feat(dnn): add algo interface for rocm&fallback matmul and batched matrix mul

GitOrigin-RevId: dea03a0f7a

dnn/src/fallback/batched_matrix_mul/algos.cpp

@@ -0,0 +1,107 @@
+/**
+ * \file dnn/src/fallback/batched_matrix_mul/algos.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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 "src/fallback/batched_matrix_mul/algos.h"
+#include "src/common/algo_base.h"
+#include "src/naive/handle.h"
+
+using namespace megdnn;
+using namespace fallback;
+
+BatchedMatrixMulForwardImpl::AlgoPack::AlgoPack() {
+    all_algos.push_back(&algo_default);
+
+    for (auto&& algo : all_algos) {
+        m_all_algos_map.emplace(algo->info().desc, algo);
+    }
+}
+
+BatchedMatrixMulForwardImpl::AlgoPack BatchedMatrixMulForwardImpl::sm_algo_pack;
+
+MEGDNN_DEF_GET_ALGO_FROM_DESC(BatchedMatrixMulForwardImpl)
+
+BatchedMatrixMulForwardImpl::AlgoBase::SizeArgs::SizeArgs(
+        BatchedMatrixMulForwardImpl* o, const TensorLayout& A,
+        const TensorLayout& B, const TensorLayout& C)
+        : opr{o}, layout_a{A}, layout_b{B}, layout_c{C} {}
+
+BatchedMatrixMulForwardImpl::AlgoBase::ExecArgs::ExecArgs(
+        BatchedMatrixMulForwardImpl* opr, _megdnn_tensor_in A,
+        _megdnn_tensor_in B, _megdnn_tensor_out C, _megdnn_workspace workspace)
+        : SizeArgs(opr, A.layout, B.layout, C.layout),
+          tensor_a{A},
+          tensor_b{B},
+          tensor_c{C},
+          workspace{workspace} {}
+
+std::string BatchedMatrixMulForwardImpl::AlgoBase::SizeArgs::to_string() const {
+    auto&& param = opr->param();
+    size_t m = layout_a.shape[0], n = layout_b.shape[1],
+           k = layout_a.shape[param.transposeA ? 0 : 1];
+    MEGDNN_MARK_USED_VAR(m);
+    MEGDNN_MARK_USED_VAR(n);
+    MEGDNN_MARK_USED_VAR(k);
+    return megdnn_mangle(ssprintf(
+            "A={%zux%zu},B={%zux%zu},C={%zux%zu},Transpose A=%d,Transpose "
+            "B=%d,ldA=%zu,ldB=%zu,ldC=%zu",
+            m, k, k, n, m, n, param.transposeA, param.transposeB,
+            layout_a.stride[0], layout_b.stride[0], layout_c.stride[0]));
+}
+
+/* ===================== default algo ===================== */
+size_t BatchedMatrixMulForwardImpl::AlgoDefault::get_workspace_in_bytes(
+        const SizeArgs& args) const {
+    auto opr = inplace_cpu_handle()->create_operator<MatrixMul>();
+    auto A_ = args.layout_a.remove_axis(0), B_ = args.layout_b.remove_axis(0),
+         C_ = args.layout_c.remove_axis(0);
+    opr->param() = args.opr->param();
+    return opr->get_workspace_in_bytes(A_, B_, C_);
+}
+
+void BatchedMatrixMulForwardImpl::AlgoDefault::exec(
+        const ExecArgs& args) const {
+    //! As megbrain may modify param when checking all transpose situations, so
+    //! here we should copy the param when dispatching kern
+    auto param = args.opr->param();
+    auto kern = [args, param]() {
+        auto N = args.layout_a.shape[0];
+        TensorND A_, B_, C_;
+        A_.raw_ptr = args.tensor_a.raw_ptr;
+        A_.layout = args.layout_a.remove_axis(0);
+        B_.raw_ptr = args.tensor_b.raw_ptr;
+        B_.layout = args.layout_b.remove_axis(0);
+        C_.raw_ptr = args.tensor_c.raw_ptr;
+        C_.layout = args.layout_c.remove_axis(0);
+
+        auto Astrd = args.layout_a.dtype.size() * args.layout_a.stride[0],
+             Bstrd = args.layout_b.dtype.size() * args.layout_b.stride[0],
+             Cstrd = args.layout_c.dtype.size() * args.layout_c.stride[0];
+
+        auto advance_ptr = [](TensorND& dest, ptrdiff_t d) {
+            dest.raw_ptr =
+                    static_cast<void*>(static_cast<dt_byte*>(dest.raw_ptr) + d);
+        };
+
+        auto opr = inplace_cpu_handle()->create_operator<MatrixMul>();
+        opr->param() = param;
+        rep(n, N) {
+            opr->exec(A_, B_, C_, args.workspace);
+            advance_ptr(A_, Astrd);
+            advance_ptr(B_, Bstrd);
+            advance_ptr(C_, Cstrd);
+        }
+    };
+
+    static_cast<naive::HandleImpl*>(args.opr->handle())->dispatch_kern(kern);
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/fallback/batched_matrix_mul/algos.h

@@ -0,0 +1,109 @@
+/**
+ * \file dnn/src/fallback/batched_matrix_mul/algos.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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/oprs.h"
+#include "src/common/algo_base.h"
+#include "src/common/metahelper.h"
+#include "src/common/utils.h"
+#include "src/fallback/batched_matrix_mul/opr_impl.h"
+
+#include <memory>
+#include <unordered_map>
+
+namespace megdnn {
+namespace fallback {
+
+/*!
+ * \brief base class for matrix mul algos
+ *
+ */
+class BatchedMatrixMulForwardImpl::AlgoBase : public Algorithm {
+protected:
+    ~AlgoBase() = default;
+
+public:
+    enum class AlgoType : uint32_t {
+        fallback_BLAS,
+    };
+    using Mapper = std::unordered_map<AlgorithmDesc, AlgoBase*>;
+
+    AlgoBase() : Algorithm() { m_handle_type = Handle::HandleType::FALLBACK; }
+
+    struct SizeArgs {
+        BatchedMatrixMulForwardImpl* opr;
+        TensorLayout layout_a, layout_b, layout_c;
+
+        std::string to_string() const;
+        SizeArgs(BatchedMatrixMulForwardImpl* opr, const TensorLayout& A,
+                 const TensorLayout& B, const TensorLayout& C);
+    };
+    struct ExecArgs : public SizeArgs {
+        TensorND tensor_a, tensor_b, tensor_c;
+        Workspace workspace;
+
+        ExecArgs(BatchedMatrixMulForwardImpl* opr, _megdnn_tensor_in A,
+                 _megdnn_tensor_in B, _megdnn_tensor_out C,
+                 _megdnn_workspace workspace);
+    };
+
+    virtual bool is_available(const SizeArgs& args) const = 0;
+    virtual size_t get_workspace_in_bytes(const SizeArgs& args) const = 0;
+    virtual void exec(const ExecArgs&) const = 0;
+
+    bool is_available_wk(const SizeArgs& args, size_t limit) const {
+        return is_available(args) && get_workspace_in_bytes(args) <= limit;
+    }
+    bool is_available_reproducible(
+            const SizeArgs& args, bool reproducible = true,
+            size_t limit = std::numeric_limits<size_t>::max()) const {
+        return (!reproducible || is_reproducible()) &&
+               is_available_wk(args, limit);
+    }
+    AlgoBase& check_workspace(const SizeArgs& args,
+                              const Workspace& workspace) {
+        auto req = get_workspace_in_bytes(args);
+        megdnn_assert(
+                req <= workspace.size,
+                "matrix mul fwd algo %s: required workspace %zu bytes, got %zu",
+                name(), req, workspace.size);
+        return *this;
+    }
+};
+
+class BatchedMatrixMulForwardImpl::AlgoDefault final : public AlgoBase {
+public:
+    AlgoDefault() = default;
+    bool is_available(const SizeArgs&) const override { return true; }
+    size_t get_workspace_in_bytes(const SizeArgs& /* args */) const override;
+    const char* name() const override { return "DEFAULT"; }
+    virtual void exec(const ExecArgs&) const override;
+    bool is_reproducible() const override { return true; }
+    MEGDNN_DECL_ALGO_TYPE(fallback_BLAS)
+};
+
+class BatchedMatrixMulForwardImpl::AlgoPack : NonCopyableObj {
+private:
+    AlgoBase::Mapper m_all_algos_map;
+
+public:
+    AlgoPack();
+    AlgoDefault algo_default;
+    std::vector<AlgoBase*> all_algos;
+
+    const AlgoBase::Mapper& all_algos_map() const { return m_all_algos_map; }
+};
+
+}  // namespace fallback
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/fallback/batched_matrix_mul/opr_impl.cpp

@@ -6,67 +6,61 @@
  *
  * 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.
  */
 #include "./opr_impl.h"
-#include "src/naive/handle.h"
+#include "./algos.h"
+#include "hcc_detail/hcc_defs_prologue.h"
+
+#include "src/common/algo_chooser.h"
+#include "src/common/utils.cuh"
+#include "src/fallback/handle.h"
 
 using namespace megdnn;
 using namespace fallback;
 
-BatchedMatrixMulImpl::BatchedMatrixMulImpl(Handle *handle):
-    BatchedMatrixMulForwardImpl(handle),
-    m_storage(new CpuOprDelegationStorage<>),
-    m_opr(m_storage->get<MatrixMul>())
-{
+std::vector<BatchedMatrixMulForwardImpl::Algorithm*>
+BatchedMatrixMulForwardImpl::get_all_algorithms(const TensorLayout& A,
+                                                const TensorLayout& B,
+                                                const TensorLayout& C) {
+    AlgoBase::SizeArgs args{this, A, B, C};
+    return megdnn::get_all_algorithms<BatchedMatrixMulForwardImpl>(args);
 }
 
-size_t BatchedMatrixMulImpl::get_workspace_in_bytes(
-        const TensorLayout &A, const TensorLayout &B,
-        const TensorLayout &C) {
-    auto A_ = A.remove_axis(0), B_ = B.remove_axis(0), C_ = C.remove_axis(0);
-    m_opr->param() = param();
-    return m_opr->get_workspace_in_bytes(A_, B_, C_);
+BatchedMatrixMulForwardImpl::Algorithm*
+BatchedMatrixMulForwardImpl::get_algorithm_heuristic(
+        const TensorLayout& A, const TensorLayout& B, const TensorLayout& C,
+        size_t workspace_limit_in_bytes, bool reproducible) {
+    AlgoBase::SizeArgs args{this, A, B, C};
+    if (sm_algo_pack.algo_default.is_available_reproducible(
+                args, reproducible, workspace_limit_in_bytes)) {
+        return &sm_algo_pack.algo_default;
+    }
+    if (reproducible) {
+        return megdnn::get_reproducible_algo<BatchedMatrixMulForwardImpl>(
+                sm_algo_pack.all_algos, args, workspace_limit_in_bytes,
+                "batched matrix mul forward");
+    } else {
+        return megdnn::get_usable_algo<BatchedMatrixMulForwardImpl>(
+                sm_algo_pack.all_algos, args, workspace_limit_in_bytes,
+                "batched matrix mul forward");
+    }
 }
 
