Merge remote-tracking branch 'origin/sycl' into llvmspirv_pulldown

Author: jsji

.github/workflows/sycl-linux-precommit.yml

@@ -69,8 +69,8 @@ jobs:
     with:
       build_ref: ${{ github.sha }}
       build_cache_root: "/__w/"
-      build_artifact_suffix: "default"
-      build_cache_suffix: "default"
+      build_artifact_suffix: "ubuntu22"
+      build_cache_suffix: "ubuntu22"
       build_image: "ghcr.io/intel/llvm/ubuntu2204_build:latest"
       changes: ${{ needs.detect_changes.outputs.filters }}
 

clang/lib/Driver/ToolChains/Clang.cpp

@@ -5930,8 +5930,16 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
         CmdArgs.push_back("-ffine-grained-bitfield-accesses");
     }
 
-    if (!Args.hasFlag(options::OPT_fsycl_unnamed_lambda,
-                      options::OPT_fno_sycl_unnamed_lambda, true))
+    // '-fsycl-unnamed-lambda' is not supported with '-fsycl-host-compiler'
+    if (Args.hasArg(options::OPT_fsycl_host_compiler_EQ)) {
+      if (Args.hasFlag(options::OPT_fsycl_unnamed_lambda,
+                       options::OPT_fno_sycl_unnamed_lambda, false))
+        D.Diag(diag::err_drv_cannot_mix_options) << "-fsycl-host-compiler"
+                                                 << "-fsycl-unnamed-lambda";
+      else // '-fsycl-host-compiler' implies '-fno-sycl-unnamed-lambda'
+        CmdArgs.push_back("-fno-sycl-unnamed-lambda");
+    } else if (!Args.hasFlag(options::OPT_fsycl_unnamed_lambda,
+                             options::OPT_fno_sycl_unnamed_lambda, true))
       CmdArgs.push_back("-fno-sycl-unnamed-lambda");
 
     if (!Args.hasFlag(options::OPT_fsycl_esimd_force_stateless_mem,

clang/test/Driver/sycl-host-compiler-old-model.cpp

@@ -62,6 +62,16 @@
 // RUN:  | FileCheck -check-prefix=HOST_COMPILER_NOARG %s
 // HOST_COMPILER_NOARG: missing argument to '-fsycl-host-compiler='
 
+/// error for -fsycl-host-compiler and -fsycl-unnamed-lambda combination 
+// RUN: not %clangxx -fsycl --no-offload-new-driver -fsycl-host-compiler=g++ -fsycl-unnamed-lambda -c -### %s 2>&1 \
+// RUN:  | FileCheck -check-prefix=HOST_COMPILER_AND_UNNAMED_LAMBDA %s
+// HOST_COMPILER_AND_UNNAMED_LAMBDA: error: cannot specify '-fsycl-unnamed-lambda' along with '-fsycl-host-compiler'
+
+// -fsycl-host-compiler implies -fno-sycl-unnamed-lambda
+// RUN: %clangxx -### -fsycl --no-offload-new-driver -fsycl-host-compiler=g++ -c -### %s 2>&1 \
+// RUN:  | FileCheck -check-prefix=IMPLY-NO-SYCL-UNNAMED-LAMBDA %s
+// IMPLY-NO-SYCL-UNNAMED-LAMBDA: clang{{.*}} "-fno-sycl-unnamed-lambda"
+
 /// Warning should not be emitted when using -fsycl-host-compiler when linking
 // RUN: touch %t.o
 // RUN: %clangxx -fsycl --no-offload-new-driver -fsycl-host-compiler=g++ %t.o -### 2>&1 \

devops/dependencies-igc-dev.json

@@ -1,10 +1,10 @@
 {
   "linux": {
     "igc_dev": {
-      "github_tag": "igc-dev-b74b7ab",
-      "version": "b74b7ab",
-      "updated_at": "2025-04-02T18:41:33Z",
-      "url": "https://api.github.com/repos/intel/intel-graphics-compiler/actions/artifacts/2869865189/zip",
+      "github_tag": "igc-dev-d1feb0f",
+      "version": "d1feb0f",
+      "updated_at": "2025-04-06T06:58:38Z",
+      "url": "https://api.github.com/repos/intel/intel-graphics-compiler/actions/artifacts/2889443018/zip",
       "root": "{DEPS_ROOT}/opencl/runtime/linux/oclgpu"
     }
   }

