wgblas 0.1.2 → 1.0.0

package/src/isamax/isamax.mjs

@@ -34,82 +34,81 @@ export async function isamax(device, n, x, incx) {
   const pipelineMain = await getPipeline(device, "isamax");
   const pipelineReduce = await getPipeline(device, "reduction/argmax");
 
-  const xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "isamax-x", false);
-  const partialsValBuffer = createStorageBuffer(
-    2 * WGS * 4,
-    "isamax-partials-val",
-  ); //to hold 2*WGS partial max values of f32
-  const partialsIdxBuffer = createStorageBuffer(
-    2 * WGS * 4,
-    "isamax-partials-idx",
-  ); //to hold 2*WGS partial max indices of u32
-  const resultBuffer = createResultBuffer(4, "isamax-result"); // u32 index
-  const paramsBuffer = createParamsBuffer(
-    [
-      { value: n, type: "u32" },
-      { value: incx, type: "u32" },
-    ],
-    "isamax-params",
-  );
+  let xBuffer = null;
+  let partialsValBuffer = null;
+  let partialsIdxBuffer = null;
+  let resultBuffer = null;
+  let paramsBuffer = null;
+  let readBuffer = null;
 
-  const bgMain = createBindGroup(pipelineMain.getBindGroupLayout(0), [
-    xBuffer,
-    partialsValBuffer,
-    partialsIdxBuffer,
-    paramsBuffer,
-  ]);
-  const { commandEncoder: enc1, ts: ts1 } = runComputePass(
-    pipelineMain,
-    bgMain,
-    2 * WGS,
-  ); //dispatch 2*WGS workgroups
-
-  submit(enc1);
-
-  const bgReduce = createBindGroup(pipelineReduce.getBindGroupLayout(0), [
-    partialsValBuffer,
-    partialsIdxBuffer,
-    resultBuffer,
-  ]);
-  const { commandEncoder: enc2, ts: ts2 } = runComputePass(
-    pipelineReduce,
-    bgReduce,
-    1,
-  ); // dispatch 1 workgroup to reduce the partial sums to a single result
-  const readBuffer = stageReadback(enc2, resultBuffer);
-
-  submit(enc2);
+  try {
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "isamax-x", false);
+    partialsValBuffer = createStorageBuffer(
+      2 * WGS * 4,
+      "isamax-partials-val",
+    ); //to hold 2*WGS partial max values of f32
+    partialsIdxBuffer = createStorageBuffer(
+      2 * WGS * 4,
+      "isamax-partials-idx",
+    ); //to hold 2*WGS partial max indices of u32
+    resultBuffer = createResultBuffer(4, "isamax-result"); // u32 index
+    paramsBuffer = createParamsBuffer(
+      [
+        { value: n, type: "u32" },
+        { value: incx, type: "u32" },
+      ],
+      "isamax-params",
+    );
 
-  const [gpuTime1, gpuTime2, idxArr] = await Promise.all([
-    extractTimestamp(ts1),
-    extractTimestamp(ts2),
-    extractResult(readBuffer, Uint32Array),
-  ]);
+    const bgMain = createBindGroup(pipelineMain.getBindGroupLayout(0), [
+      xBuffer,
+      partialsValBuffer,
+      partialsIdxBuffer,
+      paramsBuffer,
+    ]);
+    const { commandEncoder: enc1, ts: ts1 } = runComputePass(
+      pipelineMain,
+      bgMain,
+      2 * WGS,
+    ); //dispatch 2*WGS workgroups
 
-  const index = idxArr[0];
+    submit(enc1);
 
