feat(dnn): add heuristic cache

GitOrigin-RevId: 35e942b5e3

dnn/include/megdnn/heuristic_cache.h

@@ -0,0 +1,92 @@
+/**
+ * \file dnn/include/megdnn/heuristic_cache.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#pragma once
+
+#include "megdnn/basic_types.h"
+#include "megdnn/oprs/base.h"
+
+#include <mutex>
+#include <string>
+#include <unordered_map>
+
+namespace megdnn {
+
+class HeuristicCache {
+private:
+    HeuristicCache() = default;
+
+public:
+    static HeuristicCache& instance();
+
+    struct KeyStorage {
+        std::string category;
+        std::string input;
+
+        bool operator==(const KeyStorage& k) const {
+            return category == k.category && input == k.input;
+        }
+    };
+
+    class Key {
+        Handle* m_handle;
+        uint32_t m_opr_type;
+        const TensorLayout* m_inp_layouts_ptr;
+        size_t m_inp_layouts_size;
+        const void* m_param_ptr;
+        size_t m_param_size;
+
+        mutable std::string m_category;
+        mutable std::string m_input;
+
+    public:
+        Key(Handle* opr_handle, Algorithm::OprType opr_type, const TensorLayout* inp_layouts_ptr,
+            size_t inp_layouts_size, const void* param_ptr = nullptr, size_t param_size = 0)
+                : m_handle{opr_handle},
+                  m_opr_type{static_cast<uint32_t>(opr_type)},
+                  m_inp_layouts_ptr{inp_layouts_ptr},
+                  m_inp_layouts_size{inp_layouts_size},
+                  m_param_ptr{param_ptr},
+                  m_param_size{param_size} {}
+
+        KeyStorage build_key_storage() const;
+    };
+
+    struct Result {
+        ExecutionPolicy policy;
+        size_t workspace;
+    };
+
+    void put(const Key& key, Result& result);
+
+    Result get(const Key& key);
+
+    void clear();
+
+private:
+    struct Hash {
+        size_t operator()(const KeyStorage& k) const {
+            size_t h1 = std::hash<std::string>{}(k.category);
+            size_t h2 = std::hash<std::string>{}(k.input);
+            h1 ^= h2 + 0x9e3779b9 + (h1 << 6) + (h1 >> 2);
+            return h1;
+        }
+    };
+    std::unordered_map<KeyStorage, Result, Hash> m_heuristic_cache;
+#if __DEPLOY_ON_XP_SP2__
+    size_t m_mtx;
+#else
+    std::mutex m_mtx;
+#endif
+};
+
+}  // namespace megdnn

dnn/include/megdnn/oprs/linalg.h

@@ -42,6 +42,10 @@ public:
                                           const TensorLayout& B,
                                           const TensorLayout& C) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::BATCHED_MATRIX_MUL_FORWARD;
+    }
+
 protected:
     void check_exec(const TensorLayout& A, const TensorLayout& B,
                     const TensorLayout& C, size_t workspace_in_bytes);
@@ -76,6 +80,11 @@ public:
                                           const TensorLayout& C) = 0;
 
     static size_t pack_size (const Param::Format format);
+
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::MATRIX_MUL_FORWARD;
+    }
+
 protected:
     void check_exec(const TensorLayout& A, const TensorLayout& B,
                     const TensorLayout& C, size_t workspace_in_bytes);

dnn/include/megdnn/oprs/nn.h

@@ -275,6 +275,10 @@ public:
             const TensorLayout& src, const TensorLayout& filter,
             const TensorLayout& dst) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::CONVOLUTION_FORWARD;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(
             const TensorLayout& src, const TensorLayout& filter,
@@ -309,6 +313,10 @@ public:
     void deduce_layout(const TensorLayout& filter, const TensorLayout& diff,
                        TensorLayout& grad);
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::CONVOLUTION_BACKWARD_DATA;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& filter,
                                    const TensorLayout& diff,
@@ -338,6 +346,10 @@ public:
                                           const TensorLayout& diff,
                                           const TensorLayout& grad) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::CONVOLUTION_BACKWARD_FILTER;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& src,
                                    const TensorLayout& diff,
@@ -505,6 +517,10 @@ public:
             const ConvBiasForward::BiasMode bias_mode,
             const param::ConvBias::NonlineMode nonline_mode);
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::CONVBIAS_FORWARD;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(
             const TensorLayout& src, const TensorLayout& filter,
@@ -775,6 +791,10 @@ public:
     virtual size_t get_workspace_in_bytes(const TensorLayout& src,
                                           const TensorLayout& dst) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::POOLING_FORWARD;
+    }
+
 protected:
     void check_exec(const TensorLayout& src, const TensorLayout& dst,
                     size_t workspace_in_bytes);
@@ -801,6 +821,10 @@ public:
                                           const TensorLayout& diff,
                                           const TensorLayout& grad) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::POOLING_BACKWARD;
+    }
+
 protected:
     void check_exec(const TensorLayout& src, const TensorLayout& dst,
                     const TensorLayout& diff, const TensorLayout& grad,
@@ -1216,6 +1240,10 @@ public:
                                           const TensorLayout& filter,
                                           const TensorLayout& dst) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::CONVOLUTION3D_FORWARD;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& src,
                                    const TensorLayout& filter,
@@ -1244,6 +1272,10 @@ public:
     void deduce_layout(const TensorLayout& filter, const TensorLayout& diff,
                        TensorLayout& grad);
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::CONVOLUTION3D_BACKWARD_DATA;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& filter,
                                    const TensorLayout& diff,
@@ -1268,6 +1300,10 @@ public:
                                           const TensorLayout& diff,
                                           const TensorLayout& grad) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::CONVOLUTION3D_BACKWARD_FILTER;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& src,
                                    const TensorLayout& diff,
@@ -1308,6 +1344,10 @@ public:
                                           const TensorLayout& filter,
                                           const TensorLayout& dst) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::LOCAL_SHARE_FORWARD;
+    }
+
 protected:
     void check_exec(const TensorLayout& src, const TensorLayout& filter,
                     const TensorLayout& dst, size_t workspace_in_bytes);
@@ -1334,6 +1374,10 @@ public:
     void deduce_layout(const TensorLayout& filter, const TensorLayout& diff,
                        TensorLayout& grad);
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::LOCAL_SHARE_BACKWARD_DATA;
+    }
+
 protected:
     void check_exec(const TensorLayout& filter, const TensorLayout& diff,
                     const TensorLayout& grad, size_t workspace_in_bytes);
@@ -1358,6 +1402,10 @@ public:
                                           const TensorLayout& diff,
                                           const TensorLayout& grad) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::LOCAL_SHARE_BACKWARD_FILTER;
+    }
+
 protected:
     void check_exec(const TensorLayout& src, const TensorLayout& diff,
                     const TensorLayout& grad, size_t workspace_in_bytes);
@@ -1479,6 +1527,10 @@ public:
                                           const TensorLayout& mask,
                                           const TensorLayout& dst) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::DEFORMABLE_CONV_FORWARD;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& im,
                                    const TensorLayout& filter,
@@ -1520,6 +1572,10 @@ public:
                        const TensorLayout& mask, const TensorLayout& out_grad,
                        TensorLayout& filter_grad);
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::DEFORMABLE_CONV_BACKWARD_FILTER;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& im,
                                    const TensorLayout& offset,
@@ -1566,6 +1622,10 @@ public:
                        const TensorLayout& out_grad, TensorLayout& im_grad,
                        TensorLayout& offset_grad, TensorLayout& mask_grad);
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::DEFORMABLE_CONV_BACKWARD_DATA;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(
             const TensorLayout& im, const TensorLayout& filter,
@@ -1677,6 +1737,10 @@ public:
                                           const TensorLayout& z,
                                           const TensorLayout& dst) = 0;
 
