• Notes about video quality and sampling approaches

    zpekic01/07/2021 at 22:00 0 comments

    While overall the sampling of TIM-011 video signal works, and is visibly displayed on the screen, it still has problems:

    • first column pixel is displayed (black instead) because there is no real pixel pipeline (and consequently, the last pixel column is not displayed
    • there are some random "ghost artifacts" (vertical bars)

    Perhaps not clearly visible, a picture from the VGA screen (generated by sampler + VGA controller):

    Now looking at screenshot from GONBES-8200, it has different problems:

    • top few lines dancing
    • distorted height / width ratio

    So in some ways, both are "worse" but the vertical bar artifacts are very annoying. I haven't tested if those would prevent reading of text on the TIM-011 display.

    First, I thought that the image would deteriorate going from left to right, due to the skew between video signal to be sampled and the sampling clock - after all they are synchronized only once per line using hsync signal, and by 512th video signal (or 2048th sampling clock) they could be slightly off. But that is not the case, the quality is evenly good (or bad).

    So the artifacts must come from the sampling approach itself. To "debug" it, I feed the sampler with 6 extra signals that control the sampling.

    From /sys_tim011_mercury.vhd

    tim: tim_sampler port map (
    		reset => RESET,
    		clk => freq48M, -- 48MHz (4 times oversample of 12MHz)
    		hsync => TIM_HSYNC,
    		vsync => TIM_VSYNC,
    		v2 =>	TIM_VIDEO2,
    		v1 => TIM_VIDEO1,
    		a => sampler_a,
    		d => vram_dina,
    		--limit => switch(7 downto 2),
    		-- best result with sampler "algorithm"
    		-- s2 from raising edge sample
    		-- s1 from raising edge sample
    		-- 4 out of 4 sample: on
    		-- 3 out of 4 sample: on
    		-- 2 out of 4 sample: off
    		-- 1 out of 4 sample: off
    		limit => "111100",
    		we_in => we_in,
    		we_out => sampler_wr_nrd
    	);

    The two MSB of the (not very well) named "limit" are consumed by the tim_sampler.vhd to select if for the  v1, v0 video signals should be captured at the rising or falling edge of the sampling clock (s2 and s1 are 16-bit shift registers, clocked at 4*12MHz  rate and ingesting the TIM-011 v1 and v0 video signals):

    generate_s: for i in 15 downto 1 generate
    begin
    	s2(i) <= s2r(i) when (limit(5) = '1') else s2f(i);
    	s1(i) <= s1r(i) when (limit(4) = '1') else s1f(1);
    end generate;
    s2(0) <= v2;
    s1(0) <= v1;

    Answer: raising is much better for picture quality.

    The 4 LSB are consumed in the "voter" circuit which gets 4 sample bits (per pixel) and has to decide if those 4 indicated "signal on" or off. It does that by looking at the sampled patterns:

    	with value select vote <=
    		limit(3) when "1111",	--4
    		limit(2) when "1110",	--3
    		limit(2) when "1101",	--3
    		limit(1) when "1100",	--2
    		limit(2) when "1011",	--3
    		limit(1) when "1010",	--2
    		limit(1) when "1001",	--2
    		limit(0) when "1000",	--1
    		limit(2) when "0111",	--3
    		limit(1) when "0110",	--2
    		limit(1) when "0101",	--2
    		limit(0) when "0100",	--1
    		limit(1) when "0011",	--2
    		limit(0) when "0010",	--1
    		limit(0) when "0001",	--1
    		'0' when others;	--0

    Obviously, if no bit was sampled "1" then the output must be "0" - this is the last, default line. But what if it was sampled one of more times "1"? In that case, the output is controlled by selecting on or off individually each combination with 1, 2, 3, or 4 bits sampled "1" (a simple 16->1 MUX).

    Through visual experimentation, turns out the best result is enable 4 and 3 bit "1" sample combinations, but not the 2 and 1 bit ones (1 results in unstable pic, 2 is effectively a no-op).

  • Converting TIM-011 video to VGA

    zpekic11/29/2020 at 06:14 0 comments

    Converting incompatible video standards especially in retrocomputing field is a well-known problem, for which a whole cottage industry has been created (e.g. GONBES and similar). 

    To recap:

    TIM-011VGA (basic mode)
    Color infoDigital, 4 level, 2 signalsAnalog, 3 signals, theoretically 2^24 colors with 3 8-bit DACs
    Pixel clock12MHz25.125MHz (25Mhz is used for design simplicity)
    Horizonal sync15.625kHz31.25kHz
    Vertical sync50Hz60Hz

    As can be seen from the above, any simple or "passive" connectivity between the two won't work. The apparent solution is to:

    1. Capture input signal (synchronized to TIM-011 timing)
    2. Store video data in memory (buffer)
    3. Generate output signal from memory (as VGA)


    Read more »

  • Initializing video memory for test purposes

    zpekic11/29/2020 at 04:01 0 comments

    In order to see if the generated video signal is good or not, one simple trick is to pre-populate video RAM with image as if CPU had generated it. 

