onnxruntime/js/web/script/test-runner-cli-args.ts

496 lines
19 KiB
TypeScript
Raw Normal View History

// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
import minimist from 'minimist';
import npmlog from 'npmlog';
import {Env, InferenceSession} from 'onnxruntime-common';
import {Logger} from '../lib/onnxjs/instrument';
import {Test} from '../test/test-types';
/* eslint-disable max-len */
const HELP_MESSAGE = `
test-runner-cli
Run ONNX Runtime Web tests, models, benchmarks in different environments.
Usage:
test-runner-cli <mode> ... [options]
Modes:
2021-07-23 07:49:57 +00:00
suite0 Run all unittests, all operator tests and node model tests that described in suite test list
suite1 Run all operator tests and node model tests that described in suite test list
model Run a single model test
unittest Run all unittests
op Run a single operator test
Options:
*** General Options ***
-h, --help Print this message.
-d, --debug Specify to run test runner in debug mode.
Debug mode outputs verbose log for test runner, sets up environment debug flag, and keeps karma not to exit after tests completed.
-b=<...>, --backend=<...> Specify one or more backend(s) to run the test upon.
Backends can be one or more of the following, splitted by comma:
webgl
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
webgpu
wasm
[js/web] add 'xnnpack' to EP list (#12723) **Description**: This PR adds support for "XNNPACK EP" in ORTWeb and changes the behavior of how ORTWeb deals with "backends", or "EPs" in API. **Background**: Term "backend" is introduced in ONNX.js to representing a TypeScript type which implements a "backend" interface, which is a similar but different concept to ORT's EP (execution provider). There was 3 backends in ONNX.js: "cpu", "wasm" and "webgl". When ORT Web is launched, the concept is derived to help users to integrate smoothly. Technically, when "wasm" backend is used, users need to also specify "EP" in the session options. Considering it may get complicated and confused for users to figure out the difference between "backend" and "EP", the JS API hide the "backend" concept and made a mapping between names, backends and EPs: "webgl" (Name) <==> "onnxjsBackend" (Backend) "wasm" (Name) <==> "wasmBackend" (Backend) <==> "CPU" (EP) **Details**: The following changes are applied in this PR: 1. allow multi-registration for backends using the same name. This is for use scenarios where both "onnxruntime-node" and "onnxruntime-web" are consumed in a Node.js App ( so "cpu" will be registered twice in this scenario. ) 2. re-assign priority values to backends. I give 100 as base to "cpu" for node and react_native, and 10 as base to "cpu" in web. 3. add "cpu", "xnnpack" as new names of backends. 4. update onnxruntime wasm exported functions to support EP registration. 5. update implementations in ort web to handle execution providers in session options. 6. add '--use_xnnpack' as default build flag for ort-web
2022-10-03 17:38:45 +00:00
xnnpack
webnn
-e=<...>, --env=<...> Specify the environment to run the test. Should be one of the following:
chrome (default)
edge (Windows only)
firefox
electron
safari (MacOS only)
node
bs (for BrowserStack tests)
-p, --profile Enable profiler.
Profiler will generate extra logs which include the information of events time consumption
-P[=<...>], --perf[=<...>] Generate performance number. Cannot be used with flag --debug.
This flag can be used with a number as value, specifying the total count of test cases to run. The test cases may be used multiple times. Default value is 10.
-c, --file-cache Enable file cache.
[js/webgpu] support IO binding (#17480) <del> **This PR is based on a few prerequisites PRs. They are listed as below:** - #17465 - #17469 - #17470 - #17472 - #17473 - #17484 Please review the current change by only looking at commit e2e6623e673ec6de55a5c1f8edcbd3a46b535a89 and later. </del> ### Description This PR introduces WebGPU IO binding. This new feature allows onnxruntime-web users to use tensors created from GPU as model input/output so that a model inferencing can be done without unnecessary data copy between CPU and GPU for model input/output. ### Examples An E2E demo/example is being worked on. Following is some simple demo with code snippet. Let's first check today how we do: ```js // STEP.1 - create an inference session: const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'] }); // STEP.2 - create model input: (supposing myImageCpuData is a Float32Array) const feeds = { 'input_image:0': new ort.Tensor('float32', myImageCpuData, [1, 224, 224, 3]) }; // STEP.3 - run model const myResults = await mySession.run(feeds); // STEP.4 - get output data const myData = myResults['output_image:0'].data; // Float32Array ``` #### for inputs (GPU tensor): Now, with IO binding, you can create a tensor from a GPU buffer, and feed it to the model: ```js // new STEP.2.A - create model input from a GPU buffer: (supposing myInputGpuBuffer is a `GPUBuffer` object with input data) const feeds = { 'input_image:0': ort.Tensor.fromGpuBuffer(myInputGpuBuffer, { dataType: 'float32', dims: [1, 224, 224, 3] }) }; ``` ### for outputs (pre-allocated GPU tensor) you can also do that for output, **if you know the output shape**: ```js // new STEP.2.B - create model output from a GPU buffer: (supposing myOutputGpuBuffer is a pre-allocated `GPUBuffer` object) const fetches = { 'output_image:0': ort.Tensor.fromGpuBuffer(myOutputGpuBuffer, { dataType: 'float32', dims: [1, 512, 512, 3] }) }; // new STEP.3 - run model with pre-allocated output (fetches) const myResults = await mySession.run(feeds, fetches); ``` ### for outputs (specify location) if you do not know the output shape, you can specify the output location when creating the session: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: "gpu-buffer" }); ``` if the model has multiple outputs, you can specify them seperately: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: { "output_image:0": "gpu-buffer" } }); ``` now you don't need to prepare the `fetches` object and onnxruntime-web will prepare output data on the location that specified. #### read data when you get the output tensor, you can: ```js // get the gpu buffer object: const gpuBuffer = myOutputTensor.gpuBuffer; // GPUBuffer // get the CPU data asynchronizely const cpuData = await myOutputTensor.getData(); // get the CPU data asynchronizely and release the underlying GPU resources const cpuData = await myOutputTensor.getData(true); // dispose the tensor (release the underlying GPU resources). This tensor object will be invalid after dispose() is called. myOutputTensor.dispose(); ``` #### resource management JavaScript has GC so you don't need to worry about managing JavaScript objects. But there are 2 types of resources that are not managed by GC: - GPU buffer that used in tensors - Underlying ORT native resources To simplify, most of the unmanaged resources and handled inside ORT web. But there are a few resources that need users to manage: - All external GPU resources, including GPU buffers inside all tensors created by `Tensor.fromGpuBuffer()`, will not be managed by ORT. User should manage those GPU buffers themselves. - When a session is created with `preferredOutputLocation` == "gpu-buffer" specified in session options, and the corresponding output is not pre-allocated, user need to call the output tensor's `dispose()` or `getData(true)` to manually release the underlying GPU buffers. - ORT internal errors (including providing a pre-allocated output tensor with wrong type/dims) will invalidate the whole wasm memory and is not recoverable. An exception is thrown in this situation.
