Merge branch 'adapters' into eliminate-usage-of-regex-in-opencl

Author: omarahmed1111

.github/workflows/cmake.yml

@@ -203,6 +203,10 @@ jobs:
       run: LD_LIBRARY_PATH=${{github.workspace}}/dpcpp_compiler/lib
         cmake --build ${{github.workspace}}/build -j $(nproc)
 
+    - name: Test adapter specific
+      working-directory: ${{github.workspace}}/build
+      run: ctest -C ${{matrix.build_type}} --output-on-failure -L "adapter-specific" --timeout 180
+
     # Temporarily disabling platform test for L0, because of hang
     # See issue: #824
     - name: Test L0 adapter

examples/codegen/CMakeLists.txt

@@ -43,5 +43,5 @@ if(LLVM_FOUND AND PkgConfig_FOUND AND LLVMSPIRVLib_FOUND)
         )
     endif()
 else()
-    message(STATUS "The environment did not satisfy dependency requirements (LLVM, PkgConfig, LLVMSPIRVLib) for codegen example (skipping target).")
+    message(FATAL_ERROR "The environment did not satisfy dependency requirements (LLVM, PkgConfig, LLVMSPIRVLib) for codegen example (skipping target).")
 endif()

scripts/generate_code.py

@@ -315,6 +315,25 @@ def _mako_info_hpp(path, namespace, tags, version, specs, meta):
         specs=specs,
         meta=meta)
 
+"""
+Entry-point:
+    generates linker version scripts
+"""
+def _mako_linker_scripts(path, ext, namespace, tags, version, specs, meta):
+    name = "adapter"
+    filename = f"{name}.{ext}.in"
+    fin = os.path.join(templates_dir, f"{filename}.mako")
+    fout = os.path.join(path, filename)
+    print("Generating %s..." % fout)
+    return util.makoWrite(
+        fin, fout,
+        name=name,
+        ver=version,
+        namespace=namespace,
+        tags=tags,
+        specs=specs,
+        meta=meta)
+
 """
 Entry-point:
     generates lib code
@@ -349,6 +368,8 @@ def generate_adapters(path, section, namespace, tags, version, specs, meta):
 
     loc = 0
     loc += _mako_null_adapter_cpp(dstpath, namespace, tags, version, specs, meta)
+    loc += _mako_linker_scripts(dstpath, "map", namespace, tags, version, specs, meta)
+    loc += _mako_linker_scripts(dstpath, "def", namespace, tags, version, specs, meta)
     print("Generated %s lines of code.\n"%loc)
 
 """

scripts/templates/adapter.def.in.mako

@@ -0,0 +1,11 @@
+<%!
+import re
+from templates import helper as th
+%><%
+    n=namespace
+%>\
+LIBRARY @TARGET_LIBNAME@
+EXPORTS
+%for tbl in th.get_pfntables(specs, meta, n, tags):
+	${tbl['export']['name']}
+%endfor

scripts/templates/adapter.map.in.mako

@@ -0,0 +1,14 @@
+<%!
+import re
+from templates import helper as th
+%><%
+    n=namespace
+%>\
+@TARGET_LIBNAME@ {
+	global:
+%for tbl in th.get_pfntables(specs, meta, n, tags):
+		${tbl['export']['name']};
+%endfor
+	local:
+		*;
+};

source/adapters/adapter.def.in

@@ -1,20 +1,20 @@
 LIBRARY @TARGET_LIBNAME@
 EXPORTS
+	urGetGlobalProcAddrTable
 	urGetBindlessImagesExpProcAddrTable
 	urGetCommandBufferExpProcAddrTable
 	urGetContextProcAddrTable
-	urGetDeviceProcAddrTable
 	urGetEnqueueProcAddrTable
 	urGetEventProcAddrTable
-	urGetGlobalProcAddrTable
 	urGetKernelProcAddrTable
 	urGetMemProcAddrTable
 	urGetPhysicalMemProcAddrTable
 	urGetPlatformProcAddrTable
 	urGetProgramProcAddrTable
 	urGetQueueProcAddrTable
 	urGetSamplerProcAddrTable
+	urGetUSMProcAddrTable
 	urGetUSMExpProcAddrTable
 	urGetUsmP2PExpProcAddrTable
-	urGetUSMProcAddrTable
 	urGetVirtualMemProcAddrTable
+	urGetDeviceProcAddrTable

source/adapters/adapter.map.in

@@ -1,23 +1,23 @@
 @TARGET_LIBNAME@ {
 	global:
+		urGetGlobalProcAddrTable;
 		urGetBindlessImagesExpProcAddrTable;
 		urGetCommandBufferExpProcAddrTable;
 		urGetContextProcAddrTable;
-		urGetDeviceProcAddrTable;
 		urGetEnqueueProcAddrTable;
 		urGetEventProcAddrTable;
-		urGetGlobalProcAddrTable;
 		urGetKernelProcAddrTable;
 		urGetMemProcAddrTable;
 		urGetPhysicalMemProcAddrTable;
 		urGetPlatformProcAddrTable;
 		urGetProgramProcAddrTable;
 		urGetQueueProcAddrTable;
 		urGetSamplerProcAddrTable;
+		urGetUSMProcAddrTable;
 		urGetUSMExpProcAddrTable;
 		urGetUsmP2PExpProcAddrTable;
-		urGetUSMProcAddrTable;
 		urGetVirtualMemProcAddrTable;
+		urGetDeviceProcAddrTable;
 	local:
 		*;
 };

source/adapters/cuda/CMakeLists.txt

@@ -3,8 +3,6 @@
 # See LICENSE.TXT
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
-set(CUDA_DIR "${CMAKE_CURRENT_SOURCE_DIR}" CACHE PATH "CUDA adapter directory")
-
 set(TARGET_NAME ur_adapter_cuda)
 
 add_ur_adapter(${TARGET_NAME}

source/adapters/cuda/device.cpp

@@ -501,12 +501,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
       return ReturnValue(
           static_cast<uint64_t>(hDevice->getMaxChosenLocalMem()));
     } else {
-      int LocalMemSize = 0;
-      UR_CHECK_ERROR(cuDeviceGetAttribute(
-          &LocalMemSize, CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK,
-          hDevice->get()));
-      detail::ur::assertion(LocalMemSize >= 0);
-      return ReturnValue(static_cast<uint64_t>(LocalMemSize));
+      return ReturnValue(
+          static_cast<uint64_t>(hDevice->getMaxCapacityLocalMem()));
     }
   }
   case UR_DEVICE_INFO_ERROR_CORRECTION_SUPPORT: {

source/adapters/cuda/device.hpp

@@ -45,6 +45,9 @@ struct ur_device_handle_t_ {
     UR_CHECK_ERROR(cuDeviceGetAttribute(
         &MaxRegsPerBlock, CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK,
         cuDevice));
+    UR_CHECK_ERROR(cuDeviceGetAttribute(
+        &MaxCapacityLocalMem,
+        CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN, cuDevice));
 
     // Set local mem max size if env var is present
     static const char *LocalMemSizePtrUR =
@@ -56,9 +59,6 @@ struct ur_device_handle_t_ {
                           : (LocalMemSizePtrPI ? LocalMemSizePtrPI : nullptr);
 
     if (LocalMemSizePtr) {
-      cuDeviceGetAttribute(
-          &MaxCapacityLocalMem,
-          CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN, cuDevice);
       MaxChosenLocalMem = std::atoi(LocalMemSizePtr);
       MaxLocalMemSizeChosen = true;
     }