devops/scripts/benchmarks/benches/compute.py

@@ -46,7 +46,7 @@ def git_url(self) -> str:
         return "https://github.com/intel/compute-benchmarks.git"
 
     def git_hash(self) -> str:
-        return "b5cc46acf61766ab00da04e85bd4da4f7591eb21"
+        return "c10baa895b4364899e253e44127ff128a8efa5d5"
 
     def setup(self):
         if options.sycl is None:
@@ -145,7 +145,7 @@ def benchmarks(self) -> list[Benchmark]:
             benches.append(UllsKernelSwitch(self, runtime, 8, 200, 0, 0, 1, 1))
 
         # Add GraphApiSubmitGraph benchmarks
-        for runtime in self.enabled_runtimes([RUNTIMES.SYCL]):
+        for runtime in self.enabled_runtimes([RUNTIMES.SYCL, RUNTIMES.UR]):
             for in_order_queue in [0, 1]:
                 for num_kernels in [4, 10, 32]:
                     for measure_completion_time in [0, 1]:

devops/scripts/benchmarks/benches/llamacpp.py

@@ -29,7 +29,7 @@ def git_url(self) -> str:
         return "https://github.com/ggerganov/llama.cpp"
 
     def git_hash(self) -> str:
-        return "1ee9eea094fe5846c7d8d770aa7caa749d246b23"
+        return "916c83bfe7f8b08ada609c3b8e583cf5301e594b"
 
     def setup(self):
         if options.sycl is None:
@@ -47,9 +47,9 @@ def setup(self):
 
         self.model = download(
             self.models_dir,
-            "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf/resolve/main/Phi-3-mini-4k-instruct-q4.gguf",
-            "Phi-3-mini-4k-instruct-q4.gguf",
-            checksum="fc4f45c9729874a33a527465b2ec78189a18e5726b7121182623feeae38632ace4f280617b01d4a04875acf49d263ee4",
+            "https://huggingface.co/ggml-org/DeepSeek-R1-Distill-Qwen-1.5B-Q4_0-GGUF/resolve/main/deepseek-r1-distill-qwen-1.5b-q4_0.gguf",
+            "deepseek-r1-distill-qwen-1.5b-q4_0.gguf",
+            checksum="791f6091059b653a24924b9f2b9c3141c8f892ae13fff15725f77a2bf7f9b1b6b71c85718f1e9c0f26c2549aba44d191",
         )
 
         self.oneapi = get_oneapi()
@@ -64,10 +64,11 @@ def setup(self):
             f"-DGGML_SYCL=ON",
             f"-DCMAKE_C_COMPILER=clang",
             f"-DCMAKE_CXX_COMPILER=clang++",
-            f"-DDNNL_DIR={self.oneapi.dnn_cmake()}",
+            f"-DDNNL_GPU_VENDOR=INTEL",
             f"-DTBB_DIR={self.oneapi.tbb_cmake()}",
-            f'-DCMAKE_CXX_FLAGS=-I"{self.oneapi.mkl_include()}"',
-            f"-DCMAKE_SHARED_LINKER_FLAGS=-L{self.oneapi.compiler_lib()} -L{self.oneapi.mkl_lib()}",
+            f"-DDNNL_DIR={self.oneapi.dnn_cmake()}",
+            f"-DSYCL_COMPILER=ON",
+            f"-DMKL_DIR={self.oneapi.mkl_cmake()}",
         ]
 
         run(configure_command, add_sycl=True)
@@ -96,14 +97,17 @@ def __init__(self, bench):
     def setup(self):
         self.benchmark_bin = os.path.join(self.bench.build_path, "bin", "llama-bench")
 
