| 47 | |
| 48 | #if MEGDNN_WITH_BENCHMARK |
| 49 | std::vector<conv_bias::TestArg> get_conv_bias_benchmaker_args( |
| 50 | std::vector<size_t> kernel, size_t stride) { |
| 51 | using namespace conv_bias; |
| 52 | using Param = param::ConvBias; |
| 53 | using NLMode = param::ConvBias::NonlineMode; |
| 54 | |
| 55 | std::vector<TestArg> args; |
| 56 | auto pack = [&](size_t oc, size_t ic, size_t w, size_t h, size_t kernel, |
| 57 | size_t stride, NLMode nonline_mode) { |
| 58 | Param param; |
| 59 | param.stride_h = stride; |
| 60 | param.stride_w = stride; |
| 61 | param.pad_h = kernel == 1 ? 0 : kernel / 2; |
| 62 | param.pad_w = kernel == 1 ? 0 : kernel / 2; |
| 63 | param.nonlineMode = nonline_mode; |
| 64 | //! no bias |
| 65 | args.emplace_back( |
| 66 | param, TensorShape{1, ic, h, w}, TensorShape{oc, ic, kernel, kernel}, |
| 67 | TensorShape{}); |
| 68 | //! bias broadcast channle |
| 69 | args.emplace_back( |
| 70 | param, TensorShape{1, ic, h, w}, TensorShape{oc, ic, kernel, kernel}, |
| 71 | TensorShape{1, oc, 1, 1}); |
| 72 | //! bias |
| 73 | args.emplace_back( |
| 74 | param, TensorShape{1, ic, h, w}, TensorShape{oc, ic, kernel, kernel}, |
| 75 | TensorShape{ |
| 76 | 1, oc, (h + 2 * param.pad_h - kernel) / stride + 1, |
| 77 | (w + 2 * param.pad_w - kernel) / stride + 1}); |
| 78 | }; |
| 79 | |
| 80 | for (auto nlmode : {NLMode::IDENTITY, NLMode::RELU, NLMode::SIGMOID}) { |
| 81 | for (size_t k : kernel) { |
| 82 | for (size_t ic : {3, 6, 12, 24}) { |
| 83 | for (size_t oc : {3, 6, 12, 24}) { |
| 84 | for (size_t size : {4, 7, 8, 14, 16, 17, 28, 32, 34, 64, 112}) { |
| 85 | pack(oc, ic, size, size, k, stride, nlmode); |
| 86 | } |
| 87 | } |
| 88 | } |
| 89 | } |
| 90 | } |
| 91 | return args; |
| 92 | } |
| 93 | |
| 94 | void benchmarker_conv_bias( |
| 95 | std::vector<conv_bias::TestArg> args, Handle* handle, const char* algo_name, |
no test coverage detected