### Description
Previously, shape and strides were added unconditionally even they are
not used. This PR fixes this issue and only adds shape and strides when
they are required.
It's useful when some shapes are not used as uniform if the program
depends on type instead of rank.
### Description
<!-- Describe your changes. -->
Check whether the min/max inputs are provided and use default values if not provided.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Description
This PR includes a change that inspired from #18452 to resolve a
requirement: a shader may depend on an instance of `IndicesHelper` to
generate WGSL code snippet, but the IndicesHelper instance is not
necessarily an input/output of the program. So the existing
`declareVariables()` function does not work with this scenario.
In order to support this requirement, I added this "use" function to
`interface ShaderHelper`, which takes a helper-like object as parameter.
The hidden implementation `ShaderHelperImpl` class will iterate the
helpers and call `impl()` for each.
@axinging @qjia7
This change refactored matmul/conv related programs to support shape
uniforms. Currently only matmul shape uniforms are fully enabled.
TODOs: add input dependencies for conv related programs, turn clipMax
and clipMin to uniforms.
### Description
Support uniforms in Slice op
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Improve ferformance
### Description
Added Uniform support to binary ops
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
To improve performance
### Description
<!-- Describe your changes. -->
Currently, the uniform support has bugs when dims rank is larger than 4.
See https://github.com/microsoft/onnxruntime/issues/17860 item 1.
So this PR only enables shapes uniforms when shape rank is <= 4 for
transpose. Otherwise, below compilation errors are thrown:
```
1 error(s) generated while compiling the shader:
:3:50 error: uniform storage requires that array elements are aligned to 16 bytes, but array element of type 'u32' has a stride of 4 bytes. Consider using a vector or struct as the element type instead.
struct Uniforms { output_size:u32, a_shape:array<u32, 5>, a_strides:array<u32, 5>, output_shape:array<u32, 5>, output_strides:array<u32, 5> };
^^^^^^^^^^^^^
:3:7 note: see layout of struct:
/* align(4) size(84) */ struct Uniforms {
/* offset( 0) align(4) size( 4) */ output_size : u32;
/* offset( 4) align(4) size(20) */ a_shape : array<u32, 5>;
/* offset(24) align(4) size(20) */ a_strides : array<u32, 5>;
/* offset(44) align(4) size(20) */ output_shape : array<u32, 5>;
/* offset(64) align(4) size(20) */ output_strides : array<u32, 5>;
/* */ };
struct Uniforms { output_size:u32, a_shape:array<u32, 5>, a_strides:array<u32, 5>, output_shape:array<u32, 5>, output_strides:array<u32, 5> };
^^^^^^
:4:42 note: 'Uniforms' used in address space 'uniform' here
@group(0) @binding(2) var<uniform> uniforms: Uniforms;
^^^^^^^^
```
### Description
support using uniform buffer.
This PR allows to use uniform buffer in shader program, so that some
runtime information (eg. input/output shape) is no longer need to be
hardcoded into shader code.
There are 2 commits in this PR:
-
[667f31c](667f31c83d):
framework changes to support uniform buffer, as well as updates in
program manager, gpu data manager and indices helper.
-
[09e1d2a](09e1d2ad1d):
an example change for operator `Transpose` to use input's rank-only
instead of dims as shader key. With this change, model mobilenetv2-12
shader compile times dropped from 71 to 52.
Supported type: float. int32_t, uint32_t, bool.
Case where_broadcast.jsonc is not enabled due to
https://github.com/microsoft/onnxruntime/issues/17405.
### Description
<!-- Describe your changes. -->
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
---------
Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>
### Description
This PR optimizes the gather op, which is improved ~6ms in segment
anything model in ADL.
The problem in original algorithm is that it includes a for loop to
calculate a block size of data. However, the block size may be very
large, like `65536`. In GPU shader, we should try to avoid large loop in
shader and try to use more threads to do it parallelly.
Before:
```
[profiling] kernel "41771992|[Gather] 41771992" input[0]: [4,65536] | float32, input[1]: [1] | int64, output[0]: [1,65536] | float32, execution time: 6886207 ns
```
After:
```
[profiling] kernel "41771992|[Gather] 41771992" input[0]: [4,65536] | float32, input[1]: [1] | int64, output[0]: [1,65536] | float32, execution time: 11719 ns
### Description
<!-- Describe your changes. -->
In previous implementation, there are two loops to iterate H * W
elements to calculate the `mean` and `squaredNorm` value in one thread,
meanwhile it outputs H * W elements in one thread. That results it's
very very slow when H * W is a large value. And usually, H * W does be a
large value in a model. For example, in the `candy-8` model, the shapes
of [H, W] are [224,224], [112,112], [56,56] for `InstanceNormalization`
op. And in my ADL, `[1,224,224,32]` consumes 17 ms. See below:
```
[profiling] kernel "23848328|[InstanceNormalization] 23848328" input[0]: [1,224,224,32] | float32, input[1]: [32] | float32, input[2]: [32] | float32, output[0]: [1,224,224,32] | float32, execution time: 17007914 ns
```
In this PR, it uses workgroup memory to optimize the original algorithm.
