profile
viewpoint
If you are wondering where the data of this site comes from, please visit https://api.github.com/users/rsuderman/events. GitMemory does not store any data, but only uses NGINX to cache data for a period of time. The idea behind GitMemory is simply to give users a better reading experience.

google/iree 832

👻

rsuderman/llvm-project 0

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. Note: the repository does not accept github pull requests at this moment. Please submit your patches at http://reviews.llvm.org.

rsuderman/tensorflow 0

An Open Source Machine Learning Framework for Everyone

rsuderman/tosa-mlir-gen 0

Utilities for converting tosa tests to mlir and iree tests

PR opened tensorflow/tensorflow

Lowering for tfl.sparse_to_dense to tosa

Sparse to dense can be implemented using a series of reshapes, constants, numerical operations, and a final scatter. This should work to decompose into a TOSA compatible form.

+103 -0

0 comment

2 changed files

pr created time in 7 hours

push eventrsuderman/tensorflow

Rob Suderman

commit sha f901da42e3a7781add9023965eb76162cdbfe29b

Lowering for tfl.sparse_to_dense to tosa Sparse to dense can be implemented using a series of reshapes, constants, numerical operations, and a final scatter. This should work to decompose into a TOSA compatible form.

view details

push time in 7 hours

create barnchrsuderman/tensorflow

branch : SparseToDense

created branch time in 7 hours

push eventllvm/llvm-project

Kojo Acquah

commit sha 9c62bb55f473a9d0db16b894708ed09f2346ae9d

Implementation of `ReshapeNoopOptimization` canonicalizer. This canonicalizer replaces reshapes of constant tensors that contain the updated shape (skipping the reshape operation). Differential Revision: https://reviews.llvm.org/D112038

view details

push time in 2 days

PR opened tensorflow/tensorflow

Lowering dynamic case of tfl.split to tosa

In the dynamic case we can handle the split by using reshapes and splits. This allows splitting across the dynamic dimension though it assumes the dimension can be legally split.

+64 -23

0 comment

2 changed files

pr created time in 2 days

create barnchrsuderman/tensorflow

branch : SplitDynamic

created branch time in 2 days

push eventllvm/llvm-project

not-jenni

commit sha 4ada6c2aafffd90c87900cab0adbb4d43c874b9b

[mlir][tosa] Adds a canonicalization to the transpose op if the perms are a no op Reviewed By: rsuderman Differential Revision: https://reviews.llvm.org/D112037

view details

push time in 3 days

push eventgoogle/iree-samples

Rob Suderman

commit sha 180e506482268b536d18be226ffd6a860965636a

Added nasnet test (working) Disabled except under hugetest due to download size (300 mb)

view details

push time in 3 days

push eventgoogle/iree-samples

Rob Suderman

commit sha 7a709a38735781128ce29ac9683b477895054347

Added magenta test (passing)

view details

Rob Suderman

commit sha 37434531e2f0fdd27be08c9944cde2d294729ded

Fix efficientnet comparison and test_util unsigned comparison.

view details

push time in 6 days

push eventgoogle/iree-samples

Rob Suderman

commit sha ee9ae855ba1f58bc117b441ebac0d11447175dda

Added keras-ocr test (failing) and added config flag Config flag allows iree-samples to target vulkan and spirv backends.

view details

push time in 7 days

push eventrsuderman/tensorflow

Rob Suderman

commit sha 86dd26da1f7da2383c203b3156f60d642ef60d7d

Removed comment about negative strides.

view details

push time in 7 days

PR opened tensorflow/tensorflow

Fixed tfl.strided_slice lowering to tosa for negative stride

Negative strides should reverse those axis after the slice is performed. We insert tosa.reverse operations after the slicing / reshaping work is finished.

+61 -36

0 comment

2 changed files

pr created time in 7 days

PR closed tensorflow/tensorflow

Reviewers
Fixed tfl.strided_slice lowering to tosa for negative stride cla: yes awaiting review size:M

Negative strides should reverse those axis after the slice is performed. We insert tosa.reverse operations after the slicing / reshaping work is finished.

