fix(mge/build): remove MEGDNN flags from command line

GitOrigin-RevId: a9813a44e4

CMakeLists.txt

@@ -177,7 +177,7 @@ if(MGE_WITH_CUDA)
 
     if(NOT MGE_CUDA_GENCODE)
         if(${MGE_ARCH} STREQUAL "x86_64" OR ${MGE_ARCH} STREQUAL "i386")
-            set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -DMEGDNN_THREADS_512=0")
+            set(MEGDNN_THREADS_512 0)
             if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER "10.0.0" OR ${CMAKE_CUDA_COMPILER_VERSION} VERSION_EQUAL "10.0.0")
                 set(MGE_CUDA_GENCODE "${MGE_CUDA_GENCODE} -gencode arch=compute_52,code=sm_52")
                 set(MGE_CUDA_GENCODE "${MGE_CUDA_GENCODE} -gencode arch=compute_60,code=sm_60")
@@ -202,7 +202,7 @@ if(MGE_WITH_CUDA)
             message(FATAL_ERROR "Unsupported CUDA host arch.")
         endif()
     else()
-        set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -DMEGDNN_THREADS_512=1")
+        set(MEGDNN_THREADS_512 1)
     endif()
 
     set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${MGE_CUDA_GENCODE}")
@@ -287,35 +287,31 @@ option(MGE_WITH_MKLDNN "Enable Intel MKL_DNN support," ON)
 
 # MKLDNN build
 if(MGE_WITH_MKLDNN AND ${MGE_ARCH} STREQUAL "x86_64")
-    add_definitions(-DMEGDNN_X86_WITH_MKL_DNN)
     include(cmake/MKL_DNN.cmake)
+    set(MEGDNN_X86_WITH_MKL_DNN 1)
 endif()
 
 # RTTI
 if(MGE_ENABLE_RTTI)
-    add_definitions(-DMEGDNN_ENABLE_MANGLING=0 -DMEGDNN_ENABLE_RTTI=1)
+    set(MEGDNN_ENABLE_MANGLING 0)
+    set(MEGDNN_ENABLE_RTTI 1)
 else()
-    add_definitions(-DMEGDNN_ENABLE_MANGLING=1 -DMEGDNN_ENABLE_RTTI=0)
+    set(MEGDNN_ENABLE_MANGLING 1)
+    set(MEGDNN_ENABLE_RTTI 0)
 endif()
 set(MGB_VERBOSE_TYPEINFO_NAME ${MGE_ENABLE_RTTI})
 
 # Logging
-if(MGE_ENABLE_LOGGING)
-    add_definitions(-DMEGDNN_ENABLE_LOGGING=1)
-else()
-    add_definitions(-DMEGDNN_ENABLE_LOGGING=0)
-endif()
 set(MGB_ENABLE_LOGGING ${MGE_ENABLE_LOGGING})
+set(MEGDNN_ENABLE_LOGGING ${MGE_ENABLE_LOGGING})
 set(MGB_ENABLE_JSON ${MGE_ENABLE_LOGGING})
 
 # Exception
-if(MGE_ENABLE_EXCEPTIONS)
-    add_definitions(-DMEGDNN_ENABLE_EXCEPTIONS=1)
-else()
+if(NOT MGE_ENABLE_EXCEPTIONS)
     message(STATUS "Exceptions disabled; MegEngine would kill itself when it is supposed to throw an exception.")
-    add_definitions(-DMEGDNN_ENABLE_EXCEPTIONS=0)
 endif()
 set(MGB_ENABLE_EXCEPTION ${MGE_ENABLE_EXCEPTIONS})
+set(MEGDNN_ENABLE_EXCEPTIONS ${MGE_ENABLE_EXCEPTIONS})
 
 # JIT
 if(MGE_WITH_JIT AND MGE_WITH_HALIDE)
@@ -330,8 +326,15 @@ if(CMAKE_THREAD_LIBS_INIT)
     set(MGB_HAVE_THREAD 1)
 endif()
 
+if(MGE_WITH_TEST)
+    # use intra-op multi threads
+    set(MEGDNN_ENABLE_MULTI_THREADS 1)
+endif()
+
 # CUDA
 set(MGB_CUDA ${MGE_WITH_CUDA})
+set(MEGDNN_WITH_CUDA ${MGE_WITH_CUDA})
+
 
 # Debug info
 if(${CMAKE_BUILD_TYPE} STREQUAL "Debug" OR ${CMAKE_BUILD_TYPE} STREQUAL "RelWithDebInfo")
@@ -357,8 +360,46 @@ endif()
 # Distributed communication
 set(MGB_ENABLE_OPR_MM ${MGE_WITH_DISTRIBUTED})
 
