[PTEN] Add xpu context. (#39098)

paddle/fluid/framework/pten_utils.h

@@ -86,5 +86,12 @@ struct ConvertToPtenContext<platform::CPUDeviceContext> {
   using TYPE = pten::CPUContext;
 };
 
+#ifdef PADDLE_WITH_XPU
+template <>
+struct ConvertToPtenContext<platform::XPUDeviceContext> {
+  using TYPE = pten::XPUContext;
+};
+#endif
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/operators/amp/check_finite_and_unscale_op_xpu.cc

@@ -94,11 +94,11 @@ class CheckFiniteAndUnscaleXPUKernel : public framework::OpKernel<T> {
         inverse_scale = 0.0;
       }
 
-      paddle::platform::XPUVersion version = dev_ctx.xpu_version();
+      auto version = dev_ctx.xpu_version();
       framework::Tensor float_x;
       framework::Tensor float_out;
       if (std::is_same<T, paddle::platform::float16>::value &&
-          (version == paddle::platform::XPUVersion::XPU1)) {
+          (version == pten::backends::xpu::XPUVersion::XPU1)) {
         float_x.mutable_data<MPDType>(dev_ctx.GetPlace(),
                                       x->numel() * sizeof(MPDType));
         float_out.mutable_data<MPDType>(dev_ctx.GetPlace(),

paddle/fluid/operators/dropout_op_xpu.cc

@@ -107,8 +107,8 @@ class DropoutGradXPUKernel : public framework::OpKernel<T> {
       return;
     }
 
-    paddle::platform::XPUVersion version = dev_ctx.xpu_version();
-    if (version == paddle::platform::XPUVersion::XPU1) {
+    auto version = dev_ctx.xpu_version();
+    if (version == pten::backends::xpu::XPUVersion::XPU1) {
       xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
       XPUType* mask_new = RAII_GUARD.alloc_l3_or_gm<XPUType>(mask->numel());
       float scale =

paddle/fluid/operators/reshape_op.cc

@@ -448,7 +448,8 @@ class ReshapeKernel {
 #ifdef PADDLE_WITH_XPU
     if (platform::is_xpu_place(ctx.GetPlace())) {
       auto &dev_ctx = ctx.device_context<platform::XPUDeviceContext>();
-      pten::ReshapeKernel(dev_ctx, *pt_x.get(), pt_scalar_shape, pt_out);
+      pten::ReshapeKernel(static_cast<const pten::XPUContext &>(dev_ctx),
+                          *pt_x.get(), pt_scalar_shape, pt_out);
     }
 #endif
     // non-inplace need move all result from pt_out to out, inplace need set
@@ -485,7 +486,8 @@ class ReshapeGradKernel {
 #ifdef PADDLE_WITH_XPU
     if (platform::is_xpu_place(ctx.GetPlace())) {
       auto &dev_ctx = ctx.device_context<platform::XPUDeviceContext>();
-      pten::ReshapeGradKernel(dev_ctx, *pt_d_out.get(), pt_d_x.get());
+      pten::ReshapeGradKernel(static_cast<const pten::XPUContext &>(dev_ctx),
+                              *pt_d_out.get(), pt_d_x.get());
     }
 #endif
   }
@@ -516,7 +518,9 @@ class ReshapeDoubleGradKernel {
 #ifdef PADDLE_WITH_XPU
     if (platform::is_xpu_place(ctx.GetPlace())) {
       auto &dev_ctx = ctx.device_context<platform::XPUDeviceContext>();
-      pten::ReshapeDoubleGradKernel(dev_ctx, *pt_dd_x.get(), pt_dd_out.get());
+      pten::ReshapeDoubleGradKernel(
+          static_cast<const pten::XPUContext &>(dev_ctx), *pt_dd_x.get(),
+          pt_dd_out.get());
     }
 #endif
   }

paddle/fluid/operators/softmax_op_xpu.cc

@@ -45,8 +45,8 @@ class SoftmaxXPUKernel : public framework::OpKernel<T> {
     auto& dev_ctx = context.template device_context<DeviceContext>();
 
     int r = XPU_SUCCESS;
-    paddle::platform::XPUVersion version = dev_ctx.xpu_version();
-    if (version == paddle::platform::XPUVersion::XPU1) {
+    auto version = dev_ctx.xpu_version();
+    if (version == pten::backends::xpu::XPUVersion::XPU1) {
       xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
       XPUType* clip_x_data_l3 = RAII_GUARD.alloc_l3_or_gm<XPUType>(x->numel());
       r = xpu::clip_v2(dev_ctx.x_context(),

paddle/fluid/platform/CMakeLists.txt

@@ -121,6 +121,9 @@ cc_library(cudnn_workspace_helper SRCS cudnn_workspace_helper.cc DEPS boost)
 cc_library(device_context SRCS device_context.cc init.cc DEPS simple_threadpool malloc xxhash ${STREAM_CALLBACK_DEPS}
     place pten_place eigen3 stringpiece cpu_helper cpu_info framework_proto ${IPU_CTX_DEPS} ${GPU_CTX_DEPS} ${NPU_CTX_DEPS} ${MKLDNN_CTX_DEPS}
     ${dgc_deps} dlpack cudnn_workspace_helper ${XPU_CTX_DEPS} ${MLU_CTX_DEPS} cpu_context)
+if(WITH_XPU)
+  target_link_libraries(device_context xpu_context)
+endif()
 
 cc_library(collective_helper SRCS collective_helper.cc gen_comm_id_helper.cc DEPS framework_proto device_context enforce)
 if(WITH_ASCEND_CL)

paddle/fluid/platform/device/xpu/CMakeLists.txt

@@ -4,7 +4,7 @@ endif()
 
 set(XPU_CTX_DEPS xpulib ssl crypto rt z resolv dl)
 
