Merge branch 'develop' into support_acti

PaddlePaddle · May 11, 2023 · 7463161 · 7463161
2 parents bcc59de + 04e5e7b
commit 7463161
Show file tree

Hide file tree

Showing 102 changed files with 2,212 additions and 955 deletions.
diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake
@@ -23,7 +23,8 @@ set(CBLAS_TAG v0.3.7)
 # https://github.com/PaddlePaddle/Paddle/pull/52983
 if(UNIX
    AND NOT APPLE
-   AND NOT WITH_ROCM)
+   AND NOT WITH_ROCM
+   AND NOT WITH_XPU)
   set(CBLAS_TAG v0.3.18)
 endif()
 

diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
@@ -9,7 +9,7 @@ set(XPU_RT_LIB_NAME "libxpurt.so")
 set(XPU_XFT_LIB_NAME "libxft.so")
 
 set(XPU_BASE_DATE "20230427")
-set(XPU_XCCL_BASE_VERSION "1.0.13")
+set(XPU_XCCL_BASE_VERSION "1.0.49.2")
 set(XPU_XFT_BASE_VERSION "latest")
 
 if(NOT DEFINED XPU_BASE_URL)
@@ -30,35 +30,41 @@ if(NOT XPU_XFT_BASE_URL)
   )
 endif()
 
+if(WITH_XCCL_RDMA)
+  set(XPU_XCCL_PREFIX "xccl_rdma")
+else()
+  set(XPU_XCCL_PREFIX "xccl_socket")
+endif()
+
 if(WITH_AARCH64)
   set(XPU_XRE_DIR_NAME "xre-kylin_aarch64")
   set(XPU_XDNN_DIR_NAME "xdnn-kylin_aarch64")
-  set(XPU_XCCL_DIR_NAME "xccl-kylin_aarch64")
+  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}-kylin_aarch64")
   set(XPU_XFT_DIR_NAME "") # TODO: xft has no kylin output at now.
 elseif(WITH_SUNWAY)
   set(XPU_XRE_DIR_NAME "xre-deepin_sw6_64")
   set(XPU_XDNN_DIR_NAME "xdnn-deepin_sw6_64")
-  set(XPU_XCCL_DIR_NAME "xccl-deepin_sw6_64")
+  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}-deepin_sw6_64")
   set(XPU_XFT_DIR_NAME "") # TODO: xft has no deepin output at now.
 elseif(WITH_BDCENTOS)
   set(XPU_XRE_DIR_NAME "xre-bdcentos_x86_64")
   set(XPU_XDNN_DIR_NAME "xdnn-bdcentos_x86_64")
-  set(XPU_XCCL_DIR_NAME "xccl-bdcentos_x86_64")
+  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}-bdcentos_x86_64")
   set(XPU_XFT_DIR_NAME "xft_bdcentos6u3_x86_64_gcc82")
 elseif(WITH_UBUNTU)
   set(XPU_XRE_DIR_NAME "xre-ubuntu_x86_64")
   set(XPU_XDNN_DIR_NAME "xdnn-ubuntu_x86_64")
-  set(XPU_XCCL_DIR_NAME "xccl-ubuntu_x86_64")
+  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}-ubuntu_x86_64")
   set(XPU_XFT_DIR_NAME "xft_ubuntu1604_x86_64")
 elseif(WITH_CENTOS)
   set(XPU_XRE_DIR_NAME "xre-centos7_x86_64")
   set(XPU_XDNN_DIR_NAME "xdnn-centos7_x86_64")
-  set(XPU_XCCL_DIR_NAME "xccl-bdcentos_x86_64")
+  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}-bdcentos_x86_64")
   set(XPU_XFT_DIR_NAME "xft_bdcentos6u3_x86_64_gcc82")
 else()
   set(XPU_XRE_DIR_NAME "xre-ubuntu_x86_64")
   set(XPU_XDNN_DIR_NAME "xdnn-ubuntu_x86_64")
-  set(XPU_XCCL_DIR_NAME "xccl-ubuntu_x86_64")
+  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}-ubuntu_x86_64")
   set(XPU_XFT_DIR_NAME "xft_ubuntu1604_x86_64")
 endif()
 
