Fix dispatch logic to use both ICL and SPR fp16

Author: sterrettm2

lib/x86simdsort-icl.cpp

@@ -50,6 +50,8 @@ namespace avx512 {
     {
         x86simdsortStatic::partial_qsort(arr, k, arrsize, hasnan, descending);
     }
+} // namespace avx512
+namespace fp16_icl {
 #ifdef __FLT16_MAX__
     template <>
     void qsort(_Float16 *arr, size_t size, bool hasnan, bool descending)
@@ -75,5 +77,5 @@ namespace avx512 {
         x86simdsortStatic::partial_qsort(arr, k, arrsize, hasnan, descending);
     }
 #endif
-} // namespace avx512
+} // namespace fp16_icl
 } // namespace xss

lib/x86simdsort-internal.h

@@ -164,5 +164,45 @@ namespace scalar {
     XSS_HIDE_SYMBOL std::vector<size_t>
     argselect(T *arr, size_t k, size_t arrsize, bool hasnan = false);
 } // namespace scalar
+namespace fp16_spr {
+    // quicksort
+    template <typename T>
+    XSS_HIDE_SYMBOL void
+    qsort(T *arr, size_t arrsize, bool hasnan = false, bool descending = false);
+    // quickselect
+    template <typename T>
+    XSS_HIDE_SYMBOL void qselect(T *arr,
+                                 size_t k,
+                                 size_t arrsize,
+                                 bool hasnan = false,
+                                 bool descending = false);
+    // partial sort
+    template <typename T>
+    XSS_HIDE_SYMBOL void partial_qsort(T *arr,
+                                       size_t k,
+                                       size_t arrsize,
+                                       bool hasnan = false,
+                                       bool descending = false);
+} // namespace fp16_spr
+namespace fp16_icl {
+    // quicksort
+    template <typename T>
+    XSS_HIDE_SYMBOL void
+    qsort(T *arr, size_t arrsize, bool hasnan = false, bool descending = false);
+    // quickselect
+    template <typename T>
+    XSS_HIDE_SYMBOL void qselect(T *arr,
+                                 size_t k,
+                                 size_t arrsize,
+                                 bool hasnan = false,
+                                 bool descending = false);
+    // partial sort
+    template <typename T>
+    XSS_HIDE_SYMBOL void partial_qsort(T *arr,
+                                       size_t k,
+                                       size_t arrsize,
+                                       bool hasnan = false,
+                                       bool descending = false);
+} // namespace fp16_icl
 } // namespace xss
 #endif

lib/x86simdsort-spr.cpp

@@ -3,7 +3,7 @@
 #include "x86simdsort-internal.h"
 
 namespace xss {
-namespace avx512 {
+namespace fp16_spr {
     template <>
     void qsort(_Float16 *arr, size_t size, bool hasnan, bool descending)
     {
@@ -27,5 +27,5 @@ namespace avx512 {
     {
         x86simdsortStatic::partial_qsort(arr, k, arrsize, hasnan, descending);
     }
-} // namespace avx512
+} // namespace fp16_spr
 } // namespace xss

lib/x86simdsort.cpp

@@ -130,16 +130,37 @@ namespace x86simdsort {
         } \
     }
 
+#define DISPATCH_FP16(func, TYPE, ISA) \
+    DECLARE_INTERNAL_##func(TYPE) static __attribute__((constructor)) void \
+    CAT(CAT(resolve_, func), TYPE)(void) \
+    { \
+        CAT(CAT(internal_, func), TYPE) = &xss::scalar::func<TYPE>; \
+        __builtin_cpu_init(); \
+        std::string_view preferred_cpu = find_preferred_cpu(ISA); \
+        if constexpr (dispatch_requested("avx512_spr", ISA)) { \
+            if (preferred_cpu.find("avx512_spr") != std::string_view::npos) { \
+                CAT(CAT(internal_, func), TYPE) = &xss::fp16_spr::func<TYPE>; \
+                return; \
+            } \
+        } \
+        if constexpr (dispatch_requested("avx512_icl", ISA)) { \
+            if (preferred_cpu.find("avx512_icl") != std::string_view::npos) { \
+                CAT(CAT(internal_, func), TYPE) = &xss::fp16_icl::func<TYPE>; \
+                return; \
+            } \
+        } \
+    }
+
 #define ISA_LIST(...) \
     std::initializer_list<std::string_view> \
     { \
         __VA_ARGS__ \
     }
 
 #ifdef __FLT16_MAX__
-DISPATCH(qsort, _Float16, ISA_LIST("avx512_spr", "avx512_icl"))
-DISPATCH(qselect, _Float16, ISA_LIST("avx512_spr", "avx512_icl"))
-DISPATCH(partial_qsort, _Float16, ISA_LIST("avx512_spr", "avx512_icl"))
+DISPATCH_FP16(qsort, _Float16, ISA_LIST("avx512_spr", "avx512_icl"))
+DISPATCH_FP16(qselect, _Float16, ISA_LIST("avx512_spr", "avx512_icl"))
+DISPATCH_FP16(partial_qsort, _Float16, ISA_LIST("avx512_spr", "avx512_icl"))
 DISPATCH(argsort, _Float16, ISA_LIST("none"))
 DISPATCH(argselect, _Float16, ISA_LIST("none"))
 #endif