Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1 | |
Luke Hutton | 261b7b6 | 2023-01-10 14:50:31 +0000 | [diff] [blame] | 2 | // Copyright (c) 2020-2023, ARM Limited. |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 3 | // |
| 4 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | // you may not use this file except in compliance with the License. |
| 6 | // You may obtain a copy of the License at |
| 7 | // |
| 8 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | // |
| 10 | // Unless required by applicable law or agreed to in writing, software |
| 11 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | // See the License for the specific language governing permissions and |
| 14 | // limitations under the License. |
| 15 | |
| 16 | #include "op_factory.h" |
| 17 | #include "activation_funcs.h" |
| 18 | #include "comparison.h" |
| 19 | #include "control_flow.h" |
| 20 | #include "custom.h" |
| 21 | #include "data_layout.h" |
| 22 | #include "data_nodes.h" |
| 23 | #include "ewise_binary.h" |
| 24 | #include "ewise_ternary.h" |
| 25 | #include "ewise_unary.h" |
| 26 | #include "image.h" |
| 27 | #include "reduction.h" |
| 28 | #include "scatter_gather.h" |
| 29 | #include "tensor_ops.h" |
| 30 | #include "type_conversion.h" |
| 31 | |
| 32 | using namespace TosaReference; |
| 33 | using namespace tosa; |
| 34 | |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 35 | GraphNode* OpFactory::newOp(SubgraphTraverser* sgt, |
| 36 | TosaSerializationHandler* tsh, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 37 | Op opType, |
| 38 | TosaAttributeBase* attribute, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 39 | uint64_t id, |
| 40 | DType inputDType, |
| 41 | int inputRank, |
| 42 | DType outputDType, |
| 43 | int outputRank, |
| 44 | DType weightDType, |
| 45 | int weightRank) |
| 46 | { |
| 47 | switch (opType) |
| 48 | { |
| 49 | // tensor_ops |
| 50 | case Op_ARGMAX: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 51 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 52 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 53 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, FP32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 54 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 55 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, INT16); |
| 56 | break; |
| 57 | case Op_AVG_POOL2D: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 58 | DEF_FACTORY_ONE_TYPE_ONE_ACCUM(OpAvgPool2d, Pool, FP16, FP16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 59 | DEF_FACTORY_ONE_TYPE_ONE_ACCUM(OpAvgPool2d, Pool, FP16, FP32); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 60 | DEF_FACTORY_ONE_TYPE_ONE_ACCUM(OpAvgPool2d, Pool, BF16, FP32); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 61 | DEF_FACTORY_ONE_TYPE_ONE_ACCUM(OpAvgPool2d, Pool, FP32, FP32); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 62 | DEF_FACTORY_ONE_TYPE_ONE_ACCUM(OpAvgPool2d, Pool, INT8, INT32); |
| 63 | DEF_FACTORY_ONE_TYPE_ONE_ACCUM(OpAvgPool2d, Pool, INT16, INT32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 64 | break; |
| 65 | case Op_CONV2D: |
James Ward | d34b3fc | 2023-01-18 14:51:25 +0000 | [diff] [blame] | 66 | DEF_FACTORY_THREE_TYPE(OpConv2d, FP16, FP16, FP16); |
| 67 | DEF_FACTORY_THREE_TYPE(OpConv2d, FP16, FP16, FP32); |
| 68 | DEF_FACTORY_THREE_TYPE(OpConv2d, BF16, BF16, FP32); |
| 69 | DEF_FACTORY_THREE_TYPE(OpConv2d, FP32, FP32, FP32); |
| 70 | DEF_FACTORY_THREE_TYPE(OpConv2d, INT8, INT4, INT32); |
| 71 | DEF_FACTORY_THREE_TYPE(OpConv2d, INT8, INT8, INT32); |
| 72 | DEF_FACTORY_THREE_TYPE(OpConv2d, INT16, INT8, INT48); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 73 | break; |
Kevin Cheng | 1533b85 | 2021-09-01 12:51:58 -0700 | [diff] [blame] | 74 | case Op_CONV3D: |
James Ward | d34b3fc | 2023-01-18 14:51:25 +0000 | [diff] [blame] | 75 | DEF_FACTORY_THREE_TYPE(OpConv3d, FP16, FP16, FP16); |
| 76 | DEF_FACTORY_THREE_TYPE(OpConv3d, FP16, FP16, FP32); |
| 77 | DEF_FACTORY_THREE_TYPE(OpConv3d, BF16, BF16, FP32); |
| 78 | DEF_FACTORY_THREE_TYPE(OpConv3d, FP32, FP32, FP32); |
| 79 | DEF_FACTORY_THREE_TYPE(OpConv3d, INT8, INT4, INT32); |
| 80 | DEF_FACTORY_THREE_TYPE(OpConv3d, INT8, INT8, INT32); |
| 81 | DEF_FACTORY_THREE_TYPE(OpConv3d, INT16, INT8, INT48); |
Kevin Cheng | 1533b85 | 2021-09-01 12:51:58 -0700 | [diff] [blame] | 82 | break; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 83 | case Op_DEPTHWISE_CONV2D: |
James Ward | d34b3fc | 2023-01-18 14:51:25 +0000 | [diff] [blame] | 84 | DEF_FACTORY_THREE_TYPE(OpDepthwiseConv2d, FP16, FP16, FP16); |
| 85 | DEF_FACTORY_THREE_TYPE(OpDepthwiseConv2d, FP16, FP16, FP32); |
| 86 | DEF_FACTORY_THREE_TYPE(OpDepthwiseConv2d, BF16, BF16, FP32); |
| 87 | DEF_FACTORY_THREE_TYPE(OpDepthwiseConv2d, FP32, FP32, FP32); |
| 88 | DEF_FACTORY_THREE_TYPE(OpDepthwiseConv2d, INT8, INT4, INT32); |
| 89 | DEF_FACTORY_THREE_TYPE(OpDepthwiseConv2d, INT8, INT8, INT32); |
| 90 | DEF_FACTORY_THREE_TYPE(OpDepthwiseConv2d, INT16, INT8, INT48); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 91 | break; |
Luke Hutton | 5728713 | 2023-02-06 14:54:18 +0000 | [diff] [blame] | 92 | case Op_FFT2D: |
| 93 | DEF_FACTORY_ONE_TYPE(OpFFT2d, FP32); |
| 94 | break; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 95 | case Op_FULLY_CONNECTED: |
James Ward | d34b3fc | 2023-01-18 14:51:25 +0000 | [diff] [blame] | 96 | DEF_FACTORY_THREE_TYPE(OpFullyConnected, FP16, FP16, FP16); |
| 97 | DEF_FACTORY_THREE_TYPE(OpFullyConnected, FP16, FP16, FP32); |
| 98 | DEF_FACTORY_THREE_TYPE(OpFullyConnected, BF16, BF16, FP32); |
| 99 | DEF_FACTORY_THREE_TYPE(OpFullyConnected, FP32, FP32, FP32); |
| 100 | DEF_FACTORY_THREE_TYPE(OpFullyConnected, INT8, INT4, INT32); |
| 101 | DEF_FACTORY_THREE_TYPE(OpFullyConnected, INT8, INT8, INT32); |
| 102 | DEF_FACTORY_THREE_TYPE(OpFullyConnected, INT16, INT8, INT48); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 103 | break; |
| 104 | case Op_MATMUL: |
James Ward | d34b3fc | 2023-01-18 14:51:25 +0000 | [diff] [blame] | 105 | DEF_FACTORY_TWO_TYPE_IN_OUT(OpMatMul, FP16, FP16); |
| 106 | DEF_FACTORY_TWO_TYPE_IN_OUT(OpMatMul, FP16, FP32); |
| 107 | DEF_FACTORY_TWO_TYPE_IN_OUT(OpMatMul, BF16, FP32); |
| 108 | DEF_FACTORY_TWO_TYPE_IN_OUT(OpMatMul, FP32, FP32); |
| 109 | DEF_FACTORY_TWO_TYPE_IN_OUT(OpMatMul, INT8, INT32); |
| 110 | DEF_FACTORY_TWO_TYPE_IN_OUT(OpMatMul, INT16, INT48); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 111 | break; |
