[SYCL] populate PI (and direct SYCL RT to PI)

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

Author: smaslov-intel

sycl/doc/SYCL_environment_variables.md

@@ -0,0 +1,10 @@
+# Overview
+
+This file describes environment variables that are having effect on SYCL compiler and run-time.
+
+# Controlling SYCL RT
+
+| Environment variable | Description |
+| ----------- | ----------- |
+| SYCL_PI_TRACE | If set forces tracing of PI calls to stdout. |
+| SYCL_BE={PI_OPENCL,PI_OTHER} | When SYCL RT is buils with PI this controls which plugin to use. |

sycl/include/CL/sycl/buffer.hpp

@@ -128,8 +128,10 @@ class buffer {
          event AvailableEvent = {}) {
 
     size_t BufSize = 0;
-    CHECK_OCL_CODE(clGetMemObjectInfo(MemObject, CL_MEM_SIZE, sizeof(size_t),
-                                      &BufSize, nullptr));
+    PI_CALL(detail::RT::piMemGetInfo(
+        detail::pi_cast<detail::RT::PiMem>(MemObject), CL_MEM_SIZE,
+        sizeof(size_t), &BufSize, nullptr));
+
     Range[0] = BufSize / sizeof(T);
     MemRange[0] = BufSize / sizeof(T);
     impl = std::make_shared<detail::buffer_impl<AllocatorT>>(

sycl/include/CL/sycl/detail/buffer_impl.hpp

@@ -13,6 +13,7 @@
 #include <CL/sycl/context.hpp>
 #include <CL/sycl/detail/aligned_allocator.hpp>
 #include <CL/sycl/detail/common.hpp>
+#include <CL/sycl/detail/pi.hpp>
 #include <CL/sycl/detail/helpers.hpp>
 #include <CL/sycl/detail/memory_manager.hpp>
 #include <CL/sycl/detail/scheduler/scheduler.hpp>
@@ -182,13 +183,15 @@ template <typename AllocatorT> class buffer_impl : public SYCLMemObjT {
           "Creation of interoperability buffer using host context is not "
           "allowed");
 
-    cl_context Context = nullptr;
-    CHECK_OCL_CODE(clGetMemObjectInfo(MInteropMemObject, CL_MEM_CONTEXT,
-                                      sizeof(Context), &Context, nullptr));
+    RT::PiMem Mem = pi_cast<RT::PiMem>(MInteropMemObject);
+    RT::PiContext Context = nullptr;
+    PI_CALL(RT::piMemGetInfo(
+        Mem, CL_MEM_CONTEXT, sizeof(Context), &Context, nullptr));
+
     if (MInteropContext->getHandleRef() != Context)
       throw cl::sycl::invalid_parameter_error(
           "Input context must be the same as the context of cl_mem");
-    CHECK_OCL_CODE(clRetainMemObject(MInteropMemObject));
+    PI_CALL(RT::piMemRetain(Mem));
   }
 
   size_t get_size() const { return MSizeInBytes; }
@@ -206,7 +209,7 @@ template <typename AllocatorT> class buffer_impl : public SYCLMemObjT {
     releaseHostMem(MShadowCopy);
 
     if (MOpenCLInterop)
-      CHECK_OCL_CODE_NO_EXC(clReleaseMemObject(MInteropMemObject));
+      PI_CALL(RT::piMemRelease(pi_cast<RT::PiMem>(MInteropMemObject)));
   }
 
   void set_final_data(std::nullptr_t) { MUploadDataFn = nullptr; }
@@ -243,7 +246,7 @@ template <typename AllocatorT> class buffer_impl : public SYCLMemObjT {
       typename std::enable_if<std::is_pointer<Destination>::value>::type * =
           0) {
     static_assert(!std::is_const<Destination>::value,
-                  "Сan not write in a constant Destination. Destination should "
+                  "Do not write in a constant Destination. Destination should "
                   "not be const.");
     MUploadDataFn = [this, FinalData]() mutable {
 
