@gmod/hic 1.3.2 → 2.1.0

package/dist/hicFile.js

@@ -10,7 +10,6 @@ const generic_filehandle2_1 = require("generic-filehandle2");
 const binWindow_ts_1 = require("./binWindow.js");
 const binary_ts_1 = __importDefault(require("./binary.js"));
 const bufferedFile_ts_1 = __importDefault(require("./bufferedFile.js"));
-const lru_ts_1 = __importDefault(require("./lru.js"));
 const matrix_ts_1 = __importDefault(require("./matrix.js"));
 const normalizationVector_ts_1 = __importDefault(require("./normalizationVector.js"));
 const reader_ts_1 = require("./reader.js");
@@ -25,6 +24,18 @@ const Short_MIN_VALUE = -32768;
 const DOUBLE = 8;
 const FLOAT = 4;
 const INT = 4;
+function getExpectedKey(type, unit, binSize) {
+    return `${type}_${unit}_${binSize}`;
+}
+// The range of `|bin2 - bin1|` over two half-open bin windows.
+function distanceRange([x1, x2], [y1, y2]) {
+    const overlap = x1 < y2 && y1 < x2;
+    const nearest = overlap
+        ? 0
+        : Math.min(Math.abs(y1 - (x2 - 1)), Math.abs(x1 - (y2 - 1)));
+    const farthest = Math.max(Math.abs(y2 - 1 - x1), Math.abs(x2 - 1 - y1));
+    return [nearest, farthest];
+}
 /**
  * Copy a partially-filled decode buffer down to its true length, or hand it
  * back untouched when the filter dropped nothing. A `subarray` would be free
@@ -65,15 +76,13 @@ function truncateRecords(records, n) {
 // The cliffs are exactly these three capacities. Note the first column: 41% of a
 // cold whole-genome fetch's reads were re-reads of normalization vectors the
 // same fetch had already fetched.
-const BLOCK_CACHE_MAX_ENTRIES = 1024;
-// The bound the entry cap was standing in for. Blocks became struct-of-arrays
-// (see contactRecords.ts), so a cached contact is 12 bytes of untraced
-// ArrayBuffer rather than ~55 on the GC-traced heap — but a block still holds
-// every contact in its bin square, which is the one thing here that can hold
-// real memory, and block sizes vary by more than an order of magnitude with
-// binsize. Measured on the file above: 0.05 MB at 2.5 Mb, 0.23 MB at 100 kb. So
-// the entry cap now tracks the working set and this tracks the memory, instead
-// of one number answering both badly. See LRU's WeightOpts.
+//
+// Blocks became struct-of-arrays (see contactRecords.ts), so a cached contact is
+// 12 bytes of untraced ArrayBuffer rather than ~55 on the GC-traced heap — but a
+// block still holds every contact in its bin square, which is the one thing here
+// that can hold real memory, and block sizes vary by more than an order of
+// magnitude with binsize. Measured on the file above: 0.05 MB at 2.5 Mb, 0.23 MB
+// at 100 kb. So the block cache's only bound is its byte budget.
 //
 // This is the one capacity here a caller can override (`blockCacheMaxBytes`),
 // because it is the only one that bounds MEMORY. The two entry caps below are
@@ -95,6 +104,9 @@ const NORM_VECTOR_CACHE_SIZE = 64;
 // hic-straw's extra `resolution` field only made the cache single-resolution
 // (every zoom step threw the previous level away).
 function blockCacheWeight(block) {
+    if (!block) {
+        return 0;
+    }
     const { bin1, bin2, counts } = block.records;
     return bin1.byteLength + bin2.byteLength + counts.byteLength;
 }
@@ -102,7 +114,7 @@ function getNormalizationVectorKey(type, chrIdx, unit, resolution) {
     return `${type}_${chrIdx}_${unit}_${resolution}`;
 }
 class HicFile {
-    config;
+    nvi;
     file;
     // These caches share the in-flight read, not just its result. A
     // multi-region fetch runs its region pairs concurrently and they share
@@ -112,6 +124,9 @@ class HicFile {
     // it has aborted, and drops a rejection so the next caller retries.
     headerCache = new shared_read_cache_1.SharedReadCache();
     normVectorIndexCache = new shared_read_cache_1.SharedReadCache();
+    // 'raw' and 'normalized', each walked once: the sections carry no index, so
+    // a chunk's position is known only by walking every chunk before it.
