| 63 | } |
| 64 | |
| 65 | void alloc( |
| 66 | size_t N, size_t IC, size_t ID, size_t IH, size_t IW, size_t CHL_MUL, |
| 67 | size_t FD, size_t FH, size_t FW, size_t PD, size_t PH, size_t PW) { |
| 68 | pad_d = PD; |
| 69 | pad_h = PH; |
| 70 | pad_w = PW; |
| 71 | auto mkly = [](const TensorShape& s) { |
| 72 | return TensorLayout{s, dtype::Float32()}; |
| 73 | }; |
| 74 | lsrc = mkly({N, IC, ID, IH, IW}); |
| 75 | lflt0 = mkly({CHL_MUL * IC, IC, FD, FH, FW}); |
| 76 | lflt1 = mkly({IC, CHL_MUL, 1, FD, FH, FW}); |
| 77 | ldst = |
| 78 | mkly({N, IC * CHL_MUL, ID - FD + 1 + PD * 2, IH - FH + 1 + PH * 2, |
| 79 | IW - FW + 1 + PW * 2}); |
| 80 | src0.reset(new Tensor<>(handle, lsrc)); |
| 81 | src1.reset(new Tensor<>(handle, lsrc)); |
| 82 | flt0.reset(new Tensor<>(handle, lflt0)); |
| 83 | flt0_cpu.reset(new Tensor<>(handle_cpu, lflt0)); |
| 84 | flt1.reset(new Tensor<>(handle, lflt1)); |
| 85 | flt1_cpu.reset(new Tensor<>(handle_cpu, lflt1)); |
| 86 | dst0.reset(new Tensor<>(handle, ldst)); |
| 87 | dst1.reset(new Tensor<>(handle, ldst)); |
| 88 | } |
| 89 | |
| 90 | void fill_src() { |
| 91 | rng->exec(src0->tensornd(), {}); |