COMP: Recover lost CI modernization from #74 (ITK 5.4.6, OpenCL ICD/headers 2025.07.22, OpenCL 3.0)

.github/workflows/test-gpu.yml

@@ -1,19 +1,23 @@
 name: Test GPU
 
-on: [pull_request]
+on:
+  workflow_dispatch:
+  pull_request:
+    types: [opened, synchronize, reopened, labeled]
 
 env:
   itk-git-tag: "abf5fa10522a36bc51f42f20f426a622f42ed90d"
 
 jobs:
   build-test-gpu:
+    if: github.event_name == 'workflow_dispatch' || contains(github.event.pull_request.labels.*.name, 'gpu-ci')
     runs-on: [self-hosted, gpu]
     strategy:
       matrix:
         include:
           - opencl-icd-loader-git-tag: "v2021.04.29"
             opencl-headers-git-tag: "v2021.04.29"
-            opencl-version: "120"
+            opencl-version: "300"
             vkfft-backend: 3
             cmake-build-type: "MinSizeRel"
             platform-name: "ubuntu-nvidia-gpu"

.github/workflows/test-notebooks.yml

@@ -1,9 +1,13 @@
 name: Notebook tests
 
-on: [push, pull_request]
+on:
+  workflow_dispatch:
+  pull_request:
+    types: [opened, synchronize, reopened, labeled]
 
 jobs:
   nbmake:
+    if: github.event_name == 'workflow_dispatch' || contains(github.event.pull_request.labels.*.name, 'gpu-ci')
     runs-on: [self-hosted, notebook-gpu]
     name: Test notebooks with nbmake
     strategy:

CMakeLists.txt

@@ -22,7 +22,8 @@ if(VKFFT_BACKEND EQUAL 1)
   find_package(CUDAToolkit REQUIRED)
   list(APPEND VkFFTBackend_SYSTEM_INCLUDE_DIRS ${CUDAToolkit_INCLUDE_DIRS})
 elseif(VKFFT_BACKEND EQUAL 3)
-  add_compile_definitions(CL_TARGET_OPENCL_VERSION=120)
+  set(CL_TARGET_OPENCL_VERSION 300 CACHE STRING "OpenCL API version to target (e.g. 120, 300)")
+  add_compile_definitions(CL_TARGET_OPENCL_VERSION=${CL_TARGET_OPENCL_VERSION})
 
   ## When this module is loaded by an app, load OpenCL too.
   set(VkFFTBackend_EXPORT_CODE_INSTALL "
@@ -53,7 +54,10 @@ elseif(VKFFT_BACKEND EQUAL 4)
   if(NOT LevelZero_INCLUDE_DIR OR NOT LevelZero_LIBRARY)
     message(FATAL_ERROR "VKFFT_BACKEND=4 (Level Zero) requires the oneAPI Level Zero loader (ze_loader) and headers (level_zero/ze_api.h). Install the 'level-zero' package or set LEVEL_ZERO_ROOT.")
   endif()
-  list(APPEND VkFFTBackend_SYSTEM_INCLUDE_DIRS ${LevelZero_INCLUDE_DIR})
+  # VkFFT includes <ze_api.h> bare; this module uses <level_zero/ze_api.h> — both dirs needed.
+  list(APPEND VkFFTBackend_SYSTEM_INCLUDE_DIRS
+    ${LevelZero_INCLUDE_DIR}
+    ${LevelZero_INCLUDE_DIR}/level_zero)
   list(APPEND VkFFTBackend_SYSTEM_LIBRARIES ${LevelZero_LIBRARY})
 
   set(VkFFTBackend_EXPORT_CODE_INSTALL "

include/itkVkDefinitions.h

@@ -26,4 +26,12 @@
 #define LEVEL_ZERO 4
 #define METAL 5
 
+// Defensive default: when VKFFT_BACKEND is not set on the command line
+// (e.g. castxml wrapping invocations that bypass our top-level
+// add_compile_definitions), fall back to OpenCL so vkFFT.h does not
+// take the Vulkan branch and try to #include <vulkan/vulkan.h>.
+#ifndef VKFFT_BACKEND
+#  define VKFFT_BACKEND OPENCL
+#endif
+
 #endif // itkVkDefinitions_h

