fix(imperative): add alloc TensorPtr in imperative

GitOrigin-RevId: 1b438fc436

imperative/src/impl/dnn_op_helper.h

@@ -27,10 +27,16 @@ struct DnnOprCaller {
         return mgb::opr::intl::create_megdnn_opr<Opr>(cn);
     }
 
-    megdnn::Workspace create_workspace(TensorLayout layout) {
-        dev_tensor = Tensor::make(layout, cn)->dev_tensor();
-        workspace =
-                megdnn::Workspace(dev_tensor.raw_ptr(), dev_tensor.storage().size());
+    Workspace create_workspace(size_t sz) {
+        if (workspace.raw_ptr) {
+            mgb_throw(MegBrainError, "workspace should not be applicated many times");
+        }
+        if (sz) {
+            TensorLayout layout({sz}, dtype::Byte());
+            dev_tensor = Tensor::make(layout, cn)->dev_tensor();
+            workspace = megdnn::Workspace(
+                    dev_tensor.raw_ptr(), dev_tensor.storage().size());
+        }
         return workspace;
     }
 

imperative/src/impl/ops/adaptive_pooling.cpp

@@ -135,21 +135,16 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     param.window_w = IW - (OW - 1) * param.stride_w;
 
     TensorND src = inputs[0]->dnn_tensor();
-    DeviceTensorND dst =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, dst_layout);
+    auto dst = Tensor::make(dst_layout, cn);
 
     size_t sz = setup_algo<megdnn::Pooling>(
             {src_layout, dst_layout}, dnn_opr.op.get(), 0, false, false, cn,
             ::megdnn::param::ExecutionPolicy{}, false);
 
-    megdnn::Workspace dnn_wk;
-    if (sz) {
-        TensorLayout w_layout({sz}, dtype::Byte());
-        dnn_wk = dnn_opr.create_workspace(w_layout);
-    }
-    dnn_opr.op->exec(src, dst.as_megdnn(), dnn_wk);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
+    dnn_opr.op->exec(src, dst->dnn_tensor(), dnn_wk);
 
-    return {Tensor::make(dst)};
+    return {dst};
 }
 
 OP_TRAIT_REG(AdaptivePooling, AdaptivePooling)

imperative/src/impl/ops/batch_norm.cpp

@@ -160,10 +160,8 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     bool empty_input = src_layout.is_empty();
     size_t nr_inp = inputs.size();
 
-    DeviceTensorND reserve;
     size_t sz = 0, rsz = 0;
 
-    TensorLayout w_layout({sz}, dtype::Byte());
     TensorLayout r_layout({rsz}, dtype::Byte());
 
     if (!empty_input) {
@@ -172,79 +170,71 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
                 src_layout, src_layout, src_layout);
         rsz = dnn_opr.op->get_reserve_in_bytes(src_layout);
 
-        w_layout = TensorLayout({sz}, dtype::Byte());
         r_layout = TensorLayout({rsz}, dtype::Byte());
     }
-    auto dnn_wk = dnn_opr.create_workspace(w_layout);
-    reserve = BlobManager::inst()->alloc_workspace_with_defrag(comp_node, r_layout);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
+    auto reserve = Tensor::make(r_layout, comp_node);
 
     // alloc memory
-    DeviceTensorND y =
-            BlobManager::inst()->alloc_workspace_with_defrag(comp_node, src_layout);
+    auto y = Tensor::make(src_layout, comp_node);
 
-    DeviceTensorND save_mean =
-            BlobManager::inst()->alloc_workspace_with_defrag(comp_node, scale_layout);
-    DeviceTensorND save_variance =
-            BlobManager::inst()->alloc_workspace_with_defrag(comp_node, scale_layout);
+    auto save_mean = Tensor::make(scale_layout, comp_node);
+
+    auto save_variance = Tensor::make(scale_layout, comp_node);
 
     if (op_def.fwd_mode == ::megdnn::param::BN::FwdMode::INFERENCE) {
         if (!empty_input)
             dnn_opr.op->exec(
                     inp_tensornds[0], inp_tensornds[1], inp_tensornds[2],
-                    inp_tensornds[3], inp_tensornds[4], save_mean.as_megdnn(),
-                    save_variance.as_megdnn(), reserve.as_megdnn(), y.as_megdnn(),
+                    inp_tensornds[3], inp_tensornds[4], save_mean->dnn_tensor(),
+                    save_variance->dnn_tensor(), reserve->dnn_tensor(), y->dnn_tensor(),
                     dnn_wk);
-        return {inputs[3], inputs[4], Tensor::make(reserve), Tensor::make(y)};
+        return {inputs[3], inputs[4], reserve, y};
     } else {
-        DeviceTensorND mean, variance;
         if (nr_inp == 5) {
-            mean = BlobManager::inst()->alloc_workspace_with_defrag(
-                    comp_node, scale_layout);
-            variance = BlobManager::inst()->alloc_workspace_with_defrag(
-                    comp_node, scale_layout);
+            auto mean = Tensor::make(scale_layout, comp_node);
+
+            auto variance = Tensor::make(scale_layout, comp_node);
 
