### 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. -->
1. Update CK to its latest develop branch
2. `-mllvm -amdgpu-early-inline-all=true` is critical to CK's
performance, ensure it is properly configured.
- The flags are propagated from target `hip-lang::device`'s
`INTERFACE_COMPILE_OPTIONS`, we must not manually add the flags.
- Instead, we must ensure this target is properly configured by checking
_CMAKE_HIP_DEVICE_RUNTIME_TARGET is set.
TL,DR
`hip-lang::device` sometime will be not be properly configured if our
`CMAKE_PREFIX_PATH` is not configured carefully. In the CI docker, the
configuration is in good state, but on dev machine it is not, which then
silently result poor performance for kernels. We fixed it in this PR and
add a guard to avoid unsuccessful future editing and to prevent
convoluted debugging process.
`_CMAKE_HIP_DEVICE_RUNTIME_TARGET ` is shared in
`/opt/rocm/lib/cmake/hip-lang/hip-lang-config.cmake` and it is internal
to
[CMake](https://gitlab.kitware.com/cmake/cmake/-/merge_requests/6121/diffs),
the variable name will not be changed in the foreseeable future.
**Description**: Subgraph-level recompute
This PR adds an optional capability trading additional re-computation
for better memory efficiency. Specifically, a pre-defined operator list
used to iterate the Graph to find some subgraphs for recompute, to
reduce some stashed activations whose lifetime across forward and
backward pass.
When training with ORTModule, by default, the graph transformer will
scan the execution graph to find all eligible subgraph to recompute,
along with sizes that can save. An example looks like below.
If we want to enable some of them to recompute, we can define env
variable this way:
`export
ORTMODULE_ENABLE_MEMORY_ALLEVIATION="Mul+FusedMatMul+Cast+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Add+BiasSoftmaxDropout+Cast+:1:-1,BiasGelu+:1:-1,BitmaskDropout+Cast+:1:-1,FusedMatMul+:1:-1,Cast+:1:-1,Mul+Add+:1:-1,Mul+Sub+:1:-1"`
```
[1,0]<stderr>:2,022-10-12 14:47:39.302,954,530 [W:onnxruntime:, memory_alleviation.cc:595 PrintSummary]
[1,0]<stderr>:MemoryAlleviation Summary:
[1,0]<stderr>: User config:
[1,0]<stderr>: Mul+FusedMatMul+Cast+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Add+BiasSoftmaxDropout+Cast+:1,BiasGelu+:1,BitmaskDropout+Cast+:1,FusedMatMul+:1,Cast+:1,Mul+Add+:1,Mul+Sub+:1
[1,0]<stderr>: =================================
[1,0]<stderr>: Subgraph: BitmaskDropout+
[1,0]<stderr>: AlleviationType: Disabled
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x 1,024 x Frequency:1
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: BiasGelu+
[1,0]<stderr>: AlleviationType: Recompute
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x input_ids_dim1 x 4,096 x Frequency:24
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: Reshape[1,0]<stderr>:+
[1,0]<stderr>: AlleviationType: Disabled
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:labels_dim0 x Frequency:1
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: Unsqueeze+Unsqueeze+Cast+Sub+Mul+Mul+FusedMatMul+Cast+Add+BiasSoftmaxDropout+Cast+
[1,0]<stderr>: AlleviationType: Disabled
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x 16 x input_ids_dim1 x input_ids_dim1 x Frequency:23
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: Mul+FusedMatMul+Cast+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Add+BiasSoftmaxDropout+Cast+
[1,0]<stderr>: AlleviationType: Recompute
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x 16 x input_ids_dim1 x input_ids_dim1 x Frequency:1
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: Mul+Add+
[1,0]<stderr>: AlleviationType: Recompute
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x 16 x input_ids_dim1 x 1 x Frequency:24
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: FusedMatMul+Cast+Add+Reshape+Cast+
[1,0]<stderr>: AlleviationType: Disabled
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x 16 x input_ids_dim1 x 2 x 4 x Frequency:24
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: Mul+Sub+
[1,0]<stderr>: AlleviationType: Recompute
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x 16 x input_ids_dim1 x 1 x Frequency:24
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: Cast+
[1,0]<stderr>: AlleviationType: Recompute
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:1,024 x 1,024 x Frequency:97
[1,0]<stderr>: PatternShape:3 x 1,024 x Frequency:1
[1,0]<stderr>: PatternShape:8 x 64 x Frequency:24
[1,0]<stderr>: PatternShape:1,024 x 4,096 x Frequency:24
[1,0]<stderr>: PatternShape:4,096 x Frequency:24
[1,0]<stderr>: PatternShape:4,096 x 1,024 x Frequency:24
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: Subgraph: FusedMatMul+
[1,0]<stderr>: AlleviationType: Recompute
[1,0]<stderr>: Patterns:
[1,0]<stderr>: PatternShape:input_ids_dim0 x input_ids_dim1 x 4,096 x Frequency:24
[1,0]<stderr>: --------------------------------
[1,0]<stderr>: =================================
```
"Type config:" whether recompute is enabled by users. 0 - disable, 1-
enable.
