Autotune the workspace_size_limit in conv. (#40338)

* Using the maximum workspace_size of all alogirhms to limit the workspace size in exhaustive search mode.

* Use the system cudaMalloc and cudaFree to allocate workspace during searching.

* Enable switch of two kind of workspace setting methods.
Co-authored-by: NLiu Yiqun <liuyiqun01@baidu.com>

paddle/fluid/framework/conv_search_cache.h

@@ -16,7 +16,6 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator_kernel_configs.h"
-
 #include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
 
 namespace paddle {

paddle/fluid/operators/conv_base_helper.h

+/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <memory>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/conv_search_cache.h"
+#include "paddle/fluid/operators/conv_cudnn_op_cache.h"
+#include "paddle/phi/backends/gpu/gpu_context.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using DataLayout = platform::DataLayout;
+using framework::AlgorithmsCache;
+using framework::ConvSearchCache;
+
+template <typename T>
+using ScalingParamType = typename platform::CudnnDataType<T>::ScalingParamType;
+
+// As the basic for SearchAlgorithm struct.
+template <typename PerfT>
+struct SearchAlgorithm {};
+
+// As the container of searchAlgorithm::Find() result.
+template <typename AlgoT>
+struct SearchResult {
+ public:
+  AlgoT algo = static_cast<AlgoT>(0);
+  float time = -1.f;
+  size_t workspace_size = 0;
+};
+
+// As the container of conv relevant descriptors.
+template <typename HandleT, typename DataT>
+struct ConvArgsBase {
+  HandleT handle;
+  platform::TensorDescriptor idesc, odesc;
+  platform::FilterDescriptor wdesc;
+  platform::ConvolutionDescriptor cdesc;
+  const framework::Tensor *x, *w, *o;
+  DataT cudnn_dtype;
+
+  // strides
+  std::vector<int> s;
+  // paddings
+  std::vector<int> p;
+  // dilations
+  std::vector<int> d;
+
+  ConvArgsBase(const framework::Tensor* x, const framework::Tensor* w,
+               const framework::Tensor* o, const std::vector<int> s,
+               const std::vector<int> p, const std::vector<int> d, DataT dtype)
+      : x(x), w(w), o(o), s(s), p(p), d(d), cudnn_dtype(dtype) {}
+};
+
+static inline void GetNCDHW(const framework::DDim& dims,
+                            const DataLayout& layout, int* N, int* C, int* D,
+                            int* H, int* W) {
+  *N = dims[0];
+  *C = layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1];
+  int i = layout == DataLayout::kNCHW ? 0 : 1;
+  if (dims.size() == 5) {
+    *D = dims[2 - i];
+    *H = dims[3 - i];
+    *W = dims[4 - i];
+  } else {
+    *D = 1;
+    *H = dims[2 - i];
+    *W = dims[3 - i];
+  }
+}
+
+template <typename T>
+static std::ostream& operator<<(std::ostream& out, const std::vector<T>& v) {
+  out << "[";
+  for (auto const& tmp : v) out << tmp << ",";
+  out << "]";
+  return out;
+}
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/conv_cudnn_helper.h

@@ -14,44 +14,15 @@ limitations under the License. */
 
 #pragma once
 
-#include <algorithm>
-#include <array>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "paddle/fluid/framework/conv_search_cache.h"
-#include "paddle/fluid/framework/operator_kernel_configs.h"
-#include "paddle/fluid/operators/conv_cudnn_op_cache.h"
-#include "paddle/fluid/operators/eigen/eigen_function.h"
+#include "paddle/fluid/operators/conv_base_helper.h"
 #include "paddle/fluid/platform/cuda_graph_with_memory_pool.h"
-#include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
-#include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/fluid/platform/device/gpu/gpu_info.h"
+#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
 
 namespace paddle {
 namespace operators {
 
-using Tensor = framework::Tensor;
-using DataLayout = platform::DataLayout;
-template <typename T>
-using ScalingParamType = typename platform::CudnnDataType<T>::ScalingParamType;
-using framework::AlgorithmsCache;
-static inline void GetNCDHW(const framework::DDim& dims,
-                            const DataLayout& layout, int* N, int* C, int* D,
-                            int* H, int* W) {
-  *N = dims[0];
-  *C = layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1];
-  int i = layout == DataLayout::kNCHW ? 0 : 1;
-  if (dims.size() == 5) {
-    *D = dims[2 - i];
-    *H = dims[3 - i];
-    *W = dims[4 - i];
-  } else {
-    *D = 1;
-    *H = dims[2 - i];
-    *W = dims[3 - i];
-  }
-}
+using ConvArgs = ConvArgsBase<cudnnHandle_t, cudnnDataType_t>;
 
 template <typename DeviceContext, typename T, size_t D>
 static void RemovePaddingSlice(const phi::GPUContext& context,
@@ -68,121 +39,103 @@ static void RemovePaddingSlice(const phi::GPUContext& context,
     extents[i] = new_out_dims[i];
   }
 
-  int start;
   for (size_t i = 0; i < axes.size(); ++i) {
-    start = starts[i];
+    int start = starts[i];
     if (start < 0) {
       start = (start + in_dims[axes[i]]);
     }
     start = std::max(start, 0);
     offsets[axes[i]] = start;
   }
+
   auto in_t =
       framework::EigenTensor<T, D, Eigen::RowMajor, Eigen::DenseIndex>::From(
           *input);
-
   auto out_t =
       framework::EigenTensor<T, D, Eigen::RowMajor, Eigen::DenseIndex>::From(
           *out, new_out_dims);
-  EigenSlice<std::decay_t<decltype(place)>, T, D>::Eval(place, out_t, in_t,
-                                                        offsets, extents);
+
+  phi::funcs::EigenSlice<std::decay_t<decltype(place)>, T, D>::Eval(
+      place, out_t, in_t, offsets, extents);
 }
 
-template <typename T>
-std::ostream& operator<<(std::ostream& out, const std::vector<T>& v) {
-  out << "[";
-  for (auto const& tmp : v) out << tmp << ",";
-  out << "]";
-  return out;
+static inline double ToMegaBytes(size_t bytes) {
+  return static_cast<double>(bytes) / (1 << 20);
 }
 
-inline int MaxBwdFilterAlgos(cudnnHandle_t cudnn_handle) {
-  int max_algos = 0;
-#if CUDNN_VERSION_MIN(7, 0, 1)
-  PADDLE_ENFORCE_GPU_SUCCESS(
-      platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithmMaxCount(
-          cudnn_handle, &max_algos));
-#endif
-  return max_algos;
+static inline bool UseFixedWorkspace() {
+  return FLAGS_conv_workspace_size_limit >= 0;
 }
 
-template <typename PerfType, typename AlgoType>
-void ChooseAlgoByWorkspace(PerfType* perf_results, size_t perf_num,
-                           size_t workspace_byte, AlgoType* algo) {
-  for (size_t i = 0; i < perf_num; ++i) {
-    auto result = perf_results[i];
-    if (result.status == CUDNN_STATUS_SUCCESS &&
-        result.memory < workspace_byte) {
-      *algo = result.algo;
-      VLOG(3) << "    algo: " << result.algo << ", time: " << result.time
-              << " ms, wksp = " << result.memory
-              << ", status = " << result.status;
-      return;
-    }
+static size_t CaclWorkspaceLimitInBytes(const phi::GPUContext& ctx) {
+  if (!UseFixedWorkspace()) {
+    int device_id = platform::GetCurrentDeviceId();
+    int64_t allocated = memory::StatGetCurrentValue("Allocated", device_id);
+    int64_t reserved = memory::StatGetCurrentValue("Reserved", device_id);
+    int64_t availble = platform::GpuAvailableMemToAlloc();
+    int64_t cur_workspace_size = ctx.cudnn_workspace_handle().WorkspaceSize();
+    VLOG(3) << "[memory] allocated=" << ToMegaBytes(allocated)
+            << " MB, reserved=" << ToMegaBytes(reserved)
+            << " MB, available_to_alloc=" << ToMegaBytes(availble)
+            << " MB, current_workspace_size=" << ToMegaBytes(cur_workspace_size)
+            << " MB.";
+    return std::max(std::max(availble, cur_workspace_size),
+                    reserved - allocated);
+  } else {
+    return FLAGS_conv_workspace_size_limit * 1024 * 1024;
   }
-  VLOG(3) << "Can not find alog that requires memory < "
-          << static_cast<double>(workspace_byte) / (1 << 20) << " MB";
 }
 
-template <typename PerfType, typename AlgoType>
-void ChooseAlgo(const std::vector<PerfType>& perf_results,
-                size_t workspace_byte, AlgoType* algo) {
-  VLOG(3) << "=========BwdFilterAlgo Perf result=========";
-  for (const auto& result : perf_results) {
-    auto math_type_str = "False";
-    if (result.mathType == CUDNN_TENSOR_OP_MATH) {
-      math_type_str = "True";
-    }
-    VLOG(3) << "    algo: " << result.algo << ", TensorCore: " << math_type_str
-            << ", time: " << result.time << " ms"
-            << ", wksp = " << result.memory << ", status = " << result.status;
+template <typename PerfT>
+std::string GetPerfResultString(std::string prefix,
+                                const std::vector<PerfT>& perf_results,
+                                int actual_algo_count, size_t workspace_limit) {
+  std::ostringstream out;
+  out << prefix << " (workspace limit=" << ToMegaBytes(workspace_limit)
+      << " MB):\n";
+  for (int i = 0; i < actual_algo_count; ++i) {
+    const auto& result = perf_results[i];
+    auto math_type_str = (result.mathType == CUDNN_TENSOR_OP_MATH) ? "T" : "F";
+    out << "  algo=" << result.algo << ": tensor_core=" << math_type_str
+        << ", time=" << result.time
+        << " ms, memory=" << ToMegaBytes(result.memory)
+        << " MB, status=" << result.status << "\n";
   }
+  return out.str();
+}
 
-  for (size_t i = 0; i != perf_results.size(); ++i) {
-    const auto& result = perf_results[i];
+template <typename PerfT, typename AlgoT>
+void ChooseAlgoByWorkspace(const std::vector<PerfT>& perf_results,
+                           size_t workspace_limit,
+                           SearchResult<AlgoT>* algo_result) {
+  for (size_t i = 0; i < perf_results.size(); ++i) {
+    auto result = perf_results[i];
     if (result.status == CUDNN_STATUS_SUCCESS &&
-        (result.memory <= workspace_byte)) {
-      if ((result.mathType == CUDNN_TENSOR_OP_MATH) &&
-          (i != perf_results.size() - 1)) {
-        const auto& next_result = perf_results[i + 1];
-        if (next_result.status == CUDNN_STATUS_SUCCESS &&
-            next_result.algo == result.algo &&
-            next_result.memory == result.memory &&
-            next_result.mathType != CUDNN_TENSOR_OP_MATH &&
-            next_result.time < 1.01 * result.time) {
-          // Skip over this result- it's not really a Tensor Core algo.
-          // Because it is only 1% performance difference.
-          // Prefer to choose the next equivalent non-Tensor Core algo.
-          continue;
-        }
-      }
-      *algo = result.algo;
-      auto math_type_str = "0";
-      if (result.mathType == CUDNN_TENSOR_OP_MATH) {
-        math_type_str = "1";
-      }
-      VLOG(3) << "    choose algo: " << result.algo << ", TC: " << math_type_str
-              << ", time: " << result.time << " ms"
-              << ", wksp = " << result.memory << ", status = " << result.status;
-      break;
+        result.memory < workspace_limit) {
+      algo_result->algo = result.algo;
+      algo_result->time = result.time;
+      algo_result->workspace_size = result.memory;
+      VLOG(3) << "  algo=" << result.algo << ", time=" << result.time
+              << " ms, memory=" << ToMegaBytes(result.memory)
+              << " MB (limit=" << ToMegaBytes(workspace_limit)
+              << " MB), status=" << result.status;
+      return;
     }
   }
+  VLOG(3) << "Can not find an algorithm that requires memory < "
+          << ToMegaBytes(workspace_limit) << " MB";
 }
 