The advantage is that it can parallelly utilize the 64 (workgroupSize)
threads in one workgroup to calculate `mean` and `squaredNorm` value.
Meanwhile, it only outputs `H * W / workgroupSize` outputs for one
thread, which greatly reduces the overhead for one thread. With this
optimization, `[1,224,224,32]` becomes 3 ms and the main overhead is the
extra two `transpose`. The `createInstanceNormProgramInfo` only needs
`0.64` ms. See below:
```
[profiling] kernel "23003600|[InstanceNormalization] 23003600" input[0]: [1,224,224,32] | float32, output[0]: [1,32,224,224] | float32, execution time: 1543792 ns
program-manager.ts:115
[profiling] kernel "23003600|[InstanceNormalization] 23003600" input[0]: [1,32,224,224] | float32, input[1]: [32] | float32, input[2]: [32] | float32, output[0]: [1,32,224,224] | float32, execution time: 642652 ns
program-manager.ts:115
[profiling] kernel "23003600|[InstanceNormalization] 23003600" input[0]: [1,32,224,224] | float32, output[0]: [1,224,224,32] | float32, execution time: 991608 ns
```
This PR currently only applies the new algorithm to NCHW format. For
NHWC format, one way is to transpose the input so that it can use the
new algorithm. But the disadvantage is that 2 extra transpose are added.
@dakenf also gives another way to optimize NHWC. Details see
[here](d45a96616d/js/web/lib/wasm/jsep/webgpu/ops/instance-norm.ts).
I checked @dakenf's method. The perf is similar with transpose +
optimized NCHW. But on different GPUs, one is a little better than
another or vice versa. So I prefer this PR only does the NCHW part.
@dakenf can submit his optimization on NHWC.
### Description
Changes in this PR:
1) use the optimized version `makeMatMulPacked[Vec4]Source` to support
matmul.
2) enable the conv2dByMatMul path.
3) support broadcast
4) use IndicesHelper.
MatMul with M = 512, K = 512, N = 512 becomes 2ms from 15ms when
enabling profilingMode on my ADL.
### Description
This PR adds kernel implementation for operator "Not" and "Equal". Also
removed download cache in gpu data manager.
**Why removing download cache**
The following test case failed. ("Or" is on CPU, "Greater" and "Equal"
are on JSEP)
after debugging, I found that both "Equal" and "Greater" are using the
same output GPU Data ID. This is because when ORT executes the graph, it
first run "Equal", allowing its shader to write into GPU Data ID 2; then
a Gpu2Cpu copy for it is issued (because currently "Or" is on CPU EP);
at this point, ORT thinks GPU Data ID=2 is free to use; so it reuse it
as output for "Greater". This means there is no allocation for output of
"Greater" kernel, and both kernel writes to GPU Data ID=2.
For gpu data manager, there will be 2 downloads from the same GPU
buffer. Previously I think this is a waste of resource so I cached the
data. But now it shoes that we need to perform 2 downloads because the
GPU data is already different. The download data cache should be
removed.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Description
This PR introduces the new incides helper.
IndicesHelper is a helper class for generating WGSL code for
manipulating indices and data for a shader's input or output.
This class is designed to offer a unified way to generate WGSL code for
manipulating indices and data for a shader's input or output. The
following is a list of terminologies used in this class:
- `offset`: a uint32 value representing the offset of an element in the
data buffer.
- `indices`: an abstraction of a multi-dimensional array's indices
representing the data's index on each dimension.
- `value`: a value of a data element.
Users are expected to create an instance of this class for each shader's
input or output, and use the instance to generate WGSL code for
manipulating indices and data. The following 2 exported functions are
for users to call to create an instance of an indices helper:
- `inputVariable()`: create an indices helper instance for an input.
- `outputVariable()`: create an indices helper instance for an output.
An indices helper instance contains helper functions for the following
operations:
- access readonly basic information, including: `name`(the name of the
input or output), `usage`(whether it's an input or an output) and
`shape`(the passed in shape).
- `type`: access readonly type information, including: `indices`(the
type of indices), `value`(the type of value at runtime), `storage`(the
type of value at storage) and `tensor`(the tensor type as represented in
TensorView).
- generate WGSL code for getting indices from offset. Use
`offsetToIndices()` for WGSL code snippet to calculate incides from
offset, and use `indicesToOffset()` for WGSL code snippet to calculate
offset from indices.
- to manipulate an instance of indices, use `setIndices()` and
`getIndices()` to set and get the indices on an indices variable.
- to manipulate data, use `set()`/`get()` to access data at the given
indices from parameter list, use `setByIndices()`/`getByIndices()` to
access data at the given indices from an indices variable, and use
`setByOffset()`/`getByOffset()` to access data at the given offset.
- `impl`: get WGSL code of function implementation for the util
functions mentioned above.
This change applies the usage of new IndicesHelper through the code, but
not necessary for all code.
### 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>
