### 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
Two contrib kernels that supposed to speed-up StableDiffusion according
to this doc
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/transformers/models/stable_diffusion/README.md
However, there is no noticable effect in speed or memory consumption. So
i guess the only way to make it faster is to implement
MultiHeadAttention but i'm not capable of doing that right now. So i'll
focus on existing PRs and finding the JSEP kernel that produces
incorrect results. It should be one of the old ones (i suspect Conv or
ConvTranspose), as SD was not generating images correctly on webgpu
since i started working on it. I hoped someone else would fix that by
the time i finish with kernels/optimizations 😅
---------
Co-authored-by: Guenther Schmuelling <guschmue@microsoft.com>
Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>
The patch also introduces the method which copies
data from GPU to CPU synchronously.
### 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. -->
### Description
Another three ops for fp16
---------
Co-authored-by: Guenther Schmuelling <guschmue@microsoft.com>
Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>
<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.
### Description
Add ConvTranspose implementation using MatMul to increase perf.
### 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 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
1. For binary ops, the components is always 4. So the dispatchGroup
should be : `{x: Math.ceil(outputSize / 64 /* workgroup size */ / 4 /*
component size */)}` instead of `{x: Math.ceil(outputSize / 64 /*
workgroup size */ / (vectorize ? 4 : 1) /* vec size */)}`.
2. If any of a or b only has one element, we still can use the vectorize
path since the same value will be broadcasted.
### 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
<!-- 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. -->
### Description
Added Einsum operator support to JSEP.
### 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
<!-- 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. -->
### Description
<!-- Describe your changes. -->
For the conv2dByMatMul path, the simulated matmul output shape is the
reshape of the original conv2d. So we should pass this information to
`createMatmulProgramInfo` so that it can process it correctly.
### 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
<!-- Describe your changes. -->
With the label, it's more easier to identify which op causes the error.
Without the label, the error message is like below:
```
Tint WGSL reader failure: :12:5 error: return statement type must match its function return type, returned 'vec4<f32>', expected 'f32'
return W[i2o_W(indices)];
^^^^^^
- While validating [ShaderModuleDescriptor]
- While calling [Device].CreateShaderModule([ShaderModuleDescriptor]).
```
With the label, the error message is like below:
```
Tint WGSL reader failure: :12:5 error: return statement type must match its function return type, returned 'vec4<f32>', expected 'f32'
return W[i2o_W(indices)];
^^^^^^
- While validating [ShaderModuleDescriptor "ConvTranspose2D"]
- While calling [Device].CreateShaderModule([ShaderModuleDescriptor "ConvTranspose2D"]).
```
### Motivation and Context
This change is mainly for debugging. With this change, we can easily
know that `ConvTranspose2D`'s shader has problem from above message.
### 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. -->
### Description
This PR contains changes to support error pop and kernel name.
- Add a function `JsepGetNodeName` to allow reading kernel name from JS
to C++
- When in debug mode ( `env.debug = true;` ) or in profiling mode (
`env.webgpu.profilingMode = 'default';` ), kernel name will be read from
ORT; otherwise use the kernel pointer ( a number ) as kernel name to
save calls from JS to C++.
- When in debug mode, WebGPU validation errors will be recorded and if
any error occurs, `inferenceSession.run()` will fail (Promise get
rejected). Behavior when not in debug mode is not changed. This is
because recording errors are not zero-overhead, and GPU validation
errors should occur consistently in and not in debug mode.
- Add `jsepOnRunStart()` and `jsepOnRunEnd()` hook to:
- allow implementation of the features mentioned above.
- pass session ID to backend.
### Description
Fix JSEP ConvTranspose shader code errors.
### 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
Enable typed binary and support int32 type for binary.
Co-authored-by: Xing Xu <xing.xu@intel.com>
---------
Co-authored-by: Xing Xu <xing.xu@intel.com>
### Description
Add SkipLayerNormalization operator to JSEP.
### 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
Fix a typo. LayerNormalization takes 2 or 3 inputs. The third input,
bias, is optional.
### 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
Since WebGPU supports only float32 and int32, having Gather, Reshape,
Shape, Squeeze and Unsqueeze ops with other data types create additional
MemCpy ops and slow down the overall execution as all other OPs with
other tensor types will be done on CPU.
Before this patch SD Unet had these numbers:
Node(s) placed on [CPUExecutionProvider]. Number of nodes: 1141
Node(s) placed on [JsExecutionProvider]. Number of nodes: 4025
memcpy tokens: 2001
After patch:
Node(s) placed on [CPUExecutionProvider]. Number of nodes: 1735
Node(s) placed on [JsExecutionProvider]. Number of nodes: 2243
memcpu tokens: 813
It also gives more than 5X performance benefit. From 12sec for one Unet
step to 2.2sec on RTX 3090 Ti, so we are almost getting to native
performance.
UPD: with latest changes from main branch and multi-threading it went
down to 1.6sec. Will try re-exporting my model to onnx with maximum
optimizations, like using MultiHeadAttention to decrease node count.
Maybe after implementing that it can go in less than 1 sec
### 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. -->
### 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
Added two kernels for Layer and Instance norm
Also added maximum limits for `maxBufferSize` when requesting GPU device
as by default it's limited to 256mb and it fails allocating 600mb buffer
while running fp32 StableDiffusion weights.
### Motivation and Context
These two are used in StableDiffusion and many other networks
### Description
<!-- Describe your changes. -->
This PR makes sure that only storage buffers are reused. Previously, the
query buffer might also get from the freeBuffers list if there is a
matching size in it. But they are different usage, which results errors.
Fixed ArgMin and ArgMax and refactored using functionality from Reduce
operator code.
### Description
Removed code/functionality duplication and fixed some issue.
### 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
Make CacheHint mechanism, which is designed to avoid running the same
test multiple times saving the result mapped against a key, working by
adding input dims.
### 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
Added Gather op that works with both i32 and i64 indices, assuming that
values fall into i32 limit. The assumption is safe because it's not
possible to allocate more than 2gb buffer for inputs.
It treats all data from input tensor as u32, copying 1 or 2 elements for
i64, u64 and double.
---------
Co-authored-by: Guenther Schmuelling <guschmue@microsoft.com>
### Description
If Expand inputs has rank < 2, `inputIndicesHelper` and
`outputIndicesHelper` create indices as u32 instead if array<u32> and
`calculateInputIndex` throws an error
### Motivation and Context
I've encountered this error while making StableDiffusion work with JSEP
argmax and argmin are similar to reduce. Eventually we need to add
optimized flavors of the shader.
softmax is optimized but only works on the last axis for now which
should be the common use case.
todo: enable more ut for argmax/argmin
