[SYCL][NFC] Detach compiler interfaces from PI (#14700)

Our compiler interfaces do not depend on runtime bits and therefore they
shouldn't be tied to PI, UR or anything else. This PR is a by-product of
#14145 intended to refactor the code in advance so once we move to UR
there are less changes required.

Moreover, compiler interfaces are not used anywhere in public SYCL
headers and therefore they were moved into library.

As you can see, there are still some `pi` naming present in filenames
and namespaces, but this patch is huge enough already, so further
changes will be done as separate PRs.

Author: AlexeySachkov

sycl/include/sycl/detail/pi.h

@@ -988,186 +988,11 @@ using pi_device_exec_capabilities = pi_bitfield;
 static constexpr pi_device_exec_capabilities PI_EXEC_KERNEL = (1 << 0);
 static constexpr pi_device_exec_capabilities PI_EXEC_NATIVE_KERNEL = (1 << 1);
 
-// Entry type, matches OpenMP for compatibility
-struct _pi_offload_entry_struct {
-  void *addr;
-  char *name;
-  size_t size;
-  int32_t flags;
-  int32_t reserved;
-};
-
-using _pi_offload_entry = _pi_offload_entry_struct *;
-
-// A type of a binary image property.
-typedef enum {
-  PI_PROPERTY_TYPE_UNKNOWN,
-  PI_PROPERTY_TYPE_UINT32,     // 32-bit integer
-  PI_PROPERTY_TYPE_BYTE_ARRAY, // byte array
-  PI_PROPERTY_TYPE_STRING      // null-terminated string
-} pi_property_type;
-
-// Device binary image property.
-// If the type size of the property value is fixed and is no greater than
-// 64 bits, then ValAddr is 0 and the value is stored in the ValSize field.
-// Example - PI_PROPERTY_TYPE_UINT32, which is 32-bit
-struct _pi_device_binary_property_struct {
-  char *Name;       // null-terminated property name
-  void *ValAddr;    // address of property value
-  uint32_t Type;    // _pi_property_type
-  uint64_t ValSize; // size of property value in bytes
-};
-
-typedef _pi_device_binary_property_struct *pi_device_binary_property;
-
-// Named array of properties.
-struct _pi_device_binary_property_set_struct {
-  char *Name;                                // the name
-  pi_device_binary_property PropertiesBegin; // array start
-  pi_device_binary_property PropertiesEnd;   // array end
-};
-
-typedef _pi_device_binary_property_set_struct *pi_device_binary_property_set;
-
-/// Types of device binary.
-using pi_device_binary_type = uint8_t;
-// format is not determined
-static constexpr pi_device_binary_type PI_DEVICE_BINARY_TYPE_NONE = 0;
-// specific to a device
-static constexpr pi_device_binary_type PI_DEVICE_BINARY_TYPE_NATIVE = 1;
-// portable binary types go next
-// SPIR-V
-static constexpr pi_device_binary_type PI_DEVICE_BINARY_TYPE_SPIRV = 2;
-// LLVM bitcode
-static constexpr pi_device_binary_type PI_DEVICE_BINARY_TYPE_LLVMIR_BITCODE = 3;
-
-// Device binary descriptor version supported by this library.
-static const uint16_t PI_DEVICE_BINARY_VERSION = 1;
-
-// The kind of offload model the binary employs; must be 4 for SYCL
-static const uint8_t PI_DEVICE_BINARY_OFFLOAD_KIND_SYCL = 4;
-
-/// Target identification strings for
-/// pi_device_binary_struct.DeviceTargetSpec
-///
-/// A device type represented by a particular target
-/// triple requires specific binary images. We need
-/// to map the image type onto the device target triple
-///
-#define __SYCL_PI_DEVICE_BINARY_TARGET_UNKNOWN "<unknown>"
-/// SPIR-V 32-bit image <-> "spir", 32-bit OpenCL device
-#define __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV32 "spir"
-/// SPIR-V 64-bit image <-> "spir64", 64-bit OpenCL device
-#define __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64 "spir64"
-/// Device-specific binary images produced from SPIR-V 64-bit <->
-/// various "spir64_*" triples for specific 64-bit OpenCL devices
-#define __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_X86_64 "spir64_x86_64"
-#define __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_GEN "spir64_gen"
-#define __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_FPGA "spir64_fpga"
-/// PTX 64-bit image <-> "nvptx64", 64-bit NVIDIA PTX device
-#define __SYCL_PI_DEVICE_BINARY_TARGET_NVPTX64 "nvptx64"
-#define __SYCL_PI_DEVICE_BINARY_TARGET_AMDGCN "amdgcn"
-#define __SYCL_PI_DEVICE_BINARY_TARGET_NATIVE_CPU "native_cpu"
-
 /// Extension to denote native support of assert feature by an arbitrary device
 /// piDeviceGetInfo call should return this extension when the device supports
 /// native asserts if supported extensions' names are requested
 #define PI_DEVICE_INFO_EXTENSION_DEVICELIB_ASSERT "cl_intel_devicelib_assert"
 
