12-bit vp9 video support

cc/resources/video_resource_updater.cc

Index: cc/resources/video_resource_updater.cc
diff --git a/cc/resources/video_resource_updater.cc b/cc/resources/video_resource_updater.cc
index 2584e0b3a83308af23ac83893c244a9badc04111..6d9e0d51321a356c8107215e5373c02f483ba4f9 100644
--- a/cc/resources/video_resource_updater.cc
+++ b/cc/resources/video_resource_updater.cc
@@ -84,6 +84,9 @@ VideoFrameExternalResources::ResourceType ResourceTypeForVideoFrame(
     case media::PIXEL_FORMAT_YUV420P10:
     case media::PIXEL_FORMAT_YUV422P10:
     case media::PIXEL_FORMAT_YUV444P10:
+    case media::PIXEL_FORMAT_YUV420P12:
+    case media::PIXEL_FORMAT_YUV422P12:
+    case media::PIXEL_FORMAT_YUV444P12:
     case media::PIXEL_FORMAT_UNKNOWN:
       break;
   }
@@ -289,6 +292,26 @@ static gfx::Size SoftwarePlaneDimension(media::VideoFrame* input_frame,
   return gfx::Size(plane_width, plane_height);
 }
 
+// Values are expected to be in [0.0..1.0]
+void VideoResourceUpdater::MakeHalfFloats(const uint16_t* src,

[danakj, 2016/09/26 22:10:39]

Maybe MultiplyHalfFloats would be a more accurate/descriptive name?

[hubbe, 2016/09/27 00:03:24]

I think I did a terrible job of explaining what this function does.
Trying again....

+                                          float mult,
+                                          size_t num,
+                                          uint16_t* dst) {
+  // TODO(hubbe): Make AVX and neon versions of this code.
+
+  // This magic constant is 2^-112. Multiplying by this
+  // is the same as subtracting 112 from the exponent, which
+  // is he difference in exponent bias between 32-bit and

[danakj, 2016/09/26 22:10:39]

is the

[hubbe, 2016/09/27 00:03:24]

Done.

+  // 16-bit floats. Once we've done this subtraction, we can
+  // simply extract the low bits of the exponent and the high
+  // bits of the mantissa from our float and we're done.
+  mult *= 1.9259299444e-34f;
+  for (size_t i = 0; i < num; i++) {
+    float value = src[i] * mult;
+    dst[i] = (*(uint32_t*)&value) >> 13;
+  }
+}
+
 VideoFrameExternalResources VideoResourceUpdater::CreateForSoftwarePlanes(
     scoped_refptr<media::VideoFrame> video_frame) {
   TRACE_EVENT0("cc", "VideoResourceUpdater::CreateForSoftwarePlanes");
@@ -327,6 +350,11 @@ VideoFrameExternalResources VideoResourceUpdater::CreateForSoftwarePlanes(
     case media::PIXEL_FORMAT_YUV444P10:
       bits_per_channel = 10;
       break;
+    case media::PIXEL_FORMAT_YUV420P12:
+    case media::PIXEL_FORMAT_YUV422P12:
+    case media::PIXEL_FORMAT_YUV444P12:
+      bits_per_channel = 12;
+      break;
   }
 
   // Only YUV software video frames are supported.
@@ -482,9 +510,6 @@ VideoFrameExternalResources VideoResourceUpdater::CreateForSoftwarePlanes(
       // LUMINANCE_F16 uses half-floats, so we always need a conversion step.
       if (plane_resource.resource_format() == LUMINANCE_F16) {
         needs_conversion = true;
-        // Note that the current method of converting integers to half-floats
-        // stops working if you have more than 10 bits of data.
-        DCHECK_LE(bits_per_channel, 10);
       } else if (bits_per_channel > 8) {
         // If bits_per_channel > 8 and we can't use LUMINANCE_F16, we need to
         // shift the data down and create an 8-bit texture.
@@ -508,13 +533,18 @@ VideoFrameExternalResources VideoResourceUpdater::CreateForSoftwarePlanes(
                 &upload_pixels_[upload_image_stride * row]);
             const uint16_t* src = reinterpret_cast<uint16_t*>(
                 video_frame->data(i) + (video_stride_bytes * row));
-            // Micro-benchmarking indicates that the compiler does
-            // a good enough job of optimizing this loop that trying
-            // to manually operate on one uint64 at a time is not
-            // actually helpful.
-            // Note to future optimizers: Benchmark your optimizations!
-            for (size_t i = 0; i < bytes_per_row / 2; i++)
-              dst[i] = src[i] | 0x3800;
+            if (bits_per_channel <= 10) {
+              // Micro-benchmarking indicates that the compiler does
+              // a good enough job of optimizing this loop that trying
+              // to manually operate on one uint64 at a time is not
+              // actually helpful.
+              // Note to future optimizers: Benchmark your optimizations!
+              for (size_t i = 0; i < bytes_per_row / 2; i++)
+                dst[i] = src[i] | 0x3800;
+            } else {
+              MakeHalfFloats(src, 1.0f / ((1 << bits_per_channel) - 1),
+                             bytes_per_row / 2, dst);
+            }
           } else if (shift != 0) {
             // We have more-than-8-bit input which we need to shift
             // down to fit it into an 8-bit texture.
@@ -540,7 +570,8 @@ VideoFrameExternalResources VideoResourceUpdater::CreateForSoftwarePlanes(
       plane_resource.SetUniqueId(video_frame->unique_id(), i);
     }
 
-    if (plane_resource.resource_format() == LUMINANCE_F16) {
+    if (plane_resource.resource_format() == LUMINANCE_F16 &&
+        bits_per_channel <= 10) {
       // By OR-ing with 0x3800, 10-bit numbers become half-floats in the
       // range [0.5..1) and 9-bit numbers get the range [0.5..0.75).
       //