             megdnn::RefPtr src_ptr1(
                     inp_tensornds[3].get_ref_ptr().get_ptr(), inputs[3]->offset());
             megdnn::RefPtr dst_ptr1(
-                    mean.storage().get_ref_ptr(), mean.storage().offset(), false);
+                    mean->dev_tensor().storage().get_ref_ptr(),
+                    mean->dev_tensor().storage().offset(), false);
             comp_node.peer_copy_to_ref(
                     comp_node, dst_ptr1, src_ptr1, scale_layout.span().high_byte);
 
             megdnn::RefPtr src_ptr2(
                     inp_tensornds[4].get_ref_ptr().get_ptr(), inputs[4]->offset());
             megdnn::RefPtr dst_ptr2(
-                    variance.storage().get_ref_ptr(), variance.storage().offset(),
-                    false);
+                    variance->dev_tensor().storage().get_ref_ptr(),
+                    variance->dev_tensor().storage().offset(), false);
             comp_node.peer_copy_to_ref(
                     comp_node, dst_ptr2, src_ptr2, scale_layout.span().high_byte);
 
             if (!empty_input)
                 dnn_opr.op->exec(
                         inp_tensornds[0], inp_tensornds[1], inp_tensornds[2],
-                        mean.as_megdnn(), variance.as_megdnn(), save_mean.as_megdnn(),
-                        save_variance.as_megdnn(), reserve.as_megdnn(), y.as_megdnn(),
-                        dnn_wk);
+                        mean->dnn_tensor(), variance->dnn_tensor(),
+                        save_mean->dnn_tensor(), save_variance->dnn_tensor(),
+                        reserve->dnn_tensor(), y->dnn_tensor(), dnn_wk);
 
-            return {Tensor::make(mean),      Tensor::make(variance),
-                    Tensor::make(save_mean), Tensor::make(save_variance),
-                    Tensor::make(reserve),   Tensor::make(y)};
+            return {mean, variance, save_mean, save_variance, reserve, y};
         }
 
         TensorLayout m_layout({0}, scale_layout.dtype);
-        mean = BlobManager::inst()->alloc_workspace_with_defrag(comp_node, m_layout);
-        variance =
-                BlobManager::inst()->alloc_workspace_with_defrag(comp_node, m_layout);
+        auto mean = Tensor::make(m_layout, comp_node);
+        auto variance = Tensor::make(m_layout, comp_node);
 
         if (!empty_input) {
             dnn_opr.op->exec(
                     inp_tensornds[0], inp_tensornds[1], inp_tensornds[2],
-                    mean.as_megdnn(), variance.as_megdnn(), save_mean.as_megdnn(),
-                    save_variance.as_megdnn(), reserve.as_megdnn(), y.as_megdnn(),
+                    mean->dnn_tensor(), variance->dnn_tensor(), save_mean->dnn_tensor(),
+                    save_variance->dnn_tensor(), reserve->dnn_tensor(), y->dnn_tensor(),
                     dnn_wk);
         }
 
-        return {Tensor::make(save_mean), Tensor::make(save_variance),
-                Tensor::make(reserve), Tensor::make(y)};
+        return {save_mean, save_variance, reserve, y};
     }
 }
 

imperative/src/impl/ops/cond_take.cpp

@@ -44,10 +44,9 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
         DnnOprCaller<megdnn::CondTake> dnn_op(inp->comp_node());
         dnn_op.op->param().val = 1;
 
-        TensorLayout m_layout(
-                {dnn_op.op->get_workspace_in_bytes(inp->layout())}, dtype::Byte());
+        size_t sz = dnn_op.op->get_workspace_in_bytes(inp->layout());
 
-        auto dnn_workspace = dnn_op.create_workspace(m_layout);
+        auto dnn_workspace = dnn_op.create_workspace(sz);
 
         dnn_op.op->exec(
                 inp->dev_tensor().as_megdnn(), msk->dev_tensor().as_megdnn(),

imperative/src/impl/ops/convolution.cpp

@@ -165,11 +165,10 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     TensorLayout empty_shp({0}, inputs[0]->dtype());
     empty_shp.ndim = 0;
 
