Performance improvements (75% faster) in data processing.

The accelerometer data has to be processed before the sending it
to the model.

For the ML Trainer filters, before:
~5843 microsec (~373072 cycles)
After:
~1405 microsec (~ 80872 cycles)

1/4 of the time.

Author: microbit-carlos

mlrunner/mldataprocessor.c

@@ -69,11 +69,13 @@ MldpReturn_t filterStdDev(const float *data_in, const int in_size, float *data_o
     }
 
     float std = 0;
+    float f = 0;
     for (int i = 0; i < in_size; i++) {
-        std += (data_in[i] - mean) * (data_in[i] - mean);
+        f = data_in[i] - mean;
+        std += f * f;
     }
     std /= in_size;
-    *data_out = sqrt(std);
+    *data_out = sqrtf(std);
 
     return MLDP_SUCCESS;
 }
@@ -112,8 +114,8 @@ MldpReturn_t filterPeaks(const float *data_in, const int in_size, float *data_ou
     stdFilter[lag - 1] = stdDev_lag;
 
     for (int i = lag; i < in_size; i++) {
-        if (fabs(data_in[i] - avgFilter[i - 1]) > 0.1 &&
-            fabs(data_in[i] - avgFilter[i - 1]) > threshold * stdFilter[i - 1]
+        if (fabsf(data_in[i] - avgFilter[i - 1]) > 0.1f &&
+            fabsf(data_in[i] - avgFilter[i - 1]) > threshold * stdFilter[i - 1]
         ) {
             if (data_in[i] > avgFilter[i - 1]) {
                 signals[i] = +1; // positive signal
@@ -124,7 +126,7 @@ MldpReturn_t filterPeaks(const float *data_in, const int in_size, float *data_ou
                 signals[i] = -1; // negative signal
             }
             // make influence lower
-            filteredY[i] = influence * data_in[i] + (1 - influence) * filteredY[i - 1];
+            filteredY[i] = influence * data_in[i] + (1.0f - influence) * filteredY[i - 1];
         } else {
             signals[i] = 0; // no signal
             filteredY[i] = data_in[i];
@@ -151,7 +153,7 @@ MldpReturn_t filterTotalAcc(const float *data_in, const int in_size, float *data
 
     float total = 0;
     for (int i = 0; i < in_size; i++) {
-        total += fabs(data_in[i]);
+        total += fabsf(data_in[i]);
     }
     *data_out = total;
 
@@ -186,7 +188,7 @@ MldpReturn_t filterRms(const float *data_in, const int in_size, float *data_out,
     for (int i = 0; i < in_size; i++) {
         rms += data_in[i] * data_in[i];
     }
-    *data_out = sqrt(rms / in_size);
+    *data_out = sqrtf(rms / in_size);
 
     return MLDP_SUCCESS;
 }

testextension.cpp

@@ -18,6 +18,40 @@
 #define DEBUG_PRINT(...)
 #endif
 
+
+static inline void start_ticks_cpu() {
+    CoreDebug->DEMCR |= CoreDebug_DEMCR_TRCENA_Msk;
+    DWT->CYCCNT = 0;
+    DWT->CTRL |= DWT_CTRL_CYCCNTENA_Msk;
+}
+
+static inline uint32_t ticks_cpu() {
+    return DWT->CYCCNT;
+}
+
+static uint32_t ticks[10];
+static uint32_t ticks_index = 0;
+static bool ticks_start_average = false;
+static inline uint32_t calcTicks(uint32_t ticks_start, uint32_t ticks_end) {
+    ticks[ticks_index] = ticks_end - ticks_start;
+    ticks_index++;
+    if (ticks_index >= 9) {
+        ticks_start_average = true;
+        ticks_index = 0;
+    }
+
+    if (!ticks_start_average) {
+        return 0;
+    }
+
+    uint32_t ticksAverage = 0;
+    for (size_t i = 0; i < 10; i++) {
+        ticksAverage += ticks[i];
+    }
+    return ticksAverage / 10;
+}
+
+
 namespace testrunner {
     static ml_actions_t *actions = NULL;
     static ml_predictions_t *predictions = NULL;
@@ -43,7 +77,9 @@ namespace testrunner {
 
         unsigned int time_start = system_timer_current_time_us();
 
+        int32_t ticks_start = ticks_cpu() & 0x7FFFFFFF;
         float *modelData = mlDataProcessor.getProcessedData();
+        int32_t ticks_end = ticks_cpu() & 0x7FFFFFFF;
         if (modelData == NULL) {
             DEBUG_PRINT("Failed to processed data for the model\n");
             uBit.panic(TEST_RUNNER_ERROR + 21);
@@ -60,8 +96,8 @@ namespace testrunner {
 
         unsigned int time_end = system_timer_current_time_us();
 
-        DEBUG_PRINT("Prediction (%d micros + %d micros): ",
-                    time_mid - time_start, time_end - time_mid);
+        DEBUG_PRINT("Prediction (%d micros + %d micros, %d ticks): ",
+                    time_mid - time_start, time_end - time_mid, calcTicks(ticks_start, ticks_end));
         if (predictions->index >= 0) {
             DEBUG_PRINT("%d %s\n",
                         predictions->index,
@@ -218,6 +254,8 @@ namespace testrunner {
         uBit.messageBus.listen(TEST_RUNNER_ID_TIMER, ML_CODAL_TIMER_VALUE, &recordAccData, MESSAGE_BUS_LISTENER_DROP_IF_BUSY);
         uBit.timer.eventEvery(samplesPeriodMillisec, TEST_RUNNER_ID_TIMER, ML_CODAL_TIMER_VALUE);
 
+        start_ticks_cpu();
+
         initialised = true;
 
         DEBUG_PRINT("\tModel loaded\n");