H010ToARGB_AVX2 optimized conversion

AVX2 optimized 10 bit YUV to ARGB.

Bug: libyuv:751
Test: H010ToARGB unittest
Change-Id: I705630beb62714b52042c2a5dcdb8b7859e734ae
Reviewed-on: https://chromium-review.googlesource.com/852563
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Miguel Casas <mcasas@chromium.org>
Reviewed-by: richard winterton <rrwinterton@gmail.com>

include/libyuv/row.h

@@ -279,6 +279,7 @@ extern "C" {
 #define HAS_ARGBTOAR30ROW_AVX2
 #define HAS_CONVERT16TO8ROW_AVX2
 #define HAS_CONVERT8TO16ROW_AVX2
+#define HAS_I210TOARGBROW_AVX2
 #define HAS_MERGEUVROW_16_AVX2
 #define HAS_MULTIPLYROW_16_AVX2
 #endif
@@ -1850,6 +1851,12 @@ void I210ToARGBRow_SSSE3(const uint16* src_y,
                          uint8* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
+void I210ToARGBRow_AVX2(const uint16* src_y,
+                        const uint16* src_u,
+                        const uint16* src_v,
+                        uint8* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
 void I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
                               const uint8* u_buf,
                               const uint8* v_buf,
@@ -2000,6 +2007,12 @@ void I210ToARGBRow_Any_SSSE3(const uint16* src_y,
                              uint8* dst_argb,
                              const struct YuvConstants* yuvconstants,
                              int width);
+void I210ToARGBRow_Any_AVX2(const uint16* src_y,
+                            const uint16* src_u,
+                            const uint16* src_v,
+                            uint8* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I422AlphaToARGBRow_Any_SSSE3(const uint8* y_buf,
                                   const uint8* u_buf,
                                   const uint8* v_buf,

source/convert_argb.cc

@@ -448,6 +448,14 @@ static int H010ToAR30Matrix(const uint16* src_y,
     }
   }
 #endif
+#if defined(HAS_I210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210ToARGBRow = I210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210ToARGBRow = I210ToARGBRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOAR30ROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToAR30Row = ARGBToAR30Row_Any_SSSE3;
@@ -537,7 +545,14 @@ static int I010ToARGBMatrix(const uint16* src_y,
     }
   }
 #endif
-
+#if defined(HAS_I210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210ToARGBRow = I210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210ToARGBRow = I210ToARGBRow_AVX2;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     I210ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
     dst_argb += dst_stride_argb;

source/row_any.cc

@@ -194,9 +194,8 @@ ANY31C(I422ToRGB565Row_Any_MSA, I422ToRGB565Row_MSA, 1, 0, 2, 7)
 #endif
 #undef ANY31C
 
-// 64 byte per row for future AVX2
 // Any 3 planes of 16 bit to 1 with yuvconstants
-// TODO(fbarchard): consider
+// TODO(fbarchard): consider sharing this code with ANY31C
 #define ANY31CT(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, T, SBPP, BPP, MASK)      \
   void NAMEANY(const T* y_buf, const T* u_buf, const T* v_buf, uint8* dst_ptr, \
                const struct YuvConstants* yuvconstants, int width) {           \
@@ -218,6 +217,9 @@ ANY31C(I422ToRGB565Row_Any_MSA, I422ToRGB565Row_MSA, 1, 0, 2, 7)
 #ifdef HAS_I210TOARGBROW_SSSE3
 ANY31CT(I210ToARGBRow_Any_SSSE3, I210ToARGBRow_SSSE3, 1, 0, uint16, 2, 4, 7)
 #endif
+#ifdef HAS_I210TOARGBROW_AVX2
+ANY31CT(I210ToARGBRow_Any_AVX2, I210ToARGBRow_AVX2, 1, 0, uint16, 2, 4, 15)
+#endif
 #undef ANY31CT
 
 // Any 2 planes to 1.

