From d09f4289b5788d6a8b34e424841292e2b8529e56 Mon Sep 17 00:00:00 2001
From: Clyne Sullivan <clyne@bitgloo.com>
Date: Sat, 1 Feb 2025 13:31:20 -0500
Subject: exec from ram; process all samples

---
 Core/Src/main.c | 78 ++++++++++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 61 insertions(+), 17 deletions(-)

(limited to 'Core/Src')

diff --git a/Core/Src/main.c b/Core/Src/main.c
index 45d16a6..ed49703 100644
--- a/Core/Src/main.c
+++ b/Core/Src/main.c
@@ -32,7 +32,7 @@ typedef uint32_t sample_t;
 
 /* Private define ------------------------------------------------------------*/
 /* USER CODE BEGIN PD */
-#define SAMPLE_COUNT        (256)   // (Stereo!) Sample count per half-transfer
+#define SAMPLE_COUNT        (1024)   // (Stereo!) Sample count per half-transfer
 #define SAMPLES_PER_REPORT  (48000) // Report every second given our 48 kHz rate
 #define MIC_OFFSET_DB       (  0.f) // Linear offset
 #define MIC_SENSITIVITY     (-26.f) // dBFS value expected at MIC_REF_DB
@@ -79,7 +79,9 @@ static HAL_StatusTypeDef HAL_SPI_TransmitReceive_DMA_Mixed(
     uint8_t *pRxData,
     uint16_t TxSize,
     uint16_t RxSize);
+__RAM_FUNC
 static void SPI_DMATransmitReceiveCplt(DMA_HandleTypeDef *hdma);
+__RAM_FUNC
 static void SPI_DMAHalfTransmitReceiveCplt(DMA_HandleTypeDef *hdma);
 /* USER CODE END PFP */
 
@@ -91,6 +93,14 @@ int __io_putchar(int ch)
     HAL_UART_Transmit(&huart2, &buf, sizeof(buf), HAL_TIMEOUT);
     return buf;
 }
+
+__RAM_FUNC
+void fvar_init(void)
+{
+  ln10 = qfp_fln(10.f);
+  MIC_REF_AMPL = qfp_fmul(qfp_int2float((1u << (MIC_BITS - 2)) - 1),
+    qfp_fpow(10.f, MIC_SENSITIVITY / 20.f));
+}
 /* USER CODE END 0 */
 
 /**
@@ -101,9 +111,7 @@ int main(void)
 {
 
   /* USER CODE BEGIN 1 */
-  ln10 = qfp_fln(10.f);
-  MIC_REF_AMPL = qfp_fmul(qfp_int2float((1u << (MIC_BITS - 1)) - 1),
-    qfp_fpow(10.f, MIC_SENSITIVITY / 20.f));
+  fvar_init();
   /* USER CODE END 1 */
 
   /* MCU Configuration--------------------------------------------------------*/
@@ -160,7 +168,7 @@ void SystemClock_Config(void)
 
   /** Configure the main internal regulator output voltage
   */
-  HAL_PWREx_ControlVoltageScaling(PWR_REGULATOR_VOLTAGE_SCALE2);
+  HAL_PWREx_ControlVoltageScaling(PWR_REGULATOR_VOLTAGE_SCALE1);
 
   /** Initializes the RCC Oscillators according to the specified parameters
   * in the RCC_OscInitTypeDef structure.
@@ -180,10 +188,10 @@ void SystemClock_Config(void)
   RCC_ClkInitStruct.ClockType = RCC_CLOCKTYPE_HCLK|RCC_CLOCKTYPE_SYSCLK
                               |RCC_CLOCKTYPE_PCLK1;
   RCC_ClkInitStruct.SYSCLKSource = RCC_SYSCLKSOURCE_MSI;
-  RCC_ClkInitStruct.AHBCLKDivider = RCC_SYSCLK_DIV2;
-  RCC_ClkInitStruct.APB1CLKDivider = RCC_HCLK_DIV2;
+  RCC_ClkInitStruct.AHBCLKDivider = RCC_SYSCLK_DIV1;
+  RCC_ClkInitStruct.APB1CLKDivider = RCC_HCLK_DIV4;
 
