Merge pull request #143 from r-devulap/maint

Simplify use of ISA specific routines through a unified API for static methods

Author: sterrettm2

examples/Makefile

@@ -1,30 +1,24 @@
-CXX ?= g++-12
-CFLAGS = -I../src -std=c++17 -O3 $(if $(CXXFLAGS),$(CXXFLAGS),)
-EXE = qsort32avx2 argsort kvsort qsortfp16 qsort16 qsort32 qsort64
+CXX ?= g++-13
+CFLAGS = -I../src -std=c++17 -O3
+EXE = kvsort qsortavx2 qsortavx512 qsortspr qsorticl
 
 default: all
 all : $(EXE)
 
-qsortfp16: avx512fp-16bit-qsort.cpp
-	     $(CXX) -o qsortfp16 -march=sapphirerapids $(CFLAGS) avx512fp-16bit-qsort.cpp
-
-qsort16: avx512-16bit-qsort.cpp
-	     $(CXX) -o qsort16 -march=icelake-client $(CFLAGS) avx512-16bit-qsort.cpp
-
-qsort32: avx512-32bit-qsort.cpp
-	     $(CXX) -o qsort32 -march=skylake-avx512 $(CFLAGS) avx512-32bit-qsort.cpp
+kvsort: avx512-kv.cpp
+	     $(CXX) -o kvsort -mavx512vl -mavx512dq $(CFLAGS) avx512-kv.cpp
 
-qsort32avx2: avx2-32bit-qsort.cpp
-	     $(CXX) -o qsort32avx2 -march=haswell $(CFLAGS) avx2-32bit-qsort.cpp
+qsortavx512: skx-avx2.cpp
+	     $(CXX) -o qsortavx512 -mavx512vl -mavx512dq $(CFLAGS) skx-avx2.cpp
 
-qsort64: avx512-64bit-qsort.cpp
-	     $(CXX) -o qsort64 -march=skylake-avx512 $(CFLAGS) avx512-64bit-qsort.cpp
+qsortavx2: skx-avx2.cpp
+	     $(CXX) -o qsortavx2 -mavx2 $(CFLAGS) skx-avx2.cpp
 
-argsort: avx512-argsort.cpp
-	     $(CXX) -o argsort -march=skylake-avx512 $(CFLAGS) avx512-argsort.cpp
+qsorticl: icl-16bit.cpp
+	     $(CXX) -o qsorticl -mavx512vl -mavx512bw -mavx512dq -mavx512vbmi2 $(CFLAGS) icl-16bit.cpp
 
-kvsort: avx512-kv.cpp
-	     $(CXX) -o kvsort -march=skylake-avx512 $(CFLAGS) avx512-kv.cpp
+qsortspr: spr-16bit.cpp
+	     $(CXX) -o qsortspr -mavx512vl -mavx512dq -mavx512vbmi2 -mavx512fp16  $(CFLAGS) spr-16bit.cpp
 
 clean:
 	$(RM) $(EXE)

examples/avx2-32bit-qsort.cpp

@@ -1,4 +1,4 @@
-#include "avx512-64bit-keyvaluesort.hpp"
+#include "x86simdsort-static-incl.h"
 
 int main()
 {
@@ -7,17 +7,17 @@ int main()
     uint64_t arr2[size];
     double arr3[size];
     float arr4[size];
-    avx512_qsort_kv(arr1, arr1, size);
-    avx512_qsort_kv(arr1, arr2, size);
-    avx512_qsort_kv(arr1, arr3, size);
-    avx512_qsort_kv(arr2, arr1, size);
-    avx512_qsort_kv(arr2, arr2, size);
-    avx512_qsort_kv(arr2, arr3, size);
-    avx512_qsort_kv(arr3, arr1, size);
-    avx512_qsort_kv(arr3, arr2, size);
-    avx512_qsort_kv(arr1, arr4, size);
-    avx512_qsort_kv(arr2, arr4, size);
-    avx512_qsort_kv(arr3, arr4, size);
+    x86simdsortStatic::keyvalue_qsort(arr1, arr1, size);
+    x86simdsortStatic::keyvalue_qsort(arr1, arr2, size);
+    x86simdsortStatic::keyvalue_qsort(arr1, arr3, size);
+    x86simdsortStatic::keyvalue_qsort(arr2, arr1, size);
+    x86simdsortStatic::keyvalue_qsort(arr2, arr2, size);
+    x86simdsortStatic::keyvalue_qsort(arr2, arr3, size);
+    x86simdsortStatic::keyvalue_qsort(arr3, arr1, size);
+    x86simdsortStatic::keyvalue_qsort(arr3, arr2, size);
+    x86simdsortStatic::keyvalue_qsort(arr1, arr4, size);
+    x86simdsortStatic::keyvalue_qsort(arr2, arr4, size);
+    x86simdsortStatic::keyvalue_qsort(arr3, arr4, size);
     return 0;
     return 0;
 }

examples/avx512-16bit-qsort.cpp

@@ -0,0 +1,11 @@
+#include "x86simdsort-static-incl.h"
+
+int main()
+{
+    const int size = 1000;
+    short arr[size];
+    x86simdsortStatic::qsort(arr, size);
+    x86simdsortStatic::qselect(arr, 10, size);
+    x86simdsortStatic::partial_qsort(arr, 10, size);
+    return 0;
+}

examples/avx512-32bit-qsort.cpp

@@ -0,0 +1,19 @@
+#include "x86simdsort-static-incl.h"
+
+int main()
+{
+    const int size = 1000;
+    double arrd[size];
+    float arrf[size];
+    x86simdsortStatic::qsort(arrf, size);
+    x86simdsortStatic::qsort(arrd, size);
+    x86simdsortStatic::qselect(arrf, 10, size);
+    x86simdsortStatic::qselect(arrd, 10, size);
+    x86simdsortStatic::partial_qsort(arrf, 10, size);
+    x86simdsortStatic::partial_qsort(arrd, 10, size);
+    auto arg1 = x86simdsortStatic::argsort(arrf, size);
+    auto arg2 = x86simdsortStatic::argselect(arrf, 10, size);
+    auto arg3 = x86simdsortStatic::argsort(arrd, size);
+    auto arg4 = x86simdsortStatic::argselect(arrd, 10, size);
+    return 0;
+}