@@ -75,9 +81,6 @@ set(XPU_XFT_URL "${XPU_XFT_BASE_URL}/${XPU_XFT_DIR_NAME}.tar.gz")
 set(XPU_PACK_DEPENCE_URL
     "https://baidu-kunlun-public.su.bcebos.com/paddle_depence/pack_paddle_depence.sh"
     CACHE STRING "" FORCE)
-set(XPU_CHECK_DEPENCE_URL
-    "https://baidu-kunlun-public.su.bcebos.com/paddle_depence/check_xpu_dependence.sh"
-    CACHE STRING "" FORCE)
 set(XPU_XFT_GET_DEPENCE_URL
     "https://baidu-kunlun-public.su.bcebos.com/paddle_depence/get_xft_dependence.sh"
     CACHE STRING "" FORCE)
@@ -115,8 +118,8 @@ ExternalProject_Add(
   PREFIX ${SNAPPY_PREFIX_DIR}
   DOWNLOAD_DIR ${XPU_DOWNLOAD_DIR}
   DOWNLOAD_COMMAND
-    wget ${XPU_CHECK_DEPENCE_URL} && bash check_xpu_dependence.sh
-    ${XPU_BASE_URL} ${XPU_XCCL_BASE_URL} && wget ${XPU_PACK_DEPENCE_URL} && bash
+    bash ${CMAKE_SOURCE_DIR}/tools/xpu/check_xpu_dependence.sh ${XPU_BASE_URL}
+    ${XPU_XCCL_BASE_URL} && wget ${XPU_PACK_DEPENCE_URL} && bash
     pack_paddle_depence.sh ${XPU_XRE_URL} ${XPU_XRE_DIR_NAME} ${XPU_XDNN_URL}
     ${XPU_XDNN_DIR_NAME} ${XPU_XCCL_URL} ${XPU_XCCL_DIR_NAME} && wget
     ${XPU_XFT_GET_DEPENCE_URL} && bash get_xft_dependence.sh ${XPU_XFT_URL}

diff --git a/paddle/fluid/distributed/collective/process_group_bkcl.cc b/paddle/fluid/distributed/collective/process_group_bkcl.cc
@@ -115,7 +115,13 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Recv(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_recv";
+        VLOG(3) << "calling bkcl_recv"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", src_rank: " << src_rank << ", numel: " << output->numel()
+                << ", dtype: " << output->type() << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << use_calc_stream;
         int r = bkcl_recv(comm,
                           output->data(),
                           output->numel(),
@@ -148,7 +154,14 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Send(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_send";
+        VLOG(3) << "calling bkcl_send"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", dst_rank: " << dst_rank
+                << ", input numel: " << input.numel()
+                << ", dtype: " << input.type() << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << use_calc_stream;
         int r = bkcl_send(comm,
                           input.data(),
                           input.numel(),
@@ -276,7 +289,14 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllReduce(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_all_reduce";
+        VLOG(3) << "calling bkcl_all_reduce"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", numel: " << input.numel() << ", dtype: " << input.type()
+                << ", reduce_type: " << ToBKCLRedType(opts.reduce_op)
+                << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << use_calc_stream;
         int r =
             bkcl_all_reduce(comm,
                             input.data(),
@@ -307,7 +327,13 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Broadcast(
           BKCLContext_t comm,
           const XPUStream& stream) {
         int root = opts.source_rank + opts.source_root;
-        VLOG(3) << "bkcl_broadcast";
+        VLOG(3) << "calling bkcl_broadcast"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", root: " << root << ", numel: " << input.numel()
+                << ", dtype: " << input.type() << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << use_calc_stream;
         int r =
             bkcl_broadcast(comm,
                            input.data(),
@@ -346,7 +372,13 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllGather(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_all_gather";
+        VLOG(3) << "calling bkcl_all_gather"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", numel: " << in_tensor_maybe_partial.numel()
+                << ", dtype: " << input.type() << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << use_calc_stream;
         int r =
             bkcl_all_gather(comm,
                             in_tensor_maybe_partial.data(),
@@ -375,7 +407,15 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Reduce(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_reduce";
+        VLOG(3) << "calling bkcl_reduce"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", root: " << opts.root_rank << ", numel: " << input.numel()
+                << ", dtype: " << input.type()
+                << ", reduce_type: " << ToBKCLRedType(opts.reduce_op)
+                << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << use_calc_stream;
         int r = bkcl_reduce(comm,
                             input.data(),
                             output->data(),
@@ -405,7 +445,14 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::ReduceScatter(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_reduce_scatter";
+        VLOG(3) << "calling bkcl_reduce_scatter"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", numel: " << output->numel() << ", dtype: " << input.type()
+                << ", reduce_type: " << ToBKCLRedType(opts.reduce_op)
+                << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << use_calc_stream;
         int r = bkcl_reduce_scatter(
             comm,
             input.data(),
@@ -491,8 +538,13 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllReduce(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_all_reduce";
-
+        VLOG(3) << "calling bkcl_all_reduce"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", numel: " << input.numel() << ", dtype: " << input.type()
+                << ", reduce_type: " << ToBKCLRedType(opts.reduce_op)
+                << ", sync_op: " << true << ", use_calc_stream: " << false;
         int r =
             bkcl_all_reduce(comm,
                             input.data(),
@@ -535,7 +587,13 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllReduce(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_all_reduce";
+        VLOG(3) << "calling bkcl_all_reduce"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", numel: " << input.numel() << ", dtype: " << input.type()
+                << ", reduce_type: " << ToBKCLRedType(opts.reduce_op)
+                << ", sync_op: " << sync_op << ", use_calc_stream: " << false;
         int r =
             bkcl_all_reduce(comm,
                             input.data(),
@@ -580,7 +638,13 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Broadcast(
           const XPUStream& stream) {
         const auto root =
             opts.source_rank * in_tensors.size() + opts.source_root;
-        VLOG(3) << "bkcl_broadcast";
+        VLOG(3) << "calling bkcl_broadcast"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", root: " << root << ", numel: " << input.numel()
+                << ", dtype: " << input.type() << ", sync_op: " << true
+                << ", use_calc_stream: " << false;
         int r =
             bkcl_broadcast(comm,
                            input.data(),
@@ -626,7 +690,13 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::Broadcast(
           const XPUStream& stream) {
         const auto root =
             opts.source_rank * in_tensors.size() + opts.source_root;
-        VLOG(3) << "bkcl_broadcast";
+        VLOG(3) << "calling bkcl_broadcast"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", root: " << root << ", numel: " << input.numel()
+                << ", dtype: " << input.type() << ", sync_op: " << sync_op
+                << ", use_calc_stream: " << false;
         int r =
             bkcl_broadcast(comm,
                            input.data(),
@@ -671,7 +741,12 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllGather(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_all_gather";
+        VLOG(3) << "calling bkcl_all_gather"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", numel: " << input.numel() << ", dtype: " << input.type()
+                << ", sync_op: " << true << ", use_calc_stream: " << false;
         int r =
             bkcl_all_gather(comm,
                             input.data(),
@@ -712,7 +787,12 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupBKCL::AllGather(
           const phi::DenseTensor& input,
           BKCLContext_t comm,
           const XPUStream& stream) {
-        VLOG(3) << "bkcl_all_gather";
+        VLOG(3) << "calling bkcl_all_gather"
+                << ", rank_id: " << platform::GetBKCLRankID(comm)
+                << ", dev_id: " << platform::GetBKCLDevID(comm)
+                << ", nranks: " << platform::GetBKCLNRanks(comm)
+                << ", numel: " << input.numel() << ", dtype: " << input.type()
+                << ", sync_op: " << sync_op << ", use_calc_stream: " << false;
         int r =
             bkcl_all_gather(comm,
                             input.data(),

