• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017 Lima Project
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sub license,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the
12  * next paragraph) shall be included in all copies or substantial portions
13  * of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21  * DEALINGS IN THE SOFTWARE.
22  *
23  */
24 
25 #include <limits.h>
26 
27 #include "gpir.h"
28 
29 /* Register sensitive schedule algorithm from paper:
30  * "Register-Sensitive Selection, Duplication, and Sequencing of Instructions"
31  * Author: Vivek Sarkar,  Mauricio J. Serrano,  Barbara B. Simons
32  */
33 
schedule_calc_sched_info(gpir_node * node)34 static void schedule_calc_sched_info(gpir_node *node)
35 {
36    int n = 0;
37    float extra_reg = 1.0f;
38 
39    /* update all children's sched info */
40    gpir_node_foreach_pred(node, dep) {
41       gpir_node *pred = dep->pred;
42 
43       if (pred->rsched.reg_pressure < 0)
44          schedule_calc_sched_info(pred);
45 
46       int est = pred->rsched.est + 1;
47       if (node->rsched.est < est)
48          node->rsched.est = est;
49 
50       float reg_weight = 1.0f - 1.0f / list_length(&pred->succ_list);
51       if (extra_reg > reg_weight)
52          extra_reg = reg_weight;
53 
54       n++;
55    }
56 
57    /* leaf instr */
58    if (!n) {
59       node->rsched.reg_pressure = 0;
60       return;
61    }
62 
63    int i = 0;
64    float reg[n];
65    gpir_node_foreach_pred(node, dep) {
66       gpir_node *pred = dep->pred;
67       reg[i++] = pred->rsched.reg_pressure;
68    }
69 
70    /* sort */
71    for (i = 0; i < n - 1; i++) {
72       for (int j = 0; j < n - i - 1; j++) {
73          if (reg[j] > reg[j + 1]) {
74             float tmp = reg[j + 1];
75             reg[j + 1] = reg[j];
76             reg[j] = tmp;
77          }
78       }
79    }
80 
81    for (i = 0; i < n; i++) {
82       float pressure = reg[i] + n - (i + 1);
83       if (pressure > node->rsched.reg_pressure)
84          node->rsched.reg_pressure = pressure;
85    }
86 
87    /* If all children of this node have multi parents, then this
88     * node need an extra reg to store its result. For example,
89     * it's not fair for parent has the same reg pressure as child
90     * if n==1 and child's successor>1, because we need 2 reg for
91     * this.
92     *
93     * But we can't add a full reg to the reg_pressure, because the
94     * last parent of a multi-successor child doesn't need an extra
95     * reg. For example, a single child (with multi successor) node
96     * should has less reg pressure than a two children (with single
97     * successor) instr.
98     *
99     * extra reg = min(all child)(1.0 - 1.0 / num successor)
100     */
101    node->rsched.reg_pressure += extra_reg;
102 }
103 
schedule_insert_ready_list(struct list_head * ready_list,gpir_node * insert_node)104 static void schedule_insert_ready_list(struct list_head *ready_list,
105                                        gpir_node *insert_node)
106 {
107    struct list_head *insert_pos = ready_list;
108 
109    list_for_each_entry(gpir_node, node, ready_list, list) {
110       if (gpir_op_infos[node->op].schedule_first) {
111          continue;
112       }
113 
114       if (gpir_op_infos[insert_node->op].schedule_first ||
115           insert_node->rsched.parent_index < node->rsched.parent_index ||
116           (insert_node->rsched.parent_index == node->rsched.parent_index &&
117            (insert_node->rsched.reg_pressure < node->rsched.reg_pressure ||
118             (insert_node->rsched.reg_pressure == node->rsched.reg_pressure &&
119              (insert_node->rsched.est >= node->rsched.est))))) {
120          insert_pos = &node->list;
121          if (node == insert_node)
122             return;
123          break;
124       }
125    }
126 
127    list_del(&insert_node->list);
128    list_addtail(&insert_node->list, insert_pos);
129 }
130 
schedule_ready_list(gpir_block * block,struct list_head * ready_list)131 static void schedule_ready_list(gpir_block *block, struct list_head *ready_list)
132 {
133    if (list_is_empty(ready_list))
134       return;
