Skip to content

Sigmoid mojo rewrite #8

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
maxchisto wants to merge 6 commits into
base: sigmoid-mojo
Choose a base branch
from
Open

Sigmoid mojo rewrite #8

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
94e3462
working attempt at rewriting sigmoid gpu stays at 100%
maxchisto Mar 26, 2026
90ec744
detele unused bits
maxchisto Apr 15, 2026
920af7a
delete more unused bits
maxchisto Apr 15, 2026
c69e3e6
and even more
maxchisto Apr 15, 2026
d5355f7
docs
maxchisto Apr 15, 2026
ecc268c
restore submodules
maxchisto Apr 16, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions mojo/Makefile
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
MOJO = pixi run mojo
PLUGIN_DIR = /usr/lib/darktable/plugins

.PHONY: all build install bench clean

all: build

build: libsigmoid_mojo.so

libsigmoid_mojo.so: iop/sigmoid/lib.mojo iop/sigmoid/kernels.mojo
$(MOJO) build -I . iop/sigmoid/lib.mojo --emit shared-lib -o libsigmoid_mojo.so

install: libsigmoid_mojo.so
sudo cp libsigmoid_mojo.so $(PLUGIN_DIR)/

clean:
rm -f libsigmoid_mojo.so
51 changes: 51 additions & 0 deletions mojo/build_sigmoid_iop.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,51 @@
#!/bin/bash
# Build libsigmoid.so from the darktable source tree.
# This mimics what CMake does (run introspection, then compile + link).

set -e

SRC=/home/mc/code/darktable
BUILD=/tmp/sigmoid_build
DEST=$BUILD/libsigmoid.so

mkdir -p "$BUILD"

# Step 1: Run introspection (generates introspection_sigmoid.c)
echo "=== Step 1: Introspection ==="
perl "$SRC/tools/introspection/parser.pl" \
"$SRC/src/" \
"$SRC/src/iop/sigmoid.c" \
"$BUILD/introspection_sigmoid.c"
echo " -> $BUILD/introspection_sigmoid.c ($(wc -l < "$BUILD/introspection_sigmoid.c") lines)"

# Step 2: Collect flags
GTK_FLAGS=$(pkg-config --cflags gtk+-3.0 glib-2.0 librsvg-2.0 lcms2 json-glib-1.0)
DT_INCLUDE="-I$SRC/src -I$SRC/src/iop -I$SRC/build/bin -I/usr/lib/darktable"
DEFINES="-DHAVE_CONFIG_H -DHAVE_OPENCL -D_GNU_SOURCE -include common/module_api.h -include iop/iop_api.h"
CFLAGS="-O3 -march=native -fPIC -fvisibility=hidden -fopenmp $GTK_FLAGS $DT_INCLUDE $DEFINES"
LDFLAGS="-L/usr/lib/darktable -ldarktable -lm -lgomp -Wl,-rpath,/usr/lib/darktable"

echo ""
echo "=== Step 2: Compile ==="
gcc $CFLAGS \
-Wno-unused-function \
-Wno-deprecated-declarations \
-c "$BUILD/introspection_sigmoid.c" \
-o "$BUILD/introspection_sigmoid.o" 2>&1
echo " -> $BUILD/introspection_sigmoid.o"

echo ""
echo "=== Step 3: Link ==="
gcc -shared -fPIC -fopenmp \
"$BUILD/introspection_sigmoid.o" \
$LDFLAGS \
-o "$DEST"
echo " -> $DEST"

echo ""
echo "=== Step 4: Verify ==="
nm -D "$DEST" | grep " T " | awk '{print $3}'

echo ""
echo "Build complete: $DEST"
echo "Size: $(du -sh "$DEST" | cut -f1)"
1 change: 1 addition & 0 deletions mojo/iop/__init__.mojo
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
# iop package
1 change: 1 addition & 0 deletions mojo/iop/sigmoid/__init__.mojo
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
# iop.sigmoid package
281 changes: 281 additions & 0 deletions mojo/iop/sigmoid/kernels.mojo
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,281 @@
from std.math import isnan, sqrt, pow, max, min

