[ARM] Windows ARM64 support

openvinotoolkit · Dec 4, 2024 · fc08d8c · fc08d8c
1 parent b39ffaa
commit fc08d8c
Show file tree

Hide file tree

Showing 10 changed files with 51 additions and 55 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -64,10 +64,14 @@ if(POLICY CMP0082)
     cmake_policy(SET CMP0082 NEW)
 endif()
 
-if("${CMAKE_BUILD_TYPE}" STREQUAL "")
-    message(STATUS "CMAKE_BUILD_TYPE is unset, defaulting to Release")
-    set(CMAKE_BUILD_TYPE "Release" CACHE STRING
-        "Choose the type of build, options are: None Debug Release RelWithDebInfo MinSizeRel RelWithAssert RelWithMDd...")
+get_property(ONEDNN_GENERATOR_MULTI_CONFIG GLOBAL PROPERTY GENERATOR_IS_MULTI_CONFIG)
+if(NOT ONEDNN_GENERATOR_MULTI_CONFIG)
+    if(NOT CMAKE_BUILD_TYPE)
+        message(STATUS "CMAKE_BUILD_TYPE is unset, defaulting to 'Release'")
+        set(CMAKE_BUILD_TYPE "Release" CACHE STRING
+            "Choose the type of build, options are: None Debug Release RelWithDebInfo MinSizeRel RelWithAssert RelWithMDd...")
+    endif()
+    string(TOUPPER "${CMAKE_BUILD_TYPE}" UPPERCASE_CMAKE_BUILD_TYPE)
 endif()
 
 set(PROJECT_NAME "oneDNN")
@@ -109,8 +113,6 @@ set(CMAKE_SRC_CCXX_FLAGS)       # SRC specifics
 set(CMAKE_EXAMPLE_CCXX_FLAGS)   # EXAMPLE specifics
 set(CMAKE_TEST_CCXX_FLAGS)      # TESTS specifics
 
-string(TOUPPER "${CMAKE_BUILD_TYPE}" UPPERCASE_CMAKE_BUILD_TYPE)
-
 set(LIB_PACKAGE_NAME "dnnl")
 
 include("cmake/dnnl_compat.cmake")

diff --git a/cmake/platform.cmake b/cmake/platform.cmake
@@ -39,6 +39,7 @@ set(CMAKE_CCXX_FLAGS)
 set(CMAKE_CCXX_NOWARN_FLAGS)
 set(CMAKE_CCXX_NOEXCEPT_FLAGS)
 set(DEF_ARCH_OPT_FLAGS)
+set(DEF_ARCH_OPT_FLAGS_NON_DEBUG)
 
 # Compatibility with DNNL
 if($ENV{ONEDNN_WERROR})
@@ -119,7 +120,7 @@ endif()
 if(MSVC)
     set(USERCONFIG_PLATFORM "x64")
     append_if(DNNL_WERROR CMAKE_CCXX_FLAGS "/WX")
-    if(${CMAKE_CXX_COMPILER_ID} STREQUAL MSVC)
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
         append(CMAKE_CCXX_FLAGS "/MP")
         # increase number of sections in obj file
         append(CMAKE_CCXX_FLAGS "/bigobj")
@@ -235,9 +236,7 @@ elseif(UNIX OR MINGW)
     # compiler specific settings
     if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
         if(DNNL_TARGET_ARCH MATCHES "^(AARCH64|ARM)$")
-             if (NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-                 set(DEF_ARCH_OPT_FLAGS "-O3")
-             endif()
+             set(DEF_ARCH_OPT_FLAGS_NON_DEBUG "-O3")
              if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
                  # Defaults to a generic cpu target, equivalent to setting -mtune=generic -march=armv8-a.
                  # This ensures no implementation specific tuning, or architectural features beyond
@@ -247,21 +246,17 @@ elseif(UNIX OR MINGW)
                  append(DEF_ARCH_OPT_FLAGS "-mcpu=generic")
              endif()
         elseif(DNNL_TARGET_ARCH STREQUAL "PPC64")
-             if (NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-                 set(DEF_ARCH_OPT_FLAGS "-O3")
-             endif()
-             # For native compilation tune for the host processor
-             if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
-                 append(DEF_ARCH_OPT_FLAGS "-mcpu=native")
-             endif()
+            set(DEF_ARCH_OPT_FLAGS_NON_DEBUG "-O3")
+            # For native compilation tune for the host processor
+            if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
+                append(DEF_ARCH_OPT_FLAGS "-mcpu=native")
+            endif()
         elseif(DNNL_TARGET_ARCH STREQUAL "S390X")
-             if (NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-                 set(DEF_ARCH_OPT_FLAGS "-O3")
-             endif()
-             # For native compilation tune for the host processor
-             if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
-                 append(DEF_ARCH_OPT_FLAGS "-march=native")
-             endif()
+            set(DEF_ARCH_OPT_FLAGS_NON_DEBUG "-O3")
+            # For native compilation tune for the host processor
+            if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
+                append(DEF_ARCH_OPT_FLAGS "-march=native")
+            endif()
         elseif(DNNL_TARGET_ARCH STREQUAL "X64")
              platform_clang_x64_arch_ccxx_flags(DEF_ARCH_OPT_FLAGS)
         endif()
@@ -326,7 +321,7 @@ elseif(UNIX OR MINGW)
             endif()
         endif()
 