-using framework::ConvSearchCache;
-
 static void SetConvMathType(const phi::GPUContext& ctx, cudnnDataType_t dtype,
                             const platform::ConvolutionDescriptor& cdesc) {
 #if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
-  auto& dev_ctx = ctx;
-  if (dev_ctx.GetComputeCapability() >= 70 && dtype == CUDNN_DATA_HALF) {
+  if (ctx.GetComputeCapability() >= 70 && dtype == CUDNN_DATA_HALF) {
     PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
         cdesc.desc(), CUDNN_TENSOR_OP_MATH));
     VLOG(5) << "use cudnn_tensor_op_math";
 #if CUDA_VERSION >= 11000
 #if CUDNN_VERSION_MIN(8, 1, 0)
-  } else if (dev_ctx.GetComputeCapability() >= 80 &&
-             dtype == CUDNN_DATA_BFLOAT16) {
+  } else if (ctx.GetComputeCapability() >= 80 && dtype == CUDNN_DATA_BFLOAT16) {
     PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
         cdesc.desc(), CUDNN_TENSOR_OP_MATH));
 #endif  // CUDNN_VERSION_MIN(8, 1, 0)
@@ -198,76 +151,49 @@ static void SetConvMathType(const phi::GPUContext& ctx, cudnnDataType_t dtype,
 #endif
 }
 
-struct ConvArgs {
-  cudnnHandle_t handle;
-  platform::TensorDescriptor idesc, odesc;
-  platform::FilterDescriptor wdesc;
-  platform::ConvolutionDescriptor cdesc;
-  const framework::Tensor *x, *w, *o;
-  cudnnDataType_t cudnn_dtype;
-
-  // strides
-  std::vector<int> s;
-  // paddings
-  std::vector<int> p;
-  // dilations
-  std::vector<int> d;
-
-  ConvArgs(const framework::Tensor* x, const framework::Tensor* w,
-           const framework::Tensor* o, const std::vector<int> s,
-           const std::vector<int> p, const std::vector<int> d,
-           cudnnDataType_t dtype)
-      : x(x), w(w), o(o), s(s), p(p), d(d), cudnn_dtype(dtype) {}
-};
-
-template <typename perf_t>
-struct SearchAlgorithm {};
-
 template <>
 struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> {
-  using perf_t = cudnnConvolutionFwdAlgoPerf_t;
-  using algo_t = cudnnConvolutionFwdAlgo_t;
+  using PerfT = cudnnConvolutionFwdAlgoPerf_t;
+  using AlgoT = cudnnConvolutionFwdAlgo_t;
 
   template <typename T>
-  static algo_t Find(const ConvArgs& args, bool exhaustive_search,
-                     bool deterministic, const phi::GPUContext& ctx) {
+  static SearchResult<AlgoT> Find(const ConvArgs& args, bool exhaustive_search,
+                                  bool deterministic,
+                                  const phi::GPUContext& ctx) {
+    SearchResult<AlgoT> result;
     auto dtype = platform::CudnnDataType<T>::type;
-    bool has_got_workspace_size = true;
-    size_t workspace_size_limit = FLAGS_conv_workspace_size_limit * 1024 * 1024;
-    size_t workspace_size = 0;
-    algo_t algo;
+    size_t workspace_size_limit = CaclWorkspaceLimitInBytes(ctx);
     SetConvMathType(ctx, dtype, args.cdesc);
 
     if (!exhaustive_search && !deterministic) {
 #if CUDNN_VERSION >= 7001
-      int perf_count;
+      int actual_perf_count;
       int best_algo_idx = 0;
-      std::unique_ptr<perf_t[]> perf_results(new perf_t[kNUM_CUDNN_FWD_ALGS]);
+      std::vector<PerfT> perf_results(kNUM_CUDNN_FWD_ALGS);
       PADDLE_ENFORCE_GPU_SUCCESS(
           platform::dynload::cudnnGetConvolutionForwardAlgorithm_v7(
               args.handle, args.idesc.desc(), args.wdesc.desc(),
               args.cdesc.desc(), args.odesc.desc(), kNUM_CUDNN_FWD_ALGS,
-              &perf_count, perf_results.get()));
-      algo = (perf_results.get())[best_algo_idx].algo;
-      workspace_size = (perf_results.get())[best_algo_idx].memory;
+              &actual_perf_count, perf_results.data()));
+      result.algo = perf_results[best_algo_idx].algo;
+      result.workspace_size = perf_results[best_algo_idx].memory;
 
-      if (workspace_size > workspace_size_limit) {
+      if (result.workspace_size > workspace_size_limit) {
 #if CUDNN_VERSION >= 8000
         // cudnnGetConvolutionForwardAlgorithm is removed in CUDNN-8
-        ChooseAlgoByWorkspace<perf_t, algo_t>(perf_results.get(),
-                                              kNUM_CUDNN_FWD_ALGS,
-                                              workspace_size_limit, &algo);
+        ChooseAlgoByWorkspace<PerfT, AlgoT>(perf_results, workspace_size_limit,
+                                            &result);
 #else
-        VLOG(1) << "Fallback to non-v7 method to find conv algorithm becasue "
-                   "the workspace size request("
-                << workspace_size << ") exceeds the limit("
+        VLOG(3) << "Fallback to non-v7 method to find conv algorithm "
+                   "becasue the workspace size request("
+                << result.workspace_size << ") exceeds the limit("
                 << workspace_size_limit << ")";
         PADDLE_ENFORCE_GPU_SUCCESS(
             platform::dynload::cudnnGetConvolutionForwardAlgorithm(
                 args.handle, args.idesc.desc(), args.wdesc.desc(),
                 args.cdesc.desc(), args.odesc.desc(),
                 CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
-                workspace_size_limit, &algo));
+                workspace_size_limit, &(result.algo)));
 #endif
       }
 #else
@@ -276,30 +202,30 @@ struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> {
               args.handle, args.idesc.desc(), args.wdesc.desc(),
               args.cdesc.desc(), args.odesc.desc(),
               CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
-              workspace_size_limit, &algo));
+              workspace_size_limit, &(result.algo)));
 #endif
-      VLOG(3) << "choose algo " << algo;
     } else if (deterministic) {
-      algo = static_cast<cudnnConvolutionFwdAlgo_t>(1);
+      result.algo = static_cast<AlgoT>(1);
     } else {
-      auto& dev_ctx = ctx;
-      auto workspace_handle = dev_ctx.cudnn_workspace_handle();
-
-      AlgorithmsCache<algo_t>& algo_cache =
-          *(framework::ConvSearchCache::Instance().GetForward());
-
+      auto workspace_handle = ctx.cudnn_workspace_handle();
       auto x_dims = phi::vectorize(args.x->dims());
       auto w_dims = phi::vectorize(args.w->dims());
-
       VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t:"
                << ", x_dims:" << x_dims << ", w_dims:" << w_dims << ", args.s"
                << args.s << ", args.p" << args.p << ", args.d" << args.d;
 
-      algo = algo_cache.GetAlgorithm(
+      AlgorithmsCache<AlgoT>& algo_cache =
+          *(framework::ConvSearchCache::Instance().GetForward());
+
+      result.algo = algo_cache.GetAlgorithm(
           x_dims, w_dims, args.s, args.p, args.d, 0,
           static_cast<int64_t>(args.cudnn_dtype), [&]() {
             int returned_algo_count;
-            std::array<perf_t, kNUM_CUDNN_FWD_ALGS> perf_stat;
+            std::vector<PerfT> perf_results(kNUM_CUDNN_FWD_ALGS);
+            size_t max_workspace_size =
+                FindMaxWorkspaceSize(args, workspace_size_limit);
+            VLOG(4) << "max_workspace_size=" << ToMegaBytes(max_workspace_size)
+                    << " MB";
 
             auto cudnn_find_func = [&](void* cudnn_workspace_ptr) {
               PADDLE_ENFORCE_GPU_SUCCESS(
@@ -308,25 +234,28 @@ struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> {
                       args.wdesc.desc(), args.w->data<T>(), args.cdesc.desc(),
                       args.odesc.desc(), const_cast<T*>(args.o->data<T>()),
                       kNUM_CUDNN_FWD_ALGS, &returned_algo_count,
-                      perf_stat.data(), cudnn_workspace_ptr,
-                      workspace_size_limit));
+                      perf_results.data(), cudnn_workspace_ptr,
+                      max_workspace_size));
             };
-            workspace_handle.RunFuncSync(cudnn_find_func, workspace_size_limit);
-
-            VLOG(3) << "FwdAlgo Perf result: (algo: stat, time, memory)";
-            for (int i = 0; i < returned_algo_count; ++i) {
-              const auto& stat = perf_stat[i];
-              VLOG(3) << stat.algo << ": " << stat.status << " " << stat.time
-                      << " " << stat.memory;
-            }
-            return perf_stat[0].algo;
+            workspace_handle.RunFuncSync(cudnn_find_func, max_workspace_size,
+                                         UseFixedWorkspace());
+
+            VLOG(4) << GetPerfResultString<PerfT>(
+                "[Exhaustive Search] FwdAlgo Perf result", perf_results,
+                returned_algo_count, workspace_size_limit);
+            result.time = perf_results[0].time;
+            return perf_results[0].algo;
           });
     }
-    VLOG(3) << "choose algo " << algo;
-    return algo;
+    VLOG(3) << "[cuDNN Convoltion] exhaustive_search=" << exhaustive_search
+            << ", deterministic=" << deterministic
+            << ", choose algo=" << result.algo << ", workspace="
+            << ToMegaBytes(GetWorkspaceSize(args, result.algo)) << " MB";
+    return result;
   }
 