@always_inline
fn generalized_loglogistic_sigmoid_scalar(
value: Float32,
magnitude: Float32,
paper_exp: Float32,
film_fog: Float32,
film_power: Float32,
paper_power: Float32,
) -> Float32:
var clamped_value = max(value, Float32(0.0))
var film_response = pow(film_fog + clamped_value, film_power)
var paper_response = magnitude * pow(
film_response / (paper_exp + film_response), paper_power
)
if isnan(paper_response):
return magnitude
return paper_response

@always_inline
fn apply_sigmoid_rgb_ratio(
in_r: Float32,
in_g: Float32,
in_b: Float32,
in_a: Float32,
white_target: Float32,
black_target: Float32,
paper_exp: Float32,
film_fog: Float32,
film_power: Float32,
paper_power: Float32,
) -> SIMD[DType.float32, 4]:
# Desaturate negative values
var avg = max((in_r + in_g + in_b) / 3.0, Float32(0.0))
var min_v = min(min(in_r, in_g), in_b)
var sat = Float32(1.0)
if min_v < 0.0:
sat = -avg / (min_v - avg)

var p_r = avg + sat * (in_r - avg)
var p_g = avg + sat * (in_g - avg)
var p_b = avg + sat * (in_b - avg)

var luma = (p_r + p_g + p_b) / 3.0
var mapped_luma = generalized_loglogistic_sigmoid_scalar(
luma, white_target, paper_exp, film_fog, film_power, paper_power
)

if luma > 1e-9:
var scale = mapped_luma / luma
p_r *= scale
p_g *= scale
p_b *= scale
else:
p_r = mapped_luma
p_g = mapped_luma
p_b = mapped_luma

var p_min = min(min(p_r, p_g), p_b)
var p_max = max(max(p_r, p_g), p_b)
var eps = Float32(1e-6)
var d_white = (white_target - mapped_luma) / (p_max - mapped_luma + eps)
var d_black = (black_target - mapped_luma) / (p_min - mapped_luma - eps)
var db_vs_chroma = min(d_white, d_black)
var cvm_border = (mapped_luma - p_min) / (mapped_luma + eps)
var p_chr_adj = 1.0 / (cvm_border * db_vs_chroma + eps)
var h_chr = (
2.0 * cvm_border / (1.0 - cvm_border * cvm_border + eps)
) * p_chr_adj
var h_z = sqrt(h_chr * h_chr + 1.0)
var chroma_f = h_chr / (1.0 + h_z) * db_vs_chroma

return SIMD[DType.float32, 4](
mapped_luma + chroma_f * (p_r - mapped_luma),
mapped_luma + chroma_f * (p_g - mapped_luma),
mapped_luma + chroma_f * (p_b - mapped_luma),
in_a,
)

@always_inline
fn apply_sigmoid_per_channel(
in_r: Float32,
in_g: Float32,
in_b: Float32,
in_a: Float32,
white_target: Float32,
paper_exp: Float32,
film_fog: Float32,
contrast_power: Float32,
skew_power: Float32,
hue_preservation: Float32,
pipe_to_base: SIMD[DType.float32, 16],
base_to_rendering: SIMD[DType.float32, 16],
rendering_to_pipe: SIMD[DType.float32, 16],
) -> SIMD[DType.float32, 4]:
# 1. Transform to base space
var i_r = (
pipe_to_base[0] * in_r + pipe_to_base[1] * in_g + pipe_to_base[2] * in_b
)
var i_g = (
pipe_to_base[4] * in_r + pipe_to_base[5] * in_g + pipe_to_base[6] * in_b
)
var i_b = (
pipe_to_base[8] * in_r
+ pipe_to_base[9] * in_g
+ pipe_to_base[10] * in_b
)

# 2. Desaturate negative
var avg = max((i_r + i_g + i_b) / 3.0, Float32(0.0))
var min_v = min(min(i_r, i_g), i_b)
var sat = Float32(1.0)
if min_v < 0.0:
sat = -avg / (min_v - avg)
i_r = avg + sat * (i_r - avg)
i_g = avg + sat * (i_g - avg)
i_b = avg + sat * (i_b - avg)

# 3. Transform to rendering space
var r_r = (
base_to_rendering[0] * i_r
+ base_to_rendering[1] * i_g
+ base_to_rendering[2] * i_b
)
var r_g = (
base_to_rendering[4] * i_r
+ base_to_rendering[5] * i_g
+ base_to_rendering[6] * i_b
)
var r_b = (
base_to_rendering[8] * i_r
+ base_to_rendering[9] * i_g
+ base_to_rendering[10] * i_b
)