-    elseif("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
+    elseif(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
         # XXX: Suppress a warning that pops up when using a function pointer
         # to an OpenCL function as a template argument (GCC Bugzilla – Bug 71463).
         if (DNNL_GPU_RUNTIME STREQUAL "OCL")
@@ -340,9 +335,8 @@ elseif(UNIX OR MINGW)
         endif()
 
         if(DNNL_TARGET_ARCH MATCHES "^(AARCH64|ARM)$")
-            if (NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-                set(DEF_ARCH_OPT_FLAGS "-O3")
-            endif()
+            set(DEF_ARCH_OPT_FLAGS_NON_DEBUG "-O3")
+            # For native compilation tune for the host processor
             if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
                  # Defaults to a generic cpu target, equivalent to setting -mtune=generic -march=armv8-a.
                  # This ensures no implementation specific tuning, or architectural features beyond
@@ -352,18 +346,14 @@ elseif(UNIX OR MINGW)
                  append(DEF_ARCH_OPT_FLAGS "-mcpu=generic")
             endif()
         elseif(DNNL_TARGET_ARCH STREQUAL "PPC64")
-            if (NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-                set(DEF_ARCH_OPT_FLAGS "-O3")
-            endif()
+            set(DEF_ARCH_OPT_FLAGS_NON_DEBUG "-O3")
             # In GCC, -ftree-vectorize is turned on under -O3 since 2007.
             # For native compilation tune for the host processor
             if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
                 append(DEF_ARCH_OPT_FLAGS "-mcpu=native")
             endif()
         elseif(DNNL_TARGET_ARCH STREQUAL "S390X")
-            if (NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-                set(DEF_ARCH_OPT_FLAGS "-O3")
-            endif()
+            set(DEF_ARCH_OPT_FLAGS_NON_DEBUG "-O3")
             # In GCC, -ftree-vectorize is turned on under -O3 since 2007.
             # For native compilation tune for the host processor
             if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
@@ -418,8 +408,13 @@ if(DNNL_ARCH_OPT_FLAGS STREQUAL "HostOpts")
     set(DNNL_ARCH_OPT_FLAGS "${DEF_ARCH_OPT_FLAGS}")
 endif()
 
-append(CMAKE_C_FLAGS "${CMAKE_CCXX_FLAGS} ${DNNL_ARCH_OPT_FLAGS}")
-append(CMAKE_CXX_FLAGS "${CMAKE_CCXX_FLAGS} ${DNNL_ARCH_OPT_FLAGS}")
+append(CMAKE_C_FLAGS "${CMAKE_CCXX_FLAGS}")
+append(CMAKE_CXX_FLAGS "${CMAKE_CCXX_FLAGS}")
+
+foreach(config RELEASE RELWITHDEBINFO MINSIZEREL)
+    append(CMAKE_CXX_FLAGS_${config} "${DEF_ARCH_OPT_FLAGS_NON_DEBUG}")
+    append(CMAKE_C_FLAGS_${config} "${DEF_ARCH_OPT_FLAGS_NON_DEBUG}")
+endforeach()
 
 if(APPLE)
     set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)

diff --git a/src/cpu/CMakeLists.txt b/src/cpu/CMakeLists.txt
@@ -40,10 +40,8 @@ if(DNNL_TARGET_ARCH STREQUAL "PPC64")
     file(GLOB FILES_REQUIRED_OPT
         ${CMAKE_CURRENT_SOURCE_DIR}/gemm/*.[ch]pp
     )
-    if(NOT UPPERCASE_CMAKE_BUILD_TYPE STREQUAL "DEBUG")
-        set_source_files_properties(${FILES_REQUIRED_OPT}
-            PROPERTIES COMPILE_FLAGS "-O3 -funroll-loops")
-    endif()
+    set_source_files_properties(${FILES_REQUIRED_OPT}
+        PROPERTIES COMPILE_FLAGS <$<NOT:$<CONFIG:Debug>>,"-O3 -funroll-loops">)
 endif()
 
 if(NOT DNNL_ENABLE_JIT_PROFILING)

diff --git a/src/cpu/aarch64/acl_convolution_utils.cpp b/src/cpu/aarch64/acl_convolution_utils.cpp
@@ -14,6 +14,8 @@
 * limitations under the License.
 *******************************************************************************/
 
