@mni-ml/framework 0.0.1

package/dist/gpu_backend.d.ts

@@ -0,0 +1,37 @@
+export declare function getDevice(): Promise<GPUDevice>;
+export declare function destroyDevice(): void;
+/**
+ * Try to obtain a GPUDevice with experimental subgroup matrix support.
+ * Returns null when the feature is not available on this adapter/platform.
+ * Dawn (the backend for the `webgpu` npm package) exposes Apple simdgroup_matrix
+ * (8x8 f32 on Metal/Apple7+) and Vulkan VK_KHR_cooperative_matrix behind the
+ * `allow_unsafe_apis` toggle and the `chromium-experimental-subgroup-matrix`
+ * device feature.
+ */
+export declare function getTensorCoreDevice(): Promise<GPUDevice | null>;
+export declare function destroyTensorCoreDevice(): void;
+/**
+ * Upload a Float64Array to the GPU as f32. Returns a STORAGE | COPY_SRC buffer.
+ */
+export declare function uploadBuffer(device: GPUDevice, data: Float64Array): GPUBuffer;
+/**
+ * Create a GPU buffer for compute output (read-write storage, copyable for readback).
+ */
+export declare function createOutputBuffer(device: GPUDevice, count: number): GPUBuffer;
+/**
+ * Read a GPU buffer back to a Float64Array (f32 -> f64).
+ */
+export declare function readbackBuffer(device: GPUDevice, srcBuffer: GPUBuffer, count: number): Promise<Float64Array>;
+/**
+ * Create a uniform buffer from an ArrayBuffer of packed metadata.
+ * Only suitable for simple structs without array members (WGSL alignment rules).
+ */
+export declare function createUniformBuffer(device: GPUDevice, data: ArrayBuffer): GPUBuffer;
+/**
+ * Create a read-only storage buffer for params that contain arrays.
+ * Storage buffers don't have the 16-byte array element alignment requirement
+ * that uniform buffers do in WGSL.
+ */
+export declare function createStorageParamsBuffer(device: GPUDevice, data: ArrayBuffer): GPUBuffer;
+export declare function getOrCreatePipeline(device: GPUDevice, shaderCode: string, entryPoint?: string): GPUComputePipeline;
+//# sourceMappingURL=gpu_backend.d.ts.map

package/dist/gpu_backend.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"gpu_backend.d.ts","sourceRoot":"","sources":["../src/gpu_backend.ts"],"names":[],"mappings":"AASA,wBAAsB,SAAS,IAAI,OAAO,CAAC,SAAS,CAAC,CAOpD;AAED,wBAAgB,aAAa,IAAI,IAAI,CAMpC;AAOD;;;;;;;GAOG;AACH,wBAAsB,mBAAmB,IAAI,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC,CA2BrE;AAED,wBAAgB,uBAAuB,IAAI,IAAI,CAM9C;AAID;;GAEG;AACH,wBAAgB,YAAY,CAAC,MAAM,EAAE,SAAS,EAAE,IAAI,EAAE,YAAY,GAAG,SAAS,CAa7E;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,MAAM,EAAE,SAAS,EAAE,KAAK,EAAE,MAAM,GAAG,SAAS,CAK9E;AAED;;GAEG;AACH,wBAAsB,cAAc,CAChC,MAAM,EAAE,SAAS,EACjB,SAAS,EAAE,SAAS,EACpB,KAAK,EAAE,MAAM,GACd,OAAO,CAAC,YAAY,CAAC,CAkBvB;AAED;;;GAGG;AACH,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,SAAS,EAAE,IAAI,EAAE,WAAW,GAAG,SAAS,CAUnF;AAED;;;;GAIG;AACH,wBAAgB,yBAAyB,CAAC,MAAM,EAAE,SAAS,EAAE,IAAI,EAAE,WAAW,GAAG,SAAS,CASzF;AAID,wBAAgB,mBAAmB,CAC/B,MAAM,EAAE,SAAS,EACjB,UAAU,EAAE,MAAM,EAClB,UAAU,GAAE,MAAe,GAC5B,kBAAkB,CAYpB"}

package/dist/gpu_backend.js

@@ -0,0 +1,163 @@
+import { create, globals } from 'webgpu';
+Object.assign(globalThis, globals);
+let _device = null;
+let _gpu = null;
+const pipelineCache = new Map();
+export async function getDevice() {
+    if (_device)
+        return _device;
+    _gpu = create([]);
+    const adapter = await _gpu.requestAdapter();
+    if (!adapter)
+        throw new Error('No WebGPU adapter found');
+    _device = await adapter.requestDevice();
+    return _device;
+}
+export function destroyDevice() {
+    _device?.destroy();
+    _device = null;
+    _gpu = null;
+    destroyTensorCoreDevice();
+    pipelineCache.clear();
+}
+// ---- Experimental tensor core (subgroup matrix) device ----
+let _tcDevice = null;
+let _tcProbed = false;
+/**
+ * Try to obtain a GPUDevice with experimental subgroup matrix support.