-    DeviceTensorND empty_bias =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, empty_shp);
+    auto empty_bias = Tensor::make(empty_shp, cn);
 
-    inp_tensornds[2] = empty_bias.as_megdnn();
-    inp_tensornds[3] = empty_bias.as_megdnn();
+    inp_tensornds[2] = empty_bias->dnn_tensor();
+    inp_tensornds[3] = empty_bias->dnn_tensor();
 
     size_t sz = setup_algo<megdnn::ConvBiasForward>(
             {inp_shapes[0], inp_shapes[1], empty_shp, empty_shp, oup_shapes[0]},
@@ -177,17 +176,15 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
             &inp_tensornds);
 
     // alloc memory
-    DeviceTensorND out =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, out_layout);
+    auto out = Tensor::make(out_layout, cn);
 
-    TensorLayout w_layout({sz}, dtype::Byte());
-    auto dnn_wk = dnn_opr.create_workspace(w_layout);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
 
     // exeucte
     dnn_opr.op->exec(
-            inp_tensornds[0], inp_tensornds[1], empty_bias.as_megdnn(),
-            empty_bias.as_megdnn(), out.as_megdnn(), nullptr, dnn_wk);
-    return {Tensor::make(out)};
+            inp_tensornds[0], inp_tensornds[1], inp_tensornds[2], inp_tensornds[3],
+            out->dnn_tensor(), nullptr, dnn_wk);
+    return {out};
 }
 
 OP_TRAIT_REG(Convolution, Convolution, opr::Convolution)
@@ -368,6 +365,8 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
                 def, inputs[1]->layout().ndim, inputs[0]->layout(), inputs[1]->layout(),
                 cn);
 
+    auto out = Tensor::make(out_layout, cn);
+
     using TensorND = megdnn::TensorND;
     SmallVector<TensorND> inp_tensornds(inputs.size());
     TensorLayoutArray inp_shapes(inputs.size()), oup_shapes(output_descs.size());
@@ -383,16 +382,11 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
             {inp_shapes[0], inp_shapes[1], oup_shapes[0]}, dnn_opr.op.get(), 0, false,
             false, cn, convbwd.policy(), false, &inp_tensornds);
 
-    DeviceTensorND out =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, out_layout);
-
-    auto wk = Blob::make(cn, sz);
-    auto ptr = wk->storage().get();
-    megdnn::Workspace dnn_wk(ptr, sz);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
 
     // exeucte
-    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out.as_megdnn(), dnn_wk);
-    return {Tensor::make(out)};
+    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out->dnn_tensor(), dnn_wk);
+    return {out};
 }
 
 OP_TRAIT_REG(ConvolutionBackwardData, ConvolutionBackwardData)
@@ -549,18 +543,13 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
             false, cn, conv.policy(), false, &inp_tensornds);
 
     // alloc memory
-    DeviceTensorND out =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, out_layout);
+    auto out = Tensor::make(out_layout, cn);
 
-    megdnn::Workspace dnn_wk;
-    if (sz != 0) {
-        TensorLayout w_layout({sz}, dtype::Byte());
-        dnn_wk = dnn_opr.create_workspace(w_layout);
-    }
+    auto dnn_wk = dnn_opr.create_workspace(sz);
 
     // exeucte
-    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out.as_megdnn(), dnn_wk);
-    return {Tensor::make(out)};
+    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out->dnn_tensor(), dnn_wk);
+    return {out};
 }
 
 OP_TRAIT_REG(Convolution3D, Convolution3D, opr::Convolution3D)
@@ -615,8 +604,7 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
         megdnn::Convolution3DBackwardData::deduce_layout_impl(
                 wlayout, dlayout, op_def.param(), oup_layout);
     }
-    DeviceTensorND oup =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, oup_layout);
+    auto oup = Tensor::make(oup_layout, cn);
 
     SmallVector<megdnn::TensorND> inp_tensornds(inputs.size());
     inp_tensornds[0] = inputs[0]->dnn_tensor();
@@ -624,14 +612,10 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     size_t wk_size = setup_algo<megdnn::Convolution3DBackwardData>(
             {wlayout, dlayout, oup_layout}, dnn_opr.get(), 0, false, false, cn,
             op_def.policy(), false, &inp_tensornds);
-    megdnn::Workspace dnn_wk;
-    if (wk_size != 0) {
-        TensorLayout w_layout({wk_size}, dtype::Byte());
-        dnn_wk = caller.create_workspace(w_layout);
-    }
+    auto dnn_wk = caller.create_workspace(wk_size);
 