itk-module-init.cmake

@@ -8,8 +8,14 @@ if(${VKFFT_BACKEND} EQUAL 1)
 elseif(${VKFFT_BACKEND} EQUAL 3)
   find_package(OpenCL REQUIRED)
 elseif(${VKFFT_BACKEND} EQUAL 4)
-  find_path(LevelZero_INCLUDE_DIR NAMES level_zero/ze_api.h)
-  find_library(LevelZero_LIBRARY NAMES ze_loader)
+  find_path(LevelZero_INCLUDE_DIR
+    NAMES level_zero/ze_api.h
+    HINTS ENV LEVEL_ZERO_ROOT ENV CMPLR_ROOT
+    PATH_SUFFIXES include)
+  find_library(LevelZero_LIBRARY
+    NAMES ze_loader
+    HINTS ENV LEVEL_ZERO_ROOT ENV CMPLR_ROOT
+    PATH_SUFFIXES lib lib64 lib/x64)
   if(NOT LevelZero_INCLUDE_DIR OR NOT LevelZero_LIBRARY)
     message(FATAL_ERROR "VKFFT_BACKEND=4 (Level Zero) requires the oneAPI Level Zero loader (ze_loader) and headers (level_zero/ze_api.h).")
   endif()

src/itkVkCommon.cxx

@@ -114,8 +114,22 @@ VkCommon::ConfigureBackend()
   uint64_t k{ 0 };
   for (uint64_t j{ 0 }; j < numPlatforms; j++)
   {
-    cl_uint numDevices;
+    // First probe: how many devices does this platform expose? An OpenCL
+    // platform with zero compute devices is legitimate (e.g. Apple's
+    // deprecated OpenCL framework on macOS 15 returns CL_DEVICE_NOT_FOUND
+    // for CL_DEVICE_TYPE_ALL). Skip such platforms; calling clGetDeviceIDs
+    // again with num_entries=0 would return CL_INVALID_VALUE.
+    cl_uint numDevices{ 0 };
     resCL = clGetDeviceIDs(platforms[j], CL_DEVICE_TYPE_ALL, 0, nullptr, &numDevices);
+    if (resCL == CL_DEVICE_NOT_FOUND || numDevices == 0)
+    {
+      continue;
+    }
+    if (resCL != CL_SUCCESS)
+    {
+      std::cerr << __FILE__ "(" << __LINE__ << "): clGetDeviceIDs(count) returned " << resCL << std::endl;
+      return VkFFTResult{ VKFFT_ERROR_FAILED_TO_GET_DEVICE };
+    }
     std::unique_ptr<cl_device_id[]> deviceListArray{ std::make_unique<cl_device_id[]>(numDevices) };
     cl_device_id *                  deviceList{ &deviceListArray[0] };
     if (!deviceList)

test/CMakeLists.txt

@@ -170,6 +170,17 @@ itk_add_test(NAME itkVkForwardInverseFFTImageFilterTestDouble
   )
 _vkfft_disable_on_unsupported_fp64(itkVkForwardInverseFFTImageFilterTestDouble)
 
+# pocl (CPU OpenCL) computes VkFFT's size-19 Bluestein inverse incorrectly, so
+# the hosted pocl-backed leg runs the round-trip tests capped at size 16
+# (radix-2/3/5/7 plus Bluestein primes 11/13). Real GPUs run the full sweep above.
+itk_add_test(NAME itkVkForwardInverseFFTImageFilterTestFloatPoclSafe
+  COMMAND VkFFTBackendTestDriver itkVkForwardInverseFFTImageFilterTest float 16
+  )
+itk_add_test(NAME itkVkForwardInverseFFTImageFilterTestDoublePoclSafe
+  COMMAND VkFFTBackendTestDriver itkVkForwardInverseFFTImageFilterTest double 16
+  )
+_vkfft_disable_on_unsupported_fp64(itkVkForwardInverseFFTImageFilterTestDoublePoclSafe)
+
 # -----------------------------------------------------------------------------
 # ForwardInverse1DFFTImageFilterTest
 # -----------------------------------------------------------------------------
@@ -181,6 +192,14 @@ itk_add_test(NAME itkVkForwardInverse1DFFTImageFilterTestDouble
   )
 _vkfft_disable_on_unsupported_fp64(itkVkForwardInverse1DFFTImageFilterTestDouble)
 
