diff --git a/include/common/core/memory.hpp b/include/common/core/memory.hpp
index b9cfa2542..ed75cf935 100644
--- a/include/common/core/memory.hpp
+++ b/include/common/core/memory.hpp
@@ -477,16 +477,15 @@ __XETLA_API xetla_vector<T, N> xetla_load_global(
     return ret.xetla_format<T>();
   } else if constexpr (BlockWidth * sizeof(T) < sizeof(uint32_t)) {
     constexpr auto scale_factor = sizeof(uint32_t) / sizeof(T);
-    xetla_vector<uint32_t, N> ret = __ESIMD_ENS::lsc_load_2d<
+    xetla_vector<uint32_t, N> ret = xetla_load_global<
         uint32_t,
         BlockWidth,
         BlockHeight,
         NBlocks,
         Transposed,
         Transformed,
-        gpu::xetla::detail::get_cache_hint(L1H),
-        gpu::xetla::detail::get_cache_hint(L2H),
-        N>(
+        L1H,
+        L2H>(
         reinterpret_cast<const uint32_t*>(Ptr),
         SurfaceWidth,
         SurfaceHeight,
@@ -505,7 +504,7 @@ __XETLA_API xetla_vector<T, N> xetla_load_global(
         Transformed,
         gpu::xetla::detail::get_cache_hint(L1H),
         gpu::xetla::detail::get_cache_hint(L2H),
-        N>(Ptr, SurfaceWidth, SurfaceHeight, SurfacePitch, X, Y);
+        N>(Ptr, SurfaceWidth - 1, SurfaceHeight - 1, SurfacePitch - 1, X, Y);
   }
 }
 
@@ -788,7 +787,7 @@ __XETLA_API void xetla_store_global(
         BlockHeight,
         gpu::xetla::detail::get_cache_hint(L1H),
         gpu::xetla::detail::get_cache_hint(L2H)>(
-        Ptr, SurfaceWidth, SurfaceHeight, SurfacePitch, X, Y, Vals);
+        Ptr, SurfaceWidth - 1, SurfaceHeight - 1, SurfacePitch - 1, X, Y, Vals);
   }
 }
 /// template <typename T, int N, int VS = 1, typename OffsetT,
diff --git a/include/subgroup/tile/impl/load_xe.hpp b/include/subgroup/tile/impl/load_xe.hpp
index 948505031..6da99f07c 100644
--- a/include/subgroup/tile/impl/load_xe.hpp
+++ b/include/subgroup/tile/impl/load_xe.hpp
@@ -95,7 +95,7 @@ tile_load(tile_t& tile, payload_t& payload) {
 
   static constexpr reg_layout reg_layout_ = tile_desc::register_layout;
   static constexpr bool is_vnni_reverse =
-      payload_t::mem_dword_qword_transpose &&
+      payload_t::mem_transpose_dtype_less4bytes &&
       ((reg_layout_ == reg_layout::tiled) ||
        (reg_layout_ == reg_layout::transpose_tiled));
   static constexpr bool reg_transpose = tile_desc::reg_transpose;
@@ -180,6 +180,7 @@ tile_load(tile_t& tile, payload_t& payload) {
 #pragma unroll
   for (uint32_t i = 0; i < num_block_y; ++i) {
     constexpr uint32_t load_block_elems = block_elems * arr_len;
+    int offset_y = i * block_size_y;
     auto payload_row =
         payload_2d.xetla_select<num_block_x, 1, 16, 1>(i * num_block_x, 0);
     detail::reset_tile_desc_core<
@@ -191,6 +192,7 @@ tile_load(tile_t& tile, payload_t& payload) {
         mem_transpose>(payload_row);
 #pragma unroll
     for (uint32_t j = 0; j < num_block_x; j += arr_len) {
+      uint32_t offset_x = j * block_size_x;
       xetla_tdescriptor tdesc = payload_row.row(j);
       auto reg_blk = tile.reg.xetla_select<load_block_elems, 1>(
           (i * num_block_x + j) * block_elems);
@@ -201,21 +203,13 @@ tile_load(tile_t& tile, payload_t& payload) {
       xetla_vector<dtype, tmp_size> reg_tmp;
 #pragma unroll
       for (uint32_t ii = 0; ii < block_size_y / ld_blk_size_y; ++ii) {
+        offset_y += ld_blk_size_y;
         constexpr uint32_t load_elems = ld_blk_size_y * block_size_x * arr_len;
 
