1 /*
2 * Copyright (c) 2017-2022 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/operators/ClGemmLowpMatrixMultiplyCore.h"
25
26 #include "arm_compute/core/Log.h"
27 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
28
29 #include "src/core/helpers/AutoConfiguration.h"
30 #include "src/core/helpers/MemoryHelpers.h"
31 #include "src/gpu/cl/kernels/ClCastKernel.h"
32 #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.h"
33 #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.h"
34 #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel.h"
35 #include "src/gpu/cl/kernels/ClGemmLowpOffsetContributionKernel.h"
36 #include "src/gpu/cl/kernels/ClGemmLowpOffsetContributionOutputStageKernel.h"
37 #include "src/gpu/cl/kernels/ClGemmLowpReductionKernel.h"
38 #include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h"
39 #include "src/gpu/cl/utils/ClAuxTensorHandler.h"
40 #include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h"
41
42 namespace arm_compute
43 {
44 namespace opencl
45 {
46 using namespace arm_compute::misc::shape_calculator;
47 using namespace arm_compute::cl_gemm;
48 using namespace arm_compute::opencl::kernels;
49 using namespace arm_compute::experimental;
50
51 namespace
52 {
validate_gemm_kernel(CLGEMMKernelType kernel_type)53 inline bool validate_gemm_kernel(CLGEMMKernelType kernel_type)
54 {
55 switch(kernel_type)
56 {
57 case CLGEMMKernelType::NATIVE:
58 case CLGEMMKernelType::RESHAPED_ONLY_RHS:
59 case CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL:
60 {
61 return true;
62 }
63 default:
64 {
65 return false;
66 }
67 }
68 }
69
70 //Automatically select between mlgo (prioritized) and default heuristics for gemm kernel type
auto_select_gemm_kernel(auto_heuristics::CommonQuery query,bool reshape_b_only_on_first_run)71 inline CLGEMMKernelType auto_select_gemm_kernel(auto_heuristics::CommonQuery query, bool reshape_b_only_on_first_run)
72 {
73 auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run);
74 if(bool(gemm_kernel))
75 {
76 if(validate_gemm_kernel(gemm_kernel.gemm_type))
77 {
78 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from mlgo heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str());
79 return gemm_kernel.gemm_type;
80 }
81 }
82 gemm_kernel = auto_heuristics::select_default_gemm_kernel(query, reshape_b_only_on_first_run);
83 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from default heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str());
84 return gemm_kernel.gemm_type;
85 }
86
87 // Validate lhs_info and rhs_info for native kernel
validate_lhs_rhs_info_native(const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const ITensorInfo * a,const ITensorInfo * b,const GEMMReshapeInfo & reshape_info)88 inline bool validate_lhs_rhs_info_native(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const GEMMReshapeInfo &reshape_info)
89 {
90 // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel
91 TensorInfo mm_result_s32_info{};
92 // Output tensor auto initialization if not yet initialized
93 auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*a, *b, false, reshape_info)).set_data_type(DataType::S32));
94 // Validate mm kernel
95 // NOTE: Ignore all other parameters (eg. output stage etc.) and only validate lhs and rhs info
96 // NOTE: This assumes:
97 // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in CLGEMMLowpMatrixMultiplyNativeKernel.cpp validate_arguments).
98 // 2. lhs and rhs info does not cause window and padding issues through side effects (in CLGEMMLowpMatrixMultiplyNativeKernel.cpp validate_and_configure_window).
99 if(!bool(ClGemmLowpMatrixMultiplyNativeKernel::validate(a, b, &mm_result_s32_info, lhs_info, rhs_info, reshape_info)))
100 {
101 return false;
102 }
103 return true;
104 }
105
106 // Automatically select between mlgo (prioritized) and default heuristics for native kernel configs
auto_select_gemm_config_native(auto_heuristics::CommonQuery query,const ITensorInfo * a,const ITensorInfo * b,const GEMMReshapeInfo & reshape_info)107 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_native(auto_heuristics::CommonQuery query, const ITensorInfo *a, const ITensorInfo *b, const GEMMReshapeInfo &reshape_info)
108 {
109 auto config = auto_heuristics::select_mlgo_gemm_config_native(query);
110 if(config)
111 {
112 if(validate_lhs_rhs_info_native(config.lhs_info, config.rhs_info, a, b, reshape_info))
113 {
114 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use native config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
115 return { config.lhs_info, config.rhs_info };
116 }
117 }
118 config = auto_heuristics::select_default_gemm_config_native(query);
119 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use native config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
120 return { config.lhs_info, config.rhs_info };
121 }
122
123 // Validate lhs_info and rhs_info for reshaped only rhs kernel
validate_lhs_rhs_info_reshaped_only_rhs(const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const ITensorInfo * a,const ITensorInfo * b,const ITensorInfo * output,unsigned int m,unsigned int n,unsigned int k,bool reinterpret_input_as_3d,int depth_output_gemm3d)124 inline bool validate_lhs_rhs_info_reshaped_only_rhs(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output,
125 unsigned int m, unsigned int n, unsigned int k, bool reinterpret_input_as_3d, int depth_output_gemm3d)
126 {
127 // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel
128 TensorInfo tmp_b_info{};
129 // Validate reshape RHS kernel
130 auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
131 if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
132 {
133 return false;
134 }
135 // Validate mm kernel
136 // NOTE: Ignore all other parameters (eg. depth_output_gemm3d, output stage etc.) and only validate lhs and rhs info
137 // NOTE: This assumes:
138 // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_arguments).
