| 75 | } |
| 76 | |
| 77 | void ConvBiasImpl::AlgoS8MatrixMul::kimpl( |
| 78 | const NCBKernParam& param, const NCBKernIndex& ncb_index) { |
| 79 | auto is_xcorr = !param.filter_meta.should_flip; |
| 80 | UNPACK_CONV_NCB_KERN_SIZES(param); |
| 81 | auto bundle = get_bundle(param); |
| 82 | bundle.set(param.workspace_ptr); |
| 83 | auto IH2 = IH + 2 * PH; |
| 84 | auto IW2 = IW + 2 * PW; |
| 85 | size_t group_id = ncb_index.ndrange_id[0]; |
| 86 | // workspace = tmp..src2 |
| 87 | for (size_t n = 0; n < N; ++n) { |
| 88 | dt_int8* src = const_cast<dt_int8*>(param.src<dt_int8>(n, group_id)); |
| 89 | dt_int8* filter = const_cast<dt_int8*>(param.filter<dt_int8>(group_id)); |
| 90 | dt_int8* dst = static_cast<dt_int8*>(param.dst<dt_int8>(n, group_id)); |
| 91 | dt_int32* bias = const_cast<dt_int32*>(param.bias<dt_int32>(n, group_id)); |
| 92 | |
| 93 | dt_int8 *B, *src2; |
| 94 | if (FH == 1 && FW == 1 && SH == 1 && SW == 1 && PH == 0 && PW == 0) { |
| 95 | // special case: 1x1 |
| 96 | B = const_cast<dt_int8*>(src); |
| 97 | } else { |
| 98 | src2 = static_cast<dt_int8*>(bundle.get(0)); |
| 99 | // copy src to src2; |
| 100 | dt_int8* src2_ptr = src2; |
| 101 | const dt_int8* src_ptr = src; |
| 102 | rep(ic, IC) { |
| 103 | if (PH != 0) { |
| 104 | std::memset(src2_ptr, 0, sizeof(dt_int8) * PH * IW2); |
| 105 | src2_ptr += PH * IW2; |
| 106 | } |
| 107 | rep(ih, IH) { |
| 108 | if (PW != 0) |
| 109 | rep(pw, PW) { *(src2_ptr++) = 0.0f; } |
| 110 | std::memcpy(src2_ptr, src_ptr, sizeof(dt_int8) * IW); |
| 111 | src2_ptr += IW; |
| 112 | src_ptr += IW; |
| 113 | if (PW != 0) |
| 114 | rep(pw, PW) { *(src2_ptr++) = 0.0f; } |
| 115 | } |
| 116 | if (PH != 0) { |
| 117 | std::memset(src2_ptr, 0, sizeof(dt_int8) * PH * IW2); |
| 118 | src2_ptr += PH * IW2; |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | B = static_cast<dt_int8*>(bundle.get(1)); |
| 123 | if (SH == 1 && SW == 1) { |
| 124 | if (is_xcorr) |
| 125 | img2col<true>(src2, B, OC, OH, OW, IC, IH2, IW2, FH, FW); |
| 126 | else |
| 127 | img2col<false>(src2, B, OC, OH, OW, IC, IH2, IW2, FH, FW); |
| 128 | } else { |
| 129 | if (is_xcorr) |
| 130 | img2col_stride<true>( |
| 131 | src2, B, OC, OH, OW, IC, IH2, IW2, FH, FW, SH, SW); |
| 132 | else |
| 133 | img2col_stride<false>( |
| 134 | src2, B, OC, OH, OW, IC, IH2, IW2, FH, FW, SH, SW); |
nothing calls this directly
no test coverage detected