Back to Search Document details
32nd Meeting: Hannover, DE, October 2023 2023-10-17 15:24
AHG10: MTT split modes early termination
Authors: W. Ahmad Ericsson P. Wennersten Ericsson K. Andersson Ericsson
Abstract
This contribution presents a further study on the early termination method presented in JVET-AE0057t was requested to investigate the performance of the proposed scheme for different bit depths and for low QP values.
Meeting Notes
JVET-AF0111 AHG10 MTT split modes early termination [W. Ahmad, P. Wennersten, K. Andersson (Ericsson)]

This contribution presents a further study on the early termination method presented in JVET-AE0057 which was adopted for ECM and VTM. The contribution JVET-AE0057 was enabled for ECM CTC and to enable it in VTM CTC, it was requested to investigate the performance of the proposed scheme for different input bit depths and for low QP values.

Additionally, the proposal is also tested for large QP values and a modification is proposed in the threshold calculation to obtain a better trade-off between encoding time reduction and coding loss for a large range of QP values.

As in JVET-AE0057, the cost of coding a CU without splitting the CU at all is compared to a threshold. If the cost is higher than the threshold, MTT split will not be considered, which is claimed to reduce encoding runtime. The proposal is implemented on top of VTM 20.2 software and compared to VTM 20.2, the average BDR loss and runtimes reported in this contribution are as follows:

  • Test 1.1 (VTM SDR, Proposed condition is applied to CTC QPs, i.e., 22, 27, 32, 37)

Y 0.02%, U 0.00%, V -0.01%, EncTime 96.4%, DecTime 99.7% for the RA configuration,

Y 0.04%, U 0.12%, V 0.03%, EncTime 96.2%, DecTime 100.0% for the LDB configuration,

Y 0.02%, U 0.04%, V 0.05%, EncTime 95.8%, DecTime 100.1% for the RA configuration,

Y 0.05%, U 0.50%, V 0.19%, EncTime 95.5%, DecTime 99.8% for the LDB configuration,

  • Test 1.2 (VTM SDR, Proposed condition is applied to low QP values, i.e., 12, 17, 19, 22)

Y 0.06%, U 0.16%, V 0.16%, EncTime 90.7%, DecTime 100.1% for the RA configuration,

Y 0.06%, U 0.14%, V 0.14%, EncTime 91.3%, DecTime 99.9% for the RA configuration,

  • Test 1.3 (VTM SDR, Proposed condition is applied to high QP values, i.e., 37, 42, 47, 52)

Y 0.00%, U -0.03%, V -0.07%, EncTime 99.7%, DecTime 99.7% for the RA configuration,

Y 0.24%, U 0.43%, V 0.38%, EncTime 96.6%, DecTime 99.9% for the RA configuration,

  • Test 2 (VTM HDR, Proposed condition is applied to HDR CTC QPs, i.e., 22, 27, 32, 37)

Y 0.01%, U 0.08%, V 0.09% in PSNR, 0.02% 0.12% 0.08% in wPSNR, EncTime 98.3%, DecTime 99.7% for the RA configuration,

Y 0.02%, U -0.04%, V 0.04% in PSNR, 0.02% -0.02% 0.09% in wPSNR, EncTime 98.1%, DecTime 100.0% for the RA configuration,

  • Test 3 (VTM HBD, Proposed condition is applied to high bit depth normal QP range)

Y 0.03%, U 0.09%, V 0.08% in PSNR, 0.01% 0.06% 012% in wPSNR, EncTime 97.8%, DecTime 100.1% for the RA configuration,

Y 0.03%, U 0.09%, V 0.14% in PSNR, 0.01% -0.02% 011% in wPSNR, EncTime 97.2%, DecTime 99.7% for the RA configuration,

  • Test 4.1 (ECM SDR, Proposed condition is applied to CTC QPs, i.e., 22, 27, 32, 37)

Y -0.01%, U 0.00%, V -0.01%, EncTime 100.6%, DecTime 99.8% for the RA configuration,

  • Test 4.2 (ECM SDR, Proposed condition is applied to high QP values, i.e., 37, 42, 47, 52)

Y -0.41%, U -0.30%, V -0.40%, EncTime 102.9%, DecTime 99.8% for the RA configuration,

Compared to JVET-AE0057, this contribution proposes a new threshold for non-split cost of 64x64 luma block, with the main motivation being to reduce the coding loss for high QP ranges (37-52). Without the new threshold (i.e., if the existing threshold is used), there would be a 0.24% performance penalty with about 3-4% encoding time reduction for these QPs. With the new threshold, the proposed method would be effectively “turned off” for these higher QP ranges, resulting in little change in performance and encoding time.

