HalfFloat_SSE2 port from C algorithm to SSE2

Low level support for 12 bit 420, 422 and 444 YUV video frame conversion.

BUG=libyuv:560, chromium:445071
TEST=untested
R=hubbe@chromium.org

Review URL: https://codereview.chromium.org/2381493006 .

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1621
+Version: 1622
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -262,6 +262,13 @@ extern "C" {
 #define HAS_I422TOARGBROW_SSSE3
 #endif
 
+// The following are available on gcc x86 platforms:
+// TODO(fbarchard): Port to Visual C.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#define HAS_HALFFLOATROW_SSE2
+#endif
+
 // The following are available on Neon platforms:
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
@@ -1938,6 +1945,9 @@ void HalfFloatRow_C(const uint16* src, uint16* dst, float scale, int width);
 void HalfFloatRow_AVX2(const uint16* src, uint16* dst, float scale, int width);
 void HalfFloatRow_Any_AVX2(const uint16* src, uint16* dst, float scale,
                            int width);
+void HalfFloatRow_SSE2(const uint16* src, uint16* dst, float scale, int width);
+void HalfFloatRow_Any_SSE2(const uint16* src, uint16* dst, float scale,
+                           int width);
 
 void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
                              const uint8* luma, uint32 lumacoeff);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1621
+#define LIBYUV_VERSION 1622
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/planar_functions.cc

@@ -2470,6 +2470,14 @@ int HalfFloatPlane(const uint16* src_y, int src_stride_y,
     height = 1;
     src_stride_y = dst_stride_y = 0;
   }
+#if defined(HAS_HALFFLOATROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    HalfFloatRow = HalfFloatRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      HalfFloatRow = HalfFloatRow_SSE2;
+    }
+  }
+#endif
 #if defined(HAS_HALFFLOATROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     HalfFloatRow = HalfFloatRow_Any_AVX2;
@@ -2478,6 +2486,14 @@ int HalfFloatPlane(const uint16* src_y, int src_stride_y,
     }
   }
 #endif
+#if defined(HAS_HALFFLOATROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+//    HalfFloatRow = HalfFloatRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      HalfFloatRow = HalfFloatRow_AVX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     HalfFloatRow(src_y, dst_y, scale, width);

source/row_any.cc

@@ -562,6 +562,9 @@ ANY11P(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON, const uint8*, 4, 4, 3)
       memcpy(dst_ptr + n * BPP, temp + 64, r * BPP);                           \
     }
 
+#ifdef HAS_HALFFLOATROW_SSE2
+ANY11P16(HalfFloatRow_Any_SSE2, HalfFloatRow_SSE2, float, 1, 1, 15)
+#endif
 #ifdef HAS_HALFFLOATROW_AVX2
 ANY11P16(HalfFloatRow_Any_AVX2, HalfFloatRow_AVX2, float, 1, 1, 15)
 #endif

source/row_gcc.cc

@@ -5366,6 +5366,43 @@ void ARGBPolynomialRow_AVX2(const uint8* src_argb,
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_AVX2
 
+#ifdef HAS_HALFFLOATROW_SSE2
+void HalfFloatRow_SSE2(const uint16* src, uint16* dst, float scale, int width) {
+  float mult = 1.9259299444e-34f * scale;
+  asm volatile (
+    "movd        %3,%%xmm4                     \n"
+    "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+    "pxor        %%xmm5,%%xmm5                 \n"
+
+    // 16 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu      " MEMACCESS(0) ",%%xmm0       \n"  // 8 shorts
+    "lea         " MEMLEA(0x10,0) ",%0         \n"
+    "movdqa      %%xmm0,%%xmm1                 \n"
+    "punpcklwd   %%xmm5,%%xmm0                 \n"  // 8 ints in xmm0/1
+    "cvtdq2ps    %%xmm0,%%xmm0                 \n"  // 8 floats
+    "punpckhwd   %%xmm5,%%xmm1                 \n"
+    "cvtdq2ps    %%xmm1,%%xmm1                 \n"
+    "mulps       %%xmm4,%%xmm0                 \n"
+    "mulps       %%xmm4,%%xmm1                 \n"
+    "psrld       $0xd,%%xmm0                   \n"
+    "psrld       $0xd,%%xmm1                   \n"
+    "packssdw    %%xmm1,%%xmm0                 \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub         $0x8,%2                       \n"
+    "jg          1b                            \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  : "rm"(mult)   // %3
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_HALFFLOATROW_SSE2
+
 #ifdef HAS_HALFFLOATROW_AVX2
 void HalfFloatRow_AVX2(const uint16* src, uint16* dst, float scale, int width) {
   asm volatile (
@@ -5394,7 +5431,7 @@ void HalfFloatRow_AVX2(const uint16* src, uint16* dst, float scale, int width) {
     "+r"(width)  // %2
   : "x"(scale)   // %3
   : "memory", "cc",
-    "xmm0", "xmm4"
+    "xmm0", "xmm1", "xmm4"
   );
 }
 #endif  // HAS_HALFFLOATROW_AVX2