-cc_library(xpu_info SRCS xpu_info.cc DEPS gflags glog enforce xpulib device_context place)
+cc_library(xpu_info SRCS xpu_info.cc DEPS gflags glog enforce xpulib device_context place pten_xpu_info)
 cc_library(xpu_op_list SRCS xpu_op_list.cc DEPS gflags glog enforce xpulib device_context)
 
 add_subdirectory(tests)

paddle/fluid/platform/device/xpu/enforce_xpu.h

@@ -15,177 +15,36 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/fluid/platform/device/xpu/xpu_header.h"
-#include "paddle/fluid/platform/enforce.h"
-#include "xpu/bkcl.h"
+
+#include "paddle/pten/backends/xpu/enforce_xpu.h"
 
 namespace paddle {
 namespace platform {
 
 // Note: XPU runtime api return int, not XPUError_t
 inline const char* xpuGetErrorString(int stat) {
-  switch (stat) {
-    case XPU_SUCCESS:
-      return "Success";
-    case XPUERR_INVALID_DEVICE:
-      return "Invalid XPU device";
-    case XPUERR_UNINIT:
-      return "XPU runtime not properly inited";
-    case XPUERR_NOMEM:
-      return "Device memory not enough";
-    case XPUERR_NOCPUMEM:
-      return "CPU memory not enough";
-    case XPUERR_INVALID_PARAM:
-      return "Invalid parameter";
-    case XPUERR_NOXPUFUNC:
-      return "Cannot get XPU Func";
-    case XPUERR_LDSO:
-      return "Error loading dynamic library";
-    case XPUERR_LDSYM:
-      return "Error loading func from dynamic library";
-    case XPUERR_SIMULATOR:
-      return "Error from XPU Simulator";
-    case XPUERR_NOSUPPORT:
-      return "Operation not supported";
-    case XPUERR_ABNORMAL:
-      return "Device abnormal due to previous error";
-    case XPUERR_KEXCEPTION:
-      return "Exception in kernel execution";
-    case XPUERR_TIMEOUT:
-      return "Kernel execution timed out";
-    case XPUERR_BUSY:
-      return "Resource busy";
-    case XPUERR_USEAFCLOSE:
-      return "Use a stream after closed";
-    case XPUERR_UCECC:
-      return "Uncorrectable ECC";
-    case XPUERR_OVERHEAT:
-      return "Overheat";
-    case XPUERR_UNEXPECT:
-      return "Execution error, reach unexpected control flow";
-    case XPUERR_DEVRESET:
-      return "Device is being reset, try again later";
-    case XPUERR_HWEXCEPTION:
-      return "Hardware module exception";
-    case XPUERR_HBM_INIT:
-      return "Error init HBM";
-    case XPUERR_DEVINIT:
-      return "Error init device";
-    case XPUERR_PEERRESET:
-      return "Device is being reset, try again later";
-    case XPUERR_MAXDEV:
-      return "Device count exceed limit";
-    case XPUERR_NOIOC:
-      return "Unknown IOCTL command";
-    case XPUERR_DMATIMEOUT:
-      return "DMA timed out, a reboot maybe needed";
-    case XPUERR_DMAABORT:
-      return "DMA aborted due to error, possibly wrong address or hardware "
-             "state";
-    case XPUERR_MCUUNINIT:
-      return "Firmware not initialized";
-    case XPUERR_OLDFW:
-      return "Firmware version too old (<15), please update.";
-    case XPUERR_PCIE:
-      return "Error in PCIE";
-    case XPUERR_FAULT:
-      return "Error copy between kernel and user space";
-    case XPUERR_INTERRUPTED:
-      return "Execution interrupted by user";
-    default:
-      return "unkonwn error";
-  }
+  return pten::backends::xpu::xpuGetErrorString(stat);
 }
 
 inline const char* bkclGetErrorString(BKCLResult_t stat) {
-  switch (stat) {
-    case BKCL_SUCCESS:
-      return "BKCL_SUCCESS";
-    case BKCL_INVALID_ARGUMENT:
-      return "BKCL_INVALID_ARGUMENT";
-    case BKCL_RUNTIME_ERROR:
-      return "BKCL_RUNTIME_ERROR";
-    case BKCL_SYSTEM_ERROR:
-      return "BKCL_SYSTEM_ERROR";
-    case BKCL_INTERNAL_ERROR:
-      return "BKCL_INTERNAL_ERROR";
-    default:
-      return "Unknown BKCL status";
-  }
+  return pten::backends::xpu::bkclGetErrorString(stat);
 }
 
