Rebased onto recent changes

Author: sterrettm2

lib/x86simdsort-avx2.cpp

@@ -38,32 +38,32 @@
     template <> \
     void keyvalue_qsort(type *key, uint64_t *val, size_t arrsize, bool hasnan) \
     { \
-        avx2_qsort_kv(key, val, arrsize, hasnan); \
+        x86simdsortStatic::keyvalue_qsort(key, val, arrsize, hasnan); \
     } \
     template <> \
     void keyvalue_qsort(type *key, int64_t *val, size_t arrsize, bool hasnan) \
     { \
-        avx2_qsort_kv(key, val, arrsize, hasnan); \
+        x86simdsortStatic::keyvalue_qsort(key, val, arrsize, hasnan); \
     } \
     template <> \
     void keyvalue_qsort(type *key, double *val, size_t arrsize, bool hasnan) \
     { \
-        avx2_qsort_kv(key, val, arrsize, hasnan); \
+        x86simdsortStatic::keyvalue_qsort(key, val, arrsize, hasnan); \
     } \
     template <> \
     void keyvalue_qsort(type *key, uint32_t *val, size_t arrsize, bool hasnan) \
     { \
-        avx2_qsort_kv(key, val, arrsize, hasnan); \
+        x86simdsortStatic::keyvalue_qsort(key, val, arrsize, hasnan); \
     } \
     template <> \
     void keyvalue_qsort(type *key, int32_t *val, size_t arrsize, bool hasnan) \
     { \
-        avx2_qsort_kv(key, val, arrsize, hasnan); \
+        x86simdsortStatic::keyvalue_qsort(key, val, arrsize, hasnan); \
     } \
     template <> \
     void keyvalue_qsort(type *key, float *val, size_t arrsize, bool hasnan) \
     { \
-        avx2_qsort_kv(key, val, arrsize, hasnan); \
+        x86simdsortStatic::keyvalue_qsort(key, val, arrsize, hasnan); \
     }
 
 namespace xss {

src/x86simdsort-static-incl.h

@@ -100,20 +100,26 @@ keyvalue_qsort(T1 *key, T2 *val, size_t size, bool hasnan = false);
         std::iota(indices.begin(), indices.end(), 0); \
         x86simdsortStatic::argselect(arr, indices.data(), k, size, hasnan); \
         return indices; \
-    }
+    } \
+    template <typename T1, typename T2> \
+    X86_SIMD_SORT_FINLINE void \
+    x86simdsortStatic::keyvalue_qsort(T1 *key, T2 *val, size_t size, bool hasnan) \
+    { \
+        ISA##_qsort_kv(key, val, size, hasnan); \
+    } \
 
 /*
  * qsort, qselect, partial, argsort key-value sort template functions.
  */
 #include "xss-common-qsort.h"
 #include "xss-common-argsort.h"
+#include "xss-common-keyvaluesort.hpp"
 
 #if defined(__AVX512DQ__) && defined(__AVX512VL__)
 /* 32-bit and 64-bit dtypes vector definitions on SKX */
 #include "avx512-32bit-qsort.hpp"
 #include "avx512-64bit-qsort.hpp"
 #include "avx512-64bit-argsort.hpp"
-#include "avx512-64bit-keyvaluesort.hpp"
 
 /* 16-bit dtypes vector definitions on ICL */
 #if defined(__AVX512BW__) && defined(__AVX512VBMI2__)
@@ -126,14 +132,6 @@ keyvalue_qsort(T1 *key, T2 *val, size_t size, bool hasnan = false);
 
 XSS_METHODS(avx512)
 
-// key-value currently only on avx512
-template <typename T1, typename T2>
-X86_SIMD_SORT_FINLINE void
-x86simdsortStatic::keyvalue_qsort(T1 *key, T2 *val, size_t size, bool hasnan)
-{
-    avx512_qsort_kv(key, val, size, hasnan);
-}
-
 #elif defined(__AVX512F__)
 #error "x86simdsort requires AVX512DQ and AVX512VL to be enabled in addition to AVX512F to use AVX512"
 

src/xss-common-keyvaluesort.hpp

@@ -441,7 +441,7 @@ xss_qsort_kv(T1 *keys, T2 *indexes, arrsize_t arrsize, bool hasnan)
                                       indexes,
                                       0,
                                       arrsize - 1,
-                                      2 * (arrsize_t)log2(arrsize));
+                                      maxiters);
             replace_inf_with_nan(keys, arrsize, nan_count);
         }
         else {
@@ -450,7 +450,7 @@ xss_qsort_kv(T1 *keys, T2 *indexes, arrsize_t arrsize, bool hasnan)
                                       indexes,
                                       0,
                                       arrsize - 1,
-                                      2 * (arrsize_t)log2(arrsize));
+                                      maxiters);
         }
     }
 }