/* SPDX-License-Identifier: BSD-3-Clause

*

* Copyright(c) 2019 Intel Corporation. All rights reserved.

*

* Author: Marcin Maka <marcin.maka@linux.intel.com>

*/

/**

* \file xtos/include/sof/lib/perf_cnt.h

* \brief Simple performance counters

* \author Marcin Maka <marcin.maka@linux.intel.com>

*/

#ifndef __SOF_LIB_PERF_CNT_H__

#define __SOF_LIB_PERF_CNT_H__

#include <rtos/timer.h>

struct perf_cnt_data {

uint32_t plat_ts;

uint32_t cpu_ts;

uint32_t plat_delta_last;

uint32_t plat_delta_peak;

uint32_t cpu_delta_last;

uint32_t cpu_delta_peak;

uint32_t cpu_delta_sum;

uint32_t sample_cnt;

};

#if CONFIG_PERFORMANCE_COUNTERS

#define perf_cnt_trace(ctx, pcd) \

tr_info(ctx, "perf plat last %u peak %u cpu last %u, peak %u", \

(uint32_t)((pcd)->plat_delta_last), \

(uint32_t)((pcd)->plat_delta_peak), \

(uint32_t)((pcd)->cpu_delta_last), \

(uint32_t)((pcd)->cpu_delta_peak))

/** \brief Clears performance counters data. */

#define perf_cnt_clear(pcd) memset((pcd), 0, sizeof(struct perf_cnt_data))

/* NOTE: Zephyr's arch_timing_counter_get() might not be implemented

* for a particular platform. In this case let's fallback to use

* Zephyr's k_cycle_get_64(). This will result in both "platform" and

* "cpu" timestamps to be equal.

*/

#ifdef __ZEPHYR__

#ifdef CONFIG_TIMING_FUNCTIONS

#define perf_cnt_get_cpu_ts arch_timing_counter_get

#else

#define perf_cnt_get_cpu_ts sof_cycle_get_64

#endif /* CONFIG_TIMING_FUNCTIONS */

#else

#define perf_cnt_get_cpu_ts() timer_get_system(cpu_timer_get())

#endif /* __ZEPHYR__ */

/** \brief Initializes timestamps with current timer values. */

#define perf_cnt_init(pcd) do { \

(pcd)->plat_ts = sof_cycle_get_64(); \

(pcd)->cpu_ts = perf_cnt_get_cpu_ts(); \

} while (0)

/* Trace macros that can be used as trace_m argument of the perf_cnt_stamp()

* to trace PCD values if the last arch timer reading exceeds the previous

* peak value.

*

* arg passed to perf_cnt_stamp() is forwarded to the trace_m() macro

* as the second argument.

*/

/** \brief No trace when detecting peak value. */

#define perf_trace_null(pcd, arg)

/** \brief Simple trace, all values are printed, arg should be a tr_ctx address.

*/

#define perf_trace_simple(pcd, arg) perf_cnt_trace(arg, pcd)

/* perf measurement windows size 2^x */

#define PERF_CNT_CHECK_WINDOW_SIZE 10

#define task_perf_avg_info(pcd, task_p, class) \

tr_info(task_p, "perf_cycle task %p, %pU cpu avg %u peak %u",\

class, (class)->uid, \

(uint32_t)((pcd)->cpu_delta_sum), \

(uint32_t)((pcd)->cpu_delta_peak))

#define task_perf_cnt_avg(pcd, trace_m, arg, class) do { \

(pcd)->cpu_delta_sum += (pcd)->cpu_delta_last; \

if (++(pcd)->sample_cnt == 1 << PERF_CNT_CHECK_WINDOW_SIZE) { \

(pcd)->cpu_delta_sum >>= PERF_CNT_CHECK_WINDOW_SIZE; \

trace_m(pcd, arg, class); \

(pcd)->cpu_delta_sum = 0; \

(pcd)->sample_cnt = 0; \

(pcd)->plat_delta_peak = 0; \

(pcd)->cpu_delta_peak = 0; \

} \

} while (0)

/** \brief Accumulates cpu timer delta samples calculated by perf_cnt_stamp().

