Merge pull request #93 from sterrettm2/avx2-64bit

AVX2 64-bit support
intel · Oct 25, 2023 · 3c9bf9a · 3c9bf9a
2 parents fa4afb8 + 6d111f5
commit 3c9bf9a
Show file tree

Hide file tree

Showing 5 changed files with 728 additions and 9 deletions.
diff --git a/lib/x86simdsort-avx2.cpp b/lib/x86simdsort-avx2.cpp
@@ -1,5 +1,6 @@
 // AVX2 specific routines:
 #include "avx2-32bit-qsort.hpp"
+#include "avx2-64bit-qsort.hpp"
 #include "x86simdsort-internal.h"
 
 #define DEFINE_ALL_METHODS(type) \
@@ -24,5 +25,8 @@ namespace avx2 {
     DEFINE_ALL_METHODS(uint32_t)
     DEFINE_ALL_METHODS(int32_t)
     DEFINE_ALL_METHODS(float)
+    DEFINE_ALL_METHODS(uint64_t)
+    DEFINE_ALL_METHODS(int64_t)
+    DEFINE_ALL_METHODS(double)
 } // namespace avx2
 } // namespace xss
diff --git a/lib/x86simdsort.cpp b/lib/x86simdsort.cpp
@@ -150,15 +150,15 @@ DISPATCH(argselect, _Float16, ISA_LIST("none"))
 DISPATCH_ALL(qsort,
              (ISA_LIST("avx512_icl")),
              (ISA_LIST("avx512_skx", "avx2")),
-             (ISA_LIST("avx512_skx")))
+             (ISA_LIST("avx512_skx", "avx2")))
 DISPATCH_ALL(qselect,
              (ISA_LIST("avx512_icl")),
              (ISA_LIST("avx512_skx", "avx2")),
-             (ISA_LIST("avx512_skx")))
+             (ISA_LIST("avx512_skx", "avx2")))
 DISPATCH_ALL(partial_qsort,
              (ISA_LIST("avx512_icl")),
              (ISA_LIST("avx512_skx", "avx2")),
-             (ISA_LIST("avx512_skx")))
+             (ISA_LIST("avx512_skx", "avx2")))
 DISPATCH_ALL(argsort,
              (ISA_LIST("none")),
              (ISA_LIST("avx512_skx")),

diff --git a/src/avx2-32bit-qsort.hpp b/src/avx2-32bit-qsort.hpp
@@ -135,7 +135,7 @@ struct avx2_vector<int32_t> {
     }
     static void mask_compressstoreu(void *mem, opmask_t mask, reg_t x)
     {
-        return avx2_emu_mask_compressstoreu<type_t>(mem, mask, x);
+        return avx2_emu_mask_compressstoreu32<type_t>(mem, mask, x);
     }
     static reg_t maskz_loadu(opmask_t mask, void const *mem)
     {
@@ -289,7 +289,7 @@ struct avx2_vector<uint32_t> {
     }
     static void mask_compressstoreu(void *mem, opmask_t mask, reg_t x)
     {
-        return avx2_emu_mask_compressstoreu<type_t>(mem, mask, x);
+        return avx2_emu_mask_compressstoreu32<type_t>(mem, mask, x);
     }
     static reg_t mask_loadu(reg_t x, opmask_t mask, void const *mem)
     {
@@ -459,7 +459,7 @@ struct avx2_vector<float> {
     }
     static void mask_compressstoreu(void *mem, opmask_t mask, reg_t x)
     {
-        return avx2_emu_mask_compressstoreu<type_t>(mem, mask, x);
+        return avx2_emu_mask_compressstoreu32<type_t>(mem, mask, x);
     }
     static reg_t mask_loadu(reg_t x, opmask_t mask, void const *mem)
     {