wgblas 0.1.2 → 1.1.0

package/src/ssyr/ssyr.d.mts

@@ -0,0 +1,100 @@
+import { GpuVector } from "../classes/GpuVector.mjs";
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+
+/**
+ * Performs the symmetric rank-1 update A = alpha * x * x^T + A
+ *
+ * A is an n×n symmetric matrix stored in row-major order, updated in place.
+ * Only the triangle specified by `uplo` is referenced and updated; the other
+ * triangle is left untouched (implied by symmetry).
+ *
+ * {@includeCode ../../examples/ssyr/ssyr.js}
+ *
+ * **Browser (standalone HTML):**
+ * {@includeCode ../../examples/ssyr/web/ssyr.html}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param n      - order of the matrix A (number of rows and columns)
+ * @param alpha  - scalar multiplier for x*x^T
+ * @param x      - Float32Array input vector, length at least (n-1)*incx+1
+ * @param incx   - stride for x (must be a positive integer)
+ * @param A      - Float32Array, row-major or column-major (see `layout`), at least (n-1)*lda+n elements
+ * @param lda    - leading dimension of A (>= n either way — A is square)
+ * @param layout - storage layout of `A` (default: `'row-major'`); for a symmetric
+ *   matrix, column-major storage just means the *other* triangle is the one
+ *   physically referenced for a given `uplo`
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/ssyr/ssyr.mjs#L15">Source code: ssyr.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function ssyr(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  n: number,
+  alpha: number,
+  x: Float32Array,
+  incx: number,
+  A: Float32Array,
+  lda: number,
+  layout?: 'row-major' | 'column-major',
+): Promise<{ A: Float32Array; gpuTimeMs?: number }>;
+
+/**
+ * Performs the symmetric rank-1 update A = alpha * x * x^T + A
+ *
+ * A is kept GPU-resident; x is a CPU Float32Array. `A`'s own `layout` (set at
+ * `GpuMatrix.from` time) determines the operation — there is no separate
+ * `layout` argument here.
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param n      - order of the matrix A
+ * @param alpha  - scalar multiplier for x*x^T
+ * @param x      - Float32Array input vector
+ * @param incx   - stride for x (must be a positive integer)
+ * @param A      - GpuMatrix, GPU-resident
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/ssyr/ssyr.mjs#L15">Source code: ssyr.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function ssyr(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  n: number,
+  alpha: number,
+  x: Float32Array,
+  incx: number,
+  A: GpuMatrix,
+  lda: number,
+): Promise<{ gpuTimeMs?: number }>;
+
+/**
+ * Performs the symmetric rank-1 update A = alpha * x * x^T + A
+ *
+ * x and A are both kept resident on the GPU. `A`'s own `layout` (set at
+ * `GpuMatrix.from` time) determines the operation — there is no separate
+ * `layout` argument here.
+ *
+ * {@includeCode ../../examples/ssyr/gpuvec.ssyr.js}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param n      - order of the matrix A
+ * @param alpha  - scalar multiplier for x*x^T
+ * @param x      - GpuVector input vector (not mutated)
+ * @param incx   - stride for x (must be a positive integer)
+ * @param A      - GpuMatrix, mutated in place
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/ssyr/ssyr.mjs#L15">Source code: ssyr.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function ssyr(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  n: number,
+  alpha: number,
+  x: GpuVector,
+  incx: number,
+  A: GpuMatrix,
+  lda: number,
+): Promise<{ gpuTimeMs?: number }>;

