramTileDistribution<Problem, decltype(gemm_0)>());
+        q_dram_window.init_raw();
 
         // TODO: we use async Copy for K, which is inline asm
         // a side effect is we have to use inline asm for q as well
@@ -293,6 +300,17 @@ struct BlockFmhaPipelineQRKSVSAsync
             k_dram_block_window.get_window_origin(),
             Policy::template MakeKDramTileDistribution<Problem>()); // K DRAM tile window for
                                                                     // load
+        k_dram_window.init_raw();
+        constexpr auto k_oob_ck = bool_constant<true>{};
+        constexpr auto k_pre_np = [&]() {
+            if constexpr(kPadSeqLenK &&
+                         (BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS ||
+                          (BiasEnum != BlockAttentionBiasEnum::NO_BIAS && kHasDropout)))
+                return bool_constant<true>{};
+            else
+                return bool_constant<false>{};
+        }();
+
         const auto bias_origin = bias_dram_block_window_tmp.get_window_origin();
         auto bias_dram_window  = make_tile_window(
             bias_dram_block_window_tmp.get_bottom_tensor_view(),
@@ -310,7 +328,7 @@ struct BlockFmhaPipelineQRKSVSAsync
                              Policy::template MakeVDramTileDistribution<Problem>());
 
         // prefetch K tile
-        async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), k_dram_window);
+        async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, k_oob_ck, k_pre_np);
         move_tile_window(k_dram_window, {0, kK0});
         __builtin_amdgcn_sched_barrier(0);
 
@@ -333,7 +351,9 @@ struct BlockFmhaPipelineQRKSVSAsync
             {
                 static_for<0, k0_loops - 1, 1>{}([&](auto i_k0) {
                     async_load_tile_raw(k_lds_store(number<LdsSeq.at(number<i_k0 + 1>{})>{}),
-                                        k_dram_window);
+                                        k_dram_window,
+                                        k_oob_ck,
+                                        k_pre_np);
                     if constexpr(i_k0 < k0_loops - 1)
                         move_tile_window(k_dram_window, {0, kK0});
 
@@ -637,16 +657,13 @@ struct BlockFmhaPipelineQRKSVSAsync
             {
                 // move K tile windows
                 move_tile_window(k_dram_block_window, {kN0, 0});
-                k_dram_window =
-                    make_tile_window(k_dram_block_window.get_bottom_tensor_view(),
-                                     k_dram_block_window.get_window_lengths(),
-                                     k_dram_block_window.get_window_origin(),
-                                     Policy::template MakeKDramTileDistribution<Problem>());
+                k_dram_window.set_window_origin(k_dram_block_window.get_window_origin());
 
                 if constexpr(k1_loops >= 2 &&
                              LdsSeq.at(number<0>{}) == LdsSeq.at(number<k0_loops + k1_loops - 2>{}))
                     __builtin_amdgcn_s_barrier();
-                async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), k_dram_window);
+                async_load_tile_raw(
+                    k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, k_oob_ck, k_pre_np);
                 move_tile_window(k_dram_window, {0, kK0});
             }
             // tail