feat(jit): add memfwd in jit executor opr

GitOrigin-RevId: b58860bbe87582023d96fdfe9d2cb5c6c93b8731

dnn/src/common/pooling.cpp

@@ -92,7 +92,7 @@ void PoolingBase::deduce_layout_fwd(const TensorLayout& src,
     size_t sw = this->param().stride_w;
     size_t ph = this->param().pad_h;
     size_t pw = this->param().pad_w;
-    if (ph < fh && pw < fw) {
+    if (ph >= fh || pw >= fw) {
         megdnn_log_error(
                 "pooling padding size (%zu %zu) should not be bigger than "
                 "window size (%zu %zu), it only can be used in CaffePooling",

src/jit/impl/executor_opr.cpp

@@ -135,6 +135,13 @@ void JITExecutor::init_output_mem_plan(bool dynamic) {
     m_args.need_update = true;
 }
 
+void JITExecutor::mem_plan_fwd_in2out_writable() {
+    //! currently mem fwd only support elemwise fusion
+    if (m_feature_bits != JITFeatureBits::NONE) return;
+    mixin_mem_plan_fwd_in2out_writable(*this);
+}
+
+
 SymbolVar JITExecutor::make(const InternalGraphPtr& internal_graph,
                             const VarNodeArray& inputs,
                             const OperatorNodeConfig& config) {

src/jit/include/megbrain/jit/executor_opr.h

@@ -13,6 +13,7 @@
 
 #include "megbrain/graph/operator_node.h"
 #include "megbrain/jit/internal_graph.h"
+#include "megbrain/opr/internal/identical_fwd.h"
 
 #if MGB_JIT
 
@@ -31,7 +32,8 @@ class Compiler;
  * JITExecutor generates runtime Args for this specific inputs, and calls
  * methods in Compiler to get the Executable object for actual computing.
  */
-MGB_DEFINE_OPR_CLASS(JITExecutor, cg::SingleCNOperatorNodeBase) // {
+MGB_DEFINE_OPR_CLASS(JITExecutor, cg::SingleCNOperatorNodeBase,
+        opr::mixin::FwdIn2OutWritableHelper) // {
     using ModeTrait = megdnn::Elemwise::ModeTrait;
 
     InternalGraphPtr m_internal_graph;
@@ -57,6 +59,8 @@ public:
 
     void init_output_mem_plan(bool dynamic) override;
 
+    void mem_plan_fwd_in2out_writable() override;
+
     const InternalGraph& internal_graph() const { return *m_internal_graph; }
 
     const InternalGraphPtr internal_graph_ptr() const {

src/jit/test/fusion.cpp

@@ -137,6 +137,12 @@ void run<basic>(Backend backend, CompNode cn) {
     // only one broadcast is allowed in JIT fusion
     ASSERT_EQ(1u, jits[0]->input().size());
     ASSERT_EQ(4u, jits[1]->input().size());
+
+    //! check memfwd
+    ASSERT_EQ(prev_dev_ptr(jits[0]->input(0)),
+              prev_dev_ptr(jits[0]->output(0)));
+    ASSERT_EQ(prev_dev_ptr(jits[1]->input(0)),
+              prev_dev_ptr(jits[1]->output(0)));
 }
 