-void BatchedMatrixMulImpl::exec(_megdnn_tensor_in A,
-        _megdnn_tensor_in B,
-        _megdnn_tensor_out C,
-        _megdnn_workspace workspace) {
-    check_exec(A.layout, B.layout, C.layout, workspace.size);
-
-    m_opr->param() = this->param();
-    auto kern = [this, A, B, C, workspace]() {
-        auto N = A.layout.shape[0];
-        TensorND A_, B_, C_;
-        A_.raw_ptr = A.raw_ptr;
-        A_.layout = A.layout.remove_axis(0);
-        B_.raw_ptr = B.raw_ptr;
-        B_.layout = B.layout.remove_axis(0);
-        C_.raw_ptr = C.raw_ptr;
-        C_.layout = C.layout.remove_axis(0);
-
-        auto Astrd = A.layout.dtype.size() * A.layout.stride[0],
-             Bstrd = B.layout.dtype.size() * B.layout.stride[0],
-             Cstrd = C.layout.dtype.size() * C.layout.stride[0];
-
-        auto advance_ptr = [](TensorND &dest, ptrdiff_t d) {
-            dest.raw_ptr = static_cast<void*>(
-                    static_cast<dt_byte*>(dest.raw_ptr) + d);
-        };
-
-        rep(n, N) {
-            m_opr->exec(A_, B_, C_, workspace);
-            advance_ptr(A_, Astrd);
-            advance_ptr(B_, Bstrd);
-            advance_ptr(C_, Cstrd);
-        }
-    };
-
-    static_cast<naive::HandleImpl*>(handle())->dispatch_kern(kern);
+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);
 }
 
+void BatchedMatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
+                                       _megdnn_tensor_out C,
+                                       _megdnn_workspace workspace) {
+    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);
+}
 
 // vim: syntax=cpp.doxygen
-
-

dnn/src/fallback/batched_matrix_mul/opr_impl.h

@@ -15,26 +15,42 @@
 namespace megdnn {
 namespace fallback {
 
-class BatchedMatrixMulImpl: public naive::BatchedMatrixMulForwardImpl {
-    public:
-        BatchedMatrixMulImpl(Handle *handle);
-        void exec(
-                _megdnn_tensor_in A,
-                _megdnn_tensor_in B,
-                _megdnn_tensor_out C,
-                _megdnn_workspace workspace) override;
-
-        size_t get_workspace_in_bytes(const TensorLayout &A,
-                const TensorLayout &B,
-                const TensorLayout &C) override;
-
-    private:
-        std::unique_ptr<CpuOprDelegationStorage<>> m_storage;
-        MatrixMulForward* m_opr;
+class BatchedMatrixMulForwardImpl: public naive::BatchedMatrixMulForwardImpl {
+public:
+    using naive::BatchedMatrixMulForwardImpl::BatchedMatrixMulForwardImpl;
+    void exec(_megdnn_tensor_in A, _megdnn_tensor_in B, _megdnn_tensor_out C,
+              _megdnn_workspace workspace) override;
+    size_t get_workspace_in_bytes(const TensorLayout&, const TensorLayout&,
+                                  const TensorLayout&) override;
+
+    bool is_thread_safe() const override { return true; }
+
+    class AlgoBase;
+    class AlgoDefault;
+    class AlgoPack;
+    static const AlgoPack& algo_pack() { return sm_algo_pack; }
+
+    static AlgoBase* get_algo_from_desc(const AlgorithmDesc& desc);
+
+private:
+    std::vector<Algorithm*> get_all_algorithms(
+            const TensorLayout& /*A*/, const TensorLayout& /*B*/,
+            const TensorLayout& /*C*/) override;
+
+    Algorithm* get_algorithm_heuristic(const TensorLayout& /*A*/,
+                                       const TensorLayout& /*B*/,
+                                       const TensorLayout& /*C*/,
+                                       size_t /*workspace_limit_in_bytes*/,
+                                       bool /*reproducible*/) override;
+
+    const char* get_algorithm_set_name() const override {
+        return "FALLBACK BATCHED MATMUL";
+    }
+
+    static AlgoPack sm_algo_pack;
 };
 
-} // namespace fallback
-} // namespace megdnn
+}  // namespace fallback
+}  // namespace megdnn
 
 // vim: syntax=cpp.doxygen
-

dnn/src/fallback/conv_bias/im2col/factory.h

@@ -473,6 +473,13 @@ public:
                     PostprocessMode::NO_PROCESS,
                     "NoPackStrategyType::FLOAT16_FLOAT16"_hash);
                 break;
+#endif
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+            case StrategyType::FLOAT_FP16:
+                cb1(NCHW, NO_PACK, dt_float16, __fp16,
+                    PostprocessMode::NO_PROCESS,
+                    "NoPackStrategyType::FLOAT_FP16"_hash);
+                break;
 #endif
             case StrategyType::INT8x8x16:
                 cb3(NCHW, NO_PACK, dt_int8, dt_int16, dt_int16, dt_int8,

dnn/src/fallback/conv_bias/im2col/strategy_nopack.cpp

@@ -169,6 +169,10 @@ INSTANTIAL_CLASS(dt_int8, dt_int32, dt_int32, dt_int32, dt_int32,
 INSTANTIAL_CLASS(dt_float16, dt_float16, dt_float16, dt_float16, dt_float16,
                  megdnn::PostprocessMode::NO_PROCESS)
 #endif
+#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+INSTANTIAL_CLASS(dt_float16, dt_float16, dt_float16, __fp16, __fp16,
+                 megdnn::PostprocessMode::NO_PROCESS)
+#endif
 #undef INSTANTIAL_CLASS
 }  // namespace megdnn
 

dnn/src/fallback/handle.cpp

@@ -67,7 +67,7 @@ MEGDNN_SPECIALIZE_CREATE_OPERATOR(ElemwiseMultiType)
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(AddUpdate)
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(MaskConvForward)
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(Resize)
-MEGDNN_SPECIALIZE_CREATE_OPERATOR(BatchedMatrixMul)
+MEGDNN_SPECIALIZE_CREATE_OPERATOR(BatchedMatrixMulForward)
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(ConvBias)
 MEGDNN_SPECIALIZE_CREATE_OPERATOR(PowC)
 

dnn/src/fallback/matrix_mul/algos.cpp

@@ -10,13 +10,18 @@
  */
 
 #include "src/fallback/matrix_mul/algos.h"
+#include "megdnn/opr_param_defs.h"
 #include "src/fallback/matrix_mul/gemm_impl.h"
 #include "src/fallback/matrix_mul/gemv.h"
 #include "src/fallback/matrix_mul/generic_strategy.h"
+
+#include "src/naive/matrix_mul/matrix_mul_helper.h"
+
 #include "midout.h"
 
 MIDOUT_DECL(megdnn_fb_matmul_f32_kern)
 MIDOUT_DECL(megdnn_fb_matmul_f32_gemm_gemv_like)
+MIDOUT_DECL(megdnn_fb_matmul_naive)
 
 using namespace megdnn;
 using namespace fallback;
@@ -39,6 +44,32 @@ void f32_8x12x1_kern(const MatrixMulImpl::KernParam& kern_param) {
     }
     MIDOUT_END();
 }
+
+void kern_naive(const MatrixMulImpl::KernParam& kern_param) {
+    MIDOUT_BEGIN(megdnn_fb_matmul_naive, void) {
+        size_t M = kern_param.M, N = kern_param.N, K = kern_param.K;
+        size_t LDA = kern_param.LDA, LDB = kern_param.LDB, LDC = kern_param.LDC;
+
+#define DISPATCH(TA, TB)                                                 \
+    if (kern_param.trA == TA && kern_param.trB == TB) {                  \
+        naive::dispatch_ta_tb<TA, TB>(                                   \
+                kern_param.A_ptr, kern_param.B_ptr, kern_param.C_ptr,    \
+                kern_param.workspace_ptr, M, N, K, LDA, LDB, LDC,        \
+                kern_param.A_type, kern_param.B_type, kern_param.C_type, \
+                kern_param.format, kern_param.compute_mode);             \
+        return;                                                          \
+    }
+        DISPATCH(true, true);
+        DISPATCH(true, false);
+        DISPATCH(false, true);
+        DISPATCH(false, false);
+#undef DISPATCH
+    megdnn_assert_internal(0);
+
+    }
+    MIDOUT_END();
+
+}
 }  // anonymous namespace
 
 ////////////////////// AlgoF32K8x12x1 ///////////////////////////
@@ -84,11 +115,14 @@ MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL(AlgoF32K8x12x1, megdnn_fb_matmul_f32_kern,
 bool MatrixMulImpl::AlgoGemv::usable(
         const KernSizeParam& kern_size_param) const {
     return !kern_size_param.trA && !kern_size_param.trB &&
-           kern_size_param.format == param::MatrixMul::Format::DEFAULT &&
-           !((kern_size_param.A_type.enumv() ==
-              kern_size_param.B_type.enumv()) &&
-             (kern_size_param.A_type.enumv() == DTypeEnum::Int16) &&
-             (kern_size_param.C_type.enumv() == DTypeEnum::Int32));
+                   kern_size_param.format ==
+                           param::MatrixMul::Format::DEFAULT &&
+                   kern_size_param.compute_mode ==
+                   param::MatrixMul::ComputeMode::DEFAULT &&
+                   !((kern_size_param.A_type.enumv() ==
+                      kern_size_param.B_type.enumv()) &&
+                     (kern_size_param.A_type.enumv() == DTypeEnum::Int16) &&
+                     (kern_size_param.C_type.enumv() == DTypeEnum::Int32));
 }
 
 bool MatrixMulImpl::AlgoGemv::preferred(
@@ -128,4 +162,44 @@ MatrixMulImpl::kern_t MatrixMulImpl::AlgoGemv::get_kern(
     megdnn_assert(0);
 }
 
+/* ===================== naive algo ===================== */
+bool MatrixMulImpl::AlgoNaive::usable(const KernSizeParam&) const {
+    return true;
+}
+
+bool MatrixMulImpl::AlgoNaive::preferred(const KernSizeParam&) const {
+    return false;
+}
+
+size_t MatrixMulImpl::AlgoNaive::get_workspace(
+        const KernSizeParam& kern_param) const {
+    MIDOUT_BEGIN(
+            megdnn_fb_matmul_naive,
+            midout_iv("MatrixMulForwardImpl::get_workspace_in_bytes"_hash)) {
+        if (kern_param.A_type.enumv() == DTypeEnum::Quantized4Asymm ||
+            kern_param.A_type.enumv() == DTypeEnum::QuantizedS4) {
+            size_t ret = 0;
+            if (kern_param.trA) {
+                ret += kern_param.LDA * kern_param.K;
+            } else {
+                ret += kern_param.LDA * kern_param.M;
+            }
+            if (kern_param.trB) {
+                ret += kern_param.LDB * kern_param.N;
+            } else {
+                ret += kern_param.LDB * kern_param.K;
+            }
+            return ret;
+        }
+        return 0;
+    }
+    MIDOUT_END();
+
+}
+
+MatrixMulImpl::kern_t MatrixMulImpl::AlgoNaive::get_kern(
+        const KernSizeParam&) const {
+    return kern_naive;
+}
+
 // vim: syntax=cpp.doxygen

dnn/src/fallback/matrix_mul/algos.h

@@ -52,6 +52,28 @@ public:
             DEFAULT)
 };
 
