feat: add d4r inference engines and Windows parity tooling

Add native K/M/RR inference engines, HIP kernels, replay and validation
harnesses, and Windows DLSS setup and runtime support. Include GPU
profiling tools, build and packaging scripts, platform integration
patches, and implementation documentation.
This commit is contained in:
ayo committed 2026-10-16 08:54:47 -04:00
1 parent 25493b3198
commit 7c5d3c9157
430 files changed
+76191

No files matched your search

+364
View File
@@ -0,0 +1,364 @@
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)
# Architecture parameterization adapted from realdody/d4r ae530b0.
# Keep explicit subtargets: no gfx11 substitution or untested future gfx12 ISA.
file(READ "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/gpu-targets.json" d4r_target_json)
set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS tools/windows/gpu-targets.json)
string(JSON d4r_target_count LENGTH "${d4r_target_json}" targets)
math(EXPR d4r_target_last "${d4r_target_count} - 1")
set(d4r_supported_targets)
set(d4r_target_cpp_entries "")
foreach(index RANGE ${d4r_target_last})
foreach(field architecture elfMachine isaMajor family)
string(JSON target_${field} GET "${d4r_target_json}" targets ${index} ${field})
endforeach()
list(APPEND d4r_supported_targets "${target_architecture}")
string(APPEND d4r_target_cpp_entries " GpuTarget{\"${target_architecture}\", ${target_elfMachine}, ${target_isaMajor}, \"${target_family}\"},\n")
endforeach()
configure_file(tools/windows/gpu_targets.generated.h.in gpu_targets.generated.h @ONLY)
include_directories("${CMAKE_CURRENT_BINARY_DIR}")
set(D4R_GPU_ARCH "gfx1201" CACHE STRING "Native Windows package GPU target")
set_property(CACHE D4R_GPU_ARCH PROPERTY STRINGS ${d4r_supported_targets})
if(NOT D4R_GPU_ARCH IN_LIST d4r_supported_targets)
message(FATAL_ERROR "D4R_GPU_ARCH must be one of: ${d4r_supported_targets}")
endif()
add_compile_definitions(D4R_TARGET_ARCH="${D4R_GPU_ARCH}")
configure_file(tools/windows/gpu-target.json.in gpu-target.json @ONLY)
install(FILES "${CMAKE_CURRENT_BINARY_DIR}/gpu-target.json" DESTINATION .)
install(FILES tools/windows/gpu-targets.json DESTINATION .)
set(d4r_device_flags)
if(D4R_GPU_ARCH MATCHES "^gfx11")
# Windows-only workaround informed by realdody/d4r 1897b4a. Keep the real
# gfx11 layout; do not route RDNA3 through gfx12 fragments.
list(APPEND d4r_device_flags -DD4R_WINDOWS_GFX11_LANE_EXCHANGE)
endif()
set(probe_object "${CMAKE_CURRENT_BINARY_DIR}/probe_${D4R_GPU_ARCH}.hsaco")
add_custom_command(OUTPUT "${probe_object}"
COMMAND "${D4R_HIP_CLANG}" -x hip --offload-arch=${D4R_GPU_ARCH}
--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 "${probe_object}")
set(pixel_object "${CMAKE_CURRENT_BINARY_DIR}/pixel_convert_${D4R_GPU_ARCH}.hsaco")
add_custom_command(OUTPUT "${pixel_object}"
COMMAND "${D4R_HIP_CLANG}" -x hip -std=c++17 --offload-arch=${D4R_GPU_ARCH}
--offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64
-nogpuinc -nogpulib -O2 "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/pixel_convert_gfx1201.hip"
-o "${pixel_object}"
DEPENDS tools/windows/pixel_convert_gfx1201.hip tools/windows/pixel_format.h kernels/common/hip_device_minimal.h VERBATIM)
add_custom_target(d4r_pixel_code_object ALL DEPENDS "${pixel_object}")
set(wmma_object "${CMAKE_CURRENT_BINARY_DIR}/wmma_${D4R_GPU_ARCH}.hsaco")
add_custom_command(OUTPUT "${wmma_object}"
COMMAND "${D4R_HIP_CLANG}" -x hip --offload-arch=${D4R_GPU_ARCH}
--offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64
-nogpuinc -nogpulib -O2 ${d4r_device_flags} "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/wmma_gfx1201.hip"
-o "${wmma_object}"
DEPENDS tools/windows/wmma_gfx1201.hip kernels/common/wmma_backend.h
kernels/common/wmma_layout.h kernels/common/wave32_exchange.h VERBATIM)
add_custom_target(d4r_wmma_code_object ALL DEPENDS "${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}/${kernel_name}_${D4R_GPU_ARCH}.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 "${D4R_HIP_CLANG}" -x hip -std=c++17 --offload-arch=${D4R_GPU_ARCH}
--offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64
-nogpuinc -nogpulib -O3 -DD4R_DEVICE_ONLY_MINIMAL ${d4r_device_flags} ${native_flags}
"${source}" -o "${native_object}"
DEPENDS "${source}" ${native_headers} kernels/common/wmma_layout.h kernels/common/wave32_exchange.h
kernels/common/hip_device_minimal.h VERBATIM)
list(APPEND native_objects "${native_object}")
install(FILES "${native_object}" DESTINATION "experimental/${family}")
endforeach()
endforeach()
add_custom_target(d4r_native_compile ALL DEPENDS ${native_objects})
