2 * Copyright (c) 2017-2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/Helpers.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Types.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
33 #include "arm_compute/runtime/CL/CLScheduler.h"
35 using namespace arm_compute;
36 using namespace arm_compute::misc::shape_calculator;
40 inline bool is_interleaved_transposed(int m, int n, int k, bool reshape_b_only_on_first_run, GPUTarget gpu_target)
44 if(gpu_target == GPUTarget::BIFROST)
47 if(k > 256 && m > 4 && reshape_b_only_on_first_run)
49 flag = ((0.72f + n * 0.10766f) < (n * 0.1284f));
61 CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager)
62 : _memory_group(std::move(memory_manager)), _mm_kernel(), _mtx_a_reshape_kernel(), _mtx_b_reshape_kernel(), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _offset_contribution_kernel(),
63 _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _a_offset(0), _b_offset(0), _is_interleaved_transposed(true), _is_first_run(true), _reshape_b_only_on_first_run(false)
67 void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output, const GEMMInfo &gemm_info)
69 ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
70 ARM_COMPUTE_UNUSED(gemm_info);
71 ARM_COMPUTE_ERROR_THROW_ON(CLGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), output->info(), gemm_info));
73 _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
74 _a_offset = a->info()->quantization_info().offset;
75 _b_offset = b->info()->quantization_info().offset;
78 const GPUTarget gpu_target = CLScheduler::get().target();
80 // Set the target for the kernels
81 _mtx_a_reshape_kernel.set_target(gpu_target);
82 _mm_kernel.set_target(gpu_target);
84 const ICLTensor *matrix_a = a;
85 const ICLTensor *matrix_b = b;
87 // Arguments used by GEMMReshapeInfo
88 // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
89 // in order to know how the matrices have been reshaped
90 const int m = a->info()->dimension(1);
91 const int n = b->info()->dimension(0);
92 const int k = a->info()->dimension(0);
93 constexpr int mult_transpose1xW_width = 1;
94 constexpr int mult_interleave4x4_height = 1;
96 // Check if we need to reshape the matrix A and matrix B
97 _is_interleaved_transposed = is_interleaved_transposed(m, n, k, _reshape_b_only_on_first_run, gpu_target);
99 if(_is_interleaved_transposed)
104 _memory_group.manage(&_tmp_a);
105 _memory_group.manage(&_tmp_b);
107 // Configure interleave kernel
108 _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height);
110 // Configure transpose kernel
111 _mtx_b_reshape_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
114 // Configure matrix multiply kernel
115 _mm_kernel.configure(matrix_a, matrix_b, output, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height));
117 // Initialize matrix B reduction kernel only if _a_offset is not equal to 0
120 TensorInfo info_vector_sum_col(compute_reductionA_shape(*b->info()), 1, DataType::S32);
121 _vector_sum_col.allocator()->init(info_vector_sum_col);
122 _memory_group.manage(&_vector_sum_col);
124 // Configure Matrix B reduction kernel
125 _mtx_b_reduction_kernel.configure(b, &_vector_sum_col);
128 // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0
131 TensorInfo info_vector_sum_row(compute_reductionB_shape(*a->info()), 1, DataType::S32);
132 _vector_sum_row.allocator()->init(info_vector_sum_row);
133 _memory_group.manage(&_vector_sum_row);
135 // Configure matrix A reduction kernel
136 _mtx_a_reduction_kernel.configure(a, &_vector_sum_row);
139 // Configure offset contribution kernel
140 _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, a->info()->dimension(0), _a_offset, _b_offset);
143 if(_is_interleaved_transposed)
145 _tmp_a.allocator()->allocate();
146 _tmp_b.allocator()->allocate();
151 _vector_sum_col.allocator()->allocate();
156 _vector_sum_row.allocator()->allocate();
160 Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, const GEMMInfo &gemm_info)
162 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8);
163 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
164 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
165 ARM_COMPUTE_RETURN_ERROR_ON_MSG((a)->dimension(0) != (b)->dimension(1),
166 "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
167 ARM_COMPUTE_RETURN_ERROR_ON_MSG((a)->dimension(1) != (output)->dimension(1),
168 "The output matrix must have the same number of rows as the matrix A");
169 ARM_COMPUTE_RETURN_ERROR_ON_MSG((b)->dimension(0) != (output)->dimension(0),
170 "The output matrix must have the same number of columns as the matrix B");
171 ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
172 ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
174 int32_t a_offset = a->quantization_info().offset;
175 int32_t b_offset = b->quantization_info().offset;
177 const int m = a->dimension(1);
178 const int n = b->dimension(0);
179 const int k = a->dimension(0);
180 constexpr int mult_transpose1xW_width = 1;
181 constexpr int mult_interleave4x4_height = 1;
182 const GEMMReshapeInfo reshape_info(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height);
184 bool reshape_matrices = is_interleaved_transposed(m, n, k, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
188 TensorInfo info_a(compute_interleaved_shape(*a, mult_interleave4x4_height), 1, a->data_type());
189 TensorInfo info_b(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width), 1, b->data_type());
191 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &info_a, mult_interleave4x4_height));
192 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &info_b, mult_transpose1xW_width));
193 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(&info_a, &info_b, output, reshape_matrices, reshape_info));
197 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(a, b, output, reshape_matrices, reshape_info));
200 TensorInfo info_vector_sum_col, info_vector_sum_row;
202 // Validate matrix B reduction kernel only if _a_offset is not equal to 0
205 info_vector_sum_col = TensorInfo(compute_reductionA_shape(*b), 1, DataType::S32);
207 // Configure Matrix B reduction kernel
208 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixBReductionKernel::validate(b, &info_vector_sum_col));
211 // Validate Matrix A reduction kernel only if _b_offset is not equal to 0
214 info_vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32);
216 // Configure matrix A reduction kernel
217 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row));
220 // Validate offset contribution kernel
221 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionKernel::validate(output,
222 a_offset == 0 ? nullptr : &info_vector_sum_col,
223 b_offset == 0 ? nullptr : &info_vector_sum_row,
224 a_offset, b_offset));
229 void CLGEMMLowpMatrixMultiplyCore::run()
231 _memory_group.acquire();
233 if(_is_interleaved_transposed)
235 // Run reshape matrix A
236 CLScheduler::get().enqueue(_mtx_a_reshape_kernel, false);
238 if(_is_first_run || !_reshape_b_only_on_first_run)
240 // Run reshape matrix B
241 CLScheduler::get().enqueue(_mtx_b_reshape_kernel, false);
245 // Note: if _reshape_b_only_on_first_run = true, the reduction kernel can be executed only once
246 if(_is_first_run || !_reshape_b_only_on_first_run)
248 // Run matrix B reduction kernel only if _a_offset is not equal to 0
251 CLScheduler::get().enqueue(_mtx_b_reduction_kernel, false);
255 // Run matrix multiply
256 CLScheduler::get().enqueue(_mm_kernel, false);
258 // Run matrix A reduction kernel only if _b_offset is not equal to 0
261 CLScheduler::get().enqueue(_mtx_a_reduction_kernel, false);
264 // Run offset contribution kernel
265 CLScheduler::get().enqueue(_offset_contribution_kernel, true);
267 _memory_group.release();
269 _is_first_run = false;