upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-10-03 03:17:51 -05:00
Code Issues Packages Projects Releases Wiki Activity

perf: accelerate VAE direct 3D convolutions (#1996)

Browse Source
This commit is contained in:
leejet
2026-09-19 17:50:01 +08:00
committed by GitHub
parent 2ea8aff7ef
commit 3e037a81e4
9 changed files with 24 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/core/ggml_extend.cpp
+7 -2
View File
@@ -452,8 +452,13 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
int d0,
int d1,
int d2,
bool force_prec_f32) {
if (force_prec_f32) {
bool force_prec_f32,
bool direct) {
if (direct) {
int64_t OC = w->ne[3] / IC;
int64_t N = x->ne[3] / IC;
x = ggml_conv_3d_direct(ctx, w, x, s0, s1, s2, p0, p1, p2, d0, d1, d2, (int)IC, (int)N, (int)OC);
} else if (force_prec_f32) {
ggml_tensor* im2col = ggml_im2col_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, w->type);
int64_t OC = w->ne[3] / IC;