+# MGE_ARCH related flags
+if(MGE_ARCH STREQUAL "x86_64" OR MGE_ARCH STREQUAL "i386")
+    if(MGE_BLAS STREQUAL "MKL")
+        set(MEGDNN_X86_WITH_MKL 1)
+    elseif(MGE_BLAS STREQUAL "OpenBLAS")
+        set(MEGDNN_X86_WITH_OPENBLAS 1)
+    endif()
+endif()
+
+# Enable Naive
+if(MGE_ARCH STREQUAL "naive")
+    set(MEGDNN_NAIVE 1)
+    message(WARNING "MEGDNN_NAIVE is enabled; MegDNN performance is degraded.")
+endif()
+
+if(MGE_ARCH STREQUAL "x86_64" OR MGE_ARCH STREQUAL "i386")
+    set(MEGDNN_X86 1)
+    if(MGE_ARCH STREQUAL "x86_64")
+        set(MEGDNN_X86_64 1)
+        set(MEGDNN_64_BIT 1)
+        if(NOT MSVC)
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
+        endif()
+    else()
+        set(MEGDNN_X86_32 1)
+        if(NOT MSVC)
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m32")
+        endif()
+    endif()
+    if(NOT MSVC)
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.2 -mfpmath=sse")
+    endif()
+endif()
+
+
+set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${MARCH}")
+
 # Write out megbrain_build_config.h
-configure_file(src/core/include/megbrain_build_config.h.in ${CMAKE_CURRENT_BINARY_DIR}/genfiles/megbrain_build_config.h)
+# It defines macros needed by both megbrain and dnn
+configure_file(src/megbrain_build_config.h.in ${CMAKE_CURRENT_BINARY_DIR}/genfiles/megbrain_build_config.h)
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/genfiles/megbrain_build_config.h DESTINATION include)
 
 add_subdirectory(dnn)

dnn/CMakeLists.txt

@@ -1,40 +1,3 @@
-if(${MGE_ARCH} STREQUAL "x86_64" OR ${MGE_ARCH} STREQUAL "i386")
-    if(${MGE_BLAS} STREQUAL "MKL")
-        add_definitions(-DMEGDNN_X86_WITH_MKL)
-    elseif(${MGE_BLAS} STREQUAL "OpenBLAS")
-        add_definitions(-DMEGDNN_X86_WITH_OPENBLAS)
-    endif()
-endif()
-
-# Enable Naive
-if(${MGE_ARCH} STREQUAL "naive")
-    add_definitions(-DMEGDNN_NAIVE=1)
-    message(WARNING "MEGDNN_NAIVE is enabled; MegDNN performance is degraded.")
-else()
-    add_definitions(-DMEGDNN_NAIVE=0)
-endif()
-
-
-if(${MGE_ARCH} STREQUAL "x86_64" OR ${MGE_ARCH} STREQUAL "i386")
-    add_definitions(-DMEGDNN_X86=1)
-    if(${MGE_ARCH} STREQUAL "x86_64")
-        add_definitions(-DMEGDNN_X86_64 -DMEGDNN_64_BIT)
-        if(NOT MSVC)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
-        endif()
-    else()
-        add_definitions(-DMEGDNN_X86_32)
-        if(NOT MSVC)
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m32")
-        endif()
-    endif()
-    if(NOT MSVC)
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse4.2 -mfpmath=sse")
-    endif()
-endif()
-
-
-set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${MARCH}")
 
 list(APPEND OPR_PARAM_DEFS_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/scripts/opr_param_defs.py)
 set(OPR_PARAM_DEFS_SCRIPT ${CMAKE_CURRENT_SOURCE_DIR}/scripts/gen_param_defs.py)
@@ -89,8 +52,6 @@ add_dependencies(opr_param_defs _opr_param_defs)
 
 
 if(MGE_WITH_TEST)
-    # use multi threads
-    add_definitions (-DMEGDNN_ENABLE_MULTI_THREADS=1)
     add_subdirectory(test)
 endif()
 

dnn/include/megdnn/config/config.h

@@ -9,22 +9,10 @@
  * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  */
 
+#include "megbrain_build_config.h"
+
 #if !defined(__CUDACC__)
 