diff --git a/paddle/fluid/distributed/collective/process_group_custom.cc b/paddle/fluid/distributed/collective/process_group_custom.cc
@@ -125,12 +125,14 @@ void ProcessGroupCustom::BroadcastUniqueCustomID(
     std::vector<phi::ccl::CCLRootId>& ccl_ids) {  // NOLINT
   if (rank_ == 0) {
     for (size_t i = 0; i < ccl_ids.size(); i++) {
-      auto key = "ProcessGroupCustom/ccl_ids/" + std::to_string(i);
+      auto key = "ProcessGroupCustom/ccl_ids/" + std::to_string(gid_) + "/" +
+                 std::to_string(i);
       store_->set(key, ccl_ids[i]);
     }
   } else {
     for (size_t i = 0; i < ccl_ids.size(); i++) {
-      auto key = "ProcessGroupCustom/ccl_ids/" + std::to_string(i);
+      auto key = "ProcessGroupCustom/ccl_ids/" + std::to_string(gid_) + "/" +
+                 std::to_string(i);
       ccl_ids[i] = store_->get(key);
     }
   }

diff --git a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
@@ -68,7 +68,6 @@
     "matmul_double_grad",
     "tanh_double_grad",
     "add_double_grad",
-    "multiply_double_grad",
     "subtract_double_grad",
 ]
 

diff --git a/paddle/fluid/framework/new_executor/interpreter/stream_analyzer.cc b/paddle/fluid/framework/new_executor/interpreter/stream_analyzer.cc
@@ -150,7 +150,7 @@ DeviceContext* StreamAnalyzer::ParseDeviceContext(
 
   DeviceContext* dev_ctx = nullptr;
 
-  // only gpu needs update. xpu not need, because xpu memcpy op kernel is
+  // only gpu need update. xpu not need, because xpu memcpy op kernel is
   // synchronous.
   if (platform::is_gpu_place(place_) || platform::is_custom_place(place_)) {
     VLOG(6) << "Parse DeviceContext for " << op_type