wgblas 1.2.1 → 2.1.0

package/src/strmm/strmm.d.mts

@@ -0,0 +1,100 @@
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+
+/**
+ * Performs the triangular matrix-matrix operation
+ * B := alpha * op(A) * B (`side='left'`) or
+ * B := alpha * B * op(A) (`side='right'`) — `A` is triangular, only its
+ * `uplo` triangle stored; `B` is a general m×n matrix, overwritten in place.
+ *
+ * - `side='left'`:  `A` is m×m — `A` premultiplies `B`
+ * - `side='right'`: `A` is n×n — `A` postmultiplies `B`
+ *
+ * No dedicated fused kernel — a `triangularize` pass materializes a dense
+ * copy of `op(A)` (zero-filling the unstored triangle, substituting the
+ * implicit diagonal when `diag='unit'`), then a plain `sgemm` pass
+ * (`sgemm_small.wgsl`/`sgemm_large.wgsl`, unmodified) does the actual
+ * multiply, both on one command encoder.
+ *
+ * {@includeCode ../../examples/strmm/strmm.js}
+ *
+ * **Browser (standalone HTML):**
+ * {@includeCode ../../examples/strmm/web/strmm.html}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param side   - `'left'` for op(A)*B, `'right'` for B*op(A)
+ * @param uplo   - `'lower'` if only `A`'s lower triangle is stored, `'upper'` for upper
+ * @param transA - `'no-transpose'` for A, `'transpose'` for A^T
+ * @param diag   - `'unit'` to treat A's diagonal as all-ones (A's diagonal is not read), `'non-unit'` to read it
+ * @param m      - rows of B
+ * @param n      - columns of B
+ * @param alpha  - scalar multiplier for the matrix product
+ * @param A      - Float32Array, triangular, row-major or column-major (see `layout`)
+ * @param lda    - leading dimension of A as stored
+ * @param B      - Float32Array input/output matrix, overwritten with the result, row-major or column-major
+ * @param ldb    - leading dimension of B as stored
+ * @param layout - storage layout shared by A/B when they're Float32Array
+ *   (default: `'row-major'`) — column-major A is a genuine transpose (A isn't
+ *   symmetric like ssymm's), so both `transA` and `uplo` are adjusted
+ *   internally to compensate
+ * @returns updated B as a Float32Array
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/strmm/strmm.mjs#L25">Source code: strmm.mjs (L25)</a>
+ * @category BLAS Level 3
+ */
+export declare function strmm(
+  device: GPUDevice,
+  side: 'left' | 'right',
+  uplo: 'lower' | 'upper',
+  transA: 'no-transpose' | 'transpose',
+  diag: 'unit' | 'non-unit',
+  m: number,
+  n: number,
+  alpha: number,
+  A: Float32Array,
+  lda: number,
+  B: Float32Array,
+  ldb: number,
+  layout?: 'row-major' | 'column-major',
+): Promise<{ B: Float32Array; gpuTimeMs?: number }>;
+
+/**
+ * Performs the triangular matrix-matrix operation
+ * B := alpha * op(A) * B (`side='left'`) or
+ * B := alpha * B * op(A) (`side='right'`)
+ *
+ * A and B are both kept GPU-resident; B is mutated in place. Each matrix's
+ * own `layout` (set at `GpuMatrix.from` time) determines the operation —
+ * there is no separate `layout` argument here. A and B must both be
+ * GpuMatrix or both be Float32Array — mixing is not supported.
+ *
+ * {@includeCode ../../examples/strmm/gpu.strmm.js}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param side   - `'left'` for op(A)*B, `'right'` for B*op(A)
+ * @param uplo   - `'lower'` if only `A`'s lower triangle is stored, `'upper'` for upper
+ * @param transA - `'no-transpose'` for A, `'transpose'` for A^T
+ * @param diag   - `'unit'` to treat A's diagonal as all-ones (A's diagonal is not read), `'non-unit'` to read it
+ * @param m      - rows of B
+ * @param n      - columns of B
+ * @param alpha  - scalar multiplier for the matrix product
+ * @param A      - GpuMatrix, triangular
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @param B      - GpuMatrix (mutated in place)
+ * @param ldb    - leading dimension of B (must equal B.lda)
+ * @returns no B — it stays GPU-resident; call `B.read()` yourself for a CPU readback (see the example)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/strmm/strmm.mjs#L25">Source code: strmm.mjs (L25)</a>
+ * @category BLAS Level 3
+ */
+export declare function strmm(
+  device: GPUDevice,
+  side: 'left' | 'right',
+  uplo: 'lower' | 'upper',
+  transA: 'no-transpose' | 'transpose',
+  diag: 'unit' | 'non-unit',
+  m: number,
+  n: number,
+  alpha: number,
+  A: GpuMatrix,
+  lda: number,
+  B: GpuMatrix,
+  ldb: number,
+): Promise<{ gpuTimeMs?: number }>;

