add patches to fix PyTorch 1.11 on POWER

easybuild/easyconfigs/p/PyTorch/PyTorch-1.10.0_fix-XNNPACK-tests.patch

@@ -0,0 +1,236 @@
+When XNNPACK is disabled tests run into assertion failures.
+Fix by using (most of) https://github.com/pytorch/pytorch/pull/72642
+
+Author: Digant Desai <digantdesai@fb.com>
+
+Backported by Alexander Grund (TU Dresden)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 0c11507838..5b74b7f63c 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -275,7 +275,14 @@ option(USE_LITE_INTERPRETER_PROFILER "Enable " ON)
+ option(USE_VULKAN_FP16_INFERENCE "Vulkan - Use fp16 inference" OFF)
+ option(USE_VULKAN_RELAXED_PRECISION "Vulkan - Use relaxed precision math in the kernels (mediump)" OFF)
+ option(USE_VULKAN_SHADERC_RUNTIME "Vulkan - Use runtime shader compilation as opposed to build-time (needs libshaderc)" OFF)
+-option(USE_XNNPACK "Use XNNPACK" ON)
++# option USE_XNNPACK: try to enable xnnpack by default.
++set(XNNPACK_MIN_CMAKE_VER 3.12)
++cmake_dependent_option(
++    USE_XNNPACK "Use XNNPACK. Requires cmake >= ${XNNPACK_MIN_CMAKE_VER}." ON
++    "CMAKE_VERSION VERSION_GREATER_EQUAL ${XNNPACK_MIN_CMAKE_VER}" OFF)
++if(NOT USE_XNNPACK AND CMAKE_VERSION VERSION_LESS ${XNNPACK_MIN_CMAKE_VER})
++  message(WARNING "USE_XNNPACK is set to OFF. XNNPACK requires CMake version ${XNNPACK_MIN_CMAKE_VER} or greater.")
++endif()
+ option(USE_ZMQ "Use ZMQ" OFF)
+ option(USE_ZSTD "Use ZSTD" OFF)
+ # Ensure that an MKLDNN build is the default for x86 CPUs
+diff --git a/cmake/Summary.cmake b/cmake/Summary.cmake
+index 6d021536c1..7c1ac64539 100644
+--- a/cmake/Summary.cmake
++++ b/cmake/Summary.cmake
+@@ -165,6 +165,7 @@ function(caffe2_print_configuration_summary)
+   message(STATUS "  USE_PROF              : ${USE_PROF}")
+   message(STATUS "  USE_QNNPACK           : ${USE_QNNPACK}")
+   message(STATUS "  USE_PYTORCH_QNNPACK   : ${USE_PYTORCH_QNNPACK}")
++  message(STATUS "  USE_XNNPACK           : ${USE_XNNPACK}")
+   message(STATUS "  USE_REDIS             : ${USE_REDIS}")
+   message(STATUS "  USE_ROCKSDB           : ${USE_ROCKSDB}")
+   message(STATUS "  USE_ZMQ               : ${USE_ZMQ}")
+diff --git a/test/jit/test_optimize_for_mobile_preserve_debug_info.py b/test/jit/test_optimize_for_mobile_preserve_debug_info.py
+index c08f3b5838..1b93f54c15 100644
+--- a/test/jit/test_optimize_for_mobile_preserve_debug_info.py
++++ b/test/jit/test_optimize_for_mobile_preserve_debug_info.py
+@@ -1,8 +1,8 @@
+ import torch
+ import torch._C
+-import torch.backends.xnnpack
+ import torch.nn.functional as F
+ from torch.testing._internal.jit_utils import JitTestCase
++from torch.testing._internal.common_utils import skipIfNoXNNPACK
+
+ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+     def check_replacement(
+@@ -34,6 +34,7 @@ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+                     original_source_ranges[replacements[node.kind()]],
+                 )
+
++    @skipIfNoXNNPACK
+     def test_replace_conv1d_with_conv2d(self):
+         class TestConv1d(torch.nn.Module):
+             def __init__(self, weight, bias):
+@@ -61,6 +62,7 @@ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+             jit_pass=torch._C._jit_pass_transform_conv1d_to_conv2d,
+         )
+
++    @skipIfNoXNNPACK
+     def test_insert_pre_packed_linear_before_inline_and_conv_2d_op(self):
+         class TestPrepackedLinearBeforeInlineAndConv2dOp(torch.nn.Module):
+             def __init__(
+@@ -137,6 +139,7 @@ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+             jit_pass=torch._C._jit_pass_insert_prepacked_ops,
