mgras: zoomed glDrawPixels and 1-2 texel RGB textures draw right

[atomchild411]

Two IMPACT bugs found with an OpenGL 1.2 demo running through SGI's own libGL on an emulated Max IMPACT (IP28, IRIX 6.5.22). Its 320x240 "retro" frames are drawn back with glPixelZoom(3.2, 3.2), and they came out as a smeared ~240-pixel strip down the left side. Its Amiga-style ball chain, textured with a 2x2 GL_RGB red/white checker, came out yellow, red and black. Both bugs are in the board model and shared by every IMPACT board.

1. glDrawPixels under a pixel zoom (43fb5a6)

Traced (mgras trace hq,rss): the driver sends a zoomed DrawPixels by DMA in column chunks (156 RGBA or 144 RGB pixels at whole zooms, 96 at 3.2). Before each chunk, pixel state 0xDA8 gives the chunk's rectangle on the screen, already zoomed (relative to the raster position), and the DMA brings its rows unzoomed: the board stretches them over the rectangle. 0x192 holds 1 / the y zoom only (0x191 stays 0); the x zoom has no state of its own.

The model applied 0x192's zoom to both axes. It also placed a chunk only when 0xDA8's transfer size matched the rows that came, which never happens under a y zoom, so every chunk landed at the origin on top of the last. In addition, at x zoom 3.2 the transfer size's width is the pixel count times the zoom rounded up (0x180 for 96 pixels), and the "narrow host pixel" guess read that as 1-byte pixels.

Now a chunk whose 0xDA8 rectangle is its rows times the y zoom high (zoomed_chunk) is placed at that rectangle, with its rows stretched over it, nearest pixel. The host's pixel size there comes from the row words, the transfer size and the rectangle (zoomed_host_bpp). FIFO images keep 0x192's zoom both ways, so snoop's magnifier and gl_draw_pixels_zoomed are unchanged.

2. Texture rows by DMA padded to words (5b50b38)

The host DMA pads each line to whole words. A texture row of 1 or 2 RGB texels (3 or 6 bytes) reached the loader as 4 or 8 bytes, and the loader took bytes-per-texel as the row's bytes over its width: 4, i.e. RGBA. So the 2x2 checker drew its white texels as f5 f5 00, the blue being a pad byte. Wider RGB rows fill their words (4 texels = 12 bytes) and were right. The HQ now passes the DMA's LINE_BYTES (6 here) to the GL emulation, and the loader takes the texel size from it when the padded row is longer (dma_row_bpt). Rows sent by FIFO are as before.

Tested

Two small GLX programs on IRIX through SGI's libGL only (no host GL), on Max IMPACT. Each draws known patterns, reads them back, and checks every source pixel or texel:

	before	after
glDrawPixels 320x240, RGBA and RGB at zoom 1, 2, 3.2, plus 3.2 x 1 and 1 x 3.2	2 / 8 (zoom 1 only)	8 / 8
N x N red/white checker textures, RGB 2, 4, 8, 16, 64 and RGBA 2, 4, 64	7 / 8 (2x2 RGB)	8 / 8

The demo now draws its full-screen retro eras and red-and-white ball chain correctly (22 fps). mgras tests: 155 pass, including two new ones built from the traced values (gl_zoomed_chunk_host_pixel_size, gl_dma_texture_row_texel_size). The repro programs are small; happy to add them under test/gltest if you'd like.

🤖 Generated with Claude Code

[atomchild411]

cancel spotfix for overlap against techomancer's mega fix 🙌