v3.9 update (#2203)

* v3.9 update

* voidD

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>

include/cute/arch/mma_sm120.hpp

@@ -3245,7 +3245,7 @@ rr_blockscaled_op_selector_sm120()
 {
   if constexpr (UseF8F6F4) {
     return SM120::BLOCKSCALED::SM120_16x8x32_TN_VS<ElementA, ElementB, ElementC, ElementSF, SFVecSize>{};
-  } 
+  }
   else{
     return SM120::BLOCKSCALED::SM120_16x8x64_TN_VS<ElementA, ElementB, ElementC, ElementSF, SFVecSize>{};
   }

include/cute/arch/tmem_allocator_sm100.hpp

@@ -57,7 +57,7 @@ public:
    * @pre Must never be issued by more than one warp at the same time.
    * @pre For repeated allocations, the same warp must be used to issue all allocations.
   **/
-  __device__ void
+  CUTE_HOST_DEVICE void
   allocate(int num_columns, uint32_t* dst_ptr) {
   #if defined(CUTE_ARCH_TCGEN05_TMEM_ENABLED)
     uint32_t dst_intptr = cute::cast_smem_ptr_to_uint(dst_ptr);
@@ -116,7 +116,7 @@ public:
    * @pre For repeated allocations, the same warp must be used to issue all allocations.
    * @pre The 2 warps from participating CTAs have the same logical warp ID.
   **/
-  __device__ void
+  CUTE_HOST_DEVICE void
   allocate(int num_columns, uint32_t* dst_ptr) {
   #if defined(CUTE_ARCH_TCGEN05_TMEM_ENABLED)
     uint32_t dst_intptr = cute::cast_smem_ptr_to_uint(dst_ptr);

include/cute/arch/util.hpp

@@ -88,7 +88,7 @@ namespace cute
 {
 
 /// CUTE helper to cast SMEM pointer to unsigned
-CUTE_DEVICE
+CUTE_HOST_DEVICE
 uint32_t
 cast_smem_ptr_to_uint(void const* const ptr)
 {