+         )
+
++    @skipIfNoXNNPACK
+     def test_insert_pre_packed_linear_op(self):
+         self.check_replacement(
+             model=torch.jit.trace(torch.nn.Linear(5, 4), torch.rand(3, 2, 5)),
+@@ -228,6 +231,7 @@ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+             jit_pass=torch._C._jit_pass_fuse_clamp_w_prepacked_linear_conv,
+         )
+
++    @skipIfNoXNNPACK
+     def test_fuse_activation_with_pack_ops_linear_conv2d_1(self):
+         self.run_test_fuse_activation_with_pack_ops_linear_conv2d(
+             linear_activation=F.hardtanh,
+@@ -236,6 +240,7 @@ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+             conv2d_activation_kind="aten::hardtanh_"
+         )
+
++    @skipIfNoXNNPACK
+     def test_fuse_activation_with_pack_ops_linear_conv2d_2(self):
+         self.run_test_fuse_activation_with_pack_ops_linear_conv2d(
+             linear_activation=F.hardtanh_,
+@@ -244,6 +249,7 @@ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+             conv2d_activation_kind="aten::hardtanh"
+         )
+
++    @skipIfNoXNNPACK
+     def test_fuse_activation_with_pack_ops_linear_conv2d_3(self):
+         self.run_test_fuse_activation_with_pack_ops_linear_conv2d(
+             linear_activation=F.relu,
+@@ -252,6 +258,7 @@ class TestOptimizeForMobilePreserveDebugInfo(JitTestCase):
+             conv2d_activation_kind="aten::relu_"
+         )
+
++    @skipIfNoXNNPACK
+     def test_fuse_activation_with_pack_ops_linear_conv2d_4(self):
+         self.run_test_fuse_activation_with_pack_ops_linear_conv2d(
+             linear_activation=F.relu_,
+diff --git a/test/test_mobile_optimizer.py b/test/test_mobile_optimizer.py
+index 19f07e2454..21ae5d3ee9 100644
+--- a/test/test_mobile_optimizer.py
++++ b/test/test_mobile_optimizer.py
+@@ -1,9 +1,8 @@
+ import unittest
+ import torch
+ import torch.nn as nn
+-import torch.backends.xnnpack
+ import torch.utils.bundled_inputs
+-from torch.testing._internal.common_utils import TestCase, run_tests
++from torch.testing._internal.common_utils import TestCase, run_tests, skipIfNoXNNPACK
+ from torch.testing._internal.jit_utils import get_forward, get_forward_graph
+ from torch.utils.mobile_optimizer import (LintCode,
+                                           generate_mobile_module_lints,
+@@ -22,9 +21,7 @@ FileCheck = torch._C.FileCheck
+
+ class TestOptimizer(TestCase):
+
+-    @unittest.skipUnless(torch.backends.xnnpack.enabled,
+-                         " XNNPACK must be enabled for these tests."
+-                         " Please build with USE_XNNPACK=1.")
++    @skipIfNoXNNPACK
+     def test_optimize_for_mobile(self):
+         batch_size = 2
+         input_channels_per_group = 6
+@@ -263,9 +260,7 @@ class TestOptimizer(TestCase):
+             rtol=1e-2,
+             atol=1e-3)
+
+-    @unittest.skipUnless(torch.backends.xnnpack.enabled,
+-                         " XNNPACK must be enabled for these tests."
+-                         " Please build with USE_XNNPACK=1.")
++    @skipIfNoXNNPACK
+     def test_quantized_conv_no_asan_failures(self):
+         # There were ASAN failures when fold_conv_bn was run on
+         # already quantized conv modules. Verifying that this does
+@@ -359,6 +354,7 @@ class TestOptimizer(TestCase):
+         bi_module_lint_list = generate_mobile_module_lints(bi_module)
+         self.assertEqual(len(bi_module_lint_list), 0)
+
++    @skipIfNoXNNPACK
+     def test_preserve_bundled_inputs_methods(self):
