Merge branch 'open-mpi:main' into 29Aug2024_nocid_array

.ci/community-jenkins/Jenkinsfile

@@ -54,7 +54,7 @@ println('Tests Completed')
 def prepare_check_stages() {
     def configure_options = ["--disable-dlopen", "--disable-oshmem", "--enable-builtin-atomic", "--enable-ipv6"]
     def compilers = ["clang10", "gcc5", "gcc6", "gcc7", "gcc8", "gcc9", "gcc10"]
-    def platforms = ["amazon_linux_2", "amazon_linux_2-arm64", "rhel7", "rhel8", "ubuntu_18.04"]
+    def platforms = ["amazon_linux_2", "amazon_linux_2-arm64", "rhel8", "ubuntu_18.04"]
     def check_stages_list = []
 
     // Build everything stage
@@ -71,7 +71,7 @@ def prepare_check_stages() {
 
       for (configure_option in configure_options) {
         def name = "Configure: ${configure_option}".replaceAll("-", "")
-        build_parallel_map.put(name, prepare_build(name, "(ec2&&linux)", "--configure-args \\\"${configure_option}\\\""))
+        build_parallel_map.put(name, prepare_build(name, "amazon_linux_2", "--configure-args \\\"${configure_option}\\\""))
     }
 
     build_parallel_map.put("distcheck", prepare_build("distcheck", "tarball_build", "--distcheck"))

ompi/mca/coll/accelerator/coll_accelerator.h

@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2024      NVIDIA Corporation. All rights reserved.
  * Copyright (c) 2014      The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
@@ -45,6 +46,11 @@ mca_coll_accelerator_allreduce(const void *sbuf, void *rbuf, size_t count,
                         struct ompi_communicator_t *comm,
                         mca_coll_base_module_t *module);
 
+int mca_coll_accelerator_reduce_local(const void *sbuf, void *rbuf, size_t count,
+                         struct ompi_datatype_t *dtype,
+                         struct ompi_op_t *op,
+                         mca_coll_base_module_t *module);
+
 int mca_coll_accelerator_reduce(const void *sbuf, void *rbuf, size_t count,
                          struct ompi_datatype_t *dtype,
                          struct ompi_op_t *op,

ompi/mca/coll/accelerator/coll_accelerator_module.c

@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2024      NVIDIA Corporation. All rights reserved.
  * Copyright (c) 2014-2017 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
@@ -94,6 +95,7 @@ mca_coll_accelerator_comm_query(struct ompi_communicator_t *comm,
 
     accelerator_module->super.coll_allreduce  = mca_coll_accelerator_allreduce;
     accelerator_module->super.coll_reduce     = mca_coll_accelerator_reduce;
+    accelerator_module->super.coll_reduce_local         = mca_coll_accelerator_reduce_local;
     accelerator_module->super.coll_reduce_scatter_block = mca_coll_accelerator_reduce_scatter_block;
     if (!OMPI_COMM_IS_INTER(comm)) {
         accelerator_module->super.coll_scan       = mca_coll_accelerator_scan;
@@ -141,6 +143,7 @@ mca_coll_accelerator_module_enable(mca_coll_base_module_t *module,
 
     ACCELERATOR_INSTALL_COLL_API(comm, s, allreduce);
     ACCELERATOR_INSTALL_COLL_API(comm, s, reduce);
+    ACCELERATOR_INSTALL_COLL_API(comm, s, reduce_local);
     ACCELERATOR_INSTALL_COLL_API(comm, s, reduce_scatter_block);
     if (!OMPI_COMM_IS_INTER(comm)) {
         /* MPI does not define scan/exscan on intercommunicators */
@@ -159,6 +162,7 @@ mca_coll_accelerator_module_disable(mca_coll_base_module_t *module,
 
     ACCELERATOR_UNINSTALL_COLL_API(comm, s, allreduce);
     ACCELERATOR_UNINSTALL_COLL_API(comm, s, reduce);
+    ACCELERATOR_UNINSTALL_COLL_API(comm, s, reduce_local);
     ACCELERATOR_UNINSTALL_COLL_API(comm, s, reduce_scatter_block);
     if (!OMPI_COMM_IS_INTER(comm))
     {

