Merge branch 'main' into yc/0829-buffer-alignment-ub

Author: yingcong-wu

scripts/benchmarks/benches/compute.py

@@ -82,7 +82,7 @@ def run(self, env_vars) -> Result:
 
         result = self.run_bench(command, env_vars)
         (label, mean) = self.parse_output(result)
-        return Result(label=label, value=mean, command=command, env=env_vars, stdout=result)
+        return Result(label=label, value=mean, command=command, env=env_vars, stdout=result, lower_is_better=self.lower_is_better())
 
     def parse_output(self, output):
         csv_file = io.StringIO(output)

scripts/benchmarks/benches/velocity.py

@@ -61,7 +61,7 @@ def run(self, env_vars) -> Result:
 
         result = self.run_bench(command, env_vars)
 
-        return Result(label=self.bench_name, value=self.parse_output(result), command=command, env=env_vars, stdout=result)
+        return Result(label=self.bench_name, value=self.parse_output(result), command=command, env=env_vars, stdout=result, lower_is_better=self.lower_is_better())
 
     def teardown(self):
         return

scripts/benchmarks/main.py

@@ -52,34 +52,46 @@ def main(directory, additional_env_vars, save_name, compare_names, filter):
         benchmarks = [benchmark for benchmark in benchmarks if filter.search(benchmark.name())]
 
     for benchmark in benchmarks:
-        print(f"setting up {benchmark.name()}... ", end='', flush=True)
-        benchmark.setup()
-        print("complete.")
+        try:
+            print(f"setting up {benchmark.name()}... ", end='', flush=True)
+            benchmark.setup()
+            print("complete.")
+        except Exception as e:
+            if options.exit_on_failure:
+                raise e
+            else:
+                print(f"failed: {e}")
 
     results = []
     for benchmark in benchmarks:
-        merged_env_vars = {**additional_env_vars}
-        iteration_results = []
-        for iter in range(options.iterations):
-            print(f"running {benchmark.name()}, iteration {iter}... ", end='', flush=True)
-            bench_results = benchmark.run(merged_env_vars)
-            if bench_results is not None:
-                print(f"complete ({bench_results.value} {benchmark.unit()}).")
-                iteration_results.append(bench_results)
+        try:
+            merged_env_vars = {**additional_env_vars}
+            iteration_results = []
+            for iter in range(options.iterations):
+                print(f"running {benchmark.name()}, iteration {iter}... ", end='', flush=True)
+                bench_results = benchmark.run(merged_env_vars)
+                if bench_results is not None:
+                    print(f"complete ({bench_results.value} {benchmark.unit()}).")
+                    iteration_results.append(bench_results)
+                else:
+                    print(f"did not finish.")
+
+            if len(iteration_results) == 0:
+                continue
+
+            iteration_results.sort(key=lambda res: res.value)
+            median_index = len(iteration_results) // 2
+            median_result = iteration_results[median_index]
+
+            median_result.unit = benchmark.unit()
+            median_result.name = benchmark.name()
+
+            results.append(median_result)
+        except Exception as e:
+            if options.exit_on_failure:
+                raise e
             else:
-                print(f"did not finish.")
-
-        if len(iteration_results) == 0:
-            continue
-
-        iteration_results.sort(key=lambda res: res.value)
-        median_index = len(iteration_results) // 2
-        median_result = iteration_results[median_index]
-
-        median_result.unit = benchmark.unit()
-        median_result.name = benchmark.name()
-
-        results.append(median_result)
+                print(f"failed: {e}")
 
     for benchmark in benchmarks:
         print(f"tearing down {benchmark.name()}... ", end='', flush=True)
@@ -126,6 +138,7 @@ def validate_and_parse_env_args(env_args):
     parser.add_argument("--timeout", type=int, help='Timeout for individual benchmarks in seconds.', default=600)
     parser.add_argument("--filter", type=str, help='Regex pattern to filter benchmarks by name.', default=None)
     parser.add_argument("--verbose", help='Print output of all the commands.', action="store_true")
+    parser.add_argument("--exit_on_failure", help='Exit on first failure.', action="store_true")
 
     args = parser.parse_args()
     additional_env_vars = validate_and_parse_env_args(args.env)
@@ -137,6 +150,7 @@ def validate_and_parse_env_args(env_args):
     options.timeout = args.timeout
     options.ur_dir = args.ur_dir
     options.ur_adapter_name = args.ur_adapter_name
+    options.exit_on_failure = args.exit_on_failure
 
