Bine alltoall and alltoallv

config/algorithms/MPI/Cray-MPICH/alltoallv.json

@@ -0,0 +1,21 @@
+{
+  "default_mpich": {
+    "desc": "MPICH default algorithm selection",
+    "version": "8.0.0",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      }
+    ],
+    "selection": "auto",
+    "tags": [
+      "default"
+    ]
+  }
+}

config/algorithms/MPI/LibPico/alltoall.json

@@ -28,6 +28,36 @@
       "external"
     ]
   },
+  "bine_dh_over": {
+    "desc": "LibPico external Alltoall algorithm . This algorithm follows the distance-halving Bine butterfly communication pattern described in the Bine paper.",
+    "version": "1.0.0",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      },
+      {
+        "key": "comm_sz",
+        "conditions": [
+          {
+            "operator": "is_power_of_two",
+            "value": true
+          }
+        ]
+      }
+    ],
+    "selection": "pico",
+    "tags": [
+      "bine",
+      "external",
+      "distance-halving"
+    ]
+  },
   "pairwise_ompi_over": {
     "desc": "libpico pairwise algorithm, copied from Open MPI.",
     "version": "1.0.0",
@@ -57,4 +87,4 @@
       "external"
     ]
   }
-}
+}

config/algorithms/MPI/LibPico/alltoallv.json

@@ -0,0 +1,32 @@
+{
+  "bine_dh_over": {
+    "desc": "LibPico external Alltoall algorithm . This algorithm follows the distance-halving Bine butterfly communication pattern described in the Bine paper.",
+    "version": "1.0.0",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      },
+      {
+        "key": "comm_sz",
+        "conditions": [
+          {
+            "operator": "is_power_of_two",
+            "value": true
+          }
+        ]
+      }
+    ],
+    "selection": "pico",
+    "tags": [
+      "bine",
+      "external",
+      "distance-halving"
+    ]
+  }
+}

config/algorithms/MPI/MPICH/alltoallv.json

@@ -0,0 +1,21 @@
+{
+  "default_mpich": {
+    "desc": "MPICH default Alltoallv algorithm selection",
+    "version": "4.3.0",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      }
+    ],
+    "selection": "auto",
+    "tags": [
+      "default"
+    ]
+  }
+}

config/algorithms/MPI/Open-MPI/alltoallv.json

@@ -0,0 +1,66 @@
+{
+  "default_ompi": {
+    "desc": "Open MPI default Alltoallv algorithm selection.",
+    "version": "4.1.5",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      }
+    ],
+    "selection": 0,
+    "cuda_support": "yes",
+    "rocm_support": "ucx",
+    "tags": [
+      "ignore"
+    ]
+  },
+  "basic_linear_ompi": {
+    "desc": "Open MPI basic linear Alltoallv algorithm.",
+    "version": "4.1.5",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      }
+    ],
+    "selection": 1,
+    "cuda_support": "yes",
+    "rocm_support": "ucx",
+    "tags": [
+      "basic_linear",
+      "linear"
+    ]
+  },
+  "pairwise_ompi": {
+    "desc": "Open MPI pairwise Alltoallv algorithm.",
+    "version": "4.1.5",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      }
+    ],
+    "selection": 2,
+    "cuda_support": "yes",
+    "rocm_support": "ucx",
+    "tags": [
+      "pairwise"
+    ]
+  }
+}

config/algorithms/NCCL/LibPico/alltoallv.json

@@ -0,0 +1,53 @@
+{
+  "nccl_spreadout_alltoallv_over": {
+    "desc": "LibPico external NCCL spread-out Alltoallv algorithm. This implementation uses NCCL point-to-point ncclSend/ncclRecv operations arranged in size - 1 communication rounds.",
+    "version": "1.0.0",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      }
+    ],
+    "selection": "pico",
+    "tags": [
+      "alltoallv",
+      "nccl",
+      "cuda",
+      "gpu",
+      "spread-out",
+      "sendrecv",
+      "external"
+    ]
+  },
+
+  "nccl_fanout_alltoallv_over": {
+    "desc": "LibPico external NCCL fan-out Alltoallv algorithm. This implementation posts all ncclSend/ncclRecv operations inside a single ncclGroupStart/ncclGroupEnd region.",
+    "version": "1.0.0",
+    "constraints": [
+      {
+        "key": "count",
+        "conditions": [
+          {
+            "operator": ">=",
+            "value": "comm_sz"
+          }
+        ]
+      }
+    ],
+    "selection": "pico",
+    "tags": [
+      "alltoallv",
+      "nccl",
+      "cuda",
+      "gpu",
+      "fan-out",
+      "sendrecv",
+      "external"
+    ]
+  }
+}