+class MatrixMulImpl::AlgoNaive final : public AlgoBase {
+public:
+    bool is_reproducible() const override { return true; }
+    const char* name() const override { return "FB_NAIVE"; }
+    bool usable(const KernSizeParam&) const override;
+    bool preferred(const KernSizeParam&) const override;
+    size_t get_workspace(const KernSizeParam&) const override;
+    kern_t get_kern(const KernSizeParam&) const override;
+    AlgoSet algoset() const override { return AlgoSet::ALGO_TYPE_GEMM; }
+    PackMode packmode() const override { return PackMode::NO_PACK; }
+    MEGDNN_DECL_ALGO_TYPE(FB_NAIVE)
+    MEGDNN_OVERRIDE_MATMUL_DESC(
+            8, 16, 1, 4,
+            static_cast<AlgoDataType>(
+                    static_cast<uint32_t>(AlgoDataType::FLOAT16) |
+                    static_cast<uint32_t>(AlgoDataType::FLOAT32) |
+                    static_cast<uint32_t>(AlgoDataType::INT8X8X16) |
+                    static_cast<uint32_t>(AlgoDataType::QINT8X8X32) |
+                    static_cast<uint32_t>(AlgoDataType::QUINT8X8X32)),
+            DEFAULT)
+};
+
 }  // namespace fallback
 }  // namespace megdnn
 

dnn/src/fallback/matrix_mul/opr_impl.cpp

@@ -35,6 +35,7 @@ using namespace fallback;
 class MatrixMulImpl::AlgoPack : NonCopyableObj {
     AlgoF32K8x12x1 f32_k8x12x1;
     AlgoGemv gemv;
+    AlgoNaive naive;
     SmallVector<AlgoBase*> m_all_algos;
     AlgoBase::Mapper m_all_algos_map;
 
@@ -42,6 +43,7 @@ public:
     AlgoPack() {
         m_all_algos.emplace_back(&gemv);
         m_all_algos.emplace_back(&f32_k8x12x1);
+        m_all_algos.emplace_back(&naive);
         for (auto&& algo : m_all_algos) {
             m_all_algos_map.emplace(algo->info().desc, algo);
         }
@@ -147,19 +149,26 @@ MatrixMul::Algorithm* MatrixMulImpl::get_algorithm_heuristic(
     algo_type.format = kern_size_param.format;
     auto algos = select_algo_type(algo_type);
     Algorithm *heuristic_algo = nullptr;
+    Algorithm *usable_algo = nullptr;
     for (auto&& algo : algos) {
         if (static_cast<AlgoBase*>(algo)->usable(kern_size_param) &&
-            static_cast<AlgoBase*>(algo)->preferred_reproducible(
-                    kern_size_param, reproducible) &&
             static_cast<AlgoBase*>(algo)->get_workspace(kern_size_param) <=
                     workspace_limit_in_bytes) {
-            if (algo->algoset() == AlgoBase::AlgoSet::ALGO_TYPE_GEMV) {
-                return algo;
-            } else if (!heuristic_algo) {
-                heuristic_algo = algo;
+            if (static_cast<AlgoBase*>(algo)->preferred_reproducible(
+                        kern_size_param, reproducible)) {
+                //! use gemv algo if it's prefered
+                if (algo->algoset() == AlgoBase::AlgoSet::ALGO_TYPE_GEMV) {
+                    return algo;
+                } else if (!heuristic_algo) {
+                    heuristic_algo = algo;
+                }
+            } else if (!usable_algo) {
+                usable_algo = algo;
             }
         }
     }
+    if (!heuristic_algo) heuristic_algo = usable_algo;
+    megdnn_assert(heuristic_algo, "No usable algorithm found");
     return heuristic_algo;
 }
 

dnn/src/fallback/matrix_mul/opr_impl.h

@@ -110,6 +110,7 @@ public:
             //! fallback
             FB_F32K8x12x1 = 1 << 0,
             FB_GEMV,
+            FB_NAIVE,
 
 #if MEGDNN_X86
             //! x86
@@ -233,6 +234,7 @@ public:
 private:
     class AlgoF32K8x12x1;  // Fallback F32 Kernel 8x12x1
     class AlgoGemv;
+    class AlgoNaive;
     class AlgoPack;
     //! maintain all the algos of in the opr of fallback
     static const AlgoPack& algo_pack();

dnn/src/naive/matrix_mul/matrix_mul_helper.h

@@ -141,20 +141,39 @@ void run_matrix_mul_mk8_tpl(const itype* A, const itype* B, otype* C, size_t M,
 }
 
 template <bool transA, bool transB>
-void exec_matrix_mul_quint4x4x32_helper(_megdnn_tensor_in A,
-                                        _megdnn_tensor_in B,
-                                        _megdnn_tensor_out C,
-                                        _megdnn_workspace workspace,
-                                        const param::MatrixMul& param) {
+void exec_matrix_mul_quint4x4x32_helper(
+        const void* A, const void* B, void* C, void* workspace, size_t M,
+        size_t N, size_t K, ptrdiff_t LDA, ptrdiff_t LDB, ptrdiff_t LDC,
+        DType A_type, DType B_type, DType C_type,
+        const MatrixMul::Param::Format& format,
+        const MatrixMul::Param::ComputeMode& compute_mode) {
+    MEGDNN_MARK_USED_VAR(C_type);
+    MEGDNN_MARK_USED_VAR(format);
+    MEGDNN_MARK_USED_VAR(compute_mode);
     auto convert_layout = [](const TensorLayout& layout) {
         auto ret = layout;
         auto param = layout.dtype.param<dtype::Quantized4Asymm>();
         ret.dtype = dtype::Quantized8Asymm(param.scale, param.zero_point);
         return ret;
     };
-    TensorND nA = {workspace.raw_ptr, convert_layout(A.layout)};
-    TensorND nB = {workspace.raw_ptr + nA.layout.span().dist_byte(),
-                   convert_layout(B.layout)};
+    TensorLayout A_layout, B_layout;
+    if (transA) {
+        A_layout = TensorLayout({K, M}, {LDA, 1}, A_type);
+    } else {
+        A_layout = TensorLayout({M, K}, {LDA, 1}, A_type);
+    }
+    if (transB) {
+        B_layout = TensorLayout({N, K}, {LDB, 1}, B_type);
+    } else {
+        B_layout = TensorLayout({K, N}, {LDB, 1}, B_type);
+    }
+
+    TensorND tensorA{const_cast<void*>(A), A_layout};
+    TensorND tensorB{const_cast<void*>(B), B_layout};
+    TensorND nA = {workspace, convert_layout(A_layout)};
+    TensorND nB = {
+            static_cast<uint8_t*>(workspace) + nA.layout.span().dist_byte(),
+            convert_layout(B_layout)};
     auto convert_4to8 = [](const TensorND& in, const TensorND& out) {
         auto ptr =
                 static_cast<uint8_t*>(in.raw_ptr) + in.layout.span().low_byte;
@@ -168,31 +187,48 @@ void exec_matrix_mul_quint4x4x32_helper(_megdnn_tensor_in A,
             out_ptr[i + 1] = val1;
         }
     };
-    convert_4to8(A, nA);
-    convert_4to8(B, nB);
-    auto M = C.layout.shape[0], N = C.layout.shape[1];
-    auto K = A.layout.shape[param.transposeA ? 0 : 1];
-    auto LDA = A.layout.stride[0], LDB = B.layout.stride[0],
-         LDC = C.layout.stride[0];
+    convert_4to8(tensorA, nA);
+    convert_4to8(tensorB, nB);
     run_matrix_mul_tpl<uint8_t, dt_int32, transA, transB, dt_int32>(
             nA.compatible_ptr<uint8_t>(), nB.compatible_ptr<uint8_t>(),
-            C.compatible_ptr<dt_int32>(), M, N, K, LDA, LDB, LDC,
-            nA.layout.dtype, nB.layout.dtype);
+            static_cast<dt_int32*>(C), M, N, K, LDA, LDB, LDC, nA.layout.dtype,
+            nB.layout.dtype);
 }
 template <bool transA, bool transB>
-void exec_matrix_mul_qint4x4x16_helper(_megdnn_tensor_in A, _megdnn_tensor_in B,
-                                       _megdnn_tensor_out C,
-                                       _megdnn_workspace workspace,
-                                       const param::MatrixMul& param) {
+void exec_matrix_mul_qint4x4x16_helper(
+        const void* A, const void* B, void* C, void* workspace, size_t M,
+        size_t N, size_t K, ptrdiff_t LDA, ptrdiff_t LDB, ptrdiff_t LDC,
+        DType A_type, DType B_type, DType C_type,
+        const MatrixMul::Param::Format& format,
+        const MatrixMul::Param::ComputeMode& compute_mode) {
+    MEGDNN_MARK_USED_VAR(C_type);
+    MEGDNN_MARK_USED_VAR(format);
+    MEGDNN_MARK_USED_VAR(compute_mode);
     auto convert_layout = [](const TensorLayout& layout) {
         auto ret = layout;
         auto param = layout.dtype.param<dtype::QuantizedS4>();
         ret.dtype = dtype::QuantizedS8(param.scale);
         return ret;
     };
-    TensorND nA = {workspace.raw_ptr, convert_layout(A.layout)};
-    TensorND nB = {workspace.raw_ptr + nA.layout.span().dist_byte(),
-                   convert_layout(B.layout)};
+    TensorLayout A_layout, B_layout;
+    if (transA) {
+        A_layout = TensorLayout({K, M}, {LDA, 1}, A_type);
+    } else {
+        A_layout = TensorLayout({M, K}, {LDA, 1}, A_type);
+    }
+    if (transB) {
+        B_layout = TensorLayout({N, K}, {LDB, 1}, B_type);
+    } else {
+        B_layout = TensorLayout({K, N}, {LDB, 1}, B_type);
+    }
+
+    TensorND tensorA{const_cast<void*>(A), A_layout};
+    TensorND tensorB{const_cast<void*>(B), B_layout};
+
+    TensorND nA = {workspace, convert_layout(A_layout)};
+    TensorND nB = {
+            static_cast<uint8_t*>(workspace) + nA.layout.span().dist_byte(),
+            convert_layout(B_layout)};
     auto convert_4to8 = [](const TensorND& in, const TensorND& out) {
         auto ptr = static_cast<int8_t*>(in.raw_ptr) + in.layout.span().low_byte;
         auto out_ptr =
@@ -204,18 +240,98 @@ void exec_matrix_mul_qint4x4x16_helper(_megdnn_tensor_in A, _megdnn_tensor_in B,
             out_ptr[i + 1] = cur >> 4;
         }
     };
-    convert_4to8(A, nA);
-    convert_4to8(B, nB);
-    auto M = C.layout.shape[0], N = C.layout.shape[1];
-    auto K = A.layout.shape[param.transposeA ? 0 : 1];
-    auto LDA = A.layout.stride[0], LDB = B.layout.stride[0],
-         LDC = C.layout.stride[0];
+    convert_4to8(tensorA, nA);
+    convert_4to8(tensorB, nB);
     run_matrix_mul_tpl<int8_t, dt_int16, transA, transB, dt_int16>(
             nA.compatible_ptr<int8_t>(), nB.compatible_ptr<int8_t>(),
-            C.compatible_ptr<dt_int16>(), M, N, K, LDA, LDB, LDC,
-            nA.layout.dtype, nB.layout.dtype);
+            static_cast<dt_int16*>(C), M, N, K, LDA, LDB, LDC, nA.layout.dtype,
+            nB.layout.dtype);
 }
 