139 // 2. lhs and rhs info does not cause window and padding issues through side effects (in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_and_configure_window).
140 GEMMKernelInfo gemm_kernel_info;
141 gemm_kernel_info.m = m;
142 gemm_kernel_info.n = n;
143 gemm_kernel_info.k = k;
144 gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
145 gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d;
146 gemm_kernel_info.lhs_info = lhs_info;
147 gemm_kernel_info.rhs_info = rhs_info;
148 // Since we ignore the output stage, output data type has to be S32 to pass the validation
149 TensorInfo output_info_copy(*output);
150 output_info_copy.set_data_type(DataType::S32);
151 if(!bool(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, &output_info_copy, gemm_kernel_info)))
152 {
153 return false;
154 }
155 return true;
156 }
157
158 // Validate lhs_info and rhs_info for reshaped only rhs kernel
validate_lhs_rhs_info_reshaped_only_rhs_mmul(const GEMMLHSMatrixInfo & lhs_info,const GEMMRHSMatrixInfo & rhs_info,const ITensorInfo * a,const ITensorInfo * b,const ITensorInfo * output,unsigned int m,unsigned int n,unsigned int k,bool reinterpret_input_as_3d,int depth_output_gemm3d)159 inline bool validate_lhs_rhs_info_reshaped_only_rhs_mmul(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output,
160 unsigned int m, unsigned int n, unsigned int k, bool reinterpret_input_as_3d, int depth_output_gemm3d)
161 {
162 // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel
163 TensorInfo tmp_b_info{};
164 // Validate reshape RHS kernel
165 auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info)));
166 if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)))
167 {
168 return false;
169 }
170 // Validate mm kernel
171 // NOTE: Ignore all other parameters (eg. depth_output_gemm3d, output stage etc.) and only validate lhs and rhs info
172 // NOTE: This assumes:
173 // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_arguments).
174 // 2. lhs and rhs info does not cause window and padding issues through side effects (in ClGemmLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_and_configure_window).
175 GEMMKernelInfo gemm_kernel_info;
176 gemm_kernel_info.m = m;
177 gemm_kernel_info.n = n;
178 gemm_kernel_info.k = k;
179 gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
180 gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d;
181 gemm_kernel_info.lhs_info = lhs_info;
182 gemm_kernel_info.rhs_info = rhs_info;
183 // Since we ignore the output stage, output data type has to be S32 to pass the validation
184 TensorInfo output_info_copy(*output);
185 output_info_copy.set_data_type(DataType::S32);
186 if(!bool(ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel::validate(a, &tmp_b_info, &output_info_copy, gemm_kernel_info)))
187 {
188 return false;
189 }
190 return true;
191 }
192
193 // Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs
auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query,bool reinterpret_input_as_3d,int depth_output_gemm3d,const ITensorInfo * a,const ITensorInfo * b,const ITensorInfo * output)194 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query, bool reinterpret_input_as_3d, int depth_output_gemm3d,
195 const ITensorInfo *a,
196 const ITensorInfo *b, const ITensorInfo *output)
197 {
198 auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(query);
199 if(config)
200 {
201 if(validate_lhs_rhs_info_reshaped_only_rhs(config.lhs_info, config.rhs_info, a, b, output, query.m, query.n, query.k, reinterpret_input_as_3d, depth_output_gemm3d))
202 {
203 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
204 return { config.lhs_info, config.rhs_info };
205 }
206 }
207 config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query);
208 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str());
209 return { config.lhs_info, config.rhs_info };
210 }
211
212 // Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs
auto_select_gemm_config_reshaped_only_rhs_mmul(auto_heuristics::CommonQuery query,bool reinterpret_input_as_3d,int depth_output_gemm3d,const ITensorInfo * a,const ITensorInfo * b,const ITensorInfo * output)213 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped_only_rhs_mmul(auto_heuristics::CommonQuery query, bool reinterpret_input_as_3d, int depth_output_gemm3d,
214 const ITensorInfo *a,
215 const ITensorInfo *b, const ITensorInfo *output)
216 {
217 ARM_COMPUTE_UNUSED(a, b, output, reinterpret_input_as_3d, depth_output_gemm3d);
218 auto config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query);
