Faster swizzle code #1608
Open
Faster swizzle code #1608
+51
−42
Conversation
This file contains 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
|
run time in seconds for each implementation (using MSVC): int main() {
int width = 2048;
int height = 2048;
int bpp = 1;
int pitch = bpp * width;
uint8_t* garbage_in = (uint8_t*)malloc(width * height * bpp);
uint8_t* garbage_out = (uint8_t*)malloc(width * height * bpp);
for (int i = 0; i < 1000; i++) {
swizzle_rect(garbage_in, width, height, garbage_out, pitch, bpp);
unswizzle_rect(garbage_out, width, height, garbage_in, pitch, bpp);
}
} |
mborgerson
reviewed
Mar 20, 2024
There was a problem hiding this comment.
Looks like a big improvement in this swizzle logic performance, nice! In addition to the minor feedback on code, a couple things:
- Similar to the benchmark, please also test to confirm that the swizzle/unswizzle results are the same in all cases: current impl, new faster sw impl, and new instruction impl
- Please determine how well this instruction will be supported on user systems. Find out which generation of x86 processors includes this instruction, and if it is relatively commonplace now, confirm that our release builds are actually building for it (related: Enable additional architectural optimizations in release #1492)
hw/xbox/nv2a/swizzle.c
Outdated
| mask = (mask ^ mv) | (mv >> (1 << i)); // Compress m. | ||
| mk = mk & ~mp; | ||
| } | ||
|
|
hw/xbox/nv2a/swizzle.c
Outdated
| generate_swizzle_masks(width, height, depth, &mask_x, &mask_y, &mask_z); | ||
| expand_mask mask_x, mask_y, mask_z; | ||
| generate_swizzle_masks(width, height, depth, &mask_x.mask, &mask_y.mask, &mask_z.mask); | ||
| generate_expand_mask_moves(&mask_x); |
There was a problem hiding this comment.
Calculation unnecessary in case of pdep instruction?
hw/xbox/nv2a/swizzle.c
Outdated
| * expand(0000abcd, 10011010) = a00bc0d0. | ||
| * | ||
| * Implementation from Hacker's delight chapter 7 "Expand" | ||
| * https://stackoverflow.com/questions/77834169/what-is-a-fast-fallback-algorithm-which-emulates-pdep-and-pext-in-software |
There was a problem hiding this comment.
Citing the book should be enough, don't need the stackoverflow.com link
hw/xbox/nv2a/swizzle.c
Outdated
| } | ||
|
|
||
| static uint32_t expand(uint32_t x, expand_mask* expand_mask) { | ||
| #ifdef __BMI2__ |
There was a problem hiding this comment.
nit: Readability would be better with these macros unindented
mborgerson
reviewed
Mar 20, 2024
hw/xbox/nv2a/swizzle.c
Outdated
| uint32_t moves[5]; | ||
| } expand_mask; | ||
|
|
||
| static void generate_expand_mask_moves(expand_mask* expand_mask) { |
There was a problem hiding this comment.
nit: Asterisk goes next to identifier
based on https://fgiesen.wordpress.com/2011/01/17/texture-tiling-and-swizzling/ - Increment swizzled offsets, instead of calculating them from scratch for each pixel. This seems to be around 30X faster. - Add/update comments, including high-level comment on what swizzle.c is for. This improves FPS in certain areas in Fable: The Lost Chapters
NZJenkins
force-pushed
the
quick_swizzle
branch
from
af61911 to
2d8596c
Compare
|
I've changed the implementation to get the same performance without BMI2, and have updated the PR description |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
This speeds up swizzling/unswizzling textures.
fill_patternshowed up in profiling Fable: The Lost Chapters.It seems to implement the equivalent of thepdepinstruction.It turns out there is a slightly faster method to do this in a fixed number of steps.
This increases fps a bit in certain scenes in Fable TLC.
Also use nativepdepif available.Instead of calculating these offsets from masks, we can increment through the offsets.
See https://fgiesen.wordpress.com/2011/01/17/texture-tiling-and-swizzling/
I wrote a small testbench for different methods to ensure it's working the same way (not integrated into xemu)
It should be around 30X faster and is matching both swizzle/unswizzle behaviour of the original code.
Gives a few more FPS in certain situations in Fable TLC:
Fable TLC FPS comparison
Before:
After:
