Our AutoTVM use debug_skip_region to skip our precompute stage (like precompute kernel). However, debug_skip_region can not reflect practical running, which just only create empty tensor for it. Empty tensor will make AutoTVM record the performance incorrect. This condition will become worse like Winograd. In Winograd, our weight shape will be changed from (KH, KW, CO, CI) to (alpha, alpha, CO, CI). Alpha equals to tile_size + KH - 1. If tile_size is 4 (KH is 3), then alpha is 6. So to avoid this issue, we should avoid using debug_skip_region in all platforms of AutoTVM template and use the technique like this PR: #5200.
Open this issue to track this.
Our AutoTVM use
debug_skip_regionto skip our precompute stage (like precompute kernel). However,debug_skip_regioncan not reflect practical running, which just only create empty tensor for it. Empty tensor will make AutoTVM record the performance incorrect. This condition will become worse like Winograd. In Winograd, our weight shape will be changed from (KH, KW, CO, CI) to (alpha, alpha, CO, CI). Alpha equals totile_size + KH - 1. If tile_size is 4 (KH is 3), then alpha is 6. So to avoid this issue, we should avoid usingdebug_skip_regionin all platforms of AutoTVM template and use the technique like this PR: #5200.Open this issue to track this.