GPUEngine.cu does not compile with CUDA 13: cudaDeviceProp::computeMode was removed

[Dileep-Kumar-5]

Summary

GPU/GPUEngine.cu does not compile against CUDA 13. cudaDeviceProp::computeMode was removed from the struct, and PrintCudaInfo() reads it.

Error

GPU/GPUEngine.cu(371): error: class "cudaDeviceProp" has no member "computeMode"
        sComputeMode[deviceProp.computeMode]);
                                ^

1 error detected in the compilation of "GPU/GPUEngine.cu".

This is the only source-level error nvcc 13.3.73 reports; with it fixed, GPUEngine.cu compiles clean.

Cause

The field is gone from cudaDeviceProp in CUDA 13. The cudaDevAttrComputeMode device attribute still exists (driver_types.h: cudaDevAttrComputeMode = 20), so the information is still available — just not through the struct.

The use is purely informational, one field of the startup banner.

Fix

     cudaDeviceProp deviceProp;
     cudaGetDeviceProperties(&deviceProp,i);
+
+    // cudaDeviceProp::computeMode was removed in CUDA 13; the device
+    // attribute still exists, so query that instead of the struct field.
+    int computeMode = 0;
+    cudaDeviceGetAttribute(&computeMode,cudaDevAttrComputeMode,i);
+    if(computeMode < 0 || computeMode > 3) computeMode = 4;   // "Unknown"
+
     printf("GPU #%d %s (%dx%d cores) (Cap %d.%d) (%.1f MB) (%s)\n",
       i,deviceProp.name,deviceProp.multiProcessorCount,
       _ConvertSMVer2Cores(deviceProp.major,deviceProp.minor),
       deviceProp.major,deviceProp.minor,(double)deviceProp.totalGlobalMem / 1048576.0,
-      sComputeMode[deviceProp.computeMode]);
+      sComputeMode[computeMode]);

The bounds guard matters because sComputeMode[] has 5 entries and nothing constrains the attribute to that range.

cudaDeviceGetAttribute is long-standing API, so this also builds on older toolkits — no version guard needed.

Environment

• CUDA 13.3.73, MSVC 14.44, Windows, sm_86 (RTX 3050 Laptop)
• Verified after the fix: builds clean, -l prints the device correctly, and solves the sample in.txt key

Note

This is independent of everything else and stands alone. It is also included in PR #158, but that PR carries a breaking format change you may not want — this one-hunk fix is worth taking on its own regardless, since without it nobody can build the GPU version against a current toolkit.

Unrelated to the compile error, while in this file: _ConvertSMVer2Cores() has no entry for sm_86, so the banner prints (16x0 cores) on Ampere consumer cards. Cosmetic, but easy to fold in if you touch this function.

🤖 Generated with Claude Code

[Dileep-Kumar-5]

Correcting something I wrote in the original report: I called the missing sm_86 entry in _ConvertSMVer2Cores() cosmetic. It isn't.

GetGridSize() uses it to pick the default grid:

if(*y <= 0) *y = 2 * _ConvertSMVer2Cores(deviceProp.major,deviceProp.minor);
if(*y <= 0) *y = 128;

The table stops at sm_75, so Ampere and newer fall through to return 0, *y becomes 0, and the run silently lands on the 128 fallback. On an sm_86 card the default grid Y is therefore 128 instead of the intended 2 x 128 = 256 — half the thread groups. No error, it just quietly runs a smaller grid than the code intends.

Measured on an RTX 3050 Laptop, same binary, grid forced with -g, back to back:

grid	throughput
32x128 (what you get today)	338.9 MK/s
32x256 (with sm_86 in the table)	345.5 MK/s (+2.0%)

So the effect is real but small — worth having for correctness more than for speed, and users can already work around it with -g. Kangaroo memory doubles with the grid, 57 -> 105 MB.

Fix is the same shape as the computeMode one, just table entries for sm_80/86/87/89/90 and Blackwell. Included in #158 as a separate commit, and equally independent of that PR's format change.

[Dileep-Kumar-5]

If you just need a GPU build that compiles against CUDA 13 today, a branch with this fix applied is here:

https://github.com/Dileep-Kumar-5/Kangaroo (branch feat/256bit-distance-field, also the default branch)

It carries the computeMode fix and the sm_86 table fix as two separate, self-contained commits, so they are easy to cherry-pick if you would rather not take the rest:

git remote add dk https://github.com/Dileep-Kumar-5/Kangaroo.git
git fetch dk
git cherry-pick e4a2d8e   # CUDA 13: cudaDeviceProp::computeMode removed
git cherry-pick 4276449   # _ConvertSMVer2Cores: Ampere and newer

Neither touches the distance width or any file/wire format, so cherry-picking them onto stock master leaves everything else byte-identical to upstream.

Note e4a2d8e also adjusts build_gpu.bat, a Windows helper that does not exist upstream — on Linux the source change is the only part that matters and the Makefile is untouched.

Built and run with CUDA 13.3.73 / MSVC 14.44 on an RTX 3050 Laptop: solves the sample in.txt key correctly, and the banner reads (16x128 cores) with a Grid(32x256) default instead of (16x0 cores) / Grid(32x128).

Status, to be clear: my branch, unreviewed, #158 is still open. If the maintainer takes these, use upstream instead.