-  if (xIsGpu) {
-    destroyBuffers(
+    const bgReduce = createBindGroup(pipelineReduce.getBindGroupLayout(0), [
       partialsValBuffer,
       partialsIdxBuffer,
       resultBuffer,
-      paramsBuffer,
-      readBuffer,
-    );
+    ]);
+    const { commandEncoder: enc2, ts: ts2 } = runComputePass(
+      pipelineReduce,
+      bgReduce,
+      1,
+    ); // dispatch 1 workgroup to reduce the partial sums to a single result
+    readBuffer = stageReadback(enc2, resultBuffer);
+
+    submit(enc2);
+
+    const resultPromise = extractResult(readBuffer, Uint32Array);
+    readBuffer = null; // ownership transferred — extractResult's own finally destroys it
+
+    const [gpuTime1, gpuTime2, idxArr] = await Promise.all([
+      extractTimestamp(ts1),
+      extractTimestamp(ts2),
+      resultPromise,
+    ]);
+
+    const index = idxArr[0];
+
     if (gpuTime1 !== undefined && gpuTime2 !== undefined)
       return { index, gpuTimeMs: gpuTime1 + gpuTime2 };
     return { index };
+  } finally {
+    if (!xIsGpu && xBuffer) destroyBuffers(xBuffer);
+    if (partialsValBuffer) destroyBuffers(partialsValBuffer);
+    if (partialsIdxBuffer) destroyBuffers(partialsIdxBuffer);
+    if (resultBuffer) destroyBuffers(resultBuffer);
+    if (paramsBuffer) destroyBuffers(paramsBuffer);
+    // Only reached if submit(enc2) threw before ownership was transferred above.
+    if (readBuffer) destroyBuffers(readBuffer);
   }
-
-  destroyBuffers(
-    xBuffer,
-    partialsValBuffer,
-    partialsIdxBuffer,
-    resultBuffer,
-    paramsBuffer,
-    readBuffer,
-  );
-  if (gpuTime1 !== undefined && gpuTime2 !== undefined)
-    return { index, gpuTimeMs: gpuTime1 + gpuTime2 };
-  return { index };
 }

package/src/random/random.d.mts

@@ -59,3 +59,40 @@ export declare function randomFloat64Array(
   low?: number,
   high?: number,
 ): Float64Array;
+
+/**
+ * Returns a Float32Array of n*lda elements, row-major with leading dimension
+ * `lda`, representing an actual lower- or upper-triangular matrix for a
+ * triangular routine (`strmv`/`strsv`) example: entries in the `uplo`
+ * triangle are uniform in `[low, high)`, the n diagonal entries
+ * (`A[i*lda+i]`) are uniform in `[diagLow, diagHigh)` — kept well away from 0
+ * so a triangular solve doesn't divide by a near-zero pivot — and every entry
+ * in the other triangle is 0.
+ *
+ * @param n - matrix order (rows/cols read by the triangular routine)
+ * @param lda - leading dimension; throws if `lda < n`
+ * @param uplo - `'lower'` to fill the lower triangle, `'upper'` to fill the upper triangle (default: `'lower'`)
+ * @param low - lower bound for off-diagonal entries (default: -1)
+ * @param high - upper bound for off-diagonal entries (default: 1)
+ * @param diagLow - lower bound for diagonal entries (default: 5)
+ * @param diagHigh - upper bound for diagonal entries (default: 15)
+ *
+ * @example
+ * ```js
+ * import { randomTriangularFloat32Array } from "wgblas";
+ *
+ * const n = 4, lda = n;
+ * const A = randomTriangularFloat32Array(n, lda, "lower");
+ * ```
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/random/random.mjs#L13">Source code: random.mjs (L13)</a>
+ * @category Utilities
+ */
+export declare function randomTriangularFloat32Array(
+  n: number,
+  lda: number,
+  uplo?: 'lower' | 'upper',
+  low?: number,
+  high?: number,
+  diagLow?: number,
+  diagHigh?: number,
+): Float32Array;

package/src/random/random.mjs

@@ -9,3 +9,20 @@ export function randomFloat64Array(n, low = -1, high = 1) {
   for (let i = 0; i < n; i++) x[i] = low + Math.random() * (high - low);
   return x;
 }
+
+export function randomTriangularFloat32Array(n, lda, uplo = "lower", low = -1, high = 1, diagLow = 5, diagHigh = 15) {
+  if (uplo !== "lower" && uplo !== "upper")
+    throw new Error("uplo must be 'lower' or 'upper'.");
+  if (lda < n) throw new Error("lda must be >= n.");
+
+  const A = new Float32Array(n * lda);
+  for (let i = 0; i < n; i++) {
+    for (let j = 0; j < n; j++) {
+      if (i === j) continue;
+      const inTriangle = uplo === "lower" ? j < i : j > i;
+      if (inTriangle) A[i * lda + j] = low + Math.random() * (high - low);
+    }
+    A[i * lda + i] = diagLow + Math.random() * (diagHigh - diagLow);
+  }
+  return A;
+}

package/src/sasum/sasum.mjs

@@ -34,65 +34,69 @@ export async function sasum(device, n, x, incx) {
   const pipelineMain = await getPipeline(device, "sasum");
   const pipelineReduce = await getPipeline(device, "reduction/sum");
 