*

* If current sample count reaches the window size, compute the average and run trace_m.

* \param pcd Performance counters data.

* \param trace_m Trace function trace_m(pcd, arg) or trace macro if a

* more precise line number is desired in the logs.

* \param arg Argument passed to trace_m as arg.

*/

#define perf_cnt_average(pcd, trace_m, arg) do { \

(pcd)->cpu_delta_sum += (pcd)->cpu_delta_last; \

if (++(pcd)->sample_cnt == 1 << PERF_CNT_CHECK_WINDOW_SIZE) {\

(pcd)->cpu_delta_sum >>= PERF_CNT_CHECK_WINDOW_SIZE; \

trace_m(pcd, arg); \

(pcd)->cpu_delta_sum = 0; \

(pcd)->sample_cnt = 0; \

(pcd)->plat_delta_peak = 0; \

(pcd)->cpu_delta_peak = 0; \

} \

} while (0)

/** \brief Reads the timers and computes delta to the previous readings.

*

* If current arch delta exceeds the previous peak value, trace_m is run.

* \param pcd Performance counters data.

* \param trace_m Trace function trace_m(pcd, arg) or trace macro if a

* more precise line number is desired in the logs.

* \param arg Argument passed to trace_m as arg.

*/

#define perf_cnt_stamp(pcd, trace_m, arg) do { \

uint32_t plat_ts = \

(uint32_t)sof_cycle_get_64(); \

uint32_t cpu_ts = \

(uint32_t)perf_cnt_get_cpu_ts(); \

if (plat_ts > (pcd)->plat_ts) \

(pcd)->plat_delta_last = plat_ts - (pcd)->plat_ts; \

else \

(pcd)->plat_delta_last = UINT32_MAX - (pcd)->plat_ts \

+ plat_ts; \

if (cpu_ts > (pcd)->cpu_ts) \

(pcd)->cpu_delta_last = cpu_ts - (pcd)->cpu_ts; \

else \

(pcd)->cpu_delta_last = UINT32_MAX - (pcd)->cpu_ts \

+ cpu_ts;\

if ((pcd)->plat_delta_last > (pcd)->plat_delta_peak) \

(pcd)->plat_delta_peak = (pcd)->plat_delta_last; \

if ((pcd)->cpu_delta_last > (pcd)->cpu_delta_peak) { \

(pcd)->cpu_delta_peak = (pcd)->cpu_delta_last; \

trace_m(pcd, arg); \

} \

} while (0)

/**

* For simple performance measurement and optimization in development stage,

* tic-toc api is provided. Performance data are traced at each tok call,

* to allow fast clocks usage deviation estimation. Example:

*

* \code{.c}

* void foo(struct comp_dev *dev) {

* static struct perf_cnt_data pcd;

*

* perf_tic(&pcd);

* bar();

* perf_toc(&pcd, dev);

* }

* \endcode

*/

/** \brief Save start timestamp in pcd structure

*

* \param pcd Performance counters data.

*/

#define perf_tic(pcd) \

perf_cnt_init(pcd)

/** \brief Save start timestamp in pcd structure

*

* \param pcd Performance counters data.

* \param comp Component used to get corresponding trace context.

*/

#define perf_toc(pcd, comp) do { \

perf_cnt_stamp(pcd, perf_trace_null, NULL); \

perf_trace_simple(pcd, trace_comp_get_tr_ctx(comp)); \

} while (0)

#else

#define perf_cnt_clear(pcd)

#define perf_cnt_init(pcd)

#define perf_cnt_stamp(pcd, trace_m, arg)

#define perf_cnt_average(pcd, trace_m, arg)

#endif

#endif /* __SOF_LIB_PERF_CNT_H__ */