 inline const char* xdnnGetErrorString(int stat) {
-  switch (stat) {
-    case xpu::Error_t::SUCCESS:
-      return "XDNN_SUCCESS";
-    case xpu::Error_t::INVALID_PARAM:
-      return "XDNN_INVALID_PARAM";
-    case xpu::Error_t::RUNTIME_ERROR:
-      return "XDNN_RUNTIME_ERROR";
-    case xpu::Error_t::NO_ENOUGH_WORKSPACE:
-      return "XDNN_NO_ENOUGH_WORKSPACE";
-    case xpu::Error_t::NOT_IMPLEMENT:
-      return "XDNN_NOT_IMPLEMENT";
-    default:
-      return "Unknown XDNN status";
-  }
+  return pten::backends::xpu::xdnnGetErrorString(stat);
 }
 
 inline std::string build_xpu_error_msg(int stat) {
-  std::string msg("XPU Error <" + std::to_string(stat) + ">, ");
-  return msg + xpuGetErrorString(stat) + " ";
+  return pten::backends::xpu::build_xpu_error_msg(stat);
 }
 
 inline std::string build_xpu_error_msg(BKCLResult_t stat) {
-  std::string msg("BKCL Error, ");
-  return msg + bkclGetErrorString(stat) + " ";
+  return pten::backends::xpu::build_xpu_error_msg(stat);
 }
 
 inline std::string build_xpu_xdnn_error_msg(int stat, std::string msg) {
-  return msg + " XDNN Error, " + xdnnGetErrorString(stat) + " ";
+  return pten::backends::xpu::build_xpu_xdnn_error_msg(stat, msg);
 }
 
-namespace details {
-
-template <typename T>
-struct ExternalApiType {};
-
-#define DEFINE_EXTERNAL_API_TYPE(type, success_value) \
-  template <>                                         \
-  struct ExternalApiType<type> {                      \
-    using Type = type;                                \
-    static constexpr Type kSuccess = success_value;   \
-  }
-
-DEFINE_EXTERNAL_API_TYPE(int, XPU_SUCCESS);
-DEFINE_EXTERNAL_API_TYPE(BKCLResult_t, BKCL_SUCCESS);
-
-#undef DEFINE_EXTERNAL_API_TYPE
-
-}  // namespace details
-
-#define PADDLE_ENFORCE_XPU_SUCCESS(COND)                      \
-  do {                                                        \
-    auto __cond__ = (COND);                                   \
-    using __XPU_STATUS_TYPE__ = decltype(__cond__);           \
-    constexpr auto __success_type__ =                         \
-        ::paddle::platform::details::ExternalApiType<         \
-            __XPU_STATUS_TYPE__>::kSuccess;                   \
-    if (UNLIKELY(__cond__ != __success_type__)) {             \
-      auto __summary__ = paddle::platform::errors::External(  \
-          ::paddle::platform::build_xpu_error_msg(__cond__)); \
-      __THROW_ERROR_INTERNAL__(__summary__);                  \
-    }                                                         \
-  } while (0)
-
-#define PADDLE_ENFORCE_XDNN_SUCCESS(COND, MSG)                          \
-  do {                                                                  \
-    auto __cond__ = (COND);                                             \
-    if (UNLIKELY(__cond__ != xpu::Error_t::SUCCESS)) {                  \
-      auto __summary__ = paddle::platform::errors::External(            \
-          ::paddle::platform::build_xpu_xdnn_error_msg(__cond__, MSG)); \
-      __THROW_ERROR_INTERNAL__(__summary__);                            \
-    }                                                                   \
-  } while (0)
-
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/device/xpu/xpu_header.h

@@ -15,42 +15,5 @@ limitations under the License. */
 #pragma once
 
 #ifdef PADDLE_WITH_XPU
-#include <map>
-#include <string>
-#include <unordered_map>
-
-#include "paddle/fluid/platform/bfloat16.h"
-#include "paddle/fluid/platform/enforce.h"
-#include "paddle/fluid/platform/float16.h"
-
-#include "xpu/runtime.h"
-#include "xpu/runtime_ex.h"
-#include "xpu/xdnn.h"
-
-namespace xpu = baidu::xpu::api;
-
-static std::map<int, std::string> XPUAPIErrorMsg = {
-    {xpu::Error_t::SUCCESS, "xpu api success"},
-    {xpu::Error_t::INVALID_PARAM, "xpu api invalid param"},
-    {xpu::Error_t::RUNTIME_ERROR, "xpu api runtime error"},
-    {xpu::Error_t::NO_ENOUGH_WORKSPACE, "xpu api no enough workspace"}};
-
-template <typename T>
-class XPUTypeTrait {
- public:
-  using Type = T;
-};
-
-template <>
-class XPUTypeTrait<paddle::platform::float16> {
- public:
-  using Type = float16;
-};
-
-template <>
-class XPUTypeTrait<paddle::platform::bfloat16> {
- public:
-  using Type = bfloat16;
-};
-
+#include "paddle/pten/backends/xpu/xpu_header.h"
 #endif