diff options
author | Quincey Koziol <koziol@hdfgroup.org> | 2016-04-02 09:39:32 (GMT) |
---|---|---|
committer | Quincey Koziol <koziol@hdfgroup.org> | 2016-04-02 09:39:32 (GMT) |
commit | 9d2178ab886ae957cfe11b6fe09f9e7f0e9ce369 (patch) | |
tree | d4256c01e3f5364ca39ec230a6a897d16f51273d /src/H5Dchunk.c | |
parent | e45885dea912a18b9fd6b1450d3ff196dcb749eb (diff) | |
download | hdf5-9d2178ab886ae957cfe11b6fe09f9e7f0e9ce369.zip hdf5-9d2178ab886ae957cfe11b6fe09f9e7f0e9ce369.tar.gz hdf5-9d2178ab886ae957cfe11b6fe09f9e7f0e9ce369.tar.bz2 |
[svn-r29607] Description:
Bring "don't filter partial edge chunks" capability from revise_chunks
to trunk.
Tested on:
MacOSX/64 10.11.4 (amazon) w/debug, production & parallel
(h5committest forthcoming)
Diffstat (limited to 'src/H5Dchunk.c')
-rw-r--r-- | src/H5Dchunk.c | 651 |
1 files changed, 601 insertions, 50 deletions
diff --git a/src/H5Dchunk.c b/src/H5Dchunk.c index d8e8eb5..309c7a8 100644 --- a/src/H5Dchunk.c +++ b/src/H5Dchunk.c @@ -103,6 +103,11 @@ /*#define H5D_CHUNK_DEBUG */ +/* Flags for the "edge_chunk_state" field below */ +#define H5D_RDCC_DISABLE_FILTERS 0x01u /* Disable filters on this chunk */ +#define H5D_RDCC_NEWLY_DISABLED_FILTERS 0x02u /* Filters have been disabled since + * the last flush */ + /******************/ /* Local Typedefs */ @@ -113,6 +118,7 @@ typedef struct H5D_rdcc_ent_t { hbool_t locked; /*entry is locked in cache */ hbool_t dirty; /*needs to be written to disk? */ hbool_t deleted; /*chunk about to be deleted */ + unsigned edge_chunk_state; /*states related to edge chunks (see above) */ hsize_t scaled[H5O_LAYOUT_NDIMS]; /*scaled chunk 'name' (coordinates) */ uint32_t rd_count; /*bytes remaining to be read */ uint32_t wr_count; /*bytes remaining to be written */ @@ -174,6 +180,8 @@ typedef struct H5D_chunk_it_ud3_t { /* needed for compressed variable-length data */ const H5O_pline_t *pline; /* Filter pipeline */ + unsigned dset_ndims; /* Number of dimensions in dataset */ + const hsize_t *dset_dims; /* Dataset dimensions */ /* needed for copy object pointed by refs */ H5O_copy_t *cpy_info; /* Copy options */ @@ -264,17 +272,18 @@ static herr_t H5D__chunk_flush_entry(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t *dxpl_cache, H5D_rdcc_ent_t *ent, hbool_t reset); static herr_t H5D__chunk_cache_evict(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t *dxpl_cache, H5D_rdcc_ent_t *ent, hbool_t flush); +static hbool_t H5D__chunk_is_partial_edge_chunk(unsigned dset_ndims, + const uint32_t *chunk_dims, const hsize_t *chunk_scaled, const hsize_t *dset_dims); static void *H5D__chunk_lock(const H5D_io_info_t *io_info, - H5D_chunk_ud_t *udata, hbool_t relax); + H5D_chunk_ud_t *udata, hbool_t relax, hbool_t prev_unfilt_chunk); static herr_t H5D__chunk_unlock(const H5D_io_info_t *io_info, const H5D_chunk_ud_t *udata, hbool_t dirty, void *chunk, uint32_t naccessed); static herr_t H5D__chunk_cache_prune(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t *dxpl_cache, size_t size); -static herr_t H5D__chunk_prune_fill(H5D_chunk_it_ud1_t *udata); +static herr_t H5D__chunk_prune_fill(H5D_chunk_it_ud1_t *udata, hbool_t new_unfilt_chunk); static herr_t H5D__chunk_file_alloc(const H5D_chk_idx_info_t *idx_info, - const H5F_block_t *old_chunk, H5F_block_t *new_chunk, hbool_t *need_insert, - hsize_t scaled[]); + const H5F_block_t *old_chunk, H5F_block_t *new_chunk, hbool_t *need_insert); #ifdef H5_HAVE_PARALLEL static herr_t H5D__chunk_collective_fill(const H5D_t *dset, hid_t dxpl_id, H5D_chunk_coll_info_t *chunk_info, size_t chunk_size, const void *fill_buf); @@ -420,7 +429,7 @@ H5D__chunk_direct_write(const H5D_t *dset, hid_t dxpl_id, uint32_t filters, /* Create the chunk it if it doesn't exist, or reallocate the chunk * if its size changed. */ - if(H5D__chunk_file_alloc(&idx_info, &old_chunk, &udata.chunk_block, &need_insert, scaled) < 0) + if(H5D__chunk_file_alloc(&idx_info, &old_chunk, &udata.chunk_block, &need_insert) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "unable to allocate chunk") /* Make sure the address of the chunk is returned. */ @@ -1786,6 +1795,7 @@ htri_t H5D__chunk_cacheable(const H5D_io_info_t *io_info, haddr_t caddr, hbool_t write_op) { const H5D_t *dataset = io_info->dset; /* Local pointer to dataset info */ + hbool_t has_filters = FALSE; /* Whether there are filters on the chunk or not */ htri_t ret_value = FAIL; /* Return value */ FUNC_ENTER_PACKAGE @@ -1794,8 +1804,23 @@ H5D__chunk_cacheable(const H5D_io_info_t *io_info, haddr_t caddr, hbool_t write_ HDassert(io_info); HDassert(dataset); - /* Must bring the whole chunk in if there are any filters */ - if(dataset->shared->dcpl_cache.pline.nused > 0) + /* Must bring the whole chunk in if there are any filters on the chunk. + * Make sure to check if filters are on the dataset but disabled for the + * chunk because it is a partial edge chunk. */ + if(dataset->shared->dcpl_cache.pline.nused > 0) { + if(dataset->shared->layout.u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS) { + has_filters = !H5D__chunk_is_partial_edge_chunk( + io_info->dset->shared->ndims, + io_info->dset->shared->layout.u.chunk.dim, + io_info->store->chunk.scaled, + io_info->dset->shared->curr_dims); + } /* end if */ + else + has_filters = TRUE; + } /* end if */ + + if(has_filters) ret_value = TRUE; else { #ifdef H5_HAVE_PARALLEL @@ -1957,7 +1982,7 @@ H5D__chunk_read(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, src_accessed_bytes = chunk_info->chunk_points * (uint32_t)type_info->src_type_size; /* Lock the chunk into the cache */ - if(NULL == (chunk = H5D__chunk_lock(io_info, &udata, FALSE))) + if(NULL == (chunk = H5D__chunk_lock(io_info, &udata, FALSE, FALSE))) HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") /* Set up the storage buffer information for this chunk */ @@ -2091,7 +2116,7 @@ H5D__chunk_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, entire_chunk = FALSE; /* Lock the chunk into the cache */ - if(NULL == (chunk = H5D__chunk_lock(io_info, &udata, entire_chunk))) + if(NULL == (chunk = H5D__chunk_lock(io_info, &udata, entire_chunk, FALSE))) HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL, "unable to read raw data chunk") /* Set up the storage buffer information for this chunk */ @@ -2114,7 +2139,7 @@ H5D__chunk_write(H5D_io_info_t *io_info, const H5D_type_info_t *type_info, udata.chunk_block.length = io_info->dset->shared->layout.u.chunk.size; /* Allocate the chunk */ - if(H5D__chunk_file_alloc(&idx_info, NULL, &udata.chunk_block, &need_insert, chunk_info->scaled) < 0) + if(H5D__chunk_file_alloc(&idx_info, NULL, &udata.chunk_block, &need_insert) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert/resize chunk on chunk level") /* Make sure the address of the chunk is returned. */ @@ -2614,6 +2639,7 @@ H5D__chunk_lookup(const H5D_t *dset, hid_t dxpl_id, const hsize_t *scaled, udata->chunk_block.offset = HADDR_UNDEF; udata->chunk_block.length = 0; udata->filter_mask = 0; + udata->new_unfilt_chunk = FALSE; /* Check for chunk in cache */ if(dset->shared->cache.chunk.nslots > 0) { @@ -2742,7 +2768,8 @@ H5D__chunk_flush_entry(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t udata.chunk_idx = ent->chunk_idx; /* Should the chunk be filtered before writing it to disk? */ - if(dset->shared->dcpl_cache.pline.nused) { + if(dset->shared->dcpl_cache.pline.nused + && !(ent->edge_chunk_state & H5D_RDCC_DISABLE_FILTERS)) { size_t alloc = udata.chunk_block.length; /* Bytes allocated for BUF */ size_t nbytes; /* Chunk size (in bytes) */ @@ -2781,10 +2808,27 @@ H5D__chunk_flush_entry(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t /* Indicate that the chunk must be allocated */ must_alloc = TRUE; } /* end if */ - else if(!H5F_addr_defined(udata.chunk_block.offset)) + else if(!H5F_addr_defined(udata.chunk_block.offset)) { /* Indicate that the chunk must be allocated */ must_alloc = TRUE; + /* This flag could be set for this chunk, just remove and ignore it + */ + ent->edge_chunk_state &= ~H5D_RDCC_NEWLY_DISABLED_FILTERS; + } /* end else */ + else if(ent->edge_chunk_state & H5D_RDCC_NEWLY_DISABLED_FILTERS) { + /* Chunk on disk is still filtered, must insert to allocate correct + * size */ + must_alloc = TRUE; + + /* Set the disable filters field back to the standard disable + * filters setting, as it no longer needs to be inserted with every + * flush */ + ent->edge_chunk_state &= ~H5D_RDCC_NEWLY_DISABLED_FILTERS; + } /* end else */ + + HDassert(!(ent->edge_chunk_state & H5D_RDCC_NEWLY_DISABLED_FILTERS)); + /* Check if the chunk needs to be allocated (it also could exist already * and the chunk alloc operation could resize it) */ @@ -2799,7 +2843,7 @@ H5D__chunk_flush_entry(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t /* Create the chunk it if it doesn't exist, or reallocate the chunk * if its size changed. */ - if(H5D__chunk_file_alloc(&idx_info, &(ent->chunk_block), &udata.chunk_block, &need_insert, ent->scaled) < 0) + if(H5D__chunk_file_alloc(&idx_info, &(ent->chunk_block), &udata.chunk_block, &need_insert) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert/resize chunk on chunk level") /* Update the chunk entry's info, in case it was allocated or relocated */ @@ -2834,7 +2878,9 @@ H5D__chunk_flush_entry(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t if(buf == ent->chunk) buf = NULL; if(ent->chunk != NULL) - ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, &(dset->shared->dcpl_cache.pline)); + ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, + ((ent->edge_chunk_state & H5D_RDCC_DISABLE_FILTERS) ? NULL + : &(dset->shared->dcpl_cache.pline))); } /* end if */ done: @@ -2850,7 +2896,9 @@ done: */ if(ret_value < 0 && point_of_no_return) if(ent->chunk) - ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, &(dset->shared->dcpl_cache.pline)); + ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, + ((ent->edge_chunk_state & H5D_RDCC_DISABLE_FILTERS) ? NULL + : &(dset->shared->dcpl_cache.pline))); FUNC_LEAVE_NOAPI_TAG(ret_value, FAIL) } /* end H5D__chunk_flush_entry() */ @@ -2892,7 +2940,9 @@ H5D__chunk_cache_evict(const H5D_t *dset, hid_t dxpl_id, const H5D_dxpl_cache_t else { /* Don't flush, just free chunk */ if(ent->chunk != NULL) - ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, &(dset->shared->dcpl_cache.pline)); + ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, + ((ent->edge_chunk_state & H5D_RDCC_DISABLE_FILTERS) ? NULL + : &(dset->shared->dcpl_cache.pline))); } /* end else */ /* Unlink from list */ @@ -3071,10 +3121,11 @@ done: */ static void * H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, - hbool_t relax) + hbool_t relax, hbool_t prev_unfilt_chunk) { const H5D_t *dset = io_info->dset; /* Local pointer to the dataset info */ const H5O_pline_t *pline = &(dset->shared->dcpl_cache.pline); /* I/O pipeline info - always equal to the pline passed to H5D__chunk_mem_alloc */ + const H5O_pline_t *old_pline = pline; /* Old pipeline, i.e. pipeline used to read the chunk */ const H5O_layout_t *layout = &(dset->shared->layout); /* Dataset layout */ const H5O_fill_t *fill = &(dset->shared->dcpl_cache.fill); /* Fill value info */ H5D_fill_buf_info_t fb_info; /* Dataset's fill buffer info */ @@ -3082,6 +3133,7 @@ H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, H5D_rdcc_t *rdcc = &(dset->shared->cache.chunk); /*raw data chunk cache*/ H5D_rdcc_ent_t *ent; /*cache entry */ size_t chunk_size; /*size of a chunk */ + hbool_t disable_filters = FALSE; /* Whether to disable filters (when adding to cache) */ void *chunk = NULL; /*the file chunk */ void *ret_value = NULL; /* Return value */ @@ -3094,6 +3146,7 @@ H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, HDassert(dset); HDassert(TRUE == H5P_isa_class(io_info->md_dxpl_id, H5P_DATASET_XFER)); HDassert(TRUE == H5P_isa_class(io_info->raw_dxpl_id, H5P_DATASET_XFER)); + HDassert(!