219 validate_lhs_rhs_info_reshaped_only_rhs_mmul(config.lhs_info, config.rhs_info, a, b, output, query.m, query.n, query.k, reinterpret_input_as_3d, depth_output_gemm3d);
220 ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs_mmul config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(),
221 to_string(config.rhs_info).c_str());
222 return { config.lhs_info, config.rhs_info };
223 }
224
is_gemm_reshaped(CLGEMMKernelType kernel_type)225 inline bool is_gemm_reshaped(CLGEMMKernelType kernel_type)
226 {
227 switch(kernel_type)
228 {
229 case CLGEMMKernelType::NATIVE:
230 return false;
231 case CLGEMMKernelType::RESHAPED_ONLY_RHS:
232 case CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL:
233 return true;
234 default:
235 ARM_COMPUTE_ERROR("Not supported gemmlowp kernel!");
236 }
237 }
238 } // namespace
239
ClGemmLowpMatrixMultiplyCore()240 ClGemmLowpMatrixMultiplyCore::ClGemmLowpMatrixMultiplyCore()
241 : _weights_to_qasymm8(std::make_unique<ClCastKernel>()),
242 _mm_native_kernel(std::make_unique<ClGemmLowpMatrixMultiplyNativeKernel>()),
243 _mm_reshaped_only_rhs_kernel(std::make_unique<ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel>()),
244 _mm_reshaped_only_rhs_mmul_kernel(std::make_unique<ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel>()),
245 _mtx_b_reshape_kernel(std::make_unique<ClGemmReshapeRhsMatrixKernel>()),
246 _mtx_a_reduction_kernel(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
247 _mtx_b_reduction_kernel(std::make_unique<ClGemmLowpMatrixBReductionKernel>()),
248 _offset_contribution_kernel(std::make_unique<ClGemmLowpOffsetContributionKernel>()),
249 _offset_contribution_output_stage_kernel(std::make_unique<ClGemmLowpOffsetContributionOutputStageKernel>()),
250 _aux_mem(AuxTensorIdx::Count)
251 {
252 }
253
254 ClGemmLowpMatrixMultiplyCore::~ClGemmLowpMatrixMultiplyCore() = default;
255
configure(const CLCompileContext & compile_context,ITensorInfo * a,ITensorInfo * b,ITensorInfo * c,ITensorInfo * output,const GEMMInfo & gemm_info)256 void ClGemmLowpMatrixMultiplyCore::configure(const CLCompileContext &compile_context,
257 ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output,
258 const GEMMInfo &gemm_info)
259 {
260 ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
261 ARM_COMPUTE_ERROR_THROW_ON(ClGemmLowpMatrixMultiplyCore::validate(a, b, c, output, gemm_info));
262 ARM_COMPUTE_LOG_PARAMS(a, b, c, output, gemm_info);
263
264 _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
265 _a_offset = a->quantization_info().uniform().offset;
266 _convert_to_qasymm8 = is_data_type_quantized_per_channel(b->data_type()) && is_data_type_quantized_symmetric(b->data_type())
267 && a->data_type() == DataType::QASYMM8;
268 _b_offset = _convert_to_qasymm8 ? -128 : b->quantization_info().uniform().offset;
269 _gemm_info = gemm_info;
270
271 // Get the GPU target
272 const GPUTarget gpu_target = CLScheduler::get().target();
273
274 // Set the target for the kernels
275 _mm_native_kernel->set_target(gpu_target);
276 _mm_reshaped_only_rhs_kernel->set_target(gpu_target);
277 _mm_reshaped_only_rhs_mmul_kernel->set_target(gpu_target);
278
279 GEMMRHSMatrixInfo rhs_info;
280 GEMMLHSMatrixInfo lhs_info;
281
282 // Arguments used by GEMMReshapeInfo
283 // in order to know how the matrices have been reshaped
284 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
285 const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
286 const unsigned int n = b->dimension(0);
287 const unsigned int k = a->dimension(0);
288 const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2);
289 const int depth_output_gemm3d = gemm_info.depth_output_gemm3d();
290
291 const auto reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d);
292
293 _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ gpu_target, a->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run);
294
295 if(_convert_to_qasymm8)
296 {
297 // Set data type for converted weights
298 _qasymm8_weights = *b;
299 _qasymm8_weights.set_data_type(DataType::QASYMM8);
300 _weights_to_qasymm8->configure(compile_context, b, &_qasymm8_weights, ConvertPolicy::WRAP);
301 }
302
303 ITensorInfo *matrix_b = _convert_to_qasymm8 ? &_qasymm8_weights : b;
304 if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS)
305 {
306 matrix_b = &_tmp_b;
307
308 // Pick up the GEMM configuration
309 // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration
310 std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, reinterpret_input_as_3d,
311 depth_output_gemm3d,
312 a, _convert_to_qasymm8 ? &_qasymm8_weights : b, output);
313
314 // Configure reshape RHS kernel