-// Try to detect if no architecture flags defined.
-#if !defined(MEGDNN_NAIVE) && !defined(MEGDNN_X86) &&         \
-        !defined(MEGDNN_X86_64) && !defined(MEGDNN_X86_32) && \
-        !defined(MEGDNN_64_BIT) && !defined(MEGDNN_MIPS) &&   \
-        !defined(MEGDNN_ARMV7) && !defined(MEGDNN_AARCH64)
-#if defined(__x86_64__) || defined(_M_X64)
-#define MEGDNN_X86 1
-#define MEGDNN_X86_64 1
-#define MEGDNN_64_BIT 1
-#elif defined(__i386) || defined(_M_IX86)
-#define MEGDNN_X86 1
-#define MEGDNN_X86_32 1
-#endif
-#endif
 
 #endif  // !defined(__CUDACC__)
 

dnn/src/CMakeLists.txt

@@ -1,7 +1,9 @@
-
 set(LIBMEGDNN_DEF)
 file(GLOB_RECURSE SOURCES common/*.cpp naive/*.cpp)
 
+# Build configure
+list(APPEND SOURCES ${PROJECT_BINARY_DIR}/genfiles/megbrain_build_config.h)
+
 if(NOT ${MGE_ARCH} STREQUAL "naive")
     file(GLOB_RECURSE SOURCES_ fallback/*.cpp)
     list(APPEND SOURCES ${SOURCES_})
@@ -24,7 +26,6 @@ if(MGE_WITH_CUDA)
 
     file(GLOB_RECURSE CUSOURCES cuda/*.cu)
     list(APPEND SOURCES ${CUSOURCES})
-    list(APPEND LIBMEGDNN_DEF -DMEGDNN_WITH_CUDA=1)
 endif()
 
 
@@ -33,7 +34,7 @@ add_definitions(${LIBMEGDNN_DEF})
 add_library(megdnn EXCLUDE_FROM_ALL STATIC ${SOURCES})
 
 target_link_libraries(megdnn opr_param_defs)
-target_include_directories(megdnn PUBLIC ${PROJECT_SOURCE_DIR}/dnn/include)
+target_include_directories(megdnn PUBLIC ${PROJECT_BINARY_DIR}/genfiles ${PROJECT_SOURCE_DIR}/dnn/include)
 target_include_directories(megdnn PRIVATE ${PROJECT_SOURCE_DIR}/dnn ${PROJECT_SOURCE_DIR}/third_party/midout/src)
 
 install(DIRECTORY ${PROJECT_SOURCE_DIR}/dnn/include DESTINATION . FILES_MATCHING PATTERN "*.h*")

dnn/src/x86/conv_bias/f32/algos.cpp

@@ -645,7 +645,7 @@ void ConvBiasImpl::AlgoMatrixMul::kimpl(const NCBKernParam& param,
     }
 }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 static inline void mkldnn_fp32_conv_instance(
         const ConvBiasImpl::NCBKernParam& param, const uint32_t ocpg,
         const uint32_t icpg, const uint32_t group, const uint32_t in,

dnn/src/x86/conv_bias/f32/algos.h

@@ -186,7 +186,7 @@ public:
     void* type() const override;
 };
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 class ConvBiasImpl::AlgoMkldnnConv final : public AlgoBase {
     static void kern_mkldnn_fp32(const NCBKernParam& param,
                                  const NCBKernIndex&);

dnn/src/x86/conv_bias/int8/algos.cpp

@@ -20,13 +20,13 @@
 #include "src/x86/conv_bias/postprocess_helper.h"
 #include "src/x86/handle.h"
 #include "src/x86/utils.h"
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 #include <mkldnn.hpp>
 #endif
 
 #include <cstring>
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 using namespace dnnl;
 #endif
 using namespace megdnn;
@@ -161,7 +161,7 @@ ConvBiasImpl::AlgoDirectAvx2Stride1Int8::get_kimpls(
     return direct_conv_avx2_stride1::get_kimpls(param, bundle);
 }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 bool ConvBiasImpl::AlgoMkldnnQint8::usable(FallbackConvBiasImpl*,
                                            const NCBKernSizeParam& param,
                                            AlgoSelectionStrategy) const {
@@ -353,7 +353,7 @@ void ConvBiasImpl::AlgoMkldnnQint8::kern_mkldnn_s8x8x32(
 #undef REORDER_MEMORY
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 /* ===================== mkldnn qint8 matmul algo ===================== */
 bool ConvBiasImpl::AlgoMkldnnMatmulQint8::usable(FallbackConvBiasImpl*,
                                                  const NCBKernSizeParam& param,