+ * Returns null when the feature is not available on this adapter/platform.
+ * Dawn (the backend for the `webgpu` npm package) exposes Apple simdgroup_matrix
+ * (8x8 f32 on Metal/Apple7+) and Vulkan VK_KHR_cooperative_matrix behind the
+ * `allow_unsafe_apis` toggle and the `chromium-experimental-subgroup-matrix`
+ * device feature.
+ */
+export async function getTensorCoreDevice() {
+    if (_tcDevice)
+        return _tcDevice;
+    if (_tcProbed)
+        return null;
+    _tcProbed = true;
+    try {
+        const gpu = create([
+            'enable-dawn-features=allow_unsafe_apis',
+        ]);
+        const adapter = await gpu.requestAdapter();
+        if (!adapter)
+            return null;
+        const featureName = adapter.features.has('chromium-experimental-subgroup-matrix')
+            ? 'chromium-experimental-subgroup-matrix'
+            : adapter.features.has('subgroup-matrix')
+                ? 'subgroup-matrix'
+                : null;
+        if (!featureName)
+            return null;
+        _tcDevice = await adapter.requestDevice({
+            requiredFeatures: [featureName],
+        });
+        return _tcDevice;
+    }
+    catch {
+        return null;
+    }
+}
+export function destroyTensorCoreDevice() {
+    // Intentionally skip device.destroy() -- Dawn's experimental subgroup matrix
+    // path can SIGSEGV during teardown (the compute results are correct).
+    // Letting the device be GC'd avoids the crash in test/process-exit scenarios.
+    _tcDevice = null;
+    _tcProbed = false;
+}
+// Buffer helpers 
+/**
+ * Upload a Float64Array to the GPU as f32. Returns a STORAGE | COPY_SRC buffer.
+ */
+export function uploadBuffer(device, data) {
+    const f32 = new Float32Array(data.length);
+    for (let i = 0; i < data.length; i++) {
+        f32[i] = Math.fround(data[i]);
+    }
+    const buf = device.createBuffer({
+        size: f32.byteLength,
+        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,
+        mappedAtCreation: true,
+    });
+    new Float32Array(buf.getMappedRange()).set(f32);
+    buf.unmap();
+    return buf;
+}
+/**
+ * Create a GPU buffer for compute output (read-write storage, copyable for readback).
+ */
+export function createOutputBuffer(device, count) {
+    return device.createBuffer({
+        size: count * Float32Array.BYTES_PER_ELEMENT,
+        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,
+    });
+}
+/**
+ * Read a GPU buffer back to a Float64Array (f32 -> f64).
+ */
+export async function readbackBuffer(device, srcBuffer, count) {
+    const byteLen = count * Float32Array.BYTES_PER_ELEMENT;
+    const staging = device.createBuffer({
+        size: byteLen,
+        usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST,
+    });
+    const encoder = device.createCommandEncoder();
+    encoder.copyBufferToBuffer(srcBuffer, 0, staging, 0, byteLen);
+    device.queue.submit([encoder.finish()]);
+    await staging.mapAsync(GPUMapMode.READ);
+    const f32 = new Float32Array(staging.getMappedRange());
+    const result = new Float64Array(count);
+    for (let i = 0; i < count; i++) {
+        result[i] = f32[i];
+    }
+    staging.unmap();
+    staging.destroy();
+    return result;
+}
+/**
+ * Create a uniform buffer from an ArrayBuffer of packed metadata.
