PR test preloads llvm

.github/workflows/docs.yml

@@ -3,7 +3,7 @@ name: GitHubPages
 
 on:
   push:
-    branches: ["main"]
+    branches: ["test_prealoads_llvm"]
   workflow_dispatch:
 
 # Cancel previous in-progress workflow, only the latest run is relevant
@@ -15,8 +15,19 @@ permissions:
   contents: read
 
 jobs:
+  Benchmarks:
+    uses: ./.github/workflows/reusable_benchmarks.yml
+    permissions:
+      contents: read
+      pull-requests: write
+    with:
+      pr_no: '0'
+      bench_script_params: '--save baseline'
+      upload_report: true
+
   DocsBuild:
     uses: ./.github/workflows/reusable_docs_build.yml
+    needs: Benchmarks
     with:
       upload: true
 

.github/workflows/reusable_benchmarks.yml

@@ -1,6 +1,5 @@
-# Executes benchmarks implemented in this repository
-# using scripts for benchmark results visualization,
-# which are downloaded from Unified Runtime repository. 
+# Executes benchmarks implemented in this repository using scripts
+# for results visualization from intel/llvm (unified-runtime dir).
 name: Benchmarks
 
 on:
@@ -31,9 +30,10 @@ env:
 jobs:
   benchmarks:
     name: Benchmarks
+    runs-on: ubuntu-latest
     # run only on upstream; forks will not have the HW
-    if: github.repository == 'oneapi-src/unified-memory-framework'
-    runs-on: L0_PERF
+    # if: github.repository == 'oneapi-src/unified-memory-framework'
+    # runs-on: L0_PERF
 
     steps:
     # Workspace on self-hosted runners is not cleaned automatically.
@@ -44,6 +44,9 @@ jobs:
         ls -la ./
         rm -rf ./* || true
 
+    - name: Install things
+      run: sudo apt-get install libhwloc-dev libjemalloc-dev libtbb-dev
+
     - name: Add comment to PR
       uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
       if: ${{ always() && inputs.pr_no != 0 }}
@@ -98,50 +101,53 @@ jobs:
     - name: Build UMF
       run: cmake --build ${{env.BUILD_DIR}} -j $(nproc)
 
-    # We are going to clone Unified Runtime repository in order to run
-    # the most up-to-date UR scripts for benchmark data visualization
-    - name: Checkout UR
+    # Get scripts for benchmark data visualization.
+    # Use specific tag, as the scripts or files' location may change.
+    - name: Checkout SYCL
       uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
       with:
-        repository: oneapi-src/unified-runtime
-        path: ur-repo
+        repository: EuphoricThinking/llvm
+        ref: umf_preloaded_bench
+        path: sycl-repo
         fetch-depth: 1
-        fetch-tags: false
-
-    - name: Install pip packages for benchmarking scripts from UR
-      run: |
-        pip install --force-reinstall -r ${{github.workspace}}/ur-repo/third_party/benchmark_requirements.txt
 
-    - name: Set core range and GPU mask
+    - name: Install benchmarking scripts deps
       run: |
-        # Compute the core range for the second NUMA node; first node is for UR jobs.
-        # Skip the first 4 cores - the kernel is likely to schedule more work on these.
-        CORES=$(lscpu | awk '
-          /NUMA node1 CPU|On-line CPU/ {line=$0}
-          END {
-            split(line, a, " ")
-            split(a[4], b, ",")
-            sub(/^0/, "4", b[1])
-            print b[1]
-          }')
-        echo "Selected core: $CORES"
-        echo "CORES=$CORES" >> $GITHUB_ENV
-
-        ZE_AFFINITY_MASK=1
-        echo "ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK" >> $GITHUB_ENV
-
-    - name: Run UMF benchmarks (using scripts from UR)
+        pip install --force-reinstall -r ${{github.workspace}}/sycl-repo/unified-runtime/third_party/benchmark_requirements.txt
+
+    # - name: Set core range and GPU mask
+    #   run: |
+    #     # Compute the core range for the second NUMA node; first node is for SYCL/UR jobs.
+    #     # Skip the first 4 cores - the kernel is likely to schedule more work on these.
+    #     CORES=$(lscpu | awk '
+    #       /NUMA node1 CPU|On-line CPU/ {line=$0}
+    #       END {
+    #         split(line, a, " ")
+    #         split(a[4], b, ",")
+    #         sub(/^0/, "4", b[1])
+    #         print b[1]
+    #       }')
+    #     echo "Selected core: $CORES"
+    #     echo "CORES=$CORES" >> $GITHUB_ENV
+
+    #     ZE_AFFINITY_MASK=1
+    #     echo "ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK" >> $GITHUB_ENV
+
+    - name: Run UMF benchmarks
       id: benchmarks
       working-directory: ${{env.BUILD_DIR}}
       run: >
-        taskset -c ${{ env.CORES }} ${{ github.workspace }}/ur-repo/scripts/benchmarks/main.py
+        ${{ github.workspace }}/sycl-repo/unified-runtime/scripts/benchmarks/main.py
         ~/bench_workdir_umf
         --umf ${{env.BUILD_DIR}}
+        --compare baseline
         ${{ inputs.upload_report && '--output-html' || '' }}
+        ${{ inputs.pr_no != 0 && '--output-markdown' || '' }}
         ${{ inputs.bench_script_params }}
 