-/// Device binary image property set names recognized by the SYCL runtime.
-/// Name must be consistent with
-/// PropertySetRegistry::SYCL_SPECIALIZATION_CONSTANTS defined in
-/// PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SPEC_CONST_MAP "SYCL/specialization constants"
-/// PropertySetRegistry::SYCL_SPEC_CONSTANTS_DEFAULT_VALUES defined in
-/// PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SPEC_CONST_DEFAULT_VALUES_MAP                   \
-  "SYCL/specialization constants default values"
-/// PropertySetRegistry::SYCL_DEVICELIB_REQ_MASK defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_DEVICELIB_REQ_MASK "SYCL/devicelib req mask"
-/// PropertySetRegistry::SYCL_KERNEL_PARAM_OPT_INFO defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_KERNEL_PARAM_OPT_INFO "SYCL/kernel param opt"
-/// PropertySetRegistry::SYCL_KERNEL_PROGRAM_METADATA defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_PROGRAM_METADATA "SYCL/program metadata"
-/// PropertySetRegistry::SYCL_MISC_PROP defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_MISC_PROP "SYCL/misc properties"
-/// PropertySetRegistry::SYCL_ASSERT_USED defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_ASSERT_USED "SYCL/assert used"
-/// PropertySetRegistry::SYCL_EXPORTED_SYMBOLS defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_EXPORTED_SYMBOLS "SYCL/exported symbols"
-/// PropertySetRegistry::SYCL_IMPORTED_SYMBOLS defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_IMPORTED_SYMBOLS "SYCL/imported symbols"
-/// PropertySetRegistry::SYCL_DEVICE_GLOBALS defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_DEVICE_GLOBALS "SYCL/device globals"
-/// PropertySetRegistry::SYCL_DEVICE_REQUIREMENTS defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_DEVICE_REQUIREMENTS                        \
-  "SYCL/device requirements"
-/// PropertySetRegistry::SYCL_HOST_PIPES defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_HOST_PIPES "SYCL/host pipes"
-/// PropertySetRegistry::SYCL_VIRTUAL_FUNCTIONS defined in PropertySetIO.h
-#define __SYCL_PI_PROPERTY_SET_SYCL_VIRTUAL_FUNCTIONS "SYCL/virtual functions"
-
-/// Program metadata tags recognized by the PI backends. For kernels the tag
-/// must appear after the kernel name.
-#define __SYCL_PI_PROGRAM_METADATA_TAG_REQD_WORK_GROUP_SIZE                    \
-  "@reqd_work_group_size"
-#define __SYCL_PI_PROGRAM_METADATA_GLOBAL_ID_MAPPING "@global_id_mapping"
-
-#define __SYCL_PI_PROGRAM_METADATA_TAG_NEED_FINALIZATION "Requires finalization"
-
-/// This struct is a record of the device binary information. If the Kind field
-/// denotes a portable binary type (SPIR-V or LLVM IR), the DeviceTargetSpec
-/// field can still be specific and denote e.g. FPGA target. It must match the
-/// __tgt_device_image structure generated by the clang-offload-wrapper tool
-/// when their Version field match.
-struct pi_device_binary_struct {
-  /// version of this structure - for backward compatibility;
-  /// all modifications which change order/type/offsets of existing fields
-  /// should increment the version.
-  uint16_t Version;
-  /// the type of offload model the binary employs; must be 4 for SYCL
-  uint8_t Kind;
-  /// format of the binary data - SPIR-V, LLVM IR bitcode,...
-  uint8_t Format;
-  /// null-terminated string representation of the device's target architecture
-  /// which holds one of:
-  /// __SYCL_PI_DEVICE_BINARY_TARGET_UNKNOWN - unknown
-  /// __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV32 - general value for 32-bit OpenCL
-  /// devices
-  /// __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64 - general value for 64-bit OpenCL
-  /// devices
-  /// __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_X86_64 - 64-bit OpenCL CPU device
-  /// __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_GEN - GEN GPU device (64-bit
-  /// OpenCL)
-  /// __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_FPGA - 64-bit OpenCL FPGA device
-  const char *DeviceTargetSpec;
-  /// a null-terminated string; target- and compiler-specific options
-  /// which are suggested to use to "compile" program at runtime
-  const char *CompileOptions;
-  /// a null-terminated string; target- and compiler-specific options
-  /// which are suggested to use to "link" program at runtime
-  const char *LinkOptions;
-  /// Pointer to the manifest data start
-  const char *ManifestStart;
-  /// Pointer to the manifest data end
-  const char *ManifestEnd;
-  /// Pointer to the target code start
-  const unsigned char *BinaryStart;
-  /// Pointer to the target code end
-  const unsigned char *BinaryEnd;
-  /// the offload entry table
-  _pi_offload_entry EntriesBegin;
-  _pi_offload_entry EntriesEnd;
-  // Array of preperty sets; e.g. specialization constants symbol-int ID map is
-  // propagated to runtime with this mechanism.
-  pi_device_binary_property_set PropertySetsBegin;
-  pi_device_binary_property_set PropertySetsEnd;
-  // TODO Other fields like entries, link options can be propagated using
-  // the property set infrastructure. This will improve binary compatibility and
-  // add flexibility.
-};
-using pi_device_binary = pi_device_binary_struct *;
-
 // pi_buffer_region structure repeats cl_buffer_region, used for sub buffers.
 struct pi_buffer_region_struct {
   size_t origin;
@@ -1211,27 +1036,6 @@ struct pi_image_region_struct {
 };
 using pi_image_region = pi_image_region_struct *;
 
