/* before */
static inline void
ucc_knomial_pattern_init(ucc_rank_t size, ucc_rank_t rank, ucc_kn_radix_t radix,
 ucc_knomial_pattern_t *p)
{
 ucc_knomial_pattern_init_impl(size, rank, radix, p, 0, 1);
}
/* after */
static inline void ucc_knomial_pattern_init(
 ucc_rank_t size, ucc_rank_t rank, ucc_kn_radix_t radix,
 ucc_knomial_pattern_t *p)
{
 ucc_knomial_pattern_init_impl(size, rank, radix, p, 0, 1);
}

/* before */
 status =
 ucc_mc_memcpy(rbuf, scratch_header->addr, tsize * data_size,
 rmem, UCC_MEMORY_TYPE_HOST);
 if (ucc_unlikely(status != UCC_OK)) {
 tl_error(UCC_TASK_LIB(task),
 "failed to copy from scratch buffer to dst");
 task->super.status = status;
 return;
 }
/* after */
 status = ucc_mc_memcpy(
 rbuf,
 scratch_header->addr,
 tsize * data_size,
 rmem,
 UCC_MEMORY_TYPE_HOST);
 if (ucc_unlikely(status != UCC_OK)) {
 tl_error(
 UCC_TASK_LIB(task),
 "failed to copy from scratch buffer to dst");
 task->super.status = status;
 return;
 }

/* before */
/**
 * Computes actual radix at current iteration
 * @param [in] p ucc_knomial_pattern
 * @return radix
 */
static inline
ucc_rank_t ucc_kn_compute_step_radix(ucc_knomial_pattern_t *p);
/* after */
/**
 * Computes actual radix at current iteration
 * @param [in] p ucc_knomial_pattern
 * @return radix
 */
static inline ucc_rank_t ucc_kn_compute_step_radix(ucc_knomial_pattern_t *p)

/*with inline comments*/
typedef struct ucc_knomial_pattern {
 ucc_kn_radix_t radix; /* knomial tree radix */
 uint8_t type; /* pattern type */
 uint8_t iteration; /* current iteration */
 uint8_t n_iters; /* number of iterations in knomial algorithm */
 uint8_t
 pow_radix_sup; /* smallest integer N such that (radix ** N) >= size */
 uint8_t node_type; /* type of current rank: BASE, PROXY or EXTRA */
 uint8_t backward; /* boolean, iteration direction */
 ucc_rank_t
 radix_pow; /* power of radix for current algorithm iteration
 * forward: initial value is 1
 * backward: initial valus is full_pow_size if have >1 full subtrees, OR
 * (full_pow_size / radix) otherwise
 */
 ucc_rank_t
 full_pow_size; /* largest power of radix <= size. It is equal to
 * (radix ** pow_radix_sup) if (radix ** pow_radix_sup) == size, OR
 * (radix ** (_pow_radix_sup - 1)) otherwise
 */
 ucc_rank_t size; /* total number of ranks */
 ucc_rank_t rank; /* process rank */
 ucc_rank_t n_extra; /* number of "extra" ranks to be served by "proxies" */
 size_t block_size_counts;
 size_t count; /* collective buffer size */
 ucc_count_t *counts;
 ucc_rank_t block_size;
 ptrdiff_t block_offset;
 int is64;
} ucc_knomial_pattern_t;
/* no inline comments */
typedef struct ucc_knomial_pattern {
 /* knomial tree radix */
 ucc_kn_radix_t radix;
 /* pattern type */
 uint8_t type;
 /* current iteration */
 uint8_t iteration;
 /* number of iterations in knomial algorithm */
 uint8_t n_iters;
 /* smallest integer N such that (radix ** N) >= size */
 uint8_t pow_radix_sup;
 /* type of current rank: BASE, PROXY or EXTRA */
 uint8_t node_type;
 /* boolean, iteration direction */
 uint8_t backward;
 /* power of radix for current algorithm iteration
 * forward: initial value is 1
 * backward: initial valus is full_pow_size if have >1 full subtrees, OR
 * (full_pow_size / radix) otherwise
 */
 ucc_rank_t radix_pow;
 /* largest power of radix <= size. It is equal to
 * (radix ** pow_radix_sup) if (radix ** pow_radix_sup) == size, OR
 * (radix ** (_pow_radix_sup - 1)) otherwise
 */
 ucc_rank_t full_pow_size;
 /* total number of ranks */
 ucc_rank_t size;
 /* process rank */
 ucc_rank_t rank;
 /* number of "extra" ranks to be served by "proxies" */
 ucc_rank_t n_extra;
 size_t block_size_counts;
 /* collective buffer size */
 size_t count;
 ucc_count_t *counts;
 ucc_rank_t block_size;
 ptrdiff_t block_offset;
 int is64;
} ucc_knomial_pattern_t;