+    # In case it failed to add a comment, we can still print the results.
     - name: Print benchmark results
-      if: ${{ always() }}
+      if: ${{ always() && inputs.pr_no != 0 }}
       run: cat ${{env.BUILD_DIR}}/benchmark_results.md
 
     - name: Add comment to PR

.github/workflows/reusable_compatibility.yml

@@ -97,7 +97,6 @@ jobs:
         -DUMF_FORMAT_CODE_STYLE=OFF
         -DUMF_DEVELOPER_MODE=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
-        -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
         -DUMF_TESTS_FAIL_ON_SKIP=ON
 
     - name: Build latest UMF
@@ -197,7 +196,6 @@ jobs:
         -DUMF_FORMAT_CODE_STYLE=OFF
         -DUMF_DEVELOPER_MODE=ON
         -DUMF_BUILD_LIBUMF_POOL_JEMALLOC=ON
-        -DUMF_BUILD_LIBUMF_POOL_DISJOINT=ON
         -DUMF_TESTS_FAIL_ON_SKIP=ON
 
     - name: Build latest UMF

CMakeLists.txt

@@ -186,11 +186,12 @@ else()
         # --disable-initial-exec-tls - Disable the initial-exec TLS model for
         # jemalloc's internal thread-local storage (on those platforms that
         # support explicit settings). This can allow jemalloc to be dynamically
-        # loaded after program startup (e.g. using dlopen).
+        # loaded after program startup (e.g. using dlopen). --disable-doc -
+        # Disable building and installing the documentation.
         COMMAND
             ./configure --prefix=${jemalloc_targ_BINARY_DIR}
             --with-jemalloc-prefix=je_ --disable-cxx --disable-initial-exec-tls
-            CFLAGS=-fPIC
+            --disable-doc CFLAGS=-fPIC
         WORKING_DIRECTORY ${jemalloc_targ_SOURCE_DIR}
         OUTPUT ${jemalloc_targ_SOURCE_DIR}/Makefile
         DEPENDS ${jemalloc_targ_SOURCE_DIR}/configure)
@@ -282,7 +283,8 @@ else()
 
         message(STATUS "hwloc CMAKE_GENERATOR: ${CMAKE_GENERATOR}")
 
-        if(CMAKE_GENERATOR STREQUAL "Ninja")
+        if(CMAKE_GENERATOR STREQUAL "Ninja" OR CMAKE_GENERATOR STREQUAL
+                                               "Unix Makefiles")
             add_custom_command(
                 COMMAND ${CMAKE_COMMAND}
                         -DCMAKE_INSTALL_PREFIX=${hwloc_targ_BINARY_DIR} -B build
@@ -407,7 +409,7 @@ if(UMF_BUILD_LEVEL_ZERO_PROVIDER AND (NOT UMF_LEVEL_ZERO_INCLUDE_DIR))
     include(FetchContent)
 
     set(LEVEL_ZERO_LOADER_REPO "https://github.com/oneapi-src/level-zero.git")
-    set(LEVEL_ZERO_LOADER_TAG v1.19.2)
+    set(LEVEL_ZERO_LOADER_TAG v1.20.2)
 
     message(
         STATUS

benchmark/benchmark.cpp

@@ -35,6 +35,10 @@ static void multithreaded(benchmark::internal::Benchmark *benchmark) {
     benchmark->Threads(1);
 }
 
+static void singlethreaded(benchmark::internal::Benchmark *benchmark) {
+    benchmark->Threads(1);
+}
+
 static void
 default_multiple_alloc_fix_size(benchmark::internal::Benchmark *benchmark) {
     benchmark->Args({10000, 1, 4096});
@@ -68,15 +72,17 @@ UMF_BENCHMARK_TEMPLATE_DEFINE(multiple_malloc_free_benchmark, proxy_pool,
 UMF_BENCHMARK_REGISTER_F(multiple_malloc_free_benchmark, proxy_pool)
     ->Apply(&default_multiple_alloc_fix_size)
     // reduce iterations, as this benchmark is slower than others
-    ->Iterations(50000);
+    ->Iterations(50000)
+    ->Apply(&singlethreaded);
 
 UMF_BENCHMARK_TEMPLATE_DEFINE(multiple_malloc_free_benchmark, os_provider,
                               fixed_alloc_size,
                               provider_allocator<os_provider>);
 UMF_BENCHMARK_REGISTER_F(multiple_malloc_free_benchmark, os_provider)
     ->Apply(&default_multiple_alloc_fix_size)
     // reduce iterations, as this benchmark is slower than others
-    ->Iterations(50000);
+    ->Iterations(50000)
+    ->Apply(&singlethreaded);
 