package/src/ssyr/ssyr.mjs

@@ -0,0 +1,106 @@
+import {
+  uploadBuffer,
+  createParamsBuffer,
+  stageReadback,
+  destroyBuffers,
+} from "../util/buffer.mjs";
+import { createBindGroup } from "../util/bindgroup.mjs";
+import { runComputePass, submit } from "../util/compute.mjs";
+import { extractResult } from "../util/result.mjs";
+import { extractTimestamp } from "../util/benchmark.mjs";
+import { getPipeline } from "../util/pipeline.mjs";
+import { GpuVector } from "../classes/GpuVector.mjs";
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+
+export async function ssyr(device, uplo, n, alpha, x, incx, A, lda, layout = "row-major") {
+  const xIsGpu = x instanceof GpuVector;
+  const AIsGpu = A instanceof GpuMatrix;
+
+  if (!(device instanceof GPUDevice))
+    throw new Error("device must be a GPUDevice.");
+  if (uplo !== "lower" && uplo !== "upper")
+    throw new Error("uplo must be 'lower' or 'upper'.");
+  if (layout !== "row-major" && layout !== "column-major")
+    throw new Error("layout must be 'row-major' or 'column-major'.");
+  if (!Number.isInteger(n) || !Number.isInteger(incx) || !Number.isInteger(lda))
+    throw new Error("n, incx, and lda must be integers.");
+  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) throw new Error("incx must be positive.");
+  if (lda < n) throw new Error("lda must be >= n.");
+  if (!AIsGpu && !(A instanceof Float32Array))
+    throw new Error("A must be a Float32Array or GpuMatrix.");
+  if (!xIsGpu && !(x instanceof Float32Array))
+    throw new Error("x must be a Float32Array or GpuVector.");
+  if (xIsGpu && !AIsGpu)
+    throw new Error("A must be a GpuMatrix when x is a GpuVector.");
+  if (AIsGpu && xIsGpu && A._buf === x._buf)
+    throw new Error("A and x must not reference the same GPU buffer.");
+  if (AIsGpu && lda !== A.lda)
+    throw new Error("lda must match A.lda when A is a GpuMatrix.");
+  if (AIsGpu && (A.rows < n || A.cols < n))
+    throw new Error("A is too small for the given n.");
+  if (n < 0) throw new Error("n must be non-negative.");
+  if (n === 0) return AIsGpu ? {} : { A };
+
+  if (!AIsGpu && A.length < (n - 1) * lda + n)
+    throw new Error("A does not have enough elements for the given n and lda.");
+  if (x.length < (n - 1) * incx + 1)
+    throw new Error("x does not have enough elements for the given n and incx.");
+
+  // GpuMatrix's own layout wins over the argument; A is symmetric, so column-major A reinterpreted row-major just flips which triangle is stored — flip uplo to match.
+  const effLayout = AIsGpu ? A.layout : layout;
+  const isLower = effLayout === "column-major" ? uplo === "upper" : uplo === "lower";
+
+  const pipeline = await getPipeline(device, "ssyr");
+
+  let xBuffer = null;
+  let ABuffer = null;
+  let paramsBuffer = null;
+
+  try {
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "ssyr-x", false);
+    ABuffer = AIsGpu ? A._buf : uploadBuffer(A, "ssyr-A", true);
+    paramsBuffer = createParamsBuffer(
+      [
+        { value: n,               type: "u32" },
+        { value: alpha,           type: "f32" },
+        { value: incx,            type: "u32" },
+        { value: lda,             type: "u32" },
+        { value: isLower ? 0 : 1, type: "u32" },
+      ],
+      "ssyr-params",
+    );
+
+    const bindGroup = createBindGroup(pipeline.getBindGroupLayout(0), [
+      xBuffer,
+      ABuffer,
+      paramsBuffer,
+    ]);
+
+    // One workgroup per row of A; clamped to device limit — the shader's
+    // grid-stride loop handles remaining rows when n > dispatch count.
+    const wgCount = Math.min(n, device.limits.maxComputeWorkgroupsPerDimension);
+    const { commandEncoder, ts } = runComputePass(pipeline, bindGroup, wgCount);
+    const readBuffer = AIsGpu ? null : stageReadback(commandEncoder, ABuffer);
+
+    submit(commandEncoder);
+
+    const gpuTimeMs = await extractTimestamp(ts);
+
+    if (AIsGpu) {
+      if (gpuTimeMs !== undefined) return { gpuTimeMs };
+      return {};
+    }
+
+    const result = await extractResult(readBuffer, Float32Array);
+    if (gpuTimeMs !== undefined) return { A: result, gpuTimeMs };
+    return { A: result };
+  } finally {
+    if (!xIsGpu && xBuffer) destroyBuffers(xBuffer);
+    if (!AIsGpu && ABuffer) destroyBuffers(ABuffer);
+    if (paramsBuffer) destroyBuffers(paramsBuffer);
+  }
+}

