[SYCL] Improve error reporting for kernel enqueue (#860)

AlexeySachkov · bader · commit d27cff25dcf4 · 2019-11-25T23:37:10.000+03:00
The idea is to handle error code returned from PI and based on input
arguments, emit meaningful exception with error explanation.
One of the side-effects of this is that this patch effectively allows to
use non-uniform work-groups if underlying OpenCL supports this
functionality.

Signed-off-by: Alexey Sachkov &lt;alexey.sachkov@intel.com&gt;
diff --git a/sycl/include/CL/sycl/detail/cg.hpp b/sycl/include/CL/sycl/detail/cg.hpp
@@ -237,7 +237,7 @@ class HostKernel : public HostKernelBase {
     for (int I = 0; I < Dims; ++I) {
       if (NDRDesc.LocalSize[I] == 0 ||
           NDRDesc.GlobalSize[I] % NDRDesc.LocalSize[I] != 0)
-        throw sycl::runtime_error("Invalid local size for global size");
+        throw sycl::nd_range_error("Invalid local size for global size");
       GroupSize[I] = NDRDesc.GlobalSize[I] / NDRDesc.LocalSize[I];
     }
 
@@ -278,7 +278,7 @@ class HostKernel : public HostKernelBase {
     for (int I = 0; I < Dims; ++I) {
       if (NDRDesc.LocalSize[I] == 0 ||
           NDRDesc.GlobalSize[I] % NDRDesc.LocalSize[I] != 0)
-        throw sycl::runtime_error("Invalid local size for global size");
+        throw sycl::nd_range_error("Invalid local size for global size");
       NGroups[I] = NDRDesc.GlobalSize[I] / NDRDesc.LocalSize[I];
     }
 
@@ -385,15 +385,6 @@ class CGExecKernel : public CG {
         MStreams(std::move(Streams)) {
     assert((getType() == RUN_ON_HOST_INTEL || getType() == KERNEL) &&
            "Wrong type of exec kernel CG.");
-
-    if (MNDRDesc.LocalSize.size() > 0) {
-      range<3> Excess = (MNDRDesc.GlobalSize % MNDRDesc.LocalSize);
-      for (int I = 0; I < 3; I++) {
-        if (Excess[I] != 0)
-          throw nd_range_error("Global size is not a multiple of local size",
-              CL_INVALID_WORK_GROUP_SIZE);
-      }
-    }
   }
 
   std::vector<ArgDesc> getArguments() const { return MArgs; }
diff --git a/sycl/include/CL/sycl/detail/pi.h b/sycl/include/CL/sycl/detail/pi.h
@@ -62,7 +62,8 @@ typedef enum {
   PI_INVALID_DEVICE = CL_INVALID_DEVICE,
   PI_INVALID_BINARY = CL_INVALID_BINARY,
   PI_MISALIGNED_SUB_BUFFER_OFFSET = CL_MISALIGNED_SUB_BUFFER_OFFSET,
-  PI_OUT_OF_HOST_MEMORY = CL_OUT_OF_HOST_MEMORY
+  PI_OUT_OF_HOST_MEMORY = CL_OUT_OF_HOST_MEMORY,
+  PI_INVALID_WORK_GROUP_SIZE = CL_INVALID_WORK_GROUP_SIZE
 } _pi_result;
 
 typedef enum {
@@ -88,7 +89,9 @@ typedef enum {
   PI_DEVICE_INFO_PARENT         = CL_DEVICE_PARENT_DEVICE,
   PI_DEVICE_INFO_PLATFORM       = CL_DEVICE_PLATFORM,
   PI_DEVICE_INFO_PARTITION_TYPE = CL_DEVICE_PARTITION_TYPE,
-  PI_DEVICE_INFO_NAME           = CL_DEVICE_NAME
+  PI_DEVICE_INFO_NAME           = CL_DEVICE_NAME,
+  PI_DEVICE_VERSION             = CL_DEVICE_VERSION,
+  PI_DEVICE_MAX_WORK_GROUP_SIZE = CL_DEVICE_MAX_WORK_GROUP_SIZE
 } _pi_device_info;
 