315 _mtx_b_reshape_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_tmp_b, rhs_info);
316 }
317 if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL)
318 {
319 matrix_b = &_tmp_b;
320
321 // Pick up the GEMM configuration
322 // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration
323 std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs_mmul(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, reinterpret_input_as_3d,
324 depth_output_gemm3d,
325 a, _convert_to_qasymm8 ? &_qasymm8_weights : b, output);
326
327 // Configure reshape RHS kernel
328 _mtx_b_reshape_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_tmp_b, rhs_info);
329 }
330
331 // Using default reduction info
332 const GEMMLowpReductionKernelInfo reduction_info {};
333
334 // Initialize matrix B reduction kernel only if _a_offset is not equal to 0
335 if(_a_offset != 0)
336 {
337 _vector_sum_col = TensorInfo(compute_reductionA_shape(*b), 1, DataType::S32);
338
339 // Configure Matrix B reduction kernel
340 _mtx_b_reduction_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_vector_sum_col, reduction_info);
341 }
342
343 // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0
344 if(_b_offset != 0)
345 {
346 _vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32);
347
348 // Configure matrix A reduction kernel
349 _mtx_a_reduction_kernel->configure(compile_context, a, &_vector_sum_row, reduction_info);
350 }
351
352 GEMMKernelInfo gemm_kernel_info;
353 gemm_kernel_info.m = m;
354 gemm_kernel_info.n = n;
355 gemm_kernel_info.k = k;
356 gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d;
357 gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
358 gemm_kernel_info.lhs_info = lhs_info;
359 gemm_kernel_info.rhs_info = rhs_info;
360 gemm_kernel_info.a_offset = _a_offset;
361 gemm_kernel_info.b_offset = _b_offset;
362 // If GEMMLowpOutputStage != NONE, fuse the offset contribution with the output stage
363 if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE)
364 {
365 // Configure offset contribution kernel
366 const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1;
367
368 _gemm_output_stage_multipliers = TensorInfo(TensorShape(num_filters), 1, DataType::S32);
369 _gemm_output_stage_shifts = TensorInfo(TensorShape(num_filters), 1, DataType::S32);
370
371 GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage();
372 gemmlowp_output_stage.output_data_type = a->data_type();
373 if(num_filters == 1)
374 {
375 // Per-channel quantization with OFM == 1 is equivalent to uniform quantization.
376 // Setting this flag to false prevents the kernel from adding useless padding to the output multipliers and shifts
377 gemmlowp_output_stage.is_quantized_per_channel = false;
378 }
379
380 gemm_kernel_info.output_stage = gemmlowp_output_stage;
381
382 if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS && gemmlowp_output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
383 {
384 // Configure and tune matrix multiply kernel with fused output stage
385 _mm_reshaped_only_rhs_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info, _a_offset == 0 ? nullptr : &_vector_sum_col,
386 _b_offset == 0 ? nullptr : &_vector_sum_row, c != nullptr ? c : nullptr, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts);
387 }
388 else if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL && gemmlowp_output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
389 {
390 // Configure and tune matrix multiply kernel with fused output stage
391 _mm_reshaped_only_rhs_mmul_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info, _a_offset == 0 ? nullptr : &_vector_sum_col,
392 _b_offset == 0 ? nullptr : &_vector_sum_row, c != nullptr ? c : nullptr, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts);
393 }
394 else
395 {
396 _run_output_stage = true;
397
398 if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS)
399 {
400 _mm_reshaped_only_rhs_kernel->configure(compile_context, a, matrix_b, &_mm_result_s32, gemm_kernel_info);
401 }
402 if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL)
403 {
404 _mm_reshaped_only_rhs_mmul_kernel->configure(compile_context, a, matrix_b, &_mm_result_s32, gemm_kernel_info);
405 }
406 else
407 {
408 // Pick up the GEMM configuration
409 // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration
410 std::tie(lhs_info, rhs_info) = auto_select_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size },
411 a, _convert_to_qasymm8 ? &_qasymm8_weights : matrix_b, reshape_info);
412
413 // Configure matrix multiply kernel