135 
136    gpir_node *node = list_first_entry(ready_list, gpir_node, list);
137    list_del(&node->list);
138 
139    /* schedule the node to the block node list */
140    list_add(&node->list, &block->node_list);
141    node->rsched.scheduled = true;
142    block->rsched.node_index--;
143 
144    gpir_node_foreach_pred(node, dep) {
145       gpir_node *pred = dep->pred;
146       pred->rsched.parent_index = block->rsched.node_index;
147 
148       bool ready = true;
149       gpir_node_foreach_succ(pred, dep) {
150          gpir_node *succ = dep->succ;
151          if (!succ->rsched.scheduled) {
152             ready = false;
153             break;
154          }
155       }
156       /* all successor have been scheduled */
157       if (ready)
158          schedule_insert_ready_list(ready_list, pred);
159    }
160 
161    schedule_ready_list(block, ready_list);
162 }
163 
schedule_block(gpir_block * block)164 static void schedule_block(gpir_block *block)
165 {
166    /* move all nodes to node_list, block->node_list will
167     * contain schedule result */
168    struct list_head node_list;
169    list_replace(&block->node_list, &node_list);
170    list_inithead(&block->node_list);
171 
172    /* step 2 & 3 */
173    list_for_each_entry(gpir_node, node, &node_list, list) {
174       if (gpir_node_is_root(node))
175          schedule_calc_sched_info(node);
176       block->rsched.node_index++;
177    }
178 
179    /* step 4 */
180    struct list_head ready_list;
181    list_inithead(&ready_list);
182 
183    /* step 5 */
184    list_for_each_entry_safe(gpir_node, node, &node_list, list) {
185       if (gpir_node_is_root(node)) {
186          node->rsched.parent_index = INT_MAX;
187          schedule_insert_ready_list(&ready_list, node);
188       }
189    }
190 
191    /* step 6 */
192    schedule_ready_list(block, &ready_list);
193 }
194 
195 /* Due to how we translate from NIR, we never read a register written in the
196  * same block (we just pass the node through instead), so we don't have to
197  * worry about read-after-write dependencies. We do have to worry about
198  * write-after-read though, so we add those dependencies now. For example in a
199  * loop like this we need a dependency between the write and the read of i:
200  *
201  * i = ...
202  * while (...) {
203  *    ... = i;
204  *    i = i + 1;
205  * }
206  */
207 
add_false_dependencies(gpir_compiler * comp)208 static void add_false_dependencies(gpir_compiler *comp)
209 {
210    /* Make sure we allocate this only once, in case there are many values and
211     * many blocks.
212     */
213    gpir_node **last_written = calloc(comp->cur_reg, sizeof(gpir_node *));
214 
215    list_for_each_entry(gpir_block, block, &comp->block_list, list) {
216       list_for_each_entry_rev(gpir_node, node, &block->node_list, list) {
217          if (node->op == gpir_op_load_reg) {
218             gpir_load_node *load = gpir_node_to_load(node);
219             gpir_node *store = last_written[load->reg->index];
220             if (store && store->block == block) {
221                gpir_node_add_dep(store, node, GPIR_DEP_WRITE_AFTER_READ);
222             }
223          } else if (node->op == gpir_op_store_reg) {
224             gpir_store_node *store = gpir_node_to_store(node);
225             last_written[store->reg->index] = node;
226          }
227       }
228    }
229 
230    free(last_written);
231 }
232 
gpir_reduce_reg_pressure_schedule_prog(gpir_compiler * comp)233 bool gpir_reduce_reg_pressure_schedule_prog(gpir_compiler *comp)
234 {
235    add_false_dependencies(comp);
236 
237    list_for_each_entry(gpir_block, block, &comp->block_list, list) {
238       block->rsched.node_index = 0;
239       list_for_each_entry_safe(gpir_node, node, &block->node_list, list) {
240          node->rsched.reg_pressure = -1;
241          node->rsched.est = 0;
242          node->rsched.scheduled = false;
243       }
244    }
245 
246    list_for_each_entry(gpir_block, block, &comp->block_list, list) {
247       schedule_block(block);
248    }
249 
250    gpir_debug("after reduce scheduler\n");
251    gpir_node_print_prog_seq(comp);
252    return true;
253 }
254