package/src/strmm/strmm.mjs

@@ -0,0 +1,226 @@
+import {
+  uploadBuffer,
+  createParamsBuffer,
+  createStorageBuffer,
+  stageReadback,
+  destroyBuffers,
+  vec4ViewBinding,
+} from "../util/buffer.mjs";
+import { createBindGroup } from "../util/bindgroup.mjs";
+import { beginTimedEncoder, encodePass, submit } from "../util/compute.mjs";
+import { extractResult } from "../util/result.mjs";
+import { resolveTimestamp, extractTimestamp } from "../util/benchmark.mjs";
+import { getPipeline } from "../util/pipeline.mjs";
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+import { requireWorkgroupCount } from "../util/workgroup.mjs";
+import { BM_SMALL, BN_SMALL, BM_LARGE, BN_LARGE, LARGE_TILE_WORKGROUP_THRESHOLD } from "../util/constants.mjs";
+import { TILE_WG_2D } from "../util/constants.mjs";
+import { requireSameDevice } from "../util/device.mjs";
+
+
+// strmm: B := alpha*op(A)*B (side='left') or alpha*B*op(A) (side='right'), A
+// triangular. Triangularize then sgemm, one command encoder. B is both
+// input and output, so gemm writes to a fresh buffer (no aliasing race),
+// copied back into B (GpuMatrix) or read back directly (Float32Array).
+export async function strmm(
+  device, side, uplo, transA, diag, m, n, alpha, A, lda, B, ldb, layout = "row-major",
+) {
+  const AIsGpu = A instanceof GpuMatrix;
+  const BIsGpu = B instanceof GpuMatrix;
+  const isUnit = diag === "unit";
+
+  if (!(device instanceof GPUDevice))
+    throw new Error("device must be a GPUDevice.");
+  requireSameDevice(device, "strmm", { A, B });
+  if (side !== "left" && side !== "right")
+    throw new Error("side must be 'left' or 'right'.");
+  if (uplo !== "lower" && uplo !== "upper")
+    throw new Error("uplo must be 'lower' or 'upper'.");
+  if (transA !== "no-transpose" && transA !== "transpose")
+    throw new Error("transA must be 'no-transpose' or 'transpose'.");
+  if (!isUnit && diag !== "non-unit")
+    throw new Error("diag must be 'unit' or 'non-unit'.");
+  if (layout !== "row-major" && layout !== "column-major")
+    throw new Error("layout must be 'row-major' or 'column-major'.");
+  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 (!Number.isInteger(m) || !Number.isInteger(n) || !Number.isInteger(lda) || !Number.isInteger(ldb))
+    throw new Error("m, n, lda, and ldb must be integers.");
+  if (!AIsGpu && !(A instanceof Float32Array))
+    throw new Error("A must be a Float32Array or GpuMatrix.");
+  if (!BIsGpu && !(B instanceof Float32Array))
+    throw new Error("B must be a Float32Array or GpuMatrix.");
+  if (AIsGpu !== BIsGpu)
+    throw new Error("A and B must both be GpuMatrix or both be Float32Array.");
+  if (m < 0 || n < 0) throw new Error("m and n must be non-negative.");
+  if (m === 0 || n === 0) return BIsGpu ? {} : { B };
+
+  const effLayoutA = AIsGpu ? A.layout : layout;
+  const effLayoutB = BIsGpu ? B.layout : layout;
+
+  // A: triangular, order = m (side='left') or n (side='right').
+  const aOrder = side === "left" ? m : n;
+  if (lda < aOrder) throw new Error("lda must be >= " + (side === "left" ? "m" : "n") + ".");
+  if (AIsGpu) {
+    if (lda !== A.lda) throw new Error("lda must match A.lda when A is a GpuMatrix.");
+    if (A.rows < aOrder || A.cols < aOrder) throw new Error("A is too small for the given m/n and side.");
+  } else if (A.length < (aOrder - 1) * lda + aOrder) {
+    throw new Error("A does not have enough elements for the given dimensions and lda.");
+  }
+
+  // B: always m x n, overwritten in place with the same ldb.
+  const bOuter = effLayoutB === "column-major" ? n : m;