# 4. Per-channel sigmoid curves
var pc_r = generalized_loglogistic_sigmoid_scalar(
r_r, white_target, paper_exp, film_fog, contrast_power, skew_power
)
var pc_g = generalized_loglogistic_sigmoid_scalar(
r_g, white_target, paper_exp, film_fog, contrast_power, skew_power
)
var pc_b = generalized_loglogistic_sigmoid_scalar(
r_b, white_target, paper_exp, film_fog, contrast_power, skew_power
)

# 5. Preserve hue & energy
var p_min: Float32
var p_mid: Float32
var p_max: Float32
var pc_min: Float32
var pc_mid: Float32
var pc_max: Float32

if r_r >= r_g:
if r_g >= r_b: # R G B
p_max = r_r
p_mid = r_g
p_min = r_b
pc_max = pc_r
pc_mid = pc_g
pc_min = pc_b
elif r_b >= r_r: # B R G
p_max = r_b
p_mid = r_r
p_min = r_g
pc_max = pc_b
pc_mid = pc_r
pc_min = pc_g
else: # R B G
p_max = r_r
p_mid = r_b
p_min = r_g
pc_max = pc_r
pc_mid = pc_b
pc_min = pc_g
else:
if r_r >= r_b: # G R B
p_max = r_g
p_mid = r_r
p_min = r_b
pc_max = pc_g
pc_mid = pc_r
pc_min = pc_b
elif r_b >= r_g: # B G R
p_max = r_b
p_mid = r_g
p_min = r_r
pc_max = pc_b
pc_mid = pc_g
pc_min = pc_r
else: # G B R
p_max = r_g
p_mid = r_b
p_min = r_r
pc_max = pc_g
pc_mid = pc_b
pc_min = pc_r

var chroma = p_max - p_min
var midscale = Float32(0.0)
if chroma != 0.0:
midscale = (p_mid - p_min) / chroma

var f_hc = pc_min + (pc_max - pc_min) * midscale
var n_mid = (1.0 - hue_preservation) * pc_mid + hue_preservation * f_hc

var blend = 2.0 * p_min / (p_min + p_mid + 1e-9)
var target = blend * (pc_r + pc_g + pc_b) + (1.0 - blend) * (
pc_min + n_mid + pc_max
)

var res_min: Float32
var res_mid: Float32
var res_max: Float32
if n_mid <= pc_mid:
res_mid = (
(1.0 - hue_preservation) * pc_mid
+ hue_preservation
* (midscale * pc_max + (1.0 - midscale) * (target - pc_max))
) / (1.0 + hue_preservation * (1.0 - midscale))
res_min = target - pc_max - res_mid
res_max = pc_max
else:
res_mid = (
(1.0 - hue_preservation) * pc_mid
+ hue_preservation
* (pc_min * (1.0 - midscale) + midscale * (target - pc_min))
) / (1.0 + hue_preservation * midscale)
res_min = pc_min
res_max = target - pc_min - res_mid

var res_r: Float32
var res_g: Float32
var res_b: Float32
if r_r >= r_g:
if r_g >= r_b: # R G B
res_r = res_max
res_g = res_mid
res_b = res_min
elif r_b >= r_r: # B R G
res_b = res_max
res_r = res_mid
res_g = res_min
else: # R B G
res_r = res_max
res_b = res_mid
res_g = res_min
else:
if r_r >= r_b: # G R B
res_g = res_max
res_r = res_mid
res_b = res_min
elif r_b >= r_g: # B G R
res_b = res_max
res_g = res_mid
res_r = res_min
else: # G B R
res_g = res_max
res_b = res_mid
res_r = res_min

# 6. Transform to pipe space
var out_r = (
rendering_to_pipe[0] * res_r
+ rendering_to_pipe[1] * res_g
+ rendering_to_pipe[2] * res_b
)
var out_g = (
rendering_to_pipe[4] * res_r
+ rendering_to_pipe[5] * res_g
+ rendering_to_pipe[6] * res_b
)
var out_b = (
rendering_to_pipe[8] * res_r
+ rendering_to_pipe[9] * res_g
+ rendering_to_pipe[10] * res_b
)

return SIMD[DType.float32, 4](out_r, out_g, out_b, in_a)
Loading