SSD2119: make use of DMA

This is mostly a copy from Eddie's work posted here:
http://forum.chibios.org/phpbb/viewtopic.php?f=11&t=851#p11054
No work was done towards making it work as fast as possible.
Tested with:
https://github.com/etmatrix/ChibiOS-GFX-Example/blob/master/bench/main.c
Results show performance of ~5.34 Mpx/s with use of DMA compared to
~4.78 Mpx/s without.

2 files changed, 134 insertions, 42 deletions

diff --git a/drivers/gdisp/SSD2119/gdisp_lld.c b/drivers/gdisp/SSD2119/gdisp_lld.c
index fe28ca45..d7e08504 100644
--- a/drivers/gdisp/SSD2119/gdisp_lld.c
+++ b/drivers/gdisp/SSD2119/gdisp_lld.c
@@ -240,13 +240,13 @@ bool_t gdisp_lld_init(void) {
 	write_reg(SSD2119_REG_Y_RAM_ADDR, 0x00);
 	delay(5);
 
- 	// Release the bus
+	// Release the bus
 	release_bus();
 
 	/* Turn on the backlight */
 	set_backlight(GDISP_INITIAL_BACKLIGHT);
 
-   /* Initialise the GDISP structure */
+	/* Initialise the GDISP structure */
 	GDISP.Width = GDISP_SCREEN_WIDTH;
 	GDISP.Height = GDISP_SCREEN_HEIGHT;
 	GDISP.Orientation = GDISP_ROTATE_0;
@@ -309,14 +309,33 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 	 * @notapi
 	 */
 	void gdisp_lld_clear(color_t color) {
-		unsigned i;
+		unsigned area;
+
+		area = GDISP_SCREEN_WIDTH * GDISP_SCREEN_HEIGHT;
 
 		acquire_bus();
 		reset_viewport();
 		set_cursor(0, 0);
 		stream_start();
-		for(i = 0; i < GDISP_SCREEN_WIDTH * GDISP_SCREEN_HEIGHT; i++)
-			write_data(color);
+
+		#if defined(GDISP_USE_DMA) && defined(GDISP_DMA_STREAM)
+			uint8_t i;
+			dmaStreamSetPeripheral(GDISP_DMA_STREAM, &color);
+			dmaStreamSetMode(GDISP_DMA_STREAM, STM32_DMA_CR_PL(0) | STM32_DMA_CR_PSIZE_HWORD | STM32_DMA_CR_MSIZE_HWORD | STM32_DMA_CR_DIR_M2M);
+			for (i = area / 65535; i; i--) {
+				dmaStreamSetTransactionSize(GDISP_DMA_STREAM, 65535);
+				dmaStreamEnable(GDISP_DMA_STREAM);
+				dmaWaitCompletion(GDISP_DMA_STREAM);
+			}
+			dmaStreamSetTransactionSize(GDISP_DMA_STREAM, area % 65535);
+			dmaStreamEnable(GDISP_DMA_STREAM);
+			dmaWaitCompletion(GDISP_DMA_STREAM);
+		#else
+			uint32_t index;
+			for(index = 0; index < area; index++)
+				write_data(color);
+		#endif // defined(GDISP_USE_DMA) && defined(GDISP_DMA_STREAM)
+
 		stream_stop();
 		release_bus();
 	}
@@ -334,7 +353,7 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 	 * @notapi
 	 */
 	void gdisp_lld_fill_area(coord_t x, coord_t y, coord_t cx, coord_t cy, color_t color) {
-		unsigned i, area;
+		unsigned area;
 
