Port HalfFloatRow_SSE2 to AVX2 but not using F16C.

source/row_win.cc

 /*
  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
@@ skipping 6041 matching lines @@
   }
 }
 #endif  // HAS_HALFFLOATROW_SSE2
 
 #ifdef HAS_HALFFLOATROW_AVX2
 __declspec(naked)
 void HalfFloatRow_AVX2(const uint16* src, uint16* dst, float scale, int width) {
   __asm {
     mov        eax, [esp + 4]      /* src */
     mov        edx, [esp + 8]      /* dst */
+    movd       xmm4, dword ptr [esp + 12]  /* scale */
+    mov        ecx, [esp + 16]     /* width */
+
+    vmulss     xmm4, xmm4, kExpBias
+    vbroadcastss ymm4, xmm4
+    vpxor      ymm5, ymm5, ymm5
+
+    // 16 pixel loop.
+ convertloop:
+    vmovdqu     ymm2, [eax]  // 16 shorts
+    lea         eax, [eax + 32]
+    vpunpckhwd  ymm3, ymm2, ymm5 // convert 16 shorts to 16 ints
+    vpunpcklwd  ymm2, ymm2, ymm5
+    vcvtdq2ps   ymm3, ymm3  // convert 16 ints to floats
+    vcvtdq2ps   ymm2, ymm2
+    vmulps      ymm3, ymm3, ymm4  // scale to adjust exponent for 5 bit range.
+    vmulps      ymm2, ymm2, ymm4
+    vpsrld      ymm3, ymm3, 13  // float convert to 8 half floats truncate
+    vpsrld      ymm2, ymm2, 13
+    vpackssdw   ymm2, ymm2, ymm3
+    vmovdqu     [edx], ymm2
+    lea         edx, [edx + 32]
+    sub         ecx, 16
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_HALFFLOATROW_AVX2
+
+#ifdef HAS_HALFFLOATROW_F16C
+__declspec(naked)
+void HalfFloatRow_F16C(const uint16* src, uint16* dst, float scale, int width) {
+  __asm {
+    mov        eax, [esp + 4]      /* src */
+    mov        edx, [esp + 8]      /* dst */
     vbroadcastss ymm4, [esp + 12]  /* scale */
     mov        ecx, [esp + 16]     /* width */
 
-    // 8 pixel loop.
+    // 16 pixel loop.
  convertloop:
     vpmovzxwd   ymm2, xmmword ptr [eax]  // 8 shorts -> 8 ints
     vpmovzxwd   ymm3, xmmword ptr [eax + 16]  // 8 more shorts
     lea         eax, [eax + 32]
     vcvtdq2ps   ymm2, ymm2        // convert 8 ints to floats
     vcvtdq2ps   ymm3, ymm3
     vmulps      ymm2, ymm2, ymm4  // scale to normalized range 0 to 1
     vmulps      ymm3, ymm3, ymm4
     vcvtps2ph   xmm2, ymm2, 3     // float convert to 8 half floats truncate
     vcvtps2ph   xmm3, ymm3, 3
     vmovdqu     [edx], xmm2
     vmovdqu     [edx + 16], xmm3
     lea         edx, [edx + 32]
     sub         ecx, 16
     jg          convertloop
     vzeroupper
     ret
   }
 }
-#endif  // HAS_HALFFLOATROW_AVX2
+#endif  // HAS_HALFFLOATROW_F16C
 
 #ifdef HAS_ARGBCOLORTABLEROW_X86
 // Tranform ARGB pixels with color table.
 __declspec(naked)
 void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
                            int width) {
   __asm {
     push       esi
     mov        eax, [esp + 4 + 4]   /* dst_argb */
     mov        esi, [esp + 4 + 8]   /* table_argb */
@@ skipping 153 matching lines @@
 #endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
 
 #endif  // defined(_M_X64)
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif
 
 #endif  // !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64))