+template <bool TA, bool TB>
+void dispatch_ta_tb(const void* A, const void* B, void* C, void* workspace,
+                    size_t M, size_t N, size_t K, ptrdiff_t LDA, ptrdiff_t LDB,
+                    ptrdiff_t LDC, DType A_type, DType B_type, DType C_type,
+                    const MatrixMul::Param::Format& format,
+                    const MatrixMul::Param::ComputeMode& compute_mode) {
+#define cb(_itype, _otype, _comp_type)                                         \
+    if (format == param::MatrixMul::Format::DEFAULT) {                         \
+        return run_matrix_mul_tpl<_itype, _otype, TA, TB, _comp_type>(         \
+                static_cast<const _itype*>(A), static_cast<const _itype*>(B),  \
+                static_cast<_otype*>(C), M, N, K, LDA, LDB, LDC, A_type,       \
+                B_type);                                                       \
+    } else if (format == param::MatrixMul::Format::MK4) {                      \
+        return run_matrix_mul_mk4_tpl<_itype, _otype, TA, TB, _comp_type>(     \
+                static_cast<const _itype*>(A), static_cast<const _itype*>(B),  \
+                static_cast<_otype*>(C), M, N, K, LDA, LDB, LDC, A_type,       \
+                B_type);                                                       \
+    } else if (format == param::MatrixMul::Format::MK4_DOT) {                  \
+        return run_matrix_mul_mk4_dot_tpl<_itype, _otype, TA, TB, _comp_type>( \
+                static_cast<const _itype*>(A), static_cast<const _itype*>(B),  \
+                static_cast<_otype*>(C), M, N, K, LDA, LDB, LDC, A_type,       \
+                B_type);                                                       \
+    } else if (format == param::MatrixMul::Format::MK8) {                      \
+        return run_matrix_mul_mk8_tpl<_itype, _otype, TA, TB, _comp_type>(     \
+                static_cast<const _itype*>(A), static_cast<const _itype*>(B),  \
+                static_cast<_otype*>(C), M, N, K, LDA, LDB, LDC, A_type,       \
+                B_type);                                                       \
+    }
+
+    if (A_type == dtype::Float32()) {
+        cb(dt_float32, dt_float32, dt_float32);
+#if !MEGDNN_DISABLE_FLOAT16
+    } else if (A_type == dtype::Float16()) {
+        using Param = MatrixMul::Param;
+        if (compute_mode == Param::ComputeMode::DEFAULT) {
+            cb(dt_float16, dt_float16, dt_float16);
+        } else if (compute_mode == Param::ComputeMode::FLOAT32) {
+            cb(dt_float16, dt_float16, dt_float32);
+        }
+    } else if (A_type == dtype::BFloat16()) {
+        using Param = MatrixMul::Param;
+        if (compute_mode == Param::ComputeMode::DEFAULT) {
+            cb(dt_bfloat16, dt_bfloat16, dt_bfloat16);
+        } else if (compute_mode == Param::ComputeMode::FLOAT32) {
+            cb(dt_bfloat16, dt_bfloat16, dt_float32);
+        }
+#endif
+    } else if (A_type == dtype::Int8() &&
+               C_type == dtype::Int16()) {
+        cb(dt_int8, dt_int16, dt_int16);
+    } else if (A_type == dtype::Int16() &&
+               C_type == dtype::Int32()) {
+        cb(dt_int16, dt_int32, dt_int32);
+    } else if ((A_type == dtype::Int8() ||
+                A_type.enumv() == DTypeEnum::QuantizedS8) &&
+               (C_type == dtype::Int32() ||
+                C_type.enumv() == DTypeEnum::QuantizedS32)) {
+        cb(dt_int8, dt_int32, dt_int32);
+    } else if (A_type.enumv() == DTypeEnum::Quantized8Asymm &&
+               C_type.enumv() == DTypeEnum::QuantizedS32) {
+        cb(uint8_t, dt_int32, dt_int32);
+    } else if (A_type.enumv() == DTypeEnum::Quantized4Asymm &&
+               C_type.enumv() == DTypeEnum::QuantizedS32 &&
+               format == param::MatrixMul::Format::DEFAULT) {
+        exec_matrix_mul_quint4x4x32_helper<TA, TB>(
+                A, B, C, workspace, M, N, K, LDA, LDB, LDC, A_type, B_type,
+                C_type, format, compute_mode);
+        return;
+    } else if (A_type.enumv() == DTypeEnum::QuantizedS4 &&
+               C_type.enumv() == DTypeEnum::QuantizedS16 &&
+               format == param::MatrixMul::Format::DEFAULT) {
+        exec_matrix_mul_qint4x4x16_helper<TA, TB>(
+                A, B, C, workspace, M, N, K, LDA, LDB, LDC, A_type, B_type,
+                C_type, format, compute_mode);
+        return;
+    }
+#undef cb
+    megdnn_throw(
+            ssprintf("unsupported naive MatrixMul(%s, %s) -> %s (cmode = %d)",
+                     A_type.name(), B_type.name(), C_type.name(),
+                     static_cast<int>(compute_mode)));
+}
+
+
 }  // namespace naive
 }  // namespace megdnn
 

dnn/src/naive/matrix_mul/opr_impl.cpp

@@ -45,77 +45,10 @@ void dispatch_ta_tb(_megdnn_tensor_in A, _megdnn_tensor_in B,
     auto LDA = A.layout.stride[0], LDB = B.layout.stride[0],
          LDC = C.layout.stride[0];
 
-#define cb(_itype, _otype, _comp_type)                                         \
-    if (param.format == param::MatrixMul::Format::DEFAULT) {                   \
-        return run_matrix_mul_tpl<_itype, _otype, TA, TB, _comp_type>(         \
-                A.compatible_ptr<_itype>(), B.compatible_ptr<_itype>(),        \
-                C.compatible_ptr<_otype>(), M, N, K, LDA, LDB, LDC,            \
-                A.layout.dtype, B.layout.dtype);                               \
-    } else if (param.format == param::MatrixMul::Format::MK4) {                \
-        return run_matrix_mul_mk4_tpl<_itype, _otype, TA, TB, _comp_type>(     \
-                A.compatible_ptr<_itype>(), B.compatible_ptr<_itype>(),        \
-                C.compatible_ptr<_otype>(), M, N, K, LDA, LDB, LDC,            \
-                A.layout.dtype, B.layout.dtype);                               \
-    } else if (param.format == param::MatrixMul::Format::MK4_DOT) {            \
-        return run_matrix_mul_mk4_dot_tpl<_itype, _otype, TA, TB, _comp_type>( \
-                A.compatible_ptr<_itype>(), B.compatible_ptr<_itype>(),        \
-                C.compatible_ptr<_otype>(), M, N, K, LDA, LDB, LDC,            \
-                A.layout.dtype, B.layout.dtype);                               \
-    } else if (param.format == param::MatrixMul::Format::MK8) {                \
-        return run_matrix_mul_mk8_tpl<_itype, _otype, TA, TB, _comp_type>(     \
-                A.compatible_ptr<_itype>(), B.compatible_ptr<_itype>(),        \
-                C.compatible_ptr<_otype>(), M, N, K, LDA, LDB, LDC,            \
-                A.layout.dtype, B.layout.dtype);                               \
-    }
-
-    if (A.layout.dtype == dtype::Float32()) {
-        cb(dt_float32, dt_float32, dt_float32);
-#if !MEGDNN_DISABLE_FLOAT16
-    } else if (A.layout.dtype == dtype::Float16()) {
-        using Param = MatrixMul::Param;
-        if (param.compute_mode == Param::ComputeMode::DEFAULT) {
-            cb(dt_float16, dt_float16, dt_float16);
-        } else if (param.compute_mode == Param::ComputeMode::FLOAT32) {
-            cb(dt_float16, dt_float16, dt_float32);
-        }
-    } else if (A.layout.dtype == dtype::BFloat16()) {
-        using Param = MatrixMul::Param;
-        if (param.compute_mode == Param::ComputeMode::DEFAULT) {
-            cb(dt_bfloat16, dt_bfloat16, dt_bfloat16);
-        } else if (param.compute_mode == Param::ComputeMode::FLOAT32) {
-            cb(dt_bfloat16, dt_bfloat16, dt_float32);
-        }
-#endif
-    } else if (A.layout.dtype == dtype::Int8() &&
-               C.layout.dtype == dtype::Int16()) {
-        cb(dt_int8, dt_int16, dt_int16);
-    } else if (A.layout.dtype == dtype::Int16() &&
-               C.layout.dtype == dtype::Int32()) {
-        cb(dt_int16, dt_int32, dt_int32);
-    } else if ((A.layout.dtype == dtype::Int8() ||
-                A.layout.dtype.enumv() == DTypeEnum::QuantizedS8) &&
-               (C.layout.dtype == dtype::Int32() ||
-                C.layout.dtype.enumv() == DTypeEnum::QuantizedS32)) {
-        cb(dt_int8, dt_int32, dt_int32);
-    } else if (A.layout.dtype.enumv() == DTypeEnum::Quantized8Asymm &&
-               C.layout.dtype.enumv() == DTypeEnum::QuantizedS32) {
-        cb(uint8_t, dt_int32, dt_int32);
-    } else if (A.layout.dtype.enumv() == DTypeEnum::Quantized4Asymm &&
-               C.layout.dtype.enumv() == DTypeEnum::QuantizedS32 &&
-               param.format == param::MatrixMul::Format::DEFAULT) {
-        exec_matrix_mul_quint4x4x32_helper<TA, TB>(A, B, C, workspace, param);
-        return;
-    } else if (A.layout.dtype.enumv() == DTypeEnum::QuantizedS4 &&
-               C.layout.dtype.enumv() == DTypeEnum::QuantizedS16 &&
-               param.format == param::MatrixMul::Format::DEFAULT) {
-        exec_matrix_mul_qint4x4x16_helper<TA, TB>(A, B, C, workspace, param);
-        return;
-    }
-#undef cb
-    megdnn_throw(ssprintf(
-            "unsupported naive MatrixMul(%s, %s) -> %s (cmode = %d)",
-            A.layout.dtype.name(), B.layout.dtype.name(), C.layout.dtype.name(),
-            static_cast<int>(param.compute_mode)));
+    dispatch_ta_tb<TA, TB>(A.raw_ptr, B.raw_ptr, C.raw_ptr, workspace.raw_ptr,
+                           M, N, K, LDA, LDB, LDC, A.layout.dtype,
+                           B.layout.dtype, C.layout.dtype, param.format,
+                           param.compute_mode);
 }
 
 void MatrixMulForwardImpl::exec_internal(_megdnn_tensor_in A,

dnn/src/rocm/batched_matrix_mul/algos.cpp

@@ -0,0 +1,59 @@
+/**
+ * \file dnn/src/rocm/batched_matrix_mul/algos.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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 "src/rocm/batched_matrix_mul/algos.h"
+#include "src/common/algo_base.h"
+
+using namespace megdnn;
+using namespace rocm;
+
+BatchedMatrixMulForwardImpl::AlgoPack::AlgoPack() {
+    all_algos.push_back(&blas);
+
+    for (auto&& algo : all_algos) {
+        m_all_algos_map.emplace(algo->info().desc, algo);
+    }
+}
+
+BatchedMatrixMulForwardImpl::AlgoPack BatchedMatrixMulForwardImpl::sm_algo_pack;
+
+MEGDNN_DEF_GET_ALGO_FROM_DESC(BatchedMatrixMulForwardImpl)
+
+BatchedMatrixMulForwardImpl::AlgoBase::SizeArgs::SizeArgs(
+        BatchedMatrixMulForwardImpl* o, const TensorLayout& A,
+        const TensorLayout& B, const TensorLayout& C)
+        : opr{o}, layout_a{A}, layout_b{B}, layout_c{C} {}
+
+BatchedMatrixMulForwardImpl::AlgoBase::ExecArgs::ExecArgs(
+        BatchedMatrixMulForwardImpl* opr, _megdnn_tensor_in A,
+        _megdnn_tensor_in B, _megdnn_tensor_out C, _megdnn_workspace workspace)
+        : SizeArgs(opr, A.layout, B.layout, C.layout),
+          tensor_a{A},
+          tensor_b{B},
+          tensor_c{C},
+          workspace{workspace} {}
+
+std::string BatchedMatrixMulForwardImpl::AlgoBase::SizeArgs::to_string() const {
+    auto&& param = opr->param();
+    size_t m = layout_a.shape[0], n = layout_b.shape[1],
+           k = layout_a.shape[param.transposeA ? 0 : 1];
+    MEGDNN_MARK_USED_VAR(m);
+    MEGDNN_MARK_USED_VAR(n);
+    MEGDNN_MARK_USED_VAR(k);
+    return megdnn_mangle(ssprintf(
+            "A={%zux%zu},B={%zux%zu},C={%zux%zu},Transpose A=%d,Transpose "
+            "B=%d,ldA=%zu,ldB=%zu,ldC=%zu",
+            m, k, k, n, m, n, param.transposeA, param.transposeB,
+            layout_a.stride[0], layout_b.stride[0], layout_c.stride[0]));
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/rocm/batched_matrix_mul/algos.h

