Initial implementation of Deformable Convolution workload #168

AlexanderKalistratov · 2022-12-01T20:27:41Z

No description provided.

AlexanderKalistratov · 2022-12-01T20:28:28Z

adarshyoga · 2023-01-05T21:25:02Z

dpbench/benchmarks/deformable_convolution/__init__.py

+    "deformable_convolution_sycl",
+]
+
+"""l2-norm calculation of n vectors


Comments need to be added for this workload

adarshyoga · 2023-01-05T21:26:14Z

dpbench/benchmarks/deformable_convolution/deformable_convolution_initialize.py

+
+    return (
+        default_rng.random((batch, in_channels, in_height, in_width)).astype(dtype),
+        # np.ones((batch, in_channels, in_height, in_width)).astype(dtype),


Please cleanup commented out code.

adarshyoga · 2023-01-05T21:42:46Z

dpbench/benchmarks/deformable_convolution/deformable_convolution_numpy.py

@@ -0,0 +1,81 @@
+# Copyright 2022 Intel Corp.


Looks like this _numpy implementation is the same as the _numba_npr version. If so, we should not be adding this.

adarshyoga · 2023-01-05T21:43:19Z

dpbench/benchmarks/deformable_convolution/deformable_convolution_sycl_native_ext/CMakeLists.txt

+)
+find_package(MKL CONFIG REQUIRED)
+
+# target_include_directories(${py_module_name} PUBLIC ${Dpctl_INCLUDE_DIRS} $)


Please remove commented out lines

adarshyoga · 2023-01-05T21:43:56Z

...able_convolution/deformable_convolution_sycl_native_ext/deformable_convolution_sycl/impl.cpp

@@ -0,0 +1,297 @@
+//==- _l2_norm_sycl.cpp - Python native extension of l2-norm   ===//


Update comments to reflect this workload

adarshyoga · 2023-01-05T21:44:11Z

...able_convolution/deformable_convolution_sycl_native_ext/deformable_convolution_sycl/impl.cpp

+// SPDX - License - Identifier : Apache 2.0
+///
+/// \file
+/// The files implements a SYCL-based Python native extension for the


Update comments

adarshyoga · 2023-01-05T21:44:38Z

...able_convolution/deformable_convolution_sycl_native_ext/deformable_convolution_sycl/impl.cpp

+
+        auto offset_y = *get_ptr_5d(offset, k_height, k_width, 2, out_height, out_width, kh, kw, 1, h, w) + h*stride_y + (kh - k_h_m)*dilation_y - (pad_y - k_h_m);
+        auto offset_x = *get_ptr_5d(offset, k_height, k_width, 2, out_height, out_width, kh, kw, 0, h, w) + w*stride_x + (kw - k_w_m)*dilation_x - (pad_x - k_w_m);
+        // auto offset_y = h*stride_y + (kh - k_h_m)*dilation_y - (pad_y - k_h_m);


Remove commented lines

adarshyoga · 2023-01-05T21:44:49Z

...able_convolution/deformable_convolution_sycl_native_ext/deformable_convolution_sycl/impl.cpp

+        auto _input = get_ptr_3d(input, in_channels, in_height, in_width, c, 0, 0);
+
+        _output[w] = bilinear(_input, in_height, in_width, offset_y, offset_x);
+        // _output[w] = offset_y;


Remove line

adarshyoga · 2023-01-05T21:48:27Z

dpbench/runner.py

        + " ========================"
    )
-    if result.error_state == 0:
+    if result.error_state == enums.ErrorCodes.SUCCESS:


This change does not appear to be part of this workload PR. Consider adding separate PR.

ZzEeKkAa · 2023-05-05T17:42:46Z

@AlexanderKalistratov thank you for you PR. Could you please update implementation (@diptorupd said that you've improved it). Please also get rid of unrelated comments, update comments to match current implementation and provide PR description.

Hoping to see updates in the nearest feature!

AlexanderKalistratov · 2023-06-28T00:09:18Z

@ZzEeKkAa deformable convolution can't pass CI because it uses sycl implementation as reference, and there are no-sycl runs.
Also have to modify benchmark_runner.py. Otherwise it simply hangs on no-sycl runs.
https://github.com/IntelPython/dpbench/pull/168/files#diff-84ed6cf60ee660c2164668c4d4296513ccb0a6b01e124b2de6a35d6133653ad8R360

Hardcode84 · 2023-06-29T12:52:18Z

M size is too small compared to other WLs, I suggest to move L to M and make L even bigger

ZzEeKkAa · 2023-06-29T15:04:59Z

Thank @AlexanderKalistratov and @Hardcode84. Is it hard to add numpy implementation (using numpy features)?