dnn/src/x86/conv_bias/int8/algos.h

@@ -58,7 +58,7 @@ public:
     void* type() const override;
 };
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 /* ===================== mkldnn qint8 algo ===================== */
 class ConvBiasImpl::AlgoMkldnnQint8 final : public AlgoBase {
     static void kern_mkldnn_s8x8x32(const NCBKernParam& param,

dnn/src/x86/conv_bias/opr_impl.cpp

@@ -25,7 +25,7 @@ namespace {
 uint8_t x86_algo_type_storage;
 void* x86_algo_type = &x86_algo_type_storage;
 }  // anonymous namespace
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 void* ConvBiasImpl::AlgoMkldnnQint8::type() const {
     return x86_algo_type;
 }
@@ -78,7 +78,7 @@ class ConvBiasImpl::AlgoPack : NonCopyableObj {
     AlgoAVX2DirectConvStride2 avx2_stride2_direct;
     AlgoChanWiseAvx2Stride1Qint8 avx2_stride1_chanwsie_qint8;
     AlgoMatrixMul matmul;
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     AlgoMkldnnMatmulQint8 mkldnn_matmul_qint8;
     //! Because the mkldnnconv need handle
     AlgoMkldnnQint8 mkldnn_qint8;
@@ -87,7 +87,7 @@ class ConvBiasImpl::AlgoPack : NonCopyableObj {
     SmallVector<std::unique_ptr<AlgoBase>> refhold;
 public:
     AlgoPack() {
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
         //! Create the mkldnn algo
         all_algos.emplace_back(&mkldnn_conv_fp32);
         all_algos.emplace_back(&mkldnn_matmul_qint8);

dnn/src/x86/conv_bias/opr_impl.h

@@ -32,7 +32,7 @@ public:
     class AlgoDirectAvx2Stride1Int8;
     class AlgoAVX2DirectConvStride2;
     class AlgoChanWiseAvx2Stride1Qint8;
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     class AlgoMkldnnConv;
     class AlgoMkldnnQint8;
     class AlgoMkldnnMatmulQint8;

dnn/src/x86/handle.cpp

@@ -32,7 +32,7 @@
 #include "src/x86/warp_affine/opr_impl.h"
 #include "src/x86/warp_perspective/opr_impl.h"
 
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
 
 #include <mkl.h>
 #define STR_HELPER(x) #x
@@ -57,11 +57,11 @@ HandleImpl::HandleImpl(megcoreComputingHandle_t computing_handle,
                        HandleType type)
         : fallback::HandleImpl::HandleImpl(computing_handle, type) {
     disable_denorm();
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
     vmlSetMode(VML_LA | VML_FTZDAZ_ON | VML_ERRMODE_ERRNO);
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     m_mkldnn_engine = dnnl::engine(dnnl::engine::kind::cpu, 0);
     m_mkldnn_stream = dnnl::stream(m_mkldnn_engine);
 #endif

dnn/src/x86/handle.h

@@ -13,7 +13,7 @@
 
 #include "src/x86/profile.h"
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 #include <mkldnn.hpp>
 #endif
 
@@ -31,14 +31,14 @@ public:
     std::unique_ptr<Opr> create_operator();
 
     size_t alignment_requirement() const override;
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     dnnl::engine mkldnn_engine() { return m_mkldnn_engine; }
     dnnl::stream mkldnn_stream() { return m_mkldnn_stream; }
 #endif
 
 private:
     ProfileCache m_profile_cache = get_profile_cache();
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     dnnl::engine m_mkldnn_engine;
     dnnl::stream m_mkldnn_stream;
 #endif

dnn/src/x86/matrix_mul/algos.cpp

@@ -18,15 +18,15 @@
 
 #include "src/x86/matrix_mul/f32/strategy.h"
 
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
 #include <mkl.h>
 #include <mkl_cblas.h>
-#elif defined(MEGDNN_X86_WITH_OPENBLAS)
+#elif MEGDNN_X86_WITH_OPENBLAS
 #include <cblas.h>
 #else
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 #include <mkldnn.h>
 #endif
 
@@ -39,7 +39,7 @@ using namespace x86;
 namespace {
 
 void f32_blas_kern(const MatrixMulImpl::KernParam& kern_param) {
-#if defined(MEGDNN_X86_WITH_MKL) || defined(MEGDNN_X86_WITH_OPENBLAS)
+#if MEGDNN_X86_WITH_MKL || MEGDNN_X86_WITH_OPENBLAS
     auto m = kern_param.M, n = kern_param.N, k = kern_param.K;
     bool trA = kern_param.trA, trB = kern_param.trB;
     const auto Aptr = kern_param.A<dt_float32>(),
@@ -55,7 +55,7 @@ void f32_blas_kern(const MatrixMulImpl::KernParam& kern_param) {
 #endif
 }
 
