Add 32-bit ARM Neon SIMD support

[alexxy/gromacs.git] / cmake / gmxTestSimd.cmake

diff --git a/cmake/gmxTestSimd.cmake b/cmake/gmxTestSimd.cmake
index f99a43f6de8303caa41c2f5da0f4b108cdceedc9..47213b67f08c01941ddad6d44ce1aa4815d0f780 100644 (file)
--- a/cmake/gmxTestSimd.cmake
+++ b/cmake/gmxTestSimd.cmake
@@ -249,6 +249,26 @@ elseif(${GMX_SIMD} STREQUAL "AVX2_256")
     set(GMX_SIMD_X86_AVX2_256 1)
     set(SIMD_STATUS_MESSAGE "Enabling 256-bit AVX2 SIMD instructions")
 
+elseif(${GMX_SIMD} STREQUAL "ARM_NEON")
+
+    gmx_find_cflag_for_source(CFLAGS_ARM_NEON "C compiler 32-bit ARM NEON flag"
+                              "#include<arm_neon.h>
+                              int main(){float32x4_t x=vdupq_n_f32(0.5);x=vmlaq_f32(x,x,x);return vgetq_lane_f32(x,0)>0;}"
+                              SIMD_C_FLAGS
+                              "-mfpu=neon" "")
+    gmx_find_cxxflag_for_source(CXXFLAGS_ARM_NEON "C++ compiler 32-bit ARM NEON flag"
+                                "#include<arm_neon.h>
+                                int main(){float32x4_t x=vdupq_n_f32(0.5);x=vmlaq_f32(x,x,x);return vgetq_lane_f32(x,0)>0;}"
+                                SIMD_CXX_FLAGS
+                                "-mfpu=neon" "")
+
+    if(NOT CFLAGS_ARM_NEON OR NOT CXXFLAGS_ARM_NEON)
+        message(FATAL_ERROR "Cannot find ARM 32-bit NEON compiler flag. Use a newer compiler, or disable NEON SIMD.")
+    endif()
+
+    set(GMX_SIMD_ARM_NEON 1)
+    set(SIMD_STATUS_MESSAGE "Enabling 32-bit ARM NEON SIMD instructions")
+
 elseif(${GMX_SIMD} STREQUAL "IBM_QPX")
 
     try_compile(TEST_QPX ${CMAKE_BINARY_DIR}
@@ -265,6 +285,8 @@ elseif(${GMX_SIMD} STREQUAL "IBM_QPX")
 
 elseif(${GMX_SIMD} STREQUAL "SPARC64_HPC_ACE")
 
+    # Note that GMX_RELAXED_DOUBLE_PRECISION is enabled by default in the top-level CMakeLists.txt
+
     set(GMX_SIMD_SPARC64_HPC_ACE 1)
     set(SIMD_STATUS_MESSAGE "Enabling Sparc64 HPC-ACE SIMD instructions")
 
@@ -293,5 +315,24 @@ if (SIMD_CHANGED AND DEFINED SIMD_STATUS_MESSAGE)
     message(STATUS "${SIMD_STATUS_MESSAGE}")
 endif()
 
+# By default, 32-bit windows cannot pass SIMD (SSE/AVX) arguments in registers,
+# and even on 64-bit (all platforms) it is only used for a handful of arguments.
+# The __vectorcall (MSVC, from MSVC2013) or __regcall (ICC) calling conventions
+# enable this, which is critical to enable 32-bit SIMD and improves performance
+# for 64-bit SIMD.
+# Check if the compiler supports one of these, and in that case set gmx_simdcall
+# to that string. If we do not have any such calling convention modifier, set it
+# to an empty string.
+if(NOT DEFINED GMX_SIMD_CALLING_CONVENTION)
+    foreach(callconv __vectorcall __regcall "")
+        set(callconv_compile_var "_callconv_${callconv}")
+        check_c_source_compiles("int ${callconv} f(int i) {return i;} int main(void) {return f(0);}" ${callconv_compile_var})
+        if(${callconv_compile_var})
+            set(GMX_SIMD_CALLING_CONVENTION "${callconv}" CACHE INTERNAL "Calling convention for SIMD routines" FORCE)
+            break()
+        endif()
+    endforeach()
+endif()
+
 endmacro()