+ * Only suitable for simple structs without array members (WGSL alignment rules).
+ */
+export function createUniformBuffer(device, data) {
+    const aligned = Math.ceil(data.byteLength / 16) * 16;
+    const buf = device.createBuffer({
+        size: Math.max(aligned, 16),
+        usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,
+        mappedAtCreation: true,
+    });
+    new Uint8Array(buf.getMappedRange(0, data.byteLength)).set(new Uint8Array(data));
+    buf.unmap();
+    return buf;
+}
+/**
+ * Create a read-only storage buffer for params that contain arrays.
+ * Storage buffers don't have the 16-byte array element alignment requirement
+ * that uniform buffers do in WGSL.
+ */
+export function createStorageParamsBuffer(device, data) {
+    const buf = device.createBuffer({
+        size: Math.max(data.byteLength, 4),
+        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
+        mappedAtCreation: true,
+    });
+    new Uint8Array(buf.getMappedRange(0, data.byteLength)).set(new Uint8Array(data));
+    buf.unmap();
+    return buf;
+}
+// Pipeline cache
+export function getOrCreatePipeline(device, shaderCode, entryPoint = 'main') {
+    const key = shaderCode;
+    let pipeline = pipelineCache.get(key);
+    if (pipeline)
+        return pipeline;
+    const module = device.createShaderModule({ code: shaderCode });
+    pipeline = device.createComputePipeline({
+        layout: 'auto',
+        compute: { module, entryPoint },
+    });
+    pipelineCache.set(key, pipeline);
+    return pipeline;
+}
+//# sourceMappingURL=gpu_backend.js.map

package/dist/gpu_backend.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"gpu_backend.js","sourceRoot":"","sources":["../src/gpu_backend.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,QAAQ,CAAC;AAEzC,MAAM,CAAC,MAAM,CAAC,UAAU,EAAE,OAAO,CAAC,CAAC;AAEnC,IAAI,OAAO,GAAqB,IAAI,CAAC;AACrC,IAAI,IAAI,GAAe,IAAI,CAAC;AAE5B,MAAM,aAAa,GAAG,IAAI,GAAG,EAA8B,CAAC;AAE5D,MAAM,CAAC,KAAK,UAAU,SAAS;IAC3B,IAAI,OAAO;QAAE,OAAO,OAAO,CAAC;IAC5B,IAAI,GAAG,MAAM,CAAC,EAAE,CAAmB,CAAC;IACpC,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,cAAc,EAAE,CAAC;IAC5C,IAAI,CAAC,OAAO;QAAE,MAAM,IAAI,KAAK,CAAC,yBAAyB,CAAC,CAAC;IACzD,OAAO,GAAG,MAAM,OAAO,CAAC,aAAa,EAAE,CAAC;IACxC,OAAO,OAAO,CAAC;AACnB,CAAC;AAED,MAAM,UAAU,aAAa;IACzB,OAAO,EAAE,OAAO,EAAE,CAAC;IACnB,OAAO,GAAG,IAAI,CAAC;IACf,IAAI,GAAG,IAAI,CAAC;IACZ,uBAAuB,EAAE,CAAC;IAC1B,aAAa,CAAC,KAAK,EAAE,CAAC;AAC1B,CAAC;AAED,8DAA8D;AAE9D,IAAI,SAAS,GAAqB,IAAI,CAAC;AACvC,IAAI,SAAS,GAAG,KAAK,CAAC;AAEtB;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,mBAAmB;IACrC,IAAI,SAAS;QAAE,OAAO,SAAS,CAAC;IAChC,IAAI,SAAS;QAAE,OAAO,IAAI,CAAC;IAC3B,SAAS,GAAG,IAAI,CAAC;