update

Author: taozha2

include/cute/arch/xe_copy_1B.hpp

@@ -484,13 +484,35 @@ struct XE_2D_U8x32x8_LD_T {
 #if defined(SYCL_INTEL_TARGET)
     static_assert(sizeof(T) == 1, "Expected T to have size 1");
     *reinterpret_cast<intel::ushort8 *>(dst) =
-       __builtin_IB_subgroup_block_read_cacheopts_transpose_u8_m32k8(
+        __builtin_IB_subgroup_block_read_cacheopts_transpose_u8_m32k8(
           (intptr_t)(baseoffset), width - 1, height - 1, pitch - 1, coord);
 #else
     CUTE_INVALID_CONTROL_PATH("Trying to use block loads on non-PVC hardware");
 #endif
   }
 };
+
+struct XE_2D_U8x16x32_LD_T {
+  using BlockShape = Shape<_32, _16>;
+  using inst_dtype = uint32_t;
+  static constexpr bool is_transpose = true;
+
+  template <class T>
+  CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
+                                    int height, int pitch, intel::coord_t coord,
+                                    T *dst) {
+#if defined(SYCL_INTEL_TARGET)
+    static_assert(sizeof(T) == 1, "Expected T to have size 2");
+    *reinterpret_cast<intel::uint8 *>(dst) =
+        __builtin_IB_subgroup_block_read_flat_transpose_u32_k8(
+          (intptr_t)(baseoffset), width - 1, height - 1, pitch - 1, coord);
+#else
+    CUTE_INVALID_CONTROL_PATH("Trying to use block loads on non-PVC hardware");
+#endif
+  }
+};
+
+
 struct XE_2D_U4x16x16_LD_T {
   using BlockShape = Shape<_16, _16>;
   using inst_dtype = uint32_t;

include/cute/atom/copy_traits_xe.hpp

@@ -1417,7 +1417,7 @@ struct Copy_Traits_<XE_2D_TF32x8x8_LD_T, args_t...>
 };
 
 template <class... args_t>
-struct Copy_Traits<XE_2D_U32x1x16_LD_N, args_t...>
+struct Copy_Traits_<XE_2D_U32x1x16_LD_N, args_t...>
     : XE_2D_LD_Unpack<XE_2D_U32x1x16_LD_N, args_t...> {
   using ThrID = Layout<_16>;
   // Map from (src-thr,src-val) to bit
@@ -1707,8 +1707,8 @@ struct Copy_Traits_<XE_2D_U16x16x16_LD_T, args_t...>
 };
 
 template <class... args_t>
-struct Copy_Traits_<XE_2D_U8x32x16_LD_T, args_t...>
-    : XE_2D_LD_Unpack<XE_2D_U8x32x16_LD_T, args_t...> {
+struct Copy_Traits_<XE_2D_U8x16x32_LD_T, args_t...>
+    : XE_2D_LD_Unpack<XE_2D_U8x16x32_LD_T, args_t...> {
   using ThrID = Layout<_16>;
   // Map from (src-thr,src-val) to bit
   using SrcLayout = Layout<Shape <_16,_16>,
@@ -1721,7 +1721,7 @@ struct Copy_Traits_<XE_2D_U8x32x16_LD_T, args_t...>
 
   template <class... ArgT>
   Copy_Traits_(ArgT... args)
-      : XE_2D_LD_Unpack<XE_2D_U8x32x16_LD_T, args_t...>(args...) {}
+      : XE_2D_LD_Unpack<XE_2D_U8x16x32_LD_T, args_t...>(args...) {}
 };
 
 template <class... args_t>
@@ -1743,7 +1743,7 @@ struct Copy_Traits_<XE_2D_U8x32x8_LD_T, args_t...>
 };
 
 template <class... args_t>
-struct Copy_Traits<XE_2D_U8x32x4_LD_T, args_t...>
+struct Copy_Traits_<XE_2D_U8x32x4_LD_T, args_t...>
     : XE_2D_LD_Unpack<XE_2D_U8x32x4_LD_T, args_t...> {
   using ThrID = Layout<_16>;
   // Map from (src-thr,src-val) to bit
@@ -1756,10 +1756,9 @@ struct Copy_Traits<XE_2D_U8x32x4_LD_T, args_t...>
   using RefLayout = DstLayout;
 
   template <class... ArgT>
-  Copy_Traits(ArgT... args)
+  Copy_Traits_(ArgT... args)
       : XE_2D_LD_Unpack<XE_2D_U8x32x4_LD_T, args_t...>(args...) {}
 };
-};
 
 // template<class... args_t>
 // struct Copy_Traits<XE_2D_U32x16x1_LD_T, args_t...>
@@ -2307,6 +2306,7 @@ COPY_TRAIT_LD_DEF(XE_2D_U8x16x64_LD_N)
 COPY_TRAIT_LD_DEF(XE_2D_U8x32x64_LD_N)
 COPY_TRAIT_LD_DEF(XE_2D_U8x32x8_LD_T)
 COPY_TRAIT_LD_DEF(XE_2D_U8x32x4_LD_T)
+COPY_TRAIT_LD_DEF(XE_2D_U8x16x32_LD_T)
 COPY_TRAIT_LD_DEF(XE_2D_U16x1x16_LD_N)
 COPY_TRAIT_LD_DEF(XE_2D_U16x2x16_LD_N)
 COPY_TRAIT_LD_DEF(XE_2D_U16x4x16_LD_N)