Merge pull request #13 from HuangShiqing/paddlebox

abacus-aibox-841 update the xtdk version
jack603047588 · Oct 31, 2023 · b511644 · b511644
2 parents a5a42d2 + 92c657f
commit b511644
Show file tree

Hide file tree

Showing 13 changed files with 47 additions and 36 deletions.
diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
@@ -111,6 +111,9 @@ if (WITH_BOX_PS OR WITH_XPU_KP)
     set(SCALOPUS_URL
         "https://klx-sdk-release-public.su.bcebos.com/xdnn_train/dev/paddlebox/20230306/scalopus.tar.gz"
         CACHE STRING "" FORCE)
+    set(XPU_XTDK_URL
+        "https://klx-sdk-release-public.su.bcebos.com/xdnn_train/dev/paddlebox/xtdk/20230914/output.tar.gz"
+        CACHE STRING "" FORCE)
 endif()
 
 set(SNAPPY_PREFIX_DIR "${THIRD_PARTY_PATH}/xpu")
@@ -141,7 +144,8 @@ ExternalProject_Add(
     wget --no-check-certificate ${XPU_PACK_DEPENCE_URL} -O pack_paddle_box_depence.sh && bash pack_paddle_box_depence.sh ${XPU_XRE_URL}
     ${XPU_XRE_DIR_NAME} ${XPU_XDNN_URL} ${XPU_XDNN_DIR_NAME} ${XPU_XCCL_URL}
     ${XPU_XCCL_DIR_NAME} ${XPU_XCTR_URL} ${XPU_XCTR_DIR_NAME} &&
-    wget --no-check-certificate ${SCALOPUS_URL} && tar zxvf scalopus.tar.gz
+    wget --no-check-certificate ${SCALOPUS_URL} && tar zxvf scalopus.tar.gz &&
+    wget --no-check-certificate ${XPU_XTDK_URL} && tar zxvf output.tar.gz
   DOWNLOAD_NO_PROGRESS 1
   UPDATE_COMMAND ""
   CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XPU_INSTALL_ROOT}

diff --git a/cmake/xpu_kp.cmake b/cmake/xpu_kp.cmake
@@ -35,13 +35,15 @@ set(CMAKE_SHARED_LINKER_FLAGS "${LINK_FLAGS}")
 
 if(NOT XPU_TOOLCHAIN)
   #set(XPU_TOOLCHAIN /workspace/output/XTDK-ubuntu_x86_64)
-  set(XPU_TOOLCHAIN /home/work/xtdk/XTDK-centos7_GCC8u3)
+  set(XPU_TOOLCHAIN ${PROJECT_BINARY_DIR}/third_party/xpu/src/extern_xpu/output/xtdk-centos7_x86_64_gcc8u3)
+  # set(XPU_TOOLCHAIN /home/work/xtdk/XTDK-centos7_GCC8u3)
   #set(XPU_TOOLCHAIN /home/work/xtdk/XTDK-bdcentos_x86_64)
   #set(XPU_TOOLCHAIN /home/work/xtdk/XTDK-centos_GCCABI1_x86_64)
   get_filename_component(XPU_TOOLCHAIN ${XPU_TOOLCHAIN} REALPATH)
-endif()
-if(NOT IS_DIRECTORY ${XPU_TOOLCHAIN})
-  message(FATAL_ERROR "Directory ${XPU_TOOLCHAIN} not found!")
+else()
+  if(NOT IS_DIRECTORY ${XPU_TOOLCHAIN})
+    message(FATAL_ERROR "Directory ${XPU_TOOLCHAIN} not found!")
+  endif()
 endif()
 message(STATUS "Build with XPU_TOOLCHAIN=" ${XPU_TOOLCHAIN})
 set(XPU_CLANG ${XPU_TOOLCHAIN}/bin/clang++)
@@ -196,7 +198,7 @@ macro(compile_kernel COMPILE_ARGS)
     COMMAND ${CMAKE_COMMAND} -E make_directory kernel_build
     COMMAND
       ${XPU_CLANG} --sysroot=${CXX_DIR} -std=c++11 ${ABI_VERSION} ${OPT_LEVEL}
-      -fno-builtin -mcpu=xpu2 -fPIC ${XPU_CXX_DEFINES} ${XPU_CXX_FLAGS}
+      -fno-builtin --xpu-arch=xpu2 -fPIC ${XPU_CXX_DEFINES} ${XPU_CXX_FLAGS}
       ${XPU_CXX_INCLUDES} -I. -o kernel_build/${kernel_name}.bin.o.sec
       kernel_build/${kernel_name}.xpu --xpu-device-only -c -v
     COMMAND ${XTDK_DIR}/bin/xpu2-elfconv kernel_build/${kernel_name}.bin.o.sec
@@ -212,7 +214,7 @@ macro(compile_kernel COMPILE_ARGS)
     COMMAND ${CMAKE_COMMAND} -E make_directory kernel_build
     COMMAND
       ${XPU_CLANG} --sysroot=${CXX_DIR} -std=c++11 ${ABI_VERSION} ${OPT_LEVEL}
-      -fno-builtin -mcpu=xpu2 -fPIC ${XPU_CXX_DEFINES} ${XPU_CXX_FLAGS}
+      -fno-builtin --xpu-arch=xpu2 -fPIC ${XPU_CXX_DEFINES} ${XPU_CXX_FLAGS}
       ${XPU_CXX_INCLUDES} -I. -o kernel_build/${kernel_name}.host.o
       kernel_build/${kernel_name}.xpu --xpu-host-only -c -v
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}

