[Documentation]: No Clear path for TheRock ROCm Bootstrapping

[daryltucker]

Unable to Determine Targets or Families

In the scenario that you're trying to build ROCm without an existing instance of ROCm, I found it incredibly difficult to determine a proper THEROCK_AMDGPU_FAMILIES or THEROCK_AMDGPU_TARGETS.

All documentation suggests using rocminfo to determine these values for your hardware. Though I found it bizarre I couldn't just target "all", I tried determining this by compiling rocminfo.

This has been moved to rocm-systems/projects/rocminfo.

cd rocm-systems/projects/rocminfo
mkdir -p build
cd build

$ cmake -DCMAKE_PREFIX_PATH=/opt/rocm ..
CMake Deprecation Warning at CMakeLists.txt:37 (cmake_minimum_required):
  Compatibility with CMake < 3.10 will be removed from a future version of
  CMake.
  Update the VERSION argument <min> value.  Or, use the <min>...<max> syntax
  to tell CMake that the project requires at least <min> but has been updated
  to work with policies introduced by <max> or earlier.

CMake Error at CMakeLists.txt:66 (find_package):
  By not providing "Findhsa-runtime64.cmake" in CMAKE_MODULE_PATH this
  project has asked CMake to find a package configuration file provided by
  "hsa-runtime64", but CMake did not find one.
  Could not find a package configuration file provided by "hsa-runtime64"
  (requested version 1.0) with any of the following names:
    hsa-runtime64Config.cmake
    hsa-runtime64-config.cmake
  Add the installation prefix of "hsa-runtime64" to CMAKE_PREFIX_PATH or set
  "hsa-runtime64_DIR" to a directory containing one of the above files.  If
  "hsa-runtime64" provides a separate development package or SDK, be sure it
  has been installed.

-- Configuring incomplete, errors occurred!

and then I thought I found where to build libhsakmt via rocr-runtime ...