-  static size_t GetWorkspaceSize(const ConvArgs& args, algo_t algo) {
+  static size_t GetWorkspaceSize(const ConvArgs& args,
+                                 cudnnConvolutionFwdAlgo_t algo) {
     size_t workspace_size = 0;
     PADDLE_ENFORCE_GPU_SUCCESS(
         platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
@@ -334,68 +263,84 @@ struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> {
             args.cdesc.desc(), args.odesc.desc(), algo, &workspace_size));
     return workspace_size;
   }
+
+ private:
+  static size_t FindMaxWorkspaceSize(const ConvArgs& args,
+                                     size_t workspace_size_limit) {
+    if (!UseFixedWorkspace()) {
+      size_t max_workspace_size = 0;
+      for (size_t algo = 0; algo < kNUM_CUDNN_FWD_ALGS; ++algo) {
+        size_t workspace_size = 0;
+        auto status =
+            platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
+                args.handle, args.idesc.desc(), args.wdesc.desc(),
+                args.cdesc.desc(), args.odesc.desc(),
+                static_cast<cudnnConvolutionFwdAlgo_t>(algo), &workspace_size);
+        if (status == CUDNN_STATUS_SUCCESS) {
+          max_workspace_size = std::max(workspace_size, max_workspace_size);
+        }
+      }
+      return std::min(max_workspace_size, workspace_size_limit);
+    } else {
+      return workspace_size_limit;
+    }
+  }
 };
 
 template <>
 struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> {
-  using perf_t = cudnnConvolutionBwdDataAlgoPerf_t;
-  using algo_t = cudnnConvolutionBwdDataAlgo_t;
+  using PerfT = cudnnConvolutionBwdDataAlgoPerf_t;
+  using AlgoT = cudnnConvolutionBwdDataAlgo_t;
 
   template <typename T>
-  static algo_t Find(const ConvArgs& args, bool exhaustive_search,
-                     bool deterministic, const phi::GPUContext& ctx) {
+  static SearchResult<AlgoT> Find(const ConvArgs& args, bool exhaustive_search,
+                                  bool deterministic,
+                                  const phi::GPUContext& ctx) {
+    SearchResult<AlgoT> result;
     auto dtype = platform::CudnnDataType<T>::type;
-    size_t workspace_size_limit = FLAGS_conv_workspace_size_limit * 1024 * 1024;
-    size_t workspace_size = 0;
-    bool has_got_workspace_size = true;
-    algo_t algo;
+    size_t workspace_size_limit = CaclWorkspaceLimitInBytes(ctx);
     SetConvMathType(ctx, dtype, args.cdesc);
 
     if (!exhaustive_search && !deterministic) {
 #if CUDNN_VERSION >= 7001
-      int perf_count;
+      int actual_perf_count;
       int best_algo_idx = 0;
-      std::unique_ptr<perf_t[]> perf_results(
-          new perf_t[kNUM_CUDNN_BWD_DATA_ALGS]);
+      std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_DATA_ALGS);
       PADDLE_ENFORCE_GPU_SUCCESS(
           platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm_v7(
               args.handle, args.wdesc.desc(), args.odesc.desc(),
               args.cdesc.desc(), args.idesc.desc(), kNUM_CUDNN_BWD_DATA_ALGS,
-              &perf_count, perf_results.get()));
-      algo = (perf_results.get())[best_algo_idx].algo;
+              &actual_perf_count, perf_results.data()));
+      result.algo = perf_results[best_algo_idx].algo;
 
 #if CUDNN_VERSION < 7500
       int stride_dim = args.x->dims().size() - 2;
       bool blacklist = std::any_of(args.s.begin(), args.s.begin() + stride_dim,
                                    [=](int n) { return n != 1; });
-      if (blacklist && (static_cast<cudnnConvolutionBwdDataAlgo_t>(
-                            perf_results[best_algo_idx].algo) ==
+      if (blacklist && (perf_results[best_algo_idx].algo ==
                             CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING ||
-                        static_cast<cudnnConvolutionBwdDataAlgo_t>(
-                            perf_results[best_algo_idx].algo) ==
+                        perf_results[best_algo_idx].algo ==
                             CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT)) {
-        algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
+        result.algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
       }
 #endif
-      workspace_size = GetWorkspaceSize(args, algo);
-      if (workspace_size > workspace_size_limit) {
-        has_got_workspace_size = false;
+      result.workspace_size = GetWorkspaceSize(args, result.algo);
+      if (result.workspace_size > workspace_size_limit) {
 #if CUDNN_VERSION >= 8000
         // cudnnGetConvolutionBackwardDataAlgorithm is removed in CUDNN-8
-        ChooseAlgoByWorkspace<perf_t, algo_t>(perf_results.get(),
-                                              kNUM_CUDNN_BWD_DATA_ALGS,
-                                              workspace_size_limit, &algo);
+        ChooseAlgoByWorkspace<PerfT, AlgoT>(perf_results, workspace_size_limit,
+                                            &result);
 #else
         VLOG(1) << "Fallback to non-v7 method to find conv algorithm becasue "
                    "the workspace size request("
-                << workspace_size << ") exceeds the limit("
+                << result.workspace_size << ") exceeds the limit("
                 << workspace_size_limit << ")";
         PADDLE_ENFORCE_GPU_SUCCESS(
             platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
                 args.handle, args.wdesc.desc(), args.odesc.desc(),
                 args.cdesc.desc(), args.idesc.desc(),
                 CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
-                workspace_size_limit, &algo));
+                workspace_size_limit, &(result.algo)));
 #endif
       }
 #else
@@ -404,29 +349,29 @@ struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> {
               args.handle, args.wdesc.desc(), args.odesc.desc(),
               args.cdesc.desc(), args.idesc.desc(),
               CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
-              workspace_size_limit, &algo));
+              workspace_size_limit, &(result.algo)));
 #endif
     } else if (deterministic) {
-      return CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
+      result.algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
     } else {
-      auto& dev_ctx = ctx;
-      auto workspace_handle = dev_ctx.cudnn_workspace_handle();
-
-      AlgorithmsCache<algo_t>& algo_cache =
-          *(framework::ConvSearchCache::Instance().GetBackwardData());
-
+      auto workspace_handle = ctx.cudnn_workspace_handle();
       auto x_dims = phi::vectorize(args.x->dims());
       auto w_dims = phi::vectorize(args.w->dims());
-
       VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t"
                << ", x_dims:" << x_dims << ", w_dims:" << w_dims << ", args.s"
                << args.s << ", args.p" << args.p << ", args.d" << args.d;
 
-      algo = algo_cache.GetAlgorithm(
+      AlgorithmsCache<AlgoT>& algo_cache =
+          *(framework::ConvSearchCache::Instance().GetBackwardData());
+      result.algo = algo_cache.GetAlgorithm(
           x_dims, w_dims, args.s, args.p, args.d, 0,
           static_cast<int64_t>(args.cudnn_dtype), [&]() {
             int returned_algo_count;
-            std::array<perf_t, kNUM_CUDNN_BWD_DATA_ALGS> perf_stat;
+            std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_DATA_ALGS);
+            size_t max_workspace_size =
+                FindMaxWorkspaceSize(args, workspace_size_limit);
+            VLOG(3) << "max_workspace_size=" << ToMegaBytes(max_workspace_size)
+                    << " MB";
 
             auto cudnn_find_func = [&](void* cudnn_workspace_ptr) {
               PADDLE_ENFORCE_GPU_SUCCESS(
@@ -437,26 +382,28 @@ struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> {
                           args.cdesc.desc(), args.idesc.desc(),
                           const_cast<T*>(args.x->data<T>()),
                           kNUM_CUDNN_BWD_DATA_ALGS, &returned_algo_count,
-                          perf_stat.data(), cudnn_workspace_ptr,
-                          workspace_size_limit));
+                          perf_results.data(), cudnn_workspace_ptr,
+                          max_workspace_size));
             };
-            workspace_handle.RunFuncSync(cudnn_find_func, workspace_size_limit);
-
-            VLOG(3) << "BwdDataAlgo Perf result: (algo: stat, time, memory)";
-            for (int i = 0; i < returned_algo_count; ++i) {
-              const auto& stat = perf_stat[i];
-              VLOG(3) << stat.algo << ": " << stat.status << " " << stat.time
-                      << " " << stat.memory;
-            }
-
-            return perf_stat[0].algo;
+            workspace_handle.RunFuncSync(cudnn_find_func, max_workspace_size,
+                                         UseFixedWorkspace());
+
+            VLOG(3) << GetPerfResultString<PerfT>(
+                "[Exhaustive Search] BwdDataAlgo Perf result", perf_results,
+                returned_algo_count, workspace_size_limit);
+            result.time = perf_results[0].time;
+            return perf_results[0].algo;
           });
     }
-    VLOG(3) << "choose algo " << algo;
-    return algo;
+    VLOG(3) << "[cuDNN Convoltion] exhaustive_search=" << exhaustive_search
+            << ", deterministic=" << deterministic
+            << ", choose algo=" << result.algo << ", workspace="
+            << ToMegaBytes(GetWorkspaceSize(args, result.algo)) << " MB";
+    return result;
   }
 
-  static size_t GetWorkspaceSize(const ConvArgs& args, algo_t algo) {
+  static size_t GetWorkspaceSize(const ConvArgs& args,
+                                 cudnnConvolutionBwdDataAlgo_t algo) {
     size_t workspace_size = 0;
     PADDLE_ENFORCE_GPU_SUCCESS(
         platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
@@ -464,57 +411,75 @@ struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> {
             args.cdesc.desc(), args.idesc.desc(), algo, &workspace_size));
     return workspace_size;
   }
+
+ private:
+  static size_t FindMaxWorkspaceSize(const ConvArgs& args,
+                                     size_t workspace_size_limit) {
+    if (!UseFixedWorkspace()) {
+      size_t max_workspace_size = 0;
+      for (size_t algo = 0; algo < kNUM_CUDNN_BWD_DATA_ALGS; ++algo) {
+        size_t workspace_size = 0;
+        auto status =
+            platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
+                args.handle, args.wdesc.desc(), args.odesc.desc(),
+                args.cdesc.desc(), args.idesc.desc(),
+                static_cast<cudnnConvolutionBwdDataAlgo_t>(algo),
+                &workspace_size);
+        if (status == CUDNN_STATUS_SUCCESS) {
+          max_workspace_size = std::max(workspace_size, max_workspace_size);
+        }
+      }
+      return std::min(max_workspace_size, workspace_size_limit);
+    } else {
+      return workspace_size_limit;
+    }
+  }
 };
 
 template <>
 struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
-  using perf_t = cudnnConvolutionBwdFilterAlgoPerf_t;
-  using algo_t = cudnnConvolutionBwdFilterAlgo_t;
+  using PerfT = cudnnConvolutionBwdFilterAlgoPerf_t;
+  using AlgoT = cudnnConvolutionBwdFilterAlgo_t;
 
   template <typename T>
-  static algo_t Find(const ConvArgs& args, bool exhaustive_search,
-                     bool deterministic, const phi::GPUContext& ctx) {
+  static SearchResult<AlgoT> Find(const ConvArgs& args, bool exhaustive_search,
+                                  bool deterministic,
+                                  const phi::GPUContext& ctx) {
     platform::CUDAGraphCaptureModeGuard guard;
+    SearchResult<AlgoT> result;
     auto dtype = platform::CudnnDataType<T>::type;
-    size_t workspace_size_limit = FLAGS_conv_workspace_size_limit * 1024 * 1024;
-    size_t workspace_size = 0;
-    bool has_got_workspace_size = true;
+    size_t workspace_size_limit = CaclWorkspaceLimitInBytes(ctx);
     SetConvMathType(ctx, dtype, args.cdesc);
 
-    algo_t algo;
     if (!exhaustive_search && !deterministic) {
 #if CUDNN_VERSION >= 7001
-      using perf_t = cudnnConvolutionBwdFilterAlgoPerf_t;
-      int perf_count;
+      int actual_perf_count;
       int best_algo_idx = 0;
-      std::unique_ptr<perf_t[]> perf_results(
-          new perf_t[kNUM_CUDNN_BWD_FILTER_ALGS]);
+      std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_FILTER_ALGS);
       PADDLE_ENFORCE_GPU_SUCCESS(
           platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm_v7(
               args.handle, args.idesc.desc(), args.odesc.desc(),
               args.cdesc.desc(), args.wdesc.desc(), kNUM_CUDNN_BWD_FILTER_ALGS,
-              &perf_count, perf_results.get()));
-      algo = (perf_results.get())[best_algo_idx].algo;
-      workspace_size = (perf_results.get())[best_algo_idx].memory;
+              &actual_perf_count, perf_results.data()));
+      result.algo = perf_results[best_algo_idx].algo;
+      result.workspace_size = perf_results[best_algo_idx].memory;
 
-      if (workspace_size > workspace_size_limit) {
-        workspace_size = workspace_size_limit;
+      if (result.workspace_size > workspace_size_limit) {
 #if CUDNN_VERSION >= 8000
         // cudnnGetConvolutionBackwardFilterAlgorithm is removed in CUDNN-8
-        ChooseAlgoByWorkspace<perf_t, algo_t>(perf_results.get(),
-                                              kNUM_CUDNN_BWD_FILTER_ALGS,
-                                              workspace_size_limit, &algo);
+        ChooseAlgoByWorkspace<PerfT, AlgoT>(perf_results, workspace_size_limit,
+                                            &result);
 #else
         VLOG(1) << "Fallback to non-v7 method to find conv algorithm becasue "
                    "the workspace size request("
-                << workspace_size << ") exceeds the limit("
+                << result.workspace_size << ") exceeds the limit("
                 << workspace_size_limit << ")";
         PADDLE_ENFORCE_GPU_SUCCESS(
             platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
                 args.handle, args.idesc.desc(), args.odesc.desc(),
                 args.cdesc.desc(), args.wdesc.desc(),
                 CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
-                workspace_size_limit, &algo));
+                workspace_size_limit, &(result.algo)));
 #endif
       }
 #else
@@ -523,28 +488,32 @@ struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
               args.handle, args.idesc.desc(), args.odesc.desc(),
               args.cdesc.desc(), args.wdesc.desc(),
               CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
-              workspace_size_limit, &algo));
+              workspace_size_limit, &(result.algo)));
 #endif
     } else if (deterministic) {
-      return CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
+      result.algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
     } else {
-      auto& dev_ctx = ctx;
-      auto workspace_handle = dev_ctx.cudnn_workspace_handle();
-      AlgorithmsCache<algo_t>& algo_cache =
-          *(framework::ConvSearchCache::Instance().GetBackwardFilter());
-
+      auto workspace_handle = ctx.cudnn_workspace_handle();
       auto x_dims = phi::vectorize(args.x->dims());
       auto w_dims = phi::vectorize(args.w->dims());
-
       VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t:"
                << ", x_dims:" << x_dims << ", w_dims:" << w_dims << ", args.s"
                << args.s << ", args.p" << args.p << ", args.d" << args.d;
+
+      AlgorithmsCache<AlgoT>& algo_cache =
+          *(framework::ConvSearchCache::Instance().GetBackwardFilter());
+
       if (dtype != CUDNN_DATA_HALF) {
-        algo = algo_cache.GetAlgorithm(
+        result.algo = algo_cache.GetAlgorithm(
             x_dims, w_dims, args.s, args.p, args.d, 0,
             static_cast<int64_t>(args.cudnn_dtype), [&]() {
               int returned_algo_count;
-              std::array<perf_t, kNUM_CUDNN_BWD_FILTER_ALGS> perf_stat;
+              std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_FILTER_ALGS);
+              size_t max_workspace_size =
+                  FindMaxWorkspaceSize(args, workspace_size_limit);
+              VLOG(3) << "max_workspace_size="
+                      << ToMegaBytes(max_workspace_size) << " MB";
+
               auto cudnn_find_func = [&](void* cudnn_workspace_ptr) {
                 PADDLE_ENFORCE_GPU_SUCCESS(
                     platform::dynload::
@@ -554,29 +523,26 @@ struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
                             args.cdesc.desc(), args.wdesc.desc(),
                             const_cast<T*>(args.w->data<T>()),
                             kNUM_CUDNN_BWD_FILTER_ALGS, &returned_algo_count,
-                            perf_stat.data(), cudnn_workspace_ptr,
-                            workspace_size_limit));
+                            perf_results.data(), cudnn_workspace_ptr,
+                            max_workspace_size));
               };
-              workspace_handle.RunFuncSync(cudnn_find_func,
-                                           workspace_size_limit);
-
-              VLOG(3)
-                  << "BwdFilterAlgo Perf result: (algo: stat, time, memory)";
-              for (int i = 0; i < returned_algo_count; ++i) {
-                const auto& stat = perf_stat[i];
-                VLOG(3) << stat.algo << ": " << stat.status << " " << stat.time
-                        << " " << stat.memory;
-              }
-              return perf_stat[0].algo;
+              workspace_handle.RunFuncSync(cudnn_find_func, max_workspace_size,
+                                           UseFixedWorkspace());
+
+              VLOG(3) << GetPerfResultString<PerfT>(
+                  "[Exhaustive Search] BwdFilterAlgo Perf result", perf_results,
+                  returned_algo_count, workspace_size_limit);
+              result.time = perf_results[0].time;
+              return perf_results[0].algo;
             });
       } else {
-        auto max_algos = MaxBwdFilterAlgos(args.handle);
-        algo = algo_cache.GetAlgorithm(
+        result.algo = algo_cache.GetAlgorithm(
             x_dims, w_dims, args.s, args.p, args.d, 0,
             static_cast<int64_t>(args.cudnn_dtype), [&]() {
-              algo_t chosen_algo;
-              std::vector<perf_t> perf_results(max_algos);
+              SearchResult<AlgoT> algo_result;
               int actual_algos = 0;
+              std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_FILTER_ALGS);
+
               PADDLE_ENFORCE_GPU_SUCCESS(
                   platform::dynload::
                       cudnnFindConvolutionBackwardFilterAlgorithm(
@@ -585,17 +551,21 @@ struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
                           perf_results.size(), &actual_algos,
                           perf_results.data()));
               perf_results.resize(actual_algos);
-              ChooseAlgo<perf_t, algo_t>(perf_results, workspace_size_limit,
-                                         &chosen_algo);
-              return chosen_algo;
+              ChooseAlgo(perf_results, workspace_size_limit, &algo_result);
+              result.time = algo_result.time;
+              return algo_result.algo;
             });
       }
     }
-    VLOG(3) << "choose algo " << algo;
-    return algo;
+    VLOG(3) << "[cuDNN Convoltion] exhaustive_search=" << exhaustive_search
+            << ", deterministic=" << deterministic
+            << ", choose algo=" << result.algo << ", workspace="
+            << ToMegaBytes(GetWorkspaceSize(args, result.algo)) << " MB";
+    return result;
   }
 
-  static size_t GetWorkspaceSize(const ConvArgs& args, algo_t algo) {
+  static size_t GetWorkspaceSize(const ConvArgs& args,
+                                 cudnnConvolutionBwdFilterAlgo_t algo) {
     platform::CUDAGraphCaptureModeGuard guard;
     size_t workspace_size = 0;
     PADDLE_ENFORCE_GPU_SUCCESS(
@@ -604,6 +574,69 @@ struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
             args.cdesc.desc(), args.wdesc.desc(), algo, &workspace_size));
     return workspace_size;
   }
+
+ private:
+  static size_t FindMaxWorkspaceSize(const ConvArgs& args,
+                                     size_t workspace_size_limit) {
+    if (!UseFixedWorkspace()) {
+      size_t max_workspace_size = 0;
+      for (size_t algo = 0; algo < kNUM_CUDNN_BWD_FILTER_ALGS; ++algo) {
+        size_t workspace_size = 0;
+        auto status =
+            platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
+                args.handle, args.idesc.desc(), args.odesc.desc(),
+                args.cdesc.desc(), args.wdesc.desc(),
+                static_cast<cudnnConvolutionBwdFilterAlgo_t>(algo),
+                &workspace_size);
+        if (status == CUDNN_STATUS_SUCCESS) {
+          max_workspace_size = std::max(workspace_size, max_workspace_size);
+        }
+      }
+      return std::min(max_workspace_size, workspace_size_limit);
+    } else {
+      return workspace_size_limit;
+    }
+  }
+
+  static void ChooseAlgo(const std::vector<PerfT>& perf_results,
+                         size_t workspace_limit,
+                         SearchResult<AlgoT>* algo_result) {
+    VLOG(3) << GetPerfResultString<PerfT>(
+        "[Exhaustive Search] BwdFilterAlgo Perf result", perf_results,
+        perf_results.size(), workspace_limit);
+
+    for (size_t i = 0; i != perf_results.size(); ++i) {
+      const auto& result = perf_results[i];
+      if (result.status == CUDNN_STATUS_SUCCESS &&
+          (result.memory <= workspace_limit)) {
+        if ((result.mathType == CUDNN_TENSOR_OP_MATH) &&
+            (i != perf_results.size() - 1)) {
+          const auto& next_result = perf_results[i + 1];
+          if (next_result.status == CUDNN_STATUS_SUCCESS &&
+              next_result.algo == result.algo &&
+              next_result.memory == result.memory &&
+              next_result.mathType != CUDNN_TENSOR_OP_MATH &&
+              next_result.time < 1.01 * result.time) {
+            // Skip over this result- it's not really a Tensor Core algo.
+            // Because it is only 1% performance difference.
+            // Prefer to choose the next equivalent non-Tensor Core algo.
+            continue;
+          }
+        }
+        algo_result->algo = result.algo;
+        algo_result->time = result.time;
+        auto math_type_str = "0";
+        if (result.mathType == CUDNN_TENSOR_OP_MATH) {
+          math_type_str = "1";
+        }
+        VLOG(3) << "    choose algo: " << result.algo
+                << ", TC: " << math_type_str << ", time: " << result.time
+                << " ms, wksp = " << result.memory
+                << ", status = " << result.status;
+        break;
+      }
+    }
+  }
 };
 