 // TODO: populate
diff --git a/sycl/source/CMakeLists.txt b/sycl/source/CMakeLists.txt
@@ -46,6 +46,7 @@ set(SYCL_SOURCES
     "detail/context_impl.cpp"
     "detail/device_impl.cpp"
     "detail/device_info.cpp"
+    "detail/error_handling/enqueue_kernel.cpp"
     "detail/event_impl.cpp"
     "detail/force_device.cpp"
     "detail/helpers.cpp"
diff --git a/sycl/source/detail/error_handling/enqueue_kernel.cpp b/sycl/source/detail/error_handling/enqueue_kernel.cpp
@@ -0,0 +1,183 @@
+//===------------------- enqueue_kernel.cpp ---------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// SYCL error handling of enqueue kernel operations
+//
+//===----------------------------------------------------------------------===//
+
+#include "error_handling.hpp"
+
+#include <CL/sycl/detail/pi.hpp>
+
+namespace cl {
+namespace sycl {
+namespace detail {
+
+namespace enqueue_kernel_launch {
+
+bool handleInvalidWorkGroupSize(pi_device Device, pi_kernel Kernel,
+                                const NDRDescT &NDRDesc) {
+  const bool HasLocalSize = (NDRDesc.LocalSize[0] != 0);
+
+  size_t VerSize = 0;
+  PI_CALL(piDeviceGetInfo)(Device, PI_DEVICE_VERSION, 0, nullptr, &VerSize);
+  assert(VerSize >= 10 &&
+         "Unexpected device version string"); // strlen("OpenCL X.Y")
+  string_class VerStr(VerSize, '\0');
+  PI_CALL(piDeviceGetInfo)(Device, PI_DEVICE_VERSION, VerSize, &VerStr.front(),
+                           nullptr);
+  const char *Ver = &VerStr[7]; // strlen("OpenCL ")
+
+  size_t CompileWGSize[3] = {0};
+  PI_CALL(piKernelGetGroupInfo)(Kernel, Device,
+                                CL_KERNEL_COMPILE_WORK_GROUP_SIZE,
+                                sizeof(size_t) * 3, CompileWGSize, nullptr);
+
+  if (CompileWGSize[0] != 0) {
+    // OpenCL 1.x && 2.0:
+    // CL_INVALID_WORK_GROUP_SIZE if local_work_size is NULL and the
+    // reqd_work_group_size attribute is used to declare the work-group size
+    // for kernel in the program source.
+    if (!HasLocalSize && (Ver[0] == '1' || (Ver[0] == '2' && Ver[2] == '0')))
+      throw sycl::nd_range_error(
+          "OpenCL 1.x and 2.0 requires to pass local size argument even if "
+          "required work-group size was specified in the program source",
+          PI_INVALID_WORK_GROUP_SIZE);
+
+    // Any OpenCL version:
+    // CL_INVALID_WORK_GROUP_SIZE if local_work_size is specified and does not
+    // match the required work-group size for kernel in the program source.
+    if (NDRDesc.LocalSize[0] != CompileWGSize[0] ||
+        NDRDesc.LocalSize[1] != CompileWGSize[1] ||
+        NDRDesc.LocalSize[2] != CompileWGSize[2])
+      throw sycl::nd_range_error(
+          "Specified local size doesn't match the required work-group size "
+          "specified in the program source",
+          PI_INVALID_WORK_GROUP_SIZE);
+  }
+
+  if (Ver[0] == '1') {
+    // OpenCL 1.x:
+    // CL_INVALID_WORK_GROUP_SIZE if local_work_size is specified and the
+    // total number of work-items in the work-group computed as
+    // local_work_size[0] * ... * local_work_size[work_dim – 1] is greater
+    // than the value specified by CL_DEVICE_MAX_WORK_GROUP_SIZE in
+    // table 4.3
+    size_t MaxWGSize = 0;
+    PI_CALL(piDeviceGetInfo)(Device, PI_DEVICE_MAX_WORK_GROUP_SIZE,
+                             sizeof(size_t), &MaxWGSize, nullptr);
+    const size_t TotalNumberOfWIs =
+        NDRDesc.LocalSize[0] * NDRDesc.LocalSize[1] * NDRDesc.LocalSize[2];
+    if (TotalNumberOfWIs > MaxWGSize)
+      throw sycl::nd_range_error(
+          "Total number of work-items in a work-group cannot exceed "
+          "info::device::max_work_group_size which is equal to " +
+              std::to_string(MaxWGSize),
+          PI_INVALID_WORK_GROUP_SIZE);
+  } else {
+    // OpenCL 2.x:
+    // CL_INVALID_WORK_GROUP_SIZE if local_work_size is specified and the
+    // total number of work-items in the work-group computed as
+    // local_work_size[0] * ... * local_work_size[work_dim – 1] is greater
+    // than the value specified by CL_KERNEL_WORK_GROUP_SIZE in table 5.21.
+    size_t KernelWGSize = 0;
+    PI_CALL(piKernelGetGroupInfo)(Kernel, Device, CL_KERNEL_WORK_GROUP_SIZE,
+                                  sizeof(size_t), &KernelWGSize, nullptr);
+    const size_t TotalNumberOfWIs =
+        NDRDesc.LocalSize[0] * NDRDesc.LocalSize[1] * NDRDesc.LocalSize[2];
+    if (TotalNumberOfWIs > KernelWGSize)
+      throw sycl::nd_range_error(
+          "Total number of work-items in a work-group cannot exceed "
+          "info::kernel_work_group::work_group_size which is equal to " +
+              std::to_string(KernelWGSize) + " for this kernel",
+          PI_INVALID_WORK_GROUP_SIZE);
+  }
+
+  if (HasLocalSize) {
