e3488b3ac757b05c984b7b62dcaa7f3fcacd7ded
[mesa.git] / src / gallium / drivers / radeon / r600_gpu_load.c
1 /*
2 * Copyright 2015 Advanced Micro Devices, Inc.
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, sublicense,
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 next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * 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 NONINFRINGEMENT. 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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors: Marek Olšák <maraeo@gmail.com>
24 *
25 */
26
27 /* The GPU load is measured as follows.
28 *
29 * There is a thread which samples the GRBM_STATUS register at a certain
30 * frequency and the "busy" or "idle" counter is incremented based on
31 * whether the GUI_ACTIVE bit is set or not.
32 *
33 * Then, the user can sample the counters twice and calculate the average
34 * GPU load between the two samples.
35 */
36
37 #include "r600_pipe_common.h"
38 #include "os/os_time.h"
39
40 /* For good accuracy at 1000 fps or lower. This will be inaccurate for higher
41 * fps (there are too few samples per frame). */
42 #define SAMPLES_PER_SEC 10000
43
44 #define GRBM_STATUS 0x8010
45 #define SPI_BUSY(x) (((x) >> 22) & 0x1)
46 #define GUI_ACTIVE(x) (((x) >> 31) & 0x1)
47
48 static void r600_update_grbm_counters(struct r600_common_screen *rscreen,
49 union r600_grbm_counters *counters)
50 {
51 uint32_t value = 0;
52
53 rscreen->ws->read_registers(rscreen->ws, GRBM_STATUS, 1, &value);
54
55 if (SPI_BUSY(value))
56 p_atomic_inc(&counters->named.spi_busy);
57 else
58 p_atomic_inc(&counters->named.spi_idle);
59
60 if (GUI_ACTIVE(value))
61 p_atomic_inc(&counters->named.gui_busy);
62 else
63 p_atomic_inc(&counters->named.gui_idle);
64 }
65
66 static PIPE_THREAD_ROUTINE(r600_gpu_load_thread, param)
67 {
68 struct r600_common_screen *rscreen = (struct r600_common_screen*)param;
69 const int period_us = 1000000 / SAMPLES_PER_SEC;
70 int sleep_us = period_us;
71 int64_t cur_time, last_time = os_time_get();
72
73 while (!p_atomic_read(&rscreen->gpu_load_stop_thread)) {
74 if (sleep_us)
75 os_time_sleep(sleep_us);
76
77 /* Make sure we sleep the ideal amount of time to match
78 * the expected frequency. */
79 cur_time = os_time_get();
80
81 if (os_time_timeout(last_time, last_time + period_us,
82 cur_time))
83 sleep_us = MAX2(sleep_us - 1, 1);
84 else
85 sleep_us += 1;
86
87 /*printf("Hz: %.1f\n", 1000000.0 / (cur_time - last_time));*/
88 last_time = cur_time;
89
90 /* Update the counters. */
91 r600_update_grbm_counters(rscreen, &rscreen->grbm_counters);
92 }
93 p_atomic_dec(&rscreen->gpu_load_stop_thread);
94 return 0;
95 }
96
97 void r600_gpu_load_kill_thread(struct r600_common_screen *rscreen)
98 {
99 if (!rscreen->gpu_load_thread)
100 return;
101
102 p_atomic_inc(&rscreen->gpu_load_stop_thread);
103 pipe_thread_wait(rscreen->gpu_load_thread);
104 rscreen->gpu_load_thread = 0;
105 }
106
107 static uint64_t r600_read_counter(struct r600_common_screen *rscreen,
108 unsigned busy_index)
109 {
110 /* Start the thread if needed. */
111 if (!rscreen->gpu_load_thread) {
112 pipe_mutex_lock(rscreen->gpu_load_mutex);
113 /* Check again inside the mutex. */
114 if (!rscreen->gpu_load_thread)
115 rscreen->gpu_load_thread =
116 pipe_thread_create(r600_gpu_load_thread, rscreen);
117 pipe_mutex_unlock(rscreen->gpu_load_mutex);
118 }
119
120 unsigned busy = p_atomic_read(&rscreen->grbm_counters.array[busy_index]);
121 unsigned idle = p_atomic_read(&rscreen->grbm_counters.array[busy_index + 1]);
122
123 return busy | ((uint64_t)idle << 32);
124 }
125
126 static unsigned r600_end_counter(struct r600_common_screen *rscreen,
127 uint64_t begin, unsigned busy_index)
128 {
129 uint64_t end = r600_read_counter(rscreen, busy_index);
130 unsigned busy = (end & 0xffffffff) - (begin & 0xffffffff);
131 unsigned idle = (end >> 32) - (begin >> 32);
132
133 /* Calculate the % of time the busy counter was being incremented.
134 *
135 * If no counters were incremented, return the current counter status.
136 * It's for the case when the load is queried faster than
137 * the counters are updated.
138 */
139 if (idle || busy) {
140 return busy*100 / (busy + idle);
141 } else {
142 union r600_grbm_counters counters;
143
144 memset(&counters, 0, sizeof(counters));
145 r600_update_grbm_counters(rscreen, &counters);
146 return counters.array[busy_index] ? 100 : 0;
147 }
148 }
149
150 #define BUSY_INDEX(rscreen, field) (&rscreen->grbm_counters.named.field##_busy - \
151 rscreen->grbm_counters.array)
152
153 uint64_t r600_begin_counter_spi(struct r600_common_screen *rscreen)
154 {
155 return r600_read_counter(rscreen, BUSY_INDEX(rscreen, spi));
156 }
157
158 unsigned r600_end_counter_spi(struct r600_common_screen *rscreen, uint64_t begin)
159 {
160 return r600_end_counter(rscreen, begin, BUSY_INDEX(rscreen, spi));
161 }
162
163 uint64_t r600_begin_counter_gui(struct r600_common_screen *rscreen)
164 {
165 return r600_read_counter(rscreen, BUSY_INDEX(rscreen, gui));
166 }
167
168 unsigned r600_end_counter_gui(struct r600_common_screen *rscreen, uint64_t begin)
169 {
170 return r600_end_counter(rscreen, begin, BUSY_INDEX(rscreen, gui));
171 }