| 112 | case Op_MAX_POOL2D: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 113 | DEF_FACTORY_ONE_TYPE(OpMaxPool2d, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 114 | DEF_FACTORY_ONE_TYPE(OpMaxPool2d, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 115 | DEF_FACTORY_ONE_TYPE(OpMaxPool2d, FP32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 116 | DEF_FACTORY_ONE_TYPE(OpMaxPool2d, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 117 | DEF_FACTORY_ONE_TYPE(OpMaxPool2d, INT16); |
| 118 | break; |
Luke Hutton | 261b7b6 | 2023-01-10 14:50:31 +0000 | [diff] [blame] | 119 | case Op_RFFT2D: |
| 120 | DEF_FACTORY_ONE_TYPE(OpRFFT2d, FP32); |
| 121 | break; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 122 | case Op_TRANSPOSE_CONV2D: |
James Ward | d34b3fc | 2023-01-18 14:51:25 +0000 | [diff] [blame] | 123 | DEF_FACTORY_THREE_TYPE(OpTransposeConv2d, FP16, FP16, FP16); |
| 124 | DEF_FACTORY_THREE_TYPE(OpTransposeConv2d, FP16, FP16, FP32); |
| 125 | DEF_FACTORY_THREE_TYPE(OpTransposeConv2d, BF16, BF16, FP32); |
| 126 | DEF_FACTORY_THREE_TYPE(OpTransposeConv2d, FP32, FP32, FP32); |
| 127 | DEF_FACTORY_THREE_TYPE(OpTransposeConv2d, INT8, INT4, INT32); |
| 128 | DEF_FACTORY_THREE_TYPE(OpTransposeConv2d, INT8, INT8, INT32); |
| 129 | DEF_FACTORY_THREE_TYPE(OpTransposeConv2d, INT16, INT8, INT48); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 130 | break; |
| 131 | |
| 132 | // activation_funcs |
| 133 | case Op_CLAMP: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 134 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpClamp, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 135 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpClamp, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 136 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpClamp, FP32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 137 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpClamp, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 138 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpClamp, INT16); |
| 139 | break; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 140 | case Op_SIGMOID: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 141 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSigmoid, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 142 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSigmoid, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 143 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSigmoid, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 144 | break; |
| 145 | case Op_TANH: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 146 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpTanh, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 147 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpTanh, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 148 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpTanh, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 149 | break; |
| 150 | |
| 151 | // ewise_binary |
| 152 | case Op_ADD: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 153 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAdd, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 154 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAdd, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 155 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAdd, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 156 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAdd, INT32); |
| 157 | break; |
| 158 | case Op_ARITHMETIC_RIGHT_SHIFT: |
| 159 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpArithmeticRightShift, INT8); |
| 160 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpArithmeticRightShift, INT16); |
| 161 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpArithmeticRightShift, INT32); |
| 162 | break; |
| 163 | case Op_BITWISE_AND: |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 164 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseAnd, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 165 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseAnd, INT16); |
| 166 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseAnd, INT32); |
| 167 | break; |
| 168 | case Op_BITWISE_OR: |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 169 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseOr, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 170 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseOr, INT16); |
| 171 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseOr, INT32); |
| 172 | break; |
| 173 | case Op_BITWISE_XOR: |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 174 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseXor, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 175 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseXor, INT16); |
| 176 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseXor, INT32); |
| 177 | break; |
Matthew Haddon | 459443c | 2021-08-23 16:43:13 +0100 | [diff] [blame] | 178 | case Op_INTDIV: |
| 179 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIntdiv, INT32); |
Kevin Cheng | 14d7f7a | 2021-05-12 10:44:49 -0700 | [diff] [blame] | 180 | break; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 181 | case Op_LOGICAL_AND: |
| 182 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalAnd, BOOL); |
| 183 | break; |
| 184 | case Op_LOGICAL_LEFT_SHIFT: |
| 185 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalLeftShift, INT8); |
| 186 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalLeftShift, INT16); |
| 187 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalLeftShift, INT32); |
| 188 | break; |
| 189 | case Op_LOGICAL_RIGHT_SHIFT: |
| 190 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalRightShift, INT8); |
| 191 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalRightShift, INT16); |
| 192 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalRightShift, INT32); |
| 193 | break; |
| 194 | case Op_LOGICAL_OR: |
| 195 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalOr, BOOL); |
| 196 | break; |
| 197 | case Op_LOGICAL_XOR: |
Jeremy Johnson | 615feee | 2022-03-10 16:00:49 +0000 | [diff] [blame] | 198 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalXor, BOOL); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 199 | break; |