diff --git a/paddle/fluid/framework/data_feed.kps b/paddle/fluid/framework/data_feed.kps
@@ -1,8 +1,8 @@
 #ifdef PADDLE_WITH_XPU_KP
 #include "paddle/fluid/framework/data_feed_xpu_kernel_helper.h"
-#include "xpu/kernel/cluster_header.h"
+#include "xpu/kernel/xtdk.h"
 // #include "xpu/kernel/debug.h"
-#include "xpu/kernel/math.h"
+#include "xpu/kernel/xtdk_math.h"
 
 namespace paddle {
 namespace framework {

diff --git a/paddle/fluid/framework/fleet/box_wrapper_kernel.kps b/paddle/fluid/framework/fleet/box_wrapper_kernel.kps
@@ -23,10 +23,11 @@ limitations under the License. */
 #include "paddle/fluid/platform/device/xpu/enforce_xpu.h"
 #include "paddle/fluid/framework/fleet/box_wrapper_kernel.h"
 #include "paddle/fluid/platform/device_context.h"
-#include "xpu/kernel/cluster_header.h"  // NOLINT
-//#include "xpu/kernel/debug.h"           // NOLINT
-#include "xpu/kernel/math.h"            // NOLINT
-#include "xpu/kernel/simd.h"
+
+#include "xpu/kernel/xtdk.h"  // NOLINT
+// #include "xpu/kernel/debug.h"           // NOLINT
+#include "xpu/kernel/xtdk_math.h"            // NOLINT
+#include "xpu/kernel/xtdk_simd.h"
 
 #ifdef TRACE_PROFILE
 // The producer side.

diff --git a/paddle/fluid/framework/fleet/heter_ps/hashtable.h b/paddle/fluid/framework/fleet/heter_ps/hashtable.h
@@ -38,9 +38,9 @@ limitations under the License. */
 #elif defined(__xpu__)
 #include <xpu/runtime.h>
 
-#include "xpu/kernel/cluster_header.h"
-#include "xpu/kernel/math.h"
-#include "xpu/kernel/simd.h"
+#include "xpu/kernel/xtdk.h"
+#include "xpu/kernel/xtdk_math.h"
+#include "xpu/kernel/xtdk_simd.h"
 #endif
 
 #include "paddle/fluid/framework/fleet/heter_ps/optimizer_conf.h"

diff --git a/paddle/fluid/framework/fleet/heter_ps/heter_comm_kernel.kps b/paddle/fluid/framework/fleet/heter_ps/heter_comm_kernel.kps
@@ -17,10 +17,10 @@ limitations under the License. */
 