2023-09-29 18:24:42 +00:00
-i=<...>, --io-binding=<...> Specify the IO binding testing type. Should be one of the following:
none (default)
gpu-tensor use pre-allocated GPU tensors for inputs and outputs
gpu-location use pre-allocated GPU tensors for inputs and set preferredOutputLocation to 'gpu-buffer'
*** Session Options ***
-u=<...>, --optimized-model-file-path=<...> Specify whether to dump the optimized model.
-o=<...>, --graph-optimization-level=<...> Specify graph optimization level.
Default is 'all'. Valid values are 'disabled', 'basic', 'extended', 'all'.
*** Logging Options ***
--log-verbose=<...> Set log level to verbose
--log-info=<...> Set log level to info
--log-warning=<...> Set log level to warning
--log-error=<...> Set log level to error
The 4 flags above specify the logging configuration. Each flag allows to specify one or more category(s), splitted by comma. If use the flags without value, the log level will be applied to all category.
*** Backend Options ***
-x, --wasm-number-threads Set the WebAssembly number of threads
--wasm-init-timeout Set the timeout for WebAssembly backend initialization, in milliseconds
--wasm-enable-simd Set whether to enable SIMD
--wasm-enable-proxy Set whether to enable proxy worker
--webgl-context-id Set the WebGL context ID (webgl/webgl2)
--webgl-matmul-max-batch-size Set the WebGL matmulMaxBatchSize
--webgl-texture-cache-mode Set the WebGL texture cache mode (initializerOnly/full)
--webgl-texture-pack-mode Set the WebGL texture pack mode (true/false)
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
--webgpu-profiling-mode Set the WebGPU profiling mode (off/default)
*** Browser Options ***
--no-sandbox This flag will be passed to Chrome.
Sometimes Chrome need this flag to work together with Karma.
--chromium-flags=<...> This flag will be passed to Chrome and Edge browsers. Can be used multiple times.
Examples:
Run all suite0 tests:
> test-runner-cli suite0
Run single model test (test_relu) on WebAssembly backend
> test-runner-cli model test_relu --backend=wasm
Debug unittest
> test-runner-cli unittest --debug
Debug operator matmul, highlight verbose log from BaseGlContext and WebGLBackend
> test-runner-cli op matmul --backend=webgl --debug --log-verbose=BaseGlContext,WebGLBackend
Profile an ONNX model on WebGL backend
> test-runner-cli model <model_folder> --profile --backend=webgl
Run perf testing of an ONNX model on WebGL backend
> test-runner-cli model <model_folder> -b=webgl -P
`;
/* eslint-enable max-len */
export declare namespace TestRunnerCliArgs {
type Mode = 'suite0'|'suite1'|'model'|'unittest'|'op';
type Backend = 'cpu'|'webgl'|'webgpu'|'wasm'|'onnxruntime'|'xnnpack'|'webnn';
type Environment = 'chrome'|'edge'|'firefox'|'electron'|'safari'|'node'|'bs';
type BundleMode = 'dev'|'perf';
[js/webgpu] support IO binding (#17480) <del> **This PR is based on a few prerequisites PRs. They are listed as below:** - #17465 - #17469 - #17470 - #17472 - #17473 - #17484 Please review the current change by only looking at commit e2e6623e673ec6de55a5c1f8edcbd3a46b535a89 and later. </del> ### Description This PR introduces WebGPU IO binding. This new feature allows onnxruntime-web users to use tensors created from GPU as model input/output so that a model inferencing can be done without unnecessary data copy between CPU and GPU for model input/output. ### Examples An E2E demo/example is being worked on. Following is some simple demo with code snippet. Let's first check today how we do: ```js // STEP.1 - create an inference session: const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'] }); // STEP.2 - create model input: (supposing myImageCpuData is a Float32Array) const feeds = { 'input_image:0': new ort.Tensor('float32', myImageCpuData, [1, 224, 224, 3]) }; // STEP.3 - run model const myResults = await mySession.run(feeds); // STEP.4 - get output data const myData = myResults['output_image:0'].data; // Float32Array ``` #### for inputs (GPU tensor): Now, with IO binding, you can create a tensor from a GPU buffer, and feed it to the model: ```js // new STEP.2.A - create model input from a GPU buffer: (supposing myInputGpuBuffer is a `GPUBuffer` object with input data) const feeds = { 'input_image:0': ort.Tensor.fromGpuBuffer(myInputGpuBuffer, { dataType: 'float32', dims: [1, 224, 224, 3] }) }; ``` ### for outputs (pre-allocated GPU tensor) you can also do that for output, **if you know the output shape**: ```js // new STEP.2.B - create model output from a GPU buffer: (supposing myOutputGpuBuffer is a pre-allocated `GPUBuffer` object) const fetches = { 'output_image:0': ort.Tensor.fromGpuBuffer(myOutputGpuBuffer, { dataType: 'float32', dims: [1, 512, 512, 3] }) }; // new STEP.3 - run model with pre-allocated output (fetches) const myResults = await mySession.run(feeds, fetches); ``` ### for outputs (specify location) if you do not know the output shape, you can specify the output location when creating the session: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: "gpu-buffer" }); ``` if the model has multiple outputs, you can specify them seperately: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: { "output_image:0": "gpu-buffer" } }); ``` now you don't need to prepare the `fetches` object and onnxruntime-web will prepare output data on the location that specified. #### read data when you get the output tensor, you can: ```js // get the gpu buffer object: const gpuBuffer = myOutputTensor.gpuBuffer; // GPUBuffer // get the CPU data asynchronizely const cpuData = await myOutputTensor.getData(); // get the CPU data asynchronizely and release the underlying GPU resources const cpuData = await myOutputTensor.getData(true); // dispose the tensor (release the underlying GPU resources). This tensor object will be invalid after dispose() is called. myOutputTensor.dispose(); ``` #### resource management JavaScript has GC so you don't need to worry about managing JavaScript objects. But there are 2 types of resources that are not managed by GC: - GPU buffer that used in tensors - Underlying ORT native resources To simplify, most of the unmanaged resources and handled inside ORT web. But there are a few resources that need users to manage: - All external GPU resources, including GPU buffers inside all tensors created by `Tensor.fromGpuBuffer()`, will not be managed by ORT. User should manage those GPU buffers themselves. - When a session is created with `preferredOutputLocation` == "gpu-buffer" specified in session options, and the corresponding output is not pre-allocated, user need to call the output tensor's `dispose()` or `getData(true)` to manually release the underlying GPU buffers. - ORT internal errors (including providing a pre-allocated output tensor with wrong type/dims) will invalidate the whole wasm memory and is not recoverable. An exception is thrown in this situation.
2023-09-29 18:24:42 +00:00
type IOBindingMode = 'none'|'gpu-tensor'|'gpu-location';
}
export interface TestRunnerCliArgs {
debug: boolean;
mode: TestRunnerCliArgs.Mode;
/**
* The parameter that used when in mode 'model' or 'op', specifying the search string for the model or op test
*/
param?: string;
backends: [TestRunnerCliArgs.Backend];
env: TestRunnerCliArgs.Environment;
/**
* Bundle Mode
*
* this field affects the behavior of Karma and build script.
*
* Mode "perf":
* - use "dist/ort.all.min.js" as main file
* - use "test/ort.test.min.js" as test file
* Mode "dev":
* - use "dist/ort.all.js" as main file
* - use "test/ort.test.js" as test file
*/
bundleMode: TestRunnerCliArgs.BundleMode;
[js/webgpu] support IO binding (#17480) <del> **This PR is based on a few prerequisites PRs. They are listed as below:** - #17465 - #17469 - #17470 - #17472 - #17473 - #17484 Please review the current change by only looking at commit e2e6623e673ec6de55a5c1f8edcbd3a46b535a89 and later. </del> ### Description This PR introduces WebGPU IO binding. This new feature allows onnxruntime-web users to use tensors created from GPU as model input/output so that a model inferencing can be done without unnecessary data copy between CPU and GPU for model input/output. ### Examples An E2E demo/example is being worked on. Following is some simple demo with code snippet. Let's first check today how we do: ```js // STEP.1 - create an inference session: const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'] }); // STEP.2 - create model input: (supposing myImageCpuData is a Float32Array) const feeds = { 'input_image:0': new ort.Tensor('float32', myImageCpuData, [1, 224, 224, 3]) }; // STEP.3 - run model const myResults = await mySession.run(feeds); // STEP.4 - get output data const myData = myResults['output_image:0'].data; // Float32Array ``` #### for inputs (GPU tensor): Now, with IO binding, you can create a tensor from a GPU buffer, and feed it to the model: ```js // new STEP.2.A - create model input from a GPU buffer: (supposing myInputGpuBuffer is a `GPUBuffer` object with input data) const feeds = { 'input_image:0': ort.Tensor.fromGpuBuffer(myInputGpuBuffer, { dataType: 'float32', dims: [1, 224, 224, 3] }) }; ``` ### for outputs (pre-allocated GPU tensor) you can also do that for output, **if you know the output shape**: ```js // new STEP.2.B - create model output from a GPU buffer: (supposing myOutputGpuBuffer is a pre-allocated `GPUBuffer` object) const fetches = { 'output_image:0': ort.Tensor.fromGpuBuffer(myOutputGpuBuffer, { dataType: 'float32', dims: [1, 512, 512, 3] }) }; // new STEP.3 - run model with pre-allocated output (fetches) const myResults = await mySession.run(feeds, fetches); ``` ### for outputs (specify location) if you do not know the output shape, you can specify the output location when creating the session: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: "gpu-buffer" }); ``` if the model has multiple outputs, you can specify them seperately: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: { "output_image:0": "gpu-buffer" } }); ``` now you don't need to prepare the `fetches` object and onnxruntime-web will prepare output data on the location that specified. #### read data when you get the output tensor, you can: ```js // get the gpu buffer object: const gpuBuffer = myOutputTensor.gpuBuffer; // GPUBuffer // get the CPU data asynchronizely const cpuData = await myOutputTensor.getData(); // get the CPU data asynchronizely and release the underlying GPU resources const cpuData = await myOutputTensor.getData(true); // dispose the tensor (release the underlying GPU resources). This tensor object will be invalid after dispose() is called. myOutputTensor.dispose(); ``` #### resource management JavaScript has GC so you don't need to worry about managing JavaScript objects. But there are 2 types of resources that are not managed by GC: - GPU buffer that used in tensors - Underlying ORT native resources To simplify, most of the unmanaged resources and handled inside ORT web. But there are a few resources that need users to manage: - All external GPU resources, including GPU buffers inside all tensors created by `Tensor.fromGpuBuffer()`, will not be managed by ORT. User should manage those GPU buffers themselves. - When a session is created with `preferredOutputLocation` == "gpu-buffer" specified in session options, and the corresponding output is not pre-allocated, user need to call the output tensor's `dispose()` or `getData(true)` to manually release the underlying GPU buffers. - ORT internal errors (including providing a pre-allocated output tensor with wrong type/dims) will invalidate the whole wasm memory and is not recoverable. An exception is thrown in this situation.