     benchmark_filter = re.compile(args.filter) if args.filter else None
 

scripts/benchmarks/output.py

@@ -116,7 +116,7 @@ def generate_summary_table(chart_data: dict[str, list[Result]]):
             if key in results:
                 value = results[key].value
                 if key == best_key:
-                    row += f" `**{value}**` |"  # Highlight the best value
+                    row += f" <ins>{value}</ins> |"  # Highlight the best value
                 else:
                     row += f" {value} |"
             else:
@@ -132,6 +132,7 @@ def generate_markdown(chart_data: dict[str, list[Result]]):
 
     return f"""
 # Summary
+<ins>result</ins> is better\n
 {summary_table}
 # Charts
 {mermaid_script}

source/adapters/level_zero/v2/common.hpp

@@ -85,70 +85,6 @@ struct ze_handle_wrapper {
   bool ownZeHandle;
 };
 
-template <typename URHandle, ur_result_t (*retain)(URHandle),
-          ur_result_t (*release)(URHandle)>
-struct ur_shared_handle {
-  using handle_t = URHandle;
-
-  ur_shared_handle() : handle(nullptr) {}
-  explicit ur_shared_handle(handle_t handle) : handle(handle) {}
-  ~ur_shared_handle() {
-    try {
-      reset();
-    } catch (...) {
-    }
-  }
-
-  ur_shared_handle(const ur_shared_handle &other) : handle(other.handle) {
-    retain(handle);
-  }
-  ur_shared_handle(ur_shared_handle &&other) : handle(other.handle) {
-    other.handle = nullptr;
-  }
-  ur_shared_handle(std::nullptr_t) : handle(nullptr) {}
-
-  void reset() {
-    if (!handle) {
-      return;
-    }
-
-    UR_CALL_THROWS(release(handle));
-    handle = nullptr;
-  }
-
-  ur_shared_handle &operator=(const ur_shared_handle &other) {
-    if (handle) {
-      release(handle);
-    }
-    handle = other.handle;
-    retain(handle);
-    return *this;
-  }
-  ur_shared_handle &operator=(ur_shared_handle &&other) {
-    if (handle) {
-      release(handle);
-    }
-    handle = other.handle;
-    other.handle = nullptr;
-    return *this;
-  }
-  ur_shared_handle &operator=(std::nullptr_t) {
-    if (handle) {
-      release(handle);
-    }
-    new (this) ur_shared_handle(nullptr);
-    return *this;
-  }
-
-  handle_t *ptr() { return &handle; }
-  handle_t get() const { return handle; }
-  handle_t operator->() { return handle; }
-  operator handle_t() { return handle; }
-
-private:
-  handle_t handle;
-};
-
 using ze_kernel_handle_t =
     ze_handle_wrapper<::ze_kernel_handle_t, zeKernelDestroy>;
 
@@ -158,11 +94,5 @@ using ze_event_handle_t =
 using ze_event_pool_handle_t =
     ze_handle_wrapper<::ze_event_pool_handle_t, zeEventPoolDestroy>;
 
-using ur_queue_shared_handle_t =
-    ur_shared_handle<ur_queue_handle_t, urQueueRetain, urQueueRelease>;
-
-using ur_kernel_shared_handle_t =
-    ur_shared_handle<ur_kernel_handle_t, urKernelRetain, urKernelRelease>;
-
 } // namespace raii
 } // namespace v2

source/adapters/level_zero/v2/kernel.cpp

@@ -32,10 +32,12 @@ ur_result_t ur_single_device_kernel_t::release() {
   return UR_RESULT_SUCCESS;
 }
 
-ur_kernel_handle_t_::ur_kernel_handle_t_(ur_program_shared_handle_t hProgram,
+ur_kernel_handle_t_::ur_kernel_handle_t_(ur_program_handle_t hProgram,
                                          const char *kernelName)
     : hProgram(hProgram),
       deviceKernels(hProgram->Context->getPlatform()->getNumDevices()) {
+  urProgramRetain(hProgram);
+
   for (auto [zeDevice, zeModule] : hProgram->ZeModuleMap) {
     ZeStruct<ze_kernel_desc_t> zeKernelDesc;
     zeKernelDesc.pKernelName = kernelName;
@@ -78,7 +80,8 @@ ur_result_t ur_kernel_handle_t_::release() {
       singleDeviceKernelOpt.value().hKernel.reset();
     }
   }
-  hProgram.reset();
+
+  UR_CALL_THROWS(urProgramRelease(hProgram));
 
   return UR_RESULT_SUCCESS;
 }
@@ -190,14 +193,13 @@ ur_result_t ur_kernel_handle_t_::setArgPointer(
 }
 