+    def model(self):
+        return "DeepSeek-R1-Distill-Qwen-1.5B-Q4_0.gguf"
+
     def name(self):
-        return f"llama.cpp"
+        return f"llama.cpp {self.model()}"
 
     def description(self) -> str:
         return (
             "Performance testing tool for llama.cpp that measures LLM inference speed in tokens per second. "
             "Runs both prompt processing (initial context processing) and text generation benchmarks with "
-            "different batch sizes. Higher values indicate better performance. Uses the Phi-3-mini-4k-instruct "
+            f"different batch sizes. Higher values indicate better performance. Uses the {self.model()} "
             "quantized model and leverages SYCL with oneDNN for acceleration."
         )
 
@@ -122,12 +126,18 @@ def run(self, env_vars) -> list[Result]:
             "128",
             "-p",
             "512",
-            "-b",
-            "128,256,512",
+            "-pg",
+            "0,0",
+            "-sm",
+            "none",
+            "-ngl",
+            "99",
             "--numa",
             "isolate",
             "-t",
-            "56",  # TODO: use only as many threads as numa node 0 has cpus
+            "8",
+            "--mmap",
+            "0",
             "--model",
             f"{self.bench.model}",
         ]

devops/scripts/benchmarks/options.py

@@ -40,7 +40,7 @@ class Options:
     build_compute_runtime: bool = False
     extra_ld_libraries: list[str] = field(default_factory=list)
     extra_env_vars: dict = field(default_factory=dict)
-    compute_runtime_tag: str = "25.09.32961.7"
+    compute_runtime_tag: str = "25.09.32961.8"
     build_igc: bool = False
     current_run_name: str = "This PR"
     preset: str = "Full"

devops/scripts/benchmarks/utils/compute_runtime.py

@@ -135,6 +135,8 @@ def build_igc(self, repo, commit):
         self.igc_install = os.path.join(options.workdir, "igc-install")
         configure_command = [
             "cmake",
+            "-DCMAKE_C_FLAGS=-Wno-error",
+            "-DCMAKE_CXX_FLAGS=-Wno-error",
             f"-B {self.igc_build}",
             f"-S {self.igc_repo}",
             f"-DCMAKE_INSTALL_PREFIX={self.igc_install}",

devops/scripts/benchmarks/utils/oneapi.py

@@ -16,16 +16,10 @@ def __init__(self):
         Path(self.oneapi_dir).mkdir(parents=True, exist_ok=True)
         self.oneapi_instance_id = self.generate_unique_oneapi_id(self.oneapi_dir)
 
-        # can we just hardcode these links?
         self.install_package(
-            "dnnl",
-            "https://registrationcenter-download.intel.com/akdlm/IRC_NAS/87e117ab-039b-437d-9c80-dcd5c9e675d5/intel-onednn-2025.0.0.862_offline.sh",
-            "6866feb5b8dfefd6ff45d6bfabed44f01d7fba8fd452480ae1fd86b92e9481ae052c24842da14f112f672f5c4859945b",
-        )
-        self.install_package(
-            "mkl",
-            "https://registrationcenter-download.intel.com/akdlm/IRC_NAS/79153e0f-74d7-45af-b8c2-258941adf58a/intel-onemkl-2025.0.0.940_offline.sh",
-            "122bb84cf943ea27753cb399c81ab2ae218ebd51b789c74d273240157722925ab4d5a43cb0b5de41b854f2c5a59a4002",
+            "base",
+            "https://registrationcenter-download.intel.com/akdlm/IRC_NAS/cca951e1-31e7-485e-b300-fe7627cb8c08/intel-oneapi-base-toolkit-2025.1.0.651_offline.sh",
+            "98cad2489f2c90a2b328568a59371cf35855a3338643f61a9fc2d16a265d29f22feb2d673916dd7be18fa12a5e6d2475",
         )
         return
 

devops/scripts/benchmarks/utils/utils.py

@@ -9,10 +9,11 @@
 import subprocess
 
 import tarfile
-import urllib  # nosec B404
 from options import options
 from pathlib import Path
 import hashlib
+from urllib.request import urlopen  # nosec B404
+from shutil import copyfileobj
 
