Merge pull request #71 from HDFGroup/feature/select_io_tconv

Feature/select io tconv
HDFGroup · Mar 16, 2023 · a25b005 · a25b005
2 parents 4716568 + 08b8e82
commit a25b005
Show file tree

Hide file tree

Showing 7 changed files with 73 additions and 25 deletions.
diff --git a/src/H5Dchunk.c b/src/H5Dchunk.c
@@ -2562,15 +2562,12 @@ H5D__chunk_may_use_select_io(const H5D_io_info_t *io_info, const H5D_dset_io_inf
     dataset = dset_info->dset;
     HDassert(dataset);
 
-    /* Don't use selection I/O if it's globally disabled, there is a type
-     * conversion, or if there are filters on the dataset (for now) */
-    if (dset_info->io_ops.single_read != H5D__select_read || dataset->shared->dcpl_cache.pline.nused > 0)
+    /* Don't use selection I/O if there are filters on the dataset (for now) */
+    if (dataset->shared->dcpl_cache.pline.nused > 0)
         ret_value = FALSE;
     else {
         hbool_t page_buf_enabled;
 
-        HDassert(dset_info->io_ops.single_write == H5D__select_write);
-
         /* Check if the page buffer is enabled */
         if (H5PB_enabled(io_info->f_sh, H5FD_MEM_DRAW, &page_buf_enabled) < 0)
             HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if page buffer is enabled")
@@ -2768,7 +2765,7 @@ H5D__chunk_read(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)
             chunk_node = H5D_CHUNK_GET_NEXT_NODE(dset_info, chunk_node);
         } /* end while */
 
-        /* Only perform I/O if not performing multi dataset I/O, otherwise the
+        /* Only perform I/O if not performing multi dataset I/O or type conversion, otherwise the
          * higher level will handle it after all datasets have been processed */
         if (H5D_LAYOUT_CB_PERFORM_IO(io_info)) {
             /* Issue selection I/O call (we can skip the page buffer because we've
@@ -2789,7 +2786,13 @@ H5D__chunk_read(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)
                 chunk_addrs       = H5MM_xfree(chunk_addrs);
             } /* end if */
         }     /* end if */
-    }         /* end if */
+
+#ifdef H5_HAVE_PARALLEL
+        /* Report that collective chunk I/O was used (will only be set on the DXPL if collective I/O was
+         * requested) */
+        io_info->actual_io_mode |= H5D_MPIO_CHUNK_COLLECTIVE;
+#endif /* H5_HAVE_PARALLEL */
+    }  /* end if */
     else {
         H5D_io_info_t ctg_io_info; /* Contiguous I/O info object */
         H5D_storage_t ctg_store;   /* Chunk storage information as contiguous dataset */
@@ -3151,7 +3154,7 @@ H5D__chunk_write(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)
             chunk_node = H5D_CHUNK_GET_NEXT_NODE(dset_info, chunk_node);
         } /* end while */
 
-        /* Only perform I/O if not performing multi dataset I/O, otherwise the
+        /* Only perform I/O if not performing multi dataset I/O or type conversion, otherwise the
          * higher level will handle it after all datasets have been processed */
         if (H5D_LAYOUT_CB_PERFORM_IO(io_info)) {
             /* Issue selection I/O call (we can skip the page buffer because we've
@@ -3172,7 +3175,13 @@ H5D__chunk_write(H5D_io_info_t *io_info, H5D_dset_io_info_t *dset_info)
                 chunk_addrs       = H5MM_xfree(chunk_addrs);
             } /* end if */
         }     /* end if */
-    }         /* end if */
+
+#ifdef H5_HAVE_PARALLEL
+        /* Report that collective chunk I/O was used (will only be set on the DXPL if collective I/O was
+         * requested) */
+        io_info->actual_io_mode |= H5D_MPIO_CHUNK_COLLECTIVE;
+#endif /* H5_HAVE_PARALLEL */
+    }  /* end if */
     else {
         /* Iterate through nodes in chunk skip list */
         chunk_node = H5D_CHUNK_GET_FIRST_NODE(dset_info);

