fix alignment of embedded tensor weight blobs

Author: atafra

.github/workflows/internal.ci.yml

@@ -958,6 +958,16 @@ jobs:
       artifact-path: build install deps
       cmd: scripts/build.py install --full -D OIDN_SANITIZER=Thread
 
+  build-macos-arm-ubsan:
+    secrets: inherit
+    uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/macos.yml@main
+    with:
+      project: oidn
+      runs-on: '[ "macOS", "ARM64", "build" ]'
+      artifact-out: build-macos-arm-ubsan
+      artifact-path: build install deps
+      cmd: scripts/build.py install --full -D OIDN_SANITIZER=Undefined
+
   build-ios-arm:
     secrets: inherit
     uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/macos.yml@main
@@ -1076,6 +1086,21 @@ jobs:
         export TBB_ENABLE_SANITIZERS=1
         scripts/test.py --device cpu --minimal --log test.log
 
+  test-macos-arm-ubsan:
+    needs: build-macos-arm-ubsan
+    secrets: inherit
+    uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/macos.yml@main
+    with:
+      project: oidn
+      runs-on: '[ "macOS", "ARM64", "build" ]' # match the build job due to UBSAN
+      artifact-in: build-macos-arm-ubsan
+      artifact-out: test-macos-arm-ubsan
+      artifact-path: test.log
+      artifact-on-failure: true
+      cmd: |
+        export TBB_ENABLE_SANITIZERS=1
+        scripts/test.py --device cpu --minimal --log test.log
+
   test-macos-arm-metal:
     needs: build-macos-arm
     secrets: inherit
@@ -1158,6 +1183,21 @@ jobs:
         export TBB_ENABLE_SANITIZERS=1
         scripts/test.py --device metal --minimal --log test.log
 
+  test-macos-arm-ubsan-metal:
+    needs: build-macos-arm-ubsan
+    secrets: inherit
+    uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/macos.yml@main
+    with:
+      project: oidn
+      runs-on: '[ "macOS", "ARM64", "build" ]' # match the build job due to UBSAN
+      artifact-in: build-macos-arm-ubsan
+      artifact-out: test-macos-arm-ubsan-metal
+      artifact-path: test.log
+      artifact-on-failure: true
+      cmd: |
+        export TBB_ENABLE_SANITIZERS=1
+        scripts/test.py --device metal --minimal --log test.log
+
   build-macos:
     secrets: inherit
     uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/macos.yml@main

apps/oidnTest.cpp

@@ -334,11 +334,14 @@ TEST_CASE("buffer creation", "[buffer]")
   }
 #endif
 
+#if !defined(OIDN_SANITIZER)
+  // Passing an invalid enum value is a false positive for UBSan
   SECTION("invalid buffer storage")
   {
     BufferRef buffer = device.newBuffer(bufferSize, static_cast<Storage>(-42));
     REQUIRE(device.getError() == Error::InvalidArgument);
   }
+#endif
 
   SECTION("device released before buffer")
   {
@@ -632,9 +635,12 @@ TEST_CASE("single filter", "[single_filter][minimal]")
   setFilterImage(filter, nullptr, input);
   REQUIRE(device.getError() == Error::InvalidArgument);
 
-  // Try setting an image with invalid format
+#if !defined(OIDN_SANITIZER)
+  // Try setting an image with invalid format (passing an invalid enum value is a false positive for
+  // UBSan)
   filter.setImage("color", input->getBuffer(), static_cast<Format>(-1), W, H);
   REQUIRE(device.getError() == Error::InvalidArgument);
+#endif
 
   // Try setting an image with buffer overflow
   filter.setImage("color", input->getBuffer(), input->getFormat(), W+1, H);

scripts/blob_to_cpp.py

@@ -8,6 +8,10 @@
 import argparse
 from array import array
 
+# Alignment of the generated blob in bytes. This ensures that data stored at aligned offsets within
+# the blob ends up at correctly aligned absolute addresses.
+BLOB_ALIGNMENT = 64
+
 def is_git_lfs_pointer(data):
   HEADER = array('B', b'version https://git-lfs.github.com/spec/')
   return data[:len(HEADER)] == HEADER
@@ -50,7 +54,7 @@ def generate(in_path, cpp_path, hpp_path, namespace):
     write_prologue(cpp_file, in_name)
     write_namespace_begin(cpp_file, scopes)
 
-    cpp_file.write('extern const unsigned char %s[%d] = {' % (var_name, in_size))
+    cpp_file.write('alignas(%d) extern const unsigned char %s[%d] = {' % (BLOB_ALIGNMENT, var_name, in_size))
     line_length = 1000
     for i in range(in_size):
       c = in_data[i]
@@ -69,7 +73,7 @@ def generate(in_path, cpp_path, hpp_path, namespace):
     with open(hpp_path, 'w') as hpp_file:
       write_prologue(hpp_file, in_name)
       write_namespace_begin(hpp_file, scopes)
-      hpp_file.write('extern const unsigned char %s[%d];\n' % (var_name, in_size))
+      hpp_file.write('alignas(%d) extern const unsigned char %s[%d];\n' % (BLOB_ALIGNMENT, var_name, in_size))
       write_namespace_end(hpp_file, scopes)
 
 if __name__ == '__main__':