Blender  V3.3
work_balancer.cpp
Go to the documentation of this file.
1 /* SPDX-License-Identifier: Apache-2.0
2  * Copyright 2011-2022 Blender Foundation */
3 
5 
6 #include "util/math.h"
7 
8 #include "util/log.h"
9 
11 
13 {
14  const int num_infos = work_balance_infos.size();
15 
16  if (num_infos == 1) {
17  work_balance_infos[0].weight = 1.0;
18  return;
19  }
20  else if (num_infos == 0) {
21  return;
22  }
23 
24  /* There is no statistics available, so start with an equal distribution. */
25  const double weight = 1.0 / num_infos;
26  for (WorkBalanceInfo &balance_info : work_balance_infos) {
27  balance_info.weight = weight;
28  }
29 }
30 
31 static double calculate_total_time(const vector<WorkBalanceInfo> &work_balance_infos)
32 {
33  double total_time = 0;
34  for (const WorkBalanceInfo &info : work_balance_infos) {
35  total_time += info.time_spent;
36  }
37  return total_time;
38 }
39 
40 /* The balance is based on equalizing time which devices spent performing a task. Assume that
41  * average of the observed times is usable for estimating whether more or less work is to be
42  * scheduled, and how difference in the work scheduling is needed. */
43 
45 {
46  const int num_infos = work_balance_infos.size();
47 
48  const double total_time = calculate_total_time(work_balance_infos);
49  const double time_average = total_time / num_infos;
50 
51  double total_weight = 0;
52  vector<double> new_weights;
53  new_weights.reserve(num_infos);
54 
55  /* Equalize the overall average time. This means that we don't make it so every work will perform
56  * amount of work based on the current average, but that after the weights changes the time will
57  * equalize.
58  * Can think of it that if one of the devices is 10% faster than another, then one device needs
59  * to do 5% less of the current work, and another needs to do 5% more. */
60  const double lerp_weight = 1.0 / num_infos;
61 
62  bool has_big_difference = false;
63 
64  for (const WorkBalanceInfo &info : work_balance_infos) {
65  const double time_target = lerp(info.time_spent, time_average, lerp_weight);
66  const double new_weight = info.weight * time_target / info.time_spent;
67  new_weights.push_back(new_weight);
68  total_weight += new_weight;
69 
70  if (std::fabs(1.0 - time_target / time_average) > 0.02) {
71  has_big_difference = true;
72  }
73  }
74 
75  if (!has_big_difference) {
76  return false;
77  }
78 
79  const double total_weight_inv = 1.0 / total_weight;
80  for (int i = 0; i < num_infos; ++i) {
81  WorkBalanceInfo &info = work_balance_infos[i];
82  info.weight = new_weights[i] * total_weight_inv;
83  info.time_spent = 0;
84  }
85 
86  return true;
87 }
88 
#define CCL_NAMESPACE_END
Definition: cuda/compat.h:9
ccl_device_inline float2 fabs(const float2 &a)
Definition: math_float2.h:222
static float lerp(float t, float a, float b)
CCL_NAMESPACE_BEGIN void work_balance_do_initial(vector< WorkBalanceInfo > &work_balance_infos)
bool work_balance_do_rebalance(vector< WorkBalanceInfo > &work_balance_infos)
static double calculate_total_time(const vector< WorkBalanceInfo > &work_balance_infos)