+    expectedSectionCache = new shared_read_cache_1.SharedReadCache();
     normVectorCache = new shared_read_cache_1.SharedReadCache({ maxSize: NORM_VECTOR_CACHE_SIZE });
     matrixCache = new shared_read_cache_1.SharedReadCache({
         maxSize: MATRIX_CACHE_SIZE,
@@ -124,22 +139,23 @@ class HicFile {
     version = 0;
     genomeId = '';
     footerPosition = 0;
+    // Where the raw (NONE) expected values begin, right after the master index.
+    expectedValuesPosition = 0;
     // Where the file's normalization data begins, and so also where a file
     // carrying none of it ends. Public so a test can cut a real file there to
     // make that second shape.
     normExpectedValueVectorsPosition;
     chromosomes = [];
-    chromosomeIndexMap = {};
-    chrAliasTable = {};
+    chromosomeIndexMap = new Map();
     bpResolutions = [];
     masterIndex = {};
     meta;
     constructor(config) {
-        const { filehandle, path, reader, blockCacheMaxBytes } = config;
-        this.config = config;
-        this.blockCache = new lru_ts_1.default(BLOCK_CACHE_MAX_ENTRIES, {
-            maxBytes: blockCacheMaxBytes ?? BLOCK_CACHE_MAX_BYTES,
-            weigh: blockCacheWeight,
+        const { filehandle, path, reader, nvi, blockCacheMaxBytes } = config;
+        this.nvi = nvi;
+        this.blockCache = new shared_read_cache_1.SharedReadCache({
+            maxSize: blockCacheMaxBytes ?? BLOCK_CACHE_MAX_BYTES,
+            sizeOf: blockCacheWeight,
         });
         if (reader) {
             this.file = reader;
@@ -177,7 +193,7 @@ class HicFile {
             throw new Error(`Unsupported hic version: ${version}`);
         }
         const footerPosition = binaryParser.getLong();
-        const { masterIndex, normExpectedValueVectorsPosition } = await this.readFooter(version, footerPosition, signal);
+        const { masterIndex, expectedValuesPosition, normExpectedValueVectorsPosition, } = await this.readFooter(version, footerPosition, signal);
         const bodyPosition = Object.values(masterIndex).reduce((min, entry) => Math.min(min, entry.start), Number.MAX_VALUE);
         data = await this.file.read(16, bodyPosition - 16, { signal });
         binaryParser = new binary_ts_1.default(new DataView(data));
@@ -194,7 +210,7 @@ class HicFile {
             binaryParser.getString(); // attribute value
         }
         const chromosomes = [];
-        const chromosomeIndexMap = {};
+        const chromosomeIndexMap = new Map();
         let nChrs = binaryParser.getInt();
         let i = 0;
         while (nChrs-- > 0) {
@@ -204,7 +220,7 @@ class HicFile {
                 size: version < 9 ? binaryParser.getInt() : binaryParser.getLong(),
             };
             chromosomes.push(chr);
-            chromosomeIndexMap[chr.name] = chr.index;
+            chromosomeIndexMap.set(chr.name, chr.index);
             i++;
         }
         const bpResolutions = [];
@@ -212,21 +228,10 @@ class HicFile {
         while (nBpResolutions-- > 0) {
             bpResolutions.push(binaryParser.getInt());
         }
-        const chrAliasTable = {};
-        for (const chrName of Object.keys(chromosomeIndexMap)) {
-            if (chrName.startsWith('chr')) {
-                chrAliasTable[chrName.slice(3)] = chrName;
-            }
-            else if (chrName === 'MT') {
-                chrAliasTable.chrM = chrName;
-            }
-            else {
-                chrAliasTable[`chr${chrName}`] = chrName;
-            }
-        }
         (0, shared_read_cache_1.throwIfAborted)(signal);
         this.version = version;
         this.footerPosition = footerPosition;
+        this.expectedValuesPosition = expectedValuesPosition;
         this.masterIndex = masterIndex;
         this.normExpectedValueVectorsPosition = normExpectedValueVectorsPosition;
         this.genomeId = genomeId;
@@ -235,7 +240,6 @@ class HicFile {
         this.chromosomes = chromosomes;
         this.chromosomeIndexMap = chromosomeIndexMap;
         this.bpResolutions = bpResolutions;
-        this.chrAliasTable = chrAliasTable;
         this.meta = {
             version,
             genome: genomeId,
@@ -262,9 +266,14 @@ class HicFile {
             const size = binaryParser.getInt();
             masterIndex[key] = { start: pos, size };
