AVX2 port of H010ToAR30_AVX2

Was SSSE3 H010ToAR30_Opt (635 ms)
Now AVX2  H010ToAR30_Opt (448 ms)

Bug: libyuv:751
Test:  LibYUVConvertTest.H010ToAR30_Opt
Change-Id: I17b1a0e3268c4a9836e09683dd3377fb1ce60932
Reviewed-on: https://chromium-review.googlesource.com/889906
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Miguel Casas <mcasas@chromium.org>

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1693
+Version: 1694
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -272,6 +272,7 @@ extern "C" {
 #define HAS_CONVERT16TO8ROW_AVX2
 #define HAS_CONVERT8TO16ROW_AVX2
 #define HAS_I210TOARGBROW_AVX2
+#define HAS_I210TOAR30ROW_AVX2
 #define HAS_I422TOAR30ROW_AVX2
 #define HAS_MERGEUVROW_16_AVX2
 #define HAS_MULTIPLYROW_16_AVX2
@@ -1918,6 +1919,12 @@ void I210ToARGBRow_AVX2(const uint16_t* src_y,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
+void I210ToAR30Row_AVX2(const uint16_t* src_y,
+                        const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
 void I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
                               const uint8_t* u_buf,
                               const uint8_t* v_buf,
@@ -2092,6 +2099,12 @@ void I210ToARGBRow_Any_AVX2(const uint16_t* src_y,
                             uint8_t* dst_argb,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void I210ToAR30Row_Any_AVX2(const uint16_t* src_y,
+                            const uint16_t* src_u,
+                            const uint16_t* src_v,
+                            uint8_t* dst_ar30,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I422AlphaToARGBRow_Any_SSSE3(const uint8_t* y_buf,
                                   const uint8_t* u_buf,
                                   const uint8_t* v_buf,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1693
+#define LIBYUV_VERSION 1694
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/row_any.cc

@@ -230,6 +230,9 @@ ANY31CT(I210ToARGBRow_Any_SSSE3, I210ToARGBRow_SSSE3, 1, 0, uint16_t, 2, 4, 7)
 #ifdef HAS_I210TOARGBROW_AVX2
 ANY31CT(I210ToARGBRow_Any_AVX2, I210ToARGBRow_AVX2, 1, 0, uint16_t, 2, 4, 15)
 #endif
+#ifdef HAS_I210TOAR30ROW_AVX2
+ANY31CT(I210ToAR30Row_Any_AVX2, I210ToAR30Row_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
 #undef ANY31CT
 
 // Any 2 planes to 1.

source/row_gcc.cc

@@ -2501,6 +2501,46 @@ void OMITFP I210ToARGBRow_AVX2(const uint16_t* y_buf,
 }
 #endif  // HAS_I210TOARGBROW_AVX2
 
+#if defined(HAS_I210TOAR30ROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP I210ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"  // AR30 constants
+    "vpsrlw    $14,%%ymm5,%%ymm5               \n"
+    "vpsllw    $4,%%ymm5,%%ymm5                \n"  // 2 alpha bits
+    "vpxor     %%ymm6,%%ymm6,%%ymm6            \n"  // 0 for min
+    "vpcmpeqb  %%ymm7,%%ymm7,%%ymm7            \n"  // 1023 for max
+    "vpsrlw    $6,%%ymm7,%%ymm7                \n"
+
+    LABELALIGN
+    "1:                                        \n"
+    READYUV210_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_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_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I210TOAR30ROW_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.

unit_test/convert_test.cc

@@ -2083,6 +2083,7 @@ TEST_F(LibYUVConvertTest, ARGBToAR30Row_Opt) {
 
 TESTPLANAR16TOB(I010, 2, 2, ARGB, 4, 4, 1, 2)
 TESTPLANAR16TOB(I010, 2, 2, ABGR, 4, 4, 1, 2)
+TESTPLANAR16TOB(I010, 2, 2, AR30, 4, 4, 1, 2)
 TESTPLANAR16TOB(H010, 2, 2, ARGB, 4, 4, 1, 2)
 TESTPLANAR16TOB(H010, 2, 2, ABGR, 4, 4, 1, 2)
 TESTPLANAR16TOB(H010, 2, 2, AR30, 4, 4, 1, 2)