-  const xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "sasum-x", false);
-  const partialsBuffer = createStorageBuffer(2 * WGS * 4, "sasum-partials"); // 2*WGS partial sums of f32
-  const resultBuffer = createResultBuffer(4, "sasum-result"); // final f32 scalar
-  const paramsBuffer = createParamsBuffer(
-    [
-      { value: n, type: "u32" },
-      { value: incx, type: "u32" },
-    ],
-    "sasum-params",
-  );
+  let xBuffer = null;
+  let partialsBuffer = null;
+  let resultBuffer = null;
+  let paramsBuffer = null;
+  let readBuffer = null;
 
-  const bgMain = createBindGroup(pipelineMain.getBindGroupLayout(0), [
-    xBuffer,
-    partialsBuffer,
-    paramsBuffer,
-  ]);
-  const { commandEncoder: enc1, ts: ts1 } = runComputePass(
-    pipelineMain,
-    bgMain,
-    2 * WGS,
-  ); // dispatch 2*WGS workgroups
+  try {
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "sasum-x", false);
+    partialsBuffer = createStorageBuffer(2 * WGS * 4, "sasum-partials"); // 2*WGS partial sums of f32
+    resultBuffer = createResultBuffer(4, "sasum-result"); // final f32 scalar
+    paramsBuffer = createParamsBuffer(
+      [
+        { value: n, type: "u32" },
+        { value: incx, type: "u32" },
+      ],
+      "sasum-params",
+    );
+
+    const bgMain = createBindGroup(pipelineMain.getBindGroupLayout(0), [
+      xBuffer,
+      partialsBuffer,
+      paramsBuffer,
+    ]);
+    const { commandEncoder: enc1, ts: ts1 } = runComputePass(
+      pipelineMain,
+      bgMain,
+      2 * WGS,
+    ); // dispatch 2*WGS workgroups
 
-  submit(enc1);
+    submit(enc1);
 
-  const bgReduce = createBindGroup(pipelineReduce.getBindGroupLayout(0), [
-    partialsBuffer,
-    resultBuffer,
-  ]);
-  const { commandEncoder: enc2, ts: ts2 } = runComputePass(
-    pipelineReduce,
-    bgReduce,
-    1,
-  ); // dispatch 1 workgroup to reduce the partial sums to a single result
-  const readBuffer = stageReadback(enc2, resultBuffer);
+    const bgReduce = createBindGroup(pipelineReduce.getBindGroupLayout(0), [
+      partialsBuffer,
+      resultBuffer,
+    ]);
+    const { commandEncoder: enc2, ts: ts2 } = runComputePass(
+      pipelineReduce,
+      bgReduce,
+      1,
+    ); // dispatch 1 workgroup to reduce the partial sums to a single result
+    readBuffer = stageReadback(enc2, resultBuffer);
 
-  submit(enc2);
+    submit(enc2);
 
-  const [gpuTime1, gpuTime2, asumArr] = await Promise.all([
-    extractTimestamp(ts1),
-    extractTimestamp(ts2),
-    extractResult(readBuffer, Float32Array),
-  ]);
+    const resultPromise = extractResult(readBuffer, Float32Array);
+    readBuffer = null; // ownership transferred — extractResult's own finally destroys it
 
-  // asum is always a scalar readback — both paths return { asum }
-  if (xIsGpu) {
-    destroyBuffers(partialsBuffer, resultBuffer, paramsBuffer, readBuffer);
+    const [gpuTime1, gpuTime2, asumArr] = await Promise.all([
+      extractTimestamp(ts1),
+      extractTimestamp(ts2),
+      resultPromise,
+    ]);
+
+    // asum is always a scalar readback — both paths return { asum }
     if (gpuTime1 !== undefined && gpuTime2 !== undefined)
       return { asum: asumArr[0], gpuTimeMs: gpuTime1 + gpuTime2 };
     return { asum: asumArr[0] };
+  } finally {
+    if (!xIsGpu && xBuffer) destroyBuffers(xBuffer);
+    if (partialsBuffer) destroyBuffers(partialsBuffer);
+    if (resultBuffer) destroyBuffers(resultBuffer);
+    if (paramsBuffer) destroyBuffers(paramsBuffer);
+    // Only reached if submit(enc2) threw before ownership was transferred above.
+    if (readBuffer) destroyBuffers(readBuffer);
   }
-
-  destroyBuffers(
-    xBuffer,
-    partialsBuffer,
-    resultBuffer,
-    paramsBuffer,
-    readBuffer,
-  );
-  if (gpuTime1 !== undefined && gpuTime2 !== undefined)
-    return { asum: asumArr[0], gpuTimeMs: gpuTime1 + gpuTime2 };
-  return { asum: asumArr[0] };
 }