+#include <cstdint>
+
 #include "acl_convolution_utils.hpp"
 #include "common/convolution_pd.hpp"
 #include "common/utils.hpp"

diff --git a/src/cpu/aarch64/acl_gemm_convolution.cpp b/src/cpu/aarch64/acl_gemm_convolution.cpp
@@ -55,7 +55,7 @@ status_t acl_gemm_convolution_fwd_t<src_t, wei_t, dst_t, bia_t>::pd_t::init(
     using smask_t = primitive_attr_t::skip_mask_t;
 
     bool ok = is_fwd() && set_default_alg_kind(alg_kind::convolution_direct)
-            && expect_data_types(src_t, wei_t, bia_t, dst_t, undef)
+            && expect_data_types(src_t, wei_t, bia_t, dst_t, data_type::undef)
             && !has_zero_dim_memory()
             && attr()->has_default_values(
                     smask_t::post_ops | smask_t::fpmath_mode, dst_t);

diff --git a/src/cpu/aarch64/acl_indirect_gemm_convolution.cpp b/src/cpu/aarch64/acl_indirect_gemm_convolution.cpp
@@ -94,9 +94,9 @@ status_t acl_indirect_gemm_convolution_fwd_t::pd_t::init(engine_t *engine) {
     using namespace data_type;
     using smask_t = primitive_attr_t::skip_mask_t;
 
-    const bool is_fp16_ok = expect_data_types(f16, f16, f16, f16, undef)
+    const bool is_fp16_ok = expect_data_types(f16, f16, f16, f16, data_type::undef)
             && attr()->has_default_values(smask_t::post_ops, f16);
-    const bool is_fp32_ok = expect_data_types(f32, f32, f32, f32, undef)
+    const bool is_fp32_ok = expect_data_types(f32, f32, f32, f32, data_type::undef)
             && attr()->has_default_values(
                     smask_t::post_ops | smask_t::fpmath_mode, f32);
     bool ok = is_fwd() && set_default_alg_kind(alg_kind::convolution_direct)

diff --git a/src/cpu/aarch64/cpu_isa_traits.hpp b/src/cpu/aarch64/cpu_isa_traits.hpp
@@ -25,11 +25,11 @@
 #include "common/utils.hpp"
 #include "dnnl_types.h"
 
-/* in order to make selinux happy memory that would be marked with X-bit should
- * be obtained with mmap */
-#if !defined(_WIN32)
-#define XBYAK_USE_MMAP_ALLOCATOR
-#endif
+// /* in order to make selinux happy memory that would be marked with X-bit should
+//  * be obtained with mmap */
+// #if !defined(_WIN32)
+// #define XBYAK_USE_MMAP_ALLOCATOR
+// #endif
 
 #include "cpu/aarch64/xbyak_aarch64/xbyak_aarch64/xbyak_aarch64.h"
 #include "cpu/aarch64/xbyak_aarch64/xbyak_aarch64/xbyak_aarch64_util.h"

diff --git a/src/cpu/aarch64/jit_uni_deconv_zp_pad_str_kernel.cpp b/src/cpu/aarch64/jit_uni_deconv_zp_pad_str_kernel.cpp
@@ -16,6 +16,7 @@
 *******************************************************************************/
 
 #include <cassert>
+#include <cstdint>
 #include "common/dnnl_thread.hpp"
 #include "common/utils.hpp"
 #include "cpu/aarch64/jit_primitive_conf.hpp"
@@ -128,7 +129,7 @@ template <cpu_isa_t isa>
 uint32_t jit_uni_deconv_zp_pad_str_kernel_t<isa>::get_next_vmm_idx() {
     static constexpr int max_v_regs = cpu_isa_traits<isa>::n_vregs;
 
-    const ZReg vmm {static_cast<unsigned int>(current_vmm_++)};
+    const ZReg vmm {static_cast<uint32_t>(current_vmm_++)};
 
     if (current_vmm_ == max_v_regs) current_vmm_ = number_reserved_vmms_;
 

diff --git a/src/cpu/x64/CMakeLists.txt b/src/cpu/x64/CMakeLists.txt
@@ -50,12 +50,8 @@ if(MSVC)
     set_source_files_properties(${FILES_WITHOUT_OPT}
         PROPERTIES COMPILE_FLAGS "/Od")
 else()
-    if(UPPERCASE_CMAKE_BUILD_TYPE STREQUAL "DEBUG")
-        # Some compilers enable optimizations by default.
-        set(OPT_LEVEL "-O0")
-    else()
-        set(OPT_LEVEL "-O1")
-    endif()
+    # Some compilers enable optimizations by default.
+    set(OPT_LEVEL $<IF:$<CONFIG:Debug>,-O0,-O1>)
     set_source_files_properties(${FILES_WITHOUT_OPT}
         PROPERTIES COMPILE_FLAGS "${OPT_LEVEL}")
 endif()

diff --git a/src/cpu/x64/xbyak/xbyak_util.h b/src/cpu/x64/xbyak/xbyak_util.h
@@ -303,6 +303,7 @@ class Cpu {
 			}
 			return;
 		}
+#ifdef XBYAK_INTEL_CPU_SPECIFIC
 		// intel
 		const uint32_t NO_CACHE = 0;
 		const uint32_t DATA_CACHE = 1;
@@ -347,6 +348,7 @@ class Cpu {
 				dataCacheLevels_++;
 			}
 		}
+#endif // XBYAK_INTEL_CPU_SPECIFIC
 	}
 
 public: