Cycles: Switch to HIP SDK 6.4.2 on Windows

This also reverts 367d5b7eabd53229fb7e79465b4761e65e531741,
as the math flags workaround is no longer needed.

Fix #139796
Fix #138646
Fix #139071
Fix #139070

Ref #140278

Co-authored-by: Brecht Van Lommel <brecht@blender.org>
Pull Request: https://projects.blender.org/blender/blender/pulls/145311

build_files/config/pipeline_config.yaml

@@ -14,7 +14,7 @@ buildbot:
         version: '12.8.0'
     hip:
         windows:
-            version: '6.3.42560'
+            version: '6.4.2'
         linux:
             version: '6.4.0'
     optix:

intern/cycles/device/hip/device_impl.cpp

@@ -263,11 +263,6 @@ string HIPDevice::compile_kernel(const uint kernel_features, const char *name, c
 
   const char *const kernel_ext = "genco";
   std::string options = "-Wno-parentheses-equality -Wno-unused-value -ffast-math";
-  if (hipNeedPreciseMath(arch)) {
-    options.append(
-        " -fhip-fp32-correctly-rounded-divide-sqrt -fno-gpu-approx-transcendentals "
-        "-fgpu-flush-denormals-to-zero -ffp-contract=off");
-  }
 
 #  ifndef NDEBUG
   options.append(" -save-temps");

intern/cycles/device/hip/util.h

@@ -77,18 +77,6 @@ static inline bool hipIsRDNA2OrNewer(const int hipDevId)
   return (major > 10 || (major == 10 && minor >= 3));
 }
 
-static inline bool hipNeedPreciseMath(const std::string &arch)
-{
-#  ifdef _WIN32
-  /* Enable stricter math options for RDNA2 GPUs (compiler bug on Windows). */
-  return (arch == "gfx1030" || arch == "gfx1031" || arch == "gfx1032" || arch == "gfx1033" ||
-          arch == "gfx1034" || arch == "gfx1035" || arch == "gfx1036");
-#  else
-  (void)arch;
-  return false;
-#  endif
-}
-
 static inline bool hipSupportsDeviceOIDN(const int hipDevId)
 {
   /* Matches HIPDevice::getArch in HIP. */

intern/cycles/device/hiprt/device_impl.cpp

@@ -225,11 +225,6 @@ string HIPRTDevice::compile_kernel(const uint kernel_features, const char *name,
   options.append(
       "-Wno-parentheses-equality -Wno-unused-value -ffast-math -O3 -std=c++17 -D __HIPRT__");
   options.append(" --offload-arch=").append(arch.c_str());
-  if (hipNeedPreciseMath(arch)) {
-    options.append(
-        " -fhip-fp32-correctly-rounded-divide-sqrt -fno-gpu-approx-transcendentals "
-        "-fgpu-flush-denormals-to-zero -ffp-contract=off");
-  }
 #  ifdef WITH_NANOVDB
   options.append(" -D WITH_NANOVDB");
 #  endif

intern/cycles/kernel/CMakeLists.txt

@@ -643,21 +643,6 @@ endif()
 
 # HIP module
 
-# Workaround for SDK 6.3.42560-881c2d702 compiler bug on RDNA2
-# Issue #136138
-function(get_hip_math_flag arch math_flag)
-  if(WIN32 AND (${arch} MATCHES "gfx1030|gfx1031|gfx1032|gfx1033|gfx1034|gfx1035|gfx1036"))
-    set(result
-      -fhip-fp32-correctly-rounded-divide-sqrt
-      -fno-gpu-approx-transcendentals
-      -fgpu-flush-denormals-to-zero
-      -ffp-contract=off)
-  else()
-    set(result "")
-  endif()
-  set(${math_flag} "${result}" PARENT_SCOPE)
-endfunction()
-
 if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIP)
   # build for each arch
   set(hip_sources device/hip/kernel.cpp
@@ -698,8 +683,6 @@ if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIP)
       set(hip_opt_flags)
     endif()
 
-    get_hip_math_flag(${arch} math_flag)
-
     set(hip_flags
       ${hip_flags}
       --offload-arch=${arch}
@@ -716,7 +699,6 @@ if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIP)
       -Wno-unused-value
       -ffast-math
       -std=c++17
-      ${math_flag}
       ${hip_opt_flags}
       -o ${CMAKE_CURRENT_BINARY_DIR}/${hip_file})
 
@@ -812,12 +794,10 @@ if(WITH_CYCLES_HIP_BINARIES AND WITH_CYCLES_DEVICE_HIPRT)
 
   set(hiprt_hipfb)
   foreach(arch ${CYCLES_HIP_BINARIES_ARCH})
-    get_hip_math_flag(${arch} math_flag)
     set(hiprt_file ${CMAKE_CURRENT_BINARY_DIR}/kernel_rt_${arch}.hipfb)
     set(hiprt_file_compressed ${hiprt_file}.zst)
     set(hiprt_flags
     ${hiprt_compile_flags}
-    ${math_flag}
     --offload-arch=${arch}
     -o ${hiprt_file})