+add NEON optimizations of function DescrIntDecode16f (Decode16f8).

ermig1979 · Sep 27, 2023 · 7b7a0fd · 7b7a0fd
1 parent 6626a70
commit 7b7a0fd
Show file tree

Hide file tree

Showing 4 changed files with 39 additions and 6 deletions.
diff --git a/docs/2023.html b/docs/2023.html
@@ -59,6 +59,7 @@ <h5>New features</h5>
  <li>NEON optimizations of function DescrIntEncode32f.</li>
  <li>NEON optimizations of function DescrIntEncode16f.</li>
  <li>NEON optimizations of function DescrIntDecode32f.</li>
+ <li>NEON optimizations of function DescrIntDecode16f.</li>
 </ul>
 <h5>Bug fixing</h5>
 <ul>

diff --git a/src/Simd/SimdNeonDescrInt.cpp b/src/Simd/SimdNeonDescrInt.cpp
@@ -102,7 +102,7 @@ namespace Simd
             _encode16f = GetEncode16f(_depth);
 
             _decode32f = GetDecode32f(_depth);
-            if (_depth >= 9) _decode16f = GetDecode16f(_depth);
+            if (_depth >= 8) _decode16f = GetDecode16f(_depth);
 
             //_cosineDistance = GetCosineDistance(_depth);
             //_macroCosineDistancesDirect = GetMacroCosineDistancesDirect(_depth);

diff --git a/src/Simd/SimdNeonDescrIntDec.cpp b/src/Simd/SimdNeonDescrIntDec.cpp
@@ -219,6 +219,38 @@ namespace Simd
 
         //-------------------------------------------------------------------------------------------------
 
+        static void Decode16f8(const uint8_t* src, float scale, float shift, size_t size, uint16_t * dst)
+        {
+            assert(size % 8 == 0);
+            float32x4_t _scale = vdupq_n_f32(scale);
+            float32x4_t _shift = vdupq_n_f32(shift);
+            size_t i = 0;
+            if (Aligned(src) && Aligned(dst))
+            {
+                for (; i < size; i += 8)
+                {
+                    uint16x8_t u16 = vmovl_u8(LoadHalf<true>(src));
+                    Store<true>(dst + 0, (uint16x4_t)vcvt_f16_f32(vmlaq_f32(_shift, _scale, vcvtq_f32_u32(UnpackU16<0>(u16)))));
+                    Store<true>(dst + 4, (uint16x4_t)vcvt_f16_f32(vmlaq_f32(_shift, _scale, vcvtq_f32_u32(UnpackU16<1>(u16)))));
+                    src += 8;
+                    dst += 8;
+                }
+            }
+            else
+            {
+                for (; i < size; i += 8)
+                {
+                    uint16x8_t u16 = vmovl_u8(LoadHalf<false>(src));
+                    Store<false>(dst + 0, (uint16x4_t)vcvt_f16_f32(vmlaq_f32(_shift, _scale, vcvtq_f32_u32(UnpackU16<0>(u16)))));
+                    Store<false>(dst + 4, (uint16x4_t)vcvt_f16_f32(vmlaq_f32(_shift, _scale, vcvtq_f32_u32(UnpackU16<1>(u16)))));
+                    src += 8;
+                    dst += 8;
+                }
+            }
+        }
+
+        //-------------------------------------------------------------------------------------------------
+
         Base::DescrInt::Decode32fPtr GetDecode32f(size_t depth)
         {
             switch (depth)
@@ -240,7 +272,7 @@ namespace Simd
             //case 5: return Decode16f5;
             //case 6: return Decode16f6;
             //case 7: return Decode16f7;
-            //case 8: return Decode16f8;
+            case 8: return Decode16f8;
             default: assert(0); return NULL;
             }
         }

diff --git a/src/Test/TestDescrInt.cpp b/src/Test/TestDescrInt.cpp
@@ -453,10 +453,10 @@ namespace Test
             result = result && DescrIntDecode16fAutoTest(FUNC_DI(Simd::Avx512bw::DescrIntInit), FUNC_DI(SimdDescrIntInit));
 #endif 
 
-//#ifdef SIMD_NEON_ENABLE
-//        if (Simd::Neon::Enable)
-//            result = result && DescrIntDecode16fAutoTest(FUNC_DI(Simd::Neon::DescrIntInit), FUNC_DI(SimdDescrIntInit));
-//#endif 
+#ifdef SIMD_NEON_ENABLE
+        if (Simd::Neon::Enable)
+            result = result && DescrIntDecode16fAutoTest(FUNC_DI(Simd::Neon::DescrIntInit), FUNC_DI(SimdDescrIntInit));
+#endif 
 
         return result;
     }