(udata->new_unfilt_chunk && prev_unfilt_chunk)); HDassert(!rdcc->tmp_head); /* Get the chunk's size */ @@ -3124,6 +3177,66 @@ H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, */ rdcc->stats.nhits++; + /* Make adjustments if the edge chunk status changed recently */ + if(pline->nused) { + /* If the chunk recently became an unfiltered partial edge chunk + * while in cache, we must make some changes to the entry */ + if(udata->new_unfilt_chunk) { + /* If this flag is set then partial chunk filters must be + * disabled, and the chunk must not have previously been a + * partial chunk (with disabled filters) */ + HDassert(layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + HDassert(!(ent->edge_chunk_state & H5D_RDCC_DISABLE_FILTERS)); + HDassert(old_pline->nused); + + /* Disable filters. Set pline to NULL instead of just the + * default pipeline to make a quick failure more likely if the + * code is changed in an inappropriate/incomplete way. */ + pline = NULL; + + /* Reallocate the chunk so H5D__chunk_mem_xfree doesn't get confused + */ + if(NULL == (chunk = H5D__chunk_mem_alloc(chunk_size, pline))) + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for raw data chunk") + HDmemcpy(chunk, ent->chunk, chunk_size); + ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, old_pline); + ent->chunk = (uint8_t *)chunk; + chunk = NULL; + + /* Mark the chunk as having filters disabled as well as "newly + * disabled" so it is inserted on flush */ + ent->edge_chunk_state |= H5D_RDCC_DISABLE_FILTERS; + ent->edge_chunk_state |= H5D_RDCC_NEWLY_DISABLED_FILTERS; + } /* end if */ + else if(prev_unfilt_chunk) { + /* If this flag is set then partial chunk filters must be + * disabled, and the chunk must have previously been a partial + * chunk (with disabled filters) */ + HDassert(layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + HDassert((ent->edge_chunk_state & H5D_RDCC_DISABLE_FILTERS)); + HDassert(pline->nused); + + /* Mark the old pipeline as having been disabled */ + old_pline = NULL; + + /* Reallocate the chunk so H5D__chunk_mem_xfree doesn't get confused + */ + if(NULL == (chunk = H5D__chunk_mem_alloc(chunk_size, pline))) + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for raw data chunk") + HDmemcpy(chunk, ent->chunk, chunk_size); + + ent->chunk = (uint8_t *)H5D__chunk_mem_xfree(ent->chunk, old_pline); + ent->chunk = (uint8_t *)chunk; + chunk = NULL; + + /* Mark the chunk as having filters enabled */ + ent->edge_chunk_state &= ~(H5D_RDCC_DISABLE_FILTERS + | H5D_RDCC_NEWLY_DISABLED_FILTERS); + } /* end else */ + } /* end if */ + /* * If the chunk is not at the beginning of the cache; move it backward * by one slot. This is how we implement the LRU preemption @@ -3152,6 +3265,39 @@ H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, chunk_addr = udata->chunk_block.offset; chunk_alloc = udata->chunk_block.length; + /* Check if we should disable filters on this chunk */ + if(pline->nused) { + if(udata->new_unfilt_chunk) { + HDassert(layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + + /* Disable the filters for writing */ + disable_filters = TRUE; + pline = NULL; + } /* end if */ + else if(prev_unfilt_chunk) { + HDassert(layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + + /* Mark the filters as having been previously disabled (for the + * chunk as currently on disk) - disable the filters for reading + */ + old_pline = NULL; + } /* end if */ + else if(layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS) { + /* Check if this is an edge chunk */ + if(H5D__chunk_is_partial_edge_chunk(io_info->dset->shared->ndims, + layout->u.chunk.dim, io_info->store->chunk.scaled, + io_info->dset->shared->curr_dims)) { + /* Disable the filters for both writing and reading */ + disable_filters = TRUE; + old_pline = NULL; + pline = NULL; + } /* end if */ + } /* end if */ + } /* end if */ + if(relax) { /* * Not in the cache, but we're about to overwrite the whole thing @@ -3181,16 +3327,32 @@ H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, /* Chunk size on disk isn't [likely] the same size as the final chunk * size in memory, so allocate memory big enough. */ - if(NULL == (chunk = H5D__chunk_mem_alloc(my_chunk_alloc, pline))) + if(NULL == (chunk = H5D__chunk_mem_alloc(my_chunk_alloc, (udata->new_unfilt_chunk ? old_pline : pline)))) HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for raw data chunk") if(H5F_block_read(dset->oloc.file, H5FD_MEM_DRAW, chunk_addr, my_chunk_alloc, io_info->raw_dxpl_id, chunk) < 0) HGOTO_ERROR(H5E_IO, H5E_READERROR, NULL, "unable to read raw data chunk") - if(pline->nused) - if(H5Z_pipeline(pline, H5Z_FLAG_REVERSE, &(udata->filter_mask), io_info->dxpl_cache->err_detect, - io_info->dxpl_cache->filter_cb, &my_chunk_alloc, &buf_alloc, &chunk) < 0) + if(old_pline && old_pline->nused) { + if(H5Z_pipeline(old_pline, H5Z_FLAG_REVERSE, + &(udata->filter_mask), + io_info->dxpl_cache->err_detect, + io_info->dxpl_cache->filter_cb, + &my_chunk_alloc, &buf_alloc, &chunk) < 0) HGOTO_ERROR(H5E_PLINE, H5E_CANTFILTER, NULL, "data pipeline read failed") + /* Reallocate chunk if necessary */ + if(udata->new_unfilt_chunk) { + void *tmp_chunk = chunk; + + if(NULL == (chunk = H5D__chunk_mem_alloc(my_chunk_alloc, pline))) { + (void)H5D__chunk_mem_xfree(tmp_chunk, old_pline); + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, NULL, "memory allocation failed for raw data chunk") + } /* end if */ + HDmemcpy(chunk, tmp_chunk, chunk_size); + (void)H5D__chunk_mem_xfree(tmp_chunk, old_pline); + } /* end if */ + } /* end if */ + /* Increment # of cache misses */ rdcc->stats.nmisses++; } /* end if */ @@ -3259,6 +3421,10 @@ H5D__chunk_lock(const H5D_io_info_t *io_info, H5D_chunk_ud_t *udata, if(NULL == (ent = H5FL_CALLOC(H5D_rdcc_ent_t))) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, NULL, "can't allocate raw data chunk entry") + ent->edge_chunk_state = disable_filters ? H5D_RDCC_DISABLE_FILTERS : 0; + if(udata->new_unfilt_chunk) + ent->edge_chunk_state |= H5D_RDCC_NEWLY_DISABLED_FILTERS; + /* Initialize the new entry */ ent->chunk_block.offset = chunk_addr; ent->chunk_block.length = chunk_alloc; @@ -3367,12 +3533,30 @@ H5D__chunk_unlock(const H5D_io_info_t *io_info, const H5D_chunk_ud_t *udata, * It's not in the cache, probably because it's too big. If it's * dirty then flush it to disk. In any case, free the chunk. */ + hbool_t is_unfiltered_edge_chunk = FALSE; /* Whether the chunk is an unfiltered edge chunk */ + + /* Check if we should disable filters on this chunk */ + if(udata->new_unfilt_chunk) { + HDassert(layout->u.chunk.flags & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + + is_unfiltered_edge_chunk = TRUE; + } /* end if */ + else if(layout->u.chunk.flags & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS) { + /* Check if the chunk is an edge chunk, and disable filters if so */ + is_unfiltered_edge_chunk = H5D__chunk_is_partial_edge_chunk( + io_info->dset->shared->ndims, layout->u.chunk.dim, + io_info->store->chunk.scaled, io_info->dset->shared->curr_dims); + } /* end if */ + if(dirty) { H5D_rdcc_ent_t fake_ent; /* "fake" chunk cache entry */ HDmemset(&fake_ent, 0, sizeof(fake_ent)); fake_ent.dirty = TRUE; - HDmemcpy(fake_ent.scaled, udata->common.scaled, sizeof(hsize_t) * layout->u.chunk.ndims); + if(is_unfiltered_edge_chunk) + fake_ent.edge_chunk_state = H5D_RDCC_DISABLE_FILTERS; + if(udata->new_unfilt_chunk) + fake_ent.edge_chunk_state |= H5D_RDCC_NEWLY_DISABLED_FILTERS; HDmemcpy(fake_ent.scaled, udata->common.scaled, sizeof(hsize_t) * layout->u.chunk.ndims); HDassert(layout->u.chunk.size > 0); fake_ent.chunk_idx = udata->chunk_idx; @@ -3385,7 +3569,8 @@ H5D__chunk_unlock(const H5D_io_info_t *io_info, const H5D_chunk_ud_t *udata, } /* end if */ else { if(chunk) - chunk = H5D__chunk_mem_xfree(chunk, &(io_info->dset->shared->dcpl_cache.pline)); + chunk = H5D__chunk_mem_xfree(chunk, (is_unfiltered_edge_chunk ? NULL + : &(io_info->dset->shared->dcpl_cache.pline))); } /* end else */ } /* end if */ else { @@ -3531,9 +3716,12 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ unsigned filter_mask = 0; /* Filter mask for chunks that have them */ const H5O_layout_t *layout = &(dset->shared->layout); /* Dataset layout */ const H5O_pline_t *pline = &(dset->shared->dcpl_cache.pline); /* I/O pipeline info */ + const H5O_pline_t def_pline = H5O_CRT_PIPELINE_DEF; /* Default pipeline */ const H5O_fill_t *fill = &(dset->shared->dcpl_cache.fill); /* Fill value info */ H5D_fill_value_t fill_status; /* The fill value status */ hbool_t should_fill = FALSE; /* Whether fill values should be written */ + void *unfilt_fill_buf = NULL; /* Unfiltered fill value buffer */ + void **fill_buf = NULL; /* Pointer to the fill buffer to use for a chunk */ H5D_dxpl_cache_t _dxpl_cache; /* Data transfer property cache buffer */ H5D_dxpl_cache_t *dxpl_cache = &_dxpl_cache; /* Data transfer property cache */ #ifdef H5_HAVE_PARALLEL @@ -3550,6 +3738,10 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ unsigned op_dim; /* Current operating dimension */ H5D_fill_buf_info_t fb_info; /* Dataset's fill buffer info */ hbool_t fb_info_init = FALSE; /* Whether the fill value buffer has been initialized */ + hbool_t has_unfilt_edge_chunks = FALSE; /* Whether there are partial edge chunks with disabled filters */ + hbool_t unfilt_edge_chunk_dim[H5O_LAYOUT_NDIMS]; /* Whether there are unfiltered edge chunks at the edge of each dimension */ + hsize_t edge_chunk_scaled[H5O_LAYOUT_NDIMS]; /* Offset of the unfiltered edge chunks at the edge of each dimension */ + unsigned nunfilt_edge_chunk_dims = 0; /* Number of dimensions on an edge */ const H5O_storage_chunk_t *sc = &(layout->storage.u.chunk); herr_t ret_value = SUCCEED; /* Return value */ @@ -3594,6 +3786,29 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ if(H5D__get_dxpl_cache(raw_dxpl_id, &dxpl_cache) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't fill dxpl cache") + /* Calculate the minimum and maximum chunk offsets in each dimension, and + * determine if there are any unfiltered partial edge chunks. Note that we + * assume here that all elements of space_dim are > 0. This is checked at + * the top of this function. */ + for(op_dim=0; op_dim<space_ndims; op_dim++) { + min_unalloc[op_dim] = (old_dim[op_dim] + chunk_dim[op_dim] - 1) / chunk_dim[op_dim]; + max_unalloc[op_dim] = (space_dim[op_dim] - 1) / chunk_dim[op_dim]; + + /* Calculate if there are unfiltered edge chunks at the edge of this + * dimension. Note the edge_chunk_scaled is uninitialized for + * dimensions where unfilt_edge_chunk_dim is FALSE. Also */ + if((layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS) + && pline->nused > 0 + && space_dim[op_dim] % chunk_dim[op_dim] != 0) { + has_unfilt_edge_chunks = TRUE; + unfilt_edge_chunk_dim[op_dim] = TRUE; + edge_chunk_scaled[op_dim] = max_unalloc[op_dim]; + } /* end if */ + else + unfilt_edge_chunk_dim[op_dim] = FALSE; + } /* end for */ + /* Get original chunk size */ H5_CHECKED_ASSIGN(orig_chunk_size, size_t, layout->u.chunk.size, uint32_t); @@ -3625,6 +3840,11 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ HGOTO_ERROR(H5E_DATASET, H5E_CANTINIT, FAIL, "can't initialize fill buffer info") fb_info_init = TRUE; + /* Initialize the fill_buf pointer to the buffer in fb_info. If edge + * chunk filters are disabled, we will switch the buffer as appropriate + * for each chunk. */ + fill_buf = &fb_info.fill_buf; + /* Check if there are filters which need to be applied to the chunk */ /* (only do this in advance when the chunk info can be re-used (i.e. * it doesn't contain any non-default VL datatype fill values) @@ -3632,6 +3852,14 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ if(!fb_info.has_vlen_fill_type && pline->nused > 0) { size_t buf_size = orig_chunk_size; + /* If the dataset has disabled partial chunk filters, create a copy + * of the unfiltered fill_buf to use for partial chunks */ + if(has_unfilt_edge_chunks) { + if(NULL == (unfilt_fill_buf = H5D__chunk_mem_alloc(orig_chunk_size, &def_pline))) + HGOTO_ERROR(H5E_RESOURCE, H5E_NOSPACE, FAIL, "memory