 		#if GDISP_NEED_VALIDATION || GDISP_NEED_CLIP
 			if (x < GDISP.clipx0) { cx -= GDISP.clipx0 - x; x = GDISP.clipx0; }
@@ -349,8 +368,25 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 		acquire_bus();
 		set_viewport(x, y, cx, cy);
 		stream_start();
-		for(i = 0; i < area; i++)
-			write_data(color);
+
+		#if defined(GDISP_USE_DMA) && defined(GDISP_DMA_STREAM)
+			uint8_t i;
+				dmaStreamSetPeripheral(GDISP_DMA_STREAM, &color);
+				dmaStreamSetMode(GDISP_DMA_STREAM, STM32_DMA_CR_PL(0) | STM32_DMA_CR_PSIZE_HWORD | STM32_DMA_CR_MSIZE_HWORD | STM32_DMA_CR_DIR_M2M);
+			for (i = area / 65535; i; i--) {
+				dmaStreamSetTransactionSize(GDISP_DMA_STREAM, 65535);
+				dmaStreamEnable(GDISP_DMA_STREAM);
+				dmaWaitCompletion(GDISP_DMA_STREAM);
+			}
+			dmaStreamSetTransactionSize(GDISP_DMA_STREAM, area % 65535);
+			dmaStreamEnable(GDISP_DMA_STREAM);
+			dmaWaitCompletion(GDISP_DMA_STREAM);
+		#else
+			uint32_t index;
+			for(index = 0; index < area; index++)
+				write_data(color);
+		#endif // defined(GDISP_USE_DMA) && defined(GDISP_DMA_STREAM)
+
 		stream_stop();
 		release_bus();
 	}
@@ -370,8 +406,6 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 	 * @notapi
 	 */
 	void gdisp_lld_blit_area_ex(coord_t x, coord_t y, coord_t cx, coord_t cy, coord_t srcx, coord_t srcy, coord_t srccx, const pixel_t *buffer) {
-		coord_t endx, endy;
-		unsigned lg;
 
 		#if GDISP_NEED_VALIDATION || GDISP_NEED_CLIP
 			if (x < GDISP.clipx0) { cx -= GDISP.clipx0 - x; srcx += GDISP.clipx0 - x; x = GDISP.clipx0; }
@@ -382,17 +416,36 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 			if (y+cy > GDISP.clipy1)	cy = GDISP.clipy1 - y;
 		#endif
 
+		buffer += srcx + srcy * srccx;
+
 		acquire_bus();
 		set_viewport(x, y, cx, cy);
 		stream_start();
 
-		endx = srcx + cx;
-		endy = y + cy;
-		lg = srccx - cx;
-		buffer += srcx + srcy * srccx;
-		for(; y < endy; y++, buffer += lg)
-			for(x=srcx; x < endx; x++)
-				write_data(*buffer++);
+		#if defined(GDISP_USE_DMA) && defined(GDISP_DMA_STREAM)
+			uint32_t area = cx * cy;
+			uint8_t i;
+			dmaStreamSetPeripheral(GDISP_DMA_STREAM, buffer);
+			dmaStreamSetMode(GDISP_DMA_STREAM, STM32_DMA_CR_PL(0) | STM32_DMA_CR_PINC | STM32_DMA_CR_PSIZE_HWORD | STM32_DMA_CR_MSIZE_HWORD | STM32_DMA_CR_DIR_M2M);
+			for (i = area / 65535; i; i--) {
+				dmaStreamSetTransactionSize(GDISP_DMA_STREAM, 65535);
+				dmaStreamEnable(GDISP_DMA_STREAM);
+				dmaWaitCompletion(GDISP_DMA_STREAM);
+			}
+			dmaStreamSetTransactionSize(GDISP_DMA_STREAM, area % 65535);
+			dmaStreamEnable(GDISP_DMA_STREAM);
+			dmaWaitCompletion(GDISP_DMA_STREAM);
+		#else
+			coord_t endx, endy;
+			uint32_t lg;
+			endx = srcx + cx;
+			endy = y + cy;
+			lg = srccx - cx;
+			for(; y < endy; y++, buffer += lg)
+				for(x=srcx; x < endx; x++)
+					write_data(*buffer++);
+		#endif // defined(GDISP_USE_DMA) && defined(GDISP_DMA_STREAM)
+
 		stream_stop();
 		release_bus();
 	}
@@ -418,8 +471,16 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 		acquire_bus();
 		set_cursor(x, y);
 		stream_start();
-		color = read_data();			// dummy read
+
+		/* FSMC timing */
+		FSMC_Bank1->BTCR[FSMC_Bank + 1] = FSMC_BTR1_ADDSET_3 | FSMC_BTR1_DATAST_3 | FSMC_BTR1_BUSTURN_0;
+
+		color = read_data(); // dummy read
 		color = read_data();
+
+		/* FSMC timing */
+		FSMC_Bank1->BTCR[FSMC_Bank + 1] = FSMC_BTR1_ADDSET_0 | FSMC_BTR1_DATAST_2 | FSMC_BTR1_BUSTURN_0;
+
 		stream_stop();
 		release_bus();
 