2023-09-29 18:24:42 +00:00
ioBindingMode: TestRunnerCliArgs.IOBindingMode;
logConfig: Test.Config['log'];
/**
* Whether to enable InferenceSession's profiler
*/
profile: boolean;
/**
* Whether to enable file cache
*/
fileCache: boolean;
/**
* Specify the times that test cases to run
*/
times?: number;
/**
* whether to dump the optimized model
*/
optimizedModelFilePath?: string;
/**
* Specify graph optimization level
*/
graphOptimizationLevel: 'disabled'|'basic'|'extended'|'all';
cpuOptions?: InferenceSession.CpuExecutionProviderOption;
cudaOptions?: InferenceSession.CudaExecutionProviderOption;
cudaFlags?: Record<string, unknown>;
wasmOptions?: InferenceSession.WebAssemblyExecutionProviderOption;
webglOptions?: InferenceSession.WebGLExecutionProviderOption;
globalEnvFlags?: Test.Options['globalEnvFlags'];
noSandbox?: boolean;
chromiumFlags: string[];
}
function parseBooleanArg(arg: unknown, defaultValue: boolean): boolean;
function parseBooleanArg(arg: unknown): boolean|undefined;
function parseBooleanArg(arg: unknown, defaultValue?: boolean): boolean|undefined {
if (typeof arg === 'undefined') {
return defaultValue;
}
if (typeof arg === 'boolean') {
return arg;
}
if (typeof arg === 'number') {
return arg !== 0;
}
if (typeof arg === 'string') {
if (arg.toLowerCase() === 'true') {
return true;
}
if (arg.toLowerCase() === 'false') {
return false;
}
}
throw new TypeError(`invalid boolean arg: ${arg}`);
}
function parseLogLevel<T>(arg: T) {
let v: string[]|boolean;
if (typeof arg === 'string') {
v = arg.split(',');
} else if (Array.isArray(arg)) {
v = [];
for (const e of arg) {
v.push(...e.split(','));
}
} else {
v = arg ? true : false;
}
return v;
}
function parseLogConfig(args: minimist.ParsedArgs) {
const config: Array<{category: string; config: Logger.Config}> = [];
const verbose = parseLogLevel(args['log-verbose']);
const info = parseLogLevel(args['log-info']);
const warning = parseLogLevel(args['log-warning']);
const error = parseLogLevel(args['log-error']);
if (typeof error === 'boolean' && error) {
config.push({category: '*', config: {minimalSeverity: 'error'}});
} else if (typeof warning === 'boolean' && warning) {
config.push({category: '*', config: {minimalSeverity: 'warning'}});
} else if (typeof info === 'boolean' && info) {
config.push({category: '*', config: {minimalSeverity: 'info'}});
} else if (typeof verbose === 'boolean' && verbose) {
config.push({category: '*', config: {minimalSeverity: 'verbose'}});
}
if (Array.isArray(error)) {
config.push(...error.map(i => ({category: i, config: {minimalSeverity: 'error' as Logger.Severity}})));
}
if (Array.isArray(warning)) {
config.push(...warning.map(i => ({category: i, config: {minimalSeverity: 'warning' as Logger.Severity}})));
}
if (Array.isArray(info)) {
config.push(...info.map(i => ({category: i, config: {minimalSeverity: 'info' as Logger.Severity}})));
}
if (Array.isArray(verbose)) {
config.push(...verbose.map(i => ({category: i, config: {minimalSeverity: 'verbose' as Logger.Severity}})));
}
return config;
}
function parseCpuOptions(_args: minimist.ParsedArgs): InferenceSession.CpuExecutionProviderOption {
return {name: 'cpu'};
}
function parseCpuFlags(_args: minimist.ParsedArgs): Record<string, unknown> {
return {};
}
function parseWasmOptions(_args: minimist.ParsedArgs): InferenceSession.WebAssemblyExecutionProviderOption {
return {name: 'wasm'};
}
function parseWasmFlags(args: minimist.ParsedArgs): Env.WebAssemblyFlags {
const numThreads = args.x || args['wasm-number-threads'];
if (typeof numThreads !== 'undefined' && typeof numThreads !== 'number') {
throw new Error('Flag "x"/"wasm-number-threads" must be a number value');
}
const initTimeout = args['wasm-init-timeout'];
if (typeof initTimeout !== 'undefined' && typeof initTimeout !== 'number') {
throw new Error('Flag "wasm-init-timeout" must be a number value');
}
let simd = args['wasm-enable-simd'];
if (simd === 'true') {
simd = true;
} else if (simd === 'false') {
simd = false;
} else if (typeof simd !== 'undefined' && typeof simd !== 'boolean') {
throw new Error('Flag "wasm-enable-simd" must be a boolean value');
}
let proxy = args['wasm-enable-proxy'];
if (proxy === 'true') {
proxy = true;
} else if (proxy === 'false') {
proxy = false;
} else if (typeof proxy !== 'undefined' && typeof proxy !== 'boolean') {
throw new Error('Flag "wasm-enable-proxy" must be a boolean value');
}
return {numThreads, initTimeout, simd, proxy};
}
function parseWebglOptions(_args: minimist.ParsedArgs): InferenceSession.WebGLExecutionProviderOption {
return {name: 'webgl'};
}
[js/api] introducing IO binding for tensor (#16452) [//]: # (## Work In Progress. Feedbacks are welcome!) ### Description This PR adds a few properties, methods and factories to Tensor type to support IO-binding feature. This will allow user to create tensor from GPU/CPU bound data without a force transferring of data between CPU and GPU. This change is a way to resolve #15312 ### Change Summary 1. Add properties to `Tensor` type: a. `location`: indicating where the data is sitting. valid values are `cpu`, `cpu-pinned`, `texture`, `gpu-buffer`. b. `texture`: sit side to `data`, a readonly property of `WebGLTexture` type. available only when `location === 'texture'` c. `gpuBuffer`: sit side to `data`, a readonly property of `GPUBuffer` type. available only when `location === 'gpu-buffer'` 2. Add methods to `Tensor` type (usually dealing with inference outputs): - async function `getData()` allows user to download data from GPU to CPU manually. - function `dispose()` allows user to release GPU resources manually. 3. Add factories for creating `Tensor` instances: a. `fromTexture()` to create a WebGL texture bound tensor data b. `fromGpuBuffer()` to create a WebGPUBuffer bound tensor data c. `fromPinnedBuffer()` to create a tensor using a CPU pinned buffer ### Examples: create tensors from texture and pass to inference session as inputs ```js // when create session, specify we prefer 'image_output:0' to be stored on GPU as texture const session = await InferenceSession.create('./my_model.onnx', { executionProviders: [ 'webgl' ], preferredOutputLocation: { 'image_output:0': 'texture' } }); ... const myImageTexture = getTexture(); // user's function to get a texture const myFeeds = { input0: Tensor.fromTexture(myImageTexture, { width: 224, height: 224 }) }; // shape [1, 224, 224, 4], RGBA format. const results = await session.run(myFeeds); const myOutputTexture = results['image_output:0'].texture; ```
2023-08-29 19:58:26 +00:00
function parseWebglFlags(args: minimist.ParsedArgs): Partial<Env.WebGLFlags> {
const contextId = args['webgl-context-id'];
if (contextId !== undefined && contextId !== 'webgl' && contextId !== 'webgl2') {
throw new Error('Flag "webgl-context-id" is invalid');
}
const matmulMaxBatchSize = args['webgl-matmul-max-batch-size'];
if (matmulMaxBatchSize !== undefined && typeof matmulMaxBatchSize !== 'number') {
throw new Error('Flag "webgl-matmul-max-batch-size" must be a number value');
}
const textureCacheMode = args['webgl-texture-cache-mode'];
if (textureCacheMode !== undefined && textureCacheMode !== 'initializerOnly' && textureCacheMode !== 'full') {
throw new Error('Flag "webgl-texture-cache-mode" is invalid');
}
const pack = args['webgl-texture-pack-mode'];
if (pack !== undefined && typeof pack !== 'boolean') {
throw new Error('Flag "webgl-texture-pack-mode" is invalid');
}
const async = args['webgl-async'];
if (async !== undefined && typeof async !== 'boolean') {
throw new Error('Flag "webgl-async" is invalid');
}
return {contextId, matmulMaxBatchSize, textureCacheMode, pack};
}