-    dnn_opr->exec(inp_tensornds[0], inp_tensornds[1], oup.as_megdnn(), dnn_wk);
-    return {Tensor::make(oup)};
+    dnn_opr->exec(inp_tensornds[0], inp_tensornds[1], oup->dnn_tensor(), dnn_wk);
+    return {oup};
 }
 
 auto apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {

imperative/src/impl/ops/elemwise.cpp

@@ -121,10 +121,10 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     megdnn::Elemwise::deduce_shape(inp_shapes, layout);
     layout.init_contiguous_stride();
 
-    DeviceTensorND out =
-            BlobManager::inst()->alloc_workspace_with_defrag(comp_node, layout);
+    auto out = Tensor::make(layout, comp_node);
+
     if (is_empty) {
-        return {Tensor::make(out)};
+        return {out};
     }
     DnnOprCaller<megdnn::Elemwise> dnn_opr(comp_node);
 
@@ -133,12 +133,13 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
         dnn_opr.op->param().mode == Mode::FUSE_MUL_ADD4 ||
         (inp_tensornds.size() &&
          inp_tensornds[0].layout.dtype.category() == DTypeCategory::QUANTIZED)) {
-        opr::Elemwise::perform_dnn(comp_node, out, inp_tensornds, dnn_opr.op);
+        opr::Elemwise::perform_dnn(
+                comp_node, out->dnn_tensor(), inp_tensornds, dnn_opr.op);
     } else {
-        dnn_opr.op->exec(inp_tensornds, out.as_megdnn());
+        dnn_opr.op->exec(inp_tensornds, out->dnn_tensor());
     }
 
-    return {Tensor::make(out)};
+    return {out};
 }
 
 MGB_DEFINE_OPR_CLASS(

imperative/src/impl/ops/indexing.cpp

@@ -85,10 +85,9 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     TensorPtr out = Tensor::make(tlayout, inp->comp_node());
     megdnn::TensorND in = inp->dnn_tensor();
     megdnn::TensorND ind = index->dnn_tensor();
-    TensorLayout m_layout(
-            {dnn_op.op->get_workspace_in_bytes(layout, index_layout, tlayout)},
-            dtype::Byte());
-    auto dnn_workspace = dnn_op.create_workspace(m_layout);
+    size_t sz = dnn_op.op->get_workspace_in_bytes(layout, index_layout, tlayout);
+
+    auto dnn_workspace = dnn_op.create_workspace(sz);
     dnn_op.op->exec(in, ind, out->dnn_tensor(), dnn_workspace);
     return {out};
 }
@@ -152,10 +151,9 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     megdnn::TensorND in = inp->dnn_tensor();
     megdnn::TensorND ind = index->dnn_tensor();
     megdnn::TensorND su = sub->dnn_tensor();
-    TensorLayout m_layout(
-            {dnn_op.op->get_workspace_in_bytes(layout, index_layout, tlayout)},
-            dtype::Byte());
-    auto dnn_workspace = dnn_op.create_workspace(m_layout);
+
+    size_t sz = dnn_op.op->get_workspace_in_bytes(layout, index_layout, tlayout);
+    auto dnn_workspace = dnn_op.create_workspace(sz);
     dnn_op.op->exec(out->dnn_tensor(), ind, su, dnn_workspace);
     return {out};
 }

imperative/src/impl/ops/lamb.cpp

@@ -45,29 +45,25 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     TensorLayout v_t_1_layout{v_t_1->layout()};
     TensorLayout lamb_param_layout{lamb_param->layout()};
 
-    DeviceTensorND m_t = BlobManager::inst()->alloc_workspace_with_defrag(
-            m_t_1->comp_node(), m_t_1_layout);
+    auto m_t = Tensor::make(m_t_1_layout, m_t_1->comp_node());
 
-    DeviceTensorND v_t = BlobManager::inst()->alloc_workspace_with_defrag(
-            v_t_1->comp_node(), v_t_1_layout);
+    auto v_t = Tensor::make(v_t_1_layout, v_t_1->comp_node());
 
-    DeviceTensorND new_param = BlobManager::inst()->alloc_workspace_with_defrag(
-            lamb_param->comp_node(), lamb_param_layout);
+    auto new_param = Tensor::make(lamb_param_layout, lamb_param->comp_node());
 