package/src/saxpy/saxpy.mjs

@@ -24,8 +24,10 @@ export async function saxpy(device, n, alpha, x, incx, y, incy) {
     !Number.isInteger(incy)
   )
     throw new Error("n, incx, and incy must be integers.");
-  if (isNaN(alpha)) throw new Error("alpha must not be NaN.");
-  if (!isFinite(alpha)) throw new Error("alpha must be finite.");
+  if (typeof alpha !== "number")
+    throw new Error("alpha must be a number.");
+  if (Number.isNaN(alpha)) throw new Error("alpha must not be NaN.");
+  if (!Number.isFinite(alpha)) throw new Error("alpha must be finite.");
   if (incx <= 0 || incy <= 0)
     throw new Error("incx and incy must be positive.");
   if (!xIsGpu && !(x instanceof Float32Array))
@@ -48,43 +50,54 @@ export async function saxpy(device, n, alpha, x, incx, y, incy) {
 
   const pipeline = await getPipeline(device, "saxpy");
 
-  const xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "saxpy-x", false);
-  const yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "saxpy-y", true);
-  const paramsBuffer = createParamsBuffer(
-    [
-      { value: n, type: "u32" },
-      { value: alpha, type: "f32" },
-      { value: incx, type: "u32" },
-      { value: incy, type: "u32" },
-    ],
-    "saxpy-params",
-  );
+  let xBuffer = null;
+  let yBuffer = null;
+  let paramsBuffer = null;
+  let readBuffer = null;
 
-  const bindGroup = createBindGroup(pipeline.getBindGroupLayout(0), [
-    xBuffer,
-    yBuffer,
-    paramsBuffer,
-  ]);
-  const { commandEncoder, ts } = runComputePass(
-    pipeline,
-    bindGroup,
-    calcWorkgroups(n),
-  );
-  const readBuffer = yIsGpu ? null : stageReadback(commandEncoder, yBuffer);
+  try {
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "saxpy-x", false);
+    yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "saxpy-y", true);
+    paramsBuffer = createParamsBuffer(
+      [
+        { value: n, type: "u32" },
+        { value: alpha, type: "f32" },
+        { value: incx, type: "u32" },
+        { value: incy, type: "u32" },
+      ],
+      "saxpy-params",
+    );
 
-  submit(commandEncoder);
+    const bindGroup = createBindGroup(pipeline.getBindGroupLayout(0), [
+      xBuffer,
+      yBuffer,
+      paramsBuffer,
+    ]);
+    const { commandEncoder, ts } = runComputePass(
+      pipeline,
+      bindGroup,
+      calcWorkgroups(n),
+    );
+    readBuffer = yIsGpu ? null : stageReadback(commandEncoder, yBuffer);
 
-  const gpuTimeMs = await extractTimestamp(ts);
+    submit(commandEncoder);
 
-  if (yIsGpu && xIsGpu) {
-    destroyBuffers(paramsBuffer);
-    if (gpuTimeMs !== undefined) return { gpuTimeMs };
-    return {};
-  }
+    const gpuTimeMs = await extractTimestamp(ts);
 
-  const result = await extractResult(readBuffer, Float32Array);
-  destroyBuffers(xBuffer, yBuffer, paramsBuffer, readBuffer);
+    if (yIsGpu && xIsGpu) {
+      if (gpuTimeMs !== undefined) return { gpuTimeMs };
+      return {};
+    }
 
-  if (gpuTimeMs !== undefined) return { y: result, gpuTimeMs };
-  return { y: result };
+    const result = await extractResult(readBuffer, Float32Array);
+    readBuffer = null; // extractResult already destroyed it
+    if (gpuTimeMs !== undefined) return { y: result, gpuTimeMs };
+    return { y: result };
+  } finally {
+    if (!xIsGpu && xBuffer) destroyBuffers(xBuffer);
+    if (!yIsGpu && yBuffer) destroyBuffers(yBuffer);
+    if (paramsBuffer) destroyBuffers(paramsBuffer);
+    // Only reached if extractTimestamp threw before extractResult ran.
+    if (readBuffer) destroyBuffers(readBuffer);
+  }
 }

package/src/scopy/scopy.mjs

@@ -46,42 +46,53 @@ export async function scopy(device, n, x, incx, y, incy) {
 
   const pipeline = await getPipeline(device, "scopy");
 
