Commit: ee351cb74d905368195ad5c663581e646241f90b
Author: Brecht Van Lommel
Date: Wed Aug 5 14:45:54 2020 +0200
Branches: blender-v2.90-release
https://developer.blender.org/rBee351cb74d905368195ad5c663581e646241f90b
Fix T78869: denoising performance regression on Windows
Optimization was disabled in this function to work around a bug in MSVC, use
a different solution that does not come with such a big performance regression.
===================================================================
M intern/cycles/util/util_math_fast.h
===================================================================
diff --git a/intern/cycles/util/util_math_fast.h
b/intern/cycles/util/util_math_fast.h
index 4f92d2f10d3..07b0878e3d5 100644
--- a/intern/cycles/util/util_math_fast.h
+++ b/intern/cycles/util/util_math_fast.h
@@ -445,12 +445,10 @@ ccl_device_inline float fast_expf(float x)
return fast_exp2f(x / M_LN2_F);
}
-#ifndef __KERNEL_GPU__
-/* MSVC seems to have a code-gen bug here in at least SSE41/AVX
- * see T78047 for details. */
-# ifdef _MSC_VER
-# pragma optimize("", off)
-# endif
+#if defined(__KERNEL_CPU__) && !defined(_MSC_VER)
+/* MSVC seems to have a code-gen bug here in at least SSE41/AVX, see
+ * T78047 and T78869 for details. Just disable for now, it only makes
+ * a small difference in denoising performance. */
ccl_device float4 fast_exp2f4(float4 x)
{
const float4 one = make_float4(1.0f);
@@ -466,14 +464,16 @@ ccl_device float4 fast_exp2f4(float4 x)
r = madd4(x, r, make_float4(1.0f));
return __int4_as_float4(__float4_as_int4(r) + (m << 23));
}
-# ifdef _MSC_VER
-# pragma optimize("", on)
-# endif
ccl_device_inline float4 fast_expf4(float4 x)
{
return fast_exp2f4(x / M_LN2_F);
}
+#else
+ccl_device_inline float4 fast_expf4(float4 x)
+{
+ return make_float4(fast_expf(x.x), fast_expf(x.y), fast_expf(x.z),
fast_expf(x.w));
+}
#endif
ccl_device_inline float fast_exp10(float x)
_______________________________________________
Bf-blender-cvs mailing list
[email protected]
https://lists.blender.org/mailman/listinfo/bf-blender-cvs
