avx512: add _mm512_{,mask_,maskz_}permutexvar_epi{8,16} intrinsics

Add _mm512_{,mask_,maskz_}permutexvar_epi{8,16} intrinsics to permutexvar.[ch]

Add NEON translations for _mm512_permutexvar_epi{8,16,32}
This could easily be done for _mm512_permutexvar_epi64 too.
However it may not be faster than the portable version.
There are decreasing returns as the size increases.

x86/avx512/permutexvar.h

@@ -29,12 +29,89 @@
 #define SIMDE_X86_AVX512_PERMUTEXVAR_H
 
 #include "types.h"
+#include "and.h"
+#include "andnot.h"
+#include "blend.h"
 #include "mov.h"
+#include "or.h"
+#include "set1.h"
+#include "slli.h"
+#include "srli.h"
+#include "test.h"
 
 HEDLEY_DIAGNOSTIC_PUSH
 SIMDE_DISABLE_UNWANTED_DIAGNOSTICS
 SIMDE_BEGIN_DECLS_
 
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512i
+simde_mm512_permutexvar_epi16 (simde__m512i idx, simde__m512i a) {
+  #if defined(SIMDE_X86_AVX512BW_NATIVE)
+    return _mm512_permutexvar_epi16(idx, a);
+  #else
+    simde__m512i_private
+      idx_ = simde__m512i_to_private(idx),
+      a_ = simde__m512i_to_private(a),
+      r_;
+
+    #if defined(SIMDE_ARM_NEON_A64V8_NATIVE)
+      uint8x16x4_t table = { { a_.m128i_private[0].neon_u8,
+                               a_.m128i_private[1].neon_u8,
+                               a_.m128i_private[2].neon_u8,
+                               a_.m128i_private[3].neon_u8 } };
+      uint16x8_t mask16 = vdupq_n_u16(0x001F);
+      uint16x8_t byte_index16 = vdupq_n_u16(0x0100);
+
+      SIMDE_VECTORIZE
+      for (size_t i = 0 ; i < (sizeof(r_.m128i_private) / sizeof(r_.m128i_private[0])) ; i++) {
+        uint16x8_t index16 = vandq_u16(idx_.m128i_private[i].neon_u16, mask16);
+        index16 = vmulq_n_u16(index16, 0x0202);
+        index16 = vaddq_u16(index16, byte_index16);
+        r_.m128i_private[i].neon_u8 = vqtbl4q_u8(table, vreinterpretq_u8_u16(index16));
+      }
+    #else
+      SIMDE_VECTORIZE
+      for (size_t i = 0 ; i < (sizeof(r_.i16) / sizeof(r_.i16[0])) ; i++) {
+        r_.i16[i] = a_.i16[idx_.i16[i] & 0x1F];
+      }
+    #endif
+
+    return simde__m512i_from_private(r_);
+  #endif
+}
+#if defined(SIMDE_X86_AVX512BW_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_permutexvar_epi16
+  #define _mm512_permutexvar_epi16(idx, a) simde_mm512_permutexvar_epi16(idx, a)
+#endif
+
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512i
+simde_mm512_mask_permutexvar_epi16 (simde__m512i src, simde__mmask32 k, simde__m512i idx, simde__m512i a) {
+  #if defined(SIMDE_X86_AVX512BW_NATIVE)
+    return _mm512_mask_permutexvar_epi16(src, k, idx, a);
+  #else
+    return simde_mm512_mask_mov_epi16(src, k, simde_mm512_permutexvar_epi16(idx, a));
+  #endif
+}
+#if defined(SIMDE_X86_AVX512BW_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_mask_permutexvar_epi16
+  #define _mm512_mask_permutexvar_epi16(src, k, idx, a) simde_mm512_mask_permutexvar_epi16(src, k, idx, a)
+#endif
+
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512i
+simde_mm512_maskz_permutexvar_epi16 (simde__mmask32 k, simde__m512i idx, simde__m512i a) {
+  #if defined(SIMDE_X86_AVX512BW_NATIVE)
+    return _mm512_maskz_permutexvar_epi16(k, idx, a);
+  #else
+    return simde_mm512_maskz_mov_epi16(k, simde_mm512_permutexvar_epi16(idx, a));
+  #endif
+}
+#if defined(SIMDE_X86_AVX512BW_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_maskz_permutexvar_epi16
+  #define _mm512_maskz_permutexvar_epi16(k, idx, a) simde_mm512_maskz_permutexvar_epi16(k, idx, a)
+#endif
+
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m512i
 simde_mm512_permutexvar_epi32 (simde__m512i idx, simde__m512i a) {
@@ -46,12 +123,29 @@ simde_mm512_permutexvar_epi32 (simde__m512i idx, simde__m512i a) {
       a_ = simde__m512i_to_private(a),
       r_;
 
-    #if !defined(__INTEL_COMPILER)
+    #if defined(SIMDE_ARM_NEON_A64V8_NATIVE)
+      uint8x16x4_t table = { { a_.m128i_private[0].neon_u8,
+                               a_.m128i_private[1].neon_u8,
+                               a_.m128i_private[2].neon_u8,
+                               a_.m128i_private[3].neon_u8 } };
+      uint32x4_t mask32 = vdupq_n_u32(0x0000000F);
+      uint32x4_t byte_index32 = vdupq_n_u32(0x03020100);
+
       SIMDE_VECTORIZE
+      for (size_t i = 0 ; i < (sizeof(r_.m128i_private) / sizeof(r_.m128i_private[0])) ; i++) {
+        uint32x4_t index32 = vandq_u32(idx_.m128i_private[i].neon_u32, mask32);
+        index32 = vmulq_n_u32(index32, 0x04040404);
+        index32 = vaddq_u32(index32, byte_index32);
+        r_.m128i_private[i].neon_u8 = vqtbl4q_u8(table, vreinterpretq_u8_u32(index32));
+      }
+    #else
+      #if !defined(__INTEL_COMPILER)
+        SIMDE_VECTORIZE
+      #endif
+      for (size_t i = 0 ; i < (sizeof(r_.i32) / sizeof(r_.i32[0])) ; i++) {
+        r_.i32[i] = a_.i32[idx_.i32[i] & 0x0F];
+      }
     #endif
-    for (size_t i = 0 ; i < (sizeof(r_.i32) / sizeof(r_.i32[0])) ; i++) {
-      r_.i32[i] = a_.i32[idx_.i32[i] & 0x0F];
-    }
 
