[SYCL] Start running OpenCL through PI

Signed-off-by: Sergey V Maslov <[email protected]>

Author: smaslov-intel

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

@@ -47,7 +47,8 @@ typedef pi_uint32   pi_bool;
 // TODO: populate PI enums.
 //
 typedef enum {
-  PI_SUCCESS = CL_SUCCESS
+  PI_SUCCESS = CL_SUCCESS,
+  PI_RESULT_INVALID_KERNEL_NAME = CL_INVALID_KERNEL_NAME
 } _pi_result;
 
 typedef enum {

sycl/include/CL/sycl/detail/pi.hpp

@@ -12,7 +12,6 @@
 
 #include <CL/sycl/detail/pi.h>
 #include <CL/sycl/detail/common.hpp>
-#include <CL/sycl/detail/pi_opencl.hpp> // TODO: remove when switched to PI
 
 namespace cl {
 namespace sycl {
@@ -114,11 +113,7 @@ namespace pi {
   #undef _PI_API
 } // namespace pi
 
-// Select underlying runtime interface in compile-time (OpenCL or PI).
-// As such only one path (OpenCL today) is being regularily tested.
-// TODO: change to
-// namespace RT = cl::sycl::detail::pi;
-namespace RT = cl::sycl::detail::pi_opencl;
+namespace RT = cl::sycl::detail::pi;
 
 // Report error and no return (keeps compiler happy about no return statements).
 [[noreturn]] void piDie(const char *Message);

sycl/include/CL/sycl/detail/pi_opencl.hpp

@@ -395,7 +395,7 @@ class program_impl {
     RT::PiResult Err;
     Err = PI_CALL_RESULT((Kernel = RT::piKernelCreate(
         Program, KernelName.c_str(), &Err), Err));
-    if (Err == CL_INVALID_KERNEL_NAME) {
+    if (Err == PI_RESULT_INVALID_KERNEL_NAME) {
       throw invalid_object_error(
           "This instance of program does not contain the kernel requested");
     }

sycl/include/CL/sycl/detail/program_impl.hpp

@@ -18,15 +18,7 @@ platform_impl_pi::get_platforms() {
   vector_class<platform> platforms;
 
   pi_uint32 num_platforms = 0;
-  auto Err = PI_CALL_RESULT(RT::piPlatformsGet(0, 0, &num_platforms));
-
-  // TODO: remove this check when switch to PI, which will just return  0 in
-  // num_platforms.
-  //
-  if (Err == CL_PLATFORM_NOT_FOUND_KHR)
-    return platforms;
-
-  PI_CHECK(Err);
+  PI_CALL(RT::piPlatformsGet(0, 0, &num_platforms));
   info::device_type forced_type = detail::get_forced_type();
 
   if (num_platforms) {
@@ -61,20 +53,12 @@ platform_impl_pi::get_devices(info::device_type deviceType) const {
     return res;
 
   pi_uint32 num_devices;
-  auto err = PI_CALL_RESULT(RT::piDevicesGet(
-    m_platform, pi_cast<RT::PiDeviceType>(deviceType), 0, 0, &num_devices));
-
-  // TODO: remove this check when switched to PI as it would just return
-  // zero in num_devices.
-  if (err == CL_DEVICE_NOT_FOUND)
-    return res;
+  PI_CALL(RT::piDevicesGet(
+      m_platform, pi_cast<RT::PiDeviceType>(deviceType), 0, 0, &num_devices));
 
   if (num_devices == 0)
     return res;
 
-  // TODO catch an exception and put it to list of asynchronous exceptions
-  PI_CHECK(err);
-
   vector_class<RT::PiDevice> pi_devices(num_devices);
   // TODO catch an exception and put it to list of asynchronous exceptions
   PI_CALL(RT::piDevicesGet(

sycl/source/detail/platform_impl.cpp

@@ -60,7 +60,7 @@ RT::PiSampler sampler_impl::getOrCreateSampler(const context &Context) {
       getSyclObjImpl(Context)->getHandleRef(), sprops, &errcode_ret),
       errcode_ret));
 #else
-  // TODO: do we reall need this old interface into PI and here?
+  // TODO: do we really need this old interface into PI and here?
   cl_int cl_errcode_ret;
   m_contextToSampler[Context] =
       clCreateSampler(getSyclObjImpl(Context)->getHandleRef(),

sycl/source/detail/sampler_impl.cpp

@@ -265,8 +265,8 @@ cl_int MemCpyCommandHost::enqueueImp() {
 // number of work - groups, such that the size of each group is chosen by the
 // runtime, or by the number of work - groups and number of work - items for
 // users who need more control.
-static void adjustNDRangePerKernel(NDRDescT &NDR, cl_kernel Kernel,
-                                   cl_device_id Device) {
+static void adjustNDRangePerKernel(NDRDescT &NDR, RT::PiKernel Kernel,
+                                   RT::PiDevice Device) {
   if (NDR.GlobalSize[0] != 0)
     return; // GlobalSize is set - no need to adjust
   // check the prerequisites:
@@ -422,7 +422,10 @@ cl_int ExecCGCommand::enqueueImp() {
         assert(!"Unhandled");
       }
     }
-    adjustNDRangePerKernel(NDRDesc, Kernel, MQueue->get_device().get());
+    adjustNDRangePerKernel(NDRDesc, Kernel,
+                           detail::getSyclObjImpl(
+                               MQueue->get_device())->getHandleRef());
+
     PI_CALL(RT::piEnqueueKernelLaunch(
         MQueue->getHandleRef(), Kernel, NDRDesc.Dims, &NDRDesc.GlobalOffset[0],
         &NDRDesc.GlobalSize[0],