 UMF_BENCHMARK_TEMPLATE_DEFINE(multiple_malloc_free_benchmark, disjoint_pool_fix,
                               fixed_alloc_size,
@@ -89,7 +95,10 @@ UMF_BENCHMARK_TEMPLATE_DEFINE(multiple_malloc_free_benchmark,
                               disjoint_pool_uniform, uniform_alloc_size,
                               pool_allocator<disjoint_pool<os_provider>>);
 UMF_BENCHMARK_REGISTER_F(multiple_malloc_free_benchmark, disjoint_pool_uniform)
-    ->Apply(&default_multiple_alloc_uniform_size);
+    ->Apply(&default_multiple_alloc_uniform_size)
+    ->Apply(&singlethreaded);
+// TODO: change to multithreaded
+//->Apply(&multithreaded);
 
 #ifdef UMF_POOL_JEMALLOC_ENABLED
 UMF_BENCHMARK_TEMPLATE_DEFINE(multiple_malloc_free_benchmark, jemalloc_pool_fix,

benchmark/benchmark_umf.hpp

@@ -199,7 +199,7 @@ struct disjoint_pool : public pool_interface<Provider> {
             return {nullptr, [](void *) {}};
         }
 
-        ret = umfDisjointPoolParamsSetMinBucketSize(raw_params, 4096);
+        ret = umfDisjointPoolParamsSetMinBucketSize(raw_params, 8);
         if (ret != UMF_RESULT_SUCCESS) {
             state.SkipWithError("Failed to set min bucket size");
             return {nullptr, [](void *) {}};

examples/ipc_level_zero/CMakeLists.txt

@@ -24,7 +24,7 @@ endif()
 include(FetchContent)
 
 set(LEVEL_ZERO_LOADER_REPO "https://github.com/oneapi-src/level-zero.git")
-set(LEVEL_ZERO_LOADER_TAG v1.19.2)
+set(LEVEL_ZERO_LOADER_TAG v1.20.2)
 
 message(
     STATUS

examples/level_zero_shared_memory/CMakeLists.txt

@@ -24,7 +24,7 @@ endif()
 include(FetchContent)
 
 set(LEVEL_ZERO_LOADER_REPO "https://github.com/oneapi-src/level-zero.git")
-set(LEVEL_ZERO_LOADER_TAG v1.19.2)
+set(LEVEL_ZERO_LOADER_TAG v1.20.2)
 
 message(
     STATUS

src/coarse/coarse.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2024 Intel Corporation
+ * Copyright (C) 2024-2025 Intel Corporation
  *
  * Under the Apache License v2.0 with LLVM Exceptions. See LICENSE.TXT.
  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
@@ -278,7 +278,7 @@ static block_t *node_list_rm_first(ravl_free_blocks_head_t *head_node,
     assert(node->prev == NULL);
     struct block_t *block = node->block;
 
-    if (IS_NOT_ALIGNED(block->size, alignment)) {
+    if (IS_NOT_ALIGNED(((uintptr_t)block->data), alignment)) {
         return NULL;
     }
 
@@ -303,7 +303,7 @@ static block_t *node_list_rm_with_alignment(ravl_free_blocks_head_t *head_node,
 
     ravl_free_blocks_elem_t *node;
     for (node = head_node->head; node != NULL; node = node->next) {
-        if (IS_ALIGNED(node->block->size, alignment)) {
+        if (IS_ALIGNED(((uintptr_t)node->block->data), alignment)) {
             return node_list_rm(head_node, node);
         }
     }
@@ -1170,10 +1170,13 @@ umf_result_t coarse_free(coarse_t *coarse, void *ptr, size_t bytes) {
     }
 
     block_t *block = get_node_block(node);
-    assert(block->used);
+    if (!block->used) {
+        LOG_ERR("double free");
+        utils_mutex_unlock(&coarse->lock);
+        return UMF_RESULT_ERROR_INVALID_ARGUMENT;
+    }
 
     if (bytes > 0 && bytes != block->size) {
-        // wrong size of allocation
         LOG_ERR("wrong size of allocation");
         utils_mutex_unlock(&coarse->lock);
         return UMF_RESULT_ERROR_INVALID_ARGUMENT;

src/ipc_cache.c

@@ -144,6 +144,8 @@ umfIpcOpenedCacheCreate(ipc_opened_cache_eviction_cb_t eviction_cb) {
 
 void umfIpcOpenedCacheDestroy(ipc_opened_cache_handle_t cache) {
     ipc_opened_cache_entry_t *entry, *tmp;
+
+    utils_mutex_lock(&(cache->global->cache_lock));
     HASH_ITER(hh, cache->hash_table, entry, tmp) {
         DL_DELETE(cache->global->lru_list, entry);
         HASH_DEL(cache->hash_table, entry);
@@ -153,6 +155,7 @@ void umfIpcOpenedCacheDestroy(ipc_opened_cache_handle_t cache) {
         umf_ba_free(cache->global->cache_allocator, entry);
     }
     HASH_CLEAR(hh, cache->hash_table);
+    utils_mutex_unlock(&(cache->global->cache_lock));
 
     umf_ba_global_free(cache);
 }