Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2023 Arm Limited. |
| 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 | #include "src/gpu/cl/kernels/ClMatMulLowpNativeMMULKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLHelpers.h" |
| 27 | #include "arm_compute/core/CL/ICLTensor.h" |
| 28 | #include "arm_compute/core/Error.h" |
| 29 | #include "arm_compute/core/ITensorPack.h" |
| 30 | #include "arm_compute/core/TensorInfo.h" |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 31 | #include "arm_compute/core/utils/helpers/AdjustVecSize.h" |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 33 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 34 | #include "arm_compute/core/utils/StringUtils.h" |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 35 | |
| 36 | #include "src/common/utils/Log.h" |
| 37 | #include "src/core/helpers/AutoConfiguration.h" |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 38 | #include "src/gpu/cl/ClCompileContext.h" |
| 39 | #include "src/gpu/cl/kernels/helpers/MatMulKernelHelpers.h" |
| 40 | #include "support/Cast.h" |
| 41 | #include "support/StringSupport.h" |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 42 | #include "utils/TypePrinter.h" |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 43 | |
| 44 | namespace arm_compute |
| 45 | { |
| 46 | namespace opencl |
| 47 | { |
| 48 | namespace kernels |
| 49 | { |
| 50 | namespace |
| 51 | { |
| 52 | // Block size dimensions for the MMUL extension |
| 53 | constexpr int mmul_m0 = 4; |
| 54 | constexpr int mmul_n0 = 4; |
| 55 | constexpr int mmul_k0 = 16; |
| 56 | |
| 57 | Status validate_matmul_kernel_info(const MatMulKernelInfo &matmul_kernel_info) |
| 58 | { |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 59 | const bool adj_lhs = matmul_kernel_info.adj_lhs; |
| 60 | const int m0 = matmul_kernel_info.m0; |
| 61 | const int n0 = matmul_kernel_info.n0; |
| 62 | const int k0 = matmul_kernel_info.k0; |
| 63 | |
| 64 | // Validate M0 |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(m0 < 1, "Only positive integers are supported for M0"); |
| 66 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 67 | if (adj_lhs) |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 68 | { |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 69 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((m0 != 1) && (m0 != 2) && (m0 != 3) && (m0 != 4) && (m0 != 8) && (m0 != 16), |
| 70 | "Only 1,2,3,4,8,16 are supported for M0 for Lhs transposed"); |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 71 | } |
| 72 | |
| 73 | // Validate N0 |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 74 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((n0 != 1) && (n0 != 2) && (n0 != 3) && (n0 != 4) && (n0 != 8) && (n0 != 16), |
| 75 | "Only 1,2,3,4,8,16 are supported for N0"); |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 76 | |
| 77 | // Validate K0 |
| 78 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((k0 != 4), "Only 4 is supported for k0"); |
| 79 | |
Gunes Bayir | 2ad0a6b | 2023-09-19 15:37:38 +0100 | [diff] [blame] | 80 | // Validate ExportToCLImage |
| 81 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(matmul_kernel_info.export_rhs_to_cl_image, "Export to CLImage is not supported!"); |
| 82 | |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 83 | return Status{}; |
| 84 | } |
| 85 | } // namespace |
| 86 | |
| 87 | ClMatMulLowpNativeMMULKernel::ClMatMulLowpNativeMMULKernel() |
| 88 | { |
| 89 | _type = CLKernelType::GEMM; |
| 90 | } |
| 91 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 92 | Status ClMatMulLowpNativeMMULKernel::validate(const ITensorInfo *lhs, |
| 93 | const ITensorInfo *rhs, |
| 94 | const ITensorInfo *bias, |
| 95 | const ITensorInfo *dst, |
| 96 | const MatMulKernelInfo &matmul_kernel_info, |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 97 | const ActivationLayerInfo &act_info) |
| 98 | { |
| 99 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs, dst); |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 100 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(!arm_matrix_multiply_supported(CLKernelLibrary::get().get_device()), |
| 101 | "The extension cl_arm_matrix_multiply is not supported on the target platform"); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 102 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); |
| 103 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, rhs); |
| 104 | ARM_COMPUTE_RETURN_ON_ERROR(validate_matmul_kernel_info(matmul_kernel_info)); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 105 | |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 106 | const TensorShape &lhs_shape = lhs->tensor_shape(); |