@@ -450,8 +511,8 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 			if (x < GDISP.clipx0) { cx -= GDISP.clipx0 - x; x = GDISP.clipx0; }
 			if (y < GDISP.clipy0) { cy -= GDISP.clipy0 - y; y = GDISP.clipy0; }
 			if (!lines || cx <= 0 || cy <= 0 || x >= GDISP.clipx1 || y >= GDISP.clipy1) return;
-			if (x+cx > GDISP.clipx1)	cx = GDISP.clipx1 - x;
-			if (y+cy > GDISP.clipy1)	cy = GDISP.clipy1 - y;
+			if (x+cx > GDISP.clipx1) cx = GDISP.clipx1 - x;
+			if (y+cy > GDISP.clipy1) cy = GDISP.clipy1 - y;
 		#endif
 
 		abslines = lines < 0 ? -lines : lines;
@@ -474,9 +535,17 @@ void gdisp_lld_draw_pixel(coord_t x, coord_t y, color_t color) {
 				/* read row0 into the buffer and then write at row1*/
 				set_viewport(x, row0, cx, 1);
 				stream_start();
-				j = read_data();			// dummy read
+
+				/* FSMC timing */
+				FSMC_Bank1->BTCR[FSMC_Bank + 1] = FSMC_BTR1_ADDSET_3 | FSMC_BTR1_DATAST_3 | FSMC_BTR1_BUSTURN_0;
+
+				j = read_data(); // dummy read
 				for (j = 0; (coord_t)j < cx; j++)
 					buf[j] = read_data();
+
+				/* FSMC timing */
+				FSMC_Bank1->BTCR[FSMC_Bank + 1] = FSMC_BTR1_ADDSET_0 | FSMC_BTR1_DATAST_2 | FSMC_BTR1_BUSTURN_0;
+
 				stream_stop();
 
 				set_viewport(x, row1, cx, 1);
diff --git a/drivers/gdisp/SSD2119/gdisp_lld_board_embest_dmstf4bb.h b/drivers/gdisp/SSD2119/gdisp_lld_board_embest_dmstf4bb.h
index ccbba2a3..a223ad18 100644
--- a/drivers/gdisp/SSD2119/gdisp_lld_board_embest_dmstf4bb.h
+++ b/drivers/gdisp/SSD2119/gdisp_lld_board_embest_dmstf4bb.h
@@ -16,38 +16,60 @@
 #ifndef _GDISP_LLD_BOARD_H
 #define _GDISP_LLD_BOARD_H
 
+#define GDISP_USE_DMA
+#define GDISP_DMA_STREAM STM32_DMA2_STREAM6
+
 /* Using FSMC A19 (PE3) as DC */
-#define GDISP_REG	(*((volatile uint16_t *) 0x60000000)) /* DC = 0 */
-#define GDISP_RAM	(*((volatile uint16_t *) 0x60100000)) /* DC = 1 */
+#define GDISP_REG (*((volatile uint16_t *) 0x60000000)) /* DC = 0 */
+#define GDISP_RAM (*((volatile uint16_t *) 0x60100000)) /* DC = 1 */
+
+#define SET_RST palSetPad(GPIOD, 3);
+#define CLR_RST palClearPad(GPIOD, 3);
 
-#define SET_RST		palSetPad(GPIOD, 3);
-#define CLR_RST		palClearPad(GPIOD, 3);
+const unsigned char FSMC_Bank = 0;
 