414 _mm_native_kernel->configure(compile_context, a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, reshape_info);
415
416 _offset_contribution_output_stage_kernel->configure(compile_context, &_mm_result_s32, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row,
417 c != nullptr ? c : nullptr, output, a->dimension(0), _a_offset, _b_offset, gemmlowp_output_stage,
418 &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts);
419 }
420 }
421 }
422 else
423 {
424 _run_offset_contribution = true;
425 if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS)
426 {
427 // Configure and tune matrix multiply kernel
428 _mm_reshaped_only_rhs_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info);
429 }
430 else if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL)
431 {
432 // Configure and tune matrix multiply kernel
433 _mm_reshaped_only_rhs_mmul_kernel->configure(compile_context, a, matrix_b, output, gemm_kernel_info);
434 }
435 else
436 {
437 // Pick up the GEMM configuration
438 // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration
439 std::tie(lhs_info, rhs_info) = auto_select_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size },
440 a, _convert_to_qasymm8 ? &_qasymm8_weights : b, reshape_info);
441
442 // Configure matrix multiply kernel
443 _mm_native_kernel->configure(compile_context, a, matrix_b, output, lhs_info, rhs_info, reshape_info);
444 }
445
446 // Configure offset contribution kernel
447 _offset_contribution_kernel->configure(compile_context, output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row,
448 c != nullptr ? c : nullptr, a->dimension(0), _a_offset, _b_offset);
449 }
450
451 // Request memory
452 _aux_mem[RhsQAsymm8] = MemoryInfo(offset_int_vec(RhsQAsymm8), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _qasymm8_weights.total_size());
453 if(is_gemm_reshaped(_gemm_kernel_type))
454 {
455 // Overwrite Rhs as prepare if gemm is reshaped as there will be a two-step transformation
456 _aux_mem[RhsQAsymm8] = MemoryInfo(offset_int_vec(RhsQAsymm8), _reshape_b_only_on_first_run ? MemoryLifetime::Prepare : MemoryLifetime::Temporary, _qasymm8_weights.total_size());
457 _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size());
458 }
459 if(_a_offset != 0)
460 {
461 _aux_mem[VecSumCol] = MemoryInfo(offset_int_vec(VecSumCol), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _vector_sum_col.total_size());
462 }
463 if(_b_offset != 0)
464 {
465 _aux_mem[VecSumRow] = MemoryInfo(offset_int_vec(VecSumRow), MemoryLifetime::Temporary, _vector_sum_row.total_size());
466 }
467 _aux_mem[ResultS32] = MemoryInfo(offset_int_vec(ResultS32), MemoryLifetime::Temporary, _mm_result_s32.total_size());
468 _aux_mem[Multipliers] = MemoryInfo(offset_int_vec(Multipliers), MemoryLifetime::Persistent, _gemm_output_stage_multipliers.total_size());
469 _aux_mem[Shifts] = MemoryInfo(offset_int_vec(Shifts), MemoryLifetime::Persistent, _gemm_output_stage_shifts.total_size());
470 }
471
validate(const ITensorInfo * a,const ITensorInfo * b,const ITensorInfo * c,const ITensorInfo * output,const GEMMInfo & gemm_info)472 Status ClGemmLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info)
473 {
474 ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
475 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
476 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8, DataType::QSYMM8_PER_CHANNEL);
477 ARM_COMPUTE_RETURN_ERROR_ON(a->data_type() == DataType::QASYMM8 && b->data_type() == DataType::QASYMM8_SIGNED);
478 ARM_COMPUTE_RETURN_ERROR_ON(a->data_type() == DataType::QASYMM8_SIGNED && b->data_type() == DataType::QASYMM8);
479 ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
480 ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
481
482 int32_t a_offset = a->quantization_info().uniform().offset;
483 int32_t b_offset = b->quantization_info().uniform().offset;
484
485 const ITensorInfo *matrix_a_info = a;
486
487 TensorInfo tmp_b_info{};
488 GEMMRHSMatrixInfo rhs_info;
489 GEMMLHSMatrixInfo lhs_info;
490
491 // Get the GPU target
492 const GPUTarget gpu_target = CLScheduler::get().target();
493
494 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
495 const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
496 const unsigned int n = b->dimension(0);
497 const unsigned int k = a->dimension(0);
498 const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2);
499 const int depth_output_gemm3d = gemm_info.depth_output_gemm3d();
500
501 bool reshape_matrix_b = is_gemm_reshaped(auto_select_gemm_kernel(auto_heuristics::CommonQuery{ gpu_target, a->data_type(), m, n, k, batch_size }, gemm_info.reshape_b_only_on_first_run()));