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
 void f32_blas_kern_only_packA(const MatrixMulImpl::KernParam& kern_param,
                                  const void* a_panel, const void* b_panel) {
   MEGDNN_MARK_USED_VAR(b_panel);
@@ -75,7 +75,7 @@ void f32_blas_kern_only_packA(const MatrixMulImpl::KernParam& kern_param,
 
 bool MatrixMulImpl::AlgoF32Blas::usable(
         const KernSizeParam& kern_size_param) const {
-#if defined(MEGDNN_X86_WITH_MKL) || defined(MEGDNN_X86_WITH_OPENBLAS)
+#if MEGDNN_X86_WITH_MKL || MEGDNN_X86_WITH_OPENBLAS
     return kern_size_param.compute_mode == Param::ComputeMode::DEFAULT &&
            kern_size_param.format == param::MatrixMul::Format::DEFAULT &&
            kern_size_param.B_type == kern_size_param.A_type &&
@@ -93,7 +93,7 @@ MatrixMulImpl::kern_t MatrixMulImpl::AlgoF32Blas::get_kern(
 }
 
 /* ===================== AlgoF32BlasPackA====================== */
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
 bool MatrixMulImpl::AlgoF32MKLPackA::usable(
         const KernSizeParam& kern_size_param) const {
     return kern_size_param.compute_mode == Param::ComputeMode::DEFAULT &&
@@ -202,7 +202,7 @@ MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL_DETAIL(AlgoInt8x8x32Vnni,
 #endif
 
 /* ===================== Int8 mkldnn algo ===================== */
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 namespace {
 void int8x8x32_kern_mkldnn(const MatrixMulImpl::KernParam& kern_param) {
     MEGDNN_MARK_USED_VAR(kern_param);

dnn/src/x86/matrix_mul/algos.h

@@ -28,7 +28,7 @@ public:
     PackMode packmode() const override { return PackMode::NO_PACK; }
 };
 
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
 class MatrixMulImpl::AlgoF32MKLPackA : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
@@ -106,7 +106,7 @@ public:
 };
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 class MatrixMulImpl::AlgoInt8x8x32Mkldnn : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }

dnn/src/x86/matrix_mul/opr_impl.cpp

@@ -25,13 +25,13 @@ void* const MatrixMulImpl::sm_x86_algo_type = &x86_algo_type_storage;
 class MatrixMulImpl::AlgoPack : NonCopyableObj {
     AlgoF32Blas f32blas;
 
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
     AlgoF32MKLPackA f32mkl_packa;
 #endif
 #if MEGDNN_X86_WITH_VNNI
     AlgoInt8x8x32Vnni algoint8x8x32vnni;
 #endif
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     AlgoInt8x8x32Mkldnn algoint8x8x32mkldnn;
 #endif
     AlgoInt8x8x32AVX2M4N16K2 algoint8x8x32avx2_m4n16k2;
@@ -42,7 +42,7 @@ class MatrixMulImpl::AlgoPack : NonCopyableObj {
 public:
     AlgoPack() {
         if (is_supported(SIMDType::VNNI)) {
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
             all_algos.emplace_back(&algoint8x8x32mkldnn);
 #endif
 #if MEGDNN_X86_WITH_VNNI
@@ -53,11 +53,11 @@ public:
         all_algos.emplace_back(&algoint8x8x32avx2_m2n4k16);
         all_algos.emplace_back(&algoint8x8x32sse_m4n8k2);
         all_algos.emplace_back(&algof32mk8_8x8);
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
         all_algos.emplace_back(&algoint8x8x32mkldnn);
 #endif
         all_algos.emplace_back(&f32blas);
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
         all_algos.emplace_back(&f32mkl_packa);
 #endif
     }

dnn/src/x86/matrix_mul/opr_impl.h

@@ -26,14 +26,14 @@ public:
 protected:
     static void* const sm_x86_algo_type;
     class AlgoF32Blas;
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
     class AlgoF32MKLPackA;
 #endif
 #if MEGDNN_X86_WITH_VNNI
     class AlgoInt8x8x32Vnni;
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     class AlgoInt8x8x32Mkldnn;
 #endif
 

dnn/src/x86/pooling/opr_impl.cpp

@@ -17,7 +17,7 @@
 #include "src/x86/pooling/pooling_special_cases.h"
 #include "src/x86/utils.h"
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 #include "mkldnn.hpp"
 #endif
 