         }
+        const expectedValuesPosition = footerPosition + skip + binaryParser.position;
         // Normalized expected values start after the expected values
         const normExpectedValueVectorsPosition = version > 5 ? footerPosition + (version < 9 ? 4 : 8) + nBytes : undefined;
-        return { masterIndex, normExpectedValueVectorsPosition };
+        return {
+            masterIndex,
+            expectedValuesPosition,
+            normExpectedValueVectorsPosition,
+        };
     }
     async getMatrix(chrIdx1, chrIdx2, opts) {
         await this.init(opts);
@@ -279,10 +288,10 @@ class HicFile {
         }
         return matrix;
     }
-    async getContactRecords(normalization, region1, region2, units, binsize, opts) {
+    async getContactRecords(normalization, region1, region2, binsize, opts) {
         await this.init(opts);
-        const idx1 = this.chromosomeIndexMap[this.getFileChrName(region1.chr)];
-        const idx2 = this.chromosomeIndexMap[this.getFileChrName(region2.chr)];
+        const idx1 = this.chromosomeIndexMap.get(region1.chr);
+        const idx2 = this.chromosomeIndexMap.get(region2.chr);
         // A `.hic` stores only `bin1 <= bin2`, so a query whose x window sits to the
         // right of its y window has to be swapped or it asks for the half of the
         // matrix that does not exist. The same-chromosome test compares **starts**,
@@ -318,7 +327,7 @@ class HicFile {
         // resolved once per pair rather than per block, each paired with the bin
         // offset its values start at.
         const [norm, blocks] = await Promise.all([
-            this.getNormVectors(normalization, r1, r2, units, binsize, opts),
+            this.getNormVectors(normalization, r1, r2, binsize, opts),
             // The measurable half of the pair: blocks are counted and the two norm
             // vectors are not, because the vector chain is two hops whatever the
             // query and the block chain is the one that grows with the region.
@@ -393,45 +402,41 @@ class HicFile {
         //
         // `transposed` says the query was swapped, so `bin1` runs along `region2`.
         // Reported rather than left for the caller to re-derive: it is decided here
-        // from this file's own alias table and chromosome indices, and a caller
-        // re-deriving it off a divergent chr-name scheme would silently un-swap the
-        // wrong axis.
+        // from this file's chromosome indices, which a caller would otherwise have
+        // to look up again to know which axis to un-swap.
         return {
             records: contactRecords,
             appliedNormalization: norm ? normalization : 'NONE',
             transposed: transpose,
         };
     }
-    async getNormVectors(normalization, r1, r2, units, binsize, opts) {
+    async getNormVectors(normalization, r1, r2, binsize, opts) {
+        if (!normalization || normalization === 'NONE') {
+            return undefined;
+        }
         const signalOpts = { signal: opts?.signal };
-        let result;
-        if (normalization && normalization !== 'NONE') {
-            const chr1 = this.getFileChrName(r1.chr);
-            const chr2 = this.getFileChrName(r2.chr);
-            const offset1 = Math.floor(r1.start / binsize);
-            const offset2 = Math.floor(r2.start / binsize);
-            const nv1 = await this.getNormalizationVector(normalization, chr1, units, binsize, signalOpts);
-            const nv2 = chr1 === chr2
-                ? nv1
-                : await this.getNormalizationVector(normalization, chr2, units, binsize, signalOpts);
-            if (nv1 && nv2) {
-                result = {
-                    v1: await nv1.getValues(offset1, Math.ceil(r1.end / binsize), signalOpts),
-                    v2: await nv2.getValues(offset2, Math.ceil(r2.end / binsize), signalOpts),
-                    offset1,
-                    offset2,
-                };
-            }
+        const [offset1, end1] = (0, binWindow_ts_1.binWindow)(r1, binsize);
+        const [offset2, end2] = (0, binWindow_ts_1.binWindow)(r2, binsize);
+        // Awaited in sequence, an inter-chromosomal pair would pay four
+        // round-trip waves for what is two hops per vector.