 ur_program_handle_t ur_kernel_handle_t_::getProgramHandle() const {
-  return hProgram.get();
+  return hProgram;
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL
 urKernelCreate(ur_program_handle_t hProgram, const char *pKernelName,
                ur_kernel_handle_t *phKernel) {
-  *phKernel = new ur_kernel_handle_t_(
-      ur_kernel_handle_t_::ur_program_shared_handle_t(hProgram), pKernelName);
+  *phKernel = new ur_kernel_handle_t_(hProgram, pKernelName);
   return UR_RESULT_SUCCESS;
 }
 

source/adapters/level_zero/v2/kernel.hpp

@@ -26,21 +26,8 @@ struct ur_single_device_kernel_t {
 
 struct ur_kernel_handle_t_ : _ur_object {
 private:
-  static inline ur_result_t
-  internalProgramRelease(ur_program_handle_t hProgram) {
-    // do a release on the program this kernel was part of without delete of the
-    // program handle.
-    hProgram->ur_release_program_resources(false);
-    return UR_RESULT_SUCCESS;
-  }
-
 public:
-  using ur_program_shared_handle_t =
-      v2::raii::ur_shared_handle<ur_program_handle_t, urProgramRetain,
-                                 internalProgramRelease>;
-
-  ur_kernel_handle_t_(ur_program_shared_handle_t hProgram,
-                      const char *kernelName);
+  ur_kernel_handle_t_(ur_program_handle_t hProgram, const char *kernelName);
 
   // From native handle
   ur_kernel_handle_t_(ur_native_handle_t hNativeKernel,
@@ -75,7 +62,7 @@ struct ur_kernel_handle_t_ : _ur_object {
 
 private:
   // Keep the program of the kernel.
-  ur_program_shared_handle_t hProgram;
+  ur_program_handle_t hProgram;
 
   // Vector of ur_single_device_kernel_t indexed by device->Id
   std::vector<std::optional<ur_single_device_kernel_t>> deviceKernels;

test/conformance/kernel/kernel_adapter_native_cpu.match

@@ -104,6 +104,7 @@ urKernelGetSubGroupInfoTest.InvalidEnumeration/SYCL_NATIVE_CPU___SYCL_Native_CPU
 urKernelGetSubGroupInfoTest.InvalidEnumeration/SYCL_NATIVE_CPU___SYCL_Native_CPU__{{.*}}__UR_KERNEL_SUB_GROUP_INFO_SUB_GROUP_SIZE_INTEL
 urKernelGetSubGroupInfoSingleTest.CompileNumSubgroupsIsZero/SYCL_NATIVE_CPU___SYCL_Native_CPU__{{.*}}
 urKernelReleaseTest.Success/SYCL_NATIVE_CPU___SYCL_Native_CPU__{{.*}}
+urKernelReleaseTest.KernelReleaseAfterProgramRelease/SYCL_NATIVE_CPU___SYCL_Native_CPU__{{.*}}
 urKernelReleaseTest.InvalidNullHandleKernel/SYCL_NATIVE_CPU___SYCL_Native_CPU__{{.*}}
 urKernelRetainTest.Success/SYCL_NATIVE_CPU___SYCL_Native_CPU__{{.*}}
 urKernelRetainTest.InvalidNullHandleKernel/SYCL_NATIVE_CPU___SYCL_Native_CPU__{{.*}}

test/conformance/kernel/urKernelRelease.cpp

@@ -13,6 +13,13 @@ TEST_P(urKernelReleaseTest, Success) {
     ASSERT_SUCCESS(urKernelRelease(kernel));
 }
 
+TEST_P(urKernelReleaseTest, KernelReleaseAfterProgramRelease) {
+    ASSERT_SUCCESS(urKernelRetain(kernel));
+    ASSERT_SUCCESS(urProgramRelease(program));
+    program = nullptr;
+    ASSERT_SUCCESS(urKernelRelease(kernel));
+}
+
 TEST_P(urKernelReleaseTest, InvalidNullHandleKernel) {
     ASSERT_EQ_RESULT(UR_RESULT_ERROR_INVALID_NULL_HANDLE,
                      urKernelRelease(nullptr));

third_party/deps.yml

@@ -25,7 +25,7 @@ dependencies:
   - libssh2=1.11.0
   - libstdcxx-ng=13.1.0
   - libuv=1.44.2
-  - libzlib=1.2.13
+  - libzlib=1.3.1
   - llvm-spirv=14.0.0
   - llvm-tools=14.0.6
   - llvmdev=14.0.6
@@ -35,5 +35,5 @@ dependencies:
   - rhash=1.4.3
   # don't upgrade xz utils due to CVE-2024-3094
   - xz=5.2.6
-  - zlib=1.2.13
+  - zlib=1.3.1
   - zstd=1.5.2