-// Offload binaries descriptor version supported by this library.
-static const uint16_t PI_DEVICE_BINARIES_VERSION = 1;
-
-/// This struct is a record of all the device code that may be offloaded.
-/// It must match the __tgt_bin_desc structure generated by
-/// the clang-offload-wrapper tool when their Version field match.
-struct pi_device_binaries_struct {
-  /// version of this structure - for backward compatibility;
-  /// all modifications which change order/type/offsets of existing fields
-  /// should increment the version.
-  uint16_t Version;
-  /// Number of device binaries in this descriptor
-  uint16_t NumDeviceBinaries;
-  /// Device binaries data
-  pi_device_binary DeviceBinaries;
-  /// the offload entry table (not used, for compatibility with OpenMP)
-  _pi_offload_entry *HostEntriesBegin;
-  _pi_offload_entry *HostEntriesEnd;
-};
-using pi_device_binaries = pi_device_binaries_struct *;
-
 // This union encapsulates the two external handles we currently support.
 // When choosing the correct field from the union we need to look at the value
 // of the enum `pi_external_mem_handle_type` or
@@ -1369,6 +1173,11 @@ typedef struct {
 
 using pi_launch_property = _pi_launch_property;
 
+struct sycl_device_binaries_struct;
+typedef sycl_device_binaries_struct *pi_device_binary;
+struct _sycl_device_binary_property_struct;
+typedef _sycl_device_binary_property_struct *pi_device_binary_property;
+
 //
 // Following section contains SYCL RT Plugin Interface (PI) functions.
 // They are 3 distinct categories:

sycl/include/sycl/detail/pi.hpp

@@ -105,7 +105,6 @@ using PiPlatformBackend = ::pi_platform_backend;
 using PiDevice = ::pi_device;
 using PiDeviceType = ::pi_device_type;
 using PiDeviceInfo = ::pi_device_info;
-using PiDeviceBinaryType = ::pi_device_binary_type;
 using PiContext = ::pi_context;
 using PiContextInfo = ::pi_context_info;
 using PiProgram = ::pi_program;
@@ -238,12 +237,6 @@ uint64_t emitFunctionWithArgsBeginTrace(uint32_t FuncID, const char *FName,
 void emitFunctionWithArgsEndTrace(uint64_t CorrelationID, uint32_t FuncID,
                                   const char *FName, unsigned char *ArgsData,
                                   pi_result Result, pi_plugin Plugin);
-
