Optimize SSE2 packing intrinsics (#12301)

Avoid scalarization in implementation of
_mm_packs_epi16/_mm_packs_epi32/_mm_packus_epi16

AUTHORS

@@ -508,3 +508,4 @@ a license to everyone to use it as detailed in LICENSE.)
 * Georg Rottensteiner <georg@georg-rottensteiner.de>
 * Julien Jorge <julien.jorge@gmx.fr>
 * Attila Oláh <atl@google.com> (copyright owned by Google, LLC)
+* Marat Dukhan <maratek@google.com> (copyright owned by Google, LLC)

site/source/docs/porting/simd.rst

@@ -546,11 +546,11 @@ The following table highlights the availability and expected performance of diff
    * - _mm_or_si128
      - 🟡 wasm_v128_or. VM must guess type.
    * - _mm_packs_epi16
-     - ❌ scalarized
+     - ✅ wasm_i8x16_narrow_i16x8
    * - _mm_packs_epi32
-     - ❌ scalarized
+     - ✅ wasm_i16x8_narrow_i32x4
    * - _mm_packus_epi16
-     - ❌ scalarized
+     - ✅ wasm_u8x16_narrow_i16x8
    * - _mm_pause
      - 💭 No-op.
    * - _mm_sad_epu8

system/include/SSE/emmintrin.h

@@ -14,7 +14,6 @@
 #include <xmmintrin.h>
 #include <emscripten/emscripten.h>
 
-#define __SATURATE(x, Min, Max) ((x) >= Min ? ((x) <= Max ? (x) : Max) : Min)
 #define __MIN(x, y) ((x) <= (y) ? (x) : (y))
 #define __MAX(x, y) ((x) >= (y) ? (x) : (y))
 
@@ -1366,67 +1365,19 @@ _mm_mfence(void)
 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_packs_epi16(__m128i __a, __m128i __b)
 {
-  // TODO: optimize
-  union {
-    signed short x[8];
-    __m128i m;
-  } src, src2;
-  union {
-    signed char x[16];
-    __m128i m;
-  } dst;
-  src.m = __a;
-  src2.m = __b;
-  for(int i = 0; i < 8; ++i)
-  {
-    dst.x[i] = __SATURATE(src.x[i], -128, 127);
-    dst.x[8+i] = __SATURATE(src2.x[i], -128, 127);
-  }
-  return dst.m;
+  return (__m128i)wasm_i8x16_narrow_i16x8(__a, __b);
 }
 
 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_packs_epi32(__m128i __a, __m128i __b)
 {
-  // TODO: optimize
-  union {
-    signed int x[4];
-    __m128i m;
-  } src, src2;
-  union {
-    signed short x[8];
-    __m128i m;
-  } dst;
-  src.m = __a;
-  src2.m = __b;
-  for(int i = 0; i < 4; ++i)
-  {
-    dst.x[i] = __SATURATE(src.x[i], -32768, 32767);
-    dst.x[4+i] = __SATURATE(src2.x[i], -32768, 32767);
-  }
-  return dst.m;
+  return (__m128i)wasm_i16x8_narrow_i32x4(__a, __b);
 }
 
 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_packus_epi16(__m128i __a, __m128i __b)
 {
-  // TODO: optimize
-  union {
-    signed short x[8];
-    __m128i m;
-  } src, src2;
-  union {
-    unsigned char x[16];
-    __m128i m;
-  } dst;
-  src.m = __a;
-  src2.m = __b;
-  for(int i = 0; i < 8; ++i)
-  {
-    dst.x[i] = __SATURATE(src.x[i], 0, 255);
-    dst.x[8+i] = __SATURATE(src2.x[i], 0, 255);
-  }
-  return dst.m;
+  return (__m128i)wasm_u8x16_narrow_i16x8(__a, __b);
 }
 
 #define _mm_extract_epi16(__a, __imm) wasm_u16x8_extract_lane((v128_t)(__a), (__imm) & 7)