[SYCL][CUDA] accessor_api_image CTS test is failing

[againull]

Build compiler
git clone https://github.com/intel/llvm
Hash: b00fb7c

Includes: #1990, #1977

python /localdisk2/ws/againull/sycl/llvm/buildbot/configure.py --cuda -o /localdisk2/ws/againull/sycl/build
python /localdisk2/ws/againull/sycl/llvm/buildbot/compile.py -o /localdisk2/ws/againull/sycl/build

Build accessor CTS tests
git clone https://github.com/KhronosGroup/SYCL-CTS.git
Hash: 9cbe1a719b25c269ef78a2ee08f2e5ed12a1cc6d

Applied: KhronosGroup/SYCL-CTS#52

cmake -G Ninja -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang -DINTEL_SYCL_ROOT=/localdisk2/ws/againull/sycl/build -DINTEL_SYCL_TRIPLE=nvptx64-nvidia-cuda-sycldevice -DSYCL_IMPLEMENTATION=Intel_SYCL -DSYCL_CTS_ENABLE_OPENCL_INTEROP_TESTS=Off -DSYCL_CTS_ENABLE_DOUBLE_TESTS=On -DSYCL_CTS_ENABLE_HALF_TESTS=On -DINTEL_SYCL_FLAGS="-Xsycl-target-backend;--cuda-gpu-arch=sm_50" -DOpenCL_INCLUDE_DIR=/localdisk2/ws/againull/sycl/build/include/sycl -DOpenCL_LIBRARY=/localdisk2/ws/againull/sycl/build/lib/libOpenCL.so ..

ninja test_accessor -j 12

Run accessor_api_image CTS test
=> ./bin/test_accessor -p nvidia -d opencl_gpu --test accessor_api_image
--- accessor_api_image
. accessor<vec<int32_t, 4>, 1, mode{1024}, target{2017}>
. Checking get_range

PI CUDA ERROR:
Value: 500
Name: CUDA_ERROR_NOT_FOUND
Description: named symbol not found
Function: build_program
Source Location: /iusers/againull/sycl/llvm/sycl/plugins/cuda/pi_cuda.cpp:468

PI CUDA ERROR:
Value: 400
Name: CUDA_ERROR_INVALID_HANDLE
Description: invalid resource handle
Function: cuda_piProgramRelease
Source Location: /iusers/againull/sycl/llvm/sycl/plugins/cuda/pi_cuda.cpp:2807

. sycl exception caught
. what - The program was built for 1 devices
Build program log for 'GeForce GTX 1060 6GB':
-999 (Unknown OpenCL error code)
. line: 63
. a SYCL exception was caught: The program was built for 1 devices
Build program log for 'GeForce GTX 1060 6GB':
-999 (Unknown OpenCL error code)

• fail

--- accessor_api_image_fp16
. Device does not support half precision floating point operations

• pass

. Passed 1/14 tests (7%)

[bader]

@againull, @pvchupin, I think we figured out that this was caused by the regression in the driver. Can we close this one?

[againull]

@bader
Yes. Test passes with 435.21 nvidia driver when KhronosGroup/SYCL-CTS#52 is applied. Do you know when this PR it is going to be merged?

[pvchupin]

It looks like issue is still reproducible on latest driver 450.102.04.
Let's reopen it, at least for the tracking purpose.

[pgorlani]

The 500 CUDA_ERROR_NOT_FOUND (that turns in a 801 CUDA_ERROR_NOT_SUPPORTED for CUDA toolkit 11.3 and above) is caused by the suq.depth PTX instruction in 3d sampled readings at

llvm/libclc/ptx-nvidiacl/libspirv/images/image.cl

Line 151 in 90c8f05

int __nvvm_suq_depth(long) __asm("llvm.nvvm.suq.depth");

llvm/libclc/ptx-nvidiacl/libspirv/images/image.cl

Line 158 in 90c8f05

int __nvvm_suq_depth_3i(read_only image3d_t) __asm("llvm.nvvm.suq.depth");

NVIDIA tells that this error is expected since this PTX instructions work just in case they are used within the 'OpenCL driver'.
Unfortunately, they added that there are no ways to use this instruction with CUDA. They promised to update the PTX documentation in order to make it clear. If suq.depth is present in the fatbin, it produces the aforementioned error even if it is not actually executed. For this reason, it has been removed with #5378.

Another error emerged when the one above is factored out, a 700 CUDA_ERROR_ILLEGAL_ADDRESS. Which appeared in sampled readings with linear filtering. This has been fixed with #5204.

Unfortunately, the previous two errors are just the tip of the iceberg. Passing the accessor_api_image_core test implies the support of (u)int 8bit channels and their related conversion functions, which are completely missing right now: (u)int accessors can read (u)int_{32,16,8}b channels. Further details can be found in Section 6.12.14 and 8.3 of the OpenCL 1.2 Specification.

In order to let this test pass, the image support has been marked as experimental and deactivated by default with #5204.

In summary:

• the only type supported are (u)int (32bit), float and half,
• 8bit channels and related conversion functions are missing,
• writings and non-sampled readings are supported,
• 1d/2d sampled readings are supported,
• 3d sampled readings are not functioning due to suq.depth.

[pvchupin]

@pgorlani, thanks a lot for nice and detailed summary!

[bader]

Recent version of test_all (and test_accessor_legacy) aborts on NVIDIA GPU.

pi_die: PI CUDA kernels only support images with channel types int32, uint32, float, and half.
terminate called without an active exception
-------------------------------------------------------------------------------
accessor_api_image_core
-------------------------------------------------------------------------------
SYCL-CTS/tests/accessor_legacy/../common/../../util/proxy.h:35
...............................................................................

SYCL-CTS/tests/accessor_legacy/../common/../../util/proxy.h:35: FAILED:
due to a fatal error condition:
  SIGABRT - Abort (abnormal termination) signal

Can we change pi_die to an exception? pi_die aborts the execution of the whole test suite, whereas exception will just fail a single test.

[pgorlani]

Recent version of test_all (and test_accessor_legacy) aborts on NVIDIA GPU.

pi_die: PI CUDA kernels only support images with channel types int32, uint32, float, and half.
terminate called without an active exception
-------------------------------------------------------------------------------
accessor_api_image_core
-------------------------------------------------------------------------------
SYCL-CTS/tests/accessor_legacy/../common/../../util/proxy.h:35
...............................................................................

SYCL-CTS/tests/accessor_legacy/../common/../../util/proxy.h:35: FAILED:
due to a fatal error condition:
  SIGABRT - Abort (abnormal termination) signal

Can we change pi_die to an exception? pi_die aborts the execution of the whole test suite, whereas exception will just fail a single test.

Hi @bader, here is #6521, it should fit our needs.