 }  // namespace operators

paddle/fluid/operators/conv_cudnn_op_cache.h

@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
 
-DECLARE_uint64(conv_workspace_size_limit);
+DECLARE_int64(conv_workspace_size_limit);
 DECLARE_bool(cudnn_exhaustive_search);
 DECLARE_int64(cudnn_exhaustive_search_times);
 

paddle/fluid/operators/conv_miopen_helper.h

@@ -14,42 +14,12 @@ limitations under the License. */
 
 #pragma once
 
-#include <algorithm>
-#include <array>
-#include <memory>
-#include <string>
-#include <vector>
-
-#include "paddle/fluid/framework/conv_search_cache.h"
-#include "paddle/fluid/framework/operator_kernel_configs.h"
-#include "paddle/fluid/operators/conv_cudnn_op_cache.h"
-#include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
-#include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/fluid/operators/conv_base_helper.h"
 
 namespace paddle {
 namespace operators {
 
-using Tensor = framework::Tensor;
-using DataLayout = platform::DataLayout;
-template <typename T>
-using ScalingParamType = typename platform::CudnnDataType<T>::ScalingParamType;
-using framework::AlgorithmsCache;
-static inline void GetNCDHW(const framework::DDim& dims,
-                            const DataLayout& layout, int* N, int* C, int* D,
-                            int* H, int* W) {
-  *N = dims[0];
-  *C = layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1];
-  int i = layout == DataLayout::kNCHW ? 0 : 1;
-  if (dims.size() == 5) {
-    *D = dims[2 - i];
-    *H = dims[3 - i];
-    *W = dims[4 - i];
-  } else {
-    *D = 1;
-    *H = dims[2 - i];
-    *W = dims[3 - i];
-  }
-}
+using ConvArgs = ConvArgsBase<miopenHandle_t, miopenDataType_t>;
 
 template <typename DeviceContext, typename T, size_t D>
 static void RemovePaddingSlice(const phi::GPUContext& context,
@@ -66,9 +36,8 @@ static void RemovePaddingSlice(const phi::GPUContext& context,
     extents[i] = new_out_dims[i];
   }
 
-  int start;
   for (size_t i = 0; i < axes.size(); ++i) {
-    start = starts[i];
+    int start = starts[i];
     if (start < 0) {
       start = (start + in_dims[axes[i]]);
     }
@@ -85,41 +54,6 @@ static void RemovePaddingSlice(const phi::GPUContext& context,
   out_t.device(place) = in_t.slice(offsets, extents);
 }
 
-template <typename T>
-std::ostream& operator<<(std::ostream& out, const std::vector<T>& v) {
-  out << "[";
-  for (auto const& tmp : v) out << tmp << ",";
-  out << "]";
-  return out;
-}
-
-using framework::ConvSearchCache;
-
-struct ConvArgs {
-  miopenHandle_t handle;
-  platform::TensorDescriptor idesc, odesc;
-  platform::FilterDescriptor wdesc;
-  platform::ConvolutionDescriptor cdesc;
-  const framework::Tensor *x, *w, *o;
-  miopenDataType_t cudnn_dtype;
-
-  // strides
-  std::vector<int> s;
-  // paddings
-  std::vector<int> p;
-  // dilations
-  std::vector<int> d;
-
-  ConvArgs(const framework::Tensor* x, const framework::Tensor* w,
-           const framework::Tensor* o, const std::vector<int> s,
-           const std::vector<int> p, const std::vector<int> d,
-           miopenDataType_t dtype)
-      : x(x), w(w), o(o), s(s), p(p), d(d), cudnn_dtype(dtype) {}
-};
-
-template <typename algo_t>
-struct SearchAlgorithm {};
-
 template <>
 struct SearchAlgorithm<miopenConvFwdAlgorithm_t> {
   using perf_t = miopenConvAlgoPerf_t;

paddle/fluid/operators/fused/fusion_conv_inception_op.cu

@@ -16,8 +16,6 @@ limitations under the License. */
 #include "paddle/fluid/operators/conv_cudnn_op_cache.h"
 #include "paddle/fluid/platform/device/gpu/gpu_dnn.h"
 
-DECLARE_uint64(conv_workspace_size_limit);
-
 namespace paddle {
 namespace operators {
 

paddle/fluid/platform/device/gpu/gpu_info.cc

@@ -188,6 +188,8 @@ class RecordedGpuMallocHelper {
     if (UNLIKELY(malloc_managed_memory)) {
       result = cudaMallocManaged(ptr, size);
     } else {
+      VLOG(10) << "[cudaMalloc] size=" << static_cast<double>(size) / (1 << 20)
+               << " MB";
       result = cudaMalloc(ptr, size);
     }
 #endif
@@ -226,6 +228,8 @@ class RecordedGpuMallocHelper {
     if (err != hipErrorDeinitialized) {
 #else
     auto err = cudaFree(ptr);
+    VLOG(10) << "[cudaFree] size=" << static_cast<double>(size) / (1 << 20)
+             << " MB";
     if (err != cudaErrorCudartUnloading) {
 #endif
       PADDLE_ENFORCE_GPU_SUCCESS(err);

paddle/fluid/platform/device_context.cc

@@ -522,8 +522,8 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : phi::GPUContext(place) {
   cuda_stream_.reset(new stream::CUDAStream(phi::GPUContext::stream(), place));
   auto& instance = memory::allocation::AllocatorFacade::Instance();
   instance.SetDefaultStream(place, phi::GPUContext::stream());
-  workspace_.reset(
-      new phi::DnnWorkspaceHandle(instance.GetAllocator(place).get()));
+  workspace_.reset(new phi::DnnWorkspaceHandle(
+      instance.GetAllocator(place).get(), stream()));
 }
 
 CUDADeviceContext::~CUDADeviceContext() = default;
@@ -623,7 +623,8 @@ phi::DnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const {
     return phi::DnnWorkspaceHandle(
         memory::allocation::AllocatorFacade::Instance()
             .GetAllocator(GetPlace())
-            .get());
+            .get(),
+        stream());
   }
   return phi::GPUContext::cudnn_workspace_handle();
 }

paddle/fluid/platform/flags.cc

@@ -161,10 +161,9 @@ PADDLE_DEFINE_EXPORTED_bool(
  * increased.
  *       Users need to balance memory and speed.
  */
-PADDLE_DEFINE_EXPORTED_uint64(
-    conv_workspace_size_limit,
-    paddle::platform::kDefaultConvWorkspaceSizeLimitMB,
-    "cuDNN convolution workspace limit in MB unit.");
+PADDLE_DEFINE_EXPORTED_int64(conv_workspace_size_limit,
+                             paddle::platform::kDefaultConvWorkspaceSizeLimitMB,
+                             "cuDNN convolution workspace limit in MB unit.");
 
 /**
  * CUDNN related FLAG

paddle/phi/backends/gpu/gpu_context.cc

@@ -12,6 +12,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include <algorithm>
 #include <array>
@@ -155,6 +156,39 @@ static void StreamCallbackFunc(gpuStream_t stream,
 
 }  // namespace internal
 
+void DnnWorkspaceHandle::RunFuncSync(
+    const std::function<void(void*)>& cudnn_func,
+    size_t required_workspace_bytes,
+    bool use_cached_allocation) {
+  bool need_realloc = required_workspace_bytes > WorkspaceSize();
+  if (need_realloc && !use_cached_allocation) {
+    void* workspace_ptr = nullptr;
+    size_t size = ((required_workspace_bytes + 255) >> 8) << 8;
+    std::lock_guard<std::mutex> guard(*mtx_);
+#ifdef PADDLE_WITH_HIP
+    auto status = hipMalloc(&workspace_ptr, size);
+#else
+    auto status = cudaMalloc(&workspace_ptr, size);
+#endif
+    if (status == gpuSuccess) {
+      cudnn_func(workspace_ptr);
+      phi::backends::gpu::GpuStreamSync(stream_);
+#ifdef PADDLE_WITH_HIP
+      PADDLE_ENFORCE_GPU_SUCCESS(hipFree(workspace_ptr));
+#else
+      PADDLE_ENFORCE_GPU_SUCCESS(cudaFree(workspace_ptr));
+#endif
+      return;
+    }
+  }
+
+  RunFunc(cudnn_func, required_workspace_bytes);
+  if (need_realloc) {
+    // Release the workspace allocated in this running.
+    ResetWorkspace();
+  }
+}
+
 void DnnWorkspaceHandle::ResetWorkspace() { allocation_ = nullptr; }
 
 void DnnWorkspaceHandle::ReallocWorkspace(size_t required_workspace_bytes) {
@@ -295,13 +329,13 @@ struct GPUContext::Impl {
   void InitDnnWorkspace() {
     PD_CHECK(allocator_ != nullptr,
              "the device allocator for gpu context is nullptr.");
-    workspace_ = new DnnWorkspaceHandle(allocator_);
+    workspace_ = new DnnWorkspaceHandle(allocator_, stream_);
   }
 
   void DestoryInternalWorkspace() {
     if (owned_ && workspace_ != nullptr) {
       delete workspace_;
-      stream_ = nullptr;
+      workspace_ = nullptr;
     }
   }
 
@@ -313,7 +347,7 @@ struct GPUContext::Impl {
   DnnWorkspaceHandle GetDnnWorkspace() {
     PD_CHECK(allocator_ != nullptr,
              "the device allocator for gpu context is nullptr.");
-    return DnnWorkspaceHandle(allocator_);
+    return DnnWorkspaceHandle(allocator_, stream_);
   }
 
   void InitStream() {

paddle/phi/backends/gpu/gpu_context.h

@@ -21,6 +21,7 @@ limitations under the License. */
 #include "paddle/phi/backends/gpu/forwards.h"
 #include "paddle/phi/backends/gpu/gpu_decls.h"
 #include "paddle/phi/backends/gpu/gpu_helper.h"
+#include "paddle/phi/backends/gpu/gpu_info.h"
 #include "paddle/phi/common/place.h"
 #include "paddle/phi/core/device_context.h"
 
@@ -28,8 +29,8 @@ namespace phi {
 
 class DnnWorkspaceHandle {
  public:
-  explicit inline DnnWorkspaceHandle(Allocator* allocator)
-      : allocator_(allocator) {
+  inline DnnWorkspaceHandle(Allocator* allocator, gpuStream_t stream)
+      : allocator_(allocator), stream_(stream) {
     mtx_.reset(new std::mutex());
   }
 
@@ -48,11 +49,9 @@ class DnnWorkspaceHandle {
    *  running the function. Currently this function is only used when cudnn
    *  exhaustive searching and callers have to guarantee that the input function
    *  is host blocking */
-  inline void RunFuncSync(const std::function<void(void*)>& cudnn_func,
-                          size_t required_workspace_bytes) {
-    RunFunc(cudnn_func, required_workspace_bytes);
-    ResetWorkspace();
-  }
+  void RunFuncSync(const std::function<void(void*)>& cudnn_func,
+                   size_t required_workspace_bytes,
+                   bool use_cached_allocation = true);
 
   inline size_t WorkspaceSize() {
     if (allocation_ == nullptr) {
@@ -70,7 +69,8 @@ class DnnWorkspaceHandle {
 
  private:
   Allocator::AllocationPtr allocation_{nullptr};
-  Allocator* allocator_{nullptr};
+  Allocator* allocator_{nullptr};  // Not owned
+  gpuStream_t stream_{nullptr};    // Not owned
   std::unique_ptr<std::mutex> mtx_;
 };
 

paddle/phi/kernels/autotune/CMakeLists.txt

 if (WITH_GPU)
-     nv_test(gpu_timer_test SRCS gpu_timer_test.cu DEPS gtest)
-     nv_test(auto_tune_test SRCS auto_tune_test.cu DEPS gtest)
+    nv_test(gpu_timer_test SRCS gpu_timer_test.cu DEPS gtest)
+    nv_test(auto_tune_test SRCS auto_tune_test.cu DEPS gtest)
 elseif (WITH_ROCM)
     hip_test(gpu_timer_test SRCS gpu_timer_test.cu DEPS gtest)
     hip_test(auto_tune_test SRCS auto_tune_test.cu DEPS gtest)

paddle/phi/kernels/gpudnn/conv_grad_grad_kernel.cu

@@ -289,21 +289,17 @@ void ConvCudnnGradGradKernel(
                                     dtype};
 