-        // reg_tmp.xetla_format<native_type_t<load_dtype>>() =
-        // xetla_tload_global<
-        //     load_dtype,
-        //     ld_blk_height * block_size_x * arr_len / scale_factor,
-        //     L1,
-        //     L2,
-        //     trans,
-        //     mem_transform,
-        //     arch_tag>(tdesc);
         reg_tmp.xetla_format<native_type_t<load_dtype>>() = xetla_load_global<
             native_type_t<load_dtype>,
-            (mem_transpose ? ld_blk_size_y : block_size_x) / scale_factor,
-            (mem_transpose ? block_size_x : ld_blk_size_y),
+            (trans ? ld_blk_size_y : block_size_x) / scale_factor,
+            (trans ? block_size_x : ld_blk_size_y),
             // block_size_x / scale_factor,
             // ld_blk_size_y,
             arr_len,
@@ -223,13 +217,15 @@ tile_load(tile_t& tile, payload_t& payload) {
             mem_transform,
             L1,
             L2>(
-            (native_type_t<load_dtype>*)::gpu::xetla::detail::
-                xetla_get_tensor_base_address(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_x(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_y(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_pitch_x(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc));
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
+            mem_transpose
+                // ? ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc)
+                ? (payload.offset_x + offset_y / scale_factor)
+                : (payload.offset_x + offset_x / scale_factor),
+            payload.offset_y + (mem_transpose ? offset_x : offset_y));
         if constexpr (reg_transpose && trans) {
           reg_blk.xetla_select<load_elems, 1>(ii * load_elems)
               .xetla_format<native_type_t<load_dtype>>() =
@@ -246,7 +242,6 @@ tile_load(tile_t& tile, payload_t& payload) {
         } else {
           reg_blk.xetla_select<tmp_size, 1>(ii * tmp_size) = reg_tmp;
         }
-
         if constexpr (mem_transpose) {
           xetla_update_tdesc_offsetx(
               tdesc.xetla_format<uint32_t>(), ld_blk_size_y / scale_factor);
@@ -284,11 +279,10 @@ tile_load(tile_t& tile, payload_t& payload) {
             mem_transform,
             L1,
             L2>(
-            (native_type_t<load_dtype>*)::gpu::xetla::detail::
-                xetla_get_tensor_base_address(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_x(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_y(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_pitch_x(tdesc),
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
             ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
             ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc));
         // xetla_tload_global<
@@ -345,11 +339,10 @@ tile_load(tile_t& tile, payload_t& payload) {
             mem_transform,
             L1,
             L2>(
-            (native_type_t<load_dtype>*)::gpu::xetla::detail::
-                xetla_get_tensor_base_address(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_x(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_y(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_pitch_x(tdesc),
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
             ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
             ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc));
         //  xetla_tload_global<
@@ -412,11 +405,10 @@ tile_load(tile_t& tile, payload_t& payload) {
             mem_transform,
             L1,
             L2>(
-            (native_type_t<load_dtype>*)::gpu::xetla::detail::
-                xetla_get_tensor_base_address(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_x(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_y(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_pitch_x(tdesc),
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
             ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
             ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc));
         // xetla_tload_global<
diff --git a/include/subgroup/tile/impl/payload_xe.hpp b/include/subgroup/tile/impl/payload_xe.hpp
index ea47e9e15..408e79d18 100644
--- a/include/subgroup/tile/impl/payload_xe.hpp
+++ b/include/subgroup/tile/impl/payload_xe.hpp
@@ -78,11 +78,11 @@ struct mem_payload_t<
   static constexpr bool mem_transform = (sizeof(dtype) <= 2) && !trans &&
       (register_layout == reg_layout::vnni_tiled ||
        register_layout == reg_layout::vnni_tiled_col_major);
-  static constexpr bool mem_dword_qword_transpose =
+  static constexpr bool mem_transpose_dtype_less4bytes =
       (sizeof(dtype) < 4) && trans;
 