+        const [nv1, nv2] = await Promise.all([
+            this.getNormalizationVector(normalization, r1.chr, binsize, signalOpts),
+            this.getNormalizationVector(normalization, r2.chr, binsize, signalOpts),
+        ]);
+        if (!nv1 || !nv2) {
+            return undefined;
         }
-        return result;
+        const [v1, v2] = await Promise.all([
+            nv1.getValues(offset1, end1, signalOpts),
+            nv2.getValues(offset2, end2, signalOpts),
+        ]);
+        return { v1, v2, offset1, offset2 };
     }
     async getBlocks(region1, region2, binSize, opts) {
         const blockKey = (blockNumber, zd) => `${zd.getKey()}_${blockNumber}`;
         await this.init(opts);
-        const chr1 = this.getFileChrName(region1.chr);
-        const chr2 = this.getFileChrName(region2.chr);
-        const idx1 = this.chromosomeIndexMap[chr1];
-        const idx2 = this.chromosomeIndexMap[chr2];
+        const idx1 = this.chromosomeIndexMap.get(region1.chr);
+        const idx2 = this.chromosomeIndexMap.get(region2.chr);
         let blocks = [];
         if (idx1 === undefined) {
             console.warn(`No chromosome named: ${region1.chr}`);
@@ -452,43 +457,37 @@ class HicFile {
                     throw new Error(`${exports.NO_DATA_FOR_RESOLUTION}: ${binSize} for map ${region1.chr}-${region2.chr}`);
                 }
                 const blockNumbers = zd.getBlockNumbers(region1, region2, this.version);
-                const blockNumbersToQuery = [];
-                for (const num of blockNumbers) {
-                    const cached = this.blockCache.get(blockKey(num, zd));
-                    if (cached) {
-                        blocks.push(cached);
-                    }
-                    else {
-                        blockNumbersToQuery.push(num);
-                    }
-                }
+                const keys = blockNumbers.map(num => blockKey(num, zd));
+                const cached = keys.map(key => this.blockCache.has(key));
                 // Cached blocks are already done, not absent: a pan that reuses most
                 // of its blocks reads most of the way along rather than restarting at
                 // zero, which is what the bar should say — the work left is the reads.
+                // A block another fetch has in flight counts as cached too.
                 //
                 // A query covering no blocks at all reports nothing rather than `0, 0`:
                 // there is no fraction in it, and a caller dividing would get NaN.
                 const { onProgress } = opts ?? {};
-                let done = blocks.length;
-                if (blockNumbers.length > 0) {
-                    onProgress?.(done, blockNumbers.length);
+                const total = blockNumbers.length;
+                let done = cached.filter(Boolean).length;
+                if (total > 0) {
+                    onProgress?.(done, total);
                 }
-                // Still one `Promise.all`, so every block is still issued in one wave:
-                // what a remote file pays is round-trip DEPTH, and counting the
-                // completions must not turn one wave into a queue. `readChainDepth`
-                // pins that.
-                const newBlocks = await Promise.all(blockNumbersToQuery.map(async (blockNumber) => {
-                    const block = await this.readBlock(blockNumber, zd, signalOpts);
-                    done++;
-                    onProgress?.(done, blockNumbers.length);
+                // One `Promise.all`, so every block is issued in one wave: what a
+                // remote file pays is round-trip DEPTH, and counting the completions
+                // must not turn one wave into a queue. `readChainDepth` pins that.
+                //
+                // A block number missing from the index has no data, so it skips the
+                // cache, which would otherwise keep a weightless entry for it forever.