IAEjB,IAAI,CAAC;QACD,MAAM,GAAG,GAAG,MAAM,CAAC;YACf,wCAAwC;SAC3C,CAAmB,CAAC;QACrB,MAAM,OAAO,GAAG,MAAM,GAAG,CAAC,cAAc,EAAE,CAAC;QAC3C,IAAI,CAAC,OAAO;YAAE,OAAO,IAAI,CAAC;QAE1B,MAAM,WAAW,GACb,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAC,uCAAuC,CAAC;YACzD,CAAC,CAAC,uCAAuC;YACzC,CAAC,CAAC,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAC,iBAAiB,CAAC;gBACrC,CAAC,CAAC,iBAAiB;gBACnB,CAAC,CAAC,IAAI,CAAC;QACnB,IAAI,CAAC,WAAW;YAAE,OAAO,IAAI,CAAC;QAE9B,SAAS,GAAG,MAAM,OAAO,CAAC,aAAa,CAAC;YACpC,gBAAgB,EAAE,CAAC,WAA6B,CAAC;SACpD,CAAC,CAAC;QACH,OAAO,SAAS,CAAC;IACrB,CAAC;IAAC,MAAM,CAAC;QACL,OAAO,IAAI,CAAC;IAChB,CAAC;AACL,CAAC;AAED,MAAM,UAAU,uBAAuB;IACnC,6EAA6E;IAC7E,sEAAsE;IACtE,8EAA8E;IAC9E,SAAS,GAAG,IAAI,CAAC;IACjB,SAAS,GAAG,KAAK,CAAC;AACtB,CAAC;AAED,kBAAkB;AAElB;;GAEG;AACH,MAAM,UAAU,YAAY,CAAC,MAAiB,EAAE,IAAkB;IAC9D,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAC1C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QACnC,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAE,CAAC,CAAC;IACnC,CAAC;IACD,MAAM,GAAG,GAAG,MAAM,CAAC,YAAY,CAAC;QAC5B,IAAI,EAAE,GAAG,CAAC,UAAU;QACpB,KAAK,EAAE,cAAc,CAAC,OAAO,GAAG,cAAc,CAAC,QAAQ;QACvD,gBAAgB,EAAE,IAAI;KACzB,CAAC,CAAC;IACH,IAAI,YAAY,CAAC,GAAG,CAAC,cAAc,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAChD,GAAG,CAAC,KAAK,EAAE,CAAC;IACZ,OAAO,GAAG,CAAC;AACf,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,kBAAkB,CAAC,MAAiB,EAAE,KAAa;IAC/D,OAAO,MAAM,CAAC,YAAY,CAAC;QACvB,IAAI,EAAE,KAAK,GAAG,YAAY,CAAC,iBAAiB;QAC5C,KAAK,EAAE,cAAc,CAAC,OAAO,GAAG,cAAc,CAAC,QAAQ;KAC1D,CAAC,CAAC;AACP,CAAC;AAED;;GAEG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc,CAChC,MAAiB,EACjB,SAAoB,EACpB,KAAa;IAEb,MAAM,OAAO,GAAG,KAAK,GAAG,YAAY,CAAC,iBAAiB,CAAC;IACvD,MAAM,OAAO,GAAG,MAAM,CAAC,YAAY,CAAC;QAChC,IAAI,EAAE,OAAO;QACb,KAAK,EAAE,cAAc,CAAC,QAAQ,GAAG,cAAc,CAAC,QAAQ;KAC3D,CAAC,CAAC;IACH,MAAM,OAAO,GAAG,MAAM,CAAC,oBAAoB,EAAE,CAAC;IAC9C,OAAO,CAAC,kBAAkB,CAAC,SAAS,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,OAAO,CAAC,CAAC;IAC9D,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC;IACxC,MAAM,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC;IACxC,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,OAAO,CAAC,cAAc,EAAE,CAAC,CAAC;IACvD,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,KAAK,CAAC,CAAC;IACvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,EAAE,EAAE,CAAC;QAC7B,MAAM,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,CAAE,CAAC;IACxB,CAAC;IACD,OAAO,CAAC,KAAK,EAAE,CAAC;IAChB,OAAO,CAAC,OAAO,EAAE,CAAC;IAClB,OAAO,MAAM,CAAC;AAClB,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAAC,MAAiB,EAAE,IAAiB;IACpE,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,GAAG,EAAE,CAAC,GAAG,EAAE,