/* Additonal optimizations: 1. slice covers all of exactly one chunk: we can just tranfer whole chunk to/from memory */ /********************************************************************* * Copyright 2018, UCAR/Unidata * See netcdf/COPYRIGHT file for copying and redistribution conditions. *********************************************************************/ #include "zincludes.h" #define WDEBUG #undef DFALTOPTIMIZE static int initialized = 0; static unsigned int optimize = 0; extern int NCZ_buildchunkkey(size_t R, const size64_t* chunkindices, char** keyp); /* 0 => no debug */ static unsigned int wdebug = 1; /* Forward */ static int NCZ_walk(NCZProjection** projv, NCZOdometer* chunkodom, NCZOdometer* slpodom, NCZOdometer* memodom, const struct Common* common, void* chunkdata); static int rangecount(NCZChunkRange range); static int readfromcache(void* source, size64_t* chunkindices, void** chunkdata); static int iswholechunk(struct Common* common,NCZSlice*); static int wholechunk_indices(struct Common* common, NCZSlice* slices, size64_t* chunkindices); const char* astype(int typesize, void* ptr) { switch(typesize) { case 4: { static char is[8]; snprintf(is,sizeof(is),"%u",*((unsigned int*)ptr)); return is; } break; default: break; } return "?"; } /**************************************************/ int ncz_chunking_init(void) { const char* val = NULL; #ifdef DFALTOPTIMIZE val = getenv("NCZ_NOOPTIMIZATION"); optimize = (val == NULL ? 1 : 0); #else optimize = 0; #endif val = getenv("NCZ_WDEBUG"); wdebug = (val == NULL ? 0 : atoi(val)); #ifdef WDEBUG if(wdebug > 0) fprintf(stderr,"wdebug=%u\n",wdebug); #endif initialized = 1; return NC_NOERR; } /**************************************************/ /** Goal: Given the slices being applied to the variable, create and walk all possible combinations of projection vectors that can be evaluated to provide the output data. Note that we do not actually pass NCZSlice but rather (start,count,stride) vectors. @param var Controlling variable @param usreading reading vs writing @param start start vector @param stop stop vector @param stride stride vector @param memory target or source of data @param typecode nc_type of type being written @param walkfcn fcn parameter to actually transfer data */ int NCZ_transferslice(NC_VAR_INFO_T* var, int reading, size64_t* start, size64_t* count, size64_t* stride, void* memory, nc_type typecode) { int r,stat = NC_NOERR; size64_t dimlens[NC_MAX_VAR_DIMS]; size64_t chunklens[NC_MAX_VAR_DIMS]; size64_t memshape[NC_MAX_VAR_DIMS]; NCZSlice slices[NC_MAX_VAR_DIMS]; struct Common common; NCZ_FILE_INFO_T* zfile = NULL; NCZ_VAR_INFO_T* zvar = NULL; size_t typesize; if(!initialized) ncz_chunking_init(); if((stat = NC4_inq_atomic_type(typecode, NULL, &typesize))) goto done; if(wdebug >= 1) { size64_t stopvec[NC_MAX_VAR_DIMS]; for(r=0;rndims;r++) stopvec[r] = start[r]+(count[r]*stride[r]); fprintf(stderr,"var: name=%s",var->hdr.name); fprintf(stderr," start=%s",nczprint_vector(var->ndims,start)); fprintf(stderr," count=%s",nczprint_vector(var->ndims,count)); fprintf(stderr," stop=%s",nczprint_vector(var->ndims,stopvec)); fprintf(stderr," stride=%s\n",nczprint_vector(var->ndims,stride)); } /* Fill in common */ memset(&common,0,sizeof(common)); common.var = var; common.file = (var->container)->nc4_info; zfile = common.file->format_file_info; zvar = common.var->format_var_info; common.reading = reading; common.memory = memory; common.typesize = typesize; common.cache = zvar->cache; /* We need to talk scalar into account */ common.rank = var->ndims; common.scalar = zvar->scalar; common.swap = (zfile->native_endianness == var->endianness ? 