diff --git a/src/H5Dcontig.c b/src/H5Dcontig.c
@@ -677,7 +677,7 @@ H5D__contig_io_init(H5D_io_info_t *io_info, H5D_dset_io_info_t *dinfo)
     /* Check if we're performing selection I/O if it hasn't been disabled
      * already */
     if (io_info->use_select_io) {
-        if ((use_selection_io = H5D__contig_may_use_select_io(io_info, dinfo, H5D_IO_OP_READ)) < 0)
+        if ((use_selection_io = H5D__contig_may_use_select_io(io_info, dinfo, io_info->op_type)) < 0)
             HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't check if selection I/O is possible")
         io_info->use_select_io = (hbool_t)use_selection_io;
     }
@@ -808,8 +808,8 @@ H5D__contig_read(H5D_io_info_t *io_info, H5D_dset_io_info_t *dinfo)
     HDassert(dinfo->file_space);
 
     if (io_info->use_select_io) {
-        /* Only perform I/O if not performing multi dataset I/O with selection
-         * I/O, otherwise the higher level will handle it after all datasets
+        /* Only perform I/O if not performing multi dataset I/O or type conversion,
+         * otherwise the higher level will handle it after all datasets
          * have been processed */
         if (H5D_LAYOUT_CB_PERFORM_IO(io_info)) {
             size_t dst_type_size = dinfo->type_info.dst_type_size;
@@ -843,7 +843,12 @@ H5D__contig_read(H5D_io_info_t *io_info, H5D_dset_io_info_t *dinfo)
                 io_info->pieces_added++;
             }
         }
-    } /* end if */
+
+#ifdef H5_HAVE_PARALLEL
+        /* Report that collective contiguous I/O was used */
+        io_info->actual_io_mode |= H5D_MPIO_CONTIGUOUS_COLLECTIVE;
+#endif /* H5_HAVE_PARALLEL */
+    }  /* end if */
     else
         /* Read data through legacy (non-selection I/O) pathway */
         if ((dinfo->io_ops.single_read)(io_info, dinfo) < 0)
@@ -880,8 +885,8 @@ H5D__contig_write(H5D_io_info_t *io_info, H5D_dset_io_info_t *dinfo)
     HDassert(dinfo->file_space);
 
     if (io_info->use_select_io) {
-        /* Only perform I/O if not performing multi dataset I/O with selection
-         * I/O, otherwise the higher level will handle it after all datasets
+        /* Only perform I/O if not performing multi dataset I/O or type conversion,
+         * otherwise the higher level will handle it after all datasets
          * have been processed */
         if (H5D_LAYOUT_CB_PERFORM_IO(io_info)) {
             size_t dst_type_size = dinfo->type_info.dst_type_size;
@@ -915,7 +920,12 @@ H5D__contig_write(H5D_io_info_t *io_info, H5D_dset_io_info_t *dinfo)
                 io_info->pieces_added++;
             }
         }
-    } /* end if */
+
+#ifdef H5_HAVE_PARALLEL
+        /* Report that collective contiguous I/O was used */
+        io_info->actual_io_mode |= H5D_MPIO_CONTIGUOUS_COLLECTIVE;
+#endif /* H5_HAVE_PARALLEL */
+    }  /* end if */
     else
         /* Write data through legacy (non-selection I/O) pathway */
         if ((dinfo->io_ops.single_write)(io_info, dinfo) < 0)

diff --git a/src/H5Dio.c b/src/H5Dio.c
@@ -397,6 +397,21 @@ H5D__read(size_t count, H5D_dset_io_info_t *dset_info)
                     HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "selection read failed")
             }
         }
+
+#ifdef H5_HAVE_PARALLEL
+        /* Report the actual I/O mode to the application if appropriate */
+        if (io_info.using_mpi_vfd) {
+            H5FD_mpio_xfer_t xfer_mode; /* Parallel transfer for this request */
+
+            /* Get the parallel I/O transfer mode */
+            if (H5CX_get_io_xfer_mode(&xfer_mode) < 0)
+                HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode")
+
+            /* Only report the collective I/O mode if we're actually performing collective I/O */
+            if (xfer_mode == H5FD_MPIO_COLLECTIVE)
+                H5CX_set_mpio_actual_io_mode(io_info.actual_io_mode);
+        }
+#endif /* H5_HAVE_PARALLEL */
     }
 
