Change half-float conversion to use 1.0 multiplier

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 278 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) {
-  // Source and dest stride can be zero since we're only copying
-  // one row at a time.
-  int stride = 0;
-  // Maximum value used in |src|.
-  int max_value = (1 << bits_per_channel) - 1;
-  int rows = 1;
-  libyuv::HalfFloatPlane(src, stride, dst, stride, 1.0f / max_value, num, rows);
+namespace {
+// 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
+class HalfFloatMaker_xor : public VideoResourceUpdater::HalfFloatMaker {
+ public:
+  explicit HalfFloatMaker_xor(int bits_per_channel)
+      : bits_per_channel_(bits_per_channel) {}
+  float Offset() const override { return 0.5; }
+  float Multiplier() const override {
+    int max_input_value = (1 << bits_per_channel_) - 1;
+    // 2 << 11 = 2048 would be 1.0 with our exponent.
+    return 2048.0 / max_input_value;
+  }
+  void MakeHalfFloats(const uint16_t* src, size_t num, uint16_t* dst) override {
+    // 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 < num; i++)
+      dst[i] = src[i] | 0x3800;

[fbarchard1, 2016/10/21 22:46:06]

this will be slow for compilers that dont vectorize, or if you allow 'official'
build to use -Os.

[hubbe, 2016/10/21 22:55:04]

Can we get rid of this code and move the decision of how to convert the floats
into half-floats into libyuv? Libyuv could just return the resource multiplier
and offset.

+  }
+
+ private:
+  int bits_per_channel_;
+};
+
+class HalfFloatMaker_libyuv : public VideoResourceUpdater::HalfFloatMaker {
+ public:
+  explicit HalfFloatMaker_libyuv(int bits_per_channel) {
+    int max_value = (1 << bits_per_channel) - 1;
+    // For less than 15 bits, we can give libyuv a multiplier of
+    // 1.0, which is faster on some platforms. If bits is 16 or larger,
+    // a multiplier of 1.0 would cause overflows. However, a multiplier
+    // of 1/max_value would cause subnormal floats, which perform
+    // very poorly on some platforms.
+    if (bits_per_channel <= 15) {
+      libyuv_multiplier_ = 1.0f;
+    } else {
+      // This multiplier makes sure that we avoid subnormal values.
+      libyuv_multiplier_ = 1.0f / 4096.0f;

[fbarchard1, 2016/10/21 22:46:06]

have you tested using 1.0?
values near 65535 will potentially become inf as a halffloat?  but we should
test this to see what really happens?
even with /4096 you can get denormals, which are value but hurt performance
dramatically on intel.

[hubbe, 2016/10/21 22:55:04]

I have not tested the inf case.

The smallest possible normal value is 2^-14, 1/4096 is 4 times larger, so how
can you get subnormal values?

+    }
+    resource_multiplier_ = 1.0f / libyuv_multiplier_ / max_value;
+  }
+  float Offset() const override { return 0.0f; }
+  float Multiplier() const override { return resource_multiplier_; }
+  void MakeHalfFloats(const uint16_t* src, size_t num, uint16_t* dst) override {
+    // Source and dest stride can be zero since we're only copying
+    // one row at a time.
+    int stride = 0;
+    int rows = 1;
+    libyuv::HalfFloatPlane(src, stride, dst, stride, libyuv_multiplier_, num,
+                           rows);
+  }
+
+ private:
+  float libyuv_multiplier_;
+  float resource_multiplier_;
+};
+
+}  // namespace
+
+std::unique_ptr<VideoResourceUpdater::HalfFloatMaker>
+VideoResourceUpdater::NewHalfFloatMaker(int bits_per_channel) {
+  if (bits_per_channel < 11) {
+    return std::unique_ptr<VideoResourceUpdater::HalfFloatMaker>(
+        new HalfFloatMaker_xor(bits_per_channel));
+  } else {
+    return std::unique_ptr<VideoResourceUpdater::HalfFloatMaker>(
+        new HalfFloatMaker_libyuv(bits_per_channel));
+  }
 }
 
 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) {
@@ skipping 157 matching lines @@
                                  plane_resource.resource_id()));
       mailbox.set_color_space(video_frame->ColorSpace());
       external_resources.mailboxes.push_back(mailbox);
       external_resources.release_callbacks.push_back(base::Bind(
           &RecycleResource, AsWeakPtr(), plane_resource.resource_id()));
       external_resources.type = VideoFrameExternalResources::RGBA_RESOURCE;
     }
     return external_resources;
   }
 
+  std::unique_ptr<HalfFloatMaker> half_float_maker;
+  if (resource_provider_->YuvResourceFormat(bits_per_channel) ==
+      LUMINANCE_F16) {
+    half_float_maker = NewHalfFloatMaker(bits_per_channel);
+    external_resources.offset = half_float_maker->Offset();
+    external_resources.multiplier = half_float_maker->Multiplier();
+  }
+
   for (size_t i = 0; i < plane_resources.size(); ++i) {
     PlaneResource& plane_resource = *plane_resources[i];
     // Update each plane's resource id with its content.
     DCHECK_EQ(plane_resource.resource_format(),
               resource_provider_->YuvResourceFormat(bits_per_channel));
 
     if (!plane_resource.Matches(video_frame->unique_id(), i)) {
       // TODO(hubbe): Move all conversion (and upload?) code to media/.
       // We need to transfer data from |video_frame| to the plane resource.
       // TODO(reveman): Can use GpuMemoryBuffers here to improve performance.
@@ skipping 34 matching lines @@
             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));
-            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);
-            }
+            half_float_maker->MakeHalfFloats(src, bytes_per_row / 2, dst);

[fbarchard1, 2016/10/21 22:46:06]

libyuv:HalfFloatPlane has some overhead to detect cpu/alignment.  Suggest
calling it once per plane with width, height and stride.
If the stride == width * 2 it will row coalesce into a single call.

[hubbe, 2016/10/21 22:55:04]

Seems reasonable, but I would prefer to fix that after we move the
offset/multiplier decisions into libyuv.

           } 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 (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:
-        // 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.
     TextureMailbox mailbox(plane_resource.mailbox(), gpu::SyncToken(),
                            resource_provider_->GetResourceTextureTarget(
                                plane_resource.resource_id()));
     mailbox.set_color_space(video_frame->ColorSpace());
     external_resources.mailboxes.push_back(mailbox);
     external_resources.release_callbacks.push_back(base::Bind(
         &RecycleResource, AsWeakPtr(), plane_resource.resource_id()));
   }
@@ skipping 146 matching lines @@
   if (lost_resource) {
     resource_it->clear_refs();
     updater->DeleteResource(resource_it);
     return;
   }
 
   resource_it->remove_ref();
 }
 
 }  // namespace cc