arrayfire · umar456 · Apr 3, 2023 · Mar 31, 2023 · Apr 2, 2023 · Apr 2, 2023
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -45,8 +45,8 @@ if(AF_WITH_EXTERNAL_PACKAGES_ONLY)
  set(AF_REQUIRED REQUIRED)
 endif()

-#Set Intel OpenMP as default MKL thread layer
-if(CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM" OR CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
+if(CXX_COMPILER_NAME STREQUAL "dpcpp" OR CXX_COMPILER_NAME STREQUAL "dpcpp.exe"
+   OR CXX_COMPILER_NAME STREQUAL "icpx" OR CXX_COMPILER_NAME STREQUAL "icx.exe")
  set(MKL_THREAD_LAYER "TBB" CACHE STRING "The thread layer to choose for MKL")
  set(MKL_INTERFACE "ilp64")
  set(MKL_INTERFACE_INTEGER_SIZE 8)
@@ -125,7 +125,14 @@ elseif(MKL_THREAD_LAYER STREQUAL "TBB")
  set(MKL_THREADING "tbb_thread")
 else()
 endif()
-find_package(MKL)
+
+if(CMAKE_VERSION VERSION_GREATER_EQUAL 3.13)
+  # VCPKG overrides the find_package command and the PATH parameter is currently
+  # broken with the current version of VCPKG so we are setting the MKL_ROOT
+  # directory to the MKLROOT environment variable.
+  set(MKL_ROOT "$ENV{MKLROOT}")
+  find_package(MKL)
+endif()

 af_multiple_option(NAME        AF_COMPUTE_LIBRARY
                   DEFAULT     ${default_compute_library}
@@ -218,7 +225,7 @@ if(${AF_BUILD_CPU} OR ${AF_BUILD_OPENCL})
  if("${AF_COMPUTE_LIBRARY}" STREQUAL "Intel-MKL"
      OR "${AF_COMPUTE_LIBRARY}" STREQUAL "MKL")
    af_mkl_batch_check()
-    dependency_check(MKL_FOUND "Please ensure Intel-MKL / oneAPI-oneMKL is installed")
+    dependency_check(MKL_Shared_FOUND "Please ensure Intel-MKL / oneAPI-oneMKL is installed")
    set(BUILD_WITH_MKL ON)
  elseif("${AF_COMPUTE_LIBRARY}" STREQUAL "FFTW/LAPACK/BLAS")
    dependency_check(FFTW_FOUND "FFTW not found")

diff --git a/src/backend/opencl/Array.cpp b/src/backend/opencl/Array.cpp
@@ -334,11 +334,15 @@ kJITHeuristics passesJitHeuristics(span<Node *> root_nodes) {
            (3 * sizeof(uint));

        const cl::Device &device = getDevice();
-        size_t max_param_size = device.getInfo<CL_DEVICE_MAX_PARAMETER_SIZE>();
        // typical values:
        //   NVIDIA     = 4096
        //   AMD        = 3520  (AMD A10 iGPU = 1024)
        //   Intel iGPU = 1024
+        //
+        // Setting the maximum to 5120 bytes to keep the compile times
+        // resonable. This still results in large kernels but its not excessive.
+        size_t max_param_size =
+            min(5120UL, device.getInfo<CL_DEVICE_MAX_PARAMETER_SIZE>());
        max_param_size -= base_param_size;

        struct tree_info {

diff --git a/src/backend/opencl/compile_module.cpp b/src/backend/opencl/compile_module.cpp
@@ -230,7 +230,10 @@ Module loadModuleFromDisk(const int device, const string &moduleKey,
    try {
        std::ifstream in(cacheFile, std::ios::binary);
        if (!in.is_open()) {
-            AF_ERROR("Unable to open binary cache file", AF_ERR_INTERNAL);
+            AF_TRACE("{{{:<20} : Unable to open {} for {}}}", moduleKey,
+                     cacheFile, dev.getInfo<CL_DEVICE_NAME>());
+            removeFile(cacheFile);
+            return retVal;
        }
        in.exceptions(std::ios::failbit | std::ios::badbit);

@@ -247,24 +250,18 @@ Module loadModuleFromDisk(const int device, const string &moduleKey,
        const size_t recomputedHash =
            deterministicHash(clbin.data(), clbinSize);
        if (recomputedHash != clbinHash) {
-            AF_ERROR("Binary on disk seems to be corrupted", AF_ERR_LOAD_SYM);
+            AF_TRACE(
+                "{{{:<20} : Corrupt binary({}) found on disk for {}, removed}}",
+                moduleKey, cacheFile, dev.getInfo<CL_DEVICE_NAME>());
+            removeFile(cacheFile);
+            return retVal;
        }
        program = Program(arrayfire::opencl::getContext(), {dev}, {clbin});
        program.build();

        AF_TRACE("{{{:<20} : loaded from {} for {} }}", moduleKey, cacheFile,
                 dev.getInfo<CL_DEVICE_NAME>());
        retVal.set(program);
-    } catch (const AfError &e) {
-        if (e.getError() == AF_ERR_LOAD_SYM) {
-            AF_TRACE(
-                "{{{:<20} : Corrupt binary({}) found on disk for {}, removed}}",
-                moduleKey, cacheFile, dev.getInfo<CL_DEVICE_NAME>());
-        } else {
-            AF_TRACE("{{{:<20} : Unable to open {} for {}}}", moduleKey,
-                     cacheFile, dev.getInfo<CL_DEVICE_NAME>());
-        }
-        removeFile(cacheFile);
    } catch (const std::ios_base::failure &e) {
        AF_TRACE("{{{:<20} : IO failure while loading {} for {}; {}}}",
                 moduleKey, cacheFile, dev.getInfo<CL_DEVICE_NAME>(), e.what());

diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
@@ -76,7 +76,9 @@ function(af_add_test target backend is_serial)
    if(${is_serial})
      set_tests_properties(${target}
        PROPERTIES
-        RUN_SERIAL ON)
+          ENVIRONMENT AF_PRINT_ERRORS=1
+          TIMEOUT 900
+          RUN_SERIAL ON)
    endif(${is_serial})
  endif()
 endfunction()