// libjingle // Copyright 2010 Google Inc. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are met: // // 1. Redistributions of source code must retain the above copyright notice, // this list of conditions and the following disclaimer. // 2. Redistributions in binary form must reproduce the above copyright notice, // this list of conditions and the following disclaimer in the documentation // and/or other materials provided with the distribution. // 3. The name of the author may not be used to endorse or promote products // derived from this software without specific prior written permission. // // THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED // WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF // MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO // EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; // OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, // WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR // OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF // ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #include "talk/media/base/videoadapter.h" #include // For INT_MAX #include "talk/base/logging.h" #include "talk/base/timeutils.h" #include "talk/media/base/constants.h" #include "talk/media/base/videoframe.h" namespace cricket { // TODO(fbarchard): Make downgrades settable static const int kMaxCpuDowngrades = 2; // Downgrade at most 2 times for CPU. // The number of milliseconds of data to require before acting on cpu sampling // information. static const size_t kCpuLoadMinSampleTime = 5000; // The amount of weight to give to each new cpu load sample. The lower the // value, the slower we'll adapt to changing cpu conditions. static const float kCpuLoadWeightCoefficient = 0.4f; // The seed value for the cpu load moving average. static const float kCpuLoadInitialAverage = 0.5f; // TODO(fbarchard): Consider making scale factor table settable, to allow // application to select quality vs performance tradeoff. // TODO(fbarchard): Add framerate scaling to tables for 1/2 framerate. // List of scale factors that adapter will scale by. #if defined(IOS) || defined(ANDROID) // Mobile needs 1/4 scale for VGA (640 x 360) to QQVGA (160 x 90) // or 1/4 scale for HVGA (480 x 270) to QQHVGA (120 x 67) static const int kMinNumPixels = 120 * 67; static float kScaleFactors[] = { 1.f / 1.f, // Full size. 3.f / 4.f, // 3/4 scale. 1.f / 2.f, // 1/2 scale. 3.f / 8.f, // 3/8 scale. 1.f / 4.f, // 1/4 scale. }; #else // Desktop needs 1/8 scale for HD (1280 x 720) to QQVGA (160 x 90) static const int kMinNumPixels = 160 * 100; static float kScaleFactors[] = { 1.f / 1.f, // Full size. 3.f / 4.f, // 3/4 scale. 1.f / 2.f, // 1/2 scale. 3.f / 8.f, // 3/8 scale. 1.f / 4.f, // 1/4 scale. 3.f / 16.f, // 3/16 scale. 1.f / 8.f // 1/8 scale. }; #endif static const int kNumScaleFactors = ARRAY_SIZE(kScaleFactors); // Find the scale factor that, when applied to width and height, is closest // to num_pixels. float VideoAdapter::FindClosestScale(int width, int height, int target_num_pixels) { if (!target_num_pixels) { return 0.f; } int best_distance = INT_MAX; int best_index = kNumScaleFactors - 1; // Default to max scale. for (int i = 0; i < kNumScaleFactors; ++i) { int test_num_pixels = static_cast(width * kScaleFactors[i] * height * kScaleFactors[i]); int diff = test_num_pixels - target_num_pixels; if (diff < 0) { diff = -diff; } if (diff < best_distance) { best_distance = diff; best_index = i; if (best_distance == 0) { // Found exact match. break; } } } return kScaleFactors[best_index]; } // Finds the scale factor that, when applied to width and height, produces // fewer than num_pixels. float VideoAdapter::FindLowerScale(int width, int height, int target_num_pixels) { if (!target_num_pixels) { return 0.f; } int best_distance = INT_MAX; int best_index = kNumScaleFactors - 1; // Default to max scale. for (int i = 0; i < kNumScaleFactors; ++i) { int test_num_pixels = static_cast(width * kScaleFactors[i] * height * kScaleFactors[i]); int diff = target_num_pixels - test_num_pixels; if (diff >= 0 && diff < best_distance) { best_distance = diff; best_index = i; if (best_distance == 0) { // Found exact match. break; } } } return kScaleFactors[best_index]; } // There are several frame sizes used by Adapter. This explains them // input_format - set once by server to frame size expected from the camera. // output_format - size that output would like to be. Includes framerate. // output_num_pixels - size that output should be constrained to. Used to // compute output_format from in_frame. // in_frame - actual camera captured frame size, which is typically the same // as input_format. This can also be rotated or cropped for aspect ratio. // out_frame - actual frame output by adapter. Should be a direct scale of // in_frame maintaining rotation and aspect ratio. // OnOutputFormatRequest - server requests you send this resolution based on // view requests. // OnEncoderResolutionRequest - encoder requests you send this resolution based // on bandwidth // OnCpuLoadUpdated - cpu monitor requests you send this resolution based on // cpu load. /////////////////////////////////////////////////////////////////////// // Implementation of VideoAdapter VideoAdapter::VideoAdapter() : output_num_pixels_(INT_MAX), black_output_(false), is_black_(false), interval_next_frame_(0) { } VideoAdapter::~VideoAdapter() { } void VideoAdapter::SetInputFormat(const VideoFrame& in_frame) { talk_base::CritScope cs(&critical_section_); input_format_.width = static_cast(in_frame.GetWidth()); input_format_.height = static_cast(in_frame.GetHeight()); } void VideoAdapter::SetInputFormat(const VideoFormat& format) { talk_base::CritScope cs(&critical_section_); input_format_ = format; output_format_.interval = talk_base::_max( output_format_.interval, input_format_.interval); } void VideoAdapter::SetOutputFormat(const VideoFormat& format) { talk_base::CritScope cs(&critical_section_); output_format_ = format; output_num_pixels_ = output_format_.width * output_format_.height; output_format_.interval = talk_base::_max( output_format_.interval, input_format_.interval); } const VideoFormat& VideoAdapter::input_format() { talk_base::CritScope cs(&critical_section_); return input_format_; } const VideoFormat& VideoAdapter::output_format() { talk_base::CritScope cs(&critical_section_); return output_format_; } void VideoAdapter::SetBlackOutput(bool black) { talk_base::CritScope cs(&critical_section_); black_output_ = black; } // Constrain output resolution to this many pixels overall void VideoAdapter::SetOutputNumPixels(int num_pixels) { output_num_pixels_ = num_pixels; } int VideoAdapter::GetOutputNumPixels() const { return output_num_pixels_; } // TODO(fbarchard): Add AdaptFrameRate function that only drops frames but // not resolution. bool VideoAdapter::AdaptFrame(const VideoFrame* in_frame, const VideoFrame** out_frame) { talk_base::CritScope cs(&critical_section_); if (!in_frame || !out_frame) { return false; } // Update input to actual frame dimensions. SetInputFormat(*in_frame); // Drop the input frame if necessary. bool should_drop = false; if (!output_num_pixels_) { // Drop all frames as the output format is 0x0. should_drop = true; } else { // Drop some frames based on input fps and output fps. // Normally output fps is less than input fps. // TODO(fbarchard): Consider adjusting interval to reflect the adjusted // interval between frames after dropping some frames. interval_next_frame_ += input_format_.interval; if (output_format_.interval > 0) { if (interval_next_frame_ >= output_format_.interval) { interval_next_frame_ %= output_format_.interval; } else { should_drop = true; } } } if (should_drop) { *out_frame = NULL; return true; } if (output_num_pixels_) { float scale = VideoAdapter::FindClosestScale( static_cast(in_frame->GetWidth()), static_cast(in_frame->GetHeight()), output_num_pixels_); output_format_.width = static_cast(in_frame->GetWidth() * scale + .5f); output_format_.height = static_cast(in_frame->GetHeight() * scale + .5f); } if (!StretchToOutputFrame(in_frame)) { return false; } *out_frame = output_frame_.get(); return true; } bool VideoAdapter::StretchToOutputFrame(const VideoFrame* in_frame) { int output_width = output_format_.width; int output_height = output_format_.height; // Create and stretch the output frame if it has not been created yet or its // size is not same as the expected. bool stretched = false; if (!output_frame_ || output_frame_->GetWidth() != static_cast(output_width) || output_frame_->GetHeight() != static_cast(output_height)) { output_frame_.reset( in_frame->Stretch(output_width, output_height, true, true)); if (!output_frame_) { LOG(LS_WARNING) << "Adapter failed to stretch frame to " << output_width << "x" << output_height; return false; } stretched = true; is_black_ = false; } if (!black_output_) { if (!stretched) { // The output frame does not need to be blacken and has not been stretched // from the input frame yet, stretch the input frame. This is the most // common case. in_frame->StretchToFrame(output_frame_.get(), true, true); } is_black_ = false; } else { if (!is_black_) { output_frame_->SetToBlack(); is_black_ = true; } output_frame_->SetElapsedTime(in_frame->GetElapsedTime()); output_frame_->SetTimeStamp(in_frame->GetTimeStamp()); } return true; } /////////////////////////////////////////////////////////////////////// // Implementation of CoordinatedVideoAdapter CoordinatedVideoAdapter::CoordinatedVideoAdapter() : cpu_adaptation_(false), cpu_smoothing_(false), gd_adaptation_(true), view_adaptation_(true), view_switch_(false), cpu_downgrade_count_(0), cpu_adapt_wait_time_(0), high_system_threshold_(kHighSystemCpuThreshold), low_system_threshold_(kLowSystemCpuThreshold), process_threshold_(kProcessCpuThreshold), view_desired_num_pixels_(INT_MAX), view_desired_interval_(0), encoder_desired_num_pixels_(INT_MAX), cpu_desired_num_pixels_(INT_MAX), adapt_reason_(0), system_load_average_(kCpuLoadInitialAverage) { } // Helper function to UPGRADE or DOWNGRADE a number of pixels void CoordinatedVideoAdapter::StepPixelCount( CoordinatedVideoAdapter::AdaptRequest request, int* num_pixels) { switch (request) { case CoordinatedVideoAdapter::DOWNGRADE: *num_pixels /= 2; break; case CoordinatedVideoAdapter::UPGRADE: *num_pixels *= 2; break; default: // No change in pixel count break; } return; } // Find the adaptation request of the cpu based on the load. Return UPGRADE if // the load is low, DOWNGRADE if the load is high, and KEEP otherwise. CoordinatedVideoAdapter::AdaptRequest CoordinatedVideoAdapter::FindCpuRequest( int current_cpus, int max_cpus, float process_load, float system_load) { // Downgrade if system is high and plugin is at least more than midrange. if (system_load >= high_system_threshold_ * max_cpus && process_load >= process_threshold_ * current_cpus) { return CoordinatedVideoAdapter::DOWNGRADE; // Upgrade if system is low. } else if (system_load < low_system_threshold_ * max_cpus) { return CoordinatedVideoAdapter::UPGRADE; } return CoordinatedVideoAdapter::KEEP; } // A remote view request for a new resolution. void CoordinatedVideoAdapter::OnOutputFormatRequest(const VideoFormat& format) { talk_base::CritScope cs(&request_critical_section_); if (!view_adaptation_) { return; } // Set output for initial aspect ratio in mediachannel unittests. int old_num_pixels = GetOutputNumPixels(); SetOutputFormat(format); SetOutputNumPixels(old_num_pixels); view_desired_num_pixels_ = format.width * format.height; view_desired_interval_ = format.interval; int new_width, new_height; bool changed = AdaptToMinimumFormat(&new_width, &new_height); LOG(LS_INFO) << "VAdapt View Request: " << format.width << "x" << format.height << " Pixels: " << view_desired_num_pixels_ << " Changed: " << (changed ? "true" : "false") << " To: " << new_width << "x" << new_height; } // A Bandwidth GD request for new resolution void CoordinatedVideoAdapter::OnEncoderResolutionRequest( int