502
503 const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d);
504
505 bool convert_to_qasymm8 = is_data_type_quantized_per_channel(b->data_type()) && is_data_type_quantized_symmetric(b->data_type())
506 && is_data_type_quantized_asymmetric(a->data_type());
507 TensorInfo weights_info(*b);
508 if(convert_to_qasymm8)
509 {
510 b_offset = -128;
511 weights_info.set_data_type(DataType::QASYMM8);
512 ARM_COMPUTE_RETURN_ON_ERROR(ClCastKernel::validate(b, &weights_info, ConvertPolicy::WRAP));
513 }
514 const ITensorInfo *matrix_b_info = &weights_info;
515 if(reshape_matrix_b)
516 {
517 matrix_b_info = &tmp_b_info;
518
519 // Pick up the GEMM configuration
520 // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails
521 // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration
522 const auto res = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size });
523 lhs_info = res.lhs_info;
524 rhs_info = res.rhs_info;
525
526 // Validate reshape RHS kernel
527 auto_init_if_empty(tmp_b_info, weights_info.clone()->set_tensor_shape(compute_rhs_reshaped_shape(weights_info, rhs_info)));
528 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(&weights_info, &tmp_b_info, rhs_info));
529 }
530
531 TensorInfo info_vector_sum_col{};
532 TensorInfo info_vector_sum_row{};
533
534 const GEMMLowpReductionKernelInfo reduction_info;
535 // Validate matrix B reduction kernel only if _a_offset is not equal to 0
536 if(a_offset != 0)
537 {
538 info_vector_sum_col = TensorInfo(compute_reductionA_shape(weights_info), 1, DataType::S32);
539
540 // Configure Matrix B reduction kernel
541 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixBReductionKernel::validate(&weights_info, &info_vector_sum_col, reduction_info));
542 }
543
544 // Validate Matrix A reduction kernel only if _b_offset is not equal to 0
545 if(b_offset != 0)
546 {
547 info_vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32);
548
549 // Configure matrix A reduction kernel
550 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row, reduction_info));
551 }
552
553 GEMMKernelInfo gemm_kernel_info;
554 gemm_kernel_info.m = m;
555 gemm_kernel_info.n = n;
556 gemm_kernel_info.k = k;
557 gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d;
558 gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d;
559 gemm_kernel_info.lhs_info = lhs_info;
560 gemm_kernel_info.rhs_info = rhs_info;
561 gemm_kernel_info.a_offset = a_offset;
562 gemm_kernel_info.b_offset = b_offset;
563 if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE)
564 {
565 const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1;
566
567 const TensorInfo gemm_output_stage_multipliers_shifts_info(TensorInfo(TensorShape(num_filters), 1, DataType::S32));
568
569 GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage();
570 gemmlowp_output_stage.output_data_type = a->data_type();
571
572 gemm_kernel_info.output_stage = gemmlowp_output_stage;
573 if(reshape_matrix_b && gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT)
574 {
575 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info,
576 a_offset == 0 ? nullptr : &info_vector_sum_col,
577 b_offset == 0 ? nullptr : &info_vector_sum_row,
578 c,
579 &gemm_output_stage_multipliers_shifts_info,
580 &gemm_output_stage_multipliers_shifts_info));
581 }
582 else
583 {
584 TensorInfo mm_result_s32_info{};
585
586 if(reshape_matrix_b)
587 {
588 // Output tensor auto inizialitation if not yet initialized
589 auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_info)).set_data_type(DataType::S32));
590
591 // Validate matrix multiply
592 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, gemm_kernel_info));
593 }
594 else
595 {
596 // Output tensor auto inizialitation if not yet initialized
597 auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, false, reshape_info)).set_data_type(DataType::S32));
598
599 // Pick up the GEMM configuration
600 // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails
601 // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration
602 const auto res = select_default_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size });
603 lhs_info = res.lhs_info;
604 rhs_info = res.rhs_info;
605
606 // Validate matrix multiply
607 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info));
608 }
609
610 // Validate offset contribution kernel
611 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpOffsetContributionOutputStageKernel::validate(&mm_result_s32_info,
612 a_offset == 0 ? nullptr : &info_vector_sum_col,
613 b_offset == 0 ? nullptr : &info_vector_sum_row,
614 c,
615 output,
