MLECO-1870: Cherry pick profiling changes from dev to open source repo
* Documentation update
Change-Id: If85e7ebc44498840b291c408f14e66a5a5faa424
Signed-off-by: Isabella Gottardi <isabella.gottardi@arm.com>
diff --git a/source/application/hal/platforms/bare-metal/bsp/bsp-core/include/bsp_core_log.h b/source/application/hal/platforms/bare-metal/bsp/bsp-core/include/bsp_core_log.h
index f049209..ffb55e7 100644
--- a/source/application/hal/platforms/bare-metal/bsp/bsp-core/include/bsp_core_log.h
+++ b/source/application/hal/platforms/bare-metal/bsp/bsp-core/include/bsp_core_log.h
@@ -32,31 +32,31 @@
#endif /*LOG_LEVEL*/
#if (LOG_LEVEL == LOG_LEVEL_TRACE)
- #define trace(...) printf("[TRACE] "); printf(__VA_ARGS__)
+ #define trace(...) printf("TRACE - "); printf(__VA_ARGS__)
#else
#define trace(...)
#endif /* LOG_LEVEL == LOG_LEVEL_TRACE */
#if (LOG_LEVEL <= LOG_LEVEL_DEBUG)
- #define debug(...) printf("[DEBUG] "); printf(__VA_ARGS__)
+ #define debug(...) printf("DEBUG - "); printf(__VA_ARGS__)
#else
#define debug(...)
#endif /* LOG_LEVEL > LOG_LEVEL_TRACE */
#if (LOG_LEVEL <= LOG_LEVEL_INFO)
- #define info(...) printf("[INFO] "); printf(__VA_ARGS__)
+ #define info(...) printf("INFO - "); printf(__VA_ARGS__)
#else
#define info(...)
#endif /* LOG_LEVEL > LOG_LEVEL_DEBUG */
#if (LOG_LEVEL <= LOG_LEVEL_WARN)
- #define warn(...) printf("[WARN] "); printf(__VA_ARGS__)
+ #define warn(...) printf("WARN - "); printf(__VA_ARGS__)
#else
#define warn(...)
#endif /* LOG_LEVEL > LOG_LEVEL_INFO */
#if (LOG_LEVEL <= LOG_LEVEL_ERROR)
- #define printf_err(...) printf("[ERROR] "); printf(__VA_ARGS__)
+ #define printf_err(...) printf("ERROR - "); printf(__VA_ARGS__)
#else
#define printf_err(...)
#endif /* LOG_LEVEL > LOG_LEVEL_INFO */
diff --git a/source/application/hal/platforms/bare-metal/timer/baremetal_timer.c b/source/application/hal/platforms/bare-metal/timer/baremetal_timer.c
index 7257c1d..ef31a71 100644
--- a/source/application/hal/platforms/bare-metal/timer/baremetal_timer.c
+++ b/source/application/hal/platforms/bare-metal/timer/baremetal_timer.c
@@ -50,6 +50,53 @@
static uint64_t bm_get_npu_active_cycle_diff(time_counter *st,
time_counter *end);
+/** @brief Gets the difference in idle NPU cycle counts
+ * @param[in] st Pointer to time_counter value at start time.
+ * @param[in] end Pointer to time_counter value at end.
+ * @return Idle NPU cycle counts difference between the arguments expressed
+ * as unsigned 64 bit integer.
+ **/
+static uint64_t bm_get_npu_idle_cycle_diff(time_counter *st,
+ time_counter *end);
+
+/** @brief Gets the difference in axi0 bus reads cycle counts
+ * @param[in] st Pointer to time_counter value at start time.
+ * @param[in] end Pointer to time_counter value at end.
+ * @return NPU AXI0 read cycle counts difference between the arguments expressed
+ * as unsigned 64 bit integer.
+ **/
+static uint64_t bm_get_npu_axi0_read_cycle_diff(time_counter *st,
+ time_counter *end);
+
+/** @brief Gets the difference in axi0 bus writes cycle counts
+ * @param[in] st Pointer to time_counter value at start time.
+ * @param[in] end Pointer to time_counter value at end.
+ * @return NPU AXI0 write cycle counts difference between the arguments expressed
+ * as unsigned 64 bit integer.
+ **/
+static uint64_t bm_get_npu_axi0_write_cycle_diff(time_counter *st,
+ time_counter *end);
+
+/** @brief Gets the difference in axi1 bus reads cycle counts
+ * @param[in] st Pointer to time_counter value at start time.