 
 def run(
@@ -147,7 +148,9 @@ def download(dir, url, file, untar=False, unzip=False, checksum=""):
     data_file = os.path.join(dir, file)
     if not Path(data_file).exists():
         print(f"{data_file} does not exist, downloading")
-        urllib.request.urlretrieve(url, data_file)
+        with urlopen(url) as in_stream, open(data_file, "wb") as out_file:
+            copyfileobj(in_stream, out_file)
+
         calculated_checksum = calculate_checksum(data_file)
         if calculated_checksum != checksum:
             print(

llvm/lib/SYCLNativeCPUUtils/CMakeLists.txt

@@ -35,15 +35,15 @@ endif()
 if(NATIVECPU_USE_OCK)
   if(NATIVECPU_OCK_USE_FETCHCONTENT)
     set(OCK_GIT_INTERNAL_REPO "https://github.com/uxlfoundation/oneapi-construction-kit.git")
-    # commit 652f9943b0bfcf61123a0f82b2bdd499738c5391
-    # Merge: c24c971bcb e0f02ae739
-    # Author: Colin Davidson <colin.davidson@codeplay.com>
-    # Date:   Fri Feb 28 15:59:50 2025 +0000
+    # commit ffef31717750d3f15bfd2f18d9cd7f4677fe9d3e
+    # Merge: 574307afde 0b00238554
+    # Author: Harald van Dijk <harald.vandijk@codeplay.com>
+    # Date:   Fri Apr 4 16:26:38 2025 +0100
     # 
-    #     Merge pull request #677 from coldav/colin/build_installed_llvm
+    #     Merge pull request #751 from hvdijk/llvm21-address-space
     #     
-    #     Change llvm over to a choice of install or cache, default PRs to install
-    set(OCK_GIT_INTERNAL_TAG 652f9943b0bfcf61123a0f82b2bdd499738c5391)
+    #     [LLVM 21] Take address space into account for legality.
+    set(OCK_GIT_INTERNAL_TAG ffef31717750d3f15bfd2f18d9cd7f4677fe9d3e)
 
     # Overwrite OCK_GIT_INTERNAL_REPO/OCK_GIT_INTERNAL_TAG if the corresponding options are set
     if(OCK_GIT_REPO)

sycl/include/sycl/handler.hpp

@@ -1192,11 +1192,8 @@ class __SYCL_EXPORT handler {
         typename detail::get_kernel_name_t<KernelName, KernelType>::name;
 
     // Range rounding can be disabled by the user.
-    // Range rounding is not done on the host device.
     // Range rounding is supported only for newer SYCL standards.
 #if !defined(__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__) &&                  \
-    !defined(DPCPP_HOST_DEVICE_OPENMP) &&                                      \
-    !defined(DPCPP_HOST_DEVICE_PERF_NATIVE) &&                                 \
     SYCL_LANGUAGE_VERSION >= 202012L
     auto [RoundedRange, HasRoundedRange] = getRoundedRange(UserRange);
     if (HasRoundedRange) {
@@ -1226,7 +1223,6 @@ class __SYCL_EXPORT handler {
 #endif
     } else
 #endif // !__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__ &&
-       // !DPCPP_HOST_DEVICE_OPENMP && !DPCPP_HOST_DEVICE_PERF_NATIVE &&
        // SYCL_LANGUAGE_VERSION >= 202012L
     {
       (void)UserRange;

sycl/source/backend.cpp

@@ -299,7 +299,8 @@ make_kernel_bundle(ur_native_handle_t NativeHandle,
   // symbols (e.g. when kernel_bundle is supposed to be joined with another).
   auto KernelIDs = std::make_shared<std::vector<kernel_id>>();
   auto DevImgImpl = std::make_shared<device_image_impl>(
-      nullptr, TargetContext, Devices, State, KernelIDs, UrProgram);
+      nullptr, TargetContext, Devices, State, KernelIDs, UrProgram,
+      ImageOriginInterop);
   device_image_plain DevImg{DevImgImpl};
 
   return std::make_shared<kernel_bundle_impl>(TargetContext, Devices, DevImg);