| 70 | } |
| 71 | |
| 72 | void alloc( |
| 73 | size_t N, size_t IC, size_t IH, size_t IW, size_t CHL_MUL, size_t FH, |
| 74 | size_t FW, size_t PH, size_t PW) { |
| 75 | pad_h = PH; |
| 76 | pad_w = PW; |
| 77 | auto mkly = [](const TensorShape& s) { |
| 78 | return TensorLayout{s, dtype::Float32()}; |
| 79 | }; |
| 80 | lsrc = mkly({N, IC, IH, IW}); |
| 81 | lflt0 = mkly({CHL_MUL * IC, IC, FH, FW}); |
| 82 | lflt1 = mkly({IC, CHL_MUL, 1, FH, FW}); |
| 83 | ldst = mkly({N, IC * CHL_MUL, IH - FH + 1 + PH * 2, IW - FW + 1 + PW * 2}); |
| 84 | src0.reset(new Tensor<>(handle, lsrc)); |
| 85 | src1.reset(new Tensor<>(handle, lsrc)); |
| 86 | flt0.reset(new Tensor<>(handle, lflt0)); |
| 87 | flt0_cpu.reset(new Tensor<>(handle_cpu, lflt0)); |
| 88 | flt1.reset(new Tensor<>(handle, lflt1)); |
| 89 | flt1_cpu.reset(new Tensor<>(handle_cpu, lflt1)); |
| 90 | dst0.reset(new Tensor<>(handle, ldst)); |
| 91 | dst1.reset(new Tensor<>(handle, ldst)); |
| 92 | } |
| 93 | |
| 94 | void fill_src() { |
| 95 | rng->exec(src0->tensornd(), {}); |