   using mem_dtype = typename std::
-      conditional<mem_dword_qword_transpose, uint32_t, dtype>::type;
+      conditional_t<mem_transpose_dtype_less4bytes, uint32_t, dtype>;
   static constexpr uint32_t scale_factor = sizeof(mem_dtype) / sizeof(dtype);
   mem_dtype* base_ptr;
   uint32_t surface_width;
@@ -111,8 +111,9 @@ struct mem_payload_t<
     this->surface_height =
         (mem_transpose ? mem_desc.shape.x : mem_desc.shape.y);
     this->surface_pitch = mem_desc.shape.stride * sizeof(dtype);
-    this->offset_x = mem_desc.coord.x;
-    this->offset_y = mem_desc.coord.y;
+    this->offset_x =
+        (mem_transpose ? mem_desc.coord.y : mem_desc.coord.x) / scale_factor;
+    this->offset_y = mem_transpose ? mem_desc.coord.x : mem_desc.coord.y;
 
     xetla_tdescriptor base_tdesc = mem_desc.get_tdesc();
     int32_t offset = gpu::xetla::detail::xetla_get_tensor_offset_x(base_tdesc) /
@@ -130,13 +131,14 @@ struct mem_payload_t<
       int32_t surface_offset_x = 0,
       int32_t surface_offset_y = 0) {
     this->base_ptr = (mem_dtype*)p;
-    this->surface_width = surface_width;
+    this->surface_width = surface_width * sizeof(dtype);
     this->surface_height = surface_height;
-    this->surface_pitch = surface_pitch;
-    this->offset_x = surface_offset_x;
+    this->surface_pitch = surface_pitch * sizeof(dtype);
+    this->offset_x = surface_offset_x / scale_factor;
     this->offset_y = surface_offset_y;
 
     xetla_tdescriptor base_tdesc;
+
     xetla_fill_tdesc(
         base_tdesc.xetla_format<uint32_t>(),
         p,
@@ -155,8 +157,9 @@ struct mem_payload_t<
     this->surface_height =
         (mem_transpose ? mem_desc.shape.x : mem_desc.shape.y);
     this->surface_pitch = mem_desc.shape.stride * sizeof(dtype);
-    this->offset_x = mem_desc.coord.x;
-    this->offset_y = mem_desc.coord.y;
+    this->offset_x =
+        (mem_transpose ? mem_desc.coord.y : mem_desc.coord.x) / scale_factor;
+    this->offset_y = (mem_transpose ? mem_desc.coord.x : mem_desc.coord.y);
 
     xetla_tdescriptor base_tdesc = mem_desc.get_tdesc();
     int32_t offset = gpu::xetla::detail::xetla_get_tensor_offset_x(base_tdesc) /
@@ -181,11 +184,11 @@ struct mem_payload_t<
       uint32_t surface_pitch,
       int32_t surface_offset_x = 0,
       int32_t surface_offset_y = 0) {
-    this->base_ptr = (mem_dtype)p;
-    this->surface_width = surface_width;
+    this->base_ptr = (mem_dtype*)p;
+    this->surface_width = surface_width * sizeof(dtype);
     this->surface_height = surface_height;
-    this->surface_pitch = surface_pitch;
-    this->offset_x = surface_offset_x;
+    this->surface_pitch = surface_pitch * sizeof(dtype);
+    this->offset_x = surface_offset_x / scale_factor;
     this->offset_y = surface_offset_y;
 
