Appendix-CfL.md

Chroma from Luma Prediction

1. Description of the algorithm

The general idea behind the chroma from luma (CfL) prediction feature is to exploit the correlation between luma and chroma to express the Intra prediction of chroma sample values as an affine function of the corresponding reconstructed luma sample values, where the reconstructed luma samples are sub-sampled to match the chroma sub-sampling. The chroma prediction is given by

where and are predicted chroma and reconstructed luma samples, respectively. The parameters and can be determined (at least theoretically) using least squares regression. The feature provides gains in screen sharing applications.

In practice, the CfL prediction is performed as illustrated in Figure 1 below.

Figure 1. Block diagram of the chroma from luma prediction process.

The steps illustrated in the diagram above can be summarized as follows:

• Consider the reconstructed luma sample values.

• Reconstructed luma samples are sub-sampled to match the chroma sub-sampling.

• Calculate the (i.e. average) of the reconstructed luma sample values.

• Subtract the from the reconstructed luma sample values to generate the AC reconstructed luma sample values, , which has a zero average.

• Compute using the AC reconstructed luma sample values.

• Compute the intra DC mode chroma prediction, . The final chroma from luma prediction is then given by:

2. Implementation of the algorithm

Inputs: luma inverse quantized residuals

Outputs: Best and chroma residuals

Control macros/flags:

Table 1. Control flags related to the CfL prediction.

Flag	Level	Description
cfl_level	Picture	Describes the CfL level of the encoder.

Details of the implementation

Figure 2. High-level encoder pipeline dataflow with CfL feature.

Figure 3. The main function calls leading to CfL prediction. The functions highlighted in blue are where CfL prediction takes place.

Figure 4. Continuation of Figure 2 showing the details of CfL processing in the function CfLPrediction.

The high level dataflow of CfL in SVT-AV1 is shown in Figure 2. CfL prediction takes place in MD through the function CflPrediction and in the encode pass through the function Av1EncodeLoop/Av1EncodeLoop16bit. The details of the CfL prediction in the function CflPrediction are presented in Figure 4. Similar flow is also followed in the function Av1EncodeLoop/Av1EncodeLoop16bit, except for the fact that is calculated only in MD and the encode pass would use the same to perform the final CfL prediction. In the following, the details of the CfL processing in the function CflPrediction are presented.

For an intra coded block, the function CflPrediction is called when the intra_chroma_mode is set to UV_CFL_PRED. There are four steps in the function:

Step 1: Reconstruct the Luma samples (AV1PerformInverseTransformReconLuma)

The first step is to reconstruct the luma samples, since the latter would be used to generate the chroma prediction. At this stage in the encoder pipeline, the luma residuals are transformed, quantized and inverse quantized. In this step, the inverse transform is applied, and the reconstructed luma residuals are added to the prediction to build the reconstructed samples.

Step 2: Compute the AC component of the luma intra prediction

In this step, the luma reconstructed samples are down sampled to match the size of chroma samples using the cfl_luma_subsampling_420 function. Then the AC luma values are calculated by subtracting the DC luma value using the svt_subtract_average function. The resulting AC values are stored in the pred_buf_q3 buffer.

Step 3: Find the best

The best values for the chroma components are calculated by minimizing the overall full cost. The algorithm performs a search over the 16 possible values of and finds the best value that minimizes the joint prediction cost. The search is performed in the context of a joint sign between the two chroma components. After the best value for is calculated, the joint cost is compared with the cost of DC prediction and the winner is selected.

Step 4: Generate the chroma prediction

After the best is selected, the prediction using the CfL mode is performed using the svt_cfl_predict function. The chroma residuals are then calculated using the function residual_kernel.

3. Optimization of the algorithm

Finding the best requires searching different values in the set of allowed values and calculating the cost associated with each value. Performing this search process in MD for every luma mode and block size at MD would be very costly. In order to find the best quality-speed trade offs for the feature, CfL and UV (i.e. chroma) control signals are defined with multiple levels. Table 2 shows the CfL control signals and their descriptions. The CfL control signals are set in the function set_cfl_ctrls based on the cfl_level value.

Table 2. CfL control signals description.

Signal	Description
enabled	0/1: Disable/Enable CfL candidate injection
itr_th	Threshold to indicate the minimum number of α values to try. However if a large enough number of α values are evaluated without improvements in the overall rate-distortion cost, the search would stop.

Table 3 shows the CfL-related UV control signal and its description. The signal is set in the function set_chroma_controls based on the chroma level uv_level.

Table 3. CfL-related UV control signals description.

Signal	Description
uv_cfl_th	Threshold to skip CfL if the ratio of the best intra cost to the best inter cost is greater than uv_cfl_th.

The CfL and UV levels are set according to the encoder preset, PD_PASS, temporal layer index, slice type and screen content class.

4. Signaling

CfL is an Intra chroma mode that is allowed only for blocks with height and width of 32 or smaller. The entropy encoder signals the chroma mode per block and if the mode is CfL, extra parameters are included in the bit stream:

• cfl_alpha_signs contains the sign of the alpha values for U and V packed together into a single syntax element with 8 possible values. (The combination of two zero signs is prohibited as it is redundant with DC Intra prediction.)

Notes

The feature settings that are described in this document were compiled at v0.9.0 of the code and may not reflect the current status of the code. The description in this document represents an example showing how features would interact with the SVT architecture. For the most up-to-date settings, it's recommended to review the section of the code implementing this feature.

References

[1] Luc N. Trudeau, Nathan E. Egge and David Barr, “Predicting Chroma from Luma in AV1”, Data Compression Conference, 2017.