-  const xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "scopy-x", false);
-  const yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "scopy-y", true);
-  const paramsBuffer = createParamsBuffer(
-    [
-      { value: n, type: "u32" },
-      { value: incx, type: "u32" },
-      { value: incy, type: "u32" },
-    ],
-    "scopy-params",
-  );
+  let xBuffer = null;
+  let yBuffer = null;
+  let paramsBuffer = null;
+  let readBuffer = null;
 
-  const bindGroup = createBindGroup(pipeline.getBindGroupLayout(0), [
-    xBuffer,
-    yBuffer,
-    paramsBuffer,
-  ]);
-  const { commandEncoder, ts } = runComputePass(
-    pipeline,
-    bindGroup,
-    calcWorkgroups(n),
-  );
-  const readBuffer = yIsGpu ? null : stageReadback(commandEncoder, yBuffer);
+  try {
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "scopy-x", false);
+    yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "scopy-y", true);
+    paramsBuffer = createParamsBuffer(
+      [
+        { value: n, type: "u32" },
+        { value: incx, type: "u32" },
+        { value: incy, type: "u32" },
+      ],
+      "scopy-params",
+    );
 
-  submit(commandEncoder);
+    const bindGroup = createBindGroup(pipeline.getBindGroupLayout(0), [
+      xBuffer,
+      yBuffer,
+      paramsBuffer,
+    ]);
+    const { commandEncoder, ts } = runComputePass(
+      pipeline,
+      bindGroup,
+      calcWorkgroups(n),
+    );
+    readBuffer = yIsGpu ? null : stageReadback(commandEncoder, yBuffer);
 
-  const gpuTimeMs = await extractTimestamp(ts);
+    submit(commandEncoder);
 
-  if (yIsGpu && xIsGpu) {
-    destroyBuffers(paramsBuffer);
-    if (gpuTimeMs !== undefined) return { gpuTimeMs };
-    return {};
-  }
+    const gpuTimeMs = await extractTimestamp(ts);
 
-  const result = await extractResult(readBuffer, Float32Array);
-  destroyBuffers(xBuffer, yBuffer, paramsBuffer, readBuffer);
+    if (yIsGpu && xIsGpu) {
+      if (gpuTimeMs !== undefined) return { gpuTimeMs };
+      return {};
+    }
 
-  if (gpuTimeMs !== undefined) return { y: result, gpuTimeMs };
-  return { y: result };
+    const result = await extractResult(readBuffer, Float32Array);
+    readBuffer = null; // extractResult already destroyed it
+    if (gpuTimeMs !== undefined) return { y: result, gpuTimeMs };
+    return { y: result };
+  } finally {
+    if (!xIsGpu && xBuffer) destroyBuffers(xBuffer);
+    if (!yIsGpu && yBuffer) destroyBuffers(yBuffer);
+    if (paramsBuffer) destroyBuffers(paramsBuffer);
+    // Only reached if extractTimestamp threw before extractResult ran.
+    if (readBuffer) destroyBuffers(readBuffer);
+  }
 }

package/src/sdot/sdot.mjs

@@ -50,69 +50,74 @@ export async function sdot(device, n, x, incx, y, incy) {
   const pipelineMain = await getPipeline(device, "sdot");
   const pipelineReduce = await getPipeline(device, "reduction/sum");
 
-  const xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "sdot-x", false);
-  const yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "sdot-y", false);
-  const partialsBuffer = createStorageBuffer(2 * WGS * 4, "sdot-partials"); //to hold 2*WGS partial sums of f32
-  const resultBuffer = createResultBuffer(4, "sdot-result"); //to hold the final float32 dot product
-  const paramsBuffer = createParamsBuffer(
-    [
-      { value: n, type: "u32" },
-      { value: incx, type: "u32" },
-      { value: incy, type: "u32" },
-    ],
-    "sdot-params",
-  );
+  let xBuffer = null;
+  let yBuffer = null;
+  let partialsBuffer = null;
+  let resultBuffer = null;
+  let paramsBuffer = null;
+  let readBuffer = null;
 
