avx512/shuffle: mm512_{shuffle_epi32,shuffle{hi,lo}_epi16}

x86/avx512/shuffle.h

@@ -23,6 +23,7 @@
  * Copyright:
  *   2020      Evan Nemerson <evan@nemerson.com>
  *   2020      Christopher Moore <moore@free.fr>
+ *   2023      Michael R. Crusoe <crusoe@debian.org>
  */
 
 #if !defined(SIMDE_X86_AVX512_SHUFFLE_H)
@@ -31,6 +32,7 @@
 #include "types.h"
 #include "../avx2.h"
 #include "mov.h"
+#include "extract.h"
 
 HEDLEY_DIAGNOSTIC_PUSH
 SIMDE_DISABLE_UNWANTED_DIAGNOSTICS
@@ -94,6 +96,35 @@ simde_mm512_maskz_shuffle_epi8 (simde__mmask64 k, simde__m512i a, simde__m512i b
   #define _mm512_maskz_shuffle_epi8(k, a, b) simde_mm512_maskz_shuffle_epi8(k, a, b)
 #endif
 
+#if defined(SIMDE_X86_AVX512F_NATIVE)
+#  define simde_mm512_shuffle_epi32(a, imm8) _mm512_shuffle_epi32((a), (imm8))
+#elif defined(SIMDE_STATEMENT_EXPR_)
+#  define simde_mm512_shuffle_epi32(a, imm8) SIMDE_STATEMENT_EXPR_(({ \
+     simde__m512i_private simde_mm512_shuffle_epi32_r_, \
+       simde_mm512_shuffle_epi32_a_ = simde__m512i_to_private((a)); \
+     simde_mm512_shuffle_epi32_r_.m128i[0] = simde_mm_shuffle_epi32( \
+       simde_mm512_shuffle_epi32_a_.m128i[0], (imm8)); \
+     simde_mm512_shuffle_epi32_r_.m128i[1] = simde_mm_shuffle_epi32( \
+       simde_mm512_shuffle_epi32_a_.m128i[1], (imm8)); \
+     simde_mm512_shuffle_epi32_r_.m128i[2] = simde_mm_shuffle_epi32( \
+       simde_mm512_shuffle_epi32_a_.m128i[2], (imm8)); \
+     simde_mm512_shuffle_epi32_r_.m128i[3] = simde_mm_shuffle_epi32( \
+       simde_mm512_shuffle_epi32_a_.m128i[3], (imm8)); \
+     simde__m512i_from_private(simde_mm512_shuffle_epi32_r_); \
+   }))
+#else
+#  define simde_mm512_shuffle_epi32(a, imm8) \
+     simde_x_mm512_set_m128i( \
+       simde_mm_shuffle_epi32(simde_mm512_extracti32x4_epi32(a, 3), (imm8)), \
+       simde_mm_shuffle_epi32(simde_mm512_extracti32x4_epi32(a, 2), (imm8)), \
+       simde_mm_shuffle_epi32(simde_mm512_extracti32x4_epi32(a, 1), (imm8)), \
+       simde_mm_shuffle_epi32(simde_mm512_extracti32x4_epi32(a, 0), (imm8)))
+#endif
+#if defined(SIMDE_X86_AVX512F_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_shuffle_epi32
+  #define _mm512_shuffle_epi32(a, imm8) simde_mm512_shuffle_epi32((a), (imm8))
+#endif
+
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m256i
 simde_mm256_shuffle_i32x4 (simde__m256i a, simde__m256i b, const int imm8)
@@ -266,6 +297,64 @@ simde_mm512_shuffle_pd(simde__m512d a, simde__m512d b, int imm8)
   #define _mm512_shuffle_pd(a, b, imm8) simde_mm512_shuffle_pd(a, b, imm8)
 #endif
 