| 200 | case Op_MAXIMUM: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 201 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMaximum, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 202 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMaximum, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 203 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMaximum, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 204 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMaximum, INT32); |
| 205 | break; |
| 206 | case Op_MINIMUM: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 207 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMinimum, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 208 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMinimum, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 209 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMinimum, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 210 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpMinimum, INT32); |
| 211 | break; |
| 212 | case Op_MUL: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 213 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpMul, FP16, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 214 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpMul, BF16, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 215 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpMul, FP32, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 216 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpMul, INT8, INT32); |
| 217 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpMul, INT16, INT32); |
| 218 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpMul, INT32, INT32); |
| 219 | break; |
| 220 | case Op_POW: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 221 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpPow, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 222 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpPow, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 223 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpPow, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 224 | break; |
| 225 | case Op_SUB: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 226 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSub, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 227 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSub, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 228 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSub, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 229 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSub, INT32); |
| 230 | break; |
| 231 | case Op_TABLE: |
Kevin Cheng | 571f718 | 2021-05-24 17:20:01 -0700 | [diff] [blame] | 232 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpTable, INT8); |
| 233 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpTable, INT16); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 234 | break; |
| 235 | |
| 236 | // ewise_unary |
| 237 | case Op_ABS: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 238 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAbs, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 239 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAbs, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 240 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAbs, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 241 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpAbs, INT32); |
| 242 | break; |
| 243 | case Op_BITWISE_NOT: |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 244 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseNot, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 245 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseNot, INT16); |
| 246 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpBitwiseNot, INT32); |
| 247 | break; |
| 248 | case Op_CEIL: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 249 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpCeil, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 250 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpCeil, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 251 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpCeil, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 252 | break; |
| 253 | case Op_CLZ: |
| 254 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpClz, INT32); |
| 255 | break; |
| 256 | case Op_EXP: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 257 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpExp, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 258 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpExp, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 259 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpExp, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 260 | break; |
| 261 | case Op_FLOOR: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 262 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpFloor, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 263 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpFloor, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 264 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpFloor, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 265 | break; |
| 266 | case Op_LOG: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 267 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLog, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 268 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLog, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 269 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLog, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 270 | break; |
| 271 | case Op_LOGICAL_NOT: |
| 272 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpLogicalNot, BOOL); |
| 273 | break; |
| 274 | case Op_NEGATE: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 275 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpNegate, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 276 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpNegate, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 277 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpNegate, FP32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 278 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpNegate, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 279 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpNegate, INT16); |
| 280 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpNegate, INT32); |
| 281 | break; |
