Sprite doubling and graphic acceleration stuff

Hardware sprite now scales with video output, courtesy of mntmn! Some fixes and changes to mouse cursor position when scaled, plus some graphics acceleration changes to make it easier to use.

ZZ9000_proto.sdk/ZZ9000OS/src/gfx.c

@@ -944,26 +944,22 @@ void template_fill_rect(uint32_t color_format, uint16_t rect_x1, uint16_t rect_y
 }
 
 // Generic graphics acceleration functionality
-void acc_clear_buffer(uint32_t addr, uint16_t w, uint16_t h, uint16_t pitch_, uint32_t fg_color, uint32_t color_format)
+void acc_clear_buffer(uint32_t addr, uint16_t w, uint16_t h, uint16_t pitch_, uint32_t fg_color, uint32_t color_format_)
 {
 	if (!w || !h || !addr)
 		return;
 
-	uint16_t pitch = pitch_;
+	uint16_t pitch = pitch_ * color_format_;
 	uint8_t* dp = (uint8_t*)((uint32_t)addr);
 	uint8_t u8_fg = fg_color >> 24;
 
-	if (pitch == 0) {
-		switch (color_format) {
-			case MNTVA_COLOR_8BIT: pitch = w; break;
-			case MNTVA_COLOR_16BIT565: pitch = w * 2; break;
-			case MNTVA_COLOR_32BIT: pitch = w * 4; break;
-			default: return; break;
-		}
+	uint8_t color_format = 0;
+	if (color_format_ == 2) {
+		color_format = MNTVA_COLOR_16BIT565;
+	} else if (color_format == 4) {
+		color_format = MNTVA_COLOR_32BIT;
 	}
 
-	//printf("Clearing %dx%d pixels, %d bytes (%d).\n", w, h, h * pitch, pitch);
-
 	switch(color_format) {
 		case MNTVA_COLOR_8BIT:
 			memset(dp, u8_fg, h * pitch);
@@ -992,27 +988,11 @@ void acc_flip_to_fb(uint32_t src, uint32_t dest, uint16_t w, uint16_t h, uint16_
 	if (!w || !h || !src || !dest)
 		return;
 
-	uint16_t pitch = pitch_;
+	uint16_t pitch = pitch_ * color_format;
 	uint8_t* sp = (uint8_t*)((uint32_t)src);
 	uint8_t* dp = (uint8_t *)((uint32_t)dest);
-	//uint8_t* dp = (uint8_t*)((uint32_t)(framebuffer + framebuffer_pan_offset + 0x10000));
-
-	if (pitch == 0) {
-		switch (color_format) {
-			case MNTVA_COLOR_8BIT: pitch = w; break;
-			case MNTVA_COLOR_16BIT565: pitch = w * 2; break;
-			case MNTVA_COLOR_32BIT: pitch = w * 4; break;
-			default: return; break;
-		}
-	}
 
-	/*for (int i = 0; i < h; i++) {
-		memcpy(dp, sp, pitch);
-		sp += pitch;
-		dp += pitch;
-	}*/
 	memcpy (dp, sp, h * pitch);
-	//printf("Flipping %dx%d pixels, %d bytes (%d).\n", w, h, h * pitch, pitch);
 }
 
 void acc_blit_rect(uint32_t src, uint32_t dest, uint16_t dx, uint16_t dy, uint16_t w, uint16_t h, uint16_t src_pitch, uint16_t dest_pitch, uint8_t draw_mode, uint8_t mask_color)

ZZ9000_proto.sdk/ZZ9000OS/src/main.c

@@ -70,6 +70,8 @@ int hdmi_ctrl_write_byte(u8 addr, u8 value) {
 	buffer[1] = value;
 	int status;
 
+	malloc(50);
+
 	while (XIicPs_BusIsBusy(&Iic)) {
 	};
 	usleep(I2C_PAUSE);
@@ -449,13 +451,13 @@ void pixelclock_init(int mhz) {
 	//XClk_Wiz_WriteReg(XPAR_CLK_WIZ_0_BASEADDR,  0x25C, 0x00000001);
 
 	phase = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x20C);
-	printf("CLK phase: %lu\n", phase);
+	//printf("CLK phase: %lu\n", phase);
 	duty = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x210);
-	printf("CLK duty: %lu\n", duty);
+	//printf("CLK duty: %lu\n", duty);
 	divide = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x208);
-	printf("CLK divide: %lu\n", divide);
+	//printf("CLK divide: %lu\n", divide);
 	muldiv = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x200);
-	printf("CLK muldiv: %lu\n", muldiv);
+	//printf("CLK muldiv: %lu\n", muldiv);
 }
 
 void pixelclock_init_2(struct zz_video_mode *mode) {
@@ -479,13 +481,13 @@ void pixelclock_init_2(struct zz_video_mode *mode) {
 	//XClk_Wiz_WriteReg(XPAR_CLK_WIZ_0_BASEADDR,  0x25C, 0x00000001);
 
