Merge branch 'main' into tensormap-exp-api

Author: Hugh Delaney

.github/workflows/multi_device.yml

@@ -63,4 +63,4 @@ jobs:
 
     - name: Test adapters
       working-directory: ${{github.workspace}}/build
-      run: env UR_CTS_ADAPTER_PLATFORM="${{matrix.adapter.platform}}" ctest -C ${{matrix.build_type}} --output-on-failure -L "conformance" -E "enqueue|kernel|integration|exp_command_buffer|exp_enqueue_native|exp_launch_properties|exp_usm_p2p" --timeout 180
+      run: env UR_CTS_ADAPTER_PLATFORM="${{matrix.adapter.platform}}" ctest -C ${{matrix.build_type}} --output-on-failure -L "conformance" -E "exp_command_buffer" --timeout 180

CMakeLists.txt

@@ -137,7 +137,7 @@ if(UR_ENABLE_TRACING)
 
     if (UR_BUILD_XPTI_LIBS)
         # fetch xpti proxy library for the tracing layer
-        FetchContentSparse_Declare(xpti https://github.com/intel/llvm.git "sycl-nightly/20230703" "xpti")
+        FetchContentSparse_Declare(xpti https://github.com/intel/llvm.git "nightly-2024-10-22" "xpti")
         FetchContent_MakeAvailable(xpti)
 
         # set -fPIC for xpti since we are linking it with a shared library
@@ -149,7 +149,7 @@ if(UR_ENABLE_TRACING)
         set(XPTI_DIR ${xpti_SOURCE_DIR})
         set(XPTI_ENABLE_TESTS OFF CACHE INTERNAL "Turn off xptifw tests")
 
-        FetchContentSparse_Declare(xptifw https://github.com/intel/llvm.git "sycl-nightly/20230703" "xptifw")
+        FetchContentSparse_Declare(xptifw https://github.com/intel/llvm.git "nightly-2024-10-22" "xptifw")
 
         FetchContent_MakeAvailable(xptifw)
 

cmake/helpers.cmake

@@ -58,7 +58,10 @@ macro(add_sanitizer_flag flag)
     set(CMAKE_REQUIRED_LIBRARIES ${SAVED_CMAKE_REQUIRED_LIBRARIES})
 endmacro()
 
-check_cxx_compiler_flag("-fcf-protection=full" CXX_HAS_FCF_PROTECTION_FULL)
+if(CMAKE_SYSTEM_NAME STREQUAL Linux)
+    check_cxx_compiler_flag("-fcf-protection=full" CXX_HAS_FCF_PROTECTION_FULL)
+    check_cxx_compiler_flag("-fstack-clash-protection" CXX_HAS_FSTACK_CLASH_PROTECTION)
+endif()
 
 function(add_ur_target_compile_options name)
     if(NOT MSVC)
@@ -81,9 +84,7 @@ function(add_ur_target_compile_options name)
             # -flto
             # $<$<CXX_COMPILER_ID:Clang,AppleClang>:-fsanitize=cfi>
             $<$<BOOL:${CXX_HAS_FCF_PROTECTION_FULL}>:-fcf-protection=full>
-            # -fstack-clash-protection is not supported in apple clang or GCC < 8
-            $<$<AND:$<CXX_COMPILER_ID:GNU>,$<VERSION_GREATER_EQUAL:$<CXX_COMPILER_VERSION>,8>>:-fstack-clash-protection>
-            $<$<CXX_COMPILER_ID:Clang>:-fstack-clash-protection>
+            $<$<BOOL:${CXX_HAS_FSTACK_CLASH_PROTECTION}>:-fstack-clash-protection>
 
             # Colored output
             $<$<CXX_COMPILER_ID:GNU>:-fdiagnostics-color=always>

examples/collector/collector.cpp

@@ -125,8 +125,7 @@ XPTI_CALLBACK_API void xptiTraceInit(unsigned int major_version,
         return;
     }
     if (std::string_view(stream_name) != UR_STREAM_NAME) {
-        std::cout << "Invalid stream name: " << stream_name << ". Expected "
-                  << UR_STREAM_NAME << ". Aborting." << std::endl;
+        // we expect ur.call, but this can also be xpti.framework.
         return;
     }
 

