From: cgmb@slerp.xyz   
      
   This is a multi-part message in MIME format.   
   Hi Helmut,   
      
   Thank you for the help. I'm afraid your patch has a bug in it, though.   
      
   On 2026-02-04 06:50, Helmut Grohne wrote:   
   > I looked into this and discovered that rocblas abuses nocheck. This   
   > seems to be a pattern in debian-ai and I have filed #1126934 to have   
   > lintian complain about this.   
      
   Ah. I'm not sure who introduced that pattern, but I copied it widely.   
   Apologies. While I did read the description of nocheck, my   
   interpretation was that source packages should not modify the contents   
   of the binary packages based on nocheck, but that they were free to skip   
   producing some binary packages at all. At least, that's how I reconciled   
   the description with examples I saw. Opps.   
      
   > The rocblas package did not disable the build of benchmarks in nocheck   
   > builds and therefore failed locating gtest there.   
      
   The rocblas-bench utility historically didn't use gtest. My guess would   
   be that if gtest is required, it's probably a minor mistake upstream in   
   their CMake. With that said, building the benchmarks without the tests   
   are probably a set of options that don't get exercised often, so from a   
   practical standpoint perhaps we should save ourselves some trouble and   
   avoid that pathway. I can see an argument that a benchmark utility is a   
   performance-testing tool, so I suppose it could reasonably be controlled   
   by the same options.   
      
   > I am attaching a patch that fixes both the abuse of nocheck introducing   
   > the noinsttest build profile and also fixes the build with both profiles   
   > or either.   
      
   When copying your solution to apply the same fix to hiprand, I found a   
   warning about a reference to an non-existant directory named "ON" from   
   CMake. A similar error can be found in the rocblas 6.4.4-4 buildd logs   
   due to "-DBUILD_CLIENTS_TESTS=ON ON":   
      
       dh_auto_configure -- -DCMAKE_BUILD_TYPE=Release   
       -DCMAKE_SKIP_INSTALL_RPATH=ON -DROCM_SYMLINK_LIBS=OFF   
       -DBUILD_CLIENTS_BENCHMARKS=ON   
       -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF   
       -DGPU_TARGETS="gfx803;gfx900;gfx906;gfx908;gfx90a;gfx1010;gf   
   1030;gfx1100;gfx1101;gfx1102"   
       -DBUILD_WITH_HIPBLASLT=OFF   
       -DROCBLAS_TENSILE_LIBRARY_DIR=/usr/lib/aarch64-linux-gnu/rocblas/4.4.1   
       -DINSTALL_TEST_DATA_DIR=/usr/share/librocblas4-tests/data   
       -DTensile_COMPILER=hipcc -DTensile_TEST_LOCAL_PATH="../tensile"   
       -DTensile_ROOT="/build/reproducible-path/rocblas-6.4.4/tensile/Tensile"   
       -DTensile_LOGIC=asm_full -DTensile_LIBRARY_FORMAT=msgpack   
       -DTensile_LAZY_LIBRARY_LOADING=ON   
       -DTensile_SEPARATE_ARCHITECTURES=ON -DBUILD_CLIENTS_TESTS=ON ON    
       -DBUILD_WITH_PIP=false -DOptions=--ignore-asm-cap-cache   
       -DRUN_HEADER_TESTING=OFF -DBUILD_OFFLOAD_COMPRESS=OFF   
       ...   
       CMake Warning:   
          Ignoring extra path from command line:   
      
         "/build/reproducible-path/rocblas-6.4.4/obj-aarch64-linux-gnu/ON"   
      
   Sincerely,   
   Cory Bloor   
      
      
      
        
          
        
        
       
Hi Helmut,
   
       
Thank you for the help. I'm afraid your patch has a bug in it,   
         though.
   
       
On 2026-02-04 06:50, Helmut Grohne   
         wrote:
   
       
   
       
   
         
I looked into this and discovered   
   that rocblas abuses nocheck. This   
   seems to be a pattern in debian-ai and I have filed #1126934 to have   
   lintian complain about this.
   
       
   
       
Ah. I'm not sure who introduced that pattern, but I copied it   
         widely. Apologies. While I did read the description of nocheck, my   
         interpretation was that source packages should not modify the   
         contents of the binary packages based on nocheck, but that they   
         were free to skip producing some binary packages at all. At least,   
         that's how I reconciled the description with examples I saw. Opps.
   
       
   
         
The rocblas package did not disable   
   the build of benchmarks in nocheck   
   builds and therefore failed locating gtest there.
   
       
   
       
The rocblas-bench utility historically didn't use gtest. My guess   
         would be that if gtest is required, it's probably a minor mistake   
         upstream in their CMake. With that said, building the benchmarks   
         without the tests are probably a set of options that don't get   
         exercised often, so from a practical standpoint perhaps we should   
         save ourselves some trouble and avoid that pathway. I can see an   
         argument that a benchmark utility is a performance-testing tool,   
         so I suppose it could reasonably be controlled by the same   
         options.
   
       
   
         
I am attaching a patch that fixes   
   both the abuse of nocheck introducing   
   the noinsttest build profile and also fixes the build with both profiles   
   or either.   
   
   
       
   
       
When copying your solution to apply the same fix to hiprand, I   
         found a warning about a reference to an non-existant directory   
         named "ON" from CMake. A similar error can be found in the rocblas   
         6.4.4-4 buildd logs due to "-DBUILD_CLIENTS_TESTS=ON ON":
   
      
   [continued in next message]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)