@@ -45,7 +45,7 @@ WorkspaceBundle get_bundle(const TensorLayout& src, const TensorLayout& dst,
     return ws;
 }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 template <dnnl::memory::format_tag format_tag, bool use_mkl_mem>
 dnnl::memory tensor_to_mkl_memory(_megdnn_tensor_in src,
                                   const dnnl::engine& mkldnn_eng,
@@ -164,7 +164,7 @@ void PoolingImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_out dst,
         return;
     }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 
     // Mkldnn provide optimized code for nhwc int8 pooling now.
     // Mkldnn can not change the layout automatic.

dnn/src/x86/utils.cpp

@@ -18,7 +18,7 @@
 #include <intrin.h>
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL) || defined(MEGDNN_X86_WITH_OPENBLAS)
+#if MEGDNN_X86_WITH_MKL || MEGDNN_X86_WITH_OPENBLAS
 #include <pmmintrin.h>
 #endif
 

dnn/test/x86/conv_bias.cpp

@@ -777,7 +777,7 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_INT8x8x32) {
                 .execs({arg.src, arg.filter, {}, {}, {}});                     \
     }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     if (megdnn::x86::is_supported(x86::SIMDType::VNNI)) {
         cb("IM2COLMATMUL:X86_INT8X8X32_MKLDNN");
     }
@@ -846,14 +846,14 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_FP32) {
                 {arg.src, arg.filter, arg.bias, {}, {}});         \
     }
 
-#if defined(MEGDNN_X86_WITH_MKL) || defined(MEGDNN_X86_WITH_OPENBLAS)
+#if MEGDNN_X86_WITH_MKL || MEGDNN_X86_WITH_OPENBLAS
     cb("IM2COLMATMUL:X86_F32_BLAS");
 #endif
 
 #undef cb
 }
 
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
 TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_FP32_PACKA) {
     using namespace conv_bias;
     std::vector<TestArg> args;
@@ -973,7 +973,7 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_QINT8) {
                 .execs({arg.src, arg.filter, {}, {}, {}});        \
     }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     if (x86::is_supported(x86::SIMDType::VNNI)) {
         cb("IM2COLMATMUL:X86_INT8X8X32_MKLDNN");
     }
@@ -1057,7 +1057,7 @@ TEST_F(X86, CONV_BIAS_MATMUL) {
     }
 }
 #if MEGDNN_WITH_BENCHMARK
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 static void x86_benchmark_fp32_mkldnn(Handle* handle) {
     constexpr size_t RUNS = 30;
     param::ConvBias param;
@@ -1304,7 +1304,7 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_WINOGRAD_WEIGHT_PREPROCESS) {
 }
 
 /*********************************** End winograd ************************/
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 static void x86_correctness_fp32_mkldnn_run(
         Checker<ConvBias>& checker, UniformIntRNG& rng, Handle* handle,
         ConvBiasForward::BiasMode bias_mode,

dnn/test/x86/convolution.cpp

@@ -20,7 +20,7 @@
 #include "test/common/workspace_wrapper.h"
 
 namespace {
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 struct ConvArg {
     size_t batch_size, fh, sh, ph, ic, ih, iw, oc, groups;
 };
@@ -224,7 +224,7 @@ TEST_F(X86, DEFAULT_CONV_MATMUL) {
     }
 }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 TEST_F(X86, CONVOLUTION_FORWARD_INT8) {
     Checker<ConvolutionForward> checker(handle());
     checker.set_before_exec_callback(
@@ -369,7 +369,7 @@ TEST_F(X86, CONVOLUTION_DIRECT_MKLDNN_C8) {
 #endif
 
 #if MEGDNN_WITH_BENCHMARK
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 TEST_F(X86, BENCHMARK_CONVOLUTION_I8x8x32_MKLDNN) {
     using namespace convolution;
     using Param = param::Convolution;

dnn/test/x86/matrix_mul.cpp

@@ -26,7 +26,7 @@ TEST_F(X86, MATRIX_MUL_VNNI_8X8X32) {
 }
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 TEST_F(X86, MATRIX_MUL_MKLDNN_8X8X32) {
     if (is_supported(SIMDType::VNNI)) {
         matrix_mul::check_matrix_mul(dtype::Int8{}, dtype::Int8{},
@@ -52,7 +52,7 @@ TEST_F(X86, MATRIX_MUL_SSE_8X8X32) {
                                  handle(), "X86_INT8X8X32_SSE_4X8X2");
 }
 
