TL/UCP: topo aware ring algo for reduce_scatter

Replace the default ring reduce_scatter with a topo aware
multi ring implementation that uses team->cuda_ring to route data
along NVLink optimal paths (up to 8 parallel rings).

Algorithm changes:
- Ring rank, peer, and block indices are now derived from the
  cuda_ring topology pattern instead of flat team rank ordering.
- Each ring handles its own sub block slice, with per ring GPU
  reductions via the executor before forwarding to the next peer.
- Scratch buffer management simplified to a single mc_alloc/free
  per task lifetime (removed fragmentation logic).

Signed-off-by: Juee Himalbhai Desai <jueehimalbha@nvidia.com>

Author: Juee14Desai

src/components/tl/ucp/reduce_scatter/reduce_scatter.c

@@ -7,6 +7,9 @@
 #include "tl_ucp.h"
 #include "reduce_scatter.h"
 #include "utils/ucc_coll_utils.h"
+#include "utils/ucc_string.h"
+
+#define REDUCE_SCATTER_MAX_PATTERN_SIZE 256
 
 ucc_base_coll_alg_info_t
     ucc_tl_ucp_reduce_scatter_algs[UCC_TL_UCP_REDUCE_SCATTER_ALG_LAST + 1] = {
@@ -20,3 +23,50 @@ ucc_base_coll_alg_info_t
              .desc = "recursive k-ing with arbitrary radix"},
         [UCC_TL_UCP_REDUCE_SCATTER_ALG_LAST] = {
             .id = 0, .name = NULL, .desc = NULL}};
+
+char *ucc_tl_ucp_reduce_scatter_score_str_get(ucc_tl_ucp_team_t *team)
+{
+    int                   max_size = REDUCE_SCATTER_MAX_PATTERN_SIZE;
+    char                 *str      = ucc_malloc(max_size * sizeof(char));
+    ucc_tl_ucp_context_t *ctx      = UCC_TL_UCP_TEAM_CTX(team);
+    uint64_t              cuda_types =
+        ctx->ucp_memory_types &
+        (UCC_BIT(UCC_MEMORY_TYPE_CUDA) |
+         UCC_BIT(UCC_MEMORY_TYPE_CUDA_MANAGED));
+    uint64_t  non_cuda_types = ctx->ucp_memory_types & (~cuda_types);
+    char     *non_cuda_str;
+    char     *cuda_str;
+
+    if (team->cuda_ring && cuda_types) {
+        cuda_str = ucc_malloc(max_size * sizeof(char));
+        ucc_mtype_map_to_str(cuda_types, ",", cuda_str, max_size);
+        if (non_cuda_types) {
+            non_cuda_str = ucc_malloc(max_size * sizeof(char));
+            ucc_mtype_map_to_str(non_cuda_types, ",", non_cuda_str, max_size);
+            ucc_snprintf_safe(str, max_size,
+                "reduce_scatter:0-4k:@%d"
+                "#reduce_scatter:4k-inf:%s:@%d"
+                "#reduce_scatter:4k-inf:%s:@%d",
+                UCC_TL_UCP_REDUCE_SCATTER_ALG_KNOMIAL,
+                cuda_str, UCC_TL_UCP_REDUCE_SCATTER_ALG_RING,
+                non_cuda_str, UCC_TL_UCP_REDUCE_SCATTER_ALG_KNOMIAL);
+            ucc_free(cuda_str);
+            ucc_free(non_cuda_str);
+            return str;
+        }
+        ucc_snprintf_safe(str, max_size,
+            "reduce_scatter:0-4k:@%d"
+            "#reduce_scatter:4k-inf:%s:@%d"
+            "#reduce_scatter:4k-inf:@%d",
+            UCC_TL_UCP_REDUCE_SCATTER_ALG_KNOMIAL,
+            cuda_str, UCC_TL_UCP_REDUCE_SCATTER_ALG_RING,
+            UCC_TL_UCP_REDUCE_SCATTER_ALG_KNOMIAL);
+        ucc_free(cuda_str);
+        return str;
+    }
+
+    ucc_snprintf_safe(str, max_size,
+                      UCC_TL_UCP_REDUCE_SCATTER_DEFAULT_ALG_SELECT_STR,
+                      UCC_TL_UCP_REDUCE_SCATTER_ALG_KNOMIAL);
+    return str;
+}

src/components/tl/ucp/reduce_scatter/reduce_scatter.h

@@ -19,7 +19,9 @@ extern ucc_base_coll_alg_info_t
     ucc_tl_ucp_reduce_scatter_algs[UCC_TL_UCP_REDUCE_SCATTER_ALG_LAST + 1];
 
 #define UCC_TL_UCP_REDUCE_SCATTER_DEFAULT_ALG_SELECT_STR                       \
-    "reduce_scatter:@ring"
+    "reduce_scatter:@%d"
+
+char *ucc_tl_ucp_reduce_scatter_score_str_get(ucc_tl_ucp_team_t *team);
 
 static inline int ucc_tl_ucp_reduce_scatter_alg_from_str(const char *str)
 {
@@ -48,4 +50,9 @@ ucc_status_t
 ucc_tl_ucp_reduce_scatter_ring_init(ucc_base_coll_args_t *coll_args,
                                     ucc_base_team_t *     team,
                                     ucc_coll_task_t **    task_h);
+
+ucc_status_t ucc_tl_ucp_reduce_scatter_ring_init_common(
+    ucc_tl_ucp_task_t *task);
+
+void ucc_tl_ucp_reduce_scatter_ring_progress(ucc_coll_task_t *coll_task);
 #endif