CAAC;IACrD,MAAM,GAAG,GAAG,MAAM,CAAC,YAAY,CAAC;QAC5B,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,OAAO,EAAE,EAAE,CAAC;QAC3B,KAAK,EAAE,cAAc,CAAC,OAAO,GAAG,cAAc,CAAC,QAAQ;QACvD,gBAAgB,EAAE,IAAI;KACzB,CAAC,CAAC;IACH,IAAI,UAAU,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC;IACjF,GAAG,CAAC,KAAK,EAAE,CAAC;IACZ,OAAO,GAAG,CAAC;AACf,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,yBAAyB,CAAC,MAAiB,EAAE,IAAiB;IAC1E,MAAM,GAAG,GAAG,MAAM,CAAC,YAAY,CAAC;QAC5B,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,EAAE,CAAC,CAAC;QAClC,KAAK,EAAE,cAAc,CAAC,OAAO,GAAG,cAAc,CAAC,QAAQ;QACvD,gBAAgB,EAAE,IAAI;KACzB,CAAC,CAAC;IACH,IAAI,UAAU,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC;IACjF,GAAG,CAAC,KAAK,EAAE,CAAC;IACZ,OAAO,GAAG,CAAC;AACf,CAAC;AAED,iBAAiB;AAEjB,MAAM,UAAU,mBAAmB,CAC/B,MAAiB,EACjB,UAAkB,EAClB,aAAqB,MAAM;IAE3B,MAAM,GAAG,GAAG,UAAU,CAAC;IACvB,IAAI,QAAQ,GAAG,aAAa,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IACtC,IAAI,QAAQ;QAAE,OAAO,QAAQ,CAAC;IAE9B,MAAM,MAAM,GAAG,MAAM,CAAC,kBAAkB,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,CAAC,CAAC;IAC/D,QAAQ,GAAG,MAAM,CAAC,qBAAqB,CAAC;QACpC,MAAM,EAAE,MAAM;QACd,OAAO,EAAE,EAAE,MAAM,EAAE,UAAU,EAAE;KAClC,CAAC,CAAC;IACH,aAAa,CAAC,GAAG,CAAC,GAAG,EAAE,QAAQ,CAAC,CAAC;IACjC,OAAO,QAAQ,CAAC;AACpB,CAAC"}

package/dist/gpu_kernels.d.ts

@@ -0,0 +1,74 @@
+export declare const WORKGROUP_SIZE = 256;
+export declare const BLOCK_SIZE = 16;
+export declare const UNARY_OPS: Record<string, string>;
+export declare const BINARY_OPS: Record<string, string>;
+export declare function resolveUnaryOp(fn: Function): string;
+export declare function resolveBinaryOp(fn: Function): string;
+export declare const REDUCE_IDENTITY: Record<string, string>;
+/**
+ * Aligned map: shapes & strides match, simple 1:1 element mapping.
+ */
+export declare function buildAlignedMapShader(opBody: string): string;
+/**
+ * Broadcast map: output and input may differ in shape.
+ * Uses storage buffer for params to avoid WGSL uniform array alignment rules.
+ */
+export declare function buildBroadcastMapShader(opBody: string): string;
+/**
+ * Aligned zip: all three tensors share shape & strides.
+ */
+export declare function buildAlignedZipShader(opBody: string): string;
+/**
+ * Broadcast zip: output, a, b may differ in shape.
+ * Uses storage buffer for params to avoid WGSL uniform array alignment rules.
+ */
+export declare function buildBroadcastZipShader(opBody: string): string;
+/**
+ * Sum practice: block-level partial sums using shared memory.
+ * Input: array of length size. Output: array of length ceil(size / WORKGROUP_SIZE).
+ * Each workgroup sums WORKGROUP_SIZE contiguous elements into one output cell.