+    static Algorithm::OprType get_opr_type() {
+        return Algorithm::OprType::BATCH_CONV_FORWARD;
+    }
+
 protected:
     CanonizedFilterMeta check_exec(const TensorLayout& src,
                                    const TensorLayout& filter,

dnn/src/arm_common/pooling/opr_impl.cpp

@@ -101,6 +101,15 @@ PoolingImpl::PoolingKernParam PoolingImpl::make_pooling_kern_param(
 
 size_t PoolingImpl::get_workspace_in_bytes(const TensorLayout& src,
                                            const TensorLayout& dst) {
+    TensorLayoutArray layouts{src, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     auto param = make_pooling_kern_szie_param(this, src, dst);
     auto algo = get_algorithm(this, src, dst);
     if (!is_fallback_algo(algo)) {

dnn/src/common/algo_chooser.h

@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #pragma once
@@ -17,10 +18,28 @@
 #include <vector>
 
 #include "megdnn/common.h"
+#include "megdnn/heuristic_cache.h"
 #include "utils.h"
 
 namespace megdnn {
 
+template <class Opr, typename... Args>
+size_t get_dnn_workspace(Opr* opr, Args&&... args) {
+    TensorLayoutArray layouts{{args...}};
+    HeuristicCache::Key key{opr->handle(), opr->get_opr_type(),
+                            layouts.data(), layouts.size(), &opr->param(),
+                            sizeof(opr->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
+    typename Opr::AlgoBase::SizeArgs size_args(opr,
+                                               std::forward<Args>(args)...);
+    return get_algorithm(opr, std::forward<Args>(args)...)
+            ->get_workspace_in_bytes(size_args);
+}
+
 /*!
  * \brief get user-configured algorithm, or heuristic algorithm
  */
@@ -31,9 +50,20 @@ typename Opr::AlgoBase* get_algorithm(Opr* opr, Args&&... args) {
     if (set.valid()) {
         ret = set;
     } else {
-        ret = opr->get_algorithm_info_heuristic(
-                std::forward<Args>(args)..., std::numeric_limits<size_t>::max(),
-                AlgoAttribute::DEFAULT, AlgoAttribute::DEFAULT).desc;
+        TensorLayoutArray layouts{{args...}};
+        HeuristicCache::Key key{opr->handle(), opr->get_opr_type(),
+                                layouts.data(), layouts.size(), &opr->param(),
+                                sizeof(opr->param())};
+        auto rst = HeuristicCache::instance().get(key);
+        if (rst.policy.algo.valid()) {
+            ret = rst.policy.algo;
+        } else {
+            ret = opr->get_algorithm_info_heuristic(
+                             std::forward<Args>(args)...,
+                             std::numeric_limits<size_t>::max(),
+                             AlgoAttribute::DEFAULT, AlgoAttribute::DEFAULT)
+                          .desc;
+        }
     }
     return static_cast<typename Opr::AlgoBase*>(
             opr->get_algorithm_from_desc(ret));

dnn/src/common/deformable_conv.cpp

@@ -250,13 +250,9 @@ CanonizedFilterMeta DeformableConvBackwardData::check_exec(
     megdnn_assert_eq_dtype(im, mask_grad);
 
     // check layout
-    megdnn_assert(im.shape == im_grad.shape, "invalid im_grad shape: %s",
-                  megdnn_layout_msg(im_grad).c_str());
-    megdnn_assert(offset.shape == offset_grad.shape,
-                  "invalid offset_grad shape: %s",
-                  megdnn_layout_msg(offset_grad).c_str());
-    megdnn_assert(mask.shape == mask_grad.shape, "invalid mask_grad shape: %s",
-                  megdnn_layout_msg(mask_grad).c_str());
+    megdnn_assert_eq_shape(im, im_grad);
+    megdnn_assert_eq_shape(offset, offset_grad);
+    megdnn_assert_eq_shape(mask, mask_grad);
 
     auto ret = make_canonized_filter_meta(im.ndim, filter, offset);
     auto required_workspace_in_bytes =

dnn/src/common/heuristic_cache.cpp

@@ -0,0 +1,142 @@
+/**
+ * \file dnn/src/common/heuristic_cache.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megdnn/heuristic_cache.h"
+#include "src/common/utils.h"
+#include "src/naive/handle.h"
+
+#if MEGDNN_WITH_CUDA
+#include "src/cuda/utils.h"
+#endif
+
+#if MEGDNN_WITH_ROCM
+#include "hcc_detail/hcc_defs_prologue.h"
+#include "megcore_rocm.h"
+#include "src/rocm/utils.h"
+#endif
+
+using namespace megdnn;
+
+HeuristicCache& HeuristicCache::instance() {
+    static HeuristicCache ins;
+    return ins;
+}
+
+HeuristicCache::KeyStorage HeuristicCache::Key::build_key_storage() const {
+    auto&& ctg = m_category;
+    auto&& inp = m_input;
+
+    if (!m_category.empty() && !m_input.empty())
+        return {ctg, inp};
+
+    inp.reserve(sizeof(TensorLayout) * 3 * m_inp_layouts_size + m_param_size);
+    for (size_t i = 0; i < m_inp_layouts_size; i++) {
+        auto&& ly = m_inp_layouts_ptr[i];
+        for (size_t j = 0; j < ly.ndim; j++) {
+            if (j)
+                inp.push_back(',');
+            inp.append(std::to_string(ly.shape[j]));
+        }
+        inp.push_back(';');
+        for (size_t j = 0; j < ly.ndim; j++) {
+            if (j)
+                inp.push_back(',');
+            inp.append(std::to_string(ly.stride[j]));
+        }
+        inp.push_back(';');
+        inp.append(ly.dtype.name());
+        inp.push_back(';');
+        inp.append(ly.format.to_string().c_str());
+        inp.push_back('|');
+    }
+    if (m_param_size) {
+        inp.append(reinterpret_cast<const char*>(m_param_ptr), m_param_size);
+    }
+
+    ctg = "plat:";
+    ctg.append(std::to_string(static_cast<uint32_t>(m_handle->type())));
+    switch (m_handle->type()) {
+#if MEGDNN_WITH_CUDA
+        case Handle::HandleType::CUDA: {
+            int cuda_rt = -1;
+            cuda_check(cudaRuntimeGetVersion(&cuda_rt));
+            cuda_rt /= 1000;
+            auto&& handle = static_cast<megdnn::cuda::HandleImpl*>(m_handle);
+            auto&& prop = handle->device_prop();
+            ctg.append(ssprintf(";dev=%s;cap=%d.%d;runtime=%d;",
+                            prop.name, prop.major, prop.minor, cuda_rt));
+            break;
+        }
+#endif
+#if MEGDNN_WITH_ROCM
+        case Handle::HandleType::ROCM: {
+            auto&& handle = static_cast<megdnn::rocm::HandleImpl*>(m_handle);
+            auto&& prop = handle->device_prop();
+            int drv = -1, hip_rt = -1;
+            hip_check(hipDriverGetVersion(&drv));
+            hip_check(hipRuntimeGetVersion(&hip_rt));
+            ctg.append(ssprintf(";dev=%s;cap=%d.%d,drv=%d;runtime=%d;",
+                            prop.name, prop.major, prop.minor, drv, hip_rt));
+            break;
+        }
+#endif
+        case Handle::HandleType::FALLBACK:
+#if MEGDNN_X86
+        case Handle::HandleType::X86:
+#endif
+#if MEGDNN_AARCH64 || MEGDNN_ARMV7
+        case Handle::HandleType::ARM_COMMON:
+#endif
+#if MEGDNN_AARCH64
+        case Handle::HandleType::AARCH64:
+#endif
+#if MEGDNN_ARMV7
+        case Handle::HandleType::ARMV7:
+#endif
+        {
+            size_t nr_threads =
+                    static_cast<megdnn::naive::HandleImpl*>(m_handle)
+                            ->megcore_dispatcher()
+                            ->nr_threads();
+            ctg.append(";");
+            ctg.append(std::to_string(nr_threads));
+            ctg.append(";");
+            break;
+        }
+        default:
+            ctg.append(";");
+    }
+    ctg.append(std::to_string(m_opr_type));
+    return {ctg, inp};
+}
+
+void HeuristicCache::put(const Key& key, Result& result) {
+    MEGDNN_LOCK_GUARD(m_mtx);
+    if (result.policy.algo.valid())
+        m_heuristic_cache[key.build_key_storage()] = result;
+}
+
+HeuristicCache::Result HeuristicCache::get(const Key& key) {
+    MEGDNN_LOCK_GUARD(m_mtx);
+    KeyStorage ks = key.build_key_storage();
+    auto iter = m_heuristic_cache.find(ks);
+    if (iter == m_heuristic_cache.end()) {
+        return {};
+    } else {
+        return iter->second;
+    }
+}
+
+void HeuristicCache::clear() {
+    MEGDNN_LOCK_GUARD(m_mtx);
+    m_heuristic_cache.clear();
+}

