12-bit vp9 video support

cc/resources/video_resource_updater.cc

 // Copyright 2013 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "cc/resources/video_resource_updater.h"
 
 #include <stddef.h>
 #include <stdint.h>
 
 #include <algorithm>
@@ skipping 66 matching lines @@
     case media::PIXEL_FORMAT_RGB24:
     case media::PIXEL_FORMAT_RGB32:
     case media::PIXEL_FORMAT_MJPEG:
     case media::PIXEL_FORMAT_MT21:
     case media::PIXEL_FORMAT_YUV420P9:
     case media::PIXEL_FORMAT_YUV422P9:
     case media::PIXEL_FORMAT_YUV444P9:
     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;
   }
   return VideoFrameExternalResources::NONE;
 }
 
 class SyncTokenClientImpl : public media::VideoFrame::SyncTokenClient {
  public:
   SyncTokenClientImpl(gpu::gles2::GLES2Interface* gl,
                       const gpu::SyncToken& sync_token)
@@ skipping 185 matching lines @@
   if (software_compositor)
     return coded_size;
 
   int plane_width = media::VideoFrame::Columns(
       plane_index, input_frame->format(), coded_size.width());
   int plane_height = media::VideoFrame::Rows(plane_index, input_frame->format(),
                                              coded_size.height());
   return gfx::Size(plane_width, plane_height);
 }
 
+void VideoResourceUpdater::MakeHalfFloats(const uint16_t* src,
+                                          int bits_per_channel,
+                                          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 the difference in exponent bias between 32-bit and
+  // 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.
+  float mult = 1.9259299444e-34f / ((1 << bits_per_channel) - 1);
+  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");
   const media::VideoPixelFormat input_frame_format = video_frame->format();
 
   // TODO(hubbe): Make this a video frame method.
   int bits_per_channel = 0;
   switch (input_frame_format) {
     case media::PIXEL_FORMAT_UNKNOWN:
       NOTREACHED();
@@ skipping 18 matching lines @@
     case media::PIXEL_FORMAT_YUV420P9:
     case media::PIXEL_FORMAT_YUV422P9:
     case media::PIXEL_FORMAT_YUV444P9:
       bits_per_channel = 9;
       break;
     case media::PIXEL_FORMAT_YUV420P10:
     case media::PIXEL_FORMAT_YUV422P10:
     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.
   if (!media::IsYuvPlanar(input_frame_format)) {
     NOTREACHED() << media::VideoPixelFormatToString(input_frame_format);
     return VideoFrameExternalResources();
   }
 
   const bool software_compositor = context_provider_ == NULL;
 
@@ skipping 135 matching lines @@
       // Assuming that GL_UNPACK_ALIGNMENT has not changed from default.
       size_t upload_image_stride =
           MathUtil::CheckedRoundUp<size_t>(bytes_per_row, 4u);
 
       bool needs_conversion = false;
       int shift = 0;
 
       // 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.
         needs_conversion = true;
         shift = bits_per_channel - 8;
       }
       const uint8_t* pixels;
       if (static_cast<int>(upload_image_stride) == video_stride_bytes &&
           !needs_conversion) {
         pixels = video_frame->data(i);
       } else {
         // Avoid malloc for each frame/plane if possible.
         size_t needed_size =
             upload_image_stride * resource_size_pixels.height();
         if (upload_pixels_.size() < needed_size)
           upload_pixels_.resize(needed_size);
 
         for (int row = 0; row < resource_size_pixels.height(); ++row) {
           if (plane_resource.resource_format() == LUMINANCE_F16) {
             uint16_t* dst = reinterpret_cast<uint16_t*>(
                 &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, bits_per_channel, 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.
             uint8_t* dst = &upload_pixels_[upload_image_stride * row];
             const uint16_t* src = reinterpret_cast<uint16_t*>(
                 video_frame->data(i) + (video_stride_bytes * row));
             for (size_t i = 0; i < bytes_per_row; i++)
               dst[i] = src[i] >> shift;
           } else {
             // Input and output are the same size and format, but
             // differ in stride, copy one row at a time.
             uint8_t* dst = &upload_pixels_[upload_image_stride * row];
             const uint8_t* src =
                 video_frame->data(i) + (video_stride_bytes * row);
             memcpy(dst, src, bytes_per_row);
           }
         }
         pixels = &upload_pixels_[0];
       }
 
       resource_provider_->CopyToResource(plane_resource.resource_id(), pixels,
                                          resource_size_pixels);
       plane_resource.SetUniqueId(video_frame->unique_id(), i);
     }
 
-    if (plane_resource.resource_format() == LUMINANCE_F16) {
+    // If the input data was 9 or 10 bit, and we output to half-floats,
+    // then we used the OR path above, which means that we need to
+    // adjust the resource offset and multiplier accordingly. If the
+    // input data uses more than 10 bits, it will already be normalized
+    // to 0.0..1.0, so there is no need to do anything.
+    if (plane_resource.resource_format() == LUMINANCE_F16 &&

[danakj, 2016/09/27 20:49:11]

One more thought, we're doing this if thing twice essentially, which means they
need to not diverge. Above we have also checked the same things. Would it be
more clear if we set a bool use_multiplier_for_half_float_output = false up
above or something, and set it to true at the place where we do the OR path?

[hubbe, 2016/09/27 22:27:36]

I don't really want to set a bool in every row, but how about I put it above the
loops where we do some similar if statements?

[danakj, 2016/09/27 22:37:51]

That's cool too if you agree it's more clear

+        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).
       //
       // Half-floats are evaluated as:
       // float value = pow(2.0, exponent - 25) * (0x400 + fraction);
       //
       // In our case the exponent is 14 (since we or with 0x3800) and
       // pow(2.0, 14-25) * 0x400 evaluates to 0.5 (our offset) and
       // pow(2.0, 14-25) * fraction is [0..0.49951171875] for 10-bit and
       // [0..0.24951171875] for 9-bit.
       //
       // (https://en.wikipedia.org/wiki/Half-precision_floating-point_format)
       //
-      // PLEASE NOTE: This doesn't work if bits_per_channel is > 10.
       // PLEASE NOTE: All planes are assumed to use the same multiplier/offset.
       external_resources.offset = 0.5f;
       // Max value from input data.
       int max_input_value = (1 << bits_per_channel) - 1;
       // 2 << 11 = 2048 would be 1.0 with our exponent.
       external_resources.multiplier = 2048.0 / max_input_value;
     }
 
     // VideoResourceUpdater shares a context with the compositor so a
     // sync token is not required.
@@ skipping 154 matching lines @@
   if (lost_resource) {
     resource_it->clear_refs();
     updater->DeleteResource(resource_it);
     return;
   }
 
   resource_it->remove_ref();
 }
 
 }  // namespace cc