include/ur_api.h

@@ -1056,6 +1056,8 @@ typedef enum ur_adapter_backend_t {
 ///         + `NULL == phAdapters`
 ///     - ::UR_RESULT_ERROR_INVALID_SIZE
 ///         + `NumEntries == 0 && phPlatforms != NULL`
+///     - ::UR_RESULT_ERROR_INVALID_VALUE
+///         + `pNumPlatforms == NULL && phPlatforms == NULL`
 UR_APIEXPORT ur_result_t UR_APICALL
 urPlatformGet(
     ur_adapter_handle_t *phAdapters,   ///< [in][range(0, NumAdapters)] array of adapters to query for platforms.
@@ -1639,6 +1641,9 @@ typedef enum ur_device_info_t {
     UR_DEVICE_INFO_GLOBAL_VARIABLE_SUPPORT = 118,                    ///< [::ur_bool_t] return true if the device supports the
                                                                      ///< `EnqueueDeviceGlobalVariableWrite` and
                                                                      ///< `EnqueueDeviceGlobalVariableRead` entry points.
+    UR_DEVICE_INFO_USM_POOL_SUPPORT = 119,                           ///< [::ur_bool_t] return true if the device supports USM pooling. Pertains
+                                                                     ///< to the `USMPool` entry points and usage of the `pool` parameter of the
+                                                                     ///< USM alloc entry points.
     UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP = 0x1000,              ///< [::ur_bool_t] Returns true if the device supports the use of
                                                                      ///< command-buffers.
     UR_DEVICE_INFO_COMMAND_BUFFER_UPDATE_CAPABILITIES_EXP = 0x1001,  ///< [::ur_device_command_buffer_update_capability_flags_t] Command-buffer
@@ -3503,7 +3508,6 @@ typedef struct ur_usm_pool_limits_desc_t {
 /// @brief USM allocate host memory
 ///
 /// @details
-///     - This function must support memory pooling.
 ///     - If pUSMDesc is not NULL and pUSMDesc->pool is not NULL the allocation
 ///       will be served from a specified memory pool.
 ///     - Otherwise, the behavior is implementation-defined.
@@ -3536,6 +3540,8 @@ typedef struct ur_usm_pool_limits_desc_t {
 ///         + `size` is greater than ::UR_DEVICE_INFO_MAX_MEM_ALLOC_SIZE.
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
+///         + If any device associated with `hContext` reports `false` for ::UR_DEVICE_INFO_USM_POOL_SUPPORT
 UR_APIEXPORT ur_result_t UR_APICALL
 urUSMHostAlloc(
     ur_context_handle_t hContext,  ///< [in] handle of the context object
@@ -3549,7 +3555,6 @@ urUSMHostAlloc(
 /// @brief USM allocate device memory
 ///
 /// @details
-///     - This function must support memory pooling.
 ///     - If pUSMDesc is not NULL and pUSMDesc->pool is not NULL the allocation
 ///       will be served from a specified memory pool.
 ///     - Otherwise, the behavior is implementation-defined.
@@ -3583,6 +3588,8 @@ urUSMHostAlloc(
 ///         + `size` is greater than ::UR_DEVICE_INFO_MAX_MEM_ALLOC_SIZE.
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
+///         + If any device associated with `hContext` reports `false` for ::UR_DEVICE_INFO_USM_POOL_SUPPORT
 UR_APIEXPORT ur_result_t UR_APICALL
 urUSMDeviceAlloc(
     ur_context_handle_t hContext,  ///< [in] handle of the context object
@@ -3597,7 +3604,6 @@ urUSMDeviceAlloc(
 /// @brief USM allocate shared memory
 ///
 /// @details
-///     - This function must support memory pooling.
 ///     - If pUSMDesc is not NULL and pUSMDesc->pool is not NULL the allocation
 ///       will be served from a specified memory pool.
 ///     - Otherwise, the behavior is implementation-defined.
@@ -3632,6 +3638,8 @@ urUSMDeviceAlloc(
 ///         + If `UR_DEVICE_INFO_USM_SINGLE_SHARED_SUPPORT` and `UR_DEVICE_INFO_USM_CROSS_SHARED_SUPPORT` are both false.
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
+///         + If any device associated with `hContext` reports `false` for ::UR_DEVICE_INFO_USM_POOL_SUPPORT
 UR_APIEXPORT ur_result_t UR_APICALL
 urUSMSharedAlloc(
     ur_context_handle_t hContext,  ///< [in] handle of the context object
@@ -3713,6 +3721,8 @@ urUSMGetMemAllocInfo(
 ///         + `::UR_USM_POOL_FLAGS_MASK & pPoolDesc->flags`
 ///     - ::UR_RESULT_ERROR_INVALID_VALUE
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
+///         + If any device associated with `hContext` reports `false` for ::UR_DEVICE_INFO_USM_POOL_SUPPORT
 UR_APIEXPORT ur_result_t UR_APICALL
 urUSMPoolCreate(
     ur_context_handle_t hContext,  ///< [in] handle of the context object
@@ -3731,6 +3741,7 @@ urUSMPoolCreate(
 ///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
 ///         + `NULL == pPool`
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
 UR_APIEXPORT ur_result_t UR_APICALL
 urUSMPoolRetain(
     ur_usm_pool_handle_t pPool ///< [in][retain] pointer to USM memory pool
@@ -3753,6 +3764,7 @@ urUSMPoolRetain(
 ///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
 ///         + `NULL == pPool`
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
 UR_APIEXPORT ur_result_t UR_APICALL
 urUSMPoolRelease(
     ur_usm_pool_handle_t pPool ///< [in][release] pointer to USM memory pool
@@ -3794,6 +3806,7 @@ typedef enum ur_usm_pool_info_t {
 ///         + `pPropValue == NULL && pPropSizeRet == NULL`
 ///     - ::UR_RESULT_ERROR_OUT_OF_HOST_MEMORY
 ///     - ::UR_RESULT_ERROR_OUT_OF_RESOURCES
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
 UR_APIEXPORT ur_result_t UR_APICALL
 urUSMPoolGetInfo(
     ur_usm_pool_handle_t hPool,  ///< [in] handle of the USM memory pool
@@ -9819,9 +9832,9 @@ urUSMReleaseExp(
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Supported peer info
 typedef enum ur_exp_peer_info_t {
-    UR_EXP_PEER_INFO_UR_PEER_ACCESS_SUPPORTED = 0,  ///< [uint32_t] 1 if P2P access is supported otherwise P2P access is not
+    UR_EXP_PEER_INFO_UR_PEER_ACCESS_SUPPORTED = 0,  ///< [int] 1 if P2P access is supported otherwise P2P access is not
                                                     ///< supported.
-    UR_EXP_PEER_INFO_UR_PEER_ATOMICS_SUPPORTED = 1, ///< [uint32_t] 1 if atomic operations are supported over the P2P link,
+    UR_EXP_PEER_INFO_UR_PEER_ATOMICS_SUPPORTED = 1, ///< [int] 1 if atomic operations are supported over the P2P link,
                                                     ///< otherwise such operations are not supported.
     /// @cond
     UR_EXP_PEER_INFO_FORCE_UINT32 = 0x7fffffff