 	phase = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x20C);
-	printf("CLK phase: %lu\n", phase);
+	//printf("CLK phase: %lu\n", phase);
 	duty = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x210);
-	printf("CLK duty: %lu\n", duty);
+	//printf("CLK duty: %lu\n", duty);
 	divide = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x208);
-	printf("CLK divide: %lu\n", divide);
+	//printf("CLK divide: %lu\n", divide);
 	muldiv = XClk_Wiz_ReadReg(XPAR_CLK_WIZ_0_BASEADDR, 0x200);
-	printf("CLK muldiv: %lu\n", muldiv);
+	//printf("CLK muldiv: %lu\n", muldiv);
 }
 
 // FIXME!
@@ -564,19 +566,19 @@ void video_system_init(int hres, int vres, int htotal, int vtotal, int mhz,
 	printf("VSI: %d x %d [%d x %d] %d MHz %d Hz, hdiv: %d vdiv: %d\n", hres,
 			vres, htotal, vtotal, mhz, vhz, hdiv, vdiv);
 
-	printf("pixelclock_init()...\n");
+	//printf("pixelclock_init()...\n");
 	pixelclock_init(mhz);
-	printf("...done.\n");
+	//printf("...done.\n");
 
-	printf("hdmi_set_video_mode()...\n");
+	//printf("hdmi_set_video_mode()...\n");
 	//hdmi_set_video_mode(hres, vres, mhz, vhz, hdmi);
 
-	printf("hdmi_ctrl_init()...\n");
+	//printf("hdmi_ctrl_init()...\n");
 	hdmi_ctrl_init();
 
-	printf("init_vdma()...\n");
+	//printf("init_vdma()...\n");
 	init_vdma(hres, vres, hdiv, vdiv);
-	printf("...done.\n");
+	//printf("...done.\n");
 
 	//dump_vdma_status(&vdma);
 }
@@ -586,19 +588,19 @@ void video_system_init_2(struct zz_video_mode *mode, int hdiv, int vdiv) {
 	printf("VSI: %d x %d [%d x %d] %d MHz %d Hz, hdiv: %d vdiv: %d\n", mode->hres,
 			mode->vres, mode->hmax, mode->vmax, mode->mhz, mode->vhz, hdiv, vdiv);
 
-	printf("pixelclock_init()...\n");
+	//printf("pixelclock_init()...\n");
 	pixelclock_init_2(mode);
-	printf("...done.\n");
+	//printf("...done.\n");
 
-	printf("hdmi_set_video_mode()...\n");
+	//printf("hdmi_set_video_mode()...\n");
 	//hdmi_set_video_mode(hres, vres, mhz, vhz, hdmi);
 
-	printf("hdmi_ctrl_init()...\n");
+	//printf("hdmi_ctrl_init()...\n");
 	hdmi_ctrl_init();
 
-	printf("init_vdma()...\n");
+	//printf("init_vdma()...\n");
 	init_vdma(mode->hres, mode->vres, hdiv, vdiv);
-	printf("...done.\n");
+	//printf("...done.\n");
 
 	//dump_vdma_status(&vdma);
 }
@@ -707,15 +709,19 @@ void sprite_reset() {
 }
 
 void update_hw_sprite_pos(int16_t x, int16_t y) {
-	sprite_x = x + sprite_x_offset + 3;
+	sprite_x = x + sprite_x_offset + 1;
 	// horizontally doubled mode
-	if (scalemode & 1) sprite_x *=2;
-	sprite_x_adj = sprite_x;
+	if (scalemode & 1)
+		sprite_x_adj = (sprite_x * 2) + 1;
+	else
+		sprite_x_adj = sprite_x + 2;
 
 	sprite_y = y + sprite_y_offset + 1;
 	// vertically doubled mode
-	if (scalemode & 2) sprite_y *= 2;
-	sprite_y_adj = sprite_y;
+	if (scalemode & 2)
+		sprite_y_adj = sprite_y *= 2;
+	else
+		sprite_y_adj = sprite_y;
 
 	if (sprite_x < 0 || sprite_y < 0) {
 		if (sprite_clip_x != sprite_x || sprite_clip_y != sprite_y) {
@@ -1329,7 +1335,6 @@ int main() {
 							SWAP32(data->offset[0]);
 							data->offset[0] += ADDR_ADJ;
 
-							//printf("clearbuf: %.8X, %dx%d, %d size: %d\n", data->offset[0], data->x[0], data->y[0], data->pitch[0], sizeof(struct GFXData));
 							acc_clear_buffer(data->offset[0], data->x[0], data->y[0], data->pitch[0], data->rgb[0], data->u8_user[GFXDATA_U8_COLORMODE]);
 							break;
 						}
@@ -1343,7 +1348,6 @@ int main() {
 							data->offset[0] += ADDR_ADJ;
 							data->offset[1] += ADDR_ADJ;
 