The new threshold is more “aggressive” for lower QP values, in that it would result in more coding time reduction. But for these lower QP values, the new threshold wouldn’t incur more coding performance loss.

Comparing to the existing threshold in ECM, using the new threshold would result in 0.41% luma gain but with 3% encoding time increase for high QPs. Several experts commented that this is not desirable tradeoff.

However, it was also commented that a little encoder runtime increase for high QPs should be tolerable, as simulations are much faster for high QPs anyway, and that performance loss incurred by the existing threshold isn’t desirable.

It was also commented that the new threshold is just a refinement of an existing algorithm and commented that if the new threshold is accepted, then it should be applied to the VTM and ECM, as well as VTM-11ecm.

Decision (SW): Adopt the new threshold in JVET-AF0111 to the next releases of ECM, VTM and VTM-11ecm.

Cross checker reported that results are matched for VTM, low QP, SDR. Cross check of other combination of QP ranges and SDR/HDR/HDP has not been performed.

The main reasons for not turning the proposed speedup algorithm on for VTM were: 1) VTM results were not complete, 2) it was desirable to check the algorithm’s behaviour for low QP values and other types of content (high bit-depth, and HDR).

The results in this contribution provided answers to the above questions from last meeting.

It was commented that the CTC of VTM and VTM-11ecm (which serves as the anchor for ECM) should be kept consistent.

Decision (CTC): Enable JVET-AF0111 in VTM CTC. Also turn it on for VTM-11ecm.

JVET-AF0111 (New Threshold)

Y 0.02%, U 0.00%,V -0.01%, EncTime 96.4%, DecTime 99.7% for the RA configuration,

Y 0.04%, U 0.12%, V 0.03%, EncTime 96.2%, DecTime 100.0% for the LDB configuration,

JVET-AF0111 (New Threshold)

Y 0.06%, U 0.16%,V 0.16%, EncTime 90.7%, DecTime 100.1% for the RA configuration,

JVET-AF0111 (New Threshold)

Y 0.00%, U -0.03%,V -0.07%, EncTime 99.7%, DecTime 99.7% for the RA configuration,

JVET-AF0111 (New Threshold)

Y 0.01%, U 0.08%, V 0.09% in PSNR, 0.02% 0.12% 0.08% in wPSNR, EncTime 98.3%, DecTime 99.7% for the RA configuration,

JVET-AF0111 (New Threshold)

Y 0.03%, U 0.09%, V 0.08% in PSNR, 0.01% 0.06% 012% in wPSNR, EncTime 97.8%, DecTime 100.1% for the RA configuration,

Cited:
JVET-AF0247 X
ARCHIVE :
JVET-AF0111-v3.pptx JVET_AF0111-v3.docx Test1.1_VTM-SDR_JVET-AE0057_QP22to37.xlsm Test1.1_VTM-SDR_JVET-AF0111_QP22to37.xlsm Test1.2_VTM-SDR_JVET-AE0057_QP12to22.xlsm Test1.2_VTM-SDR_JVET-AF0111_QP12to22.xlsm Test1.3_VTM-SDR_JVET-AE0057_QP37to52.xlsm Test1.3_VTM-SDR_JVET-AF0111_QP37to52.xlsm Test2_VTM-HDR_JVET-AE0057_QP22to37.xlsm Test2_VTM-HDR_JVET-AF0111_QP22to37.xlsm Test3_VTM-HBD_JVET-AE0057_QP22to37.xlsm Test3_VTM-HBD_JVET-AF0111_QP22to37.xlsm Test4.1_ECM10-SDR_JVET-AF0111_QP22to37.xlsm Test4.2_ECM10-SDR_JVET-AF0111_QP37to52.xlsm
Decisions
Y 0.04%, U 0.12%, V 0.03%, EncTime 96.2%, DecTime 100.0% for the LDB configuration,
Y 0.04%, U 0.12%, V 0.03%, EncTime 96.2%, DecTime 100.0% for the LDB configuration,
Y 0.06%, U 0.16%,V 0.16%, EncTime 90.7%, DecTime 100.1% for the RA configuration,
Y 0.00%, U -0.03%,V -0.07%, EncTime 99.7%, DecTime 99.7% for the RA configuration,
Y 0.01%, U 0.08%, V 0.09% in PSNR, 0.02% 0.12% 0.08% in wPSNR, EncTime 98.3%, DecTime 99.7% for the RA configuration,
Y 0.03%, U 0.09%, V 0.08% in PSNR, 0.01% 0.06% 012% in wPSNR, EncTime 97.8%, DecTime 100.1% for the RA configuration,
Citation