0 : 1); common.chunkcount = 1; if(common.scalar) { dimlens[0] = 1; chunklens[0] = 1; slices[0].start = 0; slices[0].stride = 1; slices[0].stop = 0; slices[0].len = 1; common.chunkcount = 1; memshape[0] = 1; } else for(r=0;rdim[r]->len; chunklens[r] = var->chunksizes[r]; slices[r].start = start[r]; slices[r].stride = stride[r]; slices[r].stop = minimum(start[r]+(count[r]*stride[r]),dimlens[r]); slices[r].len = dimlens[r]; common.chunkcount *= chunklens[r]; memshape[r] = count[r]; } if(wdebug >= 1) { fprintf(stderr,"\trank=%d",common.rank); if(!common.scalar) { fprintf(stderr," dimlens=%s",nczprint_vector(common.rank,dimlens)); fprintf(stderr," chunklens=%s",nczprint_vector(common.rank,chunklens)); fprintf(stderr," memshape=%s",nczprint_vector(common.rank,memshape)); } fprintf(stderr,"\n"); } common.dimlens = dimlens; /* BAD: storing stack vector in a pointer; do not free */ common.chunklens = chunklens; /* ditto */ common.memshape = memshape; /* ditto */ common.reader.source = ((NCZ_VAR_INFO_T*)(var->format_var_info))->cache; common.reader.read = readfromcache; if(common.scalar) { if((stat = NCZ_transferscalar(&common))) goto done; } else { if((stat = NCZ_transfer(&common, slices))) goto done; } done: NCZ_clearcommon(&common); return stat; } /* Walk the possible projections. Broken out so we can use it for unit testing @param common common parameters @param slices */ int NCZ_transfer(struct Common* common, NCZSlice* slices) { int stat = NC_NOERR; NCZOdometer* chunkodom = NULL; NCZOdometer* slpodom = NULL; NCZOdometer* memodom = NULL; void* chunkdata = NULL; int wholechunk = 0; /* We will need three sets of odometers. 1. Chunk odometer to walk the chunk ranges to get all possible combinations of chunkranges over all dimensions. 2. For each chunk odometer set of indices, we need a projection odometer that walks the set of projection slices for a given set of chunk ranges over all dimensions. 3. A memory odometer that walks the memory data to specify the locations in memory for read/write */ if(wdebug >= 2) { fprintf(stderr,"slices=%s\n",nczprint_slices(common->rank,slices)); } if((stat = NCZ_projectslices(common->dimlens, common->chunklens, slices, common, &chunkodom))) goto done; if(wdebug >= 4) { fprintf(stderr,"allprojections:\n%s",nczprint_allsliceprojections(common->rank,common->allprojections)); fflush(stderr); } wholechunk = iswholechunk(common,slices); if(wholechunk) { /* Implement a whole chunk read optimization; this is a rare occurrence where the the slices cover all of a single chunk. */ size64_t chunkindices[NC_MAX_VAR_DIMS]; unsigned char* memptr; unsigned char* slpptr; /* Which chunk are we getting? */ if((stat=wholechunk_indices(common,slices,chunkindices))) goto done; if(wdebug >= 1) fprintf(stderr,"case: wholechunk: chunkindices: %s\n",nczprint_vector(common->rank,chunkindices)); /* Read the chunk; handles fixed vs char* strings*/ switch ((stat = common->reader.read(common->reader.source, chunkindices, &chunkdata))) { case NC_EEMPTY: /* cache created the chunk */ break; case NC_NOERR: break; default: goto done; } /* Figure out memory address */ memptr = ((unsigned char*)common->memory); slpptr = ((unsigned char*)chunkdata); if(common->reading) { if((stat=NCZ_copy_data(common->file,common->var->type_info,slpptr,common->chunkcount,!ZCLEAR,memptr))) goto done; } else { if((stat=NCZ_copy_data(common->file,common->var->type_info,memptr,common->chunkcount,ZCLEAR,slpptr))) goto done; } // transfern(common,slpptr,memptr,common->chunkcount,1,chunkdata); if(zutest && zutest->tests & UTEST_WHOLECHUNK) zutest->print(UTEST_WHOLECHUNK, common, chunkindices); goto done; } /* iterate over the odometer: all combination of chunk indices in the projections */ for(;nczodom_more(chunkodom);) { int r; size64_t* chunkindices = NULL; NCZSlice slpslices[NC_MAX_VAR_DIMS]; NCZSlice memslices[NC_MAX_VAR_DIMS]; NCZProjection* proj[NC_MAX_VAR_DIMS]; size64_t shape[NC_MAX_VAR_DIMS]; chunkindices = nczodom_indices(chunkodom); if(wdebug >= 1) { fprintf(stderr,"chunkindices: %s\n",nczprint_vector(common->rank,chunkindices)); } for(r=0;rrank;r++) { NCZSliceProjections* slp = &common->allprojections[r]; NCZProjection* projlist = slp->projections; size64_t indexr = chunkindices[r]; /* use chunkindices[r] to find the corresponding projection slice */ /* We must take into account that the chunkindex of projlist[r] may be greater than zero */ /* note the 2 level indexing */ indexr -= slp->range.start; NCZProjection* pr = &projlist[indexr]; proj[r] = pr; } if(wdebug > 0) { fprintf(stderr,"Selected projections:\n"); for(r=0;rrank;r++) { fprintf(stderr,"\t[%d] %s\n",r,nczprint_projection(*proj[r])); shape[r] = proj[r]->iocount; } fprintf(stderr,"\tshape=%s\n",nczprint_vector(common->rank,shape)); } /* See if any of the projections is a skip; if so, then move to the next chunk indices */ for(r=0;rrank;r++) { if(proj[r]->skip) goto next; } for(r=0;rrank;r++) { slpslices[r] = proj[r]->chunkslice; memslices[r] = proj[r]->memslice; } if(zutest && zutest->tests & UTEST_TRANSFER) zutest->print(UTEST_TRANSFER, common, chunkodom, slpslices, memslices); /* Read from cache */ stat = common->reader.read(common->reader.source, chunkindices, &chunkdata); switch (stat) { case NC_EEMPTY: /* cache created the chunk */ break; case NC_NOERR: break; default: goto done; } slpodom = nczodom_fromslices(common->rank,slpslices); memodom = nczodom_fromslices(common->rank,memslices); { /* walk with odometer */ if(wdebug >= 1) fprintf(stderr,"case: odometer:\n"); /* This is the key action: walk this set of slices and transfer data */ if((stat = NCZ_walk(proj,chunkodom,slpodom,memodom,common,chunkdata))) goto done; } next: nczodom_free(slpodom); slpodom = NULL; nczodom_free(memodom); memodom = NULL; nczodom_next(chunkodom); } done: nczodom_free(slpodom); nczodom_free(memodom); nczodom_free(chunkodom); return stat; } #ifdef WDEBUG static void wdebug2(const struct Common* common, unsigned char* slpptr, unsigned char* memptr, size_t avail, size_t stride, void* chunkdata) { unsigned char* slpbase = chunkdata; unsigned char* membase = common->memory; unsigned slpoff = (unsigned)(slpptr - slpbase); unsigned memoff = (unsigned)(memptr - membase); unsigned slpidx = slpoff / common->typesize; unsigned memidx = memoff / common->typesize; unsigned value; fprintf(stderr,"wdebug2: %s: [%u/%d] %u->%u", common->reading?"read":"write", (unsigned)avail, (unsigned)stride, (unsigned)(common->reading?slpidx:memidx), (unsigned)(common->reading?memidx:slpidx) ); if(common->reading) value = ((unsigned*)slpptr)[0]; else value = ((unsigned*)memptr)[0]; fprintf(stderr,"; [%u]=%u",(unsigned)(common->reading?slpidx:memidx),value); fprintf(stderr,"\n"); } #else #define wdebug2(common,slpptr,memptr,avail,stride,chunkdata) #endif /* @param projv @param chunkodom @param slpodom @param memodom @param common @param chunkdata @return NC_NOERR */ static int NCZ_walk(NCZProjection** projv, NCZOdometer* chunkodom, NCZOdometer* slpodom, NCZOdometer* memodom, const struct Common* common, void* chunkdata) { int stat = NC_NOERR; for(;;) { size64_t slpoffset = 0; size64_t memoffset = 0; size64_t slpavail = 0; size64_t memavail = 0; size64_t laststride = 0; unsigned char* memptr0 = NULL; unsigned char* slpptr0 = NULL; if(!nczodom_more(slpodom)) break; if(wdebug >= 3) { fprintf(stderr,"xx.slp: odom: %s\n",nczprint_odom(slpodom)); fprintf(stderr,"xx.mem: odom: %s\n",nczprint_odom(memodom)); } /* Convert the indices to a linear offset WRT to chunk indices */ slpoffset = nczodom_offset(slpodom); memoffset = nczodom_offset(memodom); /* transfer data between these addresses */ memptr0 = ((unsigned char*)common->memory)+(memoffset * common->typesize); slpptr0 = ((unsigned char*)chunkdata)+(slpoffset * common->typesize); LOG((1,"%s: slpptr0=%p memptr0=%p slpoffset=%llu memoffset=%lld",__func__,slpptr0,memptr0,slpoffset,memoffset)); if(zutest && zutest->tests & UTEST_WALK) zutest->print(UTEST_WALK, common, chunkodom, slpodom, memodom); /* See if we can transfer multiple values at