* Ported changes / bug fixes from torch/ort.
* Fixed formatting
* Renamed function
* Renamed module_ to module.
* Revert "Renamed module_ to module."
This reverts commit b17fc114b3db20d174283811d90592b5b8154c19.
* Include pybind common header to fix linker errors on windows debug.
* Fix to generation of > 1 custom op.
Co-authored-by: Ashwin Hari <ashari@microsoft.com>
* Remove APIs unavailable in Store in #8349, #8178, #8065
* Add UWP stubs of C runtime functions
* Remove UWP incompatible tests from UWP build
* Remove incompatible tests from Store
* Use UWP stubs in store only
* Skip partition check outside of Windows
* Remove unused WRL include
* Workaround Windows header not including what it uses
* Fix precompiled header name clash
* Workaround SDK bugs
* DXCore workaround in Win7
* Fix warning
* Fix more warnings
* Bump WinML to target Windows 8
* Fix more warnings
* Remove unnecessary workarounds
* Remove Desktop only APIs from DML adapter
* Implement Gemm op for DNNL execution provider
Signed-off-by: George Nash <george.nash@intel.com>
* Remove KernelRegistry and Gemm op for dnnl ep
The KernelRegistry for the dnnl execution provider only
registered a Gemm op that as best we can tell was never
actually used and also was not using the dnnl library.
We have implemented a Gemm op in the DNNL execution provider
subgraph code and thus are removing the unused Gemm op that
was in the dnnl KernelRegistry.
Signed-off-by: George Nash <george.nash@intel.com>
* Fix duplicated output and kernelshape inference
fix getcapability to make sure subgraph outputs do not have duplicates
fix kernelshape inference in pool
Signed-off-by: Wang <zhaoyang.wang@intel.com>
* Removed most dnnl specialized ifdefs from gradient_ops_test code
Re-enable GlobalAveragePoolGrad test for dnnl ep
The bugs that were exposed by the GlobalAveragePoolGrad test have
been fixed and this test no longer needs to be disabled for DNNL.
Removed the ReluGradDnnl test. We are getting the testing from the
already existing ReluGrad test.
MaxPoolGrad test no longer has specialized execution provider
enabling for DNNL execution provider. It will now run without
the extra enabling.
ConvGrad is the only test that still has dnnl specialized ifdefs
However, the ConvGrad code was not being executed by the code
unless it was listed first in the list of execution providers.
Signed-off-by: George Nash <george.nash@intel.com>
* Fix transpose issue on Gemm
On transposing square matrices, getmemoryandreshape will fail to reshape
fix by adding a bool
Signed-off-by: Wang <zhaoyang.wang@intel.com>
* Save memory space by reusing internal tensor for output
The intermediat matmul output tensor can be used as the output
tensor for the binary calculation.
Remove the unused IsAttributeSupported from the
DnnlGemmNodeCapability class since we now support all of the
Gemm attributes in our implementation.
Signed-off-by: George Nash <george.nash@intel.com>
Co-authored-by: Wang <zhaoyang.wang@intel.com>
* adding support for tracing to sqldb instead of files
* use compiled statements
* script to pull tensors from db
* link sqlite3
* remove node info redundant with onnx graph
* addressing PR comments
* address PR comments and include program counter
* third party notice
* use find_pacakge
* add to cgmanifests.json
* address thread safety and add pid suffix
* build fi
* python script to select on devicetype
* remove unpopulated and redundant Shape and Type fields
* comment
* comment
* PR comments
* add graph execution counter to session state
* move increment to inference session
* std::endl to \n
* ifdef on graph execution counter
* add ifdef to inference session
* move DEBUG_NODE_INPUTS_OUTPUTS to CMakeLists.txt
* GridSample OP implementation for CPU and CUDA
**Description**: This change contains implementation for torch grid_sample OP.
Cuda implementation contains contribution from Muscle Wu.
* Use interpolation for out-of-bound points in zero padding mode
Out-of-bound points in zeros padding mode changed from constant 0 to
interpolation of surrounding pixels. This aligns with Pytorch implementation.
A bug in CUDA batch offset calculation is fixed.
Custom op exporter type is added.
