SSE2 MM21->YUY2 conversion

Add SSE2 optimization for MM21ToYUY2 conversion.

Bug: b/238137982
Change-Id: I189f712514308322f651b082b496bce9c015c4ee
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/3832525
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Justin Green <greenjustin@google.com>

5 files changed, 51 insertions, 6 deletions

diff --git a/source/convert.cc b/source/convert.cc
index 37066721..a740d5ca 100644
--- a/source/convert.cc
+++ b/source/convert.cc
@@ -262,10 +262,10 @@ int I210ToI420(const uint16_t* src_y,
                       height);
     ScalePlaneVertical_16To8(height, uv_width, uv_height, src_stride_u,
                              dst_stride_u, src_u, dst_u, 0, 32768, dy,
-                             /*wpp=*/1, scale, kFilterBilinear);
+                             /*bpp=*/1, scale, kFilterBilinear);
     ScalePlaneVertical_16To8(height, uv_width, uv_height, src_stride_v,
                              dst_stride_v, src_v, dst_v, 0, 32768, dy,
-                             /*wpp=*/1, scale, kFilterBilinear);
+                             /*bpp=*/1, scale, kFilterBilinear);
   }
   return 0;
 }
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index fae8630e..1de71dbb 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -915,7 +915,7 @@ int NV21ToNV12(const uint8_t* src_y,
 // tile width is 16 and assumed.
 // tile_height is 16 or 32 for MM21.
 // src_stride_y is bytes per row of source ignoring tiling. e.g. 640
-// TODO(fbarchard): More detile row functions.
+// TODO: More detile row functions.
 
 LIBYUV_API
 void DetilePlane(const uint8_t* src_y,
@@ -1074,6 +1074,15 @@ void DetileToYUY2(const uint8_t* src_y,
   }
 #endif
 
+#if defined(HAS_DETILETOYUY2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    DetileToYUY2 = DetileToYUY2_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      DetileToYUY2 = DetileToYUY2_SSE2;
+    }
+  }
+#endif
+
   // Detile plane
   for (y = 0; y < height; ++y) {
     DetileToYUY2(src_y, src_y_tile_stride, src_uv, src_uv_tile_stride,
@@ -1081,9 +1090,8 @@ void DetileToYUY2(const uint8_t* src_y,
     dst_yuy2 += dst_stride_yuy2;
     src_y += 16;
 
-    if (y & 0x1) {
+    if (y & 0x1)
       src_uv += 16;
-    }
 
     // Advance to next row of tiles.
     if ((y & (tile_height - 1)) == (tile_height - 1)) {
diff --git a/source/row_any.cc b/source/row_any.cc
index 5270e86c..bd46ba1b 100644
--- a/source/row_any.cc
+++ b/source/row_any.cc
@@ -2272,6 +2272,10 @@ ANYDETILESPLITUV(DetileSplitUVRow_Any_SSSE3, DetileSplitUVRow_SSSE3, 15)
 ANYDETILEMERGE(DetileToYUY2_Any_NEON, DetileToYUY2_NEON, 15)
 #endif
 
+#ifdef HAS_DETILETOYUY2_SSE2
+ANYDETILEMERGE(DetileToYUY2_Any_SSE2, DetileToYUY2_SSE2, 15)
+#endif
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/source/row_gcc.cc b/source/row_gcc.cc
index 88766785..8d0f477c 100644
--- a/source/row_gcc.cc
+++ b/source/row_gcc.cc
@@ -4968,6 +4968,39 @@ void DetileRow_SSE2(const uint8_t* src,
 }
 #endif  // HAS_DETILEROW_SSE2
 
+#ifdef HAS_DETILETOYUY2_SSE2
+// Read 16 Y, 8 UV, and write 8 YUYV.
+void DetileToYUY2_SSE2(const uint8_t* src_y,
+                       ptrdiff_t src_y_tile_stride,
+                       const uint8_t* src_uv,
+                       ptrdiff_t src_uv_tile_stride,
+                       uint8_t* dst_yuy2,
+                       int width) {
+  asm volatile(
+      "1:                                       \n"
+      "movdqu     (%0),%%xmm0                   \n" // Load 16 Y
+      "sub        $0x10,%3                      \n"
+      "lea        (%0,%4),%0                    \n"
+      "movdqu     (%1),%%xmm1                   \n" // Load 8 UV
+      "lea        (%1,%5),%1                    \n"
+      "movdqu     %%xmm0,%%xmm2                 \n"
+      "punpcklbw  %%xmm1,%%xmm0                 \n"
+      "punpckhbw  %%xmm1,%%xmm2                 \n"
+      "movdqu     %%xmm0,(%2)                   \n"
+      "movdqu     %%xmm2,0x10(%2)               \n"
+      "lea        0x20(%2),%2                   \n"
+      "jg         1b                            \n"
+      : "+r"(src_y),                            // %0
+        "+r"(src_uv),                           // %1
+        "+r"(dst_yuy2),                         // %2
+        "+r"(width)                             // %3
+      : "r"(src_y_tile_stride),                 // %4
+        "r"(src_uv_tile_stride)                 // %5
+      : "cc", "memory", "xmm0", "xmm1", "xmm2"  // Clobber list
+  );
+}
+#endif
+
 #ifdef HAS_DETILESPLITUVROW_SSSE3
 // TODO(greenjustin): Look into generating these constants instead of loading
 // them since this can cause branch mispredicts for fPIC code on 32-bit
diff --git a/source/scale.cc b/source/scale.cc
index 4980f42d..e1335f1e 100644
--- a/source/scale.cc
+++ b/source/scale.cc
@@ -2039,7 +2039,7 @@ void ScalePlane_16(const uint16_t* src,
     }
     // Arbitrary scale vertically, but unscaled horizontally.
     ScalePlaneVertical_16(src_height, dst_width, dst_height, src_stride,
-                          dst_stride, src, dst, 0, y, dy, /*wpp=*/1, filtering);
+                          dst_stride, src, dst, 0, y, dy, /*bpp=*/1, filtering);
     return;
   }
   if (dst_width <= Abs(src_width) && dst_height <= src_height) {