Merge pull request #1959 from igchor/v2_enqueue_kernel_separate_adapter

[L0 v2] implement enqueueKernelLaunch and queueFinish

Author: igchor

source/adapters/level_zero/CMakeLists.txt

@@ -179,11 +179,15 @@ if(UR_BUILD_ADAPTER_L0_V2)
         ${CMAKE_CURRENT_SOURCE_DIR}/common.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/device.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/platform.hpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/program.hpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/helpers/kernel_helpers.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/adapter.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/common.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/ur_interface_loader.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/platform.cpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/program.cpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/helpers/kernel_helpers.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/../../ur/ur.cpp
         # v2-only sources
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/command_list_cache.hpp
@@ -194,8 +198,10 @@ if(UR_BUILD_ADAPTER_L0_V2)
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/event_provider_normal.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/event_provider.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/event.hpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/v2/kernel.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_api.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_in_order.hpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/v2/usm.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/api.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/command_list_cache.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/context.cpp
@@ -204,9 +210,11 @@ if(UR_BUILD_ADAPTER_L0_V2)
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/event_provider_counter.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/event_provider_normal.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/event.cpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/v2/kernel.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_api.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_create.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_in_order.cpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/v2/usm.cpp
     )
 
     if(NOT WIN32)
@@ -220,6 +228,8 @@ if(UR_BUILD_ADAPTER_L0_V2)
         )
     endif()
 
+    target_compile_definitions(ur_adapter_level_zero_v2 PUBLIC UR_ADAPTER_LEVEL_ZERO_V2)
+
     # TODO: fix level_zero adapter conversion warnings
     target_compile_options(ur_adapter_level_zero_v2 PRIVATE
         $<$<CXX_COMPILER_ID:MSVC>:/wd4805 /wd4244>
@@ -245,6 +255,7 @@ if(UR_BUILD_ADAPTER_L0_V2)
 
     target_include_directories(ur_adapter_level_zero_v2 PRIVATE
         "${CMAKE_CURRENT_SOURCE_DIR}/../.."
+        "${CMAKE_CURRENT_SOURCE_DIR}/../../ur"
         LevelZeroLoader-Headers
     )
 endif()

source/adapters/level_zero/command_buffer.hpp

@@ -17,6 +17,7 @@
 #include "common.hpp"
 
 #include "context.hpp"
+#include "kernel.hpp"
 #include "queue.hpp"
 
 struct command_buffer_profiling_t {

source/adapters/level_zero/common.hpp

@@ -25,6 +25,8 @@
 
 #include <umf_pools/disjoint_pool_config_parser.hpp>
 
+#include "logger/ur_logger.hpp"
+
 struct _ur_platform_handle_t;
 
 static auto getUrResultString = [](ur_result_t Result) {
@@ -168,7 +170,7 @@ static auto getUrResultString = [](ur_result_t Result) {
   }
 };
 
-// Trace an internal PI call; returns in case of an error.
+// Trace an internal UR call; returns in case of an error.
 #define UR_CALL(Call)                                                          \
   {                                                                            \
     if (PrintTrace)                                                            \
@@ -180,6 +182,18 @@ static auto getUrResultString = [](ur_result_t Result) {
       return Result;                                                           \
   }
 
+// Trace an internal UR call; throw in case of an error.
+#define UR_CALL_THROWS(Call)                                                   \
+  {                                                                            \
+    if (PrintTrace)                                                            \
+      logger::always("UR ---> {}", #Call);                                     \
+    ur_result_t Result = (Call);                                               \
+    if (PrintTrace)                                                            \
+      logger::always("UR <--- {}({})", #Call, getUrResultString(Result));      \
+    if (Result != UR_RESULT_SUCCESS)                                           \
+      throw Result;                                                            \
+  }
+
 // Controls UR L0 calls tracing.
 enum UrDebugLevel {
   UR_L0_DEBUG_NONE = 0x0,

source/adapters/level_zero/context.cpp

@@ -829,3 +829,12 @@ bool ur_context_handle_t_::isValidDevice(ur_device_handle_t Device) const {
   }
   return false;
 }
+
+const std::vector<ur_device_handle_t> &
+ur_context_handle_t_::getDevices() const {
+  return Devices;
+}
+
+ze_context_handle_t ur_context_handle_t_::getZeHandle() const {
+  return ZeContext;
+}

source/adapters/level_zero/context.hpp

@@ -193,6 +193,9 @@ struct ur_context_handle_t_ : _ur_object {
   // Return the Platform, which is the same for all devices in the context
   ur_platform_handle_t getPlatform() const;
 
+  // Get vector of devices from this context
+  const std::vector<ur_device_handle_t> &getDevices() const;
+
   // Get index of the free slot in the available pool. If there is no available
   // pool then create new one. The HostVisible parameter tells if we need a
   // slot for a host-visible event. The ProfilingEnabled tells is we need a
@@ -303,6 +306,9 @@ struct ur_context_handle_t_ : _ur_object {
   // For that the Device or its root devices need to be in the context.
   bool isValidDevice(ur_device_handle_t Device) const;
 
+  // Get handle to the L0 context
+  ze_context_handle_t getZeHandle() const;
+
 private:
   // Get the cache of events for a provided scope and profiling mode.
   auto getEventCache(bool HostVisible, bool WithProfiling,

source/adapters/level_zero/helpers/kernel_helpers.cpp

@@ -12,9 +12,14 @@
 #include "logger/ur_logger.hpp"
 
