Add support for SSE3 (new default) and option to target SSE4.2

Author: jeremyong

.travis.yml

@@ -123,5 +123,5 @@ before_script:
   - ${CMAKE} --build .
 
 script:
-  - if [ "${COVERITY_SCAN_BRANCH}" != 1 ]; then ./klein_test; fi
+  - if [ "${COVERITY_SCAN_BRANCH}" != 1 ]; then ./klein_test && ./klein_test_sse42; fi
   - if [ "${ENABLE_GCOV}" = 1 ]; then bash <(curl -s https://codecov.io/bash) -x gcov-9 -a "-s `pwd`"; fi

CMakeLists.txt

@@ -27,11 +27,20 @@ endif()
 option(KLEIN_BUILD_SYM "Enable compilation of symbolic Klein utility" ON)
 option(KLEIN_BUILD_C_BINDINGS "Enable compilation of the Klein C bindings" ON)
 
+# The default platform and instruction set is x86 SSE3
 add_library(klein INTERFACE)
 add_library(klein::klein ALIAS klein)
 target_include_directories(klein INTERFACE public)
 target_compile_features(klein INTERFACE cxx_std_17)
-# SSE4.2 has > 96% market penetration according to the Steam hardware survey
+if(NOT MSVC)
+    target_compile_options(klein INTERFACE -msse3)
+endif()
+
+add_library(klein_sse42 INTERFACE)
+add_library(klein::klein_sse42 ALIAS klein_sse42)
+target_include_directories(klein_sse42 INTERFACE public)
+target_compile_features(klein_sse42 INTERFACE cxx_std_17)
+# SSE4.1 has > 97% market penetration according to the Steam hardware survey
 # queried as of December 2019 while AVX2 is around 70%. Thus, we can assume
 # FMA support is at least 70%, but perhaps not much more beyond that.
 # TODO: Optionally support FMA
@@ -40,7 +49,8 @@ if(MSVC)
     # AVX extensions). This is on by default.
 else()
     # Unlike MSVC, FMA instructions are enabled with a separate feature flag
-    target_compile_options(klein INTERFACE -msse4.2)
+    target_compile_options(klein_sse42 INTERFACE -msse4.1)
+    target_compile_definitions(klein_sse42 INTERFACE KLEIN_SSE_4_1)
 endif()
 
 if(KLEIN_ENABLE_PERF)
@@ -54,7 +64,7 @@ if(KLEIN_ENABLE_TESTS)
 endif()
 
 if(KLEIN_BUILD_SYM)
-    add_subdirectory(src)
+    add_subdirectory(sym)
 endif()
 
 if(KLEIN_BUILD_C_BINDINGS)

README.md

@@ -19,8 +19,9 @@ SIMD Extensions) for maximum throughput.
 
 ## Requirements
 
-- Machine with a processor that supports SSE4.2 or later (has ~97% market penetration)
+- Machine with a processor that supports SSE3 or later (Steam hardware survey reports 100% market penetration)
 - C++17 compliant compiler (tested with GCC 9.2.1, Clang 9.0.1, and Visual Studio 2019)
+- Optional SSE4.1 support
 
 ## Usage
 

appveyor.yml

@@ -30,3 +30,4 @@ build_script:
   - cmake --build .
   - dir
   - C:\projects\klein\%configuration%\klein_test.exe
+  - C:\projects\klein\%configuration%\klein_test_sse42.exe

bench/rtm_bench.cpp

@@ -28,7 +28,7 @@ art kinematic and math libraries built with traditional vector and quaternion fo
 
 - Geometric computing library suitable for use with realtime graphics and animation applications
 - Header-only core libary with an optional lightweight symbolic computer algebra system
-- SSE4.2-optimized implementation (SSE2 fallback on the way)
+- SSE3 or SSE4.1-optimized implementations
 - Tested on Linux, MacOS, and Windows
 - Requires no third-party dependencies
 - Permissively licensed

docs/index.md

@@ -17,6 +17,8 @@ FetchContent_Declare(
 FetchContent_MakeAvailable(klein)
 
