xswr)
llvm_require_version $LLVM_REQUIRED_SWR "swr"
- swr_require_cxx_feature_flags "C++14" "__cplusplus >= 201402L" \
- "-std=c++14" \
- SWR_CXX14_CXXFLAGS
- AC_SUBST([SWR_CXX14_CXXFLAGS])
+ swr_require_cxx_feature_flags "C++11" "__cplusplus >= 201103L" \
+ ",-std=c++11" \
+ SWR_CXX11_CXXFLAGS
+ AC_SUBST([SWR_CXX11_CXXFLAGS])
swr_require_cxx_feature_flags "AVX" "defined(__AVX__)" \
",-mavx,-march=core-avx" \
include Makefile.sources
include $(top_srcdir)/src/gallium/Automake.inc
-AM_CXXFLAGS = $(GALLIUM_DRIVER_CFLAGS) $(SWR_CXX14_CXXFLAGS)
+AM_CXXFLAGS = $(GALLIUM_DRIVER_CFLAGS) $(SWR_CXX11_CXXFLAGS)
noinst_LTLIBRARIES = libmesaswr.la
-fno-strict-aliasing \
$(GALLIUM_DRIVER_CFLAGS) \
$(LLVM_CXXFLAGS) \
- $(SWR_CXX14_CXXFLAGS) \
+ $(SWR_CXX11_CXXFLAGS) \
-I$(builddir)/rasterizer/codegen \
-I$(builddir)/rasterizer/jitter \
-I$(builddir)/rasterizer/archrast \
private:
+ template <typename MaskT>
+ INLINE __m128i expandThenBlend4(uint32_t* min, uint32_t* max) // @llvm_func_start
+ {
+ __m128i vMin = _mm_set1_epi32(*min);
+ __m128i vMax = _mm_set1_epi32(*max);
+ return _simd_blend4_epi32<MaskT::value>(vMin, vMax);
+ } // @llvm_func_end
+
INLINE void CalcTileSampleOffsets(int numSamples) // @llvm_func_start
- {
- auto expandThenBlend4 = [](uint32_t* min, uint32_t* max, auto mask)
- {
- __m128i vMin = _mm_set1_epi32(*min);
- __m128i vMax = _mm_set1_epi32(*max);
- return _simd_blend4_epi32<decltype(mask)::value>(vMin, vMax);
- };
-
+ {
auto minXi = std::min_element(std::begin(_xi), &_xi[numSamples]);
auto maxXi = std::max_element(std::begin(_xi), &_xi[numSamples]);
- std::integral_constant<int, 0xA> xMask;
+ using xMask = std::integral_constant<int, 0xA>;
// BR(max), BL(min), UR(max), UL(min)
- tileSampleOffsetsX = expandThenBlend4(minXi, maxXi, xMask);
-
+ tileSampleOffsetsX = expandThenBlend4<xMask>(minXi, maxXi);
+
auto minYi = std::min_element(std::begin(_yi), &_yi[numSamples]);
auto maxYi = std::max_element(std::begin(_yi), &_yi[numSamples]);
- std::integral_constant<int, 0xC> yMask;
+ using yMask = std::integral_constant<int, 0xC>;
// BR(max), BL(min), UR(max), UL(min)
- tileSampleOffsetsY = expandThenBlend4(minYi, maxYi, yMask);
+ tileSampleOffsetsY = expandThenBlend4<yMask>(minYi, maxYi);
}; // @llvm_func_end
// scalar sample values
uint32_t _xi[SWR_MAX_NUM_MULTISAMPLES];