+    const bool NonUniformWGs =
+        (NDRDesc.LocalSize[0] != 0 &&
+         NDRDesc.GlobalSize[0] % NDRDesc.LocalSize[0] != 0) ||
+        (NDRDesc.LocalSize[1] != 0 &&
+         NDRDesc.GlobalSize[1] % NDRDesc.LocalSize[1] != 0) ||
+        (NDRDesc.LocalSize[2] != 0 &&
+         NDRDesc.GlobalSize[2] % NDRDesc.LocalSize[2] != 0);
+
+    if (Ver[0] == '1') {
+      // OpenCL 1.x:
+      // CL_INVALID_WORK_GROUP_SIZE if local_work_size is specified and
+      // number of workitems specified by global_work_size is not evenly
+      // divisible by size of work-group given by local_work_size
+
+      if (NonUniformWGs)
+        throw sycl::nd_range_error(
+            "Non-uniform work-groups are not supported by the target device",
+            PI_INVALID_WORK_GROUP_SIZE);
+    } else {
+      // OpenCL 2.x:
+      // CL_INVALID_WORK_GROUP_SIZE if the program was compiled with
+      // –cl-uniform-work-group-size and the number of work-items specified
+      // by global_work_size is not evenly divisible by size of work-group
+      // given by local_work_size
+
+      pi_program Program = nullptr;
+      PI_CALL(piKernelGetInfo)(Kernel, CL_KERNEL_PROGRAM, sizeof(pi_program),
+                               &Program, nullptr);
+      size_t OptsSize = 0;
+      PI_CALL(piProgramGetBuildInfo)(Program, Device, CL_PROGRAM_BUILD_OPTIONS,
+                                     0, nullptr, &OptsSize);
+      string_class Opts(OptsSize, '\0');
+      PI_CALL(piProgramGetBuildInfo)(Program, Device, CL_PROGRAM_BUILD_OPTIONS,
+                                     OptsSize, &Opts.front(), nullptr);
+      if (NonUniformWGs) {
+        const bool HasStd20 = Opts.find("-cl-std=CL2.0") != string_class::npos;
+        if (!HasStd20)
+          throw sycl::nd_range_error(
+              "Non-uniform work-groups are not allowed by default. Underlying "
+              "OpenCL 2.x implementation supports this feature and to enable "
+              "it, build device program with -cl-std=CL2.0",
+              PI_INVALID_WORK_GROUP_SIZE);
+        else
+          throw sycl::nd_range_error(
+              "Non-uniform work-groups are not allowed by default. Underlying "
+              "OpenCL 2.x implementation supports this feature, but it is "
+              "disabled by -cl-uniform-work-group-size build flag",
+              PI_INVALID_WORK_GROUP_SIZE);
+      }
+    }
+  }
+
+  // TODO: required number of sub-groups, OpenCL 2.1:
+  // CL_INVALID_WORK_GROUP_SIZE if local_work_size is specified and is not
+  // consistent with the required number of sub-groups for kernel in the
+  // program source.
+
+  // Fallback
+  constexpr pi_result Error = PI_INVALID_WORK_GROUP_SIZE;
+  throw runtime_error(
+      "OpenCL API failed. OpenCL API returns: " + codeToString(Error), Error);
+}
+
+bool handleError(pi_result Error, pi_device Device, pi_kernel Kernel,
+                 const NDRDescT &NDRDesc) {
+  assert(Error != PI_SUCCESS &&
+         "Success is expected to be handled on caller side");
+  switch (Error) {
+  case PI_INVALID_WORK_GROUP_SIZE:
+    return handleInvalidWorkGroupSize(Device, Kernel, NDRDesc);
+  // TODO: Handle other error codes
+  default:
+    throw runtime_error(
+        "OpenCL API failed. OpenCL API returns: " + codeToString(Error), Error);
+  }
+}
+
+} // namespace enqueue_kernel_launch
+
+} // namespace detail
+} // namespace sycl
+} // namespace cl
diff --git a/sycl/source/detail/error_handling/error_handling.hpp b/sycl/source/detail/error_handling/error_handling.hpp
@@ -0,0 +1,32 @@
+//===-------- error_handling.hpp - SYCL error handling  ---------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include <CL/sycl/detail/pi.h>
+#include <CL/sycl/detail/cg.hpp>
+
+namespace cl {
+namespace sycl {
+namespace detail {
+
+namespace enqueue_kernel_launch {
+/// Analyzes error code and arguments of piEnqueueKernelLaunch to emit
+/// user-friendly exception describing the problem.
+///
+/// This function is expected to be called only for non-success error codes,
+/// i.e. the first argument must not be equal to PI_SUCCESS.
+///
+/// This function actually never returns and always throws an exception with
+/// error description.
+bool handleError(pi_result, pi_device, pi_kernel, const NDRDescT &);
+} // namespace enqueue_kernel_launch
+
+} // namespace detail
+} // namespace sycl
+} // namespace cl
diff --git a/sycl/source/detail/scheduler/commands.cpp b/sycl/source/detail/scheduler/commands.cpp
@@ -6,6 +6,8 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include <detail/error_handling/error_handling.hpp>
+
 #include "CL/sycl/access/access.hpp"
 #include <CL/cl.h>
 #include <CL/sycl/detail/clusm.hpp>
@@ -940,11 +942,19 @@ cl_int ExecCGCommand::enqueueImp() {
 