 # Now, you can use target_link_libraries(your_lib PUBLIC klein::klein)
+# If you can target SSE4.1 (~97% market penetration), you can link against
+# the target klein::klein_sse42 instead.
 ```
 
 The primary "catch-all" header provided can be included using `#include <klein/klein.hpp>`.

docs/quickstart.md

@@ -31,7 +31,7 @@ FetchContent_MakeAvailable(mc_ruler)
 
 include(MCRuler)
 
-if(${CMAKE_BUILD_TYPE} STREQUAL "Release")
+if(CMAKE_BUILD_TYPE EQUAL "RELEASE")
     add_library(klein_perf klein_perf.cpp)
     target_link_libraries(klein_perf PRIVATE mc_ruler::mc_ruler klein)
     mc_ruler(klein_perf SOURCES klein_perf.cpp)

perf/CMakeLists.txt

@@ -9,7 +9,7 @@
 
 namespace kln
 {
-inline namespace detail
+namespace detail
 {
     // Partition memory layouts
     //     LSB --> MSB
@@ -48,8 +48,8 @@ inline namespace detail
 
         // Broadcast dot(a, a) ignoring the scalar component to all components
         // of a2
-        __m128 a2 = _mm_dp_ps(a, a, 0b11101111);
-        __m128 ab = _mm_dp_ps(a, b, 0b11101111);
+        __m128 a2 = hi_dp_bc(a, a);
+        __m128 ab = hi_dp_bc(a, b);
 
         // Next, we need the sqrt of that quantity. Since e0123 squares to 0,
         // this has a closed form solution.
@@ -144,9 +144,9 @@ inline namespace detail
         __m128 b       = _mm_mul_ps(bv_mask, p2);
 
         // Next, we need to compute the norm as in the exponential.
-        __m128 a2 = _mm_dp_ps(a, a, 0b11101111);
+        __m128 a2 = hi_dp_bc(a, a);
         // TODO: handle case when a2 is 0
-        __m128 ab          = _mm_dp_ps(a, b, 0b11101111);
+        __m128 ab          = hi_dp_bc(a, b);
         __m128 s           = _mm_sqrt_ps(a2);
         __m128 a2_sqrt_rcp = _mm_rcp_ps(s);
         __m128 minus_t     = _mm_mul_ps(ab, a2_sqrt_rcp);

public/klein/detail/x86/x86_exp_log.hpp

@@ -4,7 +4,7 @@
 
 namespace kln
 {
-inline namespace detail
+namespace detail
 {
     // Partition memory layouts
     //     LSB --> MSB
@@ -60,7 +60,7 @@ inline namespace detail
         p3_out = _mm_mul_ps(_mm_mul_ps(KLN_SWIZZLE(a, 0, 0, 0, 1), b),
                             _mm_set_ps(-1.f, -1.f, -1.f, 0.f));
 
-        p3_out = _mm_add_ss(p3_out, _mm_dp_ps(a, b, 0b11100001));
+        p3_out = _mm_add_ss(p3_out, hi_dp(a, b));
     }
 
     // p0 ^ p2 = p2 ^ p0
@@ -86,7 +86,7 @@ inline namespace detail
                                        __m128& p2_out) noexcept
     {
         // (a0 b0 + a1 b1 + a2 b2 + a3 b3) e0123
-        p2_out = _mm_dp_ps(a, b, 0b11110001);
+        p2_out = dp(a, b);
         if constexpr (Flip)
         {
             p2_out = _mm_xor_ps(p2_out, _mm_set_ss(-0.f));
@@ -117,7 +117,7 @@ inline namespace detail
         // (a0 b2) e02 +
         // (a0 b3) e03
         p2_out = _mm_mul_ps(KLN_SWIZZLE(a, 0, 0, 0, 0), b);
-        p2_out = _mm_add_ps(p2_out, _mm_dp_ps(a, b, 0b11100001));
+        p2_out = _mm_add_ps(p2_out, hi_dp(a, b));
     }
 
     // p1 ^ p3 = p3 ^ p1