width, int height, AdaptRequest request) { talk_base::CritScope cs(&request_critical_section_); if (!gd_adaptation_) { return; } int old_encoder_desired_num_pixels = encoder_desired_num_pixels_; if (KEEP != request) { int new_encoder_desired_num_pixels = width * height; int old_num_pixels = GetOutputNumPixels(); if (new_encoder_desired_num_pixels != old_num_pixels) { LOG(LS_VERBOSE) << "VAdapt GD resolution stale. Ignored"; } else { // Update the encoder desired format based on the request. encoder_desired_num_pixels_ = new_encoder_desired_num_pixels; StepPixelCount(request, &encoder_desired_num_pixels_); } } int new_width, new_height; bool changed = AdaptToMinimumFormat(&new_width, &new_height); // Ignore up or keep if no change. if (DOWNGRADE != request && view_switch_ && !changed) { encoder_desired_num_pixels_ = old_encoder_desired_num_pixels; LOG(LS_VERBOSE) << "VAdapt ignoring GD request."; } LOG(LS_INFO) << "VAdapt GD Request: " << (DOWNGRADE == request ? "down" : (UPGRADE == request ? "up" : "keep")) << " From: " << width << "x" << height << " Pixels: " << encoder_desired_num_pixels_ << " Changed: " << (changed ? "true" : "false") << " To: " << new_width << "x" << new_height; } // A CPU request for new resolution void CoordinatedVideoAdapter::OnCpuLoadUpdated( int current_cpus, int max_cpus, float process_load, float system_load) { talk_base::CritScope cs(&request_critical_section_); if (!cpu_adaptation_) { return; } // Update the moving average of system load. Even if we aren't smoothing, // we'll still calculate this information, in case smoothing is later enabled. system_load_average_ = kCpuLoadWeightCoefficient * system_load + (1.0f - kCpuLoadWeightCoefficient) * system_load_average_; if (cpu_smoothing_) { system_load = system_load_average_; } // If we haven't started taking samples yet, wait until we have at least // the correct number of samples per the wait time. if (cpu_adapt_wait_time_ == 0) { cpu_adapt_wait_time_ = talk_base::TimeAfter(kCpuLoadMinSampleTime); } AdaptRequest request = FindCpuRequest(current_cpus, max_cpus, process_load, system_load); // Make sure we're not adapting too quickly. if (request != KEEP) { if (talk_base::TimeIsLater(talk_base::Time(), cpu_adapt_wait_time_)) { LOG(LS_VERBOSE) << "VAdapt CPU load high/low but do not adapt until " << talk_base::TimeUntil(cpu_adapt_wait_time_) << " ms"; request = KEEP; } } // Update how many times we have downgraded due to the cpu load. switch (request) { case DOWNGRADE: // Ignore downgrades if we have downgraded the maximum times. if (cpu_downgrade_count_ < kMaxCpuDowngrades) { ++cpu_downgrade_count_; } else { LOG(LS_VERBOSE) << "VAdapt CPU load high but do not downgrade " "because maximum downgrades reached"; SignalCpuAdaptationUnable(); } break; case UPGRADE: if (cpu_downgrade_count_ > 0) { bool is_min = IsMinimumFormat(cpu_desired_num_pixels_); if (is_min) { --cpu_downgrade_count_; } else { LOG(LS_VERBOSE) << "VAdapt CPU load low but do not upgrade " "because cpu is not limiting resolution"; } } else { LOG(LS_VERBOSE) << "VAdapt CPU load low but do not upgrade " "because minimum downgrades reached"; } break; case KEEP: default: break; } if (KEEP != request) { // TODO(fbarchard): compute stepping up/down from OutputNumPixels but // clamp to inputpixels / 4 (2 steps) cpu_desired_num_pixels_ = cpu_downgrade_count_ == 0 ? INT_MAX : static_cast(input_format().width * input_format().height >> cpu_downgrade_count_); } int new_width, new_height; bool changed = AdaptToMinimumFormat(&new_width, &new_height); LOG(LS_INFO) << "VAdapt CPU Request: " << (DOWNGRADE == request ? "down" : (UPGRADE == request ? "up" : "keep")) << " Process: " << process_load << " System: " << system_load << " Steps: " << cpu_downgrade_count_ << " Changed: " << (changed ? "true" : "false") << " To: " << new_width << "x" << new_height; } // Called by cpu adapter on up requests. bool CoordinatedVideoAdapter::IsMinimumFormat(int pixels) { // Find closest scale factor that