Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 1 | /* |
Renato Arantes | 36a75da | 2024-01-26 17:31:18 +0000 | [diff] [blame] | 2 | * Copyright (c) 2023-2024 Arm Limited. |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | |
| 25 | #include "src/cpu/operators/CpuMatMul.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 26 | |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 27 | #include "arm_compute/core/experimental/Types.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/Types.h" |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 29 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 31 | #include "arm_compute/core/Validate.h" |
SiCong Li | 9129549 | 2023-07-21 18:16:13 +0100 | [diff] [blame] | 32 | #include "arm_compute/function_info/MatMulInfo.h" |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 33 | #include "arm_compute/runtime/NEON/functions/NEMatMul.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 34 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 35 | |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 36 | #include "src/common/utils/Log.h" |
| 37 | #include "src/core/CPP/Validate.h" |
| 38 | #include "src/core/helpers/AutoConfiguration.h" |
| 39 | #include "src/core/helpers/MemoryHelpers.h" |
Viet-Hoa Do | a62129a | 2023-04-26 15:38:45 +0100 | [diff] [blame] | 40 | #include "src/core/utils/quantization/AsymmHelpers.h" |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 41 | #include "src/cpu/utils/CpuAuxTensorHandler.h" |
| 42 | |
| 43 | using namespace arm_compute::experimental; |
| 44 | |
| 45 | namespace arm_compute |
| 46 | { |
| 47 | namespace cpu |
| 48 | { |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 49 | namespace |
| 50 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 51 | Status get_gemmlowp_output_stage_info(const ITensorInfo *src, |
| 52 | const ITensorInfo *weights, |
| 53 | const ITensorInfo *dst, |
| 54 | const ActivationLayerInfo &act, |
| 55 | GEMMLowpOutputStageInfo &gemmlowp_output_stage_info) |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 56 | { |
| 57 | const auto data_type = src->data_type(); |
| 58 | const QuantizationInfo oq_info = dst->quantization_info(); |
| 59 | const UniformQuantizationInfo iq_unif = src->quantization_info().uniform(); |
| 60 | const UniformQuantizationInfo wq_unif = weights->quantization_info().uniform(); |
| 61 | const UniformQuantizationInfo oq_unif = oq_info.uniform(); |
| 62 | |
| 63 | float multiplier = (iq_unif.scale * wq_unif.scale) / oq_unif.scale; |
| 64 | int32_t output_multiplier; |
| 65 | int32_t output_shift; |
| 66 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 67 | ARM_COMPUTE_RETURN_ON_ERROR( |
| 68 | quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift)); |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 69 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 70 | int32_t type_min = 0; |
| 71 | int32_t type_max = 0; |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 72 | std::tie(type_min, type_max) = quantization::get_quantized_asymmetric_output_min_max(oq_info, act, data_type); |
| 73 | |
| 74 | gemmlowp_output_stage_info.gemmlowp_multiplier = output_multiplier; |
| 75 | gemmlowp_output_stage_info.gemmlowp_shift = output_shift; |
| 76 | gemmlowp_output_stage_info.gemmlowp_offset = oq_unif.offset; |
| 77 | gemmlowp_output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; |
Viet-Hoa Do | a62129a | 2023-04-26 15:38:45 +0100 | [diff] [blame] | 78 | gemmlowp_output_stage_info.gemmlowp_min_bound = type_min; |
| 79 | gemmlowp_output_stage_info.gemmlowp_max_bound = type_max; |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 80 | |
| 81 | return Status{}; |
| 82 | } |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 83 | } // namespace |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 84 | |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 85 | CpuMatMul::CpuMatMul() |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 86 | : _transpose_kernel_lhs(), |