package/src/ssyr2/ssyr2.d.mts

@@ -0,0 +1,112 @@
+import { GpuVector } from "../classes/GpuVector.mjs";
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+
+/**
+ * Performs the symmetric rank-2 update A = alpha * x * y^T + alpha * y * x^T + A
+ *
+ * A is an n×n symmetric matrix stored in row-major order, updated in place.
+ * Only the triangle specified by `uplo` is referenced and updated; the other
+ * triangle is left untouched (implied by symmetry).
+ *
+ * {@includeCode ../../examples/ssyr2/ssyr2.js}
+ *
+ * **Browser (standalone HTML):**
+ * {@includeCode ../../examples/ssyr2/web/ssyr2.html}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param n      - order of the matrix A (number of rows and columns)
+ * @param alpha  - scalar multiplier for x*y^T + y*x^T
+ * @param x      - Float32Array input vector, length at least (n-1)*incx+1
+ * @param incx   - stride for x (must be a positive integer)
+ * @param y      - Float32Array input vector, length at least (n-1)*incy+1
+ * @param incy   - stride for y (must be a positive integer)
+ * @param A      - Float32Array, row-major or column-major (see `layout`), at least (n-1)*lda+n elements
+ * @param lda    - leading dimension of A (>= n either way — A is square)
+ * @param layout - storage layout of `A` (default: `'row-major'`); for a symmetric
+ *   matrix, column-major storage just means the *other* triangle is the one
+ *   physically referenced for a given `uplo`
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/ssyr2/ssyr2.mjs#L15">Source code: ssyr2.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function ssyr2(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  n: number,
+  alpha: number,
+  x: Float32Array,
+  incx: number,
+  y: Float32Array,
+  incy: number,
+  A: Float32Array,
+  lda: number,
+  layout?: 'row-major' | 'column-major',
+): Promise<{ A: Float32Array; gpuTimeMs?: number }>;
+
+/**
+ * Performs the symmetric rank-2 update A = alpha * x * y^T + alpha * y * x^T + A
+ *
+ * A is kept GPU-resident; x and y are CPU Float32Arrays. `A`'s own `layout`
+ * (set at `GpuMatrix.from` time) determines the operation — there is no
+ * separate `layout` argument here.
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param n      - order of the matrix A
+ * @param alpha  - scalar multiplier for x*y^T + y*x^T
+ * @param x      - Float32Array input vector
+ * @param incx   - stride for x (must be a positive integer)
+ * @param y      - Float32Array input vector
+ * @param incy   - stride for y (must be a positive integer)
+ * @param A      - GpuMatrix, row-major, GPU-resident, Float32-backed
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/ssyr2/ssyr2.mjs#L15">Source code: ssyr2.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function ssyr2(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  n: number,
+  alpha: number,
+  x: Float32Array,
+  incx: number,
+  y: Float32Array,
+  incy: number,
+  A: GpuMatrix,
+  lda: number,
+): Promise<{ gpuTimeMs?: number }>;
+
+/**
+ * Performs the symmetric rank-2 update A = alpha * x * y^T + alpha * y * x^T + A
+ *
+ * x, y, and A are all kept resident on the GPU. `A`'s own `layout` (set at
+ * `GpuMatrix.from` time) determines the operation — there is no separate
+ * `layout` argument here.
+ *
+ * {@includeCode ../../examples/ssyr2/gpuvec.ssyr2.js}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param n      - order of the matrix A
+ * @param alpha  - scalar multiplier for x*y^T + y*x^T
+ * @param x      - GpuVector input vector (not mutated), Float32-backed
+ * @param incx   - stride for x (must be a positive integer)
+ * @param y      - GpuVector input vector (not mutated), Float32-backed
+ * @param incy   - stride for y (must be a positive integer)
+ * @param A      - GpuMatrix, row-major, mutated in place, Float32-backed
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/ssyr2/ssyr2.mjs#L15">Source code: ssyr2.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function ssyr2(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  n: number,
+  alpha: number,
+  x: GpuVector,
+  incx: number,
+  y: GpuVector,
+  incy: number,
+  A: GpuMatrix,
+  lda: number,
+): Promise<{ gpuTimeMs?: number }>;