ompi/mca/coll/accelerator/coll_accelerator_reduce.c

@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2024      NVIDIA Corporation.  All rights reserved.
  * Copyright (c) 2004-2023 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
@@ -84,3 +85,60 @@ mca_coll_accelerator_reduce(const void *sbuf, void *rbuf, size_t count,
     }
     return rc;
 }
+
+int
+mca_coll_accelerator_reduce_local(const void *sbuf, void *rbuf, size_t count,
+                                  struct ompi_datatype_t *dtype,
+                                  struct ompi_op_t *op,
+                                  mca_coll_base_module_t *module)
+{
+    ptrdiff_t gap;
+    char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2 = NULL;
+    size_t bufsize;
+    int rc;
+
+    bufsize = opal_datatype_span(&dtype->super, count, &gap);
+
+    rc = mca_coll_accelerator_check_buf((void *)sbuf);
+    if (rc < 0) {
+        return rc;
+    }
+
+    if ((MPI_IN_PLACE != sbuf) && (rc > 0)) {
+        sbuf1 = (char*)malloc(bufsize);
+        if (NULL == sbuf1) {
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        mca_coll_accelerator_memcpy(sbuf1, sbuf, bufsize);
+        sbuf = sbuf1 - gap;
+    }
+
+    rc = mca_coll_accelerator_check_buf(rbuf);
+    if (rc < 0) {
+        return rc;
+    }
+
+    if (rc > 0) {
+        rbuf1 = (char*)malloc(bufsize);
+        if (NULL == rbuf1) {
+            if (NULL != sbuf1) free(sbuf1);
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        mca_coll_accelerator_memcpy(rbuf1, rbuf, bufsize);
+        rbuf2 = rbuf; /* save away original buffer */
+        rbuf = rbuf1 - gap;
+    }
+
+    ompi_op_reduce(op, (void *)sbuf, rbuf, count, dtype);
+    rc = OMPI_SUCCESS;
+
+    if (NULL != sbuf1) {
+        free(sbuf1);
+    }
+    if (NULL != rbuf1) {
+        rbuf = rbuf2;
+        mca_coll_accelerator_memcpy(rbuf, rbuf1, bufsize);
+        free(rbuf1);
+    }
+    return rc;
+}

ompi/mca/coll/base/coll_base_allgather.c

@@ -14,6 +14,7 @@
  *                         reserved.
  * Copyright (c) 2014-2016 Research Organization for Information Science
  *                         and Technology (RIST). All rights reserved.
+ * Copyright (c) 2024      Jeffrey M. Squyres.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -776,7 +777,7 @@ int ompi_coll_base_allgather_intra_k_bruck(const void *sbuf, size_t scount,
     int recvcount, distance;
     ptrdiff_t rlb, rextent;
     ptrdiff_t rsize, rgap = 0;
-    ompi_request_t **reqs;
+    ompi_request_t **reqs = NULL;
     int num_reqs, max_reqs = 0;
 
     char *tmpsend = NULL;
@@ -937,7 +938,7 @@ int ompi_coll_base_allgather_direct_messaging(const void *sbuf, size_t scount,
     int line = -1, rank, comm_size, err = MPI_SUCCESS;
     ptrdiff_t rlb, rextent;
     ptrdiff_t incr;
-    ompi_request_t **reqs;
+    ompi_request_t **reqs = NULL;
     int max_reqs = 0, reqs_needed = 0;
     int peer_rank = 0;
 

ompi/mca/coll/han/coll_han_alltoall.c

@@ -207,8 +207,8 @@ int mca_coll_han_alltoall_using_smsc(
             /* all ranks will pull from the other ranks' sbuf */
             gather_buf_in[0] = (void*)sbuf;
         }
-        gather_buf_in[1] = *(void**)&send_needs_bounce;
-        gather_buf_in[2] = *(void**)&ii_push_data;
+        gather_buf_in[1] = (void*)(intptr_t)send_needs_bounce;
+        gather_buf_in[2] = (void*)(intptr_t)ii_push_data;
 
         rc = low_comm->c_coll->coll_allgather(gather_buf_in, nptrs_gather, MPI_AINT,
                     gather_buf_out, nptrs_gather, MPI_AINT, low_comm,
@@ -385,5 +385,4 @@ int mca_coll_han_alltoall_using_smsc(
     OPAL_OUTPUT_VERBOSE((40, mca_coll_han_component.han_output,
                 "Alltoall Complete with %d\n",rc));
     return rc;
-
-}
+}

ompi/mca/coll/xhc/coll_xhc_module.c

@@ -1,6 +1,7 @@
 /*
  * Copyright (c) 2021-2023 Computer Architecture and VLSI Systems (CARV)
  *                         Laboratory, ICS Forth. All rights reserved.
+ * Copyright (c) 2024      Jeffrey M. Squyres.  All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -523,7 +524,7 @@ static int xhc_module_create_hierarchy(xhc_module_t *module,
             continue;
         }
 
-        int member_id;
+        int member_id = -1;
         int members = 0;
 
         // If working with rank list, set the ranks from the list as "local"
@@ -560,6 +561,8 @@ static int xhc_module_create_hierarchy(xhc_module_t *module,
             }
         }
 
+        assert(member_id != -1);
+
         /* If split or max ranks was specified, math partition the locality
          * and remove the previously added locality mapping to some ranks */
         if(my_def->split > 1) {

ompi/mca/common/ompio/common_ompio.h

@@ -262,6 +262,8 @@ OMPI_DECLSPEC int mca_common_ompio_file_write_at (ompio_file_t *fh, OMPI_MPI_OFF
 OMPI_DECLSPEC int mca_common_ompio_file_iwrite (ompio_file_t *fh, const void *buf, size_t count,
                                                 struct ompi_datatype_t *datatype, ompi_request_t **request);
 
