Coverage remained the same at 74.162% when pulling d1ffeb6 on spokhode/enh/py39 into b180d59 on master.

I had to disable intel channel for building because of CMake compiler check fail.
It should be investigated.

Is it perhaps better to restrict the version of cmake than to remove the intel channel. We need to be sure we are picking correct run time bits.

Is it perhaps better to restrict the version of cmake than to remove the intel channel. We need to be sure we are picking correct run time bits.

Thank for the idea. I will check that the problem is really in cmake version.

Problem is not related to CMake.
I compared py38 build log and py39. In py38 dpcpp_cpp_rt is installed in host environment. It contains libintlc.so.5 which is required by clang we use for building. Actually, dpcpp_cpp_rt is installed in build environment during building but for some reason it has no effect.
In py38 dpcpp_cpp_rt is installed in host environment because it is installed with other packages like mkl and numpy from intel channel. Now this packages are not provided by intel channel for py39 and so they are installed from defaults. Packages on defaults channel does not depend on dpcpp_cpp_rt.

@AndresGuzman-Ballen Why isn't the compiler package carrying the run-time? If run-time is not present, how can it be a viable compiler package?

@oleksandr-pavlyk Sergey already answered why this is appearing in Python 3.9 and not Python 3.8. What I imagine is happening is that dpctl is expecting the runtimes to be found in the host prefix but since {{ compiler('dpcpp') }} is a build dependency, it is being installed in the build prefix instead

@AndresGuzman-Ballen Per @PokhodenkoSA's explanation, it is the compiler that requires intlc library. The compiler is configured by cmake, and so I suspect there may be an issue with how Python driver sets paths for cmake.

dpctl Python extensions should not be built using dpcpp/clang at present, only the backend library.

I am trying to build the recipe locally.

The build environment is indeed viable, as once I activate it, I am able to build a simple C++ program that uses SYCL. This means that build environment already contains all the components in dpcpp_cpp_rt.

DPC++ launched by cmake does not see those components, because conda build isolates LD_LIBRARY_PATH to only $PREFIX subfolders, specifically: LD_LIBRARY_PATH=$PREFIX/compiler/lib/intel64_lin:$PREFIX/compiler/lib:$PREFIX/lib:.

I was able to build Python 3.9 package using conda build with the following set of changes atop the tip of master branch:

diff --git a/conda-recipe/build.sh b/conda-recipe/build.sh
index 37ae591..f2df115 100755
--- a/conda-recipe/build.sh
+++ b/conda-recipe/build.sh
@@ -3,10 +3,11 @@
 # Workaround to Klocwork overwriting LD_LIBRARY_PATH that was modified
 # by DPC++ compiler conda packages. Will need to be added to DPC++ compiler
 # activation scripts.
-export LDFLAGS="$LDFLAGS -Wl,-rpath,$CONDA_PREFIX/lib"
+export LDFLAGS="$LDFLAGS -Wl,-rpath,$PREFIX/lib"

 ${PYTHON} setup.py clean --all
-INSTALL_CMD="install --sycl-compiler-prefix=$CONDA_PREFIX"
+INSTALL_CMD="install --sycl-compiler-prefix=$BUILD_PREFIX"
+export LD_LIBRARY_PATH=$BUILD_PREFIX/lib:${LD_LIBRARY_PATH}

 if [ -n "${WHEELS_OUTPUT_FOLDER}" ]; then
     # Install packages and assemble wheel package from built bits
diff --git a/conda-recipe/meta.yaml b/conda-recipe/meta.yaml
index 77c7814..ce4da0b 100644
--- a/conda-recipe/meta.yaml
+++ b/conda-recipe/meta.yaml
@@ -22,7 +22,7 @@ requirements:
         - python
         - make  # [unix]
         - ninja  # [win]
-        - numpy >=1.17 # [win or osx or py==38]
+        - numpy >=1.17 # [win or osx or py>=38]
         - numpy 1.17 # [linux and py==37]
         - wheel
     run:

I used standard conda build --no-test --python 3.9 -c intel -c defaults --override-channels conda-recipe.

I have applied Sasha's changes. LGTM.
I will merge this PR as I need it for continuing work for other packages.
If you need other improvements, please, open another PR.

I have made job "Generate coverage data / Generate coverage and push to Coveralls.io" as not required because it fails on master too.