dnn/src/cuda/batch_conv_bias/opr_impl.cpp

@@ -56,9 +56,7 @@ size_t BatchConvBiasForwardImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& bias, const TensorLayout& z,
         const TensorLayout& dst) {
-    AlgoBase::SizeArgs args(this, src, filter, bias, z, dst);
-    return get_algorithm(this, src, filter, bias, z, dst)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, filter, bias, z, dst);
 }
 
 void BatchConvBiasForwardImpl::exec(_megdnn_tensor_in src,
@@ -66,10 +64,12 @@ void BatchConvBiasForwardImpl::exec(_megdnn_tensor_in src,
                                     _megdnn_tensor_in bias, _megdnn_tensor_in z,
                                     _megdnn_tensor_out dst,
                                     _megdnn_workspace workspace) {
+    check_exec(src.layout, filter.layout, bias.layout, z.layout, dst.layout,
+               workspace.size);
     AlgoBase::ExecArgs args(this, src, filter, bias, z, dst, workspace);
     auto algo = get_algorithm(this, src.layout, filter.layout, bias.layout,
                               z.layout, dst.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 const char* BatchConvBiasForwardImpl::get_algorithm_set_name() const {

dnn/src/cuda/batched_matrix_mul/opr_impl.cpp

@@ -33,13 +33,12 @@ void BatchedMatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
     AlgoBase::ExecArgs args(this, A, B, C, workspace);
     check_exec(A.layout, B.layout, C.layout, workspace.size);
     auto&& algo = megdnn::get_algorithm(this, A.layout, B.layout, C.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 size_t BatchedMatrixMulForwardImpl::get_workspace_in_bytes(
         const TensorLayout& A, const TensorLayout& B, const TensorLayout& C) {
-    AlgoBase::SizeArgs args(this, A, B, C);
-    return megdnn::get_algorithm(this, A, B, C)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, A, B, C);
 }
 
 std::vector<Algorithm*> BatchedMatrixMulForwardImpl::get_all_algorithms(

dnn/src/cuda/conv_bias/opr_impl.cpp

@@ -36,7 +36,7 @@ void ConvBiasForwardImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                             preprocessed_filter);
     auto algo = get_algorithm(this, src.layout, filter.layout, bias.layout,
                               z.layout, dst.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 };
 
 std::vector<ConvBiasForward::Algorithm*>
@@ -228,6 +228,15 @@ size_t ConvBiasForwardImpl::get_workspace_in_bytes(
         const TensorLayout& bias, const TensorLayout& z,
         const TensorLayout& dst,
         const PreprocessedFilter* preprocessed_filter) {
+    TensorLayoutArray layouts{src, filter, bias, z, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     AlgoBase::SizeArgs args{
             this, src, filter, bias, z, dst, preprocessed_filter};
     return get_algorithm(this, src, filter, bias, z, dst)

dnn/src/cuda/convolution/opr_impl.cpp

@@ -58,9 +58,7 @@ size_t ConvolutionForwardImpl::get_workspace_in_bytes(
         const TensorLayout& dst,
         const PreprocessedFilter* preprocessed_filter) {
     MEGDNN_MARK_USED_VAR(preprocessed_filter);
-    AlgoBase::SizeArgs args{this, src, filter, dst};
-    return megdnn::get_algorithm(this, src, filter, dst)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, filter, dst);
 }
 
 void ConvolutionForwardImpl::exec(_megdnn_tensor_in src,
@@ -72,7 +70,7 @@ void ConvolutionForwardImpl::exec(_megdnn_tensor_in src,
                preprocessed_filter);
     AlgoBase::ExecArgs args(this, src, filter, dst, workspace);
     auto&& algo = get_algorithm(this, src.layout, filter.layout, dst.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 const char* ConvolutionForwardImpl::get_algorithm_set_name() const {
@@ -85,9 +83,10 @@ void ConvolutionBackwardDataImpl::exec(_megdnn_tensor_in filter,
                                        _megdnn_tensor_in diff,
                                        _megdnn_tensor_out grad,
                                        _megdnn_workspace workspace) {
+    check_exec(filter.layout, diff.layout, grad.layout, workspace.size);
     AlgoBase::ExecArgs args(this, filter, diff, grad, workspace);
     auto algo = get_algorithm(this, filter.layout, diff.layout, grad.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 std::vector<ConvolutionBackwardDataImpl::Algorithm*>
@@ -196,9 +195,7 @@ ConvolutionBackwardDataImpl::get_algorithm_heuristic(
 size_t ConvolutionBackwardDataImpl::get_workspace_in_bytes(
         const TensorLayout& filter, const TensorLayout& diff,
         const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, filter, diff, grad);
-    return get_algorithm(this, filter, diff, grad)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, filter, diff, grad);
 }
 
 const char* ConvolutionBackwardDataImpl::get_algorithm_set_name() const {
@@ -211,9 +208,10 @@ void ConvolutionBackwardFilterImpl::exec(_megdnn_tensor_in src,
                                          _megdnn_tensor_in diff,
                                          _megdnn_tensor_out grad,
                                          _megdnn_workspace workspace) {
+    check_exec(src.layout, diff.layout, grad.layout, workspace.size);
     AlgoBase::ExecArgs args(this, src, diff, grad, workspace);
     auto algo = get_algorithm(this, src.layout, diff.layout, grad.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 std::vector<ConvolutionBackwardFilterImpl::Algorithm*>
@@ -324,9 +322,7 @@ ConvolutionBackwardFilterImpl::get_algorithm_heuristic(
 size_t ConvolutionBackwardFilterImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& diff,
         const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, src, diff, grad);
-    return get_algorithm(this, src, diff, grad)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, diff, grad);
 }
 