| 87 | _transpose_kernel_rhs(), |
| 88 | _asm_glue(), |
| 89 | _lhs_transposed(), |
| 90 | _rhs_transposed(), |
| 91 | _original_lhs_shape(), |
| 92 | _original_rhs_shape(), |
| 93 | _original_dst_shape() |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 94 | { |
| 95 | } |
| 96 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 97 | Status CpuMatMul::validate(const ITensorInfo *lhs, |
| 98 | const ITensorInfo *rhs, |
| 99 | const ITensorInfo *dst, |
| 100 | const MatMulInfo &info, |
| 101 | const CpuMatMulSettings &settings, |
| 102 | const ActivationLayerInfo &act_info) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 103 | { |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 104 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, rhs, dst); |
Renato Arantes | 36a75da | 2024-01-26 17:31:18 +0000 | [diff] [blame] | 105 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::F32, DataType::F16, DataType::BFLOAT16, |
| 106 | DataType::QASYMM8, DataType::QASYMM8_SIGNED); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 107 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs->are_values_constant(), "LHS Tensor must be dynamic."); |
| 108 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs->are_values_constant(), "RHS Tensor must be dynamic."); |
| 109 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(lhs); |
| 110 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(lhs); |
| 111 | |
| 112 | const auto adj_lhs = info.adj_lhs(); |
| 113 | const auto adj_rhs = info.adj_rhs(); |
| 114 | |
| 115 | const ITensorInfo *lhs_to_use = lhs; |
| 116 | const ITensorInfo *rhs_to_use = rhs; |
| 117 | TensorInfo lhs_transposed{}; |
| 118 | TensorInfo rhs_transposed{}; |
| 119 | |
| 120 | auto gemm_info = AsmGemmInfo(); |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 121 | gemm_info.activation_info = act_info; |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 122 | gemm_info.fast_mode = settings.fast_math(); |
Renato Arantes | 36a75da | 2024-01-26 17:31:18 +0000 | [diff] [blame] | 123 | gemm_info.fixed_format = settings.fixed_format(); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 124 | |
| 125 | // Validate and then permute a/b |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 126 | if (adj_lhs) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 127 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 128 | auto_init_if_empty(lhs_transposed, |
| 129 | lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_transposed_shape(*lhs))); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 130 | ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuTransposeKernel::validate(lhs_to_use, &lhs_transposed)); |
| 131 | // Assign lhs_to_use pointer to use transposed TensorInfo |
| 132 | lhs_to_use = &lhs_transposed; |
| 133 | } |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 134 | if (adj_rhs) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 135 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 136 | auto_init_if_empty(rhs_transposed, |
| 137 | rhs->clone()->set_tensor_shape(misc::shape_calculator::compute_transposed_shape(*rhs))); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 138 | ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuTransposeKernel::validate(rhs_to_use, &rhs_transposed)); |
| 139 | // Assign rhs_to_use pointer to use transposed TensorInfo |
| 140 | rhs_to_use = &rhs_transposed; |
| 141 | } |
| 142 | |
| 143 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_to_use->dimension(0) != rhs_to_use->dimension(1), |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 144 | "The product AB is defined only if the number of columns in A is equal to the " |
| 145 | "number of rows in B (after transpose)"); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 146 | |