+ * @param[in] end Pointer to time_counter value at end.
+ * @return NPU AXI1 read cycle counts difference between the arguments expressed
+ * as unsigned 64 bit integer.
+ **/
+static uint64_t bm_get_npu_axi1_read_cycle_diff(time_counter *st,
+ time_counter *end);
+
+/** @brief Gets the difference for 6 collected cycle counts:
+ * 1) total NPU
+ * 2) active NPU
+ * 3) idle NPU
+ * 4) axi0 read
+ * 5) axi0 write
+ * 6) axi1 read
+ * */
+static int bm_get_npu_cycle_diff(time_counter *st, time_counter *end,
+ uint64_t* pmu_counters_values, const size_t size);
+
#endif /* defined (ARM_NPU) */
#if defined(MPS3_PLATFORM)
@@ -126,8 +173,7 @@
#if defined (ARM_NPU)
/* We are capable of reporting npu cycle counts. */
timer->cap.npu_cycles = 1;
- timer->get_npu_total_cycle_diff = bm_get_npu_total_cycle_diff;
- timer->get_npu_active_cycle_diff = bm_get_npu_active_cycle_diff;
+ timer->get_npu_cycles_diff = bm_get_npu_cycle_diff;
_init_ethosu_cyclecounter();
#endif /* defined (ARM_NPU) */
@@ -149,37 +195,89 @@
/* Reset overflow status. */
ETHOSU_PMU_Set_CNTR_OVS(ETHOSU_PMU_CNT1_Msk | ETHOSU_PMU_CCNT_Msk);
- /* Set the counter #0 to count idle cycles. */
+ /* We can retrieve only 4 PMU counters: */
ETHOSU_PMU_Set_EVTYPER(0, ETHOSU_PMU_NPU_IDLE);
+ ETHOSU_PMU_Set_EVTYPER(1, ETHOSU_PMU_AXI0_RD_DATA_BEAT_RECEIVED);
+ ETHOSU_PMU_Set_EVTYPER(2, ETHOSU_PMU_AXI0_WR_DATA_BEAT_WRITTEN);
+ ETHOSU_PMU_Set_EVTYPER(3, ETHOSU_PMU_AXI1_RD_DATA_BEAT_RECEIVED);
/* Enable PMU. */
ETHOSU_PMU_Enable();
/* Enable counters for cycle and counter# 0. */
- ETHOSU_PMU_CNTR_Enable(ETHOSU_PMU_CNT1_Msk | ETHOSU_PMU_CCNT_Msk);
-
+ ETHOSU_PMU_CNTR_Enable(ETHOSU_PMU_CNT1_Msk | ETHOSU_PMU_CNT2_Msk | ETHOSU_PMU_CNT3_Msk | ETHOSU_PMU_CNT4_Msk| ETHOSU_PMU_CCNT_Msk);
_reset_ethosu_counters();
}
+static int bm_get_npu_cycle_diff(time_counter *st, time_counter *end,
+ uint64_t* pmu_counters_values, const size_t size)
+{
+ if (size == 6) {
+ pmu_counters_values[0] = bm_get_npu_total_cycle_diff(st, end);
+ pmu_counters_values[1] = bm_get_npu_active_cycle_diff(st, end);
+ pmu_counters_values[2] = bm_get_npu_idle_cycle_diff(st, end);
+ pmu_counters_values[3] = bm_get_npu_axi0_read_cycle_diff(st, end);
+ pmu_counters_values[4] = bm_get_npu_axi0_write_cycle_diff(st, end);
+ pmu_counters_values[5] = bm_get_npu_axi1_read_cycle_diff(st, end);
+ return 0;
+ } else {
+ return 1;
+ }
+}
+
static uint64_t bm_get_npu_total_cycle_diff(time_counter *st, time_counter *end)
{
return end->npu_total_ccnt - st->npu_total_ccnt;
}
-static uint64_t bm_get_npu_active_cycle_diff(time_counter *st, time_counter *end)
+static uint32_t counter_overflow(uint32_t pmu_counter_mask)
{
/* Check for overflow: The idle counter is 32 bit while the
total cycle count is 64 bit. */
const uint32_t overflow_status = ETHOSU_PMU_Get_CNTR_OVS();
+ return pmu_counter_mask & overflow_status;
+}
- if (ETHOSU_PMU_CNT1_Msk & overflow_status) {
+static uint64_t bm_get_npu_idle_cycle_diff(time_counter *st, time_counter *end)
+{
