feat(dtr): add sqrt sampling

GitOrigin-RevId: 8cb2ceb520

imperative/python/megengine/dtr/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import mprop
+
+from .dtr import *
+
+mprop.init()

imperative/python/megengine/dtr.py → imperative/python/megengine/dtr/dtr.py

@@ -9,13 +9,12 @@
 import re
 from typing import Union
 
-from mprop import mproperty
-
-from .core._imperative_rt.core2 import set_option as _set_option
-from .core._imperative_rt.utils import _set_defrag
+from ..core._imperative_rt.core2 import set_option as _set_option
+from ..core._imperative_rt.utils import _set_defrag
 
 _eviction_threshold = 0
 _evictee_minimum_size = 1024 ** 2
+_enable_sqrt_sampling = False
 
 
 def _str2bytes(text: str) -> int:
@@ -29,7 +28,7 @@ def _str2bytes(text: str) -> int:
     return int(float(result[0][0]) * 1024 ** order.index(result[0][1].lower()))
 
 
-@mproperty
+@property
 def eviction_threshold(mod):
     r"""
     Get or set the eviction threshold in bytes. It can also be set to a string,
@@ -50,21 +49,22 @@ def eviction_threshold(mod):
         mge.dtr.eviction_threshold = "2GB"
 
     """
-    return mod._eviction_threshold
+    return _eviction_threshold
 
 
 @eviction_threshold.setter
 def eviction_threshold(mod, value: Union[int, str]):
+    global _eviction_threshold
     if isinstance(value, str):
-        mod._eviction_threshold = mod._str2bytes(value)
+        _eviction_threshold = _str2bytes(value)
     elif isinstance(value, int):
-        mod._eviction_threshold = value
+        _eviction_threshold = value
     else:
         raise TypeError("`value` should be a str or an int")
-    _set_option("dtr_eviction_threshold", mod._eviction_threshold)
+    _set_option("dtr_eviction_threshold", _eviction_threshold)
 
 
-@mproperty
+@property
 def evictee_minimum_size(mod):
     r"""
     Get or set the memory threshold of tensors in bytes. It can also be set to a
@@ -85,18 +85,45 @@ def evictee_minimum_size(mod):
         mge.dtr.evictee_minimum_size = "2MB"
 
     """
-    return mod._evictee_minimum_size
+    return _evictee_minimum_size
 
 
 @evictee_minimum_size.setter
 def evictee_minimum_size(mod, value: Union[int, str]):
+    global _evictee_minimum_size
     if isinstance(value, str):
-        mod._evictee_minimum_size = mod._str2bytes(value)
+        _evictee_minimum_size = _str2bytes(value)
     elif isinstance(value, int):
-        mod._evictee_minimum_size = value
+        _evictee_minimum_size = value
     else:
         raise TypeError("`value` should be a str or an int")
-    _set_option("dtr_evictee_minimum_size", mod._evictee_minimum_size)
+    _set_option("dtr_evictee_minimum_size", _evictee_minimum_size)
+
+
+@property
+def enable_sqrt_sampling(mod):
+    r"""
+    Get or set whether sqrt sampling is allowed. Sqrt sampling means that given
+    the size of the candidate set is N, only enumerate sqrt(N) tensors. When
+    the number of tensors is very high, enabling this optimization will speed
+    up the training.
+
+    Examples:
+
+    .. code-block::
+
+        import megengine as mge
+        mge.dtr.enable_sqrt_sampling = True
+
+    """
+    return _enable_sqrt_sampling
+
+
+@enable_sqrt_sampling.setter
+def enable_sqrt_sampling(mod, value: bool):
+    global _enable_sqrt_sampling
+    _enable_sqrt_sampling = value
+    _set_option("enable_dtr_sqrt_sampling", _enable_sqrt_sampling)
 
 
 def enable():