"Subgraph" means what kind of subgraph will be recomputed, in this case,
it is a single node "Gelu", and it will be "Recompute".
"Shape && Frequency" means, for this recompute, one tensor of size
(batch size, 500) will be saved because it will be recomputed.
**Baseline**
On a 1P model (DEBERTA V2), sequence length 256, training with 16 A100
GPUs. With latest main branch, we can run batch size 16, and the maximum
batch size < 32. So 16 is usually chosen by data scientists. 65% of 40GB
memory is used during training. The SamplesPerSec=479.2543353561354.
**With this PR**
Gelu is recomputed for saving memory peak, batch size 32 can be run. The
97% of 40GB A100 is used, the SamplesPerSec=562.041593991271 (**1.17X**
of baseline).
**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.
The oneDNN 2.7.1 release includes multiple functional and performance
improvements.
Signed-off-by: George Nash <george.nash@intel.com>
### Description
Update the oneDNN library from 2.7.0 to 2.7.1. This contains multiple
functional and performance improvements.
### Motivation and Context
This is a minor point release from the oneDNN library that gives
performance and functional fixes that were found in the oneDNN 2.7
library shortly after release.
Signed-off-by: George Nash <george.nash@intel.com>
### Description
It missed a space there.
### Motivation and Context
Right now the pipeline is failing because GSL was just converted from a
submodule to a cmake external project.
This PR enables ORT to execute graphs captured by TorchDynamo. Major compilation code is in `OrtBackend.compile` in ort_backend.py. `register_backend.py` is for plugging `OrtBackend` into TorchDynamo as a compiler.
Current NoOpElimination can support only Add node. This PR adds support
for: x-0, x*1, 1*x and x/1 besides x+0 and 0+x.
With this PR, all Div(x,1) and their gradients (also Div(x,1)) in
Huggingface's diffusers model can be removed, which takes ~1% of compute
time in total previously.
### Description
Fix variable shadowing in the DML EP's implementation of EinSum
### Motivation and Context
An SDL bug was opened because of shadowing of the variable `i` in a
nested loop of the EinSum operator.
### Description
Enables the DML provider in the op tests to allow for better CI
coverage.
### Motivation and Context
Some of the CI tests for DML were actually running on the CPU because
there was no default DML provider, so it was returning a `nullptr`. This
should add better coverage, and it already uncovered some failures and
asserts hitting in a few tests, which need to be investigated
separately.
- Reverts change to CustomOpApi::GetTensorData introduced by commit 5dae0c477d,
which causes infinite recursion.
- Moves EndsProfilingAllocated to non-const session implementation
(C++ API header).
### Description
<!-- Describe your changes. -->
Fix the input shape of FastGelu
Minor improvement for the documentation of kernel explorer
### 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. -->
1. Move DML packaging pipelines to aiinfra-dml-winbuild machine pool
2. Delete
tools/ci_build/github/azure-pipelines/templates/windowsai-nuget-build.yml
because the pipeline has been migrated to Onebranch. I monitored it for
months, it worked well.
### Description
<!-- Describe your changes. -->
Fix the input shape of FastGelu
### 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. -->
Some models have QuickGelu(x)=x*sigmoid(1.702x), which has 3 Ops for
forward and 5 Ops for backward. The PR is to fuse this to a single Op
named QuickGelu and its gradient QuickGeluGrad.
For CUDA, tested in V100 using input tensor with shape [64,128,2048] and
float16 type:
Before, FW takes 335us, BW takes 614us
After, FW takes 115us, BW takes 139us, which is much faster.
For CPU kernel, using same shape and float type:
Before, FW takes 10us, BW takes 49us
Mul: 3480[µs]
Sigmoid: 1996[µs]
Mul: 4789[µs]
Mul: 4642[µs]
Mul: 4195[µs]
SigmoidGrad: 18328[µs]
Mul: 2988[µs]
Sum: 18576[µs]
After, FW takes 4us, BW takes 5us, which is also much faster.
