Merge branch 'private/dvodopya/non-native-sycl2020-spec-consts' into sycl

Author: dm-vodopyanov

sycl/include/CL/sycl/detail/pi.h

@@ -515,6 +515,7 @@ typedef enum {
 using pi_mem_flags = pi_bitfield;
 // Access
 constexpr pi_mem_flags PI_MEM_FLAGS_ACCESS_RW = CL_MEM_READ_WRITE;
+constexpr pi_mem_flags PI_MEM_FLAGS_ACCESS_RO = CL_MEM_READ_ONLY;
 // Host pointer
 constexpr pi_mem_flags PI_MEM_FLAGS_HOST_PTR_USE = CL_MEM_USE_HOST_PTR;
 constexpr pi_mem_flags PI_MEM_FLAGS_HOST_PTR_COPY = CL_MEM_COPY_HOST_PTR;

sycl/source/detail/device_image_impl.hpp

@@ -169,6 +169,19 @@ class device_image_impl {
     return MSpecConstsBlob;
   }
 
+  RT::PiMem &get_spec_const_buffer_ref() noexcept {
+    std::lock_guard<std::mutex> Lock{MSpecConstAccessMtx};
+    if (nullptr == MSpecConstsBuffer) {
+      const detail::plugin &Plugin = getSyclObjImpl(MContext)->getPlugin();
+      Plugin.call<PiApiKind::piMemBufferCreate>(
+          detail::getSyclObjImpl(MContext)->getHandleRef(),
+          PI_MEM_FLAGS_ACCESS_RO | PI_MEM_FLAGS_HOST_PTR_USE,
+          MSpecConstsBlob.size(), MSpecConstsBlob.data(), &MSpecConstsBuffer,
+          nullptr);
+    }
+    return MSpecConstsBuffer;
+  }
+
   const std::map<std::string, std::vector<SpecConstDescT>> &
   get_spec_const_data_ref() const noexcept {
     return MSpecConstSymMap;
@@ -262,6 +275,10 @@ class device_image_impl {
   // Binary blob which can have values of all specialization constants in the
   // image
   std::vector<unsigned char> MSpecConstsBlob;
+  // Buffer containing binary blob which can have values of all specialization
+  // constants in the image, it is using for storing non-native specialization
+  // constants
+  RT::PiMem MSpecConstsBuffer = nullptr;
   // Contains map of spec const names to their descriptions + offsets in
   // the MSpecConstsBlob
   std::map<std::string, std::vector<SpecConstDescT>> MSpecConstSymMap;

sycl/source/detail/kernel_bundle_impl.hpp

@@ -442,7 +442,11 @@ class kernel_bundle_impl {
     return SetInDevImg || MSpecConstValues.count(std::string{SpecName}) != 0;
   }
 
-  const device_image_plain *begin() const { return &MDeviceImages.front(); }
+  const device_image_plain *begin() const {
+    assert(!MDeviceImages.empty() && "MDeviceImages can't be empty");
+    // UB in case MDeviceImages is empty
+    return &MDeviceImages.front();
+  }
 
   const device_image_plain *end() const { return &MDeviceImages.back() + 1; }
 

sycl/source/detail/program_manager/program_manager.cpp

@@ -1309,7 +1309,8 @@ void ProgramManager::bringSYCLDeviceImagesToState(
         break;
       }
       case bundle_state::executable:
-        // Device image is already in the desired state.
+        DevImage = build(DevImage, getSyclObjImpl(DevImage)->get_devices(),
+                         /*PropList=*/{});
         break;
       }
       break;

sycl/source/detail/scheduler/commands.cpp

@@ -1636,8 +1636,9 @@ static void ReverseRangeDimensionsForKernel(NDRDescT &NDR) {
 }
 