+itk_add_test(NAME itkVkForwardInverse1DFFTImageFilterTestFloatPoclSafe
+  COMMAND VkFFTBackendTestDriver itkVkForwardInverse1DFFTImageFilterTest float 16
+  )
+itk_add_test(NAME itkVkForwardInverse1DFFTImageFilterTestDoublePoclSafe
+  COMMAND VkFFTBackendTestDriver itkVkForwardInverse1DFFTImageFilterTest double 16
+  )
+_vkfft_disable_on_unsupported_fp64(itkVkForwardInverse1DFFTImageFilterTestDoublePoclSafe)
+
 # -----------------------------------------------------------------------------
 # HalfHermitianFFTImageFilterTest
 # -----------------------------------------------------------------------------
@@ -192,6 +211,14 @@ itk_add_test(NAME itkVkHalfHermitianFFTImageFilterTestDouble
   )
 _vkfft_disable_on_unsupported_fp64(itkVkHalfHermitianFFTImageFilterTestDouble)
 
+itk_add_test(NAME itkVkHalfHermitianFFTImageFilterTestFloatPoclSafe
+  COMMAND VkFFTBackendTestDriver itkVkHalfHermitianFFTImageFilterTest float 16
+  )
+itk_add_test(NAME itkVkHalfHermitianFFTImageFilterTestDoublePoclSafe
+  COMMAND VkFFTBackendTestDriver itkVkHalfHermitianFFTImageFilterTest double 16
+  )
+_vkfft_disable_on_unsupported_fp64(itkVkHalfHermitianFFTImageFilterTestDoublePoclSafe)
+
 # -----------------------------------------------------------------------------
 # FFTImageFilterFactoryTest (instantiation only — runs on all platforms)
 # -----------------------------------------------------------------------------
@@ -356,7 +383,7 @@ if(VKFFT_BACKEND EQUAL 4 AND NOT VkFFTBackend_LEVEL_ZERO_RUNTIME_AVAILABLE)
     itkVkDiscreteGaussianImageFilterTest
     itkVkDiscreteGaussianImageFilterTest2
   )
-    foreach(_suffix Float Double)
+    foreach(_suffix Float Double FloatPoclSafe DoublePoclSafe)
       if(TEST ${_stem}${_suffix})
         set_tests_properties(${_stem}${_suffix} PROPERTIES DISABLED TRUE)
       endif()

test/itkVkForwardInverse1DFFTImageFilterTest.cxx

@@ -57,7 +57,7 @@ class ShowProgress : public itk::Command
 
 template <typename PrecisionType>
 int
-runVkForwardInverse1DFFTImageFilterTest()
+runVkForwardInverse1DFFTImageFilterTest(unsigned int maxSize = 20)
 {
   int  testNumber{ 0 };
   bool testsPassed{ true };
@@ -72,7 +72,7 @@ runVkForwardInverse1DFFTImageFilterTest()
     bool                              firstPass{ true };
 