 #include "../common.hpp"
-#include "../context.hpp"
 #include "../device.hpp"
 
+#ifdef UR_ADAPTER_LEVEL_ZERO_V2
+#include "../context.hpp"
+#else
+#include "../v2/context.hpp"
+#endif
+
 ur_result_t getSuggestedLocalWorkSize(ur_device_handle_t hDevice,
                                       ze_kernel_handle_t hZeKernel,
                                       size_t GlobalWorkSize3D[3],

source/adapters/level_zero/platform.cpp

@@ -503,6 +503,8 @@ ur_result_t ur_platform_handle_t_::populateDeviceCacheIfNeeded() {
   return UR_RESULT_SUCCESS;
 }
 
+size_t ur_platform_handle_t_::getNumDevices() { return URDevicesCache.size(); }
+
 ur_device_handle_t ur_platform_handle_t_::getDeviceById(DeviceId id) {
   for (auto &dev : URDevicesCache) {
     if (dev->Id == id) {

source/adapters/level_zero/platform.hpp

@@ -56,6 +56,8 @@ struct ur_platform_handle_t_ : public _ur_platform {
   // Check the device cache and load it if necessary.
   ur_result_t populateDeviceCacheIfNeeded();
 
+  size_t getNumDevices();
+
   ur_device_handle_t getDeviceById(DeviceId);
 
   // Return the PI device from cache that represents given native device.

source/adapters/level_zero/program.cpp

@@ -9,8 +9,14 @@
 //===----------------------------------------------------------------------===//
 
 #include "program.hpp"
+#include "device.hpp"
 #include "logger/ur_logger.hpp"
-#include "ur_level_zero.hpp"
+
+#ifdef UR_ADAPTER_LEVEL_ZERO_V2
+#include "context.hpp"
+#else
+#include "v2/context.hpp"
+#endif
 
 extern "C" {
 // Check to see if a Level Zero module has any unresolved symbols.
@@ -115,8 +121,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramBuild(
     const char *Options          ///< [in][optional] pointer to build options
                                  ///< null-terminated string.
 ) {
-  return urProgramBuildExp(Program, Context->Devices.size(),
-                           Context->Devices.data(), Options);
+  std::vector<ur_device_handle_t> Devices = Context->getDevices();
+  return urProgramBuildExp(Program, Devices.size(), Devices.data(), Options);
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL urProgramBuildExp(
@@ -174,7 +180,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramBuildExp(
 
   for (uint32_t i = 0; i < numDevices; i++) {
     ze_device_handle_t ZeDevice = phDevices[i]->ZeDevice;
-    ze_context_handle_t ZeContext = hProgram->Context->ZeContext;
+    ze_context_handle_t ZeContext = hProgram->Context->getZeHandle();
     ze_module_handle_t ZeModuleHandle = nullptr;
     ze_module_build_log_handle_t ZeBuildLog{};
 
@@ -266,7 +272,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramCompile(
     // ze-opt-greater-than-4GB-buffer-required to disable
     // stateful optimizations and be able to use larger than
     // 4GB allocations on these kernels.
-    if (Context->Devices[0]->useRelaxedAllocationLimits()) {
+    if (Context->getDevices()[0]->useRelaxedAllocationLimits()) {
       Program->BuildFlags += " -ze-opt-greater-than-4GB-buffer-required";
     }
   }
@@ -285,9 +291,9 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramLink(
     ur_program_handle_t
         *Program ///< [out] pointer to handle of program object created.
 ) {
-  return urProgramLinkExp(Context, Context->Devices.size(),
-                          Context->Devices.data(), Count, Programs, Options,
-                          Program);
+  std::vector<ur_device_handle_t> Devices = Context->getDevices();
+  return urProgramLinkExp(Context, Devices.size(), Devices.data(), Count,
+                          Programs, Options, Program);
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL urProgramLinkExp(
@@ -426,7 +432,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramLinkExp(
 
       // Call the Level Zero API to compile, link, and create the module.
       ze_device_handle_t ZeDevice = phDevices[i]->ZeDevice;
-      ze_context_handle_t ZeContext = hContext->ZeContext;
+      ze_context_handle_t ZeContext = hContext->getZeHandle();
       ze_module_handle_t ZeModule = nullptr;
       ze_module_build_log_handle_t ZeBuildLog = nullptr;
       ze_result_t ZeResult =
@@ -652,7 +658,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramGetInfo(
       for (auto &ZeModulePair : Program->ZeModuleMap) {
         auto It = Program->ZeModuleMap.find(ZeModulePair.first);
         if (It != Program->ZeModuleMap.end()) {
-          for (auto &Device : Program->Context->Devices) {
+          for (auto &Device : Program->Context->getDevices()) {
             if (Device->ZeDevice == ZeModulePair.first) {
               devices.push_back(Device);
             }
@@ -661,7 +667,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramGetInfo(
       }
       return ReturnValue(devices.data(), devices.size());
     } else {
-      return ReturnValue(Program->Context->Devices[0]);
+      return ReturnValue(Program->Context->getDevices()[0]);
     }
   case UR_PROGRAM_INFO_BINARY_SIZES: {
     std::shared_lock<ur_shared_mutex> Guard(Program->Mutex);