[js/api] introducing IO binding for tensor (#16452) [//]: # (## Work In Progress. Feedbacks are welcome!) ### Description This PR adds a few properties, methods and factories to Tensor type to support IO-binding feature. This will allow user to create tensor from GPU/CPU bound data without a force transferring of data between CPU and GPU. This change is a way to resolve #15312 ### Change Summary 1. Add properties to `Tensor` type: a. `location`: indicating where the data is sitting. valid values are `cpu`, `cpu-pinned`, `texture`, `gpu-buffer`. b. `texture`: sit side to `data`, a readonly property of `WebGLTexture` type. available only when `location === 'texture'` c. `gpuBuffer`: sit side to `data`, a readonly property of `GPUBuffer` type. available only when `location === 'gpu-buffer'` 2. Add methods to `Tensor` type (usually dealing with inference outputs): - async function `getData()` allows user to download data from GPU to CPU manually. - function `dispose()` allows user to release GPU resources manually. 3. Add factories for creating `Tensor` instances: a. `fromTexture()` to create a WebGL texture bound tensor data b. `fromGpuBuffer()` to create a WebGPUBuffer bound tensor data c. `fromPinnedBuffer()` to create a tensor using a CPU pinned buffer ### Examples: create tensors from texture and pass to inference session as inputs ```js // when create session, specify we prefer 'image_output:0' to be stored on GPU as texture const session = await InferenceSession.create('./my_model.onnx', { executionProviders: [ 'webgl' ], preferredOutputLocation: { 'image_output:0': 'texture' } }); ... const myImageTexture = getTexture(); // user's function to get a texture const myFeeds = { input0: Tensor.fromTexture(myImageTexture, { width: 224, height: 224 }) }; // shape [1, 224, 224, 4], RGBA format. const results = await session.run(myFeeds); const myOutputTexture = results['image_output:0'].texture; ```
2023-08-29 19:58:26 +00:00
function parseWebgpuFlags(args: minimist.ParsedArgs): Partial<Env.WebGpuFlags> {
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
const profilingMode = args['webgpu-profiling-mode'];
if (profilingMode !== undefined && profilingMode !== 'off' && profilingMode !== 'default') {
throw new Error('Flag "webgpu-profiling-mode" is invalid');
}
const validateInputContent = args['webgpu-validate-input-content'];
if (validateInputContent !== undefined && typeof validateInputContent !== 'boolean') {
throw new Error('Flag "webgpu-validate-input-content" is invalid');
}
return {profilingMode, validateInputContent};
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
}
function parseGlobalEnvFlags(args: minimist.ParsedArgs): NonNullable<TestRunnerCliArgs['globalEnvFlags']> {
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
const wasm = parseWasmFlags(args);
const webgl = parseWebglFlags(args);
const webgpu = parseWebgpuFlags(args);
const cpuFlags = parseCpuFlags(args);
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
return {webgl, wasm, webgpu, ...cpuFlags};
}
export function parseTestRunnerCliArgs(cmdlineArgs: string[]): TestRunnerCliArgs {
const args = minimist(cmdlineArgs);
if (args.help || args.h) {
console.log(HELP_MESSAGE);
process.exit();
}
// Option: -d, --debug
const debug = parseBooleanArg(args.debug || args.d, false);
if (debug) {
npmlog.level = 'verbose';
}
npmlog.verbose('TestRunnerCli.Init', 'Parsing commandline arguments...');
const mode = args._.length === 0 ? 'suite0' : args._[0];
// Option: -e=<...>, --env=<...>
const envArg = args.env || args.e;
const env = (typeof envArg !== 'string') ? 'chrome' : envArg;
if (['chrome', 'edge', 'firefox', 'electron', 'safari', 'node', 'bs'].indexOf(env) === -1) {
throw new Error(`not supported env ${env}`);
}
// Option: -b=<...>, --backend=<...>
const browserBackends = ['webgl', 'webgpu', 'wasm', 'xnnpack', 'webnn'];
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
// TODO: remove this when Chrome support WebNN.
// we need this for now because Chrome does not support webnn yet,
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
// and ChromeCanary is not in CI.
const defaultBrowserBackends = ['webgl', 'webgpu', 'wasm', 'xnnpack' /*, 'webnn'*/];
const nodejsBackends = ['cpu', 'wasm'];
const backendArgs = args.backend || args.b;
[js/web] WebGPU backend via JSEP (#14579) ### Description This change introduced the following new components into ONNX Runtime Web: - JavaScript Execution Provider (JSEP) - Asynchronized inferencing execution powered by Emscripten's Asyncify - WebGPU backend implemented in TypeScript - initial implementation of kernels: - elementwise operators (22) - binary operators (5) - tensor: Shape, Reshape, Transpose, Gemm - nn: Conv, {Global}Maxpool, {Global}AveragePool Code need to be polished. still working on it. ## Q&A What is JSEP? > JSEP, aka JavaScript Execution Provider, is a new ONNXRuntime execution provider that specifically works on Web environment (browsers). JSEP allows JavaScript code to kick in from various places when ONNX Runtime inferences a model. Why JSEP? > JSEP is a hybrid mode EP that contains both C/C++ and TypeScript/JavaScript implementation. There are 2 strong reasons why we introduces JSEP: > 1. the C/C++ part helps JSEP to leverage ONNX Runtime's capabilities as much as possible including graph transformer, optimizers and also the capabilities to fallback to CPU EP. TypeScript/JavaScript helps JSEP to develop and debug much easier in the browser for the kernel implementation. > 2. the requirement of asynchronized execution from JavaScript API (eg. `buffer.mapAsync()`) makes it impossible to run `OrtRun()` in a synchronized context (see "async problem" section below). This is done by using Emscripten's Asyncify. What is WebGPU? > WebGPU is the new GPU API that available in browser. It's one of the only 2 APIs that currently available to access the GPU from browser (the other is WebGL). > WebGPU is designed with more advanced and stronger features comparing to WebGL and is potentially solution that offer the best GPU performance for model inferencing that currently available. What is the async problem and why we have the problem? > The "async problem" is a problem that you cannot call an async function in a synchronous context. Think about the following C++ code: > ```c > // C-style declarations (API) > typedef void (*ON_COMPLETE)(PVOID state, DATA *data); > void read_data_from_file(FILEHANDLE file, ON_COMPLETE on_complete); > > // implementation > DATA * my_impl_read_data_from_file_sync(FILEHANDLE file) { > // how to implement? > } > ``` > The answer is, it's impossible to implement this function. Usually we try to find a sync version API, or launch a thread to call the async function and sync-wait on the main thread. Unfortunately, in browser environment, neither is possible. > > WebGPU does not offer any synchronized API for data downloading (GPU to CPU). This is the only operation that MUST be async. As `OrtRun()` will eventually call into DataTransfer for copy data from GPU to CPU, and `OrtRun()` is a synchronized function, this cannot be done in normal way. What is Emscripten? How is the Asyncify feature resolved the problem? > Emscripten is the C/C++ compiler for WebAssembly. It's what we use to compile ORT and generates the WebAssembly artifacts which runs on browsers. > > Asyncify is a [compiler feature](https://emscripten.org/docs/porting/asyncify.html) that allows calling async functions from a synchronized context. In short, it generates code to unwind and rewind call stack to emulate async execution. With this feature, we are able to call