config/parse_test.py

@@ -15,7 +15,7 @@
     "type": "object",
     "properties": {
         "libpico_version": {"type": "string", "pattern": "^\\d+\\.\\d+\\.\\d+$"},
-        "collective": {"type": "string", "enum": ["ALLREDUCE", "ALLTOALL", "ALLGATHER", "BCAST", "GATHER", "REDUCE", "REDUCE_SCATTER", "SCATTER"]},
+        "collective": {"type": "string", "enum": ["ALLREDUCE", "ALLTOALL","ALLTOALLV", "ALLGATHER", "BCAST", "GATHER", "REDUCE", "REDUCE_SCATTER", "SCATTER"]},
         "MPI_Op": {"type": "string"},
         "tags": {
             "type": "object",

config/test/alltoallv.json

@@ -0,0 +1,16 @@
+{
+  "libpico_version": "1.0.0",
+  "collective": "ALLTOALLV",
+  "MPI_Op": "MPI_SUM",
+  "tags": {
+    "include": [
+      "internal",
+      "external"
+    ],
+    "exclude": []
+  },
+  "specific": {
+    "include": [],
+    "exclude": []
+  }
+}

include/libpico.h

@@ -19,6 +19,8 @@
                                   void* rbuf, size_t rcount, MPI_Datatype rdtype, MPI_Comm comm
 #define ALLTOALL_MPI_ARGS         const void *sbuf, size_t scount, MPI_Datatype sdtype, \
                                   void *rbuf, size_t rcount, MPI_Datatype rdtype, MPI_Comm comm
+#define ALLTOALLV_MPI_ARGS        const void *sbuf, const int scounts[], const int sdispls[], MPI_Datatype sdtype, \
+                                  void *rbuf, const int rcounts[], const int rdispls[], MPI_Datatype rdtype, MPI_Comm comm
 #define BCAST_MPI_ARGS            void *buf, size_t count, MPI_Datatype dtype, int root, MPI_Comm comm
 #define GATHER_MPI_ARGS           const void *sbuf, size_t scount, MPI_Datatype sdtype, \
                                   void *rbuf, size_t rcount, MPI_Datatype rdtype, int root, MPI_Comm comm
@@ -38,6 +40,9 @@
                                   ncclDataType_t dtype, ncclComm_t nccl_comm, cudaStream_t stream
 #define ALLTOALL_NCCL_ARGS        const void *sbuf, void* rbuf, size_t count,\
                                   ncclDataType_t dtype, ncclComm_t nccl_comm, cudaStream_t stream
+#define ALLTOALLV_NCCL_ARGS       const void *sbuf, const int scounts[], const int sdispls[], \
+                                  void *rbuf, const int rcounts[], const int rdispls[], \
+                                  ncclDataType_t dtype, ncclComm_t nccl_comm, cudaStream_t stream
 #define BCAST_NCCL_ARGS           void *buf, size_t count, ncclDataType_t dtype, \
                                   int root, ncclComm_t nccl_comm, cudaStream_t stream
 #define GATHER_NCCL_ARGS          const void *sbuf, void *rbuf, size_t count, ncclDataType_t dtype, \
@@ -87,11 +92,19 @@ int allgather_bine_block_by_block_nccl(ALLGATHER_NCCL_ARGS);
 
 int alltoall_pairwise_ompi(ALLTOALL_MPI_ARGS);
 int alltoall_bine(ALLTOALL_MPI_ARGS);
+int alltoall_bine_DH(ALLTOALL_MPI_ARGS);
 // write here the prototype of NCCL alltoall functions
 #ifdef PICO_NCCL
 // int alltoall_nccl_custom(ALLTOALL_NCCL_ARGS)
 #endif
 
+int alltoallv_bine_DH(ALLTOALLV_MPI_ARGS);
+
+#ifdef PICO_NCCL
+int alltoallv_nccl_spreadout(ALLTOALLV_NCCL_ARGS);
+int alltoallv_nccl_fanout(ALLTOALLV_NCCL_ARGS);
+#endif
+
 int bcast_linear(BCAST_MPI_ARGS);
 int bcast_binomial_halving(BCAST_MPI_ARGS);
 int bcast_binomial_doubling(BCAST_MPI_ARGS);