Viet-Hoa Do | 54e52a9 | 2023-05-02 16:20:58 +0100 | [diff] [blame] | 147 | // Iterate over dimensions to be collapsed in operator - check dimensions are equivalent between tensors |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 148 | for (unsigned int i = 2; i < Coordinates::num_max_dimensions; i++) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 149 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 150 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_to_use->dimension(i) != rhs_to_use->dimension(i), |
| 151 | "Broadcasting in Batch dimension is unsupported by this operator."); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 152 | } |
| 153 | |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 154 | // Quantized-specific configuration |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 155 | if (is_data_type_quantized(lhs->data_type())) |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 156 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 157 | ARM_COMPUTE_RETURN_ON_ERROR(get_gemmlowp_output_stage_info(lhs_to_use, rhs_to_use, dst, |
| 158 | gemm_info.activation_info, gemm_info.output_stage)); |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 159 | } |
| 160 | |
Renato Arantes | 36a75da | 2024-01-26 17:31:18 +0000 | [diff] [blame] | 161 | if (gemm_info.fixed_format) |
| 162 | { |
| 163 | gemm_info.weight_format = WeightFormat::ANY; |
| 164 | arm_compute::WeightFormat expected_weight_format = WeightFormat::ANY; |
| 165 | ARM_COMPUTE_RETURN_ON_ERROR(cpu::CpuGemmAssemblyDispatch::has_opt_impl(expected_weight_format, lhs_to_use, |
| 166 | rhs_to_use, nullptr, dst, gemm_info)); |
| 167 | } |
| 168 | |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 169 | cpu::CpuGemmAssemblyDispatch::validate(lhs_to_use, rhs_to_use, nullptr, dst, gemm_info); |
| 170 | |
| 171 | return Status{}; |
| 172 | } |
| 173 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 174 | void CpuMatMul::configure(ITensorInfo *lhs, |
| 175 | ITensorInfo *rhs, |
| 176 | ITensorInfo *dst, |
| 177 | const MatMulInfo &info, |
| 178 | const CpuMatMulSettings &settings, |
| 179 | const ActivationLayerInfo &act_info) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 180 | { |
| 181 | ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst); |
| 182 | ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst, info, settings); |
| 183 | ARM_COMPUTE_ERROR_THROW_ON(CpuMatMul::validate(lhs, rhs, dst, info, settings)); |
| 184 | |
| 185 | _adj_lhs = info.adj_lhs(); |
| 186 | _adj_rhs = info.adj_rhs(); |
| 187 | _fast_math = settings.fast_math(); |
| 188 | |
| 189 | // 1. Create and reshape tensors |
| 190 | // ------------------------------------------------------ |
| 191 | // a. Clone TensorInfo to prevent changing original tensor values during setup |
| 192 | // b. Change shape of lhs/dst to [x, y, 1, collapsed(z)] to match assembly kernel configuration |
| 193 | // c. For rhs collapse all dimensions larger than 3 to z dimension |
| 194 | TensorInfo lhs_to_use = *lhs->clone(); |
| 195 | TensorInfo dst_to_use = *dst->clone(); |
| 196 | TensorInfo rhs_to_use = *rhs->clone(); |
| 197 | |
| 198 | // Save starting shape of tensors |
| 199 | _original_lhs_shape = lhs_to_use.tensor_shape(); |
| 200 | _original_dst_shape = dst_to_use.tensor_shape(); |
| 201 | _original_rhs_shape = rhs_to_use.tensor_shape(); |
| 202 | |
| 203 | // Reshape lhs for use with assembly kernels. |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 204 | lhs_to_use.set_tensor_shape( |
| 205 | TensorShape(_original_lhs_shape.x(), _original_lhs_shape.y(), 1, _original_lhs_shape.collapsed_from(2).z())); |
| 206 | dst_to_use.set_tensor_shape( |
| 207 | TensorShape(_original_dst_shape.x(), _original_dst_shape.y(), 1, _original_dst_shape.collapsed_from(2).z())); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 208 | rhs_to_use.set_tensor_shape(_original_rhs_shape.collapsed_from(2)); |