the async function inside `OrtRun()` call. ## Design Overview **Inter-op** JSEP is doing pretty much same thing to just another EP. It exposes an interface for inter-op with JavaScript, which is defined in onnxruntime/wasm/js_internal_api.js: ```js // init JSEP Module["jsepInit"] = function (backend, alloc, free, copy, copyAsync, createKernel, releaseKernel, run) { Module.jsepBackend = backend; Module.jsepAlloc = alloc; Module.jsepFree = free; Module.jsepCopy = copy; Module.jsepCopyAsync = copyAsync; Module.jsepCreateKernel = createKernel; Module.jsepReleaseKernel = releaseKernel; Module.jsepRun = run; }; ``` This simple JavaScript snippet defines all language barrier level functions that requires by JSEP to achieve implementing kernels and data transfers using JavaScript inside ONNX Runtime: - `jsepBackend`: assign the singleton object to webassembly module - `jsepAlloc` and `jsepFree`: implementation of data transfer's Alloc() and Free() - `jsepCopy`: synchronized copy ( GPU to GPU, CPU to GPU) - `jsepCopyAsync`: asynchronized copy ( GPU to CPU) - `jsepCreateKernel` and `jsepReleaseKernel`: a corresponding object that maintained in JS to match lifecycle of Kernel in ORT - `jsepRun`: OpKernel::Compute() should call into this The abstraction above allows to tie as little as possible connections and dependencies between C/C++ and TypeScript/JavaScript. **Resource Management** Lifecycle of tensor data and kernels are managed by ORT(C/C++) but the implementation are left to JavaScript. JavaScript code are responsible to implement the callbacks correctly. For WebGPU, the GPU data is managed by JavaScript using a singleton map (tensot_data_id => GPUBuffer). GPU pipeline is managed as singleton. Shaders are managed using a singletonmap (shader_key => gpu_program), while shader_key is generated by cache_key (OP specific, including attributes) and input shapes. **about data transfer** `js::DataTransfer::CopyTensor` implemented to call either synchronized or asynchronized copy callback, depending on the destination is GPU or not. Emscripten's macro `EM_ASYNC_JS` is used to wrap the async function to be called in the synchronized context. **run kernel in JS** Kernel class constructor calls once `jsepCreateKernel()` with an optional per-kernel specific serialization to pass attributes into JavaScript. `Compute()` are implemented in a way that a metadata serialization is performed in a base class and JavaScript code can access the data using the Emscripten specific builtin macro `EM_ASM_*`. **disabled features** memory pattern is force disabled, because the WebGPU data is not presented by a general memory model (a buffer can be represented by offset + size). concurrent run support is disabled. WebGPU is stateful and it also has async function call. To support concurrent run will significantly increase the complexity and we don't get any real benefit from it. **prefer channels last** JSEP prefers channels last and returns `DataLayout::NHWC` in method `GetPreferredLayout()`. This will let the graph transformers to preprocess the graph into a channels last form so that a more optimized WebGPU shader can be used. **Testing code** It's impossible to test JSEP directly because JSEP itself does not contain any kernel implementation. However, it has the kernel registration which need to work together with the corresponding JavaScript code. There are unit tests that run onnx models from JavaScript API. --------- Co-authored-by: Scott McKay <skottmckay@gmail.com>
2023-04-24 22:21:18 +00:00
const backend = (typeof backendArgs !== 'string') ? (env === 'node' ? nodejsBackends : defaultBrowserBackends) :
backendArgs.split(',');
for (const b of backend) {
if ((env !== 'node' && browserBackends.indexOf(b) === -1) || (env === 'node' && nodejsBackends.indexOf(b) === -1)) {
throw new Error(`backend ${b} is not supported in env ${env}`);
}
}
const globalEnvFlags = parseGlobalEnvFlags(args);
if (backend.includes('webnn') && !globalEnvFlags.wasm!.proxy) {
throw new Error('Backend webnn requires flag "wasm-enable-proxy" to be set to true.');
}
// Options:
// --log-verbose=<...>
// --log-info=<...>
// --log-warning=<...>
// --log-error=<...>
const logConfig = parseLogConfig(args);
globalEnvFlags.logLevel = logConfig[0]?.config.minimalSeverity;
// Option: -p, --profile
const profile = (args.profile || args.p) ? true : false;
if (profile) {
logConfig.push({category: 'Profiler.session', config: {minimalSeverity: 'verbose'}});
logConfig.push({category: 'Profiler.node', config: {minimalSeverity: 'verbose'}});
logConfig.push({category: 'Profiler.op', config: {minimalSeverity: 'verbose'}});
logConfig.push({category: 'Profiler.backend', config: {minimalSeverity: 'verbose'}});
globalEnvFlags.logLevel = 'verbose';
}
// Option: -P[=<...>], --perf[=<...>]
const perfArg = (args.perf || args.P);
const perf = perfArg ? true : false;
const times = (typeof perfArg === 'number') ? perfArg : 10;
if (debug && perf) {
throw new Error('Flag "perf" cannot be used together with flag "debug".');
}
if (perf && (mode !== 'model')) {
throw new Error('Flag "perf" can only be used in mode "model".');
}
if (perf) {
logConfig.push({category: 'TestRunner.Perf', config: {minimalSeverity: 'verbose'}});
}
[js/webgpu] support IO binding (#17480) <del> **This PR is based on a few prerequisites PRs. They are listed as below:** - #17465 - #17469 - #17470 - #17472 - #17473 - #17484 Please review the current change by only looking at commit e2e6623e673ec6de55a5c1f8edcbd3a46b535a89 and later. </del> ### Description This PR introduces WebGPU IO binding. This new feature allows onnxruntime-web users to use tensors created from GPU as model input/output so that a model inferencing can be done without unnecessary data copy between CPU and GPU for model input/output. ### Examples An E2E demo/example is being worked on. Following is some simple demo with code snippet. Let's first check today how we do: ```js // STEP.1 - create an inference session: const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'] }); // STEP.2 - create model input: (supposing myImageCpuData is a Float32Array) const feeds = { 'input_image:0': new ort.Tensor('float32', myImageCpuData, [1, 224, 224, 3]) }; // STEP.3 - run model const myResults = await mySession.run(feeds); // STEP.4 - get output data const myData = myResults['output_image:0'].data; // Float32Array ``` #### for inputs (GPU tensor): Now, with IO binding, you can create a tensor from a GPU buffer, and feed it to the model: ```js // new STEP.2.A - create model input from a GPU buffer: (supposing myInputGpuBuffer is a `GPUBuffer` object with input data) const feeds = { 'input_image:0': ort.Tensor.fromGpuBuffer(myInputGpuBuffer, { dataType: 'float32', dims: [1, 224, 224, 3] }) }; ``` ### for outputs (pre-allocated GPU tensor) you can also do that for output, **if you know the output shape**: ```js // new STEP.2.B - create model output from a GPU buffer: (supposing myOutputGpuBuffer is a pre-allocated `GPUBuffer` object) const fetches = { 'output_image:0': ort.Tensor.fromGpuBuffer(myOutputGpuBuffer, { dataType: 'float32', dims: [1, 512, 512, 3] }) }; // new STEP.3 - run model with pre-allocated output (fetches) const myResults = await mySession.run(feeds, fetches); ``` ### for outputs (specify location) if you do not know the output shape, you can specify the output location when creating the session: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: "gpu-buffer" }); ``` if the model has multiple outputs, you can specify them seperately: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: { "output_image:0": "gpu-buffer" } }); ``` now