matches input resolution to min_num_pixels // and set that for output resolution. This is not needed for VideoAdapter, // but provides feedback to unittests and users on expected resolution. // Actual resolution is based on input frame. VideoFormat new_output = output_format(); VideoFormat input = input_format(); if (input_format().IsSize0x0()) { input = new_output; } float scale = 1.0f; if (!input.IsSize0x0()) { scale = FindClosestScale(input.width, input.height, pixels); } new_output.width = static_cast(input.width * scale + .5f); new_output.height = static_cast(input.height * scale + .5f); int new_pixels = new_output.width * new_output.height; int num_pixels = GetOutputNumPixels(); return new_pixels <= num_pixels; } // Called by all coordinators when there is a change. bool CoordinatedVideoAdapter::AdaptToMinimumFormat(int* new_width, int* new_height) { VideoFormat new_output = output_format(); VideoFormat input = input_format(); if (input_format().IsSize0x0()) { input = new_output; } int old_num_pixels = GetOutputNumPixels(); // Find resolution that respects ViewRequest or less pixels. int view_desired_num_pixels = view_desired_num_pixels_; int min_num_pixels = view_desired_num_pixels_; if (!input.IsSize0x0()) { float scale = FindLowerScale(input.width, input.height, min_num_pixels); min_num_pixels = view_desired_num_pixels = static_cast(input.width * input.height * scale * scale + .5f); } // Reduce resolution further, if necessary, based on encoder bandwidth (GD). if (encoder_desired_num_pixels_ && (encoder_desired_num_pixels_ < min_num_pixels)) { min_num_pixels = encoder_desired_num_pixels_; } // Reduce resolution further, if necessary, based on CPU. if (cpu_adaptation_ && cpu_desired_num_pixels_ && (cpu_desired_num_pixels_ < min_num_pixels)) { min_num_pixels = cpu_desired_num_pixels_; } // Determine which factors are keeping adapter resolution low. // Caveat: Does not consider framerate. adapt_reason_ = static_cast(0); if (view_desired_num_pixels == min_num_pixels) { adapt_reason_ |= ADAPTREASON_VIEW; } if (encoder_desired_num_pixels_ == min_num_pixels) { adapt_reason_ |= ADAPTREASON_BANDWIDTH; } if (cpu_desired_num_pixels_ == min_num_pixels) { adapt_reason_ |= ADAPTREASON_CPU; } // Prevent going below QQVGA. if (min_num_pixels > 0 && min_num_pixels < kMinNumPixels) { min_num_pixels = kMinNumPixels; } SetOutputNumPixels(min_num_pixels); // Find closest scale factor that matches input resolution to min_num_pixels // and set that for output resolution. This is not needed for VideoAdapter, // but provides feedback to unittests and users on expected resolution. // Actual resolution is based on input frame. float scale = 1.0f; if (!input.IsSize0x0()) { scale = FindClosestScale(input.width, input.height, min_num_pixels); } if (scale == 1.0f) { adapt_reason_ = 0; } *new_width = new_output.width = static_cast(input.width * scale + .5f); *new_height = new_output.height = static_cast(input.height * scale + .5f); new_output.interval = view_desired_interval_; SetOutputFormat(new_output); int new_num_pixels = GetOutputNumPixels(); bool changed = new_num_pixels != old_num_pixels; static const char* kReasons[8] = { "None", "CPU", "BANDWIDTH", "CPU+BANDWIDTH", "VIEW", "CPU+VIEW", "BANDWIDTH+VIEW", "CPU+BANDWIDTH+VIEW", }; LOG(LS_VERBOSE) << "VAdapt Status View: " << view_desired_num_pixels_ << " GD: " << encoder_desired_num_pixels_ << " CPU: " << cpu_desired_num_pixels_ << " Pixels: " << min_num_pixels << " Input: " << input.width << "x" << input.height << " Scale: " << scale << " Resolution: " << new_output.width << "x" << new_output.height << " Changed: " << (changed ? "true" : "false") << " Reason: " << kReasons[adapt_reason_]; if (changed) { // When any adaptation occurs, historic CPU load levels are no longer // accurate. Clear out our state so we can re-learn at the new normal. cpu_adapt_wait_time_ = talk_base::TimeAfter(kCpuLoadMinSampleTime); system_load_average_ = kCpuLoadInitialAverage; } return changed; } } // namespace cricket