Separates mask and bias copy operations into distinct types

Splits the combined GmemTiledCopyMaskBias type into separate GmemTiledCopyMask and GmemTiledCopyBias types in both forward and backward kernel traits.

This separation improves code clarity and allows for independent handling of mask and bias copy operations, enabling more flexible memory access patterns and potential optimizations.

Author: algo-home

csrc/flash_dmattn/src/kernel_traits.h

@@ -183,7 +183,14 @@ struct Flash_fwd_kernel_traits : public Base {
             Layout<Shape<_1, _8>>{}
         )
     );      // Val layout, 8 vals per read
-    using GmemTiledCopyMaskBias = decltype(
+    using GmemTiledCopyMask = decltype(
+        make_tiled_copy(
+            Copy_Atom<AutoVectorizingCopyWithAssumedAlignment<128>, Element>{},
+            GmemLayoutAtom{},
+            Layout<Shape<_1, _8>>{}
+        )
+    );      // Val layout, 8 vals per read
+    using GmemTiledCopyBias = decltype(
         make_tiled_copy(
             Copy_Atom<AutoVectorizingCopyWithAssumedAlignment<128>, Element>{},
             GmemLayoutAtom{},
@@ -442,7 +449,14 @@ struct Flash_bwd_kernel_traits : public Base {
             Layout<Shape < _1, _8>>{}
         )
     );      // Val layout, 8 vals per store
-    using GmemTiledCopyMaskBias = decltype(
+    using GmemTiledCopyMask = decltype(
+        make_tiled_copy(
+            Copy_Atom<AutoVectorizingCopyWithAssumedAlignment<128>, elem_type>{},
+            GmemLayoutAtom{},
+            Layout<Shape<_1, _8>>{}
+        )
+    );      // Val layout, 8 vals per read
+    using GmemTiledCopyBias = decltype(
         make_tiled_copy(
             Copy_Atom<AutoVectorizingCopyWithAssumedAlignment<128>, elem_type>{},
             GmemLayoutAtom{},