[SYCL][ESIMD] Software-emulation preparation

* This PR adds new Plug-In module for ESIMD Software-emulation,
ESIMD_CPU

* The new feature will use CM emulation library for runtime support
and ESIMD intrinsics such as Software-multithreaded kernel launching
and buffer management/access

* pi* functions are filled with dummy codes as placeholder

* Steps for importing CM will be added later

* Two PI APIs are added - piextGetPluginSpecificData,
piEnqueueHostKernelLaunch

Author: dongkyunahn-intel

buildbot/configure.py

@@ -27,6 +27,7 @@ def do_configure(args):
     llvm_enable_projects = 'clang;' + llvm_external_projects
     libclc_targets_to_build = ''
     sycl_build_pi_cuda = 'OFF'
+    sycl_build_pi_esimd_cpu = 'ON'
     sycl_werror = 'ON'
     llvm_enable_assertions = 'ON'
     llvm_enable_doxygen = 'OFF'
@@ -46,6 +47,9 @@ def do_configure(args):
         libclc_targets_to_build = 'nvptx64--;nvptx64--nvidiacl'
         sycl_build_pi_cuda = 'ON'
 
+    if args.disable_esimd_cpu:
+        sycl_build_pi_esimd_cpu = 'OFF'
+
     if args.no_werror:
         sycl_werror = 'OFF'
 
@@ -84,6 +88,7 @@ def do_configure(args):
         "-DLLVM_ENABLE_DOXYGEN={}".format(llvm_enable_doxygen),
         "-DLLVM_ENABLE_SPHINX={}".format(llvm_enable_sphinx),
         "-DBUILD_SHARED_LIBS={}".format(llvm_build_shared_libs),
+        "-DSYCL_BUILD_PI_ESIMD_CPU={}".format(sycl_build_pi_esimd_cpu),
         "-DSYCL_ENABLE_XPTI_TRACING={}".format(sycl_enable_xpti_tracing)
     ]
 
@@ -148,6 +153,7 @@ def main():
                         metavar="BUILD_TYPE", default="Release", help="build type: Debug, Release")
     parser.add_argument("--cuda", action='store_true', help="switch from OpenCL to CUDA")
     parser.add_argument("--arm", action='store_true', help="build ARM support rather than x86")
+    parser.add_argument("--disable-esimd-cpu", action='store_true', help="build without ESIMD_CPU support")
     parser.add_argument("--no-assertions", action='store_true', help="build without assertions")
     parser.add_argument("--docs", action='store_true', help="build Doxygen documentation")
     parser.add_argument("--system-ocl", action='store_true', help="use OpenCL deps from system (no download)")

sycl/CMakeLists.txt

@@ -357,6 +357,7 @@ set( SYCL_TOOLCHAIN_DEPLOY_COMPONENTS
      sycl
      pi_opencl
      pi_level_zero
+     pi_esimd_cpu
      libsycldevice
      ${XPTIFW_LIBS}
 )

sycl/include/CL/sycl/INTEL/esimd/detail/emu/esimdcpu_device_interface.hpp

@@ -0,0 +1,37 @@
+//==----- esimdcpu_device_interface.hpp - DPC++ Explicit SIMD API ---------==//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+/// \file esimdcpu_device_interface.hpp
+/// Declarations for ESIMD_CPU-device specific definitions.
+/// ESIMD intrinsic and LibCM functionalities required by intrinsic defined
+///
+/// \ingroup sycl_pi_esimd_cpu
+
+#pragma once
+
+#include <cstdint>
+
+#ifdef _MSC_VER
+typedef unsigned int uint;
+typedef unsigned short ushort;
+typedef unsigned char uchar;
+#endif // _MSC_VER
+
+struct ESIMDDeviceInterface {
+
+  // Intrinsics
+  virtual void mt_barrier() = 0;
+  virtual void split_barrier(uint) = 0;
+  virtual void fence() = 0;
+
+  // libcm functionalities used for intrinsics such as
+  // surface/buffer/slm access
+  virtual char *get_surface_base(int surfaceID) = 0;
+  virtual char *get_slm() = 0;
+  virtual void set_slm_size(size_t size) = 0;
+};

sycl/include/CL/sycl/detail/pi.def

@@ -29,6 +29,7 @@ _PI_API(piextDeviceSelectBinary)
 _PI_API(piextGetDeviceFunctionPointer)
 _PI_API(piextDeviceGetNativeHandle)
 _PI_API(piextDeviceCreateWithNativeHandle)
+_PI_API(piextGetPluginSpecificData)
 // Context
 _PI_API(piContextCreate)
 _PI_API(piContextGetInfo)
@@ -97,6 +98,7 @@ _PI_API(piSamplerRetain)
 _PI_API(piSamplerRelease)
 // Queue commands
 _PI_API(piEnqueueKernelLaunch)
+_PI_API(piEnqueueHostKernelLaunch)
 _PI_API(piEnqueueNativeKernel)
 _PI_API(piEnqueueEventsWait)
 _PI_API(piEnqueueEventsWaitWithBarrier)

