From e4a265f05856eac40159b34b17d97c2ae1e14bb3 Mon Sep 17 00:00:00 2001
From: Sergey Sharybin <sergey.vfx@gmail.com>
Date: Fri, 25 Mar 2016 16:09:05 +0100
Subject: Cycles: Add an option to build single kernel only which fits current
 CPU

This seems quite useful for the development, so you don't need to wait
all the kernels to be re-compiled when working on a new feature, which
speeds up re-iteration.

Marked as an advanced option, so if it doesn't work so well in practice
it's safe to revert anyway.
---
 CMakeLists.txt                              |  2 ++
 intern/cycles/CMakeLists.txt                | 10 +++++++++-
 intern/cycles/kernel/kernels/cpu/kernel.cpp | 30 ++++++++++++++++++++++++++++-
 intern/cycles/util/util_simd.cpp            |  3 ++-
 4 files changed, 42 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 4010ee6416f..12f58706103 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -390,8 +390,10 @@ mark_as_advanced(CYCLES_CUDA_BINARIES_ARCH)
 unset(PLATFORM_DEFAULT)
 option(WITH_CYCLES_LOGGING	"Build Cycles with logging support" ON)
 option(WITH_CYCLES_DEBUG	"Build Cycles with extra debug capabilities" OFF)
+option(WITH_CYCLES_NATIVE_ONLY	"Build Cycles with native kernel only (which fits current CPU, use for development only)" OFF)
 mark_as_advanced(WITH_CYCLES_LOGGING)
 mark_as_advanced(WITH_CYCLES_DEBUG)
+mark_as_advanced(WITH_CYCLES_NATIVE_ONLY)
 
 option(WITH_CUDA_DYNLOAD "Dynamically load CUDA libraries at runtime" ON)
 mark_as_advanced(WITH_CUDA_DYNLOAD)
diff --git a/intern/cycles/CMakeLists.txt b/intern/cycles/CMakeLists.txt
index 13b5de360d5..3b6c25c370e 100644
--- a/intern/cycles/CMakeLists.txt
+++ b/intern/cycles/CMakeLists.txt
@@ -14,7 +14,15 @@ include(cmake/external_libs.cmake)
 # todo: this code could be refactored a bit to avoid duplication
 # note: CXX_HAS_SSE is needed in case passing SSE flags fails altogether (gcc-arm)
 
-if(NOT WITH_CPU_SSE)
+if(WITH_CYCLES_NATIVE_ONLY)
+	set(CXX_HAS_SSE FALSE)
+	set(CXX_HAS_AVX FALSE)
+	set(CXX_HAS_AVX2 FALSE)
+	add_definitions(
+		-DWITH_KERNEL_NATIVE
+	)
+	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
+elseif(NOT WITH_CPU_SSE)
 	set(CXX_HAS_SSE FALSE)
 	set(CXX_HAS_AVX FALSE)
 	set(CXX_HAS_AVX2 FALSE)
diff --git a/intern/cycles/kernel/kernels/cpu/kernel.cpp b/intern/cycles/kernel/kernels/cpu/kernel.cpp
index 643eefcdc6c..45091f6f33d 100644
--- a/intern/cycles/kernel/kernels/cpu/kernel.cpp
+++ b/intern/cycles/kernel/kernels/cpu/kernel.cpp
@@ -16,11 +16,39 @@
 
 /* CPU kernel entry points */
 
-/* On x86-64, we can assume SSE2, so avoid the extra kernel and compile this one with SSE2 intrinsics */
+/* On x86-64, we can assume SSE2, so avoid the extra kernel and compile this
+ * one with SSE2 intrinsics.
+ */
 #if defined(__x86_64__) || defined(_M_X64)
 #  define __KERNEL_SSE2__
 #endif
 
+/* When building kernel for native machine detect kernel features from the flags
+ * set by compiler.
+ */
+#ifdef WITH_KERNEL_NATIVE
+#  ifdef __SSE2__
+#    ifndef __KERNEL_SSE2__
+#      define __KERNEL_SSE2__
+#    endif
+#  endif
+#  ifdef __SSE3__
+#    define __KERNEL_SSE3__
+#  endif
+#  ifdef __SSSE3__
+#    define __KERNEL_SSSE3__
+#  endif
+#  ifdef __SSE4_1__
+#    define __KERNEL_SSE41__
+#  endif
+#  ifdef __AVX__
+#    define __KERNEL_AVX__
+#  endif
+#  ifdef __AVX2__
+#    define __KERNEL_AVX2__
+#  endif
+#endif
+
 /* quiet unused define warnings */
 #if defined(__KERNEL_SSE2__)
     /* do nothing */
diff --git a/intern/cycles/util/util_simd.cpp b/intern/cycles/util/util_simd.cpp
index eb9e32800e1..de2df612578 100644
--- a/intern/cycles/util/util_simd.cpp
+++ b/intern/cycles/util/util_simd.cpp
@@ -15,7 +15,8 @@
  * limitations under the License.
  */
 
-#ifdef WITH_KERNEL_SSE2
+#if (defined(WITH_KERNEL_SSE2)) || \
+    (defined(WITH_KERNEL_NATIVE) && defined(__SSE2__))
 
 #define __KERNEL_SSE2__
 #include "util_simd.h"
-- 
cgit v1.2.3