+  const bInner = effLayoutB === "column-major" ? m : n;
+  if (ldb < bInner)
+    throw new Error(`ldb must be >= ${effLayoutB === "column-major" ? "rows" : "cols"} of B as stored.`);
+  if (BIsGpu) {
+    if (ldb !== B.lda) throw new Error("ldb must match B.lda when B is a GpuMatrix.");
+    if (B.rows < m || B.cols < n) throw new Error("B is too small for the given m and n.");
+  } else if (B.length < (bOuter - 1) * ldb + bInner) {
+    throw new Error("B does not have enough elements for the given dimensions and ldb.");
+  }
+
+  // A isn't symmetric: column-major = genuine transpose, so flip transA;
+  // transposing also swaps which triangle looks stored, so flip uplo too.
+  const uploEffA = effLayoutA === "column-major" ? (uplo === "lower" ? "upper" : "lower") : uplo;
+  const transEffA = effLayoutA === "column-major" ? (transA === "no-transpose" ? "transpose" : "no-transpose") : transA;
+
+  const transB = effLayoutB === "column-major" ? "transpose" : "no-transpose";
+  const transDense = "no-transpose"; // Adense already embodies op(A)
+
+  // X*Y (X=Adense,Y=B for side='left', swapped for 'right'). Column-major
+  // output: compute (B_out)^T instead — sgemm's own trick, same as ssymm's.
+  let mg = m, ng = n;
+  const kg = aOrder;
+  let transX = side === "left" ? transDense : transB;
+  let transY = side === "left" ? transB : transDense;
+  const flip = (t) => (t === "no-transpose" ? "transpose" : "no-transpose");
+  let swapXY = side === "right";
+  if (effLayoutB === "column-major") {
+    [transX, transY] = [flip(transY), flip(transX)];
+    swapXY = !swapXY;
+    [mg, ng] = [ng, mg];
+  }
+
+  const ldDense = aOrder; // Adense is tightly packed, row-major
+  const largeWgX = Math.ceil(ng / BN_LARGE);
+  const largeWgY = Math.ceil(mg / BM_LARGE);
+  const useLargeTile = largeWgX * largeWgY >= LARGE_TILE_WORKGROUP_THRESHOLD;
+  const gemmPipeline = await getPipeline(device, useLargeTile ? "sgemm_large" : "sgemm_small");
+  const triPipeline = await getPipeline(device, "triangularize");
+  const gemmWgCount = useLargeTile
+    ? {
+      x: requireWorkgroupCount(device, largeWgX, "strmm", "x"),
+      y: requireWorkgroupCount(device, largeWgY, "strmm", "y"),
+    }
+    : {
+      x: requireWorkgroupCount(device, Math.ceil(ng / BN_SMALL), "strmm", "x"),
+      y: requireWorkgroupCount(device, Math.ceil(mg / BM_SMALL), "strmm", "y"),
+    };
+
+  // Null-init here and allocate inside the try below, so a throw partway
+  // through the sequence still reaches finally with every handle visible
+  // (strsv.mjs is the reference for this pattern).
+  let ABuffer = null, BBuffer = null;
+  let AdenseBuffer = null, outBuffer = null;
+  let triParams = null, gemmParams = null;
+  let outBufferAdopted = false; // true once B._buf is repointed at outBuffer
+
+  try {
+    ABuffer = AIsGpu ? A._buf : uploadBuffer(device, A, "strmm-A", false);
+    // readback=true (COPY_SRC): BBuffer is also the source that seeds outBuffer.
+    BBuffer = BIsGpu ? B._buf : uploadBuffer(device, B, "strmm-B", true);
+    AdenseBuffer = createStorageBuffer(device, aOrder * ldDense * 4, "strmm-Adense");
+    // COPY_DST: seeded from B's own content before gemm runs, so stride-padding
+    // gaps (never written by gemm's tight m x n loop) keep B's original bytes
+    // instead of reading back as zero. COPY_SRC: read back / adopted by B after.
+    outBuffer = createStorageBuffer(device,
+      bOuter * ldb * 4, "strmm-out", GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST,
+    );
+
+    triParams = createParamsBuffer(device,
+      [
+        { value: aOrder, type: "u32" },