imperative/src/impl/interpreter/interpreter_impl.cpp

@@ -761,7 +761,7 @@ bool ChannelImpl::auto_evict(size_t force_num=0) {
     while ((state.options.dtr_eviction_threshold > 0 && current_memory > state.options.dtr_eviction_threshold) || force_num > 0) {
         RECORD_EVENT(AutoEvictEvent);
         sample_on_device(m_dtr.comp_node, false);
-        auto best = m_dtr.find_best_tensor();
+        auto best = m_dtr.find_best_tensor(state.options.enable_dtr_sqrt_sampling && !force_num);
         if (!best) {
             break;
         }
@@ -988,8 +988,15 @@ void ChannelImpl::process_one_task(IdentifiedCommand& icmd) {
                     if (!inplace && !cross_cn && !m_dtr.is_bad_op(get_name(*cmd.op))) {
                         TensorInfo::ComputePath::make(cmd.id, cmd.op, cmd.inputs, cmd.outputs);
                         size_t detach_cnt = 0;
+                        if (!strcmp(get_name(*cmd.op), "BatchNorm") && cmd.outputs.size() == 5) {
+                            cmd.outputs[0]->detach_producer(); // detach running_mean
+                            cmd.outputs[1]->detach_producer(); // detach running_var
+                            for (auto input : cmd.inputs) {
+                                input->ref_cnt -= 2;
+                            }
+                        }
                         for (auto output : cmd.outputs) {
-                            if (!output->size_exceeds_thd(state.options.dtr_evictee_minimum_size)) {
+                            if (output->producer && !output->size_exceeds_thd(state.options.dtr_evictee_minimum_size)) {
                                 output->detach_producer();
                                 detach_cnt ++;
                             }
@@ -1339,9 +1346,15 @@ double ChannelImpl::DynamicSublinear::estimate_neighbor_cost(TensorInfo* ptr) {
     return cost;
 }
 
-TensorInfo* ChannelImpl::DynamicSublinear::find_best_tensor() {
+TensorInfo* ChannelImpl::DynamicSublinear::find_best_tensor(bool enable_dtr_sqrt_sampling=false) {
     double min_msps = -1;
     TensorInfo* best = nullptr;
+    size_t sz = 1;
+    if (enable_dtr_sqrt_sampling) {
+        while (sz * sz <= candidates.size()) sz ++;
+    } else {
+        sz = candidates.size();
+    }
     for (auto i : candidates) {
         if (i->producer && i->ptr && !i->pinned && i->evict_type == EvictType::NONE) {
             double neighbor_cost = estimate_neighbor_cost(i);
@@ -1354,6 +1367,7 @@ TensorInfo* ChannelImpl::DynamicSublinear::find_best_tensor() {
                 best = i;
             }
         }
+        if (--sz == 0) break;
     }
     return best;
 }

imperative/src/impl/interpreter/interpreter_impl.h

@@ -323,7 +323,7 @@ private:
          * \return the pointer of the best tensor; nullptr is returned if no 
          * available tensor is found
          */
-        TensorInfo* find_best_tensor();
+        TensorInfo* find_best_tensor(bool);
 
         /*!
          * \brief estimate the cost of recomputing tensor ptr

imperative/src/impl/interpreter/option_manager.h

@@ -41,6 +41,7 @@ public:
     DEF_OPTION(enable_host_compute,     "MEGENGINE_HOST_COMPUTE",           1,
         "enable host compute, thus computation may be done in host event if it's device is gpu.");
     DEF_OPTION(enable_dtr_auto_drop,    "MEGENGINE_DTR_AUTO_DROP",          0, "");
+    DEF_OPTION(enable_dtr_sqrt_sampling, "MEGENGINE_DTR_SQRT_SAMPLING",     0, "");
     DEF_OPTION(dtr_eviction_threshold,  "MEGENGINE_DTR_EVICTION_THRESHOLD", 0,
         "auto drop will start whenever gpu memory usage exceeds this value.");
     DEF_OPTION(dtr_evictee_minimum_size, "MEGENGINE_DTR_EVICTEE_MINIMUM_SIZE", 1048576,