+ if (counter_overflow(ETHOSU_PMU_CNT1_Msk)) {
printf_err("EthosU PMU idle counter overflow.\n");
return 0;
}
+ return (uint64_t)(end->npu_idle_ccnt - st->npu_idle_ccnt);
+}
+static uint64_t bm_get_npu_active_cycle_diff(time_counter *st, time_counter *end)
+{
/* Active NPU time = total time - idle time */
- return (bm_get_npu_total_cycle_diff(st, end) +
- (uint64_t)(st->npu_idle_ccnt)) - (uint64_t)(end->npu_idle_ccnt);
+ return bm_get_npu_total_cycle_diff(st, end) - bm_get_npu_idle_cycle_diff(st, end);
+}
+
+static uint64_t bm_get_npu_axi0_read_cycle_diff(time_counter *st, time_counter *end)
+{
+ if (counter_overflow(ETHOSU_PMU_CNT2_Msk)) {
+ printf_err("EthosU PMU axi0 read counter overflow.\n");
+ return 0;
+ }
+ return (uint64_t)(end->npu_axi0_read_ccnt - st->npu_axi0_read_ccnt);
+}
+
+static uint64_t bm_get_npu_axi0_write_cycle_diff(time_counter *st, time_counter *end)
+{
+ if (counter_overflow(ETHOSU_PMU_CNT3_Msk)) {
+ printf_err("EthosU PMU axi0 write counter overflow.\n");
+ return 0;
+ }
+ return (uint64_t)(end->npu_axi0_write_ccnt - st->npu_axi0_write_ccnt);
+}
+
+static uint64_t bm_get_npu_axi1_read_cycle_diff(time_counter *st, time_counter *end)
+{
+ if (counter_overflow(ETHOSU_PMU_CNT4_Msk)) {
+ printf_err("EthosU PMU axi1 read counter overflow.\n");
+ return 0;
+ }
+ return (uint64_t)(end->npu_axi1_read_ccnt - st->npu_axi1_read_ccnt);
}
#endif /* defined (ARM_NPU) */
@@ -199,15 +297,22 @@
.counter = get_time_counter(),
#if defined (ARM_NPU)
- .npu_idle_ccnt = ETHOSU_PMU_Get_EVCNTR(0),
- .npu_total_ccnt = ETHOSU_PMU_Get_CCNTR()
+ .npu_total_ccnt = ETHOSU_PMU_Get_CCNTR(),
+ .npu_idle_ccnt = ETHOSU_PMU_Get_EVCNTR(0),
+ .npu_axi0_read_ccnt = ETHOSU_PMU_Get_EVCNTR(1),
+ .npu_axi0_write_ccnt = ETHOSU_PMU_Get_EVCNTR(2),
+ .npu_axi1_read_ccnt = ETHOSU_PMU_Get_EVCNTR(3)
#endif /* defined (ARM_NPU) */
};
#if defined (ARM_NPU)
- debug("NPU total cc: %llu; NPU idle cc: %u\n",
- t.npu_total_ccnt, t.npu_idle_ccnt);
+ debug("NPU total cc: %llu; NPU idle cc: %u; NPU axi0 read cc: %u; NPU axi0 write cc: %u; NPU axi1 read cc: %u\n",
+ t.npu_total_ccnt,
+ t.npu_idle_ccnt,
+ t.npu_axi0_read_ccnt,
+ t.npu_axi0_write_ccnt,
+ t.npu_axi1_read_ccnt);
#endif /* defined (ARM_NPU) */
return t;
diff --git a/source/application/hal/platforms/bare-metal/timer/include/baremetal_timer.h b/source/application/hal/platforms/bare-metal/timer/include/baremetal_timer.h
index c8fc32c..3020dac 100644
--- a/source/application/hal/platforms/bare-metal/timer/include/baremetal_timer.h
+++ b/source/application/hal/platforms/bare-metal/timer/include/baremetal_timer.h
@@ -34,6 +34,9 @@
#if defined (ARM_NPU)
uint64_t npu_total_ccnt;
uint32_t npu_idle_ccnt;
+ uint32_t npu_axi0_read_ccnt;
+ uint32_t npu_axi0_write_ccnt;
+ uint32_t npu_axi1_read_ccnt;
#endif /* ARM_NPU */
} time_counter;
diff --git a/source/application/hal/platforms/native/data_presentation/log/log.c b/source/application/hal/platforms/native/data_presentation/log/log.c
index 48e8b95..6ce6684 100644
