2 files changed, 35 insertions, 29 deletions

diff --git a/embassy-stm32/src/dma/dma_bdma.rs b/embassy-stm32/src/dma/dma_bdma.rs
index 1945c3587..8b4b454c0 100644
--- a/embassy-stm32/src/dma/dma_bdma.rs
+++ b/embassy-stm32/src/dma/dma_bdma.rs
@@ -340,7 +340,8 @@ impl AnyChannel {
         mem_addr: *mut u32,
         mem_len: usize,
         incr_mem: bool,
-        data_size: WordSize,
+        mem_size: WordSize,
+        peripheral_size: WordSize,
         options: TransferOptions,
     ) {
         let info = self.info();
@@ -380,8 +381,8 @@ impl AnyChannel {
                 });
                 ch.cr().write(|w| {
                     w.set_dir(dir.into());
-                    w.set_msize(data_size.into());
-                    w.set_psize(data_size.into());
+                    w.set_msize(mem_size.into());
+                    w.set_psize(peripheral_size.into());
                     w.set_pl(options.priority.into());
                     w.set_minc(incr_mem);
                     w.set_pinc(false);
@@ -414,8 +415,8 @@ impl AnyChannel {
                 ch.mar().write_value(mem_addr as u32);
                 ch.ndtr().write(|w| w.set_ndt(mem_len as u16));
                 ch.cr().write(|w| {
-                    w.set_psize(data_size.into());
-                    w.set_msize(data_size.into());
+                    w.set_psize(peripheral_size.into());
+                    w.set_msize(mem_size.into());
                     w.set_minc(incr_mem);
                     w.set_dir(dir.into());
                     w.set_teie(true);
@@ -602,27 +603,28 @@ impl<'a> Transfer<'a> {
             buf.len(),
             true,
             W::size(),
+            W::size(),
             options,
         )
     }
 
     /// Create a new write DMA transfer (memory to peripheral).
-    pub unsafe fn new_write<W: Word>(
+    pub unsafe fn new_write<MW: Word, PW: Word>(
         channel: impl Peripheral<P = impl Channel> + 'a,
         request: Request,
-        buf: &'a [W],
-        peri_addr: *mut W,
+        buf: &'a [MW],
+        peri_addr: *mut PW,
         options: TransferOptions,
     ) -> Self {
         Self::new_write_raw(channel, request, buf, peri_addr, options)
     }
 
     /// Create a new write DMA transfer (memory to peripheral), using raw pointers.
-    pub unsafe fn new_write_raw<W: Word>(
+    pub unsafe fn new_write_raw<W: Word, PW: Word>(
         channel: impl Peripheral<P = impl Channel> + 'a,
         request: Request,
         buf: *const [W],
-        peri_addr: *mut W,
+        peri_addr: *mut PW,
         options: TransferOptions,
     ) -> Self {
         into_ref!(channel);
@@ -636,6 +638,7 @@ impl<'a> Transfer<'a> {
             buf.len(),
             true,
             W::size(),
+            W::size(),
             options,
         )
     }
@@ -660,6 +663,7 @@ impl<'a> Transfer<'a> {
             count,
             false,
             W::size(),
+            W::size(),
             options,
         )
     }
@@ -673,15 +677,23 @@ impl<'a> Transfer<'a> {
         mem_len: usize,
         incr_mem: bool,
         data_size: WordSize,
+        peripheral_size: WordSize,
         options: TransferOptions,
     ) -> Self {
         assert!(mem_len > 0 && mem_len <= 0xFFFF);
 
         channel.configure(
-            _request, dir, peri_addr, mem_addr, mem_len, incr_mem, data_size, options,
+            _request,
+            dir,
+            peri_addr,
+            mem_addr,
+            mem_len,
+            incr_mem,
+            data_size,
+            peripheral_size,
+            options,
         );
         channel.start();
-
         Self { channel }
     }
 
@@ -814,6 +826,7 @@ impl<'a, W: Word> ReadableRingBuffer<'a, W> {
             len,
             true,
             data_size,
+            data_size,
             options,
         );
 
@@ -966,6 +979,7 @@ impl<'a, W: Word> WritableRingBuffer<'a, W> {
             len,
             true,
             data_size,
+            data_size,
             options,
         );
 
diff --git a/embassy-stm32/src/timer/simple_pwm.rs b/embassy-stm32/src/timer/simple_pwm.rs
index f36fa026c..8d3c9a131 100644
--- a/embassy-stm32/src/timer/simple_pwm.rs
+++ b/embassy-stm32/src/timer/simple_pwm.rs
@@ -334,7 +334,7 @@ impl<'d, T: GeneralInstance4Channel> SimplePwm<'d, T> {
                 &mut dma,
                 req,
                 duty,
-                self.inner.regs_1ch().ccr(channel.index()).as_ptr() as *mut _,
+                self.inner.regs_1ch().ccr(channel.index()).as_ptr() as *mut u16,
                 dma_transfer_option,
             )
             .await
@@ -362,13 +362,7 @@ macro_rules! impl_waveform_chx {
     ($fn_name:ident, $dma_ch:ident, $cc_ch:ident) => {
         impl<'d, T: GeneralInstance4Channel> SimplePwm<'d, T> {
             /// Generate a sequence of PWM waveform
-            ///
-            /// Note:
-            /// 1. you will need to provide corresponding TIMx_CHy DMA channel to use this method.
-            /// 2. Please make sure the duty data length is aligned to the timer data width(16-bit or 32-bit).
-            /// 3. Please notice the endianess of the duty data. STM32 use little endian,
-            ///    for example, 0x12345678 as u32 will be stored as [0x78, 0x56, 0x34, 0x12] in memory.
-            pub async fn $fn_name(&mut self, dma: impl Peripheral<P = impl super::$dma_ch<T>>, duty: &[u8]) {
+            pub async fn $fn_name(&mut self, dma: impl Peripheral<P = impl super::$dma_ch<T>>, duty: &[u16]) {
                 use crate::pac::timer::vals::Ccds;
 
                 into_ref!(dma);
@@ -411,32 +405,30 @@ macro_rules! impl_waveform_chx {
 
                     match self.inner.bits() {
                         TimerBits::Bits16 => {
-                            // the data must be aligned to double words
-                            assert!(duty.len() % 2 == 0);
-                            let duty = core::slice::from_raw_parts(duty.as_ptr() as *const u16, duty.len() / 2);
                             Transfer::new_write(
                                 &mut dma,
                                 req,
                                 duty,
-                                self.inner.regs_gp16().ccr(cc_channel.index()).as_ptr() as *mut _,
+                                self.inner.regs_gp16().ccr(cc_channel.index()).as_ptr() as *mut u16,
                                 dma_transfer_option,
                             )
                             .await
                         }
-                        #[cfg(not(stm32l0))]
+                        #[cfg(not(any(stm32l0, bdma, gpdma)))]
                         TimerBits::Bits32 => {
-                            // the data must be aligned to quad words
-                            assert!(duty.len() % 4 == 0);
-                            let duty = core::slice::from_raw_parts(duty.as_ptr() as *const u32, duty.len() / 4);
                             Transfer::new_write(
                                 &mut dma,
                                 req,
                                 duty,
-                                self.inner.regs_gp16().ccr(cc_channel.index()).as_ptr() as *mut _,
+                                self.inner.regs_gp16().ccr(cc_channel.index()).as_ptr() as *mut u32,
                                 dma_transfer_option,
                             )
                             .await
                         }
+                        #[cfg(any(stm32l0, bdma, gpdma))]
+                        _ => {
+                            panic!("unsupported timer bits")
+                        }
                     };
                 };