Integrate AOCL-BLAS as CPU reference BLAS for TheRock CI (blocked on AOCL team)

BUILD_TOPOLOGY.toml

@@ -373,6 +373,13 @@ type = "target-neutral"
 artifact_deps = []
 feature_group = "HOST_MATH"  # Controlled by THEROCK_ENABLE_HOST_MATH (OFF by default)
 
+[artifacts.host-aocl-blas]
+artifact_group = "third-party-libs"
+type = "target-neutral"
+artifact_deps = []  # No dependencies - uses system compiler like OpenBLAS
+feature_group = "HOST_MATH"  # Controlled by THEROCK_ENABLE_HOST_MATH (OFF by default)
+# AOCL-BLAS provides BLIS/BLAS/CBLAS functions for CPU reference BLAS in testing
+
 [artifacts.host-suite-sparse]
 artifact_group = "third-party-libs"
 type = "target-neutral"
@@ -494,7 +501,7 @@ feature_group = "CORE"  # Part of core, enabled by default
 [artifacts.blas]
 artifact_group = "math-libs"
 type = "target-specific"
-artifact_deps = ["core-runtime", "core-hip", "core-amdsmi", "host-blas", "host-suite-sparse", "rocprofiler-sdk"]
+artifact_deps = ["core-runtime", "core-hip", "core-amdsmi", "host-blas", "host-aocl-blas", "host-suite-sparse", "rocprofiler-sdk"]
 split_databases = ["rocblas", "hipblaslt"]
 
 [artifacts.fft]

CMakeLists.txt

@@ -89,6 +89,12 @@ if(CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME AND NOT DEFINED THEROCK_BUILD_TESTIN
 endif()
 if(THEROCK_BUILD_TESTING)
   message(STATUS "Enabling building tests")
+  # Enable host math libraries (OpenBLAS, AOCL-BLAS, SuiteSparse) when testing is enabled
+  # These provide CPU reference BLAS implementations for rocBLAS clients/tests
+  if(NOT DEFINED THEROCK_ENABLE_HOST_MATH)
+    set(THEROCK_ENABLE_HOST_MATH ON)
+    message(STATUS "Enabling host math libraries for testing")
+  endif()
 endif()
 
 ################################################################################

base/CMakeLists.txt

@@ -56,7 +56,6 @@ therock_cmake_subproject_provide_package(rocm-core rocm-core lib/cmake/rocm-core
 therock_cmake_subproject_activate(rocm-core)
 
 
-
 ################################################################################
 # rocm_smi_lib
 ################################################################################

build_tools/github_actions/test_executable_scripts/test_rocblas.py

@@ -26,6 +26,11 @@
 
 logging.basicConfig(level=logging.INFO)
 
+# Cap OMP_NUM_THREADS for AOCL-BLAS in CI to avoid 60-100x slowdown from oversubscription.
+if "OMP_NUM_THREADS" not in environ_vars:
+    environ_vars["OMP_NUM_THREADS"] = "48"
+    logging.info("Set OMP_NUM_THREADS=48 for rocBLAS/AOCL tests")
+
 # If smoke tests are enabled, we run smoke tests only.
 # Otherwise, we run the normal test suite
 test_type = os.getenv("TEST_TYPE", "full")
@@ -41,5 +46,6 @@
 subprocess.run(
     cmd,
     cwd=THEROCK_DIR,
+    env=environ_vars,
     check=True,
 )

