Eval bug: Image processing on Metal takes significant amount of time

[qnixsynapse]

Name and Version

Reproducible on Master

Operating systems

Mac

GGML backends

Metal

Hardware

Apple Mac M2 silicon

Models

gemma-3-4B-it (Any gemma model)

Problem description & steps to reproduce

The image encoding on Gemma-3 models seems to take significant amount of time (2-5 mins) because of missing im2col kernel implementation.

Unfortunately, I am not well versed in Metal code. Can any kind person take a look at it and add support for necessary conditions?

Please use test-backend-ops -b Metal -o IM2COL to see which conditions are missing.

First Bad Commit

No response

Relevant log output

Backend 1/3: Metal
  Device description: Apple M2 Pro
  Device memory: 21845 MB (21839 MB free)

  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0): OK
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=0,p1=0,d0=1,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=0,p1=0,d0=3,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=3,p1=0,d0=1,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=3,p1=0,d0=3,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=0,p1=0,d0=1,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=0,p1=0,d0=3,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=3,p1=0,d0=1,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=3,p1=0,d0=3,d1=0,is_2D=0): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=1,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=1,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=3,d1=1,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=3,d1=3,is_2D=1): not supported [Metal]
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,32],ne_kernel=[3,3,1,32],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,32],ne_kernel=[3,3,2,32],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,1024],ne_kernel=[3,3,1,1024],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,1024],ne_kernel=[3,3,2,1024],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,2048],ne_kernel=[3,3,1,2048],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,2048],ne_kernel=[3,3,2,2048],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,2560],ne_kernel=[3,3,1,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,2560],ne_kernel=[3,3,2,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1): OK
  IM2COL(type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[5,5,1,32],ne_kernel=[3,4,1,32],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1): OK
  10844/10844 tests passed
  Backend Metal: OK
ggml_metal_free: deallocating
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
ggml_metal_mem_pool_free: freeing memory pool, num heaps = 0 (total = 0)
Backend 2/3: BLAS
  Skipping
Backend 3/3: CPU
  Skipping
3/3 backends passed
OK

[ngxson]

I think it's likely due to this condition of src and dst type:

llama.cpp/ggml/src/ggml-metal/ggml-metal.m

Lines 4706 to 4708 in 1bc664a

	GGML_ASSERT(src0->type == GGML_TYPE_F16);
	GGML_ASSERT(src1->type == GGML_TYPE_F32);
	GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);

Combined with the fact that ggml_conv_2d force the output type to be the same as dtype of the kernel:

llama.cpp/ggml/src/ggml.c

Line 4348 in 1bc664a

struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, a->type); // [N, OH, OW, IC * KH * KW]

Interestingly, the kernel is generally f32 so it won't be supported by Metal.

I'm thinking of a quick fix is to make a custom version of ggml_conv_2d in libmtmd, will draft a PR for this.

[ngxson]

Hmm no, seems like it's more simple that that: just the condition of ggml_metal_supports_op is not matching what's inside the kernel code

[ngxson]

The im2col is now run on Metal. But unfortunately, it still doesn't improve the pref. Need to dig deeper...

[qnixsynapse]

Interesting. I have approved that PR. I will be keeping an eye on this.

Thank you!

[qnixsynapse]

Reopening this because the image encoding still takes a significant amount of time compared to other backends.

image slice encoded in 299527 ms

That's almost 5 mins on Apple M3 Pro 16 GB unified memory

[altunenes]

same problem. Tried both Vulkan + Metal
Image encoding takes too much time. like 30 seconds. :(