+        { value: lda, type: "u32" },
+        { value: ldDense, type: "u32" },
+        { value: uploEffA === "upper" ? 1 : 0, type: "u32" },
+        { value: transEffA === "transpose" ? 1 : 0, type: "u32" },
+        { value: isUnit ? 1 : 0, type: "u32" },
+      ],
+      "strmm-tri-params",
+    );
+    const triBindGroup = createBindGroup(device, triPipeline.getBindGroupLayout(0), [ABuffer, AdenseBuffer, triParams]);
+
+    // X/Y buffers and their own ld, matching swapXY above.
+    const XBuffer = swapXY ? BBuffer : AdenseBuffer;
+    const ldX = swapXY ? ldb : ldDense;
+    const YBuffer = swapXY ? AdenseBuffer : BBuffer;
+    const ldY = swapXY ? ldDense : ldb;
+
+    gemmParams = createParamsBuffer(device,
+      [
+        { value: mg,  type: "u32" },
+        { value: ng,  type: "u32" },
+        { value: kg,  type: "u32" },
+        { value: alpha, type: "f32" },
+        { value: 0.0,   type: "f32" }, // beta — strmm has no C accumulation term
+        { value: ldX, type: "u32" },
+        { value: ldY, type: "u32" },
+        { value: ldb, type: "u32" },
+        { value: transX === "transpose" ? 1 : 0, type: "u32" },
+        { value: transY === "transpose" ? 1 : 0, type: "u32" },
+      ],
+      "strmm-gemm-params",
+    );
+    const gemmBindGroup = createBindGroup(device, gemmPipeline.getBindGroupLayout(0), [
+      XBuffer,
+      vec4ViewBinding(device, XBuffer),
+      YBuffer,
+      vec4ViewBinding(device, YBuffer),
+      outBuffer,
+      gemmParams,
+    ]);
+
+    const { commandEncoder, querySet } = beginTimedEncoder(device);
+    // Seed outBuffer with B's own bytes first, so gemm's tight m x n write
+    // leaves stride-padding gaps holding B's original content, not zero.
+    // BBuffer may be larger than outBuffer (e.g. a validation-test baseline
+    // over-provisioned for a bigger ldb it might later be substituted with).
+    commandEncoder.copyBufferToBuffer(BBuffer, 0, outBuffer, 0, Math.min(BBuffer.size, outBuffer.size));
+    const triDesc = querySet ? { timestampWrites: { querySet, beginningOfPassWriteIndex: 0 } } : undefined;
+    const gemmDesc = querySet ? { timestampWrites: { querySet, endOfPassWriteIndex: 1 } } : undefined;
+    encodePass(commandEncoder, triPipeline, triBindGroup, { x: Math.ceil(aOrder / TILE_WG_2D), y: Math.ceil(aOrder / TILE_WG_2D) }, triDesc);
+    encodePass(commandEncoder, gemmPipeline, gemmBindGroup, gemmWgCount, gemmDesc);
+
+    const ts = resolveTimestamp(device, commandEncoder, querySet);
+    const readBuffer = BIsGpu ? null : stageReadback(device, commandEncoder, outBuffer);
+
+    submit(device, commandEncoder);
+
+    const gpuTimeMs = await extractTimestamp(ts);
+
+    if (BIsGpu) {
+      // Adopt outBuffer as B's own backing buffer instead of copying into the
+      // old one (B._buf has no COPY_DST usage) — cheaper and avoids needing
+      // an extra buffer-usage flag on every GpuMatrix for this one routine.
+      destroyBuffers(B._buf);
+      B._buf = outBuffer;
+      outBufferAdopted = true;
+      if (gpuTimeMs !== undefined) return { gpuTimeMs };
+      return {};
+    }
+
+    const result = await extractResult(readBuffer, Float32Array);
+    if (gpuTimeMs !== undefined) return { B: result, gpuTimeMs };
+    return { B: result };
+  } finally {
+    if (!AIsGpu && ABuffer) destroyBuffers(ABuffer);
+    if (!BIsGpu && BBuffer) destroyBuffers(BBuffer);
+    if (AdenseBuffer) destroyBuffers(AdenseBuffer);
+    if (outBuffer && !outBufferAdopted) destroyBuffers(outBuffer);
+    if (triParams) destroyBuffers(triParams);
+    if (gemmParams) destroyBuffers(gemmParams);
+  }
+}

