PPU bus activity traces

[lidnariq]

Tried to get all of AWJ's requests.

Rather than a separate trace for each PPU mode, I used HDMA to switch to the next mode every 32 scanlines. Layer 1 is using 16x16 tiles, layers 2-4 are using 8x8 tiles. Sprites #0-95 are 16x16, sprites #96-127 are 8x8.

Every sixteenth word of the nametable for layer 1 has the "flipped horizontally" bit set.

Sample rate is now 12MHz, so the aliasing will be a bit more annoying (alternating between 167ns and 250ns instead of the true value of 186ns)

I couldn't get all the logic analyzer clips on one RAM, so this is a random mixture of lines on U5 and U4.

Pertinent details: BGnSC of $70, $74, $78, $7C; BGnNBA of 0, 3, 5, 6; OBSEL=$0A. HDMA changes BGMODE from $10...$16, each value lasting for 32 scanlines. HDMA changes BG3VOFS to get offset-per-tile to be visible.

[lidnariq]

Well, let's try to recapitulate (edit: add link) AWJ's previous analysis, with the extra data:

Mode 0: no new insights.
Modes 1 & 3: Lower-addressed bitplane is fetched first
Mode 2: Also lower-addressed OPT row is fetched first
Mode 5: Fetch cadence appears to be completely identical to mode 3. Horizontal flip flag does reverse left-right fetch order.
Modes 4,6: No new insights beyond above.

Sprites: Lower-addressed bitplane is fetched first. Horizontally flipped sprites reverse sliver fetch order.

Each scanline fetches 33 slivers for tiles (taking 8 cycles per tile, for 264 total cycles), followed by 8 idle cycles (bus unchanging), followed by 34 slivers for sprites (taking 2 cycles per sliver, 68 total cycles).
That's only 340 cycles, but there's 341 usually; two extra half-cycles appear to be be inserted during hsync and immediately after it ends.

Hsync timing is not obviously aligned to sprite fetch cycles. It's hard to tell anything more precise with the comparatively low sample rate.

[creaothceann]

somewhat relevant?

[whicker]

somewhat relevant?

Damn cool seeing the triple 15-bit DAC of PPU2 in the upper left corner, and the 15-bit palette RAM in the bottom left corner.

As personally expected, the All-in-one chip is a lump of logic sausage squeezed around the earlier-tested handmade RAM blocks.

[psycopathicteen]

Well, let's try to recapitulate AWJ's previous analysis, with the extra data:

Mode 0: no new insights.
Modes 1 & 3: Lower-addressed bitplane is fetched first
Mode 2: Also lower-addressed OPT row is fetched first
Mode 5: Fetch cadence appears to be completely identical to mode 3. Horizontal flip flag does reverse left-right fetch order.
Modes 4,6: No new insights beyond above.

Sprites: Lower-addressed bitplane is fetched first. Horizontally flipped sprites reverse sliver fetch order.

Each scanline fetches 33 slivers for tiles (taking 8 cycles per tile, for 264 total cycles), followed by 8 idle cycles (bus unchanging), followed by 34 slivers for sprites (taking 2 cycles per sliver, 68 total cycles).
That's only 340 cycles, but there's 341 usually; two extra half-cycles appear to be be inserted during hsync and immediately after it ends.

Hsync timing is not obviously aligned to sprite fetch cycles. It's hard to tell anything more precise with the comparatively low sample rate.

What happens when you write to $2118 and $2119 during those 8 idle cycles?

[Señor Ventura]

Well, let's try to recapitulate AWJ's previous analysis, with the extra data:

Mode 0: no new insights.
Modes 1 & 3: Lower-addressed bitplane is fetched first
Mode 2: Also lower-addressed OPT row is fetched first
Mode 5: Fetch cadence appears to be completely identical to mode 3. Horizontal flip flag does reverse left-right fetch order.
Modes 4,6: No new insights beyond above.

Sprites: Lower-addressed bitplane is fetched first. Horizontally flipped sprites reverse sliver fetch order.

Each scanline fetches 33 slivers for tiles (taking 8 cycles per tile, for 264 total cycles), followed by 8 idle cycles (bus unchanging), followed by 34 slivers for sprites (taking 2 cycles per sliver, 68 total cycles).
That's only 340 cycles, but there's 341 usually; two extra half-cycles appear to be be inserted during hsync and immediately after it ends.

Hsync timing is not obviously aligned to sprite fetch cycles. It's hard to tell anything more precise with the comparatively low sample rate.

But it only draws 256 sprite pixels, right?

[tepples]

Up to 272 sprite pixels are rendered into the line buffer. Some of them are transparent, and some are occluded by other sprite pixels, giving 256 pixels of output from the sprite layer.

Sprites aren't hi-res. Only one sprite pixel color can enter the compositor per pixel, unlike backgrounds where two different layers or two different pixels of a mode 5/6 layer can be displayed as half-pixels.