| 107 | ARM_COMPUTE_RETURN_ON_ERROR(validate_matmul_input_shapes(lhs_shape, rhs->tensor_shape(), matmul_kernel_info)); |
| 108 | |
| 109 | const size_t lhs_k = matmul_kernel_info.adj_lhs ? lhs_shape.y() : lhs_shape.x(); |
| 110 | ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR((lhs_k % mmul_k0) != 0, "K dimension must be a multiple of %d", mmul_k0); |
| 111 | |
| 112 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.activation() != ActivationFunction::IDENTITY), |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 113 | "Activation Function specified is unsupported."); |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 114 | const TensorShape expected_output_shape = |
| 115 | misc::shape_calculator::compute_matmul_shape(lhs_shape, rhs->tensor_shape(), matmul_kernel_info); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 116 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 117 | if (dst->total_size() != 0) |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 118 | { |
| 119 | const TensorInfo tensor_info_output = dst->clone()->set_tensor_shape(expected_output_shape); |
| 120 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_output); |
| 121 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, dst); |
| 122 | } |
| 123 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 124 | if (bias != nullptr) |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 125 | { |
| 126 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); |
| 127 | ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); |
| 128 | ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != bias->dimension(0)); |
| 129 | } |
| 130 | |
| 131 | return Status{}; |
| 132 | } |
| 133 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 134 | void ClMatMulLowpNativeMMULKernel::configure(const ClCompileContext &compile_context, |
| 135 | ITensorInfo *lhs, |
| 136 | ITensorInfo *rhs, |
| 137 | ITensorInfo *bias, |
| 138 | ITensorInfo *dst, |
| 139 | const MatMulKernelInfo &matmul_kernel_info, |
| 140 | const ActivationLayerInfo &act_info) |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 141 | { |
| 142 | ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst); |
| 143 | ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst, matmul_kernel_info, act_info); |
| 144 | ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, bias, dst, matmul_kernel_info)); |
| 145 | |
| 146 | // dst tensor auto initialization if not yet initialized |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 147 | auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape( |
| 148 | lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info))); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 149 | |
| 150 | ARM_COMPUTE_UNUSED(compile_context, lhs, rhs, bias, matmul_kernel_info, act_info); |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 151 | CLBuildOptions build_opts; |
| 152 | |
| 153 | const int m = dst->dimension(1); |
| 154 | const int n = dst->dimension(0); |
| 155 | const int k = matmul_kernel_info.adj_lhs ? lhs->tensor_shape().y() : lhs->tensor_shape().x(); |
| 156 | |
| 157 | const int m0 = std::min(matmul_kernel_info.m0, m); |
| 158 | const int n0 = adjust_vec_size(matmul_kernel_info.n0, n); |
| 159 | |
| 160 | // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks |
| 161 | // at the end of a row/column if any. This is to avoid padding. |
| 162 | const unsigned int m0_leftover = m % m0; |
| 163 | const unsigned int n0_leftover = n % n0; |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 164 | |
| 165 | // Configure kernel window |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 166 | const auto win_config = |
| 167 | validate_and_configure_window_for_mmul_kernels(lhs, rhs, dst, matmul_kernel_info, mmul_m0, mmul_n0); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 168 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 169 | IClKernel::configure_internal(win_config.second); |
| 170 | |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 171 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(lhs->data_type())); |
| 172 | build_opts.add_option("-DM=" + support::cpp11::to_string(m)); |
| 173 | build_opts.add_option("-DN=" + support::cpp11::to_string(n)); |
| 174 | build_opts.add_option("-DK=" + support::cpp11::to_string(k)); |
| 175 | build_opts.add_option("-DM0=" + support::cpp11::to_string(m0)); |
| 176 | build_opts.add_option("-DN0=" + support::cpp11::to_string(n0)); |
| 177 | build_opts.add_option("-DK0=" + support::cpp11::to_string(matmul_kernel_info.k0)); |