package/src/strmv/strmv.d.mts

@@ -44,41 +44,6 @@ export declare function strmv(
   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
  *
@@ -86,7 +51,7 @@ export declare function strmv(
  * `layout` (set at `GpuMatrix.from` time) determines the operation — there is
  * no separate `layout` argument here.
  *
- * {@includeCode ../../examples/strmv/gpuvec.strmv.js}
+ * {@includeCode ../../examples/strmv/gpu.strmv.js}
  *
  * @param device - GPUDevice from `init()`
  * @param uplo   - `'lower'` to use the lower triangle, `'upper'` to use the upper triangle

package/src/strmv/strmv.mjs

@@ -11,6 +11,7 @@ import { extractTimestamp } from "../util/benchmark.mjs";
 import { getPipeline } from "../util/pipeline.mjs";
 import { GpuVector } from "../classes/GpuVector.mjs";
 import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+import { requireSameDevice } from "../util/device.mjs";
 
 export async function strmv(device, uplo, trans, diag, n, A, lda, x, incx, y, incy, layout = "row-major") {
   const xIsGpu = x instanceof GpuVector;
@@ -20,6 +21,7 @@ export async function strmv(device, uplo, trans, diag, n, A, lda, x, incx, y, in
 
   if (!(device instanceof GPUDevice))
     throw new Error("device must be a GPUDevice.");
+  requireSameDevice(device, "strmv", { A, x, y });
   if (uplo !== "lower" && uplo !== "upper")
     throw new Error("uplo must be 'lower' or 'upper'.");
   if (trans !== "no-transpose" && trans !== "transpose")
@@ -54,6 +56,8 @@ export async function strmv(device, uplo, trans, diag, n, A, lda, x, incx, y, in
     );
   if (xIsGpu && !AIsGpu)
     throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");
+  if (AIsGpu && !xIsGpu)
+    throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");
   if (AIsGpu && yIsGpu && A._buf === y._buf)
     throw new Error("A and y must not reference the same GPU buffer.");
   if (AIsGpu && lda !== A.lda)
@@ -90,10 +94,10 @@ export async function strmv(device, uplo, trans, diag, n, A, lda, x, incx, y, in
   let paramsBuffer = null;
 
   try {
-    ABuffer = AIsGpu ? A._buf : uploadBuffer(A, "strmv-A", false);
-    xBuffer = xIsGpu ? x._buf : uploadBuffer(x, "strmv-x", false);
-    yBuffer = yIsGpu ? y._buf : uploadBuffer(y, "strmv-y", true);
-    paramsBuffer = createParamsBuffer(
+    ABuffer = AIsGpu ? A._buf : uploadBuffer(device, A, "strmv-A", false);
+    xBuffer = xIsGpu ? x._buf : uploadBuffer(device, x, "strmv-x", false);
+    yBuffer = yIsGpu ? y._buf : uploadBuffer(device, y, "strmv-y", true);
+    paramsBuffer = createParamsBuffer(device,
       [
         { value: n,             type: "u32" },
         { value: incx,          type: "u32" },
@@ -106,7 +110,7 @@ export async function strmv(device, uplo, trans, diag, n, A, lda, x, incx, y, in
       "strmv-params",
     );
 
-    const bindGroup = createBindGroup(pipeline.getBindGroupLayout(0), [
+    const bindGroup = createBindGroup(device, pipeline.getBindGroupLayout(0), [
       ABuffer,
       xBuffer,
       yBuffer,
@@ -114,10 +118,10 @@ export async function strmv(device, uplo, trans, diag, n, A, lda, x, incx, y, in
     ]);
 
     const wgCount = Math.min(n, device.limits.maxComputeWorkgroupsPerDimension);
-    const { commandEncoder, ts } = runComputePass(pipeline, bindGroup, wgCount);
-    const readBuffer = yIsGpu ? null : stageReadback(commandEncoder, yBuffer);
+    const { commandEncoder, ts } = runComputePass(device, pipeline, bindGroup, wgCount);
+    const readBuffer = yIsGpu ? null : stageReadback(device, commandEncoder, yBuffer);
 
-    submit(commandEncoder);
+    submit(device, commandEncoder);
 
     const gpuTimeMs = await extractTimestamp(ts);
 