-/// Tries to determine the device binary image foramat. Returns
-/// PI_DEVICE_BINARY_TYPE_NONE if unsuccessful.
-PiDeviceBinaryType getBinaryImageFormat(const unsigned char *ImgData,
-                                        size_t ImgSize);
-
 } // namespace pi
 
 // Workaround for build with GCC 5.x

sycl/plugins/cuda/CMakeLists.txt

@@ -46,6 +46,7 @@ add_sycl_plugin(cuda
     ${XPTI_INCLUDE}
     ${CMAKE_CURRENT_SOURCE_DIR}/../unified_runtime # for Unified Runtime
     ${UNIFIED_RUNTIME_SOURCE_DIR}/source/ # for adapters/cuda
+    "${SYCL_SOURCE_DIR}/source" # for compiler.hpp
   LIBRARIES
     cudadrv
     ${XPTI_LIBS}

sycl/plugins/hip/CMakeLists.txt

@@ -122,6 +122,7 @@ add_sycl_plugin(hip
     ${sycl_plugin_dir}
     ${CMAKE_CURRENT_SOURCE_DIR}/../unified_runtime
     ${UNIFIED_RUNTIME_SOURCE_DIR}/source/ # for adapters/hip
+    "${SYCL_SOURCE_DIR}/source" # for compiler.hpp
   LIBRARIES
     UnifiedRuntime-Headers
     UnifiedRuntimeCommon

sycl/plugins/level_zero/CMakeLists.txt

@@ -47,6 +47,7 @@ add_sycl_plugin(level_zero
     ${CMAKE_CURRENT_SOURCE_DIR}/../unified_runtime # for Unified Runtime
     ${UNIFIED_RUNTIME_SOURCE_DIR}/source/ # for adapters/level_zero
     ${XPTI_INCLUDE}
+    "${SYCL_SOURCE_DIR}/source" # for compiler.hpp
   LIBRARIES
     LevelZeroLoader-Headers
     UnifiedRuntime-Headers

sycl/plugins/native_cpu/CMakeLists.txt

@@ -17,6 +17,7 @@ add_sycl_plugin(native_cpu
     ${sycl_inc_dir}
     ${CMAKE_CURRENT_SOURCE_DIR}/../unified_runtime # for Unified Runtime
     ${UNIFIED_RUNTIME_SOURCE_DIR}/source/ # for adapters/native_cpu
+    "${SYCL_SOURCE_DIR}/source" # for compiler.hpp
   LIBRARIES
     sycl
     UnifiedRuntime-Headers

sycl/plugins/opencl/CMakeLists.txt

@@ -17,6 +17,7 @@ add_sycl_plugin(opencl
     ${sycl_inc_dir}
     ${CMAKE_CURRENT_SOURCE_DIR}/../unified_runtime # for Unified Runtime
     ${UNIFIED_RUNTIME_SOURCE_DIR}/source/ # for adapters/opencl
+    "${SYCL_SOURCE_DIR}/source" # for compiler.hpp
   LIBRARIES
     UnifiedRuntime-Headers
     UnifiedRuntimeCommon

sycl/plugins/unified_runtime/CMakeLists.txt

@@ -229,6 +229,7 @@ set(UNIFIED_RUNTIME_PLUGIN_ARGS
   INCLUDE_DIRS
     "${UNIFIED_RUNTIME_SRC_INCLUDE_DIR}"
     "${UNIFIED_RUNTIME_COMMON_INCLUDE_DIR}"
+    "${SYCL_SOURCE_DIR}/source" # for compiler.hpp
 )
 
