A specialized Winograd Conv2d op#971

Closed

bssrdf wants to merge 23 commits intoggml-org:masterfrom

bssrdf:add-winograd-conv2d

Commits on Sep 26, 2024

added two winograd ops
bssrdf
committed

Commits on Sep 27, 2024

added more checking for precondition to use winograd; switch to im2co if not satisfied
bssrdf
committed
added source code for winograd kernel
bssrdf
committed
winograd build ok
bssrdf
committed
test now passed; for some reason, ggml_conv_2d didn't output correct results
bssrdf
committed
remove debugging printouts
bssrdf
committed

Commits on Sep 28, 2024

added a FP16 FX kernel to deal with fp16 filter data; no need to use FP32 buffer
bssrdf
committed

Commits on Sep 29, 2024

sync from sd.cpp
bssrdf
committed
restore test-conv2d.cpp test
bssrdf
committed
restore src/CMakeLists.txt
bssrdf
committed
change mask to unsigned int; add __restric__ to various pointers
bssrdf
committed
fix indentation
bssrdf
committed

Commits on Oct 1, 2024

skip if already computed in a preprocess step
bssrdf
committed
added winograd conv2d to backend op tests
bssrdf
committed

Commits on Oct 2, 2024

add conv2d as a test case in test-backend-op
bssrdf
committed

Commits on Dec 18, 2024

further refine super tile size; implement a proper performance comparison
bssrdf
committed

Commits on Dec 25, 2024

fix test case
bssrdf
committed

Commits on Dec 26, 2024

changing #define macros to template arguments, step 1
bssrdf
committed
sync. winograd impl. from sd.cpp
bssrdf
committed

Commits on Dec 30, 2024

test performance for configs from sd.cpp
bssrdf
committed

Commits on Jan 1, 2025

refactor test-conv2d-winograd
bssrdf
committed

Commits on Jan 5, 2025

add the missing ggml_gallocr_free(allocr)
bssrdf
committed

Commits on Jan 6, 2025

get rid of #defines by using template arguments
bssrdf
committed