[MTAI-484] feat(build): support cpp test for MUSA (PaddlePaddle#53)

cmake/generic.cmake

@@ -425,6 +425,9 @@ function(cc_binary TARGET_NAME)
   if(WITH_ROCM)
     target_link_libraries(${TARGET_NAME} ${ROCM_HIPRTC_LIB})
   endif()
+  if(WITH_MUSA)
+    target_link_libraries(${TARGET_NAME} ${MUSARTC_LIB})
+  endif()
 
   check_coverage_opt(${TARGET_NAME} ${cc_binary_SRCS})
 
@@ -452,6 +455,9 @@ function(cc_test_build TARGET_NAME)
     if(WITH_ROCM)
       target_link_libraries(${TARGET_NAME} ${ROCM_HIPRTC_LIB})
     endif()
+    if(WITH_MUSA)
+      target_link_libraries(${TARGET_NAME} ${MUSARTC_LIB})
+    endif()
     check_coverage_opt(${TARGET_NAME} ${cc_test_SRCS})
   endif()
 endfunction()
@@ -840,17 +846,43 @@ function(musa_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(musa_test "${options}" "${oneValueArgs}"
                           "${multiValueArgs}" ${ARGN})
-    add_executable(${TARGET_NAME} ${musa_test_SRCS})
+    musa_add_executable(${TARGET_NAME} ${musa_test_SRCS})
+    # "-pthread -ldl -lrt" is defined in CMAKE_CXX_LINK_EXECUTABLE
+    target_link_options(${TARGET_NAME} PRIVATE -pthread -ldl -lrt)
     get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
-    target_link_libraries(${TARGET_NAME} ${musa_test_DEPS}
-                          ${os_dependency_modules} paddle_gtest_main phi)
-    add_dependencies(${TARGET_NAME} ${musa_test_DEPS} paddle_gtest_main)
+    target_link_libraries(
+      ${TARGET_NAME}
+      ${musa_test_DEPS}
+      paddle_gtest_main
+      lod_tensor
+      memory
+      gtest
+      glog
+      phi
+      ${os_dependency_modules})
+    add_dependencies(
+      ${TARGET_NAME}
+      ${musa_test_DEPS}
+      paddle_gtest_main
+      lod_tensor
+      memory
+      gtest
+      phi
+      glog)
     common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT
                                               FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT
                                               FLAGS_init_allocated_mem=true)
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT
+                                              FLAGS_cudnn_deterministic=true)
+    set_property(
+      TEST ${TARGET_NAME}
+      PROPERTY
+        ENVIRONMENT
+        "LD_LIBRARY_PATH=${CMAKE_BINARY_DIR}/python/paddle/libs:$LD_LIBRARY_PATH"
+    )
   endif()
 endfunction()
 

cmake/musa.cmake

@@ -117,3 +117,7 @@ if(CMAKE_BUILD_TYPE MATCHES Debug)
   list(APPEND MUSA_MCC_FLAGS -g2)
   list(APPEND MUSA_MCC_FLAGS -O0)
 endif()
+
+set(musa_runtime_library_name musart)
+find_library(MUSARTC_LIB ${musa_runtime_library_name} HINTS ${MUSA_PATH}/lib)
+message(STATUS "MUSARTC_LIB: ${MUSARTC_LIB}")

paddle/fluid/memory/allocation/CMakeLists.txt

@@ -91,6 +91,10 @@ if(WITH_ROCM)
     SRCS thread_local_allocator_test.cc
     DEPS allocator)
 endif()
+if(WITH_MUSA)
+  musa_test(thread_local_allocator_test SRCS thread_local_allocator_test.cc
+            DEPS allocator)
+endif()
 
 if(WITH_GPU)
   nv_test(
@@ -102,6 +106,15 @@ elseif(WITH_ROCM)
     best_fit_allocator_test
     SRCS best_fit_allocator_test.cc best_fit_allocator_test.cu
     DEPS allocator memcpy)
+elseif(WITH_MUSA)
+  musa_test(
+    best_fit_allocator_test
+    SRCS
+    best_fit_allocator_test.cc
+    best_fit_allocator_test.cu
+    DEPS
+    allocator
+    memcpy)
 else()
   cc_test_old(best_fit_allocator_test SRCS best_fit_allocator_test.cc DEPS
               allocator)

paddle/fluid/memory/allocation/buddy_allocator_test.cc

@@ -246,6 +246,9 @@ TEST(BuddyAllocator, AllocFromAvailable) {
 #ifdef PADDLE_WITH_HIP
   hipError_t result = hipMalloc(&p, available >> 1);
   EXPECT_TRUE(result == hipSuccess);
+#elif defined(PADDLE_WITH_MUSA)
+  musaError_t result = musaMalloc(&p, available >> 1);
+  EXPECT_TRUE(result == musaSuccess);
 #else
   cudaError_t result = cudaMalloc(&p, available >> 1);
   EXPECT_TRUE(result == cudaSuccess);
@@ -265,6 +268,8 @@ TEST(BuddyAllocator, AllocFromAvailable) {
   if (p) {
 #ifdef PADDLE_WITH_HIP
     EXPECT_TRUE(hipFree(p) == hipSuccess);
+#elif defined(PADDLE_WITH_MUSA)
+    EXPECT_TRUE(musaFree(p) == musaSuccess);
 #else
     EXPECT_TRUE(cudaFree(p) == cudaSuccess);
 #endif
@@ -280,6 +285,8 @@ TEST(BuddyAllocator, AllocFromAvailableWhenFractionIsOne) {
 