[  7%] Building C object libhsakmt/CMakeFiles/hsakmt.dir/src/topology.c.o
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cc -DHAVE_MEMFD_CREATE=1 -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/include -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src -I/usr/include/libdrm -I/usr/include -I/usr/include/libdrm -fPIC -W -Wall -Wextra -Wno-unused-parameter -Wformat-security -Wswitch-default -Wundef -Wshadow -Wpointer-arith -Wbad-function-cast -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -Wmissing-declarations -Wredundant-decls -Wunreachable-code -std=gnu99 -fvisibility=hidden -Wlogical-op -g -MD -MT libhsakmt/CMakeFiles/hsakmt.dir/src/topology.c.o -MF CMakeFiles/hsakmt.dir/src/topology.c.o.d -o CMakeFiles/hsakmt.dir/src/topology.c.o -c /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src/topology.c
[  8%] Building C object libhsakmt/CMakeFiles/hsakmt.dir/src/rbtree.c.o
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cc -DHAVE_MEMFD_CREATE=1 -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/include -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src -I/usr/include/libdrm -I/usr/include -I/usr/include/libdrm -fPIC -W -Wall -Wextra -Wno-unused-parameter -Wformat-security -Wswitch-default -Wundef -Wshadow -Wpointer-arith -Wbad-function-cast -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -Wmissing-declarations -Wredundant-decls -Wunreachable-code -std=gnu99 -fvisibility=hidden -Wlogical-op -g -MD -MT libhsakmt/CMakeFiles/hsakmt.dir/src/rbtree.c.o -MF CMakeFiles/hsakmt.dir/src/rbtree.c.o.d -o CMakeFiles/hsakmt.dir/src/rbtree.c.o -c /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src/rbtree.c
[  8%] Building C object libhsakmt/CMakeFiles/hsakmt.dir/src/spm.c.o
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cc -DHAVE_MEMFD_CREATE=1 -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/include -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src -I/usr/include/libdrm -I/usr/include -I/usr/include/libdrm -fPIC -W -Wall -Wextra -Wno-unused-parameter -Wformat-security -Wswitch-default -Wundef -Wshadow -Wpointer-arith -Wbad-function-cast -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -Wmissing-declarations -Wredundant-decls -Wunreachable-code -std=gnu99 -fvisibility=hidden -Wlogical-op -g -MD -MT libhsakmt/CMakeFiles/hsakmt.dir/src/spm.c.o -MF CMakeFiles/hsakmt.dir/src/spm.c.o.d -o CMakeFiles/hsakmt.dir/src/spm.c.o -c /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src/spm.c
[  9%] Building C object libhsakmt/CMakeFiles/hsakmt.dir/src/version.c.o
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cc -DHAVE_MEMFD_CREATE=1 -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/include -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src -I/usr/include/libdrm -I/usr/include -I/usr/include/libdrm -fPIC -W -Wall -Wextra -Wno-unused-parameter -Wformat-security -Wswitch-default -Wundef -Wshadow -Wpointer-arith -Wbad-function-cast -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -Wmissing-declarations -Wredundant-decls -Wunreachable-code -std=gnu99 -fvisibility=hidden -Wlogical-op -g -MD -MT libhsakmt/CMakeFiles/hsakmt.dir/src/version.c.o -MF CMakeFiles/hsakmt.dir/src/version.c.o.d -o CMakeFiles/hsakmt.dir/src/version.c.o -c /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src/version.c
[  9%] Building C object libhsakmt/CMakeFiles/hsakmt.dir/src/svm.c.o
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cc -DHAVE_MEMFD_CREATE=1 -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/include -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src -I/usr/include/libdrm -I/usr/include -I/usr/include/libdrm -fPIC -W -Wall -Wextra -Wno-unused-parameter -Wformat-security -Wswitch-default -Wundef -Wshadow -Wpointer-arith -Wbad-function-cast -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -Wmissing-declarations -Wredundant-decls -Wunreachable-code -std=gnu99 -fvisibility=hidden -Wlogical-op -g -MD -MT libhsakmt/CMakeFiles/hsakmt.dir/src/svm.c.o -MF CMakeFiles/hsakmt.dir/src/svm.c.o.d -o CMakeFiles/hsakmt.dir/src/svm.c.o -c /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src/svm.c
[ 10%] Building C object libhsakmt/CMakeFiles/hsakmt.dir/src/pc_sampling.c.o
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cc -DHAVE_MEMFD_CREATE=1 -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/include -I/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src -I/usr/include/libdrm -I/usr/include -I/usr/include/libdrm -fPIC -W -Wall -Wextra -Wno-unused-parameter -Wformat-security -Wswitch-default -Wundef -Wshadow -Wpointer-arith -Wbad-function-cast -Wcast-qual -Wstrict-prototypes -Wmissing-prototypes -Wmissing-declarations -Wredundant-decls -Wunreachable-code -std=gnu99 -fvisibility=hidden -Wlogical-op -g -MD -MT libhsakmt/CMakeFiles/hsakmt.dir/src/pc_sampling.c.o -MF CMakeFiles/hsakmt.dir/src/pc_sampling.c.o.d -o CMakeFiles/hsakmt.dir/src/pc_sampling.c.o -c /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/libhsakmt/src/pc_sampling.c
[ 10%] Linking C static library archive/libhsakmt.a
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cmake -P CMakeFiles/hsakmt.dir/cmake_clean_target.cmake
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/libhsakmt && /usr/bin/cmake -E cmake_link_script CMakeFiles/hsakmt.dir/link.txt --verbose=1
/usr/bin/ar qc archive/libhsakmt.a CMakeFiles/hsakmt.dir/src/debug.c.o CMakeFiles/hsakmt.dir/src/events.c.o CMakeFiles/hsakmt.dir/src/fmm.c.o CMakeFiles/hsakmt.dir/src/globals.c.o CMakeFiles/hsakmt.dir/src/hsakmtmodel.c.o CMakeFiles/hsakmt.dir/src/libhsakmt.c.o CMakeFiles/hsakmt.dir/src/memory.c.o CMakeFiles/hsakmt.dir/src/openclose.c.o CMakeFiles/hsakmt.dir/src/perfctr.c.o CMakeFiles/hsakmt.dir/src/pmc_table.c.o CMakeFiles/hsakmt.dir/src/queues.c.o CMakeFiles/hsakmt.dir/src/time.c.o CMakeFiles/hsakmt.dir/src/topology.c.o CMakeFiles/hsakmt.dir/src/rbtree.c.o CMakeFiles/hsakmt.dir/src/spm.c.o CMakeFiles/hsakmt.dir/src/version.c.o CMakeFiles/hsakmt.dir/src/svm.c.o CMakeFiles/hsakmt.dir/src/pc_sampling.c.o
/usr/bin/ranlib archive/libhsakmt.a
make[2]: Leaving directory '/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build'
[ 10%] Built target hsakmt
make  -f runtime/hsa-runtime/image/blit_src/CMakeFiles/opencl_blit_objects.dir/build.make runtime/hsa-runtime/image/blit_src/CMakeFiles/opencl_blit_objects.dir/depend
make[2]: Entering directory '/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build'
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build && /usr/bin/cmake -E cmake_depends "Unix Makefiles" /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/runtime/hsa-runtime/image/blit_src /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/runtime/hsa-runtime/image/blit_src /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/runtime/hsa-runtime/image/blit_src/CMakeFiles/opencl_blit_objects.dir/DependInfo.cmake "--color="
make[2]: Leaving directory '/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build'
make  -f runtime/hsa-runtime/image/blit_src/CMakeFiles/opencl_blit_objects.dir/build.make runtime/hsa-runtime/image/blit_src/CMakeFiles/opencl_blit_objects.dir/build
make[2]: Entering directory '/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build'
[ 10%] BUILDING bitcode for ocl_blit_object_gfx700...
cd /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build/runtime/hsa-runtime/image/blit_src && /usr/lib/llvm-19/bin/clang-19 -O2 -x cl -Xclang -finclude-default-header -cl-denorms-are-zero -cl-std=CL2.0 -target amdgcn-amd-amdhsa -mcpu=gfx700 -mcode-object-version=4 -o ocl_blit_object_gfx700 /home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/runtime/hsa-runtime/image/blit_src/imageblit_kernels.cl
clang-19: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
make[2]: *** [runtime/hsa-runtime/image/blit_src/CMakeFiles/opencl_blit_objects.dir/build.make:225: runtime/hsa-runtime/image/blit_src/ocl_blit_object_gfx700] Error 1
make[2]: Leaving directory '/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build'
make[1]: *** [CMakeFiles/Makefile2:565: runtime/hsa-runtime/image/blit_src/CMakeFiles/opencl_blit_objects.dir/all] Error 2
make[1]: Leaving directory '/home/daryl/Source/therock/rocm-systems/projects/rocr-runtime/build'
make: *** [Makefile:159: all] Error 2