     return simde__m512i_from_private(r_);
   #endif
@@ -143,6 +237,91 @@ simde_mm512_maskz_permutexvar_epi64 (simde__mmask8 k, simde__m512i idx, simde__m
   #define _mm512_maskz_permutexvar_epi64(k, idx, a) simde_mm512_maskz_permutexvar_epi64(k, idx, a)
 #endif
 
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512i
+simde_mm512_permutexvar_epi8 (simde__m512i idx, simde__m512i a) {
+  #if defined(SIMDE_X86_AVX512VBMI_NATIVE)
+    return _mm512_permutexvar_epi8(idx, a);
+  #elif defined(SIMDE_X86_AVX512BW_NATIVE) && defined(SIMDE_X86_AVX512VL_NATIVE)
+    simde__m512i hilo, hi, lo, hi2, lo2, idx2;
+    simde__m512i ones = simde_mm512_set1_epi8(1);
+    simde__m512i low_bytes = simde_mm512_set1_epi16(0x00FF);
+
+    idx2 = simde_mm512_srli_epi16(idx, 1);
+    hilo = simde_mm512_permutexvar_epi16(idx2, a);
+    simde__mmask64 mask = simde_mm512_test_epi8_mask(idx, ones);
+    lo = simde_mm512_and_si512(hilo, low_bytes);
+    hi = simde_mm512_srli_epi16(hilo, 8);
+
+    idx2 = simde_mm512_srli_epi16(idx, 9);
+    hilo = simde_mm512_permutexvar_epi16(idx2, a);
+    lo2 = simde_mm512_slli_epi16(hilo, 8);
+    hi2 = simde_mm512_andnot_si512(low_bytes, hilo);
+
+    lo = simde_mm512_or_si512(lo, lo2);
+    hi = simde_mm512_or_si512(hi, hi2);
+
+    return simde_mm512_mask_blend_epi8(mask, lo, hi);
+  #else
+    simde__m512i_private
+      idx_ = simde__m512i_to_private(idx),
+      a_ = simde__m512i_to_private(a),
+      r_;
+
+    #if defined(SIMDE_ARM_NEON_A64V8_NATIVE)
+      uint8x16x4_t table = { { a_.m128i_private[0].neon_u8,
+                               a_.m128i_private[1].neon_u8,
+                               a_.m128i_private[2].neon_u8,
+                               a_.m128i_private[3].neon_u8 } };
+      uint8x16_t mask = vdupq_n_u8(0x3F);
+
+      SIMDE_VECTORIZE
+      for (size_t i = 0 ; i < (sizeof(r_.m128i_private) / sizeof(r_.m128i_private[0])) ; i++) {
+        r_.m128i_private[i].neon_u8 = vqtbl4q_u8(table, vandq_u8(idx_.m128i_private[i].neon_u8, mask));
+      }
+    #else
+      SIMDE_VECTORIZE
+      for (size_t i = 0 ; i < (sizeof(r_.i8) / sizeof(r_.i8[0])) ; i++) {
+        r_.i8[i] = a_.i8[idx_.i8[i] & 0x3F];
+      }
+    #endif
+
+    return simde__m512i_from_private(r_);
+  #endif
+}
+#if defined(SIMDE_X86_AVX512VBMI_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_permutexvar_epi8
+  #define _mm512_permutexvar_epi8(idx, a) simde_mm512_permutexvar_epi8(idx, a)
+#endif
+
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512i
+simde_mm512_mask_permutexvar_epi8 (simde__m512i src, simde__mmask64 k, simde__m512i idx, simde__m512i a) {
+  #if defined(SIMDE_X86_AVX512VBMI_NATIVE)
+    return _mm512_mask_permutexvar_epi8(src, k, idx, a);
+  #else
+    return simde_mm512_mask_mov_epi8(src, k, simde_mm512_permutexvar_epi8(idx, a));
+  #endif
+}
+#if defined(SIMDE_X86_AVX512VBMI_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_mask_permutexvar_epi8
+  #define _mm512_mask_permutexvar_epi8(src, k, idx, a) simde_mm512_mask_permutexvar_epi8(src, k, idx, a)
+#endif
+
+SIMDE_FUNCTION_ATTRIBUTES
+simde__m512i
+simde_mm512_maskz_permutexvar_epi8 (simde__mmask64 k, simde__m512i idx, simde__m512i a) {
+  #if defined(SIMDE_X86_AVX512VBMI_NATIVE)
+    return _mm512_maskz_permutexvar_epi8(k, idx, a);
+  #else
+    return simde_mm512_maskz_mov_epi8(k, simde_mm512_permutexvar_epi8(idx, a));
+  #endif
+}
+#if defined(SIMDE_X86_AVX512VBMI_ENABLE_NATIVE_ALIASES)
+  #undef _mm512_maskz_permutexvar_epi8
+  #define _mm512_maskz_permutexvar_epi8(k, idx, a) simde_mm512_maskz_permutexvar_epi8(k, idx, a)
+#endif
+
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m512d
 simde_mm512_permutexvar_pd (simde__m512i idx, simde__m512d a) {