Merge develop

Author: jim19930609

.gitignore

@@ -7,9 +7,11 @@ paddle/fluid/op_use_default_grad_maker_DEV.spec
 paddle/fluid/op_use_default_grad_maker_PR.spec
 paddle/phi/api/backward/backward_api.h
 paddle/phi/api/include/api.h
+paddle/phi/api/include/sparse_api.h
 paddle/phi/api/lib/api.cc
 paddle/phi/api/lib/dygraph_api.*
 paddle/phi/api/lib/backward_api.cc
+paddle/phi/api/lib/sparse_api.cc
 paddle/phi/extension.h
 paddle/phi/include/*
 paddle/phi/infermeta/generated.*
@@ -49,6 +51,9 @@ tools/__pycache__
 # This file is automatically generated.
 # TODO(zhiqiang) Move this file to build directory.
 paddle/infrt/dialect/pd_ops.td
+paddle/infrt/dialect/phi/ir/phi_cpu_kernels.td
+paddle/infrt/dialect/phi/ir/phi_gpu_kernels.td
+tools/infrt/kernels.json
 paddle/infrt/dialect/pd_ops_info.h
 .lit_test_times.txt
 paddle/infrt/tests/dialect/Output

CMakeLists.txt

@@ -238,7 +238,8 @@ option(WITH_MIPS "Compile PaddlePaddle with mips support" OFF)
 option(WITH_MUSL "Compile with musl libc instead of gblic" OFF)
 option(WITH_UNITY_BUILD "Compile with UnityBuild mode" OFF)
 option(WITH_STRIP "Strip so files of Whl packages" OFF)
-option(NEW_RELEASE_CUBIN "PaddlePaddle next-level release strategy for pypi cubin package" OFF)
+option(NEW_RELEASE_PYPI "PaddlePaddle next-level release strategy for pypi cubin package" OFF)
+option(NEW_RELEASE_ALL "PaddlePaddle next-level release strategy for all arches cubin package" OFF)
 option(NEW_RELEASE_JIT "PaddlePaddle next-level release strategy for backup jit package" OFF)
 option(WITH_ASCEND_INT64 "Compile with int64 kernel for ascend NPU" OFF)
 option(WITH_POCKETFFT "Compile with pocketfft support" ON)

cmake/cuda.cmake

@@ -6,16 +6,22 @@ if(WITH_NV_JETSON)
  add_definitions(-DWITH_NV_JETSON)
  set(paddle_known_gpu_archs "53 62 72")
  set(paddle_known_gpu_archs10 "53 62 72")
-elseif(NEW_RELEASE_CUBIN)
+elseif(NEW_RELEASE_ALL)
+ message("Using New Release Strategy - All Arches Packge")
+ add_definitions(-DNEW_RELEASE_ALL)
+ set(paddle_known_gpu_archs "35 50 52 60 61 70 75 80 86")
+ set(paddle_known_gpu_archs10 "35 50 52 60 61 70 75")
+ set(paddle_known_gpu_archs11 "35 50 52 60 61 70 75 80")
+elseif(NEW_RELEASE_PYPI)
  message("Using New Release Strategy - Cubin Packge")
- add_definitions(-DNEW_RELEASE_CUBIN)
- set(paddle_known_gpu_archs "35 37 50 52 60 61 70 75 80 86")
- set(paddle_known_gpu_archs10 "50 60 70 75")
- set(paddle_known_gpu_archs11 "60 70 75 80")
+ add_definitions(-DNEW_RELEASE_PYPI)
+ set(paddle_known_gpu_archs "35 50 52 60 61 70 75 80 86")
+ set(paddle_known_gpu_archs10 "")
+ set(paddle_known_gpu_archs11 "60 61 70 75 80")
 elseif(NEW_RELEASE_JIT)
  message("Using New Release Strategy - JIT Packge")
  add_definitions(-DNEW_RELEASE_JIT)
- set(paddle_known_gpu_archs "35 37 50 52 60 61 70 75 80 86")
+ set(paddle_known_gpu_archs "35 50 52 60 61 70 75 80 86")
  set(paddle_known_gpu_archs10 "35 50 60 70 75")
  set(paddle_known_gpu_archs11 "35 50 60 70 75 80")
 else()
@@ -148,7 +154,7 @@ function(select_nvcc_arch_flags out_variable)
 
  # remove dots and convert to lists
  string(REGEX REPLACE "\\." "" cuda_arch_bin "${cuda_arch_bin}")
- string(REGEX REPLACE "\\." "" cuda_arch_ptx "${CUDA_ARCH_PTX}")
+ string(REGEX REPLACE "\\." "" cuda_arch_ptx "${cuda_arch_ptx}")
  string(REGEX MATCHALL "[0-9()]+" cuda_arch_bin "${cuda_arch_bin}")
  string(REGEX MATCHALL "[0-9]+" cuda_arch_ptx "${cuda_arch_ptx}")
 

paddle/fluid/distributed/collective/CMakeLists.txt

@@ -1,4 +1,7 @@
 cc_library(processgroup SRCS ProcessGroup.cc DEPS phi phi_api eager_api)
+if (WITH_DISTRIBUTE)
+ cc_library(processgroup_gloo SRCS ProcessGroupGloo.cc DEPS phi phi_api eager_api gloo_wrapper)
+endif()
 cc_library(eager_reducer SRCS reducer.cc DEPS eager_api processgroup)
 
 if(WITH_NCCL)

paddle/fluid/distributed/collective/ProcessGroup.h

@@ -117,6 +117,35 @@ class ProcessGroup {
  "ProcessGroup%s does not support receive", GetBackendName()));
  }
 
+ virtual std::shared_ptr<ProcessGroup::Task> AllGather(
+ std::vector<Tensor>& in_tensors /* tensors */, // NOLINT
+ std::vector<Tensor>& out_tensors /* tensors */) { // NOLINT
+ PADDLE_THROW(platform::errors::InvalidArgument(
+ "ProcessGroup%s does not support AllGather", GetBackendName()));
+ }
+
+ virtual std::shared_ptr<ProcessGroup::Task> AllToAll(
+ std::vector<Tensor>& in /* tensors */, // NOLINT
+ std::vector<Tensor>& out /* tensors */) { // NOLINT
+ PADDLE_THROW(platform::errors::InvalidArgument(
+ "ProcessGroup%s does not support AllToAll", GetBackendName()));
+ }
+
+ virtual std::shared_ptr<ProcessGroup::Task> Reduce(
+ std::vector<Tensor>& tensors /* tensors */, // NOLINT
+ const ReduceOptions& opts) { // NOLINT
+ PADDLE_THROW(platform::errors::InvalidArgument(
+ "ProcessGroup%s does not support Reduce", GetBackendName()));
+ }
+
+ virtual std::shared_ptr<ProcessGroup::Task> Scatter(
+ std::vector<Tensor>& in_tensors /* tensors */, // NOLINT
+ std::vector<Tensor>& out_tensors /* tensors */, // NOLINT
+ const ScatterOptions&) { // NOLINT
+ PADDLE_THROW(platform::errors::InvalidArgument(
+ "ProcessGroup%s does not support Scatter", GetBackendName()));
+ }
+
  protected:
  const int rank_;
  const int size_;