... but it needed something else.

I started looking to see if TheRock contained a way of generating this, looking into rocm-libraries/projects, but I didn't see anything, but I tried throwing that at the rocminfo configuration to see if it found what it needed (which did not work)

cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm -DDCMAKE_CXX_FLAGS="--rocm-device-lib-path=../../../../rocm-libraries" ..

At this point, I tried to adjust my approach to simply find targets and families within the codebase itself. I found cmake/therock_amdgpu_targets.cmake and tried extracting the values.

I tried the following (before realizing I was pulling the wrong strings):

$ cmake -GNinja .. -DTHEROCK_AMDGPU_FAMILIES=gfx101,gfx1010,gfx1011,gfx1012,gfx103,gfx1030,gfx1032,gfx1035,gfx1036,gfx110,gfx1100,gfx1101,gfx1102,gfx1103,gfx115,gfx1150,gfx1151,gfx120,gfx1200,gfx1201,gfx90,gfx906,gfx908,gfx94,gfx942,gfx950
-- Configuring background job pool for 2 concurrent jobs
-- Enabling building tests
-- ROCm version: 7.0.0
-- HIP version: 7.1.0
-- Implicitly enabled features: HOST_SUITE_SPARSE HOST_BLAS
-- Enabled features:
--   * HOST_BLAS (-DTHEROCK_ENABLE_HOST_BLAS=ON)
--   * HOST_SUITE_SPARSE (-DTHEROCK_ENABLE_HOST_SUITE_SPARSE=ON)
--   * COMPILER (-DTHEROCK_ENABLE_COMPILER=ON)
--   * HIPIFY (-DTHEROCK_ENABLE_HIPIFY=ON)
--   * CORE_RUNTIME (-DTHEROCK_ENABLE_CORE_RUNTIME=ON)
--   * HIP_RUNTIME (-DTHEROCK_ENABLE_HIP_RUNTIME=ON)
--   * ROCPROF_TRACE_DECODER_BINARY (-DTHEROCK_ENABLE_ROCPROF_TRACE_DECODER_BINARY=ON)
--   * ROCPROFV3 (-DTHEROCK_ENABLE_ROCPROFV3=ON)
--   * RCCL (-DTHEROCK_ENABLE_RCCL=ON)
--   * PRIM (-DTHEROCK_ENABLE_PRIM=ON)
--   * BLAS (-DTHEROCK_ENABLE_BLAS=ON)
--   * RAND (-DTHEROCK_ENABLE_RAND=ON)
--   * FFT (-DTHEROCK_ENABLE_FFT=ON)
--   * SPARSE (-DTHEROCK_ENABLE_SPARSE=ON)
--   * SOLVER (-DTHEROCK_ENABLE_SOLVER=ON)
--   * COMPOSABLE_KERNEL (-DTHEROCK_ENABLE_COMPOSABLE_KERNEL=ON)
--   * MIOPEN (-DTHEROCK_ENABLE_MIOPEN=ON)
-- Configured AMDGPU Targets:
CMake Error at cmake/therock_amdgpu_targets.cmake:135 (message):
  THEROCK_AMDGPU_FAMILIES value
  'gfx101,gfx1010,gfx1011,gfx1012,gfx103,gfx1030,gfx1032,gfx1035,gfx1036,gfx110,gfx1100,gfx1101,gfx1102,gfx1103,gfx115,gfx1150,gfx1151,gfx120,gfx1200,gfx1201,gfx90,gfx906,gfx908,gfx94,gfx942,gfx950'
  unknown.  Available: gfx906 dgpu-all gfx90X-all gfx90X-dgpu gfx90X-dcgpu
  gfx908 dcgpu-all gfx90a gfx942 gfx94X-all gfx94X-dcgpu gfx950 gfx950-all
  gfx950-dcgpu gfx1010 gfx101X-all gfx101X-dgpu gfx1011 gfx1012 gfx1030
  gfx103X-all gfx103X-dgpu gfx1032 gfx1035 gfx103X-igpu gfx1036 igpu-all
  gfx1100 gfx110X-all gfx110X-dgpu gfx1101 gfx1102 gfx1103 gfx110X-igpu
  gfx1150 gfx115X-all gfx115X-igpu gfx1151 gfx1200 gfx120X-all gfx1201
Call Stack (most recent call first):
  CMakeLists.txt:319 (therock_validate_amdgpu_targets)