one shot */ laststride = slpodom->stride[common->rank-1]; if(laststride == 1) { slpavail = nczodom_avail(slpodom); /* How much can we read? */ memavail = nczodom_avail(memodom); assert(memavail == slpavail); nczodom_skipavail(slpodom); nczodom_skipavail(memodom); } else { slpavail = 1; memavail = 1; } if(slpavail > 0) { if(wdebug > 0) wdebug2(common,slpptr0,memptr0,slpavail,laststride,chunkdata); if(common->reading) { if((stat=NCZ_copy_data(common->file,common->var->type_info,slpptr0,slpavail,!ZCLEAR,memptr0))) goto done; } else { if((stat=NCZ_copy_data(common->file,common->var->type_info,memptr0,slpavail,ZCLEAR,slpptr0))) goto done; } } // if((stat = transfern(common,slpptr0,memptr0,avail,nczodom_laststride(slpodom),chunkdata)))goto done; nczodom_next(memodom); nczodom_next(slpodom); } done: return stat; } #if 0 static void wdebug1(const struct Common* common, unsigned char* srcptr, unsigned char* dstptr, size_t count, size_t stride, void* chunkdata, const char* tag) { unsigned char* dstbase = (common->reading?common->memory:chunkdata); unsigned char* srcbase = (common->reading?chunkdata:common->memory); unsigned dstoff = (unsigned)(dstptr - dstbase); unsigned srcoff = (unsigned)(srcptr - srcbase); unsigned srcidx = srcoff / sizeof(unsigned); fprintf(stderr,"%s: %s: [%u/%d] %u->%u", tag, common->reading?"read":"write", (unsigned)count, (unsigned)stride, (unsigned)(srcoff/common->typesize), (unsigned)(dstoff/common->typesize) ); #if 0 fprintf(stderr,"\t%s[%u]=%u\n",(common->reading?"chunkdata":"memdata"), // 0,((unsigned*)srcptr)[0] srcidx,((unsigned*)srcbase)[srcidx] ); #endif #if 0 { size_t len = common->typesize*count; fprintf(stderr," | [%u] %u->%u\n",(unsigned)len,(unsigned)srcoff,(unsigned)dstoff); } #endif fprintf(stderr,"\n"); } #else #define wdebug1(common,srcptr,dstptr,count,srcstride,dststride,chunkdata,tag) #endif #if 0 static int transfern(const struct Common* common, unsigned char* slpptr, unsigned char* memptr, size_t avail, size_t slpstride, void* chunkdata) { int stat = NC_NOERR; size_t typesize = common->typesize; size_t len = typesize*avail; size_t m,s; if(common->reading) { if(slpstride == 1) memcpy(memptr,slpptr,len); /* straight copy */ else { for(m=0,s=0;sswap) NCZ_swapatomicdata(len,memptr,common->typesize); } else { /*writing*/ unsigned char* srcbase = (common->reading?chunkdata:common->memory); unsigned srcoff = (unsigned)(memptr - srcbase); unsigned srcidx = srcoff / sizeof(unsigned); (void)srcidx; if(slpstride == 1) memcpy(slpptr,memptr,len); /* straight copy */ else { for(m=0,s=0;sswap) NCZ_swapatomicdata(len,slpptr,common->typesize); } return THROW(stat); } #endif #if 0 /* This function may not be necessary if code in zvar does it instead */ static int NCZ_fillchunk(void* chunkdata, struct Common* common) { int stat = NC_NOERR; if(common->fillvalue == NULL) { memset(chunkdata,0,common->chunkcount*common->typesize); goto done; } if(common->cache->fillchunk == NULL) { /* Get fill chunk*/ if((stat = NCZ_create_fill_chunk(common->cache->chunksize, common->typesize, common->fillvalue, &common->cache->fillchunk))) goto done; } memcpy(chunkdata,common->cache->fillchunk,common->cache->chunksize); done: return stat; } #endif /* Break out this piece so we can use it for unit testing */ int NCZ_projectslices(size64_t* dimlens, size64_t* chunklens, NCZSlice* slices, struct Common* common, NCZOdometer** odomp) { int stat = NC_NOERR; int r; NCZOdometer* odom = NULL; NCZSliceProjections* allprojections = NULL; NCZChunkRange ranges[NC_MAX_VAR_DIMS]; size64_t start[NC_MAX_VAR_DIMS]; size64_t stop[NC_MAX_VAR_DIMS]; size64_t stride[NC_MAX_VAR_DIMS]; size64_t len[NC_MAX_VAR_DIMS]; if((allprojections = calloc(common->rank,sizeof(NCZSliceProjections))) == NULL) {stat = NC_ENOMEM; goto done;} memset(ranges,0,sizeof(ranges)); /* Package common arguments */ common->dimlens = dimlens; common->chunklens = chunklens; /* Compute