     ReverseRangeDimensionsForKernel(NDRDesc);
 
-    PI_CALL(piEnqueueKernelLaunch)(
+    pi_result Error = PI_CALL_NOCHECK(piEnqueueKernelLaunch)(
         MQueue->getHandleRef(), Kernel, NDRDesc.Dims, &NDRDesc.GlobalOffset[0],
         &NDRDesc.GlobalSize[0], HasLocalSize ? &NDRDesc.LocalSize[0] : nullptr,
         RawEvents.size(), RawEvents.empty() ? nullptr : &RawEvents[0], &Event);
 
+    if (PI_SUCCESS != Error) {
+      // If we have got non-success error code, let's analyze it to emit nice
+      // exception explaining what was wrong
+      pi_device Device =
+          detail::getSyclObjImpl(MQueue->get_device())->getHandleRef();
+      return detail::enqueue_kernel_launch::handleError(Error, Device, Kernel,
+                                                        NDRDesc);
+    }
     return PI_SUCCESS;
   }
   case CG::CGTYPE::COPY_USM: {
diff --git a/sycl/test/basic_tests/parallel_for_range.cpp b/sycl/test/basic_tests/parallel_for_range.cpp
diff --git a/sycl/test/basic_tests/parallel_for_range_host.cpp b/sycl/test/basic_tests/parallel_for_range_host.cpp