     // Skip trivial case where 1D image of size 1 fails.
-    for (unsigned int mySize{ 2 }; mySize <= 20; ++mySize)
+    for (unsigned int mySize{ 2 }; mySize <= maxSize; ++mySize)
     {
       // We expect that anything evenly divisible by a prime number greater than 13
       // will succeed with Bluestein's Algorithm implementation in VkFFT, though
@@ -184,14 +184,15 @@ runVkForwardInverse1DFFTImageFilterTest()
 int
 itkVkForwardInverse1DFFTImageFilterTest(int argc, char * argv[])
 {
-  const std::string precision{ (argc > 1) ? argv[1] : "float" };
+  const std::string  precision{ (argc > 1) ? argv[1] : "float" };
+  const unsigned int maxSize{ (argc > 2) ? static_cast<unsigned int>(std::stoul(argv[2])) : 20 };
   if (precision == "double")
   {
-    return runVkForwardInverse1DFFTImageFilterTest<double>();
+    return runVkForwardInverse1DFFTImageFilterTest<double>(maxSize);
   }
   if (precision == "float")
   {
-    return runVkForwardInverse1DFFTImageFilterTest<float>();
+    return runVkForwardInverse1DFFTImageFilterTest<float>(maxSize);
   }
   std::cerr << "Unknown precision '" << precision << "'. Expected 'float' or 'double'." << std::endl;
   return EXIT_FAILURE;

test/itkVkForwardInverseFFTImageFilterTest.cxx

@@ -57,7 +57,7 @@ class ShowProgress : public itk::Command
 
 template <typename PrecisionType>
 int
-runVkForwardInverseFFTImageFilterTest()
+runVkForwardInverseFFTImageFilterTest(unsigned int maxSize = 20)
 {
   int  testNumber{ 0 };
   bool testsPassed{ true };
@@ -72,7 +72,7 @@ runVkForwardInverseFFTImageFilterTest()
     bool                              firstPass{ true };
 
     // Skip trivial case where 1D image of size 1 fails.
-    for (unsigned int mySize{ 2 }; mySize <= 20; ++mySize, firstPass = false)
+    for (unsigned int mySize{ 2 }; mySize <= maxSize; ++mySize, firstPass = false)
     {
       // We expect that anything evenly divisible by a prime number greater than 13
       // will succeed with Bluestein's Algorithm implementation in VkFFT, though
@@ -185,14 +185,15 @@ runVkForwardInverseFFTImageFilterTest()
 int
 itkVkForwardInverseFFTImageFilterTest(int argc, char * argv[])
 {
-  const std::string precision{ (argc > 1) ? argv[1] : "float" };
+  const std::string  precision{ (argc > 1) ? argv[1] : "float" };
+  const unsigned int maxSize{ (argc > 2) ? static_cast<unsigned int>(std::stoul(argv[2])) : 20 };
   if (precision == "double")
   {
-    return runVkForwardInverseFFTImageFilterTest<double>();
+    return runVkForwardInverseFFTImageFilterTest<double>(maxSize);
   }
   if (precision == "float")
   {
-    return runVkForwardInverseFFTImageFilterTest<float>();
+    return runVkForwardInverseFFTImageFilterTest<float>(maxSize);
   }
   std::cerr << "Unknown precision '" << precision << "'. Expected 'float' or 'double'." << std::endl;
   return EXIT_FAILURE;

test/itkVkHalfHermitianFFTImageFilterTest.cxx

@@ -57,7 +57,7 @@ class ShowProgress : public itk::Command
 
 template <typename PrecisionType>
 int
-runVkHalfHermitianFFTImageFilterTest()
+runVkHalfHermitianFFTImageFilterTest(unsigned int maxSize = 20)
 {
   int  testNumber{ 0 };
   bool testsPassed{ true };
@@ -71,7 +71,7 @@ runVkHalfHermitianFFTImageFilterTest()
     typename RealImageType::IndexType index;
     bool                              firstPass{ true };
     // Skip trivial case where 1D image of size 1 fails.
-    for (unsigned int mySize{ 2 }; mySize <= 20; ++mySize, firstPass = false)
+    for (unsigned int mySize{ 2 }; mySize <= maxSize; ++mySize, firstPass = false)
     {
       // We expect that anything evenly divisible by a prime number greater than 13
       // will succeed with Bluestein's Algorithm implementation in VkFFT, though
@@ -189,14 +189,15 @@ runVkHalfHermitianFFTImageFilterTest()
 int
 itkVkHalfHermitianFFTImageFilterTest(int argc, char * argv[])
 {
-  const std::string precision{ (argc > 1) ? argv[1] : "float" };
+  const std::string  precision{ (argc > 1) ? argv[1] : "float" };
+  const unsigned int maxSize{ (argc > 2) ? static_cast<unsigned int>(std::stoul(argv[2])) : 20 };
   if (precision == "double")
   {
-    return runVkHalfHermitianFFTImageFilterTest<double>();
+    return runVkHalfHermitianFFTImageFilterTest<double>(maxSize);
   }
   if (precision == "float")
   {
-    return runVkHalfHermitianFFTImageFilterTest<float>();
+    return runVkHalfHermitianFFTImageFilterTest<float>(maxSize);
   }
   std::cerr << "Unknown precision '" << precision << "'. Expected 'float' or 'double'." << std::endl;
   return EXIT_FAILURE;