-  const bgMain = createBindGroup(pipelineMain.getBindGroupLayout(0), [
-    xBuffer,
-    yBuffer,
-    partialsBuffer,
-    paramsBuffer,
-  ]);
-  const { commandEncoder: enc1, ts: ts1 } = runComputePass(
-    pipelineMain,
-    bgMain,
-    2 * WGS,
-  ); //dispatch 2*WGS workgroups
+  try {
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "sdot-x", false);
+    yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "sdot-y", false);
+    partialsBuffer = createStorageBuffer(2 * WGS * 4, "sdot-partials"); //to hold 2*WGS partial sums of f32
+    resultBuffer = createResultBuffer(4, "sdot-result"); //to hold the final float32 dot product
+    paramsBuffer = createParamsBuffer(
+      [
+        { value: n, type: "u32" },
+        { value: incx, type: "u32" },
+        { value: incy, type: "u32" },
+      ],
+      "sdot-params",
+    );
+
+    const bgMain = createBindGroup(pipelineMain.getBindGroupLayout(0), [
+      xBuffer,
+      yBuffer,
+      partialsBuffer,
+      paramsBuffer,
+    ]);
+    const { commandEncoder: enc1, ts: ts1 } = runComputePass(
+      pipelineMain,
+      bgMain,
+      2 * WGS,
+    ); //dispatch 2*WGS workgroups
 
-  submit(enc1);
+    submit(enc1);
 
-  const bgReduce = createBindGroup(pipelineReduce.getBindGroupLayout(0), [
-    partialsBuffer,
-    resultBuffer,
-  ]);
-  const { commandEncoder: enc2, ts: ts2 } = runComputePass(
-    pipelineReduce,
-    bgReduce,
-    1,
-  ); // dispatch 1 workgroup to reduce the partial sums to a single result
-  const readBuffer = stageReadback(enc2, resultBuffer);
+    const bgReduce = createBindGroup(pipelineReduce.getBindGroupLayout(0), [
+      partialsBuffer,
+      resultBuffer,
+    ]);
+    const { commandEncoder: enc2, ts: ts2 } = runComputePass(
+      pipelineReduce,
+      bgReduce,
+      1,
+    ); // dispatch 1 workgroup to reduce the partial sums to a single result
+    readBuffer = stageReadback(enc2, resultBuffer);
 
-  submit(enc2);
+    submit(enc2);
 
-  const [gpuTime1, gpuTime2, dotArr] = await Promise.all([
-    extractTimestamp(ts1),
-    extractTimestamp(ts2),
-    extractResult(readBuffer, Float32Array),
-  ]);
+    const resultPromise = extractResult(readBuffer, Float32Array);
+    readBuffer = null; // ownership transferred — extractResult's own finally destroys it
 
-  // dot is always a scalar readback — both paths return { dot }
-  if (xIsGpu && yIsGpu) {
-    destroyBuffers(partialsBuffer, resultBuffer, paramsBuffer, readBuffer);
+    const [gpuTime1, gpuTime2, dotArr] = await Promise.all([
+      extractTimestamp(ts1),
+      extractTimestamp(ts2),
+      resultPromise,
+    ]);
+
+    // dot is always a scalar readback — both paths return { dot }
     if (gpuTime1 !== undefined && gpuTime2 !== undefined)
       return { dot: dotArr[0], gpuTimeMs: gpuTime1 + gpuTime2 };
     return { dot: dotArr[0] };
+  } finally {
+    if (!xIsGpu && xBuffer) destroyBuffers(xBuffer);
+    if (!yIsGpu && yBuffer) destroyBuffers(yBuffer);
+    if (partialsBuffer) destroyBuffers(partialsBuffer);
+    if (resultBuffer) destroyBuffers(resultBuffer);
+    if (paramsBuffer) destroyBuffers(paramsBuffer);
+    // Only reached if submit(enc2) threw before ownership was transferred above.
+    if (readBuffer) destroyBuffers(readBuffer);
   }
-
-  destroyBuffers(
-    xBuffer,
-    yBuffer,
-    partialsBuffer,
-    resultBuffer,
-    paramsBuffer,
-    readBuffer,
-  );
-  if (gpuTime1 !== undefined && gpuTime2 !== undefined)
-    return { dot: dotArr[0], gpuTimeMs: gpuTime1 + gpuTime2 };
-  return { dot: dotArr[0] };
 }