-- Configuring incomplete, errors occurred!

At least at this point I was getting somewhere and had more information, and I could figure out a way to pull out the appropriate strings.

However, I've searched the TheRock codebase for various segments of my CPU "Model name", but there are no hits. My THEROCK_AMDGPU_FAMILIES doesn't exist in cmake/therock_amdgpu_targets.cmake, but when I use it, the build seems to accept it and move forward.

These questions are not about my specific GPU or architecture and are general questions.

---

1. What's the design purpose of selecting device targets and families instead of building for all supported devices by default?
   • I understand wanting to have the capability of creating ROCm for a subset of devices, making more efficient builds, or those targeting specific devices for debugging, but this default behavior doesn't make sense and I am trying to understand
2. How does one bootstrap ROCm via TheRock in such a way that they can build ROCm without knowing, or have a way of discovering the proper device targets and families?
3. Is there a way to build rocminfo as a starting point, either to then build ROCm, or to simply have rocminfo available with no plans on building ROCm?
4. Would it make sense to build a tool to assist people in determining their target or family, one that has zero to no dependencies (ie: .sh), to help provide such crucial information for one of the first steps of the TheRock ROCm build steps?

Thank you for your help and consideration.

Attach any links, screenshots, or additional evidence you think will be helpful.

No response

[JohnLoveJoy]

Broken -> https://github.com/ROCm/ROCm/docs/compatibility/compatibility-matrix.rst
Correct - > https://rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html#compatibility-matrix

And as a side note, please stop using broken links in releases...

[harkgill-amd]

@daryltucker appreciate the feedback on TheRock. Scott will be able to answer your questions more thoroughly.

@JohnLoveJoy, the broken links have been updated for both the ROCm 7.0.0 and 7.0.1 releases - thanks for the catch!

[marbre]

@daryltucker appreciate the feedback on TheRock. Scott will be able to answer your questions more thoroughly.

We have https://github.com/ROCm/TheRock/blob/main/RELEASES.md#index-page-listing in TheRock for pre-build releases as well as a note in the README.md when building from source:

Not all family and targets are currently supported. See therock_amdgpu_targets.cmake file for available options.

The latter has targets as well as the project name, e.g.:

TheRock/cmake/therock_amdgpu_targets.cmake

Line 107 in 80198e0

therock_add_amdgpu_target(gfx1150 "AMD Strix Point iGPU" FAMILY igpu-all gfx115X-all gfx115X-igpu

[ScottTodd]

In the scenario that you're trying to build ROCm without an existing instance of ROCm

FYI many ROCm developers are accustomed to starting from a ROCm install and developing on top of that, but with TheRock the expectation is that ROCm is not installed on the build system already. Building everything from scratch is more reliable than bootstrapping from an existing install, but the process for "building everything from scratch" was more complicated and time consuming before we set up this unified build system.

All documentation suggests using rocminfo to determine these values for your hardware. Though I found it bizarre I couldn't just target "all", I tried determining this by compiling rocminfo.

