# Generated by dnn/scripts/cutlass_generator/gen_list.py

cutlass_gen_list = [
    "cutlass_simt_sgemm_8x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_16x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_16x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_32x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_32x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_64x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_16x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_32x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_64x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_128x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_64x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_128x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_32x256_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_64x256_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_128x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_256x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_256x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_nn_align1.cu",
    "cutlass_simt_sgemm_8x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_16x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_16x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_32x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_32x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_64x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_16x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_32x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_64x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_128x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_64x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_128x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_32x256_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_64x256_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_128x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_256x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_256x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_nt_align1.cu",
    "cutlass_simt_sgemm_8x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_16x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_16x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_32x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_32x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_64x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_16x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_32x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_64x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_128x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_64x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_128x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_32x256_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_64x256_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_128x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_256x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_256x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_tn_align1.cu",
    "cutlass_simt_sgemm_8x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_8x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_16x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_16x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_32x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_32x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_64x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_16x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_16x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_32x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_64x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_128x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_64x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_128x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_32x256_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_32x256_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_64x256_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_64x256_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_128x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_128x128_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_256x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x32_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_256x64_8x2_tt_align1.cu",
    "cutlass_simt_sgemm_split_k_parallel_256x64_8x2_tt_align1.cu",
    "all_gemm_simt_operations.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_f16_128x128_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s1688gemm_split_k_parallel_f16_128x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h1688gemm_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h1688gemm_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_h1688gemm_128x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h1688gemm_split_k_parallel_128x128_32x2_tt_align2.cu",
    "all_gemm_tensorop1688_operations.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_f16_s884gemm_f16_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_f16_128x128_32x2_tt_align2.cu",
    "cutlass_tensorop_f16_s884gemm_split_k_parallel_f16_128x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nn_align8.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nn_align8.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nn_align4.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nn_align4.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nn_align2.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nn_align2.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nt_align8.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nt_align8.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nt_align4.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nt_align4.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_nt_align2.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_nt_align2.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tn_align8.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tn_align8.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tn_align4.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tn_align4.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tn_align2.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tn_align2.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tt_align8.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tt_align8.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tt_align4.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tt_align4.cu",
    "cutlass_tensorop_h884gemm_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_256x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h884gemm_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x256_32x2_tt_align2.cu",
    "cutlass_tensorop_h884gemm_128x128_32x2_tt_align2.cu",
    "cutlass_tensorop_h884gemm_split_k_parallel_128x128_32x2_tt_align2.cu",
    "all_gemm_tensorop884_operations.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_32_tt_align4x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_16_tt_align4x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align4x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_16_tt_align2x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align2x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_4_tt_align2x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_8_tt_align1x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_4_tt_align1x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x128_2_tt_align1x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_64_tt_align4x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_32_tt_align4x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align4x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_32_tt_align2x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align2x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_8_tt_align2x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_16_tt_align1x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_8_tt_align1x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x64_4_tt_align1x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_128_tt_align4x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_64_tt_align4x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align4x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_64_tt_align2x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align2x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_16_tt_align2x1.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_32_tt_align1x4.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_16_tt_align1x2.cu",
    "cutlass_simt_sgemv_batched_strided_1x32_8_tt_align1x1.cu",
    "cutlass_simt_s8_idgrad_id_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4_align4x16.cu",
    "cutlass_simt_s8_idgrad_s2_id_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4_align4x16.cu",
    "cutlass_simt_s8_idgrad_id_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4_align4x4.cu",
    "cutlass_simt_s8_idgrad_s2_id_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4_align4x4.cu",
    "cutlass_simt_s8_idgrad_id_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4_align4x8.cu",
    "cutlass_simt_s8_idgrad_s2_id_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4_align4x8.cu",
    "cutlass_simt_s8_idgrad_id_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4_align4x4.cu",
    "cutlass_simt_s8_idgrad_s2_id_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4_align4x4.cu",
    "all_deconv_simt_operations.cu",
    "cutlass_tensorop_s8_i8816dgrad_id_s8_128x32x32_64x32x32_1_nhwc_ck4rs4_align4x4.cu",