+                blocks = await Promise.all(blockNumbers.map(async (num, i) => {
+                    const block = zd.blockIndex[num]
+                        ? await this.blockCache.get(keys[i], opts?.signal, signal => this.readBlock(num, zd, { signal }))
+                        : undefined;
+                    if (!cached[i]) {
+                        done++;
+                        onProgress?.(done, total);
+                    }
                     return block;
                 }));
-                for (const block of newBlocks) {
-                    if (block) {
-                        this.blockCache.set(blockKey(block.blockNumber, zd), block);
-                    }
-                }
-                blocks = blocks.concat(newBlocks);
             }
         }
         return blocks;
@@ -595,13 +594,13 @@ class HicFile {
         }
         throw new Error(`Unknown block type: ${type}`);
     }
-    async getNormalizationVector(type, chr, unit, binSize, opts) {
+    async getNormalizationVector(type, chr, binSize, opts) {
         await this.init(opts);
-        const chrIdx = this.chromosomeIndexMap[this.getFileChrName(chr)];
+        const chrIdx = this.chromosomeIndexMap.get(chr);
         if (chrIdx === undefined) {
             return undefined;
         }
-        const key = getNormalizationVectorKey(type, chrIdx, unit, binSize);
+        const key = getNormalizationVectorKey(type, chrIdx, 'BP', binSize);
         // Sharing the in-flight read is what keeps concurrent region pairs sharing
         // one vector: a chromosome appears in every pair it takes part in, so with a
         // result-only cache each of those pairs missed while the first was still in
@@ -658,8 +657,8 @@ class HicFile {
                 size: this.normVectorIndexSize,
             }, opts.signal);
         }
-        else if (this.config.nvi) {
-            const nviArray = decodeURIComponent(this.config.nvi).split(',');
+        else if (this.nvi) {
+            const nviArray = decodeURIComponent(this.nvi).split(',');
             return this.readNormVectorIndex({
                 start: parseInt(nviArray[0]),
                 size: parseInt(nviArray[1]),
@@ -702,102 +701,149 @@ class HicFile {
         const { signal } = opts;
         let index;
         if (this.normExpectedValueVectorsPosition !== undefined) {
-            const nviStart = await this.skipExpectedValues(this.normExpectedValueVectorsPosition, opts);
-            let byteCount = INT;
-            let data = await this.file.read(nviStart, INT, { signal });
+            const nviStart = (await this.getExpectedSection('normalized', opts)).end;
+            const data = await this.file.read(nviStart, INT, { signal });
             // Possible if there are no norm vectors. A legal v8 file, though uncommon.
             if (data.byteLength !== 0) {
-                const binaryParser = new binary_ts_1.default(new DataView(data));
-                const nEntries = binaryParser.getInt();
-                const sizeEstimate = nEntries * 30;
-                data = await this.file.read(nviStart + byteCount, sizeEstimate, {
-                    signal,
-                });
+                const nEntries = new binary_ts_1.default(new DataView(data)).getInt();
                 const found = { entries: {}, types: ['NONE'] };
-                const processEntries = async (remaining, buf) => {
-                    const parser = new binary_ts_1.default(new DataView(buf));
-                    let n = remaining;
-                    while (n-- > 0) {
-                        if (parser.available() < 100) {
-                            n++; // Reset counter as entry is not processed
-                            byteCount += parser.position;
-                            const est = Math.max(1000, n * 30);
-                            const more = await this.file.read(nviStart + byteCount, est, {
-                                signal,
-                            });
-                            await processEntries(n, more);
-                            return;
-                        }
-                        this.parseNormVectorEntry(parser, found);
+                let position = nviStart + INT;
+                let requested = nEntries * 30;
+                let buf = await this.file.read(position, requested, { signal });
+                let parser = new binary_ts_1.default(new DataView(buf));
+                for (let n = nEntries; n > 0; n--) {
+                    // A short read means the buffer already runs to EOF, and refilling
+                    // it would only return the same bytes again.
+                    if (parser.available() < 100 && buf.byteLength === requested) {
+                        position += parser.position;
+                        requested = Math.max(1000, n * 30);
+                        buf = await this.file.read(position, requested, { signal });
+                        parser = new binary_ts_1.default(new DataView(buf));
                     }
-                    byteCount += parser.position;
-                };
-                await processEntries(nEntries, data);
+                    this.parseNormVectorEntry(parser, found);
+                }
                 (0, shared_read_cache_1.throwIfAborted)(signal);
-                this.config.nvi = `${nviStart},${byteCount}`;
                 index = found;
             }
         }
         return index;
     }
-    // Used when the position of the norm vector index is unknown: skip the
-    // normalized expected values to find the index.
-    async skipExpectedValues(start, opts) {
+    /**
+     * The expected count for a region pair at `binsize`, the denominator of
+     * observed/expected, as straw computes it. `normalization` must be the one
+     * the pair's records carry (`getContactRecords`' `appliedNormalization`),
+     * since each normalization has its own vector. Undefined where the file has
+     * no expected value for it: no vector at this binsize, or a version 5 file
+     * asked for a normalized one.
+     *
+     * Within a chromosome the vector is read for only the distances the pair
+     * spans. Between chromosomes straw divides by the map's average contact, the
+     * raw `sumCounts` over its bins, whatever the normalization.