package/src/ssyr2/ssyr2.mjs

@@ -0,0 +1,130 @@
+import {
+  uploadBuffer,
+  createParamsBuffer,
+  stageReadback,
+  destroyBuffers,
+} from "../util/buffer.mjs";
+import { createBindGroup } from "../util/bindgroup.mjs";
+import { runComputePass, submit } from "../util/compute.mjs";
+import { extractResult } from "../util/result.mjs";
+import { extractTimestamp } from "../util/benchmark.mjs";
+import { getPipeline } from "../util/pipeline.mjs";
+import { GpuVector } from "../classes/GpuVector.mjs";
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+
+export async function ssyr2(device, uplo, n, alpha, x, incx, y, incy, A, lda, layout = "row-major") {
+  const xIsGpu = x instanceof GpuVector;
+  const yIsGpu = y instanceof GpuVector;
+  const AIsGpu = A instanceof GpuMatrix;
+
+  if (!(device instanceof GPUDevice))
+    throw new Error("device must be a GPUDevice.");
+  if (uplo !== "lower" && uplo !== "upper")
+    throw new Error("uplo must be 'lower' or 'upper'.");
+  if (layout !== "row-major" && layout !== "column-major")
+    throw new Error("layout must be 'row-major' or 'column-major'.");
+  if (
+    !Number.isInteger(n) ||
+    !Number.isInteger(incx) ||
+    !Number.isInteger(incy) ||
+    !Number.isInteger(lda)
+  )
+    throw new Error("n, incx, incy, and lda must be integers.");
+  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 (lda < n) throw new Error("lda must be >= n.");
+  if (!AIsGpu && !(A instanceof Float32Array))
+    throw new Error("A must be a Float32Array or GpuMatrix.");
+  if (!xIsGpu && !(x instanceof Float32Array))
+    throw new Error("x must be a Float32Array or GpuVector.");
+  if (!yIsGpu && !(y instanceof Float32Array))
+    throw new Error("y must be a Float32Array or GpuVector.");
+  if (xIsGpu !== yIsGpu)
+    throw new Error(
+      "x and y must be the same type (both Float32Array or both GpuVector).",
+    );
+  if (xIsGpu && !AIsGpu)
+    throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");
+  if (AIsGpu && xIsGpu && A._buf === x._buf)
+    throw new Error("A and x must not reference the same GPU buffer.");
+  if (AIsGpu && yIsGpu && A._buf === y._buf)
+    throw new Error("A and y must not reference the same GPU buffer.");
+  if (xIsGpu && x._buf === y._buf)
+    throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");
+  if (AIsGpu && lda !== A.lda)
+    throw new Error("lda must match A.lda when A is a GpuMatrix.");
+  if (AIsGpu && (A.rows < n || A.cols < n))
+    throw new Error("A is too small for the given n.");
+  if (n < 0) throw new Error("n must be non-negative.");
+  if (n === 0) return AIsGpu ? {} : { A };
+
+  if (!AIsGpu && A.length < (n - 1) * lda + n)
+    throw new Error("A does not have enough elements for the given n and lda.");
+  if (x.length < (n - 1) * incx + 1)
+    throw new Error("x does not have enough elements for the given n and incx.");
+  if (y.length < (n - 1) * incy + 1)
+    throw new Error("y does not have enough elements for the given n and incy.");
+
+  // GpuMatrix's own layout wins over the argument; A is symmetric, so column-major A reinterpreted row-major just flips which triangle is stored (no x/y swap needed — x*y^T+y*x^T is already symmetric under swapping them).
+  const effLayout = AIsGpu ? A.layout : layout;
+  const isLower = effLayout === "column-major" ? uplo === "upper" : uplo === "lower";
+
+  const pipeline = await getPipeline(device, "ssyr2");
+
+  let xBuffer = null;
+  let yBuffer = null;
+  let ABuffer = null;
+  let paramsBuffer = null;
+
+  try {
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "ssyr2-x", false);
+    yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "ssyr2-y", false);
+    ABuffer = AIsGpu ? A._buf : uploadBuffer(A, "ssyr2-A", true);
+    paramsBuffer = createParamsBuffer(
+      [
+        { value: n,               type: "u32" },
+        { value: alpha,           type: "f32" },
+        { value: incx,            type: "u32" },
+        { value: incy,            type: "u32" },
+        { value: lda,             type: "u32" },
+        { value: isLower ? 0 : 1, type: "u32" },
+      ],
+      "ssyr2-params",
+    );
+
+    const bindGroup = createBindGroup(pipeline.getBindGroupLayout(0), [
+      xBuffer,
+      yBuffer,
+      ABuffer,
+      paramsBuffer,
+    ]);
+
+    // One workgroup per row of A; clamped to device limit — the shader's
+    // grid-stride loop handles remaining rows when n > dispatch count.
+    const wgCount = Math.min(n, device.limits.maxComputeWorkgroupsPerDimension);
+    const { commandEncoder, ts } = runComputePass(pipeline, bindGroup, wgCount);
+    const readBuffer = AIsGpu ? null : stageReadback(commandEncoder, ABuffer);
+
+    submit(commandEncoder);
+
+    const gpuTimeMs = await extractTimestamp(ts);
+
+    if (AIsGpu) {
+      if (gpuTimeMs !== undefined) return { gpuTimeMs };
+      return {};
+    }
+
+    const result = await extractResult(readBuffer, Float32Array);
+    if (gpuTimeMs !== undefined) return { A: result, gpuTimeMs };
+    return { A: result };
+  } finally {
+    if (!xIsGpu && xBuffer) destroyBuffers(xBuffer);
+    if (!yIsGpu && yBuffer) destroyBuffers(yBuffer);
+    if (!AIsGpu && ABuffer) destroyBuffers(ABuffer);
+    if (paramsBuffer) destroyBuffers(paramsBuffer);
+  }
+}

