cmake_minimum_required(VERSION 3.24) project(d4r_windows_diagnostics LANGUAGES C CXX) if(NOT WIN32 OR NOT CMAKE_SIZEOF_VOID_P EQUAL 8) message(FATAL_ERROR "These diagnostics require native Windows x64") endif() set(CMAKE_CXX_STANDARD 17) set(CMAKE_CXX_STANDARD_REQUIRED ON) set(D4R_HIP_ROOT "$ENV{HIP_PATH}" CACHE PATH "One stable HIP SDK or TheRock installation") if(NOT D4R_HIP_ROOT AND EXISTS "C:/Program Files/AMD/ROCm/7.2") set(D4R_HIP_ROOT "C:/Program Files/AMD/ROCm/7.2") endif() if(NOT EXISTS "${D4R_HIP_ROOT}/include/hip/hip_runtime_api.h") message(FATAL_ERROR "Set D4R_HIP_ROOT to a Windows HIP SDK (headers and bin/clang++.exe required)") endif() find_program(D4R_HIP_CLANG NAMES clang++.exe PATHS "${D4R_HIP_ROOT}/bin" "${D4R_HIP_ROOT}/lib/llvm/bin" NO_DEFAULT_PATH REQUIRED) set(probe_object "${CMAKE_CURRENT_BINARY_DIR}/probe_gfx1201.hsaco") add_custom_command(OUTPUT "${probe_object}" COMMAND "${D4R_HIP_CLANG}" -x hip --offload-arch=gfx1201 --offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64 -nogpuinc -nogpulib -O2 "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/probe_kernel.hip" -o "${probe_object}" DEPENDS tools/windows/probe_kernel.hip VERBATIM) add_custom_target(d4r_probe_code_object ALL DEPENDS "${probe_object}") set(wmma_object "${CMAKE_CURRENT_BINARY_DIR}/wmma_gfx1201.hsaco") add_custom_command(OUTPUT "${wmma_object}" COMMAND "${D4R_HIP_CLANG}" -x hip --offload-arch=gfx1201 --offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64 -nogpuinc -nogpulib -O2 "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/wmma_gfx1201.hip" -o "${wmma_object}" DEPENDS tools/windows/wmma_gfx1201.hip kernels/common/wmma_backend.h kernels/common/wmma_layout.h VERBATIM) add_custom_target(d4r_wmma_code_object ALL DEPENDS "${wmma_object}") # K/M objects stay in experimental/, outside the ZLUDA native override # directory, until each layer and then the complete network pass replay. # M uses the existing FP16 widening baseline; native FP8 is not enabled here. set(native_objects) option(D4R_K_FP16_BASELINE "Round every K16 accumulation step like NVIDIA f16 WMMA" ON) option(D4R_M_FP16_BASELINE "Preserve M e4m3 quantisation and f16 accumulation per K32" ON) foreach(family k m) file(GLOB native_sources CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${family}/*.hip") file(GLOB native_headers CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${family}/*.h") foreach(source IN LISTS native_sources) get_filename_component(kernel_name "${source}" NAME_WE) set(native_object "${CMAKE_CURRENT_BINARY_DIR}/${kernel_name}_gfx1201.hsaco") set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${source}") file(READ "${source}" kernel_source) string(REGEX MATCH "// d4r-build-flags: ([^\r\n]+)" flags_line "${kernel_source}") set(native_flags) if(family STREQUAL "k" AND D4R_K_FP16_BASELINE) list(APPEND native_flags -DD4R_K_FP16_BASELINE) endif() if(family STREQUAL "m" AND D4R_M_FP16_BASELINE) list(APPEND native_flags -DSWIN_EXACT -DSWIN_EXACT_PV) endif() if(flags_line) separate_arguments(source_flags NATIVE_COMMAND "${CMAKE_MATCH_1}") list(APPEND native_flags ${source_flags}) endif() add_custom_command(OUTPUT "${native_object}" COMMAND "${D4R_HIP_CLANG}" -x hip -std=c++17 --offload-arch=gfx1201 --offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64 -nogpuinc -nogpulib -O3 -DD4R_DEVICE_ONLY_MINIMAL ${native_flags} "${source}" -o "${native_object}" DEPENDS "${source}" ${native_headers} kernels/common/wmma_layout.h kernels/common/hip_device_minimal.h VERBATIM) list(APPEND native_objects "${native_object}") install(FILES "${native_object}" DESTINATION "experimental/${family}") endforeach() endforeach() add_custom_target(d4r_native_gfx1201_compile ALL DEPENDS ${native_objects}) foreach(probe hip_gfx1201_probe gfx12_wmma_probe k_module_probe native_replay_probe cuda_driver_probe cuda_mma_probe cuda_image_probe d3d12_command_probe d3d12_hip_interop_probe hip_stream_lifecycle_probe ngx_cuda_init_probe d3d12_dlss_probe) add_executable(d4r_${probe} tools/windows/${probe}.cpp) target_compile_definitions(d4r_${probe} PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__) target_include_directories(d4r_${probe} PRIVATE "${D4R_HIP_ROOT}/include") target_link_libraries(d4r_${probe} PRIVATE version dbghelp) if(MINGW) target_link_options(d4r_${probe} PRIVATE -static) endif() if(MSVC) target_compile_options(d4r_${probe} PRIVATE /W4 /EHsc) set_property(TARGET d4r_${probe} PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$:Debug>") else() target_compile_options(d4r_${probe} PRIVATE -Wall -Wextra) endif() install(TARGETS d4r_${probe} RUNTIME DESTINATION bin) endforeach() set(D4R_MINHOOK_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/external/MinHook" CACHE PATH "MinHook 1.3.4 source checkout") if(NOT EXISTS "${D4R_MINHOOK_ROOT}/include/MinHook.h") message(FATAL_ERROR "Clone MinHook v1.3.4 into external/MinHook or set D4R_MINHOOK_ROOT") endif() add_subdirectory("${D4R_MINHOOK_ROOT}" "${CMAKE_CURRENT_BINARY_DIR}/minhook" EXCLUDE_FROM_ALL) add_library(d4r_nvngx_windows SHARED tools/windows/d4r_nvngx_windows.cpp tools/windows/d3d12_command_hooks.cpp) set_target_properties(d4r_nvngx_windows PROPERTIES PREFIX "" OUTPUT_NAME "d4r_nvngx") target_compile_definitions(d4r_nvngx_windows PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__) target_include_directories(d4r_nvngx_windows PRIVATE "${D4R_HIP_ROOT}/include") target_link_libraries(d4r_nvngx_windows PRIVATE version dbghelp d3d12 dxguid minhook) install(FILES "${D4R_MINHOOK_ROOT}/LICENSE.txt" DESTINATION licenses RENAME MinHook.txt) if(MINGW) target_link_options(d4r_nvngx_windows PRIVATE -static) endif() install(TARGETS d4r_nvngx_windows RUNTIME DESTINATION bin) if(MSVC) target_sources(d4r_ngx_cuda_init_probe PRIVATE tools/d4r_ngx_param_msvc.cpp tools/d4r_ngx_param_host.cpp) target_sources(d4r_d3d12_dlss_probe PRIVATE tools/d4r_ngx_param_msvc.cpp tools/d4r_ngx_param_host.cpp) target_sources(d4r_nvngx_windows PRIVATE tools/d4r_ngx_param_msvc.cpp tools/d4r_ngx_param_host.cpp) else() find_program(D4R_CLANG_CL NAMES clang-cl.exe PATHS "${D4R_HIP_ROOT}/bin" "${D4R_HIP_ROOT}/lib/llvm/bin" NO_DEFAULT_PATH REQUIRED) set(ngx_parameter_object "${CMAKE_CURRENT_BINARY_DIR}/ngx_parameter_msvc.obj") add_custom_command(OUTPUT "${ngx_parameter_object}" COMMAND "${D4R_CLANG_CL}" /c /O2 /Zl /GS- /GR- /EHs-c- "${CMAKE_CURRENT_SOURCE_DIR}/tools/d4r_ngx_param_msvc.cpp" "/Fo${ngx_parameter_object}" DEPENDS tools/d4r_ngx_param_msvc.cpp VERBATIM) set_source_files_properties("${ngx_parameter_object}" PROPERTIES GENERATED TRUE EXTERNAL_OBJECT TRUE) target_sources(d4r_ngx_cuda_init_probe PRIVATE "${ngx_parameter_object}" tools/d4r_ngx_param_host.cpp) target_sources(d4r_d3d12_dlss_probe PRIVATE "${ngx_parameter_object}" tools/d4r_ngx_param_host.cpp) target_sources(d4r_nvngx_windows PRIVATE "${ngx_parameter_object}" tools/d4r_ngx_param_host.cpp) endif() add_library(d4r_nvapi_trace SHARED tools/windows/nvapi_trace.cpp) set_target_properties(d4r_nvapi_trace PROPERTIES PREFIX "" OUTPUT_NAME "nvapi64" ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-trace" RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-trace") target_compile_definitions(d4r_nvapi_trace PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX) target_link_libraries(d4r_nvapi_trace PRIVATE version dbghelp) if(MINGW) target_link_options(d4r_nvapi_trace PRIVATE -static) endif() install(TARGETS d4r_nvapi_trace RUNTIME DESTINATION nvapi-trace) add_library(d4r_nvapi_compat SHARED tools/windows/nvapi_trace.cpp) set_target_properties(d4r_nvapi_compat PROPERTIES PREFIX "" OUTPUT_NAME "nvapi64" ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-compat" RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-compat") target_compile_definitions(d4r_nvapi_compat PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__ D4R_NVAPI_COMPAT) target_include_directories(d4r_nvapi_compat PRIVATE "${D4R_HIP_ROOT}/include") target_link_libraries(d4r_nvapi_compat PRIVATE version dbghelp) if(MINGW) target_link_options(d4r_nvapi_compat PRIVATE -static) endif() install(TARGETS d4r_nvapi_compat RUNTIME DESTINATION nvapi-compat) target_link_libraries(d4r_d3d12_hip_interop_probe PRIVATE d3d12 dxgi dxguid) target_link_libraries(d4r_d3d12_dlss_probe PRIVATE d3d12 dxgi dxguid) target_link_libraries(d4r_d3d12_command_probe PRIVATE d3d12 dxgi dxguid) add_dependencies(d4r_d3d12_command_probe d4r_nvngx_windows) add_dependencies(d4r_hip_gfx1201_probe d4r_probe_code_object) add_dependencies(d4r_gfx12_wmma_probe d4r_wmma_code_object) add_dependencies(d4r_k_module_probe d4r_native_gfx1201_compile) install(FILES "${probe_object}" DESTINATION bin) install(FILES "${wmma_object}" DESTINATION bin) install(FILES scripts/windows/test-windows-rdna4.ps1 DESTINATION .) install(FILES scripts/windows/test-k-replay.ps1 DESTINATION .) install(FILES scripts/windows/test-k-captures.ps1 DESTINATION .) install(FILES scripts/windows/test-native-overrides.ps1 DESTINATION .) install(FILES tools/windows/k_layer_validate.py DESTINATION .) install(FILES tools/windows/k_replay_validate.py DESTINATION .) install(FILES tools/windows/k_ptx_validate.py DESTINATION .) install(FILES tools/windows/frame_compare.py DESTINATION .) install(FILES tools/windows/m_replay_validate.py DESTINATION .) install(FILES tools/windows/m_ptx_validate.py DESTINATION .) install(FILES tools/windows/m_temporal_validate.py tools/windows/m_capture_compare.py DESTINATION .) install(FILES scripts/windows/test-m-replay.ps1 DESTINATION .) install(FILES