* Fix nearest bug in CPU
* Update per CI build finding and review comments
* Force float to avoid potential integer T issue
* Style update
* PR update
* Remove c++17 feature from cuda code
* fix build - python.h not found
* disable --build_shared_lib for ortmodule tests
* fix
* fix the build flag
* disable --build_shared_lib for training path (not only for ortmodule)
* fix missing test model files
* disable test CApiTest.test_custom_op_library when ENABLE_TRAINING_TORCH_INTEROP is ON
* enable custom_op_library build
* fix build
* fix
* merge master and fix build failure
* build onnx_test_runner when onnxruntime_ENABLE_TRAINING_TORCH_INTEROP is ON
* resolve comments
* use --enable_training_torch_interop to replace "onnxruntime_ENABLE_TRAINING_TORCH_INTEROP=ON"
* Merge CPU/GPU nuget pipeline
* Include TensorRT EP libraries into existing GPU nuget package pipeline
* modify to use correct YAML
* Modify for test
* modify for test
* Add depedance
* Add depedance (cont.)
* modify for test
* Add create TensorRT nuget package
* modify for test
* modify for test
* Merge CPU/GPU nuget pipeline
* Include TensorRT EP libraries into existing GPU nuget package pipeline
* modify to use correct YAML
* Modify for test
* modify for test
* Add depedance
* Add depedance (cont.)
* modify for test
* Add create TensorRT nuget package
* modify for test
* fix merge bug
* code refactor
* code refactor
* modify for test
* modify for test
* modify for test
* modify for test
* modify for test
* modify for test
* cleanup
* modify for test
* fix bug
* modify for test
* refactor
* fix bug and test
* Modify for test
* Modify for test
* Modify for test
* Modify for test
* Prepare for PR
* Prepare for PR
* code refacotr from review
* Remove naming 'Microsoft.ML.OnnxRuntime.TensorRT' to avoid confusion
* Add linux TensorRT libraries
* Remove redundant variable in YMAL
* revert file
* undo revert file
* Modify regular expression so that it can capture the correct file
* Remove newline at end of file
* small fix
* Revert to CUDA11.1 on Windows
* Add unit tests for nuget package on Linux
Co-authored-by: Changming Sun <chasun@microsoft.com>
Using signed int, qgemm kernel avoids extending uint8 to int16 while computing matrix multiplication, achieving higher performance. We also find that by using only lower 64b of vector registers to load A and B matrix, we can get further performance improvements. We also experimented with using ldp to load two 64b in one shot, vs using two ldr to load one 64b at a time, in both Big and little cores, there is no noticeable differences.
Submitting the LDP version. At this point we don't need to choose kernel based on micro-architecture.
Inference time of resnet50, thread count 2
Big Core on Pixel 3a
Current master: 292.947 ms
First iteration S8S8: 188.239 ms
LDP load two 64b reg: 178.715 ms
LDR load one 64b reg: 179.536 ms
Little Core
Master: 546.317 ms
S8S8: 513.332 ms
LDP: 489.19 ms
LDR: 497.865 ms
Raspberry Pi 3B+
Master: 660.08 ms
S8S8: 608.577 ms
LDP: 603.675 ms
LDR 602.075 ms
* initial update from 11.1 to 11.4
* change 11.4.1 to 11.4.0
* adjusting to match nvidia/cuda image tags
* adjusting to match nvidia/cuda image tags centos7
* correction to 11.4.0
* correction to 11.4.0
* update to cuda 11.4
* change training back to 11.1
* change training back to 11.1
* point to correct nvcr.io/nvidia/cuda 11.4.1 image
* change centos8 to centos7
* correct cudnn path
* Update linux-gpu-ci-pipeline.yml for Azure Pipelines
* Update c-api-noopenmp-packaging-pipelines.yml
* need to resolve centos images but remove space and change to 11.4
* Update linux-gpu-ci-pipeline.yml
* add cudnn to docker image
* bump devtoolset to 10
* revert cuda 11.4 change to setup_env_trt
* orttraining back to 11.1
* use nvcr.io
* Fix previous change back to cuda 11.1
* update cudnn path
* use cudnn image (revert if failure)
* Fetching frontier tensors to frontend
* Move before session initialize call
* Fetch tensor and add to cache
* Rest of the changes for using cache
* Review comments
* Review changes
* Review comments
* switch to shared_ptr
* Fix bug after rebase
* FE docstring change
Create a new M8 loop processing A[8x8] B[8x8] per iteration.