     DnnOprCaller<megdnn::LAMBUpdate> caller{lamb_param->comp_node()};
-    TensorLayout m_layout(
-            {caller.op->get_workspace_in_bytes(
-                    m_t_1->layout(), v_t_1->layout(), lamb_param->layout(),
-                    grad->layout(), m_t.layout(), v_t.layout(), new_param.layout())},
-            dtype::Byte());
+    size_t sz = caller.op->get_workspace_in_bytes(
+            m_t_1->layout(), v_t_1->layout(), lamb_param->layout(), grad->layout(),
+            m_t->layout(), v_t->layout(), new_param->layout());
 
-    auto dnn_workspace = caller.create_workspace(m_layout);
+    auto dnn_workspace = caller.create_workspace(sz);
     caller.op->param() = op.param();
     caller.op->exec(
             m_t_1->dev_tensor().as_megdnn(), v_t_1->dev_tensor().as_megdnn(),
             lamb_param->dev_tensor().as_megdnn(), grad->dev_tensor().as_megdnn(),
-            m_t.as_megdnn(), v_t.as_megdnn(), new_param.as_megdnn(), dnn_workspace);
-    return {Tensor::make(m_t), Tensor::make(v_t), Tensor::make(new_param)};
+            m_t->dnn_tensor(), v_t->dnn_tensor(), new_param->dnn_tensor(),
+            dnn_workspace);
+    return {m_t, v_t, new_param};
 }
 
 OP_TRAIT_REG(LAMBUpdate, LAMBUpdate)

imperative/src/impl/ops/layer_norm.cpp

@@ -77,32 +77,25 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     megdnn::LayerNorm::deduce_layout_fwd_impl(
             inputs[0]->dnn_tensor().layout, p, oup_layout, mean_layout, rstd_layout);
 
-    DeviceTensorND out_devtensor =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, oup_layout);
-    DeviceTensorND mean_devtensor =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, mean_layout);
-    DeviceTensorND rstd_devtensor =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, rstd_layout);
-
-    megdnn::Workspace dnn_wk;
+    auto out = Tensor::make(oup_layout, cn);
+
+    auto mean = Tensor::make(mean_layout, cn);
+
+    auto rstd = Tensor::make(rstd_layout, cn);
+
     auto wk_size = caller.op->get_workspace_in_bytes(
             inputs[0]->dnn_tensor().layout,
             p.affine ? inputs[1]->dnn_tensor().layout : TensorLayout(),
             p.affine ? inputs[2]->dnn_tensor().layout : TensorLayout(), oup_layout,
             mean_layout, rstd_layout);
-    if (wk_size != 0) {
-        TensorLayout w_layout({wk_size}, dtype::Byte());
-        dnn_wk = caller.create_workspace(w_layout);
-    }
+    auto dnn_wk = caller.create_workspace(wk_size);
 
-    dnn_opr->exec(
+    caller.op->exec(
             inputs[0]->dnn_tensor(),
             p.affine ? inputs[1]->dnn_tensor() : megdnn::TensorND(),
-            p.affine ? inputs[2]->dnn_tensor() : megdnn::TensorND(),
-            out_devtensor.as_megdnn(), mean_devtensor.as_megdnn(),
-            rstd_devtensor.as_megdnn(), dnn_wk);
-    return {Tensor::make(out_devtensor), Tensor::make(mean_devtensor),
-            Tensor::make(rstd_devtensor)};
+            p.affine ? inputs[2]->dnn_tensor() : megdnn::TensorND(), out->dnn_tensor(),
+            mean->dnn_tensor(), rstd->dnn_tensor(), dnn_wk);
+    return {out, mean, rstd};
 }
 
 OP_TRAIT_REG(LayerNorm, LayerNorm)

imperative/src/impl/ops/matmul.cpp

@@ -185,12 +185,12 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     }
 
     if (dim1 == 0 || dim2 == 0 || layout1[layout1.ndim - 1] == 0) {
-        DeviceTensorND out =
-                BlobManager::inst()->alloc_workspace_with_defrag(cn, real_dst_layout);
-        if (!out.empty()) {
-            dev_tensor_memset(out, 0);
+        auto out = Tensor::make(real_dst_layout, cn);
+
+        if (!out->empty()) {
+            dev_tensor_memset(out->dev_tensor(), 0);
         }
-        return {Tensor::make(out)};
+        return {out};
     }
 