    "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_128x32x32_64x32x32_1_nhwc_ck4rs4_align4x4.cu",
    "cutlass_tensorop_s8_i8816dgrad_id_s8_64x16x32_64x16x32_2_nhwc_ck4rs4_align4x4.cu",
    "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_64x16x32_64x16x32_2_nhwc_ck4rs4_align4x4.cu",
    "cutlass_tensorop_s8_i8816dgrad_id_s8_128x32x32_64x32x32_1_nhwc_ck8rs8_align8x8.cu",
    "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_128x32x32_64x32x32_1_nhwc_ck8rs8_align8x8.cu",
    "cutlass_tensorop_s8_i8816dgrad_id_s8_64x16x32_64x16x32_2_nhwc_ck8rs8_align8x8.cu",
    "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_64x16x32_64x16x32_2_nhwc_ck8rs8_align8x8.cu",
    "cutlass_tensorop_s8_i8816dgrad_id_s8_128x32x32_64x32x32_1_nhwc_ck16rs16_align16x16.cu",
    "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_128x32x32_64x32x32_1_nhwc_ck16rs16_align16x16.cu",
    "cutlass_tensorop_s8_i8816dgrad_id_s8_64x16x32_64x16x32_2_nhwc_ck16rs16_align16x16.cu",
    "cutlass_tensorop_s8_i8816dgrad_s2_id_s8_64x16x32_64x16x32_2_nhwc_ck16rs16_align16x16.cu",
    "all_deconv_tensorop8816_operations.cu",
    "cutlass_simt_s8_ifprop_id_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
    "cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_align4x8.cu",
    "cutlass_simt_s8_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
    "cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_align4x4.cu",
    "cutlass_simt_s8_ifprop_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_id_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32_align4x16.cu",
    "cutlass_simt_u4_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
    "cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
    "cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
    "cutlass_simt_u4_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
    "cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
    "cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
    "cutlass_simt_s4_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
    "cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
    "cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc_align4x8.cu",
    "cutlass_simt_s4_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
    "cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
    "cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc_align4x4.cu",
    "cutlass_simt_f32_ifprop_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
    "cutlass_simt_f32_ifprop_1x1_id_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
    "cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
    "cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
    "cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
    "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw_align4x8.cu",
    "cutlass_simt_f32_ifprop_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
    "cutlass_simt_f32_ifprop_1x1_id_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
    "cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
    "cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
    "cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
    "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw_align4x4.cu",
    "all_conv2d_simt_operations.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x64x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nc32hw32_c32rsk32_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x128x32_32x64x32_1_nc32hw32_c32rsk32_nc4hw4_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc4hw4_align4x4.cu",
    "cutlass_tensorop_s4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_128x32x32_64x32x32_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_id_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_relu_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_f32_i8816fprop_1x1_roc_hswish_s8_64x16x32_64x16x32_2_nhwc_nc16hw16_align16x16.cu",
    "all_conv2d_tensorop8816_operations.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x256x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x128_64x64x128_2_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nc64hw64_c64rsk64_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s4_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_u4_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_hswish_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc8hw8_align8x8.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc16hw16_align16x16.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x16x64_128x16x64_2_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x32x64_64x32x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_id_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "cutlass_tensorop_s8_i8832fprop_1x1_roc_relu_u4_s4_128x64x64_64x64x64_1_nhwc_nc32hw32_align32x32.cu",
    "all_conv2d_tensorop8832_operations.cu",
    "cutlass_simt_sdwfprop_id_f32_32x32x8_32x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_32x32x8_32x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_32x32x8_32x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_32x32x8_32x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_32x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_32x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_32x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_32x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_64x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_64x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_64x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_64x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_32x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_32x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_32x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_32x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_64x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_64x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_64x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_64x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_128x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_128x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_128x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_128x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_64x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_64x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_64x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_64x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_128x64x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_128x64x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_128x64x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_128x64x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_id_f32_128x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_128x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwfprop_id_f32_128x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwfprop_relu_f32_128x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "all_dwconv2d_fprop_simt_operations.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwfprop_relu_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "all_dwconv2d_fprop_tensorop884_operations.cu",
    "cutlass_simt_sdwdgrad_id_f32_32x32x8_32x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_32x32x8_32x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_32x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_32x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_64x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_64x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_32x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_32x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_64x64x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_64x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_128x32x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_128x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_64x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_64x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_128x64x8_64x32x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_128x64x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_128x128x8_32x64x8_2_nchw_nchw_align4x1.cu",
    "cutlass_simt_sdwdgrad_id_f32_128x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "all_dwconv2d_dgrad_simt_operations.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_f16_s884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_h884dwdgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "all_dwconv2d_dgrad_tensorop884_operations.cu",
    "cutlass_simt_sdwwgrad_id_f32_32x32x8_32x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_32x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_64x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_32x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_64x64x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_128x32x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_64x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_128x64x8_64x32x8_2_nchw_nchw_align1x1.cu",
    "cutlass_simt_sdwwgrad_id_f32_128x128x8_32x64x8_2_nchw_nchw_align1x1.cu",
    "all_dwconv2d_wgrad_simt_operations.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align8x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align2x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x8.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x2.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x256x32_64x64x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x128x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_128x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "cutlass_tensorop_s884dwwgrad_id_f16_64x64x32_32x32x32_2_nchw_nchw_align1x1.cu",
    "all_dwconv2d_wgrad_tensorop884_operations.cu",
]