# Retain the old manual build target for existing scripts.
add_custom_target(d4r_native_gfx1201_compile DEPENDS d4r_native_compile)
# Standalone Vulkan inference does not depend on the D3D12 NGX shim. Volk loads
# the Windows Vulkan runtime dynamically, so no import library is required.
option(D4R_BUILD_NATIVE_ENGINE "Build standalone K/M/L Vulkan inference runners" ON)
set(D4R_VULKAN_HEADERS_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/.tools/Vulkan-Headers-v1.4.341" CACHE PATH "Pinned Vulkan-Headers checkout")
set(D4R_VOLK_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/.tools/volk" CACHE PATH "Pinned Volk checkout")
if(D4R_BUILD_NATIVE_ENGINE)
if(EXISTS "${D4R_VULKAN_HEADERS_ROOT}/include/vulkan/vulkan.h" AND EXISTS "${D4R_VOLK_ROOT}/volk.c")
add_library(d4r_volk STATIC "${D4R_VOLK_ROOT}/volk.c")
target_include_directories(d4r_volk PUBLIC "${D4R_VOLK_ROOT}" "${D4R_VULKAN_HEADERS_ROOT}/include")
target_compile_definitions(d4r_volk PUBLIC VK_NO_PROTOTYPES)
add_library(d4r_engine STATIC engine/runtime.cpp engine/game.cpp engine/game_m.cpp engine/game_rr.cpp)
target_include_directories(d4r_engine PUBLIC "${D4R_VULKAN_HEADERS_ROOT}/include")
target_compile_definitions(d4r_engine PUBLIC VK_NO_PROTOTYPES)
target_precompile_headers(d4r_engine PRIVATE engine/volk_pch.h)
target_link_libraries(d4r_engine PUBLIC d4r_volk)
foreach(runner IN ITEMS k m rr)
if(runner STREQUAL "k")
set(runner_source engine/replay.cpp)
elseif(runner STREQUAL "m")
set(runner_source engine/replay_m.cpp)
else()
set(runner_source engine/replay_rr.cpp)
endif()
add_executable(d4r_engine_${runner} "${runner_source}")
target_include_directories(d4r_engine_${runner} PRIVATE "${D4R_VULKAN_HEADERS_ROOT}/include")
target_compile_definitions(d4r_engine_${runner} PRIVATE VK_NO_PROTOTYPES D4R_ENGINE_VOLK)
target_precompile_headers(d4r_engine_${runner} PRIVATE engine/volk_pch.h)
target_link_libraries(d4r_engine_${runner} PRIVATE d4r_engine)
if(MINGW)
target_link_options(d4r_engine_${runner} PRIVATE -static)
endif()
set_target_properties(d4r_engine_${runner} PROPERTIES OUTPUT_NAME "d4r-${runner}")
install(TARGETS d4r_engine_${runner} RUNTIME DESTINATION engine/bin)
endforeach()
else()
message(STATUS "Standalone engine skipped: run scripts/windows/setup-native-engine.ps1 for pinned Vulkan-Headers and Volk")
endif()
endif()
foreach(probe gpu_inventory 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_pixel_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_${probe} PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
else()
target_compile_options(d4r_${probe} PRIVATE -Wall -Wextra)
endif()
install(TARGETS d4r_${probe} RUNTIME DESTINATION bin)
endforeach()
foreach(probe binary16_probe gpu_target_probe)
add_executable(d4r_${probe} tools/windows/${probe}.cpp)
if(MINGW)
target_link_options(d4r_${probe} PRIVATE -static)
elseif(MSVC)
set_property(TARGET d4r_${probe} PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
endif()
install(TARGETS d4r_${probe} RUNTIME DESTINATION bin)
endforeach()
add_dependencies(d4r_gpu_target_probe d4r_probe_code_object d4r_pixel_code_object)
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)
add_dependencies(d4r_nvngx_windows d4r_pixel_code_object)
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 "${D4R_HIP_ROOT}/include")
target_link_libraries(d4r_nvngx_windows PRIVATE version dbghelp d3d12 dxguid minhook bcrypt)
install(FILES "${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)