 #ifdef PADDLE_WITH_HIP
-  miopenConvFwdAlgorithm_t fwd_algo1 = static_cast<miopenConvFwdAlgorithm_t>(0);
-  miopenConvFwdAlgorithm_t fwd_algo2 = static_cast<miopenConvFwdAlgorithm_t>(0);
-  miopenConvBwdDataAlgorithm_t data_algo =
-      static_cast<miopenConvBwdDataAlgorithm_t>(0);
-  miopenConvBwdWeightsAlgorithm_t filter_algo =
-      static_cast<miopenConvBwdWeightsAlgorithm_t>(0);
+  paddle::operators::SearchResult<miopenConvFwdAlgorithm_t> fwd_result1;
+  paddle::operators::SearchResult<miopenConvFwdAlgorithm_t> fwd_result2;
+  paddle::operators::SearchResult<miopenConvBwdDataAlgorithm_t> data_result;
+  paddle::operators::SearchResult<miopenConvBwdWeightsAlgorithm_t>
+      filter_result;
 #else
-  cudnnConvolutionFwdAlgo_t fwd_algo1 =
-      static_cast<cudnnConvolutionFwdAlgo_t>(0);
-  cudnnConvolutionFwdAlgo_t fwd_algo2 =
-      static_cast<cudnnConvolutionFwdAlgo_t>(0);
-  cudnnConvolutionBwdDataAlgo_t data_algo =
-      static_cast<cudnnConvolutionBwdDataAlgo_t>(0);
-  cudnnConvolutionBwdFilterAlgo_t filter_algo =
-      static_cast<cudnnConvolutionBwdFilterAlgo_t>(0);
+  paddle::operators::SearchResult<cudnnConvolutionFwdAlgo_t> fwd_result1;
+  paddle::operators::SearchResult<cudnnConvolutionFwdAlgo_t> fwd_result2;
+  paddle::operators::SearchResult<cudnnConvolutionBwdDataAlgo_t> data_result;
+  paddle::operators::SearchResult<cudnnConvolutionBwdFilterAlgo_t>
+      filter_result;
 #endif
 
   auto layout = paddle::platform::GetCudnnTensorFormat(
@@ -332,13 +328,13 @@ void ConvCudnnGradGradKernel(
       using search1 =
           paddle::operators::SearchAlgorithm<miopenConvFwdAlgorithm_t>;
       workspace_size = search1::GetWorkspaceSize(args1);
-      fwd_algo1 = search1::Find<T>(
+      fwd_result1.algo = search1::Find<T>(
           args1, exhaustive_search, false, workspace_size, ctx);
 #else
       using search1 =
           paddle::operators::SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
-      fwd_algo1 = search1::Find<T>(args1, exhaustive_search, false, ctx);
-      workspace_size = search1::GetWorkspaceSize(args1, fwd_algo1);
+      fwd_result1 = search1::Find<T>(args1, exhaustive_search, false, ctx);
+      workspace_size = search1::GetWorkspaceSize(args1, fwd_result1.algo);
 #endif
     }
 
@@ -360,14 +356,14 @@ void ConvCudnnGradGradKernel(
           paddle::operators::SearchAlgorithm<miopenConvFwdAlgorithm_t>;
       workspace_size =
           std::max(workspace_size, search2::GetWorkspaceSize(args2));
-      fwd_algo2 = search2::Find<T>(
+      fwd_result2.algo = search2::Find<T>(
           args2, exhaustive_search, false, workspace_size, ctx);
 #else
       using search2 =
           paddle::operators::SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
-      fwd_algo2 = search2::Find<T>(args2, exhaustive_search, false, ctx);
-      workspace_size =
-          std::max(workspace_size, search2::GetWorkspaceSize(args2, fwd_algo2));
+      fwd_result2 = search2::Find<T>(args2, exhaustive_search, false, ctx);
+      workspace_size = std::max(
+          workspace_size, search2::GetWorkspaceSize(args2, fwd_result2.algo));
 #endif
     }
   }
@@ -389,15 +385,15 @@ void ConvCudnnGradGradKernel(
     using search3 =
         paddle::operators::SearchAlgorithm<miopenConvBwdWeightsAlgorithm_t>;
     workspace_size = std::max(workspace_size, search3::GetWorkspaceSize(args3));
-    filter_algo = search3::Find<T>(
+    filter_result.algo = search3::Find<T>(
         args3, exhaustive_search, deterministic, workspace_size, ctx);
 #else
     using search3 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t>;
-    filter_algo =
+    filter_result =
         search3::Find<T>(args3, exhaustive_search, deterministic, ctx);
-    workspace_size =
-        std::max(workspace_size, search3::GetWorkspaceSize(args3, filter_algo));
+    workspace_size = std::max(
+        workspace_size, search3::GetWorkspaceSize(args3, filter_result.algo));
 #endif
   }
 
@@ -419,14 +415,15 @@ void ConvCudnnGradGradKernel(
     using search4 =
         paddle::operators::SearchAlgorithm<miopenConvBwdDataAlgorithm_t>;
     workspace_size = std::max(workspace_size, search4::GetWorkspaceSize(args4));
-    data_algo = search4::Find<T>(
+    data_result.algo = search4::Find<T>(
         args4, exhaustive_search, deterministic, workspace_size, ctx);
 #else
     using search4 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
-    data_algo = search4::Find<T>(args4, exhaustive_search, deterministic, ctx);
-    workspace_size =
-        std::max(workspace_size, search4::GetWorkspaceSize(args4, data_algo));
+    data_result =
+        search4::Find<T>(args4, exhaustive_search, deterministic, ctx);
+    workspace_size = std::max(
+        workspace_size, search4::GetWorkspaceSize(args4, data_result.algo));
 #endif
   }
 
@@ -471,7 +468,7 @@ void ConvCudnnGradGradKernel(
                     args1.wdesc.desc(),
                     w,
                     args1.cdesc.desc(),
-                    fwd_algo1,
+                    fwd_result1.algo,
                     &beta,
                     args1.odesc.desc(),
                     transformed_ddy_channel,
@@ -492,7 +489,7 @@ void ConvCudnnGradGradKernel(
                       args1.wdesc.desc(),
                       w + i * group_offset_filter,
                       args1.cdesc.desc(),
-                      fwd_algo1,
+                      fwd_result1.algo,
                       workspace_ptr,
                       workspace_size,
                       &beta,
@@ -517,7 +514,7 @@ void ConvCudnnGradGradKernel(
                     args2.wdesc.desc(),
                     ddw,
                     args2.cdesc.desc(),
-                    fwd_algo2,
+                    fwd_result2.algo,
                     &beta,
                     args2.odesc.desc(),
                     transformed_ddy_channel,
@@ -538,7 +535,7 @@ void ConvCudnnGradGradKernel(
                       args2.wdesc.desc(),
                       ddw + i * group_offset_filter,
                       args2.cdesc.desc(),
-                      fwd_algo2,
+                      fwd_result2.algo,
                       workspace_ptr,
                       workspace_size,
                       &alpha,
@@ -568,7 +565,7 @@ void ConvCudnnGradGradKernel(
                   args3.idesc.desc(),
                   ddx,
                   args3.cdesc.desc(),
-                  filter_algo,
+                  filter_result.algo,
                   &beta,
                   args3.wdesc.desc(),
                   dw,
@@ -589,7 +586,7 @@ void ConvCudnnGradGradKernel(
                     args3.odesc.desc(),
                     transformed_dy_channel + i * group_offset_out,
                     args3.cdesc.desc(),
-                    filter_algo,
+                    filter_result.algo,
                     workspace_ptr,
                     workspace_size,
                     &beta,
@@ -615,7 +612,7 @@ void ConvCudnnGradGradKernel(
                   args4.wdesc.desc(),
                   ddw,
                   args4.cdesc.desc(),
-                  data_algo,
+                  data_result.algo,
                   &beta,
                   args4.idesc.desc(),
                   transformed_dx,
@@ -636,7 +633,7 @@ void ConvCudnnGradGradKernel(
                     args4.odesc.desc(),
                     transformed_dy_channel + i * group_offset_out,
                     args4.cdesc.desc(),
-                    data_algo,
+                    data_result.algo,
                     workspace_ptr,
                     workspace_size,
                     &beta,

paddle/phi/kernels/gpudnn/conv_grad_kernel.cu

@@ -322,17 +322,16 @@ void ConvCudnnGradKernel(const Context& ctx,
   int group_offset_in = i_c / groups * i_h * i_w * i_d;
   int group_offset_out = o_c / groups * o_h * o_w * o_d;
   int group_offset_filter = transformed_filter_channel.numel() / groups;
+
 // ------------------- cudnn backward algorithm ---------------------
 #ifdef PADDLE_WITH_HIP
-  miopenConvBwdDataAlgorithm_t data_algo =
-      static_cast<miopenConvBwdDataAlgorithm_t>(0);
-  miopenConvBwdWeightsAlgorithm_t filter_algo =
-      static_cast<miopenConvBwdWeightsAlgorithm_t>(0);
+  paddle::operators::SearchResult<miopenConvBwdDataAlgorithm_t> bwd_result;
+  paddle::operators::SearchResult<miopenConvBwdWeightsAlgorithm_t>
+      filter_result;
 #else
-  cudnnConvolutionBwdDataAlgo_t data_algo =
-      static_cast<cudnnConvolutionBwdDataAlgo_t>(0);
-  cudnnConvolutionBwdFilterAlgo_t filter_algo =
-      static_cast<cudnnConvolutionBwdFilterAlgo_t>(0);
+  paddle::operators::SearchResult<cudnnConvolutionBwdDataAlgo_t> bwd_result;
+  paddle::operators::SearchResult<cudnnConvolutionBwdFilterAlgo_t>
+      filter_result;
 #endif
   // input data workspace_size
   size_t workspace_size_d = 0;
@@ -368,14 +367,14 @@ void ConvCudnnGradKernel(const Context& ctx,
         paddle::operators::SearchAlgorithm<miopenConvBwdDataAlgorithm_t>;
     workspace_size_d =
         std::max(workspace_size_d, search1::GetWorkspaceSize(args1));
-    data_algo = search1::Find<T>(
+    bwd_result.algo = search1::Find<T>(
         args1, exhaustive_search, deterministic, workspace_size_d, ctx);
 #else
     using search1 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
-    data_algo = search1::Find<T>(args1, exhaustive_search, deterministic, ctx);
-    workspace_size_d =
-        std::max(workspace_size_d, search1::GetWorkspaceSize(args1, data_algo));
+    bwd_result = search1::Find<T>(args1, exhaustive_search, deterministic, ctx);
+    workspace_size_d = std::max(
+        workspace_size_d, search1::GetWorkspaceSize(args1, bwd_result.algo));
 #endif
   }
 