-#if defined(MEGDNN_X86_WITH_MKL)
+#if MEGDNN_X86_WITH_MKL
 TEST_F(X86, MATRIX_MUL_MKL_PACKA) {
     matrix_mul::check_matrix_mul(dtype::Float32{}, dtype::Float32{},
                                  dtype::Float32{}, handle(),
@@ -93,7 +93,7 @@ TEST_F(X86, BENCHMARK_MATRIX_MUL_8X8X32) {
             AlgoChecker<MatrixMul>("X86_INT8X8X32_VNNI"));
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
     Benchmarker<MatrixMul> benchmarker_mkldnn(handle());
     benchmarker_mkldnn.set_times(RUNS)
             .set_dtype(0, dtype::Int8{})
@@ -162,7 +162,7 @@ TEST_F(X86, BENCHMARK_MATRIX_MUL_8X8X32) {
         }
 #endif
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
         if (is_supported(SIMDType::VNNI)) {
             auto mkldnn_used =
                     benchmarker_mkldnn.exec({{M, K}, {K, N}, {}}) / RUNS;

dnn/test/x86/pooling.cpp

@@ -24,7 +24,7 @@ TEST_F(X86, POOLING) {
     }
 }
 
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 TEST_F(X86, POOLING88) {
     Checker<Pooling> checker(handle());
     auto args = pooling::get_args();
@@ -105,7 +105,7 @@ TEST_F(X86_MULTI_THREADS, BENCHMARK_POOLING) {
     test_x86_megdnn_pooling(handle());
 }
 #endif
-#if defined(MEGDNN_X86_WITH_MKL_DNN)
+#if MEGDNN_X86_WITH_MKL_DNN
 TEST_F(X86, POOLING_INT8) {
     auto args = pooling::get_args();
     for (auto&& arg : args) {