you don't need to prepare the `fetches` object and onnxruntime-web will prepare output data on the location that specified. #### read data when you get the output tensor, you can: ```js // get the gpu buffer object: const gpuBuffer = myOutputTensor.gpuBuffer; // GPUBuffer // get the CPU data asynchronizely const cpuData = await myOutputTensor.getData(); // get the CPU data asynchronizely and release the underlying GPU resources const cpuData = await myOutputTensor.getData(true); // dispose the tensor (release the underlying GPU resources). This tensor object will be invalid after dispose() is called. myOutputTensor.dispose(); ``` #### resource management JavaScript has GC so you don't need to worry about managing JavaScript objects. But there are 2 types of resources that are not managed by GC: - GPU buffer that used in tensors - Underlying ORT native resources To simplify, most of the unmanaged resources and handled inside ORT web. But there are a few resources that need users to manage: - All external GPU resources, including GPU buffers inside all tensors created by `Tensor.fromGpuBuffer()`, will not be managed by ORT. User should manage those GPU buffers themselves. - When a session is created with `preferredOutputLocation` == "gpu-buffer" specified in session options, and the corresponding output is not pre-allocated, user need to call the output tensor's `dispose()` or `getData(true)` to manually release the underlying GPU buffers. - ORT internal errors (including providing a pre-allocated output tensor with wrong type/dims) will invalidate the whole wasm memory and is not recoverable. An exception is thrown in this situation.
2023-09-29 18:24:42 +00:00
// Option: -i=<...>, --io-binding=<...>
const ioBindingArg = args['io-binding'] || args.i;
const ioBindingMode = (typeof ioBindingArg !== 'string') ? 'none' : ioBindingArg;
if (['none', 'gpu-tensor', 'gpu-location'].indexOf(ioBindingMode) === -1) {
throw new Error(`not supported io binding mode ${ioBindingMode}`);
}
// Option: -u, --optimized-model-file-path
const optimizedModelFilePath = args['optimized-model-file-path'] || args.u || undefined;
if (typeof optimizedModelFilePath !== 'undefined' && typeof optimizedModelFilePath !== 'string') {
throw new Error('Flag "optimized-model-file-path" need to be either empty or a valid file path.');
}
// Option: -o, --graph-optimization-level
const graphOptimizationLevel = args['graph-optimization-level'] || args.o || 'all';
if (typeof graphOptimizationLevel !== 'string' ||
['disabled', 'basic', 'extended', 'all'].indexOf(graphOptimizationLevel) === -1) {
throw new Error(`graph optimization level is invalid: ${graphOptimizationLevel}`);
}
// Option: -c, --file-cache
const fileCache = parseBooleanArg(args['file-cache'] || args.c, false);
const cpuOptions = parseCpuOptions(args);
const wasmOptions = parseWasmOptions(args);
const webglOptions = parseWebglOptions(args);
// Option: --no-sandbox
const noSandbox = !!args['no-sandbox'];
// parse chromium flags
let chromiumFlags = args['chromium-flags'];
if (!chromiumFlags) {
chromiumFlags = [];
} else if (typeof chromiumFlags === 'string') {
chromiumFlags = [chromiumFlags];
} else if (!Array.isArray(chromiumFlags)) {
throw new Error(`Invalid command line arg: --chromium-flags: ${chromiumFlags}`);
}
npmlog.verbose('TestRunnerCli.Init', ` Mode: ${mode}`);
npmlog.verbose('TestRunnerCli.Init', ` Env: ${env}`);
npmlog.verbose('TestRunnerCli.Init', ` Debug: ${debug}`);
npmlog.verbose('TestRunnerCli.Init', ` Backend: ${backend}`);
[js/webgpu] support IO binding (#17480) <del> **This PR is based on a few prerequisites PRs. They are listed as below:** - #17465 - #17469 - #17470 - #17472 - #17473 - #17484 Please review the current change by only looking at commit e2e6623e673ec6de55a5c1f8edcbd3a46b535a89 and later. </del> ### Description This PR introduces WebGPU IO binding. This new feature allows onnxruntime-web users to use tensors created from GPU as model input/output so that a model inferencing can be done without unnecessary data copy between CPU and GPU for model input/output. ### Examples An E2E demo/example is being worked on. Following is some simple demo with code snippet. Let's first check today how we do: ```js // STEP.1 - create an inference session: const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'] }); // STEP.2 - create model input: (supposing myImageCpuData is a Float32Array) const feeds = { 'input_image:0': new ort.Tensor('float32', myImageCpuData, [1, 224, 224, 3]) }; // STEP.3 - run model const myResults = await mySession.run(feeds); // STEP.4 - get output data const myData = myResults['output_image:0'].data; // Float32Array ``` #### for inputs (GPU tensor): Now, with IO binding, you can create a tensor from a GPU buffer, and feed it to the model: ```js // new STEP.2.A - create model input from a GPU buffer: (supposing myInputGpuBuffer is a `GPUBuffer` object with input data) const feeds = { 'input_image:0': ort.Tensor.fromGpuBuffer(myInputGpuBuffer, { dataType: 'float32', dims: [1, 224, 224, 3] }) }; ``` ### for outputs (pre-allocated GPU tensor) you can also do that for output, **if you know the output shape**: ```js // new STEP.2.B - create model output from a GPU buffer: (supposing myOutputGpuBuffer is a pre-allocated `GPUBuffer` object) const fetches = { 'output_image:0': ort.Tensor.fromGpuBuffer(myOutputGpuBuffer, { dataType: 'float32', dims: [1, 512, 512, 3] }) }; // new STEP.3 - run model with pre-allocated output (fetches) const myResults = await mySession.run(feeds, fetches); ``` ### for outputs (specify location) if you do not know the output shape, you can specify the output location when creating the session: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: "gpu-buffer" }); ``` if the model has multiple outputs, you can specify them seperately: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: { "output_image:0": "gpu-buffer" } }); ``` now you don't need to prepare the `fetches` object and onnxruntime-web will prepare output data on the location that specified. #### read data when you get the output tensor, you can: ```js // get the gpu buffer object: const gpuBuffer = myOutputTensor.gpuBuffer; // GPUBuffer // get the CPU data asynchronizely const cpuData = await myOutputTensor.getData(); // get the CPU data asynchronizely and release the underlying GPU resources const cpuData = await myOutputTensor.getData(true); // dispose the tensor (release the underlying GPU resources). This tensor object will be invalid after dispose() is called. myOutputTensor.dispose(); ``` #### resource management JavaScript has GC so you don't need to worry about managing JavaScript objects. But there are 2 types of resources that are not managed by GC: - GPU buffer that used in tensors - Underlying ORT native resources To simplify, most of the unmanaged resources and handled inside ORT web. But there are a few resources that need users to manage: - All external GPU resources, including GPU buffers inside all tensors created by `Tensor.fromGpuBuffer()`, will not be managed by ORT. User should manage those GPU buffers themselves. - When a session is created with `preferredOutputLocation` == "gpu-buffer" specified in session options, and the corresponding output is not pre-allocated, user need to call the output tensor's `dispose()` or `getData(true)` to manually release the underlying GPU buffers. - ORT internal errors (including providing a pre-allocated output tensor with wrong type/dims) will invalidate the whole wasm memory and is not recoverable. An exception is thrown in this situation.