# Target-independent _nvngx.dll of drag-in releases: loads _nvngx_<arch>.dll
# for the game's actual HIP architecture. Its own D4R_TARGET_ARCH is unused.
add_library(d4r_nvngx_loader SHARED tools/windows/d4r_nvngx_loader.cpp)
set_target_properties(d4r_nvngx_loader PROPERTIES PREFIX "" OUTPUT_NAME "d4r_nvngx_loader")
target_compile_definitions(d4r_nvngx_loader PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__)
target_include_directories(d4r_nvngx_loader PRIVATE "${D4R_HIP_ROOT}/include")
target_link_libraries(d4r_nvngx_loader PRIVATE version dbghelp)
if(MINGW)
target_link_options(d4r_nvngx_loader PRIVATE -static)
endif()
install(TARGETS d4r_nvngx_loader RUNTIME DESTINATION bin)
if(MSVC)
target_sources(d4r_ngx_cuda_init_probe PRIVATE tools/d4r_ngx_param_msvc.cpp tools/d4r_ngx_param_host.cpp)
target_sources(d4r_d3d12_dlss_probe PRIVATE tools/d4r_ngx_param_msvc.cpp tools/d4r_ngx_param_host.cpp)
target_sources(d4r_nvngx_windows PRIVATE tools/d4r_ngx_param_msvc.cpp tools/d4r_ngx_param_host.cpp)
else()
find_program(D4R_CLANG_CL NAMES clang-cl.exe PATHS "${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 "${D4R_HIP_ROOT}/include")
target_link_libraries(d4r_nvapi_compat PRIVATE version dbghelp)
if(MINGW)
target_link_options(d4r_nvapi_compat PRIVATE -static)
endif()
install(TARGETS d4r_nvapi_compat RUNTIME DESTINATION nvapi-compat)
target_link_libraries(d4r_d3d12_hip_interop_probe PRIVATE d3d12 dxgi dxguid)
target_link_libraries(d4r_d3d12_dlss_probe PRIVATE d3d12 dxgi dxguid d3dcompiler)
add_executable(d4r_debug_launcher tools/windows/debug_launcher.cpp)
target_include_directories(d4r_debug_launcher PRIVATE tools/windows)
target_link_libraries(d4r_debug_launcher PRIVATE version dbghelp)
install(TARGETS d4r_debug_launcher RUNTIME DESTINATION bin)
target_link_libraries(d4r_d3d12_command_probe PRIVATE d3d12 dxgi dxguid)
target_link_libraries(d4r_d3d12_pixel_probe PRIVATE d3d12 dxgi dxguid)
add_dependencies(d4r_d3d12_pixel_probe d4r_pixel_code_object)
add_dependencies(d4r_d3d12_command_probe d4r_nvngx_windows)
set(D4R_AGILITY_ROOT "" CACHE PATH "Optional unpacked Microsoft.Direct3D.D3D12 1.619.5 debug SDK")
if(EXISTS "${D4R_AGILITY_ROOT}/build/native/bin/x64/d3d12SDKLayers.dll")
add_executable(d4r_d3d12_command_debug_probe EXCLUDE_FROM_ALL tools/windows/d3d12_command_probe.cpp)
target_compile_definitions(d4r_d3d12_command_debug_probe PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__ D4R_DIAGNOSTIC_AGILITY)
target_include_directories(d4r_d3d12_command_debug_probe PRIVATE "${D4R_HIP_ROOT}/include")
target_link_libraries(d4r_d3d12_command_debug_probe PRIVATE version dbghelp d3d12 dxgi dxguid)
if(MINGW)
target_link_options(d4r_d3d12_command_debug_probe PRIVATE -static)
endif()
foreach(dll D3D12Core.dll d3d12SDKLayers.dll)
configure_file("${D4R_AGILITY_ROOT}/build/native/bin/x64/${dll}" "${CMAKE_CURRENT_BINARY_DIR}/debug-d3d12/${dll}" COPYONLY)
endforeach()
configure_file("${D4R_AGILITY_ROOT}/LICENSE.txt" "${CMAKE_CURRENT_BINARY_DIR}/debug-d3d12/LICENSE.txt" COPYONLY)
endif()
add_dependencies(d4r_hip_gfx1201_probe d4r_probe_code_object)
add_dependencies(d4r_gfx12_wmma_probe d4r_wmma_code_object)
add_dependencies(d4r_k_module_probe d4r_native_compile)
install(FILES "${probe_object}" DESTINATION bin)
install(FILES "${pixel_object}" DESTINATION bin)
install(FILES "${wmma_object}" DESTINATION bin)
install(FILES scripts/windows/test-windows-rdna4.ps1 DESTINATION .)
install(FILES scripts/windows/gpu-target.ps1 DESTINATION .)
install(FILES scripts/windows/gpu-preflight.ps1 DESTINATION .)
install(FILES scripts/windows/test-gpu-target-contract.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)
install(FILES docs/windows-gpu-support.md DESTINATION docs)
include(CTest)
add_test(NAME binary16_rounding COMMAND d4r_binary16_probe)
set_tests_properties(binary16_rounding PROPERTIES TIMEOUT 30 LABELS "software;numerics")
add_test(NAME gpu_target_contract COMMAND d4r_gpu_target_probe "${probe_object}" "${pixel_object}")
set_tests_properties(gpu_target_contract PROPERTIES TIMEOUT 30 LABELS "software;architecture")
find_program(D4R_POWERSHELL NAMES pwsh.exe powershell.exe REQUIRED)
add_test(NAME gpu_target_script_contract COMMAND "${D4R_POWERSHELL}" -NoProfile -File
"${CMAKE_CURRENT_SOURCE_DIR}/scripts/windows/test-gpu-target-contract.ps1"
-PackageRoot "${CMAKE_CURRENT_BINARY_DIR}" -CodeObject "${probe_object}")
set_tests_properties(gpu_target_script_contract PROPERTIES TIMEOUT 30 LABELS "software;architecture")
add_test(NAME d3d12_gpu_pixel_formats COMMAND d4r_d3d12_pixel_probe
--hip-root "${D4R_HIP_ROOT}" --module "${pixel_object}" --iterations 2)
set_tests_properties(d3d12_gpu_pixel_formats PROPERTIES TIMEOUT 120 LABELS "hardware;${D4R_GPU_ARCH};formats")
add_test(NAME d3d12_gpu_enhanced_formats COMMAND d4r_d3d12_pixel_probe
--hip-root "${D4R_HIP_ROOT}" --module "${pixel_object}" --interop-mode enhanced --iterations 2)
set_tests_properties(d3d12_gpu_enhanced_formats PROPERTIES TIMEOUT 120 LABELS "hardware;${D4R_GPU_ARCH};formats;enhanced_barriers")
add_test(NAME d3d12_command_backend COMMAND d4r_d3d12_command_probe
--hip-root "${D4R_HIP_ROOT}" --module "$<TARGET_FILE:d4r_nvngx_windows>" --interop-mode indirect --iterations 32)
set_tests_properties(d3d12_command_backend PROPERTIES TIMEOUT 120 LABELS "hardware;d3d12;command_lists;lifetime")
add_test(NAME d3d12_command_initial_pso COMMAND d4r_d3d12_command_probe
--hip-root "${D4R_HIP_ROOT}" --module "$<TARGET_FILE:d4r_nvngx_windows>" --interop-mode initial-pso --iterations 32)
set_tests_properties(d3d12_command_initial_pso PROPERTIES TIMEOUT 120 LABELS "hardware;d3d12;command_lists;lifetime")
add_test(NAME d3d12_command_root_reset_warp COMMAND d4r_d3d12_command_probe
--hip-root "${D4R_HIP_ROOT}" --module "$<TARGET_FILE:d4r_nvngx_windows>" --interop-mode indirect-root-reset --iterations 32)
set_tests_properties(d3d12_command_root_reset_warp PROPERTIES TIMEOUT 120 LABELS "software;d3d12;command_lists;lifetime"
ENVIRONMENT "D4R_COMMAND_PROBE_WARP=1;D4R_COMMAND_PROBE_NO_HIP=1")
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;${D4R_GPU_ARCH}")
add_test(NAME gfx12_wmma COMMAND d4r_gfx12_wmma_probe
--hip-root "${D4R_HIP_ROOT}" --module "${wmma_object}" --iterations 16)
set_tests_properties(gfx12_wmma PROPERTIES TIMEOUT 90 LABELS "hardware;${D4R_GPU_ARCH};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_${D4R_GPU_ARCH}.hsaco"
--kernel-name ${layer})
set_tests_properties(k_${layer}_execution PROPERTIES TIMEOUT 90 LABELS "hardware;${D4R_GPU_ARCH};k_execution")
endforeach()
add_test(NAME d3d12_hip_interop COMMAND d4r_d3d12_hip_interop_probe
--hip-root "${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 "${D4R_ZLUDA_ROOT}/nvcuda.dll" --iterations 32)
set_tests_properties(cuda_images PROPERTIES TIMEOUT 90 LABELS "hardware;zluda;descriptors")
add_test(NAME cuda_images_async COMMAND d4r_cuda_image_probe
--hip-root "${D4R_HIP_ROOT}" --cuda-dll "${D4R_ZLUDA_ROOT}/nvcuda.dll" --iterations 32)
set_tests_properties(cuda_images_async PROPERTIES TIMEOUT 90 LABELS "hardware;zluda;descriptors;async"
ENVIRONMENT "D4R_DIAG_ASYNC_ARRAY_COPY=1")
add_test(NAME d3d12_cuda_images COMMAND d4r_d3d12_dlss_probe
--hip-root "${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()