Repository navigation
Mgrasfaster - #193
Merged
Merged
Mgrasfaster#193
Conversation
IMPACT's per-pixel pipeline, compiled with Cranelift per configuration, modelled on the REX3 shader JIT. - RSS and TE1 state is plain #[repr(C)] data: the stipple and transfer Options became a plain Slot, Block/Stipple/Xfer/Sampler/Buffer are repr(C), and triangle and GL line setup are shared structs (Rss::tri_setup, Rss::gl_line_setup) that both engines draw from. - PipeKey packs a primitive's pipeline into 64 bits (mixed radix, about 2^58 at most): primitive, clipping, draw target, pixel format, logic op, alpha/stencil/depth tests, blending, texture environment, format, filtering and wrap modes, fog. Keys are normalised so configurations that draw alike share a shader. - Emitters mirror the interpreter expression for expression (Rust round, clamp, min and saturating casts spelled out; libm and divisions through the interpreter's own code), so shaders are bit-exact. - The board keeps a persistent context; only writes to pipeline state registers invalidate its cached part, and each primitive memoises its shader. Row values are pinned against Cranelift's rematerialisation. - rss_jit_tests.rs sweeps the key space coverage-first (every value of every field, greedy pair coverage, random data) on two boards and compares pixel memory, clip IDs and stipple state word for word. - mgras jit [on|off|sync|list] in the monitor; IRIS_GR4_JIT, GR4_JIT_DISASM. 2.5-6x faster than the interpreter on representative primitives. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The level of detail was hypot and libm log2, which the raster JIT could only reach by calling back into Rust mid-fragment. Both engines now use te1::lod: the square root of the larger sum of squares, and lod_log2 (the exponent plus a cubic in the mantissa, exact on powers of two, within 0.0011 elsewhere). The JIT emits the same operations inline, so shaders make no calls and stay bit-exact with the interpreter. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The RE4 and TE1 are integer machines; colour from the iterators to the framebuffer is now modelled that way (src/dev/mgras/fixed.rs): - colour and fog iterators as 12.16 planes in the shared triangle and GL line setup, evaluated base + dx (i - xs) + dy (yref - j), wrapping; - texels widened to 12 bits by repeating bits (4: x0x111, 8: v<<4|v>>4); - bilinear with 8-bit weights (the four products sum to 12.16 exactly), trilinear with an 8-bit mip fraction; - texture environment, fog and blending through mul = (a*b) >> 28; - shift down on write: >> 20 to bytes, >> 16 to a 12-bit index. The JIT emits the same integer operations: the equivalence sweep and the mgras suite under IRIS_GR4_JIT=sync pass. Half-alpha blends round down now (0x7F, was 0x80); the blend test expects that. Bilinear textured triangles: 68 -> 42 ns/px (5x the interpreter); trilinear 136 -> 110; shaded and depth-tested 17 -> 12. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Z registers are z.12 fixed point; depth now stays that way. The triangle setup turns the depth plane into a z.12 i64 plane at pixels (fixed::zplane, evaluated like the colour planes), GL lines step a z.12 base per pixel, and fixed::zbuf rounds to the 24-bit buffer value ((z + 2048) >> 12, clamped), which is what z.round().clamp() gave for every depth that can be stored. The JIT emits the same i64 arithmetic. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
S/W, T/W and 1/W are i64 planes at 2^32 (their registers' scale), evaluated like the colour and depth planes. Per fragment: - a reciprocal unit divides out W: normalise 1/W, a 1024-entry seed table and one Newton step (fixed::recip, good to about 2^-22), then s and t in Q31 by one 64x64->128 multiply each; - per level, texels in Q16 (s >> (15 - lw)): the integer part picks the texel, bits 15:8 are the bilinear weight; no float floor or saturating conversions; - the level of detail in Q8: Q16 texel derivatives, the larger sum of squares, a 256-entry log2 table (fixed::lod_q8), replacing the float cubic log2; it picks magnification, the nearest level and the trilinear fraction. A fragment whose 1/W is not positive is left untextured (before: only zero or non-finite). The JIT emits the same integer operations; the equivalence sweep and the mgras suite under IRIS_GR4_JIT=sync pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.
added mgras jit and converted raster pipeline to ints
moar speed