--- a/source/application/hal/platforms/native/data_presentation/log/log.c
+++ b/source/application/hal/platforms/native/data_presentation/log/log.c
@@ -30,11 +30,11 @@
const uint32_t pos_x, const uint32_t pos_y,
const uint32_t downsample_factor)
{
- info("Image details\n");
- info("Data: %p\n", data);
- info("WxHxC: %dx%dx%d\n", width, height, channels);
- info("Pos (x,y): (%d,%d)\n", pos_x, pos_y);
- info("Downsampling factor: %u\n", downsample_factor);
+ debug("Image details\n");
+ debug("Data: %p\n", data);
+ debug("WxHxC: %dx%dx%d\n", width, height, channels);
+ debug("Pos (x,y): (%d,%d)\n", pos_x, pos_y);
+ debug("Downsampling factor: %u\n", downsample_factor);
return 0;
}
@@ -43,8 +43,8 @@
const bool allow_multiple_lines)
{
UNUSED(allow_multiple_lines);
- info("%s\n", str);
- info("Text size: %lu, x: %d, y: %d\n", str_sz, pos_x, pos_y);
+ debug("%s\n", str);
+ debug("Text size: %lu, x: %d, y: %d\n", str_sz, pos_x, pos_y);
return 0;
}
@@ -53,19 +53,19 @@
const uint32_t width, const uint32_t height,
const uint16_t color)
{
- info("Showing rectangular, width: %d, height: %d, color: %d, x: %d, y: %d\n",
+ debug("Showing rectangular, width: %d, height: %d, color: %d, x: %d, y: %d\n",
width, height, color, pos_x, pos_y);
return 0;
}
int log_clear(const uint16_t color)
{
- info("Clearing with color: %d\n", color);
+ debug("Clearing with color: %d\n", color);
return 0;
}
int log_set_text_color (const uint16_t color)
{
- info("Setting text color: %d\n", color);
+ debug("Setting text color: %d\n", color);
return 0;
}
diff --git a/source/application/hal/platforms/native/utils/include/dummy_log.h b/source/application/hal/platforms/native/utils/include/dummy_log.h
index 626436a..3df5c5c 100644
--- a/source/application/hal/platforms/native/utils/include/dummy_log.h
+++ b/source/application/hal/platforms/native/utils/include/dummy_log.h
@@ -32,31 +32,31 @@
#define UNUSED(x) ((void)(x))
#if (LOG_LEVEL == LOG_LEVEL_TRACE)
- #define trace(...) printf("[TRACE] "); printf(__VA_ARGS__)
+ #define trace(...) printf("TRACE - "); printf(__VA_ARGS__)
#else
#define trace(...)
#endif /* LOG_LEVEL == LOG_LEVEL_TRACE */
#if (LOG_LEVEL <= LOG_LEVEL_DEBUG)
- #define debug(...) printf("[DEBUG] "); printf(__VA_ARGS__)
+ #define debug(...) printf("DEBUG - "); printf(__VA_ARGS__)
#else
#define debug(...)
#endif /* LOG_LEVEL > LOG_LEVEL_TRACE */
#if (LOG_LEVEL <= LOG_LEVEL_INFO)
- #define info(...) printf("[INFO] "); printf(__VA_ARGS__)
+ #define info(...) printf("INFO - "); printf(__VA_ARGS__)
#else
#define info(...)
#endif /* LOG_LEVEL > LOG_LEVEL_DEBUG */
#if (LOG_LEVEL <= LOG_LEVEL_WARN)
- #define warn(...) printf("[WARN] "); printf(__VA_ARGS__)
+ #define warn(...) printf("WARN - "); printf(__VA_ARGS__)
#else
#define warn(...)
#endif /* LOG_LEVEL > LOG_LEVEL_INFO */
#if (LOG_LEVEL <= LOG_LEVEL_ERROR)
- #define printf_err(...) printf("[ERROR] "); printf(__VA_ARGS__)
+ #define printf_err(...) printf("ERROR - "); printf(__VA_ARGS__)
#else
#define printf_err(...)
#endif /* LOG_LEVEL > LOG_LEVEL_INFO */