allocation failed for raw data chunk") + HDmemcpy(unfilt_fill_buf, fb_info.fill_buf, orig_chunk_size); + } /* end if */ + /* Push the chunk through the filters */ if(H5Z_pipeline(pline, 0, &filter_mask, dxpl_cache->err_detect, dxpl_cache->filter_cb, &orig_chunk_size, &buf_size, &fb_info.fill_buf) < 0) HGOTO_ERROR(H5E_PLINE, H5E_WRITEERROR, FAIL, "output pipeline failed") @@ -3650,14 +3878,6 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ idx_info.layout = &dset->shared->layout.u.chunk; idx_info.storage = &dset->shared->layout.storage.u.chunk; - /* Calculate the minimum and maximum chunk offsets in each dimension. Note - * that we assume here that all elements of space_dim are > 0. This is - * checked at the top of this function. */ - for(op_dim = 0; op_dim < (unsigned)space_ndims; op_dim++) { - min_unalloc[op_dim] = (old_dim[op_dim] + chunk_dim[op_dim] - 1) / chunk_dim[op_dim]; - max_unalloc[op_dim] = (space_dim[op_dim] - 1) / chunk_dim[op_dim]; - } /* end for */ - /* Loop over all chunks */ /* The algorithm is: * For each dimension: @@ -3682,6 +3902,7 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ chunk_size = orig_chunk_size; for(op_dim = 0; op_dim < space_ndims; op_dim++) { H5D_chunk_ud_t udata; /* User data for querying chunk info */ + unsigned u; /* Local index variable */ int i; /* Local index variable */ /* Check if allocation along this dimension is really necessary */ @@ -3692,6 +3913,29 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ HDmemset(scaled, 0, (space_ndims * sizeof(scaled[0]))); scaled[op_dim] = min_unalloc[op_dim]; + if(has_unfilt_edge_chunks) { + /* Initialize nunfilt_edge_chunk_dims */ + nunfilt_edge_chunk_dims = 0; + for(u = 0; u < space_ndims; u++) + if(unfilt_edge_chunk_dim[u] && scaled[u] + == edge_chunk_scaled[u]) + nunfilt_edge_chunk_dims++; + + /* Initialize chunk_size and fill_buf */ + if(should_fill && !fb_info.has_vlen_fill_type) { + HDassert(fb_info_init); + HDassert(unfilt_fill_buf); + if(nunfilt_edge_chunk_dims) { + fill_buf = &unfilt_fill_buf; + chunk_size = layout->u.chunk.size; + } /* end if */ + else { + fill_buf = &fb_info.fill_buf; + chunk_size = orig_chunk_size; + } /* end else */ + } /* end if */ + } /* end if */ + carry = FALSE; } /* end else */ @@ -3707,12 +3951,12 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ /* Make sure the chunk is really in the dataset and outside the * original dimensions */ { - unsigned u; /* Local index variable */ + unsigned v; /* Local index variable */ hbool_t outside_orig = FALSE; - for(u = 0; u < space_ndims; u++) { - HDassert((scaled[u] * chunk_dim[u]) < space_dim[u]); - if((scaled[u] * chunk_dim[u]) >= old_dim[u]) + for(v = 0; v < space_ndims; v++) { + HDassert((scaled[v] * chunk_dim[v]) < space_dim[v]); + if((scaled[v] * chunk_dim[v]) >= old_dim[v]) outside_orig = TRUE; } /* end for */ HDassert(outside_orig); @@ -3723,6 +3967,7 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ if(fb_info_init && fb_info.has_vlen_fill_type) { /* Sanity check */ HDassert(should_fill); + HDassert(!unfilt_fill_buf); #ifdef H5_HAVE_PARALLEL HDassert(!using_mpi); /* Can't write VL datatypes in parallel currently */ #endif @@ -3741,7 +3986,7 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ HGOTO_ERROR(H5E_DATASET, H5E_CANTCONVERT, FAIL, "can't refill fill value buffer") /* Check if there are filters which need to be applied to the chunk */ - if(pline->nused > 0) { + if((pline->nused > 0) && !nunfilt_edge_chunk_dims) { size_t nbytes = orig_chunk_size; /* Push the chunk through the filters */ @@ -3759,6 +4004,8 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ } /* end if */ else chunk_size = layout->u.chunk.size; + + HDassert(*fill_buf == fb_info.fill_buf); } /* end if */ /* Initialize the chunk information */ @@ -3770,7 +4017,7 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ udata.filter_mask = filter_mask; /* Allocate the chunk (with all processes) */ - if(H5D__chunk_file_alloc(&idx_info, NULL, &udata.chunk_block, &need_insert, scaled) < 0) + if(H5D__chunk_file_alloc(&idx_info, NULL, &udata.chunk_block, &need_insert) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert/resize chunk on chunk level") HDassert(H5F_addr_defined(udata.chunk_block.offset)); @@ -3786,6 +4033,9 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ /* collect all chunk addresses to be written to write collectively at the end */ /* allocate/resize address array if no more space left */ + /* Note that if we add support for parallel filters we must + * also store an array of chunk sizes and pass it to the + * apporpriate collective write function */ if(0 == chunk_info.num_io % 1024) if(NULL == (chunk_info.addr = (haddr_t *)H5MM_realloc(chunk_info.addr, (chunk_info.num_io + 1024) * sizeof(haddr_t)))) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "memory allocation failed for chunk addresses") @@ -3799,7 +4049,7 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ } /* end if */ else { #endif /* H5_HAVE_PARALLEL */ - if(H5F_block_write(dset->oloc.file, H5FD_MEM_DRAW, udata.chunk_block.offset, chunk_size, raw_dxpl_id, fb_info.fill_buf) < 0) + if(H5F_block_write(dset->oloc.file, H5FD_MEM_DRAW, udata.chunk_block.offset, chunk_size, raw_dxpl_id, *fill_buf) < 0) HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "unable to write raw data to file") #ifdef H5_HAVE_PARALLEL } /* end else */ @@ -3820,8 +4070,31 @@ H5D__chunk_allocate(const H5D_io_info_t *io_info, hbool_t full_overwrite, hsize_ scaled[i] = min_unalloc[i]; else scaled[i] = 0; - } /* end if */ + + /* Check if we just left the edge in this dimension */ + if(unfilt_edge_chunk_dim[i] + && edge_chunk_scaled[i] == max_unalloc[i] + && scaled[i] < edge_chunk_scaled[i]) { + nunfilt_edge_chunk_dims--; + if(should_fill && nunfilt_edge_chunk_dims == 0 && !fb_info.has_vlen_fill_type) { + HDassert(!H5D__chunk_is_partial_edge_chunk(space_ndims, chunk_dim, scaled, space_dim)); + fill_buf = &fb_info.fill_buf; + chunk_size = orig_chunk_size; + } /* end