Documentation elsewhere may still be assuming independent source builds for each subproject. TheRock's documentation is here: https://github.com/ROCm/TheRock?tab=readme-ov-file#build-configuration . Those docs are very sparse and do not help answer your questions without requiring some reading of the source code though.

What I've been doing is:

1. Find your GPU name or gfx target in https://github.com/ROCm/TheRock/blob/main/cmake/therock_amdgpu_targets.cmake
2. See the family names at the end of the therock_add_amdgpu_target, like any of dgpu-all, gfx110X-all, or gfx110X-dgpu for these:

   TheRock/cmake/therock_amdgpu_targets.cmake

   Lines 94 to 100 in 038d4fc

   	# gfx110X family
   	therock_add_amdgpu_target(gfx1100 "AMD RX 7900 XTX" FAMILY dgpu-all gfx110X-all gfx110X-dgpu)
   	therock_add_amdgpu_target(gfx1101 "AMD RX 7800 XT" FAMILY dgpu-all gfx110X-all gfx110X-dgpu)
   	therock_add_amdgpu_target(gfx1102 "AMD RX 7700S/Framework Laptop 16" FAMILY dgpu-all gfx110X-all gfx110X-dgpu
   	EXCLUDE_TARGET_PROJECTS
   	hipBLASLt # https://github.com/ROCm/TheRock/issues/1062
   	)

3. Configure CMake with the family name of your choice: -DTHEROCK_AMDGPU_FAMILIES=gfx110X-dgpu

We can definitely add more docs, scripts, and CMake validation for that.

---

To the other specific questions:

1. What's the design purpose of selecting device targets and families instead of building for all supported devices by default?

• I understand wanting to have the capability of creating ROCm for a subset of devices, making more efficient builds, or those targeting specific devices for debugging, but this default behavior doesn't make sense and I am trying to understand

The math libraries include kernels for specific targets and including all targets would dramatically increase the build time and release binary size. We have more build system refactoring in mind for how to scale this better to modular releases that dynamically include what is needed based on the user's system (without the need for users to know on their own what GPUs they have).

2. How does one bootstrap ROCm via TheRock in such a way that they can build ROCm without knowing, or have a way of discovering the proper device targets and families?

The simplest way is to just know what devices you want to build for ahead of time. We run all of our CI/CD builds on machines with no GPU, so in those cases we have to be explicit and we can't rely on any bootstrapped auto-detection like "just build for my current GPU".

3. Is there a way to build rocminfo as a starting point, either to then build ROCm, or to simply have rocminfo available with no plans on building ROCm?

There are a few ways to approach this. At this level the systems libraries are mostly generic (one build portable across all GPUs). In the Python packages that we ship, we have:

• Linux: rocminfo
• Windows; hipinfo
• Both: amdgpu-arch

For example, let's say I don't know that my system is gfx1100. I can install gfx1151 wheels and run those tools:

python -m pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ rocm[devel]

amdgpu-arch
# gfx1100
# gfx1100

hipinfo
# device#                           0
# Name:                             AMD Radeon PRO W7900 Dual Slot
# ...

I think you could similarly build from source with the wrong target and run hipinfo / rocminfo / amdgpu-arch from the build directory. Maybe using -DTHEROCK_ENABLE_ALL=OFF -DTHEROCK_ENABLE_CORE=ON to just build those tools and their deps?

4. Would it make sense to build a tool to assist people in determining their target or family, one that has zero to no dependencies (ie: .sh), to help provide such crucial information for one of the first steps of the TheRock ROCm build steps?

We have something sort of like this in https://github.com/ROCm/TheRock/blob/main/build_tools/hack/diagnose.py. Running that script cross references the GPU names in the registry with https://github.com/ROCm/TheRock/blob/main/build_tools/hack/env_check/AMDGPU_LLVM_TARGET.py ... at least on Windows:

TheRock/build_tools/hack/env_check/device.py

Lines 195 to 202 in 038d4fc

	for i in range(0, gpu_count):
	_GPU_REG_KEY = str(
	r"SYSTEM\CurrentControlSet\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}"
	+ f"\\000{i}\\"
	)
	_GPU_CORE_NAME = RepoInfo.amdgpu_llvm_target(
	get_regedit("HKLM", _GPU_REG_KEY, "DriverDesc")
	)

we could do something similar on Linux, or find a way to bootstrap with rocminfo/hipinfo/amdgpu-arch instead.

[daryltucker]

Thanks so much for the information! I'll take a look over the weekend.