#10270: yolov4 implementation

[punithsekar]

This PR contains ttnn yolov4 implementation with new conv API.

[punithsekar]

The performance sheets of the following submodules are added herewith:

• Downsample1: ops_perf_results_downsample1_2024_07_11_05_59_20.csv

• Downsample2: ops_perf_results_downsample2_2024_07_11_06_02_41.csv

• Downsample3: ops_perf_results_downsample3_2024_07_11_06_03_58.csv

• Downsample4: ops_perf_results_downsample4_2024_07_11_06_05_30.csv

• Downsample5: ops_perf_results_downsample5_2024_07_11_06_06_58.csv

[punithsekar]

Performance sheet for neck sub-module : ops_perf_results_neck_2024_07_12_16_43_36.csv

We are not able to generate perf sheets for head sub-module as the test hangs after completing the sub-module and also the yolov4 pipeline fails due to OOM issue in DS5 sub-module first conv itself ,but DS5 sub-module is passing when tested separately.

Currently, Using torch maxpool in neck sub-module as ttnn.maxpool are making issues for those configuration #7746.
And the current pipeline of neck sub-modules outputs face low PCC.

[punithsekar]

Latest neck and head sub-module perf sheets:
Neck sub-module: ops_perf_results_neck_latest_2024_07_15_17_20_08.csv
Head sub_module: ops_perf_results_head_2024_07_15_17_21_58.csv

[punithsekar]

All the sub-modules work fine, the YOLOv4 model produces three outputs, and their PCC are greater than 0.95, 0.96, and 0.98, respectively.

Performance sheets of all modules:

Downsample1 sub_module : ops_perf_results_downsample1_17_2024_07_16_12_42_56.csv

Downsample2 sub_module : ops_perf_results_downsample2_17_2024_07_16_12_44_45.csv

Downsample3 sub_module : ops_perf_results_downsample3_17_2024_07_16_12_45_27.csv

Downsample4 sub_module : ops_perf_results_downsample4_17_2024_07_16_12_46_15.csv

Downsample15sub_module : ops_perf_results_downsample5_17_2024_07_16_12_46_55.csv

Neck sub_module : ops_perf_results_neck_17_2024_07_16_12_48_17.csv

Head sub_module : ops_perf_results_head_17_2024_07_16_12_47_31.csv

yolov4 module : ops_perf_results_yolov4_17_2024_07_16_12_49_50.csv

[punithsekar]

Latest performance sheet,
Downsample 1: ops_perf_results_down_1_18_2024_07_18_17_35_49.csv
Downsample 2: ops_perf_results_down_2_18_2024_07_18_17_36_29.csv
Downsample 3: ops_perf_results_down_3_18_2024_07_18_17_37_22.csv
Downsample 4: ops_perf_results_down_4_18_2024_07_18_17_37_55.csv
Downsample 5: ops_perf_results_down_5_18_2024_07_18_17_38_27.csv
Neck : ops_perf_results_neck_sub_18_2024_07_18_17_39_51.csv
Head : ops_perf_results_head_sub_18_2024_07_18_17_39_12.csv
Yolov4 : ops_perf_results_yolov4_18_latest_2024_07_18_17_41_31.csv

Below are the current FPS,

• FPS (MatMul/Conv Ops only): 263.297
• FPS (Other Device Ops): 415.307
• FPS (All Ops): 187.65

Note : Currently using 3 torch maxpool which is not visible in the perf sheets.

[punithsekar]

Latest perf sheet,
yolov4 model: yolov4_19_07_2024.csv

Below are the current FPS,

• FPS (MatMul/Conv Ops only): 263.459
• FPS (Other Device Ops): 15.494
• FPS (All Ops): 14.826

[punithsekar]

Closing this PR since it's outdated and the YOLOv4 model is in the main.