@@ -0,0 +1,118 @@
+/**
+ * \file dnn/src/rocm/batched_matrix_mul/algos.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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/oprs.h"
+#include "src/common/algo_base.h"
+#include "src/common/metahelper.h"
+#include "src/common/utils.h"
+#include "src/rocm/batched_matrix_mul/opr_impl.h"
+
+#include <memory>
+#include <unordered_map>
+
+namespace megdnn {
+namespace rocm {
+
+/*!
+ * \brief base class for matrix mul algos
+ *
+ */
+class BatchedMatrixMulForwardImpl::AlgoBase : public Algorithm {
+protected:
+    ~AlgoBase() = default;
+
+public:
+    enum class AlgoType : uint32_t {
+        ROCM_BLAS,
+    };
+    using Mapper = std::unordered_map<AlgorithmDesc, AlgoBase*>;
+
+    AlgoBase() : Algorithm() { m_handle_type = Handle::HandleType::ROCM; }
+    struct SizeArgs {
+        BatchedMatrixMulForwardImpl* opr;
+        TensorLayout layout_a, layout_b, layout_c;
+
+        std::string to_string() const;
+        SizeArgs(BatchedMatrixMulForwardImpl* opr, const TensorLayout& A,
+                 const TensorLayout& B, const TensorLayout& C);
+
+        bool can_be_treated_as_int8x8x32() const {
+            return layout_a.dtype.enumv() == layout_b.dtype.enumv() &&
+                   (layout_a.dtype.enumv() == DTypeEnum::Int8 ||
+                    layout_a.dtype.enumv() == DTypeEnum::QuantizedS8) &&
+                   (layout_c.dtype.enumv() == DTypeEnum::Int32 ||
+                    layout_c.dtype.enumv() == DTypeEnum::QuantizedS32) &&
+                   opr->param().format == param::MatrixMul::Format::DEFAULT;
+        }
+    };
+    struct ExecArgs : public SizeArgs {
+        TensorND tensor_a, tensor_b, tensor_c;
+        Workspace workspace;
+
+        ExecArgs(BatchedMatrixMulForwardImpl* opr, _megdnn_tensor_in A,
+                 _megdnn_tensor_in B, _megdnn_tensor_out C,
+                 _megdnn_workspace workspace);
+    };
+    virtual bool is_available(const SizeArgs& args) const = 0;
+    virtual size_t get_workspace_in_bytes(const SizeArgs& args) const = 0;
+    virtual void exec(const ExecArgs& args) const = 0;
+
+    bool is_available_wk(const SizeArgs& args, size_t limit) const {
+        return is_available(args) && get_workspace_in_bytes(args) <= limit;
+    }
+    bool is_available_reproducible(
+            const SizeArgs& args, bool reproducible = true,
+            size_t limit = std::numeric_limits<size_t>::max()) const {
+        return (!reproducible || is_reproducible()) &&
+               is_available_wk(args, limit);
+    }
+    AlgoBase& check_workspace(const SizeArgs& args,
+                              const Workspace& workspace) {
+        auto req = get_workspace_in_bytes(args);
+        megdnn_assert(
+                req <= workspace.size,
+                "matrix mul fwd algo %s: required workspace %zu bytes, got %zu",
+                name(), req, workspace.size);
+        return *this;
+    }
+};
+
+class BatchedMatrixMulForwardImpl::AlgoBlas final : public AlgoBase {
+public:
+    AlgoBlas() = default;
+    bool is_available(const SizeArgs& args) const override;
+    size_t get_workspace_in_bytes(const SizeArgs& /* args */) const override {
+        return 0_z;
+    }
+    const char* name() const override { return "BLAS"; }
+    void exec(const ExecArgs& args) const override;
+    bool is_reproducible() const override { return true; }
+    MEGDNN_DECL_ALGO_TYPE(ROCM_BLAS)
+};
+
+class BatchedMatrixMulForwardImpl::AlgoPack : NonCopyableObj {
+private:
+    AlgoBase::Mapper m_all_algos_map;
+
+public:
+    AlgoPack();
+    AlgoBlas blas;
+    std::vector<AlgoBase*> all_algos;
+
+    const AlgoBase::Mapper& all_algos_map() const { return m_all_algos_map; }
+};
+
+}  // namespace rocm
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/rocm/batched_matrix_mul/blas.cpp

@@ -0,0 +1,140 @@
+/**
+ * \file dnn/src/rocm/batched_matrix_mul/Blas.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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 "src/rocm/batched_matrix_mul/algos.h"
+
+#include "hcc_detail/hcc_defs_prologue.h"
+#include "src/rocm/handle.h"
+#include "src/rocm/utils.h"
+
+using namespace megdnn;
+using namespace rocm;
+
+bool BatchedMatrixMulForwardImpl::AlgoBlas::is_available(
+        const SizeArgs& args) const {
+    if (args.opr->param().format != param::MatrixMul::Format::DEFAULT)
+        return false;
+    if (args.layout_a.dtype == dtype::Float32() ||
+        args.layout_a.dtype == dtype::Float16()) {
+        return true;
+    }
+    return false;
+}
+
+void BatchedMatrixMulForwardImpl::AlgoBlas::exec(const ExecArgs& args) const {
+    auto batch = args.layout_a.shape[0];
+    auto m = args.layout_c.shape[1], n = args.layout_c.shape[2];
+    auto k = args.layout_a.shape[args.opr->param().transposeA ? 1 : 2];
+    auto&& handle = concrete_handle(args.opr->handle());
+    auto rocblas_handle_ = handle->get_rocblas_handle();
+
+    auto sgemm = [&]() {
+        auto zero = handle->zero_device();
+        auto one = handle->one_device();
+        rocblas_check(rocblas_sgemm_strided_batched(
+                rocblas_handle_,
+                args.opr->param().transposeB ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                args.opr->param().transposeA ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                n, m, k, one, args.tensor_b.ptr<dt_float32>(),
+                (rocblas_int)(args.layout_b.stride[1]),
+                (rocblas_int)(args.layout_b.stride[0]),
+                args.tensor_a.ptr<dt_float32>(),
+                (rocblas_int)(args.layout_a.stride[1]),
+                (rocblas_int)(args.layout_a.stride[0]), zero,
+                args.tensor_c.ptr<dt_float32>(),
+                (rocblas_int)(args.layout_c.stride[1]),
+                (rocblas_int)(args.layout_c.stride[0]), (rocblas_int)(batch)));
+
+    };
+
+#if !MEGDNN_DISABLE_FLOAT16
+    //! used for FLOAT_IO16xC32, not tested
+    auto gemm_ex = [&]() {
+        auto zero = handle->zero_device();
+        auto one = handle->one_device();
+        //! These two arguments for future use, see
+        //! https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/library/src/blas_ex/rocblas_gemm_ex.cpp
+        int32_t solution_index = 0;
+        uint32_t flags = 1;
+        size_t ws_size = 0;
+
+        rocblas_check(rocblas_gemm_strided_batched_ex(
+                rocblas_handle_,
+                args.opr->param().transposeB ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                args.opr->param().transposeA ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                n, m, k, one, args.tensor_b.raw_ptr, rocblas_datatype_i8_r,
+                args.layout_b.stride[1], args.layout_b.stride[0],
+                args.tensor_a.raw_ptr, rocblas_datatype_i8_r,
+                args.layout_a.stride[1], args.layout_a.stride[0], zero,
+                args.tensor_c.raw_ptr, rocblas_datatype_i32_r,
+                args.layout_c.stride[1], args.layout_c.stride[0],
+                args.tensor_c.raw_ptr, rocblas_datatype_i32_r,
+                args.layout_c.stride[1], args.layout_c.stride[0], batch,
+                rocblas_datatype_i32_r, rocblas_gemm_algo_standard,
+                solution_index, flags, &ws_size, nullptr));
+
+        MEGDNN_MARK_USED_VAR(ws_size);
+    };
+
+    auto hgemm = [&]() {
+        auto one_half = handle->one_device_h();
+        auto zero_half = handle->zero_device_h();
+        rocblas_check(rocblas_hgemm_strided_batched(
+                rocblas_handle_,
+                args.opr->param().transposeB ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                args.opr->param().transposeA ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                n, m, k, reinterpret_cast<const rocblas_half*>(one_half),
+                static_cast<const rocblas_half*>(args.tensor_b.raw_ptr),
+                args.layout_b.stride[1], args.layout_b.stride[0],
+                static_cast<const rocblas_half*>(args.tensor_a.raw_ptr),
+                args.layout_a.stride[1], args.layout_a.stride[0],
+                reinterpret_cast<const rocblas_half*>(zero_half),
+                static_cast<rocblas_half*>(args.tensor_c.raw_ptr),
+                args.layout_c.stride[1], args.layout_c.stride[0], batch));
+
+    };
+#endif
+
+    if (args.opr->param().compute_mode == Param::ComputeMode::DEFAULT) {
+        if (args.layout_a.dtype == dtype::Float32()) {
+            sgemm();
+        }
+#if !MEGDNN_DISABLE_FLOAT16
+        else {
+            megdnn_assert(args.layout_a.dtype == dtype::Float16(),
+                          "invalid matmul data type");
+            hgemm();
+        }
+#endif
+    }
+#if !MEGDNN_DISABLE_FLOAT16
+    else if (args.opr->param().compute_mode == Param::ComputeMode::FLOAT32) {
+        megdnn_assert(args.layout_b.dtype == dtype::Float16() &&
+                              args.layout_c.dtype == dtype::Float16() &&
+                              args.layout_a.dtype == dtype::Float16(),
+                      "DataType::FLOAT_IO16xC32 is supported, when dtype of A, "
+                      "B, C are all Float16");
+        gemm_ex();
+    }
+#endif
+    else {
+        megdnn_throw("Unsupported data_type of matrix mul on rocm.");
+    }
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/rocm/batched_matrix_mul/opr_impl.cpp

@@ -10,111 +10,58 @@
  * implied.
  */
 #include "./opr_impl.h"
+#include "./algos.h"
 #include "hcc_detail/hcc_defs_prologue.h"
 