| 282 | case Op_RECIPROCAL: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 283 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpReciprocal, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 284 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpReciprocal, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 285 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpReciprocal, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 286 | break; |
| 287 | case Op_RSQRT: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 288 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpRsqrt, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 289 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpRsqrt, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 290 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpRsqrt, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 291 | break; |
| 292 | |
| 293 | // ewise_ternary |
| 294 | case Op_SELECT: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 295 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSelect, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 296 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSelect, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 297 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSelect, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 298 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSelect, INT8); |
| 299 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSelect, INT16); |
| 300 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSelect, INT32); |
| 301 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpSelect, BOOL); |
| 302 | break; |
| 303 | |
| 304 | // comparison |
| 305 | case Op_EQUAL: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 306 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpEqual, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 307 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpEqual, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 308 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpEqual, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 309 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpEqual, INT32); |
| 310 | break; |
| 311 | case Op_GREATER: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 312 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreater, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 313 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreater, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 314 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreater, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 315 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreater, INT32); |
| 316 | break; |
| 317 | case Op_GREATER_EQUAL: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 318 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreaterEqual, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 319 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreaterEqual, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 320 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreaterEqual, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 321 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpGreaterEqual, INT32); |
| 322 | break; |
| 323 | |
| 324 | // reduction |
| 325 | case Op_REDUCE_ALL: |
| 326 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceAll, BOOL); |
| 327 | break; |
| 328 | case Op_REDUCE_ANY: |
| 329 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceAny, BOOL); |
| 330 | break; |
| 331 | case Op_REDUCE_MAX: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 332 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 333 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 334 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, FP32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 335 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 336 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, INT16); |
| 337 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMax, INT32); |
| 338 | break; |
| 339 | case Op_REDUCE_MIN: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 340 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 341 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 342 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, FP32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 343 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 344 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, INT16); |
| 345 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceMin, INT32); |
| 346 | break; |
| 347 | case Op_REDUCE_PRODUCT: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 348 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceProduct, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 349 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceProduct, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 350 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceProduct, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 351 | break; |
| 352 | case Op_REDUCE_SUM: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 353 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 354 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 355 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSum, FP32); |
Jeremy Johnson | 7de9b45 | 2022-04-05 14:31:37 +0100 | [diff] [blame] | 356 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReduceSumInt, INT32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 357 | break; |
| 358 | |
| 359 | // data layout |
| 360 | case Op_CONCAT: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 361 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpConcat, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 362 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpConcat, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 363 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpConcat, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 364 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpConcat, INT8); |
| 365 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpConcat, INT16); |
| 366 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpConcat, INT32); |
| 367 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpConcat, BOOL); |
| 368 | break; |