@@ -397,15 +396,17 @@ void ConvCudnnGradKernel(const Context& ctx,
         paddle::operators::SearchAlgorithm<miopenConvBwdWeightsAlgorithm_t>;
     workspace_size_w =
         std::max(workspace_size_w, search2::GetWorkspaceSize(args2));
-    filter_algo = search2::Find<T>(
+    filter_result.algo = search2::Find<T>(
         args2, exhaustive_search, deterministic, workspace_size_w, ctx);
 #else
     using search2 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t>;
-    filter_algo =
+    filter_result =
         search2::Find<T>(args2, exhaustive_search, deterministic, ctx);
-    workspace_size_w = std::max(workspace_size_w,
-                                search2::GetWorkspaceSize(args2, filter_algo));
+    VLOG(3) << "filter algo: " << filter_result.algo << ", time "
+            << filter_result.time;
+    workspace_size_w = std::max(
+        workspace_size_w, search2::GetWorkspaceSize(args2, filter_result.algo));
 #endif
   }
 
@@ -439,7 +440,7 @@ void ConvCudnnGradKernel(const Context& ctx,
                     args1.wdesc.desc(),
                     filter_data,
                     args1.cdesc.desc(),
-                    data_algo,
+                    bwd_result.algo,
                     &beta,
                     args1.idesc.desc(),
                     temp_tensor_data,
@@ -471,7 +472,7 @@ void ConvCudnnGradKernel(const Context& ctx,
                     args1.wdesc.desc(),
                     filter_data,
                     args1.cdesc.desc(),
-                    data_algo,
+                    bwd_result.algo,
                     &beta,
                     args1.idesc.desc(),
                     transformed_input_grad_data,
@@ -494,7 +495,7 @@ void ConvCudnnGradKernel(const Context& ctx,
                     args1.odesc.desc(),
                     output_grad_data + i * group_offset_out,
                     args1.cdesc.desc(),
-                    data_algo,
+                    bwd_result.algo,
                     cudnn_workspace_ptr,
                     workspace_size_d,
                     &beta,
@@ -554,7 +555,7 @@ void ConvCudnnGradKernel(const Context& ctx,
                   args2.idesc.desc(),
                   input_data,
                   args2.cdesc.desc(),
-                  filter_algo,
+                  filter_result.algo,
                   &beta,
                   args2.wdesc.desc(),
                   filter_grad_data,
@@ -575,7 +576,7 @@ void ConvCudnnGradKernel(const Context& ctx,
                     args2.odesc.desc(),
                     output_grad_data + i * group_offset_out,
                     args2.cdesc.desc(),
-                    filter_algo,
+                    filter_result.algo,
                     cudnn_workspace_ptr,
                     workspace_size_w,
                     &beta_filter,

paddle/phi/kernels/gpudnn/conv_kernel.cu

@@ -18,7 +18,6 @@
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/core/kernel_registry.h"
 
-#include "paddle/fluid/framework/eigen.h"
 #ifdef PADDLE_WITH_HIP
 #include "paddle/fluid/operators/conv_miopen_helper.h"
 #else
@@ -68,7 +67,6 @@ void ConvCudnnKernel(const Context& ctx,
                         "FLAGS_cudnn_deterministic True at same time."));
 
   const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");
-
   auto dtype = paddle::platform::CudnnDataType<T>::type;
 
 #ifdef PADDLE_WITH_HIP
@@ -309,17 +307,17 @@ void ConvCudnnKernel(const Context& ctx,
   size_t workspace_size = 0;  // final workspace to allocate.
 // ------------------- cudnn conv algorithm ---------------------
 #ifdef PADDLE_WITH_HIP
-  miopenConvFwdAlgorithm_t algo{};
+  paddle::operators::SearchResult<miopenConvFwdAlgorithm_t> fwd_result;
   using search = paddle::operators::SearchAlgorithm<miopenConvFwdAlgorithm_t>;
   workspace_size = search::GetWorkspaceSize(args);
-  algo = search::Find<T>(
+  fwd_result.algo = search::Find<T>(
       args, exhaustive_search, deterministic, workspace_size, ctx);
 #else
-  cudnnConvolutionFwdAlgo_t algo{};
+  paddle::operators::SearchResult<cudnnConvolutionFwdAlgo_t> fwd_result;
   using search =
       paddle::operators::SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
-  algo = search::Find<T>(args, exhaustive_search, deterministic, ctx);
-  workspace_size = search::GetWorkspaceSize(args, algo);
+  fwd_result = search::Find<T>(args, exhaustive_search, deterministic, ctx);
+  workspace_size = search::GetWorkspaceSize(args, fwd_result.algo);
 #endif
 
 #if defined(PADDLE_WITH_CUDA) && CUDNN_VERSION_MIN(7, 0, 1)
@@ -328,7 +326,7 @@ void ConvCudnnKernel(const Context& ctx,
   // in forward computation, so change the algorithm to CUDNN_CONVOLUTION_\
     // FWD_ALGO_IMPLICIT_GEMM manually.
   if (groups > 1) {
-    algo = static_cast<cudnnConvolutionFwdAlgo_t>(0);
+    fwd_result.algo = static_cast<cudnnConvolutionFwdAlgo_t>(0);
   }
 #endif
 
@@ -352,7 +350,7 @@ void ConvCudnnKernel(const Context& ctx,
                 args.wdesc.desc(),
                 filter_data,
                 args.cdesc.desc(),
-                algo,
+                fwd_result.algo,
                 &beta,
                 args.odesc.desc(),
                 output_data,
@@ -373,7 +371,7 @@ void ConvCudnnKernel(const Context& ctx,
                   args.wdesc.desc(),
                   filter_data + i * group_offset_filter,
                   args.cdesc.desc(),
-                  algo,
+                  fwd_result.algo,
                   workspace_ptr,
                   workspace_size,
                   &beta,

paddle/phi/kernels/gpudnn/conv_transpose_grad_kernel.cu

@@ -188,11 +188,13 @@ void ConvTransposeGradRawGPUDNNKernel(const Context& ctx,
                                     dtype};
 
 #ifdef PADDLE_WITH_HIP
-  miopenConvFwdAlgorithm_t data_algo{};
-  miopenConvBwdWeightsAlgorithm_t filter_algo{};
+  paddle::operators::SearchResult<miopenConvFwdAlgorithm_t> fwd_result;
+  paddle::operators::SearchResult<miopenConvBwdWeightsAlgorithm_t>
+      filter_result;
 #else
-  cudnnConvolutionFwdAlgo_t data_algo{};
-  cudnnConvolutionBwdFilterAlgo_t filter_algo{};
+  paddle::operators::SearchResult<cudnnConvolutionFwdAlgo_t> fwd_result;
+  paddle::operators::SearchResult<cudnnConvolutionBwdFilterAlgo_t>
+      filter_result;
 #endif
 
   auto layout_tensor = paddle::platform::GetCudnnTensorFormat(layout);
@@ -218,14 +220,14 @@ void ConvTransposeGradRawGPUDNNKernel(const Context& ctx,
     using search1 =
         paddle::operators::SearchAlgorithm<miopenConvFwdAlgorithm_t>;
     workspace_size = std::max(workspace_size, search1::GetWorkspaceSize(args1));
-    data_algo =
+    fwd_result.algo =
         search1::Find<T>(args1, false, deterministic, workspace_size, ctx);
 #else
     using search1 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
-    data_algo = search1::Find<T>(args1, false, deterministic, ctx);
-    workspace_size =
-        std::max(workspace_size, search1::GetWorkspaceSize(args1, data_algo));
+    fwd_result = search1::Find<T>(args1, false, deterministic, ctx);
+    workspace_size = std::max(
+        workspace_size, search1::GetWorkspaceSize(args1, fwd_result.algo));
 #endif
   }
 
@@ -245,14 +247,14 @@ void ConvTransposeGradRawGPUDNNKernel(const Context& ctx,
     using search2 =
         paddle::operators::SearchAlgorithm<miopenConvBwdWeightsAlgorithm_t>;
     workspace_size = std::max(workspace_size, search2::GetWorkspaceSize(args2));
-    filter_algo =
+    filter_result.algo =
         search2::Find<T>(args2, false, deterministic, workspace_size, ctx);
 #else
     using search2 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t>;
-    filter_algo = search2::Find<T>(args2, false, deterministic, ctx);
-    workspace_size =
-        std::max(workspace_size, search2::GetWorkspaceSize(args2, filter_algo));
+    filter_result = search2::Find<T>(args2, false, deterministic, ctx);
+    workspace_size = std::max(
+        workspace_size, search2::GetWorkspaceSize(args2, filter_result.algo));
 #endif
   }
 
@@ -278,7 +280,7 @@ void ConvTransposeGradRawGPUDNNKernel(const Context& ctx,
                                               args1.wdesc.desc(),
                                               filter_data + filter_offset * g,
                                               args1.cdesc.desc(),
-                                              data_algo,
+                                              fwd_result.algo,
                                               &beta,
                                               args1.odesc.desc(),
                                               dx_data + x_offset * g,
@@ -295,7 +297,7 @@ void ConvTransposeGradRawGPUDNNKernel(const Context& ctx,
                                              args1.wdesc.desc(),
                                              filter_data + filter_offset * g,
                                              args1.cdesc.desc(),
-                                             data_algo,
+                                             fwd_result.algo,
                                              cudnn_workspace,
                                              workspace_size,
                                              &beta,
@@ -338,7 +340,7 @@ void ConvTransposeGradRawGPUDNNKernel(const Context& ctx,
             args2.idesc.desc(),
             dout_data + dout_offset * g,
             args2.cdesc.desc(),
-            filter_algo,
+            filter_result.algo,
             &beta,
             args2.wdesc.desc(),
             dfilter_data + filter_offset * g,
@@ -355,7 +357,7 @@ void ConvTransposeGradRawGPUDNNKernel(const Context& ctx,
             args2.odesc.desc(),
             x_data + x_offset * g,
             args2.cdesc.desc(),
-            filter_algo,
+            filter_result.algo,
             cudnn_workspace,
             workspace_size,
             &beta,
@@ -653,22 +655,17 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                                     dilations_,
                                     dtype};
 #ifdef PADDLE_WITH_HIP
-  miopenConvBwdDataAlgorithm_t bwd_algo1 =
-      static_cast<miopenConvBwdDataAlgorithm_t>(0);
-  miopenConvBwdDataAlgorithm_t bwd_algo2 =
-      static_cast<miopenConvBwdDataAlgorithm_t>(0);
-  miopenConvFwdAlgorithm_t data_algo = static_cast<miopenConvFwdAlgorithm_t>(0);
-  miopenConvBwdWeightsAlgorithm_t filter_algo =
-      static_cast<miopenConvBwdWeightsAlgorithm_t>(0);
+  paddle::operators::SearchResult<miopenConvBwdDataAlgorithm_t> bwd_result1;
+  paddle::operators::SearchResult<miopenConvBwdDataAlgorithm_t> bwd_result2;
+  paddle::operators::SearchResult<miopenConvBwdWeightsAlgorithm_t>
+      filter_result;
+  paddle::operators::SearchResult<miopenConvFwdAlgorithm_t> fwd_result;
 #else
-  cudnnConvolutionBwdDataAlgo_t bwd_algo1 =
-      static_cast<cudnnConvolutionBwdDataAlgo_t>(0);
-  cudnnConvolutionBwdDataAlgo_t bwd_algo2 =
-      static_cast<cudnnConvolutionBwdDataAlgo_t>(0);
-  cudnnConvolutionFwdAlgo_t data_algo =
-      static_cast<cudnnConvolutionFwdAlgo_t>(0);
-  cudnnConvolutionBwdFilterAlgo_t filter_algo =
-      static_cast<cudnnConvolutionBwdFilterAlgo_t>(0);
+  paddle::operators::SearchResult<cudnnConvolutionBwdDataAlgo_t> bwd_result1;
+  paddle::operators::SearchResult<cudnnConvolutionBwdDataAlgo_t> bwd_result2;
+  paddle::operators::SearchResult<cudnnConvolutionBwdFilterAlgo_t>
+      filter_result;
+  paddle::operators::SearchResult<cudnnConvolutionFwdAlgo_t> fwd_result;
 #endif
 
