Misc. bug: CONV_2D test cases failed on CUDA backend

[bssrdf]

Name and Version

ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
version: 7839 (8f80d1b)
built with MSVC 19.39.33522.0 for x64

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

Test code

Command line

test-backend-ops.exe test -o CONV_2D

Problem description & steps to reproduce

20 test cases failed to pass for CONV_2D op on cuda backend.

First Bad Commit

No response

Relevant log output

Logs

Failing tests:
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  CONV_2D(ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0)
  Backend CUDA0: FAIL

[am17an]

Are the differences very large or can this be fixed by adjusting the max_nmse_err?

[bssrdf]

Are the differences very large or can this be fixed by adjusting the max_nmse_err?

The failure is not related to the precision. We got some diagnosing done at #15805 (comment)

[bssrdf]

This is fixed by #19165