Changed how partitioning small arrays is handled

sterrettm2 · sterrettm2 · commit 3dd2d13e7ad7 · 2023-10-11T10:44:53.000-07:00
diff --git a/src/avx512-common-qsort.h b/src/avx512-common-qsort.h
@@ -369,6 +369,8 @@ X86_SIMD_SORT_INLINE arrsize_t partition_avx512(type_t *arr,
     return l_store;
 }
 
+#include <iostream>
+
 template <typename vtype,
           int num_unroll,
           typename type_t = typename vtype::type_t>
@@ -384,9 +386,8 @@ X86_SIMD_SORT_INLINE arrsize_t partition_avx512_unrolled(type_t *arr,
                 arr, left, right, pivot, smallest, biggest);
     }
     
-    if (right - left < 4 * vtype::numlanes){
-        return partition_avx512<vtype>(
-                arr, left, right, pivot, smallest, biggest);
+    if (right - left < 3 * num_unroll * vtype::numlanes){
+        return partition_avx512<vtype>(arr, left, right, pivot, smallest, biggest);
     }
 
     /* make array length divisible by vtype::numlanes , shortening the array */

Original file line number	Diff line number	Diff line change
`@@ -369,6 +369,8 @@ X86_SIMD_SORT_INLINE arrsize_t partition_avx512(type_t *arr,`
`369`	`369`	`return l_store;`
`370`	`370`	`}`
`371`	`371`
	`372`	`+#include <iostream>`
	`373`	`+`
`372`	`374`	`template <typename vtype,`
`373`	`375`	`int num_unroll,`
`374`	`376`	`typename type_t = typename vtype::type_t>`
`@@ -384,9 +386,8 @@ X86_SIMD_SORT_INLINE arrsize_t partition_avx512_unrolled(type_t *arr,`
`384`	`386`	`arr, left, right, pivot, smallest, biggest);`
`385`	`387`	`}`
`386`	`388`
`387`		`- if (right - left < 4 * vtype::numlanes){`
`388`		`- return partition_avx512<vtype>(`
`389`		`- arr, left, right, pivot, smallest, biggest);`
	`389`	`+ if (right - left < 3 * num_unroll * vtype::numlanes){`
	`390`	`+ return partition_avx512<vtype>(arr, left, right, pivot, smallest, biggest);`
`390`	`391`	`}`
`391`	`392`
`392`	`393`	`/* make array length divisible by vtype::numlanes , shortening the array */`