+OMPI_DECLSPEC int mca_common_ompio_file_iwrite_pregen (ompio_file_t *fh, ompi_request_t *request);
+
 OMPI_DECLSPEC int mca_common_ompio_file_iwrite_at (ompio_file_t *fh,  OMPI_MPI_OFFSET_TYPE offset,
                                                    const void *buf,  size_t count,  struct ompi_datatype_t *datatype,
                                                    ompi_request_t **request);
@@ -297,6 +299,8 @@ OMPI_DECLSPEC int mca_common_ompio_file_read_at (ompio_file_t *fh, OMPI_MPI_OFFS
 OMPI_DECLSPEC int mca_common_ompio_file_iread (ompio_file_t *fh, void *buf, size_t count,
                                                struct ompi_datatype_t *datatype, ompi_request_t **request);
 
+OMPI_DECLSPEC int mca_common_ompio_file_iread_pregen (ompio_file_t *fh, ompi_request_t *request);
+
 OMPI_DECLSPEC int mca_common_ompio_file_iread_at (ompio_file_t *fh, OMPI_MPI_OFFSET_TYPE offset,
                                                   void *buf, size_t count, struct ompi_datatype_t *datatype,
                                                   ompi_request_t **request);

ompi/mca/common/ompio/common_ompio_file_read.c

@@ -12,7 +12,7 @@
  *  Copyright (c) 2008-2019 University of Houston. All rights reserved.
  *  Copyright (c) 2018      Research Organization for Information Science
  *                          and Technology (RIST). All rights reserved.
- *  Copyright (c) 2022-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *  Copyright (c) 2022-2024 Advanced Micro Devices, Inc. All rights reserved.
  *  Copyright (c) 2024      Triad National Security, LLC. All rights
  *                          reserved.
  *  $COPYRIGHT$
@@ -530,6 +530,68 @@ int mca_common_ompio_file_iread (ompio_file_t *fh,
     return ret;
 }
 
+/*
+** This routine is invoked from file_read_all.
+** It is only used if the temporary buffer is a gpu buffer,
+** and the fbtl supports the ipreadv operation.
+**
+** The io-array has already been generated in file_read_all,
+** and we use the pre-computed offsets to created a pseudo fview.
+** The position of the file pointer is updated in the file_read_all
+** operation, not here.
+*/
+
+int mca_common_ompio_file_iread_pregen (ompio_file_t *fh,
+                                        ompi_request_t *request)
+{
+    uint32_t i;
+    size_t max_data;
+    size_t pipeline_buf_size;
+    mca_ompio_request_t *ompio_req = (mca_ompio_request_t *) request;
+
+    max_data = fh->f_io_array[0].length;
+    pipeline_buf_size = OMPIO_MCA_GET(fh, pipeline_buffer_size);
+
+    mca_common_ompio_register_progress ();
+
+    OMPIO_PREPARE_READ_BUF (fh, fh->f_io_array[0].memory_address, max_data, MPI_BYTE,
+                            ompio_req->req_tbuf, &ompio_req->req_convertor, max_data,
+                            pipeline_buf_size, NULL, i);
+
+    ompio_req->req_num_subreqs = ceil((double)max_data/pipeline_buf_size);
+    ompio_req->req_size        = pipeline_buf_size;
+    ompio_req->req_max_data    = max_data;
+    ompio_req->req_post_next_subreq = mca_common_ompio_post_next_read_subreq;
+    ompio_req->req_fh          = fh;
+    ompio_req->req_ompi.req_status.MPI_ERROR = MPI_SUCCESS;
+
+    ompio_req->req_fview = (struct ompio_fview_t *) calloc(1, sizeof(struct ompio_fview_t));
+    if (NULL == ompio_req->req_fview) {
+        opal_output(1, "common_ompio: error allocating memory\n");
+        return OMPI_ERR_OUT_OF_RESOURCE;
+    }
+
+   ompio_req->req_fview->f_decoded_iov = (struct iovec*) malloc (fh->f_num_of_io_entries *
+                                                                 sizeof(struct iovec));
+    if (NULL == ompio_req->req_fview->f_decoded_iov) {
+        opal_output(1, "common_ompio_file_iread_pregen: could not allocate memory\n");
+        return OMPI_ERR_OUT_OF_RESOURCE;
+    }
+
+    ompio_req->req_fview->f_iov_count = fh->f_num_of_io_entries;
+    for (i=0; i < ompio_req->req_fview->f_iov_count; i++) {
+        ompio_req->req_fview->f_decoded_iov[i].iov_base = fh->f_io_array[i].offset;
+        ompio_req->req_fview->f_decoded_iov[i].iov_len  = fh->f_io_array[i].length ;
+    }
+
+    fh->f_num_of_io_entries = 0;
+    free (fh->f_io_array);
+    fh->f_io_array = NULL;
+
+    mca_common_ompio_post_next_read_subreq(ompio_req, 0);
+    return OMPI_SUCCESS;
+}
+
 int mca_common_ompio_file_iread_at (ompio_file_t *fh,
 				  OMPI_MPI_OFFSET_TYPE offset,
 				  void *buf,