+#include "src/common/algo_chooser.h"
 #include "src/common/utils.cuh"
 #include "src/rocm/handle.h"
 #include "src/rocm/utils.h"
 
-namespace megdnn {
-namespace rocm {
+using namespace megdnn;
+using namespace rocm;
+
+std::vector<BatchedMatrixMulForwardImpl::Algorithm*>
+BatchedMatrixMulForwardImpl::get_all_algorithms(const TensorLayout& A,
+                                                const TensorLayout& B,
+                                                const TensorLayout& C) {
+    AlgoBase::SizeArgs args{this, A, B, C};
+    return megdnn::get_all_algorithms<BatchedMatrixMulForwardImpl>(args);
+}
+
+BatchedMatrixMulForwardImpl::Algorithm*
+BatchedMatrixMulForwardImpl::get_algorithm_heuristic(
+        const TensorLayout& A, const TensorLayout& B, const TensorLayout& C,
+        size_t workspace_limit_in_bytes, bool reproducible) {
+    AlgoBase::SizeArgs args{this, A, B, C};
+    if (sm_algo_pack.blas.is_available_reproducible(args, reproducible,
+                                                    workspace_limit_in_bytes)) {
+        return &sm_algo_pack.blas;
+    }
+    if (reproducible) {
+        return megdnn::get_reproducible_algo<BatchedMatrixMulForwardImpl>(
+                sm_algo_pack.all_algos, args, workspace_limit_in_bytes,
+                "batched matrix mul forward");
+    } else {
+        return megdnn::get_usable_algo<BatchedMatrixMulForwardImpl>(
+                sm_algo_pack.all_algos, args, workspace_limit_in_bytes,
+                "batched matrix mul forward");
+    }
+}
+
+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);
+}
 
 void BatchedMatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
                                        _megdnn_tensor_out C,
                                        _megdnn_workspace workspace) {
     check_exec(A.layout, B.layout, C.layout, workspace.size);
-    auto dtype = A.layout.dtype;
-    megdnn_assert(dtype.category() == DTypeCategory::FLOAT &&
-                  param().format == param::MatrixMul::Format::DEFAULT);
-
-    if (dtype == dtype::Float32() ||
-        MEGDNN_FLOAT16_SELECT(dtype == dtype::Float16(), false)) {
-        auto batch = A.layout.shape[0];
-        auto m = C.layout.shape[1], n = C.layout.shape[2];
-        auto k = A.layout.shape[param().transposeA ? 1 : 2];
-        auto handle = concrete_handle(this->handle());
-        auto rocblas_handle_ = handle->get_rocblas_handle();
-
-        auto io32_c32 = [&]() {
-            auto zero = handle->zero_device();
-            auto one = handle->one_device();
-            rocblas_check(rocblas_sgemm_strided_batched(
-                    rocblas_handle_,
-                    param().transposeB ? rocblas_operation_transpose
-                                       : rocblas_operation_none,
-                    param().transposeA ? rocblas_operation_transpose
-                                       : rocblas_operation_none,
-                    n, m, k, one, B.ptr<dt_float32>(),
-                    (rocblas_int)(B.layout.stride[1]),
-                    (rocblas_int)(B.layout.stride[0]), A.ptr<dt_float32>(),
-                    (rocblas_int)(A.layout.stride[1]),
-                    (rocblas_int)(A.layout.stride[0]), zero,
-                    C.ptr<dt_float32>(), (rocblas_int)(C.layout.stride[1]),
-                    (rocblas_int)(C.layout.stride[0]), (rocblas_int)(batch)));
-        };
-
-#if !MEGDNN_DISABLE_FLOAT16
-        auto io16_c32 = [&]() {
-            auto zero = handle->zero_device();
-            auto one = handle->one_device();
-            int32_t solution_index = 0;
-            uint32_t flags = 1;
-            size_t ws_size = 0;
-
-            rocblas_check(rocblas_gemm_strided_batched_ex(
-                    rocblas_handle_,
-                    param().transposeB ? rocblas_operation_transpose
-                                       : rocblas_operation_none,
-                    param().transposeA ? rocblas_operation_transpose
-                                       : rocblas_operation_none,
-                    n, m, k, one, B.raw_ptr, rocblas_datatype_i8_r,
-                    B.layout.stride[1], B.layout.stride[0], A.raw_ptr,
-                    rocblas_datatype_i8_r, A.layout.stride[1],
-                    A.layout.stride[0], zero, C.raw_ptr, rocblas_datatype_i32_r,
-                    C.layout.stride[1], C.layout.stride[0], C.raw_ptr,
-                    rocblas_datatype_i32_r, C.layout.stride[1],
-                    C.layout.stride[0], batch, rocblas_datatype_i32_r,
-                    rocblas_gemm_algo_standard, solution_index, flags, &ws_size,
-                    nullptr));
-        };
-
-        auto io16_c16 = [&]() {
-            auto zero_half = handle->zero_device_h();
-            auto one_half = handle->one_device_h();
-            rocblas_check(rocblas_hgemm_strided_batched(
-                    rocblas_handle_,
-                    param().transposeB ? rocblas_operation_transpose
-                                       : rocblas_operation_none,
-                    param().transposeA ? rocblas_operation_transpose
-                                       : rocblas_operation_none,
-                    n, m, k, reinterpret_cast<const rocblas_half*>(one_half),
-                    static_cast<const rocblas_half*>(B.raw_ptr),
-                    B.layout.stride[1], B.layout.stride[0],
-                    static_cast<const rocblas_half*>(A.raw_ptr),
-                    A.layout.stride[1], A.layout.stride[0],
-                    reinterpret_cast<const rocblas_half*>(zero_half),
-                    static_cast<rocblas_half*>(C.raw_ptr), C.layout.stride[1],
-                    C.layout.stride[0], batch));
-
-        };
-#endif
-
-        if (dtype == dtype::Float32()) {
-            io32_c32();
-        }
-#if !MEGDNN_DISABLE_FLOAT16
-        else {
-            if (param().compute_mode == Param::ComputeMode::FLOAT32) {
-                io16_c32();
-            } else {
-                io16_c16();
-            }
-        }
-#endif
-    }
+    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);
 }
 
-}  // namespace rocm
-}  // namespace megdnn
-
 // vim: syntax=cpp.doxygen

dnn/src/rocm/batched_matrix_mul/opr_impl.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
 #include "megdnn/oprs.h"
@@ -17,36 +18,35 @@ namespace rocm {
 class BatchedMatrixMulForwardImpl : public BatchedMatrixMulForward {
 public:
     using BatchedMatrixMulForward::BatchedMatrixMulForward;
-    BatchedMatrixMulForwardImpl(Handle* handle)
-            : BatchedMatrixMul(handle),
-              m_opr(handle->create_operator<MatrixMul>()) {}
     void exec(_megdnn_tensor_in A, _megdnn_tensor_in B, _megdnn_tensor_out C,
               _megdnn_workspace workspace) override;
     size_t get_workspace_in_bytes(const TensorLayout&, const TensorLayout&,
-                                  const TensorLayout&) override {
-        return 0;
-    }
+                                  const TensorLayout&) override;
+
+    bool is_thread_safe() const override { return true; }
 
+    class AlgoBase;
+    class AlgoBlas;
+    class AlgoPack;
+    static const AlgoPack& algo_pack() { return sm_algo_pack; }
+
+    static AlgoBase* get_algo_from_desc(const AlgorithmDesc& desc);
+private:
     std::vector<Algorithm*> get_all_algorithms(
             const TensorLayout& /*A*/, const TensorLayout& /*B*/,
-            const TensorLayout& /*C*/) override {
-        return {};
-    }
+            const TensorLayout& /*C*/) override;
 
     Algorithm* get_algorithm_heuristic(const TensorLayout& /*A*/,
                                        const TensorLayout& /*B*/,
                                        const TensorLayout& /*C*/,
                                        size_t /*workspace_limit_in_bytes*/,
-                                       bool /* reproducible */) override {
-        return nullptr;
-    }
-
-    const char* get_algorithm_set_name() const override { return "DEFAULT"; }
+                                       bool /*reproducible*/) override;
 
-    bool is_thread_safe() const override { return true; }
+    const char* get_algorithm_set_name() const override {
+        return "ROCM BATCHED MATMUL";
+    }
 
-private:
-    std::unique_ptr<MatrixMul> m_opr;
+    static AlgoPack sm_algo_pack;
 };
 
 }  // namespace rocm

dnn/src/rocm/matrix_mul/algos.cpp

@@ -0,0 +1,62 @@
+/**
+ * \file dnn/src/rocm/matrix_mul/algos.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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 "src/rocm/matrix_mul/algos.h"
+#include "src/common/algo_base.h"
+
+using namespace megdnn;
+using namespace rocm;
+
+MatrixMulForwardImpl::AlgoPack::AlgoPack() {
+    all_algos.push_back(&blas);
+
+    for (auto&& algo : all_algos) {
+        m_all_algos_map.emplace(algo->info().desc, algo);
+    }
+}
+
+MatrixMulForwardImpl::AlgoPack MatrixMulForwardImpl::sm_algo_pack;
+
+MEGDNN_DEF_GET_ALGO_FROM_DESC(MatrixMulForwardImpl)
+
+MatrixMulForwardImpl::AlgoBase::SizeArgs::SizeArgs(MatrixMulForwardImpl* o,
+                                                   const TensorLayout& A,
+                                                   const TensorLayout& B,
+                                                   const TensorLayout& C)
+        : opr{o}, layout_a{A}, layout_b{B}, layout_c{C} {}
+
+MatrixMulForwardImpl::AlgoBase::ExecArgs::ExecArgs(MatrixMulForwardImpl* opr,
+                                                   _megdnn_tensor_in A,
+                                                   _megdnn_tensor_in B,
+                                                   _megdnn_tensor_out C,
+                                                   _megdnn_workspace workspace)
+        : SizeArgs(opr, A.layout, B.layout, C.layout),
+          tensor_a{A},
+          tensor_b{B},
+          tensor_c{C},
+          workspace{workspace} {}
+
+std::string MatrixMulForwardImpl::AlgoBase::SizeArgs::to_string() const {
+    auto&& param = opr->param();
+    size_t m = layout_a.shape[0], n = layout_b.shape[1],
+           k = layout_a.shape[param.transposeA ? 0 : 1];
+    MEGDNN_MARK_USED_VAR(m);
+    MEGDNN_MARK_USED_VAR(n);
+    MEGDNN_MARK_USED_VAR(k);
+    return megdnn_mangle(ssprintf(
+            "A={%zux%zu},B={%zux%zu},C={%zux%zu},Transpose A=%d,Transpose "
+            "B=%d,ldA=%zu,ldB=%zu,ldC=%zu",
+            m, k, k, n, m, n, param.transposeA, param.transposeB,
+            layout_a.stride[0], layout_b.stride[0], layout_c.stride[0]));
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/rocm/matrix_mul/algos.h