ZzEeKkAa · 2023-06-29T15:11:34Z

Are MKL AND TBB cmake files patched? If no and compilation is failing we need to update workflow, not provide copy of cmake files. I guess dpnp or dpctl does it in conda recepie, so we can refer there.

Hardcode84 · 2023-06-29T15:15:58Z

Is it hard to add numpy implementation (using numpy features)?

Numpy version will be prohibitively slow even for S size.

AlexanderKalistratov · 2023-06-29T15:56:12Z

@ZzEeKkAa

Is it hard to add numpy implementation (using numpy features)?

As Ivan said straightforward implementation would always fail with timeout even on S preset. I had one and then remove it.
Probably it is possible to make faster implementation, but it is quiet tricky and has no practical value.

I'd say we need option to exclude workload from run. Right now we have option to list workloads to run and now we need also option to list workloads NOT to run. And then simply exclude deformable convolution from no-sycl runs.

Are MKL AND TBB cmake files patched? If no and compilation is failing we need to update workflow, not provide copy of cmake files. I guess dpnp or dpctl does it in conda recepie, so we can refer there.

They are patched. These cmakes taken from dpnp repo.

AlexanderKalistratov · 2023-06-29T15:56:55Z

Also we need to handle the case when reference implementation not found more gracefully than now

adarshyoga · 2024-04-26T16:18:50Z

Closing this PR. Please reopen if there are further updates.

adarshyoga suggested changes Jan 5, 2023

View reviewed changes

adarshyoga force-pushed the deformable_convolution branch from 5ac57f4 to 031e9f8 Compare February 9, 2023 16:52

adarshyoga assigned AlexanderKalistratov Apr 4, 2023

AlexanderKalistratov force-pushed the deformable_convolution branch from 031e9f8 to 5c6c720 Compare April 10, 2023 21:30

AlexanderKalistratov force-pushed the deformable_convolution branch from 3c0758e to d525b2c Compare June 1, 2023 09:01

AlexanderKalistratov mentioned this pull request Jun 1, 2023

Only python and numpy implementations available as reference #271

Closed

AlexanderKalistratov force-pushed the deformable_convolution branch 2 times, most recently from bee3199 to 01ce90e Compare June 24, 2023 16:15

AlexanderKalistratov force-pushed the deformable_convolution branch from cc17ce1 to 424694e Compare June 27, 2023 12:58

AlexanderKalistratov added 19 commits June 27, 2023 23:22

Initial implementation of Deformable Convolution workload

a595aea

Fix pre-commit

2963944

refactor

809047d

Small fixes

7d39e0e

Add config & pre-commit

0de01b1

Fix mlir_p version & result

0e1b858

Add TBBConfig.cmake

ff457c0

Fix TBB discovery

95fa4a9

Add tbb and mkl to the environment

00a0944

pre-commit

46d90dc

Add MKLConfig.cmake

5badbc4

Fix mkl discovery

a1a88a5

Add tbb and mkl-devel to win conda env

bce5efa

Test cmake changes

2642393

Revert some changes

a85dcf4

Fix hang

4e2b42b

Fix hang

909d103

pre-commit

e7face2

Some cleanup

c5bddb3

AlexanderKalistratov force-pushed the deformable_convolution branch from 424694e to c5bddb3 Compare June 27, 2023 22:51

Remove auto convert list to vector

0c4f1f2

AlexanderKalistratov mentioned this pull request Jul 7, 2023

Add way to exclude workload from run #286

Open

adarshyoga closed this Apr 26, 2024

		@@ -0,0 +1,297 @@
		//==- _l2_norm_sycl.cpp - Python native extension of l2-norm ===//

Initial implementation of Deformable Convolution workload #168

Initial implementation of Deformable Convolution workload #168

Uh oh!

Conversation

AlexanderKalistratov commented Dec 1, 2022

Uh oh!

AlexanderKalistratov commented Dec 1, 2022

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ZzEeKkAa commented May 5, 2023

Uh oh!

AlexanderKalistratov commented Jun 28, 2023

Uh oh!

Hardcode84 commented Jun 29, 2023

Uh oh!

ZzEeKkAa commented Jun 29, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ZzEeKkAa commented Jun 29, 2023

Uh oh!

Hardcode84 commented Jun 29, 2023

Uh oh!

AlexanderKalistratov commented Jun 29, 2023

Uh oh!

AlexanderKalistratov commented Jun 29, 2023

Uh oh!

adarshyoga commented Apr 26, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

ZzEeKkAa commented Jun 29, 2023 •

edited

Loading