     TensorLayout layout_a = layout1, layout_b = layout2;
@@ -232,13 +232,11 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     size_t sz = setup_algo<megdnn::MatrixMul>(
             {layout_a, layout_b, dst_layout}, dnn_opr.op.get(), 0, false, false, cn,
             matmul.policy(), false, &inp_tensornds);
-    DeviceTensorND out =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, dst_layout);
-    TensorLayout w_layout({sz}, dtype::Byte());
-    auto dnn_wk = dnn_opr.create_workspace(w_layout);
+    auto out = Tensor::make(dst_layout, cn);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
 
-    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out.as_megdnn(), dnn_wk);
-    return {Tensor::make(out.sub(SubTensorSpec::make_from_layout(real_dst_layout)))};
+    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out->dnn_tensor(), dnn_wk);
+    return {out->sub(0, real_dst_layout)};
 }
 
 SmallVector<VarNode::LayoutConstraintCallback> get_input_layout_constraint(
@@ -461,12 +459,12 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     dst_layout.init_contiguous_stride();
 
     if (dim1 == 0 || dim2 == 0 || layout1[layout1.ndim - 1] == 0) {
-        DeviceTensorND out =
-                BlobManager::inst()->alloc_workspace_with_defrag(cn, dst_layout);
-        if (!out.empty()) {
-            dev_tensor_memset(out, 0);
+        auto out = Tensor::make(dst_layout, cn);
+
+        if (!out->empty()) {
+            dev_tensor_memset(out->dev_tensor(), 0);
         }
-        return {Tensor::make(out)};
+        return {out};
     }
 
     SmallVector<megdnn::TensorND> inp_tensornds(2u);
@@ -479,19 +477,17 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
             {layout1, layout2, dst_layout}, dnn_opr.op.get(), 0, false, false, cn,
             matmul.policy(), false, &inp_tensornds);
 
-    DeviceTensorND out =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, dst_layout);
+    auto out = Tensor::make(dst_layout, cn);
 
-    TensorLayout w_layout({sz}, dtype::Byte());
-    auto dnn_wk = dnn_opr.create_workspace(w_layout);
-    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out.as_megdnn(), dnn_wk);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
+    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out->dnn_tensor(), dnn_wk);
 
     shp1[shp1.ndim - 2] = dst_layout[dst_layout.ndim - 2];
     shp1[shp1.ndim - 1] = dst_layout[dst_layout.ndim - 1];
     if (maxdim > 3) {
         dst_layout = dst_layout.reshape(shp1);
     }
-    return {Tensor::make(out.sub(SubTensorSpec::make_from_layout(dst_layout)))};
+    return {out->sub(0, dst_layout)};
 }
 
 SmallVector<VarNode::LayoutConstraintCallback> get_input_layout_constraint(
@@ -540,27 +536,23 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     dnn_opr.op->deduce_layout(inp1_tensor.layout, inp2_tensor.layout, oup_layout);
 
     if (inputs[0]->layout().is_empty() || inputs[1]->layout().is_empty()) {
-        DeviceTensorND out =
-                BlobManager::inst()->alloc_workspace_with_defrag(comp_node, oup_layout);
-        if (!out.empty()) {
-            dev_tensor_memset(out, 0);
+        auto out = Tensor::make(oup_layout, comp_node);
+        if (!out->empty()) {
+            dev_tensor_memset(out->dev_tensor(), 0);
         }
-        return {Tensor::make(out)};
+        return {out};
     }
 
     auto sz = dnn_opr.op->get_workspace_in_bytes(
             inp_tensornds[0].layout, inp_tensornds[1].layout, output_descs[0].layout);
 
-    DeviceTensorND out_devtensor =
-            BlobManager::inst()->alloc_workspace_with_defrag(comp_node, oup_layout);
+    auto out = Tensor::make(oup_layout, comp_node);
 
-    TensorLayout w_layout({sz}, dtype::Byte());
-    auto dnn_wk = dnn_opr.create_workspace(w_layout);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
 
-    dnn_opr.op->exec(
-            inp_tensornds[0], inp_tensornds[1], out_devtensor.as_megdnn(), dnn_wk);
+    dnn_opr.op->exec(inp_tensornds[0], inp_tensornds[1], out->dnn_tensor(), dnn_wk);
 