if */ + } /* end if */ + } /* end if */ else { + /* Check if we just entered the edge in this dimension */ + if(unfilt_edge_chunk_dim[i] && scaled[i] == edge_chunk_scaled[i]) { + HDassert(edge_chunk_scaled[i] == max_unalloc[i]); + nunfilt_edge_chunk_dims++; + if(should_fill && nunfilt_edge_chunk_dims == 1 && !fb_info.has_vlen_fill_type) { + HDassert(H5D__chunk_is_partial_edge_chunk(space_ndims, chunk_dim, scaled, space_dim)); + fill_buf = &unfilt_fill_buf; + chunk_size = layout->u.chunk.size; + } /* end if */ + } /* end if */ + carry = FALSE; break; } /* end else */ @@ -3852,6 +4125,9 @@ done: if(fb_info_init && H5D__fill_term(&fb_info) < 0) HDONE_ERROR(H5E_DATASET, H5E_CANTFREE, FAIL, "Can't release fill buffer info") + /* Free the unfiltered fill value buffer */ + unfilt_fill_buf = H5D__chunk_mem_xfree(unfilt_fill_buf, &def_pline); + #ifdef H5_HAVE_PARALLEL if(using_mpi && chunk_info.addr) H5MM_free(chunk_info.addr); @@ -3860,6 +4136,185 @@ done: FUNC_LEAVE_NOAPI_TAG(ret_value, FAIL) } /* end H5D__chunk_allocate() */ + +/*------------------------------------------------------------------------- + * Function: H5D__chunk_update_old_edge_chunks + * + * Purpose: Update all chunks which were previously partial edge + * chunks and are now complete. Determines exactly which + * chunks need to be updated and locks each into cache using + * the 'prev_unfilt_chunk' flag, then unlocks it, causing + * filters to be applied as necessary. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Neil Fortner + * April 14, 2010 + * + *------------------------------------------------------------------------- + */ +herr_t +H5D__chunk_update_old_edge_chunks(H5D_t *dset, hid_t dxpl_id, hsize_t old_dim[]) +{ + hsize_t old_edge_chunk_sc[H5O_LAYOUT_NDIMS]; /* Offset of first previously incomplete chunk in each dimension */ + hsize_t max_edge_chunk_sc[H5O_LAYOUT_NDIMS]; /* largest offset of chunks that might need to be modified in each dimension */ + hbool_t new_full_dim[H5O_LAYOUT_NDIMS]; /* Whether the plane of chunks in this dimension needs to be modified */ + const H5O_layout_t *layout = &(dset->shared->layout); /* Dataset layout */ + const H5O_pline_t *pline = &(dset->shared->dcpl_cache.pline); /* I/O pipeline info */ + hsize_t chunk_sc[H5O_LAYOUT_NDIMS]; /* Offset of current chunk */ + const uint32_t *chunk_dim = layout->u.chunk.dim; /* Convenience pointer to chunk dimensions */ + unsigned space_ndims; /* Dataset's space rank */ + const hsize_t *space_dim; /* Dataset's dataspace dimensions */ + unsigned op_dim; /* Current operationg dimension */ + H5D_io_info_t chk_io_info; /* Chunked I/O info object */ + H5D_chunk_ud_t chk_udata; /* User data for locking chunk */ + H5D_storage_t chk_store; /* Chunk storage information */ + H5D_dxpl_cache_t _dxpl_cache; /* Data transfer property cache buffer */ + H5D_dxpl_cache_t *dxpl_cache = &_dxpl_cache; /* Data transfer property cache */ + void *chunk; /* The file chunk */ + hbool_t carry; /* Flag to indicate that chunk increment carrys to higher dimension (sorta) */ + const H5O_storage_chunk_t *sc = &(layout->storage.u.chunk); + herr_t ret_value = SUCCEED; /* Return value */ + + FUNC_ENTER_PACKAGE + + /* Check args */ + HDassert(dset && H5D_CHUNKED == layout->type); + HDassert(layout->u.chunk.ndims > 0 && layout->u.chunk.ndims <= H5O_LAYOUT_NDIMS); + H5D_CHUNK_STORAGE_INDEX_CHK(sc); + HDassert(TRUE == H5P_isa_class(dxpl_id, H5P_DATASET_XFER)); + HDassert(pline->nused > 0); + HDassert(layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS); + + /* Retrieve the dataset dimensions */ + space_dim = dset->shared->curr_dims; + space_ndims = dset->shared->ndims; + + /* The last dimension in chunk_offset is always 0 */ + chunk_sc[space_ndims] = (hsize_t)0; + + /* Check if any current dimensions are smaller than the chunk size, or if + * any old dimensions are 0. If so we do not have to do anything. */ + for(op_dim=0; op_dim<space_ndims; op_dim++) + if((space_dim[op_dim] < chunk_dim[op_dim]) || old_dim[op_dim] == 0) { + /* Reset any cached chunk info for this dataset */ + H5D__chunk_cinfo_cache_reset(&dset->shared->cache.chunk.last); + HGOTO_DONE(SUCCEED) + } /* end if */ + + /* + * Initialize structures needed to lock chunks into cache + */ + /* Fill the DXPL cache values for later use */ + if(H5D__get_dxpl_cache(dxpl_id, &dxpl_cache) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't fill dxpl cache") + + /* Set up chunked I/O info object, for operations on chunks (in callback). + * Note that we only need to set chunk_offset once, as the array's address + * will never change. */ + chk_store.chunk.scaled = chunk_sc; + H5D_BUILD_IO_INFO_RD(&chk_io_info, dset, dxpl_cache, dxpl_id, H5AC_rawdata_dxpl_id, &chk_store, NULL); + + /* + * Determine the edges of the dataset which need to be modified + */ + for(op_dim=0; op_dim<space_ndims; op_dim++) { + /* Start off with this dimension marked as not needing to be modified */ + new_full_dim[op_dim] = FALSE; + + /* Calulate offset of first previously incomplete chunk in this + * dimension */ + old_edge_chunk_sc[op_dim] = (old_dim[op_dim] / chunk_dim[op_dim]); + + /* Calculate the largest offset of chunks that might need to be + * modified in this dimension */ + max_edge_chunk_sc[op_dim] = MIN((old_dim[op_dim] - 1) / chunk_dim[op_dim], + MAX((space_dim[op_dim] / chunk_dim[op_dim]), 1) - 1); + + /* Check for old_dim aligned with chunk boundary in this dimension, if + * so we do not need to modify chunks along the edge in this dimension + */ + if(old_dim[op_dim] % chunk_dim[op_dim] == 0) + continue; + + /* Check if the dataspace expanded enough to cause the old edge chunks + * in this dimension to become full */ + if((space_dim[op_dim]/chunk_dim[op_dim]) >= (old_edge_chunk_sc[op_dim] + 1)) + new_full_dim[op_dim] = TRUE; + } /* end for */ + + /* Main loop: fix old edge chunks */ + for(op_dim=0; op_dim<space_ndims; op_dim++) { + /* Check if allocation along this dimension is really necessary */ + if(!new_full_dim[op_dim]) + continue; + else { + HDassert(max_edge_chunk_sc[op_dim] == old_edge_chunk_sc[op_dim]); + + /* Reset the chunk offset indices */ + HDmemset(chunk_sc, 0, (space_ndims * sizeof(chunk_sc[0]))); + chunk_sc[op_dim] = old_edge_chunk_sc[op_dim]; + + carry = FALSE; + } /* end if */ + + while(!carry) { + int i; /* Local index variable */ + + /* Make sure the chunk is really a former edge chunk */ + HDassert(H5D__chunk_is_partial_edge_chunk(space_ndims, chunk_dim, chunk_sc, old_dim) + && !H5D__chunk_is_partial_edge_chunk(space_ndims, chunk_dim, chunk_sc, space_dim)); + + /* Lookup the chunk */ + if(H5D__chunk_lookup(dset, dxpl_id, chunk_sc, &chk_udata) < 0) + HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") + + /* If this chunk does not exist in cache or on disk, no need to do + * anything */ + if(H5F_addr_defined(chk_udata.chunk_block.offset) + || (UINT_MAX != chk_udata.idx_hint)) { + /* Lock the chunk into cache. H5D__chunk_lock will take care of + * updating the chunk to no longer be an edge chunk. */ + if(NULL == (chunk = (void *)H5D__chunk_lock(&chk_io_info, &chk_udata, FALSE, TRUE))) + HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "unable to lock raw data chunk") + + /* Unlock the chunk */ + if(H5D__chunk_unlock(&chk_io_info, &chk_udata, TRUE, chunk, (uint32_t)0) < 0) + HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL, "unable to unlock raw data chunk") + } /* end if */ + + /* Increment indices */ + carry = TRUE; + for(i = ((int)space_ndims - 1); i >= 0; --i) { + if((unsigned)i != op_dim) { + ++chunk_sc[i]; + if(chunk_sc[i] > (hsize_t) max_edge_chunk_sc[i]) + chunk_sc[i] = 0; + else { + carry = FALSE; + break; + } /* end else */ + } /* end if */ + } /* end for */ + } /* end while(!carry) */ + + /* Adjust max_edge_chunk_sc so we don't modify the same chunk twice. + * Also check if this dimension started from 0 (and hence modified all + * of the old edge chunks. */ + if(old_edge_chunk_sc[op_dim] == 0) + break; + else + --max_edge_chunk_sc[op_dim]; + } /* end for(op_dim=0...) */ + + /* Reset any cached chunk info for this dataset */ + H5D__chunk_cinfo_cache_reset(&dset->shared->cache.chunk.last); + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* end H5D__chunk_update_old_edge_chunks() */ + #ifdef H5_HAVE_PARALLEL /*------------------------------------------------------------------------- @@ -4030,7 +4485,7 @@ done: *------------------------------------------------------------------------- */ static herr_t -H5D__chunk_prune_fill(H5D_chunk_it_ud1_t *udata) +H5D__chunk_prune_fill(H5D_chunk_it_ud1_t *udata, hbool_t new_unfilt_chunk) { const H5D_io_info_t *io_info = udata->io_info; /* Local pointer to I/O info */ const H5D_t *dset = io_info->dset; /* Local pointer to the dataset info */ @@ -4057,6 +4512,7 @@ H5D__chunk_prune_fill(H5D_chunk_it_ud1_t *udata) /* Get the info for the chunk in the file */ if(H5D__chunk_lookup(dset, io_info->md_dxpl_id, scaled, &chk_udata) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "error looking up chunk address") + chk_udata.new_unfilt_chunk = new_unfilt_chunk; /* If this chunk does not exist in cache or on disk, no need to do anything */ if(!H5F_addr_defined(chk_udata.chunk_block.offset) && UINT_MAX == chk_udata.idx_hint) @@ -4088,7 +4544,7 @@ H5D__chunk_prune_fill(H5D_chunk_it_ud1_t *udata) HGOTO_ERROR(H5E_DATASET, H5E_CANTSELECT, FAIL, "unable to select hyperslab") /* Lock the chunk into the cache, to get a pointer to the chunk buffer */ - if(NULL == (chunk = (void *)H5D__chunk_lock(io_info, &chk_udata, FALSE))) + if(NULL == (chunk = (void *)H5D__chunk_lock(io_info, &chk_udata, FALSE, FALSE))) HGOTO_ERROR(H5E_DATASET, H5E_READERROR, FAIL, "unable to lock raw data chunk") @@ -4242,6 +4698,8 @@ H5D__chunk_prune_by_extent(H5D_t *dset, hid_t dxpl_id, const hsize_t *old_dim) hsize_t max_mod_chunk_sc[H5O_LAYOUT_NDIMS]; /* Scaled offset of last chunk to modify in each dimension */ hssize_t max_fill_chunk_sc[H5O_LAYOUT_NDIMS]; /* Scaled offset of last chunk that might be filled in each dimension */ hbool_t fill_dim[H5O_LAYOUT_NDIMS]; /* Whether the plane of edge chunks in this dimension needs to be filled */ + hsize_t min_partial_chunk_sc[H5O_LAYOUT_NDIMS]; /* Offset of first partial (or empty) chunk in each dimension */ + hbool_t new_unfilt_dim[H5O_LAYOUT_NDIMS]; /* Whether the plane of edge chunks in this dimension are newly unfiltered */ H5D_chk_idx_info_t idx_info; /* Chunked index info */ H5D_io_info_t chk_io_info; /* Chunked I/O info object */ H5D_storage_t chk_store; /* Chunk storage information */ @@ -4261,6 +4719,8 @@ H5D__chunk_prune_by_extent(H5D_t *dset, hid_t dxpl_id, const hsize_t *old_dim) hsize_t scaled[H5O_LAYOUT_NDIMS]; /* Scaled offset of current chunk */ hsize_t hyper_start[H5O_LAYOUT_NDIMS]; /* Starting location of hyperslab */ uint32_t elmts_per_chunk; /* Elements in chunk */ + hbool_t disable_edge_filters = FALSE; /* Whether to disable filters on partial edge chunks */ + hbool_t new_unfilt_chunk = FALSE; /* Whether the chunk is newly unfiltered */ unsigned u; /* Local index variable */ const H5O_storage_chunk_t *sc = &(layout->storage.u.chunk); herr_t ret_value = SUCCEED; /* Return value */ @@ -4344,6 +4804,11 @@ H5D__chunk_prune_by_extent(H5D_t *dset, hid_t dxpl_id, const hsize_t *old_dim) idx_udata.layout = &layout->u.chunk; idx_udata.storage = &layout->storage.u.chunk; + /* Determine if partial edge chunk filters are disabled */ + disable_edge_filters = (layout->u.chunk.flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS) + && (idx_info.pline->nused > 0); + /* * Determine the chunks which need to be filled or removed */ @@ -4370,13 +4835,31 @@ H5D__chunk_prune_by_extent(H5D_t *dset, hid_t dxpl_id, const hsize_t *old_dim) min_mod_chunk_sc[op_dim] = space_dim[op_dim] / chunk_dim[op_dim]; /* Determine if we need to fill chunks in this dimension */ - if((hssize_t)min_mod_chunk_sc[op_dim] == max_fill_chunk_sc[op_dim]) + if((hssize_t)min_mod_chunk_sc[op_dim] == max_fill_chunk_sc[op_dim]) { fill_dim[op_dim] = TRUE; - else + + /* If necessary, check if chunks in this dimension that need to + * be filled are new partial edge chunks */ + if(disable_edge_filters && old_dim[op_dim] >= (min_mod_chunk_sc[op_dim] + 1)) + new_unfilt_dim[op_dim] = TRUE; + else + new_unfilt_dim[op_dim] = FALSE; + } /* end if */ + else { fill_dim[op_dim] = FALSE; + new_unfilt_dim[op_dim] = FALSE; + } /* end else */ } /* end if */ - else + else { fill_dim[op_dim] = FALSE; + new_unfilt_dim[op_dim] = FALSE; + } /* end else */ + + /* If necessary, calculate the smallest offset of non-previously full + * chunks in this dimension, so we know these chunks were previously + * unfiltered */ + if(disable_edge_filters) + min_partial_chunk_sc[op_dim] = old_dim[op_dim] / chunk_dim[op_dim]; } /* end for */ /* Main loop: fill or remove chunks */ @@ -4416,8 +4899,27 @@ H5D__chunk_prune_by_extent(H5D_t *dset, hid_t dxpl_id, const hsize_t *old_dim) HDassert(fill_dim[op_dim]); HDassert(scaled[op_dim] == min_mod_chunk_sc[op_dim]); + /* Make sure this is an edge chunk */ + HDassert(H5D__chunk_is_partial_edge_chunk(space_ndims, layout->u.chunk.dim, scaled, space_dim)); + + /* Determine if the chunk just became an unfiltered chunk */ + if(new_unfilt_dim[op_dim]) { + new_unfilt_chunk = TRUE; + for(u = 0; u < space_ndims; u++) + if(scaled[u] == min_partial_chunk_sc[u]) { + new_unfilt_chunk = FALSE; + break; + } /* end if */ + } /* end if */ + + /* Make sure that, if we think this is a new unfiltered chunk, + * it was previously not an edge chunk */ + HDassert(!new_unfilt_dim[op_dim] || (!new_unfilt_chunk != + !H5D__chunk_is_partial_edge_chunk(space_ndims, layout->u.chunk.dim, scaled, old_dim))); + HDassert(!new_unfilt_chunk || new_unfilt_dim[op_dim]); + /* Fill the unused parts of the chunk */ - if(H5D__chunk_prune_fill(&udata) < 0) + if(H5D__chunk_prune_fill(&udata, new_unfilt_chunk) < 0) HGOTO_ERROR(H5E_DATASET, H5E_WRITEERROR, FAIL, "unable to write fill value") } /* end if */ else { @@ -4854,8 +5356,15 @@ H5D__chunk_copy_cb(const H5D_chunk_rec_t *chunk_rec, void *_udata) /* Check for filtered chunks */ if((is_vlen || fix_ref) && pline && pline->nused) { - must_filter = TRUE; - cb_struct.func = NULL; /* no callback function when failed */ + /* Check if we should disable filters on this chunk */ + if(udata->common.layout->flags + & H5O_LAYOUT_CHUNK_DONT_FILTER_PARTIAL_BOUND_CHUNKS) { + /* Check if the chunk is an edge chunk, and disable filters if so */ + if(!H5D__chunk_is_partial_edge_chunk(udata->dset_ndims, udata->common.layout->dim, chunk_rec->scaled, udata->dset_dims)) + must_filter = TRUE; + } /* end if */ + else + must_filter = TRUE; } /* end if */ /* Resize the buf if it is too small to hold the data */ @@ -4888,6 +5397,7 @@ H5D__chunk_copy_cb(const H5D_chunk_rec_t *chunk_rec, void *_udata) if(must_filter) { unsigned filter_mask = chunk_rec->filter_mask; + cb_struct.func = NULL; /* no callback function when failed */ if(H5Z_pipeline(pline, H5Z_FLAG_REVERSE, &filter_mask, H5Z_NO_EDC, cb_struct, &nbytes, &buf_size, &buf) < 0) HGOTO_ERROR(H5E_PLINE, H5E_CANTFILTER, H5_ITER_ERROR, "data pipeline read failed") } /* end if */ @@ -4966,7 +5476,7 @@ H5D__chunk_copy_cb(const H5D_chunk_rec_t *chunk_rec, void *_udata) udata_dst.common.layout->down_chunks, udata_dst.common.scaled); /* Allocate chunk in the file */ - if(H5D__chunk_file_alloc(udata->idx_info_dst, NULL, &udata_dst.chunk_block, &need_insert, udata_dst.common.scaled) < 0) + if(H5D__chunk_file_alloc(udata->idx_info_dst, NULL, &udata_dst.chunk_block, &need_insert) < 0) HGOTO_ERROR(H5E_DATASET, H5E_CANTINSERT, FAIL, "unable to insert/resize chunk on chunk level") /* Write chunk data to destination file */ @@ -5603,6 +6113,45 @@ done: /*------------------------------------------------------------------------- + * Function: H5D__chunk_is_partial_edge_chunk + * + * Purpose: Checks to see if the chunk is a partial edge chunk. + * Either dset or (dset_dims and dset_ndims) must be + * provided. + * + * Return: Non-negative on success/Negative on failure + * + * Programmer: Neil Fortner + * 19 Nov 2009 + * + *------------------------------------------------------------------------- + */ +static hbool_t +H5D__chunk_is_partial_edge_chunk(unsigned dset_ndims, const uint32_t *chunk_dims, + const hsize_t scaled[], const hsize_t *dset_dims) +{ + unsigned u; /* Local index variable */ + hbool_t ret_value = FALSE; /* Return value */ + + FUNC_ENTER_STATIC_NOERR + + /* Check args */ + HDassert(scaled); + HDassert(dset_ndims > 0); + HDassert(dset_dims); + HDassert(chunk_dims); + + /* check if this is a partial edge chunk */ + for(u = 0; u < dset_ndims; u++) + if(((scaled[u] + 1) * chunk_dims[u]) > dset_dims[u]) + HGOTO_DONE(TRUE); + +done: + FUNC_LEAVE_NOAPI(ret_value) +} /* H5D__chunk_is_partial_edge_chunk() */ + + +/*------------------------------------------------------------------------- * Function: H5D__chunk_file_alloc() * * Purpose: Chunk allocation: @@ -5618,7 +6167,7 @@ done: */ static herr_t H5D__chunk_file_alloc(const H5D_chk_idx_info_t *idx_info, const H5F_block_t *old_chunk, - H5F_block_t *new_chunk, hbool_t *need_insert, hsize_t scaled[]) + H5F_block_t *new_chunk, hbool_t *need_insert) { hbool_t alloc_chunk = FALSE; /* Whether to allocate chunk */ herr_t ret_value = SUCCEED; /* Return value */ @@ -5635,6 +6184,8 @@ H5D__chunk_file_alloc(const H5D_chk_idx_info_t *idx_info, const H5F_block_t *old HDassert(new_chunk); HDassert(need_insert); + *need_insert = FALSE; + /* Check for filters on chunks */ if(idx_info->pline->nused > 0) { /* Sanity/error checking block */ @@ -5696,7 +6247,7 @@ H5D__chunk_file_alloc(const H5D_chk_idx_info_t *idx_info, const H5F_block_t *old new_chunk->offset = H5MF_alloc(idx_info->f, H5FD_MEM_DRAW, idx_info->dxpl_id, (hsize_t)new_chunk->length); if(!H5F_addr_defined(new_chunk->offset)) HGOTO_ERROR(H5E_DATASET, H5E_CANTALLOC, FAIL, "file allocation failed") - *need_insert = TRUE; + *need_insert = TRUE; break; case H5D_CHUNK_IDX_NTYPES: |