wouter/d4r
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Validate native Windows M against real PTX and correct patch-merge bounds

Browse Source
This commit is contained in:
Codex committed 2026-09-30 19:01:18 +03:00
1 parent 5592b7a921
commit 2db93adb1a
13 files changed
+611 -17
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

CMakeLists.txt
+9
View File
@@ -35,6 +35,7 @@ add_custom_target(d4r_wmma_code_object ALL DEPENDS "${wmma_object}")
# M uses the existing FP16 widening baseline; native FP8 is not enabled here.
set(native_objects)
option(D4R_K_FP16_BASELINE "Round every K16 accumulation step like NVIDIA f16 WMMA" ON)
option(D4R_M_FP16_BASELINE "Preserve M e4m3 quantisation and f16 accumulation per K32" ON)
foreach(family k m)
file(GLOB native_sources CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${family}/*.hip")
file(GLOB native_headers CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${family}/*.h")
@@ -48,6 +49,9 @@ foreach(family k m)
if(family STREQUAL "k" AND D4R_K_FP16_BASELINE)
list(APPEND native_flags -DD4R_K_FP16_BASELINE)
endif()
if(family STREQUAL "m" AND D4R_M_FP16_BASELINE)
list(APPEND native_flags -DSWIN_EXACT)
endif()
if(flags_line)
separate_arguments(source_flags NATIVE_COMMAND "${CMAKE_MATCH_1}")
list(APPEND native_flags ${source_flags})
@@ -127,8 +131,13 @@ 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 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 hip_gfx1201 COMMAND d4r_hip_gfx1201_probe