2 * Copyright (C) 2012-2014 ARM Limited. All rights reserved.
4 * This program is free software and is provided to you under the terms of the GNU General Public License version 2
5 * as published by the Free Software Foundation, and any use by you of this program is subject to the terms of such GNU licence.
7 * A copy of the licence is included with the program, and can also be obtained from Free Software
8 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
11 #include <linux/module.h>
13 #include <mali_profiling_gator_api.h>
14 #include "mali_kernel_common.h"
17 #include "mali_uk_types.h"
18 #include "mali_osk_profiling.h"
19 #include "mali_linux_trace.h"
22 #include "mali_l2_cache.h"
23 #include "mali_user_settings_db.h"
24 #include "mali_executor.h"
26 _mali_osk_errcode_t _mali_osk_profiling_init(mali_bool auto_start)
28 if (MALI_TRUE == auto_start) {
29 mali_set_user_setting(_MALI_UK_USER_SETTING_SW_EVENTS_ENABLE, MALI_TRUE);
32 return _MALI_OSK_ERR_OK;
35 void _mali_osk_profiling_term(void)
40 void _mali_osk_profiling_report_sw_counters(u32 *counters)
42 trace_mali_sw_counters(_mali_osk_get_pid(), _mali_osk_get_tid(), NULL, counters);
45 void _mali_osk_profiling_memory_usage_get(u32 *memory_usage)
47 *memory_usage = _mali_ukk_report_memory_usage();
50 _mali_osk_errcode_t _mali_ukk_profiling_add_event(_mali_uk_profiling_add_event_s *args)
52 /* Always add process and thread identificator in the first two data elements for events from user space */
53 _mali_osk_profiling_add_event(args->event_id, _mali_osk_get_pid(), _mali_osk_get_tid(), args->data[2], args->data[3], args->data[4]);
55 return _MALI_OSK_ERR_OK;
58 _mali_osk_errcode_t _mali_ukk_sw_counters_report(_mali_uk_sw_counters_report_s *args)
60 u32 *counters = (u32 *)(uintptr_t)args->counters;
62 _mali_osk_profiling_report_sw_counters(counters);
64 return _MALI_OSK_ERR_OK;
67 _mali_osk_errcode_t _mali_ukk_profiling_memory_usage_get(_mali_uk_profiling_memory_usage_get_s *args)
69 _mali_osk_profiling_memory_usage_get(&args->memory_usage);
70 return _MALI_OSK_ERR_OK;
74 * Called by gator.ko to set HW counters
76 * @param counter_id The counter ID.
77 * @param event_id Event ID that the counter should count (HW counter value from TRM).
79 * @return 1 on success, 0 on failure.
81 int _mali_profiling_set_event(u32 counter_id, s32 event_id)
83 if (COUNTER_VP_0_C0 == counter_id) {
84 mali_gp_job_set_gp_counter_src0(event_id);
85 } else if (COUNTER_VP_0_C1 == counter_id) {
86 mali_gp_job_set_gp_counter_src1(event_id);
87 } else if (COUNTER_FP_0_C0 <= counter_id && COUNTER_FP_7_C1 >= counter_id) {
89 * Two compatibility notes for this function:
91 * 1) Previously the DDK allowed per core counters.
93 * This did not make much sense on Mali-450 with the "virtual PP core" concept,
94 * so this option was removed, and only the same pair of HW counters was allowed on all cores,
95 * beginning with r3p2 release.
97 * Starting with r4p0, it is now possible to set different HW counters for the different sub jobs.
98 * This should be almost the same, since sub job 0 is designed to run on core 0,
99 * sub job 1 on core 1, and so on.
101 * The scheduling of PP sub jobs is not predictable, and this often led to situations where core 0 ran 2
102 * sub jobs, while for instance core 1 ran zero. Having the counters set per sub job would thus increase
103 * the predictability of the returned data (as you would be guaranteed data for all the selected HW counters).
105 * PS: Core scaling needs to be disabled in order to use this reliably (goes for both solutions).
107 * The framework/#defines with Gator still indicates that the counter is for a particular core,
108 * but this is internally used as a sub job ID instead (no translation needed).
110 * 2) Global/default vs per sub job counters
112 * Releases before r3p2 had only per PP core counters.
113 * r3p2 releases had only one set of default/global counters which applied to all PP cores
114 * Starting with r4p0, we have both a set of default/global counters,
115 * and individual counters per sub job (equal to per core).
117 * To keep compatibility with Gator/DS-5/streamline, the following scheme is used:
119 * r3p2 release; only counters set for core 0 is handled,
120 * this is applied as the default/global set of counters, and will thus affect all cores.
122 * r4p0 release; counters set for core 0 is applied as both the global/default set of counters,
123 * and counters for sub job 0.
124 * Counters set for core 1-7 is only applied for the corresponding sub job.
126 * This should allow the DS-5/Streamline GUI to have a simple mode where it only allows setting the
127 * values for core 0, and thus this will be applied to all PP sub jobs/cores.