package/src/strsm/strsm.d.mts

@@ -0,0 +1,99 @@
+import { GpuMatrix } from "../classes/GpuMatrix.mjs";
+
+/**
+ * Solves the triangular matrix equation
+ * op(A) * X = alpha * B (`side='left'`) or
+ * X * op(A) = alpha * B (`side='right'`), overwriting `B` with the
+ * solution `X` — `A` is triangular, only its `uplo` triangle stored; `B` is
+ * a general m×n matrix.
+ *
+ * - `side='left'`:  `A` is m×m — solves `op(A)*X = alpha*B`
+ * - `side='right'`: `A` is n×n — solves `X*op(A) = alpha*B`
+ *
+ * Blocked substitution (strsv's own technique, generalized to a matrix
+ * RHS): strsv_invert_block + sgemm, both unmodified. A near-zero diagonal
+ * entry amplifies error, same as any triangular solve.
+ *
+ * {@includeCode ../../examples/strsm/strsm.js}
+ *
+ * **Browser (standalone HTML):**
+ * {@includeCode ../../examples/strsm/web/strsm.html}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param side   - `'left'` to solve op(A)*X=alpha*B, `'right'` to solve X*op(A)=alpha*B
+ * @param uplo   - `'lower'` if only `A`'s lower triangle is stored, `'upper'` for upper
+ * @param transA - `'no-transpose'` for A, `'transpose'` for A^T
+ * @param diag   - `'unit'` to treat A's diagonal as all-ones (A's diagonal is not read), `'non-unit'` to read it
+ * @param m      - rows of B
+ * @param n      - columns of B
+ * @param alpha  - scalar multiplier for B
+ * @param A      - Float32Array, triangular, row-major or column-major (see `layout`)
+ * @param lda    - leading dimension of A as stored
+ * @param B      - Float32Array input/output matrix, overwritten with the solution, row-major or column-major
+ * @param ldb    - leading dimension of B as stored
+ * @param layout - storage layout shared by A/B when they're Float32Array
+ *   (default: `'row-major'`) — column-major A is a genuine transpose (A isn't
+ *   symmetric like ssymm's), so both `transA` and `uplo` are adjusted
+ *   internally to compensate
+ * @returns the solution X, written into B, as a Float32Array
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/strsm/strsm.mjs#L25">Source code: strsm.mjs (L25)</a>
+ * @category BLAS Level 3
+ */
+export declare function strsm(
+  device: GPUDevice,
+  side: 'left' | 'right',
+  uplo: 'lower' | 'upper',
+  transA: 'no-transpose' | 'transpose',
+  diag: 'unit' | 'non-unit',
+  m: number,
+  n: number,
+  alpha: number,
+  A: Float32Array,
+  lda: number,
+  B: Float32Array,
+  ldb: number,
+  layout?: 'row-major' | 'column-major',
+): Promise<{ B: Float32Array; gpuTimeMs?: number }>;
+
+/**
+ * Solves the triangular matrix equation
+ * op(A) * X = alpha * B (`side='left'`) or
+ * X * op(A) = alpha * B (`side='right'`), overwriting `B` in place with `X`.
+ *
+ * A and B are both kept GPU-resident. Each matrix's own `layout` (set at
+ * `GpuMatrix.from` time) determines the operation — there is no separate
+ * `layout` argument here. A and B must both be GpuMatrix or both be
+ * Float32Array — mixing is not supported.
+ *
+ * {@includeCode ../../examples/strsm/gpu.strsm.js}
+ *
+ * @param device - GPUDevice from `init()`
+ * @param side   - `'left'` to solve op(A)*X=alpha*B, `'right'` to solve X*op(A)=alpha*B
+ * @param uplo   - `'lower'` if only `A`'s lower triangle is stored, `'upper'` for upper
+ * @param transA - `'no-transpose'` for A, `'transpose'` for A^T
+ * @param diag   - `'unit'` to treat A's diagonal as all-ones (A's diagonal is not read), `'non-unit'` to read it
+ * @param m      - rows of B
+ * @param n      - columns of B
+ * @param alpha  - scalar multiplier for B
+ * @param A      - GpuMatrix, triangular
+ * @param lda    - leading dimension of A (must equal A.lda)
+ * @param B      - GpuMatrix (overwritten in place with the solution)
+ * @param ldb    - leading dimension of B (must equal B.lda)
+ * @returns no B — it stays GPU-resident; call `B.read()` yourself for a CPU readback (see the example)
+ * @see <a href="https://github.com/manit2004/wgblas/blob/main/src/strsm/strsm.mjs#L25">Source code: strsm.mjs (L25)</a>
+ * @category BLAS Level 3
+ */
+export declare function strsm(
+  device: GPUDevice,
+  side: 'left' | 'right',
+  uplo: 'lower' | 'upper',
+  transA: 'no-transpose' | 'transpose',
+  diag: 'unit' | 'non-unit',
+  m: number,
+  n: number,
+  alpha: number,
+  A: GpuMatrix,
+  lda: number,
+  B: GpuMatrix,
+  ldb: number,
+): Promise<{ gpuTimeMs?: number }>;