Add test for I010ToNV12

- Add support for negative height to invert
- Fix off by 1 on odd width and height
- Bump version to 1895

Initial I010 is 2 step planar conversion

libyuv_test '--gunit_filter=*010ToNV12_Opt' --gunit_also_run_disabled_tests --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=1000 --libyuv_flags=-1 --libyuv_cpu_info=-1

Skylake Xeon
[       OK ] LibYUVConvertTest.I010ToNV12_Opt (2675 ms)
[       OK ] LibYUVConvertTest.P010ToNV12_Opt (1547 ms)
Pixel 7
[       OK ] LibYUVConvertTest.I010ToNV12_Opt (464 ms)
[       OK ] LibYUVConvertTest.P010ToNV12_Opt (125 ms)

Bug: b/357721018, b/357439226
Change-Id: I2ae59783cf328a6592d0ab80c374ae4dc281daf3
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5778595
Reviewed-by: Wan-Teh Chang <wtc@google.com>

README.chromium

@@ -1,6 +1,6 @@
 Name: libyuv
 URL: https://chromium.googlesource.com/libyuv/libyuv/
-Version: 1894
+Version: 1895
 License: BSD
 License File: LICENSE
 Shipped: yes

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1894
+#define LIBYUV_VERSION 1895
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/convert.cc

@@ -11,6 +11,7 @@
 #include "libyuv/convert.h"
 
 #include "libyuv/basic_types.h"
+#include "libyuv/convert_from.h"  // For I420ToNV12()
 #include "libyuv/cpu_id.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
@@ -658,40 +659,45 @@ int I010ToNV12(const uint16_t* src_y,
                int dst_stride_uv,
                int width,
                int height) {
-  // Allocate temporary buffers for 3 planes in 8 bit.
+  int r;
-  uint8_t* temp_y = (uint8_t*)malloc(width * height);
+  if (height < 0) {
-  uint8_t* temp_u = (uint8_t*)malloc((width / 2) * (height / 2));
+    height = -height;
-  uint8_t* temp_v = (uint8_t*)malloc((width / 2) * (height / 2));
+    int halfheight = (height + 1) >> 1;
-  int temp_stride_y = width;
+    src_y = src_y + (height - 1) * src_stride_y;
-  int temp_stride_u = width / 2;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
-  int temp_stride_v = width / 2;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
-
+    src_stride_y = -src_stride_y;
-  // The first step is to convert 10 bit YUV I010 to 8 bit I420 with 3 planes.
+    src_stride_u = -src_stride_u;
-  if (I010ToI420(src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+    src_stride_v = -src_stride_v;
-                 temp_y, temp_stride_y, temp_u, temp_stride_u, temp_v,
-                 temp_stride_v, width, height) == -1) {
-    free(temp_y);
-    free(temp_u);
-    free(temp_v);
-    return -1;
   }
 
+  // Allocate temporary buffers for 3 planes in 8 bit.
+  {
+    align_buffer_64(temp_y, width * height);
+    align_buffer_64(temp_u, ((width + 1) / 2) * ((height + 1) / 2));
+    align_buffer_64(temp_v, ((width + 1) / 2) * ((height + 1) / 2));
+
+    int temp_stride_y = width;
+    int temp_stride_u = (width + 1) / 2;
+    int temp_stride_v = (width + 1) / 2;
+
+    // The first step is to convert 10 bit YUV I010 to 8 bit I420 with 3 planes.
+    r = I010ToI420(src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+                   temp_y, temp_stride_y, temp_u, temp_stride_u, temp_v,
+                   temp_stride_v, width, height);
+    if (!r) {
        // The second step is to convert 8 bit I420 with 3 planes to 8 bit NV12 with 2 planes.
-  if (I420ToNV12(temp_y, temp_stride_y, temp_u, temp_stride_u, temp_v,
+      r = I420ToNV12(temp_y, temp_stride_y, temp_u, temp_stride_u, temp_v,
                      temp_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
-                 width, height) == -1) {
+                     width, height);
-    free(temp_y);
-    free(temp_u);
-    free(temp_v);
-    return -1;
     }
 
     // Free temporary buffers.
-  free(temp_y);
+    free_aligned_buffer_64(temp_y);
-  free(temp_u);
+    free_aligned_buffer_64(temp_u);
-  free(temp_v);
+    free_aligned_buffer_64(temp_v);
-
+  }
-  return 0;
+  return r;
 }
 
 LIBYUV_API

unit_test/convert_test.cc

@@ -452,6 +452,7 @@ TESTPLANARTOBP(I422, uint8_t, 1, 2, 1, NV21, uint8_t, 1, 2, 2, 8)
 TESTPLANARTOBP(I444, uint8_t, 1, 1, 1, NV12, uint8_t, 1, 2, 2, 8)
 TESTPLANARTOBP(I444, uint8_t, 1, 1, 1, NV21, uint8_t, 1, 2, 2, 8)
 TESTPLANARTOBP(I400, uint8_t, 1, 2, 2, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I010, uint16_t, 2, 2, 2, NV12, uint8_t, 1, 2, 2, 8)
 TESTPLANARTOBP(I010, uint16_t, 2, 2, 2, P010, uint16_t, 2, 2, 2, 10)
 TESTPLANARTOBP(I210, uint16_t, 2, 2, 1, P210, uint16_t, 2, 2, 1, 10)
 TESTPLANARTOBP(I012, uint16_t, 2, 2, 2, P012, uint16_t, 2, 2, 2, 12)