scripts/windows/test-m-captures.ps1 DESTINATION .) install(FILES kernels/tools/pwin_model.py kernels/tools/pwin_check.py kernels/tools/ptxsim.py kernels/tools/extract_dlss_ptx.py kernels/tools/kernel_manifest.py DESTINATION reference) install(FILES kernels/tools/model_enc3.py kernels/tools/swin_model.py kernels/tools/swin_check.py DESTINATION reference) install(FILES docs/windows-rdna4-port.md DESTINATION docs) include(CTest) add_test(NAME d3d12_command_backend COMMAND d4r_d3d12_command_probe --hip-root "${D4R_HIP_ROOT}" --module "$" --iterations 32) set_tests_properties(d3d12_command_backend PROPERTIES TIMEOUT 120 LABELS "hardware;d3d12;command_lists;lifetime") add_test(NAME hip_gfx1201 COMMAND d4r_hip_gfx1201_probe --hip-root "${D4R_HIP_ROOT}" --module "${probe_object}") set_tests_properties(hip_gfx1201 PROPERTIES TIMEOUT 90 LABELS "hardware;gfx1201") add_test(NAME gfx12_wmma COMMAND d4r_gfx12_wmma_probe --hip-root "${D4R_HIP_ROOT}" --module "${wmma_object}" --iterations 16) set_tests_properties(gfx12_wmma PROPERTIES TIMEOUT 90 LABELS "hardware;gfx1201;wmma") foreach(layer enc1 enc2) add_test(NAME k_${layer}_execution COMMAND d4r_k_module_probe --hip-root "${D4R_HIP_ROOT}" --module "${CMAKE_CURRENT_BINARY_DIR}/dltss_pwin_${layer}_layer_gfx1201.hsaco" --kernel-name ${layer}) set_tests_properties(k_${layer}_execution PROPERTIES TIMEOUT 90 LABELS "hardware;gfx1201;k_execution") endforeach() add_test(NAME d3d12_hip_interop COMMAND d4r_d3d12_hip_interop_probe --hip-root "${D4R_HIP_ROOT}" --module "${probe_object}" --iterations 32) set_tests_properties(d3d12_hip_interop PROPERTIES TIMEOUT 120 LABELS "hardware;interop") add_test(NAME d3d12_hip_map_lifetime COMMAND d4r_d3d12_hip_interop_probe --hip-root "${D4R_HIP_ROOT}" --interop-mode map --iterations 64) set_tests_properties(d3d12_hip_map_lifetime PROPERTIES TIMEOUT 120 LABELS "hardware;interop;lifetime") set(D4R_ZLUDA_ROOT "" CACHE PATH "Local Windows ZLUDA directory (nvcuda.dll)") if(EXISTS "${D4R_ZLUDA_ROOT}/nvcuda.dll") foreach(context primary created) add_test(NAME cuda_${context} COMMAND d4r_cuda_driver_probe --hip-root "${D4R_HIP_ROOT}" --cuda-dll "${D4R_ZLUDA_ROOT}/nvcuda.dll" --context ${context} --iterations 32) set_tests_properties(cuda_${context} PROPERTIES TIMEOUT 180 LABELS "hardware;zluda") endforeach() if(EXISTS "${D4R_ZLUDA_ROOT}/build-info.json") add_test(NAME cuda_f16_mma_reference COMMAND d4r_cuda_mma_probe --hip-root "${D4R_HIP_ROOT}" --cuda-dll "${D4R_ZLUDA_ROOT}/nvcuda.dll") set_tests_properties(cuda_f16_mma_reference PROPERTIES TIMEOUT 120 LABELS "hardware;zluda;mma;reference" ENVIRONMENT "D4R_ZLUDA_WMMA=1;D4R_ZLUDA_WMMA_F16_REFERENCE=1") add_test(NAME cuda_images COMMAND d4r_cuda_image_probe --hip-root "${D4R_HIP_ROOT}" --cuda-dll "${D4R_ZLUDA_ROOT}/nvcuda.dll" --iterations 32) set_tests_properties(cuda_images PROPERTIES TIMEOUT 90 LABELS "hardware;zluda;descriptors") add_test(NAME d3d12_cuda_images COMMAND d4r_d3d12_dlss_probe --hip-root "${D4R_HIP_ROOT}" --cuda-dll "${D4R_ZLUDA_ROOT}/nvcuda.dll" --interop-mode images --iterations 1) set_tests_properties(d3d12_cuda_images PROPERTIES TIMEOUT 90 LABELS "hardware;interop;zluda;descriptors") endif() endif()