@@ -265,7 +268,7 @@ template <typename AllocatorT> class buffer_impl : public SYCLMemObjT {
       typename std::enable_if<!std::is_pointer<Destination>::value>::type * =
           0) {
     static_assert(!std::is_const<Destination>::value,
-                  "Сan not write in a constant Destination. Destination should "
+                  "Do not write in a constant Destination. Destination should "
                   "not be const.");
     MUploadDataFn = [this, FinalData]() mutable {
       using FinalDataType =
@@ -340,7 +343,7 @@ template <typename AllocatorT> class buffer_impl : public SYCLMemObjT {
   }
 
   void *allocateMem(ContextImplPtr Context, bool InitFromUserData,
-                    cl_event &OutEventToWait) override {
+                    RT::PiEvent &OutEventToWait) override {
 
     void *UserPtr = InitFromUserData ? getUserPtr() : nullptr;
 

sycl/include/CL/sycl/detail/common.hpp

@@ -17,28 +17,14 @@
 #include <string>
 #include <type_traits>
 
-// Select underlying runtime interface in compile-time (OpenCL or PI).
-// Comment the define of the FORCE_SYCL_BE_OPENCL below to switch to PI.
-// As such only one path (OpenCL today) is being regularily tested.
-//
-// TODO: we can just remove this when switch to PI completely.
-//
-#define FORCE_SYCL_BE_OPENCL
-
-#ifdef FORCE_SYCL_BE_OPENCL
-#include <CL/sycl/detail/pi_opencl.hpp>
-#else
-#include <CL/sycl/detail/pi.hpp>
-#endif
+#define STRINGIFY_LINE_HELP(s) #s
+#define STRINGIFY_LINE(s) STRINGIFY_LINE_HELP(s)
 
 const char *stringifyErrorCode(cl_int error);
 
 #define OCL_CODE_TO_STR(code)                                                  \
   std::string(std::to_string(code) + " (" + stringifyErrorCode(code) + ")")
 
-#define STRINGIFY_LINE_HELP(s) #s
-#define STRINGIFY_LINE(s) STRINGIFY_LINE_HELP(s)
-
 #define OCL_ERROR_REPORT                                                       \
   "OpenCL API failed. " __FILE__                                               \
   ":" STRINGIFY_LINE(__LINE__) ": "                                            \
@@ -92,15 +78,6 @@ namespace cl {
 namespace sycl {
 namespace detail {
 
-// Select underlying runtime interface (RT) in compile-time (OpenCL or PI).
-// As such only one path (OpenCL today) is being regularily tested.
-//
-#ifdef FORCE_SYCL_BE_OPENCL
-using RT = cl::sycl::detail::opencl;
-#else
-using RT = cl::sycl::detail::pi;
-#endif
-
 // Helper function for extracting implementation from SYCL's interface objects.
 // Note! This function relies on the fact that all SYCL interface classes
 // contain "impl" field that points to implementation object. "impl" field

sycl/include/CL/sycl/detail/context_impl.hpp

@@ -8,6 +8,7 @@
 
 #pragma once
 #include <CL/sycl/detail/common.hpp>
+#include <CL/sycl/detail/pi.hpp>
 #include <CL/sycl/exception.hpp>
 #include <CL/sycl/info/info_desc.hpp>
 #include <CL/sycl/platform.hpp>
@@ -50,13 +51,13 @@ class context_impl {
   // modification. Caller must ensure the returned object lives on stack only.
   // It can also be safely passed to the underlying native runtime API.
   // Warning. Returned reference will be invalid if context_impl was destroyed.
-  cl_context &getHandleRef();
-  const cl_context &getHandleRef() const;
+  RT::PiContext &getHandleRef();
+  const RT::PiContext &getHandleRef() const;
 
 private:
   async_handler m_AsyncHandler;
   vector_class<device> m_Devices;
-  cl_context m_ClContext;
+  RT::PiContext m_Context;
   platform m_Platform;
   bool m_OpenCLInterop;
   bool m_HostContext;

sycl/include/CL/sycl/detail/context_info.hpp

@@ -15,15 +15,15 @@ namespace cl {
 namespace sycl {
 namespace detail {
 
-template <info::context param> struct get_context_info_cl {
+template <info::context param> struct get_context_info {
   using RetType =
       typename info::param_traits<info::context, param>::return_type;
 