@@ -0,0 +1,118 @@
+/**
+ * \file dnn/src/rocm/matrix_mul/algos.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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/oprs.h"
+#include "src/common/algo_base.h"
+#include "src/common/metahelper.h"
+#include "src/common/utils.h"
+#include "src/rocm/matrix_mul/opr_impl.h"
+
+#include <memory>
+#include <unordered_map>
+
+namespace megdnn {
+namespace rocm {
+
+/*!
+ * \brief base class for matrix mul algos
+ *
+ */
+class MatrixMulForwardImpl::AlgoBase : public Algorithm {
+protected:
+    ~AlgoBase() = default;
+
+public:
+    enum class AlgoType : uint32_t {
+        ROCM_BLAS,
+    };
+    using Mapper = std::unordered_map<AlgorithmDesc, AlgoBase*>;
+
+    AlgoBase() : Algorithm() { m_handle_type = Handle::HandleType::ROCM; }
+    struct SizeArgs {
+        MatrixMulForwardImpl* opr;
+        TensorLayout layout_a, layout_b, layout_c;
+
+        std::string to_string() const;
+        SizeArgs(MatrixMulForwardImpl* opr, const TensorLayout& A,
+                 const TensorLayout& B, const TensorLayout& C);
+
+        bool can_be_treated_as_int8x8x32() const {
+            return layout_a.dtype.enumv() == layout_b.dtype.enumv() &&
+                   (layout_a.dtype.enumv() == DTypeEnum::Int8 ||
+                    layout_a.dtype.enumv() == DTypeEnum::QuantizedS8) &&
+                   (layout_c.dtype.enumv() == DTypeEnum::Int32 ||
+                    layout_c.dtype.enumv() == DTypeEnum::QuantizedS32) &&
+                   opr->param().format == param::MatrixMul::Format::DEFAULT;
+        }
+    };
+    struct ExecArgs : public SizeArgs {
+        TensorND tensor_a, tensor_b, tensor_c;
+        Workspace workspace;
+
+        ExecArgs(MatrixMulForwardImpl* opr, _megdnn_tensor_in A,
+                 _megdnn_tensor_in B, _megdnn_tensor_out C,
+                 _megdnn_workspace workspace);
+    };
+    virtual bool is_available(const SizeArgs& args) const = 0;
+    virtual size_t get_workspace_in_bytes(const SizeArgs& args) const = 0;
+    virtual void exec(const ExecArgs& args) const = 0;
+
+    bool is_available_wk(const SizeArgs& args, size_t limit) const {
+        return is_available(args) && get_workspace_in_bytes(args) <= limit;
+    }
+    bool is_available_reproducible(
+            const SizeArgs& args, bool reproducible = true,
+            size_t limit = std::numeric_limits<size_t>::max()) const {
+        return (!reproducible || is_reproducible()) &&
+               is_available_wk(args, limit);
+    }
+    AlgoBase& check_workspace(const SizeArgs& args,
+                              const Workspace& workspace) {
+        auto req = get_workspace_in_bytes(args);
+        megdnn_assert(
+                req <= workspace.size,
+                "matrix mul fwd algo %s: required workspace %zu bytes, got %zu",
+                name(), req, workspace.size);
+        return *this;
+    }
+};
+
+class MatrixMulForwardImpl::AlgoBlas final : public AlgoBase {
+public:
+    AlgoBlas() = default;
+    bool is_available(const SizeArgs& args) const override;
+    size_t get_workspace_in_bytes(const SizeArgs& /* args */) const override {
+        return 0_z;
+    }
+    const char* name() const override { return "BLAS"; }
+    void exec(const ExecArgs& args) const override;
+    bool is_reproducible() const override { return true; }
+    MEGDNN_DECL_ALGO_TYPE(ROCM_BLAS)
+};
+
+class MatrixMulForwardImpl::AlgoPack : NonCopyableObj {
+private:
+    AlgoBase::Mapper m_all_algos_map;
+
+public:
+    AlgoPack();
+    AlgoBlas blas;
+    std::vector<AlgoBase*> all_algos;
+
+    const AlgoBase::Mapper& all_algos_map() const { return m_all_algos_map; }
+};
+
+}  // namespace rocm
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/rocm/matrix_mul/blas.cpp

@@ -0,0 +1,162 @@
+/**
+ * \file dnn/src/rocm/matrix_mul/Blas.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 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 "src/rocm/matrix_mul/algos.h"
+
+#include "hcc_detail/hcc_defs_prologue.h"
+#include "src/rocm/handle.h"
+#include "src/rocm/utils.h"
+
+using namespace megdnn;
+using namespace rocm;
+
+bool MatrixMulForwardImpl::AlgoBlas::is_available(
+        const SizeArgs& args) const {
+    if (args.opr->param().format != param::MatrixMul::Format::DEFAULT)
+        return false;
+    if (args.layout_a.dtype == dtype::Float32() ||
+        args.layout_a.dtype == dtype::Float16()) {
+        return true;
+    } else if (args.layout_a.dtype.enumv() == DTypeEnum::Int8 ||
+               args.layout_a.dtype.enumv() == DTypeEnum::QuantizedS8) {
+        auto k = args.layout_a.shape[args.opr->param().transposeA ? 0 : 1];
+        //! see
+        //! https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/library/src/blas_ex/rocblas_gemm_ex.cpp:470
+        bool rocblas_int8x8x32_valid = true;
+        rocblas_int8x8x32_valid &= (k % 4 == 0);
+        rocblas_int8x8x32_valid &= (!args.opr->param().transposeB ||
+                                    args.layout_b.stride[0] % 4 == 0);
+        rocblas_int8x8x32_valid &= (!args.opr->param().transposeA ||
+                                    args.layout_a.stride[0] % 4 == 0);
+        return rocblas_int8x8x32_valid;
+    }
+    return false;
+}
+
+void MatrixMulForwardImpl::AlgoBlas::exec(const ExecArgs& args) const {
+    auto m = args.layout_c.shape[0], n = args.layout_c.shape[1];
+    auto k = args.layout_a.shape[args.opr->param().transposeA ? 0 : 1];
+    auto&& handle = concrete_handle(args.opr->handle());
+    auto rocblas_handle_ = handle->get_rocblas_handle();
+
+    auto sgemm = [&]() {
+        auto zero = handle->zero_device();
+        auto one = handle->one_device();
+        rocblas_check(rocblas_sgemm(
+                rocblas_handle_,
+                args.opr->param().transposeB ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                args.opr->param().transposeA ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                n, m, k, one, args.tensor_b.ptr<dt_float32>(),
+                args.layout_b.stride[0], args.tensor_a.ptr<dt_float32>(),
+                args.layout_a.stride[0], zero, args.tensor_c.ptr<dt_float32>(),
+                args.layout_c.stride[0]));
+    };
+
+#if !MEGDNN_DISABLE_FLOAT16
+    //! used for FLOAT_IO16xC32, not tested
+    auto gemm_ex = [&]() {
+        auto zero = handle->zero_device();
+        auto one = handle->one_device();
+        //! These two arguments for future use, see
+        //! https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/library/src/blas_ex/rocblas_gemm_ex.cpp
+        int32_t solution_index = 0;
+        uint32_t flags = 1;
+        size_t ws_size = 0;
+        auto gemm_ex_err = rocblas_gemm_ex(
+                rocblas_handle_,
+                args.opr->param().transposeB ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                args.opr->param().transposeA ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                n, m, k, one, args.tensor_b.raw_ptr, rocblas_datatype_f16_r,
+                args.layout_b.stride[0], args.tensor_a.raw_ptr,
+                rocblas_datatype_f16_r, args.layout_a.stride[0], zero,
+                args.tensor_c.raw_ptr, rocblas_datatype_f16_r,
+                args.layout_c.stride[0], args.tensor_c.raw_ptr,
+                rocblas_datatype_f16_r, args.layout_c.stride[0],
+                rocblas_datatype_f32_r, rocblas_gemm_algo_standard,
+                solution_index, flags, &ws_size, nullptr);
+        rocblas_check(gemm_ex_err);
+        MEGDNN_MARK_USED_VAR(ws_size);
+    };
+
+    auto hgemm = [&]() {
+        auto one_half = handle->one_device_h();
+        auto zero_half = handle->zero_device_h();
+        auto hgemm_err = rocblas_hgemm(
+                rocblas_handle_,
+                args.opr->param().transposeB ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                args.opr->param().transposeA ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                n, m, k, reinterpret_cast<const rocblas_half*>(one_half),
+                static_cast<const rocblas_half*>(args.tensor_b.raw_ptr),
+                args.layout_b.stride[0],
+                static_cast<const rocblas_half*>(args.tensor_a.raw_ptr),
+                args.layout_a.stride[0],
+                reinterpret_cast<const rocblas_half*>(zero_half),
+                static_cast<rocblas_half*>(args.tensor_c.raw_ptr),
+                args.layout_c.stride[0]);
+        rocblas_check(hgemm_err);
+    };
+#endif
+
+    if (args.opr->param().compute_mode == Param::ComputeMode::DEFAULT) {
+        if (args.layout_a.dtype == dtype::Float32()) {
+            sgemm();
+        }
+#if !MEGDNN_DISABLE_FLOAT16
+        else {
+            megdnn_assert(args.layout_a.dtype == dtype::Float16(),
+                          "invalid matmul data type");
+            hgemm();
+        }
+#endif
+    }
+#if !MEGDNN_DISABLE_FLOAT16
+    else if (args.opr->param().compute_mode == Param::ComputeMode::FLOAT32) {
+        megdnn_assert(args.layout_b.dtype == dtype::Float16() &&
+                              args.layout_c.dtype == dtype::Float16() &&
+                              args.layout_a.dtype == dtype::Float16(),
+                      "DataType::FLOAT_IO16xC32 is supported, when dtype of A, "
+                      "B, C are all Float16");
+        gemm_ex();
+    }
+#endif
+    else {
+        megdnn_assert(args.can_be_treated_as_int8x8x32());
+        int32_t solution_index = 0;
+        uint32_t flags = 1;
+        size_t ws_size = 0;
+        auto zero = handle->zero_device_i32();
+        auto one = handle->one_device_i32();
+        rocblas_check(rocblas_gemm_ex(
+                rocblas_handle_,
+                args.opr->param().transposeB ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                args.opr->param().transposeA ? rocblas_operation_transpose
+                                             : rocblas_operation_none,
+                n, m, k, one, args.tensor_b.raw_ptr, rocblas_datatype_i8_r,
+                args.layout_b.stride[0], args.tensor_a.raw_ptr,
+                rocblas_datatype_i8_r, args.layout_a.stride[0], zero,
+                args.tensor_c.raw_ptr, rocblas_datatype_i32_r,
+                args.layout_c.stride[0], args.tensor_c.raw_ptr,
+                rocblas_datatype_i32_r, args.layout_c.stride[0],
+                rocblas_datatype_i32_r, rocblas_gemm_algo_standard,
+                solution_index, flags, &ws_size, nullptr));
+        MEGDNN_MARK_USED_VAR(ws_size);
+    }
+
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/rocm/matrix_mul/opr_impl.cpp

@@ -13,147 +13,53 @@
 
 #include "src/rocm/utils.h"
 #include "src/rocm/handle.h"
+#include "./algos.h"
+#include "src/common/algo_chooser.h"
 
-namespace megdnn {
-namespace rocm {
+using namespace megdnn;
+using namespace rocm;
 