+61 -36

0 comment

2 changed files

rsuderman

pr closed time in 7 days

push eventrsuderman/tensorflow

Rob Suderman

commit sha 11bc67250951a543cdf1a7f820bd46ce7d87c3a0

git-clang-format

view details

push time in 7 days

create barnchrsuderman/tensorflow

branch : StridedSliceReverse

created branch time in 7 days

PR opened tensorflow/tensorflow

Fixed tfl.strided_slice lowering to tosa for negative stride

Negative strides should reverse those axis after the slice is performed. We insert tosa.reverse operations after the slicing / reshaping work is finished.

+58 -36

0 comment

2 changed files

pr created time in 7 days

push eventgoogle/iree-samples

Rob Suderman

commit sha 37d6beae4c765c53ef9fa105143c98d30f64e54c

Added efficient net test and avoided extraneuos local copy in test_util.py

view details

push time in 7 days

push eventgoogle/iree-samples

Rob Suderman

commit sha e15f7e874a459f1b3e813df4507903414a3c40e3

Added efficient net test and avoided extraneuos local copy in test_util.py

view details

push time in 7 days

push eventgoogle/iree-samples

Rob Suderman

commit sha 29b6310ad625040225626f8555bdee4f8201f80f

Renamed deeplab_v3 and add densenet

view details

push time in 7 days

push eventllvm/llvm-project

Rob Suderman

commit sha 59dd418e8941434a65fbace26dcd405a88b3775c

[mlir][tosa] Fix tosa.cast UiToFp32 for tosa-to-linalg Part of the arith update broke UiToFp32. Fixed the lowering and included a new test to detect a regression. Differential Revision: https://reviews.llvm.org/D111772

view details

push time in 7 days

push eventgoogle/iree

rsuderman

commit sha 361c405139ab5a0ffc58cdfab1a45e3974701c42

Added TosaToStandard to iree-opt by default (#7338)

view details

push time in 8 days

delete branch rsuderman/iree

delete branch : AddedTosaStandard

delete time in 8 days

PR merged google/iree

Added TosaToStandard to iree-opt by default cla: yes

Useful for testing the TOSA front-end to have both TosaTo* passes included.

+1 -0

0 comment

1 changed file

rsuderman

pr closed time in 8 days

Pull request review commentgoogle/iree

Added a series of TFLite Benchmarks to the IREE's Perf Benchmark

 set(DEEPLABV3_FP32_MODULE   "1x257x257x3xf32"               # FUNCTION_INPUTS ) +#TFHub model: https://tfhub.dev/tensorflow/lite-model/mobilebert/1/default/1+set(MOBILEBERT_FP32_MODULE+  "MobileBert"                    # MODULE_NAME+  "fp32"                          # MODULE_TAGS+  "https://storage.googleapis.com/iree-model-artifacts/MobileBertTosa-5f984ec.tar.gz" # MLIR_SOURCE+  "main"                          # ENTRY_FUNCTION+  "1x384xi32,1x384xi32,1x384xi32" # FUNCTION_INPUTS+)++# TFHub model: https://tfhub.dev/tensorflow/lite-model/mobilenet_v1_1.0_160/1/default/1+set(MOBILENETV1_FP32_MODULE+  "MobileNetV1"                   # MODULE_NAME+  "fp32"                          # MODULE_TAGS+  "https://storage.googleapis.com/iree-model-artifacts/MobileNetV1Tosa-5f984ec.tar.gz" # MLIR_SOURCE+  "main"                          # ENTRY_FUNCTION+  "1x160x160x3xf32"               # FUNCTION_INPUTS+)++# TFHub model: https://tfhub.dev/svampeatlas/lite-model/vision/classifier/fungi_mobile_V1/1/default/1

Its possible they are incorrect. The TFHub site is dependent on user specified data which seems likely to be wrong. My perspective is a V2, V3 and quantized model are the main goals. Nothing matters beyond that.