-  static RetType _(cl_context ctx) {
+  static RetType _(RT::PiContext ctx) {
     RetType Result = 0;
     // TODO catch an exception and put it to list of asynchronous exceptions
-    CHECK_OCL_CODE(clGetContextInfo(ctx, cl_context_info(param), sizeof(Result),
-                                    &Result, nullptr));
+    PI_CALL(RT::piContextGetInfo(ctx, pi_cast<pi_context_info>(param),
+                                 sizeof(Result), &Result, nullptr));
     return Result;
   }
 };

sycl/include/CL/sycl/detail/device_impl.hpp

@@ -8,6 +8,7 @@
 
 #pragma once
 
+#include <CL/sycl/detail/pi.hpp>
 #include <CL/sycl/detail/device_info.hpp>
 #include <CL/sycl/stl.hpp>
 #include <algorithm>
@@ -34,10 +35,8 @@ class device_impl {
   // It can also be safely passed to the underlying native runtime API.
   // Warning. Returned reference will be invalid if device_impl was destroyed.
   //
-  // TODO: change all uses of getHandleRef to get_handle, and remove the
-  //       getHandleRef after that.
-  virtual cl_device_id &getHandleRef() = 0;
-  virtual RT::pi_device get_handle() const = 0;
+  virtual RT::PiDevice &getHandleRef() = 0;
+  virtual const RT::PiDevice &getHandleRef() const = 0;
 
   virtual bool is_host() const = 0;
 
@@ -68,7 +67,7 @@ class device_impl {
     }
     return get_device_info<
         typename info::param_traits<info::device, param>::return_type,
-        param>::_(this->get_handle());
+        param>::_(this->getHandleRef());
   }
 
   bool is_partition_supported(info::partition_property Prop) const {
@@ -93,15 +92,15 @@ class device_impl {
 // TODO: Make code thread-safe
 class device_impl_pi : public device_impl {
 public:
-  explicit device_impl_pi(RT::pi_device a_device) : m_device(a_device) {
+  explicit device_impl_pi(RT::PiDevice a_device) : m_device(a_device) {
     // TODO catch an exception and put it to list of asynchronous exceptions
     PI_CALL(RT::piDeviceGetInfo(
-      m_device, PI_DEVICE_INFO_TYPE, sizeof(RT::pi_device_type), &m_type, 0));
+      m_device, PI_DEVICE_INFO_TYPE, sizeof(RT::PiDeviceType), &m_type, 0));
 
-    RT::pi_device parent;
+    RT::PiDevice parent;
     // TODO catch an exception and put it to list of asynchronous exceptions
     PI_CALL(RT::piDeviceGetInfo(
-      m_device, PI_DEVICE_INFO_PARENT, sizeof(RT::pi_device), &parent, 0));
+      m_device, PI_DEVICE_INFO_PARENT, sizeof(RT::PiDevice), &parent, 0));
 
     m_isRootDevice = (nullptr == parent);
     if (!m_isRootDevice) {
@@ -126,14 +125,8 @@ class device_impl_pi : public device_impl {
     return pi_cast<cl_device_id>(m_device);
   }
 
-  cl_device_id &getHandleRef() override {
-    // TODO: check that device is an OpenCL interop one before cast, or just
-    // remove when all the users are moved to get_handle.
-    return (cl_device_id&)(m_device);
-  }
-  RT::pi_device get_handle() const override {
-    return m_device;
-  }
+  RT::PiDevice &getHandleRef() override { return m_device; }
+  const RT::PiDevice &getHandleRef() const override { return m_device; }
 
   bool is_host() const override { return false; }
 
@@ -146,7 +139,7 @@ class device_impl_pi : public device_impl {
   }
 
   platform get_platform() const override {
-    RT::pi_platform plt;
+    RT::PiPlatform plt;
     // TODO catch an exception and put it to list of asynchronous exceptions
     PI_CALL(RT::piDeviceGetInfo(
       m_device, PI_DEVICE_INFO_PLATFORM, sizeof(plt), &plt, 0));
@@ -178,8 +171,8 @@ class device_impl_pi : public device_impl {
   create_sub_devices(info::partition_affinity_domain AffinityDomain) const override;
 
 private:
-  RT::pi_device m_device = 0;
-  RT::pi_device_type m_type;
+  RT::PiDevice m_device = 0;
+  RT::PiDeviceType m_type;
   bool m_isRootDevice = false;
 }; // class device_impl_pi
 
@@ -192,11 +185,11 @@ class device_host : public device_impl {
   cl_device_id get() const override {
     throw invalid_object_error("This instance of device is a host instance");
   }
-  cl_device_id &getHandleRef() override {
+  RT::PiDevice &getHandleRef() override {
     throw invalid_object_error("This instance of device is a host instance");
   }
-  RT::pi_device get_handle() const override {
-    pi_die("This instance of device is a host instance");
+  const RT::PiDevice &getHandleRef() const override {
+    throw invalid_object_error("This instance of device is a host instance");
   }
 
   bool is_host() const override { return true; }