From 5687196bdd2d9122086fb9749e31824c5779040a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ond=C5=99ej=20Hru=C5=A1ka?= <ondra@ondrovo.com>
Date: Mon, 5 Mar 2018 10:21:23 +0100
Subject: [PATCH] small speed up using larger buffer and multipart Tx

---
 TinyFrame/TF_Integration.c | 44 ++++++++++++++++++++++++++++++++++++++
 freertos.c                 |  2 +-
 platform/plat_compat.h     |  4 ++--
 units/adc/_adc_core.c      | 12 ++++++-----
 4 files changed, 54 insertions(+), 8 deletions(-)

diff --git a/TinyFrame/TF_Integration.c b/TinyFrame/TF_Integration.c
index 6aa1b18..22e0b9f 100644
--- a/TinyFrame/TF_Integration.c
+++ b/TinyFrame/TF_Integration.c
@@ -14,8 +14,40 @@
 extern osSemaphoreId semVcomTxReadyHandle;
 extern osMutexId mutTinyFrameTxHandle;
 
+static volatile bool first_tx = false; // XXX global
+
 void TF_WriteImpl(TinyFrame *tf, const uint8_t *buff, uint32_t len)
 {
+#if 1
+//    if (!first_tx) {
+//        // wait for the last USB transmission to be finished
+//        int32_t mxStatus;
+//        mxStatus = osSemaphoreWait(semVcomTxReadyHandle, 100);
+//        if (mxStatus != osOK) {
+//            TF_Error("Tx stalled");
+//            return;
+//        }
+//    }
+//    first_tx = false;
+
+    // Padding to a multiple of 64 bytes
+    if (len&0x3F) {
+        uint32_t pad = (64 - (len&0x3F));
+        memset((void *) (buff + len), 0, pad);
+        len += pad; // padding to a multiple of 64 (size of the endpoint)
+    }
+
+    assert_param(HAL_OK == HAL_PCD_EP_Transmit(hUsbDeviceFS.pData, CDC_IN_EP, (uint8_t *) buff, len));
+
+    // Wait for the semaphore - HAL keeps a pointer to the buffer, and it's the TinyFrame Tx buffer,
+    // so if we let it process it in the background, it could get corrupted before the Tx is completed.
+    int32_t mxStatus;
+    mxStatus = osSemaphoreWait(semVcomTxReadyHandle, 100);
+    if (mxStatus != osOK) {
+        TF_Error("Tx stalled");
+        return;
+    }
+#else
     (void) tf;
 #define CHUNK 64 // same as TF_SENDBUF_LEN, so we should always have only one run of the loop
     int32_t total = (int32_t) len;
@@ -37,6 +69,7 @@ void TF_WriteImpl(TinyFrame *tf, const uint8_t *buff, uint32_t len)
         buff += chunksize;
         total -= chunksize;
     }
+#endif
 }
 
 /** Claim the TX interface before composing and sending a frame */
@@ -47,6 +80,17 @@ bool TF_ClaimTx(TinyFrame *tf)
     assert_param(!inIRQ());
 
     assert_param(osOK == osMutexWait(mutTinyFrameTxHandle, 5000));
+
+//    // wait for the last USB transmission to be finished
+//    int32_t mxStatus;
+//    mxStatus = osSemaphoreWait(semVcomTxReadyHandle, 100);
+//    if (mxStatus != osOK) {
+//        TF_Error("Tx stalled");
+//        return false;
+//    }
+
+    first_tx = true;
+
     return true;
 }
 
diff --git a/freertos.c b/freertos.c
index 6213905..10cebcf 100644
--- a/freertos.c
+++ b/freertos.c
@@ -169,7 +169,7 @@ void MX_FREERTOS_Init(void) {
 
   /* USER CODE BEGIN RTOS_SEMAPHORES */
   /* add semaphores, ... */
-  xSemaphoreGive(semVcomTxReadyHandle);
+//  xSemaphoreGive(semVcomTxReadyHandle);
   /* USER CODE END RTOS_SEMAPHORES */
 
   /* USER CODE BEGIN RTOS_TIMERS */
diff --git a/platform/plat_compat.h b/platform/plat_compat.h
index 74f1025..585aeb3 100644
--- a/platform/plat_compat.h
+++ b/platform/plat_compat.h
@@ -30,7 +30,7 @@
 
 
 #define BULK_READ_BUF_LEN 256   // Buffer for TF bulk reads
-#define UNIT_TMP_LEN      512   // Buffer for internal unit operations
+#define UNIT_TMP_LEN      256   // Buffer for internal unit operations
 
 #define FLASH_SAVE_BUF_LEN  128 // Malloc'd buffer for saving to flash
 
@@ -38,7 +38,7 @@
 #define RX_QUE_CAPACITY    16 // TinyFrame rx queue size (64 bytes each)
 
 #define TF_MAX_PAYLOAD_RX 512 // TF max Rx payload
-#define TF_SENDBUF_LEN     64 // TF transmit buffer (can be less than a full frame)
+#define TF_SENDBUF_LEN    512 // TF transmit buffer (can be less than a full frame)
 
 #define TF_MAX_ID_LST   4 // Frame ID listener count
 #define TF_MAX_TYPE_LST 6 // Frame Type listener count
diff --git a/units/adc/_adc_core.c b/units/adc/_adc_core.c
index 8a59d8e..023de49 100644
--- a/units/adc/_adc_core.c
+++ b/units/adc/_adc_core.c
@@ -171,7 +171,7 @@ static void handle_httc(Unit *unit, bool tc)
     const bool m_fixcpt = priv->opmode == ADC_OPMODE_BLCAP;
 
     if (ht) {
-        end = (priv->buf_itemcount / 2);
+        end = (priv->buf_itemcount >> 1); // div2
     }
     else {
         end = priv->buf_itemcount;
@@ -234,7 +234,7 @@ static void handle_httc(Unit *unit, bool tc)
         priv->stream_startpos = 0;
     }
     else {
-        priv->stream_startpos = priv->buf_itemcount / 2;
+        priv->stream_startpos = priv->buf_itemcount >> 1; // div2
     }
 }
 
@@ -289,7 +289,7 @@ void UADC_DMA_Handler(void *arg)
         const bool m_stream = priv->opmode == ADC_OPMODE_STREAM;
         const bool m_fixcpt = priv->opmode == ADC_OPMODE_BLCAP;
         if (m_trigd || m_stream || m_fixcpt) {
-            const uint32_t half = (uint32_t) (priv->buf_itemcount / 2);
+            const uint32_t half = (uint32_t) (priv->buf_itemcount >> 1); // div2
             if (ht && tc) {
                 // dual event interrupt - may happen if we missed both and they were pending after
                 // interrupts became enabled again (this can happen due to the EOS or other higher prio irq's)
@@ -336,8 +336,10 @@ void UADC_ADC_EOS_Handler(void *arg)
     if (priv->opmode == ADC_OPMODE_UNINIT) return;
 
     // Wait for the DMA to complete copying the last sample
-    uint32_t dmapos;
-    hw_wait_while((dmapos = DMA_POS(priv)) % priv->nb_channels != 0, 100); // XXX this could be changed to reading it from the DR instead
+    uint32_t dmapos = DMA_POS(priv);
+    if ((DMA_POS(priv) % priv->nb_channels) != 0) {
+        hw_wait_while((dmapos = DMA_POS(priv)) % priv->nb_channels != 0, 100); // XXX this could be changed to reading it from the DR instead
+    }
 
     uint32_t sample_pos;
     if (dmapos == 0) {