avx512 set: add simde_x_mm512_set_m256{,d}

add tests for all `simde_mm512_set_*` functions
simd-everywhere · Oct 25, 2023 · 67e0c50 · 67e0c50
1 parent 2eedece
commit 67e0c50
Show file tree

Hide file tree

Showing 3 changed files with 756 additions and 1 deletion.
diff --git a/simde/x86/avx512/gather.h b/simde/x86/avx512/gather.h
@@ -29,6 +29,7 @@
 
 #include "types.h"
 #include "../avx2.h"
+#include "extract.h"
 
 HEDLEY_DIAGNOSTIC_PUSH
 SIMDE_DISABLE_UNWANTED_DIAGNOSTICS

diff --git a/simde/x86/avx512/set.h b/simde/x86/avx512/set.h
@@ -415,11 +415,27 @@ simde_x_mm512_set_m128i (simde__m128i a, simde__m128i b, simde__m128i c, simde__
   #endif
 }
 
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512
+simde_x_mm512_set_m256 (simde__m256 a, simde__m256 b) {
+  #if defined(SIMDE_X86_AVX512F_NATIVE)
+    SIMDE_ALIGN_TO_64 simde__m256 v[] = { b, a };
+    return simde_mm512_load_ps(HEDLEY_STATIC_CAST(__m512 *, HEDLEY_STATIC_CAST(void *, v)));
+  #else
+    simde__m512_private r_;
+
+    r_.m256[0] = b;
+    r_.m256[1] = a;
+
+    return simde__m512_from_private(r_);
+  #endif
+}
+
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m512i
 simde_x_mm512_set_m256i (simde__m256i a, simde__m256i b) {
   #if defined(SIMDE_X86_AVX512F_NATIVE)
-    SIMDE_ALIGN_LIKE_32(simde__m256i) simde__m256i v[] = { b, a };
+    SIMDE_ALIGN_TO_64 simde__m256i v[] = { b, a };
     return simde_mm512_load_si512(HEDLEY_STATIC_CAST(__m512i *, HEDLEY_STATIC_CAST(void *, v)));
   #else
     simde__m512i_private r_;
@@ -431,6 +447,22 @@ simde_x_mm512_set_m256i (simde__m256i a, simde__m256i b) {
   #endif
 }
 
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512d
+simde_x_mm512_set_m256d (simde__m256d a, simde__m256d b) {
+  #if defined(SIMDE_X86_AVX512F_NATIVE)
+    SIMDE_ALIGN_TO_64 simde__m256d v[] = { b, a };
+    return simde_mm512_load_pd(HEDLEY_STATIC_CAST(__m512d *, HEDLEY_STATIC_CAST(void *, v)));
+  #else
+    simde__m512d_private r_;
+
+    r_.m256d[0] = b;
+    r_.m256d[1] = a;
+
+    return simde__m512d_from_private(r_);
+  #endif
+}
+
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m512
 simde_mm512_set_ps (simde_float32 e15, simde_float32 e14, simde_float32 e13, simde_float32 e12,