 pi_result ExecCGCommand::SetKernelParamsAndLaunch(
-    CGExecKernel *ExecKernel, RT::PiKernel Kernel, NDRDescT &NDRDesc,
-    std::vector<RT::PiEvent> &RawEvents, RT::PiEvent &Event,
+    CGExecKernel *ExecKernel,
+    std::shared_ptr<device_image_impl> DeviceImageImpl, RT::PiKernel Kernel,
+    NDRDescT &NDRDesc, std::vector<RT::PiEvent> &RawEvents, RT::PiEvent &Event,
     ProgramManager::KernelArgMask EliminatedArgMask) {
   vector_class<ArgDesc> &Args = ExecKernel->MArgs;
   // TODO this is not necessary as long as we can guarantee that the arguments
@@ -1692,9 +1693,21 @@ pi_result ExecCGCommand::SetKernelParamsAndLaunch(
       break;
     }
     case kernel_param_kind_t::kind_specialization_constants_buffer: {
-      throw cl::sycl::feature_not_supported(
-          "SYCL2020 specialization constants are not yet fully supported",
-          PI_INVALID_OPERATION);
+      if (MQueue->is_host()) {
+        throw cl::sycl::feature_not_supported(
+            "SYCL2020 specialization constants are not yet supported on host "
+            "device",
+            PI_INVALID_OPERATION);
+      }
+      if (DeviceImageImpl != nullptr) {
+        RT::PiMem SpecConstsBuffer =
+            DeviceImageImpl->get_spec_const_buffer_ref();
+        Plugin.call<PiApiKind::piKernelSetArg>(
+            Kernel, NextTrueIndex, sizeof(RT::PiMem), &SpecConstsBuffer);
+      } else {
+        Plugin.call<PiApiKind::piKernelSetArg>(Kernel, NextTrueIndex,
+                                               sizeof(RT::PiMem), nullptr);
+      }
       break;
     }
     }
@@ -1916,6 +1929,8 @@ cl_int ExecCGCommand::enqueueImp() {
     bool KnownProgram = true;
 
     std::shared_ptr<kernel_impl> SyclKernelImpl;
+    std::shared_ptr<device_image_impl> DeviceImageImpl;
+
     // Use kernel_bundle is available
     if (KernelBundleImplPtr) {
 
@@ -1929,9 +1944,7 @@ cl_int ExecCGCommand::enqueueImp() {
       SyclKernelImpl = detail::getSyclObjImpl(SyclKernel);
 
       Kernel = SyclKernelImpl->getHandleRef();
-
-      std::shared_ptr<device_image_impl> DeviceImageImpl =
-          SyclKernelImpl->getDeviceImage();
+      DeviceImageImpl = SyclKernelImpl->getDeviceImage();
 
       Program = DeviceImageImpl->get_program_ref();
 
@@ -1979,11 +1992,13 @@ cl_int ExecCGCommand::enqueueImp() {
     if (KernelMutex != nullptr) {
       // For cacheable kernels, we use per-kernel mutex
       std::lock_guard<std::mutex> Lock(*KernelMutex);
-      Error = SetKernelParamsAndLaunch(ExecKernel, Kernel, NDRDesc, RawEvents,
-                                       Event, EliminatedArgMask);
+      Error =
+          SetKernelParamsAndLaunch(ExecKernel, DeviceImageImpl, Kernel, NDRDesc,
+                                   RawEvents, Event, EliminatedArgMask);
     } else {
-      Error = SetKernelParamsAndLaunch(ExecKernel, Kernel, NDRDesc, RawEvents,
-                                       Event, EliminatedArgMask);
+      Error =
+          SetKernelParamsAndLaunch(ExecKernel, DeviceImageImpl, Kernel, NDRDesc,
+                                   RawEvents, Event, EliminatedArgMask);
     }
 
     if (PI_SUCCESS != Error) {

sycl/source/detail/scheduler/commands.hpp

@@ -519,9 +519,10 @@ class ExecCGCommand : public Command {
   AllocaCommandBase *getAllocaForReq(Requirement *Req);
 
   pi_result SetKernelParamsAndLaunch(
-      CGExecKernel *ExecKernel, RT::PiKernel Kernel, NDRDescT &NDRDesc,
-      std::vector<RT::PiEvent> &RawEvents, RT::PiEvent &Event,
-      ProgramManager::KernelArgMask EliminatedArgMask);
+      CGExecKernel *ExecKernel,
+      std::shared_ptr<device_image_impl> DeviceImageImpl, RT::PiKernel Kernel,
+      NDRDescT &NDRDesc, std::vector<RT::PiEvent> &RawEvents,
+      RT::PiEvent &Event, ProgramManager::KernelArgMask EliminatedArgMask);
 
   std::unique_ptr<detail::CG> MCommandGroup;
 