-							//printf("flipbuf: %.8X to %.8X %dx%d, %d\n", data->offset[0], data->offset[1], data->x[0], data->y[0], data->pitch[0]);
 							acc_flip_to_fb(data->offset[0], data->offset[1], data->x[0], data->y[0], data->pitch[0], data->u8_user[GFXDATA_U8_COLORMODE]);
 							break;
 						case ACC_OP_BLIT_RECT:
@@ -1384,7 +1388,7 @@ int main() {
 								break;
 							}
 
-							printf ("Allocating %dx%d surface, %d bytes per pixel, %d bytes.\n", data->x[0], data->y[0], cf_bpp[data->u8_user[GFXDATA_U8_COLORMODE]], sfc_size);
+							printf ("Allocating %dx%d surface, %d bytes per pixel, %d bytes.\n", data->x[0], data->y[0], data->u8_user[GFXDATA_U8_COLORMODE], sfc_size);
 							uint8_t *p = calloc(1, sfc_size);
 							printf ("Surface allocated at offset %p, or %p on the Amiga side.\n", p, p - ADDR_ADJ);
 							data->offset[0] = (uint32_t)(p - ADDR_ADJ);

video_formatter.v

@@ -53,6 +53,7 @@ localparam OP_UNUSED1=12;
 localparam OP_SPRITEXY=13;
 localparam OP_SPRITE_ADDR=14;
 localparam OP_SPRITE_DATA=15;
+localparam OP_VIDEOCAP=16; // we ignore this here, it's snooped by MNTZorro
 
 localparam CMODE_8BIT=0;
 localparam CMODE_16BIT=1;
@@ -112,9 +113,14 @@ reg [23:0] sprite_buffer[SPRITE_SIZE-1:0];
 reg [11:0] sprite_addr_in;
 reg [11:0] sprite_x;
 reg [11:0] sprite_y;
+reg sprite_dbl;
+reg vga_sprite_dbl; // vga_domain
 reg [11:0] vga_sprite_x; // vga domain
 reg [11:0] vga_sprite_y; // vga domain
+reg [11:0] vga_sprite_x2; // vga domain
+reg [11:0] vga_sprite_y2; // vga domain
 reg [11:0] sprite_px; // vga domain
+reg [11:0] sprite_py; // vga domain
 reg [23:0] sprite_pix; // vga domain
 reg sprite_on; // vga domain
 
@@ -216,6 +222,7 @@ begin
     OP_SCALE: begin
         scale_x  <= control_data_in[0];
         scale_y  <= control_data_in[1];
+        sprite_dbl <= control_data_in[1];
       end
     OP_COLORMODE: colormode  <= control_data_in[1:0]; // FIXME
     OP_VSYNC: vsync_request <= 1; //control_data[0];
@@ -313,6 +320,9 @@ always @(posedge dvi_clk) begin
     vga_sprite_x <= sprite_x;
     vga_sprite_y <= sprite_y;
   end
+  vga_sprite_x2 <= vga_sprite_x+(SPRITE_W<<sprite_dbl);
+  vga_sprite_y2 <= vga_sprite_y+(SPRITE_H<<sprite_dbl);
+  vga_sprite_dbl <= sprite_dbl;
   
   // FIXME there is some non-determinism in the relationship
   // between this process and the fetching process
@@ -395,11 +405,16 @@ always @(posedge dvi_clk) begin
     CMODE_32BIT: pixout <= pixout32_dly2;
   endcase
   
-  sprite_pix <= sprite_buffer[sprite_px];
-  if (counter_y >= vga_sprite_y && counter_y < (vga_sprite_y+SPRITE_H) 
-      && counter_x >= vga_sprite_x && counter_x < (vga_sprite_x+SPRITE_W)) begin
+  sprite_pix <= sprite_buffer[((sprite_py>>sprite_dbl)<<5)+(sprite_px>>sprite_dbl)];
+  if (counter_y >= vga_sprite_y && counter_y < vga_sprite_y2 
+      && counter_x >= vga_sprite_x && counter_x < vga_sprite_x2) begin
     sprite_on <= 1;
-    sprite_px <= sprite_px + 1;
+    if (sprite_px < (SPRITE_W<<sprite_dbl)-1'b1)
+      sprite_px <= sprite_px + 1'b1;
+    else begin
+      sprite_px <= 0;
+      sprite_py <= sprite_py + 1'b1;
+    end
   end else begin
     sprite_on <= 0;
   end
@@ -411,6 +426,7 @@ always @(posedge dvi_clk) begin
     if (counter_y > vga_v_max) begin
       counter_y <= 0;
       sprite_px <= 0;
+      sprite_py <= 0;
     end else begin
       counter_y <= counter_y + 1'b1;
     end