+         class MyBundledInputModule(torch.nn.Module):
+             def __init__(self):
+@@ -413,9 +409,7 @@ class TestOptimizer(TestCase):
+         incomplete_bi_module_optim = optimize_for_mobile(incomplete_bi_module, preserved_methods=['get_all_bundled_inputs'])
+         self.assertTrue(hasattr(incomplete_bi_module_optim, 'get_all_bundled_inputs'))
+
+-    @unittest.skipUnless(torch.backends.xnnpack.enabled,
+-                         " XNNPACK must be enabled for these tests."
+-                         " Please build with USE_XNNPACK=1.")
++    @skipIfNoXNNPACK
+     def test_hoist_conv_packed_params(self):
+
+         if 'qnnpack' not in torch.backends.quantized.supported_engines:
+@@ -509,6 +503,7 @@ class TestOptimizer(TestCase):
+             m_optim_res = m_optim(data)
+             torch.testing.assert_close(m_res, m_optim_res, rtol=1e-2, atol=1e-3)
+
++    @skipIfNoXNNPACK
+     @unittest.skipUnless(HAS_TORCHVISION, "Needs torchvision")
+     def test_mobilenet_optimize_for_mobile(self):
+         m = torchvision.models.mobilenet_v3_small()
+diff --git a/test/test_model_dump.py b/test/test_model_dump.py
+index 417bb2a91a..e4a9ffa1bd 100644
+--- a/test/test_model_dump.py
++++ b/test/test_model_dump.py
+@@ -8,9 +8,10 @@ import urllib
+ import unittest
+
+ import torch
++import torch.backends.xnnpack
+ import torch.utils.model_dump
+ import torch.utils.mobile_optimizer
+-from torch.testing._internal.common_utils import TestCase, run_tests, IS_WINDOWS
++from torch.testing._internal.common_utils import TestCase, run_tests, IS_WINDOWS, skipIfNoXNNPACK
+ from torch.testing._internal.common_quantized import supported_qengines
+
+
+@@ -170,6 +171,7 @@ class TestModelDump(TestCase):
+         qmodel = self.get_quant_model()
+         self.do_dump_model(torch.jit.script(qmodel))
+
++    @skipIfNoXNNPACK
+     @unittest.skipUnless("qnnpack" in supported_qengines, "QNNPACK not available")
+     def test_optimized_quantized_model(self):
+         qmodel = self.get_quant_model()
+diff --git a/test/test_xnnpack_integration.py b/test/test_xnnpack_integration.py
+index a0f8328ec6..a737ed4b8b 100644
+--- a/test/test_xnnpack_integration.py
++++ b/test/test_xnnpack_integration.py
+@@ -51,7 +51,6 @@ class TestXNNPACKOps(TestCase):
+         output_linearprepacked = torch.ops.prepacked.linear_clamp_run(input_data, packed_weight_bias)
+         torch.testing.assert_close(ref_result, output_linearprepacked, rtol=1e-2, atol=1e-3)
+
+-
+     @given(batch_size=st.integers(0, 3),
+            input_channels_per_group=st.integers(1, 32),
+            height=st.integers(5, 64),
+diff --git a/torch/testing/_internal/common_utils.py b/torch/testing/_internal/common_utils.py
+index b4ef79620d..18d8925d6a 100644
+--- a/torch/testing/_internal/common_utils.py
++++ b/torch/testing/_internal/common_utils.py
+@@ -56,6 +56,7 @@ from torch._six import string_classes
+ from torch import Tensor
+ import torch.backends.cudnn
+ import torch.backends.mkl
++import torch.backends.xnnpack
+ from enum import Enum
+
+ torch.backends.disable_global_flags()
+@@ -908,6 +909,14 @@ def _test_function(fn, device):
+         return fn(self, device)
+     return run_test_function
+
++def skipIfNoXNNPACK(fn):
++    @wraps(fn)
++    def wrapper(*args, **kwargs):
++        if not torch.backends.xnnpack.enabled:
++            raise unittest.SkipTest('XNNPACK must be enabled for these tests. Please build with USE_XNNPACK=1.')
++        else:
++            fn(*args, **kwargs)
++    return wrapper
+
+ def skipIfNoLapack(fn):
+     @wraps(fn)