| 178 | build_opts.add_option("-DM0_LEFTOVER=" + support::cpp11::to_string(m0_leftover)); |
| 179 | build_opts.add_option("-DN0_LEFTOVER=" + support::cpp11::to_string(n0_leftover)); |
| 180 | build_opts.add_option("-DMMUL_M0=" + support::cpp11::to_string(mmul_m0)); |
| 181 | build_opts.add_option("-DMMUL_N0=" + support::cpp11::to_string(mmul_n0)); |
| 182 | build_opts.add_option("-DMMUL_K0=" + support::cpp11::to_string(mmul_k0)); |
| 183 | build_opts.add_option_if(bias != nullptr, "-DBIAS"); |
| 184 | |
| 185 | const UniformQuantizationInfo lqinfo = lhs->quantization_info().uniform(); |
| 186 | const UniformQuantizationInfo rqinfo = rhs->quantization_info().uniform(); |
| 187 | const UniformQuantizationInfo dqinfo = dst->quantization_info().uniform(); |
| 188 | |
| 189 | float multiplier = lqinfo.scale * rqinfo.scale / dqinfo.scale; |
| 190 | int output_multiplier = 0; |
| 191 | int output_shift = 0; |
| 192 | arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift); |
| 193 | |
| 194 | build_opts.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier)); |
| 195 | build_opts.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift)); |
| 196 | |
| 197 | // Note : Offset is not negated, unlike gemmlowp kernels |
| 198 | build_opts.add_option("-DLHS_OFFSET=" + support::cpp11::to_string(lqinfo.offset)); |
| 199 | build_opts.add_option("-DRHS_OFFSET=" + support::cpp11::to_string(rqinfo.offset)); |
| 200 | build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(dqinfo.offset)); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 201 | |
| 202 | std::string kernel_name("mat_mul_native_quantized_mmul"); |
| 203 | kernel_name += matmul_kernel_info.adj_lhs ? "_t" : "_nt"; |
| 204 | kernel_name += matmul_kernel_info.adj_rhs ? "_t" : "_nt"; |
| 205 | |
| 206 | // A macro guard to compile ONLY the kernel of interest |
| 207 | build_opts.add_option("-D" + upper_string(kernel_name)); |
| 208 | |
| 209 | // Create kernel |
| 210 | _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); |
| 211 | |
Gunes Bayir | a116cd3 | 2023-09-13 11:59:34 +0100 | [diff] [blame] | 212 | // Set config_id for enabling LWS tuning |
| 213 | _config_id = kernel_name; |
| 214 | _config_id += "_"; |
| 215 | _config_id += lower_string(string_from_data_type(lhs->data_type())); |
| 216 | _config_id += "_"; |
| 217 | _config_id += support::cpp11::to_string(m); |
| 218 | _config_id += "_"; |
| 219 | _config_id += support::cpp11::to_string(n); |
| 220 | _config_id += "_"; |
| 221 | _config_id += support::cpp11::to_string(k); |
| 222 | _config_id += "_"; |
| 223 | _config_id += support::cpp11::to_string(dst->dimension(2)); |
| 224 | _config_id += "_"; |
| 225 | _config_id += support::cpp11::to_string(m0); |
| 226 | _config_id += "_"; |
| 227 | _config_id += support::cpp11::to_string(n0); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 228 | } |
| 229 | |
| 230 | void ClMatMulLowpNativeMMULKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) |
| 231 | { |
| 232 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 233 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 234 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 235 | const auto *lhs = |
| 236 | utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); |
| 237 | const auto *rhs = |
| 238 | utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); |
| 239 | const auto *bias = utils::cast::polymorphic_downcast<const ICLTensor *>( |
| 240 | tensors.get_const_tensor(TensorType::ACL_SRC_2)); // nullptr if bias is not present |
| 241 | auto *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 242 | |
| 243 | ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst); |
| 244 | ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst); |
| 245 | |
| 246 | unsigned int idx = 0; |
| 247 | add_3d_tensor_nhw_argument(idx, lhs); |
| 248 | add_3d_tensor_nhw_argument(idx, rhs); |
| 249 | |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 250 | if (bias != nullptr) |
Gunes Bayir | e87fa66 | 2023-09-07 12:20:33 +0100 | [diff] [blame] | 251 | { |
| 252 | add_3d_tensor_nhw_argument(idx, bias); |
| 253 | } |
| 254 | add_3d_tensor_nhw_argument(idx, dst); |
| 255 | |
| 256 | // LWS_x should be multiple of 16 at least. (32, 2) has been chosen to have more work-items on a single core |
| 257 | // LWS also enforces the order of execution of the work items which improves cache utilization |
| 258 | enqueue(queue, *this, window, cl::NDRange(32, 2), false); |
| 259 | } |
| 260 | |
| 261 | } // namespace kernels |
| 262 | } // namespace opencl |
| 263 | } // namespace arm_compute |