| 369 | case Op_PAD: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 370 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpPad, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 371 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpPad, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 372 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpPad, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 373 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpPad, INT32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 374 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpPad, INT8); |
| 375 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpPad, INT16); |
| 376 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpPad, BOOL); |
| 377 | break; |
| 378 | case Op_RESHAPE: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 379 | DEF_FACTORY_RESHAPE(OpReshape, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 380 | DEF_FACTORY_RESHAPE(OpReshape, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 381 | DEF_FACTORY_RESHAPE(OpReshape, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 382 | DEF_FACTORY_RESHAPE(OpReshape, INT8); |
| 383 | DEF_FACTORY_RESHAPE(OpReshape, INT16); |
| 384 | DEF_FACTORY_RESHAPE(OpReshape, INT32); |
| 385 | DEF_FACTORY_RESHAPE(OpReshape, BOOL); |
| 386 | break; |
| 387 | case Op_REVERSE: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 388 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReverse, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 389 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReverse, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 390 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReverse, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 391 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReverse, INT8); |
| 392 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReverse, INT16); |
| 393 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReverse, INT32); |
| 394 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpReverse, BOOL); |
| 395 | break; |
| 396 | case Op_SLICE: |
Luke Hutton | a4e48ca | 2023-02-22 11:53:48 +0000 | [diff] [blame] | 397 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpSlice, FP16); |
| 398 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpSlice, BF16); |
| 399 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpSlice, FP32); |
| 400 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpSlice, INT8); |
| 401 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpSlice, INT16); |
| 402 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpSlice, INT32); |
| 403 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpSlice, BOOL); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 404 | break; |
| 405 | case Op_TILE: |
Luke Hutton | a4e48ca | 2023-02-22 11:53:48 +0000 | [diff] [blame] | 406 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTile, FP16); |
| 407 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTile, BF16); |
| 408 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTile, FP32); |
| 409 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTile, INT8); |
| 410 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTile, INT16); |
| 411 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTile, INT32); |
| 412 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTile, BOOL); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 413 | break; |
| 414 | case Op_TRANSPOSE: |
Luke Hutton | a4e48ca | 2023-02-22 11:53:48 +0000 | [diff] [blame] | 415 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTranspose, BOOL); |
| 416 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTranspose, FP16); |
| 417 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTranspose, BF16); |
| 418 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTranspose, FP32); |
| 419 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTranspose, INT8); |
| 420 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTranspose, INT16); |
| 421 | DEF_FACTORY_RANK1_6_ONE_RANK_ONE_TYPE(OpTranspose, INT32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 422 | break; |
| 423 | |
| 424 | // scatter_gather |
| 425 | case Op_GATHER: |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 426 | DEF_FACTORY_ONE_TYPE(OpGather, INT8); |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 427 | DEF_FACTORY_ONE_TYPE(OpGather, INT16); |
| 428 | DEF_FACTORY_ONE_TYPE(OpGather, INT32); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 429 | DEF_FACTORY_ONE_TYPE(OpGather, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 430 | DEF_FACTORY_ONE_TYPE(OpGather, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 431 | DEF_FACTORY_ONE_TYPE(OpGather, FP32); |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 432 | break; |
| 433 | case Op_SCATTER: |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 434 | DEF_FACTORY_ONE_TYPE(OpScatter, INT8); |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 435 | DEF_FACTORY_ONE_TYPE(OpScatter, INT16); |
| 436 | DEF_FACTORY_ONE_TYPE(OpScatter, INT32); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 437 | DEF_FACTORY_ONE_TYPE(OpScatter, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 438 | DEF_FACTORY_ONE_TYPE(OpScatter, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 439 | DEF_FACTORY_ONE_TYPE(OpScatter, FP32); |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 440 | break; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 441 | |
| 442 | // image |
| 443 | case Op_RESIZE: |
TatWai Chong | f732609 | 2022-06-08 12:17:14 -0700 | [diff] [blame] | 444 | DEF_FACTORY_TWO_TYPE_RESIZE_INT16(OpResize, INT8, INT32); |
| 445 | DEF_FACTORY_TWO_TYPE_RESIZE_INT16(OpResize, INT8, INT8); |
| 446 | DEF_FACTORY_TWO_TYPE_RESIZE_INT16(OpResize, INT16, INT48); |
| 447 | DEF_FACTORY_TWO_TYPE_RESIZE_INT16(OpResize, INT16, INT16); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 448 | DEF_FACTORY_TWO_TYPE_RESIZE_FP16(OpResize, FP16, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 449 | DEF_FACTORY_TWO_TYPE_RESIZE_BF16(OpResize, BF16, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 450 | DEF_FACTORY_TWO_TYPE_RESIZE_FP32(OpResize, FP32, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 451 | break; |
| 452 | |
| 453 | // data_nodes |
| 454 | case Op_CONST: |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 455 | return new OpConst(sgt, id); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 456 | case Op_IDENTITY: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 457 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIdentity, FP16); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 458 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIdentity, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 459 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIdentity, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 460 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIdentity, INT32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 461 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIdentity, INT8); |