| 209 | |
| 210 | // 2. Configuration for transpose of lhs/rhs |
| 211 | // ------------------------------------------------------ |
| 212 | // Initialise transposed TensorInfo class for aux tensors (intermediary tensors) |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 213 | if (_adj_lhs) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 214 | { |
| 215 | // Setup transpose LHS |
| 216 | _transpose_kernel_lhs = std::make_unique<cpu::kernels::CpuTransposeKernel>(); |
| 217 | _transpose_kernel_lhs->configure(&lhs_to_use, &_lhs_transposed); |
| 218 | } |
| 219 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 220 | if (_adj_rhs) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 221 | { |
| 222 | // Setup transpose RHS |
| 223 | _transpose_kernel_rhs = std::make_unique<cpu::kernels::CpuTransposeKernel>(); |
| 224 | _transpose_kernel_rhs->configure(&rhs_to_use, &_rhs_transposed); |
| 225 | } |
| 226 | |
| 227 | // 3. Configure assembly kernel using transposed tensors. |
| 228 | // ----------------------------------------------------- |
| 229 | // Use transposed tensors if the corresponding transpose flags are set |
| 230 | // Fill AsmGemmInfo class object before configuration |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 231 | _gemm_info.activation_info = act_info; |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 232 | _gemm_info.fast_mode = settings.fast_math(); |
Renato Arantes | 36a75da | 2024-01-26 17:31:18 +0000 | [diff] [blame] | 233 | _gemm_info.fixed_format = settings.fixed_format(); |
Jakub Sujak | e9b3ee2 | 2023-04-17 12:08:48 +0100 | [diff] [blame] | 234 | _gemm_info.negated_offsets = false; |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 235 | |
| 236 | lhs_to_use = (_adj_lhs) ? _lhs_transposed : lhs_to_use; |
| 237 | rhs_to_use = (_adj_rhs) ? _rhs_transposed : rhs_to_use; |
| 238 | |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 239 | // Quantized-specific configuration |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 240 | if (is_data_type_quantized(lhs->data_type())) |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 241 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 242 | get_gemmlowp_output_stage_info(&lhs_to_use, &rhs_to_use, &dst_to_use, _gemm_info.activation_info, |
| 243 | _gemm_info.output_stage); |
Viet-Hoa Do | 9c7c2d2 | 2023-04-11 17:16:27 +0100 | [diff] [blame] | 244 | } |
| 245 | |
Renato Arantes | 36a75da | 2024-01-26 17:31:18 +0000 | [diff] [blame] | 246 | if (_gemm_info.fixed_format) |
| 247 | { |
| 248 | _gemm_info.weight_format = WeightFormat::ANY; |
| 249 | arm_compute::WeightFormat expected_weight_format = WeightFormat::ANY; |
| 250 | ARM_COMPUTE_ERROR_THROW_ON(cpu::CpuGemmAssemblyDispatch::has_opt_impl(expected_weight_format, &lhs_to_use, |
| 251 | &rhs_to_use, nullptr, dst, _gemm_info)); |
| 252 | // Set gemm weights info to the one returned by has_opt_impl |
| 253 | _gemm_info.weight_format = expected_weight_format; |
| 254 | // has_opt_impl may return a non fast math kernel, even if we requested one |
| 255 | _gemm_info.fast_mode = arm_compute::is_fixed_format_fast_math(expected_weight_format); |
| 256 | } |
| 257 | |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 258 | // Configure Asm Kernel |
| 259 | _asm_glue = std::make_unique<cpu::CpuGemmAssemblyDispatch>(); |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 260 | _asm_glue->configure(&lhs_to_use, &rhs_to_use, nullptr, &dst_to_use, |
| 261 | _gemm_info); // c is nullptr as bias not supported in MatMul |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 262 | |
| 263 | // Specify memory requirements for intermediate tensors |
| 264 | auto asm_mem_req = _asm_glue->workspace(); |
| 265 | // Specify memory required by gemm kernel |
| 266 | int idx = 0; |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 267 | for (const auto &aux : asm_mem_req) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 268 | { |
| 269 | _aux_mem[idx] = aux; |
| 270 | idx++; |
| 271 | } |
| 272 | // Memory requirements for transposed tensors |
| 273 | _aux_mem[TransposeLHS] = MemoryInfo(offset_int_vec(TransposeLHS), MemoryLifetime::Temporary, lhs->total_size()); |
| 274 | _aux_mem[TransposeRHS] = MemoryInfo(offset_int_vec(TransposeRHS), MemoryLifetime::Temporary, rhs->total_size()); |
| 275 | } |
| 276 | |
| 277 | void CpuMatMul::run(ITensorPack &tensors) |
| 278 | { |
| 279 | // Retrieve tensors from tensor pack |
| 280 | auto lhs = tensors.get_tensor(ACL_SRC_0); |
| 281 | auto rhs = tensors.get_const_tensor(ACL_SRC_1); |
| 282 | auto dst = tensors.get_tensor(ACL_DST); |
| 283 | |
| 284 | // Reshape LHS and DST to ensure compatibility with GEMM asm kernel (Batch dimensions is 4th for lhs and dst within asm) |
| 285 | // Collapse RHS (necessary to support dimensions larger than 3 in gemm assembly) |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 286 | lhs->info()->set_tensor_shape( |
| 287 | TensorShape(_original_lhs_shape.x(), _original_lhs_shape.y(), 1, |
| 288 | _original_lhs_shape.collapsed_from(2).z())); // Collapsed 3+ dimensions into z |
| 289 | dst->info()->set_tensor_shape( |
| 290 | TensorShape(_original_dst_shape.x(), _original_dst_shape.y(), 1, |
| 291 | _original_dst_shape.collapsed_from(2).z())); // Collapsed 3+ dimensions into z |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 292 | rhs->info()->set_tensor_shape(_original_rhs_shape.collapsed_from(2)); |
| 293 | |
| 294 | // Initialise object to handle stored transposed tensors in auxillary memory |
| 295 | CpuAuxTensorHandler lhs_transposed(offset_int_vec(TransposeLHS), _lhs_transposed, tensors, true); |
| 296 | CpuAuxTensorHandler rhs_transposed(offset_int_vec(TransposeRHS), _rhs_transposed, tensors, true); |
| 297 | |
| 298 | // Create tensor pack for asm kernel |
| 299 | ITensorPack asm_tensors(tensors); |
| 300 | |
| 301 | // Run transpose lhs if necessary |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 302 | if (_adj_lhs) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 303 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 304 | ITensorPack lhs_transpose_pack = {{TensorType::ACL_SRC, lhs}, {TensorType::ACL_DST, lhs_transposed.get()}}; |
| 305 | NEScheduler::get().schedule_op(_transpose_kernel_lhs.get(), Window::DimY, _transpose_kernel_lhs->window(), |
| 306 | lhs_transpose_pack); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 307 | asm_tensors.add_const_tensor(TensorType::ACL_SRC_0, lhs_transposed.get()); |
| 308 | } |
| 309 | // Run transpose rhs if necessary |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 310 | if (_adj_rhs) |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 311 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 312 | ITensorPack rhs_transpose_pack = {{TensorType::ACL_SRC, rhs}, {TensorType::ACL_DST, rhs_transposed.get()}}; |
| 313 | NEScheduler::get().schedule_op(_transpose_kernel_rhs.get(), Window::DimY, _transpose_kernel_rhs->window(), |
| 314 | rhs_transpose_pack); |
Mohammed Suhail Munshi | a1b1e41 | 2023-03-23 22:21:31 +0000 | [diff] [blame] | 315 | asm_tensors.add_const_tensor(TensorType::ACL_SRC_1, rhs_transposed.get()); |
| 316 | } |
| 317 | // Run asm kernel |
| 318 | _asm_glue->run(asm_tensors); |
| 319 | |
| 320 | // Undo reshape of tensors |
| 321 | dst->info()->set_tensor_shape(_original_dst_shape); |
| 322 | lhs->info()->set_tensor_shape(_original_lhs_shape); |
| 323 | rhs->info()->set_tensor_shape(_original_rhs_shape); |
| 324 | } |
| 325 | |
| 326 | experimental::MemoryRequirements CpuMatMul::workspace() const |
| 327 | { |
| 328 | return _aux_mem; |
| 329 | } |
| 330 | } // namespace cpu |
| 331 | } // namespace arm_compute |