include/ur_print.hpp

@@ -2578,6 +2578,9 @@ inline std::ostream &operator<<(std::ostream &os, enum ur_device_info_t value) {
     case UR_DEVICE_INFO_GLOBAL_VARIABLE_SUPPORT:
         os << "UR_DEVICE_INFO_GLOBAL_VARIABLE_SUPPORT";
         break;
+    case UR_DEVICE_INFO_USM_POOL_SUPPORT:
+        os << "UR_DEVICE_INFO_USM_POOL_SUPPORT";
+        break;
     case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP:
         os << "UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP";
         break;
@@ -4080,6 +4083,18 @@ inline ur_result_t printTagged(std::ostream &os, const void *ptr, ur_device_info
 
         os << ")";
     } break;
+    case UR_DEVICE_INFO_USM_POOL_SUPPORT: {
+        const ur_bool_t *tptr = (const ur_bool_t *)ptr;
+        if (sizeof(ur_bool_t) > size) {
+            os << "invalid size (is: " << size << ", expected: >=" << sizeof(ur_bool_t) << ")";
+            return UR_RESULT_ERROR_INVALID_SIZE;
+        }
+        os << (const void *)(tptr) << " (";
+
+        os << *tptr;
+
+        os << ")";
+    } break;
     case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP: {
         const ur_bool_t *tptr = (const ur_bool_t *)ptr;
         if (sizeof(ur_bool_t) > size) {
@@ -10336,9 +10351,9 @@ inline ur_result_t printTagged(std::ostream &os, const void *ptr, ur_exp_peer_in
 