 #ifdef PADDLE_WITH_HIP
   EXPECT_TRUE(hipMalloc(&p, static_cast<size_t>(1) << 30) == hipSuccess);
+#elif defined(PADDLE_WITH_MUSA)
+  EXPECT_TRUE(musaMalloc(&p, static_cast<size_t>(1) << 30) == musaSuccess);
 #else
   EXPECT_TRUE(cudaMalloc(&p, static_cast<size_t>(1) << 30) == cudaSuccess);
 #endif
@@ -296,6 +303,8 @@ TEST(BuddyAllocator, AllocFromAvailableWhenFractionIsOne) {
   if (p) {
 #ifdef PADDLE_WITH_HIP
     EXPECT_TRUE(hipFree(p) == hipSuccess);
+#elif defined(PADDLE_WITH_MUSA)
+    EXPECT_TRUE(musaFree(p) == musaSuccess);
 #else
     EXPECT_TRUE(cudaFree(p) == cudaSuccess);
 #endif

paddle/fluid/platform/enforce_test.cc

@@ -396,6 +396,54 @@ TEST(enforce, hip_success) {
   EXPECT_TRUE(CheckCudaStatusFailure(ncclSystemError, "Rccl error"));
 #endif
 }
+#elif defined(PADDLE_WITH_MUSA)
+TEST(enforce, musa_success) {
+  EXPECT_TRUE(CheckCudaStatusSuccess(musaSuccess));
+  EXPECT_TRUE(CheckCudaStatusFailure(musaErrorInvalidValue, "MUSA error"));
+
+  EXPECT_TRUE(CheckCudaStatusFailure(musaErrorMemoryAllocation, "MUSA error"));
+
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      musaErrorInsufficientDriver,
+      "This indicates that the installed MooreThreads MUSA driver is older "
+      "than the "
+      "MUSA runtime library. This is not a supported configuration.Users "
+      "should install an updated MooreThreads display driver to allow the "
+      "application to run"));
+  EXPECT_TRUE(CheckCudaStatusFailure(
+      musaErrorContextIsDestroyed,
+      "This error indicates that the context current to the calling thread has "
+      "been destroyed using muCtxDestroy, or is a primary context which has "
+      "not yet been initialized"));
+
+  EXPECT_TRUE(CheckCudaStatusSuccess(MURAND_STATUS_SUCCESS));
+  EXPECT_TRUE(
+      CheckCudaStatusFailure(MURAND_STATUS_VERSION_MISMATCH, "MURAND error"));
+  EXPECT_TRUE(
+      CheckCudaStatusFailure(MURAND_STATUS_NOT_CREATED, "MURAND error"));
+  EXPECT_TRUE(
+      CheckCudaStatusFailure(MURAND_STATUS_LENGTH_NOT_MULTIPLE,
+                             "Length requested is not a multple of dimension"));
+
+  EXPECT_TRUE(CheckCudaStatusSuccess(MUBLAS_STATUS_SUCCESS));
+  EXPECT_TRUE(
+      CheckCudaStatusFailure(MUBLAS_STATUS_NOT_IMPLEMENTED, "MUBLAS error"));
+  EXPECT_TRUE(
+      CheckCudaStatusFailure(MUBLAS_STATUS_INVALID_VALUE, "MUBLAS error"));
+
+#if !defined(__APPLE__) && defined(PADDLE_WITH_MCCL)
+  EXPECT_TRUE(CheckCudaStatusSuccess(mcclSuccess));
+  EXPECT_TRUE(CheckCudaStatusFailure(mcclUnhandledMusaError, "MCCL error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(mcclSystemError, "MCCL error"));
+  EXPECT_TRUE(CheckCudaStatusFailure(mcclInternalError,
+                                     "An internal check failed. This is either "
+                                     "a bug in MCCL or due to memory "
+                                     "corruption"));
+  EXPECT_TRUE(CheckCudaStatusFailure(mcclInvalidUsage,
+                                     "The call to MCCL is incorrect. This is "
+                                     "usually reflecting a programming error"));
+#endif
+}
 #else
 TEST(enforce, cuda_success) {
   EXPECT_TRUE(CheckCudaStatusSuccess(cudaSuccess));

paddle/phi/backends/dynload/dynamic_loader.cc

@@ -468,6 +468,8 @@ void* GetNVRTCDsoHandle() {
   return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libnvrtc.dylib", false);
 #elif defined(PADDLE_WITH_HIP)
   return GetDsoHandleFromSearchPath(FLAGS_rocm_dir, "libamdhip64.so", false);
+#elif defined(PADDLE_WITH_MUSA)
+  return GetDsoHandleFromSearchPath(FLAGS_musa_dir, "libmusart.so", false);
 #else
   return GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libnvrtc.so", false);
 #endif

test/CMakeLists.txt

@@ -189,6 +189,9 @@ if(${len} GREATER_EQUAL 1)
       if(WITH_ROCM)
         target_link_libraries(${test_name} ${ROCM_HIPRTC_LIB})
       endif()
+      if(WITH_MUSA)
+        target_link_libraries(${test_name} ${MUSARTC_LIB})
+      endif()
       if(APPLE)
         target_link_libraries(
           ${test_name}