-  if (HAL_RCC_ClockConfig(&RCC_ClkInitStruct, FLASH_LATENCY_1) != HAL_OK)
+  if (HAL_RCC_ClockConfig(&RCC_ClkInitStruct, FLASH_LATENCY_0) != HAL_OK)
   {
     Error_Handler();
   }
@@ -323,7 +331,7 @@ static void MX_GPIO_Init(void)
   HAL_GPIO_WritePin(LED_RED_GPIO_Port, LED_RED_Pin, GPIO_PIN_RESET);
 
   /*Configure GPIO pin Output Level */
-  HAL_GPIO_WritePin(IDLE_GPIO_Port, IDLE_Pin, GPIO_PIN_RESET);
+  HAL_GPIO_WritePin(IDLE_GPIO_Port, IDLE_Pin, GPIO_PIN_SET);
 
   /*Configure GPIO pin : LED_GREEN_Pin */
   GPIO_InitStruct.Pin = LED_GREEN_Pin;
@@ -443,6 +451,43 @@ error :
 
 #define BITO (10)
 
+__RAM_FUNC
+__attribute__((naked))
+int64_t lmul(int64_t a, int64_t b)
+{
+    asm(
+"	push	{r4, lr}\n"
+"	mul	r1, r2\n"
+"	mul	r3, r0\n"
+"	add	r1, r3\n"
+
+"	lsr	r3, r0, #16\n"
+"	lsr	r4, r2, #16\n"
+"	mul	r3, r4\n"
+"	add	r1, r3\n"
+
+"	lsr	r3, r0, #16\n"
+"	uxth	r0, r0\n"
+"	uxth	r2, r2\n"
+"	mul	r3, r2\n"
+"	mul	r4, r0\n"
+"	mul	r0, r2\n"
+
+"	mov	r2, #0\n"
+"	add	r3, r4\n"
+"	adc	r2, r2\n"
+"	lsl	r2, #16\n"
+"	add	r1, r2\n"
+
+"	lsl	r2, r3, #16\n"
+"	lsr	r3, #16\n"
+"	add	r0, r2\n"
+"	adc	r1, r3\n"
+"	pop	{r4, pc}\n"
+    );
+}
+
+__RAM_FUNC
 static inline void process(int64_t in_div4)
 {
     static int64_t z[4] = {0, 0, 0, 0};
@@ -458,25 +503,24 @@ static inline void process(int64_t in_div4)
     int64_t out3 = (out2 + z[3]);
     z[3] = ((out3 * 0x3f1 /*0.985f*/) >> BITO) - out2;
 
-    DB_Sum_Squares += (out3 * out3) >> BITO;
+    DB_Sum_Squares += lmul(out3, out3) >> BITO;
 }
 
+__RAM_FUNC
 static void processSampleBlock(sample_t *sample)
 {
-  HAL_GPIO_WritePin(IDLE_GPIO_Port, IDLE_Pin, GPIO_PIN_RESET);
-
-#define STRIDE 2
+  IDLE_GPIO_Port->ODR ^= IDLE_Pin;
 
-  for (int i = 0; i < SAMPLE_COUNT; i += 2 * STRIDE) {
+  for (int i = 0; i < SAMPLE_COUNT; i += 2) {
     // 18-bit sample comes in as big-endian with right padding.
     // Use REVSH to extract 18-bit reading divided by four for process().
     int samp;
     asm("revsh %0, %1" : "=l" (samp) : "l" (sample[i]));
     process(samp);
   }
-  DB_Count += SAMPLE_COUNT / (2 * STRIDE);
+  DB_Count += SAMPLE_COUNT / 2;
 
-  if (DB_Count >= SAMPLES_PER_REPORT / STRIDE) {
+  if (DB_Count >= SAMPLES_PER_REPORT) {
     float rms = qfp_fsqrt(qfp_int2float((DB_Sum_Squares >> BITO) / DB_Count));
     float db = qfp_fadd(qfp_fmul(qfp_flog10(qfp_fdiv(rms, MIC_REF_AMPL)), 20.f),
         MIC_OFFSET_DB + MIC_REF_DB);
@@ -486,7 +530,7 @@ static void processSampleBlock(sample_t *sample)
     printf("%d dB\r\n", qfp_float2int(db));
   }
 
-  HAL_GPIO_WritePin(IDLE_GPIO_Port, IDLE_Pin, GPIO_PIN_SET);
+  IDLE_GPIO_Port->ODR ^= IDLE_Pin;
 }
 
 void SPI_DMATransmitReceiveCplt(DMA_HandleTypeDef *hdma)
-- 
cgit v1.2.3