+#if defined(SIMDE_X86_AVX512BW_NATIVE)
+#  define simde_mm512_shufflehi_epi16(a, imm8) _mm512_shufflehi_epi16(a, imm8)
+#elif defined(SIMDE_STATEMENT_EXPR_)
+#  define simde_mm512_shufflehi_epi16(a, imm8) SIMDE_STATEMENT_EXPR_(({ \
+    simde__m512i_private simde_mm512_shufflehi_epi16_r_, \
+      simde_mm512_shufflehi_epi16_a_ = simde__m512i_to_private((a)); \
+    simde_mm512_shufflehi_epi16_r_.m128i[0] = simde_mm_shufflehi_epi16( \
+      simde_mm512_shufflehi_epi16_a_.m128i[0], (imm8)); \
+    simde_mm512_shufflehi_epi16_r_.m128i[1] = simde_mm_shufflehi_epi16( \
+      simde_mm512_shufflehi_epi16_a_.m128i[1], (imm8)); \
+    simde_mm512_shufflehi_epi16_r_.m128i[2] = simde_mm_shufflehi_epi16( \
+      simde_mm512_shufflehi_epi16_a_.m128i[2], (imm8)); \
+    simde_mm512_shufflehi_epi16_r_.m128i[3] = simde_mm_shufflehi_epi16( \
+      simde_mm512_shufflehi_epi16_a_.m128i[3], (imm8)); \
+    simde__m512i_from_private(simde_mm512_shufflehi_epi16_r_); \
+  }))
+#else
+#  define simde_mm512_shufflehi_epi16(a, imm8) \
+     simde_x_mm512_set_m128i( \
+       simde_mm_shufflehi_epi16(simde_mm512_extracti32x4_epi32((a), 3), (imm8)), \
+       simde_mm_shufflehi_epi16(simde_mm512_extracti32x4_epi32((a), 2), (imm8)), \
+       simde_mm_shufflehi_epi16(simde_mm512_extracti32x4_epi32((a), 1), (imm8)), \
+       simde_mm_shufflehi_epi16(simde_mm512_extracti32x4_epi32((a), 0), (imm8)))
+#endif
+#if defined(SIMDE_X86_AVX512BW_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_shufflehi_epi16
+  #define _mm512_shufflehi_epi16(a, imm8) simde_mm512_shufflehi_epi16(a, imm8)
+#endif
+
+#if defined(SIMDE_X86_AVX512BW_NATIVE)
+#  define simde_mm512_shufflelo_epi16(a, imm8) _mm512_shufflelo_epi16(a, imm8)
+#elif defined(SIMDE_STATEMENT_EXPR_)
+#  define simde_mm512_shufflelo_epi16(a, imm8) SIMDE_STATEMENT_EXPR_(({ \
+    simde__m512i_private simde_mm512_shufflelo_epi16_r_, \
+      simde_mm512_shufflelo_epi16_a_ = simde__m512i_to_private((a)); \
+    simde_mm512_shufflelo_epi16_r_.m128i[0] = simde_mm_shufflelo_epi16( \
+      simde_mm512_shufflelo_epi16_a_.m128i[0], (imm8)); \
+    simde_mm512_shufflelo_epi16_r_.m128i[1] = simde_mm_shufflelo_epi16( \
+      simde_mm512_shufflelo_epi16_a_.m128i[1], (imm8)); \
+    simde_mm512_shufflelo_epi16_r_.m128i[2] = simde_mm_shufflelo_epi16( \
+      simde_mm512_shufflelo_epi16_a_.m128i[2], (imm8)); \
+    simde_mm512_shufflelo_epi16_r_.m128i[3] = simde_mm_shufflelo_epi16( \
+      simde_mm512_shufflelo_epi16_a_.m128i[3], (imm8)); \
+    simde__m512i_from_private(simde_mm512_shufflelo_epi16_r_); \
+  }))
+#else
+#  define simde_mm512_shufflelo_epi16(a, imm8) \
+     simde_x_mm512_set_m128i( \
+       simde_mm_shufflelo_epi16(simde_mm512_extracti32x4_epi32((a), 3), (imm8)), \
+       simde_mm_shufflelo_epi16(simde_mm512_extracti32x4_epi32((a), 2), (imm8)), \
+       simde_mm_shufflelo_epi16(simde_mm512_extracti32x4_epi32((a), 1), (imm8)), \
+       simde_mm_shufflelo_epi16(simde_mm512_extracti32x4_epi32((a), 0), (imm8)))
+#endif
+#if defined(SIMDE_X86_AVX512BW_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_shufflelo_epi16
+  #define _mm512_shufflelo_epi16(a, imm8) simde_mm512_shufflelo_epi16(a, imm8)
+#endif
+
 SIMDE_END_DECLS_
 HEDLEY_DIAGNOSTIC_POP