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

@@ -913,6 +913,9 @@ piextDeviceGetNativeHandle(pi_device device, pi_native_handle *nativeHandle);
 __SYCL_EXPORT pi_result piextDeviceCreateWithNativeHandle(
     pi_native_handle nativeHandle, pi_platform platform, pi_device *device);
 
+/// Returns plug-in specific data
+__SYCL_EXPORT pi_result piextGetPluginSpecificData(void **PlugInData);
+
 /// Selects the most appropriate device binary based on runtime information
 /// and the IR characteristics.
 ///
@@ -1309,6 +1312,12 @@ __SYCL_EXPORT pi_result piEnqueueKernelLaunch(
     const size_t *local_work_size, pi_uint32 num_events_in_wait_list,
     const pi_event *event_wait_list, pi_event *event);
 
+__SYCL_EXPORT pi_result piEnqueueHostKernelLaunch(
+    void *Kernel, pi_uint32 WorkDim, const size_t *GlobalWorkOffset,
+    const size_t *GlobalWorkSize, const size_t *LocalWorkSize,
+    pi_uint32 NumEventsInWaitList, const pi_event *EventWaitList,
+    pi_event *Event);
+
 __SYCL_EXPORT pi_result piEnqueueNativeKernel(
     pi_queue queue, void (*user_func)(void *), void *args, size_t cb_args,
     pi_uint32 num_mem_objects, const pi_mem *mem_list,

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

@@ -57,14 +57,19 @@ enum TraceLevel {
 // Return true if we want to trace PI related activities.
 bool trace(TraceLevel level);
 
+// Return PluginSpecificData referenced by kernel code
+__SYCL_EXPORT void *getPluginSpecificData(backend be);
+
 #ifdef __SYCL_RT_OS_WINDOWS
 #define __SYCL_OPENCL_PLUGIN_NAME "pi_opencl.dll"
 #define __SYCL_LEVEL_ZERO_PLUGIN_NAME "pi_level_zero.dll"
 #define __SYCL_CUDA_PLUGIN_NAME "pi_cuda.dll"
+#define __SYCL_ESIMD_CPU_PLUGIN_NAME "pi_esimd_cpu.dll"
 #else
 #define __SYCL_OPENCL_PLUGIN_NAME "libpi_opencl.so"
 #define __SYCL_LEVEL_ZERO_PLUGIN_NAME "libpi_level_zero.so"
 #define __SYCL_CUDA_PLUGIN_NAME "libpi_cuda.so"
+#define __SYCL_ESIMD_CPU_PLUGIN_NAME "libpi_esimd_cpu.so"
 #endif
 
 // Report error and no return (keeps compiler happy about no return statements).

sycl/plugins/CMakeLists.txt

@@ -10,3 +10,7 @@ endif()
 
 add_subdirectory(opencl)
 add_subdirectory(level_zero)
+
+if (SYCL_BUILD_PI_ESIMD_CPU)
+  add_subdirectory(esimd_cpu)
+endif()

sycl/plugins/esimd_cpu/CMakeLists.txt

@@ -0,0 +1,26 @@
+
+# PI Esimd CPU  library
+# Create Shared library for libpi_esimd_cpu.so.
+
+include_directories("${sycl_inc_dir}")
+include_directories(${OPENCL_INCLUDE})
+include_directories(${LIBCMRT_INCLUDE})
+
+add_library(pi_esimd_cpu SHARED
+  "${sycl_inc_dir}/CL/sycl/detail/pi.h"
+  "pi_esimd_cpu.cpp"
+)
+
+add_dependencies(sycl-toolchain pi_esimd_cpu)
+
+add_dependencies(pi_esimd_cpu
+  ocl-headers)
+
+target_link_libraries(pi_esimd_cpu PRIVATE sycl ${CMAKE_DL_LIBS})
+set_target_properties(pi_esimd_cpu PROPERTIES LINKER_LANGUAGE CXX)
+
+add_common_options(pi_esimd_cpu)
+
+install(TARGETS pi_esimd_cpu
+    LIBRARY DESTINATION "lib${LLVM_LIBDIR_SUFFIX}" COMPONENT pi_esimd_cpu
+    RUNTIME DESTINATION "bin" COMPONENT pi_esimd_cpu)