From 8f6aa50caa0de4b551d83629a6eed378a0451ec7 Mon Sep 17 00:00:00 2001
From: Clyne Sullivan <clyne@bitgloo.com>
Date: Fri, 31 Jan 2025 19:46:56 -0500
Subject: [PATCH] reduce process cycles; fix sample bit-order

---
 Core/Src/main.c | 73 +++++++++++++++++++++++++++----------------------
 Makefile        |  2 +-
 2 files changed, 42 insertions(+), 33 deletions(-)

diff --git a/Core/Src/main.c b/Core/Src/main.c
index 67e6ab8..7c41e3f 100644
--- a/Core/Src/main.c
+++ b/Core/Src/main.c
@@ -27,19 +27,17 @@
 
 /* Private typedef -----------------------------------------------------------*/
 /* USER CODE BEGIN PTD */
-typedef struct  {
-    int32_t value : 18;
-} sample_t;
+typedef uint32_t sample_t;
 /* USER CODE END PTD */
 
 /* Private define ------------------------------------------------------------*/
 /* USER CODE BEGIN PD */
-#define SAMPLE_COUNT        (256)   // Sample count per half-transfer
+#define SAMPLE_COUNT        (256)   // (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
 #define MIC_REF_DB          ( 94.f) // dB where sensitivity is specified
-#define MIC_BITS            (18)
+#define MIC_BITS            (18u)
 /* USER CODE END PD */
 
 /* Private macro -------------------------------------------------------------*/
@@ -104,7 +102,7 @@ int main(void)
 
   /* USER CODE BEGIN 1 */
   ln10 = qfp_fln(10.f);
-  MIC_REF_AMPL = qfp_fmul(qfp_int2float((1 << (MIC_BITS - 1)) - 1),
+  MIC_REF_AMPL = qfp_fmul(qfp_int2float((1u << (MIC_BITS - 1)) - 1),
     qfp_fpow(10.f, MIC_SENSITIVITY / 20.f));
   /* USER CODE END 1 */
 
@@ -443,53 +441,64 @@ error :
   return errorcode;
 }
 
-void SPI_DMATransmitReceiveCplt(DMA_HandleTypeDef *hdma)
-{
-  SPI_HandleTypeDef *hspi = (SPI_HandleTypeDef *)hdma->Parent;
-  (void)hspi;
-}
-
-static float process(float in_div4)
+static inline float process(float in_div4)
 {
-    static float z1 = 0, z2 = 0, z3 = 0, z5 = 0;
+    static float z[4] = {0.f, 0.f, 0.f, 0.f};
 
-    float out1 = qfp_fadd(in_div4, z1);
-    z1 = qfp_fadd(qfp_fmul(1.062f, out1), z2);
-    z2 = qfp_fsub(qfp_fmul(-0.14f, out1), in_div4);
+    float out1 = qfp_fadd(in_div4, z[0]);
+    z[0] = qfp_fadd(qfp_fmul(out1, 1.062f), z[1]);
+    z[1] = qfp_fsub(qfp_fmul(out1, -0.14f), in_div4);
 
-    float out2 = qfp_fadd(out1, z3);
-    z3 = out1;
+    float out2 = qfp_fadd(out1, z[2]);
+    z[2] = out1;
 
-    float out3 = qfp_fadd(out2, z5);
-    z5 = qfp_fsub(qfp_fmul(0.985f, out3), out2);
+    float out3 = qfp_fadd(out2, z[3]);
+    z[3] = qfp_fsub(qfp_fmul(out3, 0.985f), out2);
 
     return out3;
 }
 
-void SPI_DMAHalfTransmitReceiveCplt(DMA_HandleTypeDef *hdma)
+static void processSampleBlock(sample_t *sample)
 {
   HAL_GPIO_WritePin(IDLE_GPIO_Port, IDLE_Pin, GPIO_PIN_RESET);
 
-  SPI_HandleTypeDef *hspi = (SPI_HandleTypeDef *)hdma->Parent;
-  (void)hspi;
+  // SAMPLE_COUNT is number of samples in this block
+  // Divide by 2 for left channel only
+  // Divide by 8 for stride to reduce compute time
+#define STRIDE (2 * 8)
+
+  for (int i = 0; i < SAMPLE_COUNT; i += STRIDE) {
+    // 18-bit sample comes in as big-endian with right padding.
+    // 1. Convert to little-endian:
+    int samp;
+    asm("rev %0, %1" : "=r" (samp) : "r" (sample[i]));
+    // 2. Arithmetic shift right to remove padding, +2 to make "in_div4"
+    float f = process(qfp_int2float(samp >> (32 - MIC_BITS + 2)));
 
-  sample_t *sample = (sample_t *)I2S_Receive_Buffer;
-  for (int i = 0; i < SAMPLE_COUNT / 8; i++) {
-    float f = process(qfp_int2float(sample[i].value / 4));
-    DB_Sum_Squares = qfp_fadd(DB_Sum_Squares, qfp_fmul(f, f));
+    DB_Sum_Squares = qfp_fadd(qfp_fmul(f, f), DB_Sum_Squares);
   }
-  DB_Count += SAMPLE_COUNT * 2; // Pretend we sampled the entire I2S buffer
+  DB_Count += SAMPLE_COUNT / STRIDE;
 
-  if (DB_Count > SAMPLES_PER_REPORT) {
+  if (DB_Count >= SAMPLES_PER_REPORT / STRIDE * 2) {
     float rms = qfp_fsqrt(qfp_fdiv(DB_Sum_Squares, qfp_uint2float(DB_Count)));
-    float db = qfp_fadd(MIC_OFFSET_DB + MIC_REF_DB, qfp_fmul(20.f,
-      qfp_flog10(qfp_fdiv(rms, MIC_REF_AMPL))));
+    float db = qfp_fadd(qfp_fmul(qfp_flog10(qfp_fdiv(rms, MIC_REF_AMPL)), 20.f),
+        MIC_OFFSET_DB + MIC_REF_DB);
     DB_Sum_Squares = 0.f;
     DB_Count = 0;
 
     printf("%d dB\r\n", qfp_float2int(db));
   }
 }
+
+void SPI_DMATransmitReceiveCplt(DMA_HandleTypeDef *hdma)
+{
+  processSampleBlock((sample_t *)I2S_Receive_Buffer + SAMPLE_COUNT);
+}
+
+void SPI_DMAHalfTransmitReceiveCplt(DMA_HandleTypeDef *hdma)
+{
+  processSampleBlock((sample_t *)I2S_Receive_Buffer);
+}
 /* USER CODE END 4 */
 
 /**
diff --git a/Makefile b/Makefile
index ab0211a..b26e4da 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 ##########################################################################################################################
-# File automatically-generated by tool: [projectgenerator] version: [4.5.0-RC5] date: [Thu Jan 30 17:24:56 EST 2025] 
+# File automatically-generated by tool: [projectgenerator] version: [4.5.0-RC5] date: [Fri Jan 31 17:56:51 EST 2025] 
 ##########################################################################################################################
 
 # ------------------------------------------------