Repository navigation
Imageproc - #222
Merged
Merged
Imageproc#222
Conversation
The kernel evicts a context by copying its slot once per GE: reads at S and S - 1 (GE1's output runs a word late), then both halves back to S. With one ERAM modelled, GE1's half overwrote GE0's a word off, shifting the saved GL state by a register: Matador came back from being iconified drawing into the wrong buffer. A transfer repeating GE0's last one now goes to GE1's ERAM. Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A negative pixel zoom shows only in GE pixel state 0xDA8, whose rows then run down from the raster position; draw them that way (Mirai's splash was upside down). Answer GE readback 0xB9C, glIsEnabled(GL_SCISSOR_TEST): Mirai disables scissoring for its front-to-back copy only when it is on, and our 0 clipped the copy to the last widget's box, leaving a stale menu bar in one bank. Co-Authored-By: Claude Opus 5.5 <[email protected]>
Mirai's RGB16 copies into RGBA16 send narrow strips whose rows do not fill whole words (93 pixels = 558 bytes in 140 words). The host pixel size was only recognised when the row bytes divided by the width, so these were drawn as 8-byte pixels: short, channels mixed, and without their 0xDA8 offset (the welcome dialog's right edge at its left, menu labels garbled after a menu closed). Co-Authored-By: Claude Opus 5.5 <[email protected]>
Pixel data commands whose data starts mid-word (libGLcore's rows at UNPACK_ALIGNMENT 1) carry one more word than their byte count plus a pad; the HQ took the declared count only, parsed the rest as commands and the client hung. Realign them. GE_DATA and the readback answer now belong to the context: another client clearing the flag wiped a pending answer and its owner waited forever. test/gltest/pixtest.c: self-checking tests of every GL 1.1 pixel path (formats, types, store and transfer modes, zoom, copies, buffers, textures). tools/mgras-pixdecode.py splits a recording by pixtest case and names the GE pixel pipeline's modules from the microcode's line table. Co-Authored-By: Claude Opus 5.5 <[email protected]>
libGLcore programs the GE11's pixel path as a list of microcode modules at GE 0x13B (module k's state address in pair k + 1), named from the microcode's line table: inttofloat, shiftadd, unpack, scalebias, expand, LUT, clamp, mask, colmatrix, pipecontrol. pixpipe.rs runs them on a row's component stream with their traced state layouts, so draws convert every GL format and type, pixel transfer scale/bias and maps as the board does, and hand the raster engine the transfer type libGLcore asked for. State blocks too big for a token, and colour tables (INIT_COLOR_TABLE), come as the next pixel data command: kept. The raster engine now converts host pixels by XFRMODE format (3-6 one component, 7 RGB, 8 RGBA) and type (unsigned and signed 8, 16, 32 bits), with RGBorder (bit 22, which the stride-skip field swallowed) choosing red- or alpha-first RGBA; RGBA16 keeps alpha, in the JIT too. Co-Authored-By: Claude Opus 5.5 <[email protected]>
The draw copied each raster pixel through an 8-byte buffer; RGBA FLOAT draws now reach the raster engine as four 32-bit components and the HQ thread panicked. Co-Authored-By: Claude Opus 5.5 <[email protected]>
- GET_PIXELS 0x400A is a PIO read (packed-type rows, one at a time): arm a PIO read block, not a DMA one. - Zoom: the GE repeats rows by 1/zy and, for a fractional x zoom, replicates ceil(zx) (xzoom); the raster engine zooms x by XFRFACTOR (XFRMODE bit 19) or decimates by it (bit 20). XFRMODE's stride skip is bits 18:14. - shiftadd shifts (left, or right when negative) and adds (index shift and offset); expand mode 7 copies an index to all four components for the I_TO_* maps; element size from the row when the type word does not say (stencil, colour index). - glBitmap LSB_FIRST: format word bit 17. - RGBA in wider types honours RGBorder (alpha first without it); host components to the framebuffer keep their top bits, 12:12:12 reads all twelve. - pixtest: XFAIL list for libGLcore's own deviations, zoom within a source pixel, textures through MODULATE with a white fragment. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…float - Texture loads (SEND_PIXELS 0x511A) run the GE pipeline too, with its compress module (0x399A: the internal format's components out of RGBA), so every format and type converts into every internal format. - Stencil draws: SAVE_RSS has the planes in bits 20:18 for a draw (24:22 for a read). - bitmaptofloat (0x24D7) for GL_BITMAP colour indices. - PIO reads of RGB(A) come last component first. - The formatter's byte swap with byte components swaps pairs (16-bit packed pixels). - Zoom: the decimation samples at each destination pixel's centre; the xzoom replication is its index table's first run. - pixtest: glTexSubImage2D error check and a resident-texture variant; libGLcore's glBitmap sub-byte skip as an expected failure. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…ex outputs wrap PIO reads of RGB(A) come in the reverse of the DMA order unless XFRMODE bit 18 is set (the X server's XGetImage sets it, libGLcore never does); stride skip is bits 17:14. Texture loads that compress to fewer components take four input components. Stencil and colour-index draws keep the low bits of the pipeline's output. pixtest: legal glTexSubImage2D geometry (libGLcore wants >= 32 texels wide or full width), error checks, XFAILs for libGLcore's own departures. 348 pass, 0 fail, 12 XFAIL on MaxImpact. Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Colour-index transfers in every integer type (format 0, types 3/4/5/7) in the raster engine; the raster JIT leaves them to the interpreter. Draws take their pixel count from 0xDA8's transfer width rather than the word-padded row (13 16-bit indices drew 14 wide). The GE's bitmaptoint module (GL_BITMAP on CI visuals). The LUT's offsets end at the 2^23 word (the 12-bit CI state's cycle word reads 14), and the table memory moves from the per-context record to the HQ engine, 4096 entries, so a 12-bit visual's full I_TO_I map fits. pixtest --ci 8|12 runs a colour-index suite: draw and read in every type, shift/offset, I_TO_I maps, unpack/pack layouts, zoom, GL_BITMAP, glBitmap, glIndexMask, glCopyPixels. 8-bit: 36 pass, 13 XFAIL; 12-bit: 44 pass, 5 XFAIL; the RGBA run is unchanged (348 pass). Co-Authored-By: Claude Opus 5.5 <[email protected]>
Cases for what PR #221 fixed in the old draw path: 200-pixel-wide RGBA and RGB images (DMA column chunks of 156 / 144) at zooms 1, 2, 3.2 and fractional x zooms, and 1x1 / 2x2 / 4x4 GL_RGB textures with unpack alignment 1 (word-padded DMA rows). All pass on the GE pipeline path: 365 pass, 0 fail, 12 XFAIL. Co-Authored-By: Claude Opus 5.5 <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
image processing series
and some eram fixes for state management