-/* PWM configuration structure. We use timer 4 channel 2 (orange LED on board). */
+/*
+ * PWM configuration structure. We use timer 4 channel 2 (orange LED on board).
+ * The reason for so high clock is that with any lower, onboard coil is squeaking.
+ * The major disadvantage of this clock is a lack of linearity between PWM duty
+ * cycle width and brightness. In fact only with low preset one sees any change
+ * (eg. duty cycle between 1-20). Feel free to adjust this, maybe only my board
+ * behaves like this.
+ */
 static const PWMConfig pwmcfg = {
-  1000000,		/* 1 MHz PWM clock frequency. */
-  100,			/* PWM period is 100 cycles. */
-  NULL,
-  {
-   {PWM_OUTPUT_ACTIVE_HIGH, NULL},
-   {PWM_OUTPUT_ACTIVE_HIGH, NULL},
-   {PWM_OUTPUT_ACTIVE_HIGH, NULL},
-   {PWM_OUTPUT_ACTIVE_HIGH, NULL}
-  },
-  0
+	1000000,       /* 1 MHz PWM clock frequency. */
+	100,           /* PWM period is 100 cycles. */
+	NULL,
+	{
+		{PWM_OUTPUT_ACTIVE_HIGH, NULL},
+		{PWM_OUTPUT_ACTIVE_HIGH, NULL},
+		{PWM_OUTPUT_ACTIVE_HIGH, NULL},
+		{PWM_OUTPUT_ACTIVE_HIGH, NULL}
+	},
+	0
 };
 
 /**
  * @brief   Initialise the board for the display.
- * @notes	This board definition uses GPIO and assumes exclusive access to these GPIO pins
+ * @notes   This board definition uses GPIO and assumes exclusive access to these GPIO pins
  *
  * @notapi
  */
 static inline void init_board(void) {
-	unsigned char FSMC_Bank;
 
-	/* STM32F4 FSMC init */
-	rccEnableAHB3(RCC_AHB3ENR_FSMCEN, 0);
+	#if defined(STM32F4XX) || defined(STM32F2XX)
+		/* STM32F4 FSMC init */
+		rccEnableAHB3(RCC_AHB3ENR_FSMCEN, 0);
+
+		#if defined(GDISP_USE_DMA) && defined(GDISP_DMA_STREAM)
+			if (dmaStreamAllocate(GDISP_DMA_STREAM, 0, NULL, NULL))
+				gfxExit();
+			dmaStreamSetMemory0(GDISP_DMA_STREAM, &GDISP_RAM);
+			dmaStreamSetMode(GDISP_DMA_STREAM, STM32_DMA_CR_PL(0) | STM32_DMA_CR_PSIZE_HWORD | STM32_DMA_CR_MSIZE_HWORD | STM32_DMA_CR_DIR_M2M);
+		#endif
+	#else
+		#error "FSMC not implemented for this device"
+	#endif
 
 	/* Group pins */
 	IOBus busD = {GPIOD, (1 << 0) | (1 << 1) | (1 << 4) | (1 << 5) | (1 << 7) | (1 << 8) |
@@ -60,10 +82,11 @@ static inline void init_board(void) {
 	palSetBusMode(&busD, PAL_MODE_ALTERNATE(12));
 	palSetBusMode(&busE, PAL_MODE_ALTERNATE(12));
 
-	FSMC_Bank = 0;
+	/* FSMC timing */
+//	FSMC_Bank1->BTCR[FSMC_Bank+1] = FSMC_BTR1_ADDSET_0 | FSMC_BTR1_DATAST_2 | FSMC_BTR1_BUSTURN_0 ;
 
 	/* FSMC timing */
-	FSMC_Bank1->BTCR[FSMC_Bank+1] = (FSMC_BTR1_ADDSET_1 | FSMC_BTR1_ADDSET_3) \
+	FSMC_Bank1->BTCR[FSMC_Bank + 1] = (FSMC_BTR1_ADDSET_1 | FSMC_BTR1_ADDSET_3) \
 			| (FSMC_BTR1_DATAST_1 | FSMC_BTR1_DATAST_3) \
 			| (FSMC_BTR1_BUSTURN_1 | FSMC_BTR1_BUSTURN_3) ;