rsuderman

comment created time in 8 days

PullRequestReviewEvent

pull request commentgoogle/iree

Added a series of TFLite Benchmarks to the IREE's Perf Benchmark

I actually wonder if it makes sense to hold off on these until we can benchmark from tflite flatbuffers directly. Having just updated the IR files in GCS, I wouldn't say it was exactly smooth, and I don't love greatly expanding the scope before fixing that problem

It is certainly an option. I stuck with IR files for the sake of consistency and wanted to avoid doing anything special for the TFLite tests.

rsuderman

comment created time in 8 days

Pull request review commentgoogle/iree

Added a series of TFLite Benchmarks to the IREE's Perf Benchmark

 set(DEEPLABV3_FP32_MODULE   "1x257x257x3xf32"               # FUNCTION_INPUTS ) +#TFHub model: https://tfhub.dev/tensorflow/lite-model/mobilebert/1/default/1+set(MOBILEBERT_FP32_MODULE+  "MobileBert"                    # MODULE_NAME+  "fp32"                          # MODULE_TAGS+  "https://storage.googleapis.com/iree-model-artifacts/MobileBertTosa-5f984ec.tar.gz" # MLIR_SOURCE+  "main"                          # ENTRY_FUNCTION+  "1x384xi32,1x384xi32,1x384xi32" # FUNCTION_INPUTS+)++# TFHub model: https://tfhub.dev/tensorflow/lite-model/mobilenet_v1_1.0_160/1/default/1+set(MOBILENETV1_FP32_MODULE+  "MobileNetV1"                   # MODULE_NAME+  "fp32"                          # MODULE_TAGS+  "https://storage.googleapis.com/iree-model-artifacts/MobileNetV1Tosa-5f984ec.tar.gz" # MLIR_SOURCE+  "main"                          # ENTRY_FUNCTION+  "1x160x160x3xf32"               # FUNCTION_INPUTS+)++# TFHub model: https://tfhub.dev/svampeatlas/lite-model/vision/classifier/fungi_mobile_V1/1/default/1+set(MOBILENETV2_FP32_MODULE+  "MobileNetV2"                   # MODULE_NAME+  "fp32"                          # MODULE_TAGS+  "https://storage.googleapis.com/iree-model-artifacts/MobileNetV2Tosa-5f984ec.tar.gz" # MLIR_SOURCE+  "main"                          # ENTRY_FUNCTION+  "1x299x299x3xi8"                # FUNCTION_INPUTS+)++# TFHub model: https://tfhub.dev/tensorflow/lite-model/mobilenet_v2_1.0_224_quantized/1/default/1+set(MOBILENETV2_I8_MODULE+  "MobileNetV2Quant"              # MODULE_NAME+  "i8"                            # MODULE_TAGS+  "https://storage.googleapis.com/iree-model-artifacts/MobileNetV2QuantTosa-5f984ec.tar.gz" # MLIR_SOURCE+  "main"                          # ENTRY_FUNCTION+  "1x224x224x3xi8"                # FUNCTION_INPUTS+)++# TFHub Model: https://tfhub.dev/bohemian-visual-recognition-alliance/lite-model/models/mushroom-identification_v1/1?

Oh probably. The model contents are actually the same between them. The metadata version just has additional metadata specified.

rsuderman

comment created time in 8 days

PullRequestReviewEvent

issue closedgoogle/iree

ConvertStreamOps fails on transient buffer used for tosa.argmax

Describe the bug 'tosa.argmax' uses a transient buffer to store the running maximum separate from the index. It appears that during --iree-convert-to-hal for llvm/spirv it fails to allocate this transient buffer causing an assertion/segmentation-fault failure.

It appears we are missing a case to allocate buffers used by dispatches that are not returned with flow.return.

To Reproduce iree/tools/iree-opt /tmp/test.mlir.txt --iree-convert-to-hal

test.mlir.txt

closed time in 8 days

rsuderman

issue commentgoogle/iree

ConvertStreamOps fails on transient buffer used for tosa.argmax

This should be fixed. I haven't seen it arise again so best to close.

rsuderman

comment created time in 8 days