src/megbrain_build_config.h.in

@@ -0,0 +1,198 @@
+/**
+ * \file src/core/include/megbrain_build_config.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#ifndef _HEADER_MGB_BUILD_CONFIG
+#define _HEADER_MGB_BUILD_CONFIG
+
+#cmakedefine01 MGB_CUDA
+#cmakedefine01 MGB_ASSERT_LOC
+#cmakedefine01 MGB_ENABLE_DEBUG_UTIL
+#cmakedefine01 MGB_ENABLE_LOGGING
+#cmakedefine01 MGB_ENABLE_GRAD
+#cmakedefine01 MGB_VERBOSE_TYPEINFO_NAME
+#cmakedefine01 MGB_BUILD_SLIM_SERVING
+#cmakedefine01 MGB_ENABLE_EXCEPTION
+#cmakedefine01 MGB_JIT
+#cmakedefine01 MGB_JIT_HALIDE
+#cmakedefine01 MGB_ENABLE_TENSOR_RT
+#cmakedefine01 MGB_ENABLE_JSON
+#cmakedefine01 MGB_HAVE_THREAD
+#cmakedefine01 MGB_ENABLE_OPR_MM
+#cmakedefine01 MEGDNN_ENABLE_MANGLING
+
+// DNN related flags
+// Platform macro's
+#cmakedefine01 MEGDNN_WITH_CUDA
+
+
+#cmakedefine01 MEGDNN_X86_WITH_MKL
+#cmakedefine01 MEGDNN_X86_WITH_OPENBLAS
+#cmakedefine01 MEGDNN_X86_WITH_MKL_DNN
+#cmakedefine01 MEGDNN_ENABLE_RTTI
+#cmakedefine01 MEGDNN_ENABLE_LOGGING
+#cmakedefine01 MEGDNN_ENABLE_LOGGING
+#cmakedefine01 MEGDNN_ENABLE_EXCEPTIONS
+#cmakedefine01 MEGDNN_NAIVE
+#cmakedefine01 MEGDNN_X86
+#cmakedefine01 MEGDNN_X86_64
+#cmakedefine01 MEGDNN_64_BIT
+#cmakedefine01 MEGDNN_THREADS_512
+#cmakedefine01 MEGDNN_ENABLE_MULTI_THREADS
+
+// whether cuda is available
+#ifndef MGB_CUDA
+#define MGB_CUDA    1
+#endif
+
+
+// whether to include file/line location for assert message
+#ifndef MGB_ASSERT_LOC
+#define MGB_ASSERT_LOC          1
+#endif
+
+// whether to enable utils/debug.h and other debug methods
+#ifndef MGB_ENABLE_DEBUG_UTIL
+#define MGB_ENABLE_DEBUG_UTIL   1
+#endif
+
+// whether to enable logging
+#ifndef MGB_ENABLE_LOGGING
+#define MGB_ENABLE_LOGGING      1
+#endif
+
+// whether to enable registering opr grad functions
+#ifndef MGB_ENABLE_GRAD
+#define MGB_ENABLE_GRAD         1
+#endif
+
+// whether to include actual class name in mgb::Typeinfo object; if this is
+// disabled, mgb::serialization::OprRegistry::find_opr_by_name would not work.
+#ifndef MGB_VERBOSE_TYPEINFO_NAME
+#define MGB_VERBOSE_TYPEINFO_NAME   1
+#endif
+
+// whether to enbale configuing megbrain internals through env vars
+#ifndef MGB_ENABLE_GETENV
+#define MGB_ENABLE_GETENV       1
+#endif
+
+// whether to remove unnecessary features when used for serving
+#ifndef MGB_BUILD_SLIM_SERVING
+#define MGB_BUILD_SLIM_SERVING  0
+#endif
+
+// whether to enable exception
+#ifndef MGB_ENABLE_EXCEPTION
+#if __EXCEPTIONS
+#define MGB_ENABLE_EXCEPTION    1
+#else
+#define MGB_ENABLE_EXCEPTION    0
+#endif
+#endif
+
+// whether <thread> is available and usable
+#ifndef MGB_HAVE_THREAD
+#define MGB_HAVE_THREAD         1
+#endif
+
+// whether to trade thread safety for memory usage
+#ifndef MGB_THREAD_SAFE
+#define MGB_THREAD_SAFE MGB_HAVE_THREAD
+#endif
+
+// whether to enable JIT
+#ifndef MGB_JIT
+#define MGB_JIT     1
+#endif
+#ifndef MGB_JIT_HALIDE
+#define MGB_JIT_HALIDE 0
+#endif
+
+
+// whether to enable TensorRT support
+#ifndef MGB_ENABLE_TENSOR_RT
+#define MGB_ENABLE_TENSOR_RT    MGB_CUDA
+#endif
+
+// whether to enable fastrun profile
+#ifndef MGB_ENABLE_FASTRUN
+#define MGB_ENABLE_FASTRUN 1
+#endif
+
+
+/* ================= following are more finegrind controls ================= */
+
+// whether to enable json dumper
+#ifndef MGB_ENABLE_JSON
+#define MGB_ENABLE_JSON !MGB_BUILD_SLIM_SERVING
+#endif
+
+// whether to enable distributed communication
+#ifndef MGB_ENABLE_OPR_MM
+#define MGB_ENABLE_OPR_MM 0
+#endif
+
+/* ================= DNN related flags ================= */
+
+// whether to use mkl lib
+#ifndef MEGDNN_X86_WITH_MKL
+#define MEGDNN_X86_WITH_MKL 0
+#endif
+
+// whether to enable rtti
+#ifndef MEGDNN_ENABLE_RTTI
+#define MEGDNN_ENABLE_RTTI 1
+#endif
+
+// whether to enable mangling
+#ifndef MEGDNN_ENABLE_MANGLING
+#define MEGDNN_ENABLE_MANGLING !MEGDNN_ENABLE_RTTI
+#endif
+
+// whether to enable logging
+#ifndef MEGDNN_ENABLE_LOGGING
+#define MEGDNN_ENABLE_LOGGING MGB_ENABLE_LOGGING
+#endif
+
+// whether to enable exception
+#ifndef MEGDNN_ENABLE_EXCEPTIONS
+#define MEGDNN_ENABLE_EXCEPTIONS MGB_ENABLE_EXCEPTION
+#endif
+
+// whether to build naive
+#ifndef MEGDNN_NAIVE
+#define MEGDNN_NAIVE 0
+#endif
+
+// whether to build x86
+#ifndef MEGDNN_X86
+#define MEGDNN_X86 0
+#endif
+
+// whether to use cuda thread 512
+#ifndef MEGDNN_THREADS_512
+#define MEGDNN_THREADS_512 0
+#endif
+
+// whether to enable intra-op multi threads
+#ifndef MEGDNN_ENABLE_MULTI_THREADS
+#define MEGDNN_ENABLE_MULTI_THREADS 1
+#endif
+
+#ifndef MEGDNN_X86_WITH_OPENBLAS
+#define MEGDNN_X86_WITH_OPENBLAS 0
+#endif
+
+#ifndef MEGDNN_X86_WITH_MKL_DNN
+#define MEGDNN_X86_WITH_MKL_DNN 0
+#endif
+
+#endif // _HEADER_MGB_BUILD_CONFIG