diffuse or sharpen CPU performance tuning#20468

Draft

kofa73 wants to merge 19 commits intodarktable-org:masterfrom

kofa73:diffuse-perf

Commits on Mar 11, 2026

diffuse: inline structural convolution matrices for 45.631 CPU execution sum
kofa73
committed
diffuse: scalar gradient fetching optimization for 40.958s execution sum
kofa73
committed
diffuse: spatial lookup loop unroll for 40.024s execution sum
kofa73
committed
housekeeping files
kofa73
committed
diffuse: eliminate intermediate neighbour pixel arrays for 39.215s execution sum
kofa73
committed
diffuse: squared-magnitude angle computation for 38.847s execution sum
kofa73
committed
diffuse: paired convolution loops for shared angle data for 36.877s execution sum
kofa73
committed
Add diffuse integration tests and anisotropy code path documentation
kofa73
committed
diffuse: unroll derivative accumulation loop for 36.231s execution sum
kofa73
committed
diffuse: merge variance regularization into output loop for 35.972s execution sum
kofa73
committed
diffuse: remove 0.5f scaling from gradient/laplacian, absorb into half_anisotropy for 35.599s execution sum
kofa73
committed
diffuse: record failed shared diag intermediates experiment
kofa73
committed
diffuse: record failed pre-halve cs experiment and sync tracking files
kofa73
committed
diffuse: record failed a11+a22 identity, neg_magnitude, center reuse, and isotropic precompute experiments
kofa73
committed
diffuse: reorder sums loop before angle loop for L1 cache locality, yielding 35.224s execution sum
kofa73
committed
diffuse: record failed fast-math pragma and fused dt_vector_exp experiments
kofa73
committed
perf: Reduced baseline from 35.224s to 33.355s (~5.3% drop). Fused angle, ex
kofa73
committed
diffuse: removed unrelated files
kofa73
committed
diffuse: updated progress
kofa73
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

diffuse or sharpen CPU performance tuning#20468

diffuse or sharpen CPU performance tuning#20468
kofa73 wants to merge 19 commits intodarktable-org:masterfrom
kofa73:diffuse-perf

Commits on Mar 11, 2026

diffuse: inline structural convolution matrices for 45.631 CPU execution sum

diffuse: scalar gradient fetching optimization for 40.958s execution sum

diffuse: spatial lookup loop unroll for 40.024s execution sum

housekeeping files

diffuse: eliminate intermediate neighbour pixel arrays for 39.215s execution sum

diffuse: squared-magnitude angle computation for 38.847s execution sum

diffuse: paired convolution loops for shared angle data for 36.877s execution sum

Add diffuse integration tests and anisotropy code path documentation

diffuse: unroll derivative accumulation loop for 36.231s execution sum

diffuse: merge variance regularization into output loop for 35.972s execution sum

diffuse: remove 0.5f scaling from gradient/laplacian, absorb into half_anisotropy for 35.599s execution sum

diffuse: record failed shared diag intermediates experiment

diffuse: record failed pre-halve cs experiment and sync tracking files

diffuse: record failed a11+a22 identity, neg_magnitude, center reuse, and isotropic precompute experiments

diffuse: reorder sums loop before angle loop for L1 cache locality, yielding 35.224s execution sum

diffuse: record failed fast-math pragma and fused dt_vector_exp experiments

perf: Reduced baseline from 35.224s to 33.355s (~5.3% drop). Fused angle, ex

diffuse: removed unrelated files

diffuse: updated progress