refine code

codeplaysoftware · Jan 10, 2025 · 72e8400 · 72e8400
1 parent b8586e6
commit 72e8400
Show file tree

Hide file tree

Showing 12 changed files with 249 additions and 176 deletions.
diff --git a/include/cute/arch/xe_copy_2B.hpp b/include/cute/arch/xe_copy_2B.hpp
@@ -733,6 +733,8 @@ struct XE_2D_U16x16x8_LD_T {
   using BlockShape = Shape<_8, _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,
@@ -750,8 +752,10 @@ struct XE_2D_U16x16x8_LD_T {
 
 struct XE_2D_U16x16x16_LD_T {
   using BlockShape = Shape<_16, _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,

diff --git a/include/cute/arch/xe_copy_4B.hpp b/include/cute/arch/xe_copy_4B.hpp
@@ -599,6 +599,8 @@ struct XE_2D_TF32x32x16_LD_N {
 
 
 struct XE_2D_U32x16x1_LD_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,
@@ -617,6 +619,8 @@ struct XE_2D_U32x16x1_LD_T {
 struct XE_2D_U32x16x2_LD_T {
   using BlockShape = Shape<_2, _16>;
 
+  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,
@@ -635,6 +639,8 @@ struct XE_2D_U32x16x2_LD_T {
 struct XE_2D_U32x16x4_LD_T {
   using BlockShape = Shape<_4, _16>;
 
+  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,
@@ -653,6 +659,8 @@ struct XE_2D_U32x16x4_LD_T {
 struct XE_2D_U32x16x8_LD_T {
   using BlockShape = Shape<_8, _16>;
 
+  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,

diff --git a/include/cute/arch/xe_copy_8B.hpp b/include/cute/arch/xe_copy_8B.hpp
@@ -81,6 +81,8 @@ SYCL_DEVICE_OCL(intel::ulong4 intel_sub_group_block_read_transpose_64b_8r4c(
 namespace cute
 {
 struct XE_2D_U64x8x1_LD_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,
@@ -97,6 +99,8 @@ struct XE_2D_U64x8x1_LD_T {
 };
 
 struct XE_2D_U64x8x2_LD_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,
@@ -113,6 +117,8 @@ struct XE_2D_U64x8x2_LD_T {
 };
 
 struct XE_2D_U64x8x4_LD_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,