Support mips (#29943)

CMakeLists.txt

@@ -144,6 +144,7 @@ option(WITH_NCCL   "Compile PaddlePaddle with NCCL support"             ON)
 option(WITH_CRYPTO   "Compile PaddlePaddle with crypto support"         ON)
 option(WITH_ARM   "Compile PaddlePaddle with arm support"         OFF)
 option(WITH_SW   "Compile PaddlePaddle with sw support"         OFF)
+option(WITH_MIPS   "Compile PaddlePaddle with mips support"         OFF)
 option(WITH_MUSL        "Compile with musl libc instead of gblic"  OFF)
 
 # PY_VERSION
@@ -264,6 +265,11 @@ if (WITH_SW)
     add_definitions(-DPADDLE_WITH_SW)
 endif()
 
+if (WITH_MIPS)
+    set(WITH_XBYAK OFF CACHE STRING "Disable XBYAK when compiling WITH_MIPS=ON" FORCE)
+    add_definitions(-DPADDLE_WITH_MIPS)
+endif()
+
 set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
 
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")

cmake/external/openblas.cmake

@@ -19,6 +19,9 @@ SET(CBLAS_SOURCE_DIR  ${THIRD_PARTY_PATH}/openblas/src/extern_openblas)
 SET(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)
 SET(CBLAS_REPOSITORY  ${GIT_URL}/xianyi/OpenBLAS.git)
 SET(CBLAS_TAG         v0.3.7)
+if(WITH_MIPS)
+  SET(CBLAS_TAG         v0.3.13)
+endif()
 
 cache_third_party(extern_openblas
     REPOSITORY    ${CBLAS_REPOSITORY}

cmake/flags.cmake

@@ -183,7 +183,7 @@ set(GPU_COMMON_FLAGS
     -Wno-error=unused-function  # Warnings in Numpy Header.
     -Wno-error=array-bounds # Warnings in Eigen::array
 )
-if (NOT WITH_NV_JETSON AND NOT WITH_ARM AND NOT WITH_SW)
+if (NOT WITH_NV_JETSON AND NOT WITH_ARM AND NOT WITH_SW AND NOT WITH_MIPS)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
 endif()
 endif(NOT WIN32)

paddle/fluid/operators/search_compute.h

@@ -14,7 +14,8 @@ limitations under the License. */
 
 #pragma once
 
-#if !defined(PADDLE_WITH_ARM) && !defined(PADDLE_WITH_SW)
+#if !defined(PADDLE_WITH_ARM) && !defined(PADDLE_WITH_SW) && \
+    !defined(PADDLE_WITH_MIPS)
 #include <immintrin.h>
 #endif
 #include <cfloat>
@@ -74,7 +75,8 @@ void call_gemm_batched(const framework::ExecutionContext& ctx,
   }
 }
 
-#if !defined(PADDLE_WITH_ARM) && !defined(PADDLE_WITH_SW)
+#if !defined(PADDLE_WITH_ARM) && !defined(PADDLE_WITH_SW) && \
+    !defined(PADDLE_WITH_MIPS)
 
 #define __m256x __m256
 
@@ -114,7 +116,8 @@ inline void axpy(const T* x, T* y, size_t len, const T alpha) {
         _mm256_add_px(_mm256_load_px(y + jjj),
                       _mm256_mul_px(mm_alpha, _mm256_load_px(x + jjj))));
   }
-#elif defined(PADDLE_WITH_ARM) || defined(PADDLE_WITH_SW)
+#elif defined(PADDLE_WITH_ARM) || defined(PADDLE_WITH_SW) || \
+    defined(PADDLE_WITH_MIPS)
   PADDLE_THROW(platform::errors::Unimplemented("axpy is not supported"));
 #else
   lll = len & ~SSE_CUT_LEN_MASK;
@@ -143,7 +146,8 @@ inline void axpy_noadd(const T* x, T* y, size_t len, const T alpha) {
   for (jjj = 0; jjj < lll; jjj += AVX_STEP_SIZE) {
     _mm256_store_px(y + jjj, _mm256_mul_px(mm_alpha, _mm256_load_px(x + jjj)));
   }
-#elif defined(PADDLE_WITH_ARM) || defined(PADDLE_WITH_SW)
+#elif defined(PADDLE_WITH_ARM) || defined(PADDLE_WITH_SW) || \
+    defined(PADDLE_WITH_MIPS)
   PADDLE_THROW(platform::errors::Unimplemented("axpy_noadd is not supported"));
 #else
   lll = len & ~SSE_CUT_LEN_MASK;

paddle/fluid/platform/cpu_info.cc

@@ -141,7 +141,7 @@ bool MayIUse(const cpu_isa_t cpu_isa) {
     return true;
   } else {
 #if !defined(WITH_NV_JETSON) && !defined(PADDLE_WITH_ARM) && \
-    !defined(PADDLE_WITH_SW)
+    !defined(PADDLE_WITH_SW) && !defined(PADDLE_WITH_MIPS)
     int reg[4];
     cpuid(reg, 0);
     int nIds = reg[0];

paddle/fluid/platform/cpu_info.h

@@ -41,7 +41,7 @@ limitations under the License. */
 #define cpuid(reg, x) __cpuidex(reg, x, 0)
 #else
 #if !defined(WITH_NV_JETSON) && !defined(PADDLE_WITH_ARM) && \
-    !defined(PADDLE_WITH_SW)
+    !defined(PADDLE_WITH_SW) && !defined(PADDLE_WITH_MIPS)
 #include <cpuid.h>
 inline void cpuid(int reg[4], int x) {
   __cpuid_count(x, 0, reg[0], reg[1], reg[2], reg[3]);

python/setup.py.in

@@ -347,7 +347,7 @@ if '${CMAKE_BUILD_TYPE}' == 'Release':
         # The dynamic library compiled under aarch64 is greater than 64M,
         # and an oversize error will be reported when using patchelf.
         # The sw_64 not suppot patchelf, so we just disable that.
-        if platform.machine() != 'aarch64' and platform.machine() != 'sw_64':
+        if platform.machine() != 'aarch64' and platform.machine() != 'sw_64' and platform.machine() != 'mips64':
           if os.system(command) != 0:
               raise Exception("patch ${FLUID_CORE_NAME}.%s failed, command: %s" % (ext_name, command))