the chunk ranges for each slice in a given dim */ if((stat = NCZ_compute_chunk_ranges(common->rank,slices,common->chunklens,ranges))) goto done; /* Compute the slice index vector */ if((stat=NCZ_compute_all_slice_projections(common,slices,ranges,allprojections))) goto done; /* Verify */ for(r=0;rrank;r++) { assert(rangecount(ranges[r]) == allprojections[r].count); } /* Compute the shape vector */ for(r=0;rrank;r++) { int j; size64_t iocount = 0; NCZProjection* projections = allprojections[r].projections; for(j=0;jiocount; } common->shape[r] = iocount; } common->allprojections = allprojections; allprojections = NULL; /* Create an odometer to walk all the range combinations */ for(r=0;rrank;r++) { start[r] = ranges[r].start; stop[r] = ranges[r].stop; stride[r] = 1; len[r] = ceildiv(common->dimlens[r],common->chunklens[r]); } if((odom = nczodom_new(common->rank,start,stop,stride,len)) == NULL) {stat = NC_ENOMEM; goto done;} if(odomp) *odomp = odom; done: /* reclaim allprojections if !NULL */ if(allprojections != NULL) { NCZ_clearsliceprojections(common->rank,allprojections); nullfree(allprojections); } return stat; } /***************************************************/ /* Utilities */ static int rangecount(NCZChunkRange range) { return (range.stop - range.start); } /* Goal: Given a set of per-dimension indices, compute the corresponding linear position. */ size64_t NCZ_computelinearoffset(size_t R, const size64_t* indices, const size64_t* dimlens) { size64_t offset; int i; offset = 0; for(i=0;irank,common->allprojections); nullfree(common->allprojections); } /* Does the User want all of one and only chunk? */ static int iswholechunk(struct Common* common, NCZSlice* slices) { int i; /* Check that slices cover a whole chunk */ for(i=0;irank;i++) { if(!(slices[i].stride == 1 /* no point skipping */ && (slices[i].start % common->chunklens[i]) == 0 /* starting at beginning of chunk */ && (slices[i].stop - slices[i].start) /* stop-start = edge length */ == common->chunklens[i] /* edge length == chunk length */ )) return 0; /* slices do not cover a whole chunk */ } return 1; } static int wholechunk_indices(struct Common* common, NCZSlice* slices, size64_t* chunkindices) { int i; for(i=0;irank;i++) chunkindices[i] = (slices[i].start / common->chunklens[i]); return NC_NOERR; } /**************************************************/ /* Scalar variable support */ /* @param common common parameters */ int NCZ_transferscalar(struct Common* common) { int stat = NC_NOERR; void* chunkdata = NULL; size64_t chunkindices[NC_MAX_VAR_DIMS]; unsigned char* memptr, *slpptr; /* Read from single chunk from cache */ chunkindices[0] = 0; switch ((stat = common->reader.read(common->reader.source, chunkindices, &chunkdata))) { case NC_EEMPTY: /* cache created the chunk */ break; case NC_NOERR: break; default: goto done; } /* Figure out memory address */ memptr = ((unsigned char*)common->memory); slpptr = ((unsigned char*)chunkdata); if(common->reading) { if((stat=NCZ_copy_data(common->file,common->var->type_info,slpptr,common->chunkcount,!ZCLEAR,memptr))) goto done; } else { if((stat=NCZ_copy_data(common->file,common->var->type_info,memptr,common->chunkcount,ZCLEAR,slpptr))) goto done; } done: return stat; } /* Debugging Interface: return the contents of a specified chunk */ EXTERNL int NCZ_read_chunk(int ncid, int varid, size64_t* zindices, void* chunkdata) { int stat = NC_NOERR; NC_VAR_INFO_T* var = NULL; NCZ_VAR_INFO_T* zvar = NULL; struct NCZChunkCache* cache = NULL; void* cachedata = NULL; if ((stat = nc4_find_grp_h5_var(ncid, varid, NULL, NULL, &var))) return THROW(stat); zvar = (NCZ_VAR_INFO_T*)var->format_var_info; cache = zvar->cache; if((stat = NCZ_read_cache_chunk(cache,zindices,&cachedata))) goto done; if(chunkdata) { if((stat = nc_copy_data(ncid,var->type_info->hdr.id,cachedata,cache->chunkcount,chunkdata))) goto done; } done: return stat; }