easybuild/easyconfigs/p/PyTorch/PyTorch-1.10.0_fix-kineto-crash.patch

@@ -0,0 +1,24 @@
+Fix a crash during application shutdown visible in test_profiler on some machines.
+See https://github.com/pytorch/kineto/pull/642
+
+Author: Alexander Grund (TU Dresden)
+
+diff -aur a/third_party/kineto/libkineto/src/EventProfilerController.cpp b/third_party/kineto/libkineto/src/EventProfilerController.cpp
+--- a/third_party/kineto/libkineto/src/EventProfilerController.cpp	2022-08-05 13:10:46.175716618 +0200
++++ b/third_party/kineto/libkineto/src/EventProfilerController.cpp	2022-08-05 13:16:00.654118490 +0200
+@@ -233,9 +233,14 @@ EventProfilerController::~EventProfilerController() {
+
+ // Must be called under lock
+ void EventProfilerController::start(CUcontext ctx, ConfigLoader& configLoader) {
+-  profilerMap()[ctx] = unique_ptr<EventProfilerController>(
++  // Avoid static initialization order fiasco:
++  // We need the profilerMap and with it all controllers to be destroyed
++  // before everything the controller accesses gets destroyed.
++  // Hence access the profilerMap after initialization of the controller.
++  auto controller = unique_ptr<EventProfilerController>(
+       new EventProfilerController(
+           ctx, configLoader, detail::HeartbeatMonitor::instance()));
++  profilerMap()[ctx] = std::move(controller);
+ }
+
+ // Must be called under lock

easybuild/easyconfigs/p/PyTorch/PyTorch-1.10.0_fix-test-model_dump.patch

@@ -0,0 +1,36 @@
+From https://github.com/pytorch/pytorch/pull/84744
+
+Author: Alexander Grund (TU Dresden)
+
+From 4d99f72dcf71bffa7ff4275750cb86d872ac653f Mon Sep 17 00:00:00 2001
+From: Alexander Grund <alexander.grund@tu-dresden.de>
+Date: Fri, 9 Sep 2022 11:10:05 +0200
+Subject: [PATCH] Fix failing test_model_dump due to empty file
+
+The `torch.jit.save` call on a file object may not actually write the
+data to disk due to buffering. The call to `model_dump.main` on that
+file will when fail with an error like
+
+> zipfile.BadZipFile: File is not a zip file
+
+Inspecting the file confirms that it is either empty (usually) or
+incomplete (possible).
+
+Fix this by flushing the file after saving the model.
+---
+ test/test_model_dump.py | 2 ++
+ 1 file changed, 2 insertions(+)
+
+diff --git a/test/test_model_dump.py b/test/test_model_dump.py
+index a8add0e2cd923..3c682b6ce6805 100644
+--- a/test/test_model_dump.py
++++ b/test/test_model_dump.py
+@@ -131,6 +131,8 @@ def test_main(self):
+
+         with tempfile.NamedTemporaryFile() as tf:
+             torch.jit.save(torch.jit.script(SimpleModel()), tf)
++            # Actually write contents to disk so we can read it below
++            tf.flush()
+
+             stdout = io.StringIO()
+             torch.utils.model_dump.main(