sycl/source/handler.cpp

@@ -57,6 +57,10 @@ handler::getOrInsertHandlerKernelBundle(bool Insert) const {
   if (!KernelBundleImpPtr && Insert) {
     KernelBundleImpPtr = detail::getSyclObjImpl(
         get_kernel_bundle<bundle_state::input>(MQueue->get_context()));
+    if (KernelBundleImpPtr->empty()) {
+      KernelBundleImpPtr = detail::getSyclObjImpl(
+          get_kernel_bundle<bundle_state::executable>(MQueue->get_context()));
+    }
 
     detail::ExtendedMemberT EMember = {
         detail::ExtendedMembersType::HANDLER_KERNEL_BUNDLE, KernelBundleImpPtr};
@@ -340,9 +344,9 @@ void handler::processArg(void *Ptr, const detail::kernel_param_kind_t &Kind,
     break;
   }
   case kernel_param_kind_t::kind_specialization_constants_buffer: {
-    throw cl::sycl::feature_not_supported(
-        "SYCL2020 specialization constants are not yet fully supported",
-        PI_INVALID_OPERATION);
+    MArgs.emplace_back(
+        kernel_param_kind_t::kind_specialization_constants_buffer, Ptr, Size,
+        Index + IndexShift);
     break;
   }
   }

sycl/test/on-device/basic_tests/specialization_constants/kernel_lambda_with_kernel_handler_arg.cpp

@@ -5,10 +5,6 @@
 // and parallel_for_work_group to verify that this code compiles and runs
 // correctly with user's lambda with and without sycl::kernel_handler argument
 
-// TODO: enable cuda support when non-native spec constants started to be
-// supported
-// UNSUPPORTED: cuda
-
 #include <CL/sycl.hpp>
 
 int main() {

sycl/test/on-device/basic_tests/specialization_constants/non_native/Inputs/common.cpp

@@ -0,0 +1,54 @@
+#include <sycl/sycl.hpp>
+
+#include <cmath>
+
+class Kernel1Name;
+class Kernel2Name;
+
+struct TestStruct {
+  int a;
+  int b;
+};
+
+const static sycl::specialization_id<int> SpecConst1{42};
+const static sycl::specialization_id<int> SpecConst2{42};
+const static sycl::specialization_id<TestStruct> SpecConst3{TestStruct{42, 42}};
+const static sycl::specialization_id<short> SpecConst4{42};
+
+int main() {
+  sycl::queue Q;
+
+  // No support for host device so far
+  if (Q.is_host())
+    return 0;
+
+  {
+    sycl::buffer<int, 1> Buf{sycl::range{1}};
+    Q.submit([&](sycl::handler &CGH) {
+      CGH.set_specialization_constant<SpecConst2>(1);
+      auto Acc = Buf.get_access<sycl::access::mode::read_write>(CGH);
+      CGH.single_task<class Kernel1Name>([=](sycl::kernel_handler KH) {
+        Acc[0] = KH.get_specialization_constant<SpecConst2>();
+      });
+    });
+    auto Acc = Buf.get_access<sycl::access::mode::read>();
+    assert(Acc[0] == 1);
+  }
+
+  {
+    sycl::buffer<TestStruct, 1> Buf{sycl::range{1}};
+    Q.submit([&](sycl::handler &CGH) {
+      auto Acc = Buf.get_access<sycl::access::mode::read_write>(CGH);
+      CGH.set_specialization_constant<SpecConst3>(TestStruct{1, 2});
+      const auto SC = CGH.get_specialization_constant<SpecConst4>();
+      assert(SC == 42);
+      CGH.single_task<class Kernel2Name>([=](sycl::kernel_handler KH) {
+        Acc[0] = KH.get_specialization_constant<SpecConst3>();
+      });
+    });
+    auto Acc = Buf.get_access<sycl::access::mode::read>();
+    assert(Acc[0].a == 1 && Acc[0].b == 2);
+  }
+
+  return 0;
+}

sycl/test/on-device/basic_tests/specialization_constants/non_native/accelerator.cpp

@@ -0,0 +1,7 @@
+// REQUIRES: aoc, accelerator
+
+// RUN: %clangxx -fsycl -fsycl-targets=spir64_fpga-unknown-unknown-sycldevice %S/Inputs/common.cpp -o %t.out
+// RUN: %ACC_RUN_PLACEHOLDER %t.out
+
+// This test checks correctness of SYCL2020 non-native specialization constants
+// on accelerator device