[WIP] gpu: intel: add Level Zero backend

README.md

@@ -279,9 +279,9 @@ Runtime-specific dependencies:
 | `ONEDNN_CPU_RUNTIME=OMP`   | Intel C/C++ Compiler          | Intel OpenMP runtime (`libiomp5.so`)
 | `ONEDNN_CPU_RUNTIME=OMP`   | Clang                         | Intel OpenMP runtime (`libiomp5.so`)
 | `ONEDNN_CPU_RUNTIME=TBB`   | any                           | TBB (`libtbb.so`)
-| `ONEDNN_CPU_RUNTIME=SYCL`  | Intel oneAPI DPC++ Compiler   | Intel oneAPI DPC++ Compiler runtime (`libsycl.so`), TBB (`libtbb.so`), OpenCL loader (`libOpenCL.so`)
+| `ONEDNN_CPU_RUNTIME=SYCL`  | Intel oneAPI DPC++ Compiler   | Intel oneAPI DPC++ Compiler runtime (`libsycl.so`), TBB (`libtbb.so`), OpenCL loader (`libOpenCL.so`), oneAPI Level Zero loader (`libze_loader.so`)
 | `ONEDNN_GPU_RUNTIME=OCL`   | any                           | OpenCL loader (`libOpenCL.so`)
-| `ONEDNN_GPU_RUNTIME=SYCL`  | Intel oneAPI DPC++ Compiler   | Intel oneAPI DPC++ Compiler runtime (`libsycl.so`), OpenCL loader (`libOpenCL.so`), oneAPI Level Zero loader (`libze_loader.so`)
+| `ONEDNN_GPU_RUNTIME=L0`    | any                           | oneAPI Level Zero loader (`libze_loader.so`)
 
 #### Windows
 
@@ -298,6 +298,7 @@ Runtime-specific dependencies:
 | `ONEDNN_CPU_RUNTIME=SYCL`  | Intel oneAPI DPC++ Compiler   | Intel oneAPI DPC++ Compiler runtime (`sycl.dll`), TBB (`tbb.dll`), OpenCL loader (`OpenCL.dll`)
 | `ONEDNN_GPU_RUNTIME=OCL`   | any                           | OpenCL loader (`OpenCL.dll`)
 | `ONEDNN_GPU_RUNTIME=SYCL`  | Intel oneAPI DPC++ Compiler   | Intel oneAPI DPC++ Compiler runtime (`sycl.dll`), OpenCL loader (`OpenCL.dll`), oneAPI Level Zero loader (`ze_loader.dll`)
+| `ONEDNN_GPU_RUNTIME=L0`    | any                           | oneAPI Level Zero loader (`ze_loader.dll`)
 
 #### macOS
 

cmake/OpenCL.cmake

@@ -25,7 +25,7 @@ set(OpenCL_cmake_included true)
 if(DNNL_GPU_SYCL AND DNNL_GPU_VENDOR STREQUAL "INTEL")
     add_definitions_with_host_compiler(-DCL_TARGET_OPENCL_VERSION=300)
 else()
-    add_definitions(-DCL_TARGET_OPENCL_VERSION=120)
+    add_definitions(-DCL_TARGET_OPENCL_VERSION=300)
 endif()
 
 if(OpenCL_INCLUDE_DIR)

cmake/options.cmake

@@ -287,7 +287,7 @@ set(DNNL_GPU_RUNTIME "NONE" CACHE STRING
 
     Using OpenCL for GPU requires setting OPENCLROOT if the libraries are
     installed in a non-standard location.")
-if(NOT "${DNNL_GPU_RUNTIME}" MATCHES "^(OCL|NONE|DPCPP|SYCL)$")
+if(NOT "${DNNL_GPU_RUNTIME}" MATCHES "^(OCL|NONE|DPCPP|SYCL|L0)$")
     message(FATAL_ERROR "Unsupported GPU runtime: ${DNNL_GPU_RUNTIME}")
 endif()
 

doc/build/build_options.md

@@ -7,7 +7,7 @@ oneDNN supports the following build-time options.
 |:--------------------------------|:----------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------|
 | ONEDNN_LIBRARY_TYPE             | **SHARED**, STATIC                                  | Defines the resulting library type                                                                                 |
 | ONEDNN_CPU_RUNTIME              | NONE, **OMP**, TBB, SEQ, THREADPOOL, SYCL           | Defines the threading runtime for CPU engines                                                                      |
-| ONEDNN_GPU_RUNTIME              | **NONE**, OCL, SYCL                                 | Defines the offload runtime for GPU engines                                                                        |
+| ONEDNN_GPU_RUNTIME              | **NONE**, OCL, SYCL, L0                             | Defines the offload runtime for GPU engines                                                                        |
 | ONEDNN_BUILD_DOC                | **ON**, OFF                                         | Controls building the documentation                                                                                |
 | ONEDNN_DOC_VERSIONS_JSON        | **""**, *string*                                    | Location of JSON file for [PyData Sphinx Theme version switcher]. Enables documentation version switcher when set. |
 | ONEDNN_BUILD_EXAMPLES           | **ON**, OFF                                         | Controls building the examples                                                                                     |

examples/CMakeLists.txt

@@ -147,6 +147,9 @@ foreach(f ${sources})
     if(NOT DNNL_WITH_SYCL AND ${f_name} MATCHES "^sycl")
         list(REMOVE_ITEM sources "${f}")
     endif()
+    if(NOT DNNL_GPU_RUNTIME STREQUAL "L0" AND ${f_name} MATCHES ".*level_zero")
+        list(REMOVE_ITEM sources "${f}")
+    endif()
 endforeach()
 