616 a_offset, b_offset,
617 gemmlowp_output_stage,
618 &gemm_output_stage_multipliers_shifts_info,
619 &gemm_output_stage_multipliers_shifts_info));
620 }
621 }
622 else
623 {
624 if(reshape_matrix_b)
625 {
626 // Validate matrix multiply
627 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info));
628 }
629 else
630 {
631 // Pick up the GEMM configuration
632 // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration
633 const auto res = select_default_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size });
634 lhs_info = res.lhs_info;
635 rhs_info = res.rhs_info;
636
637 // Validate matrix multiply
638 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info));
639 }
640
641 if(output->total_size() != 0)
642 {
643 // Validate offset contribution kernel
644 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpOffsetContributionKernel::validate(output,
645 a_offset == 0 ? nullptr : &info_vector_sum_col,
646 b_offset == 0 ? nullptr : &info_vector_sum_row,
647 c,
648 a_offset, b_offset));
649 }
650 }
651
652 return Status{};
653 }
654
run(ITensorPack & tensors)655 void ClGemmLowpMatrixMultiplyCore::run(ITensorPack &tensors)
656 {
657 const ITensor *a = tensors.get_const_tensor(ACL_SRC_0);
658 const ITensor *b = tensors.get_const_tensor(ACL_SRC_1);
659 const ITensor *c = tensors.get_const_tensor(ACL_SRC_2);
660 ITensor *dst = tensors.get_tensor(ACL_DST);
661
662 ARM_COMPUTE_ERROR_ON_NULLPTR(a, dst);
663
664 CLAuxTensorHandler vec_sum_col(offset_int_vec(VecSumCol), _vector_sum_col, tensors, true);
665 CLAuxTensorHandler vec_sum_row(offset_int_vec(VecSumRow), _vector_sum_row, tensors, true);
666 CLAuxTensorHandler rhs_qasymm8(offset_int_vec(RhsQAsymm8), _qasymm8_weights, tensors, true);
667 CLAuxTensorHandler tmp_b(offset_int_vec(RhsReshape), _tmp_b, tensors, true);
668 CLAuxTensorHandler res32(offset_int_vec(ResultS32), _mm_result_s32, tensors, true);
669 CLAuxTensorHandler shifts(offset_int_vec(Shifts), _gemm_output_stage_shifts, tensors, true);
670 CLAuxTensorHandler multipliers(offset_int_vec(Multipliers), _gemm_output_stage_multipliers, tensors, true);
671
672 // Prepare the consts if needed
673 prepare(tensors);
674
675 const ITensor *matrix_a = a;
676 const ITensor *matrix_b = _convert_to_qasymm8 ? rhs_qasymm8.get() : b;
677
678 if(is_gemm_reshaped(_gemm_kernel_type))
679 {
680 matrix_b = tmp_b.get();
681 if(!_reshape_b_only_on_first_run)
682 {
683 // Run reshape matrix B
684 ITensorPack mtx_b_reshape_pack =
685 {
686 { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b },
687 { TensorType::ACL_DST, tmp_b.get() }
688 };
689 CLScheduler::get().enqueue_op(*_mtx_b_reshape_kernel, mtx_b_reshape_pack, false);
690 }
691 }
692
693 // Run matrix B reduction kernel only if _a_offset is not equal to 0
694 if(_a_offset != 0 && !_reshape_b_only_on_first_run)
695 {
696 ITensorPack mtx_b_red_pack =
697 {
698 { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b },
699 { TensorType::ACL_DST, vec_sum_col.get() }
700 };
701 CLScheduler::get().enqueue_op(*_mtx_b_reduction_kernel, mtx_b_red_pack, false);
702 }
703
704 // Run matrix A reduction kernel only if _b_offset is not equal to 0
705 if(_b_offset != 0)
706 {
707 ITensorPack mtx_a_red_pack =
708 {
709 { TensorType::ACL_SRC, matrix_a },
710 { TensorType::ACL_DST, vec_sum_row.get() }
711 };
712 CLScheduler::get().enqueue_op(*_mtx_a_reduction_kernel, mtx_a_red_pack, false);
713 }
714
715 // Run matrix multiply
716 if(is_gemm_reshaped(_gemm_kernel_type))
717 {
718 ITensorPack gemm_reshaped_pack;
719 if(_run_offset_contribution)
720 {
721 gemm_reshaped_pack = ITensorPack({ { TensorType::ACL_SRC_0, matrix_a },
722 { TensorType::ACL_SRC_1, matrix_b },
723 { TensorType::ACL_DST, _run_output_stage ? res32.get() : dst }
724 });
725 }
726 else
727 {
728 gemm_reshaped_pack = ITensorPack(
729 {
730 { TensorType::ACL_SRC, matrix_a },
731 { TensorType::ACL_SRC_1, matrix_b },
732 { TensorType::ACL_BIAS, c },
733 { TensorType::ACL_VEC_ROW_SUM, _b_offset == 0 ? nullptr : vec_sum_row.get() },
734 { TensorType::ACL_VEC_COL_SUM, _a_offset == 0 ? nullptr : vec_sum_col.get() },
735 { TensorType::ACL_SHIFTS, shifts.get() },
736 { TensorType::ACL_MULTIPLIERS, multipliers.get() },
737 { TensorType::ACL_DST, dst },
738 });
739 }
740 if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS)
741 {
742 CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_kernel, gemm_reshaped_pack, false);
743 }
744 else if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL)
745 {
746 CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_mmul_kernel, gemm_reshaped_pack, false);
747 }
748 else