 template <>

src/opr/impl/basic_arith.cpp

@@ -338,32 +338,7 @@ void Elemwise::broadcast_collective_collapse(
 }
 
 void Elemwise::mem_plan_fwd_in2out_writable() {
-    auto &&inp = input();
-    auto isize = inp.size();
-    mgb_assert(isize <= 6);
-    bool have_conflict[6] = {false};
-    for (size_t i = 0; i < isize; ++i) {
-        for (size_t j = i + 1; j < isize; ++j) {
-            auto type = cg::get_mem_plan_intersection_type(inp[i], inp[j]);
-            using Type = cg::MemPlanIntersectionType;
-            bool overlap = type == Type::OVERLAP;
-            bool self_fwd = type == Type::IDENTICAL &&
-                            (!inp[i]->layout().is_contiguous() ||
-                             !inp[j]->layout().is_contiguous());
-            if (overlap || self_fwd) {
-                have_conflict[i] = true;
-                have_conflict[j] = true;
-            }
-        }
-    }
-    auto o = output(0);
-    for (size_t idx = 0; idx < isize; ++ idx) {
-        auto i = inp[idx];
-        // equal shape means no broadcast
-        if (!have_conflict[idx] &&
-                o->shape().eq_shape(i->shape()) && i->layout().is_contiguous())
-            o->set_fwd_in2out_writable(i);
-    }
+    mixin_mem_plan_fwd_in2out_writable(*this);
 }
 
 void Elemwise::scn_do_execute() {

src/opr/impl/internal/identical_fwd.cpp

@@ -33,6 +33,37 @@ void mixin::init_rt_force_dynamic_mem_alloc_imply_chain_for_dyn_pass_i2o(
     valid_out->add_rt_force_dynamic_mem_alloc_imply_chain(opr.input(0));
 }
 
+/* ===================== FwdIn2OutWritableHelper  ===================== */
+void FwdIn2OutWritableHelper::mixin_mem_plan_fwd_in2out_writable(
+        OperatorNodeBase& opr) {
+    auto&& inp = opr.input();
+    auto isize = inp.size();
+    std::vector<bool> have_conflict(isize, false);
+    for (size_t i = 0; i < isize; ++i) {
+        for (size_t j = i + 1; j < isize; ++j) {
+            auto type = cg::get_mem_plan_intersection_type(inp[i], inp[j]);
+            using Type = cg::MemPlanIntersectionType;
+            bool overlap = type == Type::OVERLAP;
+            bool self_fwd = type == Type::IDENTICAL &&
+                            (!inp[i]->layout().is_contiguous() ||
+                             !inp[j]->layout().is_contiguous());
+            if (overlap || self_fwd) {
+                have_conflict[i] = true;
+                have_conflict[j] = true;
+            }
+        }
+    }
+    auto o = opr.output(0);
+    for (size_t idx = 0; idx < isize; ++ idx) {
+        auto i = inp[idx];
+        // equal shape means no broadcast
+        if (!have_conflict[idx] && o->shape().eq_shape(i->shape()) &&
+            o->dtype().enumv() == i->dtype().enumv() &&
+            i->layout().is_contiguous())
+            o->set_fwd_in2out_writable(i);
+    }
+}
+
 /* ===================== ReadonlyFwdHelper ===================== */
 
 void ReadonlyFwdHelper::mixin_rofwd_init_mem_plan(OperatorNodeBase &opr) {

src/opr/include/megbrain/opr/basic_arith.h

@@ -58,7 +58,8 @@ namespace intl {
  * The operands are broadcasted automatically on dimensions of shape one to
  * match shapes of each other; it works like broadcasting in numpy.
  */
-MGB_DEFINE_OPR_CLASS(Elemwise, intl::ElemwiseBase) // {
+MGB_DEFINE_OPR_CLASS(Elemwise, intl::ElemwiseBase,
+                     mixin::FwdIn2OutWritableHelper) // {
     using ModeTrait = megdnn::Elemwise::ModeTrait;
 
     public:

src/opr/include/megbrain/opr/internal/identical_fwd.h

@@ -19,6 +19,19 @@ namespace opr {
 
 namespace mixin {
 
+/*!
+ * \brief mixin for operators which essentially works by forward input to output
+ */
+class FwdIn2OutWritableHelper : public cg::OperatorNodeMixinBase {
+protected:
+    /*!
+     * \brief call this function in mem_plan_fwd_in2out_writable(),
+     *     this function will check if the input have conflict to find if the
+     *     output can be forward.
+     */
+    void mixin_mem_plan_fwd_in2out_writable(OperatorNodeBase &opr);
+};
+
 //! for internal use by DynamicOutputIfInputDynamic
 void init_rt_force_dynamic_mem_alloc_imply_chain_for_dyn_pass_i2o(
         OperatorNodeBase &opr);