+ */
+export declare function buildSumPracticeShader(): string;
+/**
+ * General reduce along one dimension.
+ * One workgroup per output element. Threads cooperatively reduce
+ * the reduction dimension using shared memory.
+ * Uses storage buffer for params to avoid WGSL uniform array alignment rules.
+ */
+export declare function buildReduceShader(opBody: string, identity: string): string;
+/**
+ * Tiled matrix multiplication using workgroup shared memory.
+ * Dispatched as 3D: (ceil(N/BLOCK), ceil(M/BLOCK), batchSize).
+ * Each 16x16 workgroup computes one output tile, loading tiles of A and B
+ * into shared memory to satisfy:
+ *   - all data read from shared memory (not global) during accumulation
+ *   - each global cell of A and B read exactly once
+ *   - each thread writes to global memory exactly once
+ * Supports arbitrary broadcast batch dimensions via stride-based indexing.
+ */
+export declare function buildMatMulShader(): string;
+export declare const TC_TILE = 8;
+/**
+ * Experimental tensor-core-accelerated matmul shader using the Dawn/Chrome
+ * `chromium_experimental_subgroup_matrix` extension.
+ *
+ * On Apple Silicon (Metal), this maps to `simdgroup_matrix` 8x8 f32 hardware
+ * instructions. On Vulkan with VK_KHR_cooperative_matrix it maps to the
+ * equivalent SPIR-V ops.
+ *
+ * @param workgroupX - must equal the device's maxSubgroupSize (e.g. 64 on
+ *   Apple M-series via Dawn, 32 or 64 on Vulkan). The WebGPU spec requires
+ *   the x-dimension of workgroup_size to be a multiple of maxSubgroupSize
+ *   when the shader uses subgroup matrices.
+ *
+ * Constraints (checked by the caller on the TS side):
+ *   - All tensors must be contiguous (natural row-major strides)
+ *   - M, N, K must all be multiples of TC_TILE (8)
+ *   - Batch dimensions of A and B must match exactly (no broadcasting)
+ *
+ * Dispatch: (N/TC_TILE, M/TC_TILE, batchSize).
+ * Each workgroup computes one 8x8 output tile by tiling over the K dimension
+ * with subgroupMatrixLoad / MultiplyAccumulate / Store.
+ */
+export declare function buildTensorCoreMatMulShader(workgroupX: number): string;
+//# sourceMappingURL=gpu_kernels.d.ts.map

package/dist/gpu_kernels.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"gpu_kernels.d.ts","sourceRoot":"","sources":["../src/gpu_kernels.ts"],"names":[],"mappings":"AAEA,eAAO,MAAM,cAAc,MAAM,CAAC;AAClC,eAAO,MAAM,UAAU,KAAK,CAAC;AAK7B,eAAO,MAAM,SAAS,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAQ5C,CAAC;AAEF,eAAO,MAAM,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAO7C,CAAC;AAoBF,wBAAgB,cAAc,CAAC,EAAE,EAAE,QAAQ,GAAG,MAAM,CAInD;AAED,wBAAgB,eAAe,CAAC,EAAE,EAAE,QAAQ,GAAG,MAAM,CAIpD;AAED,eAAO,MAAM,eAAe,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAIlD,CAAC;AA4CF;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAiB5D;AAED;;;GAGG;AACH,wBAAgB,uBAAuB,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAqC9D;AAED;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAkB5D;AAED;;;GAGG;AACH,wBAAgB,uBAAuB,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CA4C9D;AAED;;;;GAIG;AACH,wBAAgB,sBAAsB,IAAI,MAAM,CAqC/C;AAED;;;;;GAKG;AACH,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,MAAM,CAkE1E;AAED;;;;;;;;;GASG;AACH,wBAAgB,iBAAiB,IAAI,MAAM,CAsI1C;AAID,eAAO,MAAM,OAAO,IAAI,CAAC;AAEzB;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAgB,2BAA2B,CAAC,UAAU,EAAE,MAAM,GAAG,MAAM,CAwDtE"}