Avoid saving registers on paths that are not needed.
Adjusted M2 and M1 loop, using more registers to relax the loop carrying dependencies.
Nearly 7% improvement observed on Surface Pro X 2 with model ssd_mobilenet_v2_300
About 4.5% improvement on resnet50 on Surface Pro X 2.
Add IsSparseTensor
Add CreateSparseTensor
Add utilities and test fully sparse instantiation
Fully sparse blocksparse
Add test and docs for fully sparse tensor instantiation
Rework creation API
Use API
Non string API
Retrofit of existing String API
Add tests
Add documentation
Address build issues (Winml pending)
Add inference test
Bump binary size
Add ifdef DISABLE CONTRIB
* dnnl ep rework
rework DnnlTensor,DnnlNode,DnnlSubgraph to support arbitrary graph topology and tensor data types
rework GetCapability to claim nodes in graph greedily from node topological ordering and delay creation of DnnlSubgraph until Compile
rework compile to have DnnlSubgraphPrimitive as the object to handle primitive creation and execution
instead of thread local primitive pool which duplicates intermediate memory allocated by the EP across threads
DnnlSubgraphPrimitive provides helpers to handle many common functions for each dnnl primitive builder and become the centralized place to store input, output, intermediate memories, initializer memories and etc
it provides functions to obtain input memories with automatic reordering/reshaping and moving between engines
it provides interfaces to add primitive, set output memory for single node and etc
add CONCURRENT_EXEC compile flag for dnnl library as without it, convolution primitive cannot be created and executed on different threads
enable unit tests to run on dnnl ep as well if built with dnnl ep
add dnnl ep support for Matmulinteger
* Add Relu to the DNNL refactor
Signed-off-by: George Nash <george.nash@intel.com>
* Add Convolution op to the DNNL rework
Signed-off-by: George Nash <george.nash@intel.com>
* Add Pooling ops to the DNNL rework
This adds the following ops:
- AveragePool
- GlobalAveragePool
- GlobalMaxPool
- MaxPool
Note: Pooling with dilation is not yet supported.
Note: GlobalLpPool, LpPool, MaxRoiPool, and MaxUnpool are not supported yet.
Signed-off-by: George Nash <george.nash@intel.com>
* Add Sum op to the DNNL rework
Signed-off-by: George Nash <george.nash@intel.com>
* Add ConvGrad op to the DNNL rework
Signed-off-by: George Nash <george.nash@intel.com>
* Add MaxPoolGrad and AveragePoolGrad ops to DNNL rework
Signed-off-by: George Nash <george.nash@intel.com>
* Added lrn operator to the refactored code
Signed-off by chethan.palangoutu.keshava@intel.com
* Added ReduceMean DNNL op to the refactor code
Signed-off-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
* Added Softmax DNNL op for the refactored code
Signed-off-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
* Added BatchNorm DNNL op inference-only for refactored code
Signed-off-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
* Added Binary Ops to DNNL rework
Signed-off-by: Wang <zhaoyang.wang@intel.com>
* Added ReluGrad to DNNL Rework
Signed-off-by: Wang <zhaoyang.wang@intel.com>
* Update OneDNN tag to v2.3
Signed-off-by: Wang <zhaoyang.wang@intel.com>
* Added support for memory upto dim size 12
this is to fix the CI test cases that contain binary ops of input dim
size > 5
Signed-off-by: Wang <zhaoyang.wang@intel.com>
* Prevent claiming support for float16 and bfloat16 when only float is suppoted
By using The string.find used was causing the code to claiming support
for float16 and bfloat16 when we only supported float. We now explicitly
check the code for the data type or the data type with a 7 letter prefix
basically prefixed with "tensor("
Signed-off-by: George Nash <george.nash@intel.com>
* Disable uint8 mul and div, improve type conversion
Disable mul_uint8 and div_uint8 test cases as they use modulo for
overflow handling while onednn uses saturation
improve ype conversion using enum instead of string comparsion as well
as adding more types
Signed-off-by: Wang <zhaoyang.wang@intel.com>
Co-authored-by: Wang <zhaoyang.wang@intel.com>
Co-authored-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