package/src/strmv/strmv.d.mts

@@ -0,0 +1,117 @@
+import { GpuVector } from "../classes/GpuVector.mjs";
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+
+/**
+ * Performs the triangular matrix-vector operation y = op(A) * x
+ *
+ * A is an n×n triangular matrix stored in row-major order. Only the triangle
+ * specified by `uplo` is referenced; the other triangle is not accessed.
+ *
+ * {@includeCode ../../examples/strmv/strmv.js}
+ *
+ * **Browser (standalone HTML):**
+ * {@includeCode ../../examples/strmv/web/strmv.html}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param trans  - `'no-transpose'` for A, `'transpose'` for A^T
+ * @param diag   - `'unit'` to treat the diagonal as all-ones (A's diagonal is not read), `'non-unit'` to read it
+ * @param n      - order of the matrix A (number of rows and columns)
+ * @param A      - Float32Array, row-major or column-major (see `layout`), at least (n-1)*lda+n elements
+ * @param lda    - leading dimension of A (>= n either way — A is square)
+ * @param x      - Float32Array input vector, length at least (n-1)*incx+1
+ * @param incx   - stride for x (must be a positive integer)
+ * @param y      - Float32Array output vector, length at least (n-1)*incy+1
+ * @param incy   - stride for y (must be a positive integer)
+ * @param layout - storage layout of `A` (default: `'row-major'`); column-major
+ *   flips both the stored triangle and the effective `trans` (op(A) stays
+ *   what you asked for either way)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/strmv/strmv.mjs#L15">Source code: strmv.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function strmv(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  trans: 'no-transpose' | 'transpose',
+  diag: 'unit' | 'non-unit',
+  n: number,
+  A: Float32Array,
+  lda: number,
+  x: Float32Array,
+  incx: number,
+  y: Float32Array,
+  incy: number,
+  layout?: 'row-major' | 'column-major',
+): Promise<{ y: Float32Array; gpuTimeMs?: number }>;
+
+/**
+ * Performs the triangular matrix-vector operation y = op(A) * x
+ *
+ * A is kept GPU-resident; x and y are CPU Float32Arrays. `A`'s own `layout`
+ * (set at `GpuMatrix.from` time) determines the operation — there is no
+ * separate `layout` argument here.
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param trans  - `'no-transpose'` for A, `'transpose'` for A^T
+ * @param diag   - `'unit'` to treat the diagonal as all-ones (A's diagonal is not read), `'non-unit'` to read it
+ * @param n      - order of the matrix A
+ * @param A      - GpuMatrix, GPU-resident
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @param x      - Float32Array input vector
+ * @param incx   - stride for x (must be a positive integer)
+ * @param y      - Float32Array output vector
+ * @param incy   - stride for y (must be a positive integer)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/strmv/strmv.mjs#L15">Source code: strmv.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function strmv(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  trans: 'no-transpose' | 'transpose',
+  diag: 'unit' | 'non-unit',
+  n: number,
+  A: GpuMatrix,
+  lda: number,
+  x: Float32Array,
+  incx: number,
+  y: Float32Array,
+  incy: number,
+): Promise<{ y: Float32Array; gpuTimeMs?: number }>;
+
+/**
+ * Performs the triangular matrix-vector operation y = op(A) * x
+ *
+ * x and y are kept resident on the GPU. A must be a GpuMatrix; its own
+ * `layout` (set at `GpuMatrix.from` time) determines the operation — there is
+ * no separate `layout` argument here.
+ *
+ * {@includeCode ../../examples/strmv/gpuvec.strmv.js}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle
+ * @param trans  - `'no-transpose'` for A, `'transpose'` for A^T
+ * @param diag   - `'unit'` to treat the diagonal as all-ones (A's diagonal is not read), `'non-unit'` to read it
+ * @param n      - order of the matrix A
+ * @param A      - GpuMatrix, GPU-resident
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @param x      - GpuVector input vector (not mutated)
+ * @param incx   - stride for x (must be a positive integer)
+ * @param y      - GpuVector output vector (mutated in place)
+ * @param incy   - stride for y (must be a positive integer)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/strmv/strmv.mjs#L15">Source code: strmv.mjs (L15)</a>
+ * @category BLAS Level 2
+ */
+export declare function strmv(
+  device: GPUDevice,
+  uplo: 'lower' | 'upper',
+  trans: 'no-transpose' | 'transpose',
+  diag: 'unit' | 'non-unit',
+  n: number,
+  A: GpuMatrix,
+  lda: number,
+  x: GpuVector,
+  incx: number,
+  y: GpuVector,
+  incy: number,
+): Promise<{ gpuTimeMs?: number }>;