     xetla_tdescriptor base_tdesc;
@@ -1708,11 +1711,12 @@ struct prefetch_payload_t<
         reg_layout_>,
     num_coop_sg_,
     arch_tag_,
-    std::enable_if_t<(!arch_has_2d_load_store<arch_tag_>)&&(
-        ((block_size_y_ != 1 || tile_size_y_ != 1) &&
-         mem_layout_ == mem_layout::row_major) ||
-        ((block_size_x_ != 1 || tile_size_x_ != 1) &&
-         mem_layout_ == mem_layout::col_major))>> {
+    std::enable_if_t<
+        (!arch_has_2d_load_store<arch_tag_>) &&
+        (((block_size_y_ != 1 || tile_size_y_ != 1) &&
+          mem_layout_ == mem_layout::row_major) ||
+         ((block_size_x_ != 1 || tile_size_x_ != 1) &&
+          mem_layout_ == mem_layout::col_major))>> {
   using dtype = native_type_t<dtype_>;
   using mem_desc_t =
       mem_desc_t<dtype_, mem_layout_, mem_space::global, alignment_>;
@@ -1968,9 +1972,10 @@ struct prefetch_payload_t<
         reg_layout_>,
     num_coop_sg_,
     arch_tag_,
-    std::enable_if_t<(arch_has_2d_load_store<arch_tag_>)&&(
-        ((tile_size_y_ != 1) && mem_layout_ == mem_layout::row_major) ||
-        ((tile_size_x_ != 1) && mem_layout_ == mem_layout::col_major))>> {
+    std::enable_if_t<
+        (arch_has_2d_load_store<arch_tag_>) &&
+        (((tile_size_y_ != 1) && mem_layout_ == mem_layout::row_major) ||
+         ((tile_size_x_ != 1) && mem_layout_ == mem_layout::col_major))>> {
   using dtype = dtype_;
   using mem_desc_t =
       mem_desc_t<dtype_, mem_layout_, mem_space::global, alignment_>;
diff --git a/include/subgroup/tile/impl/store_xe.hpp b/include/subgroup/tile/impl/store_xe.hpp
index 813052455..83f07653b 100644
--- a/include/subgroup/tile/impl/store_xe.hpp
+++ b/include/subgroup/tile/impl/store_xe.hpp
@@ -171,10 +171,10 @@ tile_store(tile_t& tile, payload_t& payload) {
             st_block_size_y,
             L1,
             L2>(
-            (dtype*)::gpu::xetla::detail::xetla_get_tensor_base_address(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_x(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_width_y(tdesc),
-            ::gpu::xetla::detail::xetla_get_tensor_pitch_x(tdesc),
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
             ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
             ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc),
             st_blk);
@@ -194,12 +194,25 @@ tile_store(tile_t& tile, payload_t& payload) {
             (block_size_x * arr_len - 1) | ((blk_remained_y - 1) << 8);
         gpu::xetla::detail::xetla_set_block_widthx_widthy_arrlen(
             tdesc.xetla_format<uint32_t>(), block_widthx_widthy_arrlen);
-        xetla_tstore_global<
+        // xetla_tstore_global<
+        //     dtype,
+        //     blk_remained_elems,
+        //     L1,
+        //     L2,
+        //     payload_t::arch_tag>(tdesc, st_blk);
+        xetla_store_global<
             dtype,
-            blk_remained_elems,
+            block_size_x * arr_len,
+            blk_remained_y,
             L1,
-            L2,
-            payload_t::arch_tag>(tdesc, st_blk);
+            L2>(
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
+            ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
+            ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc),
+            st_blk);
       }
     }
   }
@@ -244,8 +257,21 @@ tile_store(tile_t& tile, payload_t& payload) {
             remained_st_blk_size_y * block_size_x * arr_len;
         auto st_blk =
             combine_blk.xetla_select<store_elems, 1>(ii * store_elems);
-        xetla_tstore_global<dtype, store_elems, L1, L2, payload_t::arch_tag>(
-            tdesc, st_blk);
+        // xetla_tstore_global<dtype, store_elems, L1, L2, payload_t::arch_tag>(
+        //     tdesc, st_blk);
+        xetla_store_global<
+            dtype,
+            block_size_x * arr_len,
+            remained_st_blk_size_y,
+            L1,
+            L2>(
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
+            ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
+            ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc),
+            st_blk);
         xetla_update_tdesc_offsety(
             tdesc.xetla_format<uint32_t>(), remained_st_blk_size_y);
       }
@@ -263,12 +289,25 @@ tile_store(tile_t& tile, payload_t& payload) {
             (block_size_x * arr_len - 1) | ((final_st_blk_size_y - 1) << 8);
         gpu::xetla::detail::xetla_set_block_widthx_widthy_arrlen(
             tdesc.xetla_format<uint32_t>(), block_widthx_widthy_arrlen);
-        xetla_tstore_global<
+        // xetla_tstore_global<
+        //     dtype,
+        //     final_store_elems,
+        //     L1,
+        //     L2,
+        //     payload_t::arch_tag>(tdesc, st_blk);
+        xetla_store_global<
             dtype,
-            final_store_elems,
+            block_size_x * arr_len,
+            final_st_blk_size_y,
             L1,
-            L2,
-            payload_t::arch_tag>(tdesc, st_blk);
+            L2>(
+            payload.base_ptr,
+            payload.surface_width,
+            payload.surface_height,
+            payload.surface_pitch,
+            ::gpu::xetla::detail::xetla_get_tensor_offset_x(tdesc),
+            ::gpu::xetla::detail::xetla_get_tensor_offset_y(tdesc),
+            st_blk);
       }
     }
   }