+     */
+    async getExpectedValues(normalization, region1, region2, binsize, opts) {
+        await this.init(opts);
+        const idx1 = this.chromosomeIndexMap.get(region1.chr);
+        const idx2 = this.chromosomeIndexMap.get(region2.chr);
+        if (idx1 === undefined || idx2 === undefined) {
+            return undefined;
+        }
+        if (idx1 !== idx2) {
+            const matrix = await this.getMatrix(idx1, idx2, opts);
+            const zd = matrix?.getZoomData(binsize);
+            if (!zd) {
+                return undefined;
+            }
+            // straw's bin counts leave out a chromosome's partial last bin
+            const nBins1 = Math.max(1, Math.floor(this.chromosomes[idx1].size / binsize));
+            const nBins2 = Math.max(1, Math.floor(this.chromosomes[idx2].size / binsize));
+            return { kind: 'inter', average: zd.sumCounts / nBins1 / nBins2 };
+        }
+        const normalized = !!normalization && normalization !== 'NONE';
+        if (normalized && this.normExpectedValueVectorsPosition === undefined) {
+            return undefined;
+        }
+        const section = await this.getExpectedSection(normalized ? 'normalized' : 'raw', opts ?? {});
+        const chunk = section.chunks.get(getExpectedKey(normalized ? normalization : 'NONE', 'BP', binsize));
+        if (!chunk || chunk.nValues === 0) {
+            return undefined;
+        }
+        const [first, last] = distanceRange((0, binWindow_ts_1.binWindow)(region1, binsize), (0, binWindow_ts_1.binWindow)(region2, binsize));
+        const signalOpts = { signal: opts?.signal };
+        const stored = await chunk.vector.getValues(first, last + 1, signalOpts);
+        const tail = stored.length < last + 1 - first
+            ? (await chunk.vector.getValues(chunk.nValues - 1, chunk.nValues, signalOpts))[0]
+            : 0;
+        // straw divides by 1 where the section names no factor for the chromosome
+        const factor = chunk.chrScaleFactors.get(idx1) ?? 1;
+        const values = new Float64Array(last + 1 - first);
+        for (let i = 0; i < values.length; i++) {
+            values[i] = (i < stored.length ? stored[i] : tail) / factor;
+        }
+        return { kind: 'intra', firstDistance: first, values };
+    }
+    async getExpectedSection(section, opts) {
+        return this.expectedSectionCache.get(section, opts.signal, signal => this.walkExpectedSection(section === 'raw'
+            ? this.expectedValuesPosition
+            : this.normExpectedValueVectorsPosition, section === 'normalized', { signal, onProgress: opts.onProgress }));
+    }
+    // Walk one expected-value section, recording where each chunk's values sit
+    // and its chromosome scale factors. The normalized section's end is also
+    // where a pre-v9 file's normalization vector index begins.
+    //
+    // A chunk's two reads land a whole expected-value vector apart — one per bin
+    // of the genome's longest chromosome, so megabytes at a fine binsize — which
+    // is further than the buffer reaches. On a remote file that is two round
+    // trips per chunk that nothing can merge, and it is why this walk, not the
+    // index it leads to, is the slow part of opening a pre-v9 file. It is also
+    // why the chunk is the unit worth counting: known before the first read, each
+    // costing about the same. Chunks are walked in order, since each one's size
+    // is known only once the one before it has been parsed.
+    async walkExpectedSection(start, normalized, opts) {
         const { signal, onProgress } = opts;
         const version = this.version;
+        const valueSize = version < 9 ? DOUBLE : FLOAT;
         const file = new bufferedFile_ts_1.default({ file: this.file, size: 256000 });
+        const chunks = new Map();
         const data = await file.read(start, INT, { signal });
-        // A file with no normalization at all ends where this section would start,
-        // so the count this is about to read is past EOF. That is the same "no norm
-        // vectors" case `readNormExpectedValuesAndNormVectorIndex` guards its own
-        // read for, and it has to be caught HERE too, because this runs first:
-        // a v9 file that records no norm-vector-index position takes the v8
-        // discovery path, and that path starts by skipping a section that isn't
-        // there. Answering `start` leaves the caller's guard to do the rest.
+        // A file with no normalization at all ends where the normalized section
+        // would start, so the count this is about to read is past EOF. A v9 file
+        // that records no norm-vector-index position takes the v8 discovery path,
+        // which starts by walking a section that isn't there, so the empty answer
+        // has to come from here.
         if (data.byteLength < INT) {
-            return start;
+            return { chunks, end: start };
         }
-        const binaryParser = new binary_ts_1.default(new DataView(data));
-        const nEntries = binaryParser.getInt(); // Total # of expected value chunks
-        // Skip one chunk, answering where the next one starts.