2023-09-29 18:24:42 +00:00
npmlog.verbose('TestRunnerCli.Init', ` IO Binding Mode: ${ioBindingMode}`);
npmlog.verbose('TestRunnerCli.Init', 'Parsing commandline arguments... DONE');
return {
debug,
mode: mode as TestRunnerCliArgs['mode'],
param: args._.length > 1 ? args._[1] : undefined,
backends: backend as TestRunnerCliArgs['backends'],
bundleMode: perf ? 'perf' : 'dev',
env: env as TestRunnerCliArgs['env'],
logConfig,
profile,
times: perf ? times : undefined,
[js/webgpu] support IO binding (#17480) <del> **This PR is based on a few prerequisites PRs. They are listed as below:** - #17465 - #17469 - #17470 - #17472 - #17473 - #17484 Please review the current change by only looking at commit e2e6623e673ec6de55a5c1f8edcbd3a46b535a89 and later. </del> ### Description This PR introduces WebGPU IO binding. This new feature allows onnxruntime-web users to use tensors created from GPU as model input/output so that a model inferencing can be done without unnecessary data copy between CPU and GPU for model input/output. ### Examples An E2E demo/example is being worked on. Following is some simple demo with code snippet. Let's first check today how we do: ```js // STEP.1 - create an inference session: const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'] }); // STEP.2 - create model input: (supposing myImageCpuData is a Float32Array) const feeds = { 'input_image:0': new ort.Tensor('float32', myImageCpuData, [1, 224, 224, 3]) }; // STEP.3 - run model const myResults = await mySession.run(feeds); // STEP.4 - get output data const myData = myResults['output_image:0'].data; // Float32Array ``` #### for inputs (GPU tensor): Now, with IO binding, you can create a tensor from a GPU buffer, and feed it to the model: ```js // new STEP.2.A - create model input from a GPU buffer: (supposing myInputGpuBuffer is a `GPUBuffer` object with input data) const feeds = { 'input_image:0': ort.Tensor.fromGpuBuffer(myInputGpuBuffer, { dataType: 'float32', dims: [1, 224, 224, 3] }) }; ``` ### for outputs (pre-allocated GPU tensor) you can also do that for output, **if you know the output shape**: ```js // new STEP.2.B - create model output from a GPU buffer: (supposing myOutputGpuBuffer is a pre-allocated `GPUBuffer` object) const fetches = { 'output_image:0': ort.Tensor.fromGpuBuffer(myOutputGpuBuffer, { dataType: 'float32', dims: [1, 512, 512, 3] }) }; // new STEP.3 - run model with pre-allocated output (fetches) const myResults = await mySession.run(feeds, fetches); ``` ### for outputs (specify location) if you do not know the output shape, you can specify the output location when creating the session: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: "gpu-buffer" }); ``` if the model has multiple outputs, you can specify them seperately: ```js // new STEP.1 - create an inference session with an option "preferredOutputLocation": const mySession = await ort.InferenceSession.create('./my_model.onnx', { executionProviders: ['webgpu'], preferredOutputLocation: { "output_image:0": "gpu-buffer" } }); ``` now you don't need to prepare the `fetches` object and onnxruntime-web will prepare output data on the location that specified. #### read data when you get the output tensor, you can: ```js // get the gpu buffer object: const gpuBuffer = myOutputTensor.gpuBuffer; // GPUBuffer // get the CPU data asynchronizely const cpuData = await myOutputTensor.getData(); // get the CPU data asynchronizely and release the underlying GPU resources const cpuData = await myOutputTensor.getData(true); // dispose the tensor (release the underlying GPU resources). This tensor object will be invalid after dispose() is called. myOutputTensor.dispose(); ``` #### resource management JavaScript has GC so you don't need to worry about managing JavaScript objects. But there are 2 types of resources that are not managed by GC: - GPU buffer that used in tensors - Underlying ORT native resources To simplify, most of the unmanaged resources and handled inside ORT web. But there are a few resources that need users to manage: - All external GPU resources, including GPU buffers inside all tensors created by `Tensor.fromGpuBuffer()`, will not be managed by ORT. User should manage those GPU buffers themselves. - When a session is created with `preferredOutputLocation` == "gpu-buffer" specified in session options, and the corresponding output is not pre-allocated, user need to call the output tensor's `dispose()` or `getData(true)` to manually release the underlying GPU buffers. - ORT internal errors (including providing a pre-allocated output tensor with wrong type/dims) will invalidate the whole wasm memory and is not recoverable. An exception is thrown in this situation.
2023-09-29 18:24:42 +00:00
ioBindingMode: ioBindingMode as TestRunnerCliArgs['ioBindingMode'],
optimizedModelFilePath,
graphOptimizationLevel: graphOptimizationLevel as TestRunnerCliArgs['graphOptimizationLevel'],
fileCache,
cpuOptions,
webglOptions,
wasmOptions,
globalEnvFlags,
noSandbox,
chromiumFlags
};
}