 # We need for #include <ze_api.h> in common.h

sycl/plugins/unified_runtime/pi2ur.hpp

@@ -9,6 +9,7 @@
 
 #include "ur_api.h"
 #include <cstdarg>
+#include <detail/compiler.hpp>
 #include <sycl/detail/cuda_definitions.hpp>
 #include <sycl/detail/pi.h>
 #include <ur/ur.hpp>
@@ -785,15 +786,15 @@ mapPIMetadataToUR(const pi_device_binary_property *pi_metadata,
   ur_metadata->pName = (*pi_metadata)->Name;
   ur_metadata->size = (*pi_metadata)->ValSize;
   switch ((*pi_metadata)->Type) {
-  case PI_PROPERTY_TYPE_UINT32:
+  case SYCL_PROPERTY_TYPE_UINT32:
     ur_metadata->type = UR_PROGRAM_METADATA_TYPE_UINT32;
     ur_metadata->value.data32 = (*pi_metadata)->ValSize;
     return UR_RESULT_SUCCESS;
-  case PI_PROPERTY_TYPE_BYTE_ARRAY:
+  case SYCL_PROPERTY_TYPE_BYTE_ARRAY:
     ur_metadata->type = UR_PROGRAM_METADATA_TYPE_BYTE_ARRAY;
     ur_metadata->value.pData = (*pi_metadata)->ValAddr;
     return UR_RESULT_SUCCESS;
-  case PI_PROPERTY_TYPE_STRING:
+  case SYCL_PROPERTY_TYPE_STRING:
     ur_metadata->type = UR_PROGRAM_METADATA_TYPE_STRING;
     ur_metadata->value.pString =
         reinterpret_cast<char *>((*pi_metadata)->ValAddr);
@@ -1487,47 +1488,48 @@ inline pi_result piGetDeviceAndHostTimer(pi_device Device, uint64_t *DeviceTime,
 
 inline pi_result
 piextDeviceSelectBinary(pi_device Device, // TODO: does this need to be context?
-                        pi_device_binary *Binaries, pi_uint32 NumBinaries,
+                        pi_device_binary *PIBinaries, pi_uint32 NumBinaries,
                         pi_uint32 *SelectedBinaryInd) {
 
   auto UrDevice = reinterpret_cast<ur_device_handle_t>(Device);
   std::vector<ur_device_binary_t> UrBinaries(NumBinaries);
+  auto *Binaries = reinterpret_cast<sycl_device_binary *>(PIBinaries);
 
   for (uint32_t BinaryCount = 0; BinaryCount < NumBinaries; BinaryCount++) {
     if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-               __SYCL_PI_DEVICE_BINARY_TARGET_UNKNOWN) == 0)
+               __SYCL_DEVICE_BINARY_TARGET_UNKNOWN) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_UNKNOWN;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV32) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_SPIRV32) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_SPIRV32;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_SPIRV64) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_SPIRV64;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_X86_64) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_SPIRV64_X86_64) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_SPIRV64_X86_64;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_GEN) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_SPIRV64_GEN) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_SPIRV64_GEN;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_SPIRV64_FPGA) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_SPIRV64_FPGA) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_SPIRV64_FPGA;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_NVPTX64) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_NVPTX64) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_NVPTX64;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_AMDGCN) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_AMDGCN) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           UR_DEVICE_BINARY_TARGET_AMDGCN;
     else if (strcmp(Binaries[BinaryCount]->DeviceTargetSpec,
-                    __SYCL_PI_DEVICE_BINARY_TARGET_NATIVE_CPU) == 0)
+                    __SYCL_DEVICE_BINARY_TARGET_NATIVE_CPU) == 0)
       UrBinaries[BinaryCount].pDeviceTargetSpec =
           "native_cpu"; // todo: define UR_DEVICE_BINARY_TARGET_NATIVE_CPU;
     else