lm4f: Improve read/write endpoint performance

Modify lm4f_ep_read/write_packet() to take advantage of 32-bit and 16-bit
accesses to the USB FIFO, as opposed to using only 8-bit accesses. This
change increases endpoint performance in a number of configurations.

On the host side, we use usb_bulk_bench to test
https://github.com/mrnuke/usb_bulk_bench

With the following invocations:

usb_bulk_bench -d c03e:b007 -a -t 64 -q 32 -e [ep] [-I/-O]

On the device side, we use the usb_bulk_dev example:

https://github.com/mrnuke/libopencm3-examples/tree/stellaris
The example is in: examples/lm4f/stellaris-ek-lm4f120xl/usb_bulk_dev

(This example will be available in libopencm3-examples in the near future)

The endpoints configuration is the following:
* EP1 OUT - interrupt driven RX endpoint
* EP2 IN  - interrupt driven TX endpoint
* EP3 OUT - polled RX endpoint
* EP4 IN  - polled TX endpoint
* EP5 OUT - polled RX endpoint with unaligned buffer
* EP6 IN  - polled TX endpoint with unaligned buffer

We test the speed in each configuration, using different system clock
frequencies. We run the tests once without the patch applied, and once with
the patch applied. The results are given below:

Before patch (numbers in KiB/s):
freq:	80 MHz	57 MHz	40 MHz	30 MHz	20 MHz	16 MHz
EP1	562	562	562	562	562	550
EP2	936	872	812	812	687	625
EP3	1062	890	700	600	562	562
EP4	900	812	812	750	625	562
EP5	1062	890	700	600	562	562
EP6	930	812	812	750	625	562

With patch  (numbers in KiB/s):
freq:	80 MHz	57 MHz	40 MHz	30 MHz	20 MHz	16 MHz
EP1	1062	1062	1062	690	562	562
EP2	1125	936	936	936	870	812
EP3	1062	960	750	750	562	562
EP4	936	936	870	870	770	700
EP5	1062	900	700	630	562	562
EP6	930	930	870	870	740	650

Percent change in speed (*):
freq:	80 MHz	57 MHz	40 MHz	30 MHz	20 MHz	16 MHz
EP1	89.0	89.0	89.0	22.8	0.0	2.2
EP2	20.2	7.3	15.3	15.3	26.6	29.9
EP3	0.0	7.9	7.1	25.0	0.0	0.0
EP4	4.0	15.3	7.1	16.0	23.2	24.6
EP5	0.0	1.1	0.0	5.0	0.0	0.0
EP6	0.0	14.5	7.1	16.0	18.4	15.7

(*) Numbers given as percent change relative to speed before applying this
patch.

We see throughput increases across the board.

Signed-off-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>

lib/lm4f/usb_lm4f.c

@@ -364,12 +364,18 @@ static u16 lm4f_ep_write_packet(usbd_device *usbd_dev, u8 addr,
 	}
 
 	/*
-	 * For some reason, using 16 or 32-bit transfers to the FIFO does not
+	 * We don't need to worry about buf not being aligned. If it's not,
-	 * work well.
+	 * the reads are downgraded to 8-bit in hardware. We lose a bit of
+	 * performance, but we don't crash.
 	 */
-	for (i = 0; i < len; i++)
+	for (i = 0; i < (len & ~0x3); i += 4)
-		USB_FIFO8(ep) = ((u8 *)buf)[i];
+		USB_FIFO32(ep) = *((u32 *)(buf + i));
-
+	if (len & 0x2) {
+		USB_FIFO16(ep) = *((u16 *)(buf + i));
+		i += 2;
+	}
+	if (len & 0x1)
+		USB_FIFO8(ep)  = *((u8 *)(buf + i));
 
 	if (ep == 0) {
 		/*
@@ -394,7 +400,6 @@ static u16 lm4f_ep_read_packet(usbd_device *usbd_dev, u8 addr, void *buf, u16 le
 {
 	(void)usbd_dev;
 
-	u8 * buffy = buf;
 	u16 rlen;
 	u8 ep = addr & 0xf;
 
@@ -402,8 +407,19 @@ static u16 lm4f_ep_read_packet(usbd_device *usbd_dev, u8 addr, void *buf, u16 le
 
 	rlen = (fifoin > len) ? len : fifoin;
 
-	for (len = 0; len < rlen; len++)
+	/*
-		buffy[len] = USB_FIFO8(ep);
+	 * We don't need to worry about buf not being aligned. If it's not,
+	 * the writes are downgraded to 8-bit in hardware. We lose a bit of
+	 * performance, but we don't crash.
+	 */
+	for (len = 0; len < (rlen & ~0x3); len += 4)
+		*((u32 *)(buf + len)) = USB_FIFO32(ep);
+	if (rlen & 0x2) {
+		*((u16 *)(buf + len)) = USB_FIFO16(ep);
+		len += 2;
+	}
+	if (rlen & 0x1)
+		*((u8 *)(buf + len)) = USB_FIFO8(ep);
 
 	if (ep == 0) {
 		/*