    I used two images:

    TIM-011 signal generator on Anvyl - my lame MSPaint image:

    TIM-011 sampler on Mercury - model students busy learning Basic from original 1988 ad (I hope I am not breaking their copyright!)

    Of course, both of these converted to 512*256*4 colors look a bit less impressive...

    All FPGAs support initializing RAM/ROM which is part of the design. However, the file format varies. As I was using Xilinx ISE14.7, I needed .coe file to include into project.  

    Here are steps to generate the .coe:

    1. Convert image to .bin file using Img2Tim utility which will:

    • Ingest .png, .jpg, .bmp file and instantiate a memory image object from it
    • Resize (squish) to 512*256
    • Sample the colors using simple thresholds to get only 2 bits per pixel
    • Use the knowledge of TIM-011 video memory organization to generate 32kB .bin file

    (use Img2Tim.cmd)

    2. Once we have the .bin file, use file conversion mode of my microcode compiler to generate .coe from .bin

    (use mcc.cmd)

    3. Finally, use ISE native tooling to generate the RAM component and point it to .coe file

    There are other approaches too - for example, tooling could generate VHDL source code directly that initializes the memory and include that file into the project to be compiled, but such file would be very large and slow down compilation. mcc is able to generate VHDL from .bin file in case non-Xilinx approach is needed.

  • TIM-011 video signal generation ("grafika" component)

    zpekic11/29/2020 at 03:35 0 comments

    In order to convert video signal coming from TIM-011 to VGA, one has to first have it... As I don't have a TIM-011, I had to "simulate" one. Key component of this simulation is to re-create the custom graphics system of TIM-011 on FPGA.

    Luckily, I had the schematics from the magazine, so I could reverse-engineer it.

    First, a bit about TIM-011 graphics implementation:

    • Implemented using 74XX TTL and CDXXXX CMOS ICs - no VDP!
    • Resolution is 512 (H) * 256 (V)
    • Graphics mode only - text is "painted" by software similar to classic Macintosh computer approach of the same era
    • 4 "colors" - or shades of gray per pixel. This means 1 byte contains 4 pixels. Total memory needed for picture is then 512*256/4 = 32k
    • Video RAM is a single generic 32kB static RAM, it is mapped into I/O space of HD64180 CPU (which has 64k I/O space) from 0x8000 to 0xFFFF
    • In addition, simple scroll up/down is supported by varying the offset location of row 0 in video memory.

     As a result, following video signal is generated:

    • Dot clock is 12MHz
    • 1 scan line contains positive HSYC signal, 256 dot periods long (21us), after which 512 dot periods (43us) is the video signal (V2 and V1)
    • 1 scan line takes therefore 64us which gives the horizontal frequency of 15.625kHz
    • VSYC positive signal comes after every 256 scan lines, and is 2.04ms long (24480 dot clock periods)
    • VSYNC repeats at 50Hz, which is the vertical refresh rate (there is no interlacing)

    There are two interesting facts about TIM-011 video:

    • V2 and V1 are "anded" with dotclk - this means that their intensity is halved. Not sure why this was done, maybe it improves the image quality on old home computer monitors preventing the pixels to bleed to each other due to phosphorus luminance.
    • The pixels are "packed" in a byte in a strange way:
    Bit:76543210
    Pixel:1032
    Intensity:V2V1V2V1V2V1V2V1

    VDHL implemenation:

    To follow description below, refer to this source code file

    (to be updated)

  • Demo video

    zpekic11/29/2020 at 02:21 0 comments

    Really low-quality, terrible video, but hopefully illustrating what is working so far:

    Anvl-board generates video signal:

    • Converted to CSYNC (composite sync) + RGB digital, goes to GBS8200 board towards VGA
    • Raw (HSYNC+CSYNC+V2+V1) goes to Mercury board
    • Scroll up/down is supported like real TIM
    • Test signal mode has been added, it displays a 4 color box pattern

    Mercury board captures the video signal:

    • Incoming video signal is converted to memory bytes and stored in 32kB onboard RAM
    • VGA controller reads the same RAM, applies a 4-color palette and outputs as 640*480, 60Hz
    • Display outside TIM-011 512*256 window is filled with hard-coded ASCII characters (text display in the future)
    • TIM window can be moved on VGA screen (it is centered initially)

    Upcoming project logs will describe these in more details.