 const char* ConvolutionBackwardFilterImpl::get_algorithm_set_name() const {

dnn/src/cuda/convolution3d/opr_impl.cpp

@@ -111,18 +111,17 @@ Convolution3DForwardImpl::get_all_algorithms(const TensorLayout& src,
 size_t Convolution3DForwardImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst) {
-    AlgoBase::SizeArgs args(this, src, filter, dst);
-    return get_algorithm(this, src, filter, dst)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, filter, dst);
 }
 
 void Convolution3DForwardImpl::exec(_megdnn_tensor_in src,
                                     _megdnn_tensor_in filter,
                                     _megdnn_tensor_out dst,
                                     _megdnn_workspace workspace) {
+    check_exec(src.layout, filter.layout, dst.layout, workspace.size);
     AlgoBase::ExecArgs args(this, src, filter, dst, workspace);
     auto algo = get_algorithm(this, src.layout, filter.layout, dst.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 const char* Convolution3DForwardImpl::get_algorithm_set_name() const {
@@ -133,9 +132,10 @@ void Convolution3DBackwardDataImpl::exec(_megdnn_tensor_in filter,
                                          _megdnn_tensor_in diff,
                                          _megdnn_tensor_out grad,
                                          _megdnn_workspace workspace) {
+    check_exec(filter.layout, diff.layout, grad.layout, workspace.size);
     AlgoBase::ExecArgs args(this, filter, diff, grad, workspace);
     auto algo = get_algorithm(this, filter.layout, diff.layout, grad.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 std::vector<Convolution3DBackwardDataImpl::Algorithm*>
@@ -200,9 +200,7 @@ Convolution3DBackwardDataImpl::get_algorithm_heuristic(
 size_t Convolution3DBackwardDataImpl::get_workspace_in_bytes(
         const TensorLayout& filter, const TensorLayout& diff,
         const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, filter, diff, grad);
-    return get_algorithm(this, filter, diff, grad)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, filter, diff, grad);
 }
 
 const char* Convolution3DBackwardDataImpl::get_algorithm_set_name() const {
@@ -213,10 +211,11 @@ void Convolution3DBackwardFilterImpl::exec(_megdnn_tensor_in src,
                                            _megdnn_tensor_in diff,
                                            _megdnn_tensor_out grad,
                                            _megdnn_workspace workspace) {
+    check_exec(src.layout, diff.layout, grad.layout, workspace.size);
     AlgoBase::ExecArgs args(this, src, diff, grad, workspace);
     auto algo =
             get_algorithm(this, src.layout, diff.layout, grad.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 std::vector<Convolution3DBackwardFilterImpl::Algorithm*>
@@ -281,9 +280,7 @@ Convolution3DBackwardFilterImpl::get_algorithm_heuristic(
 size_t Convolution3DBackwardFilterImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& diff,
         const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, src, diff, grad);
-    return get_algorithm(this, src, diff, grad)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, diff , grad);
 }
 
 const char* Convolution3DBackwardFilterImpl::get_algorithm_set_name() const {

dnn/src/cuda/deformable_conv/opr_impl.cpp

@@ -36,8 +36,7 @@ size_t Fwd::get_workspace_in_bytes(const TensorLayout& im,
                                    const TensorLayout& offset,
                                    const TensorLayout& mask,
                                    const TensorLayout& dst) {
-    auto algo = get_algorithm(this, im, filter, offset, mask, dst);
-    return algo->get_workspace_in_bytes({this, im, filter, offset, mask, dst});
+    return get_dnn_workspace(this, im, filter, offset, mask, dst);
 }
 
 std::vector<AlgoFwd*> Fwd::get_all_algorithms(const TensorLayout& /* im */,
@@ -96,13 +95,13 @@ const char* Fwd::get_algorithm_set_name() const {
 void Fwd::exec(_megdnn_tensor_in im, _megdnn_tensor_in filter,
                _megdnn_tensor_in offset, _megdnn_tensor_in mask,
                _megdnn_tensor_out out, _megdnn_workspace workspace) {
+    check_exec(im.layout, filter.layout, offset.layout, mask.layout, out.layout,
+               workspace.size);
     auto algo = get_algorithm(this, im.layout, filter.layout, offset.layout,
                               mask.layout, out.layout);
 
     AlgoBase::ExecArgs args(this, im, filter, offset, mask, out, workspace);
-
-    algo->check_workspace(args, workspace).exec(args);
-    return;
+    algo->exec(args);
 }
 
 /* ============== BwdFlt Implementation ============== */
@@ -152,21 +151,23 @@ AlgoBwdFlt* BwdFlt::get_algorithm_heuristic(
 size_t BwdFlt::get_workspace_in_bytes(
         const TensorLayout& im, const TensorLayout& offset, const TensorLayout& mask,
         const TensorLayout& out_grad, const TensorLayout& filter_grad) {
-    auto algo = get_algorithm(this, im, offset, mask, out_grad, filter_grad);
-    return algo->get_workspace_in_bytes({this, im, offset, mask, out_grad, filter_grad});
+    return get_dnn_workspace(this, im, offset, mask, out_grad, filter_grad);
 }
 
 const char* BwdFlt::get_algorithm_set_name() const {
     return "DEFORMABLE_CONV_BWD_FILTER_CUDA";
 };
 
-void BwdFlt::exec(_megdnn_tensor_in im, _megdnn_tensor_in offset, _megdnn_tensor_in mask,
-               _megdnn_tensor_in out_grad, _megdnn_tensor_out filter_grad,
-               _megdnn_workspace workspace) {
-    AlgoBase::ExecArgs args(this, im, offset, mask, out_grad, filter_grad, workspace);
-    auto algo = get_algorithm(this, im.layout, offset.layout, mask.layout, out_grad.layout,
-                              filter_grad.layout);
-    algo->check_workspace(args, workspace).exec(args);
+void BwdFlt::exec(_megdnn_tensor_in im, _megdnn_tensor_in offset,
+                  _megdnn_tensor_in mask, _megdnn_tensor_in out_grad,
+                  _megdnn_tensor_out filter_grad, _megdnn_workspace workspace) {
+    check_exec(im.layout, offset.layout, mask.layout, out_grad.layout,
+               filter_grad.layout, workspace.size);
+    AlgoBase::ExecArgs args(this, im, offset, mask, out_grad, filter_grad,
+                            workspace);
+    auto algo = get_algorithm(this, im.layout, offset.layout, mask.layout,
+                              out_grad.layout, filter_grad.layout);
+    algo->exec(args);
 }
 