wrapping/dockcross-manylinux-build-module-wheels-opencl.sh

@@ -14,9 +14,12 @@
 # Generate dockcross scripts
 
 MANYLINUX_VERSION=${MANYLINUX_VERSION:=_2_28}
-IMAGE_TAG=${IMAGE_TAG:=20221108-102ebcc}
-OPENCL_ICD_LOADER_TAG=v2021.04.29
-OPENCL_HEADERS_TAG=v2021.04.29
+# Match ITKPythonPackage v6.0b02's dockcross image, which ships a recent
+# enough `python -m build` to accept --verbose (the older 20221108 image
+# fails with "unrecognized arguments: --verbose").
+IMAGE_TAG=${IMAGE_TAG:=20260203-3dfb3ff}
+OPENCL_ICD_LOADER_TAG=${OPENCL_ICD_LOADER_TAG:=v2025.07.22}
+OPENCL_HEADERS_TAG=${OPENCL_HEADERS_TAG:=v2025.07.22}
 
 docker run --rm dockcross/manylinux${MANYLINUX_VERSION}-x64:${IMAGE_TAG} > /tmp/dockcross-manylinux-x64
 chmod u+x /tmp/dockcross-manylinux-x64
@@ -49,10 +52,22 @@ if ! test -d ./OpenCL-ICD-Loader; then
 fi
 
 # Build wheels
+#
+# Resolve the actual versioned libOpenCL.so filename produced by the
+# OpenCL-ICD-Loader build. v2021.04.29 produced libOpenCL.so.1.2; the
+# v2025.07.22 build produces libOpenCL.so.1.0.0. Bind-mount whichever the
+# loader actually wrote.
+OPENCL_SO=$(ls $(pwd)/OpenCL-ICD-Loader-build/libOpenCL.so.1.* 2>/dev/null | head -1)
+if [[ -z "${OPENCL_SO}" ]]; then
+  echo "ERROR: could not find OpenCL-ICD-Loader-build/libOpenCL.so.1.* — did the loader build?" >&2
+  exit 1
+fi
+echo "Using OpenCL loader: ${OPENCL_SO}"
+
 DOCKER_ARGS="-v $(pwd)/dist:/work/dist/ -v $script_dir/../ITKPythonPackage:/ITKPythonPackage -v $(pwd)/tools:/tools"
 DOCKER_ARGS+=" -v $(pwd)/OpenCL-ICD-Loader/inc/CL:/usr/include/CL"
-DOCKER_ARGS+=" -v $(pwd)/OpenCL-ICD-Loader-build/libOpenCL.so.1.2:/usr/lib64/libOpenCL.so.1"
-DOCKER_ARGS+=" -v $(pwd)/OpenCL-ICD-Loader-build/libOpenCL.so.1.2:/usr/lib64/libOpenCL.so"
+DOCKER_ARGS+=" -v ${OPENCL_SO}:/usr/lib64/libOpenCL.so.1"
+DOCKER_ARGS+=" -v ${OPENCL_SO}:/usr/lib64/libOpenCL.so"
 DOCKER_ARGS+=" -e MANYLINUX_VERSION"
 
 /tmp/dockcross-manylinux-x64 \