build_tools/validate_static_library.py

@@ -0,0 +1,56 @@
+#!/usr/bin/env python
+"""Validates that a static library exists and is a valid archive."""
+
+import argparse
+import os
+import subprocess
+import sys
+
+
+def run(args: argparse.Namespace):
+    for static_lib in args.static_libs:
+        print(f"Validating static library: {static_lib}", end="")
+
+        # Check if file exists
+        if not os.path.isfile(static_lib):
+            print(f" : ERROR - File does not exist")
+            sys.exit(1)
+
+        # Check if it's a valid archive using 'ar'
+        try:
+            result = subprocess.run(
+                ["ar", "t", static_lib],
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                check=True,
+            )
+            # Filter empty strings to correctly count object files
+            object_files = [
+                line for line in result.stdout.decode().splitlines() if line
+            ]
+            num_objects = len(object_files)
+            if num_objects == 0:
+                print(f" : ERROR - Archive is empty (0 object files)")
+                sys.exit(1)
+            print(f" : OK (contains {num_objects} object files)")
+        except subprocess.CalledProcessError as e:
+            print(f" : ERROR - Not a valid archive: {e.stderr.decode()}")
+            sys.exit(1)
+        except FileNotFoundError:
+            # 'ar' command not found, just check file size
+            size_mb = os.path.getsize(static_lib) / (1024 * 1024)
+            if size_mb <= 0:
+                print(f" : ERROR - File size is 0 MB, not a valid static library")
+                sys.exit(1)
+            print(f" : OK ({size_mb:.1f} MB)")
+
+
+def main(argv):
+    p = argparse.ArgumentParser()
+    p.add_argument("static_libs", nargs="*", help="Static libraries to validate")
+    args = p.parse_args(argv)
+    run(args)
+
+
+if __name__ == "__main__":
+    main(sys.argv[1:])

cmake/therock_subproject.cmake

@@ -710,6 +710,43 @@ function(therock_cmake_subproject_activate target_name)
   list(APPEND _build_env_pairs "--unset=HIP_PATH")
   list(APPEND _build_env_pairs "--unset=HIP_DIR")
 
+  # On Windows, preserve critical MSVC and Windows SDK environment variables
+  # These are needed for CMake's compiler detection and FindOpenMP to work correctly
+  # in nested CMake invocations (like AOCL's build system).
+  if(WIN32)
+    # Preserve key Windows SDK and MSVC environment variables
+    foreach(_win_env_var
+      PATH
+      WindowsSdkDir
+      WindowsSDKVersion
+      WindowsSDKLibVersion
+      WindowsLibPath
+      VCToolsInstallDir
+      VCToolsRedistDir
+      UniversalCRTSdkDir
+      UCRTVersion
+      INCLUDE
+      LIB
+      LIBPATH
+      DevEnvDir
+      VisualStudioVersion
+      VSCMD_VER
+      VSCMD_ARG_HOST_ARCH
+      VSCMD_ARG_TGT_ARCH
+    )
+      if(DEFINED ENV{${_win_env_var}})
+        # Escape semicolons in the environment variable value to prevent CMake from
+        # treating them as list separators (critical for PATH, INCLUDE, LIB, etc.)
+        string(REPLACE ";" "\\;" _escaped_value "$ENV{${_win_env_var}}")
+        # Remove trailing backslashes to prevent them from escaping the closing quote
+        # This is critical for paths like "C:\Program Files\...\10\" where the trailing
+        # backslash would escape the quote and break argument parsing in cmake -E env
+        string(REGEX REPLACE "\\\\+$" "" _escaped_value "${_escaped_value}")
+        list(APPEND _build_env_pairs "${_win_env_var}=${_escaped_value}")
+      endif()
+    endforeach()
+  endif()
+
   # Handle compiler toolchain.
   set(_compiler_toolchain_addl_depends)
   set(_compiler_toolchain_init_contents)