-        //
-        // Its two reads land a whole expected-value vector apart — one per bin of
-        // the genome, so tens of MB at a fine binsize — which is further than the
-        // buffer reaches. On a remote file that is two round trips per chunk that
-        // nothing can merge, and it is why this walk, not the index it is looking
-        // for, is the slow part of opening a pre-v9 file. It is also why the chunk
-        // is the unit worth counting: `nEntries` of them, known before the first
-        // read, each costing about the same.
-        const skipChunk = async (chunkStart) => {
-            let chunkSize = 0;
-            let buf = await file.read(chunkStart, 500, { signal });
-            let parser = new binary_ts_1.default(new DataView(buf));
-            parser.getString(); // type
-            parser.getString(); // unit
-            parser.getInt(); // binSize
-            const nValues = version < 9 ? parser.getInt() : parser.getLong();
-            chunkSize += parser.position + nValues * (version < 9 ? DOUBLE : FLOAT);
-            buf = await file.read(chunkStart + chunkSize, INT, { signal });
-            parser = new binary_ts_1.default(new DataView(buf));
-            const nChrScaleFactors = parser.getInt();
-            chunkSize +=
-                INT + nChrScaleFactors * (INT + (version < 9 ? DOUBLE : FLOAT));
-            return chunkStart + chunkSize;
-        };
-        // A loop rather than the tail recursion this was, because the position is
-        // the only thing carried between chunks and a counter now rides along with
-        // it. Chunks must be walked in order — each one's size is only known once
-        // the one before it has been parsed — so the awaits are sequential by
-        // nature, not by oversight.
+        const nEntries = new binary_ts_1.default(new DataView(data)).getInt();
         if (nEntries > 0) {
             onProgress?.(0, nEntries);
         }
         let position = start + INT;
         for (let done = 0; done < nEntries; done++) {
-            position = await skipChunk(position);
+            let parser = new binary_ts_1.default(new DataView(await file.read(position, 500, { signal })));
+            const type = normalized ? parser.getString() : 'NONE';
+            const unit = parser.getString();
+            const binSize = parser.getInt();
+            const nValues = version < 9 ? parser.getInt() : parser.getLong();
+            const valuesPosition = position + parser.position;
+            const factorsPosition = valuesPosition + nValues * valueSize;
+            parser = new binary_ts_1.default(new DataView(await file.read(factorsPosition, INT, { signal })));
+            const nFactors = parser.getInt();
+            const factorBytes = nFactors * (INT + valueSize);
+            parser = new binary_ts_1.default(new DataView(await file.read(factorsPosition + INT, factorBytes, { signal })));
+            const chrScaleFactors = new Map();
+            for (let i = 0; i < nFactors; i++) {
+                const chrIdx = parser.getInt();
+                chrScaleFactors.set(chrIdx, version < 9 ? parser.getDouble() : parser.getFloat());
+            }
+            chunks.set(getExpectedKey(type, unit, binSize), {
+                vector: new normalizationVector_ts_1.default(this.file, valuesPosition, nValues, valueSize),
+                nValues,
+                chrScaleFactors,
+            });
+            position = factorsPosition + INT + factorBytes;
             onProgress?.(done + 1, nEntries);
         }
-        return position;
+        return { chunks, end: position };
     }
     parseNormVectorEntry(binaryParser, index) {
         const type = binaryParser.getString(); // 15
@@ -806,15 +852,12 @@ class HicFile {
         const binSize = binaryParser.getInt(); // 4
         const filePosition = binaryParser.getLong(); // 8
         const sizeInBytes = this.version < 9 ? binaryParser.getInt() : binaryParser.getLong(); // 4:8
-        const key = `${type}_${chrIdx}_${unit}_${binSize}`;
+        const key = getNormalizationVectorKey(type, chrIdx, unit, binSize);
         if (!index.types.includes(type)) {
             index.types.push(type);
         }
         index.entries[key] = { filePosition, size: sizeInBytes };
     }
-    getFileChrName(chrAlias) {
-        return this.chrAliasTable[chrAlias] ?? chrAlias;
-    }
 }
 exports.HicFile = HicFile;
 //# sourceMappingURL=hicFile.js.map