QuickGelu: 3939[µs]
QuickGeluGrad: 5089[µs]
Co-authored-by: Vincent Wang <weicwang@microsoft.com>
### Description
support building xnnpack for IOS
### 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
Round 2 of fixing C4244
### 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. -->
Fix document generation CI. It's not currently updating the docs as
we're skipping the tests, which is the invocation of build.py that would
have generated the documentation.
Setup specific task to generate documentation for greater clarity.
### 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. -->
Operator kernel documentation is not getting updated and is now out of
date.
### Description
Enables all datatypes supported for DML for `Abs` and `Sign`.
### Motivation and Context
`Abs` and `Sign` haven't been updated since DML started to support all
datatypes for them. These ops are used in some transformer models and
were forcing unnecessary copies between the CPU and the GPU.
### Description
First round of fixes for C4244 error.
### 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
DML EP was a special EP w.r.t. capability fusion. It used to fuse a
capability outside the IExecutionProvider::Compile() call. But after
recent re-architecture #13131, it is no longer a special case.
### Motivation and Context
Why is this change required? What problem does it solve?
To make DML EP consistent with the ORT design.
- If it fixes an open issue, please link to the issue here. N/A
Co-authored-by: Sumit Agarwal <sumitagarwal@microsoft.com>
### Description
Fix logging for affinity failures on Linux.
Make `GetCpuCores()` consistently return the number of physical cores.
Use `CpuInfo` library to correctly set affinities for Linux where
supported.
Make windows generate affinity masks as ordinals and convert them to
masks at the setting site.
Allow setting multiple logical processors affinity masks per thread.
We continue to set all logical processors as thread affinity per
physical core.
### Motivation and Context
Error logging on Linux uses `pthread_self()` which does not return
Thread ID.
Fix default affinity mask generation on Windows. The following are the
issues with Windows:
- `GetThreadAffinityMasks()` returns bitmasks, but on other platforms it
returns ordinals generated for the hardware concurrency
- The maximum number of processors supported for requires a mask of
64-bits, but `size_t` type used is not always 64-bit
- The masks returned per physical core may have multiple bits set,
because the mask applies to several logical cores hosted by the physical
core. In the past, customers complained that their threads jump from one
core to another which adversely affects performance. The decision was
made to stay this way.
- 64-bit masks do not allow for logical processors with IDs that are
outside of 0-63 range.
### Description
Updates naming scheme for docker images built by the EP Perf pipeline.
Specifically, the docker image name is no longer based on the branch
name.
### Motivation and Context
The docker image name used by EP Perf pipeline is built from the branch
name. This makes the pipeline fail for branches with uppercase letters
because docker image names can only contain lower-case letters.
### Description
Implements a Python script for quick analysis of a generated JSON
profile from ORT.
### Motivation and Context
This PR implements a script that lists kernels that take up the most
time in a JSON profile, from both the CPU and GPU points-of-view. The
script also supports various options for CSV output, grouping of kernels
wrt shape of input tensors and wrt kernel dimensions.
Co-authored-by: Abhishek Udupa <abhishek.udupa@microsoft.com>
### Description
<!-- Describe your changes. -->
The subgraph below meet the SkipLayerNorm fusion pattern, but the fusion
rules also required every input dimension has a certain value. So the
subgraph below cannot fused to SkipLayerNorm.
subgraph we want to fuse
fusion pattern 3
[Sub1] [Sub2]
\ /
\ /
\ /
Add1
|
LayerNormalization
This change allow inputs of FirstAdd operator has dimension which only
has dim_param.
### 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: peixuanzuo <peixuanzuo@linmif39a000004.zvflicr54joexhdgnhvmxrxygg.phxx.internal.cloudapp.net>
### Description
This adds the "NHCW" format support for einsum MatMul. The logic is
basically a merge of the existing Transpose and MatMul Einsum
implementations.
### Motivation and Context
Some transformer models that I'm tracking use Einsum quite often during
a single inference, and about half of those were "NHCW" MatMul Einsums.
Supporting them will reduce the number of copies to the CPU.
### Description
Register all datatypes for DML's `Where` operator since DML now supports
everything.
### Motivation and Context
Some transformer models use the `Where` operator on int64 data, but
since DML wasn't supporting it, it needed to fall back to the CPU.
**Description**: Describe your changes.
fuse MatMul + FastGelu -> GemmFastGelu
prepare for AMD optimized fused operator GemmFastGelu
usage:
python benchmark.py -g -m bert-base-cased --sequence_length 384
--batch_sizes 128 --provider=rocm -p fp16 --disable_embed_layer_norm
--enable_gemm_fast_gelu
**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.