     switch (value) {
     case UR_EXP_PEER_INFO_UR_PEER_ACCESS_SUPPORTED: {
-        const uint32_t *tptr = (const uint32_t *)ptr;
-        if (sizeof(uint32_t) > size) {
-            os << "invalid size (is: " << size << ", expected: >=" << sizeof(uint32_t) << ")";
+        const int *tptr = (const int *)ptr;
+        if (sizeof(int) > size) {
+            os << "invalid size (is: " << size << ", expected: >=" << sizeof(int) << ")";
             return UR_RESULT_ERROR_INVALID_SIZE;
         }
         os << (const void *)(tptr) << " (";
@@ -10348,9 +10363,9 @@ inline ur_result_t printTagged(std::ostream &os, const void *ptr, ur_exp_peer_in
         os << ")";
     } break;
     case UR_EXP_PEER_INFO_UR_PEER_ATOMICS_SUPPORTED: {
-        const uint32_t *tptr = (const uint32_t *)ptr;
-        if (sizeof(uint32_t) > size) {
-            os << "invalid size (is: " << size << ", expected: >=" << sizeof(uint32_t) << ")";
+        const int *tptr = (const int *)ptr;
+        if (sizeof(int) > size) {
+            os << "invalid size (is: " << size << ", expected: >=" << sizeof(int) << ")";
             return UR_RESULT_ERROR_INVALID_SIZE;
         }
         os << (const void *)(tptr) << " (";

scripts/benchmarks/benches/compute.py

@@ -20,7 +20,7 @@ def setup(self):
         if self.built:
             return
 
-        repo_path = git_clone(self.directory, "compute-benchmarks-repo", "https://github.com/intel/compute-benchmarks.git", "08c41bb8bc1762ad53c6194df6d36bfcceff4aa2")
+        repo_path = git_clone(self.directory, "compute-benchmarks-repo", "https://github.com/intel/compute-benchmarks.git", "aa6a3b2108bb86202b654ad28129156fa746d41d")
         build_path = create_build_path(self.directory, 'compute-benchmarks-build')
 
         configure_command = [
@@ -231,3 +231,26 @@ def bin_args(self) -> list[str]:
             "--numberOfElementsY=256",
             "--numberOfElementsZ=256",
         ]
+
+class MemcpyExecute(ComputeBenchmark):
+    def __init__(self, bench, numOpsPerThread, numThreads, allocSize, iterations):
+        self.numOpsPerThread = numOpsPerThread
+        self.numThreads = numThreads
+        self.allocSize = allocSize
+        self.iterations = iterations
+        super().__init__(bench, "multithread_benchmark_ur", "MemcpyExecute")
+
+    def name(self):
+        return f"multithread_benchmark_ur MemcpyExecute opsPerThread:{self.numOpsPerThread}, numThreads:{self.numThreads}, allocSize:{self.allocSize}"
+
+    def bin_args(self) -> list[str]:
+        return [
+            "--Ioq=1",
+            "--UseEvents=1",
+            "--MeasureCompletion=1",
+            "--UseQueuePerThread=1",
+            f"--AllocSize={self.allocSize}",
+            f"--NumThreads={self.numThreads}",
+            f"--NumOpsPerThread={self.numOpsPerThread}",
+            f"--iterations={self.iterations}"
+        ]

scripts/benchmarks/main.py

@@ -44,6 +44,12 @@ def main(directory, additional_env_vars, save_name, compare_names, filter):
         ExecImmediateCopyQueue(cb, 0, 1, 'Device', 'Device', 1024),
         ExecImmediateCopyQueue(cb, 1, 1, 'Device', 'Host', 1024),
         VectorSum(cb),
+        MemcpyExecute(cb, 400, 8, 1024, 100),
+        MemcpyExecute(cb, 400, 8, 102400, 10),
+        MemcpyExecute(cb, 500, 8, 102400, 10),
+        MemcpyExecute(cb, 400, 1, 1024, 1000),
+        MemcpyExecute(cb, 10, 16, 1024, 1000),
+        MemcpyExecute(cb, 10, 16, 102400, 100),
 
         # *** Velocity benchmarks
         Hashtable(vb),

scripts/core/CONTRIB.rst

@@ -129,7 +129,7 @@ Adapter Change Process
 .. _UNIFIED_RUNTIME_REPO:
    https://github.com/intel/llvm/blob/sycl/sycl/cmake/modules/FetchUnifiedRuntime.cmake#L119
 .. _UNIFIED_RUNTIME_TAG:
-   https://github.com/intel/llvm/blob/sycl/sycl/cmake/modules/FetchUnifiedRuntime.cmake#L126
+   https://github.com/intel/llvm/blob/sycl/sycl/cmake/modules/UnifiedRuntimeTag.cmake
 
 Build Environment
 =================

scripts/core/EXP-USM-P2P.rst

@@ -69,6 +69,8 @@ Changelog
 +-----------+---------------------------------------------+
 | 1.1       | Added USM_P2P_EXTENSION_STRING_EXP ID Macro |
 +-----------+---------------------------------------------+
+| 1.2       | Switch Info types from uint32_t to int      |
++-----------+---------------------------------------------+
 
 Contributors
 --------------------------------------------------------------------------------