TL/UCP: topo aware ring algo for allreduce

Add a new monolithic allreduce ring that fuses reduce_scatter and
allgather into a single task using team->cuda_ring for topo aware
multi ring transfers (up to 8 parallel rings).

Algorithm changes:
- each step receives into scratch, reduces with the local dst block
  via GPU executor, then forwards the accumulated result to the
  next ring peer.
- in-place ring allgather distributes all fully reduced blocks
  across ranks.
- Both process runs in one progress function, with tagged send/recv
  counters reset at the algo transition.
- Auto selected for CUDA memory >4KB when cuda_ring is available.

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

Author: Juee14Desai

src/components/tl/ucp/Makefile.am

@@ -47,7 +47,8 @@ allreduce =                           \
 	allreduce/allreduce_sliding_window.h       \
 	allreduce/allreduce_sliding_window.c       \
 	allreduce/allreduce_sliding_window_setup.c \
-	allreduce/allreduce_dbt.c
+	allreduce/allreduce_dbt.c                  \
+	allreduce/allreduce_ring.c
 
 barrier =                     \
 	barrier/barrier.h         \

src/components/tl/ucp/allreduce/allreduce.c

@@ -7,6 +7,9 @@
 #include "tl_ucp.h"
 #include "allreduce.h"
 #include "utils/ucc_coll_utils.h"
+#include "utils/ucc_string.h"
+
+#define ALLREDUCE_MAX_PATTERN_SIZE 256
 
 ucc_base_coll_alg_info_t
     ucc_tl_ucp_allreduce_algs[UCC_TL_UCP_ALLREDUCE_ALG_LAST + 1] = {
@@ -29,9 +32,59 @@ ucc_base_coll_alg_info_t
             {.id   = UCC_TL_UCP_ALLREDUCE_ALG_SLIDING_WINDOW,
              .name = "sliding_window",
              .desc = "sliding window allreduce (optimized for running on DPU)"},
+        [UCC_TL_UCP_ALLREDUCE_ALG_RING] =
+            {.id   = UCC_TL_UCP_ALLREDUCE_ALG_RING,
+             .name = "ring",
+             .desc = "reduce-scatter ring followed by allgather ring "
+                     "(topology-aware, optimized for BW)"},
         [UCC_TL_UCP_ALLREDUCE_ALG_LAST] = {
             .id = 0, .name = NULL, .desc = NULL}};
 
+char *ucc_tl_ucp_allreduce_score_str_get(ucc_tl_ucp_team_t *team)
+{
+    int                   max_size = ALLREDUCE_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,
+                "allreduce:0-4k:@%d#allreduce:4k-inf:%s:@%d"
+                "#allreduce:4k-inf:%s:@%d",
+                UCC_TL_UCP_ALLREDUCE_ALG_KNOMIAL,
+                cuda_str, UCC_TL_UCP_ALLREDUCE_ALG_RING,
+                non_cuda_str, UCC_TL_UCP_ALLREDUCE_ALG_SRA_KNOMIAL);
+            ucc_free(cuda_str);
+            ucc_free(non_cuda_str);
+            return str;
+        }
+        ucc_snprintf_safe(str, max_size,
+            "allreduce:0-4k:@%d#allreduce:4k-inf:%s:@%d"
+            "#allreduce:4k-inf:@%d",
+            UCC_TL_UCP_ALLREDUCE_ALG_KNOMIAL,
+            cuda_str, UCC_TL_UCP_ALLREDUCE_ALG_RING,
+            UCC_TL_UCP_ALLREDUCE_ALG_SRA_KNOMIAL);
+        ucc_free(cuda_str);
+        return str;
+    }
+
+    ucc_snprintf_safe(str, max_size,
+                      UCC_TL_UCP_ALLREDUCE_DEFAULT_ALG_SELECT_STR,
+                      UCC_TL_UCP_ALLREDUCE_ALG_SRA_KNOMIAL);
+    return str;
+}
+
 ucc_status_t ucc_tl_ucp_allreduce_init(ucc_tl_ucp_task_t *task)
 {
     ucc_status_t status;

src/components/tl/ucp/allreduce/allreduce.h

@@ -13,6 +13,7 @@ enum {
     UCC_TL_UCP_ALLREDUCE_ALG_SRA_KNOMIAL,
     UCC_TL_UCP_ALLREDUCE_ALG_SLIDING_WINDOW,
     UCC_TL_UCP_ALLREDUCE_ALG_DBT,
+    UCC_TL_UCP_ALLREDUCE_ALG_RING,
     UCC_TL_UCP_ALLREDUCE_ALG_LAST
 };
 
@@ -21,7 +22,9 @@ extern ucc_base_coll_alg_info_t
 ucc_status_t ucc_tl_ucp_allreduce_init(ucc_tl_ucp_task_t *task);
 
 #define UCC_TL_UCP_ALLREDUCE_DEFAULT_ALG_SELECT_STR                            \
-    "allreduce:0-4k:@0#allreduce:4k-inf:@1"
+    "allreduce:0-4k:@0#allreduce:4k-inf:@%d"
+
+char *ucc_tl_ucp_allreduce_score_str_get(ucc_tl_ucp_team_t *team);
 
 #define CHECK_SAME_MEMTYPE(_args, _team)                                       \
     do {                                                                       \
@@ -77,6 +80,11 @@ ucc_status_t ucc_tl_ucp_allreduce_dbt_start(ucc_coll_task_t *task);
 
 ucc_status_t ucc_tl_ucp_allreduce_dbt_progress(ucc_coll_task_t *task);
 
+ucc_status_t
+ucc_tl_ucp_allreduce_ring_init(ucc_base_coll_args_t *coll_args,
+                               ucc_base_team_t      *team,
+                               ucc_coll_task_t     **task_h);
+
 static inline int ucc_tl_ucp_allreduce_alg_from_str(const char *str)
 {
     int i;