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 "r600_query.h"
39#include "util/os_time.h"
40
41/* For good accuracy at 1000 fps or lower. This will be inaccurate for higher
42 * fps (there are too few samples per frame). */
43#define SAMPLES_PER_SEC 10000
44
45#define GRBM_STATUS		0x8010
46#define TA_BUSY(x)		(((x) >> 14) & 0x1)
47#define GDS_BUSY(x)		(((x) >> 15) & 0x1)
48#define VGT_BUSY(x)		(((x) >> 17) & 0x1)
49#define IA_BUSY(x)		(((x) >> 19) & 0x1)
50#define SX_BUSY(x)		(((x) >> 20) & 0x1)
51#define WD_BUSY(x)		(((x) >> 21) & 0x1)
52#define SPI_BUSY(x)		(((x) >> 22) & 0x1)
53#define BCI_BUSY(x)		(((x) >> 23) & 0x1)
54#define SC_BUSY(x)		(((x) >> 24) & 0x1)
55#define PA_BUSY(x)		(((x) >> 25) & 0x1)
56#define DB_BUSY(x)		(((x) >> 26) & 0x1)
57#define CP_BUSY(x)		(((x) >> 29) & 0x1)
58#define CB_BUSY(x)		(((x) >> 30) & 0x1)
59#define GUI_ACTIVE(x)		(((x) >> 31) & 0x1)
60
61#define SRBM_STATUS2		0x0e4c
62#define SDMA_BUSY(x)		(((x) >> 5) & 0x1)
63
64#define CP_STAT                 0x8680
65#define PFP_BUSY(x)		(((x) >> 15) & 0x1)
66#define MEQ_BUSY(x)		(((x) >> 16) & 0x1)
67#define ME_BUSY(x)		(((x) >> 17) & 0x1)
68#define SURFACE_SYNC_BUSY(x)	(((x) >> 21) & 0x1)
69#define DMA_BUSY(x)		(((x) >> 22) & 0x1)
70#define SCRATCH_RAM_BUSY(x)	(((x) >> 24) & 0x1)
71
72#define IDENTITY(x) x
73
74#define UPDATE_COUNTER(field, mask)					\
75	do {								\
76		if (mask(value))					\
77			p_atomic_inc(&counters->named.field.busy);	\
78		else							\
79			p_atomic_inc(&counters->named.field.idle);	\
80	} while (0)
81
82static void r600_update_mmio_counters(struct r600_common_screen *rscreen,
83				      union r600_mmio_counters *counters)
84{
85	uint32_t value = 0;
86	bool gui_busy, sdma_busy = false;
87
88	/* GRBM_STATUS */
89	rscreen->ws->read_registers(rscreen->ws, GRBM_STATUS, 1, &value);
90
91	UPDATE_COUNTER(ta, TA_BUSY);
92	UPDATE_COUNTER(gds, GDS_BUSY);
93	UPDATE_COUNTER(vgt, VGT_BUSY);
94	UPDATE_COUNTER(ia, IA_BUSY);
95	UPDATE_COUNTER(sx, SX_BUSY);
96	UPDATE_COUNTER(wd, WD_BUSY);
97	UPDATE_COUNTER(spi, SPI_BUSY);
98	UPDATE_COUNTER(bci, BCI_BUSY);
99	UPDATE_COUNTER(sc, SC_BUSY);
100	UPDATE_COUNTER(pa, PA_BUSY);
101	UPDATE_COUNTER(db, DB_BUSY);
102	UPDATE_COUNTER(cp, CP_BUSY);
103	UPDATE_COUNTER(cb, CB_BUSY);
104	UPDATE_COUNTER(gui, GUI_ACTIVE);
105	gui_busy = GUI_ACTIVE(value);
106
107	value = gui_busy || sdma_busy;
108	UPDATE_COUNTER(gpu, IDENTITY);
109}
110
111#undef UPDATE_COUNTER
112
113static int
114r600_gpu_load_thread(void *param)
115{
116	struct r600_common_screen *rscreen = (struct r600_common_screen*)param;
117	const int period_us = 1000000 / SAMPLES_PER_SEC;
118	int sleep_us = period_us;
119	int64_t cur_time, last_time = os_time_get();
120
121	while (!p_atomic_read(&rscreen->gpu_load_stop_thread)) {
122		if (sleep_us)
123			os_time_sleep(sleep_us);
124
125		/* Make sure we sleep the ideal amount of time to match
126		 * the expected frequency. */
127		cur_time = os_time_get();
128
129		if (os_time_timeout(last_time, last_time + period_us,
130				    cur_time))
131			sleep_us = MAX2(sleep_us - 1, 1);
132		else
133			sleep_us += 1;
134
135		/*printf("Hz: %.1f\n", 1000000.0 / (cur_time - last_time));*/
136		last_time = cur_time;
137
138		/* Update the counters. */
139		r600_update_mmio_counters(rscreen, &rscreen->mmio_counters);
140	}
141	p_atomic_dec(&rscreen->gpu_load_stop_thread);
142	return 0;
143}
144
145void r600_gpu_load_kill_thread(struct r600_common_screen *rscreen)
146{
147	if (!rscreen->gpu_load_thread)
148		return;
149
150	p_atomic_inc(&rscreen->gpu_load_stop_thread);
151	thrd_join(rscreen->gpu_load_thread, NULL);
152	rscreen->gpu_load_thread = 0;
153}
154
155static uint64_t r600_read_mmio_counter(struct r600_common_screen *rscreen,
156				       unsigned busy_index)
157{
158	/* Start the thread if needed. */
159	if (!rscreen->gpu_load_thread) {
160		mtx_lock(&rscreen->gpu_load_mutex);
161		/* Check again inside the mutex. */
162		if (!rscreen->gpu_load_thread)
163			rscreen->gpu_load_thread =
164				u_thread_create(r600_gpu_load_thread, rscreen);
165		mtx_unlock(&rscreen->gpu_load_mutex);
166	}
167
168	unsigned busy = p_atomic_read(&rscreen->mmio_counters.array[busy_index]);
169	unsigned idle = p_atomic_read(&rscreen->mmio_counters.array[busy_index + 1]);
170
171	return busy | ((uint64_t)idle << 32);
172}
173
174static unsigned r600_end_mmio_counter(struct r600_common_screen *rscreen,
175				      uint64_t begin, unsigned busy_index)
176{
177	uint64_t end = r600_read_mmio_counter(rscreen, busy_index);
178	unsigned busy = (end & 0xffffffff) - (begin & 0xffffffff);
179	unsigned idle = (end >> 32) - (begin >> 32);
180
181	/* Calculate the % of time the busy counter was being incremented.