-    return {Tensor::make(out_devtensor)};
+    return {out};
 }
 
 std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(

imperative/src/impl/ops/misc.cpp

@@ -36,9 +36,8 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     megdnn::CheckNonFinite::Param param({op.scale});
     dnn_opr.op->param() = param;
     size_t sz = dnn_opr.op->get_workspace_in_bytes(srcs, dest->layout());
-    TensorLayout w_layout({sz}, dtype::Byte());
-    auto dnn_wk = dnn_opr.create_workspace(w_layout);
-    dnn_opr.op->exec(srcs, dest->dev_tensor().as_megdnn(), dnn_wk);
+    auto dnn_wk = dnn_opr.create_workspace(sz);
+    dnn_opr.op->exec(srcs, dest->dnn_tensor(), dnn_wk);
     return outputs;
 }
 

imperative/src/impl/ops/pooling.cpp

@@ -66,17 +66,12 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
             {inp_tensornds[0].layout, oup_layout}, dnn_opr.get(), 0, false, false, cn,
             op_def.policy(), false, &inp_tensornds);
 
-    DeviceTensorND out_devtensor =
-            BlobManager::inst()->alloc_workspace_with_defrag(cn, oup_layout);
+    auto out = Tensor::make(oup_layout, cn);
 
-    megdnn::Workspace dnn_wk;
-    if (wk_size) {
-        TensorLayout w_layout({wk_size}, dtype::Byte());
-        dnn_wk = caller.create_workspace(w_layout);
-    }
+    auto dnn_wk = caller.create_workspace(wk_size);
 
-    dnn_opr->exec(inp_tensornds[0], out_devtensor.as_megdnn(), dnn_wk);
-    return {Tensor::make(out_devtensor)};
+    caller.op->exec(inp_tensornds[0], out->dnn_tensor(), dnn_wk);
+    return {out};
 }
 
 OP_TRAIT_REG(Pooling, Pooling)

imperative/src/impl/ops/reduce.cpp

@@ -117,20 +117,20 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
             layout.remove_axis_inplace(axis);
             layout.init_contiguous_stride();
         }
-        DeviceTensorND out =
-                BlobManager::inst()->alloc_workspace_with_defrag(comp_node, layout);
+        auto out = Tensor::make(layout, comp_node);
+
         std::string err_msg;
         switch (mode) {
             case Reduce::Mode::SUM:
-                if (!out.empty()) {
-                    dev_tensor_memset(out, 0);
+                if (!out->empty()) {
+                    dev_tensor_memset(out->dev_tensor(), 0);
                 }
                 break;
             case Reduce::Mode::PRODUCT:
-                if (!out.empty()) {
+                if (!out->empty()) {
                     DnnOprCaller<megdnn::Fill> fill_op(comp_node);
                     fill_op.op->param() = 1;
-                    fill_op.op->exec(out.as_megdnn(), {});
+                    fill_op.op->exec(out->dnn_tensor(), {});
                 }
                 break;
             case Reduce::Mode::MEAN:
@@ -153,34 +153,29 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
                     MegBrainError, "empty input is not allowed for reduce mode: %s",
                     err_msg.c_str());
         }
-        return {Tensor::make(out)};
+        return {out};
     }
 
     auto dnn_ten = inputs[0]->dnn_tensor();
     dnn_ten.layout = src;
     inp_tensornds.push_back(dnn_ten);
 
-    megdnn::Workspace dnn_wk;
-
     auto wk_size = dnn_op.op->get_workspace_in_bytes(src, layout);
-    if (wk_size) {
-        TensorLayout w_layout({wk_size}, dtype::Byte());
-        dnn_wk = dnn_op.create_workspace(w_layout);
-    }
-
-    DeviceTensorND out =
-            BlobManager::inst()->alloc_workspace_with_defrag(comp_node, layout);
-
-    dnn_op.op->exec(inp_tensornds[0], out.as_megdnn(), dnn_wk);
+    auto dnn_wk = dnn_op.create_workspace(wk_size);
+    TensorLayout ori_layout = layout;
 
     if (!keepdim && src.ndim > 1) {
-        auto out_layout = out.layout();
-        out_layout.remove_axis_inplace(axis);
-        out_layout.init_contiguous_stride();
-        out.resize(out_layout);
+        layout.remove_axis_inplace(axis);
+        layout.init_contiguous_stride();
     }
 