 #if defined(PADDLE_WITH_XPU_KP)
 #include <xpu/runtime.h>
-#include "xpu/kernel/cluster_header.h"
+#include "xpu/kernel/xtdk.h"
 //#include "xpu/kernel/debug.h"  // NOLINT
-#include "xpu/kernel/math.h"
-#include "xpu/kernel/simd.h"
+#include "xpu/kernel/xtdk_math.h"
+#include "xpu/kernel/xtdk_simd.h"
 #endif
 
 namespace paddle {

diff --git a/paddle/fluid/framework/fleet/ps_gpu_wrapper.kps b/paddle/fluid/framework/fleet/ps_gpu_wrapper.kps
@@ -20,10 +20,10 @@ limitations under the License. */
 #include <numeric>
 #include "paddle/fluid/framework/fleet/ps_gpu_wrapper.h"
 #include "paddle/fluid/framework/lod_tensor.h"
-#include "xpu/kernel/cluster_header.h"  // NOLINT
-#include "xpu/kernel/debug.h"           // NOLINT
-#include "xpu/kernel/math.h"            // NOLINT
-#include "xpu/kernel/simd.h"
+#include "xpu/kernel/xtdk.h"  // NOLINT
+// #include "xpu/kernel/debug.h"           // NOLINT
+#include "xpu/kernel/xtdk_math.h"            // NOLINT
+#include "xpu/kernel/xtdk_simd.h"
 
 namespace paddle {
 namespace framework {

diff --git a/paddle/phi/core/hostdevice.h b/paddle/phi/core/hostdevice.h
@@ -21,9 +21,9 @@
 #if defined(__xpu__)
 #include <xpu/runtime.h>
 
-#include "xpu/kernel/cluster_header.h"
+#include "xpu/kernel/xtdk.h"
 //#include "xpu/kernel/debug.h"
-#include "xpu/kernel/math.h"
+#include "xpu/kernel/xtdk_math.h"
 #endif
 
 #if (defined(__CUDACC__) || defined(__HIPCC__) || defined(__xpu__))

diff --git a/paddle/phi/kernels/funcs/activation_functor.h b/paddle/phi/kernels/funcs/activation_functor.h
@@ -2164,7 +2164,11 @@ template <typename T>
 struct CudaExpFunctor : public BaseActivationFunctor<T> {
   // exp(x) = expf(x)
   __device__ __forceinline__ T operator()(const T x) const {
+#if defined(__xpu__)
+    return static_cast<T>(exp(static_cast<float>(x)));
+#else
     return static_cast<T>(expf(static_cast<float>(x)));
+#endif
   }
 };
 

diff --git a/paddle/phi/kernels/funcs/elementwise_functor.h b/paddle/phi/kernels/funcs/elementwise_functor.h
@@ -21,7 +21,7 @@ limitations under the License. */
 #if defined(__xpu__)
 #include <xpu/runtime.h>
 
-#include "xpu/kernel/math_xpu2.h"  //pow()
+#include "xpu/kernel/xtdk_math.h"  //pow()
 #endif
 
 namespace phi {

diff --git a/paddle/phi/kernels/primitive/compute_primitives_xpu2.h b/paddle/phi/kernels/primitive/compute_primitives_xpu2.h
@@ -14,10 +14,10 @@
 
 #pragma once
 #include "paddle/phi/common/float16.h"
-#include "xpu/kernel/cluster_header.h"
-#include "xpu/kernel/debug.h"
-#include "xpu/kernel/math.h"
-#include "xpu/kernel/simd_header.h"
+#include "xpu/kernel/xtdk.h"
+// #include "xpu/kernel/debug.h"
+#include "xpu/kernel/xtdk_math.h"
+#include "xpu/kernel/xtdk_simd.h"
 
 namespace phi {
 namespace kps {

diff --git a/paddle/phi/kernels/primitive/datamover_primitives_xpu2.h b/paddle/phi/kernels/primitive/datamover_primitives_xpu2.h
@@ -13,9 +13,9 @@
 // limitations under the License.
 
 #pragma once
-#include "xpu/kernel/cluster_header.h"
-#include "xpu/kernel/debug.h"
-#include "xpu/kernel/math.h"
+#include "xpu/kernel/xtdk.h"
+// #include "xpu/kernel/debug.h"
+#include "xpu/kernel/xtdk_math.h"
 
 namespace phi {
 namespace kps {

diff --git a/paddle/phi/kernels/primitive/functor_primitives_xpu2.h b/paddle/phi/kernels/primitive/functor_primitives_xpu2.h
@@ -14,9 +14,9 @@
 
 #pragma once
 
-#include "xpu/kernel/cluster_header.h"
-#include "xpu/kernel/debug.h"
-#include "xpu/kernel/math.h"
+#include "xpu/kernel/xtdk.h"
+// #include "xpu/kernel/debug.h"
+#include "xpu/kernel/xtdk_math.h"
 
 namespace phi {
 namespace kps {