I am getting poor performance (in terms of schedule efficiency) of autoscheduler for simple consecutive subtraction kernel: ``` in = te.placeholder((N, H, W), dtype='float') out = te.compute((N-1, H, W), lambda n, y, x: in[n+1, y, x] - in[n, y, x]) return [in, out] ```
Example (tir) of schedule "found" for particular input sizes: ``` primfn(image_in_1: handle, subtracted_1: handle) -> () attr = {"from_legacy_te_schedule": True, "global_symbol": "main", "tir.noalias": True} buffers = {subtracted: Buffer(subtracted_2: Pointer(float32), float32, [5, 2160, 3840], []), image_in: Buffer(image_in_2: Pointer(float32), float32, [6, 2160, 3840], [])} buffer_map = {image_in_1: image_in, subtracted_1: subtracted} { attr [IterVar(blockIdx.x: int32, (nullptr), "ThreadIndex", "blockIdx.x")] "thread_extent" = 648000; attr [IterVar(threadIdx.x: int32, (nullptr), "ThreadIndex", "threadIdx.x")] "thread_extent" = 64; subtracted_2[((blockIdx.x*64) + threadIdx.x)] = ((float32*)image_in_2[(((blockIdx.x*64) + threadIdx.x) + 8294400)] - (float32*)image_in_2[((blockIdx.x*64) + threadIdx.x)]) } ``` Cuda kernel for this schedule runs on my gpu for 1250us, while simple handmade kernel runs for 950us. Is that expected, or any changes in tuning options/operator "phrasing" can be made to gain better performance? tvm commit 10fca9c --- [Visit Topic](https://discuss.tvm.apache.org/t/autoscheduler-cuda-poor-perfomance-for-subtraction-kernel/10869/1) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.apache.org/email/unsubscribe/fcb886dab61bcc6ecf9f1f617164604b62737c9e15eb6710939687d7b861efc7).