source/row_gcc.cc

@@ -1627,7 +1627,7 @@ void RGBAToUVRow_SSSE3(const uint8* src_rgba0,
 // TODO(fbarchard): Consider shufb to replace pack/unpack
 // TODO(fbarchard): Consider pmulhuw to replace psraw
 // TODO(fbarchard): Consider pmullw to replace psllw and allow different bits.
-#define READYUV422_10 \
+#define READYUV210 \
   "movq       " MEMACCESS([u_buf]) ",%%xmm0                     \n"            \
     MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
     "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
@@ -1637,7 +1637,7 @@ void RGBAToUVRow_SSSE3(const uint8* src_rgba0,
     "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
     "movdqu     " MEMACCESS([y_buf]) ",%%xmm4                   \n"            \
     "psllw      $0x6,%%xmm4                                     \n"            \
-    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]               \n"
+    "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]              \n"
 
 // Read 4 UV from 422, upsample to 8 UV.  With 8 Alpha.
 #define READYUVA422 \
@@ -1892,7 +1892,7 @@ void OMITFP I210ToARGBRow_SSSE3(const uint16* y_buf,
 
     LABELALIGN
     "1:                                        \n"
-    READYUV422_10
+    READYUV210
     YUVTORGB(yuvconstants)
     STOREARGB
     "sub       $0x8,%[width]                   \n"
@@ -1968,7 +1968,7 @@ void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
-    : "memory", "cc", YUVTORGB_REGS  // Does not use r14.
+    : "memory", "cc", YUVTORGB_REGS
       "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
   );
   // clang-format on
@@ -2116,6 +2116,23 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
     "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n"        \
     "lea        " MEMLEA(0x10, [y_buf]) ",%[y_buf]                  \n"
 
+// Read 8 UV from 210 10 bit, upsample to 16 UV
+// TODO(fbarchard): Consider vshufb to replace pack/unpack
+// TODO(fbarchard): Consider vunpcklpd to combine the 2 registers into 1.
+#define READYUV210_AVX2 \
+  "vmovdqu     " MEMACCESS([u_buf]) ",%%xmm0                     \n"           \
+    MEMOPREG(vmovdqu, 0x00, [u_buf], [v_buf], 1, xmm1)                         \
+    "lea        " MEMLEA(0x10, [u_buf]) ",%[u_buf]               \n"           \
+    "vpermq     $0xd8,%%ymm0,%%ymm0                              \n"           \
+    "vpermq     $0xd8,%%ymm1,%%ymm1                              \n"           \
+    "vpunpcklwd %%ymm1,%%ymm0,%%ymm0                             \n"           \
+    "vpsraw     $0x2,%%ymm0,%%ymm0                               \n"           \
+    "vpackuswb  %%ymm0,%%ymm0,%%ymm0                             \n"           \
+    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                             \n"           \
+    "vmovdqu    " MEMACCESS([y_buf]) ",%%ymm4                    \n"           \
+    "vpsllw     $0x6,%%ymm4,%%ymm4                               \n"           \
+    "lea        " MEMLEA(0x20, [y_buf]) ",%[y_buf]               \n"
+
 // Read 8 UV from 422, upsample to 16 UV.  With 16 Alpha.
 #define READYUVA422_AVX2 \
   "vmovq      " MEMACCESS([u_buf]) ",%%xmm0                         \n"        \
@@ -2308,6 +2325,41 @@ void OMITFP I422ToARGBRow_AVX2(const uint8* y_buf,
 }
 #endif  // HAS_I422TOARGBROW_AVX2
 
+#if defined(HAS_I210TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I210ToARGBRow_AVX2(const uint16* y_buf,
+                               const uint16* u_buf,
+                               const uint16* v_buf,
+                               uint8* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+
+    LABELALIGN
+    "1:                                        \n"
+    READYUV210_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I210TOARGBROW_AVX2
+
 #if defined(HAS_I422ALPHATOARGBROW_AVX2)
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y and 16 A producing 16 ARGB.