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