 /* ============== BwdData Implementation ============== */
@@ -222,10 +223,8 @@ size_t BwdData::get_workspace_in_bytes(
         const TensorLayout& offset, const TensorLayout& mask,
         const TensorLayout& out_grad, const TensorLayout& im_grad,
         const TensorLayout& offset_grad, const TensorLayout& mask_grad) {
-    auto algo = get_algorithm(this, im, filter, offset, mask, out_grad, 
-                              im_grad, offset_grad, mask_grad);
-    return algo->get_workspace_in_bytes({this, im, filter, offset, mask, out_grad, 
-                                         im_grad, offset_grad, mask_grad});
+    return get_dnn_workspace(this, im, filter, offset, mask, out_grad, im_grad,
+                             offset_grad, mask_grad);
 }
 
 const char* BwdData::get_algorithm_set_name() const {
@@ -233,16 +232,19 @@ const char* BwdData::get_algorithm_set_name() const {
 };
 
 void BwdData::exec(_megdnn_tensor_in im, _megdnn_tensor_in filter,
-               _megdnn_tensor_in offset, _megdnn_tensor_in mask,
-               _megdnn_tensor_in out_grad, _megdnn_tensor_out im_grad,
-               _megdnn_tensor_out offset_grad, _megdnn_tensor_out mask_grad,
-               _megdnn_workspace workspace) {
+                   _megdnn_tensor_in offset, _megdnn_tensor_in mask,
+                   _megdnn_tensor_in out_grad, _megdnn_tensor_out im_grad,
+                   _megdnn_tensor_out offset_grad, _megdnn_tensor_out mask_grad,
+                   _megdnn_workspace workspace) {
+    check_exec(im.layout, filter.layout, offset.layout, mask.layout,
+               out_grad.layout, im_grad.layout, offset_grad.layout,
+               mask_grad.layout, workspace.size);
     AlgoBase::ExecArgs args(this, im, filter, offset, mask, out_grad, im_grad,
                             offset_grad, mask_grad, workspace);
     auto algo = get_algorithm(this, im.layout, filter.layout, offset.layout,
                               mask.layout, out_grad.layout, im_grad.layout,
                               offset_grad.layout, mask_grad.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 // vim: syntax=cpp.doxygen

dnn/src/cuda/local_share/opr_impl.cpp

@@ -59,17 +59,17 @@ LocalShareForwardImpl::get_all_algorithms(const TensorLayout& src,
 size_t LocalShareForwardImpl::get_workspace_in_bytes(const TensorLayout& src,
                                                      const TensorLayout& filter,
                                                      const TensorLayout& dst) {
-    AlgoBase::SizeArgs args(this, src, filter, dst);
-    return get_algorithm(this, src, filter, dst)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, filter, dst);
 }
 
 void LocalShareForwardImpl::exec(_megdnn_tensor_in src,
                                  _megdnn_tensor_in filter,
                                  _megdnn_tensor_out dst,
                                  _megdnn_workspace workspace) {
+    check_exec(src.layout, filter.layout, dst.layout, workspace.size);
     AlgoBase::ExecArgs args(this, src, filter, dst, workspace);
     auto algo = get_algorithm(this, src.layout, filter.layout, dst.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 const char* LocalShareForwardImpl::get_algorithm_set_name() const {
@@ -112,8 +112,7 @@ LocalShareBackwardDataImpl::get_all_algorithms(const TensorLayout& filter,
 size_t LocalShareBackwardDataImpl::get_workspace_in_bytes(const TensorLayout& filter,
                                                      const TensorLayout& diff,
                                                      const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, filter, diff, grad);
-    return get_algorithm(this, filter, diff, grad)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, filter, diff, grad);
 }
 
 void LocalShareBackwardDataImpl::exec(_megdnn_tensor_in filter,
@@ -166,8 +165,7 @@ LocalShareBackwardFilterImpl::get_all_algorithms(const TensorLayout& src,
 size_t LocalShareBackwardFilterImpl::get_workspace_in_bytes(const TensorLayout& src,
                                                      const TensorLayout& diff,
                                                      const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, src, diff, grad);
-    return get_algorithm(this, src, diff, grad)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, diff, grad);
 }
 
 void LocalShareBackwardFilterImpl::exec(_megdnn_tensor_in src,

dnn/src/cuda/matrix_mul/opr_impl.cpp

@@ -59,8 +59,7 @@ MatrixMulForwardImpl::Algorithm* MatrixMulForwardImpl::get_algorithm_heuristic(
 size_t MatrixMulForwardImpl::get_workspace_in_bytes(const TensorLayout& A,
                                                     const TensorLayout& B,
                                                     const TensorLayout& C) {
-    AlgoBase::SizeArgs args{this, A, B, C};
-    return megdnn::get_algorithm(this, A, B, C)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, A, B, C);
 }
 
 void MatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
@@ -69,7 +68,7 @@ void MatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
     check_exec(A.layout, B.layout, C.layout, workspace.size);
     AlgoBase::ExecArgs args(this, A, B, C, workspace);
     auto&& algo = get_algorithm(this, A.layout, B.layout, C.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 }  // namespace cuda

dnn/src/cuda/pooling/opr_impl.cpp

@@ -21,8 +21,7 @@ namespace cuda {
 
 size_t PoolingForwardImpl::get_workspace_in_bytes(const TensorLayout& src,
                                                   const TensorLayout& dst) {
-    AlgoBase::SizeArgs args(this, src, dst);
-    return get_algorithm(this, src, dst)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, dst);
 }
 
 const char* PoolingForwardImpl::get_algorithm_set_name() const {
@@ -117,9 +116,7 @@ size_t PoolingBackwardImpl::get_workspace_in_bytes(const TensorLayout& src,
                                                    const TensorLayout& dst,
                                                    const TensorLayout& diff,
                                                    const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, src, dst, diff, grad);
-    return get_algorithm(this, src, dst, diff, grad)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, dst, diff, grad);
 }
 
 }  // namespace cuda

dnn/src/fallback/batched_matrix_mul/opr_impl.cpp

@@ -44,8 +44,7 @@ BatchedMatrixMulForwardImpl::get_algorithm_heuristic(
 
 size_t BatchedMatrixMulForwardImpl::get_workspace_in_bytes(
         const TensorLayout& A, const TensorLayout& B, const TensorLayout& C) {
-    AlgoBase::SizeArgs args{this, A, B, C};
-    return megdnn::get_algorithm(this, A, B, C)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, A, B, C);
 }
 
 void BatchedMatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
@@ -54,7 +53,7 @@ void BatchedMatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
     check_exec(A.layout, B.layout, C.layout, workspace.size);
     AlgoBase::ExecArgs args(this, A, B, C, workspace);
     auto&& algo = get_algorithm(this, A.layout, B.layout, C.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 // vim: syntax=cpp.doxygen

dnn/src/fallback/conv_bias/opr_impl.cpp

@@ -224,6 +224,15 @@ size_t ConvBiasImpl::get_workspace_in_bytes(
         const TensorLayout& bias, const TensorLayout& z,
         const TensorLayout& dst,
         const PreprocessedFilter* preprocessed_filter) {
+    TensorLayoutArray layouts{src, filter, bias, z, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     auto fparam = make_ncb_kern_size_param(src, filter, bias, dst,
                                            preprocessed_filter);
     auto&& algo = get_algorithm(fparam);

dnn/src/fallback/convolution/opr_impl.cpp

@@ -146,6 +146,15 @@ size_t ConvolutionImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst,
         const PreprocessedFilter* preprocessed_filter) {
+    TensorLayoutArray layouts{src, filter, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     auto fparam =
             make_ncb_kern_size_param(src, filter, dst, preprocessed_filter);
     auto&& algo = get_algorithm(fparam);
@@ -494,6 +503,15 @@ void ConvolutionBackwardDataImpl::exec(_megdnn_tensor_in filter,
 size_t ConvolutionBackwardDataImpl::get_workspace_in_bytes(
         const TensorLayout& filter, const TensorLayout& diff,
         const TensorLayout& grad) {
+    TensorLayoutArray layouts{filter, diff, grad};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     if (param().format == param::Convolution::Format::NHWCD4 ||
         param().format == param::Convolution::Format::NCHW4 ||
         (param().format == param::Convolution::Format::NCHW &&