easybuild/easyconfigs/p/PyTorch/PyTorch-1.10.0_fix-vsx-vector-functions.patch

@@ -0,0 +1,39 @@
+This fixes the remaining bug introduced by the VSX optimized code in https://github.com/pytorch/pytorch/pull/41541
+
+See https://github.com/pytorch/pytorch/pull/59382
+
+Author: Alexander Grund (TU Dresden)
+
+diff --git a/aten/src/ATen/cpu/vec/vec256/vsx/vec256_float_vsx.h b/aten/src/ATen/cpu/vec/vec256/vsx/vec256_float_vsx.h
+index 2427276bce..2b46e0a662 100644
+--- a/aten/src/ATen/cpu/vec/vec256/vsx/vec256_float_vsx.h
++++ b/aten/src/ATen/cpu/vec/vec256/vsx/vec256_float_vsx.h
+@@ -558,27 +558,7 @@ class Vectorized<float> {
+   }
+
+   Vectorized<float> C10_ALWAYS_INLINE pow(const Vectorized<float>& exp) const {
+-    auto x = *this;
+-    auto sign_bit = (*this) & sign_mask;
+-    // |b|
+-    auto exp_abs = exp.abs();
+-    auto exp_trunc = exp.trunc();
+-    Vectorized<float> odd_mask;
+-    odd_mask._vecb0 = (vec_signed(exp._vec0) & vi_1) != vi_0;
+-    odd_mask._vecb1 = (vec_signed(exp._vec1) & vi_1) != vi_0;
+-    // using ln fuction
+-    auto temp = (abs().log() * exp).exp();
+-
+-    // is odd or even check from Sleef
+-    auto is_int = (exp == exp_trunc) | (exp_abs >= vcheck);
+-    auto is_odd = odd_mask & is_int & (exp_abs < vcheck);
+-    // if even then then pow result should be absolute
+-    auto temp_sign = temp | sign_bit; // copy_sign
+-    auto out = blendv(temp, temp_sign, is_odd);
+-    // x<0 and y != N, then NAN
+-    auto out1 = blendv(out, v_nan, ((exp.floor() != exp) & (x < zero)));
+-    // y = 0 then 1
+-    return blendv(out1, one, (exp_abs == zero));
++    return {Sleef_powf4_u10vsx(_vec0, exp._vec0), Sleef_powf4_u10vsx(_vec1, exp._vec1)};
+   }
+
+   Vectorized<float> fmod(const Vectorized<float>& b) const {

easybuild/easyconfigs/p/PyTorch/PyTorch-1.10.0_skip-nnapi-test-without-qnnpack.patch

@@ -0,0 +1,30 @@
+When QNNPACK is not available the test will fail trying to use it.
+Skip them conditionally.
+See https://github.com/pytorch/pytorch/pull/82882
+
+Author: Alexander Grund (TU Dresden)
+
+diff --git a/test/test_nnapi.py b/test/test_nnapi.py
+index f8db7e1a3d..1d4fd52e28 100644
+--- a/test/test_nnapi.py
++++ b/test/test_nnapi.py
+@@ -2,8 +2,10 @@
+ import os
+ import ctypes
+ import torch
++import unittest
+ from typing import Tuple
+ from torch.backends._nnapi.prepare import convert_model_to_nnapi
++from torch.testing._internal.common_quantized import supported_qengines
+ from torch.testing._internal.common_utils import TestCase, run_tests
+
+
+@@ -18,6 +20,8 @@ def nhwc(t):
+     return t
+
+
++@unittest.skipUnless('qnnpack' in supported_qengines,
++                     "This Pytorch Build has not been built with or does not support QNNPACK")
+ class TestNNAPI(TestCase):
+
+     def setUp(self):