128 * Advanced mode will also be supported, where individual pairs of HW counters can be selected.
130 * The GUI will (until it is updated) still refer to cores instead of sub jobs, but this is probably
131 * something we can live with!
133 * Mali-450 note: Each job is not divided into a deterministic number of sub jobs, as the HW DLBU
134 * automatically distributes the load between whatever number of cores is available at this particular time.
135 * A normal PP job on Mali-450 is thus considered a single (virtual) job, and it will thus only be possible
136 * to use a single pair of HW counters (even if the job ran on multiple PP cores).
137 * In other words, only the global/default pair of PP HW counters will be used for normal Mali-450 jobs.
139 u32 sub_job = (counter_id - COUNTER_FP_0_C0) >> 1;
140 u32 counter_src = (counter_id - COUNTER_FP_0_C0) & 1;
141 if (0 == counter_src) {
142 mali_pp_job_set_pp_counter_sub_job_src0(sub_job, event_id);
144 mali_pp_job_set_pp_counter_global_src0(event_id);
147 mali_pp_job_set_pp_counter_sub_job_src1(sub_job, event_id);
149 mali_pp_job_set_pp_counter_global_src1(event_id);
152 } else if (COUNTER_L2_0_C0 <= counter_id && COUNTER_L2_2_C1 >= counter_id) {
153 u32 core_id = (counter_id - COUNTER_L2_0_C0) >> 1;
154 struct mali_l2_cache_core *l2_cache_core = mali_l2_cache_core_get_glob_l2_core(core_id);
156 if (NULL != l2_cache_core) {
157 u32 counter_src = (counter_id - COUNTER_L2_0_C0) & 1;
158 mali_l2_cache_core_set_counter_src(l2_cache_core,
159 counter_src, event_id);
162 return 0; /* Failure, unknown event */
165 return 1; /* success */
169 * Called by gator.ko to retrieve the L2 cache counter values for all L2 cache cores.
170 * The L2 cache counters are unique in that they are polled by gator, rather than being
171 * transmitted via the tracepoint mechanism.
173 * @param values Pointer to a _mali_profiling_l2_counter_values structure where
174 * the counter sources and values will be output
175 * @return 0 if all went well; otherwise, return the mask with the bits set for the powered off cores
177 u32 _mali_profiling_get_l2_counters(_mali_profiling_l2_counter_values *values)
179 u32 l2_cores_num = mali_l2_cache_core_get_glob_num_l2_cores();
182 MALI_DEBUG_ASSERT(l2_cores_num <= 3);
184 for (i = 0; i < l2_cores_num; i++) {
185 struct mali_l2_cache_core *l2_cache = mali_l2_cache_core_get_glob_l2_core(i);
187 if (NULL == l2_cache) {
191 mali_l2_cache_core_get_counter_values(l2_cache,
192 &values->cores[i].source0,
193 &values->cores[i].value0,
194 &values->cores[i].source1,
195 &values->cores[i].value1);
202 * Called by gator to control the production of profiling information at runtime.
204 void _mali_profiling_control(u32 action, u32 value)
207 case FBDUMP_CONTROL_ENABLE:
208 mali_set_user_setting(_MALI_UK_USER_SETTING_COLORBUFFER_CAPTURE_ENABLED, (value == 0 ? MALI_FALSE : MALI_TRUE));
210 case FBDUMP_CONTROL_RATE:
211 mali_set_user_setting(_MALI_UK_USER_SETTING_BUFFER_CAPTURE_N_FRAMES, value);
213 case SW_COUNTER_ENABLE:
214 mali_set_user_setting(_MALI_UK_USER_SETTING_SW_COUNTER_ENABLED, value);
216 case FBDUMP_CONTROL_RESIZE_FACTOR:
217 mali_set_user_setting(_MALI_UK_USER_SETTING_BUFFER_CAPTURE_RESIZE_FACTOR, value);
220 break; /* Ignore unimplemented actions */
225 * Called by gator to get mali api version.
227 u32 _mali_profiling_get_api_version(void)
229 return MALI_PROFILING_API_VERSION;
233 * Called by gator to get the data about Mali instance in use:
234 * product id, version, number of cores
236 void _mali_profiling_get_mali_version(struct _mali_profiling_mali_version *values)
238 values->mali_product_id = (u32)mali_kernel_core_get_product_id();
239 values->mali_version_major = mali_kernel_core_get_gpu_major_version();
240 values->mali_version_minor = mali_kernel_core_get_gpu_minor_version();
241 values->num_of_l2_cores = mali_l2_cache_core_get_glob_num_l2_cores();
242 values->num_of_fp_cores = mali_executor_get_num_cores_total();
243 values->num_of_vp_cores = 1;
247 EXPORT_SYMBOL(_mali_profiling_set_event);
248 EXPORT_SYMBOL(_mali_profiling_get_l2_counters);
249 EXPORT_SYMBOL(_mali_profiling_control);
250 EXPORT_SYMBOL(_mali_profiling_get_api_version);
251 EXPORT_SYMBOL(_mali_profiling_get_mali_version);