cmake/therock_testing.cmake

@@ -33,3 +33,38 @@ function(therock_test_validate_shared_lib)
     )
   endforeach()
 endfunction()
+
+# Adds a test for static libraries under a common path.
+# PATH: Common path (relative to CMAKE_CURRENT_BINARY_DIR if not absolute)
+# LIB_NAMES: Library names to validate
+function(therock_test_validate_static_lib)
+  cmake_parse_arguments(
+    PARSE_ARGV 0 ARG
+    ""
+    "PATH"
+    "LIB_NAMES"
+  )
+
+  # Skip static library validation for sanitizer builds (matches shared lib behavior).
+  if(NOT "${THEROCK_SANITIZER}" STREQUAL "")
+    return()
+  endif()
+
+  if(WIN32)
+    # This helper is Linux only.
+    return()
+  endif()
+
+  if(NOT IS_ABSOLUTE ARG_PATH)
+    cmake_path(ABSOLUTE_PATH ARG_PATH BASE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}")
+  endif()
+
+  foreach(lib_name ${ARG_LIB_NAMES})
+    add_test(
+      NAME therock-validate-static-lib-${lib_name}
+      COMMAND
+        "${Python3_EXECUTABLE}" "${THEROCK_SOURCE_DIR}/build_tools/validate_static_library.py"
+          "${ARG_PATH}/${lib_name}"
+    )
+  endforeach()
+endfunction()

math-libs/BLAS/CMakeLists.txt

@@ -166,12 +166,14 @@ list(APPEND _blas_subproject_names hipBLASLt)
 
 set(rocBLAS_optional_runtime_deps)
 if(NOT WIN32)
-  # rocBLAS is hard-coded to not expect rocm-smi.
   if(TARGET rocm_smi_lib)
     list(APPEND rocBLAS_optional_runtime_deps rocm_smi_lib)
   endif()
-elseif(THEROCK_BUILD_TESTING)
-  list(APPEND rocBLAS_optional_runtime_deps therock-host-blas)
+endif()
+if(THEROCK_BUILD_TESTING)
+  # CPU reference BLAS libraries for rocBLAS clients/tests (mirror OpenBLAS: runtime dep only)
+  list(APPEND rocBLAS_optional_runtime_deps therock-host-blas)       # OpenBLAS
+  list(APPEND rocBLAS_optional_runtime_deps therock-aocl-blas)       # AOCL-BLAS (Linux & Windows)
 endif()
 
 therock_cmake_subproject_declare(rocBLAS
@@ -183,6 +185,9 @@ therock_cmake_subproject_declare(rocBLAS
     -DROCM_PATH=
     -DROCM_DIR=
     -Dhipblaslt_path=
+    # BUILD_DIR: Custom variable used by rocBLAS clients to locate bundled dependencies
+    # Set to rocBLAS's build directory so it can find deps/ subdirectory
+    "-DBUILD_DIR=${CMAKE_CURRENT_BINARY_DIR}/rocBLAS/build"
     -DBUILD_WITH_TENSILE=ON
     -DBUILD_WITH_HIPBLASLT=ON
     # TODO: With `Tensile_TEST_LOCAL_PATH` set, the resulting build path is ${Tensile_TEST_LOCAL_PATH}/build.
@@ -192,7 +197,7 @@ therock_cmake_subproject_declare(rocBLAS
     -DTENSILE_VERSION=
     -DBUILD_CLIENTS_TESTS=${THEROCK_BUILD_TESTING}
     -DBUILD_CLIENTS_BENCHMARKS=${THEROCK_BUILD_TESTING}
-    -DLINK_BLIS=OFF
+    -DLINK_BLIS=${THEROCK_BUILD_TESTING}  # Links AOCL-BLAS (libaocl.a) for CPU BLAS
   CMAKE_INCLUDES
     therock_explicit_finders.cmake
   COMPILER_TOOLCHAIN

third-party/CMakeLists.txt

@@ -21,10 +21,13 @@ add_subdirectory(frugally-deep)
 # spdlog depends on fmt.
 add_subdirectory(spdlog)
 
-# Host math libraries.
+# Host math libraries (CPU reference BLAS for testing).
 if(THEROCK_ENABLE_HOST_BLAS)
   add_subdirectory(host-blas)
 endif()
+if(THEROCK_ENABLE_HOST_AOCL_BLAS)
+  add_subdirectory(aocl)  # AOCL-BLAS: Alternative to OpenBLAS
+endif()
 if(THEROCK_ENABLE_HOST_SUITE_SPARSE)
   add_subdirectory(SuiteSparse)
 endif()