 # In case of SYCL, skip CPU examples that directly work with raw pointers
@@ -202,7 +205,7 @@ foreach(src ${sources})
         endif()
     else()
         set(cpu_rt_pattern "(SEQ|OMP|TBB|SYCL|DPCPP)")
-        set(gpu_rt_pattern "(OCL|SYCL|DPCPP)")
+        set(gpu_rt_pattern "(OCL|L0|SYCL|DPCPP)")
         if(${example_name} MATCHES "sycl.*")
             set(cpu_rt_pattern "(SYCL|DPCPP)")
             set(gpu_rt_pattern "(SYCL|DPCPP)")

examples/CMakeLists.txt.in

@@ -214,7 +214,7 @@ foreach(src ${sources})
         endif()
     else()
         set(cpu_rt_pattern "(SEQ|OMP|TBB|SYCL|DPCPP)")
-        set(gpu_rt_pattern "(OCL|SYCL|DPCPP)")
+        set(gpu_rt_pattern "(OCL|L0|SYCL|DPCPP)")
         if(${example_name} MATCHES "sycl.*")
             set(cpu_rt_pattern "(SYCL|DPCPP)")
             set(gpu_rt_pattern "(SYCL|DPCPP)")

examples/example_utils.h

@@ -29,6 +29,9 @@
 #if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
 #include "dnnl_ocl.h"
 #endif
+#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_L0
+#include "dnnl_l0.h"
+#endif
 
 #define COMPLAIN_DNNL_ERROR_AND_EXIT(what, status) \
     do { \
@@ -160,7 +163,7 @@ static inline void write_to_dnnl_memory(void *handle, dnnl_memory_t mem) {
     }
 #endif
 
-#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
+#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL || DNNL_GPU_RUNTIME == DNNL_RUNTIME_L0
     if (eng_kind == dnnl_gpu) {
         void *mapped_ptr = NULL;
         CHECK(dnnl_memory_map_data(mem, &mapped_ptr));

examples/example_utils.hpp

@@ -35,7 +35,11 @@
 
 #if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
 #include "dnnl_ocl.hpp"
-#elif DNNL_GPU_RUNTIME == DNNL_RUNTIME_SYCL
+#endif
+#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_L0
+#include "dnnl_l0.hpp"
+#endif
+#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_SYCL
 #include "dnnl_sycl.hpp"
 #endif
 
@@ -228,7 +232,7 @@ inline void read_from_dnnl_memory(void *handle, dnnl::memory &mem) {
         return;
     }
 #endif
-#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
+#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL || DNNL_GPU_RUNTIME == DNNL_RUNTIME_L0
     if (eng.get_kind() == dnnl::engine::kind::gpu) {
         void *mapped_ptr = mem.map_data();
         if (mapped_ptr) std::memcpy(handle, mapped_ptr, size);
@@ -287,7 +291,7 @@ inline void write_to_dnnl_memory(void *handle, dnnl::memory &mem) {
         return;
     }
 #endif
-#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
+#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL || DNNL_GPU_RUNTIME == DNNL_RUNTIME_L0
     if (eng.get_kind() == dnnl::engine::kind::gpu) {
         void *mapped_ptr = mem.map_data();
         if (mapped_ptr) std::memcpy(mapped_ptr, handle, size);

include/dnnl_l0.h

@@ -0,0 +1,22 @@
+/*******************************************************************************
+* Copyright 2025 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef DNNL_L0_H
+#define DNNL_L0_H
+
+#include "oneapi/dnnl/dnnl_l0.h"
+
+#endif /* DNNL_L0_H */

include/dnnl_l0.hpp

@@ -0,0 +1,22 @@
+/*******************************************************************************
+* Copyright 2025 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef DNNL_L0_HPP
+#define DNNL_L0_HPP
+
+#include "oneapi/dnnl/dnnl_l0.hpp"
+
+#endif /* DNNL_L0_HPP */

include/oneapi/dnnl/dnnl_config.h.in

@@ -82,6 +82,9 @@
 /// DPC++ runtime
 #define DNNL_RUNTIME_DPCPP DNNL_RUNTIME_SYCL
 
+/// L0 runtime
+#define DNNL_RUNTIME_L0 1024u
+
 /// No vendor (corresponding runtime is disabled)
 #define DNNL_VENDOR_NONE 0u
 
@@ -119,7 +122,8 @@
 #endif
 #if (DNNL_GPU_RUNTIME != DNNL_RUNTIME_NONE) \
         && (DNNL_GPU_RUNTIME != DNNL_RUNTIME_OCL) \
-        && (DNNL_GPU_RUNTIME != DNNL_RUNTIME_SYCL)
+        && (DNNL_GPU_RUNTIME != DNNL_RUNTIME_SYCL) \
+        && (DNNL_GPU_RUNTIME != DNNL_RUNTIME_L0)
 #error "Unexpected DNNL_GPU_RUNTIME"
 #endif
 #if (DNNL_CPU_RUNTIME == DNNL_RUNTIME_NONE \
@@ -145,9 +149,6 @@
 // When defined, DPCPP is supported.
 #cmakedefine DNNL_WITH_SYCL
 
-// When defined, Level Zero is supported.
-#cmakedefine DNNL_WITH_LEVEL_ZERO
-
 // When defined, SYCL CUDA backend is used.
 #cmakedefine DNNL_SYCL_CUDA