-void MatrixMulForwardImpl::exec(_megdnn_tensor_in A,
-        _megdnn_tensor_in B,
-        _megdnn_tensor_out C,
-        _megdnn_workspace workspace)
-{
-    check_exec(A.layout, B.layout, C.layout, workspace.size);
-
-    auto m = C.layout.shape[0], n = C.layout.shape[1];
-    auto k = A.layout.shape[param().transposeA ? 0 : 1];
-    auto handle = concrete_handle(this->handle());
-    auto rocblas_handle_ = handle->get_rocblas_handle();
-
-    auto sgemm = [&]() {
-        auto zero = handle->zero_device();
-        auto one = handle->one_device();
-        rocblas_check(rocblas_sgemm(
-                rocblas_handle_,
-                param().transposeB ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                param().transposeA ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                n, m, k, one, B.ptr<dt_float32>(), B.layout.stride[0],
-                A.ptr<dt_float32>(), A.layout.stride[0], zero,
-                C.ptr<dt_float32>(), C.layout.stride[0]));
-    };
-
-#if !MEGDNN_DISABLE_FLOAT16
-    //! used for FLOAT_IO16xC32, not tested
-    auto gemm_ex = [&]() {
-        auto zero = handle->zero_device();
-        auto one = handle->one_device();
-        //! These two arguments for future use, see
-        //! https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/library/src/blas_ex/rocblas_gemm_ex.cpp
-        int32_t solution_index = 0;
-        uint32_t flags = 1;
-        size_t ws_size = 0;
-        auto gemm_ex_err = rocblas_gemm_ex(
-                rocblas_handle_,
-                param().transposeB ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                param().transposeA ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                n, m, k, one, B.raw_ptr, rocblas_datatype_f16_r,
-                B.layout.stride[0], A.raw_ptr, rocblas_datatype_f16_r,
-                A.layout.stride[0], zero, C.raw_ptr, rocblas_datatype_f16_r,
-                C.layout.stride[0], C.raw_ptr, rocblas_datatype_f16_r,
-                C.layout.stride[0], rocblas_datatype_f32_r,
-                rocblas_gemm_algo_standard, solution_index, flags, &ws_size,
-                nullptr);
-        rocblas_check(gemm_ex_err);
-    };
-
-    auto hgemm = [&]() {
-        auto one_half = handle->one_device_h();
-        auto zero_half = handle->zero_device_h();
-        auto hgemm_err = rocblas_hgemm(
-                rocblas_handle_,
-                param().transposeB ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                param().transposeA ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                n, m, k, reinterpret_cast<const rocblas_half*>(one_half),
-                static_cast<const rocblas_half*>(B.raw_ptr), B.layout.stride[0],
-                static_cast<const rocblas_half*>(A.raw_ptr), A.layout.stride[0],
-                reinterpret_cast<const rocblas_half*>(zero_half),
-                static_cast<rocblas_half*>(C.raw_ptr), C.layout.stride[0]);
-        rocblas_check(hgemm_err);
-    };
-#endif
+std::vector<MatrixMulForwardImpl::Algorithm*>
+MatrixMulForwardImpl::get_all_algorithms(const TensorLayout& A,
+                                         const TensorLayout& B,
+                                         const TensorLayout& C) {
+    AlgoBase::SizeArgs args{this, A, B, C};
+    return megdnn::get_all_algorithms<MatrixMulForwardImpl>(args);
+}
 
-    if (param().compute_mode == Param::ComputeMode::DEFAULT) {
-        if (A.layout.dtype == dtype::Float32()) {
-            sgemm();
-        }
-#if !MEGDNN_DISABLE_FLOAT16
-        else {
-            megdnn_assert(A.layout.dtype == dtype::Float16(),
-                          "invalid matmul data type");
-            hgemm();
-        }
-#endif
+MatrixMulForwardImpl::Algorithm* MatrixMulForwardImpl::get_algorithm_heuristic(
+        const TensorLayout& A, const TensorLayout& B, const TensorLayout& C,
+        size_t workspace_limit_in_bytes, bool reproducible) {
+    AlgoBase::SizeArgs args{this, A, B, C};
+    if (sm_algo_pack.blas.is_available_reproducible(
+                args, reproducible, workspace_limit_in_bytes)) {
+        return &sm_algo_pack.blas;
     }
-#if !MEGDNN_DISABLE_FLOAT16
-    else if (param().compute_mode == Param::ComputeMode::FLOAT32) {
-        megdnn_assert(B.layout.dtype == dtype::Float16() &&
-                              C.layout.dtype == dtype::Float16() &&
-                              A.layout.dtype == dtype::Float16(),
-                      "DataType::FLOAT_IO16xC32 is supported, when dtype of A, "
-                      "B, C are all Float16");
-        gemm_ex();
-    }
-#endif
-    else if (A.layout.dtype == dtype::Int8() &&
-             B.layout.dtype == dtype::Int8() &&
-             C.layout.dtype == dtype::Int32()) {
-        //! see
-        //! https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/library/src/blas_ex/rocblas_gemm_ex.cpp:470
-        bool rocblas_int8x8x32_valid = true;
-        rocblas_int8x8x32_valid &= (k % 4 == 0);
-        rocblas_int8x8x32_valid &=
-                (!param().transposeB || B.layout.stride[0] % 4 == 0);
-        rocblas_int8x8x32_valid &=
-                (!param().transposeA || A.layout.stride[0] % 4 == 0);
-        megdnn_assert(rocblas_int8x8x32_valid,
-                      "rocblas int8x8x32 matmul requires K must be a multiple "
-                      "of 4, and/or LDA/LDB based on transpose mode"
-                      "get: %zu, is_trans_b = %d, %zu, is_trans_a = %d, %zu",
-                      k, param().transposeB, B.layout.stride[0],
-                      param().transposeA, A.layout.stride[0]);
-        int32_t solution_index = 0;
-        uint32_t flags = 1;
-        size_t ws_size = 0;
-        auto zero = handle->zero_device_i32();
-        auto one = handle->one_device_i32();
-        rocblas_check(rocblas_gemm_ex(
-                rocblas_handle_,
-                param().transposeB ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                param().transposeA ? rocblas_operation_transpose
-                                   : rocblas_operation_none,
-                n, m, k, one, B.raw_ptr, rocblas_datatype_i8_r,
-                B.layout.stride[0], A.raw_ptr, rocblas_datatype_i8_r,
-                A.layout.stride[0], zero, C.raw_ptr, rocblas_datatype_i32_r,
-                C.layout.stride[0], C.raw_ptr, rocblas_datatype_i32_r,
-                C.layout.stride[0], rocblas_datatype_i32_r,
-                rocblas_gemm_algo_standard, solution_index, flags, &ws_size,
-                nullptr));
+    if (reproducible) {
+        return megdnn::get_reproducible_algo<MatrixMulForwardImpl>(
+                sm_algo_pack.all_algos, args, workspace_limit_in_bytes,
+                "matrix mul forward");
     } else {
-        megdnn_assert((A.layout.dtype == dtype::Int8() &&
-                       B.layout.dtype == dtype::Int8() &&
-                       C.layout.dtype == dtype::Int16()),
-                      "invalid matmul data type");
-        megdnn_throw("cuda matmul does not support INT8x8x16 now");
+        return megdnn::get_usable_algo<MatrixMulForwardImpl>(
+                sm_algo_pack.all_algos, args, workspace_limit_in_bytes,
+                "matrix mul forward");
     }
 }
 
-} // namespace rocm
-} // namespace megdnn
+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);
+}
+
+void MatrixMulForwardImpl::exec(_megdnn_tensor_in A, _megdnn_tensor_in B,
+                                _megdnn_tensor_out C,
+                                _megdnn_workspace workspace) {
+    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);
+}
 
 // vim: syntax=cpp.doxygen

dnn/src/rocm/matrix_mul/opr_impl.h

@@ -20,29 +20,32 @@ public:
     void exec(_megdnn_tensor_in A, _megdnn_tensor_in B, _megdnn_tensor_out C,
               _megdnn_workspace workspace) override;
     size_t get_workspace_in_bytes(const TensorLayout&, const TensorLayout&,
-                                  const TensorLayout&) override {
-        return 0;
-    }
+                                  const TensorLayout&) override;
 
     bool is_thread_safe() const override { return true; }
+
+    class AlgoBase;
+    class AlgoBlas;
+    class AlgoPack;
+    static const AlgoPack& algo_pack() { return sm_algo_pack; }
+
+    static AlgoBase* get_algo_from_desc(const AlgorithmDesc& desc);
+
 private:
     std::vector<Algorithm*> get_all_algorithms(
             const TensorLayout& /*A*/, const TensorLayout& /*B*/,
-            const TensorLayout& /*C*/) override {
-        return {};
-    }
+            const TensorLayout& /*C*/) override;
 
     Algorithm* get_algorithm_heuristic(const TensorLayout& /*A*/,
                                        const TensorLayout& /*B*/,
                                        const TensorLayout& /*C*/,
                                        size_t /*workspace_limit_in_bytes*/,
-                                       bool /*reproducible*/) override {
-        return nullptr;
-    }
-
+                                       bool /*reproducible*/) override;
     const char* get_algorithm_set_name() const override {
         return "ROCM MATMUL";
     }
+
+    static AlgoPack sm_algo_pack;
 };
 
 }  // namespace rocm

dnn/test/fallback/matrix_mul.cpp

@@ -46,6 +46,37 @@ TEST_F(FALLBACK, MATRIX_MUL) {
     }
 }
 
+TEST_F(FALLBACK, MATRIX_MUL_NAIVE) {
+    Checker<MatrixMul> checker(handle());
+    checker.set_before_exec_callback(AlgoChecker<MatrixMul>("FB_NAIVE"));
+    using Param = MatrixMul::Param;
+    auto args = matrix_mul::get_matmul_args();
+    for (auto arg : args) {
+        auto m = arg.m, n = arg.n, k = arg.k;
+        auto mask = arg.mask;
+        Param param;
+        param.transposeA = mask & 1;
+        param.transposeB = mask & 2;
+        TensorShape AS, BS, CS;
+
+        if (param.transposeA)
+            AS = TensorShape{k, m};
+        else
+            AS = TensorShape{m, k};
+        if (param.transposeB)
+            BS = TensorShape{n, k};
+        else
+            BS = TensorShape{k, n};
+        CS = TensorShape{m, n};
+        TensorLayout AL, BL, CL;
+        AL = TensorLayout(AS, dtype::Float32());
+        BL = TensorLayout(BS, dtype::Float32());
+        CL = TensorLayout(CS, dtype::Float32());
+        checker.set_param(param);
+        checker.execl({AL, BL, CL});
+    }
+}
+
 TEST_F(FALLBACK, BATCHED_MATRIX_MUL) {
 
     Checker<BatchedMatrixMul> checker(handle());

dnn/test/naive/matrix_mul.cpp

@@ -232,7 +232,7 @@ TEST_F(NAIVE, MATRIX_MUL_QUANTIZEDS4_4x4x16) {
                               2,  5,  3,  3,  7,   4, -7,  1,
                              -5,  7, -4, -1, -1,   2,  4,  1,
                               7,  2, -6, -2, -6,   3,  4,  4,
-                             -2,  2,  3,  0,  6,   5,  3,  4,  
+                             -2,  2,  3,  0,  6,   5,  3,  4,
                              -1, -1, -5,  5,  2,   5,  1,  4,
                               6,  2,  0,  0,  3,   2,  2,  1,
                              -4, -3,  7,  5,  0,   3,  2,  3}),
@@ -243,7 +243,7 @@ TEST_F(NAIVE, MATRIX_MUL_QUANTIZEDS4_4x4x16) {
                              3,  -1,  2,  2,   7,  3,  6,  0,
                              5,   4,  0,  2,   2,  3,  3,  2,
                              1,  -8, -7, -6,   0, -5, -4,  4,
-                            -3,   7,  1,  6,  -2,  2, -1,  5,  
+                            -3,   7,  1,  6,  -2,  2, -1,  5,
                              2,   0,  7,  6,   5,  4,  3,  2,
                              0,   0,  1,  0,   5,  2,  2,  6}),
                     {}},