refactor(megdnn): refactor bfloat16 matmul to recursive inteface

GitOrigin-RevId: 641c508aec

dnn/src/cuda/matrix_mul/algos.cpp

@@ -31,8 +31,7 @@ MatrixMulForwardImpl::AlgoPack::AlgoPack() {
 #endif
     all_algos.push_back(&naive);
 #if !MEGDNN_DISABLE_FLOAT16
-    cublas_bfloat16 = std::make_unique<AlgoBFloat16>(&cublas);
-    all_algos.push_back(cublas_bfloat16.get());
+    all_algos.push_back(&bfloat16);
 #endif
 
     for (auto&& algo : all_algos) {

dnn/src/cuda/matrix_mul/algos.h

@@ -148,25 +148,20 @@ public:
 #if !MEGDNN_DISABLE_FLOAT16
 class MatrixMulForwardImpl::AlgoBFloat16 final : public AlgoBase {
 public:
-    AlgoBFloat16(MatrixMulForwardImpl::AlgoBase*);
     bool is_available(const SizeArgs& args) const override;
     size_t get_workspace_in_bytes(const SizeArgs& args) const override;
-    const char* name() const override { return m_name.c_str(); }
     void exec(const ExecArgs& args) const override;
-    bool is_reproducible() const override { return true; }
-    MEGDNN_DECL_ALGO_TYPE(CUDA_NAIVE)
+    MEGDNN_DECL_ALGO_TYPE(CUDA_BFLOAT16)
 
-    std::string param() const override {
-        std::string ret;
-        serialize_write_pod(m_algorithm, ret);
-        return ret;
-    }
+    std::vector<SearchItem> get_subopr_list(
+            const TensorLayoutArray& layouts,
+            const OperatorBase* opr) const override;
+
+    const char* name() const override { return "MATMUL_BFLOAT16"; }
+    bool is_reproducible() const override { return true; }
 
 private:
-    MatrixMulForwardImpl::AlgoBase* m_algorithm = nullptr;
-    std::string m_name;
     WorkspaceBundle get_workspace_bundle(void* ptr, const SizeArgs& args) const;
-    SizeArgs float_args(const SizeArgs& args) const;
 };
 #endif
 
@@ -185,7 +180,7 @@ public:
     AlgoCuBlasLt cublas_lt;
 #endif
 #if !MEGDNN_DISABLE_FLOAT16
-    std::unique_ptr<AlgoBFloat16> cublas_bfloat16;
+    AlgoBFloat16 bfloat16;
 #endif
     std::vector<AlgoBase*> all_algos;
 

dnn/src/cuda/matrix_mul/bfloat16.cpp

@@ -6,59 +6,87 @@
  *
  * 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 "src/cuda/handle.h"
 #include "src/cuda/matrix_mul/algos.h"
 #include "src/cuda/utils.h"
+#include "src/common/algo_chooser.h"
 
 using namespace megdnn;
 using namespace cuda;
 
-MatrixMulForwardImpl::AlgoBFloat16::AlgoBFloat16(
-        MatrixMulForwardImpl::AlgoBase* algorithm)
-        : m_algorithm(algorithm) {
-    megdnn_assert_internal(algorithm);
-    m_name = ssprintf("MATMUL_BFLOAT16:%s", m_algorithm->name());
-}
-
-MatrixMulForwardImpl::AlgoBase::SizeArgs
-MatrixMulForwardImpl::AlgoBFloat16::float_args(const SizeArgs& args) const {
-    auto new_args = args;
+namespace {
+std::pair<TensorLayoutArray, MatrixMulForwardImpl::Param> sub_opr_config(
+        const TensorLayoutArray& layouts, const MatrixMulForwardImpl* opr) {
+    megdnn_assert(layouts.size() == 3);
+    std::pair<TensorLayoutArray, MatrixMulForwardImpl::Param> ret;
+    ret.first = layouts;
     auto change_dtype = [](TensorLayout& layout) {
         if (layout.dtype == dtype::BFloat16()) {
             layout.dtype = dtype::Float32();
         }
     };
-    change_dtype(new_args.layout_a);
-    change_dtype(new_args.layout_b);
-    change_dtype(new_args.layout_c);
-    return new_args;
+    change_dtype(ret.first[0]);
+    change_dtype(ret.first[1]);
+    change_dtype(ret.first[2]);
+
+    ret.second = opr->param();
+    ret.second.compute_mode = MatrixMulForwardImpl::Param::ComputeMode::DEFAULT;
+    return ret;
+}
+}  // namespace
+
+std::vector<Algorithm::SearchItem>
+MatrixMulForwardImpl::AlgoBFloat16::get_subopr_list(
+        const TensorLayoutArray& layouts, const OperatorBase* opr) const {
+    auto&& config = sub_opr_config(
+            layouts, static_cast<const MatrixMulForwardImpl*>(opr));
+
+    std::string param_str;
+    Algorithm::serialize_write_pod(config.second, param_str);
+    return {{Algorithm::OprType::MATRIX_MUL_FORWARD, param_str, config.first}};
 }
 