| 462 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIdentity, INT16); |
| 463 | DEF_FACTORY_RANK0_6_ONE_RANK_ONE_TYPE(OpIdentity, BOOL); |
| 464 | break; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 465 | |
| 466 | // type_conversion |
| 467 | case Op_CAST: |
| 468 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BOOL, INT8); |
| 469 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BOOL, INT16); |
| 470 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BOOL, INT32); |
| 471 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, BOOL); |
| 472 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, INT16); |
| 473 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, INT32); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 474 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, FP16); |
James Ward | 736fd1a | 2023-01-23 17:13:37 +0000 | [diff] [blame] | 475 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 476 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 477 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, BOOL); |
| 478 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, INT8); |
| 479 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, INT32); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 480 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, FP16); |
James Ward | 736fd1a | 2023-01-23 17:13:37 +0000 | [diff] [blame] | 481 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 482 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, FP32); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 483 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, BOOL); |
| 484 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, INT8); |
| 485 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, INT16); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 486 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, FP16); |
James Ward | 736fd1a | 2023-01-23 17:13:37 +0000 | [diff] [blame] | 487 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, BF16); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 488 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, FP32); |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 489 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT8); |
| 490 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT16); |
| 491 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT32); |
James Ward | 736fd1a | 2023-01-23 17:13:37 +0000 | [diff] [blame] | 492 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, FP32); |
James Ward | 24dbc42 | 2022-10-19 12:20:31 +0100 | [diff] [blame] | 493 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT8); |
| 494 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT16); |
| 495 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT32); |
James Ward | 736fd1a | 2023-01-23 17:13:37 +0000 | [diff] [blame] | 496 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, FP32); |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 497 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT8); |
| 498 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT16); |
| 499 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT32); |
James Ward | 736fd1a | 2023-01-23 17:13:37 +0000 | [diff] [blame] | 500 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, FP16); |
| 501 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, BF16); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 502 | break; |
| 503 | case Op_RESCALE: |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 504 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT8, INT8); |
| 505 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT8, INT16); |
| 506 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT8, INT32); |
| 507 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT16, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 508 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT16, INT16); |
| 509 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT16, INT32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 510 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT32, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 511 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT32, INT16); |
| 512 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT32, INT32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 513 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT48, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 514 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT48, INT16); |
| 515 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT48, INT32); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 516 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, UINT8, INT8); |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 517 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, UINT8, INT16); |
| 518 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, UINT16, INT16); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 519 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT8, UINT8); |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 520 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT16, UINT8); |
| 521 | DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT16, UINT16); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 522 | break; |
| 523 | |
| 524 | // custom |
| 525 | case Op_CUSTOM: |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 526 | return new OpCustom(sgt, id); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 527 | |
| 528 | // control_flow |
| 529 | case Op_COND_IF: |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 530 | return new OpCondIf(sgt, tsh, attribute, id); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 531 | case Op_WHILE_LOOP: |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 532 | return new OpWhileLoop(sgt, tsh, attribute, id); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 533 | |
| 534 | // Ops not recognized |
| 535 | default: |
| 536 | goto done; |
| 537 | |
| 538 | } // End of switch(opType) |
| 539 | |
| 540 | done: |
| 541 | return nullptr; |
| 542 | } |