dnn/src/fallback/matrix_mul/opr_impl.cpp

@@ -219,6 +219,15 @@ MatrixMulImpl::KernParam MatrixMulImpl::make_kern_param(
 size_t MatrixMulImpl::get_workspace_in_bytes(const TensorLayout& A,
                                              const TensorLayout& B,
                                              const TensorLayout& C) {
+    TensorLayoutArray layouts{A, B, C};
+    HeuristicCache::Key key{this->handle(),this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     if (auto algo = get_algorithm_heuristic(
                 A, B, C, std::numeric_limits<size_t>::max(),
                 AlgoAttribute::DEFAULT, AlgoAttribute::DEFAULT)) {

dnn/src/naive/batch_conv_bias/opr_impl.cpp

@@ -15,6 +15,7 @@
 #include "src/naive/convolution/helper.h"
 
 #include <cstring>
+#include "megdnn/heuristic_cache.h"
 #include "src/common/utils.h"
 #include "src/naive/handle.h"
 
@@ -56,6 +57,14 @@ size_t BatchConvBiasForwardImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& flt,
         const TensorLayout& bias, const TensorLayout& z,
         const TensorLayout& dst) {
+    TensorLayoutArray layouts{src, flt, bias, z, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
     return get_workspace_bundle(nullptr, src, flt, bias, z, dst)
             .total_size_in_bytes();
 }

dnn/src/naive/conv_bias/opr_impl.cpp

@@ -13,6 +13,7 @@
 #include "src/naive/convolution/helper.h"
 
 #include <cstring>
+#include "megdnn/heuristic_cache.h"
 #include "megdnn/dtype.h"
 #include "src/common/conv_bias.h"
 #include "src/common/opr_delegate.h"
@@ -201,6 +202,15 @@ size_t ConvBiasForwardImpl::get_workspace_in_bytes(const TensorLayout& src,
                                                    const TensorLayout& z,
                                                    const TensorLayout& dst,
                                                    const PreprocessedFilter*) {
+    TensorLayoutArray layouts{src, flt, bias, z, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     size_t float_workspace_size = 0;
 
     if (z.ndim > 0 && z.dtype.category() != DTypeCategory::FLOAT) {

dnn/src/naive/convolution/convolution.cpp

@@ -11,7 +11,7 @@
 #include "./opr_impl.h"
 #include "./helper.h"
 
-#include "src/naive/handle.h"
+#include "megdnn/heuristic_cache.h"
 #include "src/naive/handle.h"
 #include "src/common/utils.h"
 #include "megdnn/dtype.h"
@@ -78,6 +78,15 @@ void ConvolutionForwardImpl::exec(_megdnn_tensor_in src,
 size_t ConvolutionBackwardDataImpl::get_workspace_in_bytes(const TensorLayout& filter,
                                                    const TensorLayout& diff,
                                                    const TensorLayout& grad) {
+    TensorLayoutArray layouts{filter, diff, grad};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     size_t workspace_size = 0;
     auto flt_dt = filter.dtype.enumv();
     auto grad_dt = grad.dtype.enumv();
@@ -191,6 +200,15 @@ size_t ConvolutionBackwardFilterImpl::get_workspace_in_bytes(
         const TensorLayout& grad) {
     size_t workspace_size = 0;
 #if !MEGDNN_DISABLE_FLOAT16
+    TensorLayoutArray layouts{src, diff, grad};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     auto src_dt = src.dtype.enumv();
     auto grad_dt = grad.dtype.enumv();
     auto diff_dt = diff.dtype.enumv();

dnn/src/naive/pooling/opr_impl.cpp

@@ -12,6 +12,7 @@
 #include "src/naive/pooling/opr_impl.h"
 
 #include <cstring>
+#include "megdnn/heuristic_cache.h"
 #include "megdnn/dtype.h"
 #include "src/common/utils.h"
 #include "src/naive/handle.h"
@@ -402,6 +403,14 @@ WorkspaceBundle PoolingForwardImpl::get_workspace_bundle(
 
 size_t PoolingForwardImpl::get_workspace_in_bytes(const TensorLayout& src,
                                                   const TensorLayout& dst) {
+    TensorLayoutArray layouts{src, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
     return get_workspace_bundle(nullptr, src, dst).total_size_in_bytes();
 }
 namespace {
@@ -652,6 +661,14 @@ WorkspaceBundle PoolingBackwardImpl::get_workspace_bundle(
 size_t PoolingBackwardImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& dst,
         const TensorLayout& diff, const TensorLayout& grad) {
+    TensorLayoutArray layouts{src, dst, diff, grad};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
     return get_workspace_bundle(nullptr, src, dst, diff, grad)
             .total_size_in_bytes();
 }

dnn/src/rocm/batched_matrix_mul/opr_impl.cpp

@@ -47,8 +47,7 @@ BatchedMatrixMulForwardImpl::get_algorithm_heuristic(
 
 size_t BatchedMatrixMulForwardImpl::get_workspace_in_bytes(
         const TensorLayout& A, const TensorLayout& B, const TensorLayout& C) {
-    AlgoBase::SizeArgs args{this, A, B, C};
-    return megdnn::get_algorithm(this, A, B, C)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, A, B, C);
 }
 
 void BatchedMatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
@@ -57,7 +56,7 @@ void BatchedMatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
     check_exec(A.layout, B.layout, C.layout, workspace.size);
     AlgoBase::ExecArgs args(this, A, B, C, workspace);
     auto&& algo = get_algorithm(this, A.layout, B.layout, C.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 // vim: syntax=cpp.doxygen

dnn/src/rocm/convolution/opr_impl.cpp

@@ -112,19 +112,30 @@ ConvolutionForwardImpl::get_all_algorithms(const TensorLayout& src,
 size_t ConvolutionForwardImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst, const PreprocessedFilter*) {
+    TensorLayoutArray layouts{src, filter, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     AlgoBase::SizeArgs args(this, src, filter, dst);
-    return get_algorithm(this, src, args.filter_meta, dst)
+    return get_algorithm(this, src, filter, dst)
             ->get_workspace_in_bytes(args);
 }
 
 void ConvolutionForwardImpl::exec(_megdnn_tensor_in src,
                                   _megdnn_tensor_in filter,
                                   _megdnn_tensor_out dst,
-                                  const PreprocessedFilter*,
+                                  const PreprocessedFilter* preprocessed_filter,
                                   _megdnn_workspace workspace) {
+    check_exec(src.layout, filter.layout, dst.layout, workspace.size,
+               preprocessed_filter);
     AlgoBase::ExecArgs args(this, src, filter, dst, workspace);
-    auto algo = get_algorithm(this, src.layout, args.filter_meta, dst.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    auto algo = get_algorithm(this, src.layout, filter.layout, dst.layout);
+    algo->exec(args);
 }
 