   auto layout = paddle::platform::GetCudnnTensorFormat(GPUDNNDataLayout::kNCHW);
@@ -696,13 +693,13 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
     using search1 =
         paddle::operators::SearchAlgorithm<miopenConvBwdDataAlgorithm_t>;
     workspace_size = search1::GetWorkspaceSize(args1);
-    bwd_algo1 =
+    bwd_result1.algo =
         search1::Find<T>(args1, false, deterministic, workspace_size, ctx);
 #else
     using search1 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
-    bwd_algo1 = search1::Find<T>(args1, false, deterministic, ctx);
-    workspace_size = search1::GetWorkspaceSize(args1, bwd_algo1);
+    bwd_result1 = search1::Find<T>(args1, false, deterministic, ctx);
+    workspace_size = search1::GetWorkspaceSize(args1, bwd_result1.algo);
 #endif
 
     ddfilter_ = ddfilter.data<T>();
@@ -720,14 +717,14 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
     using search2 =
         paddle::operators::SearchAlgorithm<miopenConvBwdDataAlgorithm_t>;
     workspace_size = std::max(workspace_size, search2::GetWorkspaceSize(args2));
-    bwd_algo2 =
+    bwd_result2.algo =
         search2::Find<T>(args2, false, deterministic, workspace_size, ctx);
 #else
     using search2 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
-    bwd_algo2 = search2::Find<T>(args2, false, deterministic, ctx);
-    workspace_size =
-        std::max(workspace_size, search2::GetWorkspaceSize(args2, bwd_algo2));
+    bwd_result2 = search2::Find<T>(args2, false, deterministic, ctx);
+    workspace_size = std::max(
+        workspace_size, search2::GetWorkspaceSize(args2, bwd_result2.algo));
 #endif
   }
 
@@ -736,9 +733,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
     args3.handle = handle;
     args3.idesc.set(transformed_dout, iwo_group);
     args3.wdesc.set(*dfilter, layout, iwo_group);
-
     args3.odesc.set(transformed_ddx_channel, iwo_group);
-
     args3.cdesc.set(dtype,
                     padding_common,
                     strides,
@@ -749,14 +744,14 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
     using search3 =
         paddle::operators::SearchAlgorithm<miopenConvBwdWeightsAlgorithm_t>;
     workspace_size = std::max(workspace_size, search3::GetWorkspaceSize(args3));
-    filter_algo =
+    filter_result.algo =
         search3::Find<T>(args3, false, deterministic, workspace_size, ctx);
 #else
     using search3 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t>;
-    filter_algo = search3::Find<T>(args3, false, deterministic, ctx);
-    workspace_size =
-        std::max(workspace_size, search3::GetWorkspaceSize(args3, filter_algo));
+    filter_result = search3::Find<T>(args3, false, deterministic, ctx);
+    workspace_size = std::max(
+        workspace_size, search3::GetWorkspaceSize(args3, filter_result.algo));
 #endif
   }
 
@@ -777,14 +772,14 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
     using search4 =
         paddle::operators::SearchAlgorithm<miopenConvFwdAlgorithm_t>;
     workspace_size = std::max(workspace_size, search4::GetWorkspaceSize(args4));
-    data_algo =
+    fwd_result.algo =
         search4::Find<T>(args4, false, deterministic, workspace_size, ctx);
 #else
     using search4 =
         paddle::operators::SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
-    data_algo = search4::Find<T>(args4, false, deterministic, ctx);
-    workspace_size =
-        std::max(workspace_size, search4::GetWorkspaceSize(args4, data_algo));
+    fwd_result = search4::Find<T>(args4, false, deterministic, ctx);
+    workspace_size = std::max(
+        workspace_size, search4::GetWorkspaceSize(args4, fwd_result.algo));
 #endif
   }
 
@@ -831,7 +826,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                 args1.wdesc.desc(),
                 filter_ + i * group_offset_filter,
                 args1.cdesc.desc(),
-                bwd_algo1,
+                bwd_result1.algo,
                 &beta,
                 args1.idesc.desc(),
                 transformed_ddout_channel_ + i * group_offset_out,
@@ -850,7 +845,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                 args1.odesc.desc(),
                 ddx_ + i * group_offset_in,
                 args1.cdesc.desc(),
-                bwd_algo1,
+                bwd_result1.algo,
                 workspace_ptr,
                 workspace_size,
                 &beta,
@@ -877,7 +872,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                 args2.wdesc.desc(),
                 ddfilter_ + i * group_offset_filter,
                 args2.cdesc.desc(),
-                bwd_algo2,
+                bwd_result2.algo,
                 &beta,
                 args2.idesc.desc(),
                 conv_x_ddfilter_data + i * group_offset_out,
@@ -908,7 +903,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                 args2.odesc.desc(),
                 x_ + i * group_offset_in,
                 args2.cdesc.desc(),
-                bwd_algo2,
+                bwd_result2.algo,
                 workspace_ptr,
                 workspace_size,
                 &alpha,
@@ -964,7 +959,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                     args3.idesc.desc(),
                     transformed_dout_channel_ + i * group_offset_out,
                     args3.cdesc.desc(),
-                    filter_algo,
+                    filter_result.algo,
                     &beta,
                     args3.wdesc.desc(),
                     dfilter_ + i * group_offset_filter,
@@ -983,7 +978,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                 args3.odesc.desc(),
                 ddx_ + i * group_offset_in,
                 args3.cdesc.desc(),
-                filter_algo,
+                filter_result.algo,
                 workspace_ptr,
                 workspace_size,
                 &beta,
@@ -1009,7 +1004,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                 args4.wdesc.desc(),
                 ddfilter_ + i * group_offset_filter,
                 args4.cdesc.desc(),
-                data_algo,
+                fwd_result.algo,
                 &beta,
                 args4.odesc.desc(),
                 transformed_dx_ + i * group_offset_in,
@@ -1028,7 +1023,7 @@ void Conv2dTransposeDoubleGradGPUDNNKernel(
                 args4.wdesc.desc(),
                 ddfilter_ + i * group_offset_filter,
                 args4.cdesc.desc(),
-                data_algo,
+                fwd_result.algo,
                 workspace_ptr,
                 workspace_size,
                 &beta,

paddle/phi/kernels/gpudnn/conv_transpose_kernel.cu

@@ -217,16 +217,19 @@ void ConvTransposeRawGPUDNNKernel(const Context& ctx,
                  c_groups);
 
 #ifdef PADDLE_WITH_HIP
+  paddle::operators::SearchResult<miopenConvBwdDataAlgorithm_t> bwd_result;
   using search =
       paddle::operators::SearchAlgorithm<miopenConvBwdDataAlgorithm_t>;
   workspace_size = std::max(workspace_size, search::GetWorkspaceSize(args));
-  algo = search::Find<T>(args, false, deterministic, workspace_size, ctx);
+  bwd_result.algo =
+      search::Find<T>(args, false, deterministic, workspace_size, ctx);
 #else
+  paddle::operators::SearchResult<cudnnConvolutionBwdDataAlgo_t> bwd_result;
   using search =
       paddle::operators::SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
-  algo = search::Find<T>(args, false, deterministic, ctx);
+  bwd_result = search::Find<T>(args, false, deterministic, ctx);
   workspace_size =
-      std::max(workspace_size, search::GetWorkspaceSize(args, algo));
+      std::max(workspace_size, search::GetWorkspaceSize(args, bwd_result.algo));
 #endif
 
   // ------------------- cudnn conv transpose forward ---------------------
@@ -247,7 +250,7 @@ void ConvTransposeRawGPUDNNKernel(const Context& ctx,
           args.wdesc.desc(),
           filter_data + filter_offset * g,
           args.cdesc.desc(),
-          algo,
+          bwd_result.algo,
           &beta,
           args.idesc.desc(),
           transformed_out_data + out_offset * g,
@@ -264,7 +267,7 @@ void ConvTransposeRawGPUDNNKernel(const Context& ctx,
           args.odesc.desc(),
           x_data + x_offset * g,
           args.cdesc.desc(),
-          algo,
+          bwd_result.algo,
           cudnn_workspace,
           workspace_size,
           &beta,

paddle/phi/kernels/impl/conv_cudnn_impl.h

@@ -36,7 +36,7 @@
 #include "paddle/phi/kernels/funcs/batch_norm_utils.h"
 
 DECLARE_bool(cudnn_deterministic);
-DECLARE_uint64(conv_workspace_size_limit);
+DECLARE_int64(conv_workspace_size_limit);
 DECLARE_bool(cudnn_exhaustive_search);
 
 namespace phi {

python/paddle/fluid/tests/unittests/test_switch_autotune.py

@@ -43,6 +43,16 @@ def static_program(net, data):
     return loss
 
 
+def set_flags(enable_autotune):
+    if paddle.is_compiled_with_cuda():
+        if enable_autotune:
+            paddle.set_flags({'FLAGS_conv_workspace_size_limit': -1})
+            paddle.set_flags({'FLAGS_cudnn_exhaustive_search': 1})
+        else:
+            paddle.set_flags({'FLAGS_conv_workspace_size_limit': 512})
+            paddle.set_flags({'FLAGS_cudnn_exhaustive_search': 0})
+
+
 class TestAutoTune(unittest.TestCase):
     def test_autotune(self):
         paddle.fluid.core.disable_autotune()
@@ -61,6 +71,7 @@ class TestAutoTune(unittest.TestCase):
 
 class TestDygraphAutoTuneStatus(TestAutoTune):
     def run_program(self, enable_autotune):
+        set_flags(enable_autotune)
         if enable_autotune:
             paddle.fluid.core.enable_autotune()
         else:
@@ -107,6 +118,7 @@ class TestDygraphAutoTuneStatus(TestAutoTune):
 class TestStaticAutoTuneStatus(TestAutoTune):
     def run_program(self, enable_autotune):
         paddle.enable_static()
+        set_flags(enable_autotune)
         if enable_autotune:
             paddle.fluid.core.enable_autotune()
         else: