griefith/test2
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Fix: OSL shadeops module fails to load in OptiX

Browse Source 
The OSL dependency is built with target sm_50 currently, which means
that LLVM defaults to generating PTX version 4.0. However, due to an
apparent bug in LLVM 20 it still uses instructions that were only
introduced in PTX version 6.0. As a result OptiX refuses to load the
shadeops PTX with an `OPTIX_ERROR_INVALID_INPUT` error.

To fix this, raise the PTX version generated by LLVM to 6.0 for both the
shadeops module (which previously used 4.0) and also any generated code
(which previously used 5.0) to be safe. PTX version 6.0 was introduced with
CUDA 9, so it has pretty long driver backwards compatibility still.

This commit contains fixes for the OSL, to fully fix the original report
a recompiled OSL libraries would need to land for the affected platforms.

Ref #147361

Pull Request: https://projects.blender.org/blender/blender/pulls/147620
This commit is contained in:
Patrick Mours
2025-10-15 16:18:46 +02:00
committed by Sergey Sharybin
parent 1621950441
commit d9cce547e6
2 changed files with 29 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
build_files/build_environment/cmake/osl.cmake
View File

@@ -81,6 +81,9 @@ ExternalProject_Add(external_osl
${PATCH_CMD} -p 1 -d
${BUILD_DIR}/osl/src/external_osl <
${PATCH_DIR}/osl.diff &&
${PATCH_CMD} -p 1 -d
${BUILD_DIR}/osl/src/external_osl <
${PATCH_DIR}/osl_ptx_version.diff &&
${PATCH_CMD} -p 1 -d
${BUILD_DIR}/osl/src/external_osl <
${PATCH_DIR}/osl_supports_isa_thread.diff

26
build_files/build_environment/patches/osl_ptx_version.diff Normal file
View File

@@ -0,0 +1,26 @@
diff --git a/src/cmake/cuda_macros.cmake b/src/cmake/cuda_macros.cmake
index cf38234e..78c7580f 100644
--- a/src/cmake/cuda_macros.cmake
+++ b/src/cmake/cuda_macros.cmake
@@ -225,7 +225,7 @@ function ( CUDA_SHADEOPS_COMPILE prefix output_bc output_ptx input_srcs headers
add_custom_command ( OUTPUT ${linked_bc} ${linked_ptx}
COMMAND ${LLVM_LINK_TOOL} ${shadeops_bc_list} -o ${linked_bc}
COMMAND ${LLVM_OPT_TOOL} ${opt_tool_flags} ${linked_bc} -o ${linked_bc}
- COMMAND ${LLVM_LLC_TOOL} --march=nvptx64 -mcpu=${CUDA_TARGET_ARCH} ${linked_bc} -o ${linked_ptx}
+ COMMAND ${LLVM_LLC_TOOL} -march=nvptx64 -mcpu=${CUDA_TARGET_ARCH} -mattr=+ptx60 ${linked_bc} -o ${linked_ptx}
# This script converts all of the .weak functions defined in the PTX into .visible functions.
COMMAND ${Python3_EXECUTABLE} "${CMAKE_SOURCE_DIR}/src/build-scripts/process-ptx.py"
${linked_ptx} ${linked_ptx}
diff --git a/src/liboslexec/llvm_util.cpp b/src/liboslexec/llvm_util.cpp
index ad85cec1..8da4149c 100644
--- a/src/liboslexec/llvm_util.cpp
+++ b/src/liboslexec/llvm_util.cpp
@@ -1831,7 +1831,7 @@ LLVM_Util::nvptx_target_machine()
&& "PTX compile error: LLVM Target is not initialized");
m_nvptx_target_machine = llvm_target->createTargetMachine(
- ModuleTriple.str(), CUDA_TARGET_ARCH, "+ptx50", options,
+ ModuleTriple.str(), CUDA_TARGET_ARCH, "+ptx60", options,
llvm::Reloc::Static, llvm::CodeModel::Small,
#if OSL_LLVM_VERSION >= 180
llvm::CodeGenOptLevel::Default