 const char* ConvolutionForwardImpl::get_algorithm_set_name() const {
@@ -137,9 +148,10 @@ void ConvolutionBackwardDataImpl::exec(_megdnn_tensor_in filter,
                                        _megdnn_tensor_in diff,
                                        _megdnn_tensor_out grad,
                                        _megdnn_workspace workspace) {
+    check_exec(filter.layout, diff.layout, grad.layout, workspace.size);
     AlgoBase::ExecArgs args(this, filter, diff, grad, workspace);
-    auto algo = get_algorithm(this, args.filter_meta, diff.layout, grad.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    auto algo = get_algorithm(this, filter.layout, diff.layout, grad.layout);
+    algo->exec(args);
 }
 
 std::vector<ConvolutionBackwardDataImpl::Algorithm*>
@@ -192,8 +204,17 @@ ConvolutionBackwardDataImpl::get_algorithm_heuristic(
 size_t ConvolutionBackwardDataImpl::get_workspace_in_bytes(
         const TensorLayout& filter, const TensorLayout& diff,
         const TensorLayout& grad) {
+    TensorLayoutArray layouts{filter, diff, grad};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     AlgoBase::SizeArgs args(this, filter, diff, grad);
-    return get_algorithm(this, args.filter_meta, diff, grad)
+    return get_algorithm(this, filter, diff, grad)
             ->get_workspace_in_bytes(args);
 }
 
@@ -207,10 +228,11 @@ void ConvolutionBackwardFilterImpl::exec(_megdnn_tensor_in src,
                                          _megdnn_tensor_in diff,
                                          _megdnn_tensor_out grad,
                                          _megdnn_workspace workspace) {
+    check_exec(src.layout, diff.layout, grad.layout, workspace.size);
     AlgoBase::ExecArgs args(this, src, diff, grad, workspace);
     auto algo =
-            get_algorithm(this, src.layout, diff.layout, args.grad_filter_meta);
-    algo->check_workspace(args, workspace).exec(args);
+            get_algorithm(this, src.layout, diff.layout, grad.layout);
+    algo->exec(args);
 }
 
 std::vector<ConvolutionBackwardFilterImpl::Algorithm*>
@@ -264,8 +286,17 @@ ConvolutionBackwardFilterImpl::get_algorithm_heuristic(
 size_t ConvolutionBackwardFilterImpl::get_workspace_in_bytes(
         const TensorLayout& src, const TensorLayout& diff,
         const TensorLayout& grad) {
+    TensorLayoutArray layouts{src, diff, grad};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     AlgoBase::SizeArgs args(this, src, diff, grad);
-    return get_algorithm(this, src, diff, args.grad_filter_meta)
+    return get_algorithm(this, src, diff, grad)
             ->get_workspace_in_bytes(args);
 }
 

dnn/src/rocm/convolution/opr_impl.h

@@ -24,7 +24,7 @@ public:
               const PreprocessedFilter* preprocessed_filter,
               _megdnn_workspace workspace) override;
     AlgorithmInfo get_algorithm_info_heuristic(
-            const TensorLayout& src, const CanonizedFilterMeta& filter,
+            const TensorLayout& src, const TensorLayout& filter,
             const TensorLayout& dst, size_t workspace_limit_in_bytes,
             const AlgoAttribute& positive_attr,
             const AlgoAttribute& negative_attr) {
@@ -95,7 +95,7 @@ public:
     void exec(_megdnn_tensor_in filter, _megdnn_tensor_in diff,
               _megdnn_tensor_out grad, _megdnn_workspace workspace) override;
     AlgorithmInfo get_algorithm_info_heuristic(
-            const CanonizedFilterMeta& filter, const TensorLayout& diff,
+            const TensorLayout& filter, const TensorLayout& diff,
             const TensorLayout& grad, size_t workspace_limit_in_bytes,
             const AlgoAttribute& positive_attr,
             const AlgoAttribute& negative_attr) {
@@ -145,7 +145,7 @@ public:
               _megdnn_tensor_out grad, _megdnn_workspace workspace) override;
     AlgorithmInfo get_algorithm_info_heuristic(
             const TensorLayout& src, const TensorLayout& diff,
-            const CanonizedFilterMeta& grad, size_t workspace_limit_in_bytes,
+            const TensorLayout& grad, size_t workspace_limit_in_bytes,
             const AlgoAttribute& positive_attr,
             const AlgoAttribute& negative_attr) {
         return get_algorithm_heuristic(src, diff, grad,

dnn/src/rocm/matrix_mul/opr_impl.cpp

@@ -44,8 +44,7 @@ MatrixMulForwardImpl::Algorithm* MatrixMulForwardImpl::get_algorithm_heuristic(
 size_t MatrixMulForwardImpl::get_workspace_in_bytes(const TensorLayout& A,
                                                     const TensorLayout& B,
                                                     const TensorLayout& C) {
-    AlgoBase::SizeArgs args{this, A, B, C};
-    return megdnn::get_algorithm(this, A, B, C)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, A, B, C);
 }
 
 void MatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
@@ -54,7 +53,7 @@ void MatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
     check_exec(A.layout, B.layout, C.layout, workspace.size);
     AlgoBase::ExecArgs args(this, A, B, C, workspace);
     auto&& algo = get_algorithm(this, A.layout, B.layout, C.layout);
-    algo->check_workspace(args, workspace).exec(args);
+    algo->exec(args);
 }
 
 // vim: syntax=cpp.doxygen

dnn/src/rocm/pooling/opr_impl.cpp

@@ -19,8 +19,7 @@ namespace rocm {
 
 size_t PoolingForwardImpl::get_workspace_in_bytes(const TensorLayout& src,
                               const TensorLayout& dst) {
-    AlgoBase::SizeArgs args(this, src, dst);
-    return get_algorithm(this, src, dst)->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, dst);
 }
 
 const char* PoolingForwardImpl::get_algorithm_set_name() const {
@@ -69,9 +68,7 @@ size_t PoolingBackwardImpl::get_workspace_in_bytes(const TensorLayout& src,
                 const TensorLayout& dst,
                 const TensorLayout& diff,
                 const TensorLayout& grad) {
-    AlgoBase::SizeArgs args(this, src, dst, diff, grad);
-    return get_algorithm(this, src, dst, diff, grad)
-            ->get_workspace_in_bytes(args);
+    return get_dnn_workspace(this, src, dst, diff, grad);
 };
 
 const char* PoolingBackwardImpl::get_algorithm_set_name() const {

dnn/src/x86/pooling/opr_impl.cpp

@@ -46,6 +46,15 @@ WorkspaceBundle megdnn::x86::get_bundle(const TensorLayout& src,
 
 size_t PoolingImpl::get_workspace_in_bytes(const TensorLayout& src,
                                            const TensorLayout& dst) {
+    TensorLayoutArray layouts{src, dst};
+    HeuristicCache::Key key{this->handle(), this->get_opr_type(),
+                            layouts.data(), layouts.size(), &this->param(),
+                            sizeof(this->param())};
+    auto rst = HeuristicCache::instance().get(key);
+    if (rst.policy.algo.valid()) {
+        return rst.workspace;
+    }
+
     auto algo = get_algorithm(this, src, dst);
     if (!is_fallback_algo(algo)) {
         if (is_supported(SIMDType::SSE) && src.dtype == dtype::Float32() &&

src/core/test/graph/misc.cpp

@@ -29,6 +29,7 @@
 #include "megbrain/plugin/profiler.h"
 
 #include "megbrain/test/helper.h"
+#include "megdnn/heuristic_cache.h"
 #include "megdnn/oprs/base.h"
 
 #include <atomic>
@@ -2075,10 +2076,12 @@ void test_free_memory_in_weight_preprocess(int record_level, CompNode cn) {
 
 TEST(TestGraph, FreeMemoryInWeightPreprocess) {
     test_free_memory_in_weight_preprocess(0, CompNode::load("xpu0"));
+    megdnn::HeuristicCache::instance().clear();
 }
 
 TEST(TestGraph, RecordFreeMemoryInWeightPreprocess) {
     test_free_memory_in_weight_preprocess(1, CompNode::load("cpu0"));
+    megdnn::HeuristicCache::instance().clear();
 }
 
 namespace {
@@ -2157,6 +2160,7 @@ TEST(TestGraph, FreeMemoryInWeightPreprocessWithValueInfer) {
                          ->cast_final_safe<opr::SharedDeviceTensor>()
                          .get_dev_tensor()
                          .empty());
+    megdnn::HeuristicCache::instance().clear();
 }
 
 TEST(TestGraph, FreeMemoryInWeightPreprocessWithMultiReader) {
@@ -2200,6 +2204,7 @@ TEST(TestGraph, FreeMemoryInWeightPreprocessWithMultiReader) {
                         ->cast_final_safe<opr::SharedDeviceTensor>()
                         .get_dev_tensor()
                         .empty());
+    megdnn::HeuristicCache::instance().clear();
 }
 
 TEST(TestGraph, FreeBias) {

src/opr/impl/search_policy/algo_chooser.cpp

@@ -24,6 +24,7 @@
 
 //! TODO: here has to be know some megdnn::opr when there is produced midout.h
 //! fix it if there is another graceful way.
+#include "megdnn/heuristic_cache.h"
 #include "megdnn/opr_param_defs.h"
 #include "megdnn/oprs.h"
 #include "megdnn/oprs/base.h"
@@ -1156,6 +1157,15 @@ template <typename Opr>
 size_t AlgoChooser<Opr>::setup_algo(const FixedTensorLayouts& layouts,
                                     Opr* megdnn_opr, const MGBOpr* mgb_opr,
                                     bool allow_weight_preprocess) {
+    HeuristicCache::Key cache_key(
+            megdnn_opr->handle(), megdnn_opr->get_opr_type(), layouts.data(),
+            layouts.size(), &megdnn_opr->param(), sizeof(megdnn_opr->param()));
+    auto rst = HeuristicCache::instance().get(cache_key);
+    if (rst.policy.algo.valid()) {
+        megdnn_opr->execution_policy() = rst.policy;
+        return rst.workspace;
+    }
+
     if (WorkspaceLimitGetter::is_prealloc_run(mgb_opr->owner_graph())) {
         return 0;
     }
@@ -1192,6 +1202,11 @@ size_t AlgoChooser<Opr>::setup_algo(const FixedTensorLayouts& layouts,
     mgb_log_debug("%s", ret.c_str());
 
     megdnn_opr->execution_policy() = policy;
+
+    if (mgb_opr->execution_policy().strategy & ExecutionStrategy::HEURISTIC) {
+        HeuristicCache::Result cache_result{policy, workspace};
+        HeuristicCache::instance().put(cache_key, cache_result);
+    }
     return workspace;
 }
 

src/opr/test/algo_chooser.cpp

@@ -22,6 +22,7 @@
 #include "megbrain/opr/tensor_manip.h"
 #include "megdnn/oprs/base.h"
 #include "megdnn/dtype.h"
+#include "megdnn/heuristic_cache.h"
 
 #include <cmath>
 #include <random>
@@ -337,6 +338,7 @@ void test_no_profiling_on_shape_change(const TensorShapeArray& inps0,
 
 TEST(TestOprDNN, FastrunNoProfilingOnShapeChange) {
     REQUIRE_GPU(1);
+    megdnn::HeuristicCache::instance().clear();
 
     test_no_profiling_on_shape_change<opr::Convolution>(
             {{12, 3, 36, 36}, {4, 3, 3, 3}}, {{32, 3, 28, 28}, {4, 3, 3, 3}});

src/opr/test/dnn/convolution.cpp

@@ -21,6 +21,7 @@
 #include "megbrain/gopt/inference.h"
 #include "megbrain/opr/tensor_manip.h"
 #include "megdnn/dtype.h"
+#include "megdnn/heuristic_cache.h"
 #include "megdnn/oprs/base.h"
 
 #include <gmock/gmock.h>
@@ -396,6 +397,7 @@ TEST(TestOprDNN, ConvBiasExePolicy) {
 #endif
         run(strategy);
     }
+    megdnn::HeuristicCache::instance().clear();
     ASSERT_THROW(run(S::OPTIMIZED | S::PROFILE), MegBrainError);
     PersistentCache::set_impl(orig_impl);
 }
@@ -460,6 +462,7 @@ TEST(TestOprDNN, ConvolutionExePolicy) {
     for (auto strategy :
          SmallVector<S>{S : HEURISTIC, S::PROFILE | S::HEURISTIC}) {
 #endif
+        megdnn::HeuristicCache::instance().clear();
         using Checker = AutoOprChecker<2, 1>;
 
         auto make_graph = [&](const Checker::SymInpArray& inputs)
@@ -489,6 +492,7 @@ TEST(TestOprDNN, ConvolutionExePolicy) {
         } else {
             ASSERT_LT(0, nr_get);
         }
+        megdnn::HeuristicCache::instance().clear();
     }
 }
 
@@ -544,6 +548,7 @@ TEST(TestOprDNN, ConvolutionBackwardDataBfloat16ExePolicy) {
 #else
     for (auto strategy: {S:HEURISTIC, S(S::PROFILE | S::HEURISTIC)}) {
 #endif
+        megdnn::HeuristicCache::instance().clear();
         using Checker = AutoOprChecker<2, 1>;
 
         auto make_graph = [&](const Checker::SymInpArray& inputs)
@@ -1835,6 +1840,7 @@ TEST(TestOprDNN, LocalShareForwardExecPolicy) {
         auto run_with_param = [&](size_t fh = 3, size_t fw = 3, size_t sh = 1,
                                   size_t sw = 1, size_t sgh = 3,
                                   size_t sgw = 3) {
+            megdnn::HeuristicCache::instance().clear();
             size_t ph = fh / 2, pw = fw / 2;
             param.pad_h = ph, param.pad_w = pw;
             param.stride_h = sh, param.stride_w = sw,
@@ -2289,6 +2295,7 @@ TEST(TestOprDNN, HeuristicReproducible) {
             }
             algo_name0 = palgo->name();
         }
+        megdnn::HeuristicCache::instance().clear();
         {
             Checker checker(make_graph, fwd);
             checker.run(inp_tensor(2, 3, 4, 9, 8, 3, 3), opt)
@@ -2306,6 +2313,7 @@ TEST(TestOprDNN, HeuristicReproducible) {
             algo_name1 = palgo->name();
         }
         EXPECT_TRUE(algo_name0 == algo_name1);
+        megdnn::HeuristicCache::instance().clear();
     }
 #undef inp_tensor
 #undef get_shp
@@ -2585,6 +2593,7 @@ TEST_F(TestWeightPreprocess, NoPreprocessNeeded) {
 }
 
 TEST_F(TestWeightPreprocess, PreprocessCalledOnlyOnce) {
+    megdnn::HeuristicCache::instance().clear();
     using ::testing::_;
     using ::testing::Return;
     using ::testing::Field;