-    return {Tensor::make(out)};
+    auto out = Tensor::make(layout, comp_node);
+    auto dnn_out = out->dnn_tensor();
+    dnn_out.layout = ori_layout;
+
+    dnn_op.op->exec(inp_tensornds[0], dnn_out, dnn_wk);
+
+    return {out};
 }
 
 std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(

imperative/src/impl/ops/tensor_manip.cpp

@@ -252,9 +252,8 @@ SmallVector<TensorPtr> param_pack_concat_apply_on_physical_tensor(
     HostTensorStorage srcs_storage;
     srcs_storage.reset(comp_node, srcs_size, srcs_ptr);
     caller.op->exec(
-            {srcs_raw_ptr, srcs_layout}, inputs.back()->dev_tensor().as_megdnn(),
-            output->dev_tensor().as_megdnn(),
-            caller.create_workspace({{ws_size}, dtype::Byte()}));
+            {srcs_raw_ptr, srcs_layout}, inputs.back()->dnn_tensor(),
+            output->dnn_tensor(), caller.create_workspace(ws_size));
     async_release(HostTensorND{comp_node, srcs_layout}.storage(srcs_storage));
     return {output};
 }

imperative/src/impl/ops/vision.cpp

@@ -89,8 +89,8 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
 
     size_t sz = dnn_opr.op->get_workspace_in_bytes(
             inputs[0]->layout(), inputs[1]->layout(), out_layout, ind_layout);
-    TensorLayout w_layout({sz}, dtype::Byte());
-    auto dnn_wk = dnn_opr.create_workspace(w_layout);
+
+    auto dnn_wk = dnn_opr.create_workspace(sz);
 
     dnn_opr.op->exec(
             inputs[0]->dnn_tensor(), inputs[1]->dnn_tensor(), out.as_megdnn(),

imperative/src/impl/physical_tensor.cpp

@@ -566,9 +566,13 @@ DeviceTensorND Tensor::dev_tensor(bool contiguous) {
     return ret;
 }
 
+bool Tensor::empty() {
+    return !m_blob->size();
+}
+
 megdnn::TensorND Tensor::dnn_tensor() {
     mgb_assert(m_blob, "uninitialized tensor.");
-    return {m_layout, {m_blob->storage().get(), m_offset}};
+    return DnnTensorND{m_layout, m_blob->storage(), m_offset};
 }
 
 void Tensor::fetch_value() {

imperative/src/include/megbrain/imperative/physical_tensor.h

@@ -10,6 +10,7 @@
 #include "megbrain/imperative/resource_manager.h"
 #include "megbrain/tensor.h"
 #include "megbrain/utils/metahelper.h"
+#include "megdnn/basic_types.h"
 
 namespace mgb {
 namespace imperative {
@@ -87,6 +88,22 @@ using EventPtr = std::unique_ptr<CompNode::Event, EventDeleter>;
 
 class Tensor;
 using TensorPtr = std::shared_ptr<Tensor>;
+
+/*
+    using DnnTensorND to save the reference count of workspace
+    allocted by blobmanager to prevent invalidation
+*/
+struct DnnTensorND : megdnn::TensorND {
+private:
+    std::shared_ptr<dt_byte> m_reference;
+
+public:
+    DnnTensorND(TensorLayout& layout_, std::shared_ptr<dt_byte> ref_ptr, size_t offset)
+            : megdnn::TensorND(layout_, {ref_ptr.get(), offset}) {
+        m_reference = ref_ptr;
+    }
+};
+
 class Tensor : public NonCopyableObj {
 public:
     Tensor() = default;
@@ -131,6 +148,8 @@ public:
 
     void to_contiguous_inplace();
 
+    bool empty();
+
     DeviceTensorND dev_tensor(bool contiguous = true);
 
     void assign_from_dev_tensor(DeviceTensorND);

src/opr/impl/basic_arith.cpp

@@ -258,9 +258,9 @@ void Elemwise::perform(
 }
 
 void Elemwise::perform_dnn(
-        CompNode cn, DeviceTensorND& dest, megdnn::TensorNDArray& inputs,
+        CompNode cn, const megdnn::TensorND& dest, megdnn::TensorNDArray& inputs,
         intl::UniqPtrWithCN<megdnn::Elemwise>& opr) {
-    call_megdnn_opr_exec(cn, inputs, dest.as_megdnn(), opr.get(), nullptr);
+    call_megdnn_opr_exec(cn, inputs, dest, opr.get(), nullptr);
 }
 
 TensorLayoutArray Elemwise::collective_collapse(const TensorLayoutArray& layouts) {

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

@@ -78,7 +78,7 @@ public:
             intl::UniqPtrWithCN<megdnn::Elemwise>& opr);
 
     MGE_WIN_DECLSPEC_FUC static void perform_dnn(
-            CompNode cn, DeviceTensorND& dest, megdnn::TensorNDArray& inputs,
+            CompNode cn, const megdnn::TensorND& dest, megdnn::TensorNDArray& inputs,
             intl::UniqPtrWithCN<megdnn::Elemwise>& opr);
 
     using TensorLayoutPtrArray = SmallVector<TensorLayout*>;