182	 *
183	 * If no counters were incremented, return the current counter status.
184	 * It's for the case when the load is queried faster than
185	 * the counters are updated.
186	 */
187	if (idle || busy) {
188		return busy*100 / (busy + idle);
189	} else {
190		union r600_mmio_counters counters;
191
192		memset(&counters, 0, sizeof(counters));
193		r600_update_mmio_counters(rscreen, &counters);
194		return counters.array[busy_index] ? 100 : 0;
195	}
196}
197
198#define BUSY_INDEX(rscreen, field) (&rscreen->mmio_counters.named.field.busy - \
199				    rscreen->mmio_counters.array)
200
201static unsigned busy_index_from_type(struct r600_common_screen *rscreen,
202				     unsigned type)
203{
204	switch (type) {
205	case R600_QUERY_GPU_LOAD:
206		return BUSY_INDEX(rscreen, gpu);
207	case R600_QUERY_GPU_SHADERS_BUSY:
208		return BUSY_INDEX(rscreen, spi);
209	case R600_QUERY_GPU_TA_BUSY:
210		return BUSY_INDEX(rscreen, ta);
211	case R600_QUERY_GPU_GDS_BUSY:
212		return BUSY_INDEX(rscreen, gds);
213	case R600_QUERY_GPU_VGT_BUSY:
214		return BUSY_INDEX(rscreen, vgt);
215	case R600_QUERY_GPU_IA_BUSY:
216		return BUSY_INDEX(rscreen, ia);
217	case R600_QUERY_GPU_SX_BUSY:
218		return BUSY_INDEX(rscreen, sx);
219	case R600_QUERY_GPU_WD_BUSY:
220		return BUSY_INDEX(rscreen, wd);
221	case R600_QUERY_GPU_BCI_BUSY:
222		return BUSY_INDEX(rscreen, bci);
223	case R600_QUERY_GPU_SC_BUSY:
224		return BUSY_INDEX(rscreen, sc);
225	case R600_QUERY_GPU_PA_BUSY:
226		return BUSY_INDEX(rscreen, pa);
227	case R600_QUERY_GPU_DB_BUSY:
228		return BUSY_INDEX(rscreen, db);
229	case R600_QUERY_GPU_CP_BUSY:
230		return BUSY_INDEX(rscreen, cp);
231	case R600_QUERY_GPU_CB_BUSY:
232		return BUSY_INDEX(rscreen, cb);
233	case R600_QUERY_GPU_SDMA_BUSY:
234		return BUSY_INDEX(rscreen, sdma);
235	case R600_QUERY_GPU_PFP_BUSY:
236		return BUSY_INDEX(rscreen, pfp);
237	case R600_QUERY_GPU_MEQ_BUSY:
238		return BUSY_INDEX(rscreen, meq);
239	case R600_QUERY_GPU_ME_BUSY:
240		return BUSY_INDEX(rscreen, me);
241	case R600_QUERY_GPU_SURF_SYNC_BUSY:
242		return BUSY_INDEX(rscreen, surf_sync);
243	case R600_QUERY_GPU_CP_DMA_BUSY:
244		return BUSY_INDEX(rscreen, cp_dma);
245	case R600_QUERY_GPU_SCRATCH_RAM_BUSY:
246		return BUSY_INDEX(rscreen, scratch_ram);
247	default:
248		unreachable("invalid query type");
249	}
250}
251
252uint64_t r600_begin_counter(struct r600_common_screen *rscreen, unsigned type)
253{
254	unsigned busy_index = busy_index_from_type(rscreen, type);
255	return r600_read_mmio_counter(rscreen, busy_index);
256}
257
258unsigned r600_end_counter(struct r600_common_screen *rscreen, unsigned type,
259			  uint64_t begin)
260{
261	unsigned busy_index = busy_index_from_type(rscreen, type);
262	return r600_end_mmio_counter(rscreen, begin, busy_index);
263}
264