 bool MatrixMulForwardImpl::AlgoBFloat16::is_available(
         const SizeArgs& args) const {
-    auto fargs = float_args(args);
+    auto&& config = sub_opr_config(
+            {args.layout_a, args.layout_b, args.layout_c}, args.opr);
+    auto matmul_opr = args.opr->handle()->create_operator<MatrixMulForward>();
+    matmul_opr->param() = config.second;
+
     return args.layout_a.dtype == dtype::BFloat16() &&
-        m_algorithm->is_available(fargs);
+           get_algorithm(static_cast<MatrixMulForwardImpl*>(matmul_opr.get()),
+                         config.first[0], config.first[1], config.first[2]);
 }
 
 WorkspaceBundle MatrixMulForwardImpl::AlgoBFloat16::get_workspace_bundle(
         void* ptr, const SizeArgs& args) const {
-    auto fargs = float_args(args);
+    auto matmul_opr = args.opr->handle()->create_operator<MatrixMulForward>();
+    if (args.opr->execution_policy().algo.valid()) {
+        megdnn_assert(args.opr->execution_policy().sub_policy.size() == 1);
+        matmul_opr->execution_policy() =
+                args.opr->execution_policy().sub_policy[0];
+    }
+    auto&& config = sub_opr_config(
+            {args.layout_a, args.layout_b, args.layout_c}, args.opr);
+    matmul_opr->param() = config.second;
+
     SmallVector<size_t> sizes;
-    auto get_workspace = [&sizes](const TensorLayout& src) {
-        TensorLayout dst = src;
-        if (dst.dtype == dtype::BFloat16()) {
-            dst.dtype = dtype::Float32();
+    auto get_workspace = [&sizes](const TensorLayout& src,
+                                  const TensorLayout& dst) {
+        if (src.dtype != dst.dtype) {
             sizes.push_back(dst.span().dist_byte());
         }
     };
-    get_workspace(args.layout_a);
-    get_workspace(args.layout_b);
-    get_workspace(args.layout_c);
-    sizes.push_back(m_algorithm->get_workspace_in_bytes(fargs));
+
+    get_workspace(args.layout_a, config.first[0]);
+    get_workspace(args.layout_b, config.first[1]);
+    get_workspace(args.layout_c, config.first[2]);
+    sizes.push_back(matmul_opr->get_workspace_in_bytes(
+            config.first[0], config.first[1], config.first[2]));
     return {ptr, std::move(sizes)};
 }
 
@@ -82,7 +110,12 @@ void MatrixMulForwardImpl::AlgoBFloat16::exec(const ExecArgs& args) const {
                 args.opr->handle()->create_operator<MatrixMulForward>();
         matmul_opr->param() = args.opr->param();
         matmul_opr->param().compute_mode = Param::ComputeMode::DEFAULT;
-        matmul_opr->execution_policy() = {m_algorithm->desc(), {}};
+        if (args.opr->execution_policy().algo.valid()) {
+            megdnn_assert(args.opr->execution_policy().sub_policy.size() == 1);
+            matmul_opr->execution_policy() =
+                    args.opr->execution_policy().sub_policy[0];
+        }
+
         matmul_opr->exec(a, b, c, ctypecvt.workspace());
     }
     ctypecvt.comp_to_dst_type(c, args.tensor_c);

dnn/test/cuda/matrix_mul.cpp

@@ -218,6 +218,9 @@ TEST_F(CUDA, MATRIX_MUL) {
                 B = TensorShape{k, n};
             if (dtype == dtype::BFloat16()) {
                 param.compute_mode = param::MatrixMul::ComputeMode::FLOAT32;
+                checker.set_before_exec_callback(
+                        AlgoChecker<MatrixMulForward>(ExecutionPolicyAlgoName{
+                                "MATMUL_BFLOAT16", {{"CUBLAS", {}}}}));
             }
             checker.set_param(param)
                     .set_dtype(0, stype)
@@ -228,6 +231,10 @@ TEST_F(CUDA, MATRIX_MUL) {
                                          ? 5e-2
                                          : 5e-3)
                     .execs({A, B, {}});
+            if (dtype == dtype::BFloat16()) {
+                checker.reset_before_exec_callback();
+                checker.opr()->execution_policy() = {};
+            }
         }
     }