CUDA: fix im2col_3d to respect non-contiguous inputs (views)

[jakekarnes42]

Problem

IM2COL_3D fails on CUDA when the input is a non‑contiguous view (v=1). CPU backend passes. The CUDA kernel computed source addresses assuming compact layout (products of dims), ignoring nb[] strides.

Reproduction steps:

C:\Users\jakek\llama-dev\>git clone git@github.com:ggml-org/llama.cpp.git
C:\Users\jakek\llama-dev\llama.cpp>cd llama.cpp
C:\Users\jakek\llama-dev\llama.cpp>cmake -S . -B build-cuda -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=120 -DCMAKE_TOOLCHAIN_FILE=%VCPKG_ROOT%\scripts\buildsystems\vcpkg.cmake -DVCPKG_TARGET_TRIPLET=x64-windows -DCMAKE_FIND_PACKAGE_PREFER_CONFIG=ON
C:\Users\jakek\llama-dev\llama.cpp>cmake --build build-cuda --config Release --parallel
C:\Users\jakek\llama-dev\llama.cpp>build-cuda\bin\Release\llama-cli.exe --list-devices
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
Available devices:
  CUDA0: NVIDIA GeForce RTX 5090 (32606 MiB, 30841 MiB free)
C:\Users\jakek\llama-dev\llama.cpp>cd build-cuda
C:\Users\jakek\llama-dev\llama.cpp\build-cuda>ctest --output-on-failure -C Release -j 8
  ...
22: [IM2COL_3D] NMSE = 1.987277064 > 0.000000100 IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=1,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=1,d2=1,v=1): FAIL 
22: IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=3,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=1,d2=1,v=0): OK 
22: IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=3,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=1,d2=1,v=1): OK 
22: IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=1,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=1,d2=3,v=0): OK 
22: [IM2COL_3D] NMSE = 1.991386136 > 0.000000100 IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=1,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=1,d2=3,v=1): FAIL 
22: IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=3,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=1,d2=3,v=0): OK 
22: IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=3,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=1,d2=3,v=1): OK 
22: IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=1,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=3,d2=1,v=0): OK 
22: [IM2COL_3D] NMSE = 1.985351196 > 0.000000100 IM2COL_3D(type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,10,3],ne_kernel=[3,3,3,3],IC=1,s0=1,s1=3,s2=1,p0=0,p1=3,p2=3,d0=1,d1=3,d2=1,v=1): FAIL 
  ...
  13959/14471 tests passed
  Backend CUDA0: FAIL
Backend 2/2: CPU
  Skipping CPU backend
1/2 backends passed
FAIL


97% tests passed, 1 tests failed out of 30

Label Time Summary:
curl             =   0.73 sec*proc (1 test)
eval-callback    =   0.73 sec*proc (1 test)
main             =  99.60 sec*proc (27 tests)
model            =   0.16 sec*proc (2 tests)

Total Test time (real) =  70.36 sec

The following tests FAILED:
         22 - test-backend-ops (Failed)                         main
Errors while running CTest
C:\Users\jakek\llama-dev\llama.cpp\build-cuda>.\bin\Release\test-backend-ops.exe test -b CUDA0
  ...
  13959/14471 tests passed
  Backend CUDA0: FAIL
Backend 2/2: CPU
  Skipping
1/2 backends passed
FAIL
C:\Users\jakek\llama-dev\llama.cpp\build-cuda>.\bin\Release\test-backend-ops.exe test -b CPU
 ...
   14471/14471 tests passed
  Backend CPU: OK
2/2 backends passed
OK

Fix

This patch switches im2col_3d source indexing to use true strides derived from src1->nb[] (in elements), mirroring the approach used in the 2D CUDA im2col path. Destination indexing is unchanged.

After the patch, the tests pass for both CUDA and CPU backends.

C:\Users\jakek\llama-dev\llama.cpp>cmake --build build-cuda --config Release --parallel
C:\Users\jakek\llama-dev\llama.cpp>build-cuda\bin\Release\llama-cli.exe --list-devices
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
Available devices:
  CUDA0: NVIDIA GeForce RTX 5090 (32606 MiB, 30841 MiB free)
C:\Users\jakek\llama-dev\llama.cpp>cd build-cuda
C:\Users\jakek\llama-dev\llama.cpp\build-cuda>ctest --output-on-failure -C Release -j 8
Test project C:/Users/jakek/llama-dev/llama.cpp/build-cuda
      Start 22: test-backend-ops
      ...
30/30 Test #22: test-backend-ops ..................   Passed   71.29 sec

100% tests passed, 0 tests failed out of 30

Label Time Summary:
curl             =   0.79 sec*proc (1 test)
eval-callback    =   0.79 sec*proc (1 test)
main             = 102.28 sec*proc (27 tests)
model            =   0.17 sec*proc (2 tests)

Total Test time (real) =  71.31 sec
C:\Users\jakek\llama-dev\llama.cpp\build-cuda>.\bin\Release\test-backend-ops.exe test -b CUDA0
  ...
    14471/14471 tests passed
  Backend CUDA0: OK
Backend 2/2: CPU
  Skipping
2/2 backends passed
OK
C:\Users\jakek\llama-dev\llama.cpp\build-cuda>.\bin\Release\test-backend-ops.exe test -b CPU
 ...
   14471/14471 tests passed
  Backend CPU: OK
2/2 backends passed
OK

[jakekarnes42]

@JohannesGaessler - Thanks for the review. I agree with your suggestion to use ggml_element_size rather than a fixed value and I've applied that change.