 done:
@@ -781,6 +796,21 @@ H5D__write(size_t count, H5D_dset_io_info_t *dset_info)
                     HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "selection write failed")
             }
         }
+
+#ifdef H5_HAVE_PARALLEL
+        /* Report the actual I/O mode to the application if appropriate */
+        if (io_info.using_mpi_vfd) {
+            H5FD_mpio_xfer_t xfer_mode; /* Parallel transfer for this request */
+
+            /* Get the parallel I/O transfer mode */
+            if (H5CX_get_io_xfer_mode(&xfer_mode) < 0)
+                HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't get MPI-I/O transfer mode")
+
+            /* Only report the collective I/O mode if we're actually performing collective I/O */
+            if (xfer_mode == H5FD_MPIO_COLLECTIVE)
+                H5CX_set_mpio_actual_io_mode(io_info.actual_io_mode);
+        }
+#endif /* H5_HAVE_PARALLEL */
     }
 
 #ifdef OLD_WAY

diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c
@@ -1437,13 +1437,8 @@ H5D__link_piece_collective_io(H5D_io_info_t *io_info, int mpi_rank)
         HDassert(io_info->dsets_info[i].dset->shared->dcpl_cache.pline.nused == 0);
         if (io_info->dsets_info[i].layout->type == H5D_CHUNKED)
             actual_io_mode |= H5D_MPIO_CHUNK_COLLECTIVE;
-        else if (io_info->dsets_info[i].layout->type == H5D_CONTIGUOUS) {
+        else if (io_info->dsets_info[i].layout->type == H5D_CONTIGUOUS)
             actual_io_mode |= H5D_MPIO_CONTIGUOUS_COLLECTIVE;
-
-            /* if only single-dset */
-            if (1 == io_info->count)
-                actual_chunk_opt_mode = H5D_MPIO_NO_CHUNK_OPTIMIZATION;
-        }
         else
             HGOTO_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL, "unsupported storage layout")
     }

diff --git a/src/H5Dpkg.h b/src/H5Dpkg.h
@@ -274,6 +274,9 @@ typedef struct H5D_io_info_t {
                                    conversion */
     hbool_t
         must_fill_bkg; /* Whether any datasets need a background buffer filled with destination contents */
+#ifdef H5_HAVE_PARALLEL
+    H5D_mpio_actual_io_mode_t actual_io_mode; /* Actual type of collective or independent I/O */
+#endif                                        /* H5_HAVE_PARALLEL */
 } H5D_io_info_t;
 
 /* Created to pass both at once for callback func */

diff --git a/src/H5Ppublic.h b/src/H5Ppublic.h
@@ -328,7 +328,6 @@ typedef enum H5D_mpio_actual_io_mode_t {
     H5D_MPIO_CHUNK_MIXED = 0x1 | 0x2,
     /**< HDF5 performed one the chunk collective optimization schemes and some
          chunks were accessed independently, some collectively. */
-    /** \internal The contiguous case is separate from the bit field. */
     H5D_MPIO_CONTIGUOUS_COLLECTIVE = 0x4
     /**< Collective I/O was performed on a contiguous dataset */
 } H5D_mpio_actual_io_mode_t;

diff --git a/testpar/t_select_io_dset.c b/testpar/t_select_io_dset.c
@@ -135,13 +135,15 @@ check_io_mode(hid_t dxpl, unsigned chunked)
         if (actual_io_mode != H5D_MPIO_CHUNK_COLLECTIVE) {
             nerrors++;
             if (MAINPROCESS)
-                HDprintf("\n     Failed: Incorrect I/O mode (chunked)     ");
+                HDprintf("\n     Failed: Incorrect I/O mode (expected chunked, returned %u)",
+                         (unsigned)actual_io_mode);
         }
     }
     else if (actual_io_mode != H5D_MPIO_CONTIGUOUS_COLLECTIVE) {
         nerrors++;
         if (MAINPROCESS)
-            HDprintf("\n     Failed: Incorrect I/O mode (contiguous)     ");
+            HDprintf("\n     Failed: Incorrect I/O mode (expected contiguous, returned %u)",
+                     (unsigned)actual_io_mode);
     }
 
 } /* check_io_mode() */