749 {
750 ARM_COMPUTE_ERROR("Invalid reshaped kernel");
751 }
752 }
753 else
754 {
755 ITensorPack gemm_native_pack =
756 {
757 { TensorType::ACL_SRC_0, matrix_a },
758 { TensorType::ACL_SRC_1, matrix_b },
759 { TensorType::ACL_DST, _run_offset_contribution ? dst : res32.get() }
760 };
761 CLScheduler::get().enqueue_op(*_mm_native_kernel, gemm_native_pack, false);
762 }
763 if(_run_output_stage)
764 {
765 // Run offset contribution/output stage kernel
766 ITensorPack output_stage_pack =
767 {
768 { TensorType::ACL_SRC, res32.get() },
769 { TensorType::ACL_BIAS, c },
770 { TensorType::ACL_VEC_ROW_SUM, _b_offset == 0 ? nullptr : vec_sum_row.get() },
771 { TensorType::ACL_VEC_COL_SUM, _a_offset == 0 ? nullptr : vec_sum_col.get() },
772 { TensorType::ACL_SHIFTS, shifts.get() },
773 { TensorType::ACL_MULTIPLIERS, multipliers.get() },
774 { TensorType::ACL_DST, dst },
775 };
776 CLScheduler::get().enqueue_op(*_offset_contribution_output_stage_kernel, output_stage_pack, true);
777 }
778 if(_run_offset_contribution)
779 {
780 // Run offset contribution kernel
781 ITensorPack offset_contrib_pack =
782 {
783 { TensorType::ACL_SRC_DST, dst },
784 { TensorType::ACL_BIAS, c },
785 { TensorType::ACL_VEC_ROW_SUM, _b_offset == 0 ? nullptr : vec_sum_row.get() },
786 { TensorType::ACL_VEC_COL_SUM, _a_offset == 0 ? nullptr : vec_sum_col.get() }
787 };
788 CLScheduler::get().enqueue_op(*_offset_contribution_kernel, offset_contrib_pack, true);
789 }
790 }
791
prepare(ITensorPack & tensors)792 void ClGemmLowpMatrixMultiplyCore::prepare(ITensorPack &tensors)
793 {
794 if(!_is_prepared)
795 {
796 auto b = tensors.get_const_tensor(TensorType::ACL_SRC_1);
797 CLAuxTensorHandler tmp_b(offset_int_vec(RhsReshape), _tmp_b, tensors, true);
798 CLAuxTensorHandler vec_sum_col(offset_int_vec(VecSumCol), _vector_sum_col, tensors, true);
799 CLAuxTensorHandler rhs_qasymm8(offset_int_vec(RhsQAsymm8), _qasymm8_weights, tensors, false);
800
801 ARM_COMPUTE_ERROR_ON_NULLPTR(b);
802
803 if(_convert_to_qasymm8)
804 {
805 ITensorPack convert_to_qs8_pack = { { ACL_SRC, b }, { ACL_DST, rhs_qasymm8.get() } };
806 CLScheduler::get().enqueue_op(*_weights_to_qasymm8, convert_to_qs8_pack, false);
807 b->mark_as_unused();
808 }
809
810 if(is_gemm_reshaped(_gemm_kernel_type) && _reshape_b_only_on_first_run)
811 {
812 // Run reshape kernel and mark original weights tensor as unused
813 ITensorPack mtx_b_pack =
814 {
815 { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b },
816 { TensorType::ACL_DST, tmp_b.get() }
817 };
818 CLScheduler::get().enqueue_op(*_mtx_b_reshape_kernel, mtx_b_pack, false);
819 b->mark_as_unused();
820 }
821
822 // Run matrix B reduction kernel only if _a_offset is not equal to 0
823 if(_a_offset != 0 && _reshape_b_only_on_first_run)
824 {
825 ITensorPack mtx_b_red_pack =
826 {
827 { TensorType::ACL_SRC, _convert_to_qasymm8 ? rhs_qasymm8.get() : b },
828 { TensorType::ACL_DST, vec_sum_col.get() }
829 };
830 CLScheduler::get().enqueue_op(*_mtx_b_reduction_kernel, mtx_b_red_pack, false);
831 }
832
833 // Compute GEMM output multipliers and shifts for output stage
834 {
835 const size_t num_filters = (_gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? _gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1;
836
837 CLAuxTensorHandler multipliers(offset_int_vec(Multipliers), _gemm_output_stage_multipliers, tensors, false);
838 CLAuxTensorHandler shifts(offset_int_vec(Shifts), _gemm_output_stage_shifts, tensors, false);
839
840 ICLTensor *multiplier_tensor = multipliers.get();
841 if(multiplier_tensor != nullptr && multiplier_tensor->info()->total_size() > 0)
842 {
843 multiplier_tensor->map(CLScheduler::get().queue(), true);
844 std::memcpy(multiplier_tensor->ptr_to_element(Coordinates(0)), _gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.data(), num_filters * sizeof(int32_t));
845 multiplier_tensor->unmap(CLScheduler::get().queue());
846 }
847
848 ICLTensor *shifts_tensor = shifts.get();
849 if(shifts.get() != nullptr && shifts_tensor->info()->total_size() > 0)
850 {
851 shifts_tensor->map(CLScheduler::get().queue(), true);
852 std::memcpy(shifts_tensor->ptr_to_element(Coordinates(0)), _gemm_info.gemmlowp_output_stage().gemmlowp_shifts.data(), num_filters * sizeof(int32_t));
853 shifts_tensor->unmap(CLScheduler::get().queue());
854 }
855 }
856 CLScheduler::get().queue().finish();
857 _is_prepared = true;
858 }
859 }
860
workspace() const861 experimental::MemoryRequirements ClGemmLowpMatrixMultiplyCore::workspace() const
862 {
863 return _aux_mem;
864 }
865 } // namespace opencl
866 } // namespace arm_compute
867